diff --git a/.devops/cpu.Dockerfile b/.devops/cpu.Dockerfile index 522ee8147..aa2aa0312 100644 --- a/.devops/cpu.Dockerfile +++ b/.devops/cpu.Dockerfile @@ -14,9 +14,9 @@ WORKDIR /app COPY . . RUN if [ "$TARGETARCH" = "amd64" ]; then \ - cmake -S . -B build -DCMAKE_BUILD_TYPE=Release -DLLAMA_CURL=ON -DGGML_NATIVE=OFF -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON; \ + cmake -S . -B build -DCMAKE_BUILD_TYPE=Release -DGGML_NATIVE=OFF -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON; \ elif [ "$TARGETARCH" = "arm64" ]; then \ - cmake -S . -B build -DCMAKE_BUILD_TYPE=Release -DLLAMA_CURL=ON -DGGML_NATIVE=OFF -DGGML_CPU_ARM_ARCH=${GGML_CPU_ARM_ARCH}; \ + cmake -S . -B build -DCMAKE_BUILD_TYPE=Release -DGGML_NATIVE=OFF -DGGML_CPU_ARM_ARCH=${GGML_CPU_ARM_ARCH}; \ else \ echo "Unsupported architecture"; \ exit 1; \ diff --git a/.devops/cuda.Dockerfile b/.devops/cuda.Dockerfile index 974dd78a8..8ae57d2e2 100644 --- a/.devops/cuda.Dockerfile +++ b/.devops/cuda.Dockerfile @@ -1,6 +1,6 @@ ARG UBUNTU_VERSION=22.04 # This needs to generally match the container host's environment. -ARG CUDA_VERSION=12.6.0 +ARG CUDA_VERSION=12.4.0 # Target the CUDA build image ARG BASE_CUDA_DEV_CONTAINER=nvidia/cuda:${CUDA_VERSION}-devel-ubuntu${UBUNTU_VERSION} @@ -21,7 +21,7 @@ COPY . . RUN if [ "${CUDA_DOCKER_ARCH}" != "default" ]; then \ export CMAKE_ARGS="-DCMAKE_CUDA_ARCHITECTURES=${CUDA_DOCKER_ARCH}"; \ fi && \ - cmake -B build -DGGML_NATIVE=OFF -DGGML_CUDA=ON -DLLAMA_CURL=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \ + cmake -B build -DGGML_NATIVE=OFF -DGGML_CUDA=ON -DLLAMA_CURL=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \ cmake --build build --config Release -j$(nproc) RUN mkdir -p /app/lib && \ diff --git a/.devops/intel.Dockerfile b/.devops/intel.Dockerfile index af783f5e9..091e1dc5d 100644 --- a/.devops/intel.Dockerfile +++ b/.devops/intel.Dockerfile @@ -17,7 +17,7 @@ RUN if [ "${GGML_SYCL_F16}" = "ON" ]; then \ && export OPT_SYCL_F16="-DGGML_SYCL_F16=ON"; \ fi && \ echo "Building with dynamic libs" && \ - cmake -B build -DGGML_NATIVE=OFF -DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA_CURL=ON ${OPT_SYCL_F16} && \ + cmake -B build -DGGML_NATIVE=OFF -DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA_CURL=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON ${OPT_SYCL_F16} && \ cmake --build build --config Release -j$(nproc) RUN mkdir -p /app/lib && \ diff --git a/.devops/llama-cli-cann.Dockerfile b/.devops/llama-cli-cann.Dockerfile index 02dce501c..0eb1af87c 100644 --- a/.devops/llama-cli-cann.Dockerfile +++ b/.devops/llama-cli-cann.Dockerfile @@ -1,4 +1,4 @@ -ARG ASCEND_VERSION=8.0.rc2.alpha003-910b-openeuler22.03-py3.8 +ARG ASCEND_VERSION=8.1.RC1.alpha001-910b-openeuler22.03-py3.10 FROM ascendai/cann:$ASCEND_VERSION AS build @@ -6,7 +6,7 @@ WORKDIR /app COPY . . -RUN yum install -y gcc g++ cmake make +RUN yum install -y gcc g++ cmake make libcurl-devel ENV ASCEND_TOOLKIT_HOME=/usr/local/Ascend/ascend-toolkit/latest ENV LIBRARY_PATH=${ASCEND_TOOLKIT_HOME}/lib64:$LIBRARY_PATH ENV LD_LIBRARY_PATH=${ASCEND_TOOLKIT_HOME}/lib64:${ASCEND_TOOLKIT_HOME}/lib64/plugin/opskernel:${ASCEND_TOOLKIT_HOME}/lib64/plugin/nnengine:${ASCEND_TOOLKIT_HOME}/opp/built-in/op_impl/ai_core/tbe/op_tiling:${LD_LIBRARY_PATH} diff --git a/.devops/llama-cpp-cuda.srpm.spec b/.devops/llama-cpp-cuda.srpm.spec index 7425d3a9d..3bbf4a4de 100644 --- a/.devops/llama-cpp-cuda.srpm.spec +++ b/.devops/llama-cpp-cuda.srpm.spec @@ -17,10 +17,10 @@ Version: %( date "+%%Y%%m%%d" ) Release: 1%{?dist} Summary: CPU Inference of LLaMA model in pure C/C++ (no CUDA/OpenCL) License: MIT -Source0: https://github.com/ggerganov/llama.cpp/archive/refs/heads/master.tar.gz +Source0: https://github.com/ggml-org/llama.cpp/archive/refs/heads/master.tar.gz BuildRequires: coreutils make gcc-c++ git cuda-toolkit Requires: cuda-toolkit -URL: https://github.com/ggerganov/llama.cpp +URL: https://github.com/ggml-org/llama.cpp %define debug_package %{nil} %define source_date_epoch_from_changelog 0 diff --git a/.devops/llama-cpp.srpm.spec b/.devops/llama-cpp.srpm.spec index 4d5560089..45902dcf8 100644 --- a/.devops/llama-cpp.srpm.spec +++ b/.devops/llama-cpp.srpm.spec @@ -18,10 +18,10 @@ Version: %( date "+%%Y%%m%%d" ) Release: 1%{?dist} Summary: CPU Inference of LLaMA model in pure C/C++ (no CUDA/OpenCL) License: MIT -Source0: https://github.com/ggerganov/llama.cpp/archive/refs/heads/master.tar.gz +Source0: https://github.com/ggml-org/llama.cpp/archive/refs/heads/master.tar.gz BuildRequires: coreutils make gcc-c++ git libstdc++-devel Requires: libstdc++ -URL: https://github.com/ggerganov/llama.cpp +URL: https://github.com/ggml-org/llama.cpp %define debug_package %{nil} %define source_date_epoch_from_changelog 0 diff --git a/.devops/musa.Dockerfile b/.devops/musa.Dockerfile index bfd7fc1c1..261a2823a 100644 --- a/.devops/musa.Dockerfile +++ b/.devops/musa.Dockerfile @@ -1,6 +1,6 @@ ARG UBUNTU_VERSION=22.04 # This needs to generally match the container host's environment. -ARG MUSA_VERSION=rc3.1.0 +ARG MUSA_VERSION=rc3.1.1 # Target the MUSA build image ARG BASE_MUSA_DEV_CONTAINER=mthreads/musa:${MUSA_VERSION}-devel-ubuntu${UBUNTU_VERSION} @@ -35,7 +35,7 @@ COPY . . RUN if [ "${MUSA_DOCKER_ARCH}" != "default" ]; then \ export CMAKE_ARGS="-DMUSA_ARCHITECTURES=${MUSA_DOCKER_ARCH}"; \ fi && \ - cmake -B build -DGGML_NATIVE=OFF -DGGML_MUSA=ON -DLLAMA_CURL=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \ + cmake -B build -DGGML_NATIVE=OFF -DGGML_MUSA=ON -DLLAMA_CURL=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \ cmake --build build --config Release -j$(nproc) RUN mkdir -p /app/lib && \ diff --git a/.devops/nix/package.nix b/.devops/nix/package.nix index 043c4364b..6e8050a49 100644 --- a/.devops/nix/package.nix +++ b/.devops/nix/package.nix @@ -133,12 +133,12 @@ effectiveStdenv.mkDerivation (finalAttrs: { --replace '[bundle pathForResource:@"default" ofType:@"metallib"];' "@\"$out/bin/default.metallib\";" ''; - # With PR#6015 https://github.com/ggerganov/llama.cpp/pull/6015, + # With PR#6015 https://github.com/ggml-org/llama.cpp/pull/6015, # `default.metallib` may be compiled with Metal compiler from XCode # and we need to escape sandbox on MacOS to access Metal compiler. # `xcrun` is used find the path of the Metal compiler, which is varible # and not on $PATH - # see https://github.com/ggerganov/llama.cpp/pull/6118 for discussion + # see https://github.com/ggml-org/llama.cpp/pull/6118 for discussion __noChroot = effectiveStdenv.isDarwin && useMetalKit && precompileMetalShaders; nativeBuildInputs = @@ -220,7 +220,7 @@ effectiveStdenv.mkDerivation (finalAttrs: { broken = (useMetalKit && !effectiveStdenv.isDarwin); description = "Inference of LLaMA model in pure C/C++${descriptionSuffix}"; - homepage = "https://github.com/ggerganov/llama.cpp/"; + homepage = "https://github.com/ggml-org/llama.cpp/"; license = lib.licenses.mit; # Accommodates `nix run` and `lib.getExe` diff --git a/.devops/rocm.Dockerfile b/.devops/rocm.Dockerfile index a8088ea00..a1b34723a 100644 --- a/.devops/rocm.Dockerfile +++ b/.devops/rocm.Dockerfile @@ -11,14 +11,14 @@ ARG BASE_ROCM_DEV_CONTAINER=rocm/dev-ubuntu-${UBUNTU_VERSION}:${ROCM_VERSION}-co FROM ${BASE_ROCM_DEV_CONTAINER} AS build # Unless otherwise specified, we make a fat build. -# List from https://github.com/ggerganov/llama.cpp/pull/1087#issuecomment-1682807878 +# List from https://github.com/ggml-org/llama.cpp/pull/1087#issuecomment-1682807878 # This is mostly tied to rocBLAS supported archs. # gfx803, gfx900, gfx1032, gfx1101, gfx1102,not officialy supported # gfx906 is deprecated #check https://rocm.docs.amd.com/projects/install-on-linux/en/docs-6.2.4/reference/system-requirements.html -#ARG ROCM_DOCKER_ARCH='gfx803,gfx900,gfx906,gfx908,gfx90a,gfx942,gfx1010,gfx1030,gfx1032,gfx1100,gfx1101,gfx1102' -ARG ROCM_DOCKER_ARCH=gfx1100 +ARG ROCM_DOCKER_ARCH='gfx803,gfx900,gfx906,gfx908,gfx90a,gfx942,gfx1010,gfx1030,gfx1032,gfx1100,gfx1101,gfx1102' +#ARG ROCM_DOCKER_ARCH=gfx1100 # Set nvcc architectured ENV AMDGPU_TARGETS=${ROCM_DOCKER_ARCH} @@ -40,7 +40,7 @@ WORKDIR /app COPY . . RUN HIPCXX="$(hipconfig -l)/clang" HIP_PATH="$(hipconfig -R)" \ - cmake -S . -B build -DGGML_HIP=ON -DAMDGPU_TARGETS=$ROCM_DOCKER_ARCH -DCMAKE_BUILD_TYPE=Release -DLLAMA_CURL=ON \ + cmake -S . -B build -DGGML_HIP=ON -DAMDGPU_TARGETS=$ROCM_DOCKER_ARCH -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DCMAKE_BUILD_TYPE=Release -DLLAMA_CURL=ON \ && cmake --build build --config Release -j$(nproc) RUN mkdir -p /app/lib \ diff --git a/.devops/vulkan.Dockerfile b/.devops/vulkan.Dockerfile index 9064f3838..f8f3072e9 100644 --- a/.devops/vulkan.Dockerfile +++ b/.devops/vulkan.Dockerfile @@ -16,7 +16,7 @@ WORKDIR /app COPY . . -RUN cmake -B build -DGGML_NATIVE=OFF -DGGML_VULKAN=1 -DLLAMA_CURL=1 && \ +RUN cmake -B build -DGGML_NATIVE=OFF -DGGML_VULKAN=1 -DLLAMA_CURL=1 -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON && \ cmake --build build --config Release -j$(nproc) RUN mkdir -p /app/lib && \ diff --git a/.github/ISSUE_TEMPLATE/020-enhancement.yml b/.github/ISSUE_TEMPLATE/020-enhancement.yml index 02dd4f575..cee1446f5 100644 --- a/.github/ISSUE_TEMPLATE/020-enhancement.yml +++ b/.github/ISSUE_TEMPLATE/020-enhancement.yml @@ -6,7 +6,7 @@ body: - type: markdown attributes: value: | - [Please post your idea first in Discussion if there is not yet a consensus for this enhancement request. This will help to keep this issue tracker focused on enhancements that the community has agreed needs to be implemented.](https://github.com/ggerganov/llama.cpp/discussions/categories/ideas) + [Please post your idea first in Discussion if there is not yet a consensus for this enhancement request. This will help to keep this issue tracker focused on enhancements that the community has agreed needs to be implemented.](https://github.com/ggml-org/llama.cpp/discussions/categories/ideas) - type: checkboxes id: prerequisites @@ -16,11 +16,11 @@ body: options: - label: I am running the latest code. Mention the version if possible as well. required: true - - label: I carefully followed the [README.md](https://github.com/ggerganov/llama.cpp/blob/master/README.md). + - label: I carefully followed the [README.md](https://github.com/ggml-org/llama.cpp/blob/master/README.md). required: true - label: I searched using keywords relevant to my issue to make sure that I am creating a new issue that is not already open (or closed). required: true - - label: I reviewed the [Discussions](https://github.com/ggerganov/llama.cpp/discussions), and have a new and useful enhancement to share. + - label: I reviewed the [Discussions](https://github.com/ggml-org/llama.cpp/discussions), and have a new and useful enhancement to share. required: true - type: textarea diff --git a/.github/ISSUE_TEMPLATE/030-research.yml b/.github/ISSUE_TEMPLATE/030-research.yml index 18975dbbf..e774550d5 100644 --- a/.github/ISSUE_TEMPLATE/030-research.yml +++ b/.github/ISSUE_TEMPLATE/030-research.yml @@ -6,7 +6,7 @@ body: - type: markdown attributes: value: | - Don't forget to check for any [duplicate research issue tickets](https://github.com/ggerganov/llama.cpp/issues?q=is%3Aopen+is%3Aissue+label%3A%22research+%F0%9F%94%AC%22) + Don't forget to check for any [duplicate research issue tickets](https://github.com/ggml-org/llama.cpp/issues?q=is%3Aopen+is%3Aissue+label%3A%22research+%F0%9F%94%AC%22) - type: checkboxes id: research-stage diff --git a/.github/ISSUE_TEMPLATE/040-refactor.yml b/.github/ISSUE_TEMPLATE/040-refactor.yml index b6e6ab36d..2fe94e26c 100644 --- a/.github/ISSUE_TEMPLATE/040-refactor.yml +++ b/.github/ISSUE_TEMPLATE/040-refactor.yml @@ -6,8 +6,8 @@ body: - type: markdown attributes: value: | - Don't forget to [check for existing refactor issue tickets](https://github.com/ggerganov/llama.cpp/issues?q=is%3Aopen+is%3Aissue+label%3Arefactoring) in case it's already covered. - Also you may want to check [Pull request refactor label as well](https://github.com/ggerganov/llama.cpp/pulls?q=is%3Aopen+is%3Apr+label%3Arefactoring) for duplicates too. + Don't forget to [check for existing refactor issue tickets](https://github.com/ggml-org/llama.cpp/issues?q=is%3Aopen+is%3Aissue+label%3Arefactoring) in case it's already covered. + Also you may want to check [Pull request refactor label as well](https://github.com/ggml-org/llama.cpp/pulls?q=is%3Aopen+is%3Apr+label%3Arefactoring) for duplicates too. - type: textarea id: background-description diff --git a/.github/ISSUE_TEMPLATE/config.yml b/.github/ISSUE_TEMPLATE/config.yml index eb8c4b472..0d246533c 100644 --- a/.github/ISSUE_TEMPLATE/config.yml +++ b/.github/ISSUE_TEMPLATE/config.yml @@ -1,11 +1,11 @@ blank_issues_enabled: true contact_links: - name: Got an idea? - url: https://github.com/ggerganov/llama.cpp/discussions/categories/ideas + url: https://github.com/ggml-org/llama.cpp/discussions/categories/ideas about: Pop it there. It may then become an enhancement ticket. - name: Got a question? - url: https://github.com/ggerganov/llama.cpp/discussions/categories/q-a + url: https://github.com/ggml-org/llama.cpp/discussions/categories/q-a about: Ask a question there! - name: Want to contribute? - url: https://github.com/ggerganov/llama.cpp/wiki/contribute + url: https://github.com/ggml-org/llama.cpp/wiki/contribute about: Head to the contribution guide page of the wiki for areas you can help with diff --git a/.github/actions/windows-setup-curl/action.yml b/.github/actions/windows-setup-curl/action.yml new file mode 100644 index 000000000..5d76da3d7 --- /dev/null +++ b/.github/actions/windows-setup-curl/action.yml @@ -0,0 +1,25 @@ +name: 'Windows - Setup CURL' +description: 'Composite action, to be reused in other workflow' +inputs: + curl_version: + description: 'CURL version' + required: false + default: '8.6.0_6' +outputs: + curl_path: + description: "Path to the downloaded libcurl" + value: ${{ steps.get_libcurl.outputs.curl_path }} + +runs: + using: "composite" + steps: + - name: libCURL + id: get_libcurl + shell: powershell + env: + CURL_VERSION: ${{ inputs.curl_version }} + run: | + curl.exe -o $env:RUNNER_TEMP/curl.zip -L "https://curl.se/windows/dl-${env:CURL_VERSION}/curl-${env:CURL_VERSION}-win64-mingw.zip" + mkdir $env:RUNNER_TEMP/libcurl + tar.exe -xvf $env:RUNNER_TEMP/curl.zip --strip-components=1 -C $env:RUNNER_TEMP/libcurl + echo "curl_path=$env:RUNNER_TEMP/libcurl" >> $env:GITHUB_OUTPUT diff --git a/.github/pull_request_template.md b/.github/pull_request_template.md index d9f5bdc23..d0bdd73c4 100644 --- a/.github/pull_request_template.md +++ b/.github/pull_request_template.md @@ -1 +1 @@ -*Make sure to read the [contributing guidelines](https://github.com/ggerganov/llama.cpp/blob/master/CONTRIBUTING.md) before submitting a PR* +*Make sure to read the [contributing guidelines](https://github.com/ggml-org/llama.cpp/blob/master/CONTRIBUTING.md) before submitting a PR* diff --git a/.github/workflows/bench.yml.disabled b/.github/workflows/bench.yml.disabled index 1c8787ef7..75d271479 100644 --- a/.github/workflows/bench.yml.disabled +++ b/.github/workflows/bench.yml.disabled @@ -1,5 +1,5 @@ # TODO: there have been some issues with the workflow, so disabling for now -# https://github.com/ggerganov/llama.cpp/issues/7893 +# https://github.com/ggml-org/llama.cpp/issues/7893 # # Benchmark name: Benchmark @@ -57,17 +57,7 @@ jobs: if: | inputs.gpu-series == 'Standard_NC4as_T4_v3' - || ( - github.event_name == 'schedule' - && github.ref_name == 'master' - && github.repository_owner == 'ggerganov' - ) || github.event_name == 'pull_request_target' - || ( - github.event_name == 'push' - && github.event.ref == 'refs/heads/master' - && github.repository_owner == 'ggerganov' - ) steps: - name: Clone id: checkout @@ -114,7 +104,6 @@ jobs: cmake -B build \ -DGGML_NATIVE=OFF \ -DLLAMA_BUILD_SERVER=ON \ - -DLLAMA_CURL=ON \ -DLLAMA_CUBLAS=ON \ -DCUDAToolkit_ROOT=/usr/local/cuda \ -DCMAKE_CUDA_COMPILER=/usr/local/cuda/bin/nvcc \ diff --git a/.github/workflows/build-linux-cross.yml b/.github/workflows/build-linux-cross.yml new file mode 100644 index 000000000..e8639913e --- /dev/null +++ b/.github/workflows/build-linux-cross.yml @@ -0,0 +1,124 @@ +name: Build on Linux using cross-compiler +on: + workflow_dispatch: + workflow_call: + +jobs: + ubuntu-latest-riscv64-cpu-cross: + runs-on: ubuntu-latest + + steps: + - uses: actions/checkout@v4 + - name: Setup Riscv + run: | + sudo dpkg --add-architecture riscv64 + sudo sed -i 's|http://azure.archive.ubuntu.com/ubuntu|http://ports.ubuntu.com/ubuntu-ports|g' \ + /etc/apt/sources.list /etc/apt/apt-mirrors.txt + sudo apt-get clean + sudo apt-get update + sudo apt-get install -y --no-install-recommends \ + build-essential \ + gcc-14-riscv64-linux-gnu \ + g++-14-riscv64-linux-gnu \ + libcurl4-openssl-dev:riscv64 + + - name: Build + run: | + cmake -B build -DCMAKE_BUILD_TYPE=Release \ + -DGGML_OPENMP=OFF \ + -DLLAMA_BUILD_EXAMPLES=ON \ + -DLLAMA_BUILD_TESTS=OFF \ + -DCMAKE_SYSTEM_NAME=Linux \ + -DCMAKE_SYSTEM_PROCESSOR=riscv64 \ + -DCMAKE_C_COMPILER=riscv64-linux-gnu-gcc-14 \ + -DCMAKE_CXX_COMPILER=riscv64-linux-gnu-g++-14 \ + -DCMAKE_POSITION_INDEPENDENT_CODE=ON \ + -DCMAKE_FIND_ROOT_PATH=/usr/lib/riscv64-linux-gnu \ + -DCMAKE_FIND_ROOT_PATH_MODE_PROGRAM=NEVER \ + -DCMAKE_FIND_ROOT_PATH_MODE_LIBRARY=ONLY \ + -DCMAKE_FIND_ROOT_PATH_MODE_INCLUDE=BOTH + + cmake --build build --config Release -j $(nproc) + + ubuntu-latest-riscv64-vulkan-cross: + runs-on: ubuntu-latest + + steps: + - uses: actions/checkout@v4 + with: + fetch-depth: 0 + + - name: Setup Riscv + run: | + sudo dpkg --add-architecture riscv64 + sudo sed -i 's|http://azure.archive.ubuntu.com/ubuntu|http://ports.ubuntu.com/ubuntu-ports|g' \ + /etc/apt/sources.list /etc/apt/apt-mirrors.txt + sudo apt-get clean + sudo apt-get update + sudo apt-get install -y --no-install-recommends \ + build-essential \ + glslc \ + gcc-14-riscv64-linux-gnu \ + g++-14-riscv64-linux-gnu \ + libvulkan-dev:riscv64 \ + libcurl4-openssl-dev:riscv64 + + - name: Build + run: | + cmake -B build -DCMAKE_BUILD_TYPE=Release \ + -DGGML_VULKAN=ON \ + -DGGML_OPENMP=OFF \ + -DLLAMA_BUILD_EXAMPLES=ON \ + -DLLAMA_BUILD_TESTS=OFF \ + -DCMAKE_SYSTEM_NAME=Linux \ + -DCMAKE_SYSTEM_PROCESSOR=riscv64 \ + -DCMAKE_C_COMPILER=riscv64-linux-gnu-gcc-14 \ + -DCMAKE_CXX_COMPILER=riscv64-linux-gnu-g++-14 \ + -DCMAKE_POSITION_INDEPENDENT_CODE=ON \ + -DCMAKE_FIND_ROOT_PATH=/usr/lib/riscv64-linux-gnu \ + -DCMAKE_FIND_ROOT_PATH_MODE_PROGRAM=NEVER \ + -DCMAKE_FIND_ROOT_PATH_MODE_LIBRARY=ONLY \ + -DCMAKE_FIND_ROOT_PATH_MODE_INCLUDE=BOTH + + cmake --build build --config Release -j $(nproc) + + ubuntu-latest-arm64-vulkan-cross: + runs-on: ubuntu-latest + + steps: + - uses: actions/checkout@v4 + with: + fetch-depth: 0 + + - name: Setup Arm64 + run: | + sudo dpkg --add-architecture arm64 + sudo sed -i 's|http://azure.archive.ubuntu.com/ubuntu|http://ports.ubuntu.com/ubuntu-ports|g' \ + /etc/apt/sources.list /etc/apt/apt-mirrors.txt + sudo apt-get clean + sudo apt-get update + sudo apt-get install -y --no-install-recommends \ + build-essential \ + glslc \ + crossbuild-essential-arm64 \ + libvulkan-dev:arm64 \ + libcurl4-openssl-dev:arm64 + + - name: Build + run: | + cmake -B build -DCMAKE_BUILD_TYPE=Release \ + -DGGML_VULKAN=ON \ + -DGGML_OPENMP=OFF \ + -DLLAMA_BUILD_EXAMPLES=ON \ + -DLLAMA_BUILD_TESTS=OFF \ + -DCMAKE_SYSTEM_NAME=Linux \ + -DCMAKE_SYSTEM_PROCESSOR=aarch64 \ + -DCMAKE_C_COMPILER=aarch64-linux-gnu-gcc \ + -DCMAKE_CXX_COMPILER=aarch64-linux-gnu-g++ \ + -DCMAKE_POSITION_INDEPENDENT_CODE=ON \ + -DCMAKE_FIND_ROOT_PATH=/usr/lib/aarch64-linux-gnu \ + -DCMAKE_FIND_ROOT_PATH_MODE_PROGRAM=NEVER \ + -DCMAKE_FIND_ROOT_PATH_MODE_LIBRARY=ONLY \ + -DCMAKE_FIND_ROOT_PATH_MODE_INCLUDE=BOTH + + cmake --build build --config Release -j $(nproc) diff --git a/.github/workflows/build.yml b/.github/workflows/build.yml index 6841ba589..bcfcf08ac 100644 --- a/.github/workflows/build.yml +++ b/.github/workflows/build.yml @@ -10,7 +10,7 @@ on: push: branches: - master - paths: ['.github/workflows/build.yml', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.cuh', '**/*.swift', '**/*.m', '**/*.metal', '**/*.comp'] + paths: ['.github/workflows/build.yml', '.github/workflows/build-linux-cross.yml', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.cuh', '**/*.swift', '**/*.m', '**/*.metal', '**/*.comp'] pull_request: types: [opened, synchronize, reopened] paths: ['.github/workflows/build.yml', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.cuh', '**/*.swift', '**/*.m', '**/*.metal', '**/*.comp'] @@ -54,6 +54,7 @@ jobs: continue-on-error: true run: | brew update + brew install curl - name: Build id: cmake_build @@ -62,7 +63,6 @@ jobs: cmake -B build \ -DCMAKE_BUILD_RPATH="@loader_path" \ -DLLAMA_FATAL_WARNINGS=ON \ - -DLLAMA_CURL=ON \ -DGGML_METAL_USE_BF16=ON \ -DGGML_METAL_EMBED_LIBRARY=ON \ -DGGML_RPC=ON @@ -92,7 +92,6 @@ jobs: if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }} run: | cp LICENSE ./build/bin/ - cp examples/run/linenoise.cpp/LICENSE ./build/bin/LICENSE.linenoise.cpp zip -r llama-${{ steps.tag.outputs.name }}-bin-macos-arm64.zip ./build/bin/* - name: Upload artifacts @@ -123,17 +122,17 @@ jobs: continue-on-error: true run: | brew update + brew install curl - name: Build id: cmake_build run: | sysctl -a # Metal is disabled due to intermittent failures with Github runners not having a GPU: - # https://github.com/ggerganov/llama.cpp/actions/runs/8635935781/job/23674807267#step:5:2313 + # https://github.com/ggml-org/llama.cpp/actions/runs/8635935781/job/23674807267#step:5:2313 cmake -B build \ -DCMAKE_BUILD_RPATH="@loader_path" \ -DLLAMA_FATAL_WARNINGS=ON \ - -DLLAMA_CURL=ON \ -DGGML_METAL=OFF \ -DGGML_RPC=ON cmake --build build --config Release -j $(sysctl -n hw.logicalcpu) @@ -162,7 +161,6 @@ jobs: if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }} run: | cp LICENSE ./build/bin/ - cp examples/run/linenoise.cpp/LICENSE ./build/bin/LICENSE.linenoise.cpp zip -r llama-${{ steps.tag.outputs.name }}-bin-macos-x64.zip ./build/bin/* - name: Upload artifacts @@ -173,7 +171,15 @@ jobs: name: llama-bin-macos-x64.zip ubuntu-cpu-cmake: - runs-on: ubuntu-22.04 + strategy: + matrix: + include: + - build: 'x64' + os: ubuntu-22.04 + - build: 'arm64' + os: ubuntu-22.04-arm + + runs-on: ${{ matrix.os }} steps: - name: Clone @@ -199,7 +205,6 @@ jobs: run: | cmake -B build \ -DLLAMA_FATAL_WARNINGS=ON \ - -DLLAMA_CURL=ON \ -DGGML_RPC=ON cmake --build build --config Release -j $(nproc) @@ -238,15 +243,14 @@ jobs: if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }} run: | cp LICENSE ./build/bin/ - cp examples/run/linenoise.cpp/LICENSE ./build/bin/LICENSE.linenoise.cpp - zip -r llama-${{ steps.tag.outputs.name }}-bin-ubuntu-x64.zip ./build/bin/* + zip -r llama-${{ steps.tag.outputs.name }}-bin-ubuntu-${{ matrix.build }}.zip ./build/bin/* - name: Upload artifacts if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }} uses: actions/upload-artifact@v4 with: - path: llama-${{ steps.tag.outputs.name }}-bin-ubuntu-x64.zip - name: llama-bin-ubuntu-x64.zip + path: llama-${{ steps.tag.outputs.name }}-bin-ubuntu-${{ matrix.build }}.zip + name: llama-bin-ubuntu-${{ matrix.build }}.zip ubuntu-latest-cmake-sanitizer: runs-on: ubuntu-latest @@ -273,7 +277,7 @@ jobs: id: depends run: | sudo apt-get update - sudo apt-get install build-essential + sudo apt-get install build-essential libcurl4-openssl-dev - name: Build id: cmake_build @@ -314,7 +318,7 @@ jobs: id: depends run: | sudo apt-get update - sudo apt-get install build-essential + sudo apt-get install build-essential libcurl4-openssl-dev - name: Build id: cmake_build @@ -352,7 +356,7 @@ jobs: id: depends run: | sudo apt-get update - sudo apt-get install build-essential + sudo apt-get install build-essential libcurl4-openssl-dev - name: Build id: cmake_build @@ -374,6 +378,8 @@ jobs: - name: Clone id: checkout uses: actions/checkout@v4 + with: + fetch-depth: 0 - name: ccache uses: hendrikmuhs/ccache-action@v1.2.16 @@ -387,7 +393,7 @@ jobs: wget -qO - https://packages.lunarg.com/lunarg-signing-key-pub.asc | sudo apt-key add - sudo wget -qO /etc/apt/sources.list.d/lunarg-vulkan-jammy.list https://packages.lunarg.com/vulkan/lunarg-vulkan-jammy.list sudo apt-get update -y - sudo apt-get install -y build-essential mesa-vulkan-drivers vulkan-sdk + sudo apt-get install -y build-essential mesa-vulkan-drivers vulkan-sdk libcurl4-openssl-dev - name: Build id: cmake_build @@ -401,7 +407,34 @@ jobs: run: | cd build # This is using llvmpipe and runs slower than other backends - ctest -L main --verbose --timeout 1800 + ctest -L main --verbose --timeout 2700 + + - name: Determine tag name + id: tag + shell: bash + run: | + BUILD_NUMBER="$(git rev-list --count HEAD)" + SHORT_HASH="$(git rev-parse --short=7 HEAD)" + if [[ "${{ env.BRANCH_NAME }}" == "master" ]]; then + echo "name=b${BUILD_NUMBER}" >> $GITHUB_OUTPUT + else + SAFE_NAME=$(echo "${{ env.BRANCH_NAME }}" | tr '/' '-') + echo "name=${SAFE_NAME}-b${BUILD_NUMBER}-${SHORT_HASH}" >> $GITHUB_OUTPUT + fi + + - name: Pack artifacts + id: pack_artifacts + if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }} + run: | + cp LICENSE ./build/bin/ + zip -r llama-${{ steps.tag.outputs.name }}-bin-ubuntu-vulkan-x64.zip ./build/bin/* + + - name: Upload artifacts + if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }} + uses: actions/upload-artifact@v4 + with: + path: llama-${{ steps.tag.outputs.name }}-bin-ubuntu-vulkan-x64.zip + name: llama-bin-ubuntu-vulkan-x64.zip ubuntu-22-cmake-hip: runs-on: ubuntu-22.04 @@ -416,7 +449,7 @@ jobs: id: depends run: | sudo apt-get update - sudo apt-get install -y build-essential git cmake rocblas-dev hipblas-dev + sudo apt-get install -y build-essential git cmake rocblas-dev hipblas-dev libcurl4-openssl-dev - name: ccache uses: hendrikmuhs/ccache-action@v1.2.16 @@ -429,6 +462,7 @@ jobs: run: | cmake -B build -S . \ -DCMAKE_HIP_COMPILER="$(hipconfig -l)/clang" \ + -DGGML_HIP_ROCWMMA_FATTN=ON \ -DGGML_HIP=ON cmake --build build --config Release -j $(nproc) @@ -438,12 +472,13 @@ jobs: cmake -B build2 -S . \ -DCMAKE_C_COMPILER=hipcc \ -DCMAKE_CXX_COMPILER=hipcc \ + -DGGML_HIP_ROCWMMA_FATTN=ON \ -DGGML_HIP=ON cmake --build build2 --config Release -j $(nproc) ubuntu-22-cmake-musa: runs-on: ubuntu-22.04 - container: mthreads/musa:rc3.1.0-devel-ubuntu22.04 + container: mthreads/musa:rc3.1.1-devel-ubuntu22.04 steps: - name: Clone @@ -490,7 +525,7 @@ jobs: shell: bash run: | sudo apt update - sudo apt install intel-oneapi-compiler-dpcpp-cpp + sudo apt install intel-oneapi-compiler-dpcpp-cpp libcurl4-openssl-dev - name: install oneAPI MKL library shell: bash @@ -538,7 +573,7 @@ jobs: shell: bash run: | sudo apt update - sudo apt install intel-oneapi-compiler-dpcpp-cpp + sudo apt install intel-oneapi-compiler-dpcpp-cpp libcurl4-openssl-dev - name: install oneAPI MKL library shell: bash @@ -566,6 +601,9 @@ jobs: -DGGML_SYCL_F16=ON cmake --build build --config Release -j $(nproc) + build-linux-cross: + uses: ./.github/workflows/build-linux-cross.yml + macOS-latest-cmake-ios: runs-on: macos-latest @@ -593,6 +631,7 @@ jobs: cmake -B build -G Xcode \ -DGGML_METAL_USE_BF16=ON \ -DGGML_METAL_EMBED_LIBRARY=ON \ + -DLLAMA_BUILD_COMMON=OFF \ -DLLAMA_BUILD_EXAMPLES=OFF \ -DLLAMA_BUILD_TESTS=OFF \ -DLLAMA_BUILD_SERVER=OFF \ @@ -628,6 +667,7 @@ jobs: cmake -B build -G Xcode \ -DGGML_METAL_USE_BF16=ON \ -DGGML_METAL_EMBED_LIBRARY=ON \ + -DLLAMA_BUILD_COMMON=OFF \ -DLLAMA_BUILD_EXAMPLES=OFF \ -DLLAMA_BUILD_TESTS=OFF \ -DLLAMA_BUILD_SERVER=OFF \ @@ -636,6 +676,36 @@ jobs: -DCMAKE_XCODE_ATTRIBUTE_DEVELOPMENT_TEAM=ggml cmake --build build --config Release -j $(sysctl -n hw.logicalcpu) -- CODE_SIGNING_ALLOWED=NO + macOS-latest-cmake-visionos: + runs-on: macos-latest + + steps: + - name: Clone + id: checkout + uses: actions/checkout@v4 + + - name: Dependencies + id: depends + continue-on-error: true + run: | + brew update + + - name: Build + id: cmake_build + run: | + sysctl -a + cmake -B build -G Xcode \ + -DGGML_METAL_USE_BF16=ON \ + -DGGML_METAL_EMBED_LIBRARY=ON \ + -DLLAMA_BUILD_COMMON=OFF \ + -DLLAMA_BUILD_EXAMPLES=OFF \ + -DLLAMA_BUILD_TESTS=OFF \ + -DLLAMA_BUILD_SERVER=OFF \ + -DCMAKE_SYSTEM_NAME=visionOS \ + -DCMAKE_OSX_DEPLOYMENT_TARGET=1.0 \ + -DCMAKE_XCODE_ATTRIBUTE_DEVELOPMENT_TEAM=ggml + cmake --build build --config Release -j $(sysctl -n hw.logicalcpu) -- CODE_SIGNING_ALLOWED=NO + macOS-latest-swift: runs-on: macos-latest @@ -667,17 +737,17 @@ jobs: cmake -B build -G Xcode \ -DGGML_METAL_USE_BF16=ON \ -DGGML_METAL_EMBED_LIBRARY=ON \ + -DLLAMA_CURL=OFF \ -DLLAMA_BUILD_EXAMPLES=OFF \ -DLLAMA_BUILD_TESTS=OFF \ -DLLAMA_BUILD_SERVER=OFF \ -DCMAKE_OSX_ARCHITECTURES="arm64;x86_64" cmake --build build --config Release -j $(sysctl -n hw.logicalcpu) - sudo cmake --install build --config Release - name: xcodebuild for swift package id: xcodebuild run: | - xcodebuild -scheme llama-Package -destination "${{ matrix.destination }}" + ./build-xcframework.sh windows-msys2: runs-on: windows-latest @@ -735,7 +805,7 @@ jobs: env: OPENBLAS_VERSION: 0.3.23 SDE_VERSION: 9.33.0-2024-01-07 - VULKAN_VERSION: 1.3.261.1 + VULKAN_VERSION: 1.4.309.0 strategy: matrix: @@ -828,10 +898,17 @@ jobs: -DCMAKE_INSTALL_PREFIX="$env:RUNNER_TEMP/opencl-arm64-release" cmake --build build-arm64-release --target install --config release + - name: libCURL + id: get_libcurl + uses: ./.github/actions/windows-setup-curl + - name: Build id: cmake_build + env: + CURL_PATH: ${{ steps.get_libcurl.outputs.curl_path }} run: | - cmake -S . -B build ${{ matrix.defines }} + cmake -S . -B build ${{ matrix.defines }} ` + -DCURL_LIBRARY="$env:CURL_PATH/lib/libcurl.dll.a" -DCURL_INCLUDE_DIR="$env:CURL_PATH/include" cmake --build build --config Release -j ${env:NUMBER_OF_PROCESSORS} - name: Add libopenblas.dll @@ -891,9 +968,10 @@ jobs: - name: Pack artifacts id: pack_artifacts if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }} + env: + CURL_PATH: ${{ steps.get_libcurl.outputs.curl_path }} run: | - Copy-Item LICENSE .\build\bin\Release\llama.cpp.txt - Copy-Item .\examples\run\linenoise.cpp\LICENSE .\build\bin\Release\linenoise.cpp.txt + Copy-Item $env:CURL_PATH\bin\libcurl-x64.dll .\build\bin\Release\libcurl-x64.dll 7z a llama-${{ steps.tag.outputs.name }}-bin-win-${{ matrix.build }}.zip .\build\bin\Release\* - name: Upload artifacts @@ -919,7 +997,7 @@ jobs: DEBIAN_FRONTEND: noninteractive run: | apt update - apt install -y cmake build-essential ninja-build libgomp1 git + apt install -y cmake build-essential ninja-build libgomp1 git libcurl4-openssl-dev - name: ccache uses: hendrikmuhs/ccache-action@v1.2.16 @@ -1021,16 +1099,23 @@ jobs: run: | choco install ninja + - name: libCURL + id: get_libcurl + uses: ./.github/actions/windows-setup-curl + - name: Build id: cmake_build shell: cmd + env: + CURL_PATH: ${{ steps.get_libcurl.outputs.curl_path }} run: | call "C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise\VC\Auxiliary\Build\vcvars64.bat" cmake -S . -B build -G "Ninja Multi-Config" ^ -DLLAMA_BUILD_SERVER=ON ^ -DGGML_NATIVE=OFF ^ -DGGML_CUDA=ON ^ - -DGGML_RPC=ON + -DGGML_RPC=ON ^ + -DCURL_LIBRARY="%CURL_PATH%/lib/libcurl.dll.a" -DCURL_INCLUDE_DIR="%CURL_PATH%/include" set /A NINJA_JOBS=%NUMBER_OF_PROCESSORS%-1 cmake --build build --config Release -j %NINJA_JOBS% -t ggml cmake --build build --config Release @@ -1051,7 +1136,10 @@ jobs: - name: Pack artifacts id: pack_artifacts if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }} + env: + CURL_PATH: ${{ steps.get_libcurl.outputs.curl_path }} run: | + cp $env:CURL_PATH\bin\libcurl-x64.dll .\build\bin\Release\libcurl-x64.dll 7z a llama-${{ steps.tag.outputs.name }}-bin-win-${{ matrix.build }}-cu${{ matrix.cuda }}-x64.zip .\build\bin\Release\* - name: Upload artifacts @@ -1106,6 +1194,8 @@ jobs: run: | scripts/install-oneapi.bat $WINDOWS_BASEKIT_URL $WINDOWS_DPCPP_MKL + # TODO: add libcurl support ; we will also need to modify win-build-sycl.bat to accept user-specified args + - name: Build id: cmake_build run: examples/sycl/win-build-sycl.bat @@ -1165,6 +1255,11 @@ jobs: id: checkout uses: actions/checkout@v4 + - name: Clone rocWMMA repository + id: clone_rocwmma + run: | + git clone https://github.com/rocm/rocwmma --branch rocm-6.2.4 --depth 1 + - name: Install id: depends run: | @@ -1186,19 +1281,29 @@ jobs: key: ${{ github.job }} evict-old-files: 1d + - name: libCURL + id: get_libcurl + uses: ./.github/actions/windows-setup-curl + - name: Build id: cmake_build + env: + CURL_PATH: ${{ steps.get_libcurl.outputs.curl_path }} run: | $env:HIP_PATH=$(Resolve-Path 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' | split-path | split-path) $env:CMAKE_PREFIX_PATH="${env:HIP_PATH}" cmake -G "Unix Makefiles" -B build -S . ` -DCMAKE_C_COMPILER="${env:HIP_PATH}\bin\clang.exe" ` -DCMAKE_CXX_COMPILER="${env:HIP_PATH}\bin\clang++.exe" ` + -DCMAKE_CXX_FLAGS="-I$($PWD.Path.Replace('\', '/'))/rocwmma/library/include/" ` -DCMAKE_BUILD_TYPE=Release ` -DGGML_HIP=ON ` - -DGGML_RPC=ON + -DGGML_HIP_ROCWMMA_FATTN=ON ` + -DGGML_RPC=ON ` + -DCURL_LIBRARY="$env:CURL_PATH/lib/libcurl.dll.a" -DCURL_INCLUDE_DIR="$env:CURL_PATH/include" cmake --build build -j ${env:NUMBER_OF_PROCESSORS} + # TODO: reuse windows-latest-cmake-hip instead of duplicating this job windows-latest-cmake-hip-release: if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }} runs-on: windows-latest @@ -1214,6 +1319,11 @@ jobs: with: fetch-depth: 0 + - name: Clone rocWMMA repository + id: clone_rocwmma + run: | + git clone https://github.com/rocm/rocwmma --branch rocm-6.2.4 --depth 1 + - name: ccache uses: hendrikmuhs/ccache-action@v1.2.16 with: @@ -1235,18 +1345,27 @@ jobs: run: | & 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' --version + - name: libCURL + id: get_libcurl + uses: ./.github/actions/windows-setup-curl + - name: Build id: cmake_build + env: + CURL_PATH: ${{ steps.get_libcurl.outputs.curl_path }} run: | $env:HIP_PATH=$(Resolve-Path 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' | split-path | split-path) $env:CMAKE_PREFIX_PATH="${env:HIP_PATH}" cmake -G "Unix Makefiles" -B build -S . ` -DCMAKE_C_COMPILER="${env:HIP_PATH}\bin\clang.exe" ` -DCMAKE_CXX_COMPILER="${env:HIP_PATH}\bin\clang++.exe" ` + -DCMAKE_CXX_FLAGS="-I$($PWD.Path.Replace('\', '/'))/rocwmma/library/include/" ` -DCMAKE_BUILD_TYPE=Release ` -DAMDGPU_TARGETS=${{ matrix.gpu_target }} ` + -DGGML_HIP_ROCWMMA_FATTN=ON ` -DGGML_HIP=ON ` - -DGGML_RPC=ON + -DGGML_RPC=ON ` + -DCURL_LIBRARY="$env:CURL_PATH/lib/libcurl.dll.a" -DCURL_INCLUDE_DIR="$env:CURL_PATH/include" cmake --build build -j ${env:NUMBER_OF_PROCESSORS} md "build\bin\rocblas\library\" cp "${env:HIP_PATH}\bin\hipblas.dll" "build\bin\" @@ -1268,7 +1387,10 @@ jobs: - name: Pack artifacts id: pack_artifacts + env: + CURL_PATH: ${{ steps.get_libcurl.outputs.curl_path }} run: | + cp $env:CURL_PATH\bin\libcurl-x64.dll .\build\bin\libcurl-x64.dll 7z a llama-${{ steps.tag.outputs.name }}-bin-win-hip-x64-${{ matrix.gpu_target }}.zip .\build\bin\* - name: Upload artifacts @@ -1283,6 +1405,8 @@ jobs: steps: - name: Checkout code uses: actions/checkout@v4 + with: + fetch-depth: 0 - name: Build id: cmake_build @@ -1291,6 +1415,7 @@ jobs: cmake -B build -G Xcode \ -DGGML_METAL_USE_BF16=ON \ -DGGML_METAL_EMBED_LIBRARY=ON \ + -DLLAMA_CURL=OFF \ -DLLAMA_BUILD_EXAMPLES=OFF \ -DLLAMA_BUILD_TESTS=OFF \ -DLLAMA_BUILD_SERVER=OFF \ @@ -1298,15 +1423,40 @@ jobs: -DCMAKE_OSX_DEPLOYMENT_TARGET=14.0 \ -DCMAKE_XCODE_ATTRIBUTE_DEVELOPMENT_TEAM=ggml cmake --build build --config Release -j $(sysctl -n hw.logicalcpu) -- CODE_SIGNING_ALLOWED=NO - sudo cmake --install build --config Release - name: xcodebuild for swift package id: xcodebuild run: | - xcodebuild -scheme llama-Package -destination 'generic/platform=iOS' + ./build-xcframework.sh - name: Build Xcode project - run: xcodebuild -project examples/llama.swiftui/llama.swiftui.xcodeproj -scheme llama.swiftui -sdk iphoneos CODE_SIGNING_REQUIRED=NO CODE_SIGN_IDENTITY= -destination 'generic/platform=iOS' build + run: xcodebuild -project examples/llama.swiftui/llama.swiftui.xcodeproj -scheme llama.swiftui -sdk iphoneos CODE_SIGNING_REQUIRED=NO CODE_SIGN_IDENTITY= -destination 'generic/platform=iOS' FRAMEWORK_FOLDER_PATH=./build-ios build + + - name: Determine tag name + id: tag + shell: bash + run: | + BUILD_NUMBER="$(git rev-list --count HEAD)" + SHORT_HASH="$(git rev-parse --short=7 HEAD)" + if [[ "${{ env.BRANCH_NAME }}" == "master" ]]; then + echo "name=b${BUILD_NUMBER}" >> $GITHUB_OUTPUT + else + SAFE_NAME=$(echo "${{ env.BRANCH_NAME }}" | tr '/' '-') + echo "name=${SAFE_NAME}-b${BUILD_NUMBER}-${SHORT_HASH}" >> $GITHUB_OUTPUT + fi + + - name: Pack artifacts + id: pack_artifacts + if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }} + run: | + zip --symlinks -r llama-${{ steps.tag.outputs.name }}-xcframework.zip build-apple/llama.xcframework + + - name: Upload artifacts + if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }} + uses: actions/upload-artifact@v4 + with: + path: llama-${{ steps.tag.outputs.name }}-xcframework.zip + name: llama-${{ steps.tag.outputs.name }}-xcframework android-build: runs-on: ubuntu-latest @@ -1345,8 +1495,10 @@ jobs: needs: - ubuntu-cpu-cmake + - ubuntu-22-cmake-vulkan - windows-latest-cmake - windows-2019-cmake-cuda + - windows-latest-cmake-sycl - windows-latest-cmake-hip-release - macOS-latest-cmake-arm64 - macOS-latest-cmake-x64 @@ -1619,7 +1771,7 @@ jobs: strategy: matrix: cann: - - '8.0.rc3.beta1-910b-openeuler22.03-py3.10' + - '8.1.RC1.alpha001-910b-openeuler22.03-py3.10' device: - 'ascend910b3' build: @@ -1632,7 +1784,7 @@ jobs: - name: Dependencies run: | yum update -y - yum install -y git gcc gcc-c++ make cmake + yum install -y git gcc gcc-c++ make cmake libcurl-devel - name: Build run: | diff --git a/.github/workflows/docker.yml b/.github/workflows/docker.yml index 6955a7dc8..114cbf83e 100644 --- a/.github/workflows/docker.yml +++ b/.github/workflows/docker.yml @@ -36,13 +36,13 @@ jobs: matrix: config: # Multi-stage build - - { tag: "cpu", dockerfile: ".devops/cpu.Dockerfile", platforms: "linux/amd64,linux/arm64", full: true, light: true, server: true, freediskspace: false} - - { tag: "cuda", dockerfile: ".devops/cuda.Dockerfile", platforms: "linux/amd64", full: true, light: true, server: true, freediskspace: false} - - { tag: "musa", dockerfile: ".devops/musa.Dockerfile", platforms: "linux/amd64", full: true, light: true, server: true, freediskspace: false} - - { tag: "intel", dockerfile: ".devops/intel.Dockerfile", platforms: "linux/amd64", full: true, light: true, server: true, freediskspace: false} - - { tag: "vulkan", dockerfile: ".devops/vulkan.Dockerfile", platforms: "linux/amd64", full: true, light: true, server: true, freediskspace: false} + - { tag: "cpu", dockerfile: ".devops/cpu.Dockerfile", platforms: "linux/amd64,linux/arm64", full: true, light: true, server: true, free_disk_space: false } + - { tag: "cuda", dockerfile: ".devops/cuda.Dockerfile", platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: false } + - { tag: "musa", dockerfile: ".devops/musa.Dockerfile", platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: true } + - { tag: "intel", dockerfile: ".devops/intel.Dockerfile", platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: false } + - { tag: "vulkan", dockerfile: ".devops/vulkan.Dockerfile", platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: false } # Note: the rocm images are failing due to a compiler error and are disabled until this is fixed to allow the workflow to complete - #- {tag: "rocm", dockerfile: ".devops/rocm.Dockerfile", platforms: "linux/amd64,linux/arm64", full: true, light: true, server: true, freediskspace: true } + #- {tag: "rocm", dockerfile: ".devops/rocm.Dockerfile", platforms: "linux/amd64,linux/arm64", full: true, light: true, server: true, free_disk_space: true } steps: - name: Check out the repo uses: actions/checkout@v4 @@ -51,6 +51,8 @@ jobs: - name: Set up QEMU uses: docker/setup-qemu-action@v3 + with: + image: tonistiigi/binfmt:qemu-v7.0.0-28 - name: Set up Docker Buildx uses: docker/setup-buildx-action@v3 diff --git a/.github/workflows/labeler.yml b/.github/workflows/labeler.yml index 368dbdbe5..0b0f300aa 100644 --- a/.github/workflows/labeler.yml +++ b/.github/workflows/labeler.yml @@ -11,7 +11,7 @@ jobs: steps: - uses: actions/checkout@v4 with: - repository: "ggerganov/llama.cpp" + repository: "ggml-org/llama.cpp" - uses: actions/labeler@v5 with: configuration-path: '.github/labeler.yml' diff --git a/.github/workflows/server.yml b/.github/workflows/server.yml index 3a29107d0..6c9b51322 100644 --- a/.github/workflows/server.yml +++ b/.github/workflows/server.yml @@ -129,7 +129,6 @@ jobs: cmake -B build \ -DGGML_NATIVE=OFF \ -DLLAMA_BUILD_SERVER=ON \ - -DLLAMA_CURL=ON \ -DCMAKE_BUILD_TYPE=${{ matrix.build_type }} \ -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON \ -DGGML_OPENMP=OFF ; @@ -142,7 +141,6 @@ jobs: cmake -B build \ -DGGML_NATIVE=OFF \ -DLLAMA_BUILD_SERVER=ON \ - -DLLAMA_CURL=ON \ -DCMAKE_BUILD_TYPE=${{ matrix.build_type }} \ -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON ; cmake --build build --config ${{ matrix.build_type }} -j $(nproc) --target llama-server @@ -154,13 +152,14 @@ jobs: cmake -B build \ -DGGML_NATIVE=OFF \ -DLLAMA_BUILD_SERVER=ON \ - -DLLAMA_CURL=ON \ -DCMAKE_BUILD_TYPE=${{ matrix.build_type }} ; cmake --build build --config ${{ matrix.build_type }} -j $(nproc) --target llama-server - name: Tests id: server_integration_tests if: ${{ matrix.sanitizer == '' }} + env: + GITHUB_ACTIONS: "true" run: | cd examples/server/tests ./tests.sh @@ -193,17 +192,14 @@ jobs: - name: libCURL id: get_libcurl - env: - CURL_VERSION: 8.6.0_6 - run: | - curl.exe -o $env:RUNNER_TEMP/curl.zip -L "https://curl.se/windows/dl-${env:CURL_VERSION}/curl-${env:CURL_VERSION}-win64-mingw.zip" - mkdir $env:RUNNER_TEMP/libcurl - tar.exe -xvf $env:RUNNER_TEMP/curl.zip --strip-components=1 -C $env:RUNNER_TEMP/libcurl + uses: ./.github/actions/windows-setup-curl - name: Build id: cmake_build + env: + CURL_PATH: ${{ steps.get_libcurl.outputs.curl_path }} run: | - cmake -B build -DLLAMA_CURL=ON -DCURL_LIBRARY="$env:RUNNER_TEMP/libcurl/lib/libcurl.dll.a" -DCURL_INCLUDE_DIR="$env:RUNNER_TEMP/libcurl/include" + cmake -B build -DCURL_LIBRARY="$env:CURL_PATH/lib/libcurl.dll.a" -DCURL_INCLUDE_DIR="$env:CURL_PATH/include" cmake --build build --config Release -j ${env:NUMBER_OF_PROCESSORS} --target llama-server - name: Python setup @@ -219,8 +215,10 @@ jobs: - name: Copy Libcurl id: prepare_libcurl + env: + CURL_PATH: ${{ steps.get_libcurl.outputs.curl_path }} run: | - cp $env:RUNNER_TEMP/libcurl/bin/libcurl-x64.dll ./build/bin/Release/libcurl-x64.dll + cp $env:CURL_PATH/bin/libcurl-x64.dll ./build/bin/Release/libcurl-x64.dll - name: Tests id: server_integration_tests diff --git a/.gitignore b/.gitignore index 694f36e04..2c67ad7f7 100644 --- a/.gitignore +++ b/.gitignore @@ -45,6 +45,8 @@ lcov-report/ tags .build/ build* +release +debug !build-info.cmake !build-info.cpp.in !build-info.sh @@ -98,6 +100,7 @@ examples/server/*.css.hpp examples/server/*.html.hpp examples/server/*.js.hpp examples/server/*.mjs.hpp +examples/server/*.gz.hpp !build_64.sh !examples/*.bat !examples/*/*.kts diff --git a/AUTHORS b/AUTHORS index 6796b2941..0af9f44ad 100644 --- a/AUTHORS +++ b/AUTHORS @@ -1,4 +1,4 @@ -# date: Tue Feb 4 13:04:05 EET 2025 +# date: Sat Mar 8 18:23:52 EET 2025 # this file is auto-generated by scripts/gen-authors.sh 0cc4m @@ -8,10 +8,12 @@ 3ooabkhxtn <31479382+3ooabkhxtn@users.noreply.github.com> 44670 <44670@users.noreply.github.com> 65a <10104049+65a@users.noreply.github.com> +708-145 <40387547+708-145@users.noreply.github.com> AN Long AT Aarni Koskela Aaron Miller +Aaron Teo <57927438+taronaeo@users.noreply.github.com> Aaryaman Vasishta Abheek Gulati Abhilash Majumder <30946547+abhilash1910@users.noreply.github.com> @@ -20,6 +22,7 @@ Adithya Balaji AdithyanI Adrian Adrian Hesketh +Adrian Kretz Adrien Gallouët Adrien Gallouët Ahmad Tameem <113388789+Tameem-10xE@users.noreply.github.com> @@ -28,15 +31,18 @@ AidanBeltonS <87009434+AidanBeltonS@users.noreply.github.com> AidanBeltonS Aisuko Akarshan Biswas +Akarshan Biswas Akarshan Biswas Al Mochkin <14274697+amochkin@users.noreply.github.com> Albert Jin Alberto <57916483+albbus-stack@users.noreply.github.com> Alberto Cabrera Pérez Alberto Cabrera Pérez +Aleksei Nikiforov <103434461+AlekseiNikiforovIBM@users.noreply.github.com> Alex Alex Azarov Alex Azarov +Alex Brooks Alex Klinkhamer Alex Klinkhamer Alex Nguyen @@ -67,6 +73,7 @@ Andrew Minh Nguyen <40281306+amqdn@users.noreply.github.com> Andy Salerno Andy Tai Anthony Van de Gejuchte +Antoine Viallon Antonis Makropoulos Arik Poznanski Armen Kaleshian @@ -83,6 +90,7 @@ Atsushi Tatsuma Austin <77757836+teleprint-me@users.noreply.github.com> AustinMroz BADR +BB-fat <45072480+BB-fat@users.noreply.github.com> Bach Le Bailey Chittle <39804642+bachittle@users.noreply.github.com> BarfingLemurs <128182951+BarfingLemurs@users.noreply.github.com> @@ -101,6 +109,7 @@ Bert Wagner Billel Mokeddem Bingan <70050083+binganao@users.noreply.github.com> Bjarke Viksøe <164612031+bviksoe@users.noreply.github.com> +Bodhi <3882561+BodhiHu@users.noreply.github.com> Bodo Graumann Bono Lv Borislav Stanimirov @@ -128,6 +137,7 @@ CentricStorm Chad Brewbaker Changyeon Kim Chao Jiang +Charles Duffy Charles Xu <63788048+chaxu01@users.noreply.github.com> Charles Xu Chen Xi @@ -139,12 +149,14 @@ Chris Kuehl Christian Demsar Christian Demsar Christian Falch <875252+chrfalch@users.noreply.github.com> +Christian Fillion Christian Kastner Christian Kögler Christian Köhnenkamp Christian Zhou-Zheng <59622928+christianazinn@users.noreply.github.com> Christopher Nielsen <62156882+mascguy@users.noreply.github.com> Clark Saben <76020733+csaben@users.noreply.github.com> +Clauszy Clint Herron Conrad Kramer Corentin REGAL @@ -163,6 +175,7 @@ Daniel Hiltgen Daniel Illescas Romero Daniel Kleine <53251018+d-kleine@users.noreply.github.com> Daniele <57776841+daniandtheweb@users.noreply.github.com> +Danny Milosavljevic DannyDaemonic Dat Quoc Nguyen <2412555+datquocnguyen@users.noreply.github.com> Dave @@ -170,6 +183,7 @@ Dave Airlie Dave Airlie Dave Della Costa David Friehs +David Huang <1969802+hjc4869@users.noreply.github.com> David Kennedy David Pflug David Renshaw @@ -236,6 +250,7 @@ Felix Finn Voorhees Firat FirstTimeEZ <179362031+FirstTimeEZ@users.noreply.github.com> +Florent BENOIT Folko-Ven <71110216+Folko-Ven@users.noreply.github.com> Foul-Tarnished <107711110+Foul-Tarnished@users.noreply.github.com> Francisco Melo <43780565+francis2tm@users.noreply.github.com> @@ -254,6 +269,7 @@ Gary Mulder Gavin Zhao Genkagaku.GPT Georgi Gerganov +Gian-Carlo Pascutto Gilad S Gilad S. <7817232+giladgd@users.noreply.github.com> Giuseppe Scrivano @@ -267,7 +283,9 @@ Guspan Tanadi <36249910+guspan-tanadi@users.noreply.github.com> Gustavo Rocha Dias <91472747+gustrd@users.noreply.github.com> Haggai Nuchi Halalaluyafail3 <55773281+Halalaluyafail3@users.noreply.github.com> +Hale Chan Hamdoud Hakem <90524568+hamdoudhakem@users.noreply.github.com> +Han Yin HanishKVC Haohui Mai Haoxiang Fei @@ -278,6 +296,7 @@ Haus1 Henk Poley Henri Vasserman Henrik Forstén +Henry Linjamäki Herman Semenov Hesen Peng HimariO @@ -307,6 +326,7 @@ Ivan Ivan Filipov <159561759+vanaka11@users.noreply.github.com> Ivan Komarov Ivan Stepanov +JC <43374599+MrSMlT@users.noreply.github.com> JFLFY2255 JH23X <165871467+JH23X@users.noreply.github.com> Jack Mousseau @@ -325,6 +345,7 @@ Jan Ploski Jannis Schönleber Jared Van Bortel Jared Van Bortel +Jason C.H Jason McCartney Jason Stillerman Jean-Christophe Hoelt @@ -342,6 +363,7 @@ Jiahao Li Jian Liao JidongZhang-THU <1119708529@qq.com> Jinwoo Jeong <33892306+williamjeong2@users.noreply.github.com> +Jinyang He Jiří Podivín <66251151+jpodivin@users.noreply.github.com> Jiří Sejkora Joan Fontanals @@ -379,6 +401,7 @@ Justine Tunney Juuso Alasuutari KASR Kamil Tomšík +Kante Yin Karol Kontny <82021046+kkontny@users.noreply.github.com> Karsten Weiss Karthick @@ -419,6 +442,7 @@ LoganDark Loïc Carrère LostRuins <39025047+LostRuins@users.noreply.github.com> LostRuins Concedo <39025047+LostRuins@users.noreply.github.com> +Lucas Moura Belo Luciano Luo Tian Lyle Dean @@ -463,6 +487,7 @@ Matthew Tejo Matvey Soloviev Max Krasnyansky Max Krasnyansky +Maxim Evtush <154841002+maximevtush@users.noreply.github.com> Maxime <672982+maximegmd@users.noreply.github.com> Maximilian Winter Meng Zhang @@ -494,6 +519,7 @@ Miwa / Ensan <63481257+ensan-hcl@users.noreply.github.com> Mohammadreza Hendiani Mohammadreza Hendiani Molly Sophia +MoonRide303 <130458190+MoonRide303@users.noreply.github.com> MorganRO8 <47795945+MorganRO8@users.noreply.github.com> Murilo Santana Musab Gultekin @@ -524,6 +550,7 @@ Nikolas <127742645+nneubacher@users.noreply.github.com> Nindaleth Nuno OSecret <135510162+OLSecret@users.noreply.github.com> +Oleksandr Kuvshynov <661042+okuvshynov@users.noreply.github.com> Oleksandr Nikitin Oleksii Maryshchenko Olivier Chafik @@ -533,6 +560,7 @@ PAB Pablo Duboue Pascal Patry Patrice Ferlet +Patrick Peng Paul Tsochantaris Pavel Zloi Pavol Rusnak @@ -549,6 +577,7 @@ Pieter Ouwerkerk Plamen Minev Prashant Vithule <119530321+Vithulep@users.noreply.github.com> Przemysław Pawełczyk +PureJourney Qin Yue Chen <71813199+chenqiny@users.noreply.github.com> Qingyou Meng Qu Zongfu <43257352+yancaoweidaode@users.noreply.github.com> @@ -564,14 +593,17 @@ Rand Xie Randall Fitzgerald Random Fly Reinforce-II +Rémy O Rémy Oudompheng Ren Xuancheng Rene Leonhardt <65483435+reneleonhardt@users.noreply.github.com> Reza Kakhki +Reza Rahemtola <49811529+RezaRahemtola@users.noreply.github.com> RhinoDevel Riccardo Orlando Riceball LEE Rich Dougherty +Richard Richard Kiss Richard Roberson Rick G <26732651+TheFlipbook@users.noreply.github.com> @@ -588,6 +620,7 @@ Robert Sung-wook Shin Robey Holderith Robyn Roger Meier +Rohanjames1997 Roland <14355895+rbur0425@users.noreply.github.com> Romain Biessy Romain D <90720+Artefact2@users.noreply.github.com> @@ -610,6 +643,7 @@ Ryan Landay Ryder Wishart Ryuei Rőczey Barnabás <31726601+An0nie@users.noreply.github.com> +SAMI SRHMorris <69468379+SRHMorris@users.noreply.github.com> SXX SakuraUmi @@ -634,6 +668,8 @@ Shane A Shangning Xu <32517059+xushangning@users.noreply.github.com> Shankar Shanshan Shen <467638484@qq.com> +Shelby Jenkins <47464908+ShelbyJenkins@users.noreply.github.com> +Sheldon Robinson Shijie <821898965@qq.com> Shintarou Okada Shouzheng Liu <61452103+lshzh-ww@users.noreply.github.com> @@ -713,18 +749,24 @@ Victor Nogueira Victor Z. Peng Viet-Anh NGUYEN (Andrew) Vinesh Janarthanan <36610342+VJHack@users.noreply.github.com> +Vitali Lovich +Vivian Vlad Vladimir Vladimir Malyutin +Vladimir Vuksanovic <109677816+vvuksanovic@users.noreply.github.com> Vladimir Zorin VoidIsVoid <343750470@qq.com> Volodymyr Vitvitskyi <72226+signalpillar@users.noreply.github.com> +Wagner Bruna Wang Qin <37098874+wangqin0@users.noreply.github.com> Wang Ran (汪然) WangHaoranRobin <56047610+WangHaoranRobin@users.noreply.github.com> Weird Constructor +Weizhao Ouyang Welby Seely Wentai Zhang +Wilken Gottwalt <12194808+wgottwalt@users.noreply.github.com> WillCorticesAI <150854901+WillCorticesAI@users.noreply.github.com> William Tambellini William Tambellini @@ -816,6 +858,8 @@ chaihahaha chiranko <96988916+chiranko@users.noreply.github.com> clibdev <52199778+clibdev@users.noreply.github.com> clyang +cmdr2 +cmdr2 cocktailpeanut <121128867+cocktailpeanut@users.noreply.github.com> codezjx coezbek @@ -835,6 +879,7 @@ deepdiffuser <112834445+deepdiffuser@users.noreply.github.com> devojony <61173062+devojony@users.noreply.github.com> ditsuke divinity76 +dm4 dm4 dotpy314 <33351922+dotpy314@users.noreply.github.com> drbh @@ -849,6 +894,7 @@ fairydreaming <166155368+fairydreaming@users.noreply.github.com> fengerhu1 <2748250768@qq.com> fj-y-saito <85871716+fj-y-saito@users.noreply.github.com> fraxy-v <65565042+fraxy-v@users.noreply.github.com> +fxzjshm <11426482+fxzjshm@users.noreply.github.com> github-actions[bot] gliptic gn64 @@ -873,6 +919,7 @@ hydai iSma iacore <74560659+iacore@users.noreply.github.com> icppWorld <124377669+icppWorld@users.noreply.github.com> +igardev <49397134+igardev@users.noreply.github.com> igarnier intelmatt <61025942+intelmatt@users.noreply.github.com> iohub @@ -880,6 +927,7 @@ issixx <46835150+issixx@users.noreply.github.com> jacobi petrucciani <8117202+jpetrucciani@users.noreply.github.com> jaime-m-p <167997752+jaime-m-p@users.noreply.github.com> jameswu2014 <545426914@qq.com> +jason_w jdomke <28772296+jdomke@users.noreply.github.com> jiahao su jiez <373447296@qq.com> @@ -891,6 +939,7 @@ jon-chuang <9093549+jon-chuang@users.noreply.github.com> jp-x-g jukofyork <69222624+jukofyork@users.noreply.github.com> junchao-loongson <68935141+junchao-loongson@users.noreply.github.com> +junchao-zhao <68935141+junchao-loongson@users.noreply.github.com> jwj7140 <32943891+jwj7140@users.noreply.github.com> k.h.lai kaizau @@ -925,6 +974,7 @@ ltoniazzi <61414566+ltoniazzi@users.noreply.github.com> luoyu-intel m3ndax maddes8cht <55592906+maddes8cht@users.noreply.github.com> +magicse mahorozte <41834471+mahorozte@users.noreply.github.com> makomk manikbhandari @@ -935,6 +985,7 @@ matt23654 matteo mdrokz mgroeber9110 <45620825+mgroeber9110@users.noreply.github.com> +midnight minarchist mj-shifu <77107165+mj-shifu@users.noreply.github.com> mmyjona @@ -958,10 +1009,12 @@ omahs <73983677+omahs@users.noreply.github.com> oobabooga <112222186+oobabooga@users.noreply.github.com> opparco ostix360 <55257054+ostix360@users.noreply.github.com> +pascal-lc <49066376+pascal-lc@users.noreply.github.com> pculliton peidaqi pengxin99 perserk +petterreinholdtsen piDack <104877312+piDack@users.noreply.github.com> pmysl postmasters @@ -983,6 +1036,7 @@ semidark serhii-nakon <57632032+serhii-nakon@users.noreply.github.com> sharpHL <132747147+sharpHL@users.noreply.github.com> shibe2 +simon886212 <37953122+simon886212@users.noreply.github.com> singularity <12184989+singularity-s0@users.noreply.github.com> sjinzh sjxx <63994076+ylsdamxssjxxdd@users.noreply.github.com> @@ -1000,10 +1054,12 @@ tarcey tc-mb <157115220+tc-mb@users.noreply.github.com> texmex76 <40733439+texmex76@users.noreply.github.com> thement <40525767+thement@users.noreply.github.com> +theraininsky <76763719+theraininsky@users.noreply.github.com> thewh1teagle <61390950+thewh1teagle@users.noreply.github.com> tjohnman toyer <2042519524@qq.com> tslmy +tv1wnd <55383215+tv1wnd@users.noreply.github.com> ubik2 uint256_t uint256_t @@ -1014,6 +1070,7 @@ valiray <133289098+valiray@users.noreply.github.com> vb vik viric +vmobilis <75476228+vmobilis@users.noreply.github.com> vodkaslime <646329483@qq.com> vvhg1 <94630311+vvhg1@users.noreply.github.com> vxiiduu <73044267+vxiiduu@users.noreply.github.com> @@ -1028,6 +1085,8 @@ wzy <32936898+Freed-Wu@users.noreply.github.com> xaedes xaedes xctan +xiaobing318 <71554036+xiaobing318@users.noreply.github.com> +xiaofei xloem <0xloem@gmail.com> yangli2 ymcki <84055651+ymcki@users.noreply.github.com> diff --git a/CMakeLists.txt b/CMakeLists.txt index 7b2a1845e..de51c0a17 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -29,6 +29,8 @@ else() set(LLAMA_STANDALONE OFF) endif() +option(LLAMA_USE_SYSTEM_GGML "Use system libggml" OFF) + if (EMSCRIPTEN) set(BUILD_SHARED_LIBS_DEFAULT OFF) @@ -79,7 +81,7 @@ option(LLAMA_BUILD_EXAMPLES "llama: build examples" ${LLAMA_STANDALONE}) option(LLAMA_BUILD_SERVER "llama: build server example" ${LLAMA_STANDALONE}) # 3rd party libs -option(LLAMA_CURL "llama: use libcurl to download model from an URL" OFF) +option(LLAMA_CURL "llama: use libcurl to download model from an URL" ON) option(LLAMA_LLGUIDANCE "llama-common: include LLGuidance library for structured output in common utils" OFF) # Required for relocatable CMake package @@ -145,7 +147,13 @@ endif() # 3rd-party # -if (NOT TARGET ggml) +if (LLAMA_USE_SYSTEM_GGML) + message(STATUS "Using system-provided libggml, skipping ggml build") + find_package(ggml REQUIRED) + add_library(ggml ALIAS ggml::ggml) +endif() + +if (NOT TARGET ggml AND NOT LLAMA_USE_SYSTEM_GGML) add_subdirectory(ggml) # ... otherwise assume ggml is added by a parent CMakeLists.txt endif() @@ -160,6 +168,11 @@ add_subdirectory(src) # utils, programs, examples and tests # +if (NOT LLAMA_BUILD_COMMON) + message(STATUS "LLAMA_BUILD_COMMON is OFF, disabling LLAMA_CURL") + set(LLAMA_CURL OFF) +endif() + if (LLAMA_BUILD_COMMON) add_subdirectory(common) endif() @@ -234,3 +247,20 @@ configure_file(cmake/llama.pc.in install(FILES "${CMAKE_CURRENT_BINARY_DIR}/llama.pc" DESTINATION ${CMAKE_INSTALL_LIBDIR}/pkgconfig) + +# +# copy the license files +# + +# Check if running in GitHub Actions +if(DEFINED ENV{GITHUB_ACTIONS} AND "$ENV{GITHUB_ACTIONS}" STREQUAL "true") + message(STATUS "Running inside GitHub Actions - copying license files") + + # Copy all files from licenses/ to build/bin/ + file(GLOB LICENSE_FILES "${CMAKE_SOURCE_DIR}/licenses/*") + foreach(LICENSE_FILE ${LICENSE_FILES}) + get_filename_component(FILENAME ${LICENSE_FILE} NAME) + configure_file(${LICENSE_FILE} "${CMAKE_BINARY_DIR}/bin/${FILENAME}" COPYONLY) + endforeach() +endif() + diff --git a/CONTRIBUTING.md b/CONTRIBUTING.md index 8d411982b..e68ff9244 100644 --- a/CONTRIBUTING.md +++ b/CONTRIBUTING.md @@ -1,10 +1,12 @@ # Pull requests (for contributors) +- llama.cpp uses the ggml tensor library for model evaluation. If you are unfamiliar with ggml, consider taking a look at the [examples in the ggml repository](https://github.com/ggml-org/ggml/tree/master/examples/). [simple](https://github.com/ggml-org/ggml/tree/master/examples/simple) shows the bare minimum for using ggml. [gpt-2](https://github.com/ggml-org/ggml/tree/master/examples/gpt-2) has minimal implementations for language model inference using GPT-2. [mnist](https://github.com/ggml-org/ggml/tree/master/examples/mnist) demonstrates how to train and evaluate a simple image classifier - Test your changes: - Execute [the full CI locally on your machine](ci/README.md) before publishing - Verify that the perplexity and the performance are not affected negatively by your changes (use `llama-perplexity` and `llama-bench`) - If you modified the `ggml` source, run the `test-backend-ops` tool to check whether different backend implementations of the `ggml` operators produce consistent results (this requires access to at least two different `ggml` backends) - If you modified a `ggml` operator or added a new one, add the corresponding test cases to `test-backend-ops` +- Create separate PRs for each feature or fix. Avoid combining unrelated changes in a single PR - Consider allowing write access to your branch for faster reviews, as reviewers can push commits directly - If your PR becomes stale, don't hesitate to ping the maintainers in the comments @@ -12,7 +14,7 @@ - Squash-merge PRs - Use the following format for the squashed commit title: ` : (#)`. For example: `utils : fix typo in utils.py (#1234)` -- Optionally pick a `` from here: https://github.com/ggerganov/llama.cpp/wiki/Modules +- Optionally pick a `` from here: https://github.com/ggml-org/llama.cpp/wiki/Modules - Consider adding yourself to [CODEOWNERS](CODEOWNERS) # Coding guidelines @@ -37,17 +39,17 @@ _(NOTE: this guideline is yet to be applied to the `llama.cpp` codebase. New code should follow this guideline.)_ -- Try to follow the existing patterns in the code (indentation, spaces, etc.). In case of doubt use `clang-format` to format the added code +- Try to follow the existing patterns in the code (indentation, spaces, etc.). In case of doubt use `clang-format` (from clang-tools v15+) to format the added code - For anything not covered in the current guidelines, refer to the [C++ Core Guidelines](https://isocpp.github.io/CppCoreGuidelines/CppCoreGuidelines) - Tensors store data in row-major order. We refer to dimension 0 as columns, 1 as rows, 2 as matrices -- Matrix multiplication is unconventional: [`C = ggml_mul_mat(ctx, A, B)`](https://github.com/ggerganov/llama.cpp/blob/880e352277fc017df4d5794f0c21c44e1eae2b84/ggml.h#L1058-L1064) means $C^T = A B^T \Leftrightarrow C = B A^T.$ +- Matrix multiplication is unconventional: [`C = ggml_mul_mat(ctx, A, B)`](https://github.com/ggml-org/llama.cpp/blob/880e352277fc017df4d5794f0c21c44e1eae2b84/ggml.h#L1058-L1064) means $C^T = A B^T \Leftrightarrow C = B A^T.$ ![matmul](media/matmul.png) # Naming guidelines - Use `snake_case` for function, variable and type names -- Naming usually optimizes for longest common prefix (see https://github.com/ggerganov/ggml/pull/302#discussion_r1243240963) +- Naming usually optimizes for longest common prefix (see https://github.com/ggml-org/ggml/pull/302#discussion_r1243240963) ```cpp // not OK @@ -122,4 +124,4 @@ The Github issues, PRs and discussions contain a lot of information that can be useful to get familiar with the codebase. For convenience, some of the more important information is referenced from Github projects: -https://github.com/ggerganov/llama.cpp/projects +https://github.com/ggml-org/llama.cpp/projects diff --git a/Makefile b/Makefile index dc3de3cb1..1f9455eff 100644 --- a/Makefile +++ b/Makefile @@ -1,5 +1,5 @@ ifndef LLAMA_MAKEFILE -$(error The Makefile build is deprecated. Use the CMake build instead. For more details, see https://github.com/ggerganov/llama.cpp/blob/master/docs/build.md) +$(error The Makefile build is deprecated. Use the CMake build instead. For more details, see https://github.com/ggml-org/llama.cpp/blob/master/docs/build.md) endif # Define the default target now so that it is always the first target @@ -463,7 +463,7 @@ endif ifneq '' '$(findstring mingw,$(shell $(CC) -dumpmachine))' # The stack is only 16-byte aligned on Windows, so don't let gcc emit aligned moves. # https://gcc.gnu.org/bugzilla/show_bug.cgi?id=54412 - # https://github.com/ggerganov/llama.cpp/issues/2922 + # https://github.com/ggml-org/llama.cpp/issues/2922 MK_CFLAGS += -Xassembler -muse-unaligned-vector-move MK_CXXFLAGS += -Xassembler -muse-unaligned-vector-move @@ -680,6 +680,10 @@ ifdef GGML_CUDA_CCBIN MK_NVCCFLAGS += -ccbin $(GGML_CUDA_CCBIN) endif # GGML_CUDA_CCBIN +ifdef GGML_CUDA_NO_FA + MK_NVCCFLAGS += -DGGML_CUDA_NO_FA +endif # GGML_CUDA_NO_FA + ifdef GGML_CUDA_FA_ALL_QUANTS MK_NVCCFLAGS += -DGGML_CUDA_FA_ALL_QUANTS endif # GGML_CUDA_FA_ALL_QUANTS @@ -800,6 +804,10 @@ ifdef GGML_CUDA_NO_PEER_COPY HIPFLAGS += -DGGML_CUDA_NO_PEER_COPY endif # GGML_CUDA_NO_PEER_COPY +ifdef GGML_CUDA_NO_FA + HIPFLAGS += -DGGML_CUDA_NO_FA +endif # GGML_CUDA_NO_FA + OBJ_GGML_EXT += ggml/src/ggml-cuda/ggml-cuda.o OBJ_GGML_EXT += $(patsubst %.cu,%.o,$(wildcard ggml/src/ggml-cuda/*.cu)) OBJ_GGML_EXT += $(OBJ_CUDA_TMPL) @@ -828,7 +836,7 @@ ifdef GGML_MUSA else MUSA_PATH ?= /opt/musa endif - MUSA_ARCHITECTURES ?= 21;22 + MUSA_ARCHITECTURES ?= 21;22;31 MK_CPPFLAGS += -DGGML_USE_MUSA -DGGML_USE_CUDA MK_LDFLAGS += -L$(MUSA_PATH)/lib -Wl,-rpath=$(MUSA_PATH)/lib @@ -847,7 +855,7 @@ ifdef GGML_MUSA CXX := $(MUSA_PATH)/bin/clang++ MCC := $(CCACHE) $(MUSA_PATH)/bin/mcc - MUSAFLAGS = -x musa -mtgpu + MUSAFLAGS = -fsigned-char -x musa -mtgpu MUSAFLAGS += $(foreach arch,$(subst ;, ,$(MUSA_ARCHITECTURES)),--cuda-gpu-arch=mp_$(arch)) ifdef GGML_CUDA_FORCE_MMQ @@ -876,6 +884,10 @@ ifdef GGML_CUDA_NO_PEER_COPY MUSAFLAGS += -DGGML_CUDA_NO_PEER_COPY endif # GGML_CUDA_NO_PEER_COPY +ifdef GGML_CUDA_NO_FA + MUSAFLAGS += -DGGML_CUDA_NO_FA +endif # GGML_CUDA_NO_FA + ifdef GGML_CUDA_FA_ALL_QUANTS MUSAFLAGS += -DGGML_CUDA_FA_ALL_QUANTS endif # GGML_CUDA_FA_ALL_QUANTS @@ -1078,8 +1090,8 @@ endif ifdef REMOVE_WARNING $(info !!! REMOVAL WARNING !!!) $(info The following LLAMA_ options have been removed and are no longer supported) -$(info - LLAMA_DISABLE_LOGS (https://github.com/ggerganov/llama.cpp/pull/9418)) -$(info - LLAMA_SERVER_VERBOSE (https://github.com/ggerganov/llama.cpp/pull/9418)) +$(info - LLAMA_DISABLE_LOGS (https://github.com/ggml-org/llama.cpp/pull/9418)) +$(info - LLAMA_SERVER_VERBOSE (https://github.com/ggml-org/llama.cpp/pull/9418)) $(info ) endif @@ -1364,7 +1376,7 @@ llama-server: \ examples/server/index.html.hpp \ examples/server/loading.html.hpp \ common/chat.cpp \ - common/chat.hpp \ + common/chat.h \ common/chat-template.hpp \ common/json.hpp \ common/minja.hpp \ diff --git a/Package.swift b/Package.swift deleted file mode 100644 index 01c996d24..000000000 --- a/Package.swift +++ /dev/null @@ -1,19 +0,0 @@ -// swift-tools-version:5.5 - -import PackageDescription - -let package = Package( - name: "llama", - platforms: [ - .macOS(.v12), - .iOS(.v14), - .watchOS(.v4), - .tvOS(.v14) - ], - products: [ - .library(name: "llama", targets: ["llama"]), - ], - targets: [ - .systemLibrary(name: "llama", pkgConfig: "llama"), - ] -) diff --git a/README.md b/README.md index 11f3d0286..cf45f23cf 100644 --- a/README.md +++ b/README.md @@ -3,26 +3,26 @@ ![llama](https://user-images.githubusercontent.com/1991296/230134379-7181e485-c521-4d23-a0d6-f7b3b61ba524.png) [![License: MIT](https://img.shields.io/badge/license-MIT-blue.svg)](https://opensource.org/licenses/MIT) -[![Server](https://github.com/ggerganov/llama.cpp/actions/workflows/server.yml/badge.svg)](https://github.com/ggerganov/llama.cpp/actions/workflows/server.yml) +[![Server](https://github.com/ggml-org/llama.cpp/actions/workflows/server.yml/badge.svg)](https://github.com/ggml-org/llama.cpp/actions/workflows/server.yml) -[Roadmap](https://github.com/users/ggerganov/projects/7) / [Project status](https://github.com/ggerganov/llama.cpp/discussions/3471) / [Manifesto](https://github.com/ggerganov/llama.cpp/discussions/205) / [ggml](https://github.com/ggerganov/ggml) +[Roadmap](https://github.com/users/ggerganov/projects/7) / [Project status](https://github.com/ggml-org/llama.cpp/discussions/3471) / [Manifesto](https://github.com/ggml-org/llama.cpp/discussions/205) / [ggml](https://github.com/ggml-org/ggml) Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others) in pure C/C++ ## Recent API changes -- [Changelog for `libllama` API](https://github.com/ggerganov/llama.cpp/issues/9289) -- [Changelog for `llama-server` REST API](https://github.com/ggerganov/llama.cpp/issues/9291) +- [Changelog for `libllama` API](https://github.com/ggml-org/llama.cpp/issues/9289) +- [Changelog for `llama-server` REST API](https://github.com/ggml-org/llama.cpp/issues/9291) ## Hot topics -- **How to use [MTLResidencySet](https://developer.apple.com/documentation/metal/mtlresidencyset?language=objc) to keep the GPU memory active?** https://github.com/ggerganov/llama.cpp/pull/11427 +- **How to use [MTLResidencySet](https://developer.apple.com/documentation/metal/mtlresidencyset?language=objc) to keep the GPU memory active?** https://github.com/ggml-org/llama.cpp/pull/11427 - **VS Code extension for FIM completions:** https://github.com/ggml-org/llama.vscode -- Universal tool call support in `llama-server`: https://github.com/ggerganov/llama.cpp/pull/9639 +- Universal [tool call support](./docs/function-calling.md) in `llama-server` https://github.com/ggml-org/llama.cpp/pull/9639 - Vim/Neovim plugin for FIM completions: https://github.com/ggml-org/llama.vim -- Introducing GGUF-my-LoRA https://github.com/ggerganov/llama.cpp/discussions/10123 -- Hugging Face Inference Endpoints now support GGUF out of the box! https://github.com/ggerganov/llama.cpp/discussions/9669 -- Hugging Face GGUF editor: [discussion](https://github.com/ggerganov/llama.cpp/discussions/9268) | [tool](https://huggingface.co/spaces/CISCai/gguf-editor) +- Introducing GGUF-my-LoRA https://github.com/ggml-org/llama.cpp/discussions/10123 +- Hugging Face Inference Endpoints now support GGUF out of the box! https://github.com/ggml-org/llama.cpp/discussions/9669 +- Hugging Face GGUF editor: [discussion](https://github.com/ggml-org/llama.cpp/discussions/9268) | [tool](https://huggingface.co/spaces/CISCai/gguf-editor) ---- @@ -39,7 +39,7 @@ range of hardware - locally and in the cloud. - Vulkan and SYCL backend support - CPU+GPU hybrid inference to partially accelerate models larger than the total VRAM capacity -The `llama.cpp` project is the main playground for developing new features for the [ggml](https://github.com/ggerganov/ggml) library. +The `llama.cpp` project is the main playground for developing new features for the [ggml](https://github.com/ggml-org/ggml) library.
Models @@ -59,23 +59,23 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo - [X] [Falcon](https://huggingface.co/models?search=tiiuae/falcon) - [X] [Chinese LLaMA / Alpaca](https://github.com/ymcui/Chinese-LLaMA-Alpaca) and [Chinese LLaMA-2 / Alpaca-2](https://github.com/ymcui/Chinese-LLaMA-Alpaca-2) - [X] [Vigogne (French)](https://github.com/bofenghuang/vigogne) -- [X] [BERT](https://github.com/ggerganov/llama.cpp/pull/5423) +- [X] [BERT](https://github.com/ggml-org/llama.cpp/pull/5423) - [X] [Koala](https://bair.berkeley.edu/blog/2023/04/03/koala/) - [X] [Baichuan 1 & 2](https://huggingface.co/models?search=baichuan-inc/Baichuan) + [derivations](https://huggingface.co/hiyouga/baichuan-7b-sft) - [X] [Aquila 1 & 2](https://huggingface.co/models?search=BAAI/Aquila) -- [X] [Starcoder models](https://github.com/ggerganov/llama.cpp/pull/3187) +- [X] [Starcoder models](https://github.com/ggml-org/llama.cpp/pull/3187) - [X] [Refact](https://huggingface.co/smallcloudai/Refact-1_6B-fim) -- [X] [MPT](https://github.com/ggerganov/llama.cpp/pull/3417) -- [X] [Bloom](https://github.com/ggerganov/llama.cpp/pull/3553) +- [X] [MPT](https://github.com/ggml-org/llama.cpp/pull/3417) +- [X] [Bloom](https://github.com/ggml-org/llama.cpp/pull/3553) - [x] [Yi models](https://huggingface.co/models?search=01-ai/Yi) - [X] [StableLM models](https://huggingface.co/stabilityai) - [x] [Deepseek models](https://huggingface.co/models?search=deepseek-ai/deepseek) - [x] [Qwen models](https://huggingface.co/models?search=Qwen/Qwen) -- [x] [PLaMo-13B](https://github.com/ggerganov/llama.cpp/pull/3557) +- [x] [PLaMo-13B](https://github.com/ggml-org/llama.cpp/pull/3557) - [x] [Phi models](https://huggingface.co/models?search=microsoft/phi) -- [x] [PhiMoE](https://github.com/ggerganov/llama.cpp/pull/11003) +- [x] [PhiMoE](https://github.com/ggml-org/llama.cpp/pull/11003) - [x] [GPT-2](https://huggingface.co/gpt2) -- [x] [Orion 14B](https://github.com/ggerganov/llama.cpp/pull/5118) +- [x] [Orion 14B](https://github.com/ggml-org/llama.cpp/pull/5118) - [x] [InternLM2](https://huggingface.co/models?search=internlm2) - [x] [CodeShell](https://github.com/WisdomShell/codeshell) - [x] [Gemma](https://ai.google.dev/gemma) @@ -97,6 +97,7 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo - [x] [Flan T5](https://huggingface.co/models?search=flan-t5) - [x] [Open Elm models](https://huggingface.co/collections/apple/openelm-instruct-models-6619ad295d7ae9f868b759ca) - [x] [ChatGLM3-6b](https://huggingface.co/THUDM/chatglm3-6b) + [ChatGLM4-9b](https://huggingface.co/THUDM/glm-4-9b) + [GLMEdge-1.5b](https://huggingface.co/THUDM/glm-edge-1.5b-chat) + [GLMEdge-4b](https://huggingface.co/THUDM/glm-edge-4b-chat) +- [x] [GLM-4-0414](https://huggingface.co/collections/THUDM/glm-4-0414-67f3cbcb34dd9d252707cb2e) - [x] [SmolLM](https://huggingface.co/collections/HuggingFaceTB/smollm-6695016cad7167254ce15966) - [x] [EXAONE-3.0-7.8B-Instruct](https://huggingface.co/LGAI-EXAONE/EXAONE-3.0-7.8B-Instruct) - [x] [FalconMamba Models](https://huggingface.co/collections/tiiuae/falconmamba-7b-66b9a580324dd1598b0f6d4a) @@ -105,6 +106,8 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo - [x] [RWKV-6](https://github.com/BlinkDL/RWKV-LM) - [x] [QRWKV-6](https://huggingface.co/recursal/QRWKV6-32B-Instruct-Preview-v0.1) - [x] [GigaChat-20B-A3B](https://huggingface.co/ai-sage/GigaChat-20B-A3B-instruct) +- [X] [Trillion-7B-preview](https://huggingface.co/trillionlabs/Trillion-7B-preview) +- [x] [Ling models](https://huggingface.co/collections/inclusionAI/ling-67c51c85b34a7ea0aba94c32) #### Multimodal @@ -146,10 +149,11 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo - Zig: [deins/llama.cpp.zig](https://github.com/Deins/llama.cpp.zig) - Flutter/Dart: [netdur/llama_cpp_dart](https://github.com/netdur/llama_cpp_dart) - Flutter: [xuegao-tzx/Fllama](https://github.com/xuegao-tzx/Fllama) -- PHP (API bindings and features built on top of llama.cpp): [distantmagic/resonance](https://github.com/distantmagic/resonance) [(more info)](https://github.com/ggerganov/llama.cpp/pull/6326) +- PHP (API bindings and features built on top of llama.cpp): [distantmagic/resonance](https://github.com/distantmagic/resonance) [(more info)](https://github.com/ggml-org/llama.cpp/pull/6326) - Guile Scheme: [guile_llama_cpp](https://savannah.nongnu.org/projects/guile-llama-cpp) - Swift [srgtuszy/llama-cpp-swift](https://github.com/srgtuszy/llama-cpp-swift) - Swift [ShenghaiWang/SwiftLlama](https://github.com/ShenghaiWang/SwiftLlama) +- Delphi [Embarcadero/llama-cpp-delphi](https://github.com/Embarcadero/llama-cpp-delphi)
@@ -164,6 +168,7 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo - [eva](https://github.com/ylsdamxssjxxdd/eva) (MIT) - [iohub/collama](https://github.com/iohub/coLLaMA) (Apache-2.0) - [janhq/jan](https://github.com/janhq/jan) (AGPL) +- [johnbean393/Sidekick](https://github.com/johnbean393/Sidekick) (MIT) - [KanTV](https://github.com/zhouwg/kantv?tab=readme-ov-file) (Apache-2.0) - [KodiBot](https://github.com/firatkiral/kodibot) (GPL) - [llama.vim](https://github.com/ggml-org/llama.vim) (MIT) @@ -212,7 +217,7 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo - [llama_cpp_canister](https://github.com/onicai/llama_cpp_canister) - llama.cpp as a smart contract on the Internet Computer, using WebAssembly - [llama-swap](https://github.com/mostlygeek/llama-swap) - transparent proxy that adds automatic model switching with llama-server - [Kalavai](https://github.com/kalavai-net/kalavai-client) - Crowdsource end to end LLM deployment at any scale - +- [llmaz](https://github.com/InftyAI/llmaz) - ☸️ Easy, advanced inference platform for large language models on Kubernetes.
@@ -235,6 +240,8 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo | [HIP](docs/build.md#hip) | AMD GPU | | [Vulkan](docs/build.md#vulkan) | GPU | | [CANN](docs/build.md#cann) | Ascend NPU | +| [OpenCL](docs/backend/OPENCL.md) | Adreno GPU | +| [RPC](https://github.com/ggml-org/llama.cpp/tree/master/examples/rpc) | All | ## Building the project @@ -244,7 +251,7 @@ The project also includes many example programs and tools using the `llama` libr - Clone this repository and build locally, see [how to build](docs/build.md) - On MacOS or Linux, install `llama.cpp` via [brew, flox or nix](docs/install.md) - Use a Docker image, see [documentation for Docker](docs/docker.md) -- Download pre-built binaries from [releases](https://github.com/ggerganov/llama.cpp/releases) +- Download pre-built binaries from [releases](https://github.com/ggml-org/llama.cpp/releases) ## Obtaining and quantizing models @@ -253,18 +260,20 @@ The [Hugging Face](https://huggingface.co) platform hosts a [number of LLMs](htt - [Trending](https://huggingface.co/models?library=gguf&sort=trending) - [LLaMA](https://huggingface.co/models?sort=trending&search=llama+gguf) -You can either manually download the GGUF file or directly use any `llama.cpp`-compatible models from Hugging Face by using this CLI argument: `-hf /[:quant]` +You can either manually download the GGUF file or directly use any `llama.cpp`-compatible models from [Hugging Face](https://huggingface.co/) or other model hosting sites, such as [ModelScope](https://modelscope.cn/), by using this CLI argument: `-hf /[:quant]`. + +By default, the CLI would download from Hugging Face, you can switch to other options with the environment variable `MODEL_ENDPOINT`. For example, you may opt to downloading model checkpoints from ModelScope or other model sharing communities by setting the environment variable, e.g. `MODEL_ENDPOINT=https://www.modelscope.cn/`. After downloading a model, use the CLI tools to run it locally - see below. -`llama.cpp` requires the model to be stored in the [GGUF](https://github.com/ggerganov/ggml/blob/master/docs/gguf.md) file format. Models in other data formats can be converted to GGUF using the `convert_*.py` Python scripts in this repo. +`llama.cpp` requires the model to be stored in the [GGUF](https://github.com/ggml-org/ggml/blob/master/docs/gguf.md) file format. Models in other data formats can be converted to GGUF using the `convert_*.py` Python scripts in this repo. The Hugging Face platform provides a variety of online tools for converting, quantizing and hosting models with `llama.cpp`: - Use the [GGUF-my-repo space](https://huggingface.co/spaces/ggml-org/gguf-my-repo) to convert to GGUF format and quantize model weights to smaller sizes -- Use the [GGUF-my-LoRA space](https://huggingface.co/spaces/ggml-org/gguf-my-lora) to convert LoRA adapters to GGUF format (more info: https://github.com/ggerganov/llama.cpp/discussions/10123) -- Use the [GGUF-editor space](https://huggingface.co/spaces/CISCai/gguf-editor) to edit GGUF meta data in the browser (more info: https://github.com/ggerganov/llama.cpp/discussions/9268) -- Use the [Inference Endpoints](https://ui.endpoints.huggingface.co/) to directly host `llama.cpp` in the cloud (more info: https://github.com/ggerganov/llama.cpp/discussions/9669) +- Use the [GGUF-my-LoRA space](https://huggingface.co/spaces/ggml-org/gguf-my-lora) to convert LoRA adapters to GGUF format (more info: https://github.com/ggml-org/llama.cpp/discussions/10123) +- Use the [GGUF-editor space](https://huggingface.co/spaces/CISCai/gguf-editor) to edit GGUF meta data in the browser (more info: https://github.com/ggml-org/llama.cpp/discussions/9268) +- Use the [Inference Endpoints](https://ui.endpoints.huggingface.co/) to directly host `llama.cpp` in the cloud (more info: https://github.com/ggml-org/llama.cpp/discussions/9669) To learn more about model quantization, [read this documentation](examples/quantize/README.md) @@ -487,9 +496,9 @@ To learn more about model quantization, [read this documentation](examples/quant - Collaborators can push to branches in the `llama.cpp` repo and merge PRs into the `master` branch - Collaborators will be invited based on contributions - Any help with managing issues, PRs and projects is very appreciated! -- See [good first issues](https://github.com/ggerganov/llama.cpp/issues?q=is%3Aissue+is%3Aopen+label%3A%22good+first+issue%22) for tasks suitable for first contributions +- See [good first issues](https://github.com/ggml-org/llama.cpp/issues?q=is%3Aissue+is%3Aopen+label%3A%22good+first+issue%22) for tasks suitable for first contributions - Read the [CONTRIBUTING.md](CONTRIBUTING.md) for more information -- Make sure to read this: [Inference at the edge](https://github.com/ggerganov/llama.cpp/discussions/205) +- Make sure to read this: [Inference at the edge](https://github.com/ggml-org/llama.cpp/discussions/205) - A bit of backstory for those who are interested: [Changelog podcast](https://changelog.com/podcast/532) ## Other documentation @@ -504,7 +513,7 @@ To learn more about model quantization, [read this documentation](examples/quant - [Running on Docker](docs/docker.md) - [Build on Android](docs/android.md) - [Performance troubleshooting](docs/development/token_generation_performance_tips.md) -- [GGML tips & tricks](https://github.com/ggerganov/llama.cpp/wiki/GGML-Tips-&-Tricks) +- [GGML tips & tricks](https://github.com/ggml-org/llama.cpp/wiki/GGML-Tips-&-Tricks) #### Seminal papers and background on the models @@ -518,5 +527,47 @@ If your issue is with model generation quality, then please at least scan the fo - [Aligning language models to follow instructions](https://openai.com/research/instruction-following) - [Training language models to follow instructions with human feedback](https://arxiv.org/abs/2203.02155) -#### References +## XCFramework +The XCFramework is a precompiled version of the library for iOS, visionOS, tvOS, +and macOS. It can be used in Swift projects without the need to compile the +library from source. For example: +```swift +// swift-tools-version: 5.10 +// The swift-tools-version declares the minimum version of Swift required to build this package. +import PackageDescription + +let package = Package( + name: "MyLlamaPackage", + targets: [ + .executableTarget( + name: "MyLlamaPackage", + dependencies: [ + "LlamaFramework" + ]), + .binaryTarget( + name: "LlamaFramework", + url: "https://github.com/ggml-org/llama.cpp/releases/download/b5046/llama-b5046-xcframework.zip", + checksum: "c19be78b5f00d8d29a25da41042cb7afa094cbf6280a225abe614b03b20029ab" + ) + ] +) +``` +The above example is using an intermediate build `b5046` of the library. This can be modified +to use a different version by changing the URL and checksum. + +## Completions +Command-line completion is available for some environments. + +#### Bash Completion +```bash +$ build/bin/llama-cli --completion-bash > ~/.llama-completion.bash +$ source ~/.llama-completion.bash +``` +Optionally this can be added to your `.bashrc` or `.bash_profile` to load it +automatically. For example: +```console +$ echo "source ~/.llama-completion.bash" >> ~/.bashrc +``` + +## References diff --git a/SECURITY.md b/SECURITY.md index f4322c6ee..6a1bb6c32 100644 --- a/SECURITY.md +++ b/SECURITY.md @@ -62,6 +62,6 @@ Beware that none of the topics under [Using llama.cpp securely](#using-llamacpp- However, If you have discovered a security vulnerability in this project, please report it privately. **Do not disclose it as a public issue.** This gives us time to work with you to fix the issue before public exposure, reducing the chance that the exploit will be used before a patch is released. -Please disclose it as a private [security advisory](https://github.com/ggerganov/llama.cpp/security/advisories/new). +Please disclose it as a private [security advisory](https://github.com/ggml-org/llama.cpp/security/advisories/new). A team of volunteers on a reasonable-effort basis maintains this project. As such, please give us at least 90 days to work on a fix before public exposure. diff --git a/Sources/llama/llama.h b/Sources/llama/llama.h deleted file mode 100644 index 41725880e..000000000 --- a/Sources/llama/llama.h +++ /dev/null @@ -1,4 +0,0 @@ -#pragma once - -#include - diff --git a/Sources/llama/module.modulemap b/Sources/llama/module.modulemap deleted file mode 100644 index d010555b1..000000000 --- a/Sources/llama/module.modulemap +++ /dev/null @@ -1,5 +0,0 @@ -module llama [system] { - header "llama.h" - link "llama" - export * -} diff --git a/build-xcframework.sh b/build-xcframework.sh new file mode 100755 index 000000000..97001b5f7 --- /dev/null +++ b/build-xcframework.sh @@ -0,0 +1,538 @@ +#!/bin/bash +# +# Options +IOS_MIN_OS_VERSION=16.4 +MACOS_MIN_OS_VERSION=13.3 +VISIONOS_MIN_OS_VERSION=1.0 +TVOS_MIN_OS_VERSION=16.4 + +BUILD_SHARED_LIBS=OFF +LLAMA_BUILD_EXAMPLES=OFF +LLAMA_BUILD_TESTS=OFF +LLAMA_BUILD_SERVER=OFF +GGML_METAL=ON +GGML_METAL_EMBED_LIBRARY=ON +GGML_BLAS_DEFAULT=ON +GGML_METAL_USE_BF16=ON +GGML_OPENMP=OFF + +COMMON_C_FLAGS="-Wno-macro-redefined -Wno-shorten-64-to-32 -Wno-unused-command-line-argument -g" +COMMON_CXX_FLAGS="-Wno-macro-redefined -Wno-shorten-64-to-32 -Wno-unused-command-line-argument -g" + +# Common options for all builds +COMMON_CMAKE_ARGS=( + -DCMAKE_XCODE_ATTRIBUTE_CODE_SIGNING_REQUIRED=NO + -DCMAKE_XCODE_ATTRIBUTE_CODE_SIGN_IDENTITY="" + -DCMAKE_XCODE_ATTRIBUTE_CODE_SIGNING_ALLOWED=NO + -DCMAKE_XCODE_ATTRIBUTE_DEBUG_INFORMATION_FORMAT="dwarf-with-dsym" + -DCMAKE_XCODE_ATTRIBUTE_GCC_GENERATE_DEBUGGING_SYMBOLS=YES + -DCMAKE_XCODE_ATTRIBUTE_COPY_PHASE_STRIP=NO + -DCMAKE_XCODE_ATTRIBUTE_STRIP_INSTALLED_PRODUCT=NO + -DCMAKE_XCODE_ATTRIBUTE_DEVELOPMENT_TEAM=ggml + -DBUILD_SHARED_LIBS=${BUILD_SHARED_LIBS} + -DLLAMA_BUILD_EXAMPLES=${LLAMA_BUILD_EXAMPLES} + -DLLAMA_BUILD_TESTS=${LLAMA_BUILD_TESTS} + -DLLAMA_BUILD_SERVER=${LLAMA_BUILD_SERVER} + -DGGML_METAL_EMBED_LIBRARY=${GGML_METAL_EMBED_LIBRARY} + -DGGML_BLAS_DEFAULT=${GGML_BLAS_DEFAULT} + -DGGML_METAL=${GGML_METAL} + -DGGML_METAL_USE_BF16=${GGML_METAL_USE_BF16} + -DGGML_NATIVE=OFF + -DGGML_OPENMP=${GGML_OPENMP} +) + +XCODE_VERSION=$(xcodebuild -version 2>/dev/null | head -n1 | awk '{ print $2 }') +MAJOR_VERSION=$(echo $XCODE_VERSION | cut -d. -f1) +MINOR_VERSION=$(echo $XCODE_VERSION | cut -d. -f2) +echo "Detected Xcode version: $XCODE_VERSION" + +check_required_tool() { + local tool=$1 + local install_message=$2 + + if ! command -v $tool &> /dev/null; then + echo "Error: $tool is required but not found." + echo "$install_message" + exit 1 + fi +} +echo "Checking for required tools..." +check_required_tool "cmake" "Please install CMake 3.28.0 or later (brew install cmake)" +check_required_tool "xcodebuild" "Please install Xcode and Xcode Command Line Tools (xcode-select --install)" +check_required_tool "libtool" "Please install libtool which should be available with Xcode Command Line Tools (CLT). Make sure Xcode CLT is installed (xcode-select --install)" +check_required_tool "dsymutil" "Please install Xcode and Xcode Command Line Tools (xcode-select --install)" + +set -e + +## Clean up previous builds +rm -rf build-apple +rm -rf build-ios-sim +rm -rf build-ios-device +rm -rf build-macos +rm -rf build-visionos +rm -rf build-visionos-sim +rm -rf build-tvos-sim +rm -rf build-tvos-device + +# Setup the xcframework build directory structure +setup_framework_structure() { + local build_dir=$1 + local min_os_version=$2 + local platform=$3 # "ios", "macos", "visionos", or "tvos" + local framework_name="llama" + + echo "Creating ${platform}-style framework structure for ${build_dir}" + + if [[ "$platform" == "macos" ]]; then + # macOS versioned structure uses versioned directories + mkdir -p ${build_dir}/framework/${framework_name}.framework/Versions/A/Headers + mkdir -p ${build_dir}/framework/${framework_name}.framework/Versions/A/Modules + mkdir -p ${build_dir}/framework/${framework_name}.framework/Versions/A/Resources + + # Create symbolic links + ln -sf A ${build_dir}/framework/${framework_name}.framework/Versions/Current + ln -sf Versions/Current/Headers ${build_dir}/framework/${framework_name}.framework/Headers + ln -sf Versions/Current/Modules ${build_dir}/framework/${framework_name}.framework/Modules + ln -sf Versions/Current/Resources ${build_dir}/framework/${framework_name}.framework/Resources + ln -sf Versions/Current/${framework_name} ${build_dir}/framework/${framework_name}.framework/${framework_name} + + # Set header and module paths + local header_path=${build_dir}/framework/${framework_name}.framework/Versions/A/Headers/ + local module_path=${build_dir}/framework/${framework_name}.framework/Versions/A/Modules/ + else + # iOS/VisionOS/tvOS use a flat structure + mkdir -p ${build_dir}/framework/${framework_name}.framework/Headers + mkdir -p ${build_dir}/framework/${framework_name}.framework/Modules + + # Remove any existing structure to ensure clean build + rm -rf ${build_dir}/framework/${framework_name}.framework/Versions + + # Set header and module paths + local header_path=${build_dir}/framework/${framework_name}.framework/Headers/ + local module_path=${build_dir}/framework/${framework_name}.framework/Modules/ + fi + + # Copy all required headers (common for all platforms) + cp include/llama.h ${header_path} + cp ggml/include/ggml.h ${header_path} + cp ggml/include/ggml-alloc.h ${header_path} + cp ggml/include/ggml-backend.h ${header_path} + cp ggml/include/ggml-metal.h ${header_path} + cp ggml/include/ggml-cpu.h ${header_path} + cp ggml/include/ggml-blas.h ${header_path} + cp ggml/include/gguf.h ${header_path} + + # Create module map (common for all platforms) + cat > ${module_path}module.modulemap << EOF +framework module llama { + header "llama.h" + header "ggml.h" + header "ggml-alloc.h" + header "ggml-backend.h" + header "ggml-metal.h" + header "ggml-cpu.h" + header "ggml-blas.h" + header "gguf.h" + + link "c++" + link framework "Accelerate" + link framework "Metal" + link framework "Foundation" + + export * +} +EOF + + # Platform-specific settings for Info.plist + local platform_name="" + local sdk_name="" + local supported_platform="" + + case "$platform" in + "ios") + platform_name="iphoneos" + sdk_name="iphoneos${min_os_version}" + supported_platform="iPhoneOS" + local plist_path="${build_dir}/framework/${framework_name}.framework/Info.plist" + local device_family=' UIDeviceFamily + + 1 + 2 + ' + ;; + "macos") + platform_name="macosx" + sdk_name="macosx${min_os_version}" + supported_platform="MacOSX" + local plist_path="${build_dir}/framework/${framework_name}.framework/Versions/A/Resources/Info.plist" + local device_family="" + ;; + "visionos") + platform_name="xros" + sdk_name="xros${min_os_version}" + supported_platform="XRPlatform" + local plist_path="${build_dir}/framework/${framework_name}.framework/Info.plist" + local device_family="" + ;; + "tvos") + platform_name="appletvos" + sdk_name="appletvos${min_os_version}" + supported_platform="AppleTVOS" + local plist_path="${build_dir}/framework/${framework_name}.framework/Info.plist" + local device_family=' UIDeviceFamily + + 3 + ' + ;; + esac + + # Create Info.plist + cat > ${plist_path} << EOF + + + + + CFBundleDevelopmentRegion + en + CFBundleExecutable + llama + CFBundleIdentifier + org.ggml.llama + CFBundleInfoDictionaryVersion + 6.0 + CFBundleName + llama + CFBundlePackageType + FMWK + CFBundleShortVersionString + 1.0 + CFBundleVersion + 1 + MinimumOSVersion + ${min_os_version} + CFBundleSupportedPlatforms + + ${supported_platform} + ${device_family} + DTPlatformName + ${platform_name} + DTSDKName + ${sdk_name} + + +EOF +} + +# Create dynamic libraries from static libraries. +combine_static_libraries() { + local build_dir="$1" + local release_dir="$2" + local platform="$3" # "ios", "macos", "visionos", or "tvos" + local is_simulator="$4" + local base_dir="$(pwd)" + local framework_name="llama" + + # Determine output path based on platform + local output_lib="" + if [[ "$platform" == "macos" ]]; then + # macOS uses versioned structure + output_lib="${build_dir}/framework/${framework_name}.framework/Versions/A/${framework_name}" + else + # iOS, visionOS, and tvOS use a directory flat structure + output_lib="${build_dir}/framework/${framework_name}.framework/${framework_name}" + fi + + local libs=( + "${base_dir}/${build_dir}/src/${release_dir}/libllama.a" + "${base_dir}/${build_dir}/ggml/src/${release_dir}/libggml.a" + "${base_dir}/${build_dir}/ggml/src/${release_dir}/libggml-base.a" + "${base_dir}/${build_dir}/ggml/src/${release_dir}/libggml-cpu.a" + "${base_dir}/${build_dir}/ggml/src/ggml-metal/${release_dir}/libggml-metal.a" + "${base_dir}/${build_dir}/ggml/src/ggml-blas/${release_dir}/libggml-blas.a" + ) + + # Create temporary directory for processing + local temp_dir="${base_dir}/${build_dir}/temp" + mkdir -p "${temp_dir}" + + # Since we have multiple architectures libtool will find object files that do not + # match the target architecture. We suppress these warnings. + libtool -static -o "${temp_dir}/combined.a" "${libs[@]}" 2> /dev/null + + # Determine SDK, architectures, and install_name based on platform and simulator flag. + local sdk="" + local archs="" + local min_version_flag="" + local install_name="" + + case "$platform" in + "ios") + if [[ "$is_simulator" == "true" ]]; then + sdk="iphonesimulator" + archs="arm64 x86_64" + min_version_flag="-mios-simulator-version-min=${IOS_MIN_OS_VERSION}" + else + sdk="iphoneos" + archs="arm64" + min_version_flag="-mios-version-min=${IOS_MIN_OS_VERSION}" + fi + install_name="@rpath/llama.framework/llama" + ;; + "macos") + sdk="macosx" + archs="arm64 x86_64" + min_version_flag="-mmacosx-version-min=${MACOS_MIN_OS_VERSION}" + install_name="@rpath/llama.framework/Versions/Current/llama" + ;; + "visionos") + if [[ "$is_simulator" == "true" ]]; then + sdk="xrsimulator" + archs="arm64 x86_64" + min_version_flag="-mtargetos=xros${VISIONOS_MIN_OS_VERSION}-simulator" + else + sdk="xros" + archs="arm64" + min_version_flag="-mtargetos=xros${VISIONOS_MIN_OS_VERSION}" + fi + # Use flat structure for visionOS, same as iOS + install_name="@rpath/llama.framework/llama" + ;; + "tvos") + if [[ "$is_simulator" == "true" ]]; then + sdk="appletvsimulator" + archs="arm64 x86_64" + min_version_flag="-mtvos-simulator-version-min=${TVOS_MIN_OS_VERSION}" + else + sdk="appletvos" + archs="arm64" + min_version_flag="-mtvos-version-min=${TVOS_MIN_OS_VERSION}" + fi + install_name="@rpath/llama.framework/llama" + ;; + esac + + # Build architecture flags + local arch_flags="" + for arch in $archs; do + arch_flags+=" -arch $arch" + done + + # Create dynamic library + echo "Creating dynamic library for ${platform}." + xcrun -sdk $sdk clang++ -dynamiclib \ + -isysroot $(xcrun --sdk $sdk --show-sdk-path) \ + $arch_flags \ + $min_version_flag \ + -Wl,-force_load,"${temp_dir}/combined.a" \ + -framework Foundation -framework Metal -framework Accelerate \ + -install_name "$install_name" \ + -o "${base_dir}/${output_lib}" + + # Platform-specific post-processing for device builds + if [[ "$is_simulator" == "false" ]]; then + if command -v xcrun vtool &>/dev/null; then + case "$platform" in + "ios") + echo "Marking binary as a framework binary for iOS..." + xcrun vtool -set-build-version ios ${IOS_MIN_OS_VERSION} ${IOS_MIN_OS_VERSION} -replace \ + -output "${base_dir}/${output_lib}" "${base_dir}/${output_lib}" + ;; + "visionos") + echo "Marking binary as a framework binary for visionOS..." + if [[ "$MAJOR_VERSION" -gt 16 ]] || [[ "$MAJOR_VERSION" -eq 16 && "$MINOR_VERSION" -gt 2 ]]; then + echo "Xcode version greater than 16.2, using visionOS." + VISION_OS_BUILD_VERSION="visionos" + else + echo "Xcode version less than or equal to 16.2, using xros." + VISION_OS_BUILD_VERSION="xros" + fi + xcrun vtool -set-build-version ${VISION_OS_BUILD_VERSION} ${VISIONOS_MIN_OS_VERSION} ${VISIONOS_MIN_OS_VERSION} -replace \ + -output "${base_dir}/${output_lib}" "${base_dir}/${output_lib}" + ;; + "tvos") + echo "Marking binary as a framework binary for tvOS..." + xcrun vtool -set-build-version tvos ${TVOS_MIN_OS_VERSION} ${TVOS_MIN_OS_VERSION} -replace \ + -output "${base_dir}/${output_lib}" "${base_dir}/${output_lib}" + ;; + esac + else + echo "Warning: vtool not found. Binary may not pass App Store validation." + fi + fi + + echo "Creating properly formatted dSYM..." + # Create a separate directory for dSYMs for all platforms + mkdir -p "${base_dir}/${build_dir}/dSYMs" + + # iOS and visionOS style dSYM (flat structure) + if [[ "$platform" == "ios" || "$platform" == "visionos" || "$platform" == "tvos" ]]; then + # Generate dSYM in the dSYMs directory + xcrun dsymutil "${base_dir}/${output_lib}" -o "${base_dir}/${build_dir}/dSYMs/llama.dSYM" + + # Create a copy of the binary that will be stripped + cp "${base_dir}/${output_lib}" "${temp_dir}/binary_to_strip" + + # Strip debug symbols from the copy + xcrun strip -S "${temp_dir}/binary_to_strip" -o "${temp_dir}/stripped_lib" + + # Replace the original with the stripped version + mv "${temp_dir}/stripped_lib" "${base_dir}/${output_lib}" + else + # macOS style dSYM + # First strip debug info to a separate file + xcrun strip -S "${base_dir}/${output_lib}" -o "${temp_dir}/stripped_lib" + + # Generate dSYM in the dSYMs directory + xcrun dsymutil "${base_dir}/${output_lib}" -o "${base_dir}/${build_dir}/dSYMs/llama.dSYM" + + # Replace original binary with stripped version + mv "${temp_dir}/stripped_lib" "${base_dir}/${output_lib}" + fi + + # Remove any automatically generated dSYM files in the framework structure as they will + # otherwise case Invalid Bundle Structure validation errors. + if [ -d "${base_dir}/${output_lib}.dSYM" ]; then + echo "Removing generated dSYM file in framework structure: ${base_dir}/${output_lib}.dSYM" + rm -rf "${base_dir}/${output_lib}.dSYM" + fi + + # Clean up + rm -rf "${temp_dir}" +} + +echo "Building for iOS simulator..." +cmake -B build-ios-sim -G Xcode \ + "${COMMON_CMAKE_ARGS[@]}" \ + -DCMAKE_OSX_DEPLOYMENT_TARGET=${IOS_MIN_OS_VERSION} \ + -DIOS=ON \ + -DCMAKE_SYSTEM_NAME=iOS \ + -DCMAKE_OSX_SYSROOT=iphonesimulator \ + -DCMAKE_OSX_ARCHITECTURES="arm64;x86_64" \ + -DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=iphonesimulator \ + -DCMAKE_C_FLAGS="${COMMON_C_FLAGS}" \ + -DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \ + -DLLAMA_CURL=OFF \ + -S . +cmake --build build-ios-sim --config Release -- -quiet + +echo "Building for iOS devices..." +cmake -B build-ios-device -G Xcode \ + "${COMMON_CMAKE_ARGS[@]}" \ + -DCMAKE_OSX_DEPLOYMENT_TARGET=${IOS_MIN_OS_VERSION} \ + -DCMAKE_OSX_SYSROOT=iphoneos \ + -DCMAKE_OSX_ARCHITECTURES="arm64" \ + -DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=iphoneos \ + -DCMAKE_C_FLAGS="${COMMON_C_FLAGS}" \ + -DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \ + -DLLAMA_CURL=OFF \ + -S . +cmake --build build-ios-device --config Release -- -quiet + +echo "Building for macOS..." +cmake -B build-macos -G Xcode \ + "${COMMON_CMAKE_ARGS[@]}" \ + -DCMAKE_OSX_DEPLOYMENT_TARGET=${MACOS_MIN_OS_VERSION} \ + -DCMAKE_OSX_ARCHITECTURES="arm64;x86_64" \ + -DCMAKE_C_FLAGS="${COMMON_C_FLAGS}" \ + -DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \ + -DLLAMA_CURL=OFF \ + -S . +cmake --build build-macos --config Release -- -quiet + +echo "Building for visionOS..." +cmake -B build-visionos -G Xcode \ + "${COMMON_CMAKE_ARGS[@]}" \ + -DCMAKE_OSX_DEPLOYMENT_TARGET=${VISIONOS_MIN_OS_VERSION} \ + -DCMAKE_OSX_ARCHITECTURES="arm64" \ + -DCMAKE_SYSTEM_NAME=visionOS \ + -DCMAKE_OSX_SYSROOT=xros \ + -DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=xros \ + -DCMAKE_C_FLAGS="-D_XOPEN_SOURCE=700 ${COMMON_C_FLAGS}" \ + -DCMAKE_CXX_FLAGS="-D_XOPEN_SOURCE=700 ${COMMON_CXX_FLAGS}" \ + -DLLAMA_CURL=OFF \ + -S . +cmake --build build-visionos --config Release -- -quiet + +echo "Building for visionOS simulator..." +cmake -B build-visionos-sim -G Xcode \ + "${COMMON_CMAKE_ARGS[@]}" \ + -DCMAKE_OSX_DEPLOYMENT_TARGET=${VISIONOS_MIN_OS_VERSION} \ + -DCMAKE_OSX_ARCHITECTURES="arm64;x86_64" \ + -DCMAKE_SYSTEM_NAME=visionOS \ + -DCMAKE_OSX_SYSROOT=xrsimulator \ + -DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=xrsimulator \ + -DCMAKE_C_FLAGS="-D_XOPEN_SOURCE=700 ${COMMON_C_FLAGS}" \ + -DCMAKE_CXX_FLAGS="-D_XOPEN_SOURCE=700 ${COMMON_CXX_FLAGS}" \ + -DLLAMA_CURL=OFF \ + -S . +cmake --build build-visionos-sim --config Release -- -quiet + +# Add tvOS builds (might need the same u_int definitions as watchOS and visionOS) +echo "Building for tvOS simulator..." +cmake -B build-tvos-sim -G Xcode \ + "${COMMON_CMAKE_ARGS[@]}" \ + -DCMAKE_OSX_DEPLOYMENT_TARGET=${TVOS_MIN_OS_VERSION} \ + -DCMAKE_SYSTEM_NAME=tvOS \ + -DCMAKE_OSX_SYSROOT=appletvsimulator \ + -DCMAKE_OSX_ARCHITECTURES="arm64;x86_64" \ + -DGGML_METAL=ON \ + -DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=appletvsimulator \ + -DCMAKE_C_FLAGS="${COMMON_C_FLAGS}" \ + -DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \ + -DLLAMA_CURL=OFF \ + -S . +cmake --build build-tvos-sim --config Release -- -quiet + +echo "Building for tvOS devices..." +cmake -B build-tvos-device -G Xcode \ + "${COMMON_CMAKE_ARGS[@]}" \ + -DCMAKE_OSX_DEPLOYMENT_TARGET=${TVOS_MIN_OS_VERSION} \ + -DCMAKE_SYSTEM_NAME=tvOS \ + -DCMAKE_OSX_SYSROOT=appletvos \ + -DCMAKE_OSX_ARCHITECTURES="arm64" \ + -DGGML_METAL=ON \ + -DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=appletvos \ + -DCMAKE_C_FLAGS="${COMMON_C_FLAGS}" \ + -DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \ + -DLLAMA_CURL=OFF \ + -S . +cmake --build build-tvos-device --config Release -- -quiet + +# Setup frameworks and copy binaries and headers +echo "Setting up framework structures..." +setup_framework_structure "build-ios-sim" ${IOS_MIN_OS_VERSION} "ios" +setup_framework_structure "build-ios-device" ${IOS_MIN_OS_VERSION} "ios" +setup_framework_structure "build-macos" ${MACOS_MIN_OS_VERSION} "macos" +setup_framework_structure "build-visionos" ${VISIONOS_MIN_OS_VERSION} "visionos" +setup_framework_structure "build-visionos-sim" ${VISIONOS_MIN_OS_VERSION} "visionos" +setup_framework_structure "build-tvos-sim" ${TVOS_MIN_OS_VERSION} "tvos" +setup_framework_structure "build-tvos-device" ${TVOS_MIN_OS_VERSION} "tvos" + +# Create dynamic libraries from static libraries +echo "Creating dynamic libraries from static libraries..." +combine_static_libraries "build-ios-sim" "Release-iphonesimulator" "ios" "true" +combine_static_libraries "build-ios-device" "Release-iphoneos" "ios" "false" +combine_static_libraries "build-macos" "Release" "macos" "false" +combine_static_libraries "build-visionos" "Release-xros" "visionos" "false" +combine_static_libraries "build-visionos-sim" "Release-xrsimulator" "visionos" "true" +combine_static_libraries "build-tvos-sim" "Release-appletvsimulator" "tvos" "true" +combine_static_libraries "build-tvos-device" "Release-appletvos" "tvos" "false" + +# Create XCFramework with correct debug symbols paths +echo "Creating XCFramework..." +xcodebuild -create-xcframework \ + -framework $(pwd)/build-ios-sim/framework/llama.framework \ + -debug-symbols $(pwd)/build-ios-sim/dSYMs/llama.dSYM \ + -framework $(pwd)/build-ios-device/framework/llama.framework \ + -debug-symbols $(pwd)/build-ios-device/dSYMs/llama.dSYM \ + -framework $(pwd)/build-macos/framework/llama.framework \ + -debug-symbols $(pwd)/build-macos/dSYMS/llama.dSYM \ + -framework $(pwd)/build-visionos/framework/llama.framework \ + -debug-symbols $(pwd)/build-visionos/dSYMs/llama.dSYM \ + -framework $(pwd)/build-visionos-sim/framework/llama.framework \ + -debug-symbols $(pwd)/build-visionos-sim/dSYMs/llama.dSYM \ + -framework $(pwd)/build-tvos-device/framework/llama.framework \ + -debug-symbols $(pwd)/build-tvos-device/dSYMs/llama.dSYM \ + -framework $(pwd)/build-tvos-sim/framework/llama.framework \ + -debug-symbols $(pwd)/build-tvos-sim/dSYMs/llama.dSYM \ + -output $(pwd)/build-apple/llama.xcframework diff --git a/ci/README.md b/ci/README.md index 406470519..ec3f44350 100644 --- a/ci/README.md +++ b/ci/README.md @@ -1,11 +1,11 @@ # CI -In addition to [Github Actions](https://github.com/ggerganov/llama.cpp/actions) `llama.cpp` uses a custom CI framework: +In addition to [Github Actions](https://github.com/ggml-org/llama.cpp/actions) `llama.cpp` uses a custom CI framework: https://github.com/ggml-org/ci It monitors the `master` branch for new commits and runs the -[ci/run.sh](https://github.com/ggerganov/llama.cpp/blob/master/ci/run.sh) script on dedicated cloud instances. This allows us +[ci/run.sh](https://github.com/ggml-org/llama.cpp/blob/master/ci/run.sh) script on dedicated cloud instances. This allows us to execute heavier workloads compared to just using Github Actions. Also with time, the cloud instances will be scaled to cover various hardware architectures, including GPU and Apple Silicon instances. @@ -26,4 +26,43 @@ GG_BUILD_CUDA=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt # with SYCL support source /opt/intel/oneapi/setvars.sh GG_BUILD_SYCL=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt + +# with MUSA support +GG_BUILD_MUSA=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt ``` + +## Running MUSA CI in a Docker Container + +Assuming `$PWD` is the root of the `llama.cpp` repository, follow these steps to set up and run MUSA CI in a Docker container: + +### 1. Create a local directory to store cached models, configuration files and venv: + +```bash +mkdir -p $HOME/llama.cpp/ci-cache +``` + +### 2. Create a local directory to store CI run results: + +```bash +mkdir -p $HOME/llama.cpp/ci-results +``` + +### 3. Start a Docker container and run the CI: + +```bash +docker run --privileged -it \ + -v $HOME/llama.cpp/ci-cache:/ci-cache \ + -v $HOME/llama.cpp/ci-results:/ci-results \ + -v $PWD:/ws -w /ws \ + mthreads/musa:rc3.1.1-devel-ubuntu22.04 +``` + +Inside the container, execute the following commands: + +```bash +apt update -y && apt install -y bc cmake ccache git python3.10-venv time unzip wget +git config --global --add safe.directory /ws +GG_BUILD_MUSA=1 bash ./ci/run.sh /ci-results /ci-cache +``` + +This setup ensures that the CI runs within an isolated Docker environment while maintaining cached files and results across runs. diff --git a/ci/run.sh b/ci/run.sh index 77c32ce00..f463d7a8b 100755 --- a/ci/run.sh +++ b/ci/run.sh @@ -16,6 +16,9 @@ # # with VULKAN support # GG_BUILD_VULKAN=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt # +# # with MUSA support +# GG_BUILD_MUSA=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt +# if [ -z "$2" ]; then echo "usage: $0 " @@ -36,7 +39,7 @@ sd=`dirname $0` cd $sd/../ SRC=`pwd` -CMAKE_EXTRA="-DLLAMA_FATAL_WARNINGS=ON" +CMAKE_EXTRA="-DLLAMA_FATAL_WARNINGS=ON -DLLAMA_CURL=OFF" if [ ! -z ${GG_BUILD_METAL} ]; then CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_METAL=ON -DGGML_METAL_USE_BF16=ON" @@ -52,13 +55,24 @@ if [ ! -z ${GG_BUILD_SYCL} ]; then echo "source /opt/intel/oneapi/setvars.sh" exit 1 fi - + # Use only main GPU + export ONEAPI_DEVICE_SELECTOR="level_zero:0" + # Enable sysman for correct memory reporting + export ZES_ENABLE_SYSMAN=1 + # to circumvent precision issues on CPY operations + export SYCL_PROGRAM_COMPILE_OPTIONS="-cl-fp32-correctly-rounded-divide-sqrt" CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_SYCL=1 -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DGGML_SYCL_F16=ON" fi if [ ! -z ${GG_BUILD_VULKAN} ]; then CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_VULKAN=1" fi + +if [ ! -z ${GG_BUILD_MUSA} ]; then + # Use qy1 by default (MTT S80) + MUSA_ARCH=${MUSA_ARCH:-21} + CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_MUSA=ON -DMUSA_ARCHITECTURES=${MUSA_ARCH}" +fi ## helpers # download a file if it does not exist or if it is outdated @@ -352,10 +366,10 @@ function gg_run_open_llama_7b_v2 { (time ./bin/llama-imatrix --model ${model_f16} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log - (time ./bin/llama-save-load-state--model ${model_q4_0} -ngl 10 -c 0 ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log - (time ./bin/llama-save-load-state--model ${model_q4_0} -ngl 10 -c 0 -fa ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log - (time ./bin/llama-save-load-state--model ${model_q4_0} -ngl 99 -c 0 ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log - (time ./bin/llama-save-load-state--model ${model_q4_0} -ngl 99 -c 0 -fa ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log + (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 10 -c 0 ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log + (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 10 -c 0 -fa ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log + (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 99 -c 0 ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log + (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 99 -c 0 -fa ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log function check_ppl { qnt="$1" @@ -808,7 +822,7 @@ export LLAMA_LOG_PREFIX=1 export LLAMA_LOG_TIMESTAMPS=1 if [ -z ${GG_BUILD_LOW_PERF} ]; then - # Create symlink: ./llama.cpp/models-mnt -> $MNT/models/models-mnt + # Create symlink: ./llama.cpp/models-mnt -> $MNT/models rm -rf ${SRC}/models-mnt mnt_models=${MNT}/models mkdir -p ${mnt_models} @@ -826,8 +840,10 @@ if [ -z ${GG_BUILD_LOW_PERF} ]; then fi ret=0 - -test $ret -eq 0 && gg_run ctest_debug +if [ -z ${GG_BUILD_SYCL} ]; then + # SYCL build breaks with debug build flags + test $ret -eq 0 && gg_run ctest_debug +fi test $ret -eq 0 && gg_run ctest_release if [ -z ${GG_BUILD_LOW_PERF} ]; then @@ -835,7 +851,9 @@ if [ -z ${GG_BUILD_LOW_PERF} ]; then test $ret -eq 0 && gg_run rerank_tiny if [ -z ${GG_BUILD_CLOUD} ] || [ ${GG_BUILD_EXTRA_TESTS_0} ]; then - test $ret -eq 0 && gg_run test_scripts_debug + if [ -z ${GG_BUILD_SYCL} ]; then + test $ret -eq 0 && gg_run test_scripts_debug + fi test $ret -eq 0 && gg_run test_scripts_release fi @@ -846,7 +864,9 @@ if [ -z ${GG_BUILD_LOW_PERF} ]; then test $ret -eq 0 && gg_run pythia_2_8b #test $ret -eq 0 && gg_run open_llama_7b_v2 fi - test $ret -eq 0 && gg_run ctest_with_model_debug + if [ -z ${GG_BUILD_SYCL} ]; then + test $ret -eq 0 && gg_run ctest_with_model_debug + fi test $ret -eq 0 && gg_run ctest_with_model_release fi fi diff --git a/cmake/common.cmake b/cmake/common.cmake index 0f54871e4..a5bb787f1 100644 --- a/cmake/common.cmake +++ b/cmake/common.cmake @@ -1,3 +1,5 @@ +include("ggml/cmake/common.cmake") + function(llama_add_compile_flags) if (LLAMA_FATAL_WARNINGS) if (CMAKE_CXX_COMPILER_ID MATCHES "GNU" OR CMAKE_CXX_COMPILER_ID MATCHES "Clang") diff --git a/common/CMakeLists.txt b/common/CMakeLists.txt index e61015d2a..43533fc86 100644 --- a/common/CMakeLists.txt +++ b/common/CMakeLists.txt @@ -57,8 +57,7 @@ add_library(${TARGET} STATIC arg.h base64.hpp chat.cpp - chat.hpp - chat-template.hpp + chat.h common.cpp common.h console.cpp @@ -68,7 +67,8 @@ add_library(${TARGET} STATIC llguidance.cpp log.cpp log.h - minja.hpp + minja/chat-template.hpp + minja/minja.hpp ngram-cache.cpp ngram-cache.h sampling.cpp @@ -85,7 +85,10 @@ set(LLAMA_COMMON_EXTRA_LIBS build_info) # Use curl to download model url if (LLAMA_CURL) - find_package(CURL REQUIRED) + find_package(CURL) + if (NOT CURL_FOUND) + message(FATAL_ERROR "Could NOT find CURL. Hint: to disable this feature, set -DLLAMA_CURL=OFF") + endif() target_compile_definitions(${TARGET} PUBLIC LLAMA_USE_CURL) include_directories(${CURL_INCLUDE_DIRS}) find_library(CURL_LIBRARY curl REQUIRED) @@ -96,27 +99,44 @@ if (LLAMA_LLGUIDANCE) include(ExternalProject) set(LLGUIDANCE_SRC ${CMAKE_BINARY_DIR}/llguidance/source) set(LLGUIDANCE_PATH ${LLGUIDANCE_SRC}/target/release) + + # Set the correct library file extension based on platform + if (WIN32) + set(LLGUIDANCE_LIB_NAME "llguidance.lib") + # Add Windows-specific libraries + set(LLGUIDANCE_PLATFORM_LIBS + ws2_32 # Windows Sockets API + userenv # For GetUserProfileDirectoryW + ntdll # For NT functions + bcrypt # For BCryptGenRandom + ) + else() + set(LLGUIDANCE_LIB_NAME "libllguidance.a") + set(LLGUIDANCE_PLATFORM_LIBS "") + endif() + ExternalProject_Add(llguidance_ext GIT_REPOSITORY https://github.com/guidance-ai/llguidance - # v0.6.12: - GIT_TAG ced1c9023d47ec194fa977932d35ce65c2ebfc09 + # v0.7.10: + GIT_TAG 0309d2a6bf40abda35344a362edc71e06d5009f8 PREFIX ${CMAKE_BINARY_DIR}/llguidance SOURCE_DIR ${LLGUIDANCE_SRC} BUILD_IN_SOURCE TRUE CONFIGURE_COMMAND "" BUILD_COMMAND cargo build --release INSTALL_COMMAND "" - BUILD_BYPRODUCTS ${LLGUIDANCE_PATH}/libllguidance.a ${LLGUIDANCE_PATH}/llguidance.h + BUILD_BYPRODUCTS ${LLGUIDANCE_PATH}/${LLGUIDANCE_LIB_NAME} ${LLGUIDANCE_PATH}/llguidance.h UPDATE_COMMAND "" ) target_compile_definitions(${TARGET} PUBLIC LLAMA_USE_LLGUIDANCE) add_library(llguidance STATIC IMPORTED) - set_target_properties(llguidance PROPERTIES IMPORTED_LOCATION ${LLGUIDANCE_PATH}/libllguidance.a) + set_target_properties(llguidance PROPERTIES IMPORTED_LOCATION ${LLGUIDANCE_PATH}/${LLGUIDANCE_LIB_NAME}) add_dependencies(llguidance llguidance_ext) target_include_directories(${TARGET} PRIVATE ${LLGUIDANCE_PATH}) - set(LLAMA_COMMON_EXTRA_LIBS ${LLAMA_COMMON_EXTRA_LIBS} llguidance) + # Add platform libraries to the main target + set(LLAMA_COMMON_EXTRA_LIBS ${LLAMA_COMMON_EXTRA_LIBS} llguidance ${LLGUIDANCE_PLATFORM_LIBS}) endif () target_include_directories(${TARGET} PUBLIC .) diff --git a/common/arg.cpp b/common/arg.cpp index 152f671ab..0b57f9da1 100644 --- a/common/arg.cpp +++ b/common/arg.cpp @@ -1,11 +1,24 @@ +#include "gguf.h" // for reading GGUF splits #include "arg.h" +#include "common.h" #include "log.h" #include "sampling.h" +#include "chat.h" + +// fix problem with std::min and std::max +#if defined(_WIN32) +#define WIN32_LEAN_AND_MEAN +#ifndef NOMINMAX +# define NOMINMAX +#endif +#include +#endif #include #include #include +#include #include #include #include @@ -13,6 +26,14 @@ #include #include +//#define LLAMA_USE_CURL + +#if defined(LLAMA_USE_CURL) +#include +#include +#include +#endif + #include "json-schema-to-grammar.h" using json = nlohmann::ordered_json; @@ -124,47 +145,554 @@ std::string common_arg::to_string() { return ss.str(); } +// +// downloader +// + +struct common_hf_file_res { + std::string repo; // repo name with ":tag" removed + std::string ggufFile; + std::string mmprojFile; +}; + +#ifdef LLAMA_USE_CURL + +#ifdef __linux__ +#include +#elif defined(_WIN32) +# if !defined(PATH_MAX) +# define PATH_MAX MAX_PATH +# endif +#elif defined(_AIX) +#include +#else +#include +#endif +#define LLAMA_CURL_MAX_URL_LENGTH 2084 // Maximum URL Length in Chrome: 2083 + +// +// CURL utils +// + +using curl_ptr = std::unique_ptr; + +// cannot use unique_ptr for curl_slist, because we cannot update without destroying the old one +struct curl_slist_ptr { + struct curl_slist * ptr = nullptr; + ~curl_slist_ptr() { + if (ptr) { + curl_slist_free_all(ptr); + } + } +}; + +#define CURL_MAX_RETRY 3 +#define CURL_RETRY_DELAY_SECONDS 2 + +static bool curl_perform_with_retry(const std::string & url, CURL * curl, int max_attempts, int retry_delay_seconds) { + int remaining_attempts = max_attempts; + + while (remaining_attempts > 0) { + LOG_INF("%s: Trying to download from %s (attempt %d of %d)...\n", __func__ , url.c_str(), max_attempts - remaining_attempts + 1, max_attempts); + + CURLcode res = curl_easy_perform(curl); + if (res == CURLE_OK) { + return true; + } + + int exponential_backoff_delay = std::pow(retry_delay_seconds, max_attempts - remaining_attempts) * 1000; + LOG_WRN("%s: curl_easy_perform() failed: %s, retrying after %d milliseconds...\n", __func__, curl_easy_strerror(res), exponential_backoff_delay); + + remaining_attempts--; + std::this_thread::sleep_for(std::chrono::milliseconds(exponential_backoff_delay)); + } + + LOG_ERR("%s: curl_easy_perform() failed after %d attempts\n", __func__, max_attempts); + + return false; +} + +// download one single file from remote URL to local path +static bool common_download_file_single(const std::string & url, const std::string & path, const std::string & bearer_token) { + // Initialize libcurl + curl_ptr curl(curl_easy_init(), &curl_easy_cleanup); + curl_slist_ptr http_headers; + if (!curl) { + LOG_ERR("%s: error initializing libcurl\n", __func__); + return false; + } + + bool force_download = false; + + // Set the URL, allow to follow http redirection + curl_easy_setopt(curl.get(), CURLOPT_URL, url.c_str()); + curl_easy_setopt(curl.get(), CURLOPT_FOLLOWLOCATION, 1L); + + http_headers.ptr = curl_slist_append(http_headers.ptr, "User-Agent: llama-cpp"); + // Check if hf-token or bearer-token was specified + if (!bearer_token.empty()) { + std::string auth_header = "Authorization: Bearer " + bearer_token; + http_headers.ptr = curl_slist_append(http_headers.ptr, auth_header.c_str()); + } + curl_easy_setopt(curl.get(), CURLOPT_HTTPHEADER, http_headers.ptr); + +#if defined(_WIN32) + // CURLSSLOPT_NATIVE_CA tells libcurl to use standard certificate store of + // operating system. Currently implemented under MS-Windows. + curl_easy_setopt(curl.get(), CURLOPT_SSL_OPTIONS, CURLSSLOPT_NATIVE_CA); +#endif + + // Check if the file already exists locally + auto file_exists = std::filesystem::exists(path); + + // If the file exists, check its JSON metadata companion file. + std::string metadata_path = path + ".json"; + nlohmann::json metadata; + std::string etag; + std::string last_modified; + + if (file_exists) { + // Try and read the JSON metadata file (note: stream autoclosed upon exiting this block). + std::ifstream metadata_in(metadata_path); + if (metadata_in.good()) { + try { + metadata_in >> metadata; + LOG_INF("%s: previous metadata file found %s: %s\n", __func__, metadata_path.c_str(), metadata.dump().c_str()); + if (metadata.contains("url") && metadata.at("url").is_string()) { + auto previous_url = metadata.at("url").get(); + if (previous_url != url) { + LOG_ERR("%s: Model URL mismatch: %s != %s\n", __func__, url.c_str(), previous_url.c_str()); + return false; + } + } + if (metadata.contains("etag") && metadata.at("etag").is_string()) { + etag = metadata.at("etag"); + } + if (metadata.contains("lastModified") && metadata.at("lastModified").is_string()) { + last_modified = metadata.at("lastModified"); + } + } catch (const nlohmann::json::exception & e) { + LOG_ERR("%s: error reading metadata file %s: %s\n", __func__, metadata_path.c_str(), e.what()); + return false; + } + } + } else { + LOG_INF("%s: no previous model file found %s\n", __func__, path.c_str()); + } + + // Send a HEAD request to retrieve the etag and last-modified headers + struct common_load_model_from_url_headers { + std::string etag; + std::string last_modified; + }; + + common_load_model_from_url_headers headers; + + { + typedef size_t(*CURLOPT_HEADERFUNCTION_PTR)(char *, size_t, size_t, void *); + auto header_callback = [](char * buffer, size_t /*size*/, size_t n_items, void * userdata) -> size_t { + common_load_model_from_url_headers * headers = (common_load_model_from_url_headers *) userdata; + + static std::regex header_regex("([^:]+): (.*)\r\n"); + static std::regex etag_regex("ETag", std::regex_constants::icase); + static std::regex last_modified_regex("Last-Modified", std::regex_constants::icase); + + std::string header(buffer, n_items); + std::smatch match; + if (std::regex_match(header, match, header_regex)) { + const std::string & key = match[1]; + const std::string & value = match[2]; + if (std::regex_match(key, match, etag_regex)) { + headers->etag = value; + } else if (std::regex_match(key, match, last_modified_regex)) { + headers->last_modified = value; + } + } + return n_items; + }; + + curl_easy_setopt(curl.get(), CURLOPT_NOBODY, 1L); // will trigger the HEAD verb + curl_easy_setopt(curl.get(), CURLOPT_NOPROGRESS, 1L); // hide head request progress + curl_easy_setopt(curl.get(), CURLOPT_HEADERFUNCTION, static_cast(header_callback)); + curl_easy_setopt(curl.get(), CURLOPT_HEADERDATA, &headers); + + bool was_perform_successful = curl_perform_with_retry(url, curl.get(), CURL_MAX_RETRY, CURL_RETRY_DELAY_SECONDS); + if (!was_perform_successful) { + return false; + } + + long http_code = 0; + curl_easy_getinfo(curl.get(), CURLINFO_RESPONSE_CODE, &http_code); + if (http_code != 200) { + // HEAD not supported, we don't know if the file has changed + // force trigger downloading + force_download = true; + LOG_ERR("%s: HEAD invalid http status code received: %ld\n", __func__, http_code); + } + } + + bool should_download = !file_exists || force_download; + if (!should_download) { + if (!etag.empty() && etag != headers.etag) { + LOG_WRN("%s: ETag header is different (%s != %s): triggering a new download\n", __func__, etag.c_str(), headers.etag.c_str()); + should_download = true; + } else if (!last_modified.empty() && last_modified != headers.last_modified) { + LOG_WRN("%s: Last-Modified header is different (%s != %s): triggering a new download\n", __func__, last_modified.c_str(), headers.last_modified.c_str()); + should_download = true; + } + } + if (should_download) { + std::string path_temporary = path + ".downloadInProgress"; + if (file_exists) { + LOG_WRN("%s: deleting previous downloaded file: %s\n", __func__, path.c_str()); + if (remove(path.c_str()) != 0) { + LOG_ERR("%s: unable to delete file: %s\n", __func__, path.c_str()); + return false; + } + } + + // Set the output file + + struct FILE_deleter { + void operator()(FILE * f) const { + fclose(f); + } + }; + + std::unique_ptr outfile(fopen(path_temporary.c_str(), "wb")); + if (!outfile) { + LOG_ERR("%s: error opening local file for writing: %s\n", __func__, path.c_str()); + return false; + } + + typedef size_t(*CURLOPT_WRITEFUNCTION_PTR)(void * data, size_t size, size_t nmemb, void * fd); + auto write_callback = [](void * data, size_t size, size_t nmemb, void * fd) -> size_t { + return fwrite(data, size, nmemb, (FILE *)fd); + }; + curl_easy_setopt(curl.get(), CURLOPT_NOBODY, 0L); + curl_easy_setopt(curl.get(), CURLOPT_WRITEFUNCTION, static_cast(write_callback)); + curl_easy_setopt(curl.get(), CURLOPT_WRITEDATA, outfile.get()); + + // display download progress + curl_easy_setopt(curl.get(), CURLOPT_NOPROGRESS, 0L); + + // helper function to hide password in URL + auto llama_download_hide_password_in_url = [](const std::string & url) -> std::string { + std::size_t protocol_pos = url.find("://"); + if (protocol_pos == std::string::npos) { + return url; // Malformed URL + } + + std::size_t at_pos = url.find('@', protocol_pos + 3); + if (at_pos == std::string::npos) { + return url; // No password in URL + } + + return url.substr(0, protocol_pos + 3) + "********" + url.substr(at_pos); + }; + + // start the download + LOG_INF("%s: trying to download model from %s to %s (server_etag:%s, server_last_modified:%s)...\n", __func__, + llama_download_hide_password_in_url(url).c_str(), path.c_str(), headers.etag.c_str(), headers.last_modified.c_str()); + bool was_perform_successful = curl_perform_with_retry(url, curl.get(), CURL_MAX_RETRY, CURL_RETRY_DELAY_SECONDS); + if (!was_perform_successful) { + return false; + } + + long http_code = 0; + curl_easy_getinfo (curl.get(), CURLINFO_RESPONSE_CODE, &http_code); + if (http_code < 200 || http_code >= 400) { + LOG_ERR("%s: invalid http status code received: %ld\n", __func__, http_code); + return false; + } + + // Causes file to be closed explicitly here before we rename it. + outfile.reset(); + + // Write the updated JSON metadata file. + metadata.update({ + {"url", url}, + {"etag", headers.etag}, + {"lastModified", headers.last_modified} + }); + std::ofstream(metadata_path) << metadata.dump(4); + LOG_INF("%s: file metadata saved: %s\n", __func__, metadata_path.c_str()); + + if (rename(path_temporary.c_str(), path.c_str()) != 0) { + LOG_ERR("%s: unable to rename file: %s to %s\n", __func__, path_temporary.c_str(), path.c_str()); + return false; + } + } + + return true; +} + +// download multiple files from remote URLs to local paths +// the input is a vector of pairs +static bool common_download_file_multiple(const std::vector> & urls, const std::string & bearer_token) { + // Prepare download in parallel + std::vector> futures_download; + for (auto const & item : urls) { + futures_download.push_back(std::async(std::launch::async, [bearer_token](const std::pair & it) -> bool { + return common_download_file_single(it.first, it.second, bearer_token); + }, item)); + } + + // Wait for all downloads to complete + for (auto & f : futures_download) { + if (!f.get()) { + return false; + } + } + + return true; +} + +static bool common_download_model( + const common_params_model & model, + const std::string & bearer_token) { + // Basic validation of the model.url + if (model.url.empty()) { + LOG_ERR("%s: invalid model url\n", __func__); + return false; + } + + if (!common_download_file_single(model.url, model.path, bearer_token)) { + return false; + } + + // check for additional GGUFs split to download + int n_split = 0; + { + struct gguf_init_params gguf_params = { + /*.no_alloc = */ true, + /*.ctx = */ NULL, + }; + auto * ctx_gguf = gguf_init_from_file(model.path.c_str(), gguf_params); + if (!ctx_gguf) { + LOG_ERR("\n%s: failed to load input GGUF from %s\n", __func__, model.path.c_str()); + return false; + } + + auto key_n_split = gguf_find_key(ctx_gguf, LLM_KV_SPLIT_COUNT); + if (key_n_split >= 0) { + n_split = gguf_get_val_u16(ctx_gguf, key_n_split); + } + + gguf_free(ctx_gguf); + } + + if (n_split > 1) { + char split_prefix[PATH_MAX] = {0}; + char split_url_prefix[LLAMA_CURL_MAX_URL_LENGTH] = {0}; + + // Verify the first split file format + // and extract split URL and PATH prefixes + { + if (!llama_split_prefix(split_prefix, sizeof(split_prefix), model.path.c_str(), 0, n_split)) { + LOG_ERR("\n%s: unexpected model file name: %s n_split=%d\n", __func__, model.path.c_str(), n_split); + return false; + } + + if (!llama_split_prefix(split_url_prefix, sizeof(split_url_prefix), model.url.c_str(), 0, n_split)) { + LOG_ERR("\n%s: unexpected model url: %s n_split=%d\n", __func__, model.url.c_str(), n_split); + return false; + } + } + + std::vector> urls; + for (int idx = 1; idx < n_split; idx++) { + char split_path[PATH_MAX] = {0}; + llama_split_path(split_path, sizeof(split_path), split_prefix, idx, n_split); + + char split_url[LLAMA_CURL_MAX_URL_LENGTH] = {0}; + llama_split_path(split_url, sizeof(split_url), split_url_prefix, idx, n_split); + + if (std::string(split_path) == model.path) { + continue; // skip the already downloaded file + } + + urls.push_back({split_url, split_path}); + } + + // Download in parallel + common_download_file_multiple(urls, bearer_token); + } + + return true; +} + +/** + * Allow getting the HF file from the HF repo with tag (like ollama), for example: + * - bartowski/Llama-3.2-3B-Instruct-GGUF:q4 + * - bartowski/Llama-3.2-3B-Instruct-GGUF:Q4_K_M + * - bartowski/Llama-3.2-3B-Instruct-GGUF:q5_k_s + * Tag is optional, default to "latest" (meaning it checks for Q4_K_M first, then Q4, then if not found, return the first GGUF file in repo) + * + * Return pair of (with "repo" already having tag removed) + * + * Note: we use the Ollama-compatible HF API, but not using the blobId. Instead, we use the special "ggufFile" field which returns the value for "hf_file". This is done to be backward-compatible with existing cache files. + */ +static struct common_hf_file_res common_get_hf_file(const std::string & hf_repo_with_tag, const std::string & bearer_token) { + auto parts = string_split(hf_repo_with_tag, ':'); + std::string tag = parts.size() > 1 ? parts.back() : "latest"; + std::string hf_repo = parts[0]; + if (string_split(hf_repo, '/').size() != 2) { + throw std::invalid_argument("error: invalid HF repo format, expected /[:quant]\n"); + } + + // fetch model info from Hugging Face Hub API + curl_ptr curl(curl_easy_init(), &curl_easy_cleanup); + curl_slist_ptr http_headers; + std::string res_str; + + std::string model_endpoint = get_model_endpoint(); + + std::string url = model_endpoint + "v2/" + hf_repo + "/manifests/" + tag; + curl_easy_setopt(curl.get(), CURLOPT_URL, url.c_str()); + curl_easy_setopt(curl.get(), CURLOPT_NOPROGRESS, 1L); + typedef size_t(*CURLOPT_WRITEFUNCTION_PTR)(void * ptr, size_t size, size_t nmemb, void * data); + auto write_callback = [](void * ptr, size_t size, size_t nmemb, void * data) -> size_t { + static_cast(data)->append((char * ) ptr, size * nmemb); + return size * nmemb; + }; + curl_easy_setopt(curl.get(), CURLOPT_WRITEFUNCTION, static_cast(write_callback)); + curl_easy_setopt(curl.get(), CURLOPT_WRITEDATA, &res_str); +#if defined(_WIN32) + curl_easy_setopt(curl.get(), CURLOPT_SSL_OPTIONS, CURLSSLOPT_NATIVE_CA); +#endif + if (!bearer_token.empty()) { + std::string auth_header = "Authorization: Bearer " + bearer_token; + http_headers.ptr = curl_slist_append(http_headers.ptr, auth_header.c_str()); + } + // Important: the User-Agent must be "llama-cpp" to get the "ggufFile" field in the response + http_headers.ptr = curl_slist_append(http_headers.ptr, "User-Agent: llama-cpp"); + http_headers.ptr = curl_slist_append(http_headers.ptr, "Accept: application/json"); + curl_easy_setopt(curl.get(), CURLOPT_HTTPHEADER, http_headers.ptr); + + CURLcode res = curl_easy_perform(curl.get()); + + if (res != CURLE_OK) { + throw std::runtime_error("error: cannot make GET request to HF API"); + } + + long res_code; + std::string ggufFile = ""; + std::string mmprojFile = ""; + curl_easy_getinfo(curl.get(), CURLINFO_RESPONSE_CODE, &res_code); + if (res_code == 200) { + // extract ggufFile.rfilename in json, using regex + { + std::regex pattern("\"ggufFile\"[\\s\\S]*?\"rfilename\"\\s*:\\s*\"([^\"]+)\""); + std::smatch match; + if (std::regex_search(res_str, match, pattern)) { + ggufFile = match[1].str(); + } + } + // extract mmprojFile.rfilename in json, using regex + { + std::regex pattern("\"mmprojFile\"[\\s\\S]*?\"rfilename\"\\s*:\\s*\"([^\"]+)\""); + std::smatch match; + if (std::regex_search(res_str, match, pattern)) { + mmprojFile = match[1].str(); + } + } + } else if (res_code == 401) { + throw std::runtime_error("error: model is private or does not exist; if you are accessing a gated model, please provide a valid HF token"); + } else { + throw std::runtime_error(string_format("error from HF API, response code: %ld, data: %s", res_code, res_str.c_str())); + } + + // check response + if (ggufFile.empty()) { + throw std::runtime_error("error: model does not have ggufFile"); + } + + return { hf_repo, ggufFile, mmprojFile }; +} + +#else + +static bool common_download_file_single(const std::string &, const std::string &, const std::string &) { + LOG_ERR("error: built without CURL, cannot download model from internet\n"); + return false; +} + +static bool common_download_file_multiple(const std::vector> &, const std::string &) { + LOG_ERR("error: built without CURL, cannot download model from the internet\n"); + return false; +} + +static bool common_download_model( + const common_params_model &, + const std::string &) { + LOG_ERR("error: built without CURL, cannot download model from the internet\n"); + return false; +} + +static struct common_hf_file_res common_get_hf_file(const std::string &, const std::string &) { + LOG_ERR("error: built without CURL, cannot download model from the internet\n"); + return {}; +} + +#endif // LLAMA_USE_CURL + // // utils // -static void common_params_handle_model_default( - std::string & model, - const std::string & model_url, - std::string & hf_repo, - std::string & hf_file, - const std::string & hf_token, - const std::string & model_default) { - if (!hf_repo.empty()) { - // short-hand to avoid specifying --hf-file -> default it to --model - if (hf_file.empty()) { - if (model.empty()) { - auto auto_detected = common_get_hf_file(hf_repo, hf_token); - if (auto_detected.first.empty() || auto_detected.second.empty()) { - exit(1); // built without CURL, error message already printed +static void common_params_handle_model( + struct common_params_model & model, + const std::string & bearer_token, + const std::string & model_path_default, + bool is_mmproj = false) { // TODO: move is_mmproj to an enum when we have more files? + // handle pre-fill default model path and url based on hf_repo and hf_file + { + if (!model.hf_repo.empty()) { + // short-hand to avoid specifying --hf-file -> default it to --model + if (model.hf_file.empty()) { + if (model.path.empty()) { + auto auto_detected = common_get_hf_file(model.hf_repo, bearer_token); + if (auto_detected.repo.empty() || auto_detected.ggufFile.empty()) { + exit(1); // built without CURL, error message already printed + } + model.hf_repo = auto_detected.repo; + model.hf_file = is_mmproj ? auto_detected.mmprojFile : auto_detected.ggufFile; + } else { + model.hf_file = model.path; } - hf_repo = auto_detected.first; - hf_file = auto_detected.second; - } else { - hf_file = model; } + + std::string model_endpoint = get_model_endpoint(); + model.url = model_endpoint + model.hf_repo + "/resolve/main/" + model.hf_file; + // make sure model path is present (for caching purposes) + if (model.path.empty()) { + // this is to avoid different repo having same file name, or same file name in different subdirs + std::string filename = model.hf_repo + "_" + model.hf_file; + // to make sure we don't have any slashes in the filename + string_replace_all(filename, "/", "_"); + model.path = fs_get_cache_file(filename); + } + + } else if (!model.url.empty()) { + if (model.path.empty()) { + auto f = string_split(model.url, '#').front(); + f = string_split(f, '?').front(); + model.path = fs_get_cache_file(string_split(f, '/').back()); + } + + } else if (model.path.empty()) { + model.path = model_path_default; } - // make sure model path is present (for caching purposes) - if (model.empty()) { - // this is to avoid different repo having same file name, or same file name in different subdirs - std::string filename = hf_repo + "_" + hf_file; - // to make sure we don't have any slashes in the filename - string_replace_all(filename, "/", "_"); - model = fs_get_cache_file(filename); + } + + // then, download it if needed + if (!model.url.empty()) { + bool ok = common_download_model(model, bearer_token); + if (!ok) { + LOG_ERR("error: failed to download model from %s\n", model.url.c_str()); + exit(1); } - } else if (!model_url.empty()) { - if (model.empty()) { - auto f = string_split(model_url, '#').front(); - f = string_split(f, '?').front(); - model = fs_get_cache_file(string_split(f, '/').back()); - } - } else if (model.empty()) { - model = model_default; } } @@ -299,10 +827,16 @@ static bool common_params_parse_ex(int argc, char ** argv, common_params_context throw std::invalid_argument("error: --prompt-cache-all not supported in interactive mode yet\n"); } - // TODO: refactor model params in a common struct - common_params_handle_model_default(params.model, params.model_url, params.hf_repo, params.hf_file, params.hf_token, DEFAULT_MODEL_PATH); - common_params_handle_model_default(params.speculative.model, params.speculative.model_url, params.speculative.hf_repo, params.speculative.hf_file, params.hf_token, ""); - common_params_handle_model_default(params.vocoder.model, params.vocoder.model_url, params.vocoder.hf_repo, params.vocoder.hf_file, params.hf_token, ""); + common_params_handle_model(params.model, params.hf_token, DEFAULT_MODEL_PATH); + common_params_handle_model(params.speculative.model, params.hf_token, ""); + common_params_handle_model(params.vocoder.model, params.hf_token, ""); + + // allow --mmproj to be set from -hf + // assuming that mmproj is always in the same repo as text model + if (!params.model.hf_repo.empty() && ctx_arg.ex == LLAMA_EXAMPLE_LLAVA) { + params.mmproj.hf_repo = params.model.hf_repo; + } + common_params_handle_model(params.mmproj, params.hf_token, "", true); if (params.escape) { string_process_escapes(params.prompt); @@ -321,6 +855,10 @@ static bool common_params_parse_ex(int argc, char ** argv, common_params_context params.kv_overrides.back().key[0] = 0; } + if (!params.tensor_buft_overrides.empty()) { + params.tensor_buft_overrides.push_back({nullptr, nullptr}); + } + if (params.reranking && params.embedding) { throw std::invalid_argument("error: either --embedding or --reranking can be specified, but not both"); } @@ -365,6 +903,112 @@ static void common_params_print_usage(common_params_context & ctx_arg) { print_options(specific_options); } +static void common_params_print_completion(common_params_context & ctx_arg) { + std::vector common_options; + std::vector sparam_options; + std::vector specific_options; + + for (auto & opt : ctx_arg.options) { + if (opt.is_sparam) { + sparam_options.push_back(&opt); + } else if (opt.in_example(ctx_arg.ex)) { + specific_options.push_back(&opt); + } else { + common_options.push_back(&opt); + } + } + + printf("_llama_completions() {\n"); + printf(" local cur prev opts\n"); + printf(" COMPREPLY=()\n"); + printf(" cur=\"${COMP_WORDS[COMP_CWORD]}\"\n"); + printf(" prev=\"${COMP_WORDS[COMP_CWORD-1]}\"\n\n"); + + printf(" opts=\""); + auto print_options = [](const std::vector & options) { + for (const common_arg * opt : options) { + for (const char * arg : opt->args) { + printf("%s ", arg); + } + } + }; + + print_options(common_options); + print_options(sparam_options); + print_options(specific_options); + printf("\"\n\n"); + + printf(" case \"$prev\" in\n"); + printf(" --model)\n"); + printf(" COMPREPLY=( $(compgen -f -X '!*.gguf' -- \"$cur\") $(compgen -d -- \"$cur\") )\n"); + printf(" return 0\n"); + printf(" ;;\n"); + printf(" --grammar-file)\n"); + printf(" COMPREPLY=( $(compgen -f -X '!*.gbnf' -- \"$cur\") $(compgen -d -- \"$cur\") )\n"); + printf(" return 0\n"); + printf(" ;;\n"); + printf(" --chat-template-file)\n"); + printf(" COMPREPLY=( $(compgen -f -X '!*.jinja' -- \"$cur\") $(compgen -d -- \"$cur\") )\n"); + printf(" return 0\n"); + printf(" ;;\n"); + printf(" *)\n"); + printf(" COMPREPLY=( $(compgen -W \"${opts}\" -- \"$cur\") )\n"); + printf(" return 0\n"); + printf(" ;;\n"); + printf(" esac\n"); + printf("}\n\n"); + + std::set executables = { + "llama-batched", + "llama-batched-bench", + "llama-bench", + "llama-cli", + "llama-convert-llama2c-to-ggml", + "llama-cvector-generator", + "llama-embedding", + "llama-eval-callback", + "llama-export-lora", + "llama-gbnf-validator", + "llama-gen-docs", + "llama-gguf", + "llama-gguf-hash", + "llama-gguf-split", + "llama-gritlm", + "llama-imatrix", + "llama-infill", + "llama-llava-cli", + "llama-llava-clip-quantize-cli", + "llama-lookahead", + "llama-lookup", + "llama-lookup-create", + "llama-lookup-merge", + "llama-lookup-stats", + "llama-minicpmv-cli", + "llama-parallel", + "llama-passkey", + "llama-perplexity", + "llama-q8dot", + "llama-quantize", + "llama-quantize-stats", + "llama-qwen2vl-cli", + "llama-retrieval", + "llama-run", + "llama-save-load-state", + "llama-server", + "llama-simple", + "llama-simple-chat", + "llama-speculative", + "llama-speculative-simple", + "llama-tokenize", + "llama-tts", + "llama-vdot" + }; + + for (const auto& exe : executables) { + printf("complete -F _llama_completions %s\n", exe.c_str()); + } +} + static std::vector parse_device_list(const std::string & value) { std::vector devices; auto dev_names = string_split(value, ','); @@ -426,6 +1070,10 @@ bool common_params_parse(int argc, char ** argv, common_params & params, llama_e } exit(0); } + if (ctx_arg.params.completion) { + common_params_print_completion(ctx_arg); + exit(0); + } } catch (const std::invalid_argument & ex) { fprintf(stderr, "%s\n", ex.what()); ctx_arg.params = params_org; @@ -494,6 +1142,13 @@ common_params_context common_params_parser_init(common_params & params, llama_ex exit(0); } )); + add_opt(common_arg( + {"--completion-bash"}, + "print source-able bash completion script for llama.cpp", + [](common_params & params) { + params.completion = true; + } + )); add_opt(common_arg( {"--verbose-prompt"}, string_format("print a verbose prompt before generation (default: %s)", params.verbose_prompt ? "true" : "false"), @@ -646,7 +1301,11 @@ common_params_context common_params_parser_init(common_params & params, llama_ex ).set_env("LLAMA_ARG_CTX_SIZE")); add_opt(common_arg( {"-n", "--predict", "--n-predict"}, "N", - string_format("number of tokens to predict (default: %d, -1 = infinity, -2 = until context filled)", params.n_predict), + string_format( + ex == LLAMA_EXAMPLE_MAIN || ex == LLAMA_EXAMPLE_INFILL + ? "number of tokens to predict (default: %d, -1 = infinity, -2 = until context filled)" + : "number of tokens to predict (default: %d, -1 = infinity)", + params.n_predict), [](common_params & params, int value) { params.n_predict = value; } @@ -674,7 +1333,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex )); add_opt(common_arg( {"--no-context-shift"}, - string_format("disables context shift on inifinite text generation (default: %s)", params.ctx_shift ? "disabled" : "enabled"), + string_format("disables context shift on infinite text generation (default: %s)", params.ctx_shift ? "disabled" : "enabled"), [](common_params & params) { params.ctx_shift = false; } @@ -695,13 +1354,18 @@ common_params_context common_params_parser_init(common_params & params, llama_ex ).set_env("LLAMA_ARG_FLASH_ATTN")); add_opt(common_arg( {"-p", "--prompt"}, "PROMPT", - ex == LLAMA_EXAMPLE_MAIN - ? "prompt to start generation with\nif -cnv is set, this will be used as system prompt" - : "prompt to start generation with", + "prompt to start generation with; for system message, use -sys", [](common_params & params, const std::string & value) { params.prompt = value; } ).set_excludes({LLAMA_EXAMPLE_SERVER})); + add_opt(common_arg( + {"-sys", "--system-prompt"}, "PROMPT", + "system prompt to use with model (if applicable, depending on chat template)", + [](common_params & params, const std::string & value) { + params.system_prompt = value; + } + ).set_examples({LLAMA_EXAMPLE_MAIN})); add_opt(common_arg( {"--no-perf"}, string_format("disable internal libllama performance timings (default: %s)", params.no_perf ? "true" : "false"), @@ -726,6 +1390,20 @@ common_params_context common_params_parser_init(common_params & params, llama_ex } } ).set_excludes({LLAMA_EXAMPLE_SERVER})); + add_opt(common_arg( + {"-sysf", "--system-prompt-file"}, "FNAME", + "a file containing the system prompt (default: none)", + [](common_params & params, const std::string & value) { + std::ifstream file(value); + if (!file) { + throw std::runtime_error(string_format("error: failed to open file '%s'\n", value.c_str())); + } + std::copy(std::istreambuf_iterator(file), std::istreambuf_iterator(), back_inserter(params.system_prompt)); + if (!params.system_prompt.empty() && params.system_prompt.back() == '\n') { + params.system_prompt.pop_back(); + } + } + ).set_examples({LLAMA_EXAMPLE_MAIN})); add_opt(common_arg( {"--in-file"}, "FNAME", "an input file (repeat to specify multiple files)", @@ -826,6 +1504,15 @@ common_params_context common_params_parser_init(common_params & params, llama_ex params.conversation_mode = COMMON_CONVERSATION_MODE_DISABLED; } ).set_examples({LLAMA_EXAMPLE_MAIN})); + add_opt(common_arg( + {"-st", "--single-turn"}, + "run conversation for a single turn only, then exit when done\n" + "will not be interactive if first turn is predefined with --prompt\n" + "(default: false)", + [](common_params & params) { + params.single_turn = true; + } + ).set_examples({LLAMA_EXAMPLE_MAIN})); add_opt(common_arg( {"-i", "--interactive"}, string_format("run in interactive mode (default: %s)", params.interactive ? "true" : "false"), @@ -946,6 +1633,13 @@ common_params_context common_params_parser_init(common_params & params, llama_ex params.sampling.min_p = std::stof(value); } ).set_sparam()); + add_opt(common_arg( + {"--top-nsigma"}, "N", + string_format("top-n-sigma sampling (default: %.1f, -1.0 = disabled)", params.sampling.top_n_sigma), + [](common_params & params, const std::string & value) { + params.sampling.top_n_sigma = std::stof(value); + } + ).set_examples({LLAMA_EXAMPLE_MAIN}).set_sparam()); add_opt(common_arg( {"--xtc-probability"}, "N", string_format("xtc probability (default: %.1f, 0.0 = disabled)", (double)params.sampling.xtc_probability), @@ -1404,7 +2098,14 @@ common_params_context common_params_parser_init(common_params & params, llama_ex {"--mmproj"}, "FILE", "path to a multimodal projector file for LLaVA. see examples/llava/README.md", [](common_params & params, const std::string & value) { - params.mmproj = value; + params.mmproj.path = value; + } + ).set_examples({LLAMA_EXAMPLE_LLAVA})); + add_opt(common_arg( + {"--mmproj-url"}, "URL", + "URL to a multimodal projector file for LLaVA. see examples/llava/README.md", + [](common_params & params, const std::string & value) { + params.mmproj.url = value; } ).set_examples({LLAMA_EXAMPLE_LLAVA})); add_opt(common_arg( @@ -1445,7 +2146,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex "- isolate: only spawn threads on CPUs on the node that execution started on\n" "- numactl: use the CPU map provided by numactl\n" "if run without this previously, it is recommended to drop the system page cache before using this\n" - "see https://github.com/ggerganov/llama.cpp/issues/1437", + "see https://github.com/ggml-org/llama.cpp/issues/1437", [](common_params & params, const std::string & value) { /**/ if (value == "distribute" || value == "") { params.numa = GGML_NUMA_STRATEGY_DISTRIBUTE; } else if (value == "isolate") { params.numa = GGML_NUMA_STRATEGY_ISOLATE; } @@ -1490,6 +2191,41 @@ common_params_context common_params_parser_init(common_params & params, llama_ex exit(0); } )); + add_opt(common_arg( + {"--override-tensor", "-ot"}, "=,...", + "override tensor buffer type", [](common_params & params, const std::string & value) { + /* static */ std::map buft_list; + if (buft_list.empty()) { + // enumerate all the devices and add their buffer types to the list + for (size_t i = 0; i < ggml_backend_dev_count(); ++i) { + auto * dev = ggml_backend_dev_get(i); + auto * buft = ggml_backend_dev_buffer_type(dev); + if (buft) { + buft_list[ggml_backend_buft_name(buft)] = buft; + } + } + } + + for (const auto & override : string_split(value, ',')) { + std::string::size_type pos = override.find('='); + if (pos == std::string::npos) { + throw std::invalid_argument("invalid value"); + } + std::string tensor_name = override.substr(0, pos); + std::string buffer_type = override.substr(pos + 1); + + if (buft_list.find(buffer_type) == buft_list.end()) { + printf("Available buffer types:\n"); + for (const auto & it : buft_list) { + printf(" %s\n", ggml_backend_buft_name(it.second)); + } + throw std::invalid_argument("unknown buffer type"); + } + // FIXME: this leaks memory + params.tensor_buft_overrides.push_back({strdup(tensor_name.c_str()), buft_list.at(buffer_type)}); + } + } + )); add_opt(common_arg( {"-ngl", "--gpu-layers", "--n-gpu-layers"}, "N", "number of layers to store in VRAM", @@ -1633,14 +2369,14 @@ common_params_context common_params_parser_init(common_params & params, llama_ex "or `--model-url` if set, otherwise %s)", DEFAULT_MODEL_PATH ), [](common_params & params, const std::string & value) { - params.model = value; + params.model.path = value; } ).set_examples({LLAMA_EXAMPLE_COMMON, LLAMA_EXAMPLE_EXPORT_LORA}).set_env("LLAMA_ARG_MODEL")); add_opt(common_arg( {"-mu", "--model-url"}, "MODEL_URL", "model download url (default: unused)", [](common_params & params, const std::string & value) { - params.model_url = value; + params.model.url = value; } ).set_env("LLAMA_ARG_MODEL_URL")); add_opt(common_arg( @@ -1649,35 +2385,35 @@ common_params_context common_params_parser_init(common_params & params, llama_ex "example: unsloth/phi-4-GGUF:q4_k_m\n" "(default: unused)", [](common_params & params, const std::string & value) { - params.hf_repo = value; + params.model.hf_repo = value; } ).set_env("LLAMA_ARG_HF_REPO")); add_opt(common_arg( {"-hfd", "-hfrd", "--hf-repo-draft"}, "/[:quant]", "Same as --hf-repo, but for the draft model (default: unused)", [](common_params & params, const std::string & value) { - params.speculative.hf_repo = value; + params.speculative.model.hf_repo = value; } ).set_env("LLAMA_ARG_HFD_REPO")); add_opt(common_arg( {"-hff", "--hf-file"}, "FILE", "Hugging Face model file. If specified, it will override the quant in --hf-repo (default: unused)", [](common_params & params, const std::string & value) { - params.hf_file = value; + params.model.hf_file = value; } ).set_env("LLAMA_ARG_HF_FILE")); add_opt(common_arg( {"-hfv", "-hfrv", "--hf-repo-v"}, "/[:quant]", "Hugging Face model repository for the vocoder model (default: unused)", [](common_params & params, const std::string & value) { - params.vocoder.hf_repo = value; + params.vocoder.model.hf_repo = value; } ).set_env("LLAMA_ARG_HF_REPO_V")); add_opt(common_arg( {"-hffv", "--hf-file-v"}, "FILE", "Hugging Face model file for the vocoder model (default: unused)", [](common_params & params, const std::string & value) { - params.vocoder.hf_file = value; + params.vocoder.model.hf_file = value; } ).set_env("LLAMA_ARG_HF_FILE_V")); add_opt(common_arg( @@ -1728,18 +2464,11 @@ common_params_context common_params_parser_init(common_params & params, llama_ex ).set_examples({LLAMA_EXAMPLE_PASSKEY})); add_opt(common_arg( {"-o", "--output", "--output-file"}, "FNAME", - string_format("output file (default: '%s')", - ex == LLAMA_EXAMPLE_EXPORT_LORA - ? params.lora_outfile.c_str() - : ex == LLAMA_EXAMPLE_CVECTOR_GENERATOR - ? params.cvector_outfile.c_str() - : params.out_file.c_str()), + string_format("output file (default: '%s')", params.out_file.c_str()), [](common_params & params, const std::string & value) { params.out_file = value; - params.cvector_outfile = value; - params.lora_outfile = value; } - ).set_examples({LLAMA_EXAMPLE_IMATRIX, LLAMA_EXAMPLE_CVECTOR_GENERATOR, LLAMA_EXAMPLE_EXPORT_LORA})); + ).set_examples({LLAMA_EXAMPLE_IMATRIX, LLAMA_EXAMPLE_CVECTOR_GENERATOR, LLAMA_EXAMPLE_EXPORT_LORA, LLAMA_EXAMPLE_TTS})); add_opt(common_arg( {"-ofreq", "--output-frequency"}, "N", string_format("output the imatrix every N iterations (default: %d)", params.n_out_freq), @@ -1829,7 +2558,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex ).set_examples({LLAMA_EXAMPLE_EMBEDDING})); add_opt(common_arg( {"--host"}, "HOST", - string_format("ip address to listen (default: %s)", params.hostname.c_str()), + string_format("ip address to listen, or bind to an UNIX socket if the address ends with .sock (default: %s)", params.hostname.c_str()), [](common_params & params, const std::string & value) { params.hostname = value; } @@ -1975,6 +2704,17 @@ common_params_context common_params_parser_init(common_params & params, llama_ex params.use_jinja = true; } ).set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_MAIN}).set_env("LLAMA_ARG_JINJA")); + add_opt(common_arg( + {"--reasoning-format"}, "FORMAT", + "reasoning format (default: deepseek; allowed values: deepseek, none)\n" + "controls whether thought tags are extracted from the response, and in which format they're returned. 'none' leaves thoughts unparsed in `message.content`, 'deepseek' puts them in `message.reasoning_content` (for DeepSeek R1 & Command R7B only).\n" + "only supported for non-streamed responses", + [](common_params & params, const std::string & value) { + /**/ if (value == "deepseek") { params.reasoning_format = COMMON_REASONING_FORMAT_DEEPSEEK; } + else if (value == "none") { params.reasoning_format = COMMON_REASONING_FORMAT_NONE; } + else { std::invalid_argument("invalid value"); } + } + ).set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_MAIN}).set_env("LLAMA_ARG_THINK")); add_opt(common_arg( {"--chat-template"}, "JINJA_TEMPLATE", string_format( @@ -2112,7 +2852,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex ).set_env("LLAMA_LOG_VERBOSITY")); add_opt(common_arg( {"--log-prefix"}, - "Enable prefx in log messages", + "Enable prefix in log messages", [](common_params &) { common_log_set_prefix(common_log_main(), true); } @@ -2293,7 +3033,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex {"-md", "--model-draft"}, "FNAME", "draft model for speculative decoding (default: unused)", [](common_params & params, const std::string & value) { - params.speculative.model = value; + params.speculative.model.path = value; } ).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_MODEL_DRAFT")); @@ -2301,7 +3041,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex {"-mv", "--model-vocoder"}, "FNAME", "vocoder model for audio generation (default: unused)", [](common_params & params, const std::string & value) { - params.vocoder.model = value; + params.vocoder.model.path = value; } ).set_examples({LLAMA_EXAMPLE_TTS, LLAMA_EXAMPLE_SERVER})); add_opt(common_arg( @@ -2311,16 +3051,23 @@ common_params_context common_params_parser_init(common_params & params, llama_ex params.vocoder.use_guide_tokens = true; } ).set_examples({LLAMA_EXAMPLE_TTS, LLAMA_EXAMPLE_SERVER})); + add_opt(common_arg( + {"--tts-speaker-file"}, "FNAME", + "speaker file path for audio generation", + [](common_params & params, const std::string & value) { + params.vocoder.speaker_file = value; + } + ).set_examples({LLAMA_EXAMPLE_TTS})); // model-specific add_opt(common_arg( {"--tts-oute-default"}, string_format("use default OuteTTS models (note: can download weights from the internet)"), [](common_params & params) { - params.hf_repo = "OuteAI/OuteTTS-0.2-500M-GGUF"; - params.hf_file = "OuteTTS-0.2-500M-Q8_0.gguf"; - params.vocoder.hf_repo = "ggml-org/WavTokenizer"; - params.vocoder.hf_file = "WavTokenizer-Large-75-F16.gguf"; + params.model.hf_repo = "OuteAI/OuteTTS-0.2-500M-GGUF"; + params.model.hf_file = "OuteTTS-0.2-500M-Q8_0.gguf"; + params.vocoder.model.hf_repo = "ggml-org/WavTokenizer"; + params.vocoder.model.hf_file = "WavTokenizer-Large-75-F16.gguf"; } ).set_examples({LLAMA_EXAMPLE_TTS})); @@ -2328,8 +3075,8 @@ common_params_context common_params_parser_init(common_params & params, llama_ex {"--embd-bge-small-en-default"}, string_format("use default bge-small-en-v1.5 model (note: can download weights from the internet)"), [](common_params & params) { - params.hf_repo = "ggml-org/bge-small-en-v1.5-Q8_0-GGUF"; - params.hf_file = "bge-small-en-v1.5-q8_0.gguf"; + params.model.hf_repo = "ggml-org/bge-small-en-v1.5-Q8_0-GGUF"; + params.model.hf_file = "bge-small-en-v1.5-q8_0.gguf"; params.pooling_type = LLAMA_POOLING_TYPE_NONE; params.embd_normalize = 2; params.n_ctx = 512; @@ -2342,8 +3089,8 @@ common_params_context common_params_parser_init(common_params & params, llama_ex {"--embd-e5-small-en-default"}, string_format("use default e5-small-v2 model (note: can download weights from the internet)"), [](common_params & params) { - params.hf_repo = "ggml-org/e5-small-v2-Q8_0-GGUF"; - params.hf_file = "e5-small-v2-q8_0.gguf"; + params.model.hf_repo = "ggml-org/e5-small-v2-Q8_0-GGUF"; + params.model.hf_file = "e5-small-v2-q8_0.gguf"; params.pooling_type = LLAMA_POOLING_TYPE_NONE; params.embd_normalize = 2; params.n_ctx = 512; @@ -2356,8 +3103,8 @@ common_params_context common_params_parser_init(common_params & params, llama_ex {"--embd-gte-small-default"}, string_format("use default gte-small model (note: can download weights from the internet)"), [](common_params & params) { - params.hf_repo = "ggml-org/gte-small-Q8_0-GGUF"; - params.hf_file = "gte-small-q8_0.gguf"; + params.model.hf_repo = "ggml-org/gte-small-Q8_0-GGUF"; + params.model.hf_file = "gte-small-q8_0.gguf"; params.pooling_type = LLAMA_POOLING_TYPE_NONE; params.embd_normalize = 2; params.n_ctx = 512; @@ -2366,5 +3113,91 @@ common_params_context common_params_parser_init(common_params & params, llama_ex } ).set_examples({LLAMA_EXAMPLE_EMBEDDING, LLAMA_EXAMPLE_SERVER})); + add_opt(common_arg( + {"--fim-qwen-1.5b-default"}, + string_format("use default Qwen 2.5 Coder 1.5B (note: can download weights from the internet)"), + [](common_params & params) { + params.model.hf_repo = "ggml-org/Qwen2.5-Coder-1.5B-Q8_0-GGUF"; + params.model.hf_file = "qwen2.5-coder-1.5b-q8_0.gguf"; + params.port = 8012; + params.n_gpu_layers = 99; + params.flash_attn = true; + params.n_ubatch = 1024; + params.n_batch = 1024; + params.n_ctx = 0; + params.n_cache_reuse = 256; + } + ).set_examples({LLAMA_EXAMPLE_SERVER})); + + add_opt(common_arg( + {"--fim-qwen-3b-default"}, + string_format("use default Qwen 2.5 Coder 3B (note: can download weights from the internet)"), + [](common_params & params) { + params.model.hf_repo = "ggml-org/Qwen2.5-Coder-3B-Q8_0-GGUF"; + params.model.hf_file = "qwen2.5-coder-3b-q8_0.gguf"; + params.port = 8012; + params.n_gpu_layers = 99; + params.flash_attn = true; + params.n_ubatch = 1024; + params.n_batch = 1024; + params.n_ctx = 0; + params.n_cache_reuse = 256; + } + ).set_examples({LLAMA_EXAMPLE_SERVER})); + + add_opt(common_arg( + {"--fim-qwen-7b-default"}, + string_format("use default Qwen 2.5 Coder 7B (note: can download weights from the internet)"), + [](common_params & params) { + params.model.hf_repo = "ggml-org/Qwen2.5-Coder-7B-Q8_0-GGUF"; + params.model.hf_file = "qwen2.5-coder-7b-q8_0.gguf"; + params.port = 8012; + params.n_gpu_layers = 99; + params.flash_attn = true; + params.n_ubatch = 1024; + params.n_batch = 1024; + params.n_ctx = 0; + params.n_cache_reuse = 256; + } + ).set_examples({LLAMA_EXAMPLE_SERVER})); + + add_opt(common_arg( + {"--fim-qwen-7b-spec"}, + string_format("use Qwen 2.5 Coder 7B + 0.5B draft for speculative decoding (note: can download weights from the internet)"), + [](common_params & params) { + params.model.hf_repo = "ggml-org/Qwen2.5-Coder-7B-Q8_0-GGUF"; + params.model.hf_file = "qwen2.5-coder-7b-q8_0.gguf"; + params.speculative.model.hf_repo = "ggml-org/Qwen2.5-Coder-0.5B-Q8_0-GGUF"; + params.speculative.model.hf_file = "qwen2.5-coder-0.5b-q8_0.gguf"; + params.speculative.n_gpu_layers = 99; + params.port = 8012; + params.n_gpu_layers = 99; + params.flash_attn = true; + params.n_ubatch = 1024; + params.n_batch = 1024; + params.n_ctx = 0; + params.n_cache_reuse = 256; + } + ).set_examples({LLAMA_EXAMPLE_SERVER})); + + add_opt(common_arg( + {"--fim-qwen-14b-spec"}, + string_format("use Qwen 2.5 Coder 14B + 0.5B draft for speculative decoding (note: can download weights from the internet)"), + [](common_params & params) { + params.model.hf_repo = "ggml-org/Qwen2.5-Coder-14B-Q8_0-GGUF"; + params.model.hf_file = "qwen2.5-coder-14b-q8_0.gguf"; + params.speculative.model.hf_repo = "ggml-org/Qwen2.5-Coder-0.5B-Q8_0-GGUF"; + params.speculative.model.hf_file = "qwen2.5-coder-0.5b-q8_0.gguf"; + params.speculative.n_gpu_layers = 99; + params.port = 8012; + params.n_gpu_layers = 99; + params.flash_attn = true; + params.n_ubatch = 1024; + params.n_batch = 1024; + params.n_ctx = 0; + params.n_cache_reuse = 256; + } + ).set_examples({LLAMA_EXAMPLE_SERVER})); + return ctx_arg; } diff --git a/common/chat.cpp b/common/chat.cpp index ef1c6fb3d..bbc5f087c 100644 --- a/common/chat.cpp +++ b/common/chat.cpp @@ -1,8 +1,436 @@ -#include "chat.hpp" -#include "chat-template.hpp" +#include "chat.h" #include "json-schema-to-grammar.h" #include "log.h" -#include "minja.hpp" +#include "minja/chat-template.hpp" +#include "minja/minja.hpp" + +#include + +typedef minja::chat_template common_chat_template; + +struct common_chat_templates { + bool has_explicit_template; // Model had builtin template or template overridde was specified. + std::unique_ptr template_default; // always set (defaults to chatml) + std::unique_ptr template_tool_use; +}; + +struct templates_params { + json messages; + json tools; + common_chat_tool_choice tool_choice; + json json_schema; + bool parallel_tool_calls; + bool stream; + std::string grammar; + bool add_generation_prompt = true; + bool extract_reasoning = true; +}; + +common_chat_tool_choice common_chat_tool_choice_parse_oaicompat(const std::string & tool_choice) { + if (tool_choice == "auto") { + return COMMON_CHAT_TOOL_CHOICE_AUTO; + } + if (tool_choice == "none") { + return COMMON_CHAT_TOOL_CHOICE_NONE; + } + if (tool_choice == "required") { + return COMMON_CHAT_TOOL_CHOICE_REQUIRED; + } + throw std::runtime_error("Invalid tool_choice: " + tool_choice); +} + +template <> +std::vector common_chat_msgs_parse_oaicompat(const json & messages) { + std::vector msgs; + + try { + + if (!messages.is_array()) { + throw std::runtime_error("Expected 'messages' to be an array, got " + messages.dump()); + } + + for (const auto & message : messages) { + if (!message.is_object()) { + throw std::runtime_error("Expected 'message' to be an object, got " + message.dump()); + } + + common_chat_msg msg; + if (!message.contains("role")) { + throw std::runtime_error("Missing 'role' in message: " + message.dump()); + } + msg.role = message.at("role"); + + auto has_content = message.contains("content"); + auto has_tool_calls = message.contains("tool_calls"); + if (has_content) { + const auto & content = message.at("content"); + if (content.is_string()) { + msg.content = content; + } else if (content.is_array()) { + for (const auto & part : content) { + if (!part.contains("type")) { + throw std::runtime_error("Missing content part type: " + part.dump()); + } + const auto & type = part.at("type"); + if (type != "text") { + throw std::runtime_error("Unsupported content part type: " + type.dump()); + } + common_chat_msg_content_part msg_part; + msg_part.type = type; + msg_part.text = part.at("text"); + msg.content_parts.push_back(msg_part); + } + } else if (!content.is_null()) { + throw std::runtime_error("Invalid 'content' type: expected string or array, got " + content.dump() + " (ref: https://github.com/ggml-org/llama.cpp/issues/8367)"); + } + } + if (has_tool_calls) { + for (const auto & tool_call : message.at("tool_calls")) { + common_chat_tool_call tc; + if (!tool_call.contains("type")) { + throw std::runtime_error("Missing tool call type: " + tool_call.dump()); + } + const auto & type = tool_call.at("type"); + if (type != "function") { + throw std::runtime_error("Unsupported tool call type: " + tool_call.dump()); + } + if (!tool_call.contains("function")) { + throw std::runtime_error("Missing tool call function: " + tool_call.dump()); + } + const auto & fc = tool_call.at("function"); + if (!fc.contains("name")) { + throw std::runtime_error("Missing tool call name: " + tool_call.dump()); + } + tc.name = fc.at("name"); + tc.arguments = fc.at("arguments"); + if (tool_call.contains("id")) { + tc.id = tool_call.at("id"); + } + msg.tool_calls.push_back(tc); + } + } + if (!has_content && !has_tool_calls) { + throw std::runtime_error("Expected 'content' or 'tool_calls' (ref: https://github.com/ggml-org/llama.cpp/issues/8367 & https://github.com/ggml-org/llama.cpp/issues/12279)"); + } + if (message.contains("reasoning_content")) { + msg.reasoning_content = message.at("reasoning_content"); + } + if (message.contains("name")) { + msg.tool_name = message.at("name"); + } + if (message.contains("tool_call_id")) { + msg.tool_call_id = message.at("tool_call_id"); + } + + msgs.push_back(msg); + } + } catch (const std::exception & e) { + throw std::runtime_error("Failed to parse messages: " + std::string(e.what()) + "; messages = " + messages.dump(2)); + } + + return msgs; +} + +template <> +json common_chat_msgs_to_json_oaicompat(const std::vector & msgs, bool concat_typed_text) { + json messages = json::array(); + for (const auto & msg : msgs) { + if (!msg.content.empty() && !msg.content_parts.empty()) { + throw std::runtime_error("Cannot specify both content and content_parts"); + } + json jmsg { + {"role", msg.role}, + }; + if (!msg.content.empty()) { + jmsg["content"] = msg.content; + } else if (!msg.content_parts.empty()) { + if (concat_typed_text) { + std::string text; + for (const auto & part : msg.content_parts) { + if (part.type != "text") { + LOG_WRN("Ignoring content part type: %s\n", part.type.c_str()); + continue; + } + if (!text.empty()) { + text += '\n'; + } + text += part.text; + } + jmsg["content"] = text; + } else { + auto & parts = jmsg["content"] = json::array(); + for (const auto & part : msg.content_parts) { + parts.push_back({ + {"type", part.type}, + {"text", part.text}, + }); + } + } + } else { + jmsg["content"] = json(); // null + } + if (!msg.reasoning_content.empty()) { + jmsg["reasoning_content"] = msg.reasoning_content; + } + if (!msg.tool_name.empty()) { + jmsg["name"] = msg.tool_name; + } + if (!msg.tool_call_id.empty()) { + jmsg["tool_call_id"] = msg.tool_call_id; + } + if (!msg.tool_calls.empty()) { + auto & tool_calls = jmsg["tool_calls"] = json::array(); + for (const auto & tool_call : msg.tool_calls) { + json tc { + {"type", "function"}, + {"function", { + {"name", tool_call.name}, + {"arguments", tool_call.arguments}, + }}, + }; + if (!tool_call.id.empty()) { + tc["id"] = tool_call.id; + } + tool_calls.push_back(tc); + } + } + messages.push_back(jmsg); + } + return messages; +} + +template <> +std::vector common_chat_msgs_parse_oaicompat(const std::string & messages) { + return common_chat_msgs_parse_oaicompat(json::parse(messages)); +} + +template <> +std::vector common_chat_tools_parse_oaicompat(const json & tools) { + std::vector result; + + try { + if (!tools.is_null()) { + if (!tools.is_array()) { + throw std::runtime_error("Expected 'tools' to be an array, got " + tools.dump()); + } + for (const auto & tool : tools) { + if (!tool.contains("type")) { + throw std::runtime_error("Missing tool type: " + tool.dump()); + } + const auto & type = tool.at("type"); + if (!type.is_string() || type != "function") { + throw std::runtime_error("Unsupported tool type: " + tool.dump()); + } + if (!tool.contains("function")) { + throw std::runtime_error("Missing tool function: " + tool.dump()); + } + + const auto & function = tool.at("function"); + result.push_back({ + /* .name = */ function.at("name"), + /* .description = */ function.at("description"), + /* .parameters = */ function.at("parameters").dump(), + }); + } + } + } catch (const std::exception & e) { + throw std::runtime_error("Failed to parse tools: " + std::string(e.what()) + "; tools = " + tools.dump(2)); + } + + return result; +} + +template <> +std::vector common_chat_tools_parse_oaicompat(const std::string & tools) { + return common_chat_tools_parse_oaicompat(json::parse(tools)); +} + +template <> +json common_chat_tools_to_json_oaicompat(const std::vector & tools) { + if (tools.empty()) { + return json(); + } + + auto result = json::array(); + for (const auto & tool : tools) { + result.push_back({ + {"type", "function"}, + {"function", { + {"name", tool.name}, + {"description", tool.description}, + {"parameters", json::parse(tool.parameters)}, + }}, + }); + } + return result; +} + +bool common_chat_verify_template(const std::string & tmpl, bool use_jinja) { + if (use_jinja) { + try { + common_chat_msg msg; + msg.role = "user"; + msg.content = "test"; + + auto tmpls = common_chat_templates_init(/* model= */ nullptr, tmpl); + + common_chat_templates_inputs inputs; + inputs.messages = {msg}; + + common_chat_templates_apply(tmpls.get(), inputs); + return true; + } catch (const std::exception & e) { + LOG_ERR("%s: failed to apply template: %s\n", __func__, e.what()); + return false; + } + } + llama_chat_message chat[] = {{"user", "test"}}; + const int res = llama_chat_apply_template(tmpl.c_str(), chat, 1, true, nullptr, 0); + return res >= 0; +} + +std::string common_chat_format_single( + const struct common_chat_templates * tmpls, + const std::vector & past_msg, + const common_chat_msg & new_msg, + bool add_ass, + bool use_jinja) { + + common_chat_templates_inputs inputs; + inputs.use_jinja = use_jinja; + + std::string fmt_past_msg; + if (!past_msg.empty()) { + inputs.messages = past_msg; + inputs.add_generation_prompt = false; + fmt_past_msg = common_chat_templates_apply(tmpls, inputs).prompt; + } + std::ostringstream ss; + // if the past_msg ends with a newline, we must preserve it in the formatted version + if (add_ass && !fmt_past_msg.empty() && fmt_past_msg.back() == '\n') { + ss << "\n"; + }; + // format chat with new_msg + inputs.messages.push_back(new_msg); + inputs.add_generation_prompt = add_ass; + auto fmt_new_msg = common_chat_templates_apply(tmpls, inputs).prompt; + // get the diff part + ss << fmt_new_msg.substr(fmt_past_msg.size(), fmt_new_msg.size() - fmt_past_msg.size()); + return ss.str(); +} + +std::string common_chat_format_example(const struct common_chat_templates * tmpls, bool use_jinja) { + common_chat_templates_inputs inputs; + inputs.use_jinja = use_jinja; + auto add_simple_msg = [&](auto role, auto content) { + common_chat_msg msg; + msg.role = role; + msg.content = content; + inputs.messages.push_back(msg); + }; + add_simple_msg("system", "You are a helpful assistant"); + add_simple_msg("user", "Hello"); + add_simple_msg("assistant", "Hi there"); + add_simple_msg("user", "How are you?"); + return common_chat_templates_apply(tmpls, inputs).prompt; +} + +#define CHATML_TEMPLATE_SRC \ + "{%- for message in messages -%}\n" \ + " {{- '<|im_start|>' + message.role + '\n' + message.content + '<|im_end|>\n' -}}\n" \ + "{%- endfor -%}\n" \ + "{%- if add_generation_prompt -%}\n" \ + " {{- '<|im_start|>assistant\n' -}}\n" \ + "{%- endif -%}" + +void common_chat_templates_free(struct common_chat_templates * tmpls) { + delete tmpls; +} + +bool common_chat_templates_was_explicit(const struct common_chat_templates * tmpls) { + return tmpls->has_explicit_template; +} + +const char * common_chat_templates_source(const struct common_chat_templates * tmpls, const char * variant) { + if (variant != nullptr) { + if (strcmp(variant, "tool_use") == 0) { + if (tmpls->template_tool_use) { + return tmpls->template_tool_use->source().c_str(); + } + return nullptr; + } else { + LOG_DBG("%s: unknown template variant: %s\n", __func__, variant); + } + } + return tmpls->template_default->source().c_str(); +} + +common_chat_templates_ptr common_chat_templates_init( + const struct llama_model * model, + const std::string & chat_template_override, + const std::string & bos_token_override, + const std::string & eos_token_override) +{ + std::string default_template_src; + std::string template_tool_use_src; + + bool has_explicit_template = !chat_template_override.empty(); + if (chat_template_override.empty()) { + GGML_ASSERT(model != nullptr); + const auto * str = llama_model_chat_template(model, /* name */ nullptr); + if (str) { + default_template_src = str; + has_explicit_template = true; + } + str = llama_model_chat_template(model, /* name */ "tool_use"); + if (str) { + template_tool_use_src = str; + has_explicit_template = true; + } + } else { + default_template_src = chat_template_override; + } + if (default_template_src.empty() || default_template_src == "chatml") { + if (!template_tool_use_src.empty()) { + default_template_src = template_tool_use_src; + } else { + default_template_src = CHATML_TEMPLATE_SRC; + } + } + std::string token_bos = bos_token_override; + std::string token_eos = eos_token_override; + if (model) { + const auto * vocab = llama_model_get_vocab(model); + const auto get_token = [&](llama_token token, const char * name, const char * jinja_variable_name) { + if (token == LLAMA_TOKEN_NULL) { + if (default_template_src.find(jinja_variable_name) != std::string::npos + || template_tool_use_src.find(jinja_variable_name) != std::string::npos) { + LOG_WRN("common_chat_templates_init: warning: vocab does not have a %s token, jinja template won't work as intended.\n", name); + } + return std::string(); + } + return common_token_to_piece(vocab, token, true); + }; + token_bos = get_token(llama_vocab_bos(vocab), "BOS", "bos_token"); + token_eos = get_token(llama_vocab_eos(vocab), "EOS", "eos_token"); + } + common_chat_templates_ptr tmpls(new common_chat_templates()); + tmpls->has_explicit_template = has_explicit_template; + try { + tmpls->template_default = std::make_unique(default_template_src, token_bos, token_eos); + } catch (const std::exception & e) { + LOG_ERR("%s: failed to parse chat template (defaulting to chatml): %s \n", __func__, e.what()); + tmpls->template_default = std::make_unique(CHATML_TEMPLATE_SRC, token_bos, token_eos); + } + if (!template_tool_use_src.empty()) { + try { + tmpls->template_tool_use = std::make_unique(template_tool_use_src, token_bos, token_eos); + } catch (const std::exception & e) { + LOG_ERR("%s: failed to parse tool use chat template (ignoring it): %s\n", __func__, e.what()); + } + } + return tmpls; +} std::string common_chat_format_name(common_chat_format format) { switch (format) { @@ -12,22 +440,19 @@ std::string common_chat_format_name(common_chat_format format) { case COMMON_CHAT_FORMAT_LLAMA_3_X: return "Llama 3.x"; case COMMON_CHAT_FORMAT_LLAMA_3_X_WITH_BUILTIN_TOOLS: return "Llama 3.x with builtin tools"; case COMMON_CHAT_FORMAT_DEEPSEEK_R1: return "DeepSeek R1"; + case COMMON_CHAT_FORMAT_DEEPSEEK_R1_EXTRACT_REASONING: return "DeepSeek R1 (extract reasoning)"; case COMMON_CHAT_FORMAT_FIREFUNCTION_V2: return "FireFunction v2"; case COMMON_CHAT_FORMAT_FUNCTIONARY_V3_2: return "Functionary v3.2"; case COMMON_CHAT_FORMAT_FUNCTIONARY_V3_1_LLAMA_3_1: return "Functionary v3.1 Llama 3.1"; case COMMON_CHAT_FORMAT_HERMES_2_PRO: return "Hermes 2 Pro"; + case COMMON_CHAT_FORMAT_HERMES_2_PRO_EXTRACT_REASONING: return "Hermes 2 Pro (extract reasoning)"; case COMMON_CHAT_FORMAT_COMMAND_R7B: return "Command R7B"; + case COMMON_CHAT_FORMAT_COMMAND_R7B_EXTRACT_REASONING: return "Command R7B (extract reasoning)"; default: throw std::runtime_error("Unknown chat format"); } } -const common_grammar_options grammar_options { - /* .dotall = */ false, - /* .compact_spaces = */ false, - // /* .compact_spaces = */ true, -}; - static bool parse_json(std::string::const_iterator & it, const std::string::const_iterator & end, json & out) { // // https://json.nlohmann.me/features/parsing/sax_interface/ struct json_error_locator : public nlohmann::json_sax { @@ -36,22 +461,22 @@ static bool parse_json(std::string::const_iterator & it, const std::string::cons json_error_locator() : position(0), found_error(false) {} - bool parse_error(std::size_t position, const std::string &, const json::exception &) override { + bool parse_error(std::size_t position, const std::string &, const json::exception &) override { // NOLINT this->position = position - 1; this->found_error = true; return false; } - bool null() override { return true; } - bool boolean(bool) override { return true; } - bool number_integer(number_integer_t) override { return true; } - bool number_unsigned(number_unsigned_t) override { return true; } - bool number_float(number_float_t, const string_t &) override { return true; } - bool string(string_t &) override { return true; } - bool binary(binary_t &) override { return true; } - bool start_object(std::size_t) override { return true; } - bool key(string_t &) override { return true; } + bool null() override { return true; } // NOLINT + bool boolean(bool) override { return true; } // NOLINT + bool number_integer(number_integer_t) override { return true; } // NOLINT + bool number_unsigned(number_unsigned_t) override { return true; } // NOLINT + bool number_float(number_float_t, const string_t &) override { return true; } // NOLINT + bool string(string_t &) override { return true; } // NOLINT + bool binary(binary_t &) override { return true; } // NOLINT + bool start_object(std::size_t) override { return true; } // NOLINT + bool key(string_t &) override { return true; } // NOLINT bool end_object() override { return true; } - bool start_array(std::size_t) override { return true; } + bool start_array(std::size_t) override { return true; } // NOLINT bool end_array() override { return true; } }; json_error_locator err_loc; @@ -73,6 +498,34 @@ static bool parse_json(std::string::const_iterator & it, const std::string::cons } } +static bool parse_literal(std::string::const_iterator & it, const std::string::const_iterator & end, const std::string & expected) { + auto expected_it = expected.begin(); + auto tmp_it = it; + while (tmp_it != end && expected_it != expected.end() && *tmp_it == *expected_it) { + ++tmp_it; + ++expected_it; + } + if (expected_it == expected.end()) { + it = tmp_it; + return true; + } + return false; +} + +static std::optional parse_pattern(std::string::const_iterator & it, const std::string::const_iterator & end, const std::regex & expected) { + std::smatch match; + if (std::regex_match(it, end, match, expected)) { + it = match.suffix().first; + return match; + } + return std::nullopt; +} + +static void consume_spaces(std::string::const_iterator & it, const std::string::const_iterator & end) { + while (it != end && std::isspace(*it)) { + ++it; + } +} /** * Takes a prefix regex that must have 1 group to capture the function name, a closing suffix, and expects json parameters in between. @@ -82,7 +535,8 @@ static common_chat_msg parse_json_tool_calls( const std::string& input, const std::optional & trigger_opt, const std::regex & function_regex, - const std::regex & close_regex) { + const std::regex & close_regex, + bool allow_raw_python = false) { std::smatch match; common_chat_msg result; @@ -105,7 +559,6 @@ static common_chat_msg parse_json_tool_calls( std::sregex_iterator rend; std::sregex_iterator rit(it, end, function_regex); if (rit == rend) { - fprintf(stderr, "No more tool calls found\n"); result.content += std::string(it, end); break; } @@ -114,48 +567,60 @@ static common_chat_msg parse_json_tool_calls( it = rit->suffix().first; json arguments; - if (!parse_json(it, end, arguments)) { - throw std::runtime_error("Failed to parse json tool call arguments"); + if (parse_json(it, end, arguments)) { + if (!std::regex_search(it, end, match, close_regex)) { + throw std::runtime_error("Malformed input, missing closing pattern: " + input); + } + it = match.suffix().first; + result.tool_calls.push_back({name, arguments.is_string() ? arguments.get() : arguments.dump(), /* id= */ ""}); + } else { + if (allow_raw_python && name == "python") { + result.tool_calls.push_back({name, json({{"code", std::string(it, end)}}).dump(), /* id= */ ""}); + break; + } + throw std::runtime_error("Failed to parse json tool call arguments: " + input); } - if (!std::regex_search(it, end, match, close_regex)) { - throw std::runtime_error("Malformed input, missing closing pattern"); + } + + if (!result.tool_calls.empty()) { + if (!string_strip(result.content).empty()) { + LOG_WRN("Content found with tool calls: %s\n", result.content.c_str()); } - it = match.suffix().first; - result.tool_calls.push_back({name, arguments.is_string() ? arguments.get() : arguments.dump(), /* id= */ ""}); + result.content = ""; } return result; } +static common_chat_tool_call process_tool_call(const json & tool_call) { + const auto & arguments = tool_call.at("arguments"); + return { + /* .name = */ tool_call.at("name"), + /* .arguments = */ arguments.is_string() ? arguments.get() : arguments.dump(), + /* .id = */ tool_call.contains("id") ? tool_call.at("id") : "", + }; +} static common_chat_msg parse_prefixed_json_tool_call_array(const std::string& input, const std::string & prefix, size_t rstrip_prefix = 0) { auto content_end = input.find(prefix); size_t tc_start = std::string::npos; common_chat_msg result; result.role = "assistant"; - const auto process_tool_calls = [&](const json & tool_calls) { - for (const auto & tool_call : tool_calls) { - const auto & arguments = tool_call["arguments"]; - result.tool_calls.push_back({ - tool_call["name"], - arguments.is_string() ? arguments.get() : arguments.dump(), - tool_call.contains("id") ? tool_call["id"] : "", - }); - } - }; if (content_end == std::string::npos) { result.content = input; } else { tc_start = content_end + prefix.size() - rstrip_prefix; result.content = input.substr(0, content_end); auto tool_calls = json::parse(input.substr(tc_start)); - process_tool_calls(tool_calls); + for (const auto & tool_call : tool_calls) { + result.tool_calls.emplace_back(process_tool_call(tool_call)); + } } return result; } static void foreach_function(const json & tools, const std::function & fn) { for (const auto & tool : tools) { - if (!tool.contains("type") || tool["type"] != "function" || !tool.contains("function")) { + if (!tool.contains("type") || tool.at("type") != "function" || !tool.contains("function")) { LOG_INF("Skipping tool without function: %s", tool.dump(2).c_str()); continue; } @@ -179,38 +644,45 @@ static std::string apply( // tmpl_inputs.now = std::chrono::system_clock::now(); minja::chat_template_options tmpl_opts; - tmpl_opts.use_bos_token = false; - tmpl_opts.use_eos_token = false; - - return tmpl.apply(tmpl_inputs, tmpl_opts); + // To avoid double BOS / EOS tokens, we're manually removing begining / trailing tokens + // instead of using `chat_template_options.use_bos_token = false`, since these tokens + // may be needed inside the template / between messages too. + auto result = tmpl.apply(tmpl_inputs, tmpl_opts); + if (string_starts_with(result, tmpl.bos_token())) { + result = result.substr(tmpl.bos_token().size()); + } + if (string_ends_with(result, tmpl.eos_token())) { + result = result.substr(0, result.size() - tmpl.eos_token().size()); + } + return result; } -static common_chat_params common_chat_params_init_generic(const common_chat_template & tmpl, const struct common_chat_inputs & inputs) { +static common_chat_params common_chat_params_init_generic(const common_chat_template & tmpl, const struct templates_params & inputs) { common_chat_params data; auto tool_call_schemas = json::array(); foreach_function(inputs.tools, [&](const json & tool) { - const auto & function = tool["function"]; + const auto & function = tool.at("function"); auto tool_schema = json { {"type", "object"}, {"properties", { {"name", { {"type", "string"}, - {"const", function["name"]}, + {"const", function.at("name")}, }}, - {"arguments", function["parameters"]}, + {"arguments", function.at("parameters")}, }}, {"required", json::array({"name", "arguments"})}, }; if (function.contains("description")) { - tool_schema["description"] = function["description"]; + tool_schema["description"] = function.at("description"); } if (inputs.parallel_tool_calls) { - tool_schema["properties"]["id"] = { + tool_schema.at("properties")["id"] = { {"type", "string"}, {"minLength", 4}, }; - tool_schema["required"].push_back("id"); + tool_schema.at("required").push_back("id"); } tool_call_schemas.emplace_back(tool_schema); }); @@ -239,7 +711,7 @@ static common_chat_params common_chat_params_init_generic(const common_chat_temp {"required", json::array({"tool_call"})}, }; const auto schema = - inputs.tool_choice != "required" + inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_REQUIRED ? json { {"anyOf", json::array({ tool_call, @@ -260,7 +732,7 @@ static common_chat_params common_chat_params_init_generic(const common_chat_temp data.grammar_lazy = false; data.grammar = build_grammar([&](const common_grammar_builder & builder) { builder.add_schema("root", schema); - }, grammar_options); + }); auto tweaked_messages = common_chat_template::add_system( inputs.messages, @@ -275,33 +747,33 @@ static common_chat_msg common_chat_parse_generic(const std::string & input) { common_chat_msg result; result.role = "assistant"; if (data.contains("tool_calls")) { - for (const auto & tool_call : data["tool_calls"]) { + for (const auto & tool_call : data.at("tool_calls")) { result.tool_calls.push_back({ - tool_call["name"], - tool_call["arguments"].dump(), - tool_call.contains("id") ? tool_call["id"] : "", + tool_call.at("name"), + tool_call.at("arguments").dump(), + tool_call.contains("id") ? tool_call.at("id") : "", }); } } else if (data.contains("tool_call")) { result.tool_calls.push_back({ - data["tool_call"]["name"], - data["tool_call"]["arguments"].dump(), + data.at("tool_call").at("name"), + data.at("tool_call").at("arguments").dump(), /* id= */ "", }); } else if (data.contains("response")) { - const auto & response = data["response"]; + const auto & response = data.at("response"); result.content = response.is_string() ? response.get() : response.dump(2); } return result; } -static common_chat_params common_chat_params_init_mistral_nemo(const common_chat_template & tmpl, const struct common_chat_inputs & inputs) { +static common_chat_params common_chat_params_init_mistral_nemo(const common_chat_template & tmpl, const struct templates_params & inputs) { common_chat_params data; - data.grammar_lazy = inputs.tool_choice != "required"; + data.grammar_lazy = inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_REQUIRED; data.grammar = build_grammar([&](const common_grammar_builder & builder) { auto schemas = json::array(); foreach_function(inputs.tools, [&](const json & tool) { - const auto & function = tool["function"]; + const auto & function = tool.at("function"); schemas.push_back({ {"type", "object"}, {"properties", { @@ -309,9 +781,9 @@ static common_chat_params common_chat_params_init_mistral_nemo(const common_chat // It's hard to constrain that for now (while reusing the JSON schema conversion), so we're just expecting a plain object. {"name", { {"type", "string"}, - {"const", function["name"]}, + {"const", function.at("name")}, }}, - {"arguments", function["parameters"]}, + {"arguments", function.at("parameters")}, {"id", { {"type", "string"}, // Nemo's template expects a 9-character alphanumeric ID. @@ -330,8 +802,11 @@ static common_chat_params common_chat_params_init_mistral_nemo(const common_chat schema["maxItems"] = 1; } builder.add_rule("root", "\"[TOOL_CALLS]\" " + builder.add_schema("tool_calls", schema)); - }, grammar_options); - data.grammar_triggers.push_back({"[TOOL_CALLS]", /* .at_start = */ true}); + }); + data.grammar_triggers.push_back({COMMON_GRAMMAR_TRIGGER_TYPE_WORD, "[TOOL_CALLS]"}); + data.preserved_tokens = { + "[TOOL_CALLS]", + }; data.prompt = apply(tmpl, inputs.messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt); data.format = COMMON_CHAT_FORMAT_MISTRAL_NEMO; return data; @@ -340,13 +815,13 @@ static common_chat_msg common_chat_parse_mistral_nemo(const std::string & input) return parse_prefixed_json_tool_call_array(input, "[TOOL_CALLS]"); } -static common_chat_params common_chat_params_init_command_r7b(const common_chat_template & tmpl, const struct common_chat_inputs & inputs) { +static common_chat_params common_chat_params_init_command_r7b(const common_chat_template & tmpl, const struct templates_params & inputs) { common_chat_params data; - data.grammar_lazy = inputs.tool_choice != "required"; + data.grammar_lazy = inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_REQUIRED; data.grammar = build_grammar([&](const common_grammar_builder & builder) { auto schemas = json::array(); foreach_function(inputs.tools, [&](const json & tool) { - const auto & function = tool["function"]; + const auto & function = tool.at("function"); schemas.push_back({ {"type", "object"}, {"properties", { @@ -357,9 +832,9 @@ static common_chat_params common_chat_params_init_command_r7b(const common_chat_ }}, {"tool_name", { {"type", "string"}, - {"const", function["name"]}, + {"const", function.at("name")}, }}, - {"parameters", function["parameters"]}, + {"parameters", function.at("parameters")}, }}, {"required", json::array({"tool_call_id", "tool_name", "parameters"})}, }); @@ -373,58 +848,88 @@ static common_chat_params common_chat_params_init_command_r7b(const common_chat_ schema["maxItems"] = 1; } builder.add_rule("root", "\"<|START_ACTION|>\" " + builder.add_schema("tool_calls", schema) + " \"<|END_ACTION|>\""); - }, grammar_options); - data.grammar_triggers.push_back({"<|START_ACTION|>", /* .at_start = */ false}); + }); + data.grammar_triggers.push_back({ + COMMON_GRAMMAR_TRIGGER_TYPE_WORD, + "<|START_ACTION|>", + }); data.preserved_tokens = { + "<|START_ACTION|>", + "<|END_ACTION|>", "<|START_RESPONSE|>", "<|END_RESPONSE|>", "<|START_THINKING|>", "<|END_THINKING|>", - "<|END_ACTION|>", }; - data.prompt = apply(tmpl, inputs.messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt); - data.format = COMMON_CHAT_FORMAT_COMMAND_R7B; + auto adjusted_messages = json::array(); + for (const auto & msg : inputs.messages) { + auto has_reasoning_content = msg.contains("reasoning_content") && msg.at("reasoning_content").is_string(); + auto has_tool_calls = msg.contains("tool_calls") && msg.at("tool_calls").is_array(); + if (has_reasoning_content && has_tool_calls) { + auto adjusted_message = msg; + adjusted_message["tool_plan"] = msg.at("reasoning_content"); + adjusted_message.erase("reasoning_content"); + adjusted_messages.push_back(adjusted_message); + } else { + adjusted_messages.push_back(msg); + } + } + data.prompt = apply(tmpl, adjusted_messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt, {}); + data.format = inputs.extract_reasoning ? COMMON_CHAT_FORMAT_COMMAND_R7B_EXTRACT_REASONING : COMMON_CHAT_FORMAT_COMMAND_R7B; return data; } -static common_chat_msg common_chat_parse_command_r7b(const std::string & input) { - static std::regex response_regex("<\\|START_RESPONSE\\|>([\\s\\S\\n\\r]*?)<\\|END_RESPONSE\\|>"); - static std::regex thought_action_regex("<\\|START_THINKING\\|>([\\s\\S\\n\\r]*?)<\\|END_THINKING\\|><\\|START_ACTION\\|>([\\s\\S\\n\\r]*?)<\\|END_ACTION\\|>"); +static common_chat_msg common_chat_parse_command_r7b(const std::string & input, bool extract_reasoning) { + static const std::regex thought_regex("(<\\|START_THINKING\\|>([\\s\\S]*?)<\\|END_THINKING\\|>)([\\s\\S]*)"); + static const std::regex action_regex("<\\|START_ACTION\\|>([\\s\\S]*?)<\\|END_ACTION\\|>"); + static const std::regex response_regex("(?:<\\|START_RESPONSE\\|>)?([\\s\\S]*?)<\\|END_RESPONSE\\|>"); + std::smatch match; common_chat_msg result; result.role = "assistant"; - if (std::regex_match(input, match, response_regex)) { - result.content = match[1].str(); - } else if (std::regex_match(input, match, thought_action_regex)) { - result.tool_plan = match[1].str(); - auto actions_str = match[2].str(); + + std::string rest = input; + + if (std::regex_match(rest, match, thought_regex)) { + if (extract_reasoning) { + result.reasoning_content = match[2].str(); + } else if (!match[2].str().empty()) { + // Let the unparsed thinking tags through in content only if their insides aren't empty. + result.content = match[1].str(); + } + rest = match[3].str(); + } + if (std::regex_match(rest, match, action_regex)) { + auto actions_str = match[1].str(); auto actions = json::parse(actions_str); for (const auto & action : actions) { result.tool_calls.push_back({ - /* .name = */ action["tool_name"], - /* .arguments = */ action["parameters"].dump(), - /* .id = */ action["tool_call_id"], + /* .name = */ action.at("tool_name"), + /* .arguments = */ action.at("parameters").dump(), + /* .id = */ action.at("tool_call_id"), }); } + } else if (std::regex_match(rest, match, response_regex)) { + auto response = match[1].str(); + result.content += response; } else { - LOG_ERR("Failed to parse command_r output"); - result.content = input; + result.content += rest; } return result; } static void expect_tool_parameters(const std::string & name, const json & parameters, const std::vector & expected_properties) { - if (!parameters.is_object() || !parameters.contains("type") || parameters["type"] != "object" || !parameters.contains("properties") || !parameters.contains("required")) { + if (!parameters.is_object() || !parameters.contains("type") || parameters.at("type") != "object" || !parameters.contains("properties") || !parameters.contains("required")) { throw std::runtime_error("Parameters of tool " + name + " must be an object w/ required properties"); } const auto & parameters_properties = parameters.at("properties"); const auto & parameters_required = parameters.at("required"); for (const auto & prop : expected_properties) { if (!parameters_properties.contains(prop)) { - throw std::runtime_error("Parameters of tool " + name + " is missing property: " + prop); + throw std::runtime_error("Parameters of tool " + name + " is missing property: " + prop); // NOLINT } if (std::find(parameters_required.begin(), parameters_required.end(), json(prop)) == parameters_required.end()) { - throw std::runtime_error("Parameters of tool " + name + " must have property marked as required: " + prop); + throw std::runtime_error("Parameters of tool " + name + " must have property marked as required: " + prop); // NOLINT } } if (parameters_properties.size() != expected_properties.size()) { @@ -432,18 +937,16 @@ static void expect_tool_parameters(const std::string & name, const json & parame } } -static common_chat_params common_chat_params_init_llama_3_1_tool_calls(const common_chat_template & tmpl, const struct common_chat_inputs & inputs, bool allow_python_tag_builtin_tools) { +static common_chat_params common_chat_params_init_llama_3_1_tool_calls(const common_chat_template & tmpl, const struct templates_params & inputs, bool allow_python_tag_builtin_tools) { auto builtin_tools = json::array(); common_chat_params data; - data.grammar_lazy = inputs.tool_choice != "required"; + data.grammar_lazy = inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_REQUIRED; data.grammar = build_grammar([&](const common_grammar_builder & builder) { std::vector tool_rules; auto handle_builtin_tool = [&](const std::string & name, const json & parameters) { - if (name == "wolfram_alpha") { + if (name == "wolfram_alpha" || name == "web_search" || name == "brave_search") { // https://github.com/meta-llama/llama-stack/blob/main/llama_stack/providers/remote/tool_runtime/wolfram_alpha/wolfram_alpha.py - expect_tool_parameters(name, parameters, {"query"}); - } else if (name == "web_search" || name == "brave_search") { // https://github.com/meta-llama/llama-stack/blob/main/llama_stack/providers/remote/tool_runtime/brave_search/brave_search.py expect_tool_parameters(name, parameters, {"query"}); } else if (name == "python" || name == "code_interpreter") { @@ -455,7 +958,7 @@ static common_chat_params common_chat_params_init_llama_3_1_tool_calls(const com std::vector kvs; for (const auto & [key, value] : parameters.at("properties").items()) { - kvs.push_back("\"" + key + "=\" " + builder.add_schema(name + "-args-" + key, value)); + kvs.push_back("\"" + key + "=\" " + builder.add_schema(name + "-args-" + key, value)); // NOLINT } tool_rules.push_back( @@ -468,9 +971,9 @@ static common_chat_params common_chat_params_init_llama_3_1_tool_calls(const com }; foreach_function(inputs.tools, [&](const json & tool) { - const auto & function = tool["function"]; - std::string name = function["name"]; - auto parameters = function["parameters"]; + const auto & function = tool.at("function"); + std::string name = function.at("name"); + auto parameters = function.at("parameters"); builder.resolve_refs(parameters); // https://github.com/meta-llama/llama-stack/tree/main/llama_stack/providers/remote/tool_runtime @@ -481,23 +984,23 @@ static common_chat_params common_chat_params_init_llama_3_1_tool_calls(const com builder.add_rule( name + "-call", "\"{\" space " - "( \"\\\"type\\\":\" space \"\\\"function\\\",\" space )? " - "\"\\\"name\\\": \\\"" + name + "\\\", \\\"parameters\\\": \" " + - builder.add_schema(name + "-args", parameters) + - " \"}\"")); - data.grammar_triggers.push_back({"{\"name\": \"" + name + "\"", /* .at_start = */ true}); + "( \"\\\"type\\\"\" space \":\" space \"\\\"function\\\"\" space \",\" space )? " + " \"\\\"name\\\"\" space \":\" space \"\\\"" + name + "\\\"\" space \",\" space " + " \"\\\"parameters\\\"\" space \":\" space " + builder.add_schema(name + "-args", parameters) + " " + "\"}\" space")); + }); + // Small models may hallucinate function names so we match anything (*at the start*) that looks like the JSON of a function call, regardless of the name. + data.grammar_triggers.push_back({ + COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN_START, + "\\{\\s*(?:\"type\"\\s*:\\s*\"function\"\\s*,\\s*)?\"name\"\\s*:\\s*\"", // + name + "\"[\\s\\S]*", }); - data.grammar_triggers.push_back({"{\"name\":", /* .at_start = */ true}); - data.grammar_triggers.push_back({"{\n \"name\":", /* .at_start = */ true}); - data.grammar_triggers.push_back({"{\n \"name\":", /* .at_start = */ true}); - data.grammar_triggers.push_back({"{\"type\": \"function\"", /* .at_start = */ true}); - data.grammar_triggers.push_back({"{\n \"type\": \"function\"", /* .at_start = */ true}); - data.grammar_triggers.push_back({"{\n \"type\": \"function\"", /* .at_start = */ true}); if (!builtin_tools.empty()) { - data.grammar_triggers.push_back({"<|python_tag|>", /* .at_start = */ false}); + data.grammar_triggers.push_back({COMMON_GRAMMAR_TRIGGER_TYPE_WORD, "<|python_tag|>"}); + data.preserved_tokens.push_back("<|python_tag|>"); } + // Allow a few empty lines on top of the usual constrained json schema space rule. builder.add_rule("root", string_join(tool_rules, " | ")); - }, grammar_options); + }); data.additional_stops.push_back("<|eom_id|>"); data.prompt = apply(tmpl, inputs.messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt, { {"tools_in_user_message", false}, @@ -510,93 +1013,158 @@ static common_chat_params common_chat_params_init_llama_3_1_tool_calls(const com } static common_chat_msg common_chat_parse_llama_3_1(const std::string & input, bool with_builtin_tools = false) { // TODO: tighten & simplify the parser, don't accept leading text context. - static std::regex function_regex("\\{[\\s\\n\\r]*(?:\"type\"[\\s\\n\\r]*:[\\s\\n\\r]*\"function\"[\\s\\n\\r]*,[\\s\\n\\r]*|[\\s\\n\\r]*)\"name\"[\\s\\n\\r]*:[\\s\\n\\r]*\"([^\"]+)\"[\\s\\n\\r]*,[\\s\\n\\r]*\"parameters\": "); - static std::regex close_regex("\\}"); - static std::regex builtin_call_regex("<\\|python_tag\\|>([^.(]+)\\.call\\((.*)\\)"); + static const std::regex function_regex( + "\\s*\\{\\s*(?:\"type\"\\s*:\\s*\"function\"\\s*,\\s*)?\"name\"\\s*:\\s*\"([^\"]+)\"\\s*,\\s*\"parameters\"\\s*: "); + static const std::regex close_regex("\\}\\s*"); + static const std::regex builtin_call_regex("<\\|python_tag\\|>\\s*([^.(]+)\\s*\\.\\s*call\\s*\\(\\s*([\\w]+)\\s*=\\s*([\\s\\S]*?)\\)"); if (with_builtin_tools) { std::smatch match; if (std::regex_match(input, match, builtin_call_regex)) { - auto name = match[1].str(); - auto raw_args = match[2].str(); + try { + auto name = match[1].str(); + auto arg_name = match[2].str(); + auto arg_value_str = match[3].str(); + auto arg_value = json::parse(arg_value_str); - // TODO: if/when builtin tools start accepting more than 1 argument, use parse_json for real parsing. - auto it_eq = raw_args.find('='); - auto arg_name = raw_args.substr(0, it_eq); - auto arg_value_str = raw_args.substr(it_eq + 1); - auto arg_value = json::parse(arg_value_str); - - return { - /* .role = */ "assistant", - /* .content = */ match.prefix().str(), - /* .tool_calls = */ { - { - /* .name = */ match[1], - /* .arguments = */ (json { - {arg_name, arg_value}, - }).dump(), - /* .id = */ "", - }, - }, - }; + common_chat_msg msg; + msg.role = "assistant"; + msg.tool_calls.push_back({ + /* .name = */ name, + /* .arguments = */ (json { + {arg_name, arg_value}, + }).dump(), + /* .id = */ "", + }); + return msg; + } catch (const std::exception & e) { + LOG_WRN("Failed to parse builtin tool call arguments (%s): %s", e.what(), input.c_str()); + } } } return parse_json_tool_calls(input, std::nullopt, function_regex, close_regex); } -static common_chat_params common_chat_params_init_deepseek_r1(const common_chat_template & tmpl, const struct common_chat_inputs & inputs) { +static common_chat_params common_chat_params_init_deepseek_r1(const common_chat_template & tmpl, const struct templates_params & inputs) { common_chat_params data; - data.grammar_lazy = inputs.tool_choice != "required"; - data.grammar = build_grammar([&](const common_grammar_builder & builder) { - std::vector tool_rules; - foreach_function(inputs.tools, [&](const json & tool) { - const auto & function = tool["function"]; - std::string name = function["name"]; - auto parameters = function["parameters"]; - auto args_rule = builder.add_schema(name + "-args", parameters); - tool_rules.push_back(builder.add_rule(name + "-call", - "\"<｜tool▁call▁begin｜>function<｜tool▁sep｜>" + name + "\\n```json\\n\" " + args_rule + " \"```<｜tool▁call▁end｜>\"")); + if (inputs.tools.is_array() && !inputs.tools.empty()) { + data.grammar_lazy = inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_REQUIRED && inputs.json_schema.is_null(); + data.grammar = build_grammar([&](const common_grammar_builder & builder) { + std::vector tool_rules; + foreach_function(inputs.tools, [&](const json & tool) { + const auto & function = tool.at("function"); + std::string name = function.at("name"); + auto parameters = function.at("parameters"); + builder.resolve_refs(parameters); + tool_rules.push_back(builder.add_rule(name + "-call", + "\"<｜tool▁call▁begin｜>function<｜tool▁sep｜>" + name + "\\n" + "```json\\n\" " + builder.add_schema(name + "-args", parameters) + " " + "\"```<｜tool▁call▁end｜>\"")); + }); + // Distill Qwen 7B & 32B models seem confused re/ syntax of their tool call opening tag, + // so we accept common variants (then it's all constrained) + builder.add_rule("root", + "( \"<｜tool▁calls▁begin｜>\" | \"<｜tool_calls_begin｜>\" | \"<｜tool calls begin｜>\" | \"<｜tool\\\\_calls\\\\_begin｜>\" ) " + "(" + string_join(tool_rules, " | ") + ")" + (inputs.parallel_tool_calls ? "*" : "") + " " + "\"<｜tool▁calls▁end｜>\"" + " space"); + data.grammar_triggers.push_back({COMMON_GRAMMAR_TRIGGER_TYPE_WORD, "<｜tool▁calls▁begin｜>"}); + data.grammar_triggers.push_back({COMMON_GRAMMAR_TRIGGER_TYPE_WORD, "<｜tool_calls_begin｜>"}); + data.grammar_triggers.push_back({COMMON_GRAMMAR_TRIGGER_TYPE_WORD, "<｜tool calls begin｜>"}); + data.grammar_triggers.push_back({COMMON_GRAMMAR_TRIGGER_TYPE_WORD, "<｜tool\\_calls\\_begin｜>"}); + data.preserved_tokens = { + "", + "", + "<｜tool▁calls▁begin｜>", + "<｜tool▁call▁begin｜>", + "<｜tool▁sep｜>", + "<｜tool▁call▁end｜>", + "<｜tool▁calls▁end｜", + }; }); - data.grammar_triggers.push_back({"<｜tool▁calls▁begin｜>", /* .at_start = */ false}); - data.preserved_tokens = { - "<｜tool▁sep｜>", - "<｜tool▁call▁end｜>", - }; - builder.add_rule("root", "\"<｜tool▁calls▁begin｜>\" (" + string_join(tool_rules, " | ") + ")" + (inputs.parallel_tool_calls ? "*" : "") + " space"); - }, grammar_options); + } auto prompt = apply(tmpl, inputs.messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt); + + // Hacks to fix the official (broken) prompt. + // It is advisable to use --chat-template-file models/templates/llama-cpp-deepseek-r1.jinja instead, + // until the official template is fixed. + if (tmpl.source().find("{% if ns.is_tool %}{{'<｜tool▁outputs▁end｜>'}}") != std::string::npos) { + // Don't leave the chat dangling after tool results + if (string_ends_with(prompt, "<｜tool▁outputs▁end｜>")) { + prompt += "<｜end▁of▁sentence｜>"; + if (inputs.add_generation_prompt) { + prompt += "<｜Assistant｜>"; + } + } + // Fix up tool call delta example added by Minja + prompt = std::regex_replace( + prompt, + std::regex("(<｜tool▁call▁end｜>)[\\s\\r\\n]*(<｜tool▁outputs▁begin｜>|<｜User｜>)"), + "$1<｜tool▁calls▁end｜><｜end▁of▁sentence｜>$2"); + } data.prompt = prompt; - data.format = COMMON_CHAT_FORMAT_DEEPSEEK_R1; + data.format = inputs.extract_reasoning ? COMMON_CHAT_FORMAT_DEEPSEEK_R1_EXTRACT_REASONING : COMMON_CHAT_FORMAT_DEEPSEEK_R1; return data; } -static common_chat_msg common_chat_parse_deepseek_r1(const std::string & input) { - static std::regex trigger_regex("<｜tool▁calls▁begin｜>"); - static std::regex function_regex("<｜tool▁call▁begin｜>function<｜tool▁sep｜>([^\n]+)\n```json\n"); - static std::regex close_regex("```<｜tool▁call▁end｜>"); - return parse_json_tool_calls(input, trigger_regex, function_regex, close_regex); +static common_chat_msg handle_think_tag_prelude(const std::string & input, bool extract_reasoning, const std::function & rest_parser) { + std::smatch match; + static const std::regex reasoning_content_regex("((?:)?([\\s\\S\\r\\n]*?))?([\\s\\S\\r\\n]*)"); + if (std::regex_match(input, match, reasoning_content_regex)) { + auto rest = match[3].str(); + auto msg = rest_parser(rest); + auto reasoning_content = string_strip(match[2].str()); + if (extract_reasoning) { + msg.reasoning_content = reasoning_content; + } else if (!reasoning_content.empty()) { + std::ostringstream content; + content << "" << reasoning_content << "" << msg.content; + msg.content = content.str(); + } + return msg; + } + return rest_parser(input); +} +static common_chat_msg common_chat_parse_deepseek_r1(const std::string & input, bool extract_reasoning) { + return handle_think_tag_prelude(input, extract_reasoning, [](const std::string & input) { + static const std::regex function_regex("<｜tool▁call▁begin｜>function<｜tool▁sep｜>([^\n]+)\n```json\n"); + static const std::regex close_regex("```[\\s\\r\\n]*<｜tool▁call▁end｜>"); + static const std::regex tool_calls_regex("[\\s\\r\\n]*(?:<｜tool▁calls▁begin｜>|<｜tool_calls_begin｜>|<｜tool calls begin｜>|<｜tool\\\\_calls\\\\_begin｜>)([\\s\\S\\r\\n]*?)<｜tool▁calls▁end｜>"); + + common_chat_msg msg; + msg.role = "assistant"; + std::smatch match; + if (std::regex_search(input, match, tool_calls_regex)) { + auto tool_calls = match[1].str(); + auto msg2 = parse_json_tool_calls(tool_calls, std::nullopt, function_regex, close_regex); + msg.tool_calls = std::move(msg2.tool_calls); + } else { + msg.content = input; + } + return msg; + }); } -static common_chat_params common_chat_params_init_firefunction_v2(const common_chat_template & tmpl, const struct common_chat_inputs & inputs) { - fprintf(stderr, "%s\n", __func__); +static common_chat_params common_chat_params_init_firefunction_v2(const common_chat_template & tmpl, const struct templates_params & inputs) { + LOG_DBG("%s\n", __func__); common_chat_params data; data.prompt = apply(tmpl, inputs.messages, /* tools= */ nullptr, inputs.add_generation_prompt, { {"datetime", "Jan 29 2025 13:00:00 GMT"}, {"functions", json(inputs.tools.empty() ? "" : inputs.tools.dump(2))}, }); - if (!inputs.tools.is_null() && !inputs.tools.empty()) { - data.grammar_lazy = inputs.tool_choice != "required"; + if (inputs.tools.is_array() && !inputs.tools.empty()) { + data.grammar_lazy = inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_REQUIRED; data.grammar = build_grammar([&](const common_grammar_builder & builder) { auto schemas = json::array(); foreach_function(inputs.tools, [&](const json & tool) { - const auto & function = tool["function"]; + const auto & function = tool.at("function"); schemas.push_back({ {"type", "object"}, {"properties", { {"name", { {"type", "string"}, - {"const", function["name"]}, + {"const", function.at("name")}, }}, - {"arguments", function["parameters"]}, + {"arguments", function.at("parameters")}, }}, {"required", json::array({"name", "arguments", "id"})}, }); @@ -610,8 +1178,11 @@ static common_chat_params common_chat_params_init_firefunction_v2(const common_c schema["maxItems"] = 1; } builder.add_rule("root", "\" functools\"? " + builder.add_schema("tool_calls", schema)); - }, grammar_options); - data.grammar_triggers.push_back({" functools[", /* .at_start = */ false}); + }); + data.grammar_triggers.push_back({COMMON_GRAMMAR_TRIGGER_TYPE_WORD, " functools["}); + data.preserved_tokens = { + " functools[", + }; data.format = COMMON_CHAT_FORMAT_FIREFUNCTION_V2; } else { data.format = COMMON_CHAT_FORMAT_CONTENT_ONLY; @@ -622,27 +1193,45 @@ static common_chat_msg common_chat_parse_firefunction_v2(const std::string & inp return parse_prefixed_json_tool_call_array(input, " functools[", /* rstrip_prefix= */ 1); } -static common_chat_params common_chat_params_init_functionary_v3_2(const common_chat_template & tmpl, const struct common_chat_inputs & inputs) { +static common_chat_params common_chat_params_init_functionary_v3_2(const common_chat_template & tmpl, const struct templates_params & inputs) { // >>>all\nlet's call functions>>>fn1\n{"arg1": 1...}\n>>>fn2\n{"arg1": 1...}... // Using ">>>f1\n", ">>>f2\n"... as trigger words for the grammar common_chat_params data; data.prompt = apply(tmpl, inputs.messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt); data.format = COMMON_CHAT_FORMAT_FUNCTIONARY_V3_2; - if (!inputs.tools.is_null() && !inputs.tools.empty()) { - data.grammar_lazy = inputs.tool_choice != "required"; + if (inputs.tools.is_array() && !inputs.tools.empty()) { + data.grammar_lazy = inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_REQUIRED; data.grammar = build_grammar([&](const common_grammar_builder & builder) { std::vector first_tool_rules; std::vector subsequent_tool_rules; foreach_function(inputs.tools, [&](const json & tool) { - const auto & function = tool["function"]; - std::string name = function["name"]; - auto parameters = function["parameters"]; + const auto & function = tool.at("function"); + std::string name = function.at("name"); + auto parameters = function.at("parameters"); + builder.resolve_refs(parameters); auto args_rule = builder.add_schema(name + "-args", parameters); - first_tool_rules.push_back(builder.add_rule(name + "-call", "\"" + name + "\\n\" " + args_rule)); + first_tool_rules.push_back(builder.add_rule(name + "-call", "( \"assistant<|end_header_id|>\\n\" )? \"" + name + "\\n\" " + args_rule)); subsequent_tool_rules.push_back(builder.add_rule(name + "-call2", "\">>>" + name + "\\n\" " + args_rule)); - data.grammar_triggers.push_back({name, /* .at_start = */ true}); - data.grammar_triggers.push_back({">>>" + name, /* .at_start = */ false}); + data.grammar_triggers.push_back({ + COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN_START, + regex_escape(name + "\n"), + }); + data.grammar_triggers.push_back({ + COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN_START, + regex_escape("assistant<|end_header_id|>\n" + name + "\n"), + }); + data.grammar_triggers.push_back({ + COMMON_GRAMMAR_TRIGGER_TYPE_WORD, + regex_escape(">>>" + name + "\n"), + }); + data.grammar_triggers.push_back({ + COMMON_GRAMMAR_TRIGGER_TYPE_WORD, + ">>>assistant<|end_header_id|>\n" + name, + }); }); + data.preserved_tokens = { + "<|end_header_id|>", + }; auto first_rule = first_tool_rules.empty() ? "" : builder.add_rule("first_tool_call", string_join(first_tool_rules, " | ")) + " space"; if (inputs.parallel_tool_calls) { auto subsequent_rule = builder.add_rule("subsequent_tool_call", string_join(subsequent_tool_rules, " | ")) + " space"; @@ -651,34 +1240,20 @@ static common_chat_params common_chat_params_init_functionary_v3_2(const common_ builder.add_rule("root", first_rule); } - }, grammar_options); + }); } return data; } -static bool consume(std::string::const_iterator & it, const std::string::const_iterator & end, const std::string & expected) { - auto expected_it = expected.begin(); - auto tmp_it = it; - while (tmp_it != end && expected_it != expected.end() && *tmp_it == *expected_it) { - ++tmp_it; - ++expected_it; - } - if (expected_it == expected.end()) { - it = tmp_it; - return true; - } - return false; -} - static common_chat_msg common_chat_parse_functionary_v3_2(const std::string & input) { - static std::regex function_regex(R"((?:>>>)?(\w+)\n)"); - static std::regex close_regex(R"($|(?=>>>))"); + static const std::regex function_regex(R"((?:>>>)?(?:assistant<|end_header_id|>\n)?(\w+)\n)"); + static const std::regex close_regex(R"($|(?=>>>))"); std::string content; auto it = input.begin(); const auto end = input.end(); - if (consume(it, end, "all\n")) { + if (parse_literal(it, end, "all\n")) { std::smatch match; if (std::regex_search(it, end, match, function_regex)) { auto fun_it = match.prefix().second; @@ -693,7 +1268,7 @@ static common_chat_msg common_chat_parse_functionary_v3_2(const std::string & in } // TODO: tighten & simplify. try { - auto res = parse_json_tool_calls(std::string(it, end), std::nullopt, function_regex, close_regex); + auto res = parse_json_tool_calls(std::string(it, end), std::nullopt, function_regex, close_regex, /* allow_raw_python= */ true); res.content = content + res.content; return res; } catch (const std::exception & e) { @@ -705,26 +1280,26 @@ static common_chat_msg common_chat_parse_functionary_v3_2(const std::string & in } } -static common_chat_params common_chat_params_init_functionary_v3_1_llama_3_1(const common_chat_template & tmpl, const struct common_chat_inputs & inputs) { +static common_chat_params common_chat_params_init_functionary_v3_1_llama_3_1(const common_chat_template & tmpl, const struct templates_params & inputs) { // https://github.com/MeetKai/functionary/blob/main/tests/prompt_test_v3-llama3.1.txt common_chat_params data; json tools = inputs.tools.is_null() ? inputs.tools : json::array(); std::string python_code_argument_name; auto has_raw_python = false; - data.grammar_lazy = inputs.tool_choice != "required"; + data.grammar_lazy = inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_REQUIRED; data.grammar = build_grammar([&](const common_grammar_builder & builder) { std::vector tool_rules; foreach_function(inputs.tools, [&](const json & tool) { - const auto & function = tool["function"]; - const auto & parameters = function["parameters"]; - std::string name = function["name"]; + const auto & function = tool.at("function"); + const auto & parameters = function.at("parameters"); + std::string name = function.at("name"); if (name == "python" || name == "ipython") { if (!parameters.contains("type")) { throw std::runtime_error("Missing type in python tool"); } has_raw_python = true; - auto type = parameters.at("type"); + const auto & type = parameters.at("type"); if (type == "object") { auto properties = parameters.at("properties"); for (auto it = properties.begin(); it != properties.end(); ++it) { @@ -746,12 +1321,13 @@ static common_chat_params common_chat_params_init_functionary_v3_1_llama_3_1(con }); if (has_raw_python) { tool_rules.push_back(builder.add_rule("python-call", "\"<|python_tag|>\" .*")); - data.grammar_triggers.push_back({"<|python_tag|>", /* .at_start = */ false}); + data.grammar_triggers.push_back({COMMON_GRAMMAR_TRIGGER_TYPE_WORD, "<|python_tag|>"}); + data.preserved_tokens.push_back("<|python_tag|>"); } auto tool_call = builder.add_rule("tool_call", string_join(tool_rules, " | ")) + " space"; builder.add_rule("root", inputs.parallel_tool_calls ? "(" + tool_call + ")+" : tool_call); - data.grammar_triggers.push_back({"([\s\S\n]*)$)"); + static const std::regex python_tag_regex(R"(<\|python_tag\|>([\s\S\n]*)$)"); std::smatch match; if (std::regex_search(input, match, python_tag_regex)) { auto code = match[1].str(); - return { - /* .role = */ "assistant", - /* .content = */ match.prefix().str(), - /* .tool_calls = */ { - { - /* .name = */ "python", - /* .arguments = */ (json {{"code", code}}).dump(), - /* .id = */ "", - }, - } - }; + common_chat_msg msg; + msg.role = "assistant"; + msg.content = match.prefix().str(); + msg.tool_calls.push_back({ + /* .name = */ "python", + /* .arguments = */ (json {{"code", code}}).dump(), + /* .id = */ "", + }); + return msg; } - static std::regex function_regex(R"()"); - static std::regex close_regex(R"()"); + static const std::regex function_regex(R"()"); + static const std::regex close_regex(R"()"); // TODO: tighten & simplify. return parse_json_tool_calls(input, std::nullopt, function_regex, close_regex); } -static common_chat_params common_chat_params_init_hermes_2_pro(const common_chat_template & tmpl, const struct common_chat_inputs & inputs) { +static common_chat_params common_chat_params_init_hermes_2_pro(const common_chat_template & tmpl, const struct templates_params & inputs) { common_chat_params data; // (content)?({"name": "foo", "arguments": {"a": 1}})* - data.grammar_lazy = inputs.tool_choice != "required"; + data.grammar_lazy = inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_REQUIRED; data.grammar = build_grammar([&](const common_grammar_builder & builder) { std::vector tool_rules; + std::vector tool_call_alts; foreach_function(inputs.tools, [&](const json & tool) { - const auto & function = tool["function"]; - std::string name = function["name"]; - auto parameters = function["parameters"]; + const auto & function = tool.at("function"); + std::string name = function.at("name"); + auto parameters = function.at("parameters"); builder.resolve_refs(parameters); tool_rules.push_back(builder.add_schema(name + "-call", { {"type", "object"}, @@ -801,73 +1376,190 @@ static common_chat_params common_chat_params_init_hermes_2_pro(const common_chat }}, {"required", json::array({"name", "arguments"})}, })); + tool_call_alts.push_back(builder.add_rule( + name + "-function-tag", + "\"\" space " + + builder.add_schema(name + "-args", parameters) + " " + "\"\" space")); + + data.grammar_triggers.push_back({ + COMMON_GRAMMAR_TRIGGER_TYPE_WORD, + "", + }); + auto escaped_name = regex_escape(name); + data.grammar_triggers.push_back({ + COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN, + "\" space " + builder.add_rule("tool_call", string_join(tool_rules, " | ")) + " \"\" space"; + auto any_tool_call = builder.add_rule("any_tool_call", "( " + string_join(tool_rules, " | ") + " ) space"); + std::vector alt_tags { + any_tool_call, + "\"\" space " + any_tool_call + " \"\"", + // The rest is just to accommodate common "good bad" outputs. + "\"\" space " + any_tool_call + " \"\"", + "\"\" space " + any_tool_call + " \"\"", + "\"\" space " + any_tool_call + " \"\"", + "\"\" space " + any_tool_call + " \"\"", + "\"\" space " + any_tool_call + " \"\"", + "\"\" space " + any_tool_call + " \"\"", + }; + auto wrappable_tool_call = builder.add_rule("wrappable_tool_call", "( " + string_join(alt_tags, " | ") + " ) space"); + tool_call_alts.push_back(wrappable_tool_call); + tool_call_alts.push_back( + "( \"```\\n\" | \"```json\\n\" | \"```xml\\n\" ) space " + wrappable_tool_call + " space \"```\" space "); + auto tool_call = builder.add_rule("tool_call", string_join(tool_call_alts, " | ")); builder.add_rule("root", inputs.parallel_tool_calls ? "(" + tool_call + ")+" : tool_call); - data.grammar_triggers.push_back({"", /* .at_start = */ false}); - data.preserved_tokens = { "" }; - }, grammar_options); + data.grammar_triggers.push_back({COMMON_GRAMMAR_TRIGGER_TYPE_WORD, ""}); + data.grammar_triggers.push_back({COMMON_GRAMMAR_TRIGGER_TYPE_WORD, "|||)?\\s*\\{\\s*\"", //name\"\\s*:\\s*\"" + escaped_name + "\"", + }); + data.preserved_tokens = { + "", + "", + "", + "", + "", + "", + "", + "", + "", + "", + "", + "", + "", + "", + "```", + "```json", + "```xml", + }; + }); data.prompt = apply(tmpl, inputs.messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt); - data.format = COMMON_CHAT_FORMAT_HERMES_2_PRO; + data.format = inputs.extract_reasoning ? COMMON_CHAT_FORMAT_HERMES_2_PRO_EXTRACT_REASONING : COMMON_CHAT_FORMAT_HERMES_2_PRO; return data; } -static common_chat_msg common_chat_parse_hermes_2_pro(const std::string & input) { - try { - std::regex start_pattern(R"([\n\s]*)"); - std::regex middle_pattern(R"([\n\s]*[\n\s]*)"); - std::regex end_pattern(R"([\n\s]*[\n\s]*$)"); +static common_chat_msg common_chat_parse_hermes_2_pro(const std::string& input, bool extract_reasoning) { + return handle_think_tag_prelude(input, extract_reasoning, [](const std::string & input) { + static const std::regex open_regex( + "(?:" + "(```(?:xml|json)?\\n\\s*)?" // match 1 (block_start) + "(" // match 2 (open_tag) + "|" + "|" + "|" + "|" + "|" + "|" + "|" + ")?" + "(\\s*\\{\\s*\"name\"\\s*:[\\s\\S]*)" // match 3 (named tool call + rest) + ")" + "|" + "(?:]+)>" // match 4 (function name) + "|)" // match 5 (function name again) + "([\\s\\S]*)" // match 6 (function arguments + rest)})" + ); - auto end = input.end(); - std::sregex_iterator rend; - std::sregex_iterator rit(input.begin(), end, start_pattern); - if (rit == rend) { - return { - /* .role = */ "assistant", - /* .content = */ input, - /* .tool_calls = */ {}, - }; - } + try { + common_chat_msg msg; + msg.role = "assistant"; - common_chat_msg result; - result.role = "assistant"; - result.content = rit->prefix(); + std::string::const_iterator it = input.begin(); + const std::string::const_iterator end = input.end(); + std::smatch match; - auto it = rit->suffix().first; - while (it != end) { - json call; - if (!parse_json(it, end, call)) { - throw std::runtime_error("Failed to parse json tool call"); - } - const auto & arguments = call["arguments"]; - result.tool_calls.push_back({ - call["name"], - arguments.dump(), - // arguments.is_string() ? arguments.get() : arguments.dump(), - /* id= */ "", - }); - rit = {it, end, middle_pattern}; - if (rit != rend) { - it = rit->suffix().first; - } else { - rit = {it, end, end_pattern}; - if (rit == rend) { - throw std::runtime_error("Malformed input, missing "); + while (it != end) { + if (std::regex_search(it, end, match, open_regex)) { + // Add content before the match + msg.content += std::string(it, match[0].first); + + auto block_start = match[1].str(); + std::string block_end = block_start.empty() ? "" : "```"; + + auto open_tag = match[2].str(); + std::string close_tag; + + if (match[3].matched) { + close_tag = open_tag.empty() ? "" : ""; + // Start parsing from after the opening tags + auto json_it = match[6].first; + json arguments; + if (parse_json(json_it, end, arguments)) { + msg.tool_calls.emplace_back(process_tool_call({ + {"name", function_name}, + {"arguments", arguments}, + })); + it = json_it; // Move iterator past parsed JSON + + // Handle close tags + consume_spaces(it, end); + if (!close_tag.empty() && !parse_literal(it, end, close_tag)) { + throw std::runtime_error("Failed to parse closing tag"); + } + consume_spaces(it, end); + if (!block_end.empty() && !parse_literal(it, end, block_end)) { + throw std::runtime_error("Failed to parse block end"); + } + consume_spaces(it, end); + } else { + // Not a valid tool call, treat as content + msg.content += std::string(match[0].first, match[0].second); + it = match[0].second; + } + } + } else { + // Add remaining content + msg.content += std::string(it, end); + break; } - break; } + return msg; + } catch (const std::exception & e) { + LOG_ERR("Failed to parse hermes 2 pro input: %s\n", e.what()); + common_chat_msg msg; + msg.role = "assistant"; + msg.content = input; + return msg; } - return result; - } catch (const std::exception & e) { - return { - /* .role = */ "assistant", - /* .content = */ input, - /* .tool_calls = */ {}, - }; - } + }); } -static common_chat_params common_chat_params_init_without_tools(const common_chat_template & tmpl, const struct common_chat_inputs & inputs) { +static common_chat_params common_chat_params_init_without_tools(const common_chat_template & tmpl, const struct templates_params & inputs) { common_chat_params data; data.prompt = apply(tmpl, inputs.messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt); data.format = COMMON_CHAT_FORMAT_CONTENT_ONLY; @@ -878,62 +1570,177 @@ static common_chat_params common_chat_params_init_without_tools(const common_cha } data.grammar = json_schema_to_grammar(inputs.json_schema); } else { - data.grammar = inputs.grammar.empty(); + data.grammar = inputs.grammar; } return data; } -common_chat_params common_chat_params_init(const common_chat_template & tmpl, const struct common_chat_inputs & inputs) { - auto has_tools = !inputs.tools.is_null() && inputs.tool_choice != "none"; - LOG_DBG("[%s] has_tools=%s\n", __func__, has_tools ? "true" : "false"); - - if (has_tools && !inputs.grammar.empty()) { - throw std::runtime_error("Cannot specify grammar with tools"); - } - +static common_chat_params common_chat_templates_apply_jinja( + const struct common_chat_templates * tmpls, + const struct common_chat_templates_inputs & inputs) +{ + templates_params params; + params.tools = common_chat_tools_to_json_oaicompat(inputs.tools); + const auto & tmpl = params.tools.is_array() && tmpls->template_tool_use + ? *tmpls->template_tool_use + : *tmpls->template_default; const auto & src = tmpl.source(); + const auto & caps = tmpl.original_caps(); + params.messages = common_chat_msgs_to_json_oaicompat(inputs.messages, /* concat_text= */ !tmpl.original_caps().requires_typed_content); + params.add_generation_prompt = inputs.add_generation_prompt; + params.extract_reasoning = inputs.extract_reasoning; + params.tool_choice = inputs.tool_choice; + params.grammar = inputs.grammar; + if (!inputs.json_schema.empty()) { + params.json_schema = json::parse(inputs.json_schema); + } + + if (inputs.parallel_tool_calls && !tmpl.original_caps().supports_parallel_tool_calls) { + LOG_DBG("Disabling parallel_tool_calls because the template does not support it\n"); + params.parallel_tool_calls = false; + } else { + params.parallel_tool_calls = inputs.parallel_tool_calls; + } + + if (params.tools.is_array()) { + if (params.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE && !params.grammar.empty()) { + throw std::runtime_error("Cannot specify grammar with tools"); + } + if (caps.supports_tool_calls && !caps.supports_tools) { + LOG_WRN("Template supports tool calls but does not natively describe tools. The fallback behaviour used may produce bad results, inspect prompt w/ --verbose & consider overriding the template.\n"); + } + } + + // DeepSeek R1: use handler in all cases except json schema (thinking / tools). + if (src.find("<｜tool▁calls▁begin｜>") != std::string::npos && params.json_schema.is_null()) { + return common_chat_params_init_deepseek_r1(tmpl, params); + } + + // Command R7B: : use handler in all cases except json schema (thinking / tools). + if (src.find("<|END_THINKING|><|START_ACTION|>") != std::string::npos && params.json_schema.is_null()) { + return common_chat_params_init_command_r7b(tmpl, params); + } + + // Hermes 2/3 Pro, Qwen 2.5 Instruct (w/ tools) + if (src.find("") != std::string::npos && params.json_schema.is_null() && params.tools.is_array() && params.json_schema.is_null()) { + return common_chat_params_init_hermes_2_pro(tmpl, params); + } + + // Use generic handler when mixing tools + JSON schema. + // TODO: support that mix in handlers below. + if ((params.tools.is_array() && params.json_schema.is_object())) { + return common_chat_params_init_generic(tmpl, params); + } + + // Functionary prepends "all\n" to plain content outputs, so we use its handler in all cases. if (src.find(">>>all") != std::string::npos) { - // Functionary prepends "all\n" to plain content outputs, so we use the parser no matter when - return common_chat_params_init_functionary_v3_2(tmpl, inputs); + return common_chat_params_init_functionary_v3_2(tmpl, params); } + + // Firefunction v2 requires datetime and functions in the context even w/o tools, so we also use its handler in all cases. if (src.find(" functools[") != std::string::npos) { - // Firefunction v2 requires datetime and functions in the context, even w/o tools. - return common_chat_params_init_firefunction_v2(tmpl, inputs); + return common_chat_params_init_firefunction_v2(tmpl, params); } - if (!has_tools) { - return common_chat_params_init_without_tools(tmpl, inputs); + // Plain handler (no tools) + if (params.tools.is_null() || inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_NONE) { + return common_chat_params_init_without_tools(tmpl, params); } - if (src.find("") != std::string::npos) { - return common_chat_params_init_hermes_2_pro(tmpl, inputs); - } + // Functionary v3.1 (w/ tools) if (src.find("<|start_header_id|>") != std::string::npos && src.find("ipython<|end_header_id|>") != std::string::npos) { auto allow_python_tag_builtin_tools = src.find("<|python_tag|>") != std::string::npos; - return common_chat_params_init_llama_3_1_tool_calls(tmpl, inputs, allow_python_tag_builtin_tools); - } - if (src.find("<｜tool▁calls▁begin｜>") != std::string::npos) { - return common_chat_params_init_deepseek_r1(tmpl, inputs); + return common_chat_params_init_llama_3_1_tool_calls(tmpl, params, allow_python_tag_builtin_tools); } + + // Mistral Nemo (w/ tools) if (src.find("[TOOL_CALLS]") != std::string::npos) { - return common_chat_params_init_mistral_nemo(tmpl, inputs); + return common_chat_params_init_mistral_nemo(tmpl, params); } - if (src.find("<|END_THINKING|><|START_ACTION|>") != std::string::npos) { - return common_chat_params_init_command_r7b(tmpl, inputs); + + // Generic fallback + return common_chat_params_init_generic(tmpl, params); +} + +// Legacy template route (adhoc C++ implementation of known templates), forward to llama_chat_apply_template. +static common_chat_params common_chat_templates_apply_legacy( + const struct common_chat_templates * tmpls, + const struct common_chat_templates_inputs & inputs) +{ + int alloc_size = 0; + std::vector chat; + std::vector contents; + for (const auto & msg : inputs.messages) { + auto content = msg.content; + for (const auto & part : msg.content_parts) { + if (part.type != "text") { + LOG_WRN("Ignoring non-text content part: %s\n", part.type.c_str()); + continue; + } + if (!content.empty()) { + content += "\n";; + } + content += part.text; + } + contents.emplace_back(std::move(content)); } - return common_chat_params_init_generic(tmpl, inputs); + for (size_t i = 0; i < contents.size(); ++i) { + const auto & msg = inputs.messages[i]; + const auto & content = contents[i]; + chat.push_back({msg.role.c_str(), content.c_str()}); + alloc_size += (msg.role.size() + content.size()) * 1.25; + } + + std::vector buf(alloc_size); + + // run the first time to get the total output length + const auto & src = tmpls->template_default->source(); + int32_t res = llama_chat_apply_template(src.c_str(), chat.data(), chat.size(), inputs.add_generation_prompt, buf.data(), buf.size()); + + // error: chat template is not supported + if (res < 0) { + // if the custom "tmpl" is not supported, we throw an error + // this is a bit redundant (for good), since we're not sure if user validated the custom template with llama_chat_verify_template() + throw std::runtime_error("this custom template is not supported"); + } + + // if it turns out that our buffer is too small, we resize it + if ((size_t) res > buf.size()) { + buf.resize(res); + res = llama_chat_apply_template(src.c_str(), chat.data(), chat.size(), inputs.add_generation_prompt, buf.data(), buf.size()); + } + + common_chat_params params; + params.prompt = std::string(buf.data(), res); + if (!inputs.json_schema.empty()) { + params.grammar = json_schema_to_grammar(json::parse(inputs.json_schema)); + } else { + params.grammar = inputs.grammar; + } + return params; +} + +common_chat_params common_chat_templates_apply( + const struct common_chat_templates * tmpls, + const struct common_chat_templates_inputs & inputs) +{ + GGML_ASSERT(tmpls != nullptr); + return inputs.use_jinja + ? common_chat_templates_apply_jinja(tmpls, inputs) + : common_chat_templates_apply_legacy(tmpls, inputs); } static common_chat_msg common_chat_parse_content_only(const std::string & input) { - return { - /* .role = */ "assistant", - /* .content = */ input, - /* .tool_calls = */ {}, - }; + common_chat_msg msg; + msg.role = "assistant"; + msg.content = input; + return msg; } common_chat_msg common_chat_parse(const std::string & input, common_chat_format format) { @@ -949,17 +1756,23 @@ common_chat_msg common_chat_parse(const std::string & input, common_chat_format case COMMON_CHAT_FORMAT_LLAMA_3_X_WITH_BUILTIN_TOOLS: return common_chat_parse_llama_3_1(input, /* with_builtin_tools= */ true); case COMMON_CHAT_FORMAT_DEEPSEEK_R1: - return common_chat_parse_deepseek_r1(input); + return common_chat_parse_deepseek_r1(input, /* extract_reasoning= */ false); + case COMMON_CHAT_FORMAT_DEEPSEEK_R1_EXTRACT_REASONING: + return common_chat_parse_deepseek_r1(input, /* extract_reasoning= */ true); case COMMON_CHAT_FORMAT_FUNCTIONARY_V3_2: return common_chat_parse_functionary_v3_2(input); case COMMON_CHAT_FORMAT_FUNCTIONARY_V3_1_LLAMA_3_1: return common_chat_parse_functionary_v3_1_llama_3_1(input); case COMMON_CHAT_FORMAT_HERMES_2_PRO: - return common_chat_parse_hermes_2_pro(input); + return common_chat_parse_hermes_2_pro(input, /* extract_reasoning= */ false); + case COMMON_CHAT_FORMAT_HERMES_2_PRO_EXTRACT_REASONING: + return common_chat_parse_hermes_2_pro(input, /* extract_reasoning= */ true); case COMMON_CHAT_FORMAT_FIREFUNCTION_V2: return common_chat_parse_firefunction_v2(input); case COMMON_CHAT_FORMAT_COMMAND_R7B: - return common_chat_parse_command_r7b(input); + return common_chat_parse_command_r7b(input, /* extract_reasoning= */ false); + case COMMON_CHAT_FORMAT_COMMAND_R7B_EXTRACT_REASONING: + return common_chat_parse_command_r7b(input, /* extract_reasoning= */ true); default: throw std::runtime_error("Unsupported format: " + common_chat_format_name(format)); } diff --git a/common/chat.h b/common/chat.h new file mode 100644 index 000000000..9aad84e88 --- /dev/null +++ b/common/chat.h @@ -0,0 +1,135 @@ +// Chat support (incl. tool call grammar constraining & output parsing) w/ generic & custom template handlers. + +#pragma once + +#include "common.h" +#include +#include + +struct common_chat_templates; + +struct common_chat_tool_call { + std::string name; + std::string arguments; + std::string id; +}; + +struct common_chat_msg_content_part { + std::string type; + std::string text; +}; + +struct common_chat_msg { + std::string role; + std::string content; + std::vector content_parts = {}; + std::vector tool_calls = {}; + std::string reasoning_content; + std::string tool_name; + std::string tool_call_id; +}; + +struct common_chat_tool { + std::string name; + std::string description; + std::string parameters; +}; + +enum common_chat_tool_choice { + COMMON_CHAT_TOOL_CHOICE_AUTO, + COMMON_CHAT_TOOL_CHOICE_REQUIRED, + COMMON_CHAT_TOOL_CHOICE_NONE, +}; + +enum common_chat_format { + COMMON_CHAT_FORMAT_CONTENT_ONLY, + COMMON_CHAT_FORMAT_GENERIC, + COMMON_CHAT_FORMAT_MISTRAL_NEMO, + COMMON_CHAT_FORMAT_LLAMA_3_X, + COMMON_CHAT_FORMAT_LLAMA_3_X_WITH_BUILTIN_TOOLS, + COMMON_CHAT_FORMAT_DEEPSEEK_R1, + COMMON_CHAT_FORMAT_DEEPSEEK_R1_EXTRACT_REASONING, + COMMON_CHAT_FORMAT_FIREFUNCTION_V2, + COMMON_CHAT_FORMAT_FUNCTIONARY_V3_2, + COMMON_CHAT_FORMAT_FUNCTIONARY_V3_1_LLAMA_3_1, + COMMON_CHAT_FORMAT_HERMES_2_PRO, + COMMON_CHAT_FORMAT_HERMES_2_PRO_EXTRACT_REASONING, + COMMON_CHAT_FORMAT_COMMAND_R7B, + COMMON_CHAT_FORMAT_COMMAND_R7B_EXTRACT_REASONING, + + COMMON_CHAT_FORMAT_COUNT, // Not a format, just the # formats +}; + +struct common_chat_templates_inputs { + std::vector messages; + std::string grammar; + std::string json_schema; + bool add_generation_prompt = true; + bool use_jinja = true; + // Parameters below only supported when use_jinja is true + std::vector tools; + common_chat_tool_choice tool_choice = COMMON_CHAT_TOOL_CHOICE_AUTO; + bool parallel_tool_calls = false; + bool extract_reasoning = true; +}; + +struct common_chat_params { + common_chat_format format = COMMON_CHAT_FORMAT_CONTENT_ONLY; + std::string prompt; + std::string grammar; + bool grammar_lazy = false; + std::vector grammar_triggers; + std::vector preserved_tokens; + std::vector additional_stops; +}; + +// Check if the template supplied via "--chat-template" is supported or not. Returns true if it's valid +bool common_chat_verify_template(const std::string & tmpl, bool use_jinja); + +void common_chat_templates_free(struct common_chat_templates * tmpls); + +struct common_chat_templates_deleter { void operator()(common_chat_templates * tmpls) { common_chat_templates_free(tmpls); } }; + +typedef std::unique_ptr common_chat_templates_ptr; + +common_chat_templates_ptr common_chat_templates_init( + const struct llama_model * model, + const std::string & chat_template_override, + const std::string & bos_token_override = "", + const std::string & eos_token_override = ""); + +bool common_chat_templates_was_explicit(const struct common_chat_templates * tmpls); +const char * common_chat_templates_source(const struct common_chat_templates * tmpls, const char * variant = nullptr); + + +struct common_chat_params common_chat_templates_apply( + const struct common_chat_templates * tmpls, + const struct common_chat_templates_inputs & inputs); + +// Format single message, while taking into account the position of that message in chat history +std::string common_chat_format_single( + const struct common_chat_templates * tmpls, + const std::vector & past_msg, + const common_chat_msg & new_msg, + bool add_ass, + bool use_jinja); + +// Returns an example of formatted chat +std::string common_chat_format_example( + const struct common_chat_templates * tmpls, + bool use_jinja); + +std::string common_chat_format_name(common_chat_format format); +common_chat_msg common_chat_parse( const std::string & input, common_chat_format format); + +common_chat_tool_choice common_chat_tool_choice_parse_oaicompat(const std::string & tool_choice); + +// Parses a JSON array of messages in OpenAI's chat completion API format. +// T can be std::string containing JSON or nlohmann::ordered_json +template std::vector common_chat_msgs_parse_oaicompat(const T & messages); +template T common_chat_msgs_to_json_oaicompat(const std::vector & msgs, bool concat_typed_text = false); + +// Parses a JSON array of tools in OpenAI's chat completion tool call API format. +// T can be std::string containing JSON or nlohmann::ordered_json +template std::vector common_chat_tools_parse_oaicompat(const T & tools); +template T common_chat_tools_to_json_oaicompat(const std::vector & tools); diff --git a/common/chat.hpp b/common/chat.hpp deleted file mode 100644 index 33e64a430..000000000 --- a/common/chat.hpp +++ /dev/null @@ -1,52 +0,0 @@ -// Chat support (incl. tool call grammar constraining & output parsing) w/ generic & custom template handlers. - -#pragma once - -#include "common.h" -#include -#include -#include -#include - -using json = nlohmann::ordered_json; - -struct common_chat_inputs { - json messages; - json tools; - json tool_choice; - json json_schema; - bool parallel_tool_calls; - bool stream; - std::string grammar; - bool add_generation_prompt = true; -}; - -enum common_chat_format { - COMMON_CHAT_FORMAT_CONTENT_ONLY, - COMMON_CHAT_FORMAT_GENERIC, - COMMON_CHAT_FORMAT_MISTRAL_NEMO, - COMMON_CHAT_FORMAT_LLAMA_3_X, - COMMON_CHAT_FORMAT_LLAMA_3_X_WITH_BUILTIN_TOOLS, - COMMON_CHAT_FORMAT_DEEPSEEK_R1, - COMMON_CHAT_FORMAT_FIREFUNCTION_V2, - COMMON_CHAT_FORMAT_FUNCTIONARY_V3_2, - COMMON_CHAT_FORMAT_FUNCTIONARY_V3_1_LLAMA_3_1, - COMMON_CHAT_FORMAT_HERMES_2_PRO, - COMMON_CHAT_FORMAT_COMMAND_R7B, - - COMMON_CHAT_FORMAT_COUNT, // Not a format, just the # formats -}; - -struct common_chat_params { - common_chat_format format = COMMON_CHAT_FORMAT_CONTENT_ONLY; - json prompt; - std::string grammar; - bool grammar_lazy = false; - std::vector grammar_triggers; - std::vector preserved_tokens; - std::vector additional_stops; -}; - -struct common_chat_params common_chat_params_init(const common_chat_template & tmpl, const struct common_chat_inputs & params); -std::string common_chat_format_name(common_chat_format format); -common_chat_msg common_chat_parse( const std::string & input, common_chat_format format); diff --git a/common/common.cpp b/common/common.cpp index 8661e164a..94f545f81 100644 --- a/common/common.cpp +++ b/common/common.cpp @@ -7,13 +7,7 @@ #include "common.h" #include "log.h" -// Change JSON_ASSERT from assert() to GGML_ASSERT: -#define JSON_ASSERT GGML_ASSERT -#include "json.hpp" -#include "json-schema-to-grammar.h" #include "llama.h" -#include "chat.hpp" -#include "chat-template.hpp" #include #include @@ -54,47 +48,11 @@ #include #include #endif -#if defined(LLAMA_USE_CURL) -#include -#include -#include -#endif #if defined(_MSC_VER) #pragma warning(disable: 4244 4267) // possible loss of data #endif -#if defined(LLAMA_USE_CURL) -#ifdef __linux__ -#include -#elif defined(_WIN32) -# if !defined(PATH_MAX) -# define PATH_MAX MAX_PATH -# endif -#else -#include -#endif -#define LLAMA_CURL_MAX_URL_LENGTH 2084 // Maximum URL Length in Chrome: 2083 - -// -// CURL utils -// - -using curl_ptr = std::unique_ptr; - -// cannot use unique_ptr for curl_slist, because we cannot update without destroying the old one -struct curl_slist_ptr { - struct curl_slist * ptr = nullptr; - ~curl_slist_ptr() { - if (ptr) { - curl_slist_free_all(ptr); - } - } -}; -#endif // LLAMA_USE_CURL - -using json = nlohmann::ordered_json; - // // CPU utils // @@ -485,6 +443,11 @@ void string_replace_all(std::string & s, const std::string & search, const std:: s = std::move(builder); } +std::string regex_escape(const std::string & s) { + static const std::regex special_chars("[.^$|()*+?\\[\\]{}\\\\]"); + return std::regex_replace(s, special_chars, "\\$0"); +} + std::string string_join(const std::vector & values, const std::string & separator) { std::ostringstream result; for (size_t i = 0; i < values.size(); ++i) { @@ -867,7 +830,7 @@ std::string fs_get_cache_directory() { if (getenv("LLAMA_CACHE")) { cache_directory = std::getenv("LLAMA_CACHE"); } else { -#ifdef __linux__ +#if defined(__linux__) || defined(__FreeBSD__) || defined(_AIX) if (std::getenv("XDG_CACHE_HOME")) { cache_directory = std::getenv("XDG_CACHE_HOME"); } else { @@ -877,7 +840,9 @@ std::string fs_get_cache_directory() { cache_directory = std::getenv("HOME") + std::string("/Library/Caches/"); #elif defined(_WIN32) cache_directory = std::getenv("LOCALAPPDATA"); -#endif // __linux__ +#else +# error Unknown architecture +#endif cache_directory = ensure_trailing_slash(cache_directory); cache_directory += "llama.cpp"; } @@ -898,22 +863,14 @@ std::string fs_get_cache_file(const std::string & filename) { // // Model utils // + struct common_init_result common_init_from_params(common_params & params) { common_init_result iparams; auto mparams = common_model_params_to_llama(params); - llama_model * model = nullptr; - - if (!params.hf_repo.empty() && !params.hf_file.empty()) { - model = common_load_model_from_hf(params.hf_repo, params.hf_file, params.model, params.hf_token, mparams); - } else if (!params.model_url.empty()) { - model = common_load_model_from_url(params.model_url, params.model, params.hf_token, mparams); - } else { - model = llama_model_load_from_file(params.model.c_str(), mparams); - } - + llama_model * model = llama_model_load_from_file(params.model.path.c_str(), mparams); if (model == NULL) { - LOG_ERR("%s: failed to load model '%s'\n", __func__, params.model.c_str()); + LOG_ERR("%s: failed to load model '%s'\n", __func__, params.model.path.c_str()); return iparams; } @@ -948,13 +905,13 @@ struct common_init_result common_init_from_params(common_params & params) { llama_context * lctx = llama_init_from_model(model, cparams); if (lctx == NULL) { - LOG_ERR("%s: failed to create context with model '%s'\n", __func__, params.model.c_str()); + LOG_ERR("%s: failed to create context with model '%s'\n", __func__, params.model.path.c_str()); llama_model_free(model); return iparams; } - if (params.ctx_shift && !llama_kv_cache_can_shift(lctx)) { - LOG_WRN("%s: KV cache shifting is not supported for this model, disabling KV cache shifting\n", __func__); + if (params.ctx_shift && !llama_kv_self_can_shift(lctx)) { + LOG_WRN("%s: KV cache shifting is not supported for this context, disabling KV cache shifting\n", __func__); params.ctx_shift = false; } @@ -1031,6 +988,8 @@ struct common_init_result common_init_from_params(common_params & params) { if (params.warmup) { LOG_WRN("%s: warming up the model with an empty run - please wait ... (--no-warmup to disable)\n", __func__); + llama_set_warmup(lctx, true); + std::vector tmp; llama_token bos = llama_vocab_bos(vocab); llama_token eos = llama_vocab_eos(vocab); @@ -1058,9 +1017,10 @@ struct common_init_result common_init_from_params(common_params & params) { if (llama_model_has_decoder(model)) { llama_decode(lctx, llama_batch_get_one(tmp.data(), std::min(tmp.size(), (size_t) params.n_batch))); } - llama_kv_cache_clear(lctx); + llama_kv_self_clear(lctx); llama_synchronize(lctx); llama_perf_context_reset(lctx); + llama_set_warmup(lctx, false); } iparams.model.reset(model); @@ -1069,6 +1029,19 @@ struct common_init_result common_init_from_params(common_params & params) { return iparams; } +std::string get_model_endpoint() { + const char * model_endpoint_env = getenv("MODEL_ENDPOINT"); + // We still respect the use of environment-variable "HF_ENDPOINT" for backward-compatibility. + const char * hf_endpoint_env = getenv("HF_ENDPOINT"); + const char * endpoint_env = model_endpoint_env ? model_endpoint_env : hf_endpoint_env; + std::string model_endpoint = "https://huggingface.co/"; + if (endpoint_env) { + model_endpoint = endpoint_env; + if (model_endpoint.back() != '/') model_endpoint += '/'; + } + return model_endpoint; +} + void common_set_adapter_lora(struct llama_context * ctx, std::vector & lora) { llama_clear_adapter_lora(ctx); for (auto & la : lora) { @@ -1084,15 +1057,18 @@ struct llama_model_params common_model_params_to_llama(common_params & params) { if (!params.devices.empty()) { mparams.devices = params.devices.data(); } + if (params.n_gpu_layers != -1) { mparams.n_gpu_layers = params.n_gpu_layers; } + mparams.main_gpu = params.main_gpu; mparams.split_mode = params.split_mode; mparams.tensor_split = params.tensor_split; mparams.use_mmap = params.use_mmap; mparams.use_mlock = params.use_mlock; mparams.check_tensors = params.check_tensors; + if (params.kv_overrides.empty()) { mparams.kv_overrides = NULL; } else { @@ -1100,6 +1076,13 @@ struct llama_model_params common_model_params_to_llama(common_params & params) { mparams.kv_overrides = params.kv_overrides.data(); } + if (params.tensor_buft_overrides.empty()) { + mparams.tensor_buft_overrides = NULL; + } else { + GGML_ASSERT(params.tensor_buft_overrides.back().pattern == nullptr && "Tensor buffer overrides not terminated with empty pattern"); + mparams.tensor_buft_overrides = params.tensor_buft_overrides.data(); + } + return mparams; } @@ -1159,451 +1142,6 @@ struct ggml_threadpool_params ggml_threadpool_params_from_cpu_params(const cpu_p return tpp; } -#ifdef LLAMA_USE_CURL - -#define CURL_MAX_RETRY 3 -#define CURL_RETRY_DELAY_SECONDS 2 - -static bool curl_perform_with_retry(const std::string & url, CURL * curl, int max_attempts, int retry_delay_seconds) { - int remaining_attempts = max_attempts; - - while (remaining_attempts > 0) { - LOG_INF("%s: Trying to download from %s (attempt %d of %d)...\n", __func__ , url.c_str(), max_attempts - remaining_attempts + 1, max_attempts); - - CURLcode res = curl_easy_perform(curl); - if (res == CURLE_OK) { - return true; - } - - int exponential_backoff_delay = std::pow(retry_delay_seconds, max_attempts - remaining_attempts) * 1000; - LOG_WRN("%s: curl_easy_perform() failed: %s, retrying after %d milliseconds...\n", __func__, curl_easy_strerror(res), exponential_backoff_delay); - - remaining_attempts--; - std::this_thread::sleep_for(std::chrono::milliseconds(exponential_backoff_delay)); - } - - LOG_ERR("%s: curl_easy_perform() failed after %d attempts\n", __func__, max_attempts); - - return false; -} - -static bool common_download_file(const std::string & url, const std::string & path, const std::string & hf_token) { - // Initialize libcurl - curl_ptr curl(curl_easy_init(), &curl_easy_cleanup); - curl_slist_ptr http_headers; - if (!curl) { - LOG_ERR("%s: error initializing libcurl\n", __func__); - return false; - } - - bool force_download = false; - - // Set the URL, allow to follow http redirection - curl_easy_setopt(curl.get(), CURLOPT_URL, url.c_str()); - curl_easy_setopt(curl.get(), CURLOPT_FOLLOWLOCATION, 1L); - - // Check if hf-token or bearer-token was specified - if (!hf_token.empty()) { - std::string auth_header = "Authorization: Bearer " + hf_token; - http_headers.ptr = curl_slist_append(http_headers.ptr, auth_header.c_str()); - curl_easy_setopt(curl.get(), CURLOPT_HTTPHEADER, http_headers.ptr); - } - -#if defined(_WIN32) - // CURLSSLOPT_NATIVE_CA tells libcurl to use standard certificate store of - // operating system. Currently implemented under MS-Windows. - curl_easy_setopt(curl.get(), CURLOPT_SSL_OPTIONS, CURLSSLOPT_NATIVE_CA); -#endif - - // Check if the file already exists locally - auto file_exists = std::filesystem::exists(path); - - // If the file exists, check its JSON metadata companion file. - std::string metadata_path = path + ".json"; - nlohmann::json metadata; - std::string etag; - std::string last_modified; - - if (file_exists) { - // Try and read the JSON metadata file (note: stream autoclosed upon exiting this block). - std::ifstream metadata_in(metadata_path); - if (metadata_in.good()) { - try { - metadata_in >> metadata; - LOG_INF("%s: previous metadata file found %s: %s\n", __func__, metadata_path.c_str(), metadata.dump().c_str()); - if (metadata.contains("url") && metadata.at("url").is_string()) { - auto previous_url = metadata.at("url").get(); - if (previous_url != url) { - LOG_ERR("%s: Model URL mismatch: %s != %s\n", __func__, url.c_str(), previous_url.c_str()); - return false; - } - } - if (metadata.contains("etag") && metadata.at("etag").is_string()) { - etag = metadata.at("etag"); - } - if (metadata.contains("lastModified") && metadata.at("lastModified").is_string()) { - last_modified = metadata.at("lastModified"); - } - } catch (const nlohmann::json::exception & e) { - LOG_ERR("%s: error reading metadata file %s: %s\n", __func__, metadata_path.c_str(), e.what()); - return false; - } - } - } else { - LOG_INF("%s: no previous model file found %s\n", __func__, path.c_str()); - } - - // Send a HEAD request to retrieve the etag and last-modified headers - struct common_load_model_from_url_headers { - std::string etag; - std::string last_modified; - }; - - common_load_model_from_url_headers headers; - - { - typedef size_t(*CURLOPT_HEADERFUNCTION_PTR)(char *, size_t, size_t, void *); - auto header_callback = [](char * buffer, size_t /*size*/, size_t n_items, void * userdata) -> size_t { - common_load_model_from_url_headers * headers = (common_load_model_from_url_headers *) userdata; - - static std::regex header_regex("([^:]+): (.*)\r\n"); - static std::regex etag_regex("ETag", std::regex_constants::icase); - static std::regex last_modified_regex("Last-Modified", std::regex_constants::icase); - - std::string header(buffer, n_items); - std::smatch match; - if (std::regex_match(header, match, header_regex)) { - const std::string & key = match[1]; - const std::string & value = match[2]; - if (std::regex_match(key, match, etag_regex)) { - headers->etag = value; - } else if (std::regex_match(key, match, last_modified_regex)) { - headers->last_modified = value; - } - } - return n_items; - }; - - curl_easy_setopt(curl.get(), CURLOPT_NOBODY, 1L); // will trigger the HEAD verb - curl_easy_setopt(curl.get(), CURLOPT_NOPROGRESS, 1L); // hide head request progress - curl_easy_setopt(curl.get(), CURLOPT_HEADERFUNCTION, static_cast(header_callback)); - curl_easy_setopt(curl.get(), CURLOPT_HEADERDATA, &headers); - - bool was_perform_successful = curl_perform_with_retry(url, curl.get(), CURL_MAX_RETRY, CURL_RETRY_DELAY_SECONDS); - if (!was_perform_successful) { - return false; - } - - long http_code = 0; - curl_easy_getinfo(curl.get(), CURLINFO_RESPONSE_CODE, &http_code); - if (http_code != 200) { - // HEAD not supported, we don't know if the file has changed - // force trigger downloading - force_download = true; - LOG_ERR("%s: HEAD invalid http status code received: %ld\n", __func__, http_code); - } - } - - bool should_download = !file_exists || force_download; - if (!should_download) { - if (!etag.empty() && etag != headers.etag) { - LOG_WRN("%s: ETag header is different (%s != %s): triggering a new download\n", __func__, etag.c_str(), headers.etag.c_str()); - should_download = true; - } else if (!last_modified.empty() && last_modified != headers.last_modified) { - LOG_WRN("%s: Last-Modified header is different (%s != %s): triggering a new download\n", __func__, last_modified.c_str(), headers.last_modified.c_str()); - should_download = true; - } - } - if (should_download) { - std::string path_temporary = path + ".downloadInProgress"; - if (file_exists) { - LOG_WRN("%s: deleting previous downloaded file: %s\n", __func__, path.c_str()); - if (remove(path.c_str()) != 0) { - LOG_ERR("%s: unable to delete file: %s\n", __func__, path.c_str()); - return false; - } - } - - // Set the output file - - struct FILE_deleter { - void operator()(FILE * f) const { - fclose(f); - } - }; - - std::unique_ptr outfile(fopen(path_temporary.c_str(), "wb")); - if (!outfile) { - LOG_ERR("%s: error opening local file for writing: %s\n", __func__, path.c_str()); - return false; - } - - typedef size_t(*CURLOPT_WRITEFUNCTION_PTR)(void * data, size_t size, size_t nmemb, void * fd); - auto write_callback = [](void * data, size_t size, size_t nmemb, void * fd) -> size_t { - return fwrite(data, size, nmemb, (FILE *)fd); - }; - curl_easy_setopt(curl.get(), CURLOPT_NOBODY, 0L); - curl_easy_setopt(curl.get(), CURLOPT_WRITEFUNCTION, static_cast(write_callback)); - curl_easy_setopt(curl.get(), CURLOPT_WRITEDATA, outfile.get()); - - // display download progress - curl_easy_setopt(curl.get(), CURLOPT_NOPROGRESS, 0L); - - // helper function to hide password in URL - auto llama_download_hide_password_in_url = [](const std::string & url) -> std::string { - std::size_t protocol_pos = url.find("://"); - if (protocol_pos == std::string::npos) { - return url; // Malformed URL - } - - std::size_t at_pos = url.find('@', protocol_pos + 3); - if (at_pos == std::string::npos) { - return url; // No password in URL - } - - return url.substr(0, protocol_pos + 3) + "********" + url.substr(at_pos); - }; - - // start the download - LOG_INF("%s: trying to download model from %s to %s (server_etag:%s, server_last_modified:%s)...\n", __func__, - llama_download_hide_password_in_url(url).c_str(), path.c_str(), headers.etag.c_str(), headers.last_modified.c_str()); - bool was_perform_successful = curl_perform_with_retry(url, curl.get(), CURL_MAX_RETRY, CURL_RETRY_DELAY_SECONDS); - if (!was_perform_successful) { - return false; - } - - long http_code = 0; - curl_easy_getinfo (curl.get(), CURLINFO_RESPONSE_CODE, &http_code); - if (http_code < 200 || http_code >= 400) { - LOG_ERR("%s: invalid http status code received: %ld\n", __func__, http_code); - return false; - } - - // Causes file to be closed explicitly here before we rename it. - outfile.reset(); - - // Write the updated JSON metadata file. - metadata.update({ - {"url", url}, - {"etag", headers.etag}, - {"lastModified", headers.last_modified} - }); - std::ofstream(metadata_path) << metadata.dump(4); - LOG_INF("%s: file metadata saved: %s\n", __func__, metadata_path.c_str()); - - if (rename(path_temporary.c_str(), path.c_str()) != 0) { - LOG_ERR("%s: unable to rename file: %s to %s\n", __func__, path_temporary.c_str(), path.c_str()); - return false; - } - } - - return true; -} - -struct llama_model * common_load_model_from_url( - const std::string & model_url, - const std::string & local_path, - const std::string & hf_token, - const struct llama_model_params & params) { - // Basic validation of the model_url - if (model_url.empty()) { - LOG_ERR("%s: invalid model_url\n", __func__); - return NULL; - } - - if (!common_download_file(model_url, local_path, hf_token)) { - return NULL; - } - - // check for additional GGUFs split to download - int n_split = 0; - { - struct gguf_init_params gguf_params = { - /*.no_alloc = */ true, - /*.ctx = */ NULL, - }; - auto * ctx_gguf = gguf_init_from_file(local_path.c_str(), gguf_params); - if (!ctx_gguf) { - LOG_ERR("\n%s: failed to load input GGUF from %s\n", __func__, local_path.c_str()); - return NULL; - } - - auto key_n_split = gguf_find_key(ctx_gguf, LLM_KV_SPLIT_COUNT); - if (key_n_split >= 0) { - n_split = gguf_get_val_u16(ctx_gguf, key_n_split); - } - - gguf_free(ctx_gguf); - } - - if (n_split > 1) { - char split_prefix[PATH_MAX] = {0}; - char split_url_prefix[LLAMA_CURL_MAX_URL_LENGTH] = {0}; - - // Verify the first split file format - // and extract split URL and PATH prefixes - { - if (!llama_split_prefix(split_prefix, sizeof(split_prefix), local_path.c_str(), 0, n_split)) { - LOG_ERR("\n%s: unexpected model file name: %s n_split=%d\n", __func__, local_path.c_str(), n_split); - return NULL; - } - - if (!llama_split_prefix(split_url_prefix, sizeof(split_url_prefix), model_url.c_str(), 0, n_split)) { - LOG_ERR("\n%s: unexpected model url: %s n_split=%d\n", __func__, model_url.c_str(), n_split); - return NULL; - } - } - - // Prepare download in parallel - std::vector> futures_download; - for (int idx = 1; idx < n_split; idx++) { - futures_download.push_back(std::async(std::launch::async, [&split_prefix, &split_url_prefix, &n_split, hf_token](int download_idx) -> bool { - char split_path[PATH_MAX] = {0}; - llama_split_path(split_path, sizeof(split_path), split_prefix, download_idx, n_split); - - char split_url[LLAMA_CURL_MAX_URL_LENGTH] = {0}; - llama_split_path(split_url, sizeof(split_url), split_url_prefix, download_idx, n_split); - - return common_download_file(split_url, split_path, hf_token); - }, idx)); - } - - // Wait for all downloads to complete - for (auto & f : futures_download) { - if (!f.get()) { - return NULL; - } - } - } - - return llama_model_load_from_file(local_path.c_str(), params); -} - -struct llama_model * common_load_model_from_hf( - const std::string & repo, - const std::string & remote_path, - const std::string & local_path, - const std::string & hf_token, - const struct llama_model_params & params) { - // construct hugging face model url: - // - // --repo ggml-org/models --file tinyllama-1.1b/ggml-model-f16.gguf - // https://huggingface.co/ggml-org/models/resolve/main/tinyllama-1.1b/ggml-model-f16.gguf - // - // --repo TheBloke/Mixtral-8x7B-v0.1-GGUF --file mixtral-8x7b-v0.1.Q4_K_M.gguf - // https://huggingface.co/TheBloke/Mixtral-8x7B-v0.1-GGUF/resolve/main/mixtral-8x7b-v0.1.Q4_K_M.gguf - // - - std::string model_url = "https://huggingface.co/"; - model_url += repo; - model_url += "/resolve/main/"; - model_url += remote_path; - - return common_load_model_from_url(model_url, local_path, hf_token, params); -} - -/** - * Allow getting the HF file from the HF repo with tag (like ollama), for example: - * - bartowski/Llama-3.2-3B-Instruct-GGUF:q4 - * - bartowski/Llama-3.2-3B-Instruct-GGUF:Q4_K_M - * - bartowski/Llama-3.2-3B-Instruct-GGUF:q5_k_s - * Tag is optional, default to "latest" (meaning it checks for Q4_K_M first, then Q4, then if not found, return the first GGUF file in repo) - * - * Return pair of (with "repo" already having tag removed) - * - * Note: we use the Ollama-compatible HF API, but not using the blobId. Instead, we use the special "ggufFile" field which returns the value for "hf_file". This is done to be backward-compatible with existing cache files. - */ -std::pair common_get_hf_file(const std::string & hf_repo_with_tag, const std::string & hf_token) { - auto parts = string_split(hf_repo_with_tag, ':'); - std::string tag = parts.size() > 1 ? parts.back() : "latest"; - std::string hf_repo = parts[0]; - if (string_split(hf_repo, '/').size() != 2) { - throw std::invalid_argument("error: invalid HF repo format, expected /[:quant]\n"); - } - - // fetch model info from Hugging Face Hub API - json model_info; - curl_ptr curl(curl_easy_init(), &curl_easy_cleanup); - curl_slist_ptr http_headers; - std::string res_str; - std::string url = "https://huggingface.co/v2/" + hf_repo + "/manifests/" + tag; - curl_easy_setopt(curl.get(), CURLOPT_URL, url.c_str()); - curl_easy_setopt(curl.get(), CURLOPT_NOPROGRESS, 1L); - typedef size_t(*CURLOPT_WRITEFUNCTION_PTR)(void * ptr, size_t size, size_t nmemb, void * data); - auto write_callback = [](void * ptr, size_t size, size_t nmemb, void * data) -> size_t { - static_cast(data)->append((char * ) ptr, size * nmemb); - return size * nmemb; - }; - curl_easy_setopt(curl.get(), CURLOPT_WRITEFUNCTION, static_cast(write_callback)); - curl_easy_setopt(curl.get(), CURLOPT_WRITEDATA, &res_str); -#if defined(_WIN32) - curl_easy_setopt(curl.get(), CURLOPT_SSL_OPTIONS, CURLSSLOPT_NATIVE_CA); -#endif - if (!hf_token.empty()) { - std::string auth_header = "Authorization: Bearer " + hf_token; - http_headers.ptr = curl_slist_append(http_headers.ptr, auth_header.c_str()); - } - // Important: the User-Agent must be "llama-cpp" to get the "ggufFile" field in the response - http_headers.ptr = curl_slist_append(http_headers.ptr, "User-Agent: llama-cpp"); - http_headers.ptr = curl_slist_append(http_headers.ptr, "Accept: application/json"); - curl_easy_setopt(curl.get(), CURLOPT_HTTPHEADER, http_headers.ptr); - - CURLcode res = curl_easy_perform(curl.get()); - - if (res != CURLE_OK) { - throw std::runtime_error("error: cannot make GET request to HF API"); - } - - long res_code; - curl_easy_getinfo(curl.get(), CURLINFO_RESPONSE_CODE, &res_code); - if (res_code == 200) { - model_info = json::parse(res_str); - } else if (res_code == 401) { - throw std::runtime_error("error: model is private or does not exist; if you are accessing a gated model, please provide a valid HF token"); - } else { - throw std::runtime_error(string_format("error from HF API, response code: %ld, data: %s", res_code, res_str.c_str())); - } - - // check response - if (!model_info.contains("ggufFile")) { - throw std::runtime_error("error: model does not have ggufFile"); - } - json & gguf_file = model_info.at("ggufFile"); - if (!gguf_file.contains("rfilename")) { - throw std::runtime_error("error: ggufFile does not have rfilename"); - } - - return std::make_pair(hf_repo, gguf_file.at("rfilename")); -} - -#else - -struct llama_model * common_load_model_from_url( - const std::string & /*model_url*/, - const std::string & /*local_path*/, - const std::string & /*hf_token*/, - const struct llama_model_params & /*params*/) { - LOG_WRN("%s: llama.cpp built without libcurl, downloading from an url not supported.\n", __func__); - return nullptr; -} - -struct llama_model * common_load_model_from_hf( - const std::string & /*repo*/, - const std::string & /*remote_path*/, - const std::string & /*local_path*/, - const std::string & /*hf_token*/, - const struct llama_model_params & /*params*/) { - LOG_WRN("%s: llama.cpp built without libcurl, downloading from Hugging Face not supported.\n", __func__); - return nullptr; -} - -std::pair common_get_hf_file(const std::string &, const std::string &) { - LOG_WRN("%s: llama.cpp built without libcurl, downloading from Hugging Face not supported.\n", __func__); - return std::make_pair("", ""); -} - -#endif // LLAMA_USE_CURL - // // Batch utils // @@ -1768,174 +1306,6 @@ std::string common_detokenize(const struct llama_vocab * vocab, const std::vecto return text; } -// -// Chat template utils -// - -bool common_chat_verify_template(const std::string & tmpl, bool use_jinja) { - if (use_jinja) { - try { - auto chat_template = common_chat_template(tmpl, "", ""); - common_chat_inputs inputs; - inputs.messages = json::array({{ - {"role", "user"}, - {"content", "test"}, - }}); - common_chat_params_init(chat_template, inputs); - return true; - } catch (const std::exception & e) { - LOG_ERR("%s: failed to apply template: %s\n", __func__, e.what()); - return false; - } - } - llama_chat_message chat[] = {{"user", "test"}}; - const int res = llama_chat_apply_template(tmpl.c_str(), chat, 1, true, nullptr, 0); - return res >= 0; -} - -std::string common_chat_apply_template( - const common_chat_template & tmpl, - const std::vector & msgs, - bool add_ass, - bool use_jinja) { - if (use_jinja) { - auto messages = json::array(); - for (const auto & msg : msgs) { - messages.push_back({{"role", msg.role}, {"content", msg.content}}); - } - common_chat_inputs inputs; - inputs.messages = messages; - inputs.add_generation_prompt = add_ass; - return common_chat_params_init(tmpl, inputs).prompt; - } - - int alloc_size = 0; - std::vector chat; - for (const auto & msg : msgs) { - chat.push_back({msg.role.c_str(), msg.content.c_str()}); - alloc_size += (msg.role.size() + msg.content.size()) * 1.25; - } - - std::vector buf(alloc_size); - - // run the first time to get the total output length - int32_t res = llama_chat_apply_template(tmpl.source().c_str(), chat.data(), chat.size(), add_ass, buf.data(), buf.size()); - - // error: chat template is not supported - if (res < 0) { - // if the custom "tmpl" is not supported, we throw an error - // this is a bit redundant (for good), since we're not sure if user validated the custom template with llama_chat_verify_template() - throw std::runtime_error("this custom template is not supported"); - } - - // if it turns out that our buffer is too small, we resize it - if ((size_t) res > buf.size()) { - buf.resize(res); - res = llama_chat_apply_template(tmpl.source().c_str(), chat.data(), chat.size(), add_ass, buf.data(), buf.size()); - } - - std::string formatted_chat(buf.data(), res); - return formatted_chat; -} - -std::string common_chat_format_single( - const common_chat_template & tmpl, - const std::vector & past_msg, - const common_chat_msg & new_msg, - bool add_ass, - bool use_jinja) { - std::ostringstream ss; - auto fmt_past_msg = past_msg.empty() ? "" : common_chat_apply_template(tmpl, past_msg, false, use_jinja); - std::vector chat_new(past_msg); - // if the past_msg ends with a newline, we must preserve it in the formatted version - if (add_ass && !fmt_past_msg.empty() && fmt_past_msg.back() == '\n') { - ss << "\n"; - }; - // format chat with new_msg - chat_new.push_back(new_msg); - auto fmt_new_msg = common_chat_apply_template(tmpl, chat_new, add_ass, use_jinja); - // get the diff part - ss << fmt_new_msg.substr(fmt_past_msg.size(), fmt_new_msg.size() - fmt_past_msg.size()); - return ss.str(); -} - -std::string common_chat_format_example(const common_chat_template & tmpl, bool use_jinja) { - std::vector msgs = { - {"system", "You are a helpful assistant", {}}, - {"user", "Hello", {}}, - {"assistant", "Hi there", {}}, - {"user", "How are you?", {}}, - }; - return common_chat_apply_template(tmpl, msgs, true, use_jinja); -} - -#define CHATML_TEMPLATE_SRC \ - "{%- for message in messages -%}\n" \ - " {{- '<|im_start|>' + message.role + '\n' + message.content + '<|im_end|>\n' -}}\n" \ - "{%- endfor -%}\n" \ - "{%- if add_generation_prompt -%}\n" \ - " {{- '<|im_start|>assistant\n' -}}\n" \ - "{%- endif -%}" - -common_chat_templates common_chat_templates_from_model(const struct llama_model * model, const std::string & chat_template_override) -{ - std::string default_template_src; - std::string template_tool_use_src; - - bool has_explicit_template = !chat_template_override.empty(); - if (chat_template_override.empty()) { - auto str = llama_model_chat_template(model, /* name */ nullptr); - if (str) { - default_template_src = str; - has_explicit_template = true; - } - str = llama_model_chat_template(model, /* name */ "tool_use"); - if (str) { - template_tool_use_src = str; - has_explicit_template = true; - } - } else { - default_template_src = chat_template_override; - } - if (default_template_src.empty() || default_template_src == "chatml") { - if (!template_tool_use_src.empty()) { - default_template_src = template_tool_use_src; - } else { - default_template_src = CHATML_TEMPLATE_SRC; - } - } - auto vocab = llama_model_get_vocab(model); - const auto get_token = [&](llama_token token, const char * name, const char * jinja_variable_name) { - if (token == LLAMA_TOKEN_NULL) { - if (default_template_src.find(jinja_variable_name) != std::string::npos - || template_tool_use_src.find(jinja_variable_name) != std::string::npos) { - LOG_WRN("%s: warning: vocab does not have a %s token, jinja template won't work as intended.\n", __func__, name); - } - return std::string(); - } else { - return common_token_to_piece(vocab, token, true); - } - }; - auto token_bos = get_token(llama_vocab_bos(vocab), "BOS", "bos_token"); - auto token_eos = get_token(llama_vocab_eos(vocab), "EOS", "eos_token"); - try { - return { - has_explicit_template, - std::make_unique(default_template_src, token_bos, token_eos), - template_tool_use_src.empty() - ? nullptr - : std::make_unique(template_tool_use_src, token_bos, token_eos), - }; - } catch (const std::exception & e) { - LOG_ERR("%s: failed to parse chat template: %s\n", __func__, e.what()); - return { - has_explicit_template, - std::make_unique(CHATML_TEMPLATE_SRC, token_bos, token_eos), - nullptr, - }; - } -} - // // KV cache utils // @@ -2195,4 +1565,3 @@ common_control_vector_data common_control_vector_load(const std::vector grammar_trigger_words; // optional trigger words to trigger lazy grammar - std::vector grammar_trigger_tokens; // optional trigger tokens to trigger lazy grammar and print trigger special tokens. + std::vector grammar_triggers; // optional triggers (for lazy grammars) std::set preserved_tokens; std::vector logit_bias; // logit biases to apply @@ -172,36 +180,42 @@ struct common_params_sampling { std::string print() const; }; +struct common_params_model { + std::string path = ""; // model local path // NOLINT + std::string url = ""; // model url to download // NOLINT + std::string hf_repo = ""; // HF repo // NOLINT + std::string hf_file = ""; // HF file // NOLINT +}; + struct common_params_speculative { std::vector devices; // devices to use for offloading int32_t n_ctx = 0; // draft context size int32_t n_max = 16; // maximum number of tokens to draft during speculative decoding - int32_t n_min = 5; // minimum number of draft tokens to use for speculative decoding + int32_t n_min = 0; // minimum number of draft tokens to use for speculative decoding int32_t n_gpu_layers = -1; // number of layers to store in VRAM for the draft model (-1 - use default) float p_split = 0.1f; // speculative decoding split probability - float p_min = 0.9f; // minimum speculative decoding probability (greedy) + float p_min = 0.75f; // minimum speculative decoding probability (greedy) struct cpu_params cpuparams; struct cpu_params cpuparams_batch; - std::string hf_repo = ""; // HF repo // NOLINT - std::string hf_file = ""; // HF file // NOLINT - - std::string model = ""; // draft model for speculative decoding // NOLINT - std::string model_url = ""; // model url to download // NOLINT + struct common_params_model model; }; struct common_params_vocoder { - std::string hf_repo = ""; // HF repo // NOLINT - std::string hf_file = ""; // HF file // NOLINT + struct common_params_model model; - std::string model = ""; // model path // NOLINT - std::string model_url = ""; // model url to download // NOLINT + std::string speaker_file = ""; // speaker file path // NOLINT bool use_guide_tokens = false; // enable guide tokens to improve TTS accuracy // NOLINT }; +enum common_reasoning_format { + COMMON_REASONING_FORMAT_NONE, + COMMON_REASONING_FORMAT_DEEPSEEK, // Extract thinking tag contents and return as `message.reasoning_content` +}; + struct common_params { int32_t n_predict = -1; // new tokens to predict int32_t n_ctx = 4096; // context size @@ -248,13 +262,12 @@ struct common_params { struct common_params_speculative speculative; struct common_params_vocoder vocoder; - std::string model = ""; // model path // NOLINT + struct common_params_model model; + std::string model_alias = ""; // model alias // NOLINT - std::string model_url = ""; // model url to download // NOLINT std::string hf_token = ""; // HF token // NOLINT - std::string hf_repo = ""; // HF repo // NOLINT - std::string hf_file = ""; // HF file // NOLINT std::string prompt = ""; // NOLINT + std::string system_prompt = ""; // NOLINT std::string prompt_file = ""; // store the external prompt file name // NOLINT std::string path_prompt_cache = ""; // path to file for saving/loading prompt eval state // NOLINT std::string input_prefix = ""; // string to prefix user inputs with // NOLINT @@ -266,6 +279,7 @@ struct common_params { std::vector in_files; // all input files std::vector antiprompt; // strings upon which more user input is prompted (a.k.a. reverse prompts) std::vector kv_overrides; + std::vector tensor_buft_overrides; bool lora_init_without_apply = false; // only load lora to memory, but do not apply it to ctx (user can manually apply lora later using llama_adapter_lora_apply) std::vector lora_adapters; // lora adapter path with user defined scale @@ -292,6 +306,7 @@ struct common_params { bool kl_divergence = false; // compute KL divergence bool usage = false; // print usage + bool completion = false; // print source-able completion script bool use_color = false; // use color to distinguish generations and inputs bool special = false; // enable special token output bool interactive = false; // interactive mode @@ -318,13 +333,15 @@ struct common_params { bool warmup = true; // warmup run bool check_tensors = false; // validate tensor data + bool single_turn = false; // single turn chat conversation + ggml_type cache_type_k = GGML_TYPE_F16; // KV cache data type for the K ggml_type cache_type_v = GGML_TYPE_F16; // KV cache data type for the V common_conversation_mode conversation_mode = COMMON_CONVERSATION_MODE_AUTO; // multimodal models (see examples/llava) - std::string mmproj = ""; // path to multimodal projector // NOLINT + struct common_params_model mmproj; std::vector image; // path to image file(s) // embedding @@ -346,6 +363,7 @@ struct common_params { std::string chat_template = ""; // NOLINT bool use_jinja = false; // NOLINT bool enable_chat_template = true; + common_reasoning_format reasoning_format = COMMON_REASONING_FORMAT_DEEPSEEK; std::vector api_keys; @@ -383,8 +401,6 @@ struct common_params { int32_t i_pos = -1; // position of the passkey in the junk text // imatrix params - std::string out_file = "imatrix.gguf"; // save the resulting imatrix to this file - int32_t n_out_freq = 10; // output the imatrix every n_out_freq iterations int32_t n_save_freq = 0; // save the imatrix every n_save_freq iterations int32_t i_chunk = 0; // start processing from this chunk @@ -396,16 +412,16 @@ struct common_params { int n_pca_batch = 100; int n_pca_iterations = 1000; dimre_method cvector_dimre_method = DIMRE_METHOD_PCA; - std::string cvector_outfile = "control_vector.gguf"; std::string cvector_positive_file = "examples/cvector-generator/positive.txt"; std::string cvector_negative_file = "examples/cvector-generator/negative.txt"; bool spm_infill = false; // suffix/prefix/middle pattern for infill - std::string lora_outfile = "ggml-lora-merged-f16.gguf"; - // batched-bench params bool batched_bench_output_jsonl = false; + + // common params + std::string out_file; // output filename for all example programs }; // call once at the start of a program if it uses libcommon @@ -424,13 +440,13 @@ bool set_process_priority(enum ggml_sched_priority prio); // #ifdef __GNUC__ -#ifdef __MINGW32__ -#define LLAMA_COMMON_ATTRIBUTE_FORMAT(...) __attribute__((format(gnu_printf, __VA_ARGS__))) +# if defined(__MINGW32__) && !defined(__clang__) +# define LLAMA_COMMON_ATTRIBUTE_FORMAT(...) __attribute__((format(gnu_printf, __VA_ARGS__))) +# else +# define LLAMA_COMMON_ATTRIBUTE_FORMAT(...) __attribute__((format(printf, __VA_ARGS__))) +# endif #else -#define LLAMA_COMMON_ATTRIBUTE_FORMAT(...) __attribute__((format(printf, __VA_ARGS__))) -#endif -#else -#define LLAMA_COMMON_ATTRIBUTE_FORMAT(...) +# define LLAMA_COMMON_ATTRIBUTE_FORMAT(...) #endif LLAMA_COMMON_ATTRIBUTE_FORMAT(1, 2) @@ -445,6 +461,8 @@ std::string string_repeat(const std::string & str, size_t n); void string_replace_all(std::string & s, const std::string & search, const std::string & replace); +std::string regex_escape(const std::string & s); + template static std::vector string_split(const std::string & str, char delim) { static_assert(!std::is_same::value, "Please use the specialized version for std::string"); @@ -522,26 +540,11 @@ struct llama_model_params common_model_params_to_llama ( common_params struct llama_context_params common_context_params_to_llama(const common_params & params); struct ggml_threadpool_params ggml_threadpool_params_from_cpu_params(const cpu_params & params); -struct llama_model * common_load_model_from_url( - const std::string & model_url, - const std::string & local_path, - const std::string & hf_token, - const struct llama_model_params & params); - -struct llama_model * common_load_model_from_hf( - const std::string & repo, - const std::string & remote_path, - const std::string & local_path, - const std::string & hf_token, - const struct llama_model_params & params); - -std::pair common_get_hf_file( - const std::string & hf_repo_with_tag, - const std::string & hf_token); - // clear LoRA adapters from context, then apply new list of adapters void common_set_adapter_lora(struct llama_context * ctx, std::vector & lora); +std::string get_model_endpoint(); + // // Batch utils // @@ -608,62 +611,6 @@ std::string common_detokenize( const std::vector & tokens, bool special = true); -// -// Chat template utils -// - -struct common_tool_call { - std::string name; - std::string arguments; - std::string id; -}; - -// same with llama_chat_message, but uses std::string -struct common_chat_msg { - std::string role; - std::string content; - std::vector tool_calls; - std::string tool_plan = ""; -}; - -// Check if the template supplied via "--chat-template" is supported or not. Returns true if it's valid -bool common_chat_verify_template(const std::string & tmpl, bool use_jinja); - -namespace minja { - class chat_template; -} - -typedef minja::chat_template common_chat_template; - -struct common_chat_templates { - bool has_explicit_template; // Model had builtin template or template overridde was specified. - std::unique_ptr template_default; // always set (defaults to chatml) - std::unique_ptr template_tool_use; -}; - -// CPP wrapper for llama_chat_apply_template -// If the built-in template is not supported, we default to chatml -// If the custom "tmpl" is not supported, we throw an error -std::string common_chat_apply_template( - const common_chat_template & tmpl, - const std::vector & chat, - bool add_ass, - bool use_jinja); - -// Format single message, while taking into account the position of that message in chat history -std::string common_chat_format_single( - const common_chat_template & tmpl, - const std::vector & past_msg, - const common_chat_msg & new_msg, - bool add_ass, - bool use_jinja); - -// Returns an example of formatted chat -std::string common_chat_format_example( - const common_chat_template & tmpl, bool use_jinja); - -common_chat_templates common_chat_templates_from_model(const struct llama_model * model, const std::string & chat_template_override); - // // KV cache utils // diff --git a/common/json-schema-to-grammar.cpp b/common/json-schema-to-grammar.cpp index 3ebcc3d9f..906798225 100644 --- a/common/json-schema-to-grammar.cpp +++ b/common/json-schema-to-grammar.cpp @@ -264,7 +264,7 @@ static void _build_min_max_int(int min_value, int max_value, std::stringstream & throw std::runtime_error("At least one of min_value or max_value must be set"); } -const std::string SPACE_RULE = "| \" \" | \"\\n\" [ \\t]{0,20}"; +const std::string SPACE_RULE = "| \" \" | \"\\n\"{1,2} [ \\t]{0,20}"; struct BuiltinRule { std::string content; @@ -764,11 +764,10 @@ private: public: SchemaConverter( const std::function & fetch_json, - bool dotall, - bool compact_spaces) + bool dotall) : _fetch_json(fetch_json), _dotall(dotall) { - _rules["space"] = compact_spaces ? "\" \"?" : SPACE_RULE; + _rules["space"] = SPACE_RULE; } void resolve_refs(json & schema, const std::string & url) { @@ -1007,7 +1006,7 @@ std::string json_schema_to_grammar(const json & schema, bool force_gbnf) { } std::string build_grammar(const std::function & cb, const common_grammar_options & options) { - SchemaConverter converter([&](const std::string &) { return json(); }, options.dotall, options.compact_spaces); + SchemaConverter converter([&](const std::string &) { return json(); }, options.dotall); common_grammar_builder builder { /* .add_rule = */ [&](const std::string & name, const std::string & rule) { return converter._add_rule(name, rule); diff --git a/common/json-schema-to-grammar.h b/common/json-schema-to-grammar.h index 62a3b0a44..4613f5d9f 100644 --- a/common/json-schema-to-grammar.h +++ b/common/json-schema-to-grammar.h @@ -16,7 +16,6 @@ struct common_grammar_builder { struct common_grammar_options { bool dotall = false; - bool compact_spaces = false; }; std::string build_grammar(const std::function & cb, const common_grammar_options & options = {}); diff --git a/common/llguidance.cpp b/common/llguidance.cpp index 2feeb93c8..8bff89ea4 100644 --- a/common/llguidance.cpp +++ b/common/llguidance.cpp @@ -11,25 +11,24 @@ struct llama_sampler_llg { std::string grammar_kind; std::string grammar_data; LlgTokenizer * tokenizer; - LlgConstraint * grammar; - LlgMaskResult llg_res; - bool has_llg_res; + LlgMatcher * grammar; }; -static LlgConstraint * llama_sampler_llg_new(LlgTokenizer * tokenizer, const char * grammar_kind, - const char * grammar_data) { +static LlgMatcher * llama_sampler_llg_new(LlgTokenizer * tokenizer, const char * grammar_kind, + const char * grammar_data) { LlgConstraintInit cinit; llg_constraint_init_set_defaults(&cinit, tokenizer); const char * log_level = getenv("LLGUIDANCE_LOG_LEVEL"); if (log_level && *log_level) { cinit.log_stderr_level = atoi(log_level); } - auto c = llg_new_constraint_any(&cinit, grammar_kind, grammar_data); - if (llg_get_error(c)) { - LOG_ERR("llg error: %s\n", llg_get_error(c)); - llg_free_constraint(c); + auto c = llg_new_matcher(&cinit, grammar_kind, grammar_data); + if (llg_matcher_get_error(c)) { + LOG_ERR("llg error: %s\n", llg_matcher_get_error(c)); + llg_free_matcher(c); return nullptr; } + return c; } @@ -40,39 +39,29 @@ static const char * llama_sampler_llg_name(const llama_sampler * /*smpl*/) { static void llama_sampler_llg_accept_impl(llama_sampler * smpl, llama_token token) { auto * ctx = (llama_sampler_llg *) smpl->ctx; if (ctx->grammar) { - LlgCommitResult res; - llg_commit_token(ctx->grammar, token, &res); - ctx->has_llg_res = false; + llg_matcher_consume_token(ctx->grammar, token); } } static void llama_sampler_llg_apply(llama_sampler * smpl, llama_token_data_array * cur_p) { auto * ctx = (llama_sampler_llg *) smpl->ctx; if (ctx->grammar) { - if (!ctx->has_llg_res) { - if (llg_compute_mask(ctx->grammar, &ctx->llg_res) == 0) { - ctx->has_llg_res = true; + const uint32_t * mask = llg_matcher_get_mask(ctx->grammar); + if (mask == nullptr) { + if (llg_matcher_compute_mask(ctx->grammar) == 0) { + mask = llg_matcher_get_mask(ctx->grammar); } else { - LOG_ERR("llg error: %s\n", llg_get_error(ctx->grammar)); - llg_free_constraint(ctx->grammar); + LOG_ERR("llg error: %s\n", llg_matcher_get_error(ctx->grammar)); + llg_free_matcher(ctx->grammar); ctx->grammar = nullptr; + return; } } - if (ctx->has_llg_res) { - if (ctx->llg_res.is_stop) { - for (size_t i = 0; i < cur_p->size; ++i) { - if (!llama_vocab_is_eog(ctx->vocab, cur_p->data[i].id)) { - cur_p->data[i].logit = -INFINITY; - } - } - } else { - const uint32_t * mask = ctx->llg_res.sample_mask; - for (size_t i = 0; i < cur_p->size; ++i) { - auto token = cur_p->data[i].id; - if ((mask[token / 32] & (1 << (token % 32))) == 0) { - cur_p->data[i].logit = -INFINITY; - } - } + + for (size_t i = 0; i < cur_p->size; ++i) { + auto token = cur_p->data[i].id; + if ((mask[token / 32] & (1 << (token % 32))) == 0) { + cur_p->data[i].logit = -INFINITY; } } } @@ -80,14 +69,9 @@ static void llama_sampler_llg_apply(llama_sampler * smpl, llama_token_data_array static void llama_sampler_llg_reset(llama_sampler * smpl) { auto * ctx = (llama_sampler_llg *) smpl->ctx; - if (!ctx->grammar) { - return; + if (ctx->grammar) { + llg_matcher_reset(ctx->grammar); } - - auto * grammar_new = llama_sampler_llg_new(ctx->tokenizer, ctx->grammar_kind.c_str(), ctx->grammar_data.c_str()); - llg_free_constraint(ctx->grammar); - ctx->grammar = grammar_new; - ctx->has_llg_res = false; } static llama_sampler * llama_sampler_llg_clone(const llama_sampler * smpl) { @@ -102,7 +86,7 @@ static llama_sampler * llama_sampler_llg_clone(const llama_sampler * smpl) { if (ctx->grammar) { result_ctx->grammar_kind = ctx->grammar_kind; result_ctx->grammar_data = ctx->grammar_data; - result_ctx->grammar = llg_clone_constraint(ctx->grammar); + result_ctx->grammar = llg_clone_matcher(ctx->grammar); result_ctx->tokenizer = llg_clone_tokenizer(ctx->tokenizer); } } @@ -114,7 +98,7 @@ static void llama_sampler_llg_free(llama_sampler * smpl) { const auto * ctx = (llama_sampler_llg *) smpl->ctx; if (ctx->grammar) { - llg_free_constraint(ctx->grammar); + llg_free_matcher(ctx->grammar); llg_free_tokenizer(ctx->tokenizer); } @@ -239,9 +223,11 @@ llama_sampler * llama_sampler_init_llg(const llama_vocab * vocab, const char * g /* .grammar_data = */ grammar_data, /* .tokenizer = */ tokenizer, /* .grammar = */ llama_sampler_llg_new(tokenizer, grammar_kind, grammar_data), - /* .llg_res = */ {}, - /* .has_llg_res = */ false, }; + if (ctx->grammar) { + GGML_ASSERT(((size_t) llama_vocab_n_tokens(vocab) + 31) / 32 * 4 == + llg_matcher_get_mask_byte_size(ctx->grammar)); + } } else { *ctx = { /* .vocab = */ vocab, @@ -249,15 +235,12 @@ llama_sampler * llama_sampler_init_llg(const llama_vocab * vocab, const char * g /* .grammar_data = */ {}, /* .tokenizer = */ nullptr, /* .grammar = */ nullptr, - /* .llg_res = */ {}, - /* .has_llg_res = */ false, }; } return llama_sampler_init( /* .iface = */ &llama_sampler_llg_i, - /* .ctx = */ ctx - ); + /* .ctx = */ ctx); } #else diff --git a/common/log.cpp b/common/log.cpp index 4bfbecf15..52b31470c 100644 --- a/common/log.cpp +++ b/common/log.cpp @@ -1,5 +1,6 @@ #include "log.h" +#include #include #include #include diff --git a/common/log.h b/common/log.h index 4ebc6314b..c56bb50d9 100644 --- a/common/log.h +++ b/common/log.h @@ -15,7 +15,7 @@ #ifndef __GNUC__ # define LOG_ATTRIBUTE_FORMAT(...) -#elif defined(__MINGW32__) +#elif defined(__MINGW32__) && !defined(__clang__) # define LOG_ATTRIBUTE_FORMAT(...) __attribute__((format(gnu_printf, __VA_ARGS__))) #else # define LOG_ATTRIBUTE_FORMAT(...) __attribute__((format(printf, __VA_ARGS__))) diff --git a/common/chat-template.hpp b/common/minja/chat-template.hpp similarity index 92% rename from common/chat-template.hpp rename to common/minja/chat-template.hpp index 0e88fb361..237a7625c 100644 --- a/common/chat-template.hpp +++ b/common/minja/chat-template.hpp @@ -9,10 +9,19 @@ #pragma once #include "minja.hpp" -#include + +#include +#include +#include +#include +#include +#include +#include #include #include +#include + using json = nlohmann::ordered_json; namespace minja { @@ -249,16 +258,30 @@ class chat_template { inputs.add_generation_prompt = false; full = apply(inputs); } - - if (full.find(prefix) != 0) { - if (prefix.rfind(eos_token_) == prefix.size() - eos_token_.size()) { - prefix = prefix.substr(0, prefix.size() - eos_token_.size()); + auto eos_pos_last = full.rfind(eos_token_); + if (eos_pos_last == prefix.size() - eos_token_.size() || + (full[full.size() - 1] == '\n' && (eos_pos_last == full.size() - eos_token_.size() - 1))) { + full = full.substr(0, eos_pos_last); + } + size_t common_prefix_length = 0; + for (size_t i = 0; i < prefix.size() && i < full.size(); ++i) { + if (prefix[i] != full[i]) { + break; } + if (prefix[i] == '<') { + // DeepSeek R1's template (as of 20250209) adds a trailing if add_generation_prompt, + // but it removes thinking tags for past messages. + // The prefix and full strings diverge at vs. <｜tool▁calls▁begin｜>, we avoid consuming the leading <. + continue; + } + common_prefix_length = i + 1; } - if (full.find(prefix) != 0) { + auto example = full.substr(common_prefix_length); + if (example.find("tool_name") == std::string::npos && example.find("some_value") == std::string::npos) { fprintf(stderr, "Failed to infer a tool call example (possible template bug)\n"); + } else { + tool_call_example_ = example; } - tool_call_example_ = full.substr(prefix.size()); } } catch (const std::exception & e) { fprintf(stderr, "Failed to generate tool call example: %s\n", e.what()); @@ -363,7 +386,7 @@ class chat_template { if (polyfill_tools) { adjusted_messages = add_system(inputs.messages, "You can call any of the following tools to satisfy the user's requests: " + minja::Value(inputs.tools).dump(2, /* to_json= */ true) + - (!polyfill_tool_call_example || tool_call_example_.empty() ? "" : "\n\nExample tool call syntax:\n\n" + tool_call_example_)); + (!polyfill_tool_call_example || tool_call_example_.empty() ? "" : "\n\nExample tool call syntax:\n\n" + tool_call_example_ + "\n\n")); } else { adjusted_messages = inputs.messages; } @@ -411,7 +434,7 @@ class chat_template { auto obj = json { {"tool_calls", tool_calls}, }; - if (!content.is_null() && content != "") { + if (!content.is_null() && !content.empty()) { obj["content"] = content; } message["content"] = obj.dump(2); @@ -421,13 +444,12 @@ class chat_template { if (polyfill_tool_responses && role == "tool") { message["role"] = "user"; auto obj = json { - {"tool_response", { - {"content", message.at("content")}, - }}, + {"tool_response", json::object()}, }; if (message.contains("name")) { - obj["tool_response"]["name"] = message.at("name"); + obj["tool_response"]["tool"] = message.at("name"); } + obj["tool_response"]["content"] = message.at("content"); if (message.contains("tool_call_id")) { obj["tool_response"]["tool_call_id"] = message.at("tool_call_id"); } @@ -496,7 +518,7 @@ class chat_template { static nlohmann::ordered_json add_system(const nlohmann::ordered_json & messages, const std::string & system_prompt) { json messages_with_system = messages; - if (messages_with_system.size() > 0 && messages_with_system[0].at("role") == "system") { + if (!messages_with_system.empty() && messages_with_system[0].at("role") == "system") { std::string existing_system = messages_with_system.at(0).at("content"); messages_with_system[0] = json { {"role", "system"}, diff --git a/common/minja.hpp b/common/minja/minja.hpp similarity index 90% rename from common/minja.hpp rename to common/minja/minja.hpp index c304b5c66..e52e792d8 100644 --- a/common/minja.hpp +++ b/common/minja/minja.hpp @@ -8,14 +8,26 @@ // SPDX-License-Identifier: MIT #pragma once +#include +#include +#include +#include +#include +#include #include -#include -#include -#include +#include +#include +#include #include -#include +#include #include +#include +#include +#include #include +#include +#include + #include using json = nlohmann::ordered_json; @@ -731,51 +743,51 @@ public: struct TextTemplateToken : public TemplateToken { std::string text; - TextTemplateToken(const Location & location, SpaceHandling pre, SpaceHandling post, const std::string& t) : TemplateToken(Type::Text, location, pre, post), text(t) {} + TextTemplateToken(const Location & loc, SpaceHandling pre, SpaceHandling post, const std::string& t) : TemplateToken(Type::Text, loc, pre, post), text(t) {} }; struct ExpressionTemplateToken : public TemplateToken { std::shared_ptr expr; - ExpressionTemplateToken(const Location & location, SpaceHandling pre, SpaceHandling post, std::shared_ptr && e) : TemplateToken(Type::Expression, location, pre, post), expr(std::move(e)) {} + ExpressionTemplateToken(const Location & loc, SpaceHandling pre, SpaceHandling post, std::shared_ptr && e) : TemplateToken(Type::Expression, loc, pre, post), expr(std::move(e)) {} }; struct IfTemplateToken : public TemplateToken { std::shared_ptr condition; - IfTemplateToken(const Location & location, SpaceHandling pre, SpaceHandling post, std::shared_ptr && c) : TemplateToken(Type::If, location, pre, post), condition(std::move(c)) {} + IfTemplateToken(const Location & loc, SpaceHandling pre, SpaceHandling post, std::shared_ptr && c) : TemplateToken(Type::If, loc, pre, post), condition(std::move(c)) {} }; struct ElifTemplateToken : public TemplateToken { std::shared_ptr condition; - ElifTemplateToken(const Location & location, SpaceHandling pre, SpaceHandling post, std::shared_ptr && c) : TemplateToken(Type::Elif, location, pre, post), condition(std::move(c)) {} + ElifTemplateToken(const Location & loc, SpaceHandling pre, SpaceHandling post, std::shared_ptr && c) : TemplateToken(Type::Elif, loc, pre, post), condition(std::move(c)) {} }; struct ElseTemplateToken : public TemplateToken { - ElseTemplateToken(const Location & location, SpaceHandling pre, SpaceHandling post) : TemplateToken(Type::Else, location, pre, post) {} + ElseTemplateToken(const Location & loc, SpaceHandling pre, SpaceHandling post) : TemplateToken(Type::Else, loc, pre, post) {} }; struct EndIfTemplateToken : public TemplateToken { - EndIfTemplateToken(const Location & location, SpaceHandling pre, SpaceHandling post) : TemplateToken(Type::EndIf, location, pre, post) {} + EndIfTemplateToken(const Location & loc, SpaceHandling pre, SpaceHandling post) : TemplateToken(Type::EndIf, loc, pre, post) {} }; struct MacroTemplateToken : public TemplateToken { std::shared_ptr name; Expression::Parameters params; - MacroTemplateToken(const Location & location, SpaceHandling pre, SpaceHandling post, std::shared_ptr && n, Expression::Parameters && p) - : TemplateToken(Type::Macro, location, pre, post), name(std::move(n)), params(std::move(p)) {} + MacroTemplateToken(const Location & loc, SpaceHandling pre, SpaceHandling post, std::shared_ptr && n, Expression::Parameters && p) + : TemplateToken(Type::Macro, loc, pre, post), name(std::move(n)), params(std::move(p)) {} }; struct EndMacroTemplateToken : public TemplateToken { - EndMacroTemplateToken(const Location & location, SpaceHandling pre, SpaceHandling post) : TemplateToken(Type::EndMacro, location, pre, post) {} + EndMacroTemplateToken(const Location & loc, SpaceHandling pre, SpaceHandling post) : TemplateToken(Type::EndMacro, loc, pre, post) {} }; struct FilterTemplateToken : public TemplateToken { std::shared_ptr filter; - FilterTemplateToken(const Location & location, SpaceHandling pre, SpaceHandling post, std::shared_ptr && filter) - : TemplateToken(Type::Filter, location, pre, post), filter(std::move(filter)) {} + FilterTemplateToken(const Location & loc, SpaceHandling pre, SpaceHandling post, std::shared_ptr && filter) + : TemplateToken(Type::Filter, loc, pre, post), filter(std::move(filter)) {} }; struct EndFilterTemplateToken : public TemplateToken { - EndFilterTemplateToken(const Location & location, SpaceHandling pre, SpaceHandling post) : TemplateToken(Type::EndFilter, location, pre, post) {} + EndFilterTemplateToken(const Location & loc, SpaceHandling pre, SpaceHandling post) : TemplateToken(Type::EndFilter, loc, pre, post) {} }; struct ForTemplateToken : public TemplateToken { @@ -783,38 +795,38 @@ struct ForTemplateToken : public TemplateToken { std::shared_ptr iterable; std::shared_ptr condition; bool recursive; - ForTemplateToken(const Location & location, SpaceHandling pre, SpaceHandling post, const std::vector & vns, std::shared_ptr && iter, + ForTemplateToken(const Location & loc, SpaceHandling pre, SpaceHandling post, const std::vector & vns, std::shared_ptr && iter, std::shared_ptr && c, bool r) - : TemplateToken(Type::For, location, pre, post), var_names(vns), iterable(std::move(iter)), condition(std::move(c)), recursive(r) {} + : TemplateToken(Type::For, loc, pre, post), var_names(vns), iterable(std::move(iter)), condition(std::move(c)), recursive(r) {} }; struct EndForTemplateToken : public TemplateToken { - EndForTemplateToken(const Location & location, SpaceHandling pre, SpaceHandling post) : TemplateToken(Type::EndFor, location, pre, post) {} + EndForTemplateToken(const Location & loc, SpaceHandling pre, SpaceHandling post) : TemplateToken(Type::EndFor, loc, pre, post) {} }; struct GenerationTemplateToken : public TemplateToken { - GenerationTemplateToken(const Location & location, SpaceHandling pre, SpaceHandling post) : TemplateToken(Type::Generation, location, pre, post) {} + GenerationTemplateToken(const Location & loc, SpaceHandling pre, SpaceHandling post) : TemplateToken(Type::Generation, loc, pre, post) {} }; struct EndGenerationTemplateToken : public TemplateToken { - EndGenerationTemplateToken(const Location & location, SpaceHandling pre, SpaceHandling post) : TemplateToken(Type::EndGeneration, location, pre, post) {} + EndGenerationTemplateToken(const Location & loc, SpaceHandling pre, SpaceHandling post) : TemplateToken(Type::EndGeneration, loc, pre, post) {} }; struct SetTemplateToken : public TemplateToken { std::string ns; std::vector var_names; std::shared_ptr value; - SetTemplateToken(const Location & location, SpaceHandling pre, SpaceHandling post, const std::string & ns, const std::vector & vns, std::shared_ptr && v) - : TemplateToken(Type::Set, location, pre, post), ns(ns), var_names(vns), value(std::move(v)) {} + SetTemplateToken(const Location & loc, SpaceHandling pre, SpaceHandling post, const std::string & ns, const std::vector & vns, std::shared_ptr && v) + : TemplateToken(Type::Set, loc, pre, post), ns(ns), var_names(vns), value(std::move(v)) {} }; struct EndSetTemplateToken : public TemplateToken { - EndSetTemplateToken(const Location & location, SpaceHandling pre, SpaceHandling post) : TemplateToken(Type::EndSet, location, pre, post) {} + EndSetTemplateToken(const Location & loc, SpaceHandling pre, SpaceHandling post) : TemplateToken(Type::EndSet, loc, pre, post) {} }; struct CommentTemplateToken : public TemplateToken { std::string text; - CommentTemplateToken(const Location & location, SpaceHandling pre, SpaceHandling post, const std::string& t) : TemplateToken(Type::Comment, location, pre, post), text(t) {} + CommentTemplateToken(const Location & loc, SpaceHandling pre, SpaceHandling post, const std::string& t) : TemplateToken(Type::Comment, loc, pre, post), text(t) {} }; enum class LoopControlType { Break, Continue }; @@ -830,7 +842,7 @@ public: struct LoopControlTemplateToken : public TemplateToken { LoopControlType control_type; - LoopControlTemplateToken(const Location & location, SpaceHandling pre, SpaceHandling post, LoopControlType control_type) : TemplateToken(Type::Break, location, pre, post), control_type(control_type) {} + LoopControlTemplateToken(const Location & loc, SpaceHandling pre, SpaceHandling post, LoopControlType control_type) : TemplateToken(Type::Break, loc, pre, post), control_type(control_type) {} }; class TemplateNode { @@ -868,8 +880,8 @@ public: class SequenceNode : public TemplateNode { std::vector> children; public: - SequenceNode(const Location & location, std::vector> && c) - : TemplateNode(location), children(std::move(c)) {} + SequenceNode(const Location & loc, std::vector> && c) + : TemplateNode(loc), children(std::move(c)) {} void do_render(std::ostringstream & out, const std::shared_ptr & context) const override { for (const auto& child : children) child->render(out, context); } @@ -878,7 +890,7 @@ public: class TextNode : public TemplateNode { std::string text; public: - TextNode(const Location & location, const std::string& t) : TemplateNode(location), text(t) {} + TextNode(const Location & loc, const std::string& t) : TemplateNode(loc), text(t) {} void do_render(std::ostringstream & out, const std::shared_ptr &) const override { out << text; } @@ -887,7 +899,7 @@ public: class ExpressionNode : public TemplateNode { std::shared_ptr expr; public: - ExpressionNode(const Location & location, std::shared_ptr && e) : TemplateNode(location), expr(std::move(e)) {} + ExpressionNode(const Location & loc, std::shared_ptr && e) : TemplateNode(loc), expr(std::move(e)) {} void do_render(std::ostringstream & out, const std::shared_ptr & context) const override { if (!expr) throw std::runtime_error("ExpressionNode.expr is null"); auto result = expr->evaluate(context); @@ -904,8 +916,8 @@ public: class IfNode : public TemplateNode { std::vector, std::shared_ptr>> cascade; public: - IfNode(const Location & location, std::vector, std::shared_ptr>> && c) - : TemplateNode(location), cascade(std::move(c)) {} + IfNode(const Location & loc, std::vector, std::shared_ptr>> && c) + : TemplateNode(loc), cascade(std::move(c)) {} void do_render(std::ostringstream & out, const std::shared_ptr & context) const override { for (const auto& branch : cascade) { auto enter_branch = true; @@ -924,7 +936,7 @@ public: class LoopControlNode : public TemplateNode { LoopControlType control_type_; public: - LoopControlNode(const Location & location, LoopControlType control_type) : TemplateNode(location), control_type_(control_type) {} + LoopControlNode(const Location & loc, LoopControlType control_type) : TemplateNode(loc), control_type_(control_type) {} void do_render(std::ostringstream &, const std::shared_ptr &) const override { throw LoopControlException(control_type_); } @@ -938,9 +950,9 @@ class ForNode : public TemplateNode { bool recursive; std::shared_ptr else_body; public: - ForNode(const Location & location, std::vector && var_names, std::shared_ptr && iterable, + ForNode(const Location & loc, std::vector && var_names, std::shared_ptr && iterable, std::shared_ptr && condition, std::shared_ptr && body, bool recursive, std::shared_ptr && else_body) - : TemplateNode(location), var_names(var_names), iterable(std::move(iterable)), condition(std::move(condition)), body(std::move(body)), recursive(recursive), else_body(std::move(else_body)) {} + : TemplateNode(loc), var_names(var_names), iterable(std::move(iterable)), condition(std::move(condition)), body(std::move(body)), recursive(recursive), else_body(std::move(else_body)) {} void do_render(std::ostringstream & out, const std::shared_ptr & context) const override { // https://jinja.palletsprojects.com/en/3.0.x/templates/#for @@ -1025,8 +1037,8 @@ class MacroNode : public TemplateNode { std::shared_ptr body; std::unordered_map named_param_positions; public: - MacroNode(const Location & location, std::shared_ptr && n, Expression::Parameters && p, std::shared_ptr && b) - : TemplateNode(location), name(std::move(n)), params(std::move(p)), body(std::move(b)) { + MacroNode(const Location & loc, std::shared_ptr && n, Expression::Parameters && p, std::shared_ptr && b) + : TemplateNode(loc), name(std::move(n)), params(std::move(p)), body(std::move(b)) { for (size_t i = 0; i < params.size(); ++i) { const auto & name = params[i].first; if (!name.empty()) { @@ -1072,8 +1084,8 @@ class FilterNode : public TemplateNode { std::shared_ptr body; public: - FilterNode(const Location & location, std::shared_ptr && f, std::shared_ptr && b) - : TemplateNode(location), filter(std::move(f)), body(std::move(b)) {} + FilterNode(const Location & loc, std::shared_ptr && f, std::shared_ptr && b) + : TemplateNode(loc), filter(std::move(f)), body(std::move(b)) {} void do_render(std::ostringstream & out, const std::shared_ptr & context) const override { if (!filter) throw std::runtime_error("FilterNode.filter is null"); @@ -1095,8 +1107,8 @@ class SetNode : public TemplateNode { std::vector var_names; std::shared_ptr value; public: - SetNode(const Location & location, const std::string & ns, const std::vector & vns, std::shared_ptr && v) - : TemplateNode(location), ns(ns), var_names(vns), value(std::move(v)) {} + SetNode(const Location & loc, const std::string & ns, const std::vector & vns, std::shared_ptr && v) + : TemplateNode(loc), ns(ns), var_names(vns), value(std::move(v)) {} void do_render(std::ostringstream &, const std::shared_ptr & context) const override { if (!value) throw std::runtime_error("SetNode.value is null"); if (!ns.empty()) { @@ -1118,8 +1130,8 @@ class SetTemplateNode : public TemplateNode { std::string name; std::shared_ptr template_value; public: - SetTemplateNode(const Location & location, const std::string & name, std::shared_ptr && tv) - : TemplateNode(location), name(name), template_value(std::move(tv)) {} + SetTemplateNode(const Location & loc, const std::string & name, std::shared_ptr && tv) + : TemplateNode(loc), name(name), template_value(std::move(tv)) {} void do_render(std::ostringstream &, const std::shared_ptr & context) const override { if (!template_value) throw std::runtime_error("SetTemplateNode.template_value is null"); Value value { template_value->render(context) }; @@ -1132,8 +1144,8 @@ class IfExpr : public Expression { std::shared_ptr then_expr; std::shared_ptr else_expr; public: - IfExpr(const Location & location, std::shared_ptr && c, std::shared_ptr && t, std::shared_ptr && e) - : Expression(location), condition(std::move(c)), then_expr(std::move(t)), else_expr(std::move(e)) {} + IfExpr(const Location & loc, std::shared_ptr && c, std::shared_ptr && t, std::shared_ptr && e) + : Expression(loc), condition(std::move(c)), then_expr(std::move(t)), else_expr(std::move(e)) {} Value do_evaluate(const std::shared_ptr & context) const override { if (!condition) throw std::runtime_error("IfExpr.condition is null"); if (!then_expr) throw std::runtime_error("IfExpr.then_expr is null"); @@ -1150,16 +1162,16 @@ public: class LiteralExpr : public Expression { Value value; public: - LiteralExpr(const Location & location, const Value& v) - : Expression(location), value(v) {} + LiteralExpr(const Location & loc, const Value& v) + : Expression(loc), value(v) {} Value do_evaluate(const std::shared_ptr &) const override { return value; } }; class ArrayExpr : public Expression { std::vector> elements; public: - ArrayExpr(const Location & location, std::vector> && e) - : Expression(location), elements(std::move(e)) {} + ArrayExpr(const Location & loc, std::vector> && e) + : Expression(loc), elements(std::move(e)) {} Value do_evaluate(const std::shared_ptr & context) const override { auto result = Value::array(); for (const auto& e : elements) { @@ -1173,8 +1185,8 @@ public: class DictExpr : public Expression { std::vector, std::shared_ptr>> elements; public: - DictExpr(const Location & location, std::vector, std::shared_ptr>> && e) - : Expression(location), elements(std::move(e)) {} + DictExpr(const Location & loc, std::vector, std::shared_ptr>> && e) + : Expression(loc), elements(std::move(e)) {} Value do_evaluate(const std::shared_ptr & context) const override { auto result = Value::object(); for (const auto& [key, value] : elements) { @@ -1189,8 +1201,8 @@ public: class SliceExpr : public Expression { public: std::shared_ptr start, end; - SliceExpr(const Location & location, std::shared_ptr && s, std::shared_ptr && e) - : Expression(location), start(std::move(s)), end(std::move(e)) {} + SliceExpr(const Location & loc, std::shared_ptr && s, std::shared_ptr && e) + : Expression(loc), start(std::move(s)), end(std::move(e)) {} Value do_evaluate(const std::shared_ptr &) const override { throw std::runtime_error("SliceExpr not implemented"); } @@ -1200,8 +1212,8 @@ class SubscriptExpr : public Expression { std::shared_ptr base; std::shared_ptr index; public: - SubscriptExpr(const Location & location, std::shared_ptr && b, std::shared_ptr && i) - : Expression(location), base(std::move(b)), index(std::move(i)) {} + SubscriptExpr(const Location & loc, std::shared_ptr && b, std::shared_ptr && i) + : Expression(loc), base(std::move(b)), index(std::move(i)) {} Value do_evaluate(const std::shared_ptr & context) const override { if (!base) throw std::runtime_error("SubscriptExpr.base is null"); if (!index) throw std::runtime_error("SubscriptExpr.index is null"); @@ -1243,8 +1255,8 @@ public: enum class Op { Plus, Minus, LogicalNot, Expansion, ExpansionDict }; std::shared_ptr expr; Op op; - UnaryOpExpr(const Location & location, std::shared_ptr && e, Op o) - : Expression(location), expr(std::move(e)), op(o) {} + UnaryOpExpr(const Location & loc, std::shared_ptr && e, Op o) + : Expression(loc), expr(std::move(e)), op(o) {} Value do_evaluate(const std::shared_ptr & context) const override { if (!expr) throw std::runtime_error("UnaryOpExpr.expr is null"); auto e = expr->evaluate(context); @@ -1269,8 +1281,8 @@ private: std::shared_ptr right; Op op; public: - BinaryOpExpr(const Location & location, std::shared_ptr && l, std::shared_ptr && r, Op o) - : Expression(location), left(std::move(l)), right(std::move(r)), op(o) {} + BinaryOpExpr(const Location & loc, std::shared_ptr && l, std::shared_ptr && r, Op o) + : Expression(loc), left(std::move(l)), right(std::move(r)), op(o) {} Value do_evaluate(const std::shared_ptr & context) const override { if (!left) throw std::runtime_error("BinaryOpExpr.left is null"); if (!right) throw std::runtime_error("BinaryOpExpr.right is null"); @@ -1378,13 +1390,34 @@ struct ArgumentsExpression { } }; -static std::string strip(const std::string & s) { - auto start = s.find_first_not_of(" \t\n\r"); +static std::string strip(const std::string & s, const std::string & chars = "", bool left = true, bool right = true) { + auto charset = chars.empty() ? " \t\n\r" : chars; + auto start = left ? s.find_first_not_of(charset) : 0; if (start == std::string::npos) return ""; - auto end = s.find_last_not_of(" \t\n\r"); + auto end = right ? s.find_last_not_of(charset) : s.size() - 1; return s.substr(start, end - start + 1); } +static std::vector split(const std::string & s, const std::string & sep) { + std::vector result; + size_t start = 0; + size_t end = s.find(sep); + while (end != std::string::npos) { + result.push_back(s.substr(start, end - start)); + start = end + sep.length(); + end = s.find(sep, start); + } + result.push_back(s.substr(start)); + return result; +} + +static std::string capitalize(const std::string & s) { + if (s.empty()) return s; + auto result = s; + result[0] = std::toupper(result[0]); + return result; +} + static std::string html_escape(const std::string & s) { std::string result; result.reserve(s.size()); @@ -1406,8 +1439,8 @@ class MethodCallExpr : public Expression { std::shared_ptr method; ArgumentsExpression args; public: - MethodCallExpr(const Location & location, std::shared_ptr && obj, std::shared_ptr && m, ArgumentsExpression && a) - : Expression(location), object(std::move(obj)), method(std::move(m)), args(std::move(a)) {} + MethodCallExpr(const Location & loc, std::shared_ptr && obj, std::shared_ptr && m, ArgumentsExpression && a) + : Expression(loc), object(std::move(obj)), method(std::move(m)), args(std::move(a)) {} Value do_evaluate(const std::shared_ptr & context) const override { if (!object) throw std::runtime_error("MethodCallExpr.object is null"); if (!method) throw std::runtime_error("MethodCallExpr.method is null"); @@ -1460,8 +1493,29 @@ public: } else if (obj.is_string()) { auto str = obj.get(); if (method->get_name() == "strip") { - vargs.expectArgs("strip method", {0, 0}, {0, 0}); - return Value(strip(str)); + vargs.expectArgs("strip method", {0, 1}, {0, 0}); + auto chars = vargs.args.empty() ? "" : vargs.args[0].get(); + return Value(strip(str, chars)); + } else if (method->get_name() == "lstrip") { + vargs.expectArgs("lstrip method", {0, 1}, {0, 0}); + auto chars = vargs.args.empty() ? "" : vargs.args[0].get(); + return Value(strip(str, chars, /* left= */ true, /* right= */ false)); + } else if (method->get_name() == "rstrip") { + vargs.expectArgs("rstrip method", {0, 1}, {0, 0}); + auto chars = vargs.args.empty() ? "" : vargs.args[0].get(); + return Value(strip(str, chars, /* left= */ false, /* right= */ true)); + } else if (method->get_name() == "split") { + vargs.expectArgs("split method", {1, 1}, {0, 0}); + auto sep = vargs.args[0].get(); + auto parts = split(str, sep); + Value result = Value::array(); + for (const auto& part : parts) { + result.push_back(Value(part)); + } + return result; + } else if (method->get_name() == "capitalize") { + vargs.expectArgs("capitalize method", {0, 0}, {0, 0}); + return Value(capitalize(str)); } else if (method->get_name() == "endswith") { vargs.expectArgs("endswith method", {1, 1}, {0, 0}); auto suffix = vargs.args[0].get(); @@ -1484,8 +1538,8 @@ class CallExpr : public Expression { public: std::shared_ptr object; ArgumentsExpression args; - CallExpr(const Location & location, std::shared_ptr && obj, ArgumentsExpression && a) - : Expression(location), object(std::move(obj)), args(std::move(a)) {} + CallExpr(const Location & loc, std::shared_ptr && obj, ArgumentsExpression && a) + : Expression(loc), object(std::move(obj)), args(std::move(a)) {} Value do_evaluate(const std::shared_ptr & context) const override { if (!object) throw std::runtime_error("CallExpr.object is null"); auto obj = object->evaluate(context); @@ -1500,8 +1554,8 @@ public: class FilterExpr : public Expression { std::vector> parts; public: - FilterExpr(const Location & location, std::vector> && p) - : Expression(location), parts(std::move(p)) {} + FilterExpr(const Location & loc, std::vector> && p) + : Expression(loc), parts(std::move(p)) {} Value do_evaluate(const std::shared_ptr & context) const override { Value result; bool first = true; @@ -1792,7 +1846,7 @@ private: auto left = parseStringConcat(); if (!left) throw std::runtime_error("Expected left side of 'logical compare' expression"); - static std::regex compare_tok(R"(==|!=|<=?|>=?|in\b|is\b|not[\r\n\s]+in\b)"); + static std::regex compare_tok(R"(==|!=|<=?|>=?|in\b|is\b|not\s+in\b)"); static std::regex not_tok(R"(not\b)"); std::string op_str; while (!(op_str = consumeToken(compare_tok)).empty()) { @@ -2171,7 +2225,7 @@ private: using TemplateTokenIterator = TemplateTokenVector::const_iterator; std::vector parseVarNames() { - static std::regex varnames_regex(R"(((?:\w+)(?:[\r\n\s]*,[\r\n\s]*(?:\w+))*)[\r\n\s]*)"); + static std::regex varnames_regex(R"(((?:\w+)(?:\s*,\s*(?:\w+))*)\s*)"); std::vector group; if ((group = consumeTokenGroups(varnames_regex)).empty()) throw std::runtime_error("Expected variable names"); @@ -2194,13 +2248,13 @@ private: } TemplateTokenVector tokenize() { - static std::regex comment_tok(R"(\{#([-~]?)([\s\S\r\n]*?)([-~]?)#\})"); + static std::regex comment_tok(R"(\{#([-~]?)([\s\S]*?)([-~]?)#\})"); static std::regex expr_open_regex(R"(\{\{([-~])?)"); - static std::regex block_open_regex(R"(^\{%([-~])?[\s\n\r]*)"); + static std::regex block_open_regex(R"(^\{%([-~])?\s*)"); static std::regex block_keyword_tok(R"((if|else|elif|endif|for|endfor|generation|endgeneration|set|endset|block|endblock|macro|endmacro|filter|endfilter|break|continue)\b)"); static std::regex non_text_open_regex(R"(\{\{|\{%|\{#)"); - static std::regex expr_close_regex(R"([\s\n\r]*([-~])?\}\})"); - static std::regex block_close_regex(R"([\s\n\r]*([-~])?%\})"); + static std::regex expr_close_regex(R"(\s*([-~])?\}\})"); + static std::regex block_close_regex(R"(\s*([-~])?%\})"); TemplateTokenVector tokens; std::vector group; @@ -2284,7 +2338,7 @@ private: auto post_space = parseBlockClose(); tokens.push_back(std::make_unique(location, pre_space, post_space)); } else if (keyword == "set") { - static std::regex namespaced_var_regex(R"((\w+)[\s\n\r]*\.[\s\n\r]*(\w+))"); + static std::regex namespaced_var_regex(R"((\w+)\s*\.\s*(\w+))"); std::string ns; std::vector var_names; @@ -2336,6 +2390,11 @@ private: throw std::runtime_error("Unexpected block: " + keyword); } } else if (std::regex_search(it, end, match, non_text_open_regex)) { + if (!match.position()) { + if (match[0] != "{#") + throw std::runtime_error("Internal error: Expected a comment"); + throw std::runtime_error("Missing end of comment tag"); + } auto text_end = it + match.position(); text = std::string(it, text_end); it = text_end; @@ -2400,7 +2459,7 @@ private: auto text = text_token->text; if (post_space == SpaceHandling::Strip) { - static std::regex trailing_space_regex(R"((\s|\r|\n)+$)"); + static std::regex trailing_space_regex(R"(\s+$)"); text = std::regex_replace(text, trailing_space_regex, ""); } else if (options.lstrip_blocks && it != end) { auto i = text.size(); @@ -2410,10 +2469,10 @@ private: } } if (pre_space == SpaceHandling::Strip) { - static std::regex leading_space_regex(R"(^(\s|\r|\n)+)"); + static std::regex leading_space_regex(R"(^\s+)"); text = std::regex_replace(text, leading_space_regex, ""); } else if (options.trim_blocks && (it - 1) != begin && !dynamic_cast((*(it - 2)).get())) { - if (text.length() > 0 && text[0] == '\n') { + if (!text.empty() && text[0] == '\n') { text.erase(0, 1); } } @@ -2491,7 +2550,7 @@ public: TemplateTokenIterator begin = tokens.begin(); auto it = begin; TemplateTokenIterator end = tokens.end(); - return parser.parseTemplate(begin, it, end, /* full= */ true); + return parser.parseTemplate(begin, it, end, /* fully= */ true); } }; @@ -2530,7 +2589,7 @@ inline std::shared_ptr Context::builtins() { throw std::runtime_error(args.at("message").get()); })); globals.set("tojson", simple_function("tojson", { "value", "indent" }, [](const std::shared_ptr &, Value & args) { - return Value(args.at("value").dump(args.get("indent", -1), /* tojson= */ true)); + return Value(args.at("value").dump(args.get("indent", -1), /* to_json= */ true)); })); globals.set("items", simple_function("items", { "object" }, [](const std::shared_ptr &, Value & args) { auto items = Value::array(); @@ -2552,21 +2611,25 @@ inline std::shared_ptr Context::builtins() { globals.set("last", simple_function("last", { "items" }, [](const std::shared_ptr &, Value & args) { auto items = args.at("items"); if (!items.is_array()) throw std::runtime_error("object is not a list"); - if (items.size() == 0) return Value(); + if (items.empty()) return Value(); return items.at(items.size() - 1); })); globals.set("trim", simple_function("trim", { "text" }, [](const std::shared_ptr &, Value & args) { auto & text = args.at("text"); return text.is_null() ? text : Value(strip(text.get())); })); - globals.set("lower", simple_function("lower", { "text" }, [](const std::shared_ptr &, Value & args) { - auto text = args.at("text"); - if (text.is_null()) return text; - std::string res; - auto str = text.get(); - std::transform(str.begin(), str.end(), std::back_inserter(res), ::tolower); - return Value(res); - })); + auto char_transform_function = [](const std::string & name, const std::function & fn) { + return simple_function(name, { "text" }, [=](const std::shared_ptr &, Value & args) { + auto text = args.at("text"); + if (text.is_null()) return text; + std::string res; + auto str = text.get(); + std::transform(str.begin(), str.end(), std::back_inserter(res), fn); + return Value(res); + }); + }; + globals.set("lower", char_transform_function("lower", ::tolower)); + globals.set("upper", char_transform_function("upper", ::toupper)); globals.set("default", Value::callable([=](const std::shared_ptr &, ArgumentsValue & args) { args.expectArgs("default", {2, 3}, {0, 1}); auto & value = args.args[0]; @@ -2696,12 +2759,17 @@ inline std::shared_ptr Context::builtins() { return Value::callable([=](const std::shared_ptr & context, ArgumentsValue & args) { args.expectArgs(is_select ? "select" : "reject", {2, (std::numeric_limits::max)()}, {0, 0}); auto & items = args.args[0]; - if (items.is_null()) + if (items.is_null()) { return Value::array(); - if (!items.is_array()) throw std::runtime_error("object is not iterable: " + items.dump()); + } + if (!items.is_array()) { + throw std::runtime_error("object is not iterable: " + items.dump()); + } auto filter_fn = context->get(args.args[1]); - if (filter_fn.is_null()) throw std::runtime_error("Undefined filter: " + args.args[1].dump()); + if (filter_fn.is_null()) { + throw std::runtime_error("Undefined filter: " + args.args[1].dump()); + } auto filter_args = Value::array(); for (size_t i = 2, n = args.args.size(); i < n; i++) { @@ -2823,20 +2891,25 @@ inline std::shared_ptr Context::builtins() { auto v = arg.get(); startEndStep[i] = v; param_set[i] = true; - } } - for (auto & [name, value] : args.kwargs) { - size_t i; - if (name == "start") i = 0; - else if (name == "end") i = 1; - else if (name == "step") i = 2; - else throw std::runtime_error("Unknown argument " + name + " for function range"); + } + for (auto & [name, value] : args.kwargs) { + size_t i; + if (name == "start") { + i = 0; + } else if (name == "end") { + i = 1; + } else if (name == "step") { + i = 2; + } else { + throw std::runtime_error("Unknown argument " + name + " for function range"); + } - if (param_set[i]) { - throw std::runtime_error("Duplicate argument " + name + " for function range"); - } - startEndStep[i] = value.get(); - param_set[i] = true; + if (param_set[i]) { + throw std::runtime_error("Duplicate argument " + name + " for function range"); + } + startEndStep[i] = value.get(); + param_set[i] = true; } if (!param_set[1]) { throw std::runtime_error("Missing required argument 'end' for function range"); diff --git a/common/ngram-cache.cpp b/common/ngram-cache.cpp index a057ae45f..d1a4d84c4 100644 --- a/common/ngram-cache.cpp +++ b/common/ngram-cache.cpp @@ -7,6 +7,7 @@ #include #include #include +#include void common_ngram_cache_update(common_ngram_cache & ngram_cache, int ngram_min, int ngram_max, std::vector & inp, int nnew, bool print_progress) { diff --git a/common/sampling.cpp b/common/sampling.cpp index e4b21ca10..1735b6501 100644 --- a/common/sampling.cpp +++ b/common/sampling.cpp @@ -4,6 +4,7 @@ #include #include +#include // the ring buffer works similarly to std::deque, but with a fixed capacity // TODO: deduplicate with llama-impl.h @@ -134,11 +135,11 @@ std::string common_params_sampling::print() const { snprintf(result, sizeof(result), "\trepeat_last_n = %d, repeat_penalty = %.3f, frequency_penalty = %.3f, presence_penalty = %.3f\n" "\tdry_multiplier = %.3f, dry_base = %.3f, dry_allowed_length = %d, dry_penalty_last_n = %d\n" - "\ttop_k = %d, top_p = %.3f, min_p = %.3f, xtc_probability = %.3f, xtc_threshold = %.3f, typical_p = %.3f, temp = %.3f\n" + "\ttop_k = %d, top_p = %.3f, min_p = %.3f, xtc_probability = %.3f, xtc_threshold = %.3f, typical_p = %.3f, top_n_sigma = %.3f, temp = %.3f\n" "\tmirostat = %d, mirostat_lr = %.3f, mirostat_ent = %.3f", penalty_last_n, penalty_repeat, penalty_freq, penalty_present, dry_multiplier, dry_base, dry_allowed_length, dry_penalty_last_n, - top_k, top_p, min_p, xtc_probability, xtc_threshold, typ_p, temp, + top_k, top_p, min_p, xtc_probability, xtc_threshold, typ_p, top_n_sigma, temp, mirostat, mirostat_eta, mirostat_tau); return std::string(result); @@ -151,12 +152,6 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, co lparams.no_perf = params.no_perf; - std::vector trigger_words; - trigger_words.reserve(params.grammar_trigger_words.size()); - for (const auto & str : params.grammar_trigger_words) { - trigger_words.push_back(str.word.c_str()); - } - struct llama_sampler * grmr; if (params.grammar.compare(0, 11, "%llguidance") == 0) { #ifdef LLAMA_USE_LLGUIDANCE @@ -165,11 +160,57 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, co GGML_ABORT("llguidance (cmake -DLLAMA_LLGUIDANCE=ON) is not enabled"); #endif // LLAMA_USE_LLGUIDANCE } else { + std::vector patterns_at_start; + std::vector patterns_anywhere; + std::vector trigger_tokens; + for (const auto & trigger : params.grammar_triggers) { + switch (trigger.type) { + case COMMON_GRAMMAR_TRIGGER_TYPE_WORD: + { + const auto & word = trigger.value; + patterns_anywhere.push_back(regex_escape(word)); + break; + } + case COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN: + case COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN_START: + { + const auto & pattern = trigger.value; + (trigger.type == COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN_START ? patterns_at_start : patterns_anywhere).push_back(pattern); + break; + } + case COMMON_GRAMMAR_TRIGGER_TYPE_TOKEN: + { + const auto token = trigger.token; + trigger_tokens.push_back(token); + break; + } + default: + GGML_ASSERT(false && "unknown trigger type"); + } + } + + std::vector trigger_patterns; + if (!patterns_at_start.empty()) { + trigger_patterns.push_back("^(" + string_join(patterns_at_start, "|") + ")[\\s\\S]*"); + } + if (!patterns_anywhere.empty()) { + trigger_patterns.push_back("^[\\s\\S]*?(" + string_join(patterns_anywhere, "|") + ")[\\s\\S]*"); + } + + std::vector trigger_patterns_c; + trigger_patterns_c.reserve(trigger_patterns.size()); + for (const auto & regex : trigger_patterns) { + trigger_patterns_c.push_back(regex.c_str()); + } + grmr = params.grammar_lazy - ? llama_sampler_init_grammar_lazy(vocab, params.grammar.c_str(), "root", - trigger_words.data(), trigger_words.size(), - params.grammar_trigger_tokens.data(), params.grammar_trigger_tokens.size()) + ? llama_sampler_init_grammar_lazy_patterns(vocab, params.grammar.c_str(), "root", + trigger_patterns_c.data(), trigger_patterns_c.size(), + trigger_tokens.data(), trigger_tokens.size()) : llama_sampler_init_grammar(vocab, params.grammar.c_str(), "root"); + if (!grmr) { + return nullptr; + } } auto * result = new common_sampler { @@ -188,45 +229,51 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, co params.logit_bias.data())); if (params.mirostat == 0) { - for (const auto & cnstr : params.samplers) { - switch (cnstr) { - case COMMON_SAMPLER_TYPE_DRY: - { - std::vector c_breakers; - c_breakers.reserve(params.dry_sequence_breakers.size()); - for (const auto & str : params.dry_sequence_breakers) { - c_breakers.push_back(str.c_str()); - } + if (params.top_n_sigma >= 0) { + llama_sampler_chain_add(result->chain, llama_sampler_init_top_k (params.top_k)); + llama_sampler_chain_add(result->chain, llama_sampler_init_temp (params.temp)); + llama_sampler_chain_add(result->chain, llama_sampler_init_top_n_sigma (params.top_n_sigma)); + } else { + for (const auto & cnstr : params.samplers) { + switch (cnstr) { + case COMMON_SAMPLER_TYPE_DRY: + { + std::vector c_breakers; + c_breakers.reserve(params.dry_sequence_breakers.size()); + for (const auto & str : params.dry_sequence_breakers) { + c_breakers.push_back(str.c_str()); + } - llama_sampler_chain_add(result->chain, llama_sampler_init_dry (vocab, llama_model_n_ctx_train(model), params.dry_multiplier, params.dry_base, params.dry_allowed_length, params.dry_penalty_last_n, c_breakers.data(), c_breakers.size())); - } - break; - case COMMON_SAMPLER_TYPE_TOP_K: - llama_sampler_chain_add(result->chain, llama_sampler_init_top_k (params.top_k)); - break; - case COMMON_SAMPLER_TYPE_TOP_P: - llama_sampler_chain_add(result->chain, llama_sampler_init_top_p (params.top_p, params.min_keep)); - break; - case COMMON_SAMPLER_TYPE_MIN_P: - llama_sampler_chain_add(result->chain, llama_sampler_init_min_p (params.min_p, params.min_keep)); - break; - case COMMON_SAMPLER_TYPE_XTC: - llama_sampler_chain_add(result->chain, llama_sampler_init_xtc (params.xtc_probability, params.xtc_threshold, params.min_keep, params.seed)); - break; - case COMMON_SAMPLER_TYPE_TYPICAL_P: - llama_sampler_chain_add(result->chain, llama_sampler_init_typical (params.typ_p, params.min_keep)); - break; - case COMMON_SAMPLER_TYPE_TEMPERATURE: - llama_sampler_chain_add(result->chain, llama_sampler_init_temp_ext (params.temp, params.dynatemp_range, params.dynatemp_exponent)); - break; - case COMMON_SAMPLER_TYPE_INFILL: - llama_sampler_chain_add(result->chain, llama_sampler_init_infill (vocab)); - break; - case COMMON_SAMPLER_TYPE_PENALTIES: - llama_sampler_chain_add(result->chain, llama_sampler_init_penalties(params.penalty_last_n, params.penalty_repeat, params.penalty_freq, params.penalty_present)); - break; - default: - GGML_ASSERT(false && "unknown sampler type"); + llama_sampler_chain_add(result->chain, llama_sampler_init_dry (vocab, llama_model_n_ctx_train(model), params.dry_multiplier, params.dry_base, params.dry_allowed_length, params.dry_penalty_last_n, c_breakers.data(), c_breakers.size())); + } + break; + case COMMON_SAMPLER_TYPE_TOP_K: + llama_sampler_chain_add(result->chain, llama_sampler_init_top_k (params.top_k)); + break; + case COMMON_SAMPLER_TYPE_TOP_P: + llama_sampler_chain_add(result->chain, llama_sampler_init_top_p (params.top_p, params.min_keep)); + break; + case COMMON_SAMPLER_TYPE_MIN_P: + llama_sampler_chain_add(result->chain, llama_sampler_init_min_p (params.min_p, params.min_keep)); + break; + case COMMON_SAMPLER_TYPE_XTC: + llama_sampler_chain_add(result->chain, llama_sampler_init_xtc (params.xtc_probability, params.xtc_threshold, params.min_keep, params.seed)); + break; + case COMMON_SAMPLER_TYPE_TYPICAL_P: + llama_sampler_chain_add(result->chain, llama_sampler_init_typical (params.typ_p, params.min_keep)); + break; + case COMMON_SAMPLER_TYPE_TEMPERATURE: + llama_sampler_chain_add(result->chain, llama_sampler_init_temp_ext (params.temp, params.dynatemp_range, params.dynatemp_exponent)); + break; + case COMMON_SAMPLER_TYPE_INFILL: + llama_sampler_chain_add(result->chain, llama_sampler_init_infill (vocab)); + break; + case COMMON_SAMPLER_TYPE_PENALTIES: + llama_sampler_chain_add(result->chain, llama_sampler_init_penalties(params.penalty_last_n, params.penalty_repeat, params.penalty_freq, params.penalty_present)); + break; + default: + GGML_ASSERT(false && "unknown sampler type"); + } } } llama_sampler_chain_add(result->chain, llama_sampler_init_dist(params.seed)); diff --git a/common/speculative.cpp b/common/speculative.cpp index 318e96ea3..ccad70fa9 100644 --- a/common/speculative.cpp +++ b/common/speculative.cpp @@ -5,6 +5,7 @@ #include "sampling.h" #include +#include #define SPEC_VOCAB_MAX_SIZE_DIFFERENCE 128 #define SPEC_VOCAB_CHECK_START_TOKEN_ID 5 @@ -172,7 +173,7 @@ llama_tokens common_speculative_gen_draft( result.reserve(params.n_draft); if (reuse_n == 0) { - llama_kv_cache_clear(ctx); + llama_kv_self_clear(ctx); prompt.clear(); } else { @@ -191,14 +192,14 @@ llama_tokens common_speculative_gen_draft( } if (reuse_i > 0) { - llama_kv_cache_seq_rm (ctx, 0, 0, reuse_i); - llama_kv_cache_seq_add(ctx, 0, reuse_i, -1, -reuse_i); + llama_kv_self_seq_rm (ctx, 0, 0, reuse_i); + llama_kv_self_seq_add(ctx, 0, reuse_i, -1, -reuse_i); prompt.erase(prompt.begin(), prompt.begin() + reuse_i); } if (reuse_n < (int) prompt.size()) { - llama_kv_cache_seq_rm (ctx, 0, reuse_n, -1); + llama_kv_self_seq_rm (ctx, 0, reuse_n, -1); prompt.erase(prompt.begin() + reuse_n, prompt.end()); } @@ -252,11 +253,6 @@ llama_tokens common_speculative_gen_draft( // add drafted token for each sequence const llama_token id = cur_p->data[0].id; - // only collect very high-confidence draft tokens - if (cur_p->data[0].p < params.p_min) { - break; - } - common_sampler_accept(smpl, id, true); result.push_back(id); @@ -265,6 +261,11 @@ llama_tokens common_speculative_gen_draft( break; } + // only collect very high-confidence draft tokens + if (cur_p->data[0].p < params.p_min) { + break; + } + common_batch_add(batch, id, n_past + i + 1, { 0 }, true); // evaluate the drafted tokens on the draft model diff --git a/common/speculative.h b/common/speculative.h index 50ec03446..2b51a70ca 100644 --- a/common/speculative.h +++ b/common/speculative.h @@ -9,7 +9,7 @@ struct common_speculative_params { int n_draft = 16; // max drafted tokens int n_reuse = 256; - float p_min = 0.9f; // min probabiliy required to accept a token in the draft + float p_min = 0.75f; // min probability required to accept a token in the draft }; struct common_speculative * common_speculative_init(struct llama_context * ctx_dft); diff --git a/convert_hf_to_gguf.py b/convert_hf_to_gguf.py index 018a2a588..2bf97475f 100755 --- a/convert_hf_to_gguf.py +++ b/convert_hf_to_gguf.py @@ -65,6 +65,7 @@ class Model: model_name: str | None metadata_override: Path | None dir_model_card: Path + remote_hf_model_id: str | None # subclasses should define this! model_arch: gguf.MODEL_ARCH @@ -73,7 +74,7 @@ class Model: use_temp_file: bool = False, eager: bool = False, metadata_override: Path | None = None, model_name: str | None = None, split_max_tensors: int = 0, split_max_size: int = 0, dry_run: bool = False, - small_first_shard: bool = False, hparams: dict[str, Any] | None = None): + small_first_shard: bool = False, hparams: dict[str, Any] | None = None, remote_hf_model_id: str | None = None): if type(self) is Model: raise TypeError(f"{type(self).__name__!r} should not be directly instantiated") @@ -83,11 +84,24 @@ class Model: self.is_big_endian = is_big_endian self.endianess = gguf.GGUFEndian.BIG if is_big_endian else gguf.GGUFEndian.LITTLE self.use_temp_file = use_temp_file - self.lazy = not eager - self.part_names = Model.get_model_part_names(self.dir_model, "model", ".safetensors") - self.is_safetensors = len(self.part_names) > 0 - if not self.is_safetensors: - self.part_names = Model.get_model_part_names(self.dir_model, "pytorch_model", ".bin") + self.lazy = not eager or (remote_hf_model_id is not None) + self.remote_hf_model_id = remote_hf_model_id + if remote_hf_model_id is not None: + self.is_safetensors = True + + def get_remote_tensors() -> Iterator[tuple[str, Tensor]]: + logger.info(f"Using remote model with HuggingFace id: {remote_hf_model_id}") + remote_tensors = gguf.utility.SafetensorRemote.get_list_tensors_hf_model(remote_hf_model_id) + self.tensor_names = set(name for name in remote_tensors.keys()) + for name, remote_tensor in gguf.utility.SafetensorRemote.get_list_tensors_hf_model(remote_hf_model_id).items(): + yield (name, LazyTorchTensor.from_remote_tensor(remote_tensor)) + + self.get_tensors = get_remote_tensors + else: + self.part_names = Model.get_model_part_names(self.dir_model, "model", ".safetensors") + self.is_safetensors = len(self.part_names) > 0 + if not self.is_safetensors: + self.part_names = Model.get_model_part_names(self.dir_model, "pytorch_model", ".bin") self.hparams = Model.load_hparams(self.dir_model) if hparams is None else hparams self.block_count = self.find_hparam(["n_layers", "num_hidden_layers", "n_layer", "num_layers"]) self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count) @@ -180,7 +194,8 @@ class Model: extra = sorted(tensor_names_from_parts.difference(self.tensor_names)) missing_files = sorted(set(weight_map[n] for n in missing if n in weight_map)) if len(extra) == 0 and len(missing_files) > 0: - raise ValueError(f"Missing or incomplete model files: {missing_files}") + raise ValueError(f"Missing or incomplete model files: {missing_files}\n" + f"Missing tensors: {missing}") else: raise ValueError("Mismatch between weight map and model parts for tensor names:\n" f"Missing tensors: {missing}\n" @@ -392,6 +407,10 @@ class Model: self.metadata = gguf.Metadata.load(self.metadata_override, self.dir_model_card, self.model_name, total_params) + # If we are using HF model id, set the metadata name to the model id + if self.remote_hf_model_id: + self.metadata.name = self.remote_hf_model_id + # Fallback to model directory name if metadata name is still missing if self.metadata.name is None: self.metadata.name = self.dir_model.name @@ -528,6 +547,8 @@ class Model: reverse_vocab = {id_: encoded_tok for encoded_tok, id_ in tokenizer.vocab.items()} added_vocab = tokenizer.get_added_vocab() + added_tokens_decoder = tokenizer.added_tokens_decoder + for i in range(vocab_size): if i not in reverse_vocab: tokens.append(f"[PAD{i}]") @@ -537,13 +558,13 @@ class Model: if token in added_vocab: # The tokenizer in llama.cpp assumes the CONTROL and USER_DEFINED tokens are pre-normalized. # To avoid unexpected issues - we make sure to normalize non-normalized tokens - if not tokenizer.added_tokens_decoder[i].normalized: + if not added_tokens_decoder[i].normalized: previous_token = token token = tokenizer.decode(tokenizer.encode(token, add_special_tokens=False)) if previous_token != token: logger.info(f"{repr(previous_token)} is encoded and decoded back to {repr(token)} using AutoTokenizer") - if tokenizer.added_tokens_decoder[i].special or self.does_token_look_special(token): + if added_tokens_decoder[i].special or self.does_token_look_special(token): toktypes.append(gguf.TokenType.CONTROL) else: # NOTE: this was added for Gemma. @@ -558,7 +579,7 @@ class Model: # NOTE: this function is generated by convert_hf_to_gguf_update.py # do not modify it manually! - # ref: https://github.com/ggerganov/llama.cpp/pull/6920 + # ref: https://github.com/ggml-org/llama.cpp/pull/6920 # Marker: Start get_vocab_base_pre def get_vocab_base_pre(self, tokenizer) -> str: # encoding this string and hashing the resulting tokens would (hopefully) give us a unique identifier that @@ -699,6 +720,24 @@ class Model: if chkhsh == "b3f499bb4255f8ca19fccd664443283318f2fd2414d5e0b040fbdd0cc195d6c5": # ref: https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B res = "deepseek-r1-qwen" + if chkhsh == "ccc2ef013c104be7bae2965776d611e1d7a8a2a9c547dd93a682c9a9fc80352e": + # ref: https://huggingface.co/Xenova/gpt-4o + res = "gpt-4o" + if chkhsh == "7dec86086fcc38b66b7bc1575a160ae21cf705be7718b9d5598190d7c12db76f": + # ref: https://huggingface.co/UW/OLMo2-8B-SuperBPE-t180k + res = "superbpe" + if chkhsh == "1994ffd01900cfb37395608534236ecd63f2bd5995d6cb1004dda1af50240f15": + # ref: https://huggingface.co/trillionlabs/Trillion-7B-preview + res = "trillion" + if chkhsh == "96a5f08be6259352137b512d4157e333e21df7edd3fcd152990608735a65b224": + # ref: https://huggingface.co/inclusionAI/Ling-lite + res = "bailingmoe" + if chkhsh == "d353350c764d8c3b39c763113960e4fb4919bea5fbf208a0e3b22e8469dc7406": + # ref: https://huggingface.co/meta-llama/Llama-4-Scout-17B-16E-Instruct + res = "llama4" + if chkhsh == "a1336059768a55c99a734006ffb02203cd450fed003e9a71886c88acf24fdbc2": + # ref: https://huggingface.co/THUDM/glm-4-9b-hf + res = "glm4" if res is None: logger.warning("\n") @@ -708,7 +747,7 @@ class Model: logger.warning("** - the model has not been added to convert_hf_to_gguf_update.py yet") logger.warning("** - the pre-tokenization config has changed upstream") logger.warning("** Check your model files and convert_hf_to_gguf_update.py and update them accordingly.") - logger.warning("** ref: https://github.com/ggerganov/llama.cpp/pull/6920") + logger.warning("** ref: https://github.com/ggml-org/llama.cpp/pull/6920") logger.warning("**") logger.warning(f"** chkhsh: {chkhsh}") logger.warning("**************************************************************************************") @@ -858,6 +897,9 @@ class Model: for token_id, token_data in added_tokens_decoder.items(): token_id = int(token_id) token: str = token_data["content"] + if token_id >= vocab_size: + logger.warning(f'ignore token {token_id}: id is out of range, max={vocab_size - 1}') + continue if toktypes[token_id] != SentencePieceTokenTypes.UNUSED: if tokens[token_id] != token.encode("utf-8"): logger.warning(f'replacing token {token_id}: {tokens[token_id].decode("utf-8")!r} -> {token!r}') @@ -902,6 +944,40 @@ class Model: special_vocab = gguf.SpecialVocab(self.dir_model, n_vocab=len(tokens)) special_vocab.add_to_gguf(self.gguf_writer) + def _set_vocab_rwkv_world(self): + assert (self.dir_model / "rwkv_vocab_v20230424.txt").is_file() + vocab_size = self.hparams.get("vocab_size", 65536) + + tokens: list[bytes] = [''.encode("utf-8")] + toktypes: list[int] = [gguf.TokenType.CONTROL] + + with open(self.dir_model / "rwkv_vocab_v20230424.txt", "r", encoding="utf-8") as f: + lines = f.readlines() + for line in lines: + parts = line.split(' ') + assert len(parts) >= 3 + token, token_len = ast.literal_eval(' '.join(parts[1:-1])), int(parts[-1]) + token = token.encode("utf-8") if isinstance(token, str) else token + assert isinstance(token, bytes) + assert len(token) == token_len + token_text: str = repr(token)[2:-1] # "b'\xff'" -> "\xff" + tokens.append(token_text.encode("utf-8")) + toktypes.append(gguf.TokenType.NORMAL) + remainder = vocab_size - len(tokens) + assert remainder >= 0 + for i in range(len(tokens), vocab_size): + tokens.append(f"[PAD{i}]".encode("utf-8")) + toktypes.append(gguf.TokenType.UNUSED) + + self.gguf_writer.add_tokenizer_model("rwkv") + self.gguf_writer.add_token_list(tokens) + self.gguf_writer.add_token_types(toktypes) + special_vocab = gguf.SpecialVocab(self.dir_model, load_merges=False) + special_vocab.chat_template = "rwkv-world" + # hack: Add '\n\n' as the EOT token to make it chat normally + special_vocab._set_special_token("eot", 261) + special_vocab.add_to_gguf(self.gguf_writer) + def _set_vocab_builtin(self, model_name: Literal["gpt-neox", "llama-spm"], vocab_size: int): tokenizer_path = Path(sys.path[0]) / "models" / f"ggml-vocab-{model_name}.gguf" logger.warning(f"Using tokenizer from '{os.path.relpath(tokenizer_path, os.getcwd())}'") @@ -1059,13 +1135,6 @@ class BloomModel(Model): tensors.append((self.map_tensor_name(name), data_torch)) - if name == "word_embeddings.weight": - assert self.tensor_names is not None - - # TODO: tie them at runtime, don't duplicate in the model file - if all(s not in self.tensor_names for s in ("lm_head.weight", "output.weight")): - tensors.append((self.format_tensor_name(gguf.MODEL_TENSOR.OUTPUT), data_torch)) - return tensors @@ -1563,6 +1632,7 @@ class StableLMModel(Model): @Model.register("LLaMAForCausalLM", "LlamaForCausalLM", "MistralForCausalLM", "MixtralForCausalLM") class LlamaModel(Model): model_arch = gguf.MODEL_ARCH.LLAMA + undo_permute = True def set_vocab(self): try: @@ -1627,10 +1697,11 @@ class LlamaModel(Model): n_head = self.hparams["num_attention_heads"] n_kv_head = self.hparams.get("num_key_value_heads") - if name.endswith(("q_proj.weight", "q_proj.bias")): - data_torch = LlamaModel.permute(data_torch, n_head, n_head) - if name.endswith(("k_proj.weight", "k_proj.bias")): - data_torch = LlamaModel.permute(data_torch, n_head, n_kv_head) + if self.undo_permute: + if name.endswith(("q_proj.weight", "q_proj.bias")): + data_torch = LlamaModel.permute(data_torch, n_head, n_head) + if name.endswith(("k_proj.weight", "k_proj.bias")): + data_torch = LlamaModel.permute(data_torch, n_head, n_kv_head) # process the experts separately if name.find("block_sparse_moe.experts") != -1: @@ -1682,7 +1753,7 @@ class LlamaModel(Model): low_freq_wavelen = old_context_len / low_freq_factor high_freq_wavelen = old_context_len / high_freq_factor - assert low_freq_wavelen != high_freq_wavelen + # assert low_freq_wavelen != high_freq_wavelen # Errors for Llama4 rope_factors = [] for freq in freqs: @@ -1707,6 +1778,76 @@ class LlamaModel(Model): raise ValueError(f"Unprocessed experts: {experts}") +@Model.register("Llama4ForConditionalGeneration") +class Llama4Model(LlamaModel): + model_arch = gguf.MODEL_ARCH.LLAMA4 + has_vision: bool = False + undo_permute = False + + # TODO @ngxson : avoid duplicate this code everywhere by at least support "text_config" + # same with llama, but we need to merge the text_config into the root level of hparams + def __init__(self, *args, **kwargs): + hparams = kwargs["hparams"] if "hparams" in kwargs else Model.load_hparams(args[0]) + if "text_config" in hparams: + hparams = {**hparams, **hparams["text_config"]} + kwargs["hparams"] = hparams + super().__init__(*args, **kwargs) + if "vision_config" in hparams: + logger.info("Has vision encoder, but it will be ignored") + self.has_vision = True + # IMPORTANT: the normal "intermediate_size" is renamed to "intermediate_size_mlp", we need to undo this + self.hparams["intermediate_size_moe"] = self.hparams["intermediate_size"] + self.hparams["intermediate_size"] = self.hparams["intermediate_size_mlp"] + + def set_vocab(self): + self._set_vocab_gpt2() + self.gguf_writer.add_add_bos_token(True) + + def set_gguf_parameters(self): + super().set_gguf_parameters() + self.gguf_writer.add_interleave_moe_layer_step(self.hparams["interleave_moe_layer_step"]) + self.gguf_writer.add_expert_feed_forward_length(self.hparams["intermediate_size_moe"]) + + def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None): + # split the gate_up into gate and up + if "gate_up_proj" in name: + name_up = name.replace("gate_up_proj", "up_proj.weight") + name_gate = name.replace("gate_up_proj", "gate_proj.weight") + dim_half = data_torch.shape[-1] // 2 + gate_proj_weight, up_proj_weight = data_torch.transpose(-1, -2).split(dim_half, dim=-2) + return [ + (self.map_tensor_name(name_gate), gate_proj_weight), + (self.map_tensor_name(name_up), up_proj_weight) + ] + + if name.endswith("down_proj"): + name += ".weight" + data_torch = data_torch.transpose(-1, -2) + + if "multi_modal_projector" in name or "vision_model" in name: + return [] + return super().modify_tensors(data_torch, name, bid) + + +@Model.register("Mistral3ForConditionalGeneration") +class Mistral3Model(LlamaModel): + model_arch = gguf.MODEL_ARCH.LLAMA + + # we need to merge the text_config into the root level of hparams + def __init__(self, *args, **kwargs): + hparams = kwargs["hparams"] if "hparams" in kwargs else Model.load_hparams(args[0]) + if "text_config" in hparams: + hparams = {**hparams, **hparams["text_config"]} + kwargs["hparams"] = hparams + super().__init__(*args, **kwargs) + + def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None): + name = name.replace("language_model.", "") + if "multi_modal_projector" in name or "vision_tower" in name: + return [] + return super().modify_tensors(data_torch, name, bid) + + @Model.register("DeciLMForCausalLM") class DeciModel(Model): model_arch = gguf.MODEL_ARCH.DECI @@ -2211,7 +2352,7 @@ class Qwen2Model(Model): self.gguf_writer.add_rope_scaling_orig_ctx_len(self.hparams["rope_scaling"]["original_max_position_embeddings"]) -@Model.register("Qwen2VLForConditionalGeneration") +@Model.register("Qwen2VLForConditionalGeneration", "Qwen2_5_VLForConditionalGeneration") class Qwen2VLModel(Model): model_arch = gguf.MODEL_ARCH.QWEN2VL @@ -2335,6 +2476,16 @@ class Qwen2MoeModel(Model): raise ValueError(f"Unprocessed experts: {experts}") +@Model.register("Qwen3ForCausalLM") +class Qwen3Model(Qwen2Model): + model_arch = gguf.MODEL_ARCH.QWEN3 + + +@Model.register("Qwen3MoeForCausalLM") +class Qwen3MoeModel(Qwen2MoeModel): + model_arch = gguf.MODEL_ARCH.QWEN3MOE + + @Model.register("GPT2LMHeadModel") class GPT2Model(Model): model_arch = gguf.MODEL_ARCH.GPT2 @@ -2364,10 +2515,6 @@ class GPT2Model(Model): tensors.append((new_name, data_torch)) - # note: GPT2 output is tied to (same as) wte in original model - if new_name == self.format_tensor_name(gguf.MODEL_TENSOR.TOKEN_EMBD): - tensors.append((self.format_tensor_name(gguf.MODEL_TENSOR.OUTPUT), data_torch)) - return tensors @@ -2512,7 +2659,8 @@ class Phi3MiniModel(Model): rms_eps = self.find_hparam(["rms_norm_eps"]) max_pos_embds = self.find_hparam(["n_positions", "max_position_embeddings"]) orig_max_pos_embds = self.find_hparam(["original_max_position_embeddings"]) - rope_dims = n_embd // n_head + rot_pct = self.hparams.get("partial_rotary_factor", 1.0) + rope_dims = int(rot_pct * n_embd) // n_head self.gguf_writer.add_context_length(max_pos_embds) self.gguf_writer.add_rope_scaling_orig_ctx_len(orig_max_pos_embds) @@ -2536,7 +2684,8 @@ class Phi3MiniModel(Model): n_head = self.find_hparam(["num_attention_heads", "n_head"]) max_pos_embds = self.find_hparam(["n_positions", "max_position_embeddings"]) orig_max_pos_embds = self.find_hparam(["original_max_position_embeddings"]) - rope_dims = n_embd // n_head + rot_pct = self.hparams.get("partial_rotary_factor", 1.0) + rope_dims = int(rot_pct * n_embd) // n_head # write rope scaling for long context (128k) model rope_scaling = self.find_hparam(['rope_scaling'], True) @@ -2565,7 +2714,7 @@ class Phi3MiniModel(Model): raise KeyError('Missing the required key rope_scaling.long_factor or rope_scaling_short_factor') if len(long_factors) != len(short_factors) or len(long_factors) != rope_dims / 2: - raise ValueError(f'The length of rope long and short factors must be {rope_dims / 2}') + raise ValueError(f'The length of rope long and short factors must be {rope_dims / 2}. long_factors = {len(long_factors)}, short_factors = {len(short_factors)}.') yield (self.format_tensor_name(gguf.MODEL_TENSOR.ROPE_FACTORS_LONG), torch.tensor(long_factors, dtype=torch.float32)) yield (self.format_tensor_name(gguf.MODEL_TENSOR.ROPE_FACTORS_SHORT), torch.tensor(short_factors, dtype=torch.float32)) @@ -2695,21 +2844,26 @@ class CodeShellModel(Model): self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR) self.gguf_writer.add_rope_scaling_factor(1.0) + _has_tok_embd = False + def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]: del bid # unused + output_name = self.format_tensor_name(gguf.MODEL_TENSOR.OUTPUT) + tok_embd_name = self.format_tensor_name(gguf.MODEL_TENSOR.TOKEN_EMBD) + new_name = self.map_tensor_name(name) - tensors: list[tuple[str, Tensor]] = [(new_name, data_torch)] + # assuming token_embd.weight is seen before output.weight + if not self._has_tok_embd and new_name == self.format_tensor_name(gguf.MODEL_TENSOR.OUTPUT): + # even though the tensor file(s) does not contain the word embeddings they are still in the weight map + if self.tensor_names and "transformer.wte.weight" in self.tensor_names: + logger.debug(f"{tok_embd_name} not found before {output_name}, assuming they are tied") + self.tensor_names.remove("transformer.wte.weight") + elif new_name == tok_embd_name: + self._has_tok_embd = True - if new_name == self.format_tensor_name(gguf.MODEL_TENSOR.TOKEN_EMBD): - assert self.tensor_names is not None - - if all(s not in self.tensor_names for s in ("lm_head.weight", "output.weight")): - # copy tok_embd.weight to output.weight - tensors.append((self.format_tensor_name(gguf.MODEL_TENSOR.OUTPUT), data_torch)) - - return tensors + return [(new_name, data_torch)] @Model.register("InternLM2ForCausalLM") @@ -2835,7 +2989,7 @@ class InternLM2Model(Model): if chat_eos_token_id is not None: # For the chat model, we replace the eos with '<|im_end|>'. # TODO: this is a hack, should be fixed - # https://github.com/ggerganov/llama.cpp/pull/6745#issuecomment-2067687048 + # https://github.com/ggml-org/llama.cpp/pull/6745#issuecomment-2067687048 special_vocab.special_token_ids["eos"] = chat_eos_token_id logger.warning(f"Replace eos:{old_eos} with a special token:{chat_eos_token_id}" " in chat mode so that the conversation can end normally.") @@ -3317,6 +3471,83 @@ class Gemma2Model(Model): return [(self.map_tensor_name(name), data_torch)] +@Model.register("Gemma3ForCausalLM", "Gemma3ForConditionalGeneration") +class Gemma3Model(Model): + model_arch = gguf.MODEL_ARCH.GEMMA3 + has_vision: bool = False + + # we need to merge the text_config into the root level of hparams + def __init__(self, *args, **kwargs): + hparams = kwargs["hparams"] if "hparams" in kwargs else Model.load_hparams(args[0]) + if "text_config" in hparams: + hparams = {**hparams, **hparams["text_config"]} + kwargs["hparams"] = hparams + super().__init__(*args, **kwargs) + if "vision_config" in hparams: + logger.info("Has vision encoder, but it will be ignored") + self.has_vision = True + + def write(self): + super().write() + if self.has_vision: + logger.info("NOTE: this script only convert the language model to GGUF") + logger.info(" for the vision model, please use gemma3_convert_encoder_to_gguf.py") + + def set_vocab(self): + self._set_vocab_sentencepiece() + + self.gguf_writer.add_add_space_prefix(False) + + def set_gguf_parameters(self): + hparams = self.hparams + block_count = hparams["num_hidden_layers"] + + # some default values are not specified in the hparams + self.gguf_writer.add_context_length(hparams.get("max_position_embeddings", 131072)) + self.gguf_writer.add_embedding_length(hparams["hidden_size"]) + self.gguf_writer.add_block_count(block_count) + self.gguf_writer.add_feed_forward_length(hparams["intermediate_size"]) + self.gguf_writer.add_head_count(hparams.get("num_attention_heads", 8)) + self.gguf_writer.add_layer_norm_rms_eps(self.hparams.get("rms_norm_eps", 1e-6)) + self.gguf_writer.add_key_length(hparams.get("head_dim", 256)) + self.gguf_writer.add_value_length(hparams.get("head_dim", 256)) + self.gguf_writer.add_file_type(self.ftype) + self.gguf_writer.add_rope_freq_base(hparams.get("rope_theta", 1_000_000.0)) # for global layers + # both attn_logit_softcapping and final_logit_softcapping are removed in Gemma3 + assert hparams.get("attn_logit_softcapping") is None + assert hparams.get("final_logit_softcapping") is None + self.gguf_writer.add_sliding_window(hparams["sliding_window"]) + self.gguf_writer.add_head_count_kv(hparams.get("num_key_value_heads", 4)) + if hparams.get("rope_scaling") is not None: + assert hparams["rope_scaling"]["rope_type"] == "linear" + # important: this rope_scaling is only applied for global layers, and not used by 1B model + self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR) + self.gguf_writer.add_rope_scaling_factor(hparams["rope_scaling"]["factor"]) + + def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]: + del bid # unused + + if name.startswith("language_model."): + name = name.replace("language_model.", "") + elif name.startswith("multi_modal_projector.") or name.startswith("vision_tower.") \ + or name.startswith("multimodal_projector.") or name.startswith("vision_model."): # this is for old HF model, should be removed later + # ignore vision tensors + return [] + + # remove OOV (out-of-vocabulary) rows in token_embd + if "embed_tokens.weight" in name: + vocab = self._create_vocab_sentencepiece() + tokens = vocab[0] + data_torch = data_torch[:len(tokens)] + + # ref code in Gemma3RMSNorm + # output = output * (1.0 + self.weight.float()) + if name.endswith("norm.weight"): + data_torch = data_torch + 1 + + return [(self.map_tensor_name(name), data_torch)] + + @Model.register("Starcoder2ForCausalLM") class StarCoder2Model(Model): model_arch = gguf.MODEL_ARCH.STARCODER2 @@ -3327,38 +3558,7 @@ class Rwkv6Model(Model): model_arch = gguf.MODEL_ARCH.RWKV6 def set_vocab(self): - assert (self.dir_model / "rwkv_vocab_v20230424.txt").is_file() - vocab_size = self.hparams.get("vocab_size", 65536) - - tokens: list[bytes] = [''.encode("utf-8")] - toktypes: list[int] = [gguf.TokenType.CONTROL] - - with open(self.dir_model / "rwkv_vocab_v20230424.txt", "r", encoding="utf-8") as f: - lines = f.readlines() - for line in lines: - parts = line.split(' ') - assert len(parts) >= 3 - token, token_len = ast.literal_eval(' '.join(parts[1:-1])), int(parts[-1]) - token = token.encode("utf-8") if isinstance(token, str) else token - assert isinstance(token, bytes) - assert len(token) == token_len - token_text: str = repr(token)[2:-1] # "b'\xff'" -> "\xff" - tokens.append(token_text.encode("utf-8")) - toktypes.append(gguf.TokenType.NORMAL) - remainder = vocab_size - len(tokens) - assert remainder >= 0 - for i in range(len(tokens), vocab_size): - tokens.append(f"[PAD{i}]".encode("utf-8")) - toktypes.append(gguf.TokenType.UNUSED) - - self.gguf_writer.add_tokenizer_model("rwkv") - self.gguf_writer.add_token_list(tokens) - self.gguf_writer.add_token_types(toktypes) - special_vocab = gguf.SpecialVocab(self.dir_model, load_merges=False) - special_vocab.chat_template = "rwkv-world" - # hack: Add '\n\n' as the EOT token to make it chat normally - special_vocab._set_special_token("eot", 261) - special_vocab.add_to_gguf(self.gguf_writer) + self._set_vocab_rwkv_world() def set_gguf_parameters(self): block_count = self.hparams["num_hidden_layers"] @@ -3444,8 +3644,8 @@ class RWKV6Qwen2Model(Rwkv6Model): head_size = hidden_size // num_attention_heads rms_norm_eps = self.hparams["rms_norm_eps"] intermediate_size = self.hparams["intermediate_size"] - time_mix_extra_dim = 64 if hidden_size >= 4096 else 32 - time_decay_extra_dim = 128 if hidden_size >= 4096 else 64 + time_mix_extra_dim = self.hparams.get("lora_rank_tokenshift", 64 if hidden_size >= 4096 else 32) + time_decay_extra_dim = self.hparams.get("lora_rank_decay", 128 if hidden_size >= 4096 else 64) # RWKV isn't context limited self.gguf_writer.add_context_length(1048576) @@ -3480,6 +3680,168 @@ class RWKV6Qwen2Model(Rwkv6Model): yield (new_name, data) +@Model.register("Rwkv7ForCausalLM", "RWKV7ForCausalLM") +class Rwkv7Model(Model): + model_arch = gguf.MODEL_ARCH.RWKV7 + + def set_vocab(self): + self._set_vocab_rwkv_world() + + def calc_lora_rank(self, hidden_size, exponent, multiplier): + return max(1, round(hidden_size ** exponent * multiplier / 32)) * 32 + + def set_gguf_parameters(self): + block_count = self.hparams["num_hidden_layers"] + try: + head_size = self.hparams["head_size"] + layer_norm_eps = self.hparams["layer_norm_epsilon"] + except KeyError: + head_size = self.hparams["head_dim"] + layer_norm_eps = self.hparams["norm_eps"] + hidden_size = self.hparams["hidden_size"] + intermediate_size = self.hparams["intermediate_size"] if self.hparams["intermediate_size"] is not None else (hidden_size * 4) + + # ICLR: In-Context-Learning-Rate + try: + lora_rank_decay = self.hparams["lora_rank_decay"] if self.hparams["lora_rank_decay"] is not None else self.calc_lora_rank(hidden_size, 0.5, 1.8) + lora_rank_iclr = self.hparams["lora_rank_iclr"] if self.hparams["lora_rank_iclr"] is not None else self.calc_lora_rank(hidden_size, 0.5, 1.8) + lora_rank_value_residual_mix = self.hparams["lora_rank_value_residual_mix"] if self.hparams["lora_rank_value_residual_mix"] is not None else self.calc_lora_rank(hidden_size, 0.5, 1.3) + lora_rank_gate = self.hparams["lora_rank_gate"] if self.hparams["lora_rank_gate"] is not None else self.calc_lora_rank(hidden_size, 0.8, 0.6) + except KeyError: + lora_rank_decay = self.hparams["decay_low_rank_dim"] if self.hparams["decay_low_rank_dim"] is not None else self.calc_lora_rank(hidden_size, 0.5, 1.8) + lora_rank_iclr = self.hparams["a_low_rank_dim"] if self.hparams["a_low_rank_dim"] is not None else self.calc_lora_rank(hidden_size, 0.5, 1.8) + lora_rank_value_residual_mix = self.hparams["v_low_rank_dim"] if self.hparams["v_low_rank_dim"] is not None else self.calc_lora_rank(hidden_size, 0.5, 1.3) + lora_rank_gate = self.hparams["gate_low_rank_dim"] if self.hparams["gate_low_rank_dim"] is not None else self.calc_lora_rank(hidden_size, 0.8, 0.6) + + # RWKV isn't context limited + self.gguf_writer.add_context_length(1048576) + self.gguf_writer.add_embedding_length(hidden_size) + self.gguf_writer.add_block_count(block_count) + self.gguf_writer.add_layer_norm_eps(layer_norm_eps) + self.gguf_writer.add_wkv_head_size(head_size) + self.gguf_writer.add_decay_lora_rank(lora_rank_decay) + self.gguf_writer.add_iclr_lora_rank(lora_rank_iclr) + self.gguf_writer.add_value_residual_mix_lora_rank(lora_rank_value_residual_mix) + self.gguf_writer.add_gate_lora_rank(lora_rank_gate) + self.gguf_writer.add_feed_forward_length(intermediate_size) + self.gguf_writer.add_file_type(self.ftype) + + # required by llama.cpp, unused + self.gguf_writer.add_head_count(0) + + lerp_weights: dict[int, dict[str, Tensor]] = {} + lora_needs_transpose: bool = True + + def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]: + # unify tensor names here to make life easier + name = name.replace("blocks", "layers").replace("ffn", "feed_forward") + name = name.replace("self_attn", "attention").replace("attn", "attention") + name = name.replace("time_mixer.", "") + # lora layer names in fla-hub's impl + if "_lora.lora" in name: + self.lora_needs_transpose = False + name = name.replace("_lora.lora.0.weight", "1.weight") + name = name.replace("_lora.lora.2.weight", "2.weight") + name = name.replace("_lora.lora.2.bias", "0.weight") + + name = name.replace("feed_forward_norm", "ln2") + name = name.replace("g_norm", "ln_x") + + if "attention.v" in name and "value" not in self.map_tensor_name(name) and bid == 0: + # some models have dummy v0/v1/v2 on first layer while others don't + # ignore them all since they are not used + return + + wkv_has_gate = self.hparams.get("wkv_has_gate", True) + lerp_list = ["r", "w", "k", "v", "a", "g"] if wkv_has_gate else ["r", "w", "k", "v", "a"] + + if bid is not None and "attention.x_" in name: + if "attention.x_x" in name: + # already concatenated + new_name = f"blk.{bid}.time_mix_lerp_fused.weight" + data = data_torch.reshape(len(lerp_list), 1, 1, -1) + yield (new_name, data) + else: + try: + self.lerp_weights[bid][name] = data_torch + except KeyError: + self.lerp_weights[bid] = {name: data_torch} + if all(f"model.layers.{bid}.attention.x_{i}" in self.lerp_weights[bid].keys() for i in lerp_list): + new_name = f"blk.{bid}.time_mix_lerp_fused.weight" + data = torch.stack([self.lerp_weights[bid][f"model.layers.{bid}.attention.x_{i}"] for i in lerp_list], dim=0) + yield (new_name, data) + return + else: + data_torch = data_torch.squeeze() + new_name = self.map_tensor_name(name) + + if not (new_name.endswith(".weight") or new_name.endswith(".bias")): + new_name += ".weight" + + if self.lora_needs_transpose and any( + new_name.endswith(t) for t in [ + "time_mix_w1.weight", "time_mix_w2.weight", + "time_mix_a1.weight", "time_mix_a2.weight", + "time_mix_v1.weight", "time_mix_v2.weight", + "time_mix_g1.weight", "time_mix_g2.weight", + ] + ): + data_torch = data_torch.transpose(0, 1) + + if 'r_k' in new_name: + data_torch = data_torch.flatten() + + if bid == 0 and "time_mix_a" in new_name: + # dummy v0/v1/v2 on first layer + # easist way to make llama happy + yield (new_name.replace("time_mix_a", "time_mix_v"), data_torch) + + yield (new_name, data_torch) + + +@Model.register("RwkvHybridForCausalLM") +class ARwkv7Model(Rwkv7Model): + model_arch = gguf.MODEL_ARCH.ARWKV7 + + def set_vocab(self): + try: + self._set_vocab_sentencepiece() + except FileNotFoundError: + self._set_vocab_gpt2() + + def set_gguf_parameters(self): + block_count = self.hparams["num_hidden_layers"] + hidden_size = self.hparams["hidden_size"] + head_size = self.hparams["head_size"] + rms_norm_eps = self.hparams["rms_norm_eps"] + intermediate_size = self.hparams["intermediate_size"] + wkv_has_gate = self.hparams["wkv_has_gate"] + assert self.hparams["wkv_version"] == 7 + + # ICLR: In-Context-Learning-Rate + lora_rank_decay = 64 + lora_rank_iclr = 64 + lora_rank_value_residual_mix = 32 + lora_rank_gate = 128 if wkv_has_gate else 0 + + # RWKV isn't context limited + self.gguf_writer.add_context_length(1048576) + self.gguf_writer.add_embedding_length(hidden_size) + self.gguf_writer.add_block_count(block_count) + self.gguf_writer.add_layer_norm_rms_eps(rms_norm_eps) + self.gguf_writer.add_wkv_head_size(head_size) + self.gguf_writer.add_decay_lora_rank(lora_rank_decay) + self.gguf_writer.add_iclr_lora_rank(lora_rank_iclr) + self.gguf_writer.add_value_residual_mix_lora_rank(lora_rank_value_residual_mix) + self.gguf_writer.add_gate_lora_rank(lora_rank_gate) + self.gguf_writer.add_feed_forward_length(intermediate_size) + self.gguf_writer.add_file_type(self.ftype) + self.gguf_writer.add_token_shift_count(1) + + # required by llama.cpp, unused + self.gguf_writer.add_head_count(0) + + @Model.register("MambaForCausalLM", "MambaLMHeadModel", "FalconMambaForCausalLM") class MambaModel(Model): model_arch = gguf.MODEL_ARCH.MAMBA @@ -3534,8 +3896,6 @@ class MambaModel(Model): _tok_embd = None def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]: - del bid # unused - output_name = self.format_tensor_name(gguf.MODEL_TENSOR.OUTPUT) tok_embd_name = self.format_tensor_name(gguf.MODEL_TENSOR.TOKEN_EMBD) @@ -3545,6 +3905,10 @@ class MambaModel(Model): logger.debug("A_log --> A ==> " + new_name) data_torch = -torch.exp(data_torch) + # [4 1 8192 1] -> [4 8192 1 1] + if self.match_model_tensor_name(new_name, gguf.MODEL_TENSOR.SSM_CONV1D, bid): + data_torch = data_torch.squeeze() + # assuming token_embd.weight is seen before output.weight if self._tok_embd is not None and new_name == output_name: if torch.equal(self._tok_embd, data_torch): @@ -4148,6 +4512,29 @@ class DeepseekV2Model(Model): raise ValueError(f"Unprocessed experts: {experts}") +@Model.register("PLMForCausalLM") +class PLMModel(Model): + model_arch = gguf.MODEL_ARCH.PLM + + def set_vocab(self): + self._set_vocab_gpt2() + + def set_gguf_parameters(self): + super().set_gguf_parameters() + hparams = self.hparams + self.gguf_writer.add_vocab_size(hparams["vocab_size"]) + self.gguf_writer.add_kv_lora_rank(hparams["kv_lora_rank"]) + self.gguf_writer.add_key_length(hparams["qk_nope_head_dim"] + hparams["qk_rope_head_dim"]) + self.gguf_writer.add_value_length(hparams["v_head_dim"]) + self.gguf_writer.add_rope_dimension_count(hparams["qk_rope_head_dim"]) + + def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]: + return [(self.map_tensor_name(name), data_torch)] + + def prepare_tensors(self): + super().prepare_tensors() + + @Model.register("T5WithLMHeadModel") @Model.register("T5ForConditionalGeneration") @Model.register("MT5ForConditionalGeneration") @@ -4513,6 +4900,22 @@ class JaisModel(Model): self.gguf_writer.add_max_alibi_bias(self.max_alibi_bias) +@Model.register("Glm4ForCausalLM") +class Glm4Model(Model): + model_arch = gguf.MODEL_ARCH.GLM4 + + def set_vocab(self): + self._set_vocab_gpt2() + + def set_gguf_parameters(self): + super().set_gguf_parameters() + if self.hparams.get("rope_scaling") is not None and "factor" in self.hparams["rope_scaling"]: + if self.hparams["rope_scaling"].get("type") == "yarn": + self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.YARN) + self.gguf_writer.add_rope_scaling_factor(self.hparams["rope_scaling"]["factor"]) + self.gguf_writer.add_rope_scaling_orig_ctx_len(self.hparams["rope_scaling"]["original_max_position_embeddings"]) + + @Model.register("GlmForCausalLM", "ChatGLMModel", "ChatGLMForConditionalGeneration") class ChatGLMModel(Model): model_arch = gguf.MODEL_ARCH.CHATGLM @@ -4836,6 +5239,105 @@ class GraniteMoeModel(GraniteModel): return super().modify_tensors(data_torch, name, bid) +@Model.register("BailingMoeForCausalLM") +class BailingMoeModel(Model): + model_arch = gguf.MODEL_ARCH.BAILINGMOE + + def set_vocab(self): + self._set_vocab_gpt2() + + def set_gguf_parameters(self): + super().set_gguf_parameters() + hparams = self.hparams + rope_dim = hparams.get("head_dim") or hparams["hidden_size"] // hparams["num_attention_heads"] + + self.gguf_writer.add_rope_dimension_count(rope_dim) + self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.NONE) + self.gguf_writer.add_leading_dense_block_count(hparams["first_k_dense_replace"]) + self.gguf_writer.add_vocab_size(hparams["vocab_size"]) + self.gguf_writer.add_expert_feed_forward_length(hparams["moe_intermediate_size"]) + self.gguf_writer.add_expert_weights_scale(1.0) + self.gguf_writer.add_expert_count(hparams["num_experts"]) + self.gguf_writer.add_expert_shared_count(hparams["num_shared_experts"]) + self.gguf_writer.add_expert_weights_norm(hparams["norm_topk_prob"]) + + _experts: list[dict[str, Tensor]] | None = None + + @staticmethod + def permute(weights: Tensor, n_head: int, n_head_kv: int | None): + if n_head_kv is not None and n_head != n_head_kv: + n_head = n_head_kv + return (weights.reshape(n_head, 2, weights.shape[0] // n_head // 2, *weights.shape[1:]) + .swapaxes(1, 2) + .reshape(weights.shape)) + + def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]: + n_head = self.hparams["num_attention_heads"] + n_kv_head = self.hparams.get("num_key_value_heads") + n_embd = self.hparams["hidden_size"] + head_dim = self.hparams.get("head_dim") or n_embd // n_head + + output_name = self.format_tensor_name(gguf.MODEL_TENSOR.OUTPUT) + + if name.endswith("attention.dense.weight"): + return [(self.format_tensor_name(gguf.MODEL_TENSOR.ATTN_OUT, bid), data_torch)] + elif name.endswith("query_key_value.weight"): + q, k, v = data_torch.split([n_head * head_dim, n_kv_head * head_dim, n_kv_head * head_dim], dim=-2) + + return [ + (self.format_tensor_name(gguf.MODEL_TENSOR.ATTN_Q, bid), BailingMoeModel.permute(q, n_head, n_head)), + (self.format_tensor_name(gguf.MODEL_TENSOR.ATTN_K, bid), BailingMoeModel.permute(k, n_head, n_kv_head)), + (self.format_tensor_name(gguf.MODEL_TENSOR.ATTN_V, bid), v) + ] + elif name.find("mlp.experts") != -1: + n_experts = self.hparams["num_experts"] + assert bid is not None + + tensors: list[tuple[str, Tensor]] = [] + + if self._experts is None: + self._experts = [{} for _ in range(self.block_count)] + + self._experts[bid][name] = data_torch + + if len(self._experts[bid]) >= n_experts * 3: + # merge the experts into a single 3d tensor + for w_name in ["down_proj", "gate_proj", "up_proj"]: + datas: list[Tensor] = [] + + for xid in range(n_experts): + ename = f"model.layers.{bid}.mlp.experts.{xid}.{w_name}.weight" + datas.append(self._experts[bid][ename]) + del self._experts[bid][ename] + + data_torch = torch.stack(datas, dim=0) + + merged_name = f"model.layers.{bid}.mlp.experts.{w_name}.weight" + + new_name = self.map_tensor_name(merged_name) + + tensors.append((new_name, data_torch)) + + return tensors + + new_name = self.map_tensor_name(name) + + if new_name == output_name and self.hparams.get("norm_head"): + data_torch = data_torch.float() + data_torch /= torch.norm(data_torch, p=2, dim=0, keepdim=True) + 1e-7 + + return [(new_name, data_torch)] + + def prepare_tensors(self): + super().prepare_tensors() + + if self._experts is not None: + # flatten `list[dict[str, Tensor]]` into `list[str]` + experts = [k for d in self._experts for k in d.keys()] + if len(experts) > 0: + raise ValueError(f"Unprocessed experts: {experts}") + + @Model.register("ChameleonForConditionalGeneration") @Model.register("ChameleonForCausalLM") # obsolete class ChameleonModel(Model): @@ -4934,6 +5436,14 @@ class LazyTorchTensor(gguf.LazyBase): lazy = cls(meta=cls.meta_with_dtype_and_shape(dtype, shape), args=(st_slice,), func=lambda s: s[:]) return cast(torch.Tensor, lazy) + @classmethod + def from_remote_tensor(cls, remote_tensor: gguf.utility.RemoteTensor): + dtype = cls._dtype_str_map[remote_tensor.dtype] + shape = remote_tensor.shape + meta = cls.meta_with_dtype_and_shape(dtype, shape) + lazy = cls(meta=meta, args=(remote_tensor,), func=lambda r: torch.frombuffer(r.data(), dtype=dtype).reshape(shape)) + return cast(torch.Tensor, lazy) + @classmethod def __torch_function__(cls, func, types, args=(), kwargs=None): del types # unused @@ -5011,6 +5521,10 @@ def parse_args() -> argparse.Namespace: "--print-supported-models", action="store_true", help="Print the supported models" ) + parser.add_argument( + "--remote", action="store_true", + help="(Experimental) Read safetensors file remotely without downloading to disk. Config and tokenizer files will still be downloaded. To use this feature, you need to specify Hugging Face model repo name instead of a local directory. For example: 'HuggingFaceTB/SmolLM2-1.7B-Instruct'. Note: To access gated repo, set HF_TOKEN environment variable to your Hugging Face token.", + ) args = parser.parse_args() if not args.print_supported_models and args.model is None: @@ -5051,6 +5565,14 @@ def main() -> None: dir_model = args.model + if args.remote: + from huggingface_hub import snapshot_download + local_dir = snapshot_download( + repo_id=str(dir_model), + allow_patterns=["LICENSE", "*.json", "*.md", "*.txt", "tokenizer.model"]) + dir_model = Path(local_dir) + logger.info(f"Downloaded config and tokenizer to {local_dir}") + if not dir_model.is_dir(): logger.error(f'Error: {args.model} is not a directory') sys.exit(1) @@ -5072,6 +5594,9 @@ def main() -> None: if args.outfile is not None: fname_out = args.outfile + elif args.remote: + # if remote, use the model ID as the output file name + fname_out = Path("./" + str(args.model).replace("/", "-") + "-{ftype}.gguf") else: fname_out = dir_model @@ -5082,20 +5607,20 @@ def main() -> None: with torch.inference_mode(): output_type = ftype_map[args.outtype] model_architecture = hparams["architectures"][0] - try: model_class = Model.from_model_architecture(model_architecture) except NotImplementedError: logger.error(f"Model {model_architecture} is not supported") sys.exit(1) - model_instance = model_class(dir_model=dir_model, ftype=output_type, fname_out=fname_out, + model_instance = model_class(dir_model, output_type, fname_out, is_big_endian=args.bigendian, use_temp_file=args.use_temp_file, eager=args.no_lazy, metadata_override=args.metadata, model_name=args.model_name, split_max_tensors=args.split_max_tensors, split_max_size=split_str_to_n_bytes(args.split_max_size), dry_run=args.dry_run, - small_first_shard=args.no_tensor_first_split) + small_first_shard=args.no_tensor_first_split, + remote_hf_model_id=str(args.model) if args.remote else None) if args.vocab_only: logger.info("Exporting model vocab...") diff --git a/convert_hf_to_gguf_update.py b/convert_hf_to_gguf_update.py index cea34413f..160c9fe0e 100755 --- a/convert_hf_to_gguf_update.py +++ b/convert_hf_to_gguf_update.py @@ -8,7 +8,7 @@ # provide the necessary information to llama.cpp via the GGUF header in order to implement # the same pre-tokenizer. # -# ref: https://github.com/ggerganov/llama.cpp/pull/6920 +# ref: https://github.com/ggml-org/llama.cpp/pull/6920 # # Instructions: # @@ -109,6 +109,12 @@ models = [ {"name": "megrez", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/Infinigence/Megrez-3B-Instruct"}, {"name": "deepseek-v3", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/deepseek-ai/DeepSeek-V3"}, {"name": "deepseek-r1-qwen", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B"}, + {"name": "gpt-4o", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/Xenova/gpt-4o", }, + {"name": "superbpe", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/UW/OLMo2-8B-SuperBPE-t180k", }, + {"name": "trillion", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/trillionlabs/Trillion-7B-preview", }, + {"name": "bailingmoe", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/inclusionAI/Ling-lite", }, + {"name": "llama4", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/meta-llama/Llama-4-Scout-17B-16E-Instruct", }, + {"name": "glm4", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/THUDM/glm-4-9b-hf", }, ] @@ -131,6 +137,10 @@ def download_model(model): files = ["config.json", "tokenizer.json", "tokenizer_config.json"] + if name == "gpt-4o": + # Xenova/gpt-4o is tokenizer-only, it does not contain config.json + files = ["tokenizer.json", "tokenizer_config.json"] + if tokt == TOKENIZER_TYPE.SPM: files.append("tokenizer.model") @@ -246,7 +256,7 @@ src_func = f""" logger.warning("** - the model has not been added to convert_hf_to_gguf_update.py yet") logger.warning("** - the pre-tokenization config has changed upstream") logger.warning("** Check your model files and convert_hf_to_gguf_update.py and update them accordingly.") - logger.warning("** ref: https://github.com/ggerganov/llama.cpp/pull/6920") + logger.warning("** ref: https://github.com/ggml-org/llama.cpp/pull/6920") logger.warning("**") logger.warning(f"** chkhsh: {{chkhsh}}") logger.warning("**************************************************************************************") diff --git a/convert_lora_to_gguf.py b/convert_lora_to_gguf.py index 6dea14a23..bdc991533 100755 --- a/convert_lora_to_gguf.py +++ b/convert_lora_to_gguf.py @@ -395,7 +395,7 @@ if __name__ == '__main__': logger.error(f"Unexpected name '{name}': Not a lora_A or lora_B tensor") if ".embed_tokens.weight" in name or ".lm_head.weight" in name: logger.error("Embeddings is present in the adapter. This can be due to new tokens added during fine tuning") - logger.error("Please refer to https://github.com/ggerganov/llama.cpp/pull/9948") + logger.error("Please refer to https://github.com/ggml-org/llama.cpp/pull/9948") sys.exit(1) if base_name in tensor_map: @@ -419,7 +419,7 @@ if __name__ == '__main__': # some archs may have the same tensor for lm_head and output (tie word embeddings) # in this case, adapters targeting lm_head will fail when using llama-export-lora # therefore, we ignore them for now - # see: https://github.com/ggerganov/llama.cpp/issues/9065 + # see: https://github.com/ggml-org/llama.cpp/issues/9065 if name == "lm_head.weight" and len(dest) == 0: raise ValueError("lm_head is present in adapter, but is ignored in base model") for dest_name, dest_data in dest: diff --git a/docs/android.md b/docs/android.md index 47530c6c1..d2a835653 100644 --- a/docs/android.md +++ b/docs/android.md @@ -12,7 +12,7 @@ $ apt update && apt upgrade -y $ apt install git cmake ``` -Then, follow the [build instructions](https://github.com/ggerganov/llama.cpp/blob/master/docs/build.md), specifically for CMake. +Then, follow the [build instructions](https://github.com/ggml-org/llama.cpp/blob/master/docs/build.md), specifically for CMake. Once the binaries are built, download your model of choice (e.g., from Hugging Face). It's recommended to place it in the `~/` directory for best performance: diff --git a/docs/cuda-fedora.md b/docs/backend/CUDA-FEDORA.md similarity index 77% rename from docs/cuda-fedora.md rename to docs/backend/CUDA-FEDORA.md index 9c88b7694..1508faf77 100644 --- a/docs/cuda-fedora.md +++ b/docs/backend/CUDA-FEDORA.md @@ -14,9 +14,7 @@ In this guide we setup [Nvidia CUDA](https://docs.nvidia.com/cuda/) in a toolbox - [Creating a Fedora Toolbox Environment](#creating-a-fedora-toolbox-environment) - [Installing Essential Development Tools](#installing-essential-development-tools) - [Adding the CUDA Repository](#adding-the-cuda-repository) -- [Installing `nvidia-driver-libs`](#installing-nvidia-driver-libs) -- [Manually Resolving Package Conflicts](#manually-resolving-package-conflicts) -- [Finalizing the Installation of `nvidia-driver-libs`](#finalizing-the-installation-of-nvidia-driver-libs) +- [Installing Nvidia Driver Libraries](#installing-nvidia-driver-libraries) - [Installing the CUDA Meta-Package](#installing-the-cuda-meta-package) - [Configuring the Environment](#configuring-the-environment) - [Verifying the Installation](#verifying-the-installation) @@ -67,7 +65,7 @@ This guide focuses on Fedora hosts, but with small adjustments, it can work for sudo dnf distro-sync ``` -2. **Install the Default Text Editor (Optional):** +2. **Install **Vim** the default text editor (Optional):** ```bash sudo dnf install vim-default-editor --allowerasing @@ -97,36 +95,48 @@ After adding the repository, synchronize the package manager again: sudo dnf distro-sync ``` -## Installing `nvidia-driver-libs` and `nvidia-driver-cuda-libs` +## Installing Nvidia Driver Libraries -We need to detect if the host is supplying the [NVIDIA driver libraries into the toolbox](https://github.com/containers/toolbox/blob/main/src/pkg/nvidia/nvidia.go). +First, we need to detect if the host is supplying the [NVIDIA driver libraries into the toolbox](https://github.com/containers/toolbox/blob/main/src/pkg/nvidia/nvidia.go): ```bash ls -la /usr/lib64/libcuda.so.1 ``` +### If *`libcuda.so.1`* is missing: + +``` +ls: cannot access '/usr/lib64/libcuda.so.1': No such file or directory +``` + **Explanation:** +The host dose not supply the CUDA drivers, **install them now:** -- `nvidia-driver-libs` and `nvidia-driver-cuda-libs` contains necessary NVIDIA driver libraries required by CUDA, - on hosts with NVIDIA drivers installed the Fedora Container will supply the host libraries. - -### Install Nvidia Driver Libraries on Guest (if `libcuda.so.1` was NOT found). +#### Install the Nvidia Driver Libraries on Guest: ```bash -sudo dnf install nvidia-driver-libs nvidia-driver-cuda-libs +sudo dnf install nvidia-driver-cuda nvidia-driver-libs nvidia-driver-cuda-libs nvidia-persistenced ``` -### Manually Updating the RPM database for host-supplied NVIDIA drivers (if `libcuda.so.1` was found). +### If *`libcuda.so.1`* exists: +``` +lrwxrwxrwx. 1 root root 21 Mar 24 11:26 /usr/lib64/libcuda.so.1 -> libcuda.so.570.133.07 +``` -If the installation fails due to conflicts, we'll manually download and install the required packages, excluding conflicting files. +**Explanation:** +The host is supply the CUDA drivers, **we need to update the guest RPM Database accordingly:** -#### 1. Download `nvidia-driver-libs` and `nvidia-driver-cuda-libs` RPM's (with dependencies) +#### Update the Toolbox RPM Database to include the Host-Supplied Libraries: + +Note: we do not actually install the libraries, we just update the DB so that the guest system knows they are supplied by the host. + +##### 1. Download `nvidia-` parts that are supplied by the host RPM's (with dependencies) ```bash -sudo dnf download --destdir=/tmp/nvidia-driver-libs --resolve --arch x86_64 nvidia-driver-libs nvidia-driver-cuda-libs +sudo dnf download --destdir=/tmp/nvidia-driver-libs --resolve --arch x86_64 nvidia-driver-cuda nvidia-driver-libs nvidia-driver-cuda-libs nvidia-persistenced ``` -#### 2. Update the RPM database to assume the installation of these packages. +##### 2. Update the RPM database to assume the installation of these packages. ```bash sudo rpm --install --verbose --hash --justdb /tmp/nvidia-driver-libs/* @@ -134,23 +144,26 @@ sudo rpm --install --verbose --hash --justdb /tmp/nvidia-driver-libs/* **Note:** -- The `--justdb` option only updates the RPM database, without touching the filesystem. +- The `--justdb` option only updates the RPM database, without touching the filesystem elsewhere. -#### Finalizing the Installation of `nvidia-driver-libs` and `nvidia-driver-cuda-libs` +##### Check that the RPM Database has been correctly updated: + +**Note:** This is the same command as in the *"Install the Nvidia Driver Libraries on Guest"* for if *`libcuda.so.1`* was missing. -After manually installing the dependencies, run: ```bash -sudo dnf install nvidia-driver-libs nvidia-driver-cuda-libs +sudo dnf install nvidia-driver-cuda nvidia-driver-libs nvidia-driver-cuda-libs nvidia-persistenced ``` -You should receive a message indicating the package is already installed: +*(this time it will not install anything, as the database things that these packages are already installed)* ``` Updating and loading repositories: Repositories loaded. -Package "nvidia-driver-libs-3:570.86.10-1.fc41.x86_64" is already installed. -Package "nvidia-driver-cuda-libs-3:570.86.10-1.fc41.x86_64" is already installed. +Package "nvidia-driver-cuda-3:570.124.06-1.fc41.x86_64" is already installed. +Package "nvidia-driver-libs-3:570.124.06-1.fc41.x86_64" is already installed. +Package "nvidia-driver-cuda-libs-3:570.124.06-1.fc41.x86_64" is already installed. +Package "nvidia-persistenced-3:570.124.06-1.fc41.x86_64" is already installed. Nothing to do. ``` @@ -207,9 +220,9 @@ You should see output similar to: ``` nvcc: NVIDIA (R) Cuda compiler driver Copyright (c) 2005-2025 NVIDIA Corporation -Built on Wed_Jan_15_19:20:09_PST_2025 -Cuda compilation tools, release 12.8, V12.8.61 -Build cuda_12.8.r12.8/compiler.35404655_0 +Built on Fri_Feb_21_20:23:50_PST_2025 +Cuda compilation tools, release 12.8, V12.8.93 +Build cuda_12.8.r12.8/compiler.35583870_0 ``` This output confirms that the CUDA compiler is accessible and indicates the installed version. @@ -248,7 +261,7 @@ You have successfully set up CUDA on Fedora within a toolbox environment using t - **Building `llama.cpp`:** - - With CUDA installed, you can follow these [build instructions for `llama.cpp`](https://github.com/ggerganov/llama.cpp/blob/master/docs/build.md) to compile it with CUDA support. + - With CUDA installed, you can follow these [build instructions for `llama.cpp`](https://github.com/ggml-org/llama.cpp/blob/master/docs/build.md) to compile it with CUDA support. - Ensure that any CUDA-specific build flags or paths are correctly set in your build configuration. - **Using the Toolbox Environment:** diff --git a/docs/backend/OPENCL.md b/docs/backend/OPENCL.md new file mode 100644 index 000000000..07146f710 --- /dev/null +++ b/docs/backend/OPENCL.md @@ -0,0 +1,209 @@ +# llama.cpp for OpenCL + +- [Background](#background) +- [OS](#os) +- [Hardware](#hardware) +- [DataType Supports](#datatype-supports) +- [Model Preparation](#model-preparation) +- [CMake Options](#cmake-options) +- [Android](#android) +- [Windows 11 Arm64](#windows-11-arm64) +- [Known Issue](#known-issues) +- [TODO](#todo) + +## Background + +OpenCL (Open Computing Language) is an open, royalty-free standard for cross-platform, parallel programming of diverse accelerators found in supercomputers, cloud servers, personal computers, mobile devices and embedded platforms. OpenCL specifies a programming language (based on C99) for programming these devices and application programming interfaces (APIs) to control the platform and execute programs on the compute devices. Similar to CUDA, OpenCL has been widely used to program GPUs and is supported by most GPU vendors. + +### Llama.cpp + OpenCL + +The llama.cpp OpenCL backend is designed to enable llama.cpp on **Qualcomm Adreno GPU** firstly via OpenCL. Thanks to the portabilty of OpenCL, the OpenCL backend can also run on certain Intel GPUs although the performance is not optimal. + +## OS + +| OS | Status | Verified | +|---------|---------|------------------------------------------------| +| Android | Support | Snapdragon 8 Gen 3, Snapdragon 8 Elite | +| Windows | Support | Windows 11 Arm64 with Snapdragon X Elite | +| Linux | Support | Ubuntu 22.04 WSL2 with Intel 12700H | + +## Hardware + +### Adreno GPU + +**Verified devices** + +| Adreno GPU | Status | +|:------------------------------------:|:-------:| +| Adreno 750 (Snapdragon 8 Gen 3) | Support | +| Adreno 830 (Snapdragon 8 Elite) | Support | +| Adreno X85 (Snapdragon X Elite) | Support | + +## DataType Supports + +| DataType | Status | +|:----------------------:|:--------------------------:| +| Q4_0 | Support | +| Q6_K | Support, but not optimized | + +## Model Preparation + +You can refer to the general [*Prepare and Quantize*](README.md#prepare-and-quantize) guide for model prepration. + +Currently we support `Q4_0` quantization and have optimize for it. To achieve best performance on Adreno GPU, add `--pure` to `llama-quantize`. For example, + +```sh +./llama-quantize --pure ggml-model-qwen2.5-3b-f16.gguf ggml-model-qwen-3b-Q4_0.gguf Q4_0 +``` + +Since `Q6_K` is also supported, `Q4_0` quantization without `--pure` will also work. However, the performance will be worse compared to pure `Q4_0` quantization. + +## CMake Options + +The OpenCL backend has the following CMake options that control the behavior of the backend. + +| CMake options | Default value | Description | +|:---------------------------------:|:--------------:|:------------------------------------------| +| `GGML_OPENCL_EMBED_KERNELS` | `ON` | Embed OpenCL kernels into the executable. | +| `GGML_OPENCL_USE_ADRENO_KERNELS` | `ON` | Use kernels optimized for Adreno. | + +## Android + +Ubuntu 22.04 is used for targeting Android. Make sure the following tools are accessible from command line, + +* Git +* CMake 3.29 +* Ninja +* Python3 + +### I. Setup Environment + +1. **Install NDK** + +```sh +cd ~ +wget https://dl.google.com/android/repository/commandlinetools-linux-8512546_latest.zip && \ +unzip commandlinetools-linux-8512546_latest.zip && \ +mkdir -p ~/android-sdk/cmdline-tools && \ +mv cmdline-tools latest && \ +mv latest ~/android-sdk/cmdline-tools/ && \ +rm -rf commandlinetools-linux-8512546_latest.zip + +yes | ~/android-sdk/cmdline-tools/latest/bin/sdkmanager "ndk;26.3.11579264" +``` + +2. **Install OpenCL Headers and Library** + +```sh +mkdir -p ~/dev/llm +cd ~/dev/llm + +git clone https://github.com/KhronosGroup/OpenCL-Headers && \ +cd OpenCL-Headers && \ +cp -r CL ~/android-sdk/ndk/26.3.11579264/toolchains/llvm/prebuilt/linux-x86_64/sysroot/usr/include + +cd ~/dev/llm + +git clone https://github.com/KhronosGroup/OpenCL-ICD-Loader && \ +cd OpenCL-ICD-Loader && \ +mkdir build_ndk26 && cd build_ndk26 && \ +cmake .. -G Ninja -DCMAKE_BUILD_TYPE=Release \ + -DCMAKE_TOOLCHAIN_FILE=$HOME/android-sdk/ndk/26.3.11579264/build/cmake/android.toolchain.cmake \ + -DOPENCL_ICD_LOADER_HEADERS_DIR=$HOME/android-sdk/ndk/26.3.11579264/toolchains/llvm/prebuilt/linux-x86_64/sysroot/usr/include \ + -DANDROID_ABI=arm64-v8a \ + -DANDROID_PLATFORM=24 \ + -DANDROID_STL=c++_shared && \ +ninja && \ +cp libOpenCL.so ~/android-sdk/ndk/26.3.11579264/toolchains/llvm/prebuilt/linux-x86_64/sysroot/usr/lib/aarch64-linux-android +``` + +### II. Build llama.cpp + +```sh +cd ~/dev/llm + +git clone https://github.com/ggml-org/llama.cpp && \ +cd llama.cpp && \ +mkdir build-android && cd build-android + +cmake .. -G Ninja \ + -DCMAKE_TOOLCHAIN_FILE=$HOME/android-sdk/ndk/26.3.11579264/build/cmake/android.toolchain.cmake \ + -DANDROID_ABI=arm64-v8a \ + -DANDROID_PLATFORM=android-28 \ + -DBUILD_SHARED_LIBS=OFF \ + -DGGML_OPENCL=ON + +ninja +``` + +## Windows 11 Arm64 + +A Snapdragon X Elite device with Windows 11 Arm64 is used. Make sure the following tools are accessible from command line, + +* Git +* CMake 3.29 +* Clang 19 +* Ninja +* Visual Studio 2022 +* Powershell 7 + +Visual Studio provides necessary headers and libraries although it is not directly used for building. +Alternatively, Visual Studio Build Tools can be installed instead of the full Visual Studio. + +Powershell 7 is used for the following commands. +If an older version of Powershell is used, these commands may not work as they are. + +### I. Setup Environment + +1. **Install OpenCL Headers and Library** + +```powershell +mkdir -p ~/dev/llm + +cd ~/dev/llm +git clone https://github.com/KhronosGroup/OpenCL-Headers && cd OpenCL-Headers +mkdir build && cd build +cmake .. -G Ninja ` + -DBUILD_TESTING=OFF ` + -DOPENCL_HEADERS_BUILD_TESTING=OFF ` + -DOPENCL_HEADERS_BUILD_CXX_TESTS=OFF ` + -DCMAKE_INSTALL_PREFIX="$HOME/dev/llm/opencl" +cmake --build . --target install + +cd ~/dev/llm +git clone https://github.com/KhronosGroup/OpenCL-ICD-Loader && cd OpenCL-ICD-Loader +mkdir build && cd build +cmake .. -G Ninja ` + -DCMAKE_BUILD_TYPE=Release ` + -DCMAKE_PREFIX_PATH="$HOME/dev/llm/opencl" ` + -DCMAKE_INSTALL_PREFIX="$HOME/dev/llm/opencl" +cmake --build . --target install +``` + +### II. Build llama.cpp + +```powershell + +mkdir -p ~/dev/llm +cd ~/dev/llm + +git clone https://github.com/ggml-org/llama.cpp && cd llama.cpp +mkdir build && cd build + +cmake .. -G Ninja ` + -DCMAKE_TOOLCHAIN_FILE="$HOME/dev/llm/llama.cpp/cmake/arm64-windows-llvm.cmake" ` + -DCMAKE_BUILD_TYPE=Release ` + -DCMAKE_PREFIX_PATH="$HOME/dev/llm/opencl" ` + -DBUILD_SHARED_LIBS=OFF ` + -DGGML_OPENCL=ON +ninja +``` + +## Known Issues + +- Currently OpenCL backend does not work on Adreno 6xx GPUs. + +## TODO + +- Optimization for Q6_K +- Support and optimization for Q4_K diff --git a/docs/backend/SYCL.md b/docs/backend/SYCL.md index 89ddbd669..20aefec2f 100644 --- a/docs/backend/SYCL.md +++ b/docs/backend/SYCL.md @@ -20,7 +20,7 @@ **oneAPI** is an open ecosystem and a standard-based specification, supporting multiple architectures including but not limited to intel CPUs, GPUs and FPGAs. The key components of the oneAPI ecosystem include: - **DPCPP** *(Data Parallel C++)*: The primary oneAPI SYCL implementation, which includes the icpx/icx Compilers. -- **oneAPI Libraries**: A set of highly optimized libraries targeting multiple domains *(e.g. oneMKL and oneDNN)*. +- **oneAPI Libraries**: A set of highly optimized libraries targeting multiple domains *(e.g. Intel oneMKL, oneMath and oneDNN)*. - **oneAPI LevelZero**: A high performance low level interface for fine-grained control over intel iGPUs and dGPUs. - **Nvidia & AMD Plugins**: These are plugins extending oneAPI's DPCPP support to SYCL on Nvidia and AMD GPU targets. @@ -36,12 +36,22 @@ The following release is verified with good quality: |Commit ID|Tag|Release|Verified Platform| Update date| |-|-|-|-|-| -|3bcd40b3c593d14261fb2abfabad3c0fb5b9e318|b4040 |[llama-b4040-bin-win-sycl-x64.zip](https://github.com/ggerganov/llama.cpp/releases/download/b4040/llama-b4040-bin-win-sycl-x64.zip) |Arc770/Linux/oneAPI 2024.1
MTL Arc GPU/Windows 11/oneAPI 2024.1| 2024-11-19| -|fb76ec31a9914b7761c1727303ab30380fd4f05c|b3038 |[llama-b3038-bin-win-sycl-x64.zip](https://github.com/ggerganov/llama.cpp/releases/download/b3038/llama-b3038-bin-win-sycl-x64.zip) |Arc770/Linux/oneAPI 2024.1
MTL Arc GPU/Windows 11/oneAPI 2024.1|| +|3bcd40b3c593d14261fb2abfabad3c0fb5b9e318|b4040 |[llama-b4040-bin-win-sycl-x64.zip](https://github.com/ggml-org/llama.cpp/releases/download/b4040/llama-b4040-bin-win-sycl-x64.zip) |Arc770/Linux/oneAPI 2024.1
MTL Arc GPU/Windows 11/oneAPI 2024.1| 2024-11-19| +|fb76ec31a9914b7761c1727303ab30380fd4f05c|b3038 |[llama-b3038-bin-win-sycl-x64.zip](https://github.com/ggml-org/llama.cpp/releases/download/b3038/llama-b3038-bin-win-sycl-x64.zip) |Arc770/Linux/oneAPI 2024.1
MTL Arc GPU/Windows 11/oneAPI 2024.1|| ## News +- 2025.2 + - Optimize MUL_MAT Q4_0 on Intel GPU for all dGPUs and built-in GPUs since MTL. Increase the performance of LLM (llama-2-7b.Q4_0.gguf) 21%-87% on Intel GPUs (MTL, ARL-H, Arc, Flex, PVC). + |GPU|Base tokens/s|Increased tokens/s|Percent| + |-|-|-|-| + |PVC 1550|39|73|+87%| + |Flex 170|39|50|+28%| + |Arc770|42|55|+30%| + |MTL|13|16|+23%| + |ARL-H|14|17|+21%| + - 2024.11 - Use syclcompat to improve the performance on some platforms. This requires to use oneAPI 2025.0 or newer. @@ -58,7 +68,7 @@ The following release is verified with good quality: - 2024.3 - Release binary files of Windows. - A blog is published: **Run LLM on all Intel GPUs Using llama.cpp**: [intel.com](https://www.intel.com/content/www/us/en/developer/articles/technical/run-llm-on-all-gpus-using-llama-cpp-artical.html) or [medium.com](https://medium.com/@jianyu_neo/run-llm-on-all-intel-gpus-using-llama-cpp-fd2e2dcbd9bd). - - New base line is ready: [tag b2437](https://github.com/ggerganov/llama.cpp/tree/b2437). + - New base line is ready: [tag b2437](https://github.com/ggml-org/llama.cpp/tree/b2437). - Support multiple cards: **--split-mode**: [none|layer]; not support [row], it's on developing. - Support to assign main GPU by **--main-gpu**, replace $GGML_SYCL_DEVICE. - Support detecting all GPUs with level-zero and same top **Max compute units**. @@ -97,8 +107,8 @@ SYCL backend supports Intel GPU Family: | Intel Data Center Max Series | Support | Max 1550, 1100 | | Intel Data Center Flex Series | Support | Flex 170 | | Intel Arc Series | Support | Arc 770, 730M, Arc A750 | -| Intel built-in Arc GPU | Support | built-in Arc GPU in Meteor Lake | -| Intel iGPU | Support | iGPU in 13700k, i5-1250P, i7-1260P, i7-1165G7 | +| Intel built-in Arc GPU | Support | built-in Arc GPU in Meteor Lake, Arrow Lake | +| Intel iGPU | Support | iGPU in 13700k,iGPU in 13400, i5-1250P, i7-1260P, i7-1165G7 | *Notes:* @@ -217,30 +227,19 @@ Upon a successful installation, SYCL is enabled for the available intel devices, **oneAPI Plugin**: In order to enable SYCL support on Nvidia GPUs, please install the [Codeplay oneAPI Plugin for Nvidia GPUs](https://developer.codeplay.com/products/oneapi/nvidia/download). User should also make sure the plugin version matches the installed base toolkit one *(previous step)* for a seamless "oneAPI on Nvidia GPU" setup. - -**oneMKL for cuBlas**: The current oneMKL releases *(shipped with the oneAPI base-toolkit)* do not contain the cuBLAS backend. A build from source of the upstream [oneMKL](https://github.com/oneapi-src/oneMKL) with the *cuBLAS* backend enabled is thus required to run it on Nvidia GPUs. +**oneDNN**: The current oneDNN releases *(shipped with the oneAPI base-toolkit)* do not include the NVIDIA backend. Therefore, oneDNN must be compiled from source to enable the NVIDIA target: ```sh -git clone https://github.com/oneapi-src/oneMKL -cd oneMKL -cmake -B buildWithCublas -DCMAKE_CXX_COMPILER=icpx -DCMAKE_C_COMPILER=icx -DENABLE_MKLGPU_BACKEND=OFF -DENABLE_MKLCPU_BACKEND=OFF -DENABLE_CUBLAS_BACKEND=ON -DTARGET_DOMAINS=blas -cmake --build buildWithCublas --config Release +git clone https://github.com/oneapi-src/oneDNN.git +cd oneDNN +cmake -GNinja -Bbuild-nvidia -DDNNL_CPU_RUNTIME=DPCPP -DDNNL_GPU_RUNTIME=DPCPP -DDNNL_GPU_VENDOR=NVIDIA -DONEDNN_BUILD_GRAPH=OFF -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx +cmake --build build-nvidia --config Release ``` - **Adding support to AMD GPUs** **oneAPI Plugin**: In order to enable SYCL support on AMD GPUs, please install the [Codeplay oneAPI Plugin for AMD GPUs](https://developer.codeplay.com/products/oneapi/amd/download). As with Nvidia GPUs, the user should also make sure the plugin version matches the installed base toolkit. -**oneMKL for rocBlas**: The current oneMKL releases *(shipped with the oneAPI base-toolkit)* doesn't contain the rocBLAS backend. A build from source of the upstream [oneMKL](https://github.com/oneapi-src/oneMKL) with the *rocBLAS* backend enabled is thus required to run it on AMD GPUs. - -```sh -git clone https://github.com/oneapi-src/oneMKL -cd oneMKL -# Find your HIPTARGET with rocminfo, under the key 'Name:' -cmake -B buildWithrocBLAS -DCMAKE_CXX_COMPILER=icpx -DCMAKE_C_COMPILER=icx -DENABLE_MKLGPU_BACKEND=OFF -DENABLE_MKLCPU_BACKEND=OFF -DENABLE_ROCBLAS_BACKEND=ON -DHIPTARGETS=${HIPTARGET} -DTARGET_DOMAINS=blas -cmake --build buildWithrocBLAS --config Release -``` - 3. **Verify installation and environment** In order to check the available SYCL devices on the machine, please use the `sycl-ls` command. @@ -303,37 +302,39 @@ cmake -B build -DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx - cmake --build build --config Release -j -v ``` +It is possible to come across some precision issues when running tests that stem from using faster +instructions, which can be circumvented by setting the environment variable `SYCL_PROGRAM_COMPILE_OPTIONS` +as `-cl-fp32-correctly-rounded-divide-sqrt` + #### Nvidia GPU -```sh -# Export relevant ENV variables -export LD_LIBRARY_PATH=/path/to/oneMKL/buildWithCublas/lib:$LD_LIBRARY_PATH -export LIBRARY_PATH=/path/to/oneMKL/buildWithCublas/lib:$LIBRARY_PATH -export CPLUS_INCLUDE_DIR=/path/to/oneMKL/buildWithCublas/include:$CPLUS_INCLUDE_DIR -export CPLUS_INCLUDE_DIR=/path/to/oneMKL/include:$CPLUS_INCLUDE_DIR +The SYCL backend depends on [oneMath](https://github.com/uxlfoundation/oneMath) for Nvidia and AMD devices. +By default it is automatically built along with the project. A specific build can be provided by setting the CMake flag `-DoneMath_DIR=/path/to/oneMath/install/lib/cmake/oneMath`. +```sh # Build LLAMA with Nvidia BLAS acceleration through SYCL # Setting GGML_SYCL_DEVICE_ARCH is optional but can improve performance GGML_SYCL_DEVICE_ARCH=sm_80 # Example architecture # Option 1: Use FP32 (recommended for better performance in most cases) -cmake -B build -DGGML_SYCL=ON -DGGML_SYCL_TARGET=NVIDIA -DGGML_SYCL_DEVICE_ARCH=${GGML_SYCL_DEVICE_ARCH} -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx +cmake -B build -DGGML_SYCL=ON -DGGML_SYCL_TARGET=NVIDIA -DGGML_SYCL_DEVICE_ARCH=${GGML_SYCL_DEVICE_ARCH} -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DDNNL_DIR=/path/to/oneDNN/build-nvidia/install/lib/cmake/dnnl # Option 2: Use FP16 -cmake -B build -DGGML_SYCL=ON -DGGML_SYCL_TARGET=NVIDIA -DGGML_SYCL_DEVICE_ARCH=${GGML_SYCL_DEVICE_ARCH} -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DGGML_SYCL_F16=ON +cmake -B build -DGGML_SYCL=ON -DGGML_SYCL_TARGET=NVIDIA -DGGML_SYCL_DEVICE_ARCH=${GGML_SYCL_DEVICE_ARCH} -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DGGML_SYCL_F16=ON -DDNNL_DIR=/path/to/oneDNN/build-nvidia/install/lib/cmake/dnnl # build all binary cmake --build build --config Release -j -v ``` +It is possible to come across some precision issues when running tests that stem from using faster +instructions, which can be circumvented by passing the `-fno-fast-math` flag to the compiler. + #### AMD GPU -```sh -# Export relevant ENV variables -export LD_LIBRARY_PATH=/path/to/oneMKL/buildWithrocBLAS/lib:$LD_LIBRARY_PATH -export LIBRARY_PATH=/path/to/oneMKL/buildWithrocBLAS/lib:$LIBRARY_PATH -export CPLUS_INCLUDE_DIR=/path/to/oneMKL/buildWithrocBLAS/include:$CPLUS_INCLUDE_DIR +The SYCL backend depends on [oneMath](https://github.com/uxlfoundation/oneMath) for Nvidia and AMD devices. +By default it is automatically built along with the project. A specific build can be provided by setting the CMake flag `-DoneMath_DIR=/path/to/oneMath/install/lib/cmake/oneMath`. +```sh # Build LLAMA with rocBLAS acceleration through SYCL ## AMD @@ -424,13 +425,13 @@ Examples: - Use device 0: ```sh -ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -m models/llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 400 -e -ngl 33 -sm none -mg 0 +ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -no-cnv -m models/llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 400 -e -ngl 33 -sm none -mg 0 ``` - Use multiple devices: ```sh -ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -m models/llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 400 -e -ngl 33 -sm layer +ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -no-cnv -m models/llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 400 -e -ngl 33 -sm layer ``` *Notes:* @@ -474,6 +475,12 @@ b. Enable oneAPI running environment: "C:\Program Files (x86)\Intel\oneAPI\setvars.bat" intel64 ``` +- if you are using Powershell, enable the runtime environment with the following: + +``` +cmd.exe "/K" '"C:\Program Files (x86)\Intel\oneAPI\setvars.bat" && powershell' +``` + c. Verify installation In the oneAPI command line, run the following to print the available SYCL devices: @@ -504,13 +511,13 @@ You could download the release package for Windows directly, which including bin Choose one of following methods to build from source code. -1. Script +#### 1. Script ```sh .\examples\sycl\win-build-sycl.bat ``` -2. CMake +#### 2. CMake On the oneAPI command line window, step into the llama.cpp main directory and run the following: @@ -539,13 +546,84 @@ cmake --preset x64-windows-sycl-debug cmake --build build-x64-windows-sycl-debug -j --target llama-cli ``` -3. Visual Studio +#### 3. Visual Studio -You can use Visual Studio to open llama.cpp folder as a CMake project. Choose the sycl CMake presets (`x64-windows-sycl-release` or `x64-windows-sycl-debug`) before you compile the project. +You have two options to use Visual Studio to build llama.cpp: +- As CMake Project using CMake presets. +- Creating a Visual Studio solution to handle the project. + +**Note**: + +All following commands are executed in PowerShell. + +##### - Open as a CMake Project + +You can use Visual Studio to open the `llama.cpp` folder directly as a CMake project. Before compiling, select one of the SYCL CMake presets: + +- `x64-windows-sycl-release` + +- `x64-windows-sycl-debug` *Notes:* +- For a minimal experimental setup, you can build only the inference executable using: -- In case of a minimal experimental setup, the user can build the inference executable only through `cmake --build build --config Release -j --target llama-cli`. + ```Powershell + cmake --build build --config Release -j --target llama-cli + ``` + +##### - Generating a Visual Studio Solution + +You can use Visual Studio solution to build and work on llama.cpp on Windows. You need to convert the CMake Project into a `.sln` file. + +If you want to use the Intel C++ Compiler for the entire `llama.cpp` project, run the following command: + +```Powershell +cmake -B build -G "Visual Studio 17 2022" -T "Intel C++ Compiler 2025" -A x64 -DGGML_SYCL=ON -DCMAKE_BUILD_TYPE=Release +``` + +If you prefer to use the Intel C++ Compiler only for `ggml-sycl`, ensure that `ggml` and its backend libraries are built as shared libraries ( i.e. `-DBUILD_SHARED_LIBRARIES=ON`, this is default behaviour): + +```Powershell +cmake -B build -G "Visual Studio 17 2022" -A x64 -DGGML_SYCL=ON -DCMAKE_BUILD_TYPE=Release \ + -DSYCL_INCLUDE_DIR="C:\Program Files (x86)\Intel\oneAPI\compiler\latest\include" \ + -DSYCL_LIBRARY_DIR="C:\Program Files (x86)\Intel\oneAPI\compiler\latest\lib" +``` + +If successful the build files have been written to: *path/to/llama.cpp/build* +Open the project file **build/llama.cpp.sln** with Visual Studio. + +Once the Visual Studio solution is created, follow these steps: + +1. Open the solution in Visual Studio. + +2. Right-click on `ggml-sycl` and select **Properties**. + +3. In the left column, expand **C/C++** and select **DPC++**. + +4. In the right panel, find **Enable SYCL Offload** and set it to `Yes`. + +5. Apply the changes and save. + + +*Navigation Path:* + +``` +Properties -> C/C++ -> DPC++ -> Enable SYCL Offload (Yes) +``` + +Now, you can build `llama.cpp` with the SYCL backend as a Visual Studio project. +To do it from menu: `Build -> Build Solution`. +Once it is completed, final results will be in **build/Release/bin** + +*Additional Note* + +- You can avoid specifying `SYCL_INCLUDE_DIR` and `SYCL_LIBRARY_DIR` in the CMake command by setting the environment variables: + + - `SYCL_INCLUDE_DIR_HINT` + + - `SYCL_LIBRARY_DIR_HINT` + +- Above instruction has been tested with Visual Studio 17 Community edition and oneAPI 2025.0. We expect them to work also with future version if the instructions are adapted accordingly. ### III. Run the inference @@ -619,13 +697,13 @@ Examples: - Use device 0: ``` -build\bin\llama-cli.exe -m models\llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:\nStep 1:" -n 400 -e -ngl 33 -s 0 -sm none -mg 0 +build\bin\llama-cli.exe -no-cnv -m models\llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:\nStep 1:" -n 400 -e -ngl 33 -s 0 -sm none -mg 0 ``` - Use multiple devices: ``` -build\bin\llama-cli.exe -m models\llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:\nStep 1:" -n 400 -e -ngl 33 -s 0 -sm layer +build\bin\llama-cli.exe -no-cnv -m models\llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:\nStep 1:" -n 400 -e -ngl 33 -s 0 -sm layer ``` @@ -650,8 +728,9 @@ use 1 SYCL GPUs: [0] with Max compute units:512 |--------------------|---------------------------------------|---------------------------------------------| | GGML_SYCL | ON (mandatory) | Enable build with SYCL code path.
FP32 path - recommended for better perforemance than FP16 on quantized model| | GGML_SYCL_TARGET | INTEL *(default)* \| NVIDIA \| AMD | Set the SYCL target device type. | -| GGML_SYCL_DEVICE_ARCH | Optional (except for AMD) | Set the SYCL device architecture, optional except for AMD. Setting the device architecture can improve the performance. See the table [--offload-arch](https://github.com/intel/llvm/blob/sycl/sycl/doc/design/OffloadDesign.md#--offload-arch) for a list of valid architectures. | +| GGML_SYCL_DEVICE_ARCH | Optional (except for AMD) | Set the SYCL device architecture, optional except for AMD. Setting the device architecture can improve the performance. See the table [--offload-arch](https://github.com/intel/llvm/blob/sycl/sycl/doc/design/OffloadDesign.md#--offload-arch) for a list of valid architectures. | | GGML_SYCL_F16 | OFF *(default)* \|ON *(optional)* | Enable FP16 build with SYCL code path. | +| GGML_SYCL_GRAPH | ON *(default)* \|OFF *(Optional)* | Enable build with [SYCL Graph extension](https://github.com/intel/llvm/blob/sycl/sycl/doc/extensions/experimental/sycl_ext_oneapi_graph.asciidoc). | | CMAKE_C_COMPILER | `icx` *(Linux)*, `icx/cl` *(Windows)* | Set `icx` compiler for SYCL code path. | | CMAKE_CXX_COMPILER | `icpx` *(Linux)*, `icx` *(Windows)* | Set `icpx/icx` compiler for SYCL code path. | @@ -660,8 +739,11 @@ use 1 SYCL GPUs: [0] with Max compute units:512 | Name | Value | Function | |-------------------|------------------|---------------------------------------------------------------------------------------------------------------------------| | GGML_SYCL_DEBUG | 0 (default) or 1 | Enable log function by macro: GGML_SYCL_DEBUG | +| GGML_SYCL_DISABLE_OPT | 0 (default) or 1 | Disable optimize features based on Intel GPU type, to compare the performance increase | +| GGML_SYCL_DISABLE_GRAPH | 0 or 1 (default) | Disable running computations through SYCL Graphs feature. Disabled by default because graph performance isn't yet better than non-graph performance. | | ZES_ENABLE_SYSMAN | 0 (default) or 1 | Support to get free memory of GPU by sycl::aspect::ext_intel_free_memory.
Recommended to use when --split-mode = layer | + ## Known Issues - `Split-mode:[row]` is not supported. diff --git a/docs/build.md b/docs/build.md index afb7a0402..3f1b04399 100644 --- a/docs/build.md +++ b/docs/build.md @@ -3,7 +3,7 @@ **To get the Code:** ```bash -git clone https://github.com/ggerganov/llama.cpp +git clone https://github.com/ggml-org/llama.cpp cd llama.cpp ``` @@ -46,7 +46,7 @@ cmake --build build --config Release ``` - Building for Windows (x86, x64 and arm64) with MSVC or clang as compilers: - - Install Visual Studio 2022, e.g. via the [Community Edition](https://visualstudio.microsoft.com/de/vs/community/). In the installer, select at least the following options (this also automatically installs the required additional tools like CMake,...): + - Install Visual Studio 2022, e.g. via the [Community Edition](https://visualstudio.microsoft.com/vs/community/). In the installer, select at least the following options (this also automatically installs the required additional tools like CMake,...): - Tab Workload: Desktop-development with C++ - Tab Components (select quickly via search): C++-_CMake_ Tools for Windows, _Git_ for Windows, C++-_Clang_ Compiler for Windows, MS-Build Support for LLVM-Toolset (clang) - Please remember to always use a Developer Command Prompt / PowerShell for VS2022 for git, build, test @@ -132,12 +132,14 @@ You may find the official downloads here: [NVIDIA developer site](https://develo #### Compile and run inside a Fedora Toolbox Container -We also have a [guide](./cuda-fedora.md) for setting up CUDA toolkit in a Fedora [toolbox container](https://containertoolbx.org/). +We also have a [guide](./backend/CUDA-FEDORA.md) for setting up CUDA toolkit in a Fedora [toolbox container](https://containertoolbx.org/). **Recommended for:** - -- ***Particularly*** *convenient* for users of [Atomic Desktops for Fedora](https://fedoraproject.org/atomic-desktops/); such as: [Silverblue](https://fedoraproject.org/atomic-desktops/silverblue/) and [Kinoite](https://fedoraproject.org/atomic-desktops/kinoite/). -- Toolbox is installed by default: [Fedora Workstation](https://fedoraproject.org/workstation/) or [Fedora KDE Plasma Desktop](https://fedoraproject.org/spins/kde). +- ***Necessary*** for users of [Atomic Desktops for Fedora](https://fedoraproject.org/atomic-desktops/); such as: [Silverblue](https://fedoraproject.org/atomic-desktops/silverblue/) and [Kinoite](https://fedoraproject.org/atomic-desktops/kinoite/). + - (there are no supported CUDA packages for these systems) +- ***Necessary*** for users that have a host that is not a: [Supported Nvidia CUDA Release Platform](https://developer.nvidia.com/cuda-downloads). + - (for example, you may have [Fedora 42 Beta](https://fedoramagazine.org/announcing-fedora-linux-42-beta/) as your your host operating system) +- ***Convenient*** For those running [Fedora Workstation](https://fedoraproject.org/workstation/) or [Fedora KDE Plasma Desktop](https://fedoraproject.org/spins/kde), and want to keep their host system clean. - *Optionally* toolbox packages are available: [Arch Linux](https://archlinux.org/), [Red Hat Enterprise Linux >= 8.5](https://www.redhat.com/en/technologies/linux-platforms/enterprise-linux), or [Ubuntu](https://ubuntu.com/download) @@ -189,7 +191,7 @@ The following compilation options are also available to tweak performance: | Option | Legal values | Default | Description | |-------------------------------|------------------------|---------|-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------| -| GGML_CUDA_FORCE_MMQ | Boolean | false | Force the use of custom matrix multiplication kernels for quantized models instead of FP16 cuBLAS even if there is no int8 tensor core implementation available (affects V100, RDNA3). MMQ kernels are enabled by default on GPUs with int8 tensor core support. With MMQ force enabled, speed for large batch sizes will be worse but VRAM consumption will be lower. | +| GGML_CUDA_FORCE_MMQ | Boolean | false | Force the use of custom matrix multiplication kernels for quantized models instead of FP16 cuBLAS even if there is no int8 tensor core implementation available (affects V100, CDNA and RDNA3+). MMQ kernels are enabled by default on GPUs with int8 tensor core support. With MMQ force enabled, speed for large batch sizes will be worse but VRAM consumption will be lower. | | GGML_CUDA_FORCE_CUBLAS | Boolean | false | Force the use of FP16 cuBLAS instead of custom matrix multiplication kernels for quantized models | | GGML_CUDA_F16 | Boolean | false | If enabled, use half-precision floating point arithmetic for the CUDA dequantization + mul mat vec kernels and for the q4_1 and q5_1 matrix matrix multiplication kernels. Can improve performance on relatively recent GPUs. | | GGML_CUDA_PEER_MAX_BATCH_SIZE | Positive integer | 128 | Maximum batch size for which to enable peer access between multiple GPUs. Peer access requires either Linux or NVLink. When using NVLink enabling peer access for larger batch sizes is potentially beneficial. | @@ -197,21 +199,54 @@ The following compilation options are also available to tweak performance: ## MUSA -This provides GPU acceleration using the MUSA cores of your Moore Threads MTT GPU. Make sure to have the MUSA SDK installed. You can download it from here: [MUSA SDK](https://developer.mthreads.com/sdk/download/musa). +This provides GPU acceleration using a Moore Threads GPU. Make sure to have the [MUSA SDK](https://developer.mthreads.com/musa/musa-sdk) installed. -- Using `CMake`: +#### Download directly from Moore Threads - ```bash - cmake -B build -DGGML_MUSA=ON +You may find the official downloads here: [Moore Threads developer site](https://developer.mthreads.com/sdk/download/musa). + +### Compilation + +```bash +cmake -B build -DGGML_MUSA=ON +cmake --build build --config Release +``` + +#### Override Compute Capability Specifications + +By default, all supported compute capabilities are enabled. To customize this behavior, you can specify the `MUSA_ARCHITECTURES` option in the CMake command: + +```bash +cmake -B build -DGGML_MUSA=ON -DMUSA_ARCHITECTURES="21" +cmake --build build --config Release +``` + +This configuration enables only compute capability `2.1` (MTT S80) during compilation, which can help reduce compilation time. + +#### Compilation options + +Most of the compilation options available for CUDA should also be available for MUSA, though they haven't been thoroughly tested yet. + +- For static builds, add `-DBUILD_SHARED_LIBS=OFF` and `-DCMAKE_POSITION_INDEPENDENT_CODE=ON`: + ``` + cmake -B build -DGGML_MUSA=ON \ + -DBUILD_SHARED_LIBS=OFF -DCMAKE_POSITION_INDEPENDENT_CODE=ON cmake --build build --config Release ``` -The environment variable [`MUSA_VISIBLE_DEVICES`](https://docs.mthreads.com/musa-sdk/musa-sdk-doc-online/programming_guide/Z%E9%99%84%E5%BD%95/) can be used to specify which GPU(s) will be used. +### Runtime MUSA environmental variables + +You may set the [musa environmental variables](https://docs.mthreads.com/musa-sdk/musa-sdk-doc-online/programming_guide/Z%E9%99%84%E5%BD%95/) at runtime. + +```bash +# Use `MUSA_VISIBLE_DEVICES` to hide the first compute device. +MUSA_VISIBLE_DEVICES="-0" ./build/bin/llama-server --model /srv/models/llama.gguf +``` + +### Unified Memory The environment variable `GGML_CUDA_ENABLE_UNIFIED_MEMORY=1` can be used to enable unified memory in Linux. This allows swapping to system RAM instead of crashing when the GPU VRAM is exhausted. -Most of the compilation options available for CUDA should also be available for MUSA, though they haven't been thoroughly tested yet. - ## HIP This provides GPU acceleration on HIP-supported AMD GPUs. @@ -227,6 +262,12 @@ You can download it from your Linux distro's package manager or from here: [ROCm On Linux it is also possible to use unified memory architecture (UMA) to share main memory between the CPU and integrated GPU by setting `-DGGML_HIP_UMA=ON`. However, this hurts performance for non-integrated GPUs (but enables working with integrated GPUs). + To enhance flash attention performance on RDNA3+ or CDNA architectures, you can utilize the rocWMMA library by enabling the `-DGGML_HIP_ROCWMMA_FATTN=ON` option. This requires rocWMMA headers to be installed on the build system. + + The rocWMMA library is included by default when installing the ROCm SDK using the `rocm` meta package provided by AMD. Alternatively, if you are not using the meta package, you can install the library using the `rocwmma-dev` or `rocwmma-devel` package, depending on your system's package manager. + + As an alternative, you can manually install the library by cloning it from the official [GitHub repository](https://github.com/ROCm/rocWMMA), checkout the corresponding version tag (e.g. `rocm-6.2.4`) and set `-DCMAKE_CXX_FLAGS="-I/library/include/"` in CMake. This also works under Windows despite not officially supported by AMD. + Note that if you get the following error: ``` clang: error: cannot find ROCm device library; provide its path via '--rocm-path' or '--rocm-device-lib-path', or pass '-nogpulib' to build without ROCm device library @@ -395,6 +436,116 @@ llama_new_context_with_model: CANN compute buffer size = 1260.81 MiB For detailed info, such as model/device supports, CANN install, please refer to [llama.cpp for CANN](./backend/CANN.md). +## Arm® KleidiAI™ +KleidiAI is a library of optimized microkernels for AI workloads, specifically designed for Arm CPUs. These microkernels enhance performance and can be enabled for use by the CPU backend. + +To enable KleidiAI, go to the llama.cpp directory and build using CMake +```bash +cmake -B build -DGGML_CPU_KLEIDIAI=ON +cmake --build build --config Release +``` +You can verify that KleidiAI is being used by running +```bash +./build/bin/llama-cli -m PATH_TO_MODEL -p "What is a car?" +``` +If KleidiAI is enabled, the ouput will contain a line similar to: +``` +load_tensors: CPU_KLEIDIAI model buffer size = 3474.00 MiB +``` +KleidiAI's microkernels implement optimized tensor operations using Arm CPU features such as dotprod, int8mm and SME. llama.cpp selects the most efficient kernel based on runtime CPU feature detection. However, on platforms that support SME, you must manually enable SME microkernels by setting the environment variable `GGML_KLEIDIAI_SME=1`. + +Depending on your build target, other higher priority backends may be enabled by default. To ensure the CPU backend is used, you must disable the higher priority backends either at compile time, e.g. -DGGML_METAL=OFF, or during run-time using the command line option `--device none`. + +## OpenCL + +This provides GPU acceleration through OpenCL on recent Adreno GPU. +More information about OpenCL backend can be found in [OPENCL.md](./backend/OPENCL.md) for more information. + +### Android + +Assume NDK is available in `$ANDROID_NDK`. First, install OpenCL headers and ICD loader library if not available, + +```sh +mkdir -p ~/dev/llm +cd ~/dev/llm + +git clone https://github.com/KhronosGroup/OpenCL-Headers && \ +cd OpenCL-Headers && \ +cp -r CL $ANDROID_NDK/toolchains/llvm/prebuilt/linux-x86_64/sysroot/usr/include + +cd ~/dev/llm + +git clone https://github.com/KhronosGroup/OpenCL-ICD-Loader && \ +cd OpenCL-ICD-Loader && \ +mkdir build_ndk && cd build_ndk && \ +cmake .. -G Ninja -DCMAKE_BUILD_TYPE=Release \ + -DCMAKE_TOOLCHAIN_FILE=$ANDROID_NDK/build/cmake/android.toolchain.cmake \ + -DOPENCL_ICD_LOADER_HEADERS_DIR=$ANDROID_NDK/toolchains/llvm/prebuilt/linux-x86_64/sysroot/usr/include \ + -DANDROID_ABI=arm64-v8a \ + -DANDROID_PLATFORM=24 \ + -DANDROID_STL=c++_shared && \ +ninja && \ +cp libOpenCL.so $ANDROID_NDK/toolchains/llvm/prebuilt/linux-x86_64/sysroot/usr/lib/aarch64-linux-android +``` + +Then build llama.cpp with OpenCL enabled, + +```sh +cd ~/dev/llm + +git clone https://github.com/ggml-org/llama.cpp && \ +cd llama.cpp && \ +mkdir build-android && cd build-android + +cmake .. -G Ninja \ + -DCMAKE_TOOLCHAIN_FILE=$ANDROID_NDK/build/cmake/android.toolchain.cmake \ + -DANDROID_ABI=arm64-v8a \ + -DANDROID_PLATFORM=android-28 \ + -DBUILD_SHARED_LIBS=OFF \ + -DGGML_OPENCL=ON + +ninja +``` + +### Windows Arm64 + +First, install OpenCL headers and ICD loader library if not available, + +```powershell +mkdir -p ~/dev/llm + +cd ~/dev/llm +git clone https://github.com/KhronosGroup/OpenCL-Headers && cd OpenCL-Headers +mkdir build && cd build +cmake .. -G Ninja ` + -DBUILD_TESTING=OFF ` + -DOPENCL_HEADERS_BUILD_TESTING=OFF ` + -DOPENCL_HEADERS_BUILD_CXX_TESTS=OFF ` + -DCMAKE_INSTALL_PREFIX="$HOME/dev/llm/opencl" +cmake --build . --target install + +cd ~/dev/llm +git clone https://github.com/KhronosGroup/OpenCL-ICD-Loader && cd OpenCL-ICD-Loader +mkdir build && cd build +cmake .. -G Ninja ` + -DCMAKE_BUILD_TYPE=Release ` + -DCMAKE_PREFIX_PATH="$HOME/dev/llm/opencl" ` + -DCMAKE_INSTALL_PREFIX="$HOME/dev/llm/opencl" +cmake --build . --target install +``` + +Then build llama.cpp with OpenCL enabled, + +```powershell +cmake .. -G Ninja ` + -DCMAKE_TOOLCHAIN_FILE="$HOME/dev/llm/llama.cpp/cmake/arm64-windows-llvm.cmake" ` + -DCMAKE_BUILD_TYPE=Release ` + -DCMAKE_PREFIX_PATH="$HOME/dev/llm/opencl" ` + -DBUILD_SHARED_LIBS=OFF ` + -DGGML_OPENCL=ON +ninja +``` + ## Android To read documentation for how to build on Android, [click here](./android.md) diff --git a/docs/development/HOWTO-add-model.md b/docs/development/HOWTO-add-model.md index 8fcd70811..78c6f7607 100644 --- a/docs/development/HOWTO-add-model.md +++ b/docs/development/HOWTO-add-model.md @@ -104,16 +104,16 @@ Note: to debug the inference graph: you can use [llama-eval-callback](/examples/ ## GGUF specification -https://github.com/ggerganov/ggml/blob/master/docs/gguf.md +https://github.com/ggml-org/ggml/blob/master/docs/gguf.md ## Resources -- YaRN RoPE scaling https://github.com/ggerganov/llama.cpp/pull/2268 -- support Baichuan serial models https://github.com/ggerganov/llama.cpp/pull/3009 -- support attention bias https://github.com/ggerganov/llama.cpp/pull/4283 -- Mixtral support https://github.com/ggerganov/llama.cpp/pull/4406 -- BERT embeddings https://github.com/ggerganov/llama.cpp/pull/5423 -- Grok-1 support https://github.com/ggerganov/llama.cpp/pull/6204 -- Command R Plus support https://github.com/ggerganov/llama.cpp/pull/6491 -- support arch DBRX https://github.com/ggerganov/llama.cpp/pull/6515 -- How to convert HuggingFace model to GGUF format https://github.com/ggerganov/llama.cpp/discussions/2948 +- YaRN RoPE scaling https://github.com/ggml-org/llama.cpp/pull/2268 +- support Baichuan serial models https://github.com/ggml-org/llama.cpp/pull/3009 +- support attention bias https://github.com/ggml-org/llama.cpp/pull/4283 +- Mixtral support https://github.com/ggml-org/llama.cpp/pull/4406 +- BERT embeddings https://github.com/ggml-org/llama.cpp/pull/5423 +- Grok-1 support https://github.com/ggml-org/llama.cpp/pull/6204 +- Command R Plus support https://github.com/ggml-org/llama.cpp/pull/6491 +- support arch DBRX https://github.com/ggml-org/llama.cpp/pull/6515 +- How to convert HuggingFace model to GGUF format https://github.com/ggml-org/llama.cpp/discussions/2948 diff --git a/docs/docker.md b/docs/docker.md index dac9a9ec1..343146dbd 100644 --- a/docs/docker.md +++ b/docs/docker.md @@ -7,21 +7,21 @@ ## Images We have three Docker images available for this project: -1. `ghcr.io/ggerganov/llama.cpp:full`: This image includes both the main executable file and the tools to convert LLaMA models into ggml and convert into 4-bit quantization. (platforms: `linux/amd64`, `linux/arm64`) -2. `ghcr.io/ggerganov/llama.cpp:light`: This image only includes the main executable file. (platforms: `linux/amd64`, `linux/arm64`) -3. `ghcr.io/ggerganov/llama.cpp:server`: This image only includes the server executable file. (platforms: `linux/amd64`, `linux/arm64`) +1. `ghcr.io/ggml-org/llama.cpp:full`: This image includes both the main executable file and the tools to convert LLaMA models into ggml and convert into 4-bit quantization. (platforms: `linux/amd64`, `linux/arm64`) +2. `ghcr.io/ggml-org/llama.cpp:light`: This image only includes the main executable file. (platforms: `linux/amd64`, `linux/arm64`) +3. `ghcr.io/ggml-org/llama.cpp:server`: This image only includes the server executable file. (platforms: `linux/amd64`, `linux/arm64`) Additionally, there the following images, similar to the above: -- `ghcr.io/ggerganov/llama.cpp:full-cuda`: Same as `full` but compiled with CUDA support. (platforms: `linux/amd64`) -- `ghcr.io/ggerganov/llama.cpp:light-cuda`: Same as `light` but compiled with CUDA support. (platforms: `linux/amd64`) -- `ghcr.io/ggerganov/llama.cpp:server-cuda`: Same as `server` but compiled with CUDA support. (platforms: `linux/amd64`) -- `ghcr.io/ggerganov/llama.cpp:full-rocm`: Same as `full` but compiled with ROCm support. (platforms: `linux/amd64`, `linux/arm64`) -- `ghcr.io/ggerganov/llama.cpp:light-rocm`: Same as `light` but compiled with ROCm support. (platforms: `linux/amd64`, `linux/arm64`) -- `ghcr.io/ggerganov/llama.cpp:server-rocm`: Same as `server` but compiled with ROCm support. (platforms: `linux/amd64`, `linux/arm64`) -- `ghcr.io/ggerganov/llama.cpp:full-musa`: Same as `full` but compiled with MUSA support. (platforms: `linux/amd64`) -- `ghcr.io/ggerganov/llama.cpp:light-musa`: Same as `light` but compiled with MUSA support. (platforms: `linux/amd64`) -- `ghcr.io/ggerganov/llama.cpp:server-musa`: Same as `server` but compiled with MUSA support. (platforms: `linux/amd64`) +- `ghcr.io/ggml-org/llama.cpp:full-cuda`: Same as `full` but compiled with CUDA support. (platforms: `linux/amd64`) +- `ghcr.io/ggml-org/llama.cpp:light-cuda`: Same as `light` but compiled with CUDA support. (platforms: `linux/amd64`) +- `ghcr.io/ggml-org/llama.cpp:server-cuda`: Same as `server` but compiled with CUDA support. (platforms: `linux/amd64`) +- `ghcr.io/ggml-org/llama.cpp:full-rocm`: Same as `full` but compiled with ROCm support. (platforms: `linux/amd64`, `linux/arm64`) +- `ghcr.io/ggml-org/llama.cpp:light-rocm`: Same as `light` but compiled with ROCm support. (platforms: `linux/amd64`, `linux/arm64`) +- `ghcr.io/ggml-org/llama.cpp:server-rocm`: Same as `server` but compiled with ROCm support. (platforms: `linux/amd64`, `linux/arm64`) +- `ghcr.io/ggml-org/llama.cpp:full-musa`: Same as `full` but compiled with MUSA support. (platforms: `linux/amd64`) +- `ghcr.io/ggml-org/llama.cpp:light-musa`: Same as `light` but compiled with MUSA support. (platforms: `linux/amd64`) +- `ghcr.io/ggml-org/llama.cpp:server-musa`: Same as `server` but compiled with MUSA support. (platforms: `linux/amd64`) The GPU enabled images are not currently tested by CI beyond being built. They are not built with any variation from the ones in the Dockerfiles defined in [.devops/](../.devops/) and the GitHub Action defined in [.github/workflows/docker.yml](../.github/workflows/docker.yml). If you need different settings (for example, a different CUDA, ROCm or MUSA library, you'll need to build the images locally for now). @@ -32,25 +32,25 @@ The easiest way to download the models, convert them to ggml and optimize them i Replace `/path/to/models` below with the actual path where you downloaded the models. ```bash -docker run -v /path/to/models:/models ghcr.io/ggerganov/llama.cpp:full --all-in-one "/models/" 7B +docker run -v /path/to/models:/models ghcr.io/ggml-org/llama.cpp:full --all-in-one "/models/" 7B ``` On completion, you are ready to play! ```bash -docker run -v /path/to/models:/models ghcr.io/ggerganov/llama.cpp:full --run -m /models/7B/ggml-model-q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 512 +docker run -v /path/to/models:/models ghcr.io/ggml-org/llama.cpp:full --run -m /models/7B/ggml-model-q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 512 ``` or with a light image: ```bash -docker run -v /path/to/models:/models ghcr.io/ggerganov/llama.cpp:light -m /models/7B/ggml-model-q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 512 +docker run -v /path/to/models:/models ghcr.io/ggml-org/llama.cpp:light -m /models/7B/ggml-model-q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 512 ``` or with a server image: ```bash -docker run -v /path/to/models:/models -p 8000:8000 ghcr.io/ggerganov/llama.cpp:server -m /models/7B/ggml-model-q4_0.gguf --port 8000 --host 0.0.0.0 -n 512 +docker run -v /path/to/models:/models -p 8000:8000 ghcr.io/ggml-org/llama.cpp:server -m /models/7B/ggml-model-q4_0.gguf --port 8000 --host 0.0.0.0 -n 512 ``` ## Docker With CUDA @@ -69,7 +69,7 @@ You may want to pass in some different `ARGS`, depending on the CUDA environment The defaults are: -- `CUDA_VERSION` set to `12.6.0` +- `CUDA_VERSION` set to `12.4.0` - `CUDA_DOCKER_ARCH` set to the cmake build default, which includes all the supported architectures The resulting images, are essentially the same as the non-CUDA images: @@ -104,7 +104,7 @@ You may want to pass in some different `ARGS`, depending on the MUSA environment The defaults are: -- `MUSA_VERSION` set to `rc3.1.0` +- `MUSA_VERSION` set to `rc3.1.1` The resulting images, are essentially the same as the non-MUSA images: diff --git a/docs/function-calling.md b/docs/function-calling.md new file mode 100644 index 000000000..c3873c3fa --- /dev/null +++ b/docs/function-calling.md @@ -0,0 +1,394 @@ +# Function Calling + +[chat.h](../common/chat.h) (https://github.com/ggml-org/llama.cpp/pull/9639) adds support for [OpenAI-style function calling](https://platform.openai.com/docs/guides/function-calling) and is used in: +- `llama-server` when started w/ `--jinja` flag +- `llama-cli` (WIP: https://github.com/ggml-org/llama.cpp/pull/11556) + +## Universal support w/ Native & Generic handlers + +Function calling is supported for all models (see https://github.com/ggml-org/llama.cpp/pull/9639): + +- Native tool call formats supported: + - Llama 3.1 / 3.3 (including builtin tools support - tool names for `wolfram_alpha`, `web_search` / `brave_search`, `code_interpreter`), Llama 3.2 + - Functionary v3.1 / v3.2 + - Hermes 2/3, Qwen 2.5 + - Qwen 2.5 Coder (WIP: https://github.com/ggml-org/llama.cpp/pull/12034) + - Mistral Nemo + - Firefunction v2 + - Command R7B + - DeepSeek R1 (WIP / seems reluctant to call any tools?) + +- Generic tool call is supported when the template isn't recognized by native format handlers (you'll see `Chat format: Generic` in the logs). + - Use `--chat-template-file` to override the template when appropriate (see examples below) + - Generic support may consume more tokens and be less efficient than a model's native format. + +
+Show some common templates and which format handler they use + +| Template | Format | +|----------|--------| +| Almawave-Velvet-14B.jinja | Hermes 2 Pro | +| AtlaAI-Selene-1-Mini-Llama-3.1-8B.jinja | Llama 3.x | +| CohereForAI-aya-expanse-8b.jinja | Generic | +| CohereForAI-c4ai-command-r-plus-default.jinja | Generic | +| CohereForAI-c4ai-command-r-plus-rag.jinja | Generic | +| CohereForAI-c4ai-command-r-plus-tool_use.jinja | Generic | +| CohereForAI-c4ai-command-r7b-12-2024-default.jinja | Command R7B (extract reasoning) | +| CohereForAI-c4ai-command-r7b-12-2024-rag.jinja | Command R7B (extract reasoning) | +| CohereForAI-c4ai-command-r7b-12-2024-tool_use.jinja | Command R7B (extract reasoning) | +| CohereForAI-c4ai-command-r7b-12-2024.jinja | Generic | +| DavieLion-Llama-3.2-1B-SPIN-iter3.jinja | Generic | +| Delta-Vector-Rei-12B.jinja | Mistral Nemo | +| EpistemeAI-Mistral-Nemo-Instruct-12B-Philosophy-Math.jinja | Mistral Nemo | +| FlofloB-83k_continued_pretraining_Qwen2.5-0.5B-Instruct_Unsloth_merged_16bit.jinja | Hermes 2 Pro | +| FlofloB-test_continued_pretraining_Phi-3-mini-4k-instruct_Unsloth_merged_16bit.jinja | Generic | +| HelpingAI-HAI-SER.jinja | Generic | +| HuggingFaceTB-SmolLM2-1.7B-Instruct.jinja | Generic | +| HuggingFaceTB-SmolLM2-135M-Instruct.jinja | Generic | +| HuggingFaceTB-SmolLM2-360M-Instruct.jinja | Generic | +| INSAIT-Institute-BgGPT-Gemma-2-27B-IT-v1.0.jinja | Generic | +| Ihor-Text2Graph-R1-Qwen2.5-0.5b.jinja | Hermes 2 Pro | +| Infinigence-Megrez-3B-Instruct.jinja | Generic | +| Josephgflowers-TinyLlama_v1.1_math_code-world-test-1.jinja | Generic | +| LGAI-EXAONE-EXAONE-3.5-2.4B-Instruct.jinja | Generic | +| LGAI-EXAONE-EXAONE-3.5-7.8B-Instruct.jinja | Generic | +| LatitudeGames-Wayfarer-12B.jinja | Generic | +| Magpie-Align-Llama-3-8B-Magpie-Align-v0.1.jinja | Generic | +| Magpie-Align-Llama-3.1-8B-Magpie-Align-v0.1.jinja | Generic | +| MaziyarPanahi-calme-3.2-instruct-78b.jinja | Generic | +| MiniMaxAI-MiniMax-Text-01.jinja | Generic | +| MiniMaxAI-MiniMax-VL-01.jinja | Generic | +| NaniDAO-deepseek-r1-qwen-2.5-32B-ablated.jinja | DeepSeek R1 (extract reasoning) | +| NexaAIDev-Octopus-v2.jinja | Generic | +| NousResearch-Hermes-2-Pro-Llama-3-8B-default.jinja | Generic | +| NousResearch-Hermes-2-Pro-Llama-3-8B-tool_use.jinja | Hermes 2 Pro | +| NousResearch-Hermes-2-Pro-Mistral-7B-default.jinja | Generic | +| NousResearch-Hermes-2-Pro-Mistral-7B-tool_use.jinja | Hermes 2 Pro | +| NousResearch-Hermes-3-Llama-3.1-70B-default.jinja | Generic | +| NousResearch-Hermes-3-Llama-3.1-70B-tool_use.jinja | Hermes 2 Pro | +| NovaSky-AI-Sky-T1-32B-Flash.jinja | Hermes 2 Pro | +| NovaSky-AI-Sky-T1-32B-Preview.jinja | Hermes 2 Pro | +| OnlyCheeini-greesychat-turbo.jinja | Generic | +| Orenguteng-Llama-3.1-8B-Lexi-Uncensored-V2.jinja | Llama 3.x | +| OrionStarAI-Orion-14B-Chat.jinja | Generic | +| PowerInfer-SmallThinker-3B-Preview.jinja | Generic | +| PrimeIntellect-INTELLECT-1-Instruct.jinja | Generic | +| Qwen-QVQ-72B-Preview.jinja | Generic | +| Qwen-QwQ-32B-Preview.jinja | Hermes 2 Pro | +| Qwen-Qwen1.5-7B-Chat.jinja | Generic | +| Qwen-Qwen2-7B-Instruct.jinja | Generic | +| Qwen-Qwen2-VL-72B-Instruct.jinja | Generic | +| Qwen-Qwen2-VL-7B-Instruct.jinja | Generic | +| Qwen-Qwen2.5-0.5B.jinja | Hermes 2 Pro | +| Qwen-Qwen2.5-1.5B-Instruct.jinja | Hermes 2 Pro | +| Qwen-Qwen2.5-14B-Instruct-1M.jinja | Hermes 2 Pro | +| Qwen-Qwen2.5-14B.jinja | Hermes 2 Pro | +| Qwen-Qwen2.5-32B-Instruct.jinja | Hermes 2 Pro | +| Qwen-Qwen2.5-32B.jinja | Hermes 2 Pro | +| Qwen-Qwen2.5-3B-Instruct.jinja | Hermes 2 Pro | +| Qwen-Qwen2.5-72B-Instruct.jinja | Hermes 2 Pro | +| Qwen-Qwen2.5-7B-Instruct-1M.jinja | Hermes 2 Pro | +| Qwen-Qwen2.5-7B-Instruct.jinja | Hermes 2 Pro | +| Qwen-Qwen2.5-7B.jinja | Hermes 2 Pro | +| Qwen-Qwen2.5-Coder-32B-Instruct.jinja | Hermes 2 Pro | +| Qwen-Qwen2.5-Coder-7B-Instruct.jinja | Hermes 2 Pro | +| Qwen-Qwen2.5-Math-1.5B.jinja | Hermes 2 Pro | +| Qwen-Qwen2.5-Math-7B-Instruct.jinja | Hermes 2 Pro | +| Qwen-Qwen2.5-VL-3B-Instruct.jinja | Hermes 2 Pro | +| Qwen-Qwen2.5-VL-72B-Instruct.jinja | Hermes 2 Pro | +| Qwen-Qwen2.5-VL-7B-Instruct.jinja | Hermes 2 Pro | +| RWKV-Red-Team-ARWKV-7B-Preview-0.1.jinja | Hermes 2 Pro | +| SakanaAI-TinySwallow-1.5B-Instruct.jinja | Hermes 2 Pro | +| SakanaAI-TinySwallow-1.5B.jinja | Hermes 2 Pro | +| Sao10K-70B-L3.3-Cirrus-x1.jinja | Llama 3.x | +| SentientAGI-Dobby-Mini-Leashed-Llama-3.1-8B.jinja | Llama 3.x | +| SentientAGI-Dobby-Mini-Unhinged-Llama-3.1-8B.jinja | Llama 3.x | +| Steelskull-L3.3-Damascus-R1.jinja | Llama 3.x | +| Steelskull-L3.3-MS-Nevoria-70b.jinja | Llama 3.x | +| Steelskull-L3.3-Nevoria-R1-70b.jinja | Llama 3.x | +| THUDM-glm-4-9b-chat.jinja | Generic | +| THUDM-glm-edge-1.5b-chat.jinja | Generic | +| Tarek07-Progenitor-V1.1-LLaMa-70B.jinja | Llama 3.x | +| TheBloke-FusionNet_34Bx2_MoE-AWQ.jinja | Generic | +| TinyLlama-TinyLlama-1.1B-Chat-v1.0.jinja | Generic | +| UCLA-AGI-Mistral7B-PairRM-SPPO-Iter3.jinja | Generic | +| ValiantLabs-Llama3.1-8B-Enigma.jinja | Llama 3.x | +| abacusai-Fewshot-Metamath-OrcaVicuna-Mistral.jinja | Generic | +| ai21labs-AI21-Jamba-1.5-Large.jinja | Generic | +| allenai-Llama-3.1-Tulu-3-405B-SFT.jinja | Generic | +| allenai-Llama-3.1-Tulu-3-405B.jinja | Generic | +| allenai-Llama-3.1-Tulu-3-8B.jinja | Generic | +| arcee-ai-Virtuoso-Lite.jinja | Hermes 2 Pro | +| arcee-ai-Virtuoso-Medium-v2.jinja | Hermes 2 Pro | +| arcee-ai-Virtuoso-Small-v2.jinja | Hermes 2 Pro | +| avemio-GRAG-NEMO-12B-ORPO-HESSIAN-AI.jinja | Generic | +| bespokelabs-Bespoke-Stratos-7B.jinja | Hermes 2 Pro | +| bfuzzy1-acheron-m1a-llama.jinja | Generic | +| bofenghuang-vigogne-2-70b-chat.jinja | Generic | +| bytedance-research-UI-TARS-72B-DPO.jinja | Generic | +| bytedance-research-UI-TARS-7B-DPO.jinja | Generic | +| bytedance-research-UI-TARS-7B-SFT.jinja | Generic | +| carsenk-phi3.5_mini_exp_825_uncensored.jinja | Generic | +| cyberagent-DeepSeek-R1-Distill-Qwen-14B-Japanese.jinja | DeepSeek R1 (extract reasoning) | +| cyberagent-DeepSeek-R1-Distill-Qwen-32B-Japanese.jinja | DeepSeek R1 (extract reasoning) | +| databricks-dbrx-instruct.jinja | Generic | +| deepseek-ai-DeepSeek-Coder-V2-Instruct.jinja | Generic | +| deepseek-ai-DeepSeek-Coder-V2-Lite-Base.jinja | Generic | +| deepseek-ai-DeepSeek-Coder-V2-Lite-Instruct.jinja | Generic | +| deepseek-ai-DeepSeek-R1-Distill-Llama-70B.jinja | DeepSeek R1 (extract reasoning) | +| deepseek-ai-DeepSeek-R1-Distill-Llama-8B.jinja | DeepSeek R1 (extract reasoning) | +| deepseek-ai-DeepSeek-R1-Distill-Qwen-1.5B.jinja | DeepSeek R1 (extract reasoning) | +| deepseek-ai-DeepSeek-R1-Distill-Qwen-14B.jinja | DeepSeek R1 (extract reasoning) | +| deepseek-ai-DeepSeek-R1-Distill-Qwen-32B.jinja | DeepSeek R1 (extract reasoning) | +| deepseek-ai-DeepSeek-R1-Distill-Qwen-7B.jinja | DeepSeek R1 (extract reasoning) | +| deepseek-ai-DeepSeek-R1-Zero.jinja | DeepSeek R1 (extract reasoning) | +| deepseek-ai-DeepSeek-R1.jinja | DeepSeek R1 (extract reasoning) | +| deepseek-ai-DeepSeek-V2-Lite.jinja | Generic | +| deepseek-ai-DeepSeek-V2.5.jinja | DeepSeek R1 (extract reasoning) | +| deepseek-ai-DeepSeek-V3.jinja | DeepSeek R1 (extract reasoning) | +| deepseek-ai-deepseek-coder-33b-instruct.jinja | Generic | +| deepseek-ai-deepseek-coder-6.7b-instruct.jinja | Generic | +| deepseek-ai-deepseek-coder-7b-instruct-v1.5.jinja | Generic | +| deepseek-ai-deepseek-llm-67b-chat.jinja | Generic | +| deepseek-ai-deepseek-llm-7b-chat.jinja | Generic | +| dicta-il-dictalm2.0-instruct.jinja | Generic | +| ehristoforu-Falcon3-8B-Franken-Basestruct.jinja | Hermes 2 Pro | +| fireworks-ai-llama-3-firefunction-v2.jinja | FireFunction v2 | +| godlikehhd-alpaca_data_sampled_ifd_new_5200.jinja | Hermes 2 Pro | +| godlikehhd-alpaca_data_score_max_0.7_2600.jinja | Hermes 2 Pro | +| google-gemma-2-27b-it.jinja | Generic | +| google-gemma-2-2b-it.jinja | Generic | +| google-gemma-2-2b-jpn-it.jinja | Generic | +| google-gemma-7b-it.jinja | Generic | +| huihui-ai-DeepSeek-R1-Distill-Llama-70B-abliterated.jinja | DeepSeek R1 (extract reasoning) | +| huihui-ai-DeepSeek-R1-Distill-Llama-8B-abliterated.jinja | DeepSeek R1 (extract reasoning) | +| huihui-ai-DeepSeek-R1-Distill-Qwen-14B-abliterated-v2.jinja | DeepSeek R1 (extract reasoning) | +| huihui-ai-DeepSeek-R1-Distill-Qwen-32B-abliterated.jinja | DeepSeek R1 (extract reasoning) | +| huihui-ai-DeepSeek-R1-Distill-Qwen-7B-abliterated-v2.jinja | DeepSeek R1 (extract reasoning) | +| huihui-ai-Qwen2.5-14B-Instruct-1M-abliterated.jinja | Hermes 2 Pro | +| ibm-granite-granite-3.1-8b-instruct.jinja | Generic | +| indischepartij-MiniCPM-3B-OpenHermes-2.5-v2.jinja | Generic | +| inflatebot-MN-12B-Mag-Mell-R1.jinja | Generic | +| jinaai-ReaderLM-v2.jinja | Generic | +| kms7530-chemeng_qwen-math-7b_24_1_100_1_nonmath.jinja | Hermes 2 Pro | +| knifeayumu-Cydonia-v1.3-Magnum-v4-22B.jinja | Mistral Nemo | +| langgptai-qwen1.5-7b-chat-sa-v0.1.jinja | Generic | +| lightblue-DeepSeek-R1-Distill-Qwen-7B-Japanese.jinja | DeepSeek R1 (extract reasoning) | +| mattshumer-Reflection-Llama-3.1-70B.jinja | Generic | +| meetkai-functionary-medium-v3.1.jinja | Functionary v3.1 Llama 3.1 | +| meetkai-functionary-medium-v3.2.jinja | Functionary v3.2 | +| meta-llama-Llama-2-7b-chat-hf.jinja | Generic | +| meta-llama-Llama-3.1-8B-Instruct.jinja | Llama 3.x | +| meta-llama-Llama-3.2-11B-Vision-Instruct.jinja | Llama 3.x | +| meta-llama-Llama-3.2-1B-Instruct.jinja | Llama 3.x | +| meta-llama-Llama-3.2-3B-Instruct.jinja | Llama 3.x | +| meta-llama-Llama-3.3-70B-Instruct.jinja | Llama 3.x | +| meta-llama-Meta-Llama-3-8B-Instruct.jinja | Generic | +| meta-llama-Meta-Llama-3.1-8B-Instruct.jinja | Llama 3.x | +| microsoft-Phi-3-medium-4k-instruct.jinja | Generic | +| microsoft-Phi-3-mini-4k-instruct.jinja | Generic | +| microsoft-Phi-3-small-8k-instruct.jinja | Generic | +| microsoft-Phi-3.5-mini-instruct.jinja | Generic | +| microsoft-Phi-3.5-vision-instruct.jinja | Generic | +| microsoft-phi-4.jinja | Generic | +| migtissera-Tess-3-Mistral-Nemo-12B.jinja | Generic | +| ministral-Ministral-3b-instruct.jinja | Generic | +| mistralai-Codestral-22B-v0.1.jinja | Generic | +| mistralai-Mistral-7B-Instruct-v0.1.jinja | Generic | +| mistralai-Mistral-7B-Instruct-v0.2.jinja | Generic | +| mistralai-Mistral-7B-Instruct-v0.3.jinja | Mistral Nemo | +| mistralai-Mistral-Large-Instruct-2407.jinja | Mistral Nemo | +| mistralai-Mistral-Large-Instruct-2411.jinja | Generic | +| mistralai-Mistral-Nemo-Instruct-2407.jinja | Mistral Nemo | +| mistralai-Mistral-Small-24B-Instruct-2501.jinja | Generic | +| mistralai-Mixtral-8x7B-Instruct-v0.1.jinja | Generic | +| mkurman-Qwen2.5-14B-DeepSeek-R1-1M.jinja | Hermes 2 Pro | +| mlabonne-AlphaMonarch-7B.jinja | Generic | +| mlx-community-Josiefied-Qwen2.5-0.5B-Instruct-abliterated-v1-float32.jinja | Hermes 2 Pro | +| mlx-community-Qwen2.5-VL-7B-Instruct-8bit.jinja | Hermes 2 Pro | +| mobiuslabsgmbh-DeepSeek-R1-ReDistill-Qwen-1.5B-v1.1.jinja | DeepSeek R1 (extract reasoning) | +| netcat420-MFANNv0.20.jinja | Generic | +| netcat420-MFANNv0.24.jinja | Generic | +| netease-youdao-Confucius-o1-14B.jinja | Hermes 2 Pro | +| nvidia-AceMath-7B-RM.jinja | Hermes 2 Pro | +| nvidia-Eagle2-1B.jinja | Hermes 2 Pro | +| nvidia-Eagle2-9B.jinja | Hermes 2 Pro | +| nvidia-Llama-3.1-Nemotron-70B-Instruct-HF.jinja | Llama 3.x | +| onnx-community-DeepSeek-R1-Distill-Qwen-1.5B-ONNX.jinja | DeepSeek R1 (extract reasoning) | +| open-thoughts-OpenThinker-7B.jinja | Hermes 2 Pro | +| openchat-openchat-3.5-0106.jinja | Generic | +| pankajmathur-orca_mini_v6_8b.jinja | Generic | +| princeton-nlp-Mistral-7B-Base-SFT-RDPO.jinja | Generic | +| princeton-nlp-Mistral-7B-Instruct-DPO.jinja | Generic | +| princeton-nlp-Mistral-7B-Instruct-RDPO.jinja | Generic | +| prithivMLmods-Bellatrix-Tiny-1.5B-R1.jinja | Hermes 2 Pro | +| prithivMLmods-Bellatrix-Tiny-1B-R1.jinja | Llama 3.x | +| prithivMLmods-Bellatrix-Tiny-1B-v3.jinja | Generic | +| prithivMLmods-Bellatrix-Tiny-3B-R1.jinja | Llama 3.x | +| prithivMLmods-Blaze-14B-xElite.jinja | Generic | +| prithivMLmods-Calcium-Opus-14B-Elite2-R1.jinja | Hermes 2 Pro | +| prithivMLmods-Calme-Ties-78B.jinja | Generic | +| prithivMLmods-Calme-Ties2-78B.jinja | Generic | +| prithivMLmods-Calme-Ties3-78B.jinja | Generic | +| prithivMLmods-ChemQwen2-vL.jinja | Generic | +| prithivMLmods-GWQ2b.jinja | Generic | +| prithivMLmods-LatexMind-2B-Codec.jinja | Generic | +| prithivMLmods-Llama-3.2-6B-AlgoCode.jinja | Llama 3.x | +| prithivMLmods-Megatron-Opus-14B-Exp.jinja | Hermes 2 Pro | +| prithivMLmods-Megatron-Opus-14B-Stock.jinja | Hermes 2 Pro | +| prithivMLmods-Megatron-Opus-7B-Exp.jinja | Hermes 2 Pro | +| prithivMLmods-Omni-Reasoner-Merged.jinja | Hermes 2 Pro | +| prithivMLmods-Omni-Reasoner4-Merged.jinja | Hermes 2 Pro | +| prithivMLmods-Primal-Opus-14B-Optimus-v1.jinja | Hermes 2 Pro | +| prithivMLmods-QwQ-Math-IO-500M.jinja | Hermes 2 Pro | +| prithivMLmods-Qwen-7B-Distill-Reasoner.jinja | DeepSeek R1 (extract reasoning) | +| prithivMLmods-Qwen2.5-1.5B-DeepSeek-R1-Instruct.jinja | Hermes 2 Pro | +| prithivMLmods-Qwen2.5-14B-DeepSeek-R1-1M.jinja | Hermes 2 Pro | +| prithivMLmods-Qwen2.5-32B-DeepSeek-R1-Instruct.jinja | Hermes 2 Pro | +| prithivMLmods-Qwen2.5-7B-DeepSeek-R1-1M.jinja | Hermes 2 Pro | +| prithivMLmods-Triangulum-v2-10B.jinja | Hermes 2 Pro | +| qingy2024-Falcon3-2x10B-MoE-Instruct.jinja | Hermes 2 Pro | +| rubenroy-Zurich-14B-GCv2-5m.jinja | Hermes 2 Pro | +| rubenroy-Zurich-7B-GCv2-5m.jinja | Hermes 2 Pro | +| silma-ai-SILMA-Kashif-2B-Instruct-v1.0.jinja | Generic | +| simplescaling-s1-32B.jinja | Hermes 2 Pro | +| sometimesanotion-Lamarck-14B-v0.7.jinja | Hermes 2 Pro | +| sonthenguyen-zephyr-sft-bnb-4bit-DPO-mtbr-180steps.jinja | Generic | +| sthenno-tempesthenno-icy-0130.jinja | Generic | +| sumink-qwft.jinja | Hermes 2 Pro | +| teknium-OpenHermes-2.5-Mistral-7B.jinja | Generic | +| thirdeyeai-elevate360m.jinja | Generic | +| tiiuae-Falcon3-10B-Instruct.jinja | Hermes 2 Pro | +| unsloth-DeepSeek-R1-Distill-Llama-8B-unsloth-bnb-4bit.jinja | DeepSeek R1 (extract reasoning) | +| unsloth-DeepSeek-R1-Distill-Llama-8B.jinja | DeepSeek R1 (extract reasoning) | +| unsloth-DeepSeek-R1.jinja | DeepSeek R1 (extract reasoning) | +| unsloth-Mistral-Small-24B-Instruct-2501-unsloth-bnb-4bit.jinja | Generic | +| upstage-solar-pro-preview-instruct.jinja | Generic | +| whyhow-ai-PatientSeek.jinja | Generic | +| xwen-team-Xwen-72B-Chat.jinja | Hermes 2 Pro | +| xwen-team-Xwen-7B-Chat.jinja | Hermes 2 Pro | + +This table can be generated with: + +```bash +./build/bin/test-chat ../minja/build/tests/*.jinja 2>/dev/null +``` + +
+ +# Usage - need tool-aware Jinja template + +First, start a server with any model, but make sure it has a tools-enabled template: you can verify this by inspecting the `chat_template` or `chat_template_tool_use` properties in `http://localhost:8080/props`). + +Here are some models known to work (w/ chat template override when needed): + +```shell +# Native support: + +llama-server --jinja -fa -hf bartowski/Qwen2.5-7B-Instruct-GGUF:Q4_K_M +llama-server --jinja -fa -hf bartowski/Mistral-Nemo-Instruct-2407-GGUF:Q6_K_L +llama-server --jinja -fa -hf bartowski/Llama-3.3-70B-Instruct-GGUF:Q4_K_M + +# Native support for DeepSeek R1 works best w/ our template override (official template is buggy, although we do work around it) + +llama-server --jinja -fa -hf bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q6_K_L \ + --chat-template-file models/templates/llama-cpp-deepseek-r1.jinja + +llama-server --jinja -fa -hf bartowski/DeepSeek-R1-Distill-Qwen-32B-GGUF:Q4_K_M \ + --chat-template-file models/templates/llama-cpp-deepseek-r1.jinja + +# Native support requires the right template for these GGUFs: + +llama-server --jinja -fa -hf bartowski/functionary-small-v3.2-GGUF:Q4_K_M + --chat-template-file models/templates/meetkai-functionary-medium-v3.2.jinja + +llama-server --jinja -fa -hf bartowski/Hermes-2-Pro-Llama-3-8B-GGUF:Q4_K_M \ + --chat-template-file models/templates/NousResearch-Hermes-2-Pro-Llama-3-8B-tool_use.jinja + +llama-server --jinja -fa -hf bartowski/Hermes-3-Llama-3.1-8B-GGUF:Q4_K_M \ + --chat-template-file models/templates/NousResearch-Hermes-3-Llama-3.1-8B-tool_use.jinja + +llama-server --jinja -fa -hf bartowski/firefunction-v2-GGUF -hff firefunction-v2-IQ1_M.gguf \ + --chat-template-file models/templates/fireworks-ai-llama-3-firefunction-v2.jinja + +llama-server --jinja -fa -hf bartowski/c4ai-command-r7b-12-2024-GGUF:Q6_K_L \ + --chat-template-file models/templates/CohereForAI-c4ai-command-r7b-12-2024-tool_use.jinja + +# Generic format support +llama-server --jinja -fa -hf bartowski/phi-4-GGUF:Q4_0 +llama-server --jinja -fa -hf bartowski/gemma-2-2b-it-GGUF:Q8_0 +llama-server --jinja -fa -hf bartowski/c4ai-command-r-v01-GGUF:Q2_K +``` + +To get the official template from original HuggingFace repos, you can use [scripts/get_chat_template.py](../scripts/get_chat_template.py) (see examples invocations in [models/templates/README.md](../models/templates/README.md)) + +> [!TIP] +> If there is no official `tool_use` Jinja template, you may want to set `--chat-template chatml` to use a default that works with many models (YMMV!), or write your own (e.g. we provide a custom [llama-cpp-deepseek-r1.jinja](../models/templates/llama-cpp-deepseek-r1.jinja) for DeepSeek R1 distills) + +Test in CLI (or with any library / software that can use OpenAI-compatible API backends): + +```bash +curl http://localhost:8080/v1/chat/completions -d '{ +"model": "gpt-3.5-turbo", +"tools": [ + { + "type":"function", + "function":{ + "name":"python", + "description":"Runs code in an ipython interpreter and returns the result of the execution after 60 seconds.", + "parameters":{ + "type":"object", + "properties":{ + "code":{ + "type":"string", + "description":"The code to run in the ipython interpreter." + } + }, + "required":["code"] + } + } + } +], +"messages": [ + { + "role": "user", + "content": "Print a hello world message with python." + } +] +}' +``` + +
+Show output + +```json +{ +"choices": [ + { + "finish_reason": "tool", + "index": 0, + "message": { + "content": null, + "tool_calls": [ + { + "name": "python", + "arguments": "{\"code\":\" \\nprint(\\\"Hello, World!\\\")\"}" + } + ], + "role": "assistant" + } + } +], +"created": 1727287211, +"model": "gpt-3.5-turbo", +"object": "chat.completion", +"usage": { + "completion_tokens": 16, + "prompt_tokens": 44, + "total_tokens": 60 +}, +"id": "chatcmpl-Htbgh9feMmGM0LEH2hmQvwsCxq3c6Ni8" +} +``` + +
diff --git a/docs/install.md b/docs/install.md index 10a568506..4971c1828 100644 --- a/docs/install.md +++ b/docs/install.md @@ -7,7 +7,14 @@ On Mac and Linux, the homebrew package manager can be used via ```sh brew install llama.cpp ``` -The formula is automatically updated with new `llama.cpp` releases. More info: https://github.com/ggerganov/llama.cpp/discussions/7668 +The formula is automatically updated with new `llama.cpp` releases. More info: https://github.com/ggml-org/llama.cpp/discussions/7668 + +## MacPorts + +```sh +sudo port install llama.cpp +``` +see also: https://ports.macports.org/port/llama.cpp/details/ ## Nix diff --git a/docs/llguidance.md b/docs/llguidance.md index 792d20704..cda787b14 100644 --- a/docs/llguidance.md +++ b/docs/llguidance.md @@ -13,13 +13,15 @@ cmake -B build -DLLAMA_LLGUIDANCE=ON make -C build -j ``` +For Windows use `cmake --build build --config Release` instead of `make`. + This requires the Rust compiler and the `cargo` tool to be [installed](https://www.rust-lang.org/tools/install). ## Interface There are no new command-line arguments or modifications to `common_params`. When enabled, grammars starting with `%llguidance` are passed to LLGuidance instead of the [current](../grammars/README.md) llama.cpp grammars. Additionally, JSON Schema requests (e.g., using the `-j` argument in `llama-cli`) are also passed to LLGuidance. -For your existing GBNF grammars, you can use [gbnf_to_lark.py script](https://github.com/guidance-ai/llguidance/blob/main/scripts/gbnf_to_lark.py) to convert them to LLGuidance Lark-like format. +For your existing GBNF grammars, you can use [gbnf_to_lark.py script](https://github.com/guidance-ai/llguidance/blob/main/python/llguidance/gbnf_to_lark.py) to convert them to LLGuidance Lark-like format. ## Performance diff --git a/examples/batched-bench/batched-bench.cpp b/examples/batched-bench/batched-bench.cpp index 0659ab6f1..0f4019293 100644 --- a/examples/batched-bench/batched-bench.cpp +++ b/examples/batched-bench/batched-bench.cpp @@ -38,7 +38,7 @@ int main(int argc, char ** argv) { llama_model_params model_params = common_model_params_to_llama(params); - llama_model * model = llama_model_load_from_file(params.model.c_str(), model_params); + llama_model * model = llama_model_load_from_file(params.model.path.c_str(), model_params); if (model == NULL) { fprintf(stderr , "%s: error: unable to load model\n" , __func__); @@ -132,7 +132,7 @@ int main(int argc, char ** argv) { const auto t_pp_start = ggml_time_us(); - llama_kv_cache_clear(ctx); + llama_kv_self_clear(ctx); if (!decode_helper(ctx, batch, ctx_params.n_batch)) { LOG_ERR("%s: llama_decode() failed\n", __func__); @@ -141,7 +141,7 @@ int main(int argc, char ** argv) { if (is_pp_shared) { for (int32_t i = 1; i < pl; ++i) { - llama_kv_cache_seq_cp(ctx, 0, i, -1, -1); + llama_kv_self_seq_cp(ctx, 0, i, -1, -1); } } diff --git a/examples/batched.swift/Sources/main.swift b/examples/batched.swift/Sources/main.swift index 55c31166c..514989e34 100644 --- a/examples/batched.swift/Sources/main.swift +++ b/examples/batched.swift/Sources/main.swift @@ -116,7 +116,7 @@ if llama_decode(context, batch) != 0 { } for i in 1 ..< n_parallel { - llama_kv_cache_seq_cp(context, 0, Int32(i), 0, batch.n_tokens) + llama_kv_self_seq_cp(context, 0, Int32(i), 0, batch.n_tokens) } if n_parallel > 1 { diff --git a/examples/batched/batched.cpp b/examples/batched/batched.cpp index 21b95ef5e..1a5de5928 100644 --- a/examples/batched/batched.cpp +++ b/examples/batched/batched.cpp @@ -41,7 +41,7 @@ int main(int argc, char ** argv) { llama_model_params model_params = common_model_params_to_llama(params); - llama_model * model = llama_model_load_from_file(params.model.c_str(), model_params); + llama_model * model = llama_model_load_from_file(params.model.path.c_str(), model_params); if (model == NULL) { LOG_ERR("%s: error: unable to load model\n" , __func__); diff --git a/examples/cvector-generator/README.md b/examples/cvector-generator/README.md index be4dd5250..6d5fd74ad 100644 --- a/examples/cvector-generator/README.md +++ b/examples/cvector-generator/README.md @@ -3,9 +3,9 @@ This example demonstrates how to generate a control vector using gguf models. Related PRs: -- [Add support for control vectors](https://github.com/ggerganov/llama.cpp/pull/5970) -- (Issue) [Generate control vector using llama.cpp](https://github.com/ggerganov/llama.cpp/issues/6880) -- [Add cvector-generator example](https://github.com/ggerganov/llama.cpp/pull/7514) +- [Add support for control vectors](https://github.com/ggml-org/llama.cpp/pull/5970) +- (Issue) [Generate control vector using llama.cpp](https://github.com/ggml-org/llama.cpp/issues/6880) +- [Add cvector-generator example](https://github.com/ggml-org/llama.cpp/pull/7514) ## Examples diff --git a/examples/cvector-generator/cvector-generator.cpp b/examples/cvector-generator/cvector-generator.cpp index 413b71d34..2a9071550 100644 --- a/examples/cvector-generator/cvector-generator.cpp +++ b/examples/cvector-generator/cvector-generator.cpp @@ -342,7 +342,7 @@ static bool cb_eval(struct ggml_tensor * t, bool ask, void * user_data) { } static bool get_hidden_layers(llama_context * ctx, std::vector & tokens) { - llama_kv_cache_clear(ctx); + llama_kv_self_clear(ctx); if (llama_decode(ctx, llama_batch_get_one(tokens.data(), tokens.size()))) { fprintf(stderr, "%s : failed to eval\n", __func__); return false; @@ -394,6 +394,8 @@ static int prepare_entries(common_params & params, train_context & ctx_train) { int main(int argc, char ** argv) { common_params params; + params.out_file = "control_vector.gguf"; + if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_CVECTOR_GENERATOR, print_usage)) { return 1; } @@ -498,7 +500,7 @@ int main(int argc, char ** argv) { } // write output vectors to gguf - export_gguf(ctx_train.v_final, params.cvector_outfile, model_hint); + export_gguf(ctx_train.v_final, params.out_file, model_hint); llama_backend_free(); diff --git a/examples/embedding/embedding.cpp b/examples/embedding/embedding.cpp index 38d22c90f..6f0890415 100644 --- a/examples/embedding/embedding.cpp +++ b/examples/embedding/embedding.cpp @@ -4,6 +4,7 @@ #include "llama.h" #include +#include #if defined(_MSC_VER) #pragma warning(disable: 4244 4267) // possible loss of data @@ -37,7 +38,7 @@ static void batch_decode(llama_context * ctx, llama_batch & batch, float * outpu const struct llama_model * model = llama_get_model(ctx); // clear previous kv_cache values (irrelevant for embeddings) - llama_kv_cache_clear(ctx); + llama_kv_self_clear(ctx); // run model LOG_INF("%s: n_tokens = %d, n_seq = %d\n", __func__, batch.n_tokens, n_seq); diff --git a/examples/export-lora/export-lora.cpp b/examples/export-lora/export-lora.cpp index 91238e4be..24dc85cf2 100644 --- a/examples/export-lora/export-lora.cpp +++ b/examples/export-lora/export-lora.cpp @@ -413,20 +413,22 @@ static void print_usage(int, char ** argv) { int main(int argc, char ** argv) { common_params params; + params.out_file = "ggml-lora-merged-f16.gguf"; + if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_EXPORT_LORA, print_usage)) { return 1; } g_verbose = (params.verbosity > 1); try { - lora_merge_ctx ctx(params.model, params.lora_adapters, params.lora_outfile, params.cpuparams.n_threads); + lora_merge_ctx ctx(params.model.path, params.lora_adapters, params.out_file, params.cpuparams.n_threads); ctx.run_merge(); } catch (const std::exception & err) { fprintf(stderr, "%s\n", err.what()); exit(EXIT_FAILURE); } - printf("done, output file is %s\n", params.lora_outfile.c_str()); + printf("done, output file is %s\n", params.out_file.c_str()); return 0; } diff --git a/examples/gguf-split/gguf-split.cpp b/examples/gguf-split/gguf-split.cpp index ef3ceb686..30e771564 100644 --- a/examples/gguf-split/gguf-split.cpp +++ b/examples/gguf-split/gguf-split.cpp @@ -408,8 +408,6 @@ static void gguf_merge(const split_params & split_params) { exit(EXIT_FAILURE); } - std::ofstream fout(split_params.output.c_str(), std::ios::binary); - fout.exceptions(std::ofstream::failbit); // fail fast on write errors auto * ctx_out = gguf_init_empty(); @@ -453,7 +451,6 @@ static void gguf_merge(const split_params & split_params) { gguf_free(ctx_gguf); ggml_free(ctx_meta); gguf_free(ctx_out); - fout.close(); exit(EXIT_FAILURE); } @@ -466,7 +463,6 @@ static void gguf_merge(const split_params & split_params) { gguf_free(ctx_gguf); ggml_free(ctx_meta); gguf_free(ctx_out); - fout.close(); exit(EXIT_FAILURE); } @@ -479,7 +475,6 @@ static void gguf_merge(const split_params & split_params) { gguf_free(ctx_gguf); ggml_free(ctx_meta); gguf_free(ctx_out); - fout.close(); exit(EXIT_FAILURE); } @@ -500,9 +495,11 @@ static void gguf_merge(const split_params & split_params) { fprintf(stderr, "\033[3Ddone\n"); } - - // placeholder for the meta data - { + std::ofstream fout; + if (!split_params.dry_run) { + fout.open(split_params.output.c_str(), std::ios::binary); + fout.exceptions(std::ofstream::failbit); // fail fast on write errors + // placeholder for the meta data auto meta_size = gguf_get_meta_size(ctx_out); ::zeros(fout, meta_size); } @@ -518,7 +515,9 @@ static void gguf_merge(const split_params & split_params) { ggml_free(ctx_metas[i]); } gguf_free(ctx_out); - fout.close(); + if (!split_params.dry_run) { + fout.close(); + } exit(EXIT_FAILURE); } fprintf(stderr, "%s: writing tensors %s ...", __func__, split_path); @@ -540,10 +539,11 @@ static void gguf_merge(const split_params & split_params) { auto offset = gguf_get_data_offset(ctx_gguf) + gguf_get_tensor_offset(ctx_gguf, i_tensor); f_input.seekg(offset); f_input.read((char *)read_data.data(), n_bytes); - - // write tensor data + padding - fout.write((const char *)read_data.data(), n_bytes); - zeros(fout, GGML_PAD(n_bytes, GGUF_DEFAULT_ALIGNMENT) - n_bytes); + if (!split_params.dry_run) { + // write tensor data + padding + fout.write((const char *)read_data.data(), n_bytes); + zeros(fout, GGML_PAD(n_bytes, GGUF_DEFAULT_ALIGNMENT) - n_bytes); + } } gguf_free(ctx_gguf); @@ -552,16 +552,15 @@ static void gguf_merge(const split_params & split_params) { fprintf(stderr, "\033[3Ddone\n"); } - { + if (!split_params.dry_run) { // go back to beginning of file and write the updated metadata fout.seekp(0); std::vector data(gguf_get_meta_size(ctx_out)); gguf_get_meta_data(ctx_out, data.data()); fout.write((const char *)data.data(), data.size()); - fout.close(); - gguf_free(ctx_out); } + gguf_free(ctx_out); fprintf(stderr, "%s: %s merged from %d split with %d tensors.\n", __func__, split_params.output.c_str(), n_split, total_tensors); diff --git a/examples/gritlm/gritlm.cpp b/examples/gritlm/gritlm.cpp index 72eb46257..539bc4d60 100644 --- a/examples/gritlm/gritlm.cpp +++ b/examples/gritlm/gritlm.cpp @@ -45,7 +45,7 @@ static std::vector> encode(llama_context * ctx, const std::ve } // clear previous kv_cache values (irrelevant for embeddings) - llama_kv_cache_clear(ctx); + llama_kv_self_clear(ctx); llama_set_embeddings(ctx, true); llama_set_causal_attn(ctx, false); @@ -102,7 +102,7 @@ static std::string generate(llama_context * ctx, llama_sampler * smpl, const std llama_token eos_token = llama_vocab_eos(vocab); - llama_kv_cache_clear(ctx); + llama_kv_self_clear(ctx); llama_set_embeddings(ctx, false); llama_set_causal_attn(ctx, true); @@ -168,7 +168,7 @@ int main(int argc, char * argv[]) { llama_backend_init(); - llama_model * model = llama_model_load_from_file(params.model.c_str(), mparams); + llama_model * model = llama_model_load_from_file(params.model.path.c_str(), mparams); // create generation context llama_context * ctx = llama_init_from_model(model, cparams); diff --git a/examples/imatrix/README.md b/examples/imatrix/README.md index 9c056986b..9aa2b2034 100644 --- a/examples/imatrix/README.md +++ b/examples/imatrix/README.md @@ -1,7 +1,7 @@ # llama.cpp/examples/imatrix -Compute an importance matrix for a model and given text dataset. Can be used during quantization to enchance the quality of the quantized models. -More information is available here: https://github.com/ggerganov/llama.cpp/pull/4861 +Compute an importance matrix for a model and given text dataset. Can be used during quantization to enhance the quality of the quantized models. +More information is available here: https://github.com/ggml-org/llama.cpp/pull/4861 ## Usage diff --git a/examples/imatrix/imatrix.cpp b/examples/imatrix/imatrix.cpp index 99056e74c..1b537407f 100644 --- a/examples/imatrix/imatrix.cpp +++ b/examples/imatrix/imatrix.cpp @@ -4,6 +4,7 @@ #include "llama.h" #include "gguf.h" +#include #include #include #include @@ -111,7 +112,7 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void * const float * data = is_host ? (const float *) src1->data : m_src1_data.data(); // this has been adapted to the new format of storing merged experts in a single 3d tensor - // ref: https://github.com/ggerganov/llama.cpp/pull/6387 + // ref: https://github.com/ggml-org/llama.cpp/pull/6387 if (t->op == GGML_OP_MUL_MAT_ID) { // ids -> [n_experts_used, n_tokens] // src1 -> [cols, n_expert_used, n_tokens] @@ -231,9 +232,6 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void * void IMatrixCollector::save_imatrix(int32_t n_chunk) const { auto fname = m_params.out_file; - if (fname.empty()) { - fname = "imatrix.gguf"; - } if (n_chunk > 0) { fname += ".at_"; @@ -584,7 +582,7 @@ static bool compute_imatrix(llama_context * ctx, const common_params & params, c const auto t_start = std::chrono::high_resolution_clock::now(); // clear the KV cache - llama_kv_cache_clear(ctx); + llama_kv_self_clear(ctx); for (int j = 0; j < num_batches; ++j) { const int batch_start = start + j * n_batch; @@ -687,6 +685,8 @@ static bool compute_imatrix(llama_context * ctx, const common_params & params, c int main(int argc, char ** argv) { common_params params; + params.out_file = "imatrix.dat" ; + params.n_ctx = 512; params.logits_all = true; params.escape = false; diff --git a/examples/infill/infill.cpp b/examples/infill/infill.cpp index 489a208b6..4e2f7b727 100644 --- a/examples/infill/infill.cpp +++ b/examples/infill/infill.cpp @@ -332,8 +332,8 @@ int main(int argc, char ** argv) { LOG_DBG("context full, swapping: n_past = %d, n_left = %d, n_ctx = %d, n_keep = %d, n_discard = %d\n", n_past, n_left, n_ctx, params.n_keep, n_discard); - llama_kv_cache_seq_rm (ctx, 0, params.n_keep + 1 , params.n_keep + n_discard + 1); - llama_kv_cache_seq_add(ctx, 0, params.n_keep + 1 + n_discard, n_past, -n_discard); + llama_kv_self_seq_rm (ctx, 0, params.n_keep + 1 , params.n_keep + n_discard + 1); + llama_kv_self_seq_add(ctx, 0, params.n_keep + 1 + n_discard, n_past, -n_discard); n_past -= n_discard; diff --git a/examples/json_schema_to_grammar.py b/examples/json_schema_to_grammar.py index fc9f0097f..55f94c0b0 100755 --- a/examples/json_schema_to_grammar.py +++ b/examples/json_schema_to_grammar.py @@ -195,7 +195,7 @@ class BuiltinRule: self.deps = deps or [] # Constraining spaces to prevent model "running away". -SPACE_RULE = '| " " | "\\n" [ \\t]{0,20}' +SPACE_RULE = '| " " | "\\n"{1,2} [ \\t]{0,20}' PRIMITIVE_RULES = { 'boolean' : BuiltinRule('("true" | "false") space', []), diff --git a/examples/llama-bench/llama-bench.cpp b/examples/llama-bench/llama-bench.cpp index 4ac19ca86..cbcbfcee8 100644 --- a/examples/llama-bench/llama-bench.cpp +++ b/examples/llama-bench/llama-bench.cpp @@ -876,8 +876,8 @@ static std::vector get_cmd_params_instances(const cmd_param struct test { static const std::string build_commit; static const int build_number; - static const std::string cpu_info; - static const std::string gpu_info; + const std::string cpu_info; + const std::string gpu_info; std::string model_filename; std::string model_type; uint64_t model_size; @@ -903,7 +903,10 @@ struct test { std::string test_time; std::vector samples_ns; - test(const cmd_params_instance & inst, const llama_model * lmodel, const llama_context * ctx) { + test(const cmd_params_instance & inst, const llama_model * lmodel, const llama_context * ctx) : + cpu_info(get_cpu_info()), + gpu_info(get_gpu_info()) { + model_filename = inst.model; char buf[128]; llama_model_desc(lmodel, buf, sizeof(buf)); @@ -1058,8 +1061,6 @@ struct test { const std::string test::build_commit = LLAMA_COMMIT; const int test::build_number = LLAMA_BUILD_NUMBER; -const std::string test::cpu_info = get_cpu_info(); -const std::string test::gpu_info = get_gpu_info(); struct printer { virtual ~printer() {} @@ -1577,7 +1578,7 @@ int main(int argc, char ** argv) { test t(inst, lmodel, ctx); - llama_kv_cache_clear(ctx); + llama_kv_self_clear(ctx); // cool off before the test if (params.delay) { @@ -1617,7 +1618,7 @@ int main(int argc, char ** argv) { } for (int i = 0; i < params.reps; i++) { - llama_kv_cache_clear(ctx); + llama_kv_self_clear(ctx); uint64_t t_start = get_time_ns(); diff --git a/examples/llama.android/llama/build.gradle.kts b/examples/llama.android/llama/build.gradle.kts index 28dbc1904..5bb647802 100644 --- a/examples/llama.android/llama/build.gradle.kts +++ b/examples/llama.android/llama/build.gradle.kts @@ -18,6 +18,7 @@ android { } externalNativeBuild { cmake { + arguments += "-DLLAMA_CURL=OFF" arguments += "-DLLAMA_BUILD_COMMON=ON" arguments += "-DGGML_LLAMAFILE=OFF" arguments += "-DCMAKE_BUILD_TYPE=Release" diff --git a/examples/llama.android/llama/src/main/cpp/CMakeLists.txt b/examples/llama.android/llama/src/main/cpp/CMakeLists.txt index 2de496574..6119fe09b 100644 --- a/examples/llama.android/llama/src/main/cpp/CMakeLists.txt +++ b/examples/llama.android/llama/src/main/cpp/CMakeLists.txt @@ -14,7 +14,7 @@ project("llama-android") #include(FetchContent) #FetchContent_Declare( # llama -# GIT_REPOSITORY https://github.com/ggerganov/llama.cpp +# GIT_REPOSITORY https://github.com/ggml-org/llama.cpp # GIT_TAG master #) diff --git a/examples/llama.android/llama/src/main/cpp/llama-android.cpp b/examples/llama.android/llama/src/main/cpp/llama-android.cpp index 2a73983a9..9654cd53c 100644 --- a/examples/llama.android/llama/src/main/cpp/llama-android.cpp +++ b/examples/llama.android/llama/src/main/cpp/llama-android.cpp @@ -194,7 +194,7 @@ Java_android_llama_cpp_LLamaAndroid_bench_1model( } batch->logits[batch->n_tokens - 1] = true; - llama_kv_cache_clear(context); + llama_kv_self_clear(context); const auto t_pp_start = ggml_time_us(); if (llama_decode(context, *batch) != 0) { @@ -206,7 +206,7 @@ Java_android_llama_cpp_LLamaAndroid_bench_1model( LOGi("Benchmark text generation (tg)"); - llama_kv_cache_clear(context); + llama_kv_self_clear(context); const auto t_tg_start = ggml_time_us(); for (i = 0; i < tg; i++) { @@ -223,7 +223,7 @@ Java_android_llama_cpp_LLamaAndroid_bench_1model( const auto t_tg_end = ggml_time_us(); - llama_kv_cache_clear(context); + llama_kv_self_clear(context); const auto t_pp = double(t_pp_end - t_pp_start) / 1000000.0; const auto t_tg = double(t_tg_end - t_tg_start) / 1000000.0; @@ -361,7 +361,7 @@ Java_android_llama_cpp_LLamaAndroid_completion_1init( const auto tokens_list = common_tokenize(context, text, true, parse_special); auto n_ctx = llama_n_ctx(context); - auto n_kv_req = tokens_list.size() + (n_len - tokens_list.size()); + auto n_kv_req = tokens_list.size() + n_len; LOGi("n_len = %d, n_ctx = %d, n_kv_req = %d", n_len, n_ctx, n_kv_req); @@ -448,5 +448,5 @@ Java_android_llama_cpp_LLamaAndroid_completion_1loop( extern "C" JNIEXPORT void JNICALL Java_android_llama_cpp_LLamaAndroid_kv_1cache_1clear(JNIEnv *, jobject, jlong context) { - llama_kv_cache_clear(reinterpret_cast(context)); + llama_kv_self_clear(reinterpret_cast(context)); } diff --git a/examples/llama.swiftui/README.md b/examples/llama.swiftui/README.md index 96cf743d4..bd7ce3774 100644 --- a/examples/llama.swiftui/README.md +++ b/examples/llama.swiftui/README.md @@ -3,9 +3,24 @@ Local inference of llama.cpp on an iPhone. This is a sample app that can be used as a starting point for more advanced projects. -For usage instructions and performance stats, check the following discussion: https://github.com/ggerganov/llama.cpp/discussions/4508 +For usage instructions and performance stats, check the following discussion: https://github.com/ggml-org/llama.cpp/discussions/4508 -![image](https://github.com/ggerganov/llama.cpp/assets/1991296/2b40284f-8421-47a2-b634-74eece09a299) + +### Building +First llama.cpp need to be built and a XCFramework needs to be created. This can be done by running +the following script from the llama.cpp project root: +```console +$ ./build-xcframework.sh +``` +Open `llama.swiftui.xcodeproj` project in Xcode and you should be able to build and run the app on +a simulator or a real device. + +To use the framework with a different project, the XCFramework can be added to the project by +adding `build-apple/llama.xcframework` by dragging and dropping it into the project navigator, or +by manually selecting the framework in the "Frameworks, Libraries, and Embedded Content" section +of the project settings. + +![image](https://github.com/ggml-org/llama.cpp/assets/1991296/2b40284f-8421-47a2-b634-74eece09a299) Video demonstration: diff --git a/examples/llama.swiftui/llama.cpp.swift/LibLlama.swift b/examples/llama.swiftui/llama.cpp.swift/LibLlama.swift index ee7141a66..f6e31abc9 100644 --- a/examples/llama.swiftui/llama.cpp.swift/LibLlama.swift +++ b/examples/llama.swiftui/llama.cpp.swift/LibLlama.swift @@ -210,7 +210,7 @@ actor LlamaContext { } batch.logits[Int(batch.n_tokens) - 1] = 1 // true - llama_kv_cache_clear(context) + llama_kv_self_clear(context) let t_pp_start = DispatchTime.now().uptimeNanoseconds / 1000; @@ -223,7 +223,7 @@ actor LlamaContext { // bench text generation - llama_kv_cache_clear(context) + llama_kv_self_clear(context) let t_tg_start = DispatchTime.now().uptimeNanoseconds / 1000; @@ -242,7 +242,7 @@ actor LlamaContext { let t_tg_end = DispatchTime.now().uptimeNanoseconds / 1000; - llama_kv_cache_clear(context) + llama_kv_self_clear(context) let t_pp = Double(t_pp_end - t_pp_start) / 1000000.0 let t_tg = Double(t_tg_end - t_tg_start) / 1000000.0 @@ -292,7 +292,7 @@ actor LlamaContext { func clear() { tokens_list.removeAll() temporary_invalid_cchars.removeAll() - llama_kv_cache_clear(context) + llama_kv_self_clear(context) } private func tokenize(text: String, add_bos: Bool) -> [llama_token] { diff --git a/examples/llama.swiftui/llama.swiftui.xcodeproj/project.pbxproj b/examples/llama.swiftui/llama.swiftui.xcodeproj/project.pbxproj index ff3d108b2..6f08fe220 100644 --- a/examples/llama.swiftui/llama.swiftui.xcodeproj/project.pbxproj +++ b/examples/llama.swiftui/llama.swiftui.xcodeproj/project.pbxproj @@ -7,7 +7,6 @@ objects = { /* Begin PBXBuildFile section */ - 1809696D2D05A39F00400EE8 /* llama in Frameworks */ = {isa = PBXBuildFile; productRef = 1809696C2D05A39F00400EE8 /* llama */; }; 549479CB2AC9E16000E0F78B /* Metal.framework in Frameworks */ = {isa = PBXBuildFile; fileRef = 549479CA2AC9E16000E0F78B /* Metal.framework */; }; 79E1D9CD2B4CD16E005F8E46 /* InputButton.swift in Sources */ = {isa = PBXBuildFile; fileRef = 79E1D9CC2B4CD16E005F8E46 /* InputButton.swift */; }; 7FA3D2B32B2EA2F600543F92 /* DownloadButton.swift in Sources */ = {isa = PBXBuildFile; fileRef = 7FA3D2B22B2EA2F600543F92 /* DownloadButton.swift */; }; @@ -18,9 +17,25 @@ 8A3F84242AC4C891005E2EE8 /* models in Resources */ = {isa = PBXBuildFile; fileRef = 8A3F84232AC4C891005E2EE8 /* models */; }; 8A907F332AC7138A006146EA /* LibLlama.swift in Sources */ = {isa = PBXBuildFile; fileRef = 8A907F322AC7134E006146EA /* LibLlama.swift */; }; 8A9F7C4D2AC332EE008AE1EA /* LlamaState.swift in Sources */ = {isa = PBXBuildFile; fileRef = 8A9F7C4C2AC332EE008AE1EA /* LlamaState.swift */; }; + DD84C9FD2D747FED007778EC /* llama.xcframework in Frameworks */ = {isa = PBXBuildFile; fileRef = DD84C9FC2D747FED007778EC /* llama.xcframework */; }; + DD84C9FE2D747FED007778EC /* llama.xcframework in Embed Frameworks */ = {isa = PBXBuildFile; fileRef = DD84C9FC2D747FED007778EC /* llama.xcframework */; settings = {ATTRIBUTES = (CodeSignOnCopy, RemoveHeadersOnCopy, ); }; }; F1FE20E22B465ECA00B45541 /* LoadCustomButton.swift in Sources */ = {isa = PBXBuildFile; fileRef = F1FE20E12B465EC900B45541 /* LoadCustomButton.swift */; }; /* End PBXBuildFile section */ +/* Begin PBXCopyFilesBuildPhase section */ + DD84C9FF2D747FED007778EC /* Embed Frameworks */ = { + isa = PBXCopyFilesBuildPhase; + buildActionMask = 2147483647; + dstPath = ""; + dstSubfolderSpec = 10; + files = ( + DD84C9FE2D747FED007778EC /* llama.xcframework in Embed Frameworks */, + ); + name = "Embed Frameworks"; + runOnlyForDeploymentPostprocessing = 0; + }; +/* End PBXCopyFilesBuildPhase section */ + /* Begin PBXFileReference section */ 549479CA2AC9E16000E0F78B /* Metal.framework */ = {isa = PBXFileReference; lastKnownFileType = wrapper.framework; name = Metal.framework; path = System/Library/Frameworks/Metal.framework; sourceTree = SDKROOT; }; 79E1D9CC2B4CD16E005F8E46 /* InputButton.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = InputButton.swift; sourceTree = ""; }; @@ -33,6 +48,7 @@ 8A3F84232AC4C891005E2EE8 /* models */ = {isa = PBXFileReference; lastKnownFileType = folder; name = models; path = llama.swiftui/Resources/models; sourceTree = ""; }; 8A907F322AC7134E006146EA /* LibLlama.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = LibLlama.swift; sourceTree = ""; }; 8A9F7C4C2AC332EE008AE1EA /* LlamaState.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = LlamaState.swift; sourceTree = ""; }; + DD84C9FC2D747FED007778EC /* llama.xcframework */ = {isa = PBXFileReference; lastKnownFileType = wrapper.xcframework; name = llama.xcframework; path = "../../build-apple/llama.xcframework"; sourceTree = ""; }; DF2D2FE72B4A59BE00FCB72D /* llama.cpp */ = {isa = PBXFileReference; lastKnownFileType = wrapper; name = llama.cpp; path = ../..; sourceTree = ""; }; F1FE20E12B465EC900B45541 /* LoadCustomButton.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = LoadCustomButton.swift; sourceTree = ""; }; /* End PBXFileReference section */ @@ -42,9 +58,9 @@ isa = PBXFrameworksBuildPhase; buildActionMask = 2147483647; files = ( - 1809696D2D05A39F00400EE8 /* llama in Frameworks */, 549479CB2AC9E16000E0F78B /* Metal.framework in Frameworks */, 8A39BE0A2AC7601100BFEB40 /* Accelerate.framework in Frameworks */, + DD84C9FD2D747FED007778EC /* llama.xcframework in Frameworks */, ); runOnlyForDeploymentPostprocessing = 0; }; @@ -86,6 +102,7 @@ 8A39BE082AC7601000BFEB40 /* Frameworks */ = { isa = PBXGroup; children = ( + DD84C9FC2D747FED007778EC /* llama.xcframework */, 549479CA2AC9E16000E0F78B /* Metal.framework */, 8A39BE092AC7601000BFEB40 /* Accelerate.framework */, ); @@ -144,6 +161,7 @@ 8A1C836F2AC328BD0096AF73 /* Sources */, 8A1C83702AC328BD0096AF73 /* Frameworks */, 8A1C83712AC328BD0096AF73 /* Resources */, + DD84C9FF2D747FED007778EC /* Embed Frameworks */, ); buildRules = ( ); @@ -151,7 +169,6 @@ ); name = llama.swiftui; packageProductDependencies = ( - 1809696C2D05A39F00400EE8 /* llama */, ); productName = llama.swiftui; productReference = 8A1C83732AC328BD0096AF73 /* llama.swiftui.app */; @@ -427,13 +444,6 @@ defaultConfigurationName = Release; }; /* End XCConfigurationList section */ - -/* Begin XCSwiftPackageProductDependency section */ - 1809696C2D05A39F00400EE8 /* llama */ = { - isa = XCSwiftPackageProductDependency; - productName = llama; - }; -/* End XCSwiftPackageProductDependency section */ }; rootObject = 8A1C836B2AC328BD0096AF73 /* Project object */; } diff --git a/examples/llama.swiftui/llama.swiftui/UI/ContentView.swift b/examples/llama.swiftui/llama.swiftui/UI/ContentView.swift index 30c2dc431..1c3cd9d2e 100644 --- a/examples/llama.swiftui/llama.swiftui/UI/ContentView.swift +++ b/examples/llama.swiftui/llama.swiftui/UI/ContentView.swift @@ -124,15 +124,26 @@ struct ContentView: View { } } }.sheet(isPresented: $showingHelp) { // Sheet for help modal - VStack(alignment: .leading) { + NavigationView { VStack(alignment: .leading) { - Text("1. Make sure the model is in GGUF Format") - .padding() - Text("2. Copy the download link of the quantized model") - .padding() + VStack(alignment: .leading) { + Text("1. Make sure the model is in GGUF Format") + .padding() + Text("2. Copy the download link of the quantized model") + .padding() + } + Spacer() } - Spacer() - } + .navigationTitle("Help") + .navigationBarTitleDisplayMode(.inline) + .toolbar { + ToolbarItem(placement: .navigationBarTrailing) { + Button("Done") { + showingHelp = false + } + } + } + } } } } diff --git a/examples/llama.vim b/examples/llama.vim index 57eb2a977..af3fd3935 100644 --- a/examples/llama.vim +++ b/examples/llama.vim @@ -39,7 +39,7 @@ " " :call llama#init() " -" more info: https://github.com/ggerganov/llama.cpp/pull/9787 +" more info: https://github.com/ggml-org/llama.cpp/pull/9787 " " colors (adjust to your liking) diff --git a/examples/llava/CMakeLists.txt b/examples/llava/CMakeLists.txt index 319effd19..2d5061de4 100644 --- a/examples/llava/CMakeLists.txt +++ b/examples/llava/CMakeLists.txt @@ -1,3 +1,5 @@ +# llava (legacy) + add_library(llava OBJECT llava.cpp llava.h @@ -22,12 +24,41 @@ if (BUILD_SHARED_LIBS) install(TARGETS llava_shared LIBRARY) endif() +# mtmd + +add_library(mtmd OBJECT + mtmd.cpp + mtmd.h + clip.cpp + clip.h + clip-impl.h + ) + +target_link_libraries(mtmd PRIVATE ggml llama ${CMAKE_THREAD_LIBS_INIT}) + +target_include_directories(mtmd PUBLIC .) +target_include_directories(mtmd PRIVATE ../..) +target_include_directories(mtmd PRIVATE ../../common) # for stb_image.h + +target_compile_features(mtmd PRIVATE cxx_std_17) + +add_library(mtmd_static STATIC $) +if (BUILD_SHARED_LIBS) + set_target_properties(mtmd PROPERTIES POSITION_INDEPENDENT_CODE ON) + target_compile_definitions(mtmd PRIVATE LLAMA_SHARED LLAMA_BUILD) + add_library(mtmd_shared SHARED $) + target_link_libraries(mtmd_shared PRIVATE ggml llama ${CMAKE_THREAD_LIBS_INIT}) + install(TARGETS mtmd_shared LIBRARY) +endif() + if (NOT MSVC) target_compile_options(llava PRIVATE -Wno-cast-qual) # stb_image.h + target_compile_options(mtmd PRIVATE -Wno-cast-qual) # stb_image.h endif() if(TARGET BUILD_INFO) add_dependencies(llava BUILD_INFO) + add_dependencies(mtmd BUILD_INFO) endif() set(TARGET llama-llava-cli) @@ -51,6 +82,13 @@ install(TARGETS ${TARGET} RUNTIME) target_link_libraries(${TARGET} PRIVATE common llava ${CMAKE_THREAD_LIBS_INIT}) target_compile_features(${TARGET} PRIVATE cxx_std_17) +set(TARGET llama-gemma3-cli) +add_executable(${TARGET} gemma3-cli.cpp) +set_target_properties(${TARGET} PROPERTIES OUTPUT_NAME llama-gemma3-cli) +install(TARGETS ${TARGET} RUNTIME) +target_link_libraries(${TARGET} PRIVATE common mtmd ${CMAKE_THREAD_LIBS_INIT}) +target_compile_features(${TARGET} PRIVATE cxx_std_17) + set(TARGET llama-llava-clip-quantize-cli) add_executable(${TARGET} clip-quantize-cli.cpp) set_target_properties(${TARGET} PROPERTIES OUTPUT_NAME llama-llava-clip-quantize-cli) diff --git a/examples/llava/README-gemma3.md b/examples/llava/README-gemma3.md new file mode 100644 index 000000000..3c25ee258 --- /dev/null +++ b/examples/llava/README-gemma3.md @@ -0,0 +1,50 @@ +# Gemma 3 vision + +> [!IMPORTANT] +> +> This is very experimental, only used for demo purpose. + +## Quick started + +You can use pre-quantized model from [ggml-org](https://huggingface.co/ggml-org)'s Hugging Face account + +```bash +# build +cmake -B build +cmake --build build --target llama-gemma3-cli + +# alternatively, install from brew (MacOS) +brew install llama.cpp + +# run it +llama-gemma3-cli -hf ggml-org/gemma-3-4b-it-GGUF +llama-gemma3-cli -hf ggml-org/gemma-3-12b-it-GGUF +llama-gemma3-cli -hf ggml-org/gemma-3-27b-it-GGUF + +# note: 1B model does not support vision +``` + +## How to get mmproj.gguf? + +```bash +cd gemma-3-4b-it +python ../llama.cpp/examples/llava/gemma3_convert_encoder_to_gguf.py . + +# output file is mmproj.gguf +``` + +## How to run it? + +What you need: +- The text model GGUF, can be converted using `convert_hf_to_gguf.py` +- The mmproj file from step above +- An image file + +```bash +# build +cmake -B build +cmake --build build --target llama-gemma3-cli + +# run it +./build/bin/llama-gemma3-cli -m {text_model}.gguf --mmproj mmproj.gguf --image your_image.jpg +``` diff --git a/examples/llava/README-granitevision.md b/examples/llava/README-granitevision.md new file mode 100644 index 000000000..f08a21cc1 --- /dev/null +++ b/examples/llava/README-granitevision.md @@ -0,0 +1,190 @@ +# Granite Vision + +Download the model and point your `GRANITE_MODEL` environment variable to the path. + +```bash +$ git clone https://huggingface.co/ibm-granite/granite-vision-3.2-2b +$ export GRANITE_MODEL=./granite-vision-3.2-2b +``` + + +### 1. Running llava surgery v2. +First, we need to run the llava surgery script as shown below: + +`python llava_surgery_v2.py -C -m $GRANITE_MODEL` + +You should see two new files (`llava.clip` and `llava.projector`) written into your model's directory, as shown below. + +```bash +$ ls $GRANITE_MODEL | grep -i llava +llava.clip +llava.projector +``` + +We should see that the projector and visual encoder get split out into the llava files. Quick check to make sure they aren't empty: +```python +import os +import torch + +MODEL_PATH = os.getenv("GRANITE_MODEL") +if not MODEL_PATH: + raise ValueError("env var GRANITE_MODEL is unset!") + +encoder_tensors = torch.load(os.path.join(MODEL_PATH, "llava.clip")) +projector_tensors = torch.load(os.path.join(MODEL_PATH, "llava.projector")) + +assert len(encoder_tensors) > 0 +assert len(projector_tensors) > 0 +``` + +If you actually inspect the `.keys()` of the loaded tensors, you should see a lot of `vision_model` tensors in the `encoder_tensors`, and 5 tensors (`'multi_modal_projector.linear_1.bias'`, `'multi_modal_projector.linear_1.weight'`, `'multi_modal_projector.linear_2.bias'`, `'multi_modal_projector.linear_2.weight'`, `'image_newline'`) in the multimodal `projector_tensors`. + + +### 2. Creating the Visual Component GGUF +Next, create a new directory to hold the visual components, and copy the llava.clip/projector files, as shown below. + +```bash +$ ENCODER_PATH=$PWD/visual_encoder +$ mkdir $ENCODER_PATH + +$ cp $GRANITE_MODEL/llava.clip $ENCODER_PATH/pytorch_model.bin +$ cp $GRANITE_MODEL/llava.projector $ENCODER_PATH/ +``` + +Now, we need to write a config for the visual encoder. In order to convert the model, be sure to use the correct `image_grid_pinpoints`, as these may vary based on the model. You can find the `image_grid_pinpoints` in `$GRANITE_MODEL/config.json`. + +```json +{ + "_name_or_path": "siglip-model", + "architectures": [ + "SiglipVisionModel" + ], + "image_grid_pinpoints": [ + [384,384], + [384,768], + [384,1152], + [384,1536], + [384,1920], + [384,2304], + [384,2688], + [384,3072], + [384,3456], + [384,3840], + [768,384], + [768,768], + [768,1152], + [768,1536], + [768,1920], + [1152,384], + [1152,768], + [1152,1152], + [1536,384], + [1536,768], + [1920,384], + [1920,768], + [2304,384], + [2688,384], + [3072,384], + [3456,384], + [3840,384] + ], + "mm_patch_merge_type": "spatial_unpad", + "hidden_size": 1152, + "image_size": 384, + "intermediate_size": 4304, + "model_type": "siglip_vision_model", + "num_attention_heads": 16, + "num_hidden_layers": 27, + "patch_size": 14, + "layer_norm_eps": 1e-6, + "hidden_act": "gelu_pytorch_tanh", + "projection_dim": 0, + "vision_feature_layer": [-24, -20, -12, -1] +} +``` + +At this point you should have something like this: +```bash +$ ls $ENCODER_PATH +config.json llava.projector pytorch_model.bin +``` + +Now convert the components to GGUF; Note that we also override the image mean/std dev to `[.5,.5,.5]` since we use the SigLIP visual encoder - in the transformers model, you can find these numbers in the `preprocessor_config.json`. +```bash +$ python convert_image_encoder_to_gguf.py \ + -m $ENCODER_PATH \ + --llava-projector $ENCODER_PATH/llava.projector \ + --output-dir $ENCODER_PATH \ + --clip-model-is-vision \ + --clip-model-is-siglip \ + --image-mean 0.5 0.5 0.5 \ + --image-std 0.5 0.5 0.5 +``` + +This will create the first GGUF file at `$ENCODER_PATH/mmproj-model-f16.gguf`; we will refer to the absolute path of this file as the `$VISUAL_GGUF_PATH.` + + +### 3. Creating the LLM GGUF. +The granite vision model contains a granite LLM as its language model. For now, the easiest way to get the GGUF for LLM is by loading the composite model in `transformers` and exporting the LLM so that it can be directly converted with the normal conversion path. + +First, set the `LLM_EXPORT_PATH` to the path to export the `transformers` LLM to. +```bash +$ export LLM_EXPORT_PATH=$PWD/granite_vision_llm +``` + +```python +import os +import transformers + +MODEL_PATH = os.getenv("GRANITE_MODEL") +if not MODEL_PATH: + raise ValueError("env var GRANITE_MODEL is unset!") + +LLM_EXPORT_PATH = os.getenv("LLM_EXPORT_PATH") +if not LLM_EXPORT_PATH: + raise ValueError("env var LLM_EXPORT_PATH is unset!") + +tokenizer = transformers.AutoTokenizer.from_pretrained(MODEL_PATH) + +# NOTE: granite vision support was added to transformers very recently (4.49); +# if you get size mismatches, your version is too old. +# If you are running with an older version, set `ignore_mismatched_sizes=True` +# as shown below; it won't be loaded correctly, but the LLM part of the model that +# we are exporting will be loaded correctly. +model = transformers.AutoModelForImageTextToText.from_pretrained(MODEL_PATH, ignore_mismatched_sizes=True) + +tokenizer.save_pretrained(LLM_EXPORT_PATH) +model.language_model.save_pretrained(LLM_EXPORT_PATH) +``` + +Now you can convert the exported LLM to GGUF with the normal converter in the root of the llama cpp project. +```bash +$ LLM_GGUF_PATH=$LLM_EXPORT_PATH/granite_llm.gguf +... +$ python convert_hf_to_gguf.py --outfile $LLM_GGUF_PATH $LLM_EXPORT_PATH +``` + + +### 4. Quantization +If you want to quantize the LLM, you can do so with `llama-quantize` as you would any other LLM. For example: +```bash +$ ./build/bin/llama-quantize $LLM_EXPORT_PATH/granite_llm.gguf $LLM_EXPORT_PATH/granite_llm_q4_k_m.gguf Q4_K_M +$ LLM_GGUF_PATH=$LLM_EXPORT_PATH/granite_llm_q4_k_m.gguf +``` + +Note that currently you cannot quantize the visual encoder because granite vision models use SigLIP as the visual encoder, which has tensor dimensions that are not divisible by 32. + + +### 5. Running the Model in Llama cpp +Build llama cpp normally; you should have a target binary named `llama-llava-cli`, which you can pass two binaries to. As an example, we pass the the llama.cpp banner. + +```bash +$ ./build/bin/llama-llava-cli -m $LLM_GGUF_PATH \ + --mmproj $VISUAL_GGUF_PATH \ + --image ./media/llama0-banner.png \ + -c 16384 \ + -p "<|system|>\nA chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions.\n<|user|>\n\\nWhat does the text in this image say?\n<|assistant|>\n" \ + --temp 0 +``` + +Sample output: `The text in the image reads "LLAMA C++ Can it run DOOM Llama?"` diff --git a/examples/llava/README-minicpmo2.6.md b/examples/llava/README-minicpmo2.6.md index 8713a43d6..48c423238 100644 --- a/examples/llava/README-minicpmo2.6.md +++ b/examples/llava/README-minicpmo2.6.md @@ -5,13 +5,25 @@ Currently, this readme only supports minicpm-omni's image capabilities, and we w Download [MiniCPM-o-2_6](https://huggingface.co/openbmb/MiniCPM-o-2_6) PyTorch model from huggingface to "MiniCPM-o-2_6" folder. + +### Build llama.cpp +Readme modification time: 20250206 + +If there are differences in usage, please refer to the official build [documentation](https://github.com/ggerganov/llama.cpp/blob/master/docs/build.md) + Clone llama.cpp: ```bash -git clone git@github.com:OpenBMB/llama.cpp.git +git clone https://github.com/ggerganov/llama.cpp cd llama.cpp -git checkout minicpm-omni ``` +Build llama.cpp using `CMake`: +```bash +cmake -B build +cmake --build build --config Release +``` + + ### Usage of MiniCPM-o 2.6 Convert PyTorch model to gguf files (You can also download the converted [gguf](https://huggingface.co/openbmb/MiniCPM-o-2_6-gguf) by us) @@ -22,25 +34,15 @@ python ./examples/llava/minicpmv-convert-image-encoder-to-gguf.py -m ../MiniCPM- python ./convert_hf_to_gguf.py ../MiniCPM-o-2_6/model # quantize int4 version -./llama-quantize ../MiniCPM-o-2_6/model/ggml-model-f16.gguf ../MiniCPM-o-2_6/model/ggml-model-Q4_K_M.gguf Q4_K_M +./build/bin/llama-quantize ../MiniCPM-o-2_6/model/ggml-model-f16.gguf ../MiniCPM-o-2_6/model/ggml-model-Q4_K_M.gguf Q4_K_M ``` -Build llama.cpp using `CMake`: -https://github.com/ggerganov/llama.cpp/blob/master/docs/build.md - -```bash -cmake -B build -cmake --build build --config Release -``` Inference on Linux or Mac -``` +```bash # run f16 version -./llama-minicpmv-cli -m ../MiniCPM-o-2_6/model/ggml-model-f16.gguf --mmproj ../MiniCPM-o-2_6/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg -p "What is in the image?" +./build/bin/llama-minicpmv-cli -m ../MiniCPM-o-2_6/model/ggml-model-f16.gguf --mmproj ../MiniCPM-o-2_6/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg -p "What is in the image?" # run quantized int4 version -./llama-minicpmv-cli -m ../MiniCPM-o-2_6/model/ggml-model-Q4_K_M.gguf --mmproj ../MiniCPM-o-2_6/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg -p "What is in the image?" - -# or run in interactive mode -./llama-minicpmv-cli -m ../MiniCPM-o-2_6/model/ggml-model-Q4_K_M.gguf --mmproj ../MiniCPM-o-2_6/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg -i +./build/bin/llama-minicpmv-cli -m ../MiniCPM-o-2_6/model/ggml-model-Q4_K_M.gguf --mmproj ../MiniCPM-o-2_6/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg -p "What is in the image?" ``` diff --git a/examples/llava/README-minicpmv2.5.md b/examples/llava/README-minicpmv2.5.md index 1c8498ff9..6bfe7abd1 100644 --- a/examples/llava/README-minicpmv2.5.md +++ b/examples/llava/README-minicpmv2.5.md @@ -4,13 +4,26 @@ Download [MiniCPM-Llama3-V-2_5](https://huggingface.co/openbmb/MiniCPM-Llama3-V-2_5) PyTorch model from huggingface to "MiniCPM-Llama3-V-2_5" folder. + +### Build llama.cpp +Readme modification time: 20250206 + +If there are differences in usage, please refer to the official build [documentation](https://github.com/ggerganov/llama.cpp/blob/master/docs/build.md) + Clone llama.cpp: ```bash -git clone https://github.com/ggerganov/llama.cpp +git clone https://github.com/ggml-org/llama.cpp cd llama.cpp ``` -### Usage +Build llama.cpp using `CMake`: +```bash +cmake -B build +cmake --build build --config Release +``` + + +### Usage of MiniCPM-Llama3-V 2.5 Convert PyTorch model to gguf files (You can also download the converted [gguf](https://huggingface.co/openbmb/MiniCPM-Llama3-V-2_5-gguf) by us) @@ -20,80 +33,15 @@ python ./examples/llava/minicpmv-convert-image-encoder-to-gguf.py -m ../MiniCPM- python ./convert_hf_to_gguf.py ../MiniCPM-Llama3-V-2_5/model # quantize int4 version -./llama-quantize ../MiniCPM-Llama3-V-2_5/model/model-8B-F16.gguf ../MiniCPM-Llama3-V-2_5/model/ggml-model-Q4_K_M.gguf Q4_K_M +./build/bin/llama-quantize ../MiniCPM-Llama3-V-2_5/model/model-8B-F16.gguf ../MiniCPM-Llama3-V-2_5/model/ggml-model-Q4_K_M.gguf Q4_K_M ``` -Build for Linux or Mac - -```bash -make -make llama-minicpmv-cli -``` Inference on Linux or Mac -``` +```bash # run f16 version -./llama-minicpmv-cli -m ../MiniCPM-Llama3-V-2_5/model/model-8B-F16.gguf --mmproj ../MiniCPM-Llama3-V-2_5/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg -p "What is in the image?" +./build/bin/llama-minicpmv-cli -m ../MiniCPM-Llama3-V-2_5/model/model-8B-F16.gguf --mmproj ../MiniCPM-Llama3-V-2_5/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg -p "What is in the image?" # run quantized int4 version -./llama-minicpmv-cli -m ../MiniCPM-Llama3-V-2_5/model/ggml-model-Q4_K_M.gguf --mmproj ../MiniCPM-Llama3-V-2_5/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg -p "What is in the image?" - -# or run in interactive mode -./llama-minicpmv-cli -m ../MiniCPM-Llama3-V-2_5/model/ggml-model-Q4_K_M.gguf --mmproj ../MiniCPM-Llama3-V-2_5/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg -i -``` - -### Android - -#### Build on Android device using Termux -We found that build on Android device would bring better runtime performance, so we recommend to build on device. - -[Termux](https://github.com/termux/termux-app#installation) is a terminal app on Android device (no root required). - -Install tools in Termux: -``` -apt update && apt upgrade -y -apt install git make cmake -``` - -It's recommended to move your model inside the `~/` directory for best performance: -``` -cd storage/downloads -mv model.gguf ~/ -``` - -#### Building the Project using Android NDK -Obtain the [Android NDK](https://developer.android.com/ndk) and then build with CMake. - -Execute the following commands on your computer to avoid downloading the NDK to your mobile. Alternatively, you can also do this in Termux: - -```bash -mkdir build-android -cd build-android -export NDK=/your_ndk_path -cmake -DCMAKE_TOOLCHAIN_FILE=$NDK/build/cmake/android.toolchain.cmake -DANDROID_ABI=arm64-v8a -DANDROID_PLATFORM=android-23 -DCMAKE_C_FLAGS=-march=armv8.4a+dotprod .. -make -``` - -Install [termux](https://github.com/termux/termux-app#installation) on your device and run `termux-setup-storage` to get access to your SD card (if Android 11+ then run the command twice). - -Finally, copy these built `llama` binaries and the model file to your device storage. Because the file permissions in the Android sdcard cannot be changed, you can copy the executable files to the `/data/data/com.termux/files/home/bin` path, and then execute the following commands in Termux to add executable permission: - -(Assumed that you have pushed the built executable files to the /sdcard/llama.cpp/bin path using `adb push`) -``` -$cp -r /sdcard/llama.cpp/bin /data/data/com.termux/files/home/ -$cd /data/data/com.termux/files/home/bin -$chmod +x ./* -``` - -Download models and push them to `/sdcard/llama.cpp/`, then move it to `/data/data/com.termux/files/home/model/` - -``` -$mv /sdcard/llama.cpp/ggml-model-Q4_K_M.gguf /data/data/com.termux/files/home/model/ -$mv /sdcard/llama.cpp/mmproj-model-f16.gguf /data/data/com.termux/files/home/model/ -``` - -Now, you can start chatting: -``` -$cd /data/data/com.termux/files/home/bin -$./llama-minicpmv-cli -m ../model/ggml-model-Q4_K_M.gguf --mmproj ../model/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg -p "What is in the image?" +./build/bin/llama-minicpmv-cli -m ../MiniCPM-Llama3-V-2_5/model/ggml-model-Q4_K_M.gguf --mmproj ../MiniCPM-Llama3-V-2_5/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg -p "What is in the image?" ``` diff --git a/examples/llava/README-minicpmv2.6.md b/examples/llava/README-minicpmv2.6.md index c4be5e5dd..2df39cdba 100644 --- a/examples/llava/README-minicpmv2.6.md +++ b/examples/llava/README-minicpmv2.6.md @@ -4,13 +4,25 @@ Download [MiniCPM-V-2_6](https://huggingface.co/openbmb/MiniCPM-V-2_6) PyTorch model from huggingface to "MiniCPM-V-2_6" folder. + +### Build llama.cpp +Readme modification time: 20250206 + +If there are differences in usage, please refer to the official build [documentation](https://github.com/ggerganov/llama.cpp/blob/master/docs/build.md) + Clone llama.cpp: ```bash -git clone git@github.com:OpenBMB/llama.cpp.git +git clone https://github.com/ggerganov/llama.cpp cd llama.cpp -git checkout minicpmv-main ``` +Build llama.cpp using `CMake`: +```bash +cmake -B build +cmake --build build --config Release +``` + + ### Usage of MiniCPM-V 2.6 Convert PyTorch model to gguf files (You can also download the converted [gguf](https://huggingface.co/openbmb/MiniCPM-V-2_6-gguf) by us) @@ -21,87 +33,15 @@ python ./examples/llava/minicpmv-convert-image-encoder-to-gguf.py -m ../MiniCPM- python ./convert_hf_to_gguf.py ../MiniCPM-V-2_6/model # quantize int4 version -./llama-quantize ../MiniCPM-V-2_6/model/ggml-model-f16.gguf ../MiniCPM-V-2_6/model/ggml-model-Q4_K_M.gguf Q4_K_M +./build/bin/llama-quantize ../MiniCPM-V-2_6/model/ggml-model-f16.gguf ../MiniCPM-V-2_6/model/ggml-model-Q4_K_M.gguf Q4_K_M ``` -Build for Linux or Mac - -```bash -make -make llama-minicpmv-cli -``` Inference on Linux or Mac -``` +```bash # run f16 version -./llama-minicpmv-cli -m ../MiniCPM-V-2_6/model/ggml-model-f16.gguf --mmproj ../MiniCPM-V-2_6/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg -p "What is in the image?" +./build/bin/llama-minicpmv-cli -m ../MiniCPM-V-2_6/model/ggml-model-f16.gguf --mmproj ../MiniCPM-V-2_6/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg -p "What is in the image?" # run quantized int4 version -./llama-minicpmv-cli -m ../MiniCPM-V-2_6/model/ggml-model-Q4_K_M.gguf --mmproj ../MiniCPM-V-2_6/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg -p "What is in the image?" - -# or run in interactive mode -./llama-minicpmv-cli -m ../MiniCPM-V-2_6/model/ggml-model-Q4_K_M.gguf --mmproj ../MiniCPM-V-2_6/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg -i -``` - -### Video -Install FFmpeg -``` -brew install ffmpeg -brew install pkg-config -``` - -### Android - -#### Build on Android device using Termux -We found that build on Android device would bring better runtime performance, so we recommend to build on device. - -[Termux](https://github.com/termux/termux-app#installation) is a terminal app on Android device (no root required). - -Install tools in Termux: -``` -apt update && apt upgrade -y -apt install git make cmake -``` - -It's recommended to move your model inside the `~/` directory for best performance: -``` -cd storage/downloads -mv model.gguf ~/ -``` - -#### Building the Project using Android NDK -Obtain the [Android NDK](https://developer.android.com/ndk) and then build with CMake. - -Execute the following commands on your computer to avoid downloading the NDK to your mobile. Alternatively, you can also do this in Termux: - -```bash -mkdir build-android -cd build-android -export NDK=/your_ndk_path -cmake -DCMAKE_TOOLCHAIN_FILE=$NDK/build/cmake/android.toolchain.cmake -DANDROID_ABI=arm64-v8a -DANDROID_PLATFORM=android-23 -DCMAKE_C_FLAGS=-march=armv8.4a+dotprod .. -make -``` - -Install [termux](https://github.com/termux/termux-app#installation) on your device and run `termux-setup-storage` to get access to your SD card (if Android 11+ then run the command twice). - -Finally, copy these built `llama` binaries and the model file to your device storage. Because the file permissions in the Android sdcard cannot be changed, you can copy the executable files to the `/data/data/com.termux/files/home/bin` path, and then execute the following commands in Termux to add executable permission: - -(Assumed that you have pushed the built executable files to the /sdcard/llama.cpp/bin path using `adb push`) -``` -$cp -r /sdcard/llama.cpp/bin /data/data/com.termux/files/home/ -$cd /data/data/com.termux/files/home/bin -$chmod +x ./* -``` - -Download models and push them to `/sdcard/llama.cpp/`, then move it to `/data/data/com.termux/files/home/model/` - -``` -$mv /sdcard/llama.cpp/ggml-model-Q4_K_M.gguf /data/data/com.termux/files/home/model/ -$mv /sdcard/llama.cpp/mmproj-model-f16.gguf /data/data/com.termux/files/home/model/ -``` - -Now, you can start chatting: -``` -$cd /data/data/com.termux/files/home/bin -$./llama-minicpmv-cli -m ../model/ggml-model-Q4_K_M.gguf --mmproj ../model/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg -p "What is in the image?" +./build/bin/llama-minicpmv-cli -m ../MiniCPM-V-2_6/model/ggml-model-Q4_K_M.gguf --mmproj ../MiniCPM-V-2_6/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg -p "What is in the image?" ``` diff --git a/examples/llava/README.md b/examples/llava/README.md index 012451361..0e3c32032 100644 --- a/examples/llava/README.md +++ b/examples/llava/README.md @@ -101,8 +101,27 @@ python ./examples/convert_legacy_llama.py ../llava-v1.6-vicuna-7b/ --skip-unknow ``` **note** llava-1.6 needs more context than llava-1.5, at least 3000 is needed (just run it at -c 4096) + **note** llava-1.6 greatly benefits from batched prompt processing (defaults work) +**note** if the language model in step `6)` is incompatible with the legacy conversion script, the easiest way handle the LLM model conversion is to load the model in transformers, and export only the LLM from the llava next model. + +```python +import os +import transformers + +model_path = ... +llm_export_path = ... + +tokenizer = transformers.AutoTokenizer.from_pretrained(model_path) +model = transformers.AutoModelForImageTextToText.from_pretrained(model_path) + +tokenizer.save_pretrained(llm_export_path) +model.language_model.save_pretrained(llm_export_path) +``` + +Then, you can convert the LLM using the `convert_hf_to_gguf.py` script, which handles more LLM architectures. + ## llava-cli templating and llava-1.6 prompting llava-1.5 models all use the same vicuna prompt, here you can just add your image question like `-p "Provide a full description."` diff --git a/examples/llava/clip-impl.h b/examples/llava/clip-impl.h new file mode 100644 index 000000000..4d7340a56 --- /dev/null +++ b/examples/llava/clip-impl.h @@ -0,0 +1,344 @@ +#include "ggml.h" +#include "gguf.h" +#include "clip.h" + +#include "clip.h" + +#include +#include +#include +#include +#include +#include +#include + +// Internal header for clip.cpp + +#define KEY_FTYPE "general.file_type" +#define KEY_NAME "general.name" +#define KEY_DESCRIPTION "general.description" +#define KEY_HAS_TEXT_ENC "clip.has_text_encoder" +#define KEY_HAS_VIS_ENC "clip.has_vision_encoder" +#define KEY_HAS_LLAVA_PROJ "clip.has_llava_projector" +#define KEY_HAS_MINICPMV_PROJ "clip.has_minicpmv_projector" +#define KEY_HAS_GLM_PROJ "clip.has_glm_projector" +#define KEY_MINICPMV_VERSION "clip.minicpmv_version" +#define KEY_HAS_QWEN2VL_MERGER "clip.has_qwen2vl_merger" +#define KEY_USE_GELU "clip.use_gelu" +#define KEY_USE_SILU "clip.use_silu" +#define KEY_N_EMBD "clip.%s.embedding_length" +#define KEY_N_FF "clip.%s.feed_forward_length" +#define KEY_N_BLOCK "clip.%s.block_count" +#define KEY_N_HEAD "clip.%s.attention.head_count" +#define KEY_LAYER_NORM_EPS "clip.%s.attention.layer_norm_epsilon" +#define KEY_PROJ_DIM "clip.%s.projection_dim" +#define KEY_TOKENS "tokenizer.ggml.tokens" +#define KEY_N_POSITIONS "clip.text.context_length" +#define KEY_IMAGE_SIZE "clip.vision.image_size" +#define KEY_PATCH_SIZE "clip.vision.patch_size" +#define KEY_IMAGE_MEAN "clip.vision.image_mean" +#define KEY_IMAGE_STD "clip.vision.image_std" +#define KEY_PROJ_TYPE "clip.projector_type" +#define KEY_FEATURE_LAYER "clip.vision.feature_layer" + +#define KEY_MM_PATCH_MERGE_TYPE "clip.vision.mm_patch_merge_type" +#define KEY_IMAGE_GRID_PINPOINTS "clip.vision.image_grid_pinpoints" +#define KEY_IMAGE_CROP_RESOLUTION "clip.vision.image_crop_resolution" + + +// +// tensor name constants +// + +#define TN_TOKEN_EMBD "%s.token_embd.weight" +#define TN_POS_EMBD "%s.position_embd.weight" +#define TN_CLASS_EMBD "v.class_embd" +#define TN_PATCH_EMBD "v.patch_embd.weight" // not rename tensor with ".0" postfix for backwrad compat +#define TN_PATCH_EMBD_1 "v.patch_embd.weight.1" +#define TN_PATCH_BIAS "v.patch_embd.bias" +#define TN_ATTN_K "%s.blk.%d.attn_k.%s" +#define TN_ATTN_Q "%s.blk.%d.attn_q.%s" +#define TN_ATTN_V "%s.blk.%d.attn_v.%s" +#define TN_ATTN_OUTPUT "%s.blk.%d.attn_out.%s" +#define TN_FFN_DOWN "%s.blk.%d.ffn_down.%s" +#define TN_FFN_UP "%s.blk.%d.ffn_up.%s" +#define TN_LN_1 "%s.blk.%d.ln1.%s" +#define TN_LN_2 "%s.blk.%d.ln2.%s" +#define TN_LN_PRE "%s.pre_ln.%s" +#define TN_LN_POST "%s.post_ln.%s" +#define TN_TEXT_PROJ "text_projection.weight" +#define TN_VIS_PROJ "visual_projection.weight" +#define TN_LLAVA_PROJ "mm.%d.%s" +#define TN_MVLM_PROJ_MLP "mm.model.mlp.%d.%s" +#define TN_MVLM_PROJ_BLOCK "mm.model.mb_block.%d.block.%d.%s" +#define TN_MVLM_PROJ_PEG "mm.model.peg.%d.%s" +#define TN_IMAGE_NEWLINE "model.image_newline" +#define TN_MM_INP_PROJ "mm.input_projection.weight" // gemma3 +#define TN_MM_SOFT_EMB_N "mm.soft_emb_norm.weight" // gemma3 + +// mimicpmv +#define TN_MINICPMV_POS_EMBD_K "resampler.pos_embed_k" +#define TN_MINICPMV_QUERY "resampler.query" +#define TN_MINICPMV_PROJ "resampler.proj.weight" +#define TN_MINICPMV_KV_PROJ "resampler.kv.weight" +#define TN_MINICPMV_ATTN "resampler.attn.%s.%s" +#define TN_MINICPMV_LN "resampler.ln_%s.%s" + +#define TN_GLM_ADAPER_CONV "adapter.conv.%s" +#define TN_GLM_ADAPTER_LINEAR "adapter.linear.linear.%s" +#define TN_GLM_ADAPTER_NORM_1 "adapter.linear.norm1.%s" +#define TN_GLM_ADAPTER_D_H_2_4H "adapter.linear.dense_h_to_4h.%s" +#define TN_GLM_ADAPTER_GATE "adapter.linear.gate.%s" +#define TN_GLM_ADAPTER_D_4H_2_H "adapter.linear.dense_4h_to_h.%s" +#define TN_GLM_BOI_W "adapter.boi" +#define TN_GLM_EOI_W "adapter.eoi" + +enum projector_type { + PROJECTOR_TYPE_MLP, + PROJECTOR_TYPE_MLP_NORM, + PROJECTOR_TYPE_LDP, + PROJECTOR_TYPE_LDPV2, + PROJECTOR_TYPE_RESAMPLER, + PROJECTOR_TYPE_GLM_EDGE, + PROJECTOR_TYPE_MERGER, + PROJECTOR_TYPE_GEMMA3, + PROJECTOR_TYPE_UNKNOWN, +}; + +static std::map PROJECTOR_TYPE_NAMES = { + { PROJECTOR_TYPE_MLP, "mlp" }, + { PROJECTOR_TYPE_LDP, "ldp" }, + { PROJECTOR_TYPE_LDPV2, "ldpv2"}, + { PROJECTOR_TYPE_RESAMPLER, "resampler"}, + { PROJECTOR_TYPE_GLM_EDGE, "adapter"}, + { PROJECTOR_TYPE_MERGER, "qwen2vl_merger"}, + { PROJECTOR_TYPE_GEMMA3, "gemma3"}, +}; + +static projector_type clip_projector_type_from_string(const std::string & str) { + for (const auto & pair : PROJECTOR_TYPE_NAMES) { + if (pair.second == str) { + return pair.first; + } + } + return PROJECTOR_TYPE_UNKNOWN; +} + +// RGB uint8 image +struct clip_image_u8 { + int nx; + int ny; + + std::vector buf; +}; + +// RGB float32 image (NHWC) +// Memory layout: RGBRGBRGB... +struct clip_image_f32 { + int nx; + int ny; + + std::vector buf; +}; + +// +// logging +// + +static void clip_log_callback_default(enum ggml_log_level level, const char * text, void * user_data) { + (void) level; + (void) user_data; + fputs(text, stderr); + fflush(stderr); +} + +struct clip_logger_state { + ggml_log_level verbosity_thold; + ggml_log_callback log_callback; + void * log_callback_user_data; +}; + +extern struct clip_logger_state g_logger_state; + +static void clip_log_internal_v(enum ggml_log_level level, const char * format, va_list args) { + if (format == NULL) { + return; + } + va_list args_copy; + va_copy(args_copy, args); + char buffer[128]; + int len = vsnprintf(buffer, 128, format, args); + if (len < 128) { + g_logger_state.log_callback(level, buffer, g_logger_state.log_callback_user_data); + } else { + char * buffer2 = (char *) calloc(len + 1, sizeof(char)); + vsnprintf(buffer2, len + 1, format, args_copy); + buffer2[len] = 0; + g_logger_state.log_callback(level, buffer2, g_logger_state.log_callback_user_data); + free(buffer2); + } + va_end(args_copy); +} + +static void clip_log_internal(enum ggml_log_level level, const char * format, ...) { + va_list args; + va_start(args, format); + clip_log_internal_v(level, format, args); + va_end(args); +} + +#define LOG_TMPL(level, ...) \ + do { \ + if ((level) >= g_logger_state.verbosity_thold) { \ + clip_log_internal((level), __VA_ARGS__); \ + } \ + } while (0) +#define LOG_INF(...) LOG_TMPL(GGML_LOG_LEVEL_INFO, __VA_ARGS__) +#define LOG_WRN(...) LOG_TMPL(GGML_LOG_LEVEL_WARN, __VA_ARGS__) +#define LOG_ERR(...) LOG_TMPL(GGML_LOG_LEVEL_ERROR, __VA_ARGS__) +#define LOG_DBG(...) LOG_TMPL(GGML_LOG_LEVEL_DEBUG, __VA_ARGS__) +#define LOG_CNT(...) LOG_TMPL(GGML_LOG_LEVEL_CONT, __VA_ARGS__) + +// +// cpp wrappers +// + +// wrapper for clip_image_size +struct clip_image_size_deleter { + void operator()(clip_image_size * val) { clip_image_size_free(val); } +}; +typedef std::unique_ptr clip_image_size_ptr; + +// wrapper for clip_image_u8 +struct clip_image_u8_deleter { + void operator()(clip_image_u8 * val) { clip_image_u8_free(val); } +}; +typedef std::unique_ptr clip_image_u8_ptr; + +// wrapper for clip_image_f32 +struct clip_image_f32_deleter { + void operator()(clip_image_f32 * val) { clip_image_f32_free(val); } +}; +typedef std::unique_ptr clip_image_f32_ptr; + +struct clip_image_u8_batch { + std::vector entries; +}; + +struct clip_image_f32_batch { + std::vector entries; +}; + +// +// common utils +// + +static std::string string_format(const char * fmt, ...) { + va_list ap; + va_list ap2; + va_start(ap, fmt); + va_copy(ap2, ap); + int size = vsnprintf(NULL, 0, fmt, ap); + GGML_ASSERT(size >= 0 && size < INT_MAX); // NOLINT + std::vector buf(size + 1); + int size2 = vsnprintf(buf.data(), size + 1, fmt, ap2); + GGML_ASSERT(size2 == size); + va_end(ap2); + va_end(ap); + return std::string(buf.data(), buf.size()); +} + +static void string_replace_all(std::string & s, const std::string & search, const std::string & replace) { + if (search.empty()) { + return; + } + std::string builder; + builder.reserve(s.length()); + size_t pos = 0; + size_t last_pos = 0; + while ((pos = s.find(search, last_pos)) != std::string::npos) { + builder.append(s, last_pos, pos - last_pos); + builder.append(replace); + last_pos = pos + search.length(); + } + builder.append(s, last_pos, std::string::npos); + s = std::move(builder); +} + +// split string by a `std::string delim` instead of `char delim` +static std::vector string_split_str(std::string s, const std::string & delimiter) { + std::vector tokens; + size_t pos = 0; + std::string token; + while ((pos = s.find(delimiter)) != std::string::npos) { + token = s.substr(0, pos); + tokens.push_back(token); + s.erase(0, pos + delimiter.length()); + } + tokens.push_back(s); + return tokens; +} + +// +// gguf utils +// + +static std::string gguf_data_to_str(enum gguf_type type, const void * data, int i) { + switch (type) { + case GGUF_TYPE_UINT8: return std::to_string(((const uint8_t *)data)[i]); + case GGUF_TYPE_INT8: return std::to_string(((const int8_t *)data)[i]); + case GGUF_TYPE_UINT16: return std::to_string(((const uint16_t *)data)[i]); + case GGUF_TYPE_INT16: return std::to_string(((const int16_t *)data)[i]); + case GGUF_TYPE_UINT32: return std::to_string(((const uint32_t *)data)[i]); + case GGUF_TYPE_INT32: return std::to_string(((const int32_t *)data)[i]); + case GGUF_TYPE_UINT64: return std::to_string(((const uint64_t *)data)[i]); + case GGUF_TYPE_INT64: return std::to_string(((const int64_t *)data)[i]); + case GGUF_TYPE_FLOAT32: return std::to_string(((const float *)data)[i]); + case GGUF_TYPE_FLOAT64: return std::to_string(((const double *)data)[i]); + case GGUF_TYPE_BOOL: return ((const bool *)data)[i] ? "true" : "false"; + default: return string_format("unknown type %d", type); + } +} + +static std::string gguf_kv_to_str(const struct gguf_context * ctx_gguf, int i) { + const enum gguf_type type = gguf_get_kv_type(ctx_gguf, i); + + switch (type) { + case GGUF_TYPE_STRING: + return gguf_get_val_str(ctx_gguf, i); + case GGUF_TYPE_ARRAY: + { + const enum gguf_type arr_type = gguf_get_arr_type(ctx_gguf, i); + int arr_n = gguf_get_arr_n(ctx_gguf, i); + const void * data = arr_type == GGUF_TYPE_STRING ? nullptr : gguf_get_arr_data(ctx_gguf, i); + std::stringstream ss; + ss << "["; + for (int j = 0; j < arr_n; j++) { + if (arr_type == GGUF_TYPE_STRING) { + std::string val = gguf_get_arr_str(ctx_gguf, i, j); + // escape quotes + string_replace_all(val, "\\", "\\\\"); + string_replace_all(val, "\"", "\\\""); + ss << '"' << val << '"'; + } else if (arr_type == GGUF_TYPE_ARRAY) { + ss << "???"; + } else { + ss << gguf_data_to_str(arr_type, data, j); + } + if (j < arr_n - 1) { + ss << ", "; + } + } + ss << "]"; + return ss.str(); + } + default: + return gguf_data_to_str(type, gguf_get_val_data(ctx_gguf, i), 0); + } +} + +// +// API used internally with mtmd +// + +projector_type clip_get_projector_type(const struct clip_ctx * ctx); diff --git a/examples/llava/clip.cpp b/examples/llava/clip.cpp index 271cf2a2a..49c90b750 100644 --- a/examples/llava/clip.cpp +++ b/examples/llava/clip.cpp @@ -3,32 +3,14 @@ // I'll gradually clean and extend it // Note: Even when using identical normalized image inputs (see normalize_image_u8_to_f32()) we have a significant difference in resulting embeddings compared to pytorch #include "clip.h" +#include "clip-impl.h" #include "ggml.h" +#include "ggml-cpp.h" #include "ggml-cpu.h" #include "ggml-alloc.h" #include "ggml-backend.h" #include "gguf.h" -//#ifdef GGML_USE_CUDA -//#include "ggml-cuda.h" -//#endif -// -//#ifdef GGML_USE_SYCL -//#include "ggml-sycl.h" -//#endif -// -//#ifdef GGML_USE_METAL -//#include "ggml-metal.h" -//#endif -// -//#ifdef GGML_USE_CANN -//#include "ggml-cann.h" -//#endif -// -//#ifdef GGML_USE_VULKAN -//#include "ggml-vulkan.h" -//#endif - #define STB_IMAGE_IMPLEMENTATION #include "stb_image.h" @@ -40,284 +22,16 @@ #include #include #include +#include #include #include #include #include -#if defined(LLAVA_LOG_OFF) -# define LOG_INF(...) -# define LOG_WRN(...) -# define LOG_ERR(...) -# define LOG_DBG(...) -#else // defined(LLAVA_LOG_OFF) -# define LOG_INF(...) do { fprintf(stdout, __VA_ARGS__); } while (0) -# define LOG_WRN(...) do { fprintf(stderr, __VA_ARGS__); } while (0) -# define LOG_ERR(...) do { fprintf(stderr, __VA_ARGS__); } while (0) -# define LOG_DBG(...) do { fprintf(stdout, __VA_ARGS__); } while (0) -#endif // defined(LLAVA_LOG_OFF) +struct clip_logger_state g_logger_state = {GGML_LOG_LEVEL_CONT, clip_log_callback_default, NULL}; //#define CLIP_DEBUG_FUNCTIONS -// RGB uint8 image -struct clip_image_u8 { - int nx; - int ny; - - std::vector buf; -}; - -// RGB float32 image (NHWC) -// Memory layout: RGBRGBRGB... -struct clip_image_f32 { - int nx; - int ny; - - std::vector buf; -}; - -static std::string format(const char * fmt, ...) { - va_list ap; - va_list ap2; - va_start(ap, fmt); - va_copy(ap2, ap); - int size = vsnprintf(NULL, 0, fmt, ap); - GGML_ASSERT(size >= 0 && size < INT_MAX); // NOLINT - std::vector buf(size + 1); - int size2 = vsnprintf(buf.data(), size + 1, fmt, ap2); - GGML_ASSERT(size2 == size); - va_end(ap2); - va_end(ap); - return std::string(buf.data(), buf.size()); -} - -// -// key constants -// - -#define KEY_FTYPE "general.file_type" -#define KEY_NAME "general.name" -#define KEY_DESCRIPTION "general.description" -#define KEY_HAS_TEXT_ENC "clip.has_text_encoder" -#define KEY_HAS_VIS_ENC "clip.has_vision_encoder" -#define KEY_HAS_LLAVA_PROJ "clip.has_llava_projector" -#define KEY_HAS_MINICPMV_PROJ "clip.has_minicpmv_projector" -#define KEY_HAS_GLM_PROJ "clip.has_glm_projector" -#define KEY_MINICPMV_VERSION "clip.minicpmv_version" -#define KEY_HAS_QWEN2VL_MERGER "clip.has_qwen2vl_merger" -#define KEY_USE_GELU "clip.use_gelu" -#define KEY_USE_SILU "clip.use_silu" -#define KEY_N_EMBD "clip.%s.embedding_length" -#define KEY_N_FF "clip.%s.feed_forward_length" -#define KEY_N_BLOCK "clip.%s.block_count" -#define KEY_N_HEAD "clip.%s.attention.head_count" -#define KEY_LAYER_NORM_EPS "clip.%s.attention.layer_norm_epsilon" -#define KEY_PROJ_DIM "clip.%s.projection_dim" -#define KEY_TOKENS "tokenizer.ggml.tokens" -#define KEY_N_POSITIONS "clip.text.context_length" -#define KEY_IMAGE_SIZE "clip.vision.image_size" -#define KEY_PATCH_SIZE "clip.vision.patch_size" -#define KEY_IMAGE_MEAN "clip.vision.image_mean" -#define KEY_IMAGE_STD "clip.vision.image_std" -#define KEY_PROJ_TYPE "clip.projector_type" - -#define KEY_MM_PATCH_MERGE_TYPE "clip.vision.mm_patch_merge_type" -#define KEY_IMAGE_GRID_PINPOINTS "clip.vision.image_grid_pinpoints" -#define KEY_IMAGE_CROP_RESOLUTION "clip.vision.image_crop_resolution" - - -// -// tensor name constants -// - -#define TN_TOKEN_EMBD "%s.token_embd.weight" -#define TN_POS_EMBD "%s.position_embd.weight" -#define TN_CLASS_EMBD "v.class_embd" -#define TN_PATCH_EMBD "v.patch_embd.weight" // not rename tensor with ".0" postfix for backwrad compat -#define TN_PATCH_EMBD_1 "v.patch_embd.weight.1" -#define TN_PATCH_BIAS "v.patch_embd.bias" -#define TN_ATTN_K "%s.blk.%d.attn_k.%s" -#define TN_ATTN_Q "%s.blk.%d.attn_q.%s" -#define TN_ATTN_V "%s.blk.%d.attn_v.%s" -#define TN_ATTN_OUTPUT "%s.blk.%d.attn_out.%s" -#define TN_FFN_DOWN "%s.blk.%d.ffn_down.%s" -#define TN_FFN_UP "%s.blk.%d.ffn_up.%s" -#define TN_LN_1 "%s.blk.%d.ln1.%s" -#define TN_LN_2 "%s.blk.%d.ln2.%s" -#define TN_LN_PRE "%s.pre_ln.%s" -#define TN_LN_POST "%s.post_ln.%s" -#define TN_TEXT_PROJ "text_projection.weight" -#define TN_VIS_PROJ "visual_projection.weight" -#define TN_LLAVA_PROJ "mm.%d.%s" -#define TN_MVLM_PROJ_MLP "mm.model.mlp.%d.%s" -#define TN_MVLM_PROJ_BLOCK "mm.model.mb_block.%d.block.%d.%s" -#define TN_MVLM_PROJ_PEG "mm.model.peg.%d.%s" -#define TN_IMAGE_NEWLINE "model.image_newline" - -#define TN_MINICPMV_POS_EMBD_K "resampler.pos_embed_k" -#define TN_MINICPMV_QUERY "resampler.query" -#define TN_MINICPMV_PROJ "resampler.proj.weight" -#define TN_MINICPMV_KV_PROJ "resampler.kv.weight" -#define TN_MINICPMV_ATTN "resampler.attn.%s.%s" -#define TN_MINICPMV_LN "resampler.ln_%s.%s" - -#define TN_GLM_ADAPER_CONV "adapter.conv.%s" -#define TN_GLM_ADAPTER_LINEAR "adapter.linear.linear.%s" -#define TN_GLM_ADAPTER_NORM_1 "adapter.linear.norm1.%s" -#define TN_GLM_ADAPTER_D_H_2_4H "adapter.linear.dense_h_to_4h.%s" -#define TN_GLM_ADAPTER_GATE "adapter.linear.gate.%s" -#define TN_GLM_ADAPTER_D_4H_2_H "adapter.linear.dense_4h_to_h.%s" -#define TN_GLM_BOI_W "adapter.boi" -#define TN_GLM_EOI_W "adapter.eoi" - - -enum projector_type { - PROJECTOR_TYPE_MLP, - PROJECTOR_TYPE_MLP_NORM, - PROJECTOR_TYPE_LDP, - PROJECTOR_TYPE_LDPV2, - PROJECTOR_TYPE_RESAMPLER, - PROJECTOR_TYPE_GLM_EDGE, - PROJECTOR_TYPE_MERGER, - PROJECTOR_TYPE_UNKNOWN, -}; - -static std::map PROJECTOR_TYPE_NAMES = { - { PROJECTOR_TYPE_MLP, "mlp" }, - { PROJECTOR_TYPE_LDP, "ldp" }, - { PROJECTOR_TYPE_LDPV2, "ldpv2"}, - { PROJECTOR_TYPE_RESAMPLER, "resampler"}, - { PROJECTOR_TYPE_GLM_EDGE, "adapter"}, - { PROJECTOR_TYPE_MERGER, "qwen2vl_merger"}, -}; - - -// -// utilities to get data from a gguf file -// - -static int get_key_idx(const gguf_context * ctx, const char * key) { - int i = gguf_find_key(ctx, key); - if (i == -1) { - LOG_ERR("key %s not found in file\n", key); - throw std::runtime_error(format("Missing required key: %s", key)); - } - - return i; -} - -static uint32_t get_u32(const gguf_context * ctx, const std::string & key) { - const int i = get_key_idx(ctx, key.c_str()); - - return gguf_get_val_u32(ctx, i); -} - -static float get_f32(const gguf_context * ctx, const std::string & key) { - const int i = get_key_idx(ctx, key.c_str()); - - return gguf_get_val_f32(ctx, i); -} - -static struct ggml_tensor * get_tensor(struct ggml_context * ctx, const std::string & name) { - struct ggml_tensor * cur = ggml_get_tensor(ctx, name.c_str()); - if (!cur) { - throw std::runtime_error(format("%s: unable to find tensor %s\n", __func__, name.c_str())); - } - - return cur; -} - -static std::string get_ftype(int ftype) { - return ggml_type_name(static_cast(ftype)); -} - -static std::string gguf_data_to_str(enum gguf_type type, const void * data, int i) { - switch (type) { - case GGUF_TYPE_UINT8: return std::to_string(((const uint8_t *)data)[i]); - case GGUF_TYPE_INT8: return std::to_string(((const int8_t *)data)[i]); - case GGUF_TYPE_UINT16: return std::to_string(((const uint16_t *)data)[i]); - case GGUF_TYPE_INT16: return std::to_string(((const int16_t *)data)[i]); - case GGUF_TYPE_UINT32: return std::to_string(((const uint32_t *)data)[i]); - case GGUF_TYPE_INT32: return std::to_string(((const int32_t *)data)[i]); - case GGUF_TYPE_UINT64: return std::to_string(((const uint64_t *)data)[i]); - case GGUF_TYPE_INT64: return std::to_string(((const int64_t *)data)[i]); - case GGUF_TYPE_FLOAT32: return std::to_string(((const float *)data)[i]); - case GGUF_TYPE_FLOAT64: return std::to_string(((const double *)data)[i]); - case GGUF_TYPE_BOOL: return ((const bool *)data)[i] ? "true" : "false"; - default: return format("unknown type %d", type); - } -} - -static void replace_all(std::string & s, const std::string & search, const std::string & replace) { - if (search.empty()) { - return; - } - std::string builder; - builder.reserve(s.length()); - size_t pos = 0; - size_t last_pos = 0; - while ((pos = s.find(search, last_pos)) != std::string::npos) { - builder.append(s, last_pos, pos - last_pos); - builder.append(replace); - last_pos = pos + search.length(); - } - builder.append(s, last_pos, std::string::npos); - s = std::move(builder); -} - -static std::string gguf_kv_to_str(const struct gguf_context * ctx_gguf, int i) { - const enum gguf_type type = gguf_get_kv_type(ctx_gguf, i); - - switch (type) { - case GGUF_TYPE_STRING: - return gguf_get_val_str(ctx_gguf, i); - case GGUF_TYPE_ARRAY: - { - const enum gguf_type arr_type = gguf_get_arr_type(ctx_gguf, i); - int arr_n = gguf_get_arr_n(ctx_gguf, i); - const void * data = arr_type == GGUF_TYPE_STRING ? nullptr : gguf_get_arr_data(ctx_gguf, i); - std::stringstream ss; - ss << "["; - for (int j = 0; j < arr_n; j++) { - if (arr_type == GGUF_TYPE_STRING) { - std::string val = gguf_get_arr_str(ctx_gguf, i, j); - // escape quotes - replace_all(val, "\\", "\\\\"); - replace_all(val, "\"", "\\\""); - ss << '"' << val << '"'; - } else if (arr_type == GGUF_TYPE_ARRAY) { - ss << "???"; - } else { - ss << gguf_data_to_str(arr_type, data, j); - } - if (j < arr_n - 1) { - ss << ", "; - } - } - ss << "]"; - return ss.str(); - } - default: - return gguf_data_to_str(type, gguf_get_val_data(ctx_gguf, i), 0); - } -} - -static void print_tensor_info(const ggml_tensor * tensor, const char * prefix = "") { - size_t tensor_size = ggml_nbytes(tensor); - LOG_INF("%s: n_dims = %d, name = %s, tensor_size=%zu, shape:[%" PRId64 ", %" PRId64 ", %" PRId64 ", %" PRId64 "], type = %s\n", - prefix, ggml_n_dims(tensor), tensor->name, tensor_size, - tensor->ne[0], tensor->ne[1], tensor->ne[2], tensor->ne[3], ggml_type_name(tensor->type)); -} - -static projector_type clip_projector_type_from_string(const std::string & name) { - for (const auto & kv : PROJECTOR_TYPE_NAMES) { // NOLINT - if (kv.second == name) { - return kv.first; - } - } - return PROJECTOR_TYPE_UNKNOWN; -} - #ifdef CLIP_DEBUG_FUNCTIONS static void clip_image_write_image_to_ppm(const clip_image_u8& img, const std::string& filename) { std::ofstream file(filename, std::ios::binary); @@ -431,6 +145,11 @@ static void clip_image_convert_f32_to_u8(const clip_image_f32& src, clip_image_u // clip layers // +enum patch_merge_type { + PATCH_MERGE_FLAT, + PATCH_MERGE_SPATIAL_UNPAD, +}; + struct clip_hparams { int32_t image_size; int32_t patch_size; @@ -440,54 +159,55 @@ struct clip_hparams { int32_t n_head; int32_t n_layer; + patch_merge_type mm_patch_merge_type = PATCH_MERGE_FLAT; + float eps; - char mm_patch_merge_type[32] = "flat"; // spatial_unpad or flat (default) - - int32_t image_grid_pinpoints[32]; + std::vector image_grid_pinpoints; int32_t image_crop_resolution; + std::unordered_set vision_feature_layer; }; struct clip_layer { // attention - struct ggml_tensor * k_w; - struct ggml_tensor * k_b; - struct ggml_tensor * q_w; - struct ggml_tensor * q_b; - struct ggml_tensor * v_w; - struct ggml_tensor * v_b; + struct ggml_tensor * k_w = nullptr; + struct ggml_tensor * k_b = nullptr; + struct ggml_tensor * q_w = nullptr; + struct ggml_tensor * q_b = nullptr; + struct ggml_tensor * v_w = nullptr; + struct ggml_tensor * v_b = nullptr; - struct ggml_tensor * o_w; - struct ggml_tensor * o_b; + struct ggml_tensor * o_w = nullptr; + struct ggml_tensor * o_b = nullptr; // layernorm 1 - struct ggml_tensor * ln_1_w; - struct ggml_tensor * ln_1_b; + struct ggml_tensor * ln_1_w = nullptr; + struct ggml_tensor * ln_1_b = nullptr; // ff - struct ggml_tensor * ff_i_w; - struct ggml_tensor * ff_i_b; + struct ggml_tensor * ff_i_w = nullptr; + struct ggml_tensor * ff_i_b = nullptr; - struct ggml_tensor * ff_o_w; - struct ggml_tensor * ff_o_b; + struct ggml_tensor * ff_o_w = nullptr; + struct ggml_tensor * ff_o_b = nullptr; // layernorm 2 - struct ggml_tensor * ln_2_w; - struct ggml_tensor * ln_2_b; + struct ggml_tensor * ln_2_w = nullptr; + struct ggml_tensor * ln_2_b = nullptr; }; struct clip_vision_model { struct clip_hparams hparams; // embeddings - struct ggml_tensor * class_embedding; - struct ggml_tensor * patch_embeddings_0; - struct ggml_tensor * patch_embeddings_1; // second Conv2D kernel when we decouple Conv3D along temproal dimension (Qwen2VL) - struct ggml_tensor * patch_bias; - struct ggml_tensor * position_embeddings; + struct ggml_tensor * class_embedding = nullptr; + struct ggml_tensor * patch_embeddings_0 = nullptr; + struct ggml_tensor * patch_embeddings_1 = nullptr; // second Conv2D kernel when we decouple Conv3D along temproal dimension (Qwen2VL) + struct ggml_tensor * patch_bias = nullptr; + struct ggml_tensor * position_embeddings = nullptr; - struct ggml_tensor * pre_ln_w; - struct ggml_tensor * pre_ln_b; + struct ggml_tensor * pre_ln_w = nullptr; + struct ggml_tensor * pre_ln_b = nullptr; std::vector layers; @@ -497,80 +217,84 @@ struct clip_vision_model { struct ggml_tensor * projection; // LLaVA projection - struct ggml_tensor * mm_0_w = NULL; - struct ggml_tensor * mm_0_b = NULL; - struct ggml_tensor * mm_2_w = NULL; - struct ggml_tensor * mm_2_b = NULL; + struct ggml_tensor * mm_0_w = nullptr; + struct ggml_tensor * mm_0_b = nullptr; + struct ggml_tensor * mm_2_w = nullptr; + struct ggml_tensor * mm_2_b = nullptr; - struct ggml_tensor * image_newline = NULL; + struct ggml_tensor * image_newline = nullptr; // Yi type models with mlp+normalization projection - struct ggml_tensor * mm_1_w = NULL; // Yi type models have 0, 1, 3, 4 - struct ggml_tensor * mm_1_b = NULL; - struct ggml_tensor * mm_3_w = NULL; - struct ggml_tensor * mm_3_b = NULL; - struct ggml_tensor * mm_4_w = NULL; - struct ggml_tensor * mm_4_b = NULL; + struct ggml_tensor * mm_1_w = nullptr; // Yi type models have 0, 1, 3, 4 + struct ggml_tensor * mm_1_b = nullptr; + struct ggml_tensor * mm_3_w = nullptr; + struct ggml_tensor * mm_3_b = nullptr; + struct ggml_tensor * mm_4_w = nullptr; + struct ggml_tensor * mm_4_b = nullptr; //GLMV-Edge projection - struct ggml_tensor * mm_model_adapter_conv_w; - struct ggml_tensor * mm_model_adapter_conv_b; - struct ggml_tensor * boi_w; - struct ggml_tensor * eoi_w; + struct ggml_tensor * mm_model_adapter_conv_w = nullptr; + struct ggml_tensor * mm_model_adapter_conv_b = nullptr; + struct ggml_tensor * boi_w = nullptr; + struct ggml_tensor * eoi_w = nullptr; // MobileVLM projection - struct ggml_tensor * mm_model_mlp_1_w; - struct ggml_tensor * mm_model_mlp_1_b; - struct ggml_tensor * mm_model_mlp_3_w; - struct ggml_tensor * mm_model_mlp_3_b; - struct ggml_tensor * mm_model_block_1_block_0_0_w; - struct ggml_tensor * mm_model_block_1_block_0_1_w; - struct ggml_tensor * mm_model_block_1_block_0_1_b; - struct ggml_tensor * mm_model_block_1_block_1_fc1_w; - struct ggml_tensor * mm_model_block_1_block_1_fc1_b; - struct ggml_tensor * mm_model_block_1_block_1_fc2_w; - struct ggml_tensor * mm_model_block_1_block_1_fc2_b; - struct ggml_tensor * mm_model_block_1_block_2_0_w; - struct ggml_tensor * mm_model_block_1_block_2_1_w; - struct ggml_tensor * mm_model_block_1_block_2_1_b; - struct ggml_tensor * mm_model_block_2_block_0_0_w; - struct ggml_tensor * mm_model_block_2_block_0_1_w; - struct ggml_tensor * mm_model_block_2_block_0_1_b; - struct ggml_tensor * mm_model_block_2_block_1_fc1_w; - struct ggml_tensor * mm_model_block_2_block_1_fc1_b; - struct ggml_tensor * mm_model_block_2_block_1_fc2_w; - struct ggml_tensor * mm_model_block_2_block_1_fc2_b; - struct ggml_tensor * mm_model_block_2_block_2_0_w; - struct ggml_tensor * mm_model_block_2_block_2_1_w; - struct ggml_tensor * mm_model_block_2_block_2_1_b; + struct ggml_tensor * mm_model_mlp_1_w = nullptr; + struct ggml_tensor * mm_model_mlp_1_b = nullptr; + struct ggml_tensor * mm_model_mlp_3_w = nullptr; + struct ggml_tensor * mm_model_mlp_3_b = nullptr; + struct ggml_tensor * mm_model_block_1_block_0_0_w = nullptr; + struct ggml_tensor * mm_model_block_1_block_0_1_w = nullptr; + struct ggml_tensor * mm_model_block_1_block_0_1_b = nullptr; + struct ggml_tensor * mm_model_block_1_block_1_fc1_w = nullptr; + struct ggml_tensor * mm_model_block_1_block_1_fc1_b = nullptr; + struct ggml_tensor * mm_model_block_1_block_1_fc2_w = nullptr; + struct ggml_tensor * mm_model_block_1_block_1_fc2_b = nullptr; + struct ggml_tensor * mm_model_block_1_block_2_0_w = nullptr; + struct ggml_tensor * mm_model_block_1_block_2_1_w = nullptr; + struct ggml_tensor * mm_model_block_1_block_2_1_b = nullptr; + struct ggml_tensor * mm_model_block_2_block_0_0_w = nullptr; + struct ggml_tensor * mm_model_block_2_block_0_1_w = nullptr; + struct ggml_tensor * mm_model_block_2_block_0_1_b = nullptr; + struct ggml_tensor * mm_model_block_2_block_1_fc1_w = nullptr; + struct ggml_tensor * mm_model_block_2_block_1_fc1_b = nullptr; + struct ggml_tensor * mm_model_block_2_block_1_fc2_w = nullptr; + struct ggml_tensor * mm_model_block_2_block_1_fc2_b = nullptr; + struct ggml_tensor * mm_model_block_2_block_2_0_w = nullptr; + struct ggml_tensor * mm_model_block_2_block_2_1_w = nullptr; + struct ggml_tensor * mm_model_block_2_block_2_1_b = nullptr; // MobileVLM_V2 projection - struct ggml_tensor * mm_model_mlp_0_w; - struct ggml_tensor * mm_model_mlp_0_b; - struct ggml_tensor * mm_model_mlp_2_w; - struct ggml_tensor * mm_model_mlp_2_b; - struct ggml_tensor * mm_model_peg_0_w; - struct ggml_tensor * mm_model_peg_0_b; + struct ggml_tensor * mm_model_mlp_0_w = nullptr; + struct ggml_tensor * mm_model_mlp_0_b = nullptr; + struct ggml_tensor * mm_model_mlp_2_w = nullptr; + struct ggml_tensor * mm_model_mlp_2_b = nullptr; + struct ggml_tensor * mm_model_peg_0_w = nullptr; + struct ggml_tensor * mm_model_peg_0_b = nullptr; // MINICPMV projection - struct ggml_tensor * mm_model_pos_embed_k; - struct ggml_tensor * mm_model_query; - struct ggml_tensor * mm_model_proj; - struct ggml_tensor * mm_model_kv_proj; - struct ggml_tensor * mm_model_attn_q_w; - struct ggml_tensor * mm_model_attn_q_b; - struct ggml_tensor * mm_model_attn_k_w; - struct ggml_tensor * mm_model_attn_k_b; - struct ggml_tensor * mm_model_attn_v_w; - struct ggml_tensor * mm_model_attn_v_b; - struct ggml_tensor * mm_model_attn_o_w; - struct ggml_tensor * mm_model_attn_o_b; - struct ggml_tensor * mm_model_ln_q_w; - struct ggml_tensor * mm_model_ln_q_b; - struct ggml_tensor * mm_model_ln_kv_w; - struct ggml_tensor * mm_model_ln_kv_b; - struct ggml_tensor * mm_model_ln_post_w; - struct ggml_tensor * mm_model_ln_post_b; + struct ggml_tensor * mm_model_pos_embed_k = nullptr; + struct ggml_tensor * mm_model_query = nullptr; + struct ggml_tensor * mm_model_proj = nullptr; + struct ggml_tensor * mm_model_kv_proj = nullptr; + struct ggml_tensor * mm_model_attn_q_w = nullptr; + struct ggml_tensor * mm_model_attn_q_b = nullptr; + struct ggml_tensor * mm_model_attn_k_w = nullptr; + struct ggml_tensor * mm_model_attn_k_b = nullptr; + struct ggml_tensor * mm_model_attn_v_w = nullptr; + struct ggml_tensor * mm_model_attn_v_b = nullptr; + struct ggml_tensor * mm_model_attn_o_w = nullptr; + struct ggml_tensor * mm_model_attn_o_b = nullptr; + struct ggml_tensor * mm_model_ln_q_w = nullptr; + struct ggml_tensor * mm_model_ln_q_b = nullptr; + struct ggml_tensor * mm_model_ln_kv_w = nullptr; + struct ggml_tensor * mm_model_ln_kv_b = nullptr; + struct ggml_tensor * mm_model_ln_post_w = nullptr; + struct ggml_tensor * mm_model_ln_post_b = nullptr; + + // gemma3 + struct ggml_tensor * mm_input_proj_w = nullptr; + struct ggml_tensor * mm_soft_emb_norm_w = nullptr; }; struct clip_ctx { @@ -585,32 +309,211 @@ struct clip_ctx { struct clip_vision_model vision_model; projector_type proj_type = PROJECTOR_TYPE_MLP; + int32_t max_feature_layer; // unused in newer models like gemma3 float image_mean[3]; float image_std[3]; bool use_gelu = false; bool use_silu = false; - int32_t ftype = 1; - bool has_class_embedding = true; - bool has_pre_norm = true; - bool has_post_norm = false; - bool has_patch_bias = false; - - struct gguf_context * ctx_gguf; - struct ggml_context * ctx_data; + gguf_context_ptr ctx_gguf; + ggml_context_ptr ctx_data; std::vector buf_compute_meta; - // memory buffers to evaluate the model - ggml_backend_buffer_t params_buffer = NULL; + std::vector backend_ptrs; + std::vector backend_buft; - ggml_backend_t backend = NULL; - ggml_gallocr_t compute_alloc = NULL; + ggml_backend_t backend; + ggml_backend_t backend_cpu; + ggml_backend_buffer_ptr buf; - struct clip_image_size * load_image_size; + ggml_backend_sched_ptr sched; + + clip_image_size load_image_size; + + clip_ctx(clip_context_params & ctx_params) { + backend_cpu = ggml_backend_init_by_type(GGML_BACKEND_DEVICE_TYPE_CPU, nullptr); + backend = ctx_params.use_gpu + ? ggml_backend_init_by_type(GGML_BACKEND_DEVICE_TYPE_GPU, nullptr) + : nullptr; + + if (backend) { + LOG_INF("%s: CLIP using %s backend\n", __func__, ggml_backend_name(backend)); + backend_ptrs.push_back(backend); + backend_buft.push_back(ggml_backend_get_default_buffer_type(backend)); + } else { + backend = backend_cpu; + LOG_INF("%s: CLIP using CPU backend\n", __func__); + } + + backend_ptrs.push_back(backend_cpu); + backend_buft.push_back(ggml_backend_get_default_buffer_type(backend_cpu)); + + sched.reset( + ggml_backend_sched_new(backend_ptrs.data(), backend_buft.data(), backend_ptrs.size(), 8192, false) + ); + } + + ~clip_ctx() { + ggml_backend_free(backend); + if (backend != backend_cpu) { + ggml_backend_free(backend_cpu); + } + } }; -static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32_batch * imgs, struct clip_image_size * load_image_size, bool is_inf = false) { +static ggml_cgraph * clip_image_build_graph_siglip(clip_ctx * ctx, const clip_image_f32_batch & imgs) { + const auto & model = ctx->vision_model; + const auto & hparams = model.hparams; + + const int image_size = hparams.image_size; + int image_size_width = image_size; + int image_size_height = image_size; + + const int patch_size = hparams.patch_size; + const int num_patches = ((image_size_width / patch_size) * (image_size_height / patch_size)); + const int hidden_size = hparams.hidden_size; + const int n_head = hparams.n_head; + const int d_head = hidden_size / n_head; + const int n_layer = hparams.n_layer; + const float eps = hparams.eps; + + GGML_ASSERT(imgs.entries.size() == 1); // batch_size == 1 + + struct ggml_init_params params = { + /*.mem_size =*/ ctx->buf_compute_meta.size(), + /*.mem_buffer =*/ ctx->buf_compute_meta.data(), + /*.no_alloc =*/ true, + }; + + ggml_context_ptr ctx0_ptr(ggml_init(params)); + auto ctx0 = ctx0_ptr.get(); + + struct ggml_cgraph * gf = ggml_new_graph(ctx0); + + // input raw + struct ggml_tensor * inp_raw = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, image_size_width, image_size_height, 3); + ggml_set_name(inp_raw, "inp_raw"); + ggml_set_input(inp_raw); + + struct ggml_tensor * inp = ggml_conv_2d(ctx0, model.patch_embeddings_0, inp_raw, patch_size, patch_size, 0, 0, 1, 1); + inp = ggml_reshape_2d(ctx0, inp, num_patches, hidden_size); + inp = ggml_cont(ctx0, ggml_transpose(ctx0, inp)); + inp = ggml_add(ctx0, inp, model.patch_bias); + + // position embeddings + struct ggml_tensor * embeddings = ggml_add(ctx0, inp, model.position_embeddings); + + // loop over layers + for (int il = 0; il < n_layer; il++) { + struct ggml_tensor * cur = embeddings; // embeddings = residual, cur = hidden_states + + // layernorm1 + { + cur = ggml_norm(ctx0, cur, eps); + cur = ggml_add(ctx0, ggml_mul(ctx0, cur, model.layers[il].ln_1_w), model.layers[il].ln_1_b); + } + + // self-attention + { + + struct ggml_tensor * Q = + ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].q_w, cur), model.layers[il].q_b); + + Q = ggml_reshape_3d(ctx0, Q, d_head, n_head, num_patches); + Q = ggml_cont(ctx0, ggml_permute(ctx0, Q, 0, 2, 1, 3)); + + struct ggml_tensor * K = + ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].k_w, cur), model.layers[il].k_b); + + K = ggml_reshape_3d(ctx0, K, d_head, n_head, num_patches); + K = ggml_cont(ctx0, ggml_permute(ctx0, K, 0, 2, 1, 3)); + + struct ggml_tensor * V = + ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].v_w, cur), model.layers[il].v_b); + + V = ggml_reshape_3d(ctx0, V, d_head, n_head, num_patches); + V = ggml_cont(ctx0, ggml_permute(ctx0, V, 1, 2, 0, 3)); + + struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q); + KQ = ggml_soft_max_ext(ctx0, KQ, nullptr, 1.0f / sqrtf((float)d_head), 0.0f); + + struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V, KQ); + KQV = ggml_reshape_3d(ctx0, KQV, d_head, num_patches, n_head); + KQV = ggml_permute(ctx0, KQV, 0, 2, 1, 3); + + cur = ggml_cont_2d(ctx0, KQV, hidden_size, num_patches); + } + + // attention output + cur = ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].o_w, cur), model.layers[il].o_b); + + // re-add the layer input, e.g., residual + cur = ggml_add(ctx0, cur, embeddings); + + embeddings = cur; // embeddings = residual, cur = hidden_states + + // layernorm2 + { + cur = ggml_norm(ctx0, cur, eps); + cur = ggml_add(ctx0, ggml_mul(ctx0, cur, model.layers[il].ln_2_w), model.layers[il].ln_2_b); + } + + cur = ggml_mul_mat(ctx0, model.layers[il].ff_i_w, cur); + cur = ggml_add(ctx0, cur, model.layers[il].ff_i_b); + + // siglip uses gelu + cur = ggml_gelu(ctx0, cur); + + cur = ggml_mul_mat(ctx0, model.layers[il].ff_o_w, cur); + cur = ggml_add(ctx0, cur, model.layers[il].ff_o_b); + + // residual 2 + cur = ggml_add(ctx0, embeddings, cur); + + embeddings = cur; + } + + // post-layernorm + if (model.post_ln_w) { + embeddings = ggml_norm(ctx0, embeddings, eps); + ggml_set_name(embeddings, "post_ln"); + + embeddings = ggml_add(ctx0, ggml_mul(ctx0, embeddings, model.post_ln_w), model.post_ln_b); + } + + if (ctx->proj_type == PROJECTOR_TYPE_GEMMA3) { + const int batch_size = 1; + const int mm_tokens_per_image = 256; // default value for gemma3 + const int tokens_per_side = sqrt(mm_tokens_per_image); + const int patches_per_image = sqrt(num_patches); + const int kernel_size = patches_per_image / tokens_per_side; + + embeddings = ggml_cont(ctx0, ggml_transpose(ctx0, embeddings)); + embeddings = ggml_reshape_4d(ctx0, embeddings, patches_per_image, patches_per_image, hidden_size, batch_size); + + // doing a pool2d to reduce the number of output tokens to 256 + embeddings = ggml_pool_2d(ctx0, embeddings, GGML_OP_POOL_AVG, kernel_size, kernel_size, kernel_size, kernel_size, 0, 0); + embeddings = ggml_reshape_3d(ctx0, embeddings, embeddings->ne[0] * embeddings->ne[0], hidden_size, batch_size); + embeddings = ggml_cont(ctx0, ggml_transpose(ctx0, embeddings)); + + // apply norm before projection + embeddings = ggml_rms_norm(ctx0, embeddings, eps); + embeddings = ggml_mul(ctx0, embeddings, model.mm_soft_emb_norm_w); + + // apply projection + embeddings = ggml_mul_mat(ctx0, + ggml_cont(ctx0, ggml_transpose(ctx0, model.mm_input_proj_w)), + embeddings); + } + + // build the graph + ggml_build_forward_expand(gf, embeddings); + + return gf; +} + +static ggml_cgraph * clip_image_build_graph_legacy(clip_ctx * ctx, const clip_image_f32_batch & imgs, struct clip_image_size load_image_size, bool is_inf = false) { if (!ctx->has_vision_encoder) { LOG_ERR("This gguf file seems to have no vision encoder\n"); return nullptr; @@ -623,39 +526,35 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32 int image_size_width = image_size; int image_size_height = image_size; if (ctx->has_minicpmv_projector) { - if (load_image_size == nullptr) { - load_image_size = clip_image_size_init(); - } - LOG_DBG("%s: %d %d\n", __func__, load_image_size->width, load_image_size->height); - image_size_width = load_image_size->width; - image_size_height = load_image_size->height; + LOG_DBG("%s: %d %d\n", __func__, load_image_size.width, load_image_size.height); + image_size_width = load_image_size.width; + image_size_height = load_image_size.height; if (is_inf) { - image_size_width = imgs->data->nx; - image_size_height = imgs->data->ny; + image_size_width = imgs.entries[0]->nx; + image_size_height = imgs.entries[0]->ny; } } else if (ctx->has_qwen2vl_merger) { // use the image's native resolution when image is avaible if (is_inf) { // if (imgs->data->nx && imgs->data->ny) { - image_size_width = imgs->data->nx; - image_size_height = imgs->data->ny; + image_size_width = imgs.entries[0]->nx; + image_size_height = imgs.entries[0]->ny; } } const int patch_size = hparams.patch_size; const int num_patches = ((image_size_width / patch_size) * (image_size_height / patch_size)); const int patches_w = image_size_width / patch_size; const int patches_h = image_size_height / patch_size; - const int num_positions = num_patches + (ctx->has_class_embedding ? 1 : 0); + const int num_positions = num_patches + (model.class_embedding ? 1 : 0); const int num_position_ids = ctx->has_qwen2vl_merger ? num_positions * 4 : num_positions; const int hidden_size = hparams.hidden_size; const int n_head = hparams.n_head; const int d_head = hidden_size / n_head; - int n_layer = hparams.n_layer; const float eps = hparams.eps; int mrope_sections[4] = {d_head/4, d_head/4, d_head/4, d_head/4}; - const int batch_size = imgs->size; + const int batch_size = imgs.entries.size(); if (ctx->has_llava_projector || ctx->has_minicpmv_projector || ctx->has_glm_projector) { GGML_ASSERT(batch_size == 1); @@ -667,7 +566,9 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32 /*.no_alloc =*/ true, }; - struct ggml_context * ctx0 = ggml_init(params); + ggml_context_ptr ctx0_ptr(ggml_init(params)); + auto ctx0 = ctx0_ptr.get(); + struct ggml_cgraph * gf = ggml_new_graph(ctx0); struct ggml_tensor * inp_raw = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, image_size_width, image_size_height, 3, batch_size); @@ -699,7 +600,7 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32 inp = ggml_cont(ctx0, ggml_permute(ctx0, inp, 1, 0, 2, 3)); } - if (ctx->has_patch_bias) { + if (model.patch_bias) { // inp = ggml_add(ctx0, inp, ggml_repeat(ctx0, model.patch_bias, inp)); inp = ggml_add(ctx0, inp, model.patch_bias); } @@ -708,7 +609,7 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32 if (ctx->has_llava_projector) { // concat class_embeddings and patch_embeddings - if (ctx->has_class_embedding) { + if (model.class_embedding) { embeddings = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, hidden_size, num_positions, batch_size); ggml_set_name(embeddings, "embeddings"); ggml_set_input(embeddings); @@ -745,20 +646,26 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32 } // pre-layernorm - if (ctx->has_pre_norm) { + if (model.pre_ln_w) { embeddings = ggml_norm(ctx0, embeddings, eps); ggml_set_name(embeddings, "pre_ln"); embeddings = ggml_add(ctx0, ggml_mul(ctx0, embeddings, model.pre_ln_w), model.pre_ln_b); } + std::vector embedding_stack; + const auto & vision_feature_layer = hparams.vision_feature_layer; + // loop over layers - if (ctx->has_minicpmv_projector || ctx->has_glm_projector || ctx->has_qwen2vl_merger) { - n_layer += 1; - } - for (int il = 0; il < n_layer - 1; il++) { + for (int il = 0; il < ctx->max_feature_layer; il++) { struct ggml_tensor * cur = embeddings; // embeddings = residual, cur = hidden_states + // If this is an embedding feature layer, save the output. + // NOTE: 0 index here refers to the input to the encoder. + if (vision_feature_layer.find(il) != vision_feature_layer.end()) { + embedding_stack.push_back(embeddings); + } + //const size_t nb_q_w = model.layers[il].q_w->nb[0]; // layernorm1 @@ -781,7 +688,6 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32 ctx0, Q, positions, nullptr, d_head/2, mrope_sections, GGML_ROPE_TYPE_VISION, 32768, 10000, 1, 0, 1, 32, 1); } - Q = ggml_scale_inplace(ctx0, Q, 1.0f / sqrt((float)d_head)); Q = ggml_cont(ctx0, ggml_permute(ctx0, Q, 0, 2, 1, 3)); Q = ggml_reshape_3d(ctx0, Q, d_head, num_positions, n_head * batch_size); @@ -805,7 +711,7 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32 V = ggml_reshape_3d(ctx0, V, num_positions, d_head, n_head * batch_size); struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q); - KQ = ggml_soft_max_inplace(ctx0, KQ); + KQ = ggml_soft_max_ext(ctx0, KQ, nullptr, 1.0f / sqrtf((float)d_head), 0.0f); struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V, KQ); KQV = ggml_reshape_4d(ctx0, KQV, d_head, num_positions, n_head, batch_size); KQV = ggml_permute(ctx0, KQV, 0, 2, 1, 3); @@ -846,17 +752,29 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32 cur = ggml_add(ctx0, embeddings, cur); embeddings = cur; - } // post-layernorm - if (ctx->has_post_norm) { + if (model.post_ln_w) { embeddings = ggml_norm(ctx0, embeddings, eps); ggml_set_name(embeddings, "post_ln"); embeddings = ggml_add(ctx0, ggml_mul(ctx0, embeddings, model.post_ln_w), model.post_ln_b); } + // final layer is a vision feature layer + if (vision_feature_layer.find(ctx->max_feature_layer) != vision_feature_layer.end()) { + embedding_stack.push_back(embeddings); + } + + // If feature layers are explicitly set, stack them (if we have multiple) + if (!embedding_stack.empty()) { + embeddings = embedding_stack[0]; + for (size_t i = 1; i < embedding_stack.size(); i++) { + embeddings = ggml_concat(ctx0, embeddings, embedding_stack[i], 0); + } + } + // llava projector if (ctx->has_llava_projector) { embeddings = ggml_reshape_2d(ctx0, embeddings, embeddings->ne[0], embeddings->ne[1]); @@ -877,8 +795,10 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32 embeddings = ggml_add(ctx0, embeddings, model.mm_0_b); embeddings = ggml_gelu(ctx0, embeddings); - embeddings = ggml_mul_mat(ctx0, model.mm_2_w, embeddings); - embeddings = ggml_add(ctx0, embeddings, model.mm_2_b); + if (model.mm_2_w) { + embeddings = ggml_mul_mat(ctx0, model.mm_2_w, embeddings); + embeddings = ggml_add(ctx0, embeddings, model.mm_2_b); + } } else if (ctx->proj_type == PROJECTOR_TYPE_MLP_NORM) { embeddings = ggml_mul_mat(ctx0, model.mm_0_w, embeddings); @@ -1081,7 +1001,6 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32 } struct ggml_tensor * Q = ggml_add(ctx0, ggml_mul_mat(ctx0, model.mm_model_attn_q_w, q), model.mm_model_attn_q_b); - Q = ggml_scale_inplace(ctx0, Q, 1.0f / sqrt((float)d_head)); struct ggml_tensor * K = ggml_add(ctx0, ggml_mul_mat(ctx0, model.mm_model_attn_k_w, k), model.mm_model_attn_k_b); struct ggml_tensor * V = ggml_add(ctx0, ggml_mul_mat(ctx0, model.mm_model_attn_v_w, v), model.mm_model_attn_v_b); // permute @@ -1095,7 +1014,7 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32 V = ggml_cont(ctx0, ggml_permute(ctx0, V, 1, 2, 0, 3)); V = ggml_reshape_3d(ctx0, V, num_positions, d_head, n_head * batch_size); struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q); - KQ = ggml_soft_max_inplace(ctx0, KQ); + KQ = ggml_soft_max_ext(ctx0, KQ, nullptr, 1.0f / sqrtf((float)d_head), 0.0f); struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V, KQ); KQV = ggml_reshape_4d(ctx0, KQV, d_head, num_query, n_head, batch_size); KQV = ggml_permute(ctx0, KQV, 0, 2, 1, 3); @@ -1137,9 +1056,10 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32 embeddings = ggml_mul_mat(ctx0, model.mm_model_mlp_3_w, embeddings); } } else { - GGML_ABORT("fatel error"); + GGML_ABORT("fatal error"); } - } else if (ctx->proj_type == PROJECTOR_TYPE_MERGER) { + } + else if (ctx->proj_type == PROJECTOR_TYPE_MERGER) { embeddings = ggml_reshape_3d(ctx0, embeddings, hidden_size * 4, num_positions / 4, batch_size); embeddings = ggml_mul_mat(ctx0, model.mm_0_w, embeddings); @@ -1156,547 +1076,513 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32 // build the graph ggml_build_forward_expand(gf, embeddings); - ggml_free(ctx0); - return gf; } -// read and create ggml_context containing the tensors and their data -struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) { - struct ggml_context * meta = NULL; - - struct gguf_init_params params = { - /*.no_alloc = */ true, - /*.ctx = */ &meta, - }; - - struct gguf_context * ctx = gguf_init_from_file(fname, params); - if (!ctx) { - throw std::runtime_error(format("%s: failed to load CLIP model from %s. Does this file exist?\n", __func__, fname)); +static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32_batch & imgs, struct clip_image_size load_image_size, bool is_inf = false) { + if (ctx->proj_type == PROJECTOR_TYPE_GEMMA3) { + return clip_image_build_graph_siglip(ctx, imgs); + } else { + // TODO: we should have one build_* function per model + return clip_image_build_graph_legacy(ctx, imgs, load_image_size, is_inf); } +} - if (verbosity >= 1) { - const int n_tensors = gguf_get_n_tensors(ctx); - const int n_kv = gguf_get_n_kv(ctx); - const int ftype = get_u32(ctx, KEY_FTYPE); - const std::string ftype_str = get_ftype(ftype); - const int idx_desc = get_key_idx(ctx, KEY_DESCRIPTION); - const std::string description = gguf_get_val_str(ctx, idx_desc); - const int idx_name = gguf_find_key(ctx, KEY_NAME); - if (idx_name != -1) { // make name optional temporarily as some of the uploaded models missing it due to a bug - const std::string name = gguf_get_val_str(ctx, idx_name); - LOG_INF("%s: model name: %s\n", __func__, name.c_str()); - } - LOG_INF("%s: description: %s\n", __func__, description.c_str()); - LOG_INF("%s: GGUF version: %d\n", __func__, gguf_get_version(ctx)); - LOG_INF("%s: alignment: %zu\n", __func__, gguf_get_alignment(ctx)); - LOG_INF("%s: n_tensors: %d\n", __func__, n_tensors); - LOG_INF("%s: n_kv: %d\n", __func__, n_kv); - LOG_INF("%s: ftype: %s\n", __func__, ftype_str.c_str()); - LOG_INF("\n"); - } - const int n_tensors = gguf_get_n_tensors(ctx); +struct clip_model_loader { + ggml_context_ptr ctx_meta; + gguf_context_ptr ctx_gguf; - // kv - const int n_kv = gguf_get_n_kv(ctx); - LOG_INF("%s: loaded meta data with %d key-value pairs and %d tensors from %s\n", - __func__, n_kv, n_tensors, fname); - { - std::map n_type; + clip_ctx & ctx_clip; + std::string fname; - for (int i = 0; i < n_tensors; i++) { - enum ggml_type type = gguf_get_tensor_type(ctx, i); + size_t model_size; // in bytes - n_type[type]++; + // TODO @ngxson : we should not pass clip_ctx here, it should be clip_vision_model + clip_model_loader(const char * fname, clip_ctx & ctx_clip) : ctx_clip(ctx_clip), fname(fname) { + struct ggml_context * meta = nullptr; + + struct gguf_init_params params = { + /*.no_alloc = */ true, + /*.ctx = */ &meta, + }; + + ctx_gguf = gguf_context_ptr(gguf_init_from_file(fname, params)); + if (!ctx_gguf.get()) { + throw std::runtime_error(string_format("%s: failed to load CLIP model from %s. Does this file exist?\n", __func__, fname)); } - LOG_INF("%s: Dumping metadata keys/values. Note: KV overrides do not apply in this output.\n", __func__); - for (int i = 0; i < n_kv; i++) { - const char * name = gguf_get_key(ctx, i); - const enum gguf_type type = gguf_get_kv_type(ctx, i); - const std::string type_name = - type == GGUF_TYPE_ARRAY - ? format("%s[%s,%d]", gguf_type_name(type), gguf_type_name(gguf_get_arr_type(ctx, i)), gguf_get_arr_n(ctx, i)) - : gguf_type_name(type); + ctx_meta.reset(meta); - std::string value = gguf_kv_to_str(ctx, i); - const size_t MAX_VALUE_LEN = 40; - if (value.size() > MAX_VALUE_LEN) { - value = format("%s...", value.substr(0, MAX_VALUE_LEN - 3).c_str()); - } - replace_all(value, "\n", "\\n"); + const int n_tensors = gguf_get_n_tensors(ctx_gguf.get()); - LOG_INF("%s: - kv %3d: %42s %-16s = %s\n", __func__, i, name, type_name.c_str(), value.c_str()); + // print gguf info + { + std::string name; + get_string(KEY_NAME, name, false); + std::string description; + get_string(KEY_DESCRIPTION, description, false); + LOG_INF("%s: model name: %s\n", __func__, name.c_str()); + LOG_INF("%s: description: %s\n", __func__, description.c_str()); + LOG_INF("%s: GGUF version: %d\n", __func__, gguf_get_version(ctx_gguf.get())); + LOG_INF("%s: alignment: %zu\n", __func__, gguf_get_alignment(ctx_gguf.get())); + LOG_INF("%s: n_tensors: %d\n", __func__, n_tensors); + LOG_INF("%s: n_kv: %d\n", __func__, (int)gguf_get_n_kv(ctx_gguf.get())); + LOG_INF("\n"); } - // print type counts - for (auto & kv : n_type) { - if (kv.second == 0) { - continue; - } - - LOG_INF("%s: - type %4s: %4d tensors\n", __func__, ggml_type_name(kv.first), kv.second); - } - } - - // data - size_t model_size = 0; - { - for (int i = 0; i < n_tensors; ++i) { - const char * name = gguf_get_tensor_name(ctx, i); - const size_t offset = gguf_get_tensor_offset(ctx, i); - enum ggml_type type = gguf_get_tensor_type(ctx, i); - struct ggml_tensor * cur = ggml_get_tensor(meta, name); - size_t tensor_size = ggml_nbytes(cur); - model_size += tensor_size; - if (verbosity >= 3) { - LOG_INF("%s: tensor[%d]: n_dims = %d, name = %s, tensor_size=%zu, offset=%zu, shape:[%" PRIu64 ", %" PRIu64 ", %" PRIu64 ", %" PRIu64 "], type = %s\n", - __func__, i, ggml_n_dims(cur), cur->name, tensor_size, offset, cur->ne[0], cur->ne[1], cur->ne[2], cur->ne[3], ggml_type_name(type)); + // tensors + { + for (int i = 0; i < n_tensors; ++i) { + const char * name = gguf_get_tensor_name(ctx_gguf.get(), i); + const size_t offset = gguf_get_tensor_offset(ctx_gguf.get(), i); + enum ggml_type type = gguf_get_tensor_type(ctx_gguf.get(), i); + struct ggml_tensor * cur = ggml_get_tensor(meta, name); + size_t tensor_size = ggml_nbytes(cur); + model_size += tensor_size; + LOG_DBG("%s: tensor[%d]: n_dims = %d, name = %s, tensor_size=%zu, offset=%zu, shape:[%" PRIu64 ", %" PRIu64 ", %" PRIu64 ", %" PRIu64 "], type = %s\n", + __func__, i, ggml_n_dims(cur), cur->name, tensor_size, offset, cur->ne[0], cur->ne[1], cur->ne[2], cur->ne[3], ggml_type_name(type)); } } } - clip_ctx * new_clip = new clip_ctx{}; - - // update projector type - { - int idx = gguf_find_key(ctx, KEY_PROJ_TYPE); - if (idx != -1) { - const std::string proj_type = gguf_get_val_str(ctx, idx); - new_clip->proj_type = clip_projector_type_from_string(proj_type); - } else { - new_clip->proj_type = PROJECTOR_TYPE_MLP; - } - - if (new_clip->proj_type == PROJECTOR_TYPE_MLP) { - if (gguf_find_tensor(ctx, format(TN_LLAVA_PROJ, 3, "weight").c_str()) != -1) { - new_clip->proj_type = PROJECTOR_TYPE_MLP_NORM; + void load_hparams() { + // projector type + { + std::string proj_type; + get_string(KEY_PROJ_TYPE, proj_type, false); + if (!proj_type.empty()) { + ctx_clip.proj_type = clip_projector_type_from_string(proj_type); + } + if (ctx_clip.proj_type == PROJECTOR_TYPE_UNKNOWN) { + throw std::runtime_error(string_format("%s: unknown projector type: %s\n", __func__, proj_type.c_str())); } } - } -//#ifdef GGML_USE_CUDA -// new_clip->backend = ggml_backend_cuda_init(0); -// LOG_INF("%s: CLIP using CUDA backend\n", __func__); -//#endif -// -//#ifdef GGML_USE_METAL -// new_clip->backend = ggml_backend_metal_init(); -// LOG_INF("%s: CLIP using Metal backend\n", __func__); -//#endif -// -//#ifdef GGML_USE_CANN -// new_clip->backend = ggml_backend_cann_init(0); -// LOG_INF("%s: CLIP using CANN backend\n", __func__); -//#endif -// -//#ifdef GGML_USE_VULKAN -// new_clip->backend = ggml_backend_vk_init(0); -// LOG_INF("%s: CLIP using Vulkan backend\n", __func__); -//#endif -// -//#ifdef GGML_USE_SYCL -// new_clip->backend = ggml_backend_sycl_init(0); -// LOG_INF("%s: CLIP using SYCL backend\n", __func__); -//#endif + // other hparams + { + get_bool(KEY_HAS_TEXT_ENC, ctx_clip.has_text_encoder, false); + get_bool(KEY_HAS_VIS_ENC, ctx_clip.has_vision_encoder, false); + GGML_ASSERT(ctx_clip.has_vision_encoder); + GGML_ASSERT(!ctx_clip.has_text_encoder); - if (!new_clip->backend) { - new_clip->backend = ggml_backend_cpu_init(); - LOG_INF("%s: CLIP using CPU backend\n", __func__); - } + // legacy keys, use KEY_PROJ_TYPE instead + get_bool(KEY_HAS_LLAVA_PROJ, ctx_clip.has_llava_projector, false); + get_bool(KEY_HAS_MINICPMV_PROJ, ctx_clip.has_minicpmv_projector, false); + get_i32(KEY_MINICPMV_VERSION, ctx_clip.minicpmv_version, false); + get_bool(KEY_HAS_GLM_PROJ, ctx_clip.has_glm_projector, false); + get_bool(KEY_HAS_QWEN2VL_MERGER, ctx_clip.has_qwen2vl_merger, false); + // !!! do NOT extend the list above, use KEY_PROJ_TYPE instead - // model size and capabilities - { - int idx = get_key_idx(ctx, KEY_HAS_TEXT_ENC); - new_clip->has_text_encoder = gguf_get_val_bool(ctx, idx); + get_bool(KEY_USE_GELU, ctx_clip.use_gelu, false); + get_bool(KEY_USE_SILU, ctx_clip.use_silu, false); - idx = get_key_idx(ctx, KEY_HAS_VIS_ENC); - new_clip->has_vision_encoder = gguf_get_val_bool(ctx, idx); + auto & hparams = ctx_clip.vision_model.hparams; + get_u32(string_format(KEY_N_EMBD, "vision"), hparams.hidden_size); + get_u32(string_format(KEY_N_HEAD, "vision"), hparams.n_head); + get_u32(string_format(KEY_N_FF, "vision"), hparams.n_intermediate); + get_u32(string_format(KEY_N_BLOCK, "vision"), hparams.n_layer); + get_u32(string_format(KEY_PROJ_DIM, "vision"), hparams.projection_dim); + get_f32(string_format(KEY_LAYER_NORM_EPS, "vision"), hparams.eps); + get_u32(KEY_IMAGE_SIZE, hparams.image_size); + get_u32(KEY_PATCH_SIZE, hparams.patch_size); + get_u32(KEY_IMAGE_CROP_RESOLUTION, hparams.image_crop_resolution, false); + get_arr_int(KEY_IMAGE_GRID_PINPOINTS, hparams.image_grid_pinpoints, false); - idx = gguf_find_key(ctx, KEY_HAS_LLAVA_PROJ); - if (idx != -1) { - new_clip->has_llava_projector = gguf_get_val_bool(ctx, idx); - } + { + std::string mm_patch_merge_type; + get_string(KEY_MM_PATCH_MERGE_TYPE, mm_patch_merge_type, false); + if (mm_patch_merge_type == "spatial_unpad") { + hparams.mm_patch_merge_type = PATCH_MERGE_SPATIAL_UNPAD; + } + } - idx = gguf_find_key(ctx, KEY_HAS_MINICPMV_PROJ); - if (idx != -1) { - new_clip->has_minicpmv_projector = gguf_get_val_bool(ctx, idx); - } + { + int idx_mean = gguf_find_key(ctx_gguf.get(), KEY_IMAGE_MEAN); + int idx_std = gguf_find_key(ctx_gguf.get(), KEY_IMAGE_STD); + GGML_ASSERT(idx_mean >= 0 && "image_mean not found"); + GGML_ASSERT(idx_std >= 0 && "image_std not found"); + const float * mean_data = (const float *) gguf_get_arr_data(ctx_gguf.get(), idx_mean); + const float * std_data = (const float *) gguf_get_arr_data(ctx_gguf.get(), idx_std); + for (int i = 0; i < 3; ++i) { + ctx_clip.image_mean[i] = mean_data[i]; + ctx_clip.image_std[i] = std_data[i]; + } + } - idx = gguf_find_key(ctx, KEY_MINICPMV_VERSION); - if (idx != -1) { - new_clip->minicpmv_version = gguf_get_val_i32(ctx, idx); - } + // Load the vision feature layer indices if they are explicitly provided; + // if multiple vision feature layers are present, the values will be concatenated + // to form the final visual features. + // NOTE: gguf conversions should standardize the values of the vision feature layer to + // be non-negative, since we use -1 to mark values as unset here. + std::vector vision_feature_layer; + get_arr_int(KEY_FEATURE_LAYER, vision_feature_layer, false); + // convert std::vector to std::unordered_set + for (auto & layer : vision_feature_layer) { + hparams.vision_feature_layer.insert(layer); + } + // Calculate the deepest feature layer based on hparams and projector type + ctx_clip.max_feature_layer = get_deepest_feature_layer(&ctx_clip); - idx = gguf_find_key(ctx, KEY_HAS_GLM_PROJ); - if (idx != -1) { - new_clip->has_glm_projector = gguf_get_val_bool(ctx, idx); - } - - idx = gguf_find_key(ctx, KEY_HAS_QWEN2VL_MERGER); - if (idx != -1) { - new_clip->has_qwen2vl_merger = gguf_get_val_bool(ctx, idx); - } - // GGML_ASSERT(new_clip->has_llava_projector); // see monatis/clip.cpp for image and/or text encoding for semantic search - - GGML_ASSERT(new_clip->has_vision_encoder); - GGML_ASSERT(!new_clip->has_text_encoder); - - idx = get_key_idx(ctx, KEY_USE_GELU); - new_clip->use_gelu = gguf_get_val_bool(ctx, idx); - - try { - idx = get_key_idx(ctx, KEY_USE_SILU); - new_clip->use_silu = gguf_get_val_bool(ctx, idx); - } catch (std::runtime_error & /*e*/) { - new_clip->use_silu = false; - } - - if (verbosity >= 1) { - LOG_INF("%s: text_encoder: %d\n", __func__, new_clip->has_text_encoder); - LOG_INF("%s: vision_encoder: %d\n", __func__, new_clip->has_vision_encoder); - LOG_INF("%s: llava_projector: %d\n", __func__, new_clip->has_llava_projector); - LOG_INF("%s: minicpmv_projector: %d\n", __func__, new_clip->has_minicpmv_projector); - LOG_INF("%s: glm_projector: %d\n", __func__, new_clip->has_glm_projector); - LOG_INF("%s: model size: %.2f MB\n", __func__, model_size / 1024.0 / 1024.0); - LOG_INF("%s: metadata size: %.2f MB\n", __func__, ggml_get_mem_size(meta) / 1024.0 / 1024.0); + LOG_INF("%s: text_encoder: %d\n", __func__, ctx_clip.has_text_encoder); + LOG_INF("%s: vision_encoder: %d\n", __func__, ctx_clip.has_vision_encoder); + LOG_INF("%s: llava_projector: %d\n", __func__, ctx_clip.has_llava_projector); + LOG_INF("%s: minicpmv_projector: %d\n", __func__, ctx_clip.has_minicpmv_projector); + LOG_INF("%s: minicpmv_version: %d\n", __func__, ctx_clip.minicpmv_version); + LOG_INF("%s: glm_projector: %d\n", __func__, ctx_clip.has_glm_projector); + LOG_INF("%s: model size: %.2f MiB\n", __func__, model_size / 1024.0 / 1024.0); + LOG_INF("%s: metadata size: %.2f MiB\n", __func__, ggml_get_mem_size(ctx_meta.get()) / 1024.0 / 1024.0); } } - LOG_INF("%s: params backend buffer size = % 6.2f MB (%i tensors)\n", __func__, model_size / (1024.0 * 1024.0), n_tensors); + void load_tensors() { + std::map tensor_offset; + std::vector tensors_to_load; - // load tensors - { - std::vector read_buf; + // get offsets + for (int64_t i = 0; i < gguf_get_n_tensors(ctx_gguf.get()); ++i) { + const char * name = gguf_get_tensor_name(ctx_gguf.get(), i); + tensor_offset[name] = gguf_get_data_offset(ctx_gguf.get()) + gguf_get_tensor_offset(ctx_gguf.get(), i); + } + + // create data context struct ggml_init_params params = { - /*.mem_size =*/ (n_tensors + 1) * ggml_tensor_overhead(), + /*.mem_size =*/ (gguf_get_n_tensors(ctx_gguf.get()) + 1) * ggml_tensor_overhead(), /*.mem_buffer =*/ NULL, /*.no_alloc =*/ true, }; - - new_clip->ctx_data = ggml_init(params); - if (!new_clip->ctx_data) { - LOG_ERR("%s: ggml_init() failed\n", __func__); - clip_free(new_clip); - gguf_free(ctx); - return nullptr; + ctx_clip.ctx_data.reset(ggml_init(params)); + if (!ctx_clip.ctx_data) { + throw std::runtime_error(string_format("%s: failed to init ggml context\n", __func__)); } - auto fin = std::ifstream(fname, std::ios::binary); - if (!fin) { - LOG_ERR("cannot open model file for loading tensors\n"); - clip_free(new_clip); - gguf_free(ctx); - return nullptr; - } - - // add tensors to context - for (int i = 0; i < n_tensors; ++i) { - const char * name = gguf_get_tensor_name(ctx, i); - struct ggml_tensor * t = ggml_get_tensor(meta, name); - struct ggml_tensor * cur = ggml_dup_tensor(new_clip->ctx_data, t); - ggml_set_name(cur, name); - } - - // alloc memory and offload data - new_clip->params_buffer = ggml_backend_alloc_ctx_tensors(new_clip->ctx_data, new_clip->backend); - for (int i = 0; i < n_tensors; ++i) { - const char * name = gguf_get_tensor_name(ctx, i); - struct ggml_tensor * cur = ggml_get_tensor(new_clip->ctx_data, name); - const size_t offset = gguf_get_data_offset(ctx) + gguf_get_tensor_offset(ctx, i); - fin.seekg(offset, std::ios::beg); - if (!fin) { - LOG_ERR("%s: failed to seek for tensor %s\n", __func__, name); - clip_free(new_clip); - gguf_free(ctx); - return nullptr; + // helper function + auto get_tensor = [&](const std::string & name, bool required = true) { + struct ggml_tensor * cur = ggml_get_tensor(ctx_meta.get(), name.c_str()); + if (!cur && required) { + throw std::runtime_error(string_format("%s: unable to find tensor %s\n", __func__, name.c_str())); } - int num_bytes = ggml_nbytes(cur); - if (ggml_backend_buffer_is_host(new_clip->params_buffer)) { - // for the CPU and Metal backend, we can read directly into the tensor - fin.read(reinterpret_cast(cur->data), num_bytes); - } else { - // read into a temporary buffer first, then copy to device memory - read_buf.resize(num_bytes); - fin.read(reinterpret_cast(read_buf.data()), num_bytes); - ggml_backend_tensor_set(cur, read_buf.data(), 0, num_bytes); + if (cur) { + tensors_to_load.push_back(cur); + // add tensors to context + struct ggml_tensor * data_tensor = ggml_dup_tensor(ctx_clip.ctx_data.get(), cur); + ggml_set_name(data_tensor, cur->name); + cur = data_tensor; } - } - fin.close(); - } + return cur; + }; - // vision model - if (new_clip->has_vision_encoder) { - // load vision model - auto & vision_model = new_clip->vision_model; - auto & hparams = vision_model.hparams; - hparams.hidden_size = get_u32(ctx, format(KEY_N_EMBD, "vision")); - hparams.n_head = get_u32(ctx, format(KEY_N_HEAD, "vision")); - hparams.n_intermediate = get_u32(ctx, format(KEY_N_FF, "vision")); - hparams.n_layer = get_u32(ctx, format(KEY_N_BLOCK, "vision")); - hparams.image_size = get_u32(ctx, KEY_IMAGE_SIZE); - hparams.patch_size = get_u32(ctx, KEY_PATCH_SIZE); - hparams.projection_dim = get_u32(ctx, format(KEY_PROJ_DIM, "vision")); - hparams.eps = get_f32(ctx, format(KEY_LAYER_NORM_EPS, "vision")); + auto & vision_model = ctx_clip.vision_model; - try { - int idx = get_key_idx(ctx, KEY_IMAGE_GRID_PINPOINTS); - int n = gguf_get_arr_n(ctx, idx); - const int32_t * pinpoints = (const int32_t *)gguf_get_arr_data(ctx, idx); - for (int i = 0; i < 32 && i < n && pinpoints[i] != 0; ++i) { - hparams.image_grid_pinpoints[i] = pinpoints[i]; - } - if (n < 32) - hparams.image_grid_pinpoints[n] = 0; - } catch (std::runtime_error & /*e*/) { - hparams.image_grid_pinpoints[0]=0; + vision_model.class_embedding = get_tensor(TN_CLASS_EMBD, false); + + vision_model.pre_ln_w = get_tensor(string_format(TN_LN_PRE, "v", "weight"), false); + vision_model.pre_ln_b = get_tensor(string_format(TN_LN_PRE, "v", "bias"), false); + + vision_model.post_ln_w = get_tensor(string_format(TN_LN_POST, "v", "weight"), false); + vision_model.post_ln_b = get_tensor(string_format(TN_LN_POST, "v", "bias"), false); + + vision_model.patch_bias = get_tensor(TN_PATCH_BIAS, false); + vision_model.patch_embeddings_0 = get_tensor(TN_PATCH_EMBD, false); + vision_model.patch_embeddings_1 = get_tensor(TN_PATCH_EMBD_1, false); + if (vision_model.patch_embeddings_1 == nullptr) { + ctx_clip.has_qwen2vl_merger = false; } - try { - int idx = get_key_idx(ctx, KEY_MM_PATCH_MERGE_TYPE); - strcpy(hparams.mm_patch_merge_type, gguf_get_val_str(ctx, idx)); - } catch (std::runtime_error & /*e*/) { - strcpy(hparams.mm_patch_merge_type, "flat"); - } + vision_model.position_embeddings = get_tensor(string_format(TN_POS_EMBD, "v"), false); - try { - hparams.image_crop_resolution = get_u32(ctx, KEY_IMAGE_CROP_RESOLUTION); // llava-1.6 - } catch(const std::exception& /*e*/) { - hparams.image_crop_resolution = hparams.image_size; - } - - int idx_mean = get_key_idx(ctx, KEY_IMAGE_MEAN); - int idx_std = get_key_idx(ctx, KEY_IMAGE_STD); - - const float * mean_data = (const float *)gguf_get_arr_data(ctx, idx_mean); - const float * std_data = (const float *)gguf_get_arr_data(ctx, idx_std); - - for (int i = 0; i < 3; ++i) { - new_clip->image_mean[i] = mean_data[i]; - new_clip->image_std[i] = std_data[i]; - } - - if (verbosity >= 2) { - LOG_INF("\n%s: vision model hparams\n", __func__); - LOG_INF("image_size %d\n", hparams.image_size); - LOG_INF("patch_size %d\n", hparams.patch_size); - LOG_INF("v_hidden_size %d\n", hparams.hidden_size); - LOG_INF("v_n_intermediate %d\n", hparams.n_intermediate); - LOG_INF("v_projection_dim %d\n", hparams.projection_dim); - LOG_INF("v_n_head %d\n", hparams.n_head); - LOG_INF("v_n_layer %d\n", hparams.n_layer); - LOG_INF("v_eps %f\n", hparams.eps); - LOG_INF("v_image_mean %f %f %f\n", new_clip->image_mean[0], new_clip->image_mean[1], new_clip->image_mean[2]); - LOG_INF("v_image_std %f %f %f\n", new_clip->image_std[0], new_clip->image_std[1], new_clip->image_std[2]); - LOG_INF("v_image_grid_pinpoints: "); - for (int i = 0; i < 32 && (hparams.image_grid_pinpoints[i] != 0); ++i) { - LOG_INF("%d ", hparams.image_grid_pinpoints[i]); - } - LOG_INF("\n"); - LOG_INF("v_mm_patch_merge_type: %s\n", hparams.mm_patch_merge_type); - - } - - try { - vision_model.class_embedding = get_tensor(new_clip->ctx_data, TN_CLASS_EMBD); - new_clip->has_class_embedding = true; - } catch (const std::exception& /*e*/) { - new_clip->has_class_embedding = false; - } - - try { - vision_model.pre_ln_w = get_tensor(new_clip->ctx_data, format(TN_LN_PRE, "v", "weight")); - vision_model.pre_ln_b = get_tensor(new_clip->ctx_data, format(TN_LN_PRE, "v", "bias")); - new_clip->has_pre_norm = true; - } catch (std::exception & /*e*/) { - new_clip->has_pre_norm = false; - } - - try { - vision_model.post_ln_w = get_tensor(new_clip->ctx_data, format(TN_LN_POST, "v", "weight")); - vision_model.post_ln_b = get_tensor(new_clip->ctx_data, format(TN_LN_POST, "v", "bias")); - new_clip->has_post_norm = true; - } catch (std::exception & /*e*/) { - new_clip->has_post_norm = false; - } - - try { - vision_model.patch_bias = get_tensor(new_clip->ctx_data, TN_PATCH_BIAS); - new_clip->has_patch_bias = true; - } catch (std::exception & /*e*/) { - new_clip->has_patch_bias = false; - } - - try { - vision_model.patch_embeddings_0 = get_tensor(new_clip->ctx_data, TN_PATCH_EMBD); - vision_model.position_embeddings = get_tensor(new_clip->ctx_data, format(TN_POS_EMBD, "v")); - } catch(const std::exception& /*e*/) { - LOG_ERR("%s: failed to load vision model tensors\n", __func__); - } - try { - vision_model.patch_embeddings_1 = get_tensor(new_clip->ctx_data, TN_PATCH_EMBD_1); - } catch(const std::exception& /*e*/) { - new_clip->has_qwen2vl_merger = false; - } - - // LLaVA projection - if (new_clip->proj_type == PROJECTOR_TYPE_MLP || new_clip->proj_type == PROJECTOR_TYPE_MLP_NORM) { - vision_model.mm_0_w = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 0, "weight")); - vision_model.mm_0_b = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 0, "bias")); - try { - // Yi-type llava - vision_model.mm_1_w = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 1, "weight")); - vision_model.mm_1_b = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 1, "bias")); - } catch (std::runtime_error & /*e*/) { } - try { - // missing in Yi-type llava - vision_model.mm_2_w = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 2, "weight")); - vision_model.mm_2_b = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 2, "bias")); - } catch (std::runtime_error & /*e*/) { } - try { - // Yi-type llava - vision_model.mm_3_w = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 3, "weight")); - vision_model.mm_3_b = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 3, "bias")); - } catch (std::runtime_error & /*e*/) { } - try { - // Yi-type llava - vision_model.mm_4_w = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 4, "weight")); - vision_model.mm_4_b = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 4, "bias")); - } catch (std::runtime_error & /*e*/) { } - try { - vision_model.image_newline = get_tensor(new_clip->ctx_data, TN_IMAGE_NEWLINE); - // LOG_INF("%s: image_newline tensor (llava-1.6) found\n", __func__); - } catch (std::runtime_error & /*e*/) { } - } else if (new_clip->proj_type == PROJECTOR_TYPE_LDP) { - // MobileVLM projection - vision_model.mm_model_mlp_1_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_MLP, 1, "weight")); - vision_model.mm_model_mlp_1_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_MLP, 1, "bias")); - vision_model.mm_model_mlp_3_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_MLP, 3, "weight")); - vision_model.mm_model_mlp_3_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_MLP, 3, "bias")); - vision_model.mm_model_block_1_block_0_0_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 1, 0, "0.weight")); - vision_model.mm_model_block_1_block_0_1_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 1, 0, "1.weight")); - vision_model.mm_model_block_1_block_0_1_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 1, 0, "1.bias")); - vision_model.mm_model_block_1_block_1_fc1_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 1, 1, "fc1.weight")); - vision_model.mm_model_block_1_block_1_fc1_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 1, 1, "fc1.bias")); - vision_model.mm_model_block_1_block_1_fc2_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 1, 1, "fc2.weight")); - vision_model.mm_model_block_1_block_1_fc2_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 1, 1, "fc2.bias")); - vision_model.mm_model_block_1_block_2_0_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 1, 2, "0.weight")); - vision_model.mm_model_block_1_block_2_1_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 1, 2, "1.weight")); - vision_model.mm_model_block_1_block_2_1_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 1, 2, "1.bias")); - vision_model.mm_model_block_2_block_0_0_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 2, 0, "0.weight")); - vision_model.mm_model_block_2_block_0_1_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 2, 0, "1.weight")); - vision_model.mm_model_block_2_block_0_1_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 2, 0, "1.bias")); - vision_model.mm_model_block_2_block_1_fc1_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 2, 1, "fc1.weight")); - vision_model.mm_model_block_2_block_1_fc1_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 2, 1, "fc1.bias")); - vision_model.mm_model_block_2_block_1_fc2_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 2, 1, "fc2.weight")); - vision_model.mm_model_block_2_block_1_fc2_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 2, 1, "fc2.bias")); - vision_model.mm_model_block_2_block_2_0_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 2, 2, "0.weight")); - vision_model.mm_model_block_2_block_2_1_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 2, 2, "1.weight")); - vision_model.mm_model_block_2_block_2_1_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 2, 2, "1.bias")); - } - else if (new_clip->proj_type == PROJECTOR_TYPE_LDPV2) - { - // MobilVLM_V2 projection - vision_model.mm_model_mlp_0_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_MLP, 0, "weight")); - vision_model.mm_model_mlp_0_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_MLP, 0, "bias")); - vision_model.mm_model_mlp_2_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_MLP, 2, "weight")); - vision_model.mm_model_mlp_2_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_MLP, 2, "bias")); - vision_model.mm_model_peg_0_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_PEG, 0, "weight")); - vision_model.mm_model_peg_0_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_PEG, 0, "bias")); - } - else if (new_clip->proj_type == PROJECTOR_TYPE_RESAMPLER) { - // vision_model.mm_model_pos_embed = get_tensor(new_clip->ctx_data, TN_MINICPMV_POS_EMBD); - vision_model.mm_model_pos_embed_k = get_tensor(new_clip->ctx_data, TN_MINICPMV_POS_EMBD_K); - vision_model.mm_model_query = get_tensor(new_clip->ctx_data, TN_MINICPMV_QUERY); - vision_model.mm_model_proj = get_tensor(new_clip->ctx_data, TN_MINICPMV_PROJ); - vision_model.mm_model_kv_proj = get_tensor(new_clip->ctx_data, TN_MINICPMV_KV_PROJ); - vision_model.mm_model_attn_q_w = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_ATTN, "q", "weight")); - vision_model.mm_model_attn_k_w = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_ATTN, "k", "weight")); - vision_model.mm_model_attn_v_w = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_ATTN, "v", "weight")); - vision_model.mm_model_attn_q_b = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_ATTN, "q", "bias")); - vision_model.mm_model_attn_k_b = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_ATTN, "k", "bias")); - vision_model.mm_model_attn_v_b = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_ATTN, "v", "bias")); - vision_model.mm_model_attn_o_w = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_ATTN, "out", "weight")); - vision_model.mm_model_attn_o_b = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_ATTN, "out", "bias")); - vision_model.mm_model_ln_q_w = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_LN, "q", "weight")); - vision_model.mm_model_ln_q_b = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_LN, "q", "bias")); - vision_model.mm_model_ln_kv_w = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_LN, "kv", "weight")); - vision_model.mm_model_ln_kv_b = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_LN, "kv", "bias")); - vision_model.mm_model_ln_post_w = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_LN, "post", "weight")); - vision_model.mm_model_ln_post_b = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_LN, "post", "bias")); - } - else if (new_clip->proj_type == PROJECTOR_TYPE_GLM_EDGE) { - vision_model.mm_model_adapter_conv_w = get_tensor(new_clip->ctx_data, format(TN_GLM_ADAPER_CONV, "weight")); - vision_model.mm_model_adapter_conv_b = get_tensor(new_clip->ctx_data, format(TN_GLM_ADAPER_CONV, "bias")); - vision_model.mm_model_mlp_0_w = get_tensor(new_clip->ctx_data, format(TN_GLM_ADAPTER_LINEAR,"weight")); - vision_model.mm_model_ln_q_w = get_tensor(new_clip->ctx_data, format(TN_GLM_ADAPTER_NORM_1,"weight")); - vision_model.mm_model_ln_q_b = get_tensor(new_clip->ctx_data, format(TN_GLM_ADAPTER_NORM_1,"bias")); - vision_model.mm_model_mlp_1_w = get_tensor(new_clip->ctx_data, format(TN_GLM_ADAPTER_D_H_2_4H,"weight")); - vision_model.mm_model_mlp_2_w = get_tensor(new_clip->ctx_data, format(TN_GLM_ADAPTER_GATE,"weight")); - vision_model.mm_model_mlp_3_w = get_tensor(new_clip->ctx_data, format(TN_GLM_ADAPTER_D_4H_2_H,"weight")); - vision_model.boi_w = get_tensor(new_clip->ctx_data, TN_GLM_BOI_W); - vision_model.eoi_w = get_tensor(new_clip->ctx_data, TN_GLM_EOI_W); - } - else if (new_clip->proj_type == PROJECTOR_TYPE_MERGER) { - vision_model.mm_0_w = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 0, "weight")); - vision_model.mm_0_b = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 0, "bias")); - vision_model.mm_1_w = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 2, "weight")); - vision_model.mm_1_b = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 2, "bias")); - } - else { - std::string proj_type = PROJECTOR_TYPE_NAMES[new_clip->proj_type]; - throw std::runtime_error(format("%s: don't support projector with: %s currently\n", __func__, proj_type.c_str())); - } - - vision_model.layers.resize(hparams.n_layer); - - for (int il = 0; il < hparams.n_layer; ++il) { + // layers + vision_model.layers.resize(vision_model.hparams.n_layer); + for (int il = 0; il < vision_model.hparams.n_layer; ++il) { auto & layer = vision_model.layers[il]; - layer.k_w = get_tensor(new_clip->ctx_data, format(TN_ATTN_K, "v", il, "weight")); - layer.q_w = get_tensor(new_clip->ctx_data, format(TN_ATTN_Q, "v", il, "weight")); - layer.v_w = get_tensor(new_clip->ctx_data, format(TN_ATTN_V, "v", il, "weight")); - layer.o_w = get_tensor(new_clip->ctx_data, format(TN_ATTN_OUTPUT, "v", il, "weight")); - layer.ln_1_w = get_tensor(new_clip->ctx_data, format(TN_LN_1, "v", il, "weight")); - layer.ln_2_w = get_tensor(new_clip->ctx_data, format(TN_LN_2, "v", il, "weight")); - layer.ff_i_w = get_tensor(new_clip->ctx_data, format(TN_FFN_DOWN, "v", il, "weight")); - layer.ff_o_w = get_tensor(new_clip->ctx_data, format(TN_FFN_UP, "v", il, "weight")); - layer.k_b = get_tensor(new_clip->ctx_data, format(TN_ATTN_K, "v", il, "bias")); - layer.q_b = get_tensor(new_clip->ctx_data, format(TN_ATTN_Q, "v", il, "bias")); - layer.v_b = get_tensor(new_clip->ctx_data, format(TN_ATTN_V, "v", il, "bias")); - layer.o_b = get_tensor(new_clip->ctx_data, format(TN_ATTN_OUTPUT, "v", il, "bias")); - layer.ln_1_b = get_tensor(new_clip->ctx_data, format(TN_LN_1, "v", il, "bias")); - layer.ln_2_b = get_tensor(new_clip->ctx_data, format(TN_LN_2, "v", il, "bias")); - layer.ff_i_b = get_tensor(new_clip->ctx_data, format(TN_FFN_DOWN, "v", il, "bias")); - layer.ff_o_b = get_tensor(new_clip->ctx_data, format(TN_FFN_UP, "v", il, "bias")); + layer.k_w = get_tensor(string_format(TN_ATTN_K, "v", il, "weight")); + layer.q_w = get_tensor(string_format(TN_ATTN_Q, "v", il, "weight")); + layer.v_w = get_tensor(string_format(TN_ATTN_V, "v", il, "weight")); + layer.o_w = get_tensor(string_format(TN_ATTN_OUTPUT, "v", il, "weight")); + layer.ln_1_w = get_tensor(string_format(TN_LN_1, "v", il, "weight"), false); + layer.ln_2_w = get_tensor(string_format(TN_LN_2, "v", il, "weight"), false); + layer.ff_i_w = get_tensor(string_format(TN_FFN_DOWN, "v", il, "weight")); + layer.ff_o_w = get_tensor(string_format(TN_FFN_UP, "v", il, "weight")); + layer.k_b = get_tensor(string_format(TN_ATTN_K, "v", il, "bias"), false); + layer.q_b = get_tensor(string_format(TN_ATTN_Q, "v", il, "bias"), false); + layer.v_b = get_tensor(string_format(TN_ATTN_V, "v", il, "bias"), false); + layer.o_b = get_tensor(string_format(TN_ATTN_OUTPUT, "v", il, "bias"), false); + layer.ln_1_b = get_tensor(string_format(TN_LN_1, "v", il, "bias"), false); + layer.ln_2_b = get_tensor(string_format(TN_LN_2, "v", il, "bias"), false); + layer.ff_i_b = get_tensor(string_format(TN_FFN_DOWN, "v", il, "bias"), false); + layer.ff_o_b = get_tensor(string_format(TN_FFN_UP, "v", il, "bias"), false); + } + + switch (ctx_clip.proj_type) { + case PROJECTOR_TYPE_MLP: + case PROJECTOR_TYPE_MLP_NORM: + { + // LLaVA projection + vision_model.mm_0_w = get_tensor(string_format(TN_LLAVA_PROJ, 0, "weight"), false); + vision_model.mm_0_b = get_tensor(string_format(TN_LLAVA_PROJ, 0, "bias"), false); + // Yi-type llava + vision_model.mm_1_w = get_tensor(string_format(TN_LLAVA_PROJ, 1, "weight"), false); + vision_model.mm_1_b = get_tensor(string_format(TN_LLAVA_PROJ, 1, "bias"), false); + // missing in Yi-type llava + vision_model.mm_2_w = get_tensor(string_format(TN_LLAVA_PROJ, 2, "weight"), false); + vision_model.mm_2_b = get_tensor(string_format(TN_LLAVA_PROJ, 2, "bias"), false); + // Yi-type llava + vision_model.mm_3_w = get_tensor(string_format(TN_LLAVA_PROJ, 3, "weight"), false); + vision_model.mm_3_b = get_tensor(string_format(TN_LLAVA_PROJ, 3, "bias"), false); + vision_model.mm_4_w = get_tensor(string_format(TN_LLAVA_PROJ, 4, "weight"), false); + vision_model.mm_4_b = get_tensor(string_format(TN_LLAVA_PROJ, 4, "bias"), false); + if (vision_model.mm_3_w) { + // TODO: this is a hack to support Yi-type llava + ctx_clip.proj_type = PROJECTOR_TYPE_MLP_NORM; + } + vision_model.image_newline = get_tensor(TN_IMAGE_NEWLINE, false); + } break; + case PROJECTOR_TYPE_LDP: + { + // MobileVLM projection + vision_model.mm_model_mlp_1_w = get_tensor(string_format(TN_MVLM_PROJ_MLP, 1, "weight")); + vision_model.mm_model_mlp_1_b = get_tensor(string_format(TN_MVLM_PROJ_MLP, 1, "bias")); + vision_model.mm_model_mlp_3_w = get_tensor(string_format(TN_MVLM_PROJ_MLP, 3, "weight")); + vision_model.mm_model_mlp_3_b = get_tensor(string_format(TN_MVLM_PROJ_MLP, 3, "bias")); + vision_model.mm_model_block_1_block_0_0_w = get_tensor(string_format(TN_MVLM_PROJ_BLOCK, 1, 0, "0.weight")); + vision_model.mm_model_block_1_block_0_1_w = get_tensor(string_format(TN_MVLM_PROJ_BLOCK, 1, 0, "1.weight")); + vision_model.mm_model_block_1_block_0_1_b = get_tensor(string_format(TN_MVLM_PROJ_BLOCK, 1, 0, "1.bias")); + vision_model.mm_model_block_1_block_1_fc1_w = get_tensor(string_format(TN_MVLM_PROJ_BLOCK, 1, 1, "fc1.weight")); + vision_model.mm_model_block_1_block_1_fc1_b = get_tensor(string_format(TN_MVLM_PROJ_BLOCK, 1, 1, "fc1.bias")); + vision_model.mm_model_block_1_block_1_fc2_w = get_tensor(string_format(TN_MVLM_PROJ_BLOCK, 1, 1, "fc2.weight")); + vision_model.mm_model_block_1_block_1_fc2_b = get_tensor(string_format(TN_MVLM_PROJ_BLOCK, 1, 1, "fc2.bias")); + vision_model.mm_model_block_1_block_2_0_w = get_tensor(string_format(TN_MVLM_PROJ_BLOCK, 1, 2, "0.weight")); + vision_model.mm_model_block_1_block_2_1_w = get_tensor(string_format(TN_MVLM_PROJ_BLOCK, 1, 2, "1.weight")); + vision_model.mm_model_block_1_block_2_1_b = get_tensor(string_format(TN_MVLM_PROJ_BLOCK, 1, 2, "1.bias")); + vision_model.mm_model_block_2_block_0_0_w = get_tensor(string_format(TN_MVLM_PROJ_BLOCK, 2, 0, "0.weight")); + vision_model.mm_model_block_2_block_0_1_w = get_tensor(string_format(TN_MVLM_PROJ_BLOCK, 2, 0, "1.weight")); + vision_model.mm_model_block_2_block_0_1_b = get_tensor(string_format(TN_MVLM_PROJ_BLOCK, 2, 0, "1.bias")); + vision_model.mm_model_block_2_block_1_fc1_w = get_tensor(string_format(TN_MVLM_PROJ_BLOCK, 2, 1, "fc1.weight")); + vision_model.mm_model_block_2_block_1_fc1_b = get_tensor(string_format(TN_MVLM_PROJ_BLOCK, 2, 1, "fc1.bias")); + vision_model.mm_model_block_2_block_1_fc2_w = get_tensor(string_format(TN_MVLM_PROJ_BLOCK, 2, 1, "fc2.weight")); + vision_model.mm_model_block_2_block_1_fc2_b = get_tensor(string_format(TN_MVLM_PROJ_BLOCK, 2, 1, "fc2.bias")); + vision_model.mm_model_block_2_block_2_0_w = get_tensor(string_format(TN_MVLM_PROJ_BLOCK, 2, 2, "0.weight")); + vision_model.mm_model_block_2_block_2_1_w = get_tensor(string_format(TN_MVLM_PROJ_BLOCK, 2, 2, "1.weight")); + vision_model.mm_model_block_2_block_2_1_b = get_tensor(string_format(TN_MVLM_PROJ_BLOCK, 2, 2, "1.bias")); + } break; + case PROJECTOR_TYPE_LDPV2: + { + // MobilVLM_V2 projection + vision_model.mm_model_mlp_0_w = get_tensor(string_format(TN_MVLM_PROJ_MLP, 0, "weight")); + vision_model.mm_model_mlp_0_b = get_tensor(string_format(TN_MVLM_PROJ_MLP, 0, "bias")); + vision_model.mm_model_mlp_2_w = get_tensor(string_format(TN_MVLM_PROJ_MLP, 2, "weight")); + vision_model.mm_model_mlp_2_b = get_tensor(string_format(TN_MVLM_PROJ_MLP, 2, "bias")); + vision_model.mm_model_peg_0_w = get_tensor(string_format(TN_MVLM_PROJ_PEG, 0, "weight")); + vision_model.mm_model_peg_0_b = get_tensor(string_format(TN_MVLM_PROJ_PEG, 0, "bias")); + } break; + case PROJECTOR_TYPE_RESAMPLER: + { + // vision_model.mm_model_pos_embed = get_tensor(new_clip->ctx_data, TN_MINICPMV_POS_EMBD); + vision_model.mm_model_pos_embed_k = get_tensor(TN_MINICPMV_POS_EMBD_K); + vision_model.mm_model_query = get_tensor(TN_MINICPMV_QUERY); + vision_model.mm_model_proj = get_tensor(TN_MINICPMV_PROJ); + vision_model.mm_model_kv_proj = get_tensor(TN_MINICPMV_KV_PROJ); + vision_model.mm_model_attn_q_w = get_tensor(string_format(TN_MINICPMV_ATTN, "q", "weight")); + vision_model.mm_model_attn_k_w = get_tensor(string_format(TN_MINICPMV_ATTN, "k", "weight")); + vision_model.mm_model_attn_v_w = get_tensor(string_format(TN_MINICPMV_ATTN, "v", "weight")); + vision_model.mm_model_attn_q_b = get_tensor(string_format(TN_MINICPMV_ATTN, "q", "bias")); + vision_model.mm_model_attn_k_b = get_tensor(string_format(TN_MINICPMV_ATTN, "k", "bias")); + vision_model.mm_model_attn_v_b = get_tensor(string_format(TN_MINICPMV_ATTN, "v", "bias")); + vision_model.mm_model_attn_o_w = get_tensor(string_format(TN_MINICPMV_ATTN, "out", "weight")); + vision_model.mm_model_attn_o_b = get_tensor(string_format(TN_MINICPMV_ATTN, "out", "bias")); + vision_model.mm_model_ln_q_w = get_tensor(string_format(TN_MINICPMV_LN, "q", "weight")); + vision_model.mm_model_ln_q_b = get_tensor(string_format(TN_MINICPMV_LN, "q", "bias")); + vision_model.mm_model_ln_kv_w = get_tensor(string_format(TN_MINICPMV_LN, "kv", "weight")); + vision_model.mm_model_ln_kv_b = get_tensor(string_format(TN_MINICPMV_LN, "kv", "bias")); + vision_model.mm_model_ln_post_w = get_tensor(string_format(TN_MINICPMV_LN, "post", "weight")); + vision_model.mm_model_ln_post_b = get_tensor(string_format(TN_MINICPMV_LN, "post", "bias")); + } break; + case PROJECTOR_TYPE_GLM_EDGE: + { + vision_model.mm_model_adapter_conv_w = get_tensor(string_format(TN_GLM_ADAPER_CONV, "weight")); + vision_model.mm_model_adapter_conv_b = get_tensor(string_format(TN_GLM_ADAPER_CONV, "bias")); + vision_model.mm_model_mlp_0_w = get_tensor(string_format(TN_GLM_ADAPTER_LINEAR,"weight")); + vision_model.mm_model_ln_q_w = get_tensor(string_format(TN_GLM_ADAPTER_NORM_1,"weight")); + vision_model.mm_model_ln_q_b = get_tensor(string_format(TN_GLM_ADAPTER_NORM_1,"bias")); + vision_model.mm_model_mlp_1_w = get_tensor(string_format(TN_GLM_ADAPTER_D_H_2_4H,"weight")); + vision_model.mm_model_mlp_2_w = get_tensor(string_format(TN_GLM_ADAPTER_GATE,"weight")); + vision_model.mm_model_mlp_3_w = get_tensor(string_format(TN_GLM_ADAPTER_D_4H_2_H,"weight")); + vision_model.boi_w = get_tensor(TN_GLM_BOI_W); + vision_model.eoi_w = get_tensor(TN_GLM_EOI_W); + } break; + case PROJECTOR_TYPE_MERGER: + { + vision_model.mm_0_w = get_tensor(string_format(TN_LLAVA_PROJ, 0, "weight")); + vision_model.mm_0_b = get_tensor(string_format(TN_LLAVA_PROJ, 0, "bias")); + vision_model.mm_1_w = get_tensor(string_format(TN_LLAVA_PROJ, 2, "weight")); + vision_model.mm_1_b = get_tensor(string_format(TN_LLAVA_PROJ, 2, "bias")); + } break; + case PROJECTOR_TYPE_GEMMA3: + { + vision_model.mm_input_proj_w = get_tensor(TN_MM_INP_PROJ); + vision_model.mm_soft_emb_norm_w = get_tensor(TN_MM_SOFT_EMB_N); + } break; + default: + GGML_ASSERT(false && "unknown projector type"); + } + + // load data + { + std::vector read_buf; + + auto fin = std::ifstream(fname, std::ios::binary); + if (!fin) { + throw std::runtime_error(string_format("%s: failed to open %s\n", __func__, fname.c_str())); + } + + // alloc memory and offload data + ggml_backend_buffer_type_t buft = ggml_backend_get_default_buffer_type(ctx_clip.backend); + ctx_clip.buf.reset(ggml_backend_alloc_ctx_tensors_from_buft(ctx_clip.ctx_data.get(), buft)); + ggml_backend_buffer_set_usage(ctx_clip.buf.get(), GGML_BACKEND_BUFFER_USAGE_WEIGHTS); + for (auto & t : tensors_to_load) { + struct ggml_tensor * cur = ggml_get_tensor(ctx_clip.ctx_data.get(), t->name); + const size_t offset = tensor_offset[t->name]; + fin.seekg(offset, std::ios::beg); + if (!fin) { + throw std::runtime_error(string_format("%s: failed to seek for tensor %s\n", __func__, t->name)); + } + size_t num_bytes = ggml_nbytes(cur); + if (ggml_backend_buft_is_host(buft)) { + // for the CPU and Metal backend, we can read directly into the tensor + fin.read(reinterpret_cast(cur->data), num_bytes); + } else { + // read into a temporary buffer first, then copy to device memory + read_buf.resize(num_bytes); + fin.read(reinterpret_cast(read_buf.data()), num_bytes); + ggml_backend_tensor_set(cur, read_buf.data(), 0, num_bytes); + } + } + fin.close(); + + LOG_DBG("%s: loaded %zu tensors from %s\n", __func__, tensors_to_load.size(), fname.c_str()); } } - ggml_free(meta); + void alloc_compute_meta() { + ctx_clip.buf_compute_meta.resize(GGML_DEFAULT_GRAPH_SIZE * ggml_tensor_overhead() + ggml_graph_overhead()); - new_clip->ctx_gguf = ctx; - - // measure mem requirement and allocate - { - new_clip->buf_compute_meta.resize(GGML_DEFAULT_GRAPH_SIZE * ggml_tensor_overhead() + ggml_graph_overhead()); - new_clip->compute_alloc = ggml_gallocr_new(ggml_backend_get_default_buffer_type(new_clip->backend)); + // create a fake batch clip_image_f32_batch batch; - batch.size = 1; - batch.data = nullptr; - ggml_cgraph * gf = clip_image_build_graph(new_clip, &batch, nullptr, false); - ggml_gallocr_reserve(new_clip->compute_alloc, gf); - size_t compute_memory_buffer_size = ggml_gallocr_get_buffer_size(new_clip->compute_alloc, 0); - LOG_INF("%s: compute allocated memory: %.2f MB\n", __func__, compute_memory_buffer_size /1024.0/1024.0); + clip_image_f32_ptr img(clip_image_f32_init()); + clip_image_size image_size; + image_size.width = clip_get_image_size(&ctx_clip); + image_size.height = clip_get_image_size(&ctx_clip); + int n_patches = clip_get_image_size(&ctx_clip) / image_size.width; + img->nx = n_patches; + img->ny = n_patches; + img->buf.resize(n_patches * image_size.width * image_size.height * 3); + batch.entries.push_back(std::move(img)); + + ggml_cgraph * gf = clip_image_build_graph(&ctx_clip, batch, image_size, false); + ggml_backend_sched_reserve(ctx_clip.sched.get(), gf); + for (size_t i = 0; i < ctx_clip.backend_ptrs.size(); ++i) { + ggml_backend_t backend = ctx_clip.backend_ptrs[i]; + ggml_backend_buffer_type_t buft = ctx_clip.backend_buft[i]; + size_t size = ggml_backend_sched_get_buffer_size(ctx_clip.sched.get(), backend); + if (size > 1) { + LOG_INF("%s: %10s compute buffer size = %8.2f MiB\n", __func__, + ggml_backend_buft_name(buft), + size / 1024.0 / 1024.0); + } + } } - return new_clip; + void get_bool(const std::string & key, bool & output, bool required = true) { + const int i = gguf_find_key(ctx_gguf.get(), key.c_str()); + if (i < 0) { + if (required) throw std::runtime_error("Key not found: " + key); + return; + } + output = gguf_get_val_bool(ctx_gguf.get(), i); + } + + void get_i32(const std::string & key, int & output, bool required = true) { + const int i = gguf_find_key(ctx_gguf.get(), key.c_str()); + if (i < 0) { + if (required) throw std::runtime_error("Key not found: " + key); + return; + } + output = gguf_get_val_i32(ctx_gguf.get(), i); + } + + void get_u32(const std::string & key, int & output, bool required = true) { + const int i = gguf_find_key(ctx_gguf.get(), key.c_str()); + if (i < 0) { + if (required) throw std::runtime_error("Key not found: " + key); + return; + } + output = gguf_get_val_u32(ctx_gguf.get(), i); + } + + void get_f32(const std::string & key, float & output, bool required = true) { + const int i = gguf_find_key(ctx_gguf.get(), key.c_str()); + if (i < 0) { + if (required) throw std::runtime_error("Key not found: " + key); + return; + } + output = gguf_get_val_f32(ctx_gguf.get(), i); + } + + void get_string(const std::string & key, std::string & output, bool required = true) { + const int i = gguf_find_key(ctx_gguf.get(), key.c_str()); + if (i < 0) { + if (required) throw std::runtime_error("Key not found: " + key); + return; + } + output = std::string(gguf_get_val_str(ctx_gguf.get(), i)); + } + + void get_arr_int(const std::string & key, std::vector & output, bool required = true) { + const int i = gguf_find_key(ctx_gguf.get(), key.c_str()); + if (i < 0) { + if (required) throw std::runtime_error("Key not found: " + key); + return; + } + int n = gguf_get_arr_n(ctx_gguf.get(), i); + output.resize(n); + const int32_t * values = (const int32_t *)gguf_get_arr_data(ctx_gguf.get(), i); + for (int i = 0; i < n; ++i) { + output[i] = values[i]; + } + } +}; + +// read and create ggml_context containing the tensors and their data +struct clip_ctx * clip_model_load(const char * fname, const int verbosity) { + return clip_init(fname, clip_context_params{ + /* use_gpu */ true, + /* verbosity */ static_cast(verbosity), + }); +} + +struct clip_ctx * clip_init(const char * fname, struct clip_context_params ctx_params) { + g_logger_state.verbosity_thold = ctx_params.verbosity; + clip_ctx * ctx_clip = new clip_ctx(ctx_params); + + try { + clip_model_loader loader(fname, *ctx_clip); + loader.load_hparams(); + loader.load_tensors(); + loader.alloc_compute_meta(); + } catch (const std::exception & e) { + LOG_ERR("%s: failed to load model '%s': %s\n", __func__, fname, e.what()); + delete ctx_clip; + return nullptr; + } + + return ctx_clip; } void clip_add_load_image_size(struct clip_ctx * ctx_clip, struct clip_image_size * load_image_size) { - ctx_clip->load_image_size = load_image_size; + ctx_clip->load_image_size = *load_image_size; // copy } struct clip_image_size * clip_get_load_image_size(struct clip_ctx * ctx_clip) { - return ctx_clip->load_image_size; + return &ctx_clip->load_image_size; } struct clip_image_size * clip_image_size_init() { @@ -1714,26 +1600,60 @@ struct clip_image_f32 * clip_image_f32_init() { return new clip_image_f32(); } -void clip_image_u8_free(struct clip_image_u8 * img) { delete img; } -void clip_image_f32_free(struct clip_image_f32 * img) { delete img; } -void clip_image_u8_batch_free(struct clip_image_u8_batch * batch) { - if (batch->size > 0) { - delete[] batch->data; - batch->size = 0; - } -} -void clip_image_f32_batch_free(struct clip_image_f32_batch * batch) { - if (batch->size > 0) { - delete[] batch->data; - batch->size = 0; - } +struct clip_image_f32_batch * clip_image_f32_batch_init() { + return new clip_image_f32_batch(); } -static void build_clip_img_from_data(const stbi_uc * data, int nx, int ny, clip_image_u8 * img) { +unsigned char * clip_image_u8_get_data(struct clip_image_u8 * img, uint32_t * nx, uint32_t * ny) { + if (nx) *nx = img->nx; + if (ny) *ny = img->ny; + return img->buf.data(); +} + +void clip_image_size_free(struct clip_image_size * load_image_size) { + if (load_image_size == nullptr) { + return; + } + delete load_image_size; +} +void clip_image_u8_free(struct clip_image_u8 * img) { if (img) delete img; } +void clip_image_f32_free(struct clip_image_f32 * img) { if (img) delete img; } +void clip_image_u8_batch_free(struct clip_image_u8_batch * batch) { if (batch) delete batch; } +void clip_image_f32_batch_free(struct clip_image_f32_batch * batch) { if (batch) delete batch; } + +size_t clip_image_f32_batch_n_images(const struct clip_image_f32_batch * batch) { + return batch->entries.size(); +} + +size_t clip_image_f32_batch_nx(const struct clip_image_f32_batch * batch, int idx) { + if (idx < 0 || idx >= (int)batch->entries.size()) { + LOG_ERR("%s: invalid index %d\n", __func__, idx); + return 0; + } + return batch->entries[idx]->nx; +} + +size_t clip_image_f32_batch_ny(const struct clip_image_f32_batch * batch, int idx) { + if (idx < 0 || idx >= (int)batch->entries.size()) { + LOG_ERR("%s: invalid index %d\n", __func__, idx); + return 0; + } + return batch->entries[idx]->ny; +} + +clip_image_f32 * clip_image_f32_get_img(const struct clip_image_f32_batch * batch, int idx) { + if (idx < 0 || idx >= (int)batch->entries.size()) { + LOG_ERR("%s: invalid index %d\n", __func__, idx); + return nullptr; + } + return batch->entries[idx].get(); +} + +void clip_build_img_from_pixels(const unsigned char * rgb_pixels, int nx, int ny, clip_image_u8 * img) { img->nx = nx; img->ny = ny; img->buf.resize(3 * nx * ny); - memcpy(img->buf.data(), data, img->buf.size()); + memcpy(img->buf.data(), rgb_pixels, img->buf.size()); } bool clip_image_load_from_file(const char * fname, clip_image_u8 * img) { @@ -1743,7 +1663,7 @@ bool clip_image_load_from_file(const char * fname, clip_image_u8 * img) { LOG_ERR("%s: failed to load image '%s'\n", __func__, fname); return false; } - build_clip_img_from_data(data, nx, ny, img); + clip_build_img_from_pixels(data, nx, ny, img); stbi_image_free(data); return true; } @@ -1755,7 +1675,7 @@ bool clip_image_load_from_bytes(const unsigned char * bytes, size_t bytes_length LOG_ERR("%s: failed to decode image bytes\n", __func__); return false; } - build_clip_img_from_data(data, nx, ny, img); + clip_build_img_from_pixels(data, nx, ny, img); stbi_image_free(data); return true; } @@ -1800,14 +1720,15 @@ static void bilinear_resize(const clip_image_u8& src, clip_image_u8& dst, int ta } // Normalize image to float32 - careful with pytorch .to(model.device, dtype=torch.float16) - this sometimes reduces precision (32>16>32), sometimes not -static void normalize_image_u8_to_f32(const clip_image_u8* src, clip_image_f32* dst, const float mean[3], const float std[3]) { - dst->nx = src->nx; - dst->ny = src->ny; - dst->buf.resize(src->buf.size()); +static void normalize_image_u8_to_f32(const clip_image_u8 & src, clip_image_f32 & dst, const float mean[3], const float std[3]) { + dst.nx = src.nx; + dst.ny = src.ny; + dst.buf.resize(src.buf.size()); - for (size_t i = 0; i < src->buf.size(); ++i) { + // TODO @ngxson : seems like this could be done more efficiently on cgraph + for (size_t i = 0; i < src.buf.size(); ++i) { int c = i % 3; // rgb - dst->buf[i] = (static_cast(src->buf[i]) / 255.0f - mean[c]) / std[c]; + dst.buf[i] = (static_cast(src.buf[i]) / 255.0f - mean[c]) / std[c]; } } @@ -1815,7 +1736,7 @@ inline int clip(int x, int lower, int upper) { return std::max(lower, std::min(x, upper)); } -static bool bicubic_resize(const clip_image_u8 &img, clip_image_u8 &dst, int target_width, int target_height) { +static bool bicubic_resize(const clip_image_u8 & img, clip_image_u8 & dst, int target_width, int target_height) { const int nx = img.nx; const int ny = img.ny; @@ -1953,13 +1874,13 @@ static std::pair select_best_resolution(const std::pair & or return best_fit; } -static std::vector divide_to_patches_u8(const clip_image_u8 & image, int patch_size) { - std::vector patches; +static std::vector divide_to_patches_u8(const clip_image_u8 & image, int patch_size) { + std::vector patches; int width = image.nx; int height = image.ny; for (int i = 0; i < height; i += patch_size) { for (int j = 0; j < width; j += patch_size) { - clip_image_u8 *patch = clip_image_u8_init(); + clip_image_u8_ptr patch(clip_image_u8_init()); patch->nx = std::min(patch_size, width - j); patch->ny = std::min(patch_size, height - i); patch->buf.resize(3 * patch->nx * patch->ny); @@ -1970,7 +1891,7 @@ static std::vector divide_to_patches_u8(const clip_image_u8 & im } } } - patches.push_back(patch); + patches.push_back(std::move(patch)); } } return patches; @@ -2051,7 +1972,7 @@ static std::pair uhd_best_grid(const int max_slice_nums, const int mul // -> https://arxiv.org/pdf/2403.11703 // -> https://github.com/thunlp/LLaVA-UHD // -> https://github.com/thunlp/LLaVA-UHD/blob/302301bc2175f7e717fb8548516188e89f649753/llava_uhd/train/llava-uhd/slice_logic.py#L118 -static std::vector> uhd_slice_image(const clip_image_u8 * img, const int max_slice_nums=9, const int scale_resolution=448, const int patch_size=14) { +static std::vector> uhd_slice_image(const clip_image_u8 * img, const int max_slice_nums=9, const int scale_resolution=448, const int patch_size=14) { const std::pair original_size={img->nx,img->ny}; const int original_width = img->nx; const int original_height = img->ny; @@ -2059,33 +1980,33 @@ static std::vector> uhd_slice_image(const clip_imag const float ratio = 1.0 * original_width * original_height/ (scale_resolution * scale_resolution); const int multiple = fmin(ceil(ratio), max_slice_nums); - std::vector> images; - LOG_INF("%s: multiple %d\n", __func__, multiple); - images.push_back(std::vector()); + std::vector> images; + LOG_DBG("%s: multiple %d\n", __func__, multiple); + images.push_back(std::vector()); if (multiple <= 1) { auto best_size = uhd_find_best_resize(original_size, scale_resolution, patch_size, true); - clip_image_u8 * source_image = clip_image_u8_init(); + clip_image_u8_ptr source_image(clip_image_u8_init()); bicubic_resize(*img, *source_image, best_size.first, best_size.second); // source_image = image.resize(best_size, Image.Resampling.BICUBIC) - images[images.size()-1].push_back(source_image); + images.back().push_back(std::move(source_image)); } else if (multiple > 1) { auto best_size = uhd_find_best_resize(original_size, scale_resolution, patch_size); - clip_image_u8 * source_image = clip_image_u8_init(); + clip_image_u8_ptr source_image(clip_image_u8_init()); bicubic_resize(*img, *source_image, best_size.first, best_size.second); // source_image = image.copy().resize(best_resize, Image.Resampling.BICUBIC) - LOG_INF("%s: image_size: %d %d; source_image size: %d %d\n", __func__, img->nx, img->ny, best_size.first, best_size.second); - images[images.size()-1].push_back(source_image); + LOG_DBG("%s: image_size: %d %d; source_image size: %d %d\n", __func__, img->nx, img->ny, best_size.first, best_size.second); + images.back().push_back(std::move(source_image)); std::pair best_grid = uhd_best_grid(max_slice_nums, multiple, log_ratio); - LOG_INF("%s: image_size: %d %d; best_grid: %d %d\n", __func__, img->nx, img->ny, best_grid.first, best_grid.second); + LOG_DBG("%s: image_size: %d %d; best_grid: %d %d\n", __func__, img->nx, img->ny, best_grid.first, best_grid.second); auto refine_size = uhd_get_refine_size(original_size, best_grid, scale_resolution, patch_size, true); - clip_image_u8 * refine_image = clip_image_u8_init(); + clip_image_u8_ptr refine_image(clip_image_u8_init()); bicubic_resize(*img, *refine_image, refine_size.first, refine_size.second); - LOG_INF("%s: refine_image_size: %d %d; refine_size: %d %d\n", __func__, refine_image->nx, refine_image->ny, refine_size.first, refine_size.second); + LOG_DBG("%s: refine_image_size: %d %d; refine_size: %d %d\n", __func__, refine_image->nx, refine_image->ny, refine_size.first, refine_size.second); // split_to_patches int width = refine_image->nx; @@ -2093,9 +2014,9 @@ static std::vector> uhd_slice_image(const clip_imag int grid_x = int(width / best_grid.first); int grid_y = int(height / best_grid.second); for (int patches_i = 0, ic = 0; patches_i < height && ic < best_grid.second; patches_i += grid_y, ic += 1){ - images.push_back(std::vector()); + images.push_back(std::vector()); for(int patches_j = 0, jc = 0; patches_j < width && jc < best_grid.first; patches_j += grid_x, jc += 1){ - clip_image_u8 * patch = clip_image_u8_init(); + clip_image_u8_ptr patch(clip_image_u8_init()); patch->nx = grid_x; patch->ny = grid_y; patch->buf.resize(3 * patch->nx * patch->ny); @@ -2108,10 +2029,9 @@ static std::vector> uhd_slice_image(const clip_imag patch->buf[j+2] = refine_image->buf[i+2]; } } - images[images.size()-1].push_back(patch); + images.back().push_back(std::move(patch)); } } - clip_image_u8_free(refine_image); } return images; } @@ -2119,8 +2039,8 @@ static std::vector> uhd_slice_image(const clip_imag int clip_uhd_num_image_embeds_col(struct clip_ctx * ctx_clip) { const int max_slice_nums=9; const int scale_resolution=448; - const int original_width = ctx_clip->load_image_size->width; - const int original_height = ctx_clip->load_image_size->height; + const int original_width = ctx_clip->load_image_size.width; + const int original_height = ctx_clip->load_image_size.height; const float log_ratio = log(1.0*original_width/original_height); const float ratio = 1.0 * original_width * original_height/ (scale_resolution * scale_resolution); const int multiple = fmin(ceil(ratio), max_slice_nums); @@ -2130,88 +2050,64 @@ int clip_uhd_num_image_embeds_col(struct clip_ctx * ctx_clip) { // returns the normalized float tensor for llava-1.5, for spatial_unpad with anyres processing for llava-1.6 it returns the normalized image patch tensors as a vector // res_imgs memory is being allocated here, previous allocations will be freed if found -bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, clip_image_f32_batch * res_imgs) { +bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, struct clip_image_f32_batch * res_imgs) { - if(clip_is_minicpmv(ctx)){ + if (clip_is_minicpmv(ctx)) { int max_slice_nums = 9; - std::vector> imgs = uhd_slice_image(img, max_slice_nums); - res_imgs->size = 0; - for (size_t i = 0; i < imgs.size(); ++i){ - res_imgs->size += imgs[i].size(); - } - res_imgs->data = new clip_image_f32[res_imgs->size]; - int idx = 0; + std::vector> imgs = uhd_slice_image(img, max_slice_nums); for (size_t i = 0; i < imgs.size(); ++i) { for (size_t j = 0; j < imgs[i].size(); ++j) { LOG_DBG("%s: %d %d\n", __func__,imgs[i][j]->nx,imgs[i][j]->ny); - clip_image_f32 * res = clip_image_f32_init(); - normalize_image_u8_to_f32(imgs[i][j], res, ctx->image_mean, ctx->image_std); - res_imgs->data[idx++] = *res; - clip_image_f32_free(res); - } - } - for (size_t i = 0; i < imgs.size(); ++i) { - for (size_t j = 0; j < imgs[i].size(); ++j) { - if (imgs[i][j] != nullptr) { - clip_image_u8_free(imgs[i][j]); - } + clip_image_f32_ptr res(clip_image_f32_init()); + normalize_image_u8_to_f32(*imgs[i][j], *res, ctx->image_mean, ctx->image_std); + res_imgs->entries.push_back(std::move(res)); } } return true; } else if (ctx->has_qwen2vl_merger) { - clip_image_u8 * resized = clip_image_u8_init(); - auto patch_size = clip_patch_size(ctx) * 2; + clip_image_u8 resized; + auto patch_size = clip_get_patch_size(ctx) * 2; int nx = ceil((float)img->nx / patch_size) * patch_size; int ny = ceil((float)img->ny / patch_size) * patch_size; - bicubic_resize(*img, *resized, nx, ny); + bicubic_resize(*img, resized, nx, ny); - res_imgs->data = new clip_image_f32[1]; - // clip_image_f32 * res = clip_image_f32_init(); - normalize_image_u8_to_f32(resized, res_imgs->data, ctx->image_mean, ctx->image_std); + clip_image_f32_ptr img_f32(clip_image_f32_init()); + // clip_image_f32_ptr res(clip_image_f32_init()); + normalize_image_u8_to_f32(resized, *img_f32, ctx->image_mean, ctx->image_std); // res_imgs->data[0] = *res; - res_imgs->size = 1; - - // clip_image_f32_free(res); - clip_image_u8_free(resized); + res_imgs->entries.push_back(std::move(img_f32)); return true; } - if (ctx->has_glm_projector) { - res_imgs->size = 1; - res_imgs->data = new clip_image_f32[res_imgs->size]; + if (ctx->has_glm_projector || ctx->proj_type == PROJECTOR_TYPE_GEMMA3) { clip_image_u8 resized_image; int32_t sz=ctx->vision_model.hparams.image_size; bicubic_resize(*img, resized_image,sz,sz); - clip_image_f32 * res = clip_image_f32_init(); + clip_image_f32_ptr img_f32(clip_image_f32_init()); //clip_image_save_to_bmp(resized_image, "resized.bmp"); - normalize_image_u8_to_f32(&resized_image, res, ctx->image_mean, ctx->image_std); - res_imgs->data[0] = *res; - clip_image_f32_free(res); + normalize_image_u8_to_f32(resized_image, *img_f32, ctx->image_mean, ctx->image_std); + res_imgs->entries.push_back(std::move(img_f32)); return true; } bool pad_to_square = true; if (!ctx->has_vision_encoder) { - LOG_ERR("This gguf file seems to have no vision encoder\n"); + LOG_ERR("%s: This gguf file seems to have no vision encoder\n", __func__); return false; } auto & params = ctx->vision_model.hparams; // The model config actually contains all we need to decide on how to preprocess, here we automatically switch to the new llava-1.6 preprocessing - if (strcmp(params.mm_patch_merge_type, "spatial_unpad") == 0) { + if (params.mm_patch_merge_type == PATCH_MERGE_SPATIAL_UNPAD) { pad_to_square = false; } // free the previous res_imgs if any set - if (res_imgs->size > 0) { - clip_image_f32_batch_free(res_imgs); - } - res_imgs->data = nullptr; - res_imgs->size = 0; + res_imgs->entries.clear(); // the logic below is to pad the shorter side to the longer side with a background color: rgb(122, 116, 104) // see https://github.com/haotian-liu/LLaVA/blob/e854a2bf85118c504f6f16bf5c3c7c92f8fa8c6b/llava/conversation.py#L113-L156 - clip_image_u8 * temp = clip_image_u8_init(); // we will keep the input image data here temporarily + clip_image_u8_ptr temp(clip_image_u8_init()); // we will keep the input image data here temporarily if (pad_to_square && img->nx != img->ny) { int longer_side = std::max(img->nx, img->ny); temp->nx = longer_side; @@ -2235,10 +2131,10 @@ bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, cli } } } else { - if (params.image_grid_pinpoints[0] != 0) { + if (!params.image_grid_pinpoints.empty()) { // "spatial_unpad" with "anyres" processing for llava-1.6 std::vector> possible_resolutions; - for (int i = 0; i < 32 && params.image_grid_pinpoints[i] != 0; i+=2) { + for (size_t i = 0; i < params.image_grid_pinpoints.size(); i+=2) { possible_resolutions.push_back({params.image_grid_pinpoints[i], params.image_grid_pinpoints[i+1]}); } std::pair best_resolution = select_best_resolution({img->nx, img->ny}, possible_resolutions); @@ -2254,28 +2150,18 @@ bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, cli // clip_image_u8_free(temp2); // } - std::vector patches = divide_to_patches_u8(*temp, params.image_size); // prepare spatial sorted main patches of image_size each (336 in llava-1.6) + std::vector patches = divide_to_patches_u8(*temp, params.image_size); // prepare spatial sorted main patches of image_size each (336 in llava-1.6) - clip_image_u8 *image_original_resize = clip_image_u8_init(); + clip_image_u8_ptr image_original_resize(clip_image_u8_init()); // bilinear_resize(*img, *image_original_resize, params.image_size, params.image_size); // in python this is "shortest_edge", but all CLIP are square bicubic_resize(*img, *image_original_resize, params.image_size, params.image_size); // in python this is "shortest_edge", but all CLIP are square - patches.insert(patches.begin(), image_original_resize); - // clip_image_f32_batch_init(patches.size()); - res_imgs->size = patches.size(); - res_imgs->data = new clip_image_f32[res_imgs->size]; - int num=0; - for (auto& patch : patches) { - normalize_image_u8_to_f32(patch, &res_imgs->data[num], ctx->image_mean, ctx->image_std); - num++; + patches.insert(patches.begin(), std::move(image_original_resize)); + for (auto & patch : patches) { + clip_image_f32_ptr res(clip_image_f32_init()); + normalize_image_u8_to_f32(*patch, *res, ctx->image_mean, ctx->image_std); + res_imgs->entries.push_back(std::move(res)); } - for (size_t i = 0; i < patches.size(); i++) { - // LOG_DBG("patch %d: %d %d\n", i, patches[i]->nx, patches[i]->ny); - clip_image_u8_free(patches[i]); - } - - clip_image_u8_free(temp); - return true; } else { temp->nx = img->nx; @@ -2291,7 +2177,7 @@ bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, cli const int nx2 = ctx->vision_model.hparams.image_size; const int ny2 = ctx->vision_model.hparams.image_size; - clip_image_f32 * res = clip_image_f32_init(); + clip_image_f32_ptr res(clip_image_f32_init()); res->nx = nx2; res->ny = ny2; res->buf.resize(3 * nx2 * ny2); @@ -2343,7 +2229,6 @@ bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, cli } } } - clip_image_u8_free(temp); // { // clip_image_u8 * temp2 = clip_image_u8_init(); @@ -2353,10 +2238,7 @@ bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, cli // } // res_imgs.push_back(res); - res_imgs->size = 1; - res_imgs->data = new clip_image_f32[res_imgs->size]; - res_imgs->data[0] = *res; - clip_image_f32_free(res); + res_imgs->entries.push_back(std::move(res)); return true; } @@ -2366,12 +2248,9 @@ ggml_tensor * clip_get_newline_tensor(const struct clip_ctx * ctx) { } void clip_free(clip_ctx * ctx) { - ggml_free(ctx->ctx_data); - gguf_free(ctx->ctx_gguf); - - ggml_backend_buffer_free(ctx->params_buffer); - ggml_backend_free(ctx->backend); - ggml_gallocr_free(ctx->compute_alloc); + if (ctx == nullptr) { + return; + } delete ctx; } @@ -2387,24 +2266,31 @@ size_t clip_embd_nbytes_by_img(const struct clip_ctx * ctx, int img_h, int img_w return clip_n_patches_by_img(ctx, &img) * clip_n_mmproj_embd(ctx) * sizeof(float); } -int32_t clip_image_size(const struct clip_ctx * ctx) { +int32_t clip_get_image_size(const struct clip_ctx * ctx) { return ctx->vision_model.hparams.image_size; } -int32_t clip_patch_size(const struct clip_ctx * ctx) { +int32_t clip_get_patch_size(const struct clip_ctx * ctx) { return ctx->vision_model.hparams.patch_size; } -int32_t clip_hidden_size(const struct clip_ctx * ctx) { +int32_t clip_get_hidden_size(const struct clip_ctx * ctx) { return ctx->vision_model.hparams.hidden_size; } const char * clip_patch_merge_type(const struct clip_ctx * ctx) { - return ctx->vision_model.hparams.mm_patch_merge_type; + return ctx->vision_model.hparams.mm_patch_merge_type == PATCH_MERGE_SPATIAL_UNPAD ? "spatial_unpad" : "flat"; } const int32_t * clip_image_grid(const struct clip_ctx * ctx) { - return ctx->vision_model.hparams.image_grid_pinpoints; + if (ctx->vision_model.hparams.image_grid_pinpoints.size()) { + return &ctx->vision_model.hparams.image_grid_pinpoints.front(); + } + return nullptr; +} + +size_t get_clip_image_grid_size(const struct clip_ctx * ctx) { + return ctx->vision_model.hparams.image_grid_pinpoints.size(); } int clip_n_patches(const struct clip_ctx * ctx) { @@ -2436,6 +2322,8 @@ int clip_n_patches_by_img(const struct clip_ctx * ctx, struct clip_image_f32 * i int x_patch = img->nx / patch_size + (int)(img->nx % patch_size > 0); int y_patch = img->ny / patch_size + (int)(img->ny % patch_size > 0); n_patches = x_patch * y_patch; + } else if (ctx->proj_type == PROJECTOR_TYPE_GEMMA3) { + n_patches = 256; } return n_patches; @@ -2529,23 +2417,27 @@ static std::vector> get_2d_sincos_pos_embed(int embed_dim, co bool clip_image_encode(struct clip_ctx * ctx, const int n_threads, clip_image_f32 * img, float * vec) { if (!ctx->has_vision_encoder) { - LOG_ERR("This gguf file seems to have no vision encoder\n"); + LOG_ERR("%s: This gguf file seems to have no vision encoder\n", __func__); return false; } - clip_image_f32_batch imgs{}; - imgs.size = 1; - imgs.data = img; + clip_image_f32_batch imgs; + clip_image_f32_ptr img_copy(clip_image_f32_init()); + *img_copy = *img; + imgs.entries.push_back(std::move(img_copy)); + return clip_image_batch_encode(ctx, n_threads, &imgs, vec); } -bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_image_f32_batch * imgs, float * vec) { +bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_image_f32_batch * imgs_c_ptr, float * vec) { + const clip_image_f32_batch & imgs = *imgs_c_ptr; + if (!ctx->has_vision_encoder) { - LOG_ERR("This gguf file seems to have no vision encoder\n"); + LOG_ERR("%s: This gguf file seems to have no vision encoder\n", __func__); return false; } - int batch_size = imgs->size; + int batch_size = imgs.entries.size(); if (ctx->has_llava_projector) { GGML_ASSERT(batch_size == 1); // TODO: support multiple images } @@ -2560,8 +2452,9 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima } // build the inference graph + ggml_backend_sched_reset(ctx->sched.get()); ggml_cgraph * gf = clip_image_build_graph(ctx, imgs, ctx->load_image_size, true); - ggml_gallocr_alloc_graph(ctx->compute_alloc, gf); + ggml_backend_sched_alloc_graph(ctx->sched.get(), gf); // set inputs const auto & model = ctx->vision_model; @@ -2571,25 +2464,22 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima int image_size_width = image_size; int image_size_height = image_size; if (ctx->has_minicpmv_projector | ctx->has_qwen2vl_merger) { - image_size_width = imgs->data[0].nx; - image_size_height = imgs->data[0].ny; + image_size_width = imgs.entries[0]->nx; + image_size_height = imgs.entries[0]->ny; } const int patch_size = hparams.patch_size; const int num_patches = ((image_size_width / patch_size) * (image_size_height / patch_size)); - const int num_positions = num_patches + (ctx->has_class_embedding ? 1 : 0); - if(ctx->load_image_size==nullptr){ - ctx->load_image_size= clip_image_size_init(); - } - const int pos_w = ctx->load_image_size->width/patch_size; - const int pos_h = ctx->load_image_size->height/patch_size; + const int num_positions = num_patches + (model.class_embedding ? 1 : 0); + const int pos_w = ctx->load_image_size.width / patch_size; + const int pos_h = ctx->load_image_size.height / patch_size; { struct ggml_tensor * inp_raw = ggml_graph_get_tensor(gf, "inp_raw"); float * data = (float *)malloc(ggml_nbytes(inp_raw)); - for (size_t i = 0; i < imgs->size; i++) { - const int nx = imgs->data[i].nx; - const int ny = imgs->data[i].ny; + for (size_t i = 0; i < imgs.entries.size(); i++) { + const int nx = imgs.entries[i]->nx; + const int ny = imgs.entries[i]->ny; if (!(ctx->has_minicpmv_projector | ctx->has_qwen2vl_merger)) { GGML_ASSERT(nx == image_size && ny == image_size); } @@ -2600,7 +2490,7 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima for (int k = 0; k < 3; k++) { for (int y = 0; y < ny; y++) { for (int x = 0; x < nx; x++) { - data[(b * 3 * n) + k * n + y * nx + x] = imgs->data[b].buf[3 * (y * nx + x) + k]; + data[(b * 3 * n) + k * n + y * nx + x] = imgs.entries[b]->buf[3 * (y * nx + x) + k]; } } } @@ -2661,16 +2551,14 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima free(pos_embed_data); } } - else{ - { - if (ctx->has_class_embedding) { - struct ggml_tensor * embeddings = ggml_graph_get_tensor(gf, "embeddings"); + else { + if (model.class_embedding) { + struct ggml_tensor * embeddings = ggml_graph_get_tensor(gf, "embeddings"); - void* zero_mem = malloc(ggml_nbytes(embeddings)); - memset(zero_mem, 0, ggml_nbytes(embeddings)); - ggml_backend_tensor_set(embeddings, zero_mem, 0, ggml_nbytes(embeddings)); - free(zero_mem); - } + void* zero_mem = malloc(ggml_nbytes(embeddings)); + memset(zero_mem, 0, ggml_nbytes(embeddings)); + ggml_backend_tensor_set(embeddings, zero_mem, 0, ggml_nbytes(embeddings)); + free(zero_mem); } if (ctx->has_qwen2vl_merger) { @@ -2700,6 +2588,9 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima ggml_backend_tensor_set(positions, positions_data, 0, ggml_nbytes(positions)); free(positions_data); } + else if (ctx->proj_type == PROJECTOR_TYPE_GEMMA3) { + // do nothing + } else { struct ggml_tensor * positions = ggml_graph_get_tensor(gf, "positions"); @@ -2712,9 +2603,13 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima if (!ctx->has_glm_projector) { struct ggml_tensor * patches = ggml_graph_get_tensor(gf, "patches"); + // The patches vector is used to get rows to index into the embeds with; + // we should skip dim 0 only if we have CLS to avoid going out of bounds + // when retrieving the rows. + int patch_offset = model.class_embedding ? 1 : 0; int* patches_data = (int*)malloc(ggml_nbytes(patches)); for (int i = 0; i < num_patches; i++) { - patches_data[i] = i + 1; + patches_data[i] = i + patch_offset; } ggml_backend_tensor_set(patches, patches_data, 0, ggml_nbytes(patches)); free(patches_data); @@ -2722,11 +2617,13 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima } } - if (ggml_backend_is_cpu(ctx->backend)) { - ggml_backend_cpu_set_n_threads(ctx->backend, n_threads); - } + ggml_backend_cpu_set_n_threads(ctx->backend_cpu, n_threads); - ggml_backend_graph_compute(ctx->backend, gf); + auto status = ggml_backend_sched_graph_compute(ctx->sched.get(), gf); + if (status != GGML_STATUS_SUCCESS) { + LOG_ERR("%s: ggml_backend_sched_graph_compute failed with error %d\n", __func__, status); + return false; + } // the last node is the embedding tensor struct ggml_tensor * embeddings = ggml_graph_node(gf, -1); @@ -2748,10 +2645,13 @@ bool clip_model_quantize(const char * fname_inp, const char * fname_out, const i assert(itype < GGML_TYPE_COUNT); ggml_type type = static_cast(itype); - auto * ctx_clip = clip_model_load(fname_inp, 2); + auto * ctx_clip = clip_init(fname_inp, clip_context_params{ + /* use_gpu */ false, + /* verbosity */ GGML_LOG_LEVEL_ERROR, + }); - const auto & ctx_src = ctx_clip->ctx_gguf; - const auto & ctx_data = ctx_clip->ctx_data; + const auto & ctx_src = ctx_clip->ctx_gguf.get(); + const auto & ctx_data = ctx_clip->ctx_data.get(); auto * ctx_out = gguf_init_empty(); gguf_set_kv(ctx_out, ctx_src); @@ -2825,7 +2725,7 @@ bool clip_model_quantize(const char * fname_inp, const char * fname_out, const i f32_data = (float *)conv_buf.data(); break; default: - LOG_ERR("Please use an input file in f32 or f16\n"); + LOG_ERR("%s: Please use an input file in f32 or f16\n", __func__); gguf_free(ctx_out); return false; } @@ -2906,9 +2806,12 @@ int clip_n_mmproj_embd(const struct clip_ctx * ctx) { if (ctx->proj_type == PROJECTOR_TYPE_MERGER) { return ctx->vision_model.mm_1_b->ne[0]; } + if (ctx->proj_type == PROJECTOR_TYPE_GEMMA3) { + return ctx->vision_model.mm_input_proj_w->ne[0]; + } std::string proj_type = PROJECTOR_TYPE_NAMES[ctx->proj_type]; - throw std::runtime_error(format("%s: don't support projector with: %s currently\n", __func__, proj_type.c_str())); + throw std::runtime_error(string_format("%s: don't support projector with: %s currently\n", __func__, proj_type.c_str())); } int clip_is_minicpmv(const struct clip_ctx * ctx) { @@ -2921,10 +2824,41 @@ int clip_is_minicpmv(const struct clip_ctx * ctx) { bool clip_is_glm(const struct clip_ctx * ctx) { return ctx->has_glm_projector; } + bool clip_is_qwen2vl(const struct clip_ctx * ctx) { return ctx->has_qwen2vl_merger; } +bool clip_is_llava(const struct clip_ctx * ctx) { + return ctx->has_llava_projector; +} + +bool clip_is_gemma3(const struct clip_ctx * ctx) { + return ctx->proj_type == PROJECTOR_TYPE_GEMMA3; +} + +// Determine the number of encoder layers to iterate over +int get_deepest_feature_layer(const struct clip_ctx * ctx) { + // Get the index of the second to last layer; this is the + // default for models that have a llava projector + const auto & hparams = ctx->vision_model.hparams; + int n_layer = hparams.n_layer - 1; + int deepest_feature_layer = -1; + + // Handle other projectors; incrementing here indicates that we + // should use the last encoder layer for the vision features. + if (ctx->has_minicpmv_projector || ctx->has_glm_projector || ctx->has_qwen2vl_merger) { + n_layer += 1; + } + + // If we set explicit vision feature layers, only go up to the deepest one + for (const auto & feature_layer : hparams.vision_feature_layer) { + if (feature_layer > deepest_feature_layer) { + deepest_feature_layer = feature_layer; + } + } + return deepest_feature_layer < 0 ? n_layer : deepest_feature_layer; +} bool clip_encode_float_image (struct clip_ctx * ctx, int n_threads, float * img, int h, int w, float * vec) { clip_image_f32 clip_img; @@ -2938,3 +2872,11 @@ bool clip_encode_float_image (struct clip_ctx * ctx, int n_threads, float * img, clip_image_encode(ctx, n_threads, &clip_img, vec); return true; } + +// +// API used internally with mtmd +// + +projector_type clip_get_projector_type(const struct clip_ctx * ctx) { + return ctx->proj_type; +} diff --git a/examples/llava/clip.h b/examples/llava/clip.h index 841b4f6f9..cc133a58d 100644 --- a/examples/llava/clip.h +++ b/examples/llava/clip.h @@ -1,6 +1,7 @@ #ifndef CLIP_H #define CLIP_H +#include "ggml.h" #include #include @@ -29,32 +30,33 @@ struct clip_image_size { int height; }; -struct clip_image_u8_batch { - struct clip_image_u8 * data; - size_t size; +struct clip_image_u8_batch; +struct clip_image_f32_batch; + +struct clip_context_params { + bool use_gpu; + ggml_log_level verbosity; }; -struct clip_image_f32_batch { - struct clip_image_f32 * data; - size_t size; -}; +// deprecated, use clip_init +CLIP_API struct clip_ctx * clip_model_load(const char * fname, int verbosity); -CLIP_API struct clip_ctx * clip_model_load (const char * fname, int verbosity); -CLIP_API struct clip_ctx * clip_model_load_cpu(const char * fname, int verbosity); +CLIP_API struct clip_ctx * clip_init(const char * fname, struct clip_context_params ctx_params); CLIP_API void clip_free(struct clip_ctx * ctx); CLIP_API size_t clip_embd_nbytes(const struct clip_ctx * ctx); CLIP_API size_t clip_embd_nbytes_by_img(const struct clip_ctx * ctx, int img_h, int img_w); -CLIP_API int32_t clip_image_size (const struct clip_ctx * ctx); -CLIP_API int32_t clip_patch_size (const struct clip_ctx * ctx); -CLIP_API int32_t clip_hidden_size(const struct clip_ctx * ctx); +CLIP_API int32_t clip_get_image_size (const struct clip_ctx * ctx); +CLIP_API int32_t clip_get_patch_size (const struct clip_ctx * ctx); +CLIP_API int32_t clip_get_hidden_size(const struct clip_ctx * ctx); // TODO: should be enum, not string CLIP_API const char * clip_patch_merge_type(const struct clip_ctx * ctx); CLIP_API const int32_t * clip_image_grid(const struct clip_ctx * ctx); +CLIP_API size_t get_clip_image_grid_size(const struct clip_ctx * ctx); CLIP_API int clip_n_patches (const struct clip_ctx * ctx); CLIP_API int clip_n_patches_by_img (const struct clip_ctx * ctx, struct clip_image_f32 * img); @@ -64,15 +66,32 @@ CLIP_API int clip_uhd_num_image_embeds_col(struct clip_ctx * ctx_clip); CLIP_API void clip_add_load_image_size(struct clip_ctx * ctx_clip, struct clip_image_size * load_image_size); CLIP_API struct clip_image_size * clip_get_load_image_size(struct clip_ctx * ctx_clip); -CLIP_API struct clip_image_size * clip_image_size_init(); -CLIP_API struct clip_image_u8 * clip_image_u8_init (); -CLIP_API struct clip_image_f32 * clip_image_f32_init(); +CLIP_API struct clip_image_size * clip_image_size_init(); +CLIP_API struct clip_image_u8 * clip_image_u8_init (); +CLIP_API struct clip_image_f32 * clip_image_f32_init(); +CLIP_API struct clip_image_f32_batch * clip_image_f32_batch_init(); // only used by libllava +// nx, ny are the output image dimensions +CLIP_API unsigned char * clip_image_u8_get_data(struct clip_image_u8 * img, uint32_t * nx, uint32_t * ny); + +CLIP_API void clip_image_size_free (struct clip_image_size * img_size); CLIP_API void clip_image_u8_free (struct clip_image_u8 * img); CLIP_API void clip_image_f32_free(struct clip_image_f32 * img); CLIP_API void clip_image_u8_batch_free (struct clip_image_u8_batch * batch); CLIP_API void clip_image_f32_batch_free(struct clip_image_f32_batch * batch); +// use for accessing underlay data of clip_image_f32_batch +CLIP_API size_t clip_image_f32_batch_n_images(const struct clip_image_f32_batch * batch); // equivalent to batch->size() +CLIP_API size_t clip_image_f32_batch_nx(const struct clip_image_f32_batch * batch, int idx); // equivalent to batch[idx]->nx +CLIP_API size_t clip_image_f32_batch_ny(const struct clip_image_f32_batch * batch, int idx); // equivalent to batch[idx]->ny +CLIP_API clip_image_f32 * clip_image_f32_get_img(const struct clip_image_f32_batch * batch, int idx); // equivalent to batch[idx]->data + +/** + * Build image from pixels decoded by other libraries instead of stb_image.h for better performance. + * The memory layout is RGBRGBRGB..., input buffer length must be 3*nx*ny bytes + */ +CLIP_API void clip_build_img_from_pixels(const unsigned char * rgb_pixels, int nx, int ny, struct clip_image_u8 * img); + CLIP_API bool clip_image_load_from_file(const char * fname, struct clip_image_u8 * img); /** interpret bytes as an image file with length bytes_length, and use the result to populate img */ @@ -89,11 +108,15 @@ CLIP_API bool clip_image_batch_encode(struct clip_ctx * ctx, int n_threads, cons CLIP_API bool clip_model_quantize(const char * fname_inp, const char * fname_out, int itype); CLIP_API int clip_is_minicpmv(const struct clip_ctx * ctx); +CLIP_API bool clip_is_glm(const struct clip_ctx * ctx); CLIP_API bool clip_is_qwen2vl(const struct clip_ctx * ctx); +CLIP_API bool clip_is_llava(const struct clip_ctx * ctx); +CLIP_API bool clip_is_gemma3(const struct clip_ctx * ctx); + +CLIP_API int get_deepest_feature_layer(const struct clip_ctx * ctx); CLIP_API bool clip_encode_float_image (struct clip_ctx * ctx, int n_threads, float * img, int h, int w, float * vec); -CLIP_API bool clip_is_glm(const struct clip_ctx * ctx); #ifdef __cplusplus } diff --git a/examples/llava/convert_image_encoder_to_gguf.py b/examples/llava/convert_image_encoder_to_gguf.py index 4fa1d6cea..2949faec4 100644 --- a/examples/llava/convert_image_encoder_to_gguf.py +++ b/examples/llava/convert_image_encoder_to_gguf.py @@ -6,7 +6,7 @@ import re import torch import numpy as np from gguf import * -from transformers import CLIPModel, CLIPProcessor, CLIPVisionModel +from transformers import CLIPModel, CLIPProcessor, CLIPVisionModel, SiglipVisionModel TEXT = "clip.text" VISION = "clip.vision" @@ -37,6 +37,18 @@ def should_skip_tensor(name: str, has_text: bool, has_vision: bool, has_llava: b def get_tensor_name(name: str) -> str: + # Standardize the transformers llava next keys for + # image newline / mm projector with the classes in haotian-liu LLaVA + if name == "image_newline": + return "model.image_newline" + if name.startswith("multi_modal_projector"): + name = name.replace("multi_modal_projector", "mm") + if "linear_1" in name: + name = name.replace("linear_1", "0") + if "linear_2" in name: + name = name.replace("linear_2", "2") + return name + if "projection" in name: return name if "mm_projector" in name: @@ -77,14 +89,21 @@ def bytes_to_unicode(): ap = argparse.ArgumentParser() ap.add_argument("-m", "--model-dir", help="Path to model directory cloned from HF Hub", required=True) ap.add_argument("--use-f32", action="store_true", default=False, help="Use f32 instead of f16") +ap.add_argument('--bigendian', action="store_true", default=False, help="Model is executed on big-endian machine") ap.add_argument("--text-only", action="store_true", required=False, help="Save a text-only model. It can't be used to encode images") ap.add_argument("--vision-only", action="store_true", required=False, help="Save a vision-only model. It can't be used to encode texts") ap.add_argument("--clip-model-is-vision", action="store_true", required=False, help="The clip model is a pure vision model (ShareGPT4V vision extract for example)") -ap.add_argument("--clip-model-is-openclip", action="store_true", required=False, + +# Selectable visual encoders that are compatible with this script +encoder_group = ap.add_mutually_exclusive_group() +encoder_group.add_argument("--clip-model-is-openclip", action="store_true", required=False, help="The clip model is from openclip (for ViT-SO400M type))") +encoder_group.add_argument("--clip-model-is-siglip", action="store_true", required=False, + help="the visual encoder is Siglip.") + ap.add_argument("--llava-projector", help="Path to llava.projector file. If specified, save an image encoder for LLaVA models.") ap.add_argument("--projector-type", help="Type of projector. Possible values: mlp, ldp, ldpv2", choices=["mlp", "ldp", "ldpv2"], default="mlp") ap.add_argument("-o", "--output-dir", help="Directory to save GGUF files. Default is the original model directory", default=None) @@ -109,7 +128,12 @@ if args.use_f32: # output in the same directory as the model if output_dir is None dir_model = args.model_dir -if args.clip_model_is_vision or not os.path.exists(dir_model + "/vocab.json") or args.clip_model_is_openclip: +if ( + args.clip_model_is_vision or + not os.path.exists(dir_model + "/vocab.json") or + args.clip_model_is_openclip or + args.clip_model_is_siglip +): vocab = None tokens = None else: @@ -137,7 +161,10 @@ ftype = 1 if args.use_f32: ftype = 0 -if args.clip_model_is_vision or args.clip_model_is_openclip: +if args.clip_model_is_siglip: + model = SiglipVisionModel.from_pretrained(dir_model) + processor = None +elif args.clip_model_is_vision or args.clip_model_is_openclip: model = CLIPVisionModel.from_pretrained(dir_model) processor = None else: @@ -165,7 +192,7 @@ output_dir = args.output_dir if args.output_dir is not None else dir_model os.makedirs(output_dir, exist_ok=True) output_prefix = os.path.basename(output_dir).replace("ggml_", "") fname_out = os.path.join(output_dir, f"{fname_middle}model-{ftype_str[ftype]}.gguf") -fout = GGUFWriter(path=fname_out, arch="clip") +fout = GGUFWriter(path=fname_out, arch="clip", endianess=GGUFEndian.LITTLE if not args.bigendian else GGUFEndian.BIG) fout.add_bool("clip.has_text_encoder", has_text_encoder) fout.add_bool("clip.has_vision_encoder", has_vision_encoder) @@ -187,26 +214,71 @@ else: if has_text_encoder: assert t_hparams is not None assert tokens is not None + if args.clip_model_is_siglip: + text_projection_dim = 0 + else: + text_projection_dim = t_hparams.get("projection_dim", config["projection_dim"]) # text_model hparams fout.add_uint32(k(KEY_CONTEXT_LENGTH, TEXT), t_hparams["max_position_embeddings"]) fout.add_uint32(k(KEY_EMBEDDING_LENGTH, TEXT), t_hparams["hidden_size"]) fout.add_uint32(k(KEY_FEED_FORWARD_LENGTH, TEXT), t_hparams["intermediate_size"]) - fout.add_uint32("clip.text.projection_dim", t_hparams.get("projection_dim", config["projection_dim"])) + fout.add_uint32("clip.text.projection_dim", text_projection_dim) fout.add_uint32(k(KEY_ATTENTION_HEAD_COUNT, TEXT), t_hparams["num_attention_heads"]) fout.add_float32(k(KEY_ATTENTION_LAYERNORM_EPS, TEXT), t_hparams["layer_norm_eps"]) fout.add_uint32(k(KEY_BLOCK_COUNT, TEXT), t_hparams["num_hidden_layers"]) fout.add_token_list(tokens) + + +def get_non_negative_vision_feature_layers(v_hparams): + """ + Determine the vision feature layer(s) for the llava model, which are indices into the + hidden states of the visual encoder. Note that the hidden states array generally takes the + form: + + [, , ... ] + + so feature indices should be offset as n+1 to get the output of encoder block n. + We convert all vision feature layers to non-negative so that -1 can be used in + the model as an unset value. If no vision feature layer is found, we leave it unset. + """ + num_hidden_layers = v_hparams["num_hidden_layers"] + to_non_negative = lambda layer_idx: layer_idx if layer_idx >= 0 else num_hidden_layers + layer_idx + 1 + feature_layers_key = None + # Key used for llava models in transformers + if "vision_feature_layer" in config: + feature_layers_key = "vision_feature_layer" + # Key used for llava models in the original format + elif "mm_vision_select_layer" in config: + feature_layers_key = "mm_vision_select_layer" + if feature_layers_key is not None: + feature_layers = config[feature_layers_key] + if isinstance(feature_layers, int): + feature_layers = [feature_layers] + return [to_non_negative(feature_layer) for feature_layer in feature_layers] + +# Determine if we have explicitly specified vision feature layers in our config +feature_layers = get_non_negative_vision_feature_layers(v_hparams) + if has_vision_encoder: - # vision_model hparams + # Siglip does not have a visual projector; set projection dim to 0 + if args.clip_model_is_siglip: + visual_projection_dim = 0 + else: + visual_projection_dim = v_hparams.get("projection_dim", config["projection_dim"]) + + # set vision_model hparams fout.add_uint32("clip.vision.image_size", v_hparams["image_size"]) fout.add_uint32("clip.vision.patch_size", v_hparams["patch_size"]) fout.add_uint32(k(KEY_EMBEDDING_LENGTH, VISION), v_hparams["hidden_size"]) fout.add_uint32(k(KEY_FEED_FORWARD_LENGTH, VISION), v_hparams["intermediate_size"]) - fout.add_uint32("clip.vision.projection_dim", v_hparams.get("projection_dim", config["projection_dim"])) + fout.add_uint32("clip.vision.projection_dim", visual_projection_dim) fout.add_uint32(k(KEY_ATTENTION_HEAD_COUNT, VISION), v_hparams["num_attention_heads"]) fout.add_float32(k(KEY_ATTENTION_LAYERNORM_EPS, VISION), v_hparams["layer_norm_eps"]) - block_count = v_hparams["num_hidden_layers"] - 1 if has_llava_projector else v_hparams["num_hidden_layers"] + if feature_layers: + block_count = max(feature_layers) + else: + block_count = v_hparams["num_hidden_layers"] - 1 if has_llava_projector else v_hparams["num_hidden_layers"] fout.add_uint32(k(KEY_BLOCK_COUNT, VISION), block_count) # /** # "image_grid_pinpoints": [ @@ -258,7 +330,8 @@ if has_vision_encoder: fout.add_string("clip.vision.mm_patch_merge_type", v_hparams["mm_patch_merge_type"]) if "mm_projector_type" in v_hparams: fout.add_string("clip.vision.mm_projector_type", v_hparams["mm_projector_type"]) - + if feature_layers: + fout.add_array("clip.vision.feature_layer", feature_layers) if processor is not None: image_mean = processor.image_processor.image_mean if args.image_mean is None or args.image_mean == default_image_mean else args.image_mean # pyright: ignore[reportAttributeAccessIssue] @@ -274,7 +347,13 @@ fout.add_bool("clip.use_gelu", use_gelu) if has_llava_projector: - model.vision_model.encoder.layers.pop(-1) + # By default, we drop the last layer for llava projector + # models unless we have explicitly set vision feature layers + if feature_layers is None: + model.vision_model.encoder.layers.pop(-1) + else: + model.vision_model.encoder.layers = model.vision_model.encoder.layers[:max(feature_layers)] + projector = torch.load(args.llava_projector) for name, data in projector.items(): name = get_tensor_name(name) diff --git a/examples/llava/gemma3-cli.cpp b/examples/llava/gemma3-cli.cpp new file mode 100644 index 000000000..91a07e2a8 --- /dev/null +++ b/examples/llava/gemma3-cli.cpp @@ -0,0 +1,322 @@ +#include "arg.h" +#include "log.h" +#include "common.h" +#include "sampling.h" +#include "llama.h" +#include "ggml.h" +#include "console.h" +#include "chat.h" +#include "mtmd.h" + +#include +#include +#include + +#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__)) +#include +#include +#elif defined (_WIN32) +#define WIN32_LEAN_AND_MEAN +#ifndef NOMINMAX +#define NOMINMAX +#endif +#include +#include +#endif + +static bool g_is_generating = false; + +/** + * Please note that this is NOT a production-ready stuff. + * It is a playground for trying Gemma 3 vision capabilities. + * For contributors: please keep this code simple and easy to understand. + */ + +static void show_additional_info(int /*argc*/, char ** argv) { + LOG( + "Experimental CLI for using Gemma 3 vision model\n\n" + "Usage: %s [options] -m --mmproj --image -p \n\n" + " -m and --mmproj are required\n" + " --image and -p are optional, if NOT provided, the CLI will run in chat mode\n", + argv[0] + ); +} + +#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__)) || defined (_WIN32) +static void sigint_handler(int signo) { + if (signo == SIGINT) { + if (g_is_generating) { + g_is_generating = false; + } else { + console::cleanup(); + LOG("\nInterrupted by user\n"); + _exit(130); + } + } +} +#endif + +struct gemma3_context { + mtmd_context_ptr ctx_vision; + common_init_result llama_init; + + llama_model * model; + llama_context * lctx; + const llama_vocab * vocab; + llama_batch batch; + int n_batch; + + // note: we know that gemma3 template is "linear", meaning each turn is completely separated to another + // so here we don't need to keep track of chat history + common_chat_templates_ptr tmpls; + + int n_threads = 1; + llama_pos n_past = 0; + + gemma3_context(common_params & params) : llama_init(common_init_from_params(params)) { + model = llama_init.model.get(); + lctx = llama_init.context.get(); + vocab = llama_model_get_vocab(model); + n_threads = params.cpuparams.n_threads; + batch = llama_batch_init(params.n_batch, 0, 1); + n_batch = params.n_batch; + tmpls = common_chat_templates_init(model, params.chat_template); + init_vision_context(params); + } + + void init_vision_context(common_params & params) { + const char * clip_path = params.mmproj.path.c_str(); + ctx_vision.reset(mtmd_init_from_file(clip_path, model, mtmd_context_params{ + /* use_gpu */ true, + /* timings */ true, + /* n_threads */ params.cpuparams.n_threads, + /* verbosity */ GGML_LOG_LEVEL_INFO, + })); + if (!ctx_vision.get()) { + LOG_ERR("Failed to load vision model from %s\n", clip_path); + exit(1); + } + } +}; + +struct decode_embd_batch { + std::vector pos; + std::vector n_seq_id; + std::vector seq_id_0; + std::vector seq_ids; + std::vector logits; + llama_batch batch; + decode_embd_batch(float * embd, int32_t n_tokens, llama_pos pos_0, llama_seq_id seq_id) { + pos .resize(n_tokens); + n_seq_id.resize(n_tokens); + seq_ids .resize(n_tokens + 1); + logits .resize(n_tokens); + seq_id_0.resize(1); + seq_id_0[0] = seq_id; + seq_ids [n_tokens] = nullptr; + batch = { + /*n_tokens =*/ n_tokens, + /*tokens =*/ nullptr, + /*embd =*/ embd, + /*pos =*/ pos.data(), + /*n_seq_id =*/ n_seq_id.data(), + /*seq_id =*/ seq_ids.data(), + /*logits =*/ logits.data(), + }; + for (int i = 0; i < n_tokens; i++) { + batch.pos [i] = pos_0 + i; + batch.n_seq_id[i] = 1; + batch.seq_id [i] = seq_id_0.data(); + batch.logits [i] = false; + } + } +}; + +static int generate_response(gemma3_context & ctx, common_sampler * smpl, int n_predict) { + for (int i = 0; i < n_predict; i++) { + if (i > n_predict || !g_is_generating) { + printf("\n"); + break; + } + + llama_token token_id = common_sampler_sample(smpl, ctx.lctx, -1); + common_sampler_accept(smpl, token_id, true); + + if (llama_vocab_is_eog(ctx.vocab, token_id)) { + printf("\n"); + break; // end of generation + } + + printf("%s", common_token_to_piece(ctx.lctx, token_id).c_str()); + fflush(stdout); + + // eval the token + common_batch_clear(ctx.batch); + common_batch_add(ctx.batch, token_id, ctx.n_past++, {0}, true); + if (llama_decode(ctx.lctx, ctx.batch)) { + LOG_ERR("failed to decode token\n"); + return 1; + } + } + return 0; +} + +static int eval_message(gemma3_context & ctx, common_chat_msg & msg, std::vector & images_fname, bool add_bos = false) { + std::vector bitmaps; + + common_chat_templates_inputs tmpl_inputs; + tmpl_inputs.messages = {msg}; + tmpl_inputs.add_generation_prompt = true; + tmpl_inputs.use_jinja = false; // jinja is buggy here + auto formatted_chat = common_chat_templates_apply(ctx.tmpls.get(), tmpl_inputs); + LOG_DBG("formatted_chat.prompt: %s\n", formatted_chat.prompt.c_str()); + + for (auto & fname : images_fname) { + mtmd_bitmap bitmap; + if (mtmd_helper_bitmap_init_from_file(fname.c_str(), bitmap)) { + LOG_ERR("Unable to load image %s\n", fname.c_str()); + return 2; // image not found + } + bitmaps.push_back(std::move(bitmap)); + } + + mtmd_input_text text; + text.text = formatted_chat.prompt; + text.add_special = add_bos; + text.parse_special = true; + mtmd_input_chunks_ptr chunks(mtmd_tokenize(ctx.ctx_vision.get(), text, bitmaps)); + if (chunks == nullptr) { + LOG_ERR("Unable to tokenize prompt\n"); + return 1; + } + + if (mtmd_helper_eval(ctx.ctx_vision.get(), ctx.lctx, chunks.get(), ctx.n_past, 0, ctx.n_batch)) { + LOG_ERR("Unable to eval prompt\n"); + return 1; + } + + ctx.n_past += mtmd_helper_get_n_tokens(chunks.get()); + + return 0; +} + +int main(int argc, char ** argv) { + ggml_time_init(); + + common_params params; + params.sampling.temp = 0.2; // lower temp by default for better quality + + if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_LLAVA, show_additional_info)) { + return 1; + } + + common_init(); + + if (params.mmproj.path.empty()) { + show_additional_info(argc, argv); + return 1; + } + + gemma3_context ctx(params); + printf("%s: %s\n", __func__, params.model.path.c_str()); + + bool is_single_turn = !params.prompt.empty() && !params.image.empty(); + + struct common_sampler * smpl = common_sampler_init(ctx.model, params.sampling); + int n_predict = params.n_predict < 0 ? INT_MAX : params.n_predict; + + // ctrl+C handling + { +#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__)) + struct sigaction sigint_action; + sigint_action.sa_handler = sigint_handler; + sigemptyset (&sigint_action.sa_mask); + sigint_action.sa_flags = 0; + sigaction(SIGINT, &sigint_action, NULL); +#elif defined (_WIN32) + auto console_ctrl_handler = +[](DWORD ctrl_type) -> BOOL { + return (ctrl_type == CTRL_C_EVENT) ? (sigint_handler(SIGINT), true) : false; + }; + SetConsoleCtrlHandler(reinterpret_cast(console_ctrl_handler), true); +#endif + } + + if (is_single_turn) { + g_is_generating = true; + if (params.prompt.find("<__image__>") == std::string::npos) { + params.prompt += " <__image__>"; + } + common_chat_msg msg; + msg.role = "user"; + msg.content = params.prompt; + if (eval_message(ctx, msg, params.image, true)) { + return 1; + } + if (generate_response(ctx, smpl, n_predict)) { + return 1; + } + + } else { + LOG("\n Running in chat mode, available commands:"); + LOG("\n /image load an image"); + LOG("\n /clear clear the chat history"); + LOG("\n /quit or /exit exit the program"); + LOG("\n"); + + bool is_first_msg = true; + std::vector images_fname; + std::string content; + + while (true) { + g_is_generating = false; + LOG("\n> "); + console::set_display(console::user_input); + std::string line; + console::readline(line, false); + console::set_display(console::reset); + line = string_strip(line); + if (line.empty()) { + continue; + } + if (line == "/quit" || line == "/exit") { + break; + } + if (line == "/clear") { + ctx.n_past = 0; + llama_kv_self_seq_rm(ctx.lctx, 0, 1, -1); // keep BOS + LOG("Chat history cleared\n\n"); + continue; + } + g_is_generating = true; + if (line.find("/image") == 0) { + std::string image = line.substr(7); + images_fname.push_back(string_strip(image)); + content += "<__image__>"; + continue; + } else { + content += line; + } + common_chat_msg msg; + msg.role = "user"; + msg.content = content; + int ret = eval_message(ctx, msg, images_fname, is_first_msg); + if (ret == 2) { + // non-fatal error + images_fname.clear(); + content.clear(); + continue; + } + if (ret) { + return 1; + } + if (generate_response(ctx, smpl, n_predict)) { + return 1; + } + images_fname.clear(); + content.clear(); + is_first_msg = false; + } + } + + return 0; +} diff --git a/examples/llava/gemma3_convert_encoder_to_gguf.py b/examples/llava/gemma3_convert_encoder_to_gguf.py new file mode 100644 index 000000000..241b526b9 --- /dev/null +++ b/examples/llava/gemma3_convert_encoder_to_gguf.py @@ -0,0 +1,307 @@ +import gguf +import argparse +import logging +import sys +import torch +import json +import os +import numpy as np +from typing import cast, ContextManager, Any, Iterator +from pathlib import Path +from torch import Tensor + +logger = logging.getLogger("gemma3-mmproj") + + +# (copied from convert_hf_to_gguf.py) +# tree of lazy tensors +class LazyTorchTensor(gguf.LazyBase): + _tensor_type = torch.Tensor + # to keep the type-checker happy + dtype: torch.dtype + shape: torch.Size + + # only used when converting a torch.Tensor to a np.ndarray + _dtype_map: dict[torch.dtype, type] = { + torch.float16: np.float16, + torch.float32: np.float32, + } + + # used for safetensors slices + # ref: https://github.com/huggingface/safetensors/blob/079781fd0dc455ba0fe851e2b4507c33d0c0d407/bindings/python/src/lib.rs#L1046 + # TODO: uncomment U64, U32, and U16, ref: https://github.com/pytorch/pytorch/issues/58734 + _dtype_str_map: dict[str, torch.dtype] = { + "F64": torch.float64, + "F32": torch.float32, + "BF16": torch.bfloat16, + "F16": torch.float16, + # "U64": torch.uint64, + "I64": torch.int64, + # "U32": torch.uint32, + "I32": torch.int32, + # "U16": torch.uint16, + "I16": torch.int16, + "U8": torch.uint8, + "I8": torch.int8, + "BOOL": torch.bool, + "F8_E4M3": torch.float8_e4m3fn, + "F8_E5M2": torch.float8_e5m2, + } + + def numpy(self) -> gguf.LazyNumpyTensor: + dtype = self._dtype_map[self.dtype] + return gguf.LazyNumpyTensor( + meta=gguf.LazyNumpyTensor.meta_with_dtype_and_shape(dtype, self.shape), + args=(self,), + func=(lambda s: s.numpy()) + ) + + @classmethod + def meta_with_dtype_and_shape(cls, dtype: torch.dtype, shape: tuple[int, ...]) -> Tensor: + return torch.empty(size=shape, dtype=dtype, device="meta") + + @classmethod + def from_safetensors_slice(cls, st_slice: Any) -> Tensor: + dtype = cls._dtype_str_map[st_slice.get_dtype()] + shape: tuple[int, ...] = tuple(st_slice.get_shape()) + lazy = cls(meta=cls.meta_with_dtype_and_shape(dtype, shape), args=(st_slice,), func=lambda s: s[:]) + return cast(torch.Tensor, lazy) + + @classmethod + def __torch_function__(cls, func, types, args=(), kwargs=None): + del types # unused + + if kwargs is None: + kwargs = {} + + if func is torch.Tensor.numpy: + return args[0].numpy() + + return cls._wrap_fn(func)(*args, **kwargs) + + +class Gemma3VisionTower: + hparams: dict + gguf_writer: gguf.GGUFWriter + fname_out: Path + ftype: gguf.LlamaFileType + + @staticmethod + def load_hparams(dir_model: Path): + with open(dir_model / "config.json", "r", encoding="utf-8") as f: + return json.load(f) + + @staticmethod + def get_model_part_names(dir_model: Path, prefix: str, suffix: str) -> list[str]: + part_names: list[str] = [] + for filename in os.listdir(dir_model): + if filename.startswith(prefix) and filename.endswith(suffix): + part_names.append(filename) + part_names.sort() + return part_names + + def __init__(self, + dir_model: Path, + fname_out: Path, + ftype: gguf.LlamaFileType, + is_big_endian: bool,): + hparams = Gemma3VisionTower.load_hparams(dir_model) + self.hparams = hparams + self.fname_out = fname_out + self.ftype = ftype + endianess = gguf.GGUFEndian.BIG if is_big_endian else gguf.GGUFEndian.LITTLE + self.gguf_writer = gguf.GGUFWriter(path=None, arch="clip", endianess=endianess) + + text_config = hparams["text_config"] + vision_config = hparams["vision_config"] + + assert hparams["architectures"][0] == "Gemma3ForConditionalGeneration" + assert text_config is not None + assert vision_config is not None + + self.gguf_writer.add_string ("clip.projector_type", "gemma3") + self.gguf_writer.add_bool ("clip.has_text_encoder", False) + self.gguf_writer.add_bool ("clip.has_vision_encoder", True) + self.gguf_writer.add_bool ("clip.has_llava_projector", False) # legacy + self.gguf_writer.add_uint32 ("clip.vision.image_size", vision_config["image_size"]) + self.gguf_writer.add_uint32 ("clip.vision.patch_size", vision_config["patch_size"]) + self.gguf_writer.add_uint32 ("clip.vision.embedding_length", vision_config["hidden_size"]) + self.gguf_writer.add_uint32 ("clip.vision.feed_forward_length", vision_config["intermediate_size"]) + self.gguf_writer.add_uint32 ("clip.vision.projection_dim", text_config["hidden_size"]) + self.gguf_writer.add_uint32 ("clip.vision.block_count", vision_config["num_hidden_layers"]) + self.gguf_writer.add_uint32 ("clip.vision.attention.head_count", vision_config["num_attention_heads"]) + self.gguf_writer.add_float32("clip.vision.attention.layer_norm_epsilon", vision_config.get("layer_norm_eps", 1e-6)) + # default values taken from HF tranformers code + self.gguf_writer.add_array ("clip.vision.image_mean", [0.5, 0.5, 0.5]) + self.gguf_writer.add_array ("clip.vision.image_std", [0.5, 0.5, 0.5]) + self.gguf_writer.add_bool ("clip.use_gelu", True) + + # load tensors + for name, data_torch in self.get_tensors(dir_model): + # convert any unsupported data types to float32 + if data_torch.dtype not in (torch.float16, torch.float32): + data_torch = data_torch.to(torch.float32) + self.add_tensor(name, data_torch) + + def get_tensors(self, dir_model: Path) -> Iterator[tuple[str, Tensor]]: + part_names = Gemma3VisionTower.get_model_part_names(dir_model, "model", ".safetensors") + tensor_names_from_parts: set[str] = set() + for part_name in part_names: + logger.info(f"gguf: loading model part '{part_name}'") + from safetensors import safe_open + ctx = cast(ContextManager[Any], safe_open(dir_model / part_name, framework="pt", device="cpu")) + with ctx as model_part: + tensor_names_from_parts.update(model_part.keys()) + + for name in model_part.keys(): + data = model_part.get_slice(name) + data = LazyTorchTensor.from_safetensors_slice(data) + yield name, data + + def add_tensor(self, name: str, data_torch: Tensor): + is_1d = len(data_torch.shape) == 1 + is_embd = ".embeddings." in name + old_dtype = data_torch.dtype + can_quantize = not is_1d and not is_embd + data_qtype = gguf.GGMLQuantizationType.F32 + + # this is to support old checkpoint + # TODO: remove this when we have the final model + name = name.replace("vision_model.vision_model.", "vision_tower.vision_model.") + name = name.replace("multimodal_projector.", "multi_modal_projector.") + + # filter only vision tensors + if not name.startswith("vision_tower.vision_model.") and not name.startswith("multi_modal_projector."): + return + # prefix + name = name.replace("vision_tower.vision_model.encoder.layers.", "v.blk.") + name = name.replace("vision_tower.vision_model.", "v.") + # projector and input embd + name = name.replace(".embeddings.patch_embedding.", ".patch_embd.") + name = name.replace(".embeddings.position_embedding.", ".position_embd.") + name = name.replace( + "multi_modal_projector.mm_input_projection_weight", + "mm.input_projection.weight" + ) + name = name.replace( + "multi_modal_projector.mm_soft_emb_norm.weight", + "mm.soft_emb_norm.weight" + ) + name = name.replace("post_layernorm.", "post_ln.") + # each block + name = name.replace(".self_attn.k_proj.", ".attn_k.") + name = name.replace(".self_attn.v_proj.", ".attn_v.") + name = name.replace(".self_attn.q_proj.", ".attn_q.") + name = name.replace(".self_attn.out_proj.", ".attn_out.") + name = name.replace(".layer_norm1.", ".ln1.") + name = name.replace(".layer_norm2.", ".ln2.") + name = name.replace(".mlp.fc1.", ".ffn_down.") + name = name.replace(".mlp.fc2.", ".ffn_up.") + + if can_quantize: + if self.ftype == gguf.LlamaFileType.ALL_F32: + data_qtype = gguf.GGMLQuantizationType.F32 + elif self.ftype == gguf.LlamaFileType.MOSTLY_F16: + data_qtype = gguf.GGMLQuantizationType.F16 + elif self.ftype == gguf.LlamaFileType.MOSTLY_BF16: + data_qtype = gguf.GGMLQuantizationType.BF16 + elif self.ftype == gguf.LlamaFileType.MOSTLY_Q8_0: + data_qtype = gguf.GGMLQuantizationType.Q8_0 + else: + raise ValueError(f"Unsupported file type: {self.ftype}") + + # corrent norm value ; only this "soft_emb_norm" need to be corrected as it's part of Gemma projector + # the other norm values are part of SigLIP model, and they are already correct + # ref code: Gemma3RMSNorm + if "soft_emb_norm.weight" in name: + logger.info(f"Correcting norm value for '{name}'") + data_torch = data_torch + 1 + + data = data_torch.numpy() + + try: + data = gguf.quants.quantize(data, data_qtype) + except Exception as e: + logger.error(f"Error quantizing tensor '{name}': {e}, fallback to F16") + data_qtype = gguf.GGMLQuantizationType.F16 + data = gguf.quants.quantize(data, data_qtype) + + # reverse shape to make it similar to the internal ggml dimension order + shape_str = f"{{{', '.join(str(n) for n in reversed(data_torch.shape))}}}" + logger.info(f"{f'%-32s' % f'{name},'} {old_dtype} --> {data_qtype.name}, shape = {shape_str}") + + self.gguf_writer.add_tensor(name, data, raw_dtype=data_qtype) + + def write(self): + self.gguf_writer.write_header_to_file(path=self.fname_out) + self.gguf_writer.write_kv_data_to_file() + self.gguf_writer.write_tensors_to_file(progress=True) + self.gguf_writer.close() + +def parse_args() -> argparse.Namespace: + parser = argparse.ArgumentParser( + description="Convert Gemma 3 vision tower safetensors to GGUF format",) + parser.add_argument( + "--outfile", type=Path, default="mmproj.gguf", + help="path to write to", + ) + parser.add_argument( + "--outtype", type=str, choices=["f32", "f16", "bf16", "q8_0"], default="f16", + help="output format", + ) + parser.add_argument( + "--bigendian", action="store_true", + help="model is executed on big endian machine", + ) + parser.add_argument( + "model", type=Path, + help="directory containing model file", + nargs="?", + ) + parser.add_argument( + "--verbose", action="store_true", + help="increase output verbosity", + ) + + args = parser.parse_args() + if args.model is None: + parser.error("the following arguments are required: model") + return args + + +def main() -> None: + args = parse_args() + + if args.verbose: + logging.basicConfig(level=logging.DEBUG) + else: + logging.basicConfig(level=logging.INFO) + + dir_model = args.model + + if not dir_model.is_dir(): + logger.error(f'Error: {args.model} is not a directory') + sys.exit(1) + + ftype_map: dict[str, gguf.LlamaFileType] = { + "f32": gguf.LlamaFileType.ALL_F32, + "f16": gguf.LlamaFileType.MOSTLY_F16, + "bf16": gguf.LlamaFileType.MOSTLY_BF16, + "q8_0": gguf.LlamaFileType.MOSTLY_Q8_0, + } + + logger.info(f"Loading model: {dir_model.name}") + + with torch.inference_mode(): + gemma3_vision_tower = Gemma3VisionTower( + dir_model=dir_model, + fname_out=args.outfile, + ftype=ftype_map[args.outtype], + is_big_endian=args.bigendian, + ) + gemma3_vision_tower.write() + + +if __name__ == '__main__': + main() + diff --git a/examples/llava/llava-cli.cpp b/examples/llava/llava-cli.cpp index 40aa0876f..0fe0e333a 100644 --- a/examples/llava/llava-cli.cpp +++ b/examples/llava/llava-cli.cpp @@ -225,7 +225,7 @@ static struct llama_model * llava_init(common_params * params) { llama_model_params model_params = common_model_params_to_llama(*params); - llama_model * model = llama_model_load_from_file(params->model.c_str(), model_params); + llama_model * model = llama_model_load_from_file(params->model.path.c_str(), model_params); if (model == NULL) { LOG_ERR("%s: unable to load model\n" , __func__); return NULL; @@ -234,14 +234,14 @@ static struct llama_model * llava_init(common_params * params) { } static struct llava_context * llava_init_context(common_params * params, llama_model * model) { - const char * clip_path = params->mmproj.c_str(); + const char * clip_path = params->mmproj.path.c_str(); auto prompt = params->prompt; if (prompt.empty()) { prompt = "describe the image in detail."; } - auto ctx_clip = clip_model_load(clip_path, /*verbosity=*/ 1); + auto ctx_clip = clip_model_load(clip_path, GGML_LOG_LEVEL_INFO); llama_context_params ctx_params = common_context_params_to_llama(*params); ctx_params.n_ctx = params->n_ctx < 2048 ? 2048 : params->n_ctx; // we need a longer context size to process image embeddings @@ -283,7 +283,7 @@ int main(int argc, char ** argv) { common_init(); - if (params.mmproj.empty() || (params.image.empty() && !prompt_contains_image(params.prompt))) { + if (params.mmproj.path.empty() || (params.image.empty() && !prompt_contains_image(params.prompt))) { print_usage(argc, argv); return 1; } diff --git a/examples/llava/llava.cpp b/examples/llava/llava.cpp index 300714045..03a22cbb4 100644 --- a/examples/llava/llava.cpp +++ b/examples/llava/llava.cpp @@ -10,6 +10,7 @@ #include #include #include +#include #if defined(LLAVA_LOG_OFF) # define LOG_INF(...) @@ -45,6 +46,17 @@ struct clip_image_grid_shape { int second; }; +// convenience cpp wrapper +struct clip_image_f32_batch_deleter { + void operator()(clip_image_f32_batch * val) { clip_image_f32_batch_free(val); } +}; +typedef std::unique_ptr clip_image_f32_batch_ptr; + +struct clip_image_size_deleter { + void operator()(clip_image_f32_batch * val) { clip_image_f32_batch_free(val); } +}; +typedef std::unique_ptr clip_image_size_ptr; + /** * Selects the best resolution from a list of possible resolutions based on the original size. * @@ -105,8 +117,8 @@ static bool clip_llava_handle_patches(clip_ctx * ctx_clip, std::vector struct ggml_context * ctx; } model; - const int32_t image_size = clip_image_size(ctx_clip); - const int32_t patch_size = clip_patch_size(ctx_clip); + const int32_t image_size = clip_get_image_size(ctx_clip); + const int32_t patch_size = clip_get_patch_size(ctx_clip); int32_t num_patches_per_side = image_size / patch_size; // 336 / 14 = 24 - used for embedding-patching boxes (24*24 = 576 patches) @@ -246,12 +258,9 @@ static clip_image_f32 * reshape_by_patch(clip_image_f32 * image, int patch_size) static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const clip_image_u8 * img, float * image_embd, int * n_img_pos) { // std::vector img_res_v; // format VectN x H x W x RGB (N x 336 x 336 x 3), so interleaved RGB - different to the python implementation which is N x 3 x 336 x 336 - clip_image_f32_batch img_res_v; - img_res_v.size = 0; - img_res_v.data = nullptr; - if (!clip_image_preprocess(ctx_clip, img, &img_res_v)) { + clip_image_f32_batch_ptr img_res_v(clip_image_f32_batch_init()); + if (!clip_image_preprocess(ctx_clip, img, img_res_v.get())) { LOG_ERR("%s: unable to preprocess image\n", __func__); - delete[] img_res_v.data; return false; } @@ -259,66 +268,72 @@ static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const cli const char * mm_patch_merge_type = clip_patch_merge_type(ctx_clip); + const size_t n_imgs = clip_image_f32_batch_n_images(img_res_v.get()); + if (clip_is_minicpmv(ctx_clip) || clip_is_qwen2vl(ctx_clip)) { std::vector image_embd_v; - image_embd_v.resize(img_res_v.size); - struct clip_image_size * load_image_size = clip_image_size_init(); + image_embd_v.resize(n_imgs); + clip_image_size load_image_size; - for (size_t i = 0; i < img_res_v.size; i++) { + for (size_t i = 0; i < n_imgs; i++) { const int64_t t_img_enc_step_start_us = ggml_time_us(); - image_embd_v[i] = (float *)malloc(clip_embd_nbytes_by_img(ctx_clip, img_res_v.data[i].nx, img_res_v.data[i].ny)); - int patch_size=14; - load_image_size->width = img_res_v.data[i].nx; - load_image_size->height = img_res_v.data[i].ny; - clip_add_load_image_size(ctx_clip, load_image_size); + int nx = clip_image_f32_batch_nx(img_res_v.get(), i); + int ny = clip_image_f32_batch_ny(img_res_v.get(), i); + image_embd_v[i] = (float *)malloc(clip_embd_nbytes_by_img(ctx_clip, nx, ny)); + int patch_size = 14; + load_image_size.width = nx; + load_image_size.height = ny; + clip_add_load_image_size(ctx_clip, &load_image_size); bool encoded = false; + clip_image_f32 * img_res = clip_image_f32_get_img(img_res_v.get(), i); if (clip_is_qwen2vl(ctx_clip)) { - encoded = clip_image_encode(ctx_clip, n_threads, &img_res_v.data[i], image_embd_v[i]); + encoded = clip_image_encode(ctx_clip, n_threads, img_res, image_embd_v[i]); } else { - encoded = clip_image_encode(ctx_clip, n_threads, reshape_by_patch(&img_res_v.data[i], patch_size), image_embd_v[i]); + encoded = clip_image_encode(ctx_clip, n_threads, reshape_by_patch(img_res, patch_size), image_embd_v[i]); } if (!encoded) { - LOG_ERR("Unable to encode image - spatial_unpad - subimage %d of %d\n", (int) i+1, (int) img_res_v.size); + LOG_ERR("Unable to encode image - spatial_unpad - subimage %d of %d\n", (int) i+1, (int) n_imgs); return false; } const int64_t t_img_enc_steop_batch_us = ggml_time_us(); - LOG_INF("%s: step %d of %d encoded in %8.2f ms\n", __func__, (int)i+1, (int)img_res_v.size, (t_img_enc_steop_batch_us - t_img_enc_step_start_us) / 1000.0); + LOG_INF("%s: step %d of %d encoded in %8.2f ms\n", __func__, (int)i+1, (int)n_imgs, (t_img_enc_steop_batch_us - t_img_enc_step_start_us) / 1000.0); } const int64_t t_img_enc_batch_us = ggml_time_us(); - LOG_INF("%s: all %d segments encoded in %8.2f ms\n", __func__, (int)img_res_v.size, (t_img_enc_batch_us - t_img_enc_start_us) / 1000.0); + LOG_INF("%s: all %d segments encoded in %8.2f ms\n", __func__, (int)n_imgs, (t_img_enc_batch_us - t_img_enc_start_us) / 1000.0); int n_img_pos_out = 0; for (size_t i = 0; i < image_embd_v.size(); i++) { + int nx = clip_image_f32_batch_nx(img_res_v.get(), i); + int ny = clip_image_f32_batch_ny(img_res_v.get(), i); + clip_image_f32 * img_res = clip_image_f32_get_img(img_res_v.get(), i); std::memcpy( image_embd + n_img_pos_out * clip_n_mmproj_embd(ctx_clip), image_embd_v[i], - clip_embd_nbytes_by_img(ctx_clip, img_res_v.data[i].nx, img_res_v.data[i].ny)); - n_img_pos_out += clip_n_patches_by_img(ctx_clip, &img_res_v.data[i]); + clip_embd_nbytes_by_img(ctx_clip, nx, ny)); + n_img_pos_out += clip_n_patches_by_img(ctx_clip, img_res); } *n_img_pos = n_img_pos_out; for (size_t i = 0; i < image_embd_v.size(); i++) { free(image_embd_v[i]); } image_embd_v.clear(); - load_image_size->width = img->nx; - load_image_size->height = img->ny; - clip_add_load_image_size(ctx_clip, load_image_size); - LOG_INF("%s: load_image_size %d %d\n", __func__, load_image_size->width, load_image_size->height); - delete[] img_res_v.data; - img_res_v.size = 0; - img_res_v.data = nullptr; + load_image_size.width = img->nx; + load_image_size.height = img->ny; + clip_add_load_image_size(ctx_clip, &load_image_size); + LOG_INF("%s: load_image_size %d %d\n", __func__, load_image_size.width, load_image_size.height); } else if (clip_is_glm(ctx_clip)){ struct clip_image_size * load_image_size = clip_image_size_init(); - load_image_size->width = img_res_v.data[0].nx; - load_image_size->height = img_res_v.data[0].ny; + load_image_size->width = clip_image_f32_batch_nx(img_res_v.get(), 0); + load_image_size->height = clip_image_f32_batch_ny(img_res_v.get(), 0); clip_add_load_image_size(ctx_clip, load_image_size); - bool encoded = clip_image_encode(ctx_clip, n_threads, &img_res_v.data[0], image_embd); - int pos = int(load_image_size->width/clip_patch_size(ctx_clip)/2); + clip_image_f32 * img_res = clip_image_f32_get_img(img_res_v.get(), 0); + bool encoded = clip_image_encode(ctx_clip, n_threads, img_res, image_embd); + int pos = int(load_image_size->width/clip_get_patch_size(ctx_clip)/2); *n_img_pos = (pos * pos + 2); if (!encoded){ LOG_ERR("Unable to encode image \n"); @@ -328,8 +343,8 @@ static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const cli else if (strcmp(mm_patch_merge_type, "spatial_unpad") != 0) { // flat / default llava-1.5 type embedding *n_img_pos = clip_n_patches(ctx_clip); - bool encoded = clip_image_encode(ctx_clip, n_threads, &img_res_v.data[0], image_embd); // image_embd shape is 576 x 4096 - delete[] img_res_v.data; + clip_image_f32 * img_res = clip_image_f32_get_img(img_res_v.get(), 0); + bool encoded = clip_image_encode(ctx_clip, n_threads, img_res, image_embd); // image_embd shape is 576 x 4096 if (!encoded) { LOG_ERR("Unable to encode image\n"); @@ -340,31 +355,28 @@ static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const cli // spatial_unpad llava-1.6 type embedding // TODO: CLIP needs batching support - in HF the llm projection is separate after encoding, which might be a solution to quickly get batching working std::vector image_embd_v; - image_embd_v.resize(img_res_v.size); - for (size_t i = 0; i < img_res_v.size; i++) { + image_embd_v.resize(n_imgs); + for (size_t i = 0; i < n_imgs; i++) { + clip_image_f32 * img_res = clip_image_f32_get_img(img_res_v.get(), i); image_embd_v[i] = (float *)malloc(clip_embd_nbytes(ctx_clip)); // 576 patches * 4096 embeddings * 4 bytes = 9437184 - const bool encoded = clip_image_encode(ctx_clip, n_threads, &img_res_v.data[i], image_embd_v[i]); // image data is in 3x336x336 format and will be converted to 336x336x3 inside + const bool encoded = clip_image_encode(ctx_clip, n_threads, img_res, image_embd_v[i]); // image data is in 3x336x336 format and will be converted to 336x336x3 inside if (!encoded) { - LOG_ERR("Unable to encode image - spatial_unpad - subimage %d of %d\n", (int) i+1, (int) img_res_v.size); + LOG_ERR("Unable to encode image - spatial_unpad - subimage %d of %d\n", (int) i+1, (int) n_imgs); return false; } } const int64_t t_img_enc_batch_us = ggml_time_us(); - LOG_INF("%s: %d segments encoded in %8.2f ms\n", __func__, (int)img_res_v.size, (t_img_enc_batch_us - t_img_enc_start_us) / 1000.0); + LOG_INF("%s: %d segments encoded in %8.2f ms\n", __func__, (int)n_imgs, (t_img_enc_batch_us - t_img_enc_start_us) / 1000.0); const int32_t * image_grid = clip_image_grid(ctx_clip); + const size_t num_gridpoints = get_clip_image_grid_size(ctx_clip); std::vector> grid_pinpoints; - for (int i = 0; i < 32 && image_grid[i] != 0; i += 2) { + for (size_t i = 0; i < num_gridpoints; i += 2) { grid_pinpoints.push_back({image_grid[i], image_grid[i+1]}); } - // free all img_res_v - not needed anymore - delete[] img_res_v.data; - img_res_v.size = 0; - img_res_v.data = nullptr; - - const int32_t image_size = clip_image_size(ctx_clip); + const int32_t image_size = clip_get_image_size(ctx_clip); struct clip_image_grid_shape grid_shape = get_anyres_image_grid_shape({img->nx,img->ny}, grid_pinpoints, image_size); @@ -405,7 +417,8 @@ bool llava_validate_embed_size(const llama_context * ctx_llama, const clip_ctx * } bool llava_image_embed_make_with_clip_img(clip_ctx * ctx_clip, int n_threads, const clip_image_u8 * img, float ** image_embd_out, int * n_img_pos_out) { - int num_max_patches = 6; + // Granite vision uses up to 10 patches + base patch + int num_max_patches = 11; if (clip_is_minicpmv(ctx_clip)) { num_max_patches = 10; } diff --git a/examples/llava/llava_surgery_v2.py b/examples/llava/llava_surgery_v2.py index 2d5b32fe6..b07c3e323 100644 --- a/examples/llava/llava_surgery_v2.py +++ b/examples/llava/llava_surgery_v2.py @@ -33,6 +33,33 @@ def save_model(model, file_path, file_type): else: torch.save(model, file_path) +# Helpers to match weight names from specific components or +# determine if a saved shard contains that component +def is_vision_tower(weight_name): + return ( + weight_name.startswith("model.vision_tower") or + weight_name.startswith("vit.") or + weight_name.startswith("vision_tower") + ) + +def is_newline(weight_name): + return ( + weight_name.startswith("model.image_newline") or + weight_name.startswith("image_newline") + ) + +def is_mm_projector(weight_name): + return ( + weight_name.startswith("model.mm_projector") or + weight_name.startswith("vision_proj.") or + weight_name.startswith("multi_modal_projector") + ) + +def newline_criteria(checkpoint): + return any(is_newline(k) for k in checkpoint.keys()) + +def proj_criteria(checkpoint): + return any(is_mm_projector(k) for k in checkpoint.keys()) # Adapted function to clean vision tower from checkpoint def clean_vision_tower_from_checkpoint(checkpoint_path): @@ -40,7 +67,7 @@ def clean_vision_tower_from_checkpoint(checkpoint_path): # file_type = 'pytorch' model_path = os.path.dirname(checkpoint_path) print(f"Searching for vision tower tensors in {checkpoint_path}") - clip_tensors = [k for k, v in checkpoint.items() if (k.startswith("model.vision_tower") or k.startswith("vit."))] + clip_tensors = [k for k, v in checkpoint.items() if is_vision_tower(k)] if len(clip_tensors) > 0: print(f"Found {len(clip_tensors)} tensors to extract from {checkpoint_path}") @@ -84,12 +111,6 @@ def find_relevant_checkpoints(checkpoint_paths, newline_criteria, projector): return newline_checkpoint_path, projector_checkpoint_path -def newline_criteria(checkpoint): - return any(k.startswith("model.image_newline") for k in checkpoint.keys()) - -def proj_criteria(checkpoint): - return any(k.startswith("model.mm_projector") or k.startswith("vision_proj.") for k in checkpoint.keys()) - # Command-line interface setup ap = argparse.ArgumentParser() @@ -123,14 +144,14 @@ first_checkpoint = None if newline_checkpoint_path is not None: print(f"Taking newline from {newline_checkpoint_path}") first_checkpoint, file_type = load_model(newline_checkpoint_path) - first_mm_tensors = [k for k, v in first_checkpoint.items() if k.startswith("model.image_newline")] + first_mm_tensors = [k for k, v in first_checkpoint.items() if is_newline(k)] # Load the checkpoint mm_tensors = [] last_checkpoint = None if projector_checkpoint_path is not None: last_checkpoint, file_type = load_model(projector_checkpoint_path) - mm_tensors = [k for k, v in last_checkpoint.items() if k.startswith("model.mm_projector") or k.startswith("vision_proj.")] + mm_tensors = [k for k, v in last_checkpoint.items() if is_mm_projector(k)] if len(mm_tensors) == 0: if last_checkpoint is not None: @@ -155,5 +176,5 @@ if len(projector) > 0: save_model(projector, f"{args.model}/llava.projector", 'pytorch') print("Done!") -print(f"Now you can convert {args.model} to a a regular LLaMA GGUF file.") +print(f"Now you can convert {args.model} to a regular LLaMA GGUF file.") print(f"Also, use {args.model}/llava.projector to prepare a llava-encoder.gguf file.") diff --git a/examples/llava/minicpmv-cli.cpp b/examples/llava/minicpmv-cli.cpp index 53d902d61..5ad970c22 100644 --- a/examples/llava/minicpmv-cli.cpp +++ b/examples/llava/minicpmv-cli.cpp @@ -31,7 +31,7 @@ static struct llama_model * llava_init(common_params * params) { llama_model_params model_params = common_model_params_to_llama(*params); - llama_model * model = llama_model_load_from_file(params->model.c_str(), model_params); + llama_model * model = llama_model_load_from_file(params->model.path.c_str(), model_params); if (model == NULL) { LOG_ERR("%s: unable to load model\n" , __func__); return NULL; @@ -80,13 +80,17 @@ static void llava_free(struct llava_context * ctx_llava) { } static struct clip_ctx * clip_init_context(common_params * params) { - const char * clip_path = params->mmproj.c_str(); + const char * clip_path = params->mmproj.path.c_str(); auto prompt = params->prompt; if (prompt.empty()) { prompt = "describe the image in detail."; } - auto * ctx_clip = clip_model_load(clip_path, /*verbosity=*/ 1); + struct clip_context_params clip_params = { + /* use_gpu */ params->n_gpu_layers != 0, + /* verbosity */ GGML_LOG_LEVEL_INFO, // TODO: make this configurable + }; + auto * ctx_clip = clip_init(clip_path, clip_params); return ctx_clip; } @@ -148,19 +152,34 @@ static void process_image(struct llava_context * ctx_llava, struct llava_image_e process_eval_image_embed(ctx_llava, embeds, params->n_batch, &n_past, idx++); eval_string(ctx_llava->ctx_llama, std::string("").c_str(), params->n_batch, &n_past, false); if (num_image_embeds > 1) { - size_t num_image_embeds_col = clip_uhd_num_image_embeds_col(ctx_llava->ctx_clip); - eval_string(ctx_llava->ctx_llama, std::string("").c_str(), params->n_batch, &n_past, false); - for (size_t i = 0; i < (num_image_embeds-1)/num_image_embeds_col; ++i) { - for (size_t j = 0; j < num_image_embeds_col; ++j) { - eval_string(ctx_llava->ctx_llama, std::string("").c_str(), params->n_batch, &n_past, false); - process_eval_image_embed(ctx_llava, embeds, params->n_batch, &n_past, idx++); - eval_string(ctx_llava->ctx_llama, std::string("").c_str(), params->n_batch, &n_past, false); - if (j == num_image_embeds_col - 1) { - eval_string(ctx_llava->ctx_llama, std::string("\n").c_str(), params->n_batch, &n_past, false); + if (has_minicpmv_projector == 2) { + size_t num_image_embeds_col = clip_uhd_num_image_embeds_col(ctx_llava->ctx_clip); + eval_string(ctx_llava->ctx_llama, std::string("").c_str(), params->n_batch, &n_past, false); + for (size_t i = 0; i < (num_image_embeds-1)/num_image_embeds_col; ++i) { + for (size_t j = 0; j < num_image_embeds_col; ++j) { + eval_string(ctx_llava->ctx_llama, std::string("").c_str(), params->n_batch, &n_past, false); + process_eval_image_embed(ctx_llava, embeds, params->n_batch, &n_past, idx++); + eval_string(ctx_llava->ctx_llama, std::string("").c_str(), params->n_batch, &n_past, false); + if (j == num_image_embeds_col - 1) { + eval_string(ctx_llava->ctx_llama, std::string("\n").c_str(), params->n_batch, &n_past, false); + } + } + } + eval_string(ctx_llava->ctx_llama, std::string("").c_str(), params->n_batch, &n_past, false); + } + else if (has_minicpmv_projector == 3 || has_minicpmv_projector == 4) { + size_t num_image_embeds_col = clip_uhd_num_image_embeds_col(ctx_llava->ctx_clip); + for (size_t i = 0; i < (num_image_embeds-1)/num_image_embeds_col; ++i) { + for (size_t j = 0; j < num_image_embeds_col; ++j) { + eval_string(ctx_llava->ctx_llama, std::string("").c_str(), params->n_batch, &n_past, false); + process_eval_image_embed(ctx_llava, embeds, params->n_batch, &n_past, idx++); + eval_string(ctx_llava->ctx_llama, std::string("").c_str(), params->n_batch, &n_past, false); + if (j == num_image_embeds_col - 1) { + eval_string(ctx_llava->ctx_llama, std::string("\n").c_str(), params->n_batch, &n_past, false); + } } } } - eval_string(ctx_llava->ctx_llama, std::string("").c_str(), params->n_batch, &n_past, false); } LOG_INF("%s: image token past: %d\n", __func__, n_past); } @@ -271,7 +290,7 @@ int main(int argc, char ** argv) { common_init(); - if (params.mmproj.empty() || (params.image.empty())) { + if (params.mmproj.path.empty() || (params.image.empty())) { show_additional_info(argc, argv); return 1; } diff --git a/examples/llava/minicpmv-convert-image-encoder-to-gguf.py b/examples/llava/minicpmv-convert-image-encoder-to-gguf.py index 9b196757f..cfe0961f9 100644 --- a/examples/llava/minicpmv-convert-image-encoder-to-gguf.py +++ b/examples/llava/minicpmv-convert-image-encoder-to-gguf.py @@ -597,7 +597,6 @@ elif args.minicpmv_projector is not None: fname_middle = "mmproj-" has_text_encoder = False has_minicpmv_projector = True - minicpmv_version = 4 elif args.vision_only: fname_middle = "vision-" has_text_encoder = False diff --git a/examples/llava/mtmd.cpp b/examples/llava/mtmd.cpp new file mode 100644 index 000000000..114c274bc --- /dev/null +++ b/examples/llava/mtmd.cpp @@ -0,0 +1,341 @@ +#include "clip.h" +#include "clip-impl.h" +#include "mtmd.h" + +#include "llama.h" + +#include +#include +#include +#include +#include +#include +#include + +struct mtmd_context { + struct clip_ctx * ctx_clip; + const struct llama_model * text_model; + std::vector image_embd_v; // image embedding vector + bool print_timings; + int n_threads; + std::string image_marker; + + // TODO @ngxson : add timings + + mtmd_context(const char * mmproj_fname, + const llama_model * text_model, + const mtmd_context_params & ctx_params) : print_timings(ctx_params.print_timings), n_threads(ctx_params.n_threads), image_marker(ctx_params.image_marker) { + clip_context_params ctx_clip_params; + ctx_clip_params.use_gpu = ctx_params.use_gpu; + ctx_clip_params.verbosity = ctx_params.verbosity; + ctx_clip = clip_init(mmproj_fname, ctx_clip_params); + if (!ctx_clip) { + throw std::runtime_error(string_format("Failed to load CLIP model from %s\n", mmproj_fname)); + } + this->text_model = text_model; + } + + ~mtmd_context() { + clip_free(ctx_clip); + } +}; + +struct mtmd_image_tokens_data { + clip_image_f32_batch batch_f32; // preprocessed image patches +}; + +struct mtmd_image_tokens { + uint32_t nx; // number of tokens in x direction + uint32_t ny; // number of tokens in y direction + uint32_t n_tokens() const { return nx * ny; } + clip_image_f32_batch batch_f32; // preprocessed image patches +}; + +mtmd_context * mtmd_init_from_file(const char * mmproj_fname, + const struct llama_model * text_model, + const struct mtmd_context_params ctx_params) { + try { + return new mtmd_context(mmproj_fname, text_model, ctx_params); + } catch (const std::exception & e) { + LOG_ERR("%s: error: %s\n", __func__, e.what()); + return nullptr; + } +} + +void mtmd_free(mtmd_context * ctx) { + if (ctx) { + delete ctx; + } +} + +// copied from common_tokenize +static std::vector mtmd_tokenize_text_internal( + const struct llama_vocab * vocab, + const std::string & text, + bool add_special, + bool parse_special) { + // upper limit for the number of tokens + int n_tokens = text.length() + 2 * add_special; + std::vector result(n_tokens); + n_tokens = llama_tokenize(vocab, text.data(), text.length(), result.data(), result.size(), add_special, parse_special); + if (n_tokens < 0) { + result.resize(-n_tokens); + int check = llama_tokenize(vocab, text.data(), text.length(), result.data(), result.size(), add_special, parse_special); + GGML_ASSERT(check == -n_tokens); + } else { + result.resize(n_tokens); + } + return result; +} + +mtmd_input_chunks * mtmd_tokenize(mtmd_context * ctx, + const mtmd_input_text & text, + const std::vector & bitmaps) { + mtmd_input_chunks * output = new mtmd_input_chunks; + auto vocab = llama_model_get_vocab(ctx->text_model); + + std::string prompt_modified(text.text); + std::string marker_modified(ctx->image_marker); + projector_type proj_type = clip_get_projector_type(ctx->ctx_clip); + // a bit hacky here, but works for now + // for some models, we need to add prefix and suffix to the image embeddings + if (proj_type == PROJECTOR_TYPE_GEMMA3) { + // ... (image embeddings) ... + marker_modified = "" + ctx->image_marker + ""; + string_replace_all(prompt_modified, ctx->image_marker, marker_modified); + } + + std::vector parts = string_split_str(text.text, ctx->image_marker); + output->clear(); + output->reserve(parts.size()); + + size_t i_img = 0; + + for (const auto & part : parts) { + //printf("tokenizing part: %s\n", part.c_str()); + bool add_bos = &parts.front() == ∂ + auto tokens = mtmd_tokenize_text_internal(vocab, part, text.add_special && add_bos, text.parse_special); + if (tokens.empty()) { + continue; + } + mtmd_input_chunk chunk{ + MTMD_INPUT_CHUNK_TYPE_TEXT, + std::move(tokens), + {}, + }; + output->emplace_back(std::move(chunk)); + + if (&parts.back() != &part) { + // add image token to middle of 2 parts + + if (i_img >= bitmaps.size()) { + LOG_ERR("%s: error: not enough images for %d parts\n", __func__, (int)parts.size()); + return nullptr; + } + + // shim layer + clip_image_u8_ptr img_u8(clip_image_u8_init()); + img_u8->nx = bitmaps[i_img].nx; + img_u8->ny = bitmaps[i_img].ny; + img_u8->buf.resize(bitmaps[i_img].data.size()); + std::memcpy(img_u8->buf.data(), bitmaps[i_img].data.data(), img_u8->nx * img_u8->ny * 3); + + // preprocess image + clip_image_f32_batch batch_f32; + bool ok = clip_image_preprocess(ctx->ctx_clip, img_u8.get(), &batch_f32); + if (!ok) { + LOG_ERR("Unable to preprocess image\n"); + return nullptr; + } + + mtmd_image_tokens * image_tokens = new mtmd_image_tokens; + image_tokens->nx = clip_n_patches(ctx->ctx_clip); // TODO @ngxson : use clip_n_patches_by_image + image_tokens->ny = 1; // TODO + image_tokens->batch_f32 = std::move(batch_f32); + + mtmd_input_chunk chunk{ + MTMD_INPUT_CHUNK_TYPE_IMAGE, + {}, + image_tokens, + }; + output->emplace_back(std::move(chunk)); + i_img++; + } + } + + return output; +} + +void mtmd_input_chunks_free(mtmd_input_chunks * chunks) { + for (auto & chunk : *chunks) { + if (chunk.type == MTMD_INPUT_CHUNK_TYPE_IMAGE && chunk.tokens_image) { + delete chunk.tokens_image; + } + } + delete chunks; +} + +int32_t mtmd_encode(mtmd_context * ctx, const mtmd_image_tokens * image_tokens) { + int n_mmproj_embd = clip_n_mmproj_embd(ctx->ctx_clip); + ctx->image_embd_v.resize(image_tokens->n_tokens() * n_mmproj_embd); + bool ok = clip_image_batch_encode( + ctx->ctx_clip, + ctx->n_threads, + &image_tokens->batch_f32, + ctx->image_embd_v.data()); + return ok ? 0 : 1; +} + +float * mtmd_get_output_embd(mtmd_context * ctx) { + return ctx->image_embd_v.data(); +} + +size_t mtmd_helper_get_n_tokens(mtmd_input_chunks * chunks) { + size_t n_tokens = 0; + for (auto & chunk : *chunks) { + if (chunk.type == MTMD_INPUT_CHUNK_TYPE_TEXT) { + n_tokens += chunk.tokens_text.size(); + } else if (chunk.type == MTMD_INPUT_CHUNK_TYPE_IMAGE) { + n_tokens += chunk.tokens_image->n_tokens(); + } else { + GGML_ASSERT(false && "chunk type not supported"); + } + } + return n_tokens; +} + +// helper struct to make working with embd batch easier +// note: this will be removed after llama_batch_ext refactoring +struct decode_embd_batch { + std::vector pos; + std::vector n_seq_id; + std::vector seq_id_0; + std::vector seq_ids; + std::vector logits; + llama_batch batch; + decode_embd_batch(float * embd, int32_t n_tokens, llama_pos pos_0, llama_seq_id seq_id) { + pos .resize(n_tokens); + n_seq_id.resize(n_tokens); + seq_ids .resize(n_tokens + 1); + logits .resize(n_tokens); + seq_id_0.resize(1); + seq_id_0[0] = seq_id; + seq_ids [n_tokens] = nullptr; + batch = { + /*n_tokens =*/ n_tokens, + /*tokens =*/ nullptr, + /*embd =*/ embd, + /*pos =*/ pos.data(), + /*n_seq_id =*/ n_seq_id.data(), + /*seq_id =*/ seq_ids.data(), + /*logits =*/ logits.data(), + }; + for (int i = 0; i < n_tokens; i++) { + batch.pos [i] = pos_0 + i; + batch.n_seq_id[i] = 1; + batch.seq_id [i] = seq_id_0.data(); + batch.logits [i] = false; + } + } +}; + +int32_t mtmd_helper_eval(mtmd_context * ctx, + llama_context * lctx, + mtmd_input_chunks * chunks, + llama_pos pos0, + llama_seq_id seq_id, + int32_t n_batch) { + int32_t ret; + llama_pos n_past = pos0; + llama_batch text_batch = llama_batch_init(n_batch, 0, 1); + + for (auto & chunk : *chunks) { + bool is_last = &chunk == &chunks->back(); + if (chunk.type == MTMD_INPUT_CHUNK_TYPE_TEXT) { + // TODO @ngxson : may need to split into smaller batches + text_batch.n_tokens = chunk.tokens_text.size(); + for (size_t i = 0; i < chunk.tokens_text.size(); i++) { + text_batch.token [i] = chunk.tokens_text[i]; + text_batch.pos [i] = n_past++; + text_batch.n_seq_id[i] = 1; + text_batch.seq_id [i][0] = seq_id; + text_batch.logits [i] = false; + } + if (is_last) { + // always get logits for last input chunk + text_batch.logits[text_batch.n_tokens - 1] = true; + } + ret = llama_decode(lctx, text_batch); + if (ret != 0) { + LOG_ERR("failed to decode text\n"); + llama_batch_free(text_batch); + return ret; + } + + } else if (chunk.type == MTMD_INPUT_CHUNK_TYPE_IMAGE) { + GGML_ASSERT(!is_last && "logits for last image chunk is not yet support"); + GGML_ASSERT(chunk.tokens_image != nullptr); + int64_t t0 = ggml_time_ms(); + if (ctx->print_timings) { + LOG_INF("encoding image...\n"); + } + ret = mtmd_encode(ctx, chunk.tokens_image); + if (ret != 0) { + LOG_ERR("failed to encode image\n"); + llama_batch_free(text_batch); + return ret; + } + if (ctx->print_timings) { + LOG_INF("image encoded in %" PRId64 " ms\n", ggml_time_ms() - t0); + } + + int32_t n_tokens = chunk.tokens_image->n_tokens(); + float * embd = mtmd_get_output_embd(ctx); + decode_embd_batch batch_img(embd, n_tokens, n_past, 0); + int64_t t1 = ggml_time_ms(); + ret = llama_decode(lctx, batch_img.batch); + if (ret != 0) { + LOG_ERR("failed to decode image\n"); + llama_batch_free(text_batch); + return ret; + } + if (ctx->print_timings) { + LOG_INF("image decoded in %" PRId64 " ms\n", ggml_time_ms() - t1); + } + + n_past += n_tokens; + + } else { + GGML_ASSERT(false && "chunk type not supported"); + } + } + + llama_batch_free(text_batch); + return 0; +} + +int32_t mtmd_helper_bitmap_init_from_buf(const unsigned char * buf, size_t len, mtmd_bitmap & output) { + clip_image_u8_ptr img_u8(clip_image_u8_init()); + bool ok = clip_image_load_from_bytes(buf, len, img_u8.get()); + if (!ok) { + LOG_ERR("Unable to load image from buffer\n"); + return 1; + } + unsigned char * data = clip_image_u8_get_data(img_u8.get(), &output.nx, &output.ny); + output.data.resize(output.nx * output.ny * 3); + std::memcpy(output.data.data(), data, output.nx * output.ny * 3); + return 0; +} + +int32_t mtmd_helper_bitmap_init_from_file(const char * fname, mtmd_bitmap & output) { + clip_image_u8_ptr img_u8(clip_image_u8_init()); + bool ok = clip_image_load_from_file(fname, img_u8.get()); + if (!ok) { + LOG_ERR("Unable to load image %s\n", fname); + return 1; + } + unsigned char * data = clip_image_u8_get_data(img_u8.get(), &output.nx, &output.ny); + output.data.resize(output.nx * output.ny * 3); + std::memcpy(output.data.data(), data, output.nx * output.ny * 3); + return 0; +} diff --git a/examples/llava/mtmd.h b/examples/llava/mtmd.h new file mode 100644 index 000000000..598f6947b --- /dev/null +++ b/examples/llava/mtmd.h @@ -0,0 +1,146 @@ +#ifndef MTMD_H +#define MTMD_H + +#include "ggml.h" +#include "llama.h" +#include "clip.h" + +#include +#include +#include + +#ifdef LLAMA_SHARED +# if defined(_WIN32) && !defined(__MINGW32__) +# ifdef LLAMA_BUILD +# define MTMD_API __declspec(dllexport) +# else +# define MTMD_API __declspec(dllimport) +# endif +# else +# define MTMD_API __attribute__ ((visibility ("default"))) +# endif +#else +# define MTMD_API +#endif + +#ifdef __cplusplus + +enum mtmd_input_chunk_type { + MTMD_INPUT_CHUNK_TYPE_TEXT, + MTMD_INPUT_CHUNK_TYPE_IMAGE, +}; + +struct mtmd_context; +struct mtmd_image_tokens; + +// represents raw image data, layout is RGBRGBRGB... +// length of data must be nx * ny * 3 +struct mtmd_bitmap { + uint32_t nx; + uint32_t ny; + std::vector data; +}; + +struct mtmd_input_chunk { + mtmd_input_chunk_type type; + std::vector tokens_text; + mtmd_image_tokens * tokens_image = nullptr; +}; + +using mtmd_input_chunks = std::vector; + +struct mtmd_context_params { + bool use_gpu = true; + bool print_timings = true; + int n_threads = 4; + enum ggml_log_level verbosity = GGML_LOG_LEVEL_INFO; + const char * image_marker = "<__image__>"; +}; + +struct mtmd_input_text { + std::string text; + bool add_special; + bool parse_special; +}; + +// initialize the mtmd context +// return nullptr on failure +MTMD_API mtmd_context * mtmd_init_from_file(const char * mmproj_fname, + const llama_model * text_model, + const mtmd_context_params ctx_params); + +MTMD_API void mtmd_free(mtmd_context * ctx); + +// tokenize an input text prompt and an image +// the prompt must have the input image marker (default: "<__image__>") in it +// the marker will be replaced with the image tokens +// for example: +// "here is an image: <__image__>\ndescribe it in detail." +// this will gives 3 chunks: +// 1. "here is an image: " +// 2. (image tokens) +// 3. "\ndescribe it in detail." +// number of bitmaps must be equal to the number of image markers in the prompt +// this function is thread-safe (shared ctx) +MTMD_API mtmd_input_chunks * mtmd_tokenize(mtmd_context * ctx, + const mtmd_input_text & text, + const std::vector & bitmaps); + +// free image chunk data +MTMD_API void mtmd_input_chunks_free(mtmd_input_chunks * chunks); + +// returns 0 on success +MTMD_API int32_t mtmd_encode(mtmd_context * ctx, + const mtmd_image_tokens * image_tokens); + +// get output embeddings from the last encode pass +MTMD_API float * mtmd_get_output_embd(mtmd_context * ctx); + +// +// helper functions (can be implemented based on other functions) +// + +// helper to count the total number of tokens from a list of chunks, useful to keep track of n_past +MTMD_API size_t mtmd_helper_get_n_tokens(mtmd_input_chunks * chunks); + +// helper function that automatically: +// 1. run llama_decode() on text chunks +// 2. run mtmd_encode() on image chunks, then mtmd_get_output_embd() and then llama_decode() +// if any of the mtmd_encode() or llama_decode() calls return non-zero, stop and forward the error +// otherwise, returns 0 on success +MTMD_API int32_t mtmd_helper_eval(mtmd_context * ctx, + llama_context * lctx, + mtmd_input_chunks * chunks, + llama_pos pos0, + llama_seq_id seq_id, + int32_t n_batch); + +// helper function to construct a mtmd_bitmap from a file +// returns 0 on success +// this function is thread-safe +MTMD_API int32_t mtmd_helper_bitmap_init_from_file(const char * fname, mtmd_bitmap & output); + +// helper function to construct a mtmd_bitmap from a buffer +// the buffer must be an image in format supported by stb_image (jpg, png, bmp, gif, etc.) +// returns 0 on success +// this function is thread-safe +MTMD_API int32_t mtmd_helper_bitmap_init_from_buf(const unsigned char * buf, size_t len, mtmd_bitmap & output); + +// convenient unique_ptr wrappers +struct mtmd_context_deleter { + void operator()(mtmd_context * val) { mtmd_free(val); } +}; +using mtmd_context_ptr = std::unique_ptr; + +struct mtmd_input_chunks_deleter { + void operator()(mtmd_input_chunks * val) { mtmd_input_chunks_free(val); } +}; +using mtmd_input_chunks_ptr = std::unique_ptr; + +#else + +static_assert(false && "C header is not yet supported by this library"); + +#endif + +#endif diff --git a/examples/llava/qwen2vl-cli.cpp b/examples/llava/qwen2vl-cli.cpp index 132a7da54..eca7b7f10 100644 --- a/examples/llava/qwen2vl-cli.cpp +++ b/examples/llava/qwen2vl-cli.cpp @@ -314,7 +314,7 @@ static struct llama_model * llava_init(common_params * params) { llama_model_params model_params = common_model_params_to_llama(*params); - llama_model * model = llama_model_load_from_file(params->model.c_str(), model_params); + llama_model * model = llama_model_load_from_file(params->model.path.c_str(), model_params); if (model == NULL) { LOG_ERR("%s: unable to load model\n" , __func__); return NULL; @@ -323,14 +323,14 @@ static struct llama_model * llava_init(common_params * params) { } static struct llava_context * llava_init_context(common_params * params, llama_model * model) { - const char * clip_path = params->mmproj.c_str(); + const char * clip_path = params->mmproj.path.c_str(); auto prompt = params->prompt; if (prompt.empty()) { prompt = "describe the image in detail."; } - auto ctx_clip = clip_model_load(clip_path, /*verbosity=*/ 1); + auto ctx_clip = clip_model_load(clip_path, GGML_LOG_LEVEL_INFO); llama_context_params ctx_params = common_context_params_to_llama(*params); ctx_params.n_ctx = params->n_ctx < 2048 ? 2048 : params->n_ctx; // we need a longer context size to process image embeddings @@ -524,7 +524,7 @@ int main(int argc, char ** argv) { common_init(); - if (params.mmproj.empty() || (params.image.empty() && !prompt_contains_image(params.prompt))) { + if (params.mmproj.path.empty() || (params.image.empty() && !prompt_contains_image(params.prompt))) { print_usage(argc, argv); return 1; } diff --git a/examples/llava/test-1.jpeg b/examples/llava/test-1.jpeg new file mode 100644 index 000000000..7fdcaaf04 Binary files /dev/null and b/examples/llava/test-1.jpeg differ diff --git a/examples/llava/tests.sh b/examples/llava/tests.sh new file mode 100755 index 000000000..cc9bda876 --- /dev/null +++ b/examples/llava/tests.sh @@ -0,0 +1,81 @@ +#!/bin/bash + +# make sure we are in the right directory +SCRIPT_DIR=$( cd -- "$( dirname -- "${BASH_SOURCE[0]}" )" &> /dev/null && pwd ) +cd $SCRIPT_DIR + +#export LLAMA_CACHE="$SCRIPT_DIR/tmp" + +set -eux + +mkdir -p $SCRIPT_DIR/output + +PROJ_ROOT="$SCRIPT_DIR/../.." +cd $PROJ_ROOT + +############### + +arr_bin=() +arr_hf=() + +add_test() { + local bin=$1 + local hf=$2 + arr_bin+=("$bin") + arr_hf+=("$hf") +} + +add_test "llama-gemma3-cli" "ggml-org/gemma-3-4b-it-GGUF:Q4_K_M" +add_test "llama-llava-cli" "cmp-nct/Yi-VL-6B-GGUF:Q5_K" +add_test "llama-llava-cli" "guinmoon/MobileVLM-3B-GGUF:Q4_K_M" +add_test "llama-llava-cli" "THUDM/glm-edge-v-5b-gguf:Q4_K_M" +add_test "llama-llava-cli" "second-state/Llava-v1.5-7B-GGUF:Q2_K" +add_test "llama-llava-cli" "cjpais/llava-1.6-mistral-7b-gguf:Q3_K" +add_test "llama-llava-cli" "ibm-research/granite-vision-3.2-2b-GGUF:Q4_K_M" +add_test "llama-minicpmv-cli" "second-state/MiniCPM-Llama3-V-2_5-GGUF:Q2_K" # model from openbmb is corrupted +add_test "llama-minicpmv-cli" "openbmb/MiniCPM-V-2_6-gguf:Q2_K" +add_test "llama-minicpmv-cli" "openbmb/MiniCPM-o-2_6-gguf:Q4_0" +add_test "llama-qwen2vl-cli" "bartowski/Qwen2-VL-2B-Instruct-GGUF:Q4_K_M" + +############### + +cmake --build build -j --target "${arr_bin[@]}" + +arr_res=() + +for i in "${!arr_bin[@]}"; do + bin="${arr_bin[$i]}" + hf="${arr_hf[$i]}" + + echo "Running test with binary: $bin and HF model: $hf" + echo "" + echo "" + + output=$("$PROJ_ROOT/build/bin/$bin" -hf "$hf" --image $SCRIPT_DIR/test-1.jpeg -p "what is the publisher name of the newspaper?" --temp 0 2>&1 | tee /dev/tty) + + echo "$output" > $SCRIPT_DIR/output/$bin-$(echo "$hf" | tr '/' '-').log + + if echo "$output" | grep -iq "new york"; then + result="\033[32mOK\033[0m: $bin $hf" + else + result="\033[31mFAIL\033[0m: $bin $hf" + fi + echo -e "$result" + arr_res+=("$result") + + echo "" + echo "" + echo "" + echo "#################################################" + echo "#################################################" + echo "" + echo "" +done + +set +x + +for i in "${!arr_res[@]}"; do + echo -e "${arr_res[$i]}" +done +echo "" +echo "Output logs are saved in $SCRIPT_DIR/output" diff --git a/examples/lookahead/README.md b/examples/lookahead/README.md index a69a471b4..aab3cd0ca 100644 --- a/examples/lookahead/README.md +++ b/examples/lookahead/README.md @@ -4,4 +4,4 @@ Demonstration of lookahead decoding technique: https://lmsys.org/blog/2023-11-21-lookahead-decoding/ -More info: https://github.com/ggerganov/llama.cpp/pull/4207 +More info: https://github.com/ggml-org/llama.cpp/pull/4207 diff --git a/examples/lookahead/lookahead.cpp b/examples/lookahead/lookahead.cpp index 2f0898e62..7df20aee1 100644 --- a/examples/lookahead/lookahead.cpp +++ b/examples/lookahead/lookahead.cpp @@ -7,6 +7,7 @@ #include #include #include +#include struct ngram_data { bool active = false; @@ -95,7 +96,7 @@ int main(int argc, char ** argv) { llama_decode(ctx, llama_batch_get_one(&inp.back(), 1)); for (int s = 1; s < W + G + 1; ++s) { - llama_kv_cache_seq_cp(ctx, 0, s, -1, -1); + llama_kv_self_seq_cp(ctx, 0, s, -1, -1); } const auto t_enc_end = ggml_time_us(); @@ -437,17 +438,17 @@ int main(int argc, char ** argv) { // KV cache management // if no verification token matched, we simply remove all cells from this batch -> no fragmentation - llama_kv_cache_seq_rm(ctx, -1, n_past, -1); + llama_kv_self_seq_rm(ctx, -1, n_past, -1); if (seq_id_best != 0) { // if a verification token matched, we keep the best sequence and remove the rest // this leads to some KV cache fragmentation - llama_kv_cache_seq_keep(ctx, seq_id_best); - llama_kv_cache_seq_cp (ctx, seq_id_best, 0, -1, -1); - llama_kv_cache_seq_rm (ctx, seq_id_best, -1, -1); + llama_kv_self_seq_keep(ctx, seq_id_best); + llama_kv_self_seq_cp (ctx, seq_id_best, 0, -1, -1); + llama_kv_self_seq_rm (ctx, seq_id_best, -1, -1); for (int s = 1; s < W + G + 1; ++s) { - llama_kv_cache_seq_cp(ctx, 0, s, -1, -1); + llama_kv_self_seq_cp(ctx, 0, s, -1, -1); } } } diff --git a/examples/lookup/README.md b/examples/lookup/README.md index 71c345c03..07d73849b 100644 --- a/examples/lookup/README.md +++ b/examples/lookup/README.md @@ -8,5 +8,5 @@ The key parameters for lookup decoding are `ngram_min`, `ngram_max` and `n_draft More info: -https://github.com/ggerganov/llama.cpp/pull/4484 -https://github.com/ggerganov/llama.cpp/issues/4226 +https://github.com/ggml-org/llama.cpp/pull/4484 +https://github.com/ggml-org/llama.cpp/issues/4226 diff --git a/examples/lookup/lookup.cpp b/examples/lookup/lookup.cpp index dbd0444ec..4ae93b2a5 100644 --- a/examples/lookup/lookup.cpp +++ b/examples/lookup/lookup.cpp @@ -192,7 +192,7 @@ int main(int argc, char ** argv){ // KV cache management // clean the cache of draft tokens that weren't accepted - llama_kv_cache_seq_rm(ctx, 0, n_past, -1); + llama_kv_self_seq_rm(ctx, 0, n_past, -1); common_batch_clear(batch_tgt); common_batch_add(batch_tgt, draft[0], n_past, { 0 }, true); diff --git a/examples/main/README.md b/examples/main/README.md index 46f92eb7a..e4b3590b5 100644 --- a/examples/main/README.md +++ b/examples/main/README.md @@ -1,6 +1,6 @@ # llama.cpp/examples/main -This example program allows you to use various LLaMA language models easily and efficiently. It is specifically designed to work with the [llama.cpp](https://github.com/ggerganov/llama.cpp) project, which provides a plain C/C++ implementation with optional 4-bit quantization support for faster, lower memory inference, and is optimized for desktop CPUs. This program can be used to perform various inference tasks with LLaMA models, including generating text based on user-provided prompts and chat-like interactions with reverse prompts. +This example program allows you to use various LLaMA language models easily and efficiently. It is specifically designed to work with the [llama.cpp](https://github.com/ggml-org/llama.cpp) project, which provides a plain C/C++ implementation with optional 4-bit quantization support for faster, lower memory inference, and is optimized for desktop CPUs. This program can be used to perform various inference tasks with LLaMA models, including generating text based on user-provided prompts and chat-like interactions with reverse prompts. ## Table of Contents @@ -27,29 +27,53 @@ Once downloaded, place your model in the models folder in llama.cpp. ##### Input prompt (One-and-done) ```bash -./llama-cli -m models/gemma-1.1-7b-it.Q4_K_M.gguf --prompt "Once upon a time" +./llama-cli -m models/gemma-1.1-7b-it.Q4_K_M.gguf -no-cnv --prompt "Once upon a time" ``` ##### Conversation mode (Allow for continuous interaction with the model) ```bash -./llama-cli -m models/gemma-1.1-7b-it.Q4_K_M.gguf -cnv --chat-template gemma +./llama-cli -m models/gemma-1.1-7b-it.Q4_K_M.gguf --chat-template gemma +``` + +##### Conversation mode using built-in jinja chat template + +```bash +./llama-cli -m models/gemma-1.1-7b-it.Q4_K_M.gguf --jinja +``` + +##### One-and-done query using jinja with custom system prompt and a starting prompt + +```bash +./llama-cli -m models/gemma-1.1-7b-it.Q4_K_M.gguf --jinja --single-turn -sys "You are a helpful assistant" -p "Hello" ``` ##### Infinite text from a starting prompt (you can use `Ctrl-C` to stop it): ```bash -./llama-cli -m models\gemma-1.1-7b-it.Q4_K_M.gguf --ignore-eos -n -1 +./llama-cli -m models/gemma-1.1-7b-it.Q4_K_M.gguf --ignore-eos -n -1 ``` ### Windows: ##### Input prompt (One-and-done) ```powershell -./llama-cli.exe -m models\gemma-1.1-7b-it.Q4_K_M.gguf --prompt "Once upon a time" +./llama-cli.exe -m models\gemma-1.1-7b-it.Q4_K_M.gguf -no-cnv --prompt "Once upon a time" ``` ##### Conversation mode (Allow for continuous interaction with the model) ```powershell -./llama-cli.exe -m models\gemma-1.1-7b-it.Q4_K_M.gguf -cnv --chat-template gemma +./llama-cli.exe -m models\gemma-1.1-7b-it.Q4_K_M.gguf --chat-template gemma +``` + +##### Conversation mode using built-in jinja chat template + +```powershell +./llama-cli.exe -m models\gemma-1.1-7b-it.Q4_K_M.gguf --jinja +``` + +##### One-and-done query using jinja with custom system prompt and a starting prompt + +```powershell +./llama-cli.exe -m models\gemma-1.1-7b-it.Q4_K_M.gguf --jinja --single-turn -sys "You are a helpful assistant" -p "Hello" ``` #### Infinite text from a starting prompt (you can use `Ctrl-C` to stop it): @@ -77,6 +101,8 @@ The `llama-cli` program provides several ways to interact with the LLaMA models - `--prompt PROMPT`: Provide a prompt directly as a command-line option. - `--file FNAME`: Provide a file containing a prompt or multiple prompts. +- `--system-prompt PROMPT`: Provide a system prompt (will otherwise use the default one in the chat template (if provided)). +- `--system-prompt-file FNAME`: Provide a file containing a system prompt. - `--interactive-first`: Run the program in interactive mode and wait for input right away. (More on this below.) ## Interaction @@ -89,7 +115,10 @@ In interactive mode, users can participate in text generation by injecting their - `-i, --interactive`: Run the program in interactive mode, allowing users to engage in real-time conversations or provide specific instructions to the model. - `--interactive-first`: Run the program in interactive mode and immediately wait for user input before starting the text generation. -- `-cnv, --conversation`: Run the program in conversation mode (does not print special tokens and suffix/prefix, use default chat template) (default: false) +- `-cnv, --conversation`: Run the program in conversation mode (does not print special tokens and suffix/prefix, use default or provided chat template) (default: true if chat template found) +- `-no-cnv`: Disable conversation mode (default: false) +- `-st, --single-turn`: Only process a single conversation turn (user input) and then exit. +- `--jinja`: Enable jinja chat template parser, will use the model's built-in template or a user-provided one (default: false) - `--color`: Enable colorized output to differentiate visually distinguishing between prompts, user input, and generated text. By understanding and utilizing these interaction options, you can create engaging and dynamic experiences with the LLaMA models, tailoring the text generation process to your specific needs. @@ -121,10 +150,12 @@ When --in-prefix or --in-suffix options are enabled the chat template ( --chat-t ### Chat templates - `--chat-template JINJA_TEMPLATE`: This option sets a custom jinja chat template. It accepts a string, not a file name. Default: template taken from model's metadata. Llama.cpp only supports [some pre-defined templates](https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template). These include llama2, llama3, gemma, monarch, chatml, orion, vicuna, vicuna-orca, deepseek, command-r, zephyr. When --in-prefix or --in-suffix options are enabled the chat template ( --chat-template ) is disabled. + `--chat-template JINJA_TEMPLATE`: This option sets a custom jinja chat template. It accepts a string, not a file name. Default: template taken from model's metadata. Llama.cpp only supports [some pre-defined templates](https://github.com/ggml-org/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template). These include llama2, llama3, gemma, monarch, chatml, orion, vicuna, vicuna-orca, deepseek, command-r, zephyr. When --in-prefix or --in-suffix options are enabled the chat template ( --chat-template ) is disabled. Example usage: `--chat-template gemma` +`--chat-template-file FNAME`: Load a custom jinja chat template from an external file, useful if the model contains outdated or incompatible template, some examples can be found in models/templates. Up-to-date chat templates can be downloaded from Hugging Face using scripts/get_chat_template.py + ## Context Management During text generation, LLaMA models have a limited context size, which means they can only consider a certain number of tokens from the input and generated text. When the context fills up, the model resets internally, potentially losing some information from the beginning of the conversation or instructions. Context management options help maintain continuity and coherence in these situations. @@ -265,6 +296,14 @@ Being experimental and unique, XTC is disabled by default. The recommended combi Example usage: `--xtc-probability 0.5 --xtc-threshold 0.1` +### Top-nσ Sampling + +- `--top-nsigma N`: Limit the next token selection to a subset of tokens with pre-softmax logits that are within n * σ less than the max logit (default: -1, -1 = disabled). + +Top-nσ sampling is a text generation method that selects tokens based on a statistical threshold in pre-softmax logits. It works by only sampling from tokens with logits that are within n * σ of the maximum logit. This method helps maintain a stable sampling space regardless of temperature scaling, allowing it to perform well on reasoning tasks even in high temperatures. Without complex probability manipulation, it efficiently filters tokens directly on the pre-softmax logits. A higher value for top-nsigma (e.g., 5) will take more noisy tokens into consideration, while a lower value (e.g., 1) will focous on the more informative region of the sampling space. + +Example usage: `--top-nsigma 1` + ### Logit Bias - `-l TOKEN_ID(+/-)BIAS, --logit-bias TOKEN_ID(+/-)BIAS`: Modify the likelihood of a token appearing in the generated text completion. diff --git a/examples/main/main.cpp b/examples/main/main.cpp index e654d3542..fd7410a64 100644 --- a/examples/main/main.cpp +++ b/examples/main/main.cpp @@ -4,7 +4,7 @@ #include "log.h" #include "sampling.h" #include "llama.h" -#include "chat-template.hpp" +#include "chat.h" #include #include @@ -31,8 +31,6 @@ #pragma warning(disable: 4244 4267) // possible loss of data #endif -static const char * DEFAULT_SYSTEM_MESSAGE = "You are a helpful assistant"; - static llama_context ** g_ctx; static llama_model ** g_model; static common_sampler ** g_smpl; @@ -47,8 +45,8 @@ static void print_usage(int argc, char ** argv) { (void) argc; LOG("\nexample usage:\n"); - LOG("\n text generation: %s -m your_model.gguf -p \"I believe the meaning of life is\" -n 128\n", argv[0]); - LOG("\n chat (conversation): %s -m your_model.gguf -p \"You are a helpful assistant\" -cnv\n", argv[0]); + LOG("\n text generation: %s -m your_model.gguf -p \"I believe the meaning of life is\" -n 128 -no-cnv\n", argv[0]); + LOG("\n chat (conversation): %s -m your_model.gguf -sys \"You are a helpful assistant\"\n", argv[0]); LOG("\n"); } @@ -158,7 +156,7 @@ int main(int argc, char ** argv) { } const llama_vocab * vocab = llama_model_get_vocab(model); - auto chat_templates = common_chat_templates_from_model(model, params.chat_template); + auto chat_templates = common_chat_templates_init(model, params.chat_template); LOG_INF("%s: llama threadpool init, n_threads = %d\n", __func__, (int) params.cpuparams.n_threads); @@ -201,7 +199,7 @@ int main(int argc, char ** argv) { } // auto enable conversation mode if chat template is available - const bool has_chat_template = chat_templates.has_explicit_template && chat_templates.template_default; + const bool has_chat_template = common_chat_templates_was_explicit(chat_templates.get()); if (params.conversation_mode == COMMON_CONVERSATION_MODE_AUTO) { if (has_chat_template) { LOG_INF("%s: chat template is available, enabling conversation mode (disable it with -no-cnv)\n", __func__); @@ -219,7 +217,11 @@ int main(int argc, char ** argv) { // print chat template example in conversation mode if (params.conversation_mode) { if (params.enable_chat_template) { - LOG_INF("%s: chat template example:\n%s\n", __func__, common_chat_format_example(*chat_templates.template_default, params.use_jinja).c_str()); + if (!params.prompt.empty() && params.system_prompt.empty()) { + LOG_WRN("*** User-specified prompt will pre-start conversation, did you mean to set --system-prompt (-sys) instead?\n"); + } + + LOG_INF("%s: chat template example:\n%s\n", __func__, common_chat_format_example(chat_templates.get(), params.use_jinja).c_str()); } else { LOG_INF("%s: in-suffix/prefix is specified, chat template will be disabled\n", __func__); } @@ -263,21 +265,45 @@ int main(int argc, char ** argv) { std::vector embd_inp; + bool waiting_for_first_input = false; auto chat_add_and_format = [&chat_msgs, &chat_templates](const std::string & role, const std::string & content) { - common_chat_msg new_msg{role, content, {}}; - auto formatted = common_chat_format_single(*chat_templates.template_default, chat_msgs, new_msg, role == "user", g_params->use_jinja); - chat_msgs.push_back({role, content, {}}); + common_chat_msg new_msg; + new_msg.role = role; + new_msg.content = content; + auto formatted = common_chat_format_single(chat_templates.get(), chat_msgs, new_msg, role == "user", g_params->use_jinja); + chat_msgs.push_back(new_msg); LOG_DBG("formatted: '%s'\n", formatted.c_str()); return formatted; }; + std::string prompt; { - auto prompt = (params.conversation_mode && params.enable_chat_template) - // format the system prompt in conversation mode (fallback to default if empty) - ? chat_add_and_format("system", params.prompt.empty() ? DEFAULT_SYSTEM_MESSAGE : params.prompt) + if (params.conversation_mode && params.enable_chat_template) { + if (!params.system_prompt.empty()) { + // format the system prompt (will use template default if empty) + chat_add_and_format("system", params.system_prompt); + } + + if (!params.prompt.empty()) { + // format and append the user prompt + chat_add_and_format("user", params.prompt); + } else { + waiting_for_first_input = true; + } + + if (!params.system_prompt.empty() || !params.prompt.empty()) { + common_chat_templates_inputs inputs; + inputs.messages = chat_msgs; + inputs.add_generation_prompt = !params.prompt.empty(); + + prompt = common_chat_templates_apply(chat_templates.get(), inputs).prompt; + } + } else { // otherwise use the prompt as is - : params.prompt; - if (params.interactive_first || !params.prompt.empty() || session_tokens.empty()) { + prompt = params.prompt; + } + + if (params.interactive_first || !prompt.empty() || session_tokens.empty()) { LOG_DBG("tokenize the prompt\n"); embd_inp = common_tokenize(ctx, prompt, true, true); } else { @@ -290,7 +316,7 @@ int main(int argc, char ** argv) { } // Should not run without any tokens - if (embd_inp.empty()) { + if (!waiting_for_first_input && embd_inp.empty()) { if (add_bos) { embd_inp.push_back(llama_vocab_bos(vocab)); LOG_WRN("embd_inp was considered empty and bos was added: %s\n", string_from(ctx, embd_inp).c_str()); @@ -328,7 +354,7 @@ int main(int argc, char ** argv) { } // remove any "future" tokens that we might have inherited from the previous session - llama_kv_cache_seq_rm(ctx, -1, n_matching_session_tokens, -1); + llama_kv_self_seq_rm(ctx, -1, n_matching_session_tokens, -1); } LOG_DBG("recalculate the cached logits (check): embd_inp.size() %zu, n_matching_session_tokens %zu, embd_inp.size() %zu, session_tokens.size() %zu\n", @@ -350,7 +376,12 @@ int main(int argc, char ** argv) { } if (params.conversation_mode) { - params.interactive_first = true; + if (params.single_turn && !params.prompt.empty()) { + params.interactive = false; + params.interactive_first = false; + } else { + params.interactive_first = true; + } } // enable interactive mode if interactive start is specified @@ -474,8 +505,8 @@ int main(int argc, char ** argv) { LOG_INF( " - Press Ctrl+C to interject at any time.\n"); #endif LOG_INF( "%s", control_message); - if (params.conversation_mode && params.enable_chat_template && params.prompt.empty()) { - LOG_INF( " - Using default system message. To change it, set a different value via -p PROMPT or -f FILE argument.\n"); + if (params.conversation_mode && params.enable_chat_template && params.system_prompt.empty()) { + LOG_INF( " - Not using system message. To change it, set a different value via -sys PROMPT\n"); } LOG_INF("\n"); @@ -571,8 +602,8 @@ int main(int argc, char ** argv) { LOG_DBG("context full, swapping: n_past = %d, n_left = %d, n_ctx = %d, n_keep = %d, n_discard = %d\n", n_past, n_left, n_ctx, params.n_keep, n_discard); - llama_kv_cache_seq_rm (ctx, 0, params.n_keep , params.n_keep + n_discard); - llama_kv_cache_seq_add(ctx, 0, params.n_keep + n_discard, n_past, -n_discard); + llama_kv_self_seq_rm (ctx, 0, params.n_keep , params.n_keep + n_discard); + llama_kv_self_seq_add(ctx, 0, params.n_keep + n_discard, n_past, -n_discard); n_past -= n_discard; @@ -595,9 +626,9 @@ int main(int argc, char ** argv) { LOG_DBG("div: [%6d, %6d] / %6d -> [%6d, %6d]\n", ga_i + ib*bd, ga_i + ib*bd + ga_w, ga_n, (ga_i + ib*bd)/ga_n, (ga_i + ib*bd + ga_w)/ga_n); LOG_DBG("shift: [%6d, %6d] + %6d -> [%6d, %6d]\n", ga_i + ib*bd + ga_w, n_past + ib*bd, dd, ga_i + ib*bd + ga_w + dd, n_past + ib*bd + dd); - llama_kv_cache_seq_add(ctx, 0, ga_i, n_past, ib*bd); - llama_kv_cache_seq_div(ctx, 0, ga_i + ib*bd, ga_i + ib*bd + ga_w, ga_n); - llama_kv_cache_seq_add(ctx, 0, ga_i + ib*bd + ga_w, n_past + ib*bd, dd); + llama_kv_self_seq_add(ctx, 0, ga_i, n_past, ib*bd); + llama_kv_self_seq_div(ctx, 0, ga_i + ib*bd, ga_i + ib*bd + ga_w, ga_n); + llama_kv_self_seq_add(ctx, 0, ga_i + ib*bd + ga_w, n_past + ib*bd, dd); n_past -= bd; @@ -755,11 +786,14 @@ int main(int argc, char ** argv) { // check for reverse prompt using special tokens llama_token last_token = common_sampler_last(smpl); - if (std::find(antiprompt_token.begin(), antiprompt_token.end(), last_token) != antiprompt_token.end()) { - if (params.interactive) { - is_interacting = true; + for (auto token : antiprompt_token) { + if (token == last_token) { + if (params.interactive) { + is_interacting = true; + } + is_antiprompt = true; + break; } - is_antiprompt = true; } if (is_antiprompt) { @@ -768,7 +802,7 @@ int main(int argc, char ** argv) { } // deal with end of generation tokens in interactive mode - if (llama_vocab_is_eog(vocab, common_sampler_last(smpl))) { + if (!waiting_for_first_input && llama_vocab_is_eog(vocab, common_sampler_last(smpl))) { LOG_DBG("found an EOG token\n"); if (params.interactive) { @@ -788,12 +822,17 @@ int main(int argc, char ** argv) { } // if current token is not EOG, we add it to current assistant message - if (params.conversation_mode) { + if (params.conversation_mode && !waiting_for_first_input) { const auto id = common_sampler_last(smpl); assistant_ss << common_token_to_piece(ctx, id, false); + + if (!prompt.empty()) { + prompt.clear(); + is_interacting = false; + } } - if (n_past > 0 && is_interacting) { + if ((n_past > 0 || waiting_for_first_input) && is_interacting) { LOG_DBG("waiting for user input\n"); if (params.conversation_mode) { @@ -883,11 +922,17 @@ int main(int argc, char ** argv) { input_echo = false; // do not echo this again } - if (n_past > 0) { + if (n_past > 0 || waiting_for_first_input) { if (is_interacting) { common_sampler_reset(smpl); } is_interacting = false; + + if (waiting_for_first_input && params.single_turn) { + params.interactive = false; + params.interactive_first = false; + } + waiting_for_first_input = false; } } diff --git a/examples/parallel/parallel.cpp b/examples/parallel/parallel.cpp index 7ef43d5e1..80698518e 100644 --- a/examples/parallel/parallel.cpp +++ b/examples/parallel/parallel.cpp @@ -12,6 +12,7 @@ #include #include #include +#include // trim whitespace from the beginning and end of a string static std::string trim(const std::string & str) { @@ -105,6 +106,8 @@ int main(int argc, char ** argv) { common_params params; + params.n_predict = 128; + if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_PARALLEL)) { return 1; } @@ -201,7 +204,7 @@ int main(int argc, char ** argv) { // assign the system KV cache to all parallel sequences for (int32_t i = 1; i <= n_clients; ++i) { - llama_kv_cache_seq_cp(ctx, 0, i, -1, -1); + llama_kv_self_seq_cp(ctx, 0, i, -1, -1); } LOG_INF("\n"); @@ -233,9 +236,9 @@ int main(int argc, char ** argv) { if (batch.n_tokens == 0) { // all sequences have ended - clear the entire KV cache for (int i = 1; i <= n_clients; ++i) { - llama_kv_cache_seq_rm(ctx, i, -1, -1); + llama_kv_self_seq_rm(ctx, i, -1, -1); // but keep the system prompt - llama_kv_cache_seq_cp(ctx, 0, i, -1, -1); + llama_kv_self_seq_cp(ctx, 0, i, -1, -1); } LOG_INF("%s: clearing the KV cache\n", __func__); @@ -371,8 +374,8 @@ int main(int argc, char ** argv) { } // delete only the generated part of the sequence, i.e. keep the system prompt in the cache - llama_kv_cache_seq_rm(ctx, client.id + 1, -1, -1); - llama_kv_cache_seq_cp(ctx, 0, client.id + 1, -1, -1); + llama_kv_self_seq_rm(ctx, client.id + 1, -1, -1); + llama_kv_self_seq_cp(ctx, 0, client.id + 1, -1, -1); const auto t_main_end = ggml_time_us(); @@ -404,7 +407,7 @@ int main(int argc, char ** argv) { params.prompt_file = "used built-in defaults"; } LOG_INF("External prompt file: \033[32m%s\033[0m\n", params.prompt_file.c_str()); - LOG_INF("Model and path used: \033[32m%s\033[0m\n\n", params.model.c_str()); + LOG_INF("Model and path used: \033[32m%s\033[0m\n\n", params.model.path.c_str()); LOG_INF("Total prompt tokens: %6d, speed: %5.2f t/s\n", n_total_prompt, (double) (n_total_prompt ) / (t_main_end - t_main_start) * 1e6); LOG_INF("Total gen tokens: %6d, speed: %5.2f t/s\n", n_total_gen, (double) (n_total_gen ) / (t_main_end - t_main_start) * 1e6); diff --git a/examples/passkey/README.md b/examples/passkey/README.md index 2b8e910f9..2f19597c4 100644 --- a/examples/passkey/README.md +++ b/examples/passkey/README.md @@ -5,8 +5,8 @@ models ability to recall information from long contexts. See the following PRs for more info: -- https://github.com/ggerganov/llama.cpp/pull/3856 -- https://github.com/ggerganov/llama.cpp/pull/4810 +- https://github.com/ggml-org/llama.cpp/pull/3856 +- https://github.com/ggml-org/llama.cpp/pull/4810 ### Usage diff --git a/examples/passkey/passkey.cpp b/examples/passkey/passkey.cpp index 5953928d4..347ea4a69 100644 --- a/examples/passkey/passkey.cpp +++ b/examples/passkey/passkey.cpp @@ -7,6 +7,7 @@ #include #include #include +#include static void print_usage(int, char ** argv) { LOG("\nexample usage:\n"); @@ -63,7 +64,7 @@ int main(int argc, char ** argv) { llama_model_params model_params = common_model_params_to_llama(params); - llama_model * model = llama_model_load_from_file(params.model.c_str(), model_params); + llama_model * model = llama_model_load_from_file(params.model.path.c_str(), model_params); if (model == NULL) { LOG_ERR("%s: unable to load model\n" , __func__); @@ -132,11 +133,11 @@ int main(int argc, char ** argv) { const int ib = i/n_batch - 1; const int bd = n_batch_grp*(n_grp - 1); - llama_kv_cache_seq_add (ctx, 0, n_past - n_batch, n_past, ib*bd); - llama_kv_cache_seq_div (ctx, 0, n_past - n_batch + ib*bd, n_past + ib*bd, n_grp); - llama_kv_cache_update (ctx); + llama_kv_self_seq_add (ctx, 0, n_past - n_batch, n_past, ib*bd); + llama_kv_self_seq_div (ctx, 0, n_past - n_batch + ib*bd, n_past + ib*bd, n_grp); + llama_kv_self_update (ctx); - n_past = llama_kv_cache_seq_pos_max(ctx, 0) + 1; + n_past = llama_kv_self_seq_pos_max(ctx, 0) + 1; } common_batch_clear(batch); @@ -166,12 +167,12 @@ int main(int argc, char ** argv) { LOG_INF("%s: shifting KV cache with %d\n", __func__, n_discard); - llama_kv_cache_seq_rm (ctx, 0, n_keep , n_keep + n_discard); - llama_kv_cache_seq_add(ctx, 0, n_keep + n_discard, n_ctx, -n_discard); - //llama_kv_cache_defrag (ctx); - llama_kv_cache_update (ctx); + llama_kv_self_seq_rm (ctx, 0, n_keep , n_keep + n_discard); + llama_kv_self_seq_add(ctx, 0, n_keep + n_discard, n_ctx, -n_discard); + //llama_kv_self_defrag (ctx); + llama_kv_self_update (ctx); - n_past = llama_kv_cache_seq_pos_max(ctx, 0) + 1; + n_past = llama_kv_self_seq_pos_max(ctx, 0) + 1; common_batch_clear(batch); @@ -197,12 +198,12 @@ int main(int argc, char ** argv) { if (n_discard > 0) { LOG_INF("%s: shifting KV cache with %d to free space for the answer\n", __func__, n_discard); - llama_kv_cache_seq_rm (ctx, 0, n_keep , n_keep + n_discard); - llama_kv_cache_seq_add(ctx, 0, n_keep + n_discard, n_ctx, -n_discard); - //llama_kv_cache_defrag (ctx); - llama_kv_cache_update (ctx); + llama_kv_self_seq_rm (ctx, 0, n_keep , n_keep + n_discard); + llama_kv_self_seq_add(ctx, 0, n_keep + n_discard, n_ctx, -n_discard); + //llama_kv_self_defrag (ctx); + llama_kv_self_update (ctx); - n_past = llama_kv_cache_seq_pos_max(ctx, 0) + 1; + n_past = llama_kv_self_seq_pos_max(ctx, 0) + 1; } } diff --git a/examples/perplexity/perplexity.cpp b/examples/perplexity/perplexity.cpp index 9bf6c5743..175f2804b 100644 --- a/examples/perplexity/perplexity.cpp +++ b/examples/perplexity/perplexity.cpp @@ -3,6 +3,7 @@ #include "log.h" #include "llama.h" +#include #include #include #include @@ -360,7 +361,7 @@ static results_perplexity perplexity_v2(llama_context * ctx, const common_params const auto t_start = std::chrono::high_resolution_clock::now(); // clear the KV cache - llama_kv_cache_clear(ctx); + llama_kv_self_clear(ctx); llama_batch batch = llama_batch_init(n_batch, 0, 1); @@ -546,7 +547,7 @@ static results_perplexity perplexity(llama_context * ctx, const common_params & const auto t_start = std::chrono::high_resolution_clock::now(); // clear the KV cache - llama_kv_cache_clear(ctx); + llama_kv_self_clear(ctx); for (int j = 0; j < num_batches; ++j) { const int batch_start = start + j * n_batch; @@ -850,7 +851,7 @@ static void hellaswag_score(llama_context * ctx, const common_params & params) { LOG_INF("%s : calculating hellaswag score over selected tasks.\n", __func__); - LOG("\ntask\tacc_norm\n"); + LOG("\ntask\tacc_norm\t95%% confidence interval\n"); double acc = 0.0f; @@ -923,7 +924,7 @@ static void hellaswag_score(llama_context * ctx, const common_params & params) { return; } - llama_kv_cache_clear(ctx); + llama_kv_self_clear(ctx); // decode all tasks [i0, i1) if (!decode_helper(ctx, batch, batch_logits, n_batch, n_vocab)) { @@ -984,8 +985,22 @@ static void hellaswag_score(llama_context * ctx, const common_params & params) { acc += 1.0; } - // Print the accumulated accuracy mean x 100 - LOG("%zu\t%.8lf\n", i + 1, acc/double(i + 1)*100.0); + double freq = acc / double(i + 1); + + const double za = 1.95996398454; + + // // Wald normal approx + // double conf =za*sqrt(freq*(1-freq)/double(i + 1)); + // LOG("%zu\t%.8lf +/- %.8lf\n", i + 1, freq*100.0, conf*100.0); + + // Wilson score interval, more accurate + double z = za * za / double(i + 1); + double cnf = z * sqrt(double(i + 1) * (4.0 * freq * (1 - freq) + z)) / (za + za); + double a = (freq + z * 0.5 - cnf) / (1.0 + z); + double b = (freq + z * 0.5 + cnf) / (1.0 + z); + + // Print the accumulated accuracy mean x 100 and confidence interval + LOG("%zu\t%3.8lf%%\t[%3.4lf%%, %3.4lf%%]\n", i + 1, freq * 100.0, a * 100.0, b * 100.0); } i0 = i1 - 1; @@ -1202,7 +1217,7 @@ static void winogrande_score(llama_context * ctx, const common_params & params) return; } - llama_kv_cache_clear(ctx); + llama_kv_self_clear(ctx); // decode all tasks [i0, i1) if (!decode_helper(ctx, batch, batch_logits, n_batch, n_vocab)) { @@ -1574,7 +1589,7 @@ static void multiple_choice_score(llama_context * ctx, const common_params & par return; } - llama_kv_cache_clear(ctx); + llama_kv_self_clear(ctx); // decode all tasks [i0, i1) if (!decode_helper(ctx, batch, batch_logits, n_batch, n_vocab)) { @@ -1764,7 +1779,7 @@ static void kl_divergence(llama_context * ctx, const common_params & params) { } // clear the KV cache - llama_kv_cache_clear(ctx); + llama_kv_self_clear(ctx); llama_batch batch = llama_batch_init(n_batch, 0, 1); diff --git a/examples/pydantic_models_to_grammar_examples.py b/examples/pydantic_models_to_grammar_examples.py index eb000d5cc..f94b82ca4 100755 --- a/examples/pydantic_models_to_grammar_examples.py +++ b/examples/pydantic_models_to_grammar_examples.py @@ -23,7 +23,7 @@ def create_completion(host, prompt, gbnf_grammar): """Calls the /completion API on llama-server. See - https://github.com/ggerganov/llama.cpp/tree/HEAD/examples/server#api-endpoints + https://github.com/ggml-org/llama.cpp/tree/HEAD/examples/server#api-endpoints """ print(f" Request:\n Grammar:\n{textwrap.indent(gbnf_grammar, ' ')}\n Prompt:\n{textwrap.indent(prompt.rstrip(), ' ')}") headers = {"Content-Type": "application/json"} diff --git a/examples/quantize-stats/quantize-stats.cpp b/examples/quantize-stats/quantize-stats.cpp index bd2f73467..dd07ab9b3 100644 --- a/examples/quantize-stats/quantize-stats.cpp +++ b/examples/quantize-stats/quantize-stats.cpp @@ -1,6 +1,6 @@ #include "ggml.h" #include "llama.h" -#include "llama-context.h" +#include "llama-model.h" #include "common.h" #include @@ -328,7 +328,7 @@ int main(int argc, char ** argv) { } } - const auto & tensors = llama_internal_get_tensor_map(ctx); + const auto & tensors = llama_internal_get_tensor_map(model); // check layer tensors int included_layers = 0; diff --git a/examples/quantize/README.md b/examples/quantize/README.md index f9cce7b21..992d00e21 100644 --- a/examples/quantize/README.md +++ b/examples/quantize/README.md @@ -69,22 +69,22 @@ Several quantization methods are supported. They differ in the resulting model d | 13B | ms/tok @ 8th | - | 73 | 82 | 98 | 105 | 128 | | 13B | bits/weight | 16.0 | 4.5 | 5.0 | 5.5 | 6.0 | 8.5 | -- [k-quants](https://github.com/ggerganov/llama.cpp/pull/1684) +- [k-quants](https://github.com/ggml-org/llama.cpp/pull/1684) - recent k-quants improvements and new i-quants - - [#2707](https://github.com/ggerganov/llama.cpp/pull/2707) - - [#2807](https://github.com/ggerganov/llama.cpp/pull/2807) - - [#4773 - 2-bit i-quants (inference)](https://github.com/ggerganov/llama.cpp/pull/4773) - - [#4856 - 2-bit i-quants (inference)](https://github.com/ggerganov/llama.cpp/pull/4856) - - [#4861 - importance matrix](https://github.com/ggerganov/llama.cpp/pull/4861) - - [#4872 - MoE models](https://github.com/ggerganov/llama.cpp/pull/4872) - - [#4897 - 2-bit quantization](https://github.com/ggerganov/llama.cpp/pull/4897) - - [#4930 - imatrix for all k-quants](https://github.com/ggerganov/llama.cpp/pull/4930) - - [#4951 - imatrix on the GPU](https://github.com/ggerganov/llama.cpp/pull/4957) - - [#4969 - imatrix for legacy quants](https://github.com/ggerganov/llama.cpp/pull/4969) - - [#4996 - k-quants tuning](https://github.com/ggerganov/llama.cpp/pull/4996) - - [#5060 - Q3_K_XS](https://github.com/ggerganov/llama.cpp/pull/5060) - - [#5196 - 3-bit i-quants](https://github.com/ggerganov/llama.cpp/pull/5196) - - [quantization tuning](https://github.com/ggerganov/llama.cpp/pull/5320), [another one](https://github.com/ggerganov/llama.cpp/pull/5334), and [another one](https://github.com/ggerganov/llama.cpp/pull/5361) + - [#2707](https://github.com/ggml-org/llama.cpp/pull/2707) + - [#2807](https://github.com/ggml-org/llama.cpp/pull/2807) + - [#4773 - 2-bit i-quants (inference)](https://github.com/ggml-org/llama.cpp/pull/4773) + - [#4856 - 2-bit i-quants (inference)](https://github.com/ggml-org/llama.cpp/pull/4856) + - [#4861 - importance matrix](https://github.com/ggml-org/llama.cpp/pull/4861) + - [#4872 - MoE models](https://github.com/ggml-org/llama.cpp/pull/4872) + - [#4897 - 2-bit quantization](https://github.com/ggml-org/llama.cpp/pull/4897) + - [#4930 - imatrix for all k-quants](https://github.com/ggml-org/llama.cpp/pull/4930) + - [#4951 - imatrix on the GPU](https://github.com/ggml-org/llama.cpp/pull/4957) + - [#4969 - imatrix for legacy quants](https://github.com/ggml-org/llama.cpp/pull/4969) + - [#4996 - k-quants tuning](https://github.com/ggml-org/llama.cpp/pull/4996) + - [#5060 - Q3_K_XS](https://github.com/ggml-org/llama.cpp/pull/5060) + - [#5196 - 3-bit i-quants](https://github.com/ggml-org/llama.cpp/pull/5196) + - [quantization tuning](https://github.com/ggml-org/llama.cpp/pull/5320), [another one](https://github.com/ggml-org/llama.cpp/pull/5334), and [another one](https://github.com/ggml-org/llama.cpp/pull/5361) **Llama 2 7B** diff --git a/examples/quantize/quantize.cpp b/examples/quantize/quantize.cpp index 355aef4a6..7f2afe657 100644 --- a/examples/quantize/quantize.cpp +++ b/examples/quantize/quantize.cpp @@ -8,6 +8,9 @@ #include #include #include +#include +#include +#include struct quant_option { std::string name; diff --git a/examples/retrieval/README.md b/examples/retrieval/README.md index bc5f22e2f..6938a1e96 100644 --- a/examples/retrieval/README.md +++ b/examples/retrieval/README.md @@ -3,7 +3,7 @@ Demonstration of simple retrieval technique based on cosine similarity More info: -https://github.com/ggerganov/llama.cpp/pull/6193 +https://github.com/ggml-org/llama.cpp/pull/6193 ### How to use diff --git a/examples/retrieval/retrieval.cpp b/examples/retrieval/retrieval.cpp index 2439022a2..0efe20d4b 100644 --- a/examples/retrieval/retrieval.cpp +++ b/examples/retrieval/retrieval.cpp @@ -83,7 +83,7 @@ static void batch_add_seq(llama_batch & batch, const std::vector & toke static void batch_decode(llama_context * ctx, llama_batch & batch, float * output, int n_seq, int n_embd) { // clear previous kv_cache values (irrelevant for embeddings) - llama_kv_cache_clear(ctx); + llama_kv_self_clear(ctx); // run model LOG_INF("%s: n_tokens = %d, n_seq = %d\n", __func__, batch.n_tokens, n_seq); diff --git a/examples/rpc/CMakeLists.txt b/examples/rpc/CMakeLists.txt index ae48fb98d..c2c748148 100644 --- a/examples/rpc/CMakeLists.txt +++ b/examples/rpc/CMakeLists.txt @@ -1,2 +1,4 @@ -add_executable(rpc-server rpc-server.cpp) -target_link_libraries(rpc-server PRIVATE ggml llama) +set(TARGET rpc-server) +add_executable(${TARGET} rpc-server.cpp) +target_link_libraries(${TARGET} PRIVATE ggml) +target_compile_features(${TARGET} PRIVATE cxx_std_17) diff --git a/examples/rpc/README.md b/examples/rpc/README.md index 312bb634d..561f19fda 100644 --- a/examples/rpc/README.md +++ b/examples/rpc/README.md @@ -72,3 +72,14 @@ $ bin/llama-cli -m ../models/tinyllama-1b/ggml-model-f16.gguf -p "Hello, my name This way you can offload model layers to both local and remote devices. +### Local cache + +The RPC server can use a local cache to store large tensors and avoid transferring them over the network. +This can speed up model loading significantly, especially when using large models. +To enable the cache, use the `-c` option: + +```bash +$ bin/rpc-server -c +``` + +By default, the cache is stored in the `$HOME/.cache/llama.cpp/rpc` directory and can be controlled via the `LLAMA_CACHE` environment variable. diff --git a/examples/rpc/rpc-server.cpp b/examples/rpc/rpc-server.cpp index 8b1b23eda..f8f2cb90e 100644 --- a/examples/rpc/rpc-server.cpp +++ b/examples/rpc/rpc-server.cpp @@ -1,3 +1,7 @@ +#if defined(_MSC_VER) +#define _SILENCE_CXX17_CODECVT_HEADER_DEPRECATION_WARNING +#endif + #include "ggml-cpu.h" #ifdef GGML_USE_CUDA @@ -18,26 +22,144 @@ #include "ggml-rpc.h" #ifdef _WIN32 +# define DIRECTORY_SEPARATOR '\\' +# include # include +# include +# include #else +# define DIRECTORY_SEPARATOR '/' # include +# include #endif +#include #include #include +#include +#include + +namespace fs = std::filesystem; + +// NOTE: this is copied from common.cpp to avoid linking with libcommon +// returns true if successful, false otherwise +static bool fs_create_directory_with_parents(const std::string & path) { +#ifdef _WIN32 + std::wstring_convert> converter; + std::wstring wpath = converter.from_bytes(path); + + // if the path already exists, check whether it's a directory + const DWORD attributes = GetFileAttributesW(wpath.c_str()); + if ((attributes != INVALID_FILE_ATTRIBUTES) && (attributes & FILE_ATTRIBUTE_DIRECTORY)) { + return true; + } + + size_t pos_slash = 0; + + // process path from front to back, procedurally creating directories + while ((pos_slash = path.find('\\', pos_slash)) != std::string::npos) { + const std::wstring subpath = wpath.substr(0, pos_slash); + const wchar_t * test = subpath.c_str(); + + const bool success = CreateDirectoryW(test, NULL); + if (!success) { + const DWORD error = GetLastError(); + + // if the path already exists, ensure that it's a directory + if (error == ERROR_ALREADY_EXISTS) { + const DWORD attributes = GetFileAttributesW(subpath.c_str()); + if (attributes == INVALID_FILE_ATTRIBUTES || !(attributes & FILE_ATTRIBUTE_DIRECTORY)) { + return false; + } + } else { + return false; + } + } + + pos_slash += 1; + } + + return true; +#else + // if the path already exists, check whether it's a directory + struct stat info; + if (stat(path.c_str(), &info) == 0) { + return S_ISDIR(info.st_mode); + } + + size_t pos_slash = 1; // skip leading slashes for directory creation + + // process path from front to back, procedurally creating directories + while ((pos_slash = path.find('/', pos_slash)) != std::string::npos) { + const std::string subpath = path.substr(0, pos_slash); + struct stat info; + + // if the path already exists, ensure that it's a directory + if (stat(subpath.c_str(), &info) == 0) { + if (!S_ISDIR(info.st_mode)) { + return false; + } + } else { + // create parent directories + const int ret = mkdir(subpath.c_str(), 0755); + if (ret != 0) { + return false; + } + } + + pos_slash += 1; + } + + return true; +#endif // _WIN32 +} + +// NOTE: this is copied from common.cpp to avoid linking with libcommon +static std::string fs_get_cache_directory() { + std::string cache_directory = ""; + auto ensure_trailing_slash = [](std::string p) { + // Make sure to add trailing slash + if (p.back() != DIRECTORY_SEPARATOR) { + p += DIRECTORY_SEPARATOR; + } + return p; + }; + if (getenv("LLAMA_CACHE")) { + cache_directory = std::getenv("LLAMA_CACHE"); + } else { +#if defined(__linux__) || defined(__FreeBSD__) || defined(_AIX) + if (std::getenv("XDG_CACHE_HOME")) { + cache_directory = std::getenv("XDG_CACHE_HOME"); + } else { + cache_directory = std::getenv("HOME") + std::string("/.cache/"); + } +#elif defined(__APPLE__) + cache_directory = std::getenv("HOME") + std::string("/Library/Caches/"); +#elif defined(_WIN32) + cache_directory = std::getenv("LOCALAPPDATA"); +#else +# error Unknown architecture +#endif + cache_directory = ensure_trailing_slash(cache_directory); + cache_directory += "llama.cpp"; + } + return ensure_trailing_slash(cache_directory); +} struct rpc_server_params { std::string host = "127.0.0.1"; int port = 50052; size_t backend_mem = 0; + bool use_cache = false; }; static void print_usage(int /*argc*/, char ** argv, rpc_server_params params) { fprintf(stderr, "Usage: %s [options]\n\n", argv[0]); fprintf(stderr, "options:\n"); - fprintf(stderr, " -h, --help show this help message and exit\n"); - fprintf(stderr, " -H HOST, --host HOST host to bind to (default: %s)\n", params.host.c_str()); - fprintf(stderr, " -p PORT, --port PORT port to bind to (default: %d)\n", params.port); - fprintf(stderr, " -m MEM, --mem MEM backend memory size (in MB)\n"); + fprintf(stderr, " -h, --help show this help message and exit\n"); + fprintf(stderr, " -H HOST, --host HOST host to bind to (default: %s)\n", params.host.c_str()); + fprintf(stderr, " -p PORT, --port PORT port to bind to (default: %d)\n", params.port); + fprintf(stderr, " -m MEM, --mem MEM backend memory size (in MB)\n"); + fprintf(stderr, " -c, --cache enable local file cache\n"); fprintf(stderr, "\n"); } @@ -58,6 +180,8 @@ static bool rpc_server_params_parse(int argc, char ** argv, rpc_server_params & if (params.port <= 0 || params.port > 65535) { return false; } + } else if (arg == "-c" || arg == "--cache") { + params.use_cache = true; } else if (arg == "-m" || arg == "--mem") { if (++i >= argc) { return false; @@ -164,8 +288,20 @@ int main(int argc, char * argv[]) { } else { get_backend_memory(&free_mem, &total_mem); } - printf("Starting RPC server on %s, backend memory: %zu MB\n", endpoint.c_str(), free_mem / (1024 * 1024)); - ggml_backend_rpc_start_server(backend, endpoint.c_str(), free_mem, total_mem); + const char * cache_dir = nullptr; + std::string cache_dir_str = fs_get_cache_directory() + "rpc/"; + if (params.use_cache) { + if (!fs_create_directory_with_parents(cache_dir_str)) { + fprintf(stderr, "Failed to create cache directory: %s\n", cache_dir_str.c_str()); + return 1; + } + cache_dir = cache_dir_str.c_str(); + } + printf("Starting RPC server\n"); + printf(" endpoint : %s\n", endpoint.c_str()); + printf(" local cache : %s\n", cache_dir ? cache_dir : "n/a"); + printf(" backend memory : %zu MB\n", free_mem / (1024 * 1024)); + ggml_backend_rpc_start_server(backend, endpoint.c_str(), cache_dir, free_mem, total_mem); ggml_backend_free(backend); return 0; } diff --git a/examples/run/CMakeLists.txt b/examples/run/CMakeLists.txt index cd6b0520e..7cff188ca 100644 --- a/examples/run/CMakeLists.txt +++ b/examples/run/CMakeLists.txt @@ -1,5 +1,16 @@ set(TARGET llama-run) add_executable(${TARGET} run.cpp linenoise.cpp/linenoise.cpp) + +# TODO: avoid copying this code block from common/CMakeLists.txt +set(LLAMA_RUN_EXTRA_LIBS "") +if (LLAMA_CURL) + find_package(CURL REQUIRED) + target_compile_definitions(${TARGET} PUBLIC LLAMA_USE_CURL) + include_directories(${CURL_INCLUDE_DIRS}) + find_library(CURL_LIBRARY curl REQUIRED) + set(LLAMA_RUN_EXTRA_LIBS ${LLAMA_RUN_EXTRA_LIBS} ${CURL_LIBRARY}) +endif () + install(TARGETS ${TARGET} RUNTIME) -target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT}) +target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT} ${LLAMA_RUN_EXTRA_LIBS}) target_compile_features(${TARGET} PRIVATE cxx_std_17) diff --git a/examples/run/linenoise.cpp/linenoise.cpp b/examples/run/linenoise.cpp/linenoise.cpp index a68f12a1a..9cb939900 100644 --- a/examples/run/linenoise.cpp/linenoise.cpp +++ b/examples/run/linenoise.cpp/linenoise.cpp @@ -75,7 +75,7 @@ * * DSR (Device Status Report) * Sequence: ESC [ 6 n - * Effect: reports the current cusor position as ESC [ n ; m R + * Effect: reports the current cursor position as ESC [ n ; m R * where n is the row and m is the column * * When multi line mode is enabled, we also use an additional escape @@ -107,6 +107,7 @@ # include # include +# include # include # include # include @@ -205,9 +206,9 @@ class File { int fd = -1; }; -__attribute__((format(printf, 1, 2))) -/* Debugging function. */ #if 0 +/* Debugging function. */ +__attribute__((format(printf, 1, 2))) static void lndebug(const char *fmt, ...) { static File file; if (file.file == nullptr) { @@ -222,11 +223,469 @@ static void lndebug(const char *fmt, ...) { fflush(file.file); } } -#else -static void lndebug(const char *, ...) { -} #endif +/* ========================== Encoding functions ============================= */ + +/* Get length of previous UTF8 codepoint */ +static size_t prevUtf8CodePointLen(const char * buf, int pos) { + int end = pos--; + while (pos >= 0 && ((unsigned char) buf[pos] & 0xC0) == 0x80) { + pos--; + } + return end - pos; +} + +/* Convert UTF8 to Unicode code point */ +static size_t utf8BytesToCodePoint(const char * buf, size_t len, int * cp) { + if (len) { + unsigned char byte = buf[0]; + if ((byte & 0x80) == 0) { + *cp = byte; + return 1; + } else if ((byte & 0xE0) == 0xC0) { + if (len >= 2) { + *cp = (((unsigned long) (buf[0] & 0x1F)) << 6) | ((unsigned long) (buf[1] & 0x3F)); + return 2; + } + } else if ((byte & 0xF0) == 0xE0) { + if (len >= 3) { + *cp = (((unsigned long) (buf[0] & 0x0F)) << 12) | (((unsigned long) (buf[1] & 0x3F)) << 6) | + ((unsigned long) (buf[2] & 0x3F)); + return 3; + } + } else if ((byte & 0xF8) == 0xF0) { + if (len >= 4) { + *cp = (((unsigned long) (buf[0] & 0x07)) << 18) | (((unsigned long) (buf[1] & 0x3F)) << 12) | + (((unsigned long) (buf[2] & 0x3F)) << 6) | ((unsigned long) (buf[3] & 0x3F)); + return 4; + } + } + } + return 0; +} + +/* Check if the code is a wide character */ +static const unsigned long wideCharTable[][2] = { + /* BEGIN: WIDE CHAR TABLE */ + { 0x1100, 0x115F }, + { 0x231A, 0x231B }, + { 0x2329, 0x232A }, + { 0x23E9, 0x23EC }, + { 0x23F0, 0x23F0 }, + { 0x23F3, 0x23F3 }, + { 0x25FD, 0x25FE }, + { 0x2614, 0x2615 }, + { 0x2630, 0x2637 }, + { 0x2648, 0x2653 }, + { 0x267F, 0x267F }, + { 0x268A, 0x268F }, + { 0x2693, 0x2693 }, + { 0x26A1, 0x26A1 }, + { 0x26AA, 0x26AB }, + { 0x26BD, 0x26BE }, + { 0x26C4, 0x26C5 }, + { 0x26CE, 0x26CE }, + { 0x26D4, 0x26D4 }, + { 0x26EA, 0x26EA }, + { 0x26F2, 0x26F3 }, + { 0x26F5, 0x26F5 }, + { 0x26FA, 0x26FA }, + { 0x26FD, 0x26FD }, + { 0x2705, 0x2705 }, + { 0x270A, 0x270B }, + { 0x2728, 0x2728 }, + { 0x274C, 0x274C }, + { 0x274E, 0x274E }, + { 0x2753, 0x2755 }, + { 0x2757, 0x2757 }, + { 0x2795, 0x2797 }, + { 0x27B0, 0x27B0 }, + { 0x27BF, 0x27BF }, + { 0x2B1B, 0x2B1C }, + { 0x2B50, 0x2B50 }, + { 0x2B55, 0x2B55 }, + { 0x2E80, 0x2E99 }, + { 0x2E9B, 0x2EF3 }, + { 0x2F00, 0x2FD5 }, + { 0x2FF0, 0x303E }, + { 0x3041, 0x3096 }, + { 0x3099, 0x30FF }, + { 0x3105, 0x312F }, + { 0x3131, 0x318E }, + { 0x3190, 0x31E5 }, + { 0x31EF, 0x321E }, + { 0x3220, 0x3247 }, + { 0x3250, 0xA48C }, + { 0xA490, 0xA4C6 }, + { 0xA960, 0xA97C }, + { 0xAC00, 0xD7A3 }, + { 0xF900, 0xFAFF }, + { 0xFE10, 0xFE19 }, + { 0xFE30, 0xFE52 }, + { 0xFE54, 0xFE66 }, + { 0xFE68, 0xFE6B }, + { 0xFF01, 0xFF60 }, + { 0xFFE0, 0xFFE6 }, + { 0x16FE0, 0x16FE4 }, + { 0x16FF0, 0x16FF1 }, + { 0x17000, 0x187F7 }, + { 0x18800, 0x18CD5 }, + { 0x18CFF, 0x18D08 }, + { 0x1AFF0, 0x1AFF3 }, + { 0x1AFF5, 0x1AFFB }, + { 0x1AFFD, 0x1AFFE }, + { 0x1B000, 0x1B122 }, + { 0x1B132, 0x1B132 }, + { 0x1B150, 0x1B152 }, + { 0x1B155, 0x1B155 }, + { 0x1B164, 0x1B167 }, + { 0x1B170, 0x1B2FB }, + { 0x1D300, 0x1D356 }, + { 0x1D360, 0x1D376 }, + { 0x1F004, 0x1F004 }, + { 0x1F0CF, 0x1F0CF }, + { 0x1F18E, 0x1F18E }, + { 0x1F191, 0x1F19A }, + { 0x1F200, 0x1F202 }, + { 0x1F210, 0x1F23B }, + { 0x1F240, 0x1F248 }, + { 0x1F250, 0x1F251 }, + { 0x1F260, 0x1F265 }, + { 0x1F300, 0x1F320 }, + { 0x1F32D, 0x1F335 }, + { 0x1F337, 0x1F37C }, + { 0x1F37E, 0x1F393 }, + { 0x1F3A0, 0x1F3CA }, + { 0x1F3CF, 0x1F3D3 }, + { 0x1F3E0, 0x1F3F0 }, + { 0x1F3F4, 0x1F3F4 }, + { 0x1F3F8, 0x1F43E }, + { 0x1F440, 0x1F440 }, + { 0x1F442, 0x1F4FC }, + { 0x1F4FF, 0x1F53D }, + { 0x1F54B, 0x1F54E }, + { 0x1F550, 0x1F567 }, + { 0x1F57A, 0x1F57A }, + { 0x1F595, 0x1F596 }, + { 0x1F5A4, 0x1F5A4 }, + { 0x1F5FB, 0x1F64F }, + { 0x1F680, 0x1F6C5 }, + { 0x1F6CC, 0x1F6CC }, + { 0x1F6D0, 0x1F6D2 }, + { 0x1F6D5, 0x1F6D7 }, + { 0x1F6DC, 0x1F6DF }, + { 0x1F6EB, 0x1F6EC }, + { 0x1F6F4, 0x1F6FC }, + { 0x1F7E0, 0x1F7EB }, + { 0x1F7F0, 0x1F7F0 }, + { 0x1F90C, 0x1F93A }, + { 0x1F93C, 0x1F945 }, + { 0x1F947, 0x1F9FF }, + { 0x1FA70, 0x1FA7C }, + { 0x1FA80, 0x1FA89 }, + { 0x1FA8F, 0x1FAC6 }, + { 0x1FACE, 0x1FADC }, + { 0x1FADF, 0x1FAE9 }, + { 0x1FAF0, 0x1FAF8 }, + { 0x20000, 0x2FFFD }, + { 0x30000, 0x3FFFD } + /* END: WIDE CHAR TABLE */ +}; + +static const size_t wideCharTableSize = sizeof(wideCharTable) / sizeof(wideCharTable[0]); + +static bool isWideChar(unsigned long cp) { + for (size_t i = 0; i < wideCharTableSize; i++) { + auto first_code = wideCharTable[i][0]; + auto last_code = wideCharTable[i][1]; + if (first_code > cp) { + return false; + } + if (first_code <= cp && cp <= last_code) { + return true; + } + } + return false; +} + +/* Check if the code is a combining character */ +static const unsigned long combiningCharTable[] = { + /* BEGIN: COMBINING CHAR TABLE */ + 0x0300, 0x0301, 0x0302, 0x0303, 0x0304, 0x0305, 0x0306, 0x0307, 0x0308, 0x0309, 0x030A, 0x030B, 0x030C, + 0x030D, 0x030E, 0x030F, 0x0310, 0x0311, 0x0312, 0x0313, 0x0314, 0x0315, 0x0316, 0x0317, 0x0318, 0x0319, + 0x031A, 0x031B, 0x031C, 0x031D, 0x031E, 0x031F, 0x0320, 0x0321, 0x0322, 0x0323, 0x0324, 0x0325, 0x0326, + 0x0327, 0x0328, 0x0329, 0x032A, 0x032B, 0x032C, 0x032D, 0x032E, 0x032F, 0x0330, 0x0331, 0x0332, 0x0333, + 0x0334, 0x0335, 0x0336, 0x0337, 0x0338, 0x0339, 0x033A, 0x033B, 0x033C, 0x033D, 0x033E, 0x033F, 0x0340, + 0x0341, 0x0342, 0x0343, 0x0344, 0x0345, 0x0346, 0x0347, 0x0348, 0x0349, 0x034A, 0x034B, 0x034C, 0x034D, + 0x034E, 0x034F, 0x0350, 0x0351, 0x0352, 0x0353, 0x0354, 0x0355, 0x0356, 0x0357, 0x0358, 0x0359, 0x035A, + 0x035B, 0x035C, 0x035D, 0x035E, 0x035F, 0x0360, 0x0361, 0x0362, 0x0363, 0x0364, 0x0365, 0x0366, 0x0367, + 0x0368, 0x0369, 0x036A, 0x036B, 0x036C, 0x036D, 0x036E, 0x036F, 0x0483, 0x0484, 0x0485, 0x0486, 0x0487, + 0x0591, 0x0592, 0x0593, 0x0594, 0x0595, 0x0596, 0x0597, 0x0598, 0x0599, 0x059A, 0x059B, 0x059C, 0x059D, + 0x059E, 0x059F, 0x05A0, 0x05A1, 0x05A2, 0x05A3, 0x05A4, 0x05A5, 0x05A6, 0x05A7, 0x05A8, 0x05A9, 0x05AA, + 0x05AB, 0x05AC, 0x05AD, 0x05AE, 0x05AF, 0x05B0, 0x05B1, 0x05B2, 0x05B3, 0x05B4, 0x05B5, 0x05B6, 0x05B7, + 0x05B8, 0x05B9, 0x05BA, 0x05BB, 0x05BC, 0x05BD, 0x05BF, 0x05C1, 0x05C2, 0x05C4, 0x05C5, 0x05C7, 0x0610, + 0x0611, 0x0612, 0x0613, 0x0614, 0x0615, 0x0616, 0x0617, 0x0618, 0x0619, 0x061A, 0x064B, 0x064C, 0x064D, + 0x064E, 0x064F, 0x0650, 0x0651, 0x0652, 0x0653, 0x0654, 0x0655, 0x0656, 0x0657, 0x0658, 0x0659, 0x065A, + 0x065B, 0x065C, 0x065D, 0x065E, 0x065F, 0x0670, 0x06D6, 0x06D7, 0x06D8, 0x06D9, 0x06DA, 0x06DB, 0x06DC, + 0x06DF, 0x06E0, 0x06E1, 0x06E2, 0x06E3, 0x06E4, 0x06E7, 0x06E8, 0x06EA, 0x06EB, 0x06EC, 0x06ED, 0x0711, + 0x0730, 0x0731, 0x0732, 0x0733, 0x0734, 0x0735, 0x0736, 0x0737, 0x0738, 0x0739, 0x073A, 0x073B, 0x073C, + 0x073D, 0x073E, 0x073F, 0x0740, 0x0741, 0x0742, 0x0743, 0x0744, 0x0745, 0x0746, 0x0747, 0x0748, 0x0749, + 0x074A, 0x07A6, 0x07A7, 0x07A8, 0x07A9, 0x07AA, 0x07AB, 0x07AC, 0x07AD, 0x07AE, 0x07AF, 0x07B0, 0x07EB, + 0x07EC, 0x07ED, 0x07EE, 0x07EF, 0x07F0, 0x07F1, 0x07F2, 0x07F3, 0x07FD, 0x0816, 0x0817, 0x0818, 0x0819, + 0x081B, 0x081C, 0x081D, 0x081E, 0x081F, 0x0820, 0x0821, 0x0822, 0x0823, 0x0825, 0x0826, 0x0827, 0x0829, + 0x082A, 0x082B, 0x082C, 0x082D, 0x0859, 0x085A, 0x085B, 0x0897, 0x0898, 0x0899, 0x089A, 0x089B, 0x089C, + 0x089D, 0x089E, 0x089F, 0x08CA, 0x08CB, 0x08CC, 0x08CD, 0x08CE, 0x08CF, 0x08D0, 0x08D1, 0x08D2, 0x08D3, + 0x08D4, 0x08D5, 0x08D6, 0x08D7, 0x08D8, 0x08D9, 0x08DA, 0x08DB, 0x08DC, 0x08DD, 0x08DE, 0x08DF, 0x08E0, + 0x08E1, 0x08E3, 0x08E4, 0x08E5, 0x08E6, 0x08E7, 0x08E8, 0x08E9, 0x08EA, 0x08EB, 0x08EC, 0x08ED, 0x08EE, + 0x08EF, 0x08F0, 0x08F1, 0x08F2, 0x08F3, 0x08F4, 0x08F5, 0x08F6, 0x08F7, 0x08F8, 0x08F9, 0x08FA, 0x08FB, + 0x08FC, 0x08FD, 0x08FE, 0x08FF, 0x0900, 0x0901, 0x0902, 0x093A, 0x093C, 0x0941, 0x0942, 0x0943, 0x0944, + 0x0945, 0x0946, 0x0947, 0x0948, 0x094D, 0x0951, 0x0952, 0x0953, 0x0954, 0x0955, 0x0956, 0x0957, 0x0962, + 0x0963, 0x0981, 0x09BC, 0x09C1, 0x09C2, 0x09C3, 0x09C4, 0x09CD, 0x09E2, 0x09E3, 0x09FE, 0x0A01, 0x0A02, + 0x0A3C, 0x0A41, 0x0A42, 0x0A47, 0x0A48, 0x0A4B, 0x0A4C, 0x0A4D, 0x0A51, 0x0A70, 0x0A71, 0x0A75, 0x0A81, + 0x0A82, 0x0ABC, 0x0AC1, 0x0AC2, 0x0AC3, 0x0AC4, 0x0AC5, 0x0AC7, 0x0AC8, 0x0ACD, 0x0AE2, 0x0AE3, 0x0AFA, + 0x0AFB, 0x0AFC, 0x0AFD, 0x0AFE, 0x0AFF, 0x0B01, 0x0B3C, 0x0B3F, 0x0B41, 0x0B42, 0x0B43, 0x0B44, 0x0B4D, + 0x0B55, 0x0B56, 0x0B62, 0x0B63, 0x0B82, 0x0BC0, 0x0BCD, 0x0C00, 0x0C04, 0x0C3C, 0x0C3E, 0x0C3F, 0x0C40, + 0x0C46, 0x0C47, 0x0C48, 0x0C4A, 0x0C4B, 0x0C4C, 0x0C4D, 0x0C55, 0x0C56, 0x0C62, 0x0C63, 0x0C81, 0x0CBC, + 0x0CBF, 0x0CC6, 0x0CCC, 0x0CCD, 0x0CE2, 0x0CE3, 0x0D00, 0x0D01, 0x0D3B, 0x0D3C, 0x0D41, 0x0D42, 0x0D43, + 0x0D44, 0x0D4D, 0x0D62, 0x0D63, 0x0D81, 0x0DCA, 0x0DD2, 0x0DD3, 0x0DD4, 0x0DD6, 0x0E31, 0x0E34, 0x0E35, + 0x0E36, 0x0E37, 0x0E38, 0x0E39, 0x0E3A, 0x0E47, 0x0E48, 0x0E49, 0x0E4A, 0x0E4B, 0x0E4C, 0x0E4D, 0x0E4E, + 0x0EB1, 0x0EB4, 0x0EB5, 0x0EB6, 0x0EB7, 0x0EB8, 0x0EB9, 0x0EBA, 0x0EBB, 0x0EBC, 0x0EC8, 0x0EC9, 0x0ECA, + 0x0ECB, 0x0ECC, 0x0ECD, 0x0ECE, 0x0F18, 0x0F19, 0x0F35, 0x0F37, 0x0F39, 0x0F71, 0x0F72, 0x0F73, 0x0F74, + 0x0F75, 0x0F76, 0x0F77, 0x0F78, 0x0F79, 0x0F7A, 0x0F7B, 0x0F7C, 0x0F7D, 0x0F7E, 0x0F80, 0x0F81, 0x0F82, + 0x0F83, 0x0F84, 0x0F86, 0x0F87, 0x0F8D, 0x0F8E, 0x0F8F, 0x0F90, 0x0F91, 0x0F92, 0x0F93, 0x0F94, 0x0F95, + 0x0F96, 0x0F97, 0x0F99, 0x0F9A, 0x0F9B, 0x0F9C, 0x0F9D, 0x0F9E, 0x0F9F, 0x0FA0, 0x0FA1, 0x0FA2, 0x0FA3, + 0x0FA4, 0x0FA5, 0x0FA6, 0x0FA7, 0x0FA8, 0x0FA9, 0x0FAA, 0x0FAB, 0x0FAC, 0x0FAD, 0x0FAE, 0x0FAF, 0x0FB0, + 0x0FB1, 0x0FB2, 0x0FB3, 0x0FB4, 0x0FB5, 0x0FB6, 0x0FB7, 0x0FB8, 0x0FB9, 0x0FBA, 0x0FBB, 0x0FBC, 0x0FC6, + 0x102D, 0x102E, 0x102F, 0x1030, 0x1032, 0x1033, 0x1034, 0x1035, 0x1036, 0x1037, 0x1039, 0x103A, 0x103D, + 0x103E, 0x1058, 0x1059, 0x105E, 0x105F, 0x1060, 0x1071, 0x1072, 0x1073, 0x1074, 0x1082, 0x1085, 0x1086, + 0x108D, 0x109D, 0x135D, 0x135E, 0x135F, 0x1712, 0x1713, 0x1714, 0x1732, 0x1733, 0x1752, 0x1753, 0x1772, + 0x1773, 0x17B4, 0x17B5, 0x17B7, 0x17B8, 0x17B9, 0x17BA, 0x17BB, 0x17BC, 0x17BD, 0x17C6, 0x17C9, 0x17CA, + 0x17CB, 0x17CC, 0x17CD, 0x17CE, 0x17CF, 0x17D0, 0x17D1, 0x17D2, 0x17D3, 0x17DD, 0x180B, 0x180C, 0x180D, + 0x180F, 0x1885, 0x1886, 0x18A9, 0x1920, 0x1921, 0x1922, 0x1927, 0x1928, 0x1932, 0x1939, 0x193A, 0x193B, + 0x1A17, 0x1A18, 0x1A1B, 0x1A56, 0x1A58, 0x1A59, 0x1A5A, 0x1A5B, 0x1A5C, 0x1A5D, 0x1A5E, 0x1A60, 0x1A62, + 0x1A65, 0x1A66, 0x1A67, 0x1A68, 0x1A69, 0x1A6A, 0x1A6B, 0x1A6C, 0x1A73, 0x1A74, 0x1A75, 0x1A76, 0x1A77, + 0x1A78, 0x1A79, 0x1A7A, 0x1A7B, 0x1A7C, 0x1A7F, 0x1AB0, 0x1AB1, 0x1AB2, 0x1AB3, 0x1AB4, 0x1AB5, 0x1AB6, + 0x1AB7, 0x1AB8, 0x1AB9, 0x1ABA, 0x1ABB, 0x1ABC, 0x1ABD, 0x1ABF, 0x1AC0, 0x1AC1, 0x1AC2, 0x1AC3, 0x1AC4, + 0x1AC5, 0x1AC6, 0x1AC7, 0x1AC8, 0x1AC9, 0x1ACA, 0x1ACB, 0x1ACC, 0x1ACD, 0x1ACE, 0x1B00, 0x1B01, 0x1B02, + 0x1B03, 0x1B34, 0x1B36, 0x1B37, 0x1B38, 0x1B39, 0x1B3A, 0x1B3C, 0x1B42, 0x1B6B, 0x1B6C, 0x1B6D, 0x1B6E, + 0x1B6F, 0x1B70, 0x1B71, 0x1B72, 0x1B73, 0x1B80, 0x1B81, 0x1BA2, 0x1BA3, 0x1BA4, 0x1BA5, 0x1BA8, 0x1BA9, + 0x1BAB, 0x1BAC, 0x1BAD, 0x1BE6, 0x1BE8, 0x1BE9, 0x1BED, 0x1BEF, 0x1BF0, 0x1BF1, 0x1C2C, 0x1C2D, 0x1C2E, + 0x1C2F, 0x1C30, 0x1C31, 0x1C32, 0x1C33, 0x1C36, 0x1C37, 0x1CD0, 0x1CD1, 0x1CD2, 0x1CD4, 0x1CD5, 0x1CD6, + 0x1CD7, 0x1CD8, 0x1CD9, 0x1CDA, 0x1CDB, 0x1CDC, 0x1CDD, 0x1CDE, 0x1CDF, 0x1CE0, 0x1CE2, 0x1CE3, 0x1CE4, + 0x1CE5, 0x1CE6, 0x1CE7, 0x1CE8, 0x1CED, 0x1CF4, 0x1CF8, 0x1CF9, 0x1DC0, 0x1DC1, 0x1DC2, 0x1DC3, 0x1DC4, + 0x1DC5, 0x1DC6, 0x1DC7, 0x1DC8, 0x1DC9, 0x1DCA, 0x1DCB, 0x1DCC, 0x1DCD, 0x1DCE, 0x1DCF, 0x1DD0, 0x1DD1, + 0x1DD2, 0x1DD3, 0x1DD4, 0x1DD5, 0x1DD6, 0x1DD7, 0x1DD8, 0x1DD9, 0x1DDA, 0x1DDB, 0x1DDC, 0x1DDD, 0x1DDE, + 0x1DDF, 0x1DE0, 0x1DE1, 0x1DE2, 0x1DE3, 0x1DE4, 0x1DE5, 0x1DE6, 0x1DE7, 0x1DE8, 0x1DE9, 0x1DEA, 0x1DEB, + 0x1DEC, 0x1DED, 0x1DEE, 0x1DEF, 0x1DF0, 0x1DF1, 0x1DF2, 0x1DF3, 0x1DF4, 0x1DF5, 0x1DF6, 0x1DF7, 0x1DF8, + 0x1DF9, 0x1DFA, 0x1DFB, 0x1DFC, 0x1DFD, 0x1DFE, 0x1DFF, 0x20D0, 0x20D1, 0x20D2, 0x20D3, 0x20D4, 0x20D5, + 0x20D6, 0x20D7, 0x20D8, 0x20D9, 0x20DA, 0x20DB, 0x20DC, 0x20E1, 0x20E5, 0x20E6, 0x20E7, 0x20E8, 0x20E9, + 0x20EA, 0x20EB, 0x20EC, 0x20ED, 0x20EE, 0x20EF, 0x20F0, 0x2CEF, 0x2CF0, 0x2CF1, 0x2D7F, 0x2DE0, 0x2DE1, + 0x2DE2, 0x2DE3, 0x2DE4, 0x2DE5, 0x2DE6, 0x2DE7, 0x2DE8, 0x2DE9, 0x2DEA, 0x2DEB, 0x2DEC, 0x2DED, 0x2DEE, + 0x2DEF, 0x2DF0, 0x2DF1, 0x2DF2, 0x2DF3, 0x2DF4, 0x2DF5, 0x2DF6, 0x2DF7, 0x2DF8, 0x2DF9, 0x2DFA, 0x2DFB, + 0x2DFC, 0x2DFD, 0x2DFE, 0x2DFF, 0x302A, 0x302B, 0x302C, 0x302D, 0x3099, 0x309A, 0xA66F, 0xA674, 0xA675, + 0xA676, 0xA677, 0xA678, 0xA679, 0xA67A, 0xA67B, 0xA67C, 0xA67D, 0xA69E, 0xA69F, 0xA6F0, 0xA6F1, 0xA802, + 0xA806, 0xA80B, 0xA825, 0xA826, 0xA82C, 0xA8C4, 0xA8C5, 0xA8E0, 0xA8E1, 0xA8E2, 0xA8E3, 0xA8E4, 0xA8E5, + 0xA8E6, 0xA8E7, 0xA8E8, 0xA8E9, 0xA8EA, 0xA8EB, 0xA8EC, 0xA8ED, 0xA8EE, 0xA8EF, 0xA8F0, 0xA8F1, 0xA8FF, + 0xA926, 0xA927, 0xA928, 0xA929, 0xA92A, 0xA92B, 0xA92C, 0xA92D, 0xA947, 0xA948, 0xA949, 0xA94A, 0xA94B, + 0xA94C, 0xA94D, 0xA94E, 0xA94F, 0xA950, 0xA951, 0xA980, 0xA981, 0xA982, 0xA9B3, 0xA9B6, 0xA9B7, 0xA9B8, + 0xA9B9, 0xA9BC, 0xA9BD, 0xA9E5, 0xAA29, 0xAA2A, 0xAA2B, 0xAA2C, 0xAA2D, 0xAA2E, 0xAA31, 0xAA32, 0xAA35, + 0xAA36, 0xAA43, 0xAA4C, 0xAA7C, 0xAAB0, 0xAAB2, 0xAAB3, 0xAAB4, 0xAAB7, 0xAAB8, 0xAABE, 0xAABF, 0xAAC1, + 0xAAEC, 0xAAED, 0xAAF6, 0xABE5, 0xABE8, 0xABED, 0xFB1E, 0xFE00, 0xFE01, 0xFE02, 0xFE03, 0xFE04, 0xFE05, + 0xFE06, 0xFE07, 0xFE08, 0xFE09, 0xFE0A, 0xFE0B, 0xFE0C, 0xFE0D, 0xFE0E, 0xFE0F, 0xFE20, 0xFE21, 0xFE22, + 0xFE23, 0xFE24, 0xFE25, 0xFE26, 0xFE27, 0xFE28, 0xFE29, 0xFE2A, 0xFE2B, 0xFE2C, 0xFE2D, 0xFE2E, 0xFE2F, + 0x101FD, 0x102E0, 0x10376, 0x10377, 0x10378, 0x10379, 0x1037A, 0x10A01, 0x10A02, 0x10A03, 0x10A05, 0x10A06, 0x10A0C, + 0x10A0D, 0x10A0E, 0x10A0F, 0x10A38, 0x10A39, 0x10A3A, 0x10A3F, 0x10AE5, 0x10AE6, 0x10D24, 0x10D25, 0x10D26, 0x10D27, + 0x10D69, 0x10D6A, 0x10D6B, 0x10D6C, 0x10D6D, 0x10EAB, 0x10EAC, 0x10EFC, 0x10EFD, 0x10EFE, 0x10EFF, 0x10F46, 0x10F47, + 0x10F48, 0x10F49, 0x10F4A, 0x10F4B, 0x10F4C, 0x10F4D, 0x10F4E, 0x10F4F, 0x10F50, 0x10F82, 0x10F83, 0x10F84, 0x10F85, + 0x11001, 0x11038, 0x11039, 0x1103A, 0x1103B, 0x1103C, 0x1103D, 0x1103E, 0x1103F, 0x11040, 0x11041, 0x11042, 0x11043, + 0x11044, 0x11045, 0x11046, 0x11070, 0x11073, 0x11074, 0x1107F, 0x11080, 0x11081, 0x110B3, 0x110B4, 0x110B5, 0x110B6, + 0x110B9, 0x110BA, 0x110C2, 0x11100, 0x11101, 0x11102, 0x11127, 0x11128, 0x11129, 0x1112A, 0x1112B, 0x1112D, 0x1112E, + 0x1112F, 0x11130, 0x11131, 0x11132, 0x11133, 0x11134, 0x11173, 0x11180, 0x11181, 0x111B6, 0x111B7, 0x111B8, 0x111B9, + 0x111BA, 0x111BB, 0x111BC, 0x111BD, 0x111BE, 0x111C9, 0x111CA, 0x111CB, 0x111CC, 0x111CF, 0x1122F, 0x11230, 0x11231, + 0x11234, 0x11236, 0x11237, 0x1123E, 0x11241, 0x112DF, 0x112E3, 0x112E4, 0x112E5, 0x112E6, 0x112E7, 0x112E8, 0x112E9, + 0x112EA, 0x11300, 0x11301, 0x1133B, 0x1133C, 0x11340, 0x11366, 0x11367, 0x11368, 0x11369, 0x1136A, 0x1136B, 0x1136C, + 0x11370, 0x11371, 0x11372, 0x11373, 0x11374, 0x113BB, 0x113BC, 0x113BD, 0x113BE, 0x113BF, 0x113C0, 0x113CE, 0x113D0, + 0x113D2, 0x113E1, 0x113E2, 0x11438, 0x11439, 0x1143A, 0x1143B, 0x1143C, 0x1143D, 0x1143E, 0x1143F, 0x11442, 0x11443, + 0x11444, 0x11446, 0x1145E, 0x114B3, 0x114B4, 0x114B5, 0x114B6, 0x114B7, 0x114B8, 0x114BA, 0x114BF, 0x114C0, 0x114C2, + 0x114C3, 0x115B2, 0x115B3, 0x115B4, 0x115B5, 0x115BC, 0x115BD, 0x115BF, 0x115C0, 0x115DC, 0x115DD, 0x11633, 0x11634, + 0x11635, 0x11636, 0x11637, 0x11638, 0x11639, 0x1163A, 0x1163D, 0x1163F, 0x11640, 0x116AB, 0x116AD, 0x116B0, 0x116B1, + 0x116B2, 0x116B3, 0x116B4, 0x116B5, 0x116B7, 0x1171D, 0x1171F, 0x11722, 0x11723, 0x11724, 0x11725, 0x11727, 0x11728, + 0x11729, 0x1172A, 0x1172B, 0x1182F, 0x11830, 0x11831, 0x11832, 0x11833, 0x11834, 0x11835, 0x11836, 0x11837, 0x11839, + 0x1183A, 0x1193B, 0x1193C, 0x1193E, 0x11943, 0x119D4, 0x119D5, 0x119D6, 0x119D7, 0x119DA, 0x119DB, 0x119E0, 0x11A01, + 0x11A02, 0x11A03, 0x11A04, 0x11A05, 0x11A06, 0x11A07, 0x11A08, 0x11A09, 0x11A0A, 0x11A33, 0x11A34, 0x11A35, 0x11A36, + 0x11A37, 0x11A38, 0x11A3B, 0x11A3C, 0x11A3D, 0x11A3E, 0x11A47, 0x11A51, 0x11A52, 0x11A53, 0x11A54, 0x11A55, 0x11A56, + 0x11A59, 0x11A5A, 0x11A5B, 0x11A8A, 0x11A8B, 0x11A8C, 0x11A8D, 0x11A8E, 0x11A8F, 0x11A90, 0x11A91, 0x11A92, 0x11A93, + 0x11A94, 0x11A95, 0x11A96, 0x11A98, 0x11A99, 0x11C30, 0x11C31, 0x11C32, 0x11C33, 0x11C34, 0x11C35, 0x11C36, 0x11C38, + 0x11C39, 0x11C3A, 0x11C3B, 0x11C3C, 0x11C3D, 0x11C3F, 0x11C92, 0x11C93, 0x11C94, 0x11C95, 0x11C96, 0x11C97, 0x11C98, + 0x11C99, 0x11C9A, 0x11C9B, 0x11C9C, 0x11C9D, 0x11C9E, 0x11C9F, 0x11CA0, 0x11CA1, 0x11CA2, 0x11CA3, 0x11CA4, 0x11CA5, + 0x11CA6, 0x11CA7, 0x11CAA, 0x11CAB, 0x11CAC, 0x11CAD, 0x11CAE, 0x11CAF, 0x11CB0, 0x11CB2, 0x11CB3, 0x11CB5, 0x11CB6, + 0x11D31, 0x11D32, 0x11D33, 0x11D34, 0x11D35, 0x11D36, 0x11D3A, 0x11D3C, 0x11D3D, 0x11D3F, 0x11D40, 0x11D41, 0x11D42, + 0x11D43, 0x11D44, 0x11D45, 0x11D47, 0x11D90, 0x11D91, 0x11D95, 0x11D97, 0x11EF3, 0x11EF4, 0x11F00, 0x11F01, 0x11F36, + 0x11F37, 0x11F38, 0x11F39, 0x11F3A, 0x11F40, 0x11F42, 0x11F5A, 0x13440, 0x13447, 0x13448, 0x13449, 0x1344A, 0x1344B, + 0x1344C, 0x1344D, 0x1344E, 0x1344F, 0x13450, 0x13451, 0x13452, 0x13453, 0x13454, 0x13455, 0x1611E, 0x1611F, 0x16120, + 0x16121, 0x16122, 0x16123, 0x16124, 0x16125, 0x16126, 0x16127, 0x16128, 0x16129, 0x1612D, 0x1612E, 0x1612F, 0x16AF0, + 0x16AF1, 0x16AF2, 0x16AF3, 0x16AF4, 0x16B30, 0x16B31, 0x16B32, 0x16B33, 0x16B34, 0x16B35, 0x16B36, 0x16F4F, 0x16F8F, + 0x16F90, 0x16F91, 0x16F92, 0x16FE4, 0x1BC9D, 0x1BC9E, 0x1CF00, 0x1CF01, 0x1CF02, 0x1CF03, 0x1CF04, 0x1CF05, 0x1CF06, + 0x1CF07, 0x1CF08, 0x1CF09, 0x1CF0A, 0x1CF0B, 0x1CF0C, 0x1CF0D, 0x1CF0E, 0x1CF0F, 0x1CF10, 0x1CF11, 0x1CF12, 0x1CF13, + 0x1CF14, 0x1CF15, 0x1CF16, 0x1CF17, 0x1CF18, 0x1CF19, 0x1CF1A, 0x1CF1B, 0x1CF1C, 0x1CF1D, 0x1CF1E, 0x1CF1F, 0x1CF20, + 0x1CF21, 0x1CF22, 0x1CF23, 0x1CF24, 0x1CF25, 0x1CF26, 0x1CF27, 0x1CF28, 0x1CF29, 0x1CF2A, 0x1CF2B, 0x1CF2C, 0x1CF2D, + 0x1CF30, 0x1CF31, 0x1CF32, 0x1CF33, 0x1CF34, 0x1CF35, 0x1CF36, 0x1CF37, 0x1CF38, 0x1CF39, 0x1CF3A, 0x1CF3B, 0x1CF3C, + 0x1CF3D, 0x1CF3E, 0x1CF3F, 0x1CF40, 0x1CF41, 0x1CF42, 0x1CF43, 0x1CF44, 0x1CF45, 0x1CF46, 0x1D167, 0x1D168, 0x1D169, + 0x1D17B, 0x1D17C, 0x1D17D, 0x1D17E, 0x1D17F, 0x1D180, 0x1D181, 0x1D182, 0x1D185, 0x1D186, 0x1D187, 0x1D188, 0x1D189, + 0x1D18A, 0x1D18B, 0x1D1AA, 0x1D1AB, 0x1D1AC, 0x1D1AD, 0x1D242, 0x1D243, 0x1D244, 0x1DA00, 0x1DA01, 0x1DA02, 0x1DA03, + 0x1DA04, 0x1DA05, 0x1DA06, 0x1DA07, 0x1DA08, 0x1DA09, 0x1DA0A, 0x1DA0B, 0x1DA0C, 0x1DA0D, 0x1DA0E, 0x1DA0F, 0x1DA10, + 0x1DA11, 0x1DA12, 0x1DA13, 0x1DA14, 0x1DA15, 0x1DA16, 0x1DA17, 0x1DA18, 0x1DA19, 0x1DA1A, 0x1DA1B, 0x1DA1C, 0x1DA1D, + 0x1DA1E, 0x1DA1F, 0x1DA20, 0x1DA21, 0x1DA22, 0x1DA23, 0x1DA24, 0x1DA25, 0x1DA26, 0x1DA27, 0x1DA28, 0x1DA29, 0x1DA2A, + 0x1DA2B, 0x1DA2C, 0x1DA2D, 0x1DA2E, 0x1DA2F, 0x1DA30, 0x1DA31, 0x1DA32, 0x1DA33, 0x1DA34, 0x1DA35, 0x1DA36, 0x1DA3B, + 0x1DA3C, 0x1DA3D, 0x1DA3E, 0x1DA3F, 0x1DA40, 0x1DA41, 0x1DA42, 0x1DA43, 0x1DA44, 0x1DA45, 0x1DA46, 0x1DA47, 0x1DA48, + 0x1DA49, 0x1DA4A, 0x1DA4B, 0x1DA4C, 0x1DA4D, 0x1DA4E, 0x1DA4F, 0x1DA50, 0x1DA51, 0x1DA52, 0x1DA53, 0x1DA54, 0x1DA55, + 0x1DA56, 0x1DA57, 0x1DA58, 0x1DA59, 0x1DA5A, 0x1DA5B, 0x1DA5C, 0x1DA5D, 0x1DA5E, 0x1DA5F, 0x1DA60, 0x1DA61, 0x1DA62, + 0x1DA63, 0x1DA64, 0x1DA65, 0x1DA66, 0x1DA67, 0x1DA68, 0x1DA69, 0x1DA6A, 0x1DA6B, 0x1DA6C, 0x1DA75, 0x1DA84, 0x1DA9B, + 0x1DA9C, 0x1DA9D, 0x1DA9E, 0x1DA9F, 0x1DAA1, 0x1DAA2, 0x1DAA3, 0x1DAA4, 0x1DAA5, 0x1DAA6, 0x1DAA7, 0x1DAA8, 0x1DAA9, + 0x1DAAA, 0x1DAAB, 0x1DAAC, 0x1DAAD, 0x1DAAE, 0x1DAAF, 0x1E000, 0x1E001, 0x1E002, 0x1E003, 0x1E004, 0x1E005, 0x1E006, + 0x1E008, 0x1E009, 0x1E00A, 0x1E00B, 0x1E00C, 0x1E00D, 0x1E00E, 0x1E00F, 0x1E010, 0x1E011, 0x1E012, 0x1E013, 0x1E014, + 0x1E015, 0x1E016, 0x1E017, 0x1E018, 0x1E01B, 0x1E01C, 0x1E01D, 0x1E01E, 0x1E01F, 0x1E020, 0x1E021, 0x1E023, 0x1E024, + 0x1E026, 0x1E027, 0x1E028, 0x1E029, 0x1E02A, 0x1E08F, 0x1E130, 0x1E131, 0x1E132, 0x1E133, 0x1E134, 0x1E135, 0x1E136, + 0x1E2AE, 0x1E2EC, 0x1E2ED, 0x1E2EE, 0x1E2EF, 0x1E4EC, 0x1E4ED, 0x1E4EE, 0x1E4EF, 0x1E5EE, 0x1E5EF, 0x1E8D0, 0x1E8D1, + 0x1E8D2, 0x1E8D3, 0x1E8D4, 0x1E8D5, 0x1E8D6, 0x1E944, 0x1E945, 0x1E946, 0x1E947, 0x1E948, 0x1E949, 0x1E94A, 0xE0100, + 0xE0101, 0xE0102, 0xE0103, 0xE0104, 0xE0105, 0xE0106, 0xE0107, 0xE0108, 0xE0109, 0xE010A, 0xE010B, 0xE010C, 0xE010D, + 0xE010E, 0xE010F, 0xE0110, 0xE0111, 0xE0112, 0xE0113, 0xE0114, 0xE0115, 0xE0116, 0xE0117, 0xE0118, 0xE0119, 0xE011A, + 0xE011B, 0xE011C, 0xE011D, 0xE011E, 0xE011F, 0xE0120, 0xE0121, 0xE0122, 0xE0123, 0xE0124, 0xE0125, 0xE0126, 0xE0127, + 0xE0128, 0xE0129, 0xE012A, 0xE012B, 0xE012C, 0xE012D, 0xE012E, 0xE012F, 0xE0130, 0xE0131, 0xE0132, 0xE0133, 0xE0134, + 0xE0135, 0xE0136, 0xE0137, 0xE0138, 0xE0139, 0xE013A, 0xE013B, 0xE013C, 0xE013D, 0xE013E, 0xE013F, 0xE0140, 0xE0141, + 0xE0142, 0xE0143, 0xE0144, 0xE0145, 0xE0146, 0xE0147, 0xE0148, 0xE0149, 0xE014A, 0xE014B, 0xE014C, 0xE014D, 0xE014E, + 0xE014F, 0xE0150, 0xE0151, 0xE0152, 0xE0153, 0xE0154, 0xE0155, 0xE0156, 0xE0157, 0xE0158, 0xE0159, 0xE015A, 0xE015B, + 0xE015C, 0xE015D, 0xE015E, 0xE015F, 0xE0160, 0xE0161, 0xE0162, 0xE0163, 0xE0164, 0xE0165, 0xE0166, 0xE0167, 0xE0168, + 0xE0169, 0xE016A, 0xE016B, 0xE016C, 0xE016D, 0xE016E, 0xE016F, 0xE0170, 0xE0171, 0xE0172, 0xE0173, 0xE0174, 0xE0175, + 0xE0176, 0xE0177, 0xE0178, 0xE0179, 0xE017A, 0xE017B, 0xE017C, 0xE017D, 0xE017E, 0xE017F, 0xE0180, 0xE0181, 0xE0182, + 0xE0183, 0xE0184, 0xE0185, 0xE0186, 0xE0187, 0xE0188, 0xE0189, 0xE018A, 0xE018B, 0xE018C, 0xE018D, 0xE018E, 0xE018F, + 0xE0190, 0xE0191, 0xE0192, 0xE0193, 0xE0194, 0xE0195, 0xE0196, 0xE0197, 0xE0198, 0xE0199, 0xE019A, 0xE019B, 0xE019C, + 0xE019D, 0xE019E, 0xE019F, 0xE01A0, 0xE01A1, 0xE01A2, 0xE01A3, 0xE01A4, 0xE01A5, 0xE01A6, 0xE01A7, 0xE01A8, 0xE01A9, + 0xE01AA, 0xE01AB, 0xE01AC, 0xE01AD, 0xE01AE, 0xE01AF, 0xE01B0, 0xE01B1, 0xE01B2, 0xE01B3, 0xE01B4, 0xE01B5, 0xE01B6, + 0xE01B7, 0xE01B8, 0xE01B9, 0xE01BA, 0xE01BB, 0xE01BC, 0xE01BD, 0xE01BE, 0xE01BF, 0xE01C0, 0xE01C1, 0xE01C2, 0xE01C3, + 0xE01C4, 0xE01C5, 0xE01C6, 0xE01C7, 0xE01C8, 0xE01C9, 0xE01CA, 0xE01CB, 0xE01CC, 0xE01CD, 0xE01CE, 0xE01CF, 0xE01D0, + 0xE01D1, 0xE01D2, 0xE01D3, 0xE01D4, 0xE01D5, 0xE01D6, 0xE01D7, 0xE01D8, 0xE01D9, 0xE01DA, 0xE01DB, 0xE01DC, 0xE01DD, + 0xE01DE, 0xE01DF, 0xE01E0, 0xE01E1, 0xE01E2, 0xE01E3, 0xE01E4, 0xE01E5, 0xE01E6, 0xE01E7, 0xE01E8, 0xE01E9, 0xE01EA, + 0xE01EB, 0xE01EC, 0xE01ED, 0xE01EE, 0xE01EF + /* END: COMBINING CHAR TABLE */ +}; + +static const unsigned long combiningCharTableSize = sizeof(combiningCharTable) / sizeof(combiningCharTable[0]); + +static bool isCombiningChar(unsigned long cp) { + for (size_t i = 0; i < combiningCharTableSize; i++) { + auto code = combiningCharTable[i]; + if (code > cp) { + return false; + } + if (code == cp) { + return true; + } + } + return false; +} + +/* Get length of previous grapheme */ +static size_t defaultPrevCharLen(const char * buf, size_t /*buf_len*/, size_t pos, size_t * col_len) { + size_t end = pos; + while (pos > 0) { + size_t len = prevUtf8CodePointLen(buf, pos); + pos -= len; + int cp; + utf8BytesToCodePoint(buf + pos, len, &cp); + if (!isCombiningChar(cp)) { + if (col_len != NULL) { + *col_len = isWideChar(cp) ? 2 : 1; + } + return end - pos; + } + } + /* NOTREACHED */ + return 0; +} + +/* Get length of next grapheme */ +static size_t defaultNextCharLen(const char * buf, size_t buf_len, size_t pos, size_t * col_len) { + size_t beg = pos; + int cp; + size_t len = utf8BytesToCodePoint(buf + pos, buf_len - pos, &cp); + if (isCombiningChar(cp)) { + /* NOTREACHED */ + return 0; + } + if (col_len != NULL) { + *col_len = isWideChar(cp) ? 2 : 1; + } + pos += len; + while (pos < buf_len) { + int cp; + len = utf8BytesToCodePoint(buf + pos, buf_len - pos, &cp); + if (!isCombiningChar(cp)) { + return pos - beg; + } + pos += len; + } + return pos - beg; +} + +/* Read a Unicode from file. */ +static size_t defaultReadCode(int fd, char * buf, size_t buf_len, int * cp) { + if (buf_len < 1) { + return -1; + } + size_t nread = read(fd, &buf[0], 1); + if (nread <= 0) { + return nread; + } + + unsigned char byte = buf[0]; + if ((byte & 0x80) == 0) { + ; + } else if ((byte & 0xE0) == 0xC0) { + if (buf_len < 2) { + return -1; + } + nread = read(fd, &buf[1], 1); + if (nread <= 0) { + return nread; + } + } else if ((byte & 0xF0) == 0xE0) { + if (buf_len < 3) { + return -1; + } + nread = read(fd, &buf[1], 2); + if (nread <= 0) { + return nread; + } + } else if ((byte & 0xF8) == 0xF0) { + if (buf_len < 3) { + return -1; + } + nread = read(fd, &buf[1], 3); + if (nread <= 0) { + return nread; + } + } else { + return -1; + } + + return utf8BytesToCodePoint(buf, buf_len, cp); +} + +/* Set default encoding functions */ +static linenoisePrevCharLen * prevCharLen = defaultPrevCharLen; +static linenoiseNextCharLen * nextCharLen = defaultNextCharLen; +static linenoiseReadCode * readCode = defaultReadCode; + +/* Set used defined encoding functions */ +void linenoiseSetEncodingFunctions(linenoisePrevCharLen * prevCharLenFunc, linenoiseNextCharLen * nextCharLenFunc, + linenoiseReadCode * readCodeFunc) { + prevCharLen = prevCharLenFunc; + nextCharLen = nextCharLenFunc; + readCode = readCodeFunc; +} + /* ======================= Low level terminal handling ====================== */ /* Enable "mask mode". When it is enabled, instead of the input that @@ -408,6 +867,73 @@ static void refreshLineWithCompletion(struct linenoiseState *ls, linenoiseComple } } +enum ESC_TYPE { ESC_NULL = 0, ESC_DELETE, ESC_UP, ESC_DOWN, ESC_RIGHT, ESC_LEFT, ESC_HOME, ESC_END }; + +static ESC_TYPE readEscapeSequence(struct linenoiseState * l) { + /* Check if the file input has additional data. */ + struct pollfd pfd; + pfd.fd = l->ifd; + pfd.events = POLLIN; + + auto ret = poll(&pfd, 1, 1); // 1 millisecond timeout + if (ret <= 0) { // -1: error, 0: timeout + return ESC_NULL; + } + + /* Read the next two bytes representing the escape sequence. + * Use two calls to handle slow terminals returning the two + * chars at different times. */ + char seq[3]; + if (read(l->ifd, seq, 1) == -1) { + return ESC_NULL; + } + if (read(l->ifd, seq + 1, 1) == -1) { + return ESC_NULL; + } + + /* ESC [ sequences. */ + if (seq[0] == '[') { + if (seq[1] >= '0' && seq[1] <= '9') { + /* Extended escape, read additional byte. */ + if (read(l->ifd, seq + 2, 1) == -1) { + return ESC_NULL; + } + if (seq[2] == '~') { + switch (seq[1]) { + case '3': + return ESC_DELETE; + } + } + } else { + switch (seq[1]) { + case 'A': + return ESC_UP; + case 'B': + return ESC_DOWN; + case 'C': + return ESC_RIGHT; + case 'D': + return ESC_LEFT; + case 'H': + return ESC_HOME; + case 'F': + return ESC_END; + } + } + } + + /* ESC O sequences. */ + else if (seq[0] == 'O') { + switch (seq[1]) { + case 'H': + return ESC_HOME; + case 'F': + return ESC_END; + } + } + return ESC_NULL; +} + /* This is an helper function for linenoiseEdit*() and is called when the * user types the key in order to complete the string currently in the * input. @@ -418,11 +944,11 @@ static void refreshLineWithCompletion(struct linenoiseState *ls, linenoiseComple * If the function returns non-zero, the caller should handle the * returned value as a byte read from the standard input, and process * it as usually: this basically means that the function may return a byte - * read from the termianl but not processed. Otherwise, if zero is returned, + * read from the terminal but not processed. Otherwise, if zero is returned, * the input was consumed by the completeLine() function to navigate the * possible completions, and the caller should read for the next characters * from stdin. */ -static int completeLine(struct linenoiseState *ls, int keypressed) { +static int completeLine(struct linenoiseState * ls, int keypressed, ESC_TYPE esc_type) { linenoiseCompletions lc; int nwritten; char c = keypressed; @@ -432,31 +958,31 @@ static int completeLine(struct linenoiseState *ls, int keypressed) { linenoiseBeep(); ls->in_completion = 0; } else { - switch(c) { - case 9: /* tab */ - if (ls->in_completion == 0) { - ls->in_completion = 1; - ls->completion_idx = 0; - } else { - ls->completion_idx = (ls->completion_idx + 1) % (lc.len + 1); - if (ls->completion_idx == lc.len) linenoiseBeep(); + if (c == TAB) { + if (ls->in_completion == 0) { + ls->in_completion = 1; + ls->completion_idx = 0; + } else { + ls->completion_idx = (ls->completion_idx + 1) % (lc.len + 1); + if (ls->completion_idx == lc.len) { + linenoiseBeep(); } - c = 0; - break; - case 27: /* escape */ - /* Re-show original buffer */ - if (ls->completion_idx < lc.len) refreshLine(ls); - ls->in_completion = 0; - c = 0; - break; - default: - /* Update buffer and return */ - if (ls->completion_idx < lc.len) { - nwritten = snprintf(ls->buf, ls->buflen, "%s", lc.cvec[ls->completion_idx]); - ls->len = ls->pos = nwritten; - } - ls->in_completion = 0; - break; + } + c = 0; + } else if (c == ESC && esc_type == ESC_NULL) { + /* Re-show original buffer */ + if (ls->completion_idx < lc.len) { + refreshLine(ls); + } + ls->in_completion = 0; + c = 0; + } else { + /* Update buffer and return */ + if (ls->completion_idx < lc.len) { + nwritten = snprintf(ls->buf, ls->buflen, "%s", lc.cvec[ls->completion_idx]); + ls->len = ls->pos = nwritten; + } + ls->in_completion = 0; } /* Show completion or original buffer */ @@ -508,16 +1034,30 @@ void linenoiseAddCompletion(linenoiseCompletions *lc, const char *str) { lc->cvec[lc->len++] = copy.release(); } +/* Get column length from begining of buffer to current byte position */ +static size_t columnPos(const char * buf, size_t buf_len, size_t pos) { + size_t ret = 0; + size_t off = 0; + while (off < pos) { + size_t col_len; + size_t len = nextCharLen(buf, buf_len, off, &col_len); + off += len; + ret += col_len; + } + return ret; +} + /* Helper of refreshSingleLine() and refreshMultiLine() to show hints * to the right of the prompt. */ -static void refreshShowHints(std::string & ab, struct linenoiseState * l, int plen) { +static void refreshShowHints(std::string & ab, struct linenoiseState * l, int pcollen) { char seq[64]; - if (hintsCallback && plen+l->len < l->cols) { + size_t collen = pcollen + columnPos(l->buf, l->len, l->len); + if (hintsCallback && collen < l->cols) { int color = -1, bold = 0; const char *hint = hintsCallback(l->buf,&color,&bold); if (hint) { int hintlen = strlen(hint); - int hintmaxlen = l->cols-(plen+l->len); + int hintmaxlen = l->cols - collen; if (hintlen > hintmaxlen) hintlen = hintmaxlen; if (bold == 1 && color == -1) color = 37; if (color != -1 || bold != 0) @@ -535,6 +1075,50 @@ static void refreshShowHints(std::string & ab, struct linenoiseState * l, int pl } } +/* Check if text is an ANSI escape sequence */ +static int isAnsiEscape(const char * buf, size_t buf_len, size_t * len) { + if (buf_len > 2 && !memcmp("\033[", buf, 2)) { + size_t off = 2; + while (off < buf_len) { + switch (buf[off++]) { + case 'A': + case 'B': + case 'C': + case 'D': + case 'E': + case 'F': + case 'G': + case 'H': + case 'J': + case 'K': + case 'S': + case 'T': + case 'f': + case 'm': + *len = off; + return 1; + } + } + } + return 0; +} + +/* Get column length of prompt text */ +static size_t promptTextColumnLen(const char * prompt, size_t plen) { + char buf[LINENOISE_MAX_LINE]; + size_t buf_len = 0; + size_t off = 0; + while (off < plen) { + size_t len; + if (isAnsiEscape(prompt + off, plen - off, &len)) { + off += len; + continue; + } + buf[buf_len++] = prompt[off++]; + } + return columnPos(buf, buf_len, buf_len); +} + /* Single line low level line refresh. * * Rewrite the currently edited line accordingly to the buffer content, @@ -544,19 +1128,21 @@ static void refreshShowHints(std::string & ab, struct linenoiseState * l, int pl * prompt, just write it, or both. */ static void refreshSingleLine(struct linenoiseState *l, int flags) { char seq[64]; - size_t plen = strlen(l->prompt); + size_t pcollen = promptTextColumnLen(l->prompt, strlen(l->prompt)); int fd = l->ofd; char *buf = l->buf; size_t len = l->len; size_t pos = l->pos; std::string ab; - while((plen+pos) >= l->cols) { - buf++; - len--; - pos--; + + while ((pcollen + columnPos(buf, len, pos)) >= l->cols) { + int chlen = nextCharLen(buf, len, 0, NULL); + buf += chlen; + len -= chlen; + pos -= chlen; } - while (plen+len > l->cols) { - len--; + while (pcollen + columnPos(buf, len, len) > l->cols) { + len -= prevCharLen(buf, len, len, NULL); } /* Cursor to left edge */ @@ -574,7 +1160,7 @@ static void refreshSingleLine(struct linenoiseState *l, int flags) { ab.append(buf, len); } /* Show hits if any. */ - refreshShowHints(ab, l, plen); + refreshShowHints(ab, l, pcollen); } /* Erase to right */ @@ -582,13 +1168,43 @@ static void refreshSingleLine(struct linenoiseState *l, int flags) { ab.append(seq); if (flags & REFRESH_WRITE) { /* Move cursor to original position. */ - snprintf(seq,sizeof(seq),"\r\x1b[%dC", (int)(pos+plen)); + snprintf(seq, sizeof(seq), "\r\x1b[%dC", (int) (columnPos(buf, len, pos) + pcollen)); ab.append(seq); } (void) !write(fd, ab.c_str(), ab.size()); /* Can't recover from write error. */ } +/* Get column length from begining of buffer to current byte position for multiline mode*/ +static size_t columnPosForMultiLine(const char * buf, size_t buf_len, size_t pos, size_t cols, size_t ini_pos) { + size_t ret = 0; + size_t colwid = ini_pos; + + size_t off = 0; + while (off < buf_len) { + size_t col_len; + size_t len = nextCharLen(buf, buf_len, off, &col_len); + + int dif = (int) (colwid + col_len) - (int) cols; + if (dif > 0) { + ret += dif; + colwid = col_len; + } else if (dif == 0) { + colwid = 0; + } else { + colwid += col_len; + } + + if (off >= pos) { + break; + } + off += len; + ret += col_len; + } + + return ret; +} + /* Multi line low level line refresh. * * Rewrite the currently edited line accordingly to the buffer content, @@ -598,11 +1214,13 @@ static void refreshSingleLine(struct linenoiseState *l, int flags) { * prompt, just write it, or both. */ static void refreshMultiLine(struct linenoiseState *l, int flags) { char seq[64]; - int plen = strlen(l->prompt); - int rows = (plen+l->len+l->cols-1)/l->cols; /* rows used by current buf. */ - int rpos = (plen+l->oldpos+l->cols)/l->cols; /* cursor relative row. */ + size_t pcollen = promptTextColumnLen(l->prompt, strlen(l->prompt)); + int colpos = columnPosForMultiLine(l->buf, l->len, l->len, l->cols, pcollen); + int colpos2; /* cursor column position. */ + int rows = (pcollen + colpos + l->cols - 1) / l->cols; /* rows used by current buf. */ + int rpos = (pcollen + l->oldcolpos + l->cols) / l->cols; /* cursor relative row. */ int rpos2; /* rpos after refresh. */ - int col; /* colum position, zero-based. */ + int col; /* column position, zero-based. */ int old_rows = l->oldrows; int fd = l->ofd, j; std::string ab; @@ -611,15 +1229,13 @@ static void refreshMultiLine(struct linenoiseState *l, int flags) { /* First step: clear all the lines used before. To do so start by * going to the last row. */ if (flags & REFRESH_CLEAN) { - if (old_rows-rpos > 0) { - lndebug("go down %d", old_rows-rpos); + if (old_rows - rpos > 0) { snprintf(seq,64,"\x1b[%dB", old_rows-rpos); ab.append(seq); } /* Now for every row clear it, go up. */ - for (j = 0; j < old_rows-1; j++) { - lndebug("clear+up"); + for (j = 0; j < old_rows - 1; j++) { snprintf(seq,64,"\r\x1b[0K\x1b[1A"); ab.append(seq); } @@ -627,11 +1243,13 @@ static void refreshMultiLine(struct linenoiseState *l, int flags) { if (flags & REFRESH_ALL) { /* Clean the top line. */ - lndebug("clear"); snprintf(seq,64,"\r\x1b[0K"); ab.append(seq); } + /* Get column length to cursor position */ + colpos2 = columnPosForMultiLine(l->buf, l->len, l->pos, l->cols, pcollen); + if (flags & REFRESH_WRITE) { /* Write the prompt and the current buffer content */ ab.append(l->prompt); @@ -644,15 +1262,11 @@ static void refreshMultiLine(struct linenoiseState *l, int flags) { } /* Show hits if any. */ - refreshShowHints(ab, l, plen); + refreshShowHints(ab, l, pcollen); /* If we are at the very end of the screen with our prompt, we need to * emit a newline and move the prompt to the first column. */ - if (l->pos && - l->pos == l->len && - (l->pos+plen) % l->cols == 0) - { - lndebug(""); + if (l->pos && l->pos == l->len && (colpos2 + pcollen) % l->cols == 0) { ab.append("\n"); snprintf(seq,64,"\r"); ab.append(seq); @@ -661,19 +1275,16 @@ static void refreshMultiLine(struct linenoiseState *l, int flags) { } /* Move cursor to right position. */ - rpos2 = (plen+l->pos+l->cols)/l->cols; /* Current cursor relative row */ - lndebug("rpos2 %d", rpos2); + rpos2 = (pcollen + colpos2 + l->cols) / l->cols; /* Current cursor relative row */ - /* Go up till we reach the expected positon. */ - if (rows-rpos2 > 0) { - lndebug("go-up %d", rows-rpos2); + /* Go up till we reach the expected position. */ + if (rows - rpos2 > 0) { snprintf(seq,64,"\x1b[%dA", rows-rpos2); ab.append(seq); } /* Set column. */ - col = (plen+(int)l->pos) % (int)l->cols; - lndebug("set col %d", 1+col); + col = (pcollen + colpos2) % l->cols; if (col) snprintf(seq,64,"\r\x1b[%dC", col); else @@ -681,8 +1292,8 @@ static void refreshMultiLine(struct linenoiseState *l, int flags) { ab.append(seq); } - lndebug("\n"); - l->oldpos = l->pos; + l->oldcolpos = colpos2; + (void) !write(fd, ab.c_str(), ab.size()); /* Can't recover from write error. */ } @@ -720,26 +1331,36 @@ void linenoiseShow(struct linenoiseState *l) { /* Insert the character 'c' at cursor current position. * * On error writing to the terminal -1 is returned, otherwise 0. */ -static int linenoiseEditInsert(struct linenoiseState * l, char c) { - if (l->len < l->buflen) { +static int linenoiseEditInsert(struct linenoiseState * l, const char * cbuf, int clen) { + if (l->len + clen <= l->buflen) { if (l->len == l->pos) { - l->buf[l->pos] = c; - l->pos++; - l->len++; + memcpy(&l->buf[l->pos], cbuf, clen); + l->pos += clen; + l->len += clen; + ; l->buf[l->len] = '\0'; - if ((!mlmode && l->plen+l->len < l->cols && !hintsCallback)) { + if ((!mlmode && promptTextColumnLen(l->prompt, l->plen) + columnPos(l->buf, l->len, l->len) < l->cols && + !hintsCallback)) { /* Avoid a full update of the line in the * trivial case. */ - char d = (maskmode==1) ? '*' : c; - if (write(l->ofd,&d,1) == -1) return -1; + if (maskmode == 1) { + static const char d = '*'; + if (write(l->ofd, &d, 1) == -1) { + return -1; + } + } else { + if (write(l->ofd, cbuf, clen) == -1) { + return -1; + } + } } else { refreshLine(l); } } else { - memmove(l->buf+l->pos+1,l->buf+l->pos,l->len-l->pos); - l->buf[l->pos] = c; - l->len++; - l->pos++; + memmove(l->buf + l->pos + clen, l->buf + l->pos, l->len - l->pos); + memcpy(&l->buf[l->pos], cbuf, clen); + l->pos += clen; + l->len += clen; l->buf[l->len] = '\0'; refreshLine(l); } @@ -750,7 +1371,7 @@ static int linenoiseEditInsert(struct linenoiseState * l, char c) { /* Move cursor on the left. */ static void linenoiseEditMoveLeft(struct linenoiseState * l) { if (l->pos > 0) { - l->pos--; + l->pos -= prevCharLen(l->buf, l->len, l->pos, NULL); refreshLine(l); } } @@ -758,7 +1379,7 @@ static void linenoiseEditMoveLeft(struct linenoiseState * l) { /* Move cursor on the right. */ static void linenoiseEditMoveRight(struct linenoiseState * l) { if (l->pos != l->len) { - l->pos++; + l->pos += nextCharLen(l->buf, l->len, l->pos, NULL); refreshLine(l); } } @@ -810,8 +1431,9 @@ static void linenoiseEditHistoryNext(struct linenoiseState * l, int dir) { * position. Basically this is what happens with the "Delete" keyboard key. */ static void linenoiseEditDelete(struct linenoiseState * l) { if (l->len > 0 && l->pos < l->len) { - memmove(l->buf+l->pos,l->buf+l->pos+1,l->len-l->pos-1); - l->len--; + int chlen = nextCharLen(l->buf, l->len, l->pos, NULL); + memmove(l->buf + l->pos, l->buf + l->pos + chlen, l->len - l->pos - chlen); + l->len -= chlen; l->buf[l->len] = '\0'; refreshLine(l); } @@ -820,15 +1442,16 @@ static void linenoiseEditDelete(struct linenoiseState * l) { /* Backspace implementation. */ static void linenoiseEditBackspace(struct linenoiseState * l) { if (l->pos > 0 && l->len > 0) { - memmove(l->buf+l->pos-1,l->buf+l->pos,l->len-l->pos); - l->pos--; - l->len--; + int chlen = prevCharLen(l->buf, l->len, l->pos, NULL); + memmove(l->buf + l->pos - chlen, l->buf + l->pos, l->len - l->pos); + l->pos -= chlen; + l->len -= chlen; l->buf[l->len] = '\0'; refreshLine(l); } } -/* Delete the previosu word, maintaining the cursor at the start of the +/* Delete the previous word, maintaining the cursor at the start of the * current word. */ static void linenoiseEditDeletePrevWord(struct linenoiseState * l) { size_t old_pos = l->pos; @@ -855,7 +1478,7 @@ static void linenoiseEditDeletePrevWord(struct linenoiseState * l) { * each time there is some data arriving in the standard input. * * The user can also call linenoiseEditHide() and linenoiseEditShow() if it - * is required to show some input arriving asyncronously, without mixing + * is required to show some input arriving asynchronously, without mixing * it with the currently edited line. * * When linenoiseEditFeed() returns non-NULL, the user finished with the @@ -878,7 +1501,7 @@ int linenoiseEditStart(struct linenoiseState *l, int stdin_fd, int stdout_fd, ch l->buflen = buflen; l->prompt = prompt; l->plen = strlen(prompt); - l->oldpos = l->pos = 0; + l->oldcolpos = l->pos = 0; l->len = 0; /* Enter raw mode. */ @@ -890,7 +1513,7 @@ int linenoiseEditStart(struct linenoiseState *l, int stdin_fd, int stdout_fd, ch /* Buffer starts empty. */ l->buf[0] = '\0'; - l->buflen--; /* Make sure there is always space for the nulterm */ + l->buflen--; /* Make sure there is always space for the nullterm */ /* If stdin is not a tty, stop here with the initialization. We * will actually just read a line from standard input in blocking @@ -907,6 +1530,159 @@ int linenoiseEditStart(struct linenoiseState *l, int stdin_fd, int stdout_fd, ch const char* linenoiseEditMore = "If you see this, you are misusing the API: when linenoiseEditFeed() is called, if it returns linenoiseEditMore the user is yet editing the line. See the README file for more information."; +static const char * handleEnterKey(struct linenoiseState * l) { + --history_len; + free(history[history_len]); + if (mlmode) { + linenoiseEditMoveEnd(l); + } + if (hintsCallback) { + /* Force a refresh without hints to leave the previous + * line as the user typed it after a newline. */ + linenoiseHintsCallback * hc = hintsCallback; + hintsCallback = NULL; + refreshLine(l); + hintsCallback = hc; + } + + return strdup(l->buf); +} + +static const char * handleCtrlCKey() { + errno = EAGAIN; + return NULL; +} + +static const char * handleCtrlDKey(struct linenoiseState * l) { + if (l->len > 0) { + linenoiseEditDelete(l); + return linenoiseEditMore; + } + + --history_len; + free(history[history_len]); + errno = ENOENT; + return NULL; +} + +static void handleCtrlTKey(struct linenoiseState * l) { + if (l->pos > 0 && l->pos < l->len) { + auto prev_chlen = prevCharLen(l->buf, l->len, l->pos, NULL); + auto curr_chlen = nextCharLen(l->buf, l->len, l->pos, NULL); + + std::string prev_char(prev_chlen, 0); + memcpy(prev_char.data(), l->buf + l->pos - prev_chlen, prev_chlen); + memmove(l->buf + l->pos - prev_chlen, l->buf + l->pos, curr_chlen); + memmove(l->buf + l->pos - prev_chlen + curr_chlen, prev_char.data(), prev_chlen); + + l->pos = l->pos - prev_chlen + curr_chlen; + if (l->pos + prev_chlen != l->len) { + l->pos += prev_chlen; + } + + refreshLine(l); + } +} + +static void handleEscapeSequence(struct linenoiseState * l, int esc_type) { + switch (esc_type) { + case ESC_NULL: + break; + case ESC_DELETE: + linenoiseEditDelete(l); + break; + case ESC_UP: + linenoiseEditHistoryNext(l, LINENOISE_HISTORY_PREV); + break; + case ESC_DOWN: + linenoiseEditHistoryNext(l, LINENOISE_HISTORY_NEXT); + break; + case ESC_RIGHT: + linenoiseEditMoveRight(l); + break; + case ESC_LEFT: + linenoiseEditMoveLeft(l); + break; + case ESC_HOME: + linenoiseEditMoveHome(l); + break; + case ESC_END: + linenoiseEditMoveEnd(l); + break; + } +} + +static void handleCtrlUKey(struct linenoiseState * l) { + l->buf[0] = '\0'; + l->pos = l->len = 0; + refreshLine(l); +} + +static void handleCtrlKKey(struct linenoiseState * l) { + l->buf[l->pos] = '\0'; + l->len = l->pos; + refreshLine(l); +} + +static const char * processInputCharacter(struct linenoiseState * l, int c, char * cbuf, int nread, int esc_type) { + switch (c) { + case ENTER: + return handleEnterKey(l); + case CTRL_C: + return handleCtrlCKey(); + case BACKSPACE: + case CTRL_H: + linenoiseEditBackspace(l); + break; + case CTRL_D: /* ctrl-d, remove char at right of cursor, or if the + line is empty, act as end-of-file. */ + return handleCtrlDKey(l); + case CTRL_T: + handleCtrlTKey(l); + break; + case CTRL_B: + linenoiseEditMoveLeft(l); + break; + case CTRL_F: + linenoiseEditMoveRight(l); + break; + case CTRL_P: + linenoiseEditHistoryNext(l, LINENOISE_HISTORY_PREV); + break; + case CTRL_N: + linenoiseEditHistoryNext(l, LINENOISE_HISTORY_NEXT); + break; + case ESC: + handleEscapeSequence(l, esc_type); + break; + default: + if (linenoiseEditInsert(l, cbuf, nread)) { + return NULL; + } + break; + case CTRL_U: /* Ctrl+u, delete the whole line. */ + handleCtrlUKey(l); + break; + case CTRL_K: /* Ctrl+k, delete from current to end of line. */ + handleCtrlKKey(l); + break; + case CTRL_A: /* Ctrl+a, go to the start of the line */ + linenoiseEditMoveHome(l); + break; + case CTRL_E: /* ctrl+e, go to the end of the line */ + linenoiseEditMoveEnd(l); + break; + case CTRL_L: /* ctrl+l, clear screen */ + linenoiseClearScreen(); + refreshLine(l); + break; + case CTRL_W: /* ctrl+w, delete previous word */ + linenoiseEditDeletePrevWord(l); + break; + } + return linenoiseEditMore; +} + /* This function is part of the multiplexed API of linenoise, see the top * comment on linenoiseEditStart() for more information. Call this function * each time there is some data to read from the standard input file @@ -925,163 +1701,33 @@ const char* linenoiseEditMore = "If you see this, you are misusing the API: when * * Some other errno: I/O error. */ -const char *linenoiseEditFeed(struct linenoiseState *l) { - /* Not a TTY, pass control to line reading without character - * count limits. */ +const char * linenoiseEditFeed(struct linenoiseState * l) { + /* Not a TTY, pass control to line reading without character count + * limits. */ if (!isatty(l->ifd)) return linenoiseNoTTY(); - char c; + int c; int nread; - char seq[3]; + char cbuf[32]; - nread = read(l->ifd,&c,1); + nread = readCode(l->ifd, cbuf, sizeof(cbuf), &c); if (nread <= 0) return NULL; + auto esc_type = ESC_NULL; + if (c == ESC) { + esc_type = readEscapeSequence(l); + } + /* Only autocomplete when the callback is set. It returns < 0 when * there was an error reading from fd. Otherwise it will return the * character that should be handled next. */ if ((l->in_completion || c == 9) && completionCallback != NULL) { - c = completeLine(l,c); + c = completeLine(l, c, esc_type); /* Read next character when 0 */ if (c == 0) return linenoiseEditMore; } - switch(c) { - case ENTER: /* enter */ - history_len--; - free(history[history_len]); - if (mlmode) linenoiseEditMoveEnd(l); - if (hintsCallback) { - /* Force a refresh without hints to leave the previous - * line as the user typed it after a newline. */ - linenoiseHintsCallback *hc = hintsCallback; - hintsCallback = NULL; - refreshLine(l); - hintsCallback = hc; - } - return strdup(l->buf); - case CTRL_C: /* ctrl-c */ - errno = EAGAIN; - return NULL; - case BACKSPACE: /* backspace */ - case 8: /* ctrl-h */ - linenoiseEditBackspace(l); - break; - case CTRL_D: /* ctrl-d, remove char at right of cursor, or if the - line is empty, act as end-of-file. */ - if (l->len > 0) { - linenoiseEditDelete(l); - } else { - history_len--; - free(history[history_len]); - errno = ENOENT; - return NULL; - } - break; - case CTRL_T: /* ctrl-t, swaps current character with previous. */ - if (l->pos > 0 && l->pos < l->len) { - int aux = l->buf[l->pos-1]; - l->buf[l->pos-1] = l->buf[l->pos]; - l->buf[l->pos] = aux; - if (l->pos != l->len-1) l->pos++; - refreshLine(l); - } - break; - case CTRL_B: /* ctrl-b */ - linenoiseEditMoveLeft(l); - break; - case CTRL_F: /* ctrl-f */ - linenoiseEditMoveRight(l); - break; - case CTRL_P: /* ctrl-p */ - linenoiseEditHistoryNext(l, LINENOISE_HISTORY_PREV); - break; - case CTRL_N: /* ctrl-n */ - linenoiseEditHistoryNext(l, LINENOISE_HISTORY_NEXT); - break; - case ESC: /* escape sequence */ - /* Read the next two bytes representing the escape sequence. - * Use two calls to handle slow terminals returning the two - * chars at different times. */ - if (read(l->ifd,seq,1) == -1) break; - if (read(l->ifd,seq+1,1) == -1) break; - - /* ESC [ sequences. */ - if (seq[0] == '[') { - if (seq[1] >= '0' && seq[1] <= '9') { - /* Extended escape, read additional byte. */ - if (read(l->ifd,seq+2,1) == -1) break; - if (seq[2] == '~') { - switch(seq[1]) { - case '3': /* Delete key. */ - linenoiseEditDelete(l); - break; - } - } - } else { - switch(seq[1]) { - case 'A': /* Up */ - linenoiseEditHistoryNext(l, LINENOISE_HISTORY_PREV); - break; - case 'B': /* Down */ - linenoiseEditHistoryNext(l, LINENOISE_HISTORY_NEXT); - break; - case 'C': /* Right */ - linenoiseEditMoveRight(l); - break; - case 'D': /* Left */ - linenoiseEditMoveLeft(l); - break; - case 'H': /* Home */ - linenoiseEditMoveHome(l); - break; - case 'F': /* End*/ - linenoiseEditMoveEnd(l); - break; - } - } - } - - /* ESC O sequences. */ - else if (seq[0] == 'O') { - switch(seq[1]) { - case 'H': /* Home */ - linenoiseEditMoveHome(l); - break; - case 'F': /* End*/ - linenoiseEditMoveEnd(l); - break; - } - } - break; - default: - if (linenoiseEditInsert(l,c)) return NULL; - break; - case CTRL_U: /* Ctrl+u, delete the whole line. */ - l->buf[0] = '\0'; - l->pos = l->len = 0; - refreshLine(l); - break; - case CTRL_K: /* Ctrl+k, delete from current to end of line. */ - l->buf[l->pos] = '\0'; - l->len = l->pos; - refreshLine(l); - break; - case CTRL_A: /* Ctrl+a, go to the start of the line */ - linenoiseEditMoveHome(l); - break; - case CTRL_E: /* ctrl+e, go to the end of the line */ - linenoiseEditMoveEnd(l); - break; - case CTRL_L: /* ctrl+l, clear screen */ - linenoiseClearScreen(); - refreshLine(l); - break; - case CTRL_W: /* ctrl+w, delete previous word */ - linenoiseEditDeletePrevWord(l); - break; - } - return linenoiseEditMore; + return processInputCharacter(l, c, cbuf, nread, esc_type); } /* This is part of the multiplexed linenoise API. See linenoiseEditStart() @@ -1095,7 +1741,7 @@ void linenoiseEditStop(struct linenoiseState *l) { } /* This just implements a blocking loop for the multiplexed API. - * In many applications that are not event-drivern, we can just call + * In many applications that are not event-driven, we can just call * the blocking linenoise API, wait for the user to complete the editing * and return the buffer. */ static const char *linenoiseBlockingEdit(int stdin_fd, int stdout_fd, char *buf, size_t buflen, const char *prompt) @@ -1135,8 +1781,7 @@ void linenoisePrintKeyCodes(void) { quit[sizeof(quit)-1] = c; /* Insert current char on the right. */ if (memcmp(quit,"quit",sizeof(quit)) == 0) break; - printf("'%c' %02x (%d) (type quit to exit)\n", - isprint(c) ? c : '?', (int)c, (int)c); + printf("'%c' %02x (%d) (type quit to exit)\n", isprint((int) c) ? c : '?', (int) c, (int) c); printf("\r"); /* Go left edge manually, we are in raw mode. */ fflush(stdout); } diff --git a/examples/run/linenoise.cpp/linenoise.h b/examples/run/linenoise.cpp/linenoise.h index a14ec6c74..9823ca36d 100644 --- a/examples/run/linenoise.cpp/linenoise.h +++ b/examples/run/linenoise.cpp/linenoise.h @@ -47,27 +47,27 @@ extern "C" { #include /* For size_t. */ #include -extern const char *linenoiseEditMore; +extern const char * linenoiseEditMore; /* The linenoiseState structure represents the state during line editing. * We pass this state to functions implementing specific editing * functionalities. */ struct linenoiseState { - int in_completion; /* The user pressed TAB and we are now in completion + int in_completion; /* The user pressed TAB and we are now in completion * mode, so input is handled by completeLine(). */ - size_t completion_idx; /* Index of next completion to propose. */ - int ifd; /* Terminal stdin file descriptor. */ - int ofd; /* Terminal stdout file descriptor. */ - char *buf; /* Edited line buffer. */ - size_t buflen; /* Edited line buffer size. */ - const char *prompt; /* Prompt to display. */ - size_t plen; /* Prompt length. */ - size_t pos; /* Current cursor position. */ - size_t oldpos; /* Previous refresh cursor position. */ - size_t len; /* Current edited line length. */ - size_t cols; /* Number of columns in terminal. */ - size_t oldrows; /* Rows used by last refrehsed line (multiline mode) */ - int history_index; /* The history index we are currently editing. */ + size_t completion_idx; /* Index of next completion to propose. */ + int ifd; /* Terminal stdin file descriptor. */ + int ofd; /* Terminal stdout file descriptor. */ + char * buf; /* Edited line buffer. */ + size_t buflen; /* Edited line buffer size. */ + const char * prompt; /* Prompt to display. */ + size_t plen; /* Prompt length. */ + size_t pos; /* Current cursor position. */ + size_t oldcolpos; /* Previous refresh cursor column position. */ + size_t len; /* Current edited line length. */ + size_t cols; /* Number of columns in terminal. */ + size_t oldrows; /* Rows used by last refreshed line (multiline mode) */ + int history_index; /* The history index we are currently editing. */ }; struct linenoiseCompletions { @@ -89,19 +89,20 @@ struct linenoiseCompletions { }; /* Non blocking API. */ -int linenoiseEditStart(struct linenoiseState *l, int stdin_fd, int stdout_fd, char *buf, size_t buflen, const char *prompt); -const char *linenoiseEditFeed(struct linenoiseState *l); -void linenoiseEditStop(struct linenoiseState *l); -void linenoiseHide(struct linenoiseState *l); -void linenoiseShow(struct linenoiseState *l); +int linenoiseEditStart(struct linenoiseState * l, int stdin_fd, int stdout_fd, char * buf, size_t buflen, + const char * prompt); +const char * linenoiseEditFeed(struct linenoiseState * l); +void linenoiseEditStop(struct linenoiseState * l); +void linenoiseHide(struct linenoiseState * l); +void linenoiseShow(struct linenoiseState * l); /* Blocking API. */ -const char *linenoise(const char *prompt); -void linenoiseFree(void *ptr); +const char * linenoise(const char * prompt); +void linenoiseFree(void * ptr); /* Completion API. */ typedef void(linenoiseCompletionCallback)(const char *, linenoiseCompletions *); -typedef const char*(linenoiseHintsCallback)(const char *, int *color, int *bold); +typedef const char *(linenoiseHintsCallback) (const char *, int * color, int * bold); typedef void(linenoiseFreeHintsCallback)(const char *); void linenoiseSetCompletionCallback(linenoiseCompletionCallback *); void linenoiseSetHintsCallback(linenoiseHintsCallback *); @@ -109,10 +110,10 @@ void linenoiseSetFreeHintsCallback(linenoiseFreeHintsCallback *); void linenoiseAddCompletion(linenoiseCompletions *, const char *); /* History API. */ -int linenoiseHistoryAdd(const char *line); +int linenoiseHistoryAdd(const char * line); int linenoiseHistorySetMaxLen(int len); -int linenoiseHistorySave(const char *filename); -int linenoiseHistoryLoad(const char *filename); +int linenoiseHistorySave(const char * filename); +int linenoiseHistoryLoad(const char * filename); /* Other utilities. */ void linenoiseClearScreen(void); @@ -121,6 +122,14 @@ void linenoisePrintKeyCodes(void); void linenoiseMaskModeEnable(void); void linenoiseMaskModeDisable(void); +/* Encoding functions. */ +typedef size_t(linenoisePrevCharLen)(const char * buf, size_t buf_len, size_t pos, size_t * col_len); +typedef size_t(linenoiseNextCharLen)(const char * buf, size_t buf_len, size_t pos, size_t * col_len); +typedef size_t(linenoiseReadCode)(int fd, char * buf, size_t buf_len, int * c); + +void linenoiseSetEncodingFunctions(linenoisePrevCharLen * prevCharLenFunc, linenoiseNextCharLen * nextCharLenFunc, + linenoiseReadCode * readCodeFunc); + #ifdef __cplusplus } #endif diff --git a/examples/run/run.cpp b/examples/run/run.cpp index 9362da220..e63c2aac3 100644 --- a/examples/run/run.cpp +++ b/examples/run/run.cpp @@ -24,7 +24,7 @@ #include #include -#include "chat-template.hpp" +#include "chat.h" #include "common.h" #include "json.hpp" #include "linenoise.cpp/linenoise.h" @@ -38,24 +38,6 @@ } #endif -GGML_ATTRIBUTE_FORMAT(1, 2) -static std::string fmt(const char * fmt, ...) { - va_list ap; - va_list ap2; - va_start(ap, fmt); - va_copy(ap2, ap); - const int size = vsnprintf(NULL, 0, fmt, ap); - GGML_ASSERT(size >= 0 && size < INT_MAX); // NOLINT - std::string buf; - buf.resize(size); - const int size2 = vsnprintf(const_cast(buf.data()), buf.size() + 1, fmt, ap2); - GGML_ASSERT(size2 == size); - va_end(ap2); - va_end(ap); - - return buf; -} - GGML_ATTRIBUTE_FORMAT(1, 2) static int printe(const char * fmt, ...) { va_list args; @@ -79,6 +61,7 @@ class Opt { ctx_params = llama_context_default_params(); model_params = llama_model_default_params(); context_size_default = ctx_params.n_batch; + n_threads_default = ctx_params.n_threads; ngl_default = model_params.n_gpu_layers; common_params_sampling sampling; temperature_default = sampling.temp; @@ -104,6 +87,7 @@ class Opt { ctx_params.n_batch = context_size >= 0 ? context_size : context_size_default; ctx_params.n_ctx = ctx_params.n_batch; + ctx_params.n_threads = ctx_params.n_threads_batch = n_threads >= 0 ? n_threads : n_threads_default; model_params.n_gpu_layers = ngl >= 0 ? ngl : ngl_default; temperature = temperature >= 0 ? temperature : temperature_default; @@ -113,14 +97,15 @@ class Opt { llama_context_params ctx_params; llama_model_params model_params; std::string model_; + std::string chat_template_file; std::string user; bool use_jinja = false; - int context_size = -1, ngl = -1; + int context_size = -1, ngl = -1, n_threads = -1; float temperature = -1; bool verbose = false; private: - int context_size_default = -1, ngl_default = -1; + int context_size_default = -1, ngl_default = -1, n_threads_default = -1; float temperature_default = -1; bool help = false; @@ -148,48 +133,104 @@ class Opt { return 0; } + int handle_option_with_value(int argc, const char ** argv, int & i, std::string & option_value) { + if (i + 1 >= argc) { + return 1; + } + + option_value = argv[++i]; + + return 0; + } + + int parse_options_with_value(int argc, const char ** argv, int & i, bool & options_parsing) { + if (options_parsing && (strcmp(argv[i], "-c") == 0 || strcmp(argv[i], "--context-size") == 0)) { + if (handle_option_with_value(argc, argv, i, context_size) == 1) { + return 1; + } + } else if (options_parsing && + (strcmp(argv[i], "-n") == 0 || strcmp(argv[i], "-ngl") == 0 || strcmp(argv[i], "--ngl") == 0)) { + if (handle_option_with_value(argc, argv, i, ngl) == 1) { + return 1; + } + } else if (options_parsing && (strcmp(argv[i], "-t") == 0 || strcmp(argv[i], "--threads") == 0)) { + if (handle_option_with_value(argc, argv, i, n_threads) == 1) { + return 1; + } + } else if (options_parsing && strcmp(argv[i], "--temp") == 0) { + if (handle_option_with_value(argc, argv, i, temperature) == 1) { + return 1; + } + } else if (options_parsing && strcmp(argv[i], "--chat-template-file") == 0) { + if (handle_option_with_value(argc, argv, i, chat_template_file) == 1) { + return 1; + } + use_jinja = true; + } else { + return 2; + } + + return 0; + } + + int parse_options(const char ** argv, int & i, bool & options_parsing) { + if (options_parsing && (parse_flag(argv, i, "-v", "--verbose") || parse_flag(argv, i, "-v", "--log-verbose"))) { + verbose = true; + } else if (options_parsing && strcmp(argv[i], "--jinja") == 0) { + use_jinja = true; + } else if (options_parsing && parse_flag(argv, i, "-h", "--help")) { + help = true; + return 0; + } else if (options_parsing && strcmp(argv[i], "--") == 0) { + options_parsing = false; + } else { + return 2; + } + + return 0; + } + + int parse_positional_args(const char ** argv, int & i, int & positional_args_i) { + if (positional_args_i == 0) { + if (!argv[i][0] || argv[i][0] == '-') { + return 1; + } + + ++positional_args_i; + model_ = argv[i]; + } else if (positional_args_i == 1) { + ++positional_args_i; + user = argv[i]; + } else { + user += " " + std::string(argv[i]); + } + + return 0; + } + int parse(int argc, const char ** argv) { bool options_parsing = true; for (int i = 1, positional_args_i = 0; i < argc; ++i) { - if (options_parsing && (strcmp(argv[i], "-c") == 0 || strcmp(argv[i], "--context-size") == 0)) { - if (handle_option_with_value(argc, argv, i, context_size) == 1) { - return 1; - } - } else if (options_parsing && - (strcmp(argv[i], "-n") == 0 || strcmp(argv[i], "-ngl") == 0 || strcmp(argv[i], "--ngl") == 0)) { - if (handle_option_with_value(argc, argv, i, ngl) == 1) { - return 1; - } - } else if (options_parsing && strcmp(argv[i], "--temp") == 0) { - if (handle_option_with_value(argc, argv, i, temperature) == 1) { - return 1; - } - } else if (options_parsing && - (parse_flag(argv, i, "-v", "--verbose") || parse_flag(argv, i, "-v", "--log-verbose"))) { - verbose = true; - } else if (options_parsing && strcmp(argv[i], "--jinja") == 0) { - use_jinja = true; - } else if (options_parsing && parse_flag(argv, i, "-h", "--help")) { - help = true; - return 0; - } else if (options_parsing && strcmp(argv[i], "--") == 0) { - options_parsing = false; - } else if (positional_args_i == 0) { - if (!argv[i][0] || argv[i][0] == '-') { - return 1; - } + int ret = parse_options_with_value(argc, argv, i, options_parsing); + if (ret == 0) { + continue; + } else if (ret == 1) { + return ret; + } - ++positional_args_i; - model_ = argv[i]; - } else if (positional_args_i == 1) { - ++positional_args_i; - user = argv[i]; - } else { - user += " " + std::string(argv[i]); + ret = parse_options(argv, i, options_parsing); + if (ret == 0) { + continue; + } else if (ret == 1) { + return ret; + } + + if (parse_positional_args(argv, i, positional_args_i)) { + return 1; } } - if (model_.empty()){ + if (model_.empty()) { return 1; } @@ -207,10 +248,17 @@ class Opt { "Options:\n" " -c, --context-size \n" " Context size (default: %d)\n" + " --chat-template-file \n" + " Path to the file containing the chat template to use with the model.\n" + " Only supports jinja templates and implicitly sets the --jinja flag.\n" + " --jinja\n" + " Use jinja templating for the chat template of the model\n" " -n, -ngl, --ngl \n" " Number of GPU layers (default: %d)\n" " --temp \n" " Temperature (default: %.1f)\n" + " -t, --threads \n" + " Number of threads to use during generation (default: %d)\n" " -v, --verbose, --log-verbose\n" " Set verbosity level to infinity (i.e. log all messages, useful for debugging)\n" " -h, --help\n" @@ -239,7 +287,7 @@ class Opt { " llama-run file://some-file3.gguf\n" " llama-run --ngl 999 some-file4.gguf\n" " llama-run --ngl 999 some-file5.gguf Hello World\n", - context_size_default, ngl_default, temperature_default); + context_size_default, ngl_default, temperature_default, n_threads_default); } }; @@ -261,13 +309,12 @@ static int get_terminal_width() { #endif } -#ifdef LLAMA_USE_CURL class File { public: FILE * file = nullptr; FILE * open(const std::string & filename, const char * mode) { - file = fopen(filename.c_str(), mode); + file = ggml_fopen(filename.c_str(), mode); return file; } @@ -303,6 +350,20 @@ class File { return 0; } + std::string to_string() { + fseek(file, 0, SEEK_END); + const size_t size = ftell(file); + fseek(file, 0, SEEK_SET); + std::string out; + out.resize(size); + const size_t read_size = fread(&out[0], 1, size, file); + if (read_size != size) { + printe("Error reading file: %s", strerror(errno)); + } + + return out; + } + ~File() { if (fd >= 0) { # ifdef _WIN32 @@ -327,6 +388,7 @@ class File { # endif }; +#ifdef LLAMA_USE_CURL class HttpClient { public: int init(const std::string & url, const std::vector & headers, const std::string & output_file, @@ -445,11 +507,11 @@ class HttpClient { int secs = static_cast(seconds) % 60; if (hrs > 0) { - return fmt("%dh %02dm %02ds", hrs, mins, secs); + return string_format("%dh %02dm %02ds", hrs, mins, secs); } else if (mins > 0) { - return fmt("%dm %02ds", mins, secs); + return string_format("%dm %02ds", mins, secs); } else { - return fmt("%ds", secs); + return string_format("%ds", secs); } } @@ -464,7 +526,7 @@ class HttpClient { } } - return fmt("%.2f %s", dbl_size, suffix[i]); + return string_format("%.2f %s", dbl_size, suffix[i]); } static int update_progress(void * ptr, curl_off_t total_to_download, curl_off_t now_downloaded, curl_off_t, @@ -498,7 +560,9 @@ class HttpClient { return (now_downloaded_plus_file_size * 100) / total_to_download; } - static std::string generate_progress_prefix(curl_off_t percentage) { return fmt("%3ld%% |", static_cast(percentage)); } + static std::string generate_progress_prefix(curl_off_t percentage) { + return string_format("%3ld%% |", static_cast(percentage)); + } static double calculate_speed(curl_off_t now_downloaded, const std::chrono::steady_clock::time_point & start_time) { const auto now = std::chrono::steady_clock::now(); @@ -509,9 +573,9 @@ class HttpClient { static std::string generate_progress_suffix(curl_off_t now_downloaded_plus_file_size, curl_off_t total_to_download, double speed, double estimated_time) { const int width = 10; - return fmt("%*s/%*s%*s/s%*s", width, human_readable_size(now_downloaded_plus_file_size).c_str(), width, - human_readable_size(total_to_download).c_str(), width, human_readable_size(speed).c_str(), width, - human_readable_time(estimated_time).c_str()); + return string_format("%*s/%*s%*s/s%*s", width, human_readable_size(now_downloaded_plus_file_size).c_str(), + width, human_readable_size(total_to_download).c_str(), width, + human_readable_size(speed).c_str(), width, human_readable_time(estimated_time).c_str()); } static int calculate_progress_bar_width(const std::string & progress_prefix, const std::string & progress_suffix) { @@ -557,7 +621,7 @@ class LlamaData { llama_model_ptr model; llama_sampler_ptr sampler; llama_context_ptr context; - std::vector messages; + std::vector messages; // TODO: switch to common_chat_msg std::list msg_strs; std::vector fmtted; @@ -633,8 +697,10 @@ class LlamaData { std::vector headers = { "User-Agent: llama-cpp", "Accept: application/json" }; std::string url; + std::string model_endpoint = get_model_endpoint(); + if (pos == std::string::npos) { - auto [model_name, manifest_url] = extract_model_and_tag(model, "https://huggingface.co/v2/"); + auto [model_name, manifest_url] = extract_model_and_tag(model, model_endpoint + "v2/"); hfr = model_name; nlohmann::json manifest; @@ -649,7 +715,7 @@ class LlamaData { hff = model.substr(pos + 1); } - url = "https://huggingface.co/" + hfr + "/resolve/main/" + hff; + url = model_endpoint + hfr + "/resolve/main/" + hff; return download(url, bn, true, headers); } @@ -834,50 +900,29 @@ static void add_message(const char * role, const std::string & text, LlamaData & } // Function to apply the chat template and resize `formatted` if needed -static int apply_chat_template(const common_chat_template & tmpl, LlamaData & llama_data, const bool append, bool use_jinja) { - if (use_jinja) { - json messages = json::array(); - for (const auto & msg : llama_data.messages) { - messages.push_back({ - {"role", msg.role}, - {"content", msg.content}, - }); - } - try { - minja::chat_template_inputs tmpl_inputs; - tmpl_inputs.messages = messages; - tmpl_inputs.add_generation_prompt = append; - - minja::chat_template_options tmpl_opts; - tmpl_opts.use_bos_token = false; - tmpl_opts.use_eos_token = false; - - auto result = tmpl.apply(tmpl_inputs, tmpl_opts); - llama_data.fmtted.resize(result.size() + 1); - memcpy(llama_data.fmtted.data(), result.c_str(), result.size() + 1); - return result.size(); - } catch (const std::exception & e) { - printe("failed to render the chat template: %s\n", e.what()); - return -1; - } - } - int result = llama_chat_apply_template( - tmpl.source().c_str(), llama_data.messages.data(), llama_data.messages.size(), append, - append ? llama_data.fmtted.data() : nullptr, append ? llama_data.fmtted.size() : 0); - if (append && result > static_cast(llama_data.fmtted.size())) { - llama_data.fmtted.resize(result); - result = llama_chat_apply_template(tmpl.source().c_str(), llama_data.messages.data(), - llama_data.messages.size(), append, llama_data.fmtted.data(), - llama_data.fmtted.size()); +static int apply_chat_template(const struct common_chat_templates * tmpls, LlamaData & llama_data, const bool append, bool use_jinja) { + common_chat_templates_inputs inputs; + for (const auto & msg : llama_data.messages) { + common_chat_msg cmsg; + cmsg.role = msg.role; + cmsg.content = msg.content; + inputs.messages.push_back(cmsg); } + inputs.add_generation_prompt = append; + inputs.use_jinja = use_jinja; - return result; + auto chat_params = common_chat_templates_apply(tmpls, inputs); + // TODO: use other params for tool calls. + auto result = chat_params.prompt; + llama_data.fmtted.resize(result.size() + 1); + memcpy(llama_data.fmtted.data(), result.c_str(), result.size() + 1); + return result.size(); } // Function to tokenize the prompt static int tokenize_prompt(const llama_vocab * vocab, const std::string & prompt, std::vector & prompt_tokens, const LlamaData & llama_data) { - const bool is_first = llama_get_kv_cache_used_cells(llama_data.context.get()) == 0; + const bool is_first = llama_kv_self_used_cells(llama_data.context.get()) == 0; const int n_prompt_tokens = -llama_tokenize(vocab, prompt.c_str(), prompt.size(), NULL, 0, is_first, true); prompt_tokens.resize(n_prompt_tokens); @@ -893,7 +938,7 @@ static int tokenize_prompt(const llama_vocab * vocab, const std::string & prompt // Check if we have enough space in the context to evaluate this batch static int check_context_size(const llama_context_ptr & ctx, const llama_batch & batch) { const int n_ctx = llama_n_ctx(ctx.get()); - const int n_ctx_used = llama_get_kv_cache_used_cells(ctx.get()); + const int n_ctx_used = llama_kv_self_used_cells(ctx.get()); if (n_ctx_used + batch.n_tokens > n_ctx) { printf(LOG_COL_DEFAULT "\n"); printe("context size exceeded\n"); @@ -963,7 +1008,8 @@ static int generate(LlamaData & llama_data, const std::string & prompt, std::str } static int read_user_input(std::string & user_input) { - static const char * prompt_prefix = "> "; + static const char * prompt_prefix_env = std::getenv("LLAMA_PROMPT_PREFIX"); + static const char * prompt_prefix = prompt_prefix_env ? prompt_prefix_env : "> "; #ifdef WIN32 printf("\r" LOG_CLR_TO_EOL LOG_COL_DEFAULT "%s", prompt_prefix); @@ -1015,8 +1061,8 @@ static int generate_response(LlamaData & llama_data, const std::string & prompt, } // Helper function to apply the chat template and handle errors -static int apply_chat_template_with_error_handling(const common_chat_template & tmpl, LlamaData & llama_data, const bool append, int & output_length, bool use_jinja) { - const int new_len = apply_chat_template(tmpl, llama_data, append, use_jinja); +static int apply_chat_template_with_error_handling(const common_chat_templates * tmpls, LlamaData & llama_data, const bool append, int & output_length, bool use_jinja) { + const int new_len = apply_chat_template(tmpls, llama_data, append, use_jinja); if (new_len < 0) { printe("failed to apply the chat template\n"); return -1; @@ -1074,40 +1120,68 @@ static int get_user_input(std::string & user_input, const std::string & user) { return 0; } +// Reads a chat template file to be used +static std::string read_chat_template_file(const std::string & chat_template_file) { + File file; + if (!file.open(chat_template_file, "r")) { + printe("Error opening chat template file '%s': %s", chat_template_file.c_str(), strerror(errno)); + return ""; + } + + return file.to_string(); +} + +static int process_user_message(const Opt & opt, const std::string & user_input, LlamaData & llama_data, + const common_chat_templates_ptr & chat_templates, int & prev_len, + const bool stdout_a_terminal) { + add_message("user", opt.user.empty() ? user_input : opt.user, llama_data); + int new_len; + if (apply_chat_template_with_error_handling(chat_templates.get(), llama_data, true, new_len, opt.use_jinja) < 0) { + return 1; + } + + std::string prompt(llama_data.fmtted.begin() + prev_len, llama_data.fmtted.begin() + new_len); + std::string response; + if (generate_response(llama_data, prompt, response, stdout_a_terminal)) { + return 1; + } + + if (!opt.user.empty()) { + return 2; + } + + add_message("assistant", response, llama_data); + if (apply_chat_template_with_error_handling(chat_templates.get(), llama_data, false, prev_len, opt.use_jinja) < 0) { + return 1; + } + + return 0; +} + // Main chat loop function -static int chat_loop(LlamaData & llama_data, const std::string & user, bool use_jinja) { +static int chat_loop(LlamaData & llama_data, const Opt & opt) { int prev_len = 0; llama_data.fmtted.resize(llama_n_ctx(llama_data.context.get())); - auto chat_templates = common_chat_templates_from_model(llama_data.model.get(), ""); - GGML_ASSERT(chat_templates.template_default); + std::string chat_template; + if (!opt.chat_template_file.empty()) { + chat_template = read_chat_template_file(opt.chat_template_file); + } + + common_chat_templates_ptr chat_templates = common_chat_templates_init(llama_data.model.get(), chat_template); static const bool stdout_a_terminal = is_stdout_a_terminal(); while (true) { // Get user input std::string user_input; - if (get_user_input(user_input, user) == 1) { + if (get_user_input(user_input, opt.user) == 1) { return 0; } - add_message("user", user.empty() ? user_input : user, llama_data); - int new_len; - if (apply_chat_template_with_error_handling(*chat_templates.template_default, llama_data, true, new_len, use_jinja) < 0) { + const int ret = process_user_message(opt, user_input, llama_data, chat_templates, prev_len, stdout_a_terminal); + if (ret == 1) { return 1; - } - - std::string prompt(llama_data.fmtted.begin() + prev_len, llama_data.fmtted.begin() + new_len); - std::string response; - if (generate_response(llama_data, prompt, response, stdout_a_terminal)) { - return 1; - } - - if (!user.empty()) { + } else if (ret == 2) { break; } - - add_message("assistant", response, llama_data); - if (apply_chat_template_with_error_handling(*chat_templates.template_default, llama_data, false, prev_len, use_jinja) < 0) { - return 1; - } } return 0; @@ -1165,7 +1239,7 @@ int main(int argc, const char ** argv) { return 1; } - if (chat_loop(llama_data, opt.user, opt.use_jinja)) { + if (chat_loop(llama_data, opt)) { return 1; } diff --git a/examples/save-load-state/save-load-state.cpp b/examples/save-load-state/save-load-state.cpp index cf7cbd815..760ebbbf0 100644 --- a/examples/save-load-state/save-load-state.cpp +++ b/examples/save-load-state/save-load-state.cpp @@ -15,7 +15,7 @@ int main(int argc, char ** argv) { return 1; } - print_build_info(); + common_init(); if (params.n_predict < 0) { params.n_predict = 16; @@ -196,7 +196,7 @@ int main(int argc, char ** argv) { fprintf(stderr, "%s : seq 0 copied, %zd bytes\n", __func__, ncopy); // erase whole kv - llama_kv_cache_clear(ctx3); + llama_kv_self_clear(ctx3); fprintf(stderr, "%s : kv cache cleared\n", __func__); // restore kv into seq 1 diff --git a/examples/server/CMakeLists.txt b/examples/server/CMakeLists.txt index 1b7cc8c13..aee90388e 100644 --- a/examples/server/CMakeLists.txt +++ b/examples/server/CMakeLists.txt @@ -5,7 +5,7 @@ option(LLAMA_SERVER_SSL "Build SSL support for the server" OFF) include_directories(${CMAKE_CURRENT_SOURCE_DIR} ${CMAKE_CURRENT_BINARY_DIR}) if (MINGW) - # fix: https://github.com/ggerganov/llama.cpp/actions/runs/9651004652/job/26617901362?pr=8006 + # fix: https://github.com/ggml-org/llama.cpp/actions/runs/9651004652/job/26617901362?pr=8006 add_compile_definitions(_WIN32_WINNT=${GGML_WIN_VER}) endif() diff --git a/examples/server/README.md b/examples/server/README.md index d0b262f0e..a2a090326 100644 --- a/examples/server/README.md +++ b/examples/server/README.md @@ -7,14 +7,15 @@ Set of LLM REST APIs and a simple web front end to interact with llama.cpp. **Features:** * LLM inference of F16 and quantized models on GPU and CPU * [OpenAI API](https://github.com/openai/openai-openapi) compatible chat completions and embeddings routes - * Reranking endoint (WIP: https://github.com/ggerganov/llama.cpp/pull/9510) + * Reranking endoint (WIP: https://github.com/ggml-org/llama.cpp/pull/9510) * Parallel decoding with multi-user support * Continuous batching * Multimodal (wip) * Monitoring endpoints * Schema-constrained JSON response format + * [Function calling](../../docs/function-calling.md) / tool use for ~any model -The project is under active development, and we are [looking for feedback and contributors](https://github.com/ggerganov/llama.cpp/issues/4216). +The project is under active development, and we are [looking for feedback and contributors](https://github.com/ggml-org/llama.cpp/issues/4216). ## Usage @@ -65,7 +66,7 @@ The project is under active development, and we are [looking for feedback and co | `-np, --parallel N` | number of parallel sequences to decode (default: 1)
(env: LLAMA_ARG_N_PARALLEL) | | `--mlock` | force system to keep model in RAM rather than swapping or compressing
(env: LLAMA_ARG_MLOCK) | | `--no-mmap` | do not memory-map model (slower load but may reduce pageouts if not using mlock)
(env: LLAMA_ARG_NO_MMAP) | -| `--numa TYPE` | attempt optimizations that help on some NUMA systems
- distribute: spread execution evenly over all nodes
- isolate: only spawn threads on CPUs on the node that execution started on
- numactl: use the CPU map provided by numactl
if run without this previously, it is recommended to drop the system page cache before using this
see https://github.com/ggerganov/llama.cpp/issues/1437
(env: LLAMA_ARG_NUMA) | +| `--numa TYPE` | attempt optimizations that help on some NUMA systems
- distribute: spread execution evenly over all nodes
- isolate: only spawn threads on CPUs on the node that execution started on
- numactl: use the CPU map provided by numactl
if run without this previously, it is recommended to drop the system page cache before using this
see https://github.com/ggml-org/llama.cpp/issues/1437
(env: LLAMA_ARG_NUMA) | | `-dev, --device ` | comma-separated list of devices to use for offloading (none = don't offload)
use --list-devices to see a list of available devices
(env: LLAMA_ARG_DEVICE) | | `--list-devices` | print list of available devices and exit | | `-ngl, --gpu-layers, --n-gpu-layers N` | number of layers to store in VRAM
(env: LLAMA_ARG_N_GPU_LAYERS) | @@ -127,6 +128,7 @@ The project is under active development, and we are [looking for feedback and co | `--grammar-file FNAME` | file to read grammar from | | `-j, --json-schema SCHEMA` | JSON schema to constrain generations (https://json-schema.org/), e.g. `{}` for any JSON object
For schemas w/ external $refs, use --grammar + example/json_schema_to_grammar.py instead | | `--jinja` | Enable experimental Jinja templating engine (required for tool use) | +| `--reasoning-format FORMAT` | Controls extraction of model thinking traces and the format / field in which they are returned (default: `deepseek`; allowed values: `deepseek`, `none`; requires `--jinja`). `none` will leave thinking traces inline in `message.content` in a model-specific format, while `deepseek` will return them separately under `message.reasoning_content` | **Example-specific params** @@ -177,7 +179,7 @@ Example usage of docker compose with environment variables: ```yml services: llamacpp-server: - image: ghcr.io/ggerganov/llama.cpp:server + image: ghcr.io/ggml-org/llama.cpp:server ports: - 8080:8080 volumes: @@ -272,10 +274,10 @@ You can consume the endpoints with Postman or NodeJS with axios library. You can ### Docker ```bash -docker run -p 8080:8080 -v /path/to/models:/models ghcr.io/ggerganov/llama.cpp:server -m models/7B/ggml-model.gguf -c 512 --host 0.0.0.0 --port 8080 +docker run -p 8080:8080 -v /path/to/models:/models ghcr.io/ggml-org/llama.cpp:server -m models/7B/ggml-model.gguf -c 512 --host 0.0.0.0 --port 8080 # or, with CUDA: -docker run -p 8080:8080 -v /path/to/models:/models --gpus all ghcr.io/ggerganov/llama.cpp:server-cuda -m models/7B/ggml-model.gguf -c 512 --host 0.0.0.0 --port 8080 --n-gpu-layers 99 +docker run -p 8080:8080 -v /path/to/models:/models --gpus all ghcr.io/ggml-org/llama.cpp:server-cuda -m models/7B/ggml-model.gguf -c 512 --host 0.0.0.0 --port 8080 --n-gpu-layers 99 ``` ## Testing with CURL @@ -1065,7 +1067,7 @@ print(completion.choices[0].text) ### POST `/v1/chat/completions`: OpenAI-compatible Chat Completions API -Given a ChatML-formatted json description in `messages`, it returns the predicted completion. Both synchronous and streaming mode are supported, so scripted and interactive applications work fine. While no strong claims of compatibility with OpenAI API spec is being made, in our experience it suffices to support many apps. Only models with a [supported chat template](https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template) can be used optimally with this endpoint. By default, the ChatML template will be used. +Given a ChatML-formatted json description in `messages`, it returns the predicted completion. Both synchronous and streaming mode are supported, so scripted and interactive applications work fine. While no strong claims of compatibility with OpenAI API spec is being made, in our experience it suffices to support many apps. Only models with a [supported chat template](https://github.com/ggml-org/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template) can be used optimally with this endpoint. By default, the ChatML template will be used. *Options:* @@ -1119,181 +1121,9 @@ curl http://localhost:8080/v1/chat/completions \ *Tool call support* -[Function calling](https://platform.openai.com/docs/guides/function-calling) is supported for all models (see https://github.com/ggerganov/llama.cpp/pull/9639): +[OpenAI-style function calling](https://platform.openai.com/docs/guides/function-calling) is supported with the `--jinja` flag (and may require a `--chat-template-file` override to get the right tool-use compatible Jinja template; worst case, `--chat-template chatml` may also work). -- Requires `--jinja` flag -- Native tool call formats supported: - - Llama 3.1 / 3.3 (including builtin tools support - tool names for `wolfram_alpha`, `web_search` / `brave_search`, `code_interpreter`), Llama 3.2 - - Functionary v3.1 / v3.2 - - Hermes 2/3, Qwen 2.5 - - Mistral Nemo - - Firefunction v2 - - Command R7B - - DeepSeek R1 (WIP / seems reluctant to call any tools?) - -
- Show some common templates and which format handler they use - - | Template | Format | - |----------|--------| - | CohereForAI-c4ai-command-r-plus-default.jinja | generic tool calls | - | CohereForAI-c4ai-command-r-plus-rag.jinja | generic tool calls | - | CohereForAI-c4ai-command-r-plus-tool_use.jinja | generic tool calls | - | MiniMaxAI-MiniMax-Text-01.jinja | generic tool calls | - | NexaAIDev-Octopus-v2.jinja | generic tool calls | - | NousResearch-Hermes-2-Pro-Llama-3-8B-default.jinja | generic tool calls | - | NousResearch-Hermes-2-Pro-Llama-3-8B-tool_use.jinja | hermes 2 pro tool calls | - | NousResearch-Hermes-2-Pro-Mistral-7B-default.jinja | generic tool calls | - | NousResearch-Hermes-2-Pro-Mistral-7B-tool_use.jinja | hermes 2 pro tool calls | - | NousResearch-Hermes-3-Llama-3.1-70B-default.jinja | generic tool calls | - | NousResearch-Hermes-3-Llama-3.1-70B-tool_use.jinja | hermes 2 pro tool calls | - | OrionStarAI-Orion-14B-Chat.jinja | generic tool calls | - | Qwen-QwQ-32B-Preview.jinja | hermes 2 pro tool calls | - | Qwen-Qwen2-7B-Instruct.jinja | generic tool calls | - | Qwen-Qwen2-VL-7B-Instruct.jinja | generic tool calls | - | Qwen-Qwen2.5-7B-Instruct.jinja | hermes 2 pro tool calls | - | Qwen-Qwen2.5-Math-7B-Instruct.jinja | hermes 2 pro tool calls | - | TheBloke-FusionNet_34Bx2_MoE-AWQ.jinja | generic tool calls | - | abacusai-Fewshot-Metamath-OrcaVicuna-Mistral.jinja | generic tool calls | - | bofenghuang-vigogne-2-70b-chat.jinja | generic tool calls | - | databricks-dbrx-instruct.jinja | generic tool calls | - | deepseek-ai-DeepSeek-Coder-V2-Instruct.jinja | generic tool calls | - | deepseek-ai-DeepSeek-R1-Distill-Llama-8B.jinja | deepseek r1 tool calls | - | deepseek-ai-DeepSeek-R1-Distill-Qwen-32B.jinja | deepseek r1 tool calls | - | deepseek-ai-DeepSeek-R1-Distill-Qwen-7B.jinja | deepseek r1 tool calls | - | deepseek-ai-DeepSeek-V2.5.jinja | deepseek r1 tool calls | - | deepseek-ai-deepseek-coder-33b-instruct.jinja | generic tool calls | - | google-gemma-2-2b-it.jinja | generic tool calls | - | google-gemma-7b-it.jinja | generic tool calls | - | indischepartij-MiniCPM-3B-OpenHermes-2.5-v2.jinja | generic tool calls | - | mattshumer-Reflection-Llama-3.1-70B.jinja | generic tool calls | - | meetkai-functionary-medium-v3.2.jinja | functionary v3.2 tool calls | - | meta-llama-Llama-3.1-8B-Instruct.jinja | llama 3.x tool calls (w/ builtin tools) | - | meta-llama-Llama-3.2-3B-Instruct.jinja | llama 3.x tool calls | - | meta-llama-Llama-3.3-70B-Instruct.jinja | llama 3.x tool calls (w/ builtin tools) | - | meta-llama-Meta-Llama-3.1-8B-Instruct.jinja | llama 3.x tool calls (w/ builtin tools) | - | microsoft-Phi-3-medium-4k-instruct.jinja | generic tool calls | - | microsoft-Phi-3-mini-4k-instruct.jinja | generic tool calls | - | microsoft-Phi-3-small-8k-instruct.jinja | generic tool calls | - | microsoft-Phi-3.5-mini-instruct.jinja | generic tool calls | - | microsoft-Phi-3.5-vision-instruct.jinja | generic tool calls | - | mistralai-Mistral-7B-Instruct-v0.2.jinja | generic tool calls | - | mistralai-Mistral-Large-Instruct-2407.jinja | mistral nemo tool calls | - | mistralai-Mistral-Large-Instruct-2411.jinja | generic tool calls | - | mistralai-Mistral-Nemo-Instruct-2407.jinja | mistral nemo tool calls | - | mistralai-Mixtral-8x7B-Instruct-v0.1.jinja | generic tool calls | - | mlabonne-AlphaMonarch-7B.jinja | generic tool calls | - | nvidia-Llama-3.1-Nemotron-70B-Instruct-HF.jinja | llama 3.x tool calls (w/ builtin tools) | - | openchat-openchat-3.5-0106.jinja | generic tool calls | - | teknium-OpenHermes-2.5-Mistral-7B.jinja | generic tool calls | - - This table can be generated with: - - ```bash - ./build/bin/test-chat ../minja/build/tests/*.jinja 2>/dev/null - -
- -- Generic tool call is supported when the template isn't recognized by native format handlers (you'll see `Chat format: Generic` in the logs). - - Use `--chat-template-file` to override the template when appropriate (see examples below) - - Generic support may consume more tokens and be less efficient than a model's native format. - -- Run with: - - ```shell - # Native support: - llama-server --jinja -fa -hf bartowski/Qwen2.5-7B-Instruct-GGUF:Q4_K_M - llama-server --jinja -fa -hf bartowski/Mistral-Nemo-Instruct-2407-GGUF:Q6_K_L - llama-server --jinja -fa -hf bartowski/functionary-small-v3.2-GGUF:Q4_K_M - llama-server --jinja -fa -hf bartowski/Llama-3.3-70B-Instruct-GGUF:Q4_K_M - - # Native support requires the right template for these GGUFs: - - llama-server --jinja -fa -hf bartowski/Hermes-2-Pro-Llama-3-8B-GGUF:Q4_K_M \ - --chat-template-file <( python scripts/get_chat_template.py NousResearch/Hermes-2-Pro-Llama-3-8B tool_use ) - - llama-server --jinja -fa -hf bartowski/Hermes-3-Llama-3.1-8B-GGUF:Q4_K_M \ - --chat-template-file <( python scripts/get_chat_template.py NousResearch/Hermes-3-Llama-3.1-8B tool_use ) - - llama-server --jinja -fa -hf bartowski/firefunction-v2-GGUF -hff firefunction-v2-IQ1_M.gguf \ - --chat-template-file <( python scripts/get_chat_template.py fireworks-ai/llama-3-firefunction-v2 tool_use ) - - llama-server --jinja -fa -hf bartowski/c4ai-command-r7b-12-2024-GGUF:Q6_K_L \ - --chat-template-file <( python scripts/get_chat_template.py CohereForAI/c4ai-command-r7b-12-2024 tool_use ) - - # Generic format support - llama-server --jinja -fa -hf bartowski/phi-4-GGUF:Q4_0 - llama-server --jinja -fa -hf bartowski/gemma-2-2b-it-GGUF:Q8_0 - llama-server --jinja -fa -hf bartowski/c4ai-command-r-v01-GGUF:Q2_K - ``` - -- Test in CLI: - - ```bash - curl http://localhost:8080/v1/chat/completions -d '{ - "model": "gpt-3.5-turbo", - "tools": [ - { - "type":"function", - "function":{ - "name":"get_current_weather", - "description":"Get the current weather in a given location", - "parameters":{ - "type":"object", - "properties":{ - "location":{ - "type":"string", - "description":"The city and state, e.g. San Francisco, CA" - } - }, - "required":["location"] - } - } - } - ], - "messages": [ - { - "role": "user", - "content": "What is the weather like in Istanbul?." - } - ] - }' - ``` - -
- Show output - - ```json - { - "choices": [ - { - "finish_reason": "tool", - "index": 0, - "message": { - "content": null, - "tool_calls": [ - { - "name": "python", - "arguments": "{\"code\":\" \\nprint(\\\"Hello, World!\\\")\"}" - } - ], - "role": "assistant" - } - } - ], - "created": 1727287211, - "model": "gpt-3.5-turbo", - "object": "chat.completion", - "usage": { - "completion_tokens": 16, - "prompt_tokens": 44, - "total_tokens": 60 - }, - "id": "chatcmpl-Htbgh9feMmGM0LEH2hmQvwsCxq3c6Ni8" - } - ``` - -
+**See our [Function calling](../../docs/function-calling.md) docs** for more details, supported native tool call styles (generic tool call style is used as fallback) / examples of use. ### POST `/v1/embeddings`: OpenAI-compatible embeddings API @@ -1398,7 +1228,7 @@ Apart from error types supported by OAI, we also have custom types that are spec ### Legacy completion web UI -A new chat-based UI has replaced the old completion-based since [this PR](https://github.com/ggerganov/llama.cpp/pull/10175). If you want to use the old completion, start the server with `--path ./examples/server/public_legacy` +A new chat-based UI has replaced the old completion-based since [this PR](https://github.com/ggml-org/llama.cpp/pull/10175). If you want to use the old completion, start the server with `--path ./examples/server/public_legacy` For example: diff --git a/examples/server/httplib.h b/examples/server/httplib.h index c2f12dd2a..0f981dc89 100644 --- a/examples/server/httplib.h +++ b/examples/server/httplib.h @@ -1,14 +1,14 @@ // // httplib.h // -// Copyright (c) 2024 Yuji Hirose. All rights reserved. +// Copyright (c) 2025 Yuji Hirose. All rights reserved. // MIT License // #ifndef CPPHTTPLIB_HTTPLIB_H #define CPPHTTPLIB_HTTPLIB_H -#define CPPHTTPLIB_VERSION "0.18.5" +#define CPPHTTPLIB_VERSION "0.20.0" /* * Configuration @@ -66,6 +66,10 @@ #define CPPHTTPLIB_CLIENT_WRITE_TIMEOUT_USECOND 0 #endif +#ifndef CPPHTTPLIB_CLIENT_MAX_TIMEOUT_MSECOND +#define CPPHTTPLIB_CLIENT_MAX_TIMEOUT_MSECOND 0 +#endif + #ifndef CPPHTTPLIB_IDLE_INTERVAL_SECOND #define CPPHTTPLIB_IDLE_INTERVAL_SECOND 0 #endif @@ -184,14 +188,16 @@ using ssize_t = long; #include #include +// afunix.h uses types declared in winsock2.h, so has to be included after it. +#include + #ifndef WSA_FLAG_NO_HANDLE_INHERIT #define WSA_FLAG_NO_HANDLE_INHERIT 0x80 #endif +using nfds_t = unsigned long; using socket_t = SOCKET; -#ifdef CPPHTTPLIB_USE_POLL -#define poll(fds, nfds, timeout) WSAPoll(fds, nfds, timeout) -#endif +using socklen_t = int; #else // not _WIN32 @@ -211,14 +217,11 @@ using socket_t = SOCKET; #ifdef __linux__ #include #endif -#include -#ifdef CPPHTTPLIB_USE_POLL -#include -#endif #include +#include +#include #include #include -#include #include #include #include @@ -240,7 +243,6 @@ using socket_t = int; #include #include #include -#include #include #include #include @@ -313,6 +315,10 @@ using socket_t = int; #include #endif +#ifdef CPPHTTPLIB_ZSTD_SUPPORT +#include +#endif + /* * Declaration */ @@ -428,6 +434,15 @@ private: } // namespace detail +enum SSLVerifierResponse { + // no decision has been made, use the built-in certificate verifier + NoDecisionMade, + // connection certificate is verified and accepted + CertificateAccepted, + // connection certificate was processed but is rejected + CertificateRejected +}; + enum StatusCode { // Information responses Continue_100 = 100, @@ -662,6 +677,8 @@ struct Request { ContentProvider content_provider_; bool is_chunked_content_provider_ = false; size_t authorization_count_ = 0; + std::chrono::time_point start_time_ = + (std::chrono::steady_clock::time_point::min)(); }; struct Response { @@ -727,7 +744,8 @@ public: virtual ~Stream() = default; virtual bool is_readable() const = 0; - virtual bool is_writable() const = 0; + virtual bool wait_readable() const = 0; + virtual bool wait_writable() const = 0; virtual ssize_t read(char *ptr, size_t size) = 0; virtual ssize_t write(const char *ptr, size_t size) = 0; @@ -735,6 +753,8 @@ public: virtual void get_local_ip_and_port(std::string &ip, int &port) const = 0; virtual socket_t socket() const = 0; + virtual time_t duration() const = 0; + ssize_t write(const char *ptr); ssize_t write(const std::string &s); }; @@ -838,6 +858,16 @@ using Logger = std::function; using SocketOptions = std::function; +namespace detail { + +bool set_socket_opt_impl(socket_t sock, int level, int optname, + const void *optval, socklen_t optlen); +bool set_socket_opt(socket_t sock, int level, int optname, int opt); +bool set_socket_opt_time(socket_t sock, int level, int optname, time_t sec, + time_t usec); + +} // namespace detail + void default_socket_options(socket_t sock); const char *status_message(int status); @@ -858,7 +888,7 @@ public: * Captures parameters in request path and stores them in Request::path_params * * Capture name is a substring of a pattern from : to /. - * The rest of the pattern is matched agains the request path directly + * The rest of the pattern is matched against the request path directly * Parameters are captured starting from the next character after * the end of the last matched static pattern fragment until the next /. * @@ -1088,7 +1118,7 @@ private: virtual bool process_and_close_socket(socket_t sock); std::atomic is_running_{false}; - std::atomic is_decommisioned{false}; + std::atomic is_decommissioned{false}; struct MountPointEntry { std::string mount_point; @@ -1421,6 +1451,10 @@ public: template void set_write_timeout(const std::chrono::duration &duration); + void set_max_timeout(time_t msec); + template + void set_max_timeout(const std::chrono::duration &duration); + void set_basic_auth(const std::string &username, const std::string &password); void set_bearer_token_auth(const std::string &token); #ifdef CPPHTTPLIB_OPENSSL_SUPPORT @@ -1458,7 +1492,8 @@ public: #ifdef CPPHTTPLIB_OPENSSL_SUPPORT void enable_server_certificate_verification(bool enabled); void enable_server_hostname_verification(bool enabled); - void set_server_certificate_verifier(std::function verifier); + void set_server_certificate_verifier( + std::function verifier); #endif void set_logger(Logger logger); @@ -1529,6 +1564,7 @@ protected: time_t read_timeout_usec_ = CPPHTTPLIB_CLIENT_READ_TIMEOUT_USECOND; time_t write_timeout_sec_ = CPPHTTPLIB_CLIENT_WRITE_TIMEOUT_SECOND; time_t write_timeout_usec_ = CPPHTTPLIB_CLIENT_WRITE_TIMEOUT_USECOND; + time_t max_timeout_msec_ = CPPHTTPLIB_CLIENT_MAX_TIMEOUT_MSECOND; std::string basic_auth_username_; std::string basic_auth_password_; @@ -1574,7 +1610,7 @@ protected: #ifdef CPPHTTPLIB_OPENSSL_SUPPORT bool server_certificate_verification_ = true; bool server_hostname_verification_ = true; - std::function server_certificate_verifier_; + std::function server_certificate_verifier_; #endif Logger logger_; @@ -1583,9 +1619,6 @@ private: bool send_(Request &req, Response &res, Error &error); Result send_(Request &&req); -#ifdef CPPHTTPLIB_OPENSSL_SUPPORT - bool is_ssl_peer_could_be_closed(SSL *ssl) const; -#endif socket_t create_client_socket(Error &error) const; bool read_response_line(Stream &strm, const Request &req, Response &res) const; @@ -1611,8 +1644,10 @@ private: std::string adjust_host_string(const std::string &host) const; - virtual bool process_socket(const Socket &socket, - std::function callback); + virtual bool + process_socket(const Socket &socket, + std::chrono::time_point start_time, + std::function callback); virtual bool is_ssl() const; }; @@ -1854,6 +1889,10 @@ public: template void set_write_timeout(const std::chrono::duration &duration); + void set_max_timeout(time_t msec); + template + void set_max_timeout(const std::chrono::duration &duration); + void set_basic_auth(const std::string &username, const std::string &password); void set_bearer_token_auth(const std::string &token); #ifdef CPPHTTPLIB_OPENSSL_SUPPORT @@ -1884,7 +1923,8 @@ public: #ifdef CPPHTTPLIB_OPENSSL_SUPPORT void enable_server_certificate_verification(bool enabled); void enable_server_hostname_verification(bool enabled); - void set_server_certificate_verifier(std::function verifier); + void set_server_certificate_verifier( + std::function verifier); #endif void set_logger(Logger logger); @@ -1971,12 +2011,16 @@ private: void shutdown_ssl(Socket &socket, bool shutdown_gracefully) override; void shutdown_ssl_impl(Socket &socket, bool shutdown_gracefully); - bool process_socket(const Socket &socket, - std::function callback) override; + bool + process_socket(const Socket &socket, + std::chrono::time_point start_time, + std::function callback) override; bool is_ssl() const override; - bool connect_with_proxy(Socket &sock, Response &res, bool &success, - Error &error); + bool connect_with_proxy( + Socket &sock, + std::chrono::time_point start_time, + Response &res, bool &success, Error &error); bool initialize_ssl(Socket &socket, Error &error); bool load_certs(); @@ -2013,6 +2057,10 @@ inline void duration_to_sec_and_usec(const T &duration, U callback) { callback(static_cast(sec), static_cast(usec)); } +template inline constexpr size_t str_len(const char (&)[N]) { + return N - 1; +} + inline bool is_numeric(const std::string &str) { return !str.empty() && std::all_of(str.begin(), str.end(), ::isdigit); } @@ -2053,20 +2101,45 @@ inline uint64_t Response::get_header_value_u64(const std::string &key, return detail::get_header_value_u64(headers, key, def, id); } -inline void default_socket_options(socket_t sock) { - int opt = 1; +namespace detail { + +inline bool set_socket_opt_impl(socket_t sock, int level, int optname, + const void *optval, socklen_t optlen) { + return setsockopt(sock, level, optname, #ifdef _WIN32 - setsockopt(sock, SOL_SOCKET, SO_REUSEADDR, - reinterpret_cast(&opt), sizeof(opt)); + reinterpret_cast(optval), #else + optval, +#endif + optlen) == 0; +} + +inline bool set_socket_opt(socket_t sock, int level, int optname, int optval) { + return set_socket_opt_impl(sock, level, optname, &optval, sizeof(optval)); +} + +inline bool set_socket_opt_time(socket_t sock, int level, int optname, + time_t sec, time_t usec) { +#ifdef _WIN32 + auto timeout = static_cast(sec * 1000 + usec / 1000); +#else + timeval timeout; + timeout.tv_sec = static_cast(sec); + timeout.tv_usec = static_cast(usec); +#endif + return set_socket_opt_impl(sock, level, optname, &timeout, sizeof(timeout)); +} + +} // namespace detail + +inline void default_socket_options(socket_t sock) { + detail::set_socket_opt(sock, SOL_SOCKET, #ifdef SO_REUSEPORT - setsockopt(sock, SOL_SOCKET, SO_REUSEPORT, - reinterpret_cast(&opt), sizeof(opt)); + SO_REUSEPORT, #else - setsockopt(sock, SOL_SOCKET, SO_REUSEADDR, - reinterpret_cast(&opt), sizeof(opt)); -#endif + SO_REUSEADDR, #endif + 1); } inline const char *status_message(int status) { @@ -2147,9 +2220,9 @@ inline const char *status_message(int status) { inline std::string get_bearer_token_auth(const Request &req) { if (req.has_header("Authorization")) { - static std::string BearerHeaderPrefix = "Bearer "; + constexpr auto bearer_header_prefix_len = detail::str_len("Bearer "); return req.get_header_value("Authorization") - .substr(BearerHeaderPrefix.length()); + .substr(bearer_header_prefix_len); } return ""; } @@ -2238,6 +2311,14 @@ inline void ClientImpl::set_write_timeout( duration, [&](time_t sec, time_t usec) { set_write_timeout(sec, usec); }); } +template +inline void ClientImpl::set_max_timeout( + const std::chrono::duration &duration) { + auto msec = + std::chrono::duration_cast(duration).count(); + set_max_timeout(msec); +} + template inline void Client::set_connection_timeout( const std::chrono::duration &duration) { @@ -2256,6 +2337,12 @@ Client::set_write_timeout(const std::chrono::duration &duration) { cli_->set_write_timeout(duration); } +template +inline void +Client::set_max_timeout(const std::chrono::duration &duration) { + cli_->set_max_timeout(duration); +} + /* * Forward declarations and types that will be part of the .h file if split into * .h + .cc. @@ -2310,8 +2397,6 @@ std::string encode_query_param(const std::string &value); std::string decode_url(const std::string &s, bool convert_plus_to_space); -void read_file(const std::string &path, std::string &out); - std::string trim_copy(const std::string &s); void divide( @@ -2330,10 +2415,12 @@ void split(const char *b, const char *e, char d, void split(const char *b, const char *e, char d, size_t m, std::function fn); -bool process_client_socket(socket_t sock, time_t read_timeout_sec, - time_t read_timeout_usec, time_t write_timeout_sec, - time_t write_timeout_usec, - std::function callback); +bool process_client_socket( + socket_t sock, time_t read_timeout_sec, time_t read_timeout_usec, + time_t write_timeout_sec, time_t write_timeout_usec, + time_t max_timeout_msec, + std::chrono::time_point start_time, + std::function callback); socket_t create_client_socket(const std::string &host, const std::string &ip, int port, int address_family, bool tcp_nodelay, @@ -2365,7 +2452,7 @@ ssize_t send_socket(socket_t sock, const void *ptr, size_t size, int flags); ssize_t read_socket(socket_t sock, void *ptr, size_t size, int flags); -enum class EncodingType { None = 0, Gzip, Brotli }; +enum class EncodingType { None = 0, Gzip, Brotli, Zstd }; EncodingType encoding_type(const Request &req, const Response &res); @@ -2375,12 +2462,14 @@ public: ~BufferStream() override = default; bool is_readable() const override; - bool is_writable() const override; + bool wait_readable() const override; + bool wait_writable() const override; ssize_t read(char *ptr, size_t size) override; ssize_t write(const char *ptr, size_t size) override; void get_remote_ip_and_port(std::string &ip, int &port) const override; void get_local_ip_and_port(std::string &ip, int &port) const override; socket_t socket() const override; + time_t duration() const override; const std::string &get_buffer() const; @@ -2476,6 +2565,34 @@ private: }; #endif +#ifdef CPPHTTPLIB_ZSTD_SUPPORT +class zstd_compressor : public compressor { +public: + zstd_compressor(); + ~zstd_compressor(); + + bool compress(const char *data, size_t data_length, bool last, + Callback callback) override; + +private: + ZSTD_CCtx *ctx_ = nullptr; +}; + +class zstd_decompressor : public decompressor { +public: + zstd_decompressor(); + ~zstd_decompressor(); + + bool is_valid() const override; + + bool decompress(const char *data, size_t data_length, + Callback callback) override; + +private: + ZSTD_DCtx *ctx_ = nullptr; +}; +#endif + // NOTE: until the read size reaches `fixed_buffer_size`, use `fixed_buffer` // to store data. The call can set memory on stack for performance. class stream_line_reader { @@ -2494,7 +2611,7 @@ private: char *fixed_buffer_; const size_t fixed_buffer_size_; size_t fixed_buffer_used_size_ = 0; - std::string glowable_buffer_; + std::string growable_buffer_; }; class mmap { @@ -2547,7 +2664,7 @@ inline bool is_obs_text(char c) { return 128 <= static_cast(c); } inline bool is_field_vchar(char c) { return is_vchar(c) || is_obs_text(c); } inline bool is_field_content(const std::string &s) { - if (s.empty()) { return false; } + if (s.empty()) { return true; } if (s.size() == 1) { return is_field_vchar(s[0]); @@ -2835,18 +2952,9 @@ inline std::string decode_url(const std::string &s, return result; } -inline void read_file(const std::string &path, std::string &out) { - std::ifstream fs(path, std::ios_base::binary); - fs.seekg(0, std::ios_base::end); - auto size = fs.tellg(); - fs.seekg(0); - out.resize(static_cast(size)); - fs.read(&out[0], static_cast(size)); -} - inline std::string file_extension(const std::string &path) { std::smatch m; - static auto re = std::regex("\\.([a-zA-Z0-9]+)$"); + thread_local auto re = std::regex("\\.([a-zA-Z0-9]+)$"); if (std::regex_search(path, m, re)) { return m[1].str(); } return std::string(); } @@ -2930,18 +3038,18 @@ inline stream_line_reader::stream_line_reader(Stream &strm, char *fixed_buffer, fixed_buffer_size_(fixed_buffer_size) {} inline const char *stream_line_reader::ptr() const { - if (glowable_buffer_.empty()) { + if (growable_buffer_.empty()) { return fixed_buffer_; } else { - return glowable_buffer_.data(); + return growable_buffer_.data(); } } inline size_t stream_line_reader::size() const { - if (glowable_buffer_.empty()) { + if (growable_buffer_.empty()) { return fixed_buffer_used_size_; } else { - return glowable_buffer_.size(); + return growable_buffer_.size(); } } @@ -2952,7 +3060,7 @@ inline bool stream_line_reader::end_with_crlf() const { inline bool stream_line_reader::getline() { fixed_buffer_used_size_ = 0; - glowable_buffer_.clear(); + growable_buffer_.clear(); #ifndef CPPHTTPLIB_ALLOW_LF_AS_LINE_TERMINATOR char prev_byte = 0; @@ -2990,11 +3098,11 @@ inline void stream_line_reader::append(char c) { fixed_buffer_[fixed_buffer_used_size_++] = c; fixed_buffer_[fixed_buffer_used_size_] = '\0'; } else { - if (glowable_buffer_.empty()) { + if (growable_buffer_.empty()) { assert(fixed_buffer_[fixed_buffer_used_size_] == '\0'); - glowable_buffer_.assign(fixed_buffer_, fixed_buffer_used_size_); + growable_buffer_.assign(fixed_buffer_, fixed_buffer_used_size_); } - glowable_buffer_ += c; + growable_buffer_ += c; } } @@ -3171,72 +3279,43 @@ inline ssize_t send_socket(socket_t sock, const void *ptr, size_t size, }); } -inline ssize_t select_read(socket_t sock, time_t sec, time_t usec) { -#ifdef CPPHTTPLIB_USE_POLL - struct pollfd pfd_read; - pfd_read.fd = sock; - pfd_read.events = POLLIN; - - auto timeout = static_cast(sec * 1000 + usec / 1000); - - return handle_EINTR([&]() { return poll(&pfd_read, 1, timeout); }); +inline int poll_wrapper(struct pollfd *fds, nfds_t nfds, int timeout) { +#ifdef _WIN32 + return ::WSAPoll(fds, nfds, timeout); #else -#ifndef _WIN32 - if (sock >= FD_SETSIZE) { return -1; } -#endif - - fd_set fds; - FD_ZERO(&fds); - FD_SET(sock, &fds); - - timeval tv; - tv.tv_sec = static_cast(sec); - tv.tv_usec = static_cast(usec); - - return handle_EINTR([&]() { - return select(static_cast(sock + 1), &fds, nullptr, nullptr, &tv); - }); + return ::poll(fds, nfds, timeout); #endif } -inline ssize_t select_write(socket_t sock, time_t sec, time_t usec) { -#ifdef CPPHTTPLIB_USE_POLL - struct pollfd pfd_read; - pfd_read.fd = sock; - pfd_read.events = POLLOUT; +template +inline ssize_t select_impl(socket_t sock, time_t sec, time_t usec) { + struct pollfd pfd; + pfd.fd = sock; + pfd.events = (Read ? POLLIN : POLLOUT); auto timeout = static_cast(sec * 1000 + usec / 1000); - return handle_EINTR([&]() { return poll(&pfd_read, 1, timeout); }); -#else -#ifndef _WIN32 - if (sock >= FD_SETSIZE) { return -1; } -#endif + return handle_EINTR([&]() { return poll_wrapper(&pfd, 1, timeout); }); +} - fd_set fds; - FD_ZERO(&fds); - FD_SET(sock, &fds); +inline ssize_t select_read(socket_t sock, time_t sec, time_t usec) { + return select_impl(sock, sec, usec); +} - timeval tv; - tv.tv_sec = static_cast(sec); - tv.tv_usec = static_cast(usec); - - return handle_EINTR([&]() { - return select(static_cast(sock + 1), nullptr, &fds, nullptr, &tv); - }); -#endif +inline ssize_t select_write(socket_t sock, time_t sec, time_t usec) { + return select_impl(sock, sec, usec); } inline Error wait_until_socket_is_ready(socket_t sock, time_t sec, time_t usec) { -#ifdef CPPHTTPLIB_USE_POLL struct pollfd pfd_read; pfd_read.fd = sock; pfd_read.events = POLLIN | POLLOUT; auto timeout = static_cast(sec * 1000 + usec / 1000); - auto poll_res = handle_EINTR([&]() { return poll(&pfd_read, 1, timeout); }); + auto poll_res = + handle_EINTR([&]() { return poll_wrapper(&pfd_read, 1, timeout); }); if (poll_res == 0) { return Error::ConnectionTimeout; } @@ -3250,38 +3329,6 @@ inline Error wait_until_socket_is_ready(socket_t sock, time_t sec, } return Error::Connection; -#else -#ifndef _WIN32 - if (sock >= FD_SETSIZE) { return Error::Connection; } -#endif - - fd_set fdsr; - FD_ZERO(&fdsr); - FD_SET(sock, &fdsr); - - auto fdsw = fdsr; - auto fdse = fdsr; - - timeval tv; - tv.tv_sec = static_cast(sec); - tv.tv_usec = static_cast(usec); - - auto ret = handle_EINTR([&]() { - return select(static_cast(sock + 1), &fdsr, &fdsw, &fdse, &tv); - }); - - if (ret == 0) { return Error::ConnectionTimeout; } - - if (ret > 0 && (FD_ISSET(sock, &fdsr) || FD_ISSET(sock, &fdsw))) { - auto error = 0; - socklen_t len = sizeof(error); - auto res = getsockopt(sock, SOL_SOCKET, SO_ERROR, - reinterpret_cast(&error), &len); - auto successful = res >= 0 && !error; - return successful ? Error::Success : Error::Connection; - } - return Error::Connection; -#endif } inline bool is_socket_alive(socket_t sock) { @@ -3298,16 +3345,21 @@ inline bool is_socket_alive(socket_t sock) { class SocketStream final : public Stream { public: SocketStream(socket_t sock, time_t read_timeout_sec, time_t read_timeout_usec, - time_t write_timeout_sec, time_t write_timeout_usec); + time_t write_timeout_sec, time_t write_timeout_usec, + time_t max_timeout_msec = 0, + std::chrono::time_point start_time = + (std::chrono::steady_clock::time_point::min)()); ~SocketStream() override; bool is_readable() const override; - bool is_writable() const override; + bool wait_readable() const override; + bool wait_writable() const override; ssize_t read(char *ptr, size_t size) override; ssize_t write(const char *ptr, size_t size) override; void get_remote_ip_and_port(std::string &ip, int &port) const override; void get_local_ip_and_port(std::string &ip, int &port) const override; socket_t socket() const override; + time_t duration() const override; private: socket_t sock_; @@ -3315,6 +3367,8 @@ private: time_t read_timeout_usec_; time_t write_timeout_sec_; time_t write_timeout_usec_; + time_t max_timeout_msec_; + const std::chrono::time_point start_time_; std::vector read_buff_; size_t read_buff_off_ = 0; @@ -3326,18 +3380,23 @@ private: #ifdef CPPHTTPLIB_OPENSSL_SUPPORT class SSLSocketStream final : public Stream { public: - SSLSocketStream(socket_t sock, SSL *ssl, time_t read_timeout_sec, - time_t read_timeout_usec, time_t write_timeout_sec, - time_t write_timeout_usec); + SSLSocketStream( + socket_t sock, SSL *ssl, time_t read_timeout_sec, + time_t read_timeout_usec, time_t write_timeout_sec, + time_t write_timeout_usec, time_t max_timeout_msec = 0, + std::chrono::time_point start_time = + (std::chrono::steady_clock::time_point::min)()); ~SSLSocketStream() override; bool is_readable() const override; - bool is_writable() const override; + bool wait_readable() const override; + bool wait_writable() const override; ssize_t read(char *ptr, size_t size) override; ssize_t write(const char *ptr, size_t size) override; void get_remote_ip_and_port(std::string &ip, int &port) const override; void get_local_ip_and_port(std::string &ip, int &port) const override; socket_t socket() const override; + time_t duration() const override; private: socket_t sock_; @@ -3346,6 +3405,8 @@ private: time_t read_timeout_usec_; time_t write_timeout_sec_; time_t write_timeout_usec_; + time_t max_timeout_msec_; + const std::chrono::time_point start_time_; }; #endif @@ -3416,13 +3477,15 @@ process_server_socket(const std::atomic &svr_sock, socket_t sock, }); } -inline bool process_client_socket(socket_t sock, time_t read_timeout_sec, - time_t read_timeout_usec, - time_t write_timeout_sec, - time_t write_timeout_usec, - std::function callback) { +inline bool process_client_socket( + socket_t sock, time_t read_timeout_sec, time_t read_timeout_usec, + time_t write_timeout_sec, time_t write_timeout_usec, + time_t max_timeout_msec, + std::chrono::time_point start_time, + std::function callback) { SocketStream strm(sock, read_timeout_sec, read_timeout_usec, - write_timeout_sec, write_timeout_usec); + write_timeout_sec, write_timeout_usec, max_timeout_msec, + start_time); return callback(strm); } @@ -3478,7 +3541,6 @@ socket_t create_socket(const std::string &host, const std::string &ip, int port, hints.ai_flags = socket_flags; } -#ifndef _WIN32 if (hints.ai_family == AF_UNIX) { const auto addrlen = host.length(); if (addrlen > sizeof(sockaddr_un::sun_path)) { return INVALID_SOCKET; } @@ -3502,11 +3564,19 @@ socket_t create_socket(const std::string &host, const std::string &ip, int port, sizeof(addr) - sizeof(addr.sun_path) + addrlen); #ifndef SOCK_CLOEXEC +#ifndef _WIN32 fcntl(sock, F_SETFD, FD_CLOEXEC); +#endif #endif if (socket_options) { socket_options(sock); } +#ifdef _WIN32 + // Setting SO_REUSEADDR seems not to work well with AF_UNIX on windows, so + // remove the option. + detail::set_socket_opt(sock, SOL_SOCKET, SO_REUSEADDR, 0); +#endif + bool dummy; if (!bind_or_connect(sock, hints, dummy)) { close_socket(sock); @@ -3515,7 +3585,6 @@ socket_t create_socket(const std::string &host, const std::string &ip, int port, } return sock; } -#endif auto service = std::to_string(port); @@ -3569,26 +3638,10 @@ socket_t create_socket(const std::string &host, const std::string &ip, int port, } #endif - if (tcp_nodelay) { - auto opt = 1; -#ifdef _WIN32 - setsockopt(sock, IPPROTO_TCP, TCP_NODELAY, - reinterpret_cast(&opt), sizeof(opt)); -#else - setsockopt(sock, IPPROTO_TCP, TCP_NODELAY, - reinterpret_cast(&opt), sizeof(opt)); -#endif - } + if (tcp_nodelay) { set_socket_opt(sock, IPPROTO_TCP, TCP_NODELAY, 1); } if (rp->ai_family == AF_INET6) { - auto opt = ipv6_v6only ? 1 : 0; -#ifdef _WIN32 - setsockopt(sock, IPPROTO_IPV6, IPV6_V6ONLY, - reinterpret_cast(&opt), sizeof(opt)); -#else - setsockopt(sock, IPPROTO_IPV6, IPV6_V6ONLY, - reinterpret_cast(&opt), sizeof(opt)); -#endif + set_socket_opt(sock, IPPROTO_IPV6, IPV6_V6ONLY, ipv6_v6only ? 1 : 0); } if (socket_options) { socket_options(sock); } @@ -3731,36 +3784,10 @@ inline socket_t create_client_socket( } set_nonblocking(sock2, false); - - { -#ifdef _WIN32 - auto timeout = static_cast(read_timeout_sec * 1000 + - read_timeout_usec / 1000); - setsockopt(sock2, SOL_SOCKET, SO_RCVTIMEO, - reinterpret_cast(&timeout), sizeof(timeout)); -#else - timeval tv; - tv.tv_sec = static_cast(read_timeout_sec); - tv.tv_usec = static_cast(read_timeout_usec); - setsockopt(sock2, SOL_SOCKET, SO_RCVTIMEO, - reinterpret_cast(&tv), sizeof(tv)); -#endif - } - { - -#ifdef _WIN32 - auto timeout = static_cast(write_timeout_sec * 1000 + - write_timeout_usec / 1000); - setsockopt(sock2, SOL_SOCKET, SO_SNDTIMEO, - reinterpret_cast(&timeout), sizeof(timeout)); -#else - timeval tv; - tv.tv_sec = static_cast(write_timeout_sec); - tv.tv_usec = static_cast(write_timeout_usec); - setsockopt(sock2, SOL_SOCKET, SO_SNDTIMEO, - reinterpret_cast(&tv), sizeof(tv)); -#endif - } + set_socket_opt_time(sock2, SOL_SOCKET, SO_RCVTIMEO, read_timeout_sec, + read_timeout_usec); + set_socket_opt_time(sock2, SOL_SOCKET, SO_SNDTIMEO, write_timeout_sec, + write_timeout_usec); error = Error::Success; return true; @@ -3963,6 +3990,12 @@ inline EncodingType encoding_type(const Request &req, const Response &res) { if (ret) { return EncodingType::Gzip; } #endif +#ifdef CPPHTTPLIB_ZSTD_SUPPORT + // TODO: 'Accept-Encoding' has zstd, not zstd;q=0 + ret = s.find("zstd") != std::string::npos; + if (ret) { return EncodingType::Zstd; } +#endif + return EncodingType::None; } @@ -4171,6 +4204,61 @@ inline bool brotli_decompressor::decompress(const char *data, } #endif +#ifdef CPPHTTPLIB_ZSTD_SUPPORT +inline zstd_compressor::zstd_compressor() { + ctx_ = ZSTD_createCCtx(); + ZSTD_CCtx_setParameter(ctx_, ZSTD_c_compressionLevel, ZSTD_fast); +} + +inline zstd_compressor::~zstd_compressor() { ZSTD_freeCCtx(ctx_); } + +inline bool zstd_compressor::compress(const char *data, size_t data_length, + bool last, Callback callback) { + std::array buff{}; + + ZSTD_EndDirective mode = last ? ZSTD_e_end : ZSTD_e_continue; + ZSTD_inBuffer input = {data, data_length, 0}; + + bool finished; + do { + ZSTD_outBuffer output = {buff.data(), CPPHTTPLIB_COMPRESSION_BUFSIZ, 0}; + size_t const remaining = ZSTD_compressStream2(ctx_, &output, &input, mode); + + if (ZSTD_isError(remaining)) { return false; } + + if (!callback(buff.data(), output.pos)) { return false; } + + finished = last ? (remaining == 0) : (input.pos == input.size); + + } while (!finished); + + return true; +} + +inline zstd_decompressor::zstd_decompressor() { ctx_ = ZSTD_createDCtx(); } + +inline zstd_decompressor::~zstd_decompressor() { ZSTD_freeDCtx(ctx_); } + +inline bool zstd_decompressor::is_valid() const { return ctx_ != nullptr; } + +inline bool zstd_decompressor::decompress(const char *data, size_t data_length, + Callback callback) { + std::array buff{}; + ZSTD_inBuffer input = {data, data_length, 0}; + + while (input.pos < input.size) { + ZSTD_outBuffer output = {buff.data(), CPPHTTPLIB_COMPRESSION_BUFSIZ, 0}; + size_t const remaining = ZSTD_decompressStream(ctx_, &output, &input); + + if (ZSTD_isError(remaining)) { return false; } + + if (!callback(buff.data(), output.pos)) { return false; } + } + + return true; +} +#endif + inline bool has_header(const Headers &headers, const std::string &key) { return headers.find(key) != headers.end(); } @@ -4197,6 +4285,9 @@ inline bool parse_header(const char *beg, const char *end, T fn) { p++; } + auto name = std::string(beg, p); + if (!detail::fields::is_field_name(name)) { return false; } + if (p == end) { return false; } auto key_end = p; @@ -4212,22 +4303,17 @@ inline bool parse_header(const char *beg, const char *end, T fn) { if (!key_len) { return false; } auto key = std::string(beg, key_end); - auto val = case_ignore::equal(key, "Location") - ? std::string(p, end) - : decode_url(std::string(p, end), false); + auto val = std::string(p, end); - // NOTE: From RFC 9110: - // Field values containing CR, LF, or NUL characters are - // invalid and dangerous, due to the varying ways that - // implementations might parse and interpret those - // characters; a recipient of CR, LF, or NUL within a field - // value MUST either reject the message or replace each of - // those characters with SP before further processing or - // forwarding of that message. - static const std::string CR_LF_NUL("\r\n\0", 3); - if (val.find_first_of(CR_LF_NUL) != std::string::npos) { return false; } + if (!detail::fields::is_field_value(val)) { return false; } + + if (case_ignore::equal(key, "Location") || + case_ignore::equal(key, "Referer")) { + fn(key, val); + } else { + fn(key, decode_url(val, false)); + } - fn(key, val); return true; } @@ -4263,7 +4349,7 @@ inline bool read_headers(Stream &strm, Headers &headers) { auto end = line_reader.ptr() + line_reader.size() - line_terminator_len; if (!parse_header(line_reader.ptr(), end, - [&](const std::string &key, std::string &val) { + [&](const std::string &key, const std::string &val) { headers.emplace(key, val); })) { return false; @@ -4312,7 +4398,8 @@ inline bool read_content_without_length(Stream &strm, uint64_t r = 0; for (;;) { auto n = strm.read(buf, CPPHTTPLIB_RECV_BUFSIZ); - if (n <= 0) { return true; } + if (n == 0) { return true; } + if (n < 0) { return false; } if (!out(buf, static_cast(n), r, 0)) { return false; } r += static_cast(n); @@ -4355,7 +4442,7 @@ inline bool read_content_chunked(Stream &strm, T &x, assert(chunk_len == 0); - // NOTE: In RFC 9112, '7.1 Chunked Transfer Coding' mentiones "The chunked + // NOTE: In RFC 9112, '7.1 Chunked Transfer Coding' mentions "The chunked // transfer coding is complete when a chunk with a chunk-size of zero is // received, possibly followed by a trailer section, and finally terminated by // an empty line". https://www.rfc-editor.org/rfc/rfc9112.html#section-7.1 @@ -4365,8 +4452,8 @@ inline bool read_content_chunked(Stream &strm, T &x, // to be ok whether the final CRLF exists or not in the chunked data. // https://www.rfc-editor.org/rfc/rfc9112.html#section-7.1.3 // - // According to the reference code in RFC 9112, cpp-htpplib now allows - // chuncked transfer coding data without the final CRLF. + // According to the reference code in RFC 9112, cpp-httplib now allows + // chunked transfer coding data without the final CRLF. if (!line_reader.getline()) { return true; } while (strcmp(line_reader.ptr(), "\r\n") != 0) { @@ -4413,6 +4500,13 @@ bool prepare_content_receiver(T &x, int &status, #else status = StatusCode::UnsupportedMediaType_415; return false; +#endif + } else if (encoding == "zstd") { +#ifdef CPPHTTPLIB_ZSTD_SUPPORT + decompressor = detail::make_unique(); +#else + status = StatusCode::UnsupportedMediaType_415; + return false; #endif } @@ -4456,9 +4550,9 @@ bool read_content(Stream &strm, T &x, size_t payload_max_length, int &status, ret = read_content_without_length(strm, out); } else { auto is_invalid_value = false; - auto len = get_header_value_u64(x.headers, "Content-Length", - std::numeric_limits::max(), - 0, is_invalid_value); + auto len = get_header_value_u64( + x.headers, "Content-Length", + (std::numeric_limits::max)(), 0, is_invalid_value); if (is_invalid_value) { ret = false; @@ -4536,7 +4630,7 @@ inline bool write_content(Stream &strm, const ContentProvider &content_provider, data_sink.write = [&](const char *d, size_t l) -> bool { if (ok) { - if (strm.is_writable() && write_data(strm, d, l)) { + if (write_data(strm, d, l)) { offset += l; } else { ok = false; @@ -4545,10 +4639,10 @@ inline bool write_content(Stream &strm, const ContentProvider &content_provider, return ok; }; - data_sink.is_writable = [&]() -> bool { return strm.is_writable(); }; + data_sink.is_writable = [&]() -> bool { return strm.wait_writable(); }; while (offset < end_offset && !is_shutting_down()) { - if (!strm.is_writable()) { + if (!strm.wait_writable()) { error = Error::Write; return false; } else if (!content_provider(offset, end_offset - offset, data_sink)) { @@ -4586,17 +4680,17 @@ write_content_without_length(Stream &strm, data_sink.write = [&](const char *d, size_t l) -> bool { if (ok) { offset += l; - if (!strm.is_writable() || !write_data(strm, d, l)) { ok = false; } + if (!write_data(strm, d, l)) { ok = false; } } return ok; }; - data_sink.is_writable = [&]() -> bool { return strm.is_writable(); }; + data_sink.is_writable = [&]() -> bool { return strm.wait_writable(); }; data_sink.done = [&](void) { data_available = false; }; while (data_available && !is_shutting_down()) { - if (!strm.is_writable()) { + if (!strm.wait_writable()) { return false; } else if (!content_provider(offset, 0, data_sink)) { return false; @@ -4631,10 +4725,7 @@ write_content_chunked(Stream &strm, const ContentProvider &content_provider, // Emit chunked response header and footer for each chunk auto chunk = from_i_to_hex(payload.size()) + "\r\n" + payload + "\r\n"; - if (!strm.is_writable() || - !write_data(strm, chunk.data(), chunk.size())) { - ok = false; - } + if (!write_data(strm, chunk.data(), chunk.size())) { ok = false; } } } else { ok = false; @@ -4643,7 +4734,7 @@ write_content_chunked(Stream &strm, const ContentProvider &content_provider, return ok; }; - data_sink.is_writable = [&]() -> bool { return strm.is_writable(); }; + data_sink.is_writable = [&]() -> bool { return strm.wait_writable(); }; auto done_with_trailer = [&](const Headers *trailer) { if (!ok) { return; } @@ -4663,17 +4754,14 @@ write_content_chunked(Stream &strm, const ContentProvider &content_provider, if (!payload.empty()) { // Emit chunked response header and footer for each chunk auto chunk = from_i_to_hex(payload.size()) + "\r\n" + payload + "\r\n"; - if (!strm.is_writable() || - !write_data(strm, chunk.data(), chunk.size())) { + if (!write_data(strm, chunk.data(), chunk.size())) { ok = false; return; } } - static const std::string done_marker("0\r\n"); - if (!write_data(strm, done_marker.data(), done_marker.size())) { - ok = false; - } + constexpr const char done_marker[] = "0\r\n"; + if (!write_data(strm, done_marker, str_len(done_marker))) { ok = false; } // Trailer if (trailer) { @@ -4685,8 +4773,8 @@ write_content_chunked(Stream &strm, const ContentProvider &content_provider, } } - static const std::string crlf("\r\n"); - if (!write_data(strm, crlf.data(), crlf.size())) { ok = false; } + constexpr const char crlf[] = "\r\n"; + if (!write_data(strm, crlf, str_len(crlf))) { ok = false; } }; data_sink.done = [&](void) { done_with_trailer(nullptr); }; @@ -4696,7 +4784,7 @@ write_content_chunked(Stream &strm, const ContentProvider &content_provider, }; while (data_available && !is_shutting_down()) { - if (!strm.is_writable()) { + if (!strm.wait_writable()) { error = Error::Write; return false; } else if (!content_provider(offset, 0, data_sink)) { @@ -4928,13 +5016,13 @@ public: return false; } - static const std::string header_content_type = "Content-Type:"; + constexpr const char header_content_type[] = "Content-Type:"; if (start_with_case_ignore(header, header_content_type)) { file_.content_type = - trim_copy(header.substr(header_content_type.size())); + trim_copy(header.substr(str_len(header_content_type))); } else { - static const std::regex re_content_disposition( + thread_local const std::regex re_content_disposition( R"~(^Content-Disposition:\s*form-data;\s*(.*)$)~", std::regex_constants::icase); @@ -4956,8 +5044,8 @@ public: it = params.find("filename*"); if (it != params.end()) { - // Only allow UTF-8 enconnding... - static const std::regex re_rfc5987_encoding( + // Only allow UTF-8 encoding... + thread_local const std::regex re_rfc5987_encoding( R"~(^UTF-8''(.+?)$)~", std::regex_constants::icase); std::smatch m2; @@ -5029,10 +5117,10 @@ private: file_.content_type.clear(); } - bool start_with_case_ignore(const std::string &a, - const std::string &b) const { - if (a.size() < b.size()) { return false; } - for (size_t i = 0; i < b.size(); i++) { + bool start_with_case_ignore(const std::string &a, const char *b) const { + const auto b_len = strlen(b); + if (a.size() < b_len) { return false; } + for (size_t i = 0; i < b_len; i++) { if (case_ignore::to_lower(a[i]) != case_ignore::to_lower(b[i])) { return false; } @@ -5119,19 +5207,18 @@ private: }; inline std::string random_string(size_t length) { - static const char data[] = + constexpr const char data[] = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz"; - // std::random_device might actually be deterministic on some - // platforms, but due to lack of support in the c++ standard library, - // doing better requires either some ugly hacks or breaking portability. - static std::random_device seed_gen; - - // Request 128 bits of entropy for initialization - static std::seed_seq seed_sequence{seed_gen(), seed_gen(), seed_gen(), - seed_gen()}; - - static std::mt19937 engine(seed_sequence); + thread_local auto engine([]() { + // std::random_device might actually be deterministic on some + // platforms, but due to lack of support in the c++ standard library, + // doing better requires either some ugly hacks or breaking portability. + std::random_device seed_gen; + // Request 128 bits of entropy for initialization + std::seed_seq seed_sequence{seed_gen(), seed_gen(), seed_gen(), seed_gen()}; + return std::mt19937(seed_sequence); + }()); std::string result; for (size_t i = 0; i < length; i++) { @@ -5203,7 +5290,7 @@ serialize_multipart_formdata(const MultipartFormDataItems &items, inline bool range_error(Request &req, Response &res) { if (!req.ranges.empty() && 200 <= res.status && res.status < 300) { - ssize_t contant_len = static_cast( + ssize_t content_len = static_cast( res.content_length_ ? res.content_length_ : res.body.size()); ssize_t prev_first_pos = -1; @@ -5223,12 +5310,12 @@ inline bool range_error(Request &req, Response &res) { if (first_pos == -1 && last_pos == -1) { first_pos = 0; - last_pos = contant_len; + last_pos = content_len; } if (first_pos == -1) { - first_pos = contant_len - last_pos; - last_pos = contant_len - 1; + first_pos = content_len - last_pos; + last_pos = content_len - 1; } // NOTE: RFC-9110 '14.1.2. Byte Ranges': @@ -5240,13 +5327,13 @@ inline bool range_error(Request &req, Response &res) { // with a value that is one less than the current length of the selected // representation). // https://www.rfc-editor.org/rfc/rfc9110.html#section-14.1.2-6 - if (last_pos == -1 || last_pos >= contant_len) { - last_pos = contant_len - 1; + if (last_pos == -1 || last_pos >= content_len) { + last_pos = content_len - 1; } // Range must be within content length if (!(0 <= first_pos && first_pos <= last_pos && - last_pos <= contant_len - 1)) { + last_pos <= content_len - 1)) { return true; } @@ -5380,10 +5467,14 @@ write_multipart_ranges_data(Stream &strm, const Request &req, Response &res, inline bool expect_content(const Request &req) { if (req.method == "POST" || req.method == "PUT" || req.method == "PATCH" || - req.method == "PRI" || req.method == "DELETE") { + req.method == "DELETE") { return true; } - // TODO: check if Content-Length is set + if (req.has_header("Content-Length") && + req.get_header_value_u64("Content-Length") > 0) { + return true; + } + if (is_chunked_transfer_encoding(req.headers)) { return true; } return false; } @@ -5428,9 +5519,76 @@ inline std::string SHA_256(const std::string &s) { inline std::string SHA_512(const std::string &s) { return message_digest(s, EVP_sha512()); } -#endif -#ifdef CPPHTTPLIB_OPENSSL_SUPPORT +inline std::pair make_digest_authentication_header( + const Request &req, const std::map &auth, + size_t cnonce_count, const std::string &cnonce, const std::string &username, + const std::string &password, bool is_proxy = false) { + std::string nc; + { + std::stringstream ss; + ss << std::setfill('0') << std::setw(8) << std::hex << cnonce_count; + nc = ss.str(); + } + + std::string qop; + if (auth.find("qop") != auth.end()) { + qop = auth.at("qop"); + if (qop.find("auth-int") != std::string::npos) { + qop = "auth-int"; + } else if (qop.find("auth") != std::string::npos) { + qop = "auth"; + } else { + qop.clear(); + } + } + + std::string algo = "MD5"; + if (auth.find("algorithm") != auth.end()) { algo = auth.at("algorithm"); } + + std::string response; + { + auto H = algo == "SHA-256" ? detail::SHA_256 + : algo == "SHA-512" ? detail::SHA_512 + : detail::MD5; + + auto A1 = username + ":" + auth.at("realm") + ":" + password; + + auto A2 = req.method + ":" + req.path; + if (qop == "auth-int") { A2 += ":" + H(req.body); } + + if (qop.empty()) { + response = H(H(A1) + ":" + auth.at("nonce") + ":" + H(A2)); + } else { + response = H(H(A1) + ":" + auth.at("nonce") + ":" + nc + ":" + cnonce + + ":" + qop + ":" + H(A2)); + } + } + + auto opaque = (auth.find("opaque") != auth.end()) ? auth.at("opaque") : ""; + + auto field = "Digest username=\"" + username + "\", realm=\"" + + auth.at("realm") + "\", nonce=\"" + auth.at("nonce") + + "\", uri=\"" + req.path + "\", algorithm=" + algo + + (qop.empty() ? ", response=\"" + : ", qop=" + qop + ", nc=" + nc + ", cnonce=\"" + + cnonce + "\", response=\"") + + response + "\"" + + (opaque.empty() ? "" : ", opaque=\"" + opaque + "\""); + + auto key = is_proxy ? "Proxy-Authorization" : "Authorization"; + return std::make_pair(key, field); +} + +inline bool is_ssl_peer_could_be_closed(SSL *ssl, socket_t sock) { + detail::set_nonblocking(sock, true); + auto se = detail::scope_exit([&]() { detail::set_nonblocking(sock, false); }); + + char buf[1]; + return !SSL_peek(ssl, buf, 1) && + SSL_get_error(ssl, 0) == SSL_ERROR_ZERO_RETURN; +} + #ifdef _WIN32 // NOTE: This code came up with the following stackoverflow post: // https://stackoverflow.com/questions/9507184/can-openssl-on-windows-use-the-system-certificate-store @@ -5569,74 +5727,13 @@ public: static WSInit wsinit_; #endif -#ifdef CPPHTTPLIB_OPENSSL_SUPPORT -inline std::pair make_digest_authentication_header( - const Request &req, const std::map &auth, - size_t cnonce_count, const std::string &cnonce, const std::string &username, - const std::string &password, bool is_proxy = false) { - std::string nc; - { - std::stringstream ss; - ss << std::setfill('0') << std::setw(8) << std::hex << cnonce_count; - nc = ss.str(); - } - - std::string qop; - if (auth.find("qop") != auth.end()) { - qop = auth.at("qop"); - if (qop.find("auth-int") != std::string::npos) { - qop = "auth-int"; - } else if (qop.find("auth") != std::string::npos) { - qop = "auth"; - } else { - qop.clear(); - } - } - - std::string algo = "MD5"; - if (auth.find("algorithm") != auth.end()) { algo = auth.at("algorithm"); } - - std::string response; - { - auto H = algo == "SHA-256" ? detail::SHA_256 - : algo == "SHA-512" ? detail::SHA_512 - : detail::MD5; - - auto A1 = username + ":" + auth.at("realm") + ":" + password; - - auto A2 = req.method + ":" + req.path; - if (qop == "auth-int") { A2 += ":" + H(req.body); } - - if (qop.empty()) { - response = H(H(A1) + ":" + auth.at("nonce") + ":" + H(A2)); - } else { - response = H(H(A1) + ":" + auth.at("nonce") + ":" + nc + ":" + cnonce + - ":" + qop + ":" + H(A2)); - } - } - - auto opaque = (auth.find("opaque") != auth.end()) ? auth.at("opaque") : ""; - - auto field = "Digest username=\"" + username + "\", realm=\"" + - auth.at("realm") + "\", nonce=\"" + auth.at("nonce") + - "\", uri=\"" + req.path + "\", algorithm=" + algo + - (qop.empty() ? ", response=\"" - : ", qop=" + qop + ", nc=" + nc + ", cnonce=\"" + - cnonce + "\", response=\"") + - response + "\"" + - (opaque.empty() ? "" : ", opaque=\"" + opaque + "\""); - - auto key = is_proxy ? "Proxy-Authorization" : "Authorization"; - return std::make_pair(key, field); -} -#endif - inline bool parse_www_authenticate(const Response &res, std::map &auth, bool is_proxy) { auto auth_key = is_proxy ? "Proxy-Authenticate" : "WWW-Authenticate"; if (res.has_header(auth_key)) { - static auto re = std::regex(R"~((?:(?:,\s*)?(.+?)=(?:"(.*?)"|([^,]*))))~"); + thread_local auto re = + std::regex(R"~((?:(?:,\s*)?(.+?)=(?:"(.*?)"|([^,]*))))~"); auto s = res.get_header_value(auth_key); auto pos = s.find(' '); if (pos != std::string::npos) { @@ -5720,7 +5817,7 @@ inline void hosted_at(const std::string &hostname, inline std::string append_query_params(const std::string &path, const Params ¶ms) { std::string path_with_query = path; - const static std::regex re("[^?]+\\?.*"); + thread_local const std::regex re("[^?]+\\?.*"); auto delm = std::regex_match(path, re) ? '&' : '?'; path_with_query += delm + detail::params_to_query_str(params); return path_with_query; @@ -5949,23 +6046,52 @@ inline ssize_t Stream::write(const std::string &s) { namespace detail { +inline void calc_actual_timeout(time_t max_timeout_msec, time_t duration_msec, + time_t timeout_sec, time_t timeout_usec, + time_t &actual_timeout_sec, + time_t &actual_timeout_usec) { + auto timeout_msec = (timeout_sec * 1000) + (timeout_usec / 1000); + + auto actual_timeout_msec = + (std::min)(max_timeout_msec - duration_msec, timeout_msec); + + actual_timeout_sec = actual_timeout_msec / 1000; + actual_timeout_usec = (actual_timeout_msec % 1000) * 1000; +} + // Socket stream implementation -inline SocketStream::SocketStream(socket_t sock, time_t read_timeout_sec, - time_t read_timeout_usec, - time_t write_timeout_sec, - time_t write_timeout_usec) +inline SocketStream::SocketStream( + socket_t sock, time_t read_timeout_sec, time_t read_timeout_usec, + time_t write_timeout_sec, time_t write_timeout_usec, + time_t max_timeout_msec, + std::chrono::time_point start_time) : sock_(sock), read_timeout_sec_(read_timeout_sec), read_timeout_usec_(read_timeout_usec), write_timeout_sec_(write_timeout_sec), - write_timeout_usec_(write_timeout_usec), read_buff_(read_buff_size_, 0) {} + write_timeout_usec_(write_timeout_usec), + max_timeout_msec_(max_timeout_msec), start_time_(start_time), + read_buff_(read_buff_size_, 0) {} inline SocketStream::~SocketStream() = default; inline bool SocketStream::is_readable() const { - return select_read(sock_, read_timeout_sec_, read_timeout_usec_) > 0; + return read_buff_off_ < read_buff_content_size_; } -inline bool SocketStream::is_writable() const { +inline bool SocketStream::wait_readable() const { + if (max_timeout_msec_ <= 0) { + return select_read(sock_, read_timeout_sec_, read_timeout_usec_) > 0; + } + + time_t read_timeout_sec; + time_t read_timeout_usec; + calc_actual_timeout(max_timeout_msec_, duration(), read_timeout_sec_, + read_timeout_usec_, read_timeout_sec, read_timeout_usec); + + return select_read(sock_, read_timeout_sec, read_timeout_usec) > 0; +} + +inline bool SocketStream::wait_writable() const { return select_write(sock_, write_timeout_sec_, write_timeout_usec_) > 0 && is_socket_alive(sock_); } @@ -5992,7 +6118,7 @@ inline ssize_t SocketStream::read(char *ptr, size_t size) { } } - if (!is_readable()) { return -1; } + if (!wait_readable()) { return -1; } read_buff_off_ = 0; read_buff_content_size_ = 0; @@ -6017,7 +6143,7 @@ inline ssize_t SocketStream::read(char *ptr, size_t size) { } inline ssize_t SocketStream::write(const char *ptr, size_t size) { - if (!is_writable()) { return -1; } + if (!wait_writable()) { return -1; } #if defined(_WIN32) && !defined(_WIN64) size = @@ -6039,10 +6165,18 @@ inline void SocketStream::get_local_ip_and_port(std::string &ip, inline socket_t SocketStream::socket() const { return sock_; } +inline time_t SocketStream::duration() const { + return std::chrono::duration_cast( + std::chrono::steady_clock::now() - start_time_) + .count(); +} + // Buffer stream implementation inline bool BufferStream::is_readable() const { return true; } -inline bool BufferStream::is_writable() const { return true; } +inline bool BufferStream::wait_readable() const { return true; } + +inline bool BufferStream::wait_writable() const { return true; } inline ssize_t BufferStream::read(char *ptr, size_t size) { #if defined(_MSC_VER) && _MSC_VER < 1910 @@ -6067,10 +6201,12 @@ inline void BufferStream::get_local_ip_and_port(std::string & /*ip*/, inline socket_t BufferStream::socket() const { return 0; } +inline time_t BufferStream::duration() const { return 0; } + inline const std::string &BufferStream::get_buffer() const { return buffer; } inline PathParamsMatcher::PathParamsMatcher(const std::string &pattern) { - static constexpr char marker[] = "/:"; + constexpr const char marker[] = "/:"; // One past the last ending position of a path param substring std::size_t last_param_end = 0; @@ -6091,7 +6227,7 @@ inline PathParamsMatcher::PathParamsMatcher(const std::string &pattern) { static_fragments_.push_back( pattern.substr(last_param_end, marker_pos - last_param_end + 1)); - const auto param_name_start = marker_pos + 2; + const auto param_name_start = marker_pos + str_len(marker); auto sep_pos = pattern.find(separator, param_name_start); if (sep_pos == std::string::npos) { sep_pos = pattern.length(); } @@ -6398,12 +6534,12 @@ inline Server &Server::set_payload_max_length(size_t length) { inline bool Server::bind_to_port(const std::string &host, int port, int socket_flags) { auto ret = bind_internal(host, port, socket_flags); - if (ret == -1) { is_decommisioned = true; } + if (ret == -1) { is_decommissioned = true; } return ret >= 0; } inline int Server::bind_to_any_port(const std::string &host, int socket_flags) { auto ret = bind_internal(host, 0, socket_flags); - if (ret == -1) { is_decommisioned = true; } + if (ret == -1) { is_decommissioned = true; } return ret; } @@ -6417,7 +6553,7 @@ inline bool Server::listen(const std::string &host, int port, inline bool Server::is_running() const { return is_running_; } inline void Server::wait_until_ready() const { - while (!is_running_ && !is_decommisioned) { + while (!is_running_ && !is_decommissioned) { std::this_thread::sleep_for(std::chrono::milliseconds{1}); } } @@ -6429,10 +6565,10 @@ inline void Server::stop() { detail::shutdown_socket(sock); detail::close_socket(sock); } - is_decommisioned = false; + is_decommissioned = false; } -inline void Server::decommission() { is_decommisioned = true; } +inline void Server::decommission() { is_decommissioned = true; } inline bool Server::parse_request_line(const char *s, Request &req) const { auto len = strlen(s); @@ -6455,7 +6591,7 @@ inline bool Server::parse_request_line(const char *s, Request &req) const { if (count != 3) { return false; } } - static const std::set methods{ + thread_local const std::set methods{ "GET", "HEAD", "POST", "PUT", "DELETE", "CONNECT", "OPTIONS", "TRACE", "PATCH", "PRI"}; @@ -6609,6 +6745,10 @@ Server::write_content_with_provider(Stream &strm, const Request &req, } else if (type == detail::EncodingType::Brotli) { #ifdef CPPHTTPLIB_BROTLI_SUPPORT compressor = detail::make_unique(); +#endif + } else if (type == detail::EncodingType::Zstd) { +#ifdef CPPHTTPLIB_ZSTD_SUPPORT + compressor = detail::make_unique(); #endif } else { compressor = detail::make_unique(); @@ -6791,7 +6931,7 @@ Server::create_server_socket(const std::string &host, int port, inline int Server::bind_internal(const std::string &host, int port, int socket_flags) { - if (is_decommisioned) { return -1; } + if (is_decommissioned) { return -1; } if (!is_valid()) { return -1; } @@ -6818,7 +6958,7 @@ inline int Server::bind_internal(const std::string &host, int port, } inline bool Server::listen_internal() { - if (is_decommisioned) { return false; } + if (is_decommissioned) { return false; } auto ret = true; is_running_ = true; @@ -6842,7 +6982,7 @@ inline bool Server::listen_internal() { #endif #if defined _WIN32 - // sockets conneced via WASAccept inherit flags NO_HANDLE_INHERIT, + // sockets connected via WASAccept inherit flags NO_HANDLE_INHERIT, // OVERLAPPED socket_t sock = WSAAccept(svr_sock_, nullptr, nullptr, nullptr, 0); #elif defined SOCK_CLOEXEC @@ -6869,35 +7009,10 @@ inline bool Server::listen_internal() { break; } - { -#ifdef _WIN32 - auto timeout = static_cast(read_timeout_sec_ * 1000 + - read_timeout_usec_ / 1000); - setsockopt(sock, SOL_SOCKET, SO_RCVTIMEO, - reinterpret_cast(&timeout), sizeof(timeout)); -#else - timeval tv; - tv.tv_sec = static_cast(read_timeout_sec_); - tv.tv_usec = static_cast(read_timeout_usec_); - setsockopt(sock, SOL_SOCKET, SO_RCVTIMEO, - reinterpret_cast(&tv), sizeof(tv)); -#endif - } - { - -#ifdef _WIN32 - auto timeout = static_cast(write_timeout_sec_ * 1000 + - write_timeout_usec_ / 1000); - setsockopt(sock, SOL_SOCKET, SO_SNDTIMEO, - reinterpret_cast(&timeout), sizeof(timeout)); -#else - timeval tv; - tv.tv_sec = static_cast(write_timeout_sec_); - tv.tv_usec = static_cast(write_timeout_usec_); - setsockopt(sock, SOL_SOCKET, SO_SNDTIMEO, - reinterpret_cast(&tv), sizeof(tv)); -#endif - } + detail::set_socket_opt_time(sock, SOL_SOCKET, SO_RCVTIMEO, + read_timeout_sec_, read_timeout_usec_); + detail::set_socket_opt_time(sock, SOL_SOCKET, SO_SNDTIMEO, + write_timeout_sec_, write_timeout_usec_); if (!task_queue->enqueue( [this, sock]() { process_and_close_socket(sock); })) { @@ -6909,7 +7024,7 @@ inline bool Server::listen_internal() { task_queue->shutdown(); } - is_decommisioned = !ret; + is_decommissioned = !ret; return ret; } @@ -7049,6 +7164,8 @@ inline void Server::apply_ranges(const Request &req, Response &res, res.set_header("Content-Encoding", "gzip"); } else if (type == detail::EncodingType::Brotli) { res.set_header("Content-Encoding", "br"); + } else if (type == detail::EncodingType::Zstd) { + res.set_header("Content-Encoding", "zstd"); } } } @@ -7088,6 +7205,11 @@ inline void Server::apply_ranges(const Request &req, Response &res, #ifdef CPPHTTPLIB_BROTLI_SUPPORT compressor = detail::make_unique(); content_encoding = "br"; +#endif + } else if (type == detail::EncodingType::Zstd) { +#ifdef CPPHTTPLIB_ZSTD_SUPPORT + compressor = detail::make_unique(); + content_encoding = "zstd"; #endif } @@ -7143,28 +7265,6 @@ Server::process_request(Stream &strm, const std::string &remote_addr, res.version = "HTTP/1.1"; res.headers = default_headers_; -#ifdef _WIN32 - // TODO: Increase FD_SETSIZE statically (libzmq), dynamically (MySQL). -#else -#ifndef CPPHTTPLIB_USE_POLL - // Socket file descriptor exceeded FD_SETSIZE... - if (strm.socket() >= FD_SETSIZE) { - Headers dummy; - detail::read_headers(strm, dummy); - res.status = StatusCode::InternalServerError_500; - return write_response(strm, close_connection, req, res); - } -#endif -#endif - - // Check if the request URI doesn't exceed the limit - if (line_reader.size() > CPPHTTPLIB_REQUEST_URI_MAX_LENGTH) { - Headers dummy; - detail::read_headers(strm, dummy); - res.status = StatusCode::UriTooLong_414; - return write_response(strm, close_connection, req, res); - } - // Request line and headers if (!parse_request_line(line_reader.ptr(), req) || !detail::read_headers(strm, req.headers)) { @@ -7172,6 +7272,14 @@ Server::process_request(Stream &strm, const std::string &remote_addr, return write_response(strm, close_connection, req, res); } + // Check if the request URI doesn't exceed the limit + if (req.target.size() > CPPHTTPLIB_REQUEST_URI_MAX_LENGTH) { + Headers dummy; + detail::read_headers(strm, dummy); + res.status = StatusCode::UriTooLong_414; + return write_response(strm, close_connection, req, res); + } + if (req.get_header_value("Connection") == "close") { connection_closed = true; } @@ -7348,6 +7456,16 @@ inline ClientImpl::ClientImpl(const std::string &host, int port, client_cert_path_(client_cert_path), client_key_path_(client_key_path) {} inline ClientImpl::~ClientImpl() { + // Wait until all the requests in flight are handled. + size_t retry_count = 10; + while (retry_count-- > 0) { + { + std::lock_guard guard(socket_mutex_); + if (socket_requests_in_flight_ == 0) { break; } + } + std::this_thread::sleep_for(std::chrono::milliseconds{1}); + } + std::lock_guard guard(socket_mutex_); shutdown_socket(socket_); close_socket(socket_); @@ -7363,6 +7481,7 @@ inline void ClientImpl::copy_settings(const ClientImpl &rhs) { read_timeout_usec_ = rhs.read_timeout_usec_; write_timeout_sec_ = rhs.write_timeout_sec_; write_timeout_usec_ = rhs.write_timeout_usec_; + max_timeout_msec_ = rhs.max_timeout_msec_; basic_auth_username_ = rhs.basic_auth_username_; basic_auth_password_ = rhs.basic_auth_password_; bearer_token_auth_token_ = rhs.bearer_token_auth_token_; @@ -7472,9 +7591,9 @@ inline bool ClientImpl::read_response_line(Stream &strm, const Request &req, if (!line_reader.getline()) { return false; } #ifdef CPPHTTPLIB_ALLOW_LF_AS_LINE_TERMINATOR - const static std::regex re("(HTTP/1\\.[01]) (\\d{3})(?: (.*?))?\r?\n"); + thread_local const std::regex re("(HTTP/1\\.[01]) (\\d{3})(?: (.*?))?\r?\n"); #else - const static std::regex re("(HTTP/1\\.[01]) (\\d{3})(?: (.*?))?\r\n"); + thread_local const std::regex re("(HTTP/1\\.[01]) (\\d{3})(?: (.*?))?\r\n"); #endif std::cmatch m; @@ -7509,18 +7628,6 @@ inline bool ClientImpl::send(Request &req, Response &res, Error &error) { return ret; } -#ifdef CPPHTTPLIB_OPENSSL_SUPPORT -inline bool ClientImpl::is_ssl_peer_could_be_closed(SSL *ssl) const { - detail::set_nonblocking(socket_.sock, true); - auto se = detail::scope_exit( - [&]() { detail::set_nonblocking(socket_.sock, false); }); - - char buf[1]; - return !SSL_peek(ssl, buf, 1) && - SSL_get_error(ssl, 0) == SSL_ERROR_ZERO_RETURN; -} -#endif - inline bool ClientImpl::send_(Request &req, Response &res, Error &error) { { std::lock_guard guard(socket_mutex_); @@ -7535,12 +7642,14 @@ inline bool ClientImpl::send_(Request &req, Response &res, Error &error) { #ifdef CPPHTTPLIB_OPENSSL_SUPPORT if (is_alive && is_ssl()) { - if (is_ssl_peer_could_be_closed(socket_.ssl)) { is_alive = false; } + if (detail::is_ssl_peer_could_be_closed(socket_.ssl, socket_.sock)) { + is_alive = false; + } } #endif if (!is_alive) { - // Attempt to avoid sigpipe by shutting down nongracefully if it seems + // Attempt to avoid sigpipe by shutting down non-gracefully if it seems // like the other side has already closed the connection Also, there // cannot be any requests in flight from other threads since we locked // request_mutex_, so safe to close everything immediately @@ -7560,7 +7669,8 @@ inline bool ClientImpl::send_(Request &req, Response &res, Error &error) { auto &scli = static_cast(*this); if (!proxy_host_.empty() && proxy_port_ != -1) { auto success = false; - if (!scli.connect_with_proxy(socket_, res, success, error)) { + if (!scli.connect_with_proxy(socket_, req.start_time_, res, success, + error)) { return success; } } @@ -7606,7 +7716,7 @@ inline bool ClientImpl::send_(Request &req, Response &res, Error &error) { } }); - ret = process_socket(socket_, [&](Stream &strm) { + ret = process_socket(socket_, req.start_time_, [&](Stream &strm) { return handle_request(strm, req, res, close_connection, error); }); @@ -7715,7 +7825,7 @@ inline bool ClientImpl::redirect(Request &req, Response &res, Error &error) { auto location = res.get_header_value("location"); if (location.empty()) { return false; } - const static std::regex re( + thread_local const std::regex re( R"((?:(https?):)?(?://(?:\[([a-fA-F\d:]+)\]|([^:/?#]+))(?::(\d+))?)?([^?#]*)(\?[^#]*)?(?:#.*)?)"); std::smatch m; @@ -7824,6 +7934,10 @@ inline bool ClientImpl::write_request(Stream &strm, Request &req, #ifdef CPPHTTPLIB_ZLIB_SUPPORT if (!accept_encoding.empty()) { accept_encoding += ", "; } accept_encoding += "gzip, deflate"; +#endif +#ifdef CPPHTTPLIB_ZSTD_SUPPORT + if (!accept_encoding.empty()) { accept_encoding += ", "; } + accept_encoding += "zstd"; #endif req.set_header("Accept-Encoding", accept_encoding); } @@ -8015,6 +8129,9 @@ inline Result ClientImpl::send_with_content_provider( req.headers = headers; req.path = path; req.progress = progress; + if (max_timeout_msec_ > 0) { + req.start_time_ = std::chrono::steady_clock::now(); + } auto error = Error::Success; @@ -8041,7 +8158,7 @@ inline bool ClientImpl::process_request(Stream &strm, Request &req, if (is_ssl()) { auto is_proxy_enabled = !proxy_host_.empty() && proxy_port_ != -1; if (!is_proxy_enabled) { - if (is_ssl_peer_could_be_closed(socket_.ssl)) { + if (detail::is_ssl_peer_could_be_closed(socket_.ssl, socket_.sock)) { error = Error::SSLPeerCouldBeClosed_; return false; } @@ -8166,12 +8283,13 @@ inline ContentProviderWithoutLength ClientImpl::get_multipart_content_provider( }; } -inline bool -ClientImpl::process_socket(const Socket &socket, - std::function callback) { +inline bool ClientImpl::process_socket( + const Socket &socket, + std::chrono::time_point start_time, + std::function callback) { return detail::process_client_socket( socket.sock, read_timeout_sec_, read_timeout_usec_, write_timeout_sec_, - write_timeout_usec_, std::move(callback)); + write_timeout_usec_, max_timeout_msec_, start_time, std::move(callback)); } inline bool ClientImpl::is_ssl() const { return false; } @@ -8195,6 +8313,9 @@ inline Result ClientImpl::Get(const std::string &path, const Headers &headers, req.path = path; req.headers = headers; req.progress = std::move(progress); + if (max_timeout_msec_ > 0) { + req.start_time_ = std::chrono::steady_clock::now(); + } return send_(std::move(req)); } @@ -8260,6 +8381,9 @@ inline Result ClientImpl::Get(const std::string &path, const Headers &headers, return content_receiver(data, data_length); }; req.progress = std::move(progress); + if (max_timeout_msec_ > 0) { + req.start_time_ = std::chrono::steady_clock::now(); + } return send_(std::move(req)); } @@ -8305,6 +8429,9 @@ inline Result ClientImpl::Head(const std::string &path, req.method = "HEAD"; req.headers = headers; req.path = path; + if (max_timeout_msec_ > 0) { + req.start_time_ = std::chrono::steady_clock::now(); + } return send_(std::move(req)); } @@ -8722,6 +8849,9 @@ inline Result ClientImpl::Delete(const std::string &path, req.headers = headers; req.path = path; req.progress = progress; + if (max_timeout_msec_ > 0) { + req.start_time_ = std::chrono::steady_clock::now(); + } if (!content_type.empty()) { req.set_header("Content-Type", content_type); } req.body.assign(body, content_length); @@ -8769,6 +8899,9 @@ inline Result ClientImpl::Options(const std::string &path, req.method = "OPTIONS"; req.headers = headers; req.path = path; + if (max_timeout_msec_ > 0) { + req.start_time_ = std::chrono::steady_clock::now(); + } return send_(std::move(req)); } @@ -8822,6 +8955,10 @@ inline void ClientImpl::set_write_timeout(time_t sec, time_t usec) { write_timeout_usec_ = usec; } +inline void ClientImpl::set_max_timeout(time_t msec) { + max_timeout_msec_ = msec; +} + inline void ClientImpl::set_basic_auth(const std::string &username, const std::string &password) { basic_auth_username_ = username; @@ -8947,7 +9084,7 @@ inline void ClientImpl::enable_server_hostname_verification(bool enabled) { } inline void ClientImpl::set_server_certificate_verifier( - std::function verifier) { + std::function verifier) { server_certificate_verifier_ = verifier; } #endif @@ -9000,22 +9137,13 @@ inline void ssl_delete(std::mutex &ctx_mutex, SSL *ssl, socket_t sock, // Note that it is not always possible to avoid SIGPIPE, this is merely a // best-efforts. if (shutdown_gracefully) { -#ifdef _WIN32 (void)(sock); - SSL_shutdown(ssl); -#else - timeval tv; - tv.tv_sec = 1; - tv.tv_usec = 0; - setsockopt(sock, SOL_SOCKET, SO_RCVTIMEO, - reinterpret_cast(&tv), sizeof(tv)); - - auto ret = SSL_shutdown(ssl); - while (ret == 0) { - std::this_thread::sleep_for(std::chrono::milliseconds{100}); - ret = SSL_shutdown(ssl); + // SSL_shutdown() returns 0 on first call (indicating close_notify alert + // sent) and 1 on subsequent call (indicating close_notify alert received) + if (SSL_shutdown(ssl) == 0) { + // Expected to return 1, but even if it doesn't, we free ssl + SSL_shutdown(ssl); } -#endif } std::lock_guard guard(ctx_mutex); @@ -9060,51 +9188,59 @@ inline bool process_server_socket_ssl( } template -inline bool -process_client_socket_ssl(SSL *ssl, socket_t sock, time_t read_timeout_sec, - time_t read_timeout_usec, time_t write_timeout_sec, - time_t write_timeout_usec, T callback) { +inline bool process_client_socket_ssl( + SSL *ssl, socket_t sock, time_t read_timeout_sec, time_t read_timeout_usec, + time_t write_timeout_sec, time_t write_timeout_usec, + time_t max_timeout_msec, + std::chrono::time_point start_time, T callback) { SSLSocketStream strm(sock, ssl, read_timeout_sec, read_timeout_usec, - write_timeout_sec, write_timeout_usec); + write_timeout_sec, write_timeout_usec, max_timeout_msec, + start_time); return callback(strm); } -class SSLInit { -public: - SSLInit() { - OPENSSL_init_ssl( - OPENSSL_INIT_LOAD_SSL_STRINGS | OPENSSL_INIT_LOAD_CRYPTO_STRINGS, NULL); - } -}; - // SSL socket stream implementation -inline SSLSocketStream::SSLSocketStream(socket_t sock, SSL *ssl, - time_t read_timeout_sec, - time_t read_timeout_usec, - time_t write_timeout_sec, - time_t write_timeout_usec) +inline SSLSocketStream::SSLSocketStream( + socket_t sock, SSL *ssl, time_t read_timeout_sec, time_t read_timeout_usec, + time_t write_timeout_sec, time_t write_timeout_usec, + time_t max_timeout_msec, + std::chrono::time_point start_time) : sock_(sock), ssl_(ssl), read_timeout_sec_(read_timeout_sec), read_timeout_usec_(read_timeout_usec), write_timeout_sec_(write_timeout_sec), - write_timeout_usec_(write_timeout_usec) { + write_timeout_usec_(write_timeout_usec), + max_timeout_msec_(max_timeout_msec), start_time_(start_time) { SSL_clear_mode(ssl, SSL_MODE_AUTO_RETRY); } inline SSLSocketStream::~SSLSocketStream() = default; inline bool SSLSocketStream::is_readable() const { - return detail::select_read(sock_, read_timeout_sec_, read_timeout_usec_) > 0; + return SSL_pending(ssl_) > 0; } -inline bool SSLSocketStream::is_writable() const { +inline bool SSLSocketStream::wait_readable() const { + if (max_timeout_msec_ <= 0) { + return select_read(sock_, read_timeout_sec_, read_timeout_usec_) > 0; + } + + time_t read_timeout_sec; + time_t read_timeout_usec; + calc_actual_timeout(max_timeout_msec_, duration(), read_timeout_sec_, + read_timeout_usec_, read_timeout_sec, read_timeout_usec); + + return select_read(sock_, read_timeout_sec, read_timeout_usec) > 0; +} + +inline bool SSLSocketStream::wait_writable() const { return select_write(sock_, write_timeout_sec_, write_timeout_usec_) > 0 && - is_socket_alive(sock_); + is_socket_alive(sock_) && !is_ssl_peer_could_be_closed(ssl_, sock_); } inline ssize_t SSLSocketStream::read(char *ptr, size_t size) { if (SSL_pending(ssl_) > 0) { return SSL_read(ssl_, ptr, static_cast(size)); - } else if (is_readable()) { + } else if (wait_readable()) { auto ret = SSL_read(ssl_, ptr, static_cast(size)); if (ret < 0) { auto err = SSL_get_error(ssl_, ret); @@ -9118,7 +9254,7 @@ inline ssize_t SSLSocketStream::read(char *ptr, size_t size) { #endif if (SSL_pending(ssl_) > 0) { return SSL_read(ssl_, ptr, static_cast(size)); - } else if (is_readable()) { + } else if (wait_readable()) { std::this_thread::sleep_for(std::chrono::microseconds{10}); ret = SSL_read(ssl_, ptr, static_cast(size)); if (ret >= 0) { return ret; } @@ -9129,12 +9265,13 @@ inline ssize_t SSLSocketStream::read(char *ptr, size_t size) { } } return ret; + } else { + return -1; } - return -1; } inline ssize_t SSLSocketStream::write(const char *ptr, size_t size) { - if (is_writable()) { + if (wait_writable()) { auto handle_size = static_cast( std::min(size, (std::numeric_limits::max)())); @@ -9149,7 +9286,7 @@ inline ssize_t SSLSocketStream::write(const char *ptr, size_t size) { #else while (--n >= 0 && err == SSL_ERROR_WANT_WRITE) { #endif - if (is_writable()) { + if (wait_writable()) { std::this_thread::sleep_for(std::chrono::microseconds{10}); ret = SSL_write(ssl_, ptr, static_cast(handle_size)); if (ret >= 0) { return ret; } @@ -9176,7 +9313,11 @@ inline void SSLSocketStream::get_local_ip_and_port(std::string &ip, inline socket_t SSLSocketStream::socket() const { return sock_; } -static SSLInit sslinit_; +inline time_t SSLSocketStream::duration() const { + return std::chrono::duration_cast( + std::chrono::steady_clock::now() - start_time_) + .count(); +} } // namespace detail @@ -9416,16 +9557,22 @@ inline bool SSLClient::create_and_connect_socket(Socket &socket, Error &error) { } // Assumes that socket_mutex_ is locked and that there are no requests in flight -inline bool SSLClient::connect_with_proxy(Socket &socket, Response &res, - bool &success, Error &error) { +inline bool SSLClient::connect_with_proxy( + Socket &socket, + std::chrono::time_point start_time, + Response &res, bool &success, Error &error) { success = true; Response proxy_res; if (!detail::process_client_socket( socket.sock, read_timeout_sec_, read_timeout_usec_, - write_timeout_sec_, write_timeout_usec_, [&](Stream &strm) { + write_timeout_sec_, write_timeout_usec_, max_timeout_msec_, + start_time, [&](Stream &strm) { Request req2; req2.method = "CONNECT"; req2.path = host_and_port_; + if (max_timeout_msec_ > 0) { + req2.start_time_ = std::chrono::steady_clock::now(); + } return process_request(strm, req2, proxy_res, false, error); })) { // Thread-safe to close everything because we are assuming there are no @@ -9445,7 +9592,8 @@ inline bool SSLClient::connect_with_proxy(Socket &socket, Response &res, proxy_res = Response(); if (!detail::process_client_socket( socket.sock, read_timeout_sec_, read_timeout_usec_, - write_timeout_sec_, write_timeout_usec_, [&](Stream &strm) { + write_timeout_sec_, write_timeout_usec_, max_timeout_msec_, + start_time, [&](Stream &strm) { Request req3; req3.method = "CONNECT"; req3.path = host_and_port_; @@ -9453,6 +9601,9 @@ inline bool SSLClient::connect_with_proxy(Socket &socket, Response &res, req3, auth, 1, detail::random_string(10), proxy_digest_auth_username_, proxy_digest_auth_password_, true)); + if (max_timeout_msec_ > 0) { + req3.start_time_ = std::chrono::steady_clock::now(); + } return process_request(strm, req3, proxy_res, false, error); })) { // Thread-safe to close everything because we are assuming there are @@ -9536,12 +9687,18 @@ inline bool SSLClient::initialize_ssl(Socket &socket, Error &error) { } if (server_certificate_verification_) { + auto verification_status = SSLVerifierResponse::NoDecisionMade; + if (server_certificate_verifier_) { - if (!server_certificate_verifier_(ssl2)) { - error = Error::SSLServerVerification; - return false; - } - } else { + verification_status = server_certificate_verifier_(ssl2); + } + + if (verification_status == SSLVerifierResponse::CertificateRejected) { + error = Error::SSLServerVerification; + return false; + } + + if (verification_status == SSLVerifierResponse::NoDecisionMade) { verify_result_ = SSL_get_verify_result(ssl2); if (verify_result_ != X509_V_OK) { @@ -9608,13 +9765,15 @@ inline void SSLClient::shutdown_ssl_impl(Socket &socket, assert(socket.ssl == nullptr); } -inline bool -SSLClient::process_socket(const Socket &socket, - std::function callback) { +inline bool SSLClient::process_socket( + const Socket &socket, + std::chrono::time_point start_time, + std::function callback) { assert(socket.ssl); return detail::process_client_socket_ssl( socket.ssl, socket.sock, read_timeout_sec_, read_timeout_usec_, - write_timeout_sec_, write_timeout_usec_, std::move(callback)); + write_timeout_sec_, write_timeout_usec_, max_timeout_msec_, start_time, + std::move(callback)); } inline bool SSLClient::is_ssl() const { return true; } @@ -9651,8 +9810,8 @@ SSLClient::verify_host_with_subject_alt_name(X509 *server_cert) const { auto type = GEN_DNS; - struct in6_addr addr6{}; - struct in_addr addr{}; + struct in6_addr addr6 = {}; + struct in_addr addr = {}; size_t addr_len = 0; #ifndef __MINGW32__ @@ -10300,7 +10459,7 @@ inline void Client::enable_server_hostname_verification(bool enabled) { } inline void Client::set_server_certificate_verifier( - std::function verifier) { + std::function verifier) { cli_->set_server_certificate_verifier(verifier); } #endif @@ -10344,8 +10503,4 @@ inline SSL_CTX *Client::ssl_context() const { } // namespace httplib -#if defined(_WIN32) && defined(CPPHTTPLIB_USE_POLL) -#undef poll -#endif - #endif // CPPHTTPLIB_HTTPLIB_H diff --git a/examples/server/public/index.html.gz b/examples/server/public/index.html.gz index 141e80920..674e22757 100644 Binary files a/examples/server/public/index.html.gz and b/examples/server/public/index.html.gz differ diff --git a/examples/server/public_legacy/json-schema-to-grammar.mjs b/examples/server/public_legacy/json-schema-to-grammar.mjs index e67bb15c1..f767ce7b7 100644 --- a/examples/server/public_legacy/json-schema-to-grammar.mjs +++ b/examples/server/public_legacy/json-schema-to-grammar.mjs @@ -1,5 +1,5 @@ // WARNING: This file was ported from json_schema_to_grammar.py, please fix bugs / add features there first. -const SPACE_RULE = '| " " | "\\n" [ \\t]{0,20}'; +const SPACE_RULE = '| " " | "\\n"{1,2} [ \\t]{0,20}'; function _buildRepetition(itemRule, minItems, maxItems, opts={}) { if (minItems === 0 && maxItems === 1) { diff --git a/examples/server/server.cpp b/examples/server/server.cpp index 0718806c8..d87bda1a0 100644 --- a/examples/server/server.cpp +++ b/examples/server/server.cpp @@ -42,7 +42,7 @@ enum stop_type { STOP_TYPE_LIMIT, }; -// state diagram: https://github.com/ggerganov/llama.cpp/pull/9283 +// state diagram: https://github.com/ggml-org/llama.cpp/pull/9283 enum slot_state { SLOT_STATE_IDLE, SLOT_STATE_STARTED, // TODO: this state is only used for setting up the initial prompt processing; maybe merge it with launch_slot_with_task in the future @@ -131,9 +131,10 @@ struct slot_params { lora.push_back({{"id", i}, {"scale", this->lora[i].scale}}); } - std::vector grammar_trigger_words; - for (const auto & trigger : sampling.grammar_trigger_words) { - grammar_trigger_words.push_back(trigger.word); + auto grammar_triggers = json::array(); + for (const auto & trigger : sampling.grammar_triggers) { + server_grammar_trigger ct(std::move(trigger)); + grammar_triggers.push_back(ct.to_json()); } return json { @@ -170,9 +171,10 @@ struct slot_params { {"n_probs", sampling.n_probs}, {"min_keep", sampling.min_keep}, {"grammar", sampling.grammar}, - {"grammar_trigger_words", grammar_trigger_words}, - {"grammar_trigger_tokens", sampling.grammar_trigger_tokens}, + {"grammar_lazy", sampling.grammar_lazy}, + {"grammar_triggers", grammar_triggers}, {"preserved_tokens", sampling.preserved_tokens}, + {"chat_format", common_chat_format_name(oaicompat_chat_format)}, {"samplers", samplers}, {"speculative.n_max", speculative.n_max}, {"speculative.n_min", speculative.n_min}, @@ -273,7 +275,7 @@ struct server_task { params.speculative.p_min = json_value(data, "speculative.p_min", defaults.speculative.p_min); params.speculative.n_min = std::min(params.speculative.n_max, params.speculative.n_min); - params.speculative.n_min = std::max(params.speculative.n_min, 2); + params.speculative.n_min = std::max(params.speculative.n_min, 0); params.speculative.n_max = std::max(params.speculative.n_max, 0); // Use OpenAI API logprobs only if n_probs wasn't provided @@ -328,9 +330,6 @@ struct server_task { } // process "json_schema" and "grammar" - if (data.contains("json_schema") && !data.at("json_schema").is_null() && data.contains("grammar") && !data.at("grammar").is_null()) { - throw std::runtime_error("Either \"json_schema\" or \"grammar\" can be specified, but not both"); - } if (data.contains("json_schema") && !data.contains("grammar")) { try { auto schema = json_value(data, "json_schema", json::object()); @@ -358,24 +357,6 @@ struct server_task { } { - const auto grammar_triggers = data.find("grammar_triggers"); - if (grammar_triggers != data.end()) { - for (const auto & t : *grammar_triggers) { - common_grammar_trigger trigger; - trigger.word = t.at("word"); - trigger.at_start = t.at("at_start"); - - auto ids = common_tokenize(vocab, trigger.word, /* add_special= */ false, /* parse_special= */ true); - if (ids.size() == 1) { - SRV_DBG("Grammar trigger token: %d (`%s`)\n", ids[0], trigger.word.c_str()); - params.sampling.grammar_trigger_tokens.push_back(ids[0]); - params.sampling.preserved_tokens.insert(ids[0]); - continue; - } - SRV_DBG("Grammar trigger word: `%s`\n", trigger.word.c_str()); - params.sampling.grammar_trigger_words.push_back(trigger); - } - } const auto preserved_tokens = data.find("preserved_tokens"); if (preserved_tokens != data.end()) { for (const auto & t : *preserved_tokens) { @@ -385,12 +366,39 @@ struct server_task { params.sampling.preserved_tokens.insert(ids[0]); } else { // This may happen when using a tool call style meant for a model with special tokens to preserve on a model without said tokens. - SRV_WRN("Not preserved because more than 1 token (wrong chat template override?): %s\n", t.get().c_str()); + SRV_DBG("Not preserved because more than 1 token: %s\n", t.get().c_str()); } } } - if (params.sampling.grammar_lazy) { - GGML_ASSERT(params.sampling.grammar_trigger_tokens.size() > 0 || params.sampling.grammar_trigger_words.size() > 0); + const auto grammar_triggers = data.find("grammar_triggers"); + if (grammar_triggers != data.end()) { + for (const auto & t : *grammar_triggers) { + server_grammar_trigger ct(t); + if (ct.value.type == COMMON_GRAMMAR_TRIGGER_TYPE_WORD) { + const auto & word = ct.value.value; + auto ids = common_tokenize(vocab, word, /* add_special= */ false, /* parse_special= */ true); + if (ids.size() == 1) { + auto token = ids[0]; + if (std::find(params.sampling.preserved_tokens.begin(), params.sampling.preserved_tokens.end(), (llama_token) token) == params.sampling.preserved_tokens.end()) { + throw std::runtime_error("Grammar trigger word should be marked as preserved token: " + word); + } + SRV_DBG("Grammar trigger token: %d (`%s`)\n", token, word.c_str()); + common_grammar_trigger trigger; + trigger.type = COMMON_GRAMMAR_TRIGGER_TYPE_TOKEN; + trigger.value = word; + trigger.token = token; + params.sampling.grammar_triggers.push_back(std::move(trigger)); + } else { + SRV_DBG("Grammar trigger word: `%s`\n", word.c_str()); + params.sampling.grammar_triggers.push_back({COMMON_GRAMMAR_TRIGGER_TYPE_WORD, word}); + } + } else { + params.sampling.grammar_triggers.push_back(std::move(ct.value)); + } + } + } + if (params.sampling.grammar_lazy && params.sampling.grammar_triggers.empty()) { + throw std::runtime_error("Error: no triggers set for lazy grammar!"); } } @@ -482,8 +490,12 @@ struct result_timings { double predicted_per_token_ms; double predicted_per_second; + // Optional speculative metrics - only included when > 0 + int32_t draft_n = 0; + int32_t draft_n_accepted = 0; + json to_json() const { - return { + json base = { {"prompt_n", prompt_n}, {"prompt_ms", prompt_ms}, {"prompt_per_token_ms", prompt_per_token_ms}, @@ -494,6 +506,13 @@ struct result_timings { {"predicted_per_token_ms", predicted_per_token_ms}, {"predicted_per_second", predicted_per_second}, }; + + if (draft_n > 0) { + base["draft_n"] = draft_n; + base["draft_n_accepted"] = draft_n_accepted; + } + + return base; } }; @@ -724,9 +743,19 @@ struct server_task_result_cmpl_final : server_task_result { msg.content = content; } - json tool_calls; + json message { + {"role", "assistant"}, + }; + if (!msg.reasoning_content.empty()) { + message["reasoning_content"] = msg.reasoning_content; + } + if (msg.content.empty() && !msg.tool_calls.empty()) { + message["content"] = json(); + } else { + message["content"] = msg.content; + } if (!msg.tool_calls.empty()) { - tool_calls = json::array(); + auto tool_calls = json::array(); for (const auto & tc : msg.tool_calls) { tool_calls.push_back({ {"type", "function"}, @@ -734,18 +763,13 @@ struct server_task_result_cmpl_final : server_task_result { {"name", tc.name}, {"arguments", tc.arguments}, }}, - {"id", tc.id}, + // Some templates generate and require an id (sometimes in a very specific format, e.g. Mistral Nemo). + // We only generate a random id for the ones that don't generate one by themselves + // (they also won't get to see it as their template likely doesn't use it, so it's all for the client) + {"id", tc.id.empty() ? gen_tool_call_id() : tc.id}, }); } - } - - json message { - {"content", msg.content}, - {"tool_calls", tool_calls}, - {"role", "assistant"}, - }; - if (!msg.tool_plan.empty()) { - message["tool_plan"] = msg.tool_plan; + message["tool_calls"] = tool_calls; } json choice { @@ -818,6 +842,11 @@ struct server_task_result_cmpl_final : server_task_result { ret.push_back({"timings", timings.to_json()}); } + // extra fields for debugging purposes + if (verbose) { + ret["__verbose"] = to_json_non_oaicompat(); + } + return ret; } }; @@ -1282,6 +1311,10 @@ struct server_slot { std::function callback_on_release; + // Speculative decoding stats + int32_t n_draft_total = 0; // Total draft tokens generated + int32_t n_draft_accepted = 0; // Draft tokens actually accepted + void reset() { SLT_DBG(*this, "%s", "\n"); @@ -1298,13 +1331,17 @@ struct server_slot { generated_tokens.clear(); generated_token_probs.clear(); + + // clear speculative decoding stats + n_draft_total = 0; + n_draft_accepted = 0; } bool is_non_causal() const { return task_type == SERVER_TASK_TYPE_EMBEDDING || task_type == SERVER_TASK_TYPE_RERANK; } - bool can_batch_with(server_slot & other_slot) { + bool can_batch_with(server_slot & other_slot) const { return is_non_causal() == other_slot.is_non_causal() && are_lora_equal(lora, other_slot.lora); } @@ -1364,6 +1401,12 @@ struct server_slot { timings.predicted_per_token_ms = t_token_generation / n_decoded; timings.predicted_per_second = 1e3 / t_token_generation * n_decoded; + // Add speculative metrics + if (n_draft_total > 0) { + timings.draft_n = n_draft_total; + timings.draft_n_accepted = n_draft_accepted; + } + return timings; } @@ -1411,6 +1454,15 @@ struct server_slot { t_prompt_processing, n_prompt_tokens_processed, t_prompt, n_prompt_second, t_token_generation, n_decoded, t_gen, n_gen_second, t_prompt_processing + t_token_generation, n_prompt_tokens_processed + n_decoded); + + if (n_draft_total > 0) { + const float draft_ratio = (float) n_draft_accepted / n_draft_total; + SLT_INF(*this, + "\n" + "draft acceptance rate = %0.5f (%5d accepted / %5d generated)\n", + draft_ratio, n_draft_accepted, n_draft_total + ); + } } json to_json() const { @@ -1600,6 +1652,10 @@ struct server_queue { while (true) { std::unique_lock lock(mutex_tasks); + if (!running) { + QUE_DBG("%s", "terminate\n"); + return; + } if (queue_tasks.empty()) { lock.unlock(); break; @@ -1620,11 +1676,11 @@ struct server_queue { QUE_DBG("%s", "waiting for new tasks\n"); { std::unique_lock lock(mutex_tasks); + if (!running) { + QUE_DBG("%s", "terminate\n"); + return; + } if (queue_tasks.empty()) { - if (!running) { - QUE_DBG("%s", "terminate\n"); - return; - } condition_tasks.wait(lock, [&]{ return (!queue_tasks.empty() || !running); }); @@ -1649,6 +1705,8 @@ private: }; struct server_response { + bool running = true; + // for keeping track of all tasks waiting for the result std::unordered_set waiting_task_ids; @@ -1703,6 +1761,10 @@ struct server_response { while (true) { std::unique_lock lock(mutex_results); condition_results.wait(lock, [&]{ + if (!running) { + SRV_DBG("%s : queue result stop\n", __func__); + std::terminate(); // we cannot return here since the caller is HTTP code + } return !queue_results.empty(); }); @@ -1733,6 +1795,10 @@ struct server_response { } std::cv_status cr_res = condition_results.wait_for(lock, std::chrono::seconds(timeout)); + if (!running) { + SRV_DBG("%s : queue result stop\n", __func__); + std::terminate(); // we cannot return here since the caller is HTTP code + } if (cr_res == std::cv_status::timeout) { return nullptr; } @@ -1762,6 +1828,12 @@ struct server_response { } } } + + // terminate the waiting loop + void terminate() { + running = false; + condition_results.notify_all(); + } }; struct server_context { @@ -1800,7 +1872,7 @@ struct server_context { // Necessary similarity of prompt for slot selection float slot_prompt_similarity = 0.0f; - common_chat_templates chat_templates; + common_chat_templates_ptr chat_templates; ~server_context() { // Clear any sampling context @@ -1821,7 +1893,7 @@ struct server_context { } bool load_model(const common_params & params) { - SRV_INF("loading model '%s'\n", params.model.c_str()); + SRV_INF("loading model '%s'\n", params.model.path.c_str()); params_base = params; @@ -1831,7 +1903,7 @@ struct server_context { ctx = llama_init.context.get(); if (model == nullptr) { - SRV_ERR("failed to load model, '%s'\n", params_base.model.c_str()); + SRV_ERR("failed to load model, '%s'\n", params_base.model.path.c_str()); return false; } @@ -1842,31 +1914,32 @@ struct server_context { add_bos_token = llama_vocab_get_add_bos(vocab); has_eos_token = llama_vocab_eos(vocab) != LLAMA_TOKEN_NULL; - if (!params_base.speculative.model.empty() || !params_base.speculative.hf_repo.empty()) { - SRV_INF("loading draft model '%s'\n", params_base.speculative.model.c_str()); + if (!params_base.speculative.model.path.empty() || !params_base.speculative.model.hf_repo.empty()) { + SRV_INF("loading draft model '%s'\n", params_base.speculative.model.path.c_str()); auto params_dft = params_base; params_dft.devices = params_base.speculative.devices; - params_dft.hf_file = params_base.speculative.hf_file; - params_dft.hf_repo = params_base.speculative.hf_repo; params_dft.model = params_base.speculative.model; - params_dft.model_url = params_base.speculative.model_url; params_dft.n_ctx = params_base.speculative.n_ctx == 0 ? params_base.n_ctx / params_base.n_parallel : params_base.speculative.n_ctx; params_dft.n_gpu_layers = params_base.speculative.n_gpu_layers; params_dft.n_parallel = 1; + // force F16 KV cache for the draft model for extra performance + params_dft.cache_type_k = GGML_TYPE_F16; + params_dft.cache_type_v = GGML_TYPE_F16; + llama_init_dft = common_init_from_params(params_dft); model_dft = llama_init_dft.model.get(); if (model_dft == nullptr) { - SRV_ERR("failed to load draft model, '%s'\n", params_base.speculative.model.c_str()); + SRV_ERR("failed to load draft model, '%s'\n", params_base.speculative.model.path.c_str()); return false; } if (!common_speculative_are_compatible(ctx, llama_init_dft.context.get())) { - SRV_ERR("the draft model '%s' is not compatible with the target model '%s'\n", params_base.speculative.model.c_str(), params_base.model.c_str()); + SRV_ERR("the draft model '%s' is not compatible with the target model '%s'\n", params_base.speculative.model.path.c_str(), params_base.model.path.c_str()); return false; } @@ -1876,53 +1949,22 @@ struct server_context { cparams_dft = common_context_params_to_llama(params_dft); cparams_dft.n_batch = n_ctx_dft; - // force F16 KV cache for the draft model for extra performance - cparams_dft.type_k = GGML_TYPE_F16; - cparams_dft.type_v = GGML_TYPE_F16; - // the context is not needed - we will create one for each slot llama_init_dft.context.reset(); } - if (params_base.chat_template.empty() && !validate_builtin_chat_template(params.use_jinja)) { + chat_templates = common_chat_templates_init(model, params_base.chat_template); + try { + common_chat_format_example(chat_templates.get(), params.use_jinja); + } catch (const std::exception & e) { + SRV_WRN("%s: Chat template parsing error: %s\n", __func__, e.what()); SRV_WRN("%s: The chat template that comes with this model is not yet supported, falling back to chatml. This may cause the model to output suboptimal responses\n", __func__); - chat_templates = common_chat_templates_from_model(model, "chatml"); - } else { - chat_templates = common_chat_templates_from_model(model, params_base.chat_template); + chat_templates = common_chat_templates_init(model, "chatml"); } - GGML_ASSERT(chat_templates.template_default.get() != nullptr); return true; } - bool validate_builtin_chat_template(bool use_jinja) const { - llama_chat_message chat[] = {{"user", "test"}}; - - if (use_jinja) { - auto templates = common_chat_templates_from_model(model, ""); - common_chat_inputs inputs; - inputs.messages = json::array({{ - {"role", "user"}, - {"content", "test"}, - }}); - GGML_ASSERT(templates.template_default); - try { - common_chat_params_init(*templates.template_default, inputs); - if (templates.template_tool_use) { - common_chat_params_init(*templates.template_tool_use, inputs); - } - return true; - } catch (const std::exception & e) { - SRV_ERR("failed to apply template: %s\n", e.what()); - return false; - } - } else { - const char * tmpl = llama_model_chat_template(model, /* name */ nullptr); - const int32_t chat_res = llama_chat_apply_template(tmpl, chat, 1, true, nullptr, 0); - return chat_res > 0; - } - } - void init() { const int32_t n_ctx_slot = n_ctx / params_base.n_parallel; @@ -2051,6 +2093,18 @@ struct server_context { return ret; } + bool can_be_detokenized(const struct llama_context * ctx, const std::vector & tokens) { + const llama_model * model = llama_get_model(ctx); + const llama_vocab * vocab = llama_model_get_vocab(model); + const int32_t n_vocab = llama_vocab_n_tokens(vocab); + for (const auto & token : tokens) { + if (token < 0 || token >= n_vocab) { + return false; + } + } + return true; + } + bool launch_slot_with_task(server_slot & slot, const server_task & task) { slot.reset(); slot.id_task = task.id; @@ -2065,12 +2119,17 @@ struct server_context { slot.lora = task.params.lora; } + bool can_detokenize = can_be_detokenized(ctx, slot.prompt_tokens); + if (!can_detokenize) { + send_error(task, "Prompt contains invalid tokens", ERROR_TYPE_INVALID_REQUEST); + return false; + } SLT_DBG(slot, "launching slot : %s\n", safe_json_to_str(slot.to_json()).c_str()); if (slot.n_predict > 0 && slot.params.n_predict > slot.n_predict) { // Might be better to reject the request with a 400 ? + SLT_WRN(slot, "n_predict = %d exceeds server configuration, setting to %d\n", slot.params.n_predict, slot.n_predict); slot.params.n_predict = slot.n_predict; - SLT_WRN(slot, "n_predict = %d exceeds server configuration, setting to %d", slot.n_predict, slot.n_predict); } if (slot.params.ignore_eos && has_eos_token) { @@ -2107,7 +2166,7 @@ struct server_context { SRV_DBG("%s", "clearing KV cache\n"); // clear the entire KV cache - llama_kv_cache_clear(ctx); + llama_kv_self_clear(ctx); clean_kv_cache = false; } @@ -2172,14 +2231,6 @@ struct server_context { } if (slot.has_new_line) { - // if we have already seen a new line, we stop after a certain time limit - if (slot.params.t_max_predict_ms > 0 && (ggml_time_us() - slot.t_start_generation > 1000.0f*slot.params.t_max_predict_ms)) { - slot.stop = STOP_TYPE_LIMIT; - slot.has_next_token = false; - - SLT_DBG(slot, "stopped by time limit, n_decoded = %d, t_max_predict_ms = %d ms\n", slot.n_decoded, (int) slot.params.t_max_predict_ms); - } - // require that each new line has a whitespace prefix (i.e. indentation) of at least slot.params.n_indent if (slot.params.n_indent > 0) { // check the current indentation @@ -2218,6 +2269,14 @@ struct server_context { // check if there is a new line in the generated text if (result.text_to_send.find('\n') != std::string::npos) { slot.has_new_line = true; + + // if we have seen a new line, we stop after a certain time limit, but only upon another new line + if (slot.params.t_max_predict_ms > 0 && (ggml_time_us() - slot.t_start_generation > 1000.0f*slot.params.t_max_predict_ms)) { + slot.stop = STOP_TYPE_LIMIT; + slot.has_next_token = false; + + SLT_DBG(slot, "stopped by time limit, n_decoded = %d, t_max_predict_ms = %d ms\n", slot.n_decoded, (int) slot.params.t_max_predict_ms); + } } // if context shift is disabled, we stop when it reaches the context limit @@ -2275,7 +2334,7 @@ struct server_context { for (size_t i = 0; i < std::min(max_probs, n_probs); i++) { result.probs.push_back({ cur_p->data[i].id, - common_detokenize(ctx, {cur_p->data[i].id}, special), + common_token_to_piece(ctx, cur_p->data[i].id, special), cur_p->data[i].p }); } @@ -2297,7 +2356,7 @@ struct server_context { for (size_t i = 0; i < std::min(n_vocab, n_probs); i++) { result.probs.push_back({ cur[i].id, - common_detokenize(ctx, {cur[i].id}, special), + common_token_to_piece(ctx, cur[i].id, special), cur[i].p }); } @@ -2649,8 +2708,8 @@ struct server_context { res->n_tasks_deferred = queue_tasks.queue_tasks_deferred.size(); res->t_start = metrics.t_start; - res->kv_cache_tokens_count = llama_get_kv_cache_token_count(ctx); - res->kv_cache_used_cells = llama_get_kv_cache_used_cells(ctx); + res->kv_cache_tokens_count = llama_kv_self_n_tokens(ctx); + res->kv_cache_used_cells = llama_kv_self_used_cells(ctx); res->n_prompt_tokens_processed_total = metrics.n_prompt_tokens_processed_total; res->t_prompt_processing_total = metrics.t_prompt_processing_total; @@ -2766,7 +2825,7 @@ struct server_context { // Erase token cache const size_t n_erased = slot->cache_tokens.size(); - llama_kv_cache_seq_rm(ctx, slot->id, -1, -1); + llama_kv_self_seq_rm(ctx, slot->id, -1, -1); slot->cache_tokens.clear(); auto res = std::make_unique(); @@ -2834,8 +2893,8 @@ struct server_context { SLT_WRN(slot, "slot context shift, n_keep = %d, n_left = %d, n_discard = %d\n", n_keep, n_left, n_discard); - llama_kv_cache_seq_rm (ctx, slot.id, n_keep , n_keep + n_discard); - llama_kv_cache_seq_add(ctx, slot.id, n_keep + n_discard, slot.n_past, -n_discard); + llama_kv_self_seq_rm (ctx, slot.id, n_keep , n_keep + n_discard); + llama_kv_self_seq_add(ctx, slot.id, n_keep + n_discard, slot.n_past, -n_discard); if (slot.params.cache_prompt) { for (size_t i = n_keep + n_discard; i < slot.cache_tokens.size(); i++) { @@ -3026,8 +3085,8 @@ struct server_context { const int64_t kv_shift = (int64_t) head_p - (int64_t) head_c; - llama_kv_cache_seq_rm (ctx, slot.id, head_p, head_c); - llama_kv_cache_seq_add(ctx, slot.id, head_c, -1, kv_shift); + llama_kv_self_seq_rm (ctx, slot.id, head_p, head_c); + llama_kv_self_seq_add(ctx, slot.id, head_c, head_c + n_match, kv_shift); for (size_t i = 0; i < n_match; i++) { slot.cache_tokens[head_p + i] = slot.cache_tokens[head_c + i]; @@ -3065,9 +3124,9 @@ struct server_context { } // keep only the common part - if (!llama_kv_cache_seq_rm(ctx, slot.id, slot.n_past, -1)) { + if (!llama_kv_self_seq_rm(ctx, slot.id, slot.n_past, -1)) { // could not partially delete (likely using a non-Transformer model) - llama_kv_cache_seq_rm(ctx, slot.id, -1, -1); + llama_kv_self_seq_rm(ctx, slot.id, -1, -1); // there is no common part left slot.n_past = 0; @@ -3279,6 +3338,9 @@ struct server_context { llama_tokens draft = common_speculative_gen_draft(slot.spec, params_spec, slot.cache_tokens, id); + // keep track of total number of tokens generated in the draft + slot.n_draft_total += draft.size(); + // ignore small drafts if (slot.params.speculative.n_min > (int) draft.size()) { SLT_DBG(slot, "ignoring small draft: %d < %d\n", (int) draft.size(), slot.params.speculative.n_min); @@ -3304,10 +3366,13 @@ struct server_context { slot.n_past += ids.size(); slot.n_decoded += ids.size(); + // update how many tokens out of draft was accepted + slot.n_draft_accepted += ids.size() - 1; + slot.cache_tokens.push_back(id); slot.cache_tokens.insert(slot.cache_tokens.end(), ids.begin(), ids.end() - 1); - llama_kv_cache_seq_rm(ctx, slot.id, slot.n_past, -1); + llama_kv_self_seq_rm(ctx, slot.id, slot.n_past, -1); for (size_t i = 0; i < ids.size(); ++i) { completion_token_output result; @@ -3649,7 +3714,7 @@ int main(int argc, char ** argv) { }, { {"name", "n_busy_slots_per_decode"}, {"help", "Average number of busy slots per llama_decode() call"}, - {"value", (float) res_metrics->n_busy_slots_total / (float) res_metrics->n_decode_total} + {"value", (float) res_metrics->n_busy_slots_total / std::max((float) res_metrics->n_decode_total, 1.f)} }}}, {"gauge", {{ {"name", "prompt_tokens_seconds"}, @@ -3814,14 +3879,16 @@ int main(int argc, char ** argv) { json data = { { "default_generation_settings", ctx_server.default_generation_settings_for_props }, { "total_slots", ctx_server.params_base.n_parallel }, - { "model_path", ctx_server.params_base.model }, - { "chat_template", ctx_server.chat_templates.template_default->source() }, - { "bos_token", ctx_server.chat_templates.template_default->bos_token() }, - { "eos_token", ctx_server.chat_templates.template_default->eos_token() }, + { "model_path", ctx_server.params_base.model.path }, + { "chat_template", common_chat_templates_source(ctx_server.chat_templates.get()) }, + { "bos_token", common_token_to_piece(ctx_server.ctx, llama_vocab_bos(ctx_server.vocab), /* special= */ true)}, + { "eos_token", common_token_to_piece(ctx_server.ctx, llama_vocab_eos(ctx_server.vocab), /* special= */ true)}, { "build_info", build_info }, }; - if (ctx_server.params_base.use_jinja && ctx_server.chat_templates.template_tool_use) { - data["chat_template_tool_use"] = ctx_server.chat_templates.template_tool_use->source(); + if (ctx_server.params_base.use_jinja) { + if (auto tool_use_src = common_chat_templates_source(ctx_server.chat_templates.get(), "tool_use")) { + data["chat_template_tool_use"] = tool_use_src; + } } res_ok(res, data); @@ -3840,6 +3907,21 @@ int main(int argc, char ** argv) { res_ok(res, {{ "success", true }}); }; + const auto handle_api_show = [&ctx_server, &res_ok](const httplib::Request &, httplib::Response & res) { + json data = { + { + "template", common_chat_templates_source(ctx_server.chat_templates.get()), + }, + { + "model_info", { + { "llama.context_length", ctx_server.slots.back().n_ctx, }, + } + }, + }; + + res_ok(res, data); + }; + // handle completion-like requests (completion, chat, infill) // we can optionally provide a custom format for partial results and final results const auto handle_completions_impl = [&ctx_server, &res_error, &res_ok]( @@ -4056,7 +4138,7 @@ int main(int argc, char ** argv) { } auto body = json::parse(req.body); - json data = oaicompat_completion_params_parse(body, params.use_jinja, ctx_server.chat_templates); + json data = oaicompat_completion_params_parse(body, params.use_jinja, params.reasoning_format, ctx_server.chat_templates.get()); return handle_completions_impl( SERVER_TASK_TYPE_COMPLETION, @@ -4069,7 +4151,7 @@ int main(int argc, char ** argv) { // same with handle_chat_completions, but without inference part const auto handle_apply_template = [&ctx_server, ¶ms, &res_ok](const httplib::Request & req, httplib::Response & res) { auto body = json::parse(req.body); - json data = oaicompat_completion_params_parse(body, params.use_jinja, ctx_server.chat_templates); + json data = oaicompat_completion_params_parse(body, params.use_jinja, params.reasoning_format, ctx_server.chat_templates.get()); res_ok(res, {{ "prompt", std::move(data.at("prompt")) }}); }; @@ -4078,7 +4160,7 @@ int main(int argc, char ** argv) { {"object", "list"}, {"data", { { - {"id", params.model_alias.empty() ? params.model : params.model_alias}, + {"id", params.model_alias.empty() ? params.model.path : params.model_alias}, {"object", "model"}, {"created", std::time(0)}, {"owned_by", "llamacpp"}, @@ -4256,6 +4338,11 @@ int main(int argc, char ** argv) { // return; //} + // if true, use TEI API format, otherwise use Jina API format + // Jina: https://jina.ai/reranker/ + // TEI: https://huggingface.github.io/text-embeddings-inference/#/Text%20Embeddings%20Inference/rerank + bool is_tei_format = body.contains("texts"); + json query; if (body.count("query") == 1) { query = body.at("query"); @@ -4268,7 +4355,8 @@ int main(int argc, char ** argv) { return; } - std::vector documents = json_value(body, "documents", std::vector()); + std::vector documents = json_value(body, "documents", + json_value(body, "texts", std::vector())); if (documents.empty()) { res_error(res, format_error_response("\"documents\" must be a non-empty string array", ERROR_TYPE_INVALID_REQUEST)); return; @@ -4313,7 +4401,12 @@ int main(int argc, char ** argv) { } // write JSON response - json root = format_response_rerank(body, responses); + json root = format_response_rerank( + body, + responses, + is_tei_format, + documents); + res_ok(res, root); }; @@ -4393,6 +4486,7 @@ int main(int argc, char ** argv) { svr->Get ("/metrics", handle_metrics); svr->Get ("/props", handle_props); svr->Post("/props", handle_props_change); + svr->Post("/api/show", handle_api_show); svr->Get ("/models", handle_models); // public endpoint (no API key check) svr->Get ("/v1/models", handle_models); // public endpoint (no API key check) svr->Post("/completion", handle_completions); // legacy @@ -4429,27 +4523,34 @@ int main(int argc, char ** argv) { svr->new_task_queue = [¶ms] { return new httplib::ThreadPool(params.n_threads_http); }; // clean up function, to be called before exit - auto clean_up = [&svr]() { + auto clean_up = [&svr, &ctx_server]() { + SRV_INF("%s: cleaning up before exit...\n", __func__); svr->stop(); + ctx_server.queue_results.terminate(); llama_backend_free(); }; - // bind HTTP listen port bool was_bound = false; - if (params.port == 0) { - int bound_port = svr->bind_to_any_port(params.hostname); - if ((was_bound = (bound_port >= 0))) { - params.port = bound_port; - } + if (string_ends_with(std::string(params.hostname), ".sock")) { + LOG_INF("%s: setting address family to AF_UNIX\n", __func__); + svr->set_address_family(AF_UNIX); + // bind_to_port requires a second arg, any value other than 0 should + // simply get ignored + was_bound = svr->bind_to_port(params.hostname, 8080); } else { - was_bound = svr->bind_to_port(params.hostname, params.port); + LOG_INF("%s: binding port with default address family\n", __func__); + // bind HTTP listen port + if (params.port == 0) { + int bound_port = svr->bind_to_any_port(params.hostname); + if ((was_bound = (bound_port >= 0))) { + params.port = bound_port; + } + } else { + was_bound = svr->bind_to_port(params.hostname, params.port); + } } if (!was_bound) { - //LOG_ERROR("couldn't bind HTTP server socket", { - // {"hostname", params.hostname}, - // {"port", params.port}, - //}); LOG_ERR("%s: couldn't bind HTTP server socket, hostname: %s, port: %d\n", __func__, params.hostname.c_str(), params.port); clean_up(); return 1; @@ -4478,8 +4579,8 @@ int main(int argc, char ** argv) { // print sample chat example to make it clear which template is used LOG_INF("%s: chat template, chat_template: %s, example_format: '%s'\n", __func__, - ctx_server.chat_templates.template_default->source().c_str(), - common_chat_format_example(*ctx_server.chat_templates.template_default, ctx_server.params_base.use_jinja).c_str()); + common_chat_templates_source(ctx_server.chat_templates.get()), + common_chat_format_example(ctx_server.chat_templates.get(), ctx_server.params_base.use_jinja).c_str()); ctx_server.queue_tasks.on_new_task([&ctx_server](const server_task & task) { ctx_server.process_single_task(task); @@ -4490,13 +4591,10 @@ int main(int argc, char ** argv) { }); shutdown_handler = [&](int) { + // this will unblock start_loop() ctx_server.queue_tasks.terminate(); }; - LOG_INF("%s: server is listening on http://%s:%d - starting the main loop\n", __func__, params.hostname.c_str(), params.port); - - ctx_server.queue_tasks.start_loop(); - #if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__)) struct sigaction sigint_action; sigint_action.sa_handler = signal_handler; @@ -4511,6 +4609,11 @@ int main(int argc, char ** argv) { SetConsoleCtrlHandler(reinterpret_cast(console_ctrl_handler), true); #endif + LOG_INF("%s: server is listening on http://%s:%d - starting the main loop\n", __func__, params.hostname.c_str(), params.port); + + // this call blocks the main thread until queue_tasks.terminate() is called + ctx_server.queue_tasks.start_loop(); + clean_up(); t.join(); diff --git a/examples/server/tests/README.md b/examples/server/tests/README.md index 1de0eb30e..652dea038 100644 --- a/examples/server/tests/README.md +++ b/examples/server/tests/README.md @@ -17,7 +17,7 @@ To mitigate it, you can increase values in `n_predict`, `kv_size`. ```shell cd ../../.. -cmake -B build -DLLAMA_CURL=ON +cmake -B build cmake --build build --target llama-server ``` @@ -48,7 +48,7 @@ DEBUG=1 ./tests.sh -s -v -x To run all the tests in a file: ```shell -./tests.sh unit/test_chat_completion.py.py -v -x +./tests.sh unit/test_chat_completion.py -v -x ``` To run a single test: diff --git a/examples/server/tests/unit/test_chat_completion.py b/examples/server/tests/unit/test_chat_completion.py index f23d5cff4..491cb3a5d 100644 --- a/examples/server/tests/unit/test_chat_completion.py +++ b/examples/server/tests/unit/test_chat_completion.py @@ -21,6 +21,8 @@ def create_server(): (None, "Book", "What is the best book", 8, "^ blue", 23, 8, "length", True, "This is not a chat template, it is"), ("codellama70b", "You are a coding assistant.", "Write the fibonacci function in c++.", 128, "(Aside|she|felter|alonger)+", 104, 64, "length", False, None), ("codellama70b", "You are a coding assistant.", "Write the fibonacci function in c++.", 128, "(Aside|she|felter|alonger)+", 104, 64, "length", True, None), + (None, "Book", [{"type": "text", "text": "What is"}, {"type": "text", "text": "the best book"}], 8, "Whillicter", 79, 8, "length", False, None), + (None, "Book", [{"type": "text", "text": "What is"}, {"type": "text", "text": "the best book"}], 8, "Whillicter", 79, 8, "length", True, None), ] ) def test_chat_completion(model, system_prompt, user_prompt, max_tokens, re_content, n_prompt, n_predicted, finish_reason, jinja, chat_template): @@ -44,7 +46,7 @@ def test_chat_completion(model, system_prompt, user_prompt, max_tokens, re_conte assert res.body["usage"]["completion_tokens"] == n_predicted choice = res.body["choices"][0] assert "assistant" == choice["message"]["role"] - assert match_regex(re_content, choice["message"]["content"]) + assert match_regex(re_content, choice["message"]["content"]), f'Expected {re_content}, got {choice["message"]["content"]}' assert choice["finish_reason"] == finish_reason @@ -142,6 +144,7 @@ def test_apply_chat_template(): @pytest.mark.parametrize("response_format,n_predicted,re_content", [ ({"type": "json_object", "schema": {"const": "42"}}, 6, "\"42\""), ({"type": "json_object", "schema": {"items": [{"type": "integer"}]}}, 10, "[ -3000 ]"), + ({"type": "json_schema", "json_schema": {"schema": {"const": "foooooo"}}}, 10, "\"foooooo\""), ({"type": "json_object"}, 10, "(\\{|John)+"), ({"type": "sound"}, 0, None), # invalid response format (expected to fail) @@ -169,6 +172,47 @@ def test_completion_with_response_format(response_format: dict, n_predicted: int assert "error" in res.body +@pytest.mark.parametrize("jinja,json_schema,n_predicted,re_content", [ + (False, {"const": "42"}, 6, "\"42\""), + (True, {"const": "42"}, 6, "\"42\""), +]) +def test_completion_with_json_schema(jinja: bool, json_schema: dict, n_predicted: int, re_content: str): + global server + server.jinja = jinja + server.start() + res = server.make_request("POST", "/chat/completions", data={ + "max_tokens": n_predicted, + "messages": [ + {"role": "system", "content": "You are a coding assistant."}, + {"role": "user", "content": "Write an example"}, + ], + "json_schema": json_schema, + }) + assert res.status_code == 200, f'Expected 200, got {res.status_code}' + choice = res.body["choices"][0] + assert match_regex(re_content, choice["message"]["content"]), f'Expected {re_content}, got {choice["message"]["content"]}' + + +@pytest.mark.parametrize("jinja,grammar,n_predicted,re_content", [ + (False, 'root ::= "a"{5,5}', 6, "a{5,5}"), + (True, 'root ::= "a"{5,5}', 6, "a{5,5}"), +]) +def test_completion_with_grammar(jinja: bool, grammar: str, n_predicted: int, re_content: str): + global server + server.jinja = jinja + server.start() + res = server.make_request("POST", "/chat/completions", data={ + "max_tokens": n_predicted, + "messages": [ + {"role": "user", "content": "Does not matter what I say, does it?"}, + ], + "grammar": grammar, + }) + assert res.status_code == 200, res.body + choice = res.body["choices"][0] + assert match_regex(re_content, choice["message"]["content"]), choice["message"]["content"] + + @pytest.mark.parametrize("messages", [ None, "string", diff --git a/examples/server/tests/unit/test_embedding.py b/examples/server/tests/unit/test_embedding.py index 8b0eb42b0..0feb452cc 100644 --- a/examples/server/tests/unit/test_embedding.py +++ b/examples/server/tests/unit/test_embedding.py @@ -49,6 +49,26 @@ def test_embedding_multiple(): assert len(d['embedding']) > 1 +def test_embedding_multiple_with_fa(): + server = ServerPreset.bert_bge_small_with_fa() + server.pooling = 'last' + server.start() + # one of these should trigger the FA branch (i.e. context size % 256 == 0) + res = server.make_request("POST", "/v1/embeddings", data={ + "input": [ + "a "*253, + "b "*254, + "c "*255, + "d "*256, + ], + }) + assert res.status_code == 200 + assert len(res.body['data']) == 4 + for d in res.body['data']: + assert 'embedding' in d + assert len(d['embedding']) > 1 + + @pytest.mark.parametrize( "input,is_multi_prompt", [ diff --git a/examples/server/tests/unit/test_rerank.py b/examples/server/tests/unit/test_rerank.py index 7203d7943..f4f570ad5 100644 --- a/examples/server/tests/unit/test_rerank.py +++ b/examples/server/tests/unit/test_rerank.py @@ -10,17 +10,20 @@ def create_server(): server = ServerPreset.jina_reranker_tiny() +TEST_DOCUMENTS = [ + "A machine is a physical system that uses power to apply forces and control movement to perform an action. The term is commonly applied to artificial devices, such as those employing engines or motors, but also to natural biological macromolecules, such as molecular machines.", + "Learning is the process of acquiring new understanding, knowledge, behaviors, skills, values, attitudes, and preferences. The ability to learn is possessed by humans, non-human animals, and some machines; there is also evidence for some kind of learning in certain plants.", + "Machine learning is a field of study in artificial intelligence concerned with the development and study of statistical algorithms that can learn from data and generalize to unseen data, and thus perform tasks without explicit instructions.", + "Paris, capitale de la France, est une grande ville européenne et un centre mondial de l'art, de la mode, de la gastronomie et de la culture. Son paysage urbain du XIXe siècle est traversé par de larges boulevards et la Seine." +] + + def test_rerank(): global server server.start() res = server.make_request("POST", "/rerank", data={ "query": "Machine learning is", - "documents": [ - "A machine is a physical system that uses power to apply forces and control movement to perform an action. The term is commonly applied to artificial devices, such as those employing engines or motors, but also to natural biological macromolecules, such as molecular machines.", - "Learning is the process of acquiring new understanding, knowledge, behaviors, skills, values, attitudes, and preferences. The ability to learn is possessed by humans, non-human animals, and some machines; there is also evidence for some kind of learning in certain plants.", - "Machine learning is a field of study in artificial intelligence concerned with the development and study of statistical algorithms that can learn from data and generalize to unseen data, and thus perform tasks without explicit instructions.", - "Paris, capitale de la France, est une grande ville européenne et un centre mondial de l'art, de la mode, de la gastronomie et de la culture. Son paysage urbain du XIXe siècle est traversé par de larges boulevards et la Seine." - ] + "documents": TEST_DOCUMENTS, }) assert res.status_code == 200 assert len(res.body["results"]) == 4 @@ -38,6 +41,29 @@ def test_rerank(): assert least_relevant["index"] == 3 +def test_rerank_tei_format(): + global server + server.start() + res = server.make_request("POST", "/rerank", data={ + "query": "Machine learning is", + "texts": TEST_DOCUMENTS, + }) + assert res.status_code == 200 + assert len(res.body) == 4 + + most_relevant = res.body[0] + least_relevant = res.body[0] + for doc in res.body: + if doc["score"] > most_relevant["score"]: + most_relevant = doc + if doc["score"] < least_relevant["score"]: + least_relevant = doc + + assert most_relevant["score"] > least_relevant["score"] + assert most_relevant["index"] == 2 + assert least_relevant["index"] == 3 + + @pytest.mark.parametrize("documents", [ [], None, diff --git a/examples/server/tests/unit/test_tool_call.py b/examples/server/tests/unit/test_tool_call.py old mode 100644 new mode 100755 index 4a551404f..569c2a1f8 --- a/examples/server/tests/unit/test_tool_call.py +++ b/examples/server/tests/unit/test_tool_call.py @@ -1,4 +1,12 @@ +#!/usr/bin/env python import pytest + +# ensure grandparent path is in sys.path +from pathlib import Path +import sys +path = Path(__file__).resolve().parents[1] +sys.path.insert(0, str(path)) + from utils import * server: ServerProcess @@ -66,15 +74,8 @@ WEATHER_TOOL = { } -def do_test_completion_with_required_tool_tiny(template_name: str, tool: dict, argument_key: str | None): - global server - n_predict = 512 - # server = ServerPreset.stories15m_moe() - server.jinja = True - server.n_predict = n_predict - server.chat_template_file = f'../../../models/templates/{template_name}.jinja' - server.start(timeout_seconds=TIMEOUT_SERVER_START) - res = server.make_request("POST", "/chat/completions", data={ +def do_test_completion_with_required_tool_tiny(server: ServerProcess, tool: dict, argument_key: str | None, n_predict, **kwargs): + res = server.make_request("POST", "/v1/chat/completions", data={ "max_tokens": n_predict, "messages": [ {"role": "system", "content": "You are a coding assistant."}, @@ -83,15 +84,15 @@ def do_test_completion_with_required_tool_tiny(template_name: str, tool: dict, a "tool_choice": "required", "tools": [tool], "parallel_tool_calls": False, - "temperature": 0.0, - "top_k": 1, - "top_p": 1.0, + **kwargs, }) assert res.status_code == 200, f"Expected status code 200, got {res.status_code}" choice = res.body["choices"][0] tool_calls = choice["message"].get("tool_calls") assert tool_calls and len(tool_calls) == 1, f'Expected 1 tool call in {choice["message"]}' tool_call = tool_calls[0] + assert choice["message"].get("content") in (None, ""), f'Expected no content in {choice["message"]}' + assert len(tool_call.get("id", "")) > 0, f'Expected non empty tool call id in {tool_call}' expected_function_name = "python" if tool["type"] == "code_interpreter" else tool["function"]["name"] assert expected_function_name == tool_call["function"]["name"] actual_arguments = tool_call["function"]["arguments"] @@ -107,7 +108,14 @@ def do_test_completion_with_required_tool_tiny(template_name: str, tool: dict, a ("meta-llama-Llama-3.3-70B-Instruct", PYTHON_TOOL, "code"), ]) def test_completion_with_required_tool_tiny_fast(template_name: str, tool: dict, argument_key: str | None): - do_test_completion_with_required_tool_tiny(template_name, tool, argument_key) + global server + n_predict = 512 + # server = ServerPreset.stories15m_moe() + server.jinja = True + server.n_predict = n_predict + server.chat_template_file = f'../../../models/templates/{template_name}.jinja' + server.start(timeout_seconds=TIMEOUT_SERVER_START) + do_test_completion_with_required_tool_tiny(server, tool, argument_key, n_predict, temperature=0.0, top_k=1, top_p=1.0) @pytest.mark.slow @@ -129,10 +137,17 @@ def test_completion_with_required_tool_tiny_fast(template_name: str, tool: dict, ("deepseek-ai-DeepSeek-R1-Distill-Llama-8B", TEST_TOOL, "success"), ("deepseek-ai-DeepSeek-R1-Distill-Llama-8B", PYTHON_TOOL, "code"), ("fireworks-ai-llama-3-firefunction-v2", TEST_TOOL, "success"), - ("fireworks-ai-llama-3-firefunction-v2", PYTHON_TOOL, "code"), + # ("fireworks-ai-llama-3-firefunction-v2", PYTHON_TOOL, "code"), ]) def test_completion_with_required_tool_tiny_slow(template_name: str, tool: dict, argument_key: str | None): - do_test_completion_with_required_tool_tiny(template_name, tool, argument_key) + global server + n_predict = 512 + # server = ServerPreset.stories15m_moe() + server.jinja = True + server.n_predict = n_predict + server.chat_template_file = f'../../../models/templates/{template_name}.jinja' + server.start(timeout_seconds=TIMEOUT_SERVER_START) + do_test_completion_with_required_tool_tiny(server, tool, argument_key, n_predict) @pytest.mark.slow @@ -141,14 +156,22 @@ def test_completion_with_required_tool_tiny_slow(template_name: str, tool: dict, (PYTHON_TOOL, "code", "bartowski/Meta-Llama-3.1-8B-Instruct-GGUF:Q4_K_M", None), (PYTHON_TOOL, "code", "bartowski/Meta-Llama-3.1-8B-Instruct-GGUF:Q4_K_M", "chatml"), - # Note: gemma-2-2b-it knows itself as "model", not "assistant", so we don't test the ill-suited chatml on it. (TEST_TOOL, "success", "bartowski/gemma-2-2b-it-GGUF:Q4_K_M", None), (PYTHON_TOOL, "code", "bartowski/gemma-2-2b-it-GGUF:Q4_K_M", None), + (PYTHON_TOOL, "code", "bartowski/gemma-2-2b-it-GGUF:Q4_K_M", "chatml"), (TEST_TOOL, "success", "bartowski/Phi-3.5-mini-instruct-GGUF:Q4_K_M", None), (PYTHON_TOOL, "code", "bartowski/Phi-3.5-mini-instruct-GGUF:Q4_K_M", None), (PYTHON_TOOL, "code", "bartowski/Phi-3.5-mini-instruct-GGUF:Q4_K_M", "chatml"), + (TEST_TOOL, "success", "bartowski/Qwen2.5-1.5B-Instruct-GGUF:Q4_K_M", None), + (PYTHON_TOOL, "code", "bartowski/Qwen2.5-1.5B-Instruct-GGUF:Q4_K_M", None), + (PYTHON_TOOL, "code", "bartowski/Qwen2.5-1.5B-Instruct-GGUF:Q4_K_M", "chatml"), + + (TEST_TOOL, "success", "bartowski/Qwen2.5-Coder-3B-Instruct-GGUF:Q4_K_M", None), + (PYTHON_TOOL, "code", "bartowski/Qwen2.5-Coder-3B-Instruct-GGUF:Q4_K_M", None), + (PYTHON_TOOL, "code", "bartowski/Qwen2.5-Coder-3B-Instruct-GGUF:Q4_K_M", "chatml"), + (TEST_TOOL, "success", "bartowski/Qwen2.5-7B-Instruct-GGUF:Q4_K_M", None), (PYTHON_TOOL, "code", "bartowski/Qwen2.5-7B-Instruct-GGUF:Q4_K_M", None), (PYTHON_TOOL, "code", "bartowski/Qwen2.5-7B-Instruct-GGUF:Q4_K_M", "chatml"), @@ -176,9 +199,9 @@ def test_completion_with_required_tool_tiny_slow(template_name: str, tool: dict, (TEST_TOOL, "success", "bartowski/Llama-3.2-1B-Instruct-GGUF:Q4_K_M", ("meta-llama/Llama-3.2-3B-Instruct", None)), (PYTHON_TOOL, "code", "bartowski/Llama-3.2-1B-Instruct-GGUF:Q4_K_M", ("meta-llama/Llama-3.2-3B-Instruct", None)), (PYTHON_TOOL, "code", "bartowski/Llama-3.2-1B-Instruct-GGUF:Q4_K_M", "chatml"), - # TODO: fix these - # (TEST_TOOL, "success", "bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q4_K_M", None), - # (PYTHON_TOOL, "code", "bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q4_K_M", None), + + (TEST_TOOL, "success", "bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q4_K_M", None), + (PYTHON_TOOL, "code", "bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q4_K_M", None), ]) def test_completion_with_required_tool_real_model(tool: dict, argument_key: str | None, hf_repo: str, template_override: str | Tuple[str, str | None] | None): global server @@ -196,7 +219,7 @@ def test_completion_with_required_tool_real_model(tool: dict, argument_key: str elif isinstance(template_override, str): server.chat_template = template_override server.start(timeout_seconds=TIMEOUT_SERVER_START) - res = server.make_request("POST", "/chat/completions", data={ + res = server.make_request("POST", "/v1/chat/completions", data={ "max_tokens": n_predict, "messages": [ {"role": "system", "content": "You are a coding assistant."}, @@ -214,6 +237,7 @@ def test_completion_with_required_tool_real_model(tool: dict, argument_key: str tool_calls = choice["message"].get("tool_calls") assert tool_calls and len(tool_calls) == 1, f'Expected 1 tool call in {choice["message"]}' tool_call = tool_calls[0] + # assert choice["message"].get("content") in (None, ""), f'Expected no content in {choice["message"]}' expected_function_name = "python" if tool["type"] == "code_interpreter" else tool["function"]["name"] assert expected_function_name == tool_call["function"]["name"] actual_arguments = tool_call["function"]["arguments"] @@ -223,13 +247,8 @@ def test_completion_with_required_tool_real_model(tool: dict, argument_key: str assert argument_key in actual_arguments, f"tool arguments: {json.dumps(actual_arguments)}, expected: {argument_key}" -def do_test_completion_without_tool_call(template_name: str, n_predict: int, tools: list[dict], tool_choice: str | None): - global server - server.jinja = True - server.n_predict = n_predict - server.chat_template_file = f'../../../models/templates/{template_name}.jinja' - server.start(timeout_seconds=TIMEOUT_SERVER_START) - res = server.make_request("POST", "/chat/completions", data={ +def do_test_completion_without_tool_call(server: ServerProcess, n_predict: int, tools: list[dict], tool_choice: str | None, **kwargs): + res = server.make_request("POST", "/v1/chat/completions", data={ "max_tokens": n_predict, "messages": [ {"role": "system", "content": "You are a coding assistant."}, @@ -237,9 +256,7 @@ def do_test_completion_without_tool_call(template_name: str, n_predict: int, too ], "tools": tools if tools else None, "tool_choice": tool_choice, - "temperature": 0.0, - "top_k": 1, - "top_p": 1.0, + **kwargs, }, timeout=TIMEOUT_HTTP_REQUEST) assert res.status_code == 200, f"Expected status code 200, got {res.status_code}" choice = res.body["choices"][0] @@ -252,7 +269,12 @@ def do_test_completion_without_tool_call(template_name: str, n_predict: int, too ("meta-llama-Llama-3.3-70B-Instruct", 128, [PYTHON_TOOL], 'none'), ]) def test_completion_without_tool_call_fast(template_name: str, n_predict: int, tools: list[dict], tool_choice: str | None): - do_test_completion_without_tool_call(template_name, n_predict, tools, tool_choice) + global server + server.jinja = True + server.n_predict = n_predict + server.chat_template_file = f'../../../models/templates/{template_name}.jinja' + server.start(timeout_seconds=TIMEOUT_SERVER_START) + do_test_completion_without_tool_call(server, n_predict, tools, tool_choice) @pytest.mark.slow @@ -268,18 +290,28 @@ def test_completion_without_tool_call_fast(template_name: str, n_predict: int, t ("meta-llama-Llama-3.2-3B-Instruct", 256, [PYTHON_TOOL], 'none'), ]) def test_completion_without_tool_call_slow(template_name: str, n_predict: int, tools: list[dict], tool_choice: str | None): - do_test_completion_without_tool_call(template_name, n_predict, tools, tool_choice) + global server + server.jinja = True + server.n_predict = n_predict + server.chat_template_file = f'../../../models/templates/{template_name}.jinja' + server.start(timeout_seconds=TIMEOUT_SERVER_START) + do_test_completion_without_tool_call(server, n_predict, tools, tool_choice) @pytest.mark.slow @pytest.mark.parametrize("hf_repo,template_override", [ - ("bartowski/c4ai-command-r7b-12-2024-GGUF:Q4_K_M", ("CohereForAI/c4ai-command-r7b-12-2024", "tool_use")), ("bartowski/Meta-Llama-3.1-8B-Instruct-GGUF:Q4_K_M", None), ("bartowski/Meta-Llama-3.1-8B-Instruct-GGUF:Q4_K_M", "chatml"), ("bartowski/Phi-3.5-mini-instruct-GGUF:Q4_K_M", None), ("bartowski/Phi-3.5-mini-instruct-GGUF:Q4_K_M", "chatml"), + ("bartowski/Qwen2.5-1.5B-Instruct-GGUF:Q4_K_M", None), + ("bartowski/Qwen2.5-1.5B-Instruct-GGUF:Q4_K_M", "chatml"), + + ("bartowski/Qwen2.5-Coder-3B-Instruct-GGUF:Q4_K_M", None), + ("bartowski/Qwen2.5-Coder-3B-Instruct-GGUF:Q4_K_M", "chatml"), + ("bartowski/Qwen2.5-7B-Instruct-GGUF:Q4_K_M", None), ("bartowski/Qwen2.5-7B-Instruct-GGUF:Q4_K_M", "chatml"), @@ -298,13 +330,16 @@ def test_completion_without_tool_call_slow(template_name: str, n_predict: int, t ("bartowski/Llama-3.2-3B-Instruct-GGUF:Q4_K_M", ("meta-llama/Llama-3.2-3B-Instruct", None)), ("bartowski/Llama-3.2-3B-Instruct-GGUF:Q4_K_M", "chatml"), + ("bartowski/c4ai-command-r7b-12-2024-GGUF:Q6_K_L", ("CohereForAI/c4ai-command-r7b-12-2024", "tool_use")), + + ("bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q4_K_M", None), + # Note: gemma-2-2b-it knows itself as "model", not "assistant", so we don't test the ill-suited chatml on it. ("bartowski/gemma-2-2b-it-GGUF:Q4_K_M", None), # ("bartowski/Llama-3.2-1B-Instruct-GGUF:Q4_K_M", ("meta-llama/Llama-3.2-3B-Instruct", None)), - # ("bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q4_K_M", None), ]) -def test_weather(hf_repo: str, template_override: Tuple[str, str | None] | None): +def test_weather(hf_repo: str, template_override: str | Tuple[str, str | None] | None): global server n_predict = 512 server.n_slots = 1 @@ -320,66 +355,57 @@ def test_weather(hf_repo: str, template_override: Tuple[str, str | None] | None) elif isinstance(template_override, str): server.chat_template = template_override server.start(timeout_seconds=TIMEOUT_SERVER_START) - res = server.make_request("POST", "/chat/completions", data={ - "max_tokens": n_predict, + do_test_weather(server, max_tokens=n_predict) + + +def do_test_weather(server: ServerProcess, **kwargs): + res = server.make_request("POST", "/v1/chat/completions", data={ "messages": [ + {"role": "system", "content": "You are a chatbot that uses tools/functions. Dont overthink things."}, {"role": "user", "content": "What is the weather in Istanbul?"}, ], "tools": [WEATHER_TOOL], + **kwargs, }, timeout=TIMEOUT_HTTP_REQUEST) assert res.status_code == 200, f"Expected status code 200, got {res.status_code}" choice = res.body["choices"][0] tool_calls = choice["message"].get("tool_calls") assert tool_calls and len(tool_calls) == 1, f'Expected 1 tool call in {choice["message"]}' tool_call = tool_calls[0] - assert tool_call["function"]["name"] == WEATHER_TOOL["function"]["name"] + # assert choice["message"].get("content") in (None, ""), f'Expected no content in {choice["message"]}' + assert tool_call["function"]["name"] == WEATHER_TOOL["function"]["name"], f'Expected weather tool call, got {tool_call["function"]["name"]}' + assert len(tool_call.get("id", "")) > 0, f'Expected non empty tool call id in {tool_call}' actual_arguments = json.loads(tool_call["function"]["arguments"]) assert 'location' in actual_arguments, f"location not found in {json.dumps(actual_arguments)}" location = actual_arguments["location"] assert isinstance(location, str), f"Expected location to be a string, got {type(location)}: {json.dumps(location)}" - assert re.match('^Istanbul(, (TR|Turkey|Türkiye))?$', location), f'Expected Istanbul for location, got {location}' + assert re.match('^Istanbul(( |, ?)(TR|Turkey|Türkiye))?$', location), f'Expected Istanbul for location, got {location}' @pytest.mark.slow -@pytest.mark.parametrize("expected_arguments_override,hf_repo,template_override", [ - (None, "bartowski/Phi-3.5-mini-instruct-GGUF:Q4_K_M", None), - (None, "bartowski/Phi-3.5-mini-instruct-GGUF:Q4_K_M", "chatml"), +@pytest.mark.parametrize("result_override,n_predict,hf_repo,template_override", [ + (None, 128, "bartowski/Phi-3.5-mini-instruct-GGUF:Q4_K_M", "chatml"), + (None, 128, "bartowski/Qwen2.5-Coder-3B-Instruct-GGUF:Q4_K_M", None), + (None, 128, "bartowski/Qwen2.5-Coder-3B-Instruct-GGUF:Q4_K_M", "chatml"), + (None, 128, "bartowski/Qwen2.5-7B-Instruct-GGUF:Q4_K_M", "chatml"), + (None, 128, "bartowski/Hermes-2-Pro-Llama-3-8B-GGUF:Q4_K_M", ("NousResearch/Hermes-2-Pro-Llama-3-8B", "tool_use")), + (None, 128, "bartowski/Hermes-3-Llama-3.1-8B-GGUF:Q4_K_M", ("NousResearch/Hermes-3-Llama-3.1-8B", "tool_use")), + (None, 128, "bartowski/functionary-small-v3.2-GGUF:Q8_0", ("meetkai/functionary-medium-v3.2", None)), + (None, 128, "bartowski/Mistral-Nemo-Instruct-2407-GGUF:Q4_K_M", None), + (None, 128, "bartowski/Mistral-Nemo-Instruct-2407-GGUF:Q4_K_M", "chatml"), + (None, 128, "bartowski/Phi-3.5-mini-instruct-GGUF:Q4_K_M", None), + ("[\\s\\S]*?\\*\\*\\s*0.5($|\\*\\*)", 8192, "bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q4_K_M", ("llama-cpp-deepseek-r1", None)), - (None, "bartowski/functionary-small-v3.2-GGUF:Q8_0", ("meetkai-functionary-medium-v3.2", None)), - (None, "bartowski/functionary-small-v3.2-GGUF:Q8_0", "chatml"), - - (None, "bartowski/Meta-Llama-3.1-8B-Instruct-GGUF:Q4_K_M", None), - ('{"code":"print("}', "bartowski/Meta-Llama-3.1-8B-Instruct-GGUF:Q4_K_M", "chatml"), - - ('{"code":"print("}', "bartowski/Llama-3.2-1B-Instruct-GGUF:Q4_K_M", ("meta-llama-Llama-3.2-3B-Instruct", None)), - (None, "bartowski/Llama-3.2-1B-Instruct-GGUF:Q4_K_M", "chatml"), - - ('{"code":"print("}', "bartowski/Llama-3.2-3B-Instruct-GGUF:Q4_K_M", ("meta-llama-Llama-3.2-3B-Instruct", None)), - ('{"code":"print("}', "bartowski/Llama-3.2-3B-Instruct-GGUF:Q4_K_M", "chatml"), - - (None, "bartowski/Qwen2.5-7B-Instruct-GGUF:Q4_K_M", None), - (None, "bartowski/Qwen2.5-7B-Instruct-GGUF:Q4_K_M", "chatml"), - - (None, "bartowski/Hermes-2-Pro-Llama-3-8B-GGUF:Q4_K_M", ("NousResearch/Hermes-2-Pro-Llama-3-8B", "tool_use")), - (None, "bartowski/Hermes-2-Pro-Llama-3-8B-GGUF:Q4_K_M", "chatml"), - - (None, "bartowski/Hermes-3-Llama-3.1-8B-GGUF:Q4_K_M", ("NousResearch-Hermes-3-Llama-3.1-8B", "tool_use")), - (None, "bartowski/Hermes-3-Llama-3.1-8B-GGUF:Q4_K_M", "chatml"), - - (None, "bartowski/Mistral-Nemo-Instruct-2407-GGUF:Q4_K_M", None), - (None, "bartowski/Mistral-Nemo-Instruct-2407-GGUF:Q4_K_M", "chatml"), - - # Note: gemma-2-2b-it knows itself as "model", not "assistant", so we don't test the ill-suited chatml on it. - (None, "bartowski/gemma-2-2b-it-GGUF:Q4_K_M", None), - - # (None, "bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q4_K_M", None), + # TODO: fix these (wrong results, either didn't respect decimal instruction or got wrong value) + # (None, 128, "bartowski/Meta-Llama-3.1-8B-Instruct-GGUF:Q4_K_M", None), + # ("[\\s\\S]*?\\*\\*\\s*0.5($|\\*\\*)", 8192, "bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q4_K_M", None), ]) -def test_hello_world_tool_call(expected_arguments_override: str | None, hf_repo: str, template_override: str | Tuple[str, str | None] | None): +def test_calc_result(result_override: str | None, n_predict: int, hf_repo: str, template_override: str | Tuple[str, str | None] | None): global server server.n_slots = 1 server.jinja = True - server.n_ctx = 8192 - server.n_predict = 128 + server.n_ctx = 8192 * 2 + server.n_predict = n_predict server.model_hf_repo = hf_repo server.model_hf_file = None if isinstance(template_override, tuple): @@ -389,30 +415,192 @@ def test_hello_world_tool_call(expected_arguments_override: str | None, hf_repo: elif isinstance(template_override, str): server.chat_template = template_override server.start(timeout_seconds=TIMEOUT_SERVER_START) - res = server.make_request("POST", "/chat/completions", data={ - "max_tokens": 256, + do_test_calc_result(server, result_override, n_predict) + + +def do_test_calc_result(server: ServerProcess, result_override: str | None, n_predict: int, **kwargs): + res = server.make_request("POST", "/v1/chat/completions", data={ + "max_tokens": n_predict, "messages": [ - {"role": "system", "content": "You are a coding assistant."}, + {"role": "system", "content": "You are a tools-calling assistant. You express numerical values with at most two decimals."}, + {"role": "user", "content": "What's the y coordinate of a point on the unit sphere at angle 30 degrees?"}, + { + "role": "assistant", + "content": None, + "tool_calls": [ + { + "id": "call_6789", + "type": "function", + "function": { + "name": "calculate", + "arguments": "{\"expression\":\"sin(30 * pi / 180)\"}" + } + } + ] + }, + { + "role": "tool", + "name": "calculate", + "content": "0.55644242476", + "tool_call_id": "call_6789" + } + ], + "tools": [ + { + "type":"function", + "function":{ + "name":"calculate", + "description":"A calculator function that computes values of arithmetic expressions in the Python syntax", + "parameters":{ + "type":"object", + "properties":{ + "expression":{ + "type":"string", + "description":"An arithmetic expression to compute the value of (Python syntad, assuming all floats)" + } + }, + "required":["expression"] + } + } + } + ], + **kwargs, + }, timeout=TIMEOUT_HTTP_REQUEST) + assert res.status_code == 200, f"Expected status code 200, got {res.status_code}" + choice = res.body["choices"][0] + tool_calls = choice["message"].get("tool_calls") + assert tool_calls is None, f'Expected no tool call in {choice["message"]}' + content = choice["message"].get("content") + assert content is not None, f'Expected content in {choice["message"]}' + if result_override is not None: + assert re.match(result_override, content), f'Expected {result_override}, got {content}' + else: + assert re.match('^[\\s\\S]*?((That\'s|\\bis) (approximately )?)?\\b0\\.(5\\b|56\\b|556)', content), \ + f'Expected something like "The y coordinate is 0.56.", got {content}' + + +@pytest.mark.slow +@pytest.mark.parametrize("n_predict,reasoning_format,expect_content,expect_reasoning_content,hf_repo,template_override", [ + (128, 'deepseek', "^The sum of 102 and 7 is 109[\\s\\S]*", None, "bartowski/Phi-3.5-mini-instruct-GGUF:Q4_K_M", None), + (128, None, "^The sum of 102 and 7 is 109[\\s\\S]*", None, "bartowski/Phi-3.5-mini-instruct-GGUF:Q4_K_M", None), + + (1024, 'deepseek', "To find the sum of[\\s\\S]*", "I need to calculate the sum of 102 and 7[\\s\\S]*", "bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q4_K_M", None), + (1024, 'none', "^(\\s*)?I need[\\s\\S]*?\\s*To find[\\s\\S]*", None, "bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q4_K_M", None), + + (1024, 'deepseek', "To find the sum of[\\s\\S]*", "First, I [\\s\\S]*", "bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q4_K_M", ("llama-cpp-deepseek-r1", None)), +]) +def test_thoughts(n_predict: int, reasoning_format: Literal['deepseek', 'none'] | None, expect_content: str | None, expect_reasoning_content: str | None, hf_repo: str, template_override: str | Tuple[str, str | None] | None): + global server + server.n_slots = 1 + server.reasoning_format = reasoning_format + server.jinja = True + server.n_ctx = 8192 * 2 + server.n_predict = n_predict + server.model_hf_repo = hf_repo + server.model_hf_file = None + if isinstance(template_override, tuple): + (template_hf_repo, template_variant) = template_override + server.chat_template_file = f"../../../models/templates/{template_hf_repo.replace('/', '-') + ('-' + template_variant if template_variant else '')}.jinja" + assert os.path.exists(server.chat_template_file), f"Template file {server.chat_template_file} does not exist. Run `python scripts/get_chat_template.py {template_hf_repo} {template_variant} > {server.chat_template_file}` to download the template." + elif isinstance(template_override, str): + server.chat_template = template_override + server.start(timeout_seconds=TIMEOUT_SERVER_START) + res = server.make_request("POST", "/v1/chat/completions", data={ + "max_tokens": n_predict, + "messages": [ + {"role": "user", "content": "What's the sum of 102 and 7?"}, + ] + }, timeout=TIMEOUT_HTTP_REQUEST) + assert res.status_code == 200, f"Expected status code 200, got {res.status_code}" + choice = res.body["choices"][0] + assert choice["message"].get("tool_calls") is None, f'Expected no tool call in {choice["message"]}' + + content = choice["message"].get("content") + if expect_content is None: + assert choice["message"].get("content") in (None, ""), f'Expected no content in {choice["message"]}' + else: + assert re.match(expect_content, content), f'Expected {expect_content}, got {content}' + + reasoning_content = choice["message"].get("reasoning_content") + if expect_reasoning_content is None: + assert reasoning_content is None, f'Expected no reasoning content in {choice["message"]}' + else: + assert re.match(expect_reasoning_content, reasoning_content), f'Expected {expect_reasoning_content}, got {reasoning_content}' + + +@pytest.mark.slow +@pytest.mark.parametrize("hf_repo,template_override", [ + ("bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q4_K_M", None), + + ("bartowski/Phi-3.5-mini-instruct-GGUF:Q4_K_M", None), + ("bartowski/Phi-3.5-mini-instruct-GGUF:Q4_K_M", "chatml"), + + ("bartowski/functionary-small-v3.2-GGUF:Q8_0", ("meetkai-functionary-medium-v3.2", None)), + ("bartowski/functionary-small-v3.2-GGUF:Q8_0", "chatml"), + + # ("bartowski/Meta-Llama-3.1-8B-Instruct-GGUF:Q4_K_M", None), + ("bartowski/Meta-Llama-3.1-8B-Instruct-GGUF:Q4_K_M", "chatml"), + + ("bartowski/Llama-3.2-1B-Instruct-GGUF:Q4_K_M", ("meta-llama-Llama-3.2-3B-Instruct", None)), + ("bartowski/Llama-3.2-1B-Instruct-GGUF:Q4_K_M", None), + + ("bartowski/Llama-3.2-3B-Instruct-GGUF:Q4_K_M", ("meta-llama-Llama-3.2-3B-Instruct", None)), + ("bartowski/Llama-3.2-3B-Instruct-GGUF:Q4_K_M", None), + + ("bartowski/Qwen2.5-7B-Instruct-GGUF:Q4_K_M", None), + ("bartowski/Qwen2.5-7B-Instruct-GGUF:Q4_K_M", "chatml"), + + ("bartowski/Hermes-2-Pro-Llama-3-8B-GGUF:Q4_K_M", ("NousResearch/Hermes-2-Pro-Llama-3-8B", "tool_use")), + ("bartowski/Hermes-2-Pro-Llama-3-8B-GGUF:Q4_K_M", "chatml"), + + ("bartowski/Hermes-3-Llama-3.1-8B-GGUF:Q4_K_M", ("NousResearch-Hermes-3-Llama-3.1-8B", "tool_use")), + ("bartowski/Hermes-3-Llama-3.1-8B-GGUF:Q4_K_M", "chatml"), + + ("bartowski/Mistral-Nemo-Instruct-2407-GGUF:Q4_K_M", None), + ("bartowski/Mistral-Nemo-Instruct-2407-GGUF:Q4_K_M", "chatml"), + + ("bartowski/gemma-2-2b-it-GGUF:Q4_K_M", None), + ("bartowski/gemma-2-2b-it-GGUF:Q4_K_M", "chatml"), +]) +def test_hello_world(hf_repo: str, template_override: str | Tuple[str, str | None] | None): + global server + n_predict = 512 # High because of DeepSeek R1 + server.n_slots = 1 + server.jinja = True + server.n_ctx = 8192 + server.n_predict = n_predict + server.model_hf_repo = hf_repo + server.model_hf_file = None + if isinstance(template_override, tuple): + (template_hf_repo, template_variant) = template_override + server.chat_template_file = f"../../../models/templates/{template_hf_repo.replace('/', '-') + ('-' + template_variant if template_variant else '')}.jinja" + assert os.path.exists(server.chat_template_file), f"Template file {server.chat_template_file} does not exist. Run `python scripts/get_chat_template.py {template_hf_repo} {template_variant} > {server.chat_template_file}` to download the template." + elif isinstance(template_override, str): + server.chat_template = template_override + server.start(timeout_seconds=TIMEOUT_SERVER_START) + + do_test_hello_world(server, max_tokens=n_predict) + + +def do_test_hello_world(server: ServerProcess, **kwargs): + res = server.make_request("POST", "/v1/chat/completions", data={ + "messages": [ + {"role": "system", "content": "You are a tool-calling agent."}, {"role": "user", "content": "say hello world with python"}, ], "tools": [PYTHON_TOOL], - # Note: without these greedy params, Functionary v3.2 writes `def hello_world():\n print("Hello, World!")\nhello_world()` which is correct but a pain to test. - "temperature": 0.0, - "top_k": 1, - "top_p": 1.0, + **kwargs, }, timeout=TIMEOUT_HTTP_REQUEST) assert res.status_code == 200, f"Expected status code 200, got {res.status_code}" choice = res.body["choices"][0] tool_calls = choice["message"].get("tool_calls") assert tool_calls and len(tool_calls) == 1, f'Expected 1 tool call in {choice["message"]}' tool_call = tool_calls[0] + # assert choice["message"].get("content") in (None, ""), f'Expected no content in {choice["message"]}' assert tool_call["function"]["name"] == PYTHON_TOOL["function"]["name"] - actual_arguments = tool_call["function"]["arguments"] - if expected_arguments_override is not None: - assert actual_arguments == expected_arguments_override - else: - actual_arguments = json.loads(actual_arguments) - assert 'code' in actual_arguments, f"code not found in {json.dumps(actual_arguments)}" - code = actual_arguments["code"] - assert isinstance(code, str), f"Expected code to be a string, got {type(code)}: {json.dumps(code)}" - assert re.match(r'''print\(("[Hh]ello,? [Ww]orld!?"|'[Hh]ello,? [Ww]orld!?')\)''', code), f'Expected hello world, got {code}' + assert len(tool_call.get("id", "")) > 0, f'Expected non empty tool call id in {tool_call}' + actual_arguments = json.loads(tool_call["function"]["arguments"]) + assert 'code' in actual_arguments, f"code not found in {json.dumps(actual_arguments)}" + code = actual_arguments["code"] + assert isinstance(code, str), f"Expected code to be a string, got {type(code)}: {json.dumps(code)}" + assert re.match(r'''print\(("[Hh]ello,? [Ww]orld!?"|'[Hh]ello,? [Ww]orld!?')\)''', code), f'Expected hello world, got {code}' diff --git a/examples/server/tests/utils.py b/examples/server/tests/utils.py index ce0680662..4dc2062a8 100644 --- a/examples/server/tests/utils.py +++ b/examples/server/tests/utils.py @@ -26,7 +26,10 @@ from re import RegexFlag import wget -DEFAULT_HTTP_TIMEOUT = 12 if "LLAMA_SANITIZE" not in os.environ else 30 +DEFAULT_HTTP_TIMEOUT = 12 + +if "LLAMA_SANITIZE" in os.environ or "GITHUB_ACTION" in os.environ: + DEFAULT_HTTP_TIMEOUT = 30 class ServerResponse: @@ -64,6 +67,9 @@ class ServerProcess: id_slot: int | None = None cache_prompt: bool | None = None n_slots: int | None = None + ctk: str | None = None + ctv: str | None = None + fa: bool | None = None server_continuous_batching: bool | None = False server_embeddings: bool | None = False server_reranking: bool | None = False @@ -78,8 +84,10 @@ class ServerProcess: draft_max: int | None = None no_webui: bool | None = None jinja: bool | None = None + reasoning_format: Literal['deepseek', 'none'] | None = None chat_template: str | None = None chat_template_file: str | None = None + server_path: str | None = None # session variables process: subprocess.Popen | None = None @@ -93,7 +101,9 @@ class ServerProcess: self.server_port = int(os.environ["PORT"]) def start(self, timeout_seconds: int | None = DEFAULT_HTTP_TIMEOUT) -> None: - if "LLAMA_SERVER_BIN_PATH" in os.environ: + if self.server_path is not None: + server_path = self.server_path + elif "LLAMA_SERVER_BIN_PATH" in os.environ: server_path = os.environ["LLAMA_SERVER_BIN_PATH"] elif os.name == "nt": server_path = "../../../build/bin/Release/llama-server.exe" @@ -147,6 +157,12 @@ class ServerProcess: server_args.extend(["--ctx-size", self.n_ctx]) if self.n_slots: server_args.extend(["--parallel", self.n_slots]) + if self.ctk: + server_args.extend(["-ctk", self.ctk]) + if self.ctv: + server_args.extend(["-ctv", self.ctv]) + if self.fa is not None: + server_args.append("-fa") if self.n_predict: server_args.extend(["--n-predict", self.n_predict]) if self.slot_save_path: @@ -172,13 +188,15 @@ class ServerProcess: server_args.append("--no-webui") if self.jinja: server_args.append("--jinja") + if self.reasoning_format is not None: + server_args.extend(("--reasoning-format", self.reasoning_format)) if self.chat_template: server_args.extend(["--chat-template", self.chat_template]) if self.chat_template_file: server_args.extend(["--chat-template-file", self.chat_template_file]) args = [str(arg) for arg in [server_path, *server_args]] - print(f"bench: starting server with: {' '.join(args)}") + print(f"tests: starting server with: {' '.join(args)}") flags = 0 if "nt" == os.name: @@ -209,6 +227,10 @@ class ServerProcess: return # server is ready except Exception as e: pass + # Check if process died + if self.process.poll() is not None: + raise RuntimeError(f"Server process died with return code {self.process.returncode}") + print(f"Waiting for server to start...") time.sleep(0.5) raise TimeoutError(f"Server did not start within {timeout_seconds} seconds") @@ -280,7 +302,7 @@ class ServerPreset: server.model_hf_repo = "ggml-org/models" server.model_hf_file = "tinyllamas/stories260K.gguf" server.model_alias = "tinyllama-2" - server.n_ctx = 256 + server.n_ctx = 512 server.n_batch = 32 server.n_slots = 2 server.n_predict = 64 @@ -301,6 +323,21 @@ class ServerPreset: server.server_embeddings = True return server + @staticmethod + def bert_bge_small_with_fa() -> ServerProcess: + server = ServerProcess() + server.model_hf_repo = "ggml-org/models" + server.model_hf_file = "bert-bge-small/ggml-model-f16.gguf" + server.model_alias = "bert-bge-small" + server.n_ctx = 1024 + server.n_batch = 300 + server.n_ubatch = 300 + server.n_slots = 2 + server.fa = True + server.seed = 42 + server.server_embeddings = True + return server + @staticmethod def tinyllama_infill() -> ServerProcess: server = ServerProcess() diff --git a/examples/server/utils.hpp b/examples/server/utils.hpp index 5f97df5fd..aba2f27f9 100644 --- a/examples/server/utils.hpp +++ b/examples/server/utils.hpp @@ -3,18 +3,18 @@ #include "common.h" #include "log.h" #include "llama.h" -#include "common/base64.hpp" +#include "base64.hpp" // increase max payload length to allow use of larger context size #define CPPHTTPLIB_FORM_URL_ENCODED_PAYLOAD_MAX_LENGTH 1048576 +// disable Nagle's algorithm +#define CPPHTTPLIB_TCP_NODELAY true #include "httplib.h" // Change JSON_ASSERT from assert() to GGML_ASSERT: #define JSON_ASSERT GGML_ASSERT #include "json.hpp" -#include "minja.hpp" -#include "chat.hpp" -#include "chat-template.hpp" +#include "chat.h" #include #include @@ -58,6 +58,32 @@ static T json_value(const json & body, const std::string & key, const T & defaul const static std::string build_info("b" + std::to_string(LLAMA_BUILD_NUMBER) + "-" + LLAMA_COMMIT); +// thin wrapper around common_grammar_trigger with (de)serialization functions +struct server_grammar_trigger { + common_grammar_trigger value; + + server_grammar_trigger() = default; + server_grammar_trigger(const common_grammar_trigger & value) : value(value) {} + server_grammar_trigger(const json & in) { + value.type = (common_grammar_trigger_type) in.at("type").get(); + value.value = in.at("value").get(); + if (value.type == COMMON_GRAMMAR_TRIGGER_TYPE_TOKEN) { + value.token = (llama_token) in.at("token").get(); + } + } + + json to_json() const { + json out { + {"type", (int) value.type}, + {"value", value.value}, + }; + if (value.type == COMMON_GRAMMAR_TRIGGER_TYPE_TOKEN) { + out["token"] = (int) value.token; + } + return out; + } +}; + // // tokenizer and input processing utils // @@ -347,41 +373,6 @@ static llama_tokens format_infill( return embd_inp; } -// Format given chat. If tmpl is empty, we take the template from model metadata -inline std::string format_chat(const common_chat_template & tmpl, const std::vector & messages) { - std::vector chat; - - for (size_t i = 0; i < messages.size(); ++i) { - const auto & curr_msg = messages[i]; - - std::string role = json_value(curr_msg, "role", std::string("")); - - std::string content; - if (curr_msg.contains("content")) { - if (curr_msg["content"].is_string()) { - content = curr_msg["content"].get(); - } else if (curr_msg["content"].is_array()) { - for (const auto & part : curr_msg["content"]) { - if (part.contains("text")) { - content += "\n" + part["text"].get(); - } - } - } else { - throw std::runtime_error("Invalid 'content' type (ref: https://github.com/ggerganov/llama.cpp/issues/8367)"); - } - } else { - throw std::runtime_error("Missing 'content' (ref: https://github.com/ggerganov/llama.cpp/issues/8367)"); - } - - chat.push_back({role, content, /* tool_calls= */ {}}); - } - - const auto formatted_chat = common_chat_apply_template(tmpl, chat, true, /* use_jinja= */ false); - LOG_DBG("formatted_chat: '%s'\n", formatted_chat.c_str()); - - return formatted_chat; -} - // // base64 utils (TODO: move to common in the future) // @@ -470,6 +461,10 @@ static std::string gen_chatcmplid() { return "chatcmpl-" + random_string(); } +static std::string gen_tool_call_id() { + return random_string(); +} + // // other common utils // @@ -556,8 +551,13 @@ static json oaicompat_completion_params_parse(const json & body) { throw std::runtime_error("Only one completion choice is allowed"); } + // Handle "echo" field + if (json_value(body, "echo", false)) { + throw std::runtime_error("Only no echo is supported"); + } + // Params supported by OAI but unsupported by llama.cpp - static const std::vector unsupported_params { "best_of", "echo", "suffix" }; + static const std::vector unsupported_params { "best_of", "suffix" }; for (const auto & param : unsupported_params) { if (body.contains(param)) { throw std::runtime_error("Unsupported param: " + param); @@ -578,12 +578,10 @@ static json oaicompat_completion_params_parse(const json & body) { static json oaicompat_completion_params_parse( const json & body, /* openai api json semantics */ bool use_jinja, - const common_chat_templates & chat_templates) + common_reasoning_format reasoning_format, + const struct common_chat_templates * tmpls) { json llama_params; - const auto & tmpl = body.contains("tools") && chat_templates.template_tool_use - ? *chat_templates.template_tool_use - : *chat_templates.template_default; auto tools = json_value(body, "tools", json()); auto stream = json_value(body, "stream", false); @@ -609,61 +607,59 @@ static json oaicompat_completion_params_parse( llama_params["stop"] = json_value(body, "stop", json::array()); } + auto json_schema = json_value(body, "json_schema", json()); + auto grammar = json_value(body, "grammar", std::string()); + if (!json_schema.is_null() && !grammar.empty()) { + throw std::runtime_error("Cannot use both json_schema and grammar"); + } + // Handle "response_format" field if (body.contains("response_format")) { json response_format = json_value(body, "response_format", json::object()); std::string response_type = json_value(response_format, "type", std::string()); if (response_type == "json_object") { - llama_params["json_schema"] = json_value(response_format, "schema", json::object()); + json_schema = json_value(response_format, "schema", json::object()); } else if (response_type == "json_schema") { - json json_schema = json_value(response_format, "json_schema", json::object()); - llama_params["json_schema"] = json_value(json_schema, "schema", json::object()); + auto schema_wrapper = json_value(response_format, "json_schema", json::object()); + json_schema = json_value(schema_wrapper, "schema", json::object()); } else if (!response_type.empty() && response_type != "text") { throw std::runtime_error("response_format type must be one of \"text\" or \"json_object\", but got: " + response_type); } } - // Apply chat template to the list of messages - if (use_jinja) { - auto tool_choice = json_value(body, "tool_choice", std::string("auto")); - if (tool_choice != "none" && tool_choice != "auto" && tool_choice != "required") { - throw std::runtime_error("Invalid tool_choice: " + tool_choice); - } - if (tool_choice != "none" && llama_params.contains("grammar")) { - throw std::runtime_error("Cannot use custom grammar constraints with tools."); - } - common_chat_inputs inputs; - inputs.messages = body.at("messages"); - inputs.tools = tools; - inputs.tool_choice = tool_choice; - inputs.parallel_tool_calls = json_value(body, "parallel_tool_calls", false); - if (inputs.parallel_tool_calls && !tmpl.original_caps().supports_parallel_tool_calls) { - LOG_DBG("Disabling parallel_tool_calls because the template does not support it\n"); - inputs.parallel_tool_calls = false; - } - inputs.stream = stream; - // TODO: support mixing schema w/ tools beyond generic format. - inputs.json_schema = json_value(llama_params, "json_schema", json()); - auto chat_params = common_chat_params_init(tmpl, inputs); + common_chat_templates_inputs inputs; + inputs.messages = common_chat_msgs_parse_oaicompat(body.at("messages")); + inputs.tools = common_chat_tools_parse_oaicompat(tools); + inputs.tool_choice = common_chat_tool_choice_parse_oaicompat(json_value(body, "tool_choice", std::string("auto"))); + inputs.json_schema = json_schema.is_null() ? "" : json_schema.dump(); + inputs.grammar = grammar; + inputs.add_generation_prompt = json_value(body, "add_generation_prompt", true); + inputs.use_jinja = use_jinja; + inputs.parallel_tool_calls = json_value(body, "parallel_tool_calls", false); + inputs.extract_reasoning = reasoning_format != COMMON_REASONING_FORMAT_NONE; + inputs.add_generation_prompt = json_value(body, "add_generation_prompt", true); + if (!inputs.tools.empty() && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE && body.contains("grammar")) { + throw std::runtime_error("Cannot use custom grammar constraints with tools."); + } - llama_params["chat_format"] = static_cast(chat_params.format); - llama_params["prompt"] = chat_params.prompt; + // Apply chat template to the list of messages + auto chat_params = common_chat_templates_apply(tmpls, inputs); + + llama_params["chat_format"] = static_cast(chat_params.format); + llama_params["prompt"] = chat_params.prompt; + if (!chat_params.grammar.empty()) { llama_params["grammar"] = chat_params.grammar; - llama_params["grammar_lazy"] = chat_params.grammar_lazy; - auto grammar_triggers = json::array(); - for (const auto & trigger : chat_params.grammar_triggers) { - grammar_triggers.push_back({ - {"word", trigger.word}, - {"at_start", trigger.at_start}, - }); - } - llama_params["grammar_triggers"] = grammar_triggers; - llama_params["preserved_tokens"] = chat_params.preserved_tokens; - for (const auto & stop : chat_params.additional_stops) { - llama_params["stop"].push_back(stop); - } - } else { - llama_params["prompt"] = format_chat(tmpl, body.at("messages")); + } + llama_params["grammar_lazy"] = chat_params.grammar_lazy; + auto grammar_triggers = json::array(); + for (const auto & trigger : chat_params.grammar_triggers) { + server_grammar_trigger ct(trigger); + grammar_triggers.push_back(ct.to_json()); + } + llama_params["grammar_triggers"] = grammar_triggers; + llama_params["preserved_tokens"] = chat_params.preserved_tokens; + for (const auto & stop : chat_params.additional_stops) { + llama_params["stop"].push_back(stop); } // Handle "n" field @@ -735,29 +731,51 @@ static json format_embeddings_response_oaicompat(const json & request, const jso return res; } -static json format_response_rerank(const json & request, const json & ranks) { - json data = json::array(); - int32_t n_tokens = 0; - int i = 0; - for (const auto & rank : ranks) { - data.push_back(json{ - {"index", i++}, - {"relevance_score", json_value(rank, "score", 0.0)}, - }); +static json format_response_rerank( + const json & request, + const json & ranks, + bool is_tei_format, + std::vector & texts) { + json res; + if (is_tei_format) { + // TEI response format + res = json::array(); + bool return_text = json_value(request, "return_text", false); + for (const auto & rank : ranks) { + int index = json_value(rank, "index", 0); + json elem = json{ + {"index", index}, + {"score", json_value(rank, "score", 0.0)}, + }; + if (return_text) { + elem["text"] = std::move(texts[index]); + } + res.push_back(elem); + } + } else { + // Jina response format + json results = json::array(); + int32_t n_tokens = 0; + for (const auto & rank : ranks) { + results.push_back(json{ + {"index", json_value(rank, "index", 0)}, + {"relevance_score", json_value(rank, "score", 0.0)}, + }); - n_tokens += json_value(rank, "tokens_evaluated", 0); + n_tokens += json_value(rank, "tokens_evaluated", 0); + } + + res = json{ + {"model", json_value(request, "model", std::string(DEFAULT_OAICOMPAT_MODEL))}, + {"object", "list"}, + {"usage", json{ + {"prompt_tokens", n_tokens}, + {"total_tokens", n_tokens} + }}, + {"results", results} + }; } - json res = json { - {"model", json_value(request, "model", std::string(DEFAULT_OAICOMPAT_MODEL))}, - {"object", "list"}, - {"usage", json { - {"prompt_tokens", n_tokens}, - {"total_tokens", n_tokens} - }}, - {"results", data} - }; - return res; } diff --git a/examples/server/webui/package-lock.json b/examples/server/webui/package-lock.json index c6c5de3c0..b2e3cf94a 100644 --- a/examples/server/webui/package-lock.json +++ b/examples/server/webui/package-lock.json @@ -10,9 +10,12 @@ "dependencies": { "@heroicons/react": "^2.2.0", "@sec-ant/readable-stream": "^0.6.0", + "@tailwindcss/postcss": "^4.1.1", + "@tailwindcss/vite": "^4.1.1", "@vscode/markdown-it-katex": "^1.1.1", "autoprefixer": "^10.4.20", - "daisyui": "^4.12.14", + "daisyui": "^5.0.12", + "dexie": "^4.0.11", "highlight.js": "^11.10.0", "katex": "^0.16.15", "postcss": "^8.4.49", @@ -25,7 +28,7 @@ "remark-breaks": "^4.0.0", "remark-gfm": "^4.0.0", "remark-math": "^6.0.0", - "tailwindcss": "^3.4.15", + "tailwindcss": "^4.1.1", "textlinestream": "^1.1.1", "vite-plugin-singlefile": "^2.0.3" }, @@ -978,27 +981,11 @@ "url": "https://github.com/sponsors/nzakas" } }, - "node_modules/@isaacs/cliui": { - "version": "8.0.2", - "resolved": "https://registry.npmjs.org/@isaacs/cliui/-/cliui-8.0.2.tgz", - "integrity": "sha512-O8jcjabXaleOG9DQ0+ARXWZBTfnP4WNAqzuiJK7ll44AmxGKv/J2M4TPjxjY3znBCfvBXFzucm1twdyFybFqEA==", - "license": "ISC", - "dependencies": { - "string-width": "^5.1.2", - "string-width-cjs": "npm:string-width@^4.2.0", - "strip-ansi": "^7.0.1", - "strip-ansi-cjs": "npm:strip-ansi@^6.0.1", - "wrap-ansi": "^8.1.0", - "wrap-ansi-cjs": "npm:wrap-ansi@^7.0.0" - }, - "engines": { - "node": ">=12" - } - }, "node_modules/@jridgewell/gen-mapping": { "version": "0.3.8", "resolved": "https://registry.npmjs.org/@jridgewell/gen-mapping/-/gen-mapping-0.3.8.tgz", "integrity": "sha512-imAbBGkb+ebQyxKgzv5Hu2nmROxoDOXHh80evxdoXNOrvAnVx7zimzc1Oo5h9RlfV4vPXaE2iM5pOFbvOCClWA==", + "dev": true, "license": "MIT", "dependencies": { "@jridgewell/set-array": "^1.2.1", @@ -1013,6 +1000,7 @@ "version": "3.1.2", "resolved": "https://registry.npmjs.org/@jridgewell/resolve-uri/-/resolve-uri-3.1.2.tgz", "integrity": "sha512-bRISgCIjP20/tbWSPWMEi54QVPRZExkuD9lJL+UIxUKtwVJA8wW1Trb1jMs1RFXo1CBTNZ/5hpC9QvmKWdopKw==", + "dev": true, "license": "MIT", "engines": { "node": ">=6.0.0" @@ -1022,6 +1010,7 @@ "version": "1.2.1", "resolved": "https://registry.npmjs.org/@jridgewell/set-array/-/set-array-1.2.1.tgz", "integrity": "sha512-R8gLRTZeyp03ymzP/6Lil/28tGeGEzhx1q2k703KGWRAI1VdvPIXdG70VJc2pAMw3NA6JKL5hhFu1sJX0Mnn/A==", + "dev": true, "license": "MIT", "engines": { "node": ">=6.0.0" @@ -1031,12 +1020,14 @@ "version": "1.5.0", "resolved": "https://registry.npmjs.org/@jridgewell/sourcemap-codec/-/sourcemap-codec-1.5.0.tgz", "integrity": "sha512-gv3ZRaISU3fjPAgNsriBRqGWQL6quFx04YMPW/zD8XMLsU32mhCCbfbO6KZFLjvYpCZ8zyDEgqsgf+PwPaM7GQ==", + "dev": true, "license": "MIT" }, "node_modules/@jridgewell/trace-mapping": { "version": "0.3.25", "resolved": "https://registry.npmjs.org/@jridgewell/trace-mapping/-/trace-mapping-0.3.25.tgz", "integrity": "sha512-vNk6aEwybGtawWmy/PzwnGDOjCkLWSD2wqvjGGAgOAwCGWySYXfYoxt00IJkTF+8Lb57DwOb3Aa0o9CApepiYQ==", + "dev": true, "license": "MIT", "dependencies": { "@jridgewell/resolve-uri": "^3.1.0", @@ -1047,6 +1038,7 @@ "version": "2.1.5", "resolved": "https://registry.npmjs.org/@nodelib/fs.scandir/-/fs.scandir-2.1.5.tgz", "integrity": "sha512-vq24Bq3ym5HEQm2NKCr3yXDwjc7vTsEThRDnkp2DK9p1uqLR+DHurm/NOTo0KG7HYHU7eppKZj3MyqYuMBf62g==", + "dev": true, "license": "MIT", "dependencies": { "@nodelib/fs.stat": "2.0.5", @@ -1060,6 +1052,7 @@ "version": "2.0.5", "resolved": "https://registry.npmjs.org/@nodelib/fs.stat/-/fs.stat-2.0.5.tgz", "integrity": "sha512-RkhPPp2zrqDAQA/2jNhnztcPAlv64XdhIp7a7454A5ovI7Bukxgt7MX7udwAu3zg1DcpPU0rz3VV1SeaqvY4+A==", + "dev": true, "license": "MIT", "engines": { "node": ">= 8" @@ -1069,6 +1062,7 @@ "version": "1.2.8", "resolved": "https://registry.npmjs.org/@nodelib/fs.walk/-/fs.walk-1.2.8.tgz", "integrity": "sha512-oGB+UxlgWcgQkgwo8GcEGwemoTFt3FIO9ababBmaGwXIoBKZ+GTy0pP185beGg7Llih/NSHSV2XAs1lnznocSg==", + "dev": true, "license": "MIT", "dependencies": { "@nodelib/fs.scandir": "2.1.5", @@ -1078,16 +1072,6 @@ "node": ">= 8" } }, - "node_modules/@pkgjs/parseargs": { - "version": "0.11.0", - "resolved": "https://registry.npmjs.org/@pkgjs/parseargs/-/parseargs-0.11.0.tgz", - "integrity": "sha512-+1VkjdD0QBLPodGrJUeqarH8VAIvQODIbwh9XpP5Syisf7YoQgsJKPNFoqqLQlu+VQ/tVSshMR6loPMn8U+dPg==", - "license": "MIT", - "optional": true, - "engines": { - "node": ">=14" - } - }, "node_modules/@rollup/rollup-android-arm-eabi": { "version": "4.34.2", "resolved": "https://registry.npmjs.org/@rollup/rollup-android-arm-eabi/-/rollup-android-arm-eabi-4.34.2.tgz", @@ -1341,6 +1325,243 @@ "integrity": "sha512-uiBh8DrB5FN35gP6/o8JEhEQ7/ci1jUsOZO/VMUjyvTpjtV54VstOXVj1TvTj/wsT23pfX6butxxh3qufsW3+g==", "license": "MIT" }, + "node_modules/@tailwindcss/node": { + "version": "4.1.1", + "resolved": "https://registry.npmjs.org/@tailwindcss/node/-/node-4.1.1.tgz", + "integrity": "sha512-xvlh4pvfG/bkv0fEtJDABAm1tjtSmSyi2QmS4zyj1EKNI1UiOYiUq1IphSwDsNJ5vJ9cWEGs4rJXpUdCN2kujQ==", + "license": "MIT", + "dependencies": { + "enhanced-resolve": "^5.18.1", + "jiti": "^2.4.2", + "lightningcss": "1.29.2", + "tailwindcss": "4.1.1" + } + }, + "node_modules/@tailwindcss/oxide": { + "version": "4.1.1", + "resolved": "https://registry.npmjs.org/@tailwindcss/oxide/-/oxide-4.1.1.tgz", + "integrity": "sha512-7+YBgnPQ4+jv6B6WVOerJ6WOzDzNJXrRKDts674v6TKAqFqYRr9+EBtSziO7nNcwQ8JtoZNMeqA+WJDjtCM/7w==", + "license": "MIT", + "engines": { + "node": ">= 10" + }, + "optionalDependencies": { + "@tailwindcss/oxide-android-arm64": "4.1.1", + "@tailwindcss/oxide-darwin-arm64": "4.1.1", + "@tailwindcss/oxide-darwin-x64": "4.1.1", + "@tailwindcss/oxide-freebsd-x64": "4.1.1", + "@tailwindcss/oxide-linux-arm-gnueabihf": "4.1.1", + "@tailwindcss/oxide-linux-arm64-gnu": "4.1.1", + "@tailwindcss/oxide-linux-arm64-musl": "4.1.1", + "@tailwindcss/oxide-linux-x64-gnu": "4.1.1", + "@tailwindcss/oxide-linux-x64-musl": "4.1.1", + "@tailwindcss/oxide-win32-arm64-msvc": "4.1.1", + "@tailwindcss/oxide-win32-x64-msvc": "4.1.1" + } + }, + "node_modules/@tailwindcss/oxide-android-arm64": { + "version": "4.1.1", + "resolved": "https://registry.npmjs.org/@tailwindcss/oxide-android-arm64/-/oxide-android-arm64-4.1.1.tgz", + "integrity": "sha512-gTyRzfdParpoCU1yyUC/iN6XK6T0Ra4bDlF8Aeul5NP9cLzKEZDogdNVNGv5WZmCDkVol7qlex7TMmcfytMmmw==", + "cpu": [ + "arm64" + ], + "license": "MIT", + "optional": true, + "os": [ + "android" + ], + "engines": { + "node": ">= 10" + } + }, + "node_modules/@tailwindcss/oxide-darwin-arm64": { + "version": "4.1.1", + "resolved": "https://registry.npmjs.org/@tailwindcss/oxide-darwin-arm64/-/oxide-darwin-arm64-4.1.1.tgz", + "integrity": "sha512-dI0QbdMWBvLB3MtaTKetzUKG9CUUQow8JSP4Nm+OxVokeZ+N+f1OmZW/hW1LzMxpx9RQCBgSRL+IIvKRat5Wdg==", + "cpu": [ + "arm64" + ], + "license": "MIT", + "optional": true, + "os": [ + "darwin" + ], + "engines": { + "node": ">= 10" + } + }, + "node_modules/@tailwindcss/oxide-darwin-x64": { + "version": "4.1.1", + "resolved": "https://registry.npmjs.org/@tailwindcss/oxide-darwin-x64/-/oxide-darwin-x64-4.1.1.tgz", + "integrity": "sha512-2Y+NPQOTRBCItshPgY/CWg4bKi7E9evMg4bgdb6h9iZObCZLOe3doPcuSxGS3DB0dKyMFKE8pTdWtFUbxZBMSA==", + "cpu": [ + "x64" + ], + "license": "MIT", + "optional": true, + "os": [ + "darwin" + ], + "engines": { + "node": ">= 10" + } + }, + "node_modules/@tailwindcss/oxide-freebsd-x64": { + "version": "4.1.1", + "resolved": "https://registry.npmjs.org/@tailwindcss/oxide-freebsd-x64/-/oxide-freebsd-x64-4.1.1.tgz", + "integrity": "sha512-N97NGMsB/7CHShbc5ube4dcsW/bYENkBrg8yWi8ieN9boYVRdw3cZviVryV/Nfu9bKbBV9kUvduFF2qBI7rEqg==", + "cpu": [ + "x64" + ], + "license": "MIT", + "optional": true, + "os": [ + "freebsd" + ], + "engines": { + "node": ">= 10" + } + }, + "node_modules/@tailwindcss/oxide-linux-arm-gnueabihf": { + "version": "4.1.1", + "resolved": "https://registry.npmjs.org/@tailwindcss/oxide-linux-arm-gnueabihf/-/oxide-linux-arm-gnueabihf-4.1.1.tgz", + "integrity": "sha512-33Lk6KbHnUZbXqza6RWNFo9wqPQ4+H5BAn1CkUUfC1RZ1vYbyDN6+iJPj53wmnWJ3mhRI8jWt3Jt1fO02IVdUQ==", + "cpu": [ + "arm" + ], + "license": "MIT", + "optional": true, + "os": [ + "linux" + ], + "engines": { + "node": ">= 10" + } + }, + "node_modules/@tailwindcss/oxide-linux-arm64-gnu": { + "version": "4.1.1", + "resolved": "https://registry.npmjs.org/@tailwindcss/oxide-linux-arm64-gnu/-/oxide-linux-arm64-gnu-4.1.1.tgz", + "integrity": "sha512-LyW35RzSUy+80WYScv03HKasAUmMFDaSbNpWfk1gG5gEE9kuRGnDzSrqMoLAmY/kzMCYP/1kqmUiAx8EFLkI2A==", + "cpu": [ + "arm64" + ], + "license": "MIT", + "optional": true, + "os": [ + "linux" + ], + "engines": { + "node": ">= 10" + } + }, + "node_modules/@tailwindcss/oxide-linux-arm64-musl": { + "version": "4.1.1", + "resolved": "https://registry.npmjs.org/@tailwindcss/oxide-linux-arm64-musl/-/oxide-linux-arm64-musl-4.1.1.tgz", + "integrity": "sha512-1KPnDMlHdqjPTUSFjx55pafvs8RZXRgxfeRgUrukwDKkuj7gFk28vW3Mx65YdiugAc9NWs3VgueZWaM1Po6uGw==", + "cpu": [ + "arm64" + ], + "license": "MIT", + "optional": true, + "os": [ + "linux" + ], + "engines": { + "node": ">= 10" + } + }, + "node_modules/@tailwindcss/oxide-linux-x64-gnu": { + "version": "4.1.1", + "resolved": "https://registry.npmjs.org/@tailwindcss/oxide-linux-x64-gnu/-/oxide-linux-x64-gnu-4.1.1.tgz", + "integrity": "sha512-4WdzA+MRlsinEEE6yxNMLJxpw0kE9XVipbAKdTL8BeUpyC2TdA3TL46lBulXzKp3BIxh3nqyR/UCqzl5o+3waQ==", + "cpu": [ + "x64" + ], + "license": "MIT", + "optional": true, + "os": [ + "linux" + ], + "engines": { + "node": ">= 10" + } + }, + "node_modules/@tailwindcss/oxide-linux-x64-musl": { + "version": "4.1.1", + "resolved": "https://registry.npmjs.org/@tailwindcss/oxide-linux-x64-musl/-/oxide-linux-x64-musl-4.1.1.tgz", + "integrity": "sha512-q7Ugbw3ARcjCW2VMUYrcMbJ6aMQuWPArBBE2EqC/swPZTdGADvMQSlvR0VKusUM4HoSsO7ZbvcZ53YwR57+AKw==", + "cpu": [ + "x64" + ], + "license": "MIT", + "optional": true, + "os": [ + "linux" + ], + "engines": { + "node": ">= 10" + } + }, + "node_modules/@tailwindcss/oxide-win32-arm64-msvc": { + "version": "4.1.1", + "resolved": "https://registry.npmjs.org/@tailwindcss/oxide-win32-arm64-msvc/-/oxide-win32-arm64-msvc-4.1.1.tgz", + "integrity": "sha512-0KpqsovgHcIzm7eAGzzEZsEs0/nPYXnRBv+aPq/GehpNQuE/NAQu+YgZXIIof+VflDFuyXOEnaFr7T5MZ1INhA==", + "cpu": [ + "arm64" + ], + "license": "MIT", + "optional": true, + "os": [ + "win32" + ], + "engines": { + "node": ">= 10" + } + }, + "node_modules/@tailwindcss/oxide-win32-x64-msvc": { + "version": "4.1.1", + "resolved": "https://registry.npmjs.org/@tailwindcss/oxide-win32-x64-msvc/-/oxide-win32-x64-msvc-4.1.1.tgz", + "integrity": "sha512-B1mjeXNS26kBOHv5sXARf6Wd0PWHV9x1TDlW0ummrBUOUAxAy5wcy4Nii1wzNvCdvC448hgiL06ylhwAbNthmg==", + "cpu": [ + "x64" + ], + "license": "MIT", + "optional": true, + "os": [ + "win32" + ], + "engines": { + "node": ">= 10" + } + }, + "node_modules/@tailwindcss/postcss": { + "version": "4.1.1", + "resolved": "https://registry.npmjs.org/@tailwindcss/postcss/-/postcss-4.1.1.tgz", + "integrity": "sha512-GX9AEM+msH0i2Yh1b6CuDRaZRo3kmbvIrLbSfvJ53C3uaAgsQ//fTQAh9HMQ6t1a9zvoUptlYqG//plWsBQTCw==", + "license": "MIT", + "dependencies": { + "@alloc/quick-lru": "^5.2.0", + "@tailwindcss/node": "4.1.1", + "@tailwindcss/oxide": "4.1.1", + "postcss": "^8.4.41", + "tailwindcss": "4.1.1" + } + }, + "node_modules/@tailwindcss/vite": { + "version": "4.1.1", + "resolved": "https://registry.npmjs.org/@tailwindcss/vite/-/vite-4.1.1.tgz", + "integrity": "sha512-tFTkRZwXq4XKr3S2dUZBxy80wbWYHdDSsu4QOB1yE1HJFKjfxKVpXtup4dyTVdQcLInoHC9lZXFPHnjoBP774g==", + "license": "MIT", + "dependencies": { + "@tailwindcss/node": "4.1.1", + "@tailwindcss/oxide": "4.1.1", + "tailwindcss": "4.1.1" + }, + "peerDependencies": { + "vite": "^5.2.0 || ^6" + } + }, "node_modules/@types/babel__core": { "version": "7.20.5", "resolved": "https://registry.npmjs.org/@types/babel__core/-/babel__core-7.20.5.tgz", @@ -1814,22 +2035,11 @@ "url": "https://github.com/sponsors/epoberezkin" } }, - "node_modules/ansi-regex": { - "version": "6.1.0", - "resolved": "https://registry.npmjs.org/ansi-regex/-/ansi-regex-6.1.0.tgz", - "integrity": "sha512-7HSX4QQb4CspciLpVFwyRe79O3xsIZDDLER21kERQ71oaPodF8jL725AgJMFAYbooIqolJoRLuM81SpeUkpkvA==", - "license": "MIT", - "engines": { - "node": ">=12" - }, - "funding": { - "url": "https://github.com/chalk/ansi-regex?sponsor=1" - } - }, "node_modules/ansi-styles": { "version": "4.3.0", "resolved": "https://registry.npmjs.org/ansi-styles/-/ansi-styles-4.3.0.tgz", "integrity": "sha512-zbB9rCJAT1rbjiVDb2hqKFHNYLxgtk8NURxZ3IZwD3F6NtxbXZQCnnSi1Lkx+IDohdPlFp222wVALIheZJQSEg==", + "dev": true, "license": "MIT", "dependencies": { "color-convert": "^2.0.1" @@ -1841,31 +2051,6 @@ "url": "https://github.com/chalk/ansi-styles?sponsor=1" } }, - "node_modules/any-promise": { - "version": "1.3.0", - "resolved": "https://registry.npmjs.org/any-promise/-/any-promise-1.3.0.tgz", - "integrity": "sha512-7UvmKalWRt1wgjL1RrGxoSJW/0QZFIegpeGvZG9kjp8vrRu55XTHbwnqq2GpXm9uLbcuhxm3IqX9OB4MZR1b2A==", - "license": "MIT" - }, - "node_modules/anymatch": { - "version": "3.1.3", - "resolved": "https://registry.npmjs.org/anymatch/-/anymatch-3.1.3.tgz", - "integrity": "sha512-KMReFUr0B4t+D+OBkjR3KYqvocp2XaSzO55UcB6mgQMd3KbcE+mWTyvVV7D/zsdEbNnV6acZUutkiHQXvTr1Rw==", - "license": "ISC", - "dependencies": { - "normalize-path": "^3.0.0", - "picomatch": "^2.0.4" - }, - "engines": { - "node": ">= 8" - } - }, - "node_modules/arg": { - "version": "5.0.2", - "resolved": "https://registry.npmjs.org/arg/-/arg-5.0.2.tgz", - "integrity": "sha512-PYjyFOLKQ9y57JvQ6QLo8dAgNqswh8M1RMJYdQduT6xbWSgK36P/Z/v+p888pM69jMMfS8Xd8F6I1kQ/I9HUGg==", - "license": "MIT" - }, "node_modules/argparse": { "version": "2.0.1", "resolved": "https://registry.npmjs.org/argparse/-/argparse-2.0.1.tgz", @@ -1924,20 +2109,9 @@ "version": "1.0.2", "resolved": "https://registry.npmjs.org/balanced-match/-/balanced-match-1.0.2.tgz", "integrity": "sha512-3oSeUO0TMV67hN1AmbXsK4yaqU7tjiHlbxRDZOpH0KW9+CeX4bRAaX0Anxt0tx2MrpRpWwQaPwIlISEJhYU5Pw==", + "dev": true, "license": "MIT" }, - "node_modules/binary-extensions": { - "version": "2.3.0", - "resolved": "https://registry.npmjs.org/binary-extensions/-/binary-extensions-2.3.0.tgz", - "integrity": "sha512-Ceh+7ox5qe7LJuLHoY0feh3pHuUDHAcRUeyL2VYghZwfpkNIy/+8Ocg0a3UuSoYzavmylwuLWQOf3hl0jjMMIw==", - "license": "MIT", - "engines": { - "node": ">=8" - }, - "funding": { - "url": "https://github.com/sponsors/sindresorhus" - } - }, "node_modules/brace-expansion": { "version": "1.1.11", "resolved": "https://registry.npmjs.org/brace-expansion/-/brace-expansion-1.1.11.tgz", @@ -2010,15 +2184,6 @@ "node": ">=6" } }, - "node_modules/camelcase-css": { - "version": "2.0.1", - "resolved": "https://registry.npmjs.org/camelcase-css/-/camelcase-css-2.0.1.tgz", - "integrity": "sha512-QOSvevhslijgYwRx6Rv7zKdMF8lbRmx+uQGx2+vDc+KI/eBnsy9kit5aj23AgGu3pa4t9AgwbnXWqS+iOY+2aA==", - "license": "MIT", - "engines": { - "node": ">= 6" - } - }, "node_modules/caniuse-lite": { "version": "1.0.30001697", "resolved": "https://registry.npmjs.org/caniuse-lite/-/caniuse-lite-1.0.30001697.tgz", @@ -2106,46 +2271,11 @@ "url": "https://github.com/sponsors/wooorm" } }, - "node_modules/chokidar": { - "version": "3.6.0", - "resolved": "https://registry.npmjs.org/chokidar/-/chokidar-3.6.0.tgz", - "integrity": "sha512-7VT13fmjotKpGipCW9JEQAusEPE+Ei8nl6/g4FBAmIm0GOOLMua9NDDo/DWp0ZAxCr3cPq5ZpBqmPAQgDda2Pw==", - "license": "MIT", - "dependencies": { - "anymatch": "~3.1.2", - "braces": "~3.0.2", - "glob-parent": "~5.1.2", - "is-binary-path": "~2.1.0", - "is-glob": "~4.0.1", - "normalize-path": "~3.0.0", - "readdirp": "~3.6.0" - }, - "engines": { - "node": ">= 8.10.0" - }, - "funding": { - "url": "https://paulmillr.com/funding/" - }, - "optionalDependencies": { - "fsevents": "~2.3.2" - } - }, - "node_modules/chokidar/node_modules/glob-parent": { - "version": "5.1.2", - "resolved": "https://registry.npmjs.org/glob-parent/-/glob-parent-5.1.2.tgz", - "integrity": "sha512-AOIgSQCepiJYwP3ARnGx+5VnTu2HBYdzbGP45eLw1vr3zB3vZLeyed1sC9hnbcOc9/SrMyM5RPQrkGz4aS9Zow==", - "license": "ISC", - "dependencies": { - "is-glob": "^4.0.1" - }, - "engines": { - "node": ">= 6" - } - }, "node_modules/color-convert": { "version": "2.0.1", "resolved": "https://registry.npmjs.org/color-convert/-/color-convert-2.0.1.tgz", "integrity": "sha512-RRECPsj7iu/xb5oKYcsFHSppFNnsj/52OVTRKb4zP5onXwVF3zVmmToNcOfGC+CRDpfK/U584fMg38ZHCaElKQ==", + "dev": true, "license": "MIT", "dependencies": { "color-name": "~1.1.4" @@ -2158,6 +2288,7 @@ "version": "1.1.4", "resolved": "https://registry.npmjs.org/color-name/-/color-name-1.1.4.tgz", "integrity": "sha512-dOy+3AuW3a2wNbZHIuMZpTcgjGuLU/uBL/ubcZF9OXbDo8ff4O8yVp5Bf0efS8uEoYo5q4Fx7dY9OgQGXgAsQA==", + "dev": true, "license": "MIT" }, "node_modules/colorjs.io": { @@ -2213,6 +2344,7 @@ "version": "7.0.6", "resolved": "https://registry.npmjs.org/cross-spawn/-/cross-spawn-7.0.6.tgz", "integrity": "sha512-uV2QOWP2nWzsy2aMp8aRibhi9dlzF5Hgh5SHaB9OiTGEyDTiJJyx0uy51QXdyWbtAHNua4XJzUKca3OzKUd3vA==", + "dev": true, "license": "MIT", "dependencies": { "path-key": "^3.1.0", @@ -2223,60 +2355,19 @@ "node": ">= 8" } }, - "node_modules/css-selector-tokenizer": { - "version": "0.8.0", - "resolved": "https://registry.npmjs.org/css-selector-tokenizer/-/css-selector-tokenizer-0.8.0.tgz", - "integrity": "sha512-Jd6Ig3/pe62/qe5SBPTN8h8LeUg/pT4lLgtavPf7updwwHpvFzxvOQBHYj2LZDMjUnBzgvIUSjRcf6oT5HzHFg==", - "license": "MIT", - "dependencies": { - "cssesc": "^3.0.0", - "fastparse": "^1.1.2" - } - }, - "node_modules/cssesc": { - "version": "3.0.0", - "resolved": "https://registry.npmjs.org/cssesc/-/cssesc-3.0.0.tgz", - "integrity": "sha512-/Tb/JcjK111nNScGob5MNtsntNM1aCNUDipB/TkwZFhyDrrE47SOx/18wF2bbjgc3ZzCSKW1T5nt5EbFoAz/Vg==", - "license": "MIT", - "bin": { - "cssesc": "bin/cssesc" - }, - "engines": { - "node": ">=4" - } - }, "node_modules/csstype": { "version": "3.1.3", "resolved": "https://registry.npmjs.org/csstype/-/csstype-3.1.3.tgz", "integrity": "sha512-M1uQkMl8rQK/szD0LNhtqxIPLpimGm8sOBwU7lLnCpSbTyY3yeU1Vc7l4KT5zT4s/yOxHH5O7tIuuLOCnLADRw==", "license": "MIT" }, - "node_modules/culori": { - "version": "3.3.0", - "resolved": "https://registry.npmjs.org/culori/-/culori-3.3.0.tgz", - "integrity": "sha512-pHJg+jbuFsCjz9iclQBqyL3B2HLCBF71BwVNujUYEvCeQMvV97R59MNK3R2+jgJ3a1fcZgI9B3vYgz8lzr/BFQ==", - "license": "MIT", - "engines": { - "node": "^12.20.0 || ^14.13.1 || >=16.0.0" - } - }, "node_modules/daisyui": { - "version": "4.12.23", - "resolved": "https://registry.npmjs.org/daisyui/-/daisyui-4.12.23.tgz", - "integrity": "sha512-EM38duvxutJ5PD65lO/AFMpcw+9qEy6XAZrTpzp7WyaPeO/l+F/Qiq0ECHHmFNcFXh5aVoALY4MGrrxtCiaQCQ==", + "version": "5.0.12", + "resolved": "https://registry.npmjs.org/daisyui/-/daisyui-5.0.12.tgz", + "integrity": "sha512-01DU0eYBcHgPtuf5fxcrkGkIN6/Uyaqmkle5Yo3ZyW9YVAu036ALZbjv2KH5euvUbeQ4r9q3gAarGcf7Tywhng==", "license": "MIT", - "dependencies": { - "css-selector-tokenizer": "^0.8", - "culori": "^3", - "picocolors": "^1", - "postcss-js": "^4" - }, - "engines": { - "node": ">=16.9.0" - }, "funding": { - "type": "opencollective", - "url": "https://opencollective.com/daisyui" + "url": "https://github.com/saadeghi/daisyui?sponsor=1" } }, "node_modules/debug": { @@ -2325,6 +2416,15 @@ "node": ">=6" } }, + "node_modules/detect-libc": { + "version": "2.0.3", + "resolved": "https://registry.npmjs.org/detect-libc/-/detect-libc-2.0.3.tgz", + "integrity": "sha512-bwy0MGW55bG41VqxxypOsdSdGqLwXPI/focwgTYCFMbdUiBAxLg9CFzG08sz2aqzknwiX7Hkl0bQENjg8iLByw==", + "license": "Apache-2.0", + "engines": { + "node": ">=8" + } + }, "node_modules/devlop": { "version": "1.1.0", "resolved": "https://registry.npmjs.org/devlop/-/devlop-1.1.0.tgz", @@ -2338,35 +2438,30 @@ "url": "https://github.com/sponsors/wooorm" } }, - "node_modules/didyoumean": { - "version": "1.2.2", - "resolved": "https://registry.npmjs.org/didyoumean/-/didyoumean-1.2.2.tgz", - "integrity": "sha512-gxtyfqMg7GKyhQmb056K7M3xszy/myH8w+B4RT+QXBQsvAOdc3XymqDDPHx1BgPgsdAA5SIifona89YtRATDzw==", + "node_modules/dexie": { + "version": "4.0.11", + "resolved": "https://registry.npmjs.org/dexie/-/dexie-4.0.11.tgz", + "integrity": "sha512-SOKO002EqlvBYYKQSew3iymBoN2EQ4BDw/3yprjh7kAfFzjBYkaMNa/pZvcA7HSWlcKSQb9XhPe3wKyQ0x4A8A==", "license": "Apache-2.0" }, - "node_modules/dlv": { - "version": "1.1.3", - "resolved": "https://registry.npmjs.org/dlv/-/dlv-1.1.3.tgz", - "integrity": "sha512-+HlytyjlPKnIG8XuRG8WvmBP8xs8P71y+SKKS6ZXWoEgLuePxtDoUEiH7WkdePWrQ5JBpE6aoVqfZfJUQkjXwA==", - "license": "MIT" - }, - "node_modules/eastasianwidth": { - "version": "0.2.0", - "resolved": "https://registry.npmjs.org/eastasianwidth/-/eastasianwidth-0.2.0.tgz", - "integrity": "sha512-I88TYZWc9XiYHRQ4/3c5rjjfgkjhLyW2luGIheGERbNQ6OY7yTybanSpDXZa8y7VUP9YmDcYa+eyq4ca7iLqWA==", - "license": "MIT" - }, "node_modules/electron-to-chromium": { "version": "1.5.91", "resolved": "https://registry.npmjs.org/electron-to-chromium/-/electron-to-chromium-1.5.91.tgz", "integrity": "sha512-sNSHHyq048PFmZY4S90ax61q+gLCs0X0YmcOII9wG9S2XwbVr+h4VW2wWhnbp/Eys3cCwTxVF292W3qPaxIapQ==", "license": "ISC" }, - "node_modules/emoji-regex": { - "version": "9.2.2", - "resolved": "https://registry.npmjs.org/emoji-regex/-/emoji-regex-9.2.2.tgz", - "integrity": "sha512-L18DaJsXSUk2+42pv8mLs5jJT2hqFkFE4j21wOmgbUqsZ2hL72NsUU785g9RXgo3s0ZNgVl42TiHp3ZtOv/Vyg==", - "license": "MIT" + "node_modules/enhanced-resolve": { + "version": "5.18.1", + "resolved": "https://registry.npmjs.org/enhanced-resolve/-/enhanced-resolve-5.18.1.tgz", + "integrity": "sha512-ZSW3ma5GkcQBIpwZTSRAI8N71Uuwgs93IezB7mf7R60tC8ZbJideoDNKjHn2O9KIlx6rkGTTEk1xUCK2E1Y2Yg==", + "license": "MIT", + "dependencies": { + "graceful-fs": "^4.2.4", + "tapable": "^2.2.0" + }, + "engines": { + "node": ">=10.13.0" + } }, "node_modules/entities": { "version": "4.5.0", @@ -2646,6 +2741,7 @@ "version": "3.3.3", "resolved": "https://registry.npmjs.org/fast-glob/-/fast-glob-3.3.3.tgz", "integrity": "sha512-7MptL8U0cqcFdzIzwOTHoilX9x5BrNqye7Z/LuC7kCMRio1EMSyqRK3BEAUD7sXRq4iT4AzTVuZdhgQ2TCvYLg==", + "dev": true, "license": "MIT", "dependencies": { "@nodelib/fs.stat": "^2.0.2", @@ -2662,6 +2758,7 @@ "version": "5.1.2", "resolved": "https://registry.npmjs.org/glob-parent/-/glob-parent-5.1.2.tgz", "integrity": "sha512-AOIgSQCepiJYwP3ARnGx+5VnTu2HBYdzbGP45eLw1vr3zB3vZLeyed1sC9hnbcOc9/SrMyM5RPQrkGz4aS9Zow==", + "dev": true, "license": "ISC", "dependencies": { "is-glob": "^4.0.1" @@ -2684,16 +2781,11 @@ "dev": true, "license": "MIT" }, - "node_modules/fastparse": { - "version": "1.1.2", - "resolved": "https://registry.npmjs.org/fastparse/-/fastparse-1.1.2.tgz", - "integrity": "sha512-483XLLxTVIwWK3QTrMGRqUfUpoOs/0hbQrl2oz4J0pAcm3A3bu84wxTFqGqkJzewCLdME38xJLJAxBABfQT8sQ==", - "license": "MIT" - }, "node_modules/fastq": { "version": "1.19.0", "resolved": "https://registry.npmjs.org/fastq/-/fastq-1.19.0.tgz", "integrity": "sha512-7SFSRCNjBQIZH/xZR3iy5iQYR8aGBE0h3VG6/cwlbrpdciNYBMotQav8c1XI3HjHH+NikUpP53nPdlZSdWmFzA==", + "dev": true, "license": "ISC", "dependencies": { "reusify": "^1.0.4" @@ -2762,22 +2854,6 @@ "dev": true, "license": "ISC" }, - "node_modules/foreground-child": { - "version": "3.3.0", - "resolved": "https://registry.npmjs.org/foreground-child/-/foreground-child-3.3.0.tgz", - "integrity": "sha512-Ld2g8rrAyMYFXBhEqMz8ZAHBi4J4uS1i/CxGMDnjyFWddMXLVcDp051DZfu+t7+ab7Wv6SMqpWmyFIj5UbfFvg==", - "license": "ISC", - "dependencies": { - "cross-spawn": "^7.0.0", - "signal-exit": "^4.0.1" - }, - "engines": { - "node": ">=14" - }, - "funding": { - "url": "https://github.com/sponsors/isaacs" - } - }, "node_modules/fraction.js": { "version": "4.3.7", "resolved": "https://registry.npmjs.org/fraction.js/-/fraction.js-4.3.7.tgz", @@ -2805,15 +2881,6 @@ "node": "^8.16.0 || ^10.6.0 || >=11.0.0" } }, - "node_modules/function-bind": { - "version": "1.1.2", - "resolved": "https://registry.npmjs.org/function-bind/-/function-bind-1.1.2.tgz", - "integrity": "sha512-7XHNxH7qX9xG5mIwxkhumTox/MIRNcOgDrxWsMt2pAr23WHp6MrRlN7FBSFpCpr+oVO0F744iUgR82nJMfG2SA==", - "license": "MIT", - "funding": { - "url": "https://github.com/sponsors/ljharb" - } - }, "node_modules/gensync": { "version": "1.0.0-beta.2", "resolved": "https://registry.npmjs.org/gensync/-/gensync-1.0.0-beta.2.tgz", @@ -2824,30 +2891,11 @@ "node": ">=6.9.0" } }, - "node_modules/glob": { - "version": "10.4.5", - "resolved": "https://registry.npmjs.org/glob/-/glob-10.4.5.tgz", - "integrity": "sha512-7Bv8RF0k6xjo7d4A/PxYLbUCfb6c+Vpd2/mB2yRDlew7Jb5hEXiCD9ibfO7wpk8i4sevK6DFny9h7EYbM3/sHg==", - "license": "ISC", - "dependencies": { - "foreground-child": "^3.1.0", - "jackspeak": "^3.1.2", - "minimatch": "^9.0.4", - "minipass": "^7.1.2", - "package-json-from-dist": "^1.0.0", - "path-scurry": "^1.11.1" - }, - "bin": { - "glob": "dist/esm/bin.mjs" - }, - "funding": { - "url": "https://github.com/sponsors/isaacs" - } - }, "node_modules/glob-parent": { "version": "6.0.2", "resolved": "https://registry.npmjs.org/glob-parent/-/glob-parent-6.0.2.tgz", "integrity": "sha512-XxwI8EOhVQgWp6iDL+3b0r86f4d6AX6zSU55HfB4ydCEuXLXc5FcYeOu+nnGftS4TEju/11rt4KJPTMgbfmv4A==", + "dev": true, "license": "ISC", "dependencies": { "is-glob": "^4.0.3" @@ -2856,30 +2904,6 @@ "node": ">=10.13.0" } }, - "node_modules/glob/node_modules/brace-expansion": { - "version": "2.0.1", - "resolved": "https://registry.npmjs.org/brace-expansion/-/brace-expansion-2.0.1.tgz", - "integrity": "sha512-XnAIvQ8eM+kC6aULx6wuQiwVsnzsi9d3WxzV3FpWTGA19F621kwdbsAcFKXgKUHZWsy+mY6iL1sHTxWEFCytDA==", - "license": "MIT", - "dependencies": { - "balanced-match": "^1.0.0" - } - }, - "node_modules/glob/node_modules/minimatch": { - "version": "9.0.5", - "resolved": "https://registry.npmjs.org/minimatch/-/minimatch-9.0.5.tgz", - "integrity": "sha512-G6T0ZX48xgozx7587koeX9Ys2NYy6Gmv//P89sEte9V9whIapMNF4idKxnW2QtCcLiTWlb/wfCabAtAFWhhBow==", - "license": "ISC", - "dependencies": { - "brace-expansion": "^2.0.1" - }, - "engines": { - "node": ">=16 || 14 >=14.17" - }, - "funding": { - "url": "https://github.com/sponsors/isaacs" - } - }, "node_modules/globals": { "version": "15.14.0", "resolved": "https://registry.npmjs.org/globals/-/globals-15.14.0.tgz", @@ -2893,6 +2917,12 @@ "url": "https://github.com/sponsors/sindresorhus" } }, + "node_modules/graceful-fs": { + "version": "4.2.11", + "resolved": "https://registry.npmjs.org/graceful-fs/-/graceful-fs-4.2.11.tgz", + "integrity": "sha512-RbJ5/jmFcNNCcDV5o9eTnBLJ/HszWV0P73bc+Ff4nS/rJj+YaS6IGyiOL0VoBYX+l1Wrl3k63h/KrH+nhJ0XvQ==", + "license": "ISC" + }, "node_modules/graphemer": { "version": "1.4.0", "resolved": "https://registry.npmjs.org/graphemer/-/graphemer-1.4.0.tgz", @@ -2910,18 +2940,6 @@ "node": ">=8" } }, - "node_modules/hasown": { - "version": "2.0.2", - "resolved": "https://registry.npmjs.org/hasown/-/hasown-2.0.2.tgz", - "integrity": "sha512-0hJU9SCPvmMzIBdZFqNPXWa6dqh7WdH0cII9y+CyS8rG3nL48Bclra9HmKhVVUHyPWNH5Y7xDwAB7bfgSjkUMQ==", - "license": "MIT", - "dependencies": { - "function-bind": "^1.1.2" - }, - "engines": { - "node": ">= 0.4" - } - }, "node_modules/hast-util-from-dom": { "version": "5.0.1", "resolved": "https://registry.npmjs.org/hast-util-from-dom/-/hast-util-from-dom-5.0.1.tgz", @@ -3183,33 +3201,6 @@ "url": "https://github.com/sponsors/wooorm" } }, - "node_modules/is-binary-path": { - "version": "2.1.0", - "resolved": "https://registry.npmjs.org/is-binary-path/-/is-binary-path-2.1.0.tgz", - "integrity": "sha512-ZMERYes6pDydyuGidse7OsHxtbI7WVeUEozgR/g7rd0xUimYNlvZRE/K2MgZTjWy725IfelLeVcEM97mmtRGXw==", - "license": "MIT", - "dependencies": { - "binary-extensions": "^2.0.0" - }, - "engines": { - "node": ">=8" - } - }, - "node_modules/is-core-module": { - "version": "2.16.1", - "resolved": "https://registry.npmjs.org/is-core-module/-/is-core-module-2.16.1.tgz", - "integrity": "sha512-UfoeMA6fIJ8wTYFEUjelnaGI67v6+N7qXJEvQuIGa99l4xsCruSYOVSQ0uPANn4dAzm8lkYPaKLrrijLq7x23w==", - "license": "MIT", - "dependencies": { - "hasown": "^2.0.2" - }, - "engines": { - "node": ">= 0.4" - }, - "funding": { - "url": "https://github.com/sponsors/ljharb" - } - }, "node_modules/is-decimal": { "version": "2.0.1", "resolved": "https://registry.npmjs.org/is-decimal/-/is-decimal-2.0.1.tgz", @@ -3224,24 +3215,17 @@ "version": "2.1.1", "resolved": "https://registry.npmjs.org/is-extglob/-/is-extglob-2.1.1.tgz", "integrity": "sha512-SbKbANkN603Vi4jEZv49LeVJMn4yGwsbzZworEoyEiutsN3nJYdbO36zfhGJ6QEDpOZIFkDtnq5JRxmvl3jsoQ==", + "dev": true, "license": "MIT", "engines": { "node": ">=0.10.0" } }, - "node_modules/is-fullwidth-code-point": { - "version": "3.0.0", - "resolved": "https://registry.npmjs.org/is-fullwidth-code-point/-/is-fullwidth-code-point-3.0.0.tgz", - "integrity": "sha512-zymm5+u+sCsSWyD9qNaejV3DFvhCKclKdizYaJUuHA83RLjb7nSuGnddCHGv0hk+KY7BMAlsWeK4Ueg6EV6XQg==", - "license": "MIT", - "engines": { - "node": ">=8" - } - }, "node_modules/is-glob": { "version": "4.0.3", "resolved": "https://registry.npmjs.org/is-glob/-/is-glob-4.0.3.tgz", "integrity": "sha512-xelSayHH36ZgE7ZWhli7pW34hNbNl8Ojv5KVmkJD4hBdD3th8Tfk9vYasLM+mXWOZhFkgZfxhLSnrwRr4elSSg==", + "dev": true, "license": "MIT", "dependencies": { "is-extglob": "^2.1.1" @@ -3285,30 +3269,16 @@ "version": "2.0.0", "resolved": "https://registry.npmjs.org/isexe/-/isexe-2.0.0.tgz", "integrity": "sha512-RHxMLp9lnKHGHRng9QFhRCMbYAcVpn69smSGcq3f36xjgVVWThj4qqLbTLlq7Ssj8B+fIQ1EuCEGI2lKsyQeIw==", + "dev": true, "license": "ISC" }, - "node_modules/jackspeak": { - "version": "3.4.3", - "resolved": "https://registry.npmjs.org/jackspeak/-/jackspeak-3.4.3.tgz", - "integrity": "sha512-OGlZQpz2yfahA/Rd1Y8Cd9SIEsqvXkLVoSw/cgwhnhFMDbsQFeZYoJJ7bIZBS9BcamUW96asq/npPWugM+RQBw==", - "license": "BlueOak-1.0.0", - "dependencies": { - "@isaacs/cliui": "^8.0.2" - }, - "funding": { - "url": "https://github.com/sponsors/isaacs" - }, - "optionalDependencies": { - "@pkgjs/parseargs": "^0.11.0" - } - }, "node_modules/jiti": { - "version": "1.21.7", - "resolved": "https://registry.npmjs.org/jiti/-/jiti-1.21.7.tgz", - "integrity": "sha512-/imKNG4EbWNrVjoNC/1H5/9GFy+tqjGBHCaSsN+P2RnPqjsLmv6UD3Ej+Kj8nBWaRAwyk7kK5ZUc+OEatnTR3A==", + "version": "2.4.2", + "resolved": "https://registry.npmjs.org/jiti/-/jiti-2.4.2.tgz", + "integrity": "sha512-rg9zJN+G4n2nfJl5MW3BMygZX56zKPNVEYYqq7adpmMh4Jn2QNEwhvQlFy6jPVdcod7txZtKHWnyZiA3a0zP7A==", "license": "MIT", "bin": { - "jiti": "bin/jiti.js" + "jiti": "lib/jiti-cli.mjs" } }, "node_modules/js-tokens": { @@ -3417,23 +3387,233 @@ "node": ">= 0.8.0" } }, - "node_modules/lilconfig": { - "version": "3.1.3", - "resolved": "https://registry.npmjs.org/lilconfig/-/lilconfig-3.1.3.tgz", - "integrity": "sha512-/vlFKAoH5Cgt3Ie+JLhRbwOsCQePABiU3tJ1egGvyQ+33R/vcwM2Zl2QR/LzjsBeItPt3oSVXapn+m4nQDvpzw==", - "license": "MIT", + "node_modules/lightningcss": { + "version": "1.29.2", + "resolved": "https://registry.npmjs.org/lightningcss/-/lightningcss-1.29.2.tgz", + "integrity": "sha512-6b6gd/RUXKaw5keVdSEtqFVdzWnU5jMxTUjA2bVcMNPLwSQ08Sv/UodBVtETLCn7k4S1Ibxwh7k68IwLZPgKaA==", + "license": "MPL-2.0", + "dependencies": { + "detect-libc": "^2.0.3" + }, "engines": { - "node": ">=14" + "node": ">= 12.0.0" }, "funding": { - "url": "https://github.com/sponsors/antonk52" + "type": "opencollective", + "url": "https://opencollective.com/parcel" + }, + "optionalDependencies": { + "lightningcss-darwin-arm64": "1.29.2", + "lightningcss-darwin-x64": "1.29.2", + "lightningcss-freebsd-x64": "1.29.2", + "lightningcss-linux-arm-gnueabihf": "1.29.2", + "lightningcss-linux-arm64-gnu": "1.29.2", + "lightningcss-linux-arm64-musl": "1.29.2", + "lightningcss-linux-x64-gnu": "1.29.2", + "lightningcss-linux-x64-musl": "1.29.2", + "lightningcss-win32-arm64-msvc": "1.29.2", + "lightningcss-win32-x64-msvc": "1.29.2" } }, - "node_modules/lines-and-columns": { - "version": "1.2.4", - "resolved": "https://registry.npmjs.org/lines-and-columns/-/lines-and-columns-1.2.4.tgz", - "integrity": "sha512-7ylylesZQ/PV29jhEDl3Ufjo6ZX7gCqJr5F7PKrqc93v7fzSymt1BpwEU8nAUXs8qzzvqhbjhK5QZg6Mt/HkBg==", - "license": "MIT" + "node_modules/lightningcss-darwin-arm64": { + "version": "1.29.2", + "resolved": "https://registry.npmjs.org/lightningcss-darwin-arm64/-/lightningcss-darwin-arm64-1.29.2.tgz", + "integrity": "sha512-cK/eMabSViKn/PG8U/a7aCorpeKLMlK0bQeNHmdb7qUnBkNPnL+oV5DjJUo0kqWsJUapZsM4jCfYItbqBDvlcA==", + "cpu": [ + "arm64" + ], + "license": "MPL-2.0", + "optional": true, + "os": [ + "darwin" + ], + "engines": { + "node": ">= 12.0.0" + }, + "funding": { + "type": "opencollective", + "url": "https://opencollective.com/parcel" + } + }, + "node_modules/lightningcss-darwin-x64": { + "version": "1.29.2", + "resolved": "https://registry.npmjs.org/lightningcss-darwin-x64/-/lightningcss-darwin-x64-1.29.2.tgz", + "integrity": "sha512-j5qYxamyQw4kDXX5hnnCKMf3mLlHvG44f24Qyi2965/Ycz829MYqjrVg2H8BidybHBp9kom4D7DR5VqCKDXS0w==", + "cpu": [ + "x64" + ], + "license": "MPL-2.0", + "optional": true, + "os": [ + "darwin" + ], + "engines": { + "node": ">= 12.0.0" + }, + "funding": { + "type": "opencollective", + "url": "https://opencollective.com/parcel" + } + }, + "node_modules/lightningcss-freebsd-x64": { + "version": "1.29.2", + "resolved": "https://registry.npmjs.org/lightningcss-freebsd-x64/-/lightningcss-freebsd-x64-1.29.2.tgz", + "integrity": "sha512-wDk7M2tM78Ii8ek9YjnY8MjV5f5JN2qNVO+/0BAGZRvXKtQrBC4/cn4ssQIpKIPP44YXw6gFdpUF+Ps+RGsCwg==", + "cpu": [ + "x64" + ], + "license": "MPL-2.0", + "optional": true, + "os": [ + "freebsd" + ], + "engines": { + "node": ">= 12.0.0" + }, + "funding": { + "type": "opencollective", + "url": "https://opencollective.com/parcel" + } + }, + "node_modules/lightningcss-linux-arm-gnueabihf": { + "version": "1.29.2", + "resolved": "https://registry.npmjs.org/lightningcss-linux-arm-gnueabihf/-/lightningcss-linux-arm-gnueabihf-1.29.2.tgz", + "integrity": "sha512-IRUrOrAF2Z+KExdExe3Rz7NSTuuJ2HvCGlMKoquK5pjvo2JY4Rybr+NrKnq0U0hZnx5AnGsuFHjGnNT14w26sg==", + "cpu": [ + "arm" + ], + "license": "MPL-2.0", + "optional": true, + "os": [ + "linux" + ], + "engines": { + "node": ">= 12.0.0" + }, + "funding": { + "type": "opencollective", + "url": "https://opencollective.com/parcel" + } + }, + "node_modules/lightningcss-linux-arm64-gnu": { + "version": "1.29.2", + "resolved": "https://registry.npmjs.org/lightningcss-linux-arm64-gnu/-/lightningcss-linux-arm64-gnu-1.29.2.tgz", + "integrity": "sha512-KKCpOlmhdjvUTX/mBuaKemp0oeDIBBLFiU5Fnqxh1/DZ4JPZi4evEH7TKoSBFOSOV3J7iEmmBaw/8dpiUvRKlQ==", + "cpu": [ + "arm64" + ], + "license": "MPL-2.0", + "optional": true, + "os": [ + "linux" + ], + "engines": { + "node": ">= 12.0.0" + }, + "funding": { + "type": "opencollective", + "url": "https://opencollective.com/parcel" + } + }, + "node_modules/lightningcss-linux-arm64-musl": { + "version": "1.29.2", + "resolved": "https://registry.npmjs.org/lightningcss-linux-arm64-musl/-/lightningcss-linux-arm64-musl-1.29.2.tgz", + "integrity": "sha512-Q64eM1bPlOOUgxFmoPUefqzY1yV3ctFPE6d/Vt7WzLW4rKTv7MyYNky+FWxRpLkNASTnKQUaiMJ87zNODIrrKQ==", + "cpu": [ + "arm64" + ], + "license": "MPL-2.0", + "optional": true, + "os": [ + "linux" + ], + "engines": { + "node": ">= 12.0.0" + }, + "funding": { + "type": "opencollective", + "url": "https://opencollective.com/parcel" + } + }, + "node_modules/lightningcss-linux-x64-gnu": { + "version": "1.29.2", + "resolved": "https://registry.npmjs.org/lightningcss-linux-x64-gnu/-/lightningcss-linux-x64-gnu-1.29.2.tgz", + "integrity": "sha512-0v6idDCPG6epLXtBH/RPkHvYx74CVziHo6TMYga8O2EiQApnUPZsbR9nFNrg2cgBzk1AYqEd95TlrsL7nYABQg==", + "cpu": [ + "x64" + ], + "license": "MPL-2.0", + "optional": true, + "os": [ + "linux" + ], + "engines": { + "node": ">= 12.0.0" + }, + "funding": { + "type": "opencollective", + "url": "https://opencollective.com/parcel" + } + }, + "node_modules/lightningcss-linux-x64-musl": { + "version": "1.29.2", + "resolved": "https://registry.npmjs.org/lightningcss-linux-x64-musl/-/lightningcss-linux-x64-musl-1.29.2.tgz", + "integrity": "sha512-rMpz2yawkgGT8RULc5S4WiZopVMOFWjiItBT7aSfDX4NQav6M44rhn5hjtkKzB+wMTRlLLqxkeYEtQ3dd9696w==", + "cpu": [ + "x64" + ], + "license": "MPL-2.0", + "optional": true, + "os": [ + "linux" + ], + "engines": { + "node": ">= 12.0.0" + }, + "funding": { + "type": "opencollective", + "url": "https://opencollective.com/parcel" + } + }, + "node_modules/lightningcss-win32-arm64-msvc": { + "version": "1.29.2", + "resolved": "https://registry.npmjs.org/lightningcss-win32-arm64-msvc/-/lightningcss-win32-arm64-msvc-1.29.2.tgz", + "integrity": "sha512-nL7zRW6evGQqYVu/bKGK+zShyz8OVzsCotFgc7judbt6wnB2KbiKKJwBE4SGoDBQ1O94RjW4asrCjQL4i8Fhbw==", + "cpu": [ + "arm64" + ], + "license": "MPL-2.0", + "optional": true, + "os": [ + "win32" + ], + "engines": { + "node": ">= 12.0.0" + }, + "funding": { + "type": "opencollective", + "url": "https://opencollective.com/parcel" + } + }, + "node_modules/lightningcss-win32-x64-msvc": { + "version": "1.29.2", + "resolved": "https://registry.npmjs.org/lightningcss-win32-x64-msvc/-/lightningcss-win32-x64-msvc-1.29.2.tgz", + "integrity": "sha512-EdIUW3B2vLuHmv7urfzMI/h2fmlnOQBk1xlsDxkN1tCWKjNFjfLhGxYk8C8mzpSfr+A6jFFIi8fU6LbQGsRWjA==", + "cpu": [ + "x64" + ], + "license": "MPL-2.0", + "optional": true, + "os": [ + "win32" + ], + "engines": { + "node": ">= 12.0.0" + }, + "funding": { + "type": "opencollective", + "url": "https://opencollective.com/parcel" + } }, "node_modules/locate-path": { "version": "6.0.0", @@ -3834,6 +4014,7 @@ "version": "1.4.1", "resolved": "https://registry.npmjs.org/merge2/-/merge2-1.4.1.tgz", "integrity": "sha512-8q7VEgMJW4J8tcfVPy8g09NcQwZdbwFEqhe/WZkoIzjn/3TGDwtOCYtXGxA3O8tPzpczCCDgv+P2P5y00ZJOOg==", + "dev": true, "license": "MIT", "engines": { "node": ">= 8" @@ -4447,32 +4628,12 @@ "node": "*" } }, - "node_modules/minipass": { - "version": "7.1.2", - "resolved": "https://registry.npmjs.org/minipass/-/minipass-7.1.2.tgz", - "integrity": "sha512-qOOzS1cBTWYF4BH8fVePDBOO9iptMnGUEZwNc/cMWnTV2nVLZ7VoNWEPHkYczZA0pdoA7dl6e7FL659nX9S2aw==", - "license": "ISC", - "engines": { - "node": ">=16 || 14 >=14.17" - } - }, "node_modules/ms": { "version": "2.1.3", "resolved": "https://registry.npmjs.org/ms/-/ms-2.1.3.tgz", "integrity": "sha512-6FlzubTLZG3J2a/NVCAleEhjzq5oxgHyaCU9yYXvcLsvoVaHJq/s5xXI6/XXP6tz7R9xAOtHnSO/tXtF3WRTlA==", "license": "MIT" }, - "node_modules/mz": { - "version": "2.7.0", - "resolved": "https://registry.npmjs.org/mz/-/mz-2.7.0.tgz", - "integrity": "sha512-z81GNO7nnYMEhrGh9LeymoE4+Yr0Wn5McHIZMK5cfQCl+NDX08sCZgUc9/6MHni9IWuFLm1Z3HTCXu2z9fN62Q==", - "license": "MIT", - "dependencies": { - "any-promise": "^1.0.0", - "object-assign": "^4.0.1", - "thenify-all": "^1.0.0" - } - }, "node_modules/nanoid": { "version": "3.3.8", "resolved": "https://registry.npmjs.org/nanoid/-/nanoid-3.3.8.tgz", @@ -4504,15 +4665,6 @@ "integrity": "sha512-xxOWJsBKtzAq7DY0J+DTzuz58K8e7sJbdgwkbMWQe8UYB6ekmsQ45q0M/tJDsGaZmbC+l7n57UV8Hl5tHxO9uw==", "license": "MIT" }, - "node_modules/normalize-path": { - "version": "3.0.0", - "resolved": "https://registry.npmjs.org/normalize-path/-/normalize-path-3.0.0.tgz", - "integrity": "sha512-6eZs5Ls3WtCisHWp9S2GUy8dqkpGi4BVSz3GaqiE6ezub0512ESztXUwUB6C6IKbQkY2Pnb/mD4WYojCRwcwLA==", - "license": "MIT", - "engines": { - "node": ">=0.10.0" - } - }, "node_modules/normalize-range": { "version": "0.1.2", "resolved": "https://registry.npmjs.org/normalize-range/-/normalize-range-0.1.2.tgz", @@ -4522,24 +4674,6 @@ "node": ">=0.10.0" } }, - "node_modules/object-assign": { - "version": "4.1.1", - "resolved": "https://registry.npmjs.org/object-assign/-/object-assign-4.1.1.tgz", - "integrity": "sha512-rJgTQnkUnH1sFw8yT6VSU3zD3sWmu6sZhIseY8VX+GRu3P6F7Fu+JNDoXfklElbLJSnc3FUQHVe4cU5hj+BcUg==", - "license": "MIT", - "engines": { - "node": ">=0.10.0" - } - }, - "node_modules/object-hash": { - "version": "3.0.0", - "resolved": "https://registry.npmjs.org/object-hash/-/object-hash-3.0.0.tgz", - "integrity": "sha512-RSn9F68PjH9HqtltsSnqYC1XXoWe9Bju5+213R98cNGttag9q9yAOTzdbsqvIa7aNm5WffBZFpWYr2aWrklWAw==", - "license": "MIT", - "engines": { - "node": ">= 6" - } - }, "node_modules/optionator": { "version": "0.9.4", "resolved": "https://registry.npmjs.org/optionator/-/optionator-0.9.4.tgz", @@ -4590,12 +4724,6 @@ "url": "https://github.com/sponsors/sindresorhus" } }, - "node_modules/package-json-from-dist": { - "version": "1.0.1", - "resolved": "https://registry.npmjs.org/package-json-from-dist/-/package-json-from-dist-1.0.1.tgz", - "integrity": "sha512-UEZIS3/by4OC8vL3P2dTXRETpebLI2NiI5vIrjaD/5UtrkFX/tNbwjTSRAGC/+7CAo2pIcBaRgWmcBBHcsaCIw==", - "license": "BlueOak-1.0.0" - }, "node_modules/parent-module": { "version": "1.0.1", "resolved": "https://registry.npmjs.org/parent-module/-/parent-module-1.0.1.tgz", @@ -4660,39 +4788,12 @@ "version": "3.1.1", "resolved": "https://registry.npmjs.org/path-key/-/path-key-3.1.1.tgz", "integrity": "sha512-ojmeN0qd+y0jszEtoY48r0Peq5dwMEkIlCOu6Q5f41lfkswXuKtYrhgoTpLnyIcHm24Uhqx+5Tqm2InSwLhE6Q==", + "dev": true, "license": "MIT", "engines": { "node": ">=8" } }, - "node_modules/path-parse": { - "version": "1.0.7", - "resolved": "https://registry.npmjs.org/path-parse/-/path-parse-1.0.7.tgz", - "integrity": "sha512-LDJzPVEEEPR+y48z93A0Ed0yXb8pAByGWo/k5YYdYgpY2/2EsOsksJrq7lOHxryrVOn1ejG6oAp8ahvOIQD8sw==", - "license": "MIT" - }, - "node_modules/path-scurry": { - "version": "1.11.1", - "resolved": "https://registry.npmjs.org/path-scurry/-/path-scurry-1.11.1.tgz", - "integrity": "sha512-Xa4Nw17FS9ApQFJ9umLiJS4orGjm7ZzwUrwamcGQuHSzDyth9boKDaycYdDcZDuqYATXw4HFXgaqWTctW/v1HA==", - "license": "BlueOak-1.0.0", - "dependencies": { - "lru-cache": "^10.2.0", - "minipass": "^5.0.0 || ^6.0.2 || ^7.0.0" - }, - "engines": { - "node": ">=16 || 14 >=14.18" - }, - "funding": { - "url": "https://github.com/sponsors/isaacs" - } - }, - "node_modules/path-scurry/node_modules/lru-cache": { - "version": "10.4.3", - "resolved": "https://registry.npmjs.org/lru-cache/-/lru-cache-10.4.3.tgz", - "integrity": "sha512-JNAzZcXrCt42VGLuYz0zfAzDfAvJWW6AfYlDBQyDV5DClI2m5sAmK+OIO7s59XfsRsWHp02jAJrRadPRGTt6SQ==", - "license": "ISC" - }, "node_modules/picocolors": { "version": "1.1.1", "resolved": "https://registry.npmjs.org/picocolors/-/picocolors-1.1.1.tgz", @@ -4711,24 +4812,6 @@ "url": "https://github.com/sponsors/jonschlinkert" } }, - "node_modules/pify": { - "version": "2.3.0", - "resolved": "https://registry.npmjs.org/pify/-/pify-2.3.0.tgz", - "integrity": "sha512-udgsAY+fTnvv7kI7aaxbqwWNb0AHiB0qBO89PZKPkoTmGOgdbrHDKD+0B2X4uTfJ/FT1R09r9gTsjUjNJotuog==", - "license": "MIT", - "engines": { - "node": ">=0.10.0" - } - }, - "node_modules/pirates": { - "version": "4.0.6", - "resolved": "https://registry.npmjs.org/pirates/-/pirates-4.0.6.tgz", - "integrity": "sha512-saLsH7WeYYPiD25LDuLRRY/i+6HaPYr6G1OUlN39otzkSTxKnubR9RTxS3/Kk50s1g2JTgFwWQDQyplC5/SHZg==", - "license": "MIT", - "engines": { - "node": ">= 6" - } - }, "node_modules/postcss": { "version": "8.5.1", "resolved": "https://registry.npmjs.org/postcss/-/postcss-8.5.1.tgz", @@ -4757,115 +4840,6 @@ "node": "^10 || ^12 || >=14" } }, - "node_modules/postcss-import": { - "version": "15.1.0", - "resolved": "https://registry.npmjs.org/postcss-import/-/postcss-import-15.1.0.tgz", - "integrity": "sha512-hpr+J05B2FVYUAXHeK1YyI267J/dDDhMU6B6civm8hSY1jYJnBXxzKDKDswzJmtLHryrjhnDjqqp/49t8FALew==", - "license": "MIT", - "dependencies": { - "postcss-value-parser": "^4.0.0", - "read-cache": "^1.0.0", - "resolve": "^1.1.7" - }, - "engines": { - "node": ">=14.0.0" - }, - "peerDependencies": { - "postcss": "^8.0.0" - } - }, - "node_modules/postcss-js": { - "version": "4.0.1", - "resolved": "https://registry.npmjs.org/postcss-js/-/postcss-js-4.0.1.tgz", - "integrity": "sha512-dDLF8pEO191hJMtlHFPRa8xsizHaM82MLfNkUHdUtVEV3tgTp5oj+8qbEqYM57SLfc74KSbw//4SeJma2LRVIw==", - "license": "MIT", - "dependencies": { - "camelcase-css": "^2.0.1" - }, - "engines": { - "node": "^12 || ^14 || >= 16" - }, - "funding": { - "type": "opencollective", - "url": "https://opencollective.com/postcss/" - }, - "peerDependencies": { - "postcss": "^8.4.21" - } - }, - "node_modules/postcss-load-config": { - "version": "4.0.2", - "resolved": "https://registry.npmjs.org/postcss-load-config/-/postcss-load-config-4.0.2.tgz", - "integrity": "sha512-bSVhyJGL00wMVoPUzAVAnbEoWyqRxkjv64tUl427SKnPrENtq6hJwUojroMz2VB+Q1edmi4IfrAPpami5VVgMQ==", - "funding": [ - { - "type": "opencollective", - "url": "https://opencollective.com/postcss/" - }, - { - "type": "github", - "url": "https://github.com/sponsors/ai" - } - ], - "license": "MIT", - "dependencies": { - "lilconfig": "^3.0.0", - "yaml": "^2.3.4" - }, - "engines": { - "node": ">= 14" - }, - "peerDependencies": { - "postcss": ">=8.0.9", - "ts-node": ">=9.0.0" - }, - "peerDependenciesMeta": { - "postcss": { - "optional": true - }, - "ts-node": { - "optional": true - } - } - }, - "node_modules/postcss-nested": { - "version": "6.2.0", - "resolved": "https://registry.npmjs.org/postcss-nested/-/postcss-nested-6.2.0.tgz", - "integrity": "sha512-HQbt28KulC5AJzG+cZtj9kvKB93CFCdLvog1WFLf1D+xmMvPGlBstkpTEZfK5+AN9hfJocyBFCNiqyS48bpgzQ==", - "funding": [ - { - "type": "opencollective", - "url": "https://opencollective.com/postcss/" - }, - { - "type": "github", - "url": "https://github.com/sponsors/ai" - } - ], - "license": "MIT", - "dependencies": { - "postcss-selector-parser": "^6.1.1" - }, - "engines": { - "node": ">=12.0" - }, - "peerDependencies": { - "postcss": "^8.2.14" - } - }, - "node_modules/postcss-selector-parser": { - "version": "6.1.2", - "resolved": "https://registry.npmjs.org/postcss-selector-parser/-/postcss-selector-parser-6.1.2.tgz", - "integrity": "sha512-Q8qQfPiZ+THO/3ZrOrO0cJJKfpYCagtMUkXbnEfmgUjwXg6z/WBeOyS9APBBPCTSiDV+s4SwQGu8yFsiMRIudg==", - "license": "MIT", - "dependencies": { - "cssesc": "^3.0.0", - "util-deprecate": "^1.0.2" - }, - "engines": { - "node": ">=4" - } - }, "node_modules/postcss-value-parser": { "version": "4.2.0", "resolved": "https://registry.npmjs.org/postcss-value-parser/-/postcss-value-parser-4.2.0.tgz", @@ -4922,6 +4896,7 @@ "version": "1.2.3", "resolved": "https://registry.npmjs.org/queue-microtask/-/queue-microtask-1.2.3.tgz", "integrity": "sha512-NuaNSa6flKT5JaSYQzJok04JzTL1CA6aGhv5rfLW3PgqA+M2ChpZQnAC8h8i4ZFkBS8X5RqkDBHA7r4hej3K9A==", + "dev": true, "funding": [ { "type": "github", @@ -5023,27 +4998,6 @@ } } }, - "node_modules/read-cache": { - "version": "1.0.0", - "resolved": "https://registry.npmjs.org/read-cache/-/read-cache-1.0.0.tgz", - "integrity": "sha512-Owdv/Ft7IjOgm/i0xvNDZ1LrRANRfew4b2prF3OWMQLxLfu3bS8FVhCsrSCMK4lR56Y9ya+AThoTpDCTxCmpRA==", - "license": "MIT", - "dependencies": { - "pify": "^2.3.0" - } - }, - "node_modules/readdirp": { - "version": "3.6.0", - "resolved": "https://registry.npmjs.org/readdirp/-/readdirp-3.6.0.tgz", - "integrity": "sha512-hOS089on8RduqdbhvQ5Z37A0ESjsqz6qnRcffsMU3495FuTdqSm+7bhJ29JvIOsBDEEnan5DPu9t3To9VRlMzA==", - "license": "MIT", - "dependencies": { - "picomatch": "^2.2.1" - }, - "engines": { - "node": ">=8.10.0" - } - }, "node_modules/rehype-highlight": { "version": "7.0.2", "resolved": "https://registry.npmjs.org/rehype-highlight/-/rehype-highlight-7.0.2.tgz", @@ -5177,26 +5131,6 @@ "url": "https://opencollective.com/unified" } }, - "node_modules/resolve": { - "version": "1.22.10", - "resolved": "https://registry.npmjs.org/resolve/-/resolve-1.22.10.tgz", - "integrity": "sha512-NPRy+/ncIMeDlTAsuqwKIiferiawhefFJtkNSW0qZJEqMEb+qBt/77B/jGeeek+F0uOeN05CDa6HXbbIgtVX4w==", - "license": "MIT", - "dependencies": { - "is-core-module": "^2.16.0", - "path-parse": "^1.0.7", - "supports-preserve-symlinks-flag": "^1.0.0" - }, - "bin": { - "resolve": "bin/resolve" - }, - "engines": { - "node": ">= 0.4" - }, - "funding": { - "url": "https://github.com/sponsors/ljharb" - } - }, "node_modules/resolve-from": { "version": "4.0.0", "resolved": "https://registry.npmjs.org/resolve-from/-/resolve-from-4.0.0.tgz", @@ -5211,6 +5145,7 @@ "version": "1.0.4", "resolved": "https://registry.npmjs.org/reusify/-/reusify-1.0.4.tgz", "integrity": "sha512-U9nH88a3fc/ekCF1l0/UP1IosiuIjyTh7hBvXVMHYgVcfGvt897Xguj2UOLDeI5BG2m7/uwyaLVT6fbtCwTyzw==", + "dev": true, "license": "MIT", "engines": { "iojs": ">=1.0.0", @@ -5259,6 +5194,7 @@ "version": "1.2.0", "resolved": "https://registry.npmjs.org/run-parallel/-/run-parallel-1.2.0.tgz", "integrity": "sha512-5l4VyZR86LZ/lDxZTR6jqL8AFE2S0IFLMP26AbjsLVADxHdhB/c0GUsH+y39UfCi3dzz8OlQuPmnaJOMoDHQBA==", + "dev": true, "funding": [ { "type": "github", @@ -5698,6 +5634,7 @@ "version": "2.0.0", "resolved": "https://registry.npmjs.org/shebang-command/-/shebang-command-2.0.0.tgz", "integrity": "sha512-kHxr2zZpYtdmrN1qDjrrX/Z1rR1kG8Dx+gkpK1G4eXmvXswmcE1hTWBWYUzlraYw1/yZp6YuDY77YtvbN0dmDA==", + "dev": true, "license": "MIT", "dependencies": { "shebang-regex": "^3.0.0" @@ -5710,23 +5647,12 @@ "version": "3.0.0", "resolved": "https://registry.npmjs.org/shebang-regex/-/shebang-regex-3.0.0.tgz", "integrity": "sha512-7++dFhtcx3353uBaq8DDR4NuxBetBzC7ZQOhmTQInHEd6bSrXdiEyzCvG07Z44UYdLShWUyXt5M/yhz8ekcb1A==", + "dev": true, "license": "MIT", "engines": { "node": ">=8" } }, - "node_modules/signal-exit": { - "version": "4.1.0", - "resolved": "https://registry.npmjs.org/signal-exit/-/signal-exit-4.1.0.tgz", - "integrity": "sha512-bzyZ1e88w9O1iNJbKnOlvYTrWPDl46O1bG0D3XInv+9tkPrxrN8jUUTiFlDkkmKWgn1M6CfIA13SuGqOa9Korw==", - "license": "ISC", - "engines": { - "node": ">=14" - }, - "funding": { - "url": "https://github.com/sponsors/isaacs" - } - }, "node_modules/source-map-js": { "version": "1.2.1", "resolved": "https://registry.npmjs.org/source-map-js/-/source-map-js-1.2.1.tgz", @@ -5746,65 +5672,6 @@ "url": "https://github.com/sponsors/wooorm" } }, - "node_modules/string-width": { - "version": "5.1.2", - "resolved": "https://registry.npmjs.org/string-width/-/string-width-5.1.2.tgz", - "integrity": "sha512-HnLOCR3vjcY8beoNLtcjZ5/nxn2afmME6lhrDrebokqMap+XbeW8n9TXpPDOqdGK5qcI3oT0GKTW6wC7EMiVqA==", - "license": "MIT", - "dependencies": { - "eastasianwidth": "^0.2.0", - "emoji-regex": "^9.2.2", - "strip-ansi": "^7.0.1" - }, - "engines": { - "node": ">=12" - }, - "funding": { - "url": "https://github.com/sponsors/sindresorhus" - } - }, - "node_modules/string-width-cjs": { - "name": "string-width", - "version": "4.2.3", - "resolved": "https://registry.npmjs.org/string-width/-/string-width-4.2.3.tgz", - "integrity": "sha512-wKyQRQpjJ0sIp62ErSZdGsjMJWsap5oRNihHhu6G7JVO/9jIB6UyevL+tXuOqrng8j/cxKTWyWUwvSTriiZz/g==", - "license": "MIT", - "dependencies": { - "emoji-regex": "^8.0.0", - "is-fullwidth-code-point": "^3.0.0", - "strip-ansi": "^6.0.1" - }, - "engines": { - "node": ">=8" - } - }, - "node_modules/string-width-cjs/node_modules/ansi-regex": { - "version": "5.0.1", - "resolved": "https://registry.npmjs.org/ansi-regex/-/ansi-regex-5.0.1.tgz", - "integrity": "sha512-quJQXlTSUGL2LH9SUXo8VwsY4soanhgo6LNSm84E1LBcE8s3O0wpdiRzyR9z/ZZJMlMWv37qOOb9pdJlMUEKFQ==", - "license": "MIT", - "engines": { - "node": ">=8" - } - }, - "node_modules/string-width-cjs/node_modules/emoji-regex": { - "version": "8.0.0", - "resolved": "https://registry.npmjs.org/emoji-regex/-/emoji-regex-8.0.0.tgz", - "integrity": "sha512-MSjYzcWNOA0ewAHpz0MxpYFvwg6yjy1NG3xteoqz644VCo/RPgnr1/GGt+ic3iJTzQ8Eu3TdM14SawnVUmGE6A==", - "license": "MIT" - }, - "node_modules/string-width-cjs/node_modules/strip-ansi": { - "version": "6.0.1", - "resolved": "https://registry.npmjs.org/strip-ansi/-/strip-ansi-6.0.1.tgz", - "integrity": "sha512-Y38VPSHcqkFrCpFnQ9vuSXmquuv5oXOKpGeT6aGrr3o3Gc9AlVa6JBfUSOCnbxGGZF+/0ooI7KrPuUSztUdU5A==", - "license": "MIT", - "dependencies": { - "ansi-regex": "^5.0.1" - }, - "engines": { - "node": ">=8" - } - }, "node_modules/stringify-entities": { "version": "4.0.4", "resolved": "https://registry.npmjs.org/stringify-entities/-/stringify-entities-4.0.4.tgz", @@ -5819,43 +5686,6 @@ "url": "https://github.com/sponsors/wooorm" } }, - "node_modules/strip-ansi": { - "version": "7.1.0", - "resolved": "https://registry.npmjs.org/strip-ansi/-/strip-ansi-7.1.0.tgz", - "integrity": "sha512-iq6eVVI64nQQTRYq2KtEg2d2uU7LElhTJwsH4YzIHZshxlgZms/wIc4VoDQTlG/IvVIrBKG06CrZnp0qv7hkcQ==", - "license": "MIT", - "dependencies": { - "ansi-regex": "^6.0.1" - }, - "engines": { - "node": ">=12" - }, - "funding": { - "url": "https://github.com/chalk/strip-ansi?sponsor=1" - } - }, - "node_modules/strip-ansi-cjs": { - "name": "strip-ansi", - "version": "6.0.1", - "resolved": "https://registry.npmjs.org/strip-ansi/-/strip-ansi-6.0.1.tgz", - "integrity": "sha512-Y38VPSHcqkFrCpFnQ9vuSXmquuv5oXOKpGeT6aGrr3o3Gc9AlVa6JBfUSOCnbxGGZF+/0ooI7KrPuUSztUdU5A==", - "license": "MIT", - "dependencies": { - "ansi-regex": "^5.0.1" - }, - "engines": { - "node": ">=8" - } - }, - "node_modules/strip-ansi-cjs/node_modules/ansi-regex": { - "version": "5.0.1", - "resolved": "https://registry.npmjs.org/ansi-regex/-/ansi-regex-5.0.1.tgz", - "integrity": "sha512-quJQXlTSUGL2LH9SUXo8VwsY4soanhgo6LNSm84E1LBcE8s3O0wpdiRzyR9z/ZZJMlMWv37qOOb9pdJlMUEKFQ==", - "license": "MIT", - "engines": { - "node": ">=8" - } - }, "node_modules/strip-json-comments": { "version": "3.1.1", "resolved": "https://registry.npmjs.org/strip-json-comments/-/strip-json-comments-3.1.1.tgz", @@ -5878,37 +5708,6 @@ "inline-style-parser": "0.2.4" } }, - "node_modules/sucrase": { - "version": "3.35.0", - "resolved": "https://registry.npmjs.org/sucrase/-/sucrase-3.35.0.tgz", - "integrity": "sha512-8EbVDiu9iN/nESwxeSxDKe0dunta1GOlHufmSSXxMD2z2/tMZpDMpvXQGsc+ajGo8y2uYUmixaSRUc/QPoQ0GA==", - "license": "MIT", - "dependencies": { - "@jridgewell/gen-mapping": "^0.3.2", - "commander": "^4.0.0", - "glob": "^10.3.10", - "lines-and-columns": "^1.1.6", - "mz": "^2.7.0", - "pirates": "^4.0.1", - "ts-interface-checker": "^0.1.9" - }, - "bin": { - "sucrase": "bin/sucrase", - "sucrase-node": "bin/sucrase-node" - }, - "engines": { - "node": ">=16 || 14 >=14.17" - } - }, - "node_modules/sucrase/node_modules/commander": { - "version": "4.1.1", - "resolved": "https://registry.npmjs.org/commander/-/commander-4.1.1.tgz", - "integrity": "sha512-NOKm8xhkzAjzFx8B2v5OAHT+u5pRQc2UCa2Vq9jYL/31o2wi9mxBA7LIFs3sV5VSC49z6pEhfbMULvShKj26WA==", - "license": "MIT", - "engines": { - "node": ">= 6" - } - }, "node_modules/supports-color": { "version": "7.2.0", "resolved": "https://registry.npmjs.org/supports-color/-/supports-color-7.2.0.tgz", @@ -5922,18 +5721,6 @@ "node": ">=8" } }, - "node_modules/supports-preserve-symlinks-flag": { - "version": "1.0.0", - "resolved": "https://registry.npmjs.org/supports-preserve-symlinks-flag/-/supports-preserve-symlinks-flag-1.0.0.tgz", - "integrity": "sha512-ot0WnXS9fgdkgIcePe6RHNk1WA8+muPa6cSjeR3V8K27q9BB1rTE3R1p7Hv0z1ZyAc8s6Vvv8DIyWf681MAt0w==", - "license": "MIT", - "engines": { - "node": ">= 0.4" - }, - "funding": { - "url": "https://github.com/sponsors/ljharb" - } - }, "node_modules/sync-child-process": { "version": "1.0.2", "resolved": "https://registry.npmjs.org/sync-child-process/-/sync-child-process-1.0.2.tgz", @@ -5958,40 +5745,18 @@ } }, "node_modules/tailwindcss": { - "version": "3.4.17", - "resolved": "https://registry.npmjs.org/tailwindcss/-/tailwindcss-3.4.17.tgz", - "integrity": "sha512-w33E2aCvSDP0tW9RZuNXadXlkHXqFzSkQew/aIa2i/Sj8fThxwovwlXHSPXTbAHwEIhBFXAedUhP2tueAKP8Og==", + "version": "4.1.1", + "resolved": "https://registry.npmjs.org/tailwindcss/-/tailwindcss-4.1.1.tgz", + "integrity": "sha512-QNbdmeS979Efzim2g/bEvfuh+fTcIdp1y7gA+sb6OYSW74rt7Cr7M78AKdf6HqWT3d5AiTb7SwTT3sLQxr4/qw==", + "license": "MIT" + }, + "node_modules/tapable": { + "version": "2.2.1", + "resolved": "https://registry.npmjs.org/tapable/-/tapable-2.2.1.tgz", + "integrity": "sha512-GNzQvQTOIP6RyTfE2Qxb8ZVlNmw0n88vp1szwWRimP02mnTsx3Wtn5qRdqY9w2XduFNUgvOwhNnQsjwCp+kqaQ==", "license": "MIT", - "dependencies": { - "@alloc/quick-lru": "^5.2.0", - "arg": "^5.0.2", - "chokidar": "^3.6.0", - "didyoumean": "^1.2.2", - "dlv": "^1.1.3", - "fast-glob": "^3.3.2", - "glob-parent": "^6.0.2", - "is-glob": "^4.0.3", - "jiti": "^1.21.6", - "lilconfig": "^3.1.3", - "micromatch": "^4.0.8", - "normalize-path": "^3.0.0", - "object-hash": "^3.0.0", - "picocolors": "^1.1.1", - "postcss": "^8.4.47", - "postcss-import": "^15.1.0", - "postcss-js": "^4.0.1", - "postcss-load-config": "^4.0.2", - "postcss-nested": "^6.2.0", - "postcss-selector-parser": "^6.1.2", - "resolve": "^1.22.8", - "sucrase": "^3.35.0" - }, - "bin": { - "tailwind": "lib/cli.js", - "tailwindcss": "lib/cli.js" - }, "engines": { - "node": ">=14.0.0" + "node": ">=6" } }, "node_modules/textlinestream": { @@ -6000,27 +5765,6 @@ "integrity": "sha512-iBHbi7BQxrFmwZUQJsT0SjNzlLLsXhvW/kg7EyOMVMBIrlnj/qYofwo1LVLZi+3GbUEo96Iu2eqToI2+lZoAEQ==", "license": "MIT" }, - "node_modules/thenify": { - "version": "3.3.1", - "resolved": "https://registry.npmjs.org/thenify/-/thenify-3.3.1.tgz", - "integrity": "sha512-RVZSIV5IG10Hk3enotrhvz0T9em6cyHBLkH/YAZuKqd8hRkKhSfCGIcP2KUY0EPxndzANBmNllzWPwak+bheSw==", - "license": "MIT", - "dependencies": { - "any-promise": "^1.0.0" - } - }, - "node_modules/thenify-all": { - "version": "1.6.0", - "resolved": "https://registry.npmjs.org/thenify-all/-/thenify-all-1.6.0.tgz", - "integrity": "sha512-RNxQH/qI8/t3thXJDwcstUO4zeqo64+Uy/+sNVRBx4Xn2OX+OZ9oP+iJnNFqplFra2ZUVeKCSa2oVWi3T4uVmA==", - "license": "MIT", - "dependencies": { - "thenify": ">= 3.1.0 < 4" - }, - "engines": { - "node": ">=0.8" - } - }, "node_modules/to-regex-range": { "version": "5.0.1", "resolved": "https://registry.npmjs.org/to-regex-range/-/to-regex-range-5.0.1.tgz", @@ -6066,12 +5810,6 @@ "typescript": ">=4.8.4" } }, - "node_modules/ts-interface-checker": { - "version": "0.1.13", - "resolved": "https://registry.npmjs.org/ts-interface-checker/-/ts-interface-checker-0.1.13.tgz", - "integrity": "sha512-Y/arvbn+rrz3JCKl9C4kVNfTfSm2/mEp5FSz5EsZSANGPSlQrpRI5M4PKF+mJnE52jOO90PnPSc3Ur3bTQw0gA==", - "license": "Apache-2.0" - }, "node_modules/tslib": { "version": "2.8.1", "resolved": "https://registry.npmjs.org/tslib/-/tslib-2.8.1.tgz", @@ -6297,12 +6035,6 @@ "punycode": "^2.1.0" } }, - "node_modules/util-deprecate": { - "version": "1.0.2", - "resolved": "https://registry.npmjs.org/util-deprecate/-/util-deprecate-1.0.2.tgz", - "integrity": "sha512-EPD5q1uXyFxJpCrLnCc1nHnq3gOa6DZBocAIiI2TaSCA7VCJ1UJDMagCzIkXNsUYfD1daK//LTEQ8xiIbrHtcw==", - "license": "MIT" - }, "node_modules/varint": { "version": "6.0.0", "resolved": "https://registry.npmjs.org/varint/-/varint-6.0.0.tgz", @@ -6453,6 +6185,7 @@ "version": "2.0.2", "resolved": "https://registry.npmjs.org/which/-/which-2.0.2.tgz", "integrity": "sha512-BLI3Tl1TW3Pvl70l3yq3Y64i+awpwXqsGBYWkkqMtnbXgrMD+yj7rhW0kuEDxzJaYXGjEW5ogapKNMEKNMjibA==", + "dev": true, "license": "ISC", "dependencies": { "isexe": "^2.0.0" @@ -6474,94 +6207,6 @@ "node": ">=0.10.0" } }, - "node_modules/wrap-ansi": { - "version": "8.1.0", - "resolved": "https://registry.npmjs.org/wrap-ansi/-/wrap-ansi-8.1.0.tgz", - "integrity": "sha512-si7QWI6zUMq56bESFvagtmzMdGOtoxfR+Sez11Mobfc7tm+VkUckk9bW2UeffTGVUbOksxmSw0AA2gs8g71NCQ==", - "license": "MIT", - "dependencies": { - "ansi-styles": "^6.1.0", - "string-width": "^5.0.1", - "strip-ansi": "^7.0.1" - }, - "engines": { - "node": ">=12" - }, - "funding": { - "url": "https://github.com/chalk/wrap-ansi?sponsor=1" - } - }, - "node_modules/wrap-ansi-cjs": { - "name": "wrap-ansi", - "version": "7.0.0", - "resolved": "https://registry.npmjs.org/wrap-ansi/-/wrap-ansi-7.0.0.tgz", - "integrity": "sha512-YVGIj2kamLSTxw6NsZjoBxfSwsn0ycdesmc4p+Q21c5zPuZ1pl+NfxVdxPtdHvmNVOQ6XSYG4AUtyt/Fi7D16Q==", - "license": "MIT", - "dependencies": { - "ansi-styles": "^4.0.0", - "string-width": "^4.1.0", - "strip-ansi": "^6.0.0" - }, - "engines": { - "node": ">=10" - }, - "funding": { - "url": "https://github.com/chalk/wrap-ansi?sponsor=1" - } - }, - "node_modules/wrap-ansi-cjs/node_modules/ansi-regex": { - "version": "5.0.1", - "resolved": "https://registry.npmjs.org/ansi-regex/-/ansi-regex-5.0.1.tgz", - "integrity": "sha512-quJQXlTSUGL2LH9SUXo8VwsY4soanhgo6LNSm84E1LBcE8s3O0wpdiRzyR9z/ZZJMlMWv37qOOb9pdJlMUEKFQ==", - "license": "MIT", - "engines": { - "node": ">=8" - } - }, - "node_modules/wrap-ansi-cjs/node_modules/emoji-regex": { - "version": "8.0.0", - "resolved": "https://registry.npmjs.org/emoji-regex/-/emoji-regex-8.0.0.tgz", - "integrity": "sha512-MSjYzcWNOA0ewAHpz0MxpYFvwg6yjy1NG3xteoqz644VCo/RPgnr1/GGt+ic3iJTzQ8Eu3TdM14SawnVUmGE6A==", - "license": "MIT" - }, - "node_modules/wrap-ansi-cjs/node_modules/string-width": { - "version": "4.2.3", - "resolved": "https://registry.npmjs.org/string-width/-/string-width-4.2.3.tgz", - "integrity": "sha512-wKyQRQpjJ0sIp62ErSZdGsjMJWsap5oRNihHhu6G7JVO/9jIB6UyevL+tXuOqrng8j/cxKTWyWUwvSTriiZz/g==", - "license": "MIT", - "dependencies": { - "emoji-regex": "^8.0.0", - "is-fullwidth-code-point": "^3.0.0", - "strip-ansi": "^6.0.1" - }, - "engines": { - "node": ">=8" - } - }, - "node_modules/wrap-ansi-cjs/node_modules/strip-ansi": { - "version": "6.0.1", - "resolved": "https://registry.npmjs.org/strip-ansi/-/strip-ansi-6.0.1.tgz", - "integrity": "sha512-Y38VPSHcqkFrCpFnQ9vuSXmquuv5oXOKpGeT6aGrr3o3Gc9AlVa6JBfUSOCnbxGGZF+/0ooI7KrPuUSztUdU5A==", - "license": "MIT", - "dependencies": { - "ansi-regex": "^5.0.1" - }, - "engines": { - "node": ">=8" - } - }, - "node_modules/wrap-ansi/node_modules/ansi-styles": { - "version": "6.2.1", - "resolved": "https://registry.npmjs.org/ansi-styles/-/ansi-styles-6.2.1.tgz", - "integrity": "sha512-bN798gFfQX+viw3R7yrGWRqnrN2oRkEkUjjl4JNn4E8GxxbjtG3FbrEIIY3l8/hrwUwIeCZvi4QuOTP4MErVug==", - "license": "MIT", - "engines": { - "node": ">=12" - }, - "funding": { - "url": "https://github.com/chalk/ansi-styles?sponsor=1" - } - }, "node_modules/yallist": { "version": "3.1.1", "resolved": "https://registry.npmjs.org/yallist/-/yallist-3.1.1.tgz", @@ -6574,6 +6219,8 @@ "resolved": "https://registry.npmjs.org/yaml/-/yaml-2.7.0.tgz", "integrity": "sha512-+hSoy/QHluxmC9kCIJyL/uyFmLmc+e5CFR5Wa+bpIhIj85LVb9ZH2nVnqrHoSvKogwODv0ClqZkmiSSaIH5LTA==", "license": "ISC", + "optional": true, + "peer": true, "bin": { "yaml": "bin.mjs" }, diff --git a/examples/server/webui/package.json b/examples/server/webui/package.json index 3be2b14de..6ac06b1a4 100644 --- a/examples/server/webui/package.json +++ b/examples/server/webui/package.json @@ -13,9 +13,12 @@ "dependencies": { "@heroicons/react": "^2.2.0", "@sec-ant/readable-stream": "^0.6.0", + "@tailwindcss/postcss": "^4.1.1", + "@tailwindcss/vite": "^4.1.1", "@vscode/markdown-it-katex": "^1.1.1", "autoprefixer": "^10.4.20", - "daisyui": "^4.12.14", + "daisyui": "^5.0.12", + "dexie": "^4.0.11", "highlight.js": "^11.10.0", "katex": "^0.16.15", "postcss": "^8.4.49", @@ -28,7 +31,7 @@ "remark-breaks": "^4.0.0", "remark-gfm": "^4.0.0", "remark-math": "^6.0.0", - "tailwindcss": "^3.4.15", + "tailwindcss": "^4.1.1", "textlinestream": "^1.1.1", "vite-plugin-singlefile": "^2.0.3" }, diff --git a/examples/server/webui/postcss.config.js b/examples/server/webui/postcss.config.js index 2e7af2b7f..fb05b5692 100644 --- a/examples/server/webui/postcss.config.js +++ b/examples/server/webui/postcss.config.js @@ -1,6 +1,5 @@ export default { plugins: { - tailwindcss: {}, - autoprefixer: {}, + "@tailwindcss/postcss": {}, }, } diff --git a/examples/server/webui/src/App.tsx b/examples/server/webui/src/App.tsx index 2ce734682..cc4659e15 100644 --- a/examples/server/webui/src/App.tsx +++ b/examples/server/webui/src/App.tsx @@ -28,7 +28,7 @@ function AppLayout() { <>
diff --git a/examples/server/webui/src/Config.ts b/examples/server/webui/src/Config.ts index 779ed9bf7..dd1cc0e10 100644 --- a/examples/server/webui/src/Config.ts +++ b/examples/server/webui/src/Config.ts @@ -1,4 +1,4 @@ -import daisyuiThemes from 'daisyui/src/theming/themes'; +import daisyuiThemes from 'daisyui/theme/object'; import { isNumeric } from './utils/misc'; export const isDev = import.meta.env.MODE === 'development'; diff --git a/examples/server/webui/src/components/ChatMessage.tsx b/examples/server/webui/src/components/ChatMessage.tsx index ec72196ba..40ea74711 100644 --- a/examples/server/webui/src/components/ChatMessage.tsx +++ b/examples/server/webui/src/components/ChatMessage.tsx @@ -3,6 +3,7 @@ import { useAppContext } from '../utils/app.context'; import { Message, PendingMessage } from '../utils/types'; import { classNames } from '../utils/misc'; import MarkdownDisplay, { CopyButton } from './MarkdownDisplay'; +import { ChevronLeftIcon, ChevronRightIcon } from '@heroicons/react/24/outline'; interface SplitMessage { content: PendingMessage['content']; @@ -12,17 +13,24 @@ interface SplitMessage { export default function ChatMessage({ msg, + siblingLeafNodeIds, + siblingCurrIdx, id, - scrollToBottom, + onRegenerateMessage, + onEditMessage, + onChangeSibling, isPending, }: { msg: Message | PendingMessage; + siblingLeafNodeIds: Message['id'][]; + siblingCurrIdx: number; id?: string; - scrollToBottom: (requiresNearBottom: boolean) => void; + onRegenerateMessage(msg: Message): void; + onEditMessage(msg: Message, content: string): void; + onChangeSibling(sibling: Message['id']): void; isPending?: boolean; }) { - const { viewingConversation, replaceMessageAndGenerate, config } = - useAppContext(); + const { viewingChat, config } = useAppContext(); const [editingContent, setEditingContent] = useState(null); const timings = useMemo( () => @@ -37,6 +45,8 @@ export default function ChatMessage({ : null, [msg.timings] ); + const nextSibling = siblingLeafNodeIds[siblingCurrIdx + 1]; + const prevSibling = siblingLeafNodeIds[siblingCurrIdx - 1]; // for reasoning model, we split the message into content and thought // TODO: implement this as remark/rehype plugin in the future @@ -64,13 +74,7 @@ export default function ChatMessage({ return { content: actualContent, thought, isThinking }; }, [msg]); - if (!viewingConversation) return null; - - const regenerate = async () => { - replaceMessageAndGenerate(viewingConversation.id, msg.id, undefined, () => - scrollToBottom(true) - ); - }; + if (!viewingChat) return null; return (
@@ -105,13 +109,12 @@ export default function ChatMessage({ @@ -156,6 +159,35 @@ export default function ChatMessage({
)} + + {msg.extra && msg.extra.length > 0 && ( +
+ + Extra content + +
+ {msg.extra.map( + (extra, i) => + extra.type === 'textFile' ? ( +
+ {extra.name} + 
{extra.content}
+
+ ) : extra.type === 'context' ? ( +
+ 
{extra.content}
+
+ ) : null // TODO: support other extra types + )} +
+
+ )} + + {siblingLeafNodeIds && siblingLeafNodeIds.length > 1 && ( +
+ + + {siblingCurrIdx + 1} / {siblingLeafNodeIds.length} + + +
+ )} {/* user message */} {msg.role === 'user' && ( )} - )} + )} diff --git a/examples/server/webui/src/components/ChatScreen.tsx b/examples/server/webui/src/components/ChatScreen.tsx index dbc683ed1..29ab5ea64 100644 --- a/examples/server/webui/src/components/ChatScreen.tsx +++ b/examples/server/webui/src/components/ChatScreen.tsx @@ -1,28 +1,79 @@ -import { useEffect, useState } from 'react'; -import { useAppContext } from '../utils/app.context'; -import StorageUtils from '../utils/storage'; -import { useNavigate } from 'react-router'; +import { useEffect, useMemo, useState } from 'react'; +import { CallbackGeneratedChunk, useAppContext } from '../utils/app.context'; import ChatMessage from './ChatMessage'; -import { CanvasType, PendingMessage } from '../utils/types'; -import { classNames } from '../utils/misc'; +import { CanvasType, Message, PendingMessage } from '../utils/types'; +import { classNames, cleanCurrentUrl, throttle } from '../utils/misc'; import CanvasPyInterpreter from './CanvasPyInterpreter'; +import StorageUtils from '../utils/storage'; +import { useVSCodeContext } from '../utils/llama-vscode'; +import { useChatTextarea, ChatTextareaApi } from './useChatTextarea.ts'; -export default function ChatScreen() { - const { - viewingConversation, - sendMessage, - isGenerating, - stopGenerating, - pendingMessages, - canvasData, - } = useAppContext(); - const [inputMsg, setInputMsg] = useState(''); - const navigate = useNavigate(); +/** + * A message display is a message node with additional information for rendering. + * For example, siblings of the message node are stored as their last node (aka leaf node). + */ +export interface MessageDisplay { + msg: Message | PendingMessage; + siblingLeafNodeIds: Message['id'][]; + siblingCurrIdx: number; + isPending?: boolean; +} - const currConvId = viewingConversation?.id ?? ''; - const pendingMsg: PendingMessage | undefined = pendingMessages[currConvId]; +/** + * If the current URL contains "?m=...", prefill the message input with the value. + * If the current URL contains "?q=...", prefill and SEND the message. + */ +const prefilledMsg = { + content() { + const url = new URL(window.location.href); + return url.searchParams.get('m') ?? url.searchParams.get('q') ?? ''; + }, + shouldSend() { + const url = new URL(window.location.href); + return url.searchParams.has('q'); + }, + clear() { + cleanCurrentUrl(['m', 'q']); + }, +}; - const scrollToBottom = (requiresNearBottom: boolean) => { +function getListMessageDisplay( + msgs: Readonly, + leafNodeId: Message['id'] +): MessageDisplay[] { + const currNodes = StorageUtils.filterByLeafNodeId(msgs, leafNodeId, true); + const res: MessageDisplay[] = []; + const nodeMap = new Map(); + for (const msg of msgs) { + nodeMap.set(msg.id, msg); + } + // find leaf node from a message node + const findLeafNode = (msgId: Message['id']): Message['id'] => { + let currNode: Message | undefined = nodeMap.get(msgId); + while (currNode) { + if (currNode.children.length === 0) break; + currNode = nodeMap.get(currNode.children.at(-1) ?? -1); + } + return currNode?.id ?? -1; + }; + // traverse the current nodes + for (const msg of currNodes) { + const parentNode = nodeMap.get(msg.parent ?? -1); + if (!parentNode) continue; + const siblings = parentNode.children; + if (msg.type !== 'root') { + res.push({ + msg, + siblingLeafNodeIds: siblings.map(findLeafNode), + siblingCurrIdx: siblings.indexOf(msg.id), + }); + } + } + return res; +} + +const scrollToBottom = throttle( + (requiresNearBottom: boolean, delay: number = 80) => { const mainScrollElem = document.getElementById('main-scroll'); if (!mainScrollElem) return; const spaceToBottom = @@ -32,36 +83,138 @@ export default function ChatScreen() { if (!requiresNearBottom || spaceToBottom < 50) { setTimeout( () => mainScrollElem.scrollTo({ top: mainScrollElem.scrollHeight }), - 1 + delay ); } + }, + 80 +); + +export default function ChatScreen() { + const { + viewingChat, + sendMessage, + isGenerating, + stopGenerating, + pendingMessages, + canvasData, + replaceMessageAndGenerate, + } = useAppContext(); + + const textarea: ChatTextareaApi = useChatTextarea(prefilledMsg.content()); + + const { extraContext, clearExtraContext } = useVSCodeContext(textarea); + // TODO: improve this when we have "upload file" feature + const currExtra: Message['extra'] = extraContext ? [extraContext] : undefined; + + // keep track of leaf node for rendering + const [currNodeId, setCurrNodeId] = useState(-1); + const messages: MessageDisplay[] = useMemo(() => { + if (!viewingChat) return []; + else return getListMessageDisplay(viewingChat.messages, currNodeId); + }, [currNodeId, viewingChat]); + + const currConvId = viewingChat?.conv.id ?? null; + const pendingMsg: PendingMessage | undefined = + pendingMessages[currConvId ?? '']; + + useEffect(() => { + // reset to latest node when conversation changes + setCurrNodeId(-1); + // scroll to bottom when conversation changes + scrollToBottom(false, 1); + }, [currConvId]); + + const onChunk: CallbackGeneratedChunk = (currLeafNodeId?: Message['id']) => { + if (currLeafNodeId) { + setCurrNodeId(currLeafNodeId); + } + scrollToBottom(true); }; - // scroll to bottom when conversation changes - useEffect(() => { - scrollToBottom(false); - }, [viewingConversation?.id]); - const sendNewMessage = async () => { - if (inputMsg.trim().length === 0 || isGenerating(currConvId)) return; - const convId = viewingConversation?.id ?? StorageUtils.getNewConvId(); - const lastInpMsg = inputMsg; - setInputMsg(''); - if (!viewingConversation) { - // if user is creating a new conversation, redirect to the new conversation - navigate(`/chat/${convId}`); - } + const lastInpMsg = textarea.value(); + if (lastInpMsg.trim().length === 0 || isGenerating(currConvId ?? '')) + return; + textarea.setValue(''); scrollToBottom(false); - // auto scroll as message is being generated - const onChunk = () => scrollToBottom(true); - if (!(await sendMessage(convId, inputMsg, onChunk))) { + setCurrNodeId(-1); + // get the last message node + const lastMsgNodeId = messages.at(-1)?.msg.id ?? null; + if ( + !(await sendMessage( + currConvId, + lastMsgNodeId, + lastInpMsg, + currExtra, + onChunk + )) + ) { // restore the input message if failed - setInputMsg(lastInpMsg); + textarea.setValue(lastInpMsg); } + // OK + clearExtraContext(); + }; + + const handleEditMessage = async (msg: Message, content: string) => { + if (!viewingChat) return; + setCurrNodeId(msg.id); + scrollToBottom(false); + await replaceMessageAndGenerate( + viewingChat.conv.id, + msg.parent, + content, + msg.extra, + onChunk + ); + setCurrNodeId(-1); + scrollToBottom(false); + }; + + const handleRegenerateMessage = async (msg: Message) => { + if (!viewingChat) return; + setCurrNodeId(msg.parent); + scrollToBottom(false); + await replaceMessageAndGenerate( + viewingChat.conv.id, + msg.parent, + null, + msg.extra, + onChunk + ); + setCurrNodeId(-1); + scrollToBottom(false); }; const hasCanvas = !!canvasData; + useEffect(() => { + if (prefilledMsg.shouldSend()) { + // send the prefilled message if needed + sendNewMessage(); + } else { + // otherwise, focus on the input + textarea.focus(); + } + prefilledMsg.clear(); + // no need to keep track of sendNewMessage + // eslint-disable-next-line react-hooks/exhaustive-deps + }, [textarea.ref]); + + // due to some timing issues of StorageUtils.appendMsg(), we need to make sure the pendingMsg is not duplicated upon rendering (i.e. appears once in the saved conversation and once in the pendingMsg) + const pendingMsgDisplay: MessageDisplay[] = + pendingMsg && messages.at(-1)?.msg.id !== pendingMsg.id + ? [ + { + msg: pendingMsg, + siblingLeafNodeIds: [], + siblingCurrIdx: 0, + isPending: true, + }, + ] + : []; + return (
{/* placeholder to shift the message to the bottom */} - {viewingConversation ? '' : 'Send a message to start'} + {viewingChat ? '' : 'Send a message to start'}
- {viewingConversation?.messages.map((msg) => ( + {[...messages, ...pendingMsgDisplay].map((msg) => ( ))} - - {pendingMsg && ( - - )}
{/* chat input */} -
+
- {isGenerating(currConvId) ? ( + + {isGenerating(currConvId ?? '') ? ( ) : ( - )} diff --git a/examples/server/webui/src/components/Header.tsx b/examples/server/webui/src/components/Header.tsx index 505350313..4c6b291e6 100644 --- a/examples/server/webui/src/components/Header.tsx +++ b/examples/server/webui/src/components/Header.tsx @@ -2,7 +2,7 @@ import { useEffect, useState } from 'react'; import StorageUtils from '../utils/storage'; import { useAppContext } from '../utils/app.context'; import { classNames } from '../utils/misc'; -import daisyuiThemes from 'daisyui/src/theming/themes'; +import daisyuiThemes from 'daisyui/theme/object'; import { THEMES } from '../Config'; import { useNavigate } from 'react-router'; @@ -20,17 +20,16 @@ export default function Header() { document.body.setAttribute('data-theme', selectedTheme); document.body.setAttribute( 'data-color-scheme', - // @ts-expect-error daisyuiThemes complains about index type, but it should work daisyuiThemes[selectedTheme]?.['color-scheme'] ?? 'auto' ); }, [selectedTheme]); - const { isGenerating, viewingConversation } = useAppContext(); - const isCurrConvGenerating = isGenerating(viewingConversation?.id ?? ''); + const { isGenerating, viewingChat } = useAppContext(); + const isCurrConvGenerating = isGenerating(viewingChat?.conv.id ?? ''); const removeConversation = () => { - if (isCurrConvGenerating || !viewingConversation) return; - const convId = viewingConversation.id; + if (isCurrConvGenerating || !viewingChat) return; + const convId = viewingChat?.conv.id; if (window.confirm('Are you sure to delete this conversation?')) { StorageUtils.remove(convId); navigate('/'); @@ -38,9 +37,9 @@ export default function Header() { }; const downloadConversation = () => { - if (isCurrConvGenerating || !viewingConversation) return; - const convId = viewingConversation.id; - const conversationJson = JSON.stringify(viewingConversation, null, 2); + if (isCurrConvGenerating || !viewingChat) return; + const convId = viewingChat?.conv.id; + const conversationJson = JSON.stringify(viewingChat, null, 2); const blob = new Blob([conversationJson], { type: 'application/json' }); const url = URL.createObjectURL(blob); const a = document.createElement('a'); @@ -75,38 +74,41 @@ export default function Header() { {/* action buttons (top right) */}
-
- {/* "..." button */} - - {/* dropdown menu */} - -
+ + + + + {/* dropdown menu */} + +
+ )} +
))}
- Conversations are saved to browser's localStorage + Conversations are saved to browser's IndexedDB
diff --git a/examples/server/webui/src/components/useChatTextarea.ts b/examples/server/webui/src/components/useChatTextarea.ts new file mode 100644 index 000000000..42b128194 --- /dev/null +++ b/examples/server/webui/src/components/useChatTextarea.ts @@ -0,0 +1,96 @@ +import { useEffect, useRef, useState, useCallback } from 'react'; + +// Media Query for detecting "large" screens (matching Tailwind's lg: breakpoint) +const LARGE_SCREEN_MQ = '(min-width: 1024px)'; + +// Calculates and sets the textarea height based on its scrollHeight +const adjustTextareaHeight = (textarea: HTMLTextAreaElement | null) => { + if (!textarea) return; + + // Only perform auto-sizing on large screens + if (!window.matchMedia(LARGE_SCREEN_MQ).matches) { + // On small screens, reset inline height and max-height styles. + // This allows CSS (e.g., `rows` attribute or classes) to control the height, + // and enables manual resizing if `resize-vertical` is set. + textarea.style.height = ''; // Use 'auto' or '' to reset + textarea.style.maxHeight = ''; + return; // Do not adjust height programmatically on small screens + } + + const computedStyle = window.getComputedStyle(textarea); + // Get the max-height specified by CSS (e.g., from `lg:max-h-48`) + const currentMaxHeight = computedStyle.maxHeight; + + // Temporarily remove max-height to allow scrollHeight to be calculated correctly + textarea.style.maxHeight = 'none'; + // Reset height to 'auto' to measure the actual scrollHeight needed + textarea.style.height = 'auto'; + // Set the height to the calculated scrollHeight + textarea.style.height = `${textarea.scrollHeight}px`; + // Re-apply the original max-height from CSS to enforce the limit + textarea.style.maxHeight = currentMaxHeight; +}; + +// Interface describing the API returned by the hook +export interface ChatTextareaApi { + value: () => string; + setValue: (value: string) => void; + focus: () => void; + ref: React.RefObject; + onInput: (event: React.FormEvent) => void; // Input handler +} + +// This is a workaround to prevent the textarea from re-rendering when the inner content changes +// See https://github.com/ggml-org/llama.cpp/pull/12299 +// combined now with auto-sizing logic. +export function useChatTextarea(initValue: string): ChatTextareaApi { + const [savedInitValue, setSavedInitValue] = useState(initValue); + const textareaRef = useRef(null); + + // Effect to set initial value and height on mount or when initValue changes + useEffect(() => { + const textarea = textareaRef.current; + if (textarea) { + if (typeof savedInitValue === 'string' && savedInitValue.length > 0) { + textarea.value = savedInitValue; + // Call adjustTextareaHeight - it will check screen size internally + setTimeout(() => adjustTextareaHeight(textarea), 0); + setSavedInitValue(''); // Reset after applying + } else { + // Adjust height even if there's no initial value (for initial render) + setTimeout(() => adjustTextareaHeight(textarea), 0); + } + } + }, [textareaRef, savedInitValue]); // Depend on ref and savedInitValue + + const handleInput = useCallback( + (event: React.FormEvent) => { + // Call adjustTextareaHeight on every input - it will decide whether to act + adjustTextareaHeight(event.currentTarget); + }, + [] + ); + + return { + // Method to get the current value directly from the textarea + value: () => { + return textareaRef.current?.value ?? ''; + }, + // Method to programmatically set the value and trigger height adjustment + setValue: (value: string) => { + const textarea = textareaRef.current; + if (textarea) { + textarea.value = value; + // Call adjustTextareaHeight - it will check screen size internally + setTimeout(() => adjustTextareaHeight(textarea), 0); + } + }, + focus: () => { + if (textareaRef.current) { + textareaRef.current.focus(); + } + }, + ref: textareaRef, + onInput: handleInput, + }; +} diff --git a/examples/server/webui/src/index.scss b/examples/server/webui/src/index.scss index 314b68232..a18f09454 100644 --- a/examples/server/webui/src/index.scss +++ b/examples/server/webui/src/index.scss @@ -1,8 +1,13 @@ @use 'sass:meta'; +@use 'tailwindcss'; -@tailwind base; -@tailwind components; -@tailwind utilities; +@plugin 'daisyui' { + themes: all; +} + +html { + scrollbar-gutter: auto; +} .markdown { h1, diff --git a/examples/server/webui/src/utils/app.context.tsx b/examples/server/webui/src/utils/app.context.tsx index af6bd885f..54bb65b6e 100644 --- a/examples/server/webui/src/utils/app.context.tsx +++ b/examples/server/webui/src/utils/app.context.tsx @@ -5,6 +5,7 @@ import { Conversation, Message, PendingMessage, + ViewingChat, } from './types'; import StorageUtils from './storage'; import { @@ -13,24 +14,27 @@ import { getSSEStreamAsync, } from './misc'; import { BASE_URL, CONFIG_DEFAULT, isDev } from '../Config'; -import { matchPath, useLocation } from 'react-router'; +import { matchPath, useLocation, useNavigate } from 'react-router'; interface AppContextValue { // conversations and messages - viewingConversation: Conversation | null; + viewingChat: ViewingChat | null; pendingMessages: Record; isGenerating: (convId: string) => boolean; sendMessage: ( - convId: string, + convId: string | null, + leafNodeId: Message['id'] | null, content: string, - onChunk?: CallbackGeneratedChunk + extra: Message['extra'], + onChunk: CallbackGeneratedChunk ) => Promise; stopGenerating: (convId: string) => void; replaceMessageAndGenerate: ( convId: string, - origMsgId: Message['id'], - content?: string, - onChunk?: CallbackGeneratedChunk + parentNodeId: Message['id'], // the parent node of the message to be replaced + content: string | null, + extra: Message['extra'], + onChunk: CallbackGeneratedChunk ) => Promise; // canvas @@ -44,23 +48,33 @@ interface AppContextValue { setShowSettings: (show: boolean) => void; } -// for now, this callback is only used for scrolling to the bottom of the chat -type CallbackGeneratedChunk = () => void; +// this callback is used for scrolling to the bottom of the chat and switching to the last node +export type CallbackGeneratedChunk = (currLeafNodeId?: Message['id']) => void; // eslint-disable-next-line @typescript-eslint/no-explicit-any const AppContext = createContext({} as any); +const getViewingChat = async (convId: string): Promise => { + const conv = await StorageUtils.getOneConversation(convId); + if (!conv) return null; + return { + conv: conv, + // all messages from all branches, not filtered by last node + messages: await StorageUtils.getMessages(convId), + }; +}; + export const AppContextProvider = ({ children, }: { children: React.ReactElement; }) => { const { pathname } = useLocation(); + const navigate = useNavigate(); const params = matchPath('/chat/:convId', pathname); const convId = params?.params?.convId; - const [viewingConversation, setViewingConversation] = - useState(null); + const [viewingChat, setViewingChat] = useState(null); const [pendingMessages, setPendingMessages] = useState< Record >({}); @@ -75,12 +89,12 @@ export const AppContextProvider = ({ useEffect(() => { // also reset the canvas data setCanvasData(null); - const handleConversationChange = (changedConvId: string) => { + const handleConversationChange = async (changedConvId: string) => { if (changedConvId !== convId) return; - setViewingConversation(StorageUtils.getOneConversation(convId)); + setViewingChat(await getViewingChat(changedConvId)); }; StorageUtils.onConversationChanged(handleConversationChange); - setViewingConversation(StorageUtils.getOneConversation(convId ?? '')); + getViewingChat(convId ?? '').then(setViewingChat); return () => { StorageUtils.offConversationChanged(handleConversationChange); }; @@ -118,23 +132,39 @@ export const AppContextProvider = ({ const generateMessage = async ( convId: string, - onChunk?: CallbackGeneratedChunk + leafNodeId: Message['id'], + onChunk: CallbackGeneratedChunk ) => { if (isGenerating(convId)) return; const config = StorageUtils.getConfig(); - const currConversation = StorageUtils.getOneConversation(convId); + const currConversation = await StorageUtils.getOneConversation(convId); if (!currConversation) { throw new Error('Current conversation is not found'); } + const currMessages = StorageUtils.filterByLeafNodeId( + await StorageUtils.getMessages(convId), + leafNodeId, + false + ); const abortController = new AbortController(); setAbort(convId, abortController); + if (!currMessages) { + throw new Error('Current messages are not found'); + } + + const pendingId = Date.now() + 1; let pendingMsg: PendingMessage = { - id: Date.now() + 1, + id: pendingId, + convId, + type: 'text', + timestamp: pendingId, role: 'assistant', content: null, + parent: leafNodeId, + children: [], }; setPending(convId, pendingMsg); @@ -144,7 +174,7 @@ export const AppContextProvider = ({ ...(config.systemMessage.length === 0 ? [] : [{ role: 'system', content: config.systemMessage } as APIMessage]), - ...normalizeMsgsForAPI(currConversation?.messages ?? []), + ...normalizeMsgsForAPI(currMessages), ]; if (config.excludeThoughtOnReq) { messages = filterThoughtFromMsgs(messages); @@ -205,8 +235,7 @@ export const AppContextProvider = ({ const lastContent = pendingMsg.content || ''; if (addedContent) { pendingMsg = { - id: pendingMsg.id, - role: 'assistant', + ...pendingMsg, content: lastContent + addedContent, }; } @@ -221,7 +250,7 @@ export const AppContextProvider = ({ }; } setPending(convId, pendingMsg); - onChunk?.(); + onChunk(); // don't need to switch node for pending message } } catch (err) { setPending(convId, null); @@ -236,37 +265,55 @@ export const AppContextProvider = ({ } } - if (pendingMsg.content) { - StorageUtils.appendMsg(currConversation.id, { - id: pendingMsg.id, - content: pendingMsg.content, - role: pendingMsg.role, - timings: pendingMsg.timings, - }); + if (pendingMsg.content !== null) { + await StorageUtils.appendMsg(pendingMsg as Message, leafNodeId); } setPending(convId, null); - onChunk?.(); // trigger scroll to bottom + onChunk(pendingId); // trigger scroll to bottom and switch to the last node }; const sendMessage = async ( - convId: string, + convId: string | null, + leafNodeId: Message['id'] | null, content: string, - onChunk?: CallbackGeneratedChunk + extra: Message['extra'], + onChunk: CallbackGeneratedChunk ): Promise => { - if (isGenerating(convId) || content.trim().length === 0) return false; + if (isGenerating(convId ?? '') || content.trim().length === 0) return false; - StorageUtils.appendMsg(convId, { - id: Date.now(), - role: 'user', - content, - }); + if (convId === null || convId.length === 0 || leafNodeId === null) { + const conv = await StorageUtils.createConversation( + content.substring(0, 256) + ); + convId = conv.id; + leafNodeId = conv.currNode; + // if user is creating a new conversation, redirect to the new conversation + navigate(`/chat/${convId}`); + } + + const now = Date.now(); + const currMsgId = now; + StorageUtils.appendMsg( + { + id: currMsgId, + timestamp: now, + type: 'text', + convId, + role: 'user', + content, + extra, + parent: leafNodeId, + children: [], + }, + leafNodeId + ); + onChunk(currMsgId); try { - await generateMessage(convId, onChunk); + await generateMessage(convId, currMsgId, onChunk); return true; } catch (_) { - // rollback - StorageUtils.popMsg(convId); + // TODO: rollback } return false; }; @@ -279,22 +326,35 @@ export const AppContextProvider = ({ // if content is undefined, we remove last assistant message const replaceMessageAndGenerate = async ( convId: string, - origMsgId: Message['id'], - content?: string, - onChunk?: CallbackGeneratedChunk + parentNodeId: Message['id'], // the parent node of the message to be replaced + content: string | null, + extra: Message['extra'], + onChunk: CallbackGeneratedChunk ) => { if (isGenerating(convId)) return; - StorageUtils.filterAndKeepMsgs(convId, (msg) => msg.id < origMsgId); - if (content) { - StorageUtils.appendMsg(convId, { - id: Date.now(), - role: 'user', - content, - }); + if (content !== null) { + const now = Date.now(); + const currMsgId = now; + StorageUtils.appendMsg( + { + id: currMsgId, + timestamp: now, + type: 'text', + convId, + role: 'user', + content, + extra, + parent: parentNodeId, + children: [], + }, + parentNodeId + ); + parentNodeId = currMsgId; } + onChunk(parentNodeId); - await generateMessage(convId, onChunk); + await generateMessage(convId, parentNodeId, onChunk); }; const saveConfig = (config: typeof CONFIG_DEFAULT) => { @@ -306,7 +366,7 @@ export const AppContextProvider = ({ { + const [extraContext, setExtraContext] = useState( + null + ); + + // Accept setText message from a parent window and set inputMsg and extraContext + useEffect(() => { + const handleMessage = (event: MessageEvent) => { + if (event.data?.command === 'setText') { + const data: SetTextEvData = event.data; + textarea.setValue(data?.text); + if (data?.context && data.context.length > 0) { + setExtraContext({ + type: 'context', + content: data.context, + }); + } + textarea.focus(); + } + }; + + window.addEventListener('message', handleMessage); + return () => window.removeEventListener('message', handleMessage); + }, [textarea]); + + // Add a keydown listener that sends the "escapePressed" message to the parent window + useEffect(() => { + const handleKeyDown = (event: KeyboardEvent) => { + if (event.key === 'Escape') { + window.parent.postMessage({ command: 'escapePressed' }, '*'); + } + }; + + window.addEventListener('keydown', handleKeyDown); + return () => window.removeEventListener('keydown', handleKeyDown); + }, []); + + return { + extraContext, + // call once the user message is sent, to clear the extra context + clearExtraContext: () => setExtraContext(null), + }; +}; diff --git a/examples/server/webui/src/utils/misc.ts b/examples/server/webui/src/utils/misc.ts index e3153fff5..87f55b2af 100644 --- a/examples/server/webui/src/utils/misc.ts +++ b/examples/server/webui/src/utils/misc.ts @@ -4,7 +4,6 @@ import { APIMessage, Message } from './types'; // ponyfill for missing ReadableStream asyncIterator on Safari import { asyncIterator } from '@sec-ant/readable-stream/ponyfill/asyncIterator'; -import { isDev } from '../Config'; // eslint-disable-next-line @typescript-eslint/no-explicit-any export const isString = (x: any) => !!x.toLowerCase; @@ -23,7 +22,7 @@ export async function* getSSEStreamAsync(fetchResponse: Response) { .pipeThrough(new TextLineStream()); // @ts-expect-error asyncIterator complains about type, but it should work for await (const line of asyncIterator(lines)) { - if (isDev) console.log({ line }); + //if (isDev) console.log({ line }); if (line.startsWith('data:') && !line.endsWith('[DONE]')) { const data = JSON.parse(line.slice(5)); yield data; @@ -54,12 +53,23 @@ export const copyStr = (textToCopy: string) => { /** * filter out redundant fields upon sending to API + * also format extra into text */ -export function normalizeMsgsForAPI(messages: Message[]) { +export function normalizeMsgsForAPI(messages: Readonly) { return messages.map((msg) => { + let newContent = ''; + + for (const extra of msg.extra ?? []) { + if (extra.type === 'context') { + newContent += `${extra.content}\n\n`; + } + } + + newContent += msg.content; + return { role: msg.role, - content: msg.content, + content: newContent, }; }) as APIMessage[]; } @@ -88,3 +98,31 @@ export function classNames(classes: Record): string { export const delay = (ms: number) => new Promise((resolve) => setTimeout(resolve, ms)); + +export const throttle = ( + callback: (...args: T) => void, + delay: number +) => { + let isWaiting = false; + + return (...args: T) => { + if (isWaiting) { + return; + } + + callback(...args); + isWaiting = true; + + setTimeout(() => { + isWaiting = false; + }, delay); + }; +}; + +export const cleanCurrentUrl = (removeQueryParams: string[]) => { + const url = new URL(window.location.href); + removeQueryParams.forEach((param) => { + url.searchParams.delete(param); + }); + window.history.replaceState({}, '', url.toString()); +}; diff --git a/examples/server/webui/src/utils/storage.ts b/examples/server/webui/src/utils/storage.ts index 8c03fa781..1dfc9d979 100644 --- a/examples/server/webui/src/utils/storage.ts +++ b/examples/server/webui/src/utils/storage.ts @@ -2,7 +2,8 @@ // format: { [convId]: { id: string, lastModified: number, messages: [...] } } import { CONFIG_DEFAULT } from '../Config'; -import { Conversation, Message } from './types'; +import { Conversation, Message, TimingReport } from './types'; +import Dexie, { Table } from 'dexie'; const event = new EventTarget(); @@ -17,85 +18,154 @@ const dispatchConversationChange = (convId: string) => { ); }; +const db = new Dexie('LlamacppWebui') as Dexie & { + conversations: Table; + messages: Table; +}; + +// https://dexie.org/docs/Version/Version.stores() +db.version(1).stores({ + // Unlike SQL, you don’t need to specify all properties but only the one you wish to index. + conversations: '&id, lastModified', + messages: '&id, convId, [convId+id], timestamp', +}); + // convId is a string prefixed with 'conv-' const StorageUtils = { /** * manage conversations */ - getAllConversations(): Conversation[] { - const res = []; - for (const key in localStorage) { - if (key.startsWith('conv-')) { - res.push(JSON.parse(localStorage.getItem(key) ?? '{}')); - } - } - res.sort((a, b) => b.lastModified - a.lastModified); - return res; + async getAllConversations(): Promise { + await migrationLStoIDB().catch(console.error); // noop if already migrated + return (await db.conversations.toArray()).sort( + (a, b) => b.lastModified - a.lastModified + ); }, /** * can return null if convId does not exist */ - getOneConversation(convId: string): Conversation | null { - return JSON.parse(localStorage.getItem(convId) || 'null'); + async getOneConversation(convId: string): Promise { + return (await db.conversations.where('id').equals(convId).first()) ?? null; }, /** - * if convId does not exist, create one + * get all message nodes in a conversation */ - appendMsg(convId: string, msg: Message): void { - if (msg.content === null) return; - const conv = StorageUtils.getOneConversation(convId) || { - id: convId, - lastModified: Date.now(), - messages: [], + async getMessages(convId: string): Promise { + return await db.messages.where({ convId }).toArray(); + }, + /** + * use in conjunction with getMessages to filter messages by leafNodeId + * includeRoot: whether to include the root node in the result + * if node with leafNodeId does not exist, return the path with the latest timestamp + */ + filterByLeafNodeId( + msgs: Readonly, + leafNodeId: Message['id'], + includeRoot: boolean + ): Readonly { + const res: Message[] = []; + const nodeMap = new Map(); + for (const msg of msgs) { + nodeMap.set(msg.id, msg); + } + let startNode: Message | undefined = nodeMap.get(leafNodeId); + if (!startNode) { + // if not found, we return the path with the latest timestamp + let latestTime = -1; + for (const msg of msgs) { + if (msg.timestamp > latestTime) { + startNode = msg; + latestTime = msg.timestamp; + } + } + } + // traverse the path from leafNodeId to root + // startNode can never be undefined here + let currNode: Message | undefined = startNode; + while (currNode) { + if (currNode.type !== 'root' || (currNode.type === 'root' && includeRoot)) + res.push(currNode); + currNode = nodeMap.get(currNode.parent ?? -1); + } + res.sort((a, b) => a.timestamp - b.timestamp); + return res; + }, + /** + * create a new conversation with a default root node + */ + async createConversation(name: string): Promise { + const now = Date.now(); + const msgId = now; + const conv: Conversation = { + id: `conv-${now}`, + lastModified: now, + currNode: msgId, + name, }; - conv.messages.push(msg); - conv.lastModified = Date.now(); - localStorage.setItem(convId, JSON.stringify(conv)); - dispatchConversationChange(convId); + await db.conversations.add(conv); + // create a root node + await db.messages.add({ + id: msgId, + convId: conv.id, + type: 'root', + timestamp: now, + role: 'system', + content: '', + parent: -1, + children: [], + }); + return conv; }, /** - * Get new conversation id + * if convId does not exist, throw an error */ - getNewConvId(): string { - return `conv-${Date.now()}`; + async appendMsg( + msg: Exclude, + parentNodeId: Message['id'] + ): Promise { + if (msg.content === null) return; + const { convId } = msg; + await db.transaction('rw', db.conversations, db.messages, async () => { + const conv = await StorageUtils.getOneConversation(convId); + const parentMsg = await db.messages + .where({ convId, id: parentNodeId }) + .first(); + // update the currNode of conversation + if (!conv) { + throw new Error(`Conversation ${convId} does not exist`); + } + if (!parentMsg) { + throw new Error( + `Parent message ID ${parentNodeId} does not exist in conversation ${convId}` + ); + } + await db.conversations.update(convId, { + lastModified: Date.now(), + currNode: msg.id, + }); + // update parent + await db.messages.update(parentNodeId, { + children: [...parentMsg.children, msg.id], + }); + // create message + await db.messages.add({ + ...msg, + parent: parentNodeId, + children: [], + }); + }); + dispatchConversationChange(convId); }, /** * remove conversation by id */ - remove(convId: string): void { - localStorage.removeItem(convId); + async remove(convId: string): Promise { + await db.transaction('rw', db.conversations, db.messages, async () => { + await db.conversations.delete(convId); + await db.messages.where({ convId }).delete(); + }); dispatchConversationChange(convId); }, - /** - * remove all conversations - */ - filterAndKeepMsgs( - convId: string, - predicate: (msg: Message) => boolean - ): void { - const conv = StorageUtils.getOneConversation(convId); - if (!conv) return; - conv.messages = conv.messages.filter(predicate); - conv.lastModified = Date.now(); - localStorage.setItem(convId, JSON.stringify(conv)); - dispatchConversationChange(convId); - }, - /** - * remove last message from conversation - */ - popMsg(convId: string): Message | undefined { - const conv = StorageUtils.getOneConversation(convId); - if (!conv) return; - const msg = conv.messages.pop(); - conv.lastModified = Date.now(); - if (conv.messages.length === 0) { - StorageUtils.remove(convId); - } else { - localStorage.setItem(convId, JSON.stringify(conv)); - } - dispatchConversationChange(convId); - return msg; - }, // event listeners onConversationChanged(callback: CallbackConversationChanged) { @@ -136,3 +206,79 @@ const StorageUtils = { }; export default StorageUtils; + +// Migration from localStorage to IndexedDB + +// these are old types, LS prefix stands for LocalStorage +interface LSConversation { + id: string; // format: `conv-{timestamp}` + lastModified: number; // timestamp from Date.now() + messages: LSMessage[]; +} +interface LSMessage { + id: number; + role: 'user' | 'assistant' | 'system'; + content: string; + timings?: TimingReport; +} +async function migrationLStoIDB() { + if (localStorage.getItem('migratedToIDB')) return; + const res: LSConversation[] = []; + for (const key in localStorage) { + if (key.startsWith('conv-')) { + res.push(JSON.parse(localStorage.getItem(key) ?? '{}')); + } + } + if (res.length === 0) return; + await db.transaction('rw', db.conversations, db.messages, async () => { + let migratedCount = 0; + for (const conv of res) { + const { id: convId, lastModified, messages } = conv; + const firstMsg = messages[0]; + const lastMsg = messages.at(-1); + if (messages.length < 2 || !firstMsg || !lastMsg) { + console.log( + `Skipping conversation ${convId} with ${messages.length} messages` + ); + continue; + } + const name = firstMsg.content ?? '(no messages)'; + await db.conversations.add({ + id: convId, + lastModified, + currNode: lastMsg.id, + name, + }); + const rootId = messages[0].id - 2; + await db.messages.add({ + id: rootId, + convId: convId, + type: 'root', + timestamp: rootId, + role: 'system', + content: '', + parent: -1, + children: [firstMsg.id], + }); + for (let i = 0; i < messages.length; i++) { + const msg = messages[i]; + await db.messages.add({ + ...msg, + type: 'text', + convId: convId, + timestamp: msg.id, + parent: i === 0 ? rootId : messages[i - 1].id, + children: i === messages.length - 1 ? [] : [messages[i + 1].id], + }); + } + migratedCount++; + console.log( + `Migrated conversation ${convId} with ${messages.length} messages` + ); + } + console.log( + `Migrated ${migratedCount} conversations from localStorage to IndexedDB` + ); + localStorage.setItem('migratedToIDB', '1'); + }); +} diff --git a/examples/server/webui/src/utils/types.ts b/examples/server/webui/src/utils/types.ts index 7cd12b40a..0eb774001 100644 --- a/examples/server/webui/src/utils/types.ts +++ b/examples/server/webui/src/utils/types.ts @@ -5,11 +5,60 @@ export interface TimingReport { predicted_ms: number; } +/** + * What is conversation "branching"? It is a feature that allows the user to edit an old message in the history, while still keeping the conversation flow. + * Inspired by ChatGPT / Claude / Hugging Chat where you edit a message, a new branch of the conversation is created, and the old message is still visible. + * + * We use the same node-based structure like other chat UIs, where each message has a parent and children. A "root" message is the first message in a conversation, which will not be displayed in the UI. + * + * root + * ├── message 1 + * │ └── message 2 + * │ └── message 3 + * └── message 4 + * └── message 5 + * + * In the above example, assuming that user wants to edit message 2, a new branch will be created: + * + * ├── message 2 + * │ └── message 3 + * └── message 6 + * + * Message 2 and 6 are siblings, and message 6 is the new branch. + * + * We only need to know the last node (aka leaf) to get the current branch. In the above example, message 5 is the leaf of branch containing message 4 and 5. + * + * For the implementation: + * - StorageUtils.getMessages() returns list of all nodes + * - StorageUtils.filterByLeafNodeId() filters the list of nodes from a given leaf node + */ + +// Note: the term "message" and "node" are used interchangeably in this context export interface Message { id: number; + convId: string; + type: 'text' | 'root'; + timestamp: number; // timestamp from Date.now() role: 'user' | 'assistant' | 'system'; content: string; timings?: TimingReport; + extra?: MessageExtra[]; + // node based system for branching + parent: Message['id']; + children: Message['id'][]; +} + +type MessageExtra = MessageExtraTextFile | MessageExtraContext; // TODO: will add more in the future + +export interface MessageExtraTextFile { + type: 'textFile'; + name: string; + content: string; +} + +export interface MessageExtraContext { + type: 'context'; + content: string; } export type APIMessage = Pick; @@ -17,7 +66,13 @@ export type APIMessage = Pick; export interface Conversation { id: string; // format: `conv-{timestamp}` lastModified: number; // timestamp from Date.now() - messages: Message[]; + currNode: Message['id']; // the current message node being viewed + name: string; +} + +export interface ViewingChat { + conv: Readonly; + messages: Readonly; } export type PendingMessage = Omit & { diff --git a/examples/server_embd.py b/examples/server_embd.py index 0e34c6cea..f8b0ffecd 100644 --- a/examples/server_embd.py +++ b/examples/server_embd.py @@ -15,7 +15,7 @@ async def main(): model_url = "http://127.0.0.1:6900" responses: list[requests.Response] = await asyncio.gather(*[requests_post_async( url= f"{model_url}/embedding", - json= {"content": str(0)*1024} + json= {"content": "a "*1022} ) for i in range(n)]) for response in responses: diff --git a/examples/simple-chat/simple-chat.cpp b/examples/simple-chat/simple-chat.cpp index c5534cc13..84f415973 100644 --- a/examples/simple-chat/simple-chat.cpp +++ b/examples/simple-chat/simple-chat.cpp @@ -98,7 +98,7 @@ int main(int argc, char ** argv) { auto generate = [&](const std::string & prompt) { std::string response; - const bool is_first = llama_get_kv_cache_used_cells(ctx) == 0; + const bool is_first = llama_kv_self_used_cells(ctx) == 0; // tokenize the prompt const int n_prompt_tokens = -llama_tokenize(vocab, prompt.c_str(), prompt.size(), NULL, 0, is_first, true); @@ -113,7 +113,7 @@ int main(int argc, char ** argv) { while (true) { // check if we have enough space in the context to evaluate this batch int n_ctx = llama_n_ctx(ctx); - int n_ctx_used = llama_get_kv_cache_used_cells(ctx); + int n_ctx_used = llama_kv_self_used_cells(ctx); if (n_ctx_used + batch.n_tokens > n_ctx) { printf("\033[0m\n"); fprintf(stderr, "context size exceeded\n"); diff --git a/examples/simple-cmake-pkg/README.md b/examples/simple-cmake-pkg/README.md index 8b30049e2..d7430cc9c 100644 --- a/examples/simple-cmake-pkg/README.md +++ b/examples/simple-cmake-pkg/README.md @@ -1,6 +1,6 @@ # llama.cpp/example/simple-cmake-pkg -This program builds [simple](../simple) using a relocatable CMake package. It serves as an example of using the `find_package()` CMake command to conveniently include [llama.cpp](https://github.com/ggerganov/llama.cpp) in projects which live outside of the source tree. +This program builds [simple](../simple) using a relocatable CMake package. It serves as an example of using the `find_package()` CMake command to conveniently include [llama.cpp](https://github.com/ggml-org/llama.cpp) in projects which live outside of the source tree. ## Building @@ -13,7 +13,7 @@ When hardware acceleration libraries are used (e.g. CUDA, Metal, Vulkan, etc.), ### Build llama.cpp and install to llama.cpp/inst ```sh -git clone https://github.com/ggerganov/llama.cpp +git clone https://github.com/ggml-org/llama.cpp cd llama.cpp cmake -S . -B build cmake --build build diff --git a/examples/speculative-simple/speculative-simple.cpp b/examples/speculative-simple/speculative-simple.cpp index 403ba2dd2..0783ed4a4 100644 --- a/examples/speculative-simple/speculative-simple.cpp +++ b/examples/speculative-simple/speculative-simple.cpp @@ -24,7 +24,7 @@ int main(int argc, char ** argv) { common_init(); - if (params.speculative.model.empty()) { + if (params.speculative.model.path.empty()) { LOG_ERR("%s: --model-draft is required\n", __func__); return 1; } @@ -217,7 +217,7 @@ int main(int argc, char ** argv) { { LOG_DBG("clear kv cache from any extra tokens, n_past = %d\n", n_past); - llama_kv_cache_seq_rm(ctx_tgt, 0, n_past, -1); + llama_kv_self_seq_rm(ctx_tgt, 0, n_past, -1); } if ((params.n_predict >= 0 && n_predict > params.n_predict) || has_eos) { diff --git a/examples/speculative/README.md b/examples/speculative/README.md index a6608c5fe..36ab37086 100644 --- a/examples/speculative/README.md +++ b/examples/speculative/README.md @@ -4,6 +4,6 @@ Demonstration of speculative decoding and tree-based speculative decoding techni More info: -- https://github.com/ggerganov/llama.cpp/pull/2926 -- https://github.com/ggerganov/llama.cpp/pull/3624 -- https://github.com/ggerganov/llama.cpp/pull/5625 +- https://github.com/ggml-org/llama.cpp/pull/2926 +- https://github.com/ggml-org/llama.cpp/pull/3624 +- https://github.com/ggml-org/llama.cpp/pull/5625 diff --git a/examples/speculative/speculative.cpp b/examples/speculative/speculative.cpp index c7ccea50d..561c30883 100644 --- a/examples/speculative/speculative.cpp +++ b/examples/speculative/speculative.cpp @@ -46,7 +46,7 @@ int main(int argc, char ** argv) { common_init(); - if (params.speculative.model.empty()) { + if (params.speculative.model.path.empty()) { LOG_ERR("%s: --model-draft is required\n", __func__); return 1; } @@ -331,11 +331,11 @@ int main(int argc, char ** argv) { } active_seqs.erase(s); - for(int i = 0; i < n_seq_dft; i++) { + for (int i = 0; i < n_seq_dft; i++) { if (i == s) { continue; } - if (drafts[i].tokens[i_dft] == drafts[s].tokens[i_dft]) { + if (drafts[i].active && drafts[i].tokens[i_dft] == drafts[s].tokens[i_dft]) { // synchronize active status for sequences with the same drafted token drafts[i].active = drafts[i].active && accept; if (!drafts[i].active) { @@ -420,14 +420,14 @@ int main(int argc, char ** argv) { { LOG_DBG("keeping sequence %d, n_past_tgt = %d, n_past_dft = %d\n", s_keep, n_past_tgt, n_past_dft); - llama_kv_cache_seq_keep(ctx_dft, s_keep); - llama_kv_cache_seq_cp (ctx_dft, s_keep, 0, -1, -1); - llama_kv_cache_seq_keep(ctx_dft, 0); + llama_kv_self_seq_keep(ctx_dft, s_keep); + llama_kv_self_seq_cp (ctx_dft, s_keep, 0, -1, -1); + llama_kv_self_seq_keep(ctx_dft, 0); - llama_kv_cache_seq_rm (ctx_tgt, s_keep, n_past_tgt, -1); - llama_kv_cache_seq_keep(ctx_tgt, s_keep); - llama_kv_cache_seq_cp (ctx_tgt, s_keep, 0, -1, -1); - llama_kv_cache_seq_keep(ctx_tgt, 0); + llama_kv_self_seq_rm (ctx_tgt, s_keep, n_past_tgt, -1); + llama_kv_self_seq_keep(ctx_tgt, s_keep); + llama_kv_self_seq_cp (ctx_tgt, s_keep, 0, -1, -1); + llama_kv_self_seq_keep(ctx_tgt, 0); } for (int s = 0; s < n_seq_dft; ++s) { @@ -444,7 +444,7 @@ int main(int argc, char ** argv) { common_batch_clear(batch_dft); common_batch_add (batch_dft, token_id, n_past_dft, { 0 }, true); - llama_kv_cache_seq_rm(ctx_dft, 0, n_past_dft, -1); + llama_kv_self_seq_rm(ctx_dft, 0, n_past_dft, -1); // LOG_DBG("dft batch: %s\n", LOG_BATCH_TOSTR_PRETTY(ctx_dft, batch_dft).c_str()); llama_decode(ctx_dft, batch_dft); @@ -503,8 +503,8 @@ int main(int argc, char ** argv) { if (n_seq_cur < n_seq_dft && cur_p->data[f].p > p_draft_split) { LOG_DBG("splitting seq %3d into %3d\n", s, n_seq_cur); - llama_kv_cache_seq_rm(ctx_dft, n_seq_cur, -1, -1); - llama_kv_cache_seq_cp(ctx_dft, s, n_seq_cur, -1, -1); + llama_kv_self_seq_rm(ctx_dft, n_seq_cur, -1, -1); + llama_kv_self_seq_cp(ctx_dft, s, n_seq_cur, -1, -1); // all previous tokens from this branch are now also part of the new branch for (int t = 0; t < batch_tgt.n_tokens; ++t) { @@ -585,9 +585,9 @@ int main(int argc, char ** argv) { // evaluate the target model on the drafted tokens { - llama_kv_cache_seq_keep(ctx_tgt, 0); + llama_kv_self_seq_keep(ctx_tgt, 0); for (int s = 1; s < n_seq_dft; ++s) { - llama_kv_cache_seq_cp(ctx_tgt, 0, s, -1, -1); + llama_kv_self_seq_cp(ctx_tgt, 0, s, -1, -1); } // LOG_DBG("target batch: %s\n", LOG_BATCH_TOSTR_PRETTY(ctx_tgt, batch_tgt).c_str()); diff --git a/examples/sycl/run-llama2.sh b/examples/sycl/run-llama2.sh index 3b9ba3b2d..8ce71e370 100755 --- a/examples/sycl/run-llama2.sh +++ b/examples/sycl/run-llama2.sh @@ -3,7 +3,7 @@ # MIT license # Copyright (C) 2024 Intel Corporation # SPDX-License-Identifier: MIT - +export ONEAPI_DEVICE_SELECTOR="level_zero:0" source /opt/intel/oneapi/setvars.sh #export GGML_SYCL_DEBUG=1 @@ -13,7 +13,7 @@ source /opt/intel/oneapi/setvars.sh INPUT_PROMPT="Building a website can be done in 10 simple steps:\nStep 1:" MODEL_FILE=models/llama-2-7b.Q4_0.gguf NGL=33 -CONEXT=8192 +CONEXT=4096 if [ $# -gt 0 ]; then GGML_SYCL_DEVICE=$1 diff --git a/examples/sycl/win-build-sycl.bat b/examples/sycl/win-build-sycl.bat index 17dd1ff5c..6fc897b14 100644 --- a/examples/sycl/win-build-sycl.bat +++ b/examples/sycl/win-build-sycl.bat @@ -13,10 +13,10 @@ if %errorlevel% neq 0 goto ERROR :: for FP16 :: faster for long-prompt inference -:: cmake -G "MinGW Makefiles" .. -DGGML_SYCL=ON -DCMAKE_CXX_COMPILER=icx -DBUILD_SHARED_LIBS=ON -DCMAKE_BUILD_TYPE=Release -DGGML_SYCL_F16=ON +:: cmake -G "MinGW Makefiles" .. -DLLAMA_CURL=OFF -DGGML_SYCL=ON -DCMAKE_CXX_COMPILER=icx -DBUILD_SHARED_LIBS=ON -DCMAKE_BUILD_TYPE=Release -DGGML_SYCL_F16=ON :: for FP32 -cmake -G "Ninja" .. -DGGML_SYCL=ON -DCMAKE_C_COMPILER=cl -DCMAKE_CXX_COMPILER=icx -DBUILD_SHARED_LIBS=ON -DCMAKE_BUILD_TYPE=Release +cmake -G "Ninja" .. -DLLAMA_CURL=OFF -DGGML_SYCL=ON -DCMAKE_C_COMPILER=cl -DCMAKE_CXX_COMPILER=icx -DBUILD_SHARED_LIBS=ON -DCMAKE_BUILD_TYPE=Release if %errorlevel% neq 0 goto ERROR :: build example/main only :: make main diff --git a/examples/tts/tts.cpp b/examples/tts/tts.cpp index f78f76303..0f0479869 100644 --- a/examples/tts/tts.cpp +++ b/examples/tts/tts.cpp @@ -1,10 +1,11 @@ +#define _USE_MATH_DEFINES // For M_PI on MSVC + #include "arg.h" #include "common.h" #include "sampling.h" #include "log.h" #include "llama.h" - -#define _USE_MATH_DEFINES // For M_PI on MSVC +#include "json.hpp" #include #include @@ -16,6 +17,13 @@ #include #include +using json = nlohmann::ordered_json; + +enum outetts_version { + OUTETTS_V0_2, + OUTETTS_V0_3, +}; + // // Terminal utils // @@ -79,11 +87,11 @@ struct wav_header { uint32_t data_size; }; -static void save_wav16(const std::string & fname, const std::vector & data, int sample_rate) { +static bool save_wav16(const std::string & fname, const std::vector & data, int sample_rate) { std::ofstream file(fname, std::ios::binary); if (!file) { - LOG_ERR("%s: Failed to open file '%s' for writing", __func__, fname.c_str()); - return; + LOG_ERR("%s: Failed to open file '%s' for writing.\n", __func__, fname.c_str()); + return false; } wav_header header; @@ -100,7 +108,7 @@ static void save_wav16(const std::string & fname, const std::vector & dat file.write(reinterpret_cast(&pcm_sample), sizeof(pcm_sample)); } - file.close(); + return file.good(); } static void fill_hann_window(int length, bool periodic, float * output) { @@ -371,7 +379,7 @@ static std::string replace_numbers_with_words(const std::string & input_text) { } // Based on: https://github.com/edwko/OuteTTS/blob/a613e79c489d8256dd657ea9168d78de75895d82/outetts/version/v1/prompt_processor.py#L39 -static std::string process_text(const std::string & text) { +static std::string process_text(const std::string & text, const outetts_version tts_version = OUTETTS_V0_2) { // For now I skipped text romanization as I am unsure how to handle // uroman and MeCab implementations in C++ @@ -401,7 +409,8 @@ static std::string process_text(const std::string & text) { if (c == ' ') { prompt_clean += "<|text_sep|>"; */ - processed_text = std::regex_replace(processed_text, std::regex(R"(\s)"), "<|text_sep|>"); + std::string separator = (tts_version == OUTETTS_V0_3) ? "<|space|>" : "<|text_sep|>"; + processed_text = std::regex_replace(processed_text, std::regex(R"(\s)"), separator); return processed_text; } @@ -425,8 +434,8 @@ static void prompt_init(llama_tokens & prompt, const llama_vocab * vocab) { prompt_add(prompt, vocab, "<|im_start|>\n", true, true); } -static std::vector prepare_guide_tokens(const llama_vocab * vocab, const std::string & str) { - const std::string& delimiter = "<|text_sep|>"; +static std::vector prepare_guide_tokens(const llama_vocab * vocab, const std::string & str, const outetts_version tts_version = OUTETTS_V0_2) { + const std::string& delimiter = (tts_version == OUTETTS_V0_3 ? "<|space|>" : "<|text_sep|>"); std::vector result; size_t start = 0; @@ -452,9 +461,82 @@ static std::vector prepare_guide_tokens(const llama_vocab * vocab, return result; } +static json speaker_from_file(const std::string & speaker_file) { + std::ifstream file(speaker_file); + if (!file) { + LOG_ERR("%s: Failed to open file '%s' for reading\n", __func__, speaker_file.c_str()); + return json(); + } + + json speaker = json::parse(file); + return speaker; +} + +static outetts_version get_tts_version(llama_model *model, json speaker = json::object()) { + if (speaker.contains("version")) { + std::string version = speaker["version"].get(); + if (version == "0.2") { + return OUTETTS_V0_2; + } else if (version == "0.3") { + return OUTETTS_V0_3; + } else { + LOG_ERR("%s: Unsupported speaker version '%s'\n", __func__, version.c_str()); + } + } + + // Also could get version from model itself + const char *chat_template = llama_model_chat_template(model, nullptr); + if (chat_template && std::string(chat_template) == "outetts-0.3") { + return OUTETTS_V0_3; + } + + // Use 0.2 as the default version + return OUTETTS_V0_2; +} + +static std::string audio_text_from_speaker(json speaker, const outetts_version tts_version = OUTETTS_V0_2) { + std::string audio_text = "<|text_start|>"; + + if (tts_version == OUTETTS_V0_2 || tts_version == OUTETTS_V0_3) { + std::string separator = (tts_version == OUTETTS_V0_3) ? "<|space|>" : "<|text_sep|>"; + for (const auto &word : speaker["words"]) { + audio_text += word["word"].get() + separator; + } + } + + return audio_text; +} + +static std::string audio_data_from_speaker(json speaker, const outetts_version tts_version = OUTETTS_V0_2) { + std::string audio_data = "<|audio_start|>\n"; + + if (tts_version == OUTETTS_V0_2 || tts_version == OUTETTS_V0_3) { + std::string code_start = (tts_version == OUTETTS_V0_3) ? "" : "<|code_start|>"; + std::string code_end = (tts_version == OUTETTS_V0_3) ? "<|space|>" : "<|code_end|>"; + for (const auto &word : speaker["words"]) { + std::string word_text = word["word"].get(); + double duration = word["duration"].get(); + std::vector codes = word["codes"].get>(); + + // Create the audio output entry + std::ostringstream word_entry; + word_entry << word_text << "<|t_" << std::fixed << std::setprecision(2) + << duration << "|>" + code_start; + for (const auto &Code : codes) { + word_entry << "<|" << Code << "|>"; + } + word_entry << code_end << "\n"; + audio_data += word_entry.str(); + } + } + + return audio_data; +} + int main(int argc, char ** argv) { common_params params; + params.out_file = "output.wav"; params.prompt = ""; params.n_predict = 4096; @@ -489,14 +571,13 @@ int main(int argc, char ** argv) { model_ttc = llama_init_ttc.model.get(); ctx_ttc = llama_init_ttc.context.get(); + if (model_ttc == nullptr || ctx_ttc == nullptr) { + return ENOENT; + } + const llama_vocab * vocab = llama_model_get_vocab(model_ttc); - // TODO: refactor in a common struct - params.model = params.vocoder.model; - params.model_url = params.vocoder.model_url; - params.hf_repo = params.vocoder.hf_repo; - params.hf_file = params.vocoder.hf_file; - + params.model = params.vocoder.model; params.embedding = true; common_init_result llama_init_cts = common_init_from_params(params); @@ -504,6 +585,10 @@ int main(int argc, char ** argv) { model_cts = llama_init_cts.model.get(); ctx_cts = llama_init_cts.context.get(); + if (model_cts == nullptr || ctx_cts == nullptr) { + return ENOENT; + } + std::vector smpl(n_parallel); for (int i = 0; i < n_parallel; ++i) { params.sampling.no_perf = (i != 0); @@ -523,34 +608,9 @@ int main(int argc, char ** argv) { std::vector codes; std::vector guide_tokens; - // process prompt and generate voice codes - { - LOG_INF("%s: constructing prompt ..\n", __func__); - - std::vector prompt_inp; - - prompt_init(prompt_inp, vocab); - - prompt_add(prompt_inp, vocab, "<|text_start|>the<|text_sep|>overall<|text_sep|>package<|text_sep|>from<|text_sep|>just<|text_sep|>two<|text_sep|>people<|text_sep|>is<|text_sep|>pretty<|text_sep|>remarkable<|text_sep|>sure<|text_sep|>i<|text_sep|>have<|text_sep|>some<|text_sep|>critiques<|text_sep|>about<|text_sep|>some<|text_sep|>of<|text_sep|>the<|text_sep|>gameplay<|text_sep|>aspects<|text_sep|>but<|text_sep|>its<|text_sep|>still<|text_sep|>really<|text_sep|>enjoyable<|text_sep|>and<|text_sep|>it<|text_sep|>looks<|text_sep|>lovely<|text_sep|>", false, true); - - // convert the input text into the necessary format expected by OuteTTS - { - std::string prompt_clean = process_text(params.prompt); - if (params.vocoder.use_guide_tokens) { - guide_tokens = prepare_guide_tokens(vocab, prompt_clean); - } - - LOG_INF("%s: prompt: '%s'\n", __func__, prompt_clean.c_str()); - - prompt_add(prompt_inp, vocab, prompt_clean, false, true); - } - - prompt_add(prompt_inp, vocab, "<|text_end|>\n", false, true); - - // disabled to save time on tokenizing each time - // TODO: load voices from the json files -#if 0 - const std::string voice_data = R"(<|audio_start|> + // the default speaker profile is from: https://github.com/edwko/OuteTTS/blob/main/outetts/version/v1/default_speakers/en_male_1.json + std::string audio_text = "<|text_start|>the<|text_sep|>overall<|text_sep|>package<|text_sep|>from<|text_sep|>just<|text_sep|>two<|text_sep|>people<|text_sep|>is<|text_sep|>pretty<|text_sep|>remarkable<|text_sep|>sure<|text_sep|>i<|text_sep|>have<|text_sep|>some<|text_sep|>critiques<|text_sep|>about<|text_sep|>some<|text_sep|>of<|text_sep|>the<|text_sep|>gameplay<|text_sep|>aspects<|text_sep|>but<|text_sep|>its<|text_sep|>still<|text_sep|>really<|text_sep|>enjoyable<|text_sep|>and<|text_sep|>it<|text_sep|>looks<|text_sep|>lovely<|text_sep|>"; + std::string audio_data = R"(<|audio_start|> the<|t_0.08|><|code_start|><|257|><|740|><|636|><|913|><|788|><|1703|><|code_end|> overall<|t_0.36|><|code_start|><|127|><|201|><|191|><|774|><|700|><|532|><|1056|><|557|><|798|><|298|><|1741|><|747|><|1662|><|1617|><|1702|><|1527|><|368|><|1588|><|1049|><|1008|><|1625|><|747|><|1576|><|728|><|1019|><|1696|><|1765|><|code_end|> package<|t_0.56|><|code_start|><|935|><|584|><|1319|><|627|><|1016|><|1491|><|1344|><|1117|><|1526|><|1040|><|239|><|1435|><|951|><|498|><|723|><|1180|><|535|><|789|><|1649|><|1637|><|78|><|465|><|1668|><|901|><|595|><|1675|><|117|><|1009|><|1667|><|320|><|840|><|79|><|507|><|1762|><|1508|><|1228|><|1768|><|802|><|1450|><|1457|><|232|><|639|><|code_end|> @@ -582,117 +642,172 @@ it<|t_0.09|><|code_start|><|848|><|1366|><|395|><|1601|><|1513|><|593|><|1302|>< looks<|t_0.27|><|code_start|><|1281|><|1266|><|1755|><|572|><|248|><|1751|><|1257|><|695|><|1380|><|457|><|659|><|585|><|1315|><|1105|><|1776|><|736|><|24|><|736|><|654|><|1027|><|code_end|> lovely<|t_0.56|><|code_start|><|634|><|596|><|1766|><|1556|><|1306|><|1285|><|1481|><|1721|><|1123|><|438|><|1246|><|1251|><|795|><|659|><|1381|><|1658|><|217|><|1772|><|562|><|952|><|107|><|1129|><|1112|><|467|><|550|><|1079|><|840|><|1615|><|1469|><|1380|><|168|><|917|><|836|><|1827|><|437|><|583|><|67|><|595|><|1087|><|1646|><|1493|><|1677|><|code_end|>)"; - auto tmp = common_tokenize(vocab, voice_data, false, true); - printf("\n\n"); - for (int i = 0; i < tmp.size(); ++i) { - printf("%d, ", tmp[i]); + // audio data for 0.3 version + outetts_version tts_version = get_tts_version(model_ttc); + if (tts_version == OUTETTS_V0_3) { + audio_text = std::regex_replace(audio_text, std::regex(R"(<\|text_sep\|>)"), "<|space|>"); + audio_data = std::regex_replace(audio_data, std::regex(R"(<\|code_start\|>)"), ""); + audio_data = std::regex_replace(audio_data, std::regex(R"(<\|code_end\|>)"), "<|space|>"); + } + + // load speaker if given + if (!params.vocoder.speaker_file.empty()) { + LOG_INF("%s: loading speaker ..\n", __func__); + json speaker = speaker_from_file(params.vocoder.speaker_file); + if (speaker.empty()) { + LOG_ERR("%s: Failed to load speaker file '%s'\n", __func__, params.vocoder.speaker_file.c_str()); + return 1; } - printf("\n\n"); + audio_text = audio_text_from_speaker(speaker, tts_version); + audio_data = audio_data_from_speaker(speaker, tts_version); + } + + // process prompt and generate voice codes + { + LOG_INF("%s: constructing prompt ..\n", __func__); + + std::vector prompt_inp; + + prompt_init(prompt_inp, vocab); + + prompt_add(prompt_inp, vocab, audio_text, false, true); + + // convert the input text into the necessary format expected by OuteTTS + { + std::string prompt_clean = process_text(params.prompt, tts_version); + if (params.vocoder.use_guide_tokens) { + guide_tokens = prepare_guide_tokens(vocab, prompt_clean, tts_version); + } + + LOG_INF("%s: prompt: '%s'\n", __func__, prompt_clean.c_str()); + + prompt_add(prompt_inp, vocab, prompt_clean, false, true); + } + + prompt_add(prompt_inp, vocab, "<|text_end|>\n", false, true); + + if (!params.vocoder.speaker_file.empty()) { + prompt_add(prompt_inp, vocab, audio_data, false, true); + } else { + // disabled to save time on tokenizing each time +#if 1 + const std::string voice_data = audio_data; + + auto tmp = common_tokenize(vocab, voice_data, false, true); + + std::ostringstream tokens_oss; + for (size_t i = 0; i < tmp.size(); ++i) { + tokens_oss << tmp[i] << ", "; + } + LOG_INF("\n\n%s: llama tokens: %s\n\n", __func__, tokens_oss.str().c_str()); + + prompt_add(prompt_inp, tmp); #else - prompt_add(prompt_inp, llama_tokens { - 151667, 198, 1782, 155780, 151669, 151929, 152412, 152308, 152585, - 152460, 153375, 151670, 198, 74455, 155808, 151669, 151799, - 151873, 151863, 152446, 152372, 152204, 152728, 152229, 152470, - 151970, 153413, 152419, 153334, 153289, 153374, 153199, 152040, - 153260, 152721, 152680, 153297, 152419, 153248, 152400, 152691, - 153368, 153437, 151670, 198, 1722, 155828, 151669, 152607, - 152256, 152991, 152299, 152688, 153163, 153016, 152789, 153198, - 152712, 151911, 153107, 152623, 152170, 152395, 152852, 152207, - 152461, 153321, 153309, 151750, 152137, 153340, 152573, 152267, - 153347, 151789, 152681, 153339, 151992, 152512, 151751, 152179, - 153434, 153180, 152900, 153440, 152474, 153122, 153129, 151904, - 152311, 151670, 198, 1499, 155791, 151669, 152276, 152454, - 153354, 152544, 153204, 153272, 152708, 153433, 152319, 153226, - 153043, 152325, 153267, 152622, 151670, 198, 4250, 155797, - 151669, 153454, 153342, 151989, 152458, 153420, 152303, 152271, - 152827, 153036, 153196, 151708, 153263, 152561, 153207, 152213, - 152112, 153204, 151722, 152542, 151670, 198, 19789, 155796, - 151669, 153353, 153182, 152345, 152471, 152477, 153014, 152002, - 152191, 151734, 152312, 152810, 152237, 153224, 153169, 153224, - 152244, 153387, 153404, 151670, 198, 16069, 155811, 151669, - 152265, 151946, 151808, 152412, 152363, 152305, 153156, 152733, - 152810, 153157, 152016, 152100, 152069, 153234, 152317, 152589, - 152707, 153121, 153341, 152159, 152114, 153156, 153001, 153504, - 153376, 152272, 152433, 152325, 151941, 151670, 198, 285, - 155788, 151669, 152238, 152255, 153427, 152318, 153009, 152381, - 152474, 152680, 152157, 153255, 152324, 151682, 151670, 198, - 32955, 155804, 151669, 153490, 153419, 152364, 152405, 152682, - 152206, 152078, 153369, 152725, 153193, 153027, 152946, 152488, - 153070, 151883, 152890, 152489, 153144, 153375, 152358, 151685, - 152494, 152117, 152740, 151670, 198, 37448, 480, 155840, 151669, - 151902, 152720, 153377, 152027, 152378, 152821, 153207, 153459, - 153028, 153068, 152507, 153255, 152158, 152921, 151958, 152609, - 152748, 152822, 152286, 151714, 152730, 152377, 152353, 152470, - 152606, 152162, 152186, 153071, 152244, 153118, 153375, 153018, - 152712, 153098, 152976, 152336, 151843, 153202, 152297, 151736, - 153380, 153502, 152702, 152115, 153181, 152735, 153277, 153457, - 152393, 153112, 152595, 151670, 198, 19098, 155808, 151669, - 152464, 153452, 152595, 153312, 151937, 151933, 153197, 152239, - 153163, 152922, 153402, 152034, 152591, 153438, 152215, 151673, - 152005, 151785, 152642, 151924, 153278, 151805, 151974, 153482, - 152718, 152862, 153347, 151670, 198, 72, 155780, 151669, 151795, - 152111, 152746, 152377, 153471, 152309, 151670, 198, 19016, - 155788, 151669, 153181, 152271, 152190, 152842, 152224, 152701, - 152939, 152536, 152091, 151815, 152733, 151672, 151670, 198, - 14689, 155788, 151669, 152291, 152072, 152942, 151734, 153042, - 153504, 152589, 153333, 151839, 151941, 153038, 153180, 151670, - 198, 36996, 8303, 155832, 151669, 152231, 152256, 152835, - 152801, 152985, 153400, 152393, 152818, 152765, 152249, 152600, - 151699, 152302, 152752, 153018, 153009, 151992, 153054, 152847, - 153354, 153228, 152662, 153355, 152532, 153393, 151782, 152458, - 152048, 152757, 152428, 153195, 151906, 153006, 153178, 153250, - 152331, 152284, 152780, 153138, 153319, 151980, 153142, 152418, - 152228, 152733, 151670, 198, 9096, 155801, 151669, 151698, - 153321, 152217, 153039, 152935, 153400, 152122, 152531, 153106, - 152169, 152892, 152957, 151851, 152427, 152826, 152451, 151851, - 152901, 152885, 152594, 153446, 153080, 151670, 198, 14689, - 155795, 151669, 152658, 151700, 153321, 152450, 152530, 153191, - 151673, 151690, 151698, 152714, 152846, 152981, 153171, 153384, - 153364, 153188, 153246, 151670, 198, 1055, 155779, 151669, - 151869, 152388, 152711, 153334, 151736, 151670, 198, 1782, - 155780, 151669, 153483, 153240, 152241, 152558, 152697, 153046, - 151670, 198, 5804, 1363, 155820, 151669, 152941, 152764, 152605, - 153034, 153434, 153372, 153347, 151887, 152453, 152758, 152133, - 152510, 152694, 152431, 152321, 153088, 152676, 152223, 152581, - 152459, 152015, 152502, 153063, 152712, 153294, 153451, 153032, - 152903, 152859, 152989, 151748, 152669, 152661, 152650, 152409, - 151861, 151670, 198, 300, 7973, 155828, 151669, 153095, 152469, - 152988, 152894, 151819, 152391, 153019, 152058, 153062, 153230, - 151826, 152112, 152306, 152264, 152769, 153390, 152384, 152435, - 152790, 153393, 152983, 152540, 152252, 152034, 153107, 152540, - 151919, 151893, 152558, 152817, 152946, 152956, 152129, 152715, - 153131, 153490, 151734, 152271, 152707, 151734, 153321, 152450, - 151670, 198, 8088, 155792, 151669, 152452, 153497, 153353, - 152679, 152533, 152382, 152374, 152611, 153341, 153163, 152285, - 153411, 152495, 153141, 152320, 151670, 198, 1199, 155781, - 151669, 151764, 152360, 153295, 152634, 153342, 152199, 152271, - 151670, 198, 43366, 155799, 151669, 152308, 151682, 152889, - 152016, 152385, 152629, 152495, 151826, 153321, 152958, 152180, - 151886, 153432, 152922, 152128, 153024, 153040, 152593, 152287, - 151677, 151670, 198, 53660, 155808, 151669, 151727, 152092, - 152680, 153331, 151699, 152316, 152938, 152289, 152433, 153384, - 151781, 153137, 153259, 152175, 153213, 152291, 151869, 152691, - 152489, 151941, 152049, 152034, 153053, 152179, 153160, 151676, - 153367, 151670, 198, 268, 4123, 480, 155821, 151669, 152350, - 152173, 152536, 151991, 151960, 153144, 153013, 152358, 152234, - 153135, 152291, 153235, 152143, 152583, 152402, 153483, 152678, - 152192, 152533, 152946, 151797, 153103, 152310, 152293, 151825, - 152548, 153442, 152109, 152659, 153325, 152781, 152570, 152957, - 151752, 152265, 153381, 152515, 151670, 198, 437, 155787, - 151669, 152957, 152659, 151975, 152709, 152402, 152836, 152174, - 151792, 153409, 153327, 152990, 151670, 198, 275, 155781, - 151669, 152520, 153038, 152067, 153273, 153185, 152265, 152974, - 151670, 198, 94273, 155799, 151669, 152953, 152938, 153427, - 152244, 151920, 153423, 152929, 152367, 153052, 152129, 152331, - 152257, 152987, 152777, 153448, 152408, 151696, 152408, 152326, - 152699, 151670, 198, 385, 16239, 155828, 151669, 152306, 152268, - 153438, 153228, 152978, 152957, 153153, 153393, 152795, 152110, - 152918, 152923, 152467, 152331, 153053, 153330, 151889, 153444, - 152234, 152624, 151779, 152801, 152784, 152139, 152222, 152751, - 152512, 153287, 153141, 153052, 151840, 152589, 152508, 153499, - 152109, 152255, 151739, 152267, 152759, 153318, 153165, 153349, - 151670,}); + prompt_add(prompt_inp, llama_tokens { + 151667, 198, 1782, 155780, 151669, 151929, 152412, 152308, 152585, + 152460, 153375, 151670, 198, 74455, 155808, 151669, 151799, + 151873, 151863, 152446, 152372, 152204, 152728, 152229, 152470, + 151970, 153413, 152419, 153334, 153289, 153374, 153199, 152040, + 153260, 152721, 152680, 153297, 152419, 153248, 152400, 152691, + 153368, 153437, 151670, 198, 1722, 155828, 151669, 152607, + 152256, 152991, 152299, 152688, 153163, 153016, 152789, 153198, + 152712, 151911, 153107, 152623, 152170, 152395, 152852, 152207, + 152461, 153321, 153309, 151750, 152137, 153340, 152573, 152267, + 153347, 151789, 152681, 153339, 151992, 152512, 151751, 152179, + 153434, 153180, 152900, 153440, 152474, 153122, 153129, 151904, + 152311, 151670, 198, 1499, 155791, 151669, 152276, 152454, + 153354, 152544, 153204, 153272, 152708, 153433, 152319, 153226, + 153043, 152325, 153267, 152622, 151670, 198, 4250, 155797, + 151669, 153454, 153342, 151989, 152458, 153420, 152303, 152271, + 152827, 153036, 153196, 151708, 153263, 152561, 153207, 152213, + 152112, 153204, 151722, 152542, 151670, 198, 19789, 155796, + 151669, 153353, 153182, 152345, 152471, 152477, 153014, 152002, + 152191, 151734, 152312, 152810, 152237, 153224, 153169, 153224, + 152244, 153387, 153404, 151670, 198, 16069, 155811, 151669, + 152265, 151946, 151808, 152412, 152363, 152305, 153156, 152733, + 152810, 153157, 152016, 152100, 152069, 153234, 152317, 152589, + 152707, 153121, 153341, 152159, 152114, 153156, 153001, 153504, + 153376, 152272, 152433, 152325, 151941, 151670, 198, 285, + 155788, 151669, 152238, 152255, 153427, 152318, 153009, 152381, + 152474, 152680, 152157, 153255, 152324, 151682, 151670, 198, + 32955, 155804, 151669, 153490, 153419, 152364, 152405, 152682, + 152206, 152078, 153369, 152725, 153193, 153027, 152946, 152488, + 153070, 151883, 152890, 152489, 153144, 153375, 152358, 151685, + 152494, 152117, 152740, 151670, 198, 37448, 480, 155840, 151669, + 151902, 152720, 153377, 152027, 152378, 152821, 153207, 153459, + 153028, 153068, 152507, 153255, 152158, 152921, 151958, 152609, + 152748, 152822, 152286, 151714, 152730, 152377, 152353, 152470, + 152606, 152162, 152186, 153071, 152244, 153118, 153375, 153018, + 152712, 153098, 152976, 152336, 151843, 153202, 152297, 151736, + 153380, 153502, 152702, 152115, 153181, 152735, 153277, 153457, + 152393, 153112, 152595, 151670, 198, 19098, 155808, 151669, + 152464, 153452, 152595, 153312, 151937, 151933, 153197, 152239, + 153163, 152922, 153402, 152034, 152591, 153438, 152215, 151673, + 152005, 151785, 152642, 151924, 153278, 151805, 151974, 153482, + 152718, 152862, 153347, 151670, 198, 72, 155780, 151669, 151795, + 152111, 152746, 152377, 153471, 152309, 151670, 198, 19016, + 155788, 151669, 153181, 152271, 152190, 152842, 152224, 152701, + 152939, 152536, 152091, 151815, 152733, 151672, 151670, 198, + 14689, 155788, 151669, 152291, 152072, 152942, 151734, 153042, + 153504, 152589, 153333, 151839, 151941, 153038, 153180, 151670, + 198, 36996, 8303, 155832, 151669, 152231, 152256, 152835, + 152801, 152985, 153400, 152393, 152818, 152765, 152249, 152600, + 151699, 152302, 152752, 153018, 153009, 151992, 153054, 152847, + 153354, 153228, 152662, 153355, 152532, 153393, 151782, 152458, + 152048, 152757, 152428, 153195, 151906, 153006, 153178, 153250, + 152331, 152284, 152780, 153138, 153319, 151980, 153142, 152418, + 152228, 152733, 151670, 198, 9096, 155801, 151669, 151698, + 153321, 152217, 153039, 152935, 153400, 152122, 152531, 153106, + 152169, 152892, 152957, 151851, 152427, 152826, 152451, 151851, + 152901, 152885, 152594, 153446, 153080, 151670, 198, 14689, + 155795, 151669, 152658, 151700, 153321, 152450, 152530, 153191, + 151673, 151690, 151698, 152714, 152846, 152981, 153171, 153384, + 153364, 153188, 153246, 151670, 198, 1055, 155779, 151669, + 151869, 152388, 152711, 153334, 151736, 151670, 198, 1782, + 155780, 151669, 153483, 153240, 152241, 152558, 152697, 153046, + 151670, 198, 5804, 1363, 155820, 151669, 152941, 152764, 152605, + 153034, 153434, 153372, 153347, 151887, 152453, 152758, 152133, + 152510, 152694, 152431, 152321, 153088, 152676, 152223, 152581, + 152459, 152015, 152502, 153063, 152712, 153294, 153451, 153032, + 152903, 152859, 152989, 151748, 152669, 152661, 152650, 152409, + 151861, 151670, 198, 300, 7973, 155828, 151669, 153095, 152469, + 152988, 152894, 151819, 152391, 153019, 152058, 153062, 153230, + 151826, 152112, 152306, 152264, 152769, 153390, 152384, 152435, + 152790, 153393, 152983, 152540, 152252, 152034, 153107, 152540, + 151919, 151893, 152558, 152817, 152946, 152956, 152129, 152715, + 153131, 153490, 151734, 152271, 152707, 151734, 153321, 152450, + 151670, 198, 8088, 155792, 151669, 152452, 153497, 153353, + 152679, 152533, 152382, 152374, 152611, 153341, 153163, 152285, + 153411, 152495, 153141, 152320, 151670, 198, 1199, 155781, + 151669, 151764, 152360, 153295, 152634, 153342, 152199, 152271, + 151670, 198, 43366, 155799, 151669, 152308, 151682, 152889, + 152016, 152385, 152629, 152495, 151826, 153321, 152958, 152180, + 151886, 153432, 152922, 152128, 153024, 153040, 152593, 152287, + 151677, 151670, 198, 53660, 155808, 151669, 151727, 152092, + 152680, 153331, 151699, 152316, 152938, 152289, 152433, 153384, + 151781, 153137, 153259, 152175, 153213, 152291, 151869, 152691, + 152489, 151941, 152049, 152034, 153053, 152179, 153160, 151676, + 153367, 151670, 198, 268, 4123, 480, 155821, 151669, 152350, + 152173, 152536, 151991, 151960, 153144, 153013, 152358, 152234, + 153135, 152291, 153235, 152143, 152583, 152402, 153483, 152678, + 152192, 152533, 152946, 151797, 153103, 152310, 152293, 151825, + 152548, 153442, 152109, 152659, 153325, 152781, 152570, 152957, + 151752, 152265, 153381, 152515, 151670, 198, 437, 155787, + 151669, 152957, 152659, 151975, 152709, 152402, 152836, 152174, + 151792, 153409, 153327, 152990, 151670, 198, 275, 155781, + 151669, 152520, 153038, 152067, 153273, 153185, 152265, 152974, + 151670, 198, 94273, 155799, 151669, 152953, 152938, 153427, + 152244, 151920, 153423, 152929, 152367, 153052, 152129, 152331, + 152257, 152987, 152777, 153448, 152408, 151696, 152408, 152326, + 152699, 151670, 198, 385, 16239, 155828, 151669, 152306, 152268, + 153438, 153228, 152978, 152957, 153153, 153393, 152795, 152110, + 152918, 152923, 152467, 152331, 153053, 153330, 151889, 153444, + 152234, 152624, 151779, 152801, 152784, 152139, 152222, 152751, + 152512, 153287, 153141, 153052, 151840, 152589, 152508, 153499, + 152109, 152255, 151739, 152267, 152759, 153318, 153165, 153349, + 151670,}); #endif + } // print the prompt token-by-token @@ -951,8 +1066,6 @@ lovely<|t_0.56|><|code_start|><|634|><|596|><|1766|><|1556|><|1306|><|1285|><|14 } #endif - const std::string fname = "output.wav"; - const int n_sr = 24000; // sampling rate // zero out first 0.25 seconds @@ -963,11 +1076,15 @@ lovely<|t_0.56|><|code_start|><|634|><|596|><|1766|><|1556|><|1306|><|1285|><|14 LOG_INF("%s: time for spectral ops: %.3f ms\n", __func__, (ggml_time_us() - t_spec_start) / 1000.0f); LOG_INF("%s: total time: %.3f ms\n", __func__, (ggml_time_us() - t_main_start) / 1000.0f); - save_wav16(fname, audio, n_sr); + int retval = 0; - LOG_INF("%s: audio written to file '%s'\n", __func__, fname.c_str()); + if (save_wav16(params.out_file, audio, n_sr)) { + LOG_INF("%s: audio written to file '%s'\n", __func__, params.out_file.c_str()); + } else { + retval = ENOENT; + } llama_backend_free(); - return 0; + return retval; } diff --git a/flake.nix b/flake.nix index 26a258816..0b5edf911 100644 --- a/flake.nix +++ b/flake.nix @@ -36,7 +36,7 @@ # ``` # nixConfig = { # extra-substituters = [ - # # Populated by the CI in ggerganov/llama.cpp + # # Populated by the CI in ggml-org/llama.cpp # "https://llama-cpp.cachix.org" # # # A development cache for nixpkgs imported with `config.cudaSupport = true`. @@ -56,11 +56,11 @@ # }; # ``` - # For inspection, use `nix flake show github:ggerganov/llama.cpp` or the nix repl: + # For inspection, use `nix flake show github:ggml-org/llama.cpp` or the nix repl: # # ```bash # ❯ nix repl - # nix-repl> :lf github:ggerganov/llama.cpp + # nix-repl> :lf github:ggml-org/llama.cpp # Added 13 variables. # nix-repl> outputs.apps.x86_64-linux.quantize # { program = "/nix/store/00000000000000000000000000000000-llama.cpp/bin/llama-quantize"; type = "app"; } @@ -176,7 +176,7 @@ # # We could test all outputs e.g. as `checks = confg.packages`. # - # TODO: Build more once https://github.com/ggerganov/llama.cpp/issues/6346 has been addressed + # TODO: Build more once https://github.com/ggml-org/llama.cpp/issues/6346 has been addressed checks = { inherit (config.packages) default vulkan; }; diff --git a/ggml/CMakeLists.txt b/ggml/CMakeLists.txt index 75b5ea3b4..d33f843b4 100644 --- a/ggml/CMakeLists.txt +++ b/ggml/CMakeLists.txt @@ -100,11 +100,17 @@ else() set(INS_ENB ON) endif() +message(DEBUG "GGML_NATIVE : ${GGML_NATIVE}") +message(DEBUG "GGML_NATIVE_DEFAULT : ${GGML_NATIVE_DEFAULT}") +message(DEBUG "INS_ENB : ${INS_ENB}") + option(GGML_CPU_HBM "ggml: use memkind for CPU HBM" OFF) option(GGML_CPU_AARCH64 "ggml: use runtime weight conversion of Q4_0 to Q4_X_X" ON) +option(GGML_CPU_KLEIDIAI "ggml: use KleidiAI optimized kernels if applicable" OFF) option(GGML_AVX "ggml: enable AVX" ${INS_ENB}) option(GGML_AVX_VNNI "ggml: enable AVX-VNNI" OFF) option(GGML_AVX2 "ggml: enable AVX2" ${INS_ENB}) +option(GGML_BMI2 "ggml: enable BMI2" ${INS_ENB}) option(GGML_AVX512 "ggml: enable AVX512F" OFF) option(GGML_AVX512_VBMI "ggml: enable AVX512-VBMI" OFF) option(GGML_AVX512_VNNI "ggml: enable AVX512-VNNI" OFF) @@ -121,9 +127,12 @@ endif() option(GGML_LASX "ggml: enable lasx" ON) option(GGML_LSX "ggml: enable lsx" ON) option(GGML_RVV "ggml: enable rvv" ON) +option(GGML_RV_ZFH "ggml: enable riscv zfh" OFF) +option(GGML_VXE "ggml: enable vxe" ON) option(GGML_CPU_ALL_VARIANTS "ggml: build all variants of the CPU backend (requires GGML_BACKEND_DL)" OFF) -set(GGML_CPU_ARM_ARCH "" CACHE STRING "ggml: CPU architecture for ARM") +set(GGML_CPU_ARM_ARCH "" CACHE STRING "ggml: CPU architecture for ARM") +set(GGML_CPU_POWERPC_CPUTYPE "" CACHE STRING "ggml: CPU type for PowerPC") if (WIN32) @@ -150,12 +159,17 @@ set (GGML_CUDA_PEER_MAX_BATCH_SIZE "128" CACHE STRING "ggml: max. batch size for using peer access") option(GGML_CUDA_NO_PEER_COPY "ggml: do not use peer to peer copies" OFF) option(GGML_CUDA_NO_VMM "ggml: do not try to use CUDA VMM" OFF) +option(GGML_CUDA_FA "ggml: compile ggml FlashAttention CUDA kernels" ON) option(GGML_CUDA_FA_ALL_QUANTS "ggml: compile all quants for FlashAttention" OFF) option(GGML_CUDA_GRAPHS "ggml: use CUDA graphs (llama.cpp only)" ${GGML_CUDA_GRAPHS_DEFAULT}) +set (GGML_CUDA_COMPRESSION_MODE "size" CACHE STRING + "ggml: cuda link binary compression mode; requires cuda 12.8+") +set_property(CACHE GGML_CUDA_COMPRESSION_MODE PROPERTY STRINGS "none;speed;balance;size") option(GGML_HIP "ggml: use HIP" OFF) option(GGML_HIP_GRAPHS "ggml: use HIP graph, experimental, slow" OFF) option(GGML_HIP_NO_VMM "ggml: do not try to use HIP VMM" ON) +option(GGML_HIP_ROCWMMA_FATTN "ggml: enable rocWMMA for FlashAttention" OFF) option(GGML_HIP_UMA "ggml: use HIP unified memory architecture" OFF) option(GGML_VULKAN "ggml: use Vulkan" OFF) option(GGML_VULKAN_CHECK_RESULTS "ggml: run Vulkan op checks" OFF) @@ -178,6 +192,7 @@ option(GGML_OPENMP "ggml: use OpenMP" option(GGML_RPC "ggml: use RPC" OFF) option(GGML_SYCL "ggml: use SYCL" OFF) option(GGML_SYCL_F16 "ggml: use 16 bit floats for sycl calculations" OFF) +option(GGML_SYCL_GRAPH "ggml: enable graphs in the SYCL backend" ON) set (GGML_SYCL_TARGET "INTEL" CACHE STRING "ggml: sycl target device") set (GGML_SYCL_DEVICE_ARCH "" CACHE STRING @@ -187,6 +202,8 @@ option(GGML_OPENCL "ggml: use OpenCL" option(GGML_OPENCL_PROFILING "ggml: use OpenCL profiling (increases overhead)" OFF) option(GGML_OPENCL_EMBED_KERNELS "ggml: embed kernels" ON) option(GGML_OPENCL_USE_ADRENO_KERNELS "ggml: use optimized kernels for Adreno" ON) +set (GGML_OPENCL_TARGET_VERSION "300" CACHE STRING + "gmml: OpenCL API version to target") # toolchain for vulkan-shaders-gen set (GGML_VULKAN_SHADERS_GEN_TOOLCHAIN "" CACHE FILEPATH "ggml: toolchain file for vulkan-shaders-gen") @@ -209,6 +226,8 @@ set(THREADS_PREFER_PTHREAD_FLAG ON) find_package(Threads REQUIRED) +include(GNUInstallDirs) + # # build the library # @@ -232,7 +251,6 @@ endif () # install # -include(GNUInstallDirs) include(CMakePackageConfigHelpers) # all public headers @@ -243,6 +261,7 @@ set(GGML_PUBLIC_HEADERS include/ggml-backend.h include/ggml-blas.h include/ggml-cann.h + include/ggml-cpp.h include/ggml-cuda.h include/ggml-kompute.h include/ggml-opt.h diff --git a/ggml/cmake/GitVars.cmake b/ggml/cmake/GitVars.cmake new file mode 100644 index 000000000..1a4c24ebf --- /dev/null +++ b/ggml/cmake/GitVars.cmake @@ -0,0 +1,22 @@ +find_package(Git) + +# the commit's SHA1 +execute_process(COMMAND + "${GIT_EXECUTABLE}" describe --match=NeVeRmAtCh --always --abbrev=8 + WORKING_DIRECTORY "${CMAKE_SOURCE_DIR}" + OUTPUT_VARIABLE GIT_SHA1 + ERROR_QUIET OUTPUT_STRIP_TRAILING_WHITESPACE) + +# the date of the commit +execute_process(COMMAND + "${GIT_EXECUTABLE}" log -1 --format=%ad --date=local + WORKING_DIRECTORY "${CMAKE_SOURCE_DIR}" + OUTPUT_VARIABLE GIT_DATE + ERROR_QUIET OUTPUT_STRIP_TRAILING_WHITESPACE) + +# the subject of the commit +execute_process(COMMAND + "${GIT_EXECUTABLE}" log -1 --format=%s + WORKING_DIRECTORY "${CMAKE_SOURCE_DIR}" + OUTPUT_VARIABLE GIT_COMMIT_SUBJECT + ERROR_QUIET OUTPUT_STRIP_TRAILING_WHITESPACE) diff --git a/ggml/cmake/common.cmake b/ggml/cmake/common.cmake new file mode 100644 index 000000000..1976d0ae9 --- /dev/null +++ b/ggml/cmake/common.cmake @@ -0,0 +1,26 @@ +function(ggml_get_flags CCID CCVER) + set(C_FLAGS "") + set(CXX_FLAGS "") + + if (CCID MATCHES "Clang") + set(C_FLAGS -Wunreachable-code-break -Wunreachable-code-return) + set(CXX_FLAGS -Wunreachable-code-break -Wunreachable-code-return -Wmissing-prototypes -Wextra-semi) + + if ( + (CCID STREQUAL "Clang" AND CCVER VERSION_GREATER_EQUAL 3.8.0) OR + (CCID STREQUAL "AppleClang" AND CCVER VERSION_GREATER_EQUAL 7.3.0) + ) + list(APPEND C_FLAGS -Wdouble-promotion) + endif() + elseif (CCID STREQUAL "GNU") + set(C_FLAGS -Wdouble-promotion) + set(CXX_FLAGS -Wno-array-bounds) + + if (CCVER VERSION_GREATER_EQUAL 8.1.0) + list(APPEND CXX_FLAGS -Wextra-semi) + endif() + endif() + + set(GF_C_FLAGS ${C_FLAGS} PARENT_SCOPE) + set(GF_CXX_FLAGS ${CXX_FLAGS} PARENT_SCOPE) +endfunction() diff --git a/ggml/cmake/ggml-config.cmake.in b/ggml/cmake/ggml-config.cmake.in index bf39f9c00..8c2dc31c6 100644 --- a/ggml/cmake/ggml-config.cmake.in +++ b/ggml/cmake/ggml-config.cmake.in @@ -5,7 +5,7 @@ set_and_check(GGML_INCLUDE_DIR "@PACKAGE_GGML_INCLUDE_INSTALL_DIR@") set_and_check(GGML_LIB_DIR "@PACKAGE_GGML_LIB_INSTALL_DIR@") -set_and_check(GGML_BIN_DIR "@PACKAGE_GGML_BIN_INSTALL_DIR@") +#set_and_check(GGML_BIN_DIR "@PACKAGE_GGML_BIN_INSTALL_DIR@") find_package(Threads REQUIRED) @@ -112,7 +112,7 @@ foreach(_ggml_backend ${GGML_AVAILABLE_BACKENDS}) string(REGEX MATCH "^ggml-cpu" is_cpu_variant "${_ggml_backend}") if(is_cpu_variant) - list(APPEND GGML_CPU_INTERFACE_LINK_LIBRARIES "ggml::ggml" "ggml::ggml-base") + list(APPEND GGML_CPU_INTERFACE_LINK_LIBRARIES "ggml::ggml-base") set_target_properties(ggml::${_ggml_backend} PROPERTIES INTERFACE_LINK_LIBRARIES "${GGML_CPU_INTERFACE_LINK_LIBRARIES}") @@ -124,7 +124,7 @@ foreach(_ggml_backend ${GGML_AVAILABLE_BACKENDS}) endif() else() - list(APPEND ${_ggml_backend_pfx}_INTERFACE_LINK_LIBRARIES "ggml::ggml" "ggml::ggml-base") + list(APPEND ${_ggml_backend_pfx}_INTERFACE_LINK_LIBRARIES "ggml::ggml-base") set_target_properties(ggml::${_ggml_backend} PROPERTIES INTERFACE_LINK_LIBRARIES "${${_ggml_backend_pfx}_INTERFACE_LINK_LIBRARIES}") @@ -139,6 +139,11 @@ foreach(_ggml_backend ${GGML_AVAILABLE_BACKENDS}) list(APPEND _ggml_all_targets ggml::${_ggml_backend}) endforeach() +list(APPEND GGML_INTERFACE_LINK_LIBRARIES ggml::ggml-base "${_ggml_all_targets}") +set_target_properties(ggml::ggml + PROPERTIES + INTERFACE_LINK_LIBRARIES "${GGML_INTERFACE_LINK_LIBRARIES}") + add_library(ggml::all INTERFACE IMPORTED) set_target_properties(ggml::all PROPERTIES diff --git a/ggml/include/ggml-alloc.h b/ggml/include/ggml-alloc.h index 23600eea9..2cb150fd2 100644 --- a/ggml/include/ggml-alloc.h +++ b/ggml/include/ggml-alloc.h @@ -19,7 +19,7 @@ struct ggml_tallocr { }; GGML_API struct ggml_tallocr ggml_tallocr_new(ggml_backend_buffer_t buffer); -GGML_API void ggml_tallocr_alloc(struct ggml_tallocr * talloc, struct ggml_tensor * tensor); +GGML_API enum ggml_status ggml_tallocr_alloc(struct ggml_tallocr * talloc, struct ggml_tensor * tensor); // Graph allocator /* diff --git a/ggml/include/ggml-backend.h b/ggml/include/ggml-backend.h index fc9571c82..64671495b 100644 --- a/ggml/include/ggml-backend.h +++ b/ggml/include/ggml-backend.h @@ -56,7 +56,7 @@ extern "C" { GGML_API void ggml_backend_buffer_free (ggml_backend_buffer_t buffer); GGML_API void * ggml_backend_buffer_get_base (ggml_backend_buffer_t buffer); GGML_API size_t ggml_backend_buffer_get_size (ggml_backend_buffer_t buffer); - GGML_API void ggml_backend_buffer_init_tensor (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor); + GGML_API enum ggml_status ggml_backend_buffer_init_tensor (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor); GGML_API size_t ggml_backend_buffer_get_alignment (ggml_backend_buffer_t buffer); GGML_API size_t ggml_backend_buffer_get_max_size (ggml_backend_buffer_t buffer); GGML_API size_t ggml_backend_buffer_get_alloc_size(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor); @@ -342,8 +342,8 @@ extern "C" { GGML_API bool ggml_backend_compare_graph_backend(ggml_backend_t backend1, ggml_backend_t backend2, struct ggml_cgraph * graph, ggml_backend_eval_callback callback, void * user_data); // Tensor initialization - GGML_API void ggml_backend_tensor_alloc(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, void * addr); - GGML_API void ggml_backend_view_init(struct ggml_tensor * tensor); + GGML_API enum ggml_status ggml_backend_tensor_alloc(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, void * addr); + GGML_API enum ggml_status ggml_backend_view_init(struct ggml_tensor * tensor); // CPU buffer types are always available GGML_API ggml_backend_buffer_t ggml_backend_cpu_buffer_from_ptr(void * ptr, size_t size); diff --git a/ggml/include/ggml-cpu.h b/ggml/include/ggml-cpu.h index 3aa71badb..f5e11f1e1 100644 --- a/ggml/include/ggml-cpu.h +++ b/ggml/include/ggml-cpu.h @@ -8,7 +8,7 @@ extern "C" { #endif // the compute plan that needs to be prepared for ggml_graph_compute() - // since https://github.com/ggerganov/ggml/issues/287 + // since https://github.com/ggml-org/ggml/issues/287 struct ggml_cplan { size_t work_size; // size of work buffer, calculated by `ggml_graph_plan()` uint8_t * work_data; // work buffer, to be allocated by caller before calling to `ggml_graph_compute()` @@ -80,6 +80,7 @@ extern "C" { GGML_BACKEND_API int ggml_cpu_has_avx (void); GGML_BACKEND_API int ggml_cpu_has_avx_vnni (void); GGML_BACKEND_API int ggml_cpu_has_avx2 (void); + GGML_BACKEND_API int ggml_cpu_has_bmi2 (void); GGML_BACKEND_API int ggml_cpu_has_f16c (void); GGML_BACKEND_API int ggml_cpu_has_fma (void); GGML_BACKEND_API int ggml_cpu_has_avx512 (void); @@ -95,9 +96,11 @@ extern "C" { GGML_BACKEND_API int ggml_cpu_has_matmul_int8(void); GGML_BACKEND_API int ggml_cpu_has_sve (void); GGML_BACKEND_API int ggml_cpu_get_sve_cnt (void); // sve vector length in bytes + GGML_BACKEND_API int ggml_cpu_has_sme (void); // other GGML_BACKEND_API int ggml_cpu_has_riscv_v (void); GGML_BACKEND_API int ggml_cpu_has_vsx (void); + GGML_BACKEND_API int ggml_cpu_has_vxe (void); GGML_BACKEND_API int ggml_cpu_has_wasm_simd (void); GGML_BACKEND_API int ggml_cpu_has_llamafile (void); diff --git a/ggml/include/ggml-metal.h b/ggml/include/ggml-metal.h index 669c1f84a..a61069442 100644 --- a/ggml/include/ggml-metal.h +++ b/ggml/include/ggml-metal.h @@ -45,7 +45,7 @@ GGML_BACKEND_API bool ggml_backend_is_metal(ggml_backend_t backend); GGML_DEPRECATED( GGML_BACKEND_API ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data, size_t size, size_t max_size), - "obsoleted by the new device interface - https://github.com/ggerganov/llama.cpp/pull/9713"); + "obsoleted by the new device interface - https://github.com/ggml-org/llama.cpp/pull/9713"); GGML_BACKEND_API void ggml_backend_metal_set_abort_callback(ggml_backend_t backend, ggml_abort_callback abort_callback, void * user_data); diff --git a/ggml/include/ggml-rpc.h b/ggml/include/ggml-rpc.h index ade6c3b0e..4e0d210f8 100644 --- a/ggml/include/ggml-rpc.h +++ b/ggml/include/ggml-rpc.h @@ -17,7 +17,9 @@ GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_rpc_buffer_type(const c GGML_BACKEND_API void ggml_backend_rpc_get_device_memory(const char * endpoint, size_t * free, size_t * total); -GGML_BACKEND_API void ggml_backend_rpc_start_server(ggml_backend_t backend, const char * endpoint, size_t free_mem, size_t total_mem); +GGML_BACKEND_API void ggml_backend_rpc_start_server(ggml_backend_t backend, const char * endpoint, + const char * cache_dir, + size_t free_mem, size_t total_mem); GGML_BACKEND_API ggml_backend_reg_t ggml_backend_rpc_reg(void); diff --git a/ggml/include/ggml-vulkan.h b/ggml/include/ggml-vulkan.h index 53cdba072..ed5ea5f79 100644 --- a/ggml/include/ggml-vulkan.h +++ b/ggml/include/ggml-vulkan.h @@ -10,8 +10,6 @@ extern "C" { #define GGML_VK_NAME "Vulkan" #define GGML_VK_MAX_DEVICES 16 -GGML_BACKEND_API void ggml_vk_instance_init(void); - // backend API GGML_BACKEND_API ggml_backend_t ggml_backend_vk_init(size_t dev_num); diff --git a/ggml/include/ggml.h b/ggml/include/ggml.h index 5bd8d9c8b..8fcc16df9 100644 --- a/ggml/include/ggml.h +++ b/ggml/include/ggml.h @@ -198,7 +198,7 @@ #ifndef __GNUC__ # define GGML_ATTRIBUTE_FORMAT(...) -#elif defined(__MINGW32__) +#elif defined(__MINGW32__) && !defined(__clang__) # define GGML_ATTRIBUTE_FORMAT(...) __attribute__((format(gnu_printf, __VA_ARGS__))) #else # define GGML_ATTRIBUTE_FORMAT(...) __attribute__((format(printf, __VA_ARGS__))) @@ -454,6 +454,7 @@ extern "C" { GGML_OP_RMS_NORM, GGML_OP_RMS_NORM_BACK, GGML_OP_GROUP_NORM, + GGML_OP_L2_NORM, GGML_OP_MUL_MAT, GGML_OP_MUL_MAT_ID, @@ -502,20 +503,16 @@ extern "C" { GGML_OP_ADD_REL_POS, GGML_OP_RWKV_WKV6, GGML_OP_GATED_LINEAR_ATTN, + GGML_OP_RWKV_WKV7, GGML_OP_UNARY, - GGML_OP_MAP_UNARY, - GGML_OP_MAP_BINARY, - - GGML_OP_MAP_CUSTOM1_F32, - GGML_OP_MAP_CUSTOM2_F32, - GGML_OP_MAP_CUSTOM3_F32, - GGML_OP_MAP_CUSTOM1, GGML_OP_MAP_CUSTOM2, GGML_OP_MAP_CUSTOM3, + GGML_OP_CUSTOM, + GGML_OP_CROSS_ENTROPY_LOSS, GGML_OP_CROSS_ENTROPY_LOSS_BACK, GGML_OP_OPT_STEP_ADAMW, @@ -1095,6 +1092,18 @@ extern "C" { int n_groups, float eps); + // l2 normalize along rows + // used in rwkv v7 + GGML_API struct ggml_tensor * ggml_l2_norm( + struct ggml_context * ctx, + struct ggml_tensor * a, + float eps); + + GGML_API struct ggml_tensor * ggml_l2_norm_inplace( + struct ggml_context * ctx, + struct ggml_tensor * a, + float eps); + // a - x // b - dy GGML_API struct ggml_tensor * ggml_rms_norm_back( @@ -1708,24 +1717,29 @@ extern "C" { float p0, float p1); - // nearest interpolate + enum ggml_scale_mode { + GGML_SCALE_MODE_NEAREST = 0, + GGML_SCALE_MODE_BILINEAR = 1, + }; + + // interpolate // multiplies ne0 and ne1 by scale factor - // used in stable-diffusion GGML_API struct ggml_tensor * ggml_upscale( struct ggml_context * ctx, struct ggml_tensor * a, - int scale_factor); + int scale_factor, + enum ggml_scale_mode mode); - // nearest interpolate - // nearest interpolate to specified dimensions - // used in tortoise.cpp + // interpolate + // interpolate scale to specified dimensions GGML_API struct ggml_tensor * ggml_upscale_ext( struct ggml_context * ctx, struct ggml_tensor * a, int ne0, int ne1, int ne2, - int ne3); + int ne3, + enum ggml_scale_mode mode); // pad each dimension with zeros: [x, ..., x] -> [x, ..., x, 0, ..., 0] GGML_API struct ggml_tensor * ggml_pad( @@ -1777,11 +1791,11 @@ extern "C" { #define GGML_KQ_MASK_PAD 64 - // q: [n_embd, n_batch, n_head, 1] - // k: [n_embd, n_kv, n_head_kv, 1] - // v: [n_embd, n_kv, n_head_kv, 1] !! not transposed !! - // mask: [n_kv, n_batch_pad, 1, 1] !! n_batch_pad = GGML_PAD(n_batch, GGML_KQ_MASK_PAD) !! - // res: [n_embd, n_head, n_batch, 1] !! permuted !! + // q: [n_embd_k, n_batch, n_head, 1] + // k: [n_embd_k, n_kv, n_head_kv, 1] + // v: [n_embd_v, n_kv, n_head_kv, 1] !! not transposed !! + // mask: [n_kv, n_batch_pad, 1, 1] !! n_batch_pad = GGML_PAD(n_batch, GGML_KQ_MASK_PAD) !! + // res: [n_embd_v, n_head, n_batch, 1] !! permuted !! GGML_API struct ggml_tensor * ggml_flash_attn_ext( struct ggml_context * ctx, struct ggml_tensor * q, @@ -1890,85 +1904,18 @@ extern "C" { struct ggml_tensor * state, float scale); + GGML_API struct ggml_tensor * ggml_rwkv_wkv7( + struct ggml_context * ctx, + struct ggml_tensor * r, + struct ggml_tensor * w, + struct ggml_tensor * k, + struct ggml_tensor * v, + struct ggml_tensor * a, + struct ggml_tensor * b, + struct ggml_tensor * state); + // custom operators - typedef void (*ggml_unary_op_f32_t) (const int, float *, const float *); - typedef void (*ggml_binary_op_f32_t)(const int, float *, const float *, const float *); - - typedef void (*ggml_custom1_op_f32_t)(struct ggml_tensor *, const struct ggml_tensor *); - typedef void (*ggml_custom2_op_f32_t)(struct ggml_tensor *, const struct ggml_tensor *, const struct ggml_tensor *); - typedef void (*ggml_custom3_op_f32_t)(struct ggml_tensor *, const struct ggml_tensor *, const struct ggml_tensor *, const struct ggml_tensor *); - - GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_unary_f32( - struct ggml_context * ctx, - struct ggml_tensor * a, - ggml_unary_op_f32_t fun), - "use ggml_map_custom1 instead"); - - GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_unary_inplace_f32( - struct ggml_context * ctx, - struct ggml_tensor * a, - ggml_unary_op_f32_t fun), - "use ggml_map_custom1_inplace instead"); - - GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_binary_f32( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b, - ggml_binary_op_f32_t fun), - "use ggml_map_custom2 instead"); - - GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_binary_inplace_f32( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b, - ggml_binary_op_f32_t fun), - "use ggml_map_custom2_inplace instead"); - - GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_custom1_f32( - struct ggml_context * ctx, - struct ggml_tensor * a, - ggml_custom1_op_f32_t fun), - "use ggml_map_custom1 instead"); - - GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_custom1_inplace_f32( - struct ggml_context * ctx, - struct ggml_tensor * a, - ggml_custom1_op_f32_t fun), - "use ggml_map_custom1_inplace instead"); - - GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_custom2_f32( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b, - ggml_custom2_op_f32_t fun), - "use ggml_map_custom2 instead"); - - GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_custom2_inplace_f32( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b, - ggml_custom2_op_f32_t fun), - "use ggml_map_custom2_inplace instead"); - - GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_custom3_f32( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b, - struct ggml_tensor * c, - ggml_custom3_op_f32_t fun), - "use ggml_map_custom3 instead"); - - GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_custom3_inplace_f32( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b, - struct ggml_tensor * c, - ggml_custom3_op_f32_t fun), - "use ggml_map_custom3_inplace instead"); - - // custom operators v2 - typedef void (*ggml_custom1_op_t)(struct ggml_tensor * dst , const struct ggml_tensor * a, int ith, int nth, void * userdata); typedef void (*ggml_custom2_op_t)(struct ggml_tensor * dst , const struct ggml_tensor * a, const struct ggml_tensor * b, int ith, int nth, void * userdata); typedef void (*ggml_custom3_op_t)(struct ggml_tensor * dst , const struct ggml_tensor * a, const struct ggml_tensor * b, const struct ggml_tensor * c, int ith, int nth, void * userdata); @@ -2024,6 +1971,30 @@ extern "C" { int n_tasks, void * userdata); + typedef void (*ggml_custom_op_t)(struct ggml_tensor * dst , int ith, int nth, void * userdata); + + GGML_API struct ggml_tensor * ggml_custom_4d( + struct ggml_context * ctx, + enum ggml_type type, + int64_t ne0, + int64_t ne1, + int64_t ne2, + int64_t ne3, + struct ggml_tensor ** args, + int n_args, + ggml_custom_op_t fun, + int n_tasks, + void * userdata); + + GGML_API struct ggml_tensor * ggml_custom_inplace( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor ** args, + int n_args, + ggml_custom_op_t fun, + int n_tasks, + void * userdata); + // loss function GGML_API struct ggml_tensor * ggml_cross_entropy_loss( @@ -2140,7 +2111,11 @@ extern "C" { # define GGML_RESTRICT # endif #else -# define GGML_RESTRICT restrict +# if defined (_MSC_VER) && (__STDC_VERSION__ < 201112L) +# define GGML_RESTRICT __restrict +# else +# define GGML_RESTRICT restrict +# endif #endif typedef void (*ggml_to_float_t) (const void * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k); typedef void (*ggml_from_float_t)(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k); diff --git a/ggml/src/CMakeLists.txt b/ggml/src/CMakeLists.txt index 0002ac18a..f00700da7 100644 --- a/ggml/src/CMakeLists.txt +++ b/ggml/src/CMakeLists.txt @@ -1,4 +1,5 @@ include(CheckCXXCompilerFlag) +include("../cmake/common.cmake") add_compile_definitions(GGML_SCHED_MAX_COPIES=${GGML_SCHED_MAX_COPIES}) @@ -24,33 +25,6 @@ if (NOT MSVC) endif() endif() -function(ggml_get_flags CCID CCVER) - set(C_FLAGS "") - set(CXX_FLAGS "") - - if (CCID MATCHES "Clang") - set(C_FLAGS -Wunreachable-code-break -Wunreachable-code-return) - set(CXX_FLAGS -Wunreachable-code-break -Wunreachable-code-return -Wmissing-prototypes -Wextra-semi) - - if ( - (CCID STREQUAL "Clang" AND CCVER VERSION_GREATER_EQUAL 3.8.0) OR - (CCID STREQUAL "AppleClang" AND CCVER VERSION_GREATER_EQUAL 7.3.0) - ) - list(APPEND C_FLAGS -Wdouble-promotion) - endif() - elseif (CCID STREQUAL "GNU") - set(C_FLAGS -Wdouble-promotion) - set(CXX_FLAGS -Wno-array-bounds) - - if (CCVER VERSION_GREATER_EQUAL 8.1.0) - list(APPEND CXX_FLAGS -Wextra-semi) - endif() - endif() - - set(GF_C_FLAGS ${C_FLAGS} PARENT_SCOPE) - set(GF_CXX_FLAGS ${CXX_FLAGS} PARENT_SCOPE) -endfunction() - if (GGML_FATAL_WARNINGS) if (CMAKE_CXX_COMPILER_ID MATCHES "GNU" OR CMAKE_CXX_COMPILER_ID MATCHES "Clang") list(APPEND C_FLAGS -Werror) @@ -91,7 +65,7 @@ if (GGML_LTO) endif() endif() -if (GGML_CCACHE) +if (GGML_CCACHE AND NOT CMAKE_C_COMPILER_LAUNCHER AND NOT CMAKE_CXX_COMPILER_LAUNCHER) find_program(GGML_CCACHE_FOUND ccache) find_program(GGML_SCCACHE_FOUND sccache) @@ -102,7 +76,11 @@ if (GGML_CCACHE) set(GGML_CCACHE_VARIANT sccache) endif() # TODO: should not be set globally - set_property(GLOBAL PROPERTY RULE_LAUNCH_COMPILE "${GGML_CCACHE_VARIANT}") + if (GGML_SYCL AND GGML_CCACHE_FOUND AND WIN32) + set_property(GLOBAL PROPERTY RULE_LAUNCH_COMPILE "ccache compiler_type=icl") + else () + set_property(GLOBAL PROPERTY RULE_LAUNCH_COMPILE "${GGML_CCACHE_VARIANT}") + endif () set(ENV{CCACHE_SLOPPINESS} time_macros) message(STATUS "${GGML_CCACHE_VARIANT} found, compilation results will be cached. Disable with GGML_CCACHE=OFF.") else() @@ -226,6 +204,9 @@ add_library(ggml-base gguf.cpp) target_include_directories(ggml-base PRIVATE .) +if (GGML_BACKEND_DL) + target_compile_definitions(ggml-base PUBLIC GGML_BACKEND_DL) +endif() add_library(ggml ggml-backend-reg.cpp) @@ -233,7 +214,7 @@ add_library(ggml target_link_libraries(ggml PUBLIC ggml-base) if (CMAKE_SYSTEM_NAME MATCHES "Linux") - target_link_libraries(ggml PRIVATE dl) + target_link_libraries(ggml PRIVATE dl stdc++fs) endif() function(ggml_add_backend_library backend) @@ -286,7 +267,7 @@ function(ggml_add_cpu_backend_variant tag_name) set(GGML_CPU_TAG_NAME ${tag_name}) # other: OPENMP LLAMAFILE CPU_HBM foreach (feat NATIVE - AVX AVX2 AVX_VNNI FMA F16C + AVX AVX2 BMI2 AVX_VNNI FMA F16C AVX512 AVX512_VBMI AVX512_VNNI AVX512_BF16 AMX_TILE AMX_INT8 AMX_BF16) set(GGML_${feat} OFF) @@ -306,13 +287,13 @@ if (GGML_CPU_ALL_VARIANTS) message(FATAL_ERROR "GGML_CPU_ALL_VARIANTS requires GGML_BACKEND_DL") endif() ggml_add_cpu_backend_variant(sandybridge AVX) - ggml_add_cpu_backend_variant(haswell AVX F16C AVX2 FMA) - ggml_add_cpu_backend_variant(skylakex AVX F16C AVX2 FMA AVX512) - ggml_add_cpu_backend_variant(icelake AVX F16C AVX2 FMA AVX512 AVX512_VBMI AVX512_VNNI) - ggml_add_cpu_backend_variant(alderlake AVX F16C AVX2 FMA AVX_VNNI) + ggml_add_cpu_backend_variant(haswell AVX F16C AVX2 BMI2 FMA) + ggml_add_cpu_backend_variant(skylakex AVX F16C AVX2 BMI2 FMA AVX512) + ggml_add_cpu_backend_variant(icelake AVX F16C AVX2 BMI2 FMA AVX512 AVX512_VBMI AVX512_VNNI) + ggml_add_cpu_backend_variant(alderlake AVX F16C AVX2 BMI2 FMA AVX_VNNI) if (NOT MSVC) # MSVC doesn't support AMX - ggml_add_cpu_backend_variant(sapphirerapids AVX F16C AVX2 FMA AVX512 AVX512_VBMI AVX512_VNNI AVX512_BF16 AMX_TILE AMX_INT8) + ggml_add_cpu_backend_variant(sapphirerapids AVX F16C AVX2 BMI2 FMA AVX512 AVX512_VBMI AVX512_VNNI AVX512_BF16 AMX_TILE AMX_INT8) endif() elseif (GGML_CPU) ggml_add_cpu_backend_variant_impl("") @@ -348,6 +329,10 @@ if (CMAKE_SYSTEM_NAME MATCHES "Android") target_link_libraries(ggml-base PRIVATE dl) endif() +if(CMAKE_SYSTEM_NAME MATCHES "visionOS") + target_compile_definitions(ggml-base PUBLIC _DARWIN_C_SOURCE) +endif() + if (BUILD_SHARED_LIBS) foreach (target ggml-base ggml) set_target_properties(${target} PROPERTIES POSITION_INDEPENDENT_CODE ON) diff --git a/ggml/src/ggml-alloc.c b/ggml/src/ggml-alloc.c index 7244a9cbb..a3d3f6901 100644 --- a/ggml/src/ggml-alloc.c +++ b/ggml/src/ggml-alloc.c @@ -89,7 +89,7 @@ struct ggml_tallocr ggml_tallocr_new(ggml_backend_buffer_t buffer) { return talloc; } -void ggml_tallocr_alloc(struct ggml_tallocr * talloc, struct ggml_tensor * tensor) { +enum ggml_status ggml_tallocr_alloc(struct ggml_tallocr * talloc, struct ggml_tensor * tensor) { size_t size = ggml_backend_buffer_get_alloc_size(talloc->buffer, tensor); size = GGML_PAD(size, talloc->alignment); @@ -104,7 +104,7 @@ void ggml_tallocr_alloc(struct ggml_tallocr * talloc, struct ggml_tensor * tenso assert(((uintptr_t)addr % talloc->alignment) == 0); - ggml_backend_tensor_alloc(talloc->buffer, tensor, addr); + return ggml_backend_tensor_alloc(talloc->buffer, tensor, addr); } // dynamic tensor allocator @@ -933,42 +933,51 @@ size_t ggml_gallocr_get_buffer_size(ggml_gallocr_t galloc, int buffer_id) { // utils +static void free_buffers(ggml_backend_buffer_t ** buffers, const size_t * n_buffers) { + for (size_t i = 0; i < *n_buffers; i++) { + ggml_backend_buffer_free((*buffers)[i]); + } + free(*buffers); +} + static bool alloc_tensor_range(struct ggml_context * ctx, struct ggml_tensor * first, struct ggml_tensor * last, ggml_backend_buffer_type_t buft, size_t size, ggml_backend_buffer_t ** buffers, size_t * n_buffers) { + ggml_backend_buffer_t buffer = ggml_backend_buft_alloc_buffer(buft, size); if (buffer == NULL) { -#ifndef NDEBUG - GGML_LOG_DEBUG("%s: failed to allocate %s buffer of size %zu\n", __func__, ggml_backend_buft_name(buft), size); -#endif - for (size_t i = 0; i < *n_buffers; i++) { - ggml_backend_buffer_free((*buffers)[i]); - } - free(*buffers); + GGML_LOG_ERROR("%s: failed to allocate %s buffer of size %zu\n", __func__, ggml_backend_buft_name(buft), size); + free_buffers(buffers, n_buffers); return false; } - struct ggml_tallocr tallocr = ggml_tallocr_new(buffer); - - for (struct ggml_tensor * t = first; t != last; t = ggml_get_next_tensor(ctx, t)) { - if (t->data == NULL) { - if (t->view_src == NULL) { - ggml_tallocr_alloc(&tallocr, t); - } else if (t->buffer == NULL) { - ggml_backend_view_init(t); - } - } else { - if (t->view_src != NULL && t->buffer == NULL) { - // view of a pre-allocated tensor - ggml_backend_view_init(t); - } - } - } - *buffers = realloc(*buffers, sizeof(ggml_backend_buffer_t) * (*n_buffers + 1)); (*buffers)[(*n_buffers)++] = buffer; + struct ggml_tallocr tallocr = ggml_tallocr_new(buffer); + + for (struct ggml_tensor * t = first; t != last; t = ggml_get_next_tensor(ctx, t)) { + enum ggml_status status = GGML_STATUS_SUCCESS; + if (t->data == NULL) { + if (t->view_src == NULL) { + status = ggml_tallocr_alloc(&tallocr, t); + } else if (t->buffer == NULL) { + status = ggml_backend_view_init(t); + } + } else { + if (t->view_src != NULL && t->buffer == NULL) { + // view of a pre-allocated tensor + status = ggml_backend_view_init(t); + } + } + if (status != GGML_STATUS_SUCCESS) { + GGML_LOG_ERROR("%s: failed to initialize tensor %s\n", __func__, t->name); + free_buffers(buffers, n_buffers); + return false; + } + } + return true; } diff --git a/ggml/src/ggml-backend-impl.h b/ggml/src/ggml-backend-impl.h index d1c2d76d8..c36c12d65 100644 --- a/ggml/src/ggml-backend-impl.h +++ b/ggml/src/ggml-backend-impl.h @@ -44,7 +44,7 @@ extern "C" { // base address of the buffer void * (*get_base) (ggml_backend_buffer_t buffer); // (optional) initialize a tensor in the buffer (eg. add tensor extras) - void (*init_tensor) (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor); + enum ggml_status (*init_tensor)(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor); // tensor data access void (*memset_tensor)(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, uint8_t value, size_t offset, size_t size); void (*set_tensor) (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size); diff --git a/ggml/src/ggml-backend-reg.cpp b/ggml/src/ggml-backend-reg.cpp index 955ed505f..405d8e315 100644 --- a/ggml/src/ggml-backend-reg.cpp +++ b/ggml/src/ggml-backend-reg.cpp @@ -2,14 +2,13 @@ #include "ggml-backend.h" #include "ggml-impl.h" #include -#include #include #include -#include #include #include #include #include +#include #ifdef _WIN32 # define WIN32_LEAN_AND_MEAN @@ -72,14 +71,22 @@ # pragma clang diagnostic ignored "-Wdeprecated-declarations" #endif -static std::wstring utf8_to_utf16(const std::string & str) { - std::wstring_convert> converter; - return converter.from_bytes(str); -} +namespace fs = std::filesystem; -static std::string utf16_to_utf8(const std::wstring & str) { - std::wstring_convert> converter; - return converter.to_bytes(str); +static std::string path_str(const fs::path & path) { + std::string u8path; + try { +#if defined(__cpp_lib_char8_t) + // C++20 and later: u8string() returns std::u8string + std::u8string u8str = path.u8string(); + u8path = std::string(reinterpret_cast(u8str.c_str())); +#else + // C++17: u8string() returns std::string + u8path = path.u8string(); +#endif + } catch (...) { + } + return u8path; } #if defined(__clang__) @@ -96,12 +103,12 @@ struct dl_handle_deleter { } }; -static dl_handle * dl_load_library(const std::wstring & path) { +static dl_handle * dl_load_library(const fs::path & path) { // suppress error dialogs for missing DLLs DWORD old_mode = SetErrorMode(SEM_FAILCRITICALERRORS); SetErrorMode(old_mode | SEM_FAILCRITICALERRORS); - HMODULE handle = LoadLibraryW(path.c_str()); + HMODULE handle = LoadLibraryW(path.wstring().c_str()); SetErrorMode(old_mode); @@ -129,8 +136,8 @@ struct dl_handle_deleter { } }; -static void * dl_load_library(const std::wstring & path) { - dl_handle * handle = dlopen(utf16_to_utf8(path).c_str(), RTLD_NOW | RTLD_LOCAL); +static void * dl_load_library(const fs::path & path) { + dl_handle * handle = dlopen(path.string().c_str(), RTLD_NOW | RTLD_LOCAL); return handle; } @@ -217,11 +224,11 @@ struct ggml_backend_registry { devices.push_back(device); } - ggml_backend_reg_t load_backend(const std::wstring & path, bool silent) { + ggml_backend_reg_t load_backend(const fs::path & path, bool silent) { dl_handle_ptr handle { dl_load_library(path) }; if (!handle) { if (!silent) { - GGML_LOG_ERROR("%s: failed to load %s\n", __func__, utf16_to_utf8(path).c_str()); + GGML_LOG_ERROR("%s: failed to load %s\n", __func__, path_str(path).c_str()); } return nullptr; } @@ -229,7 +236,7 @@ struct ggml_backend_registry { auto score_fn = (ggml_backend_score_t) dl_get_sym(handle.get(), "ggml_backend_score"); if (score_fn && score_fn() == 0) { if (!silent) { - GGML_LOG_INFO("%s: backend %s is not supported on this system\n", __func__, utf16_to_utf8(path).c_str()); + GGML_LOG_INFO("%s: backend %s is not supported on this system\n", __func__, path_str(path).c_str()); } return nullptr; } @@ -237,7 +244,7 @@ struct ggml_backend_registry { auto backend_init_fn = (ggml_backend_init_t) dl_get_sym(handle.get(), "ggml_backend_init"); if (!backend_init_fn) { if (!silent) { - GGML_LOG_ERROR("%s: failed to find ggml_backend_init in %s\n", __func__, utf16_to_utf8(path).c_str()); + GGML_LOG_ERROR("%s: failed to find ggml_backend_init in %s\n", __func__, path_str(path).c_str()); } return nullptr; } @@ -246,16 +253,17 @@ struct ggml_backend_registry { if (!reg || reg->api_version != GGML_BACKEND_API_VERSION) { if (!silent) { if (!reg) { - GGML_LOG_ERROR("%s: failed to initialize backend from %s: ggml_backend_init returned NULL\n", __func__, utf16_to_utf8(path).c_str()); + GGML_LOG_ERROR("%s: failed to initialize backend from %s: ggml_backend_init returned NULL\n", + __func__, path_str(path).c_str()); } else { GGML_LOG_ERROR("%s: failed to initialize backend from %s: incompatible API version (backend: %d, current: %d)\n", - __func__, utf16_to_utf8(path).c_str(), reg->api_version, GGML_BACKEND_API_VERSION); + __func__, path_str(path).c_str(), reg->api_version, GGML_BACKEND_API_VERSION); } } return nullptr; } - GGML_LOG_INFO("%s: loaded %s backend from %s\n", __func__, ggml_backend_reg_name(reg), utf16_to_utf8(path).c_str()); + GGML_LOG_INFO("%s: loaded %s backend from %s\n", __func__, ggml_backend_reg_name(reg), path_str(path).c_str()); register_backend(reg, std::move(handle)); @@ -391,14 +399,14 @@ ggml_backend_t ggml_backend_init_best(void) { // Dynamic loading ggml_backend_reg_t ggml_backend_load(const char * path) { - return get_reg().load_backend(utf8_to_utf16(path), false); + return get_reg().load_backend(path, false); } void ggml_backend_unload(ggml_backend_reg_t reg) { get_reg().unload_backend(reg, true); } -static std::wstring get_executable_path() { +static fs::path get_executable_path() { #if defined(__APPLE__) // get executable path std::vector path; @@ -416,7 +424,7 @@ static std::wstring get_executable_path() { if (last_slash != std::string::npos) { base_path = base_path.substr(0, last_slash); } - return utf8_to_utf16(base_path + "/"); + return base_path + "/"; #elif defined(__linux__) || defined(__FreeBSD__) std::string base_path = "."; std::vector path(1024); @@ -442,7 +450,7 @@ static std::wstring get_executable_path() { path.resize(path.size() * 2); } - return utf8_to_utf16(base_path + "/"); + return base_path + "/"; #elif defined(_WIN32) std::vector path(MAX_PATH); DWORD len = GetModuleFileNameW(NULL, path.data(), path.size()); @@ -461,74 +469,69 @@ static std::wstring get_executable_path() { #endif } -static std::wstring backend_filename_prefix() { +static fs::path backend_filename_prefix() { #ifdef _WIN32 - return L"ggml-"; + return fs::u8path("ggml-"); #else - return L"libggml-"; + return fs::u8path("libggml-"); #endif } -static std::wstring backend_filename_suffix() { +static fs::path backend_filename_extension() { #ifdef _WIN32 - return L".dll"; + return fs::u8path(".dll"); #else - return L".so"; -#endif -} - -static std::wstring path_separator() { -#ifdef _WIN32 - return L"\\"; -#else - return L"/"; + return fs::u8path(".so"); #endif } static ggml_backend_reg_t ggml_backend_load_best(const char * name, bool silent, const char * user_search_path) { // enumerate all the files that match [lib]ggml-name-*.[so|dll] in the search paths - // TODO: search system paths - std::wstring file_prefix = backend_filename_prefix() + utf8_to_utf16(name) + L"-"; - std::vector search_paths; + const fs::path name_path = fs::u8path(name); + const fs::path file_prefix = backend_filename_prefix().native() + name_path.native() + fs::u8path("-").native(); + const fs::path file_extension = backend_filename_extension(); + + std::vector search_paths; if (user_search_path == nullptr) { - search_paths.push_back(L"." + path_separator()); + // default search paths: executable directory, current directory search_paths.push_back(get_executable_path()); + search_paths.push_back(fs::current_path()); } else { - search_paths.push_back(utf8_to_utf16(user_search_path) + path_separator()); + search_paths.push_back(fs::u8path(user_search_path)); } int best_score = 0; - std::wstring best_path; + fs::path best_path; - namespace fs = std::filesystem; for (const auto & search_path : search_paths) { if (!fs::exists(search_path)) { + GGML_LOG_DEBUG("%s: search path %s does not exist\n", __func__, path_str(search_path).c_str()); continue; } fs::directory_iterator dir_it(search_path, fs::directory_options::skip_permission_denied); for (const auto & entry : dir_it) { if (entry.is_regular_file()) { - std::wstring filename = entry.path().filename().wstring(); - std::wstring ext = entry.path().extension().wstring(); - if (filename.find(file_prefix) == 0 && ext == backend_filename_suffix()) { - dl_handle_ptr handle { dl_load_library(entry.path().wstring()) }; + auto filename = entry.path().filename(); + auto ext = entry.path().extension(); + if (filename.native().find(file_prefix) == 0 && ext == file_extension) { + dl_handle_ptr handle { dl_load_library(entry) }; if (!handle && !silent) { - GGML_LOG_ERROR("%s: failed to load %s\n", __func__, utf16_to_utf8(entry.path().wstring()).c_str()); + GGML_LOG_ERROR("%s: failed to load %s\n", __func__, path_str(entry.path()).c_str()); } if (handle) { auto score_fn = (ggml_backend_score_t) dl_get_sym(handle.get(), "ggml_backend_score"); if (score_fn) { int s = score_fn(); #ifndef NDEBUG - GGML_LOG_DEBUG("%s: %s score: %d\n", __func__, utf16_to_utf8(entry.path().wstring()).c_str(), s); + GGML_LOG_DEBUG("%s: %s score: %d\n", __func__, path_str(entry.path()).c_str(), s); #endif if (s > best_score) { best_score = s; - best_path = entry.path().wstring(); + best_path = entry.path(); } } else { if (!silent) { - GGML_LOG_INFO("%s: failed to find ggml_backend_score in %s\n", __func__, utf16_to_utf8(entry.path().wstring()).c_str()); + GGML_LOG_INFO("%s: failed to find ggml_backend_score in %s\n", __func__, path_str(entry.path()).c_str()); } } } @@ -540,7 +543,8 @@ static ggml_backend_reg_t ggml_backend_load_best(const char * name, bool silent, if (best_score == 0) { // try to load the base backend for (const auto & search_path : search_paths) { - std::wstring path = search_path + backend_filename_prefix() + utf8_to_utf16(name) + backend_filename_suffix(); + fs::path filename = backend_filename_prefix().native() + name_path.native() + backend_filename_extension().native(); + fs::path path = search_path / filename; if (fs::exists(path)) { return get_reg().load_backend(path, silent); } diff --git a/ggml/src/ggml-backend.cpp b/ggml/src/ggml-backend.cpp index dba7be33b..273075f4e 100644 --- a/ggml/src/ggml-backend.cpp +++ b/ggml/src/ggml-backend.cpp @@ -21,6 +21,7 @@ #include #include #include +#include #ifdef __APPLE__ #include @@ -126,11 +127,12 @@ void * ggml_backend_buffer_get_base(ggml_backend_buffer_t buffer) { return base; } -void ggml_backend_buffer_init_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) { +enum ggml_status ggml_backend_buffer_init_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) { // init_tensor is optional if (buffer->iface.init_tensor) { - buffer->iface.init_tensor(buffer, tensor); + return buffer->iface.init_tensor(buffer, tensor); } + return GGML_STATUS_SUCCESS; } void ggml_backend_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) { @@ -1641,7 +1643,7 @@ ggml_backend_t ggml_backend_sched_get_tensor_backend(ggml_backend_sched_t sched, // utils -void ggml_backend_view_init(struct ggml_tensor * tensor) { +enum ggml_status ggml_backend_view_init(struct ggml_tensor * tensor) { GGML_ASSERT(tensor->buffer == NULL); GGML_ASSERT(tensor->view_src != NULL); GGML_ASSERT(tensor->view_src->buffer != NULL); @@ -1649,10 +1651,10 @@ void ggml_backend_view_init(struct ggml_tensor * tensor) { tensor->buffer = tensor->view_src->buffer; tensor->data = (char *)tensor->view_src->data + tensor->view_offs; - ggml_backend_buffer_init_tensor(tensor->buffer, tensor); + return ggml_backend_buffer_init_tensor(tensor->buffer, tensor); } -void ggml_backend_tensor_alloc(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, void * addr) { +enum ggml_status ggml_backend_tensor_alloc(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, void * addr) { GGML_ASSERT(tensor->buffer == NULL); GGML_ASSERT(tensor->data == NULL); GGML_ASSERT(tensor->view_src == NULL); @@ -1662,7 +1664,7 @@ void ggml_backend_tensor_alloc(ggml_backend_buffer_t buffer, struct ggml_tensor tensor->buffer = buffer; tensor->data = addr; - ggml_backend_buffer_init_tensor(buffer, tensor); + return ggml_backend_buffer_init_tensor(buffer, tensor); } static struct ggml_tensor * graph_copy_dup_tensor(struct ggml_hash_set hash_set, struct ggml_tensor ** node_copies, @@ -1708,7 +1710,8 @@ static void graph_copy_init_tensor(struct ggml_hash_set * hash_set, struct ggml_ struct ggml_tensor * dst = node_copies[id]; if (dst->view_src != NULL) { graph_copy_init_tensor(hash_set, node_copies, node_init, src->view_src); - ggml_backend_view_init(dst); + enum ggml_status status = ggml_backend_view_init(dst); + GGML_ASSERT(status == GGML_STATUS_SUCCESS); } else { ggml_backend_tensor_copy(src, dst); @@ -1823,7 +1826,6 @@ bool ggml_backend_compare_graph_backend(ggml_backend_t backend1, ggml_backend_t assert(g1->n_nodes == g2->n_nodes); for (int i = 0; i < g1->n_nodes; i++) { - //printf("eval %d/%d\n", i, g1->n_nodes); struct ggml_tensor * t1 = g1->nodes[i]; struct ggml_tensor * t2 = g2->nodes[i]; diff --git a/ggml/src/ggml-cann/.clang-format b/ggml/src/ggml-cann/.clang-format deleted file mode 100644 index 2ad03d739..000000000 --- a/ggml/src/ggml-cann/.clang-format +++ /dev/null @@ -1,168 +0,0 @@ ---- -Language: Cpp -# BasedOnStyle: Google -AccessModifierOffset: -1 -AlignAfterOpenBracket: Align -AlignConsecutiveMacros: false -AlignConsecutiveAssignments: false -AlignConsecutiveDeclarations: false -AlignEscapedNewlines: Left -AlignOperands: true -AlignTrailingComments: true -AllowAllArgumentsOnNextLine: true -AllowAllConstructorInitializersOnNextLine: true -AllowAllParametersOfDeclarationOnNextLine: true -AllowShortBlocksOnASingleLine: Never -AllowShortCaseLabelsOnASingleLine: false -AllowShortFunctionsOnASingleLine: All -AllowShortLambdasOnASingleLine: All -AllowShortIfStatementsOnASingleLine: WithoutElse -AllowShortLoopsOnASingleLine: true -AlwaysBreakAfterDefinitionReturnType: None -AlwaysBreakAfterReturnType: None -AlwaysBreakBeforeMultilineStrings: true -AlwaysBreakTemplateDeclarations: Yes -BinPackArguments: true -BinPackParameters: true -BraceWrapping: - AfterCaseLabel: false - AfterClass: false - AfterControlStatement: false - AfterEnum: false - AfterFunction: false - AfterNamespace: false - AfterObjCDeclaration: false - AfterStruct: false - AfterUnion: false - AfterExternBlock: false - BeforeCatch: false - BeforeElse: false - IndentBraces: false - SplitEmptyFunction: true - SplitEmptyRecord: true - SplitEmptyNamespace: true -BreakBeforeBinaryOperators: None -BreakBeforeBraces: Attach -BreakBeforeInheritanceComma: false -BreakInheritanceList: BeforeColon -BreakBeforeTernaryOperators: true -BreakConstructorInitializersBeforeComma: false -BreakConstructorInitializers: BeforeColon -BreakAfterJavaFieldAnnotations: false -BreakStringLiterals: true -ColumnLimit: 80 -CommentPragmas: '^ IWYU pragma:' -CompactNamespaces: false -ConstructorInitializerAllOnOneLineOrOnePerLine: true -ConstructorInitializerIndentWidth: 4 -ContinuationIndentWidth: 4 -Cpp11BracedListStyle: true -DeriveLineEnding: true -DerivePointerAlignment: true -DisableFormat: false -ExperimentalAutoDetectBinPacking: false -FixNamespaceComments: true -ForEachMacros: - - foreach - - Q_FOREACH - - BOOST_FOREACH -IncludeBlocks: Regroup -IncludeCategories: - - Regex: '^' - Priority: 2 - SortPriority: 0 - - Regex: '^<.*\.h>' - Priority: 1 - SortPriority: 0 - - Regex: '^<.*' - Priority: 2 - SortPriority: 0 - - Regex: '.*' - Priority: 3 - SortPriority: 0 -IncludeIsMainRegex: '([-_](test|unittest))?$' -IncludeIsMainSourceRegex: '' -IndentCaseLabels: true -IndentGotoLabels: true -IndentPPDirectives: None -IndentWidth: 4 -IndentWrappedFunctionNames: false -JavaScriptQuotes: Leave -JavaScriptWrapImports: true -KeepEmptyLinesAtTheStartOfBlocks: false -MacroBlockBegin: '' -MacroBlockEnd: '' -MaxEmptyLinesToKeep: 1 -NamespaceIndentation: None -ObjCBinPackProtocolList: Never -ObjCBlockIndentWidth: 2 -ObjCSpaceAfterProperty: false -ObjCSpaceBeforeProtocolList: true -PenaltyBreakAssignment: 2 -PenaltyBreakBeforeFirstCallParameter: 1 -PenaltyBreakComment: 300 -PenaltyBreakFirstLessLess: 120 -PenaltyBreakString: 1000 -PenaltyBreakTemplateDeclaration: 10 -PenaltyExcessCharacter: 1000000 -PenaltyReturnTypeOnItsOwnLine: 200 -PointerAlignment: Left -RawStringFormats: - - Language: Cpp - Delimiters: - - cc - - CC - - cpp - - Cpp - - CPP - - 'c++' - - 'C++' - CanonicalDelimiter: '' - BasedOnStyle: google - - Language: TextProto - Delimiters: - - pb - - PB - - proto - - PROTO - EnclosingFunctions: - - EqualsProto - - EquivToProto - - PARSE_PARTIAL_TEXT_PROTO - - PARSE_TEST_PROTO - - PARSE_TEXT_PROTO - - ParseTextOrDie - - ParseTextProtoOrDie - CanonicalDelimiter: '' - BasedOnStyle: google -ReflowComments: true -SortIncludes: true -SortUsingDeclarations: true -SpaceAfterCStyleCast: false -SpaceAfterLogicalNot: false -SpaceAfterTemplateKeyword: true -SpaceBeforeAssignmentOperators: true -SpaceBeforeCpp11BracedList: false -SpaceBeforeCtorInitializerColon: true -SpaceBeforeInheritanceColon: true -SpaceBeforeParens: ControlStatements -SpaceBeforeRangeBasedForLoopColon: true -SpaceInEmptyBlock: false -SpaceInEmptyParentheses: false -SpacesBeforeTrailingComments: 2 -SpacesInAngles: false -SpacesInConditionalStatement: false -SpacesInContainerLiterals: true -SpacesInCStyleCastParentheses: false -SpacesInParentheses: false -SpacesInSquareBrackets: false -SpaceBeforeSquareBrackets: false -Standard: Auto -StatementMacros: - - Q_UNUSED - - QT_REQUIRE_VERSION -TabWidth: 8 -UseCRLF: false -UseTab: Never -... - diff --git a/ggml/src/ggml-cann/CMakeLists.txt b/ggml/src/ggml-cann/CMakeLists.txt index 05cf06bfa..0d8e483b2 100644 --- a/ggml/src/ggml-cann/CMakeLists.txt +++ b/ggml/src/ggml-cann/CMakeLists.txt @@ -51,13 +51,11 @@ if (CANN_INSTALL_DIR) ${CANN_INSTALL_DIR}/acllib/include ) - add_subdirectory(kernels) list(APPEND CANN_LIBRARIES ascendcl nnopbase opapi acl_op_compiler - ascendc_kernels ) file(GLOB GGML_SOURCES_CANN "*.cpp") diff --git a/ggml/src/ggml-cann/acl_tensor.cpp b/ggml/src/ggml-cann/acl_tensor.cpp index d120ce6ac..f5462c5a1 100644 --- a/ggml/src/ggml-cann/acl_tensor.cpp +++ b/ggml/src/ggml-cann/acl_tensor.cpp @@ -41,6 +41,8 @@ aclDataType ggml_cann_type_mapping(ggml_type type) { return ACL_INT4; case GGML_TYPE_Q8_0: return ACL_INT8; + case GGML_TYPE_I64: + return ACL_INT64; default: return ACL_DT_UNDEFINED; } @@ -54,9 +56,7 @@ aclTensor* ggml_cann_create_tensor(const ggml_tensor* tensor, int64_t* ne, // added. int64_t acl_ne[GGML_MAX_DIMS * 2], acl_stride[GGML_MAX_DIMS * 2]; - int64_t acl_storage_len = 0; if (ne == nullptr) { - acl_storage_len = ggml_nbytes(tensor); for (int i = 0; i < GGML_MAX_DIMS; i++) { acl_ne[i] = tensor->ne[i]; // The step size of acl is in elements. @@ -65,14 +65,18 @@ aclTensor* ggml_cann_create_tensor(const ggml_tensor* tensor, int64_t* ne, } else { // With bcast for (int i = 0; i < dims; i++) { - acl_storage_len += (ne[i] - 1) * nb[i]; acl_ne[i] = ne[i]; acl_stride[i] = nb[i] / ggml_element_size(tensor); } } - // Reverse ne and stride. int64_t final_dims = (dims == 0 ? GGML_MAX_DIMS : dims); + int64_t acl_storage_len = 1; + for (int i = 0; i < final_dims; i++) { + acl_storage_len += (acl_ne[i] - 1) * acl_stride[i]; + } + + // Reverse ne and stride. std::reverse(acl_ne, acl_ne + final_dims); std::reverse(acl_stride, acl_stride + final_dims); diff --git a/ggml/src/ggml-cann/acl_tensor.h b/ggml/src/ggml-cann/acl_tensor.h index 4734a9cb8..93f09937e 100644 --- a/ggml/src/ggml-cann/acl_tensor.h +++ b/ggml/src/ggml-cann/acl_tensor.h @@ -101,14 +101,14 @@ aclTensor* ggml_cann_create_tensor(void* data_ptr, aclDataType dtype, tmp_stride[i] = nb[i] / type_size; } + int64_t acl_storage_len = 1; + for (int i = 0; i < dims; i++) { + acl_storage_len += (tmp_ne[i] - 1) * tmp_stride[i]; + } + std::reverse(tmp_ne, tmp_ne + dims); std::reverse(tmp_stride, tmp_stride + dims); - int64_t acl_storage_len = 0; - for (int i = 0; i < dims; i++) { - acl_storage_len += (ne[i] - 1) * nb[i]; - } - aclTensor* acl_tensor = aclCreateTensor(tmp_ne, dims, dtype, tmp_stride, offset / type_size, format, &acl_storage_len, 1, data_ptr); diff --git a/ggml/src/ggml-cann/aclnn_ops.cpp b/ggml/src/ggml-cann/aclnn_ops.cpp index b2d857e1e..37d411797 100644 --- a/ggml/src/ggml-cann/aclnn_ops.cpp +++ b/ggml/src/ggml-cann/aclnn_ops.cpp @@ -28,8 +28,8 @@ #include #include #include -#include #include +#include #include #include #include @@ -44,12 +44,26 @@ #include #include #include -#include #include #include #include #include #include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include #include #include @@ -58,12 +72,41 @@ #include #include "ggml-impl.h" -#include "kernels/ascendc_kernels.h" #define GGML_COMMON_DECL_C #include "../ggml-common.h" +void bcast_shape(ggml_tensor * src0, ggml_tensor * src1, ggml_tensor * dst, aclTensor ** acl_src0, + aclTensor ** acl_src1, aclTensor ** acl_dst) { + GGML_ASSERT(ggml_are_same_shape(src0, dst) && ggml_can_repeat(src1, src0)); + // Need bcast + if (!ggml_are_same_shape(src0, src1) && ggml_cann_need_bcast(src0, src1)) { + BCAST_SHAPE(src0, src1) + *acl_src0 = ggml_cann_create_tensor(src0, BCAST_PARAM(src0)); + *acl_src1 = ggml_cann_create_tensor(src1, BCAST_PARAM(src1)); + *acl_dst = ggml_cann_create_tensor(dst, BCAST_PARAM(src0)); + } else { + *acl_src0 = ggml_cann_create_tensor(src0); + *acl_src1 = ggml_cann_create_tensor(src1); + *acl_dst = ggml_cann_create_tensor(dst); + } +} + +void ggml_cann_unary_op( + std::function unary_op, + ggml_backend_cann_context& ctx, ggml_tensor* dst) { + ggml_tensor* src = dst->src[0]; + + aclTensor* acl_src = ggml_cann_create_tensor(src); + aclTensor* acl_dst = ggml_cann_create_tensor(dst); + + unary_op(ctx, acl_src, acl_dst); + + ACL_CHECK(aclDestroyTensor(acl_src)); + ACL_CHECK(aclDestroyTensor(acl_dst)); +} + /** * @brief Repeats elements of a tensor along each dimension according to the * specified repeat array. @@ -79,26 +122,45 @@ static void aclnn_repeat(ggml_backend_cann_context& ctx, aclTensor* acl_src, // repeat tensor along each dim with repeat_array aclIntArray* repeats = aclCreateIntArray(repeat_array, GGML_MAX_DIMS); - uint64_t workspaceSize = 0; - aclOpExecutor* executor; - void* workspaceAddr = nullptr; - - ACL_CHECK(aclnnRepeatGetWorkspaceSize(acl_src, repeats, acl_dst, - &workspaceSize, &executor)); - - if (workspaceSize > 0) { - // Memory from allocator will "free" immediately, and this memory - // will be alloced to other pointers, but it won't access before - // this async task end because all tasks in same stream will execute - // in queue. - ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize); - workspaceAddr = workspace_allocator.get(); - } - ACL_CHECK( - aclnnRepeat(workspaceAddr, workspaceSize, executor, ctx.stream())); + GGML_CANN_CALL_ACLNN_OP(Repeat, acl_src, repeats, acl_dst); ACL_CHECK(aclDestroyIntArray(repeats)); } +/** + * @brief Casts the data type of a source tensor to a destination tensor. + * + * This function casts the data type of the source tensor `acl_src` to the + * specified data type `cast_data_type` and stores the result in the destination + * tensor `acl_dst`. + * + * @param ctx The context for the CANN backend operations. + * @param acl_src The source tensor whose data type will be casted. + * @param acl_dst The destination tensor where the casted result will be stored. + * @param cast_data_type The target data type to which the source tensor will be + * casted. + */ +static void aclnn_cast(ggml_backend_cann_context& ctx, aclTensor* acl_src, + aclTensor* acl_dst, aclDataType cast_data_type) { + GGML_CANN_CALL_ACLNN_OP(Cast, acl_src, cast_data_type, acl_dst); +} + +/** + * @brief Casts the elements of a tensor to a specified data type using the CANN backend. + * + * @details This function performs a type conversion on the elements of the input tensor `acl_src` + * and stores the results in the destination tensor `acl_dst`. The conversion type is + * determined based on the `dst` tensor's data type. + * + * @param ctx The context for the CANN backend operations. + * @param acl_src The source tensor whose elements will be cast. + * @param acl_dst The destination tensor that will store the casted elements. + * @param dst The ggml tensor specifying the target data type. + */ +static void aclnn_cast(ggml_backend_cann_context& ctx, aclTensor* acl_src, + aclTensor* acl_dst, ggml_tensor* dst) { + aclnn_cast(ctx, acl_src, acl_dst, ggml_cann_type_mapping(dst->type)); +} + void ggml_cann_repeat(ggml_backend_cann_context& ctx, ggml_tensor* dst) { ggml_tensor* src = dst->src[0]; GGML_ASSERT(ggml_can_repeat(src, dst)); @@ -114,69 +176,75 @@ void ggml_cann_repeat(ggml_backend_cann_context& ctx, ggml_tensor* dst) { ACL_CHECK(aclDestroyTensor(acl_dst)); } -/** - * @brief Adds two tensors element-wise and stores the result in a destination - * tensor. - * - * This function performs the operation: - * \f[ - * dst = acl\_src0 + alpha \times acl\_src1 - * \f] - * where alpha is a scalar value and defaults to 1.0f. - * - * @param ctx The context for the CANN backend operations. - * @param acl_src0 The first source tensor. - * @param acl_src1 The second source tensor. - * @param acl_dst The destination tensor where the result will be stored. - */ -static void aclnn_add(ggml_backend_cann_context& ctx, aclTensor* acl_src0, +void aclnn_add(ggml_backend_cann_context& ctx, aclTensor* acl_src0, aclTensor* acl_src1, aclTensor* acl_dst) { - aclScalar* alpha = nullptr; float alphaValue = 1.0f; - alpha = aclCreateScalar(&alphaValue, aclDataType::ACL_FLOAT); - - uint64_t workspaceSize = 0; - aclOpExecutor* executor; - void* workspaceAddr = nullptr; - - ACL_CHECK(aclnnAddGetWorkspaceSize(acl_src0, acl_src1, alpha, acl_dst, - &workspaceSize, &executor)); - if (workspaceSize > 0) { - ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize); - workspaceAddr = workspace_allocator.get(); - } - - ACL_CHECK(aclnnAdd(workspaceAddr, workspaceSize, executor, ctx.stream())); - + aclScalar* alpha = aclCreateScalar(&alphaValue, aclDataType::ACL_FLOAT); + if (acl_dst != nullptr) + GGML_CANN_CALL_ACLNN_OP(Add, acl_src0, acl_src1, alpha, acl_dst); + else + GGML_CANN_CALL_ACLNN_OP(InplaceAdd, acl_src0, acl_src1, alpha); ACL_CHECK(aclDestroyScalar(alpha)); } -void ggml_cann_add(ggml_backend_cann_context& ctx, ggml_tensor* dst) { - ggml_tensor* src0 = dst->src[0]; - ggml_tensor* src1 = dst->src[1]; - GGML_ASSERT(ggml_can_repeat(src1, src0) && ggml_are_same_shape(src0, dst)); +void aclnn_sub(ggml_backend_cann_context& ctx, aclTensor* acl_src0, + aclTensor* acl_src1, aclTensor* acl_dst) { + float alphaValue = 1.0f; + aclScalar* alpha = aclCreateScalar(&alphaValue, aclDataType::ACL_FLOAT); + if (acl_dst != nullptr) + GGML_CANN_CALL_ACLNN_OP(Sub, acl_src0, acl_src1, alpha, acl_dst); + else + GGML_CANN_CALL_ACLNN_OP(InplaceSub, acl_src0, acl_src1, alpha); + ACL_CHECK(aclDestroyScalar(alpha)); +} - aclTensor* acl_src0; - aclTensor* acl_src1; - aclTensor* acl_dst; +void aclnn_mul(ggml_backend_cann_context& ctx, aclTensor* acl_src, + aclTensor* acl_other, aclTensor* acl_dst) { + if (acl_dst != nullptr) + GGML_CANN_CALL_ACLNN_OP(Mul, acl_src, acl_other, acl_dst); + else + GGML_CANN_CALL_ACLNN_OP(InplaceMul, acl_src, acl_other); +} - // Need bcast - if (!ggml_are_same_shape(src0, src1) && ggml_cann_need_bcast(src0, src1)) { - BCAST_SHAPE(src0, src1) - acl_src0 = ggml_cann_create_tensor(src0, BCAST_PARAM(src0)); - acl_src1 = ggml_cann_create_tensor(src1, BCAST_PARAM(src1)); - acl_dst = ggml_cann_create_tensor(dst, BCAST_PARAM(src0)); +void aclnn_div(ggml_backend_cann_context& ctx, aclTensor* acl_src, + aclTensor* acl_other, aclTensor* acl_dst) { + if (acl_dst != nullptr) + GGML_CANN_CALL_ACLNN_OP(Div, acl_src, acl_other, acl_dst); + else + GGML_CANN_CALL_ACLNN_OP(InplaceDiv, acl_src, acl_other); +} + +/** + * @brief Multiplies elements of a tensor by a scalar value, optionally + * in-place. + * + * This function multiplies each element of the source tensor `acl_src` by the + * scalar `scale` and stores the result in the destination tensor `acl_dst`. If + * `inplace` is true, `acl_dst` will not be used and the operation is performed + * in-place on `acl_src`. + * The operation is defined as: + * \f[ + * \text {acl_dst }_i=\text {acl_src }_i \times \text {scale} + * \f] + * + * @param ctx The context for the CANN backend operations. + * @param acl_src The source tensor whose elements will be multiplied. + * @param scale The scalar value by which each element of `acl_src` will be + * multiplied. + * @param acl_dst The destination tensor where the result will be stored if + * `inplace` is false. + * @param inplace Flag indicating whether to perform the operation in-place on + * `acl_src`. + */ +static void aclnn_muls(ggml_backend_cann_context& ctx, aclTensor* acl_src, + float scale, aclTensor* acl_dst, bool inplace) { + aclScalar* acl_scale = aclCreateScalar(&scale, aclDataType::ACL_FLOAT); + if (inplace) { + GGML_CANN_CALL_ACLNN_OP(InplaceMuls, acl_src, acl_scale); } else { - acl_src0 = ggml_cann_create_tensor(src0); - acl_src1 = ggml_cann_create_tensor(src1); - acl_dst = ggml_cann_create_tensor(dst); + GGML_CANN_CALL_ACLNN_OP(Muls, acl_src, acl_scale, acl_dst); } - - aclnn_add(ctx, acl_src0, acl_src1, acl_dst); - - ACL_CHECK(aclDestroyTensor(acl_src0)); - ACL_CHECK(aclDestroyTensor(acl_src1)); - ACL_CHECK(aclDestroyTensor(acl_dst)); + ACL_CHECK(aclDestroyScalar(acl_scale)); } void ggml_cann_leaky_relu(ggml_backend_cann_context& ctx, ggml_tensor* dst) { @@ -193,19 +261,7 @@ void ggml_cann_leaky_relu(ggml_backend_cann_context& ctx, ggml_tensor* dst) { aclScalar* acl_negative_slope = aclCreateScalar(&negative_slope, aclDataType::ACL_FLOAT); - uint64_t workspaceSize = 0; - aclOpExecutor* executor; - void* workspaceAddr = nullptr; - - ACL_CHECK(aclnnLeakyReluGetWorkspaceSize( - acl_src, acl_negative_slope, acl_dst, &workspaceSize, &executor)); - if (workspaceSize > 0) { - ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize); - workspaceAddr = workspace_allocator.get(); - } - - ACL_CHECK( - aclnnLeakyRelu(workspaceAddr, workspaceSize, executor, ctx.stream())); + GGML_CANN_CALL_ACLNN_OP(LeakyRelu, acl_src, acl_negative_slope, acl_dst); ACL_CHECK(aclDestroyScalar(acl_negative_slope)); ACL_CHECK(aclDestroyTensor(acl_src)); @@ -225,18 +281,7 @@ void ggml_cann_leaky_relu(ggml_backend_cann_context& ctx, ggml_tensor* dst) { static void aclnn_concat(ggml_backend_cann_context& ctx, aclTensorList* tensorList, aclTensor* acl_dst, int64_t concat_dim) { - uint64_t workspaceSize = 0; - aclOpExecutor* executor; - void* workspaceAddr = nullptr; - - ACL_CHECK(aclnnCatGetWorkspaceSize(tensorList, concat_dim, acl_dst, - &workspaceSize, &executor)); - if (workspaceSize > 0) { - ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize); - workspaceAddr = workspace_allocator.get(); - } - - ACL_CHECK(aclnnCat(workspaceAddr, workspaceSize, executor, ctx.stream())); + GGML_CANN_CALL_ACLNN_OP(Cat, tensorList, concat_dim, acl_dst); } void ggml_cann_concat(ggml_backend_cann_context& ctx, ggml_tensor* dst) { @@ -282,24 +327,11 @@ static void aclnn_arange(ggml_backend_cann_context& ctx, aclTensor* acl_dst, int64_t steps = (int64_t)std::ceil((stop - start) / step); GGML_ASSERT(n_elements == steps); - uint64_t workspaceSize = 0; - aclOpExecutor* executor; - void* workspaceAddr = nullptr; - aclScalar* acl_start = aclCreateScalar(&start, aclDataType::ACL_FLOAT); aclScalar* acl_end = aclCreateScalar(&stop, aclDataType::ACL_FLOAT); aclScalar* acl_step = aclCreateScalar(&step, aclDataType::ACL_FLOAT); - ACL_CHECK(aclnnArangeGetWorkspaceSize(acl_start, acl_end, acl_step, acl_dst, - &workspaceSize, &executor)); - if (workspaceSize > 0) { - ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize); - workspaceAddr = workspace_allocator.get(); - } - - ACL_CHECK( - aclnnArange(workspaceAddr, workspaceSize, executor, ctx.stream())); - + GGML_CANN_CALL_ACLNN_OP(Arange, acl_start, acl_end, acl_step, acl_dst); ACL_CHECK(aclDestroyScalar(acl_start)); ACL_CHECK(aclDestroyScalar(acl_end)); ACL_CHECK(aclDestroyScalar(acl_step)); @@ -322,15 +354,8 @@ void ggml_cann_arange(ggml_backend_cann_context& ctx, ggml_tensor* dst) { ACL_CHECK(aclDestroyTensor(acl_dst)); } -void ggml_cann_sqr(ggml_backend_cann_context& ctx, ggml_tensor* dst) { - dst->src[1] = dst->src[0]; - ggml_cann_mul_div(ctx, dst); -} - void ggml_cann_clamp(ggml_backend_cann_context& ctx, ggml_tensor* dst) { ggml_tensor* src = dst->src[0]; - GGML_ASSERT(src->type == GGML_TYPE_F32); - GGML_ASSERT(dst->type == GGML_TYPE_F32); float min; float max; @@ -343,19 +368,7 @@ void ggml_cann_clamp(ggml_backend_cann_context& ctx, ggml_tensor* dst) { aclScalar* acl_min = aclCreateScalar(&min, aclDataType::ACL_FLOAT); aclScalar* acl_max = aclCreateScalar(&max, aclDataType::ACL_FLOAT); - uint64_t workspaceSize = 0; - aclOpExecutor* executor; - void* workspaceAddr = nullptr; - - ACL_CHECK(aclnnClampGetWorkspaceSize(acl_src, acl_min, acl_max, acl_dst, - &workspaceSize, &executor)); - if (workspaceSize > 0) { - ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize); - workspaceAddr = workspace_allocator.get(); - } - - ACL_CHECK(aclnnClamp(workspaceAddr, workspaceSize, executor, ctx.stream())); - + GGML_CANN_CALL_ACLNN_OP(Clamp, acl_src, acl_min, acl_max, acl_dst); ACL_CHECK(aclDestroyScalar(acl_min)); ACL_CHECK(aclDestroyScalar(acl_max)); ACL_CHECK(aclDestroyTensor(acl_src)); @@ -373,19 +386,7 @@ void ggml_cann_scale(ggml_backend_cann_context& ctx, ggml_tensor* dst) { aclTensor* acl_src = ggml_cann_create_tensor(src); aclTensor* acl_dst = ggml_cann_create_tensor(dst); - uint64_t workspaceSize = 0; - aclOpExecutor* executor; - void* workspaceAddr = nullptr; - - ACL_CHECK(aclnnMulsGetWorkspaceSize(acl_src, scale, acl_dst, &workspaceSize, - &executor)); - if (workspaceSize > 0) { - ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize); - workspaceAddr = workspace_allocator.get(); - } - - ACL_CHECK(aclnnMuls(workspaceAddr, workspaceSize, executor, ctx.stream())); - + GGML_CANN_CALL_ACLNN_OP(Muls, acl_src, scale, acl_dst); ACL_CHECK(aclDestroyScalar(scale)); ACL_CHECK(aclDestroyTensor(acl_src)); ACL_CHECK(aclDestroyTensor(acl_dst)); @@ -403,33 +404,9 @@ void ggml_cann_argsort(ggml_backend_cann_context& ctx, ggml_tensor* dst) { aclTensor* tmp_tensor = ggml_cann_create_tensor(buffer, ACL_INT64, ggml_type_size(dst->type), dst->ne, dst->nb, GGML_MAX_DIMS); - - uint64_t workspaceSize = 0; - aclOpExecutor* executor; - void* workspaceAddr = nullptr; - - ACL_CHECK(aclnnArgsortGetWorkspaceSize( - acl_src, -1, (order == GGML_SORT_ORDER_DESC ? true : false), tmp_tensor, - &workspaceSize, &executor)); - if (workspaceSize > 0) { - ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize); - workspaceAddr = workspace_allocator.get(); - } - - ACL_CHECK( - aclnnArgsort(workspaceAddr, workspaceSize, executor, ctx.stream())); - - workspaceSize = 0; - ACL_CHECK(aclnnCastGetWorkspaceSize(tmp_tensor, - ggml_cann_type_mapping(dst->type), - acl_dst, &workspaceSize, &executor)); - if (workspaceSize > 0) { - ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize); - workspaceAddr = workspace_allocator.get(); - } - - ACL_CHECK(aclnnCast(workspaceAddr, workspaceSize, executor, ctx.stream())); - + GGML_CANN_CALL_ACLNN_OP(Argsort, acl_src, -1, (order == GGML_SORT_ORDER_DESC ? true : false), + tmp_tensor); + GGML_CANN_CALL_ACLNN_OP(Cast, tmp_tensor, ggml_cann_type_mapping(dst->type), acl_dst); ACL_CHECK(aclDestroyTensor(acl_src)); ACL_CHECK(aclDestroyTensor(tmp_tensor)); ACL_CHECK(aclDestroyTensor(acl_dst)); @@ -444,24 +421,10 @@ void ggml_cann_norm(ggml_backend_cann_context& ctx, ggml_tensor* dst) { float eps; memcpy(&eps, dst->op_params, sizeof(float)); - uint64_t workspaceSize = 0; - aclOpExecutor* executor; - void* workspaceAddr = nullptr; - std::vector normData = {dst->ne[0]}; aclIntArray* norm = aclCreateIntArray(normData.data(), normData.size()); - ACL_CHECK(aclnnLayerNormGetWorkspaceSize(acl_src, norm, nullptr, nullptr, - eps, acl_dst, nullptr, nullptr, - &workspaceSize, &executor)); - - if (workspaceSize > 0) { - ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize); - workspaceAddr = workspace_allocator.get(); - } - - ACL_CHECK( - aclnnLayerNorm(workspaceAddr, workspaceSize, executor, ctx.stream())); - + GGML_CANN_CALL_ACLNN_OP(LayerNorm, acl_src, norm, nullptr, nullptr, + eps, acl_dst, nullptr, nullptr); ACL_CHECK(aclDestroyIntArray(norm)); ACL_CHECK(aclDestroyTensor(acl_src)); ACL_CHECK(aclDestroyTensor(acl_dst)); @@ -478,10 +441,6 @@ void ggml_cann_group_norm(ggml_backend_cann_context& ctx, ggml_tensor* dst) { float eps; memcpy(&eps, dst->op_params + 1, sizeof(float)); - uint64_t workspaceSize = 0; - aclOpExecutor* executor; - void* workspaceAddr = nullptr; - int64_t N = src->ne[3]; int64_t C = src->ne[2]; int64_t HxW = src->ne[1] * src->ne[0]; @@ -498,18 +457,8 @@ void ggml_cann_group_norm(ggml_backend_cann_context& ctx, ggml_tensor* dst) { aclTensor* acl_rstd_out = ggml_cann_create_tensor( (char*)buffer + n_bytes, ACL_FLOAT, type_size, ne, nb, ACL_FORMAT_ND); - ACL_CHECK(aclnnGroupNormGetWorkspaceSize( - acl_src, nullptr, nullptr, N, C, HxW, n_groups, eps, acl_dst, - acl_mean_out, acl_rstd_out, &workspaceSize, &executor)); - - if (workspaceSize > 0) { - ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize); - workspaceAddr = workspace_allocator.get(); - } - - ACL_CHECK( - aclnnGroupNorm(workspaceAddr, workspaceSize, executor, ctx.stream())); - + GGML_CANN_CALL_ACLNN_OP(GroupNorm, acl_src, nullptr, nullptr, N, C, HxW, n_groups, eps, + acl_dst, acl_mean_out, acl_rstd_out); ACL_CHECK(aclDestroyTensor(acl_src)); ACL_CHECK(aclDestroyTensor(acl_dst)); ACL_CHECK(aclDestroyTensor(acl_mean_out)); @@ -536,68 +485,57 @@ void ggml_cann_acc(ggml_backend_cann_context& ctx, ggml_tensor* dst) { float alphaValue = 1.0f; alpha = aclCreateScalar(&alphaValue, aclDataType::ACL_FLOAT); - uint64_t workspaceSize = 0; - aclOpExecutor* executor; - void* workspaceAddr = nullptr; - if (!inplace) { size_t cpy_size = ggml_nbytes(dst); ACL_CHECK(aclrtMemcpyAsync(dst->data, cpy_size, src0->data, cpy_size, ACL_MEMCPY_DEVICE_TO_DEVICE, ctx.stream())); aclTensor* acl_src0 = ggml_cann_create_tensor( src0, src1->ne, src0->nb, GGML_MAX_DIMS, ACL_FORMAT_ND, offset); - ACL_CHECK(aclnnAddGetWorkspaceSize(acl_src0, acl_src1, alpha, acl_dst, - &workspaceSize, &executor)); - if (workspaceSize > 0) { - ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize); - workspaceAddr = workspace_allocator.get(); - } - ACL_CHECK( - aclnnAdd(workspaceAddr, workspaceSize, executor, ctx.stream())); + + GGML_CANN_CALL_ACLNN_OP(Add, acl_src0, acl_src1, alpha, acl_dst); ACL_CHECK(aclDestroyTensor(acl_src0)); } else { - ACL_CHECK(aclnnInplaceAddGetWorkspaceSize(acl_dst, acl_src1, alpha, - &workspaceSize, &executor)); - if (workspaceSize > 0) { - ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize); - workspaceAddr = workspace_allocator.get(); - } - ACL_CHECK(aclnnInplaceAdd(workspaceAddr, workspaceSize, executor, - ctx.stream())); + GGML_CANN_CALL_ACLNN_OP(InplaceAdd, acl_dst, acl_src1, alpha); } ACL_CHECK(aclDestroyTensor(acl_src1)); ACL_CHECK(aclDestroyTensor(acl_dst)); } -void ggml_cann_sum_rows(ggml_backend_cann_context& ctx, ggml_tensor* dst) { - ggml_tensor* src = dst->src[0]; - - aclTensor* acl_src = ggml_cann_create_tensor(src); +/** + * @brief Performs sum reduction on a given tensor along specified dimensions. + * + * This function reduces the input tensor by summing along the specified dimensions. + * + * @param ctx The context for the CANN backend operations. + * @param dst The destination tensor where the reduced result will be stored. + * @param dim An array of dimension indices. + * @param dim_size The number of dimensions. + */ +static void aclnn_reduce_sum(ggml_backend_cann_context& ctx, ggml_tensor* dst, + int64_t* dim, size_t dim_size) { GGML_ASSERT(dst->ne[0] == 1); + ggml_tensor* src = dst->src[0]; + aclTensor* acl_src = ggml_cann_create_tensor(src); aclTensor* acl_dst = ggml_cann_create_tensor(dst); + aclIntArray* reduce_dims = aclCreateIntArray(dim, dim_size); - int64_t reduce_dims_host[] = {3}; - aclIntArray* reduce_dims = aclCreateIntArray(reduce_dims_host, 1); - - uint64_t workspaceSize = 0; - aclOpExecutor* executor; - void* workspaceAddr = nullptr; - - ACL_CHECK(aclnnReduceSumGetWorkspaceSize( - acl_src, reduce_dims, true, ggml_cann_type_mapping(src->type), acl_dst, - &workspaceSize, &executor)); - if (workspaceSize > 0) { - ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize); - workspaceAddr = workspace_allocator.get(); - } - - ACL_CHECK( - aclnnReduceSum(workspaceAddr, workspaceSize, executor, ctx.stream())); - + GGML_CANN_CALL_ACLNN_OP(ReduceSum, acl_src, reduce_dims, true, + ggml_cann_type_mapping(dst->type), acl_dst); ACL_CHECK(aclDestroyTensor(acl_src)); ACL_CHECK(aclDestroyTensor(acl_dst)); + ACL_CHECK(aclDestroyIntArray(reduce_dims)); +} + +void ggml_cann_sum_rows(ggml_backend_cann_context& ctx, ggml_tensor* dst) { + int64_t reduce_dims[] = {3}; + aclnn_reduce_sum(ctx, dst, reduce_dims, 1); +} + +void ggml_cann_sum(ggml_backend_cann_context& ctx, ggml_tensor* dst) { + int64_t reduce_dims[] = {0, 1, 2, 3}; + aclnn_reduce_sum(ctx, dst, reduce_dims, 4); } void ggml_cann_upsample_nearest2d(ggml_backend_cann_context& ctx, @@ -611,20 +549,7 @@ void ggml_cann_upsample_nearest2d(ggml_backend_cann_context& ctx, std::vector output_size{dst->ne[1], dst->ne[0]}; auto output_size_array = aclCreateIntArray(output_size.data(), 2); - uint64_t workspaceSize = 0; - aclOpExecutor* executor; - void* workspaceAddr = nullptr; - - ACL_CHECK(aclnnUpsampleNearest2dGetWorkspaceSize( - acl_src, output_size_array, acl_dst, &workspaceSize, &executor)); - if (workspaceSize > 0) { - ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize); - workspaceAddr = workspace_allocator.get(); - } - - ACL_CHECK(aclnnUpsampleNearest2d(workspaceAddr, workspaceSize, executor, - ctx.stream())); - + GGML_CANN_CALL_ACLNN_OP(UpsampleNearest2d, acl_src, output_size_array, acl_dst); ACL_CHECK(aclDestroyIntArray(output_size_array)); ACL_CHECK(aclDestroyTensor(acl_src)); ACL_CHECK(aclDestroyTensor(acl_dst)); @@ -650,21 +575,7 @@ static void aclnn_pad(ggml_backend_cann_context& ctx, aclTensor* acl_src, aclIntArray* acl_pad = aclCreateIntArray(paddings, GGML_MAX_DIMS * 2); aclScalar* acl_value = aclCreateScalar(&value, aclDataType::ACL_FLOAT); - uint64_t workspaceSize = 0; - aclOpExecutor* executor; - void* workspaceAddr = nullptr; - - ACL_CHECK(aclnnConstantPadNdGetWorkspaceSize( - acl_src, acl_pad, acl_value, acl_dst, &workspaceSize, &executor)); - - if (workspaceSize > 0) { - ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize); - workspaceAddr = workspace_allocator.get(); - } - - ACL_CHECK(aclnnConstantPadNd(workspaceAddr, workspaceSize, executor, - ctx.stream())); - + GGML_CANN_CALL_ACLNN_OP(ConstantPadNd, acl_src, acl_pad, acl_value, acl_dst); ACL_CHECK(aclDestroyIntArray(acl_pad)); ACL_CHECK(aclDestroyScalar(acl_value)); } @@ -730,23 +641,9 @@ static void ggml_cann_avg_pool2d(ggml_backend_cann_context& ctx, bool count_include_pad = true; int64_t divisor_override = 0; int8_t cube_math_type = 0; - - uint64_t workspaceSize = 0; - aclOpExecutor* executor; - void* workspaceAddr = nullptr; - - ACL_CHECK(aclnnAvgPool2dGetWorkspaceSize( - acl_src, kernel_size, strides, paddings_avg, ceil_mode, - count_include_pad, divisor_override, cube_math_type, acl_dst, - &workspaceSize, &executor)); - - if (workspaceSize > 0) { - ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize); - workspaceAddr = workspace_allocator.get(); - } - ACL_CHECK( - aclnnAvgPool2d(workspaceAddr, workspaceSize, executor, ctx.stream())); - + GGML_CANN_CALL_ACLNN_OP(AvgPool2d, acl_src, kernel_size, strides, paddings_avg, + ceil_mode, count_include_pad, divisor_override, + cube_math_type, acl_dst); ACL_CHECK(aclDestroyTensor(acl_src)); ACL_CHECK(aclDestroyTensor(acl_dst)); ACL_CHECK(aclDestroyIntArray(kernel_size)); @@ -819,22 +716,8 @@ static void ggml_cann_max_pool2d(ggml_backend_cann_context& ctx, bool ceil_mode = false; int64_t auto_pads = 0; - - uint64_t workspaceSize = 0; - aclOpExecutor* executor; - void* workspaceAddr = nullptr; - - ACL_CHECK(aclnnMaxPoolGetWorkspaceSize( - tmp_tensor, kernel_size, strides, auto_pads, paddings_max, dilations, - ceil_mode, acl_dst, &workspaceSize, &executor)); - if (workspaceSize > 0) { - ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize); - workspaceAddr = workspace_allocator.get(); - } - - ACL_CHECK( - aclnnMaxPool(workspaceAddr, workspaceSize, executor, ctx.stream())); - + GGML_CANN_CALL_ACLNN_OP(MaxPool, tmp_tensor, kernel_size, strides, auto_pads, + paddings_max, dilations, ceil_mode, acl_dst); ACL_CHECK(aclDestroyTensor(acl_src)); ACL_CHECK(aclDestroyTensor(acl_dst)); ACL_CHECK(aclDestroyTensor(tmp_tensor)); @@ -872,206 +755,81 @@ void ggml_cann_pool2d(ggml_backend_cann_context& ctx, ggml_tensor* dst) { */ static void cann_copy(ggml_backend_cann_context& ctx, aclTensor* acl_src, aclTensor* acl_dst) { - uint64_t workspaceSize = 0; - aclOpExecutor* executor; - void* workspaceAddr = nullptr; - - ACL_CHECK(aclnnInplaceCopyGetWorkspaceSize(acl_dst, acl_src, &workspaceSize, - &executor)); - - if (workspaceSize > 0) { - ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize); - workspaceAddr = workspace_allocator.get(); - } - - ACL_CHECK( - aclnnInplaceCopy(workspaceAddr, workspaceSize, executor, ctx.stream())); + GGML_CANN_CALL_ACLNN_OP(InplaceCopy, acl_dst, acl_src); } void ggml_cann_dup(ggml_backend_cann_context& ctx, ggml_tensor* dst) { - ggml_tensor* src = dst->src[0]; + ggml_tensor* src0 = dst->src[0]; - aclTensor* acl_src = ggml_cann_create_tensor(src); + aclTensor* acl_src = ggml_cann_create_tensor(src0); aclTensor* acl_dst = ggml_cann_create_tensor(dst); - - ggml_cann_pool_alloc src_extra_allocator(ctx.pool(), sizeof(ggml_tensor)); - ggml_cann_pool_alloc dst_extra_allocator(ctx.pool(), sizeof(ggml_tensor)); - src->extra = src_extra_allocator.get(); - dst->extra = dst_extra_allocator.get(); - ACL_CHECK(aclrtMemcpyAsync(src->extra, sizeof(ggml_tensor), src, - sizeof(ggml_tensor), ACL_MEMCPY_HOST_TO_DEVICE, - ctx.stream())); - ACL_CHECK(aclrtMemcpyAsync(dst->extra, sizeof(ggml_tensor), dst, - sizeof(ggml_tensor), ACL_MEMCPY_HOST_TO_DEVICE, - ctx.stream())); - - if ((dst->type == GGML_TYPE_F16 || dst->type == GGML_TYPE_F32) && - ggml_are_same_shape(src, dst)) { - cann_copy(ctx, acl_src, acl_dst); - ACL_CHECK(aclDestroyTensor(acl_src)); - ACL_CHECK(aclDestroyTensor(acl_dst)); - return; - } - // TODO: simplify - if (src->type == GGML_TYPE_F16) { - if (dst->type == GGML_TYPE_Q8_0) { - aclrtlaunch_ascendc_quantize_f16_q8_0( - 24, ctx.stream(), src->data, dst->data, - ((ggml_tensor*)src->extra)->ne, ((ggml_tensor*)src->extra)->nb, - ((ggml_tensor*)dst->extra)->ne); - return; - } - if (dst->type == GGML_TYPE_Q4_0) { - aclrtlaunch_ascendc_quantize_f16_to_q4_0( - 24, ctx.stream(), src->data, dst->data, - ((ggml_tensor*)src->extra)->ne, ((ggml_tensor*)src->extra)->nb, - ((ggml_tensor*)dst->extra)->ne); - return; - } - if (dst->type == GGML_TYPE_F16) { - if (ggml_are_same_shape(src, dst)) { - cann_copy(ctx, acl_src, acl_dst); - ACL_CHECK(aclDestroyTensor(acl_src)); - ACL_CHECK(aclDestroyTensor(acl_dst)); - return; - } - if (ggml_is_contiguous(dst)) { - const size_t src_type_size = ggml_type_size(src->type); - if (src->nb[0] == src_type_size) { - // src0 is contigous on first dimension, copy by rows - int64_t rows_num = ggml_nrows(src); - - aclrtlaunch_ascendc_dup_by_rows_fp16( - rows_num, ctx.stream(), src->data, dst->data, - ((ggml_tensor*)src->extra)->ne, - ((ggml_tensor*)src->extra)->nb, - ((ggml_tensor*)dst->extra)->ne, - ((ggml_tensor*)dst->extra)->nb); - return; - } - GGML_ABORT("fatal error"); - } - GGML_ABORT("fatal error"); - } - if (dst->type == GGML_TYPE_F32) { - if (ggml_are_same_shape(src, dst)) { - cann_copy(ctx, acl_src, acl_dst); - ACL_CHECK(aclDestroyTensor(acl_src)); - ACL_CHECK(aclDestroyTensor(acl_dst)); - return; - } - if (ggml_is_contiguous(dst)) { - const size_t src_type_size = ggml_type_size(src->type); - if (src->nb[0] == src_type_size) { - // src0 is contigous on first dimension, copy by rows - int64_t rows_num = ggml_nrows(src); - aclrtlaunch_ascendc_dup_by_rows_fp16_to_fp32( - rows_num, ctx.stream(), src->data, dst->data, - ((ggml_tensor*)src->extra)->ne, - ((ggml_tensor*)src->extra)->nb, - ((ggml_tensor*)dst->extra)->ne, - ((ggml_tensor*)dst->extra)->nb); - return; - } - GGML_ABORT("fatal error"); - } - GGML_ABORT("fatal error"); - } - // TODO - GGML_ABORT("fatal error"); - } else if (src->type == GGML_TYPE_F32) { - // TODO: if (src0->type == dst->type && ne00 == ne0 && nb00 == type_size - // && nb0 == type_size) - if (dst->type == GGML_TYPE_Q8_0) { - aclrtlaunch_ascendc_quantize_f32_q8_0( - 24, ctx.stream(), src->data, dst->data, - ((ggml_tensor*)src->extra)->ne, ((ggml_tensor*)src->extra)->nb, - ((ggml_tensor*)dst->extra)->ne); - return; - } - if (dst->type == GGML_TYPE_Q4_0) { - aclrtlaunch_ascendc_quantize_f32_to_q4_0( - 24, ctx.stream(), src->data, dst->data, - ((ggml_tensor*)src->extra)->ne, ((ggml_tensor*)src->extra)->nb, - ((ggml_tensor*)dst->extra)->ne); - return; - } - if (dst->type == GGML_TYPE_F32) { - if (ggml_are_same_shape(src, dst)) { - cann_copy(ctx, acl_src, acl_dst); - ACL_CHECK(aclDestroyTensor(acl_src)); - ACL_CHECK(aclDestroyTensor(acl_dst)); - return; - } - if (ggml_is_contiguous(dst)) { - const size_t src_type_size = ggml_type_size(src->type); - if (src->nb[0] == src_type_size) { - // src0 is contigous on first dimension, copy by rows - int64_t rows_num = ggml_nrows(src); - aclrtlaunch_ascendc_dup_by_rows_fp32( - rows_num, ctx.stream(), src->data, dst->data, - ((ggml_tensor*)src->extra)->ne, - ((ggml_tensor*)src->extra)->nb, - ((ggml_tensor*)dst->extra)->ne, - ((ggml_tensor*)dst->extra)->nb); - return; - } - GGML_ABORT("fatal error"); - } else { - // TODO: dst not contiguous - GGML_ABORT("fatal error"); - } - } - if (dst->type == GGML_TYPE_F16) { - if (ggml_are_same_shape(src, dst)) { - cann_copy(ctx, acl_src, acl_dst); - ACL_CHECK(aclDestroyTensor(acl_src)); - ACL_CHECK(aclDestroyTensor(acl_dst)); - return; - } - if (ggml_is_contiguous(dst)) { - const size_t src_type_size = ggml_type_size(src->type); - if (src->nb[0] == src_type_size) { - // src0 is contigous on first dimension, copy by rows - int64_t rows_num = ggml_nrows(src); - aclrtlaunch_ascendc_dup_by_rows_fp32_to_fp16( - rows_num, ctx.stream(), src->data, dst->data, - ((ggml_tensor*)src->extra)->ne, - ((ggml_tensor*)src->extra)->nb, - ((ggml_tensor*)dst->extra)->ne, - ((ggml_tensor*)dst->extra)->nb); - return; - } - GGML_ABORT("fatal error"); - } - } - // TODO - GGML_ABORT("fatal error"); - } else { - if (ggml_are_same_shape(src, dst)) { + if (ggml_are_same_shape(src0, dst)) { + if (dst->type == src0->type) { cann_copy(ctx, acl_src, acl_dst); - ACL_CHECK(aclDestroyTensor(acl_src)); - ACL_CHECK(aclDestroyTensor(acl_dst)); - return; + } else { + aclnn_cast(ctx, acl_src, acl_dst, dst); } - GGML_ABORT("fatal error"); - } -} + } else { + if (ggml_is_contiguous(src0) && ggml_is_contiguous(dst)) { + if (dst->type == src0->type) { + size_t cpy_size = ggml_nbytes(dst); + ACL_CHECK(aclrtMemcpyAsync( + dst->data, cpy_size, src0->data, cpy_size, + ACL_MEMCPY_DEVICE_TO_DEVICE, ctx.stream())); + return; + } else { + ggml_cann_pool_alloc src_buffer_allocator( + ctx.pool(), + ggml_nelements(dst) * ggml_type_size(dst->type)); + void* src_trans_buffer = src_buffer_allocator.get(); + size_t src_trans_nb[GGML_MAX_DIMS]; + src_trans_nb[0] = ggml_type_size(dst->type); + for (int i = 1; i < GGML_MAX_DIMS; i++) { + src_trans_nb[i] = src_trans_nb[i - 1] * src0->ne[i - 1]; + } + aclTensor* src_trans_tensor = ggml_cann_create_tensor( + src_trans_buffer, ggml_cann_type_mapping(dst->type), + ggml_type_size(dst->type), src0->ne, src_trans_nb, + GGML_MAX_DIMS); -#ifdef __cplusplus -extern "C" { -#endif -aclnnStatus aclnnRmsNormGetWorkspaceSize(const aclTensor* x, - const aclTensor* gamma, double epsilon, - const aclTensor* yOut, - const aclTensor* rstdOout, - uint64_t* workspaceSize, - aclOpExecutor** executor); -aclnnStatus aclnnRmsNorm(void* workspace, uint64_t workspaceSize, - aclOpExecutor* executor, aclrtStream stream); -#ifdef __cplusplus + aclnn_cast(ctx, acl_src, src_trans_tensor, dst); + size_t cpy_size = ggml_nbytes(dst); + ACL_CHECK(aclrtMemcpyAsync( + dst->data, cpy_size, src_trans_buffer, cpy_size, + ACL_MEMCPY_DEVICE_TO_DEVICE, ctx.stream())); + ACL_CHECK(aclDestroyTensor(src_trans_tensor)); + return; + } + } else if (ggml_is_contiguous(dst)) { + ggml_cann_pool_alloc src_buffer_allocator( + ctx.pool(), ggml_nelements(dst) * ggml_type_size(dst->type)); + void* src_trans_buffer = src_buffer_allocator.get(); + size_t src_trans_nb[GGML_MAX_DIMS]; + src_trans_nb[0] = ggml_type_size(dst->type); + for (int i = 1; i < GGML_MAX_DIMS; i++) { + src_trans_nb[i] = src_trans_nb[i - 1] * src0->ne[i - 1]; + } + aclTensor* src_trans_tensor = ggml_cann_create_tensor( + src_trans_buffer, ggml_cann_type_mapping(dst->type), + ggml_type_size(dst->type), src0->ne, src_trans_nb, + GGML_MAX_DIMS); + + aclnn_cast(ctx, acl_src, src_trans_tensor, dst); + + size_t cpy_size = ggml_nbytes(dst); + ACL_CHECK(aclrtMemcpyAsync(dst->data, cpy_size, src_trans_buffer, + cpy_size, ACL_MEMCPY_DEVICE_TO_DEVICE, + ctx.stream())); + ACL_CHECK(aclDestroyTensor(src_trans_tensor)); + return; + } else { + GGML_ABORT("Unsupport dst is not tontiguous."); + } + } + + ACL_CHECK(aclDestroyTensor(acl_src)); + ACL_CHECK(aclDestroyTensor(acl_dst)); } -#endif /** * @brief Creates an ACL tensor initialized with zeros using a provided buffer. @@ -1131,21 +889,7 @@ static aclTensor* aclnn_values(ggml_backend_cann_context& ctx, void* buffer, float alpha_host = 1.0f; aclScalar* alpha = aclCreateScalar(&alpha_host, aclDataType::ACL_FLOAT); aclScalar* other = aclCreateScalar(&value, aclDataType::ACL_FLOAT); - - uint64_t workspaceSize = 0; - aclOpExecutor* executor; - void* workspaceAddr = nullptr; - - ACL_CHECK(aclnnInplaceAddsGetWorkspaceSize(acl_tensor, other, alpha, - &workspaceSize, &executor)); - - if (workspaceSize > 0) { - ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize); - workspaceAddr = workspace_allocator.get(); - } - ACL_CHECK( - aclnnInplaceAdds(workspaceAddr, workspaceSize, executor, ctx.stream())); - + GGML_CANN_CALL_ACLNN_OP(InplaceAdds, acl_tensor, other, alpha); return acl_tensor; } @@ -1157,13 +901,6 @@ void ggml_cann_rms_norm(ggml_backend_cann_context& ctx, ggml_tensor* dst) { float eps; memcpy(&eps, dst->op_params, sizeof(float)); - - GGML_ASSERT(eps > 0.0f); - - uint64_t workspaceSize = 0; - aclOpExecutor* executor; - void* workspaceAddr = nullptr; - size_t one_tensor_n_bytes = src->ne[0] * ggml_element_size(src); ggml_cann_pool_alloc one_tensor_allocator(ctx.pool(), one_tensor_n_bytes); @@ -1178,18 +915,7 @@ void ggml_cann_rms_norm(ggml_backend_cann_context& ctx, ggml_tensor* dst) { aclnn_zero(ctx, zero_tensor_allocator.get(), zero_tensor_n_bytes, src->ne, GGML_MAX_DIMS, ggml_cann_type_mapping(src->type), ggml_element_size(src)); - - ACL_CHECK(aclnnRmsNormGetWorkspaceSize( - acl_src, acl_gamma, eps, acl_dst, acl_rstd, &workspaceSize, &executor)); - - if (workspaceSize > 0) { - ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize); - workspaceAddr = workspace_allocator.get(); - } - - ACL_CHECK( - aclnnRmsNorm(workspaceAddr, workspaceSize, executor, ctx.stream())); - + GGML_CANN_CALL_ACLNN_OP(RmsNorm, acl_src, acl_gamma, eps, acl_dst, acl_rstd); ACL_CHECK(aclDestroyTensor(acl_src)); ACL_CHECK(aclDestroyTensor(acl_dst)); ACL_CHECK(aclDestroyTensor(acl_gamma)); @@ -1215,77 +941,19 @@ void ggml_cann_diag_mask(ggml_backend_cann_context& ctx, ggml_tensor* dst, src->ne, GGML_MAX_DIMS, ggml_cann_type_mapping(src->type), ggml_element_size(src), value); - uint64_t workspaceSize = 0; - aclOpExecutor* executor; - void* workspaceAddr = nullptr; - - ACL_CHECK(aclnnInplaceTriuGetWorkspaceSize(mask_tensor, n_past + 1, - &workspaceSize, &executor)); - if (workspaceSize > 0) { - ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize); - workspaceAddr = workspace_allocator.get(); - } - - ACL_CHECK( - aclnnInplaceTriu(workspaceAddr, workspaceSize, executor, ctx.stream())); - - ACL_CHECK(aclnnTrilGetWorkspaceSize(acl_src, n_past + 1, acl_dst, - &workspaceSize, &executor)); - if (workspaceSize > 0) { - ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize); - workspaceAddr = workspace_allocator.get(); - } - - ACL_CHECK(aclnnTril(workspaceAddr, workspaceSize, executor, ctx.stream())); - aclScalar* alpha = nullptr; float alphaValue = 1.0f; alpha = aclCreateScalar(&alphaValue, aclDataType::ACL_FLOAT); - ACL_CHECK(aclnnInplaceAddGetWorkspaceSize(acl_dst, mask_tensor, alpha, - &workspaceSize, &executor)); - if (workspaceSize > 0) { - ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize); - workspaceAddr = workspace_allocator.get(); - } - ACL_CHECK( - aclnnInplaceAdd(workspaceAddr, workspaceSize, executor, ctx.stream())); - + GGML_CANN_CALL_ACLNN_OP(InplaceTriu, mask_tensor, n_past + 1); + GGML_CANN_CALL_ACLNN_OP(Tril, acl_src, n_past + 1, acl_dst); + GGML_CANN_CALL_ACLNN_OP(InplaceAdd, acl_dst, mask_tensor, alpha); ACL_CHECK(aclDestroyScalar(alpha)); ACL_CHECK(aclDestroyTensor(mask_tensor)); ACL_CHECK(aclDestroyTensor(acl_src)); ACL_CHECK(aclDestroyTensor(acl_dst)); } -/** - * @brief Casts the data type of a source tensor to a destination tensor. - * - * This function casts the data type of the source tensor `acl_src` to the - * specified data type `cast_data_type` and stores the result in the destination - * tensor `acl_dst`. - * - * @param ctx The context for the CANN backend operations. - * @param acl_src The source tensor whose data type will be casted. - * @param acl_dst The destination tensor where the casted result will be stored. - * @param cast_data_type The target data type to which the source tensor will be - * casted. - */ -static void aclnn_cast(ggml_backend_cann_context& ctx, aclTensor* acl_src, - aclTensor* acl_dst, aclDataType cast_data_type) { - uint64_t workspaceSize = 0; - aclOpExecutor* executor; - void* workspaceAddr = nullptr; - - ACL_CHECK(aclnnCastGetWorkspaceSize(acl_src, cast_data_type, acl_dst, - &workspaceSize, &executor)); - if (workspaceSize > 0) { - ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize); - workspaceAddr = workspace_allocator.get(); - } - - ACL_CHECK(aclnnCast(workspaceAddr, workspaceSize, executor, ctx.stream())); -} - /** * @brief Permutes the dimensions of a tensor according to a specified order. * @@ -1304,40 +972,9 @@ static void aclnn_cast(ggml_backend_cann_context& ctx, aclTensor* acl_src, static void aclnn_permute(ggml_backend_cann_context& ctx, aclTensor* acl_src, aclTensor* acl_dst, int64_t* new_dim, uint64_t dims) { aclIntArray* acl_dims = aclCreateIntArray(new_dim, dims); - - uint64_t workspaceSize = 0; - aclOpExecutor* executor; - void* workspaceAddr = nullptr; - - ACL_CHECK(aclnnPermuteGetWorkspaceSize(acl_src, acl_dims, acl_dst, - &workspaceSize, &executor)); - if (workspaceSize > 0) { - ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize); - workspaceAddr = workspace_allocator.get(); - } - - ACL_CHECK( - aclnnPermute(workspaceAddr, workspaceSize, executor, ctx.stream())); - - ACL_CHECK(aclDestroyIntArray(acl_dims)); + GGML_CANN_CALL_ACLNN_OP(Permute, acl_src, acl_dims, acl_dst); } -#ifdef __cplusplus -extern "C" { -#endif -aclnnStatus aclnnIm2colGetWorkspaceSize(const aclTensor* self, - const aclIntArray* kernelSize, - const aclIntArray* dilation, - const aclIntArray* padding, - const aclIntArray* stride, - aclTensor* out, uint64_t* workspaceSize, - aclOpExecutor** executor); -aclnnStatus aclnnIm2col(void* workspace, uint64_t workspaceSize, - aclOpExecutor* executor, aclrtStream stream); -#ifdef __cplusplus -} -#endif - static void ggml_cann_im2col_2d_post_process(ggml_backend_cann_context& ctx, ggml_tensor* dst, ggml_tensor* src1, @@ -1501,23 +1138,8 @@ void ggml_cann_im2col(ggml_backend_cann_context& ctx, ggml_tensor* dst) { auto* dilations = aclCreateIntArray(dilation_size.data(), 2); auto* paddings = aclCreateIntArray(padding_dims.data(), 2); auto* strides = aclCreateIntArray(stride_dims.data(), 2); - - uint64_t workspaceSize = 0; - aclOpExecutor* executor; - void* workspaceAddr = nullptr; - - ACL_CHECK(aclnnIm2colGetWorkspaceSize(acl_src1, kernel_size, dilations, - paddings, strides, tmp_im2col_tensor, - &workspaceSize, &executor)); - - ggml_cann_pool_alloc workspace_allocator(ctx.pool()); - if (workspaceSize > 0) { - workspace_allocator.alloc(workspaceSize); - workspaceAddr = workspace_allocator.get(); - } - - ACL_CHECK( - aclnnIm2col(workspaceAddr, workspaceSize, executor, ctx.stream())); + GGML_CANN_CALL_ACLNN_OP(Im2col, acl_src1, kernel_size, dilations, + paddings, strides, tmp_im2col_tensor); // Cast if dst is f16. aclTensor* tmp_cast_tensor = nullptr; @@ -1536,8 +1158,7 @@ void ggml_cann_im2col(ggml_backend_cann_context& ctx, ggml_tensor* dst) { tmp_cast_buffer, ggml_cann_type_mapping(dst->type), ggml_type_size(dst->type), tmp_im2col_ne, temp_cast_nb, GGML_MAX_DIMS - 1, ACL_FORMAT_ND); - aclnn_cast(ctx, tmp_im2col_tensor, tmp_cast_tensor, - ggml_cann_type_mapping(dst->type)); + aclnn_cast(ctx, tmp_im2col_tensor, tmp_cast_tensor, dst); } // post-processing @@ -1575,285 +1196,17 @@ void ggml_cann_im2col(ggml_backend_cann_context& ctx, ggml_tensor* dst) { * @param acl_src The tensor on which the exponential function will be applied. */ static void aclnn_exp(ggml_backend_cann_context& ctx, aclTensor* acl_src) { - uint64_t workspaceSize = 0; - aclOpExecutor* executor; - void* workspaceAddr = nullptr; - - ACL_CHECK( - aclnnInplaceExpGetWorkspaceSize(acl_src, &workspaceSize, &executor)); - if (workspaceSize > 0) { - ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize); - workspaceAddr = workspace_allocator.get(); - } - - ACL_CHECK( - aclnnInplaceExp(workspaceAddr, workspaceSize, executor, ctx.stream())); + GGML_CANN_CALL_ACLNN_OP(InplaceExp, acl_src); } -/** - * @brief Multiplies elements of a tensor by a scalar value, optionally - * in-place. - * - * This function multiplies each element of the source tensor `acl_src` by the - * scalar `scale` and stores the result in the destination tensor `acl_dst`. If - * `inplace` is true, `acl_dst` will not be used and the operation is performed - * in-place on `acl_src`. - * The operation is defined as: - * \f[ - * \text {acl_dst }_i=\text {acl_src }_i \times \text {scale} - * \f] - * - * @param ctx The context for the CANN backend operations. - * @param acl_src The source tensor whose elements will be multiplied. - * @param scale The scalar value by which each element of `acl_src` will be - * multiplied. - * @param acl_dst The destination tensor where the result will be stored if - * `inplace` is false. - * @param inplace Flag indicating whether to perform the operation in-place on - * `acl_src`. - */ -static void aclnn_muls(ggml_backend_cann_context& ctx, aclTensor* acl_src, - float scale, aclTensor* acl_dst, bool inplace) { - aclScalar* acl_scale = aclCreateScalar(&scale, aclDataType::ACL_FLOAT); - - uint64_t workspaceSize = 0; - aclOpExecutor* executor; - void* workspaceAddr = nullptr; - - if (inplace) { - ACL_CHECK(aclnnInplaceMulsGetWorkspaceSize(acl_src, acl_scale, - &workspaceSize, &executor)); - if (workspaceSize > 0) { - ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize); - workspaceAddr = workspace_allocator.get(); - } - - ACL_CHECK(aclnnInplaceMuls(workspaceAddr, workspaceSize, executor, - ctx.stream())); - } else { - ACL_CHECK(aclnnMulsGetWorkspaceSize(acl_src, acl_scale, acl_dst, - &workspaceSize, &executor)); - if (workspaceSize > 0) { - ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize); - workspaceAddr = workspace_allocator.get(); - } - - ACL_CHECK( - aclnnMuls(workspaceAddr, workspaceSize, executor, ctx.stream())); - } - - ACL_CHECK(aclDestroyScalar(acl_scale)); -} - -/** - * @brief Performs an in-place element-wise multiplication of two tensors. - * - * This function performs an element-wise multiplication of the tensors - * `acl_src` and `acl_other` and stores the result in `acl_src`. - * The operation is defined as: - * \f[ - * \text {acl_src }_i=\text {acl_src }_i \times \text {acl_other }_i - * \f] - * - * @param ctx The context for the CANN backend operations. - * @param acl_src The source tensor where the multiplication result will be - * stored. - * @param acl_other The tensor whose elements will be multiplied with `acl_src`. - */ -static void aclnn_inplace_mul(ggml_backend_cann_context& ctx, - aclTensor* acl_src, aclTensor* acl_other) { - uint64_t workspaceSize = 0; - aclOpExecutor* executor; - void* workspaceAddr = nullptr; - - ACL_CHECK(aclnnInplaceMulGetWorkspaceSize(acl_src, acl_other, - &workspaceSize, &executor)); - if (workspaceSize > 0) { - ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize); - workspaceAddr = workspace_allocator.get(); - } - - ACL_CHECK( - aclnnInplaceMul(workspaceAddr, workspaceSize, executor, ctx.stream())); -} - -/** - * @brief Performs element-wise multiplication of two tensors and stores the - * result in a destination tensor. - * - * This function performs element-wise multiplication of the tensors `acl_src` - * and `acl_other` and stores the result in the destination tensor `acl_dst`. - * The operation is defined as: - * \f[ - * \text {acl_dst }_i=\text {acl_src }_i \times \text {acl_other }_i - * \f] - * - * @param ctx The context for the CANN backend operations. - * @param acl_src The first tensor for element-wise multiplication. - * @param acl_other The second tensor for element-wise multiplication. - * @param acl_dst The destination tensor where the result will be stored. - */ -static void aclnn_mul(ggml_backend_cann_context& ctx, aclTensor* acl_src, - aclTensor* acl_other, aclTensor* acl_dst) { - uint64_t workspaceSize = 0; - aclOpExecutor* executor; - void* workspaceAddr = nullptr; - - ACL_CHECK(aclnnMulGetWorkspaceSize(acl_src, acl_other, acl_dst, - &workspaceSize, &executor)); - if (workspaceSize > 0) { - ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize); - workspaceAddr = workspace_allocator.get(); - } - - ACL_CHECK(aclnnMul(workspaceAddr, workspaceSize, executor, ctx.stream())); -} - -/** - * @brief Applies element-wise cosine function to the elements of a tensor. - * - * This function computes the cosine of each element in the source tensor - * `acl_src` and stores the result in the destination tensor `acl_dst`. The - * operation is defined as: \f[ \text {acl_dst }_i=\cos \left(\text {acl_src - * }_i\right) \f] - * - * @param ctx The context for the CANN backend operations. - * @param acl_src The source tensor on which the cosine function will be - * applied. - * @param acl_dst The destination tensor where the cosine results will be - * stored. - */ -static void aclnn_cos(ggml_backend_cann_context& ctx, aclTensor* acl_src, +void aclnn_cos(ggml_backend_cann_context& ctx, aclTensor* acl_src, aclTensor* acl_dst) { - uint64_t workspaceSize = 0; - aclOpExecutor* executor; - void* workspaceAddr = nullptr; - - ACL_CHECK( - aclnnCosGetWorkspaceSize(acl_src, acl_dst, &workspaceSize, &executor)); - if (workspaceSize > 0) { - ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize); - workspaceAddr = workspace_allocator.get(); - } - - ACL_CHECK(aclnnCos(workspaceAddr, workspaceSize, executor, ctx.stream())); + GGML_CANN_CALL_ACLNN_OP(Cos, acl_src, acl_dst); } -/** - * @brief Applies element-wise sine function to the elements of a tensor. - * - * This function computes the sine of each element in the source tensor - `acl_src` - * and stores the result in the destination tensor `acl_dst`. - * The operation is defined as: - * \f[ - * \text {acl_dst }_i=\sin \left(\text {acl_src }_i\right) - * \f] - - * @param ctx The context for the CANN backend operations. - * @param acl_src The source tensor on which the sine function will be applied. - * @param acl_dst The destination tensor where the sine results will be stored. - */ -static void aclnn_sin(ggml_backend_cann_context& ctx, aclTensor* acl_src, +void aclnn_sin(ggml_backend_cann_context& ctx, aclTensor* acl_src, aclTensor* acl_dst) { - uint64_t workspaceSize = 0; - aclOpExecutor* executor; - void* workspaceAddr = nullptr; - - ACL_CHECK( - aclnnSinGetWorkspaceSize(acl_src, acl_dst, &workspaceSize, &executor)); - if (workspaceSize > 0) { - ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize); - workspaceAddr = workspace_allocator.get(); - } - - ACL_CHECK(aclnnSin(workspaceAddr, workspaceSize, executor, ctx.stream())); -} - -/** - * @brief Performs element-wise division of tensor1 by tensor2 , multiplies the - result by the scalar value and adds it to self . - * - * Performs element-wise division of tensor1 by tensor2, - * multiplies the result by the scalar value and adds it to self . - * The operation is defined as: - * \f[ - * \text{out}_i = \text{selft}_i + \text{value} \times - \frac{\text{tensor1}_i}{\text{tensor2}_i} - * \f] - - * @param ctx The context for the CANN backend operations. - * @param acl_self The source tensor on which the addcdiv function will be - applied. - * @param tensor1 Numerator tensor. - * @param tensor2 Denominator tensor. - * @param value The value to be used for coefficient. - */ -static void aclnn_inplace_addcdiv(ggml_backend_cann_context& ctx, - aclTensor* acl_self, aclTensor* tensor1, - aclTensor* tensor2, float value) { - uint64_t workspaceSize = 0; - aclOpExecutor* executor; - void* workspaceAddr = nullptr; - aclScalar* acl_value = aclCreateScalar(&value, aclDataType::ACL_FLOAT); - - ACL_CHECK(aclnnInplaceAddcdivGetWorkspaceSize( - acl_self, tensor1, tensor2, acl_value, &workspaceSize, &executor)); - if (workspaceSize > 0) { - ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize); - workspaceAddr = workspace_allocator.get(); - } - - ACL_CHECK(aclnnInplaceAddcdiv(workspaceAddr, workspaceSize, executor, - ctx.stream())); -} - -/** - * @brief Matrix division, optionally in-place. - * - * This function division each element of the source tensor `acl_src` by the - * tensor `acl_other` and stores the result in the destination tensor `acl_dst`. - * If `inplace` is true, `acl_dst` will not be used and the operation is - * performed in-place on `acl_src`. The operation is defined as: \f[ - * \text{dst}_i = \frac{\text{acl_src}_i}{\text{acl_other}_i} - * \f] - * - * @param ctx The context for the CANN backend operations. - * @param acl_src Numerator tensor.. - * @param acl_other Denominator tensor. - * @param acl_dst The destination tensor where the result will be stored if - * `inplace` is false. - * @param inplace Flag indicating whether to perform the operation in-place on - * `acl_src`. - */ -static void aclnn_div_tensor(ggml_backend_cann_context& ctx, aclTensor* acl_src, - aclTensor* acl_other, aclTensor* acl_dst, - bool inplace) { - uint64_t workspaceSize = 0; - aclOpExecutor* executor; - void* workspaceAddr = nullptr; - - if (inplace) { - ACL_CHECK(aclnnInplaceDivGetWorkspaceSize(acl_src, acl_other, - &workspaceSize, &executor)); - if (workspaceSize > 0) { - ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize); - workspaceAddr = workspace_allocator.get(); - } - - ACL_CHECK(aclnnInplaceDiv(workspaceAddr, workspaceSize, executor, - ctx.stream())); - } else { - ACL_CHECK(aclnnDivGetWorkspaceSize(acl_src, acl_other, acl_dst, - &workspaceSize, &executor)); - if (workspaceSize > 0) { - ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize); - workspaceAddr = workspace_allocator.get(); - } - - ACL_CHECK( - aclnnDiv(workspaceAddr, workspaceSize, executor, ctx.stream())); - } + GGML_CANN_CALL_ACLNN_OP(Sin, acl_src, acl_dst); } void ggml_cann_timestep_embedding(ggml_backend_cann_context& ctx, @@ -1983,20 +1336,7 @@ void ggml_cann_timestep_embedding(ggml_backend_cann_context& ctx, static void aclnn_fill_scalar(ggml_backend_cann_context& ctx, float scalar, aclTensor* acl_dst) { auto acl_scalar = aclCreateScalar(&scalar, aclDataType::ACL_FLOAT); - - uint64_t workspaceSize = 0; - aclOpExecutor* executor; - void* workspaceAddr = nullptr; - - ACL_CHECK(aclnnInplaceFillScalarGetWorkspaceSize( - acl_dst, acl_scalar, &workspaceSize, &executor)); - if (workspaceSize > 0) { - ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize); - workspaceAddr = workspace_allocator.get(); - } - - ACL_CHECK(aclnnInplaceFillScalar(workspaceAddr, workspaceSize, executor, - ctx.stream())); + GGML_CANN_CALL_ACLNN_OP(InplaceFillScalar, acl_dst, acl_scalar); ACL_CHECK(aclDestroyScalar(acl_scalar)); } @@ -2018,19 +1358,7 @@ static void aclnn_fill_scalar(ggml_backend_cann_context& ctx, float scalar, */ static void aclnn_pow_tensor_tensor(ggml_backend_cann_context& ctx, aclTensor* acl_dst, aclTensor* acl_exp) { - uint64_t workspaceSize = 0; - aclOpExecutor* executor; - void* workspaceAddr = nullptr; - - ACL_CHECK(aclnnInplacePowTensorTensorGetWorkspaceSize( - acl_dst, acl_exp, &workspaceSize, &executor)); - if (workspaceSize > 0) { - ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize); - workspaceAddr = workspace_allocator.get(); - } - - ACL_CHECK(aclnnInplacePowTensorTensor(workspaceAddr, workspaceSize, - executor, ctx.stream())); + GGML_CANN_CALL_ACLNN_OP(InplacePowTensorTensor, acl_dst, acl_exp); } /** @@ -2197,41 +1525,6 @@ void ggml_cann_cpy(ggml_backend_cann_context& ctx, ggml_tensor* dst) { ggml_cann_dup(ctx, dst); } -/** - * @brief Performs element-wise addition of two tensors in place. - * - * This function adds the source tensor `acl_src` to the destination tensor - * `acl_dst` element-wise and stores the result in the destination tensor - * `acl_dst`. - * - * @param ctx The context for the CANN backend operations. - * @param acl_src The source tensor to be added. - * @param acl_dst The destination tensor which will hold the result of the - * addition. - */ -static void aclnn_inplace_add(ggml_backend_cann_context& ctx, - aclTensor* acl_src, aclTensor* acl_dst) { - aclScalar* alpha = nullptr; - float alphaValue = 1.0f; - alpha = aclCreateScalar(&alphaValue, aclDataType::ACL_FLOAT); - - uint64_t workspaceSize = 0; - aclOpExecutor* executor; - void* workspaceAddr = nullptr; - - ACL_CHECK(aclnnInplaceAddGetWorkspaceSize(acl_dst, acl_src, alpha, - &workspaceSize, &executor)); - if (workspaceSize > 0) { - ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize); - workspaceAddr = workspace_allocator.get(); - } - - ACL_CHECK( - aclnnInplaceAdd(workspaceAddr, workspaceSize, executor, ctx.stream())); - - ACL_CHECK(aclDestroyScalar(alpha)); -} - /** * @brief Applies the softmax function to a tensor along a specified dimension. * @@ -2248,20 +1541,7 @@ static void aclnn_inplace_add(ggml_backend_cann_context& ctx, */ static void aclnn_softmax(ggml_backend_cann_context& ctx, aclTensor* acl_src, int64_t dim, aclTensor* acl_dst) { - uint64_t workspaceSize = 0; - aclOpExecutor* executor; - void* workspaceAddr = nullptr; - - ACL_CHECK(aclnnSoftmaxGetWorkspaceSize(acl_src, dim, acl_dst, - &workspaceSize, &executor)); - - if (workspaceSize > 0) { - ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize); - workspaceAddr = workspace_allocator.get(); - } - - aclrtStream stream = ctx.stream(); - ACL_CHECK(aclnnSoftmax(workspaceAddr, workspaceSize, executor, stream)); + GGML_CANN_CALL_ACLNN_OP(Softmax, acl_src, dim, acl_dst); } void ggml_cann_softmax(ggml_backend_cann_context& ctx, ggml_tensor* dst) { @@ -2378,85 +1658,152 @@ void ggml_cann_softmax(ggml_backend_cann_context& ctx, ggml_tensor* dst) { ACL_CHECK(aclDestroyTensor(tmp_mask_tensor)); } -void ggml_cann_get_rows(ggml_backend_cann_context& ctx, ggml_tensor* dst) { - ggml_tensor* src0 = dst->src[0]; - ggml_tensor* src1 = dst->src[1]; +/** + * @brief Performs embedding operation on a 4D tensor using the CANN backend. + * + * This function extracts slices from the source tensor (`src_buffer`), + * index tensor (`index`), and destination tensor (`dst`), and performs an + * embedding operation on them. The embedding operation is applied by iterating + * over the last two dimensions of the source tensor, creating the necessary + * tensors for the source, index, and output, and executing the embedding operation. + * + * @param ctx The context for CANN backend operations. + * @param src_buffer The source buffer holding the data for the source tensor. + * @param src_ne The dimensions of the source tensor. + * @param src_nb The strides (byte offsets) of the source tensor. + * @param index The index tensor used in the embedding operation. + * @param dst The destination tensor where the result will be stored. + */ +static void aclnn_embedding_4d(ggml_backend_cann_context& ctx, void* src_buffer, + int64_t* src_ne, size_t* src_nb, ggml_tensor* index, + ggml_tensor* dst) { + for (int64_t i = 0; i < src_ne[3]; i++) { + for (int64_t j = 0; j < src_ne[2]; j++) { + // src + int64_t acl_src_ne[2] = {src_ne[0], src_ne[1]}; + size_t acl_src_nb[2] = {src_nb[0], src_nb[1]}; + aclTensor* acl_src_tensor = ggml_cann_create_tensor( + (char*)src_buffer + i * src_nb[3] + j * src_nb[2], + ggml_cann_type_mapping(dst->type), ggml_element_size(dst), + acl_src_ne, acl_src_nb, 2); - ggml_cann_pool_alloc src0_extra_allocator(ctx.pool(), sizeof(ggml_tensor)); - ggml_cann_pool_alloc src1_extra_allocator(ctx.pool(), sizeof(ggml_tensor)); - ggml_cann_pool_alloc dst_extra_allocator(ctx.pool(), sizeof(ggml_tensor)); - src0->extra = src0_extra_allocator.get(); - src1->extra = src1_extra_allocator.get(); - dst->extra = dst_extra_allocator.get(); - ACL_CHECK(aclrtMemcpyAsync(src0->extra, sizeof(ggml_tensor), src0, - sizeof(ggml_tensor), ACL_MEMCPY_HOST_TO_DEVICE, - ctx.stream())); - ACL_CHECK(aclrtMemcpyAsync(src1->extra, sizeof(ggml_tensor), src1, - sizeof(ggml_tensor), ACL_MEMCPY_HOST_TO_DEVICE, - ctx.stream())); - ACL_CHECK(aclrtMemcpyAsync(dst->extra, sizeof(ggml_tensor), dst, - sizeof(ggml_tensor), ACL_MEMCPY_HOST_TO_DEVICE, - ctx.stream())); + // index + int64_t acl_index_ne[1] = {index->ne[0]}; + size_t acl_index_nb[1] = {index->nb[0]}; + aclTensor* acl_index = ggml_cann_create_tensor( + (char*)index->data + i * index->nb[2] + j * index->nb[1], + ggml_cann_type_mapping(index->type), ggml_element_size(index), + acl_index_ne, acl_index_nb, 1); + + // out + int64_t acl_out_ne[2] = {dst->ne[0], dst->ne[1]}; + size_t acl_out_nb[2] = {dst->nb[0], dst->nb[1]}; + aclTensor* acl_out = ggml_cann_create_tensor( + (char*)dst->data + i * dst->nb[3] + j * dst->nb[2], + ggml_cann_type_mapping(dst->type), ggml_element_size(dst), + acl_out_ne, acl_out_nb, 2); + GGML_CANN_CALL_ACLNN_OP(Embedding, acl_src_tensor, acl_index, acl_out); + ACL_CHECK(aclDestroyTensor(acl_src_tensor)); + ACL_CHECK(aclDestroyTensor(acl_index)); + ACL_CHECK(aclDestroyTensor(acl_out)); + } + } +} + +void ggml_cann_get_rows(ggml_backend_cann_context& ctx, ggml_tensor* dst) { + ggml_tensor* src0 = dst->src[0]; // src + ggml_tensor* src1 = dst->src[1]; // index switch (src0->type) { case GGML_TYPE_F32: { -#ifdef ASCEND_310P - // Special operation for get_row_f32 kernel of 310P: clear the - // content of dest data buffer when row is not aligned to 32 bytes - if ((src0->ne[0] % 8) != 0) { - size_t dst_len = src1->ne[0] * src1->ne[1] * src1->ne[2] * - src0->ne[0] * ggml_type_size(GGML_TYPE_F32); - ACL_CHECK(aclrtMemset((char*)dst->data, dst_len, 0, dst_len)); - } -#endif - aclrtlaunch_ascendc_get_row_f32( - 24, ctx.stream(), src0->data, src1->data, dst->data, - ((ggml_tensor*)src0->extra)->ne, - ((ggml_tensor*)src0->extra)->nb, - ((ggml_tensor*)src1->extra)->ne, - ((ggml_tensor*)src1->extra)->nb, ((ggml_tensor*)dst->extra)->ne, - ((ggml_tensor*)dst->extra)->nb); + aclnn_embedding_4d(ctx, src0->data, src0->ne, src0->nb, src1, + dst); break; } case GGML_TYPE_F16: { -#ifdef ASCEND_310P - // Special operation for get_row_f16 kernel of 310P: clear the - // content of dest data buffer when row is not aligned to 32 bytes - if ((src0->ne[0] % 16) != 0) { - size_t dst_len = - src1->ne[0] * src1->ne[1] * src1->ne[2] * src0->ne[0] * - ggml_type_size( - GGML_TYPE_F32); // out is also f32, even input is f16 - ACL_CHECK(aclrtMemset((char*)dst->data, dst_len, 0, dst_len)); + aclTensor* acl_src0 = ggml_cann_create_tensor(src0); + ggml_cann_pool_alloc src_buffer_allocator( + ctx.pool(), ggml_nelements(src0) * sizeof(float_t)); + void* src_trans_buffer = src_buffer_allocator.get(); + size_t src_trans_nb[GGML_MAX_DIMS]; + src_trans_nb[0] = sizeof(float_t); + for (int i = 1; i < GGML_MAX_DIMS; i++) { + src_trans_nb[i] = src_trans_nb[i - 1] * src0->ne[i - 1]; } -#endif - aclrtlaunch_ascendc_get_row_f16( - 24, ctx.stream(), src0->data, src1->data, dst->data, - ((ggml_tensor*)src0->extra)->ne, - ((ggml_tensor*)src0->extra)->nb, - ((ggml_tensor*)src1->extra)->ne, - ((ggml_tensor*)src1->extra)->nb, ((ggml_tensor*)dst->extra)->ne, - ((ggml_tensor*)dst->extra)->nb); + aclTensor* src_trans_tensor = ggml_cann_create_tensor( + src_trans_buffer, ACL_FLOAT, ggml_type_size(dst->type), + src0->ne, src_trans_nb, GGML_MAX_DIMS); + aclnn_cast(ctx, acl_src0, src_trans_tensor, dst); + aclnn_embedding_4d(ctx, src_trans_buffer, src0->ne, + src_trans_nb, src1, dst); + ACL_CHECK(aclDestroyTensor(acl_src0)); + ACL_CHECK(aclDestroyTensor(src_trans_tensor)); break; } - case GGML_TYPE_Q4_0: - aclrtlaunch_ascendc_get_row_q4_0( - 24, ctx.stream(), src0->data, src1->data, dst->data, - ((ggml_tensor*)src0->extra)->ne, - ((ggml_tensor*)src1->extra)->ne, - ((ggml_tensor*)src1->extra)->nb, ((ggml_tensor*)dst->extra)->ne, - ((ggml_tensor*)dst->extra)->nb); - break; - case GGML_TYPE_Q8_0: - aclrtlaunch_ascendc_get_row_q8_0( - 24, ctx.stream(), src0->data, src1->data, dst->data, - ((ggml_tensor*)src0->extra)->ne, - ((ggml_tensor*)src1->extra)->ne, - ((ggml_tensor*)src1->extra)->nb, ((ggml_tensor*)dst->extra)->ne, - ((ggml_tensor*)dst->extra)->nb); + case GGML_TYPE_Q8_0: { + // add 1 dim for bcast mul. + size_t weight_nb[GGML_MAX_DIMS + 1], scale_nb[GGML_MAX_DIMS + 1], + dequant_nb[GGML_MAX_DIMS + 1]; + int64_t weight_ne[GGML_MAX_DIMS + 1], scale_ne[GGML_MAX_DIMS + 1], + *dequant_ne; + int64_t scale_offset = 0; + + // [3,4,5,64] -> [3,4,5,2,32] + weight_ne[0] = QK8_0; + weight_ne[1] = src0->ne[0] / QK8_0; + weight_nb[0] = sizeof(int8_t); + weight_nb[1] = weight_nb[0] * weight_ne[0]; + for (int i = 2; i < GGML_MAX_DIMS + 1; i++) { + weight_ne[i] = src0->ne[i - 1]; + weight_nb[i] = weight_nb[i - 1] * weight_ne[i - 1]; + } + + // [3,4,5,64] -> [3,4,5,2,1] + scale_ne[0] = 1; + scale_ne[1] = src0->ne[0] / QK8_0; + scale_nb[0] = sizeof(uint16_t); + scale_nb[1] = scale_nb[0] * scale_ne[0]; + for (int i = 2; i < GGML_MAX_DIMS + 1; i++) { + scale_ne[i] = src0->ne[i - 1]; + scale_nb[i] = scale_nb[i - 1] * scale_ne[i - 1]; + } + + // [3,4,5,64] -> [3,4,5,2,32] + dequant_ne = weight_ne; + dequant_nb[0] = sizeof(float_t); + for (int i = 1; i < GGML_MAX_DIMS + 1; i++) { + dequant_nb[i] = dequant_nb[i - 1] * dequant_ne[i - 1]; + } + + scale_offset = ggml_nelements(src0) * sizeof(int8_t); + ggml_cann_pool_alloc dequant_buffer_allocator( + ctx.pool(), ggml_nelements(src0) * sizeof(float_t)); + + aclTensor* acl_weight_tensor = ggml_cann_create_tensor( + src0->data, ACL_INT8, sizeof(int8_t), weight_ne, weight_nb, + GGML_MAX_DIMS + 1); + aclTensor* acl_scale_tensor = ggml_cann_create_tensor( + src0->data, ACL_FLOAT16, sizeof(float16_t), scale_ne, scale_nb, + GGML_MAX_DIMS + 1, ACL_FORMAT_ND, scale_offset); + aclTensor* dequant_tensor = ggml_cann_create_tensor( + dequant_buffer_allocator.get(), ACL_FLOAT, sizeof(float_t), + dequant_ne, dequant_nb, GGML_MAX_DIMS + 1); + + aclnn_mul(ctx, acl_weight_tensor, acl_scale_tensor, dequant_tensor); + dequant_nb[0] = sizeof(float_t); + dequant_ne = src0->ne; + for (int i = 1; i < GGML_MAX_DIMS; i++) { + dequant_nb[i] = dequant_nb[i - 1] * src0->ne[i - 1]; + } + + aclnn_embedding_4d(ctx, dequant_buffer_allocator.get(), + dequant_ne, dequant_nb, src1, dst); + + ACL_CHECK(aclDestroyTensor(dequant_tensor)); break; + } default: - GGML_ABORT("fatal error"); + GGML_ABORT("Unsupported tensor type for GGML_OP_GET_ROWS"); break; } } @@ -2480,133 +1827,8 @@ static void aclnn_repeat_interleave(ggml_backend_cann_context& ctx, aclTensor* acl_src, aclTensor* acl_dst, int64_t dim, int64_t repeats, int64_t output_size) { - uint64_t workspaceSize = 0; - aclOpExecutor* executor; - void* workspaceAddr = nullptr; - - ACL_CHECK(aclnnRepeatInterleaveIntWithDimGetWorkspaceSize( - acl_src, repeats, dim, output_size, acl_dst, &workspaceSize, - &executor)); - if (workspaceSize > 0) { - ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize); - workspaceAddr = workspace_allocator.get(); - } - - ACL_CHECK(aclnnRepeatInterleaveIntWithDim(workspaceAddr, workspaceSize, - executor, ctx.stream())); -} - -/** - * @brief Performs matrix multiplication of two tensors. - * - * This function computes the matrix multiplication of the input tensor - * `acl_input` and the weight tensor `acl_weight`, and stores the result in the - * destination tensor `acl_dst`. - * The operation is defined as: - * \f[ - * \text {acl_dst}=\text {acl_input@acl_weight} - * \f] - * - * @param ctx The context for the CANN backend operations. - * @param acl_input The input tensor for the matrix multiplication. - * @param acl_weight The weight tensor for the matrix multiplication. - * @param acl_dst The destination tensor where the result of the matrix - * multiplication will be stored. - */ -static void aclnn_mat_mul(ggml_backend_cann_context& ctx, aclTensor* acl_input, - aclTensor* acl_weight, aclTensor* acl_dst) { - int8_t cube_math_type = 1; // ALLOW_FP32_DOWN_PRECISION, when input is - // fp32, atlas a2 will transpose it to HFLOAT32. - uint64_t workspaceSize = 0; - aclOpExecutor* executor; - void* workspaceAddr = nullptr; - - ACL_CHECK(aclnnMatmulGetWorkspaceSize(acl_input, acl_weight, acl_dst, - cube_math_type, &workspaceSize, - &executor)); - - if (workspaceSize > 0) { - ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize); - workspaceAddr = workspace_allocator.get(); - } - - ACL_CHECK( - aclnnMatmul(workspaceAddr, workspaceSize, executor, ctx.stream())); -} - -/** - * @brief Performs matrix multiplication of two 2D tensors. - * - * This function computes the matrix multiplication of the input tensor - * `acl_input` and the weight tensor `acl_weight`, and stores the result in the - * destination tensor `acl_dst`. - * The operation is defined as: - * \f[ - * \text {acl_dst}=\text {acl_input@acl_weight} - * \f] - * - * @param ctx The context for the CANN backend operations. - * @param acl_input The input tensor for the matrix multiplication. - * @param acl_weight The weight tensor for the matrix multiplication. - * @param acl_dst The destination tensor where the result of the matrix - * multiplication will be stored. - */ -static void aclnn_mat_mul_2d(ggml_backend_cann_context& ctx, - aclTensor* acl_input, aclTensor* acl_weight, - aclTensor* acl_dst) { - int8_t cube_math_type = 2; - uint64_t workspaceSize = 0; - aclOpExecutor* executor; - void* workspaceAddr = nullptr; - - ACL_CHECK(aclnnMmGetWorkspaceSize(acl_input, acl_weight, acl_dst, - cube_math_type, &workspaceSize, - &executor)); - - if (workspaceSize > 0) { - ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize); - workspaceAddr = workspace_allocator.get(); - } - - ACL_CHECK(aclnnMm(workspaceAddr, workspaceSize, executor, ctx.stream())); -} - -/** - * @brief Performs matrix multiplication of two 3D tensors. - * - * This function computes the matrix multiplication of the input tensor - * `acl_input` and the weight tensor `acl_weight`, and stores the result in the - * destination tensor `acl_dst`. - * The operation is defined as: - * \f[ - * \text {acl_dst}=\text {acl_input@acl_weight} - * \f] - * - * @param ctx The context for the CANN backend operations. - * @param acl_input The input tensor for the matrix multiplication. - * @param acl_weight The weight tensor for the matrix multiplication. - * @param acl_dst The destination tensor where the result of the matrix - * multiplication will be stored. - */ -static void aclnn_mat_mul_3d(ggml_backend_cann_context& ctx, - aclTensor* acl_input, aclTensor* acl_weight, - aclTensor* acl_dst) { - int8_t cube_math_type = 2; - uint64_t workspaceSize = 0; - aclOpExecutor* executor; - void* workspaceAddr = nullptr; - - ACL_CHECK(aclnnBatchMatMulGetWorkspaceSize(acl_input, acl_weight, acl_dst, - cube_math_type, &workspaceSize, - &executor)); - - if (workspaceSize > 0) { - ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize); - workspaceAddr = workspace_allocator.get(); - } - - ACL_CHECK( - aclnnBatchMatMul(workspaceAddr, workspaceSize, executor, ctx.stream())); + GGML_CANN_CALL_ACLNN_OP(RepeatInterleaveIntWithDim, acl_src, repeats, dim, + output_size, acl_dst); } /** @@ -2654,13 +1876,15 @@ static void ggml_cann_mat_mul_fp(ggml_backend_cann_context& ctx, switch (n_dims) { case 2: - aclnn_mat_mul_2d(ctx, acl_input_tensor, acl_weight_tensor, acl_dst); + GGML_CANN_CALL_ACLNN_OP(Mm, acl_input_tensor, acl_weight_tensor, acl_dst, 2); break; case 3: - aclnn_mat_mul_3d(ctx, acl_input_tensor, acl_weight_tensor, acl_dst); + GGML_CANN_CALL_ACLNN_OP(BatchMatMul, acl_input_tensor, acl_weight_tensor, acl_dst, 2); break; default: - aclnn_mat_mul(ctx, acl_input_tensor, acl_weight_tensor, acl_dst); + // ALLOW_FP32_DOWN_PRECISION, when input is + // fp32, atlas a2 will transpose it to HFLOAT32. + GGML_CANN_CALL_ACLNN_OP(Matmul, acl_input_tensor, acl_weight_tensor, acl_dst, 1); break; } @@ -2753,9 +1977,6 @@ static void ggml_cann_mul_mat_quant(ggml_backend_cann_context& ctx, int64_t max_elem_size = 65535; int64_t split_size = (src0->ne[1] / max_elem_size) + 1; ggml_cann_pool_alloc workspace_allocator(ctx.pool()); - aclOpExecutor* executor = nullptr; - uint64_t workspaceSize = 0; - void* workspaceAddr = nullptr; for (int64_t n1 = 0; n1 < src1->ne[3]; n1++) { for (int64_t c1 = 0; c1 < src1->ne[2]; c1++) { int64_t n0 = n1 / (src1->ne[3] / src0->ne[3]); @@ -2790,17 +2011,14 @@ static void ggml_cann_mul_mat_quant(ggml_backend_cann_context& ctx, (char*)output_buffer + batch1 * output_stride, ACL_FLOAT16, output_elem_size, output_ne, output_nb, 2, ACL_FORMAT_ND, output_ne_offset); - - ACL_CHECK(aclnnWeightQuantBatchMatmulV2GetWorkspaceSize( - acl_input_tensor, acl_weight_tensor, acl_scale_tensor, nullptr, - nullptr, nullptr, nullptr, QK8_0, acl_output_tensor, - &workspaceSize, &executor)); - if (workspaceAddr == nullptr) { - workspaceAddr = workspace_allocator.alloc(workspaceSize); + int64_t antiquantGroupSize = 0; + if (src0->ne[0] > QK8_0) { + antiquantGroupSize = QK8_0; } - ACL_CHECK(aclnnWeightQuantBatchMatmulV2( - workspaceAddr, workspaceSize, executor, ctx.stream())); - + GGML_CANN_CALL_ACLNN_OP(WeightQuantBatchMatmulV2, acl_input_tensor, + acl_weight_tensor, acl_scale_tensor, nullptr, + nullptr, nullptr, nullptr, antiquantGroupSize, + acl_output_tensor); ACL_CHECK(aclDestroyTensor(acl_weight_tensor)); ACL_CHECK(aclDestroyTensor(acl_scale_tensor)); ACL_CHECK(aclDestroyTensor(acl_output_tensor)); @@ -2830,14 +2048,10 @@ static void ggml_cann_mul_mat_quant(ggml_backend_cann_context& ctx, (char*)output_buffer + batch1 * output_stride, ACL_FLOAT16, output_elem_size, output_ne, output_nb, 2, ACL_FORMAT_ND, output_ne_offset); - - ACL_CHECK(aclnnWeightQuantBatchMatmulV2GetWorkspaceSize( - acl_input_tensor, acl_weight_tensor, acl_scale_tensor, - nullptr, nullptr, nullptr, nullptr, QK8_0, - acl_output_tensor, &workspaceSize, &executor)); - ACL_CHECK(aclnnWeightQuantBatchMatmulV2( - workspaceAddr, workspaceSize, executor, ctx.stream())); - + GGML_CANN_CALL_ACLNN_OP(WeightQuantBatchMatmulV2, acl_input_tensor, + acl_weight_tensor, acl_scale_tensor, nullptr, + nullptr, nullptr, nullptr, antiquantGroupSize, + acl_output_tensor); ACL_CHECK(aclDestroyTensor(acl_weight_tensor)); ACL_CHECK(aclDestroyTensor(acl_scale_tensor)); ACL_CHECK(aclDestroyTensor(acl_output_tensor)); @@ -2860,8 +2074,7 @@ static void ggml_cann_mul_mat_quant(ggml_backend_cann_context& ctx, output_buffer, ACL_FLOAT16, output_elem_size, output_cast_ne, output_cast_nb, GGML_MAX_DIMS); aclTensor* acl_dst_tensor = ggml_cann_create_tensor(dst); - aclnn_cast(ctx, acl_output_tensor, acl_dst_tensor, - ggml_cann_type_mapping(dst->type)); + aclnn_cast(ctx, acl_output_tensor, acl_dst_tensor, dst); ACL_CHECK(aclDestroyTensor(acl_output_tensor)); ACL_CHECK(aclDestroyTensor(acl_dst_tensor)); @@ -2880,7 +2093,7 @@ void ggml_cann_mul_mat(ggml_backend_cann_context& ctx, ggml_tensor* dst) { ggml_cann_mul_mat_quant(ctx, dst, type); break; default: - GGML_ABORT("fatal error"); + GGML_ABORT("Unsupported type for mul_mat"); break; } } @@ -2905,20 +2118,7 @@ static void aclnn_roll(ggml_backend_cann_context& ctx, aclTensor* acl_src, aclTensor* acl_dst, int64_t* shifts, int64_t* dims) { aclIntArray* acl_shifts = aclCreateIntArray(shifts, 1); aclIntArray* acl_dims = aclCreateIntArray(dims, 1); - - uint64_t workspaceSize = 0; - aclOpExecutor* executor; - void* workspaceAddr = nullptr; - - ACL_CHECK(aclnnRollGetWorkspaceSize(acl_src, acl_shifts, acl_dims, acl_dst, - &workspaceSize, &executor)); - if (workspaceSize > 0) { - ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize); - workspaceAddr = workspace_allocator.get(); - } - - ACL_CHECK(aclnnRoll(workspaceAddr, workspaceSize, executor, ctx.stream())); - + GGML_CANN_CALL_ACLNN_OP(Roll, acl_src, acl_shifts, acl_dims, acl_dst); ACL_CHECK(aclDestroyIntArray(acl_shifts)); ACL_CHECK(aclDestroyIntArray(acl_dims)); } @@ -2942,21 +2142,7 @@ static void aclnn_index_fill_tensor(ggml_backend_cann_context& ctx, float value) { aclIntArray* acl_index = aclCreateIntArray(index, index_num); aclScalar* acl_value = aclCreateScalar(&value, aclDataType::ACL_FLOAT); - - uint64_t workspaceSize = 0; - aclOpExecutor* executor; - void* workspaceAddr = nullptr; - - ACL_CHECK(aclnnInplaceIndexFillTensorGetWorkspaceSize( - acl_src, dim, acl_index, acl_value, &workspaceSize, &executor)); - if (workspaceSize > 0) { - ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize); - workspaceAddr = workspace_allocator.get(); - } - - ACL_CHECK(aclnnInplaceIndexFillTensor(workspaceAddr, workspaceSize, - executor, ctx.stream())); - + GGML_CANN_CALL_ACLNN_OP(InplaceIndexFillTensor, acl_src, dim, acl_index, acl_value); ACL_CHECK(aclDestroyIntArray(acl_index)); ACL_CHECK(aclDestroyScalar(acl_value)); } @@ -3015,8 +2201,7 @@ static void aclnn_cache_init(ggml_backend_cann_context& ctx, ggml_tensor* dst, aclTensor* acl_freq_factors_tensor = ggml_cann_create_tensor( src2->data, ggml_cann_type_mapping(src2->type), ggml_type_size(src2->type), arange_ne, arange_nb, GGML_MAX_DIMS); - aclnn_div_tensor(ctx, acl_theta_scale_tensor, acl_freq_factors_tensor, - nullptr, true); + aclnn_div(ctx, acl_theta_scale_tensor, acl_freq_factors_tensor); ACL_CHECK(aclDestroyTensor(acl_freq_factors_tensor)); } @@ -3133,7 +2318,7 @@ void ggml_cann_rope(ggml_backend_cann_context& ctx, ggml_tensor* dst) { // TODO: use ascendc // Only test with LLAMA model. ggml_tensor* src0 = dst->src[0]; // input - ggml_tensor* src2 = dst->src[2]; // freq_factors + // ggml_tensor* src2 = dst->src[2]; // freq_factors, not used now. // param float freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow; @@ -3315,8 +2500,8 @@ void ggml_cann_rope(ggml_backend_cann_context& ctx, ggml_tensor* dst) { // output void* output_fp32_buffer; if (src0->type == GGML_TYPE_F32) { - aclnn_inplace_mul(ctx, acl_src, acl_cos_reshape_tensor); - aclnn_inplace_mul(ctx, acl_input_roll_mul_scale_tensor, + aclnn_mul(ctx, acl_src, acl_cos_reshape_tensor); + aclnn_mul(ctx, acl_input_roll_mul_scale_tensor, acl_sin_reshape_tensor); aclnn_add(ctx, acl_src, acl_input_roll_mul_scale_tensor, acl_dst); // TODO: ne0 != n_dims in mode2 @@ -3389,39 +2574,160 @@ void ggml_cann_rope(ggml_backend_cann_context& ctx, ggml_tensor* dst) { ggml_type_size(src0->type), sin_final_ne, sin_final_nb, GGML_MAX_DIMS); - aclnn_cast(ctx, acl_sin_reshape_tensor, acl_sin_final_tensor, - ggml_cann_type_mapping(src0->type)); - aclnn_cast(ctx, acl_cos_reshape_tensor, acl_cos_final_tensor, - ggml_cann_type_mapping(src0->type)); + aclnn_cast(ctx, acl_sin_reshape_tensor, acl_sin_final_tensor, dst); + aclnn_cast(ctx, acl_cos_reshape_tensor, acl_cos_final_tensor, dst); ACL_CHECK(aclDestroyTensor(acl_cos_reshape_tensor)); ACL_CHECK(aclDestroyTensor(acl_sin_reshape_tensor)); acl_sin_reshape_tensor = acl_sin_final_tensor; acl_cos_reshape_tensor = acl_cos_final_tensor; } - uint64_t workspaceSize = 0; - aclOpExecutor* executor; - - void* workspaceAddr = nullptr; - int acl_mode = mode; if (mode == 0) { acl_mode = 1; } - ACL_CHECK(aclnnRotaryPositionEmbeddingGetWorkspaceSize( - acl_src, acl_cos_reshape_tensor, acl_sin_reshape_tensor, acl_mode, - acl_dst, &workspaceSize, &executor)); - if (workspaceSize > 0) { - ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize); - workspaceAddr = workspace_allocator.get(); - } - - ACL_CHECK(aclnnRotaryPositionEmbedding(workspaceAddr, workspaceSize, - executor, ctx.stream())); - + GGML_CANN_CALL_ACLNN_OP(RotaryPositionEmbedding, acl_src, acl_cos_reshape_tensor, + acl_sin_reshape_tensor, acl_mode, acl_dst); ACL_CHECK(aclDestroyTensor(acl_src)); ACL_CHECK(aclDestroyTensor(acl_cos_reshape_tensor)); ACL_CHECK(aclDestroyTensor(acl_sin_reshape_tensor)); ACL_CHECK(aclDestroyTensor(acl_dst)); } + + + void ggml_cann_argmax(ggml_backend_cann_context& ctx, ggml_tensor* dst){ + ggml_tensor * src0 = dst->src[0]; + + aclTensor* acl_src = ggml_cann_create_tensor(src0); + aclTensor* acl_dst = ggml_cann_create_tensor(dst, dst->ne, dst->nb, 3); + + GGML_CANN_CALL_ACLNN_OP(ArgMax, acl_src, 3, false, acl_dst); + + ACL_CHECK(aclDestroyTensor(acl_src)); + ACL_CHECK(aclDestroyTensor(acl_dst)); +} + +void ggml_cann_conv_transpose_1d(ggml_backend_cann_context& ctx, ggml_tensor* dst){ + ggml_tensor * src0 = dst->src[0]; + ggml_tensor * src1 = dst->src[1]; + + // stride + int64_t s0 = ((const int32_t*)(dst->op_params))[0]; + + aclTensor* acl_input = ggml_cann_create_tensor(src1, src1->ne, src1->nb, 3, ACL_FORMAT_NCL); + aclTensor* acl_weight = ggml_cann_create_tensor(src0, src0->ne, src0->nb, 3, ACL_FORMAT_NCL); + aclTensor* acl_dst = ggml_cann_create_tensor(dst, dst->ne, dst->nb, 3, ACL_FORMAT_NCL); + + int64_t strideVal[1]; + strideVal[0] = s0; + aclIntArray *stride = aclCreateIntArray(strideVal, 1); + int64_t paddingVal[] = {0}; + aclIntArray *padding = aclCreateIntArray(paddingVal, 1); + int64_t dilationVal[] = {1}; + aclIntArray *dilation = aclCreateIntArray(dilationVal, 1); + bool transposed = true; + int64_t groups = 1; + int8_t cubeMathType = 0; + + GGML_CANN_CALL_ACLNN_OP(Convolution, acl_input, acl_weight, nullptr, stride, + padding, dilation, transposed, padding, groups, acl_dst, cubeMathType); + + ACL_CHECK(aclDestroyTensor(acl_weight)); + ACL_CHECK(aclDestroyTensor(acl_dst)); + ACL_CHECK(aclDestroyIntArray(stride)); + ACL_CHECK(aclDestroyIntArray(padding)); + ACL_CHECK(aclDestroyIntArray(dilation)); +} + +void ggml_cann_elu(ggml_backend_cann_context& ctx, ggml_tensor* dst){ + ggml_tensor * src0 = dst->src[0]; + + aclTensor* acl_input = ggml_cann_create_tensor(src0); + aclTensor* acl_dst = ggml_cann_create_tensor(dst); + + float alphaValue = 1.0f; + aclScalar* alpha = nullptr; + alpha = aclCreateScalar(&alphaValue, aclDataType::ACL_FLOAT); + + GGML_CANN_CALL_ACLNN_OP(Elu, acl_input, alpha, alpha, alpha, + acl_dst); + + ACL_CHECK(aclDestroyTensor(acl_input)); + ACL_CHECK(aclDestroyTensor(acl_dst)); + ACL_CHECK(aclDestroyScalar(alpha)); +} + +void ggml_cann_mean(ggml_backend_cann_context& ctx, ggml_tensor* dst){ + ggml_tensor * src0 = dst->src[0]; + + aclTensor* acl_src = ggml_cann_create_tensor(src0); + aclTensor* acl_dst = ggml_cann_create_tensor(dst); + + int64_t reduceDimValue[] = {3}; + aclIntArray* reduceDim = aclCreateIntArray(reduceDimValue, 1); + bool keepDim = true; + + GGML_CANN_CALL_ACLNN_OP(Mean, acl_src, reduceDim, keepDim, ACL_FLOAT, acl_dst); + + ACL_CHECK(aclDestroyTensor(acl_src)); + ACL_CHECK(aclDestroyTensor(acl_dst)); + ACL_CHECK(aclDestroyIntArray(reduceDim)); +} + +void ggml_cann_pad_reflect_1d(ggml_backend_cann_context& ctx, ggml_tensor* dst){ + ggml_tensor * src0 = dst->src[0]; + int32_t *opts = (int32_t *) dst->op_params; + int64_t paddingsArray[2] = {opts[0], opts[1]}; + aclIntArray* paddings = aclCreateIntArray(paddingsArray, 2); + + for (int64_t i = 0; i < src0->ne[3]; i++) { + aclTensor* acl_src = ggml_cann_create_tensor( + (char*)src0->data + i * src0->ne[3], + ggml_cann_type_mapping(src0->type), ggml_element_size(src0), + src0->ne, src0->nb, 3); + + aclTensor* acl_dst = ggml_cann_create_tensor( + (char*)dst->data + i * src0->ne[3], + ggml_cann_type_mapping(dst->type), ggml_element_size(dst), + dst->ne, dst->nb, 3); + + GGML_CANN_CALL_ACLNN_OP(ReflectionPad1d, acl_src, paddings, acl_dst); + + ACL_CHECK(aclDestroyTensor(acl_src)); + ACL_CHECK(aclDestroyTensor(acl_dst)); + } + ACL_CHECK(aclDestroyIntArray(paddings)); +} + +void ggml_cann_count_equal(ggml_backend_cann_context& ctx, ggml_tensor* dst){ + ggml_tensor * src0 = dst->src[0]; + ggml_tensor * src1 = dst->src[1]; + + aclTensor* acl_self = ggml_cann_create_tensor(src0); + aclTensor* acl_other = ggml_cann_create_tensor(src1); + + GGML_CANN_CALL_ACLNN_OP(InplaceEqTensor, acl_self, acl_other); + + ggml_cann_sum(ctx, dst); + + ACL_CHECK(aclDestroyTensor(acl_self)); + ACL_CHECK(aclDestroyTensor(acl_other)); +} + +void ggml_cann_step(ggml_backend_cann_context& ctx, ggml_tensor* dst){ + ggml_tensor * src0 = dst->src[0]; + + aclTensor* acl_src = ggml_cann_create_tensor(src0); + aclTensor* acl_dst = ggml_cann_create_tensor(dst); + + float alphaValue = 0.0f; + aclScalar* alpha = nullptr; + alpha = aclCreateScalar(&alphaValue, aclDataType::ACL_FLOAT); + + GGML_CANN_CALL_ACLNN_OP(GtScalar, acl_src, alpha, acl_dst); + + ACL_CHECK(aclDestroyTensor(acl_src)); + ACL_CHECK(aclDestroyTensor(acl_dst)); + ACL_CHECK(aclDestroyScalar(alpha)); +} diff --git a/ggml/src/ggml-cann/aclnn_ops.h b/ggml/src/ggml-cann/aclnn_ops.h index 680129c76..b2d1b3c36 100644 --- a/ggml/src/ggml-cann/aclnn_ops.h +++ b/ggml/src/ggml-cann/aclnn_ops.h @@ -1,15 +1,4 @@ -#ifndef CANN_ACLNN_OPS -#define CANN_ACLNN_OPS - /** - * @file acl_tensor - * @brief This file contains related functions of ggml_tensor and acl_tensor. - * Contains conversion from ggml_tensor to acl_tensor, broadcast and other - * functions. - * @author hipudding - * @author wangshuai09 <391746016@qq.com> - * @date July 15, 2024 - * * Copyright (c) 2023-2024 The ggml authors * * Permission is hereby granted, free of charge, to any person obtaining a copy @@ -31,20 +20,30 @@ * IN THE SOFTWARE. */ -#include +#ifndef CANN_ACLNN_OPS +#define CANN_ACLNN_OPS + +#include +#include +#include #include #include #include #include -#include #include +#include +#include #include #include #include -#include #include #include #include +#include +#include +#include +#include +#include #include "acl_tensor.h" #include "common.h" @@ -63,23 +62,6 @@ */ void ggml_cann_repeat(ggml_backend_cann_context& ctx, ggml_tensor* dst); -/** - * @brief Adds two ggml tensors using the CANN backend. - * - * @details This function performs an element-wise addition of two tensors. In - * case the tensors do not have the same shape, one or both tensors - * will be broadcasted to match the shape of the other before the - * addition is performed.The formula for the operation is given by: - * \f[ - * \text{dst} = \text{acl_src0} + \alpha \cdot \text{acl_src1} - * \f] - * - * @param ctx The CANN context used for operations. - * @param dst The ggml tensor representing the destination, result of the - * addition is stored at dst->data, and dst->op is `GGML_OP_ADD` - */ -void ggml_cann_add(ggml_backend_cann_context& ctx, ggml_tensor* dst); - /** * @brief Applies the Leaky ReLU activation function to a tensor using the CANN * backend. @@ -131,19 +113,6 @@ void ggml_cann_concat(ggml_backend_cann_context& ctx, ggml_tensor* dst); */ void ggml_cann_arange(ggml_backend_cann_context& ctx, ggml_tensor* dst); -/** - * @brief Computes the square of the elements of a ggml tensor using the CANN - * backend. - * @details The function sets the second source tensor of the destination - * tensor `dst` to be equal to the first source tensor. This is - * effectively squaring the elements since the multiplication becomes - * `element * element`. - * @param ctx The CANN context used for operations. - * @param dst The destination tensor where the squared values will be stored， - * which dst->op is `GGML_OP_SQR`. - */ -void ggml_cann_sqr(ggml_backend_cann_context& ctx, ggml_tensor* dst); - /** * @brief Applies a clamp operation to the elements of a ggml tensor using the * CANN backend. @@ -275,6 +244,20 @@ void ggml_cann_acc(ggml_backend_cann_context& ctx, ggml_tensor* dst); */ void ggml_cann_sum_rows(ggml_backend_cann_context& ctx, ggml_tensor* dst); +/** + * @brief Computes the sum of elements in a ggml tensor. + * + * @details This function performs a reduction sum operation along the last + * dimension of the input tensor `src`. The result of the sum is stored + * in the destination tensor `dst`. + * + * @param ctx The CANN context used for operations. + * @param dst The destination tensor where the reduced values will be stored。 + * + */ + +void ggml_cann_sum(ggml_backend_cann_context& ctx, ggml_tensor* dst); + /** * @brief Upsamples a ggml tensor using nearest neighbor interpolation using * the CANN backend. @@ -484,109 +467,387 @@ void ggml_cann_mul_mat(ggml_backend_cann_context& ctx, ggml_tensor* dst); */ void ggml_cann_rope(ggml_backend_cann_context& ctx, ggml_tensor* dst); -template -void ggml_cann_mul_div(ggml_backend_cann_context& ctx, ggml_tensor* dst) { +/** + * @brief Computes the index of the maximum value along the specified dimension + * of a ggml tensor using the CANN backend. + * + * @details This function performs an argmax operation on the input tensor. + * It finds the index of the maximum value along the specified axis + * and stores these indices in the destination tensor `dst`. The + * operation is executed using the CANN backend for optimized performance. + * + * @param ctx The CANN context used for operations. + * @param dst The destination tensor where the indices of the maximum values will + * be stored. dst->op is `GGML_OP_ARGMAX`. + */ +void ggml_cann_argmax(ggml_backend_cann_context& ctx, ggml_tensor* dst); + +/** + * @brief Adds two tensors element-wise and stores the result in a destination + * tensor. + * + * This function performs the operation: + * \f[ + * dst = acl\_src0 + alpha \times acl\_src1 + * \f] + * where alpha is a scalar value and defaults to 1.0f. + * + * @param ctx The context for the CANN backend operations. + * @param acl_src0 The first source tensor. + * @param acl_src1 The second source tensor. + * @param acl_dst The destination tensor where the result will be stored. + */ +void aclnn_add(ggml_backend_cann_context& ctx, aclTensor* acl_src0, + aclTensor* acl_src1, aclTensor* acl_dst = nullptr); + +/** + * @brief Sub two tensors element-wise and stores the result in a destination + * tensor. + * + * This function performs the operation: + * \f[ + * dst = acl\_src0 - alpha \times acl\_src1 + * \f] + * where alpha is a scalar value and defaults to 1.0f. + * + * @param ctx The context for the CANN backend operations. + * @param acl_src0 The first source tensor. + * @param acl_src1 The second source tensor. + * @param acl_dst The destination tensor where the result will be stored. + */ +void aclnn_sub(ggml_backend_cann_context& ctx, aclTensor* acl_src0, + aclTensor* acl_src1, aclTensor* acl_dst = nullptr); + +/** + * @brief Performs element-wise multiplication of two tensors and stores the + * result in a destination tensor. + * + * This function performs element-wise multiplication of the tensors `acl_src` + * and `acl_other` and stores the result in the destination tensor `acl_dst`. + * The operation is defined as: + * \f[ + * \text {acl_dst }_i=\text {acl_src }_i \times \text {acl_other }_i + * \f] + * + * @param ctx The context for the CANN backend operations. + * @param acl_src The first tensor for element-wise multiplication. + * @param acl_other The second tensor for element-wise multiplication. + * @param acl_dst The destination tensor where the result will be stored. + */ +void aclnn_mul(ggml_backend_cann_context& ctx, aclTensor* acl_src, + aclTensor* acl_other, aclTensor* acl_dst = nullptr); + +/** + * @brief Matrix division, optionally in-place. + * + * This function division each element of the source tensor `acl_src` by the + * tensor `acl_other` and stores the result in the destination tensor `acl_dst`. + * If `inplace` is true, `acl_dst` will not be used and the operation is + * performed in-place on `acl_src`. The operation is defined as: \f[ + * \text{dst}_i = \frac{\text{acl_src}_i}{\text{acl_other}_i} + * \f] + * + * @param ctx The context for the CANN backend operations. + * @param acl_src Numerator tensor.. + * @param acl_other Denominator tensor. + * @param acl_dst The destination tensor where the result will be stored if + * `inplace` is false. + * @param inplace Flag indicating whether to perform the operation in-place on + * `acl_src`. + */ +void aclnn_div(ggml_backend_cann_context& ctx, aclTensor* acl_src, + aclTensor* acl_other, aclTensor* acl_dst = nullptr); + +/** + * @brief Applies element-wise cosine function to the elements of a tensor. + * + * This function computes the cosine of each element in the source tensor + * `acl_src` and stores the result in the destination tensor `acl_dst`. The + * operation is defined as: \f[ \text {acl_dst }_i=\cos \left(\text {acl_src + * }_i\right) \f] + * + * @param ctx The context for the CANN backend operations. + * @param acl_src The source tensor on which the cosine function will be + * applied. + * @param acl_dst The destination tensor where the cosine results will be + * stored. + */ +void aclnn_cos(ggml_backend_cann_context& ctx, aclTensor* acl_src, + aclTensor* acl_dst); + +/** + * @brief Applies element-wise sine function to the elements of a tensor. + * + * This function computes the sine of each element in the source tensor + `acl_src` + * and stores the result in the destination tensor `acl_dst`. + * The operation is defined as: + * \f[ + * \text {acl_dst }_i=\sin \left(\text {acl_src }_i\right) + * \f] + + * @param ctx The context for the CANN backend operations. + * @param acl_src The source tensor on which the sine function will be applied. + * @param acl_dst The destination tensor where the sine results will be stored. + */ +void aclnn_sin(ggml_backend_cann_context& ctx, aclTensor* acl_src, + aclTensor* acl_dst); + +/** + * @brief Prepares broadcast-compatible ACL tensors for two input tensors and one + * output tensor. + * + * This function checks whether broadcasting is needed between `src0` and `src1`. + * If broadcasting is required, it calculates the proper shapes and creates + * ACL tensors with broadcast parameters. Otherwise, it directly creates ACL tensors + * based on the original tensor shapes. + * + * @param src0 The first input tensor (reference shape). + * @param src1 The second input tensor (possibly broadcasted). + * @param dst The destination/output tensor. + * @param acl_src0 Output pointer to the created ACL tensor corresponding to src0. + * @param acl_src1 Output pointer to the created ACL tensor corresponding to src1. + * @param acl_dst Output pointer to the created ACL tensor corresponding to dst. + */ +void bcast_shape(ggml_tensor * src0, ggml_tensor * src1, ggml_tensor * dst, + aclTensor ** acl_src0, aclTensor ** acl_src1, aclTensor ** acl_dst); + +/** + * @brief Computes the 1D transposed convolution (deconvolution) of a ggml + * tensor using the CANN backend. + * + * @details This function performs a 1D transposed convolution (also known as + * deconvolution) operation on the input tensor. The computed result is stored + * in the destination tensor `dst`. The operation is optimized using the CANN + * backend for improved performance. + * + * @param ctx The CANN context used for operations. + * @param dst The destination tensor where the transposed convolution result + * will be stored. dst->op is `GGML_OP_CONV_TRANSPOSE_1D`. + */ +void ggml_cann_conv_transpose_1d(ggml_backend_cann_context& ctx, ggml_tensor* dst); + +/** + * @brief Applies the ELU (Exponential Linear Unit) activation to a ggml tensor + * using the CANN backend. + * + * @details This function performs an element-wise ELU activation on the input + * tensor. + * The result is written to the destination tensor `dst` in-place. + * The ELU function is defined as: + * + * \text{ELU}(x) = + * \begin{cases} + * x, & \text{if } x > 0 \\ + * \alpha \left( \exp(x) - 1 \right), & \text{if } x \leq 0 + * \end{cases} + * + * where α (alpha) is a hyperparameter, typically set to 1.0. + * This operation is optimized using the CANN backend for high-performance + * inference or training. + * + * @param ctx The CANN context used for operations. + * @param dst The destination tensor where the ELU-activated result will be stored. + * dst->op is expected to be `GGML_OP_ELU`. + */ +void ggml_cann_elu(ggml_backend_cann_context& ctx, ggml_tensor* dst); + +/** + * @brief Computes the mean of a ggml tensor element-wise using the CANN backend. + * + * @details This function calculates the element-wise mean of the input tensor. + * The result is written to the destination tensor `dst`. + * The mean is computed by averaging the values across the entire tensor. + * + * This operation is optimized using the CANN backend for high-performance inference or training. + * + * @param ctx The CANN context used for operations. + * @param dst The destination tensor where the mean result will be stored. + * dst->op is expected to be `GGML_OP_MEAN`. + */ +void ggml_cann_mean(ggml_backend_cann_context& ctx, ggml_tensor* dst); + +/** + * @brief Applies 1D reflect padding to a ggml tensor using the CANN backend. + * + * @details This function performs 1D reflect padding on the input tensor. + * The amount of padding on each side is specified by parameters stored in `dst->op_params`. + * The operation reflects the values at the borders of the tensor to generate the padded output. + * + * This operation is optimized using the CANN backend for high-performance inference or training. + * + * @param ctx The CANN context used for operations. + * @param dst The destination tensor where the padded result will be stored. + * dst->op is expected to be `GGML_OP_PAD_REFLECT_1D`. + */ +void ggml_cann_pad_reflect_1d(ggml_backend_cann_context& ctx, ggml_tensor* dst); + +/** + * @brief Counts the number of equal elements in two ggml tensors using the CANN backend. + * + * @details This function performs an element-wise comparison between two input tensors, + * and counts the number of positions where the elements are equal. The result is + * stored in the destination tensor `dst` as a scalar. + * + * The operation is optimized using the CANN backend, making it suitable for + * high-performance inference or training scenarios. + * + * @param ctx The CANN context used for operations. + * @param dst The destination tensor where the result will be stored. + * dst->op is expected to be `GGML_OP_COUNT_EQUAL`. + */ +void ggml_cann_count_equal(ggml_backend_cann_context& ctx, ggml_tensor* dst); + +/** + * @brief Applies the Step activation function to a ggml tensor using the CANN backend. + * + * @details This function applies a step function element-wise to the input tensor, where + * each element is transformed to 1.0 if it is greater than 0, and 0.0 otherwise. + * The result is stored in the destination tensor `dst`. + * + * This operation is accelerated using the CANN backend to improve runtime performance. + * + * @param ctx The CANN context used for operations. + * @param dst The destination tensor where the result will be stored. + * dst->op is expected to be `GGML_OP_STEP`. + */ +void ggml_cann_step(ggml_backend_cann_context& ctx, ggml_tensor* dst); + +/** + * @brief Applies a element-wise operation to two input tensors using the CANN + * backend. + * + * This templated function takes a binary operator and applies it to two source + * tensors + * associated with the destination tensor. The function handles broadcasting as + * needed. + * + * @tparam binary_op A callable object (e.g., lambda or function pointer) representing + * the binary operation to be performed. It must take three arguments: + * (ggml_backend_cann_context&, aclTensor*, aclTensor*, aclTensor*). + * + * @param ctx The CANN backend context used to manage execution and resources. + * @param dst The destination tensor. + */ +template +void ggml_cann_binary_op(ggml_backend_cann_context& ctx, ggml_tensor* dst) { ggml_tensor* src0 = dst->src[0]; ggml_tensor* src1 = dst->src[1]; - GGML_ASSERT(ggml_can_repeat(src1, src0) && ggml_are_same_shape(src0, dst)); aclTensor* acl_src0; aclTensor* acl_src1; aclTensor* acl_dst; // Need bcast - if (!ggml_are_same_shape(src0, src1) && ggml_cann_need_bcast(src0, src1)) { - BCAST_SHAPE(src0, src1) - acl_src0 = ggml_cann_create_tensor(src0, BCAST_PARAM(src0)); - acl_src1 = ggml_cann_create_tensor(src1, BCAST_PARAM(src1)); - acl_dst = ggml_cann_create_tensor(dst, BCAST_PARAM(src0)); - } else { - acl_src0 = ggml_cann_create_tensor(src0); - acl_src1 = ggml_cann_create_tensor(src1); - acl_dst = ggml_cann_create_tensor(dst); - } - - uint64_t workspaceSize = 0; - aclOpExecutor* executor; - void* workspaceAddr = nullptr; - - ACL_CHECK(getWorkspaceSize(acl_src0, acl_src1, acl_dst, &workspaceSize, - &executor)); - if (workspaceSize > 0) { - ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize); - workspaceAddr = workspace_allocator.get(); - } - - aclrtStream main_stream = ctx.stream(); - ACL_CHECK(execute(workspaceAddr, workspaceSize, executor, main_stream)); + bcast_shape(src0, src1, dst, &acl_src0, &acl_src1, &acl_dst); + binary_op(ctx, acl_src0, acl_src1, acl_dst); ACL_CHECK(aclDestroyTensor(acl_src0)); ACL_CHECK(aclDestroyTensor(acl_src1)); ACL_CHECK(aclDestroyTensor(acl_dst)); } -// Activation functions template. -template -void ggml_cann_activation(ggml_backend_cann_context& ctx, ggml_tensor* dst) { - ggml_tensor* src = dst->src[0]; +/** + * @brief Launches an asynchronous task using the memory allocator. + * + * This macro submit an asynchronous task on the specified stream. + * The task uses memory allocated by the allocator. It is guaranteed + * that the memory will not be accessed by other tasks until this task + * completes, due to the sequential execution order within the same stream. + * + * @param OP_NAME aclnn operator name. + * @param args Additional arguments required by the task. + * + * @note + * Memory from the allocator will be "freed" immediately and can be + * reallocated to other pointers. However, it won't be accessed by any + * other task before this asynchronous task ends, because all tasks in the + * same stream are executed in queue order. + */ +#define GGML_CANN_CALL_ACLNN_OP(OP_NAME, ...) \ + do { \ + uint64_t workspaceSize = 0; \ + aclOpExecutor * executor; \ + void * workspaceAddr = nullptr; \ + \ + ACL_CHECK(aclnn##OP_NAME##GetWorkspaceSize(__VA_ARGS__, &workspaceSize, &executor)); \ + \ + if (workspaceSize > 0) { \ + ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize); \ + workspaceAddr = workspace_allocator.get(); \ + } \ + ACL_CHECK(aclnn##OP_NAME(workspaceAddr, workspaceSize, executor, ctx.stream())); \ + } while (0) - GGML_ASSERT(src->type == GGML_TYPE_F32); - GGML_ASSERT(dst->type == GGML_TYPE_F32); +/** + * @brief Applies a unary operation to an input tensor using the CANN backend. + * + * This templated function applies a unary operator to the source tensor of `dst` + * and stores the result in the destination tensor. + * + * @tparam unary_op A callable with the signature: + * void(ggml_backend_cann_context&, aclTensor*, aclTensor*) + * where the first aclTensor is the source and the second is the destination. + * @param ctx The CANN backend context for managing resources and execution. + * @param dst The destination tensor. Its src[0] is treated as the input tensor. + */ +template + void ggml_cann_unary_op(ggml_backend_cann_context& ctx, ggml_tensor* dst) { + ggml_tensor* src = dst->src[0]; aclTensor* acl_src = ggml_cann_create_tensor(src); aclTensor* acl_dst = ggml_cann_create_tensor(dst); - uint64_t workspaceSize = 0; - aclOpExecutor* executor; - void* workspaceAddr = nullptr; - - ACL_CHECK(getWorkspaceSize(acl_src, acl_dst, &workspaceSize, &executor)); - if (workspaceSize > 0) { - ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize); - workspaceAddr = workspace_allocator.get(); - } - - aclrtStream main_stream = ctx.stream(); - ACL_CHECK(execute(workspaceAddr, workspaceSize, executor, main_stream)); + unary_op(ctx, acl_src, acl_dst); ACL_CHECK(aclDestroyTensor(acl_src)); ACL_CHECK(aclDestroyTensor(acl_dst)); } -// Activation functions template for const aclTensors. -template -void ggml_cann_activation(ggml_backend_cann_context& ctx, ggml_tensor* dst) { - ggml_tensor* src = dst->src[0]; - - GGML_ASSERT(src->type == GGML_TYPE_F32); - GGML_ASSERT(dst->type == GGML_TYPE_F32); - - aclTensor* acl_src = ggml_cann_create_tensor(src); - aclTensor* acl_dst = ggml_cann_create_tensor(dst); - - uint64_t workspaceSize = 0; - aclOpExecutor* executor; - void* workspaceAddr = nullptr; - - ACL_CHECK(getWorkspaceSize(acl_src, acl_dst, &workspaceSize, &executor)); - if (workspaceSize > 0) { - ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize); - workspaceAddr = workspace_allocator.get(); - } - - aclrtStream main_stream = ctx.stream(); - ACL_CHECK(execute(workspaceAddr, workspaceSize, executor, main_stream)); - - ACL_CHECK(aclDestroyTensor(acl_src)); - ACL_CHECK(aclDestroyTensor(acl_dst)); -} +/** + * @brief Applies a unary operation to a ggml tensor using the CANN backend. + * + * @details This function performs a unary operation on the input tensor using + * a user-provided lambda or callable object `unary_op`, which accepts the CANN + * context and two ACL tensors (source and destination). Internally, this function + * creates ACL representations of the ggml tensors and invokes the unary operation. + * The result is stored in the destination tensor `dst`. This utility abstracts the + * common boilerplate of tensor conversion and cleanup when implementing unary ops. + * + * @param unary_op A callable that performs the unary operation using CANN APIs. + * @param ctx The CANN context used for operations. + * @param dst The destination tensor where the result will be stored. + * The source tensor is retrieved from `dst->src[0]`. + */ +void ggml_cann_unary_op( + std::function unary_op, + ggml_backend_cann_context& ctx, ggml_tensor* dst); +/** + * @brief Helper macro to invoke a unary ACL operation using ggml_cann_unary_op. + * + * This macro defines an inline lambda wrapping a specific ACL operation name, + * and passes it to the templated ggml_cann_unary_op function. It simplifies + * calling unary ops by hiding the lambda boilerplate. + * + * Internally, the lambda will call: + * @code + * GGML_CANN_CALL_ACLNN_OP(OP_NAME, acl_src, acl_dst); + * @endcode + * + * @param OP_NAME The name of the ACL unary operator to invoke via GGML_CANN_CALL_ACLNN_OP. + * + * @see ggml_cann_unary_op + * @see GGML_CANN_CALL_ACLNN_OP + */ +#define GGML_CANN_CALL_UNARY_OP(OP_NAME) \ + do { \ + auto lambda = [](ggml_backend_cann_context& ctx, \ + aclTensor* acl_src, \ + aclTensor* acl_dst) { \ + GGML_CANN_CALL_ACLNN_OP(OP_NAME, acl_src, acl_dst); \ + }; \ + ggml_cann_unary_op(lambda, ctx, dst); \ + } \ + while (0) #endif // CANN_ACLNN_OPS diff --git a/ggml/src/ggml-cann/ggml-cann.cpp b/ggml/src/ggml-cann/ggml-cann.cpp index d410c0244..cec36b36e 100644 --- a/ggml/src/ggml-cann/ggml-cann.cpp +++ b/ggml/src/ggml-cann/ggml-cann.cpp @@ -796,14 +796,14 @@ static bool need_transform(ggml_type type) { * @param buffer The CANN buffer from which to initialize the tensor. * @param tensor Pointer to the tensor to be initialized. */ -static void ggml_backend_cann_buffer_init_tensor( +static enum ggml_status ggml_backend_cann_buffer_init_tensor( ggml_backend_buffer_t buffer, ggml_tensor* tensor) { if (tensor->view_src != NULL && tensor->view_offs == 0) { GGML_ASSERT(tensor->view_src->buffer->buft == buffer->buft); - return; + return GGML_STATUS_SUCCESS; } - // TODO: can backend doesn't support quantized yet. Just leave the code + // TODO: cann backend doesn't support quantized yet. Just leave the code // here. if (ggml_is_quantized(tensor->type)) { // Initialize padding to 0 to avoid possible NaN values @@ -817,6 +817,7 @@ static void ggml_backend_cann_buffer_init_tensor( memset_size, 0, memset_size)); } } + return GGML_STATUS_SUCCESS; } // TODO: need handle tensor which has paddings. @@ -1299,47 +1300,69 @@ static bool ggml_cann_compute_forward(ggml_backend_cann_context& ctx, ggml_cann_dup(ctx, dst); break; case GGML_OP_ADD: - ggml_cann_add(ctx, dst); + case GGML_OP_ADD1: + ggml_cann_binary_op(ctx, dst); + break; + case GGML_OP_SUB: + ggml_cann_binary_op(ctx, dst); break; case GGML_OP_ACC: ggml_cann_acc(ctx, dst); break; case GGML_OP_MUL: - ggml_cann_mul_div(ctx, dst); + ggml_cann_binary_op(ctx, dst); break; case GGML_OP_DIV: - ggml_cann_mul_div(ctx, dst); + ggml_cann_binary_op(ctx, dst); break; case GGML_OP_UNARY: switch (ggml_get_unary_op(dst)) { + case GGML_UNARY_OP_ABS: + GGML_CANN_CALL_UNARY_OP(Abs); + break; + case GGML_UNARY_OP_NEG: + GGML_CANN_CALL_UNARY_OP(Neg); + break; case GGML_UNARY_OP_GELU: - ggml_cann_activation( - ctx, dst); + GGML_CANN_CALL_UNARY_OP(Gelu); break; case GGML_UNARY_OP_SILU: - ggml_cann_activation( - ctx, dst); - break; - // TODO: Use faster gelu?? - case GGML_UNARY_OP_GELU_QUICK: - ggml_cann_activation( - ctx, dst); + GGML_CANN_CALL_UNARY_OP(Silu); break; + case GGML_UNARY_OP_GELU_QUICK: { + auto lambda = [](ggml_backend_cann_context& ctx, + aclTensor* acl_src, + aclTensor* acl_dst) { + GGML_CANN_CALL_ACLNN_OP(GeluV2, acl_src, 0, acl_dst); + }; + ggml_cann_unary_op(lambda, ctx, dst); + } break; case GGML_UNARY_OP_TANH: - ggml_cann_activation( - ctx, dst); + GGML_CANN_CALL_UNARY_OP(Tanh); break; case GGML_UNARY_OP_RELU: - ggml_cann_activation( - ctx, dst); + GGML_CANN_CALL_UNARY_OP(Relu); + break; + case GGML_UNARY_OP_SIGMOID: + GGML_CANN_CALL_UNARY_OP(Sigmoid); break; case GGML_UNARY_OP_HARDSIGMOID: - ggml_cann_activation(ctx, dst); + GGML_CANN_CALL_UNARY_OP(Hardsigmoid); break; case GGML_UNARY_OP_HARDSWISH: - ggml_cann_activation(ctx, dst); + GGML_CANN_CALL_UNARY_OP(Hardswish); + break; + case GGML_UNARY_OP_EXP: + GGML_CANN_CALL_UNARY_OP(Exp); + break; + case GGML_UNARY_OP_ELU: + ggml_cann_elu(ctx, dst); + break; + case GGML_UNARY_OP_SGN: + GGML_CANN_CALL_UNARY_OP(Sign); + break; + case GGML_UNARY_OP_STEP: + ggml_cann_step(ctx, dst); break; default: return false; @@ -1381,7 +1404,12 @@ static bool ggml_cann_compute_forward(ggml_backend_cann_context& ctx, ggml_cann_scale(ctx, dst); break; case GGML_OP_SQR: - ggml_cann_sqr(ctx, dst); + GGML_ASSERT(dst->src[1] == nullptr); + dst->src[1] = dst->src[0]; + ggml_cann_binary_op(ctx, dst); + break; + case GGML_OP_SQRT: + GGML_CANN_CALL_UNARY_OP(Sqrt); break; case GGML_OP_CLAMP: ggml_cann_clamp(ctx, dst); @@ -1413,12 +1441,39 @@ static bool ggml_cann_compute_forward(ggml_backend_cann_context& ctx, case GGML_OP_POOL_2D: ggml_cann_pool2d(ctx, dst); break; + case GGML_OP_SUM: + ggml_cann_sum(ctx, dst); + break; case GGML_OP_SUM_ROWS: ggml_cann_sum_rows(ctx, dst); break; case GGML_OP_ARGSORT: ggml_cann_argsort(ctx, dst); break; + case GGML_OP_ARGMAX: + ggml_cann_argmax(ctx, dst); + break; + case GGML_OP_COS: + ggml_cann_unary_op(ctx, dst); + break; + case GGML_OP_SIN: + ggml_cann_unary_op(ctx, dst); + break; + case GGML_OP_CONV_TRANSPOSE_1D: + ggml_cann_conv_transpose_1d(ctx, dst); + break; + case GGML_OP_LOG: + GGML_CANN_CALL_UNARY_OP(Log); + break; + case GGML_OP_MEAN: + ggml_cann_mean(ctx, dst); + break; + case GGML_OP_PAD_REFLECT_1D: + ggml_cann_pad_reflect_1d(ctx, dst); + break; + case GGML_OP_COUNT_EQUAL: + ggml_cann_count_equal(ctx, dst); + break; default: return false; } @@ -1457,11 +1512,6 @@ static void ggml_backend_cann_free(ggml_backend_t backend) { ACL_CHECK(aclrtSynchronizeDevice()); ACL_CHECK(aclrtResetDevice(cann_ctx->device)); - // finalize when last backend freed. - if (cann_ctx->device == ggml_backend_cann_get_device_count() - 1) { - ACL_CHECK(aclFinalize()); - } - delete cann_ctx; delete backend; } @@ -1674,29 +1724,34 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev, switch (op->op) { case GGML_OP_UNARY: switch (ggml_get_unary_op(op)) { + case GGML_UNARY_OP_ABS: + case GGML_UNARY_OP_NEG: case GGML_UNARY_OP_GELU: case GGML_UNARY_OP_SILU: case GGML_UNARY_OP_RELU: + case GGML_UNARY_OP_SIGMOID: case GGML_UNARY_OP_HARDSIGMOID: case GGML_UNARY_OP_HARDSWISH: case GGML_UNARY_OP_GELU_QUICK: case GGML_UNARY_OP_TANH: + case GGML_UNARY_OP_EXP: + case GGML_UNARY_OP_ELU: + case GGML_UNARY_OP_SGN: + case GGML_UNARY_OP_STEP: return true; default: return false; } case GGML_OP_MUL_MAT: { switch (op->src[0]->type) { - case GGML_TYPE_Q8_0: - // Current groupsize should not be greater than k-1 in - // aclnnWeightQuantBatchMatmulV2GetWorkspaceSize - if (op->src[0]->ne[0] <= QK8_0) { - return false; - } case GGML_TYPE_F16: case GGML_TYPE_F32: - case GGML_TYPE_Q4_0: return true; + case GGML_TYPE_Q8_0: + case GGML_TYPE_Q4_0: + // only support contiguous for quantized types. + return ggml_is_contiguous(op->src[0]) && + ggml_is_contiguous(op->src[1]); default: return false; } @@ -1708,7 +1763,6 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev, switch (op->src[0]->type) { case GGML_TYPE_F32: case GGML_TYPE_F16: - case GGML_TYPE_Q4_0: case GGML_TYPE_Q8_0: return true; default: @@ -1716,16 +1770,21 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev, } } break; case GGML_OP_CPY: { - switch (op->type) { - case GGML_TYPE_F32: - case GGML_TYPE_F16: - case GGML_TYPE_Q8_0: - case GGML_TYPE_Q4_0: - return true; - default: - return false; + ggml_tensor *src = op->src[0]; + if ((op->type != GGML_TYPE_F32 && op->type != GGML_TYPE_F16) || + (src->type != GGML_TYPE_F32 && + src->type != GGML_TYPE_F16)) { + // only support F32 and F16. + return false; } - } + + if (!ggml_are_same_shape(op, src) && !ggml_is_contiguous(op)) { + // unsupport dst is not contiguous. + return false; + } + + return true; + } break; case GGML_OP_CONT: { // TODO: support GGML_TYPE_BF16 switch (op->src[0]->type) { @@ -1738,13 +1797,14 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev, } case GGML_OP_ROPE: { // TODO: with ops-test v == 1 - float * ext_factor = (float*)((int32_t*)op->op_params + 7); + float ext_factor = 0.0f; + memcpy(&ext_factor, (const float *) op->op_params + 7, sizeof(float)); // TODO: n_dims <= ne0 if (op->src[0]->ne[0] != op->op_params[1]) { return false; } // TODO: ext_factor != 0 - if (*ext_factor != 0) { + if (ext_factor != 0) { return false; } @@ -1764,11 +1824,25 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev, if (op->src[0]->ne[2] * op->ne[3] != op->src[0]->ne[3] * op->ne[2]) { return false; } + if (op->op_params[0] != GGML_SCALE_MODE_NEAREST) { + return false; + } return true; } + case GGML_OP_POOL_2D: { + const int32_t * opts = (const int32_t *) op->op_params; + const int k0 = opts[1]; + const int k1 = opts[2]; + const int p0 = opts[5]; + const int p1 = opts[6]; + // value of paddingH should be at most half of kernelH + // value of paddingW should be at most half of kernelW + return (p0 <= (k0 / 2)) && (p1 <= (k1 / 2)); + } + case GGML_OP_SUM: + case GGML_OP_DUP: case GGML_OP_IM2COL: case GGML_OP_CONCAT: - case GGML_OP_DUP: case GGML_OP_REPEAT: case GGML_OP_NONE: case GGML_OP_RESHAPE: @@ -1777,15 +1851,17 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev, case GGML_OP_TRANSPOSE: case GGML_OP_NORM: case GGML_OP_ADD: + case GGML_OP_ADD1: + case GGML_OP_SUB: case GGML_OP_MUL: case GGML_OP_DIV: case GGML_OP_RMS_NORM: case GGML_OP_SCALE: case GGML_OP_SQR: + case GGML_OP_SQRT: case GGML_OP_CLAMP: case GGML_OP_DIAG_MASK_INF: case GGML_OP_SOFT_MAX: - case GGML_OP_POOL_2D: case GGML_OP_SUM_ROWS: case GGML_OP_ARGSORT: case GGML_OP_ACC: @@ -1794,6 +1870,14 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev, case GGML_OP_ARANGE: case GGML_OP_TIMESTEP_EMBEDDING: case GGML_OP_LEAKY_RELU: + case GGML_OP_ARGMAX: + case GGML_OP_COS: + case GGML_OP_SIN: + case GGML_OP_CONV_TRANSPOSE_1D: + case GGML_OP_LOG: + case GGML_OP_MEAN: + case GGML_OP_PAD_REFLECT_1D: + case GGML_OP_COUNT_EQUAL: return true; default: return false; diff --git a/ggml/src/ggml-cann/kernels/CMakeLists.txt b/ggml/src/ggml-cann/kernels/CMakeLists.txt deleted file mode 100644 index d687220c3..000000000 --- a/ggml/src/ggml-cann/kernels/CMakeLists.txt +++ /dev/null @@ -1,30 +0,0 @@ -file(GLOB SRC_FILES - get_row_f32.cpp - get_row_f16.cpp - get_row_q4_0.cpp - get_row_q8_0.cpp - quantize_f32_q8_0.cpp - quantize_f16_q8_0.cpp - quantize_float_to_q4_0.cpp - dup.cpp -) - -set(ASCEND_CANN_PACKAGE_PATH ${CANN_INSTALL_DIR}) -set(RUN_MODE "npu" CACHE STRING "run mode: npu/sim") - -if(EXISTS ${ASCEND_CANN_PACKAGE_PATH}/compiler/tikcpp/ascendc_kernel_cmake) - set(ASCENDC_CMAKE_DIR ${ASCEND_CANN_PACKAGE_PATH}/compiler/tikcpp/ascendc_kernel_cmake) -elseif(EXISTS ${ASCEND_CANN_PACKAGE_PATH}/ascendc_devkit/tikcpp/samples/cmake) - set(ASCENDC_CMAKE_DIR ${ASCEND_CANN_PACKAGE_PATH}/ascendc_devkit/tikcpp/samples/cmake) -else() - message(FATAL_ERROR "ascendc_kernel_cmake does not exist, please check whether the compiler package is installed.") -endif() -include(${ASCENDC_CMAKE_DIR}/ascendc.cmake) - -ascendc_library(ascendc_kernels STATIC - ${SRC_FILES} -) - -message(STATUS "CANN: compile ascend kernels witch SOC_TYPE:${SOC_TYPE}, SOC_VERSION:${SOC_VERSION}, compile macro:-D${SOC_TYPE_COMPILE_OPTION}.") -ascendc_compile_definitions(ascendc_kernels PRIVATE "-D${SOC_TYPE_COMPILE_OPTION}") -# ascendc_compile_definitions(ascendc_kernels PRIVATE -DASCENDC_DUMP) diff --git a/ggml/src/ggml-cann/kernels/ascendc_kernels.h b/ggml/src/ggml-cann/kernels/ascendc_kernels.h deleted file mode 100644 index 7e153208c..000000000 --- a/ggml/src/ggml-cann/kernels/ascendc_kernels.h +++ /dev/null @@ -1,19 +0,0 @@ -#ifndef ASCENDC_KERNELS_H -#define ASCENDC_KERNELS_H - -#include "aclrtlaunch_ascendc_get_row_f32.h" -#include "aclrtlaunch_ascendc_get_row_f16.h" -#include "aclrtlaunch_ascendc_get_row_q8_0.h" -#include "aclrtlaunch_ascendc_get_row_q4_0.h" - -#include "aclrtlaunch_ascendc_quantize_f32_q8_0.h" -#include "aclrtlaunch_ascendc_quantize_f16_q8_0.h" -#include "aclrtlaunch_ascendc_quantize_f16_to_q4_0.h" -#include "aclrtlaunch_ascendc_quantize_f32_to_q4_0.h" - -#include "aclrtlaunch_ascendc_dup_by_rows_fp16.h" -#include "aclrtlaunch_ascendc_dup_by_rows_fp32.h" -#include "aclrtlaunch_ascendc_dup_by_rows_fp32_to_fp16.h" -#include "aclrtlaunch_ascendc_dup_by_rows_fp16_to_fp32.h" - -#endif // ASCENDC_KERNELS_H diff --git a/ggml/src/ggml-cann/kernels/dup.cpp b/ggml/src/ggml-cann/kernels/dup.cpp deleted file mode 100644 index c7ba38d10..000000000 --- a/ggml/src/ggml-cann/kernels/dup.cpp +++ /dev/null @@ -1,236 +0,0 @@ -#include "kernel_operator.h" - -#include - -using namespace AscendC; - -#define BUFFER_NUM 2 -const int64_t SUPPORTED_MAX_DIM = 65535; // currently the limit of max block dim supportted by dup kernel is 65535template - -template -class DupByRows { - public: - __aicore__ inline DupByRows() {} - __aicore__ inline void init(GM_ADDR src, GM_ADDR dst, int64_t *input_ne_ub, - size_t *input_nb_ub) { - /* Dup by rows when src is contigous on first dimension and dst is - contiguous, each kernel process one row. - */ - - // Input has four dims. - int64_t op_block_num = GetBlockNum(); - int64_t op_block_idx = GetBlockIdx(); - - // param - num_rows = input_ne_ub[1] * input_ne_ub[2] * input_ne_ub[3]; - num_elem = input_ne_ub[0]; - - // index for (ne[1], ne[2], ne[3]): (idx_ne1, idx_ne2, idx_ne3) - idx_ne3 = op_block_idx / (input_ne_ub[1] * input_ne_ub[2]); - idx_ne2 = (op_block_idx - idx_ne3 * (input_ne_ub[1] * input_ne_ub[2])) - / (input_ne_ub[1]); - idx_ne1 = op_block_idx - idx_ne3 * (input_ne_ub[1] * input_ne_ub[2]) - - idx_ne2 * input_ne_ub[1]; - - // src may not contiguous in dim [1,2,3], so stride decited by ne&nb - src_stride = input_nb_ub[3] * idx_ne3 + input_nb_ub[2] * idx_ne2 - + input_nb_ub[1] * idx_ne1; - - // dst is contiguous - dst_stride = op_block_idx * (input_ne_ub[0] * sizeof(DST_T)); - - src_gm.SetGlobalBuffer(reinterpret_cast<__gm__ SRC_T *>(src + - src_stride)); - dst_gm.SetGlobalBuffer(reinterpret_cast<__gm__ DST_T *>(dst + - dst_stride)); - - pipe.InitBuffer(src_queue, BUFFER_NUM, (sizeof(SRC_T) * num_elem + - 32 - 1) / 32 * 32); - pipe.InitBuffer(dst_queue, BUFFER_NUM, (sizeof(DST_T) * num_elem + - 32 - 1) / 32 * 32); - } - - __aicore__ inline void copy_in() { - LocalTensor src_local = src_queue.AllocTensor(); - const size_t elem_per_block = 32 / sizeof(SRC_T); - size_t tail = num_elem % elem_per_block; - size_t cpy_elements_len = tail > 0 ? num_elem + 1 : num_elem; - DataCopy(src_local, src_gm, cpy_elements_len); - src_queue.EnQue(src_local); - } - - __aicore__ inline void copy_out() { - LocalTensor dst_local = dst_queue.DeQue(); -#ifdef ASCEND_310P - const size_t elem_per_block = 32 / sizeof(DST_T); - size_t tail = num_elem % elem_per_block; - size_t len = num_elem & ~(elem_per_block - 1); - if (len > 0) { - DataCopy(dst_gm, dst_local, len); - } - if(tail != 0) { - for (size_t i = tail; i < elem_per_block; i++) { - dst_local[len + i].SetValue(0, 0); - } - SetAtomicAdd(); - DataCopy(dst_gm[len], dst_local[len], elem_per_block); - SetAtomicNone(); - } -#else - DataCopyExtParams dataCopyParams; - dataCopyParams.blockCount = 1; - dataCopyParams.blockLen = num_elem * sizeof(DST_T); - DataCopyPad(dst_gm, dst_local, dataCopyParams); -#endif - dst_queue.FreeTensor(dst_local); - } - - __aicore__ inline void dup() { - // main process, copy one row data from src to dst. - copy_in(); - - LocalTensor src_local = src_queue.DeQue(); - LocalTensor dst_local = dst_queue.AllocTensor(); - - int32_t BLOCK_NUM = 32 / sizeof(DST_T); - DataCopy(dst_local, src_local, (num_elem + BLOCK_NUM - 1) - / BLOCK_NUM * BLOCK_NUM); - dst_queue.EnQue(dst_local); - - src_queue.FreeTensor(src_local); - copy_out(); - } - - __aicore__ inline void dup_with_cast() { - // main process, copy one row data from src to dst. - // cast dtype from src to dst. - copy_in(); - - LocalTensor src_local = src_queue.DeQue(); - LocalTensor dst_local = dst_queue.AllocTensor(); - - Cast(dst_local, src_local, RoundMode::CAST_NONE, num_elem); - dst_queue.EnQue(dst_local); - - src_queue.FreeTensor(src_local); - copy_out(); - } - - private: - - TPipe pipe; - GlobalTensor src_gm; - GlobalTensor dst_gm; - - int64_t num_rows; - int64_t num_elem; - int64_t idx_ne3; - int64_t idx_ne2; - int64_t idx_ne1; - int64_t src_stride; - int64_t dst_stride; - - TQue src_queue; - TQue dst_queue; -}; - -template -__aicore__ inline void copy_to_ub(GM_ADDR gm, T *ub, size_t size) { - auto gm_ptr = (__gm__ uint8_t *)gm; - auto ub_ptr = (uint8_t *)(ub); - for (int32_t i = 0; i < size; ++i, ++ub_ptr, ++gm_ptr) { - *ub_ptr = *gm_ptr; - } -} - -extern "C" __global__ __aicore__ void ascendc_dup_by_rows_fp16( - GM_ADDR src_gm, - GM_ADDR dst_gm, - GM_ADDR input_ne_gm, - GM_ADDR input_nb_gm, - GM_ADDR output_ne_gm, - GM_ADDR output_nb_gm) { - - int64_t input_ne_ub[4]; - size_t input_nb_ub[4]; - int64_t output_ne_ub[4]; - size_t output_nb_ub[4]; - - copy_to_ub(input_ne_gm, input_ne_ub, 32); - copy_to_ub(input_nb_gm, input_nb_ub, 32); - copy_to_ub(output_ne_gm, output_ne_ub, 32); - copy_to_ub(output_nb_gm, output_nb_ub, 32); - - DupByRows op; - op.init(src_gm, dst_gm, input_ne_ub, input_nb_ub); - op.dup(); -} - -extern "C" __global__ __aicore__ void ascendc_dup_by_rows_fp32( - GM_ADDR src_gm, - GM_ADDR dst_gm, - GM_ADDR input_ne_gm, - GM_ADDR input_nb_gm, - GM_ADDR output_ne_gm, - GM_ADDR output_nb_gm) { - int64_t input_ne_ub[4]; - size_t input_nb_ub[4]; - int64_t output_ne_ub[4]; - size_t output_nb_ub[4]; - - copy_to_ub(input_ne_gm, input_ne_ub, 32); - copy_to_ub(input_nb_gm, input_nb_ub, 32); - copy_to_ub(output_ne_gm, output_ne_ub, 32); - copy_to_ub(output_nb_gm, output_nb_ub, 32); - - DupByRows op; - op.init(src_gm, dst_gm, input_ne_ub, input_nb_ub); - op.dup(); -} - -extern "C" __global__ __aicore__ void ascendc_dup_by_rows_fp32_to_fp16( - GM_ADDR src_gm, - GM_ADDR dst_gm, - GM_ADDR input_ne_gm, - GM_ADDR input_nb_gm, - GM_ADDR output_ne_gm, - GM_ADDR output_nb_gm) { - - int64_t input_ne_ub[4]; - size_t input_nb_ub[4]; - int64_t output_ne_ub[4]; - size_t output_nb_ub[4]; - - copy_to_ub(input_ne_gm, input_ne_ub, 32); - copy_to_ub(input_nb_gm, input_nb_ub, 32); - copy_to_ub(output_ne_gm, output_ne_ub, 32); - copy_to_ub(output_nb_gm, output_nb_ub, 32); - - DupByRows op; - op.init(src_gm, dst_gm, input_ne_ub, input_nb_ub); - op.dup_with_cast(); -} - -extern "C" __global__ __aicore__ void ascendc_dup_by_rows_fp16_to_fp32( - GM_ADDR src_gm, - GM_ADDR dst_gm, - GM_ADDR input_ne_gm, - GM_ADDR input_nb_gm, - GM_ADDR output_ne_gm, - GM_ADDR output_nb_gm) { - - // copy params from gm to ub. - int64_t input_ne_ub[4]; - size_t input_nb_ub[4]; - int64_t output_ne_ub[4]; - size_t output_nb_ub[4]; - - copy_to_ub(input_ne_gm, input_ne_ub, 32); - copy_to_ub(input_nb_gm, input_nb_ub, 32); - copy_to_ub(output_ne_gm, output_ne_ub, 32); - copy_to_ub(output_nb_gm, output_nb_ub, 32); - - DupByRows op; - op.init(src_gm, dst_gm, input_ne_ub, input_nb_ub); - op.dup_with_cast(); -} diff --git a/ggml/src/ggml-cann/kernels/get_row_f16.cpp b/ggml/src/ggml-cann/kernels/get_row_f16.cpp deleted file mode 100644 index 416b45104..000000000 --- a/ggml/src/ggml-cann/kernels/get_row_f16.cpp +++ /dev/null @@ -1,197 +0,0 @@ -#include "kernel_operator.h" - -// optimize me. Use template to avoid copy code. -using namespace AscendC; - -#define BUFFER_NUM 2 - -class GET_ROW_F16 { - public: - __aicore__ inline GET_ROW_F16() {} - __aicore__ inline void init(GM_ADDR input, GM_ADDR indices, GM_ADDR output, - int64_t *input_ne_ub, size_t *input_nb_ub, - int64_t *indices_ne_ub, size_t *indices_nb_ub, - int64_t *output_ne_ub, size_t *output_nb_ub) { - // TODO, use template for F16/f32 - int64_t op_block_num = GetBlockNum(); - op_block_idx = GetBlockIdx(); - - for (int i = 0; i < 4; i++) { - input_ne[i] = input_ne_ub[i]; - input_stride[i] = input_nb_ub[i] / input_nb_ub[0]; - - indices_ne[i] = indices_ne_ub[i]; - indices_stride[i] = indices_nb_ub[i] / indices_nb_ub[0]; - - output_ne[i] = output_ne_ub[i]; - output_stride[i] = output_nb_ub[i] / output_nb_ub[0]; - } - - // Indices has two dims. n_elements = all rows should get. - // dr, all rows should this thread get. - uint64_t n_elements = - indices_ne[0] * indices_ne[1] * indices_ne[2] * indices_ne[3]; - dr = n_elements / op_block_num; - - uint64_t tails = n_elements % op_block_num; - if (op_block_idx < tails) { - dr += 1; - ir = dr * op_block_idx; - } else { - ir = dr * op_block_idx + tails; - } - - input_gm.SetGlobalBuffer((__gm__ half *)input); - indices_gm.SetGlobalBuffer((__gm__ int32_t *)indices); - output_gm.SetGlobalBuffer((__gm__ float *)output); - - uint64_t input_local_buffer_size = ((input_ne[0] * sizeof(half) + 31) - & ~31); - uint64_t output_local_buffer_size = ((input_ne[0] * sizeof(float) + 31) - & ~31); - - local_buffer_elems = input_local_buffer_size / sizeof(half); - - // TODO, consider long row that can't put in UB. - // All data should asign to 32. It's ok because all data is align to 32. - pipe.InitBuffer(input_queue, BUFFER_NUM, input_local_buffer_size); - pipe.InitBuffer(output_queue, BUFFER_NUM, output_local_buffer_size); - } - - __aicore__ inline void copy_in(uint32_t offset, size_t len) { - size_t origin_len = len; - LocalTensor input_local = input_queue.AllocTensor(); - const size_t elem_per_block = 32 / sizeof(half); - size_t tail = len % elem_per_block; - len = len & ~(elem_per_block - 1); - if(tail != 0) { - len += elem_per_block; - } - DataCopy(input_local, input_gm[offset], len); - input_queue.EnQue(input_local); - } - - __aicore__ inline void copy_out(uint32_t offset, size_t len) { - LocalTensor output_local = output_queue.DeQue(); - const size_t elem_per_block = 32 / sizeof(float); - size_t tail = len % elem_per_block; - len = len & ~(elem_per_block - 1); - if (len > 0) { - DataCopy(output_gm[offset], output_local, len); - } - - if(tail != 0) { -#ifdef ASCEND_310P - for (size_t i = tail; i < elem_per_block; i++) { - output_local[len + i].SetValue(0, 0); - } - SetAtomicAdd(); - DataCopy(output_gm[offset + len], output_local[len], elem_per_block); - SetAtomicNone(); -#else - DataCopyExtParams dataCopyParams; - dataCopyParams.blockCount = 1; - dataCopyParams.blockLen = tail * sizeof(float); - DataCopyPad(output_gm[offset + len], output_local[len], - dataCopyParams); -#endif - } - output_queue.FreeTensor(output_local); - } - - __aicore__ inline void calculate_row(int64_t idx) { - const int64_t indices_ne2_idx = idx / (indices_ne[0] * indices_ne[1]); - const int64_t indices_ne1_idx = - (idx - indices_ne2_idx * indices_ne[0] * indices_ne[1]) / - indices_ne[0]; - const int64_t indices_ne0_idx = - (idx - indices_ne2_idx * indices_ne[0] * indices_ne[1] - - indices_ne1_idx * indices_ne[0]); - - const int64_t indices_offset = indices_ne0_idx * indices_stride[0] + - indices_ne1_idx * indices_stride[1] + - indices_ne2_idx * indices_stride[2]; - const int32_t selected_row_idx = indices_gm.GetValue(indices_offset); - - const int64_t input_offset = selected_row_idx * input_stride[1] + - indices_ne1_idx * input_stride[2] + - indices_ne2_idx * input_stride[3]; - - const int64_t output_offset = indices_ne0_idx * output_stride[1] + - indices_ne1_idx * output_stride[2] + - indices_ne2_idx * output_stride[3]; - - copy_in(input_offset, input_ne[0]); - LocalTensor input_local = input_queue.DeQue(); - LocalTensor output_local = output_queue.AllocTensor(); - - Cast(output_local, input_local, RoundMode::CAST_NONE, - local_buffer_elems); - output_queue.EnQue(output_local); - copy_out(output_offset, input_ne[0]); - - input_queue.FreeTensor(input_local); - } - - __aicore__ inline void calculate() { - for (int64_t i = ir; i < ir + dr; i++) { - calculate_row(i); - } - } - - private: - int64_t input_ne[4]; - size_t input_stride[4]; - - int64_t indices_ne[4]; - size_t indices_stride[4]; - - int64_t output_ne[4]; - size_t output_stride[4]; - - size_t local_buffer_elems; - - int64_t ir; - int64_t dr; - - TPipe pipe; - GlobalTensor input_gm; - GlobalTensor indices_gm; - GlobalTensor output_gm; - TQue input_queue; - TQue output_queue; - int64_t op_block_idx; -}; - -template -__aicore__ inline void copy_to_ub(GM_ADDR gm, T *ub, size_t size) { - auto gm_ptr = (__gm__ uint8_t *)gm; - auto ub_ptr = (uint8_t *)(ub); - for (int32_t i = 0; i < size; ++i, ++ub_ptr, ++gm_ptr) { - *ub_ptr = *gm_ptr; - } -} - -extern "C" __global__ __aicore__ void ascendc_get_row_f16( - GM_ADDR input_gm, GM_ADDR indices_gm, GM_ADDR output_gm, - GM_ADDR input_ne_gm, GM_ADDR input_nb_gm, GM_ADDR indices_ne_gm, - GM_ADDR indices_nb_gm, GM_ADDR output_ne_gm, GM_ADDR output_nb_gm) { - int64_t input_ne_ub[4]; - size_t input_nb_ub[4]; - int64_t indices_ne_ub[4]; - size_t indices_nb_ub[4]; - int64_t output_ne_ub[4]; - size_t output_nb_ub[4]; - - copy_to_ub(input_ne_gm, input_ne_ub, 32); - copy_to_ub(input_nb_gm, input_nb_ub, 32); - copy_to_ub(indices_ne_gm, indices_ne_ub, 32); - copy_to_ub(indices_nb_gm, indices_nb_ub, 32); - copy_to_ub(output_ne_gm, output_ne_ub, 32); - copy_to_ub(output_nb_gm, output_nb_ub, 32); - - GET_ROW_F16 op; - op.init(input_gm, indices_gm, output_gm, input_ne_ub, input_nb_ub, - indices_ne_ub, indices_nb_ub, output_ne_ub, output_nb_ub); - op.calculate(); -} diff --git a/ggml/src/ggml-cann/kernels/get_row_f32.cpp b/ggml/src/ggml-cann/kernels/get_row_f32.cpp deleted file mode 100644 index 02116905b..000000000 --- a/ggml/src/ggml-cann/kernels/get_row_f32.cpp +++ /dev/null @@ -1,190 +0,0 @@ -#include "kernel_operator.h" - -// optimize me. Use template to avoid copy code. -using namespace AscendC; - -#define BUFFER_NUM 2 - -class GET_ROW_F32 { - public: - __aicore__ inline GET_ROW_F32() {} - __aicore__ inline void init(GM_ADDR input, GM_ADDR indices, GM_ADDR output, - int64_t *input_ne_ub, size_t *input_nb_ub, - int64_t *indices_ne_ub, size_t *indices_nb_ub, - int64_t *output_ne_ub, size_t *output_nb_ub) { - int64_t op_block_num = GetBlockNum(); - op_block_idx = GetBlockIdx(); - - for (int i = 0; i < 4; i++) { - input_ne[i] = input_ne_ub[i]; - input_stride[i] = input_nb_ub[i] / input_nb_ub[0]; - - indices_ne[i] = indices_ne_ub[i]; - indices_stride[i] = indices_nb_ub[i] / indices_nb_ub[0]; - - output_ne[i] = output_ne_ub[i]; - output_stride[i] = output_nb_ub[i] / output_nb_ub[0]; - } - - // Indices has two dims. n_elements = all rows should get. - // dr, all rows should this thread get. - uint64_t n_elements = - indices_ne[0] * indices_ne[1] * indices_ne[2] * indices_ne[3]; - dr = n_elements / op_block_num; - - uint64_t tails = n_elements % op_block_num; - if (op_block_idx < tails) { - dr += 1; - ir = dr * op_block_idx; - } else { - ir = dr * op_block_idx + tails; - } - - input_gm.SetGlobalBuffer((__gm__ float *)input); - indices_gm.SetGlobalBuffer((__gm__ int32_t *)indices); - output_gm.SetGlobalBuffer((__gm__ float *)output); - - uint64_t local_buffer_size = ((input_ne[0] * sizeof(float) + 31) & ~31); - local_buffer_elems = local_buffer_size / sizeof(float); - - // TODO, consider long row that can't put in UB. - // All data should asign to 32. It's ok because all data is align to 32. - pipe.InitBuffer(input_queue, BUFFER_NUM, local_buffer_size); - pipe.InitBuffer(output_queue, BUFFER_NUM, local_buffer_size); - } - - __aicore__ inline void copy_in(uint32_t offset, size_t len) { - LocalTensor input_local = input_queue.AllocTensor(); - const size_t elem_per_block = 32 / sizeof(float); - size_t tail = len % elem_per_block; - len = len & ~(elem_per_block - 1); - if(tail != 0) { - len += elem_per_block; - } - DataCopy(input_local, input_gm[offset], len); - input_queue.EnQue(input_local); - } - - __aicore__ inline void copy_out(uint32_t offset, size_t len) { - LocalTensor output_local = output_queue.DeQue(); - const size_t elem_per_block = 32 / sizeof(float); - size_t tail = len % elem_per_block; - len = len & ~(elem_per_block - 1); - if (len > 0) { - DataCopy(output_gm[offset], output_local, len); - } - - if(tail != 0) { -#ifdef ASCEND_310P - for (size_t i = tail; i < elem_per_block; i++) { - output_local[len + i].SetValue(0, 0); - } - SetAtomicAdd(); - DataCopy(output_gm[offset + len], output_local[len], elem_per_block); - SetAtomicNone(); -#else - DataCopyExtParams dataCopyParams; - dataCopyParams.blockCount = 1; - dataCopyParams.blockLen = tail * sizeof(float); - DataCopyPad(output_gm[offset + len], output_local[len], - dataCopyParams); -#endif - } - output_queue.FreeTensor(output_local); - } - - __aicore__ inline void calculate_row(int64_t idx) { - const int64_t indices_ne2_idx = idx / (indices_ne[0] * indices_ne[1]); - const int64_t indices_ne1_idx = - (idx - indices_ne2_idx * indices_ne[0] * indices_ne[1]) / - indices_ne[0]; - const int64_t indices_ne0_idx = - (idx - indices_ne2_idx * indices_ne[0] * indices_ne[1] - - indices_ne1_idx * indices_ne[0]); - - const int64_t indices_offset = indices_ne0_idx * indices_stride[0] + - indices_ne1_idx * indices_stride[1] + - indices_ne2_idx * indices_stride[2]; - const int32_t selected_row_idx = indices_gm.GetValue(indices_offset); - - const int64_t input_offset = selected_row_idx * input_stride[1] + - indices_ne1_idx * input_stride[2] + - indices_ne2_idx * input_stride[3]; - - const int64_t output_offset = indices_ne0_idx * output_stride[1] + - indices_ne1_idx * output_stride[2] + - indices_ne2_idx * output_stride[3]; - - copy_in(input_offset, input_ne[0]); - LocalTensor input_local = input_queue.DeQue(); - LocalTensor output_local = output_queue.AllocTensor(); - - DataCopy(output_local, input_local, local_buffer_elems); - output_queue.EnQue(output_local); - copy_out(output_offset, input_ne[0]); - - input_queue.FreeTensor(input_local); - } - - __aicore__ inline void calculate() { - for (int64_t i = ir; i < ir + dr; i++) { - calculate_row(i); - } - } - - private: - int64_t input_ne[4]; - size_t input_stride[4]; - - int64_t indices_ne[4]; - size_t indices_stride[4]; - - int64_t output_ne[4]; - size_t output_stride[4]; - - size_t local_buffer_elems; - - int64_t ir; - int64_t dr; - - TPipe pipe; - GlobalTensor input_gm; - GlobalTensor indices_gm; - GlobalTensor output_gm; - TQue input_queue; - TQue output_queue; - int64_t op_block_idx; -}; - -template -__aicore__ inline void copy_to_ub(GM_ADDR gm, T *ub, size_t size) { - auto gm_ptr = (__gm__ uint8_t *)gm; - auto ub_ptr = (uint8_t *)(ub); - for (int32_t i = 0; i < size; ++i, ++ub_ptr, ++gm_ptr) { - *ub_ptr = *gm_ptr; - } -} - -extern "C" __global__ __aicore__ void ascendc_get_row_f32( - GM_ADDR input_gm, GM_ADDR indices_gm, GM_ADDR output_gm, - GM_ADDR input_ne_gm, GM_ADDR input_nb_gm, GM_ADDR indices_ne_gm, - GM_ADDR indices_nb_gm, GM_ADDR output_ne_gm, GM_ADDR output_nb_gm) { - int64_t input_ne_ub[4]; - size_t input_nb_ub[4]; - int64_t indices_ne_ub[4]; - size_t indices_nb_ub[4]; - int64_t output_ne_ub[4]; - size_t output_nb_ub[4]; - - copy_to_ub(input_ne_gm, input_ne_ub, 32); - copy_to_ub(input_nb_gm, input_nb_ub, 32); - copy_to_ub(indices_ne_gm, indices_ne_ub, 32); - copy_to_ub(indices_nb_gm, indices_nb_ub, 32); - copy_to_ub(output_ne_gm, output_ne_ub, 32); - copy_to_ub(output_nb_gm, output_nb_ub, 32); - - GET_ROW_F32 op; - op.init(input_gm, indices_gm, output_gm, input_ne_ub, input_nb_ub, - indices_ne_ub, indices_nb_ub, output_ne_ub, output_nb_ub); - op.calculate(); -} diff --git a/ggml/src/ggml-cann/kernels/get_row_q4_0.cpp b/ggml/src/ggml-cann/kernels/get_row_q4_0.cpp deleted file mode 100644 index 4fbe72208..000000000 --- a/ggml/src/ggml-cann/kernels/get_row_q4_0.cpp +++ /dev/null @@ -1,204 +0,0 @@ -#include "kernel_operator.h" - -// optimize me. Use template to avoid copy code. -using namespace AscendC; -#ifdef ASCEND_310P // 310P not support 4bit get row - extern "C" __global__ __aicore__ void ascendc_get_row_q4_0( - GM_ADDR input_gm, GM_ADDR indices_gm, GM_ADDR output_gm, - GM_ADDR input_ne_gm, GM_ADDR indices_ne_gm, GM_ADDR indices_nb_gm, - GM_ADDR output_ne_gm, GM_ADDR output_nb_gm) { - // let following test cases can continue run, here just print error information. Of Cource the test case that call this operator is failed. - printf("Ascend310P not support 4bit get row.\n"); - } -#else - -#define BUFFER_NUM 2 - -#define QK4_0 32 - -class GET_ROW_Q4_0 { - public: - __aicore__ inline GET_ROW_Q4_0() {} - __aicore__ inline void init(GM_ADDR input, GM_ADDR indices, GM_ADDR output, - int64_t *input_ne_ub, int64_t *indices_ne_ub, - size_t *indices_nb_ub, int64_t *output_ne_ub, - size_t *output_nb_ub) { - int64_t op_block_num = GetBlockNum(); - int64_t op_block_idx = GetBlockIdx(); - - for (int i = 0; i < 4; i++) { - input_ne[i] = input_ne_ub[i]; - indices_ne[i] = indices_ne_ub[i]; - indices_stride[i] = indices_nb_ub[i] / indices_nb_ub[0]; - scale_ne[i] = input_ne_ub[i]; - output_ne[i] = output_ne_ub[i]; - output_stride[i] = output_nb_ub[i] / output_nb_ub[0]; - } - - // one scale for a group. - scale_ne[0] /= QK4_0; - - input_stride[0] = 1; - scale_stride[0] = 1; - output_stride[0] = 1; - for (int i = 1; i < 4; i++) { - input_stride[i] = input_stride[i - 1] * input_ne[i - 1]; - scale_stride[i] = scale_stride[i - 1] * scale_ne[i - 1]; - } - - group_size_in_row = input_ne[0] / QK4_0; - int64_t scale_offset = input_ne[0] * input_ne[1] * input_ne[2] * - input_ne[3] / 2; - - // Indices has two dims. n_elements = all rows should get. - // dr, all rows should this thread get. - uint64_t n_elements = - indices_ne[0] * indices_ne[1] * indices_ne[2] * indices_ne[3]; - dr = n_elements / op_block_num; - - uint64_t tails = n_elements % op_block_num; - if (op_block_idx < tails) { - dr += 1; - ir = dr * op_block_idx; - } else { - ir = dr * op_block_idx + tails; - } - - input_gm.SetGlobalBuffer((__gm__ int4b_t *)input); - scale_gm.SetGlobalBuffer((__gm__ half *)(input + scale_offset)); - indices_gm.SetGlobalBuffer((__gm__ int32_t *)indices); - output_gm.SetGlobalBuffer((__gm__ float *)output); - - pipe.InitBuffer(input_queue, BUFFER_NUM, QK4_0 * sizeof(int4b_t)); - pipe.InitBuffer(cast_queue, BUFFER_NUM, QK4_0 * sizeof(half)); - pipe.InitBuffer(output_queue, BUFFER_NUM, QK4_0 * sizeof(float)); - } - - __aicore__ inline void copy_in(uint32_t offset) { - LocalTensor input_local = input_queue.AllocTensor(); - // 32 * sizeof(int4b_t) = 16, which is not aligned to 32, why no error? - DataCopy(input_local, input_gm[offset], QK4_0); - input_queue.EnQue(input_local); - } - - __aicore__ inline void copy_out(uint32_t offset) { - LocalTensor output_local = output_queue.DeQue(); - DataCopy(output_gm[offset], output_local, QK4_0); - output_queue.FreeTensor(output_local); - } - - __aicore__ inline void calculate_group(int64_t idx, int64_t group) { - const int64_t indices_ne2_idx = idx / (indices_ne[0] * indices_ne[1]); - const int64_t indices_ne1_idx = - (idx - indices_ne2_idx * indices_ne[0] * indices_ne[1]) / - indices_ne[0]; - const int64_t indices_ne0_idx = - (idx - indices_ne2_idx * indices_ne[0] * indices_ne[1] - - indices_ne1_idx * indices_ne[0]); - - const int64_t indices_offset = indices_ne0_idx * indices_stride[0] + - indices_ne1_idx * indices_stride[1] + - indices_ne2_idx * indices_stride[2]; - const int32_t selected_row_idx = indices_gm.GetValue(indices_offset); - - const int64_t input_offset = selected_row_idx * input_stride[1] + - indices_ne1_idx * input_stride[2] + - indices_ne2_idx * input_stride[3] + - group * QK4_0; - const int64_t scale_offset = selected_row_idx * scale_stride[1] + - indices_ne1_idx * scale_stride[2] + - indices_ne2_idx * scale_stride[3] + group; - const int64_t output_offset = indices_ne0_idx * output_stride[1] + - indices_ne1_idx * output_stride[2] + - indices_ne2_idx * output_stride[3] + - group * QK4_0; - - copy_in(input_offset); - LocalTensor input_local = input_queue.DeQue(); - LocalTensor cast_local = cast_queue.AllocTensor(); - LocalTensor output_local = output_queue.AllocTensor(); - - // TODO: cast more data to speed up. - Cast(cast_local, input_local, RoundMode::CAST_NONE, QK4_0); - Cast(output_local, cast_local, RoundMode::CAST_NONE, QK4_0); - - // Only mul need compile by group. - half scale = scale_gm.GetValue(scale_offset); - - Muls(output_local, output_local, (float)scale, QK4_0); - - input_queue.FreeTensor(input_local); - cast_queue.FreeTensor(cast_local); - output_queue.EnQue(output_local); - - copy_out(output_offset); - } - - __aicore__ inline void calculate() { - for (int64_t i = ir; i < ir + dr; i++) { - for (int64_t j = 0; j < group_size_in_row; j++) { - calculate_group(i, j); - } - } - } - - private: - int64_t input_ne[4]; - size_t input_stride[4]; - - int64_t scale_ne[4]; - size_t scale_stride[4]; - - int64_t indices_ne[4]; - size_t indices_stride[4]; - - int64_t output_ne[4]; - size_t output_stride[4]; - - int64_t ir; - int64_t dr; - - int64_t group_size_in_row; - - TPipe pipe; - GlobalTensor input_gm; - GlobalTensor scale_gm; - GlobalTensor indices_gm; - GlobalTensor output_gm; - TQue input_queue; - TQue output_queue; - TQue cast_queue; -}; - -template -__aicore__ inline void copy_to_ub(GM_ADDR gm, T *ub, size_t size) { - auto gm_ptr = (__gm__ uint8_t *)gm; - auto ub_ptr = (uint8_t *)(ub); - for (int32_t i = 0; i < size; ++i, ++ub_ptr, ++gm_ptr) { - *ub_ptr = *gm_ptr; - } -} - -extern "C" __global__ __aicore__ void ascendc_get_row_q4_0( - GM_ADDR input_gm, GM_ADDR indices_gm, GM_ADDR output_gm, - GM_ADDR input_ne_gm, GM_ADDR indices_ne_gm, GM_ADDR indices_nb_gm, - GM_ADDR output_ne_gm, GM_ADDR output_nb_gm) { - int64_t input_ne_ub[4]; - int64_t indices_ne_ub[4]; - size_t indices_nb_ub[4]; - int64_t output_ne_ub[4]; - size_t output_nb_ub[4]; - - copy_to_ub(input_ne_gm, input_ne_ub, 32); - copy_to_ub(indices_ne_gm, indices_ne_ub, 32); - copy_to_ub(indices_nb_gm, indices_nb_ub, 32); - copy_to_ub(output_ne_gm, output_ne_ub, 32); - copy_to_ub(output_nb_gm, output_nb_ub, 32); - - GET_ROW_Q4_0 op; - op.init(input_gm, indices_gm, output_gm, input_ne_ub, indices_ne_ub, - indices_nb_ub, output_ne_ub, output_nb_ub); - op.calculate(); -} - -#endif // #ifdef ASCEND_310P diff --git a/ggml/src/ggml-cann/kernels/get_row_q8_0.cpp b/ggml/src/ggml-cann/kernels/get_row_q8_0.cpp deleted file mode 100644 index ba9ab3c04..000000000 --- a/ggml/src/ggml-cann/kernels/get_row_q8_0.cpp +++ /dev/null @@ -1,191 +0,0 @@ -#include "kernel_operator.h" - -// optimize me. Use template to avoid copy code. -using namespace AscendC; - -#define BUFFER_NUM 2 - -#define QK8_0 32 - -class GET_ROW_Q8_0 { - public: - __aicore__ inline GET_ROW_Q8_0() {} - __aicore__ inline void init(GM_ADDR input, GM_ADDR indices, GM_ADDR output, - int64_t *input_ne_ub, int64_t *indices_ne_ub, - size_t *indices_nb_ub, int64_t *output_ne_ub, - size_t *output_nb_ub) { - int64_t op_block_num = GetBlockNum(); - int64_t op_block_idx = GetBlockIdx(); - - for (int i = 0; i < 4; i++) { - input_ne[i] = input_ne_ub[i]; - indices_ne[i] = indices_ne_ub[i]; - indices_stride[i] = indices_nb_ub[i] / indices_nb_ub[0]; - scale_ne[i] = input_ne_ub[i]; - output_ne[i] = output_ne_ub[i]; - output_stride[i] = output_nb_ub[i] / output_nb_ub[0]; - } - - // one scale for a group. - scale_ne[0] /= QK8_0; - - input_stride[0] = 1; - scale_stride[0] = 1; - output_stride[0] = 1; - for (int i = 1; i < 4; i++) { - input_stride[i] = input_stride[i - 1] * input_ne[i - 1]; - scale_stride[i] = scale_stride[i - 1] * scale_ne[i - 1]; - } - - group_size_in_row = input_ne[0] / QK8_0; - int64_t scale_offset = input_ne[0] * input_ne[1] * input_ne[2] * - input_ne[3] * sizeof(int8_t); - - // Indices has two dims. n_elements = all rows should get. - // dr, all rows should this thread get. - uint64_t n_elements = - indices_ne[0] * indices_ne[1] * indices_ne[2] * indices_ne[3]; - dr = n_elements / op_block_num; - - uint64_t tails = n_elements % op_block_num; - if (op_block_idx < tails) { - dr += 1; - ir = dr * op_block_idx; - } else { - ir = dr * op_block_idx + tails; - } - - input_gm.SetGlobalBuffer((__gm__ int8_t *)input); - scale_gm.SetGlobalBuffer((__gm__ half *)(input + scale_offset)); - indices_gm.SetGlobalBuffer((__gm__ int32_t *)indices); - output_gm.SetGlobalBuffer((__gm__ float *)output); - - pipe.InitBuffer(input_queue, BUFFER_NUM, QK8_0 * sizeof(int8_t)); - pipe.InitBuffer(cast_queue, BUFFER_NUM, QK8_0 * sizeof(half)); - pipe.InitBuffer(output_queue, BUFFER_NUM, QK8_0 * sizeof(float)); - } - - __aicore__ inline void copy_in(uint32_t offset) { - LocalTensor input_local = input_queue.AllocTensor(); - DataCopy(input_local, input_gm[offset], QK8_0); - input_queue.EnQue(input_local); - } - - __aicore__ inline void copy_out(uint32_t offset) { - LocalTensor output_local = output_queue.DeQue(); - DataCopy(output_gm[offset], output_local, QK8_0); - output_queue.FreeTensor(output_local); - } - - __aicore__ inline void calculate_group(int64_t idx, int64_t group) { - const int64_t indices_ne2_idx = idx / (indices_ne[0] * indices_ne[1]); - const int64_t indices_ne1_idx = - (idx - indices_ne2_idx * indices_ne[0] * indices_ne[1]) / - indices_ne[0]; - const int64_t indices_ne0_idx = - (idx - indices_ne2_idx * indices_ne[0] * indices_ne[1] - - indices_ne1_idx * indices_ne[0]); - - const int64_t indices_offset = indices_ne0_idx * indices_stride[0] + - indices_ne1_idx * indices_stride[1] + - indices_ne2_idx * indices_stride[2]; - const int32_t selected_row_idx = indices_gm.GetValue(indices_offset); - - const int64_t input_offset = selected_row_idx * input_stride[1] + - indices_ne1_idx * input_stride[2] + - indices_ne2_idx * input_stride[3] + - group * QK8_0; - const int64_t scale_offset = selected_row_idx * scale_stride[1] + - indices_ne1_idx * scale_stride[2] + - indices_ne2_idx * scale_stride[3] + group; - const int64_t output_offset = indices_ne0_idx * output_stride[1] + - indices_ne1_idx * output_stride[2] + - indices_ne2_idx * output_stride[3] + - group * QK8_0; - - copy_in(input_offset); - LocalTensor input_local = input_queue.DeQue(); - LocalTensor cast_local = cast_queue.AllocTensor(); - LocalTensor output_local = output_queue.AllocTensor(); - - // TODO: cast more data to speed up. - Cast(cast_local, input_local, RoundMode::CAST_NONE, QK8_0); - Cast(output_local, cast_local, RoundMode::CAST_NONE, QK8_0); - - // Only mul need compile by group. - half scale = scale_gm.GetValue(scale_offset); - Muls(output_local, output_local, (float)scale, QK8_0); - - input_queue.FreeTensor(input_local); - cast_queue.FreeTensor(cast_local); - output_queue.EnQue(output_local); - - copy_out(output_offset); - } - - __aicore__ inline void calculate() { - for (int64_t i = ir; i < ir + dr; i++) { - for (int64_t j = 0; j < group_size_in_row; j++) { - calculate_group(i, j); - } - } - } - - private: - int64_t input_ne[4]; - size_t input_stride[4]; - - int64_t scale_ne[4]; - size_t scale_stride[4]; - - int64_t indices_ne[4]; - size_t indices_stride[4]; - - int64_t output_ne[4]; - size_t output_stride[4]; - - int64_t ir; - int64_t dr; - - int64_t group_size_in_row; - - TPipe pipe; - GlobalTensor input_gm; - GlobalTensor scale_gm; - GlobalTensor indices_gm; - GlobalTensor output_gm; - TQue input_queue; - TQue output_queue; - TQue cast_queue; -}; - -template -__aicore__ inline void copy_to_ub(GM_ADDR gm, T *ub, size_t size) { - auto gm_ptr = (__gm__ uint8_t *)gm; - auto ub_ptr = (uint8_t *)(ub); - for (int32_t i = 0; i < size; ++i, ++ub_ptr, ++gm_ptr) { - *ub_ptr = *gm_ptr; - } -} - -extern "C" __global__ __aicore__ void ascendc_get_row_q8_0( - GM_ADDR input_gm, GM_ADDR indices_gm, GM_ADDR output_gm, - GM_ADDR input_ne_gm, GM_ADDR indices_ne_gm, GM_ADDR indices_nb_gm, - GM_ADDR output_ne_gm, GM_ADDR output_nb_gm) { - int64_t input_ne_ub[4]; - int64_t indices_ne_ub[4]; - size_t indices_nb_ub[4]; - int64_t output_ne_ub[4]; - size_t output_nb_ub[4]; - - copy_to_ub(input_ne_gm, input_ne_ub, 32); - copy_to_ub(indices_ne_gm, indices_ne_ub, 32); - copy_to_ub(indices_nb_gm, indices_nb_ub, 32); - copy_to_ub(output_ne_gm, output_ne_ub, 32); - copy_to_ub(output_nb_gm, output_nb_ub, 32); - - GET_ROW_Q8_0 op; - op.init(input_gm, indices_gm, output_gm, input_ne_ub, indices_ne_ub, - indices_nb_ub, output_ne_ub, output_nb_ub); - op.calculate(); -} diff --git a/ggml/src/ggml-cann/kernels/quantize_f16_q8_0.cpp b/ggml/src/ggml-cann/kernels/quantize_f16_q8_0.cpp deleted file mode 100644 index 504b43afa..000000000 --- a/ggml/src/ggml-cann/kernels/quantize_f16_q8_0.cpp +++ /dev/null @@ -1,218 +0,0 @@ -#include "kernel_operator.h" - -using namespace AscendC; -#ifdef ASCEND_310P - extern "C" __global__ __aicore__ void ascendc_quantize_f16_q8_0( - GM_ADDR input_gm, GM_ADDR output_gm, GM_ADDR input_ne_gm, - GM_ADDR input_nb_gm, GM_ADDR output_ne_gm) { - // let following test cases can continue run, here just print error information. Of Cource the test case that call this operator is failed. - printf("Ascend310P not support f16->8bit quantization.\n"); - } -#else - -#define BUFFER_NUM 2 -#define QK8_0 32 - -class QUANTIZE_F16_Q8_0 { - public: - __aicore__ inline QUANTIZE_F16_Q8_0() {} - __aicore__ inline void init(GM_ADDR input, GM_ADDR output, - int64_t *input_ne_ub, size_t *input_nb_ub, - int64_t *output_ne_ub) { - int64_t op_block_num = GetBlockNum(); - int64_t op_block_idx = GetBlockIdx(); - - for (int i = 0; i < 4; i++) { - input_ne[i] = input_ne_ub[i]; - input_stride[i] = input_nb_ub[i] / input_nb_ub[0]; - - output_ne[i] = output_ne_ub[i]; - } - - output_stride[0] = 1; - for (int i = 1; i < 4; i++) { - output_stride[i] = output_stride[i - 1] * output_ne[i - 1]; - } - - scale_ne = input_ne; - scale_stride[0] = 1; - scale_stride[1] = input_ne[0] / QK8_0; - for (int i = 2; i < 4; i++) { - scale_stride[i] = scale_stride[i - 1] * scale_ne[i - 1]; - } - - // split input tensor by rows. - uint64_t nr = input_ne[1] * input_ne[2] * input_ne[3]; - dr = nr / op_block_num; - - uint64_t tails = nr % op_block_num; - if (op_block_idx < tails) { - dr += 1; - ir = dr * op_block_idx; - } else { - ir = dr * op_block_idx + tails; - } - - group_size_in_row = scale_stride[1]; - int64_t output_size = output_ne[0] * output_ne[1] * output_ne[2] * - output_ne[3] * sizeof(uint8_t); - - input_gm.SetGlobalBuffer((__gm__ half *)input); - output_gm.SetGlobalBuffer((__gm__ int8_t *)output); - scale_gm.SetGlobalBuffer((__gm__ half *)(output + output_size + ir * - group_size_in_row * - sizeof(half))); - - pipe.InitBuffer(input_queue, BUFFER_NUM, QK8_0 * sizeof(half)); - pipe.InitBuffer(output_queue, BUFFER_NUM, QK8_0 * sizeof(int8_t)); - pipe.InitBuffer(work_queue, 1, 32); - pipe.InitBuffer(max_queue, 1, 32); - pipe.InitBuffer(abs_queue, 1, QK8_0 * sizeof(float)); - pipe.InitBuffer(scale_queue, 1, 32); - pipe.InitBuffer(cast_queue ,1 ,QK8_0 * sizeof(float)); - } - - __aicore__ inline void copy_in(uint32_t offset) { - LocalTensor input_local = input_queue.AllocTensor(); - DataCopy(input_local, input_gm[offset], QK8_0); - input_queue.EnQue(input_local); - } - - __aicore__ inline void copy_out(uint32_t offset) { - LocalTensor output_local = output_queue.DeQue(); - DataCopy(output_gm[offset], output_local, QK8_0); - output_queue.FreeTensor(output_local); - } - - __aicore__ inline half calculate_group(int64_t row, int64_t group) { - const int64_t i3 = row / (input_ne[1] * input_ne[2]); - const int64_t i2 = (row - i3 * input_ne[1] * input_ne[2]) / input_ne[1]; - const int64_t i1 = - row - i3 * input_ne[1] * input_ne[2] - i2 * input_ne[1]; - - const int64_t input_offset = i1 * input_stride[1] + - i2 * input_stride[2] + - i3 * input_stride[3] + QK8_0 * group; - - const int64_t output_offset = i1 * output_stride[1] + - i2 * output_stride[2] + - i3 * output_stride[3] + QK8_0 * group; - - copy_in(input_offset); - LocalTensor input_local = input_queue.DeQue(); - LocalTensor output_local = output_queue.AllocTensor(); - LocalTensor work_local = work_queue.AllocTensor(); - LocalTensor abs_local = abs_queue.AllocTensor(); - LocalTensor max_local = max_queue.AllocTensor(); - LocalTensor cast_local = cast_queue.AllocTensor(); - - Cast(cast_local, input_local, RoundMode::CAST_NONE, QK8_0); - Abs(abs_local, cast_local, QK8_0); - ReduceMax(max_local, abs_local, work_local, QK8_0); - - pipe_barrier(PIPE_ALL); - float d = max_local.GetValue(0); - d = d / ((1 << 7) - 1); - if (d != 0) { - Muls(cast_local, cast_local, 1.0f / d, QK8_0); - } - - Cast(cast_local, cast_local, RoundMode::CAST_ROUND, QK8_0); - Cast(input_local, cast_local, RoundMode::CAST_ROUND, QK8_0); - Cast(output_local, input_local, RoundMode::CAST_ROUND, QK8_0); - output_queue.EnQue(output_local); - copy_out(output_offset); - - input_queue.FreeTensor(input_local); - work_queue.FreeTensor(work_local); - abs_queue.FreeTensor(abs_local); - max_queue.FreeTensor(max_local); - cast_queue.FreeTensor(cast_local); - return (half)d; - } - - __aicore__ inline void calculate() { - LocalTensor scale_local = scale_queue.AllocTensor(); - uint32_t scale_local_offset = 0; - uint32_t scale_global_offset = 0; - for (int64_t i = ir; i < ir + dr; i++) { - for (int64_t j = 0; j < group_size_in_row; j++) { - half scale = calculate_group(i, j); - scale_local.SetValue(scale_local_offset++, scale); - if (scale_local_offset == 16) { - scale_local_offset = 0; - // TODO: OPTIMIZE ME - pipe_barrier(PIPE_ALL); - DataCopy(scale_gm[scale_global_offset], scale_local, 16); - pipe_barrier(PIPE_ALL); - scale_global_offset += 16; - } - } - } - - if (scale_local_offset != 0) { - pipe_barrier(PIPE_ALL); - DataCopyExtParams dataCopyParams; - dataCopyParams.blockCount = 1; - dataCopyParams.blockLen = scale_local_offset * sizeof(half); - DataCopyPad(scale_gm[scale_global_offset], scale_local, - dataCopyParams); - pipe_barrier(PIPE_ALL); - } - } - - private: - int64_t input_ne[4]; - size_t input_stride[4]; - - int64_t *scale_ne; - size_t scale_stride[4]; - - int64_t output_ne[4]; - size_t output_stride[4]; - - int64_t group_size_in_row; - - int64_t ir; - int64_t dr; - - TPipe pipe; - GlobalTensor input_gm; - GlobalTensor scale_gm; - GlobalTensor output_gm; - TQue input_queue; - TQue output_queue; - TQue work_queue; - TQue max_queue; - TQue abs_queue; - TQue scale_queue; - TQue cast_queue; - -}; - -template -__aicore__ inline void copy_to_ub(GM_ADDR gm, T *ub, size_t size) { - auto gm_ptr = (__gm__ uint8_t *)gm; - auto ub_ptr = (uint8_t *)(ub); - for (int32_t i = 0; i < size; ++i, ++ub_ptr, ++gm_ptr) { - *ub_ptr = *gm_ptr; - } -} - -extern "C" __global__ __aicore__ void ascendc_quantize_f16_q8_0( - GM_ADDR input_gm, GM_ADDR output_gm, GM_ADDR input_ne_gm, - GM_ADDR input_nb_gm, GM_ADDR output_ne_gm) { - int64_t input_ne_ub[4]; - size_t input_nb_ub[4]; - int64_t output_ne_ub[4]; - - copy_to_ub(input_ne_gm, input_ne_ub, 32); - copy_to_ub(input_nb_gm, input_nb_ub, 32); - copy_to_ub(output_ne_gm, output_ne_ub, 32); - - QUANTIZE_F16_Q8_0 op; - op.init(input_gm, output_gm, input_ne_ub, input_nb_ub, output_ne_ub); - op.calculate(); -} - -#endif // #ifdef ASCEND_310P diff --git a/ggml/src/ggml-cann/kernels/quantize_f32_q8_0.cpp b/ggml/src/ggml-cann/kernels/quantize_f32_q8_0.cpp deleted file mode 100644 index 05b0bc1df..000000000 --- a/ggml/src/ggml-cann/kernels/quantize_f32_q8_0.cpp +++ /dev/null @@ -1,216 +0,0 @@ -#include "kernel_operator.h" - -using namespace AscendC; -#ifdef ASCEND_310P // 310P not support f32->8bit quantization - extern "C" __global__ __aicore__ void ascendc_quantize_f32_q8_0( - GM_ADDR input_gm, GM_ADDR output_gm, GM_ADDR input_ne_gm, - GM_ADDR input_nb_gm, GM_ADDR output_ne_gm) { - // let following test cases can continue run, here just print error information. Of Cource the test case that call this operator is failed. - printf("Ascend310P not support f32->8bit quantization.\n"); - } -#else - -#define BUFFER_NUM 2 -#define QK8_0 32 - -class QUANTIZE_F32_Q8_0 { - public: - __aicore__ inline QUANTIZE_F32_Q8_0() {} - __aicore__ inline void init(GM_ADDR input, GM_ADDR output, - int64_t *input_ne_ub, size_t *input_nb_ub, - int64_t *output_ne_ub) { - int64_t op_block_num = GetBlockNum(); - int64_t op_block_idx = GetBlockIdx(); - - for (int i = 0; i < 4; i++) { - input_ne[i] = input_ne_ub[i]; - input_stride[i] = input_nb_ub[i] / input_nb_ub[0]; - - output_ne[i] = output_ne_ub[i]; - } - - output_stride[0] = 1; - for (int i = 1; i < 4; i++) { - output_stride[i] = output_stride[i - 1] * output_ne[i - 1]; - } - - scale_ne = input_ne; - scale_stride[0] = 1; - scale_stride[1] = input_ne[0] / QK8_0; - for (int i = 2; i < 4; i++) { - scale_stride[i] = scale_stride[i - 1] * scale_ne[i - 1]; - } - - // split input tensor by rows. - uint64_t nr = input_ne[1] * input_ne[2] * input_ne[3]; - dr = nr / op_block_num; - - uint64_t tails = nr % op_block_num; - if (op_block_idx < tails) { - dr += 1; - ir = dr * op_block_idx; - } else { - ir = dr * op_block_idx + tails; - } - - group_size_in_row = scale_stride[1]; - int64_t output_size = output_ne[0] * output_ne[1] * output_ne[2] * - output_ne[3] * sizeof(uint8_t); - - input_gm.SetGlobalBuffer((__gm__ float *)input); - output_gm.SetGlobalBuffer((__gm__ int8_t *)output); - scale_gm.SetGlobalBuffer((__gm__ half *)(output + output_size + - ir * group_size_in_row * - sizeof(half))); - - pipe.InitBuffer(input_queue, BUFFER_NUM, QK8_0 * sizeof(float)); - pipe.InitBuffer(output_queue, BUFFER_NUM, QK8_0 * sizeof(int8_t)); - pipe.InitBuffer(work_queue, 1, 32); - pipe.InitBuffer(max_queue, 1, 32); - pipe.InitBuffer(abs_queue, 1, QK8_0 * sizeof(float)); - pipe.InitBuffer(cast_queue, 1, QK8_0 * sizeof(half)); - pipe.InitBuffer(scale_queue, 1, 32); - } - - __aicore__ inline void copy_in(uint32_t offset) { - LocalTensor input_local = input_queue.AllocTensor(); - DataCopy(input_local, input_gm[offset], QK8_0); - input_queue.EnQue(input_local); - } - - __aicore__ inline void copy_out(uint32_t offset) { - LocalTensor output_local = output_queue.DeQue(); - DataCopy(output_gm[offset], output_local, QK8_0); - output_queue.FreeTensor(output_local); - } - - __aicore__ inline half calculate_group(int64_t row, int64_t group) { - const int64_t i3 = row / (input_ne[1] * input_ne[2]); - const int64_t i2 = (row - i3 * input_ne[1] * input_ne[2]) / input_ne[1]; - const int64_t i1 = - row - i3 * input_ne[1] * input_ne[2] - i2 * input_ne[1]; - - const int64_t input_offset = i1 * input_stride[1] + - i2 * input_stride[2] + - i3 * input_stride[3] + QK8_0 * group; - - const int64_t output_offset = i1 * output_stride[1] + - i2 * output_stride[2] + - i3 * output_stride[3] + QK8_0 * group; - - copy_in(input_offset); - LocalTensor input_local = input_queue.DeQue(); - LocalTensor output_local = output_queue.AllocTensor(); - LocalTensor work_local = work_queue.AllocTensor(); - LocalTensor abs_local = abs_queue.AllocTensor(); - LocalTensor max_local = max_queue.AllocTensor(); - LocalTensor cast_local = cast_queue.AllocTensor(); - - Abs(abs_local, input_local, QK8_0); - ReduceMax(max_local, abs_local, work_local, QK8_0); - pipe_barrier(PIPE_ALL); - float d = max_local.GetValue(0); - d = d / ((1 << 7) - 1); - if (d != 0) { - Muls(input_local, input_local, 1.0f / d, QK8_0); - } - - Cast(input_local, input_local, RoundMode::CAST_ROUND, QK8_0); - Cast(cast_local, input_local, RoundMode::CAST_ROUND, QK8_0); - Cast(output_local, cast_local, RoundMode::CAST_ROUND, QK8_0); - output_queue.EnQue(output_local); - copy_out(output_offset); - - input_queue.FreeTensor(input_local); - work_queue.FreeTensor(work_local); - abs_queue.FreeTensor(abs_local); - max_queue.FreeTensor(max_local); - cast_queue.FreeTensor(cast_local); - - return (half)d; - } - - __aicore__ inline void calculate() { - LocalTensor scale_local = scale_queue.AllocTensor(); - uint32_t scale_local_offset = 0; - uint32_t scale_global_offset = 0; - for (int64_t i = ir; i < ir + dr; i++) { - for (int64_t j = 0; j < group_size_in_row; j++) { - half scale = calculate_group(i, j); - scale_local.SetValue(scale_local_offset++, scale); - if (scale_local_offset == 16) { - scale_local_offset = 0; - // TODO: OPTIMIZE ME - pipe_barrier(PIPE_ALL); - DataCopy(scale_gm[scale_global_offset], scale_local, 16); - pipe_barrier(PIPE_ALL); - scale_global_offset += 16; - } - } - } - - if (scale_local_offset != 0) { - pipe_barrier(PIPE_ALL); - DataCopyExtParams dataCopyParams; - dataCopyParams.blockCount = 1; - dataCopyParams.blockLen = scale_local_offset * sizeof(half); - DataCopyPad(scale_gm[scale_global_offset], scale_local, - dataCopyParams); - pipe_barrier(PIPE_ALL); - } - } - - private: - int64_t input_ne[4]; - size_t input_stride[4]; - - int64_t *scale_ne; - size_t scale_stride[4]; - - int64_t output_ne[4]; - size_t output_stride[4]; - - int64_t group_size_in_row; - - int64_t ir; - int64_t dr; - - TPipe pipe; - GlobalTensor input_gm; - GlobalTensor scale_gm; - GlobalTensor output_gm; - TQue input_queue; - TQue output_queue; - TQue work_queue; - TQue max_queue; - TQue abs_queue; - TQue cast_queue; - TQue scale_queue; -}; - -template -__aicore__ inline void copy_to_ub(GM_ADDR gm, T *ub, size_t size) { - auto gm_ptr = (__gm__ uint8_t *)gm; - auto ub_ptr = (uint8_t *)(ub); - for (int32_t i = 0; i < size; ++i, ++ub_ptr, ++gm_ptr) { - *ub_ptr = *gm_ptr; - } -} - -extern "C" __global__ __aicore__ void ascendc_quantize_f32_q8_0( - GM_ADDR input_gm, GM_ADDR output_gm, GM_ADDR input_ne_gm, - GM_ADDR input_nb_gm, GM_ADDR output_ne_gm) { - int64_t input_ne_ub[4]; - size_t input_nb_ub[4]; - int64_t output_ne_ub[4]; - - copy_to_ub(input_ne_gm, input_ne_ub, 32); - copy_to_ub(input_nb_gm, input_nb_ub, 32); - copy_to_ub(output_ne_gm, output_ne_ub, 32); - - QUANTIZE_F32_Q8_0 op; - op.init(input_gm, output_gm, input_ne_ub, input_nb_ub, output_ne_ub); - op.calculate(); -} - -#endif // #ifdef ASCEND_310P diff --git a/ggml/src/ggml-cann/kernels/quantize_float_to_q4_0.cpp b/ggml/src/ggml-cann/kernels/quantize_float_to_q4_0.cpp deleted file mode 100644 index 1188937b7..000000000 --- a/ggml/src/ggml-cann/kernels/quantize_float_to_q4_0.cpp +++ /dev/null @@ -1,295 +0,0 @@ -#include "kernel_operator.h" - -using namespace AscendC; -#ifdef ASCEND_310P // 310P not support float->4bit quantization - extern "C" __global__ __aicore__ void ascendc_quantize_f32_to_q4_0( - GM_ADDR input_gm, GM_ADDR output_gm, GM_ADDR input_ne_gm, - GM_ADDR input_nb_gm, GM_ADDR output_ne_gm) { - // let following test cases can continue run, here just print error information. Of Cource the test case that call this operator is failed. - printf("Ascend310P not support f32->4bit quantization.\n"); - } - - extern "C" __global__ __aicore__ void ascendc_quantize_f16_to_q4_0( - GM_ADDR input_gm, GM_ADDR output_gm, GM_ADDR input_ne_gm, - GM_ADDR input_nb_gm, GM_ADDR output_ne_gm) { - // let following test cases can continue run, here just print error information. Of Cource the test case that call this operator is failed. - printf("Ascend310P not support f16->4bit quantization.\n"); - } -#else - -#define BUFFER_NUM 2 -#define Group_Size 32 - -template -class QUANTIZE_FLOAT_TO_Q4_0 { - public: - __aicore__ inline QUANTIZE_FLOAT_TO_Q4_0() {} - __aicore__ inline void init(GM_ADDR input, GM_ADDR output, - int64_t *input_ne_ub, size_t *input_nb_ub, - int64_t *output_ne_ub) { - // TODO: fix test_case CPY(type_src=f16,type_dst=q4_0,ne=[256,4,4,4], - // permute=[0,0,0,0]): - // [CPY] NMSE = 0.000008343 > 0.000001000 FAIL - int64_t op_block_num = GetBlockNum(); - int64_t op_block_idx = GetBlockIdx(); - - // input stride of data elements - for (int i = 0; i < 4; i++) { - input_ne[i] = input_ne_ub[i]; - input_stride[i] = input_nb_ub[i] / input_nb_ub[0]; - output_ne[i] = output_ne_ub[i]; - } - - // output stride of data elements - output_stride[0] = 1; - for (int i = 1; i < 4; i++) { - output_stride[i] = output_stride[i - 1] * output_ne[i - 1]; - } - - // scale saved one by one after data:. [group1_scale, group2_scale, ...] - scale_ne = input_ne; - scale_stride[0] = 1; - scale_stride[1] = input_ne[0] / Group_Size; - for (int i = 2; i < 4; i++) { - scale_stride[i] = scale_stride[i - 1] * scale_ne[i - 1]; - } - - // split input tensor by rows. - uint64_t nr = input_ne[1] * input_ne[2] * input_ne[3]; - dr = nr / op_block_num; - - uint64_t tails = nr % op_block_num; - if (op_block_idx < tails) { - dr += 1; - ir = dr * op_block_idx; - } else { - ir = dr * op_block_idx + tails; - } - - group_size_in_row = scale_stride[1]; - int64_t scale_offset = output_ne[0] * output_ne[1] * output_ne[2] * - output_ne[3] * sizeof(uint8_t) / 2; - - input_gm.SetGlobalBuffer((__gm__ SRC_T *)input); - output_gm.SetGlobalBuffer((__gm__ int8_t *)output); - scale_gm.SetGlobalBuffer((__gm__ half *)(output + scale_offset + ir * - group_size_in_row * - sizeof(half))); - - pipe.InitBuffer(input_queue, BUFFER_NUM, Group_Size * sizeof(SRC_T)); - pipe.InitBuffer(output_queue, BUFFER_NUM, - Group_Size * sizeof(int8_t) / 2); - pipe.InitBuffer(cast_queue , 1, Group_Size * sizeof(float)); - pipe.InitBuffer(work_queue, 1, Group_Size * sizeof(float)); - pipe.InitBuffer(max_queue, 1, Group_Size * sizeof(float)); - pipe.InitBuffer(min_queue, 1, Group_Size * sizeof(float)); - pipe.InitBuffer(scale_queue, 1, Group_Size / 2 * sizeof(half)); - pipe.InitBuffer(int8_queue, 1, Group_Size * sizeof(int8_t)); - pipe.InitBuffer(half_queue, 1, Group_Size * sizeof(half)); - } - - __aicore__ inline void copy_in(uint32_t offset) { - LocalTensor input_local = input_queue.AllocTensor(); - DataCopy(input_local, input_gm[offset], Group_Size); - input_queue.EnQue(input_local); - } - - __aicore__ inline void copy_out(uint32_t offset) { - // reinterpretcast Group_Size(32) * int4b_t to Group_Size / 2 * int8_t, - // and using DataCopyPad to avoid 32 bits align. - LocalTensor output_local = output_queue.DeQue(); - LocalTensor output_int8_local = - output_local.ReinterpretCast(); - - DataCopyExtParams dataCopyParams; - dataCopyParams.blockCount = 1; - dataCopyParams.blockLen = Group_Size / 2 * sizeof(int8_t); - DataCopyPad(output_gm[offset], output_int8_local, dataCopyParams); - - output_queue.FreeTensor(output_local); - } - - __aicore__ inline void input_to_cast(LocalTensor cast_local, - LocalTensor input_local) { - DataCopy(cast_local, input_local, Group_Size); - } - - __aicore__ inline void input_to_cast(LocalTensor cast_local, - LocalTensor input_local) { - Cast(cast_local, input_local, RoundMode::CAST_NONE, Group_Size); - } - - __aicore__ inline half calculate_group(int64_t row, int64_t group) { - const int64_t i3 = row / (input_ne[1] * input_ne[2]); - const int64_t i2 = (row - i3 * input_ne[1] * input_ne[2]) / input_ne[1]; - const int64_t i1 = - row - i3 * input_ne[1] * input_ne[2] - i2 * input_ne[1]; - - const int64_t input_offset = i1 * input_stride[1] + - i2 * input_stride[2] + - i3 * input_stride[3] + Group_Size * group; - - // output_offset is stride for output_gm which datatype is int8_t and - // divided by 2 is needed for int4b_t. - const int64_t output_offset = (i1 * output_stride[1] + - i2 * output_stride[2] + - i3 * output_stride[3] + - Group_Size * group) / 2; - copy_in(input_offset); - - LocalTensor input_local = input_queue.DeQue(); - LocalTensor output_local = output_queue.AllocTensor(); - LocalTensor cast_local = cast_queue.AllocTensor(); - LocalTensor work_local = work_queue.AllocTensor(); - LocalTensor max_local = max_queue.AllocTensor(); - LocalTensor min_local = min_queue.AllocTensor(); - LocalTensor int8_local = int8_queue.AllocTensor(); - LocalTensor half_local = half_queue.AllocTensor(); - - input_to_cast(cast_local, input_local); - - ReduceMax(max_local, cast_local, work_local, Group_Size); - ReduceMin(min_local, cast_local, work_local, Group_Size); - const float max_value = max_local.GetValue(0); - const float min_value = min_local.GetValue(0); - float d = max_value; - if (min_value < 0 && (-1 * min_value) > max_value) { - d = min_value; - } - - d = d / (-8); - if (d != 0) { - Muls(cast_local, cast_local, 1.0f / d, Group_Size); - } - - // range: [-8,8] -> [0.5,16.5] -> [0,16] -> [0,15] -> [-8,7] - float scalar = 8.5f; - Adds(cast_local, cast_local, scalar, Group_Size); - Cast(cast_local, cast_local, RoundMode::CAST_FLOOR, Group_Size); - scalar = 15.0f; - Mins(cast_local, cast_local, scalar, Group_Size); - scalar = -8.0f; - Adds(cast_local, cast_local, scalar, Group_Size); - - // float->half->int4b - Cast(half_local, cast_local, RoundMode::CAST_NONE, Group_Size); - Cast(output_local, half_local, RoundMode::CAST_NONE, Group_Size); - - output_queue.EnQue(output_local); - copy_out(output_offset); - - input_queue.FreeTensor(input_local); - work_queue.FreeTensor(work_local); - max_queue.FreeTensor(max_local); - min_queue.FreeTensor(min_local); - int8_queue.FreeTensor(int8_local); - half_queue.FreeTensor(half_local); - cast_queue.FreeTensor(cast_local); - return (half)d; - } - - __aicore__ inline void calculate() { - LocalTensor scale_local = scale_queue.AllocTensor(); - uint32_t scale_local_offset = 0; - uint32_t scale_global_offset = 0; - for (int64_t i = ir; i < ir + dr; i++) { - for (int64_t j = 0; j < group_size_in_row; j++) { - half scale = calculate_group(i, j); - scale_local.SetValue(scale_local_offset++, scale); - // Copy Group_Size/2 length data each time. - if (scale_local_offset == Group_Size / 2) { - scale_local_offset = 0; - // TODO: OPTIMIZE ME - pipe_barrier(PIPE_ALL); - DataCopy(scale_gm[scale_global_offset], scale_local, - Group_Size / 2); - pipe_barrier(PIPE_ALL); - scale_global_offset += Group_Size / 2; - } - } - } - - if (scale_local_offset != 0) { - pipe_barrier(PIPE_ALL); - DataCopyExtParams dataCopyParams; - dataCopyParams.blockCount = 1; - dataCopyParams.blockLen = scale_local_offset * sizeof(half); - DataCopyPad(scale_gm[scale_global_offset], scale_local, - dataCopyParams); - pipe_barrier(PIPE_ALL); - } - scale_queue.FreeTensor(scale_local); - } - - private: - int64_t input_ne[4]; - size_t input_stride[4]; - - int64_t *scale_ne; - size_t scale_stride[4]; - - int64_t output_ne[4]; - size_t output_stride[4]; - - int64_t group_size_in_row; - - int64_t ir; - int64_t dr; - - TPipe pipe; - GlobalTensor input_gm; - GlobalTensor scale_gm; - GlobalTensor output_gm; - TQue input_queue; - TQue output_queue; - TQue work_queue; - TQue max_queue; - TQue min_queue; - TQue scale_queue; - TQue cast_queue; - TQue int8_queue; - TQue half_queue; -}; - -template -__aicore__ inline void copy_to_ub(GM_ADDR gm, T *ub, size_t size) { - auto gm_ptr = (__gm__ uint8_t *)gm; - auto ub_ptr = (uint8_t *)(ub); - for (int32_t i = 0; i < size; ++i, ++ub_ptr, ++gm_ptr) { - *ub_ptr = *gm_ptr; - } -} - -extern "C" __global__ __aicore__ void ascendc_quantize_f16_to_q4_0( - GM_ADDR input_gm, GM_ADDR output_gm, GM_ADDR input_ne_gm, - GM_ADDR input_nb_gm, GM_ADDR output_ne_gm) { - int64_t input_ne_ub[4]; - size_t input_nb_ub[4]; - int64_t output_ne_ub[4]; - - copy_to_ub(input_ne_gm, input_ne_ub, 32); - copy_to_ub(input_nb_gm, input_nb_ub, 32); - copy_to_ub(output_ne_gm, output_ne_ub, 32); - - QUANTIZE_FLOAT_TO_Q4_0 op; - op.init(input_gm, output_gm, input_ne_ub, input_nb_ub, output_ne_ub); - op.calculate(); -} - -extern "C" __global__ __aicore__ void ascendc_quantize_f32_to_q4_0( - GM_ADDR input_gm, GM_ADDR output_gm, GM_ADDR input_ne_gm, - GM_ADDR input_nb_gm, GM_ADDR output_ne_gm) { - int64_t input_ne_ub[4]; - size_t input_nb_ub[4]; - int64_t output_ne_ub[4]; - - copy_to_ub(input_ne_gm, input_ne_ub, 32); - copy_to_ub(input_nb_gm, input_nb_ub, 32); - copy_to_ub(output_ne_gm, output_ne_ub, 32); - - QUANTIZE_FLOAT_TO_Q4_0 op; - op.init(input_gm, output_gm, input_ne_ub, input_nb_ub, output_ne_ub); - op.calculate(); -} - -#endif // #ifdef ASCEND_310P diff --git a/ggml/src/ggml-common.h b/ggml/src/ggml-common.h index f13fd4dea..086c822d7 100644 --- a/ggml/src/ggml-common.h +++ b/ggml/src/ggml-common.h @@ -158,6 +158,12 @@ typedef sycl::half2 ggml_half2; #endif // GGML_COMMON_DECL_CUDA || GGML_COMMON_DECL_HIP +#ifdef _MSC_VER +#define GGML_EXTENSION +#else // _MSC_VER +#define GGML_EXTENSION __extension__ +#endif // _MSC_VER + #define QK4_0 32 typedef struct { ggml_half d; // delta @@ -167,7 +173,7 @@ static_assert(sizeof(block_q4_0) == sizeof(ggml_half) + QK4_0 / 2, "wrong q4_0 b #define QK4_1 32 typedef struct { - union { + GGML_EXTENSION union { struct { ggml_half d; // delta ggml_half m; // min @@ -188,7 +194,7 @@ static_assert(sizeof(block_q5_0) == sizeof(ggml_half) + sizeof(uint32_t) + QK5_0 #define QK5_1 32 typedef struct { - union { + GGML_EXTENSION union { struct { ggml_half d; // delta ggml_half m; // min @@ -209,7 +215,7 @@ static_assert(sizeof(block_q8_0) == sizeof(ggml_half) + QK8_0, "wrong q8_0 block #define QK8_1 32 typedef struct { - union { + GGML_EXTENSION union { struct { ggml_half d; // delta ggml_half s; // d * sum(qs[i]) @@ -250,7 +256,7 @@ static_assert(sizeof(block_tq2_0) == sizeof(ggml_half) + QK_K / 4, "wrong tq2_0 typedef struct { uint8_t scales[QK_K/16]; // scales and mins, quantized with 4 bits uint8_t qs[QK_K/4]; // quants - union { + GGML_EXTENSION union { struct { ggml_half d; // super-block scale for quantized scales ggml_half dmin; // super-block scale for quantized mins @@ -277,7 +283,7 @@ static_assert(sizeof(block_q3_K) == sizeof(ggml_half) + QK_K / 4 + QK_K / 8 + 12 // weight is represented as x = a * q + b // Effectively 4.5 bits per weight typedef struct { - union { + GGML_EXTENSION union { struct { ggml_half d; // super-block scale for quantized scales ggml_half dmin; // super-block scale for quantized mins @@ -294,7 +300,7 @@ static_assert(sizeof(block_q4_K) == 2*sizeof(ggml_half) + K_SCALE_SIZE + QK_K/2, // weight is represented as x = a * q + b // Effectively 5.5 bits per weight typedef struct { - union { + GGML_EXTENSION union { struct { ggml_half d; // super-block scale for quantized scales ggml_half dmin; // super-block scale for quantized mins @@ -473,7 +479,6 @@ GGML_TABLE_BEGIN(uint8_t, ksigns_iq2xs, 128) 240, 113, 114, 243, 116, 245, 246, 119, 120, 249, 250, 123, 252, 125, 126, 255, GGML_TABLE_END() -//#if __CUDA_ARCH__ >= GGML_CUDA_CC_DP4A // lowest compute capability for integer intrinsics GGML_TABLE_BEGIN(uint64_t, ksigns64, 128) 0x0000000000000000, 0xff000000000000ff, 0xff0000000000ff00, 0x000000000000ffff, 0xff00000000ff0000, 0x0000000000ff00ff, 0x0000000000ffff00, 0xff00000000ffffff, @@ -508,7 +513,6 @@ GGML_TABLE_BEGIN(uint64_t, ksigns64, 128) 0x00ffffffff000000, 0xffffffffff0000ff, 0xffffffffff00ff00, 0x00ffffffff00ffff, 0xffffffffffff0000, 0x00ffffffffff00ff, 0x00ffffffffffff00, 0xffffffffffffffff, GGML_TABLE_END() -//#endif GGML_TABLE_BEGIN(uint64_t, iq2xxs_grid, 256) diff --git a/ggml/src/ggml-cpu/CMakeLists.txt b/ggml/src/ggml-cpu/CMakeLists.txt index 6b3641c42..e73a3b69b 100644 --- a/ggml/src/ggml-cpu/CMakeLists.txt +++ b/ggml/src/ggml-cpu/CMakeLists.txt @@ -23,6 +23,16 @@ function(ggml_add_cpu_backend_variant_impl tag_name) ggml-cpu/amx/mmq.cpp ggml-cpu/amx/mmq.h ggml-cpu/ggml-cpu-impl.h + ggml-cpu/common.h + ggml-cpu/binary-ops.h + ggml-cpu/binary-ops.cpp + ggml-cpu/unary-ops.h + ggml-cpu/unary-ops.cpp + ggml-cpu/simd-mappings.h + ggml-cpu/vec.h + ggml-cpu/vec.cpp + ggml-cpu/ops.h + ggml-cpu/ops.cpp ) target_compile_features(${GGML_CPU_NAME} PRIVATE c_std_11 cxx_std_17) @@ -111,14 +121,15 @@ function(ggml_add_cpu_backend_variant_impl tag_name) function(check_arm_feature tag code) set(CMAKE_REQUIRED_FLAGS_SAVE ${CMAKE_REQUIRED_FLAGS}) set(CMAKE_REQUIRED_FLAGS "${ARM_MCPU_FLAG}+${tag}") - check_cxx_source_runs( - "${code}" - GGML_MACHINE_SUPPORTS_${tag} - ) + check_cxx_source_runs("${code}" GGML_MACHINE_SUPPORTS_${tag}) if (GGML_MACHINE_SUPPORTS_${tag}) set(ARM_MCPU_FLAG_FIX "${ARM_MCPU_FLAG_FIX}+${tag}" PARENT_SCOPE) else() - set(ARM_MCPU_FLAG_FIX "${ARM_MCPU_FLAG_FIX}+no${tag}" PARENT_SCOPE) + set(CMAKE_REQUIRED_FLAGS "${ARM_MCPU_FLAG}+no${tag}") + check_cxx_source_compiles("int main() { return 0; }" GGML_MACHINE_SUPPORTS_no${tag}) + if (GGML_MACHINE_SUPPORTS_no${tag}) + set(ARM_MCPU_FLAG_FIX "${ARM_MCPU_FLAG_FIX}+no${tag}" PARENT_SCOPE) + endif() endif() set(CMAKE_REQUIRED_FLAGS ${CMAKE_REQUIRED_FLAGS_SAVE}) endfunction() @@ -126,6 +137,7 @@ function(ggml_add_cpu_backend_variant_impl tag_name) check_arm_feature(dotprod "#include \nint main() { int8x16_t _a, _b; volatile int32x4_t _s = vdotq_s32(_s, _a, _b); return 0; }") check_arm_feature(i8mm "#include \nint main() { int8x16_t _a, _b; volatile int32x4_t _s = vmmlaq_s32(_s, _a, _b); return 0; }") check_arm_feature(sve "#include \nint main() { svfloat32_t _a, _b; volatile svfloat32_t _c = svadd_f32_z(svptrue_b8(), _a, _b); return 0; }") + check_arm_feature(sme "#include \n__arm_locally_streaming int main() { __asm__ volatile(\"smstart; smstop;\"); return 0; }") list(APPEND ARCH_FLAGS "${ARM_MCPU_FLAG}${ARM_MCPU_FLAG_FIX}") else() @@ -150,7 +162,7 @@ function(ggml_add_cpu_backend_variant_impl tag_name) if (ARM_FEATURE_RESULT) message(WARNING "Failed to get ARM features") else() - foreach(feature DOTPROD SVE MATMUL_INT8 FMA FP16_VECTOR_ARITHMETIC) + foreach(feature DOTPROD SVE MATMUL_INT8 FMA FP16_VECTOR_ARITHMETIC SME) string(FIND "${ARM_FEATURE}" "__ARM_FEATURE_${feature} 1" feature_pos) if (NOT ${feature_pos} EQUAL -1) message(STATUS "ARM feature ${feature} enabled") @@ -217,6 +229,10 @@ function(ggml_add_cpu_backend_variant_impl tag_name) if (GGML_AVX_VNNI) list(APPEND ARCH_DEFINITIONS __AVXVNNI__ GGML_AVX_VNNI) endif() + if (GGML_BMI2) + # MSVC does not define macro __BMI2__ + list(APPEND ARCH_DEFINITIONS __BMI2__ GGML_BMI2) + endif() else () if (GGML_NATIVE) list(APPEND ARCH_FLAGS -march=native) @@ -231,6 +247,10 @@ function(ggml_add_cpu_backend_variant_impl tag_name) list(APPEND ARCH_FLAGS -mfma) list(APPEND ARCH_DEFINITIONS GGML_FMA) endif() + if (GGML_BMI2) + list(APPEND ARCH_FLAGS -mbmi2) + list(APPEND ARCH_DEFINITIONS GGML_BMI2) + endif() if (GGML_AVX) list(APPEND ARCH_FLAGS -mavx) list(APPEND ARCH_DEFINITIONS GGML_AVX) @@ -277,21 +297,31 @@ function(ggml_add_cpu_backend_variant_impl tag_name) endif() endif() endif() - elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64") + elseif ("${CMAKE_SYSTEM_PROCESSOR} " STREQUAL "ppc64le " OR "${CMAKE_SYSTEM_PROCESSOR} " STREQUAL "powerpc ") message(STATUS "PowerPC detected") - execute_process(COMMAND bash -c "grep POWER10 /proc/cpuinfo | head -n 1" OUTPUT_VARIABLE POWER10_M) - string(FIND "${POWER10_M}" "POWER10" substring_index) - if (NOT DEFINED substring_index OR "${substring_index}" STREQUAL "") - set(substring_index -1) - endif() + if (GGML_NATIVE) + if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64") + file(READ "/proc/cpuinfo" POWER10_M) + elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "powerpc") + execute_process(COMMAND bash -c "prtconf |grep 'Implementation' | head -n 1" OUTPUT_VARIABLE POWER10_M) + endif() - if (${substring_index} GREATER_EQUAL 0) - list(APPEND ARCH_FLAGS -mcpu=power10) - elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64le") - list(APPEND ARCH_FLAGS -mcpu=powerpc64le) + string(REGEX MATCHALL "POWER *([0-9]+)" MATCHED_STRING "${POWER10_M}") + string(REGEX REPLACE "POWER *([0-9]+)" "\\1" EXTRACTED_NUMBER "${MATCHED_STRING}") + + if (EXTRACTED_NUMBER GREATER_EQUAL 10) + list(APPEND ARCH_FLAGS -mcpu=power10 -mpowerpc64) + elseif (EXTRACTED_NUMBER EQUAL 9) + list(APPEND ARCH_FLAGS -mcpu=power9 -mpowerpc64) + elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64le") + list(APPEND ARCH_FLAGS -mcpu=powerpc64le -mtune=native) + else() + list(APPEND ARCH_FLAGS -mcpu=native -mtune=native -mpowerpc64) + endif() else() - list(APPEND ARCH_FLAGS -mcpu=native -mtune=native) - # TODO: Add targets for Power8/Power9 (Altivec/VSX) and Power10(MMA) and query for big endian systems (ppc64/le/be) + if (GGML_CPU_POWERPC_CPUTYPE) + list(APPEND ARCH_FLAGS -mcpu=${GGML_CPU_POWERPC_CPUTYPE}) + endif() endif() elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "loongarch64") message(STATUS "loongarch64 detected") @@ -306,7 +336,32 @@ function(ggml_add_cpu_backend_variant_impl tag_name) elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "riscv64") message(STATUS "RISC-V detected") if (GGML_RVV) - list(APPEND ARCH_FLAGS -march=rv64gcv -mabi=lp64d) + if (GGML_RV_ZFH) + list(APPEND ARCH_FLAGS -march=rv64gcv_zfhmin -DGGML_RV_ZFH -mabi=lp64d) + else() + list(APPEND ARCH_FLAGS -march=rv64gcv -mabi=lp64d) + endif() + endif() + elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "s390x") + message(STATUS "s390x detected") + file(READ "/proc/cpuinfo" CPUINFO_CONTENTS) + string(REGEX REPLACE "machine[ \t\r\n]*=[ \t\r\n]*([0-9]+)" "\\1" S390X_M ${CPUINFO_CONTENTS}) + + # TODO: Separation to determine activation of VX/VXE/VXE2 + if (${S390X_M} MATCHES "8561|8562") + message(STATUS "z15 target") + list(APPEND ARCH_FLAGS -march=z15 -mtune=z15) + elseif (${S390X_M} MATCHES "3931") + message(STATUS "z16 target") + list(APPEND ARCH_FLAGS -march=z16 -mtune=z16) + else() + message(STATUS "Unknown target") + message(WARNING "Unknown target. If you are compiling for z14 and earlier, you might have to add -DGGML_VXE=OFF.") + list(APPEND ARCH_FLAGS -march=native -mtune=native) + endif() + + if (GGML_VXE) + list(APPEND ARCH_FLAGS -mvx -mzvector) endif() else() message(STATUS "Unknown architecture") @@ -316,6 +371,94 @@ function(ggml_add_cpu_backend_variant_impl tag_name) target_compile_definitions(${GGML_CPU_NAME} PRIVATE GGML_USE_CPU_AARCH64) endif() + if (GGML_CPU_KLEIDIAI) + message(STATUS "Using KleidiAI optimized kernels if applicable") + + # Disable the KleidiAI tests + set(KLEIDIAI_BUILD_TESTS OFF) + + # Fetch KleidiAI sources: + include(FetchContent) + set(KLEIDIAI_COMMIT_TAG "v1.5.0") + set(KLEIDIAI_DOWNLOAD_URL "https://github.com/ARM-software/kleidiai/archive/refs/tags/${KLEIDIAI_COMMIT_TAG}.tar.gz") + set(KLEIDIAI_ARCHIVE_MD5 "ea22e1aefb800e9bc8c74d91633cc58e") + + if (POLICY CMP0135) + cmake_policy(SET CMP0135 NEW) + endif() + + FetchContent_Declare(KleidiAI_Download + URL ${KLEIDIAI_DOWNLOAD_URL} + DOWNLOAD_EXTRACT_TIMESTAMP NEW + URL_HASH MD5=${KLEIDIAI_ARCHIVE_MD5}) + + FetchContent_MakeAvailable(KleidiAI_Download) + FetchContent_GetProperties(KleidiAI_Download + SOURCE_DIR KLEIDIAI_SRC + POPULATED KLEIDIAI_POPULATED) + + if (NOT KLEIDIAI_POPULATED) + message(FATAL_ERROR "KleidiAI source downloaded failed.") + endif() + + add_compile_definitions(GGML_USE_CPU_KLEIDIAI) + + # Remove kleidiai target after fetching it + if (TARGET kleidiai) + set_target_properties(kleidiai PROPERTIES EXCLUDE_FROM_ALL TRUE) + endif() + + list(APPEND GGML_CPU_SOURCES + ggml-cpu/kleidiai/kleidiai.cpp + ggml-cpu/kleidiai/kernels.cpp + ggml-cpu/kleidiai/kleidiai.h + ggml-cpu/kleidiai/kernels.h + ) + + # KleidiAI + include_directories( + ${KLEIDIAI_SRC}/ + ${KLEIDIAI_SRC}/kai/ + ${KLEIDIAI_SRC}/kai/ukernels/ + ${KLEIDIAI_SRC}/kai/ukernels/matmul/ + ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/ + ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/) + + set(ARCH_FLAGS_TEMP "${ARCH_FLAGS}") + if (NOT ARCH_FLAGS_TEMP) + string(REGEX MATCH "-march=[^ ]+" ARCH_FLAGS_TEMP "${CMAKE_C_FLAGS}") + endif() + string(FIND "${ARCH_FLAGS_TEMP}" "+dotprod" DOTPROD_ENABLED) + string(FIND "${ARCH_FLAGS_TEMP}" "+i8mm" I8MM_ENABLED) + string(FIND "${ARCH_FLAGS_TEMP}" "+sme" SME_ENABLED) + + set(PRIVATE_ARCH_FLAGS ${ARCH_FLAGS}) + + list(APPEND GGML_KLEIDIAI_SOURCES ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_lhs_quant_pack_qsi8d32p_f32.c) + list(APPEND GGML_KLEIDIAI_SOURCES ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon.c) + list(APPEND GGML_KLEIDIAI_SOURCES ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_lhs_quant_pack_qsi8d32p_f32_neon.c) + list(APPEND GGML_KLEIDIAI_SOURCES ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0.c) + + if (NOT DOTPROD_ENABLED MATCHES -1) + list(APPEND GGML_KLEIDIAI_SOURCES ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod.c) + list(APPEND GGML_KLEIDIAI_SOURCES ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod.c) + list(APPEND GGML_KLEIDIAI_SOURCES ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod.c) + endif() + + if (NOT I8MM_ENABLED MATCHES -1) + list(APPEND GGML_KLEIDIAI_SOURCES ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm.c) + endif() + + if (NOT SME_ENABLED MATCHES -1) + list(APPEND GGML_KLEIDIAI_SOURCES ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa.c) + list(APPEND GGML_KLEIDIAI_SOURCES ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot.c) + set(PRIVATE_ARCH_FLAGS "${PRIVATE_ARCH_FLAGS}+sve+sve2") + endif() + + set_source_files_properties(${GGML_KLEIDIAI_SOURCES} PROPERTIES COMPILE_OPTIONS "${PRIVATE_ARCH_FLAGS}") + list(APPEND GGML_CPU_SOURCES ${GGML_KLEIDIAI_SOURCES}) + endif() + message(STATUS "Adding CPU backend variant ${GGML_CPU_NAME}: ${ARCH_FLAGS} ${ARCH_DEFINITIONS}") target_sources(${GGML_CPU_NAME} PRIVATE ${GGML_CPU_SOURCES}) target_compile_options(${GGML_CPU_NAME} PRIVATE ${ARCH_FLAGS}) diff --git a/ggml/src/ggml-cpu/amx/amx.cpp b/ggml/src/ggml-cpu/amx/amx.cpp index 5ec5263ce..0f067137d 100644 --- a/ggml/src/ggml-cpu/amx/amx.cpp +++ b/ggml/src/ggml-cpu/amx/amx.cpp @@ -50,10 +50,11 @@ static void * ggml_backend_amx_buffer_get_base(ggml_backend_buffer_t buffer) { return (void *) (buffer->context); } -static void ggml_backend_amx_buffer_init_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) { +static enum ggml_status ggml_backend_amx_buffer_init_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) { tensor->extra = (void *) ggml::cpu::amx::get_tensor_traits(buffer, tensor); GGML_UNUSED(buffer); + return GGML_STATUS_SUCCESS; } static void ggml_backend_amx_buffer_memset_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, diff --git a/ggml/src/ggml-cpu/binary-ops.cpp b/ggml/src/ggml-cpu/binary-ops.cpp new file mode 100644 index 000000000..14f5b43ae --- /dev/null +++ b/ggml/src/ggml-cpu/binary-ops.cpp @@ -0,0 +1,158 @@ +#include "binary-ops.h" + +#if defined(GGML_USE_ACCELERATE) +#include + +using vDSP_fn_t = void (*)(const float *, vDSP_Stride, const float *, vDSP_Stride, float *, vDSP_Stride, vDSP_Length); +#endif + +static inline float op_add(float a, float b) { + return a + b; +} + +static inline float op_sub(float a, float b) { + return a - b; +} + +static inline float op_mul(float a, float b) { + return a * b; +} + +static inline float op_div(float a, float b) { + return a / b; +} + +template +static inline void vec_binary_op_contiguous(const int64_t n, dst_t * z, const src0_t * x, const src1_t * y) { + constexpr auto src0_to_f32 = type_conversion_table::to_f32; + constexpr auto src1_to_f32 = type_conversion_table::to_f32; + constexpr auto f32_to_dst = type_conversion_table::from_f32; + + for (int i = 0; i < n; i++) { + z[i] = f32_to_dst(op(src0_to_f32(x[i]), src1_to_f32(y[i]))); + } +} + +template +static inline void vec_binary_op_non_contiguous(const int64_t n, const int64_t ne10, const int64_t nb10, dst_t * z, const src0_t * x, const src1_t * y) { + constexpr auto src0_to_f32 = type_conversion_table::to_f32; + constexpr auto src1_to_f32 = type_conversion_table::to_f32; + constexpr auto f32_to_dst = type_conversion_table::from_f32; + + for (int i = 0; i < n; i++) { + int i10 = i % ne10; + const src1_t * y_ptr = (const src1_t *)((const char *)y + i10*nb10); + z[i] = f32_to_dst(op(src0_to_f32(x[i]), src1_to_f32(*y_ptr))); + } +} + +template +static void apply_binary_op(const ggml_compute_params * params, ggml_tensor * dst) { + const ggml_tensor * src0 = dst->src[0]; + const ggml_tensor * src1 = dst->src[1]; + + GGML_ASSERT(ggml_can_repeat(src1, src0) && ggml_are_same_shape(src0, dst)); + + GGML_TENSOR_BINARY_OP_LOCALS + + GGML_ASSERT( nb0 == sizeof(dst_t)); + GGML_ASSERT(nb00 == sizeof(src0_t)); + + const auto [ir0, ir1] = get_thread_range(params, src0); + const bool is_src1_contiguous = (nb10 == sizeof(src1_t)); + + if (!is_src1_contiguous) { // broadcast not implemented yet for non-contiguous + GGML_ASSERT(ggml_are_same_shape(src0, src1)); + } + +#ifdef GGML_USE_ACCELERATE + vDSP_fn_t vDSP_op = nullptr; + // TODO - avoid the f32-only check using type 'trait' lookup tables and row-based src-to-float conversion functions + if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) { + if (op == op_add) { + vDSP_op = vDSP_vadd; + } else if (op == op_sub) { + vDSP_op = vDSP_vsub; + } else if (op == op_mul) { + vDSP_op = vDSP_vmul; + } else if (op == op_div) { + vDSP_op = vDSP_vdiv; + } + } +#endif + + for (int64_t ir = ir0; ir < ir1; ++ir) { + const int64_t i03 = ir/(ne02*ne01); + const int64_t i02 = (ir - i03*ne02*ne01)/ne01; + const int64_t i01 = (ir - i03*ne02*ne01 - i02*ne01); + + const int64_t i13 = i03 % ne13; + const int64_t i12 = i02 % ne12; + const int64_t i11 = i01 % ne11; + + dst_t * dst_ptr = (dst_t *) ((char *) dst->data + i03*nb3 + i02*nb2 + i01*nb1 ); + const src0_t * src0_ptr = (const src0_t *) ((const char *) src0->data + i03*nb03 + i02*nb02 + i01*nb01); + const src1_t * src1_ptr = (const src1_t *) ((const char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11); + + if (is_src1_contiguous) { + // src1 is broadcastable across src0 and dst in i1, i2, i3 + const int64_t nr0 = ne00 / ne10; + + for (int64_t r = 0; r < nr0; ++r) { +#ifdef GGML_USE_ACCELERATE + if constexpr (std::is_same_v && std::is_same_v && std::is_same_v) { + if (vDSP_op != nullptr) { + vDSP_op(src1_ptr, 1, src0_ptr + r*ne10, 1, dst_ptr + r*ne10, 1, ne10); + continue; + } + } +#endif + vec_binary_op_contiguous(ne10, dst_ptr + r*ne10, src0_ptr + r*ne10, src1_ptr); + } + } else { + vec_binary_op_non_contiguous(ne0, ne10, nb10, dst_ptr, src0_ptr, src1_ptr); + } + } +} + +// TODO: Use the 'traits' lookup table (for type conversion fns), instead of a mass of 'if' conditions with long templates +template +static void binary_op(const ggml_compute_params * params, ggml_tensor * dst) { + const ggml_tensor * src0 = dst->src[0]; + const ggml_tensor * src1 = dst->src[1]; + + /* */ if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) { // all f32 + apply_binary_op(params, dst); + } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F16) { // all f16 + apply_binary_op(params, dst); + } else if (src0->type == GGML_TYPE_BF16 && src1->type == GGML_TYPE_BF16 && dst->type == GGML_TYPE_BF16) { // all bf16 + apply_binary_op(params, dst); + } else if (src0->type == GGML_TYPE_BF16 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_BF16) { + apply_binary_op(params, dst); + } else if (src0->type == GGML_TYPE_BF16 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) { + apply_binary_op(params, dst); + } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F16) { + apply_binary_op(params, dst); + } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) { + apply_binary_op(params, dst); + } else { + GGML_ABORT("%s: unsupported types: dst: %s, src0: %s, src1: %s\n", __func__, + ggml_type_name(dst->type), ggml_type_name(src0->type), ggml_type_name(src1->type)); + } +} + +void ggml_compute_forward_add_non_quantized(const ggml_compute_params * params, ggml_tensor * dst) { + binary_op(params, dst); +} + +void ggml_compute_forward_sub(const ggml_compute_params * params, ggml_tensor * dst) { + binary_op(params, dst); +} + +void ggml_compute_forward_mul(const ggml_compute_params * params, ggml_tensor * dst) { + binary_op(params, dst); +} + +void ggml_compute_forward_div(const ggml_compute_params * params, ggml_tensor * dst) { + binary_op(params, dst); +} diff --git a/ggml/src/ggml-cpu/binary-ops.h b/ggml/src/ggml-cpu/binary-ops.h new file mode 100644 index 000000000..aca1d89be --- /dev/null +++ b/ggml/src/ggml-cpu/binary-ops.h @@ -0,0 +1,16 @@ +#pragma once + +#include "common.h" + +#ifdef __cplusplus +extern "C" { +#endif + +void ggml_compute_forward_add_non_quantized(const struct ggml_compute_params * params, struct ggml_tensor * dst); +void ggml_compute_forward_sub(const struct ggml_compute_params * params, struct ggml_tensor * dst); +void ggml_compute_forward_mul(const struct ggml_compute_params * params, struct ggml_tensor * dst); +void ggml_compute_forward_div(const struct ggml_compute_params * params, struct ggml_tensor * dst); + +#ifdef __cplusplus +} +#endif diff --git a/ggml/src/ggml-cpu/common.h b/ggml/src/ggml-cpu/common.h new file mode 100644 index 000000000..3df01c1ed --- /dev/null +++ b/ggml/src/ggml-cpu/common.h @@ -0,0 +1,72 @@ +#pragma once + +#include "ggml.h" +#include "ggml-cpu-traits.h" +#include "ggml-cpu-impl.h" +#include "ggml-impl.h" + +#ifdef __cplusplus + +#include + +// convenience functions/macros for use in template calls +// note: these won't be required after the 'traits' lookup table is used. +static inline ggml_fp16_t f32_to_f16(float x) { + return GGML_FP32_TO_FP16(x); +} + +static inline float f16_to_f32(ggml_fp16_t x) { + return GGML_FP16_TO_FP32(x); +} + +static inline ggml_bf16_t f32_to_bf16(float x) { + return GGML_FP32_TO_BF16(x); +} + +static inline float bf16_to_f32(ggml_bf16_t x) { + return GGML_BF16_TO_FP32(x); +} + +static inline float f32_to_f32(float x) { + return x; +} + +// TODO - merge this into the traits table, after using row-based conversions +template +struct type_conversion_table; + +template <> +struct type_conversion_table { + static constexpr float (*to_f32)(ggml_fp16_t) = f16_to_f32; + static constexpr ggml_fp16_t (*from_f32)(float) = f32_to_f16; +}; + +template <> +struct type_conversion_table { + static constexpr float (*to_f32)(float) = f32_to_f32; + static constexpr float (*from_f32)(float) = f32_to_f32; +}; + +template <> +struct type_conversion_table { + static constexpr float (*to_f32)(ggml_bf16_t) = bf16_to_f32; + static constexpr ggml_bf16_t (*from_f32)(float) = f32_to_bf16; +}; + +static std::pair get_thread_range(const struct ggml_compute_params * params, const struct ggml_tensor * src0) { + const int64_t ith = params->ith; + const int64_t nth = params->nth; + + const int64_t nr = ggml_nrows(src0); + + // rows per thread + const int64_t dr = (nr + nth - 1)/nth; + + // row range for this thread + const int64_t ir0 = dr*ith; + const int64_t ir1 = MIN(ir0 + dr, nr); + + return {ir0, ir1}; +} + +#endif diff --git a/ggml/src/ggml-cpu/cpu-feats-x86.cpp b/ggml/src/ggml-cpu/cpu-feats-x86.cpp index e8133d411..902ee4346 100644 --- a/ggml/src/ggml-cpu/cpu-feats-x86.cpp +++ b/ggml/src/ggml-cpu/cpu-feats-x86.cpp @@ -278,6 +278,10 @@ static int ggml_backend_cpu_x86_score() { if (!is.SSE42()) { return 0; } score += 1<<2; #endif +#ifdef GGML_BMI2 + if (!is.BMI2()) { return 0; } + score += 1<<3; +#endif #ifdef GGML_AVX if (!is.AVX()) { return 0; } score += 1<<4; diff --git a/ggml/src/ggml-cpu/ggml-cpu-aarch64.cpp b/ggml/src/ggml-cpu/ggml-cpu-aarch64.cpp index b311a5b1c..74a31abb2 100644 --- a/ggml/src/ggml-cpu/ggml-cpu-aarch64.cpp +++ b/ggml/src/ggml-cpu/ggml-cpu-aarch64.cpp @@ -45,6 +45,24 @@ using block_q4_0x8 = block<4, 8>; using block_q8_0x4 = block<8, 4>; using block_q8_0x8 = block<8, 8>; + +struct block_q4_Kx8 { + ggml_half d[8]; // super-block scale for quantized scales + ggml_half dmin[8]; // super-block scale for quantized mins + uint8_t scales[96]; // scales and mins, quantized with 6 bits + uint8_t qs[1024]; // 4--bit quants +}; + +static_assert(sizeof(block_q4_Kx8) == sizeof(ggml_half) * 16 + K_SCALE_SIZE * 8 + QK_K * 4, "wrong q4_K block size/padding"); + +struct block_q8_Kx4 { + float d[4]; // delta + int8_t qs[QK_K * 4]; // quants + int16_t bsums[QK_K / 4]; // sum of quants in groups of 16 +}; + +static_assert(sizeof(block_q8_Kx4) == sizeof(float) * 4 + QK_K * 4 + (QK_K / 4) * sizeof(int16_t), "wrong q8_K block size/padding"); + struct block_iq4_nlx4 { ggml_half d[4]; // deltas for 4 iq4_nl blocks uint8_t qs[QK4_NL * 2]; // nibbles / quants for 4 iq4_nl blocks @@ -60,6 +78,13 @@ static_assert(sizeof(block_iq4_nlx4) == 4 * sizeof(ggml_half) + QK4_NL * 2, "wro #define UNUSED GGML_UNUSED +static inline int nearest_int(float fval) { + assert(fabsf(fval) <= 4194303.f); + float val = fval + 12582912.f; + int i; memcpy(&i, &val, sizeof(int)); + return (i & 0x007fffff) - 0x00400000; +} + // Functions to create the interleaved data layout formats // interleave 4 block_q4_0s in blocks of blck_size_interleave @@ -225,7 +250,7 @@ static inline __m256i mul_sum_i8_pairs_int32x8(const __m256i x, const __m256i y) static const int8_t kvalues_iq4nl[16] = {-127, -104, -83, -65, -49, -35, -22, -10, 1, 13, 25, 38, 53, 69, 89, 113}; -static void quantize_q8_0_4x4(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) { +static void ggml_quantize_mat_q8_0_4x4(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) { assert(QK8_0 == 32); assert(k % QK8_0 == 0); const int nb = k / QK8_0; @@ -319,7 +344,7 @@ static void quantize_q8_0_4x4(const float * GGML_RESTRICT x, void * GGML_RESTRIC #endif } -static void quantize_q8_0_4x8(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) { +static void ggml_quantize_mat_q8_0_4x8(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) { assert(QK8_0 == 32); assert(k % QK8_0 == 0); const int nb = k / QK8_0; @@ -534,16 +559,289 @@ static void quantize_q8_0_4x8(const float * GGML_RESTRICT x, void * GGML_RESTRIC #endif } -static void quantize_mat_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t nrow, int64_t n_per_row, int64_t blck_size_interleave) { +static void ggml_quantize_mat_q8_K_4x8(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) { + assert(QK_K == 256); + assert(k % QK_K == 0); + const int nb = k / QK_K; + + block_q8_Kx4 * GGML_RESTRICT y = (block_q8_Kx4 *) vy; + +#if defined(__AVX2__) + float iscale[4]; + __m256 srcv[4][32]; + __m256 iscale_vec[4]; + + for (int i = 0; i < nb; i++) { + for (int row_iter = 0; row_iter < 4; row_iter++) { + // Load elements into 4 AVX vectors + __m256 v0 = _mm256_loadu_ps( x + row_iter * k + i * 256 ); + __m256 v1 = _mm256_loadu_ps( x + row_iter * k + i * 256 + 8 ); + __m256 v2 = _mm256_loadu_ps( x + row_iter * k + i * 256 + 16 ); + __m256 v3 = _mm256_loadu_ps( x + row_iter * k + i * 256 + 24 ); + + // Compute max(abs(e)) for the block + const __m256 signBit = _mm256_set1_ps( -0.0f ); + __m256 abs0 = _mm256_andnot_ps( signBit, v0 ); + __m256 abs1 = _mm256_andnot_ps( signBit, v1 ); + __m256 abs2 = _mm256_andnot_ps( signBit, v2 ); + __m256 abs3 = _mm256_andnot_ps( signBit, v3 ); + + __m256 maxAbs = _mm256_max_ps( abs0, abs1 ); + maxAbs = _mm256_max_ps( maxAbs, abs2 ); + maxAbs = _mm256_max_ps( maxAbs, abs3 ); + + __m256 mask0 = _mm256_cmp_ps( maxAbs, v0, _CMP_EQ_OQ ); + __m256 mask1 = _mm256_cmp_ps( maxAbs, v1, _CMP_EQ_OQ ); + __m256 mask2 = _mm256_cmp_ps( maxAbs, v2, _CMP_EQ_OQ ); + __m256 mask3 = _mm256_cmp_ps( maxAbs, v3, _CMP_EQ_OQ ); + + __m256 maskAbs = _mm256_or_ps(_mm256_or_ps(mask0, mask1),_mm256_or_ps(mask2, mask3)); + + srcv[row_iter][0] = v0; + srcv[row_iter][1] = v1; + srcv[row_iter][2] = v2; + srcv[row_iter][3] = v3; + + for (int sb = 1; sb < 8; sb++) { + // Temporarily stores absolute quant values + __m256 tempAbs = maxAbs; + + // Load elements into 4 AVX vectors + __m256 v0 = _mm256_loadu_ps( x + row_iter * k + i * 256 + sb * 32); + __m256 v1 = _mm256_loadu_ps( x + row_iter * k + i * 256 + sb * 32 + 8 ); + __m256 v2 = _mm256_loadu_ps( x + row_iter * k + i * 256 + sb * 32 + 16 ); + __m256 v3 = _mm256_loadu_ps( x + row_iter * k + i * 256 + sb * 32 + 24 ); + + // Compute max(abs(e)) for the block + __m256 abs0 = _mm256_andnot_ps( signBit, v0 ); + __m256 abs1 = _mm256_andnot_ps( signBit, v1 ); + __m256 abs2 = _mm256_andnot_ps( signBit, v2 ); + __m256 abs3 = _mm256_andnot_ps( signBit, v3 ); + + maxAbs = _mm256_max_ps( maxAbs, abs0 ); + maxAbs = _mm256_max_ps( maxAbs, abs1 ); + maxAbs = _mm256_max_ps( maxAbs, abs2 ); + maxAbs = _mm256_max_ps( maxAbs, abs3 ); + + __m256 mask_prev = _mm256_cmp_ps( tempAbs, maxAbs, _CMP_EQ_OQ ); + maskAbs = _mm256_and_ps( maskAbs, mask_prev ); + + mask0 = _mm256_cmp_ps( maxAbs, v0, _CMP_EQ_OQ ); + mask1 = _mm256_cmp_ps( maxAbs, v1, _CMP_EQ_OQ ); + mask2 = _mm256_cmp_ps( maxAbs, v2, _CMP_EQ_OQ ); + mask3 = _mm256_cmp_ps( maxAbs, v3, _CMP_EQ_OQ ); + + __m256 mask_curr = _mm256_or_ps(_mm256_or_ps(mask0, mask1),_mm256_or_ps(mask2, mask3)); + maskAbs = _mm256_or_ps(maskAbs, mask_curr); + + srcv[row_iter][sb * 4] = v0; + srcv[row_iter][sb * 4 + 1] = v1; + srcv[row_iter][sb * 4 + 2] = v2; + srcv[row_iter][sb * 4 + 3] = v3; + } + + __m128 max4 = _mm_max_ps( _mm256_extractf128_ps( maxAbs, 1 ), _mm256_castps256_ps128( maxAbs ) ); + max4 = _mm_max_ps( max4, _mm_movehl_ps( max4, max4 ) ); + max4 = _mm_max_ss( max4, _mm_movehdup_ps( max4 ) ); + const float maxScalar = _mm_cvtss_f32( max4 ); + + __m256 maxScalarVec = _mm256_set1_ps(maxScalar); + + __m256 mask_next = _mm256_cmp_ps( maxScalarVec, maxAbs, _CMP_EQ_OQ ); + __m256 finalMask = _mm256_and_ps(maskAbs, mask_next); + + const int mask = _mm256_movemask_ps(finalMask); + iscale[row_iter] = ( maxScalar != 0.0f ) ? 127.f / maxScalar : 0.0f; + + if(mask) { + iscale[row_iter] = ( maxScalar != 0.0f ) ? -127.f / maxScalar: 0.0f; + } + + y[i].d[row_iter] = maxScalar ? 1/iscale[row_iter] : 0; + iscale_vec[row_iter] = _mm256_set1_ps(iscale[row_iter]); + } + + __m256i quants_interleaved[32]; + for (int j = 0; j < 32; j++) { + // Apply the multiplier + __m256 v0 = _mm256_mul_ps(srcv[0][j], iscale_vec[0]); + __m256 v1 = _mm256_mul_ps(srcv[1][j], iscale_vec[1]); + __m256 v2 = _mm256_mul_ps(srcv[2][j], iscale_vec[2]); + __m256 v3 = _mm256_mul_ps(srcv[3][j], iscale_vec[3]); + + // Round to nearest integer + v0 = _mm256_round_ps( v0, _MM_ROUND_NEAREST ); + v1 = _mm256_round_ps( v1, _MM_ROUND_NEAREST ); + v2 = _mm256_round_ps( v2, _MM_ROUND_NEAREST ); + v3 = _mm256_round_ps( v3, _MM_ROUND_NEAREST ); + + // Convert floats to integers + __m256i i0 = _mm256_cvtps_epi32( v0 ); + __m256i i1 = _mm256_cvtps_epi32( v1 ); + __m256i i2 = _mm256_cvtps_epi32( v2 ); + __m256i i3 = _mm256_cvtps_epi32( v3 ); + + // Convert int32 to int16 + i0 = _mm256_packs_epi32( i0, i1 ); + i2 = _mm256_packs_epi32( i2, i3 ); + // Convert int16 to int8 + i0 = _mm256_packs_epi16( i0, i2 ); + + // Permute and store the quantized weights in the required order after the pack instruction + const __m256i perm = _mm256_setr_epi32( 0, 4, 1, 5, 2, 6, 3, 7 ); + i0 = _mm256_permutevar8x32_epi32( i0, perm ); + + _mm256_storeu_si256((__m256i *)(y[i].qs + 32 * j), i0); + quants_interleaved[j] = i0; + } + + // Masks to shuffle the quants of corresonding sub blocks for rearraning quants for vectorized bsums computation + __m256i shuffle_mask_sb2 = _mm256_castsi128_si256(_mm_setr_epi8(0, 1, 0, 1, 4, 5, 6, 7, 8, 9, 8, 9, 12, 13, 14, 15)); + shuffle_mask_sb2 = _mm256_permute2f128_si256(shuffle_mask_sb2, shuffle_mask_sb2, 0); + __m256i shuffle_mask_sb3 = _mm256_castsi128_si256(_mm_setr_epi8(0, 1, 2, 3, 0, 1, 6, 7, 8, 9, 10, 11, 8, 9, 14, 15)); + shuffle_mask_sb3 = _mm256_permute2f128_si256(shuffle_mask_sb3, shuffle_mask_sb3, 0); + __m256i shuffle_mask_sb4 = _mm256_castsi128_si256(_mm_setr_epi8(0, 1, 2, 3, 4, 5, 0, 1, 8, 9, 10, 11, 12, 13, 8, 9)); + shuffle_mask_sb4 = _mm256_permute2f128_si256(shuffle_mask_sb4, shuffle_mask_sb4, 0); + + for (int k = 0; k < 4; k++) { + // Quants from four different sub blocks are taken + __m256i q0 = quants_interleaved[k * 8 + 0]; + __m256i q1 = quants_interleaved[k * 8 + 1]; + __m256i q2 = quants_interleaved[k * 8 + 2]; + __m256i q3 = quants_interleaved[k * 8 + 3]; + __m256i q4 = quants_interleaved[k * 8 + 4]; + __m256i q5 = quants_interleaved[k * 8 + 5]; + __m256i q6 = quants_interleaved[k * 8 + 6]; + __m256i q7 = quants_interleaved[k * 8 + 7]; + + + // The below code block has the first half of different sub blocks shuffled and blended so as to process 2 values from each sub block at a time + __m256i sb2_h1_shuffled = _mm256_shuffle_epi8(q2, shuffle_mask_sb2); + __m256i sb_h1_interleaved = _mm256_blend_epi16(q0, sb2_h1_shuffled, 34); + __m256i sb3_h1_shuffled = _mm256_shuffle_epi8(q4, shuffle_mask_sb3); + sb_h1_interleaved = _mm256_blend_epi16(sb_h1_interleaved, sb3_h1_shuffled, 68); + __m256i sb4_h1_shuffled = _mm256_shuffle_epi8(q6, shuffle_mask_sb4); + sb_h1_interleaved = _mm256_blend_epi16(sb_h1_interleaved, sb4_h1_shuffled, 136); + + __m256i one = _mm256_set1_epi8(1); + __m256i bsums_r1 = _mm256_maddubs_epi16(one, sb_h1_interleaved); + + for (int l = 0; l < 3; l++) { + // Quants value shifted to process next two values from each sub block + q0 = _mm256_srli_epi64(q0, 16); + q2 = _mm256_srli_epi64(q2, 16); + q4 = _mm256_srli_epi64(q4, 16); + q6 = _mm256_srli_epi64(q6, 16); + + sb2_h1_shuffled = _mm256_shuffle_epi8(q2, shuffle_mask_sb2); + sb_h1_interleaved = _mm256_blend_epi16(q0, sb2_h1_shuffled, 34); + sb3_h1_shuffled = _mm256_shuffle_epi8(q4, shuffle_mask_sb3); + sb_h1_interleaved = _mm256_blend_epi16(sb_h1_interleaved, sb3_h1_shuffled, 68); + sb4_h1_shuffled = _mm256_shuffle_epi8(q6, shuffle_mask_sb4); + sb_h1_interleaved = _mm256_blend_epi16(sb_h1_interleaved, sb4_h1_shuffled, 136); + + bsums_r1 = _mm256_add_epi16(bsums_r1, _mm256_maddubs_epi16(one, sb_h1_interleaved)); + } + + // The below code block has the second half of different sub blocks shuffled and blended so as to process 2 values from each sub block at a time + __m256i sb2_h2_shuffled = _mm256_shuffle_epi8(q3, shuffle_mask_sb2); + __m256i sb_h2_interleaved = _mm256_blend_epi16(q1, sb2_h2_shuffled, 34); + __m256i sb3_h2_shuffled = _mm256_shuffle_epi8(q5, shuffle_mask_sb3); + sb_h2_interleaved = _mm256_blend_epi16(sb_h2_interleaved, sb3_h2_shuffled, 68); + __m256i sb4_h2_shuffled = _mm256_shuffle_epi8(q7, shuffle_mask_sb4); + sb_h2_interleaved = _mm256_blend_epi16(sb_h2_interleaved, sb4_h2_shuffled, 136); + + __m256i bsums_r2 = _mm256_maddubs_epi16(one, sb_h2_interleaved); + + for (int l = 0; l < 3; l++) { + // Quants value shifted to process next two values from each sub block + q1 = _mm256_srli_epi64(q1, 16); + q3 = _mm256_srli_epi64(q3, 16); + q5 = _mm256_srli_epi64(q5, 16); + q7 = _mm256_srli_epi64(q7, 16); + + sb2_h2_shuffled = _mm256_shuffle_epi8(q3, shuffle_mask_sb2); + sb_h2_interleaved = _mm256_blend_epi16(q1, sb2_h2_shuffled, 34); + sb3_h2_shuffled = _mm256_shuffle_epi8(q5, shuffle_mask_sb3); + sb_h2_interleaved = _mm256_blend_epi16(sb_h2_interleaved, sb3_h2_shuffled, 68); + sb4_h2_shuffled = _mm256_shuffle_epi8(q7, shuffle_mask_sb4); + sb_h2_interleaved = _mm256_blend_epi16(sb_h2_interleaved, sb4_h2_shuffled, 136); + + bsums_r2 = _mm256_add_epi16(bsums_r2, _mm256_maddubs_epi16(one, sb_h2_interleaved)); + } + + // Overall bsums in interleaved fashion computed by adding results of both halves + __m256i bsums_r = _mm256_add_epi16(bsums_r1, bsums_r2); + _mm256_storeu_si256((__m256i *)(y[i].bsums + 16 * k), bsums_r); + } + } + +#else + + // scalar + const int blck_size_interleave = 8; + float srcv[4][QK_K]; + float iscale[4]; + + for (int i = 0; i < nb; i++) { + for (int row_iter = 0; row_iter < 4; row_iter++) { + float amax = 0.0f; // absolute max + float max = 0; + + for (int j = 0; j < QK_K; j++) { + srcv[row_iter][j] = x[row_iter * k + i * QK_K + j]; + // Update the maximum value of the corresponding super block + if(amax < fabsf(srcv[row_iter][j])) { + amax = fabsf(srcv[row_iter][j]); + max = srcv[row_iter][j]; + } + } + + iscale[row_iter] = amax ? -127.f/max : 0; + + y[i].d[row_iter] = amax ? 1/iscale[row_iter] : 0; + } + + for (int j = 0; j < QK_K / 4; j++) { + y[i].bsums[j] = 0; + } + + // Quants values are interleaved in sequence of eight bytes from corresponding super blocks + // Bsums values are interleaved in sequence of four bsums from each super block taken for interleaving + // i.e first four bsums from the first super block, followed by first four bsums from second super block and so on + for (int j = 0; j < QK_K * 4; j++) { + int src_offset = (j / (4 * blck_size_interleave)) * blck_size_interleave; + int src_id = (j % (4 * blck_size_interleave)) / blck_size_interleave; + src_offset += (j % blck_size_interleave); + int index = (((j & 31) >> 3) << 2) + ((j >> 8) << 4) + ((j >> 6) & 3); + + float x0 = srcv[src_id][src_offset] * iscale[src_id]; + y[i].qs[j] = nearest_int(x0); + y[i].bsums[index] += y[i].qs[j]; + } + } +#endif +} + +template +void ggml_quantize_mat_t(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t nrow, int64_t n_per_row); + +template <> void ggml_quantize_mat_t<4, GGML_TYPE_Q8_0>(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t nrow, int64_t n_per_row) { assert(nrow == 4); UNUSED(nrow); - if (blck_size_interleave == 4) { - quantize_q8_0_4x4(x, vy, n_per_row); - } else if (blck_size_interleave == 8) { - quantize_q8_0_4x8(x, vy, n_per_row); - } else { - assert(false); - } + ggml_quantize_mat_q8_0_4x4(x, vy, n_per_row); +} + +template <> void ggml_quantize_mat_t<8, GGML_TYPE_Q8_0>(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t nrow, int64_t n_per_row) { + assert(nrow == 4); + UNUSED(nrow); + ggml_quantize_mat_q8_0_4x8(x, vy, n_per_row); +} + +template <> void ggml_quantize_mat_t<8, GGML_TYPE_Q8_K>(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t nrow, int64_t n_per_row) { + assert(nrow == 4); + UNUSED(nrow); + ggml_quantize_mat_q8_K_4x8(x, vy, n_per_row); } static void ggml_gemv_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) { @@ -994,6 +1292,281 @@ static void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, c } } +static void ggml_gemv_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) { + const int qk = QK_K; + const int nb = n / qk; + const int ncols_interleaved = 8; + const int blocklen = 8; + static const uint32_t kmask1 = 0x3f3f3f3f; + static const uint32_t kmask2 = 0x0f0f0f0f; + static const uint32_t kmask3 = 0x03030303; + + assert (n % qk == 0); + assert (nc % ncols_interleaved == 0); + + UNUSED(s); + UNUSED(bs); + UNUSED(vx); + UNUSED(vy); + UNUSED(nr); + UNUSED(nc); + UNUSED(nb); + UNUSED(ncols_interleaved); + UNUSED(blocklen); + +#if defined(__AVX2__) + // Lookup table to convert signed nibbles to signed bytes + __m256i signextendlut = _mm256_castsi128_si256(_mm_set_epi8(-1, -2, -3, -4, -5, -6, -7, -8, 7, 6, 5, 4, 3, 2, 1, 0)); + signextendlut = _mm256_permute2f128_si256(signextendlut, signextendlut, 0); + // Shuffle masks to rearrange delta and scale values to multiply with appropriate scales + __m128i deltamask = _mm_set_epi8(15, 14, 7, 6, 13, 12, 5, 4, 11, 10, 3, 2, 9, 8, 1, 0); + __m128i scalemask = _mm_set_epi8(7, 7, 3, 3, 6, 6, 2, 2, 5, 5, 1, 1, 4, 4, 0, 0); + // Permute mask used for easier vector processing at later stages + __m256i finalpermutemask = _mm256_set_epi32(7, 5, 3, 1, 6, 4, 2, 0); + + // Mask to extract nibbles from bytes + const __m256i m4b = _mm256_set1_epi8(0x0F); + + int64_t b_nb = n / QK_K; + + const block_q4_Kx8 * b_ptr_start = (const block_q4_Kx8 *)vx; + const block_q8_K * a_ptr_start = (const block_q8_K *)vy; + + // Process Q8_K blocks one by one + for (int64_t y = 0; y < nr; y++) { + + // Pointers to LHS blocks of block_q8_K format + const block_q8_K * a_ptr = a_ptr_start + (y * nb); + + // Take group of eight interleaved block_q4_K structures at each pass of the loop and perform dot product operation + for (int64_t x = 0; x < nc / 8; x++) { + + // Pointers to RHS blocks + const block_q4_Kx8 * b_ptr = b_ptr_start + (x * b_nb); + + // Master FP accumulators + __m256 acc_row = _mm256_setzero_ps(); + __m256 acc_min_rows = _mm256_setzero_ps(); + + for (int64_t b = 0; b < nb; b++) { + + // Load and convert to FP32 scale from block_q8_K + const __m256 row_scale_f32 = _mm256_set1_ps((a_ptr[b].d)); + + // Load the scale values for the 8 blocks interleaved in block_q4_Kx8 + // col_scale_f32 rearranged so as to multiply with appropriate quants + const __m256 col_scale_f32 = GGML_F32Cx8_REARRANGE_LOAD(b_ptr[b].d, deltamask); + const __m256 col_dmin_f32 = GGML_F32Cx8_LOAD(b_ptr[b].dmin); + + __m256i iacc_b = _mm256_setzero_si256(); + __m256i iacc_min_b = _mm256_setzero_si256(); + + const __m256i q8sums = _mm256_loadu_si256((const __m256i * )(a_ptr[b].bsums)); + __m256i q8s = _mm256_castsi128_si256(_mm_hadd_epi16(_mm256_castsi256_si128(q8sums), _mm256_extracti128_si256(q8sums, 1))); + q8s = _mm256_permute2f128_si256(q8s, q8s, 0); + + // Processes two sub blocks from each Q4_K in each iteration + for (int sb = 0; sb < QK_K / 64; sb++) { + + // Load the eight block_q4_K for two sub blocks quantized values interleaved with each other in chunks of eight - B0,B1 ....B6,B7 + const __m256i rhs_raw_vec_0123_0 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + sb * 256)); + const __m256i rhs_raw_vec_4567_0 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 32 + sb * 256)); + const __m256i rhs_raw_vec_0123_1 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 64 + sb * 256)); + const __m256i rhs_raw_vec_4567_1 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 96 + sb * 256)); + const __m256i rhs_raw_vec_0123_2 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 128 + sb * 256)); + const __m256i rhs_raw_vec_4567_2 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 160 + sb * 256)); + const __m256i rhs_raw_vec_0123_3 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 192 + sb * 256)); + const __m256i rhs_raw_vec_4567_3 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 224 + sb * 256)); + + // 4-bit -> 8-bit + // Values of the first sub block of eight block_q4_K structures for the sb loop + const __m256i rhs_vec_0123_00 = _mm256_and_si256(rhs_raw_vec_0123_0, m4b); + const __m256i rhs_vec_4567_00 = _mm256_and_si256(rhs_raw_vec_4567_0, m4b); + const __m256i rhs_vec_0123_01 = _mm256_and_si256(rhs_raw_vec_0123_1, m4b); + const __m256i rhs_vec_4567_01 = _mm256_and_si256(rhs_raw_vec_4567_1, m4b); + const __m256i rhs_vec_0123_02 = _mm256_and_si256(rhs_raw_vec_0123_2, m4b); + const __m256i rhs_vec_4567_02 = _mm256_and_si256(rhs_raw_vec_4567_2, m4b); + const __m256i rhs_vec_0123_03 = _mm256_and_si256(rhs_raw_vec_0123_3, m4b); + const __m256i rhs_vec_4567_03 = _mm256_and_si256(rhs_raw_vec_4567_3, m4b); + + // Values of the second sub block of eight block_q4_K structures when sb = 1 + const __m256i rhs_vec_0123_10 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_0, 4), m4b); + const __m256i rhs_vec_4567_10 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_0, 4), m4b); + const __m256i rhs_vec_0123_11 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_1, 4), m4b); + const __m256i rhs_vec_4567_11 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_1, 4), m4b); + const __m256i rhs_vec_0123_12 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_2, 4), m4b); + const __m256i rhs_vec_4567_12 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_2, 4), m4b); + const __m256i rhs_vec_0123_13 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_3, 4), m4b); + const __m256i rhs_vec_4567_13 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_3, 4), m4b); + + uint32_t utmp_0[4], utmp_1[4]; + + // Scales and Mins of corresponding sub blocks from different Q8_K structures are stored together + // The below block is for eg to extract first sub block's scales and mins from different Q4_K structures for the sb loop + memcpy(utmp_0, b_ptr[b].scales + 24 * sb, 12); + utmp_0[3] = ((utmp_0[2] >> 4) & kmask2) | (((utmp_0[1] >> 6) & kmask3) << 4); + const uint32_t uaux_0 = utmp_0[1] & kmask1; + utmp_0[1] = (utmp_0[2] & kmask2) | (((utmp_0[0] >> 6) & kmask3) << 4); + utmp_0[2] = uaux_0; + utmp_0[0] &= kmask1; + + // The below block is for eg to extract second sub block's scales and mins from different Q4_K structures for the sb loop + memcpy(utmp_1, b_ptr[b].scales + 12 + sb * 24, 12); + utmp_1[3] = ((utmp_1[2] >> 4) & kmask2) | (((utmp_1[1] >> 6) & kmask3) << 4); + const uint32_t uaux_1 = utmp_1[1] & kmask1; + utmp_1[1] = (utmp_1[2] & kmask2) | (((utmp_1[0] >> 6) & kmask3) << 4); + utmp_1[2] = uaux_1; + utmp_1[0] &= kmask1; + + // Scales of first sub block in the sb loop + const __m128i mins_and_scales_0 = _mm_set_epi32(utmp_0[3], utmp_0[2], utmp_0[1], utmp_0[0]); + __m128i scales_rearrange_0 = _mm_shuffle_epi8(mins_and_scales_0, scalemask); + __m256i scales_0 = _mm256_cvtepu8_epi16(scales_rearrange_0); + + // Scales of second sub block in the sb loop + __m128i mins_and_scales_1 = _mm_set_epi32(utmp_1[3], utmp_1[2], utmp_1[1], utmp_1[0]); + __m128i scales_rearrange_1 = _mm_shuffle_epi8(mins_and_scales_1, scalemask); + __m256i scales_1 = _mm256_cvtepu8_epi16(scales_rearrange_1); + + // Mins of first and second sub block of Q4_K block are arranged side by side + __m256i mins_01 = _mm256_cvtepu8_epi16(_mm_unpacklo_epi8(_mm_shuffle_epi32(mins_and_scales_0, 78), _mm_shuffle_epi32(mins_and_scales_1, 78))); + + // Load the two sub block values corresponding to sb in block_q8_K in batches of 16 bytes and replicate the same across 256 bit vector + __m256i lhs_vec_00 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + sb * 64))); + __m256i lhs_vec_01 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + 16 + sb * 64))); + __m256i lhs_vec_10 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + 32 + sb * 64))); + __m256i lhs_vec_11 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + 48 + sb * 64))); + + lhs_vec_00 = _mm256_permute2f128_si256(lhs_vec_00, lhs_vec_00, 0); + lhs_vec_01 = _mm256_permute2f128_si256(lhs_vec_01, lhs_vec_01, 0); + lhs_vec_10 = _mm256_permute2f128_si256(lhs_vec_10, lhs_vec_10, 0); + lhs_vec_11 = _mm256_permute2f128_si256(lhs_vec_11, lhs_vec_11, 0); + + // Dot product done within 32 bit lanes and accumulated in the same vector + // First done for first sub block and thenn for second sub block in each sb + // B0(0-3) B4(0-3) B1(0-3) B5(0-3) B2(0-3) B6(0-3) B3(0-3) B7(0-3) with A0(0-3) + // B0(4-7) B4(4-7) B1(4-7) B5(4-7) B2(4-7) B6(4-7) B3(4-7) B7(4-7) with A0(4-7) + // ........................................................................... + // B0(28-31) B4(28-31) B1(28-31) B5(28-31) B2(28-31) B6(28-31) B3(28-31) B7(28-31) with A0(28-31) + + + __m256i iacc_0 = _mm256_setzero_si256(); + __m256i iacc_1 = _mm256_setzero_si256(); + + iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_00 ,_mm256_shuffle_epi32(rhs_vec_4567_00, 177), 170), _mm256_shuffle_epi32(lhs_vec_00, 0))); + iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_00, 177) ,rhs_vec_4567_00, 170), _mm256_shuffle_epi32(lhs_vec_00, 85))); + + iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_01 ,_mm256_shuffle_epi32(rhs_vec_4567_01, 177), 170), _mm256_shuffle_epi32(lhs_vec_00, 170))); + iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_01, 177) ,rhs_vec_4567_01, 170), _mm256_shuffle_epi32(lhs_vec_00, 255))); + + iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_02 ,_mm256_shuffle_epi32(rhs_vec_4567_02, 177), 170), _mm256_shuffle_epi32(lhs_vec_01, 0))); + iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_02, 177) ,rhs_vec_4567_02, 170), _mm256_shuffle_epi32(lhs_vec_01, 85))); + + iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_03 ,_mm256_shuffle_epi32(rhs_vec_4567_03, 177), 170), _mm256_shuffle_epi32(lhs_vec_01, 170))); + iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_03, 177) ,rhs_vec_4567_03, 170), _mm256_shuffle_epi32(lhs_vec_01, 255))); + + iacc_0 = _mm256_madd_epi16(iacc_0, scales_0); + + iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_10 ,_mm256_shuffle_epi32(rhs_vec_4567_10, 177), 170), _mm256_shuffle_epi32(lhs_vec_10, 0))); + iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_10, 177) ,rhs_vec_4567_10, 170), _mm256_shuffle_epi32(lhs_vec_10, 85))); + + iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_11 ,_mm256_shuffle_epi32(rhs_vec_4567_11, 177), 170), _mm256_shuffle_epi32(lhs_vec_10, 170))); + iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_11, 177) ,rhs_vec_4567_11, 170), _mm256_shuffle_epi32(lhs_vec_10, 255))); + + iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_12 ,_mm256_shuffle_epi32(rhs_vec_4567_12, 177), 170), _mm256_shuffle_epi32(lhs_vec_11, 0))); + iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_12, 177) ,rhs_vec_4567_12, 170), _mm256_shuffle_epi32(lhs_vec_11, 85))); + + iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_13 ,_mm256_shuffle_epi32(rhs_vec_4567_13, 177), 170), _mm256_shuffle_epi32(lhs_vec_11, 170))); + iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_13, 177) ,rhs_vec_4567_13, 170), _mm256_shuffle_epi32(lhs_vec_11, 255))); + + iacc_1 = _mm256_madd_epi16(iacc_1, scales_1); + + // Accumulate the iacc value for one sb + __m256i iacc_sb = _mm256_add_epi32(iacc_0, iacc_1); + + // Broadcast the bsums of the two sub blocks of the iteration of Q8_K across the vector + // Multiply-Add with corresponding mins of Q4_Kx8 with bsums + __m256i q8s_sb = _mm256_shuffle_epi32(q8s, 0); + __m256i iacc_min_sb = _mm256_madd_epi16(q8s_sb, mins_01); + q8s = _mm256_bsrli_epi128(q8s, 4); + + // Accumulate for the complete block + iacc_b = _mm256_add_epi32(iacc_b, iacc_sb); + iacc_min_b = _mm256_add_epi32(iacc_min_b, iacc_min_sb); + } + + // Multiply-Add with scale values for the complete super block + acc_row = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_b), _mm256_mul_ps(col_scale_f32, row_scale_f32), acc_row); + acc_min_rows = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_min_b), _mm256_mul_ps(col_dmin_f32, row_scale_f32), acc_min_rows); + + } + + // Accumulated output values permuted so as to be stored in appropriate order post accumulation + acc_row = _mm256_permutevar8x32_ps(acc_row, finalpermutemask); + _mm256_storeu_ps(s + (y * nr + x * 8), _mm256_sub_ps(acc_row, acc_min_rows)); + } + } + +#else + + float sumf[8]; + float sum_minf[8]; + uint32_t utmp[32]; + int sumi1; + int sumi2; + int sumi; + + const block_q8_K * a_ptr = (const block_q8_K *) vy; + for (int x = 0; x < nc / ncols_interleaved; x++) { + const block_q4_Kx8 * b_ptr = (const block_q4_Kx8 *) vx + (x * nb); + + for (int j = 0; j < ncols_interleaved; j++) { + sumf[j] = 0.0; + sum_minf[j] = 0.0; + } + for (int l = 0; l < nb; l++) { + for (int sb = 0; sb < 8; sb++) { + memcpy(utmp + sb * 4, b_ptr[l].scales + sb * 12, 12); + utmp[sb * 4 + 3] = ((utmp[sb * 4 + 2] >> 4) & kmask2) | (((utmp[sb * 4 + 1] >> 6) & kmask3) << 4); + const uint32_t uaux_0 = utmp[sb * 4 + 1] & kmask1; + utmp[sb * 4 + 1] = (utmp[sb * 4 + 2] & kmask2) | (((utmp[sb * 4 + 0] >> 6) & kmask3) << 4); + utmp[sb * 4 + 2] = uaux_0; + utmp[sb * 4 + 0] &= kmask1; + } + for (int k = 0; k < (qk / (2 * blocklen)); k++) { + uint8_t *scales_0 = (uint8_t*) utmp + (k / 4) * 32; + uint8_t *scales_1 = (uint8_t*) utmp + (k / 4) * 32 + 16; + for (int j = 0; j < ncols_interleaved; j++) { + sumi1 = 0; + sumi2 = 0; + sumi = 0; + for (int i = 0; i < blocklen; ++i) { + const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF); + const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4); + sumi1 = (v0 * a_ptr[l].qs[(k >> 2) * 64 + (k % 4) * blocklen + i]); + sumi2 = (v1 * a_ptr[l].qs[(k >> 2) * 64 + (k % 4) * blocklen + i + 32]); + sumi1 = sumi1 * scales_0[j]; + sumi2 = sumi2 * scales_1[j]; + sumi += sumi1 + sumi2; + } + sumf[j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d; + } + } + for (int sb = 0; sb < 8; sb++) { + uint8_t *mins = (uint8_t*) utmp + 8 + sb * 16; + for (int j = 0; j < ncols_interleaved; j++) { + sum_minf[j] += mins[j] * (a_ptr[l].bsums[sb * 2] + a_ptr[l].bsums[sb * 2 + 1]) * GGML_FP16_TO_FP32(b_ptr[l].dmin[j]) * a_ptr[l].d; + } + } + } + for (int j = 0; j < ncols_interleaved; j++) { + s[x * ncols_interleaved + j] = sumf[j] - sum_minf[j]; + } + } +#endif +} + + static void ggml_gemv_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) { const int qk = QK8_0; const int nb = n / qk; @@ -3480,6 +4053,781 @@ static void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, c } } +static void ggml_gemm_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) { + const int qk = QK_K; + const int nb = n / qk; + const int ncols_interleaved = 8; + const int blocklen = 8; + static const uint32_t kmask1 = 0x3f3f3f3f; + static const uint32_t kmask2 = 0x0f0f0f0f; + static const uint32_t kmask3 = 0x03030303; + + assert (n % qk == 0); + assert (nr % 4 == 0); + assert (nc % ncols_interleaved == 0); + + UNUSED(s); + UNUSED(bs); + UNUSED(vx); + UNUSED(vy); + UNUSED(nr); + UNUSED(nc); + UNUSED(nb); + UNUSED(ncols_interleaved); + UNUSED(blocklen); + +#if defined(__AVX2__) + const block_q4_Kx8 * b_ptr_start = (const block_q4_Kx8 * ) vx; + const block_q8_Kx4 * a_ptr_start = (const block_q8_Kx4 * ) vy; + int64_t b_nb = n / QK_K; + int64_t y = 0; + + // Mask to mask out nibbles from packed bytes + const __m256i m4b = _mm256_set1_epi8(0x0F); + // Permute mask used for easier vector processing at later stages + __m256i requiredOrder = _mm256_set_epi32(3, 2, 1, 0, 7, 6, 5, 4); + + int anr = nr - nr % 16;; // Used to align nr with boundary of 16 + // Take group of four block_q8_Kx4 structures at each pass of the loop and perform dot product operation + for (; y < anr / 4; y += 4) { + + const block_q8_Kx4 * a_ptrs[4]; + + a_ptrs[0] = a_ptr_start + (y * nb); + for (int i = 0; i < 3; ++i) { + a_ptrs[i + 1] = a_ptrs[i] + nb; + } + + // Take group of eight block_q4_kx8 structures at each pass of the loop and perform dot product operation + for (int64_t x = 0; x < nc / 8; x++) { + + const block_q4_Kx8 * b_ptr = b_ptr_start + (x * b_nb); + + // Master FP accumulators + __m256 acc_rows[16]; + for (int i = 0; i < 16; i++) { + acc_rows[i] = _mm256_setzero_ps(); + } + + __m256 acc_min_rows[16]; + for (int i = 0; i < 16; i++) { + acc_min_rows[i] = _mm256_setzero_ps(); + } + + // For super block + for (int64_t b = 0; b < nb; b++) { + + // Scale values - Load the eight scale values of block_q4_kx8 + const __m256 col_scale_f32 = GGML_F32Cx8_LOAD(b_ptr[b].d); + + // dmin values - Load the eight dmin values of block_q4_kx8 + const __m256 col_dmin_f32 = GGML_F32Cx8_LOAD(b_ptr[b].dmin); + + // Loop to iterate over the eight sub blocks of a super block - two sub blocks are processed per iteration + for (int sb = 0; sb < QK_K / 64; sb++) { + + // Load the eight block_q4_K for two sub blocks quantized values interleaved with each other in chunks of eight bytes - B0,B1 ....B6,B7 + const __m256i rhs_raw_mat_0123_0 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + sb * 256)); + const __m256i rhs_raw_mat_4567_0 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 32 + sb * 256)); + const __m256i rhs_raw_mat_0123_1 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 64 + sb * 256)); + const __m256i rhs_raw_mat_4567_1 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 96 + sb * 256)); + const __m256i rhs_raw_mat_0123_2 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 128 + sb * 256)); + const __m256i rhs_raw_mat_4567_2 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 160 + sb * 256)); + const __m256i rhs_raw_mat_0123_3 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 192 + sb * 256)); + const __m256i rhs_raw_mat_4567_3 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 224 + sb * 256)); + + // Save the values in the following vectors in the formats B0B1B4B5, B2B3B6B7 for further processing and storing of values + const __m256i rhs_raw_mat_0145_0 = _mm256_blend_epi32(rhs_raw_mat_0123_0, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_0, requiredOrder), 240); + const __m256i rhs_raw_mat_2367_0 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_0, requiredOrder), rhs_raw_mat_4567_0, 240); + const __m256i rhs_raw_mat_0145_1 = _mm256_blend_epi32(rhs_raw_mat_0123_1, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_1, requiredOrder), 240); + const __m256i rhs_raw_mat_2367_1 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_1, requiredOrder), rhs_raw_mat_4567_1, 240); + const __m256i rhs_raw_mat_0145_2 = _mm256_blend_epi32(rhs_raw_mat_0123_2, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_2, requiredOrder), 240); + const __m256i rhs_raw_mat_2367_2 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_2, requiredOrder), rhs_raw_mat_4567_2, 240); + const __m256i rhs_raw_mat_0145_3 = _mm256_blend_epi32(rhs_raw_mat_0123_3, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_3, requiredOrder), 240); + const __m256i rhs_raw_mat_2367_3 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_3, requiredOrder), rhs_raw_mat_4567_3, 240); + + // 4-bit -> 8-bit + // First sub block of the two sub blocks processed in the iteration + const __m256i rhs_mat_0145_00 = _mm256_and_si256(rhs_raw_mat_0145_0, m4b); //B00(0-7) B01(0-7) B04(0-7) B05(0-7) + const __m256i rhs_mat_2367_00 = _mm256_and_si256(rhs_raw_mat_2367_0, m4b); //B02(0-7) B03(0-7) B06(0-7) B07(0-7) + + const __m256i rhs_mat_0145_01 = _mm256_and_si256(rhs_raw_mat_0145_1, m4b); //B00(8-15) B01(8-15) B04(8-15) B05(8-15) + const __m256i rhs_mat_2367_01 = _mm256_and_si256(rhs_raw_mat_2367_1, m4b); //B02(8-15) B03(8-15) B06(8-15) B07(8-15) + + const __m256i rhs_mat_0145_02 = _mm256_and_si256(rhs_raw_mat_0145_2, m4b); //B00(16-23) B01(16-23) B04(16-23) B05(16-23) + const __m256i rhs_mat_2367_02 = _mm256_and_si256(rhs_raw_mat_2367_2, m4b); //B02(16-23) B03(16-23) B06(16-23) B07(16-23) + + const __m256i rhs_mat_0145_03 = _mm256_and_si256(rhs_raw_mat_0145_3, m4b); //B00(24-31) B01(24-31) B04(24-31) B05(24-31) + const __m256i rhs_mat_2367_03 = _mm256_and_si256(rhs_raw_mat_2367_3, m4b); //B02(24-31) B03(24-31) B06(24-31) B07(24-31) + + // Second sub block of the two sub blocks processed in the iteration + const __m256i rhs_mat_0145_10 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_0, 4), m4b); //B10(0-7) B11(0-7) B14(0-7) B15(0-7) + const __m256i rhs_mat_2367_10 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_0, 4), m4b); //B12(0-7) B13(0-7) B16(0-7) B17(0-7) + + const __m256i rhs_mat_0145_11 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_1, 4), m4b); //B10(8-15) B11(8-15) B14(8-15) B15(8-15) + const __m256i rhs_mat_2367_11 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_1, 4), m4b); //B12(8-15) B13(8-15) B16(8-15) B17(8-15) + + const __m256i rhs_mat_0145_12 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_2, 4), m4b); //B10(16-23) B11(16-23) B14(16-23) B15(16-23) + const __m256i rhs_mat_2367_12 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_2, 4), m4b); //B12(16-23) B13(16-23) B16(16-23) B17(16-23) + + const __m256i rhs_mat_0145_13 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_3, 4), m4b); //B10(24-31) B11(24-31) B14(24-31) B15(24-31) + const __m256i rhs_mat_2367_13 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_3, 4), m4b); //B12(24-31) B13(24-31) B16(24-31) B17(24-31) + + // Shuffle pattern one - right side input + const __m256i rhs_mat_0145_00_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_00, 136); //B00(0-3) B01(0-3) B00(0-3) B01(0-3) B04(0-3) B05(0-3) B04(0-3) B05(0-3) + const __m256i rhs_mat_2367_00_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_00, 136); //B02(0-3) B03(0-3) B02(0-3) B03(0-3) B06(0-3) B07(0-3) B06(0-3) B07(0-3) + + const __m256i rhs_mat_0145_01_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_01, 136); //B00(8-11) B01(8-11) B00(8-11) B01(8-11) B04(8-11) B05(8-11) B04(8-11) B05(8-11) + const __m256i rhs_mat_2367_01_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_01, 136); //B02(8-11) B03(8-11) B02(8-11) B03(8-11) B06(8-11) B07(8-11) B06(8-11) B07(8-11) + + const __m256i rhs_mat_0145_02_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_02, 136); //B00(16-19) B01(16-19) B00(16-19) B01(16-19) B04(16-19) B05(16-19) B04(16-19) B05(16-19) + const __m256i rhs_mat_2367_02_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_02, 136); //B02(16-19) B03(16-19) B02(16-19) B03(16-19) B06(16-19) B07(16-19) B06(16-19) B07(16-19) + + const __m256i rhs_mat_0145_03_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_03, 136); //B00(24-27) B01(24-27) B00(24-27) B01(24-27) B04(24-27) B05(24-27) B04(24-27) B05(24-27) + const __m256i rhs_mat_2367_03_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_03, 136); //B02(24-27) B03(24-27) B02(24-27) B03(24-27) B06(24-27) B07(24-27) B06(24-27) B07(24-27) + + const __m256i rhs_mat_0145_10_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_10, 136); //B10(0-3) B11(0-3) B10(0-3) B11(0-3) B14(0-3) B15(0-3) B14(0-3) B15(0-3) + const __m256i rhs_mat_2367_10_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_10, 136); //B12(0-3) B13(0-3) B12(0-3) B13(0-3) B16(0-3) B17(0-3) B16(0-3) B17(0-3) + + const __m256i rhs_mat_0145_11_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_11, 136); //B10(8-11) B11(8-11) B10(8-11) B11(8-11) B14(8-11) B15(8-11) B14(8-11) B15(8-11) + const __m256i rhs_mat_2367_11_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_11, 136); //B12(8-11) B13(8-11) B12(8-11) B13(8-11) B16(8-11) B17(8-11) B16(8-11) B17(8-11) + + const __m256i rhs_mat_0145_12_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_12, 136); //B10(16-19) B11(16-19) B10(16-19) B11(16-19) B14(16-19) B15(16-19) B14(16-19) B15(16-19) + const __m256i rhs_mat_2367_12_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_12, 136); //B12(16-19) B13(16-19) B12(16-19) B13(16-19) B16(16-19) B17(16-19) B16(16-19) B17(16-19) + + const __m256i rhs_mat_0145_13_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_13, 136); //B10(24-27) B11(24-27) B10(24-27) B11(24-27) B14(24-27) B15(24-27) B14(24-27) B15(24-27) + const __m256i rhs_mat_2367_13_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_13, 136); //B12(24-27) B13(24-27) B12(24-27) B13(24-27) B16(24-27) B17(24-27) B16(24-27) B17(24-27) + + + // Shuffle pattern two - right side input + const __m256i rhs_mat_0145_00_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_00, 221); //B00(4-7) B01(4-7) B00(4-7) B01(4-7) B04(4-7) B05(4-7) B04(4-7) B05(4-7) + const __m256i rhs_mat_2367_00_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_00, 221); //B02(4-7) B03(4-7) B02(4-7) B03(4-7) B06(4-7) B07(4-7) B06(4-7) B07(4-7) + + const __m256i rhs_mat_0145_01_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_01, 221); //B00(12-15) B01(12-15) B00(12-15) B01(12-15) B04(12-15) B05(12-15) B04(12-15) B05(12-15) + const __m256i rhs_mat_2367_01_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_01, 221); //B02(12-15) B03(12-15) B02(12-15) B03(12-15) B06(12-15) B07(12-15) B06(12-15) B07(12-15) + + const __m256i rhs_mat_0145_02_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_02, 221); //B00(20-23) B01(20-23) B00(20-23) B01(20-23) B04(20-23) B05(20-23) B04(20-23) B05(20-23) + const __m256i rhs_mat_2367_02_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_02, 221); //B02(20-23) B03(20-23) B02(20-23) B03(20-23) B06(20-23) B07(20-23) B06(20-23) B07(20-23) + + const __m256i rhs_mat_0145_03_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_03, 221); //B00(28-31) B01(28-31) B00(28-31) B01(28-31) B04(28-31) B05(28-31) B04(28-31) B05(28-31) + const __m256i rhs_mat_2367_03_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_03, 221); //B02(28-31) B03(28-31) B02(28-31) B03(28-31) B06(28-31) B07(28-31) B06(28-31) B07(28-31) + + const __m256i rhs_mat_0145_10_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_10, 221); //B10(4-7) B11(4-7) B10(4-7) B11(4-7) B14(4-7) B15(4-7) B14(4-7) B15(4-7) + const __m256i rhs_mat_2367_10_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_10, 221); //B12(4-7) B13(4-7) B12(4-7) B13(4-7) B16(4-7) B17(4-7) B16(4-7) B17(4-7) + + const __m256i rhs_mat_0145_11_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_11, 221); //B10(12-15) B11(12-15) B10(12-15) B11(12-15) B14(12-15) B15(12-15) B14(12-15) B15(12-15) + const __m256i rhs_mat_2367_11_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_11, 221); //B12(12-15) B13(12-15) B12(12-15) B13(12-15) B16(12-15) B17(12-15) B16(12-15) B17(12-15) + + const __m256i rhs_mat_0145_12_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_12, 221); //B10(20-23) B11(20-23) B10(20-23) B11(20-23) B14(20-23) B15(20-23) B14(20-23) B15(20-23) + const __m256i rhs_mat_2367_12_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_12, 221); //B12(20-23) B13(20-23) B12(20-23) B13(20-23) B16(20-23) B17(20-23) B16(20-23) B17(20-23) + + const __m256i rhs_mat_0145_13_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_13, 221); //B10(28-31) B11(28-31) B10(28-31) B11(28-31) B14(28-31) B15(28-31) B14(28-31) B15(28-31) + const __m256i rhs_mat_2367_13_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_13, 221); //B12(28-31) B13(28-31) B12(28-31) B13(28-31) B16(28-31) B17(28-31) B16(28-31) B17(28-31) + + uint32_t utmp_0[4], utmp_1[4]; + + // Scales and Mins of corresponding sub blocks from different Q4_K structures are stored together + // The below block is for eg to extract first sub block's scales and mins from different Q4_K structures for the sb loop + memcpy(utmp_0, b_ptr[b].scales + 24 * sb, 12); + utmp_0[3] = ((utmp_0[2] >> 4) & kmask2) | (((utmp_0[1] >> 6) & kmask3) << 4); + const uint32_t uaux_0 = utmp_0[1] & kmask1; + utmp_0[1] = (utmp_0[2] & kmask2) | (((utmp_0[0] >> 6) & kmask3) << 4); + utmp_0[2] = uaux_0; + utmp_0[0] &= kmask1; + + // The below block is for eg to extract second sub block's scales and mins from different Q4_K structures for the sb loop + memcpy(utmp_1, b_ptr[b].scales + 12 + sb * 24, 12); + utmp_1[3] = ((utmp_1[2] >> 4) & kmask2) | (((utmp_1[1] >> 6) & kmask3) << 4); + const uint32_t uaux_1 = utmp_1[1] & kmask1; + utmp_1[1] = (utmp_1[2] & kmask2) | (((utmp_1[0] >> 6) & kmask3) << 4); + utmp_1[2] = uaux_1; + utmp_1[0] &= kmask1; + + // Scales of first sub block in the sb loop + const __m128i mins_and_scales_0 = _mm_set_epi32(utmp_0[3], utmp_0[2], utmp_0[1], utmp_0[0]); + const __m256i scales_0 = _mm256_cvtepu8_epi16(_mm_unpacklo_epi8(mins_and_scales_0, mins_and_scales_0)); + + // Scales of second sub block in the sb loop + const __m128i mins_and_scales_1 = _mm_set_epi32(utmp_1[3], utmp_1[2], utmp_1[1], utmp_1[0]); + const __m256i scales_1 = _mm256_cvtepu8_epi16(_mm_unpacklo_epi8(mins_and_scales_1, mins_and_scales_1)); + + // Mins of first and second sub block of Q4_K block are arranged side by side + const __m256i mins_01 = _mm256_cvtepu8_epi16(_mm_unpacklo_epi8(_mm_shuffle_epi32(mins_and_scales_0, 78), _mm_shuffle_epi32(mins_and_scales_1, 78))); + + const __m256i scale_0145_0 = _mm256_shuffle_epi32(scales_0, 68); + const __m256i scale_2367_0 = _mm256_shuffle_epi32(scales_0, 238); + + const __m256i scale_0145_1 = _mm256_shuffle_epi32(scales_1, 68); + const __m256i scale_2367_1 = _mm256_shuffle_epi32(scales_1, 238); + + for (int rp = 0; rp < 4; rp++) { + + // Load the four block_q8_k quantized values interleaved with each other in chunks of eight bytes - A0,A1,A2,A3 + // Loaded as set of 128 bit vectors and repeated into a 256 bit vector + __m256i lhs_mat_0123_00 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 256 * sb))); + __m256i lhs_mat_01_00 = _mm256_permute2f128_si256(lhs_mat_0123_00, lhs_mat_0123_00, 0); + __m256i lhs_mat_23_00 = _mm256_permute2f128_si256(lhs_mat_0123_00, lhs_mat_0123_00, 17); + __m256i lhs_mat_0123_01 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 32 + 256 * sb))); + __m256i lhs_mat_01_01 = _mm256_permute2f128_si256(lhs_mat_0123_01, lhs_mat_0123_01, 0); + __m256i lhs_mat_23_01 = _mm256_permute2f128_si256(lhs_mat_0123_01, lhs_mat_0123_01, 17); + __m256i lhs_mat_0123_02 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 64 + 256 * sb))); + __m256i lhs_mat_01_02 = _mm256_permute2f128_si256(lhs_mat_0123_02, lhs_mat_0123_02, 0); + __m256i lhs_mat_23_02 = _mm256_permute2f128_si256(lhs_mat_0123_02, lhs_mat_0123_02, 17); + __m256i lhs_mat_0123_03 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 96 + 256 * sb))); + __m256i lhs_mat_01_03 = _mm256_permute2f128_si256(lhs_mat_0123_03, lhs_mat_0123_03, 0); + __m256i lhs_mat_23_03 = _mm256_permute2f128_si256(lhs_mat_0123_03, lhs_mat_0123_03, 17); + __m256i lhs_mat_0123_10 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 128 + 256 * sb))); + __m256i lhs_mat_01_10 = _mm256_permute2f128_si256(lhs_mat_0123_10, lhs_mat_0123_10, 0); + __m256i lhs_mat_23_10 = _mm256_permute2f128_si256(lhs_mat_0123_10, lhs_mat_0123_10, 17); + __m256i lhs_mat_0123_11 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 160 + 256 * sb))); + __m256i lhs_mat_01_11 = _mm256_permute2f128_si256(lhs_mat_0123_11, lhs_mat_0123_11, 0); + __m256i lhs_mat_23_11 = _mm256_permute2f128_si256(lhs_mat_0123_11, lhs_mat_0123_11, 17); + __m256i lhs_mat_0123_12 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 192 + 256 * sb))); + __m256i lhs_mat_01_12 = _mm256_permute2f128_si256(lhs_mat_0123_12, lhs_mat_0123_12, 0); + __m256i lhs_mat_23_12 = _mm256_permute2f128_si256(lhs_mat_0123_12, lhs_mat_0123_12, 17); + __m256i lhs_mat_0123_13 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 224 + 256 * sb))); + __m256i lhs_mat_01_13 = _mm256_permute2f128_si256(lhs_mat_0123_13, lhs_mat_0123_13, 0); + __m256i lhs_mat_23_13 = _mm256_permute2f128_si256(lhs_mat_0123_13, lhs_mat_0123_13, 17); + + // Bsums are loaded - four bsums are loaded (for two sub blocks) for the different Q8_K blocks + __m256i lhs_bsums_0123_01 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].bsums + 16 * sb))); + __m256i lhs_bsums_hsum_0123_01 = _mm256_castsi128_si256(_mm_hadd_epi16(_mm256_castsi256_si128(lhs_bsums_0123_01), _mm256_extractf128_si256(lhs_bsums_0123_01, 1))); + lhs_bsums_hsum_0123_01 = _mm256_permute2x128_si256(lhs_bsums_hsum_0123_01, lhs_bsums_hsum_0123_01, 0); + + // Shuffle pattern one - left side input + const __m256i lhs_mat_01_00_sp1 = _mm256_shuffle_epi32(lhs_mat_01_00, 160); //A00(0-3) A00(0-3) A01(0-3) A01(0-3) A00(0-3) A00(0-3) A01(0-3) A01(0-3) + const __m256i lhs_mat_23_00_sp1 = _mm256_shuffle_epi32(lhs_mat_23_00, 160); //A02(0-3) A03(0-3) A02(0-3) A03(0-3) A02(0-3) A03(0-3) A02(0-3) A03(0-3) + + const __m256i lhs_mat_01_01_sp1 = _mm256_shuffle_epi32(lhs_mat_01_01, 160); //A00(8-11) A00(8-11) A01(8-11) A01(8-11) A00(8-11) A00(8-11) A01(8-11) A01(8-11) + const __m256i lhs_mat_23_01_sp1 = _mm256_shuffle_epi32(lhs_mat_23_01, 160); //A02(8-11) A03(8-11) A02(8-11) A03(8-11) A02(8-11) A03(8-11) A02(8-11) A03(8-11) + + const __m256i lhs_mat_01_02_sp1 = _mm256_shuffle_epi32(lhs_mat_01_02, 160); //A00(16-19) A00(16-19) A01(16-19) A01(16-19) A00(16-19) A00(16-19) A01(16-19) A01(16-19) + const __m256i lhs_mat_23_02_sp1 = _mm256_shuffle_epi32(lhs_mat_23_02, 160); //A02(16-19) A03(16-19) A02(16-19) A03(16-19) A02(16-19) A03(16-19) A02(16-19) A03(16-19) + + const __m256i lhs_mat_01_03_sp1 = _mm256_shuffle_epi32(lhs_mat_01_03, 160); //A00(24-27) A00(24-27) A01(24-27) A01(24-27) A00(24-27) A00(24-27) A01(24-27) A01(24-27) + const __m256i lhs_mat_23_03_sp1 = _mm256_shuffle_epi32(lhs_mat_23_03, 160); //A02(24-27) A03(24-27) A02(24-27) A03(24-27) A02(24-27) A03(24-27) A02(24-27) A03(24-27) + + const __m256i lhs_mat_01_10_sp1 = _mm256_shuffle_epi32(lhs_mat_01_10, 160); //A10(0-3) A10(0-3) A11(0-3) A11(0-3) A10(0-3) A10(0-3) A11(0-3) A11(0-3) + const __m256i lhs_mat_23_10_sp1 = _mm256_shuffle_epi32(lhs_mat_23_10, 160); //A12(0-3) A13(0-3) A12(0-3) A13(0-3) A12(0-3) A13(0-3) A12(0-3) A13(0-3) + + const __m256i lhs_mat_01_11_sp1 = _mm256_shuffle_epi32(lhs_mat_01_11, 160); //A10(8-11) A10(8-11) A11(8-11) A11(8-11) A10(8-11) A10(8-11) A11(8-11) A11(8-11) + const __m256i lhs_mat_23_11_sp1 = _mm256_shuffle_epi32(lhs_mat_23_11, 160); //A12(8-11) A13(8-11) A12(8-11) A13(8-11) A12(8-11) A13(8-11) A12(8-11) A13(8-11) + + const __m256i lhs_mat_01_12_sp1 = _mm256_shuffle_epi32(lhs_mat_01_12, 160); //A10(16-19) A10(16-19) A11(16-19) A11(16-19) A10(16-19) A10(16-19) A11(16-19) A11(16-19) + const __m256i lhs_mat_23_12_sp1 = _mm256_shuffle_epi32(lhs_mat_23_12, 160); //A12(16-19) A13(16-19) A12(16-19) A13(16-19) A12(16-19) A13(16-19) A12(16-19) A13(16-19) + + const __m256i lhs_mat_01_13_sp1 = _mm256_shuffle_epi32(lhs_mat_01_13, 160); //A10(24-27) A10(24-27) A11(24-27) A11(24-27) A10(24-27) A10(24-27) A11(24-27) A11(24-27) + const __m256i lhs_mat_23_13_sp1 = _mm256_shuffle_epi32(lhs_mat_23_13, 160); //A12(24-27) A13(24-27) A12(24-27) A13(24-27) A12(24-27) A13(24-27) A12(24-27) A13(24-27) + + // Shuffle pattern two- left side input + const __m256i lhs_mat_01_00_sp2 = _mm256_shuffle_epi32(lhs_mat_01_00, 245); //A00(4-7) A00(4-7) A01(4-7) A01(4-7) A00(4-7) A00(4-7) A01(4-7) A01(4-7) + const __m256i lhs_mat_23_00_sp2 = _mm256_shuffle_epi32(lhs_mat_23_00, 245); //A02(4-7) A03(4-7) A02(4-7) A03(4-7) A02(4-7) A03(4-7) A02(4-7) A03(4-7) + + const __m256i lhs_mat_01_01_sp2 = _mm256_shuffle_epi32(lhs_mat_01_01, 245); //A00(12-15) A00(12-15) A01(12-15) A01(12-15) A00(12-15) A00(12-15) A01(12-15) A01(12-15) + const __m256i lhs_mat_23_01_sp2 = _mm256_shuffle_epi32(lhs_mat_23_01, 245); //A02(12-15) A03(12-15) A02(12-15) A03(12-15) A02(12-15) A03(12-15) A02(12-15) A03(12-15) + + const __m256i lhs_mat_01_02_sp2 = _mm256_shuffle_epi32(lhs_mat_01_02, 245); //A00(20-23) A00(20-23) A01(20-23) A01(20-23) A00(20-23) A00(20-23) A01(20-23) A01(20-23) + const __m256i lhs_mat_23_02_sp2 = _mm256_shuffle_epi32(lhs_mat_23_02, 245); //A02(20-23) A03(20-23) A02(20-23) A03(20-23) A02(20-23) A03(20-23) A02(20-23) A03(20-23) + + const __m256i lhs_mat_01_03_sp2 = _mm256_shuffle_epi32(lhs_mat_01_03, 245); //A00(28-31) A00(28-31) A01(28-31) A01(28-31) A00(28-31) A00(28-31) A01(28-31) A01(28-31) + const __m256i lhs_mat_23_03_sp2 = _mm256_shuffle_epi32(lhs_mat_23_03, 245); //A02(28-31) A03(28-31) A02(28-31) A03(28-31) A02(28-31) A03(28-31) A02(28-31) A03(28-31) + + const __m256i lhs_mat_01_10_sp2 = _mm256_shuffle_epi32(lhs_mat_01_10, 245); //A10(4-7) A10(4-7) A11(4-7) A11(4-7) A10(4-7) A10(4-7) A11(4-7) A11(4-7) + const __m256i lhs_mat_23_10_sp2 = _mm256_shuffle_epi32(lhs_mat_23_10, 245); //A12(4-7) A13(4-7) A12(4-7) A13(4-7) A12(4-7) A13(4-7) A12(4-7) A13(4-7) + + const __m256i lhs_mat_01_11_sp2 = _mm256_shuffle_epi32(lhs_mat_01_11, 245); //A10(12-15) A10(12-15) A11(12-15) A11(12-15) A10(12-15) A10(12-15) A11(12-15) A11(12-15) + const __m256i lhs_mat_23_11_sp2 = _mm256_shuffle_epi32(lhs_mat_23_11, 245); //A12(12-15) A13(12-15) A12(12-15) A13(12-15) A12(12-15) A13(12-15) A12(12-15) A13(12-15) + + const __m256i lhs_mat_01_12_sp2 = _mm256_shuffle_epi32(lhs_mat_01_12, 245); //A10(20-23) A10(20-23) A11(20-23) A11(20-23) A10(20-23) A10(20-23) A11(20-23) A11(20-23) + const __m256i lhs_mat_23_12_sp2 = _mm256_shuffle_epi32(lhs_mat_23_12, 245); //A12(20-23) A13(20-23) A12(20-23) A13(20-23) A12(20-23) A13(20-23) A12(20-23) A13(20-23) + + const __m256i lhs_mat_01_13_sp2 = _mm256_shuffle_epi32(lhs_mat_01_13, 245); //A10(28-31) A10(28-31) A11(28-31) A11(28-31) A10(28-31) A10(28-31) A11(28-31) A11(28-31) + const __m256i lhs_mat_23_13_sp2 = _mm256_shuffle_epi32(lhs_mat_23_13, 245); //A12(28-31) A13(28-31) A12(28-31) A13(28-31) A12(28-31) A13(28-31) A12(28-31) A13(28-31) + + // The values arranged in shuffle patterns are operated with dot product operation within 32 bit lane i.e corresponding bytes and multiplied and added into 32 bit integers within 32 bit lane + __m256i iacc_mat_00_0_sp1 = _mm256_add_epi16(_mm256_add_epi16(_mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_03_sp1, lhs_mat_01_03_sp1), _mm256_maddubs_epi16(rhs_mat_0145_02_sp1, lhs_mat_01_02_sp1)), _mm256_maddubs_epi16(rhs_mat_0145_01_sp1, lhs_mat_01_01_sp1)), _mm256_maddubs_epi16(rhs_mat_0145_00_sp1, lhs_mat_01_00_sp1)); + __m256i iacc_mat_01_0_sp1 = _mm256_add_epi16(_mm256_add_epi16(_mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_03_sp1, lhs_mat_01_03_sp1), _mm256_maddubs_epi16(rhs_mat_2367_02_sp1, lhs_mat_01_02_sp1)), _mm256_maddubs_epi16(rhs_mat_2367_01_sp1, lhs_mat_01_01_sp1)), _mm256_maddubs_epi16(rhs_mat_2367_00_sp1, lhs_mat_01_00_sp1)); + __m256i iacc_mat_10_0_sp1 = _mm256_add_epi16(_mm256_add_epi16(_mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_03_sp1, lhs_mat_23_03_sp1), _mm256_maddubs_epi16(rhs_mat_0145_02_sp1, lhs_mat_23_02_sp1)), _mm256_maddubs_epi16(rhs_mat_0145_01_sp1, lhs_mat_23_01_sp1)), _mm256_maddubs_epi16(rhs_mat_0145_00_sp1, lhs_mat_23_00_sp1)); + __m256i iacc_mat_11_0_sp1 = _mm256_add_epi16(_mm256_add_epi16(_mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_03_sp1, lhs_mat_23_03_sp1), _mm256_maddubs_epi16(rhs_mat_2367_02_sp1, lhs_mat_23_02_sp1)), _mm256_maddubs_epi16(rhs_mat_2367_01_sp1, lhs_mat_23_01_sp1)), _mm256_maddubs_epi16(rhs_mat_2367_00_sp1, lhs_mat_23_00_sp1)); + __m256i iacc_mat_00_1_sp1 = _mm256_add_epi16(_mm256_add_epi16(_mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_13_sp1, lhs_mat_01_13_sp1), _mm256_maddubs_epi16(rhs_mat_0145_12_sp1, lhs_mat_01_12_sp1)), _mm256_maddubs_epi16(rhs_mat_0145_11_sp1, lhs_mat_01_11_sp1)), _mm256_maddubs_epi16(rhs_mat_0145_10_sp1, lhs_mat_01_10_sp1)); + __m256i iacc_mat_01_1_sp1 = _mm256_add_epi16(_mm256_add_epi16(_mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_13_sp1, lhs_mat_01_13_sp1), _mm256_maddubs_epi16(rhs_mat_2367_12_sp1, lhs_mat_01_12_sp1)), _mm256_maddubs_epi16(rhs_mat_2367_11_sp1, lhs_mat_01_11_sp1)), _mm256_maddubs_epi16(rhs_mat_2367_10_sp1, lhs_mat_01_10_sp1)); + __m256i iacc_mat_10_1_sp1 = _mm256_add_epi16(_mm256_add_epi16(_mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_13_sp1, lhs_mat_23_13_sp1), _mm256_maddubs_epi16(rhs_mat_0145_12_sp1, lhs_mat_23_12_sp1)), _mm256_maddubs_epi16(rhs_mat_0145_11_sp1, lhs_mat_23_11_sp1)), _mm256_maddubs_epi16(rhs_mat_0145_10_sp1, lhs_mat_23_10_sp1)); + __m256i iacc_mat_11_1_sp1 = _mm256_add_epi16(_mm256_add_epi16(_mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_13_sp1, lhs_mat_23_13_sp1), _mm256_maddubs_epi16(rhs_mat_2367_12_sp1, lhs_mat_23_12_sp1)), _mm256_maddubs_epi16(rhs_mat_2367_11_sp1, lhs_mat_23_11_sp1)), _mm256_maddubs_epi16(rhs_mat_2367_10_sp1, lhs_mat_23_10_sp1)); + + __m256i iacc_mat_00_0_sp2 = _mm256_add_epi16(_mm256_add_epi16(_mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_03_sp2, lhs_mat_01_03_sp2), _mm256_maddubs_epi16(rhs_mat_0145_02_sp2, lhs_mat_01_02_sp2)), _mm256_maddubs_epi16(rhs_mat_0145_01_sp2, lhs_mat_01_01_sp2)), _mm256_maddubs_epi16(rhs_mat_0145_00_sp2, lhs_mat_01_00_sp2)); + __m256i iacc_mat_01_0_sp2 = _mm256_add_epi16(_mm256_add_epi16(_mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_03_sp2, lhs_mat_01_03_sp2), _mm256_maddubs_epi16(rhs_mat_2367_02_sp2, lhs_mat_01_02_sp2)), _mm256_maddubs_epi16(rhs_mat_2367_01_sp2, lhs_mat_01_01_sp2)), _mm256_maddubs_epi16(rhs_mat_2367_00_sp2, lhs_mat_01_00_sp2)); + __m256i iacc_mat_10_0_sp2 = _mm256_add_epi16(_mm256_add_epi16(_mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_03_sp2, lhs_mat_23_03_sp2), _mm256_maddubs_epi16(rhs_mat_0145_02_sp2, lhs_mat_23_02_sp2)), _mm256_maddubs_epi16(rhs_mat_0145_01_sp2, lhs_mat_23_01_sp2)), _mm256_maddubs_epi16(rhs_mat_0145_00_sp2, lhs_mat_23_00_sp2)); + __m256i iacc_mat_11_0_sp2 = _mm256_add_epi16(_mm256_add_epi16(_mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_03_sp2, lhs_mat_23_03_sp2), _mm256_maddubs_epi16(rhs_mat_2367_02_sp2, lhs_mat_23_02_sp2)), _mm256_maddubs_epi16(rhs_mat_2367_01_sp2, lhs_mat_23_01_sp2)), _mm256_maddubs_epi16(rhs_mat_2367_00_sp2, lhs_mat_23_00_sp2)); + __m256i iacc_mat_00_1_sp2 = _mm256_add_epi16(_mm256_add_epi16(_mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_13_sp2, lhs_mat_01_13_sp2), _mm256_maddubs_epi16(rhs_mat_0145_12_sp2, lhs_mat_01_12_sp2)), _mm256_maddubs_epi16(rhs_mat_0145_11_sp2, lhs_mat_01_11_sp2)), _mm256_maddubs_epi16(rhs_mat_0145_10_sp2, lhs_mat_01_10_sp2)); + __m256i iacc_mat_01_1_sp2 = _mm256_add_epi16(_mm256_add_epi16(_mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_13_sp2, lhs_mat_01_13_sp2), _mm256_maddubs_epi16(rhs_mat_2367_12_sp2, lhs_mat_01_12_sp2)), _mm256_maddubs_epi16(rhs_mat_2367_11_sp2, lhs_mat_01_11_sp2)), _mm256_maddubs_epi16(rhs_mat_2367_10_sp2, lhs_mat_01_10_sp2)); + __m256i iacc_mat_10_1_sp2 = _mm256_add_epi16(_mm256_add_epi16(_mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_13_sp2, lhs_mat_23_13_sp2), _mm256_maddubs_epi16(rhs_mat_0145_12_sp2, lhs_mat_23_12_sp2)), _mm256_maddubs_epi16(rhs_mat_0145_11_sp2, lhs_mat_23_11_sp2)), _mm256_maddubs_epi16(rhs_mat_0145_10_sp2, lhs_mat_23_10_sp2)); + __m256i iacc_mat_11_1_sp2 = _mm256_add_epi16(_mm256_add_epi16(_mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_13_sp2, lhs_mat_23_13_sp2), _mm256_maddubs_epi16(rhs_mat_2367_12_sp2, lhs_mat_23_12_sp2)), _mm256_maddubs_epi16(rhs_mat_2367_11_sp2, lhs_mat_23_11_sp2)), _mm256_maddubs_epi16(rhs_mat_2367_10_sp2, lhs_mat_23_10_sp2)); + + // Output of both shuffle patterns are added in order to sum dot product outputs of all 32 values in block + __m256i iacc_mat_00_0 = _mm256_add_epi16(iacc_mat_00_0_sp1, iacc_mat_00_0_sp2); + __m256i iacc_mat_01_0 = _mm256_add_epi16(iacc_mat_01_0_sp1, iacc_mat_01_0_sp2); + __m256i iacc_mat_10_0 = _mm256_add_epi16(iacc_mat_10_0_sp1, iacc_mat_10_0_sp2); + __m256i iacc_mat_11_0 = _mm256_add_epi16(iacc_mat_11_0_sp1, iacc_mat_11_0_sp2); + + __m256i iacc_mat_00_1 = _mm256_add_epi16(iacc_mat_00_1_sp1, iacc_mat_00_1_sp2); + __m256i iacc_mat_01_1 = _mm256_add_epi16(iacc_mat_01_1_sp1, iacc_mat_01_1_sp2); + __m256i iacc_mat_10_1 = _mm256_add_epi16(iacc_mat_10_1_sp1, iacc_mat_10_1_sp2); + __m256i iacc_mat_11_1 = _mm256_add_epi16(iacc_mat_11_1_sp1, iacc_mat_11_1_sp2); + + // Output of both shuffle patterns are added in order to sum dot product outputs of all 32 values in block + iacc_mat_00_0 = _mm256_madd_epi16(iacc_mat_00_0, scale_0145_0); + iacc_mat_01_0 = _mm256_madd_epi16(iacc_mat_01_0, scale_2367_0); + iacc_mat_10_0 = _mm256_madd_epi16(iacc_mat_10_0, scale_0145_0); + iacc_mat_11_0 = _mm256_madd_epi16(iacc_mat_11_0, scale_2367_0); + + iacc_mat_00_1 = _mm256_madd_epi16(iacc_mat_00_1, scale_0145_1); + iacc_mat_01_1 = _mm256_madd_epi16(iacc_mat_01_1, scale_2367_1); + iacc_mat_10_1 = _mm256_madd_epi16(iacc_mat_10_1, scale_0145_1); + iacc_mat_11_1 = _mm256_madd_epi16(iacc_mat_11_1, scale_2367_1); + + // Straighten out to make 4 row vectors (4 for each sub block which are accumulated together in the next step) + __m256i iacc_row_0_0 = _mm256_blend_epi32(iacc_mat_00_0, _mm256_shuffle_epi32(iacc_mat_01_0, 78), 204); + __m256i iacc_row_1_0 = _mm256_blend_epi32(_mm256_shuffle_epi32(iacc_mat_00_0, 78), iacc_mat_01_0, 204); + __m256i iacc_row_2_0 = _mm256_blend_epi32(iacc_mat_10_0, _mm256_shuffle_epi32(iacc_mat_11_0, 78), 204); + __m256i iacc_row_3_0 = _mm256_blend_epi32(_mm256_shuffle_epi32(iacc_mat_10_0, 78), iacc_mat_11_0, 204); + __m256i iacc_row_0_1 = _mm256_blend_epi32(iacc_mat_00_1, _mm256_shuffle_epi32(iacc_mat_01_1, 78), 204); + __m256i iacc_row_1_1 = _mm256_blend_epi32(_mm256_shuffle_epi32(iacc_mat_00_1, 78), iacc_mat_01_1, 204); + __m256i iacc_row_2_1 = _mm256_blend_epi32(iacc_mat_10_1, _mm256_shuffle_epi32(iacc_mat_11_1, 78), 204); + __m256i iacc_row_3_1 = _mm256_blend_epi32(_mm256_shuffle_epi32(iacc_mat_10_1, 78), iacc_mat_11_1, 204); + + __m256i iacc_row_0 = _mm256_add_epi32(iacc_row_0_0, iacc_row_0_1); + __m256i iacc_row_1 = _mm256_add_epi32(iacc_row_1_0, iacc_row_1_1); + __m256i iacc_row_2 = _mm256_add_epi32(iacc_row_2_0, iacc_row_2_1); + __m256i iacc_row_3 = _mm256_add_epi32(iacc_row_3_0, iacc_row_3_1); + + // Load the scale(d) values for all the 4 Q8_k blocks and repeat it across lanes + const __m128 row_scale_f32_sse = _mm_load_ps(a_ptrs[rp][b].d); + const __m256 row_scale_f32 = _mm256_set_m128(row_scale_f32_sse, row_scale_f32_sse);//GGML_F32Cx8_REPEAT_LOAD(a_ptrs[rp][b].d, loadMask); + + // Multiply with appropiate scales and accumulate (for both d and dmin) below + acc_rows[rp * 4] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_0), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 0)), acc_rows[rp * 4]); + acc_rows[rp * 4 + 1] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_1), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 85)), acc_rows[rp * 4 + 1]); + acc_rows[rp * 4 + 2] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_2), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 170)), acc_rows[rp * 4 + 2]); + acc_rows[rp * 4 + 3] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_3), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 255)), acc_rows[rp * 4 + 3]); + + __m256i iacc_row_min_0 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_hsum_0123_01, 0), mins_01); + __m256i iacc_row_min_1 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_hsum_0123_01, 85), mins_01); + __m256i iacc_row_min_2 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_hsum_0123_01, 170), mins_01); + __m256i iacc_row_min_3 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_hsum_0123_01, 255), mins_01); + + acc_min_rows[rp * 4] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_min_0), _mm256_mul_ps(col_dmin_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 0)), acc_min_rows[rp * 4]); + acc_min_rows[rp * 4 + 1] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_min_1), _mm256_mul_ps(col_dmin_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 85)), acc_min_rows[rp * 4 + 1]); + acc_min_rows[rp * 4 + 2] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_min_2), _mm256_mul_ps(col_dmin_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 170)), acc_min_rows[rp * 4 + 2]); + acc_min_rows[rp * 4 + 3] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_min_3), _mm256_mul_ps(col_dmin_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 255)), acc_min_rows[rp * 4 + 3]); + + } + } + } + // Store the accumulated values + for (int i = 0; i < 16; i++) { + _mm256_storeu_ps((float * )(s + ((y * 4 + i) * bs + x * 8)), _mm256_sub_ps(acc_rows[i], acc_min_rows[i])); + } + } + } + for (; y < nr / 4; y++) { + + const block_q8_Kx4 * a_ptr = a_ptr_start + (y * nb); + + for (int64_t x = 0; x < nc / 8; x++) { + + const block_q4_Kx8 * b_ptr = b_ptr_start + (x * b_nb); + + // Master FP accumulators + __m256 acc_rows[4]; + for (int i = 0; i < 4; i++) { + acc_rows[i] = _mm256_setzero_ps(); + } + + __m256 acc_min_rows[4]; + for (int i = 0; i < 4; i++) { + acc_min_rows[i] = _mm256_setzero_ps(); + } + + for (int64_t b = 0; b < nb; b++) { + + // Scale values - Load the eight scale values of block_q4_Kx8 + const __m256 col_scale_f32 = GGML_F32Cx8_LOAD(b_ptr[b].d); + + // dmin values - Load the eight dmin values of block_q4_Kx8 + const __m256 col_dmin_f32 = GGML_F32Cx8_LOAD(b_ptr[b].dmin); + + // Loop to iterate over the eight sub blocks of a super block - two sub blocks are processed per iteration + for (int sb = 0; sb < QK_K / 64; sb++) { + + // Load the eight block_q4_k for two sub blocks quantized values interleaved with each other in chunks of eight bytes - B0,B1 ....B6,B7 + const __m256i rhs_raw_mat_0123_0 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + sb * 256)); + const __m256i rhs_raw_mat_4567_0 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 32 + sb * 256)); + const __m256i rhs_raw_mat_0123_1 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 64 + sb * 256)); + const __m256i rhs_raw_mat_4567_1 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 96 + sb * 256)); + const __m256i rhs_raw_mat_0123_2 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 128 + sb * 256)); + const __m256i rhs_raw_mat_4567_2 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 160 + sb * 256)); + const __m256i rhs_raw_mat_0123_3 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 192 + sb * 256)); + const __m256i rhs_raw_mat_4567_3 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 224 + sb * 256)); + + // Save the values in the following vectors in the formats B0B1B4B5, B2B3B6B7 for further processing and storing of values + const __m256i rhs_raw_mat_0145_0 = _mm256_blend_epi32(rhs_raw_mat_0123_0, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_0, requiredOrder), 240); + const __m256i rhs_raw_mat_2367_0 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_0, requiredOrder), rhs_raw_mat_4567_0, 240); + const __m256i rhs_raw_mat_0145_1 = _mm256_blend_epi32(rhs_raw_mat_0123_1, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_1, requiredOrder), 240); + const __m256i rhs_raw_mat_2367_1 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_1, requiredOrder), rhs_raw_mat_4567_1, 240); + const __m256i rhs_raw_mat_0145_2 = _mm256_blend_epi32(rhs_raw_mat_0123_2, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_2, requiredOrder), 240); + const __m256i rhs_raw_mat_2367_2 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_2, requiredOrder), rhs_raw_mat_4567_2, 240); + const __m256i rhs_raw_mat_0145_3 = _mm256_blend_epi32(rhs_raw_mat_0123_3, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_3, requiredOrder), 240); + const __m256i rhs_raw_mat_2367_3 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_3, requiredOrder), rhs_raw_mat_4567_3, 240); + + // 4-bit -> 8-bit + // First sub block of the two sub blocks processed in the iteration + const __m256i rhs_mat_0145_00 = _mm256_and_si256(rhs_raw_mat_0145_0, m4b); //B00(0-7) B01(0-7) B04(0-7) B05(0-7) + const __m256i rhs_mat_2367_00 = _mm256_and_si256(rhs_raw_mat_2367_0, m4b); //B02(0-7) B03(0-7) B06(0-7) B07(0-7) + + const __m256i rhs_mat_0145_01 = _mm256_and_si256(rhs_raw_mat_0145_1, m4b); //B00(8-15) B01(8-15) B04(8-15) B05(8-15) + const __m256i rhs_mat_2367_01 = _mm256_and_si256(rhs_raw_mat_2367_1, m4b); //B02(8-15) B03(8-15) B06(8-15) B07(8-15) + + const __m256i rhs_mat_0145_02 = _mm256_and_si256(rhs_raw_mat_0145_2, m4b); //B00(16-23) B01(16-23) B04(16-23) B05(16-23) + const __m256i rhs_mat_2367_02 = _mm256_and_si256(rhs_raw_mat_2367_2, m4b); //B02(16-23) B03(16-23) B06(16-23) B07(16-23) + + const __m256i rhs_mat_0145_03 = _mm256_and_si256(rhs_raw_mat_0145_3, m4b); //B00(24-31) B01(24-31) B04(24-31) B05(24-31) + const __m256i rhs_mat_2367_03 = _mm256_and_si256(rhs_raw_mat_2367_3, m4b); //B02(24-31) B03(24-31) B06(24-31) B07(24-31) + + // Second sub block of the two sub blocks processed in the iteration + const __m256i rhs_mat_0145_10 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_0, 4), m4b); //B10(0-7) B11(0-7) B14(0-7) B15(0-7) + const __m256i rhs_mat_2367_10 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_0, 4), m4b); //B12(0-7) B13(0-7) B16(0-7) B17(0-7) + + const __m256i rhs_mat_0145_11 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_1, 4), m4b); //B10(8-15) B11(8-15) B14(8-15) B15(8-15) + const __m256i rhs_mat_2367_11 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_1, 4), m4b); //B12(8-15) B13(8-15) B16(8-15) B17(8-15) + + const __m256i rhs_mat_0145_12 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_2, 4), m4b); //B10(16-23) B11(16-23) B14(16-23) B15(16-23) + const __m256i rhs_mat_2367_12 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_2, 4), m4b); //B12(16-23) B13(16-23) B16(16-23) B17(16-23) + + const __m256i rhs_mat_0145_13 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_3, 4), m4b); //B10(24-31) B11(24-31) B14(24-31) B15(24-31) + const __m256i rhs_mat_2367_13 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_3, 4), m4b); //B12(24-31) B13(24-31) B16(24-31) B17(24-31) + + // Shuffle pattern one - right side input + const __m256i rhs_mat_0145_00_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_00, 136); //B00(0-3) B01(0-3) B00(0-3) B01(0-3) B04(0-3) B05(0-3) B04(0-3) B05(0-3) + const __m256i rhs_mat_2367_00_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_00, 136); //B02(0-3) B03(0-3) B02(0-3) B03(0-3) B06(0-3) B07(0-3) B06(0-3) B07(0-3) + + const __m256i rhs_mat_0145_01_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_01, 136); //B00(8-11) B01(8-11) B00(8-11) B01(8-11) B04(8-11) B05(8-11) B04(8-11) B05(8-11) + const __m256i rhs_mat_2367_01_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_01, 136); //B02(8-11) B03(8-11) B02(8-11) B03(8-11) B06(8-11) B07(8-11) B06(8-11) B07(8-11) + + const __m256i rhs_mat_0145_02_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_02, 136); //B00(16-19) B01(16-19) B00(16-19) B01(16-19) B04(16-19) B05(16-19) B04(16-19) B05(16-19) + const __m256i rhs_mat_2367_02_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_02, 136); //B02(16-19) B03(16-19) B02(16-19) B03(16-19) B06(16-19) B07(16-19) B06(16-19) B07(16-19) + + const __m256i rhs_mat_0145_03_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_03, 136); //B00(24-27) B01(24-27) B00(24-27) B01(24-27) B04(24-27) B05(24-27) B04(24-27) B05(24-27) + const __m256i rhs_mat_2367_03_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_03, 136); //B02(24-27) B03(24-27) B02(24-27) B03(24-27) B06(24-27) B07(24-27) B06(24-27) B07(24-27) + + const __m256i rhs_mat_0145_10_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_10, 136); //B10(0-3) B11(0-3) B10(0-3) B11(0-3) B14(0-3) B15(0-3) B14(0-3) B15(0-3) + const __m256i rhs_mat_2367_10_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_10, 136); //B12(0-3) B13(0-3) B12(0-3) B13(0-3) B16(0-3) B17(0-3) B16(0-3) B17(0-3) + + const __m256i rhs_mat_0145_11_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_11, 136); //B10(8-11) B11(8-11) B10(8-11) B11(8-11) B14(8-11) B15(8-11) B14(8-11) B15(8-11) + const __m256i rhs_mat_2367_11_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_11, 136); //B12(8-11) B13(8-11) B12(8-11) B13(8-11) B16(8-11) B17(8-11) B16(8-11) B17(8-11) + + const __m256i rhs_mat_0145_12_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_12, 136); //B10(16-19) B11(16-19) B10(16-19) B11(16-19) B14(16-19) B15(16-19) B14(16-19) B15(16-19) + const __m256i rhs_mat_2367_12_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_12, 136); //B12(16-19) B13(16-19) B12(16-19) B13(16-19) B16(16-19) B17(16-19) B16(16-19) B17(16-19) + + const __m256i rhs_mat_0145_13_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_13, 136); //B10(24-27) B11(24-27) B10(24-27) B11(24-27) B14(24-27) B15(24-27) B14(24-27) B15(24-27) + const __m256i rhs_mat_2367_13_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_13, 136); //B12(24-27) B13(24-27) B12(24-27) B13(24-27) B16(24-27) B17(24-27) B16(24-27) B17(24-27) + + // Shuffle pattern two - right side input + const __m256i rhs_mat_0145_00_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_00, 221); //B00(4-7) B01(4-7) B00(4-7) B01(4-7) B04(4-7) B05(4-7) B04(4-7) B05(4-7) + const __m256i rhs_mat_2367_00_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_00, 221); //B02(4-7) B03(4-7) B02(4-7) B03(4-7) B06(4-7) B07(4-7) B06(4-7) B07(4-7) + + const __m256i rhs_mat_0145_01_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_01, 221); //B00(12-15) B01(12-15) B00(12-15) B01(12-15) B04(12-15) B05(12-15) B04(12-15) B05(12-15) + const __m256i rhs_mat_2367_01_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_01, 221); //B02(12-15) B03(12-15) B02(12-15) B03(12-15) B06(12-15) B07(12-15) B06(12-15) B07(12-15) + + const __m256i rhs_mat_0145_02_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_02, 221); //B00(20-23) B01(20-23) B00(20-23) B01(20-23) B04(20-23) B05(20-23) B04(20-23) B05(20-23) + const __m256i rhs_mat_2367_02_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_02, 221); //B02(20-23) B03(20-23) B02(20-23) B03(20-23) B06(20-23) B07(20-23) B06(20-23) B07(20-23) + + const __m256i rhs_mat_0145_03_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_03, 221); //B00(28-31) B01(28-31) B00(28-31) B01(28-31) B04(28-31) B05(28-31) B04(28-31) B05(28-31) + const __m256i rhs_mat_2367_03_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_03, 221); //B02(28-31) B03(28-31) B02(28-31) B03(28-31) B06(28-31) B07(28-31) B06(28-31) B07(28-31) + + const __m256i rhs_mat_0145_10_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_10, 221); //B10(4-7) B11(4-7) B10(4-7) B11(4-7) B14(4-7) B15(4-7) B14(4-7) B15(4-7) + const __m256i rhs_mat_2367_10_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_10, 221); //B12(4-7) B13(4-7) B12(4-7) B13(4-7) B16(4-7) B17(4-7) B16(4-7) B17(4-7) + + const __m256i rhs_mat_0145_11_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_11, 221); //B10(12-15) B11(12-15) B10(12-15) B11(12-15) B14(12-15) B15(12-15) B14(12-15) B15(12-15) + const __m256i rhs_mat_2367_11_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_11, 221); //B12(12-15) B13(12-15) B12(12-15) B13(12-15) B16(12-15) B17(12-15) B16(12-15) B17(12-15) + + const __m256i rhs_mat_0145_12_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_12, 221); //B10(20-23) B11(20-23) B10(20-23) B11(20-23) B14(20-23) B15(20-23) B14(20-23) B15(20-23) + const __m256i rhs_mat_2367_12_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_12, 221); //B12(20-23) B13(20-23) B12(20-23) B13(20-23) B16(20-23) B17(20-23) B16(20-23) B17(20-23) + + const __m256i rhs_mat_0145_13_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_13, 221); //B10(28-31) B11(28-31) B10(28-31) B11(28-31) B14(28-31) B15(28-31) B14(28-31) B15(28-31) + const __m256i rhs_mat_2367_13_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_13, 221); //B12(28-31) B13(28-31) B12(28-31) B13(28-31) B16(28-31) B17(28-31) B16(28-31) B17(28-31) + + uint32_t utmp_0[4], utmp_1[4]; + + // Scales and Mins of corresponding sub blocks from different Q4_K structures are stored together + // The below block is for eg to extract first sub block's scales and mins from different Q4_K structures for the sb loop + memcpy(utmp_0, b_ptr[b].scales + 24 * sb, 12); + utmp_0[3] = ((utmp_0[2] >> 4) & kmask2) | (((utmp_0[1] >> 6) & kmask3) << 4); + const uint32_t uaux_0 = utmp_0[1] & kmask1; + utmp_0[1] = (utmp_0[2] & kmask2) | (((utmp_0[0] >> 6) & kmask3) << 4); + utmp_0[2] = uaux_0; + utmp_0[0] &= kmask1; + + // The below block is for eg to extract second sub block's scales and mins from different Q4_K structures when sb = 1 + memcpy(utmp_1, b_ptr[b].scales + 12 + sb * 24, 12); + utmp_1[3] = ((utmp_1[2] >> 4) & kmask2) | (((utmp_1[1] >> 6) & kmask3) << 4); + const uint32_t uaux_1 = utmp_1[1] & kmask1; + utmp_1[1] = (utmp_1[2] & kmask2) | (((utmp_1[0] >> 6) & kmask3) << 4); + utmp_1[2] = uaux_1; + utmp_1[0] &= kmask1; + + // Scales of first sub block in the sb loop + const __m128i mins_and_scales_0 = _mm_set_epi32(utmp_0[3], utmp_0[2], utmp_0[1], utmp_0[0]); + const __m256i scales_0 = _mm256_cvtepu8_epi16(_mm_unpacklo_epi8(mins_and_scales_0, mins_and_scales_0)); + + // Scales of second sub block in the sb loop + const __m128i mins_and_scales_1 = _mm_set_epi32(utmp_1[3], utmp_1[2], utmp_1[1], utmp_1[0]); + const __m256i scales_1 = _mm256_cvtepu8_epi16(_mm_unpacklo_epi8(mins_and_scales_1, mins_and_scales_1)); + + // Mins of first and second sub block of Q4_K block are arranged side by side + const __m256i mins_01 = _mm256_cvtepu8_epi16(_mm_unpacklo_epi8(_mm_shuffle_epi32(mins_and_scales_0, 78), _mm_shuffle_epi32(mins_and_scales_1, 78))); + + const __m256i scale_0145_0 = _mm256_shuffle_epi32(scales_0, 68); + const __m256i scale_2367_0 = _mm256_shuffle_epi32(scales_0, 238); + + const __m256i scale_0145_1 = _mm256_shuffle_epi32(scales_1, 68); + const __m256i scale_2367_1 = _mm256_shuffle_epi32(scales_1, 238); + + // Load the four block_q8_k quantized values interleaved with each other in chunks of eight bytes - A0,A1,A2,A3 + // Loaded as set of 128 bit vectors and repeated into a 256 bit vector + __m256i lhs_mat_0123_00 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 256 * sb))); + __m256i lhs_mat_01_00 = _mm256_permute2f128_si256(lhs_mat_0123_00, lhs_mat_0123_00, 0); + __m256i lhs_mat_23_00 = _mm256_permute2f128_si256(lhs_mat_0123_00, lhs_mat_0123_00, 17); + __m256i lhs_mat_0123_01 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 32 + 256 * sb))); + __m256i lhs_mat_01_01 = _mm256_permute2f128_si256(lhs_mat_0123_01, lhs_mat_0123_01, 0); + __m256i lhs_mat_23_01 = _mm256_permute2f128_si256(lhs_mat_0123_01, lhs_mat_0123_01, 17); + __m256i lhs_mat_0123_02 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 64 + 256 * sb))); + __m256i lhs_mat_01_02 = _mm256_permute2f128_si256(lhs_mat_0123_02, lhs_mat_0123_02, 0); + __m256i lhs_mat_23_02 = _mm256_permute2f128_si256(lhs_mat_0123_02, lhs_mat_0123_02, 17); + __m256i lhs_mat_0123_03 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 96 + 256 * sb))); + __m256i lhs_mat_01_03 = _mm256_permute2f128_si256(lhs_mat_0123_03, lhs_mat_0123_03, 0); + __m256i lhs_mat_23_03 = _mm256_permute2f128_si256(lhs_mat_0123_03, lhs_mat_0123_03, 17); + __m256i lhs_mat_0123_10 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 128 + 256 * sb))); + __m256i lhs_mat_01_10 = _mm256_permute2f128_si256(lhs_mat_0123_10, lhs_mat_0123_10, 0); + __m256i lhs_mat_23_10 = _mm256_permute2f128_si256(lhs_mat_0123_10, lhs_mat_0123_10, 17); + __m256i lhs_mat_0123_11 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 160 + 256 * sb))); + __m256i lhs_mat_01_11 = _mm256_permute2f128_si256(lhs_mat_0123_11, lhs_mat_0123_11, 0); + __m256i lhs_mat_23_11 = _mm256_permute2f128_si256(lhs_mat_0123_11, lhs_mat_0123_11, 17); + __m256i lhs_mat_0123_12 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 192 + 256 * sb))); + __m256i lhs_mat_01_12 = _mm256_permute2f128_si256(lhs_mat_0123_12, lhs_mat_0123_12, 0); + __m256i lhs_mat_23_12 = _mm256_permute2f128_si256(lhs_mat_0123_12, lhs_mat_0123_12, 17); + __m256i lhs_mat_0123_13 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 224 + 256 * sb))); + __m256i lhs_mat_01_13 = _mm256_permute2f128_si256(lhs_mat_0123_13, lhs_mat_0123_13, 0); + __m256i lhs_mat_23_13 = _mm256_permute2f128_si256(lhs_mat_0123_13, lhs_mat_0123_13, 17); + + // Bsums are loaded - four bsums are loaded (for two sub blocks) for the different Q8_K blocks + __m256i lhs_bsums_0123_01 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].bsums + 16 * sb))); + __m256i lhs_bsums_hsum_0123_01 = _mm256_castsi128_si256(_mm_hadd_epi16(_mm256_castsi256_si128(lhs_bsums_0123_01), _mm256_extractf128_si256(lhs_bsums_0123_01, 1))); + lhs_bsums_hsum_0123_01 = _mm256_permute2x128_si256(lhs_bsums_hsum_0123_01, lhs_bsums_hsum_0123_01, 0); + + // Shuffle pattern one - left side input + const __m256i lhs_mat_01_00_sp1 = _mm256_shuffle_epi32(lhs_mat_01_00, 160); //A00(0-3) A00(0-3) A01(0-3) A01(0-3) A00(0-3) A00(0-3) A01(0-3) A01(0-3) + const __m256i lhs_mat_23_00_sp1 = _mm256_shuffle_epi32(lhs_mat_23_00, 160); //A02(0-3) A03(0-3) A02(0-3) A03(0-3) A02(0-3) A03(0-3) A02(0-3) A03(0-3) + + const __m256i lhs_mat_01_01_sp1 = _mm256_shuffle_epi32(lhs_mat_01_01, 160); //A00(8-11) A00(8-11) A01(8-11) A01(8-11) A00(8-11) A00(8-11) A01(8-11) A01(8-11) + const __m256i lhs_mat_23_01_sp1 = _mm256_shuffle_epi32(lhs_mat_23_01, 160); //A02(8-11) A03(8-11) A02(8-11) A03(8-11) A02(8-11) A03(8-11) A02(8-11) A03(8-11) + + const __m256i lhs_mat_01_02_sp1 = _mm256_shuffle_epi32(lhs_mat_01_02, 160); //A00(16-19) A00(16-19) A01(16-19) A01(16-19) A00(16-19) A00(16-19) A01(16-19) A01(16-19) + const __m256i lhs_mat_23_02_sp1 = _mm256_shuffle_epi32(lhs_mat_23_02, 160); //A02(16-19) A03(16-19) A02(16-19) A03(16-19) A02(16-19) A03(16-19) A02(16-19) A03(16-19) + + const __m256i lhs_mat_01_03_sp1 = _mm256_shuffle_epi32(lhs_mat_01_03, 160); //A00(24-27) A00(24-27) A01(24-27) A01(24-27) A00(24-27) A00(24-27) A01(24-27) A01(24-27) + const __m256i lhs_mat_23_03_sp1 = _mm256_shuffle_epi32(lhs_mat_23_03, 160); //A02(24-27) A03(24-27) A02(24-27) A03(24-27) A02(24-27) A03(24-27) A02(24-27) A03(24-27) + + const __m256i lhs_mat_01_10_sp1 = _mm256_shuffle_epi32(lhs_mat_01_10, 160); //A10(0-3) A10(0-3) A11(0-3) A11(0-3) A10(0-3) A10(0-3) A11(0-3) A11(0-3) + const __m256i lhs_mat_23_10_sp1 = _mm256_shuffle_epi32(lhs_mat_23_10, 160); //A12(0-3) A13(0-3) A12(0-3) A13(0-3) A12(0-3) A13(0-3) A12(0-3) A13(0-3) + + const __m256i lhs_mat_01_11_sp1 = _mm256_shuffle_epi32(lhs_mat_01_11, 160); //A10(8-11) A10(8-11) A11(8-11) A11(8-11) A10(8-11) A10(8-11) A11(8-11) A11(8-11) + const __m256i lhs_mat_23_11_sp1 = _mm256_shuffle_epi32(lhs_mat_23_11, 160); //A12(8-11) A13(8-11) A12(8-11) A13(8-11) A12(8-11) A13(8-11) A12(8-11) A13(8-11) + + const __m256i lhs_mat_01_12_sp1 = _mm256_shuffle_epi32(lhs_mat_01_12, 160); //A10(16-19) A10(16-19) A11(16-19) A11(16-19) A10(16-19) A10(16-19) A11(16-19) A11(16-19) + const __m256i lhs_mat_23_12_sp1 = _mm256_shuffle_epi32(lhs_mat_23_12, 160); //A12(16-19) A13(16-19) A12(16-19) A13(16-19) A12(16-19) A13(16-19) A12(16-19) A13(16-19) + + const __m256i lhs_mat_01_13_sp1 = _mm256_shuffle_epi32(lhs_mat_01_13, 160); //A10(24-27) A10(24-27) A11(24-27) A11(24-27) A10(24-27) A10(24-27) A11(24-27) A11(24-27) + const __m256i lhs_mat_23_13_sp1 = _mm256_shuffle_epi32(lhs_mat_23_13, 160); //A12(24-27) A13(24-27) A12(24-27) A13(24-27) A12(24-27) A13(24-27) A12(24-27) A13(24-27) + + // Shuffle pattern two- left side input + const __m256i lhs_mat_01_00_sp2 = _mm256_shuffle_epi32(lhs_mat_01_00, 245); //A00(4-7) A00(4-7) A01(4-7) A01(4-7) A00(4-7) A00(4-7) A01(4-7) A01(4-7) + const __m256i lhs_mat_23_00_sp2 = _mm256_shuffle_epi32(lhs_mat_23_00, 245); //A02(4-7) A03(4-7) A02(4-7) A03(4-7) A02(4-7) A03(4-7) A02(4-7) A03(4-7) + + const __m256i lhs_mat_01_01_sp2 = _mm256_shuffle_epi32(lhs_mat_01_01, 245); //A00(12-15) A00(12-15) A01(12-15) A01(12-15) A00(12-15) A00(12-15) A01(12-15) A01(12-15) + const __m256i lhs_mat_23_01_sp2 = _mm256_shuffle_epi32(lhs_mat_23_01, 245); //A02(12-15) A03(12-15) A02(12-15) A03(12-15) A02(12-15) A03(12-15) A02(12-15) A03(12-15) + + const __m256i lhs_mat_01_02_sp2 = _mm256_shuffle_epi32(lhs_mat_01_02, 245); //A00(20-23) A00(20-23) A01(20-23) A01(20-23) A00(20-23) A00(20-23) A01(20-23) A01(20-23) + const __m256i lhs_mat_23_02_sp2 = _mm256_shuffle_epi32(lhs_mat_23_02, 245); //A02(20-23) A03(20-23) A02(20-23) A03(20-23) A02(20-23) A03(20-23) A02(20-23) A03(20-23) + + const __m256i lhs_mat_01_03_sp2 = _mm256_shuffle_epi32(lhs_mat_01_03, 245); //A00(28-31) A00(28-31) A01(28-31) A01(28-31) A00(28-31) A00(28-31) A01(28-31) A01(28-31) + const __m256i lhs_mat_23_03_sp2 = _mm256_shuffle_epi32(lhs_mat_23_03, 245); //A02(28-31) A03(28-31) A02(28-31) A03(28-31) A02(28-31) A03(28-31) A02(28-31) A03(28-31) + + const __m256i lhs_mat_01_10_sp2 = _mm256_shuffle_epi32(lhs_mat_01_10, 245); //A10(4-7) A10(4-7) A11(4-7) A11(4-7) A10(4-7) A10(4-7) A11(4-7) A11(4-7) + const __m256i lhs_mat_23_10_sp2 = _mm256_shuffle_epi32(lhs_mat_23_10, 245); //A12(4-7) A13(4-7) A12(4-7) A13(4-7) A12(4-7) A13(4-7) A12(4-7) A13(4-7) + + const __m256i lhs_mat_01_11_sp2 = _mm256_shuffle_epi32(lhs_mat_01_11, 245); //A10(12-15) A10(12-15) A11(12-15) A11(12-15) A10(12-15) A10(12-15) A11(12-15) A11(12-15) + const __m256i lhs_mat_23_11_sp2 = _mm256_shuffle_epi32(lhs_mat_23_11, 245); //A12(12-15) A13(12-15) A12(12-15) A13(12-15) A12(12-15) A13(12-15) A12(12-15) A13(12-15) + + const __m256i lhs_mat_01_12_sp2 = _mm256_shuffle_epi32(lhs_mat_01_12, 245); //A10(20-23) A10(20-23) A11(20-23) A11(20-23) A10(20-23) A10(20-23) A11(20-23) A11(20-23) + const __m256i lhs_mat_23_12_sp2 = _mm256_shuffle_epi32(lhs_mat_23_12, 245); //A12(20-23) A13(20-23) A12(20-23) A13(20-23) A12(20-23) A13(20-23) A12(20-23) A13(20-23) + + const __m256i lhs_mat_01_13_sp2 = _mm256_shuffle_epi32(lhs_mat_01_13, 245); //A10(28-31) A10(28-31) A11(28-31) A11(28-31) A10(28-31) A10(28-31) A11(28-31) A11(28-31) + const __m256i lhs_mat_23_13_sp2 = _mm256_shuffle_epi32(lhs_mat_23_13, 245); //A12(28-31) A13(28-31) A12(28-31) A13(28-31) A12(28-31) A13(28-31) A12(28-31) A13(28-31) + + // The values arranged in shuffle patterns are operated with dot product operation within 32 bit lane i.e corresponding bytes and multiplied and added into 32 bit integers within 32 bit lane + __m256i iacc_mat_00_0_sp1 = _mm256_add_epi16(_mm256_add_epi16(_mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_03_sp1, lhs_mat_01_03_sp1), _mm256_maddubs_epi16(rhs_mat_0145_02_sp1, lhs_mat_01_02_sp1)), _mm256_maddubs_epi16(rhs_mat_0145_01_sp1, lhs_mat_01_01_sp1)), _mm256_maddubs_epi16(rhs_mat_0145_00_sp1, lhs_mat_01_00_sp1)); + __m256i iacc_mat_01_0_sp1 = _mm256_add_epi16(_mm256_add_epi16(_mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_03_sp1, lhs_mat_01_03_sp1), _mm256_maddubs_epi16(rhs_mat_2367_02_sp1, lhs_mat_01_02_sp1)), _mm256_maddubs_epi16(rhs_mat_2367_01_sp1, lhs_mat_01_01_sp1)), _mm256_maddubs_epi16(rhs_mat_2367_00_sp1, lhs_mat_01_00_sp1)); + __m256i iacc_mat_10_0_sp1 = _mm256_add_epi16(_mm256_add_epi16(_mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_03_sp1, lhs_mat_23_03_sp1), _mm256_maddubs_epi16(rhs_mat_0145_02_sp1, lhs_mat_23_02_sp1)), _mm256_maddubs_epi16(rhs_mat_0145_01_sp1, lhs_mat_23_01_sp1)), _mm256_maddubs_epi16(rhs_mat_0145_00_sp1, lhs_mat_23_00_sp1)); + __m256i iacc_mat_11_0_sp1 = _mm256_add_epi16(_mm256_add_epi16(_mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_03_sp1, lhs_mat_23_03_sp1), _mm256_maddubs_epi16(rhs_mat_2367_02_sp1, lhs_mat_23_02_sp1)), _mm256_maddubs_epi16(rhs_mat_2367_01_sp1, lhs_mat_23_01_sp1)), _mm256_maddubs_epi16(rhs_mat_2367_00_sp1, lhs_mat_23_00_sp1)); + __m256i iacc_mat_00_1_sp1 = _mm256_add_epi16(_mm256_add_epi16(_mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_13_sp1, lhs_mat_01_13_sp1), _mm256_maddubs_epi16(rhs_mat_0145_12_sp1, lhs_mat_01_12_sp1)), _mm256_maddubs_epi16(rhs_mat_0145_11_sp1, lhs_mat_01_11_sp1)), _mm256_maddubs_epi16(rhs_mat_0145_10_sp1, lhs_mat_01_10_sp1)); + __m256i iacc_mat_01_1_sp1 = _mm256_add_epi16(_mm256_add_epi16(_mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_13_sp1, lhs_mat_01_13_sp1), _mm256_maddubs_epi16(rhs_mat_2367_12_sp1, lhs_mat_01_12_sp1)), _mm256_maddubs_epi16(rhs_mat_2367_11_sp1, lhs_mat_01_11_sp1)), _mm256_maddubs_epi16(rhs_mat_2367_10_sp1, lhs_mat_01_10_sp1)); + __m256i iacc_mat_10_1_sp1 = _mm256_add_epi16(_mm256_add_epi16(_mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_13_sp1, lhs_mat_23_13_sp1), _mm256_maddubs_epi16(rhs_mat_0145_12_sp1, lhs_mat_23_12_sp1)), _mm256_maddubs_epi16(rhs_mat_0145_11_sp1, lhs_mat_23_11_sp1)), _mm256_maddubs_epi16(rhs_mat_0145_10_sp1, lhs_mat_23_10_sp1)); + __m256i iacc_mat_11_1_sp1 = _mm256_add_epi16(_mm256_add_epi16(_mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_13_sp1, lhs_mat_23_13_sp1), _mm256_maddubs_epi16(rhs_mat_2367_12_sp1, lhs_mat_23_12_sp1)), _mm256_maddubs_epi16(rhs_mat_2367_11_sp1, lhs_mat_23_11_sp1)), _mm256_maddubs_epi16(rhs_mat_2367_10_sp1, lhs_mat_23_10_sp1)); + + __m256i iacc_mat_00_0_sp2 = _mm256_add_epi16(_mm256_add_epi16(_mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_03_sp2, lhs_mat_01_03_sp2), _mm256_maddubs_epi16(rhs_mat_0145_02_sp2, lhs_mat_01_02_sp2)), _mm256_maddubs_epi16(rhs_mat_0145_01_sp2, lhs_mat_01_01_sp2)), _mm256_maddubs_epi16(rhs_mat_0145_00_sp2, lhs_mat_01_00_sp2)); + __m256i iacc_mat_01_0_sp2 = _mm256_add_epi16(_mm256_add_epi16(_mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_03_sp2, lhs_mat_01_03_sp2), _mm256_maddubs_epi16(rhs_mat_2367_02_sp2, lhs_mat_01_02_sp2)), _mm256_maddubs_epi16(rhs_mat_2367_01_sp2, lhs_mat_01_01_sp2)), _mm256_maddubs_epi16(rhs_mat_2367_00_sp2, lhs_mat_01_00_sp2)); + __m256i iacc_mat_10_0_sp2 = _mm256_add_epi16(_mm256_add_epi16(_mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_03_sp2, lhs_mat_23_03_sp2), _mm256_maddubs_epi16(rhs_mat_0145_02_sp2, lhs_mat_23_02_sp2)), _mm256_maddubs_epi16(rhs_mat_0145_01_sp2, lhs_mat_23_01_sp2)), _mm256_maddubs_epi16(rhs_mat_0145_00_sp2, lhs_mat_23_00_sp2)); + __m256i iacc_mat_11_0_sp2 = _mm256_add_epi16(_mm256_add_epi16(_mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_03_sp2, lhs_mat_23_03_sp2), _mm256_maddubs_epi16(rhs_mat_2367_02_sp2, lhs_mat_23_02_sp2)), _mm256_maddubs_epi16(rhs_mat_2367_01_sp2, lhs_mat_23_01_sp2)), _mm256_maddubs_epi16(rhs_mat_2367_00_sp2, lhs_mat_23_00_sp2)); + __m256i iacc_mat_00_1_sp2 = _mm256_add_epi16(_mm256_add_epi16(_mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_13_sp2, lhs_mat_01_13_sp2), _mm256_maddubs_epi16(rhs_mat_0145_12_sp2, lhs_mat_01_12_sp2)), _mm256_maddubs_epi16(rhs_mat_0145_11_sp2, lhs_mat_01_11_sp2)), _mm256_maddubs_epi16(rhs_mat_0145_10_sp2, lhs_mat_01_10_sp2)); + __m256i iacc_mat_01_1_sp2 = _mm256_add_epi16(_mm256_add_epi16(_mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_13_sp2, lhs_mat_01_13_sp2), _mm256_maddubs_epi16(rhs_mat_2367_12_sp2, lhs_mat_01_12_sp2)), _mm256_maddubs_epi16(rhs_mat_2367_11_sp2, lhs_mat_01_11_sp2)), _mm256_maddubs_epi16(rhs_mat_2367_10_sp2, lhs_mat_01_10_sp2)); + __m256i iacc_mat_10_1_sp2 = _mm256_add_epi16(_mm256_add_epi16(_mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_13_sp2, lhs_mat_23_13_sp2), _mm256_maddubs_epi16(rhs_mat_0145_12_sp2, lhs_mat_23_12_sp2)), _mm256_maddubs_epi16(rhs_mat_0145_11_sp2, lhs_mat_23_11_sp2)), _mm256_maddubs_epi16(rhs_mat_0145_10_sp2, lhs_mat_23_10_sp2)); + __m256i iacc_mat_11_1_sp2 = _mm256_add_epi16(_mm256_add_epi16(_mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_13_sp2, lhs_mat_23_13_sp2), _mm256_maddubs_epi16(rhs_mat_2367_12_sp2, lhs_mat_23_12_sp2)), _mm256_maddubs_epi16(rhs_mat_2367_11_sp2, lhs_mat_23_11_sp2)), _mm256_maddubs_epi16(rhs_mat_2367_10_sp2, lhs_mat_23_10_sp2)); + + // Output of both shuffle patterns are added in order to sum dot product outputs of all 32 values in block + __m256i iacc_mat_00_0 = _mm256_add_epi16(iacc_mat_00_0_sp1, iacc_mat_00_0_sp2); + __m256i iacc_mat_01_0 = _mm256_add_epi16(iacc_mat_01_0_sp1, iacc_mat_01_0_sp2); + __m256i iacc_mat_10_0 = _mm256_add_epi16(iacc_mat_10_0_sp1, iacc_mat_10_0_sp2); + __m256i iacc_mat_11_0 = _mm256_add_epi16(iacc_mat_11_0_sp1, iacc_mat_11_0_sp2); + + __m256i iacc_mat_00_1 = _mm256_add_epi16(iacc_mat_00_1_sp1, iacc_mat_00_1_sp2); + __m256i iacc_mat_01_1 = _mm256_add_epi16(iacc_mat_01_1_sp1, iacc_mat_01_1_sp2); + __m256i iacc_mat_10_1 = _mm256_add_epi16(iacc_mat_10_1_sp1, iacc_mat_10_1_sp2); + __m256i iacc_mat_11_1 = _mm256_add_epi16(iacc_mat_11_1_sp1, iacc_mat_11_1_sp2); + + // Output of both shuffle patterns are added in order to sum dot product outputs of all 32 values in block + iacc_mat_00_0 = _mm256_madd_epi16(iacc_mat_00_0, scale_0145_0); + iacc_mat_01_0 = _mm256_madd_epi16(iacc_mat_01_0, scale_2367_0); + iacc_mat_10_0 = _mm256_madd_epi16(iacc_mat_10_0, scale_0145_0); + iacc_mat_11_0 = _mm256_madd_epi16(iacc_mat_11_0, scale_2367_0); + + iacc_mat_00_1 = _mm256_madd_epi16(iacc_mat_00_1, scale_0145_1); + iacc_mat_01_1 = _mm256_madd_epi16(iacc_mat_01_1, scale_2367_1); + iacc_mat_10_1 = _mm256_madd_epi16(iacc_mat_10_1, scale_0145_1); + iacc_mat_11_1 = _mm256_madd_epi16(iacc_mat_11_1, scale_2367_1); + + // Straighten out to make 4 row vectors (4 for each sub block which are accumulated together in the next step) + __m256i iacc_row_0_0 = _mm256_blend_epi32(iacc_mat_00_0, _mm256_shuffle_epi32(iacc_mat_01_0, 78), 204); + __m256i iacc_row_1_0 = _mm256_blend_epi32(_mm256_shuffle_epi32(iacc_mat_00_0, 78), iacc_mat_01_0, 204); + __m256i iacc_row_2_0 = _mm256_blend_epi32(iacc_mat_10_0, _mm256_shuffle_epi32(iacc_mat_11_0, 78), 204); + __m256i iacc_row_3_0 = _mm256_blend_epi32(_mm256_shuffle_epi32(iacc_mat_10_0, 78), iacc_mat_11_0, 204); + __m256i iacc_row_0_1 = _mm256_blend_epi32(iacc_mat_00_1, _mm256_shuffle_epi32(iacc_mat_01_1, 78), 204); + __m256i iacc_row_1_1 = _mm256_blend_epi32(_mm256_shuffle_epi32(iacc_mat_00_1, 78), iacc_mat_01_1, 204); + __m256i iacc_row_2_1 = _mm256_blend_epi32(iacc_mat_10_1, _mm256_shuffle_epi32(iacc_mat_11_1, 78), 204); + __m256i iacc_row_3_1 = _mm256_blend_epi32(_mm256_shuffle_epi32(iacc_mat_10_1, 78), iacc_mat_11_1, 204); + + __m256i iacc_row_0 = _mm256_add_epi32(iacc_row_0_0, iacc_row_0_1); + __m256i iacc_row_1 = _mm256_add_epi32(iacc_row_1_0, iacc_row_1_1); + __m256i iacc_row_2 = _mm256_add_epi32(iacc_row_2_0, iacc_row_2_1); + __m256i iacc_row_3 = _mm256_add_epi32(iacc_row_3_0, iacc_row_3_1); + + // Load the scale(d) values for all the 4 Q8_k blocks and repeat it across lanes + const __m128 row_scale_f32_sse = _mm_load_ps(a_ptr[b].d); + const __m256 row_scale_f32 = _mm256_set_m128(row_scale_f32_sse, row_scale_f32_sse); //GGML_F32Cx8_REPEAT_LOAD(a_ptrs[rp][b].d, loadMask); + + // Multiply with appropiate scales and accumulate (for both d and dmin) below + acc_rows[0] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_0), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 0)), acc_rows[0]); + acc_rows[1] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_1), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 85)), acc_rows[1]); + acc_rows[2] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_2), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 170)), acc_rows[2]); + acc_rows[3] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_3), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 255)), acc_rows[3]); + + __m256i iacc_row_min_0 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_hsum_0123_01, 0), mins_01); + __m256i iacc_row_min_1 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_hsum_0123_01, 85), mins_01); + __m256i iacc_row_min_2 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_hsum_0123_01, 170), mins_01); + __m256i iacc_row_min_3 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_hsum_0123_01, 255), mins_01); + + acc_min_rows[0] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_min_0), _mm256_mul_ps(col_dmin_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 0)), acc_min_rows[0]); + acc_min_rows[1] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_min_1), _mm256_mul_ps(col_dmin_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 85)), acc_min_rows[1]); + acc_min_rows[2] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_min_2), _mm256_mul_ps(col_dmin_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 170)), acc_min_rows[2]); + acc_min_rows[3] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_min_3), _mm256_mul_ps(col_dmin_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 255)), acc_min_rows[3]); + } + } + + // Store the accumulated values + for (int i = 0; i < 4; i++) { + _mm256_storeu_ps((float * )(s + ((y * 4 + i) * bs + x * 8)), _mm256_sub_ps(acc_rows[i], acc_min_rows[i])); + } + } + } + +#else + + float sumf[4][8]; + float sum_minf[4][8]; + uint32_t utmp[32]; + int sumi1; + int sumi2; + int sumi; + + for (int y = 0; y < nr / 4; y++) { + const block_q8_Kx4 * a_ptr = (const block_q8_Kx4 *) vy + (y * nb); + for (int x = 0; x < nc / ncols_interleaved; x++) { + const block_q4_Kx8 * b_ptr = (const block_q4_Kx8 *) vx + (x * nb); + for (int m = 0; m < 4; m++) { + for (int j = 0; j < ncols_interleaved; j++) { + sumf[m][j] = 0.0; + sum_minf[m][j] = 0.0; + } + } + for (int l = 0; l < nb; l++) { + for (int sb = 0; sb < 8; sb++) { + memcpy(utmp + sb * 4, b_ptr[l].scales + sb * 12, 12); + utmp[sb * 4 + 3] = ((utmp[sb * 4 + 2] >> 4) & kmask2) | (((utmp[sb * 4 + 1] >> 6) & kmask3) << 4); + const uint32_t uaux_0 = utmp[sb * 4 + 1] & kmask1; + utmp[sb * 4 + 1] = (utmp[sb * 4 + 2] & kmask2) | (((utmp[sb * 4 + 0] >> 6) & kmask3) << 4); + utmp[sb * 4 + 2] = uaux_0; + utmp[sb * 4 + 0] &= kmask1; + } + for (int k = 0; k < (qk / (2 * blocklen)); k++) { + uint8_t *scales_0 = (uint8_t*) utmp + (k / 4) * 32; + uint8_t *scales_1 = (uint8_t*) utmp + (k / 4) * 32 + 16; + for (int m = 0; m < 4; m++) { + for (int j = 0; j < ncols_interleaved; j++) { + sumi1 = 0; + sumi2 = 0; + sumi = 0; + for (int i = 0; i < blocklen; ++i) { + const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF); + const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4); + sumi1 = (v0 * a_ptr[l].qs[(k >> 2) * 256 + (k % 4) * 4 * blocklen + m * blocklen + i]); + sumi2 = (v1 * a_ptr[l].qs[(k >> 2) * 256 + (k % 4) * 4 * blocklen + m * blocklen + i + 128]); + sumi1 = sumi1 * scales_0[j]; + sumi2 = sumi2 * scales_1[j]; + sumi += sumi1 + sumi2; + } + sumf[m][j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d[m]; + } + } + } + for (int sb = 0; sb < 8; sb++) { + uint8_t *mins = (uint8_t*) utmp + 8 + sb * 16; + for(int m = 0; m < 4; m++) { + const int16_t *bsums = a_ptr[l].bsums + (sb * 8) + (m * 4) - ((sb % 2) * 6); + for(int j = 0; j < ncols_interleaved; j++) { + sum_minf[m][j] += mins[j] * (bsums[0] + bsums[1]) * GGML_FP16_TO_FP32(b_ptr[l].dmin[j]) * a_ptr[l].d[m]; + } + } + } + } + for (int m = 0; m < 4; m++) { + for (int j = 0; j < ncols_interleaved; j++) { + s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j] - sum_minf[m][j]; + } + } + } + } +#endif +} + static void ggml_gemm_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) { const int qk = QK8_0; const int nb = n / qk; @@ -3660,6 +5008,82 @@ static block_q4_0x8 make_block_q4_0x8(block_q4_0 * in, unsigned int blck_size_in return out; } +static block_q4_Kx8 make_block_q4_Kx8(block_q4_K * in, unsigned int blck_size_interleave) { + block_q4_Kx8 out; + //Delta(scale) and dmin values of the eight Q4_K structures are copied onto the output interleaved structure + for (int i = 0; i < 8; i++) { + out.d[i] = in[i].GGML_COMMON_AGGR_U.GGML_COMMON_AGGR_S.d; + } + + for (int i = 0; i < 8; i++) { + out.dmin[i] = in[i].GGML_COMMON_AGGR_U.GGML_COMMON_AGGR_S.dmin; + } + + const int end = QK_K * 4 / blck_size_interleave; + + // Interleave Q4_K quants by taking 8 bytes at a time + for (int i = 0; i < end; ++i) { + int src_id = i % 8; + int src_offset = (i / 8) * blck_size_interleave; + int dst_offset = i * blck_size_interleave; + + uint64_t elems; + memcpy(&elems, &in[src_id].qs[src_offset], sizeof(uint64_t)); + memcpy(&out.qs[dst_offset], &elems, sizeof(uint64_t)); + } + + // The below logic is designed so as to unpack and rearrange scales and mins values in Q4_K + // Currently the Q4_K structure has 8 scales and 8 mins packed in 12 bytes ( 6 bits for each value) + // The output Q4_Kx8 structure has 96 bytes + // Every 12 byte is packed such that it contains scales and mins for corresponding sub blocks from Q4_K structure + // For eg - First 12 bytes contains 8 scales and 8 mins - each of first sub block from different Q4_K structures + uint8_t s[8], m[8]; + + for (int i = 0; i < 4; i++) { + for (int j = 0; j < 8; j++) { + s[j] = in[j].scales[i] & 63; + m[j] = in[j].scales[i + 4] & 63; + } + + out.scales[i * 12] = (s[0] & 63) + ((s[4] & 48) << 2); + out.scales[i * 12 + 1] = (s[1] & 63) + ((s[5] & 48) << 2); + out.scales[i * 12 + 2] = (s[2] & 63) + ((s[6] & 48) << 2); + out.scales[i * 12 + 3] = (s[3] & 63) + ((s[7] & 48) << 2); + out.scales[i * 12 + 4] = (m[0] & 63) + ((m[4] & 48) << 2); + out.scales[i * 12 + 5] = (m[1] & 63) + ((m[5] & 48) << 2); + out.scales[i * 12 + 6] = (m[2] & 63) + ((m[6] & 48) << 2); + out.scales[i * 12 + 7] = (m[3] & 63) + ((m[7] & 48) << 2); + out.scales[i * 12 + 8] = (s[4] & 15) + ((m[4] & 15) << 4); + out.scales[i * 12 + 9] = (s[5] & 15) + ((m[5] & 15) << 4); + out.scales[i * 12 + 10] = (s[6] & 15) + ((m[6] & 15) << 4); + out.scales[i * 12 + 11] = (s[7] & 15) + ((m[7] & 15) << 4); + + } + + for (int i = 0; i < 4; i++) { + for (int j = 0; j < 8; j++) { + s[j] = ((in[j].scales[i] & 192) >> 2) | (in[j].scales[i+8] & 15); + m[j] = ((in[j].scales[i + 4] & 192) >> 2) | ((in[j].scales[i+8] & 240) >> 4); + } + + out.scales[i * 12 + 48] = (s[0] & 63) + ((s[4] & 48) << 2); + out.scales[i * 12 + 49] = (s[1] & 63) + ((s[5] & 48) << 2); + out.scales[i * 12 + 50] = (s[2] & 63) + ((s[6] & 48) << 2); + out.scales[i * 12 + 51] = (s[3] & 63) + ((s[7] & 48) << 2); + out.scales[i * 12 + 52] = (m[0] & 63) + ((m[4] & 48) << 2); + out.scales[i * 12 + 53] = (m[1] & 63) + ((m[5] & 48) << 2); + out.scales[i * 12 + 54] = (m[2] & 63) + ((m[6] & 48) << 2); + out.scales[i * 12 + 55] = (m[3] & 63) + ((m[7] & 48) << 2); + out.scales[i * 12 + 56] = (s[4] & 15) + ((m[4] & 15) << 4); + out.scales[i * 12 + 57] = (s[5] & 15) + ((m[5] & 15) << 4); + out.scales[i * 12 + 58] = (s[6] & 15) + ((m[6] & 15) << 4); + out.scales[i * 12 + 59] = (s[7] & 15) + ((m[7] & 15) << 4); + + } + + return out; +} + static int repack_q4_0_to_q4_0_4_bl(struct ggml_tensor * t, int interleave_block, const void * GGML_RESTRICT data, size_t data_size) { GGML_ASSERT(t->type == GGML_TYPE_Q4_0); GGML_ASSERT(interleave_block == 4 || interleave_block == 8); @@ -3690,6 +5114,36 @@ static int repack_q4_0_to_q4_0_4_bl(struct ggml_tensor * t, int interleave_block GGML_UNUSED(data_size); } +static int repack_q4_K_to_q4_K_8_bl(struct ggml_tensor * t, int interleave_block, const void * GGML_RESTRICT data, size_t data_size) { + GGML_ASSERT(t->type == GGML_TYPE_Q4_K); + GGML_ASSERT(interleave_block == 8); + constexpr int nrows_interleaved = 8; + + block_q4_Kx8 * dst = (block_q4_Kx8*)t->data; + const block_q4_K * src = (const block_q4_K*) data; + block_q4_K dst_tmp[8]; + int nrow = ggml_nrows(t); + int nblocks = t->ne[0] / QK_K; + + GGML_ASSERT(data_size == nrow * nblocks * sizeof(block_q4_K)); + + if (t->ne[1] % nrows_interleaved != 0 || t->ne[0] % 8 != 0) { + return -1; + } + + for (int b = 0; b < nrow; b += nrows_interleaved) { + for (int64_t x = 0; x < nblocks; x++) { + for (int i = 0; i < nrows_interleaved; i++ ) { + dst_tmp[i] = src[x + i * nblocks]; + } + *dst++ = make_block_q4_Kx8(dst_tmp, interleave_block); + } + src += nrows_interleaved * nblocks; + } + return 0; + + GGML_UNUSED(data_size); +} static int repack_q4_0_to_q4_0_8_bl(struct ggml_tensor * t, int interleave_block, const void * GGML_RESTRICT data, size_t data_size) { GGML_ASSERT(t->type == GGML_TYPE_Q4_0); @@ -3807,6 +5261,10 @@ template <> int repack(struct ggml_tensor * t, const void * da return repack_q4_0_to_q4_0_8_bl(t, 8, data, data_size); } +template <> int repack(struct ggml_tensor * t, const void * data, size_t data_size) { + return repack_q4_K_to_q4_K_8_bl(t, 8, data, data_size); +} + template <> int repack(struct ggml_tensor * t, const void * data, size_t data_size) { return repack_iq4_nl_to_iq4_nl_4_bl(t, 4, data, data_size); } @@ -3817,44 +5275,50 @@ template <> int repack(struct ggml_tensor * t, const void * //} // gemv -template +template void gemv(int, float *, size_t, const void *, const void *, int, int); -template <> void gemv(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) { +template <> void gemv(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) { ggml_gemv_q4_0_4x4_q8_0(n, s, bs, vx, vy, nr, nc); } -template <> void gemv(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) { +template <> void gemv(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) { ggml_gemv_q4_0_4x8_q8_0(n, s, bs, vx, vy, nr, nc); } -template <> void gemv(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) { +template <> void gemv(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) { ggml_gemv_q4_0_8x8_q8_0(n, s, bs, vx, vy, nr, nc); } -template <> -void gemv(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) { +template <> void gemv(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) { + ggml_gemv_q4_K_8x8_q8_K(n, s, bs, vx, vy, nr, nc); +} + +template <> void gemv(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) { ggml_gemv_iq4_nl_4x4_q8_0(n, s, bs, vx, vy, nr, nc); } // gemm -template +template void gemm(int, float *, size_t, const void *, const void *, int, int); -template <> void gemm(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) { +template <> void gemm(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) { ggml_gemm_q4_0_4x4_q8_0(n, s, bs, vx, vy, nr, nc); } -template <> void gemm(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) { +template <> void gemm(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) { ggml_gemm_q4_0_4x8_q8_0(n, s, bs, vx, vy, nr, nc); } -template <> void gemm(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) { +template <> void gemm(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) { ggml_gemm_q4_0_8x8_q8_0(n, s, bs, vx, vy, nr, nc); } -template <> -void gemm(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) { +template <> void gemm(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) { + ggml_gemm_q4_K_8x8_q8_K(n, s, bs, vx, vy, nr, nc); +} + +template <> void gemm(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) { ggml_gemm_iq4_nl_4x4_q8_0(n, s, bs, vx, vy, nr, nc); } @@ -3863,37 +5327,37 @@ class tensor_traits_base : public ggml::cpu::tensor_traits { virtual int repack(struct ggml_tensor * t, const void * data, size_t data_size) = 0; }; -template class tensor_traits : public tensor_traits_base { +template class tensor_traits : public tensor_traits_base { bool work_size(int /* n_threads */, const struct ggml_tensor * op, size_t & size) override { // not realy a GGML_TYPE_Q8_0 but same size. switch (op->op) { - case GGML_OP_MUL_MAT: - size = ggml_row_size(GGML_TYPE_Q8_0, ggml_nelements(op->src[1])); - return true; - case GGML_OP_MUL_MAT_ID: - size = ggml_row_size(GGML_TYPE_Q8_0, ggml_nelements(op->src[1])); - size = GGML_PAD(size, sizeof(int64_t)); // + padding for next bloc. - size += sizeof(int64_t) * (1+op->src[0]->ne[2]) * op->src[1]->ne[2]; - return true; - default: - // GGML_ABORT("fatal error"); - break; + case GGML_OP_MUL_MAT: + size = ggml_row_size(PARAM_TYPE, ggml_nelements(op->src[1])); + return true; + case GGML_OP_MUL_MAT_ID: + size = ggml_row_size(PARAM_TYPE, ggml_nelements(op->src[1])); + size = GGML_PAD(size, sizeof(int64_t)); // + padding for next bloc. + size += sizeof(int64_t) * (1+op->src[0]->ne[2]) * op->src[1]->ne[2]; + return true; + default: + // GGML_ABORT("fatal error"); + break; } return false; } bool compute_forward(struct ggml_compute_params * params, struct ggml_tensor * op) override { switch (op->op) { - case GGML_OP_MUL_MAT: - forward_mul_mat(params, op); - return true; - case GGML_OP_MUL_MAT_ID: - forward_mul_mat_id(params, op); - return true; - default: - // GGML_ABORT("fatal error"); - break; + case GGML_OP_MUL_MAT: + forward_mul_mat(params, op); + return true; + case GGML_OP_MUL_MAT_ID: + forward_mul_mat_id(params, op); + return true; + default: + // GGML_ABORT("fatal error"); + break; } return false; } @@ -3925,17 +5389,17 @@ template class tensor_ // GGML_ASSERT(ggml_n_dims(op->src[1]) == 2); char * wdata = static_cast(params->wdata); - const size_t nbw1 = ggml_row_size(GGML_TYPE_Q8_0, ne10); + const size_t nbw1 = ggml_row_size(PARAM_TYPE, ne10); assert(params->wsize >= nbw1 * ne11); - const ggml_from_float_t from_float = ggml_get_type_traits_cpu(GGML_TYPE_Q8_0)->from_float; + const ggml_from_float_t from_float = ggml_get_type_traits_cpu(PARAM_TYPE)->from_float; int64_t i11_processed = 0; for (int64_t i11 = ith * 4; i11 < ne11 - ne11 % 4; i11 += nth * 4) { - quantize_mat_q8_0((float *) ((char *) src1->data + i11 * nb11), (void *) (wdata + i11 * nbw1), 4, ne10, - INTER_SIZE); + ggml_quantize_mat_t((float *) ((char *) src1->data + i11 * nb11), (void *) (wdata + i11 * nbw1), 4, ne10); } + i11_processed = ne11 - ne11 % 4; for (int64_t i11 = i11_processed + ith; i11 < ne11; i11 += nth) { from_float((float *) ((char *) src1->data + i11 * nb11), (void *) (wdata + i11 * nbw1), ne10); @@ -3944,26 +5408,28 @@ template class tensor_ ggml_barrier(params->threadpool); const void * src1_wdata = params->wdata; - const size_t src1_col_stride = ggml_row_size(GGML_TYPE_Q8_0, ne10); + const size_t src1_col_stride = ggml_row_size(PARAM_TYPE, ne10); int64_t src0_start = (ith * ne01) / nth; int64_t src0_end = ((ith + 1) * ne01) / nth; src0_start = (src0_start % NB_COLS) ? src0_start + NB_COLS - (src0_start % NB_COLS) : src0_start; - src0_end = (src0_end % NB_COLS) ? src0_end + NB_COLS - (src0_end % NB_COLS) : src0_end; + src0_end = (src0_end % NB_COLS) ? src0_end + NB_COLS - (src0_end % NB_COLS) : src0_end; if (src0_start >= src0_end) { return; } // If there are more than three rows in src1, use gemm; otherwise, use gemv. if (ne11 > 3) { - gemm(ne00, (float *) ((char *) dst->data) + src0_start, ne01, - (const char *) src0->data + src0_start * nb01, - (const char *) src1_wdata, ne11 - ne11 % 4, src0_end - src0_start); + gemm(ne00, + (float *) ((char *) dst->data) + src0_start, ne01, + (const char *) src0->data + src0_start * nb01, + (const char *) src1_wdata, ne11 - ne11 % 4, src0_end - src0_start); } for (int iter = ne11 - ne11 % 4; iter < ne11; iter++) { - gemv(ne00, (float *) ((char *) dst->data + (iter * nb1)) + src0_start, ne01, - (const char *) src0->data + src0_start * nb01, - (const char *) src1_wdata + (src1_col_stride * iter), 1, - src0_end - src0_start); + gemv(ne00, + (float *) ((char *) dst->data + (iter * nb1)) + src0_start, ne01, + (const char *) src0->data + src0_start * nb01, + (const char *) src1_wdata + (src1_col_stride * iter), 1, + src0_end - src0_start); } } @@ -3978,7 +5444,7 @@ template class tensor_ const int ith = params->ith; const int nth = params->nth; - const ggml_from_float_t from_float = ggml_get_type_traits_cpu(GGML_TYPE_Q8_0)->from_float; + const ggml_from_float_t from_float = ggml_get_type_traits_cpu(PARAM_TYPE)->from_float; // we don't support permuted src0 or src1 GGML_ASSERT(nb00 == ggml_type_size(src0->type)); @@ -4000,7 +5466,7 @@ template class tensor_ const int n_ids = ids->ne[0]; // n_expert_used const int n_as = ne02; // n_expert - const size_t nbw1 = ggml_row_size(GGML_TYPE_Q8_0, ne10); + const size_t nbw1 = ggml_row_size(PARAM_TYPE, ne10); const size_t nbw2 = nbw1*ne11; const size_t nbw3 = nbw2*ne12; @@ -4012,12 +5478,13 @@ template class tensor_ GGML_ASSERT(params->wsize >= (GGML_PAD(nbw3, sizeof(int64_t)) + n_as * sizeof(int64_t) + n_as * ne12 * sizeof(mmid_row_mapping))); - auto wdata = (char *) params->wdata; - auto wdata_src1_end = (char *) wdata + GGML_PAD(nbw3, sizeof(int64_t)); - int64_t * matrix_row_counts = (int64_t *) (wdata_src1_end); // [n_as] + auto * wdata = (char *) params->wdata; + auto * wdata_src1_end = (char *) wdata + GGML_PAD(nbw3, sizeof(int64_t)); + auto * matrix_row_counts = (int64_t *) (wdata_src1_end); // [n_as] + struct mmid_row_mapping * matrix_rows = (struct mmid_row_mapping *) (matrix_row_counts + n_as); // [n_as][ne12] - // src1: float32 => block_q8_0 + // src1: float32 => param type for (int64_t i12 = 0; i12 < ne12; ++i12) { for (int64_t i11 = ith; i11 < ne11; i11 += nth) { from_float((float *)((char *) src1->data + i12 * nb12 + i11 * nb11), @@ -4056,34 +5523,37 @@ template class tensor_ continue; } - auto src0_cur = (const char *) src0->data + cur_a*nb02; + const auto * src0_cur = (const char *) src0->data + cur_a*nb02; //const int64_t nr0 = ne01; // src0 rows const int64_t nr1 = cne1; // src1 rows int64_t src0_cur_start = (ith * ne01) / nth; int64_t src0_cur_end = ((ith + 1) * ne01) / nth; - src0_cur_start = - (src0_cur_start % NB_COLS) ? src0_cur_start + NB_COLS - (src0_cur_start % NB_COLS) : src0_cur_start; - src0_cur_end = (src0_cur_end % NB_COLS) ? src0_cur_end + NB_COLS - (src0_cur_end % NB_COLS) : src0_cur_end; - if (src0_cur_start >= src0_cur_end) return; + src0_cur_start = (src0_cur_start % NB_COLS) ? src0_cur_start + NB_COLS - (src0_cur_start % NB_COLS) : src0_cur_start; + src0_cur_end = (src0_cur_end % NB_COLS) ? src0_cur_end + NB_COLS - (src0_cur_end % NB_COLS) : src0_cur_end; + + if (src0_cur_start >= src0_cur_end) { + return; + } for (int ir1 = 0; ir1 < nr1; ir1++) { struct mmid_row_mapping row_mapping = MMID_MATRIX_ROW(cur_a, ir1); - const int id = row_mapping.i1; // selected expert index - const int64_t i11 = id % ne11; - const int64_t i12 = row_mapping.i2; // row index in src1 + const int id = row_mapping.i1; // selected expert index - const int64_t i1 = id; // selected expert index - const int64_t i2 = i12; // row + const int64_t i11 = id % ne11; + const int64_t i12 = row_mapping.i2; // row index in src1 - auto src1_col = (const char *) wdata + (i11 * nbw1 + i12 * nbw2); + const int64_t i1 = id; // selected expert index + const int64_t i2 = i12; // row - gemv( - ne00, (float *)((char *) dst->data + (i1 * nb1 + i2 * nb2)) + src0_cur_start, - ne01, src0_cur + src0_cur_start * nb01, + const auto * src1_col = (const char *) wdata + (i11 * nbw1 + i12 * nbw2); + + gemv(ne00, + (float *)((char *) dst->data + (i1 * nb1 + i2 * nb2)) + src0_cur_start, ne01, + src0_cur + src0_cur_start * nb01, src1_col, 1, src0_cur_end - src0_cur_start); } } @@ -4098,12 +5568,13 @@ template class tensor_ }; // instance for Q4 -static const tensor_traits q4_0_4x4_q8_0; -static const tensor_traits q4_0_4x8_q8_0; -static const tensor_traits q4_0_8x8_q8_0; +static const tensor_traits q4_0_4x4_q8_0; +static const tensor_traits q4_0_4x8_q8_0; +static const tensor_traits q4_0_8x8_q8_0; +static const tensor_traits q4_K_8x8_q8_K; // instance for IQ4 -static const tensor_traits iq4_nl_4x4_q8_0; +static const tensor_traits iq4_nl_4x4_q8_0; } // namespace ggml::cpu::aarch64 @@ -4124,6 +5595,12 @@ static const ggml::cpu::tensor_traits * ggml_aarch64_get_optimal_repack_type(con return &ggml::cpu::aarch64::q4_0_4x4_q8_0; } } + } else if (cur->type == GGML_TYPE_Q4_K) { + if (ggml_cpu_has_avx2()) { + if (cur->ne[1] % 8 == 0) { + return &ggml::cpu::aarch64::q4_K_8x8_q8_K; + } + } } else if (cur->type == GGML_TYPE_IQ4_NL) { if (ggml_cpu_has_neon() && ggml_cpu_has_dotprod()) { if (cur->ne[1] % 4 == 0) { @@ -4135,10 +5612,11 @@ static const ggml::cpu::tensor_traits * ggml_aarch64_get_optimal_repack_type(con return nullptr; } -static void ggml_backend_cpu_aarch64_buffer_init_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) { +static enum ggml_status ggml_backend_cpu_aarch64_buffer_init_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) { tensor->extra = (void *) const_cast(ggml_aarch64_get_optimal_repack_type(tensor)); GGML_UNUSED(buffer); + return GGML_STATUS_SUCCESS; } static void ggml_backend_cpu_aarch64_buffer_set_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, diff --git a/ggml/src/ggml-cpu/ggml-cpu-impl.h b/ggml/src/ggml-cpu/ggml-cpu-impl.h index 9ddd972a5..e4af07635 100644 --- a/ggml/src/ggml-cpu/ggml-cpu-impl.h +++ b/ggml/src/ggml-cpu/ggml-cpu-impl.h @@ -4,13 +4,13 @@ #include "ggml.h" #include "ggml-impl.h" + #include // load `stdlib.h` before other headers to work around MinGW bug: https://sourceforge.net/p/mingw-w64/bugs/192/ //#include #include #include // memcpy #include // fabsf - #ifdef __cplusplus extern "C" { #endif @@ -59,34 +59,26 @@ struct ggml_compute_params { #endif #endif +#if defined(__s390x__) && defined(__VEC__) +#ifndef __VXE__ +#define __VXE__ +#endif +#ifndef __VXE2__ +#define __VXE2__ +#endif +#endif + #if defined(__ARM_FEATURE_SVE) -#include #include #endif -// 16-bit float -// on Arm, we use __fp16 -// on x86, we use uint16_t #if defined(__ARM_NEON) -// if YCM cannot find , make a symbolic link to it, for example: -// -// $ ln -sfn /Library/Developer/CommandLineTools/usr/lib/clang/13.1.6/include/arm_neon.h ./src/ -// -#include - +// ref: https://github.com/ggml-org/llama.cpp/pull/5404 #ifdef _MSC_VER - -typedef uint16_t ggml_fp16_internal_t; - #define ggml_vld1q_u32(w,x,y,z) { ((w) + ((uint64_t)(x) << 32)), ((y) + ((uint64_t)(z) << 32)) } - #else - -typedef __fp16 ggml_fp16_internal_t; - #define ggml_vld1q_u32(w,x,y,z) { (w), (x), (y), (z) } - #endif // _MSC_VER #if !defined(__aarch64__) @@ -331,8 +323,6 @@ inline static int32x4_t ggml_vdotq_s32(int32x4_t acc, int8x16_t a, int8x16_t b) #else #ifdef __POWER9_VECTOR__ #include -#undef bool -#define bool _Bool #else #if defined(_MSC_VER) || defined(__MINGW32__) #include @@ -359,6 +349,148 @@ inline static int32x4_t ggml_vdotq_s32(int32x4_t acc, int8x16_t a, int8x16_t b) #endif #endif +#if defined(__VXE__) || defined(__VXE2__) +#include + +#define vec_neg(a) (-(a)) // Vector Negate +#define vec_add(a, b) ((a) + (b)) // Vector Add +#define vec_sub(a, b) ((a) - (b)) // Vector Subtract +#define vec_mul(a, b) ((a) * (b)) // Vector Multiply +#define vec_div(a, b) ((a) / (b)) // Vector Divide +#define vec_sl(a, b) ((a) << (b)) // Vector Shift Left +#define vec_sra(a, b) ((a) >> (b)) // Vector Shift Right +#define vec_sr(a, b) ((a) >> (b)) // Vector Shift Right Algebraic +#define vec_slo(a, b) vec_slb(a, (b) << 64) // Vector Shift Left by Octet +#define vec_sro(a, b) vec_srb(a, (b) << 64) // Vector Shift Right by Octet + +#ifndef vec_and +#define vec_and(a, b) ((a) & (b)) // Vector AND +#endif + +#ifndef vec_or +#define vec_or(a, b) ((a) | (b)) // Vector OR +#endif + +#ifndef vec_xor +#define vec_xor(a, b) ((a) ^ (b)) // Vector XOR +#endif + +typedef signed char char8x16_t __attribute__((vector_size(16))); +typedef unsigned char uchar8x16_t __attribute__((vector_size(16))); + +typedef int8_t int8x16_t __attribute__((vector_size(16))); +typedef int16_t int16x8_t __attribute__((vector_size(16))); +typedef int32_t int32x4_t __attribute__((vector_size(16))); + +typedef uint8_t uint8x16_t __attribute__((vector_size(16))); +typedef uint16_t uint16x8_t __attribute__((vector_size(16))); +typedef uint32_t uint32x4_t __attribute__((vector_size(16))); + +typedef float float32x4_t __attribute__((vector_size(16))); +typedef double double64x2_t __attribute((vector_size(16))); + +typedef signed long long long64x2_t __attribute((vector_size(16))); +typedef unsigned long long ulong64x2_t __attribute__((vector_size(16))); + +typedef struct ggml_uint8x16x2_t { + uint8x16_t val[2]; +} ggml_uint8x16x2_t; + +inline static ggml_uint8x16x2_t ggml_vec_xl_u8x2(const uint8_t * ptr) { + ggml_uint8x16x2_t res; + + res.val[0] = vec_xl( 0, ptr); + res.val[1] = vec_xl(16, ptr); + + return res; +} + +typedef struct ggml_uint8x16x4_t { + uint8x16_t val[4]; +} ggml_uint8x16x4_t; + +inline static ggml_uint8x16x4_t ggml_vec_xl_u8x4(const uint8_t * ptr) { + ggml_uint8x16x4_t res; + + res.val[0] = vec_xl( 0, ptr); + res.val[1] = vec_xl(16, ptr); + res.val[2] = vec_xl(32, ptr); + res.val[3] = vec_xl(48, ptr); + + return res; +} + +typedef struct ggml_int8x16x4_t { + int8x16_t val[4]; +} ggml_int8x16x4_t; + +inline static ggml_int8x16x4_t ggml_vec_xl_s8x4(const int8_t * ptr) { + ggml_int8x16x4_t res; + + res.val[0] = vec_xl( 0, ptr); + res.val[1] = vec_xl(16, ptr); + res.val[2] = vec_xl(32, ptr); + res.val[3] = vec_xl(48, ptr); + + return res; +} + +typedef struct ggml_int16x8x2_t { + int16x8_t val[2]; +} ggml_int16x8x2_t; + +inline static ggml_int16x8x2_t ggml_vec_xl_s16x2(const int16_t * ptr) { + ggml_int16x8x2_t res; + + res.val[0] = vec_xl( 0, ptr); + res.val[1] = vec_xl(16, ptr); + + return res; +} + +/* + ! WARNING: Very slow. Use vec_perm if possible. Refer to iq4_xs + ! or iq4_nl for example implementation. +*/ +inline static int8x16_t ggml_vec_tbl(int8x16_t a, uint8x16_t b) { + int8x16_t res; + + res[ 0] = a[b[ 0]]; + res[ 1] = a[b[ 1]]; + res[ 2] = a[b[ 2]]; + res[ 3] = a[b[ 3]]; + res[ 4] = a[b[ 4]]; + res[ 5] = a[b[ 5]]; + res[ 6] = a[b[ 6]]; + res[ 7] = a[b[ 7]]; + res[ 8] = a[b[ 8]]; + res[ 9] = a[b[ 9]]; + res[10] = a[b[10]]; + res[11] = a[b[11]]; + res[12] = a[b[12]]; + res[13] = a[b[13]]; + res[14] = a[b[14]]; + res[15] = a[b[15]]; + + return res; +} + +inline static int16x8_t vec_padd_s16(int16x8_t a, int16x8_t b) { + const uchar8x16_t v_maske = { 0, 1, 4, 5, 8, 9, 12, 13, + 16, 17, 20, 21, 24, 25, 28, 29 }; + + const int16x8_t v_abo = vec_pack((int32x4_t)a, (int32x4_t)b); + const int16x8_t v_abe = vec_perm(a, b, v_maske); + return v_abo + v_abe; +} + +inline static int32x4_t ggml_vec_dot(int32x4_t acc, int8x16_t a, int8x16_t b) { + const int16x8_t p = vec_mule(a, b) + vec_mulo(a, b); + return acc + (vec_unpackh(p) + vec_unpackl(p)); +} + +#endif + #if defined(__loongarch_asx) /* float type data load instructions */ static __m128 __lsx_vreplfr2vr_s(const float val) { diff --git a/ggml/src/ggml-cpu/ggml-cpu-quants.c b/ggml/src/ggml-cpu/ggml-cpu-quants.c index 27ec14935..91a81bdc3 100644 --- a/ggml/src/ggml-cpu/ggml-cpu-quants.c +++ b/ggml/src/ggml-cpu/ggml-cpu-quants.c @@ -562,6 +562,41 @@ static __m256i lasx_packs_h(__m256i a, __m256i b) { return __lasx_xvpickev_b(tmp1, tmp); } +static inline __m256i lasx_madd_h_b(__m256i a, __m256i b) { + __m256i tmp1, tmp2; + tmp1 = __lasx_xvmulwev_h_b(a, b); + tmp2 = __lasx_xvmulwod_h_b(a, b); + return __lasx_xvadd_h(tmp1, tmp2); +} + +static inline __m256i lasx_xvrepl128vei_h(__m256i a, const unsigned int b) { + switch (b) { + case 0: return __lasx_xvrepl128vei_h(a, 0); + case 1: return __lasx_xvrepl128vei_h(a, 1); + case 2: return __lasx_xvrepl128vei_h(a, 2); + case 3: return __lasx_xvrepl128vei_h(a, 3); + case 4: return __lasx_xvrepl128vei_h(a, 4); + case 5: return __lasx_xvrepl128vei_h(a, 5); + case 6: return __lasx_xvrepl128vei_h(a, 6); + case 7: return __lasx_xvrepl128vei_h(a, 7); + default: __builtin_unreachable(); + } +} + +static inline __m256i lasx_xvandi_b_bit(__m256i a, const unsigned int b) { + switch (b) { + case 0: return __lasx_xvandi_b(a, 1 << 0); + case 1: return __lasx_xvandi_b(a, 1 << 1); + case 2: return __lasx_xvandi_b(a, 1 << 2); + case 3: return __lasx_xvandi_b(a, 1 << 3); + case 4: return __lasx_xvandi_b(a, 1 << 4); + case 5: return __lasx_xvandi_b(a, 1 << 5); + case 6: return __lasx_xvandi_b(a, 1 << 6); + case 7: return __lasx_xvandi_b(a, 1 << 7); + default: __builtin_unreachable(); + } +} + // multiply int8_t, add results pairwise twice static inline __m128i mul_sum_i8_pairs(const __m128i x, const __m128i y) { // Get absolute values of x vectors @@ -656,13 +691,8 @@ static inline __m256 mul_sum_us8_pairs_float(const __m256i ax, const __m256i sy) // multiply int8_t, add results pairwise twice and return as float vector static inline __m256 mul_sum_i8_pairs_float(const __m256i x, const __m256i y) { - - // Get absolute values of x vectors - const __m256i ax = __lasx_xvsigncov_b(x, x); - // Sign the values of the y vectors - const __m256i sy = __lasx_xvsigncov_b(x, y); - - return mul_sum_us8_pairs_float(ax, sy); + const __m256i dot = lasx_madd_h_b(x, y); + return sum_i16_pairs_float(dot); } static inline __m128i packNibbles( __m256i bytes ) { @@ -689,28 +719,28 @@ static inline __m128i packNibbles( __m256i bytes ) { } #endif //__loongarch_asx -void quantize_row_q4_0(const float * restrict x, void * restrict y, int64_t k) { +void quantize_row_q4_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k) { quantize_row_q4_0_ref(x, y, k); } -void quantize_row_q4_1(const float * restrict x, void * restrict y, int64_t k) { +void quantize_row_q4_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k) { quantize_row_q4_1_ref(x, y, k); } -void quantize_row_q5_0(const float * restrict x, void * restrict y, int64_t k) { +void quantize_row_q5_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k) { quantize_row_q5_0_ref(x, y, k); } -void quantize_row_q5_1(const float * restrict x, void * restrict y, int64_t k) { +void quantize_row_q5_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k) { quantize_row_q5_1_ref(x, y, k); } -void quantize_row_q8_0(const float * restrict x, void * restrict vy, int64_t k) { +void quantize_row_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) { assert(QK8_0 == 32); assert(k % QK8_0 == 0); const int nb = k / QK8_0; - block_q8_0 * restrict y = vy; + block_q8_0 * GGML_RESTRICT y = vy; #if defined(__ARM_NEON) for (int i = 0; i < nb; i++) { @@ -742,7 +772,7 @@ void quantize_row_q8_0(const float * restrict x, void * restrict vy, int64_t k) y[i].qs[4*j + 3] = vgetq_lane_s32(vi, 3); } } -#elif defined(__wasm_simd128__) +#elif defined __wasm_simd128__ for (int i = 0; i < nb; i++) { v128_t srcv [8]; v128_t asrcv[8]; @@ -861,15 +891,15 @@ void quantize_row_q8_0(const float * restrict x, void * restrict vy, int64_t k) } #elif defined(__riscv_v_intrinsic) - size_t vl = __riscv_vsetvl_e32m4(QK8_0); + size_t vl = QK8_0; for (int i = 0; i < nb; i++) { // load elements - vfloat32m4_t v_x = __riscv_vle32_v_f32m4(x+i*QK8_0, vl); + vfloat32m8_t v_x = __riscv_vle32_v_f32m8(x+i*QK8_0, vl); - vfloat32m4_t vfabs = __riscv_vfabs_v_f32m4(v_x, vl); + vfloat32m8_t vfabs = __riscv_vfabs_v_f32m8(v_x, vl); vfloat32m1_t tmp = __riscv_vfmv_v_f_f32m1(0.0f, vl); - vfloat32m1_t vmax = __riscv_vfredmax_vs_f32m4_f32m1(vfabs, tmp, vl); + vfloat32m1_t vmax = __riscv_vfredmax_vs_f32m8_f32m1(vfabs, tmp, vl); float amax = __riscv_vfmv_f_s_f32m1_f32(vmax); const float d = amax / ((1 << 7) - 1); @@ -877,14 +907,14 @@ void quantize_row_q8_0(const float * restrict x, void * restrict vy, int64_t k) y[i].d = GGML_FP32_TO_FP16(d); - vfloat32m4_t x0 = __riscv_vfmul_vf_f32m4(v_x, id, vl); + vfloat32m8_t x0 = __riscv_vfmul_vf_f32m8(v_x, id, vl); // convert to integer - vint16m2_t vi = __riscv_vfncvt_x_f_w_i16m2(x0, vl); - vint8m1_t vs = __riscv_vncvt_x_x_w_i8m1(vi, vl); + vint16m4_t vi = __riscv_vfncvt_x_f_w_i16m4(x0, vl); + vint8m2_t vs = __riscv_vncvt_x_x_w_i8m2(vi, vl); // store result - __riscv_vse8_v_i8m1(y[i].qs , vs, vl); + __riscv_vse8_v_i8m2(y[i].qs , vs, vl); } #elif defined(__POWER9_VECTOR__) @@ -981,6 +1011,38 @@ void quantize_row_q8_0(const float * restrict x, void * restrict vy, int64_t k) __lsx_vst(ni4, (__m128i *)(y[i].qs + 16), 0); } +#elif defined(__VXE__) || defined(__VXE2__) + for (int i = 0; i < nb; i++) { + __vector float srcv [8]; + __vector float asrcv[8]; + __vector float amaxv[8]; + + for (int j = 0; j < 8; j++) srcv[j] = vec_xl(0, x + i*32 + 4*j); + for (int j = 0; j < 8; j++) asrcv[j] = vec_abs(srcv[j]); + for (int j = 0; j < 4; j++) amaxv[2*j] = vec_max(asrcv[2*j], asrcv[2*j+1]); + for (int j = 0; j < 2; j++) amaxv[4*j] = vec_max(amaxv[4*j], amaxv[4*j+2]); + for (int j = 0; j < 1; j++) amaxv[8*j] = vec_max(amaxv[8*j], amaxv[8*j+4]); + + const float amax = MAX(MAX(vec_extract(amaxv[0], 0), + vec_extract(amaxv[0], 1)), + MAX(vec_extract(amaxv[0], 2), + vec_extract(amaxv[0], 3))); + + const float d = amax / ((1 << 7) - 1); + const float id = d ? 1.0f / d : 0.0f; + + y[i].d = GGML_FP32_TO_FP16(d); + + for (int j = 0; j < 8; j++) { + const __vector float v = vec_mul(srcv[j], vec_splats(id)); + const __vector int32_t vi = vec_signed(v); + + y[i].qs[4*j + 0] = vec_extract(vi, 0); + y[i].qs[4*j + 1] = vec_extract(vi, 1); + y[i].qs[4*j + 2] = vec_extract(vi, 2); + y[i].qs[4*j + 3] = vec_extract(vi, 3); + } + } #else GGML_UNUSED(nb); // scalar @@ -988,11 +1050,11 @@ void quantize_row_q8_0(const float * restrict x, void * restrict vy, int64_t k) #endif } -void quantize_row_q8_1(const float * restrict x, void * restrict vy, int64_t k) { +void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) { assert(k % QK8_1 == 0); const int nb = k / QK8_1; - block_q8_1 * restrict y = vy; + block_q8_1 * GGML_RESTRICT y = vy; #if defined(__ARM_NEON) for (int i = 0; i < nb; i++) { @@ -1030,7 +1092,7 @@ void quantize_row_q8_1(const float * restrict x, void * restrict vy, int64_t k) y[i].s = GGML_FP32_TO_FP16(d * vaddvq_s32(accv)); } -#elif defined(__wasm_simd128__) +#elif defined __wasm_simd128__ for (int i = 0; i < nb; i++) { v128_t srcv [8]; v128_t asrcv[8]; @@ -1167,15 +1229,15 @@ void quantize_row_q8_1(const float * restrict x, void * restrict vy, int64_t k) } #elif defined(__riscv_v_intrinsic) - size_t vl = __riscv_vsetvl_e32m4(QK8_1); + size_t vl = QK8_1; for (int i = 0; i < nb; i++) { // load elements - vfloat32m4_t v_x = __riscv_vle32_v_f32m4(x+i*QK8_1, vl); + vfloat32m8_t v_x = __riscv_vle32_v_f32m8(x+i*QK8_1, vl); - vfloat32m4_t vfabs = __riscv_vfabs_v_f32m4(v_x, vl); + vfloat32m8_t vfabs = __riscv_vfabs_v_f32m8(v_x, vl); vfloat32m1_t tmp = __riscv_vfmv_v_f_f32m1(0.0, vl); - vfloat32m1_t vmax = __riscv_vfredmax_vs_f32m4_f32m1(vfabs, tmp, vl); + vfloat32m1_t vmax = __riscv_vfredmax_vs_f32m8_f32m1(vfabs, tmp, vl); float amax = __riscv_vfmv_f_s_f32m1_f32(vmax); const float d = amax / ((1 << 7) - 1); @@ -1183,18 +1245,18 @@ void quantize_row_q8_1(const float * restrict x, void * restrict vy, int64_t k) y[i].d = GGML_FP32_TO_FP16(d); - vfloat32m4_t x0 = __riscv_vfmul_vf_f32m4(v_x, id, vl); + vfloat32m8_t x0 = __riscv_vfmul_vf_f32m8(v_x, id, vl); // convert to integer - vint16m2_t vi = __riscv_vfncvt_x_f_w_i16m2(x0, vl); - vint8m1_t vs = __riscv_vncvt_x_x_w_i8m1(vi, vl); + vint16m4_t vi = __riscv_vfncvt_x_f_w_i16m4(x0, vl); + vint8m2_t vs = __riscv_vncvt_x_x_w_i8m2(vi, vl); // store result - __riscv_vse8_v_i8m1(y[i].qs , vs, vl); + __riscv_vse8_v_i8m2(y[i].qs , vs, vl); // compute sum for y[i].s vint16m1_t tmp2 = __riscv_vmv_v_x_i16m1(0, vl); - vint16m1_t vwrs = __riscv_vwredsum_vs_i8m1_i16m1(vs, tmp2, vl); + vint16m1_t vwrs = __riscv_vwredsum_vs_i8m2_i16m1(vs, tmp2, vl); // set y[i].s int sum = __riscv_vmv_x_s_i16m1_i16(vwrs); @@ -1307,6 +1369,44 @@ void quantize_row_q8_1(const float * restrict x, void * restrict vy, int64_t k) __lsx_vst(ni0, (__m128i *)(y[i].qs + 0), 0); __lsx_vst(ni4, (__m128i *)(y[i].qs + 16), 0); } +#elif defined(__VXE__) || defined(__VXE2__) + for (int i = 0; i < nb; i++) { + __vector float srcv [8]; + __vector float asrcv[8]; + __vector float amaxv[8]; + + for (int j = 0; j < 8; j++) srcv[j] = vec_xl(0, x + i*32 + 4*j); + for (int j = 0; j < 8; j++) asrcv[j] = vec_abs(srcv[j]); + for (int j = 0; j < 4; j++) amaxv[2*j] = vec_max(asrcv[2*j], asrcv[2*j+1]); + for (int j = 0; j < 2; j++) amaxv[4*j] = vec_max(amaxv[4*j], amaxv[4*j+2]); + for (int j = 0; j < 1; j++) amaxv[8*j] = vec_max(amaxv[8*j], amaxv[8*j+4]); + + const float amax = MAX(MAX(vec_extract(amaxv[0], 0), + vec_extract(amaxv[0], 1)), + MAX(vec_extract(amaxv[0], 2), + vec_extract(amaxv[0], 3))); + + const float d = amax / ((1 << 7) - 1); + const float id = d ? 1.0f / d : 0.0f; + + y[i].d = GGML_FP32_TO_FP16(d); + + __vector int32_t acc = vec_splats(0); + + for (int j = 0; j < 8; j++) { + const __vector float v = vec_mul(srcv[j], vec_splats(id)); + const __vector int32_t vi = vec_signed(v); + + y[i].qs[4*j + 0] = vec_extract(vi, 0); + y[i].qs[4*j + 1] = vec_extract(vi, 1); + y[i].qs[4*j + 2] = vec_extract(vi, 2); + y[i].qs[4*j + 3] = vec_extract(vi, 3); + + acc = vec_add(acc, vi); + } + + y[i].s = GGML_FP32_TO_FP16(d * (acc[0] + acc[1] + acc[2] + acc[3])); + } #else GGML_UNUSED(nb); // scalar @@ -1328,8 +1428,8 @@ static inline int nearest_int(float fval) { return (i & 0x007fffff) - 0x00400000; } -static float make_qx_quants(int n, int nmax, const float * restrict x, int8_t * restrict L, int rmse_type, - const float * restrict qw) { +static float make_qx_quants(int n, int nmax, const float * GGML_RESTRICT x, int8_t * GGML_RESTRICT L, int rmse_type, + const float * GGML_RESTRICT qw) { float max = 0; float amax = 0; for (int i = 0; i < n; ++i) { @@ -1397,7 +1497,7 @@ static float make_qx_quants(int n, int nmax, const float * restrict x, int8_t * return scale; } -static float make_q3_quants(int n, int nmax, const float * restrict x, int8_t * restrict L, bool do_rmse) { +static float make_q3_quants(int n, int nmax, const float * GGML_RESTRICT x, int8_t * GGML_RESTRICT L, bool do_rmse) { float max = 0; float amax = 0; for (int i = 0; i < n; ++i) { @@ -1456,7 +1556,7 @@ static float make_q3_quants(int n, int nmax, const float * restrict x, int8_t * return 1/iscale; } -static float make_qkx1_quants(int n, int nmax, const float * restrict x, uint8_t * restrict L, float * restrict the_min, +static float make_qkx1_quants(int n, int nmax, const float * GGML_RESTRICT x, uint8_t * GGML_RESTRICT L, float * GGML_RESTRICT the_min, int ntry, float alpha) { float min = x[0]; float max = x[0]; @@ -1499,8 +1599,8 @@ static float make_qkx1_quants(int n, int nmax, const float * restrict x, uint8_t return scale; } -static float make_qkx2_quants(int n, int nmax, const float * restrict x, const float * restrict weights, - uint8_t * restrict L, float * restrict the_min, uint8_t * restrict Laux, +static float make_qkx2_quants(int n, int nmax, const float * GGML_RESTRICT x, const float * GGML_RESTRICT weights, + uint8_t * GGML_RESTRICT L, float * GGML_RESTRICT the_min, uint8_t * GGML_RESTRICT Laux, float rmin, float rdelta, int nstep, bool use_mad) { float min = x[0]; float max = x[0]; @@ -1580,7 +1680,7 @@ static float make_qkx2_quants(int n, int nmax, const float * restrict x, const f return scale; } -static inline void get_scale_min_k4(int j, const uint8_t * restrict q, uint8_t * restrict d, uint8_t * restrict m) { +static inline void get_scale_min_k4(int j, const uint8_t * GGML_RESTRICT q, uint8_t * GGML_RESTRICT d, uint8_t * GGML_RESTRICT m) { if (j < 4) { *d = q[j] & 63; *m = q[j + 4] & 63; } else { @@ -1591,51 +1691,51 @@ static inline void get_scale_min_k4(int j, const uint8_t * restrict q, uint8_t * //========================- 2-bit (de)-quantization -void quantize_row_q2_K(const float * restrict x, void * restrict vy, int64_t k) { +void quantize_row_q2_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) { quantize_row_q2_K_ref(x, vy, k); } //========================= 3-bit (de)-quantization -void quantize_row_q3_K(const float * restrict x, void * restrict vy, int64_t k) { +void quantize_row_q3_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) { quantize_row_q3_K_ref(x, vy, k); } // ====================== 4-bit (de)-quantization -void quantize_row_q4_K(const float * restrict x, void * restrict vy, int64_t k) { +void quantize_row_q4_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) { assert(k % QK_K == 0); - block_q4_K * restrict y = vy; + block_q4_K * GGML_RESTRICT y = vy; quantize_row_q4_K_ref(x, y, k); } // ====================== 5-bit (de)-quantization -void quantize_row_q5_K(const float * restrict x, void * restrict vy, int64_t k) { +void quantize_row_q5_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) { assert(k % QK_K == 0); - block_q5_K * restrict y = vy; + block_q5_K * GGML_RESTRICT y = vy; quantize_row_q5_K_ref(x, y, k); } // ====================== 6-bit (de)-quantization -void quantize_row_q6_K(const float * restrict x, void * restrict vy, int64_t k) { +void quantize_row_q6_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) { assert(k % QK_K == 0); - block_q6_K * restrict y = vy; + block_q6_K * GGML_RESTRICT y = vy; quantize_row_q6_K_ref(x, y, k); } // ====================== Ternary (de)-quantization (BitNet b1.58 and TriLMs) -void quantize_row_tq1_0(const float * restrict x, void * restrict vy, int64_t k) { +void quantize_row_tq1_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) { assert(k % QK_K == 0); - block_tq1_0 * restrict y = vy; + block_tq1_0 * GGML_RESTRICT y = vy; quantize_row_tq1_0_ref(x, y, k); } -void quantize_row_tq2_0(const float * restrict x, void * restrict vy, int64_t k) { +void quantize_row_tq2_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) { assert(k % QK_K == 0); - block_tq2_0 * restrict y = vy; + block_tq2_0 * GGML_RESTRICT y = vy; quantize_row_tq2_0_ref(x, y, k); } @@ -1643,8 +1743,88 @@ static const int8_t kvalues_iq4nl[16] = {-127, -104, -83, -65, -49, -35, -22, -1 //===================================== Q8_K ============================================== -void quantize_row_q8_K(const float * restrict x, void * restrict y, int64_t k) { +void quantize_row_q8_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k) { +#ifdef __wasm_simd128__ + assert(k % QK_K == 0); + const int64_t nb = k / QK_K; + block_q8_K * GGML_RESTRICT yc = y; // Cast to proper type + + for (int i = 0; i < nb; i++) { + const float * x_block = x + i * QK_K; + + v128_t min_vec = wasm_v128_load(x_block); + v128_t max_vec = min_vec; + + for (int j = 4; j < QK_K; j += 4) { + v128_t x_vec = wasm_v128_load(x_block + j); + max_vec = wasm_f32x4_pmax(max_vec, x_vec); + min_vec = wasm_f32x4_pmin(min_vec, x_vec); + } + max_vec = wasm_f32x4_pmax(max_vec, wasm_i32x4_shuffle(max_vec, max_vec, 2, 3, 0, 1)); + max_vec = wasm_f32x4_pmax(max_vec, wasm_i32x4_shuffle(max_vec, max_vec, 1, 0, 3, 2)); + min_vec = wasm_f32x4_pmin(min_vec, wasm_i32x4_shuffle(min_vec, min_vec, 2, 3, 0, 1)); + min_vec = wasm_f32x4_pmin(min_vec, wasm_i32x4_shuffle(min_vec, min_vec, 1, 0, 3, 2)); + float max = wasm_f32x4_extract_lane(max_vec, 0); + float min = wasm_f32x4_extract_lane(min_vec, 0); + float amax = -min > max ? min : max; + + if (amax == 0.0f) { + yc[i].d = 0.0f; + const v128_t zero = wasm_i8x16_splat(0); + for (int j = 0; j < QK_K; j += 16) { + wasm_v128_store(yc[i].qs + j, zero); + } + continue; + } + + const float iscale = -127.0f / amax; + const v128_t scale_vec = wasm_f32x4_splat(iscale); + + // Process 16 elements per iteration + for (int j = 0, jb = 0; j < QK_K; j += 16, jb++) { + // Load and quantize 16 floats + v128_t x0 = wasm_v128_load(x_block + j); + v128_t x1 = wasm_v128_load(x_block + j + 4); + v128_t x2 = wasm_v128_load(x_block + j + 8); + v128_t x3 = wasm_v128_load(x_block + j + 12); + + v128_t q0 = wasm_f32x4_nearest(wasm_f32x4_mul(x0, scale_vec)); + v128_t q1 = wasm_f32x4_nearest(wasm_f32x4_mul(x1, scale_vec)); + v128_t q2 = wasm_f32x4_nearest(wasm_f32x4_mul(x2, scale_vec)); + v128_t q3 = wasm_f32x4_nearest(wasm_f32x4_mul(x3, scale_vec)); + + // Convert to i32 with saturation + v128_t i0 = wasm_i32x4_trunc_sat_f32x4(q0); + v128_t i1 = wasm_i32x4_trunc_sat_f32x4(q1); + v128_t i2 = wasm_i32x4_trunc_sat_f32x4(q2); + v128_t i3 = wasm_i32x4_trunc_sat_f32x4(q3); + + // Pack into 16 i8 values + v128_t i8 = wasm_i8x16_narrow_i16x8( + wasm_i16x8_narrow_i32x4(i0, i1), + wasm_i16x8_narrow_i32x4(i2, i3) + ); + wasm_v128_store(yc[i].qs + j, i8); + + // Calculate bsums using SIMD + v128_t sum16 = wasm_i16x8_add( + wasm_i16x8_extend_low_i8x16(i8), + wasm_i16x8_extend_high_i8x16(i8) + ); + v128_t sum32 = wasm_i32x4_add( + wasm_i32x4_extend_low_i16x8(sum16), + wasm_i32x4_extend_high_i16x8(sum16) + ); + sum32 = wasm_i32x4_add(sum32, wasm_i32x4_shuffle(sum32, sum32, 2, 3, 0, 1)); + sum32 = wasm_i32x4_add(sum32, wasm_i32x4_shuffle(sum32, sum32, 1, 0, 3, 2)); + yc[i].bsums[jb] = wasm_i32x4_extract_lane(sum32, 0); + } + + yc[i].d = 1.0f / iscale; + } +#else quantize_row_q8_K_ref(x, y, k); +#endif } //===================================== Dot products ================================= @@ -1729,7 +1909,7 @@ static inline __m128i get_scale_shuffle(int i) { } #endif -void ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) { +void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) { const int qk = QK8_0; const int nb = n / qk; @@ -1744,23 +1924,23 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, size_t bs, const void * r UNUSED(by); UNUSED(bs); - const block_q4_0 * restrict x = vx; - const block_q8_0 * restrict y = vy; + const block_q4_0 * GGML_RESTRICT x = vx; + const block_q8_0 * GGML_RESTRICT y = vy; #if defined(__ARM_FEATURE_MATMUL_INT8) if (nrc == 2) { - const block_q4_0 * restrict vx0 = vx; - const block_q4_0 * restrict vx1 = (const block_q4_0 *) ((const uint8_t*)vx + bx); - const block_q8_0 * restrict vy0 = vy; - const block_q8_0 * restrict vy1 = (const block_q8_0 *) ((const uint8_t*)vy + by); + const block_q4_0 * GGML_RESTRICT vx0 = vx; + const block_q4_0 * GGML_RESTRICT vx1 = (const block_q4_0 *) ((const uint8_t*)vx + bx); + const block_q8_0 * GGML_RESTRICT vy0 = vy; + const block_q8_0 * GGML_RESTRICT vy1 = (const block_q8_0 *) ((const uint8_t*)vy + by); float32x4_t sumv0 = vdupq_n_f32(0.0f); for (int i = 0; i < nb; i++) { - const block_q4_0 * restrict b_x0 = &vx0[i]; - const block_q4_0 * restrict b_x1 = &vx1[i]; - const block_q8_0 * restrict b_y0 = &vy0[i]; - const block_q8_0 * restrict b_y1 = &vy1[i]; + const block_q4_0 * GGML_RESTRICT b_x0 = &vx0[i]; + const block_q4_0 * GGML_RESTRICT b_x1 = &vx1[i]; + const block_q8_0 * GGML_RESTRICT b_y0 = &vy0[i]; + const block_q8_0 * GGML_RESTRICT b_y1 = &vy1[i]; const uint8x16_t m4b = vdupq_n_u8(0x0F); const int8x16_t s8b = vdupq_n_s8(0x8); @@ -1837,10 +2017,10 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, size_t bs, const void * r const svbool_t ph4 = svptrue_pat_b32(SV_VL4); for (; ib + 1 < nb; ib += 2) { - const block_q4_0 * restrict x0 = &x[ib + 0]; - const block_q4_0 * restrict x1 = &x[ib + 1]; - const block_q8_0 * restrict y0 = &y[ib + 0]; - const block_q8_0 * restrict y1 = &y[ib + 1]; + const block_q4_0 * GGML_RESTRICT x0 = &x[ib + 0]; + const block_q4_0 * GGML_RESTRICT x1 = &x[ib + 1]; + const block_q8_0 * GGML_RESTRICT y0 = &y[ib + 0]; + const block_q8_0 * GGML_RESTRICT y1 = &y[ib + 1]; // load x const svuint8_t qx0r = svld1rq_u8(svptrue_b8(), x0->qs); @@ -1883,10 +2063,10 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, size_t bs, const void * r const svbool_t pl16 = svnot_b_z(svptrue_b8(), ph16); for (; ib + 1 < nb; ib += 2) { - const block_q4_0 * restrict x0 = &x[ib + 0]; - const block_q4_0 * restrict x1 = &x[ib + 1]; - const block_q8_0 * restrict y0 = &y[ib + 0]; - const block_q8_0 * restrict y1 = &y[ib + 1]; + const block_q4_0 * GGML_RESTRICT x0 = &x[ib + 0]; + const block_q4_0 * GGML_RESTRICT x1 = &x[ib + 1]; + const block_q8_0 * GGML_RESTRICT y0 = &y[ib + 0]; + const block_q8_0 * GGML_RESTRICT y1 = &y[ib + 1]; // load x const svuint8_t qx0r = svld1rq_u8(svptrue_b8(), x0->qs); @@ -1924,10 +2104,10 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, size_t bs, const void * r const svbool_t pl16 = svnot_b_z(ph32, ph16); for (; ib + 1 < nb; ib += 2) { - const block_q4_0 * restrict x0 = &x[ib + 0]; - const block_q4_0 * restrict x1 = &x[ib + 1]; - const block_q8_0 * restrict y0 = &y[ib + 0]; - const block_q8_0 * restrict y1 = &y[ib + 1]; + const block_q4_0 * GGML_RESTRICT x0 = &x[ib + 0]; + const block_q4_0 * GGML_RESTRICT x1 = &x[ib + 1]; + const block_q8_0 * GGML_RESTRICT y0 = &y[ib + 0]; + const block_q8_0 * GGML_RESTRICT y1 = &y[ib + 1]; // load x const svuint8_t qx0r = svld1rq_u8(ph32, x0->qs); @@ -1964,10 +2144,10 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, size_t bs, const void * r float32x4_t sumv1 = vdupq_n_f32(0.0f); for (; ib + 1 < nb; ib += 2) { - const block_q4_0 * restrict x0 = &x[ib + 0]; - const block_q4_0 * restrict x1 = &x[ib + 1]; - const block_q8_0 * restrict y0 = &y[ib + 0]; - const block_q8_0 * restrict y1 = &y[ib + 1]; + const block_q4_0 * GGML_RESTRICT x0 = &x[ib + 0]; + const block_q4_0 * GGML_RESTRICT x1 = &x[ib + 1]; + const block_q8_0 * GGML_RESTRICT y0 = &y[ib + 0]; + const block_q8_0 * GGML_RESTRICT y1 = &y[ib + 1]; const uint8x16_t m4b = vdupq_n_u8(0x0F); const int8x16_t s8b = vdupq_n_s8(0x8); @@ -2002,6 +2182,94 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, size_t bs, const void * r } sumf = vaddvq_f32(sumv0) + vaddvq_f32(sumv1); +#elif defined __wasm_simd128__ + v128_t sumv = wasm_f32x4_splat(0.0f); + + const v128_t m4b = wasm_i8x16_splat(0x0F); + const v128_t s8b = wasm_i8x16_splat(0x8); + + for (; ib + 1 < nb; ib += 2) { + const block_q4_0 * GGML_RESTRICT x0 = &x[ib]; + const block_q4_0 * GGML_RESTRICT x1 = &x[ib + 1]; + const block_q8_0 * GGML_RESTRICT y0 = &y[ib]; + const block_q8_0 * GGML_RESTRICT y1 = &y[ib + 1]; + + // Load and process x0 + v128_t v0_0 = wasm_v128_load(x0->qs); + v128_t v0_0l = wasm_v128_and(v0_0, m4b); + v128_t v0_0h = wasm_u8x16_shr(v0_0, 4); + v128_t v0_0ls = wasm_i8x16_sub(v0_0l, s8b); + v128_t v0_0hs = wasm_i8x16_sub(v0_0h, s8b); + + // Load y0 vectors + v128_t y0_l = wasm_v128_load(y0->qs); + v128_t y0_h = wasm_v128_load(y0->qs + 16); + + // Extend to i16x8 and compute dot products + v128_t dx0l = wasm_i16x8_extend_low_i8x16(v0_0ls); + v128_t dx0h = wasm_i16x8_extend_high_i8x16(v0_0ls); + v128_t dx0hl = wasm_i16x8_extend_low_i8x16(v0_0hs); + v128_t dx0hh = wasm_i16x8_extend_high_i8x16(v0_0hs); + + v128_t dy0ll = wasm_i16x8_extend_low_i8x16(y0_l); + v128_t dy0lh = wasm_i16x8_extend_high_i8x16(y0_l); + v128_t dy0hl = wasm_i16x8_extend_low_i8x16(y0_h); + v128_t dy0hh = wasm_i16x8_extend_high_i8x16(y0_h); + + v128_t dp0 = wasm_i32x4_add( + wasm_i32x4_add( + wasm_i32x4_dot_i16x8(dx0l, dy0ll), + wasm_i32x4_dot_i16x8(dx0h, dy0lh) + ), + wasm_i32x4_add( + wasm_i32x4_dot_i16x8(dx0hl, dy0hl), + wasm_i32x4_dot_i16x8(dx0hh, dy0hh) + ) + ); + + // Load and process x1 + v128_t v0_1 = wasm_v128_load(x1->qs); + v128_t v0_1l = wasm_v128_and(v0_1, m4b); + v128_t v0_1h = wasm_u8x16_shr(v0_1, 4); + v128_t v0_1ls = wasm_i8x16_sub(v0_1l, s8b); + v128_t v0_1hs = wasm_i8x16_sub(v0_1h, s8b); + + // Load y1 vectors + v128_t y1_l = wasm_v128_load(y1->qs); + v128_t y1_h = wasm_v128_load(y1->qs + 16); + + // Extend to i16x8 and compute dot products + v128_t dx1l = wasm_i16x8_extend_low_i8x16(v0_1ls); + v128_t dx1h = wasm_i16x8_extend_high_i8x16(v0_1ls); + v128_t dx1hl = wasm_i16x8_extend_low_i8x16(v0_1hs); + v128_t dx1hh = wasm_i16x8_extend_high_i8x16(v0_1hs); + + v128_t dy1ll = wasm_i16x8_extend_low_i8x16(y1_l); + v128_t dy1lh = wasm_i16x8_extend_high_i8x16(y1_l); + v128_t dy1hl = wasm_i16x8_extend_low_i8x16(y1_h); + v128_t dy1hh = wasm_i16x8_extend_high_i8x16(y1_h); + + v128_t dp1 = wasm_i32x4_add( + wasm_i32x4_add( + wasm_i32x4_dot_i16x8(dx1l, dy1ll), + wasm_i32x4_dot_i16x8(dx1h, dy1lh) + ), + wasm_i32x4_add( + wasm_i32x4_dot_i16x8(dx1hl, dy1hl), + wasm_i32x4_dot_i16x8(dx1hh, dy1hh) + ) + ); + + // Accumulate results with scaling + float scale0 = GGML_FP16_TO_FP32(x0->d) * GGML_FP16_TO_FP32(y0->d); + float scale1 = GGML_FP16_TO_FP32(x1->d) * GGML_FP16_TO_FP32(y1->d); + + sumv = wasm_f32x4_add(sumv, wasm_f32x4_mul(wasm_f32x4_convert_i32x4(dp0), wasm_f32x4_splat(scale0))); + sumv = wasm_f32x4_add(sumv, wasm_f32x4_mul(wasm_f32x4_convert_i32x4(dp1), wasm_f32x4_splat(scale1))); + } + + sumf = wasm_f32x4_extract_lane(sumv, 0) + wasm_f32x4_extract_lane(sumv, 1) + + wasm_f32x4_extract_lane(sumv, 2) + wasm_f32x4_extract_lane(sumv, 3); #elif defined(__AVX2__) // Initialize accumulator with zeros __m256 acc = _mm256_setzero_ps(); @@ -2123,33 +2391,31 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, size_t bs, const void * r sumf = hsum_float_4x4(acc_0, acc_1, acc_2, acc_3); #elif defined(__riscv_v_intrinsic) - size_t vl = __riscv_vsetvl_e8m1(qk/2); + size_t vl = qk / 2; for (; ib < nb; ++ib) { // load elements - vuint8mf2_t tx = __riscv_vle8_v_u8mf2(x[ib].qs, vl); + vuint8m1_t tx = __riscv_vle8_v_u8m1(x[ib].qs, vl); - vint8mf2_t y0 = __riscv_vle8_v_i8mf2(y[ib].qs, vl); - vint8mf2_t y1 = __riscv_vle8_v_i8mf2(y[ib].qs+16, vl); + vint8m1_t y0 = __riscv_vle8_v_i8m1(y[ib].qs, vl); + vint8m1_t y1 = __riscv_vle8_v_i8m1(y[ib].qs+16, vl); // mask and store lower part of x, and then upper part - vuint8mf2_t x_a = __riscv_vand_vx_u8mf2(tx, 0x0F, vl); - vuint8mf2_t x_l = __riscv_vsrl_vx_u8mf2(tx, 0x04, vl); + vuint8m1_t x_a = __riscv_vand_vx_u8m1(tx, 0x0F, vl); + vuint8m1_t x_l = __riscv_vsrl_vx_u8m1(tx, 0x04, vl); - vint8mf2_t x_ai = __riscv_vreinterpret_v_u8mf2_i8mf2(x_a); - vint8mf2_t x_li = __riscv_vreinterpret_v_u8mf2_i8mf2(x_l); + vint8m1_t x_ai = __riscv_vreinterpret_v_u8m1_i8m1(x_a); + vint8m1_t x_li = __riscv_vreinterpret_v_u8m1_i8m1(x_l); // subtract offset - vint8mf2_t v0 = __riscv_vsub_vx_i8mf2(x_ai, 8, vl); - vint8mf2_t v1 = __riscv_vsub_vx_i8mf2(x_li, 8, vl); + vint8m1_t v0 = __riscv_vsub_vx_i8m1(x_ai, 8, vl); + vint8m1_t v1 = __riscv_vsub_vx_i8m1(x_li, 8, vl); - vint16m1_t vec_mul1 = __riscv_vwmul_vv_i16m1(v0, y0, vl); - vint16m1_t vec_mul2 = __riscv_vwmul_vv_i16m1(v1, y1, vl); + vint16m2_t vec_mul1 = __riscv_vwmul_vv_i16m2(v0, y0, vl); + vint16m2_t vec_mul2 = __riscv_vwmacc_vv_i16m2(vec_mul1, v1, y1, vl); vint32m1_t vec_zero = __riscv_vmv_v_x_i32m1(0, vl); - - vint32m1_t vs1 = __riscv_vwredsum_vs_i16m1_i32m1(vec_mul1, vec_zero, vl); - vint32m1_t vs2 = __riscv_vwredsum_vs_i16m1_i32m1(vec_mul2, vs1, vl); + vint32m1_t vs2 = __riscv_vwredsum_vs_i16m2_i32m1(vec_mul2, vec_zero, vl); int sumi = __riscv_vmv_x_s_i32m1_i32(vs2); @@ -2290,6 +2556,37 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, size_t bs, const void * r } sumf = hsum_float_4x4(acc_0, acc_1, acc_2, acc_3); +#elif defined(__VXE__) || defined(__VXE2__) + __vector float acc = vec_splats(0.0f); + + const __vector uint8_t v_m = vec_splats((const uint8_t)0x0F); + const __vector int8_t v_s = vec_splats( (const int8_t)0x08); + + for (; ib < nb; ++ib) { + const __vector uint8_t v_x = vec_xl(0, x[ib].qs); + const __vector int8_t v_xl = (const __vector int8_t)(v_x & v_m); + const __vector int8_t v_xh = (const __vector int8_t)(v_x >> 4); + + const __vector int8_t v_xls = vec_sub(v_xl, v_s); + const __vector int8_t v_xhs = vec_sub(v_xh, v_s); + + const __vector int8_t v_yl = vec_xl(0 , y[ib].qs); + const __vector int8_t v_yh = vec_xl(QK8_0/2, y[ib].qs); + + const __vector int16_t v_xylso = vec_mulo(v_xls, v_yl); + const __vector int16_t v_xylse = vec_mule(v_xls, v_yl); + const __vector int16_t v_xyhso = vec_mulo(v_xhs, v_yh); + const __vector int16_t v_xyhse = vec_mule(v_xhs, v_yh); + + __vector int16_t v_xy_ = v_xylso + v_xylse + v_xyhso + v_xyhse; v_xy_ += vec_reve(v_xy_); + + const __vector float v_xy = vec_float(vec_unpackh(v_xy_)); + const __vector float v_d = vec_splats(GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d)); + + acc = vec_madd(v_xy, v_d, acc); + } + + sumf = acc[0] + acc[1] + acc[2] + acc[3]; #endif for (; ib < nb; ++ib) { int sumi0 = 0; @@ -2310,7 +2607,7 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, size_t bs, const void * r *s = sumf; } -void ggml_vec_dot_q4_1_q8_1(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) { +void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) { const int qk = QK8_1; const int nb = n / qk; @@ -2325,24 +2622,24 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * restrict s, size_t bs, const void * r UNUSED(by); UNUSED(bs); - const block_q4_1 * restrict x = vx; - const block_q8_1 * restrict y = vy; + const block_q4_1 * GGML_RESTRICT x = vx; + const block_q8_1 * GGML_RESTRICT y = vy; #if defined(__ARM_FEATURE_MATMUL_INT8) if (nrc == 2) { - const block_q4_1 * restrict vx0 = vx; - const block_q4_1 * restrict vx1 = (const block_q4_1 *) ((const uint8_t*)vx + bx); - const block_q8_1 * restrict vy0 = vy; - const block_q8_1 * restrict vy1 = (const block_q8_1 *) ((const uint8_t*)vy + by); + const block_q4_1 * GGML_RESTRICT vx0 = vx; + const block_q4_1 * GGML_RESTRICT vx1 = (const block_q4_1 *) ((const uint8_t*)vx + bx); + const block_q8_1 * GGML_RESTRICT vy0 = vy; + const block_q8_1 * GGML_RESTRICT vy1 = (const block_q8_1 *) ((const uint8_t*)vy + by); float32x4_t sumv0 = vdupq_n_f32(0.0f); float32x4_t summs0 = vdupq_n_f32(0.0f); for (int i = 0; i < nb; i++) { - const block_q4_1 * restrict b_x0 = &vx0[i]; - const block_q4_1 * restrict b_x1 = &vx1[i]; - const block_q8_1 * restrict b_y0 = &vy0[i]; - const block_q8_1 * restrict b_y1 = &vy1[i]; + const block_q4_1 * GGML_RESTRICT b_x0 = &vx0[i]; + const block_q4_1 * GGML_RESTRICT b_x1 = &vx1[i]; + const block_q8_1 * GGML_RESTRICT b_y0 = &vy0[i]; + const block_q8_1 * GGML_RESTRICT b_y1 = &vy1[i]; float32_t summs_t[4] = { GGML_FP16_TO_FP32(b_x0->m) * GGML_FP16_TO_FP32(b_y0->s), @@ -2416,10 +2713,10 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * restrict s, size_t bs, const void * r float summs = 0; for (; ib + 1 < nb; ib += 2) { - const block_q4_1 * restrict x0 = &x[ib + 0]; - const block_q4_1 * restrict x1 = &x[ib + 1]; - const block_q8_1 * restrict y0 = &y[ib + 0]; - const block_q8_1 * restrict y1 = &y[ib + 1]; + const block_q4_1 * GGML_RESTRICT x0 = &x[ib + 0]; + const block_q4_1 * GGML_RESTRICT x1 = &x[ib + 1]; + const block_q8_1 * GGML_RESTRICT y0 = &y[ib + 0]; + const block_q8_1 * GGML_RESTRICT y1 = &y[ib + 1]; summs += GGML_FP16_TO_FP32(x0->m) * GGML_FP16_TO_FP32(y0->s) + GGML_FP16_TO_FP32(x1->m) * GGML_FP16_TO_FP32(y1->s); @@ -2484,29 +2781,27 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * restrict s, size_t bs, const void * r sumf = hsum_float_8(acc) + summs; #elif defined(__riscv_v_intrinsic) - size_t vl = __riscv_vsetvl_e8m1(qk/2); + size_t vl = qk / 2; for (; ib < nb; ++ib) { // load elements - vuint8mf2_t tx = __riscv_vle8_v_u8mf2(x[ib].qs, vl); + vuint8m1_t tx = __riscv_vle8_v_u8m1(x[ib].qs, vl); - vint8mf2_t y0 = __riscv_vle8_v_i8mf2(y[ib].qs, vl); - vint8mf2_t y1 = __riscv_vle8_v_i8mf2(y[ib].qs+16, vl); + vint8m1_t y0 = __riscv_vle8_v_i8m1(y[ib].qs, vl); + vint8m1_t y1 = __riscv_vle8_v_i8m1(y[ib].qs+16, vl); // mask and store lower part of x, and then upper part - vuint8mf2_t x_a = __riscv_vand_vx_u8mf2(tx, 0x0F, vl); - vuint8mf2_t x_l = __riscv_vsrl_vx_u8mf2(tx, 0x04, vl); + vuint8m1_t x_a = __riscv_vand_vx_u8m1(tx, 0x0F, vl); + vuint8m1_t x_l = __riscv_vsrl_vx_u8m1(tx, 0x04, vl); - vint8mf2_t v0 = __riscv_vreinterpret_v_u8mf2_i8mf2(x_a); - vint8mf2_t v1 = __riscv_vreinterpret_v_u8mf2_i8mf2(x_l); + vint8m1_t v0 = __riscv_vreinterpret_v_u8m1_i8m1(x_a); + vint8m1_t v1 = __riscv_vreinterpret_v_u8m1_i8m1(x_l); - vint16m1_t vec_mul1 = __riscv_vwmul_vv_i16m1(v0, y0, vl); - vint16m1_t vec_mul2 = __riscv_vwmul_vv_i16m1(v1, y1, vl); + vint16m2_t vec_mul1 = __riscv_vwmul_vv_i16m2(v0, y0, vl); + vint16m2_t vec_mul2 = __riscv_vwmacc_vv_i16m2(vec_mul1, v1, y1, vl); vint32m1_t vec_zero = __riscv_vmv_v_x_i32m1(0, vl); - - vint32m1_t vs1 = __riscv_vwredsum_vs_i16m1_i32m1(vec_mul1, vec_zero, vl); - vint32m1_t vs2 = __riscv_vwredsum_vs_i16m1_i32m1(vec_mul2, vs1, vl); + vint32m1_t vs2 = __riscv_vwredsum_vs_i16m2_i32m1(vec_mul2, vec_zero, vl); int sumi = __riscv_vmv_x_s_i32m1_i32(vs2); @@ -2583,6 +2878,35 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * restrict s, size_t bs, const void * r } sumf = hsum_float_8(acc) + summs; +#elif defined(__VXE__) || defined(__VXE2__) + float summs = 0; + float32x4_t acc = vec_splats(0.0f); + + const uint8x16_t v_m = vec_splat_u8(0x0F); + +#pragma GCC unroll 4 + for (; ib < nb; ++ib) { + __builtin_prefetch(x[ib].qs, 0, 1); + __builtin_prefetch(y[ib].qs, 0, 1); + + summs += GGML_FP16_TO_FP32(x[ib].m) * GGML_FP16_TO_FP32(y[ib].s); + + const uint8x16_t v_x = vec_xl(0, x[ib].qs); + const int8x16_t v_xl = (const int8x16_t)(v_x & v_m); + const int8x16_t v_xh = (const int8x16_t)(v_x >> 4); + + const int8x16_t v_yl = vec_xl(0 , y[ib].qs); + const int8x16_t v_yh = vec_xl(QK8_1/2, y[ib].qs); + + const int32x4_t v_xy_ = ggml_vec_dot(ggml_vec_dot(vec_splats(0), v_xl, v_yl), v_xh, v_yh); + const float32x4_t v_xy = vec_float(v_xy_); + + const float32x4_t v_d = vec_splats(GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d)); + + acc = vec_madd(v_xy, v_d, acc); + } + + sumf = acc[0] + acc[1] + acc[2] + acc[3] + summs; #endif for (; ib < nb; ++ib) { int sumi0 = 0; @@ -2603,7 +2927,7 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * restrict s, size_t bs, const void * r *s = sumf; } -void ggml_vec_dot_q5_0_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) { +void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) { const int qk = QK8_0; const int nb = n / qk; @@ -2618,8 +2942,8 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * restrict s, size_t bs, const void * r UNUSED(by); UNUSED(bs); - const block_q5_0 * restrict x = vx; - const block_q8_0 * restrict y = vy; + const block_q5_0 * GGML_RESTRICT x = vx; + const block_q8_0 * GGML_RESTRICT y = vy; #if defined(__ARM_NEON) float32x4_t sumv0 = vdupq_n_f32(0.0f); @@ -2632,10 +2956,10 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * restrict s, size_t bs, const void * r uint64_t tmp1[4]; for (; ib + 1 < nb; ib += 2) { - const block_q5_0 * restrict x0 = &x[ib]; - const block_q5_0 * restrict x1 = &x[ib + 1]; - const block_q8_0 * restrict y0 = &y[ib]; - const block_q8_0 * restrict y1 = &y[ib + 1]; + const block_q5_0 * GGML_RESTRICT x0 = &x[ib]; + const block_q5_0 * GGML_RESTRICT x1 = &x[ib + 1]; + const block_q8_0 * GGML_RESTRICT y0 = &y[ib]; + const block_q8_0 * GGML_RESTRICT y1 = &y[ib + 1]; const uint8x16_t m4b = vdupq_n_u8(0x0F); @@ -2688,26 +3012,26 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * restrict s, size_t bs, const void * r } sumf = vaddvq_f32(sumv0) + vaddvq_f32(sumv1); -#elif defined(__wasm_simd128__) +#elif defined __wasm_simd128__ v128_t sumv = wasm_f32x4_splat(0.0f); - uint32_t qh; + uint32_t qh_; uint64_t tmp[4]; // TODO: check if unrolling this is better for (; ib < nb; ++ib) { - const block_q5_0 * restrict x0 = &x[ib]; - const block_q8_0 * restrict y0 = &y[ib]; + const block_q5_0 * GGML_RESTRICT x0 = &x[ib]; + const block_q8_0 * GGML_RESTRICT y0 = &y[ib]; const v128_t m4b = wasm_i8x16_splat(0x0F); // extract the 5th bit - memcpy(&qh, x0->qh, sizeof(qh)); + memcpy(&qh_, x0->qh, sizeof(qh_)); - tmp[0] = table_b2b_1[(qh >> 0) & 0xFF]; - tmp[1] = table_b2b_1[(qh >> 8) & 0xFF]; - tmp[2] = table_b2b_1[(qh >> 16) & 0xFF]; - tmp[3] = table_b2b_1[(qh >> 24) ]; + tmp[0] = table_b2b_1[(qh_ >> 0) & 0xFF]; + tmp[1] = table_b2b_1[(qh_ >> 8) & 0xFF]; + tmp[2] = table_b2b_1[(qh_ >> 16) & 0xFF]; + tmp[3] = table_b2b_1[(qh_ >> 24) ]; const v128_t qhl = wasm_v128_load(tmp + 0); const v128_t qhh = wasm_v128_load(tmp + 2); @@ -2804,65 +3128,33 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * restrict s, size_t bs, const void * r sumf = hsum_float_8(acc); #elif defined(__riscv_v_intrinsic) - uint32_t qh; - - size_t vl = __riscv_vsetvl_e8m1(qk/2); - - // These temporary registers are for masking and shift operations - vuint32m2_t vt_1 = __riscv_vid_v_u32m2(vl); - vuint32m2_t vt_2 = __riscv_vsll_vv_u32m2(__riscv_vmv_v_x_u32m2(1, vl), vt_1, vl); - - vuint32m2_t vt_3 = __riscv_vsll_vx_u32m2(vt_2, 16, vl); - vuint32m2_t vt_4 = __riscv_vadd_vx_u32m2(vt_1, 12, vl); + size_t vl; + size_t vlenb = __riscv_vlenb(); for (; ib < nb; ++ib) { - memcpy(&qh, x[ib].qh, sizeof(uint32_t)); + vl = qk / 2; + vuint8m1_t v0 = __riscv_vle8_v_u8m1(x[ib].qs, vl); + vint8m1_t v0l = __riscv_vreinterpret_v_u8m1_i8m1(__riscv_vand_vx_u8m1(v0, 0x0F, vl)); + vint8m1_t v0h = __riscv_vreinterpret_v_u8m1_i8m1(__riscv_vsrl_vx_u8m1(v0, 4, vl)); + vint8m2_t v0c; + if (vlenb == 16) { + v0c = __riscv_vcreate_v_i8m1_i8m2(v0l, v0h); + } else { + v0l = __riscv_vslideup_vx_i8m1(v0l, v0h, 16, 32); + v0c = __riscv_vlmul_ext_v_i8m1_i8m2(v0l); + } - // ((qh & (1u << (j + 0 ))) >> (j + 0 )) << 4; - vuint32m2_t xha_0 = __riscv_vand_vx_u32m2(vt_2, qh, vl); - vuint32m2_t xhr_0 = __riscv_vsrl_vv_u32m2(xha_0, vt_1, vl); - vuint32m2_t xhl_0 = __riscv_vsll_vx_u32m2(xhr_0, 4, vl); + vl = qk; + vbool4_t qh = __riscv_vlm_v_b4(x[ib].qh, vl); + qh = __riscv_vmnand_mm_b4(qh, qh, vl); + vint8m2_t v0f = __riscv_vsub_vx_i8m2_mu(qh, v0c, v0c, 0x10, vl); + vint8m2_t v1 = __riscv_vle8_v_i8m2(y[ib].qs, vl); + vint16m4_t mul = __riscv_vwmul_vv_i16m4(v0f, v1, vl); + vint32m1_t zero = __riscv_vmv_v_x_i32m1(0, vl); + vint32m1_t sum = __riscv_vwredsum_vs_i16m4_i32m1(mul, zero, vl); + int32_t sumi = __riscv_vmv_x_s_i32m1_i32(sum); - // ((qh & (1u << (j + 16))) >> (j + 12)); - vuint32m2_t xha_1 = __riscv_vand_vx_u32m2(vt_3, qh, vl); - vuint32m2_t xhl_1 = __riscv_vsrl_vv_u32m2(xha_1, vt_4, vl); - - // narrowing - vuint16m1_t xhc_0 = __riscv_vncvt_x_x_w_u16m1(xhl_0, vl); - vuint8mf2_t xh_0 = __riscv_vncvt_x_x_w_u8mf2(xhc_0, vl); - - vuint16m1_t xhc_1 = __riscv_vncvt_x_x_w_u16m1(xhl_1, vl); - vuint8mf2_t xh_1 = __riscv_vncvt_x_x_w_u8mf2(xhc_1, vl); - - // load - vuint8mf2_t tx = __riscv_vle8_v_u8mf2(x[ib].qs, vl); - - vint8mf2_t y0 = __riscv_vle8_v_i8mf2(y[ib].qs, vl); - vint8mf2_t y1 = __riscv_vle8_v_i8mf2(y[ib].qs+16, vl); - - vuint8mf2_t x_at = __riscv_vand_vx_u8mf2(tx, 0x0F, vl); - vuint8mf2_t x_lt = __riscv_vsrl_vx_u8mf2(tx, 0x04, vl); - - vuint8mf2_t x_a = __riscv_vor_vv_u8mf2(x_at, xh_0, vl); - vuint8mf2_t x_l = __riscv_vor_vv_u8mf2(x_lt, xh_1, vl); - - vint8mf2_t x_ai = __riscv_vreinterpret_v_u8mf2_i8mf2(x_a); - vint8mf2_t x_li = __riscv_vreinterpret_v_u8mf2_i8mf2(x_l); - - vint8mf2_t v0 = __riscv_vsub_vx_i8mf2(x_ai, 16, vl); - vint8mf2_t v1 = __riscv_vsub_vx_i8mf2(x_li, 16, vl); - - vint16m1_t vec_mul1 = __riscv_vwmul_vv_i16m1(v0, y0, vl); - vint16m1_t vec_mul2 = __riscv_vwmul_vv_i16m1(v1, y1, vl); - - vint32m1_t vec_zero = __riscv_vmv_v_x_i32m1(0, vl); - - vint32m1_t vs1 = __riscv_vwredsum_vs_i16m1_i32m1(vec_mul1, vec_zero, vl); - vint32m1_t vs2 = __riscv_vwredsum_vs_i16m1_i32m1(vec_mul2, vs1, vl); - - int sumi = __riscv_vmv_x_s_i32m1_i32(vs2); - - sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d)) * sumi; + sumf += (GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d)) * sumi; } #elif defined(__POWER9_VECTOR__) @@ -2958,7 +3250,7 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * restrict s, size_t bs, const void * r *s = sumf; } -void ggml_vec_dot_q5_1_q8_1(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) { +void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) { const int qk = QK8_1; const int nb = n / qk; @@ -2973,8 +3265,8 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * restrict s, size_t bs, const void * r UNUSED(by); UNUSED(bs); - const block_q5_1 * restrict x = vx; - const block_q8_1 * restrict y = vy; + const block_q5_1 * GGML_RESTRICT x = vx; + const block_q8_1 * GGML_RESTRICT y = vy; #if defined(__ARM_NEON) float32x4_t sumv0 = vdupq_n_f32(0.0f); @@ -2990,10 +3282,10 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * restrict s, size_t bs, const void * r uint64_t tmp1[4]; for (; ib + 1 < nb; ib += 2) { - const block_q5_1 * restrict x0 = &x[ib]; - const block_q5_1 * restrict x1 = &x[ib + 1]; - const block_q8_1 * restrict y0 = &y[ib]; - const block_q8_1 * restrict y1 = &y[ib + 1]; + const block_q5_1 * GGML_RESTRICT x0 = &x[ib]; + const block_q5_1 * GGML_RESTRICT x1 = &x[ib + 1]; + const block_q8_1 * GGML_RESTRICT y0 = &y[ib]; + const block_q8_1 * GGML_RESTRICT y1 = &y[ib + 1]; const uint8x16_t m4b = vdupq_n_u8(0x0F); @@ -3049,30 +3341,30 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * restrict s, size_t bs, const void * r } sumf = vaddvq_f32(sumv0) + vaddvq_f32(sumv1) + summs0 + summs1; -#elif defined(__wasm_simd128__) +#elif defined __wasm_simd128__ v128_t sumv = wasm_f32x4_splat(0.0f); float summs = 0.0f; - uint32_t qh; + uint32_t qh_; uint64_t tmp[4]; // TODO: check if unrolling this is better for (; ib < nb; ++ib) { - const block_q5_1 * restrict x0 = &x[ib]; - const block_q8_1 * restrict y0 = &y[ib]; + const block_q5_1 * GGML_RESTRICT x0 = &x[ib]; + const block_q8_1 * GGML_RESTRICT y0 = &y[ib]; summs += GGML_FP16_TO_FP32(x0->m) * GGML_FP16_TO_FP32(y0->s); const v128_t m4b = wasm_i8x16_splat(0x0F); // extract the 5th bit - memcpy(&qh, x0->qh, sizeof(qh)); + memcpy(&qh_, x0->qh, sizeof(qh_)); - tmp[0] = table_b2b_0[(qh >> 0) & 0xFF]; - tmp[1] = table_b2b_0[(qh >> 8) & 0xFF]; - tmp[2] = table_b2b_0[(qh >> 16) & 0xFF]; - tmp[3] = table_b2b_0[(qh >> 24) ]; + tmp[0] = table_b2b_0[(qh_ >> 0) & 0xFF]; + tmp[1] = table_b2b_0[(qh_ >> 8) & 0xFF]; + tmp[2] = table_b2b_0[(qh_ >> 16) & 0xFF]; + tmp[3] = table_b2b_0[(qh_ >> 24) ]; const v128_t qhl = wasm_v128_load(tmp + 0); const v128_t qhh = wasm_v128_load(tmp + 2); @@ -3175,60 +3467,30 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * restrict s, size_t bs, const void * r sumf = hsum_float_8(acc) + summs; #elif defined(__riscv_v_intrinsic) - uint32_t qh; - - size_t vl = __riscv_vsetvl_e8m1(qk/2); - - // temporary registers for shift operations - vuint32m2_t vt_1 = __riscv_vid_v_u32m2(vl); - vuint32m2_t vt_2 = __riscv_vadd_vx_u32m2(vt_1, 12, vl); + size_t vl; + size_t vlenb = __riscv_vlenb(); for (; ib < nb; ++ib) { - memcpy(&qh, x[ib].qh, sizeof(uint32_t)); + vl = qk / 2; + vuint8m1_t v0 = __riscv_vle8_v_u8m1(x[ib].qs, vl); + vint8m1_t v0l = __riscv_vreinterpret_v_u8m1_i8m1(__riscv_vand_vx_u8m1(v0, 0x0F, vl)); + vint8m1_t v0h = __riscv_vreinterpret_v_u8m1_i8m1(__riscv_vsrl_vx_u8m1(v0, 4, vl)); + vint8m2_t v0c; + if (vlenb == 16) { + v0c = __riscv_vcreate_v_i8m1_i8m2(v0l, v0h); + } else { + v0l = __riscv_vslideup_vx_i8m1(v0l, v0h, 16, 32); + v0c = __riscv_vlmul_ext_v_i8m1_i8m2(v0l); + } - // load qh - vuint32m2_t vqh = __riscv_vmv_v_x_u32m2(qh, vl); - - // ((qh >> (j + 0)) << 4) & 0x10; - vuint32m2_t xhr_0 = __riscv_vsrl_vv_u32m2(vqh, vt_1, vl); - vuint32m2_t xhl_0 = __riscv_vsll_vx_u32m2(xhr_0, 4, vl); - vuint32m2_t xha_0 = __riscv_vand_vx_u32m2(xhl_0, 0x10, vl); - - // ((qh >> (j + 12)) ) & 0x10; - vuint32m2_t xhr_1 = __riscv_vsrl_vv_u32m2(vqh, vt_2, vl); - vuint32m2_t xha_1 = __riscv_vand_vx_u32m2(xhr_1, 0x10, vl); - - // narrowing - vuint16m1_t xhc_0 = __riscv_vncvt_x_x_w_u16m1(xha_0, vl); - vuint8mf2_t xh_0 = __riscv_vncvt_x_x_w_u8mf2(xhc_0, vl); - - vuint16m1_t xhc_1 = __riscv_vncvt_x_x_w_u16m1(xha_1, vl); - vuint8mf2_t xh_1 = __riscv_vncvt_x_x_w_u8mf2(xhc_1, vl); - - // load - vuint8mf2_t tx = __riscv_vle8_v_u8mf2(x[ib].qs, vl); - - vint8mf2_t y0 = __riscv_vle8_v_i8mf2(y[ib].qs, vl); - vint8mf2_t y1 = __riscv_vle8_v_i8mf2(y[ib].qs+16, vl); - - vuint8mf2_t x_at = __riscv_vand_vx_u8mf2(tx, 0x0F, vl); - vuint8mf2_t x_lt = __riscv_vsrl_vx_u8mf2(tx, 0x04, vl); - - vuint8mf2_t x_a = __riscv_vor_vv_u8mf2(x_at, xh_0, vl); - vuint8mf2_t x_l = __riscv_vor_vv_u8mf2(x_lt, xh_1, vl); - - vint8mf2_t v0 = __riscv_vreinterpret_v_u8mf2_i8mf2(x_a); - vint8mf2_t v1 = __riscv_vreinterpret_v_u8mf2_i8mf2(x_l); - - vint16m1_t vec_mul1 = __riscv_vwmul_vv_i16m1(v0, y0, vl); - vint16m1_t vec_mul2 = __riscv_vwmul_vv_i16m1(v1, y1, vl); - - vint32m1_t vec_zero = __riscv_vmv_v_x_i32m1(0, vl); - - vint32m1_t vs1 = __riscv_vwredsum_vs_i16m1_i32m1(vec_mul1, vec_zero, vl); - vint32m1_t vs2 = __riscv_vwredsum_vs_i16m1_i32m1(vec_mul2, vs1, vl); - - int sumi = __riscv_vmv_x_s_i32m1_i32(vs2); + vl = qk; + vbool4_t qh = __riscv_vlm_v_b4(x[ib].qh, vl); + vint8m2_t v0f = __riscv_vor_vx_i8m2_mu(qh, v0c, v0c, 0x10, vl); + vint8m2_t v1 = __riscv_vle8_v_i8m2(y[ib].qs, vl); + vint16m4_t mul = __riscv_vwmul_vv_i16m4(v0f, v1, vl); + vint32m1_t zero = __riscv_vmv_v_x_i32m1(0, vl); + vint32m1_t sum = __riscv_vwredsum_vs_i16m4_i32m1(mul, zero, vl); + int32_t sumi = __riscv_vmv_x_s_i32m1_i32(sum); sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d))*sumi + GGML_FP16_TO_FP32(x[ib].m)*GGML_FP16_TO_FP32(y[ib].s); } @@ -3332,7 +3594,7 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * restrict s, size_t bs, const void * r *s = sumf; } -void ggml_vec_dot_q8_0_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) { +void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) { const int qk = QK8_0; const int nb = n / qk; @@ -3347,24 +3609,24 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * restrict s, size_t bs, const void * r UNUSED(by); UNUSED(bs); - const block_q8_0 * restrict x = vx; - const block_q8_0 * restrict y = vy; + const block_q8_0 * GGML_RESTRICT x = vx; + const block_q8_0 * GGML_RESTRICT y = vy; #if defined(__ARM_FEATURE_MATMUL_INT8) if (nrc == 2) { - const block_q8_0 * restrict vx0 = vx; - const block_q8_0 * restrict vx1 = (const block_q8_0 *) ((const uint8_t*)vx + bx); - const block_q8_0 * restrict vy0 = vy; - const block_q8_0 * restrict vy1 = (const block_q8_0 *) ((const uint8_t*)vy + by); + const block_q8_0 * GGML_RESTRICT vx0 = vx; + const block_q8_0 * GGML_RESTRICT vx1 = (const block_q8_0 *) ((const uint8_t*)vx + bx); + const block_q8_0 * GGML_RESTRICT vy0 = vy; + const block_q8_0 * GGML_RESTRICT vy1 = (const block_q8_0 *) ((const uint8_t*)vy + by); float32x4_t sumv0 = vdupq_n_f32(0.0f); for (int i = 0; i < nb; i++) { - const block_q8_0 * restrict b_x0 = &vx0[i]; - const block_q8_0 * restrict b_y0 = &vy0[i]; + const block_q8_0 * GGML_RESTRICT b_x0 = &vx0[i]; + const block_q8_0 * GGML_RESTRICT b_y0 = &vy0[i]; - const block_q8_0 * restrict b_x1 = &vx1[i]; - const block_q8_0 * restrict b_y1 = &vy1[i]; + const block_q8_0 * GGML_RESTRICT b_x1 = &vx1[i]; + const block_q8_0 * GGML_RESTRICT b_y1 = &vy1[i]; const int8x16_t x0_l = vld1q_s8(b_x0->qs); const int8x16_t x0_h = vld1q_s8(b_x0->qs + 16); @@ -3429,10 +3691,10 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * restrict s, size_t bs, const void * r const svbool_t pl16 = svptrue_pat_b32(SV_VL4); for (; ib + 1 < nb; ib += 2) { - const block_q8_0 * restrict x0 = &x[ib + 0]; - const block_q8_0 * restrict x1 = &x[ib + 1]; - const block_q8_0 * restrict y0 = &y[ib + 0]; - const block_q8_0 * restrict y1 = &y[ib + 1]; + const block_q8_0 * GGML_RESTRICT x0 = &x[ib + 0]; + const block_q8_0 * GGML_RESTRICT x1 = &x[ib + 1]; + const block_q8_0 * GGML_RESTRICT y0 = &y[ib + 0]; + const block_q8_0 * GGML_RESTRICT y1 = &y[ib + 1]; // load x const svint8_t qx0_0 = svld1_s8(ph16, x0->qs); @@ -3460,10 +3722,10 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * restrict s, size_t bs, const void * r { //printf("sve256"); for (; ib + 1 < nb; ib += 2) { - const block_q8_0 * restrict x0 = &x[ib + 0]; - const block_q8_0 * restrict x1 = &x[ib + 1]; - const block_q8_0 * restrict y0 = &y[ib + 0]; - const block_q8_0 * restrict y1 = &y[ib + 1]; + const block_q8_0 * GGML_RESTRICT x0 = &x[ib + 0]; + const block_q8_0 * GGML_RESTRICT x1 = &x[ib + 1]; + const block_q8_0 * GGML_RESTRICT y0 = &y[ib + 0]; + const block_q8_0 * GGML_RESTRICT y1 = &y[ib + 1]; // load x const svint8_t qx0 = svld1_s8(svptrue_b8(), x0->qs); @@ -3496,10 +3758,10 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * restrict s, size_t bs, const void * r svfloat32_t sumv00 = svdup_n_f32(0.0f); for (; ib + 1 < nb; ib += 2) { - const block_q8_0 * restrict x0 = &x[ib + 0]; - const block_q8_0 * restrict x1 = &x[ib + 1]; - const block_q8_0 * restrict y0 = &y[ib + 0]; - const block_q8_0 * restrict y1 = &y[ib + 1]; + const block_q8_0 * GGML_RESTRICT x0 = &x[ib + 0]; + const block_q8_0 * GGML_RESTRICT x1 = &x[ib + 1]; + const block_q8_0 * GGML_RESTRICT y0 = &y[ib + 0]; + const block_q8_0 * GGML_RESTRICT y1 = &y[ib + 1]; //load 32 int8_t in first half of vector and put another 32 int8_t in second vector lower bits // and add them to make one 64 element vector @@ -3539,10 +3801,10 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * restrict s, size_t bs, const void * r float32x4_t sumv1 = vdupq_n_f32(0.0f); for (; ib + 1 < nb; ib += 2) { - const block_q8_0 * restrict x0 = &x[ib + 0]; - const block_q8_0 * restrict x1 = &x[ib + 1]; - const block_q8_0 * restrict y0 = &y[ib + 0]; - const block_q8_0 * restrict y1 = &y[ib + 1]; + const block_q8_0 * GGML_RESTRICT x0 = &x[ib + 0]; + const block_q8_0 * GGML_RESTRICT x1 = &x[ib + 1]; + const block_q8_0 * GGML_RESTRICT y0 = &y[ib + 0]; + const block_q8_0 * GGML_RESTRICT y1 = &y[ib + 1]; const int8x16_t x0_0 = vld1q_s8(x0->qs); const int8x16_t x0_1 = vld1q_s8(x0->qs + 16); @@ -3565,6 +3827,45 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * restrict s, size_t bs, const void * r } sumf = vaddvq_f32(sumv0) + vaddvq_f32(sumv1); +#elif defined __wasm_simd128__ + v128_t sumv = wasm_f32x4_splat(0.0f); + + for (; ib < nb; ++ib) { + const block_q8_0 * GGML_RESTRICT x0 = &x[ib]; + const block_q8_0 * GGML_RESTRICT y0 = &y[ib]; + + const v128_t x0_0 = wasm_v128_load(x0->qs); + const v128_t x0_1 = wasm_v128_load(x0->qs + 16); + const v128_t y0_0 = wasm_v128_load(y0->qs); + const v128_t y0_1 = wasm_v128_load(y0->qs + 16); + + // Extend 8-bit to 16-bit + const v128_t x0_0l = wasm_i16x8_extend_low_i8x16(x0_0); + const v128_t x0_0h = wasm_i16x8_extend_high_i8x16(x0_0); + const v128_t x0_1l = wasm_i16x8_extend_low_i8x16(x0_1); + const v128_t x0_1h = wasm_i16x8_extend_high_i8x16(x0_1); + + const v128_t y0_0l = wasm_i16x8_extend_low_i8x16(y0_0); + const v128_t y0_0h = wasm_i16x8_extend_high_i8x16(y0_0); + const v128_t y0_1l = wasm_i16x8_extend_low_i8x16(y0_1); + const v128_t y0_1h = wasm_i16x8_extend_high_i8x16(y0_1); + + // Compute dot products + const v128_t dx0_0 = wasm_i32x4_dot_i16x8(x0_0l, y0_0l); + const v128_t dx0_1 = wasm_i32x4_dot_i16x8(x0_0h, y0_0h); + const v128_t dx1_0 = wasm_i32x4_dot_i16x8(x0_1l, y0_1l); + const v128_t dx1_1 = wasm_i32x4_dot_i16x8(x0_1h, y0_1h); + + // Sum all dot products + const v128_t sum_dots = wasm_i32x4_add(wasm_i32x4_add(dx0_0, dx0_1), wasm_i32x4_add(dx1_0, dx1_1)); + + // Convert to float and accumulate + const float scale = GGML_FP16_TO_FP32(x0->d) * GGML_FP16_TO_FP32(y0->d); + sumv = wasm_f32x4_add(sumv, wasm_f32x4_mul(wasm_f32x4_convert_i32x4(sum_dots), wasm_f32x4_splat(scale))); + } + + sumf = wasm_f32x4_extract_lane(sumv, 0) + wasm_f32x4_extract_lane(sumv, 1) + + wasm_f32x4_extract_lane(sumv, 2) + wasm_f32x4_extract_lane(sumv, 3); #elif defined(__AVX2__) // Initialize accumulator with zeros __m256 acc = _mm256_setzero_ps(); @@ -3603,17 +3904,17 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * restrict s, size_t bs, const void * r sumf = hsum_float_8(accum); #elif defined(__riscv_v_intrinsic) - size_t vl = __riscv_vsetvl_e8m1(qk); + size_t vl = qk; for (; ib < nb; ++ib) { // load elements - vint8m1_t bx_0 = __riscv_vle8_v_i8m1(x[ib].qs, vl); - vint8m1_t by_0 = __riscv_vle8_v_i8m1(y[ib].qs, vl); + vint8m2_t bx_0 = __riscv_vle8_v_i8m2(x[ib].qs, vl); + vint8m2_t by_0 = __riscv_vle8_v_i8m2(y[ib].qs, vl); - vint16m2_t vw_mul = __riscv_vwmul_vv_i16m2(bx_0, by_0, vl); + vint16m4_t vw_mul = __riscv_vwmul_vv_i16m4(bx_0, by_0, vl); vint32m1_t v_zero = __riscv_vmv_v_x_i32m1(0, vl); - vint32m1_t v_sum = __riscv_vwredsum_vs_i16m2_i32m1(vw_mul, v_zero, vl); + vint32m1_t v_sum = __riscv_vwredsum_vs_i16m4_i32m1(vw_mul, v_zero, vl); int sumi = __riscv_vmv_x_s_i32m1_i32(v_sum); @@ -3678,6 +3979,27 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * restrict s, size_t bs, const void * r } sumf = hsum_float_8(acc); +#elif defined(__VXE__) || defined(__VXE2__) + __vector float acc = vec_splats(0.0f); + +#pragma GCC unroll 8 + for (; ib < nb; ++ib) { + __builtin_prefetch(x[ib].qs, 0, 1); + __builtin_prefetch(y[ib].qs, 0, 1); + + const int8x16_t v_xl = vec_xl(0 , x[ib].qs); + const int8x16_t v_xh = vec_xl(QK8_0/2, x[ib].qs); + const int8x16_t v_yl = vec_xl(0 , y[ib].qs); + const int8x16_t v_yh = vec_xl(QK8_0/2, y[ib].qs); + + const int32x4_t v_xy_ = ggml_vec_dot(ggml_vec_dot(vec_splats(0), v_xl, v_yl), v_xh, v_yh); + const float32x4_t v_xy = vec_float(v_xy_); + const float32x4_t v_d = vec_splats(GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d)); + + acc = vec_madd(v_xy, v_d, acc); + } + + sumf = acc[0] + acc[1] + acc[2] + acc[3]; #endif for (; ib < nb; ++ib) { int sumi = 0; @@ -3692,15 +4014,15 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * restrict s, size_t bs, const void * r *s = sumf; } -void ggml_vec_dot_tq1_0_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) { +void ggml_vec_dot_tq1_0_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) { assert(nrc == 1); UNUSED(nrc); UNUSED(bx); UNUSED(by); UNUSED(bs); - const block_tq1_0 * restrict x = vx; - const block_q8_K * restrict y = vy; + const block_tq1_0 * GGML_RESTRICT x = vx; + const block_q8_K * GGML_RESTRICT y = vy; const int nb = n / QK_K; @@ -4015,15 +4337,15 @@ void ggml_vec_dot_tq1_0_q8_K(int n, float * restrict s, size_t bs, const void * #endif } -void ggml_vec_dot_tq2_0_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) { +void ggml_vec_dot_tq2_0_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) { assert(nrc == 1); UNUSED(nrc); UNUSED(bx); UNUSED(by); UNUSED(bs); - const block_tq2_0 * restrict x = vx; - const block_q8_K * restrict y = vy; + const block_tq2_0 * GGML_RESTRICT x = vx; + const block_q8_K * GGML_RESTRICT y = vy; const int nb = n / QK_K; @@ -4187,19 +4509,264 @@ void ggml_vec_dot_tq2_0_q8_K(int n, float * restrict s, size_t bs, const void * #endif } -void ggml_vec_dot_q2_K_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) { +void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) { assert(nrc == 1); UNUSED(nrc); UNUSED(bx); UNUSED(by); UNUSED(bs); - const block_q2_K * restrict x = vx; - const block_q8_K * restrict y = vy; + const block_q2_K * GGML_RESTRICT x = vx; + const block_q8_K * GGML_RESTRICT y = vy; const int nb = n / QK_K; -#ifdef __ARM_NEON +#ifdef __ARM_FEATURE_SVE + const int vector_length = svcntb()*8; + const svuint8_t m3s = svdup_n_u8(0x3); + const svuint32_t m4s = svdup_n_u32(0xF); + const svint32_t vzero_sv = svdup_n_s32(0); + svfloat32_t acc_sum = svdup_n_f32(0); + svbool_t pred_s32 = svptrue_pat_b32(SV_VL4); + + switch (vector_length) { + case 128: + for (int i = 0; i < nb; ++i) { + const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); + svfloat32_t d_broad = svdup_n_f32((float32_t)d); + const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin); + svfloat32_t dmin_broad = svdup_n_f32((float32_t)dmin); + + const uint8_t * GGML_RESTRICT q2 = x[i].qs; + const int8_t * GGML_RESTRICT q8_sv = y[i].qs; + const uint8_t * GGML_RESTRICT sc = x[i].scales; + + svuint32_t mins_and_scales_sve = svld1ub_u32(svptrue_b32(), sc); + const svint32_t mins_sv_1 = svreinterpret_s32_u32(svlsr_n_u32_x(svptrue_b32(), mins_and_scales_sve, 4)); + + mins_and_scales_sve = svld1ub_u32(svptrue_b32(), sc+4); + const svint32_t mins_sv_2 = svreinterpret_s32_u32(svlsr_n_u32_x(svptrue_b32(), mins_and_scales_sve, 4)); + + svint32_t q8sums_sv_1 = svld1sh_s32(svptrue_b32(), y[i].bsums); + svint32_t q8sums_sv_2 = svld1sh_s32(svptrue_b32(), y[i].bsums+4); + + const svint32_t s0 = svadd_s32_x(svptrue_b32(), svmul_s32_x(svptrue_b32(), mins_sv_1, q8sums_sv_1), svmul_s32_x(svptrue_b32(), mins_sv_2, q8sums_sv_2)); + + mins_and_scales_sve = svld1ub_u32(svptrue_b32(), sc+8); + const svint32_t mins_sv_3 = svreinterpret_s32_u32(svlsr_n_u32_x(svptrue_b32(), mins_and_scales_sve, 4)); + + mins_and_scales_sve = svld1ub_u32(svptrue_b32(), sc+12); + const svint32_t mins_sv_4 = svreinterpret_s32_u32(svlsr_n_u32_x(svptrue_b32(), mins_and_scales_sve, 4)); + + q8sums_sv_1 = svld1sh_s32(svptrue_b32(), y[i].bsums+8); + q8sums_sv_2 = svld1sh_s32(svptrue_b32(), y[i].bsums+12); + + svint32_t s1 = svadd_s32_x(svptrue_b32(), svmul_s32_x(svptrue_b32(), mins_sv_3, q8sums_sv_1), svmul_s32_x(svptrue_b32(), mins_sv_4, q8sums_sv_2)); + + svfloat32_t temp = svcvt_f32_s32_x(svptrue_b32(), svadd_s32_x(svptrue_b32(), s0, s1)); + + acc_sum = svmla_f32_m(svptrue_b32(), acc_sum, temp, dmin_broad); + + svint32_t sumi1 = svdup_n_s32(0); + + { + const svuint8_t q2bits_1 = svld1_u8(svptrue_b8(), q2); + svint8_t q2bytes_sv = svreinterpret_s8_u8(svand_u8_x(svptrue_b8(), q2bits_1, m3s)); + svint8_t q8bytes_sv = svld1_s8(svptrue_b8(), q8_sv); q8_sv += 16; + const svint32_t scales_sv = svreinterpret_s32_u32(svand_u32_m(svptrue_b32(), svld1ub_u32(svptrue_b32(), sc), m4s)); + + sumi1 = svmla_s32_m(svptrue_b32(), sumi1, svdot_s32(vzero_sv, q2bytes_sv, q8bytes_sv), svdup_lane_s32(scales_sv, 0)); + + const svuint8_t q2bits_3 = svld1_u8(svptrue_b8(), q2+16); + q2bytes_sv = svreinterpret_s8_u8(svand_u8_x(svptrue_b8(), q2bits_3, m3s)); + q8bytes_sv = svld1_s8(svptrue_b8(), q8_sv); q8_sv += 16; + + sumi1 = svmla_s32_m(svptrue_b32(), sumi1, svdot_s32(vzero_sv, q2bytes_sv, q8bytes_sv), svdup_lane_s32(scales_sv, 1)); + + q2bytes_sv = svreinterpret_s8_u8(svand_u8_x(svptrue_b8(), svlsr_n_u8_x(svptrue_b8(), q2bits_1, 2), m3s)); + q8bytes_sv = svld1_s8(svptrue_b8(), q8_sv); q8_sv += 16; + + sumi1 = svmla_s32_m(svptrue_b32(), sumi1, svdot_s32(vzero_sv, q2bytes_sv, q8bytes_sv), svdup_lane_s32(scales_sv, 2)); + + q2bytes_sv = svreinterpret_s8_u8(svand_u8_x(svptrue_b8(), svlsr_n_u8_x(svptrue_b8(), q2bits_3, 2), m3s)); + q8bytes_sv = svld1_s8(svptrue_b8(), q8_sv); q8_sv += 16; + + sumi1 = svmla_s32_m(svptrue_b32(), sumi1, svdot_s32(vzero_sv, q2bytes_sv, q8bytes_sv), svdup_lane_s32(scales_sv, 3)); + + + const svint32_t scales_sv_1 = svreinterpret_s32_u32(svand_u32_m(svptrue_b32(), svld1ub_u32(svptrue_b32(), sc+4), m4s)); + + q2bytes_sv = svreinterpret_s8_u8(svand_u8_x(svptrue_b8(), svlsr_n_u8_x(svptrue_b8(), q2bits_1, 4), m3s)); + q8bytes_sv = svld1_s8(svptrue_b8(), q8_sv); q8_sv += 16; + + sumi1 = svmla_s32_m(svptrue_b32(), sumi1, svdot_s32(vzero_sv, q2bytes_sv, q8bytes_sv), svdup_lane_s32(scales_sv_1, 0)); + + q2bytes_sv = svreinterpret_s8_u8(svand_u8_x(svptrue_b8(), svlsr_n_u8_x(svptrue_b8(), q2bits_3, 4), m3s)); + q8bytes_sv = svld1_s8(svptrue_b8(), q8_sv); q8_sv += 16; + + sumi1 = svmla_s32_m(svptrue_b32(), sumi1, svdot_s32(vzero_sv, q2bytes_sv, q8bytes_sv), svdup_lane_s32(scales_sv_1, 1)); + + q2bytes_sv = svreinterpret_s8_u8(svand_u8_x(svptrue_b8(), svlsr_n_u8_x(svptrue_b8(), q2bits_1, 6), m3s)); + q8bytes_sv = svld1_s8(svptrue_b8(), q8_sv); q8_sv += 16; + + sumi1 = svmla_s32_m(svptrue_b32(), sumi1, svdot_s32(vzero_sv, q2bytes_sv, q8bytes_sv), svdup_lane_s32(scales_sv_1, 2)); + + q2bytes_sv = svreinterpret_s8_u8(svand_u8_x(svptrue_b8(), svlsr_n_u8_x(svptrue_b8(), q2bits_3, 6), m3s)); + q8bytes_sv = svld1_s8(svptrue_b8(), q8_sv); q8_sv += 16; + + sumi1 = svmla_s32_m(svptrue_b32(), sumi1, svdot_s32(vzero_sv, q2bytes_sv, q8bytes_sv), svdup_lane_s32(scales_sv_1, 3)); + + //------------------------------- + + q2 += 32; + const svint32_t scales_sv_2 = svreinterpret_s32_u32(svand_u32_m(svptrue_b32(), svld1ub_u32(svptrue_b32(), sc+8), m4s)); + const svuint8_t q2bits_2 = svld1_u8(svptrue_b8(), q2); + + q2bytes_sv = svreinterpret_s8_u8(svand_u8_x(svptrue_b8(), q2bits_2, m3s)); + q8bytes_sv = svld1_s8(svptrue_b8(), q8_sv); q8_sv += 16; + + sumi1 = svmla_s32_m(svptrue_b32(), sumi1, svdot_s32(vzero_sv, q2bytes_sv, q8bytes_sv), svdup_lane_s32(scales_sv_2, 0)); + + const svuint8_t q2bits_4 = svld1_u8(svptrue_b8(), q2+16); + q2bytes_sv = svreinterpret_s8_u8(svand_u8_x(svptrue_b8(), q2bits_4, m3s)); + q8bytes_sv = svld1_s8(svptrue_b8(), q8_sv); q8_sv += 16; + + sumi1 = svmla_s32_m(svptrue_b32(), sumi1, svdot_s32(vzero_sv, q2bytes_sv, q8bytes_sv), svdup_lane_s32(scales_sv_2, 1)); + + + q2bytes_sv = svreinterpret_s8_u8(svand_u8_x(svptrue_b8(), svlsr_n_u8_x(svptrue_b8(), q2bits_2, 2), m3s)); + q8bytes_sv = svld1_s8(svptrue_b8(), q8_sv); q8_sv += 16; + + sumi1 = svmla_s32_m(svptrue_b32(), sumi1, svdot_s32(vzero_sv, q2bytes_sv, q8bytes_sv), svdup_lane_s32(scales_sv_2, 2)); + + q2bytes_sv = svreinterpret_s8_u8(svand_u8_x(svptrue_b8(), svlsr_n_u8_x(svptrue_b8(), q2bits_4, 2), m3s)); + q8bytes_sv = svld1_s8(svptrue_b8(), q8_sv); q8_sv += 16; + + sumi1 = svmla_s32_m(svptrue_b32(), sumi1, svdot_s32(vzero_sv, q2bytes_sv, q8bytes_sv), svdup_lane_s32(scales_sv_2, 3)); + + + const svint32_t scales_sv_3 = svreinterpret_s32_u32(svand_u32_m(svptrue_b32(), svld1ub_u32(svptrue_b32(), sc+12), m4s)); + + q2bytes_sv = svreinterpret_s8_u8(svand_u8_x(svptrue_b8(), svlsr_n_u8_x(svptrue_b8(), q2bits_2, 4), m3s)); + q8bytes_sv = svld1_s8(svptrue_b8(), q8_sv); q8_sv += 16; + + sumi1 = svmla_s32_m(svptrue_b32(), sumi1, svdot_s32(vzero_sv, q2bytes_sv, q8bytes_sv), svdup_lane_s32(scales_sv_3, 0)); + + q2bytes_sv = svreinterpret_s8_u8(svand_u8_x(svptrue_b8(), svlsr_n_u8_x(svptrue_b8(), q2bits_4, 4), m3s)); + q8bytes_sv = svld1_s8(svptrue_b8(), q8_sv); q8_sv += 16; + + sumi1 = svmla_s32_m(svptrue_b32(), sumi1, svdot_s32(vzero_sv, q2bytes_sv, q8bytes_sv), svdup_lane_s32(scales_sv_3, 1)); + + + + q2bytes_sv = svreinterpret_s8_u8(svand_u8_x(svptrue_b8(), svlsr_n_u8_x(svptrue_b8(), q2bits_2, 6), m3s)); + q8bytes_sv = svld1_s8(svptrue_b8(), q8_sv); q8_sv += 16; + + sumi1 = svmla_s32_m(svptrue_b32(), sumi1, svdot_s32(vzero_sv, q2bytes_sv, q8bytes_sv), svdup_lane_s32(scales_sv_3, 2)); + + q2bytes_sv = svreinterpret_s8_u8(svand_u8_x(svptrue_b8(), svlsr_n_u8_x(svptrue_b8(), q2bits_4, 6), m3s)); + q8bytes_sv = svld1_s8(svptrue_b8(), q8_sv); q8_sv += 16; + + sumi1 = svmla_s32_m(svptrue_b32(), sumi1, svdot_s32(vzero_sv, q2bytes_sv, q8bytes_sv), svdup_lane_s32(scales_sv_3, 3)); + } + acc_sum = svmla_f32_m(svptrue_b32(), acc_sum, svcvt_f32_s32_x(svptrue_b32(), sumi1), d_broad); + } + *s = svaddv_f32(svptrue_b32(), acc_sum); + break; + + case 256: + case 512: + for (int i = 0; i < nb; ++i) { + const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); + svfloat32_t d_broad = svdup_n_f32((float32_t)d); + const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin); + svfloat32_t dmin_broad = svdup_n_f32((float32_t)dmin); + + const uint8_t * GGML_RESTRICT q2 = x[i].qs; + const int8_t * GGML_RESTRICT q8_sv = y[i].qs; + const uint8_t * GGML_RESTRICT sc = x[i].scales; + + const svuint32_t mins_and_scales_sve = svld1ub_u32(svptrue_pat_b32(SV_VL8), sc); sc += 8; + const svint32_t scales_sv = svreinterpret_s32_u32(svand_u32_m(svptrue_pat_b32(SV_VL8), mins_and_scales_sve, m4s)); + const svint32_t mins_sv_1 = svreinterpret_s32_u32(svlsr_n_u32_x(svptrue_pat_b32(SV_VL8), mins_and_scales_sve, 4)); + svint32_t q8sums_sv_1 = svld1sh_s32(svptrue_pat_b32(SV_VL8), y[i].bsums); + + const svuint32_t mins_and_scales_sve_1 = svld1ub_u32(svptrue_pat_b32(SV_VL8), sc); + const svint32_t scales_sv_1 = svreinterpret_s32_u32(svand_u32_m(svptrue_pat_b32(SV_VL8), mins_and_scales_sve_1, m4s)); + const svint32_t mins_sv_2 = svreinterpret_s32_u32(svlsr_n_u32_x(svptrue_pat_b32(SV_VL8), mins_and_scales_sve_1, 4)); + + svint32_t q8sums_sv_2 = svld1sh_s32(svptrue_pat_b32(SV_VL8), y[i].bsums+8); + + svfloat32_t temp = svcvt_f32_s32_x(svptrue_pat_b32(SV_VL8), svadd_s32_x(svptrue_pat_b32(SV_VL8), svmul_s32_x(svptrue_pat_b32(SV_VL8), mins_sv_1, q8sums_sv_1), svmul_s32_x(svptrue_pat_b32(SV_VL8), mins_sv_2, q8sums_sv_2))); + + acc_sum = svmla_f32_m(svptrue_pat_b32(SV_VL8), acc_sum, temp, dmin_broad); + + svint32_t sumi1 = svdup_n_s32(0); + + { + const svuint8_t q2bits_1 = svld1_u8(svptrue_pat_b8(SV_VL32), q2); + svint8_t q2bytes_sv = svreinterpret_s8_u8(svand_u8_m(svptrue_pat_b8(SV_VL32), q2bits_1, m3s)); + svint8_t q8bytes_sv = svld1_s8(svptrue_pat_b8(SV_VL32), q8_sv); q8_sv += 32; + + svint32_t scale_1 = svsel(pred_s32, svdup_lane_s32(scales_sv, 0), svdup_lane_s32(scales_sv, 1)); + sumi1 = svmla_s32_m(svptrue_pat_b32(SV_VL8), sumi1, svdot_s32(vzero_sv, q2bytes_sv, q8bytes_sv), scale_1); + + q2bytes_sv = svreinterpret_s8_u8(svand_u8_m(svptrue_pat_b8(SV_VL32), svlsr_n_u8_x(svptrue_pat_b8(SV_VL32), q2bits_1, 2), m3s)); + q8bytes_sv = svld1_s8(svptrue_pat_b8(SV_VL32), q8_sv); q8_sv += 32; + + svint32_t scale_2 = svsel(pred_s32, svdup_lane_s32(scales_sv, 2), svdup_lane_s32(scales_sv, 3)); + sumi1 = svmla_s32_m(svptrue_pat_b32(SV_VL8), sumi1, svdot_s32(svdup_n_s32(0), q2bytes_sv, q8bytes_sv), scale_2); + + q2bytes_sv = svreinterpret_s8_u8(svand_u8_m(svptrue_pat_b8(SV_VL32), svlsr_n_u8_x(svptrue_pat_b8(SV_VL32), q2bits_1, 4), m3s)); + q8bytes_sv = svld1_s8(svptrue_pat_b8(SV_VL32), q8_sv); q8_sv += 32; + + scale_1 = svsel(pred_s32, svdup_lane_s32(scales_sv, 4), svdup_lane_s32(scales_sv, 5)); + sumi1 = svmla_s32_m(svptrue_pat_b32(SV_VL8), sumi1, svdot_s32(vzero_sv, q2bytes_sv, q8bytes_sv), scale_1); + + q2bytes_sv = svreinterpret_s8_u8(svand_u8_m(svptrue_pat_b8(SV_VL32), svlsr_n_u8_x(svptrue_pat_b8(SV_VL32), q2bits_1, 6), m3s)); + q8bytes_sv = svld1_s8(svptrue_pat_b8(SV_VL32), q8_sv); q8_sv += 32; + + scale_2 = svsel(pred_s32, svdup_lane_s32(scales_sv, 6), svdup_lane_s32(scales_sv, 7)); + sumi1 = svmla_s32_m(svptrue_pat_b32(SV_VL8), sumi1, svdot_s32(vzero_sv, q2bytes_sv, q8bytes_sv), scale_2); + + q2 += 32; + + const svuint8_t q2bits_2 = svld1_u8(svptrue_pat_b8(SV_VL32), q2); + q2bytes_sv = svreinterpret_s8_u8(svand_u8_m(svptrue_pat_b8(SV_VL32), q2bits_2, m3s)); + q8bytes_sv = svld1_s8(svptrue_pat_b8(SV_VL32), q8_sv); q8_sv += 32; + + scale_1 = svsel(pred_s32, svdup_lane_s32(scales_sv_1, 0), svdup_lane_s32(scales_sv_1, 1)); + sumi1 = svmla_s32_m(svptrue_pat_b32(SV_VL8), sumi1, svdot_s32(vzero_sv, q2bytes_sv, q8bytes_sv), scale_1); + + q2bytes_sv = svreinterpret_s8_u8(svand_u8_m(svptrue_pat_b8(SV_VL32), svlsr_n_u8_x(svptrue_pat_b8(SV_VL32), q2bits_2, 2), m3s)); + q8bytes_sv = svld1_s8(svptrue_pat_b8(SV_VL32), q8_sv); q8_sv += 32; + + scale_2 = svsel(pred_s32, svdup_lane_s32(scales_sv_1, 2), svdup_lane_s32(scales_sv_1, 3)); + sumi1 = svmla_s32_m(svptrue_pat_b32(SV_VL8), sumi1, svdot_s32(vzero_sv, q2bytes_sv, q8bytes_sv), scale_2); + + q2bytes_sv = svreinterpret_s8_u8(svand_u8_m(svptrue_pat_b8(SV_VL32), svlsr_n_u8_x(svptrue_pat_b8(SV_VL32), q2bits_2, 4), m3s)); + q8bytes_sv = svld1_s8(svptrue_pat_b8(SV_VL32), q8_sv); q8_sv += 32; + + scale_1 = svsel(pred_s32, svdup_lane_s32(scales_sv_1, 4), svdup_lane_s32(scales_sv_1, 5)); + sumi1 = svmla_s32_m(svptrue_pat_b32(SV_VL8), sumi1, svdot_s32(vzero_sv, q2bytes_sv, q8bytes_sv), scale_1); + + q2bytes_sv = svreinterpret_s8_u8(svand_u8_m(svptrue_pat_b8(SV_VL32), svlsr_n_u8_x(svptrue_pat_b8(SV_VL32), q2bits_2, 6), m3s)); + q8bytes_sv = svld1_s8(svptrue_pat_b8(SV_VL32), q8_sv); q8_sv += 32; + + scale_2 = svsel(pred_s32, svdup_lane_s32(scales_sv_1, 6), svdup_lane_s32(scales_sv_1, 7)); + sumi1 = svmla_s32_m(svptrue_pat_b32(SV_VL8), sumi1, svdot_s32(vzero_sv, q2bytes_sv, q8bytes_sv), scale_2); + } + acc_sum = svmla_f32_m(svptrue_pat_b32(SV_VL8), acc_sum, svcvt_f32_s32_x(svptrue_pat_b32(SV_VL8), sumi1), d_broad); + } + *s = svaddv_f32(svptrue_pat_b32(SV_VL8), acc_sum); + break; + + default: + assert(false && "Unsupported vector length"); + break; + } + +#elif __ARM_NEON const uint8x16_t m3 = vdupq_n_u8(0x3); const uint8x16_t m4 = vdupq_n_u8(0xF); @@ -4214,9 +4781,9 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * restrict s, size_t bs, const void * r const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin); - const uint8_t * restrict q2 = x[i].qs; - const int8_t * restrict q8 = y[i].qs; - const uint8_t * restrict sc = x[i].scales; + const uint8_t * GGML_RESTRICT q2 = x[i].qs; + const int8_t * GGML_RESTRICT q8 = y[i].qs; + const uint8_t * GGML_RESTRICT sc = x[i].scales; const uint8x16_t mins_and_scales = vld1q_u8(sc); const uint8x16_t scales = vandq_u8(mins_and_scales, m4); @@ -4279,8 +4846,8 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * restrict s, size_t bs, const void * r const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin); - const uint8_t * restrict q2 = x[i].qs; - const int8_t * restrict q8 = y[i].qs; + const uint8_t * GGML_RESTRICT q2 = x[i].qs; + const int8_t * GGML_RESTRICT q8 = y[i].qs; const __m128i mins_and_scales = _mm_loadu_si128((const __m128i*)x[i].scales); const __m128i scales8 = _mm_and_si128(mins_and_scales, m4); @@ -4346,8 +4913,8 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * restrict s, size_t bs, const void * r const float dall = y[i].d * GGML_FP16_TO_FP32(x[i].d); const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin); - const uint8_t * restrict q2 = x[i].qs; - const int8_t * restrict q8 = y[i].qs; + const uint8_t * GGML_RESTRICT q2 = x[i].qs; + const int8_t * GGML_RESTRICT q8 = y[i].qs; // load mins and scales from block_q2_K.scales[QK_K/16] const __m128i mins_and_scales = _mm_loadu_si128((const __m128i*)x[i].scales); @@ -4439,86 +5006,284 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * restrict s, size_t bs, const void * r *s = hsum_float_8(acc); -#elif defined __riscv_v_intrinsic - +#elif defined __wasm_simd128__ float sumf = 0; - uint8_t temp_01[32] = {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, - 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1}; for (int i = 0; i < nb; ++i) { - const uint8_t * q2 = x[i].qs; - const int8_t * q8 = y[i].qs; + const int8_t * q8 = y[i].qs; const uint8_t * sc = x[i].scales; - const float dall = y[i].d * GGML_FP16_TO_FP32(x[i].d); - const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin); + // Vectorized summs calculation + v128_t summs_vec = wasm_i32x4_splat(0); + { + v128_t sc_vec = wasm_v128_load(sc); + v128_t sc_upper = wasm_u8x16_shr(sc_vec, 4); - size_t vl = 16; + v128_t sc_low = wasm_u16x8_extend_low_u8x16(sc_upper); + v128_t sc_high = wasm_u16x8_extend_high_u8x16(sc_upper); - vuint8m1_t scales = __riscv_vle8_v_u8m1(sc, vl); - vuint8m1_t aux = __riscv_vand_vx_u8m1(scales, 0x0F, vl); + v128_t bsums1 = wasm_v128_load(&y[i].bsums[0]); + v128_t bsums2 = wasm_v128_load(&y[i].bsums[8]); - vint16m1_t q8sums = __riscv_vle16_v_i16m1(y[i].bsums, vl); + summs_vec = wasm_i32x4_add( + wasm_i32x4_add(wasm_i32x4_dot_i16x8(sc_low, bsums1), + wasm_i32x4_dot_i16x8(sc_high, bsums2)), + summs_vec + ); - vuint8mf2_t scales_2 = __riscv_vle8_v_u8mf2(sc, vl); - vuint8mf2_t mins8 = __riscv_vsrl_vx_u8mf2(scales_2, 0x4, vl); - vint16m1_t mins = __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(mins8, vl)); - vint32m2_t prod = __riscv_vwmul_vv_i32m2(q8sums, mins, vl); - vint32m1_t vsums = __riscv_vredsum_vs_i32m2_i32m1(prod, __riscv_vmv_v_x_i32m1(0, 1), vl); + summs_vec = wasm_i32x4_add(summs_vec, wasm_i32x4_shuffle(summs_vec, summs_vec, 2, 3, 0, 1)); + summs_vec = wasm_i32x4_add(summs_vec, wasm_i32x4_shuffle(summs_vec, summs_vec, 1, 0, 3, 2)); + } + int32_t summs = wasm_i32x4_extract_lane(summs_vec, 0); - sumf += dmin * __riscv_vmv_x_s_i32m1_i32(vsums); + // Vectorized isum calculation + int32_t isum = 0; + const uint8_t * sc_ptr = sc; + const int k_iters = QK_K/128; - vl = 32; + for (int k = 0; k < k_iters; ++k) { + v128_t isum_vec = wasm_i32x4_splat(0); + int shift = 0; - vint32m1_t vzero = __riscv_vmv_v_x_i32m1(0, 1); - vuint8m1_t v_b = __riscv_vle8_v_u8m1(temp_01, vl); + for (int j = 0; j < 4; ++j) { + const int d0 = (sc_ptr[0] & 0xF); + const int d1 = (sc_ptr[1] & 0xF); + sc_ptr += 2; - uint8_t is=0; - int isum=0; + // Process first 16 elements + v128_t q2_0 = wasm_v128_load(q2); + v128_t q8_0 = wasm_v128_load(q8); + v128_t q2_shift_0 = wasm_u8x16_shr(q2_0, shift); + v128_t q2_bits_0 = wasm_v128_and(q2_shift_0, wasm_i8x16_splat(0x03)); - for (int j = 0; j < QK_K/128; ++j) { - // load Q2 - vuint8m1_t q2_x = __riscv_vle8_v_u8m1(q2, vl); + // Process next 16 elements + v128_t q2_1 = wasm_v128_load(q2 + 16); + v128_t q8_1 = wasm_v128_load(q8 + 16); + v128_t q2_shift_1 = wasm_u8x16_shr(q2_1, shift); + v128_t q2_bits_1 = wasm_v128_and(q2_shift_1, wasm_i8x16_splat(0x03)); - vuint8m1_t q2_0 = __riscv_vand_vx_u8m1(q2_x, 0x03, vl); - vuint8m1_t q2_1 = __riscv_vand_vx_u8m1(__riscv_vsrl_vx_u8m1(q2_x, 0x2, vl), 0x03 , vl); - vuint8m1_t q2_2 = __riscv_vand_vx_u8m1(__riscv_vsrl_vx_u8m1(q2_x, 0x4, vl), 0x03 , vl); - vuint8m1_t q2_3 = __riscv_vand_vx_u8m1(__riscv_vsrl_vx_u8m1(q2_x, 0x6, vl), 0x03 , vl); + // Calculate dot products + v128_t p0 = wasm_i32x4_dot_i16x8( + wasm_i16x8_extend_low_i8x16(q8_0), + wasm_i16x8_extend_low_i8x16(q2_bits_0) + ); + v128_t p1 = wasm_i32x4_dot_i16x8( + wasm_i16x8_extend_high_i8x16(q8_0), + wasm_i16x8_extend_high_i8x16(q2_bits_0) + ); + v128_t p2 = wasm_i32x4_dot_i16x8( + wasm_i16x8_extend_low_i8x16(q8_1), + wasm_i16x8_extend_low_i8x16(q2_bits_1) + ); + v128_t p3 = wasm_i32x4_dot_i16x8( + wasm_i16x8_extend_high_i8x16(q8_1), + wasm_i16x8_extend_high_i8x16(q2_bits_1) + ); - // duplicate scale elements for product - vuint8m1_t sc0 = __riscv_vrgather_vv_u8m1(aux, __riscv_vadd_vx_u8m1(v_b, 0+is, vl), vl); - vuint8m1_t sc1 = __riscv_vrgather_vv_u8m1(aux, __riscv_vadd_vx_u8m1(v_b, 2+is, vl), vl); - vuint8m1_t sc2 = __riscv_vrgather_vv_u8m1(aux, __riscv_vadd_vx_u8m1(v_b, 4+is, vl), vl); - vuint8m1_t sc3 = __riscv_vrgather_vv_u8m1(aux, __riscv_vadd_vx_u8m1(v_b, 6+is, vl), vl); + // Accumulate scaled results + v128_t scaled = wasm_i32x4_add( + wasm_i32x4_mul(wasm_i32x4_add(p0, p1), wasm_i32x4_splat(d0)), + wasm_i32x4_mul(wasm_i32x4_add(p2, p3), wasm_i32x4_splat(d1)) + ); - vint16m2_t p0 = __riscv_vreinterpret_v_u16m2_i16m2(__riscv_vwmulu_vv_u16m2(q2_0, sc0, vl)); - vint16m2_t p1 = __riscv_vreinterpret_v_u16m2_i16m2(__riscv_vwmulu_vv_u16m2(q2_1, sc1, vl)); - vint16m2_t p2 = __riscv_vreinterpret_v_u16m2_i16m2(__riscv_vwmulu_vv_u16m2(q2_2, sc2, vl)); - vint16m2_t p3 = __riscv_vreinterpret_v_u16m2_i16m2(__riscv_vwmulu_vv_u16m2(q2_3, sc3, vl)); - - // load Q8 - vint8m1_t q8_0 = __riscv_vle8_v_i8m1(q8, vl); - vint8m1_t q8_1 = __riscv_vle8_v_i8m1(q8+32, vl); - vint8m1_t q8_2 = __riscv_vle8_v_i8m1(q8+64, vl); - vint8m1_t q8_3 = __riscv_vle8_v_i8m1(q8+96, vl); - - vint32m4_t s0 = __riscv_vwmul_vv_i32m4(p0, __riscv_vwcvt_x_x_v_i16m2(q8_0, vl), vl); - vint32m4_t s1 = __riscv_vwmul_vv_i32m4(p1, __riscv_vwcvt_x_x_v_i16m2(q8_1, vl), vl); - vint32m4_t s2 = __riscv_vwmul_vv_i32m4(p2, __riscv_vwcvt_x_x_v_i16m2(q8_2, vl), vl); - vint32m4_t s3 = __riscv_vwmul_vv_i32m4(p3, __riscv_vwcvt_x_x_v_i16m2(q8_3, vl), vl); - - vint32m1_t isum0 = __riscv_vredsum_vs_i32m4_i32m1(__riscv_vadd_vv_i32m4(s0, s1, vl), vzero, vl); - vint32m1_t isum1 = __riscv_vredsum_vs_i32m4_i32m1(__riscv_vadd_vv_i32m4(s2, s3, vl), isum0, vl); - - isum += __riscv_vmv_x_s_i32m1_i32(isum1); - - q2+=32; q8+=128; is=8; + isum_vec = wasm_i32x4_add(isum_vec, scaled); + q8 += 32; + shift += 2; + } + q2 += 32; + // Horizontal sum of isum_vec + isum_vec = wasm_i32x4_add(isum_vec, wasm_i32x4_shuffle(isum_vec, isum_vec, 2, 3, 0, 1)); + isum_vec = wasm_i32x4_add(isum_vec, wasm_i32x4_shuffle(isum_vec, isum_vec, 1, 0, 3, 2)); + isum += wasm_i32x4_extract_lane(isum_vec, 0); } - sumf += dall * isum; + const float dall = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d; + sumf += dall * isum - dmin * summs; + } + *s = sumf; + +#elif defined __riscv_v_intrinsic + + const int vector_length = __riscv_vlenb() * 8; + float sumf = 0; + + uint8_t temp_01[32] = { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, + 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1 }; + uint8_t atmp[16]; + + switch (vector_length) { + case 256: + for (int i = 0; i < nb; ++i) { + const uint8_t * q2 = x[i].qs; + const int8_t * q8 = y[i].qs; + const uint8_t * sc = x[i].scales; + + const float dall = y[i].d * GGML_FP16_TO_FP32(x[i].d); + const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin); + + size_t vl = 16; + + vuint8m1_t scales = __riscv_vle8_v_u8m1(sc, vl); + vuint8m1_t aux = __riscv_vand_vx_u8m1(scales, 0x0F, vl); + + vint16m1_t q8sums = __riscv_vle16_v_i16m1(y[i].bsums, vl); + + vuint8mf2_t scales_2 = __riscv_vle8_v_u8mf2(sc, vl); + vuint8mf2_t mins8 = __riscv_vsrl_vx_u8mf2(scales_2, 0x4, vl); + vint16m1_t mins = __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(mins8, vl)); + vint32m2_t prod = __riscv_vwmul_vv_i32m2(q8sums, mins, vl); + vint32m1_t vsums = __riscv_vredsum_vs_i32m2_i32m1(prod, __riscv_vmv_v_x_i32m1(0, 1), vl); + + sumf += dmin * __riscv_vmv_x_s_i32m1_i32(vsums); + + vl = 32; + + vint32m1_t vzero = __riscv_vmv_v_x_i32m1(0, 1); + vuint8m1_t v_b = __riscv_vle8_v_u8m1(temp_01, vl); + + uint8_t is = 0; + int isum = 0; + + for (int j = 0; j < QK_K / 128; ++j) { + // load Q2 + vuint8m1_t q2_x = __riscv_vle8_v_u8m1(q2, vl); + + vuint8m1_t q2_0 = __riscv_vand_vx_u8m1(q2_x, 0x03, vl); + vuint8m1_t q2_1 = __riscv_vand_vx_u8m1(__riscv_vsrl_vx_u8m1(q2_x, 0x2, vl), 0x03, vl); + vuint8m1_t q2_2 = __riscv_vand_vx_u8m1(__riscv_vsrl_vx_u8m1(q2_x, 0x4, vl), 0x03, vl); + vuint8m1_t q2_3 = __riscv_vand_vx_u8m1(__riscv_vsrl_vx_u8m1(q2_x, 0x6, vl), 0x03, vl); + + // duplicate scale elements for product + vuint8m1_t sc0 = __riscv_vrgather_vv_u8m1(aux, __riscv_vadd_vx_u8m1(v_b, 0 + is, vl), vl); + vuint8m1_t sc1 = __riscv_vrgather_vv_u8m1(aux, __riscv_vadd_vx_u8m1(v_b, 2 + is, vl), vl); + vuint8m1_t sc2 = __riscv_vrgather_vv_u8m1(aux, __riscv_vadd_vx_u8m1(v_b, 4 + is, vl), vl); + vuint8m1_t sc3 = __riscv_vrgather_vv_u8m1(aux, __riscv_vadd_vx_u8m1(v_b, 6 + is, vl), vl); + + vint16m2_t p0 = __riscv_vreinterpret_v_u16m2_i16m2(__riscv_vwmulu_vv_u16m2(q2_0, sc0, vl)); + vint16m2_t p1 = __riscv_vreinterpret_v_u16m2_i16m2(__riscv_vwmulu_vv_u16m2(q2_1, sc1, vl)); + vint16m2_t p2 = __riscv_vreinterpret_v_u16m2_i16m2(__riscv_vwmulu_vv_u16m2(q2_2, sc2, vl)); + vint16m2_t p3 = __riscv_vreinterpret_v_u16m2_i16m2(__riscv_vwmulu_vv_u16m2(q2_3, sc3, vl)); + + // load Q8 + vint8m1_t q8_0 = __riscv_vle8_v_i8m1(q8, vl); + vint8m1_t q8_1 = __riscv_vle8_v_i8m1(q8 + 32, vl); + vint8m1_t q8_2 = __riscv_vle8_v_i8m1(q8 + 64, vl); + vint8m1_t q8_3 = __riscv_vle8_v_i8m1(q8 + 96, vl); + + vint32m4_t s0 = __riscv_vwmul_vv_i32m4(p0, __riscv_vwcvt_x_x_v_i16m2(q8_0, vl), vl); + vint32m4_t s1 = __riscv_vwmul_vv_i32m4(p1, __riscv_vwcvt_x_x_v_i16m2(q8_1, vl), vl); + vint32m4_t s2 = __riscv_vwmul_vv_i32m4(p2, __riscv_vwcvt_x_x_v_i16m2(q8_2, vl), vl); + vint32m4_t s3 = __riscv_vwmul_vv_i32m4(p3, __riscv_vwcvt_x_x_v_i16m2(q8_3, vl), vl); + + vint32m1_t isum0 = __riscv_vredsum_vs_i32m4_i32m1(__riscv_vadd_vv_i32m4(s0, s1, vl), vzero, vl); + vint32m1_t isum1 = __riscv_vredsum_vs_i32m4_i32m1(__riscv_vadd_vv_i32m4(s2, s3, vl), isum0, vl); + + isum += __riscv_vmv_x_s_i32m1_i32(isum1); + + q2 += 32; + q8 += 128; + is = 8; + } + + sumf += dall * isum; + } + break; + case 128: + for (int i = 0; i < nb; ++i) { + const uint8_t * q2 = x[i].qs; + const int8_t * q8 = y[i].qs; + const uint8_t * sc = x[i].scales; + const float dall = y[i].d * GGML_FP16_TO_FP32(x[i].d); + const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin); + uint8_t *patmp = atmp; + int vsums; + int tmp; + __asm__ __volatile__( + "vsetivli zero, 16, e8, m1\n\t" + "vmv.v.x v8, zero\n\t" + "vle8.v v1, (%[sc])\n\t" + "vand.vi v0, v1, 0xF\n\t" + "vsrl.vi v1, v1, 4\n\t" + "vse8.v v0, (%[scale])\n\t" + "vsetivli zero, 16, e16, m2\n\t" + "vle16.v v2, (%[bsums])\n\t" + "vzext.vf2 v0, v1\n\t" + "vwmul.vv v4, v0, v2\n\t" + "vsetivli zero, 16, e32, m4\n\t" + "vredsum.vs v8, v4, v8\n\t" + "vmv.x.s %[vsums], v8" + : [tmp] "=&r" (tmp), [vsums] "=&r" (vsums) + : [sc] "r" (sc), [scale] "r" (atmp), [bsums] "r" (y[i].bsums) + : "memory" + , "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7" + , "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15" + , "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23" + , "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31" + ); + sumf += dmin * vsums; + int isum = 0; + + for (int j = 0; j < QK_K/128; ++j) { + __asm__ __volatile__( + "vsetvli zero, %[vl32], e8, m2\n\t" + "vle8.v v0, (%[q2])\n\t" + "vsrl.vi v2, v0, 2\n\t" + "vsrl.vi v4, v0, 4\n\t" + "vsrl.vi v6, v0, 6\n\t" + "vand.vi v0, v0, 0x3\n\t" + "vand.vi v2, v2, 0x3\n\t" + "vand.vi v4, v4, 0x3\n\t" + "vsetvli zero, %[vl128], e8, m8\n\t" + "vle8.v v8, (%[q8])\n\t" + "vsetvli zero, %[vl64], e8, m4\n\t" + "vwmul.vv v16, v0, v8\n\t" + "vwmul.vv v24, v4, v12\n\t" + "vsetivli zero, 16, e16, m2\n\t" + "vmv.v.x v0, zero\n\t" + "vwredsum.vs v10, v16, v0\n\t" + "vwredsum.vs v9, v18, v0\n\t" + "vwredsum.vs v8, v20, v0\n\t" + "vwredsum.vs v7, v22, v0\n\t" + "vwredsum.vs v11, v24, v0\n\t" + "vwredsum.vs v12, v26, v0\n\t" + "vwredsum.vs v13, v28, v0\n\t" + "vwredsum.vs v14, v30, v0\n\t" + "vsetivli zero, 4, e32, m1\n\t" + "vslideup.vi v10, v9, 1\n\t" + "vslideup.vi v8, v7, 1\n\t" + "vslideup.vi v11, v12, 1\n\t" + "vslideup.vi v13, v14, 1\n\t" + "vslideup.vi v10, v8, 2\n\t" + "vslideup.vi v11, v13, 2\n\t" + "vsetivli zero, 8, e32, m2\n\t" + "vle8.v v15, (%[scale])\n\t" + "vzext.vf4 v12, v15\n\t" + "vmul.vv v10, v10, v12\n\t" + "vredsum.vs v0, v10, v0\n\t" + "vmv.x.s %[tmp], v0\n\t" + "add %[isum], %[isum], %[tmp]" + : [tmp] "=&r" (tmp), [isum] "+&r" (isum) + : [q2] "r" (q2), [scale] "r" (patmp), [q8] "r" (q8) + , [vl32] "r" (32), [vl64] "r" (64), [vl128] "r" (128) + : "memory" + , "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7" + , "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15" + , "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23" + , "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31" + ); + q2 += 32; q8 += 128; patmp += 8; + } + + sumf += dall * isum; + } + break; + default: + assert(false && "Unsupported vector length"); + break; } *s = sumf; @@ -4573,8 +5338,8 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * restrict s, size_t bs, const void * r vector signed int vsumi6 = v0; vector signed int vsumi7 = v0; - const uint8_t * restrict q2 = x[i].qs; - const int8_t * restrict q8 = y[i].qs; + const uint8_t * GGML_RESTRICT q2 = x[i].qs; + const int8_t * GGML_RESTRICT q8 = y[i].qs; for (int j = 0; j < QK_K/128; ++j) { __builtin_prefetch(q2, 0, 1); @@ -4658,9 +5423,6 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * restrict s, size_t bs, const void * r #elif defined __loongarch_asx - const __m256i m3 = __lasx_xvreplgr2vr_b(3); - const __m128i m4 = __lsx_vreplgr2vr_b(0xF); - __m256 acc = (__m256)__lasx_xvldi(0); for (int i = 0; i < nb; ++i) { @@ -4668,21 +5430,18 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * restrict s, size_t bs, const void * r const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin); - const uint8_t * restrict q2 = x[i].qs; - const int8_t * restrict q8 = y[i].qs; + const uint8_t * GGML_RESTRICT q2 = x[i].qs; + const int8_t * GGML_RESTRICT q8 = y[i].qs; - const __m128i mins_and_scales = __lsx_vld((const __m128i*)x[i].scales, 0); - const __m128i scales8 = __lsx_vand_v(mins_and_scales, m4); - const __m128i mins8 = __lsx_vand_v(__lsx_vsrli_h(mins_and_scales, 4), m4); - const __m256i mins = lasx_ext8_16(mins8); + const __m128i mins_and_scales128 = __lsx_vld((const __m128i*)x[i].scales, 0); + const __m128i scales128 = __lsx_vandi_b(mins_and_scales128, 0xf); + const __m256i mins = lasx_ext8_16(__lsx_vsrli_b(mins_and_scales128, 4)); const __m256i prod = lasx_madd_h(mins, __lasx_xvld((const __m256i*)y[i].bsums, 0)); acc = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(dmin), __lasx_xvffint_s_w(prod), acc); - const __m256i all_scales = lasx_ext8_16(scales8); - const __m128i l_scales = lasx_extracti128(all_scales, 0); - const __m128i h_scales = lasx_extracti128(all_scales, 1); - const __m256i scales[2] = {lasx_insertf128(l_scales, l_scales), lasx_insertf128(h_scales, h_scales)}; + const v16i8 shuffle_mask = {0, 2, 4, 6, 8, 10, 12, 14, 1, 3, 5, 7, 9, 11, 13, 15}; + const __m256i scales_shuffled = lasx_ext8_16(__lsx_vshuf_b(scales128, scales128, (__m128i)shuffle_mask)); __m256i sumi = __lasx_xvldi(0); @@ -4695,20 +5454,20 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * restrict s, size_t bs, const void * r const __m256i q8_2 = __lasx_xvld((const __m256i*)q8, 0); q8 += 32; const __m256i q8_3 = __lasx_xvld((const __m256i*)q8, 0); q8 += 32; - const __m256i q2_0 = __lasx_xvand_v(q2bits, m3); - const __m256i q2_1 = __lasx_xvand_v(__lasx_xvsrli_h(q2bits, 2), m3); - const __m256i q2_2 = __lasx_xvand_v(__lasx_xvsrli_h(q2bits, 4), m3); - const __m256i q2_3 = __lasx_xvand_v(__lasx_xvsrli_h(q2bits, 6), m3); + const __m256i q2_0 = __lasx_xvandi_b(q2bits, 3); + const __m256i q2_1 = __lasx_xvandi_b(__lasx_xvsrli_b(q2bits, 2), 3); + const __m256i q2_2 = __lasx_xvandi_b(__lasx_xvsrli_b(q2bits, 4), 3); + const __m256i q2_3 = __lasx_xvsrli_b(q2bits, 6); - __m256i p0 = lasx_maddubs_h(q2_0, q8_0); - __m256i p1 = lasx_maddubs_h(q2_1, q8_1); - __m256i p2 = lasx_maddubs_h(q2_2, q8_2); - __m256i p3 = lasx_maddubs_h(q2_3, q8_3); + __m256i p0 = lasx_madd_h_b(q2_0, q8_0); + __m256i p1 = lasx_madd_h_b(q2_1, q8_1); + __m256i p2 = lasx_madd_h_b(q2_2, q8_2); + __m256i p3 = lasx_madd_h_b(q2_3, q8_3); - p0 = lasx_madd_h(lasx_shuffle_b(scales[j], get_scale_shuffle_q3k(0)), p0); - p1 = lasx_madd_h(lasx_shuffle_b(scales[j], get_scale_shuffle_q3k(1)), p1); - p2 = lasx_madd_h(lasx_shuffle_b(scales[j], get_scale_shuffle_q3k(2)), p2); - p3 = lasx_madd_h(lasx_shuffle_b(scales[j], get_scale_shuffle_q3k(3)), p3); + p0 = lasx_madd_h(lasx_xvrepl128vei_h(scales_shuffled, 4 * j + 0), p0); + p1 = lasx_madd_h(lasx_xvrepl128vei_h(scales_shuffled, 4 * j + 1), p1); + p2 = lasx_madd_h(lasx_xvrepl128vei_h(scales_shuffled, 4 * j + 2), p2); + p3 = lasx_madd_h(lasx_xvrepl128vei_h(scales_shuffled, 4 * j + 3), p3); p0 = __lasx_xvadd_w(p0, p1); p2 = __lasx_xvadd_w(p2, p3); @@ -4765,7 +5524,7 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * restrict s, size_t bs, const void * r #endif } -void ggml_vec_dot_q3_K_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) { +void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) { assert(n % QK_K == 0); assert(nrc == 1); UNUSED(nrc); @@ -4776,12 +5535,187 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * restrict s, size_t bs, const void * r const uint32_t kmask1 = 0x03030303; const uint32_t kmask2 = 0x0f0f0f0f; - const block_q3_K * restrict x = vx; - const block_q8_K * restrict y = vy; + const block_q3_K * GGML_RESTRICT x = vx; + const block_q8_K * GGML_RESTRICT y = vy; const int nb = n / QK_K; -#ifdef __ARM_NEON +#if defined(__ARM_FEATURE_SVE) + + uint32_t aux[3]; + uint32_t utmp[4]; + + const int8_t m32 = 32; + const int vector_length = svcntb()*8; + const svuint8_t m3b_sv = svdup_n_u8(0x3); + const svint32_t vzero_sv = svdup_n_s32(0); + + const svuint8_t m0_sv = svdup_n_u8(1); + const svuint8_t m1_sv = svlsl_n_u8_x(svptrue_b8(), m0_sv, 1); + const svuint8_t m2_sv = svlsl_n_u8_x(svptrue_b8(), m0_sv, 2); + const svuint8_t m3_sv = svlsl_n_u8_x(svptrue_b8(), m0_sv, 3); + + float sum = 0; + + for (int i = 0; i < nb; ++i) { + + const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); + + const uint8_t * GGML_RESTRICT q3_sv = x[i].qs; + const uint8_t * GGML_RESTRICT qh_sv = x[i].hmask; + const int8_t * GGML_RESTRICT q8_sv = y[i].qs; + + // Set up scales + memcpy(aux, x[i].scales, 12); + utmp[3] = ((aux[1] >> 4) & kmask2) | (((aux[2] >> 6) & kmask1) << 4); + utmp[2] = ((aux[0] >> 4) & kmask2) | (((aux[2] >> 4) & kmask1) << 4); + utmp[1] = (aux[1] & kmask2) | (((aux[2] >> 2) & kmask1) << 4); + utmp[0] = (aux[0] & kmask2) | (((aux[2] >> 0) & kmask1) << 4); + + int8_t * scale = (int8_t *)utmp; + + for (int j = 0; j < 16; ++j) scale[j] -= m32; + + switch (vector_length) { + case 128: + { + svuint8_t qhbits_sv_1 = svld1_u8(svptrue_b8(), qh_sv); + svuint8_t qhbits_sv_2 = svld1_u8(svptrue_b8(), qh_sv+16); + svuint8_t q3h_sv; + + svint32_t sumi1_1 = svdup_n_s32(0); + svint8_t q3bytes_sv; + + for (int j = 0; j < QK_K/128; ++j) { + + const svuint8_t q3bits_sv = svld1_u8(svptrue_b8(), q3_sv); q3_sv += 16; + const svuint8_t q3bits_sv_1 = svld1_u8(svptrue_b8(), q3_sv); q3_sv += 16; + svint8_t q8bytes_1_sv_1 = svld1_s8(svptrue_b8(), q8_sv); q8_sv += 16; + svint8_t q8bytes_1_sv_2 = svld1_s8(svptrue_b8(), q8_sv); q8_sv += 16; + + q3h_sv = svlsl_n_u8_x(svptrue_b8(), svbic_u8_x(svptrue_b8(), m0_sv, qhbits_sv_1), 2); + q3bytes_sv = svsub_s8_x(svptrue_b8(), svreinterpret_s8_u8(svand_u8_m(svptrue_b8(), q3bits_sv, m3b_sv)), svreinterpret_s8_u8(q3h_sv)); + + sumi1_1 = svmla_s32_m(svptrue_b32(), sumi1_1, svdot_s32(vzero_sv, q3bytes_sv, q8bytes_1_sv_1), svdup_n_s32((int32_t)scale[0])); + + q3h_sv = svlsl_n_u8_x(svptrue_b8(), svbic_u8_x(svptrue_b8(), m0_sv, qhbits_sv_2), 2); + q3bytes_sv = svsub_s8_x(svptrue_b8(), svreinterpret_s8_u8(svand_u8_m(svptrue_b8(), q3bits_sv_1, m3b_sv)), svreinterpret_s8_u8(q3h_sv)); + + sumi1_1 = svmla_s32_m(svptrue_b32(), sumi1_1, svdot_s32(vzero_sv, q3bytes_sv, q8bytes_1_sv_2), svdup_n_s32((int32_t)scale[1])); + + q8bytes_1_sv_1 = svld1_s8(svptrue_b8(), q8_sv); q8_sv += 16; + q8bytes_1_sv_2 = svld1_s8(svptrue_b8(), q8_sv); q8_sv += 16; + + q3h_sv = svlsl_n_u8_x(svptrue_b8(), svbic_u8_x(svptrue_b8(), m1_sv, qhbits_sv_1), 1); + q3bytes_sv = svsub_s8_x(svptrue_b8(), svreinterpret_s8_u8(svand_u8_m(svptrue_b8(), svlsr_n_u8_x(svptrue_b8(), q3bits_sv, 2), m3b_sv)), svreinterpret_s8_u8(q3h_sv)); + + sumi1_1 = svmla_s32_m(svptrue_b32(), sumi1_1, svdot_s32(vzero_sv, q3bytes_sv, q8bytes_1_sv_1), svdup_n_s32((int32_t)scale[2])); + + q3h_sv = svlsl_n_u8_x(svptrue_b8(), svbic_u8_x(svptrue_b8(), m1_sv, qhbits_sv_2), 1); + q3bytes_sv = svsub_s8_x(svptrue_b8(), svreinterpret_s8_u8(svand_u8_m(svptrue_b8(), svlsr_n_u8_x(svptrue_b8(), q3bits_sv_1, 2), m3b_sv)), svreinterpret_s8_u8(q3h_sv)); + + sumi1_1 = svmla_s32_m(svptrue_b32(), sumi1_1, svdot_s32(vzero_sv, q3bytes_sv, q8bytes_1_sv_2), svdup_n_s32((int32_t)scale[3])); + + + scale += 4; + q8bytes_1_sv_1 = svld1_s8(svptrue_b8(), q8_sv); q8_sv += 16; + q8bytes_1_sv_2 = svld1_s8(svptrue_b8(), q8_sv); q8_sv += 16; + + q3h_sv = svbic_u8_x(svptrue_b8(), m2_sv, qhbits_sv_1); + q3bytes_sv = svsub_s8_x(svptrue_b8(), svreinterpret_s8_u8(svand_u8_m(svptrue_b8(), svlsr_n_u8_x(svptrue_b8(), q3bits_sv, 4), m3b_sv)), svreinterpret_s8_u8(q3h_sv)); + + sumi1_1 = svmla_s32_m(svptrue_b32(), sumi1_1, svdot_s32(vzero_sv, q3bytes_sv, q8bytes_1_sv_1), svdup_n_s32((int32_t)scale[0])); + + q3h_sv = svbic_u8_x(svptrue_b8(), m2_sv, qhbits_sv_2); + q3bytes_sv = svsub_s8_x(svptrue_b8(), svreinterpret_s8_u8(svand_u8_m(svptrue_b8(), svlsr_n_u8_x(svptrue_b8(), q3bits_sv_1, 4), m3b_sv)), svreinterpret_s8_u8(q3h_sv)); + + sumi1_1 = svmla_s32_m(svptrue_b32(), sumi1_1, svdot_s32(vzero_sv, q3bytes_sv, q8bytes_1_sv_2), svdup_n_s32((int32_t)scale[1])); + + + q8bytes_1_sv_1 = svld1_s8(svptrue_b8(), q8_sv); q8_sv += 16; + q8bytes_1_sv_2 = svld1_s8(svptrue_b8(), q8_sv); q8_sv += 16; + + q3h_sv = svlsr_n_u8_x(svptrue_b8(), svbic_u8_x(svptrue_b8(), m3_sv, qhbits_sv_1), 1); + q3bytes_sv = svsub_s8_x(svptrue_b8(), svreinterpret_s8_u8(svand_u8_m(svptrue_b8(), svlsr_n_u8_x(svptrue_b8(), q3bits_sv, 6), m3b_sv)), svreinterpret_s8_u8(q3h_sv)); + + sumi1_1 = svmla_s32_m(svptrue_b32(), sumi1_1, svdot_s32(vzero_sv, q3bytes_sv, q8bytes_1_sv_1), svdup_n_s32((int32_t)scale[2])); + + q3h_sv = svlsr_n_u8_x(svptrue_b8(), svbic_u8_x(svptrue_b8(), m3_sv, qhbits_sv_2), 1); + q3bytes_sv = svsub_s8_x(svptrue_b8(), svreinterpret_s8_u8(svand_u8_m(svptrue_b8(), svlsr_n_u8_x(svptrue_b8(), q3bits_sv_1, 6), m3b_sv)), svreinterpret_s8_u8(q3h_sv)); + + sumi1_1 = svmla_s32_m(svptrue_b32(), sumi1_1, svdot_s32(vzero_sv, q3bytes_sv, q8bytes_1_sv_2), svdup_n_s32((int32_t)scale[3])); + + if (j == 0) { + qhbits_sv_1 = svlsr_n_u8_x(svptrue_b8(), qhbits_sv_1, 4); + qhbits_sv_2 = svlsr_n_u8_x(svptrue_b8(), qhbits_sv_2, 4); + } + + scale += 4; + } + + sum += d * (svaddv_s32(svptrue_b32(), sumi1_1)); + } break; + case 256: + case 512: + { + svuint8_t qhbits_sv = svld1_u8(svptrue_pat_b8(SV_VL32), qh_sv); + svuint8_t q3h_sv; + + svint32_t sumi1_1 = svdup_n_s32(0); + svint8_t q3bytes_sv; + + for (int j = 0; j < QK_K/128; ++j) { + + const svuint8_t q3bits_sv = svld1_u8(svptrue_pat_b8(SV_VL32), q3_sv); q3_sv += 32; + svint8_t q8bytes_1_sv_1 = svld1_s8(svptrue_pat_b8(SV_VL32), q8_sv); q8_sv += 32; + svint8_t q8bytes_1_sv_2 = svld1_s8(svptrue_pat_b8(SV_VL32), q8_sv); q8_sv += 32; + + q3h_sv = svlsl_n_u8_x(svptrue_pat_b8(SV_VL32), svbic_u8_x(svptrue_pat_b8(SV_VL32), m0_sv, qhbits_sv), 2); + q3bytes_sv = svsub_s8_x(svptrue_pat_b8(SV_VL32), svreinterpret_s8_u8(svand_u8_m(svptrue_pat_b8(SV_VL32), q3bits_sv, m3b_sv)), svreinterpret_s8_u8(q3h_sv)); + + + svint32_t scale_1 = svsel_s32(svptrue_pat_b32(SV_VL4), svdup_n_s32((int32_t)scale[0]), svdup_n_s32((int32_t)scale[1])); + sumi1_1 = svmla_s32_m(svptrue_pat_b32(SV_VL8), sumi1_1, svdot_s32(vzero_sv, q3bytes_sv, q8bytes_1_sv_1), scale_1); + + q3h_sv = svlsl_n_u8_x(svptrue_pat_b8(SV_VL32), svbic_u8_x(svptrue_pat_b8(SV_VL32), m1_sv, qhbits_sv), 1); + q3bytes_sv = svsub_s8_x(svptrue_pat_b8(SV_VL32), svreinterpret_s8_u8(svand_u8_m(svptrue_pat_b8(SV_VL32), svlsr_n_u8_x(svptrue_pat_b8(SV_VL32), q3bits_sv, 2), m3b_sv)), svreinterpret_s8_u8(q3h_sv)); + + scale_1 = svsel_s32(svptrue_pat_b32(SV_VL4), svdup_n_s32((int32_t)scale[2]), svdup_n_s32((int32_t)scale[3])); + sumi1_1 = svmla_s32_m(svptrue_pat_b32(SV_VL8), sumi1_1, svdot_s32(vzero_sv, q3bytes_sv, q8bytes_1_sv_2), scale_1); + + scale += 4; + q8bytes_1_sv_1 = svld1_s8(svptrue_pat_b8(SV_VL32), q8_sv); q8_sv += 32; + q8bytes_1_sv_2 = svld1_s8(svptrue_pat_b8(SV_VL32), q8_sv); q8_sv += 32; + + q3h_sv = svbic_u8_x(svptrue_pat_b8(SV_VL32), m2_sv, qhbits_sv); + q3bytes_sv = svsub_s8_x(svptrue_pat_b8(SV_VL32), svreinterpret_s8_u8(svand_u8_m(svptrue_pat_b8(SV_VL32), svlsr_n_u8_x(svptrue_pat_b8(SV_VL32), q3bits_sv, 4), m3b_sv)), svreinterpret_s8_u8(q3h_sv)); + + scale_1 = svsel_s32(svptrue_pat_b32(SV_VL4), svdup_n_s32((int32_t)scale[0]), svdup_n_s32((int32_t)scale[1])); + sumi1_1 = svmla_s32_m(svptrue_pat_b32(SV_VL8), sumi1_1, svdot_s32(vzero_sv, q3bytes_sv, q8bytes_1_sv_1), scale_1); + + q3h_sv = svlsr_n_u8_x(svptrue_pat_b8(SV_VL32), svbic_u8_x(svptrue_pat_b8(SV_VL32), m3_sv, qhbits_sv), 1); + q3bytes_sv = svsub_s8_x(svptrue_pat_b8(SV_VL32), svreinterpret_s8_u8(svand_u8_m(svptrue_pat_b8(SV_VL32), svlsr_n_u8_x(svptrue_pat_b8(SV_VL32), q3bits_sv, 6), m3b_sv)), svreinterpret_s8_u8(q3h_sv)); + + scale_1 = svsel_s32(svptrue_pat_b32(SV_VL4), svdup_n_s32((int32_t)scale[2]), svdup_n_s32((int32_t)scale[3])); + sumi1_1 = svmla_s32_m(svptrue_pat_b32(SV_VL8), sumi1_1, svdot_s32(vzero_sv, q3bytes_sv, q8bytes_1_sv_2), scale_1); + + if (j == 0) { + qhbits_sv = svlsr_n_u8_x(svptrue_pat_b8(SV_VL32), qhbits_sv, 4); + } + + scale += 4; + } + + sum += d * (svaddv_s32(svptrue_pat_b32(SV_VL8), sumi1_1)); + } break; + default: + assert(false && "Unsupported vector length"); + break; + } + } + *s = sum; + +#elif __ARM_NEON uint32_t aux[3]; uint32_t utmp[4]; @@ -4803,9 +5737,9 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * restrict s, size_t bs, const void * r const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); - const uint8_t * restrict q3 = x[i].qs; - const uint8_t * restrict qh = x[i].hmask; - const int8_t * restrict q8 = y[i].qs; + const uint8_t * GGML_RESTRICT q3 = x[i].qs; + const uint8_t * GGML_RESTRICT qh = x[i].hmask; + const int8_t * GGML_RESTRICT q8 = y[i].qs; ggml_uint8x16x2_t qhbits = ggml_vld1q_u8_x2(qh); @@ -4889,8 +5823,8 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * restrict s, size_t bs, const void * r const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); - const uint8_t * restrict q3 = x[i].qs; - const int8_t * restrict q8 = y[i].qs; + const uint8_t * GGML_RESTRICT q3 = x[i].qs; + const int8_t * GGML_RESTRICT q8 = y[i].qs; // Set up scales memcpy(aux, x[i].scales, 12); @@ -4994,8 +5928,8 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * restrict s, size_t bs, const void * r const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); - const uint8_t * restrict q3 = x[i].qs; - const int8_t * restrict q8 = y[i].qs; + const uint8_t * GGML_RESTRICT q3 = x[i].qs; + const int8_t * GGML_RESTRICT q8 = y[i].qs; // Set up scales aux = (const uint32_t *)x[i].scales; @@ -5121,102 +6055,314 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * restrict s, size_t bs, const void * r *s = hsum_float_8(acc); +#elif defined __wasm_simd128__ + int8_t aux8[QK_K]; + float sums[8] = {0}; + uint32_t auxs[4]; + + float sumf = 0; + for (int i = 0; i < nb; ++i) { + const uint8_t * GGML_RESTRICT q3 = x[i].qs; + const uint8_t * GGML_RESTRICT hm = x[i].hmask; + const int8_t * GGML_RESTRICT q8 = y[i].qs; + + // Process blocks with SIMD + int8_t * a = aux8; + uint8_t m = 1; + for (int j = 0; j < QK_K; j += 128) { + for (int shift = 0; shift <= 6; shift += 2) { + v128_t v_m = wasm_i8x16_splat(m); + for (int l = 0; l < 32; l += 16) { + v128_t v_q3 = wasm_v128_load(q3 + l); + v128_t v_shift = wasm_i8x16_shr(v_q3, shift); + v128_t v_low2 = wasm_v128_and(v_shift, wasm_i8x16_splat(0x03)); + + v128_t v_hm = wasm_v128_load(hm + l); + v128_t v_mask = wasm_v128_and(v_hm, v_m); + v_mask = wasm_i8x16_ne(v_mask, wasm_i8x16_splat(0)); + + v_low2 = wasm_i8x16_sub(v_low2, wasm_v128_and(wasm_i8x16_splat(4), wasm_v128_not(v_mask))); + wasm_v128_store(a + l, v_low2); + } + a += 32; + m <<= 1; + } + q3 += 32; + } + + // Extract scales + memcpy(auxs, x[i].scales, 12); + uint32_t tmp = auxs[2]; + auxs[2] = ((auxs[0] >> 4) & kmask2) | (((tmp >> 4) & kmask1) << 4); + auxs[3] = ((auxs[1] >> 4) & kmask2) | (((tmp >> 6) & kmask1) << 4); + auxs[0] = (auxs[0] & kmask2) | (((tmp >> 0) & kmask1) << 4); + auxs[1] = (auxs[1] & kmask2) | (((tmp >> 2) & kmask1) << 4); + const int8_t * scales = (const int8_t *)auxs; + + // SIMD dot product with register accumulators + v128_t v_acc0 = wasm_i32x4_splat(0); + v128_t v_acc1 = wasm_i32x4_splat(0); + a = aux8; + for (int j = 0; j < QK_K/16; ++j) { + const v128_t v_scale = wasm_i16x8_splat(scales[j] - 32); + + // Process 16 elements per iteration + for (int k = 0; k < 2; ++k) { + const v128_t v_q8 = wasm_i16x8_load8x8(q8); + const v128_t v_a = wasm_i16x8_load8x8(a); + + v128_t v_prod = wasm_i16x8_mul(v_q8, v_a); + v_prod = wasm_i16x8_mul(v_prod, v_scale); + + v_acc0 = wasm_i32x4_add(v_acc0, wasm_i32x4_extend_low_i16x8(v_prod)); + v_acc1 = wasm_i32x4_add(v_acc1, wasm_i32x4_extend_high_i16x8(v_prod)); + + q8 += 8; + a += 8; + } + } + + // Accumulate results + const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + const v128_t v_d = wasm_f32x4_splat(d); + v128_t v_sum = wasm_f32x4_add( + wasm_f32x4_mul(wasm_f32x4_convert_i32x4(v_acc0), v_d), + wasm_f32x4_mul(wasm_f32x4_convert_i32x4(v_acc1), v_d) + ); + + // Accumulate into sums vector + wasm_v128_store(sums, wasm_f32x4_add(wasm_v128_load(sums), v_sum)); + } + + // Horizontal sum + v128_t v_sum = wasm_f32x4_add(wasm_v128_load(sums), wasm_v128_load(sums + 4)); + sumf = wasm_f32x4_extract_lane(v_sum, 0) + + wasm_f32x4_extract_lane(v_sum, 1) + + wasm_f32x4_extract_lane(v_sum, 2) + + wasm_f32x4_extract_lane(v_sum, 3); + + *s = sumf; + #elif defined __riscv_v_intrinsic uint32_t aux[3]; uint32_t utmp[4]; + const int vector_length = __riscv_vlenb() * 8; float sumf = 0; - for (int i = 0; i < nb; ++i) { - const uint8_t * restrict q3 = x[i].qs; - const uint8_t * restrict qh = x[i].hmask; - const int8_t * restrict q8 = y[i].qs; + switch (vector_length) { + case 256: + for (int i = 0; i < nb; ++i) { - memcpy(aux, x[i].scales, 12); - utmp[3] = ((aux[1] >> 4) & kmask2) | (((aux[2] >> 6) & kmask1) << 4); - utmp[2] = ((aux[0] >> 4) & kmask2) | (((aux[2] >> 4) & kmask1) << 4); - utmp[1] = (aux[1] & kmask2) | (((aux[2] >> 2) & kmask1) << 4); - utmp[0] = (aux[0] & kmask2) | (((aux[2] >> 0) & kmask1) << 4); + const uint8_t * GGML_RESTRICT q3 = x[i].qs; + const uint8_t * GGML_RESTRICT qh = x[i].hmask; + const int8_t * GGML_RESTRICT q8 = y[i].qs; - int8_t * scale = (int8_t *)utmp; - for (int j = 0; j < 16; ++j) scale[j] -= 32; + memcpy(aux, x[i].scales, 12); + utmp[3] = ((aux[1] >> 4) & kmask2) | (((aux[2] >> 6) & kmask1) << 4); + utmp[2] = ((aux[0] >> 4) & kmask2) | (((aux[2] >> 4) & kmask1) << 4); + utmp[1] = (aux[1] & kmask2) | (((aux[2] >> 2) & kmask1) << 4); + utmp[0] = (aux[0] & kmask2) | (((aux[2] >> 0) & kmask1) << 4); + + int8_t * scale = (int8_t *)utmp; + for (int j = 0; j < 16; ++j) scale[j] -= 32; - size_t vl = 32; - uint8_t m = 1; + size_t vl = 32; + uint8_t m = 1; - vint32m1_t vzero = __riscv_vmv_v_x_i32m1(0, 1); - vuint8m1_t vqh = __riscv_vle8_v_u8m1(qh, vl); + vint32m1_t vzero = __riscv_vmv_v_x_i32m1(0, 1); + vuint8m1_t vqh = __riscv_vle8_v_u8m1(qh, vl); - int sum_t = 0; + int sum_t = 0; - for (int j = 0; j < QK_K; j += 128) { + for (int j = 0; j < QK_K; j += 128) { - vl = 32; + vl = 32; - // load Q3 - vuint8m1_t q3_x = __riscv_vle8_v_u8m1(q3, vl); + // load Q3 + vuint8m1_t q3_x = __riscv_vle8_v_u8m1(q3, vl); - vint8m1_t q3_0 = __riscv_vreinterpret_v_u8m1_i8m1(__riscv_vand_vx_u8m1(q3_x, 0x03, vl)); - vint8m1_t q3_1 = __riscv_vreinterpret_v_u8m1_i8m1(__riscv_vand_vx_u8m1(__riscv_vsrl_vx_u8m1(q3_x, 0x2, vl), 0x03 , vl)); - vint8m1_t q3_2 = __riscv_vreinterpret_v_u8m1_i8m1(__riscv_vand_vx_u8m1(__riscv_vsrl_vx_u8m1(q3_x, 0x4, vl), 0x03 , vl)); - vint8m1_t q3_3 = __riscv_vreinterpret_v_u8m1_i8m1(__riscv_vand_vx_u8m1(__riscv_vsrl_vx_u8m1(q3_x, 0x6, vl), 0x03 , vl)); + vint8m1_t q3_0 = __riscv_vreinterpret_v_u8m1_i8m1(__riscv_vand_vx_u8m1(q3_x, 0x03, vl)); + vint8m1_t q3_1 = __riscv_vreinterpret_v_u8m1_i8m1(__riscv_vand_vx_u8m1(__riscv_vsrl_vx_u8m1(q3_x, 0x2, vl), 0x03 , vl)); + vint8m1_t q3_2 = __riscv_vreinterpret_v_u8m1_i8m1(__riscv_vand_vx_u8m1(__riscv_vsrl_vx_u8m1(q3_x, 0x4, vl), 0x03 , vl)); + vint8m1_t q3_3 = __riscv_vreinterpret_v_u8m1_i8m1(__riscv_vand_vx_u8m1(__riscv_vsrl_vx_u8m1(q3_x, 0x6, vl), 0x03 , vl)); - // compute mask for subtraction - vuint8m1_t qh_m0 = __riscv_vand_vx_u8m1(vqh, m, vl); - vbool8_t vmask_0 = __riscv_vmseq_vx_u8m1_b8(qh_m0, 0, vl); - vint8m1_t q3_m0 = __riscv_vsub_vx_i8m1_mu(vmask_0, q3_0, q3_0, 0x4, vl); - m <<= 1; + // compute mask for subtraction + vuint8m1_t qh_m0 = __riscv_vand_vx_u8m1(vqh, m, vl); + vbool8_t vmask_0 = __riscv_vmseq_vx_u8m1_b8(qh_m0, 0, vl); + vint8m1_t q3_m0 = __riscv_vsub_vx_i8m1_mu(vmask_0, q3_0, q3_0, 0x4, vl); + m <<= 1; - vuint8m1_t qh_m1 = __riscv_vand_vx_u8m1(vqh, m, vl); - vbool8_t vmask_1 = __riscv_vmseq_vx_u8m1_b8(qh_m1, 0, vl); - vint8m1_t q3_m1 = __riscv_vsub_vx_i8m1_mu(vmask_1, q3_1, q3_1, 0x4, vl); - m <<= 1; + vuint8m1_t qh_m1 = __riscv_vand_vx_u8m1(vqh, m, vl); + vbool8_t vmask_1 = __riscv_vmseq_vx_u8m1_b8(qh_m1, 0, vl); + vint8m1_t q3_m1 = __riscv_vsub_vx_i8m1_mu(vmask_1, q3_1, q3_1, 0x4, vl); + m <<= 1; - vuint8m1_t qh_m2 = __riscv_vand_vx_u8m1(vqh, m, vl); - vbool8_t vmask_2 = __riscv_vmseq_vx_u8m1_b8(qh_m2, 0, vl); - vint8m1_t q3_m2 = __riscv_vsub_vx_i8m1_mu(vmask_2, q3_2, q3_2, 0x4, vl); - m <<= 1; + vuint8m1_t qh_m2 = __riscv_vand_vx_u8m1(vqh, m, vl); + vbool8_t vmask_2 = __riscv_vmseq_vx_u8m1_b8(qh_m2, 0, vl); + vint8m1_t q3_m2 = __riscv_vsub_vx_i8m1_mu(vmask_2, q3_2, q3_2, 0x4, vl); + m <<= 1; - vuint8m1_t qh_m3 = __riscv_vand_vx_u8m1(vqh, m, vl); - vbool8_t vmask_3 = __riscv_vmseq_vx_u8m1_b8(qh_m3, 0, vl); - vint8m1_t q3_m3 = __riscv_vsub_vx_i8m1_mu(vmask_3, q3_3, q3_3, 0x4, vl); - m <<= 1; + vuint8m1_t qh_m3 = __riscv_vand_vx_u8m1(vqh, m, vl); + vbool8_t vmask_3 = __riscv_vmseq_vx_u8m1_b8(qh_m3, 0, vl); + vint8m1_t q3_m3 = __riscv_vsub_vx_i8m1_mu(vmask_3, q3_3, q3_3, 0x4, vl); + m <<= 1; - // load Q8 and take product with Q3 - vint16m2_t a0 = __riscv_vwmul_vv_i16m2(q3_m0, __riscv_vle8_v_i8m1(q8, vl), vl); - vint16m2_t a1 = __riscv_vwmul_vv_i16m2(q3_m1, __riscv_vle8_v_i8m1(q8+32, vl), vl); - vint16m2_t a2 = __riscv_vwmul_vv_i16m2(q3_m2, __riscv_vle8_v_i8m1(q8+64, vl), vl); - vint16m2_t a3 = __riscv_vwmul_vv_i16m2(q3_m3, __riscv_vle8_v_i8m1(q8+96, vl), vl); + // load Q8 and take product with Q3 + vint16m2_t a0 = __riscv_vwmul_vv_i16m2(q3_m0, __riscv_vle8_v_i8m1(q8, vl), vl); + vint16m2_t a1 = __riscv_vwmul_vv_i16m2(q3_m1, __riscv_vle8_v_i8m1(q8+32, vl), vl); + vint16m2_t a2 = __riscv_vwmul_vv_i16m2(q3_m2, __riscv_vle8_v_i8m1(q8+64, vl), vl); + vint16m2_t a3 = __riscv_vwmul_vv_i16m2(q3_m3, __riscv_vle8_v_i8m1(q8+96, vl), vl); - vl = 16; + vl = 16; - // retrieve lane to multiply with scale - vint32m2_t aux0_0 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(a0, 0), (scale[0]), vl); - vint32m2_t aux0_1 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(a0, 1), (scale[1]), vl); - vint32m2_t aux1_0 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(a1, 0), (scale[2]), vl); - vint32m2_t aux1_1 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(a1, 1), (scale[3]), vl); - vint32m2_t aux2_0 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(a2, 0), (scale[4]), vl); - vint32m2_t aux2_1 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(a2, 1), (scale[5]), vl); - vint32m2_t aux3_0 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(a3, 0), (scale[6]), vl); - vint32m2_t aux3_1 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(a3, 1), (scale[7]), vl); + // retrieve lane to multiply with scale + vint32m2_t aux0_0 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(a0, 0), (scale[0]), vl); + vint32m2_t aux0_1 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(a0, 1), (scale[1]), vl); + vint32m2_t aux1_0 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(a1, 0), (scale[2]), vl); + vint32m2_t aux1_1 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(a1, 1), (scale[3]), vl); + vint32m2_t aux2_0 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(a2, 0), (scale[4]), vl); + vint32m2_t aux2_1 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(a2, 1), (scale[5]), vl); + vint32m2_t aux3_0 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(a3, 0), (scale[6]), vl); + vint32m2_t aux3_1 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(a3, 1), (scale[7]), vl); - vint32m1_t isum0 = __riscv_vredsum_vs_i32m2_i32m1(__riscv_vadd_vv_i32m2(aux0_0, aux0_1, vl), vzero, vl); - vint32m1_t isum1 = __riscv_vredsum_vs_i32m2_i32m1(__riscv_vadd_vv_i32m2(aux1_0, aux1_1, vl), isum0, vl); - vint32m1_t isum2 = __riscv_vredsum_vs_i32m2_i32m1(__riscv_vadd_vv_i32m2(aux2_0, aux2_1, vl), isum1, vl); - vint32m1_t isum3 = __riscv_vredsum_vs_i32m2_i32m1(__riscv_vadd_vv_i32m2(aux3_0, aux3_1, vl), isum2, vl); + vint32m1_t isum0 = __riscv_vredsum_vs_i32m2_i32m1(__riscv_vadd_vv_i32m2(aux0_0, aux0_1, vl), vzero, vl); + vint32m1_t isum1 = __riscv_vredsum_vs_i32m2_i32m1(__riscv_vadd_vv_i32m2(aux1_0, aux1_1, vl), isum0, vl); + vint32m1_t isum2 = __riscv_vredsum_vs_i32m2_i32m1(__riscv_vadd_vv_i32m2(aux2_0, aux2_1, vl), isum1, vl); + vint32m1_t isum3 = __riscv_vredsum_vs_i32m2_i32m1(__riscv_vadd_vv_i32m2(aux3_0, aux3_1, vl), isum2, vl); - sum_t += __riscv_vmv_x_s_i32m1_i32(isum3); + sum_t += __riscv_vmv_x_s_i32m1_i32(isum3); - q3 += 32; q8 += 128; scale += 8; + q3 += 32; q8 += 128; scale += 8; + + } + + const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + + sumf += d*sum_t; } + break; + case 128: + for (int i = 0; i < nb; ++i) { + const uint8_t * restrict q3 = x[i].qs; + const uint8_t * restrict qh = x[i].hmask; + const int8_t * restrict q8 = y[i].qs; - const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + int8_t * scale = (int8_t *)utmp; + int tmp; + __asm__ __volatile__( + "vsetivli zero, 12, e8, m1\n\t" + "vle8.v v0, (%[s6b])\n\t" + "vmv1r.v v2, v0\n\t" + "vsetivli zero, 2, e64, m1\n\t" + "vmv.v.x v9, %[sh]\n\t"\ + "vslidedown.vi v1, v0, 1\n\t" + "vslide1up.vx v8, v9, zero\n\t" // {0, 0, 4, 4} + "vslideup.vi v0, v2, 1\n\t" // {aux[0], aux[1], aux[0], aux[1]} + "vsetivli zero, 4, e32, m1\n\t" + "vid.v v9\n\t" + "vmv.x.s %[tmp], v1\n\t" + "vsll.vi v9, v9, 1\n\t" // {0, 2, 4, 6} + "vmv.v.x v1, %[tmp]\n\t" // {aux[2], aux[2], aux[2], aux[2]} + "vsrl.vv v4, v1, v9\n\t" + "vsrl.vv v2, v0, v8\n\t" + "vand.vx v5, v4, %[kmask1]\n\t" + "vand.vx v3, v2, %[kmask2]\n\t" + "vsll.vi v6, v5, 4\n\t" + "vor.vv v7, v6, v3\n\t" + "vsetivli zero, 16, e8, m1\n\t" + "vsub.vx v0, v7, %[c]\n\t" + "vse8.v v0, (%[scale])" + : [tmp] "=&r" (tmp) + : [sh] "r" (0x0000000400000004), [s6b] "r" (x[i].scales), [c] "r" (32) + , [scale] "r" (scale), [kmask1] "r" (kmask1), [kmask2] "r" (kmask2) + : "memory" + , "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7" + , "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15" + , "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23" + , "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31" + ); - sumf += d*sum_t; + uint8_t m = 1; + int isum = 0; + for (int j = 0; j < QK_K; j += 128) { + __asm__ __volatile__( + "vsetvli zero, %[vl32], e8, m2, ta, mu\n\t" + "vle8.v v8, (%[q3])\n\t" + "vsrl.vi v10, v8, 2\n\t" + "vsrl.vi v12, v8, 4\n\t" + "vsrl.vi v14, v8, 6\n\t" + "vand.vi v8, v8, 3\n\t" + "vand.vi v10, v10, 3\n\t" + "vand.vi v12, v12, 3\n\t" + "vle8.v v2, (%[qh])\n\t" + "vand.vx v4, v2, %[m]\n\t" + "slli %[m], %[m], 1\n\t" + "vmseq.vx v0, v4, zero\n\t" + "vadd.vi v8, v8, -4, v0.t\n\t" + "vand.vx v4, v2, %[m]\n\t" + "slli %[m], %[m], 1\n\t" + "vmseq.vx v0, v4, zero\n\t" + "vadd.vi v10, v10, -4, v0.t\n\t" + "vand.vx v4, v2, %[m]\n\t" + "slli %[m], %[m], 1\n\t" + "vmseq.vx v0, v4, zero\n\t" + "vadd.vi v12, v12, -4, v0.t\n\t" + "vand.vx v4, v2, %[m]\n\t" + "slli %[m], %[m], 1\n\t" + "vmseq.vx v0, v4, zero\n\t" + "vadd.vi v14, v14, -4, v0.t\n\t" + "vsetvli zero, %[vl128], e8, m8\n\t" + "vle8.v v0, (%[q8])\n\t" + "vsetvli zero, %[vl64], e8, m4\n\t" + "vwmul.vv v16, v0, v8\n\t" + "vwmul.vv v24, v4, v12\n\t" + "vsetivli zero, 16, e16, m2\n\t" + "vmv.v.x v0, zero\n\t" + "vwredsum.vs v10, v16, v0\n\t" + "vwredsum.vs v9, v18, v0\n\t" + "vwredsum.vs v8, v20, v0\n\t" + "vwredsum.vs v7, v22, v0\n\t" + "vwredsum.vs v11, v24, v0\n\t" + "vwredsum.vs v12, v26, v0\n\t" + "vwredsum.vs v13, v28, v0\n\t" + "vwredsum.vs v14, v30, v0\n\t" + "vsetivli zero, 4, e32, m1\n\t" + "vslideup.vi v10, v9, 1\n\t" + "vslideup.vi v8, v7, 1\n\t" + "vslideup.vi v11, v12, 1\n\t" + "vslideup.vi v13, v14, 1\n\t" + "vslideup.vi v10, v8, 2\n\t" + "vslideup.vi v11, v13, 2\n\t" + "vsetivli zero, 8, e32, m2\n\t"\ + "vle8.v v15, (%[scale])\n\t" + "vsext.vf4 v12, v15\n\t" + "vmul.vv v10, v10, v12\n\t" + "vredsum.vs v0, v10, v0\n\t" + "vmv.x.s %[tmp], v0\n\t" + "add %[isum], %[isum], %[tmp]" + : [tmp] "=&r" (tmp), [m] "+&r" (m), [isum] "+&r" (isum) + : [vl128] "r" (128), [vl64] "r" (64), [vl32] "r" (32) + , [q3] "r" (q3), [qh] "r" (qh), [scale] "r" (scale), [q8] "r" (q8) + : "memory" + , "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7" + , "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15" + , "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23" + , "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31" + ); + q3 += 32; q8 += 128; scale += 8; + } + const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + sumf += d * isum; + } + break; + default: + assert(false && "Unsupported vector length"); + break; } *s = sumf; @@ -5271,8 +6417,8 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * restrict s, size_t bs, const void * r vector signed int vsumi6 = v0; vector signed int vsumi7 = v0; - const uint8_t * restrict q3 = x[i].qs; - const int8_t * restrict q8 = y[i].qs; + const uint8_t * GGML_RESTRICT q3 = x[i].qs; + const int8_t * GGML_RESTRICT q8 = y[i].qs; for (int j = 0; j < QK_K/128; ++j) { __builtin_prefetch(q3, 0, 1); @@ -5376,8 +6522,6 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * restrict s, size_t bs, const void * r #elif defined __loongarch_asx - const __m256i m3 = __lasx_xvreplgr2vr_b(3); - const __m256i mone = __lasx_xvreplgr2vr_b(1); const __m128i m32 = __lsx_vreplgr2vr_b(32); __m256 acc = (__m256)__lasx_xvldi(0); @@ -5387,8 +6531,8 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * restrict s, size_t bs, const void * r for (int i = 0; i < nb; ++i) { const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); - const uint8_t * restrict q3 = x[i].qs; - const int8_t * restrict q8 = y[i].qs; + const uint8_t * GGML_RESTRICT q3 = x[i].qs; + const int8_t * GGML_RESTRICT q8 = y[i].qs; // Set up scales memcpy(aux, x[i].scales, 12); __m128i scales128 = lsx_set_w( @@ -5397,10 +6541,9 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * restrict s, size_t bs, const void * r (aux[1] & kmask2) | (((aux[2] >> 2) & kmask1) << 4), (aux[0] & kmask2) | (((aux[2] >> 0) & kmask1) << 4)); scales128 = __lsx_vsub_b(scales128, m32); - const __m256i all_scales = lasx_ext8_16(scales128); - const __m128i l_scales = lasx_extracti128(all_scales, 0); - const __m128i h_scales = lasx_extracti128(all_scales, 1); - const __m256i scales[2] = {lasx_insertf128(l_scales, l_scales), lasx_insertf128(h_scales, h_scales)}; + + const v16i8 shuffle_mask = {0, 2, 4, 6, 8, 10, 12, 14, 1, 3, 5, 7, 9, 11, 13, 15}; + const __m256i scales_shuffled = lasx_ext8_16(__lsx_vshuf_b(scales128, scales128, (__m128i)shuffle_mask)); // high bit const __m256i hbits = __lasx_xvld((const __m256i*)x[i].hmask, 0); @@ -5408,35 +6551,23 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * restrict s, size_t bs, const void * r // integer accumulator __m256i sumi = __lasx_xvldi(0); - int bit = 0; - int is = 0; - __m256i xvbit; - - for (int j = 0; j < QK_K/128; ++j) { // load low 2 bits const __m256i q3bits = __lasx_xvld((const __m256i*)q3, 0); q3 += 32; - xvbit = __lasx_xvreplgr2vr_h(bit); // prepare low and high bits - const __m256i q3l_0 = __lasx_xvand_v(q3bits, m3); - const __m256i q3h_0 = __lasx_xvslli_h(__lasx_xvsrl_h(__lasx_xvandn_v(hbits, __lasx_xvsll_h(mone, xvbit)), xvbit), 2); - ++bit; - - xvbit = __lasx_xvreplgr2vr_h(bit); - const __m256i q3l_1 = __lasx_xvand_v(__lasx_xvsrli_h(q3bits, 2), m3); - const __m256i q3h_1 = __lasx_xvslli_h(__lasx_xvsrl_h(__lasx_xvandn_v(hbits, __lasx_xvsll_h(mone, xvbit)), xvbit), 2); - ++bit; - - xvbit = __lasx_xvreplgr2vr_h(bit); - const __m256i q3l_2 = __lasx_xvand_v(__lasx_xvsrli_h(q3bits, 4), m3); - const __m256i q3h_2 = __lasx_xvslli_h(__lasx_xvsrl_h(__lasx_xvandn_v(hbits, __lasx_xvsll_h(mone, xvbit)), xvbit), 2); - ++bit; - - xvbit = __lasx_xvreplgr2vr_h(bit); - const __m256i q3l_3 = __lasx_xvand_v(__lasx_xvsrli_h(q3bits, 6), m3); - const __m256i q3h_3 = __lasx_xvslli_h(__lasx_xvsrl_h(__lasx_xvandn_v(hbits, __lasx_xvsll_h(mone, xvbit)), xvbit), 2); - ++bit; + const __m256i q3l_0 = __lasx_xvandi_b(q3bits, 3); + const __m256i q3l_1 = __lasx_xvandi_b(__lasx_xvsrli_b(q3bits, 2), 3); + const __m256i q3l_2 = __lasx_xvandi_b(__lasx_xvsrli_b(q3bits, 4), 3); + const __m256i q3l_3 = __lasx_xvsrli_b(q3bits, 6); + const __m256i q3h_0 = __lasx_xvslli_b(__lasx_xvseqi_b(lasx_xvandi_b_bit(hbits, 4 * j + 0), 0), 2); + const __m256i q3h_1 = __lasx_xvslli_b(__lasx_xvseqi_b(lasx_xvandi_b_bit(hbits, 4 * j + 1), 0), 2); + const __m256i q3h_2 = __lasx_xvslli_b(__lasx_xvseqi_b(lasx_xvandi_b_bit(hbits, 4 * j + 2), 0), 2); + const __m256i q3h_3 = __lasx_xvslli_b(__lasx_xvseqi_b(lasx_xvandi_b_bit(hbits, 4 * j + 3), 0), 2); + const __m256i q3_0 = __lasx_xvor_v(q3h_0, q3l_0); + const __m256i q3_1 = __lasx_xvor_v(q3h_1, q3l_1); + const __m256i q3_2 = __lasx_xvor_v(q3h_2, q3l_2); + const __m256i q3_3 = __lasx_xvor_v(q3h_3, q3l_3); // load Q8 quants const __m256i q8_0 = __lasx_xvld((const __m256i*)q8, 0); q8 += 32; @@ -5444,29 +6575,16 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * restrict s, size_t bs, const void * r const __m256i q8_2 = __lasx_xvld((const __m256i*)q8, 0); q8 += 32; const __m256i q8_3 = __lasx_xvld((const __m256i*)q8, 0); q8 += 32; - // Dot product: we multiply the 2 low bits and 1 high bit part separately, so we can use lasx_maddubs_h, - // and then subtract. The high bit part has the 2 already subtracted (and so, it is zero if the high bit was not set, - // and 2 if the high bit was set) - __m256i q8s_0 = lasx_maddubs_h(q3h_0, q8_0); - __m256i q8s_1 = lasx_maddubs_h(q3h_1, q8_1); - __m256i q8s_2 = lasx_maddubs_h(q3h_2, q8_2); - __m256i q8s_3 = lasx_maddubs_h(q3h_3, q8_3); - - __m256i p16_0 = lasx_maddubs_h(q3l_0, q8_0); - __m256i p16_1 = lasx_maddubs_h(q3l_1, q8_1); - __m256i p16_2 = lasx_maddubs_h(q3l_2, q8_2); - __m256i p16_3 = lasx_maddubs_h(q3l_3, q8_3); - - p16_0 = __lasx_xvsub_h(p16_0, q8s_0); - p16_1 = __lasx_xvsub_h(p16_1, q8s_1); - p16_2 = __lasx_xvsub_h(p16_2, q8s_2); - p16_3 = __lasx_xvsub_h(p16_3, q8s_3); + __m256i p16_0 = lasx_madd_h_b(q8_0, q3_0); + __m256i p16_1 = lasx_madd_h_b(q8_1, q3_1); + __m256i p16_2 = lasx_madd_h_b(q8_2, q3_2); + __m256i p16_3 = lasx_madd_h_b(q8_3, q3_3); // multiply with scales - p16_0 = lasx_madd_h(lasx_shuffle_b(scales[j], get_scale_shuffle_q3k(is + 0)), p16_0); - p16_1 = lasx_madd_h(lasx_shuffle_b(scales[j], get_scale_shuffle_q3k(is + 1)), p16_1); - p16_2 = lasx_madd_h(lasx_shuffle_b(scales[j], get_scale_shuffle_q3k(is + 2)), p16_2); - p16_3 = lasx_madd_h(lasx_shuffle_b(scales[j], get_scale_shuffle_q3k(is + 3)), p16_3); + p16_0 = lasx_madd_h(lasx_xvrepl128vei_h(scales_shuffled, 4 * j + 0), p16_0); + p16_1 = lasx_madd_h(lasx_xvrepl128vei_h(scales_shuffled, 4 * j + 1), p16_1); + p16_2 = lasx_madd_h(lasx_xvrepl128vei_h(scales_shuffled, 4 * j + 2), p16_2); + p16_3 = lasx_madd_h(lasx_xvrepl128vei_h(scales_shuffled, 4 * j + 3), p16_3); // accumulate p16_0 = __lasx_xvadd_w(p16_0, p16_1); @@ -5474,7 +6592,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * restrict s, size_t bs, const void * r sumi = __lasx_xvadd_w(sumi, __lasx_xvadd_w(p16_0, p16_2)); } // multiply with block scale and accumulate - acc = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(d), __lasx_xvffint_s_w(sumi), acc);//FIXME + acc = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(d), __lasx_xvffint_s_w(sumi), acc); } *s = hsum_float_8(acc); @@ -5499,11 +6617,11 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * restrict s, size_t bs, const void * r float sumf = 0; for (int i = 0; i < nb; ++i) { - const uint8_t * restrict q3 = x[i].qs; - const uint8_t * restrict hm = x[i].hmask; - const int8_t * restrict q8 = y[i].qs; + const uint8_t * GGML_RESTRICT q3 = x[i].qs; + const uint8_t * GGML_RESTRICT hm = x[i].hmask; + const int8_t * GGML_RESTRICT q8 = y[i].qs; memset(aux32, 0, 8*sizeof(int32_t)); - int8_t * restrict a = aux8; + int8_t * GGML_RESTRICT a = aux8; uint8_t m = 1; for (int j = 0; j < QK_K; j += 128) { for (int l = 0; l < 32; ++l) a[l] = q3[l] & 3; @@ -5546,7 +6664,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * restrict s, size_t bs, const void * r } -void ggml_vec_dot_q4_K_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) { +void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) { assert(n % QK_K == 0); assert(nrc == 1); UNUSED(nrc); @@ -5554,8 +6672,8 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * restrict s, size_t bs, const void * r UNUSED(by); UNUSED(bs); - const block_q4_K * restrict x = vx; - const block_q8_K * restrict y = vy; + const block_q4_K * GGML_RESTRICT x = vx; + const block_q8_K * GGML_RESTRICT y = vy; const int nb = n / QK_K; @@ -5590,8 +6708,8 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * restrict s, size_t bs, const void * r const uint8_t * scales = (const uint8_t *)utmp; - const uint8_t * restrict q4 = x[i].qs; - const int8_t * restrict q8 = y[i].qs; + const uint8_t * GGML_RESTRICT q4 = x[i].qs; + const int8_t * GGML_RESTRICT q8 = y[i].qs; const int vector_length = ggml_cpu_get_sve_cnt()*8; const svuint8_t m4b = svdup_n_u8(0xf); @@ -5646,7 +6764,7 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * restrict s, size_t bs, const void * r } } *s = sumf; -#elif __ARM_NEON +#elif defined __ARM_NEON const uint8x16_t m4b = vdupq_n_u8(0xf); const int32x4_t mzero = vdupq_n_s32(0); @@ -5678,8 +6796,8 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * restrict s, size_t bs, const void * r const uint8_t * scales = (const uint8_t *)utmp; - const uint8_t * restrict q4 = x[i].qs; - const int8_t * restrict q8 = y[i].qs; + const uint8_t * GGML_RESTRICT q4 = x[i].qs; + const int8_t * GGML_RESTRICT q8 = y[i].qs; int32_t sumi1 = 0; int32_t sumi2 = 0; @@ -5709,6 +6827,107 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * restrict s, size_t bs, const void * r *s = sumf; +#elif defined __wasm_simd128__ + const uint8_t * scales = (const uint8_t*)&utmp[0]; + float sumf = 0; + + for (int i = 0; i < nb; ++i) { + const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); + const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin); // Corrected sign + + const uint8_t * GGML_RESTRICT q4 = x[i].qs; + const int8_t * GGML_RESTRICT q8 = y[i].qs; + + // Process scales and mins + memcpy(utmp, x[i].scales, 12); + utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4); + const uint32_t uaux = utmp[1] & kmask1; + utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4); + utmp[2] = uaux; + utmp[0] &= kmask1; + + // Sum mins * q8sums + int32_t sumi = 0; + const int16_t * GGML_RESTRICT q8sums = y[i].bsums; + const uint8_t * m = (const uint8_t *)&utmp[2]; + for (int j = 0; j < 16; j += 2) { + sumi += (q8sums[j] + q8sums[j+1]) * m[j/2]; + } + sumf -= dmin * sumi; + + int32_t sumi1 = 0; + int32_t sumi2 = 0; + + for (int j = 0; j < QK_K/64; ++j) { + // Load 64 4-bit weights (32 bytes) + const v128_t q4x0 = wasm_v128_load(q4); + const v128_t q4x1 = wasm_v128_load(q4 + 16); + q4 += 32; + + // Split into low/high nibbles + const v128_t q4l0 = wasm_v128_and(q4x0, wasm_i8x16_splat(0x0F)); + const v128_t q4h0 = wasm_u8x16_shr(q4x0, 4); + const v128_t q4l1 = wasm_v128_and(q4x1, wasm_i8x16_splat(0x0F)); + const v128_t q4h1 = wasm_u8x16_shr(q4x1, 4); + + // Load 64 8-bit values (64 bytes) + const v128_t q8x0 = wasm_v128_load(q8); + const v128_t q8x1 = wasm_v128_load(q8 + 16); + const v128_t q8x2 = wasm_v128_load(q8 + 32); + const v128_t q8x3 = wasm_v128_load(q8 + 48); + q8 += 64; + + // Low nibble products + v128_t vacc1 = wasm_i32x4_dot_i16x8( + wasm_i16x8_extend_low_i8x16(q4l0), + wasm_i16x8_extend_low_i8x16(q8x0) + ); + vacc1 = wasm_i32x4_add(vacc1, wasm_i32x4_dot_i16x8( + wasm_i16x8_extend_high_i8x16(q4l0), + wasm_i16x8_extend_high_i8x16(q8x0) + )); + vacc1 = wasm_i32x4_add(vacc1, wasm_i32x4_dot_i16x8( + wasm_i16x8_extend_low_i8x16(q4l1), + wasm_i16x8_extend_low_i8x16(q8x1) + )); + vacc1 = wasm_i32x4_add(vacc1, wasm_i32x4_dot_i16x8( + wasm_i16x8_extend_high_i8x16(q4l1), + wasm_i16x8_extend_high_i8x16(q8x1) + )); + + // High nibble products + v128_t vacc2 = wasm_i32x4_dot_i16x8( + wasm_i16x8_extend_low_i8x16(q4h0), + wasm_i16x8_extend_low_i8x16(q8x2) + ); + vacc2 = wasm_i32x4_add(vacc2, wasm_i32x4_dot_i16x8( + wasm_i16x8_extend_high_i8x16(q4h0), + wasm_i16x8_extend_high_i8x16(q8x2) + )); + vacc2 = wasm_i32x4_add(vacc2, wasm_i32x4_dot_i16x8( + wasm_i16x8_extend_low_i8x16(q4h1), + wasm_i16x8_extend_low_i8x16(q8x3) + )); + vacc2 = wasm_i32x4_add(vacc2, wasm_i32x4_dot_i16x8( + wasm_i16x8_extend_high_i8x16(q4h1), + wasm_i16x8_extend_high_i8x16(q8x3) + )); + + // Accumulate scaled results + int32_t vacc1_sum = wasm_i32x4_extract_lane(vacc1, 0) + wasm_i32x4_extract_lane(vacc1, 1) + + wasm_i32x4_extract_lane(vacc1, 2) + wasm_i32x4_extract_lane(vacc1, 3); + sumi1 += vacc1_sum * scales[2*j]; + + int32_t vacc2_sum = wasm_i32x4_extract_lane(vacc2, 0) + wasm_i32x4_extract_lane(vacc2, 1) + + wasm_i32x4_extract_lane(vacc2, 2) + wasm_i32x4_extract_lane(vacc2, 3); + sumi2 += vacc2_sum * scales[2*j+1]; + } + + sumf += d * (sumi1 + sumi2); + } + + *s = sumf; + #elif defined __AVX2__ const __m256i m4 = _mm256_set1_epi8(0xF); @@ -5728,8 +6947,8 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * restrict s, size_t bs, const void * r utmp[2] = uaux; utmp[0] &= kmask1; - const uint8_t * restrict q4 = x[i].qs; - const int8_t * restrict q8 = y[i].qs; + const uint8_t * GGML_RESTRICT q4 = x[i].qs; + const int8_t * GGML_RESTRICT q8 = y[i].qs; const __m256i mins_and_scales = _mm256_cvtepu8_epi16(_mm_set_epi32(utmp[3], utmp[2], utmp[1], utmp[0])); @@ -5787,8 +7006,8 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * restrict s, size_t bs, const void * r const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin); - const uint8_t * restrict q4 = x[i].qs; - const int8_t * restrict q8 = y[i].qs; + const uint8_t * GGML_RESTRICT q4 = x[i].qs; + const int8_t * GGML_RESTRICT q8 = y[i].qs; memcpy(utmp, x[i].scales, 12); utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4); @@ -5861,69 +7080,181 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * restrict s, size_t bs, const void * r const uint8_t * scales = (const uint8_t*)&utmp[0]; const uint8_t * mins = (const uint8_t*)&utmp[2]; + const int vector_length = __riscv_vlenb() * 8; float sumf = 0; - for (int i = 0; i < nb; ++i) { + switch (vector_length) { + case 256: + for (int i = 0; i < nb; ++i) { - size_t vl = 8; + size_t vl = 8; - const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); - const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin); + const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); + const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin); - vint16mf2_t q8sums_0 = __riscv_vlse16_v_i16mf2(y[i].bsums, 4, vl); - vint16mf2_t q8sums_1 = __riscv_vlse16_v_i16mf2(y[i].bsums+1, 4, vl); - vint16mf2_t q8sums = __riscv_vadd_vv_i16mf2(q8sums_0, q8sums_1, vl); + vint16mf2_t q8sums_0 = __riscv_vlse16_v_i16mf2(y[i].bsums, 4, vl); + vint16mf2_t q8sums_1 = __riscv_vlse16_v_i16mf2(y[i].bsums+1, 4, vl); + vint16mf2_t q8sums = __riscv_vadd_vv_i16mf2(q8sums_0, q8sums_1, vl); - memcpy(utmp, x[i].scales, 12); - utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4); - const uint32_t uaux = utmp[1] & kmask1; - utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4); - utmp[2] = uaux; - utmp[0] &= kmask1; + memcpy(utmp, x[i].scales, 12); + utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4); + const uint32_t uaux = utmp[1] & kmask1; + utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4); + utmp[2] = uaux; + utmp[0] &= kmask1; - vuint8mf4_t mins8 = __riscv_vle8_v_u8mf4(mins, vl); - vint16mf2_t v_mins = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vzext_vf2_u16mf2(mins8, vl)); - vint32m1_t prod = __riscv_vwmul_vv_i32m1(q8sums, v_mins, vl); + vuint8mf4_t mins8 = __riscv_vle8_v_u8mf4(mins, vl); + vint16mf2_t v_mins = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vzext_vf2_u16mf2(mins8, vl)); + vint32m1_t prod = __riscv_vwmul_vv_i32m1(q8sums, v_mins, vl); - vint32m1_t sumi = __riscv_vredsum_vs_i32m1_i32m1(prod, __riscv_vmv_v_x_i32m1(0, 1), vl); - sumf -= dmin * __riscv_vmv_x_s_i32m1_i32(sumi); + vint32m1_t sumi = __riscv_vredsum_vs_i32m1_i32m1(prod, __riscv_vmv_v_x_i32m1(0, 1), vl); + sumf -= dmin * __riscv_vmv_x_s_i32m1_i32(sumi); - const uint8_t * restrict q4 = x[i].qs; - const int8_t * restrict q8 = y[i].qs; + const uint8_t * GGML_RESTRICT q4 = x[i].qs; + const int8_t * GGML_RESTRICT q8 = y[i].qs; - vl = 32; + vl = 32; - int32_t sum_1 = 0; - int32_t sum_2 = 0; + int32_t sum_1 = 0; + int32_t sum_2 = 0; - vint16m1_t vzero = __riscv_vmv_v_x_i16m1(0, 1); + vint16m1_t vzero = __riscv_vmv_v_x_i16m1(0, 1); - for (int j = 0; j < QK_K/64; ++j) { - // load Q4 - vuint8m1_t q4_x = __riscv_vle8_v_u8m1(q4, vl); + for (int j = 0; j < QK_K/64; ++j) { + // load Q4 + vuint8m1_t q4_x = __riscv_vle8_v_u8m1(q4, vl); - // load Q8 and multiply it with lower Q4 nibble - vint8m1_t q8_0 = __riscv_vle8_v_i8m1(q8, vl); - vint8m1_t q4_0 = __riscv_vreinterpret_v_u8m1_i8m1(__riscv_vand_vx_u8m1(q4_x, 0x0F, vl)); - vint16m2_t qv_0 = __riscv_vwmul_vv_i16m2(q4_0, q8_0, vl); - vint16m1_t vs_0 = __riscv_vredsum_vs_i16m2_i16m1(qv_0, vzero, vl); + // load Q8 and multiply it with lower Q4 nibble + vint8m1_t q8_0 = __riscv_vle8_v_i8m1(q8, vl); + vint8m1_t q4_0 = __riscv_vreinterpret_v_u8m1_i8m1(__riscv_vand_vx_u8m1(q4_x, 0x0F, vl)); + vint16m2_t qv_0 = __riscv_vwmul_vv_i16m2(q4_0, q8_0, vl); + vint16m1_t vs_0 = __riscv_vredsum_vs_i16m2_i16m1(qv_0, vzero, vl); - sum_1 += __riscv_vmv_x_s_i16m1_i16(vs_0) * scales[2*j+0]; + sum_1 += __riscv_vmv_x_s_i16m1_i16(vs_0) * scales[2*j+0]; - // load Q8 and multiply it with upper Q4 nibble - vint8m1_t q8_1 = __riscv_vle8_v_i8m1(q8+32, vl); - vint8m1_t q4_1 = __riscv_vreinterpret_v_u8m1_i8m1(__riscv_vsrl_vx_u8m1(q4_x, 0x04, vl)); - vint16m2_t qv_1 = __riscv_vwmul_vv_i16m2(q4_1, q8_1, vl); - vint16m1_t vs_1 = __riscv_vredsum_vs_i16m2_i16m1(qv_1, vzero, vl); + // load Q8 and multiply it with upper Q4 nibble + vint8m1_t q8_1 = __riscv_vle8_v_i8m1(q8+32, vl); + vint8m1_t q4_1 = __riscv_vreinterpret_v_u8m1_i8m1(__riscv_vsrl_vx_u8m1(q4_x, 0x04, vl)); + vint16m2_t qv_1 = __riscv_vwmul_vv_i16m2(q4_1, q8_1, vl); + vint16m1_t vs_1 = __riscv_vredsum_vs_i16m2_i16m1(qv_1, vzero, vl); - sum_2 += __riscv_vmv_x_s_i16m1_i16(vs_1) * scales[2*j+1]; + sum_2 += __riscv_vmv_x_s_i16m1_i16(vs_1) * scales[2*j+1]; - q4 += 32; q8 += 64; + q4 += 32; q8 += 64; + + } + + sumf += d*(sum_1 + sum_2); } + break; + case 128: + for (int i = 0; i < nb; ++i) { + const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); + const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin); - sumf += d*(sum_1 + sum_2); + int tmp, tmp2, sumi; + __asm__ __volatile__( + "vsetivli zero, 12, e8, m1\n\t" + "vle8.v v1, (%[s6b])\n\t" // {aux[0], aux[1], aux[2]} + "vsetivli zero, 4, e32, m1\n\t" + "vslidedown.vi v2, v1, 2\n\t" + "vmv1r.v v3, v2\n\t" + "vslideup.vi v2, v3, 1\n\t" // {aux[2], aux[2]} + "vsetivli zero, 2, e32, m1\n\t" + "vmv.v.i v4, 4\n\t" + "vand.vx v8, v1, %[kmask1]\n\t" + "vslide1up.vx v5, v4, zero\n\t" // {0, 4} + "vsrl.vi v6, v1, 6\n\t" + "vsrl.vv v7, v2, v5\n\t" + "vand.vx v0, v6, %[kmask3]\n\t" + "vand.vx v2, v7, %[kmask2]\n\t" + "vsll.vi v6, v0, 4\n\t" + "li %[t2], 8\n\t" + "addi %[t1], %[utmp], 4\n\t" + "vor.vv v1, v6, v2\n\t" + "vsse32.v v8, (%[utmp]), %[t2]\n\t" + "vsse32.v v1, (%[t1]), %[t2]\n\t" + "vsetivli zero, 8, e16, m1\n\t" + "vle32.v v2, (%[bsums])\n\t" + "vnsrl.wi v0, v2, 0\n\t" + "vnsrl.wi v1, v2, 16\n\t" + "vadd.vv v2, v0, v1\n\t" + "vle8.v v3, (%[mins])\n\t" + "vzext.vf2 v4, v3\n\t" + "vwmul.vv v6, v4, v2\n\t" + "vmv.v.x v0, zero\n\t" + "vsetivli zero, 8, e32, m2\n\t" + "vredsum.vs v0, v6, v0\n\t" + "vmv.x.s %[sumi], v0" + : [t1] "=&r" (tmp), [t2] "=&r" (tmp2), [sumi] "=&r" (sumi) + : [bsums] "r" (y[i].bsums), [mins] "r" (mins), [utmp] "r" (utmp) + , [s6b] "r" (x[i].scales), [kmask1] "r" (kmask1) + , [kmask2] "r" (kmask2), [kmask3] "r" (kmask3) + : "memory" + , "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7" + , "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15" + , "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23" + , "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31" + ); + sumf -= dmin * sumi; + const uint8_t * restrict q4 = x[i].qs; + const int8_t * restrict q8 = y[i].qs; + + sumi = 0; + const uint8_t * scale = scales; + + for (int j = 0; j < QK_K/128; ++j) { + int vl128 = 128, vl64 = 64, vl32 = 32; + __asm__ __volatile__( + "vsetvli zero, %[vl128], e8, m8\n\t" + "vle8.v v8, (%[q8])\n\t" + "vsetvli zero, %[vl64], e8, m4\n\t" + "vle8.v v0, (%[q4])\n\t" + "vsrl.vi v4, v0, 4\n\t" + "vand.vi v0, v0, 0xF\n\t" + "vsetvli zero, %[vl32], e8, m2\n\t" + "vwmul.vv v28, v6, v14\n\t" + "vwmul.vv v20, v4, v10\n\t" + "vwmul.vv v24, v2, v12\n\t" + "vwmul.vv v16, v0, v8\n\t" + "vsetivli zero, 4, e32, m1\n\t" + "vle8.v v2, (%[scale])\n\t" + "vmv.v.x v0, zero\n\t" + "vzext.vf4 v1, v2\n\t" + "vsetvli zero, %[vl32], e16, m4\n\t" + "vwredsum.vs v6, v24, v0\n\t" + "vwredsum.vs v7, v28, v0\n\t" + "vwredsum.vs v4, v16, v0\n\t" + "vwredsum.vs v5, v20, v0\n\t" + "vsetivli zero, 4, e32, m1\n\t" + "vslideup.vi v6, v7, 1\n\t" + "vslideup.vi v4, v5, 1\n\t" + "vslideup.vi v4, v6, 2\n\t" + "vmul.vv v8, v4, v1\n\t" + "vredsum.vs v0, v8, v0\n\t" + "vmv.x.s %[tmp], v0\n\t" + "add %[sumi], %[sumi], %[tmp]" + : [tmp] "=&r" (tmp), [sumi] "+&r" (sumi) + : [vl128] "r" (vl128), [vl64] "r" (vl64), [vl32] "r" (vl32) + , [q4] "r" (q4), [q8] "r" (q8), [scale] "r" (scale) + : "memory" + , "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7" + , "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15" + , "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23" + , "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31" + ); + + q4 += 64; q8 += 128; scale += 4; + } + + sumf += d * sumi; + } + break; + default: + assert(false && "Unsupported vector length"); + break; } *s = sumf; @@ -5990,8 +7321,8 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * restrict s, size_t bs, const void * r vector signed int vsumi2 = v0; vector signed int vsumi3 = v0; - const uint8_t * restrict q4 = x[i].qs; - const int8_t * restrict q8 = y[i].qs; + const uint8_t * GGML_RESTRICT q4 = x[i].qs; + const int8_t * GGML_RESTRICT q8 = y[i].qs; for (int j = 0; j < QK_K/64; j+=2) { __builtin_prefetch(q4, 0, 1); @@ -6066,11 +7397,6 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * restrict s, size_t bs, const void * r *s = vec_extract(vsumf0, 0); #elif defined __loongarch_asx - GGML_UNUSED(kmask1); - GGML_UNUSED(kmask2); - GGML_UNUSED(kmask3); - - const __m256i m4 = __lasx_xvreplgr2vr_b(0xF); __m256 acc = (__m256)__lasx_xvldi(0); __m128 acc_m = (__m128)__lsx_vldi(0); @@ -6087,36 +7413,37 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * restrict s, size_t bs, const void * r utmp[2] = uaux; utmp[0] &= kmask1; - const uint8_t * restrict q4 = x[i].qs; - const int8_t * restrict q8 = y[i].qs; + const uint8_t * GGML_RESTRICT q4 = x[i].qs; + const int8_t * GGML_RESTRICT q8 = y[i].qs; - const __m256i mins_and_scales = lasx_extu8_16(lsx_set_w(utmp[3], utmp[2], utmp[1], utmp[0])); + const __m128i mins_and_scales128 = lsx_set_w(utmp[3], utmp[2], utmp[1], utmp[0]); + const __m128i mins128 = __lsx_vexth_h_b(mins_and_scales128); + const __m128i scales128 = __lsx_vsllwil_h_b(mins_and_scales128, 0); const __m256i q8sums = __lasx_xvld((const __m256i*)y[i].bsums, 0); const __m128i q8s = lsx_hadd_h(lasx_extracti128(q8sums, 0), lasx_extracti128(q8sums, 1)); - const __m128i prod = lsx_madd_h(lasx_extracti128(mins_and_scales, 1), q8s); + const __m128i prod = lsx_madd_h(mins128, q8s); acc_m = __lsx_vfmadd_s(__lsx_vreplfr2vr_s(dmin), __lsx_vffint_s_w(prod), acc_m); - const __m128i sc128 = lasx_extracti128(mins_and_scales, 0); - const __m256i scales = lasx_insertf128(sc128, sc128); + const __m256i scales = lasx_insertf128(scales128, scales128); __m256i sumi = __lasx_xvldi(0); for (int j = 0; j < QK_K/64; ++j) { - const __m256i scale_l = lasx_shuffle_b(scales, get_scale_shuffle_k4(2*j+0)); - const __m256i scale_h = lasx_shuffle_b(scales, get_scale_shuffle_k4(2*j+1)); + const __m256i scale_l = lasx_xvrepl128vei_h(scales, 2 * j + 0); + const __m256i scale_h = lasx_xvrepl128vei_h(scales, 2 * j + 1); const __m256i q4bits = __lasx_xvld((const __m256i*)q4, 0); q4 += 32; - const __m256i q4l = __lasx_xvand_v(q4bits, m4); - const __m256i q4h = __lasx_xvand_v(__lasx_xvsrli_h(q4bits, 4), m4); + const __m256i q4l = __lasx_xvandi_b(q4bits, 0xf); + const __m256i q4h = __lasx_xvsrli_b(q4bits, 4); const __m256i q8l = __lasx_xvld((const __m256i*)q8, 0); q8 += 32; - __m256i p16l = lasx_maddubs_h(q4l, q8l); + __m256i p16l = lasx_madd_h_b(q4l, q8l); p16l = lasx_madd_h(scale_l, p16l); const __m256i q8h = __lasx_xvld((const __m256i*)q8, 0); q8 += 32; - __m256i p16h = lasx_maddubs_h(q4h, q8h); + __m256i p16h = lasx_madd_h_b(q4h, q8h); p16h = lasx_madd_h(scale_h, p16h); const __m256i sumj = __lasx_xvadd_w(p16l, p16h); @@ -6134,6 +7461,77 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * restrict s, size_t bs, const void * r *s = hsum_float_8(acc) + ((v4f32)acc_m)[0]; +#elif defined(__VXE__) || defined(__VXE2__) + const uint8x16_t v_lm = vec_splat_u8(0x0F); + const int32x4_t v_z = vec_splat_s32(0); + + uint8x16_t v_x[2]; + int8x16_t v_xl[2]; + int8x16_t v_y[2]; + + float sumf = 0; + + for (int i = 0; i < nb; ++i) { + const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); + const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin); + + const int16x8_t v_ysumsl = vec_xl(0 , y[i].bsums); + const int16x8_t v_ysumsh = vec_xl(16, y[i].bsums); + const int16x8_t v_ysums = vec_padd_s16(v_ysumsl, v_ysumsh); + + memcpy(utmp, x[i].scales, 12); + + uint32x4_t v_mins8 = { 0 }; + v_mins8 = vec_insert(utmp[1] & kmask1, v_mins8, 0); + v_mins8 = vec_insert(((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4), v_mins8, 1); + + utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4); + utmp[0] &= kmask1; + + const int16x8_t v_minsh = (int16x8_t)vec_unpackh((uint8x16_t)v_mins8); + + const int32x4_t v_minso = vec_mulo(v_ysums, v_minsh); + const int32x4_t v_minse = vec_mule(v_ysums, v_minsh); + const int32x4_t v_mins = v_minso + v_minse; + sumf -= dmin * (v_mins[0] + v_mins[1] + v_mins[2] + v_mins[3]); + + const uint8_t * scales = (const uint8_t *)utmp; + const uint8_t * GGML_RESTRICT x0 = x[i].qs; + const int8_t * GGML_RESTRICT y0 = y[i].qs; + + int32_t sumi1 = 0; + int32_t sumi2 = 0; + + for (int j = 0; j < QK_K/64; ++j) { + v_x[0] = vec_xl(0 , x0); + v_x[1] = vec_xl(16, x0); + x0 += 32; + + v_y[0] = vec_xl(0 , y0); + v_y[1] = vec_xl(16, y0); + y0 += 32; + + v_xl[0] = (int8x16_t)vec_and(v_x[0], v_lm); + v_xl[1] = (int8x16_t)vec_and(v_x[1], v_lm); + + const int32x4_t p1 = ggml_vec_dot(ggml_vec_dot(v_z, v_xl[0], v_y[0]), v_xl[1], v_y[1]); + sumi1 += (p1[0] + p1[1] + p1[2] + p1[3]) * scales[2*j+0]; + + v_y[0] = vec_xl(0 , y0); + v_y[1] = vec_xl(16, y0); + y0 += 32; + + v_xl[0] = (int8x16_t)vec_sr(v_x[0], 4); + v_xl[1] = (int8x16_t)vec_sr(v_x[1], 4); + + const int32x4_t p2 = ggml_vec_dot(ggml_vec_dot(v_z, v_xl[0], v_y[0]), v_xl[1], v_y[1]); + sumi2 += (p2[0] + p2[1] + p2[2] + p2[3]) * scales[2*j+1]; + } + + sumf += d * (sumi1 + sumi2); + } + + *s = sumf; #else const uint8_t * scales = (const uint8_t*)&utmp[0]; @@ -6147,10 +7545,10 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * restrict s, size_t bs, const void * r float sumf = 0; for (int i = 0; i < nb; ++i) { - const uint8_t * restrict q4 = x[i].qs; - const int8_t * restrict q8 = y[i].qs; + const uint8_t * GGML_RESTRICT q4 = x[i].qs; + const int8_t * GGML_RESTRICT q8 = y[i].qs; memset(aux32, 0, 8*sizeof(int32_t)); - int8_t * restrict a = aux8; + int8_t * GGML_RESTRICT a = aux8; for (int j = 0; j < QK_K/64; ++j) { for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] & 0xF); a += 32; @@ -6193,7 +7591,7 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * restrict s, size_t bs, const void * r #endif } -void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) { +void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) { assert(n % QK_K == 0); assert(nrc == 1); UNUSED(nrc); @@ -6201,8 +7599,8 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, size_t bs, const void * r UNUSED(by); UNUSED(bs); - const block_q5_K * restrict x = vx; - const block_q8_K * restrict y = vy; + const block_q5_K * GGML_RESTRICT x = vx; + const block_q8_K * GGML_RESTRICT y = vy; const int nb = n / QK_K; @@ -6244,9 +7642,9 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, size_t bs, const void * r const uint8_t * scales = (const uint8_t *)utmp; - const uint8_t * restrict q5 = x[i].qs; - const uint8_t * restrict qh = x[i].qh; - const int8_t * restrict q8 = y[i].qs; + const uint8_t * GGML_RESTRICT q5 = x[i].qs; + const uint8_t * GGML_RESTRICT qh = x[i].qh; + const int8_t * GGML_RESTRICT q8 = y[i].qs; ggml_uint8x16x2_t qhbits = ggml_vld1q_u8_x2(qh); @@ -6291,8 +7689,8 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, size_t bs, const void * r float summs = 0.f; for (int i = 0; i < nb; ++i) { - const uint8_t * restrict q5 = x[i].qs; - const int8_t * restrict q8 = y[i].qs; + const uint8_t * GGML_RESTRICT q5 = x[i].qs; + const int8_t * GGML_RESTRICT q8 = y[i].qs; const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin); @@ -6375,8 +7773,8 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, size_t bs, const void * r const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin); - const uint8_t * restrict q5 = x[i].qs; - const int8_t * restrict q8 = y[i].qs; + const uint8_t * GGML_RESTRICT q5 = x[i].qs; + const int8_t * GGML_RESTRICT q8 = y[i].qs; memcpy(utmp, x[i].scales, 12); utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4); @@ -6459,6 +7857,118 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, size_t bs, const void * r *s = hsum_float_8(acc) + summs; +#elif defined __wasm_simd128__ + //const uint8_t * scales = (const uint8_t*)&utmp[0]; + float sumf = 0; + + for (int i = 0; i < nb; ++i) { + const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); + const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin); // Fixed sign + + const uint8_t * GGML_RESTRICT q5 = x[i].qs; + const uint8_t * GGML_RESTRICT qh = x[i].qh; + const int8_t * GGML_RESTRICT q8 = y[i].qs; + + // Process scales and mins + memcpy(utmp, x[i].scales, 12); + utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4); + const uint32_t uaux = utmp[1] & kmask1; + utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4); + utmp[2] = uaux; + utmp[0] &= kmask1; + + // Sum mins * q8sums + int32_t sumi_mins = 0; + const int16_t * GGML_RESTRICT q8sums = y[i].bsums; + const uint8_t * m = (const uint8_t *)&utmp[2]; + for (int j = 0; j < 16; j += 2) { + sumi_mins += (q8sums[j] + q8sums[j+1]) * m[j/2]; + } + sumf -= dmin * sumi_mins; // Correct subtraction + + v128_t qh0 = wasm_v128_load(qh); + v128_t qh1 = wasm_v128_load(qh + 16); + const uint8_t * sc = (const uint8_t *)utmp; + + int32_t sumi = 0; + + for (int j = 0; j < QK_K/64; ++j) { + const int shift = j * 2; + v128_t qh_shift0 = wasm_u8x16_shr(qh0, shift); + v128_t qh_shift1 = wasm_u8x16_shr(qh1, shift); + + v128_t qh_low0 = wasm_i8x16_shl(wasm_v128_and(qh_shift0, wasm_i8x16_splat(0x01)), 4); + v128_t qh_high0 = wasm_i8x16_shl(wasm_v128_and(qh_shift0, wasm_i8x16_splat(0x02)), 3); + v128_t qh_low1 = wasm_i8x16_shl(wasm_v128_and(qh_shift1, wasm_i8x16_splat(0x01)), 4); + v128_t qh_high1 = wasm_i8x16_shl(wasm_v128_and(qh_shift1, wasm_i8x16_splat(0x02)), 3); + + v128_t q5_0 = wasm_v128_load(q5); + v128_t q5_1 = wasm_v128_load(q5 + 16); + q5 += 32; + + v128_t q5l_0 = wasm_v128_or(wasm_v128_and(q5_0, wasm_i8x16_splat(0x0F)), qh_low0); + v128_t q5h_0 = wasm_v128_or(wasm_u8x16_shr(q5_0, 4), qh_high0); + v128_t q5l_1 = wasm_v128_or(wasm_v128_and(q5_1, wasm_i8x16_splat(0x0F)), qh_low1); + v128_t q5h_1 = wasm_v128_or(wasm_u8x16_shr(q5_1, 4), qh_high1); + + v128_t q8_0 = wasm_v128_load(q8); + v128_t q8_1 = wasm_v128_load(q8 + 16); + v128_t q8_2 = wasm_v128_load(q8 + 32); + v128_t q8_3 = wasm_v128_load(q8 + 48); + q8 += 64; + + // Process low quants + v128_t pl0 = wasm_i32x4_dot_i16x8( + wasm_i16x8_extend_low_i8x16(q5l_0), + wasm_i16x8_extend_low_i8x16(q8_0) + ); + pl0 = wasm_i32x4_add(pl0, wasm_i32x4_dot_i16x8( + wasm_i16x8_extend_high_i8x16(q5l_0), + wasm_i16x8_extend_high_i8x16(q8_0) + )); + v128_t pl1 = wasm_i32x4_dot_i16x8( + wasm_i16x8_extend_low_i8x16(q5l_1), + wasm_i16x8_extend_low_i8x16(q8_1) + ); + pl1 = wasm_i32x4_add(pl1, wasm_i32x4_dot_i16x8( + wasm_i16x8_extend_high_i8x16(q5l_1), + wasm_i16x8_extend_high_i8x16(q8_1) + )); + v128_t sum_low = wasm_i32x4_add(pl0, pl1); + + // Process high quants + v128_t ph0 = wasm_i32x4_dot_i16x8( + wasm_i16x8_extend_low_i8x16(q5h_0), + wasm_i16x8_extend_low_i8x16(q8_2) + ); + ph0 = wasm_i32x4_add(ph0, wasm_i32x4_dot_i16x8( + wasm_i16x8_extend_high_i8x16(q5h_0), + wasm_i16x8_extend_high_i8x16(q8_2) + )); + v128_t ph1 = wasm_i32x4_dot_i16x8( + wasm_i16x8_extend_low_i8x16(q5h_1), + wasm_i16x8_extend_low_i8x16(q8_3) + ); + ph1 = wasm_i32x4_add(ph1, wasm_i32x4_dot_i16x8( + wasm_i16x8_extend_high_i8x16(q5h_1), + wasm_i16x8_extend_high_i8x16(q8_3) + )); + v128_t sum_high = wasm_i32x4_add(ph0, ph1); + + // Accumulate with scale factors + int32_t sl = wasm_i32x4_extract_lane(sum_low, 0) + wasm_i32x4_extract_lane(sum_low, 1) + + wasm_i32x4_extract_lane(sum_low, 2) + wasm_i32x4_extract_lane(sum_low, 3); + int32_t sh = wasm_i32x4_extract_lane(sum_high, 0) + wasm_i32x4_extract_lane(sum_high, 1) + + wasm_i32x4_extract_lane(sum_high, 2) + wasm_i32x4_extract_lane(sum_high, 3); + + sumi += sl * sc[2*j] + sh * sc[2*j+1]; + } + + sumf += d * sumi; + } + + *s = sumf; + #elif defined __riscv_v_intrinsic const uint8_t * scales = (const uint8_t*)&utmp[0]; @@ -6473,16 +7983,16 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, size_t bs, const void * r vl = 8; - const uint8_t * restrict q5 = x[i].qs; - const uint8_t * restrict hm = x[i].qh; - const int8_t * restrict q8 = y[i].qs; + const uint8_t * GGML_RESTRICT q5 = x[i].qs; + const uint8_t * GGML_RESTRICT hm = x[i].qh; + const int8_t * GGML_RESTRICT q8 = y[i].qs; const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d; - vint16mf2_t q8sums_0 = __riscv_vlse16_v_i16mf2(y[i].bsums, 4, vl); - vint16mf2_t q8sums_1 = __riscv_vlse16_v_i16mf2(y[i].bsums+1, 4, vl); - vint16mf2_t q8sums = __riscv_vadd_vv_i16mf2(q8sums_0, q8sums_1, vl); + vint16m1_t q8sums_0 = __riscv_vlse16_v_i16m1(y[i].bsums, 4, vl); + vint16m1_t q8sums_1 = __riscv_vlse16_v_i16m1(y[i].bsums+1, 4, vl); + vint16m1_t q8sums = __riscv_vadd_vv_i16m1(q8sums_0, q8sums_1, vl); memcpy(utmp, x[i].scales, 12); utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4); @@ -6491,11 +8001,11 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, size_t bs, const void * r utmp[2] = uaux; utmp[0] &= kmask1; - vuint8mf4_t mins8 = __riscv_vle8_v_u8mf4(mins, vl); - vint16mf2_t v_mins = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vzext_vf2_u16mf2(mins8, vl)); - vint32m1_t prod = __riscv_vwmul_vv_i32m1(q8sums, v_mins, vl); + vuint8mf2_t mins8 = __riscv_vle8_v_u8mf2(mins, vl); + vint16m1_t v_mins = __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(mins8, vl)); + vint32m2_t prod = __riscv_vwmul_vv_i32m2(q8sums, v_mins, vl); - vint32m1_t sumi = __riscv_vredsum_vs_i32m1_i32m1(prod, __riscv_vmv_v_x_i32m1(0, 1), vl); + vint32m1_t sumi = __riscv_vredsum_vs_i32m2_i32m1(prod, __riscv_vmv_v_x_i32m1(0, 1), vl); sumf -= dmin * __riscv_vmv_x_s_i32m1_i32(sumi); vl = 32; @@ -6504,43 +8014,42 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, size_t bs, const void * r uint8_t m = 1; vint32m1_t vzero = __riscv_vmv_v_x_i32m1(0, 1); - vuint8m1_t vqh = __riscv_vle8_v_u8m1(hm, vl); + vuint8m2_t vqh = __riscv_vle8_v_u8m2(hm, vl); for (int j = 0; j < QK_K/64; ++j) { // load Q5 and Q8 - vuint8m1_t q5_x = __riscv_vle8_v_u8m1(q5, vl); - vint8m1_t q8_y1 = __riscv_vle8_v_i8m1(q8, vl); - vint8m1_t q8_y2 = __riscv_vle8_v_i8m1(q8+32, vl); + vuint8m2_t q5_x = __riscv_vle8_v_u8m2(q5, vl); + vint8m2_t q8_y1 = __riscv_vle8_v_i8m2(q8, vl); + vint8m2_t q8_y2 = __riscv_vle8_v_i8m2(q8+32, vl); // compute mask for addition - vint8m1_t q5_a = __riscv_vreinterpret_v_u8m1_i8m1(__riscv_vand_vx_u8m1(q5_x, 0x0F, vl)); - vuint8m1_t qh_m1 = __riscv_vand_vx_u8m1(vqh, m, vl); - vbool8_t vmask_1 = __riscv_vmsne_vx_u8m1_b8(qh_m1, 0, vl); - vint8m1_t q5_m1 = __riscv_vadd_vx_i8m1_mu(vmask_1, q5_a, q5_a, 16, vl); + vint8m2_t q5_a = __riscv_vreinterpret_v_u8m2_i8m2(__riscv_vand_vx_u8m2(q5_x, 0x0F, vl)); + vuint8m2_t qh_m1 = __riscv_vand_vx_u8m2(vqh, m, vl); + vbool4_t vmask_1 = __riscv_vmsne_vx_u8m2_b4(qh_m1, 0, vl); + vint8m2_t q5_m1 = __riscv_vadd_vx_i8m2_mu(vmask_1, q5_a, q5_a, 16, vl); m <<= 1; - vint8m1_t q5_l = __riscv_vreinterpret_v_u8m1_i8m1(__riscv_vsrl_vx_u8m1(q5_x, 0x04, vl)); - vuint8m1_t qh_m2 = __riscv_vand_vx_u8m1(vqh, m, vl); - vbool8_t vmask_2 = __riscv_vmsne_vx_u8m1_b8(qh_m2, 0, vl); - vint8m1_t q5_m2 = __riscv_vadd_vx_i8m1_mu(vmask_2, q5_l, q5_l, 16, vl); + vint8m2_t q5_l = __riscv_vreinterpret_v_u8m2_i8m2(__riscv_vsrl_vx_u8m2(q5_x, 0x04, vl)); + vuint8m2_t qh_m2 = __riscv_vand_vx_u8m2(vqh, m, vl); + vbool4_t vmask_2 = __riscv_vmsne_vx_u8m2_b4(qh_m2, 0, vl); + vint8m2_t q5_m2 = __riscv_vadd_vx_i8m2_mu(vmask_2, q5_l, q5_l, 16, vl); m <<= 1; - vint16m2_t v0 = __riscv_vwmul_vv_i16m2(q5_m1, q8_y1, vl); - vint16m2_t v1 = __riscv_vwmul_vv_i16m2(q5_m2, q8_y2, vl); + vint16m4_t v0 = __riscv_vwmul_vv_i16m4(q5_m1, q8_y1, vl); + vint16m4_t v1 = __riscv_vwmul_vv_i16m4(q5_m2, q8_y2, vl); - vint32m4_t vs1 = __riscv_vwmul_vx_i32m4(v0, scales[is++], vl); - vint32m4_t vs2 = __riscv_vwmul_vx_i32m4(v1, scales[is++], vl); + vint32m8_t vs1 = __riscv_vwmul_vx_i32m8(v0, scales[is++], vl); + vint32m8_t vs2 = __riscv_vwmul_vx_i32m8(v1, scales[is++], vl); - vint32m1_t vacc1 = __riscv_vredsum_vs_i32m4_i32m1(vs1, vzero, vl); - vint32m1_t vacc2 = __riscv_vredsum_vs_i32m4_i32m1(vs2, vzero, vl); + vint32m1_t vacc1 = __riscv_vredsum_vs_i32m8_i32m1(vs1, vzero, vl); + vint32m1_t vacc2 = __riscv_vredsum_vs_i32m8_i32m1(vs2, vacc1, vl); - aux32 += __riscv_vmv_x_s_i32m1_i32(vacc1) + __riscv_vmv_x_s_i32m1_i32(vacc2); + aux32 += __riscv_vmv_x_s_i32m1_i32(vacc2); q5 += 32; q8 += 64; } - vfloat32m1_t vaux = __riscv_vfmul_vf_f32m1(__riscv_vfmv_v_f_f32m1(aux32, 1), d, 1); - sums += __riscv_vfmv_f_s_f32m1_f32(vaux); + sums += aux32 * d; } @@ -6614,8 +8123,8 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, size_t bs, const void * r vector signed int vsumi2 = v0; vector signed int vsumi3 = v0; - const uint8_t * restrict q5 = x[i].qs; - const int8_t * restrict q8 = y[i].qs; + const uint8_t * GGML_RESTRICT q5 = x[i].qs; + const int8_t * GGML_RESTRICT q8 = y[i].qs; for (int j = 0; j < QK_K/64; ++j) { __builtin_prefetch(q5, 0, 1); @@ -6681,22 +8190,14 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, size_t bs, const void * r *s = vec_extract(vsumf0, 0); #elif defined __loongarch_asx - GGML_UNUSED(kmask1); - GGML_UNUSED(kmask2); - GGML_UNUSED(kmask3); - - const __m256i m4 = __lasx_xvreplgr2vr_b(0xF); - const __m128i mzero = __lsx_vldi(0); - const __m256i mone = __lasx_xvreplgr2vr_b(1); __m256 acc = (__m256)__lasx_xvldi(0); + __m128 acc_m = (__m128)__lsx_vldi(0); - float summs = 0.f; + for (int i = 0; i < nb; ++i) { - for (int i = 0; i < nb; ++i) { - - const uint8_t * restrict q5 = x[i].qs; - const int8_t * restrict q8 = y[i].qs; + const uint8_t * GGML_RESTRICT q5 = x[i].qs; + const int8_t * GGML_RESTRICT q8 = y[i].qs; const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin); @@ -6708,49 +8209,40 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, size_t bs, const void * r utmp[2] = uaux; utmp[0] &= kmask1; - const __m256i mins_and_scales = lasx_extu8_16(lsx_set_w(utmp[3], utmp[2], utmp[1], utmp[0])); + const __m128i mins_and_scales128 = lsx_set_w(utmp[3], utmp[2], utmp[1], utmp[0]); + const __m128i mins128 = __lsx_vexth_h_b(mins_and_scales128); + const __m128i scales128 = __lsx_vsllwil_h_b(mins_and_scales128, 0); const __m256i q8sums = __lasx_xvld((const __m256i*)y[i].bsums, 0); const __m128i q8s = lsx_hadd_h(lasx_extracti128(q8sums, 0), lasx_extracti128(q8sums, 1)); - const __m128i prod = lsx_madd_h(lasx_extracti128(mins_and_scales, 1), q8s); - const __m128i hsum = lsx_hadd_w(lsx_hadd_w(prod, mzero), mzero); - summs += dmin * __lsx_vpickve2gr_w(hsum, 0); //TODO check + const __m128i prod = lsx_madd_h(mins128, q8s); + acc_m = __lsx_vfmadd_s(__lsx_vreplfr2vr_s(dmin), __lsx_vffint_s_w(prod), acc_m); - const __m128i sc128 = lasx_extracti128(mins_and_scales, 0); - const __m256i scales = lasx_insertf128(sc128, sc128); + const __m256i scales = lasx_insertf128(scales128, scales128); const __m256i hbits = __lasx_xvld((const __m256i*)x[i].qh, 0); - __m256i hmask = mone; __m256i sumi = __lasx_xvldi(0); - int bit = 0; - __m256i xvbit; - for (int j = 0; j < QK_K/64; ++j) { - const __m256i scale_0 = lasx_shuffle_b(scales, get_scale_shuffle_k4(2*j+0)); - const __m256i scale_1 = lasx_shuffle_b(scales, get_scale_shuffle_k4(2*j+1)); + const __m256i scale_0 = lasx_xvrepl128vei_h(scales, 2 * j + 0); + const __m256i scale_1 = lasx_xvrepl128vei_h(scales, 2 * j + 1); const __m256i q5bits = __lasx_xvld((const __m256i*)q5, 0); q5 += 32; - xvbit = __lasx_xvreplgr2vr_h(bit++); - const __m256i q5l_0 = __lasx_xvand_v(q5bits, m4); - const __m256i q5h_0 = __lasx_xvslli_h(__lasx_xvsrl_h(__lasx_xvand_v(hbits, hmask), xvbit), 4); - const __m256i q5_0 = __lasx_xvadd_b(q5l_0, q5h_0); - hmask = __lasx_xvslli_h(hmask, 1); - - xvbit = __lasx_xvreplgr2vr_h(bit++); - const __m256i q5l_1 = __lasx_xvand_v(__lasx_xvsrli_h(q5bits, 4), m4); - const __m256i q5h_1 = __lasx_xvslli_h(__lasx_xvsrl_h(__lasx_xvand_v(hbits, hmask), xvbit), 4); - const __m256i q5_1 = __lasx_xvadd_b(q5l_1, q5h_1); - hmask = __lasx_xvslli_h(hmask, 1); + const __m256i q5l_0 = __lasx_xvandi_b(q5bits, 0xf); + const __m256i q5l_1 = __lasx_xvsrli_b(q5bits, 4); + const __m256i q5h_0 = __lasx_xvnori_b(__lasx_xvseqi_b(lasx_xvandi_b_bit(hbits, 2 * j + 0), 0), 0xef); + const __m256i q5h_1 = __lasx_xvnori_b(__lasx_xvseqi_b(lasx_xvandi_b_bit(hbits, 2 * j + 1), 0), 0xef); + const __m256i q5_0 = __lasx_xvor_v(q5l_0, q5h_0); + const __m256i q5_1 = __lasx_xvor_v(q5l_1, q5h_1); const __m256i q8_0 = __lasx_xvld((const __m256i*)q8, 0); q8 += 32; const __m256i q8_1 = __lasx_xvld((const __m256i*)q8, 0); q8 += 32; - __m256i p16_0 = lasx_maddubs_h(q5_0, q8_0); - __m256i p16_1 = lasx_maddubs_h(q5_1, q8_1); + __m256i p16_0 = lasx_madd_h_b(q5_0, q8_0); + __m256i p16_1 = lasx_madd_h_b(q5_1, q8_1); p16_0 = lasx_madd_h(scale_0, p16_0); p16_1 = lasx_madd_h(scale_1, p16_1); @@ -6764,8 +8256,98 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, size_t bs, const void * r } - *s = hsum_float_8(acc) + summs; + acc_m = __lsx_vfadd_s(acc_m, (__m128)__lsx_vbsrl_v(acc_m, 8)); + acc_m = __lsx_vfadd_s(acc_m, (__m128)__lsx_vbsrl_v(acc_m, 4)); + *s = hsum_float_8(acc) + ((v4f32)acc_m)[0]; +#elif defined(__VXE__) || defined(__VXE2__) + const uint8x16_t v_lm = vec_splat_u8(0x0F); + const uint8x16_t v_1m = vec_splat_u8(0x01); + const uint8x16_t v_2m = vec_splat_u8(0x02); + + const int32x4_t v_z = vec_splat_s32(0); + + const uchar8x16_t v_minsm = { + 0x08, 0x09, 0x0A, 0x0B, 0x0C, 0x0D, 0x0E, 0x0F, + 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF + }; + + int8x16_t q5b[4]; + uint8x16_t q5h[4]; + + uint8x16_t v_xl[2]; + uint8x16_t v_xh[2]; + int8x16_t v_y[4]; + + float sumf = 0; + + for (int i = 0; i < nb; ++i) { + const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); + const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin); + + const int16x8_t v_ysumsl = vec_xl(0 , y[i].bsums); + const int16x8_t v_ysumsh = vec_xl(16, y[i].bsums); + const int16x8_t v_ysums = vec_padd_s16(v_ysumsl, v_ysumsh); + + memcpy(utmp, x[i].scales, 12); + utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4); + const uint32_t uaux = utmp[1] & kmask1; + utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4); + utmp[2] = uaux; + utmp[0] &= kmask1; + + const uint8x16_t v_mins16 = vec_xl(0, (const uint8_t *)utmp); + const uint8x16_t v_mins8 = vec_perm(v_mins16, v_mins16, v_minsm); + const int16x8_t v_minsh = (int16x8_t)vec_unpackh(v_mins8); + + const int32x4_t v_minsho = vec_mulo(v_ysums, v_minsh); + const int32x4_t v_minshe = vec_mule(v_ysums, v_minsh); + const int32x4_t v_mins = vec_add(v_minsho, v_minshe); + const int32_t mins = v_mins[0] + v_mins[1] + v_mins[2] + v_mins[3]; + + const uint8_t * scales = (const uint8_t *)utmp; + const uint8_t * GGML_RESTRICT x0l = x[i].qs; + const uint8_t * GGML_RESTRICT x0h = x[i].qh; + const int8_t * GGML_RESTRICT y0 = y[i].qs; + + v_xh[0] = vec_xl(0 , x0h); + v_xh[1] = vec_xl(16, x0h); + + int32_t sumi = 0; + for (int j = 0; j < QK_K/64; ++j) { + v_xl[0] = vec_xl(0 , x0l); + v_xl[1] = vec_xl(16, x0l); + x0l += 32; + + v_y[0] = vec_xl(0 , y0); + v_y[1] = vec_xl(16, y0); + v_y[2] = vec_xl(32, y0); + v_y[3] = vec_xl(48, y0); + y0 += 64; + + q5h[0] = vec_sl(vec_and(v_1m, v_xh[0]), 4); + q5h[1] = vec_sl(vec_and(v_1m, v_xh[1]), 4); + q5h[2] = vec_sl(vec_and(v_2m, v_xh[0]), 3); + q5h[3] = vec_sl(vec_and(v_2m, v_xh[1]), 3); + v_xh[0] = vec_sr(v_xh[0], 2); + v_xh[1] = vec_sr(v_xh[1], 2); + + q5b[0] = (int8x16_t)vec_or(vec_and(v_xl[0], v_lm), q5h[0]); + q5b[1] = (int8x16_t)vec_or(vec_and(v_xl[1], v_lm), q5h[1]); + q5b[2] = (int8x16_t)vec_or(vec_sr(v_xl[0], 4), q5h[2]); + q5b[3] = (int8x16_t)vec_or(vec_sr(v_xl[1], 4), q5h[3]); + + int32x4_t sumi0 = ggml_vec_dot(ggml_vec_dot(v_z, q5b[0], v_y[0]), q5b[1], v_y[1]); + int32x4_t sumi1 = ggml_vec_dot(ggml_vec_dot(v_z, q5b[2], v_y[2]), q5b[3], v_y[3]); + + sumi += (sumi0[0] + sumi0[1] + sumi0[2] + sumi0[3]) * *scales++; + sumi += (sumi1[0] + sumi1[1] + sumi1[2] + sumi1[3]) * *scales++; + } + + sumf += d * sumi - dmin * mins; + } + + *s = sumf; #else const uint8_t * scales = (const uint8_t*)&utmp[0]; @@ -6779,11 +8361,11 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, size_t bs, const void * r float sumf = 0; for (int i = 0; i < nb; ++i) { - const uint8_t * restrict q4 = x[i].qs; - const uint8_t * restrict hm = x[i].qh; - const int8_t * restrict q8 = y[i].qs; + const uint8_t * GGML_RESTRICT q4 = x[i].qs; + const uint8_t * GGML_RESTRICT hm = x[i].qh; + const int8_t * GGML_RESTRICT q8 = y[i].qs; memset(aux32, 0, 8*sizeof(int32_t)); - int8_t * restrict a = aux8; + int8_t * GGML_RESTRICT a = aux8; uint8_t m = 1; for (int j = 0; j < QK_K/64; ++j) { for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] & 0xF); @@ -6830,7 +8412,7 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, size_t bs, const void * r #endif } -void ggml_vec_dot_q6_K_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) { +void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) { assert(n % QK_K == 0); assert(nrc == 1); UNUSED(nrc); @@ -6838,12 +8420,161 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * restrict s, size_t bs, const void * r UNUSED(by); UNUSED(bs); - const block_q6_K * restrict x = vx; - const block_q8_K * restrict y = vy; + const block_q6_K * GGML_RESTRICT x = vx; + const block_q8_K * GGML_RESTRICT y = vy; const int nb = n / QK_K; -#ifdef __ARM_NEON +#ifdef __ARM_FEATURE_SVE + const int vector_length = ggml_cpu_get_sve_cnt()*8; + float sum = 0; + svuint8_t m4b = svdup_n_u8(0xf); + svint32_t vzero = svdup_n_s32(0); + svuint8_t mone = svdup_n_u8(0x30); + svint8_t q6bytes_1, q6bytes_2, q6bytes_3, q6bytes_4; + svuint8_t q6h_1, q6h_2, q6h_3, q6h_4; + + for (int i = 0; i < nb; ++i) { + const float d_all = GGML_FP16_TO_FP32(x[i].d); + + const uint8_t * GGML_RESTRICT q6 = x[i].ql; + const uint8_t * GGML_RESTRICT qh = x[i].qh; + const int8_t * GGML_RESTRICT q8 = y[i].qs; + + const int8_t * GGML_RESTRICT scale = x[i].scales; + + const svbool_t pg16_8 = svptrue_pat_b16(SV_VL8); + const svint16_t q8sums_1 = svld1_s16(pg16_8, y[i].bsums); + const svint16_t q8sums_2 = svld1_s16(pg16_8, y[i].bsums + 8); + const svint16_t q6scales_1 = svunpklo_s16(svld1_s8(svptrue_pat_b8(SV_VL8), scale)); + const svint16_t q6scales_2 = svunpklo_s16(svld1_s8(svptrue_pat_b8(SV_VL8), scale + 8)); + const svint64_t prod = svdup_n_s64(0); + int32_t isum_mins = svaddv_s64(svptrue_b64(), svadd_s64_x(svptrue_b64(), svdot_s64(prod, q8sums_1, q6scales_1), + svdot_s64(prod, q8sums_2, q6scales_2))); + int32_t isum = 0; + + switch (vector_length) { + case 128: + { + const svbool_t pg32_4 = svptrue_pat_b32(SV_VL4); + const svbool_t pg8_16 = svptrue_pat_b8(SV_VL16); + svint32_t isum_tmp = svdup_n_s32(0); + for (int j = 0; j < QK_K/128; ++j) { + svuint8_t qhbits_1 = svld1_u8(pg8_16, qh); + svuint8_t qhbits_2 = svld1_u8(pg8_16, qh+16); + qh += 32; + svuint8_t q6bits_1 = svld1_u8(pg8_16, q6); + svuint8_t q6bits_2 = svld1_u8(pg8_16, q6+16); + svuint8_t q6bits_3 = svld1_u8(pg8_16, q6+32); + svuint8_t q6bits_4 = svld1_u8(pg8_16, q6+48); + q6 += 64; + svint8_t q8bytes_1 = svld1_s8(pg8_16, q8); + svint8_t q8bytes_2 = svld1_s8(pg8_16, q8+16); + svint8_t q8bytes_3 = svld1_s8(pg8_16, q8+32); + svint8_t q8bytes_4 = svld1_s8(pg8_16, q8+48); + q8 += 64; + + q6h_1 = svand_u8_x(pg16_8, mone, svlsl_n_u8_x(pg16_8, qhbits_1, 4)); + q6h_2 = svand_u8_x(pg16_8, mone, svlsl_n_u8_x(pg16_8, qhbits_2, 4)); + q6h_3 = svand_u8_x(pg16_8, mone, svlsl_n_u8_x(pg16_8, qhbits_1, 2)); + q6h_4 = svand_u8_x(pg16_8, mone, svlsl_n_u8_x(pg16_8, qhbits_2, 2)); + q6bytes_1 = svreinterpret_s8_u8(svorr_u8_x(pg8_16, svand_u8_x(pg8_16, q6bits_1, m4b), q6h_1)); + q6bytes_2 = svreinterpret_s8_u8(svorr_u8_x(pg8_16, svand_u8_x(pg8_16, q6bits_2, m4b), q6h_2)); + q6bytes_3 = svreinterpret_s8_u8(svorr_u8_x(pg8_16, svand_u8_x(pg8_16, q6bits_3, m4b), q6h_3)); + q6bytes_4 = svreinterpret_s8_u8(svorr_u8_x(pg8_16, svand_u8_x(pg8_16, q6bits_4, m4b), q6h_4)); + isum_tmp = svmla_n_s32_x(pg32_4, isum_tmp, svdot_s32(vzero, q6bytes_1, q8bytes_1), scale[0]); + isum_tmp = svmla_n_s32_x(pg32_4, isum_tmp, svdot_s32(vzero, q6bytes_2, q8bytes_2), scale[1]); + isum_tmp = svmla_n_s32_x(pg32_4, isum_tmp, svdot_s32(vzero, q6bytes_3, q8bytes_3), scale[2]); + isum_tmp = svmla_n_s32_x(pg32_4, isum_tmp, svdot_s32(vzero, q6bytes_4, q8bytes_4), scale[3]); + + scale += 4; + q8bytes_1 = svld1_s8(pg8_16, q8); + q8bytes_2 = svld1_s8(pg8_16, q8+16); + q8bytes_3 = svld1_s8(pg8_16, q8+32); + q8bytes_4 = svld1_s8(pg8_16, q8+48); + q8 += 64; + + q6h_1 = svand_u8_x(pg16_8, mone, qhbits_1); + q6h_2 = svand_u8_x(pg16_8, mone, qhbits_2); + q6h_3 = svand_u8_x(pg16_8, mone, svlsr_n_u8_x(pg16_8, qhbits_1, 2)); + q6h_4 = svand_u8_x(pg16_8, mone, svlsr_n_u8_x(pg16_8, qhbits_2, 2)); + q6bytes_1 = svreinterpret_s8_u8(svorr_u8_x(pg8_16, svlsr_n_u8_x(pg8_16, q6bits_1, 4), q6h_1)); + q6bytes_2 = svreinterpret_s8_u8(svorr_u8_x(pg8_16, svlsr_n_u8_x(pg8_16, q6bits_2, 4), q6h_2)); + q6bytes_3 = svreinterpret_s8_u8(svorr_u8_x(pg8_16, svlsr_n_u8_x(pg8_16, q6bits_3, 4), q6h_3)); + q6bytes_4 = svreinterpret_s8_u8(svorr_u8_x(pg8_16, svlsr_n_u8_x(pg8_16, q6bits_4, 4), q6h_4)); + isum_tmp = svmla_n_s32_x(pg32_4, isum_tmp, svdot_s32(vzero, q6bytes_1, q8bytes_1), scale[0]); + isum_tmp = svmla_n_s32_x(pg32_4, isum_tmp, svdot_s32(vzero, q6bytes_2, q8bytes_2), scale[1]); + isum_tmp = svmla_n_s32_x(pg32_4, isum_tmp, svdot_s32(vzero, q6bytes_3, q8bytes_3), scale[2]); + isum_tmp = svmla_n_s32_x(pg32_4, isum_tmp, svdot_s32(vzero, q6bytes_4, q8bytes_4), scale[3]); + scale += 4; + } + isum += svaddv_s32(pg32_4, isum_tmp); + sum += d_all * y[i].d * (isum - 32 * isum_mins); + } + break; + case 256: + case 512: + { + const svbool_t pg8_2 = svptrue_pat_b8(SV_VL2); + const svbool_t pg32_8 = svptrue_pat_b32(SV_VL8); + const svbool_t pg8_32 = svptrue_pat_b8(SV_VL32); + svint32_t isum_tmp = svdup_n_s32(0); + for (int j = 0; j < QK_K/128; j++) { + svuint8_t qhbits_1 = svld1_u8(pg8_32, qh); + qh += 32; + svuint8_t q6bits_1 = svld1_u8(pg8_32, q6); + svuint8_t q6bits_2 = svld1_u8(pg8_32, q6+32); + q6 += 64; + svint8_t q8bytes_1 = svld1_s8(pg8_32, q8); + svint8_t q8bytes_2 = svld1_s8(pg8_32, q8+32); + svint8_t q8bytes_3 = svld1_s8(pg8_32, q8+64); + svint8_t q8bytes_4 = svld1_s8(pg8_32, q8+96); + q8 += 128; + q6h_1 = svand_u8_x(pg8_32, mone, svlsl_n_u8_x(pg8_32, qhbits_1, 4)); + q6h_2 = svand_u8_x(pg8_32, mone, svlsl_n_u8_x(pg8_32, qhbits_1, 2)); + q6h_3 = svand_u8_x(pg8_32, mone, qhbits_1); + q6h_4 = svand_u8_x(pg8_32, mone, svlsr_n_u8_x(pg8_32, qhbits_1, 2)); + q6bytes_1 = svreinterpret_s8_u8(svorr_u8_x(pg8_32, svand_u8_x(pg8_32, q6bits_1, m4b), q6h_1)); + q6bytes_2 = svreinterpret_s8_u8(svorr_u8_x(pg8_32, svand_u8_x(pg8_32, q6bits_2, m4b), q6h_2)); + q6bytes_3 = svreinterpret_s8_u8(svorr_u8_x(pg8_32, svlsr_n_u8_x(pg8_32, q6bits_1, 4), q6h_3)); + q6bytes_4 = svreinterpret_s8_u8(svorr_u8_x(pg8_32, svlsr_n_u8_x(pg8_32, q6bits_2, 4), q6h_4)); + + svint8_t scale_lane_1_tmp = svld1_s8(pg8_2, scale); + scale_lane_1_tmp= svzip1_s8(scale_lane_1_tmp, scale_lane_1_tmp); + scale_lane_1_tmp= svzip1_s8(scale_lane_1_tmp, scale_lane_1_tmp); + svint8_t scale_lane_2_tmp = svld1_s8(pg8_2, scale+2); + scale_lane_2_tmp = svzip1_s8(scale_lane_2_tmp, scale_lane_2_tmp); + scale_lane_2_tmp = svzip1_s8(scale_lane_2_tmp, scale_lane_2_tmp); + svint8_t scale_lane_3_tmp = svld1_s8(pg8_2, scale+4); + scale_lane_3_tmp = svzip1_s8(scale_lane_3_tmp, scale_lane_3_tmp); + scale_lane_3_tmp = svzip1_s8(scale_lane_3_tmp, scale_lane_3_tmp); + svint8_t scale_lane_4_tmp = svld1_s8(pg8_2, scale+6); + scale_lane_4_tmp = svzip1_s8(scale_lane_4_tmp, scale_lane_4_tmp); + scale_lane_4_tmp = svzip1_s8(scale_lane_4_tmp, scale_lane_4_tmp); + svint32_t scale_lane_1 = svunpklo_s32(svunpklo_s16(scale_lane_1_tmp)); + svint32_t scale_lane_2 = svunpklo_s32(svunpklo_s16(scale_lane_2_tmp)); + svint32_t scale_lane_3 = svunpklo_s32(svunpklo_s16(scale_lane_3_tmp)); + svint32_t scale_lane_4 = svunpklo_s32(svunpklo_s16(scale_lane_4_tmp)); + + isum_tmp = svmla_s32_x(pg32_8, isum_tmp, svdot_s32(vzero, q6bytes_1, q8bytes_1), scale_lane_1); + isum_tmp = svmla_s32_x(pg32_8, isum_tmp, svdot_s32(vzero, q6bytes_2, q8bytes_2), scale_lane_2); + isum_tmp = svmla_s32_x(pg32_8, isum_tmp, svdot_s32(vzero, q6bytes_3, q8bytes_3), scale_lane_3); + isum_tmp = svmla_s32_x(pg32_8, isum_tmp, svdot_s32(vzero, q6bytes_4, q8bytes_4), scale_lane_4); + scale += 8; + } + isum += svaddv_s32(pg32_8, isum_tmp); + sum += d_all * y[i].d * (isum - 32 * isum_mins); + } + break; + default: + assert(false && "Unsupported vector length"); + break; + } + } + + *s = sum; + +#elif __ARM_NEON float sum = 0; const uint8x16_t m4b = vdupq_n_u8(0xF); @@ -6859,11 +8590,11 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * restrict s, size_t bs, const void * r const float d_all = GGML_FP16_TO_FP32(x[i].d); - const uint8_t * restrict q6 = x[i].ql; - const uint8_t * restrict qh = x[i].qh; - const int8_t * restrict q8 = y[i].qs; + const uint8_t * GGML_RESTRICT q6 = x[i].ql; + const uint8_t * GGML_RESTRICT qh = x[i].qh; + const int8_t * GGML_RESTRICT q8 = y[i].qs; - const int8_t * restrict scale = x[i].scales; + const int8_t * GGML_RESTRICT scale = x[i].scales; const ggml_int16x8x2_t q8sums = ggml_vld1q_s16_x2(y[i].bsums); const int8x16_t scales = vld1q_s8(scale); @@ -6950,9 +8681,9 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * restrict s, size_t bs, const void * r const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); - const uint8_t * restrict q4 = x[i].ql; - const uint8_t * restrict qh = x[i].qh; - const int8_t * restrict q8 = y[i].qs; + const uint8_t * GGML_RESTRICT q4 = x[i].ql; + const uint8_t * GGML_RESTRICT qh = x[i].qh; + const int8_t * GGML_RESTRICT q8 = y[i].qs; const __m128i scales = _mm_loadu_si128((const __m128i*)x[i].scales); @@ -7028,9 +8759,9 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * restrict s, size_t bs, const void * r const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); - const uint8_t * restrict q4 = x[i].ql; - const uint8_t * restrict qh = x[i].qh; - const int8_t * restrict q8 = y[i].qs; + const uint8_t * GGML_RESTRICT q4 = x[i].ql; + const uint8_t * GGML_RESTRICT qh = x[i].qh; + const int8_t * GGML_RESTRICT q8 = y[i].qs; // handle the q6_k -32 offset separately using bsums const __m128i q8sums_0 = _mm_loadu_si128((const __m128i*)y[i].bsums); @@ -7122,87 +8853,249 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * restrict s, size_t bs, const void * r *s = hsum_float_8(acc); -#elif defined __riscv_v_intrinsic +#elif defined __wasm_simd128__ + int8_t aux8[QK_K] __attribute__((aligned(16))); + int32_t aux32[8] __attribute__((aligned(16))) = {0}; + float sums[8] __attribute__((aligned(16))) = {0}; - float sumf = 0; for (int i = 0; i < nb; ++i) { - - const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; - - const uint8_t * restrict q6 = x[i].ql; - const uint8_t * restrict qh = x[i].qh; - const int8_t * restrict q8 = y[i].qs; - - const int8_t * restrict scale = x[i].scales; - - size_t vl; - - vint32m1_t vzero = __riscv_vmv_v_x_i32m1(0, 1); - - int sum_t = 0; - int is = 0; - - for (int j = 0; j < QK_K/128; ++j) { - - vl = 32; - - // load qh - vuint8m1_t qh_x = __riscv_vle8_v_u8m1(qh, vl); - - // load Q6 - vuint8m1_t q6_0 = __riscv_vle8_v_u8m1(q6, vl); - vuint8m1_t q6_1 = __riscv_vle8_v_u8m1(q6+32, vl); - - vuint8m1_t q6a_0 = __riscv_vand_vx_u8m1(q6_0, 0x0F, vl); - vuint8m1_t q6a_1 = __riscv_vand_vx_u8m1(q6_1, 0x0F, vl); - vuint8m1_t q6s_0 = __riscv_vsrl_vx_u8m1(q6_0, 0x04, vl); - vuint8m1_t q6s_1 = __riscv_vsrl_vx_u8m1(q6_1, 0x04, vl); - - vuint8m1_t qh_0 = __riscv_vand_vx_u8m1(qh_x, 0x03, vl); - vuint8m1_t qh_1 = __riscv_vand_vx_u8m1(__riscv_vsrl_vx_u8m1(qh_x, 0x2, vl), 0x03 , vl); - vuint8m1_t qh_2 = __riscv_vand_vx_u8m1(__riscv_vsrl_vx_u8m1(qh_x, 0x4, vl), 0x03 , vl); - vuint8m1_t qh_3 = __riscv_vand_vx_u8m1(__riscv_vsrl_vx_u8m1(qh_x, 0x6, vl), 0x03 , vl); - - vuint8m1_t qhi_0 = __riscv_vor_vv_u8m1(q6a_0, __riscv_vsll_vx_u8m1(qh_0, 0x04, vl), vl); - vuint8m1_t qhi_1 = __riscv_vor_vv_u8m1(q6a_1, __riscv_vsll_vx_u8m1(qh_1, 0x04, vl), vl); - vuint8m1_t qhi_2 = __riscv_vor_vv_u8m1(q6s_0, __riscv_vsll_vx_u8m1(qh_2, 0x04, vl), vl); - vuint8m1_t qhi_3 = __riscv_vor_vv_u8m1(q6s_1, __riscv_vsll_vx_u8m1(qh_3, 0x04, vl), vl); - - vint8m1_t a_0 = __riscv_vsub_vx_i8m1(__riscv_vreinterpret_v_u8m1_i8m1(qhi_0), 32, vl); - vint8m1_t a_1 = __riscv_vsub_vx_i8m1(__riscv_vreinterpret_v_u8m1_i8m1(qhi_1), 32, vl); - vint8m1_t a_2 = __riscv_vsub_vx_i8m1(__riscv_vreinterpret_v_u8m1_i8m1(qhi_2), 32, vl); - vint8m1_t a_3 = __riscv_vsub_vx_i8m1(__riscv_vreinterpret_v_u8m1_i8m1(qhi_3), 32, vl); - - // load Q8 and take product - vint16m2_t va_q_0 = __riscv_vwmul_vv_i16m2(a_0, __riscv_vle8_v_i8m1(q8, vl), vl); - vint16m2_t va_q_1 = __riscv_vwmul_vv_i16m2(a_1, __riscv_vle8_v_i8m1(q8+32, vl), vl); - vint16m2_t va_q_2 = __riscv_vwmul_vv_i16m2(a_2, __riscv_vle8_v_i8m1(q8+64, vl), vl); - vint16m2_t va_q_3 = __riscv_vwmul_vv_i16m2(a_3, __riscv_vle8_v_i8m1(q8+96, vl), vl); - - vl = 16; - - vint32m2_t vaux_0 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(va_q_0, 0), scale[is+0], vl); - vint32m2_t vaux_1 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(va_q_0, 1), scale[is+1], vl); - vint32m2_t vaux_2 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(va_q_1, 0), scale[is+2], vl); - vint32m2_t vaux_3 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(va_q_1, 1), scale[is+3], vl); - vint32m2_t vaux_4 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(va_q_2, 0), scale[is+4], vl); - vint32m2_t vaux_5 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(va_q_2, 1), scale[is+5], vl); - vint32m2_t vaux_6 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(va_q_3, 0), scale[is+6], vl); - vint32m2_t vaux_7 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(va_q_3, 1), scale[is+7], vl); - - vint32m1_t isum0 = __riscv_vredsum_vs_i32m2_i32m1(__riscv_vadd_vv_i32m2(vaux_0, vaux_1, vl), vzero, vl); - vint32m1_t isum1 = __riscv_vredsum_vs_i32m2_i32m1(__riscv_vadd_vv_i32m2(vaux_2, vaux_3, vl), isum0, vl); - vint32m1_t isum2 = __riscv_vredsum_vs_i32m2_i32m1(__riscv_vadd_vv_i32m2(vaux_4, vaux_5, vl), isum1, vl); - vint32m1_t isum3 = __riscv_vredsum_vs_i32m2_i32m1(__riscv_vadd_vv_i32m2(vaux_6, vaux_7, vl), isum2, vl); - - sum_t += __riscv_vmv_x_s_i32m1_i32(isum3); - - q6 += 64; qh += 32; q8 += 128; is=8; - + // Unpack 6-bit quantized data into aux8 (unchanged) + const uint8_t * GGML_RESTRICT q4 = x[i].ql; + const uint8_t * GGML_RESTRICT qh = x[i].qh; + int8_t * a = aux8; + for (int j = 0; j < QK_K; j += 128) { + for (int l = 0; l < 32; ++l) { + a[l + 0] = (int8_t)((q4[l + 0] & 0xF) | (((qh[l] >> 0) & 3) << 4)) - 32; + a[l + 32] = (int8_t)((q4[l + 32] & 0xF) | (((qh[l] >> 2) & 3) << 4)) - 32; + a[l + 64] = (int8_t)((q4[l + 0] >> 4) | (((qh[l] >> 4) & 3) << 4)) - 32; + a[l + 96] = (int8_t)((q4[l + 32] >> 4) | (((qh[l] >> 6) & 3) << 4)) - 32; + } + a += 128; + q4 += 64; + qh += 32; } - sumf += d * sum_t; + const int8_t * GGML_RESTRICT a_ptr = aux8; + const int8_t * GGML_RESTRICT q8 = y[i].qs; + v128_t acc0 = wasm_i32x4_splat(0); + v128_t acc1 = wasm_i32x4_splat(0); + for (int j = 0; j < QK_K/16; ++j) { + const int scale = x[i].scales[j]; + const v128_t vscale = wasm_i32x4_splat(scale); + + // Load 16 elements from a and q8 + const v128_t a_vec = wasm_v128_load(a_ptr); + const v128_t q8_vec = wasm_v128_load(q8); + + // Process low 8 elements + v128_t a_low = wasm_i16x8_extend_low_i8x16(a_vec); + v128_t q8_low = wasm_i16x8_extend_low_i8x16(q8_vec); + v128_t prod_low = wasm_i16x8_mul(a_low, q8_low); + v128_t prod_lo_lo = wasm_i32x4_extend_low_i16x8(prod_low); + v128_t prod_lo_hi = wasm_i32x4_extend_high_i16x8(prod_low); + + // Process high 8 elements + v128_t a_high = wasm_i16x8_extend_high_i8x16(a_vec); + v128_t q8_high = wasm_i16x8_extend_high_i8x16(q8_vec); + v128_t prod_high = wasm_i16x8_mul(a_high, q8_high); + v128_t prod_hi_lo = wasm_i32x4_extend_low_i16x8(prod_high); + v128_t prod_hi_hi = wasm_i32x4_extend_high_i16x8(prod_high); + + // Scale and accumulate + prod_lo_lo = wasm_i32x4_mul(prod_lo_lo, vscale); + prod_lo_hi = wasm_i32x4_mul(prod_lo_hi, vscale); + prod_hi_lo = wasm_i32x4_mul(prod_hi_lo, vscale); + prod_hi_hi = wasm_i32x4_mul(prod_hi_hi, vscale); + + acc0 = wasm_i32x4_add(acc0, wasm_i32x4_add(prod_lo_lo, prod_hi_lo)); + acc1 = wasm_i32x4_add(acc1, wasm_i32x4_add(prod_lo_hi, prod_hi_hi)); + + a_ptr += 16; + q8 += 16; + } + + // Store accumulated results + wasm_v128_store(&aux32[0], acc0); + wasm_v128_store(&aux32[4], acc1); + + const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + for (int l = 0; l < 8; ++l) { + sums[l] += d * aux32[l]; + } + } + + // Sum final results + float sumf = 0; + for (int l = 0; l < 8; ++l) { + sumf += sums[l]; + } + *s = sumf; + +#elif defined __riscv_v_intrinsic + + const int vector_length = __riscv_vlenb() * 8; + float sumf = 0; + + switch (vector_length) { + case 256: + for (int i = 0; i < nb; ++i) { + + const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + + const uint8_t * GGML_RESTRICT q6 = x[i].ql; + const uint8_t * GGML_RESTRICT qh = x[i].qh; + const int8_t * GGML_RESTRICT q8 = y[i].qs; + + const int8_t * GGML_RESTRICT scale = x[i].scales; + + size_t vl; + + vint32m1_t vzero = __riscv_vmv_v_x_i32m1(0, 1); + + int sum_t = 0; + int is = 0; + + for (int j = 0; j < QK_K/128; ++j) { + + vl = 32; + + // load qh + vuint8m1_t qh_x = __riscv_vle8_v_u8m1(qh, vl); + + // load Q6 + vuint8m1_t q6_0 = __riscv_vle8_v_u8m1(q6, vl); + vuint8m1_t q6_1 = __riscv_vle8_v_u8m1(q6+32, vl); + + vuint8m1_t q6a_0 = __riscv_vand_vx_u8m1(q6_0, 0x0F, vl); + vuint8m1_t q6a_1 = __riscv_vand_vx_u8m1(q6_1, 0x0F, vl); + vuint8m1_t q6s_0 = __riscv_vsrl_vx_u8m1(q6_0, 0x04, vl); + vuint8m1_t q6s_1 = __riscv_vsrl_vx_u8m1(q6_1, 0x04, vl); + + vuint8m1_t qh_0 = __riscv_vand_vx_u8m1(qh_x, 0x03, vl); + vuint8m1_t qh_1 = __riscv_vand_vx_u8m1(__riscv_vsrl_vx_u8m1(qh_x, 0x2, vl), 0x03 , vl); + vuint8m1_t qh_2 = __riscv_vand_vx_u8m1(__riscv_vsrl_vx_u8m1(qh_x, 0x4, vl), 0x03 , vl); + vuint8m1_t qh_3 = __riscv_vand_vx_u8m1(__riscv_vsrl_vx_u8m1(qh_x, 0x6, vl), 0x03 , vl); + + vuint8m1_t qhi_0 = __riscv_vor_vv_u8m1(q6a_0, __riscv_vsll_vx_u8m1(qh_0, 0x04, vl), vl); + vuint8m1_t qhi_1 = __riscv_vor_vv_u8m1(q6a_1, __riscv_vsll_vx_u8m1(qh_1, 0x04, vl), vl); + vuint8m1_t qhi_2 = __riscv_vor_vv_u8m1(q6s_0, __riscv_vsll_vx_u8m1(qh_2, 0x04, vl), vl); + vuint8m1_t qhi_3 = __riscv_vor_vv_u8m1(q6s_1, __riscv_vsll_vx_u8m1(qh_3, 0x04, vl), vl); + + vint8m1_t a_0 = __riscv_vsub_vx_i8m1(__riscv_vreinterpret_v_u8m1_i8m1(qhi_0), 32, vl); + vint8m1_t a_1 = __riscv_vsub_vx_i8m1(__riscv_vreinterpret_v_u8m1_i8m1(qhi_1), 32, vl); + vint8m1_t a_2 = __riscv_vsub_vx_i8m1(__riscv_vreinterpret_v_u8m1_i8m1(qhi_2), 32, vl); + vint8m1_t a_3 = __riscv_vsub_vx_i8m1(__riscv_vreinterpret_v_u8m1_i8m1(qhi_3), 32, vl); + + // load Q8 and take product + vint16m2_t va_q_0 = __riscv_vwmul_vv_i16m2(a_0, __riscv_vle8_v_i8m1(q8, vl), vl); + vint16m2_t va_q_1 = __riscv_vwmul_vv_i16m2(a_1, __riscv_vle8_v_i8m1(q8+32, vl), vl); + vint16m2_t va_q_2 = __riscv_vwmul_vv_i16m2(a_2, __riscv_vle8_v_i8m1(q8+64, vl), vl); + vint16m2_t va_q_3 = __riscv_vwmul_vv_i16m2(a_3, __riscv_vle8_v_i8m1(q8+96, vl), vl); + + vl = 16; + + vint32m2_t vaux_0 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(va_q_0, 0), scale[is+0], vl); + vint32m2_t vaux_1 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(va_q_0, 1), scale[is+1], vl); + vint32m2_t vaux_2 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(va_q_1, 0), scale[is+2], vl); + vint32m2_t vaux_3 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(va_q_1, 1), scale[is+3], vl); + vint32m2_t vaux_4 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(va_q_2, 0), scale[is+4], vl); + vint32m2_t vaux_5 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(va_q_2, 1), scale[is+5], vl); + vint32m2_t vaux_6 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(va_q_3, 0), scale[is+6], vl); + vint32m2_t vaux_7 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(va_q_3, 1), scale[is+7], vl); + + vint32m1_t isum0 = __riscv_vredsum_vs_i32m2_i32m1(__riscv_vadd_vv_i32m2(vaux_0, vaux_1, vl), vzero, vl); + vint32m1_t isum1 = __riscv_vredsum_vs_i32m2_i32m1(__riscv_vadd_vv_i32m2(vaux_2, vaux_3, vl), isum0, vl); + vint32m1_t isum2 = __riscv_vredsum_vs_i32m2_i32m1(__riscv_vadd_vv_i32m2(vaux_4, vaux_5, vl), isum1, vl); + vint32m1_t isum3 = __riscv_vredsum_vs_i32m2_i32m1(__riscv_vadd_vv_i32m2(vaux_6, vaux_7, vl), isum2, vl); + + sum_t += __riscv_vmv_x_s_i32m1_i32(isum3); + + q6 += 64; qh += 32; q8 += 128; is=8; + + } + + sumf += d * sum_t; + + } + break; + case 128: + for (int i = 0; i < nb; ++i) { + + const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; + + const uint8_t * restrict q6 = x[i].ql; + const uint8_t * restrict qh = x[i].qh; + const int8_t * restrict q8 = y[i].qs; + + const int8_t * restrict scale = x[i].scales; + + int sum_t = 0; + int t0; + + for (int j = 0; j < QK_K/128; ++j) { + __asm__ __volatile__( + "vsetvli zero, %[vl32], e8, m2\n\t" + "vle8.v v4, (%[qh])\n\t" + "vsll.vi v0, v4, 4\n\t" + "vsll.vi v2, v4, 2\n\t" + "vsrl.vi v6, v4, 2\n\t" + "vsetvli zero, %[vl64], e8, m4\n\t" + "vle8.v v8, (%[q6])\n\t" + "vsrl.vi v12, v8, 4\n\t" + "vand.vi v8, v8, 0xF\n\t" + "vsetvli zero, %[vl128], e8, m8\n\t" + "vand.vx v0, v0, %[mask]\n\t" + "vor.vv v8, v8, v0\n\t" + "vle8.v v0, (%[q8])\n\t" + "vsub.vx v8, v8, %[vl32]\n\t" + "vsetvli zero, %[vl64], e8, m4\n\t" + "vwmul.vv v16, v0, v8\n\t" + "vwmul.vv v24, v4, v12\n\t" + "vsetivli zero, 16, e16, m2\n\t" + "vmv.v.x v0, zero\n\t" + "vwredsum.vs v10, v16, v0\n\t" + "vwredsum.vs v9, v18, v0\n\t" + "vwredsum.vs v8, v20, v0\n\t" + "vwredsum.vs v7, v22, v0\n\t" + "vwredsum.vs v11, v24, v0\n\t" + "vwredsum.vs v12, v26, v0\n\t" + "vwredsum.vs v13, v28, v0\n\t" + "vwredsum.vs v14, v30, v0\n\t" + "vsetivli zero, 4, e32, m1\n\t" + "vslideup.vi v10, v9, 1\n\t" + "vslideup.vi v8, v7, 1\n\t" + "vslideup.vi v11, v12, 1\n\t" + "vslideup.vi v13, v14, 1\n\t" + "vslideup.vi v10, v8, 2\n\t" + "vslideup.vi v11, v13, 2\n\t" + "vsetivli zero, 8, e32, m2\n\t" + "vle8.v v2, (%[scale])\n\t" + "vsext.vf4 v4, v2\n\t" + "vmul.vv v2, v4, v10\n\t" + "vredsum.vs v0, v2, v0\n\t" + "vmv.x.s %[t0], v0\n\t" + "add %[sumi], %[sumi], %[t0]" + : [sumi] "+&r" (sum_t), [t0] "=&r" (t0) + : [qh] "r" (qh), [q6] "r" (q6), [q8] "r" (q8), [scale] "r" (scale) + , [vl32] "r" (32), [vl64] "r" (64), [vl128] "r" (128) + , [mask] "r" (0x30) + : "memory" + , "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7" + , "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15" + , "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23" + , "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31" + ); + q6 += 64; qh += 32; q8 += 128; scale += 8; + } + + sumf += d * sum_t; + + } + break; + default: + assert(false && "Unsupported vector length"); + break; } *s = sumf; @@ -7235,10 +9128,10 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * restrict s, size_t bs, const void * r vector signed int vsumi6 = v0; vector signed int vsumi7 = v0; - const uint8_t * restrict q6 = x[i].ql; - const uint8_t * restrict qh = x[i].qh; - const int8_t * restrict qs = x[i].scales; - const int8_t * restrict q8 = y[i].qs; + const uint8_t * GGML_RESTRICT q6 = x[i].ql; + const uint8_t * GGML_RESTRICT qh = x[i].qh; + const int8_t * GGML_RESTRICT qs = x[i].scales; + const int8_t * GGML_RESTRICT q8 = y[i].qs; for (int j = 0; j < QK_K/128; ++j) { __builtin_prefetch(q6, 0, 0); @@ -7346,8 +9239,6 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * restrict s, size_t bs, const void * r #elif defined __loongarch_asx - const __m256i m4 = __lasx_xvreplgr2vr_b(0xF); - const __m256i m2 = __lasx_xvreplgr2vr_b(3); const __m256i m32s = __lasx_xvreplgr2vr_b(32); __m256 acc = (__m256)__lasx_xvldi(0); @@ -7356,62 +9247,46 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * restrict s, size_t bs, const void * r const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d); - const uint8_t * restrict q4 = x[i].ql; - const uint8_t * restrict qh = x[i].qh; - const int8_t * restrict q8 = y[i].qs; + const uint8_t * GGML_RESTRICT q4 = x[i].ql; + const uint8_t * GGML_RESTRICT qh = x[i].qh; + const int8_t * GGML_RESTRICT q8 = y[i].qs; - const __m128i scales = __lsx_vld((const __m128i*)x[i].scales, 0); + const __m128i scales128 = __lsx_vld((const __m128i*)x[i].scales, 0); + const v16i8 shuffle_mask = {0, 2, 4, 6, 8, 10, 12, 14, 1, 3, 5, 7, 9, 11, 13, 15}; + const __m256i scales_shuffled = lasx_ext8_16(__lsx_vshuf_b(scales128, scales128, (__m128i)shuffle_mask)); __m256i sumi = __lasx_xvldi(0); - int is = 0; - for (int j = 0; j < QK_K/128; ++j) { - const __m128i scale_0 = lsx_shuffle_b(scales, get_scale_shuffle(is + 0)); - const __m128i scale_1 = lsx_shuffle_b(scales, get_scale_shuffle(is + 1)); - const __m128i scale_2 = lsx_shuffle_b(scales, get_scale_shuffle(is + 2)); - const __m128i scale_3 = lsx_shuffle_b(scales, get_scale_shuffle(is + 3)); - is += 4; - const __m256i q4bits1 = __lasx_xvld((const __m256i*)q4, 0); q4 += 32; const __m256i q4bits2 = __lasx_xvld((const __m256i*)q4, 0); q4 += 32; const __m256i q4bitsH = __lasx_xvld((const __m256i*)qh, 0); qh += 32; - const __m256i q4h_0 = __lasx_xvslli_h(__lasx_xvand_v(q4bitsH, m2), 4); - const __m256i q4h_1 = __lasx_xvslli_h(__lasx_xvand_v(__lasx_xvsrli_h(q4bitsH, 2), m2), 4); - const __m256i q4h_2 = __lasx_xvslli_h(__lasx_xvand_v(__lasx_xvsrli_h(q4bitsH, 4), m2), 4); - const __m256i q4h_3 = __lasx_xvslli_h(__lasx_xvand_v(__lasx_xvsrli_h(q4bitsH, 6), m2), 4); + const __m256i q4h_0 = __lasx_xvslli_b(__lasx_xvandi_b(q4bitsH, 3), 4); + const __m256i q4h_1 = __lasx_xvslli_b(__lasx_xvandi_b(q4bitsH, 3 << 2), 2); + const __m256i q4h_2 = __lasx_xvandi_b(q4bitsH, 3 << 4); + const __m256i q4h_3 = __lasx_xvsrli_b(__lasx_xvandi_b(q4bitsH, 3 << 6), 2); - const __m256i q4_0 = __lasx_xvor_v(__lasx_xvand_v(q4bits1, m4), q4h_0); - const __m256i q4_1 = __lasx_xvor_v(__lasx_xvand_v(q4bits2, m4), q4h_1); - const __m256i q4_2 = __lasx_xvor_v(__lasx_xvand_v(__lasx_xvsrli_h(q4bits1, 4), m4), q4h_2); - const __m256i q4_3 = __lasx_xvor_v(__lasx_xvand_v(__lasx_xvsrli_h(q4bits2, 4), m4), q4h_3); + const __m256i q4_0 = __lasx_xvor_v(__lasx_xvandi_b(q4bits1, 0xf), q4h_0); + const __m256i q4_1 = __lasx_xvor_v(__lasx_xvandi_b(q4bits2, 0xf), q4h_1); + const __m256i q4_2 = __lasx_xvor_v(__lasx_xvsrli_b(q4bits1, 4), q4h_2); + const __m256i q4_3 = __lasx_xvor_v(__lasx_xvsrli_b(q4bits2, 4), q4h_3); const __m256i q8_0 = __lasx_xvld((const __m256i*)q8, 0); q8 += 32; const __m256i q8_1 = __lasx_xvld((const __m256i*)q8, 0); q8 += 32; const __m256i q8_2 = __lasx_xvld((const __m256i*)q8, 0); q8 += 32; const __m256i q8_3 = __lasx_xvld((const __m256i*)q8, 0); q8 += 32; - __m256i q8s_0 = lasx_maddubs_h(m32s, q8_0); - __m256i q8s_1 = lasx_maddubs_h(m32s, q8_1); - __m256i q8s_2 = lasx_maddubs_h(m32s, q8_2); - __m256i q8s_3 = lasx_maddubs_h(m32s, q8_3); + __m256i p16_0 = lasx_madd_h_b(__lasx_xvsub_b(q4_0, m32s), q8_0); + __m256i p16_1 = lasx_madd_h_b(__lasx_xvsub_b(q4_1, m32s), q8_1); + __m256i p16_2 = lasx_madd_h_b(__lasx_xvsub_b(q4_2, m32s), q8_2); + __m256i p16_3 = lasx_madd_h_b(__lasx_xvsub_b(q4_3, m32s), q8_3); - __m256i p16_0 = lasx_maddubs_h(q4_0, q8_0); - __m256i p16_1 = lasx_maddubs_h(q4_1, q8_1); - __m256i p16_2 = lasx_maddubs_h(q4_2, q8_2); - __m256i p16_3 = lasx_maddubs_h(q4_3, q8_3); - - p16_0 = __lasx_xvsub_h(p16_0, q8s_0); - p16_1 = __lasx_xvsub_h(p16_1, q8s_1); - p16_2 = __lasx_xvsub_h(p16_2, q8s_2); - p16_3 = __lasx_xvsub_h(p16_3, q8s_3); - - p16_0 = lasx_madd_h(lasx_ext8_16(scale_0), p16_0); - p16_1 = lasx_madd_h(lasx_ext8_16(scale_1), p16_1); - p16_2 = lasx_madd_h(lasx_ext8_16(scale_2), p16_2); - p16_3 = lasx_madd_h(lasx_ext8_16(scale_3), p16_3); + p16_0 = lasx_madd_h(lasx_xvrepl128vei_h(scales_shuffled, 4 * j + 0), p16_0); + p16_1 = lasx_madd_h(lasx_xvrepl128vei_h(scales_shuffled, 4 * j + 1), p16_1); + p16_2 = lasx_madd_h(lasx_xvrepl128vei_h(scales_shuffled, 4 * j + 2), p16_2); + p16_3 = lasx_madd_h(lasx_xvrepl128vei_h(scales_shuffled, 4 * j + 3), p16_3); sumi = __lasx_xvadd_w(sumi, __lasx_xvadd_w(p16_0, p16_1)); sumi = __lasx_xvadd_w(sumi, __lasx_xvadd_w(p16_2, p16_3)); @@ -7421,7 +9296,130 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * restrict s, size_t bs, const void * r } *s = hsum_float_8(acc); +#elif defined(__VXE__) || defined(__VXE2__) + float sum = 0; + // Lower 4-bit and upper 2-bit masks + const uint8x16_t v_lm = vec_splat_u8(0x0F); + const uint8x16_t v_um = vec_splat_u8(0x03); + + const int32x4_t v_z = vec_splat_s32(0); + + int8x16_t q6b[4]; + uint8x16_t q6h[4]; + + uint8x16_t v_xl[4]; + uint8x16_t v_xh[2]; + int8x16_t v_y[4]; + + for (int i = 0; i < nb; ++i) { + const float d_all = GGML_FP16_TO_FP32(x[i].d); + + const uint8_t * GGML_RESTRICT x0l = x[i].ql; + const uint8_t * GGML_RESTRICT x0h = x[i].qh; + const int8_t * GGML_RESTRICT y0 = y[i].qs; + + const int8_t * GGML_RESTRICT scale = x[i].scales; + + const int16x8_t v_ysumsl = vec_xl(0 , y[i].bsums); + const int16x8_t v_ysumsh = vec_xl(16, y[i].bsums); + + const int8x16_t v_scale = vec_xl(0, scale); + const int16x8_t v_scalel = vec_unpackh(v_scale); + const int16x8_t v_scaleh = vec_unpackl(v_scale); + + const int32x4_t v_minslo = vec_mulo(v_ysumsl, v_scalel); + const int32x4_t v_minsle = vec_mule(v_ysumsl, v_scalel); + const int32x4_t v_minsho = vec_mulo(v_ysumsh, v_scaleh); + const int32x4_t v_minshe = vec_mule(v_ysumsh, v_scaleh); + const int32x4_t v_mins = v_minslo + v_minsle + v_minsho + v_minshe; + + const int32_t mins = v_mins[0] + v_mins[1] + v_mins[2] + v_mins[3]; + + int32_t isum = 0; + for (int j = 0; j < QK_K/128; ++j) { + // Load model upper 2 bits + v_xh[0] = vec_xl(0 , x0h); + v_xh[1] = vec_xl(16, x0h); + x0h += 32; + + // Load model lower 4 bits + v_xl[0] = vec_xl(0 , x0l); + v_xl[1] = vec_xl(16, x0l); + v_xl[2] = vec_xl(32, x0l); + v_xl[3] = vec_xl(48, x0l); + x0l += 64; + + // Load activation quants + v_y[0] = vec_xl(0 , y0); + v_y[1] = vec_xl(16, y0); + v_y[2] = vec_xl(32, y0); + v_y[3] = vec_xl(48, y0); + y0 += 64; + + q6h[0] = vec_sl(vec_and(v_um, v_xh[0]), 4); + q6h[1] = vec_sl(vec_and(v_um, v_xh[1]), 4); + uint8x16_t shifted = vec_sr(v_xh[0], 2); + q6h[2] = vec_sl(vec_and(v_um, shifted), 4); + shifted = vec_sr(v_xh[1], 2); + q6h[3] = vec_sl(vec_and(v_um, shifted), 4); + + q6b[0] = (int8x16_t)(vec_or(vec_and(v_xl[0], v_lm), q6h[0])); + q6b[1] = (int8x16_t)(vec_or(vec_and(v_xl[1], v_lm), q6h[1])); + q6b[2] = (int8x16_t)(vec_or(vec_and(v_xl[2], v_lm), q6h[2])); + q6b[3] = (int8x16_t)(vec_or(vec_and(v_xl[3], v_lm), q6h[3])); + + int32x4_t summs0 = ggml_vec_dot(v_z, q6b[0], v_y[0]); + int32x4_t summs1 = ggml_vec_dot(v_z, q6b[1], v_y[1]); + int32x4_t summs2 = ggml_vec_dot(v_z, q6b[2], v_y[2]); + int32x4_t summs3 = ggml_vec_dot(v_z, q6b[3], v_y[3]); + + isum += (summs0[0] + summs0[1] + summs0[2] + summs0[3]) * scale[0] + + (summs1[0] + summs1[1] + summs1[2] + summs1[3]) * scale[1] + + (summs2[0] + summs2[1] + summs2[2] + summs2[3]) * scale[2] + + (summs3[0] + summs3[1] + summs3[2] + summs3[3]) * scale[3]; + + scale += 4; + + + // Load activation quants + v_y[0] = vec_xl(0 , y0); + v_y[1] = vec_xl(16, y0); + v_y[2] = vec_xl(32, y0); + v_y[3] = vec_xl(48, y0); + y0 += 64; + + shifted = vec_sr(v_xh[0], 4); + q6h[0] = vec_sl(vec_and(v_um, shifted), 4); + shifted = vec_sr(v_xh[1], 4); + q6h[1] = vec_sl(vec_and(v_um, shifted), 4); + shifted = vec_sr(v_xh[0], 6); + q6h[2] = vec_sl(vec_and(v_um, shifted), 4); + shifted = vec_sr(v_xh[1], 6); + q6h[3] = vec_sl(vec_and(v_um, shifted), 4); + + q6b[0] = (int8x16_t)(vec_or(vec_sr(v_xl[0], 4), q6h[0])); + q6b[1] = (int8x16_t)(vec_or(vec_sr(v_xl[1], 4), q6h[1])); + q6b[2] = (int8x16_t)(vec_or(vec_sr(v_xl[2], 4), q6h[2])); + q6b[3] = (int8x16_t)(vec_or(vec_sr(v_xl[3], 4), q6h[3])); + + summs0 = ggml_vec_dot(v_z, q6b[0], v_y[0]); + summs1 = ggml_vec_dot(v_z, q6b[1], v_y[1]); + summs2 = ggml_vec_dot(v_z, q6b[2], v_y[2]); + summs3 = ggml_vec_dot(v_z, q6b[3], v_y[3]); + + isum += (summs0[0] + summs0[1] + summs0[2] + summs0[3]) * scale[0] + + (summs1[0] + summs1[1] + summs1[2] + summs1[3]) * scale[1] + + (summs2[0] + summs2[1] + summs2[2] + summs2[3]) * scale[2] + + (summs3[0] + summs3[1] + summs3[2] + summs3[3]) * scale[3]; + + scale += 4; + } + + sum += d_all * y[i].d * (isum - 32 * mins); + } + + *s = sum; #else int8_t aux8[QK_K]; @@ -7432,11 +9430,11 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * restrict s, size_t bs, const void * r float sumf = 0; for (int i = 0; i < nb; ++i) { - const uint8_t * restrict q4 = x[i].ql; - const uint8_t * restrict qh = x[i].qh; - const int8_t * restrict q8 = y[i].qs; + const uint8_t * GGML_RESTRICT q4 = x[i].ql; + const uint8_t * GGML_RESTRICT qh = x[i].qh; + const int8_t * GGML_RESTRICT q8 = y[i].qs; memset(aux32, 0, 8*sizeof(int32_t)); - int8_t * restrict a = aux8; + int8_t * GGML_RESTRICT a = aux8; for (int j = 0; j < QK_K; j += 128) { for (int l = 0; l < 32; ++l) { a[l + 0] = (int8_t)((q4[l + 0] & 0xF) | (((qh[l] >> 0) & 3) << 4)) - 32; @@ -7504,7 +9502,7 @@ static const int8_t keven_signs_q2xs[1024] = { }; #endif -void ggml_vec_dot_iq2_xxs_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) { +void ggml_vec_dot_iq2_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) { assert(n % QK_K == 0); assert(nrc == 1); UNUSED(nrc); @@ -7512,8 +9510,8 @@ void ggml_vec_dot_iq2_xxs_q8_K(int n, float * restrict s, size_t bs, const void UNUSED(by); UNUSED(bs); - const block_iq2_xxs * restrict x = vx; - const block_q8_K * restrict y = vy; + const block_iq2_xxs * GGML_RESTRICT x = vx; + const block_q8_K * GGML_RESTRICT y = vy; const int nb = n / QK_K; @@ -7531,8 +9529,8 @@ void ggml_vec_dot_iq2_xxs_q8_K(int n, float * restrict s, size_t bs, const void float sumf = 0; for (int i = 0; i < nb; ++i) { const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; - const uint16_t * restrict q2 = x[i].qs; - const int8_t * restrict q8 = y[i].qs; + const uint16_t * GGML_RESTRICT q2 = x[i].qs; + const int8_t * GGML_RESTRICT q8 = y[i].qs; float sumf1 = 0, sumf2 = 0; for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) { q8b = ggml_vld1q_s8_x4(q8); q8 += 64; @@ -7568,8 +9566,8 @@ void ggml_vec_dot_iq2_xxs_q8_K(int n, float * restrict s, size_t bs, const void __m256 accumf = _mm256_setzero_ps(); for (int i = 0; i < nb; ++i) { const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; - const uint16_t * restrict q2 = x[i].qs; - const int8_t * restrict q8 = y[i].qs; + const uint16_t * GGML_RESTRICT q2 = x[i].qs; + const int8_t * GGML_RESTRICT q8 = y[i].qs; __m256i sumi1 = _mm256_setzero_si256(); __m256i sumi2 = _mm256_setzero_si256(); for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) { @@ -7609,8 +9607,8 @@ void ggml_vec_dot_iq2_xxs_q8_K(int n, float * restrict s, size_t bs, const void __m256 accumf = _mm256_setzero_ps(); for (int i = 0; i < nb; ++i) { const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; - const uint16_t * restrict q2 = x[i].qs; - const int8_t * restrict q8 = y[i].qs; + const uint16_t * GGML_RESTRICT q2 = x[i].qs; + const int8_t * GGML_RESTRICT q8 = y[i].qs; __m128i sumi1_0 = _mm_setzero_si128(); __m128i sumi1_1 = _mm_setzero_si128(); __m128i sumi2_0 = _mm_setzero_si128(); @@ -7674,8 +9672,8 @@ void ggml_vec_dot_iq2_xxs_q8_K(int n, float * restrict s, size_t bs, const void vector signed int vsumi2 = v0; vector signed int vsumi3 = v0; - const uint16_t * restrict q2 = x[i].qs; - const int8_t * restrict q8 = y[i].qs; + const uint16_t * GGML_RESTRICT q2 = x[i].qs; + const int8_t * GGML_RESTRICT q8 = y[i].qs; for (int j = 0; j < QK_K/32; j += 2) { __builtin_prefetch(q2, 0, 1); @@ -7751,8 +9749,8 @@ void ggml_vec_dot_iq2_xxs_q8_K(int n, float * restrict s, size_t bs, const void __m256 accumf = (__m256)__lasx_xvldi(0); for (int i = 0; i < nb; ++i) { const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; - const uint16_t * restrict q2 = x[i].qs; - const int8_t * restrict q8 = y[i].qs; + const uint16_t * GGML_RESTRICT q2 = x[i].qs; + const int8_t * GGML_RESTRICT q8 = y[i].qs; __m256i sumi1 = __lasx_xvldi(0); __m256i sumi2 = __lasx_xvldi(0); for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) { @@ -7782,7 +9780,57 @@ void ggml_vec_dot_iq2_xxs_q8_K(int n, float * restrict s, size_t bs, const void } *s = 0.125f * hsum_float_8(accumf); - +//#elif defined(__VXE__) || defined(__VXE2__) +// const uint64_t * signs64 = (const uint64_t *)keven_signs_q2xs; +// +// uint32_t aux32[4]; +// const uint8_t * aux8 = (const uint8_t *)aux32; +// +// float sumf = 0; +// +// for (int i = 0; i < nb; ++i) { +// const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; +// const uint16_t * GGML_RESTRICT q2 = x[i].qs; +// const int8_t * GGML_RESTRICT q8 = y[i].qs; +// +// float sumf1 = 0, sumf2 = 0; +// +// for (int ib32 = 0; ib32 < QK_K/32; ib += 2) { +// int8x16_t q8b0 = vec_xl( 0, q8); +// int8x16_t qb81 = vec_xl(16, q8); +// int8x16_t q8b2 = vec_xl(32, q8); +// int8x16_t q8b3 = vec_xl(48, q8); +// q8 += 64; +// +// memcpy(aux32, q2, 4 * sizeof(uint32_t)); +// q2 += 8; +// +// int8x16_t q2u0 = { *(const int64_t *)(iq2xxs_grid + aux8[ 0]), *(const int64_t *)(iq2xxs_grid + aux8[ 1]) }; +// int8x16_t q2u1 = { *(const int64_t *)(iq2xxs_grid + aux8[ 2]), *(const int64_t *)(iq2xxs_grid + aux8[ 3]) }; +// int8x16_t q2u2 = { *(const int64_t *)(iq2xxs_grid + aux8[ 8]), *(const int64_t *)(iq2xxs_grid + aux8[ 9]) }; +// int8x16_t q2u3 = { *(const int64_t *)(iq2xxs_grid + aux8[10]), *(const int64_t *)(iq2xxs_grid + aux8[11]) }; +// +// int8x16_t q2s0 = { *(const int64_t *)(signs64 + ((aux32[1] >> 0) & 127)), *(const int64_t *)(signs64 + ((aux32[1] >> 7) & 127)) }; +// int8x16_t q2s1 = { *(const int64_t *)(signs64 + ((aux32[1] >> 14) & 127)), *(const int64_t *)(signs64 + ((aux32[1] >> 21) & 127)) }; +// int8x16_t q2s2 = { *(const int64_t *)(signs64 + ((aux32[3] >> 0) & 127)), *(const int64_t *)(signs64 + ((aux32[3] >> 7) & 127)) }; +// int8x16_t q2s3 = { *(const int64_t *)(signs64 + ((aux32[3] >> 14) & 127)), *(const int64_t *)(signs64 + ((aux32[3] >> 21) & 127)) }; +// +// q2u0 = vec_mul(q2u0, q2s0); +// q2u1 = vec_mul(q2u1, q2s1); +// q2u2 = vec_mul(q2u2, q2s2); +// q2u3 = vec_mul(q2u3, q2s3); +// +// const int32x4_t p1 = ggml_vec_dot(ggml_vec_dot(vec_splat_s32(0), q2u0, q8b0), q2u1, q8b1); +// const int32x4_t p2 = ggml_vec_dot(ggml_vec_dot(vec_splat_s32(0), q2u2, q8b2), q2u3, q8b3); +// +// sumf1 += (p1[0] + p1[1] + p1[2] + p1[3]) * (0.5f + (aux32[1] >> 28)); +// sumf2 += (p2[0] + p2[1] + p2[2] + p2[3]) * (0.5f + (aux32[3] >> 28)); +// } +// +// sumf += d * (sumf1 + sumf2); +// } +// +// *s = 0.25f * sumf; #else uint32_t aux32[2]; @@ -7791,8 +9839,8 @@ void ggml_vec_dot_iq2_xxs_q8_K(int n, float * restrict s, size_t bs, const void float sumf = 0.f; for (int i = 0; i < nb; ++i) { const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; - const uint16_t * restrict q2 = x[i].qs; - const int8_t * restrict q8 = y[i].qs; + const uint16_t * GGML_RESTRICT q2 = x[i].qs; + const int8_t * GGML_RESTRICT q8 = y[i].qs; int32_t bsum = 0; for (int ib32 = 0; ib32 < QK_K/32; ++ib32) { memcpy(aux32, q2, 2*sizeof(uint32_t)); @@ -7815,7 +9863,7 @@ void ggml_vec_dot_iq2_xxs_q8_K(int n, float * restrict s, size_t bs, const void #endif } -void ggml_vec_dot_iq2_xs_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) { +void ggml_vec_dot_iq2_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) { assert(n % QK_K == 0); assert(nrc == 1); UNUSED(nrc); @@ -7823,8 +9871,8 @@ void ggml_vec_dot_iq2_xs_q8_K(int n, float * restrict s, size_t bs, const void * UNUSED(by); UNUSED(bs); - const block_iq2_xs * restrict x = vx; - const block_q8_K * restrict y = vy; + const block_iq2_xs * GGML_RESTRICT x = vx; + const block_q8_K * GGML_RESTRICT y = vy; const int nb = n / QK_K; @@ -7841,8 +9889,8 @@ void ggml_vec_dot_iq2_xs_q8_K(int n, float * restrict s, size_t bs, const void * float sumf = 0; for (int i = 0; i < nb; ++i) { const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; - const uint16_t * restrict q2 = x[i].qs; - const int8_t * restrict q8 = y[i].qs; + const uint16_t * GGML_RESTRICT q2 = x[i].qs; + const int8_t * GGML_RESTRICT q8 = y[i].qs; const uint8x8_t scales8 = vld1_u8(x[i].scales); const uint8x8_t scales_l = vand_u8(scales8, vdup_n_u8(0xf)); const uint8x8_t scales_h = vshr_n_u8(scales8, 4); @@ -7919,8 +9967,8 @@ void ggml_vec_dot_iq2_xs_q8_K(int n, float * restrict s, size_t bs, const void * __m256 accumf = _mm256_setzero_ps(); for (int i = 0; i < nb; ++i) { const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; - const uint16_t * restrict q2 = x[i].qs; - const int8_t * restrict q8 = y[i].qs; + const uint16_t * GGML_RESTRICT q2 = x[i].qs; + const int8_t * GGML_RESTRICT q8 = y[i].qs; memcpy(&aux64, x[i].scales, 8); __m128i stmp = _mm_set1_epi64x(aux64); @@ -8040,8 +10088,8 @@ void ggml_vec_dot_iq2_xs_q8_K(int n, float * restrict s, size_t bs, const void * __m256 accumf = _mm256_setzero_ps(); for (int i = 0; i < nb; ++i) { const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; - const uint16_t * restrict q2 = x[i].qs; - const int8_t * restrict q8 = y[i].qs; + const uint16_t * GGML_RESTRICT q2 = x[i].qs; + const int8_t * GGML_RESTRICT q8 = y[i].qs; memcpy(&aux64, x[i].scales, 8); __m128i stmp = _mm_set1_epi64x(aux64); @@ -8195,8 +10243,8 @@ void ggml_vec_dot_iq2_xs_q8_K(int n, float * restrict s, size_t bs, const void * __m256 accumf = (__m256)__lasx_xvldi(0); for (int i = 0; i < nb; ++i) { const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; - const uint16_t * restrict q2 = x[i].qs; - const int8_t * restrict q8 = y[i].qs; + const uint16_t * GGML_RESTRICT q2 = x[i].qs; + const int8_t * GGML_RESTRICT q8 = y[i].qs; memcpy(&aux64, x[i].scales, 8); __m128i stmp = __lsx_vreplgr2vr_d(aux64); @@ -8293,9 +10341,9 @@ void ggml_vec_dot_iq2_xs_q8_K(int n, float * restrict s, size_t bs, const void * vector signed int vsumi2 = v0; vector signed int vsumi3 = v0; - const uint16_t * restrict q2 = x[i].qs; - const uint8_t * restrict sc = x[i].scales; - const int8_t * restrict q8 = y[i].qs; + const uint16_t * GGML_RESTRICT q2 = x[i].qs; + const uint8_t * GGML_RESTRICT sc = x[i].scales; + const int8_t * GGML_RESTRICT q8 = y[i].qs; for (int j = 0; j < QK_K/64; ++j) { __builtin_prefetch(q2, 0, 1); @@ -8365,9 +10413,9 @@ void ggml_vec_dot_iq2_xs_q8_K(int n, float * restrict s, size_t bs, const void * float sumf = 0.f; for (int i = 0; i < nb; ++i) { const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; - const uint16_t * restrict q2 = x[i].qs; - const uint8_t * restrict sc = x[i].scales; - const int8_t * restrict q8 = y[i].qs; + const uint16_t * GGML_RESTRICT q2 = x[i].qs; + const uint8_t * GGML_RESTRICT sc = x[i].scales; + const int8_t * GGML_RESTRICT q8 = y[i].qs; int32_t bsum = 0; for (int ib32 = 0; ib32 < QK_K/32; ++ib32) { const uint16_t ls1 = 2*(sc[ib32] & 0xf) + 1; @@ -8400,7 +10448,7 @@ void ggml_vec_dot_iq2_xs_q8_K(int n, float * restrict s, size_t bs, const void * #endif } -void ggml_vec_dot_iq2_s_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) { +void ggml_vec_dot_iq2_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) { assert(n % QK_K == 0); assert(nrc == 1); UNUSED(nrc); @@ -8408,8 +10456,8 @@ void ggml_vec_dot_iq2_s_q8_K(int n, float * restrict s, size_t bs, const void * UNUSED(by); UNUSED(bs); - const block_iq2_s * restrict x = vx; - const block_q8_K * restrict y = vy; + const block_iq2_s * GGML_RESTRICT x = vx; + const block_q8_K * GGML_RESTRICT y = vy; const int nb = n / QK_K; @@ -8435,10 +10483,10 @@ void ggml_vec_dot_iq2_s_q8_K(int n, float * restrict s, size_t bs, const void * const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; - const uint8_t * restrict qs = x[i].qs; - const uint8_t * restrict qh = x[i].qh; - const uint16_t * restrict signs = (const uint16_t *)(x[i].qs + QK_K/8); - const int8_t * restrict q8 = y[i].qs; + const uint8_t * GGML_RESTRICT qs = x[i].qs; + const uint8_t * GGML_RESTRICT qh = x[i].qh; + const uint16_t * GGML_RESTRICT signs = (const uint16_t *)(x[i].qs + QK_K/8); + const int8_t * GGML_RESTRICT q8 = y[i].qs; int sumi1 = 0, sumi2 = 0; for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) { @@ -8509,10 +10557,10 @@ void ggml_vec_dot_iq2_s_q8_K(int n, float * restrict s, size_t bs, const void * __m256 accumf = _mm256_setzero_ps(); for (int i = 0; i < nb; ++i) { const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; - const uint8_t * restrict qs = x[i].qs; - const uint8_t * restrict qh = x[i].qh; - const uint16_t * restrict signs = (const uint16_t *)(x[i].qs + QK_K/8); - const int8_t * restrict q8 = y[i].qs; + const uint8_t * GGML_RESTRICT qs = x[i].qs; + const uint8_t * GGML_RESTRICT qh = x[i].qh; + const uint16_t * GGML_RESTRICT signs = (const uint16_t *)(x[i].qs + QK_K/8); + const int8_t * GGML_RESTRICT q8 = y[i].qs; memcpy(&aux64, x[i].scales, 8); const __m128i scales8 = _mm_add_epi8(_mm_slli_epi16(_mm_and_si128(_mm_set_epi64x(aux64 >> 4, aux64), m4), 1), m1); @@ -8582,10 +10630,10 @@ void ggml_vec_dot_iq2_s_q8_K(int n, float * restrict s, size_t bs, const void * __m256 accumf = _mm256_setzero_ps(); for (int i = 0; i < nb; ++i) { const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; - const uint8_t * restrict qs = x[i].qs; - const uint8_t * restrict qh = x[i].qh; - const uint16_t * restrict signs = (const uint16_t *)(x[i].qs + QK_K/8); - const int8_t * restrict q8 = y[i].qs; + const uint8_t * GGML_RESTRICT qs = x[i].qs; + const uint8_t * GGML_RESTRICT qh = x[i].qh; + const uint16_t * GGML_RESTRICT signs = (const uint16_t *)(x[i].qs + QK_K/8); + const int8_t * GGML_RESTRICT q8 = y[i].qs; memcpy(&aux64, x[i].scales, 8); const __m128i scales8 = _mm_add_epi8(_mm_slli_epi16(_mm_and_si128(_mm_set_epi64x(aux64 >> 4, aux64), m4), 1), m1); @@ -8680,11 +10728,11 @@ void ggml_vec_dot_iq2_s_q8_K(int n, float * restrict s, size_t bs, const void * vector signed int vsumi2 = v0; vector signed int vsumi3 = v0; - const uint8_t * restrict q2 = x[i].qs; - const uint8_t * restrict qh = x[i].qh; - const uint16_t * restrict signs = (const uint16_t *)(x[i].qs + QK_K/8); - const uint8_t * restrict sc = x[i].scales; - const int8_t * restrict q8 = y[i].qs; + const uint8_t * GGML_RESTRICT q2 = x[i].qs; + const uint8_t * GGML_RESTRICT qh = x[i].qh; + const uint16_t * GGML_RESTRICT signs = (const uint16_t *)(x[i].qs + QK_K/8); + const uint8_t * GGML_RESTRICT sc = x[i].scales; + const int8_t * GGML_RESTRICT q8 = y[i].qs; for (int j = 0; j < QK_K/32; j += 2) { __builtin_prefetch(q2, 0, 1); @@ -8781,10 +10829,10 @@ void ggml_vec_dot_iq2_s_q8_K(int n, float * restrict s, size_t bs, const void * __m256 accumf = (__m256)__lasx_xvldi(0); for (int i = 0; i < nb; ++i) { const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; - const uint8_t * restrict qs = x[i].qs; - const uint8_t * restrict qh = x[i].qh; - const uint16_t * restrict signs = (const uint16_t *)(x[i].qs + QK_K/8); - const int8_t * restrict q8 = y[i].qs; + const uint8_t * GGML_RESTRICT qs = x[i].qs; + const uint8_t * GGML_RESTRICT qh = x[i].qh; + const uint16_t * GGML_RESTRICT signs = (const uint16_t *)(x[i].qs + QK_K/8); + const int8_t * GGML_RESTRICT q8 = y[i].qs; __m128i tmp1; memcpy(&aux64, x[i].scales, 8); @@ -8878,7 +10926,7 @@ void ggml_vec_dot_iq2_s_q8_K(int n, float * restrict s, size_t bs, const void * } -void ggml_vec_dot_iq3_xxs_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) { +void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) { assert(n % QK_K == 0); assert(nrc == 1); UNUSED(nrc); @@ -8886,8 +10934,8 @@ void ggml_vec_dot_iq3_xxs_q8_K(int n, float * restrict s, size_t bs, const void UNUSED(by); UNUSED(bs); - const block_iq3_xxs * restrict x = vx; - const block_q8_K * restrict y = vy; + const block_iq3_xxs * GGML_RESTRICT x = vx; + const block_q8_K * GGML_RESTRICT y = vy; const int nb = n / QK_K; @@ -8903,9 +10951,9 @@ void ggml_vec_dot_iq3_xxs_q8_K(int n, float * restrict s, size_t bs, const void float sumf = 0; for (int i = 0; i < nb; ++i) { const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; - const uint8_t * restrict q3 = x[i].qs; - const uint8_t * restrict gas = x[i].qs + QK_K/4; - const int8_t * restrict q8 = y[i].qs; + const uint8_t * GGML_RESTRICT q3 = x[i].qs; + const uint8_t * GGML_RESTRICT gas = x[i].qs + QK_K/4; + const int8_t * GGML_RESTRICT q8 = y[i].qs; float sumf1 = 0, sumf2 = 0; for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) { q8b = ggml_vld1q_s8_x4(q8); q8 += 64; @@ -8941,9 +10989,9 @@ void ggml_vec_dot_iq3_xxs_q8_K(int n, float * restrict s, size_t bs, const void __m256 accumf = _mm256_setzero_ps(); for (int i = 0; i < nb; ++i) { const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; - const uint8_t * restrict q3 = x[i].qs; - const uint8_t * restrict gas = x[i].qs + QK_K/4; - const int8_t * restrict q8 = y[i].qs; + const uint8_t * GGML_RESTRICT q3 = x[i].qs; + const uint8_t * GGML_RESTRICT gas = x[i].qs + QK_K/4; + const int8_t * GGML_RESTRICT q8 = y[i].qs; __m256i sumi1 = _mm256_setzero_si256(); __m256i sumi2 = _mm256_setzero_si256(); for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) { @@ -8986,9 +11034,9 @@ void ggml_vec_dot_iq3_xxs_q8_K(int n, float * restrict s, size_t bs, const void __m256 accumf = _mm256_setzero_ps(); for (int i = 0; i < nb; ++i) { const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; - const uint8_t * restrict q3 = x[i].qs; - const uint8_t * restrict gas = x[i].qs + QK_K/4; - const int8_t * restrict q8 = y[i].qs; + const uint8_t * GGML_RESTRICT q3 = x[i].qs; + const uint8_t * GGML_RESTRICT gas = x[i].qs + QK_K/4; + const int8_t * GGML_RESTRICT q8 = y[i].qs; __m128i sumi1_0 = _mm_setzero_si128(); __m128i sumi1_1 = _mm_setzero_si128(); __m128i sumi2_0 = _mm_setzero_si128(); @@ -9055,9 +11103,9 @@ void ggml_vec_dot_iq3_xxs_q8_K(int n, float * restrict s, size_t bs, const void vector signed int vsumi2 = v0; vector signed int vsumi3 = v0; - const uint8_t * restrict q3 = x[i].qs; - const uint32_t * restrict signs = (const uint32_t *)(x[i].qs + QK_K/4); - const int8_t * restrict q8 = y[i].qs; + const uint8_t * GGML_RESTRICT q3 = x[i].qs; + const uint32_t * GGML_RESTRICT signs = (const uint32_t *)(x[i].qs + QK_K/4); + const int8_t * GGML_RESTRICT q8 = y[i].qs; #pragma GCC unroll 1 for (int j = 0; j < QK_K/32; j += 2) { @@ -9129,9 +11177,9 @@ void ggml_vec_dot_iq3_xxs_q8_K(int n, float * restrict s, size_t bs, const void __m256 accumf = (__m256)__lasx_xvldi(0); for (int i = 0; i < nb; ++i) { const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; - const uint8_t * restrict q3 = x[i].qs; - const uint8_t * restrict gas = x[i].qs + QK_K/4; - const int8_t * restrict q8 = y[i].qs; + const uint8_t * GGML_RESTRICT q3 = x[i].qs; + const uint8_t * GGML_RESTRICT gas = x[i].qs + QK_K/4; + const int8_t * GGML_RESTRICT q8 = y[i].qs; __m256i sumi1 = __lasx_xvldi(0); __m256i sumi2 = __lasx_xvldi(0); for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) { @@ -9174,9 +11222,9 @@ void ggml_vec_dot_iq3_xxs_q8_K(int n, float * restrict s, size_t bs, const void float sumf = 0.f; for (int i = 0; i < nb; ++i) { const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; - const uint8_t * restrict q3 = x[i].qs; - const uint8_t * restrict gas = x[i].qs + QK_K/4; - const int8_t * restrict q8 = y[i].qs; + const uint8_t * GGML_RESTRICT q3 = x[i].qs; + const uint8_t * GGML_RESTRICT gas = x[i].qs + QK_K/4; + const int8_t * GGML_RESTRICT q8 = y[i].qs; int32_t bsum = 0; for (int ib32 = 0; ib32 < QK_K/32; ++ib32) { memcpy(&aux32, gas, sizeof(uint32_t)); gas += sizeof(uint32_t); @@ -9201,7 +11249,7 @@ void ggml_vec_dot_iq3_xxs_q8_K(int n, float * restrict s, size_t bs, const void #endif } -void ggml_vec_dot_iq3_s_q8_K (int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) { +void ggml_vec_dot_iq3_s_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) { assert(n % QK_K == 0); assert(nrc == 1); UNUSED(nrc); @@ -9209,8 +11257,8 @@ void ggml_vec_dot_iq3_s_q8_K (int n, float * restrict s, size_t bs, const void * UNUSED(by); UNUSED(bs); - const block_iq3_s * restrict x = vx; - const block_q8_K * restrict y = vy; + const block_iq3_s * GGML_RESTRICT x = vx; + const block_q8_K * GGML_RESTRICT y = vy; const int nb = n / QK_K; @@ -9247,10 +11295,10 @@ void ggml_vec_dot_iq3_s_q8_K (int n, float * restrict s, size_t bs, const void * float sumf = 0; for (int i = 0; i < nb; ++i) { const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; - const uint8_t * restrict qs = x[i].qs; - const uint8_t * restrict qh = x[i].qh; - const uint16_t * restrict signs = (const uint16_t *)x[i].signs; - const int8_t * restrict q8 = y[i].qs; + const uint8_t * GGML_RESTRICT qs = x[i].qs; + const uint8_t * GGML_RESTRICT qh = x[i].qh; + const uint16_t * GGML_RESTRICT signs = (const uint16_t *)x[i].signs; + const int8_t * GGML_RESTRICT q8 = y[i].qs; memcpy(scales32, x[i].scales, 4); scales32[1] = (((scales32[0] >> 4) & 0x0f0f0f0f) << 1) | 0x01010101; @@ -9329,10 +11377,10 @@ void ggml_vec_dot_iq3_s_q8_K (int n, float * restrict s, size_t bs, const void * __m256 accumf = _mm256_setzero_ps(); for (int i = 0; i < nb; ++i) { const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; - const uint8_t * restrict qs = x[i].qs; - const uint8_t * restrict qh = x[i].qh; - const uint16_t * restrict signs = (const uint16_t *)x[i].signs; - const int8_t * restrict q8 = y[i].qs; + const uint8_t * GGML_RESTRICT qs = x[i].qs; + const uint8_t * GGML_RESTRICT qh = x[i].qh; + const uint16_t * GGML_RESTRICT signs = (const uint16_t *)x[i].signs; + const int8_t * GGML_RESTRICT q8 = y[i].qs; __m256i sumi1 = _mm256_setzero_si256(); __m256i sumi2 = _mm256_setzero_si256(); for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) { @@ -9414,10 +11462,10 @@ void ggml_vec_dot_iq3_s_q8_K (int n, float * restrict s, size_t bs, const void * __m256 accumf = _mm256_setzero_ps(); for (int i = 0; i < nb; ++i) { const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; - const uint8_t * restrict qs = x[i].qs; - const uint8_t * restrict qh = x[i].qh; - const uint16_t * restrict signs = (const uint16_t *)x[i].signs; - const int8_t * restrict q8 = y[i].qs; + const uint8_t * GGML_RESTRICT qs = x[i].qs; + const uint8_t * GGML_RESTRICT qh = x[i].qh; + const uint16_t * GGML_RESTRICT signs = (const uint16_t *)x[i].signs; + const int8_t * GGML_RESTRICT q8 = y[i].qs; __m128i sumi1_0 = _mm_setzero_si128(); __m128i sumi1_1 = _mm_setzero_si128(); __m128i sumi2_0 = _mm_setzero_si128(); @@ -9515,11 +11563,11 @@ void ggml_vec_dot_iq3_s_q8_K (int n, float * restrict s, size_t bs, const void * vector float vyd = vec_splats(y[i].d); vector float vd = vec_mul(vxd, vyd); - const uint8_t * restrict q3 = x[i].qs; - const uint8_t * restrict qh = x[i].qh; - const uint16_t * restrict signs = (const uint16_t *)(x[i].signs); - const uint8_t * restrict sc = x[i].scales; - const int8_t * restrict q8 = y[i].qs; + const uint8_t * GGML_RESTRICT q3 = x[i].qs; + const uint8_t * GGML_RESTRICT qh = x[i].qh; + const uint16_t * GGML_RESTRICT signs = (const uint16_t *)(x[i].signs); + const uint8_t * GGML_RESTRICT sc = x[i].scales; + const int8_t * GGML_RESTRICT q8 = y[i].qs; vector signed int vsumi0 = v0; vector signed int vsumi1 = v0; @@ -9626,10 +11674,10 @@ void ggml_vec_dot_iq3_s_q8_K (int n, float * restrict s, size_t bs, const void * __m256 accumf = (__m256)__lasx_xvldi(0); for (int i = 0; i < nb; ++i) { const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; - const uint8_t * restrict qs = x[i].qs; - const uint8_t * restrict qh = x[i].qh; - const uint16_t * restrict signs = (const uint16_t *)x[i].signs; - const int8_t * restrict q8 = y[i].qs; + const uint8_t * GGML_RESTRICT qs = x[i].qs; + const uint8_t * GGML_RESTRICT qh = x[i].qh; + const uint16_t * GGML_RESTRICT signs = (const uint16_t *)x[i].signs; + const int8_t * GGML_RESTRICT q8 = y[i].qs; __m256i sumi1 = __lasx_xvldi(0); __m256i sumi2 = __lasx_xvldi(0); for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) { @@ -9687,10 +11735,10 @@ void ggml_vec_dot_iq3_s_q8_K (int n, float * restrict s, size_t bs, const void * float sumf = 0.f; for (int i = 0; i < nb; ++i) { const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; - const uint8_t * restrict qs = x[i].qs; - const uint8_t * restrict qh = x[i].qh; - const uint8_t * restrict signs = x[i].signs; - const int8_t * restrict q8 = y[i].qs; + const uint8_t * GGML_RESTRICT qs = x[i].qs; + const uint8_t * GGML_RESTRICT qh = x[i].qh; + const uint8_t * GGML_RESTRICT signs = x[i].signs; + const int8_t * GGML_RESTRICT q8 = y[i].qs; int32_t bsum = 0; for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) { const uint32_t ls1 = 2*(x[i].scales[ib32/2] & 0xf) + 1; @@ -9736,17 +11784,13 @@ static inline __m256i mul_add_epi8(const __m256i x, const __m256i y) { } #elif defined(__loongarch_asx) static inline __m256i mul_add_epi8(const __m256i x, const __m256i y) { - const __m256i ax = __lasx_xvsigncov_b(x, x); - const __m256i sy = __lasx_xvsigncov_b(x, y); - __m256i tmp1, tmp2, tmp3; - tmp1 = __lasx_xvmulwev_h_bu_b(ax, sy); - tmp2 = __lasx_xvmulwod_h_bu_b(ax, sy); - tmp3 = __lasx_xvadd_h(tmp1, tmp2); - return __lasx_xvsat_h(tmp3, 15); + const __m256i a = __lasx_xvmulwev_h_b(x, y); + const __m256i b = __lasx_xvmulwod_h_b(x, y); + return __lasx_xvadd_h(a, b); } #endif -void ggml_vec_dot_iq1_s_q8_K (int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) { +void ggml_vec_dot_iq1_s_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) { assert(n % QK_K == 0); assert(nrc == 1); UNUSED(nrc); @@ -9754,8 +11798,8 @@ void ggml_vec_dot_iq1_s_q8_K (int n, float * restrict s, size_t bs, const void UNUSED(by); UNUSED(bs); - const block_iq1_s * restrict x = vx; - const block_q8_K * restrict y = vy; + const block_iq1_s * GGML_RESTRICT x = vx; + const block_q8_K * GGML_RESTRICT y = vy; const int nb = n / QK_K; @@ -9817,10 +11861,19 @@ void ggml_vec_dot_iq1_s_q8_K (int n, float * restrict s, size_t bs, const void __m256i sumi = _mm256_setzero_si256(); int sumi1 = 0; for (int ib = 0; ib < QK_K/32; ib += 2) { +#ifdef __BMI2__ + const uint64_t packed_idx1 = _pdep_u64(*(const uint32_t *)qs, 0x00ff00ff00ff00ffULL) | _pdep_u64(qh[ib], 0x700070007000700ULL); + const uint64_t packed_idx2 = _pdep_u64(*(const uint32_t *)(qs + 4), 0x00ff00ff00ff00ffULL) | _pdep_u64(qh[ib + 1], 0x700070007000700ULL); + const uint16_t *idx1 = (const uint16_t *)(&packed_idx1); + const uint16_t *idx2 = (const uint16_t *)(&packed_idx2); + const __m256i q1b_1 = _mm256_set_epi64x(iq1s_grid[idx1[3]], iq1s_grid[idx1[2]], iq1s_grid[idx1[1]], iq1s_grid[idx1[0]]); + const __m256i q1b_2 = _mm256_set_epi64x(iq1s_grid[idx2[3]], iq1s_grid[idx2[2]], iq1s_grid[idx2[1]], iq1s_grid[idx2[0]]); +#else const __m256i q1b_1 = _mm256_set_epi64x(iq1s_grid[qs[3] | ((qh[ib+0] >> 1) & 0x700)], iq1s_grid[qs[2] | ((qh[ib+0] << 2) & 0x700)], iq1s_grid[qs[1] | ((qh[ib+0] << 5) & 0x700)], iq1s_grid[qs[0] | ((qh[ib+0] << 8) & 0x700)]); const __m256i q1b_2 = _mm256_set_epi64x(iq1s_grid[qs[7] | ((qh[ib+1] >> 1) & 0x700)], iq1s_grid[qs[6] | ((qh[ib+1] << 2) & 0x700)], iq1s_grid[qs[5] | ((qh[ib+1] << 5) & 0x700)], iq1s_grid[qs[4] | ((qh[ib+1] << 8) & 0x700)]); +#endif qs += 8; const __m256i q8b_1 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32; const __m256i q8b_2 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32; @@ -9913,10 +11966,10 @@ void ggml_vec_dot_iq1_s_q8_K (int n, float * restrict s, size_t bs, const void vector signed int vsumi3 = vec_splats((int32_t)0); vector signed int vsumi8 = vec_splats((int32_t)0); - const uint8_t * restrict q1 = x[i].qs; - const uint16_t * restrict qh = x[i].qh; - const int8_t * restrict q8 = y[i].qs; - const int16_t * restrict qs = y[i].bsums; + const uint8_t * GGML_RESTRICT q1 = x[i].qs; + const uint16_t * GGML_RESTRICT qh = x[i].qh; + const int8_t * GGML_RESTRICT q8 = y[i].qs; + const int16_t * GGML_RESTRICT qs = y[i].bsums; for (int j = 0; j < QK_K/32; j += 2) { __builtin_prefetch(q1, 0, 1); @@ -10077,7 +12130,7 @@ void ggml_vec_dot_iq1_s_q8_K (int n, float * restrict s, size_t bs, const void #endif } -void ggml_vec_dot_iq1_m_q8_K (int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) { +void ggml_vec_dot_iq1_m_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) { assert(n % QK_K == 0); assert(nrc == 1); UNUSED(nrc); @@ -10085,8 +12138,8 @@ void ggml_vec_dot_iq1_m_q8_K (int n, float * restrict s, size_t bs, const void UNUSED(by); UNUSED(bs); - const block_iq1_m * restrict x = vx; - const block_q8_K * restrict y = vy; + const block_iq1_m * GGML_RESTRICT x = vx; + const block_q8_K * GGML_RESTRICT y = vy; const int nb = n / QK_K; @@ -10166,6 +12219,10 @@ void ggml_vec_dot_iq1_m_q8_K (int n, float * restrict s, size_t bs, const void const __m256i mask = _mm256_set1_epi16(0x7); const __m256i mone = _mm256_set1_epi16(1); + const __m256i mone8 = _mm256_set1_epi8(1); + const __m256i mtwo8 = _mm256_set1_epi8(2); + // VPSHUFB cannot cross 128-bit lanes so odd shifts go to upper half. + const __m256i scales_shift = _mm256_set_epi64x(9, 3, 6, 0); __m256 accum1 = _mm256_setzero_ps(); __m256 accum2 = _mm256_setzero_ps(); @@ -10177,10 +12234,33 @@ void ggml_vec_dot_iq1_m_q8_K (int n, float * restrict s, size_t bs, const void const uint16_t * sc = (const uint16_t *)x[i].scales; scale.u16 = (sc[0] >> 12) | ((sc[1] >> 8) & 0x00f0) | ((sc[2] >> 4) & 0x0f00) | (sc[3] & 0xf000); + // Extract 3-bit scales (16 values) + __m256i scales = _mm256_set1_epi64x(*(const uint64_t*)sc); + scales = _mm256_srlv_epi64(scales, scales_shift); + scales = _mm256_add_epi16(_mm256_slli_epi16(_mm256_and_si256(scales, mask), 1), mone); + + // Indices to repeat each scale 8 times. + __m256i scales_idx1 = _mm256_set1_epi16(0x0100); + __m256i scales_idx2 = _mm256_add_epi8(scales_idx1, _mm256_set1_epi8(8)); __m256i sumi1 = _mm256_setzero_si256(); __m256i sumi2 = _mm256_setzero_si256(); for (int ib = 0; ib < QK_K/32; ib += 2) { +#ifdef __BMI2__ + const uint64_t packed_idx1 = _pdep_u64(*(const uint32_t *)qs, 0x00ff00ff00ff00ffULL) + | _pdep_u64(*(const uint16_t*)(qh) & 0x7777, 0xf000f000f000f00ULL); + const uint64_t packed_idx2 = _pdep_u64(*(const uint32_t *)(qs + 4), 0x00ff00ff00ff00ffULL) + | _pdep_u64(*(const uint16_t*)(qh + 2) & 0x7777, 0xf000f000f000f00ULL); + const uint16_t *idx1 = (const uint16_t *)(&packed_idx1); + const uint16_t *idx2 = (const uint16_t *)(&packed_idx2); + const __m256i q1b_1 = _mm256_set_epi64x(iq1s_grid[idx1[3]], iq1s_grid[idx1[2]], iq1s_grid[idx1[1]], iq1s_grid[idx1[0]]); + const __m256i q1b_2 = _mm256_set_epi64x(iq1s_grid[idx2[3]], iq1s_grid[idx2[2]], iq1s_grid[idx2[1]], iq1s_grid[idx2[0]]); + + // Convert signs to bytes 0x81 (negative) or 0x01 (positive) + const uint64_t delta_sign = _pdep_u64(*(const uint32_t*)(qh) & 0x88888888, 0xf0f0f0f0f0f0f0f0ULL); + const __m256i delta1 = _mm256_or_si256(mone8, _mm256_cvtepi8_epi64(_mm_set1_epi32(delta_sign))); + const __m256i delta2 = _mm256_or_si256(mone8, _mm256_cvtepi8_epi64(_mm_set1_epi32(delta_sign >> 32))); +#else const __m256i q1b_1 = _mm256_set_epi64x( iq1s_grid[qs[3] | (((uint16_t)qh[1] << 4) & 0x700)], iq1s_grid[qs[2] | (((uint16_t)qh[1] << 8) & 0x700)], iq1s_grid[qs[1] | (((uint16_t)qh[0] << 4) & 0x700)], iq1s_grid[qs[0] | (((uint16_t)qh[0] << 8) & 0x700)] @@ -10189,11 +12269,6 @@ void ggml_vec_dot_iq1_m_q8_K (int n, float * restrict s, size_t bs, const void iq1s_grid[qs[7] | (((uint16_t)qh[3] << 4) & 0x700)], iq1s_grid[qs[6] | (((uint16_t)qh[3] << 8) & 0x700)], iq1s_grid[qs[5] | (((uint16_t)qh[2] << 4) & 0x700)], iq1s_grid[qs[4] | (((uint16_t)qh[2] << 8) & 0x700)] ); - const __m256i q8b_1 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32; - const __m256i q8b_2 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32; - - const __m256i dot1 = mul_add_epi8(q1b_1, q8b_1); - const __m256i dot2 = mul_add_epi8(q1b_2, q8b_2); const __m256i delta1 = _mm256_set_epi64x(qh[1] & 0x80 ? 0xffffffffffffffff : 0x0101010101010101, qh[1] & 0x08 ? 0xffffffffffffffff : 0x0101010101010101, @@ -10203,15 +12278,21 @@ void ggml_vec_dot_iq1_m_q8_K (int n, float * restrict s, size_t bs, const void qh[3] & 0x08 ? 0xffffffffffffffff : 0x0101010101010101, qh[2] & 0x80 ? 0xffffffffffffffff : 0x0101010101010101, qh[2] & 0x08 ? 0xffffffffffffffff : 0x0101010101010101); +#endif + const __m256i q8b_1 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32; + const __m256i q8b_2 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32; - const __m256i dot3 = mul_add_epi8(delta1, q8b_1); - const __m256i dot4 = mul_add_epi8(delta2, q8b_2); + const __m256i dot1 = mul_add_epi8(q1b_1, q8b_1); + const __m256i dot2 = mul_add_epi8(q1b_2, q8b_2); + const __m256i dot3 = _mm256_maddubs_epi16(mone8, _mm256_sign_epi8(q8b_1, delta1)); + const __m256i dot4 = _mm256_maddubs_epi16(mone8, _mm256_sign_epi8(q8b_2, delta2)); - __m256i scale1 = MM256_SET_M128I(_mm_set1_epi16(sc[ib/2] >> 3), _mm_set1_epi16(sc[ib/2] >> 0)); - __m256i scale2 = MM256_SET_M128I(_mm_set1_epi16(sc[ib/2] >> 9), _mm_set1_epi16(sc[ib/2] >> 6)); + __m256i scale1 = _mm256_shuffle_epi8(scales, scales_idx1); + __m256i scale2 = _mm256_shuffle_epi8(scales, scales_idx2); + + scales_idx1 = _mm256_add_epi8(scales_idx1, mtwo8); + scales_idx2 = _mm256_add_epi8(scales_idx2, mtwo8); - scale1 = _mm256_add_epi16(_mm256_slli_epi16(_mm256_and_si256(scale1, mask), 1), mone); - scale2 = _mm256_add_epi16(_mm256_slli_epi16(_mm256_and_si256(scale2, mask), 1), mone); const __m256i p1 = _mm256_madd_epi16(dot1, scale1); const __m256i p2 = _mm256_madd_epi16(dot2, scale2); const __m256i p3 = _mm256_madd_epi16(dot3, scale1); @@ -10367,7 +12448,7 @@ void ggml_vec_dot_iq1_m_q8_K (int n, float * restrict s, size_t bs, const void #endif } -void ggml_vec_dot_iq4_nl_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) { +void ggml_vec_dot_iq4_nl_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) { assert(nrc == 1); UNUSED(nrc); UNUSED(bx); @@ -10376,8 +12457,8 @@ void ggml_vec_dot_iq4_nl_q8_0(int n, float * restrict s, size_t bs, const void * assert(n % QK4_NL == 0); static_assert(QK4_NL == QK8_0, "QK4_NL and QK8_0 must be the same"); - const block_iq4_nl * restrict x = vx; - const block_q8_0 * restrict y = vy; + const block_iq4_nl * GGML_RESTRICT x = vx; + const block_q8_0 * GGML_RESTRICT y = vy; const int nb = n / QK4_NL; @@ -10547,6 +12628,27 @@ void ggml_vec_dot_iq4_nl_q8_0(int n, float * restrict s, size_t bs, const void * sumf = hsum_float_8(__lasx_xvfadd_s(accum1, accum2)); +#elif defined(__VXE__) || defined(__VXE2__) + const int8x16_t v_k = vec_xl(0, kvalues_iq4nl); + const uint8x16_t v_m = vec_splat_u8(0x0F); + + for (; ib < nb; ++ib) { + const block_iq4_nl * GGML_RESTRICT x0 = &x[ib]; + const block_q8_0 * GGML_RESTRICT y0 = &y[ib]; + + const uint8x16_t v_x = vec_xl(0, x0->qs); + int8x16_t v_xl = (int8x16_t)vec_and(v_x, v_m); + int8x16_t v_xh = (int8x16_t)vec_sr(v_x, 4); + + v_xl = vec_perm(v_k, v_k, (uchar8x16_t)v_xl); + v_xh = vec_perm(v_k, v_k, (uchar8x16_t)v_xh); + + const int8x16_t v_yl = vec_xl(0 , y0->qs); + const int8x16_t v_yh = vec_xl(QK8_0/2, y0->qs); + const int32x4_t v_xy = ggml_vec_dot(ggml_vec_dot(vec_splats(0), v_xl, v_yl), v_xh, v_yh); + + sumf += GGML_FP16_TO_FP32(x0->d) * GGML_FP16_TO_FP32(y0->d) * (v_xy[0] + v_xy[1] + v_xy[2] + v_xy[3]); + } #endif for (; ib < nb; ++ib) { const float d = GGML_FP16_TO_FP32(y[ib].d)*GGML_FP16_TO_FP32(x[ib].d); @@ -10560,7 +12662,7 @@ void ggml_vec_dot_iq4_nl_q8_0(int n, float * restrict s, size_t bs, const void * *s = sumf; } -void ggml_vec_dot_iq4_xs_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) { +void ggml_vec_dot_iq4_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) { assert(nrc == 1); UNUSED(nrc); UNUSED(bx); @@ -10568,8 +12670,8 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * restrict s, size_t bs, const void * UNUSED(bs); assert(n % QK_K == 0); - const block_iq4_xs * restrict x = vx; - const block_q8_K * restrict y = vy; + const block_iq4_xs * GGML_RESTRICT x = vx; + const block_q8_K * GGML_RESTRICT y = vy; const int nb = n / QK_K; @@ -10726,9 +12828,9 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * restrict s, size_t bs, const void * uint16_t h = x[ibl].scales_h; - const uint8_t * restrict q4 = x[ibl].qs; - const uint8_t * restrict sc = x[ibl].scales_l; - const int8_t * restrict q8 = y[ibl].qs; + const uint8_t * GGML_RESTRICT q4 = x[ibl].qs; + const uint8_t * GGML_RESTRICT sc = x[ibl].scales_l; + const int8_t * GGML_RESTRICT q8 = y[ibl].qs; for (int ib = 0; ib < QK_K/64; ib ++ ) { __builtin_prefetch(q4, 0, 1); @@ -10792,67 +12894,31 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * restrict s, size_t bs, const void * #elif defined(__loongarch_asx) const __m128i values128 = __lsx_vld((const __m128i*)kvalues_iq4nl, 0); - const __m128i m4b = __lsx_vreplgr2vr_b(0x0f); __m256 accum = (__m256)__lasx_xvldi(0); - __m256i tmp1; - __m128i tmp0, tmp2, tmp3, tmp4, mask_8f, mask; - mask_8f = __lsx_vreplgr2vr_b(0x8f); for (int ibl = 0; ibl < nb; ++ibl) { const uint8_t * qs = x[ibl].qs; const int8_t * q8 = y[ibl].qs; uint16_t sh = x[ibl].scales_h; __m256i sumi1 = __lasx_xvldi(0); __m256i sumi2 = __lasx_xvldi(0); - __m128i zero = __lsx_vldi(0); for (int ib = 0; ib < QK_K/32; ib += 2) { - const __m128i q4bits_1 = __lsx_vld((const __m128i*)qs, 0); qs += 16; - const __m128i q4bits_2 = __lsx_vld((const __m128i*)qs, 0); qs += 16; + const __m128i q4bits_1 = __lsx_vld((const __m128i*)qs, 0); qs += 16; + const __m128i q4bits_2 = __lsx_vld((const __m128i*)qs, 0); qs += 16; const __m256i q8b_1 = __lasx_xvld((const __m256i *)q8, 0); q8 += 32; const __m256i q8b_2 = __lasx_xvld((const __m256i *)q8, 0); q8 += 32; - tmp2 = __lsx_vand_v(__lsx_vand_v(__lsx_vsrli_h(q4bits_1, 4), m4b), mask_8f); - tmp0 = __lsx_vori_b(tmp2, 0x10); - mask = __lsx_vsle_b(zero, tmp2); - tmp3 = __lsx_vand_v(tmp0, mask); - tmp3 = __lsx_vshuf_b(values128, zero, tmp3); - - tmp2 = __lsx_vand_v(__lsx_vand_v(q4bits_1, m4b), mask_8f); - tmp0 = __lsx_vori_b(tmp2, 0x10); - mask = __lsx_vsle_b(zero, tmp2); - tmp4 = __lsx_vand_v(tmp0, mask); - tmp4 = __lsx_vshuf_b(values128, zero, tmp4); - - const __m256i q4b_1 = lasx_insertf128(tmp3, tmp4); - - tmp2 = __lsx_vand_v(__lsx_vand_v(__lsx_vsrli_h(q4bits_2, 4), m4b), mask_8f); - tmp0 = __lsx_vori_b(tmp2, 0x10); - mask = __lsx_vsle_b(zero, tmp2); - tmp3 = __lsx_vand_v(tmp0, mask); - tmp3 = __lsx_vshuf_b(values128, zero, tmp3); - - tmp2 = __lsx_vand_v(__lsx_vand_v(q4bits_2, m4b), mask_8f); - tmp0 = __lsx_vori_b(tmp2, 0x10); - mask = __lsx_vsle_b(zero, tmp2); - tmp4 = __lsx_vand_v(tmp0, mask); - tmp4 = __lsx_vshuf_b(values128, zero, tmp4); - - const __m256i q4b_2 = lasx_insertf128(tmp3, tmp4); - + const __m256i q4b_1 = lasx_insertf128(__lsx_vshuf_b(values128, values128, __lsx_vsrli_b(q4bits_1, 4)), + __lsx_vshuf_b(values128, values128, __lsx_vandi_b(q4bits_1, 0xf))); + const __m256i q4b_2 = lasx_insertf128(__lsx_vshuf_b(values128, values128, __lsx_vsrli_b(q4bits_2, 4)), + __lsx_vshuf_b(values128, values128, __lsx_vandi_b(q4bits_2, 0xf))); const __m256i p16_1 = mul_add_epi8(q4b_1, q8b_1); const __m256i p16_2 = mul_add_epi8(q4b_2, q8b_2); const int16_t ls1 = ((x[ibl].scales_l[ib/2] & 0xf) | ((sh << 4) & 0x30)) - 32; const int16_t ls2 = ((x[ibl].scales_l[ib/2] >> 4) | ((sh << 2) & 0x30)) - 32; sh >>= 4; - __m256i tmp5, tmp6; - tmp1 = __lasx_xvreplgr2vr_h(ls1); - tmp5 = __lasx_xvmulwev_w_h(p16_1, tmp1); - tmp6 = __lasx_xvmulwod_w_h(p16_1, tmp1); - const __m256i p_1 = __lasx_xvadd_w(tmp5, tmp6); - tmp1 = __lasx_xvreplgr2vr_h(ls2); - tmp5 = __lasx_xvmulwev_w_h(p16_2, tmp1); - tmp6 = __lasx_xvmulwod_w_h(p16_2, tmp1); - const __m256i p_2 = __lasx_xvadd_w(tmp5, tmp6); + const __m256i p_1 = lasx_madd_h(p16_1, __lasx_xvreplgr2vr_h(ls1)); + const __m256i p_2 = lasx_madd_h(p16_2, __lasx_xvreplgr2vr_h(ls2)); sumi1 = __lasx_xvadd_w(p_1, sumi1); sumi2 = __lasx_xvadd_w(p_2, sumi2); } @@ -10861,6 +12927,56 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * restrict s, size_t bs, const void * } *s = hsum_float_8(accum); +#elif defined(__VXE__) || defined(__VXE2__) + const int8x16_t v_k = vec_xl(0, kvalues_iq4nl); + const uint8x16_t v_m = vec_splat_u8(0x0F); + + float sumf = 0; + + for (int ibl = 0; ibl < nb; ++ibl) { + const uint8_t * GGML_RESTRICT q4 = x[ibl].qs; + const int8_t * GGML_RESTRICT q8 = y[ibl].qs; + + uint16_t h = x[ibl].scales_h; + + int sumi1 = 0, sumi2 = 0; + for (int ib = 0; ib < QK_K/64; ++ib) { + const uint8x16_t v_x0 = vec_xl(0 , q4); + const uint8x16_t v_x1 = vec_xl(QK4_NL/2, q4); + q4 += 32; + + int8x16_t v_x0l = (int8x16_t)vec_and(v_x0, v_m); + int8x16_t v_x0h = (int8x16_t)vec_sr(v_x0, 4); + int8x16_t v_x1l = (int8x16_t)vec_and(v_x1, v_m); + int8x16_t v_x1h = (int8x16_t)vec_sr(v_x1, 4); + + v_x0l = vec_perm(v_k, v_k, (uchar8x16_t)v_x0l); + v_x0h = vec_perm(v_k, v_k, (uchar8x16_t)v_x0h); + v_x1l = vec_perm(v_k, v_k, (uchar8x16_t)v_x1l); + v_x1h = vec_perm(v_k, v_k, (uchar8x16_t)v_x1h); + + const int8x16_t v_y0 = vec_xl( 0, q8); + const int8x16_t v_y1 = vec_xl(16, q8); + const int8x16_t v_y2 = vec_xl(32, q8); + const int8x16_t v_y3 = vec_xl(48, q8); + q8 += 64; + + int32x4_t vsumi0 = ggml_vec_dot(ggml_vec_dot(vec_splats(0), v_x0l, v_y0), v_x0h, v_y1); + int32x4_t vsumi1 = ggml_vec_dot(ggml_vec_dot(vec_splats(0), v_x1l, v_y2), v_x1h, v_y3); + + int ls1 = ((x[ibl].scales_l[ib] & 0xF) | ((h << 4) & 0x30)) - 32; + int ls2 = ((x[ibl].scales_l[ib] >> 4) | ((h << 2) & 0x30)) - 32; + + h >>= 4; + + sumi1 += (vsumi0[0] + vsumi0[1] + vsumi0[2] + vsumi0[3]) * ls1; + sumi2 += (vsumi1[0] + vsumi1[1] + vsumi1[2] + vsumi1[3]) * ls2; + } + + sumf += GGML_FP16_TO_FP32(x[ibl].d) * y[ibl].d * (sumi1 + sumi2); + } + + *s = sumf; #else float sumf = 0; @@ -10899,12 +13015,12 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * restrict s, size_t bs, const void * // ============================ 4-bit non-linear quants -void quantize_row_iq4_nl(const float * restrict x, void * restrict y, int64_t k) { +void quantize_row_iq4_nl(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k) { assert(k % QK4_NL == 0); quantize_row_iq4_nl_ref(x, y, k); } -void quantize_row_iq4_xs(const float * restrict x, void * restrict y, int64_t k) { +void quantize_row_iq4_xs(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k) { assert(k % QK_K == 0); quantize_iq4_xs(x, y, 1, k, NULL); } diff --git a/ggml/src/ggml-cpu/ggml-cpu.c b/ggml/src/ggml-cpu/ggml-cpu.c index fdb430a43..504003287 100644 --- a/ggml/src/ggml-cpu/ggml-cpu.c +++ b/ggml/src/ggml-cpu/ggml-cpu.c @@ -7,10 +7,12 @@ #include "ggml-cpu-impl.h" #include "ggml-cpu.h" #include "ggml-impl.h" -#include "ggml-quants.h" #include "ggml-cpu-quants.h" #include "ggml-threading.h" -#include "amx/amx.h" +#include "unary-ops.h" +#include "binary-ops.h" +#include "vec.h" +#include "ops.h" #include "ggml.h" #if defined(_MSC_VER) || defined(__MINGW32__) @@ -83,30 +85,6 @@ #define UNUSED GGML_UNUSED #define SWAP(x, y, T) do { T SWAP = x; (x) = y; (y) = SWAP; } while (0) -#if defined(GGML_USE_ACCELERATE) -#include -#endif - -// floating point type used to accumulate sums -typedef double ggml_float; - -#define GGML_GELU_FP16 -#define GGML_GELU_QUICK_FP16 - -#define GGML_SOFT_MAX_UNROLL 4 -#define GGML_VEC_DOT_UNROLL 2 -#define GGML_VEC_MAD_UNROLL 32 - -// -// global data -// - -// precomputed gelu table for f16 (128 KB) -static ggml_fp16_t ggml_table_gelu_f16[1 << 16]; - -// precomputed quick gelu table for f16 (128 KB) -static ggml_fp16_t ggml_table_gelu_quick_f16[1 << 16]; - #if defined(__ARM_ARCH) struct ggml_arm_arch_features_type { int has_neon; @@ -114,7 +92,8 @@ struct ggml_arm_arch_features_type { int has_i8mm; int has_sve; int sve_cnt; -} ggml_arm_arch_features = {-1, -1, -1, -1, 0}; + int has_sme; +} ggml_arm_arch_features = {-1, -1, -1, -1, 0, -1}; #endif @@ -229,27 +208,6 @@ typedef pthread_t ggml_thread_t; #include #endif -// -// cache line -// - -#if defined(__cpp_lib_hardware_interference_size) -#define CACHE_LINE_SIZE hardware_destructive_interference_size -#else -#if defined(__POWER9_VECTOR__) -#define CACHE_LINE_SIZE 128 -#else -#define CACHE_LINE_SIZE 64 -#endif -#endif - -static const size_t CACHE_LINE_SIZE_F32 = CACHE_LINE_SIZE/sizeof(float); - - -static void ggml_vec_dot_f32(int n, float * restrict s, size_t bs, const float * restrict x, size_t bx, const float * restrict y, size_t by, int nrc); -static void ggml_vec_dot_f16(int n, float * restrict s, size_t bs, ggml_fp16_t * restrict x, size_t bx, ggml_fp16_t * restrict y, size_t by, int nrc); -static void ggml_vec_dot_bf16(int n, float * restrict s, size_t bs, ggml_bf16_t * restrict x, size_t bx, ggml_bf16_t * restrict y, size_t by, int nrc); - static const struct ggml_type_traits_cpu type_traits_cpu[GGML_TYPE_COUNT] = { [GGML_TYPE_F32] = { .vec_dot = (ggml_vec_dot_t) ggml_vec_dot_f32, @@ -421,806 +379,6 @@ const struct ggml_type_traits_cpu * ggml_get_type_traits_cpu(enum ggml_type type return &type_traits_cpu[type]; } -// -// simd mappings -// - -// we define a common set of C macros which map to specific intrinsics based on the current architecture -// we then implement the fundamental computation operations below using only these macros -// adding support for new architectures requires to define the corresponding SIMD macros -// -// GGML_F32_STEP / GGML_F16_STEP -// number of elements to process in a single step -// -// GGML_F32_EPR / GGML_F16_EPR -// number of elements to fit in a single register -// - -#if defined(__ARM_NEON) && defined(__ARM_FEATURE_FMA) - -#define GGML_SIMD - -// F32 NEON - -#define GGML_F32_STEP 16 -#define GGML_F32_EPR 4 - -#define GGML_F32x4 float32x4_t -#define GGML_F32x4_ZERO vdupq_n_f32(0.0f) -#define GGML_F32x4_SET1(x) vdupq_n_f32(x) -#define GGML_F32x4_LOAD vld1q_f32 -#define GGML_F32x4_STORE vst1q_f32 -#define GGML_F32x4_FMA(a, b, c) vfmaq_f32(a, b, c) -#define GGML_F32x4_ADD vaddq_f32 -#define GGML_F32x4_MUL vmulq_f32 -#define GGML_F32x4_REDUCE_ONE(x) vaddvq_f32(x) -#define GGML_F32x4_REDUCE(res, x) \ -{ \ - int offset = GGML_F32_ARR >> 1; \ - for (int i = 0; i < offset; ++i) { \ - (x)[i] = vaddq_f32((x)[i], (x)[offset+i]); \ - } \ - offset >>= 1; \ - for (int i = 0; i < offset; ++i) { \ - (x)[i] = vaddq_f32((x)[i], (x)[offset+i]); \ - } \ - offset >>= 1; \ - for (int i = 0; i < offset; ++i) { \ - (x)[i] = vaddq_f32((x)[i], (x)[offset+i]); \ - } \ - (res) = (ggml_float) GGML_F32x4_REDUCE_ONE((x)[0]); \ -} - -#define GGML_F32_VEC GGML_F32x4 -#define GGML_F32_VEC_ZERO GGML_F32x4_ZERO -#define GGML_F32_VEC_SET1 GGML_F32x4_SET1 -#define GGML_F32_VEC_LOAD GGML_F32x4_LOAD -#define GGML_F32_VEC_STORE GGML_F32x4_STORE -#define GGML_F32_VEC_FMA GGML_F32x4_FMA -#define GGML_F32_VEC_ADD GGML_F32x4_ADD -#define GGML_F32_VEC_MUL GGML_F32x4_MUL -#define GGML_F32_VEC_REDUCE GGML_F32x4_REDUCE - -// F16 NEON - -#if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) - #define GGML_F16_STEP 32 - #define GGML_F16_EPR 8 - - #define GGML_F16x8 float16x8_t - #define GGML_F16x8_ZERO vdupq_n_f16(0.0f) - #define GGML_F16x8_SET1(x) vdupq_n_f16(x) - #define GGML_F16x8_LOAD(x) vld1q_f16((const ggml_fp16_internal_t *)(x)) - #define GGML_F16x8_STORE vst1q_f16 - #define GGML_F16x8_FMA(a, b, c) vfmaq_f16(a, b, c) - #define GGML_F16x8_ADD vaddq_f16 - #define GGML_F16x8_MUL vmulq_f16 - #define GGML_F16x8_REDUCE(res, x) \ - do { \ - int offset = GGML_F16_ARR >> 1; \ - for (int i = 0; i < offset; ++i) { \ - (x)[i] = vaddq_f16((x)[i], (x)[offset+i]); \ - } \ - offset >>= 1; \ - for (int i = 0; i < offset; ++i) { \ - (x)[i] = vaddq_f16((x)[i], (x)[offset+i]); \ - } \ - offset >>= 1; \ - for (int i = 0; i < offset; ++i) { \ - (x)[i] = vaddq_f16((x)[i], (x)[offset+i]); \ - } \ - const float32x4_t t0 = vcvt_f32_f16(vget_low_f16 ((x)[0])); \ - const float32x4_t t1 = vcvt_f32_f16(vget_high_f16((x)[0])); \ - (res) = (ggml_float) vaddvq_f32(vaddq_f32(t0, t1)); \ - } while (0) - - #define GGML_F16_VEC GGML_F16x8 - #define GGML_F16_VEC_ZERO GGML_F16x8_ZERO - #define GGML_F16_VEC_SET1 GGML_F16x8_SET1 - #define GGML_F16_VEC_LOAD(p, i) GGML_F16x8_LOAD(p) - #define GGML_F16_VEC_STORE(p, r, i) GGML_F16x8_STORE((ggml_fp16_internal_t *)(p), (r)[i]) - #define GGML_F16_VEC_FMA GGML_F16x8_FMA - #define GGML_F16_VEC_ADD GGML_F16x8_ADD - #define GGML_F16_VEC_MUL GGML_F16x8_MUL - #define GGML_F16_VEC_REDUCE GGML_F16x8_REDUCE -#else - // if FP16 vector arithmetic is not supported, we use FP32 instead - // and take advantage of the vcvt_ functions to convert to/from FP16 - - #define GGML_F16_STEP 16 - #define GGML_F16_EPR 4 - - #define GGML_F32Cx4 float32x4_t - #define GGML_F32Cx4_ZERO vdupq_n_f32(0.0f) - #define GGML_F32Cx4_SET1(x) vdupq_n_f32(x) - #define GGML_F32Cx4_LOAD(x) vcvt_f32_f16(vld1_f16((const ggml_fp16_internal_t *)(x))) - #define GGML_F32Cx4_STORE(x, y) vst1_f16(x, vcvt_f16_f32(y)) - #define GGML_F32Cx4_FMA(a, b, c) vfmaq_f32(a, b, c) - #define GGML_F32Cx4_ADD vaddq_f32 - #define GGML_F32Cx4_MUL vmulq_f32 - #define GGML_F32Cx4_REDUCE GGML_F32x4_REDUCE - - #define GGML_F16_VEC GGML_F32Cx4 - #define GGML_F16_VEC_ZERO GGML_F32Cx4_ZERO - #define GGML_F16_VEC_SET1 GGML_F32Cx4_SET1 - #define GGML_F16_VEC_LOAD(p, i) GGML_F32Cx4_LOAD(p) - #define GGML_F16_VEC_STORE(p, r, i) GGML_F32Cx4_STORE((ggml_fp16_internal_t *)(p), r[i]) - #define GGML_F16_VEC_FMA GGML_F32Cx4_FMA - #define GGML_F16_VEC_ADD GGML_F32Cx4_ADD - #define GGML_F16_VEC_MUL GGML_F32Cx4_MUL - #define GGML_F16_VEC_REDUCE GGML_F32Cx4_REDUCE -#endif - -#elif defined(__AVX512F__) - -#define GGML_SIMD - -// F32 AVX512 - -#define GGML_F32_STEP 64 -#define GGML_F32_EPR 16 - -#define GGML_F32x16 __m512 -#define GGML_F32x16_ZERO _mm512_setzero_ps() -#define GGML_F32x16_SET1(x) _mm512_set1_ps(x) -#define GGML_F32x16_LOAD _mm512_loadu_ps -#define GGML_F32x16_STORE _mm512_storeu_ps -// _mm512_fmadd_ps is defined in AVX512F so no guard is required -#define GGML_F32x16_FMA(a, b, c) _mm512_fmadd_ps(b, c, a) -#define GGML_F32x16_ADD _mm512_add_ps -#define GGML_F32x16_MUL _mm512_mul_ps -#define GGML_F32x16_REDUCE(res, x) \ -do { \ - int offset = GGML_F32_ARR >> 1; \ - for (int i = 0; i < offset; ++i) { \ - x[i] = _mm512_add_ps(x[i], x[offset+i]); \ - } \ - offset >>= 1; \ - for (int i = 0; i < offset; ++i) { \ - x[i] = _mm512_add_ps(x[i], x[offset+i]); \ - } \ - offset >>= 1; \ - for (int i = 0; i < offset; ++i) { \ - x[i] = _mm512_add_ps(x[i], x[offset+i]); \ - } \ - res = (ggml_float) _mm512_reduce_add_ps(x[0]); \ -} while (0) - -// TODO: is this optimal ? - -#define GGML_F32_VEC GGML_F32x16 -#define GGML_F32_VEC_ZERO GGML_F32x16_ZERO -#define GGML_F32_VEC_SET1 GGML_F32x16_SET1 -#define GGML_F32_VEC_LOAD GGML_F32x16_LOAD -#define GGML_F32_VEC_STORE GGML_F32x16_STORE -#define GGML_F32_VEC_FMA GGML_F32x16_FMA -#define GGML_F32_VEC_ADD GGML_F32x16_ADD -#define GGML_F32_VEC_MUL GGML_F32x16_MUL -#define GGML_F32_VEC_REDUCE GGML_F32x16_REDUCE - -// F16 AVX512 - -// F16 AVX - -#define GGML_F16_STEP 64 -#define GGML_F16_EPR 16 - -// AVX512 has FP16 extension (AVX512_FP16) but I don't have it on my machine so I use FP32 instead - -#define GGML_F32Cx16 __m512 -#define GGML_F32Cx16_ZERO _mm512_setzero_ps() -#define GGML_F32Cx16_SET1(x) _mm512_set1_ps(x) - -// unlike _mm256_cvt intrinsics that require F16C, _mm512_cvt is defined in AVX512F -// so F16C guard isn't required -#define GGML_F32Cx16_LOAD(x) _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)(x))) -#define GGML_F32Cx16_STORE(x, y) _mm256_storeu_si256((__m256i *)(x), _mm512_cvtps_ph(y, 0)) - -#define GGML_F32Cx16_FMA(a, b, c) _mm512_fmadd_ps(b, c, a) -#define GGML_F32Cx16_ADD _mm512_add_ps -#define GGML_F32Cx16_MUL _mm512_mul_ps -#define GGML_F32Cx16_REDUCE(res, x) \ -do { \ - int offset = GGML_F32_ARR >> 1; \ - for (int i = 0; i < offset; ++i) { \ - x[i] = _mm512_add_ps(x[i], x[offset+i]); \ - } \ - offset >>= 1; \ - for (int i = 0; i < offset; ++i) { \ - x[i] = _mm512_add_ps(x[i], x[offset+i]); \ - } \ - offset >>= 1; \ - for (int i = 0; i < offset; ++i) { \ - x[i] = _mm512_add_ps(x[i], x[offset+i]); \ - } \ - res = (ggml_float) _mm512_reduce_add_ps(x[0]); \ -} while (0) - -#define GGML_F16_VEC GGML_F32Cx16 -#define GGML_F16_VEC_ZERO GGML_F32Cx16_ZERO -#define GGML_F16_VEC_SET1 GGML_F32Cx16_SET1 -#define GGML_F16_VEC_LOAD(p, i) GGML_F32Cx16_LOAD(p) -#define GGML_F16_VEC_STORE(p, r, i) GGML_F32Cx16_STORE(p, r[i]) -#define GGML_F16_VEC_FMA GGML_F32Cx16_FMA -#define GGML_F16_VEC_ADD GGML_F32Cx16_ADD -#define GGML_F16_VEC_MUL GGML_F32Cx16_MUL - -#define GGML_F16_VEC_REDUCE GGML_F32Cx16_REDUCE -#elif defined(__AVX__) - -#define GGML_SIMD - -// F32 AVX - -#define GGML_F32_STEP 32 -#define GGML_F32_EPR 8 - -#define GGML_F32x8 __m256 -#define GGML_F32x8_ZERO _mm256_setzero_ps() -#define GGML_F32x8_SET1(x) _mm256_set1_ps(x) -#define GGML_F32x8_LOAD _mm256_loadu_ps -#define GGML_F32x8_STORE _mm256_storeu_ps -#if defined(__FMA__) - #define GGML_F32x8_FMA(a, b, c) _mm256_fmadd_ps(b, c, a) -#else - #define GGML_F32x8_FMA(a, b, c) _mm256_add_ps(_mm256_mul_ps(b, c), a) -#endif -#define GGML_F32x8_ADD _mm256_add_ps -#define GGML_F32x8_MUL _mm256_mul_ps -#define GGML_F32x8_REDUCE(res, x) \ -do { \ - int offset = GGML_F32_ARR >> 1; \ - for (int i = 0; i < offset; ++i) { \ - x[i] = _mm256_add_ps(x[i], x[offset+i]); \ - } \ - offset >>= 1; \ - for (int i = 0; i < offset; ++i) { \ - x[i] = _mm256_add_ps(x[i], x[offset+i]); \ - } \ - offset >>= 1; \ - for (int i = 0; i < offset; ++i) { \ - x[i] = _mm256_add_ps(x[i], x[offset+i]); \ - } \ - const __m128 t0 = _mm_add_ps(_mm256_castps256_ps128(x[0]), \ - _mm256_extractf128_ps(x[0], 1)); \ - const __m128 t1 = _mm_hadd_ps(t0, t0); \ - res = (ggml_float) _mm_cvtss_f32(_mm_hadd_ps(t1, t1)); \ -} while (0) -// TODO: is this optimal ? - -#define GGML_F32_VEC GGML_F32x8 -#define GGML_F32_VEC_ZERO GGML_F32x8_ZERO -#define GGML_F32_VEC_SET1 GGML_F32x8_SET1 -#define GGML_F32_VEC_LOAD GGML_F32x8_LOAD -#define GGML_F32_VEC_STORE GGML_F32x8_STORE -#define GGML_F32_VEC_FMA GGML_F32x8_FMA -#define GGML_F32_VEC_ADD GGML_F32x8_ADD -#define GGML_F32_VEC_MUL GGML_F32x8_MUL -#define GGML_F32_VEC_REDUCE GGML_F32x8_REDUCE - -// F16 AVX - -#define GGML_F16_STEP 32 -#define GGML_F16_EPR 8 - -// F16 arithmetic is not supported by AVX, so we use F32 instead - -#define GGML_F32Cx8 __m256 -#define GGML_F32Cx8_ZERO _mm256_setzero_ps() -#define GGML_F32Cx8_SET1(x) _mm256_set1_ps(x) - -#if defined(__F16C__) -// the _mm256_cvt intrinsics require F16C -#define GGML_F32Cx8_LOAD(x) _mm256_cvtph_ps(_mm_loadu_si128((const __m128i *)(x))) -#define GGML_F32Cx8_STORE(x, y) _mm_storeu_si128((__m128i *)(x), _mm256_cvtps_ph(y, 0)) -#else -static inline __m256 __avx_f32cx8_load(const ggml_fp16_t * x) { - float tmp[8]; - - for (int i = 0; i < 8; i++) { - tmp[i] = GGML_FP16_TO_FP32(x[i]); - } - - return _mm256_loadu_ps(tmp); -} -static inline void __avx_f32cx8_store(ggml_fp16_t *x, __m256 y) { - float arr[8]; - - _mm256_storeu_ps(arr, y); - - for (int i = 0; i < 8; i++) - x[i] = GGML_FP32_TO_FP16(arr[i]); -} -#define GGML_F32Cx8_LOAD(x) __avx_f32cx8_load(x) -#define GGML_F32Cx8_STORE(x, y) __avx_f32cx8_store(x, y) -#endif - -#define GGML_F32Cx8_FMA GGML_F32x8_FMA -#define GGML_F32Cx8_ADD _mm256_add_ps -#define GGML_F32Cx8_MUL _mm256_mul_ps -#define GGML_F32Cx8_REDUCE GGML_F32x8_REDUCE - -#define GGML_F16_VEC GGML_F32Cx8 -#define GGML_F16_VEC_ZERO GGML_F32Cx8_ZERO -#define GGML_F16_VEC_SET1 GGML_F32Cx8_SET1 -#define GGML_F16_VEC_LOAD(p, i) GGML_F32Cx8_LOAD(p) -#define GGML_F16_VEC_STORE(p, r, i) GGML_F32Cx8_STORE(p, r[i]) -#define GGML_F16_VEC_FMA GGML_F32Cx8_FMA -#define GGML_F16_VEC_ADD GGML_F32Cx8_ADD -#define GGML_F16_VEC_MUL GGML_F32Cx8_MUL -#define GGML_F16_VEC_REDUCE GGML_F32Cx8_REDUCE - -#elif defined(__POWER9_VECTOR__) - -#define GGML_SIMD - -// F32 POWER9 - -#define GGML_F32_STEP 32 -#define GGML_F32_EPR 4 - -#define GGML_F32x4 vector float -#define GGML_F32x4_ZERO 0.0f -#define GGML_F32x4_SET1 vec_splats -#define GGML_F32x4_LOAD(p) vec_xl(0, p) -#define GGML_F32x4_STORE(p, r) vec_xst(r, 0, p) -#define GGML_F32x4_FMA(a, b, c) vec_madd(b, c, a) -#define GGML_F32x4_ADD vec_add -#define GGML_F32x4_MUL vec_mul -#define GGML_F32x4_REDUCE(res, x) \ -{ \ - int offset = GGML_F32_ARR >> 1; \ - for (int i = 0; i < offset; ++i) { \ - x[i] = vec_add(x[i], x[offset+i]); \ - } \ - offset >>= 1; \ - for (int i = 0; i < offset; ++i) { \ - x[i] = vec_add(x[i], x[offset+i]); \ - } \ - offset >>= 1; \ - for (int i = 0; i < offset; ++i) { \ - x[i] = vec_add(x[i], x[offset+i]); \ - } \ - res = vec_extract(x[0], 0) + \ - vec_extract(x[0], 1) + \ - vec_extract(x[0], 2) + \ - vec_extract(x[0], 3); \ -} - -#define GGML_F32_VEC GGML_F32x4 -#define GGML_F32_VEC_ZERO GGML_F32x4_ZERO -#define GGML_F32_VEC_SET1 GGML_F32x4_SET1 -#define GGML_F32_VEC_LOAD GGML_F32x4_LOAD -#define GGML_F32_VEC_STORE GGML_F32x4_STORE -#define GGML_F32_VEC_FMA GGML_F32x4_FMA -#define GGML_F32_VEC_ADD GGML_F32x4_ADD -#define GGML_F32_VEC_MUL GGML_F32x4_MUL -#define GGML_F32_VEC_REDUCE GGML_F32x4_REDUCE - -// F16 POWER9 -#define GGML_F16_STEP GGML_F32_STEP -#define GGML_F16_EPR GGML_F32_EPR -#define GGML_F16_VEC GGML_F32x4 -#define GGML_F16_VEC_ZERO GGML_F32x4_ZERO -#define GGML_F16_VEC_SET1 GGML_F32x4_SET1 -#define GGML_F16_VEC_FMA GGML_F32x4_FMA -#define GGML_F16_VEC_ADD GGML_F32x4_ADD -#define GGML_F16_VEC_MUL GGML_F32x4_MUL -#define GGML_F16_VEC_REDUCE GGML_F32x4_REDUCE -// Use vec_xl, not vec_ld, in case the load address is not aligned. -#define GGML_F16_VEC_LOAD(p, i) (i & 0x1) ? \ - vec_extract_fp32_from_shorth(vec_xl(0, p - GGML_F16_EPR)) : \ - vec_extract_fp32_from_shortl(vec_xl(0, p)) -#define GGML_ENDIAN_BYTE(i) ((unsigned char *)&(uint16_t){1})[i] -#define GGML_F16_VEC_STORE(p, r, i) \ - if (i & 0x1) \ - vec_xst(vec_pack_to_short_fp32(r[i - GGML_ENDIAN_BYTE(1)], \ - r[i - GGML_ENDIAN_BYTE(0)]), \ - 0, p - GGML_F16_EPR) - -#elif defined(__wasm_simd128__) - -#define GGML_SIMD - -// F32 WASM - -#define GGML_F32_STEP 16 -#define GGML_F32_EPR 4 - -#define GGML_F32x4 v128_t -#define GGML_F32x4_ZERO wasm_f32x4_splat(0.0f) -#define GGML_F32x4_SET1(x) wasm_f32x4_splat(x) -#define GGML_F32x4_LOAD wasm_v128_load -#define GGML_F32x4_STORE wasm_v128_store -#define GGML_F32x4_FMA(a, b, c) wasm_f32x4_add(wasm_f32x4_mul(b, c), a) -#define GGML_F32x4_ADD wasm_f32x4_add -#define GGML_F32x4_MUL wasm_f32x4_mul -#define GGML_F32x4_REDUCE(res, x) \ -{ \ - int offset = GGML_F32_ARR >> 1; \ - for (int i = 0; i < offset; ++i) { \ - x[i] = wasm_f32x4_add(x[i], x[offset+i]); \ - } \ - offset >>= 1; \ - for (int i = 0; i < offset; ++i) { \ - x[i] = wasm_f32x4_add(x[i], x[offset+i]); \ - } \ - offset >>= 1; \ - for (int i = 0; i < offset; ++i) { \ - x[i] = wasm_f32x4_add(x[i], x[offset+i]); \ - } \ - res = wasm_f32x4_extract_lane(x[0], 0) + \ - wasm_f32x4_extract_lane(x[0], 1) + \ - wasm_f32x4_extract_lane(x[0], 2) + \ - wasm_f32x4_extract_lane(x[0], 3); \ -} - -#define GGML_F32_VEC GGML_F32x4 -#define GGML_F32_VEC_ZERO GGML_F32x4_ZERO -#define GGML_F32_VEC_SET1 GGML_F32x4_SET1 -#define GGML_F32_VEC_LOAD GGML_F32x4_LOAD -#define GGML_F32_VEC_STORE GGML_F32x4_STORE -#define GGML_F32_VEC_FMA GGML_F32x4_FMA -#define GGML_F32_VEC_ADD GGML_F32x4_ADD -#define GGML_F32_VEC_MUL GGML_F32x4_MUL -#define GGML_F32_VEC_REDUCE GGML_F32x4_REDUCE - -// F16 WASM - -#define GGML_F16_STEP 16 -#define GGML_F16_EPR 4 - -inline static v128_t __wasm_f16x4_load(const ggml_fp16_t * p) { - float tmp[4]; - - tmp[0] = GGML_FP16_TO_FP32(p[0]); - tmp[1] = GGML_FP16_TO_FP32(p[1]); - tmp[2] = GGML_FP16_TO_FP32(p[2]); - tmp[3] = GGML_FP16_TO_FP32(p[3]); - - return wasm_v128_load(tmp); -} - -inline static void __wasm_f16x4_store(ggml_fp16_t * p, v128_t x) { - float tmp[4]; - - wasm_v128_store(tmp, x); - - p[0] = GGML_FP32_TO_FP16(tmp[0]); - p[1] = GGML_FP32_TO_FP16(tmp[1]); - p[2] = GGML_FP32_TO_FP16(tmp[2]); - p[3] = GGML_FP32_TO_FP16(tmp[3]); -} - -#define GGML_F16x4 v128_t -#define GGML_F16x4_ZERO wasm_f32x4_splat(0.0f) -#define GGML_F16x4_SET1(x) wasm_f32x4_splat(x) -#define GGML_F16x4_LOAD(x) __wasm_f16x4_load(x) -#define GGML_F16x4_STORE(x, y) __wasm_f16x4_store(x, y) -#define GGML_F16x4_FMA GGML_F32x4_FMA -#define GGML_F16x4_ADD wasm_f32x4_add -#define GGML_F16x4_MUL wasm_f32x4_mul -#define GGML_F16x4_REDUCE(res, x) \ -{ \ - int offset = GGML_F16_ARR >> 1; \ - for (int i = 0; i < offset; ++i) { \ - x[i] = wasm_f32x4_add(x[i], x[offset+i]); \ - } \ - offset >>= 1; \ - for (int i = 0; i < offset; ++i) { \ - x[i] = wasm_f32x4_add(x[i], x[offset+i]); \ - } \ - offset >>= 1; \ - for (int i = 0; i < offset; ++i) { \ - x[i] = wasm_f32x4_add(x[i], x[offset+i]); \ - } \ - res = wasm_f32x4_extract_lane(x[0], 0) + \ - wasm_f32x4_extract_lane(x[0], 1) + \ - wasm_f32x4_extract_lane(x[0], 2) + \ - wasm_f32x4_extract_lane(x[0], 3); \ -} - -#define GGML_F16_VEC GGML_F16x4 -#define GGML_F16_VEC_ZERO GGML_F16x4_ZERO -#define GGML_F16_VEC_SET1 GGML_F16x4_SET1 -#define GGML_F16_VEC_LOAD(p, i) GGML_F16x4_LOAD(p) -#define GGML_F16_VEC_STORE(p, r, i) GGML_F16x4_STORE(p, r[i]) -#define GGML_F16_VEC_FMA GGML_F16x4_FMA -#define GGML_F16_VEC_ADD GGML_F16x4_ADD -#define GGML_F16_VEC_MUL GGML_F16x4_MUL -#define GGML_F16_VEC_REDUCE GGML_F16x4_REDUCE - -#elif defined(__SSE3__) - -#define GGML_SIMD - -// F32 SSE - -#define GGML_F32_STEP 32 -#define GGML_F32_EPR 4 - -#define GGML_F32x4 __m128 -#define GGML_F32x4_ZERO _mm_setzero_ps() -#define GGML_F32x4_SET1(x) _mm_set1_ps(x) -#define GGML_F32x4_LOAD _mm_loadu_ps -#define GGML_F32x4_STORE _mm_storeu_ps -#if defined(__FMA__) - // TODO: Does this work? - #define GGML_F32x4_FMA(a, b, c) _mm_fmadd_ps(b, c, a) -#else - #define GGML_F32x4_FMA(a, b, c) _mm_add_ps(_mm_mul_ps(b, c), a) -#endif -#define GGML_F32x4_ADD _mm_add_ps -#define GGML_F32x4_MUL _mm_mul_ps -#define GGML_F32x4_REDUCE(res, x) \ -{ \ - int offset = GGML_F32_ARR >> 1; \ - for (int i = 0; i < offset; ++i) { \ - x[i] = _mm_add_ps(x[i], x[offset+i]); \ - } \ - offset >>= 1; \ - for (int i = 0; i < offset; ++i) { \ - x[i] = _mm_add_ps(x[i], x[offset+i]); \ - } \ - offset >>= 1; \ - for (int i = 0; i < offset; ++i) { \ - x[i] = _mm_add_ps(x[i], x[offset+i]); \ - } \ - const __m128 t0 = _mm_hadd_ps(x[0], x[0]); \ - res = (ggml_float) _mm_cvtss_f32(_mm_hadd_ps(t0, t0)); \ -} -// TODO: is this optimal ? - -#define GGML_F32_VEC GGML_F32x4 -#define GGML_F32_VEC_ZERO GGML_F32x4_ZERO -#define GGML_F32_VEC_SET1 GGML_F32x4_SET1 -#define GGML_F32_VEC_LOAD GGML_F32x4_LOAD -#define GGML_F32_VEC_STORE GGML_F32x4_STORE -#define GGML_F32_VEC_FMA GGML_F32x4_FMA -#define GGML_F32_VEC_ADD GGML_F32x4_ADD -#define GGML_F32_VEC_MUL GGML_F32x4_MUL -#define GGML_F32_VEC_REDUCE GGML_F32x4_REDUCE - -// F16 SSE - -#define GGML_F16_STEP 32 -#define GGML_F16_EPR 4 - -static inline __m128 __sse_f16x4_load(const ggml_fp16_t * x) { - float tmp[4]; - - tmp[0] = GGML_FP16_TO_FP32(x[0]); - tmp[1] = GGML_FP16_TO_FP32(x[1]); - tmp[2] = GGML_FP16_TO_FP32(x[2]); - tmp[3] = GGML_FP16_TO_FP32(x[3]); - - return _mm_loadu_ps(tmp); -} - -static inline void __sse_f16x4_store(ggml_fp16_t * x, __m128 y) { - float arr[4]; - - _mm_storeu_ps(arr, y); - - x[0] = GGML_FP32_TO_FP16(arr[0]); - x[1] = GGML_FP32_TO_FP16(arr[1]); - x[2] = GGML_FP32_TO_FP16(arr[2]); - x[3] = GGML_FP32_TO_FP16(arr[3]); -} - -#define GGML_F32Cx4 __m128 -#define GGML_F32Cx4_ZERO _mm_setzero_ps() -#define GGML_F32Cx4_SET1(x) _mm_set1_ps(x) -#define GGML_F32Cx4_LOAD(x) __sse_f16x4_load(x) -#define GGML_F32Cx4_STORE(x, y) __sse_f16x4_store(x, y) -#define GGML_F32Cx4_FMA GGML_F32x4_FMA -#define GGML_F32Cx4_ADD _mm_add_ps -#define GGML_F32Cx4_MUL _mm_mul_ps -#define GGML_F32Cx4_REDUCE GGML_F32x4_REDUCE - -#define GGML_F16_VEC GGML_F32Cx4 -#define GGML_F16_VEC_ZERO GGML_F32Cx4_ZERO -#define GGML_F16_VEC_SET1 GGML_F32Cx4_SET1 -#define GGML_F16_VEC_LOAD(p, i) GGML_F32Cx4_LOAD(p) -#define GGML_F16_VEC_STORE(p, r, i) GGML_F32Cx4_STORE(p, r[i]) -#define GGML_F16_VEC_FMA GGML_F32Cx4_FMA -#define GGML_F16_VEC_ADD GGML_F32Cx4_ADD -#define GGML_F16_VEC_MUL GGML_F32Cx4_MUL -#define GGML_F16_VEC_REDUCE GGML_F32Cx4_REDUCE - -#elif defined(__loongarch_asx) - -#define GGML_SIMD - -// F32 LASX -#define GGML_F32_STEP 32 -#define GGML_F32_EPR 8 - -#define GGML_F32x8 __m256 -#define GGML_F32x8_ZERO (__m256)__lasx_xvldi(0) -#define GGML_F32x8_SET1(x) (__m256)__lasx_xvreplfr2vr_s((x)) -#define GGML_F32x8_LOAD(x) (__m256)__lasx_xvld((x), 0) -#define GGML_F32x8_STORE(x,y) __lasx_xvst((y), (x), 0) -#define GGML_F32x8_FMA(a, b, c) __lasx_xvfmadd_s(b, c, a) -#define GGML_F32x8_ADD __lasx_xvfadd_s -#define GGML_F32x8_MUL __lasx_xvfmul_s -#define GGML_F32x8_REDUCE(res, x) \ -do { \ - int offset = GGML_F32_ARR >> 1; \ - for (int i = 0; i < offset; ++i) { \ - x[i] = __lasx_xvfadd_s(x[i], x[offset+i]); \ - } \ - offset >>= 1; \ - for (int i = 0; i < offset; ++i) { \ - x[i] = __lasx_xvfadd_s(x[i], x[offset+i]); \ - } \ - offset >>= 1; \ - for (int i = 0; i < offset; ++i) { \ - x[i] = __lasx_xvfadd_s(x[i], x[offset+i]); \ - } \ - float *tmp_p = (float *)&x[0]; \ - res = tmp_p[0] + tmp_p[1] + tmp_p[2] + tmp_p[3] + tmp_p[4] + tmp_p[5] + tmp_p[6] + tmp_p[7]; \ -} while (0) -// TODO: is this optimal ? - -#define GGML_F32_VEC GGML_F32x8 -#define GGML_F32_VEC_ZERO GGML_F32x8_ZERO -#define GGML_F32_VEC_SET1 GGML_F32x8_SET1 -#define GGML_F32_VEC_LOAD GGML_F32x8_LOAD -#define GGML_F32_VEC_STORE GGML_F32x8_STORE -#define GGML_F32_VEC_FMA GGML_F32x8_FMA -#define GGML_F32_VEC_ADD GGML_F32x8_ADD -#define GGML_F32_VEC_MUL GGML_F32x8_MUL -#define GGML_F32_VEC_REDUCE GGML_F32x8_REDUCE - -// F16 LASX - -#define GGML_F16_STEP 32 -#define GGML_F16_EPR 8 - -// F16 arithmetic is not supported by LASX, so we use F32 instead - -#define GGML_F32Cx8 __m256 -#define GGML_F32Cx8_ZERO (__m256)__lasx_xvldi(0) -#define GGML_F32Cx8_SET1(x) (__m256)__lasx_xvreplgr2vr_w((x)) - -static inline __m256 __lasx_f32cx8_load(const ggml_fp16_t * x) { - __m256i a; - memcpy(&a, x, sizeof(ggml_fp16_t) * 8); - a = __lasx_xvpermi_d(a, 0 | (1 << 4)); - return __lasx_xvfcvtl_s_h(a); -} - -static inline void __lasx_f32cx8_store(ggml_fp16_t * x, __m256 y) { - __m256i a = __lasx_xvfcvt_h_s(y, y); - a = __lasx_xvpermi_d(a, 0 | (2 << 2)); - memcpy(x, &a, sizeof(ggml_fp16_t) * 8); -} -#define GGML_F32Cx8_LOAD(x) __lasx_f32cx8_load(x) -#define GGML_F32Cx8_STORE(x, y) __lasx_f32cx8_store(x, y) - -#define GGML_F32Cx8_FMA GGML_F32x8_FMA -#define GGML_F32Cx8_ADD __lasx_xvfadd_s -#define GGML_F32Cx8_MUL __lasx_xvfmul_s -#define GGML_F32Cx8_REDUCE GGML_F32x8_REDUCE - -#define GGML_F16_VEC GGML_F32Cx8 -#define GGML_F16_VEC_ZERO GGML_F32Cx8_ZERO -#define GGML_F16_VEC_SET1 GGML_F32Cx8_SET1 -#define GGML_F16_VEC_LOAD(p, i) GGML_F32Cx8_LOAD(p) -#define GGML_F16_VEC_STORE(p, r, i) GGML_F32Cx8_STORE(p, r[i]) -#define GGML_F16_VEC_FMA GGML_F32Cx8_FMA -#define GGML_F16_VEC_ADD GGML_F32Cx8_ADD -#define GGML_F16_VEC_MUL GGML_F32Cx8_MUL -#define GGML_F16_VEC_REDUCE GGML_F32Cx8_REDUCE - -#elif defined(__loongarch_sx) - -#define GGML_SIMD - -// F32 LSX - -#define GGML_F32_STEP 32 -#define GGML_F32_EPR 4 - -#define GGML_F32x4 __m128 -#define GGML_F32x4_ZERO __lsx_vldi(0) -#define GGML_F32x4_SET1(x) __lsx_vinsgr2vr_w(__lsx_vldi(0),(x), 0) -#define GGML_F32x4_LOAD(x) __lsx_vld((x), 0) -#define GGML_F32x4_STORE((x),(y)) __lsx_vst((y), (x), 0) -#define GGML_F32x4_FMA(a, b, c) __lsx_vfmadd_s(b, c, a) -#define GGML_F32x4_ADD __lsx_vfadd_s -#define GGML_F32x4_MUL __lsx_vfmul_s -#define GGML_F32x4_REDUCE(res, x) \ -{ \ - int offset = GGML_F32_ARR >> 1; \ - for (int i = 0; i < offset; ++i) { \ - x[i] = __lsx_vfadd_s(x[i], x[offset + i]); \ - } \ - offset >>= 1; \ - for (int i = 0; i < offset; ++i) { \ - x[i] = __lsx_vfadd_s(x[i], x[offset + i]); \ - } \ - offset >>= 1; \ - for (int i = 0; i < offset; ++i) { \ - x[i] = __lsx_vfadd_s(x[i], x[offset + i]); \ - } \ - __m128i tmp = __lsx_vsrli_d((__m128i) x[0], 32); \ - tmp = (__m128i) __lsx_vfadd_s((__m128) tmp, x[0]); \ - tmp = __lsx_vpickev_w(__lsx_vldi(0), tmp); \ - const __m128 t0 = __lsx_vshuf4i_w(tmp, 0x88); \ - tmp = __lsx_vsrli_d((__m128i) t0, 32); \ - tmp = (__m128i) __lsx_vfadd_s((__m128) tmp, t0); \ - tmp = __lsx_vpickev_w(__lsx_vldi(0), tmp); \ - res = (ggml_float) __lsx_vpickve2gr_w(__lsx_vshuf4i_w(tmp, 0x88), 0); \ -} - -#define GGML_F32_VEC GGML_F32x4 -#define GGML_F32_VEC_ZERO GGML_F32x4_ZERO -#define GGML_F32_VEC_SET1 GGML_F32x4_SET1 -#define GGML_F32_VEC_LOAD GGML_F32x4_LOAD -#define GGML_F32_VEC_STORE GGML_F32x4_STORE -#define GGML_F32_VEC_FMA GGML_F32x4_FMA -#define GGML_F32_VEC_ADD GGML_F32x4_ADD -#define GGML_F32_VEC_MUL GGML_F32x4_MUL -#define GGML_F32_VEC_REDUCE GGML_F32x4_REDUCE - -// F16 LSX - -#define GGML_F16_STEP 32 -#define GGML_F16_EPR 4 - -static inline __m128 __lsx_f16x4_load(const ggml_fp16_t * x) { - float tmp[4]; - - tmp[0] = GGML_FP16_TO_FP32(x[0]); - tmp[1] = GGML_FP16_TO_FP32(x[1]); - tmp[2] = GGML_FP16_TO_FP32(x[2]); - tmp[3] = GGML_FP16_TO_FP32(x[3]); - - return __lsx_vld(tmp, 0); -} - -static inline void __lsx_f16x4_store(ggml_fp16_t * x, __m128 y) { - float arr[4]; - - __lsx_vst(y, arr, 0); - - x[0] = GGML_FP32_TO_FP16(arr[0]); - x[1] = GGML_FP32_TO_FP16(arr[1]); - x[2] = GGML_FP32_TO_FP16(arr[2]); - x[3] = GGML_FP32_TO_FP16(arr[3]); -} - -#define GGML_F32Cx4 __m128 -#define GGML_F32Cx4_ZERO __lsx_vldi(0) -#define GGML_F32Cx4_SET1(x) __lsx_vinsgr2vr_w(__lsx_vldi(0),(x), 0) -#define GGML_F32Cx4_LOAD(x) __lsx_f16x4_load(x) -#define GGML_F32Cx4_STORE(x, y) __lsx_f16x4_store(x, y) -#define GGML_F32Cx4_FMA GGML_F32x4_FMA -#define GGML_F32Cx4_ADD __lsx_vfadd_s -#define GGML_F32Cx4_MUL __lsx_vfmul_s -#define GGML_F32Cx4_REDUCE GGML_F32x4_REDUCE - -#define GGML_F16_VEC GGML_F32Cx4 -#define GGML_F16_VEC_ZERO GGML_F32Cx4_ZERO -#define GGML_F16_VEC_SET1 GGML_F32Cx4_SET1 -#define GGML_F16_VEC_LOAD(p, i) GGML_F32Cx4_LOAD(p) -#define GGML_F16_VEC_STORE(p, r, i) GGML_F32Cx4_STORE(p, r[i]) -#define GGML_F16_VEC_FMA GGML_F32Cx4_FMA -#define GGML_F16_VEC_ADD GGML_F32Cx4_ADD -#define GGML_F16_VEC_MUL GGML_F32Cx4_MUL -#define GGML_F16_VEC_REDUCE GGML_F32Cx4_REDUCE - -#endif - -// GGML_F32_ARR / GGML_F16_ARR -// number of registers to use per step -#ifdef GGML_SIMD -#define GGML_F32_ARR (GGML_F32_STEP/GGML_F32_EPR) -#define GGML_F16_ARR (GGML_F16_STEP/GGML_F16_EPR) -#endif - // // Threading defs // @@ -1291,7 +449,7 @@ struct ggml_threadpool { atomic_int n_graph; // incremented when there is work to be done (i.e each graph) atomic_int GGML_CACHE_ALIGN n_barrier; atomic_int GGML_CACHE_ALIGN n_barrier_passed; - atomic_int current_chunk; // currently processing chunk during Mat_Mul, shared between all the threads. + atomic_int GGML_CACHE_ALIGN current_chunk; // currently processing chunk during Mat_Mul, shared between all the threads. // these are atomic as an annotation for thread-sanitizer atomic_bool stop; // Used for stopping the threadpool altogether @@ -1320,864 +478,6 @@ struct ggml_compute_state { int ith; }; -// -// fundamental operations -// - -inline static void ggml_vec_set_i8(const int n, int8_t * x, const int8_t v) { for (int i = 0; i < n; ++i) x[i] = v; } -inline static void ggml_vec_set_i16(const int n, int16_t * x, const int16_t v) { for (int i = 0; i < n; ++i) x[i] = v; } - -inline static void ggml_vec_set_i32(const int n, int32_t * x, const int32_t v) { for (int i = 0; i < n; ++i) x[i] = v; } -inline static void ggml_vec_cpy_i32(const int n, int32_t * y, const int32_t * x) { for (int i = 0; i < n; ++i) y[i] = x[i]; } - -inline static void ggml_vec_set_f16(const int n, ggml_fp16_t * x, const int32_t v) { for (int i = 0; i < n; ++i) x[i] = v; } -inline static void ggml_vec_set_bf16(const int n, ggml_bf16_t * x, const ggml_bf16_t v) { for (int i = 0; i < n; ++i) x[i] = v; } -inline static void ggml_vec_add_f32 (const int n, float * z, const float * x, const float * y) { for (int i = 0; i < n; ++i) z[i] = x[i] + y[i]; } -inline static void ggml_vec_add1_f32(const int n, float * z, const float * x, const float v) { for (int i = 0; i < n; ++i) z[i] = x[i] + v; } -inline static void ggml_vec_acc_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] += x[i]; } -inline static void ggml_vec_acc1_f32(const int n, float * y, const float v) { for (int i = 0; i < n; ++i) y[i] += v; } -inline static void ggml_vec_sub_f32 (const int n, float * z, const float * x, const float * y) { for (int i = 0; i < n; ++i) z[i] = x[i] - y[i]; } -inline static void ggml_vec_set_f32 (const int n, float * x, const float v) { for (int i = 0; i < n; ++i) x[i] = v; } -inline static void ggml_vec_cpy_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = x[i]; } -inline static void ggml_vec_neg_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = -x[i]; } -inline static void ggml_vec_mul_f32 (const int n, float * z, const float * x, const float * y) { for (int i = 0; i < n; ++i) z[i] = x[i]*y[i]; } -inline static void ggml_vec_div_f32 (const int n, float * z, const float * x, const float * y) { for (int i = 0; i < n; ++i) z[i] = x[i]/y[i]; } - -static void ggml_vec_dot_f32(int n, float * restrict s, size_t bs, const float * restrict x, size_t bx, const float * restrict y, size_t by, int nrc) { - assert(nrc == 1); - UNUSED(nrc); - UNUSED(bx); - UNUSED(by); - UNUSED(bs); - -#if defined(GGML_SIMD) - float sumf = 0.0f; - const int np = (n & ~(GGML_F32_STEP - 1)); - - GGML_F32_VEC sum[GGML_F32_ARR] = { GGML_F32_VEC_ZERO }; - - GGML_F32_VEC ax[GGML_F32_ARR]; - GGML_F32_VEC ay[GGML_F32_ARR]; - - for (int i = 0; i < np; i += GGML_F32_STEP) { - for (int j = 0; j < GGML_F32_ARR; j++) { - ax[j] = GGML_F32_VEC_LOAD(x + i + j*GGML_F32_EPR); - ay[j] = GGML_F32_VEC_LOAD(y + i + j*GGML_F32_EPR); - - sum[j] = GGML_F32_VEC_FMA(sum[j], ax[j], ay[j]); - } - } - - // reduce sum0..sum3 to sum0 - GGML_F32_VEC_REDUCE(sumf, sum); - - // leftovers - for (int i = np; i < n; ++i) { - sumf += x[i]*y[i]; - } -#else - // scalar - ggml_float sumf = 0.0; - for (int i = 0; i < n; ++i) { - sumf += (ggml_float)(x[i]*y[i]); - } -#endif - - *s = sumf; -} - -static void ggml_vec_dot_bf16(int n, float * restrict s, size_t bs, ggml_bf16_t * restrict x, size_t bx, ggml_bf16_t * restrict y, size_t by, int nrc) { - assert(nrc == 1); - UNUSED(nrc); - UNUSED(bx); - UNUSED(by); - UNUSED(bs); - int i = 0; - ggml_float sumf = 0; - -#if defined(__AVX512BF16__) - __m512 c1 = _mm512_setzero_ps(); - __m512 c2 = _mm512_setzero_ps(); - for (; i + 64 <= n; i += 64) { - c1 = _mm512_dpbf16_ps(c1, m512bh(_mm512_loadu_si512((x + i))), - m512bh(_mm512_loadu_si512((y + i)))); - c2 = _mm512_dpbf16_ps(c2, m512bh(_mm512_loadu_si512((x + i + 32))), - m512bh(_mm512_loadu_si512((y + i + 32)))); - } - sumf += (ggml_float)_mm512_reduce_add_ps(c1); - sumf += (ggml_float)_mm512_reduce_add_ps(c2); - -#elif defined(__AVX512F__) -#define LOAD(p) _mm512_castsi512_ps(_mm512_slli_epi32(_mm512_cvtepu16_epi32(_mm256_loadu_si256((const __m256i *)(p))), 16)) - __m512 c1 = _mm512_setzero_ps(); - __m512 c2 = _mm512_setzero_ps(); - for (; i + 32 <= n; i += 32) { - c1 = _mm512_add_ps(_mm512_mul_ps(LOAD(x + i), LOAD(y + i)), c1); - c2 = _mm512_add_ps(_mm512_mul_ps(LOAD(x + i + 16), LOAD(y + i + 16)), c2); - } - sumf += (ggml_float)_mm512_reduce_add_ps(c1); - sumf += (ggml_float)_mm512_reduce_add_ps(c2); - -#undef LOAD -#elif defined(__AVX2__) || defined(__AVX__) -#if defined(__AVX2__) -#define LOAD(p) _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_cvtepu16_epi32(_mm_loadu_si128((const __m128i *)(p))), 16)) -#else -#define LOAD(p) _mm256_castsi256_ps(_mm256_insertf128_si256(_mm256_castsi128_si256(_mm_slli_epi32(_mm_cvtepu16_epi32(_mm_loadu_si128((const __m128i *)(p))), 16)), (_mm_slli_epi32(_mm_cvtepu16_epi32(_mm_bsrli_si128(_mm_loadu_si128((const __m128i *)(p)), 8)), 16)), 1)) -#endif - __m256 c1 = _mm256_setzero_ps(); - __m256 c2 = _mm256_setzero_ps(); - __m256 c3 = _mm256_setzero_ps(); - __m256 c4 = _mm256_setzero_ps(); - for (; i + 32 <= n; i += 32) { - c1 = _mm256_add_ps(_mm256_mul_ps(LOAD(x + i), LOAD(y + i)), c1); - c2 = _mm256_add_ps(_mm256_mul_ps(LOAD(x + i + 8), LOAD(y + i + 8)), c2); - c3 = _mm256_add_ps(_mm256_mul_ps(LOAD(x + i + 16), LOAD(y + i + 16)), c3); - c4 = _mm256_add_ps(_mm256_mul_ps(LOAD(x + i + 24), LOAD(y + i + 24)), c4); - } - __m128 g; - c1 = _mm256_add_ps(_mm256_add_ps(c1, c3), - _mm256_add_ps(c2, c4)); - g = _mm_add_ps(_mm256_extractf128_ps(c1, 1), - _mm256_castps256_ps128(c1)); - g = _mm_add_ps(g, _mm_movehl_ps(g, g)); - g = _mm_add_ss(g, _mm_movehdup_ps(g)); - sumf += (ggml_float)_mm_cvtss_f32(g); - -#undef LOAD -#endif - - for (; i < n; ++i) { - sumf += (ggml_float)(GGML_BF16_TO_FP32(x[i]) * - GGML_BF16_TO_FP32(y[i])); - } - *s = sumf; -} - -static void ggml_vec_dot_f16(int n, float * restrict s, size_t bs, ggml_fp16_t * restrict x, size_t bx, ggml_fp16_t * restrict y, size_t by, int nrc) { - assert(nrc == 1); - UNUSED(nrc); - UNUSED(bx); - UNUSED(by); - UNUSED(bs); - - ggml_float sumf = 0.0; - -#if defined(GGML_SIMD) - const int np = (n & ~(GGML_F16_STEP - 1)); - - GGML_F16_VEC sum[GGML_F16_ARR] = { GGML_F16_VEC_ZERO }; - - GGML_F16_VEC ax[GGML_F16_ARR]; - GGML_F16_VEC ay[GGML_F16_ARR]; - - for (int i = 0; i < np; i += GGML_F16_STEP) { - for (int j = 0; j < GGML_F16_ARR; j++) { - ax[j] = GGML_F16_VEC_LOAD(x + i + j*GGML_F16_EPR, j); - ay[j] = GGML_F16_VEC_LOAD(y + i + j*GGML_F16_EPR, j); - - sum[j] = GGML_F16_VEC_FMA(sum[j], ax[j], ay[j]); - } - } - - // reduce sum0..sum3 to sum0 - GGML_F16_VEC_REDUCE(sumf, sum); - - // leftovers - for (int i = np; i < n; ++i) { - sumf += (ggml_float)(GGML_FP16_TO_FP32(x[i])*GGML_FP16_TO_FP32(y[i])); - } -#else - for (int i = 0; i < n; ++i) { - sumf += (ggml_float)(GGML_FP16_TO_FP32(x[i])*GGML_FP16_TO_FP32(y[i])); - } -#endif - - *s = sumf; -} - -// compute GGML_VEC_DOT_UNROLL dot products at once -// xs - x row stride in bytes -inline static void ggml_vec_dot_f16_unroll(const int n, const int xs, float * restrict s, void * restrict xv, ggml_fp16_t * restrict y) { - ggml_float sumf[GGML_VEC_DOT_UNROLL] = { 0.0 }; - - ggml_fp16_t * restrict x[GGML_VEC_DOT_UNROLL]; - - for (int i = 0; i < GGML_VEC_DOT_UNROLL; ++i) { - x[i] = (ggml_fp16_t *) ((char *) xv + i*xs); - } - -#if defined(GGML_SIMD) - const int np = (n & ~(GGML_F16_STEP - 1)); - - GGML_F16_VEC sum[GGML_VEC_DOT_UNROLL][GGML_F16_ARR] = { { GGML_F16_VEC_ZERO } }; - - GGML_F16_VEC ax[GGML_F16_ARR]; - GGML_F16_VEC ay[GGML_F16_ARR]; - - for (int i = 0; i < np; i += GGML_F16_STEP) { - for (int j = 0; j < GGML_F16_ARR; j++) { - ay[j] = GGML_F16_VEC_LOAD(y + i + j*GGML_F16_EPR, j); - - for (int k = 0; k < GGML_VEC_DOT_UNROLL; ++k) { - ax[j] = GGML_F16_VEC_LOAD(x[k] + i + j*GGML_F16_EPR, j); - - sum[k][j] = GGML_F16_VEC_FMA(sum[k][j], ax[j], ay[j]); - } - } - } - - // reduce sum0..sum3 to sum0 - for (int k = 0; k < GGML_VEC_DOT_UNROLL; ++k) { - GGML_F16_VEC_REDUCE(sumf[k], sum[k]); - } - - // leftovers - for (int i = np; i < n; ++i) { - for (int j = 0; j < GGML_VEC_DOT_UNROLL; ++j) { - sumf[j] += (ggml_float)(GGML_FP16_TO_FP32(x[j][i])*GGML_FP16_TO_FP32(y[i])); - } - } -#else - for (int i = 0; i < n; ++i) { - for (int j = 0; j < GGML_VEC_DOT_UNROLL; ++j) { - sumf[j] += (ggml_float)(GGML_FP16_TO_FP32(x[j][i])*GGML_FP16_TO_FP32(y[i])); - } - } -#endif - - for (int i = 0; i < GGML_VEC_DOT_UNROLL; ++i) { - s[i] = sumf[i]; - } -} - -inline static void ggml_vec_mad_f32(const int n, float * restrict y, const float * restrict x, const float v) { -#if defined(GGML_SIMD) - const int np = (n & ~(GGML_F32_STEP - 1)); - - GGML_F32_VEC vx = GGML_F32_VEC_SET1(v); - - GGML_F32_VEC ax[GGML_F32_ARR]; - GGML_F32_VEC ay[GGML_F32_ARR]; - - for (int i = 0; i < np; i += GGML_F32_STEP) { - for (int j = 0; j < GGML_F32_ARR; j++) { - ax[j] = GGML_F32_VEC_LOAD(x + i + j*GGML_F32_EPR); - ay[j] = GGML_F32_VEC_LOAD(y + i + j*GGML_F32_EPR); - ay[j] = GGML_F32_VEC_FMA(ay[j], ax[j], vx); - - GGML_F32_VEC_STORE(y + i + j*GGML_F32_EPR, ay[j]); - } - } - - // leftovers - for (int i = np; i < n; ++i) { - y[i] += x[i]*v; - } -#else - // scalar - for (int i = 0; i < n; ++i) { - y[i] += x[i]*v; - } -#endif -} - -inline static void ggml_vec_mad_f16(const int n, ggml_fp16_t * restrict y, const ggml_fp16_t * restrict x, const float v) { -#if defined(GGML_SIMD) - const int np = (n & ~(GGML_F16_STEP - 1)); - - GGML_F16_VEC vx = GGML_F16_VEC_SET1(v); - - GGML_F16_VEC ax[GGML_F16_ARR]; - GGML_F16_VEC ay[GGML_F16_ARR]; - - for (int i = 0; i < np; i += GGML_F16_STEP) { - for (int j = 0; j < GGML_F16_ARR; j++) { - ax[j] = GGML_F16_VEC_LOAD(x + i + j*GGML_F16_EPR, j); - ay[j] = GGML_F16_VEC_LOAD(y + i + j*GGML_F16_EPR, j); - ay[j] = GGML_F16_VEC_FMA(ay[j], ax[j], vx); - - GGML_F16_VEC_STORE(y + i + j*GGML_F16_EPR, ay, j); - } - } - - // leftovers - for (int i = np; i < n; ++i) { - y[i] = GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(y[i]) + GGML_FP16_TO_FP32(x[i])*v); - } -#else - // scalar - for (int i = 0; i < n; ++i) { - y[i] = GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(y[i]) + GGML_FP16_TO_FP32(x[i])*v); - } -#endif -} - -// xs and vs are byte strides of x and v -inline static void ggml_vec_mad_f32_unroll(const int n, const int xs, const int vs, float * restrict y, const float * restrict xv, const float * restrict vv) { - - const float * restrict x[GGML_VEC_MAD_UNROLL]; - const float * restrict v[GGML_VEC_MAD_UNROLL]; - - for (int i = 0; i < GGML_VEC_MAD_UNROLL; ++i) { - x[i] = (const float *) ((const char *) xv + i*xs); - v[i] = (const float *) ((const char *) vv + i*vs); - } - -#if defined(GGML_SIMD) - const int np = (n & ~(GGML_F32_STEP - 1)); - - GGML_F32_VEC vx[GGML_VEC_MAD_UNROLL]; - - for (int k = 0; k < GGML_VEC_MAD_UNROLL; ++k) { - vx[k] = GGML_F32_VEC_SET1(v[k][0]); - } - - GGML_F32_VEC ax[GGML_VEC_MAD_UNROLL][GGML_F32_ARR]; - GGML_F32_VEC ay[GGML_F32_ARR]; - - for (int i = 0; i < np; i += GGML_F32_STEP) { - for (int j = 0; j < GGML_F32_ARR; j++) { - ay[j] = GGML_F32_VEC_LOAD(y + i + j*GGML_F32_EPR); - - for (int k = 0; k < GGML_VEC_MAD_UNROLL; ++k) { - ax[k][j] = GGML_F32_VEC_LOAD(x[k] + i + j*GGML_F32_EPR); - ay[j] = GGML_F32_VEC_FMA(ay[j], ax[k][j], vx[k]); - } - - GGML_F32_VEC_STORE(y + i + j*GGML_F32_EPR, ay[j]); - } - } - - // leftovers - for (int k = 0; k < GGML_VEC_MAD_UNROLL; ++k) { - for (int i = np; i < n; ++i) { - y[i] += x[k][i]*v[k][0]; - } - } -#else - // scalar - for (int k = 0; k < GGML_VEC_MAD_UNROLL; ++k) { - for (int i = 0; i < n; ++i) { - y[i] += x[k][i]*v[k][0]; - } - } -#endif -} - -//inline static void ggml_vec_scale_f32(const int n, float * y, const float v) { for (int i = 0; i < n; ++i) y[i] *= v; } -inline static void ggml_vec_scale_f32(const int n, float * y, const float v) { -#if defined(GGML_USE_ACCELERATE) - vDSP_vsmul(y, 1, &v, y, 1, n); -#elif defined(GGML_SIMD) - const int np = (n & ~(GGML_F32_STEP - 1)); - - GGML_F32_VEC vx = GGML_F32_VEC_SET1(v); - - GGML_F32_VEC ay[GGML_F32_ARR]; - - for (int i = 0; i < np; i += GGML_F32_STEP) { - for (int j = 0; j < GGML_F32_ARR; j++) { - ay[j] = GGML_F32_VEC_LOAD(y + i + j*GGML_F32_EPR); - ay[j] = GGML_F32_VEC_MUL(ay[j], vx); - - GGML_F32_VEC_STORE(y + i + j*GGML_F32_EPR, ay[j]); - } - } - - // leftovers - for (int i = np; i < n; ++i) { - y[i] *= v; - } -#else - // scalar - for (int i = 0; i < n; ++i) { - y[i] *= v; - } -#endif -} - -inline static void ggml_vec_scale_f16(const int n, ggml_fp16_t * y, const float v) { -#if defined(GGML_SIMD) - const int np = (n & ~(GGML_F16_STEP - 1)); - - GGML_F16_VEC vx = GGML_F16_VEC_SET1(v); - - GGML_F16_VEC ay[GGML_F16_ARR]; - - for (int i = 0; i < np; i += GGML_F16_STEP) { - for (int j = 0; j < GGML_F16_ARR; j++) { - ay[j] = GGML_F16_VEC_LOAD(y + i + j*GGML_F16_EPR, j); - ay[j] = GGML_F16_VEC_MUL(ay[j], vx); - - GGML_F16_VEC_STORE(y + i + j*GGML_F16_EPR, ay, j); - } - } - - // leftovers - for (int i = np; i < n; ++i) { - y[i] = GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(y[i])*v); - } -#else - // scalar - for (int i = 0; i < n; ++i) { - y[i] = GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(y[i])*v); - } -#endif -} - -inline static void ggml_vec_norm_f32 (const int n, float * s, const float * x) { ggml_vec_dot_f32(n, s, 0, x, 0, x, 0, 1); *s = sqrtf(*s); } -inline static void ggml_vec_sqr_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = x[i]*x[i]; } -inline static void ggml_vec_sqrt_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = sqrtf(x[i]); } -inline static void ggml_vec_log_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = logf(x[i]); } -inline static void ggml_vec_sin_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = sinf(x[i]); } -inline static void ggml_vec_cos_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = cosf(x[i]); } -inline static void ggml_vec_abs_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = fabsf(x[i]); } -inline static void ggml_vec_sgn_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? 1.f : ((x[i] < 0.f) ? -1.f : 0.f); } -inline static void ggml_vec_step_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? 1.f : 0.f; } -inline static void ggml_vec_tanh_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = tanhf(x[i]); } -inline static void ggml_vec_elu_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? x[i] : expm1f(x[i]); } -inline static void ggml_vec_relu_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? x[i] : 0.f; } -inline static void ggml_vec_leaky_relu_f32 (const int n, float * y, const float * x, const float ns) { for (int i = 0; i < n; ++i) y[i] = ((x[i] > 0.f) ? x[i] : 0.f) + ns * ((x[i] < 0.0f) ? x[i] : 0.f); } -inline static void ggml_vec_sigmoid_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = 1.f / (1.f + expf(-x[i])); } -// TODO: optimize performance -inline static void ggml_vec_hardswish_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = x[i] * fminf(1.0f, fmaxf(0.0f, (x[i] + 3.0f) / 6.0f)); } -inline static void ggml_vec_hardsigmoid_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = fminf(1.0f, fmaxf(0.0f, (x[i] + 3.0f) / 6.0f)); } -inline static void ggml_vec_exp_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = expf(x[i]); } - -static const float GELU_COEF_A = 0.044715f; -static const float GELU_QUICK_COEF = -1.702f; -static const float SQRT_2_OVER_PI = 0.79788456080286535587989211986876f; - -inline static float ggml_gelu_f32(float x) { - return 0.5f*x*(1.0f + tanhf(SQRT_2_OVER_PI*x*(1.0f + GELU_COEF_A*x*x))); -} - -inline static void ggml_vec_gelu_f16(const int n, ggml_fp16_t * y, const ggml_fp16_t * x) { - const uint16_t * i16 = (const uint16_t *) x; - for (int i = 0; i < n; ++i) { - y[i] = ggml_table_gelu_f16[i16[i]]; - } -} - -#ifdef GGML_GELU_FP16 -inline static void ggml_vec_gelu_f32(const int n, float * y, const float * x) { - uint16_t t; - for (int i = 0; i < n; ++i) { - if (x[i] <= -10.0f) { - y[i] = 0.0f; - } else if (x[i] >= 10.0f) { - y[i] = x[i]; - } else { - ggml_fp16_t fp16 = GGML_FP32_TO_FP16(x[i]); - memcpy(&t, &fp16, sizeof(uint16_t)); - y[i] = GGML_FP16_TO_FP32(ggml_table_gelu_f16[t]); - } - } -} -#else -inline static void ggml_vec_gelu_f32(const int n, float * y, const float * x) { - for (int i = 0; i < n; ++i) { - y[i] = ggml_gelu_f32(x[i]); - } -} -#endif - -inline static float ggml_gelu_quick_f32(float x) { - return x*(1.0f/(1.0f+expf(GELU_QUICK_COEF*x))); -} - -//inline static void ggml_vec_gelu_quick_f16(const int n, ggml_fp16_t * y, const ggml_fp16_t * x) { -// const uint16_t * i16 = (const uint16_t *) x; -// for (int i = 0; i < n; ++i) { -// y[i] = ggml_table_gelu_quick_f16[i16[i]]; -// } -//} - -#ifdef GGML_GELU_QUICK_FP16 -inline static void ggml_vec_gelu_quick_f32(const int n, float * y, const float * x) { - uint16_t t; - for (int i = 0; i < n; ++i) { - ggml_fp16_t fp16 = GGML_FP32_TO_FP16(x[i]); - memcpy(&t, &fp16, sizeof(uint16_t)); - y[i] = GGML_FP16_TO_FP32(ggml_table_gelu_quick_f16[t]); - } -} -#else -inline static void ggml_vec_gelu_quick_f32(const int n, float * y, const float * x) { - for (int i = 0; i < n; ++i) { - y[i] = ggml_gelu_quick_f32(x[i]); - } -} -#endif - -// Sigmoid Linear Unit (SiLU) function -inline static float ggml_silu_f32(float x) { - return x/(1.0f + expf(-x)); -} - -#if __FINITE_MATH_ONLY__ -#error "some routines in ggml.c require non-finite math arithmetics -- pass -fno-finite-math-only to the compiler to fix" -#error "ref: https://github.com/ggerganov/llama.cpp/pull/7154#issuecomment-2143844461" -#endif - -#if defined(__ARM_NEON) && defined(__aarch64__) - -// adapted from arm limited optimized routine -// the maximum error is 1.45358 plus 0.5 ulps -// numbers above 88.38 will flush to infinity -// numbers beneath -103.97 will flush to zero -inline static float32x4_t ggml_v_expf(float32x4_t x) { - const float32x4_t r = vdupq_n_f32(0x1.8p23f); - const float32x4_t z = vfmaq_f32(r, x, vdupq_n_f32(0x1.715476p+0f)); - const float32x4_t n = vsubq_f32(z, r); - const float32x4_t b = vfmsq_f32(vfmsq_f32(x, n, vdupq_n_f32(0x1.62e4p-1f)), n, - vdupq_n_f32(0x1.7f7d1cp-20f)); - const uint32x4_t e = vshlq_n_u32(vreinterpretq_u32_f32(z), 23); - const float32x4_t k = vreinterpretq_f32_u32(vaddq_u32(e, vreinterpretq_u32_f32(vdupq_n_f32(1)))); - const uint32x4_t c = vcagtq_f32(n, vdupq_n_f32(126)); - const float32x4_t u = vmulq_f32(b, b); - const float32x4_t j = vfmaq_f32( - vmulq_f32(vdupq_n_f32(0x1.ffffecp-1f), b), - vfmaq_f32(vfmaq_f32(vdupq_n_f32(0x1.fffdb6p-2f), vdupq_n_f32(0x1.555e66p-3f), b), - vfmaq_f32(vdupq_n_f32(0x1.573e2ep-5f), vdupq_n_f32(0x1.0e4020p-7f), b), u), u); - if (!vpaddd_u64(vreinterpretq_u64_u32(c))) - return vfmaq_f32(k, j, k); - const uint32x4_t d = vandq_u32(vclezq_f32(n), vdupq_n_u32(0x82000000)); - const float32x4_t s1 = vreinterpretq_f32_u32(vaddq_u32(d, vdupq_n_u32(0x7f000000))); - const float32x4_t s2 = vreinterpretq_f32_u32(vsubq_u32(e, d)); - return vbslq_f32(vcagtq_f32(n, vdupq_n_f32(192)), vmulq_f32(s1, s1), - vbslq_f32(c, vmulq_f32(vfmaq_f32(s2, s2, j), s1), vfmaq_f32(k, k, j))); -} - -// computes silu x/(1+exp(-x)) in single precision vector -inline static float32x4_t ggml_v_silu(float32x4_t x) { - const float32x4_t one = vdupq_n_f32(1.0f); - const float32x4_t zero = vdupq_n_f32(0.0f); - const float32x4_t neg_x = vsubq_f32(zero, x); - const float32x4_t exp_neg_x = ggml_v_expf(neg_x); - const float32x4_t one_plus_exp_neg_x = vaddq_f32(one, exp_neg_x); - return vdivq_f32(x, one_plus_exp_neg_x); -} - -#elif defined(__AVX512F__) && defined(__AVX512DQ__) - -// adapted from arm limited optimized routine -// the maximum error is 1.45358 plus 0.5 ulps -// numbers above 88.38 will flush to infinity -// numbers beneath -103.97 will flush to zero -inline static __m512 ggml_v_expf(__m512 x) { - const __m512 r = _mm512_set1_ps(0x1.8p23f); - const __m512 z = _mm512_fmadd_ps(x, _mm512_set1_ps(0x1.715476p+0f), r); - const __m512 n = _mm512_sub_ps(z, r); - const __m512 b = - _mm512_fnmadd_ps(n, _mm512_set1_ps(0x1.7f7d1cp-20f), - _mm512_fnmadd_ps(n, _mm512_set1_ps(0x1.62e4p-1f), x)); - const __mmask16 d = - _mm512_cmp_ps_mask(_mm512_abs_ps(n), _mm512_set1_ps(192), _CMP_GT_OQ); - const __m512 u = _mm512_mul_ps(b, b); - const __m512 j = _mm512_fmadd_ps( - _mm512_fmadd_ps(_mm512_fmadd_ps(_mm512_set1_ps(0x1.0e4020p-7f), b, - _mm512_set1_ps(0x1.573e2ep-5f)), - u, - _mm512_fmadd_ps(_mm512_set1_ps(0x1.555e66p-3f), b, - _mm512_set1_ps(0x1.fffdb6p-2f))), - u, - _mm512_fmadd_ps(_mm512_set1_ps(0x1.ffffecp-1f), b, _mm512_set1_ps(1.0F))); - const __m512 res = _mm512_scalef_ps(j, n); - if (_mm512_kortestz(d, d)) - return res; - const __m512 zero = _mm512_setzero_ps(); - const __m512 alt = _mm512_mask_blend_ps( - _mm512_cmp_ps_mask(n, zero, _CMP_LE_OQ), _mm512_set1_ps(INFINITY), zero); - return _mm512_mask_blend_ps(d, res, alt); -} - -// computes silu x/(1+exp(-x)) in single precision vector -inline static __m512 ggml_v_silu(__m512 x) { - const __m512 one = _mm512_set1_ps(1); - const __m512 zero = _mm512_setzero_ps(); - const __m512 neg_x = _mm512_sub_ps(zero, x); - const __m512 exp_neg_x = ggml_v_expf(neg_x); - const __m512 one_plus_exp_neg_x = _mm512_add_ps(one, exp_neg_x); - return _mm512_div_ps(x, one_plus_exp_neg_x); -} - -#elif defined(__AVX2__) && defined(__FMA__) - -// adapted from arm limited optimized routine -// the maximum error is 1.45358 plus 0.5 ulps -// numbers above 88.38 will flush to infinity -// numbers beneath -103.97 will flush to zero -inline static __m256 ggml_v_expf(__m256 x) { - const __m256 r = _mm256_set1_ps(0x1.8p23f); - const __m256 z = _mm256_fmadd_ps(x, _mm256_set1_ps(0x1.715476p+0f), r); - const __m256 n = _mm256_sub_ps(z, r); - const __m256 b = _mm256_fnmadd_ps(n, _mm256_set1_ps(0x1.7f7d1cp-20f), - _mm256_fnmadd_ps(n, _mm256_set1_ps(0x1.62e4p-1f), x)); - const __m256i e = _mm256_slli_epi32(_mm256_castps_si256(z), 23); - const __m256 k = _mm256_castsi256_ps( - _mm256_add_epi32(e, _mm256_castps_si256(_mm256_set1_ps(1)))); - const __m256i c = _mm256_castps_si256( - _mm256_cmp_ps(_mm256_andnot_ps(_mm256_set1_ps(-0.f), n), - _mm256_set1_ps(126), _CMP_GT_OQ)); - const __m256 u = _mm256_mul_ps(b, b); - const __m256 j = _mm256_fmadd_ps(_mm256_fmadd_ps(_mm256_fmadd_ps(_mm256_set1_ps(0x1.0e4020p-7f), b, - _mm256_set1_ps(0x1.573e2ep-5f)), u, - _mm256_fmadd_ps(_mm256_set1_ps(0x1.555e66p-3f), b, - _mm256_set1_ps(0x1.fffdb6p-2f))), - u, _mm256_mul_ps(_mm256_set1_ps(0x1.ffffecp-1f), b)); - if (!_mm256_movemask_ps(_mm256_castsi256_ps(c))) - return _mm256_fmadd_ps(j, k, k); - const __m256i g = _mm256_and_si256( - _mm256_castps_si256(_mm256_cmp_ps(n, _mm256_setzero_ps(), _CMP_LE_OQ)), - _mm256_set1_epi32(0x82000000u)); - const __m256 s1 = - _mm256_castsi256_ps(_mm256_add_epi32(g, _mm256_set1_epi32(0x7f000000u))); - const __m256 s2 = _mm256_castsi256_ps(_mm256_sub_epi32(e, g)); - const __m256i d = _mm256_castps_si256( - _mm256_cmp_ps(_mm256_andnot_ps(_mm256_set1_ps(-0.f), n), - _mm256_set1_ps(192), _CMP_GT_OQ)); - return _mm256_or_ps( - _mm256_and_ps(_mm256_castsi256_ps(d), _mm256_mul_ps(s1, s1)), - _mm256_andnot_ps( - _mm256_castsi256_ps(d), - _mm256_or_ps( - _mm256_and_ps(_mm256_castsi256_ps(c), - _mm256_mul_ps(_mm256_fmadd_ps(s2, j, s2), s1)), - _mm256_andnot_ps(_mm256_castsi256_ps(c), _mm256_fmadd_ps(k, j, k))))); -} - -// computes silu x/(1+exp(-x)) in single precision vector -inline static __m256 ggml_v_silu(__m256 x) { - const __m256 one = _mm256_set1_ps(1); - const __m256 zero = _mm256_setzero_ps(); - const __m256 neg_x = _mm256_sub_ps(zero, x); - const __m256 exp_neg_x = ggml_v_expf(neg_x); - const __m256 one_plus_exp_neg_x = _mm256_add_ps(one, exp_neg_x); - return _mm256_div_ps(x, one_plus_exp_neg_x); -} - -#elif defined(__SSE2__) // __AVX2__ / __ARM_NEON - -#if defined(__FMA__) -#define MADD128(x, y, z) _mm_fmadd_ps(x, y, z) -#define NMADD128(x, y, z) _mm_fnmadd_ps(x, y, z) -#else -#define MADD128(x, y, z) _mm_add_ps(_mm_mul_ps(x, y), z) -#define NMADD128(x, y, z) _mm_sub_ps(z, _mm_mul_ps(x, y)) -#endif - -// adapted from arm limited optimized routine -// the maximum error is 1.45358 plus 0.5 ulps -// numbers above 88.38 will flush to infinity -// numbers beneath -103.97 will flush to zero -inline static __m128 ggml_v_expf(__m128 x) { - const __m128 r = _mm_set1_ps(0x1.8p23f); - const __m128 z = MADD128(x, _mm_set1_ps(0x1.715476p+0f), r); - const __m128 n = _mm_sub_ps(z, r); - const __m128 b = - NMADD128(n, _mm_set1_ps(0x1.7f7d1cp-20f), NMADD128(n, _mm_set1_ps(0x1.62e4p-1f), x)); - const __m128i e = _mm_slli_epi32(_mm_castps_si128(z), 23); - const __m128 k = _mm_castsi128_ps(_mm_add_epi32(e, _mm_castps_si128(_mm_set1_ps(1)))); - const __m128i c = - _mm_castps_si128(_mm_cmpgt_ps(_mm_andnot_ps(_mm_set1_ps(-0.f), n), _mm_set1_ps(126))); - const __m128 u = _mm_mul_ps(b, b); - const __m128 j = - MADD128(MADD128(MADD128(_mm_set1_ps(0x1.0e4020p-7f), b, _mm_set1_ps(0x1.573e2ep-5f)), u, - MADD128(_mm_set1_ps(0x1.555e66p-3f), b, _mm_set1_ps(0x1.fffdb6p-2f))), - u, _mm_mul_ps(_mm_set1_ps(0x1.ffffecp-1f), b)); - if (!_mm_movemask_epi8(c)) - return MADD128(j, k, k); - const __m128i g = _mm_and_si128(_mm_castps_si128(_mm_cmple_ps(n, _mm_setzero_ps())), - _mm_set1_epi32(0x82000000u)); - const __m128 s1 = _mm_castsi128_ps(_mm_add_epi32(g, _mm_set1_epi32(0x7f000000u))); - const __m128 s2 = _mm_castsi128_ps(_mm_sub_epi32(e, g)); - const __m128i d = - _mm_castps_si128(_mm_cmpgt_ps(_mm_andnot_ps(_mm_set1_ps(-0.f), n), _mm_set1_ps(192))); - return _mm_or_ps( - _mm_and_ps(_mm_castsi128_ps(d), _mm_mul_ps(s1, s1)), - _mm_andnot_ps(_mm_castsi128_ps(d), - _mm_or_ps(_mm_and_ps(_mm_castsi128_ps(c), _mm_mul_ps(MADD128(s2, j, s2), s1)), - _mm_andnot_ps(_mm_castsi128_ps(c), MADD128(k, j, k))))); -} - -// computes silu x/(1+exp(-x)) in single precision vector -inline static __m128 ggml_v_silu(__m128 x) { - const __m128 one = _mm_set1_ps(1); - const __m128 zero = _mm_setzero_ps(); - const __m128 neg_x = _mm_sub_ps(zero, x); - const __m128 exp_neg_x = ggml_v_expf(neg_x); - const __m128 one_plus_exp_neg_x = _mm_add_ps(one, exp_neg_x); - return _mm_div_ps(x, one_plus_exp_neg_x); -} - -#endif // __ARM_NEON / __AVX2__ / __SSE2__ - -static void ggml_vec_silu_f32(const int n, float * y, const float * x) { - int i = 0; -#if defined(__AVX512F__) && defined(__AVX512DQ__) - for (; i + 15 < n; i += 16) { - _mm512_storeu_ps(y + i, ggml_v_silu(_mm512_loadu_ps(x + i))); - } -#elif defined(__AVX2__) && defined(__FMA__) - for (; i + 7 < n; i += 8) { - _mm256_storeu_ps(y + i, ggml_v_silu(_mm256_loadu_ps(x + i))); - } -#elif defined(__SSE2__) - for (; i + 3 < n; i += 4) { - _mm_storeu_ps(y + i, ggml_v_silu(_mm_loadu_ps(x + i))); - } -#elif defined(__ARM_NEON) && defined(__aarch64__) - for (; i + 3 < n; i += 4) { - vst1q_f32(y + i, ggml_v_silu(vld1q_f32(x + i))); - } -#endif - for (; i < n; ++i) { - y[i] = ggml_silu_f32(x[i]); - } -} - -static ggml_float ggml_vec_soft_max_f32(const int n, float * y, const float * x, float max) { - int i = 0; - ggml_float sum = 0; -#if defined(__AVX512F__) && defined(__AVX512DQ__) - for (; i + 15 < n; i += 16) { - __m512 val = ggml_v_expf(_mm512_sub_ps(_mm512_loadu_ps(x + i), - _mm512_set1_ps(max))); - _mm512_storeu_ps(y + i, val); - sum += (ggml_float)_mm512_reduce_add_ps(val); - } -#elif defined(__AVX2__) && defined(__FMA__) - for (; i + 7 < n; i += 8) { - __m256 val = ggml_v_expf(_mm256_sub_ps(_mm256_loadu_ps(x + i), - _mm256_set1_ps(max))); - _mm256_storeu_ps(y + i, val); - __m128 val2 = _mm_add_ps(_mm256_extractf128_ps(val, 1), - _mm256_castps256_ps128(val)); - val2 = _mm_add_ps(val2, _mm_movehl_ps(val2, val2)); - val2 = _mm_add_ss(val2, _mm_movehdup_ps(val2)); - sum += (ggml_float)_mm_cvtss_f32(val2); - } -#elif defined(__SSE2__) - for (; i + 3 < n; i += 4) { - __m128 val = ggml_v_expf(_mm_sub_ps(_mm_loadu_ps(x + i), - _mm_set1_ps(max))); - _mm_storeu_ps(y + i, val); -#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__) - val = _mm_add_ps(val, _mm_movehl_ps(val, val)); - val = _mm_add_ss(val, _mm_movehdup_ps(val)); -#else - __m128 tmp = _mm_shuffle_ps(val, val, _MM_SHUFFLE(2, 3, 0, 1)); - val = _mm_add_ps(val, tmp); - tmp = _mm_movehl_ps(tmp, val); - val = _mm_add_ss(val, tmp); -#endif - sum += (ggml_float)_mm_cvtss_f32(val); - } -#elif defined(__ARM_NEON) && defined(__aarch64__) - for (; i + 3 < n; i += 4) { - float32x4_t val = ggml_v_expf(vsubq_f32(vld1q_f32(x + i), - vdupq_n_f32(max))); - vst1q_f32(y + i, val); - sum += (ggml_float)vaddvq_f32(val); - } -#endif - for (; i < n; ++i) { - float val = expf(x[i] - max); - sum += (ggml_float)val; - y[i] = val; - } - return sum; -} - -static ggml_float ggml_vec_log_soft_max_f32(const int n, float * y, const float * x, float max) { - // log(soft_max) = log(soft_max_i / soft_max_sum) = log(soft_max_i) - log(soft_max_sum) = (logit_i - max) - log(soft_max_i) - - int i = 0; - ggml_float sum = 0; - for (; i < n; ++i) { - float val = x[i] - max; - y[i] = val; - sum += (ggml_float)expf(val); - } - return sum = (ggml_float)logf(sum); -} - -inline static float ggml_silu_backward_f32(float x, float dy) { - const float s = 1.0f/(1.0f + expf(-x)); - return dy*s*(1.0f + x*(1.0f - s)); -} - -inline static void ggml_vec_silu_backward_f32(const int n, float * dx, const float * x, const float * dy) { - for (int i = 0; i < n; ++i) { - dx[i] = ggml_silu_backward_f32(x[i], dy[i]); - } -} - -inline static void ggml_vec_sum_f32(const int n, float * s, const float * x) { -#ifndef GGML_USE_ACCELERATE - ggml_float sum = 0.0; - for (int i = 0; i < n; ++i) { - sum += (ggml_float)x[i]; - } - *s = sum; -#else - vDSP_sve(x, 1, s, n); -#endif -} - -inline static void ggml_vec_sum_f32_ggf(const int n, ggml_float * s, const float * x) { - ggml_float sum = 0.0; - for (int i = 0; i < n; ++i) { - sum += (ggml_float)x[i]; - } - *s = sum; -} - -inline static void ggml_vec_sum_f16_ggf(const int n, float * s, const ggml_fp16_t * x) { - float sum = 0.0f; - for (int i = 0; i < n; ++i) { - sum += GGML_FP16_TO_FP32(x[i]); - } - *s = sum; -} - -inline static void ggml_vec_sum_bf16_ggf(const int n, float * s, const ggml_bf16_t * x) { - float sum = 0.0f; - for (int i = 0; i < n; ++i) { - sum += GGML_BF16_TO_FP32(x[i]); - } - *s = sum; -} - -inline static void ggml_vec_max_f32(const int n, float * s, const float * x) { -#ifndef GGML_USE_ACCELERATE - float max = -INFINITY; - for (int i = 0; i < n; ++i) { - max = MAX(max, x[i]); - } - *s = max; -#else - vDSP_maxv(x, 1, s, n); -#endif -} - -inline static void ggml_vec_norm_inv_f32(const int n, float * s, const float * x) { - ggml_vec_norm_f32(n, s, x); - *s = 1.f/(*s); -} - -inline static void ggml_vec_argmax_f32(const int n, int * s, const float * x) { - float max = -INFINITY; - int idx = 0; - for (int i = 0; i < n; ++i) { - max = MAX(max, x[i]); - if (max == x[i]) { idx = i; } - } - *s = idx; -} - // Helpers for polling loops #if defined(__aarch64__) && ( defined(__clang__) || defined(__GNUC__) ) static inline void ggml_thread_cpu_relax(void) { @@ -2383,15 +683,20 @@ bool ggml_is_numa(void) { #define HWCAP2_I8MM (1 << 13) #endif +#if !defined(HWCAP2_SME) +#define HWCAP2_SME (1 << 23) +#endif + static void ggml_init_arm_arch_features(void) { #if defined(__linux__) && defined(__aarch64__) uint32_t hwcap = getauxval(AT_HWCAP); uint32_t hwcap2 = getauxval(AT_HWCAP2); - ggml_arm_arch_features.has_neon = !!(hwcap & HWCAP_ASIMD); + ggml_arm_arch_features.has_neon = !!(hwcap & HWCAP_ASIMD); ggml_arm_arch_features.has_dotprod = !!(hwcap & HWCAP_ASIMDDP); - ggml_arm_arch_features.has_i8mm = !!(hwcap2 & HWCAP2_I8MM); - ggml_arm_arch_features.has_sve = !!(hwcap & HWCAP_SVE); + ggml_arm_arch_features.has_i8mm = !!(hwcap2 & HWCAP2_I8MM); + ggml_arm_arch_features.has_sve = !!(hwcap & HWCAP_SVE); + ggml_arm_arch_features.has_sme = !!(hwcap2 & HWCAP2_SME); #if defined(__ARM_FEATURE_SVE) ggml_arm_arch_features.sve_cnt = PR_SVE_VL_LEN_MASK & prctl(PR_SVE_GET_VL); @@ -2414,6 +719,11 @@ static void ggml_init_arm_arch_features(void) { } ggml_arm_arch_features.has_i8mm = oldp; + if (sysctlbyname("hw.optional.arm.FEAT_SME", &oldp, &size, NULL, 0) != 0) { + oldp = 0; + } + ggml_arm_arch_features.has_sme = oldp; + ggml_arm_arch_features.has_sve = 0; ggml_arm_arch_features.sve_cnt = 0; #else @@ -2437,6 +747,12 @@ static void ggml_init_arm_arch_features(void) { ggml_arm_arch_features.has_sve = 0; ggml_arm_arch_features.sve_cnt = 0; #endif + +#if defined(__ARM_FEATURE_SME) || defined(__ARM_FEATURE_SME2) + ggml_arm_arch_features.has_sme = 1; +#else + ggml_arm_arch_features.has_sme = 0; +#endif #endif } #endif @@ -2855,4484 +1171,6 @@ void ggml_set_f32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, //////////////////////////////////////////////////////////////////////////////// -// ggml_compute_forward_dup - -static void ggml_compute_forward_dup_same_cont( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - GGML_ASSERT(ggml_nelements(dst) == ggml_nelements(src0)); - GGML_ASSERT(ggml_is_contiguous(dst) && ggml_is_contiguous(src0)); - GGML_ASSERT(src0->type == dst->type); - - const size_t nb0 = ggml_type_size(src0->type); - - const int ith = params->ith; // thread index - const int nth = params->nth; // number of threads - - // parallelize by elements - const int ne = ggml_nelements(dst); - const int dr = (ne + nth - 1) / nth; - const int ie0 = dr * ith; - const int ie1 = MIN(ie0 + dr, ne); - - if (ie0 < ie1) { - memcpy( - ((char *) dst->data + ie0*nb0), - ((char *) src0->data + ie0*nb0), - (ie1 - ie0) * nb0); - } -} - -static void ggml_compute_forward_dup_f16( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - GGML_ASSERT(ggml_nelements(dst) == ggml_nelements(src0)); - - GGML_TENSOR_UNARY_OP_LOCALS - - const int ith = params->ith; // thread index - const int nth = params->nth; // number of threads - - // parallelize by rows - const int nr = ne01; - // number of rows per thread - const int dr = (nr + nth - 1) / nth; - // row range for this thread - const int ir0 = dr * ith; - const int ir1 = MIN(ir0 + dr, nr); - - if (src0->type == dst->type && - ne00 == ne0 && - nb00 == ggml_type_size(src0->type) && nb0 == ggml_type_size(dst->type)) { - // copy by rows - const size_t rs = ne00*nb00; - for (int64_t i03 = 0; i03 < ne03; i03++) { - for (int64_t i02 = 0; i02 < ne02; i02++) { - for (int64_t i01 = ir0; i01 < ir1; i01++) { - memcpy( - ((char *) dst->data + i01*nb1 + i02*nb2 + i03*nb3), - ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03), - rs); - } - } - } - return; - } - - // TODO: add more special-case implementations for tensor shapes/strides that can benefit from memcpy - - if (ggml_is_contiguous(dst)) { - if (nb00 == sizeof(ggml_fp16_t)) { - if (dst->type == GGML_TYPE_F16) { - size_t id = 0; - const size_t rs = ne00 * nb00; - char * dst_ptr = (char *) dst->data; - - for (int i03 = 0; i03 < ne03; i03++) { - for (int i02 = 0; i02 < ne02; i02++) { - id += rs * ir0; - for (int i01 = ir0; i01 < ir1; i01++) { - const char * src0_ptr = (char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03; - memcpy(dst_ptr + id, src0_ptr, rs); - id += rs; - } - id += rs * (ne01 - ir1); - } - } - } else if (dst->type == GGML_TYPE_F32) { - size_t id = 0; - float * dst_ptr = (float *) dst->data; - - for (int i03 = 0; i03 < ne03; i03++) { - for (int i02 = 0; i02 < ne02; i02++) { - id += ne00 * ir0; - for (int i01 = ir0; i01 < ir1; i01++) { - const ggml_fp16_t * src0_ptr = (ggml_fp16_t *) ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03); - for (int i00 = 0; i00 < ne00; i00++) { - dst_ptr[id] = GGML_FP16_TO_FP32(src0_ptr[i00]); - id++; - } - } - id += ne00 * (ne01 - ir1); - } - } - } else if (ggml_get_type_traits_cpu(dst->type)->from_float) { - ggml_from_float_t const quantize_row_q = ggml_get_type_traits_cpu(dst->type)->from_float; - float * src0_f32 = (float *) params->wdata + (ne00 + CACHE_LINE_SIZE_F32) * ith; - - size_t id = 0; - size_t rs = nb0 * (ne00 / ggml_blck_size(dst->type)); - char * dst_ptr = (char *) dst->data; - - for (int i03 = 0; i03 < ne03; i03++) { - for (int i02 = 0; i02 < ne02; i02++) { - id += rs * ir0; - for (int i01 = ir0; i01 < ir1; i01++) { - const ggml_fp16_t * src0_ptr = (ggml_fp16_t *) ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03); - - for (int i00 = 0; i00 < ne00; i00++) { - src0_f32[i00] = GGML_FP16_TO_FP32(src0_ptr[i00]); - } - - quantize_row_q(src0_f32, dst_ptr + id, ne00); - id += rs; - } - id += rs * (ne01 - ir1); - } - } - } else { - GGML_ABORT("fatal error"); // TODO: implement - } - } else { - //printf("%s: this is not optimal - fix me\n", __func__); - - if (dst->type == GGML_TYPE_F32) { - size_t id = 0; - float * dst_ptr = (float *) dst->data; - - for (int i03 = 0; i03 < ne03; i03++) { - for (int i02 = 0; i02 < ne02; i02++) { - id += ne00 * ir0; - for (int i01 = ir0; i01 < ir1; i01++) { - for (int i00 = 0; i00 < ne00; i00++) { - const ggml_fp16_t * src0_ptr = (ggml_fp16_t *) ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03); - - dst_ptr[id] = GGML_FP16_TO_FP32(*src0_ptr); - id++; - } - } - id += ne00 * (ne01 - ir1); - } - } - } else if (dst->type == GGML_TYPE_F16) { - size_t id = 0; - ggml_fp16_t * dst_ptr = (ggml_fp16_t *) dst->data; - - for (int i03 = 0; i03 < ne03; i03++) { - for (int i02 = 0; i02 < ne02; i02++) { - id += ne00 * ir0; - for (int i01 = ir0; i01 < ir1; i01++) { - for (int i00 = 0; i00 < ne00; i00++) { - const ggml_fp16_t * src0_ptr = (ggml_fp16_t *) ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03); - - dst_ptr[id] = *src0_ptr; - id++; - } - } - id += ne00 * (ne01 - ir1); - } - } - } else { - GGML_ABORT("fatal error"); // TODO: implement - } - } - return; - } - - // dst counters - int64_t i10 = 0; - int64_t i11 = 0; - int64_t i12 = 0; - int64_t i13 = 0; - - if (dst->type == GGML_TYPE_F16) { - for (int64_t i03 = 0; i03 < ne03; i03++) { - for (int64_t i02 = 0; i02 < ne02; i02++) { - i10 += ne00 * ir0; - while (i10 >= ne0) { - i10 -= ne0; - if (++i11 == ne1) { - i11 = 0; - if (++i12 == ne2) { - i12 = 0; - if (++i13 == ne3) { - i13 = 0; - } - } - } - } - for (int64_t i01 = ir0; i01 < ir1; i01++) { - for (int64_t i00 = 0; i00 < ne00; i00++) { - const char * src0_ptr = ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03); - char * dst_ptr = ((char *) dst->data + i10*nb0 + i11*nb1 + i12*nb2 + i13*nb3); - - memcpy(dst_ptr, src0_ptr, sizeof(ggml_fp16_t)); - - if (++i10 == ne00) { - i10 = 0; - if (++i11 == ne01) { - i11 = 0; - if (++i12 == ne02) { - i12 = 0; - if (++i13 == ne03) { - i13 = 0; - } - } - } - } - } - } - i10 += ne00 * (ne01 - ir1); - while (i10 >= ne0) { - i10 -= ne0; - if (++i11 == ne1) { - i11 = 0; - if (++i12 == ne2) { - i12 = 0; - if (++i13 == ne3) { - i13 = 0; - } - } - } - } - } - } - } else if (dst->type == GGML_TYPE_F32) { - for (int64_t i03 = 0; i03 < ne03; i03++) { - for (int64_t i02 = 0; i02 < ne02; i02++) { - i10 += ne00 * ir0; - while (i10 >= ne0) { - i10 -= ne0; - if (++i11 == ne1) { - i11 = 0; - if (++i12 == ne2) { - i12 = 0; - if (++i13 == ne3) { - i13 = 0; - } - } - } - } - for (int64_t i01 = ir0; i01 < ir1; i01++) { - for (int64_t i00 = 0; i00 < ne00; i00++) { - const char * src0_ptr = ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03); - char * dst_ptr = ((char *) dst->data + i10*nb0 + i11*nb1 + i12*nb2 + i13*nb3); - - *(float *) dst_ptr = GGML_FP16_TO_FP32(*(const ggml_fp16_t *) src0_ptr); - - if (++i10 == ne0) { - i10 = 0; - if (++i11 == ne1) { - i11 = 0; - if (++i12 == ne2) { - i12 = 0; - if (++i13 == ne3) { - i13 = 0; - } - } - } - } - } - } - i10 += ne00 * (ne01 - ir1); - while (i10 >= ne0) { - i10 -= ne0; - if (++i11 == ne1) { - i11 = 0; - if (++i12 == ne2) { - i12 = 0; - if (++i13 == ne3) { - i13 = 0; - } - } - } - } - } - } - } else { - GGML_ABORT("fatal error"); // TODO: implement - } -} - -static void ggml_compute_forward_dup_bf16( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - GGML_ASSERT(ggml_nelements(dst) == ggml_nelements(src0)); - - GGML_TENSOR_UNARY_OP_LOCALS - - const int ith = params->ith; // thread index - const int nth = params->nth; // number of threads - - // parallelize by rows - const int nr = ne01; - // number of rows per thread - const int dr = (nr + nth - 1) / nth; - // row range for this thread - const int ir0 = dr * ith; - const int ir1 = MIN(ir0 + dr, nr); - - if (src0->type == dst->type && - ne00 == ne0 && - nb00 == ggml_type_size(src0->type) && nb0 == ggml_type_size(dst->type)) { - // copy by rows - const size_t rs = ne00*nb00; - for (int64_t i03 = 0; i03 < ne03; i03++) { - for (int64_t i02 = 0; i02 < ne02; i02++) { - for (int64_t i01 = ir0; i01 < ir1; i01++) { - memcpy( - ((char *) dst->data + i01*nb1 + i02*nb2 + i03*nb3), - ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03), - rs); - } - } - } - return; - } - - // TODO: add more special-case implementations for tensor shapes/strides that can benefit from memcpy - - if (ggml_is_contiguous(dst)) { - if (nb00 == sizeof(ggml_bf16_t)) { - if (dst->type == GGML_TYPE_BF16) { - size_t id = 0; - const size_t rs = ne00 * nb00; - char * dst_ptr = (char *) dst->data; - - for (int i03 = 0; i03 < ne03; i03++) { - for (int i02 = 0; i02 < ne02; i02++) { - id += rs * ir0; - for (int i01 = ir0; i01 < ir1; i01++) { - const char * src0_ptr = (char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03; - memcpy(dst_ptr + id, src0_ptr, rs); - id += rs; - } - id += rs * (ne01 - ir1); - } - } - } else if (dst->type == GGML_TYPE_F16) { - size_t id = 0; - ggml_fp16_t * dst_ptr = (ggml_fp16_t *) dst->data; - - for (int i03 = 0; i03 < ne03; i03++) { - for (int i02 = 0; i02 < ne02; i02++) { - id += ne00 * ir0; - for (int i01 = ir0; i01 < ir1; i01++) { - const ggml_bf16_t * src0_ptr = (ggml_bf16_t *) ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03); - for (int i00 = 0; i00 < ne00; i00++) { - dst_ptr[id] = GGML_FP32_TO_FP16(GGML_BF16_TO_FP32(src0_ptr[i00])); - id++; - } - } - id += ne00 * (ne01 - ir1); - } - } - } else if (dst->type == GGML_TYPE_F32) { - size_t id = 0; - float * dst_ptr = (float *) dst->data; - - for (int i03 = 0; i03 < ne03; i03++) { - for (int i02 = 0; i02 < ne02; i02++) { - id += ne00 * ir0; - for (int i01 = ir0; i01 < ir1; i01++) { - const ggml_bf16_t * src0_ptr = (ggml_bf16_t *) ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03); - for (int i00 = 0; i00 < ne00; i00++) { - dst_ptr[id] = GGML_BF16_TO_FP32(src0_ptr[i00]); - id++; - } - } - id += ne00 * (ne01 - ir1); - } - } - } else if (ggml_get_type_traits_cpu(dst->type)->from_float) { - ggml_from_float_t const quantize_row_q = ggml_get_type_traits_cpu(dst->type)->from_float; - float * src0_f32 = (float *) params->wdata + (ne00 + CACHE_LINE_SIZE_F32) * ith; - - size_t id = 0; - size_t rs = nb0 * (ne00 / ggml_blck_size(dst->type)); - char * dst_ptr = (char *) dst->data; - - for (int i03 = 0; i03 < ne03; i03++) { - for (int i02 = 0; i02 < ne02; i02++) { - id += rs * ir0; - for (int i01 = ir0; i01 < ir1; i01++) { - const ggml_bf16_t * src0_ptr = (ggml_bf16_t *) ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03); - - for (int i00 = 0; i00 < ne00; i00++) { - src0_f32[i00] = GGML_BF16_TO_FP32(src0_ptr[i00]); - } - - quantize_row_q(src0_f32, dst_ptr + id, ne00); - id += rs; - } - id += rs * (ne01 - ir1); - } - } - } else { - GGML_ABORT("fatal error"); // TODO: implement - } - } else { - //printf("%s: this is not optimal - fix me\n", __func__); - - if (dst->type == GGML_TYPE_F32) { - size_t id = 0; - float * dst_ptr = (float *) dst->data; - - for (int i03 = 0; i03 < ne03; i03++) { - for (int i02 = 0; i02 < ne02; i02++) { - id += ne00 * ir0; - for (int i01 = ir0; i01 < ir1; i01++) { - for (int i00 = 0; i00 < ne00; i00++) { - const ggml_bf16_t * src0_ptr = (ggml_bf16_t *) ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03); - - dst_ptr[id] = GGML_BF16_TO_FP32(*src0_ptr); - id++; - } - } - id += ne00 * (ne01 - ir1); - } - } - } else if (dst->type == GGML_TYPE_BF16) { - size_t id = 0; - ggml_bf16_t * dst_ptr = (ggml_bf16_t *) dst->data; - - for (int i03 = 0; i03 < ne03; i03++) { - for (int i02 = 0; i02 < ne02; i02++) { - id += ne00 * ir0; - for (int i01 = ir0; i01 < ir1; i01++) { - for (int i00 = 0; i00 < ne00; i00++) { - const ggml_bf16_t * src0_ptr = (ggml_bf16_t *) ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03); - - dst_ptr[id] = *src0_ptr; - id++; - } - } - id += ne00 * (ne01 - ir1); - } - } - } else if (dst->type == GGML_TYPE_F16) { - size_t id = 0; - ggml_fp16_t * dst_ptr = (ggml_fp16_t *) dst->data; - - for (int i03 = 0; i03 < ne03; i03++) { - for (int i02 = 0; i02 < ne02; i02++) { - id += ne00 * ir0; - for (int i01 = ir0; i01 < ir1; i01++) { - for (int i00 = 0; i00 < ne00; i00++) { - const ggml_bf16_t * src0_ptr = (ggml_bf16_t *) ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03); - - dst_ptr[id] = GGML_FP32_TO_FP16(GGML_BF16_TO_FP32(*src0_ptr)); - id++; - } - } - id += ne00 * (ne01 - ir1); - } - } - } else { - GGML_ABORT("fatal error"); // TODO: implement - } - } - return; - } - - // dst counters - int64_t i10 = 0; - int64_t i11 = 0; - int64_t i12 = 0; - int64_t i13 = 0; - - if (dst->type == GGML_TYPE_BF16) { - for (int64_t i03 = 0; i03 < ne03; i03++) { - for (int64_t i02 = 0; i02 < ne02; i02++) { - i10 += ne00 * ir0; - while (i10 >= ne0) { - i10 -= ne0; - if (++i11 == ne1) { - i11 = 0; - if (++i12 == ne2) { - i12 = 0; - if (++i13 == ne3) { - i13 = 0; - } - } - } - } - for (int64_t i01 = ir0; i01 < ir1; i01++) { - for (int64_t i00 = 0; i00 < ne00; i00++) { - const char * src0_ptr = ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03); - char * dst_ptr = ((char *) dst->data + i10*nb0 + i11*nb1 + i12*nb2 + i13*nb3); - - memcpy(dst_ptr, src0_ptr, sizeof(ggml_bf16_t)); - - if (++i10 == ne00) { - i10 = 0; - if (++i11 == ne01) { - i11 = 0; - if (++i12 == ne02) { - i12 = 0; - if (++i13 == ne03) { - i13 = 0; - } - } - } - } - } - } - i10 += ne00 * (ne01 - ir1); - while (i10 >= ne0) { - i10 -= ne0; - if (++i11 == ne1) { - i11 = 0; - if (++i12 == ne2) { - i12 = 0; - if (++i13 == ne3) { - i13 = 0; - } - } - } - } - } - } - } else if (dst->type == GGML_TYPE_F16) { - for (int64_t i03 = 0; i03 < ne03; i03++) { - for (int64_t i02 = 0; i02 < ne02; i02++) { - i10 += ne00 * ir0; - while (i10 >= ne0) { - i10 -= ne0; - if (++i11 == ne1) { - i11 = 0; - if (++i12 == ne2) { - i12 = 0; - if (++i13 == ne3) { - i13 = 0; - } - } - } - } - for (int64_t i01 = ir0; i01 < ir1; i01++) { - for (int64_t i00 = 0; i00 < ne00; i00++) { - const char * src0_ptr = ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03); - char * dst_ptr = ((char *) dst->data + i10*nb0 + i11*nb1 + i12*nb2 + i13*nb3); - - *(ggml_fp16_t *) dst_ptr = GGML_FP32_TO_FP16(GGML_BF16_TO_FP32(*(const ggml_bf16_t *) src0_ptr)); - - if (++i10 == ne0) { - i10 = 0; - if (++i11 == ne1) { - i11 = 0; - if (++i12 == ne2) { - i12 = 0; - if (++i13 == ne3) { - i13 = 0; - } - } - } - } - } - } - i10 += ne00 * (ne01 - ir1); - while (i10 >= ne0) { - i10 -= ne0; - if (++i11 == ne1) { - i11 = 0; - if (++i12 == ne2) { - i12 = 0; - if (++i13 == ne3) { - i13 = 0; - } - } - } - } - } - } - } else if (dst->type == GGML_TYPE_F32) { - for (int64_t i03 = 0; i03 < ne03; i03++) { - for (int64_t i02 = 0; i02 < ne02; i02++) { - i10 += ne00 * ir0; - while (i10 >= ne0) { - i10 -= ne0; - if (++i11 == ne1) { - i11 = 0; - if (++i12 == ne2) { - i12 = 0; - if (++i13 == ne3) { - i13 = 0; - } - } - } - } - for (int64_t i01 = ir0; i01 < ir1; i01++) { - for (int64_t i00 = 0; i00 < ne00; i00++) { - const char * src0_ptr = ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03); - char * dst_ptr = ((char *) dst->data + i10*nb0 + i11*nb1 + i12*nb2 + i13*nb3); - - *(float *) dst_ptr = GGML_BF16_TO_FP32(*(const ggml_bf16_t *) src0_ptr); - - if (++i10 == ne0) { - i10 = 0; - if (++i11 == ne1) { - i11 = 0; - if (++i12 == ne2) { - i12 = 0; - if (++i13 == ne3) { - i13 = 0; - } - } - } - } - } - } - i10 += ne00 * (ne01 - ir1); - while (i10 >= ne0) { - i10 -= ne0; - if (++i11 == ne1) { - i11 = 0; - if (++i12 == ne2) { - i12 = 0; - if (++i13 == ne3) { - i13 = 0; - } - } - } - } - } - } - } else { - GGML_ABORT("fatal error"); // TODO: implement - } -} - -static void ggml_compute_forward_dup_f32( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - GGML_ASSERT(ggml_nelements(dst) == ggml_nelements(src0)); - - GGML_TENSOR_UNARY_OP_LOCALS - - const int ith = params->ith; // thread index - const int nth = params->nth; // number of threads - - // parallelize by rows - const int nr = ne01; - // number of rows per thread - const int dr = (nr + nth - 1) / nth; - // row range for this thread - const int ir0 = dr * ith; - const int ir1 = MIN(ir0 + dr, nr); - - if (src0->type == dst->type && - ne00 == ne0 && - nb00 == ggml_type_size(src0->type) && nb0 == ggml_type_size(dst->type)) { - // copy by rows - const size_t rs = ne00*nb00; - for (int64_t i03 = 0; i03 < ne03; i03++) { - for (int64_t i02 = 0; i02 < ne02; i02++) { - for (int64_t i01 = ir0; i01 < ir1; i01++) { - memcpy( - ((char *) dst->data + i01*nb1 + i02*nb2 + i03*nb3), - ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03), - rs); - } - } - } - return; - } - - if (ggml_is_contiguous(dst)) { - // TODO: simplify - if (nb00 == sizeof(float)) { - if (dst->type == GGML_TYPE_F32) { - size_t id = 0; - const size_t rs = ne00 * nb00; - char * dst_ptr = (char *) dst->data; - - for (int i03 = 0; i03 < ne03; i03++) { - for (int i02 = 0; i02 < ne02; i02++) { - id += rs * ir0; - for (int i01 = ir0; i01 < ir1; i01++) { - const char * src0_ptr = (char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03; - memcpy(dst_ptr + id, src0_ptr, rs); - id += rs; - } - id += rs * (ne01 - ir1); - } - } - } else if (ggml_get_type_traits_cpu(dst->type)->from_float) { - ggml_from_float_t const quantize_row_q = ggml_get_type_traits_cpu(dst->type)->from_float; - - size_t id = 0; - size_t rs = nb0 * (ne00 / ggml_blck_size(dst->type)); - char * dst_ptr = (char *) dst->data; - - for (int i03 = 0; i03 < ne03; i03++) { - for (int i02 = 0; i02 < ne02; i02++) { - id += rs * ir0; - for (int i01 = ir0; i01 < ir1; i01++) { - const float * src0_ptr = (float *) ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03); - quantize_row_q(src0_ptr, dst_ptr + id, ne00); - id += rs; - } - id += rs * (ne01 - ir1); - } - } - } else { - GGML_ABORT("fatal error"); // TODO: implement - } - } else { - //printf("%s: this is not optimal - fix me\n", __func__); - - if (dst->type == GGML_TYPE_F32) { - size_t id = 0; - float * dst_ptr = (float *) dst->data; - - for (int i03 = 0; i03 < ne03; i03++) { - for (int i02 = 0; i02 < ne02; i02++) { - id += ne00 * ir0; - for (int i01 = ir0; i01 < ir1; i01++) { - for (int i00 = 0; i00 < ne00; i00++) { - const float * src0_ptr = (float *) ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03); - - dst_ptr[id] = *src0_ptr; - id++; - } - } - id += ne00 * (ne01 - ir1); - } - } - } else if (dst->type == GGML_TYPE_F16) { - size_t id = 0; - ggml_fp16_t * dst_ptr = (ggml_fp16_t *) dst->data; - - for (int i03 = 0; i03 < ne03; i03++) { - for (int i02 = 0; i02 < ne02; i02++) { - id += ne00 * ir0; - for (int i01 = ir0; i01 < ir1; i01++) { - for (int i00 = 0; i00 < ne00; i00++) { - const float * src0_ptr = (float *) ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03); - - dst_ptr[id] = GGML_FP32_TO_FP16(*src0_ptr); - id++; - } - } - id += ne00 * (ne01 - ir1); - } - } - } else if (dst->type == GGML_TYPE_BF16) { - size_t id = 0; - ggml_bf16_t * dst_ptr = (ggml_bf16_t *) dst->data; - - for (int i03 = 0; i03 < ne03; i03++) { - for (int i02 = 0; i02 < ne02; i02++) { - id += ne00 * ir0; - for (int i01 = ir0; i01 < ir1; i01++) { - for (int i00 = 0; i00 < ne00; i00++) { - const float * src0_ptr = (float *) ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03); - - dst_ptr[id] = GGML_FP32_TO_BF16(*src0_ptr); - id++; - } - } - id += ne00 * (ne01 - ir1); - } - } - } else { - GGML_ABORT("fatal error"); // TODO: implement - } - } - - return; - } - - // dst counters - - int64_t i10 = 0; - int64_t i11 = 0; - int64_t i12 = 0; - int64_t i13 = 0; - - if (dst->type == GGML_TYPE_F32) { - for (int64_t i03 = 0; i03 < ne03; i03++) { - for (int64_t i02 = 0; i02 < ne02; i02++) { - i10 += ne00 * ir0; - while (i10 >= ne0) { - i10 -= ne0; - if (++i11 == ne1) { - i11 = 0; - if (++i12 == ne2) { - i12 = 0; - if (++i13 == ne3) { - i13 = 0; - } - } - } - } - for (int64_t i01 = ir0; i01 < ir1; i01++) { - for (int64_t i00 = 0; i00 < ne00; i00++) { - const char * src0_ptr = ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03); - char * dst_ptr = ((char *) dst->data + i10*nb0 + i11*nb1 + i12*nb2 + i13*nb3); - - memcpy(dst_ptr, src0_ptr, sizeof(float)); - - if (++i10 == ne0) { - i10 = 0; - if (++i11 == ne1) { - i11 = 0; - if (++i12 == ne2) { - i12 = 0; - if (++i13 == ne3) { - i13 = 0; - } - } - } - } - } - } - i10 += ne00 * (ne01 - ir1); - while (i10 >= ne0) { - i10 -= ne0; - if (++i11 == ne1) { - i11 = 0; - if (++i12 == ne2) { - i12 = 0; - if (++i13 == ne3) { - i13 = 0; - } - } - } - } - } - } - } else if (dst->type == GGML_TYPE_F16) { - for (int64_t i03 = 0; i03 < ne03; i03++) { - for (int64_t i02 = 0; i02 < ne02; i02++) { - i10 += ne00 * ir0; - while (i10 >= ne0) { - i10 -= ne0; - if (++i11 == ne1) { - i11 = 0; - if (++i12 == ne2) { - i12 = 0; - if (++i13 == ne3) { - i13 = 0; - } - } - } - } - for (int64_t i01 = ir0; i01 < ir1; i01++) { - for (int64_t i00 = 0; i00 < ne00; i00++) { - const char * src0_ptr = ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03); - char * dst_ptr = ((char *) dst->data + i10*nb0 + i11*nb1 + i12*nb2 + i13*nb3); - - *(ggml_fp16_t *) dst_ptr = GGML_FP32_TO_FP16(*(const float *) src0_ptr); - - if (++i10 == ne0) { - i10 = 0; - if (++i11 == ne1) { - i11 = 0; - if (++i12 == ne2) { - i12 = 0; - if (++i13 == ne3) { - i13 = 0; - } - } - } - } - } - } - i10 += ne00 * (ne01 - ir1); - while (i10 >= ne0) { - i10 -= ne0; - if (++i11 == ne1) { - i11 = 0; - if (++i12 == ne2) { - i12 = 0; - if (++i13 == ne3) { - i13 = 0; - } - } - } - } - } - } - } else if (dst->type == GGML_TYPE_BF16) { - for (int64_t i03 = 0; i03 < ne03; i03++) { - for (int64_t i02 = 0; i02 < ne02; i02++) { - i10 += ne00 * ir0; - while (i10 >= ne0) { - i10 -= ne0; - if (++i11 == ne1) { - i11 = 0; - if (++i12 == ne2) { - i12 = 0; - if (++i13 == ne3) { - i13 = 0; - } - } - } - } - for (int64_t i01 = ir0; i01 < ir1; i01++) { - for (int64_t i00 = 0; i00 < ne00; i00++) { - const char * src0_ptr = ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03); - char * dst_ptr = ((char *) dst->data + i10*nb0 + i11*nb1 + i12*nb2 + i13*nb3); - - *(ggml_bf16_t *) dst_ptr = GGML_FP32_TO_BF16(*(const float *) src0_ptr); - - if (++i10 == ne0) { - i10 = 0; - if (++i11 == ne1) { - i11 = 0; - if (++i12 == ne2) { - i12 = 0; - if (++i13 == ne3) { - i13 = 0; - } - } - } - } - } - } - i10 += ne00 * (ne01 - ir1); - while (i10 >= ne0) { - i10 -= ne0; - if (++i11 == ne1) { - i11 = 0; - if (++i12 == ne2) { - i12 = 0; - if (++i13 == ne3) { - i13 = 0; - } - } - } - } - } - } - } else { - GGML_ABORT("fatal error"); // TODO: implement - } -} - -// A simplified version of ggml_compute_forward_dup that doesn't do float upcasting, and just plain old memcpy. -static void ggml_compute_forward_dup_bytes( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - GGML_ASSERT(ggml_nelements(dst) == ggml_nelements(src0)); - GGML_ASSERT(src0->type == dst->type); - - GGML_TENSOR_UNARY_OP_LOCALS; - - if (ggml_is_contiguous(src0) && ggml_is_contiguous(dst)) { - ggml_compute_forward_dup_same_cont(params, dst); - return; - } - - const size_t type_size = ggml_type_size(src0->type); - const int ith = params->ith; // thread index - const int nth = params->nth; // number of threads - - - // parallelize by rows - const int nr = ne01; - // number of rows per thread - const int dr = (nr + nth - 1) / nth; - // row range for this thread - const int ir0 = dr * ith; - const int ir1 = MIN(ir0 + dr, nr); - - if (src0->type == dst->type && - ne00 == ne0 && - nb00 == type_size && nb0 == type_size) { - // copy by rows - const size_t rs = ne00 * type_size; - for (int64_t i03 = 0; i03 < ne03; i03++) { - for (int64_t i02 = 0; i02 < ne02; i02++) { - for (int64_t i01 = ir0; i01 < ir1; i01++) { - memcpy( - ((char *) dst->data + i01*nb1 + i02*nb2 + i03*nb3), - ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03), - rs); - } - } - } - return; - } - - if (ggml_is_contiguous(dst)) { - size_t id = 0; - char * dst_ptr = (char *) dst->data; - const size_t rs = ne00 * type_size; - - if (nb00 == type_size) { - // src0 is contigous on first dimension, copy by rows - for (int64_t i03 = 0; i03 < ne03; i03++) { - for (int64_t i02 = 0; i02 < ne02; i02++) { - id += rs * ir0; - for (int64_t i01 = ir0; i01 < ir1; i01++) { - const char * src0_ptr = (char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03; - memcpy(dst_ptr + id, src0_ptr, rs); - id += rs; - } - id += rs * (ne01 - ir1); - } - } - } else { - //printf("%s: this is not optimal - fix me\n", __func__); - - for (int64_t i03 = 0; i03 < ne03; i03++) { - for (int64_t i02 = 0; i02 < ne02; i02++) { - id += rs * ir0; - for (int64_t i01 = ir0; i01 < ir1; i01++) { - for (int64_t i00 = 0; i00 < ne00; i00++) { - const char * src0_ptr = (char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03; - memcpy(dst_ptr + id, src0_ptr, type_size); - - id += type_size; - } - } - id += rs * (ne01 - ir1); - } - } - } - - return; - } - - // dst counters - - int64_t i10 = 0; - int64_t i11 = 0; - int64_t i12 = 0; - int64_t i13 = 0; - - for (int64_t i03 = 0; i03 < ne03; i03++) { - for (int64_t i02 = 0; i02 < ne02; i02++) { - i10 += ne00 * ir0; - while (i10 >= ne0) { - i10 -= ne0; - if (++i11 == ne1) { - i11 = 0; - if (++i12 == ne2) { - i12 = 0; - if (++i13 == ne3) { - i13 = 0; - } - } - } - } - for (int64_t i01 = ir0; i01 < ir1; i01++) { - for (int64_t i00 = 0; i00 < ne00; i00++) { - const char * src0_ptr = ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03); - char * dst_ptr = ((char *) dst->data + i10*nb0 + i11*nb1 + i12*nb2 + i13*nb3); - - memcpy(dst_ptr, src0_ptr, type_size); - - if (++i10 == ne0) { - i10 = 0; - if (++i11 == ne1) { - i11 = 0; - if (++i12 == ne2) { - i12 = 0; - if (++i13 == ne3) { - i13 = 0; - } - } - } - } - } - } - i10 += ne00 * (ne01 - ir1); - while (i10 >= ne0) { - i10 -= ne0; - if (++i11 == ne1) { - i11 = 0; - if (++i12 == ne2) { - i12 = 0; - if (++i13 == ne3) { - i13 = 0; - } - } - } - } - } - } -} - -static void ggml_compute_forward_dup_q( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - const struct ggml_tensor * src1 = dst->src[1]; - - GGML_TENSOR_BINARY_OP_LOCALS - - const enum ggml_type type = src0->type; - ggml_to_float_t const dequantize_row_q = ggml_get_type_traits(type)->to_float; - - size_t qk = ggml_blck_size(type); - const int64_t nr = ggml_nelements(src1) / qk; - - // destination must be contiguous in the first dimension - GGML_ASSERT(nb10 == ggml_type_size(dst->type)); - // must either have first dimension large enough to hold a row, or fully contiguous - GGML_ASSERT((ne10 % qk) == 0 || ggml_is_contiguous(dst)); - - const int ith = params->ith; - const int nth = params->nth; - - const int dr = (nr + nth - 1)/nth; - - // row range for this thread - const int ir0 = dr*ith; - const int ir1 = MIN(ir0 + dr, nr); - - for (int64_t ir = ir0; ir < ir1; ++ir) { - - uint32_t i = ir * qk; - - const int64_t i03 = i/(ne00 * ne01 * ne02); - const int64_t i02 = (i - i03*ne00*ne01*ne02 )/ (ne00*ne01); - const int64_t i01 = (i - i03*ne00*ne01*ne02 - i02*ne01*ne00) / ne00; - const int64_t i00 = i - i03*ne00*ne01*ne02 - i02*ne01*ne00 - i01*ne00; - const int64_t x_offset = (i00/qk)*nb00 + i01*nb01 + i02*nb02 + i03 * nb03; - - const int64_t i13 = i/(ne10 * ne11 * ne12); - const int64_t i12 = (i - i13*ne10*ne11*ne12) / (ne10*ne11); - const int64_t i11 = (i - i13*ne10*ne11*ne12 - i12*ne10*ne11) / ne10; - const int64_t i10 = i - i13*ne10*ne11*ne12 - i12*ne10*ne11 - i11*ne10; - const int64_t dst_offset = i10*nb10 + i11*nb11 + i12*nb12 + i13*nb13; - - dequantize_row_q( - (const void *) ((char *) src0->data + x_offset), - (float *) ((char *) dst->data + dst_offset), qk); - } -} - -static void ggml_compute_forward_dup( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - if (src0->type == dst->type) { - ggml_compute_forward_dup_bytes(params, dst); - return; - } - - switch (src0->type) { - case GGML_TYPE_F16: - { - ggml_compute_forward_dup_f16(params, dst); - } break; - case GGML_TYPE_BF16: - { - ggml_compute_forward_dup_bf16(params, dst); - } break; - case GGML_TYPE_F32: - { - ggml_compute_forward_dup_f32(params, dst); - } break; - default: - { - if (ggml_is_quantized(src0->type) && dst->type == GGML_TYPE_F32) { - ggml_compute_forward_dup_q(params, dst); - break; - } - GGML_ABORT("fatal error"); - } - } -} - -// ggml_compute_forward_add - -static void ggml_compute_forward_add_f32( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - const struct ggml_tensor * src1 = dst->src[1]; - - GGML_ASSERT(ggml_can_repeat(src1, src0) && ggml_are_same_shape(src0, dst)); - - const int ith = params->ith; - const int nth = params->nth; - - const int nr = ggml_nrows(src0); - - GGML_TENSOR_BINARY_OP_LOCALS - - GGML_ASSERT( nb0 == sizeof(float)); - GGML_ASSERT(nb00 == sizeof(float)); - - // rows per thread - const int dr = (nr + nth - 1)/nth; - - // row range for this thread - const int ir0 = dr*ith; - const int ir1 = MIN(ir0 + dr, nr); - - if (nb10 == sizeof(float)) { - for (int ir = ir0; ir < ir1; ++ir) { - // src1 is broadcastable across src0 and dst in i1, i2, i3 - const int64_t i03 = ir/(ne02*ne01); - const int64_t i02 = (ir - i03*ne02*ne01)/ne01; - const int64_t i01 = (ir - i03*ne02*ne01 - i02*ne01); - - const int64_t i13 = i03 % ne13; - const int64_t i12 = i02 % ne12; - const int64_t i11 = i01 % ne11; - const int64_t nr0 = ne00 / ne10; - - float * dst_ptr = (float *) ((char *) dst->data + i03*nb3 + i02*nb2 + i01*nb1 ); - float * src0_ptr = (float *) ((char *) src0->data + i03*nb03 + i02*nb02 + i01*nb01); - float * src1_ptr = (float *) ((char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11); - - for (int64_t r = 0; r < nr0; ++r) { -#ifdef GGML_USE_ACCELERATE - vDSP_vadd(src0_ptr + r*ne10, 1, src1_ptr, 1, dst_ptr + r*ne10, 1, ne10); -#else - ggml_vec_add_f32(ne10, dst_ptr + r*ne10, src0_ptr + r*ne10, src1_ptr); -#endif - } - } - } else { - // src1 is not contiguous - for (int ir = ir0; ir < ir1; ++ir) { - // src1 is broadcastable across src0 and dst in i1, i2, i3 - const int64_t i03 = ir/(ne02*ne01); - const int64_t i02 = (ir - i03*ne02*ne01)/ne01; - const int64_t i01 = (ir - i03*ne02*ne01 - i02*ne01); - - const int64_t i13 = i03 % ne13; - const int64_t i12 = i02 % ne12; - const int64_t i11 = i01 % ne11; - - float * dst_ptr = (float *) ((char *) dst->data + i03*nb3 + i02*nb2 + i01*nb1 ); - float * src0_ptr = (float *) ((char *) src0->data + i03*nb03 + i02*nb02 + i01*nb01); - - for (int64_t i0 = 0; i0 < ne0; ++i0) { - const int64_t i10 = i0 % ne10; - float * src1_ptr = (float *) ((char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11 + i10*nb10); - - dst_ptr[i0] = src0_ptr[i0] + *src1_ptr; - } - } - } -} - -static void ggml_compute_forward_add_f16_f32( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - const struct ggml_tensor * src1 = dst->src[1]; - - GGML_ASSERT(ggml_are_same_shape(src0, src1) && ggml_are_same_shape(src0, dst)); - - const int ith = params->ith; - const int nth = params->nth; - - const int nr = ggml_nrows(src0); - - GGML_TENSOR_BINARY_OP_LOCALS - - GGML_ASSERT(src0->type == GGML_TYPE_F16); - GGML_ASSERT(src1->type == GGML_TYPE_F32); - - if (dst->type == GGML_TYPE_F32) { - GGML_ASSERT( nb0 == sizeof(float)); - } - else { - GGML_ASSERT(dst->type == GGML_TYPE_F16); - GGML_ASSERT( nb0 == sizeof(ggml_fp16_t)); - } - - GGML_ASSERT(nb00 == sizeof(ggml_fp16_t)); - - // rows per thread - const int dr = (nr + nth - 1)/nth; - - // row range for this thread - const int ir0 = dr*ith; - const int ir1 = MIN(ir0 + dr, nr); - - if (nb10 == sizeof(float)) { - if (dst->type == GGML_TYPE_F16) { - for (int ir = ir0; ir < ir1; ++ir) { - // src0, src1 and dst are same shape => same indices - const int i3 = ir/(ne2*ne1); - const int i2 = (ir - i3*ne2*ne1)/ne1; - const int i1 = (ir - i3*ne2*ne1 - i2*ne1); - - ggml_fp16_t * dst_ptr = (ggml_fp16_t *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1); - ggml_fp16_t * src0_ptr = (ggml_fp16_t *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01); - float * src1_ptr = (float *) ((char *) src1->data + i3*nb13 + i2*nb12 + i1*nb11); - - for (int i = 0; i < ne0; i++) { - dst_ptr[i] = GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(src0_ptr[i]) + src1_ptr[i]); - } - } - } else { - for (int ir = ir0; ir < ir1; ++ir) { - // src0, src1 and dst are same shape => same indices - const int i3 = ir/(ne2*ne1); - const int i2 = (ir - i3*ne2*ne1)/ne1; - const int i1 = (ir - i3*ne2*ne1 - i2*ne1); - - float * dst_ptr = (float *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1); - ggml_fp16_t * src0_ptr = (ggml_fp16_t *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01); - float * src1_ptr = (float *) ((char *) src1->data + i3*nb13 + i2*nb12 + i1*nb11); - - for (int i = 0; i < ne0; i++) { - dst_ptr[i] = GGML_FP16_TO_FP32(src0_ptr[i]) + src1_ptr[i]; - } - } - } - } - else { - // src1 is not contiguous - GGML_ABORT("fatal error"); - } -} - -static void ggml_compute_forward_add_bf16_f32( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - const struct ggml_tensor * src1 = dst->src[1]; - - GGML_ASSERT(ggml_are_same_shape(src0, src1) && ggml_are_same_shape(src0, dst)); - - const int ith = params->ith; - const int nth = params->nth; - - const int nr = ggml_nrows(src0); - - GGML_TENSOR_BINARY_OP_LOCALS - - GGML_ASSERT(src0->type == GGML_TYPE_BF16); - GGML_ASSERT(src1->type == GGML_TYPE_F32); - - if (dst->type == GGML_TYPE_F32) { - GGML_ASSERT( nb0 == sizeof(float)); - } - else { - GGML_ASSERT(dst->type == GGML_TYPE_BF16); - GGML_ASSERT( nb0 == sizeof(ggml_bf16_t)); - } - - GGML_ASSERT(nb00 == sizeof(ggml_bf16_t)); - - // rows per thread - const int dr = (nr + nth - 1)/nth; - - // row range for this thread - const int ir0 = dr*ith; - const int ir1 = MIN(ir0 + dr, nr); - - if (nb10 == sizeof(float)) { - if (dst->type == GGML_TYPE_BF16) { - for (int ir = ir0; ir < ir1; ++ir) { - // src0, src1 and dst are same shape => same indices - const int i3 = ir/(ne2*ne1); - const int i2 = (ir - i3*ne2*ne1)/ne1; - const int i1 = (ir - i3*ne2*ne1 - i2*ne1); - - ggml_bf16_t * dst_ptr = (ggml_bf16_t *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1); - ggml_bf16_t * src0_ptr = (ggml_bf16_t *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01); - float * src1_ptr = (float *) ((char *) src1->data + i3*nb13 + i2*nb12 + i1*nb11); - - for (int i = 0; i < ne0; i++) { - dst_ptr[i] = GGML_FP32_TO_BF16(GGML_BF16_TO_FP32(src0_ptr[i]) + src1_ptr[i]); - } - } - } else { - for (int ir = ir0; ir < ir1; ++ir) { - // src0, src1 and dst are same shape => same indices - const int i3 = ir/(ne2*ne1); - const int i2 = (ir - i3*ne2*ne1)/ne1; - const int i1 = (ir - i3*ne2*ne1 - i2*ne1); - - float * dst_ptr = (float *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1); - ggml_bf16_t * src0_ptr = (ggml_bf16_t *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01); - float * src1_ptr = (float *) ((char *) src1->data + i3*nb13 + i2*nb12 + i1*nb11); - - for (int i = 0; i < ne0; i++) { - dst_ptr[i] = GGML_BF16_TO_FP32(src0_ptr[i]) + src1_ptr[i]; - } - } - } - } - else { - // src1 is not contiguous - GGML_ABORT("fatal error"); - } -} - -static void ggml_compute_forward_add_f16_f16( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - const struct ggml_tensor * src1 = dst->src[1]; - - GGML_ASSERT(ggml_are_same_shape(src0, src1) && ggml_are_same_shape(src0, dst)); - - const int ith = params->ith; - const int nth = params->nth; - - const int nr = ggml_nrows(src0); - - GGML_TENSOR_BINARY_OP_LOCALS - - GGML_ASSERT(src0->type == GGML_TYPE_F16); - GGML_ASSERT(src1->type == GGML_TYPE_F16); - GGML_ASSERT(dst->type == GGML_TYPE_F16); - - GGML_ASSERT( nb0 == sizeof(ggml_fp16_t)); - GGML_ASSERT(nb00 == sizeof(ggml_fp16_t)); - - // rows per thread - const int dr = (nr + nth - 1)/nth; - - // row range for this thread - const int ir0 = dr*ith; - const int ir1 = MIN(ir0 + dr, nr); - - if (nb10 == sizeof(ggml_fp16_t)) { - for (int ir = ir0; ir < ir1; ++ir) { - // src0, src1 and dst are same shape => same indices - const int i3 = ir/(ne2*ne1); - const int i2 = (ir - i3*ne2*ne1)/ne1; - const int i1 = (ir - i3*ne2*ne1 - i2*ne1); - - ggml_fp16_t * dst_ptr = (ggml_fp16_t *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1); - ggml_fp16_t * src0_ptr = (ggml_fp16_t *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01); - ggml_fp16_t * src1_ptr = (ggml_fp16_t *) ((char *) src1->data + i3*nb13 + i2*nb12 + i1*nb11); - - for (int i = 0; i < ne0; i++) { - dst_ptr[i] = GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(src0_ptr[i]) + GGML_FP16_TO_FP32(src1_ptr[i])); - } - } - } - else { - // src1 is not contiguous - GGML_ABORT("fatal error"); - } -} - -static void ggml_compute_forward_add_bf16_bf16( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - const struct ggml_tensor * src1 = dst->src[1]; - - GGML_ASSERT(ggml_are_same_shape(src0, src1) && ggml_are_same_shape(src0, dst)); - - const int ith = params->ith; - const int nth = params->nth; - - const int nr = ggml_nrows(src0); - - GGML_TENSOR_BINARY_OP_LOCALS - - GGML_ASSERT(src0->type == GGML_TYPE_BF16); - GGML_ASSERT(src1->type == GGML_TYPE_BF16); - GGML_ASSERT(dst->type == GGML_TYPE_BF16); - - GGML_ASSERT( nb0 == sizeof(ggml_bf16_t)); - GGML_ASSERT(nb00 == sizeof(ggml_bf16_t)); - - // rows per thread - const int dr = (nr + nth - 1)/nth; - - // row range for this thread - const int ir0 = dr*ith; - const int ir1 = MIN(ir0 + dr, nr); - - if (nb10 == sizeof(ggml_bf16_t)) { - for (int ir = ir0; ir < ir1; ++ir) { - // src0, src1 and dst are same shape => same indices - const int i3 = ir/(ne2*ne1); - const int i2 = (ir - i3*ne2*ne1)/ne1; - const int i1 = (ir - i3*ne2*ne1 - i2*ne1); - - ggml_bf16_t * dst_ptr = (ggml_bf16_t *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1); - ggml_bf16_t * src0_ptr = (ggml_bf16_t *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01); - ggml_bf16_t * src1_ptr = (ggml_bf16_t *) ((char *) src1->data + i3*nb13 + i2*nb12 + i1*nb11); - - for (int i = 0; i < ne0; i++) { - dst_ptr[i] = GGML_FP32_TO_BF16(GGML_BF16_TO_FP32(src0_ptr[i]) + GGML_BF16_TO_FP32(src1_ptr[i])); - } - } - } - else { - // src1 is not contiguous - GGML_ABORT("fatal error"); - } -} - -static void ggml_compute_forward_add_q_f32( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - const struct ggml_tensor * src1 = dst->src[1]; - - GGML_ASSERT(ggml_are_same_shape(src0, src1) && ggml_are_same_shape(src0, dst)); - - const int nr = ggml_nrows(src0); - - GGML_TENSOR_BINARY_OP_LOCALS - - const int ith = params->ith; - const int nth = params->nth; - - const enum ggml_type type = src0->type; - const enum ggml_type dtype = dst->type; - ggml_to_float_t const dequantize_row_q = ggml_get_type_traits(type)->to_float; - ggml_from_float_t const quantize_row_q = ggml_get_type_traits_cpu(dtype)->from_float; - - // we don't support permuted src0 or src1 - GGML_ASSERT(nb00 == ggml_type_size(type)); - GGML_ASSERT(nb10 == sizeof(float)); - - // dst cannot be transposed or permuted - GGML_ASSERT(nb0 <= nb1); - GGML_ASSERT(nb1 <= nb2); - GGML_ASSERT(nb2 <= nb3); - - GGML_ASSERT(ggml_is_quantized(src0->type)); - GGML_ASSERT(src1->type == GGML_TYPE_F32); - - // rows per thread - const int dr = (nr + nth - 1)/nth; - - // row range for this thread - const int ir0 = dr*ith; - const int ir1 = MIN(ir0 + dr, nr); - - float * wdata = (float *) params->wdata + (ne00 + CACHE_LINE_SIZE_F32) * ith; - - for (int ir = ir0; ir < ir1; ++ir) { - // src0 indices - const int i03 = ir/(ne02*ne01); - const int i02 = (ir - i03*ne02*ne01)/ne01; - const int i01 = (ir - i03*ne02*ne01 - i02*ne01); - - // src1 and dst are same shape as src0 => same indices - const int i13 = i03; - const int i12 = i02; - const int i11 = i01; - - const int i3 = i03; - const int i2 = i02; - const int i1 = i01; - - void * src0_row = (void *) ((char *) src0->data + (i01*nb01 + i02*nb02 + i03*nb03)); - float * src1_row = (float *)((char *) src1->data + (i11*nb11 + i12*nb12 + i13*nb13)); - void * dst_row = (void *) ((char *) dst->data + ( i1*nb1 + i2*nb2 + i3*nb3)); - - assert(ne00 % 32 == 0); - - // unquantize row from src0 to temp buffer - dequantize_row_q(src0_row, wdata, ne00); - // add src1 - ggml_vec_acc_f32(ne00, wdata, src1_row); - // quantize row to dst - if (quantize_row_q != NULL) { - quantize_row_q(wdata, dst_row, ne00); - } else { - memcpy(dst_row, wdata, ne0*nb0); - } - } -} - -static void ggml_compute_forward_add( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - const struct ggml_tensor * src1 = dst->src[1]; - - switch (src0->type) { - case GGML_TYPE_F32: - { - if (src1->type == GGML_TYPE_F32) { - ggml_compute_forward_add_f32(params, dst); - } - else { - GGML_ABORT("fatal error"); - } - } break; - case GGML_TYPE_F16: - { - if (src1->type == GGML_TYPE_F16) { - ggml_compute_forward_add_f16_f16(params, dst); - } - else if (src1->type == GGML_TYPE_F32) { - ggml_compute_forward_add_f16_f32(params, dst); - } - else { - GGML_ABORT("fatal error"); - } - } break; - case GGML_TYPE_BF16: - { - if (src1->type == GGML_TYPE_BF16) { - ggml_compute_forward_add_bf16_bf16(params, dst); - } - else if (src1->type == GGML_TYPE_F32) { - ggml_compute_forward_add_bf16_f32(params, dst); - } - else { - GGML_ABORT("fatal error"); - } - } break; - case GGML_TYPE_Q4_0: - case GGML_TYPE_Q4_1: - case GGML_TYPE_Q5_0: - case GGML_TYPE_Q5_1: - case GGML_TYPE_Q8_0: - case GGML_TYPE_Q2_K: - case GGML_TYPE_Q3_K: - case GGML_TYPE_Q4_K: - case GGML_TYPE_Q5_K: - case GGML_TYPE_Q6_K: - case GGML_TYPE_TQ1_0: - case GGML_TYPE_TQ2_0: - case GGML_TYPE_IQ2_XXS: - case GGML_TYPE_IQ2_XS: - case GGML_TYPE_IQ3_XXS: - case GGML_TYPE_IQ1_S: - case GGML_TYPE_IQ1_M: - case GGML_TYPE_IQ4_NL: - case GGML_TYPE_IQ4_XS: - case GGML_TYPE_IQ3_S: - case GGML_TYPE_IQ2_S: - { - ggml_compute_forward_add_q_f32(params, dst); - } break; - default: - { - GGML_ABORT("fatal error"); - } - } -} - -// ggml_compute_forward_add1 - -static void ggml_compute_forward_add1_f32( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - const struct ggml_tensor * src1 = dst->src[1]; - - GGML_ASSERT(ggml_are_same_shape(src0, dst)); - GGML_ASSERT(ggml_is_scalar(src1)); - - const int ith = params->ith; - const int nth = params->nth; - - const int nr = ggml_nrows(src0); - - GGML_TENSOR_UNARY_OP_LOCALS - - GGML_ASSERT( nb0 == sizeof(float)); - GGML_ASSERT(nb00 == sizeof(float)); - - // rows per thread - const int dr = (nr + nth - 1)/nth; - - // row range for this thread - const int ir0 = dr*ith; - const int ir1 = MIN(ir0 + dr, nr); - - for (int ir = ir0; ir < ir1; ++ir) { - // src0 and dst are same shape => same indices - const int i3 = ir/(ne2*ne1); - const int i2 = (ir - i3*ne2*ne1)/ne1; - const int i1 = (ir - i3*ne2*ne1 - i2*ne1); - -#ifdef GGML_USE_ACCELERATE - UNUSED(ggml_vec_add1_f32); - - vDSP_vadd( - (float *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01), 1, - (float *) ((char *) src1->data), 0, - (float *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 ), 1, - ne0); -#else - ggml_vec_add1_f32(ne0, - (float *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 ), - (float *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01), - *(float *) src1->data); -#endif - } -} - -static void ggml_compute_forward_add1_f16_f32( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - const struct ggml_tensor * src1 = dst->src[1]; - - GGML_ASSERT(ggml_are_same_shape(src0, dst)); - GGML_ASSERT(ggml_is_scalar(src1)); - - // scalar to add - const float v = *(float *) src1->data; - - const int ith = params->ith; - const int nth = params->nth; - - const int nr = ggml_nrows(src0); - - GGML_TENSOR_UNARY_OP_LOCALS - - GGML_ASSERT(src0->type == GGML_TYPE_F16); - GGML_ASSERT(src1->type == GGML_TYPE_F32); - GGML_ASSERT(dst->type == GGML_TYPE_F16); - - GGML_ASSERT( nb0 == sizeof(ggml_fp16_t)); - GGML_ASSERT(nb00 == sizeof(ggml_fp16_t)); - - // rows per thread - const int dr = (nr + nth - 1)/nth; - - // row range for this thread - const int ir0 = dr*ith; - const int ir1 = MIN(ir0 + dr, nr); - - for (int ir = ir0; ir < ir1; ++ir) { - // src0 and dst are same shape => same indices - const int i3 = ir/(ne2*ne1); - const int i2 = (ir - i3*ne2*ne1)/ne1; - const int i1 = (ir - i3*ne2*ne1 - i2*ne1); - - ggml_fp16_t * dst_ptr = (ggml_fp16_t *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 ); - ggml_fp16_t * src0_ptr = (ggml_fp16_t *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01); - for (int i = 0; i < ne0; i++) { - dst_ptr[i] = GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(src0_ptr[i]) + v); - } - } -} - -static void ggml_compute_forward_add1_f16_f16( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - const struct ggml_tensor * src1 = dst->src[1]; - - GGML_ASSERT(ggml_are_same_shape(src0, dst)); - GGML_ASSERT(ggml_is_scalar(src1)); - - // scalar to add - const float v = GGML_FP16_TO_FP32(*(ggml_fp16_t *) src1->data); - - const int ith = params->ith; - const int nth = params->nth; - - const int nr = ggml_nrows(src0); - - GGML_TENSOR_UNARY_OP_LOCALS - - GGML_ASSERT(src0->type == GGML_TYPE_F16); - GGML_ASSERT(src1->type == GGML_TYPE_F16); - GGML_ASSERT(dst->type == GGML_TYPE_F16); - - GGML_ASSERT( nb0 == sizeof(ggml_fp16_t)); - GGML_ASSERT(nb00 == sizeof(ggml_fp16_t)); - - // rows per thread - const int dr = (nr + nth - 1)/nth; - - // row range for this thread - const int ir0 = dr*ith; - const int ir1 = MIN(ir0 + dr, nr); - - for (int ir = ir0; ir < ir1; ++ir) { - // src0 and dst are same shape => same indices - const int i3 = ir/(ne2*ne1); - const int i2 = (ir - i3*ne2*ne1)/ne1; - const int i1 = (ir - i3*ne2*ne1 - i2*ne1); - - ggml_fp16_t * dst_ptr = (ggml_fp16_t *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 ); - ggml_fp16_t * src0_ptr = (ggml_fp16_t *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01); - for (int i = 0; i < ne0; i++) { - dst_ptr[i] = GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(src0_ptr[i]) + v); - } - } -} - -static void ggml_compute_forward_add1_q_f32( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - const struct ggml_tensor * src1 = dst->src[1]; - - GGML_ASSERT(ggml_are_same_shape(src0, dst)); - GGML_ASSERT(ggml_is_scalar(src1)); - - // scalar to add - const float v = *(float *) src1->data; - - const int ith = params->ith; - const int nth = params->nth; - - const int nr = ggml_nrows(src0); - - GGML_TENSOR_UNARY_OP_LOCALS - - const enum ggml_type type = src0->type; - ggml_to_float_t const dequantize_row_q = ggml_get_type_traits(type)->to_float; - ggml_from_float_t const quantize_row_q = ggml_get_type_traits_cpu(type)->from_float; - - // we don't support permuted src0 - GGML_ASSERT(nb00 == ggml_type_size(type)); - - // dst cannot be transposed or permuted - GGML_ASSERT(nb0 <= nb1); - GGML_ASSERT(nb1 <= nb2); - GGML_ASSERT(nb2 <= nb3); - - GGML_ASSERT(ggml_is_quantized(src0->type)); - GGML_ASSERT(dst->type == src0->type); - GGML_ASSERT(src1->type == GGML_TYPE_F32); - - // rows per thread - const int dr = (nr + nth - 1)/nth; - - // row range for this thread - const int ir0 = dr*ith; - const int ir1 = MIN(ir0 + dr, nr); - - float * wdata = (float *) params->wdata + (ne0 + CACHE_LINE_SIZE_F32) * ith; - - for (int ir = ir0; ir < ir1; ++ir) { - // src0 and dst are same shape => same indices - const int i3 = ir/(ne2*ne1); - const int i2 = (ir - i3*ne2*ne1)/ne1; - const int i1 = (ir - i3*ne2*ne1 - i2*ne1); - - void * src0_row = (void *) ((char *) src0->data + (i1*nb01 + i2*nb02 + i3*nb03)); - void * dst_row = (void *) ((char *) dst->data + (i1*nb1 + i2*nb2 + i3*nb0 )); - - assert(ne0 % 32 == 0); - - // unquantize row from src0 to temp buffer - dequantize_row_q(src0_row, wdata, ne0); - // add src1 - ggml_vec_acc1_f32(ne0, wdata, v); - // quantize row to dst - quantize_row_q(wdata, dst_row, ne0); - } -} - -static void ggml_compute_forward_add1_bf16_f32( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - const struct ggml_tensor * src1 = dst->src[1]; - - GGML_ASSERT(ggml_are_same_shape(src0, dst)); - GGML_ASSERT(ggml_is_scalar(src1)); - - // scalar to add - const float v = *(float *) src1->data; - - const int ith = params->ith; - const int nth = params->nth; - - const int nr = ggml_nrows(src0); - - GGML_TENSOR_UNARY_OP_LOCALS - - GGML_ASSERT(src0->type == GGML_TYPE_BF16); - GGML_ASSERT(src1->type == GGML_TYPE_F32); - GGML_ASSERT(dst->type == GGML_TYPE_BF16); - - GGML_ASSERT( nb0 == sizeof(ggml_bf16_t)); - GGML_ASSERT(nb00 == sizeof(ggml_bf16_t)); - - // rows per thread - const int dr = (nr + nth - 1)/nth; - - // row range for this thread - const int ir0 = dr*ith; - const int ir1 = MIN(ir0 + dr, nr); - - for (int ir = ir0; ir < ir1; ++ir) { - // src0 and dst are same shape => same indices - const int i3 = ir/(ne2*ne1); - const int i2 = (ir - i3*ne2*ne1)/ne1; - const int i1 = (ir - i3*ne2*ne1 - i2*ne1); - - ggml_bf16_t * dst_ptr = (ggml_bf16_t *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 ); - ggml_bf16_t * src0_ptr = (ggml_bf16_t *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01); - for (int i = 0; i < ne0; i++) { - dst_ptr[i] = GGML_FP32_TO_BF16(GGML_BF16_TO_FP32(src0_ptr[i]) + v); - } - } -} - -static void ggml_compute_forward_add1_bf16_bf16( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - const struct ggml_tensor * src1 = dst->src[1]; - - GGML_ASSERT(ggml_are_same_shape(src0, dst)); - GGML_ASSERT(ggml_is_scalar(src1)); - - // scalar to add - const float v = GGML_BF16_TO_FP32(*(ggml_bf16_t *) src1->data); - - const int ith = params->ith; - const int nth = params->nth; - - const int nr = ggml_nrows(src0); - - GGML_TENSOR_UNARY_OP_LOCALS - - GGML_ASSERT(src0->type == GGML_TYPE_BF16); - GGML_ASSERT(src1->type == GGML_TYPE_BF16); - GGML_ASSERT(dst->type == GGML_TYPE_BF16); - - GGML_ASSERT( nb0 == sizeof(ggml_bf16_t)); - GGML_ASSERT(nb00 == sizeof(ggml_bf16_t)); - - // rows per thread - const int dr = (nr + nth - 1)/nth; - - // row range for this thread - const int ir0 = dr*ith; - const int ir1 = MIN(ir0 + dr, nr); - - for (int ir = ir0; ir < ir1; ++ir) { - // src0 and dst are same shape => same indices - const int i3 = ir/(ne2*ne1); - const int i2 = (ir - i3*ne2*ne1)/ne1; - const int i1 = (ir - i3*ne2*ne1 - i2*ne1); - - ggml_bf16_t * dst_ptr = (ggml_bf16_t *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 ); - ggml_bf16_t * src0_ptr = (ggml_bf16_t *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01); - for (int i = 0; i < ne0; i++) { - dst_ptr[i] = GGML_FP32_TO_BF16(GGML_BF16_TO_FP32(src0_ptr[i]) + v); - } - } -} - -static void ggml_compute_forward_add1( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - const struct ggml_tensor * src1 = dst->src[1]; - - switch (src0->type) { - case GGML_TYPE_F32: - { - ggml_compute_forward_add1_f32(params, dst); - } break; - case GGML_TYPE_F16: - { - if (src1->type == GGML_TYPE_F16) { - ggml_compute_forward_add1_f16_f16(params, dst); - } - else if (src1->type == GGML_TYPE_F32) { - ggml_compute_forward_add1_f16_f32(params, dst); - } - else { - GGML_ABORT("fatal error"); - } - } break; - case GGML_TYPE_BF16: - { - if (src1->type == GGML_TYPE_BF16) { - ggml_compute_forward_add1_bf16_bf16(params, dst); - } - else if (src1->type == GGML_TYPE_F32) { - ggml_compute_forward_add1_bf16_f32(params, dst); - } - else { - GGML_ABORT("fatal error"); - } - } break; - case GGML_TYPE_Q4_0: - case GGML_TYPE_Q4_1: - case GGML_TYPE_Q5_0: - case GGML_TYPE_Q5_1: - case GGML_TYPE_Q8_0: - case GGML_TYPE_Q8_1: - case GGML_TYPE_Q2_K: - case GGML_TYPE_Q3_K: - case GGML_TYPE_Q4_K: - case GGML_TYPE_Q5_K: - case GGML_TYPE_Q6_K: - case GGML_TYPE_TQ1_0: - case GGML_TYPE_TQ2_0: - case GGML_TYPE_IQ2_XXS: - case GGML_TYPE_IQ2_XS: - case GGML_TYPE_IQ3_XXS: - case GGML_TYPE_IQ1_S: - case GGML_TYPE_IQ1_M: - case GGML_TYPE_IQ4_NL: - case GGML_TYPE_IQ4_XS: - case GGML_TYPE_IQ3_S: - case GGML_TYPE_IQ2_S: - { - ggml_compute_forward_add1_q_f32(params, dst); - } break; - default: - { - GGML_ABORT("fatal error"); - } - } -} - -// ggml_compute_forward_acc - -static void ggml_compute_forward_acc_f32( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - const struct ggml_tensor * src1 = dst->src[1]; - - GGML_ASSERT(ggml_are_same_shape(src0, dst)); - GGML_ASSERT(ggml_is_contiguous(dst) && ggml_is_contiguous(src0)); - - // view src0 and dst with these strides and data offset inbytes during acc - // nb0 is implicitly element_size because src0 and dst are contiguous - size_t nb1 = ((int32_t *) dst->op_params)[0]; - size_t nb2 = ((int32_t *) dst->op_params)[1]; - size_t nb3 = ((int32_t *) dst->op_params)[2]; - size_t offset = ((int32_t *) dst->op_params)[3]; - bool inplace = (bool) ((int32_t *) dst->op_params)[4]; - - if (!inplace) { - if (params->ith == 0) { - // memcpy needs to be synchronized across threads to avoid race conditions. - // => do it in INIT phase - memcpy( - ((char *) dst->data), - ((char *) src0->data), - ggml_nbytes(dst)); - } - ggml_barrier(params->threadpool); - } - - const int ith = params->ith; - const int nth = params->nth; - - const int nr = ggml_nrows(src1); - const int nc = src1->ne[0]; - - GGML_TENSOR_LOCALS(int64_t, ne1, src1, ne) - GGML_TENSOR_LOCALS(size_t, nb1, src1, nb) - - // src0 and dst as viewed during acc - const size_t nb0 = ggml_element_size(src0); - - const size_t nb00 = nb0; - const size_t nb01 = nb1; - const size_t nb02 = nb2; - const size_t nb03 = nb3; - - GGML_ASSERT(offset + (ne10 == 0 ? 0 : ne10-1)*nb0 + (ne11 == 0 ? 0 : ne11-1)*nb1 + (ne12 == 0 ? 0 : ne12-1)*nb2 + (ne13 == 0 ? 0 : ne13-1)*nb3 < ggml_nbytes(dst)); - GGML_ASSERT(offset + (ne10 == 0 ? 0 : ne10-1)*nb00 + (ne11 == 0 ? 0 : ne11-1)*nb01 + (ne12 == 0 ? 0 : ne12-1)*nb02 + (ne13 == 0 ? 0 : ne13-1)*nb03 < ggml_nbytes(src0)); - - GGML_ASSERT(nb10 == sizeof(float)); - - // rows per thread - const int dr = (nr + nth - 1)/nth; - - // row range for this thread - const int ir0 = dr*ith; - const int ir1 = MIN(ir0 + dr, nr); - - for (int ir = ir0; ir < ir1; ++ir) { - // src0 and dst are viewed with shape of src1 and offset - // => same indices - const int i3 = ir/(ne12*ne11); - const int i2 = (ir - i3*ne12*ne11)/ne11; - const int i1 = (ir - i3*ne12*ne11 - i2*ne11); - -#ifdef GGML_USE_ACCELERATE - vDSP_vadd( - (float *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + offset), 1, - (float *) ((char *) src1->data + i3*nb13 + i2*nb12 + i1*nb11), 1, - (float *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + offset), 1, nc); -#else - ggml_vec_add_f32(nc, - (float *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + offset), - (float *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + offset), - (float *) ((char *) src1->data + i3*nb13 + i2*nb12 + i1*nb11)); -#endif - } -} - -static void ggml_compute_forward_acc( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - switch (src0->type) { - case GGML_TYPE_F32: - { - ggml_compute_forward_acc_f32(params, dst); - } break; - case GGML_TYPE_F16: - case GGML_TYPE_BF16: - case GGML_TYPE_Q4_0: - case GGML_TYPE_Q4_1: - case GGML_TYPE_Q5_0: - case GGML_TYPE_Q5_1: - case GGML_TYPE_Q8_0: - case GGML_TYPE_Q8_1: - case GGML_TYPE_Q2_K: - case GGML_TYPE_Q3_K: - case GGML_TYPE_Q4_K: - case GGML_TYPE_Q5_K: - case GGML_TYPE_Q6_K: - case GGML_TYPE_TQ1_0: - case GGML_TYPE_TQ2_0: - case GGML_TYPE_IQ2_XXS: - case GGML_TYPE_IQ2_XS: - case GGML_TYPE_IQ3_XXS: - case GGML_TYPE_IQ1_S: - case GGML_TYPE_IQ1_M: - case GGML_TYPE_IQ4_NL: - case GGML_TYPE_IQ4_XS: - case GGML_TYPE_IQ3_S: - case GGML_TYPE_IQ2_S: - default: - { - GGML_ABORT("fatal error"); - } - } -} - -// ggml_compute_forward_sub - -static void ggml_compute_forward_sub_f32( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - const struct ggml_tensor * src1 = dst->src[1]; - - assert(ggml_can_repeat(src1, src0) && ggml_are_same_shape(src0, dst)); - - const int ith = params->ith; - const int nth = params->nth; - - const int nr = ggml_nrows(src0); - - GGML_TENSOR_BINARY_OP_LOCALS - - GGML_ASSERT( nb0 == sizeof(float)); - GGML_ASSERT(nb00 == sizeof(float)); - - // rows per thread - const int dr = (nr + nth - 1)/nth; - - // row range for this thread - const int ir0 = dr*ith; - const int ir1 = MIN(ir0 + dr, nr); - - if (nb10 == sizeof(float)) { - for (int ir = ir0; ir < ir1; ++ir) { - // src1 is broadcastable across src0 and dst in i1, i2, i3 - const int64_t i03 = ir/(ne02*ne01); - const int64_t i02 = (ir - i03*ne02*ne01)/ne01; - const int64_t i01 = (ir - i03*ne02*ne01 - i02*ne01); - - const int64_t i13 = i03 % ne13; - const int64_t i12 = i02 % ne12; - const int64_t i11 = i01 % ne11; - const int64_t nr0 = ne00 / ne10; - - float * dst_ptr = (float *) ((char *) dst->data + i03*nb3 + i02*nb2 + i01*nb1 ); - float * src0_ptr = (float *) ((char *) src0->data + i03*nb03 + i02*nb02 + i01*nb01); - float * src1_ptr = (float *) ((char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11); - - for (int64_t r = 0; r < nr0; ++r) { -#ifdef GGML_USE_ACCELERATE - vDSP_vsub(src1_ptr, 1, src0_ptr + r*ne10, 1, dst_ptr + r*ne10, 1, ne10); -#else - ggml_vec_sub_f32(ne10, dst_ptr + r*ne10, src0_ptr + r*ne10, src1_ptr); -#endif - } - } - } else { - // src1 is not contiguous - for (int ir = ir0; ir < ir1; ++ir) { - // src1 is broadcastable across src0 and dst in i1, i2, i3 - const int64_t i03 = ir/(ne02*ne01); - const int64_t i02 = (ir - i03*ne02*ne01)/ne01; - const int64_t i01 = (ir - i03*ne02*ne01 - i02*ne01); - - const int64_t i13 = i03 % ne13; - const int64_t i12 = i02 % ne12; - const int64_t i11 = i01 % ne11; - - float * dst_ptr = (float *) ((char *) dst->data + i03*nb3 + i02*nb2 + i01*nb1 ); - float * src0_ptr = (float *) ((char *) src0->data + i03*nb03 + i02*nb02 + i01*nb01); - - for (int64_t i0 = 0; i0 < ne0; ++i0) { - const int64_t i10 = i0 % ne10; - float * src1_ptr = (float *) ((char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11 + i10*nb10); - - dst_ptr[i0] = src0_ptr[i0] - *src1_ptr; - } - } - } -} - -static void ggml_compute_forward_sub( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - switch (src0->type) { - case GGML_TYPE_F32: - { - ggml_compute_forward_sub_f32(params, dst); - } break; - default: - { - GGML_ABORT("fatal error"); - } - } -} - -// ggml_compute_forward_mul - -static void ggml_compute_forward_mul_f32( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - const struct ggml_tensor * src1 = dst->src[1]; - - GGML_ASSERT(ggml_can_repeat(src1, src0) && ggml_are_same_shape(src0, dst)); - - const int ith = params->ith; - const int nth = params->nth; - - const int64_t nr = ggml_nrows(src0); - - GGML_TENSOR_BINARY_OP_LOCALS - - GGML_ASSERT( nb0 == sizeof(float)); - GGML_ASSERT(nb00 == sizeof(float)); - - if (nb10 == sizeof(float)) { - for (int64_t ir = ith; ir < nr; ir += nth) { - // src0 and dst are same shape => same indices - const int64_t i03 = ir/(ne02*ne01); - const int64_t i02 = (ir - i03*ne02*ne01)/ne01; - const int64_t i01 = (ir - i03*ne02*ne01 - i02*ne01); - - const int64_t i13 = i03 % ne13; - const int64_t i12 = i02 % ne12; - const int64_t i11 = i01 % ne11; - const int64_t nr0 = ne00 / ne10; - - float * dst_ptr = (float *) ((char *) dst->data + i03*nb3 + i02*nb2 + i01*nb1 ); - float * src0_ptr = (float *) ((char *) src0->data + i03*nb03 + i02*nb02 + i01*nb01); - float * src1_ptr = (float *) ((char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11); - - for (int64_t r = 0 ; r < nr0; ++r) { -#ifdef GGML_USE_ACCELERATE - UNUSED(ggml_vec_mul_f32); - - vDSP_vmul(src0_ptr + r*ne10, 1, src1_ptr, 1, dst_ptr + r*ne10, 1, ne10); -#else - ggml_vec_mul_f32(ne10, dst_ptr + r*ne10, src0_ptr + r*ne10, src1_ptr); -#endif - } - } - } else { - // src1 is not contiguous - for (int64_t ir = ith; ir < nr; ir += nth) { - // src0 and dst are same shape => same indices - // src1 is broadcastable across src0 and dst in i1, i2, i3 - const int64_t i03 = ir/(ne02*ne01); - const int64_t i02 = (ir - i03*ne02*ne01)/ne01; - const int64_t i01 = (ir - i03*ne02*ne01 - i02*ne01); - - const int64_t i13 = i03 % ne13; - const int64_t i12 = i02 % ne12; - const int64_t i11 = i01 % ne11; - - float * dst_ptr = (float *) ((char *) dst->data + i03*nb3 + i02*nb2 + i01*nb1 ); - float * src0_ptr = (float *) ((char *) src0->data + i03*nb03 + i02*nb02 + i01*nb01); - - for (int64_t i0 = 0; i0 < ne00; ++i0) { - const int64_t i10 = i0 % ne10; - float * src1_ptr = (float *) ((char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11 + i10*nb10); - - dst_ptr[i0] = src0_ptr[i0] * (*src1_ptr); - } - } - } -} - -static void ggml_compute_forward_mul( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - const struct ggml_tensor * src1 = dst->src[1]; - - GGML_ASSERT(src1->type == GGML_TYPE_F32 && "only f32 src1 supported for now"); - - switch (src0->type) { - case GGML_TYPE_F32: - { - ggml_compute_forward_mul_f32(params, dst); - } break; - default: - { - GGML_ABORT("fatal error"); - } - } -} - -// ggml_compute_forward_div - -static void ggml_compute_forward_div_f32( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - const struct ggml_tensor * src1 = dst->src[1]; - - GGML_ASSERT(ggml_can_repeat(src1, src0) && ggml_are_same_shape(src0, dst)); - - const int ith = params->ith; - const int nth = params->nth; - - const int64_t nr = ggml_nrows(src0); - - GGML_TENSOR_BINARY_OP_LOCALS - - GGML_ASSERT( nb0 == sizeof(float)); - GGML_ASSERT(nb00 == sizeof(float)); - - if (nb10 == sizeof(float)) { - for (int64_t ir = ith; ir < nr; ir += nth) { - // src0 and dst are same shape => same indices - const int64_t i03 = ir/(ne02*ne01); - const int64_t i02 = (ir - i03*ne02*ne01)/ne01; - const int64_t i01 = (ir - i03*ne02*ne01 - i02*ne01); - - const int64_t i13 = i03 % ne13; - const int64_t i12 = i02 % ne12; - const int64_t i11 = i01 % ne11; - const int64_t nr0 = ne00 / ne10; - - float * dst_ptr = (float *) ((char *) dst->data + i03*nb3 + i02*nb2 + i01*nb1 ); - float * src0_ptr = (float *) ((char *) src0->data + i03*nb03 + i02*nb02 + i01*nb01); - float * src1_ptr = (float *) ((char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11); - - for (int64_t r = 0; r < nr0; ++r) { -#ifdef GGML_USE_ACCELERATE - UNUSED(ggml_vec_div_f32); - - vDSP_vdiv(src1_ptr, 1, src0_ptr + r*ne10, 1, dst_ptr + r*ne10, 1, ne10); -#else - ggml_vec_div_f32(ne10, dst_ptr + r*ne10, src0_ptr + r*ne10, src1_ptr); -#endif - } - } - } else { - // src1 is not contiguous - for (int64_t ir = ith; ir < nr; ir += nth) { - // src0 and dst are same shape => same indices - // src1 is broadcastable across src0 and dst in i1, i2, i3 - const int64_t i03 = ir/(ne02*ne01); - const int64_t i02 = (ir - i03*ne02*ne01)/ne01; - const int64_t i01 = (ir - i03*ne02*ne01 - i02*ne01); - - const int64_t i13 = i03 % ne13; - const int64_t i12 = i02 % ne12; - const int64_t i11 = i01 % ne11; - - float * dst_ptr = (float *) ((char *) dst->data + i03*nb3 + i02*nb2 + i01*nb1 ); - float * src0_ptr = (float *) ((char *) src0->data + i03*nb03 + i02*nb02 + i01*nb01); - - for (int64_t i0 = 0; i0 < ne00; ++i0) { - const int64_t i10 = i0 % ne10; - float * src1_ptr = (float *) ((char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11 + i10*nb10); - - dst_ptr[i0] = src0_ptr[i0] / (*src1_ptr); - } - } - } -} - -static void ggml_compute_forward_div( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - switch (src0->type) { - case GGML_TYPE_F32: - { - ggml_compute_forward_div_f32(params, dst); - } break; - default: - { - GGML_ABORT("fatal error"); - } - } -} - -// ggml_compute_forward_sqr - -static void ggml_compute_forward_sqr_f32( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - if (params->ith != 0) { - return; - } - - assert(ggml_are_same_shape(src0, dst)); - - const int n = ggml_nrows(src0); - const int nc = src0->ne[0]; - - assert( dst->nb[0] == sizeof(float)); - assert(src0->nb[0] == sizeof(float)); - - for (int i = 0; i < n; i++) { - ggml_vec_sqr_f32(nc, - (float *) ((char *) dst->data + i*( dst->nb[1])), - (float *) ((char *) src0->data + i*(src0->nb[1]))); - } -} - -static void ggml_compute_forward_sqr( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - switch (src0->type) { - case GGML_TYPE_F32: - { - ggml_compute_forward_sqr_f32(params, dst); - } break; - default: - { - GGML_ABORT("fatal error"); - } - } -} - -// ggml_compute_forward_sqrt - -static void ggml_compute_forward_sqrt_f32( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - if (params->ith != 0) { - return; - } - - assert(ggml_are_same_shape(src0, dst)); - - const int n = ggml_nrows(src0); - const int nc = src0->ne[0]; - - assert( dst->nb[0] == sizeof(float)); - assert(src0->nb[0] == sizeof(float)); - - for (int i = 0; i < n; i++) { - ggml_vec_sqrt_f32(nc, - (float *) ((char *) dst->data + i*( dst->nb[1])), - (float *) ((char *) src0->data + i*(src0->nb[1]))); - } -} - -static void ggml_compute_forward_sqrt( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - switch (src0->type) { - case GGML_TYPE_F32: - { - ggml_compute_forward_sqrt_f32(params, dst); - } break; - default: - { - GGML_ABORT("fatal error"); - } - } -} - -// ggml_compute_forward_log - -static void ggml_compute_forward_log_f32( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - if (params->ith != 0) { - return; - } - - GGML_ASSERT(ggml_are_same_shape(src0, dst)); - - const int n = ggml_nrows(src0); - const int nc = src0->ne[0]; - - GGML_ASSERT( dst->nb[0] == sizeof(float)); - GGML_ASSERT(src0->nb[0] == sizeof(float)); - - for (int i = 0; i < n; i++) { - ggml_vec_log_f32(nc, - (float *) ((char *) dst->data + i*( dst->nb[1])), - (float *) ((char *) src0->data + i*(src0->nb[1]))); - } -} - -static void ggml_compute_forward_log( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - switch (src0->type) { - case GGML_TYPE_F32: - { - ggml_compute_forward_log_f32(params, dst); - } break; - default: - { - GGML_ABORT("fatal error"); - } - } -} - -// ggml_compute_forward_sin - -static void ggml_compute_forward_sin_f32( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - if (params->ith != 0) { - return; - } - - GGML_ASSERT(ggml_are_same_shape(src0, dst)); - - const int n = ggml_nrows(src0); - const int nc = src0->ne[0]; - - GGML_ASSERT( dst->nb[0] == sizeof(float)); - GGML_ASSERT(src0->nb[0] == sizeof(float)); - - for (int i = 0; i < n; i++) { - ggml_vec_sin_f32(nc, - (float *) ((char *) dst->data + i*( dst->nb[1])), - (float *) ((char *) src0->data + i*(src0->nb[1]))); - } -} - -static void ggml_compute_forward_sin( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - switch (src0->type) { - case GGML_TYPE_F32: - { - ggml_compute_forward_sin_f32(params, dst); - } break; - default: - { - GGML_ABORT("fatal error"); - } - } -} - -// ggml_compute_forward_cos - -static void ggml_compute_forward_cos_f32( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - if (params->ith != 0) { - return; - } - - GGML_ASSERT(ggml_are_same_shape(src0, dst)); - - const int n = ggml_nrows(src0); - const int nc = src0->ne[0]; - - GGML_ASSERT( dst->nb[0] == sizeof(float)); - GGML_ASSERT(src0->nb[0] == sizeof(float)); - - for (int i = 0; i < n; i++) { - ggml_vec_cos_f32(nc, - (float *) ((char *) dst->data + i*( dst->nb[1])), - (float *) ((char *) src0->data + i*(src0->nb[1]))); - } -} - -static void ggml_compute_forward_cos( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - switch (src0->type) { - case GGML_TYPE_F32: - { - ggml_compute_forward_cos_f32(params, dst); - } break; - default: - { - GGML_ABORT("fatal error"); - } - } -} - -// ggml_compute_forward_sum - -static void ggml_compute_forward_sum_f32( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - if (params->ith != 0) { - return; - } - - assert(ggml_is_scalar(dst)); - assert(src0->nb[0] == sizeof(float)); - - GGML_TENSOR_LOCALS(int64_t, ne0, src0, ne) - GGML_TENSOR_LOCALS(size_t, nb0, src0, nb) - - ggml_float sum = 0; - ggml_float row_sum = 0; - - for (int64_t i03 = 0; i03 < ne03; i03++) { - for (int64_t i02 = 0; i02 < ne02; i02++) { - for (int64_t i01 = 0; i01 < ne01; i01++) { - ggml_vec_sum_f32_ggf(ne00, - &row_sum, - (float *) ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03)); - sum += row_sum; - } - } - } - ((float *) dst->data)[0] = sum; -} - -static void ggml_compute_forward_sum_f16( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - if (params->ith != 0) { - return; - } - - assert(ggml_is_scalar(dst)); - - assert(src0->nb[0] == sizeof(ggml_fp16_t)); - - GGML_TENSOR_LOCALS(int64_t, ne0, src0, ne) - GGML_TENSOR_LOCALS(size_t, nb0, src0, nb) - - float sum = 0; - float row_sum = 0; - - for (int64_t i03 = 0; i03 < ne03; i03++) { - for (int64_t i02 = 0; i02 < ne02; i02++) { - for (int64_t i01 = 0; i01 < ne01; i01++) { - ggml_vec_sum_f16_ggf(ne00, - &row_sum, - (ggml_fp16_t *) ((char *) src0->data + i01 * nb01 + i02 * nb02 + i03 * nb03)); - sum += row_sum; - } - } - } - ((ggml_fp16_t *) dst->data)[0] = GGML_FP32_TO_FP16(sum); -} - -static void ggml_compute_forward_sum_bf16( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - if (params->ith != 0) { - return; - } - - assert(ggml_is_scalar(dst)); - - assert(src0->nb[0] == sizeof(ggml_bf16_t)); - - GGML_TENSOR_LOCALS(int64_t, ne0, src0, ne) - GGML_TENSOR_LOCALS(size_t, nb0, src0, nb) - - float sum = 0; - float row_sum = 0; - - for (int64_t i03 = 0; i03 < ne03; i03++) { - for (int64_t i02 = 0; i02 < ne02; i02++) { - for (int64_t i01 = 0; i01 < ne01; i01++) { - ggml_vec_sum_bf16_ggf(ne00, - &row_sum, - (ggml_bf16_t *) ((char *) src0->data + i01 * nb01 + i02 * nb02 + i03 * nb03)); - sum += row_sum; - } - } - } - ((ggml_bf16_t *) dst->data)[0] = GGML_FP32_TO_BF16(sum); -} - -static void ggml_compute_forward_sum( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - switch (src0->type) { - case GGML_TYPE_F32: - { - ggml_compute_forward_sum_f32(params, dst); - } break; - case GGML_TYPE_F16: - { - ggml_compute_forward_sum_f16(params, dst); - } break; - case GGML_TYPE_BF16: - { - ggml_compute_forward_sum_bf16(params, dst); - } break; - default: - { - GGML_ABORT("fatal error"); - } - } -} - -// ggml_compute_forward_sum_rows - -static void ggml_compute_forward_sum_rows_f32( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - if (params->ith != 0) { - return; - } - - GGML_ASSERT(src0->nb[0] == sizeof(float)); - GGML_ASSERT(dst->nb[0] == sizeof(float)); - - GGML_TENSOR_UNARY_OP_LOCALS - - GGML_ASSERT(ne0 == 1); - GGML_ASSERT(ne1 == ne01); - GGML_ASSERT(ne2 == ne02); - GGML_ASSERT(ne3 == ne03); - - for (int64_t i3 = 0; i3 < ne03; i3++) { - for (int64_t i2 = 0; i2 < ne02; i2++) { - for (int64_t i1 = 0; i1 < ne01; i1++) { - float * src_row = (float *) ((char *) src0->data + i1*nb01 + i2*nb02 + i3*nb03); - float * dst_row = (float *) ((char *) dst->data + i1*nb1 + i2*nb2 + i3*nb3); - float row_sum = 0; - ggml_vec_sum_f32(ne00, &row_sum, src_row); - dst_row[0] = row_sum; - } - } - } -} - -static void ggml_compute_forward_sum_rows( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - switch (src0->type) { - case GGML_TYPE_F32: - { - ggml_compute_forward_sum_rows_f32(params, dst); - } break; - default: - { - GGML_ABORT("fatal error"); - } - } -} - -// ggml_compute_forward_mean - -static void ggml_compute_forward_mean_f32( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - if (params->ith != 0) { - return; - } - - assert(src0->nb[0] == sizeof(float)); - - GGML_TENSOR_UNARY_OP_LOCALS - - assert(ne0 == 1); - assert(ne1 == ne01); - assert(ne2 == ne02); - assert(ne3 == ne03); - - UNUSED(ne0); - UNUSED(ne1); - UNUSED(ne2); - UNUSED(ne3); - - for (int64_t i03 = 0; i03 < ne03; i03++) { - for (int64_t i02 = 0; i02 < ne02; i02++) { - for (int64_t i01 = 0; i01 < ne01; i01++) { - ggml_vec_sum_f32(ne00, - (float *) ((char *) dst->data + i01*nb1 + i02*nb2 + i03*nb3), - (float *) ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03)); - - *(float *) ((char *) dst->data + i01*nb1 + i02*nb2 + i03*nb3) /= (float) ne00; - } - } - } -} - -static void ggml_compute_forward_mean( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - switch (src0->type) { - case GGML_TYPE_F32: - { - ggml_compute_forward_mean_f32(params, dst); - } break; - default: - { - GGML_ABORT("fatal error"); - } - } -} - -// ggml_compute_forward_argmax - -static void ggml_compute_forward_argmax_f32( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - if (params->ith != 0) { - return; - } - - assert(src0->nb[0] == sizeof(float)); - assert(dst->nb[0] == sizeof(float)); - - const int64_t ne00 = src0->ne[0]; - const int64_t ne01 = src0->ne[1]; - - const size_t nb01 = src0->nb[1]; - const size_t nb0 = dst->nb[0]; - - for (int64_t i1 = 0; i1 < ne01; i1++) { - float * src = (float *) ((char *) src0->data + i1*nb01); - int32_t * dst_ = (int32_t *) ((char *) dst->data + i1*nb0); - int v = 0; - ggml_vec_argmax_f32(ne00, &v, src); - dst_[0] = v; - } -} - -static void ggml_compute_forward_argmax( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - switch (src0->type) { - case GGML_TYPE_F32: - { - ggml_compute_forward_argmax_f32(params, dst); - } break; - default: - { - GGML_ABORT("fatal error"); - } - } -} - -// ggml_compute_forward_count_equal - -static void ggml_compute_forward_count_equal_i32( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - const struct ggml_tensor * src1 = dst->src[1]; - - GGML_TENSOR_BINARY_OP_LOCALS; - - GGML_ASSERT(src0->type == GGML_TYPE_I32); - GGML_ASSERT(src1->type == GGML_TYPE_I32); - GGML_ASSERT(ggml_are_same_shape(src0, src1)); - GGML_ASSERT(ggml_is_scalar(dst)); - GGML_ASSERT(dst->type == GGML_TYPE_I64); - - const int64_t nr = ggml_nrows(src0); - - const int ith = params->ith; - const int nth = params->nth; - - int64_t * sums = (int64_t *) params->wdata; - int64_t sum_thread = 0; - - // rows per thread - const int64_t dr = (nr + nth - 1)/nth; - - // row range for this thread - const int64_t ir0 = dr*ith; - const int64_t ir1 = MIN(ir0 + dr, nr); - - for (int64_t ir = ir0; ir < ir1; ++ir) { - const int64_t i03 = ir / (ne02*ne01); - const int64_t i02 = (ir - i03*ne03) / ne01; - const int64_t i01 = ir - i03*ne03 - i02*ne02; - - const char * data0 = (const char *) src0->data + i03*nb03 + i02*nb02 + i01*nb01; - const char * data1 = (const char *) src1->data + i03*nb13 + i02*nb12 + i01*nb11; - - for (int64_t i00 = 0; i00 < ne00; ++i00) { - const int32_t val0 = *((const int32_t *) (data0 + i00*nb00)); - const int32_t val1 = *((const int32_t *) (data1 + i00*nb10)); - - sum_thread += val0 == val1; - } - } - if (ith != 0) { - sums[ith] = sum_thread; - } - ggml_barrier(params->threadpool); - - if (ith != 0) { - return; - } - - for (int ith_other = 1; ith_other < nth; ++ith_other) { - sum_thread += sums[ith_other]; - } - *((int64_t *) dst->data) = sum_thread; -} - -static void ggml_compute_forward_count_equal( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - switch (src0->type) { - case GGML_TYPE_I32: - { - ggml_compute_forward_count_equal_i32(params, dst); - } break; - default: - { - GGML_ABORT("fatal error"); - } - } -} - -// ggml_compute_forward_repeat - -static void ggml_compute_forward_repeat_f32( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - if (params->ith != 0) { - return; - } - - GGML_ASSERT(ggml_can_repeat(src0, dst)); - - GGML_TENSOR_UNARY_OP_LOCALS - - // guaranteed to be an integer due to the check in ggml_can_repeat - const int nr0 = (int)(ne0/ne00); - const int nr1 = (int)(ne1/ne01); - const int nr2 = (int)(ne2/ne02); - const int nr3 = (int)(ne3/ne03); - - // TODO: support for transposed / permuted tensors - GGML_ASSERT(nb0 == sizeof(float)); - GGML_ASSERT(nb00 == sizeof(float)); - - // TODO: maybe this is not optimal? - for (int i3 = 0; i3 < nr3; i3++) { - for (int k3 = 0; k3 < ne03; k3++) { - for (int i2 = 0; i2 < nr2; i2++) { - for (int k2 = 0; k2 < ne02; k2++) { - for (int i1 = 0; i1 < nr1; i1++) { - for (int k1 = 0; k1 < ne01; k1++) { - for (int i0 = 0; i0 < nr0; i0++) { - ggml_vec_cpy_f32(ne00, - (float *) ((char *) dst->data + (i3*ne03 + k3)*nb3 + (i2*ne02 + k2)*nb2 + (i1*ne01 + k1)*nb1 + (i0*ne00)*nb0), - (float *) ((char *) src0->data + ( k3)*nb03 + ( k2)*nb02 + ( k1)*nb01)); - } - } - } - } - } - } - } -} - -static void ggml_compute_forward_repeat_f16( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - if (params->ith != 0) { - return; - } - - GGML_ASSERT(ggml_can_repeat(src0, dst)); - - GGML_TENSOR_UNARY_OP_LOCALS - - // guaranteed to be an integer due to the check in ggml_can_repeat - const int nr0 = (int)(ne0/ne00); - const int nr1 = (int)(ne1/ne01); - const int nr2 = (int)(ne2/ne02); - const int nr3 = (int)(ne3/ne03); - - // TODO: support for transposed / permuted tensors - GGML_ASSERT(nb0 == sizeof(ggml_fp16_t)); - GGML_ASSERT(nb00 == sizeof(ggml_fp16_t)); - - // TODO: maybe this is not optimal? - for (int i3 = 0; i3 < nr3; i3++) { - for (int k3 = 0; k3 < ne03; k3++) { - for (int i2 = 0; i2 < nr2; i2++) { - for (int k2 = 0; k2 < ne02; k2++) { - for (int i1 = 0; i1 < nr1; i1++) { - for (int k1 = 0; k1 < ne01; k1++) { - for (int i0 = 0; i0 < nr0; i0++) { - ggml_fp16_t * y = (ggml_fp16_t *) ((char *) dst->data + (i3*ne03 + k3)*nb3 + (i2*ne02 + k2)*nb2 + (i1*ne01 + k1)*nb1 + (i0*ne00)*nb0); - ggml_fp16_t * x = (ggml_fp16_t *) ((char *) src0->data + ( k3)*nb03 + ( k2)*nb02 + ( k1)*nb01); - // ggml_vec_cpy_f16(ne00, y, x) - for (int i = 0; i < ne00; ++i) { - y[i] = x[i]; - } - } - } - } - } - } - } - } -} - -static void ggml_compute_forward_repeat( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - switch (src0->type) { - case GGML_TYPE_F16: - case GGML_TYPE_BF16: - case GGML_TYPE_I16: - { - ggml_compute_forward_repeat_f16(params, dst); - } break; - case GGML_TYPE_F32: - case GGML_TYPE_I32: - { - ggml_compute_forward_repeat_f32(params, dst); - } break; - default: - { - GGML_ABORT("fatal error"); - } - } -} - -// ggml_compute_forward_repeat_back - -static void ggml_compute_forward_repeat_back_f32( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - if (params->ith != 0) { - return; - } - - GGML_ASSERT(ggml_can_repeat(dst, src0)); - - GGML_TENSOR_UNARY_OP_LOCALS - - // guaranteed to be an integer due to the check in ggml_can_repeat - const int nr0 = (int)(ne00/ne0); - const int nr1 = (int)(ne01/ne1); - const int nr2 = (int)(ne02/ne2); - const int nr3 = (int)(ne03/ne3); - - // TODO: support for transposed / permuted tensors - GGML_ASSERT(nb0 == sizeof(float)); - GGML_ASSERT(nb00 == sizeof(float)); - - if (ggml_is_contiguous(dst)) { - ggml_vec_set_f32(ne0*ne1*ne2*ne3, dst->data, 0); - } else { - for (int k3 = 0; k3 < ne3; k3++) { - for (int k2 = 0; k2 < ne2; k2++) { - for (int k1 = 0; k1 < ne1; k1++) { - ggml_vec_set_f32(ne0, - (float *) ((char *) dst->data + k1*nb1 + k2*nb2 + k3*nb3), - 0); - } - } - } - } - - // TODO: maybe this is not optimal? - for (int i3 = 0; i3 < nr3; i3++) { - for (int k3 = 0; k3 < ne3; k3++) { - for (int i2 = 0; i2 < nr2; i2++) { - for (int k2 = 0; k2 < ne2; k2++) { - for (int i1 = 0; i1 < nr1; i1++) { - for (int k1 = 0; k1 < ne1; k1++) { - for (int i0 = 0; i0 < nr0; i0++) { - ggml_vec_acc_f32(ne0, - (float *) ((char *) dst->data + ( k3)*nb3 + ( k2)*nb2 + ( k1)*nb1), - (float *) ((char *) src0->data + (i3*ne3 + k3)*nb03 + (i2*ne2 + k2)*nb02 + (i1*ne1 + k1)*nb01 + (i0*ne0)*nb00)); - } - } - } - } - } - } - } -} - -static void ggml_compute_forward_repeat_back( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - switch (src0->type) { - case GGML_TYPE_F32: - { - ggml_compute_forward_repeat_back_f32(params, dst); - } break; - default: - { - GGML_ABORT("fatal error"); - } - } -} - -// ggml_compute_forward_concat - -static void ggml_compute_forward_concat_f32( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - const struct ggml_tensor * src1 = dst->src[1]; - - GGML_ASSERT(src0->nb[0] == sizeof(float)); - - const int ith = params->ith; - const int nth = params->nth; - - GGML_TENSOR_BINARY_OP_LOCALS - - const int32_t dim = ggml_get_op_params_i32(dst, 0); - - GGML_ASSERT(dim >= 0 && dim < 4); - - int64_t o[4] = {0, 0, 0, 0}; - o[dim] = src0->ne[dim]; - - const float * x; - - // TODO: smarter multi-theading - for (int i3 = 0; i3 < ne3; i3++) { - for (int i2 = ith; i2 < ne2; i2 += nth) { - for (int i1 = 0; i1 < ne1; i1++) { - for (int i0 = 0; i0 < ne0; i0++) { - if (i0 < ne00 && i1 < ne01 && i2 < ne02 && i3 < ne03) { - x = (const float *) ((const char *)src0->data + (i0 )*nb00 + (i1 )*nb01 + (i2 )*nb02 + (i3 )*nb03); - } else { - x = (const float *) ((const char *)src1->data + (i0 - o[0])*nb10 + (i1 - o[1])*nb11 + (i2 - o[2])*nb12 + (i3 - o[3])*nb13); - } - - float * y = (float *)((char *)dst->data + i0*nb0 + i1*nb1 + i2*nb2 + i3*nb3); - - *y = *x; - } - } - } - } -} - -static void ggml_compute_forward_concat( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - switch (src0->type) { - case GGML_TYPE_F32: - case GGML_TYPE_I32: - { - ggml_compute_forward_concat_f32(params, dst); - } break; - default: - { - GGML_ABORT("fatal error"); - } - } -} - -// ggml_compute_forward_abs - -static void ggml_compute_forward_abs_f32( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - if (params->ith != 0) { - return; - } - - assert(ggml_is_contiguous_1(src0)); - assert(ggml_is_contiguous_1(dst)); - assert(ggml_are_same_shape(src0, dst)); - - const int n = ggml_nrows(src0); - const int nc = src0->ne[0]; - - for (int i = 0; i < n; i++) { - ggml_vec_abs_f32(nc, - (float *) ((char *) dst->data + i*( dst->nb[1])), - (float *) ((char *) src0->data + i*(src0->nb[1]))); - } -} - -static void ggml_compute_forward_abs( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - switch (src0->type) { - case GGML_TYPE_F32: - { - ggml_compute_forward_abs_f32(params, dst); - } break; - default: - { - GGML_ABORT("fatal error"); - } - } -} - -// ggml_compute_forward_sgn - -static void ggml_compute_forward_sgn_f32( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - if (params->ith != 0) { - return; - } - - assert(ggml_is_contiguous_1(src0)); - assert(ggml_is_contiguous_1(dst)); - assert(ggml_are_same_shape(src0, dst)); - - const int n = ggml_nrows(src0); - const int nc = src0->ne[0]; - - for (int i = 0; i < n; i++) { - ggml_vec_sgn_f32(nc, - (float *) ((char *) dst->data + i*( dst->nb[1])), - (float *) ((char *) src0->data + i*(src0->nb[1]))); - } -} - -static void ggml_compute_forward_sgn( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - switch (src0->type) { - case GGML_TYPE_F32: - { - ggml_compute_forward_sgn_f32(params, dst); - } break; - default: - { - GGML_ABORT("fatal error"); - } - } -} - -// ggml_compute_forward_neg - -static void ggml_compute_forward_neg_f32( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - if (params->ith != 0) { - return; - } - - assert(ggml_is_contiguous_1(src0)); - assert(ggml_is_contiguous_1(dst)); - assert(ggml_are_same_shape(src0, dst)); - - const int n = ggml_nrows(src0); - const int nc = src0->ne[0]; - - for (int i = 0; i < n; i++) { - ggml_vec_neg_f32(nc, - (float *) ((char *) dst->data + i*( dst->nb[1])), - (float *) ((char *) src0->data + i*(src0->nb[1]))); - } -} - -static void ggml_compute_forward_neg( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - switch (src0->type) { - case GGML_TYPE_F32: - { - ggml_compute_forward_neg_f32(params, dst); - } break; - default: - { - GGML_ABORT("fatal error"); - } - } -} - -// ggml_compute_forward_step - -static void ggml_compute_forward_step_f32( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - if (params->ith != 0) { - return; - } - - assert(ggml_is_contiguous_1(src0)); - assert(ggml_is_contiguous_1(dst)); - assert(ggml_are_same_shape(src0, dst)); - - const int n = ggml_nrows(src0); - const int nc = src0->ne[0]; - - for (int i = 0; i < n; i++) { - ggml_vec_step_f32(nc, - (float *) ((char *) dst->data + i*( dst->nb[1])), - (float *) ((char *) src0->data + i*(src0->nb[1]))); - } -} - -static void ggml_compute_forward_step( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - switch (src0->type) { - case GGML_TYPE_F32: - { - ggml_compute_forward_step_f32(params, dst); - } break; - default: - { - GGML_ABORT("fatal error"); - } - } -} - -// ggml_compute_forward_tanh - -static void ggml_compute_forward_tanh_f32( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - if (params->ith != 0) { - return; - } - - assert(ggml_is_contiguous_1(src0)); - assert(ggml_is_contiguous_1(dst)); - assert(ggml_are_same_shape(src0, dst)); - - const int n = ggml_nrows(src0); - const int nc = src0->ne[0]; - - for (int i = 0; i < n; i++) { - ggml_vec_tanh_f32(nc, - (float *) ((char *) dst->data + i*( dst->nb[1])), - (float *) ((char *) src0->data + i*(src0->nb[1]))); - } -} - -static void ggml_compute_forward_tanh( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - switch (src0->type) { - case GGML_TYPE_F32: - { - ggml_compute_forward_tanh_f32(params, dst); - } break; - default: - { - GGML_ABORT("fatal error"); - } - } -} - -// ggml_compute_forward_elu - -static void ggml_compute_forward_elu_f32( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - if (params->ith != 0) { - return; - } - - assert(ggml_is_contiguous_1(src0)); - assert(ggml_is_contiguous_1(dst)); - assert(ggml_are_same_shape(src0, dst)); - - const int n = ggml_nrows(src0); - const int nc = src0->ne[0]; - - for (int i = 0; i < n; i++) { - ggml_vec_elu_f32(nc, - (float *) ((char *) dst->data + i*( dst->nb[1])), - (float *) ((char *) src0->data + i*(src0->nb[1]))); - } -} - -static void ggml_compute_forward_elu( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - switch (src0->type) { - case GGML_TYPE_F32: - { - ggml_compute_forward_elu_f32(params, dst); - } break; - default: - { - GGML_ABORT("fatal error"); - } - } -} - -// ggml_compute_forward_relu - -static void ggml_compute_forward_relu_f32( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - if (params->ith != 0) { - return; - } - - assert(ggml_is_contiguous_1(src0)); - assert(ggml_is_contiguous_1(dst)); - assert(ggml_are_same_shape(src0, dst)); - - const int n = ggml_nrows(src0); - const int nc = src0->ne[0]; - - for (int i = 0; i < n; i++) { - ggml_vec_relu_f32(nc, - (float *) ((char *) dst->data + i*( dst->nb[1])), - (float *) ((char *) src0->data + i*(src0->nb[1]))); - } -} - -static void ggml_compute_forward_relu( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - switch (src0->type) { - case GGML_TYPE_F32: - { - ggml_compute_forward_relu_f32(params, dst); - } break; - default: - { - GGML_ABORT("fatal error"); - } - } -} - -// ggml_compute_forward_sigmoid - -static void ggml_compute_forward_sigmoid_f32( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - if (params->ith != 0) { - return; - } - - assert(ggml_is_contiguous_1(src0)); - assert(ggml_is_contiguous_1(dst)); - assert(ggml_are_same_shape(src0, dst)); - - const int n = ggml_nrows(src0); - const int nc = src0->ne[0]; - - for (int i = 0; i < n; i++) { - ggml_vec_sigmoid_f32(nc, - (float *) ((char *) dst->data + i*( dst->nb[1])), - (float *) ((char *) src0->data + i*(src0->nb[1]))); - } -} - -static void ggml_compute_forward_sigmoid( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - switch (src0->type) { - case GGML_TYPE_F32: - { - ggml_compute_forward_sigmoid_f32(params, dst); - } break; - default: - { - GGML_ABORT("fatal error"); - } - } -} - -// ggml_compute_forward_gelu - -static void ggml_compute_forward_gelu_f32( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - assert(ggml_is_contiguous_1(src0)); - assert(ggml_is_contiguous_1(dst)); - assert(ggml_are_same_shape(src0, dst)); - - const int ith = params->ith; - const int nth = params->nth; - - const int nc = src0->ne[0]; - const int nr = ggml_nrows(src0); - - // rows per thread - const int dr = (nr + nth - 1)/nth; - - // row range for this thread - const int ir0 = dr*ith; - const int ir1 = MIN(ir0 + dr, nr); - - for (int i1 = ir0; i1 < ir1; i1++) { - ggml_vec_gelu_f32(nc, - (float *) ((char *) dst->data + i1*( dst->nb[1])), - (float *) ((char *) src0->data + i1*(src0->nb[1]))); - -#ifndef NDEBUG - for (int k = 0; k < nc; k++) { - const float x = ((float *) ((char *) dst->data + i1*( dst->nb[1])))[k]; - UNUSED(x); - assert(!isnan(x)); - assert(!isinf(x)); - } -#endif - } -} - -static void ggml_compute_forward_gelu( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - switch (src0->type) { - case GGML_TYPE_F32: - { - ggml_compute_forward_gelu_f32(params, dst); - } break; - default: - { - GGML_ABORT("fatal error"); - } - } -} - -// ggml_compute_forward_gelu_quick - -static void ggml_compute_forward_gelu_quick_f32( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - assert(ggml_is_contiguous_1(src0)); - assert(ggml_is_contiguous_1(dst)); - assert(ggml_are_same_shape(src0, dst)); - - const int ith = params->ith; - const int nth = params->nth; - - const int nc = src0->ne[0]; - const int nr = ggml_nrows(src0); - - // rows per thread - const int dr = (nr + nth - 1)/nth; - - // row range for this thread - const int ir0 = dr*ith; - const int ir1 = MIN(ir0 + dr, nr); - - for (int i1 = ir0; i1 < ir1; i1++) { - ggml_vec_gelu_quick_f32(nc, - (float *) ((char *) dst->data + i1*( dst->nb[1])), - (float *) ((char *) src0->data + i1*(src0->nb[1]))); - -#ifndef NDEBUG - for (int k = 0; k < nc; k++) { - const float x = ((float *) ((char *) dst->data + i1*( dst->nb[1])))[k]; - UNUSED(x); - assert(!isnan(x)); - assert(!isinf(x)); - } -#endif - } -} - -static void ggml_compute_forward_gelu_quick( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - switch (src0->type) { - case GGML_TYPE_F32: - { - ggml_compute_forward_gelu_quick_f32(params, dst); - } break; - default: - { - GGML_ABORT("fatal error"); - } - } -} - -// ggml_compute_forward_silu - -static void ggml_compute_forward_silu_f32( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - assert(ggml_is_contiguous_1(src0)); - assert(ggml_is_contiguous_1(dst)); - assert(ggml_are_same_shape(src0, dst)); - - const int ith = params->ith; - const int nth = params->nth; - - const int nc = src0->ne[0]; - const int nr = ggml_nrows(src0); - - // rows per thread - const int dr = (nr + nth - 1)/nth; - - // row range for this thread - const int ir0 = dr*ith; - const int ir1 = MIN(ir0 + dr, nr); - - for (int i1 = ir0; i1 < ir1; i1++) { - ggml_vec_silu_f32(nc, - (float *) ((char *) dst->data + i1*( dst->nb[1])), - (float *) ((char *) src0->data + i1*(src0->nb[1]))); - -#ifndef NDEBUG - for (int k = 0; k < nc; k++) { - const float x = ((float *) ((char *) dst->data + i1*(dst->nb[1])))[k]; - UNUSED(x); - assert(!isnan(x)); - assert(!isinf(x)); - } -#endif - } -} - -static void ggml_compute_forward_silu( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - switch (src0->type) { - case GGML_TYPE_F32: - { - ggml_compute_forward_silu_f32(params, dst); - } break; - default: - { - GGML_ABORT("fatal error"); - } - } -} -// ggml_compute_forward_leaky_relu - -static void ggml_compute_forward_leaky_relu_f32( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - if (params->ith != 0) { - return; - } - - assert(ggml_is_contiguous_1(src0)); - assert(ggml_is_contiguous_1(dst)); - assert(ggml_are_same_shape(src0, dst)); - - const int n = ggml_nrows(src0); - const int nc = src0->ne[0]; - - float negative_slope; - memcpy(&negative_slope, dst->op_params, sizeof(float)); - - assert(dst->nb[0] == sizeof(float)); - assert(src0->nb[0] == sizeof(float)); - - for (int i = 0; i < n; i++) { - ggml_vec_leaky_relu_f32(nc, - (float *) ((char *) dst->data + i*( dst->nb[1])), - (float *) ((char *) src0->data + i*(src0->nb[1])), negative_slope); - } -} - -static void ggml_compute_forward_leaky_relu( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - switch (src0->type) { - case GGML_TYPE_F32: - { - ggml_compute_forward_leaky_relu_f32(params, dst); - } break; - default: - { - GGML_ABORT("fatal error"); - } - } -} - -// ggml_compute_forward_silu_back - -static void ggml_compute_forward_silu_back_f32( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * grad = dst->src[0]; - const struct ggml_tensor * src1 = dst->src[1]; - - assert(ggml_is_contiguous_1(grad)); - assert(ggml_is_contiguous_1(src1)); - assert(ggml_is_contiguous_1(dst)); - assert(ggml_are_same_shape(src1, dst)); - assert(ggml_are_same_shape(src1, grad)); - - const int ith = params->ith; - const int nth = params->nth; - - const int nc = src1->ne[0]; - const int nr = ggml_nrows(src1); - - // rows per thread - const int dr = (nr + nth - 1)/nth; - - // row range for this thread - const int ir0 = dr*ith; - const int ir1 = MIN(ir0 + dr, nr); - - for (int i1 = ir0; i1 < ir1; i1++) { - ggml_vec_silu_backward_f32(nc, - (float *) ((char *) dst->data + i1*( dst->nb[1])), - (float *) ((char *) src1->data + i1*(src1->nb[1])), - (float *) ((char *) grad->data + i1*(grad->nb[1]))); - -#ifndef NDEBUG - for (int k = 0; k < nc; k++) { - const float x = ((float *) ((char *) dst->data + i1*( dst->nb[1])))[k]; - UNUSED(x); - assert(!isnan(x)); - assert(!isinf(x)); - } -#endif - } -} - -static void ggml_compute_forward_silu_back( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - switch (src0->type) { - case GGML_TYPE_F32: - { - ggml_compute_forward_silu_back_f32(params, dst); - } break; - default: - { - GGML_ABORT("fatal error"); - } - } -} - - -static void ggml_compute_forward_hardswish_f32( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - if (params->ith != 0) { - return; - } - - assert(ggml_is_contiguous_1(src0)); - assert(ggml_is_contiguous_1(dst)); - assert(ggml_are_same_shape(src0, dst)); - - const int n = ggml_nrows(src0); - const int nc = src0->ne[0]; - - for (int i = 0; i < n; i++) { - ggml_vec_hardswish_f32(nc, - (float *) ((char *) dst->data + i*( dst->nb[1])), - (float *) ((char *) src0->data + i*(src0->nb[1]))); - } -} -static void ggml_compute_forward_hardswish( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - switch (src0->type) { - case GGML_TYPE_F32: - { - ggml_compute_forward_hardswish_f32(params, dst); - } break; - default: - { - GGML_ABORT("fatal error"); - } - } -} - -static void ggml_compute_forward_hardsigmoid_f32( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - if (params->ith != 0) { - return; - } - - assert(ggml_is_contiguous_1(src0)); - assert(ggml_is_contiguous_1(dst)); - assert(ggml_are_same_shape(src0, dst)); - - const int n = ggml_nrows(src0); - const int nc = src0->ne[0]; - - for (int i = 0; i < n; i++) { - ggml_vec_hardsigmoid_f32(nc, - (float *) ((char *) dst->data + i*( dst->nb[1])), - (float *) ((char *) src0->data + i*(src0->nb[1]))); - } -} - -static void ggml_compute_forward_hardsigmoid( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - switch (src0->type) { - case GGML_TYPE_F32: - { - ggml_compute_forward_hardsigmoid_f32(params, dst); - } break; - default: - { - GGML_ABORT("fatal error"); - } - } -} - -static void ggml_compute_forward_exp_f32( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - if (params->ith != 0) { - return; - } - - assert(ggml_is_contiguous_1(src0)); - assert(ggml_is_contiguous_1(dst)); - assert(ggml_are_same_shape(src0, dst)); - - const int n = ggml_nrows(src0); - const int nc = src0->ne[0]; - - for (int i = 0; i < n; i++) { - ggml_vec_exp_f32(nc, - (float *) ((char *) dst->data + i*( dst->nb[1])), - (float *) ((char *) src0->data + i*(src0->nb[1]))); - } -} - -static void ggml_compute_forward_exp( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - switch (src0->type) { - case GGML_TYPE_F32: - { - ggml_compute_forward_exp_f32(params, dst); - } break; - default: - { - GGML_ABORT("fatal error"); - } - } -} - - -// ggml_compute_forward_norm - -static void ggml_compute_forward_norm_f32( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - GGML_ASSERT(ggml_are_same_shape(src0, dst)); - - GGML_ASSERT(src0->nb[0] == sizeof(float)); - - const int ith = params->ith; - const int nth = params->nth; - - GGML_TENSOR_UNARY_OP_LOCALS - - float eps; - memcpy(&eps, dst->op_params, sizeof(float)); - - GGML_ASSERT(eps >= 0.0f); - - // TODO: optimize - for (int64_t i03 = 0; i03 < ne03; i03++) { - for (int64_t i02 = 0; i02 < ne02; i02++) { - for (int64_t i01 = ith; i01 < ne01; i01 += nth) { - const float * x = (float *) ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03); - - ggml_float sum = 0.0; - for (int64_t i00 = 0; i00 < ne00; i00++) { - sum += (ggml_float)x[i00]; - } - - float mean = sum/ne00; - - float * y = (float *) ((char *) dst->data + i01*nb1 + i02*nb2 + i03*nb3); - - ggml_float sum2 = 0.0; - for (int64_t i00 = 0; i00 < ne00; i00++) { - float v = x[i00] - mean; - y[i00] = v; - sum2 += (ggml_float)(v*v); - } - - float variance = sum2/ne00; - const float scale = 1.0f/sqrtf(variance + eps); - - ggml_vec_scale_f32(ne00, y, scale); - } - } - } -} - -static void ggml_compute_forward_norm( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - switch (src0->type) { - case GGML_TYPE_F32: - { - ggml_compute_forward_norm_f32(params, dst); - } break; - default: - { - GGML_ABORT("fatal error"); - } - } -} - -// ggml_compute_forward_group_rms_norm - -static void ggml_compute_forward_rms_norm_f32( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - GGML_ASSERT(ggml_are_same_shape(src0, dst)); - - GGML_ASSERT(src0->nb[0] == sizeof(float)); - - const int ith = params->ith; - const int nth = params->nth; - - GGML_TENSOR_UNARY_OP_LOCALS - - float eps; - memcpy(&eps, dst->op_params, sizeof(float)); - - GGML_ASSERT(eps >= 0.0f); - - // TODO: optimize - for (int64_t i03 = 0; i03 < ne03; i03++) { - for (int64_t i02 = 0; i02 < ne02; i02++) { - for (int64_t i01 = ith; i01 < ne01; i01 += nth) { - const float * x = (float *) ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03); - - ggml_float sum = 0.0; - for (int64_t i00 = 0; i00 < ne00; i00++) { - sum += (ggml_float)(x[i00] * x[i00]); - } - - const float mean = sum/ne00; - - float * y = (float *) ((char *) dst->data + i01*nb1 + i02*nb2 + i03*nb3); - - memcpy(y, x, ne00 * sizeof(float)); - // for (int i00 = 0; i00 < ne00; i00++) { - // y[i00] = x[i00]; - // } - - const float scale = 1.0f/sqrtf(mean + eps); - - ggml_vec_scale_f32(ne00, y, scale); - } - } - } -} - -static void ggml_compute_forward_rms_norm( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - switch (src0->type) { - case GGML_TYPE_F32: - { - ggml_compute_forward_rms_norm_f32(params, dst); - } break; - default: - { - GGML_ABORT("fatal error"); - } - } -} - -static void ggml_compute_forward_rms_norm_back_f32( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; // gradients from forward pass output - const struct ggml_tensor * src1 = dst->src[1]; // src1 from forward pass - - GGML_ASSERT(ggml_are_same_shape(src0, dst) && ggml_are_same_shape(src0, src1)); - - GGML_ASSERT(src0->nb[0] == sizeof(float)); - GGML_ASSERT(src1->nb[0] == sizeof(float)); - - const int ith = params->ith; - const int nth = params->nth; - - GGML_TENSOR_BINARY_OP_LOCALS - - float eps; - memcpy(&eps, dst->op_params, sizeof(float)); - - // TODO: optimize - for (int64_t i03 = 0; i03 < ne03; i03++) { - for (int64_t i02 = 0; i02 < ne02; i02++) { - for (int64_t i01 = ith; i01 < ne01; i01 += nth) { - // src1 is same shape as src0 => same indices - const int64_t i11 = i01; - const int64_t i12 = i02; - const int64_t i13 = i03; - - const float * dz = (float *) ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03); - const float * x = (float *) ((char *) src1->data + i11*nb11 + i12*nb12 + i13*nb13); - - ggml_float sum_xx = 0.0; - ggml_float sum_xdz = 0.0; - - for (int64_t i00 = 0; i00 < ne00; i00++) { - sum_xx += (ggml_float)(x[i00] * x[i00]); - sum_xdz += (ggml_float)(x[i00] * dz[i00]); - } - - //const float mean = (float)(sum_xx)/ne00; - const float mean_eps = (float)(sum_xx)/ne00 + eps; - const float sum_eps = (float)(sum_xx) + eps*ne00; - //const float mean_xdz = (float)(sum_xdz)/ne00; - // we could cache rms from forward pass to improve performance. - // to do this implement ggml_rms and compose ggml_rms_norm using ggml_rms. - //const float rms = sqrtf(mean_eps); - const float rrms = 1.0f / sqrtf(mean_eps); - //const float scale = -rrms/(ne00 * mean_eps); // -1/(n*rms**3) - - { - // z = rms_norm(x) - // - // rms_norm(src1) = - // scale( - // src1, - // div( - // 1, - // sqrt( - // add( - // scale( - // sum( - // sqr( - // src1)), - // (1.0/N)), - // eps)))); - - // postorder: - // ## op args grad - // 00 param src1 grad[#00] - // 01 const 1 - // 02 sqr (#00) grad[#02] - // 03 sum (#02) grad[#03] - // 04 const 1/N - // 05 scale (#03, #04) grad[#05] - // 06 const eps - // 07 add (#05, #06) grad[#07] - // 08 sqrt (#07) grad[#08] - // 09 div (#01,#08) grad[#09] - // 10 scale (#00,#09) grad[#10] - // - // backward pass, given grad[#10] - // #10: scale - // grad[#00] += scale(grad[#10],#09) - // grad[#09] += sum(mul(grad[#10],#00)) - // #09: div - // grad[#08] += neg(mul(grad[#09], div(#09,#08))) - // #08: sqrt - // grad[#07] += mul(grad[#08], div(0.5, #08)) - // #07: add - // grad[#05] += grad[#07] - // #05: scale - // grad[#03] += scale(grad[#05],#04) - // #03: sum - // grad[#02] += repeat(grad[#03], #02) - // #02: - // grad[#00] += scale(mul(#00, grad[#02]), 2.0) - // - // substitute and simplify: - // grad[#00] = scale(grad(#10), #09) + scale(mul(#00, grad[#02]), 2.0) - // grad[#02] = repeat(grad[#03], #02) - // grad[#02] = repeat(scale(grad[#05],#04), #02) - // grad[#02] = repeat(scale(grad[#07],#04), #02) - // grad[#02] = repeat(scale(mul(grad[#08], div(0.5, #08)),#04), #02) - // grad[#02] = repeat(scale(mul(neg(mul(grad[#09], div(#09,#08))), div(0.5, #08)),#04), #02) - // grad[#02] = repeat(scale(mul(neg(mul(sum(mul(grad[#10],#00)), div(#09,#08))), div(0.5, #08)),#04), #02) - // grad[#02] = repeat(-(sum(mul(grad[#10],#00)) * div(#09,#08) * div(0.5, #08) * (1/N)), #02) - // grad[#02] = repeat(-(sum(mul(grad[#10],#00)) * div(div(#01,#08),#08) * div(0.5, #08) * (1/N)), #02) - // grad[#02] = repeat(-(sum(mul(grad[#10],#00)) * div(1,#08*#08) * div(0.5, #08) * (1/N)), #02) - // grad[#02] = repeat(-(sum(mul(grad[#10],#00)) * div(1,#07) * div(0.5, #08) * (1/N)), #02) - // grad[#00] = scale(grad(#10), #09) + scale(mul(#00, grad[#02]), 2.0) - // grad[#00] = scale(grad(#10), #09) + scale(mul(#00, repeat(-(sum(mul(grad[#10],#00)) * div(1,#07) * div(0.5, #08) * (1/N)), #02)), 2.0) - // grad[#00] = scale(grad(#10), #09) + scale(scale(#00, -(sum(mul(grad[#10],#00)) * div(1,#07) * div(0.5, #08) * (1/N))), 2.0) - // grad[#00] = scale(grad(#10), #09) + scale(#00, -(sum(mul(grad[#10],#00)) * div(1,#07) * div(1,#08) * (1/N))) - // grad[#00] = scale(grad(#10), #09) + scale(#00, sum(mul(grad[#10],#00)) * div(1,#07*#08) * (-1/N)) - // grad[#00] = scale(grad(#10), #09) + scale(#00, sum(mul(grad[#10],#00)) * div(1,#07*#08) * (-1/N)) - // grad[#00] = scale(grad(#10), #09) + scale(#00, sum(mul(grad[#10],#00)) * div(1,mean_eps*rms) * (-1/N)) - // grad[#00] = scale(grad(#10), #09) + scale(#00, sum(mul(grad[#10],#00)) * div(-1,rms*N*mean_eps)) - // grad[#00] = scale(grad(#10), #09) + scale(#00, sum(mul(grad[#10],#00)) * div(-1,rms*N*(sum_xx/N+eps))) - // grad[#00] = scale(grad(#10), #09) + scale(#00, sum(mul(grad[#10],#00)) * div(-1,rms*N*sum_xx+rms*N*eps)) - // grad[#00] = scale(dz, rrms) + scale(x, sum(mul(dz,x)) * div(-1,rms*N*mean_eps)) - // grad[#00] = scale(dz, rrms) + scale(x, sum_xdz * div(-1,rms*N*mean_eps)) - // a = b*c + d*e - // a = b*c*f/f + d*e*f/f - // a = (b*c*f + d*e*f)*(1/f) - // a = (b*c*(1/c) + d*e*(1/c))*(1/(1/c)) - // a = (b + d*e/c)*c - // b = dz, c = rrms, d = x, e = sum_xdz * div(-1,rms*N*mean_eps) - // a = (dz + x*sum_xdz * div(-1,rms*N*mean_eps)/rrms)*rrms - // a = (dz + x*sum_xdz * div(-1,rms*N*mean_eps)*rms)*rrms - // a = (dz + x*sum_xdz * div(-rms,rms*N*mean_eps))*rrms - // a = (dz + x*sum_xdz * div(-1,N*mean_eps))*rrms - // a = (dz + x*div(-sum_xdz,N*mean_eps))*rrms - // a = (dz + x*div(-mean_xdz,mean_eps))*rrms - // grad[#00] = scale(dz + scale(x, div(-mean_xdz,mean_eps)),rrms) - // grad[#00] = scale(dz + scale(x, -mean_xdz/mean_eps),rrms) - // dx = scale(dz + scale(x, -mean_xdz/mean_eps),rrms) - } - // dx = scale(dz + scale(x, -mean_xdz/mean_eps),rrms) - // post-order: - // dx := x - // dx := scale(dx,-mean_xdz/mean_eps) - // dx := add(dx, dz) - // dx := scale(dx, rrms) - float * dx = (float *) ((char *) dst->data + i01*nb1 + i02*nb2 + i03*nb3); - - // dx[i00] = (x*(-sum_xdz/sum_eps) + dz) / sqrtf(mean_eps) - ggml_vec_cpy_f32 (ne00, dx, x); - // ggml_vec_scale_f32(ne00, dx, -mean_xdz/mean_eps); - ggml_vec_scale_f32(ne00, dx, (float)(-sum_xdz)/sum_eps); - ggml_vec_acc_f32 (ne00, dx, dz); - ggml_vec_scale_f32(ne00, dx, rrms); - } - } - } -} - -static void ggml_compute_forward_rms_norm_back( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - switch (src0->type) { - case GGML_TYPE_F32: - { - ggml_compute_forward_rms_norm_back_f32(params, dst); - } break; - default: - { - GGML_ABORT("fatal error"); - } - } -} - -// ggml_compute_forward_group_norm - -static void ggml_compute_forward_group_norm_f32( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - GGML_ASSERT(ggml_are_same_shape(src0, dst)); - - GGML_ASSERT(src0->nb[0] == sizeof(float)); - - const int ith = params->ith; - const int nth = params->nth; - - GGML_TENSOR_UNARY_OP_LOCALS - - // TODO: optimize - - float eps; - memcpy(&eps, dst->op_params + 1, sizeof(float)); - - int n_channels = src0->ne[2]; - int n_groups = dst->op_params[0]; - int n_channels_per_group = (n_channels + n_groups - 1) / n_groups; - for (int i = ith; i < n_groups; i += nth) { - int start = i * n_channels_per_group; - int end = start + n_channels_per_group; - if (end > n_channels) { - end = n_channels; - } - int step = end - start; - - for (int64_t i03 = 0; i03 < ne03; i03++) { - ggml_float sum = 0.0; - for (int64_t i02 = start; i02 < end; i02++) { - for (int64_t i01 = 0; i01 < ne01; i01++) { - const float * x = (float *)((char *) src0->data + i01 * nb01 + i02 * nb02 + i03 * nb03); - - ggml_float sumr = 0.0; - for (int64_t i00 = 0; i00 < ne00; i00++) { - sumr += (ggml_float)x[i00]; - } - sum += sumr; - } - } - const float mean = sum / (ne00 * ne01 * step); - - ggml_float sum2 = 0.0; - for (int64_t i02 = start; i02 < end; i02++) { - for (int64_t i01 = 0; i01 < ne01; i01++) { - const float * x = (float *)((char *) src0->data + i01 * nb01 + i02 * nb02 + i03 * nb03); - - float * y = (float *)((char *) dst->data + i01 * nb1 + i02 * nb2 + i03 * nb3); - - ggml_float sumr = 0.0; - for (int64_t i00 = 0; i00 < ne00; i00++) { - float v = x[i00] - mean; - y[i00] = v; - sumr += (ggml_float)(v * v); - } - sum2 += sumr; - } - } - const float variance = sum2 / (ne00 * ne01 * step); - const float scale = 1.0f / sqrtf(variance + eps); - - for (int64_t i02 = start; i02 < end; i02++) { - for (int64_t i01 = 0; i01 < ne01; i01++) { - float * y = (float *)((char *) dst->data + i01 * nb1 + i02 * nb2 + i03 * nb3); - ggml_vec_scale_f32(ne00, y, scale); - } - } - } - } -} - -static void ggml_compute_forward_group_norm( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - switch (src0->type) { - case GGML_TYPE_F32: - { - ggml_compute_forward_group_norm_f32(params, dst); - } break; - default: - { - GGML_ABORT("fatal error"); - } - } -} - // ggml_compute_forward_mul_mat static void ggml_compute_forward_mul_mat_one_chunk( @@ -7490,6 +1328,7 @@ UseGgmlGemm1:; if (src1->type != vec_dot_type) { char * wdata = params->wdata; + const size_t nbw0 = ggml_type_size(vec_dot_type); const size_t nbw1 = ggml_row_size(vec_dot_type, ne10); const size_t nbw2 = nbw1*ne11; const size_t nbw3 = nbw2*ne12; @@ -7497,6 +1336,7 @@ UseGgmlGemm1:; assert(params->wsize >= ne13*nbw3); GGML_ASSERT(src1->type == GGML_TYPE_F32); + #if 0 for (int64_t i13 = 0; i13 < ne13; ++i13) { for (int64_t i12 = 0; i12 < ne12; ++i12) { for (int64_t i11 = ith; i11 < ne11; i11 += nth) { @@ -7506,6 +1346,20 @@ UseGgmlGemm1:; } } } + #else + for (int64_t i13 = 0; i13 < ne13; ++i13) { + for (int64_t i12 = 0; i12 < ne12; ++i12) { + for (int64_t i11 = 0; i11 < ne11; ++i11) { + size_t bs = ggml_blck_size(vec_dot_type); + int64_t ne10_block_start = (ith * ne10/bs) / nth; + int64_t ne10_block_end = ((ith + 1) * ne10/bs) / nth; + from_float((float *)((char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11 + ne10_block_start*bs*nb10), + (void *) (wdata + i13*nbw3 + i12*nbw2 + i11*nbw1 + ne10_block_start*nbw0), + (ne10_block_end - ne10_block_start) * bs); + } + } + } + #endif } if (ith == 0) { @@ -7560,7 +1414,7 @@ UseGgmlGemm2:; int64_t nchunk1 = (nr1 + chunk_size - 1) / chunk_size; // If the chunking is poor for the number of threads on this setup, scrap the whole plan. Re-chunk it by thread. - // Also, chunking by thread was measured to have perform better on NUMA systems. See https://github.com/ggerganov/llama.cpp/pull/6915 + // Also, chunking by thread was measured to have perform better on NUMA systems. See https://github.com/ggml-org/llama.cpp/pull/6915 // In theory, chunking should be just as useful on NUMA and non NUMA systems, but testing disagreed with that. if (nchunk0 * nchunk1 < nth * 4 || ggml_is_numa()) { // distribute the thread work across the inner or outer loop based on which one is larger @@ -7593,7 +1447,6 @@ UseGgmlGemm2:; if ((nr0 % 2 != 0) || (ne11 % 2 != 0) || ((ir0_end - ir0_start) % 2 != 0) || ((ir1_end - ir1_start) % 2 != 0)) { num_rows_per_vec_dot = 1; } - ggml_compute_forward_mul_mat_one_chunk(params, dst, src0->type, num_rows_per_vec_dot, ir0_start, ir0_end, ir1_start, ir1_end); if (nth >= nchunk0 * nchunk1) { @@ -7606,6 +1459,84 @@ UseGgmlGemm2:; // ggml_compute_forward_mul_mat_id +#define MMID_MATRIX_ROW(row_id, i1) matrix_rows[(row_id)*ids->ne[0]*ids->ne[1] + (i1)] + +struct mmid_row_mapping { + int32_t i1; + int32_t i2; +}; + +static void ggml_compute_forward_mul_mat_id_one_chunk( + struct ggml_tensor * dst, + const struct ggml_tensor * src0, + const struct ggml_tensor * src1, + const struct ggml_tensor * ids, + const int64_t cur_a, + const int64_t ir0_start, + const int64_t ir0_end, + const int64_t ir1_start, + const int64_t ir1_end, + const char * src0_cur, + const struct mmid_row_mapping * matrix_rows, + const size_t row_size, + const bool src1_cont, + const void * wdata) { + + GGML_TENSOR_BINARY_OP_LOCALS + + const enum ggml_type type = src0->type; + + ggml_vec_dot_t const vec_dot = type_traits_cpu[type].vec_dot; + enum ggml_type const vec_dot_type = type_traits_cpu[type].vec_dot_type; + + const int64_t blck_0 = 16; + const int64_t blck_1 = 16; + + float tmp[16]; + + for (int64_t iir1 = ir1_start; iir1 < ir1_end; iir1 += blck_1) { + for (int64_t iir0 = ir0_start; iir0 < ir0_end; iir0 += blck_0) { + for (int64_t ir1 = iir1; ir1 < iir1 + blck_1 && ir1 < ir1_end; ++ir1) { + const int64_t _i12 = ir1; // logical row index for this expert + + struct mmid_row_mapping row_mapping = MMID_MATRIX_ROW(cur_a, _i12); + const int id = row_mapping.i1; // selected expert index + + const int64_t i11 = id % ne11; + const int64_t i12 = row_mapping.i2; // row index in src1 + + const int64_t i1 = id; // selected expert index + const int64_t i2 = i12; // row + + // desc: when src1 is not a contiguous memory block we have to calculate the offset using the strides + // if it is, then we have either copied the data to params->wdata and made it contiguous or we are using + // the original src1 data pointer, so we should index using the indices directly + // TODO: this is a bit of a hack, we should probably have a better way to handle this + const char * src1_col = (const char *) wdata + + (src1_cont || src1->type != vec_dot_type + ? (i11 + i12*ne11)*row_size + : (i11*nb11 + i12*nb12)); + + float * dst_col = (float *) ((char *) dst->data + (i1*nb1 + i2*nb2)); + + for (int64_t ir0 = iir0; ir0 < iir0 + blck_0 && ir0 < ir0_end; ++ir0) { + vec_dot(ne00, &tmp[ir0 - iir0], 0, src0_cur + ir0*nb01, 0, src1_col, 0, 1); + } + + memcpy(&dst_col[iir0], tmp, (MIN(iir0 + blck_0, ir0_end) - iir0)*sizeof(float)); + } + } + } +} + +static void * incr_ptr_aligned(void ** p, size_t size, size_t align) { + + void * ptr = *p; + ptr = (void *) GGML_PAD((uintptr_t) ptr, align); + *p = (void *) ((char *) ptr + size); + return ptr; +} + static void ggml_compute_forward_mul_mat_id( const struct ggml_compute_params * params, struct ggml_tensor * dst) { @@ -7623,7 +1554,6 @@ static void ggml_compute_forward_mul_mat_id( const bool src1_cont = ggml_is_contiguous(src1); - ggml_vec_dot_t const vec_dot = type_traits_cpu[type].vec_dot; enum ggml_type const vec_dot_type = type_traits_cpu[type].vec_dot_type; ggml_from_float_t const from_float = type_traits_cpu[vec_dot_type].from_float; @@ -7641,21 +1571,27 @@ static void ggml_compute_forward_mul_mat_id( const int n_ids = ids->ne[0]; // n_expert_used const int n_as = ne02; // n_expert - char * wdata_src1_end = (src1->type == vec_dot_type) ? - (char *) params->wdata : - (char *) params->wdata + GGML_PAD(ggml_row_size(vec_dot_type, ggml_nelements(src1)), sizeof(int64_t)); + void * wdata_cur = params->wdata; - struct mmid_row_mapping { - int32_t i1; - int32_t i2; - }; + if (src1->type != vec_dot_type) { + incr_ptr_aligned(&wdata_cur, ggml_row_size(vec_dot_type, ggml_nelements(src1)), sizeof(int64_t)); + } - int64_t * matrix_row_counts = (int64_t *) (wdata_src1_end); // [n_as] - struct mmid_row_mapping * matrix_rows = (struct mmid_row_mapping *)(matrix_row_counts + n_as); // [n_as][ne11] + int64_t * matrix_row_counts = // [n_as] + incr_ptr_aligned(&wdata_cur, n_as*sizeof(int64_t), sizeof(int64_t)); + + struct mmid_row_mapping * matrix_rows = // [n_as][ids->ne[0]*ids->ne[1]] + incr_ptr_aligned(&wdata_cur, n_as*ids->ne[0]*ids->ne[1]*sizeof(struct mmid_row_mapping), sizeof(int64_t)); + + char (*atomic_current_chunk)[CACHE_LINE_SIZE] = // [n_as] + incr_ptr_aligned(&wdata_cur, CACHE_LINE_SIZE * n_as, CACHE_LINE_SIZE); + + GGML_ASSERT(params->wsize >= (size_t)((char *) wdata_cur - (char *) params->wdata)); if (src1->type != vec_dot_type) { char * wdata = params->wdata; + const size_t nbw0 = ggml_type_size(vec_dot_type); const size_t nbw1 = ggml_row_size(vec_dot_type, ne10); const size_t nbw2 = nbw1*ne11; const size_t nbw3 = nbw2*ne12; @@ -7663,19 +1599,32 @@ static void ggml_compute_forward_mul_mat_id( assert(params->wsize >= ne13*nbw3); GGML_ASSERT(src1->type == GGML_TYPE_F32); +#if 0 for (int64_t i13 = 0; i13 < ne13; ++i13) { - for (int64_t i12 = 0; i12 < ne12; ++i12) { - for (int64_t i11 = ith; i11 < ne11; i11 += nth) { + for (int64_t i12 = ith; i12 < ne12; i12 += nth) { + for (int64_t i11 = 0; i11 < ne11; ++i11) { from_float((float *)((char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11), (void *) (wdata + i13*nbw3 + i12*nbw2 + i11*nbw1), ne10); } } } +#else + for (int64_t i13 = 0; i13 < ne13; ++i13) { + for (int64_t i12 = 0; i12 < ne12; ++i12) { + for (int64_t i11 = 0; i11 < ne11; ++i11) { + size_t bs = ggml_blck_size(vec_dot_type); + int64_t ne10_block_start = (ith * ne10/bs) / nth; + int64_t ne10_block_end = ((ith + 1) * ne10/bs) / nth; + from_float((float *)((char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11 + ne10_block_start*bs*nb10), + (void *) (wdata + i13*nbw3 + i12*nbw2 + i11*nbw1 + ne10_block_start*nbw0), + (ne10_block_end - ne10_block_start) * bs); + } + } + } +#endif } -#define MMID_MATRIX_ROW(row_id, i1) matrix_rows[(row_id)*ne12 + (i1)] - if (ith == 0) { // initialize matrix_row_counts memset(matrix_row_counts, 0, n_as*sizeof(int64_t)); @@ -7693,9 +1642,14 @@ static void ggml_compute_forward_mul_mat_id( } } + // reset current_chunk + for (int cur_a = ith; cur_a < n_as; cur_a += nth) { + atomic_int * current_chunk_ctr = (atomic_int *)(atomic_current_chunk + cur_a); + *current_chunk_ctr = nth; + } + ggml_barrier(params->threadpool); - // compute each matrix multiplication in sequence for (int cur_a = 0; cur_a < n_as; ++cur_a) { const int64_t cne1 = matrix_row_counts[cur_a]; @@ -7703,5013 +1657,66 @@ static void ggml_compute_forward_mul_mat_id( continue; } - const char * src0_cur = (const char *) src0->data + cur_a*nb02; - - const void * wdata = (src1->type == vec_dot_type) ? src1->data : params->wdata; + const char * src0_cur = (const char *) src0->data + cur_a * nb02; + const void * wdata = (src1->type == vec_dot_type) ? src1->data : params->wdata; const size_t row_size = ggml_row_size(vec_dot_type, ne10); - const int64_t nr0 = ne01; // src0 rows - const int64_t nr1 = cne1; // src1 rows + const int64_t nr0 = ne01; + const int64_t nr1 = cne1; - // distribute the thread work across the inner or outer loop based on which one is larger - - const int64_t nth0 = nr0 > nr1 ? nth : 1; // parallelize by src0 rows - const int64_t nth1 = nr0 > nr1 ? 1 : nth; // parallelize by src1 rows - - const int64_t ith0 = ith % nth0; - const int64_t ith1 = ith / nth0; - - const int64_t dr0 = (nr0 + nth0 - 1)/nth0; - const int64_t dr1 = (nr1 + nth1 - 1)/nth1; - - const int64_t ir010 = dr0*ith0; - const int64_t ir011 = MIN(ir010 + dr0, nr0); - - const int64_t ir110 = dr1*ith1; - const int64_t ir111 = MIN(ir110 + dr1, nr1); - - // threads with no work simply yield (not sure if it helps) - //if (ir010 >= ir011 || ir110 >= ir111) { - // sched_yield(); - // continue; - //} - - // block-tiling attempt - const int64_t blck_0 = 16; - const int64_t blck_1 = 16; - - // attempt to reduce false-sharing (does not seem to make a difference) - float tmp[16]; - - for (int64_t iir1 = ir110; iir1 < ir111; iir1 += blck_1) { - for (int64_t iir0 = ir010; iir0 < ir011; iir0 += blck_0) { - for (int64_t ir1 = iir1; ir1 < iir1 + blck_1 && ir1 < ir111; ++ir1) { - const int64_t _i12 = ir1; // logical row index for this expert - - struct mmid_row_mapping row_mapping = MMID_MATRIX_ROW(cur_a, _i12); - const int id = row_mapping.i1; // selected expert index - - const int64_t i11 = id % ne11; - const int64_t i12 = row_mapping.i2; // row index in src1 - - const int64_t i1 = id; // selected expert index - const int64_t i2 = i12; // row - - // desc: when src1 is not a contiguous memory block we have to calculate the offset using the strides - // if it is, then we have either copied the data to params->wdata and made it contiguous or we are using - // the original src1 data pointer, so we should index using the indices directly - // TODO: this is a bit of a hack, we should probably have a better way to handle this - const char * src1_col = (const char *) wdata + - (src1_cont || src1->type != vec_dot_type - ? (i11 + i12*ne11)*row_size - : (i11*nb11 + i12*nb12)); - - float * dst_col = (float *) ((char *) dst->data + (i1*nb1 + i2*nb2)); - - //for (int64_t ir0 = iir0; ir0 < iir0 + blck_0 && ir0 < ir011; ++ir0) { - // vec_dot(ne00, &dst_col[ir0], src0_row + ir0*nb01, src1_col); - //} - - for (int64_t ir0 = iir0; ir0 < iir0 + blck_0 && ir0 < ir011; ++ir0) { - vec_dot(ne00, &tmp[ir0 - iir0], 0, src0_cur + ir0*nb01, 0, src1_col, 0, 1); - } - - memcpy(&dst_col[iir0], tmp, (MIN(iir0 + blck_0, ir011) - iir0)*sizeof(float)); - } - } + int chunk_size = 16; + if (nr0 == 1 || nr1 == 1) { + chunk_size = 64; } - } -#undef MMID_MATRIX_ROW -} - -// ggml_compute_forward_out_prod - -static void ggml_compute_forward_out_prod_f32( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - const struct ggml_tensor * src1 = dst->src[1]; - - GGML_TENSOR_BINARY_OP_LOCALS - - GGML_ASSERT(dst->type == GGML_TYPE_F32); - GGML_ASSERT(src0->type == GGML_TYPE_F32); - GGML_ASSERT(src1->type == GGML_TYPE_F32); - - const int ith = params->ith; - const int nth = params->nth; - - GGML_ASSERT(ne0 == ne00); - GGML_ASSERT(ne1 == ne10); - GGML_ASSERT(ne2 == ne12); - GGML_ASSERT(ne3 == ne13); - - GGML_ASSERT(ne2 % ne02 == 0); - GGML_ASSERT(ne3 % ne03 == 0); - - // we don't support permuted src0 or src1 - GGML_ASSERT(nb00 == sizeof(float)); - - // dst cannot be transposed or permuted - GGML_ASSERT(nb0 == sizeof(float)); - // GGML_ASSERT(nb0 <= nb1); - // GGML_ASSERT(nb1 <= nb2); - // GGML_ASSERT(nb2 <= nb3); - - // nb01 >= nb00 - src0 is not transposed - // compute by src0 rows - - if (ith == 0) { - ggml_vec_set_f32(ne0*ne1*ne2*ne3, dst->data, 0); - } - ggml_barrier(params->threadpool); - - // dst[:,:,:,:] = 0 - // for i2,i3: - // for i1: - // for i01: - // for i0: - // dst[i0,i1,i2,i3] += src0[i0,i01,i2,i3] * src1[i1,i01,i2,i3] - - // parallelize by last three dimensions - - // total rows in dst - const int64_t nr = ne1*ne2*ne3; - - // rows per thread - const int64_t dr = (nr + nth - 1)/nth; - - // row range for this thread - const int64_t ir0 = dr*ith; - const int64_t ir1 = MIN(ir0 + dr, nr); - - // block-tiling attempt - const int64_t blck_0 = MAX(GGML_VEC_MAD_UNROLL, 32); - const int64_t blck_1 = 16; - - // dps == dst per src0, used for group query attention - const int64_t dps2 = ne2 / ne02; - const int64_t dps3 = ne3 / ne03; - - for (int64_t bir = ir0; bir < ir1; bir += blck_1) { - const int64_t bir1 = MIN(bir + blck_1, ir1); - for (int64_t bi01 = 0; bi01 < ne01; bi01 += blck_0) { - const int64_t bne01 = MIN(bi01 + blck_0, ne01); - for (int64_t ir = bir; ir < bir1; ++ir) { - // dst indices - const int64_t i3 = ir/(ne2*ne1); - const int64_t i2 = (ir - i3*ne2*ne1)/ne1; - const int64_t i1 = (ir - i3*ne2*ne1 - i2*ne1); - - const int64_t i02 = i2 / dps2; - const int64_t i03 = i3 / dps3; - - //const int64_t i10 = i1; - const int64_t i12 = i2; - const int64_t i13 = i3; - -#if GGML_VEC_MAD_UNROLL > 2 - const int64_t bne01_unroll = bne01 - (bne01 % GGML_VEC_MAD_UNROLL); - for (int64_t i01 = bi01; i01 < bne01_unroll; i01 += GGML_VEC_MAD_UNROLL) { - const int64_t i11 = i01; - - float * s0 = (float *) ((char *) src0->data + ( i01*nb01 + i02*nb02 + i03*nb03)); - float * s1 = (float *) ((char *) src1->data + (i1*nb10 + i11*nb11 + i12*nb12 + i13*nb13)); - float * d = (float *) ((char *) dst->data + ( i1*nb1 + i2*nb2 + i3*nb3)); - - ggml_vec_mad_f32_unroll(ne0, nb01, nb11, d, s0, s1); - } - for (int64_t i01 = bne01_unroll; i01 < bne01; ++i01) { - const int64_t i11 = i01; - - float * s0 = (float *) ((char *) src0->data + ( i01*nb01 + i02*nb02 + i03*nb03)); - float * s1 = (float *) ((char *) src1->data + (i1*nb10 + i11*nb11 + i12*nb12 + i13*nb13)); - float * d = (float *) ((char *) dst->data + ( i1*nb1 + i2*nb2 + i3*nb3)); - - ggml_vec_mad_f32(ne0, d, s0, *s1); - } +#if defined(__aarch64__) + // disable for ARM + const bool disable_chunking = true; #else - for (int64_t i01 = bi01; i01 < bne01; ++i01) { - const int64_t i11 = i01; + // disable for NUMA + const bool disable_chunking = ggml_is_numa(); +#endif // defined(__aarch64__) - float * s0 = (float *) ((char *) src0->data + ( i01*nb01 + i02*nb02 + i03*nb03)); - float * s1 = (float *) ((char *) src1->data + (i1*nb10 + i11*nb11 + i12*nb12 + i13*nb13)); - float * d = (float *) ((char *) dst->data + ( i1*nb1 + i2*nb2 + i3*nb3)); + int64_t nchunk0 = (nr0 + chunk_size - 1) / chunk_size; + int64_t nchunk1 = (nr1 + chunk_size - 1) / chunk_size; - ggml_vec_mad_f32(ne0, d, s0, *s1); - } -#endif + if (nchunk0 * nchunk1 < nth * 4 || disable_chunking) { + nchunk0 = nr0 > nr1 ? nth : 1; + nchunk1 = nr0 > nr1 ? 1 : nth; + } + + const int64_t dr0 = (nr0 + nchunk0 - 1) / nchunk0; + const int64_t dr1 = (nr1 + nchunk1 - 1) / nchunk1; + + int current_chunk = ith; + + atomic_int * current_chunk_ctr = (atomic_int *)(atomic_current_chunk + cur_a); + + while (current_chunk < nchunk0 * nchunk1) { + const int64_t ith0 = current_chunk % nchunk0; + const int64_t ith1 = current_chunk / nchunk0; + + const int64_t ir0_start = dr0 * ith0; + const int64_t ir0_end = MIN(ir0_start + dr0, nr0); + + const int64_t ir1_start = dr1 * ith1; + const int64_t ir1_end = MIN(ir1_start + dr1, nr1); + + ggml_compute_forward_mul_mat_id_one_chunk( + dst, src0, src1, ids, cur_a, + ir0_start, ir0_end, ir1_start, ir1_end, + src0_cur, matrix_rows, row_size, src1_cont, wdata + ); + + if (nth >= nchunk0 * nchunk1) { + break; } + + current_chunk = atomic_fetch_add_explicit(current_chunk_ctr, 1, memory_order_relaxed); } } } -static void ggml_compute_forward_out_prod_q_f32( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - const struct ggml_tensor * src1 = dst->src[1]; - - GGML_TENSOR_BINARY_OP_LOCALS; - - const int ith = params->ith; - const int nth = params->nth; - - const enum ggml_type type = src0->type; - ggml_to_float_t const dequantize_row_q = ggml_get_type_traits(type)->to_float; - - GGML_ASSERT(ne02 == ne12); - GGML_ASSERT(ne03 == ne13); - GGML_ASSERT(ne2 == ne12); - GGML_ASSERT(ne3 == ne13); - - // we don't support permuted src0 dim0 - GGML_ASSERT(nb00 == ggml_type_size(type)); - - // dst dim0 cannot be transposed or permuted - GGML_ASSERT(nb0 == sizeof(float)); - // GGML_ASSERT(nb0 <= nb1); - // GGML_ASSERT(nb1 <= nb2); - // GGML_ASSERT(nb2 <= nb3); - - GGML_ASSERT(ne0 == ne00); - GGML_ASSERT(ne1 == ne10); - GGML_ASSERT(ne2 == ne02); - GGML_ASSERT(ne3 == ne03); - - // nb01 >= nb00 - src0 is not transposed - // compute by src0 rows - - if (ith == 0) { - ggml_vec_set_f32(ne0*ne1*ne2*ne3, dst->data, 0); - } - ggml_barrier(params->threadpool); - - // parallelize by last three dimensions - - // total rows in dst - const int64_t nr = ne1*ne2*ne3; - - // rows per thread - const int64_t dr = (nr + nth - 1)/nth; - - // row range for this thread - const int64_t ir0 = dr*ith; - const int64_t ir1 = MIN(ir0 + dr, nr); - - // dst[:,:,:,:] = 0 - // for i2,i3: - // for i1: - // for i01: - // for i0: - // dst[i0,i1,i2,i3] += src0[i0,i01,i2,i3] * src1[i1,i01,i2,i3] - - float * wdata = (float *) params->wdata + (ne0 + CACHE_LINE_SIZE_F32) * ith; - - for (int64_t ir = ir0; ir < ir1; ++ir) { - // dst indices - const int64_t i3 = ir/(ne2*ne1); - const int64_t i2 = (ir - i3*ne2*ne1)/ne1; - const int64_t i1 = (ir - i3*ne2*ne1 - i2*ne1); - - const int64_t i02 = i2; - const int64_t i03 = i3; - - //const int64_t i10 = i1; - const int64_t i12 = i2; - const int64_t i13 = i3; - - for (int64_t i01 = 0; i01 < ne01; ++i01) { - const int64_t i11 = i01; - - float * s0 = (float *) ((char *) src0->data + ( i01*nb01 + i02*nb02 + i03*nb03)); - float * s1 = (float *) ((char *) src1->data + (i1*nb10 + i11*nb11 + i12*nb12 + i13*nb13)); - float * d = (float *) ((char *) dst->data + ( i1*nb1 + i2*nb2 + i3*nb3)); - - dequantize_row_q(s0, wdata, ne0); - ggml_vec_mad_f32(ne0, d, wdata, *s1); - } - } -} - -static void ggml_compute_forward_out_prod( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - switch (src0->type) { - case GGML_TYPE_Q4_0: - case GGML_TYPE_Q4_1: - case GGML_TYPE_Q5_0: - case GGML_TYPE_Q5_1: - case GGML_TYPE_Q8_0: - case GGML_TYPE_Q2_K: - case GGML_TYPE_Q3_K: - case GGML_TYPE_Q4_K: - case GGML_TYPE_Q5_K: - case GGML_TYPE_Q6_K: - case GGML_TYPE_TQ1_0: - case GGML_TYPE_TQ2_0: - case GGML_TYPE_IQ2_XXS: - case GGML_TYPE_IQ2_XS: - case GGML_TYPE_IQ3_XXS: - case GGML_TYPE_IQ1_S: - case GGML_TYPE_IQ1_M: - case GGML_TYPE_IQ4_NL: - case GGML_TYPE_IQ4_XS: - case GGML_TYPE_IQ3_S: - case GGML_TYPE_IQ2_S: - { - ggml_compute_forward_out_prod_q_f32(params, dst); - } break; - case GGML_TYPE_F16: - { - GGML_ABORT("fatal error"); // todo - // ggml_compute_forward_out_prod_f16_f32(params, dst); - } - case GGML_TYPE_F32: - { - ggml_compute_forward_out_prod_f32(params, dst); - } break; - default: - { - GGML_ABORT("fatal error"); - } - } -} - -// ggml_compute_forward_scale - -static void ggml_compute_forward_scale_f32( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - GGML_ASSERT(ggml_is_contiguous(src0)); - GGML_ASSERT(ggml_is_contiguous(dst)); - GGML_ASSERT(ggml_are_same_shape(src0, dst)); - - // scale factor - float v; - memcpy(&v, dst->op_params, sizeof(float)); - - const int ith = params->ith; - const int nth = params->nth; - - const int nc = src0->ne[0]; - const int nr = ggml_nrows(src0); - - // rows per thread - const int dr = (nr + nth - 1)/nth; - - // row range for this thread - const int ir0 = dr*ith; - const int ir1 = MIN(ir0 + dr, nr); - - const size_t nb01 = src0->nb[1]; - - const size_t nb1 = dst->nb[1]; - - for (int i1 = ir0; i1 < ir1; i1++) { - if (dst->data != src0->data) { - // src0 is same shape as dst => same indices - memcpy((char *)dst->data + i1*nb1, (char *)src0->data + i1*nb01, nc * sizeof(float)); - } - ggml_vec_scale_f32(nc, (float *) ((char *) dst->data + i1*nb1), v); - } -} - -static void ggml_compute_forward_scale( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - switch (src0->type) { - case GGML_TYPE_F32: - { - ggml_compute_forward_scale_f32(params, dst); - } break; - default: - { - GGML_ABORT("fatal error"); - } - } -} - -// ggml_compute_forward_set - -static void ggml_compute_forward_set_f32( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - const struct ggml_tensor * src1 = dst->src[1]; - - GGML_ASSERT(ggml_are_same_shape(src0, dst)); - GGML_ASSERT(ggml_is_contiguous(dst) && ggml_is_contiguous(src0)); - - // view src0 and dst with these strides and data offset inbytes during set - // nb0 is implicitly element_size because src0 and dst are contiguous - size_t nb1 = ((int32_t *) dst->op_params)[0]; - size_t nb2 = ((int32_t *) dst->op_params)[1]; - size_t nb3 = ((int32_t *) dst->op_params)[2]; - size_t offset = ((int32_t *) dst->op_params)[3]; - bool inplace = (bool) ((int32_t *) dst->op_params)[4]; - - if (!inplace) { - if (params->ith == 0) { - // memcpy needs to be synchronized across threads to avoid race conditions. - // => do it in INIT phase - memcpy( - ((char *) dst->data), - ((char *) src0->data), - ggml_nbytes(dst)); - } - ggml_barrier(params->threadpool); - } - - const int ith = params->ith; - const int nth = params->nth; - - const int nr = ggml_nrows(src1); - const int nc = src1->ne[0]; - - GGML_TENSOR_LOCALS(int64_t, ne1, src1, ne) - GGML_TENSOR_LOCALS(size_t, nb1, src1, nb) - - // src0 and dst as viewed during set - const size_t nb0 = ggml_element_size(src0); - - const int im0 = (ne10 == 0 ? 0 : ne10-1); - const int im1 = (ne11 == 0 ? 0 : ne11-1); - const int im2 = (ne12 == 0 ? 0 : ne12-1); - const int im3 = (ne13 == 0 ? 0 : ne13-1); - - GGML_ASSERT(offset + im0*nb0 + im1*nb1 + im2*nb2 + im3*nb3 <= ggml_nbytes(dst)); - - GGML_ASSERT(nb10 == sizeof(float)); - - // rows per thread - const int dr = (nr + nth - 1)/nth; - - // row range for this thread - const int ir0 = dr*ith; - const int ir1 = MIN(ir0 + dr, nr); - - for (int ir = ir0; ir < ir1; ++ir) { - // src0 and dst are viewed with shape of src1 and offset - // => same indices - const int i3 = ir/(ne12*ne11); - const int i2 = (ir - i3*ne12*ne11)/ne11; - const int i1 = (ir - i3*ne12*ne11 - i2*ne11); - - ggml_vec_cpy_f32(nc, - (float *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + offset), - (float *) ((char *) src1->data + i3*nb13 + i2*nb12 + i1*nb11)); - } -} - -static void ggml_compute_forward_set_i32( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - const struct ggml_tensor * src1 = dst->src[1]; - - GGML_ASSERT(ggml_are_same_shape(src0, dst)); - GGML_ASSERT(ggml_is_contiguous(dst) && ggml_is_contiguous(src0)); - - // view src0 and dst with these strides and data offset inbytes during set - // nb0 is implicitly element_size because src0 and dst are contiguous - size_t nb1 = ((int32_t *) dst->op_params)[0]; - size_t nb2 = ((int32_t *) dst->op_params)[1]; - size_t nb3 = ((int32_t *) dst->op_params)[2]; - size_t offset = ((int32_t *) dst->op_params)[3]; - bool inplace = (bool) ((int32_t *) dst->op_params)[4]; - - if (!inplace) { - if (params->ith == 0) { - // memcpy needs to be synchronized across threads to avoid race conditions. - // => do it in INIT phase - memcpy( - ((char *) dst->data), - ((char *) src0->data), - ggml_nbytes(dst)); - } - ggml_barrier(params->threadpool); - } - - const int ith = params->ith; - const int nth = params->nth; - - const int nr = ggml_nrows(src1); - const int nc = src1->ne[0]; - - GGML_TENSOR_LOCALS(int64_t, ne1, src1, ne) - GGML_TENSOR_LOCALS(size_t, nb1, src1, nb) - - // src0 and dst as viewed during set - const size_t nb0 = ggml_element_size(src0); - - const int im0 = (ne10 == 0 ? 0 : ne10-1); - const int im1 = (ne11 == 0 ? 0 : ne11-1); - const int im2 = (ne12 == 0 ? 0 : ne12-1); - const int im3 = (ne13 == 0 ? 0 : ne13-1); - - GGML_ASSERT(offset + im0*nb0 + im1*nb1 + im2*nb2 + im3*nb3 <= ggml_nbytes(dst)); - - GGML_ASSERT(nb10 == sizeof(int32_t)); - - // rows per thread - const int dr = (nr + nth - 1)/nth; - - // row range for this thread - const int ir0 = dr*ith; - const int ir1 = MIN(ir0 + dr, nr); - - for (int ir = ir0; ir < ir1; ++ir) { - // src0 and dst are viewed with shape of src1 and offset - // => same indices - const int i3 = ir/(ne12*ne11); - const int i2 = (ir - i3*ne12*ne11)/ne11; - const int i1 = (ir - i3*ne12*ne11 - i2*ne11); - - ggml_vec_cpy_i32(nc, - (int32_t *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + offset), - (int32_t *) ((char *) src1->data + i3*nb13 + i2*nb12 + i1*nb11)); - } -} - -static void ggml_compute_forward_set( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - switch (src0->type) { - case GGML_TYPE_F32: - { - ggml_compute_forward_set_f32(params, dst); - } break; - case GGML_TYPE_I32: - { - ggml_compute_forward_set_i32(params, dst); - } break; - case GGML_TYPE_F16: - case GGML_TYPE_BF16: - case GGML_TYPE_Q4_0: - case GGML_TYPE_Q4_1: - case GGML_TYPE_Q5_0: - case GGML_TYPE_Q5_1: - case GGML_TYPE_Q8_0: - case GGML_TYPE_Q8_1: - case GGML_TYPE_Q2_K: - case GGML_TYPE_Q3_K: - case GGML_TYPE_Q4_K: - case GGML_TYPE_Q5_K: - case GGML_TYPE_Q6_K: - case GGML_TYPE_TQ1_0: - case GGML_TYPE_TQ2_0: - case GGML_TYPE_IQ2_XXS: - case GGML_TYPE_IQ2_XS: - case GGML_TYPE_IQ3_XXS: - case GGML_TYPE_IQ1_S: - case GGML_TYPE_IQ1_M: - case GGML_TYPE_IQ4_NL: - case GGML_TYPE_IQ4_XS: - case GGML_TYPE_IQ3_S: - case GGML_TYPE_IQ2_S: - default: - { - GGML_ABORT("fatal error"); - } - } -} - -// ggml_compute_forward_cpy - -static void ggml_compute_forward_cpy( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - ggml_compute_forward_dup(params, dst); -} - -// ggml_compute_forward_cont - -static void ggml_compute_forward_cont( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - ggml_compute_forward_dup(params, dst); -} - -// ggml_compute_forward_reshape - -static void ggml_compute_forward_reshape( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - // NOP - UNUSED(params); - UNUSED(dst); -} - -// ggml_compute_forward_view - -static void ggml_compute_forward_view( - const struct ggml_compute_params * params, - const struct ggml_tensor * dst) { - // NOP - UNUSED(params); - UNUSED(dst); -} - -// ggml_compute_forward_permute - -static void ggml_compute_forward_permute( - const struct ggml_compute_params * params, - const struct ggml_tensor * dst) { - // NOP - UNUSED(params); - UNUSED(dst); -} - -// ggml_compute_forward_transpose - -static void ggml_compute_forward_transpose( - const struct ggml_compute_params * params, - const struct ggml_tensor * dst) { - // NOP - UNUSED(params); - UNUSED(dst); -} - -// ggml_compute_forward_get_rows - -static void ggml_compute_forward_get_rows_q( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - const struct ggml_tensor * src1 = dst->src[1]; - - GGML_TENSOR_BINARY_OP_LOCALS - - const int64_t nc = ne00; - const int64_t nr = ggml_nelements(src1); - - const enum ggml_type type = src0->type; - ggml_to_float_t const dequantize_row_q = ggml_get_type_traits(type)->to_float; - - assert(ne0 == nc); - assert(ne02 == ne11); - assert(nb00 == ggml_type_size(type)); - assert(ggml_nrows(dst) == nr); - - const int ith = params->ith; - const int nth = params->nth; - - // rows per thread - const int dr = (nr + nth - 1)/nth; - - // row range for this thread - const int ir0 = dr*ith; - const int ir1 = MIN(ir0 + dr, nr); - - for (int64_t i = ir0; i < ir1; ++i) { - const int64_t i12 = i/(ne11*ne10); - const int64_t i11 = (i - i12*ne11*ne10)/ne10; - const int64_t i10 = (i - i12*ne11*ne10 - i11*ne10); - const int64_t i01 = *(int32_t *) ((char *) src1->data + i10*nb10 + i11*nb11 + i12*nb12); - - GGML_ASSERT(i01 >= 0 && i01 < ne01); - - dequantize_row_q( - (const void *) ((char *) src0->data + i01*nb01 + i11*nb02 + i12*nb03), - (float *) ((char *) dst->data + i10*nb1 + i11*nb2 + i12*nb3), nc); - } -} - -static void ggml_compute_forward_get_rows_f16( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - const struct ggml_tensor * src1 = dst->src[1]; - - GGML_TENSOR_BINARY_OP_LOCALS - - const int64_t nc = ne00; - const int64_t nr = ggml_nelements(src1); - - assert(ne0 == nc); - assert(ne02 == ne11); - assert(nb00 == sizeof(ggml_fp16_t)); - assert(ggml_nrows(dst) == nr); - - const int ith = params->ith; - const int nth = params->nth; - - // rows per thread - const int dr = (nr + nth - 1)/nth; - - // row range for this thread - const int ir0 = dr*ith; - const int ir1 = MIN(ir0 + dr, nr); - - for (int64_t i = ir0; i < ir1; ++i) { - const int64_t i12 = i/(ne11*ne10); - const int64_t i11 = (i - i12*ne11*ne10)/ne10; - const int64_t i10 = (i - i12*ne11*ne10 - i11*ne10); - const int64_t i01 = *(int32_t *) ((char *) src1->data + i10*nb10 + i11*nb11 + i12*nb12); - - GGML_ASSERT(i01 >= 0 && i01 < ne01); - - ggml_fp16_to_fp32_row( - (const void *) ((char *) src0->data + i01*nb01 + i11*nb02 + i12*nb03), - (float *) ((char *) dst->data + i10*nb1 + i11*nb2 + i12*nb3), nc); - } -} - -static void ggml_compute_forward_get_rows_bf16( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - const struct ggml_tensor * src1 = dst->src[1]; - - GGML_TENSOR_BINARY_OP_LOCALS - - const int64_t nc = ne00; - const int64_t nr = ggml_nelements(src1); - - assert(ne0 == nc); - assert(ne02 == ne11); - assert(nb00 == sizeof(ggml_bf16_t)); - assert(ggml_nrows(dst) == nr); - - const int ith = params->ith; - const int nth = params->nth; - - // rows per thread - const int dr = (nr + nth - 1)/nth; - - // row range for this thread - const int ir0 = dr*ith; - const int ir1 = MIN(ir0 + dr, nr); - - for (int64_t i = ir0; i < ir1; ++i) { - const int64_t i12 = i/(ne11*ne10); - const int64_t i11 = (i - i12*ne11*ne10)/ne10; - const int64_t i10 = (i - i12*ne11*ne10 - i11*ne10); - const int64_t i01 = *(int32_t *) ((char *) src1->data + i10*nb10 + i11*nb11 + i12*nb12); - - GGML_ASSERT(i01 >= 0 && i01 < ne01); - - ggml_bf16_to_fp32_row( - (const void *) ((char *) src0->data + i01*nb01 + i11*nb02 + i12*nb03), - (float *) ((char *) dst->data + i10*nb1 + i11*nb2 + i12*nb3), nc); - } -} - -static void ggml_compute_forward_get_rows_f32( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - const struct ggml_tensor * src1 = dst->src[1]; - - GGML_TENSOR_BINARY_OP_LOCALS - - const int64_t nc = ne00; - const int64_t nr = ggml_nelements(src1); - - assert(ne0 == nc); - assert(ne02 == ne11); - assert(nb00 == sizeof(float)); - assert(ggml_nrows(dst) == nr); - - const int ith = params->ith; - const int nth = params->nth; - - // rows per thread - const int dr = (nr + nth - 1)/nth; - - // row range for this thread - const int ir0 = dr*ith; - const int ir1 = MIN(ir0 + dr, nr); - - for (int64_t i = ir0; i < ir1; ++i) { - const int64_t i12 = i/(ne11*ne10); - const int64_t i11 = (i - i12*ne11*ne10)/ne10; - const int64_t i10 = (i - i12*ne11*ne10 - i11*ne10); - const int64_t i01 = *(int32_t *) ((char *) src1->data + i10*nb10 + i11*nb11 + i12*nb12); - - GGML_ASSERT(i01 >= 0 && i01 < ne01); - - ggml_vec_cpy_f32(nc, - (float *) ((char *) dst->data + i10*nb1 + i11*nb2 + i12*nb3), - (float *) ((char *) src0->data + i01*nb01 + i11*nb02 + i12*nb03)); - } -} - -static void ggml_compute_forward_get_rows( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - switch (src0->type) { - case GGML_TYPE_Q4_0: - case GGML_TYPE_Q4_1: - case GGML_TYPE_Q5_0: - case GGML_TYPE_Q5_1: - case GGML_TYPE_Q8_0: - case GGML_TYPE_Q8_1: - case GGML_TYPE_Q2_K: - case GGML_TYPE_Q3_K: - case GGML_TYPE_Q4_K: - case GGML_TYPE_Q5_K: - case GGML_TYPE_Q6_K: - case GGML_TYPE_TQ1_0: - case GGML_TYPE_TQ2_0: - case GGML_TYPE_IQ2_XXS: - case GGML_TYPE_IQ2_XS: - case GGML_TYPE_IQ3_XXS: - case GGML_TYPE_IQ1_S: - case GGML_TYPE_IQ1_M: - case GGML_TYPE_IQ4_NL: - case GGML_TYPE_IQ4_XS: - case GGML_TYPE_IQ3_S: - case GGML_TYPE_IQ2_S: - { - ggml_compute_forward_get_rows_q(params, dst); - } break; - case GGML_TYPE_F16: - { - ggml_compute_forward_get_rows_f16(params, dst); - } break; - case GGML_TYPE_BF16: - { - ggml_compute_forward_get_rows_bf16(params, dst); - } break; - case GGML_TYPE_F32: - case GGML_TYPE_I32: - { - ggml_compute_forward_get_rows_f32(params, dst); - } break; - default: - { - GGML_ABORT("fatal error"); - } - } - - //static bool first = true; - //printf("ne0 = %d, ne1 = %d, ne2 = %d\n", dst->ne[0], dst->ne[1], dst->ne[2]); - //if (first) { - // first = false; - //} else { - // for (int k = 0; k < dst->ne[1]; ++k) { - // for (int j = 0; j < dst->ne[0]/16; ++j) { - // for (int i = 0; i < 16; ++i) { - // printf("%8.4f ", ((float *) dst->data)[k*dst->ne[0] + j*16 + i]); - // } - // printf("\n"); - // } - // printf("\n"); - // } - // printf("\n"); - // exit(0); - //} -} - -// ggml_compute_forward_get_rows_back - -static void ggml_compute_forward_get_rows_back_f32_f16( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - const struct ggml_tensor * src1 = dst->src[1]; - - if (params->ith != 0) { - return; - } - - GGML_ASSERT(ggml_is_contiguous(dst)); - - // ggml_compute_forward_dup_same_cont(params, opt0, dst); - - memset(dst->data, 0, ggml_nbytes(dst)); - - const int nc = src0->ne[0]; - const int nr = ggml_nelements(src1); - - GGML_ASSERT( dst->ne[0] == nc); - GGML_ASSERT(src0->nb[0] == sizeof(ggml_fp16_t)); - - for (int i = 0; i < nr; ++i) { - const int r = ((int32_t *) src1->data)[i]; - - for (int j = 0; j < nc; ++j) { - ggml_fp16_t v = ((ggml_fp16_t *) ((char *) src0->data + i*src0->nb[1]))[j]; - ((float *) ((char *) dst->data + r*dst->nb[1]))[j] += GGML_FP16_TO_FP32(v); - } - } -} - -static void ggml_compute_forward_get_rows_back_f32( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - const struct ggml_tensor * src1 = dst->src[1]; - - if (params->ith != 0) { - return; - } - - GGML_ASSERT(ggml_is_contiguous(dst)); - - // ggml_compute_forward_dup_same_cont(params, opt0, dst); - - memset(dst->data, 0, ggml_nbytes(dst)); - - const int nc = src0->ne[0]; - const int nr = ggml_nelements(src1); - - GGML_ASSERT( dst->ne[0] == nc); - GGML_ASSERT(src0->nb[0] == sizeof(float)); - - for (int i = 0; i < nr; ++i) { - const int r = ((int32_t *) src1->data)[i]; - - ggml_vec_add_f32(nc, - (float *) ((char *) dst->data + r*dst->nb[1]), - (float *) ((char *) dst->data + r*dst->nb[1]), - (float *) ((char *) src0->data + i*src0->nb[1])); - } -} - -static void ggml_compute_forward_get_rows_back( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - switch (src0->type) { - case GGML_TYPE_F16: - { - ggml_compute_forward_get_rows_back_f32_f16(params, dst); - } break; - case GGML_TYPE_F32: - { - ggml_compute_forward_get_rows_back_f32(params, dst); - } break; - default: - { - GGML_ABORT("fatal error"); - } - } - - //static bool first = true; - //printf("ne0 = %d, ne1 = %d, ne2 = %d\n", dst->ne[0], dst->ne[1], dst->ne[2]); - //if (first) { - // first = false; - //} else { - // for (int k = 0; k < dst->ne[1]; ++k) { - // for (int j = 0; j < dst->ne[0]/16; ++j) { - // for (int i = 0; i < 16; ++i) { - // printf("%8.4f ", ((float *) dst->data)[k*dst->ne[0] + j*16 + i]); - // } - // printf("\n"); - // } - // printf("\n"); - // } - // printf("\n"); - // exit(0); - //} -} - -// ggml_compute_forward_diag - -static void ggml_compute_forward_diag_f32( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - if (params->ith != 0) { - return; - } - - // TODO: handle transposed/permuted matrices - - GGML_TENSOR_UNARY_OP_LOCALS - - GGML_ASSERT(ne00 == ne0); - GGML_ASSERT(ne00 == ne1); - GGML_ASSERT(ne01 == 1); - GGML_ASSERT(ne02 == ne2); - GGML_ASSERT(ne03 == ne3); - - GGML_ASSERT(nb00 == sizeof(float)); - GGML_ASSERT(nb0 == sizeof(float)); - - for (int i3 = 0; i3 < ne3; i3++) { - for (int i2 = 0; i2 < ne2; i2++) { - for (int i1 = 0; i1 < ne1; i1++) { - float * d = (float *)((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1); - float * s = (float *)((char *) src0->data + i3*nb03 + i2*nb02); - for (int i0 = 0; i0 < i1; i0++) { - d[i0] = 0; - } - d[i1] = s[i1]; - for (int i0 = i1+1; i0 < ne0; i0++) { - d[i0] = 0; - } - } - } - } -} - -static void ggml_compute_forward_diag( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - switch (src0->type) { - case GGML_TYPE_F32: - { - ggml_compute_forward_diag_f32(params, dst); - } break; - default: - { - GGML_ABORT("fatal error"); - } - } -} - -// ggml_compute_forward_diag_mask_inf - -static void ggml_compute_forward_diag_mask_f32( - const struct ggml_compute_params * params, - struct ggml_tensor * dst, - const float value) { - - const struct ggml_tensor * src0 = dst->src[0]; - - const int ith = params->ith; - const int nth = params->nth; - - const int n_past = ((int32_t *) dst->op_params)[0]; - const bool inplace = src0->data == dst->data; - - GGML_ASSERT(n_past >= 0); - - if (!inplace) { - if (ith == 0) { - // memcpy needs to be synchronized across threads to avoid race conditions. - // => do it in INIT phase - GGML_ASSERT(ggml_nelements(dst) == ggml_nelements(src0)); - GGML_ASSERT(ggml_is_contiguous(dst) && ggml_is_contiguous(src0)); - memcpy( - ((char *) dst->data), - ((char *) src0->data), - ggml_nbytes(dst)); - } - ggml_barrier(params->threadpool); - } - - // TODO: handle transposed/permuted matrices - - const int n = ggml_nrows(src0); - const int nc = src0->ne[0]; - const int nr = src0->ne[1]; - const int nz = n/nr; - - GGML_ASSERT( dst->nb[0] == sizeof(float)); - GGML_ASSERT(src0->nb[0] == sizeof(float)); - - for (int k = 0; k < nz; k++) { - for (int j = ith; j < nr; j += nth) { - for (int i = n_past; i < nc; i++) { - if (i > n_past + j) { - *(float *)((char *) dst->data + k*dst->nb[2] + j*dst->nb[1] + i*dst->nb[0]) = value; - } - } - } - } -} - -static void ggml_compute_forward_diag_mask_inf( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - switch (src0->type) { - case GGML_TYPE_F32: - { - ggml_compute_forward_diag_mask_f32(params, dst, -INFINITY); - } break; - default: - { - GGML_ABORT("fatal error"); - } - } -} - -static void ggml_compute_forward_diag_mask_zero( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - switch (src0->type) { - case GGML_TYPE_F32: - { - ggml_compute_forward_diag_mask_f32(params, dst, 0); - } break; - default: - { - GGML_ABORT("fatal error"); - } - } -} - -// ggml_compute_forward_soft_max - -static void ggml_compute_forward_soft_max_f32( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - const struct ggml_tensor * src1 = dst->src[1]; - - assert(ggml_is_contiguous(dst)); - assert(ggml_are_same_shape(src0, dst)); - - float scale = 1.0f; - float max_bias = 0.0f; - - memcpy(&scale, (float *) dst->op_params + 0, sizeof(float)); - memcpy(&max_bias, (float *) dst->op_params + 1, sizeof(float)); - - // TODO: handle transposed/permuted matrices - - const int ith = params->ith; - const int nth = params->nth; - - GGML_TENSOR_UNARY_OP_LOCALS - - //const int64_t ne11 = src1 ? src1->ne[1] : 1; - - // TODO: is this supposed to be ceil instead of floor? - // https://huggingface.co/mosaicml/mpt-7b/blob/main/attention.py#L370 - const uint32_t n_head = ne02; - const uint32_t n_head_log2 = 1u << (uint32_t) floor(log2(n_head)); - - const float m0 = powf(2.0f, -(max_bias ) / n_head_log2); - const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2); - - const int nc = src0->ne[0]; - const int nr = ggml_nrows(src0); - - // rows per thread - const int dr = (nr + nth - 1)/nth; - - // row range for this thread - const int ir0 = dr*ith; - const int ir1 = MIN(ir0 + dr, nr); - - float * wp = (float *) params->wdata + (nc + CACHE_LINE_SIZE_F32) * ith; - - const bool use_f16 = (src1 && src1->type == GGML_TYPE_F16); - - for (int i1 = ir0; i1 < ir1; i1++) { - // ALiBi - const uint32_t h = (i1/ne01)%ne02; // head - const float slope = (max_bias > 0.0f) ? h < n_head_log2 ? powf(m0, h + 1) : powf(m1, 2*(h - n_head_log2) + 1) : 1.0f; - - float * sp = (float *)((char *) src0->data + i1*src0->nb[1]); - float * dp = (float *)((char *) dst->data + i1*dst->nb[1]); - - // broadcast the mask across rows - ggml_fp16_t * mp_f16 = src1 ? (ggml_fp16_t *)((char *) src1->data) + (i1%ne01)*ne00 : NULL; - float * mp_f32 = src1 ? (float *)((char *) src1->data) + (i1%ne01)*ne00 : NULL; - - ggml_vec_cpy_f32 (nc, wp, sp); - ggml_vec_scale_f32(nc, wp, scale); - if (mp_f32) { - if (use_f16) { - for (int i = 0; i < nc; ++i) { - wp[i] += slope*GGML_FP16_TO_FP32(mp_f16[i]); - } - } else { - for (int i = 0; i < nc; ++i) { - wp[i] += slope*mp_f32[i]; - } - } - } - -#ifndef NDEBUG - for (int i = 0; i < nc; ++i) { - //printf("p[%d] = %f\n", i, p[i]); - assert(!isnan(wp[i])); - } -#endif - - float max = -INFINITY; - ggml_vec_max_f32(nc, &max, wp); - - ggml_float sum = ggml_vec_soft_max_f32(nc, dp, wp, max); - assert(sum > 0.0); - - sum = 1.0/sum; - ggml_vec_scale_f32(nc, dp, sum); - -#ifndef NDEBUG - for (int i = 0; i < nc; ++i) { - assert(!isnan(dp[i])); - assert(!isinf(dp[i])); - } -#endif - } -} - -static void ggml_compute_forward_soft_max( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - switch (src0->type) { - case GGML_TYPE_F32: - { - ggml_compute_forward_soft_max_f32(params, dst); - } break; - default: - { - GGML_ABORT("fatal error"); - } - } -} - - -// ggml_compute_forward_soft_max_ext_back - -static void ggml_compute_forward_soft_max_ext_back_f32( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - const struct ggml_tensor * src1 = dst->src[1]; - - GGML_ASSERT(ggml_is_contiguous(src0)); - GGML_ASSERT(ggml_is_contiguous(src1)); - GGML_ASSERT(ggml_is_contiguous(dst)); - GGML_ASSERT(ggml_are_same_shape(src0, dst)); - GGML_ASSERT(ggml_are_same_shape(src1, dst)); - - float scale = 1.0f; - float max_bias = 0.0f; - - memcpy(&scale, (const float *) dst->op_params + 0, sizeof(float)); - memcpy(&max_bias, (const float *) dst->op_params + 1, sizeof(float)); - - GGML_ASSERT(max_bias == 0.0f); - - // TODO: handle transposed/permuted matrices - - const int ith = params->ith; - const int nth = params->nth; - - const int nc = src0->ne[0]; - const int nr = ggml_nrows(src0); - - // rows per thread - const int dr = (nr + nth - 1)/nth; - - // row range for this thread - const int ir0 = dr*ith; - const int ir1 = MIN(ir0 + dr, nr); - - for (int i1 = ir0; i1 < ir1; i1++) { - float *dy = (float *)((char *) src0->data + i1*src0->nb[1]); - float *y = (float *)((char *) src1->data + i1*src1->nb[1]); - float *dx = (float *)((char *) dst->data + i1*dst->nb[1]); - -#ifndef NDEBUG - for (int i = 0; i < nc; ++i) { - //printf("p[%d] = %f\n", i, p[i]); - assert(!isnan(dy[i])); - assert(!isnan(y[i])); - } -#endif - // Jii = yi - yi*yi - // Jij = -yi*yj - // J = diag(y)-y.T*y - // dx = J * dy - // dxk = sum_i(Jki * dyi) - // dxk = sum_i(-yk*yi * dyi) - (-yk*yk)*dyk + (yk - yk*yk)*dyk - // dxk = sum_i(-yk*yi * dyi) + yk*yk*dyk + yk*dyk - yk*yk*dyk - // dxk = sum_i(-yk*yi * dyi) + yk*dyk - // dxk = -yk * sum_i(yi * dyi) + yk*dyk - // dxk = -yk * dot(y, dy) + yk*dyk - // dxk = yk * (- dot(y, dy) + dyk) - // dxk = yk * (dyk - dot(y, dy)) - // - // post-order: - // dot_y_dy := dot(y, dy) - // dx := dy - // dx := dx - dot_y_dy - // dx := dx * y - - // linear runtime, no additional memory - float dot_y_dy = 0; - ggml_vec_dot_f32 (nc, &dot_y_dy, 0, y, 0, dy, 0, 1); - ggml_vec_cpy_f32 (nc, dx, dy); - ggml_vec_acc1_f32 (nc, dx, -dot_y_dy); - ggml_vec_mul_f32 (nc, dx, dx, y); - ggml_vec_scale_f32(nc, dx, scale); - -#ifndef NDEBUG - for (int i = 0; i < nc; ++i) { - assert(!isnan(dx[i])); - assert(!isinf(dx[i])); - } -#endif - } -} - -static void ggml_compute_forward_soft_max_ext_back( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - switch (src0->type) { - case GGML_TYPE_F32: - { - ggml_compute_forward_soft_max_ext_back_f32(params, dst); - } break; - default: - { - GGML_ABORT("fatal error"); - } - } -} - -// ggml_compute_forward_clamp - -static void ggml_compute_forward_clamp_f32( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - if (params->ith != 0) { - return; - } - - float min; - float max; - memcpy(&min, (float *) dst->op_params + 0, sizeof(float)); - memcpy(&max, (float *) dst->op_params + 1, sizeof(float)); - - const int ith = params->ith; - const int nth = params->nth; - - const int n = ggml_nrows(src0); - const int nc = src0->ne[0]; - - const size_t nb00 = src0->nb[0]; - const size_t nb01 = src0->nb[1]; - - const size_t nb0 = dst->nb[0]; - const size_t nb1 = dst->nb[1]; - - GGML_ASSERT( nb0 == sizeof(float)); - GGML_ASSERT(nb00 == sizeof(float)); - - for (int j = ith; j < n; j += nth) { - float * dst_ptr = (float *) ((char *) dst->data + j*nb1); - float * src0_ptr = (float *) ((char *) src0->data + j*nb01); - - for (int i = 0; i < nc; i++) { - dst_ptr[i] = MAX(MIN(src0_ptr[i], max), min); - } - } -} - -static void ggml_compute_forward_clamp( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - switch (src0->type) { - case GGML_TYPE_F32: - { - ggml_compute_forward_clamp_f32(params, dst); - } break; - case GGML_TYPE_F16: - case GGML_TYPE_BF16: - case GGML_TYPE_Q4_0: - case GGML_TYPE_Q4_1: - case GGML_TYPE_Q5_0: - case GGML_TYPE_Q5_1: - case GGML_TYPE_Q8_0: - case GGML_TYPE_Q8_1: - case GGML_TYPE_Q2_K: - case GGML_TYPE_Q3_K: - case GGML_TYPE_Q4_K: - case GGML_TYPE_Q5_K: - case GGML_TYPE_Q6_K: - case GGML_TYPE_TQ1_0: - case GGML_TYPE_TQ2_0: - case GGML_TYPE_IQ2_XXS: - case GGML_TYPE_IQ2_XS: - case GGML_TYPE_IQ3_XXS: - case GGML_TYPE_IQ1_S: - case GGML_TYPE_IQ1_M: - case GGML_TYPE_IQ4_NL: - case GGML_TYPE_IQ4_XS: - case GGML_TYPE_IQ3_S: - case GGML_TYPE_IQ2_S: - case GGML_TYPE_Q8_K: - case GGML_TYPE_I8: - case GGML_TYPE_I16: - case GGML_TYPE_I32: - case GGML_TYPE_I64: - case GGML_TYPE_F64: - case GGML_TYPE_COUNT: - { - GGML_ABORT("fatal error"); - } - } -} - -// ggml_compute_forward_rope - -static float rope_yarn_ramp(const float low, const float high, const int i0) { - const float y = (i0 / 2 - low) / MAX(0.001f, high - low); - return 1 - MIN(1, MAX(0, y)); -} - -// YaRN algorithm based on LlamaYaRNScaledRotaryEmbedding.py from https://github.com/jquesnelle/yarn -// MIT licensed. Copyright (c) 2023 Jeffrey Quesnelle and Bowen Peng. -static void rope_yarn( - float theta_extrap, float freq_scale, float corr_dims[2], int64_t i0, float ext_factor, float mscale, - float * cos_theta, float * sin_theta) { - // Get n-d rotational scaling corrected for extrapolation - float theta_interp = freq_scale * theta_extrap; - float theta = theta_interp; - if (ext_factor != 0.0f) { - float ramp_mix = rope_yarn_ramp(corr_dims[0], corr_dims[1], i0) * ext_factor; - theta = theta_interp * (1 - ramp_mix) + theta_extrap * ramp_mix; - - // Get n-d magnitude scaling corrected for interpolation - mscale *= 1.0f + 0.1f * logf(1.0f / freq_scale); - } - *cos_theta = cosf(theta) * mscale; - *sin_theta = sinf(theta) * mscale; -} - -static void ggml_rope_cache_init( - float theta_base, float freq_scale, const float * freq_factors, float corr_dims[2], int64_t ne0, float ext_factor, float mscale, - float * cache, float sin_sign, float theta_scale) { - // ref: https://github.com/jquesnelle/yarn/blob/master/scaled_rope/LlamaYaRNScaledRotaryEmbedding.py - float theta = theta_base; - for (int64_t i0 = 0; i0 < ne0; i0 += 2) { - const float ff = freq_factors ? freq_factors[i0/2] : 1.0f; - rope_yarn( - theta/ff, freq_scale, corr_dims, i0, ext_factor, mscale, &cache[i0 + 0], &cache[i0 + 1] - ); - cache[i0 + 1] *= sin_sign; - - theta *= theta_scale; - } -} - -static void ggml_mrope_cache_init( - float theta_base_t, float theta_base_h, float theta_base_w, float theta_base_e, int sections[4], bool indep_sects, - float freq_scale, const float * freq_factors, float corr_dims[2], int64_t ne0, float ext_factor, float mscale, - float * cache, float sin_sign, float theta_scale) { - // ref: https://github.com/jquesnelle/yarn/blob/master/scaled_rope/LlamaYaRNScaledRotaryEmbedding.py - float theta_t = theta_base_t; - float theta_h = theta_base_h; - float theta_w = theta_base_w; - float theta_e = theta_base_e; // extra position id for vision encoder - int sect_dims = sections[0] + sections[1] + sections[2] + sections[3]; - int sec_w = sections[1] + sections[0]; - int sec_e = sections[2] + sec_w; - GGML_ASSERT(sect_dims <= ne0); - - for (int64_t i0 = 0; i0 < ne0; i0 += 2) { - const float ff = freq_factors ? freq_factors[i0/2] : 1.0f; - - int sector = (i0 / 2) % sect_dims; - if (indep_sects) { - // compute theta independently for each dim sections - // (i.e. reset corresponding theta when `i0` go from one section to another) - if (sector == 0) { - theta_t = theta_base_t; - } - else if (sector == sections[0]) { - theta_h = theta_base_h;; - } - else if (sector == sec_w) { - theta_w = theta_base_w; - } - else if (sector == sec_e) { - theta_e = theta_base_e; - } - } - - float theta = theta_t; - if (sector >= sections[0] && sector < sec_w) { - theta = theta_h; - } - else if (sector >= sec_w && sector < sec_w + sections[2]) { - theta = theta_w; - } - else if (sector >= sec_w + sections[2]) { - theta = theta_e; - } - - rope_yarn( - theta/ff, freq_scale, corr_dims, i0, ext_factor, mscale, &cache[i0 + 0], &cache[i0 + 1] - ); - cache[i0 + 1] *= sin_sign; - - theta_t *= theta_scale; - theta_w *= theta_scale; - theta_h *= theta_scale; - theta_e *= theta_scale; - } -} - -static void ggml_compute_forward_rope_f32( - const struct ggml_compute_params * params, - struct ggml_tensor * dst, - const bool forward) { - - const struct ggml_tensor * src0 = dst->src[0]; - const struct ggml_tensor * src1 = dst->src[1]; - const struct ggml_tensor * src2 = dst->src[2]; - - float freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow; - int sections[4]; - - //const int n_past = ((int32_t *) dst->op_params)[0]; - const int n_dims = ((int32_t *) dst->op_params)[1]; - const int mode = ((int32_t *) dst->op_params)[2]; - //const int n_ctx = ((int32_t *) dst->op_params)[3]; - const int n_ctx_orig = ((int32_t *) dst->op_params)[4]; - - memcpy(&freq_base, (int32_t *) dst->op_params + 5, sizeof(float)); - memcpy(&freq_scale, (int32_t *) dst->op_params + 6, sizeof(float)); - memcpy(&ext_factor, (int32_t *) dst->op_params + 7, sizeof(float)); - memcpy(&attn_factor, (int32_t *) dst->op_params + 8, sizeof(float)); - memcpy(&beta_fast, (int32_t *) dst->op_params + 9, sizeof(float)); - memcpy(&beta_slow, (int32_t *) dst->op_params + 10, sizeof(float)); - memcpy(§ions, (int32_t *) dst->op_params + 11, sizeof(int)*4); - - GGML_TENSOR_UNARY_OP_LOCALS - - //printf("ne0: %d, ne1: %d, ne2: %d, ne3: %d\n", ne0, ne1, ne2, ne3); - //printf("n_past = %d, ne2 = %d\n", n_past, ne2); - - GGML_ASSERT(nb00 == sizeof(float)); - - const int ith = params->ith; - const int nth = params->nth; - - const int nr = ggml_nrows(dst); - - GGML_ASSERT(n_dims <= ne0); - GGML_ASSERT(n_dims % 2 == 0); - - // rows per thread - const int dr = (nr + nth - 1)/nth; - - // row range for this thread - const int ir0 = dr*ith; - const int ir1 = MIN(ir0 + dr, nr); - - // row index used to determine which thread to use - int ir = 0; - - const float theta_scale = powf(freq_base, -2.0f/n_dims); - - float corr_dims[2]; - ggml_rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow, corr_dims); - - const bool is_neox = mode & GGML_ROPE_TYPE_NEOX; - const bool is_mrope = mode & GGML_ROPE_TYPE_MROPE; // ggml_rope_multi, multimodal rotary position embedding - const bool is_vision = mode == GGML_ROPE_TYPE_VISION; - - if (is_mrope) { - GGML_ASSERT(sections[0] > 0 || sections[1] > 0 || sections[2] > 0); - } - - if (is_vision) { - GGML_ASSERT(n_dims == ne0/2); - } - - const float * freq_factors = NULL; - if (src2 != NULL) { - GGML_ASSERT(src2->type == GGML_TYPE_F32); - GGML_ASSERT(src2->ne[0] >= n_dims / 2); - freq_factors = (const float *) src2->data; - } - - // backward process uses inverse rotation by cos and sin. - // cos and sin build a rotation matrix, where the inverse is the transpose. - // this essentially just switches the sign of sin. - const float sin_sign = forward ? 1.0f : -1.0f; - - const int32_t * pos = (const int32_t *) src1->data; - - for (int64_t i3 = 0; i3 < ne3; i3++) { // batch - for (int64_t i2 = 0; i2 < ne2; i2++) { // seq-len - - float * cache = (float *) params->wdata + (ne0 + CACHE_LINE_SIZE_F32)*ith; - if (!is_mrope) { - const int64_t p = pos[i2]; - ggml_rope_cache_init(p, freq_scale, freq_factors, corr_dims, ne0, ext_factor, attn_factor, cache, sin_sign, theta_scale); - } - else { - const int64_t p_t = pos[i2]; - const int64_t p_h = pos[i2 + ne2]; - const int64_t p_w = pos[i2 + ne2 * 2]; - const int64_t p_e = pos[i2 + ne2 * 3]; - ggml_mrope_cache_init( - p_t, p_h, p_w, p_e, sections, is_vision, - freq_scale, freq_factors, corr_dims, ne0, ext_factor, attn_factor, cache, sin_sign, theta_scale); - } - - for (int64_t i1 = 0; i1 < ne1; i1++) { // attn-heads - if (ir++ < ir0) continue; - if (ir > ir1) break; - - if (is_neox || is_mrope) { - if (is_vision){ - for (int64_t i0 = 0; i0 < n_dims; i0 += 2) { - const int64_t ic = i0/2; - - const float cos_theta = cache[i0 + 0]; - const float sin_theta = cache[i0 + 1]; - - const float * const src = (float *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + ic*nb00); - float * dst_data = (float *)((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + ic*nb0); - - const float x0 = src[0]; - const float x1 = src[n_dims]; - - dst_data[0] = x0*cos_theta - x1*sin_theta; - dst_data[n_dims] = x0*sin_theta + x1*cos_theta; - } - } else { - for (int64_t i0 = 0; i0 < n_dims; i0 += 2) { - const int64_t ic = i0/2; - - const float cos_theta = cache[i0 + 0]; - const float sin_theta = cache[i0 + 1]; - - const float * const src = (float *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + ic*nb00); - float * dst_data = (float *)((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + ic*nb0); - - const float x0 = src[0]; - const float x1 = src[n_dims/2]; - - dst_data[0] = x0*cos_theta - x1*sin_theta; - dst_data[n_dims/2] = x0*sin_theta + x1*cos_theta; - } - } - } else { - for (int64_t i0 = 0; i0 < n_dims; i0 += 2) { - const float cos_theta = cache[i0 + 0]; - const float sin_theta = cache[i0 + 1]; - - const float * const src = (float *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00); - float * dst_data = (float *)((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0); - - const float x0 = src[0]; - const float x1 = src[1]; - - dst_data[0] = x0*cos_theta - x1*sin_theta; - dst_data[1] = x0*sin_theta + x1*cos_theta; - } - } - - if (is_vision) { - for (int64_t i0 = n_dims; i0 < ne0; i0 += 2) { - const int64_t ic = i0/2; - - const float cos_theta = cache[i0 + 0]; - const float sin_theta = cache[i0 + 1]; - - const float * const src = (float *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + ic*nb00); - float * dst_data = (float *)((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + ic*nb0); - - const float x0 = src[0]; - const float x1 = src[n_dims]; - - dst_data[0] = x0*cos_theta - x1*sin_theta; - dst_data[n_dims] = x0*sin_theta + x1*cos_theta; - } - } else { - // fill the remain channels with data from src tensor - for (int64_t i0 = n_dims; i0 < ne0; i0 += 2) { - const float * const src = (float *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00); - float * dst_data = (float *)((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0); - - dst_data[0] = src[0]; - dst_data[1] = src[1]; - } - } - } - } - } -} - -// TODO: deduplicate f16/f32 code -static void ggml_compute_forward_rope_f16( - const struct ggml_compute_params * params, - struct ggml_tensor * dst, - const bool forward) { - - const struct ggml_tensor * src0 = dst->src[0]; - const struct ggml_tensor * src1 = dst->src[1]; - const struct ggml_tensor * src2 = dst->src[2]; - - float freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow; - int sections[4]; - - //const int n_past = ((int32_t *) dst->op_params)[0]; - const int n_dims = ((int32_t *) dst->op_params)[1]; - const int mode = ((int32_t *) dst->op_params)[2]; - //const int n_ctx = ((int32_t *) dst->op_params)[3]; - const int n_ctx_orig = ((int32_t *) dst->op_params)[4]; - memcpy(&freq_base, (int32_t *) dst->op_params + 5, sizeof(float)); - memcpy(&freq_scale, (int32_t *) dst->op_params + 6, sizeof(float)); - memcpy(&ext_factor, (int32_t *) dst->op_params + 7, sizeof(float)); - memcpy(&attn_factor, (int32_t *) dst->op_params + 8, sizeof(float)); - memcpy(&beta_fast, (int32_t *) dst->op_params + 9, sizeof(float)); - memcpy(&beta_slow, (int32_t *) dst->op_params + 10, sizeof(float)); - memcpy(§ions, (int32_t *) dst->op_params + 11, sizeof(int)*4); - - - GGML_TENSOR_UNARY_OP_LOCALS - - //printf("ne0: %d, ne1: %d, ne2: %d, ne3: %d\n", ne0, ne1, ne2, ne3); - //printf("n_past = %d, ne2 = %d\n", n_past, ne2); - - GGML_ASSERT(nb0 == sizeof(ggml_fp16_t)); - - const int ith = params->ith; - const int nth = params->nth; - - const int nr = ggml_nrows(dst); - - GGML_ASSERT(n_dims <= ne0); - GGML_ASSERT(n_dims % 2 == 0); - - // rows per thread - const int dr = (nr + nth - 1)/nth; - - // row range for this thread - const int ir0 = dr*ith; - const int ir1 = MIN(ir0 + dr, nr); - - // row index used to determine which thread to use - int ir = 0; - - const float theta_scale = powf(freq_base, -2.0f/n_dims); - - float corr_dims[2]; - ggml_rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow, corr_dims); - - const bool is_neox = mode & GGML_ROPE_TYPE_NEOX; - const bool is_mrope = mode & GGML_ROPE_TYPE_MROPE; - const bool is_vision = mode == GGML_ROPE_TYPE_VISION; - - if (is_mrope) { - GGML_ASSERT(sections[0] > 0 || sections[1] > 0 || sections[2] > 0); - } - - if (is_vision) { - GGML_ASSERT(n_dims == ne0/2); - } - - const float * freq_factors = NULL; - if (src2 != NULL) { - GGML_ASSERT(src2->type == GGML_TYPE_F32); - GGML_ASSERT(src2->ne[0] >= n_dims / 2); - freq_factors = (const float *) src2->data; - } - - // backward process uses inverse rotation by cos and sin. - // cos and sin build a rotation matrix, where the inverse is the transpose. - // this essentially just switches the sign of sin. - const float sin_sign = forward ? 1.0f : -1.0f; - - const int32_t * pos = (const int32_t *) src1->data; - - for (int64_t i3 = 0; i3 < ne3; i3++) { - for (int64_t i2 = 0; i2 < ne2; i2++) { - - float * cache = (float *) params->wdata + (ne0 + CACHE_LINE_SIZE_F32)*ith; - if (!is_mrope) { - const int64_t p = pos[i2]; - ggml_rope_cache_init(p, freq_scale, freq_factors, corr_dims, ne0, ext_factor, attn_factor, cache, sin_sign, theta_scale); - } - else { - const int64_t p_t = pos[i2]; - const int64_t p_h = pos[i2 + ne2]; - const int64_t p_w = pos[i2 + ne2 * 2]; - const int64_t p_e = pos[i2 + ne2 * 3]; - ggml_mrope_cache_init( - p_t, p_h, p_w, p_e, sections, is_vision, - freq_scale, freq_factors, corr_dims, ne0, ext_factor, attn_factor, cache, sin_sign, theta_scale); - } - - for (int64_t i1 = 0; i1 < ne1; i1++) { - if (ir++ < ir0) continue; - if (ir > ir1) break; - - if (is_neox || is_mrope) { - if (is_vision) { - for (int64_t i0 = 0; i0 < n_dims; i0 += 2) { - const int64_t ic = i0/2; - - const float cos_theta = cache[i0 + 0]; - const float sin_theta = cache[i0 + 1]; - - const ggml_fp16_t * const src = (ggml_fp16_t *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + ic*nb00); - ggml_fp16_t * dst_data = (ggml_fp16_t *)((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + ic*nb0); - - const float x0 = GGML_FP16_TO_FP32(src[0]); - const float x1 = GGML_FP16_TO_FP32(src[n_dims]); - - dst_data[0] = GGML_FP32_TO_FP16(x0*cos_theta - x1*sin_theta); - dst_data[n_dims] = GGML_FP32_TO_FP16(x0*sin_theta + x1*cos_theta); - } - } else { - for (int64_t i0 = 0; i0 < n_dims; i0 += 2) { - const int64_t ic = i0/2; - - const float cos_theta = cache[i0 + 0]; - const float sin_theta = cache[i0 + 1]; - - const ggml_fp16_t * const src = (ggml_fp16_t *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + ic*nb00); - ggml_fp16_t * dst_data = (ggml_fp16_t *)((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + ic*nb0); - - const float x0 = GGML_FP16_TO_FP32(src[0]); - const float x1 = GGML_FP16_TO_FP32(src[n_dims/2]); - - dst_data[0] = GGML_FP32_TO_FP16(x0*cos_theta - x1*sin_theta); - dst_data[n_dims/2] = GGML_FP32_TO_FP16(x0*sin_theta + x1*cos_theta); - } - } - } else { - for (int64_t i0 = 0; i0 < n_dims; i0 += 2) { - const float cos_theta = cache[i0 + 0]; - const float sin_theta = cache[i0 + 1]; - - const ggml_fp16_t * const src = (ggml_fp16_t *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00); - ggml_fp16_t * dst_data = (ggml_fp16_t *)((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0); - - const float x0 = GGML_FP16_TO_FP32(src[0]); - const float x1 = GGML_FP16_TO_FP32(src[1]); - - dst_data[0] = GGML_FP32_TO_FP16(x0*cos_theta - x1*sin_theta); - dst_data[1] = GGML_FP32_TO_FP16(x0*sin_theta + x1*cos_theta); - } - } - - if (is_vision) { - for (int64_t i0 = n_dims; i0 < ne0; i0 += 2) { - const int64_t ic = i0/2; - - const float cos_theta = cache[i0 + 0]; - const float sin_theta = cache[i0 + 1]; - - const ggml_fp16_t * const src = (ggml_fp16_t *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + ic*nb00); - ggml_fp16_t * dst_data = (ggml_fp16_t *)((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + ic*nb0); - - const float x0 = GGML_FP16_TO_FP32(src[0]); - const float x1 = GGML_FP16_TO_FP32(src[n_dims]); - - dst_data[0] = GGML_FP32_TO_FP16(x0*cos_theta - x1*sin_theta); - dst_data[n_dims] = GGML_FP32_TO_FP16(x0*sin_theta + x1*cos_theta); - } - } else { - for (int64_t i0 = n_dims; i0 < ne0; i0 += 2) { - const ggml_fp16_t * const src = (ggml_fp16_t *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00); - ggml_fp16_t * dst_data = (ggml_fp16_t *)((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0); - - dst_data[0] = src[0]; - dst_data[1] = src[1]; - } - } - } - } - } -} - -static void ggml_compute_forward_rope( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - switch (src0->type) { - case GGML_TYPE_F16: - { - ggml_compute_forward_rope_f16(params, dst, true); - } break; - case GGML_TYPE_F32: - { - ggml_compute_forward_rope_f32(params, dst, true); - } break; - default: - { - GGML_ABORT("fatal error"); - } - } -} - -// ggml_compute_forward_rope_back - -static void ggml_compute_forward_rope_back( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - switch (src0->type) { - case GGML_TYPE_F16: - { - ggml_compute_forward_rope_f16(params, dst, false); - } break; - case GGML_TYPE_F32: - { - ggml_compute_forward_rope_f32(params, dst, false); - } break; - default: - { - GGML_ABORT("fatal error"); - } - } -} - -// ggml_compute_forward_conv_transpose_1d - -static void ggml_compute_forward_conv_transpose_1d_f16_f32( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - const struct ggml_tensor * src1 = dst->src[1]; - - GGML_ASSERT(src0->type == GGML_TYPE_F16); - GGML_ASSERT(src1->type == GGML_TYPE_F32); - GGML_ASSERT( dst->type == GGML_TYPE_F32); - - GGML_TENSOR_BINARY_OP_LOCALS - - const int ith = params->ith; - const int nth = params->nth; - - const int nk = ne00*ne01*ne02; - - GGML_ASSERT(nb00 == sizeof(ggml_fp16_t)); - GGML_ASSERT(nb10 == sizeof(float)); - - if (ith == 0) { - memset(params->wdata, 0, params->wsize); - - // permute kernel data (src0) from (K x Cout x Cin) to (Cin x K x Cout) - { - ggml_fp16_t * const wdata = (ggml_fp16_t *) params->wdata + 0; - - for (int64_t i02 = 0; i02 < ne02; i02++) { - for (int64_t i01 = 0; i01 < ne01; i01++) { - const ggml_fp16_t * const src = (ggml_fp16_t *)((char *) src0->data + i02*nb02 + i01*nb01); - ggml_fp16_t * dst_data = wdata + i01*ne00*ne02; - for (int64_t i00 = 0; i00 < ne00; i00++) { - dst_data[i00*ne02 + i02] = src[i00]; - } - } - } - } - - // permute source data (src1) from (L x Cin) to (Cin x L) - { - ggml_fp16_t * const wdata = (ggml_fp16_t *) params->wdata + nk; - ggml_fp16_t * dst_data = wdata; - - for (int64_t i11 = 0; i11 < ne11; i11++) { - const float * const src = (float *)((char *) src1->data + i11*nb11); - for (int64_t i10 = 0; i10 < ne10; i10++) { - dst_data[i10*ne11 + i11] = GGML_FP32_TO_FP16(src[i10]); - } - } - } - - // need to zero dst since we are accumulating into it - memset(dst->data, 0, ggml_nbytes(dst)); - } - ggml_barrier(params->threadpool); - - const int32_t s0 = ((const int32_t*)(dst->op_params))[0]; - - // total rows in dst - const int nr = ne1; - - // rows per thread - const int dr = (nr + nth - 1)/nth; - - // row range for this thread - const int ir0 = dr*ith; - const int ir1 = MIN(ir0 + dr, nr); - - ggml_fp16_t * const wdata = (ggml_fp16_t *) params->wdata + 0; - ggml_fp16_t * const wdata_src = wdata + nk; - - for (int i1 = ir0; i1 < ir1; i1++) { - float * dst_data = (float *)((char *) dst->data + i1*nb1); - ggml_fp16_t * wdata_kernel = wdata + i1*ne02*ne00; - for (int i10 = 0; i10 < ne10; i10++) { - const int i1n = i10*ne11; - for (int i00 = 0; i00 < ne00; i00++) { - float v = 0; - ggml_vec_dot_f16(ne02, &v, 0, - (ggml_fp16_t *) wdata_src + i1n, 0, - (ggml_fp16_t *) wdata_kernel + i00*ne02, 0, 1); - dst_data[i10*s0 + i00] += v; - } - } - } -} - -static void ggml_compute_forward_conv_transpose_1d_f32( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - const struct ggml_tensor * src1 = dst->src[1]; - - GGML_ASSERT(src0->type == GGML_TYPE_F32); - GGML_ASSERT(src1->type == GGML_TYPE_F32); - GGML_ASSERT( dst->type == GGML_TYPE_F32); - - GGML_TENSOR_BINARY_OP_LOCALS - - const int ith = params->ith; - const int nth = params->nth; - - const int nk = ne00*ne01*ne02; - - GGML_ASSERT(nb00 == sizeof(float)); - GGML_ASSERT(nb10 == sizeof(float)); - - if (ith == 0) { - memset(params->wdata, 0, params->wsize); - - // prepare kernel data (src0) from (K x Cout x Cin) to (Cin x K x Cout) - { - float * const wdata = (float *) params->wdata + 0; - - for (int64_t i02 = 0; i02 < ne02; i02++) { - for (int64_t i01 = 0; i01 < ne01; i01++) { - const float * const src = (float *)((char *) src0->data + i02*nb02 + i01*nb01); - float * dst_data = wdata + i01*ne00*ne02; - for (int64_t i00 = 0; i00 < ne00; i00++) { - dst_data[i00*ne02 + i02] = src[i00]; - } - } - } - } - - // prepare source data (src1) - { - float * const wdata = (float *) params->wdata + nk; - float * dst_data = wdata; - - for (int64_t i11 = 0; i11 < ne11; i11++) { - const float * const src = (float *)((char *) src1->data + i11*nb11); - for (int64_t i10 = 0; i10 < ne10; i10++) { - dst_data[i10*ne11 + i11] = src[i10]; - } - } - } - - // need to zero dst since we are accumulating into it - memset(dst->data, 0, ggml_nbytes(dst)); - } - ggml_barrier(params->threadpool); - - const int32_t s0 = ((const int32_t*)(dst->op_params))[0]; - - // total rows in dst - const int nr = ne1; - - // rows per thread - const int dr = (nr + nth - 1)/nth; - - // row range for this thread - const int ir0 = dr*ith; - const int ir1 = MIN(ir0 + dr, nr); - - float * const wdata = (float *) params->wdata + 0; - float * const wdata_src = wdata + nk; - - for (int i1 = ir0; i1 < ir1; i1++) { - float * dst_data = (float *)((char *) dst->data + i1*nb1); - float * wdata_kernel = wdata + i1*ne02*ne00; - for (int i10 = 0; i10 < ne10; i10++) { - const int i1n = i10*ne11; - for (int i00 = 0; i00 < ne00; i00++) { - float v = 0; - ggml_vec_dot_f32(ne02, &v, 0, - wdata_src + i1n, 0, - wdata_kernel + i00*ne02, 0, 1); - dst_data[i10*s0 + i00] += v; - } - } - } -} - -static void ggml_compute_forward_conv_transpose_1d( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - switch (src0->type) { - case GGML_TYPE_F16: - { - ggml_compute_forward_conv_transpose_1d_f16_f32(params, dst); - } break; - case GGML_TYPE_F32: - { - ggml_compute_forward_conv_transpose_1d_f32(params, dst); - } break; - default: - { - GGML_ABORT("fatal error"); - } - } -} - -// ggml_compute_forward_im2col_f32 -// src0: kernel [OC, IC, KH, KW] -// src1: image [N, IC, IH, IW] -// dst: result [N, OH, OW, IC*KH*KW] -static void ggml_compute_forward_im2col_f32( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - const struct ggml_tensor * src1 = dst->src[1]; - - GGML_ASSERT(src1->type == GGML_TYPE_F32); - GGML_ASSERT( dst->type == GGML_TYPE_F32); - - GGML_TENSOR_BINARY_OP_LOCALS; - - const int32_t s0 = ((const int32_t *)(dst->op_params))[0]; - const int32_t s1 = ((const int32_t *)(dst->op_params))[1]; - const int32_t p0 = ((const int32_t *)(dst->op_params))[2]; - const int32_t p1 = ((const int32_t *)(dst->op_params))[3]; - const int32_t d0 = ((const int32_t *)(dst->op_params))[4]; - const int32_t d1 = ((const int32_t *)(dst->op_params))[5]; - const bool is_2D = ((const int32_t *)(dst->op_params))[6] == 1; - - const int ith = params->ith; - const int nth = params->nth; - - const int64_t N = is_2D ? ne13 : ne12; - const int64_t IC = is_2D ? ne12 : ne11; - const int64_t IH = is_2D ? ne11 : 1; - const int64_t IW = ne10; - - const int64_t KH = is_2D ? ne01 : 1; - const int64_t KW = ne00; - - const int64_t OH = is_2D ? ne2 : 1; - const int64_t OW = ne1; - - int ofs0 = is_2D ? nb13 : nb12; - int ofs1 = is_2D ? nb12 : nb11; - - GGML_ASSERT(nb10 == sizeof(float)); - - // im2col: [N, IC, IH, IW] => [N, OH, OW, IC*KH*KW] - { - float * const wdata = (float *) dst->data; - - for (int64_t in = 0; in < N; in++) { - for (int64_t ioh = 0; ioh < OH; ioh++) { // 1 - for (int64_t iow = 0; iow < OW; iow++) { - for (int64_t iic = ith; iic < IC; iic += nth) { - - // micro kernel - float * dst_data = wdata + (in*OH*OW + ioh*OW + iow)*(IC*KH*KW); // [IC, KH, KW] - const float * const src_data = (float *)((char *) src1->data + in*ofs0 + iic*ofs1); // [IH, IW] - - for (int64_t ikh = 0; ikh < KH; ikh++) { // 1 - for (int64_t ikw = 0; ikw < KW; ikw++) { - const int64_t iiw = iow*s0 + ikw*d0 - p0; - const int64_t iih = ioh*s1 + ikh*d1 - p1; - - if (iih < 0 || iih >= IH || iiw < 0 || iiw >= IW) { - dst_data[iic*(KH*KW) + ikh*KW + ikw] = 0; - } else { - dst_data[iic*(KH*KW) + ikh*KW + ikw] = (src_data[iih*IW + iiw]); - } - } - } - } - } - } - } - } -} - - -// ggml_compute_forward_im2col_f16 -// src0: kernel [OC, IC, KH, KW] -// src1: image [N, IC, IH, IW] -// dst: result [N, OH, OW, IC*KH*KW] -static void ggml_compute_forward_im2col_f16( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - const struct ggml_tensor * src1 = dst->src[1]; - - GGML_ASSERT(src0->type == GGML_TYPE_F16); - GGML_ASSERT(src1->type == GGML_TYPE_F32); - GGML_ASSERT( dst->type == GGML_TYPE_F16); - - GGML_TENSOR_BINARY_OP_LOCALS; - - const int32_t s0 = ((const int32_t *)(dst->op_params))[0]; - const int32_t s1 = ((const int32_t *)(dst->op_params))[1]; - const int32_t p0 = ((const int32_t *)(dst->op_params))[2]; - const int32_t p1 = ((const int32_t *)(dst->op_params))[3]; - const int32_t d0 = ((const int32_t *)(dst->op_params))[4]; - const int32_t d1 = ((const int32_t *)(dst->op_params))[5]; - const bool is_2D = ((const int32_t *)(dst->op_params))[6] == 1; - - const int ith = params->ith; - const int nth = params->nth; - - const int64_t N = is_2D ? ne13 : ne12; - const int64_t IC = is_2D ? ne12 : ne11; - const int64_t IH = is_2D ? ne11 : 1; - const int64_t IW = ne10; - - const int64_t KH = is_2D ? ne01 : 1; - const int64_t KW = ne00; - - const int64_t OH = is_2D ? ne2 : 1; - const int64_t OW = ne1; - - int ofs0 = is_2D ? nb13 : nb12; - int ofs1 = is_2D ? nb12 : nb11; - - GGML_ASSERT(nb00 == sizeof(ggml_fp16_t)); - GGML_ASSERT(nb10 == sizeof(float)); - - // im2col: [N, IC, IH, IW] => [N, OH, OW, IC*KH*KW] - { - ggml_fp16_t * const wdata = (ggml_fp16_t *) dst->data; - - for (int64_t in = 0; in < N; in++) { - for (int64_t ioh = 0; ioh < OH; ioh++) { // 1 - for (int64_t iow = 0; iow < OW; iow++) { - for (int64_t iic = ith; iic < IC; iic += nth) { - - // micro kernel - ggml_fp16_t * dst_data = wdata + (in*OH*OW + ioh*OW + iow)*(IC*KH*KW); // [IC, KH, KW] - const float * const src_data = (float *)((char *) src1->data + in*ofs0 + iic*ofs1); // [IH, IW] - - for (int64_t ikh = 0; ikh < KH; ikh++) { // 1 - for (int64_t ikw = 0; ikw < KW; ikw++) { - const int64_t iiw = iow*s0 + ikw*d0 - p0; - const int64_t iih = ioh*s1 + ikh*d1 - p1; - - if (iih < 0 || iih >= IH || iiw < 0 || iiw >= IW) { - dst_data[iic*(KH*KW) + ikh*KW + ikw] = 0; - } else { - dst_data[iic*(KH*KW) + ikh*KW + ikw] = GGML_FP32_TO_FP16(src_data[iih*IW + iiw]); - } - } - } - } - } - } - } - } -} - -static void ggml_compute_forward_im2col( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - switch (dst->type) { - case GGML_TYPE_F16: - { - ggml_compute_forward_im2col_f16(params, dst); - } break; - case GGML_TYPE_F32: - { - ggml_compute_forward_im2col_f32(params, dst); - } break; - default: - { - GGML_ABORT("fatal error"); - } - } -} - -// ggml_compute_forward_im2col_back_f32 - -static void ggml_compute_forward_im2col_back_f32( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; // gradients of forward pass output - const struct ggml_tensor * src1 = dst->src[1]; // convolution kernel - - GGML_ASSERT(src0->type == GGML_TYPE_F32); - GGML_ASSERT(src1->type == GGML_TYPE_F32); - GGML_ASSERT( dst->type == GGML_TYPE_F32); - - GGML_TENSOR_BINARY_OP_LOCALS; - - const int32_t s0 = ((const int32_t *)(dst->op_params))[0]; - const int32_t s1 = ((const int32_t *)(dst->op_params))[1]; - const int32_t p0 = ((const int32_t *)(dst->op_params))[2]; - const int32_t p1 = ((const int32_t *)(dst->op_params))[3]; - const int32_t d0 = ((const int32_t *)(dst->op_params))[4]; - const int32_t d1 = ((const int32_t *)(dst->op_params))[5]; - const bool is_2D = ((const int32_t *)(dst->op_params))[6] == 1; - - const int ith = params->ith; - const int nth = params->nth; - - const int64_t N = is_2D ? ne3 : ne2; - const int64_t IC = is_2D ? ne2 : ne1; - const int64_t IH = is_2D ? ne1 : 1; - const int64_t IW = ne0; - - const int64_t KH = is_2D ? ne11 : 1; - const int64_t KW = ne10; - - const int64_t OH = is_2D ? ne02 : 1; - const int64_t OW = ne01; - - int ofs0 = is_2D ? nb3 : nb2; - int ofs1 = is_2D ? nb2 : nb1; - - GGML_ASSERT(nb0 == sizeof(float)); - - // im2col: [N, IC, IH, IW] => [N, OH, OW, IC*KH*KW] - { - float * const wdata = (float *) dst->data; - - for (int64_t in = 0; in < N; in++) { - for (int64_t iic = ith; iic < IC; iic += nth) { - for (int64_t iih = 0; iih < IH; iih++) { - for (int64_t iiw = 0; iiw < IW; iiw++) { - - // micro kernel - float grad = 0.0f; - for (int64_t ikh = 0; ikh < KH; ikh++) { - for (int64_t ikw = 0; ikw < KW; ikw++) { - // For s0 > 1 some values were skipped over in the forward pass. - // These values have tmpw % s0 != 0 and need to be skipped in the backwards pass as well. - const int64_t tmpw = (iiw + p0 - ikw*d0); - if (tmpw % s0 != 0) { - continue; - } - const int64_t iow = tmpw / s0; - - // Equivalent logic as above except for s1. - int64_t ioh; - if (is_2D) { - const int64_t tmph = iih + p1 - ikh*d1; - - if (tmph % s1 != 0) { - continue; - } - - ioh = tmph / s1; - } else { - ioh = 0; - } - - if (iow < 0 || iow >= OW || ioh < 0 || ioh >= OH) { - continue; - } - - const float * const grad_in = (const float *) src0->data - + (in*OH*OW + ioh*OW + iow)*(IC*KH*KW); // [IC, KH, KW] - grad += grad_in[iic*(KH*KW) + ikh*KW + ikw]; - } - } - float * dst_data = (float *)((char *) wdata + (in*ofs0 + iic*ofs1)); // [IH, IW] - dst_data[iih*IW + iiw] = grad; - } - } - } - } - } -} - -// ggml_compute_forward_conv_transpose_2d - -static void ggml_compute_forward_conv_transpose_2d( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - const struct ggml_tensor * src1 = dst->src[1]; - - GGML_ASSERT(src0->type == GGML_TYPE_F16); - GGML_ASSERT(src1->type == GGML_TYPE_F32); - GGML_ASSERT( dst->type == GGML_TYPE_F32); - - GGML_TENSOR_BINARY_OP_LOCALS - - const int ith = params->ith; - const int nth = params->nth; - - const int nk = ne00*ne01*ne02*ne03; - - GGML_ASSERT(nb00 == sizeof(ggml_fp16_t)); - GGML_ASSERT(nb10 == sizeof(float)); - - if (ith == 0) { - memset(params->wdata, 0, params->wsize); - - // permute kernel data (src0) from (Kw x Kh x Cout x Cin) to (Cin x Kw x Kh x Cout) - { - ggml_fp16_t * const wdata = (ggml_fp16_t *) params->wdata + 0; - - for (int64_t i03 = 0; i03 < ne03; i03++) { - for (int64_t i02 = 0; i02 < ne02; i02++) { - const ggml_fp16_t * const src = (ggml_fp16_t *)((char *) src0->data + i03*nb03 + i02*nb02); - ggml_fp16_t * dst_data = wdata + i02*ne01*ne00*ne03; - for (int64_t i01 = 0; i01 < ne01; i01++) { - for (int64_t i00 = 0; i00 < ne00; i00++) { - dst_data[i01*ne00*ne03 + i00*ne03 + i03] = src[i01 * ne00 + i00]; - } - } - } - } - } - - // permute source data (src1) from (Sw x Sh x Cin) to (Cin x Sw x Sh) - { - ggml_fp16_t * const wdata = (ggml_fp16_t *) params->wdata + nk; - for (int i12 = 0; i12 < ne12; i12++) { - for (int i11 = 0; i11 < ne11; i11++) { - const float * const src = (float *)((char *) src1->data + i12*nb12 + i11*nb11); - ggml_fp16_t * dst_data = wdata + i11*ne10*ne12; - for (int i10 = 0; i10 < ne10; i10++) { - dst_data[i10*ne12 + i12] = GGML_FP32_TO_FP16(src[i10]); - } - } - } - } - - memset(dst->data, 0, ggml_nbytes(dst)); - } - ggml_barrier(params->threadpool); - - const int32_t stride = ggml_get_op_params_i32(dst, 0); - - // total patches in dst - const int np = ne2; - - // patches per thread - const int dp = (np + nth - 1)/nth; - - // patch range for this thread - const int ip0 = dp*ith; - const int ip1 = MIN(ip0 + dp, np); - - ggml_fp16_t * const wdata = (ggml_fp16_t *) params->wdata + 0; - ggml_fp16_t * const wdata_src = wdata + nk; - - for (int i2 = ip0; i2 < ip1; i2++) { // Cout - float * dst_data = (float *)((char *) dst->data + i2*nb2); - ggml_fp16_t * wdata_kernel = wdata + i2*ne01*ne00*ne03; - for (int i11 = 0; i11 < ne11; i11++) { - for (int i10 = 0; i10 < ne10; i10++) { - const int i1n = i11*ne10*ne12 + i10*ne12; - for (int i01 = 0; i01 < ne01; i01++) { - for (int i00 = 0; i00 < ne00; i00++) { - float v = 0; - ggml_vec_dot_f16(ne03, &v, 0, - wdata_src + i1n, 0, - wdata_kernel + i01*ne00*ne03 + i00*ne03, 0, 1); - dst_data[(i11*stride + i01)*ne0 + i10*stride + i00] += v; - } - } - } - } - } -} - -// ggml_compute_forward_pool_1d_sk_p0 - -static void ggml_compute_forward_pool_1d_sk_p0( - const struct ggml_compute_params * params, - const enum ggml_op_pool op, - const int k, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src = dst->src[0]; - - assert(src->type == GGML_TYPE_F32 || src->type == GGML_TYPE_F16); - - if (params->ith != 0) { - return; - } - - const char * cdata = (const char *)src->data; - const char * const data_end = cdata + ggml_nbytes(src); - float * drow = (float *)dst->data; - - const int64_t rs = dst->ne[0]; - - while (cdata < data_end) { - const void * srow = (const void *)cdata; - int j = 0; - for (int64_t i = 0; i < rs; ++i) { - switch (op) { - case GGML_OP_POOL_AVG: drow[i] = 0; break; - case GGML_OP_POOL_MAX: drow[i] = -FLT_MAX; break; - case GGML_OP_POOL_COUNT: GGML_ABORT("fatal error"); - } - for (int ki = 0; ki < k; ++ki) { - const float srow_j = (src->type == GGML_TYPE_F32) ? ((const float*)srow)[j] : GGML_FP16_TO_FP32(((const ggml_fp16_t*)srow)[j]); - switch (op) { - case GGML_OP_POOL_AVG: drow[i] += srow_j; break; - case GGML_OP_POOL_MAX: if (srow_j > drow[i]) drow[i] = srow_j; break; - case GGML_OP_POOL_COUNT: GGML_ABORT("fatal error"); - } - ++j; - } - switch (op) { - case GGML_OP_POOL_AVG: drow[i] /= k; break; - case GGML_OP_POOL_MAX: break; - case GGML_OP_POOL_COUNT: GGML_ABORT("fatal error"); - } - } - - cdata += src->nb[1]; - drow += rs; - } -} - -// ggml_compute_forward_pool_1d - -static void ggml_compute_forward_pool_1d( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const int32_t * opts = (const int32_t *)dst->op_params; - enum ggml_op_pool op = opts[0]; - const int k0 = opts[1]; - const int s0 = opts[2]; - const int p0 = opts[3]; - GGML_ASSERT(p0 == 0); // padding not supported - GGML_ASSERT(k0 == s0); // only s = k supported - - ggml_compute_forward_pool_1d_sk_p0(params, op, k0, dst); -} - -// ggml_compute_forward_pool_2d - -static void ggml_compute_forward_pool_2d( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src = dst->src[0]; - - assert(src->type == GGML_TYPE_F32 || src->type == GGML_TYPE_F16); - - if (params->ith != 0) { - return; - } - - const int32_t * opts = (const int32_t *)dst->op_params; - enum ggml_op_pool op = opts[0]; - const int k0 = opts[1]; - const int k1 = opts[2]; - const int s0 = opts[3]; - const int s1 = opts[4]; - const int p0 = opts[5]; - const int p1 = opts[6]; - const char * cdata = (const char*)src->data; - const char * const data_end = cdata + ggml_nbytes(src); - - const int64_t px = dst->ne[0]; - const int64_t py = dst->ne[1]; - const int64_t pa = px * py; - - float * dplane = (float *)dst->data; - - const int ka = k0 * k1; - const int offset0 = -p0; - const int offset1 = -p1; - - while (cdata < data_end) { - for (int oy = 0; oy < py; ++oy) { - float * const drow = dplane + oy * px; - for (int ox = 0; ox < px; ++ox) { - float * const out = drow + ox; - switch (op) { - case GGML_OP_POOL_AVG: *out = 0; break; - case GGML_OP_POOL_MAX: *out = -FLT_MAX; break; - case GGML_OP_POOL_COUNT: GGML_ABORT("fatal error"); - } - - const int ix = offset0 + ox * s0; - const int iy = offset1 + oy * s1; - - for (int ky = 0; ky < k1; ++ky) { - if (iy + ky < 0 || iy + ky >= src->ne[1]) continue; - const void * srow = (const void *)(cdata + src->nb[1] * (iy + ky)); - for (int kx = 0; kx < k0; ++kx) { - int j = ix + kx; - if (j < 0 || j >= src->ne[0]) continue; - const float srow_j = (src->type == GGML_TYPE_F32) ? ((const float*)srow)[j] : GGML_FP16_TO_FP32(((const ggml_fp16_t*)srow)[j]); - switch (op) { - case GGML_OP_POOL_AVG: *out += srow_j; break; - case GGML_OP_POOL_MAX: if (srow_j > *out) *out = srow_j; break; - case GGML_OP_POOL_COUNT: GGML_ABORT("fatal error"); - } - } - } - switch (op) { - case GGML_OP_POOL_AVG: *out /= ka; break; - case GGML_OP_POOL_MAX: break; - case GGML_OP_POOL_COUNT: GGML_ABORT("fatal error"); - } - } - } - - cdata += src->nb[2]; - dplane += pa; - } -} - -// ggml_compute_forward_pool_2d_back - -static void ggml_compute_forward_pool_2d_back( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src = dst->src[0]; - const struct ggml_tensor * dstf = dst->src[1]; // forward tensor of dst - - assert(dst->type == GGML_TYPE_F32 || dst->type == GGML_TYPE_F16); - - if (params->ith != 0) { - return; - } - - const int32_t * opts = (const int32_t *)dst->op_params; - enum ggml_op_pool op = opts[0]; - const int k0 = opts[1]; - const int k1 = opts[2]; - const int s0 = opts[3]; - const int s1 = opts[4]; - const int p0 = opts[5]; - const int p1 = opts[6]; - - char * cdata = (char *) dst->data; - const char * cdataf = (const char *) dstf->data; - const char * const data_end = cdata + ggml_nbytes(dst); - - GGML_ASSERT(params->ith == 0); - memset(cdata, 0, ggml_nbytes(dst)); - - const int64_t px = src->ne[0]; - const int64_t py = src->ne[1]; - const int64_t pa = px * py; - - const float * splane = (const float *) src->data; - - const int ka = k0 * k1; - const int offset0 = -p0; - const int offset1 = -p1; - - while (cdata < data_end) { - for (int oy = 0; oy < py; ++oy) { - const float * const srow = splane + oy * px; - for (int ox = 0; ox < px; ++ox) { - const float grad0 = srow[ox]; - - const int ix = offset0 + ox * s0; - const int iy = offset1 + oy * s1; - - if (op == GGML_OP_POOL_MAX) { - float maxval = -FLT_MAX; - int kxmax = -1; - int kymax = -1; - - for (int ky = 0; ky < k1; ++ky) { - if (iy + ky < 0 || iy + ky >= dst->ne[1]) { - continue; - } - const void * drowf = (const void *)(cdataf + dst->nb[1] * (iy + ky)); - for (int kx = 0; kx < k0; ++kx) { - int j = ix + kx; - if (j < 0 || j >= dst->ne[0]) { - continue; - } - - const float val = dst->type == GGML_TYPE_F32 ? - ((const float *) drowf)[j] : GGML_FP16_TO_FP32(((const ggml_fp16_t *) drowf)[j]); - if (val <= maxval) { - continue; - } - - maxval = val; - kxmax = kx; - kymax = ky; - } - } - - if (kxmax == -1 || kymax == -1) { - continue; - } - - void * drow = (void *)(cdata + dst->nb[1] * (iy + kymax)); - const int j = ix + kxmax; - if (dst->type == GGML_TYPE_F32) { - ((float *) drow)[j] += grad0; - } else { - ((ggml_fp16_t *) drow)[j] = GGML_FP32_TO_FP16(grad0 + GGML_FP16_TO_FP32(((const ggml_fp16_t *) drow)[j])); - } - } else if (op == GGML_OP_POOL_AVG) { - const float grad = grad0 / ka; - - for (int ky = 0; ky < k1; ++ky) { - if (iy + ky < 0 || iy + ky >= dst->ne[1]) { - continue; - } - void * drow = (void *)(cdata + dst->nb[1] * (iy + ky)); - for (int kx = 0; kx < k0; ++kx) { - int j = ix + kx; - if (j < 0 || j >= dst->ne[0]) { - continue; - } - - if (dst->type == GGML_TYPE_F32) { - ((float *) drow)[j] += grad; - } else { - ((ggml_fp16_t *) drow)[j] += GGML_FP32_TO_FP16(grad); - } - } - } - } else { - GGML_ASSERT(false); - } - } - } - - cdata += dst->nb[2]; - cdataf += dst->nb[2]; - splane += pa; - } -} - -// ggml_compute_forward_upscale - -static void ggml_compute_forward_upscale_f32( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - GGML_ASSERT(src0->type == GGML_TYPE_F32); - - const int ith = params->ith; - const int nth = params->nth; - - GGML_TENSOR_UNARY_OP_LOCALS - - const float sf0 = (float)ne0/src0->ne[0]; - const float sf1 = (float)ne1/src0->ne[1]; - const float sf2 = (float)ne2/src0->ne[2]; - const float sf3 = (float)ne3/src0->ne[3]; - - // TODO: optimize - - for (int64_t i3 = 0; i3 < ne3; i3++) { - const int64_t i03 = i3 / sf3; - for (int64_t i2 = ith; i2 < ne2; i2 += nth) { - const int64_t i02 = i2 / sf2; - for (int64_t i1 = 0; i1 < ne1; i1++) { - const int64_t i01 = i1 / sf1; - for (int64_t i0 = 0; i0 < ne0; i0++) { - const int64_t i00 = i0 / sf0; - - const float * x = (float *)((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03); - float * y = (float *)((char *) dst->data + i0*nb0 + i1*nb1 + i2*nb2 + i3*nb3); - - *y = *x; - } - } - } - } -} - -static void ggml_compute_forward_upscale( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - switch (src0->type) { - case GGML_TYPE_F32: - { - ggml_compute_forward_upscale_f32(params, dst); - } break; - default: - { - GGML_ABORT("fatal error"); - } - } -} - - -// ggml_compute_forward_pad - -static void ggml_compute_forward_pad_f32( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - GGML_ASSERT(src0->nb[0] == sizeof(float)); - GGML_ASSERT( dst->nb[0] == sizeof(float)); - - const int ith = params->ith; - const int nth = params->nth; - - GGML_TENSOR_UNARY_OP_LOCALS - - float * dst_ptr = (float *) dst->data; - - // TODO: optimize - - for (int64_t i2 = 0; i2 < ne2; ++i2) { - for (int64_t i1 = ith; i1 < ne1; i1 += nth) { - for (int64_t i0 = 0; i0 < ne0; ++i0) { - for (int64_t i3 = 0; i3 < ne3; ++i3) { - const int64_t dst_idx = i3*(ne0*ne1*ne2) + i2*(ne0*ne1) + i1*ne0 + i0; - - const float * src_ptr = (const float *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00); - - if (i0 < ne00 && i1 < ne01 && i2 < ne02 && i3 < ne03) { - dst_ptr[dst_idx] = *src_ptr; - } else { - dst_ptr[dst_idx] = 0; - } - } - } - } - } -} - -static void ggml_compute_forward_pad( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - switch (src0->type) { - case GGML_TYPE_F32: - { - ggml_compute_forward_pad_f32(params, dst); - } break; - default: - { - GGML_ABORT("fatal error"); - } - } -} - -// ggml_compute_forward_pad_reflect_1d - -static void ggml_compute_forward_pad_reflect_1d( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - GGML_ASSERT(src0->type == GGML_TYPE_F32); - GGML_ASSERT( dst->type == GGML_TYPE_F32); - - const int ith = params->ith; - const int nth = params->nth; - - const int32_t * opts = (const int32_t *) dst->op_params; - const int p0 = opts[0]; - const int p1 = opts[1]; - - GGML_TENSOR_UNARY_OP_LOCALS - - for (int64_t i3 = 0; i3 < ne3; i3++) { - for (int64_t i2 = 0; i2 < ne2; i2++) { - for (int64_t i1 = ith; i1 < ne1; i1 += nth) { - float * left = (float *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + p0*nb0); - float * right = (float *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + (ne0-p1-1)*nb0); - - ggml_vec_cpy_f32(ne00, left, (float *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01)); - - for (int i0 = 1; i0 <= p0; i0++) { left[-i0] = left[i0]; } - for (int i0 = 1; i0 <= p1; i0++) { right[i0] = right[-i0]; } - } - } - } -} - -// ggml_compute_forward_arange - -static void ggml_compute_forward_arange_f32( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - GGML_ASSERT(dst->nb[0] == sizeof(float)); - - const int ith = params->ith; - const int nth = params->nth; - - const float start = ggml_get_op_params_f32(dst, 0); - const float stop = ggml_get_op_params_f32(dst, 1); - const float step = ggml_get_op_params_f32(dst, 2); - - const int64_t steps = (int64_t) ceilf((stop - start) / step); - - GGML_ASSERT(ggml_nelements(dst) == steps); - - for (int64_t i = ith; i < steps; i+= nth) { - float value = start + step * i; - ((float *)dst->data)[i] = value; - } -} - -static void ggml_compute_forward_arange( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - switch (dst->type) { - case GGML_TYPE_F32: - { - ggml_compute_forward_arange_f32(params, dst); - } break; - default: - { - GGML_ABORT("fatal error"); - } - } -} - -static void ggml_compute_forward_timestep_embedding_f32( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - GGML_ASSERT(src0->nb[0] == sizeof(float)); - - const int ith = params->ith; - const int nth = params->nth; - - GGML_TENSOR_UNARY_OP_LOCALS - - const int dim = ggml_get_op_params_i32(dst, 0); - const int max_period = ggml_get_op_params_i32(dst, 1); - - int half = dim / 2; - - for (int64_t i = 0; i < ne00; i++) { - float * embed_data = (float *)((char *) dst->data + i*nb1); - for (int64_t j = ith; j < half; j += nth) { - float timestep = ((float *)src0->data)[i]; - float freq = (float)expf(-logf(max_period) * j / half); - float arg = timestep * freq; - embed_data[j] = cosf(arg); - embed_data[j + half] = sinf(arg); - } - if (dim % 2 != 0 && ith == 0) { - embed_data[dim] = 0.f; - } - } -} - -static void ggml_compute_forward_timestep_embedding( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - switch (src0->type) { - case GGML_TYPE_F32: - { - ggml_compute_forward_timestep_embedding_f32(params, dst); - } break; - default: - { - GGML_ABORT("fatal error"); - } - } -} - -// ggml_compute_forward_argsort - -static void ggml_compute_forward_argsort_f32( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - GGML_TENSOR_UNARY_OP_LOCALS - - GGML_ASSERT(nb0 == sizeof(float)); - - const int ith = params->ith; - const int nth = params->nth; - - const int64_t nr = ggml_nrows(src0); - - enum ggml_sort_order order = (enum ggml_sort_order) ggml_get_op_params_i32(dst, 0); - - for (int64_t i = ith; i < nr; i += nth) { - int32_t * dst_data = (int32_t *)((char *) dst->data + i*nb1); - const float * src_data = (float *)((char *) src0->data + i*nb01); - - for (int64_t j = 0; j < ne0; j++) { - dst_data[j] = j; - } - - // C doesn't have a functional sort, so we do a bubble sort instead - for (int64_t j = 0; j < ne0; j++) { - for (int64_t k = j + 1; k < ne0; k++) { - if ((order == GGML_SORT_ORDER_ASC && src_data[dst_data[j]] > src_data[dst_data[k]]) || - (order == GGML_SORT_ORDER_DESC && src_data[dst_data[j]] < src_data[dst_data[k]])) { - int32_t tmp = dst_data[j]; - dst_data[j] = dst_data[k]; - dst_data[k] = tmp; - } - } - } - } -} - -static void ggml_compute_forward_argsort( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - switch (src0->type) { - case GGML_TYPE_F32: - { - ggml_compute_forward_argsort_f32(params, dst); - } break; - default: - { - GGML_ABORT("fatal error"); - } - } -} - -// ggml_compute_forward_flash_attn_ext - -static void ggml_compute_forward_flash_attn_ext_f16( - const struct ggml_compute_params * params, - const struct ggml_tensor * q, - const struct ggml_tensor * k, - const struct ggml_tensor * v, - const struct ggml_tensor * mask, - struct ggml_tensor * dst) { - - GGML_TENSOR_LOCALS(int64_t, neq, q, ne) - GGML_TENSOR_LOCALS(size_t, nbq, q, nb) - GGML_TENSOR_LOCALS(int64_t, nek, k, ne) - GGML_TENSOR_LOCALS(size_t, nbk, k, nb) - GGML_TENSOR_LOCALS(int64_t, nev, v, ne) - GGML_TENSOR_LOCALS(size_t, nbv, v, nb) - GGML_TENSOR_LOCALS(int64_t, ne, dst, ne) - GGML_TENSOR_LOCALS(size_t, nb, dst, nb) - - const int ith = params->ith; - const int nth = params->nth; - - const int64_t D = neq0; - const int64_t N = neq1; - - GGML_ASSERT(ne0 == D); - GGML_ASSERT(ne2 == N); - - // input tensor rows must be contiguous - GGML_ASSERT(nbq0 == ggml_type_size(q->type)); - GGML_ASSERT(nbk0 == ggml_type_size(k->type)); - GGML_ASSERT(nbv0 == ggml_type_size(v->type)); - - GGML_ASSERT(neq0 == D); - GGML_ASSERT(nek0 == D); - GGML_ASSERT(nev0 == D); - - GGML_ASSERT(neq1 == N); - GGML_ASSERT(nev0 == D); - - // dst cannot be transposed or permuted - GGML_ASSERT(nb0 == sizeof(float)); - GGML_ASSERT(nb0 <= nb1); - GGML_ASSERT(nb1 <= nb2); - GGML_ASSERT(nb2 <= nb3); - - // broadcast factors - const int64_t rk2 = neq2/nek2; - const int64_t rk3 = neq3/nek3; - - const int64_t rv2 = neq2/nev2; - const int64_t rv3 = neq3/nev3; - - // parallelize by q rows using ggml_vec_dot_f32 - - // total rows in q - const int nr = neq1*neq2*neq3; - - // rows per thread - const int dr = (nr + nth - 1)/nth; - - // row range for this thread - const int ir0 = dr*ith; - const int ir1 = MIN(ir0 + dr, nr); - - float scale = 1.0f; - float max_bias = 0.0f; - float logit_softcap = 0.0f; - - memcpy(&scale, (float *) dst->op_params + 0, sizeof(float)); - memcpy(&max_bias, (float *) dst->op_params + 1, sizeof(float)); - memcpy(&logit_softcap, (float *) dst->op_params + 2, sizeof(float)); - - if (logit_softcap != 0) { - scale /= logit_softcap; - } - - const uint32_t n_head = neq2; - const uint32_t n_head_log2 = 1u << (uint32_t) floor(log2(n_head)); - - const float m0 = powf(2.0f, -(max_bias ) / n_head_log2); - const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2); - - enum ggml_type const k_vec_dot_type = type_traits_cpu[k->type].vec_dot_type; - ggml_from_float_t const q_to_vec_dot = type_traits_cpu[k_vec_dot_type].from_float; - ggml_vec_dot_t const kq_vec_dot = type_traits_cpu[k->type].vec_dot; - ggml_to_float_t const v_to_float = ggml_get_type_traits(v->type)->to_float; - - GGML_ASSERT(q_to_vec_dot && "fattn: unsupported K-type"); - GGML_ASSERT(v_to_float && "fattn: unsupported V-type"); - - // loop over n_batch and n_head - for (int ir = ir0; ir < ir1; ++ir) { - // q indices - const int iq3 = ir/(neq2*neq1); - const int iq2 = (ir - iq3*neq2*neq1)/neq1; - const int iq1 = (ir - iq3*neq2*neq1 - iq2*neq1); - - const uint32_t h = iq2; // head index - const float slope = (max_bias > 0.0f) ? h < n_head_log2 ? powf(m0, h + 1) : powf(m1, 2*(h - n_head_log2) + 1) : 1.0f; - - float S = 0.0f; // sum - float M = -INFINITY; // maximum KQ value - - float * VKQ32 = (float *) params->wdata + ith*(3*D + CACHE_LINE_SIZE_F32); // FP32 VKQ accumulator - float * V32 = (VKQ32 + 1*D); // (temporary) FP32 V buffer - ggml_fp16_t * VKQ16 = (ggml_fp16_t *) (VKQ32 + 1*D); // (temporary) FP16 VKQ accumulator - ggml_fp16_t * Q_q = (ggml_fp16_t *) (VKQ32 + 2*D); // (temporary) buffer for Q converted to quantized/FP16 - - if (v->type == GGML_TYPE_F16) { - memset(VKQ16, 0, D*sizeof(ggml_fp16_t)); - } else { - memset(VKQ32, 0, D*sizeof(float)); - } - - const ggml_fp16_t * mp = mask ? (ggml_fp16_t *)((char *) mask->data + iq1*mask->nb[1]) : NULL; - - // k indices - const int ik3 = iq3 / rk3; - const int ik2 = iq2 / rk2; - - // v indices - const int iv3 = iq3 / rv3; - const int iv2 = iq2 / rv2; - - const float * pq = (const float *) ((char *) q->data + (iq1*nbq1 + iq2*nbq2 + iq3*nbq3)); - q_to_vec_dot(pq, Q_q, D); - - // online softmax / attention - // loop over n_kv and n_head_kv - // ref: https://arxiv.org/pdf/2112.05682.pdf - for (int64_t ic = 0; ic < nek1; ++ic) { - const float mv = mp ? slope*GGML_FP16_TO_FP32(mp[ic]) : 0.0f; - if (mv == -INFINITY) { - continue; - } - - float s; // KQ value - - const char * k_data = (const char *) k->data + ( ic*nbk1 + ik2*nbk2 + ik3*nbk3); - kq_vec_dot(D, &s, 0, k_data, 0, Q_q, 0, 1); - - s = s*scale; // scale KQ value - - if (logit_softcap != 0.0f) { - s = logit_softcap*tanhf(s); - } - - s += mv; // apply mask - - const float Mold = M; - - float ms = 1.0f; // upon new higher max val, scale VKQ and KQ sum with this value - float vs = 1.0f; // post-softmax KQ value, expf(s - M) - - const char * v_data = ((const char *) v->data + (ic*nbv1 + iv2*nbv2 + iv3*nbv3)); - - if (v->type == GGML_TYPE_F16) { - if (s > M) { - // s is new maximum, ms < 1.0f, vs == expf(s - s) == 1.0f - M = s; - ms = expf(Mold - M); - - // V = V*expf(Mold - M) - ggml_vec_scale_f16(D, VKQ16, ms); - } else { - // no new maximum, ms == 1.0f, vs != 1.0f - vs = expf(s - M); - } - - // V += v*expf(s - M) - ggml_vec_mad_f16(D, VKQ16, (const ggml_fp16_t *) v_data, vs); - } else { - if (s > M) { - // s is new maximum, ms < 1.0f, vs == expf(s - s) == 1.0f - M = s; - ms = expf(Mold - M); - - // V = V*expf(Mold - M) - ggml_vec_scale_f32(D, VKQ32, ms); - } else { - // no new maximum, ms == 1.0f, vs != 1.0f - vs = expf(s - M); - } - - v_to_float(v_data, V32, D); - - // V += v*expf(s - M) - ggml_vec_mad_f32(D, VKQ32, V32, vs); - } - - S = S*ms + vs; // scale and increment sum with partial sum - } - - if (v->type == GGML_TYPE_F16) { - for (int64_t d = 0; d < D; ++d) { - VKQ32[d] = GGML_FP16_TO_FP32(VKQ16[d]); - } - } - - // V /= S - const float S_inv = 1.0f/S; - ggml_vec_scale_f32(D, VKQ32, S_inv); - - // dst indices - const int i1 = iq1; - const int i2 = iq2; - const int i3 = iq3; - - // original - //memcpy((char *) dst->data + (i1*nb1 + i2*nb2 + i3*nb3), V, nev0*sizeof(float)); - - // permute(0, 2, 1, 3) - memcpy((char *) dst->data + (i3*ne2*ne1 + i2 + i1*ne1)*nb1, VKQ32, nb1); - } -} - -static void ggml_compute_forward_flash_attn_ext( - const struct ggml_compute_params * params, - const struct ggml_tensor * q, - const struct ggml_tensor * k, - const struct ggml_tensor * v, - const struct ggml_tensor * mask, - struct ggml_tensor * dst) { - switch (dst->op_params[3]) { - case GGML_PREC_DEFAULT: - case GGML_PREC_F32: - { - // uses F32 accumulators - ggml_compute_forward_flash_attn_ext_f16(params, q, k, v, mask, dst); - } break; - default: - { - GGML_ABORT("fatal error"); - } - } -} - -// ggml_compute_forward_flash_attn_back - -static void ggml_compute_forward_flash_attn_back_f32( - const struct ggml_compute_params * params, - const bool masked, - struct ggml_tensor * dst) { - - const struct ggml_tensor * q = dst->src[0]; - const struct ggml_tensor * k = dst->src[1]; - const struct ggml_tensor * v = dst->src[2]; - const struct ggml_tensor * d = dst->src[3]; - - GGML_TENSOR_LOCALS(int64_t, neq, q, ne) - GGML_TENSOR_LOCALS(size_t, nbq, q, nb) - GGML_TENSOR_LOCALS(int64_t, nek, k, ne) - GGML_TENSOR_LOCALS(size_t, nbk, k, nb) - GGML_TENSOR_LOCALS(int64_t, nev, v, ne) - GGML_TENSOR_LOCALS(size_t, nbv, v, nb) - GGML_TENSOR_LOCALS(int64_t, ned, d, ne) - GGML_TENSOR_LOCALS(size_t, nbd, d, nb) - GGML_TENSOR_LOCALS(int64_t, ne, dst, ne) - GGML_TENSOR_LOCALS(size_t, nb, dst, nb) - - const int ith = params->ith; - const int nth = params->nth; - - const int64_t D = neq0; - const int64_t N = neq1; - const int64_t P = nek1 - N; - const int64_t M = P + N; - - const int Mup = ggml_up(M, GGML_SOFT_MAX_UNROLL); - const int mxDM = MAX(D, Mup); - - // GGML_ASSERT(ne0 == D); - // GGML_ASSERT(ne1 == N); - GGML_ASSERT(P >= 0); - - GGML_ASSERT(nbq0 == sizeof(float)); - GGML_ASSERT(nbk0 == sizeof(float)); - GGML_ASSERT(nbv0 == sizeof(float)); - - GGML_ASSERT(neq0 == D); - GGML_ASSERT(nek0 == D); - GGML_ASSERT(nev1 == D); - GGML_ASSERT(ned0 == D); - - GGML_ASSERT(neq1 == N); - GGML_ASSERT(nek1 == N + P); - GGML_ASSERT(nev1 == D); - GGML_ASSERT(ned1 == N); - - // dst cannot be transposed or permuted - GGML_ASSERT(nb0 == sizeof(float)); - GGML_ASSERT(nb0 <= nb1); - GGML_ASSERT(nb1 <= nb2); - GGML_ASSERT(nb2 <= nb3); - - if (ith == 0) { - memset(dst->data, 0, nb0*ne0*ne1*ne2*ne3); - } - ggml_barrier(params->threadpool); - - const int64_t elem_q = ggml_nelements(q); - const int64_t elem_k = ggml_nelements(k); - - enum ggml_type result_type = dst->type; - GGML_ASSERT(ggml_blck_size(result_type) == 1); - const size_t tsize = ggml_type_size(result_type); - - const size_t offs_q = 0; - const size_t offs_k = offs_q + GGML_PAD(elem_q * tsize, GGML_MEM_ALIGN); - const size_t offs_v = offs_k + GGML_PAD(elem_k * tsize, GGML_MEM_ALIGN); - - void * grad_q = (char *) dst->data; - void * grad_k = (char *) dst->data + offs_k; - void * grad_v = (char *) dst->data + offs_v; - - const size_t nbgq1 = nb0*neq0; - const size_t nbgq2 = nb0*neq0*neq1; - const size_t nbgq3 = nb0*neq0*neq1*neq2; - - const size_t nbgk1 = nb0*nek0; - const size_t nbgk2 = nb0*nek0*nek1; - const size_t nbgk3 = nb0*nek0*nek1*neq2; - - const size_t nbgv1 = nb0*nev0; - const size_t nbgv2 = nb0*nev0*nev1; - const size_t nbgv3 = nb0*nev0*nev1*neq2; - - // parallelize by k rows using ggml_vec_dot_f32 - - // total rows in k - const int nr = nek2*nek3; - - // rows per thread - const int dr = (nr + nth - 1)/nth; - - // row range for this thread - const int ir0 = dr*ith; - const int ir1 = MIN(ir0 + dr, nr); - - const float scale = 1.0f/sqrtf(D); - - //printf("P=%d N=%d D=%d ir0=%d ir1=%d scale = %f\n", P, N, D, ir0, ir1, scale); - - // how often k2 (and v2) is repeated in q2 - int nrep = neq2/nek2; - - for (int ir = ir0; ir < ir1; ++ir) { - // q indices - const int ik3 = ir/(nek2); - const int ik2 = ir - ik3*nek2; - - const int iq3 = ik3; - const int id3 = ik3; - const int iv3 = ik3; - const int iv2 = ik2; - - for (int irep = 0; irep < nrep; ++irep) { - const int iq2 = ik2 + irep*nek2; - const int id2 = iq2; - - // (ik2 + irep*nek2) % nek2 == ik2 - for (int iq1 = 0; iq1 < neq1; ++iq1) { - const int id1 = iq1; - - // not sure about CACHE_LINE_SIZE_F32.. - // - maybe it must not be multiplied by 2 and excluded from .. in SM 1*(..) offset? - float * S = (float *) params->wdata + ith*2*(mxDM + CACHE_LINE_SIZE_F32) + 0*(mxDM+CACHE_LINE_SIZE_F32); - float * SM = (float *) params->wdata + ith*2*(mxDM + CACHE_LINE_SIZE_F32) + 1*(mxDM+CACHE_LINE_SIZE_F32); - - for (int i = M; i < Mup; ++i) { - S[i] = -INFINITY; - } - - const int64_t masked_begin = masked ? (P + iq1 + 1) : M; - for (int64_t ic = 0; ic < masked_begin; ++ic) { - // k indices - const int ik1 = ic; - - // S indices - const int i1 = ik1; - - ggml_vec_dot_f32(neq0, - S + i1, 0, - (float *) ((char *) k->data + (ik1*nbk1 + ik2*nbk2 + ik3*nbk3)), 0, - (float *) ((char *) q->data + (iq1*nbq1 + iq2*nbq2 + iq3*nbq3)), 0, 1); - } - - // scale - ggml_vec_scale_f32(masked_begin, S, scale); - - for (int64_t i = masked_begin; i < M; i++) { - S[i] = -INFINITY; - } - - // softmax - // exclude known -INF S[..] values from max and loop - // dont forget to set their SM values to zero - { - float max = -INFINITY; - ggml_vec_max_f32(masked_begin, &max, S); - - ggml_float sum = 0.0; - { -#ifdef GGML_SOFT_MAX_ACCELERATE - max = -max; - vDSP_vsadd(SM, 1, &max, SM, 1, Mup); - vvexpf(SM, SM, &Mup); - ggml_vec_sum_f32(Mup, &sum, SM); -#else - sum = ggml_vec_soft_max_f32(Mup, SM, S, max); -#endif - } - - assert(sum > 0.0); - - sum = 1.0/sum; - ggml_vec_scale_f32(masked_begin, SM, sum); - - } - - // step-by-step explanation - { - // forward-process shape grads from backward process - // parallel_for ik2,ik3: - // for irep: - // iq2 = ik2 + irep*nek2 - // k[:D,:M,:,:] [D,M,:,:] grad[k][:D,:M,ik2,ik3] += grad[kcur] - // q[:D,:N,:,:] [D,N,:,:] grad[q][:D,iq1,iq2,iq3] += grad[qcur] - // v[:M,:D,:,:] [M,D,:,:] grad[v][:M,:D,iv2,iv3] += grad[vcur] - // for iq1: - // kcur = k[:D,:M,ik2,ik3] [D,M,1,1] grad[kcur] = grad[S1].T @ qcur - // qcur = q[:D,iq1,iq2,iq3] [D,1,1,1] grad[qcur] = grad[S1] @ kcur - // vcur = v[:M,:D,iv2,iv3] [M,D,1,1] grad[vcur] = grad[S5].T @ S4 - // S0 = -Inf [D,1,1,1] - // ~S1[i] = dot(kcur[:D,i], qcur) - // S1 = qcur @ kcur.T [M,1,1,1] grad[S1] = grad[S2] * scale - // S2 = S1 * scale [M,1,1,1] grad[S2] = diag_mask_zero(grad[S3], P) - // S3 = diag_mask_inf(S2, P) [M,1,1,1] grad[S3] = S4 * (grad[S4] - dot(S4, grad[S4])) - // S4 = softmax(S3) [M,1,1,1] grad[S4] = grad[S5] @ vcur - // ~S5[i] = dot(vcur[:,i], S4) - // S5 = S4 @ vcur.T [D,1,1,1] grad[S5] = d[:D,id1,id2,id3] - // ~dst[i,iq1,iq2,iq3] = S5[i] ^ - // dst[:D,iq1,iq2,iq3] = S5 | grad[dst[:D,iq1,iq2,iq3]] = d[:D,id1,id2,id3] - // dst backward-/ grad[dst] = d - // - // output gradients with their dependencies: - // - // grad[kcur] = grad[S1].T @ qcur - // grad[S1] = diag_mask_zero(grad[S3], P) * scale - // grad[S3] = S4 * (grad[S4] - dot(S4, grad[S4])) - // grad[S4] = grad[S5] @ vcur - // grad[S4] = d[:D,id1,id2,id3] @ vcur - // grad[qcur] = grad[S1] @ kcur - // grad[vcur] = grad[S5].T @ S4 - // grad[vcur] = d[:D,id1,id2,id3].T @ S4 - // - // in post-order: - // - // S1 = qcur @ kcur.T - // S2 = S1 * scale - // S3 = diag_mask_inf(S2, P) - // S4 = softmax(S3) - // grad[S4] = d[:D,id1,id2,id3] @ vcur - // grad[S3] = S4 * (grad[S4] - dot(S4, grad[S4])) - // grad[S1] = diag_mask_zero(grad[S3], P) * scale - // grad[qcur] = grad[S1] @ kcur - // grad[kcur] = grad[S1].T @ qcur - // grad[vcur] = d[:D,id1,id2,id3].T @ S4 - // - // using less variables (SM=S4): - // - // S = diag_mask_inf(qcur @ kcur.T * scale, P) - // SM = softmax(S) - // S = d[:D,iq1,iq2,iq3] @ vcur - // dot_SM_gradSM = dot(SM, S) - // S = SM * (S - dot(SM, S)) - // S = diag_mask_zero(S, P) * scale - // - // grad[q][:D,iq1,iq2,iq3] += S @ kcur - // grad[k][:D,:M,ik2,ik3] += S.T @ qcur - // grad[v][:M,:D,iv2,iv3] += d[:D,id1,id2,id3].T @ SM - } - - // S = gradSM = d[:D,id1,id2,id3] @ vcur[:,:,iv2,iv3] - // S = d[:D,id1,id2,id3] @ vcur[:,:,iv2,iv3] - // for ic: - // S[:M] += vcur[:M,ic,iv2,iv3] * d[ic,id1,id2,id3] - // exclude known future zero S[..] values from operation - ggml_vec_set_f32(masked_begin, S, 0); - for (int64_t ic = 0; ic < D; ++ic) { - ggml_vec_mad_f32(masked_begin, - S, - (float *) ((char *) v->data + ( ic*nbv1 + iv2*nbv2 + iv3*nbv3)), - *(float *) ((char *) d->data + (ic*nbd0 + id1*nbd1 + id2*nbd2 + id3*nbd3))); - } - - // S = SM * (S - dot(SM, S)) - float dot_SM_gradSM = 0; - ggml_vec_dot_f32 (masked_begin, &dot_SM_gradSM, 0, SM, 0, S, 0, 1); - ggml_vec_acc1_f32(M, S, -dot_SM_gradSM); - ggml_vec_mul_f32 (masked_begin, S, S, SM); - - // S = diag_mask_zero(S, P) * scale - // already done by above ggml_vec_set_f32 - - // exclude known zero S[..] values from operation - ggml_vec_scale_f32(masked_begin, S, scale); - - // S shape [M,1] - // SM shape [M,1] - // kcur shape [D,M] - // qcur shape [D,1] - // vcur shape [M,D] - - // grad[q][:D,iq1,iq2,iq3] += S @ kcur - // grad[q][:D,iq1,iq2,iq3] += shape[M,1] @ shape[D,M] - // for ic: - // grad[q][:D,iq1,iq2,iq3] += S[ic] * kcur[:D,ic,ik2,ik3] - // exclude known zero S[..] values from loop - for (int64_t ic = 0; ic < masked_begin; ++ic) { - ggml_vec_mad_f32(D, - (float *) ((char *) grad_q + (iq1*nbgq1 + iq2*nbgq2 + iq3*nbgq3)), - (float *) ((char *) k->data + (ic*nbk1 + ik2*nbk2 + ik3*nbk3)), - S[ic]); - } - - // grad[k][:D,:M,iq2,iq3] += S.T @ qcur - // for ic: - // grad[k][:D,ic,iq2,iq3] += S.T[0,ic] * qcur[:D,0] - // grad[k][:D,ic,iq2,iq3] += S[ic] * qcur[:D,0] - // exclude known zero S[..] values from loop - for (int64_t ic = 0; ic < masked_begin; ++ic) { - ggml_vec_mad_f32(D, - (float *) ((char *) grad_k + (ic*nbgk1 + ik2*nbgk2 + ik3*nbgk3)), - (float *) ((char *) q->data + (iq1*nbq1 + iq2*nbq2 + iq3*nbq3)), - S[ic]); - } - - // grad[v][:M,:D,iv2,iv3] += d[:D,id1,id2,id3].T @ SM - // for ic: - // grad[v][:M,ic,iv2,iv3] += d[:D,id1,id2,id3].T[0,ic] * SM[:M] - // grad[v][:M,ic,iv2,iv3] += d[ic,id1,id2,id3] * SM[:M] - // exclude known zero SM[..] values from mad - for (int64_t ic = 0; ic < D; ++ic) { - ggml_vec_mad_f32(masked_begin, - (float *) ((char *) grad_v + ( ic*nbgv1 + iv2*nbgv2 + iv3*nbgv3)), - SM, - *(float *) ((char *) d->data + (ic*nbd0 + id1*nbd1 + id2*nbd2 + id3*nbd3))); - } - } - } - } -} - -static void ggml_compute_forward_flash_attn_back( - const struct ggml_compute_params * params, - const bool masked, - struct ggml_tensor * dst) { - - const struct ggml_tensor * q = dst->src[0]; - - switch (q->type) { - case GGML_TYPE_F32: - { - ggml_compute_forward_flash_attn_back_f32(params, masked, dst); - } break; - default: - { - GGML_ABORT("fatal error"); - } - } -} - -// ggml_compute_forward_ssm_conv - -static void ggml_compute_forward_ssm_conv_f32( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - const struct ggml_tensor * src0 = dst->src[0]; // conv_x - const struct ggml_tensor * src1 = dst->src[1]; // conv1d.weight - - const int ith = params->ith; - const int nth = params->nth; - - const int nc = src1->ne[0]; // d_conv - const int ncs = src0->ne[0]; // d_conv - 1 + n_t - const int nr = src0->ne[1]; // d_inner - const int n_t = dst->ne[1]; // tokens per sequence - const int n_s = dst->ne[2]; // number of sequences in the batch - - GGML_ASSERT( dst->ne[0] == nr); - GGML_ASSERT(src0->nb[0] == sizeof(float)); - GGML_ASSERT(src1->nb[0] == sizeof(float)); - GGML_ASSERT(src0->nb[1] == src0->ne[0]*sizeof(float)); - - // rows per thread - const int dr = (nr + nth - 1)/nth; - - // row range for this thread - const int ir0 = dr*ith; - const int ir1 = MIN(ir0 + dr, nr); - const int ir = ir1 - ir0; - - for (int i3 = 0; i3 < n_s; ++i3) { - for (int i2 = 0; i2 < n_t; ++i2) { - // {d_conv - 1 + n_t, d_inner, n_seqs} - // sliding window - const float * s = (const float *) ((const char *) src0->data + ir0*(src0->nb[1]) + i2*(src0->nb[0]) + i3*(src0->nb[2])); // {d_conv, d_inner, n_s} - const float * c = (const float *) ((const char *) src1->data + ir0*(src1->nb[1])); // {d_conv, d_inner} - float * x = (float *) ((char *) dst->data + ir0*(dst->nb[0]) + i2*(dst->nb[1]) + i3*(dst->nb[2])); // {d_inner, n_t, n_s} - - // TODO: transpose the output for smaller strides for big batches? - // d_inner - for (int i1 = 0; i1 < ir; ++i1) { - // rowwise dot product - // NOTE: not using ggml_vec_dot_f32, because its sum is in double precision - float sumf = 0.0f; - - // d_conv - for (int i0 = 0; i0 < nc; ++i0) { - sumf += s[i0 + i1*ncs] * c[i0 + i1*nc]; - } - x[i1] = sumf; - } - } - } -} - -static void ggml_compute_forward_ssm_conv( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - switch (dst->src[0]->type) { - case GGML_TYPE_F32: - { - ggml_compute_forward_ssm_conv_f32(params, dst); - } break; - default: - { - GGML_ABORT("fatal error"); - } - } -} - -// ggml_compute_forward_ssm_scan - -static void ggml_compute_forward_ssm_scan_f32( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - const struct ggml_tensor * src0 = dst->src[0]; // s - const struct ggml_tensor * src1 = dst->src[1]; // x - const struct ggml_tensor * src2 = dst->src[2]; // dt - const struct ggml_tensor * src3 = dst->src[3]; // A - const struct ggml_tensor * src4 = dst->src[4]; // B - const struct ggml_tensor * src5 = dst->src[5]; // C - - const int ith = params->ith; - const int nth = params->nth; - - const int64_t nc = src0->ne[0]; // d_state - const int64_t nr = src0->ne[1]; // d_inner - const int64_t n_t = src1->ne[1]; // number of tokens per sequence - const int64_t n_s = src0->ne[2]; // number of sequences in the batch - - GGML_ASSERT(ggml_nelements(src1) + ggml_nelements(src0) == ggml_nelements(dst)); - GGML_ASSERT(src0->nb[0] == sizeof(float)); - GGML_ASSERT(src1->nb[0] == sizeof(float)); - GGML_ASSERT(src2->nb[0] == sizeof(float)); - GGML_ASSERT(src3->nb[0] == sizeof(float)); - GGML_ASSERT(src4->nb[0] == sizeof(float)); - GGML_ASSERT(src5->nb[0] == sizeof(float)); - // required for the dot product between s and C - GGML_ASSERT(src0->nb[1] == src0->ne[0]*sizeof(float)); - // required for per-sequence offsets for states - GGML_ASSERT(src0->nb[2] == src0->ne[0]*src0->ne[1]*sizeof(float)); - // required to get correct offset for state destination (i.e. src1->nb[3]) - GGML_ASSERT(src1->nb[3] == src1->ne[0]*src1->ne[1]*src1->ne[2]*sizeof(float)); - - // rows per thread - const int dr = (nr + nth - 1)/nth; - - // row range for this thread - const int ir0 = dr*ith; - const int ir1 = MIN(ir0 + dr, nr); - const int ir = ir1 - ir0; - - for (int i3 = 0; i3 < n_s; ++i3) { - for (int i2 = 0; i2 < n_t; ++i2) { - const float * s0 = (const float *) ((const char *) src0->data + ir0*(src0->nb[1]) + i3*(src0->nb[2])); // {d_state, d_inner, n_s} - const float * x = (const float *) ((const char *) src1->data + ir0*(src1->nb[0]) + i2*(src1->nb[1]) + i3*(src1->nb[2])); // {d_inner, n_t, n_s} - const float * dt = (const float *) ((const char *) src2->data + ir0*(src2->nb[0]) + i2*(src2->nb[1]) + i3*(src2->nb[2])); // {d_inner, n_t, n_s} - const float * A = (const float *) ((const char *) src3->data + ir0*(src3->nb[1])); // {d_state, d_inner} - const float * B = (const float *) ((const char *) src4->data + i2*(src4->nb[1]) + i3*(src4->nb[2])); // {d_state, n_t, n_s} - const float * C = (const float *) ((const char *) src5->data + i2*(src5->nb[1]) + i3*(src5->nb[2])); // {d_state, n_t, n_s} - float * y = ( float *) (( char *) dst->data + ir0*(src1->nb[0]) + i2*(src1->nb[1]) + i3*(src1->nb[2])); // {d_inner, n_t, n_s} - float * s = ( float *) (( char *) dst->data + ir0*(src0->nb[1]) + i3*(src0->nb[2]) + src1->nb[3]); // {d_state, d_inner, n_s} - - // use the output as the source for the next token-wise iterations - if (i2 > 0) { s0 = s; } - - // d_inner - for (int i1 = 0; i1 < ir; ++i1) { - // ref: https://github.com/state-spaces/mamba/blob/34076d664838588a3c97727b263478ab9f621a07/mamba_ssm/ops/triton/selective_state_update.py#L78 - float dt_soft_plus = dt[i1] <= 20.0f ? log1pf(expf(dt[i1])) : dt[i1]; - float x_dt = x[i1] * dt_soft_plus; - float sumf = 0.0f; - // d_state - for (int i0 = 0; i0 < nc; ++i0) { - int i = i0 + i1*nc; - // state = prev_state * dA + dB * x - float state = (s0[i] * expf(dt_soft_plus * A[i])) + (B[i0] * x_dt); - // y = rowwise_dotprod(state, C) - sumf += state * C[i0]; - s[i] = state; - } - y[i1] = sumf; - } - } - } -} - -static void ggml_compute_forward_ssm_scan( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - switch (dst->src[0]->type) { - case GGML_TYPE_F32: - { - ggml_compute_forward_ssm_scan_f32(params, dst); - } break; - default: - { - GGML_ABORT("fatal error"); - } - } -} - -// ggml_compute_forward_win_part - -static void ggml_compute_forward_win_part_f32( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - UNUSED(params); - - const struct ggml_tensor * src0 = dst->src[0]; - - GGML_TENSOR_LOCALS(int64_t, ne0, src0, ne) - GGML_TENSOR_LOCALS(int64_t, ne, dst, ne) - - const int32_t nep0 = ((const int32_t *)(dst->op_params))[0]; - const int32_t nep1 = ((const int32_t *)(dst->op_params))[1]; - const int32_t w = ((const int32_t *)(dst->op_params))[2]; - - assert(ne00 == ne0); - assert(ne3 == nep0*nep1); - - // TODO: optimize / multi-thread - for (int py = 0; py < nep1; ++py) { - for (int px = 0; px < nep0; ++px) { - const int64_t i3 = py*nep0 + px; - for (int64_t i2 = 0; i2 < ne2; ++i2) { - for (int64_t i1 = 0; i1 < ne1; ++i1) { - for (int64_t i0 = 0; i0 < ne0; ++i0) { - const int64_t i02 = py*w + i2; - const int64_t i01 = px*w + i1; - const int64_t i00 = i0; - - const int64_t i = i3*ne2*ne1*ne0 + i2*ne1*ne0 + i1*ne0 + i0; - const int64_t j = i02*ne01*ne00 + i01*ne00 + i00; - - if (py*w + i2 >= ne02 || px*w + i1 >= ne01) { - ((float *) dst->data)[i] = 0.0f; - } else { - ((float *) dst->data)[i] = ((float *) src0->data)[j]; - } - } - } - } - } - } -} - -static void ggml_compute_forward_win_part( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - switch (src0->type) { - case GGML_TYPE_F32: - { - ggml_compute_forward_win_part_f32(params, dst); - } break; - default: - { - GGML_ABORT("fatal error"); - } - } -} - -// ggml_compute_forward_win_unpart - -static void ggml_compute_forward_win_unpart_f32( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - UNUSED(params); - - const struct ggml_tensor * src0 = dst->src[0]; - - GGML_TENSOR_LOCALS(int64_t, ne0, src0, ne) - GGML_TENSOR_LOCALS(int64_t, ne, dst, ne) - - const int32_t w = ((const int32_t *)(dst->op_params))[0]; - - // padding - const int px = (w - ne1%w)%w; - //const int py = (w - ne2%w)%w; - - const int npx = (px + ne1)/w; - //const int npy = (py + ne2)/w; - - assert(ne0 == ne00); - - // TODO: optimize / multi-thread - for (int64_t i2 = 0; i2 < ne2; ++i2) { - for (int64_t i1 = 0; i1 < ne1; ++i1) { - for (int64_t i0 = 0; i0 < ne0; ++i0) { - const int ip2 = i2/w; - const int ip1 = i1/w; - - const int64_t i02 = i2%w; - const int64_t i01 = i1%w; - const int64_t i00 = i0; - - const int64_t i = (ip2*npx + ip1)*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00 + i00; - const int64_t j = i2*ne1*ne0 + i1*ne0 + i0; - - ((float *) dst->data)[j] = ((float *) src0->data)[i]; - } - } - } -} - -static void ggml_compute_forward_win_unpart( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - switch (src0->type) { - case GGML_TYPE_F32: - { - ggml_compute_forward_win_unpart_f32(params, dst); - } break; - default: - { - GGML_ABORT("fatal error"); - } - } -} - -//gmml_compute_forward_unary - -static void ggml_compute_forward_unary( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const enum ggml_unary_op op = ggml_get_unary_op(dst); - - switch (op) { - case GGML_UNARY_OP_ABS: - { - ggml_compute_forward_abs(params, dst); - } break; - case GGML_UNARY_OP_SGN: - { - ggml_compute_forward_sgn(params, dst); - } break; - case GGML_UNARY_OP_NEG: - { - ggml_compute_forward_neg(params, dst); - } break; - case GGML_UNARY_OP_STEP: - { - ggml_compute_forward_step(params, dst); - } break; - case GGML_UNARY_OP_TANH: - { - ggml_compute_forward_tanh(params, dst); - } break; - case GGML_UNARY_OP_ELU: - { - ggml_compute_forward_elu(params, dst); - } break; - case GGML_UNARY_OP_RELU: - { - ggml_compute_forward_relu(params, dst); - } break; - case GGML_UNARY_OP_SIGMOID: - { - ggml_compute_forward_sigmoid(params, dst); - } break; - case GGML_UNARY_OP_GELU: - { - ggml_compute_forward_gelu(params, dst); - } break; - case GGML_UNARY_OP_GELU_QUICK: - { - ggml_compute_forward_gelu_quick(params, dst); - } break; - case GGML_UNARY_OP_SILU: - { - ggml_compute_forward_silu(params, dst); - } break; - case GGML_UNARY_OP_HARDSWISH: - { - ggml_compute_forward_hardswish(params, dst); - } break; - case GGML_UNARY_OP_HARDSIGMOID: - { - ggml_compute_forward_hardsigmoid(params, dst); - } break; - case GGML_UNARY_OP_EXP: - { - ggml_compute_forward_exp(params, dst); - } break; - default: - { - GGML_ABORT("fatal error"); - } - } -} - -// ggml_compute_forward_get_rel_pos - -static void ggml_compute_forward_get_rel_pos_f16( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - UNUSED(params); - - const struct ggml_tensor * src0 = dst->src[0]; - - // ref: https://github.com/facebookresearch/segment-anything/blob/main/segment_anything/modeling/image_encoder.py#L292-L322 - - GGML_TENSOR_UNARY_OP_LOCALS - - const int64_t w = ne1; - - ggml_fp16_t * src0_data = (ggml_fp16_t *) src0->data; - ggml_fp16_t * dst_data = (ggml_fp16_t *) dst->data; - - for (int64_t i2 = 0; i2 < ne2; ++i2) { - for (int64_t i1 = 0; i1 < ne1; ++i1) { - const int64_t pos = (w - i1 - 1) + i2; - for (int64_t i0 = 0; i0 < ne0; ++i0) { - dst_data[i2*ne1*ne0 + i1*ne0 + i0] = src0_data[pos*ne00 + i0]; - } - } - } -} - -static void ggml_compute_forward_get_rel_pos( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - switch (src0->type) { - case GGML_TYPE_F16: - case GGML_TYPE_BF16: - { - ggml_compute_forward_get_rel_pos_f16(params, dst); - } break; - default: - { - GGML_ABORT("fatal error"); - } - } -} - -// ggml_compute_forward_add_rel_pos - -static void ggml_compute_forward_add_rel_pos_f32( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - const struct ggml_tensor * src1 = dst->src[1]; - const struct ggml_tensor * src2 = dst->src[2]; - - const bool inplace = (bool) ((int32_t *) dst->op_params)[0]; - if (!inplace) { - if (params->ith == 0) { - memcpy((char *) dst->data, (char *) src0->data, ggml_nbytes(dst)); - } - ggml_barrier(params->threadpool); - } - // ref: https://github.com/facebookresearch/segment-anything/blob/main/segment_anything/modeling/image_encoder.py#L357-L359 - - float * src1_data = (float *) src1->data; - float * src2_data = (float *) src2->data; - float * dst_data = (float *) dst->data; - - const int64_t ne10 = src1->ne[0]; - const int64_t ne11 = src1->ne[1]; - const int64_t ne12 = src1->ne[2]; - const int64_t ne13 = src1->ne[3]; - - const int ith = params->ith; - const int nth = params->nth; - - // total patches in dst - const int np = ne13; - - // patches per thread - const int dp = (np + nth - 1)/nth; - - // patch range for this thread - const int ip0 = dp*ith; - const int ip1 = MIN(ip0 + dp, np); - - for (int64_t i13 = ip0; i13 < ip1; ++i13) { - for (int64_t i12 = 0; i12 < ne12; ++i12) { - for (int64_t i11 = 0; i11 < ne11; ++i11) { - const int64_t jp1 = i13*ne12*ne11*ne10 + i12*ne11*ne10 + i11*ne10; - for (int64_t i10 = 0; i10 < ne10; ++i10) { - const int64_t jp0 = jp1 + i10; - const float src1_e = src1_data[jp0]; - const float src2_e = src2_data[jp0]; - - const int64_t jdh = jp0 * ne10; - const int64_t jdw = jdh - (ne10 - 1) * i10; - - for (int64_t j = 0; j < ne10; ++j) { - dst_data[jdh + j ] += src2_e; - dst_data[jdw + j*ne10] += src1_e; - } - } - } - } - } -} - -static void ggml_compute_forward_add_rel_pos( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - switch (src0->type) { - case GGML_TYPE_F32: - { - ggml_compute_forward_add_rel_pos_f32(params, dst); - } break; - default: - { - GGML_ABORT("fatal error"); - } - } -} - -// ggml_compute_forward_rwkv_wkv6 - -static void ggml_compute_forward_rwkv_wkv6_f32( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - const int64_t T = dst->src[1]->ne[2]; - const int64_t C = dst->ne[0]; - const int64_t HEADS = dst->src[1]->ne[1]; - const int64_t n_seqs = dst->src[5]->ne[1]; - const int64_t head_size = C / HEADS; - - float * dst_data = (float *) dst->data; - float * state = ((float *) dst->data) + C * T; - - const int ith = params->ith; - const int nth = params->nth; - - if (ith >= HEADS) { - return; - } - - const int h_start = (HEADS * ith) / nth; - const int h_end = ((HEADS * (ith + 1)) / nth < HEADS) ? - (HEADS * (ith + 1)) / nth : HEADS; - - float * k = (float *) dst->src[0]->data; - float * v = (float *) dst->src[1]->data; - float * r = (float *) dst->src[2]->data; - float * time_faaaa = (float *) dst->src[3]->data; - float * time_decay = (float *) dst->src[4]->data; - - size_t t_stride = HEADS * head_size; // Same to C - - size_t h_stride = C / HEADS; - GGML_ASSERT(C % HEADS == 0); // C must be divisible by HEADS - size_t h_stride_2d = head_size * head_size; - - if (ith == 0) { - memset(dst_data, 0, T * C * sizeof(float)); - } - ggml_barrier(params->threadpool); - - - #if defined(__AVX__) && !defined(__AVX512F__) - #define GGML_F32X GGML_F32x8 - #define GGML_F32X_SET1 GGML_F32x8_SET1 - #define GGML_F32X_LOAD GGML_F32x8_LOAD - #define GGML_F32X_STORE GGML_F32x8_STORE - #define GGML_F32X_MUL GGML_F32x8_MUL - #define GGML_F32X_FMA GGML_F32x8_FMA - #define WKV_VECTOR_SIZE 8 - #elif defined(__AVX512F__) - #define GGML_F32X GGML_F32x16 - #define GGML_F32X_SET1 GGML_F32x16_SET1 - #define GGML_F32X_LOAD GGML_F32x16_LOAD - #define GGML_F32X_STORE GGML_F32x16_STORE - #define GGML_F32X_MUL GGML_F32x16_MUL - #define GGML_F32X_FMA GGML_F32x16_FMA - #define WKV_VECTOR_SIZE 16 - #elif defined(__ARM_NEON) && defined(__aarch64__) - #define GGML_F32X GGML_F32x4 - #define GGML_F32X_SET1 GGML_F32x4_SET1 - #define GGML_F32X_LOAD GGML_F32x4_LOAD - #define GGML_F32X_STORE GGML_F32x4_STORE - #define GGML_F32X_MUL GGML_F32x4_MUL - #define GGML_F32X_FMA GGML_F32x4_FMA - #define WKV_VECTOR_SIZE 4 - #endif - - #ifdef WKV_VECTOR_SIZE - const int64_t vec_count = head_size / WKV_VECTOR_SIZE; - - for (int64_t t = 0; t < T; t++) { - size_t t_offset = t * t_stride; - size_t state_offset = head_size * C * (t / (T / n_seqs)); - float * state_cur = state + state_offset; - float * state_prev = t % (T / n_seqs) ? state_cur : (float*)dst->src[5]->data + state_offset; - - for (int64_t h = h_start; h < h_end; h++) { - size_t h_offset = h * h_stride; - size_t t_h_offset = t_offset + h_offset; - size_t h_2d_offset = h * h_stride_2d; - - for (int64_t i = 0; i < head_size; i++) { - size_t t_h_i_offset = t_h_offset + i; - size_t h_i_offset = h_offset + i; - size_t h_2d_i_offset = h_2d_offset + i * h_stride; - - float k_val = k[t_h_i_offset]; - float r_val = r[t_h_i_offset]; - float time_faaaa_val = time_faaaa[h_i_offset]; - float time_decay_val = time_decay[t_h_i_offset]; - - // Broadcast scalar values to vectors - GGML_F32X k_vec = GGML_F32X_SET1(k_val); - GGML_F32X r_vec = GGML_F32X_SET1(r_val); - GGML_F32X time_faaaa_vec = GGML_F32X_SET1(time_faaaa_val); - GGML_F32X time_decay_vec = GGML_F32X_SET1(time_decay_val); - - for (int64_t j = 0; j < vec_count; j++) { - size_t base_j = j * WKV_VECTOR_SIZE; - size_t t_h_j_offset = t_h_offset + base_j; - size_t h_2d_i_j_offset = h_2d_i_offset + base_j; - - // Load x elements at once - GGML_F32X v_vec = GGML_F32X_LOAD(&v[t_h_j_offset]); - GGML_F32X prev_state_vec = GGML_F32X_LOAD(&state_prev[h_2d_i_j_offset]); - GGML_F32X dst_vec = GGML_F32X_LOAD(&dst_data[t_h_j_offset]); - - // Compute kv = v * k - GGML_F32X kv_vec = GGML_F32X_MUL(v_vec, k_vec); - - // Compute temp = kv * time_faaaa + prev_state - GGML_F32X temp_vec = GGML_F32X_FMA(prev_state_vec, kv_vec, time_faaaa_vec); - - // Update dst: dst += temp * r - dst_vec = GGML_F32X_FMA(dst_vec, temp_vec, r_vec); - GGML_F32X_STORE(&dst_data[t_h_j_offset], dst_vec); - - // Update state: state = prev_state * time_decay + kv - GGML_F32X new_state_vec = GGML_F32X_FMA(kv_vec, prev_state_vec, time_decay_vec); - GGML_F32X_STORE(&state_cur[h_2d_i_j_offset], new_state_vec); - } - - // Handle remaining elements, this will not be used. - for (int64_t j = vec_count * WKV_VECTOR_SIZE; j < head_size; j++) { - size_t t_h_j_offset = t_h_offset + j; - size_t h_2d_i_j_offset = h_2d_i_offset + j; - float v_val = v[t_h_j_offset]; - float kv_val = v_val * k_val; - float prev_state_val = state_prev[h_2d_i_j_offset]; - float temp_val = kv_val * time_faaaa_val + prev_state_val; - dst_data[t_h_j_offset] += temp_val * r_val; - state_cur[h_2d_i_j_offset] = prev_state_val * time_decay_val + kv_val; - } - } - } - } - - #else - // basically fused operations: - // dst = r @ (time_faaaa * (k @ v) + state), - // state = time_decay * state + (k @ v), - // recursive through each token - for (int64_t t = 0; t < T; t++) { - size_t t_offset = t * t_stride; - size_t state_offset = head_size * C * (t / (T / n_seqs)); - float * state_cur = state + state_offset; - float * state_prev = t % (T / n_seqs) ? state_cur : (float*)dst->src[5]->data + state_offset; - - for (int64_t h = h_start; h < h_end; h++) { - size_t h_offset = h * h_stride; - size_t t_h_offset = t_offset + h_offset; - size_t h_2d_offset = h * h_stride_2d; - - for (int64_t i = 0; i < head_size; i++) { - size_t t_h_i_offset = t_h_offset + i; - size_t h_i_offset = h_offset + i; - size_t h_2d_i_offset = h_2d_offset + i * h_stride; - - float k_val = k[t_h_i_offset]; - float r_val = r[t_h_i_offset]; - float time_faaaa_val = time_faaaa[h_i_offset]; - // RWKV v6: different time_decay for each token. - float time_decay_val = time_decay[t_h_i_offset]; - - for (int64_t j = 0; j < head_size; j++) { - size_t t_h_j_offset = t_h_offset + j; - size_t h_2d_i_j_offset = h_2d_i_offset + j; - - float v_val = v[t_h_j_offset]; - float kv_val = v_val * k_val; - float prev_state_val = state_prev[h_2d_i_j_offset]; - float temp_val = kv_val * time_faaaa_val + prev_state_val; - dst_data[t_h_j_offset] += temp_val * r_val; - state_cur[h_2d_i_j_offset] = prev_state_val * time_decay_val + kv_val; - } - } - } - } - #endif -} - - -static void ggml_compute_forward_rwkv_wkv6( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - switch (src0->type) { - case GGML_TYPE_F32: - { - ggml_compute_forward_rwkv_wkv6_f32(params, dst); - } break; - default: - { - GGML_ABORT("fatal error"); - } - } -} - -// ggml_compute_forward_gla - -static void ggml_compute_forward_gla_f32( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - const int64_t T = dst->src[1]->ne[2]; - const int64_t C = dst->ne[0]; - const int64_t HEADS = dst->src[1]->ne[1]; - const int64_t n_seqs = dst->src[4]->ne[1]; - const int64_t head_size = C / HEADS; - const float scale = ggml_get_op_params_f32(dst, 0); - - float * dst_data = (float *) dst->data; - float * state = ((float *) dst->data) + C * T; - - const int ith = params->ith; - const int nth = params->nth; - - if (ith >= HEADS) { - return; - } - - const int h_start = (HEADS * ith) / nth; - const int h_end = ((HEADS * (ith + 1)) / nth < HEADS) ? - (HEADS * (ith + 1)) / nth : HEADS; - - float * k = (float *) dst->src[0]->data; - float * v = (float *) dst->src[1]->data; - float * q = (float *) dst->src[2]->data; - float * g = (float *) dst->src[3]->data; - - size_t t_stride = HEADS * head_size; // Same to C - - size_t h_stride = C / HEADS; - GGML_ASSERT(C % HEADS == 0); // C must be divisible by HEADS - size_t h_stride_2d = head_size * head_size; - - if (ith == 0) { - memset(dst_data, 0, T * C * sizeof(float)); - } - ggml_barrier(params->threadpool); - - - #if defined(__AVX__) && !defined(__AVX512F__) - #define GGML_F32X GGML_F32x8 - #define GGML_F32X_SET1 GGML_F32x8_SET1 - #define GGML_F32X_LOAD GGML_F32x8_LOAD - #define GGML_F32X_STORE GGML_F32x8_STORE - #define GGML_F32X_MUL GGML_F32x8_MUL - #define GGML_F32X_FMA GGML_F32x8_FMA - #define GLA_VECTOR_SIZE 8 - #elif defined(__AVX512F__) - #define GGML_F32X GGML_F32x16 - #define GGML_F32X_SET1 GGML_F32x16_SET1 - #define GGML_F32X_LOAD GGML_F32x16_LOAD - #define GGML_F32X_STORE GGML_F32x16_STORE - #define GGML_F32X_MUL GGML_F32x16_MUL - #define GGML_F32X_FMA GGML_F32x16_FMA - #define GLA_VECTOR_SIZE 16 - #elif defined(__ARM_NEON) && defined(__aarch64__) - #define GGML_F32X GGML_F32x4 - #define GGML_F32X_SET1 GGML_F32x4_SET1 - #define GGML_F32X_LOAD GGML_F32x4_LOAD - #define GGML_F32X_STORE GGML_F32x4_STORE - #define GGML_F32X_MUL GGML_F32x4_MUL - #define GGML_F32X_FMA GGML_F32x4_FMA - #define GLA_VECTOR_SIZE 4 - #endif - - #ifdef GLA_VECTOR_SIZE - const int64_t vec_count = head_size / GLA_VECTOR_SIZE; - - for (int64_t t = 0; t < T; t++) { - size_t t_offset = t * t_stride; - size_t state_offset = head_size * C * (t / (T / n_seqs)); - float * state_cur = state + state_offset; - float * state_prev = t % (T / n_seqs) ? state_cur : (float*)dst->src[4]->data + state_offset; - - for (int64_t h = h_start; h < h_end; h++) { - size_t h_offset = h * h_stride; - size_t t_h_offset = t_offset + h_offset; - size_t h_2d_offset = h * h_stride_2d; - - for (int64_t i = 0; i < head_size; i++) { - size_t t_h_i_offset = t_h_offset + i; - size_t h_2d_i_offset = h_2d_offset + i * h_stride; - - float k_val = k[t_h_i_offset]; - float q_val = q[t_h_i_offset] * scale; - float g_val = g[t_h_i_offset]; - - // Broadcast scalar values to vectors - GGML_F32X k_vec = GGML_F32X_SET1(k_val); - GGML_F32X q_vec = GGML_F32X_SET1(q_val); - GGML_F32X g_vec = GGML_F32X_SET1(g_val); - - for (int64_t j = 0; j < vec_count; j++) { - size_t base_j = j * GLA_VECTOR_SIZE; - size_t t_h_j_offset = t_h_offset + base_j; - size_t h_2d_i_j_offset = h_2d_i_offset + base_j; - - // Load x elements at once - GGML_F32X v_vec = GGML_F32X_LOAD(&v[t_h_j_offset]); - GGML_F32X prev_state_vec = GGML_F32X_LOAD(&state_prev[h_2d_i_j_offset]); - GGML_F32X dst_vec = GGML_F32X_LOAD(&dst_data[t_h_j_offset]); - - // Compute kv = v * k - GGML_F32X kv_vec = GGML_F32X_MUL(v_vec, k_vec); - - // Compute temp = prev_state * g + kv - GGML_F32X temp_vec = GGML_F32X_FMA(kv_vec, prev_state_vec, g_vec); - - // Update dst: dst += temp * q - dst_vec = GGML_F32X_FMA(dst_vec, temp_vec, q_vec); - GGML_F32X_STORE(&dst_data[t_h_j_offset], dst_vec); - - // Update state - GGML_F32X_STORE(&state_cur[h_2d_i_j_offset], temp_vec); - } - - // Handle remaining elements, this will not be used. - for (int64_t j = vec_count * GLA_VECTOR_SIZE; j < head_size; j++) { - size_t t_h_j_offset = t_h_offset + j; - size_t h_2d_i_j_offset = h_2d_i_offset + j; - float v_val = v[t_h_j_offset]; - float kv_val = v_val * k_val; - float prev_state_val = state_prev[h_2d_i_j_offset]; - float temp_val = kv_val + prev_state_val * g_val; - dst_data[t_h_j_offset] += temp_val * q_val; - state_cur[h_2d_i_j_offset] = temp_val; - } - } - } - } - - #else - for (int64_t t = 0; t < T; t++) { - size_t t_offset = t * t_stride; - size_t state_offset = head_size * C * (t / (T / n_seqs)); - float * state_cur = state + state_offset; - float * state_prev = t % (T / n_seqs) ? state_cur : (float*)dst->src[4]->data + state_offset; - - for (int64_t h = h_start; h < h_end; h++) { - size_t h_offset = h * h_stride; - size_t t_h_offset = t_offset + h_offset; - size_t h_2d_offset = h * h_stride_2d; - - for (int64_t i = 0; i < head_size; i++) { - size_t t_h_i_offset = t_h_offset + i; - size_t h_2d_i_offset = h_2d_offset + i * h_stride; - - float k_val = k[t_h_i_offset]; - float q_val = q[t_h_i_offset] * scale; - float g_val = g[t_h_i_offset]; - - for (int64_t j = 0; j < head_size; j++) { - size_t t_h_j_offset = t_h_offset + j; - size_t h_2d_i_j_offset = h_2d_i_offset + j; - - float v_val = v[t_h_j_offset]; - float kv_val = v_val * k_val; - float prev_state_val = state_prev[h_2d_i_j_offset]; - float temp_val = prev_state_val * g_val + kv_val; - dst_data[t_h_j_offset] += temp_val * q_val; - state_cur[h_2d_i_j_offset] = temp_val; - } - } - } - } - #endif -} - - -static void ggml_compute_forward_gla( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - switch (src0->type) { - case GGML_TYPE_F32: - { - ggml_compute_forward_gla_f32(params, dst); - } break; - default: - { - GGML_ABORT("fatal error"); - } - } -} - -// ggml_compute_forward_map_unary - -static void ggml_compute_forward_map_unary_f32( - const struct ggml_compute_params * params, - struct ggml_tensor * dst, - const ggml_unary_op_f32_t fun) { - - const struct ggml_tensor * src0 = dst->src[0]; - - if (params->ith != 0) { - return; - } - - assert(ggml_is_contiguous_1(src0)); - assert(ggml_is_contiguous_1(dst)); - assert(ggml_are_same_shape(src0, dst)); - - const int n = ggml_nrows(src0); - const int nc = src0->ne[0]; - - for (int i = 0; i < n; i++) { - fun(nc, - (float *) ((char *) dst->data + i*( dst->nb[1])), - (float *) ((char *) src0->data + i*(src0->nb[1]))); - } -} - -static void ggml_compute_forward_map_unary( - const struct ggml_compute_params * params, - struct ggml_tensor * dst, - const ggml_unary_op_f32_t fun) { - - const struct ggml_tensor * src0 = dst->src[0]; - - switch (src0->type) { - case GGML_TYPE_F32: - { - ggml_compute_forward_map_unary_f32(params, dst, fun); - } break; - default: - { - GGML_ABORT("fatal error"); - } - } -} - -// ggml_compute_forward_map_binary - -static void ggml_compute_forward_map_binary_f32( - const struct ggml_compute_params * params, - struct ggml_tensor * dst, - const ggml_binary_op_f32_t fun) { - - const struct ggml_tensor * src0 = dst->src[0]; - const struct ggml_tensor * src1 = dst->src[1]; - - if (params->ith != 0) { - return; - } - - assert(ggml_is_contiguous_1(src0)); - assert(ggml_is_contiguous_1(src1)); - assert(ggml_is_contiguous_1(dst)); - assert(ggml_are_same_shape(src0, src1) && ggml_are_same_shape(src0, dst)); - - const int n = ggml_nrows(src0); - const int nc = src0->ne[0]; - - for (int i = 0; i < n; i++) { - fun(nc, - (float *) ((char *) dst->data + i*( dst->nb[1])), - (float *) ((char *) src0->data + i*(src0->nb[1])), - (float *) ((char *) src1->data + i*(src1->nb[1]))); - } -} - -static void ggml_compute_forward_map_binary( - const struct ggml_compute_params * params, - struct ggml_tensor * dst, - const ggml_binary_op_f32_t fun) { - - const struct ggml_tensor * src0 = dst->src[0]; - - switch (src0->type) { - case GGML_TYPE_F32: - { - ggml_compute_forward_map_binary_f32(params, dst, fun); - } break; - default: - { - GGML_ABORT("fatal error"); - } - } -} - -// ggml_compute_forward_map_custom1 - -static void ggml_compute_forward_map_custom1_f32( - const struct ggml_compute_params * params, - struct ggml_tensor * dst, - const ggml_custom1_op_f32_t fun) { - - const struct ggml_tensor * a = dst->src[0]; - - if (params->ith != 0) { - return; - } - - fun(dst, a); -} - -// ggml_compute_forward_map_custom2 - -static void ggml_compute_forward_map_custom2_f32( - const struct ggml_compute_params * params, - struct ggml_tensor * dst, - const ggml_custom2_op_f32_t fun) { - - const struct ggml_tensor * a = dst->src[0]; - const struct ggml_tensor * b = dst->src[1]; - - if (params->ith != 0) { - return; - } - - fun(dst, a, b); -} - -// ggml_compute_forward_map_custom3 - -static void ggml_compute_forward_map_custom3_f32( - const struct ggml_compute_params * params, - struct ggml_tensor * dst, - const ggml_custom3_op_f32_t fun) { - - const struct ggml_tensor * a = dst->src[0]; - const struct ggml_tensor * b = dst->src[1]; - const struct ggml_tensor * c = dst->src[1]; - - if (params->ith != 0) { - return; - } - - fun(dst, a, b, c); -} - -// ggml_compute_forward_map_custom1 - -static void ggml_compute_forward_map_custom1( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * a = dst->src[0]; - - struct ggml_map_custom1_op_params p; - memcpy(&p, dst->op_params, sizeof(p)); - - p.fun(dst, a, params->ith, params->nth, p.userdata); -} - -// ggml_compute_forward_map_custom2 - -static void ggml_compute_forward_map_custom2( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * a = dst->src[0]; - const struct ggml_tensor * b = dst->src[1]; - - struct ggml_map_custom2_op_params p; - memcpy(&p, dst->op_params, sizeof(p)); - - p.fun(dst, a, b, params->ith, params->nth, p.userdata); -} - -// ggml_compute_forward_map_custom3 - -static void ggml_compute_forward_map_custom3( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * a = dst->src[0]; - const struct ggml_tensor * b = dst->src[1]; - const struct ggml_tensor * c = dst->src[2]; - - struct ggml_map_custom3_op_params p; - memcpy(&p, dst->op_params, sizeof(p)); - - p.fun(dst, a, b, c, params->ith, params->nth, p.userdata); -} - -// ggml_compute_forward_cross_entropy_loss - -static void ggml_compute_forward_cross_entropy_loss_f32( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - const struct ggml_tensor * src1 = dst->src[1]; - - GGML_ASSERT(src0->type == GGML_TYPE_F32); - GGML_ASSERT(src1->type == GGML_TYPE_F32); - GGML_ASSERT(src0->nb[0] == ggml_type_size(src0->type)); - GGML_ASSERT(src1->nb[0] == ggml_type_size(src1->type)); - GGML_ASSERT(ggml_are_same_shape(src0, src1)); - GGML_ASSERT(ggml_is_scalar(dst)); - GGML_ASSERT(dst->type == GGML_TYPE_F32); - - // TODO: handle transposed/permuted matrices - const int64_t nc = src0->ne[0]; - const int64_t nr = ggml_nrows(src0); - - const int ith = params->ith; - const int nth = params->nth; - - float * sums = (float *) params->wdata; - float * st = ((float *) params->wdata) + nth + ith*nc; - float sum_thread = 0.0f; - - GGML_ASSERT(params->wsize >= sizeof(float) * (nth + nth * nc)); - - // rows per thread - const int64_t dr = (nr + nth - 1)/nth; - - // row range for this thread - const int64_t ir0 = dr*ith; - const int64_t ir1 = MIN(ir0 + dr, nr); - - for (int64_t i1 = ir0; i1 < ir1; ++i1) { - const float * s0 = (const float *)((const char *) src0->data + i1*src0->nb[1]); - const float * s1 = (const float *)((const char *) src1->data + i1*src1->nb[1]); - -#ifndef NDEBUG - for (int64_t i = 0; i < nc; ++i) { - //printf("p[%d] = %f\n", i, p[i]); - assert(!isnan(s0[i])); - assert(!isnan(s1[i])); - } -#endif - - float max = -INFINITY; - ggml_vec_max_f32(nc, &max, s0); - const ggml_float sum_softmax = ggml_vec_log_soft_max_f32(nc, st, s0, max); - assert(sum_softmax >= 0.0); - - ggml_vec_add1_f32(nc, st, st, -sum_softmax); - ggml_vec_mul_f32(nc, st, st, s1); - - float sum_st = 0.0f; - ggml_vec_sum_f32(nc, &sum_st, st); - sum_thread += sum_st; - -#ifndef NDEBUG - for (int64_t i = 0; i < nc; ++i) { - assert(!isnan(st[i])); - assert(!isinf(st[i])); - } -#endif - } - sums[ith] = sum_thread; - ggml_barrier(params->threadpool); - - if (ith == 0) { - float * dp = (float *) dst->data; - ggml_vec_sum_f32(nth, dp, sums); - dp[0] *= -1.0f / (float) nr; - } -} - -static void ggml_compute_forward_cross_entropy_loss( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - switch (src0->type) { - case GGML_TYPE_F32: - { - ggml_compute_forward_cross_entropy_loss_f32(params, dst); - } break; - default: - { - GGML_ABORT("fatal error"); - } - } -} - -// ggml_compute_forward_cross_entropy_loss_back - -static void ggml_compute_forward_cross_entropy_loss_back_f32( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * grad = dst->src[0]; // gradient of forward pass output - const struct ggml_tensor * src0f = dst->src[1]; // src0 of forward pass - const struct ggml_tensor * src1f = dst->src[2]; // src1 of forward pass - - GGML_ASSERT(ggml_is_contiguous(dst)); - GGML_ASSERT(ggml_is_contiguous(src0f)); - GGML_ASSERT(ggml_is_contiguous(src1f)); - GGML_ASSERT(ggml_is_contiguous(grad)); - GGML_ASSERT(ggml_are_same_shape(src0f, src1f) && ggml_are_same_shape(src0f, dst)); - - const int64_t ith = params->ith; - const int64_t nth = params->nth; - - // TODO: handle transposed/permuted matrices - const int64_t nc = src0f->ne[0]; - const int64_t nr = ggml_nrows(src0f); - - // rows per thread - const int64_t dr = (nr + nth - 1)/nth; - - // row range for this thread - const int64_t ir0 = dr*ith; - const int64_t ir1 = MIN(ir0 + dr, nr); - - const float d_by_nr = ((const float *) grad->data)[0] / (float) nr; - - for (int64_t i1 = ir0; i1 < ir1; i1++) { - float * ds0 = (float *)((char *) dst->data + i1*dst->nb[1]); - const float * s0 = (const float *)((const char *) src0f->data + i1*src0f->nb[1]); - const float * s1 = (const float *)((const char *) src1f->data + i1*src1f->nb[1]); - -#ifndef NDEBUG - for (int64_t i = 0; i < nc; ++i) { - //printf("p[%d] = %f\n", i, p[i]); - assert(!isnan(s0[i])); - assert(!isnan(s1[i])); - } -#endif - - // soft_max - float max = -INFINITY; - ggml_vec_max_f32(nc, &max, s0); - const ggml_float sum = ggml_vec_soft_max_f32(nc, ds0, s0, max); - assert(sum > 0.0); - ggml_vec_scale_f32(nc, ds0, 1.0/sum); - - // grad(src0f) = (softmax(src0f) - src1f) * grad(cross_entropy_loss(src0f, src1f)) / nr - ggml_vec_sub_f32(nc, ds0, ds0, s1); - ggml_vec_scale_f32(nc, ds0, d_by_nr); - -#ifndef NDEBUG - for (int64_t i = 0; i < nc; ++i) { - assert(!isnan(ds0[i])); - assert(!isinf(ds0[i])); - } -#endif - } -} - -static void ggml_compute_forward_cross_entropy_loss_back( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - switch (src0->type) { - case GGML_TYPE_F32: - { - ggml_compute_forward_cross_entropy_loss_back_f32(params, dst); - } break; - default: - { - GGML_ABORT("fatal error"); - } - } -} - -static void ggml_compute_forward_opt_step_adamw_f32( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - const struct ggml_tensor * src0_grad = dst->src[1]; - const struct ggml_tensor * src0_grad_m = dst->src[2]; - const struct ggml_tensor * src0_grad_v = dst->src[3]; - const struct ggml_tensor * adamw_params = dst->src[4]; - - GGML_ASSERT(ggml_are_same_shape(src0, src0_grad)); - GGML_ASSERT(ggml_are_same_shape(src0, src0_grad_m)); - GGML_ASSERT(ggml_are_same_shape(src0, src0_grad_v)); - GGML_ASSERT(ggml_nelements(adamw_params) == 7); - - const int ith = params->ith; - const int nth = params->nth; - - const int nr = ggml_nrows(src0); - - GGML_TENSOR_UNARY_OP_LOCALS - GGML_ASSERT(nb00 == sizeof(float)); - - // rows per thread - const int dr = (nr + nth - 1)/nth; - - // row range for this thread - const int ir0 = dr*ith; - const int ir1 = MIN(ir0 + dr, nr); - - const float * adamw_params_ptr = ggml_get_data_f32(adamw_params); - const float alpha = adamw_params_ptr[0]; - const float beta1 = adamw_params_ptr[1]; - const float beta2 = adamw_params_ptr[2]; - const float eps = adamw_params_ptr[3]; - const float wd = adamw_params_ptr[4]; - const float beta1h = adamw_params_ptr[5]; - const float beta2h = adamw_params_ptr[6]; - - for (int ir = ir0; ir < ir1; ++ir) { - const int64_t i03 = ir/(ne02*ne01); - const int64_t i02 = (ir - i03*ne02*ne01)/ne01; - const int64_t i01 = (ir - i03*ne02*ne01 - i02*ne01); - - const size_t offset = i03*nb03 + i02*nb02 + i01*nb01; - - float * w = (float *) ((char *) src0->data + offset); // weight - const float * g = (const float *) ((const char *) src0_grad->data + offset); // grad - float * m = (float *) ((char *) src0_grad_m->data + offset); - float * v = (float *) ((char *) src0_grad_v->data + offset); - - for (int i00 = 0; i00 < ne00; ++i00) { - m[i00] = m[i00]*beta1 + g[i00]*(1.0f - beta1); - v[i00] = v[i00]*beta2 + g[i00]*g[i00]*(1.0f - beta2); - - const float mh = m[i00]*beta1h; - const float vh = sqrtf(v[i00]*beta2h) + eps; - - // The weight decay is applied independently of the Adam momenta m and v. - // This is NOT equivalent to l2 regularization that adds w[i00]*w[i00] to the loss. - // See: https://arxiv.org/pdf/1711.05101v3.pdf - w[i00] = w[i00]*(1.0f - alpha*wd) - alpha*mh/vh; - } - } -} - -static void ggml_compute_forward_opt_step_adamw( - const struct ggml_compute_params * params, - struct ggml_tensor * dst) { - - const struct ggml_tensor * src0 = dst->src[0]; - - switch (src0->type) { - case GGML_TYPE_F32: - { - ggml_compute_forward_opt_step_adamw_f32(params, dst); - } break; - default: - { - GGML_ABORT("fatal error"); - } - } -} ///////////////////////////////// static void ggml_compute_forward(struct ggml_compute_params * params, struct ggml_tensor * tensor) { @@ -12720,7 +1727,9 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm } // extra_buffer op? - if (ggml_cpu_extra_compute_forward(params, tensor)) return; + if (ggml_cpu_extra_compute_forward(params, tensor)) { + return; + } switch (tensor->op) { case GGML_OP_DUP: @@ -12823,6 +1832,10 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm { ggml_compute_forward_group_norm(params, tensor); } break; + case GGML_OP_L2_NORM: + { + ggml_compute_forward_l2_norm(params, tensor); + } break; case GGML_OP_MUL_MAT: { ggml_compute_forward_mul_mat(params, tensor); @@ -13010,41 +2023,10 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm { ggml_compute_forward_gla(params, tensor); } break; - case GGML_OP_MAP_UNARY: + case GGML_OP_RWKV_WKV7: { - ggml_unary_op_f32_t fun; - memcpy(&fun, tensor->op_params, sizeof(fun)); - ggml_compute_forward_map_unary(params, tensor, fun); - } - break; - case GGML_OP_MAP_BINARY: - { - ggml_binary_op_f32_t fun; - memcpy(&fun, tensor->op_params, sizeof(fun)); - ggml_compute_forward_map_binary(params, tensor, fun); - } - break; - case GGML_OP_MAP_CUSTOM1_F32: - { - ggml_custom1_op_f32_t fun; - memcpy(&fun, tensor->op_params, sizeof(fun)); - ggml_compute_forward_map_custom1_f32(params, tensor, fun); - } - break; - case GGML_OP_MAP_CUSTOM2_F32: - { - ggml_custom2_op_f32_t fun; - memcpy(&fun, tensor->op_params, sizeof(fun)); - ggml_compute_forward_map_custom2_f32(params, tensor, fun); - } - break; - case GGML_OP_MAP_CUSTOM3_F32: - { - ggml_custom3_op_f32_t fun; - memcpy(&fun, tensor->op_params, sizeof(fun)); - ggml_compute_forward_map_custom3_f32(params, tensor, fun); - } - break; + ggml_compute_forward_rwkv_wkv7(params, tensor); + } break; case GGML_OP_MAP_CUSTOM1: { ggml_compute_forward_map_custom1(params, tensor); @@ -13060,6 +2042,11 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm ggml_compute_forward_map_custom3(params, tensor); } break; + case GGML_OP_CUSTOM: + { + ggml_compute_forward_custom(params, tensor); + } + break; case GGML_OP_CROSS_ENTROPY_LOSS: { ggml_compute_forward_cross_entropy_loss(params, tensor); @@ -13235,6 +2222,7 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) { case GGML_OP_NORM: case GGML_OP_RMS_NORM: case GGML_OP_RMS_NORM_BACK: + case GGML_OP_L2_NORM: case GGML_OP_GROUP_NORM: case GGML_OP_CONCAT: case GGML_OP_MUL_MAT: @@ -13301,19 +2289,15 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) { case GGML_OP_FLASH_ATTN_BACK: case GGML_OP_SSM_CONV: case GGML_OP_SSM_SCAN: + case GGML_OP_RWKV_WKV6: + case GGML_OP_GATED_LINEAR_ATTN: + case GGML_OP_RWKV_WKV7: { n_tasks = n_threads; } break; case GGML_OP_WIN_PART: case GGML_OP_WIN_UNPART: case GGML_OP_GET_REL_POS: - case GGML_OP_RWKV_WKV6: - case GGML_OP_GATED_LINEAR_ATTN: - case GGML_OP_MAP_UNARY: - case GGML_OP_MAP_BINARY: - case GGML_OP_MAP_CUSTOM1_F32: - case GGML_OP_MAP_CUSTOM2_F32: - case GGML_OP_MAP_CUSTOM3_F32: { n_tasks = 1; } break; @@ -13347,6 +2331,16 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) { n_tasks = MIN(p.n_tasks, n_threads); } } break; + case GGML_OP_CUSTOM: + { + struct ggml_custom_op_params p; + memcpy(&p, node->op_params, sizeof(p)); + if (p.n_tasks == GGML_N_TASKS_MAX) { + n_tasks = n_threads; + } else { + n_tasks = MIN(p.n_tasks, n_threads); + } + } break; case GGML_OP_CROSS_ENTROPY_LOSS: case GGML_OP_CROSS_ENTROPY_LOSS_BACK: case GGML_OP_OPT_STEP_ADAMW: @@ -13675,7 +2669,6 @@ struct ggml_cplan ggml_graph_plan( size_t cur = 0; if (!ggml_cpu_extra_work_size(n_threads, node, &cur)) { - switch (node->op) { case GGML_OP_CPY: case GGML_OP_DUP: @@ -13717,14 +2710,19 @@ struct ggml_cplan ggml_graph_plan( cur = 0; const struct ggml_tensor * src0 = node->src[0]; const struct ggml_tensor * src1 = node->src[1]; + const struct ggml_tensor * ids = node->src[2]; const enum ggml_type vec_dot_type = type_traits_cpu[src0->type].vec_dot_type; - if (src1->type != vec_dot_type) { - cur += ggml_row_size(vec_dot_type, ggml_nelements(src1)); - } const int n_as = src0->ne[2]; - cur += GGML_PAD(cur, sizeof(int64_t)); // align - cur += n_as * sizeof(int64_t); // matrix_row_counts - cur += n_as * src1->ne[2] * sizeof(int64_t); // matrix_rows + // src1 + if (src1->type != vec_dot_type) { + cur += ggml_row_size(vec_dot_type, ggml_nelements(src1)) + sizeof(int64_t); + } + // matrix_row_counts + cur += n_as * sizeof(int64_t) + sizeof(int64_t); + // matrix_rows + cur += n_as*ids->ne[0]*ids->ne[1]*sizeof(struct mmid_row_mapping) + sizeof(int64_t); + // atomic_current_chunk + cur += CACHE_LINE_SIZE*n_as + CACHE_LINE_SIZE; } break; case GGML_OP_OUT_PROD: { @@ -13779,9 +2777,10 @@ struct ggml_cplan ggml_graph_plan( } break; case GGML_OP_FLASH_ATTN_EXT: { - const int64_t ne00 = node->src[0]->ne[0]; // D + const int64_t ne10 = node->src[1]->ne[0]; // DK + const int64_t ne20 = node->src[2]->ne[0]; // DV - cur = 3*sizeof(float)*ne00*n_tasks; // 3x head size/thread + cur = sizeof(float)*(1*ne10 + 2*ne20)*n_tasks; // 1x head size K + 2x head size V (per thread) } break; case GGML_OP_FLASH_ATTN_BACK: { @@ -14227,6 +3226,14 @@ int ggml_cpu_has_amx_int8(void) { #endif } +int ggml_cpu_has_bmi2(void) { +#if defined(__BMI2__) + return 1; +#else + return 0; +#endif +} + int ggml_cpu_has_fma(void) { #if defined(__FMA__) return 1; @@ -14307,6 +3314,14 @@ int ggml_cpu_has_vsx(void) { #endif } +int ggml_cpu_has_vxe(void) { +#if defined(__VXE__) || defined(__VXE2__) + return 1; +#else + return 0; +#endif +} + int ggml_cpu_has_neon(void) { #if defined(__ARM_ARCH) && defined(__ARM_NEON) return ggml_arm_arch_features.has_neon; @@ -14347,6 +3362,14 @@ int ggml_cpu_get_sve_cnt(void) { #endif } +int ggml_cpu_has_sme(void) { +#if defined(__ARM_ARCH) && defined(__ARM_FEATURE_SME) + return ggml_arm_arch_features.has_sme; +#else + return 0; +#endif +} + void ggml_cpu_init(void) { // needed to initialize f16 tables { diff --git a/ggml/src/ggml-cpu/ggml-cpu.cpp b/ggml/src/ggml-cpu/ggml-cpu.cpp index 2ccb4b472..09f8382b9 100644 --- a/ggml/src/ggml-cpu/ggml-cpu.cpp +++ b/ggml/src/ggml-cpu/ggml-cpu.cpp @@ -14,6 +14,10 @@ #include "ggml-cpu-hbm.h" #endif +#ifdef GGML_USE_CPU_KLEIDIAI +#include "kleidiai/kleidiai.h" +#endif + #if defined(__APPLE__) #include #include @@ -39,6 +43,12 @@ std::vector& ggml_backend_cpu_get_extra_buffers_type } #endif +#ifdef GGML_USE_CPU_KLEIDIAI + if (ggml_backend_cpu_kleidiai_buffer_type()) { + bufts.push_back(ggml_backend_cpu_kleidiai_buffer_type()); + } +#endif + #ifdef GGML_USE_CPU_AARCH64 if (ggml_backend_cpu_aarch64_buffer_type()) { bufts.push_back(ggml_backend_cpu_aarch64_buffer_type()); @@ -284,14 +294,14 @@ struct ggml_backend_cpu_device_context { &hKey) == ERROR_SUCCESS) { DWORD cpu_brand_size = 0; if (RegQueryValueExA(hKey, - TEXT("ProcessorNameString"), + "ProcessorNameString", NULL, NULL, NULL, &cpu_brand_size) == ERROR_SUCCESS) { description.resize(cpu_brand_size); if (RegQueryValueExA(hKey, - TEXT("ProcessorNameString"), + "ProcessorNameString", NULL, NULL, (LPBYTE)&description[0], // NOLINT @@ -501,6 +511,9 @@ static ggml_backend_feature * ggml_backend_cpu_get_features(ggml_backend_reg_t r if (ggml_cpu_has_fma()) { features.push_back({ "FMA", "1" }); } + if (ggml_cpu_has_bmi2()) { + features.push_back({ "BMI2", "1" }); + } if (ggml_cpu_has_avx512()) { features.push_back({ "AVX512", "1" }); } @@ -534,19 +547,22 @@ static ggml_backend_feature * ggml_backend_cpu_get_features(ggml_backend_reg_t r if (ggml_cpu_has_dotprod()) { features.push_back({ "DOTPROD", "1" }); } - if (ggml_cpu_has_matmul_int8()) { - features.push_back({ "MATMUL_INT8", "1" }); - } if (ggml_cpu_get_sve_cnt() > 0) { static std::string sve_cnt = std::to_string(ggml_cpu_get_sve_cnt()); features.push_back({ "SVE_CNT", sve_cnt.c_str() }); } + if (ggml_cpu_has_sme()) { + features.push_back({ "SME", "1" }); + } if (ggml_cpu_has_riscv_v()) { features.push_back({ "RISCV_V", "1" }); } if (ggml_cpu_has_vsx()) { features.push_back({ "VSX", "1" }); } + if (ggml_cpu_has_vxe()) { + features.push_back({ "VXE", "1" }); + } if (ggml_cpu_has_wasm_simd()) { features.push_back({ "WASM_SIMD", "1" }); } @@ -562,6 +578,9 @@ static ggml_backend_feature * ggml_backend_cpu_get_features(ggml_backend_reg_t r #ifdef GGML_USE_OPENMP features.push_back({ "OPENMP", "1" }); #endif + #ifdef GGML_USE_CPU_KLEIDIAI + features.push_back({ "KLEIDIAI", "1" }); + #endif #ifdef GGML_USE_CPU_AARCH64 features.push_back({ "AARCH64_REPACK", "1" }); #endif diff --git a/ggml/src/ggml-cpu/kleidiai/kernels.cpp b/ggml/src/ggml-cpu/kleidiai/kernels.cpp new file mode 100644 index 000000000..aacc2bb5e --- /dev/null +++ b/ggml/src/ggml-cpu/kleidiai/kernels.cpp @@ -0,0 +1,254 @@ +// SPDX-FileCopyrightText: Copyright 2025 Arm Limited and/or its affiliates +// SPDX-License-Identifier: MIT +// + +// KleidiAI micro-kernels +#include "kai_matmul_clamp_f32_qsi8d32p_qsi4c32p_interface.h" +#include "kai_lhs_quant_pack_qsi8d32p_f32.h" +#include "kai_lhs_quant_pack_qsi8d32p_f32_neon.h" +#include "kai_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0.h" +#include "kai_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon.h" +#include "kai_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod.h" +#include "kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod.h" +#include "kai_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod.h" +#include "kai_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm.h" +#include "kai_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa.h" +#include "kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot.h" +#include "kai_common.h" + +#include "kernels.h" + +#define NELEMS(x) sizeof(x) / sizeof(*x) +static ggml_kleidiai_kernels gemm_gemv_kernels[] = { +#if defined(__ARM_FEATURE_SME) + { + /* SME GEMM */ + /* .kern_info = */ { + /* .get_m_step = */ kai_get_m_step_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa, + /* .get_n_step = */ kai_get_n_step_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa, + /* .get_mr = */ kai_get_mr_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa, + /* .get_nr = */ kai_get_nr_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa, + /* .get_kr = */ kai_get_kr_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa, + /* .get_sr = */ kai_get_sr_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa, + /* .get_lhs_offset = */ kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa, + /* .get_rhs_packed_offset = */ kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa, + /* .get_dst_offset = */ kai_get_dst_offset_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa, + /* .get_dst_size = */ kai_get_dst_size_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa, + /* .run_kernel = */ kai_run_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa, + }, + /* SME GEMV */ + /* .kern_info = */ { + /* .get_m_step = */ kai_get_m_step_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot, + /* .get_n_step = */ kai_get_n_step_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot, + /* .get_mr = */ kai_get_mr_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot, + /* .get_nr = */ kai_get_nr_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot, + /* .get_kr = */ kai_get_kr_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot, + /* .get_sr = */ kai_get_sr_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot, + /* .get_lhs_offset = */ kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot, + /* .get_rhs_packed_offset = */ kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot, + /* .get_dst_offset = */ kai_get_dst_offset_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot, + /* .get_dst_size = */ kai_get_dst_size_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot, + /* .run_kernel = */ kai_run_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot, + }, + /* .lhs_info = */ { + /* .get_offset = */ kai_get_lhs_offset_lhs_quant_pack_qsi8d32p_f32_neon, + /* .get_packed_offset = */ kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32p_f32_neon, + /* .packed_size = */ kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32p_f32_neon, + /* .pack_func = */ kai_run_lhs_quant_pack_qsi8d32p_f32_neon, + }, + /* .rhs_info = */ { + /* .packed_size = */ kai_get_rhs_packed_size_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon, + /* .pack_func = */ kai_run_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon, + }, + /* .required_cpu = */ CPU_FEATURE_SME, + }, +#endif +#if defined(__APPLE__) +#if defined(__ARM_FEATURE_DOTPROD) + { + /* DOTPROD GEMM */ + /* .kern_info = */ { + /* .get_m_step = */ kai_get_m_step_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod, + /* .get_n_step = */ kai_get_n_step_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod, + /* .get_mr = */ kai_get_mr_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod, + /* .get_nr = */ kai_get_nr_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod, + /* .get_kr = */ kai_get_kr_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod, + /* .get_sr = */ kai_get_sr_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod, + /* .get_lhs_offset = */ kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod, + /* .get_rhs_packed_offset = */ kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod, + /* .get_dst_offset = */ kai_get_dst_offset_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod, + /* .get_dst_size = */ kai_get_dst_size_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod, + /* .run_kernel = */ kai_run_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod, + }, + /* DOTPROD GEMV */ + /* .kern_info = */ { + /* .get_m_step = */ kai_get_m_step_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod, + /* .get_n_step = */ kai_get_n_step_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod, + /* .get_mr = */ kai_get_mr_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod, + /* .get_nr = */ kai_get_nr_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod, + /* .get_kr = */ kai_get_kr_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod, + /* .get_sr = */ kai_get_sr_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod, + /* .get_lhs_offset = */ kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod, + /* .get_rhs_packed_offset = */ kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod, + /* .get_dst_offset = */ kai_get_dst_offset_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod, + /* .get_dst_size = */ kai_get_dst_size_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod, + /* .run_kernel = */ kai_run_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod, + }, + /* .lhs_info = */ { + /* .get_offset = */ kai_get_lhs_offset_lhs_quant_pack_qsi8d32p_f32, + /* .get_packed_offset = */ kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32p_f32, + /* .packed_size = */ kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32p_f32, + /* .pack_func = */ kai_run_lhs_quant_pack_qsi8d32p_f32, + }, + /* .rhs_info = */ { + /* .packed_size = */ kai_get_rhs_packed_size_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0, + /* .pack_func = */ kai_run_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0, + }, + /* .required_cpu = */ CPU_FEATURE_DOTPROD, + }, +#endif +#if defined(__ARM_FEATURE_MATMUL_INT8) + { + /* i8mm GEMM */ + /* .kern_info = */ { + /* .get_m_step = */ kai_get_m_step_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm, + /* .get_n_step = */ kai_get_n_step_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm, + /* .get_mr = */ kai_get_mr_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm, + /* .get_nr = */ kai_get_nr_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm, + /* .get_kr = */ kai_get_kr_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm, + /* .get_sr = */ kai_get_sr_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm, + /* .get_lhs_offset = */ kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm, + /* .get_rhs_packed_offset = */ kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm, + /* .get_dst_offset = */ kai_get_dst_offset_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm, + /* .get_dst_size = */ kai_get_dst_size_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm, + /* .run_kernel = */ kai_run_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm, + }, + /* i8mm GEMV */ + /* .kern_info = */ { + /* .get_m_step = */ kai_get_m_step_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod, + /* .get_n_step = */ kai_get_n_step_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod, + /* .get_mr = */ kai_get_mr_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod, + /* .get_nr = */ kai_get_nr_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod, + /* .get_kr = */ kai_get_kr_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod, + /* .get_sr = */ kai_get_sr_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod, + /* .get_lhs_offset = */ kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod, + /* .get_rhs_packed_offset = */ kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod, + /* .get_dst_offset = */ kai_get_dst_offset_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod, + /* .get_dst_size = */ kai_get_dst_size_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod, + /* .run_kernel = */ kai_run_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod, + }, + /* .lhs_info = */ { + /* .get_offset = */ kai_get_lhs_offset_lhs_quant_pack_qsi8d32p_f32, + /* .get_packed_offset = */ kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32p_f32, + /* .packed_size = */ kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32p_f32, + /* .pack_func = */ kai_run_lhs_quant_pack_qsi8d32p_f32, + }, + /* .rhs_info = */ { + /* .packed_size = */ kai_get_rhs_packed_size_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0, + /* .pack_func = */ kai_run_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0, + }, + /* .required_cpu = */ CPU_FEATURE_DOTPROD | CPU_FEATURE_I8MM, + }, +#endif +#else +#if defined(__ARM_FEATURE_MATMUL_INT8) + { + /* i8mm GEMM */ + /* .kern_info = */ { + /* .get_m_step = */ kai_get_m_step_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm, + /* .get_n_step = */ kai_get_n_step_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm, + /* .get_mr = */ kai_get_mr_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm, + /* .get_nr = */ kai_get_nr_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm, + /* .get_kr = */ kai_get_kr_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm, + /* .get_sr = */ kai_get_sr_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm, + /* .get_lhs_offset = */ kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm, + /* .get_rhs_packed_offset = */ kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm, + /* .get_dst_offset = */ kai_get_dst_offset_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm, + /* .get_dst_size = */ kai_get_dst_size_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm, + /* .run_kernel = */ kai_run_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm, + }, + /* i8mm GEMV */ + /* .kern_info = */ { + /* .get_m_step = */ kai_get_m_step_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod, + /* .get_n_step = */ kai_get_n_step_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod, + /* .get_mr = */ kai_get_mr_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod, + /* .get_nr = */ kai_get_nr_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod, + /* .get_kr = */ kai_get_kr_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod, + /* .get_sr = */ kai_get_sr_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod, + /* .get_lhs_offset = */ kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod, + /* .get_rhs_packed_offset = */ kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod, + /* .get_dst_offset = */ kai_get_dst_offset_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod, + /* .get_dst_size = */ kai_get_dst_size_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod, + /* .run_kernel = */ kai_run_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod, + }, + /* .lhs_info = */ { + /* .get_offset = */ kai_get_lhs_offset_lhs_quant_pack_qsi8d32p_f32, + /* .get_packed_offset = */ kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32p_f32, + /* .packed_size = */ kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32p_f32, + /* .pack_func = */ kai_run_lhs_quant_pack_qsi8d32p_f32, + }, + /* .rhs_info = */ { + /* .packed_size = */ kai_get_rhs_packed_size_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0, + /* .pack_func = */ kai_run_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0, + }, + /* .required_cpu = */ CPU_FEATURE_DOTPROD | CPU_FEATURE_I8MM, + }, +#endif +#if defined(__ARM_FEATURE_DOTPROD) + { + /* DOTPROD GEMM */ + /* .kern_info = */ { + /* .get_m_step = */ kai_get_m_step_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod, + /* .get_n_step = */ kai_get_n_step_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod, + /* .get_mr = */ kai_get_mr_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod, + /* .get_nr = */ kai_get_nr_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod, + /* .get_kr = */ kai_get_kr_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod, + /* .get_sr = */ kai_get_sr_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod, + /* .get_lhs_offset = */ kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod, + /* .get_rhs_packed_offset = */ kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod, + /* .get_dst_offset = */ kai_get_dst_offset_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod, + /* .get_dst_size = */ kai_get_dst_size_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod, + /* .run_kernel = */ kai_run_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod, + }, + /* DOTPROD GEMV */ + /* .kern_info = */ { + /* .get_m_step = */ kai_get_m_step_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod, + /* .get_n_step = */ kai_get_n_step_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod, + /* .get_mr = */ kai_get_mr_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod, + /* .get_nr = */ kai_get_nr_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod, + /* .get_kr = */ kai_get_kr_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod, + /* .get_sr = */ kai_get_sr_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod, + /* .get_lhs_offset = */ kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod, + /* .get_rhs_packed_offset = */ kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod, + /* .get_dst_offset = */ kai_get_dst_offset_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod, + /* .get_dst_size = */ kai_get_dst_size_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod, + /* .run_kernel = */ kai_run_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod, + }, + /* .lhs_info = */ { + /* .get_offset = */ kai_get_lhs_offset_lhs_quant_pack_qsi8d32p_f32, + /* .get_packed_offset = */ kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32p_f32, + /* .packed_size = */ kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32p_f32, + /* .pack_func = */ kai_run_lhs_quant_pack_qsi8d32p_f32, + }, + /* .rhs_info = */ { + /* .packed_size = */ kai_get_rhs_packed_size_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0, + /* .pack_func = */ kai_run_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0, + }, + /* .required_cpu = */ CPU_FEATURE_DOTPROD, + }, +#endif +#endif +}; + +ggml_kleidiai_kernels * ggml_kleidiai_select_kernels(cpu_feature features) { + ggml_kleidiai_kernels * kernels = nullptr; + + for (size_t i = 0; i < NELEMS(gemm_gemv_kernels); ++i) { + if ((features & gemm_gemv_kernels[i].required_cpu) == gemm_gemv_kernels[i].required_cpu) { + kernels = &gemm_gemv_kernels[i]; + break; + } + } + + return kernels; +} diff --git a/ggml/src/ggml-cpu/kleidiai/kernels.h b/ggml/src/ggml-cpu/kleidiai/kernels.h new file mode 100644 index 000000000..2ffe97eb4 --- /dev/null +++ b/ggml/src/ggml-cpu/kleidiai/kernels.h @@ -0,0 +1,60 @@ +// SPDX-FileCopyrightText: Copyright 2025 Arm Limited and/or its affiliates +// SPDX-License-Identifier: MIT +// + +#pragma once + +enum cpu_feature { + CPU_FEATURE_NONE = 0, + CPU_FEATURE_DOTPROD = 1, + CPU_FEATURE_I8MM = 2, + CPU_FEATURE_SVE = 4, + CPU_FEATURE_SME = 8 +}; +inline cpu_feature& operator|=(cpu_feature& lhs, cpu_feature rhs) { + lhs = static_cast(lhs | rhs); + return lhs; +} +inline cpu_feature operator|(cpu_feature lhs, cpu_feature rhs) { + return static_cast(static_cast(lhs) | static_cast(rhs)); +} + +struct kernel_info { + size_t (*get_m_step)(void); + size_t (*get_n_step)(void); + size_t (*get_mr)(void); + size_t (*get_nr)(void); + size_t (*get_kr)(void); + size_t (*get_sr)(void); + size_t (*get_lhs_offset)(size_t m_idx, size_t k, size_t bl); + size_t (*get_rhs_packed_offset)(size_t n_idx, size_t k, size_t bl); + size_t (*get_dst_offset)(size_t m_idx, size_t n_idx, size_t stride); + size_t (*get_dst_size)(size_t m, size_t n); + void (*run_kernel)(size_t m, size_t n, size_t k, size_t bl, const void* lhs_packed, const void* rhs_packed, + float* dst, size_t dst_stride_row, size_t dst_stride_col, float scalar_min, float scalar_max); +}; + +struct lhs_packing_info { + size_t (*get_offset)(size_t m_idx, size_t lhs_stride); + size_t (*get_packed_offset)(size_t m_idx, size_t k, size_t bl, size_t mr, size_t kr, size_t sr); + size_t (*packed_size)(size_t m, size_t k, size_t bl, size_t mr, size_t kr, size_t sr); + void (*pack_func)(size_t m, size_t k, size_t bl, size_t mr, size_t kr, size_t sr, size_t m_idx_start, const float* lhs, + size_t lhs_stride, void* lhs_packed); +}; + +struct rhs_packing_info { + size_t (*packed_size)(size_t n, size_t k, size_t nr, size_t kr, size_t bl); + void (*pack_func)(size_t num_groups, size_t n, size_t k, size_t nr, size_t kr, size_t sr, size_t bl, const uint8_t* rhs, + const float* bias, void* rhs_packed, size_t extra_bytes, const struct kai_rhs_pack_qs4cxs1s0_param* params); +}; + +struct ggml_kleidiai_kernels { + kernel_info gemm; + kernel_info gemv; + lhs_packing_info lhs_info; + rhs_packing_info rhs_info; + + cpu_feature required_cpu; +}; + +ggml_kleidiai_kernels * ggml_kleidiai_select_kernels(cpu_feature cpu_features); diff --git a/ggml/src/ggml-cpu/kleidiai/kleidiai.cpp b/ggml/src/ggml-cpu/kleidiai/kleidiai.cpp new file mode 100644 index 000000000..4e89ca0fa --- /dev/null +++ b/ggml/src/ggml-cpu/kleidiai/kleidiai.cpp @@ -0,0 +1,287 @@ +// SPDX-FileCopyrightText: Copyright 2025 Arm Limited and/or its affiliates +// SPDX-License-Identifier: MIT +// +#include +#include +#include +#include +#include +#if defined(__linux__) +#include +#include +#elif defined(__APPLE__) +#include +#include +#include +#elif defined(_WIN32) +#include +#include +#endif + +#include "kleidiai.h" + +#include "ggml-cpu.h" +#include "ggml-impl.h" +#include "ggml-backend-impl.h" +#include "ggml-threading.h" +#include "ggml-cpu-traits.h" + +#include "kernels.h" + +#include "kai_common.h" + +#define GGML_COMMON_DECL_CPP +#include "ggml-common.h" + +struct ggml_kleidiai_context { + ggml_kleidiai_kernels * kernels; +} static ctx = { NULL }; + +static void init_kleidiai_context(void) { + + ggml_critical_section_start(); + static bool initialized = false; + + if (!initialized) { + initialized = true; + const char *env_var = getenv("GGML_KLEIDIAI_SME"); + int sme_enabled = 0; + + cpu_feature features = (ggml_cpu_has_dotprod() ? CPU_FEATURE_DOTPROD : CPU_FEATURE_NONE) | + (ggml_cpu_has_matmul_int8() ? CPU_FEATURE_I8MM : CPU_FEATURE_NONE) | + (ggml_cpu_has_sve() ? CPU_FEATURE_SVE : CPU_FEATURE_NONE); + + if (env_var) { + sme_enabled = atoi(env_var); + } + + if (sme_enabled != 0) { + features |= ggml_cpu_has_sme() ? CPU_FEATURE_SME : CPU_FEATURE_NONE; + } + ctx.kernels = ggml_kleidiai_select_kernels(features); + } + ggml_critical_section_end(); +} + +static inline int64_t ggml_ne(const ggml_tensor * tensor, int dim) { + GGML_ASSERT(dim >= 0 && dim < GGML_MAX_DIMS); + return tensor->ne[dim]; +} + +namespace ggml::cpu::kleidiai { +class tensor_traits : public ggml::cpu::tensor_traits { + bool work_size(int /* n_threads */, const struct ggml_tensor * op, size_t & size) override { + GGML_ASSERT(ctx.kernels); + kernel_info * kernel = op->src[1]->ne[1] == 1 ? &ctx.kernels->gemv : &ctx.kernels->gemm; + + size_t k = op->src[0]->ne[0]; + size_t m = op->src[1]->ne[1]; + + size_t mr = kernel->get_mr(); + size_t kr = kernel->get_kr(); + size_t sr = kernel->get_sr(); + + size = ctx.kernels->lhs_info.packed_size(m, k, QK4_0, mr, kr, sr); + + return true; + } + + bool compute_forward(struct ggml_compute_params * params, struct ggml_tensor * dst) override { + if (dst->op == GGML_OP_MUL_MAT) { + const ggml_tensor * src0 = dst->src[0]; + const ggml_tensor * src1 = dst->src[1]; + + GGML_TENSOR_BINARY_OP_LOCALS + + GGML_ASSERT(ctx.kernels); + kernel_info * kernel = src1->ne[1] == 1 ? &ctx.kernels->gemv : &ctx.kernels->gemm; + lhs_packing_info * lhs_info = &ctx.kernels->lhs_info; + + GGML_ASSERT(kernel); + + const int ith = params->ith; + const int nth = params->nth; + + const size_t k = ne00; + const size_t m = ne11; + const size_t n = ne01; + + const size_t n_step = kernel->get_n_step(); + const size_t num_n_per_thread = kai_roundup(kai_roundup(n, nth) / nth, n_step); + const size_t n_start = ith * num_n_per_thread; + + size_t n_to_process = num_n_per_thread; + if ((n_start + n_to_process) > n) { + n_to_process = n - n_start; + } + + const uint8_t * lhs = static_cast(src1->data); + uint8_t * lhs_packed = (uint8_t*)params->wdata; + const uint8_t * rhs_packed = static_cast(src0->data); + + size_t mr = kernel->get_mr(); + size_t kr = kernel->get_kr(); + size_t sr = kernel->get_sr(); + + // Calculate number of columns to be processed per thread + const size_t num_m_per_thread = kai_roundup(m, mr * nth) / nth; + const size_t m_start = ith * num_m_per_thread; + size_t m_to_process = num_m_per_thread; + if ((m_start + m_to_process) > m) { + m_to_process = m - m_start; + } + + if(m_start < m) { + // Transform LHS + const size_t src_stride = src1->nb[1]; + const float * src_ptr = reinterpret_cast(lhs + lhs_info->get_offset(m_start, dst->src[1]->nb[1])); + const size_t lhs_packed_offset = lhs_info->get_packed_offset(m_start, k, QK4_0, mr, kr, sr); + void * lhs_packed_ptr = static_cast(lhs_packed + lhs_packed_offset); + + lhs_info->pack_func(m_to_process, k, QK4_0, mr, kr, sr, 0, src_ptr, src_stride, lhs_packed_ptr); + } + + ggml_barrier(params->threadpool); + + // Perform the operation + const size_t dst_stride = dst->nb[1]; + const size_t lhs_packed_offset = lhs_info->get_packed_offset(0, k, QK4_0, mr, kr, sr); + const size_t rhs_packed_offset = kernel->get_rhs_packed_offset(n_start, k, QK4_0); + const size_t dst_offset = kernel->get_dst_offset(0, n_start, dst_stride); + const void * rhs_ptr = static_cast(rhs_packed + rhs_packed_offset); + const void* lhs_ptr = (const void*)((const char *)lhs_packed + lhs_packed_offset); + float *dst_ptr = reinterpret_cast(static_cast(dst->data) + dst_offset); + + kernel->run_kernel(m, n_to_process, k, QK4_0, lhs_ptr, rhs_ptr, dst_ptr, + dst_stride, sizeof(float), -FLT_MAX, FLT_MAX); + return true; + } + return false; + } + +public: + int repack(struct ggml_tensor * tensor, const void * data, size_t data_size) { + GGML_ASSERT(ctx.kernels); + const size_t n = tensor->ne[1]; + const size_t k = tensor->ne[0]; + size_t nr = ctx.kernels->gemm.get_nr(); + size_t kr = ctx.kernels->gemm.get_kr(); + size_t sr = ctx.kernels->gemm.get_sr(); + +#ifndef NDEBUG + const size_t repacked_size = ctx.kernels->rhs_info.packed_size(n, k, nr, kr, QK4_0); + GGML_ASSERT(repacked_size <= data_size && "repacked size larger than the packed size!"); +#endif + struct kai_rhs_pack_qs4cxs1s0_param params; + params.lhs_zero_point = 1; + params.rhs_zero_point = 8; + ctx.kernels->rhs_info.pack_func(1, n, k, nr, kr, sr, QK4_0, (const uint8_t *)data, NULL, tensor->data, 0, ¶ms); + + return 0; + + GGML_UNUSED(data_size); + } +}; + +static ggml::cpu::tensor_traits * get_tensor_traits(ggml_backend_buffer_t, struct ggml_tensor *) { + static tensor_traits traits; + return &traits; +} +} // namespace ggml::cpu::kleidiai + +GGML_API enum ggml_status ggml_backend_cpu_kleidiai_buffer_init_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) { + tensor->extra = (void *) ggml::cpu::kleidiai::get_tensor_traits(buffer, tensor); + + GGML_UNUSED(buffer); + return GGML_STATUS_SUCCESS; +} + +static void ggml_backend_cpu_kleidiai_buffer_set_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, + const void * data, size_t offset, size_t size) { + GGML_ASSERT(offset == 0); + GGML_ASSERT(size == ggml_nbytes(tensor)); + + auto tensor_traits = (ggml::cpu::kleidiai::tensor_traits *) tensor->extra; + auto OK = tensor_traits->repack(tensor, data, size); + + GGML_ASSERT(OK == 0); + GGML_UNUSED(buffer); +} + +static const char * ggml_backend_cpu_kleidiai_buffer_type_get_name(ggml_backend_buffer_type_t buft) { + return "CPU_KLEIDIAI"; + + GGML_UNUSED(buft); +} + +static ggml_backend_buffer_t ggml_backend_cpu_kleidiai_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) { + ggml_backend_buffer_t buffer = ggml_backend_buft_alloc_buffer(ggml_backend_cpu_buffer_type(), size); + + if (buffer == nullptr) { + return nullptr; + } + + buffer->buft = buft; + buffer->iface.init_tensor = ggml_backend_cpu_kleidiai_buffer_init_tensor; + buffer->iface.set_tensor = ggml_backend_cpu_kleidiai_buffer_set_tensor; + buffer->iface.get_tensor = nullptr; + buffer->iface.cpy_tensor = nullptr; + return buffer; +} + +static size_t ggml_backend_cpu_kleidiai_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) { + return TENSOR_ALIGNMENT; + + GGML_UNUSED(buft); +} + +namespace ggml::cpu::kleidiai { +class extra_buffer_type : ggml::cpu::extra_buffer_type { + bool supports_op(ggml_backend_dev_t, const struct ggml_tensor * op) override { + if ( op->op == GGML_OP_MUL_MAT && + op->src[0]->type == GGML_TYPE_Q4_0 && + op->src[0]->buffer && + (ggml_n_dims(op->src[0]) == 2) && + op->src[0]->buffer->buft == ggml_backend_cpu_kleidiai_buffer_type() && ctx.kernels + ) { + if (op->src[1]->buffer && !ggml_backend_buft_is_host(op->src[1]->buffer->buft)) { + return false; + } + if (op->src[1]->type == GGML_TYPE_F32 && + ggml_ne(op->src[1], 2) == 1 && ggml_ne(op->src[1], 3) == 1) { + return true; + } + } + return false; + } + + ggml::cpu::tensor_traits * get_tensor_traits(const struct ggml_tensor * op) override { + if (op->op == GGML_OP_MUL_MAT) { + if (op->src[0]->buffer && op->src[0]->buffer->buft == ggml_backend_cpu_kleidiai_buffer_type()) { + return (ggml::cpu::tensor_traits *) op->src[0]->extra; + } + } + return nullptr; + } +}; +} // namespace ggml::cpu::kleidiai + +ggml_backend_buffer_type_t ggml_backend_cpu_kleidiai_buffer_type(void) { + static ggml::cpu::kleidiai::extra_buffer_type ctx; + static struct ggml_backend_buffer_type ggml_backend_cpu_buffer_type_kleidiai = { + /* .iface = */ { + /* .get_name = */ ggml_backend_cpu_kleidiai_buffer_type_get_name, + /* .alloc_buffer = */ ggml_backend_cpu_kleidiai_buffer_type_alloc_buffer, + /* .get_alignment = */ ggml_backend_cpu_kleidiai_buffer_type_get_alignment, + /* .get_max_size = */ nullptr, // defaults to SIZE_MAX + /* .get_alloc_size = */ nullptr, // defaults to ggml_nbytes + /* .is_host = */ nullptr, + }, + /* .device = */ ggml_backend_reg_dev_get(ggml_backend_cpu_reg(), 0), + /* .context = */ &ctx, + }; + + init_kleidiai_context(); + + return &ggml_backend_cpu_buffer_type_kleidiai; +} diff --git a/ggml/src/ggml-cpu/kleidiai/kleidiai.h b/ggml/src/ggml-cpu/kleidiai/kleidiai.h new file mode 100644 index 000000000..38eac58f7 --- /dev/null +++ b/ggml/src/ggml-cpu/kleidiai/kleidiai.h @@ -0,0 +1,17 @@ +// SPDX-FileCopyrightText: Copyright 2025 Arm Limited and/or its affiliates +// SPDX-License-Identifier: MIT +// + +#pragma once + +#include "ggml-alloc.h" + +#ifdef __cplusplus +extern "C" { +#endif + +ggml_backend_buffer_type_t ggml_backend_cpu_kleidiai_buffer_type(void); + +#ifdef __cplusplus +} +#endif diff --git a/ggml/src/ggml-cpu/llamafile/sgemm.cpp b/ggml/src/ggml-cpu/llamafile/sgemm.cpp index c22a66287..f6374f789 100644 --- a/ggml/src/ggml-cpu/llamafile/sgemm.cpp +++ b/ggml/src/ggml-cpu/llamafile/sgemm.cpp @@ -55,6 +55,7 @@ #include #include +#include #ifdef _MSC_VER #define NOINLINE __declspec(noinline) @@ -280,14 +281,6 @@ template <> inline __m256bh load(const float *p) { } #endif -//////////////////////////////////////////////////////////////////////////////////////////////////// -// CONSTANTS - -#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__) -static const int8_t kvalues_iq4nl[16] = {-127, -104, -83, -65, -49, -35, -22, -10, 1, 13, 25, 38, 53, 69, 89, 113}; -static const __m128i iq4nlt = _mm_loadu_si128((const __m128i *) kvalues_iq4nl); -#endif - //////////////////////////////////////////////////////////////////////////////////////////////////// // FLOATING POINT MATRIX MULTIPLICATION @@ -614,6 +607,14 @@ class tinyBLAS_Q0_AVX { TC *C, int64_t ldc, int ith, int nth) : A(A), B(B), C(C), k(k), lda(lda), ldb(ldb), ldc(ldc), ith(ith), nth(nth) { + const int8_t kvalues_iq4nl[16] = { + -127, -104, -83, -65, + -49, -35, -22, -10, + 1, 13, 25, 38, + 53, 69, 89, 113 + }; + + iq4nlt = _mm_loadu_si128((const __m128i *)kvalues_iq4nl); } void matmul(int64_t m, int64_t n) { @@ -1038,6 +1039,7 @@ class tinyBLAS_Q0_AVX { const int64_t ldc; const int ith; const int nth; + __m128i iq4nlt; }; #endif // __AVX__ @@ -1091,13 +1093,403 @@ class tinyBLAS_Q0_PPC { } } - template - void packNormal(const TA* a, int64_t lda, int rows, int cols, VA* vec, bool flip) { + template + void packNormalInt4(const TA* a, int64_t lda, int rows, int cols, VA* vec, std::array& comparray) { int64_t i, j; TA *aoffset = NULL; VA *vecOffset = NULL; TA *aoffset1 = NULL, *aoffset2 = NULL, *aoffset3 = NULL, *aoffset4 = NULL; TA *aoffset5 = NULL, *aoffset6 = NULL, *aoffset7 = NULL, *aoffset8 = NULL; + VB c1[2] = {0}, c2[2] = {0}, c3[2] = {0}, c4[2] = {0}; + VB c5[2] = {0}, c6[2] = {0}, c7[2] = {0}, c8[2] = {0}; + VB t1, t2, t3, t4, t5, t6, t7, t8; + const vector signed char lowMask = vec_splats((signed char)0xF); + const vector unsigned char v4 = vec_splats((unsigned char)0x4); + const vector signed char v8 = vec_splats((signed char)0x8); + aoffset = const_cast(a); + vecOffset = vec; + vector unsigned char swiz1 = {0, 1, 2, 3, 4, 5, 6, 7, 16, 17, 18, 19, 20, 21, 22, 23}; + vector unsigned char swiz2 = {8, 9, 10, 11, 12, 13, 14, 15, 24, 25, 26, 27, 28, 29, 30, 31}; + vector unsigned char swiz3 = {0, 1, 2, 3, 8, 9, 10, 11, 16, 17, 18, 19, 24, 25, 26, 27}; + vector unsigned char swiz4 = {4, 5, 6, 7, 12, 13, 14, 15, 20, 21, 22, 23, 28, 29, 30, 31}; + vector signed int vsum = {0}; + vector signed int vsum2 = {0}; + + j = (rows >> 3); + if (j > 0) { + do { + aoffset1 = aoffset; + aoffset2 = aoffset1 + lda; + aoffset3 = aoffset2 + lda; + aoffset4 = aoffset3 + lda; + aoffset5 = aoffset4 + lda; + aoffset6 = aoffset5 + lda; + aoffset7 = aoffset6 + lda; + aoffset8 = aoffset7 + lda; + aoffset += 8 * lda; + + i = (cols >> 2); + if (i > 0) { + do { + c1[1] = reinterpret_cast(vec_xl(0, aoffset1->qs)); + c2[1] = reinterpret_cast(vec_xl(0, aoffset2->qs)); + c3[1] = reinterpret_cast(vec_xl(0, aoffset3->qs)); + c4[1] = reinterpret_cast(vec_xl(0, aoffset4->qs)); + c5[1] = reinterpret_cast(vec_xl(0, aoffset5->qs)); + c6[1] = reinterpret_cast(vec_xl(0, aoffset6->qs)); + c7[1] = reinterpret_cast(vec_xl(0, aoffset7->qs)); + c8[1] = reinterpret_cast(vec_xl(0, aoffset8->qs)); + + c1[0] = vec_and(c1[1], lowMask); + c1[1] = vec_sr(c1[1], v4); + c1[0] = vec_sub(c1[0], v8); + c1[1] = vec_sub(c1[1], v8); + vsum = vec_sum4s(c1[0], vsum); + vsum2 = vec_sum4s(c1[1], vsum2); + vsum = vec_add(vsum, vsum2); + comparray[0] = vsum[0] + vsum[1] + vsum[2] + vsum[3]; + vsum = vec_splats(0); + vsum2 = vec_splats(0); + + c2[0] = vec_and(c2[1], lowMask); + c2[1] = vec_sr(c2[1], v4); + c2[0] = vec_sub(c2[0], v8); + c2[1] = vec_sub(c2[1], v8); + vsum = vec_sum4s(c2[0], vsum); + vsum2 = vec_sum4s(c2[1], vsum2); + vsum = vec_add(vsum, vsum2); + comparray[1] = vsum[0] + vsum[1] + vsum[2] + vsum[3]; + vsum = vec_splats(0); + vsum2 = vec_splats(0); + + c3[0] = vec_and(c3[1], lowMask); + c3[1] = vec_sr(c3[1], v4); + c3[0] = vec_sub(c3[0], v8); + c3[1] = vec_sub(c3[1], v8); + vsum = vec_sum4s(c3[0], vsum); + vsum2 = vec_sum4s(c3[1], vsum2); + vsum = vec_add(vsum, vsum2); + comparray[2] = vsum[0] + vsum[1] + vsum[2] + vsum[3]; + vsum = vec_splats(0); + vsum2 = vec_splats(0); + + c4[0] = vec_and(c4[1], lowMask); + c4[1] = vec_sr(c4[1], v4); + c4[0] = vec_sub(c4[0], v8); + c4[1] = vec_sub(c4[1], v8); + vsum = vec_sum4s(c4[0], vsum); + vsum2 = vec_sum4s(c4[1], vsum2); + vsum = vec_add(vsum, vsum2); + comparray[3] = vsum[0] + vsum[1] + vsum[2] + vsum[3]; + vsum = vec_splats(0); + vsum2 = vec_splats(0); + + c5[0] = vec_and(c5[1], lowMask); + c5[1] = vec_sr(c5[1], v4); + c5[0] = vec_sub(c5[0], v8); + c5[1] = vec_sub(c5[1], v8); + vsum = vec_sum4s(c5[0], vsum); + vsum2 = vec_sum4s(c5[1], vsum2); + vsum = vec_add(vsum, vsum2); + comparray[4] = vsum[0] + vsum[1] + vsum[2] + vsum[3]; + vsum = vec_splats(0); + vsum2 = vec_splats(0); + + c6[0] = vec_and(c6[1], lowMask); + c6[1] = vec_sr(c6[1], v4); + c6[0] = vec_sub(c6[0], v8); + c6[1] = vec_sub(c6[1], v8); + vsum = vec_sum4s(c6[0], vsum); + vsum2 = vec_sum4s(c6[1], vsum2); + vsum = vec_add(vsum, vsum2); + comparray[5] = vsum[0] + vsum[1] + vsum[2] + vsum[3]; + vsum = vec_splats(0); + vsum2 = vec_splats(0); + + c7[0] = vec_and(c7[1], lowMask); + c7[1] = vec_sr(c7[1], v4); + c7[0] = vec_sub(c7[0], v8); + c7[1] = vec_sub(c7[1], v8); + vsum = vec_sum4s(c7[0], vsum); + vsum2 = vec_sum4s(c7[1], vsum2); + vsum = vec_add(vsum, vsum2); + comparray[6] = vsum[0] + vsum[1] + vsum[2] + vsum[3]; + vsum = vec_splats(0); + vsum2 = vec_splats(0); + + c8[0] = vec_and(c8[1], lowMask); + c8[1] = vec_sr(c8[1], v4); + c8[0] = vec_sub(c8[0], v8); + c8[1] = vec_sub(c8[1], v8); + vsum = vec_sum4s(c8[0], vsum); + vsum2 = vec_sum4s(c8[1], vsum2); + vsum = vec_add(vsum, vsum2); + comparray[7] = vsum[0] + vsum[1] + vsum[2] + vsum[3]; + vsum = vec_splats(0); + vsum2 = vec_splats(0); + + t1 = vec_perm(c1[0], c2[0], swiz1); + t2 = vec_perm(c1[0], c2[0], swiz2); + t3 = vec_perm(c3[0], c4[0], swiz1); + t4 = vec_perm(c3[0], c4[0], swiz2); + t5 = vec_perm(t1, t3, swiz3); + t6 = vec_perm(t1, t3, swiz4); + t7 = vec_perm(t2, t4, swiz3); + t8 = vec_perm(t2, t4, swiz4); + vec_xst(t5, 0, vecOffset); + vec_xst(t6, 0, vecOffset+16); + vec_xst(t7, 0, vecOffset+32); + vec_xst(t8, 0, vecOffset+48); + + t1 = vec_perm(c1[1], c2[1], swiz1); + t2 = vec_perm(c1[1], c2[1], swiz2); + t3 = vec_perm(c3[1], c4[1], swiz1); + t4 = vec_perm(c3[1], c4[1], swiz2); + t5 = vec_perm(t1, t3, swiz3); + t6 = vec_perm(t1, t3, swiz4); + t7 = vec_perm(t2, t4, swiz3); + t8 = vec_perm(t2, t4, swiz4); + vec_xst(t5, 0, vecOffset+64); + vec_xst(t6, 0, vecOffset+80); + vec_xst(t7, 0, vecOffset+96); + vec_xst(t8, 0, vecOffset+112); + + t1 = vec_perm(c5[0], c6[0], swiz1); + t2 = vec_perm(c5[0], c6[0], swiz2); + t3 = vec_perm(c7[0], c8[0], swiz1); + t4 = vec_perm(c7[0], c8[0], swiz2); + t5 = vec_perm(t1, t3, swiz3); + t6 = vec_perm(t1, t3, swiz4); + t7 = vec_perm(t2, t4, swiz3); + t8 = vec_perm(t2, t4, swiz4); + vec_xst(t5, 0, vecOffset+128); + vec_xst(t6, 0, vecOffset+144); + vec_xst(t7, 0, vecOffset+160); + vec_xst(t8, 0, vecOffset+176); + + t1 = vec_perm(c5[1], c6[1], swiz1); + t2 = vec_perm(c5[1], c6[1], swiz2); + t3 = vec_perm(c7[1], c8[1], swiz1); + t4 = vec_perm(c7[1], c8[1], swiz2); + t5 = vec_perm(t1, t3, swiz3); + t6 = vec_perm(t1, t3, swiz4); + t7 = vec_perm(t2, t4, swiz3); + t8 = vec_perm(t2, t4, swiz4); + vec_xst(t5, 0, vecOffset+192); + vec_xst(t6, 0, vecOffset+208); + vec_xst(t7, 0, vecOffset+224); + vec_xst(t8, 0, vecOffset+240); + + aoffset1 += lda; + aoffset2 += lda; + aoffset3 += lda; + aoffset4 += lda; + aoffset5 += lda; + aoffset6 += lda; + aoffset7 += lda; + aoffset8 += lda; + vecOffset += 256; + i--; + } while (i > 0); + } + j--; + } while (j > 0); + } + + if (rows & 4) { + aoffset1 = aoffset; + aoffset2 = aoffset1 + lda; + aoffset3 = aoffset2 + lda; + aoffset4 = aoffset3 + lda; + aoffset += 4 * lda; + + i = (cols >> 2); + if (i > 0) { + do { + c1[1] = reinterpret_cast(vec_xl(0, aoffset1->qs)); + c2[1] = reinterpret_cast(vec_xl(0, aoffset2->qs)); + c3[1] = reinterpret_cast(vec_xl(0, aoffset3->qs)); + c4[1] = reinterpret_cast(vec_xl(0, aoffset4->qs)); + + c1[0] = vec_and(c1[1], lowMask); + c1[1] = vec_sr(c1[1], v4); + c1[0] = vec_sub(c1[0], v8); + c1[1] = vec_sub(c1[1], v8); + vsum = vec_sum4s(c1[0], vsum); + vsum2 = vec_sum4s(c1[1], vsum2); + vsum = vec_add(vsum, vsum2); + comparray[0] = vsum[0] + vsum[1] + vsum[2] + vsum[3]; + vsum = vec_splats(0); + vsum2 = vec_splats(0); + + c2[0] = vec_and(c2[1], lowMask); + c2[1] = vec_sr(c2[1], v4); + c2[0] = vec_sub(c2[0], v8); + c2[1] = vec_sub(c2[1], v8); + vsum = vec_sum4s(c2[0], vsum); + vsum2 = vec_sum4s(c2[1], vsum2); + vsum = vec_add(vsum, vsum2); + comparray[1] = vsum[0] + vsum[1] + vsum[2] + vsum[3]; + vsum = vec_splats(0); + vsum2 = vec_splats(0); + + c3[0] = vec_and(c3[1], lowMask); + c3[1] = vec_sr(c3[1], v4); + c3[0] = vec_sub(c3[0], v8); + c3[1] = vec_sub(c3[1], v8); + vsum = vec_sum4s(c3[0], vsum); + vsum2 = vec_sum4s(c3[1], vsum2); + vsum = vec_add(vsum, vsum2); + comparray[2] = vsum[0] + vsum[1] + vsum[2] + vsum[3]; + vsum = vec_splats(0); + vsum2 = vec_splats(0); + + c4[0] = vec_and(c4[1], lowMask); + c4[1] = vec_sr(c4[1], v4); + c4[0] = vec_sub(c4[0], v8); + c4[1] = vec_sub(c4[1], v8); + vsum = vec_sum4s(c4[0], vsum); + vsum2 = vec_sum4s(c4[1], vsum2); + vsum = vec_add(vsum, vsum2); + comparray[3] = vsum[0] + vsum[1] + vsum[2] + vsum[3]; + vsum = vec_splats(0); + vsum2 = vec_splats( 0); + + t1 = vec_perm(c1[0], c2[0], swiz1); + t2 = vec_perm(c1[0], c2[0], swiz2); + t3 = vec_perm(c3[0], c4[0], swiz1); + t4 = vec_perm(c3[0], c4[0], swiz2); + t5 = vec_perm(t1, t3, swiz3); + t6 = vec_perm(t1, t3, swiz4); + t7 = vec_perm(t2, t4, swiz3); + t8 = vec_perm(t2, t4, swiz4); + vec_xst(t5, 0, vecOffset); + vec_xst(t6, 0, vecOffset+16); + vec_xst(t7, 0, vecOffset+32); + vec_xst(t8, 0, vecOffset+48); + + t1 = vec_perm(c1[1], c2[1], swiz1); + t2 = vec_perm(c1[1], c2[1], swiz2); + t3 = vec_perm(c3[1], c4[1], swiz1); + t4 = vec_perm(c3[1], c4[1], swiz2); + t5 = vec_perm(t1, t3, swiz3); + t6 = vec_perm(t1, t3, swiz4); + t7 = vec_perm(t2, t4, swiz3); + t8 = vec_perm(t2, t4, swiz4); + vec_xst(t5, 0, vecOffset+64); + vec_xst(t6, 0, vecOffset+80); + vec_xst(t7, 0, vecOffset+96); + vec_xst(t8, 0, vecOffset+112); + + aoffset1 += lda; + aoffset2 += lda; + aoffset3 += lda; + aoffset4 += lda; + vecOffset += 128; + i--; + } while (i > 0); + } + } + + if (rows & 3) { + aoffset1 = aoffset; + aoffset2 = aoffset1 + lda; + aoffset3 = aoffset2 + lda; + i = (cols >> 2); + if (i > 0) { + do { + switch(rows) { + case 3: c3[1] = reinterpret_cast(vec_xl(0, aoffset3->qs)); + case 2: c2[1] = reinterpret_cast(vec_xl(0, aoffset2->qs)); + case 1: c1[1] = reinterpret_cast(vec_xl(0, aoffset1->qs)); + break; + } + c1[0] = vec_and(c1[1], lowMask); + c1[1] = vec_sr(c1[1], v4); + c1[0] = vec_sub(c1[0], v8); + c1[1] = vec_sub(c1[1], v8); + vsum = vec_sum4s(c1[0], vsum); + vsum2 = vec_sum4s(c1[1], vsum2); + vsum = vec_add(vsum, vsum2); + comparray[0] = vsum[0] + vsum[1] + vsum[2] + vsum[3]; + vsum = vec_splats(0); + vsum2 = vec_splats(0); + + c2[0] = vec_and(c2[1], lowMask); + c2[1] = vec_sr(c2[1], v4); + c2[0] = vec_sub(c2[0], v8); + c2[1] = vec_sub(c2[1], v8); + vsum = vec_sum4s(c2[0], vsum); + vsum2 = vec_sum4s(c2[1], vsum2); + vsum = vec_add(vsum, vsum2); + comparray[1] = vsum[0] + vsum[1] + vsum[2] + vsum[3]; + vsum = vec_splats(0); + vsum2 = vec_splats(0); + + c3[0] = vec_and(c3[1], lowMask); + c3[1] = vec_sr(c3[1], v4); + c3[0] = vec_sub(c3[0], v8); + c3[1] = vec_sub(c3[1], v8); + vsum = vec_sum4s(c3[0], vsum); + vsum2 = vec_sum4s(c3[1], vsum2); + vsum = vec_add(vsum, vsum2); + comparray[2] = vsum[0] + vsum[1] + vsum[2] + vsum[3]; + vsum = vec_splats(0); + vsum2 = vec_splats(0); + + c4[0] = vec_and(c4[1], lowMask); + c4[1] = vec_sr(c4[1], v4); + c4[0] = vec_sub(c4[0], v8); + c4[1] = vec_sub(c4[1], v8); + vsum = vec_sum4s(c4[0], vsum); + vsum2 = vec_sum4s(c4[1], vsum2); + vsum = vec_add(vsum, vsum2); + comparray[3] = vsum[0] + vsum[1] + vsum[2] + vsum[3]; + vsum = vec_splats(0); + vsum2 = vec_splats(0); + + t1 = vec_perm(c1[0], c2[0], swiz1); + t2 = vec_perm(c1[0], c2[0], swiz2); + t3 = vec_perm(c3[0], c4[0], swiz1); + t4 = vec_perm(c3[0], c4[0], swiz2); + t5 = vec_perm(t1, t3, swiz3); + t6 = vec_perm(t1, t3, swiz4); + t7 = vec_perm(t2, t4, swiz3); + t8 = vec_perm(t2, t4, swiz4); + vec_xst(t5, 0, vecOffset); + vec_xst(t6, 0, vecOffset+16); + vec_xst(t7, 0, vecOffset+32); + vec_xst(t8, 0, vecOffset+48); + + t1 = vec_perm(c1[1], c2[1], swiz1); + t2 = vec_perm(c1[1], c2[1], swiz2); + t3 = vec_perm(c3[1], c4[1], swiz1); + t4 = vec_perm(c3[1], c4[1], swiz2); + t5 = vec_perm(t1, t3, swiz3); + t6 = vec_perm(t1, t3, swiz4); + t7 = vec_perm(t2, t4, swiz3); + t8 = vec_perm(t2, t4, swiz4); + vec_xst(t5, 0, vecOffset+64); + vec_xst(t6, 0, vecOffset+80); + vec_xst(t7, 0, vecOffset+96); + vec_xst(t8, 0, vecOffset+112); + aoffset1 += lda; + aoffset2 += lda; + aoffset3 += lda; + vecOffset += 128; + i--; + } while(i > 0); + } + } + } + + template + void packNormal(const TB* a, int64_t lda, int rows, int cols, VA* vec, bool flip) { + int64_t i, j; + TB *aoffset = NULL; + VA *vecOffset = NULL; + TB *aoffset1 = NULL, *aoffset2 = NULL, *aoffset3 = NULL, *aoffset4 = NULL; + TB *aoffset5 = NULL, *aoffset6 = NULL, *aoffset7 = NULL, *aoffset8 = NULL; __vector_pair C1, C2, C3, C4, C5, C6, C7, C8; VB c1[2] = {0}, c2[2] = {0}, c3[2] = {0}, c4[2]={0}; VB c5[2] = {0}, c6[2] = {0}, c7[2] = {0}, c8[2]={0}; @@ -1110,24 +1502,24 @@ class tinyBLAS_Q0_PPC { vector unsigned char swiz3 = {0, 1, 2, 3, 8, 9, 10, 11, 16, 17, 18, 19, 24, 25, 26, 27}; vector unsigned char swiz4 = {4, 5, 6, 7, 12, 13, 14, 15, 20, 21, 22, 23, 28, 29, 30, 31}; - aoffset = const_cast(a); + aoffset = const_cast(a); vecOffset = vec; j = (rows >> 3); if (j > 0) { do { - aoffset1 = aoffset; - aoffset2 = aoffset1 + lda; - aoffset3 = aoffset2 + lda; - aoffset4 = aoffset3 + lda; - aoffset5 = aoffset4 + lda; - aoffset6 = aoffset5 + lda; - aoffset7 = aoffset6 + lda; - aoffset8 = aoffset7 + lda; - aoffset += 8 * lda; + aoffset1 = aoffset; + aoffset2 = aoffset1 + lda; + aoffset3 = aoffset2 + lda; + aoffset4 = aoffset3 + lda; + aoffset5 = aoffset4 + lda; + aoffset6 = aoffset5 + lda; + aoffset7 = aoffset6 + lda; + aoffset8 = aoffset7 + lda; + aoffset += 8 * lda; - i = (cols >> 3); - if (i > 0) { - do { + i = (cols >> 3); + if (i > 0) { + do { C1 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset1->qs); C2 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset2->qs); C3 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset3->qs); @@ -1155,10 +1547,10 @@ class tinyBLAS_Q0_PPC { t7 = vec_perm(t2, t4, swiz3); t8 = vec_perm(t2, t4, swiz4); if (flip == true) { - t5 = vec_xor(t5, xor_vector); - t6 = vec_xor(t6, xor_vector); - t7 = vec_xor(t7, xor_vector); - t8 = vec_xor(t8, xor_vector); + t5 = vec_xor(t5, xor_vector); + t6 = vec_xor(t6, xor_vector); + t7 = vec_xor(t7, xor_vector); + t8 = vec_xor(t8, xor_vector); } vec_xst(t5, 0, vecOffset); vec_xst(t6, 0, vecOffset+16); @@ -1174,10 +1566,10 @@ class tinyBLAS_Q0_PPC { t7 = vec_perm(t2, t4, swiz3); t8 = vec_perm(t2, t4, swiz4); if (flip == true) { - t5 = vec_xor(t5, xor_vector); - t6 = vec_xor(t6, xor_vector); - t7 = vec_xor(t7, xor_vector); - t8 = vec_xor(t8, xor_vector); + t5 = vec_xor(t5, xor_vector); + t6 = vec_xor(t6, xor_vector); + t7 = vec_xor(t7, xor_vector); + t8 = vec_xor(t8, xor_vector); } vec_xst(t5, 0, vecOffset+64); vec_xst(t6, 0, vecOffset+80); @@ -1193,10 +1585,10 @@ class tinyBLAS_Q0_PPC { t7 = vec_perm(t2, t4, swiz3); t8 = vec_perm(t2, t4, swiz4); if (flip == true) { - t5 = vec_xor(t5, xor_vector); - t6 = vec_xor(t6, xor_vector); - t7 = vec_xor(t7, xor_vector); - t8 = vec_xor(t8, xor_vector); + t5 = vec_xor(t5, xor_vector); + t6 = vec_xor(t6, xor_vector); + t7 = vec_xor(t7, xor_vector); + t8 = vec_xor(t8, xor_vector); } vec_xst(t5, 0, vecOffset+128); vec_xst(t6, 0, vecOffset+144); @@ -1212,10 +1604,10 @@ class tinyBLAS_Q0_PPC { t7 = vec_perm(t2, t4, swiz3); t8 = vec_perm(t2, t4, swiz4); if (flip == true) { - t5 = vec_xor(t5, xor_vector); - t6 = vec_xor(t6, xor_vector); - t7 = vec_xor(t7, xor_vector); - t8 = vec_xor(t8, xor_vector); + t5 = vec_xor(t5, xor_vector); + t6 = vec_xor(t6, xor_vector); + t7 = vec_xor(t7, xor_vector); + t8 = vec_xor(t8, xor_vector); } vec_xst(t5, 0, vecOffset+192); vec_xst(t6, 0, vecOffset+208); @@ -1239,11 +1631,11 @@ class tinyBLAS_Q0_PPC { } if (rows & 4) { - aoffset1 = aoffset; - aoffset2 = aoffset1 + lda; - aoffset3 = aoffset2 + lda; - aoffset4 = aoffset3 + lda; - aoffset += 4 * lda; + aoffset1 = aoffset; + aoffset2 = aoffset1 + lda; + aoffset3 = aoffset2 + lda; + aoffset4 = aoffset3 + lda; + aoffset += 4 * lda; i = (cols >> 3); if (i > 0) { @@ -1310,7 +1702,7 @@ class tinyBLAS_Q0_PPC { aoffset2 = aoffset1 + lda; aoffset3 = aoffset2 + lda; i = (cols >> 3); - if (i > 0) { + if (i > 0) { do { switch(rows) { case 3: C3 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset3->qs); @@ -1526,13 +1918,18 @@ class tinyBLAS_Q0_PPC { void KERNEL_4x8(int64_t ii, int64_t jj) { vec_t vec_A[8], vec_B[16] = {0}; acc_t acc_0, acc_1; - std::array comparray; + std::array comparray {}; vector float fin_res[8] = {0}; vector float vs[8] = {0}; + bool isAblock_q4 = std::is_same_v; for (int l = 0; l < k; l++) { __builtin_mma_xxsetaccz(&acc_0); __builtin_mma_xxsetaccz(&acc_1); - packNormal((A+(ii*lda)+l), lda, 4, 8, (int8_t*)vec_A, false); + if (std::is_same_v) { + packNormalInt4((A+(ii*lda)+l), lda, 4, 4, (int8_t*)vec_A, comparray); + } else { + packNormal((const TB*)(A+(ii*lda)+l), lda, 4, 8, (int8_t*)vec_A, false); + } packNormal((B+(jj*ldb)+l), ldb, 8, 8, (uint8_t*)vec_B, true); for(int x = 0; x < 8; x++) { __builtin_mma_xvi8ger4pp(&acc_0, vec_A[x], vec_B[x]); @@ -1544,15 +1941,17 @@ class tinyBLAS_Q0_PPC { *((float*)&vs[I+4]+J) = (unhalf((A+((ii+I)*lda)+l)->d) * unhalf((B+((jj+J+4)*ldb)+l)->d)); } } - auto aoffset = A+(ii*lda)+l; - for (int i = 0; i < 4; i++) { - comparray[i] = 0; - int ca = 0; - const int8_t *at = aoffset->qs; - for (int j = 0; j < 32; j++) - ca += (int)*at++; - comparray[i] = ca; - aoffset += lda; + if (!isAblock_q4) { + auto aoffset = A+(ii*lda)+l; + for (int i = 0; i < 4; i++) { + comparray[i] = 0; + int ca = 0; + auto *at = aoffset->qs; + for (int j = 0; j < 32; j++) + ca += (int)*at++; + comparray[i] = ca; + aoffset += lda; + } } compute<4>(&acc_0, 0, 0, comparray, vs, fin_res); compute<4>(&acc_1, 0, 4, comparray, vs, fin_res); @@ -1564,13 +1963,18 @@ class tinyBLAS_Q0_PPC { void KERNEL_8x4(int64_t ii, int64_t jj) { vec_t vec_A[16], vec_B[8] = {0}; acc_t acc_0, acc_1; - std::array comparray; + std::array comparray {}; vector float fin_res[8] = {0}; vector float vs[8] = {0}; + bool isAblock_q4 = std::is_same_v; for (int l = 0; l < k; l++) { __builtin_mma_xxsetaccz(&acc_0); __builtin_mma_xxsetaccz(&acc_1); - packNormal((A+(ii*lda)+l), lda, 8, 8, (int8_t*)vec_A, false); + if (std::is_same_v) { + packNormalInt4((A+(ii*lda)+l), lda, 8, 4, (int8_t*)vec_A, comparray); + } else { + packNormal((const TB*)(A+(ii*lda)+l), lda, 8, 8, (int8_t*)vec_A, false); + } packNormal((B+(jj*ldb)+l), ldb, 4, 8, (uint8_t*)vec_B, true); for(int x = 0; x < 8; x++) { __builtin_mma_xvi8ger4pp(&acc_0, vec_A[x], vec_B[x]); @@ -1581,15 +1985,17 @@ class tinyBLAS_Q0_PPC { *((float*)&vs[I]+J) = (unhalf((A+((ii+I)*lda)+l)->d) * unhalf((B+((jj+J)*ldb)+l)->d)); } } - auto aoffset = A+(ii*lda)+l; - for (int i = 0; i < 8; i++) { - comparray[i] = 0; - int ca = 0; - const int8_t *at = aoffset->qs; - for (int j = 0; j < 32; j++) - ca += (int)*at++; - comparray[i] = ca; - aoffset += lda; + if (!isAblock_q4) { + auto aoffset = A+(ii*lda)+l; + for (int i = 0; i < 8; i++) { + comparray[i] = 0; + int ca = 0; + auto *at = aoffset->qs; + for (int j = 0; j < 32; j++) + ca += (int)*at++; + comparray[i] = ca; + aoffset += lda; + } } compute<8>(&acc_0, 0, 0, comparray, vs, fin_res); compute<8>(&acc_1, 4, 4, comparray, vs, fin_res); @@ -1601,15 +2007,20 @@ class tinyBLAS_Q0_PPC { void KERNEL_8x8(int64_t ii, int64_t jj) { vec_t vec_A[16], vec_B[16] = {0}; acc_t acc_0, acc_1, acc_2, acc_3; - std::array comparray; + std::array comparray {}; vector float fin_res[16] = {0}; vector float vs[16] = {0}; + bool isAblock_q4 = std::is_same_v; for (int l = 0; l < k; l++) { __builtin_mma_xxsetaccz(&acc_0); __builtin_mma_xxsetaccz(&acc_1); __builtin_mma_xxsetaccz(&acc_2); __builtin_mma_xxsetaccz(&acc_3); - packNormal((A+(ii*lda)+l), lda, 8, 8, (int8_t*)vec_A, false); + if (std::is_same_v) { + packNormalInt4((A+(ii*lda)+l), lda, 8, 4, (int8_t*)vec_A, comparray); + } else { + packNormal((const TB*)(A+(ii*lda)+l), lda, 8, 8, (int8_t*)vec_A, false); + } packNormal((B+(jj*ldb)+l), ldb, 8, 8, (uint8_t*)vec_B, true); for(int x = 0; x < 8; x++) { __builtin_mma_xvi8ger4pp(&acc_0, vec_A[x], vec_B[x]); @@ -1623,15 +2034,17 @@ class tinyBLAS_Q0_PPC { *((float*)&vs[I+8]+J) = (unhalf((A+((ii+I)*lda)+l)->d) * unhalf((B+((jj+J+4)*ldb)+l)->d)); } } - auto aoffset = A+(ii*lda)+l; - for (int i = 0; i < 8; i++) { - comparray[i] = 0; - int ca = 0; - const int8_t *at = aoffset->qs; - for (int j = 0; j < 32; j++) - ca += (int)*at++; - comparray[i] = ca; - aoffset += lda; + if (!isAblock_q4) { + auto aoffset = A+(ii*lda)+l; + for (int i = 0; i < 8; i++) { + comparray[i] = 0; + int ca = 0; + auto *at = aoffset->qs; + for (int j = 0; j < 32; j++) + ca += (int)*at++; + comparray[i] = ca; + aoffset += lda; + } } compute<8>(&acc_0, 0, 0, comparray, vs, fin_res); compute<8>(&acc_1, 4, 4, comparray, vs, fin_res); @@ -1652,16 +2065,17 @@ class tinyBLAS_Q0_PPC { int64_t duty = (tiles + nth - 1) / nth; int64_t start = duty * ith; int64_t end = start + duty; - vec_t vec_A[8], vec_B[8] = {0}; + vec_t vec_A[8] = {0}, vec_B[8] = {0}; vector signed int vec_C[4]; acc_t acc_0; + bool isAblock_q4 = std::is_same_v; if (end > tiles) end = tiles; for (int64_t job = start; job < end; ++job) { int64_t ii = m0 + job / xtiles * RM; int64_t jj = n0 + job % xtiles * RN; - std::array comparray; + std::array comparray{}; vector float res[4] = {0}; vector float fin_res[4] = {0}; vector float vs[4] = {0}; @@ -1672,7 +2086,11 @@ class tinyBLAS_Q0_PPC { __builtin_prefetch((A+(ii*lda)+(l+1))->qs, 0, 1); // prefetch one loop ahead __builtin_prefetch((B+(jj*ldb)+(l+1))->qs, 0, 1); // prefetch one loop ahead __builtin_mma_xxsetaccz(&acc_0); - packNormal((A+(ii*lda)+l), lda, RM, 8, (int8_t*)vec_A, false); + if (isAblock_q4) { + packNormalInt4((A+(ii*lda)+l), lda, RM, 4, (int8_t*)vec_A, comparray); + } else { + packNormal((const TB*)(A+(ii*lda)+l), lda, RM, 8, (int8_t*)vec_A, false); + } packNormal((B+(jj*ldb)+l), ldb, RN, 8, (uint8_t*)vec_B, true); for(int x = 0; x < 8; x+=4) { __builtin_mma_xvi8ger4pp(&acc_0, vec_A[x], vec_B[x]); @@ -1686,17 +2104,18 @@ class tinyBLAS_Q0_PPC { } } __builtin_mma_disassemble_acc(vec_C, &acc_0); - auto aoffset = A+(ii*lda)+l; - for (int i = 0; i < RM; i++) { - comparray[i] = 0; - int ca = 0; - const int8_t *at = aoffset->qs; - for (int j = 0; j < 32; j++) - ca += (int)*at++; - comparray[i] = ca; - aoffset += lda; + if (!isAblock_q4) { + auto aoffset = A+(ii*lda)+l; + for (int i = 0; i < RM; i++) { + comparray[i] = 0; + int ca = 0; + auto *at = aoffset->qs; + for (int j = 0; j < 32; j++) + ca += (int)*at++; + comparray[i] = ca; + aoffset += lda; + } } - for (int i = 0; i < RM; i++) { CA[i] = vec_splats((float)(((double)comparray[i]) * -128.0)); res[i] = vec_add(vec_ctf(vec_C[i], 0), CA[i]); @@ -2012,6 +2431,7 @@ class tinyBLAS_PPC { } } } + void KERNEL_4x4(int64_t ii, int64_t jj) { vec_t vec_A[4], vec_B[4], vec_C[4]; acc_t acc_0; @@ -2258,15 +2678,27 @@ class tinyBLAS_PPC { vec_t vec_C[4]; acc_t acc_0; __builtin_mma_xxsetaccz(&acc_0); - vec_t vec_A[4], vec_B[4]; + vec_t vec_A[4] {0}, vec_B[4] = {0}; for (int l=0; l= 4 && RM == 1) { + /* 'GEMV Forwarding' concept is used in first two conditional loops. + * when one of the matrix has a single row/column, the elements are + * broadcasted, instead of using packing routine to prepack the + * matrix elements. + */ + if (RM == 1) { TA* a = const_cast(A+(ii)*lda+l); - packTranspose(B+(jj*ldb)+l, ldb, 4, 4, (TA*)vec_B); + packTranspose(B+(jj*ldb)+l, ldb, RN, 4, (TA*)vec_B); vec_A[0] = (vec_t)vec_xl(0,a); vec_A[1] = (vec_t)vec_splats(*((TA*)&vec_A+1)); vec_A[2] = (vec_t)vec_splats(*((TA*)&vec_A+2)); vec_A[3] = (vec_t)vec_splats(*((TA*)&vec_A+3)); + } else if (RN == 1) { + packTranspose(A+(ii*lda)+l, lda, RM, 4, (TA*)vec_A); + TB* b = const_cast(B+(jj)*ldb+l); + vec_B[0] = (vec_t)vec_xl(0,b); + vec_B[1] = (vec_t)vec_splats(*((TB*)&vec_B+1)); + vec_B[2] = (vec_t)vec_splats(*((TB*)&vec_B+2)); + vec_B[3] = (vec_t)vec_splats(*((TB*)&vec_B+3)); } else { packTranspose(A+(ii*lda)+l, lda, RM, 4, (TA*)vec_A); packTranspose(B+(jj*ldb)+l, ldb, RN, 4, (TA*)vec_B); @@ -2370,8 +2802,10 @@ bool llamafile_sgemm(const struct ggml_compute_params * params, int64_t m, int64 assert(params->ith < params->nth); // only enable sgemm for prompt processing +#if !defined(__MMA__) if (n < 2) return false; +#endif if (Ctype != GGML_TYPE_F32) return false; @@ -2502,8 +2936,8 @@ bool llamafile_sgemm(const struct ggml_compute_params * params, int64_t m, int64 params->ith, params->nth}; tb.matmul(m, n); return true; - #elif defined(__MMA__) + //TO-DO: Remove this condition once gemv forwarding is enabled. if (n < 8 && n != 4) return false; if (m < 8 && m != 4) @@ -2515,7 +2949,6 @@ bool llamafile_sgemm(const struct ggml_compute_params * params, int64_t m, int64 params->ith, params->nth}; tb.matmul(m, n); return true; - #else return false; #endif @@ -2540,6 +2973,19 @@ bool llamafile_sgemm(const struct ggml_compute_params * params, int64_t m, int64 params->ith, params->nth}; tb.matmul(m, n); return true; +#elif defined(__MMA__) + //TO-DO: Remove this condition once gemv forwarding is enabled. + if (n < 8 && n != 4) + return false; + if (m < 8 && m != 4) + return false; + tinyBLAS_Q0_PPC tb{ + k, (const block_q4_0 *)A, lda, + (const block_q8_0 *)B, ldb, + (float *)C, ldc, + params->ith, params->nth}; + tb.matmul(m, n); + return true; #else return false; #endif diff --git a/ggml/src/ggml-cpu/ops.cpp b/ggml/src/ggml-cpu/ops.cpp new file mode 100644 index 000000000..6050147be --- /dev/null +++ b/ggml/src/ggml-cpu/ops.cpp @@ -0,0 +1,8637 @@ +#include "ops.h" + +#include "ggml-cpu.h" +#include "ggml-impl.h" +#include "binary-ops.h" +#include "unary-ops.h" +#include "vec.h" + +#include + +#if defined(_MSC_VER) +// disable "possible loss of data" to avoid hundreds of casts +// we should just be careful :) +#pragma warning(disable: 4244 4267) + +// disable POSIX deprecation warnings +// these functions are never going away, anyway +#pragma warning(disable: 4996) + +// unreachable code because of multiple instances of code after GGML_ABORT +#pragma warning(disable: 4702) +#endif + +// ggml_compute_forward_dup + +static void ggml_compute_forward_dup_same_cont( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + + GGML_ASSERT(ggml_nelements(dst) == ggml_nelements(src0)); + GGML_ASSERT(ggml_is_contiguous(dst) && ggml_is_contiguous(src0)); + GGML_ASSERT(src0->type == dst->type); + + const size_t nb0 = ggml_type_size(src0->type); + + const int ith = params->ith; // thread index + const int nth = params->nth; // number of threads + + // parallelize by blocks + const int nk = ggml_nelements(src0)/ggml_blck_size(src0->type); + const int dr = (nk + nth - 1) / nth; + const int k0 = dr * ith; + const int k1 = MIN(k0 + dr, nk); + + if (k0 < k1) { + memcpy( + ((char *) dst->data + k0*nb0), + ((char *) src0->data + k0*nb0), + (k1 - k0) * nb0); + } +} + +static void ggml_compute_forward_dup_f16( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + + GGML_ASSERT(ggml_nelements(dst) == ggml_nelements(src0)); + + GGML_TENSOR_UNARY_OP_LOCALS + + const int ith = params->ith; // thread index + const int nth = params->nth; // number of threads + + // parallelize by rows + const int nr = ne01; + // number of rows per thread + const int dr = (nr + nth - 1) / nth; + // row range for this thread + const int ir0 = dr * ith; + const int ir1 = MIN(ir0 + dr, nr); + + if (src0->type == dst->type && + ne00 == ne0 && + nb00 == ggml_type_size(src0->type) && nb0 == ggml_type_size(dst->type)) { + // copy by rows + const size_t rs = ne00*nb00; + for (int64_t i03 = 0; i03 < ne03; i03++) { + for (int64_t i02 = 0; i02 < ne02; i02++) { + for (int64_t i01 = ir0; i01 < ir1; i01++) { + memcpy( + ((char *) dst->data + i01*nb1 + i02*nb2 + i03*nb3), + ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03), + rs); + } + } + } + return; + } + + // TODO: add more special-case implementations for tensor shapes/strides that can benefit from memcpy + + if (ggml_is_contiguous(dst)) { + if (nb00 == sizeof(ggml_fp16_t)) { + if (dst->type == GGML_TYPE_F16) { + size_t id = 0; + const size_t rs = ne00 * nb00; + char * dst_ptr = (char *) dst->data; + + for (int i03 = 0; i03 < ne03; i03++) { + for (int i02 = 0; i02 < ne02; i02++) { + id += rs * ir0; + for (int i01 = ir0; i01 < ir1; i01++) { + const char * src0_ptr = (char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03; + memcpy(dst_ptr + id, src0_ptr, rs); + id += rs; + } + id += rs * (ne01 - ir1); + } + } + } else if (dst->type == GGML_TYPE_F32) { + size_t id = 0; + float * dst_ptr = (float *) dst->data; + + for (int i03 = 0; i03 < ne03; i03++) { + for (int i02 = 0; i02 < ne02; i02++) { + id += ne00 * ir0; + for (int i01 = ir0; i01 < ir1; i01++) { + const ggml_fp16_t * src0_ptr = (ggml_fp16_t *) ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03); + for (int i00 = 0; i00 < ne00; i00++) { + dst_ptr[id] = GGML_FP16_TO_FP32(src0_ptr[i00]); + id++; + } + } + id += ne00 * (ne01 - ir1); + } + } + } else if (ggml_get_type_traits_cpu(dst->type)->from_float) { + ggml_from_float_t const quantize_row_q = ggml_get_type_traits_cpu(dst->type)->from_float; + float * src0_f32 = (float *) params->wdata + (ne00 + CACHE_LINE_SIZE_F32) * ith; + + size_t id = 0; + size_t rs = nb0 * (ne00 / ggml_blck_size(dst->type)); + char * dst_ptr = (char *) dst->data; + + for (int i03 = 0; i03 < ne03; i03++) { + for (int i02 = 0; i02 < ne02; i02++) { + id += rs * ir0; + for (int i01 = ir0; i01 < ir1; i01++) { + const ggml_fp16_t * src0_ptr = (ggml_fp16_t *) ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03); + + for (int i00 = 0; i00 < ne00; i00++) { + src0_f32[i00] = GGML_FP16_TO_FP32(src0_ptr[i00]); + } + + quantize_row_q(src0_f32, dst_ptr + id, ne00); + id += rs; + } + id += rs * (ne01 - ir1); + } + } + } else { + GGML_ABORT("fatal error"); // TODO: implement + } + } else { + //printf("%s: this is not optimal - fix me\n", __func__); + + if (dst->type == GGML_TYPE_F32) { + size_t id = 0; + float * dst_ptr = (float *) dst->data; + + for (int i03 = 0; i03 < ne03; i03++) { + for (int i02 = 0; i02 < ne02; i02++) { + id += ne00 * ir0; + for (int i01 = ir0; i01 < ir1; i01++) { + for (int i00 = 0; i00 < ne00; i00++) { + const ggml_fp16_t * src0_ptr = (ggml_fp16_t *) ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03); + + dst_ptr[id] = GGML_FP16_TO_FP32(*src0_ptr); + id++; + } + } + id += ne00 * (ne01 - ir1); + } + } + } else if (dst->type == GGML_TYPE_F16) { + size_t id = 0; + ggml_fp16_t * dst_ptr = (ggml_fp16_t *) dst->data; + + for (int i03 = 0; i03 < ne03; i03++) { + for (int i02 = 0; i02 < ne02; i02++) { + id += ne00 * ir0; + for (int i01 = ir0; i01 < ir1; i01++) { + for (int i00 = 0; i00 < ne00; i00++) { + const ggml_fp16_t * src0_ptr = (ggml_fp16_t *) ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03); + + dst_ptr[id] = *src0_ptr; + id++; + } + } + id += ne00 * (ne01 - ir1); + } + } + } else { + GGML_ABORT("fatal error"); // TODO: implement + } + } + return; + } + + // dst counters + int64_t i10 = 0; + int64_t i11 = 0; + int64_t i12 = 0; + int64_t i13 = 0; + + if (dst->type == GGML_TYPE_F16) { + for (int64_t i03 = 0; i03 < ne03; i03++) { + for (int64_t i02 = 0; i02 < ne02; i02++) { + i10 += ne00 * ir0; + while (i10 >= ne0) { + i10 -= ne0; + if (++i11 == ne1) { + i11 = 0; + if (++i12 == ne2) { + i12 = 0; + if (++i13 == ne3) { + i13 = 0; + } + } + } + } + for (int64_t i01 = ir0; i01 < ir1; i01++) { + for (int64_t i00 = 0; i00 < ne00; i00++) { + const char * src0_ptr = ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03); + char * dst_ptr = ((char *) dst->data + i10*nb0 + i11*nb1 + i12*nb2 + i13*nb3); + + memcpy(dst_ptr, src0_ptr, sizeof(ggml_fp16_t)); + + if (++i10 == ne00) { + i10 = 0; + if (++i11 == ne01) { + i11 = 0; + if (++i12 == ne02) { + i12 = 0; + if (++i13 == ne03) { + i13 = 0; + } + } + } + } + } + } + i10 += ne00 * (ne01 - ir1); + while (i10 >= ne0) { + i10 -= ne0; + if (++i11 == ne1) { + i11 = 0; + if (++i12 == ne2) { + i12 = 0; + if (++i13 == ne3) { + i13 = 0; + } + } + } + } + } + } + } else if (dst->type == GGML_TYPE_F32) { + for (int64_t i03 = 0; i03 < ne03; i03++) { + for (int64_t i02 = 0; i02 < ne02; i02++) { + i10 += ne00 * ir0; + while (i10 >= ne0) { + i10 -= ne0; + if (++i11 == ne1) { + i11 = 0; + if (++i12 == ne2) { + i12 = 0; + if (++i13 == ne3) { + i13 = 0; + } + } + } + } + for (int64_t i01 = ir0; i01 < ir1; i01++) { + for (int64_t i00 = 0; i00 < ne00; i00++) { + const char * src0_ptr = ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03); + char * dst_ptr = ((char *) dst->data + i10*nb0 + i11*nb1 + i12*nb2 + i13*nb3); + + *(float *) dst_ptr = GGML_FP16_TO_FP32(*(const ggml_fp16_t *) src0_ptr); + + if (++i10 == ne0) { + i10 = 0; + if (++i11 == ne1) { + i11 = 0; + if (++i12 == ne2) { + i12 = 0; + if (++i13 == ne3) { + i13 = 0; + } + } + } + } + } + } + i10 += ne00 * (ne01 - ir1); + while (i10 >= ne0) { + i10 -= ne0; + if (++i11 == ne1) { + i11 = 0; + if (++i12 == ne2) { + i12 = 0; + if (++i13 == ne3) { + i13 = 0; + } + } + } + } + } + } + } else { + GGML_ABORT("fatal error"); // TODO: implement + } +} + +static void ggml_compute_forward_dup_bf16( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + + GGML_ASSERT(ggml_nelements(dst) == ggml_nelements(src0)); + + GGML_TENSOR_UNARY_OP_LOCALS + + const int ith = params->ith; // thread index + const int nth = params->nth; // number of threads + + // parallelize by rows + const int nr = ne01; + // number of rows per thread + const int dr = (nr + nth - 1) / nth; + // row range for this thread + const int ir0 = dr * ith; + const int ir1 = MIN(ir0 + dr, nr); + + if (src0->type == dst->type && + ne00 == ne0 && + nb00 == ggml_type_size(src0->type) && nb0 == ggml_type_size(dst->type)) { + // copy by rows + const size_t rs = ne00*nb00; + for (int64_t i03 = 0; i03 < ne03; i03++) { + for (int64_t i02 = 0; i02 < ne02; i02++) { + for (int64_t i01 = ir0; i01 < ir1; i01++) { + memcpy( + ((char *) dst->data + i01*nb1 + i02*nb2 + i03*nb3), + ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03), + rs); + } + } + } + return; + } + + // TODO: add more special-case implementations for tensor shapes/strides that can benefit from memcpy + + if (ggml_is_contiguous(dst)) { + if (nb00 == sizeof(ggml_bf16_t)) { + if (dst->type == GGML_TYPE_BF16) { + size_t id = 0; + const size_t rs = ne00 * nb00; + char * dst_ptr = (char *) dst->data; + + for (int i03 = 0; i03 < ne03; i03++) { + for (int i02 = 0; i02 < ne02; i02++) { + id += rs * ir0; + for (int i01 = ir0; i01 < ir1; i01++) { + const char * src0_ptr = (char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03; + memcpy(dst_ptr + id, src0_ptr, rs); + id += rs; + } + id += rs * (ne01 - ir1); + } + } + } else if (dst->type == GGML_TYPE_F16) { + size_t id = 0; + ggml_fp16_t * dst_ptr = (ggml_fp16_t *) dst->data; + + for (int i03 = 0; i03 < ne03; i03++) { + for (int i02 = 0; i02 < ne02; i02++) { + id += ne00 * ir0; + for (int i01 = ir0; i01 < ir1; i01++) { + const ggml_bf16_t * src0_ptr = (ggml_bf16_t *) ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03); + for (int i00 = 0; i00 < ne00; i00++) { + dst_ptr[id] = GGML_FP32_TO_FP16(GGML_BF16_TO_FP32(src0_ptr[i00])); + id++; + } + } + id += ne00 * (ne01 - ir1); + } + } + } else if (dst->type == GGML_TYPE_F32) { + size_t id = 0; + float * dst_ptr = (float *) dst->data; + + for (int i03 = 0; i03 < ne03; i03++) { + for (int i02 = 0; i02 < ne02; i02++) { + id += ne00 * ir0; + for (int i01 = ir0; i01 < ir1; i01++) { + const ggml_bf16_t * src0_ptr = (ggml_bf16_t *) ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03); + for (int i00 = 0; i00 < ne00; i00++) { + dst_ptr[id] = GGML_BF16_TO_FP32(src0_ptr[i00]); + id++; + } + } + id += ne00 * (ne01 - ir1); + } + } + } else if (ggml_get_type_traits_cpu(dst->type)->from_float) { + ggml_from_float_t const quantize_row_q = ggml_get_type_traits_cpu(dst->type)->from_float; + float * src0_f32 = (float *) params->wdata + (ne00 + CACHE_LINE_SIZE_F32) * ith; + + size_t id = 0; + size_t rs = nb0 * (ne00 / ggml_blck_size(dst->type)); + char * dst_ptr = (char *) dst->data; + + for (int i03 = 0; i03 < ne03; i03++) { + for (int i02 = 0; i02 < ne02; i02++) { + id += rs * ir0; + for (int i01 = ir0; i01 < ir1; i01++) { + const ggml_bf16_t * src0_ptr = (ggml_bf16_t *) ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03); + + for (int i00 = 0; i00 < ne00; i00++) { + src0_f32[i00] = GGML_BF16_TO_FP32(src0_ptr[i00]); + } + + quantize_row_q(src0_f32, dst_ptr + id, ne00); + id += rs; + } + id += rs * (ne01 - ir1); + } + } + } else { + GGML_ABORT("fatal error"); // TODO: implement + } + } else { + //printf("%s: this is not optimal - fix me\n", __func__); + + if (dst->type == GGML_TYPE_F32) { + size_t id = 0; + float * dst_ptr = (float *) dst->data; + + for (int i03 = 0; i03 < ne03; i03++) { + for (int i02 = 0; i02 < ne02; i02++) { + id += ne00 * ir0; + for (int i01 = ir0; i01 < ir1; i01++) { + for (int i00 = 0; i00 < ne00; i00++) { + const ggml_bf16_t * src0_ptr = (ggml_bf16_t *) ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03); + + dst_ptr[id] = GGML_BF16_TO_FP32(*src0_ptr); + id++; + } + } + id += ne00 * (ne01 - ir1); + } + } + } else if (dst->type == GGML_TYPE_BF16) { + size_t id = 0; + ggml_bf16_t * dst_ptr = (ggml_bf16_t *) dst->data; + + for (int i03 = 0; i03 < ne03; i03++) { + for (int i02 = 0; i02 < ne02; i02++) { + id += ne00 * ir0; + for (int i01 = ir0; i01 < ir1; i01++) { + for (int i00 = 0; i00 < ne00; i00++) { + const ggml_bf16_t * src0_ptr = (ggml_bf16_t *) ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03); + + dst_ptr[id] = *src0_ptr; + id++; + } + } + id += ne00 * (ne01 - ir1); + } + } + } else if (dst->type == GGML_TYPE_F16) { + size_t id = 0; + ggml_fp16_t * dst_ptr = (ggml_fp16_t *) dst->data; + + for (int i03 = 0; i03 < ne03; i03++) { + for (int i02 = 0; i02 < ne02; i02++) { + id += ne00 * ir0; + for (int i01 = ir0; i01 < ir1; i01++) { + for (int i00 = 0; i00 < ne00; i00++) { + const ggml_bf16_t * src0_ptr = (ggml_bf16_t *) ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03); + + dst_ptr[id] = GGML_FP32_TO_FP16(GGML_BF16_TO_FP32(*src0_ptr)); + id++; + } + } + id += ne00 * (ne01 - ir1); + } + } + } else { + GGML_ABORT("fatal error"); // TODO: implement + } + } + return; + } + + // dst counters + int64_t i10 = 0; + int64_t i11 = 0; + int64_t i12 = 0; + int64_t i13 = 0; + + if (dst->type == GGML_TYPE_BF16) { + for (int64_t i03 = 0; i03 < ne03; i03++) { + for (int64_t i02 = 0; i02 < ne02; i02++) { + i10 += ne00 * ir0; + while (i10 >= ne0) { + i10 -= ne0; + if (++i11 == ne1) { + i11 = 0; + if (++i12 == ne2) { + i12 = 0; + if (++i13 == ne3) { + i13 = 0; + } + } + } + } + for (int64_t i01 = ir0; i01 < ir1; i01++) { + for (int64_t i00 = 0; i00 < ne00; i00++) { + const char * src0_ptr = ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03); + char * dst_ptr = ((char *) dst->data + i10*nb0 + i11*nb1 + i12*nb2 + i13*nb3); + + memcpy(dst_ptr, src0_ptr, sizeof(ggml_bf16_t)); + + if (++i10 == ne00) { + i10 = 0; + if (++i11 == ne01) { + i11 = 0; + if (++i12 == ne02) { + i12 = 0; + if (++i13 == ne03) { + i13 = 0; + } + } + } + } + } + } + i10 += ne00 * (ne01 - ir1); + while (i10 >= ne0) { + i10 -= ne0; + if (++i11 == ne1) { + i11 = 0; + if (++i12 == ne2) { + i12 = 0; + if (++i13 == ne3) { + i13 = 0; + } + } + } + } + } + } + } else if (dst->type == GGML_TYPE_F16) { + for (int64_t i03 = 0; i03 < ne03; i03++) { + for (int64_t i02 = 0; i02 < ne02; i02++) { + i10 += ne00 * ir0; + while (i10 >= ne0) { + i10 -= ne0; + if (++i11 == ne1) { + i11 = 0; + if (++i12 == ne2) { + i12 = 0; + if (++i13 == ne3) { + i13 = 0; + } + } + } + } + for (int64_t i01 = ir0; i01 < ir1; i01++) { + for (int64_t i00 = 0; i00 < ne00; i00++) { + const char * src0_ptr = ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03); + char * dst_ptr = ((char *) dst->data + i10*nb0 + i11*nb1 + i12*nb2 + i13*nb3); + + *(ggml_fp16_t *) dst_ptr = GGML_FP32_TO_FP16(GGML_BF16_TO_FP32(*(const ggml_bf16_t *) src0_ptr)); + + if (++i10 == ne0) { + i10 = 0; + if (++i11 == ne1) { + i11 = 0; + if (++i12 == ne2) { + i12 = 0; + if (++i13 == ne3) { + i13 = 0; + } + } + } + } + } + } + i10 += ne00 * (ne01 - ir1); + while (i10 >= ne0) { + i10 -= ne0; + if (++i11 == ne1) { + i11 = 0; + if (++i12 == ne2) { + i12 = 0; + if (++i13 == ne3) { + i13 = 0; + } + } + } + } + } + } + } else if (dst->type == GGML_TYPE_F32) { + for (int64_t i03 = 0; i03 < ne03; i03++) { + for (int64_t i02 = 0; i02 < ne02; i02++) { + i10 += ne00 * ir0; + while (i10 >= ne0) { + i10 -= ne0; + if (++i11 == ne1) { + i11 = 0; + if (++i12 == ne2) { + i12 = 0; + if (++i13 == ne3) { + i13 = 0; + } + } + } + } + for (int64_t i01 = ir0; i01 < ir1; i01++) { + for (int64_t i00 = 0; i00 < ne00; i00++) { + const char * src0_ptr = ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03); + char * dst_ptr = ((char *) dst->data + i10*nb0 + i11*nb1 + i12*nb2 + i13*nb3); + + *(float *) dst_ptr = GGML_BF16_TO_FP32(*(const ggml_bf16_t *) src0_ptr); + + if (++i10 == ne0) { + i10 = 0; + if (++i11 == ne1) { + i11 = 0; + if (++i12 == ne2) { + i12 = 0; + if (++i13 == ne3) { + i13 = 0; + } + } + } + } + } + } + i10 += ne00 * (ne01 - ir1); + while (i10 >= ne0) { + i10 -= ne0; + if (++i11 == ne1) { + i11 = 0; + if (++i12 == ne2) { + i12 = 0; + if (++i13 == ne3) { + i13 = 0; + } + } + } + } + } + } + } else { + GGML_ABORT("fatal error"); // TODO: implement + } +} + +static void ggml_compute_forward_dup_f32( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + + GGML_ASSERT(ggml_nelements(dst) == ggml_nelements(src0)); + + GGML_TENSOR_UNARY_OP_LOCALS + + const int ith = params->ith; // thread index + const int nth = params->nth; // number of threads + + // parallelize by rows + const int nr = ne01; + // number of rows per thread + const int dr = (nr + nth - 1) / nth; + // row range for this thread + const int ir0 = dr * ith; + const int ir1 = MIN(ir0 + dr, nr); + + if (src0->type == dst->type && + ne00 == ne0 && + nb00 == ggml_type_size(src0->type) && nb0 == ggml_type_size(dst->type)) { + // copy by rows + const size_t rs = ne00*nb00; + for (int64_t i03 = 0; i03 < ne03; i03++) { + for (int64_t i02 = 0; i02 < ne02; i02++) { + for (int64_t i01 = ir0; i01 < ir1; i01++) { + memcpy( + ((char *) dst->data + i01*nb1 + i02*nb2 + i03*nb3), + ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03), + rs); + } + } + } + return; + } + + if (ggml_is_contiguous(dst)) { + // TODO: simplify + if (nb00 == sizeof(float)) { + if (dst->type == GGML_TYPE_F32) { + size_t id = 0; + const size_t rs = ne00 * nb00; + char * dst_ptr = (char *) dst->data; + + for (int i03 = 0; i03 < ne03; i03++) { + for (int i02 = 0; i02 < ne02; i02++) { + id += rs * ir0; + for (int i01 = ir0; i01 < ir1; i01++) { + const char * src0_ptr = (char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03; + memcpy(dst_ptr + id, src0_ptr, rs); + id += rs; + } + id += rs * (ne01 - ir1); + } + } + } else if (ggml_get_type_traits_cpu(dst->type)->from_float) { + ggml_from_float_t const quantize_row_q = ggml_get_type_traits_cpu(dst->type)->from_float; + + size_t id = 0; + size_t rs = nb0 * (ne00 / ggml_blck_size(dst->type)); + char * dst_ptr = (char *) dst->data; + + for (int i03 = 0; i03 < ne03; i03++) { + for (int i02 = 0; i02 < ne02; i02++) { + id += rs * ir0; + for (int i01 = ir0; i01 < ir1; i01++) { + const float * src0_ptr = (float *) ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03); + quantize_row_q(src0_ptr, dst_ptr + id, ne00); + id += rs; + } + id += rs * (ne01 - ir1); + } + } + } else { + GGML_ABORT("fatal error"); // TODO: implement + } + } else { + //printf("%s: this is not optimal - fix me\n", __func__); + + if (dst->type == GGML_TYPE_F32) { + size_t id = 0; + float * dst_ptr = (float *) dst->data; + + for (int i03 = 0; i03 < ne03; i03++) { + for (int i02 = 0; i02 < ne02; i02++) { + id += ne00 * ir0; + for (int i01 = ir0; i01 < ir1; i01++) { + for (int i00 = 0; i00 < ne00; i00++) { + const float * src0_ptr = (float *) ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03); + + dst_ptr[id] = *src0_ptr; + id++; + } + } + id += ne00 * (ne01 - ir1); + } + } + } else if (dst->type == GGML_TYPE_F16) { + size_t id = 0; + ggml_fp16_t * dst_ptr = (ggml_fp16_t *) dst->data; + + for (int i03 = 0; i03 < ne03; i03++) { + for (int i02 = 0; i02 < ne02; i02++) { + id += ne00 * ir0; + for (int i01 = ir0; i01 < ir1; i01++) { + for (int i00 = 0; i00 < ne00; i00++) { + const float * src0_ptr = (float *) ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03); + + dst_ptr[id] = GGML_FP32_TO_FP16(*src0_ptr); + id++; + } + } + id += ne00 * (ne01 - ir1); + } + } + } else if (dst->type == GGML_TYPE_BF16) { + size_t id = 0; + ggml_bf16_t * dst_ptr = (ggml_bf16_t *) dst->data; + + for (int i03 = 0; i03 < ne03; i03++) { + for (int i02 = 0; i02 < ne02; i02++) { + id += ne00 * ir0; + for (int i01 = ir0; i01 < ir1; i01++) { + for (int i00 = 0; i00 < ne00; i00++) { + const float * src0_ptr = (float *) ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03); + + dst_ptr[id] = GGML_FP32_TO_BF16(*src0_ptr); + id++; + } + } + id += ne00 * (ne01 - ir1); + } + } + } else { + GGML_ABORT("fatal error"); // TODO: implement + } + } + + return; + } + + // dst counters + + int64_t i10 = 0; + int64_t i11 = 0; + int64_t i12 = 0; + int64_t i13 = 0; + + if (dst->type == GGML_TYPE_F32) { + for (int64_t i03 = 0; i03 < ne03; i03++) { + for (int64_t i02 = 0; i02 < ne02; i02++) { + i10 += ne00 * ir0; + while (i10 >= ne0) { + i10 -= ne0; + if (++i11 == ne1) { + i11 = 0; + if (++i12 == ne2) { + i12 = 0; + if (++i13 == ne3) { + i13 = 0; + } + } + } + } + for (int64_t i01 = ir0; i01 < ir1; i01++) { + for (int64_t i00 = 0; i00 < ne00; i00++) { + const char * src0_ptr = ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03); + char * dst_ptr = ((char *) dst->data + i10*nb0 + i11*nb1 + i12*nb2 + i13*nb3); + + memcpy(dst_ptr, src0_ptr, sizeof(float)); + + if (++i10 == ne0) { + i10 = 0; + if (++i11 == ne1) { + i11 = 0; + if (++i12 == ne2) { + i12 = 0; + if (++i13 == ne3) { + i13 = 0; + } + } + } + } + } + } + i10 += ne00 * (ne01 - ir1); + while (i10 >= ne0) { + i10 -= ne0; + if (++i11 == ne1) { + i11 = 0; + if (++i12 == ne2) { + i12 = 0; + if (++i13 == ne3) { + i13 = 0; + } + } + } + } + } + } + } else if (dst->type == GGML_TYPE_F16) { + for (int64_t i03 = 0; i03 < ne03; i03++) { + for (int64_t i02 = 0; i02 < ne02; i02++) { + i10 += ne00 * ir0; + while (i10 >= ne0) { + i10 -= ne0; + if (++i11 == ne1) { + i11 = 0; + if (++i12 == ne2) { + i12 = 0; + if (++i13 == ne3) { + i13 = 0; + } + } + } + } + for (int64_t i01 = ir0; i01 < ir1; i01++) { + for (int64_t i00 = 0; i00 < ne00; i00++) { + const char * src0_ptr = ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03); + char * dst_ptr = ((char *) dst->data + i10*nb0 + i11*nb1 + i12*nb2 + i13*nb3); + + *(ggml_fp16_t *) dst_ptr = GGML_FP32_TO_FP16(*(const float *) src0_ptr); + + if (++i10 == ne0) { + i10 = 0; + if (++i11 == ne1) { + i11 = 0; + if (++i12 == ne2) { + i12 = 0; + if (++i13 == ne3) { + i13 = 0; + } + } + } + } + } + } + i10 += ne00 * (ne01 - ir1); + while (i10 >= ne0) { + i10 -= ne0; + if (++i11 == ne1) { + i11 = 0; + if (++i12 == ne2) { + i12 = 0; + if (++i13 == ne3) { + i13 = 0; + } + } + } + } + } + } + } else if (dst->type == GGML_TYPE_BF16) { + for (int64_t i03 = 0; i03 < ne03; i03++) { + for (int64_t i02 = 0; i02 < ne02; i02++) { + i10 += ne00 * ir0; + while (i10 >= ne0) { + i10 -= ne0; + if (++i11 == ne1) { + i11 = 0; + if (++i12 == ne2) { + i12 = 0; + if (++i13 == ne3) { + i13 = 0; + } + } + } + } + for (int64_t i01 = ir0; i01 < ir1; i01++) { + for (int64_t i00 = 0; i00 < ne00; i00++) { + const char * src0_ptr = ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03); + char * dst_ptr = ((char *) dst->data + i10*nb0 + i11*nb1 + i12*nb2 + i13*nb3); + + *(ggml_bf16_t *) dst_ptr = GGML_FP32_TO_BF16(*(const float *) src0_ptr); + + if (++i10 == ne0) { + i10 = 0; + if (++i11 == ne1) { + i11 = 0; + if (++i12 == ne2) { + i12 = 0; + if (++i13 == ne3) { + i13 = 0; + } + } + } + } + } + } + i10 += ne00 * (ne01 - ir1); + while (i10 >= ne0) { + i10 -= ne0; + if (++i11 == ne1) { + i11 = 0; + if (++i12 == ne2) { + i12 = 0; + if (++i13 == ne3) { + i13 = 0; + } + } + } + } + } + } + } else { + GGML_ABORT("fatal error"); // TODO: implement + } +} + +// A simplified version of ggml_compute_forward_dup that doesn't do float upcasting, and just plain old memcpy. +static void ggml_compute_forward_dup_bytes( + const ggml_compute_params * params, + ggml_tensor * dst) { + const ggml_tensor * src0 = dst->src[0]; + + GGML_ASSERT(ggml_nelements(dst) == ggml_nelements(src0)); + GGML_ASSERT(src0->type == dst->type); + + GGML_TENSOR_UNARY_OP_LOCALS; + + if (ggml_is_contiguous(src0) && ggml_is_contiguous(dst)) { + ggml_compute_forward_dup_same_cont(params, dst); + return; + } + + const size_t type_size = ggml_type_size(src0->type); + + const int ith = params->ith; // thread index + const int nth = params->nth; // number of threads + + // parallelize by rows + const int nr = ne01; + // number of rows per thread + const int dr = (nr + nth - 1) / nth; + // row range for this thread + const int ir0 = dr * ith; + const int ir1 = MIN(ir0 + dr, nr); + + if (src0->type == dst->type && + ggml_are_same_shape(src0, dst) && + nb00 == type_size && nb0 == type_size) { + // copy by rows + const size_t rs = ggml_row_size(src0->type, ne00); + for (int64_t i03 = 0; i03 < ne03; i03++) { + for (int64_t i02 = 0; i02 < ne02; i02++) { + for (int64_t i01 = ir0; i01 < ir1; i01++) { + memcpy( + ((char *) dst->data + i01*nb1 + i02*nb2 + i03*nb3), + ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03), + rs); + } + } + } + return; + } + + if (ggml_is_contiguous(dst)) { + size_t id = 0; + char * dst_ptr = (char *) dst->data; + const size_t rs = ne00 * type_size; + + if (nb00 == type_size) { + // src0 is contigous on first dimension, copy by rows + for (int64_t i03 = 0; i03 < ne03; i03++) { + for (int64_t i02 = 0; i02 < ne02; i02++) { + id += rs * ir0; + for (int64_t i01 = ir0; i01 < ir1; i01++) { + const char * src0_ptr = (char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03; + memcpy(dst_ptr + id, src0_ptr, rs); + id += rs; + } + id += rs * (ne01 - ir1); + } + } + } else { + //printf("%s: this is not optimal - fix me\n", __func__); + + for (int64_t i03 = 0; i03 < ne03; i03++) { + for (int64_t i02 = 0; i02 < ne02; i02++) { + id += rs * ir0; + for (int64_t i01 = ir0; i01 < ir1; i01++) { + for (int64_t i00 = 0; i00 < ne00; i00++) { + const char * src0_ptr = (char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03; + memcpy(dst_ptr + id, src0_ptr, type_size); + + id += type_size; + } + } + id += rs * (ne01 - ir1); + } + } + } + + return; + } + + // dst counters + int64_t k10 = 0; + int64_t i11 = 0; + int64_t i12 = 0; + int64_t i13 = 0; + + // number of blocks in a row + const int64_t nk00 = ne00 / ggml_blck_size(src0->type); + const int64_t nk0 = ne0 / ggml_blck_size(dst->type); + + for (int64_t i03 = 0; i03 < ne03; i03++) { + for (int64_t i02 = 0; i02 < ne02; i02++) { + k10 += nk00 * ir0; + while (k10 >= nk0) { + k10 -= nk0; + if (++i11 == ne1) { + i11 = 0; + if (++i12 == ne2) { + i12 = 0; + if (++i13 == ne3) { + i13 = 0; + } + } + } + } + for (int64_t i01 = ir0; i01 < ir1; i01++) { + for (int64_t k00 = 0; k00 < nk00; k00++) { + const char * src0_ptr = ((char *) src0->data + k00*nb00 + i01*nb01 + i02*nb02 + i03*nb03); + char * dst_ptr = ((char *) dst->data + k10*nb0 + i11*nb1 + i12*nb2 + i13*nb3); + + memcpy(dst_ptr, src0_ptr, type_size); + + if (++k10 == nk0) { + k10 = 0; + if (++i11 == ne1) { + i11 = 0; + if (++i12 == ne2) { + i12 = 0; + if (++i13 == ne3) { + i13 = 0; + } + } + } + } + } + } + k10 += nk00 * (ne01 - ir1); + while (k10 >= nk0) { + k10 -= nk0; + if (++i11 == ne1) { + i11 = 0; + if (++i12 == ne2) { + i12 = 0; + if (++i13 == ne3) { + i13 = 0; + } + } + } + } + } + } +} + +static void ggml_compute_forward_dup_q( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + const ggml_tensor * src1 = dst->src[1]; + + GGML_TENSOR_BINARY_OP_LOCALS + + const ggml_type type = src0->type; + ggml_to_float_t const dequantize_row_q = ggml_get_type_traits(type)->to_float; + + size_t qk = ggml_blck_size(type); + const int64_t nr = ggml_nelements(src1) / qk; + + // destination must be contiguous in the first dimension + GGML_ASSERT(nb10 == ggml_type_size(dst->type)); + // must either have first dimension large enough to hold a row, or fully contiguous + GGML_ASSERT((ne10 % qk) == 0 || ggml_is_contiguous(dst)); + + const int ith = params->ith; + const int nth = params->nth; + + const int dr = (nr + nth - 1)/nth; + + // row range for this thread + const int ir0 = dr*ith; + const int ir1 = MIN(ir0 + dr, nr); + + for (int64_t ir = ir0; ir < ir1; ++ir) { + + uint32_t i = ir * qk; + + const int64_t i03 = i/(ne00 * ne01 * ne02); + const int64_t i02 = (i - i03*ne00*ne01*ne02 )/ (ne00*ne01); + const int64_t i01 = (i - i03*ne00*ne01*ne02 - i02*ne01*ne00) / ne00; + const int64_t i00 = i - i03*ne00*ne01*ne02 - i02*ne01*ne00 - i01*ne00; + const int64_t x_offset = (i00/qk)*nb00 + i01*nb01 + i02*nb02 + i03 * nb03; + + const int64_t i13 = i/(ne10 * ne11 * ne12); + const int64_t i12 = (i - i13*ne10*ne11*ne12) / (ne10*ne11); + const int64_t i11 = (i - i13*ne10*ne11*ne12 - i12*ne10*ne11) / ne10; + const int64_t i10 = i - i13*ne10*ne11*ne12 - i12*ne10*ne11 - i11*ne10; + const int64_t dst_offset = i10*nb10 + i11*nb11 + i12*nb12 + i13*nb13; + + dequantize_row_q( + (const void *) ((char *) src0->data + x_offset), + (float *) ((char *) dst->data + dst_offset), qk); + } +} + +void ggml_compute_forward_dup( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + + if (src0->type == dst->type) { + ggml_compute_forward_dup_bytes(params, dst); + return; + } + + switch (src0->type) { + case GGML_TYPE_F16: + { + ggml_compute_forward_dup_f16(params, dst); + } break; + case GGML_TYPE_BF16: + { + ggml_compute_forward_dup_bf16(params, dst); + } break; + case GGML_TYPE_F32: + { + ggml_compute_forward_dup_f32(params, dst); + } break; + default: + { + if (ggml_is_quantized(src0->type) && dst->type == GGML_TYPE_F32) { + ggml_compute_forward_dup_q(params, dst); + break; + } + GGML_ABORT("fatal error"); + } + } +} + +// ggml_compute_forward_add + +static void ggml_compute_forward_add_q_f32( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + const ggml_tensor * src1 = dst->src[1]; + + GGML_ASSERT(ggml_are_same_shape(src0, src1) && ggml_are_same_shape(src0, dst)); + + const int nr = ggml_nrows(src0); + + GGML_TENSOR_BINARY_OP_LOCALS + + const int ith = params->ith; + const int nth = params->nth; + + const ggml_type type = src0->type; + const ggml_type dtype = dst->type; + ggml_to_float_t const dequantize_row_q = ggml_get_type_traits(type)->to_float; + ggml_from_float_t const quantize_row_q = ggml_get_type_traits_cpu(dtype)->from_float; + + // we don't support permuted src0 or src1 + GGML_ASSERT(nb00 == ggml_type_size(type)); + GGML_ASSERT(nb10 == sizeof(float)); + + // dst cannot be transposed or permuted + GGML_ASSERT(nb0 <= nb1); + GGML_ASSERT(nb1 <= nb2); + GGML_ASSERT(nb2 <= nb3); + + GGML_ASSERT(ggml_is_quantized(src0->type)); + GGML_ASSERT(src1->type == GGML_TYPE_F32); + + // rows per thread + const int dr = (nr + nth - 1)/nth; + + // row range for this thread + const int ir0 = dr*ith; + const int ir1 = MIN(ir0 + dr, nr); + + float * wdata = (float *) params->wdata + (ne00 + CACHE_LINE_SIZE_F32) * ith; + + for (int ir = ir0; ir < ir1; ++ir) { + // src0 indices + const int i03 = ir/(ne02*ne01); + const int i02 = (ir - i03*ne02*ne01)/ne01; + const int i01 = (ir - i03*ne02*ne01 - i02*ne01); + + // src1 and dst are same shape as src0 => same indices + const int i13 = i03; + const int i12 = i02; + const int i11 = i01; + + const int i3 = i03; + const int i2 = i02; + const int i1 = i01; + + void * src0_row = (void *) ((char *) src0->data + (i01*nb01 + i02*nb02 + i03*nb03)); + float * src1_row = (float *)((char *) src1->data + (i11*nb11 + i12*nb12 + i13*nb13)); + void * dst_row = (void *) ((char *) dst->data + ( i1*nb1 + i2*nb2 + i3*nb3)); + + assert(ne00 % 32 == 0); + + // unquantize row from src0 to temp buffer + dequantize_row_q(src0_row, wdata, ne00); + // add src1 + ggml_vec_acc_f32(ne00, wdata, src1_row); + // quantize row to dst + if (quantize_row_q != NULL) { + quantize_row_q(wdata, dst_row, ne00); + } else { + memcpy(dst_row, wdata, ne0*nb0); + } + } +} + +void ggml_compute_forward_add( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F32: + case GGML_TYPE_F16: + case GGML_TYPE_BF16: + { + ggml_compute_forward_add_non_quantized(params, dst); + } break; + case GGML_TYPE_Q4_0: + case GGML_TYPE_Q4_1: + case GGML_TYPE_Q5_0: + case GGML_TYPE_Q5_1: + case GGML_TYPE_Q8_0: + case GGML_TYPE_Q2_K: + case GGML_TYPE_Q3_K: + case GGML_TYPE_Q4_K: + case GGML_TYPE_Q5_K: + case GGML_TYPE_Q6_K: + case GGML_TYPE_TQ1_0: + case GGML_TYPE_TQ2_0: + case GGML_TYPE_IQ2_XXS: + case GGML_TYPE_IQ2_XS: + case GGML_TYPE_IQ3_XXS: + case GGML_TYPE_IQ1_S: + case GGML_TYPE_IQ1_M: + case GGML_TYPE_IQ4_NL: + case GGML_TYPE_IQ4_XS: + case GGML_TYPE_IQ3_S: + case GGML_TYPE_IQ2_S: + { + ggml_compute_forward_add_q_f32(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +// ggml_compute_forward_add1 + +static void ggml_compute_forward_add1_f32( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + const ggml_tensor * src1 = dst->src[1]; + + GGML_ASSERT(ggml_are_same_shape(src0, dst)); + GGML_ASSERT(ggml_is_scalar(src1)); + + const int ith = params->ith; + const int nth = params->nth; + + const int nr = ggml_nrows(src0); + + GGML_TENSOR_UNARY_OP_LOCALS + + GGML_ASSERT( nb0 == sizeof(float)); + GGML_ASSERT(nb00 == sizeof(float)); + + // rows per thread + const int dr = (nr + nth - 1)/nth; + + // row range for this thread + const int ir0 = dr*ith; + const int ir1 = MIN(ir0 + dr, nr); + + for (int ir = ir0; ir < ir1; ++ir) { + // src0 and dst are same shape => same indices + const int i3 = ir/(ne2*ne1); + const int i2 = (ir - i3*ne2*ne1)/ne1; + const int i1 = (ir - i3*ne2*ne1 - i2*ne1); + +#ifdef GGML_USE_ACCELERATE + GGML_UNUSED(ggml_vec_add1_f32); + + vDSP_vadd( + (float *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01), 1, + (float *) ((char *) src1->data), 0, + (float *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 ), 1, + ne0); +#else + ggml_vec_add1_f32(ne0, + (float *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 ), + (float *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01), + *(float *) src1->data); +#endif + } +} + +static void ggml_compute_forward_add1_f16_f32( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + const ggml_tensor * src1 = dst->src[1]; + + GGML_ASSERT(ggml_are_same_shape(src0, dst)); + GGML_ASSERT(ggml_is_scalar(src1)); + + // scalar to add + const float v = *(float *) src1->data; + + const int ith = params->ith; + const int nth = params->nth; + + const int nr = ggml_nrows(src0); + + GGML_TENSOR_UNARY_OP_LOCALS + + GGML_ASSERT(src0->type == GGML_TYPE_F16); + GGML_ASSERT(src1->type == GGML_TYPE_F32); + GGML_ASSERT(dst->type == GGML_TYPE_F16); + + GGML_ASSERT( nb0 == sizeof(ggml_fp16_t)); + GGML_ASSERT(nb00 == sizeof(ggml_fp16_t)); + + // rows per thread + const int dr = (nr + nth - 1)/nth; + + // row range for this thread + const int ir0 = dr*ith; + const int ir1 = MIN(ir0 + dr, nr); + + for (int ir = ir0; ir < ir1; ++ir) { + // src0 and dst are same shape => same indices + const int i3 = ir/(ne2*ne1); + const int i2 = (ir - i3*ne2*ne1)/ne1; + const int i1 = (ir - i3*ne2*ne1 - i2*ne1); + + ggml_fp16_t * dst_ptr = (ggml_fp16_t *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 ); + ggml_fp16_t * src0_ptr = (ggml_fp16_t *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01); + for (int i = 0; i < ne0; i++) { + dst_ptr[i] = GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(src0_ptr[i]) + v); + } + } +} + +static void ggml_compute_forward_add1_f16_f16( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + const ggml_tensor * src1 = dst->src[1]; + + GGML_ASSERT(ggml_are_same_shape(src0, dst)); + GGML_ASSERT(ggml_is_scalar(src1)); + + // scalar to add + const float v = GGML_FP16_TO_FP32(*(ggml_fp16_t *) src1->data); + + const int ith = params->ith; + const int nth = params->nth; + + const int nr = ggml_nrows(src0); + + GGML_TENSOR_UNARY_OP_LOCALS + + GGML_ASSERT(src0->type == GGML_TYPE_F16); + GGML_ASSERT(src1->type == GGML_TYPE_F16); + GGML_ASSERT(dst->type == GGML_TYPE_F16); + + GGML_ASSERT( nb0 == sizeof(ggml_fp16_t)); + GGML_ASSERT(nb00 == sizeof(ggml_fp16_t)); + + // rows per thread + const int dr = (nr + nth - 1)/nth; + + // row range for this thread + const int ir0 = dr*ith; + const int ir1 = MIN(ir0 + dr, nr); + + for (int ir = ir0; ir < ir1; ++ir) { + // src0 and dst are same shape => same indices + const int i3 = ir/(ne2*ne1); + const int i2 = (ir - i3*ne2*ne1)/ne1; + const int i1 = (ir - i3*ne2*ne1 - i2*ne1); + + ggml_fp16_t * dst_ptr = (ggml_fp16_t *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 ); + ggml_fp16_t * src0_ptr = (ggml_fp16_t *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01); + for (int i = 0; i < ne0; i++) { + dst_ptr[i] = GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(src0_ptr[i]) + v); + } + } +} + +static void ggml_compute_forward_add1_q_f32( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + const ggml_tensor * src1 = dst->src[1]; + + GGML_ASSERT(ggml_are_same_shape(src0, dst)); + GGML_ASSERT(ggml_is_scalar(src1)); + + // scalar to add + const float v = *(float *) src1->data; + + const int ith = params->ith; + const int nth = params->nth; + + const int nr = ggml_nrows(src0); + + GGML_TENSOR_UNARY_OP_LOCALS + + const ggml_type type = src0->type; + ggml_to_float_t const dequantize_row_q = ggml_get_type_traits(type)->to_float; + ggml_from_float_t const quantize_row_q = ggml_get_type_traits_cpu(type)->from_float; + + // we don't support permuted src0 + GGML_ASSERT(nb00 == ggml_type_size(type)); + + // dst cannot be transposed or permuted + GGML_ASSERT(nb0 <= nb1); + GGML_ASSERT(nb1 <= nb2); + GGML_ASSERT(nb2 <= nb3); + + GGML_ASSERT(ggml_is_quantized(src0->type)); + GGML_ASSERT(dst->type == src0->type); + GGML_ASSERT(src1->type == GGML_TYPE_F32); + + // rows per thread + const int dr = (nr + nth - 1)/nth; + + // row range for this thread + const int ir0 = dr*ith; + const int ir1 = MIN(ir0 + dr, nr); + + float * wdata = (float *) params->wdata + (ne0 + CACHE_LINE_SIZE_F32) * ith; + + for (int ir = ir0; ir < ir1; ++ir) { + // src0 and dst are same shape => same indices + const int i3 = ir/(ne2*ne1); + const int i2 = (ir - i3*ne2*ne1)/ne1; + const int i1 = (ir - i3*ne2*ne1 - i2*ne1); + + void * src0_row = (void *) ((char *) src0->data + (i1*nb01 + i2*nb02 + i3*nb03)); + void * dst_row = (void *) ((char *) dst->data + (i1*nb1 + i2*nb2 + i3*nb0 )); + + assert(ne0 % 32 == 0); + + // unquantize row from src0 to temp buffer + dequantize_row_q(src0_row, wdata, ne0); + // add src1 + ggml_vec_acc1_f32(ne0, wdata, v); + // quantize row to dst + quantize_row_q(wdata, dst_row, ne0); + } +} + +static void ggml_compute_forward_add1_bf16_f32( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + const ggml_tensor * src1 = dst->src[1]; + + GGML_ASSERT(ggml_are_same_shape(src0, dst)); + GGML_ASSERT(ggml_is_scalar(src1)); + + // scalar to add + const float v = *(float *) src1->data; + + const int ith = params->ith; + const int nth = params->nth; + + const int nr = ggml_nrows(src0); + + GGML_TENSOR_UNARY_OP_LOCALS + + GGML_ASSERT(src0->type == GGML_TYPE_BF16); + GGML_ASSERT(src1->type == GGML_TYPE_F32); + GGML_ASSERT(dst->type == GGML_TYPE_BF16); + + GGML_ASSERT( nb0 == sizeof(ggml_bf16_t)); + GGML_ASSERT(nb00 == sizeof(ggml_bf16_t)); + + // rows per thread + const int dr = (nr + nth - 1)/nth; + + // row range for this thread + const int ir0 = dr*ith; + const int ir1 = MIN(ir0 + dr, nr); + + for (int ir = ir0; ir < ir1; ++ir) { + // src0 and dst are same shape => same indices + const int i3 = ir/(ne2*ne1); + const int i2 = (ir - i3*ne2*ne1)/ne1; + const int i1 = (ir - i3*ne2*ne1 - i2*ne1); + + ggml_bf16_t * dst_ptr = (ggml_bf16_t *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 ); + ggml_bf16_t * src0_ptr = (ggml_bf16_t *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01); + for (int i = 0; i < ne0; i++) { + dst_ptr[i] = GGML_FP32_TO_BF16(GGML_BF16_TO_FP32(src0_ptr[i]) + v); + } + } +} + +static void ggml_compute_forward_add1_bf16_bf16( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + const ggml_tensor * src1 = dst->src[1]; + + GGML_ASSERT(ggml_are_same_shape(src0, dst)); + GGML_ASSERT(ggml_is_scalar(src1)); + + // scalar to add + const float v = GGML_BF16_TO_FP32(*(ggml_bf16_t *) src1->data); + + const int ith = params->ith; + const int nth = params->nth; + + const int nr = ggml_nrows(src0); + + GGML_TENSOR_UNARY_OP_LOCALS + + GGML_ASSERT(src0->type == GGML_TYPE_BF16); + GGML_ASSERT(src1->type == GGML_TYPE_BF16); + GGML_ASSERT(dst->type == GGML_TYPE_BF16); + + GGML_ASSERT( nb0 == sizeof(ggml_bf16_t)); + GGML_ASSERT(nb00 == sizeof(ggml_bf16_t)); + + // rows per thread + const int dr = (nr + nth - 1)/nth; + + // row range for this thread + const int ir0 = dr*ith; + const int ir1 = MIN(ir0 + dr, nr); + + for (int ir = ir0; ir < ir1; ++ir) { + // src0 and dst are same shape => same indices + const int i3 = ir/(ne2*ne1); + const int i2 = (ir - i3*ne2*ne1)/ne1; + const int i1 = (ir - i3*ne2*ne1 - i2*ne1); + + ggml_bf16_t * dst_ptr = (ggml_bf16_t *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 ); + ggml_bf16_t * src0_ptr = (ggml_bf16_t *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01); + for (int i = 0; i < ne0; i++) { + dst_ptr[i] = GGML_FP32_TO_BF16(GGML_BF16_TO_FP32(src0_ptr[i]) + v); + } + } +} + +void ggml_compute_forward_add1( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + const ggml_tensor * src1 = dst->src[1]; + + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_add1_f32(params, dst); + } break; + case GGML_TYPE_F16: + { + if (src1->type == GGML_TYPE_F16) { + ggml_compute_forward_add1_f16_f16(params, dst); + } + else if (src1->type == GGML_TYPE_F32) { + ggml_compute_forward_add1_f16_f32(params, dst); + } + else { + GGML_ABORT("fatal error"); + } + } break; + case GGML_TYPE_BF16: + { + if (src1->type == GGML_TYPE_BF16) { + ggml_compute_forward_add1_bf16_bf16(params, dst); + } + else if (src1->type == GGML_TYPE_F32) { + ggml_compute_forward_add1_bf16_f32(params, dst); + } + else { + GGML_ABORT("fatal error"); + } + } break; + case GGML_TYPE_Q4_0: + case GGML_TYPE_Q4_1: + case GGML_TYPE_Q5_0: + case GGML_TYPE_Q5_1: + case GGML_TYPE_Q8_0: + case GGML_TYPE_Q8_1: + case GGML_TYPE_Q2_K: + case GGML_TYPE_Q3_K: + case GGML_TYPE_Q4_K: + case GGML_TYPE_Q5_K: + case GGML_TYPE_Q6_K: + case GGML_TYPE_TQ1_0: + case GGML_TYPE_TQ2_0: + case GGML_TYPE_IQ2_XXS: + case GGML_TYPE_IQ2_XS: + case GGML_TYPE_IQ3_XXS: + case GGML_TYPE_IQ1_S: + case GGML_TYPE_IQ1_M: + case GGML_TYPE_IQ4_NL: + case GGML_TYPE_IQ4_XS: + case GGML_TYPE_IQ3_S: + case GGML_TYPE_IQ2_S: + { + ggml_compute_forward_add1_q_f32(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +// ggml_compute_forward_acc + +static void ggml_compute_forward_acc_f32( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + const ggml_tensor * src1 = dst->src[1]; + + GGML_ASSERT(ggml_are_same_shape(src0, dst)); + GGML_ASSERT(ggml_is_contiguous(dst) && ggml_is_contiguous(src0)); + + // view src0 and dst with these strides and data offset inbytes during acc + // nb0 is implicitly element_size because src0 and dst are contiguous + size_t nb1 = ((int32_t *) dst->op_params)[0]; + size_t nb2 = ((int32_t *) dst->op_params)[1]; + size_t nb3 = ((int32_t *) dst->op_params)[2]; + size_t offset = ((int32_t *) dst->op_params)[3]; + bool inplace = (bool) ((int32_t *) dst->op_params)[4]; + + if (!inplace) { + if (params->ith == 0) { + // memcpy needs to be synchronized across threads to avoid race conditions. + // => do it in INIT phase + memcpy( + ((char *) dst->data), + ((char *) src0->data), + ggml_nbytes(dst)); + } + ggml_barrier(params->threadpool); + } + + const int ith = params->ith; + const int nth = params->nth; + + const int nr = ggml_nrows(src1); + const int nc = src1->ne[0]; + + GGML_TENSOR_LOCALS(int64_t, ne1, src1, ne) + GGML_TENSOR_LOCALS(size_t, nb1, src1, nb) + + // src0 and dst as viewed during acc + const size_t nb0 = ggml_element_size(src0); + + const size_t nb00 = nb0; + const size_t nb01 = nb1; + const size_t nb02 = nb2; + const size_t nb03 = nb3; + + GGML_ASSERT(offset + (ne10 == 0 ? 0 : ne10-1)*nb0 + (ne11 == 0 ? 0 : ne11-1)*nb1 + (ne12 == 0 ? 0 : ne12-1)*nb2 + (ne13 == 0 ? 0 : ne13-1)*nb3 < ggml_nbytes(dst)); + GGML_ASSERT(offset + (ne10 == 0 ? 0 : ne10-1)*nb00 + (ne11 == 0 ? 0 : ne11-1)*nb01 + (ne12 == 0 ? 0 : ne12-1)*nb02 + (ne13 == 0 ? 0 : ne13-1)*nb03 < ggml_nbytes(src0)); + + GGML_ASSERT(nb10 == sizeof(float)); + + // rows per thread + const int dr = (nr + nth - 1)/nth; + + // row range for this thread + const int ir0 = dr*ith; + const int ir1 = MIN(ir0 + dr, nr); + + for (int ir = ir0; ir < ir1; ++ir) { + // src0 and dst are viewed with shape of src1 and offset + // => same indices + const int i3 = ir/(ne12*ne11); + const int i2 = (ir - i3*ne12*ne11)/ne11; + const int i1 = (ir - i3*ne12*ne11 - i2*ne11); + +#ifdef GGML_USE_ACCELERATE + vDSP_vadd( + (float *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + offset), 1, + (float *) ((char *) src1->data + i3*nb13 + i2*nb12 + i1*nb11), 1, + (float *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + offset), 1, nc); +#else + ggml_vec_add_f32(nc, + (float *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + offset), + (float *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + offset), + (float *) ((char *) src1->data + i3*nb13 + i2*nb12 + i1*nb11)); +#endif + } +} + +void ggml_compute_forward_acc( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_acc_f32(params, dst); + } break; + case GGML_TYPE_F16: + case GGML_TYPE_BF16: + case GGML_TYPE_Q4_0: + case GGML_TYPE_Q4_1: + case GGML_TYPE_Q5_0: + case GGML_TYPE_Q5_1: + case GGML_TYPE_Q8_0: + case GGML_TYPE_Q8_1: + case GGML_TYPE_Q2_K: + case GGML_TYPE_Q3_K: + case GGML_TYPE_Q4_K: + case GGML_TYPE_Q5_K: + case GGML_TYPE_Q6_K: + case GGML_TYPE_TQ1_0: + case GGML_TYPE_TQ2_0: + case GGML_TYPE_IQ2_XXS: + case GGML_TYPE_IQ2_XS: + case GGML_TYPE_IQ3_XXS: + case GGML_TYPE_IQ1_S: + case GGML_TYPE_IQ1_M: + case GGML_TYPE_IQ4_NL: + case GGML_TYPE_IQ4_XS: + case GGML_TYPE_IQ3_S: + case GGML_TYPE_IQ2_S: + default: + { + GGML_ABORT("fatal error"); + } + } +} + +// ggml_compute_forward_sum + +static void ggml_compute_forward_sum_f32( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + + if (params->ith != 0) { + return; + } + + assert(ggml_is_scalar(dst)); + assert(src0->nb[0] == sizeof(float)); + + GGML_TENSOR_LOCALS(int64_t, ne0, src0, ne) + GGML_TENSOR_LOCALS(size_t, nb0, src0, nb) + + ggml_float sum = 0; + ggml_float row_sum = 0; + + for (int64_t i03 = 0; i03 < ne03; i03++) { + for (int64_t i02 = 0; i02 < ne02; i02++) { + for (int64_t i01 = 0; i01 < ne01; i01++) { + ggml_vec_sum_f32_ggf(ne00, + &row_sum, + (float *) ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03)); + sum += row_sum; + } + } + } + ((float *) dst->data)[0] = sum; +} + +static void ggml_compute_forward_sum_f16( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + + if (params->ith != 0) { + return; + } + + assert(ggml_is_scalar(dst)); + + assert(src0->nb[0] == sizeof(ggml_fp16_t)); + + GGML_TENSOR_LOCALS(int64_t, ne0, src0, ne) + GGML_TENSOR_LOCALS(size_t, nb0, src0, nb) + + float sum = 0; + float row_sum = 0; + + for (int64_t i03 = 0; i03 < ne03; i03++) { + for (int64_t i02 = 0; i02 < ne02; i02++) { + for (int64_t i01 = 0; i01 < ne01; i01++) { + ggml_vec_sum_f16_ggf(ne00, + &row_sum, + (ggml_fp16_t *) ((char *) src0->data + i01 * nb01 + i02 * nb02 + i03 * nb03)); + sum += row_sum; + } + } + } + ((ggml_fp16_t *) dst->data)[0] = GGML_FP32_TO_FP16(sum); +} + +static void ggml_compute_forward_sum_bf16( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + + if (params->ith != 0) { + return; + } + + assert(ggml_is_scalar(dst)); + + assert(src0->nb[0] == sizeof(ggml_bf16_t)); + + GGML_TENSOR_LOCALS(int64_t, ne0, src0, ne) + GGML_TENSOR_LOCALS(size_t, nb0, src0, nb) + + float sum = 0; + float row_sum = 0; + + for (int64_t i03 = 0; i03 < ne03; i03++) { + for (int64_t i02 = 0; i02 < ne02; i02++) { + for (int64_t i01 = 0; i01 < ne01; i01++) { + ggml_vec_sum_bf16_ggf(ne00, + &row_sum, + (ggml_bf16_t *) ((char *) src0->data + i01 * nb01 + i02 * nb02 + i03 * nb03)); + sum += row_sum; + } + } + } + ((ggml_bf16_t *) dst->data)[0] = GGML_FP32_TO_BF16(sum); +} + +void ggml_compute_forward_sum( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_sum_f32(params, dst); + } break; + case GGML_TYPE_F16: + { + ggml_compute_forward_sum_f16(params, dst); + } break; + case GGML_TYPE_BF16: + { + ggml_compute_forward_sum_bf16(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +// ggml_compute_forward_sum_rows + +static void ggml_compute_forward_sum_rows_f32( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + + if (params->ith != 0) { + return; + } + + GGML_ASSERT(src0->nb[0] == sizeof(float)); + GGML_ASSERT(dst->nb[0] == sizeof(float)); + + GGML_TENSOR_UNARY_OP_LOCALS + + GGML_ASSERT(ne0 == 1); + GGML_ASSERT(ne1 == ne01); + GGML_ASSERT(ne2 == ne02); + GGML_ASSERT(ne3 == ne03); + + for (int64_t i3 = 0; i3 < ne03; i3++) { + for (int64_t i2 = 0; i2 < ne02; i2++) { + for (int64_t i1 = 0; i1 < ne01; i1++) { + float * src_row = (float *) ((char *) src0->data + i1*nb01 + i2*nb02 + i3*nb03); + float * dst_row = (float *) ((char *) dst->data + i1*nb1 + i2*nb2 + i3*nb3); + float row_sum = 0; + ggml_vec_sum_f32(ne00, &row_sum, src_row); + dst_row[0] = row_sum; + } + } + } +} + +void ggml_compute_forward_sum_rows( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_sum_rows_f32(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +// ggml_compute_forward_mean + +static void ggml_compute_forward_mean_f32( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + + if (params->ith != 0) { + return; + } + + assert(src0->nb[0] == sizeof(float)); + + GGML_TENSOR_UNARY_OP_LOCALS + + assert(ne0 == 1); + assert(ne1 == ne01); + assert(ne2 == ne02); + assert(ne3 == ne03); + + GGML_UNUSED(ne0); + GGML_UNUSED(ne1); + GGML_UNUSED(ne2); + GGML_UNUSED(ne3); + + for (int64_t i03 = 0; i03 < ne03; i03++) { + for (int64_t i02 = 0; i02 < ne02; i02++) { + for (int64_t i01 = 0; i01 < ne01; i01++) { + ggml_vec_sum_f32(ne00, + (float *) ((char *) dst->data + i01*nb1 + i02*nb2 + i03*nb3), + (float *) ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03)); + + *(float *) ((char *) dst->data + i01*nb1 + i02*nb2 + i03*nb3) /= (float) ne00; + } + } + } +} + +void ggml_compute_forward_mean( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_mean_f32(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +// ggml_compute_forward_argmax + +static void ggml_compute_forward_argmax_f32( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + + if (params->ith != 0) { + return; + } + + assert(src0->nb[0] == sizeof(float)); + assert(dst->nb[0] == sizeof(float)); + + const int64_t ne00 = src0->ne[0]; + const int64_t ne01 = src0->ne[1]; + + const size_t nb01 = src0->nb[1]; + const size_t nb0 = dst->nb[0]; + + for (int64_t i1 = 0; i1 < ne01; i1++) { + float * src = (float *) ((char *) src0->data + i1*nb01); + int32_t * dst_ = (int32_t *) ((char *) dst->data + i1*nb0); + int v = 0; + ggml_vec_argmax_f32(ne00, &v, src); + dst_[0] = v; + } +} + +void ggml_compute_forward_argmax( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_argmax_f32(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +// ggml_compute_forward_count_equal + +static void ggml_compute_forward_count_equal_i32( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + const ggml_tensor * src1 = dst->src[1]; + + GGML_TENSOR_BINARY_OP_LOCALS; + + GGML_ASSERT(src0->type == GGML_TYPE_I32); + GGML_ASSERT(src1->type == GGML_TYPE_I32); + GGML_ASSERT(ggml_are_same_shape(src0, src1)); + GGML_ASSERT(ggml_is_scalar(dst)); + GGML_ASSERT(dst->type == GGML_TYPE_I64); + + const int64_t nr = ggml_nrows(src0); + + const int ith = params->ith; + const int nth = params->nth; + + int64_t * sums = (int64_t *) params->wdata; + int64_t sum_thread = 0; + + // rows per thread + const int64_t dr = (nr + nth - 1)/nth; + + // row range for this thread + const int64_t ir0 = dr*ith; + const int64_t ir1 = MIN(ir0 + dr, nr); + + for (int64_t ir = ir0; ir < ir1; ++ir) { + const int64_t i03 = ir / (ne02*ne01); + const int64_t i02 = (ir - i03*ne03) / ne01; + const int64_t i01 = ir - i03*ne03 - i02*ne02; + + const char * data0 = (const char *) src0->data + i03*nb03 + i02*nb02 + i01*nb01; + const char * data1 = (const char *) src1->data + i03*nb13 + i02*nb12 + i01*nb11; + + for (int64_t i00 = 0; i00 < ne00; ++i00) { + const int32_t val0 = *((const int32_t *) (data0 + i00*nb00)); + const int32_t val1 = *((const int32_t *) (data1 + i00*nb10)); + + sum_thread += val0 == val1; + } + } + if (ith != 0) { + sums[ith] = sum_thread; + } + ggml_barrier(params->threadpool); + + if (ith != 0) { + return; + } + + for (int ith_other = 1; ith_other < nth; ++ith_other) { + sum_thread += sums[ith_other]; + } + *((int64_t *) dst->data) = sum_thread; +} + +void ggml_compute_forward_count_equal( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_I32: + { + ggml_compute_forward_count_equal_i32(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +// ggml_compute_forward_repeat + +static void ggml_compute_forward_repeat_f32( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + + if (params->ith != 0) { + return; + } + + GGML_ASSERT(ggml_can_repeat(src0, dst)); + + GGML_TENSOR_UNARY_OP_LOCALS + + // guaranteed to be an integer due to the check in ggml_can_repeat + const int nr0 = (int)(ne0/ne00); + const int nr1 = (int)(ne1/ne01); + const int nr2 = (int)(ne2/ne02); + const int nr3 = (int)(ne3/ne03); + + // TODO: support for transposed / permuted tensors + GGML_ASSERT(nb0 == sizeof(float)); + GGML_ASSERT(nb00 == sizeof(float)); + + // TODO: maybe this is not optimal? + for (int i3 = 0; i3 < nr3; i3++) { + for (int k3 = 0; k3 < ne03; k3++) { + for (int i2 = 0; i2 < nr2; i2++) { + for (int k2 = 0; k2 < ne02; k2++) { + for (int i1 = 0; i1 < nr1; i1++) { + for (int k1 = 0; k1 < ne01; k1++) { + for (int i0 = 0; i0 < nr0; i0++) { + ggml_vec_cpy_f32(ne00, + (float *) ((char *) dst->data + (i3*ne03 + k3)*nb3 + (i2*ne02 + k2)*nb2 + (i1*ne01 + k1)*nb1 + (i0*ne00)*nb0), + (float *) ((char *) src0->data + ( k3)*nb03 + ( k2)*nb02 + ( k1)*nb01)); + } + } + } + } + } + } + } +} + +static void ggml_compute_forward_repeat_f16( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + + if (params->ith != 0) { + return; + } + + GGML_ASSERT(ggml_can_repeat(src0, dst)); + + GGML_TENSOR_UNARY_OP_LOCALS + + // guaranteed to be an integer due to the check in ggml_can_repeat + const int nr0 = (int)(ne0/ne00); + const int nr1 = (int)(ne1/ne01); + const int nr2 = (int)(ne2/ne02); + const int nr3 = (int)(ne3/ne03); + + // TODO: support for transposed / permuted tensors + GGML_ASSERT(nb0 == sizeof(ggml_fp16_t)); + GGML_ASSERT(nb00 == sizeof(ggml_fp16_t)); + + // TODO: maybe this is not optimal? + for (int i3 = 0; i3 < nr3; i3++) { + for (int k3 = 0; k3 < ne03; k3++) { + for (int i2 = 0; i2 < nr2; i2++) { + for (int k2 = 0; k2 < ne02; k2++) { + for (int i1 = 0; i1 < nr1; i1++) { + for (int k1 = 0; k1 < ne01; k1++) { + for (int i0 = 0; i0 < nr0; i0++) { + ggml_fp16_t * y = (ggml_fp16_t *) ((char *) dst->data + (i3*ne03 + k3)*nb3 + (i2*ne02 + k2)*nb2 + (i1*ne01 + k1)*nb1 + (i0*ne00)*nb0); + ggml_fp16_t * x = (ggml_fp16_t *) ((char *) src0->data + ( k3)*nb03 + ( k2)*nb02 + ( k1)*nb01); + // ggml_vec_cpy_f16(ne00, y, x) + for (int i = 0; i < ne00; ++i) { + y[i] = x[i]; + } + } + } + } + } + } + } + } +} + +void ggml_compute_forward_repeat( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F16: + case GGML_TYPE_BF16: + case GGML_TYPE_I16: + { + ggml_compute_forward_repeat_f16(params, dst); + } break; + case GGML_TYPE_F32: + case GGML_TYPE_I32: + { + ggml_compute_forward_repeat_f32(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +// ggml_compute_forward_repeat_back + +static void ggml_compute_forward_repeat_back_f32( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + + if (params->ith != 0) { + return; + } + + GGML_ASSERT(ggml_can_repeat(dst, src0)); + + GGML_TENSOR_UNARY_OP_LOCALS + + // guaranteed to be an integer due to the check in ggml_can_repeat + const int nr0 = (int)(ne00/ne0); + const int nr1 = (int)(ne01/ne1); + const int nr2 = (int)(ne02/ne2); + const int nr3 = (int)(ne03/ne3); + + // TODO: support for transposed / permuted tensors + GGML_ASSERT(nb0 == sizeof(float)); + GGML_ASSERT(nb00 == sizeof(float)); + + if (ggml_is_contiguous(dst)) { + ggml_vec_set_f32(ne0*ne1*ne2*ne3, (float *)dst->data, 0); + } else { + for (int k3 = 0; k3 < ne3; k3++) { + for (int k2 = 0; k2 < ne2; k2++) { + for (int k1 = 0; k1 < ne1; k1++) { + ggml_vec_set_f32(ne0, + (float *) ((char *) dst->data + k1*nb1 + k2*nb2 + k3*nb3), + 0); + } + } + } + } + + // TODO: maybe this is not optimal? + for (int i3 = 0; i3 < nr3; i3++) { + for (int k3 = 0; k3 < ne3; k3++) { + for (int i2 = 0; i2 < nr2; i2++) { + for (int k2 = 0; k2 < ne2; k2++) { + for (int i1 = 0; i1 < nr1; i1++) { + for (int k1 = 0; k1 < ne1; k1++) { + for (int i0 = 0; i0 < nr0; i0++) { + ggml_vec_acc_f32(ne0, + (float *) ((char *) dst->data + ( k3)*nb3 + ( k2)*nb2 + ( k1)*nb1), + (float *) ((char *) src0->data + (i3*ne3 + k3)*nb03 + (i2*ne2 + k2)*nb02 + (i1*ne1 + k1)*nb01 + (i0*ne0)*nb00)); + } + } + } + } + } + } + } +} + +void ggml_compute_forward_repeat_back( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_repeat_back_f32(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +// ggml_compute_forward_concat + +static void ggml_compute_forward_concat_any( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + const ggml_tensor * src1 = dst->src[1]; + + const size_t len = ggml_type_size(src0->type); + + const int ith = params->ith; + const int nth = params->nth; + + GGML_TENSOR_BINARY_OP_LOCALS + + const int32_t dim = ggml_get_op_params_i32(dst, 0); + + GGML_ASSERT(dim >= 0 && dim < 4); + + int64_t o[4] = {0, 0, 0, 0}; + o[dim] = src0->ne[dim]; + + const char * x; + + // TODO: smarter multi-theading + for (int i3 = 0; i3 < ne3; i3++) { + for (int i2 = ith; i2 < ne2; i2 += nth) { + for (int i1 = 0; i1 < ne1; i1++) { + for (int i0 = 0; i0 < ne0; i0++) { + if (i0 < ne00 && i1 < ne01 && i2 < ne02 && i3 < ne03) { + x = (const char *)src0->data + (i0 )*nb00 + (i1 )*nb01 + (i2 )*nb02 + (i3 )*nb03; + } else { + x = (const char *)src1->data + (i0 - o[0])*nb10 + (i1 - o[1])*nb11 + (i2 - o[2])*nb12 + (i3 - o[3])*nb13; + } + + char * y = (char *)dst->data + i0*nb0 + i1*nb1 + i2*nb2 + i3*nb3; + + memcpy(y, x, len); + } + } + } + } +} + +static void ggml_compute_forward_concat_i8( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + const ggml_tensor * src1 = dst->src[1]; + + GGML_ASSERT(ggml_type_size(src0->type) == sizeof(int8_t)); + + const int ith = params->ith; + const int nth = params->nth; + + GGML_TENSOR_BINARY_OP_LOCALS + + const int32_t dim = ggml_get_op_params_i32(dst, 0); + + GGML_ASSERT(dim >= 0 && dim < 4); + + int64_t o[4] = {0, 0, 0, 0}; + o[dim] = src0->ne[dim]; + + const int8_t * x; + + // TODO: smarter multi-theading + for (int i3 = 0; i3 < ne3; i3++) { + for (int i2 = ith; i2 < ne2; i2 += nth) { + for (int i1 = 0; i1 < ne1; i1++) { + for (int i0 = 0; i0 < ne0; i0++) { + if (i0 < ne00 && i1 < ne01 && i2 < ne02 && i3 < ne03) { + x = (const int8_t *) ((const char *)src0->data + (i0 )*nb00 + (i1 )*nb01 + (i2 )*nb02 + (i3 )*nb03); + } else { + x = (const int8_t *) ((const char *)src1->data + (i0 - o[0])*nb10 + (i1 - o[1])*nb11 + (i2 - o[2])*nb12 + (i3 - o[3])*nb13); + } + + int8_t * y = (int8_t *)((char *)dst->data + i0*nb0 + i1*nb1 + i2*nb2 + i3*nb3); + + *y = *x; + } + } + } + } +} + +static void ggml_compute_forward_concat_f16( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + const ggml_tensor * src1 = dst->src[1]; + + GGML_ASSERT(ggml_type_size(src0->type) == sizeof(ggml_fp16_t)); + + const int ith = params->ith; + const int nth = params->nth; + + GGML_TENSOR_BINARY_OP_LOCALS + + const int32_t dim = ggml_get_op_params_i32(dst, 0); + + GGML_ASSERT(dim >= 0 && dim < 4); + + int64_t o[4] = {0, 0, 0, 0}; + o[dim] = src0->ne[dim]; + + const ggml_fp16_t * x; + + // TODO: smarter multi-theading + for (int i3 = 0; i3 < ne3; i3++) { + for (int i2 = ith; i2 < ne2; i2 += nth) { + for (int i1 = 0; i1 < ne1; i1++) { + for (int i0 = 0; i0 < ne0; i0++) { + if (i0 < ne00 && i1 < ne01 && i2 < ne02 && i3 < ne03) { + x = (const ggml_fp16_t *) ((const char *)src0->data + (i0 )*nb00 + (i1 )*nb01 + (i2 )*nb02 + (i3 )*nb03); + } else { + x = (const ggml_fp16_t *) ((const char *)src1->data + (i0 - o[0])*nb10 + (i1 - o[1])*nb11 + (i2 - o[2])*nb12 + (i3 - o[3])*nb13); + } + + ggml_fp16_t * y = (ggml_fp16_t *)((char *)dst->data + i0*nb0 + i1*nb1 + i2*nb2 + i3*nb3); + + *y = *x; + } + } + } + } +} + +static void ggml_compute_forward_concat_f32( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + const ggml_tensor * src1 = dst->src[1]; + + GGML_ASSERT(ggml_type_size(src0->type) == sizeof(float)); + + const int ith = params->ith; + const int nth = params->nth; + + GGML_TENSOR_BINARY_OP_LOCALS + + const int32_t dim = ggml_get_op_params_i32(dst, 0); + + GGML_ASSERT(dim >= 0 && dim < 4); + + int64_t o[4] = {0, 0, 0, 0}; + o[dim] = src0->ne[dim]; + + const float * x; + + // TODO: smarter multi-theading + for (int i3 = 0; i3 < ne3; i3++) { + for (int i2 = ith; i2 < ne2; i2 += nth) { + for (int i1 = 0; i1 < ne1; i1++) { + for (int i0 = 0; i0 < ne0; i0++) { + if (i0 < ne00 && i1 < ne01 && i2 < ne02 && i3 < ne03) { + x = (const float *) ((const char *)src0->data + (i0 )*nb00 + (i1 )*nb01 + (i2 )*nb02 + (i3 )*nb03); + } else { + x = (const float *) ((const char *)src1->data + (i0 - o[0])*nb10 + (i1 - o[1])*nb11 + (i2 - o[2])*nb12 + (i3 - o[3])*nb13); + } + + float * y = (float *)((char *)dst->data + i0*nb0 + i1*nb1 + i2*nb2 + i3*nb3); + + *y = *x; + } + } + } + } +} + +void ggml_compute_forward_concat( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F16: + case GGML_TYPE_BF16: + case GGML_TYPE_I16: + { + ggml_compute_forward_concat_f16(params, dst); + } break; + case GGML_TYPE_I8: + { + ggml_compute_forward_concat_i8(params, dst); + } break; + case GGML_TYPE_F32: + case GGML_TYPE_I32: + { + ggml_compute_forward_concat_f32(params, dst); + } break; + default: + { + ggml_compute_forward_concat_any(params, dst); + } + } +} + +// ggml_compute_forward_gelu + +static void ggml_compute_forward_gelu_f32( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + + assert(ggml_is_contiguous_1(src0)); + assert(ggml_is_contiguous_1(dst)); + assert(ggml_are_same_shape(src0, dst)); + + const int ith = params->ith; + const int nth = params->nth; + + const int nc = src0->ne[0]; + const int nr = ggml_nrows(src0); + + // rows per thread + const int dr = (nr + nth - 1)/nth; + + // row range for this thread + const int ir0 = dr*ith; + const int ir1 = MIN(ir0 + dr, nr); + + for (int i1 = ir0; i1 < ir1; i1++) { + ggml_vec_gelu_f32(nc, + (float *) ((char *) dst->data + i1*( dst->nb[1])), + (float *) ((char *) src0->data + i1*(src0->nb[1]))); + +#ifndef NDEBUG + for (int k = 0; k < nc; k++) { + const float x = ((float *) ((char *) dst->data + i1*( dst->nb[1])))[k]; + GGML_UNUSED(x); + assert(!isnan(x)); + assert(!isinf(x)); + } +#endif + } +} + +static void ggml_compute_forward_gelu_f16( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + + assert(ggml_is_contiguous_1(src0)); + assert(ggml_is_contiguous_1(dst)); + assert(ggml_are_same_shape(src0, dst)); + + const int ith = params->ith; + const int nth = params->nth; + + const int nc = src0->ne[0]; + const int nr = ggml_nrows(src0); + + // rows per thread + const int dr = (nr + nth - 1)/nth; + + // row range for this thread + const int ir0 = dr*ith; + const int ir1 = MIN(ir0 + dr, nr); + + for (int i1 = ir0; i1 < ir1; i1++) { + ggml_vec_gelu_f16(nc, + (ggml_fp16_t *) ((char *) dst->data + i1*( dst->nb[1])), + (ggml_fp16_t *) ((char *) src0->data + i1*(src0->nb[1]))); + +#ifndef NDEBUG + for (int k = 0; k < nc; k++) { + const ggml_fp16_t x = ((ggml_fp16_t *) ((char *) dst->data + i1*( dst->nb[1])))[k]; + const float v = GGML_FP16_TO_FP32(x); + GGML_UNUSED(v); + assert(!isnan(v)); + assert(!isinf(v)); + } +#endif + } +} + +static void ggml_compute_forward_gelu( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_gelu_f32(params, dst); + } break; + case GGML_TYPE_F16: + { + ggml_compute_forward_gelu_f16(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +// ggml_compute_forward_gelu_quick + +static void ggml_compute_forward_gelu_quick_f32( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + + assert(ggml_is_contiguous_1(src0)); + assert(ggml_is_contiguous_1(dst)); + assert(ggml_are_same_shape(src0, dst)); + + const int ith = params->ith; + const int nth = params->nth; + + const int nc = src0->ne[0]; + const int nr = ggml_nrows(src0); + + // rows per thread + const int dr = (nr + nth - 1)/nth; + + // row range for this thread + const int ir0 = dr*ith; + const int ir1 = MIN(ir0 + dr, nr); + + for (int i1 = ir0; i1 < ir1; i1++) { + ggml_vec_gelu_quick_f32(nc, + (float *) ((char *) dst->data + i1*( dst->nb[1])), + (float *) ((char *) src0->data + i1*(src0->nb[1]))); + +#ifndef NDEBUG + for (int k = 0; k < nc; k++) { + const float x = ((float *) ((char *) dst->data + i1*( dst->nb[1])))[k]; + GGML_UNUSED(x); + assert(!isnan(x)); + assert(!isinf(x)); + } +#endif + } +} + +static void ggml_compute_forward_gelu_quick_f16( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + + assert(ggml_is_contiguous_1(src0)); + assert(ggml_is_contiguous_1(dst)); + assert(ggml_are_same_shape(src0, dst)); + + const int ith = params->ith; + const int nth = params->nth; + + const int nc = src0->ne[0]; + const int nr = ggml_nrows(src0); + + // rows per thread + const int dr = (nr + nth - 1)/nth; + + // row range for this thread + const int ir0 = dr*ith; + const int ir1 = MIN(ir0 + dr, nr); + + for (int i1 = ir0; i1 < ir1; i1++) { + ggml_vec_gelu_quick_f16(nc, + (ggml_fp16_t *) ((char *) dst->data + i1*( dst->nb[1])), + (ggml_fp16_t *) ((char *) src0->data + i1*(src0->nb[1]))); + +#ifndef NDEBUG + for (int k = 0; k < nc; k++) { + const ggml_fp16_t x = ((ggml_fp16_t *) ((char *) dst->data + i1*( dst->nb[1])))[k]; + const float v = GGML_FP16_TO_FP32(x); + GGML_UNUSED(v); + assert(!isnan(v)); + assert(!isinf(v)); + } +#endif + } +} + +static void ggml_compute_forward_gelu_quick( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_gelu_quick_f32(params, dst); + } break; + case GGML_TYPE_F16: + { + ggml_compute_forward_gelu_quick_f16(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +// ggml_compute_forward_silu + +static void ggml_compute_forward_silu_f32( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + + assert(ggml_is_contiguous_1(src0)); + assert(ggml_is_contiguous_1(dst)); + assert(ggml_are_same_shape(src0, dst)); + + const int ith = params->ith; + const int nth = params->nth; + + const int nc = src0->ne[0]; + const int nr = ggml_nrows(src0); + + // rows per thread + const int dr = (nr + nth - 1)/nth; + + // row range for this thread + const int ir0 = dr*ith; + const int ir1 = MIN(ir0 + dr, nr); + + for (int i1 = ir0; i1 < ir1; i1++) { + ggml_vec_silu_f32(nc, + (float *) ((char *) dst->data + i1*( dst->nb[1])), + (float *) ((char *) src0->data + i1*(src0->nb[1]))); + +#ifndef NDEBUG + for (int k = 0; k < nc; k++) { + const float x = ((float *) ((char *) dst->data + i1*(dst->nb[1])))[k]; + GGML_UNUSED(x); + assert(!isnan(x)); + assert(!isinf(x)); + } +#endif + } +} + +static void ggml_compute_forward_silu_f16( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + + assert(ggml_is_contiguous_1(src0)); + assert(ggml_is_contiguous_1(dst)); + assert(ggml_are_same_shape(src0, dst)); + + const int ith = params->ith; + const int nth = params->nth; + + const int nc = src0->ne[0]; + const int nr = ggml_nrows(src0); + + // rows per thread + const int dr = (nr + nth - 1)/nth; + + // row range for this thread + const int ir0 = dr*ith; + const int ir1 = MIN(ir0 + dr, nr); + + for (int i1 = ir0; i1 < ir1; i1++) { + ggml_vec_silu_f16(nc, + (ggml_fp16_t *) ((char *) dst->data + i1*( dst->nb[1])), + (ggml_fp16_t *) ((char *) src0->data + i1*(src0->nb[1]))); + +#ifndef NDEBUG + for (int k = 0; k < nc; k++) { + const ggml_fp16_t x = ((ggml_fp16_t *) ((char *) dst->data + i1*(dst->nb[1])))[k]; + const float v = GGML_FP16_TO_FP32(x); + GGML_UNUSED(v); + assert(!isnan(v)); + assert(!isinf(v)); + } +#endif + } +} + +static void ggml_compute_forward_silu( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_silu_f32(params, dst); + } break; + case GGML_TYPE_F16: + { + ggml_compute_forward_silu_f16(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} +// ggml_compute_forward_leaky_relu + +static void ggml_compute_forward_leaky_relu_f32( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + + if (params->ith != 0) { + return; + } + + assert(ggml_is_contiguous_1(src0)); + assert(ggml_is_contiguous_1(dst)); + assert(ggml_are_same_shape(src0, dst)); + + const int n = ggml_nrows(src0); + const int nc = src0->ne[0]; + + float negative_slope; + memcpy(&negative_slope, dst->op_params, sizeof(float)); + + assert(dst->nb[0] == sizeof(float)); + assert(src0->nb[0] == sizeof(float)); + + for (int i = 0; i < n; i++) { + ggml_vec_leaky_relu_f32(nc, + (float *) ((char *) dst->data + i*( dst->nb[1])), + (float *) ((char *) src0->data + i*(src0->nb[1])), negative_slope); + } +} + +static void ggml_compute_forward_leaky_relu_f16( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + + if (params->ith != 0) { + return; + } + + assert(ggml_is_contiguous_1(src0)); + assert(ggml_is_contiguous_1(dst)); + assert(ggml_are_same_shape(src0, dst)); + + const int n = ggml_nrows(src0); + const int nc = src0->ne[0]; + + float negative_slope; + memcpy(&negative_slope, dst->op_params, sizeof(float)); + + assert(dst->nb[0] == sizeof(ggml_fp16_t)); + assert(src0->nb[0] == sizeof(ggml_fp16_t)); + + for (int i = 0; i < n; i++) { + ggml_vec_leaky_relu_f16(nc, + (ggml_fp16_t *) ((char *) dst->data + i*( dst->nb[1])), + (ggml_fp16_t *) ((char *) src0->data + i*(src0->nb[1])), negative_slope); + } +} + +void ggml_compute_forward_leaky_relu( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_leaky_relu_f32(params, dst); + } break; + case GGML_TYPE_F16: + { + ggml_compute_forward_leaky_relu_f16(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +// ggml_compute_forward_silu_back + +static void ggml_compute_forward_silu_back_f32( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * grad = dst->src[0]; + const ggml_tensor * src1 = dst->src[1]; + + assert(ggml_is_contiguous_1(grad)); + assert(ggml_is_contiguous_1(src1)); + assert(ggml_is_contiguous_1(dst)); + assert(ggml_are_same_shape(src1, dst)); + assert(ggml_are_same_shape(src1, grad)); + + const int ith = params->ith; + const int nth = params->nth; + + const int nc = src1->ne[0]; + const int nr = ggml_nrows(src1); + + // rows per thread + const int dr = (nr + nth - 1)/nth; + + // row range for this thread + const int ir0 = dr*ith; + const int ir1 = MIN(ir0 + dr, nr); + + for (int i1 = ir0; i1 < ir1; i1++) { + ggml_vec_silu_backward_f32(nc, + (float *) ((char *) dst->data + i1*( dst->nb[1])), + (float *) ((char *) src1->data + i1*(src1->nb[1])), + (float *) ((char *) grad->data + i1*(grad->nb[1]))); + +#ifndef NDEBUG + for (int k = 0; k < nc; k++) { + const float x = ((float *) ((char *) dst->data + i1*( dst->nb[1])))[k]; + GGML_UNUSED(x); + assert(!isnan(x)); + assert(!isinf(x)); + } +#endif + } +} + +static void ggml_compute_forward_silu_back_f16( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * grad = dst->src[0]; + const ggml_tensor * src1 = dst->src[1]; + + assert(ggml_is_contiguous_1(grad)); + assert(ggml_is_contiguous_1(src1)); + assert(ggml_is_contiguous_1(dst)); + assert(ggml_are_same_shape(src1, dst)); + assert(ggml_are_same_shape(src1, grad)); + + const int ith = params->ith; + const int nth = params->nth; + + const int nc = src1->ne[0]; + const int nr = ggml_nrows(src1); + + // rows per thread + const int dr = (nr + nth - 1)/nth; + + // row range for this thread + const int ir0 = dr*ith; + const int ir1 = MIN(ir0 + dr, nr); + + for (int i1 = ir0; i1 < ir1; i1++) { + ggml_vec_silu_backward_f16(nc, + (ggml_fp16_t *) ((char *) dst->data + i1*( dst->nb[1])), + (ggml_fp16_t *) ((char *) src1->data + i1*(src1->nb[1])), + (ggml_fp16_t *) ((char *) grad->data + i1*(grad->nb[1]))); + + #ifndef NDEBUG + for (int k = 0; k < nc; k++) { + const float x = ((ggml_fp16_t *) ((char *) dst->data + i1*( dst->nb[1])))[k]; + const float v = GGML_FP16_TO_FP32(x); + GGML_UNUSED(v); + assert(!isnan(v)); + assert(!isinf(v)); + } + #endif + } +} + +void ggml_compute_forward_silu_back( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_silu_back_f32(params, dst); + } break; + case GGML_TYPE_F16: + { + ggml_compute_forward_silu_back_f16(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +// ggml_compute_forward_norm + +static void ggml_compute_forward_norm_f32( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + + GGML_ASSERT(ggml_are_same_shape(src0, dst)); + + GGML_ASSERT(src0->nb[0] == sizeof(float)); + + const int ith = params->ith; + const int nth = params->nth; + + GGML_TENSOR_UNARY_OP_LOCALS + + float eps; + memcpy(&eps, dst->op_params, sizeof(float)); + + GGML_ASSERT(eps >= 0.0f); + + // TODO: optimize + for (int64_t i03 = 0; i03 < ne03; i03++) { + for (int64_t i02 = 0; i02 < ne02; i02++) { + for (int64_t i01 = ith; i01 < ne01; i01 += nth) { + const float * x = (float *) ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03); + + ggml_float sum = 0.0; + for (int64_t i00 = 0; i00 < ne00; i00++) { + sum += (ggml_float)x[i00]; + } + + float mean = sum/ne00; + + float * y = (float *) ((char *) dst->data + i01*nb1 + i02*nb2 + i03*nb3); + + ggml_float sum2 = 0.0; + for (int64_t i00 = 0; i00 < ne00; i00++) { + float v = x[i00] - mean; + y[i00] = v; + sum2 += (ggml_float)(v*v); + } + + float variance = sum2/ne00; + const float scale = 1.0f/sqrtf(variance + eps); + + ggml_vec_scale_f32(ne00, y, scale); + } + } + } +} + +void ggml_compute_forward_norm( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_norm_f32(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +// ggml_compute_forward_group_rms_norm + +static void ggml_compute_forward_rms_norm_f32( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + + GGML_ASSERT(ggml_are_same_shape(src0, dst)); + + GGML_ASSERT(src0->nb[0] == sizeof(float)); + + const int ith = params->ith; + const int nth = params->nth; + + GGML_TENSOR_UNARY_OP_LOCALS + + float eps; + memcpy(&eps, dst->op_params, sizeof(float)); + + GGML_ASSERT(eps >= 0.0f); + + // TODO: optimize + for (int64_t i03 = 0; i03 < ne03; i03++) { + for (int64_t i02 = 0; i02 < ne02; i02++) { + for (int64_t i01 = ith; i01 < ne01; i01 += nth) { + const float * x = (float *) ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03); + + ggml_float sum = 0.0; + for (int64_t i00 = 0; i00 < ne00; i00++) { + sum += (ggml_float)(x[i00] * x[i00]); + } + + const float mean = sum/ne00; + + float * y = (float *) ((char *) dst->data + i01*nb1 + i02*nb2 + i03*nb3); + + memcpy(y, x, ne00 * sizeof(float)); + // for (int i00 = 0; i00 < ne00; i00++) { + // y[i00] = x[i00]; + // } + + const float scale = 1.0f/sqrtf(mean + eps); + + ggml_vec_scale_f32(ne00, y, scale); + } + } + } +} + +void ggml_compute_forward_rms_norm( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_rms_norm_f32(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +static void ggml_compute_forward_rms_norm_back_f32( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; // gradients from forward pass output + const ggml_tensor * src1 = dst->src[1]; // src1 from forward pass + + GGML_ASSERT(ggml_are_same_shape(src0, dst) && ggml_are_same_shape(src0, src1)); + + GGML_ASSERT(src0->nb[0] == sizeof(float)); + GGML_ASSERT(src1->nb[0] == sizeof(float)); + + const int ith = params->ith; + const int nth = params->nth; + + GGML_TENSOR_BINARY_OP_LOCALS + + float eps; + memcpy(&eps, dst->op_params, sizeof(float)); + + // TODO: optimize + for (int64_t i03 = 0; i03 < ne03; i03++) { + for (int64_t i02 = 0; i02 < ne02; i02++) { + for (int64_t i01 = ith; i01 < ne01; i01 += nth) { + // src1 is same shape as src0 => same indices + const int64_t i11 = i01; + const int64_t i12 = i02; + const int64_t i13 = i03; + + const float * dz = (float *) ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03); + const float * x = (float *) ((char *) src1->data + i11*nb11 + i12*nb12 + i13*nb13); + + ggml_float sum_xx = 0.0; + ggml_float sum_xdz = 0.0; + + for (int64_t i00 = 0; i00 < ne00; i00++) { + sum_xx += (ggml_float)(x[i00] * x[i00]); + sum_xdz += (ggml_float)(x[i00] * dz[i00]); + } + + //const float mean = (float)(sum_xx)/ne00; + const float mean_eps = (float)(sum_xx)/ne00 + eps; + const float sum_eps = (float)(sum_xx) + eps*ne00; + //const float mean_xdz = (float)(sum_xdz)/ne00; + // we could cache rms from forward pass to improve performance. + // to do this implement ggml_rms and compose ggml_rms_norm using ggml_rms. + //const float rms = sqrtf(mean_eps); + const float rrms = 1.0f / sqrtf(mean_eps); + //const float scale = -rrms/(ne00 * mean_eps); // -1/(n*rms**3) + + { + // z = rms_norm(x) + // + // rms_norm(src1) = + // scale( + // src1, + // div( + // 1, + // sqrt( + // add( + // scale( + // sum( + // sqr( + // src1)), + // (1.0/N)), + // eps)))); + + // postorder: + // ## op args grad + // 00 param src1 grad[#00] + // 01 const 1 + // 02 sqr (#00) grad[#02] + // 03 sum (#02) grad[#03] + // 04 const 1/N + // 05 scale (#03, #04) grad[#05] + // 06 const eps + // 07 add (#05, #06) grad[#07] + // 08 sqrt (#07) grad[#08] + // 09 div (#01,#08) grad[#09] + // 10 scale (#00,#09) grad[#10] + // + // backward pass, given grad[#10] + // #10: scale + // grad[#00] += scale(grad[#10],#09) + // grad[#09] += sum(mul(grad[#10],#00)) + // #09: div + // grad[#08] += neg(mul(grad[#09], div(#09,#08))) + // #08: sqrt + // grad[#07] += mul(grad[#08], div(0.5, #08)) + // #07: add + // grad[#05] += grad[#07] + // #05: scale + // grad[#03] += scale(grad[#05],#04) + // #03: sum + // grad[#02] += repeat(grad[#03], #02) + // #02: + // grad[#00] += scale(mul(#00, grad[#02]), 2.0) + // + // substitute and simplify: + // grad[#00] = scale(grad(#10), #09) + scale(mul(#00, grad[#02]), 2.0) + // grad[#02] = repeat(grad[#03], #02) + // grad[#02] = repeat(scale(grad[#05],#04), #02) + // grad[#02] = repeat(scale(grad[#07],#04), #02) + // grad[#02] = repeat(scale(mul(grad[#08], div(0.5, #08)),#04), #02) + // grad[#02] = repeat(scale(mul(neg(mul(grad[#09], div(#09,#08))), div(0.5, #08)),#04), #02) + // grad[#02] = repeat(scale(mul(neg(mul(sum(mul(grad[#10],#00)), div(#09,#08))), div(0.5, #08)),#04), #02) + // grad[#02] = repeat(-(sum(mul(grad[#10],#00)) * div(#09,#08) * div(0.5, #08) * (1/N)), #02) + // grad[#02] = repeat(-(sum(mul(grad[#10],#00)) * div(div(#01,#08),#08) * div(0.5, #08) * (1/N)), #02) + // grad[#02] = repeat(-(sum(mul(grad[#10],#00)) * div(1,#08*#08) * div(0.5, #08) * (1/N)), #02) + // grad[#02] = repeat(-(sum(mul(grad[#10],#00)) * div(1,#07) * div(0.5, #08) * (1/N)), #02) + // grad[#00] = scale(grad(#10), #09) + scale(mul(#00, grad[#02]), 2.0) + // grad[#00] = scale(grad(#10), #09) + scale(mul(#00, repeat(-(sum(mul(grad[#10],#00)) * div(1,#07) * div(0.5, #08) * (1/N)), #02)), 2.0) + // grad[#00] = scale(grad(#10), #09) + scale(scale(#00, -(sum(mul(grad[#10],#00)) * div(1,#07) * div(0.5, #08) * (1/N))), 2.0) + // grad[#00] = scale(grad(#10), #09) + scale(#00, -(sum(mul(grad[#10],#00)) * div(1,#07) * div(1,#08) * (1/N))) + // grad[#00] = scale(grad(#10), #09) + scale(#00, sum(mul(grad[#10],#00)) * div(1,#07*#08) * (-1/N)) + // grad[#00] = scale(grad(#10), #09) + scale(#00, sum(mul(grad[#10],#00)) * div(1,#07*#08) * (-1/N)) + // grad[#00] = scale(grad(#10), #09) + scale(#00, sum(mul(grad[#10],#00)) * div(1,mean_eps*rms) * (-1/N)) + // grad[#00] = scale(grad(#10), #09) + scale(#00, sum(mul(grad[#10],#00)) * div(-1,rms*N*mean_eps)) + // grad[#00] = scale(grad(#10), #09) + scale(#00, sum(mul(grad[#10],#00)) * div(-1,rms*N*(sum_xx/N+eps))) + // grad[#00] = scale(grad(#10), #09) + scale(#00, sum(mul(grad[#10],#00)) * div(-1,rms*N*sum_xx+rms*N*eps)) + // grad[#00] = scale(dz, rrms) + scale(x, sum(mul(dz,x)) * div(-1,rms*N*mean_eps)) + // grad[#00] = scale(dz, rrms) + scale(x, sum_xdz * div(-1,rms*N*mean_eps)) + // a = b*c + d*e + // a = b*c*f/f + d*e*f/f + // a = (b*c*f + d*e*f)*(1/f) + // a = (b*c*(1/c) + d*e*(1/c))*(1/(1/c)) + // a = (b + d*e/c)*c + // b = dz, c = rrms, d = x, e = sum_xdz * div(-1,rms*N*mean_eps) + // a = (dz + x*sum_xdz * div(-1,rms*N*mean_eps)/rrms)*rrms + // a = (dz + x*sum_xdz * div(-1,rms*N*mean_eps)*rms)*rrms + // a = (dz + x*sum_xdz * div(-rms,rms*N*mean_eps))*rrms + // a = (dz + x*sum_xdz * div(-1,N*mean_eps))*rrms + // a = (dz + x*div(-sum_xdz,N*mean_eps))*rrms + // a = (dz + x*div(-mean_xdz,mean_eps))*rrms + // grad[#00] = scale(dz + scale(x, div(-mean_xdz,mean_eps)),rrms) + // grad[#00] = scale(dz + scale(x, -mean_xdz/mean_eps),rrms) + // dx = scale(dz + scale(x, -mean_xdz/mean_eps),rrms) + } + // dx = scale(dz + scale(x, -mean_xdz/mean_eps),rrms) + // post-order: + // dx := x + // dx := scale(dx,-mean_xdz/mean_eps) + // dx := add(dx, dz) + // dx := scale(dx, rrms) + float * dx = (float *) ((char *) dst->data + i01*nb1 + i02*nb2 + i03*nb3); + + // dx[i00] = (x*(-sum_xdz/sum_eps) + dz) / sqrtf(mean_eps) + ggml_vec_cpy_f32 (ne00, dx, x); + // ggml_vec_scale_f32(ne00, dx, -mean_xdz/mean_eps); + ggml_vec_scale_f32(ne00, dx, (float)(-sum_xdz)/sum_eps); + ggml_vec_acc_f32 (ne00, dx, dz); + ggml_vec_scale_f32(ne00, dx, rrms); + } + } + } +} + +void ggml_compute_forward_rms_norm_back( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_rms_norm_back_f32(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +// ggml_compute_forward_group_norm + +static void ggml_compute_forward_group_norm_f32( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + + GGML_ASSERT(ggml_are_same_shape(src0, dst)); + + GGML_ASSERT(src0->nb[0] == sizeof(float)); + + const int ith = params->ith; + const int nth = params->nth; + + GGML_TENSOR_UNARY_OP_LOCALS + + // TODO: optimize + + float eps; + memcpy(&eps, dst->op_params + 1, sizeof(float)); + + int n_channels = src0->ne[2]; + int n_groups = dst->op_params[0]; + int n_channels_per_group = (n_channels + n_groups - 1) / n_groups; + for (int i = ith; i < n_groups; i += nth) { + int start = i * n_channels_per_group; + int end = start + n_channels_per_group; + if (end > n_channels) { + end = n_channels; + } + int step = end - start; + + for (int64_t i03 = 0; i03 < ne03; i03++) { + ggml_float sum = 0.0; + for (int64_t i02 = start; i02 < end; i02++) { + for (int64_t i01 = 0; i01 < ne01; i01++) { + const float * x = (float *)((char *) src0->data + i01 * nb01 + i02 * nb02 + i03 * nb03); + + ggml_float sumr = 0.0; + for (int64_t i00 = 0; i00 < ne00; i00++) { + sumr += (ggml_float)x[i00]; + } + sum += sumr; + } + } + const float mean = sum / (ne00 * ne01 * step); + + ggml_float sum2 = 0.0; + for (int64_t i02 = start; i02 < end; i02++) { + for (int64_t i01 = 0; i01 < ne01; i01++) { + const float * x = (float *)((char *) src0->data + i01 * nb01 + i02 * nb02 + i03 * nb03); + + float * y = (float *)((char *) dst->data + i01 * nb1 + i02 * nb2 + i03 * nb3); + + ggml_float sumr = 0.0; + for (int64_t i00 = 0; i00 < ne00; i00++) { + float v = x[i00] - mean; + y[i00] = v; + sumr += (ggml_float)(v * v); + } + sum2 += sumr; + } + } + const float variance = sum2 / (ne00 * ne01 * step); + const float scale = 1.0f / sqrtf(variance + eps); + + for (int64_t i02 = start; i02 < end; i02++) { + for (int64_t i01 = 0; i01 < ne01; i01++) { + float * y = (float *)((char *) dst->data + i01 * nb1 + i02 * nb2 + i03 * nb3); + ggml_vec_scale_f32(ne00, y, scale); + } + } + } + } +} + +void ggml_compute_forward_group_norm( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_group_norm_f32(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +// ggml_compute_forward_l2_norm + +static void ggml_compute_forward_l2_norm_f32( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + + GGML_ASSERT(ggml_are_same_shape(src0, dst)); + + GGML_ASSERT(src0->nb[0] == sizeof(float)); + + const int ith = params->ith; + const int nth = params->nth; + + GGML_TENSOR_UNARY_OP_LOCALS + + float eps; + memcpy(&eps, dst->op_params, sizeof(float)); + + GGML_ASSERT(eps >= 0.0f); + + // TODO: optimize + for (int64_t i03 = 0; i03 < ne03; i03++) { + for (int64_t i02 = 0; i02 < ne02; i02++) { + for (int64_t i01 = ith; i01 < ne01; i01 += nth) { + const float * x = (float *) ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03); + + ggml_float sum = 0.0; + for (int64_t i00 = 0; i00 < ne00; i00++) { + sum += (ggml_float)(x[i00] * x[i00]); + } + + float * y = (float *) ((char *) dst->data + i01*nb1 + i02*nb2 + i03*nb3); + + memcpy(y, x, ne00 * sizeof(float)); + + const float scale = 1.0f/fmaxf(sqrtf(sum), eps); + + ggml_vec_scale_f32(ne00, y, scale); + } + } + } +} + +void ggml_compute_forward_l2_norm( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_l2_norm_f32(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +// ggml_compute_forward_out_prod + +static void ggml_compute_forward_out_prod_f32( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + const ggml_tensor * src1 = dst->src[1]; + + GGML_TENSOR_BINARY_OP_LOCALS + + GGML_ASSERT(dst->type == GGML_TYPE_F32); + GGML_ASSERT(src0->type == GGML_TYPE_F32); + GGML_ASSERT(src1->type == GGML_TYPE_F32); + + const int ith = params->ith; + const int nth = params->nth; + + GGML_ASSERT(ne0 == ne00); + GGML_ASSERT(ne1 == ne10); + GGML_ASSERT(ne2 == ne12); + GGML_ASSERT(ne3 == ne13); + + GGML_ASSERT(ne2 % ne02 == 0); + GGML_ASSERT(ne3 % ne03 == 0); + + // we don't support permuted src0 or src1 + GGML_ASSERT(nb00 == sizeof(float)); + + // dst cannot be transposed or permuted + GGML_ASSERT(nb0 == sizeof(float)); + // GGML_ASSERT(nb0 <= nb1); + // GGML_ASSERT(nb1 <= nb2); + // GGML_ASSERT(nb2 <= nb3); + + // nb01 >= nb00 - src0 is not transposed + // compute by src0 rows + + if (ith == 0) { + ggml_vec_set_f32(ne0*ne1*ne2*ne3, (float *)dst->data, 0); + } + ggml_barrier(params->threadpool); + + // dst[:,:,:,:] = 0 + // for i2,i3: + // for i1: + // for i01: + // for i0: + // dst[i0,i1,i2,i3] += src0[i0,i01,i2,i3] * src1[i1,i01,i2,i3] + + // parallelize by last three dimensions + + // total rows in dst + const int64_t nr = ne1*ne2*ne3; + + // rows per thread + const int64_t dr = (nr + nth - 1)/nth; + + // row range for this thread + const int64_t ir0 = dr*ith; + const int64_t ir1 = MIN(ir0 + dr, nr); + + // block-tiling attempt + const int64_t blck_0 = MAX(GGML_VEC_MAD_UNROLL, 32); + const int64_t blck_1 = 16; + + // dps == dst per src0, used for group query attention + const int64_t dps2 = ne2 / ne02; + const int64_t dps3 = ne3 / ne03; + + for (int64_t bir = ir0; bir < ir1; bir += blck_1) { + const int64_t bir1 = MIN(bir + blck_1, ir1); + for (int64_t bi01 = 0; bi01 < ne01; bi01 += blck_0) { + const int64_t bne01 = MIN(bi01 + blck_0, ne01); + for (int64_t ir = bir; ir < bir1; ++ir) { + // dst indices + const int64_t i3 = ir/(ne2*ne1); + const int64_t i2 = (ir - i3*ne2*ne1)/ne1; + const int64_t i1 = (ir - i3*ne2*ne1 - i2*ne1); + + const int64_t i02 = i2 / dps2; + const int64_t i03 = i3 / dps3; + + //const int64_t i10 = i1; + const int64_t i12 = i2; + const int64_t i13 = i3; + +#if GGML_VEC_MAD_UNROLL > 2 + const int64_t bne01_unroll = bne01 - (bne01 % GGML_VEC_MAD_UNROLL); + for (int64_t i01 = bi01; i01 < bne01_unroll; i01 += GGML_VEC_MAD_UNROLL) { + const int64_t i11 = i01; + + float * s0 = (float *) ((char *) src0->data + ( i01*nb01 + i02*nb02 + i03*nb03)); + float * s1 = (float *) ((char *) src1->data + (i1*nb10 + i11*nb11 + i12*nb12 + i13*nb13)); + float * d = (float *) ((char *) dst->data + ( i1*nb1 + i2*nb2 + i3*nb3)); + + ggml_vec_mad_f32_unroll(ne0, nb01, nb11, d, s0, s1); + } + for (int64_t i01 = bne01_unroll; i01 < bne01; ++i01) { + const int64_t i11 = i01; + + float * s0 = (float *) ((char *) src0->data + ( i01*nb01 + i02*nb02 + i03*nb03)); + float * s1 = (float *) ((char *) src1->data + (i1*nb10 + i11*nb11 + i12*nb12 + i13*nb13)); + float * d = (float *) ((char *) dst->data + ( i1*nb1 + i2*nb2 + i3*nb3)); + + ggml_vec_mad_f32(ne0, d, s0, *s1); + } +#else + for (int64_t i01 = bi01; i01 < bne01; ++i01) { + const int64_t i11 = i01; + + float * s0 = (float *) ((char *) src0->data + ( i01*nb01 + i02*nb02 + i03*nb03)); + float * s1 = (float *) ((char *) src1->data + (i1*nb10 + i11*nb11 + i12*nb12 + i13*nb13)); + float * d = (float *) ((char *) dst->data + ( i1*nb1 + i2*nb2 + i3*nb3)); + + ggml_vec_mad_f32(ne0, d, s0, *s1); + } +#endif + } + } + } +} + +static void ggml_compute_forward_out_prod_q_f32( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + const ggml_tensor * src1 = dst->src[1]; + + GGML_TENSOR_BINARY_OP_LOCALS; + + const int ith = params->ith; + const int nth = params->nth; + + const ggml_type type = src0->type; + ggml_to_float_t const dequantize_row_q = ggml_get_type_traits(type)->to_float; + + GGML_ASSERT(ne02 == ne12); + GGML_ASSERT(ne03 == ne13); + GGML_ASSERT(ne2 == ne12); + GGML_ASSERT(ne3 == ne13); + + // we don't support permuted src0 dim0 + GGML_ASSERT(nb00 == ggml_type_size(type)); + + // dst dim0 cannot be transposed or permuted + GGML_ASSERT(nb0 == sizeof(float)); + // GGML_ASSERT(nb0 <= nb1); + // GGML_ASSERT(nb1 <= nb2); + // GGML_ASSERT(nb2 <= nb3); + + GGML_ASSERT(ne0 == ne00); + GGML_ASSERT(ne1 == ne10); + GGML_ASSERT(ne2 == ne02); + GGML_ASSERT(ne3 == ne03); + + // nb01 >= nb00 - src0 is not transposed + // compute by src0 rows + + if (ith == 0) { + ggml_vec_set_f32(ne0*ne1*ne2*ne3, (float *)dst->data, 0); + } + ggml_barrier(params->threadpool); + + // parallelize by last three dimensions + + // total rows in dst + const int64_t nr = ne1*ne2*ne3; + + // rows per thread + const int64_t dr = (nr + nth - 1)/nth; + + // row range for this thread + const int64_t ir0 = dr*ith; + const int64_t ir1 = MIN(ir0 + dr, nr); + + // dst[:,:,:,:] = 0 + // for i2,i3: + // for i1: + // for i01: + // for i0: + // dst[i0,i1,i2,i3] += src0[i0,i01,i2,i3] * src1[i1,i01,i2,i3] + + float * wdata = (float *) params->wdata + (ne0 + CACHE_LINE_SIZE_F32) * ith; + + for (int64_t ir = ir0; ir < ir1; ++ir) { + // dst indices + const int64_t i3 = ir/(ne2*ne1); + const int64_t i2 = (ir - i3*ne2*ne1)/ne1; + const int64_t i1 = (ir - i3*ne2*ne1 - i2*ne1); + + const int64_t i02 = i2; + const int64_t i03 = i3; + + //const int64_t i10 = i1; + const int64_t i12 = i2; + const int64_t i13 = i3; + + for (int64_t i01 = 0; i01 < ne01; ++i01) { + const int64_t i11 = i01; + + float * s0 = (float *) ((char *) src0->data + ( i01*nb01 + i02*nb02 + i03*nb03)); + float * s1 = (float *) ((char *) src1->data + (i1*nb10 + i11*nb11 + i12*nb12 + i13*nb13)); + float * d = (float *) ((char *) dst->data + ( i1*nb1 + i2*nb2 + i3*nb3)); + + dequantize_row_q(s0, wdata, ne0); + ggml_vec_mad_f32(ne0, d, wdata, *s1); + } + } +} + +void ggml_compute_forward_out_prod( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_Q4_0: + case GGML_TYPE_Q4_1: + case GGML_TYPE_Q5_0: + case GGML_TYPE_Q5_1: + case GGML_TYPE_Q8_0: + case GGML_TYPE_Q2_K: + case GGML_TYPE_Q3_K: + case GGML_TYPE_Q4_K: + case GGML_TYPE_Q5_K: + case GGML_TYPE_Q6_K: + case GGML_TYPE_TQ1_0: + case GGML_TYPE_TQ2_0: + case GGML_TYPE_IQ2_XXS: + case GGML_TYPE_IQ2_XS: + case GGML_TYPE_IQ3_XXS: + case GGML_TYPE_IQ1_S: + case GGML_TYPE_IQ1_M: + case GGML_TYPE_IQ4_NL: + case GGML_TYPE_IQ4_XS: + case GGML_TYPE_IQ3_S: + case GGML_TYPE_IQ2_S: + { + ggml_compute_forward_out_prod_q_f32(params, dst); + } break; + case GGML_TYPE_F16: + { + GGML_ABORT("fatal error"); // todo + // ggml_compute_forward_out_prod_f16_f32(params, dst); + } + case GGML_TYPE_F32: + { + ggml_compute_forward_out_prod_f32(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +// ggml_compute_forward_scale + +static void ggml_compute_forward_scale_f32( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + + GGML_ASSERT(ggml_is_contiguous(src0)); + GGML_ASSERT(ggml_is_contiguous(dst)); + GGML_ASSERT(ggml_are_same_shape(src0, dst)); + + // scale factor + float v; + memcpy(&v, dst->op_params, sizeof(float)); + + const int ith = params->ith; + const int nth = params->nth; + + const int nc = src0->ne[0]; + const int nr = ggml_nrows(src0); + + // rows per thread + const int dr = (nr + nth - 1)/nth; + + // row range for this thread + const int ir0 = dr*ith; + const int ir1 = MIN(ir0 + dr, nr); + + const size_t nb01 = src0->nb[1]; + + const size_t nb1 = dst->nb[1]; + + for (int i1 = ir0; i1 < ir1; i1++) { + if (dst->data != src0->data) { + // src0 is same shape as dst => same indices + memcpy((char *)dst->data + i1*nb1, (char *)src0->data + i1*nb01, nc * sizeof(float)); + } + ggml_vec_scale_f32(nc, (float *) ((char *) dst->data + i1*nb1), v); + } +} + +void ggml_compute_forward_scale( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_scale_f32(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +// ggml_compute_forward_set + +static void ggml_compute_forward_set_f32( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + const ggml_tensor * src1 = dst->src[1]; + + GGML_ASSERT(ggml_are_same_shape(src0, dst)); + GGML_ASSERT(ggml_is_contiguous(dst) && ggml_is_contiguous(src0)); + + // view src0 and dst with these strides and data offset inbytes during set + // nb0 is implicitly element_size because src0 and dst are contiguous + size_t nb1 = ((int32_t *) dst->op_params)[0]; + size_t nb2 = ((int32_t *) dst->op_params)[1]; + size_t nb3 = ((int32_t *) dst->op_params)[2]; + size_t offset = ((int32_t *) dst->op_params)[3]; + bool inplace = (bool) ((int32_t *) dst->op_params)[4]; + + if (!inplace) { + if (params->ith == 0) { + // memcpy needs to be synchronized across threads to avoid race conditions. + // => do it in INIT phase + memcpy( + ((char *) dst->data), + ((char *) src0->data), + ggml_nbytes(dst)); + } + ggml_barrier(params->threadpool); + } + + const int ith = params->ith; + const int nth = params->nth; + + const int nr = ggml_nrows(src1); + const int nc = src1->ne[0]; + + GGML_TENSOR_LOCALS(int64_t, ne1, src1, ne) + GGML_TENSOR_LOCALS(size_t, nb1, src1, nb) + + // src0 and dst as viewed during set + const size_t nb0 = ggml_element_size(src0); + + const int im0 = (ne10 == 0 ? 0 : ne10-1); + const int im1 = (ne11 == 0 ? 0 : ne11-1); + const int im2 = (ne12 == 0 ? 0 : ne12-1); + const int im3 = (ne13 == 0 ? 0 : ne13-1); + + GGML_ASSERT(offset + im0*nb0 + im1*nb1 + im2*nb2 + im3*nb3 <= ggml_nbytes(dst)); + + GGML_ASSERT(nb10 == sizeof(float)); + + // rows per thread + const int dr = (nr + nth - 1)/nth; + + // row range for this thread + const int ir0 = dr*ith; + const int ir1 = MIN(ir0 + dr, nr); + + for (int ir = ir0; ir < ir1; ++ir) { + // src0 and dst are viewed with shape of src1 and offset + // => same indices + const int i3 = ir/(ne12*ne11); + const int i2 = (ir - i3*ne12*ne11)/ne11; + const int i1 = (ir - i3*ne12*ne11 - i2*ne11); + + ggml_vec_cpy_f32(nc, + (float *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + offset), + (float *) ((char *) src1->data + i3*nb13 + i2*nb12 + i1*nb11)); + } +} + +static void ggml_compute_forward_set_i32( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + const ggml_tensor * src1 = dst->src[1]; + + GGML_ASSERT(ggml_are_same_shape(src0, dst)); + GGML_ASSERT(ggml_is_contiguous(dst) && ggml_is_contiguous(src0)); + + // view src0 and dst with these strides and data offset inbytes during set + // nb0 is implicitly element_size because src0 and dst are contiguous + size_t nb1 = ((int32_t *) dst->op_params)[0]; + size_t nb2 = ((int32_t *) dst->op_params)[1]; + size_t nb3 = ((int32_t *) dst->op_params)[2]; + size_t offset = ((int32_t *) dst->op_params)[3]; + bool inplace = (bool) ((int32_t *) dst->op_params)[4]; + + if (!inplace) { + if (params->ith == 0) { + // memcpy needs to be synchronized across threads to avoid race conditions. + // => do it in INIT phase + memcpy( + ((char *) dst->data), + ((char *) src0->data), + ggml_nbytes(dst)); + } + ggml_barrier(params->threadpool); + } + + const int ith = params->ith; + const int nth = params->nth; + + const int nr = ggml_nrows(src1); + const int nc = src1->ne[0]; + + GGML_TENSOR_LOCALS(int64_t, ne1, src1, ne) + GGML_TENSOR_LOCALS(size_t, nb1, src1, nb) + + // src0 and dst as viewed during set + const size_t nb0 = ggml_element_size(src0); + + const int im0 = (ne10 == 0 ? 0 : ne10-1); + const int im1 = (ne11 == 0 ? 0 : ne11-1); + const int im2 = (ne12 == 0 ? 0 : ne12-1); + const int im3 = (ne13 == 0 ? 0 : ne13-1); + + GGML_ASSERT(offset + im0*nb0 + im1*nb1 + im2*nb2 + im3*nb3 <= ggml_nbytes(dst)); + + GGML_ASSERT(nb10 == sizeof(int32_t)); + + // rows per thread + const int dr = (nr + nth - 1)/nth; + + // row range for this thread + const int ir0 = dr*ith; + const int ir1 = MIN(ir0 + dr, nr); + + for (int ir = ir0; ir < ir1; ++ir) { + // src0 and dst are viewed with shape of src1 and offset + // => same indices + const int i3 = ir/(ne12*ne11); + const int i2 = (ir - i3*ne12*ne11)/ne11; + const int i1 = (ir - i3*ne12*ne11 - i2*ne11); + + ggml_vec_cpy_i32(nc, + (int32_t *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + offset), + (int32_t *) ((char *) src1->data + i3*nb13 + i2*nb12 + i1*nb11)); + } +} + +void ggml_compute_forward_set( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_set_f32(params, dst); + } break; + case GGML_TYPE_I32: + { + ggml_compute_forward_set_i32(params, dst); + } break; + case GGML_TYPE_F16: + case GGML_TYPE_BF16: + case GGML_TYPE_Q4_0: + case GGML_TYPE_Q4_1: + case GGML_TYPE_Q5_0: + case GGML_TYPE_Q5_1: + case GGML_TYPE_Q8_0: + case GGML_TYPE_Q8_1: + case GGML_TYPE_Q2_K: + case GGML_TYPE_Q3_K: + case GGML_TYPE_Q4_K: + case GGML_TYPE_Q5_K: + case GGML_TYPE_Q6_K: + case GGML_TYPE_TQ1_0: + case GGML_TYPE_TQ2_0: + case GGML_TYPE_IQ2_XXS: + case GGML_TYPE_IQ2_XS: + case GGML_TYPE_IQ3_XXS: + case GGML_TYPE_IQ1_S: + case GGML_TYPE_IQ1_M: + case GGML_TYPE_IQ4_NL: + case GGML_TYPE_IQ4_XS: + case GGML_TYPE_IQ3_S: + case GGML_TYPE_IQ2_S: + default: + { + GGML_ABORT("fatal error"); + } + } +} + +// ggml_compute_forward_cpy + +void ggml_compute_forward_cpy( + const ggml_compute_params * params, + ggml_tensor * dst) { + ggml_compute_forward_dup(params, dst); +} + +// ggml_compute_forward_cont + +void ggml_compute_forward_cont( + const ggml_compute_params * params, + ggml_tensor * dst) { + ggml_compute_forward_dup(params, dst); +} + +// ggml_compute_forward_reshape + +void ggml_compute_forward_reshape( + const ggml_compute_params * params, + ggml_tensor * dst) { + // NOP + GGML_UNUSED(params); + GGML_UNUSED(dst); +} + +// ggml_compute_forward_view + +void ggml_compute_forward_view( + const ggml_compute_params * params, + ggml_tensor * dst) { + // NOP + GGML_UNUSED(params); + GGML_UNUSED(dst); +} + +// ggml_compute_forward_permute + +void ggml_compute_forward_permute( + const ggml_compute_params * params, + ggml_tensor * dst) { + // NOP + GGML_UNUSED(params); + GGML_UNUSED(dst); +} + +// ggml_compute_forward_transpose + +void ggml_compute_forward_transpose( + const ggml_compute_params * params, + ggml_tensor * dst) { + // NOP + GGML_UNUSED(params); + GGML_UNUSED(dst); +} + +// ggml_compute_forward_get_rows + +static void ggml_compute_forward_get_rows_q( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + const ggml_tensor * src1 = dst->src[1]; + + GGML_TENSOR_BINARY_OP_LOCALS + + const int64_t nc = ne00; + const int64_t nr = ggml_nelements(src1); + + const ggml_type type = src0->type; + ggml_to_float_t const dequantize_row_q = ggml_get_type_traits(type)->to_float; + + assert(ne0 == nc); + assert(ne02 == ne11); + assert(nb00 == ggml_type_size(type)); + assert(ggml_nrows(dst) == nr); + + const int ith = params->ith; + const int nth = params->nth; + + // rows per thread + const int dr = (nr + nth - 1)/nth; + + // row range for this thread + const int ir0 = dr*ith; + const int ir1 = MIN(ir0 + dr, nr); + + for (int64_t i = ir0; i < ir1; ++i) { + const int64_t i12 = i/(ne11*ne10); + const int64_t i11 = (i - i12*ne11*ne10)/ne10; + const int64_t i10 = (i - i12*ne11*ne10 - i11*ne10); + const int64_t i01 = *(int32_t *) ((char *) src1->data + i10*nb10 + i11*nb11 + i12*nb12); + + GGML_ASSERT(i01 >= 0 && i01 < ne01); + + dequantize_row_q( + (const void *) ((char *) src0->data + i01*nb01 + i11*nb02 + i12*nb03), + (float *) ((char *) dst->data + i10*nb1 + i11*nb2 + i12*nb3), nc); + } +} + +static void ggml_compute_forward_get_rows_f16( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + const ggml_tensor * src1 = dst->src[1]; + + GGML_TENSOR_BINARY_OP_LOCALS + + const int64_t nc = ne00; + const int64_t nr = ggml_nelements(src1); + + assert(ne0 == nc); + assert(ne02 == ne11); + assert(nb00 == sizeof(ggml_fp16_t)); + assert(ggml_nrows(dst) == nr); + + const int ith = params->ith; + const int nth = params->nth; + + // rows per thread + const int dr = (nr + nth - 1)/nth; + + // row range for this thread + const int ir0 = dr*ith; + const int ir1 = MIN(ir0 + dr, nr); + + for (int64_t i = ir0; i < ir1; ++i) { + const int64_t i12 = i/(ne11*ne10); + const int64_t i11 = (i - i12*ne11*ne10)/ne10; + const int64_t i10 = (i - i12*ne11*ne10 - i11*ne10); + const int64_t i01 = *(int32_t *) ((char *) src1->data + i10*nb10 + i11*nb11 + i12*nb12); + + GGML_ASSERT(i01 >= 0 && i01 < ne01); + + ggml_fp16_to_fp32_row( + (const ggml_fp16_t*) ((char *) src0->data + i01*nb01 + i11*nb02 + i12*nb03), + (float *) ((char *) dst->data + i10*nb1 + i11*nb2 + i12*nb3), nc); + } +} + +static void ggml_compute_forward_get_rows_bf16( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + const ggml_tensor * src1 = dst->src[1]; + + GGML_TENSOR_BINARY_OP_LOCALS + + const int64_t nc = ne00; + const int64_t nr = ggml_nelements(src1); + + assert(ne0 == nc); + assert(ne02 == ne11); + assert(nb00 == sizeof(ggml_bf16_t)); + assert(ggml_nrows(dst) == nr); + + const int ith = params->ith; + const int nth = params->nth; + + // rows per thread + const int dr = (nr + nth - 1)/nth; + + // row range for this thread + const int ir0 = dr*ith; + const int ir1 = MIN(ir0 + dr, nr); + + for (int64_t i = ir0; i < ir1; ++i) { + const int64_t i12 = i/(ne11*ne10); + const int64_t i11 = (i - i12*ne11*ne10)/ne10; + const int64_t i10 = (i - i12*ne11*ne10 - i11*ne10); + const int64_t i01 = *(int32_t *) ((char *) src1->data + i10*nb10 + i11*nb11 + i12*nb12); + + GGML_ASSERT(i01 >= 0 && i01 < ne01); + + ggml_bf16_to_fp32_row( + (const ggml_bf16_t *) ((char *) src0->data + i01*nb01 + i11*nb02 + i12*nb03), + (float *) ((char *) dst->data + i10*nb1 + i11*nb2 + i12*nb3), nc); + } +} + +static void ggml_compute_forward_get_rows_f32( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + const ggml_tensor * src1 = dst->src[1]; + + GGML_TENSOR_BINARY_OP_LOCALS + + const int64_t nc = ne00; + const int64_t nr = ggml_nelements(src1); + + assert(ne0 == nc); + assert(ne02 == ne11); + assert(nb00 == sizeof(float)); + assert(ggml_nrows(dst) == nr); + + const int ith = params->ith; + const int nth = params->nth; + + // rows per thread + const int dr = (nr + nth - 1)/nth; + + // row range for this thread + const int ir0 = dr*ith; + const int ir1 = MIN(ir0 + dr, nr); + + for (int64_t i = ir0; i < ir1; ++i) { + const int64_t i12 = i/(ne11*ne10); + const int64_t i11 = (i - i12*ne11*ne10)/ne10; + const int64_t i10 = (i - i12*ne11*ne10 - i11*ne10); + const int64_t i01 = *(int32_t *) ((char *) src1->data + i10*nb10 + i11*nb11 + i12*nb12); + + GGML_ASSERT(i01 >= 0 && i01 < ne01); + + ggml_vec_cpy_f32(nc, + (float *) ((char *) dst->data + i10*nb1 + i11*nb2 + i12*nb3), + (float *) ((char *) src0->data + i01*nb01 + i11*nb02 + i12*nb03)); + } +} + +void ggml_compute_forward_get_rows( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_Q4_0: + case GGML_TYPE_Q4_1: + case GGML_TYPE_Q5_0: + case GGML_TYPE_Q5_1: + case GGML_TYPE_Q8_0: + case GGML_TYPE_Q8_1: + case GGML_TYPE_Q2_K: + case GGML_TYPE_Q3_K: + case GGML_TYPE_Q4_K: + case GGML_TYPE_Q5_K: + case GGML_TYPE_Q6_K: + case GGML_TYPE_TQ1_0: + case GGML_TYPE_TQ2_0: + case GGML_TYPE_IQ2_XXS: + case GGML_TYPE_IQ2_XS: + case GGML_TYPE_IQ3_XXS: + case GGML_TYPE_IQ1_S: + case GGML_TYPE_IQ1_M: + case GGML_TYPE_IQ4_NL: + case GGML_TYPE_IQ4_XS: + case GGML_TYPE_IQ3_S: + case GGML_TYPE_IQ2_S: + { + ggml_compute_forward_get_rows_q(params, dst); + } break; + case GGML_TYPE_F16: + { + ggml_compute_forward_get_rows_f16(params, dst); + } break; + case GGML_TYPE_BF16: + { + ggml_compute_forward_get_rows_bf16(params, dst); + } break; + case GGML_TYPE_F32: + case GGML_TYPE_I32: + { + ggml_compute_forward_get_rows_f32(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } + + //static bool first = true; + //printf("ne0 = %d, ne1 = %d, ne2 = %d\n", dst->ne[0], dst->ne[1], dst->ne[2]); + //if (first) { + // first = false; + //} else { + // for (int k = 0; k < dst->ne[1]; ++k) { + // for (int j = 0; j < dst->ne[0]/16; ++j) { + // for (int i = 0; i < 16; ++i) { + // printf("%8.4f ", ((float *) dst->data)[k*dst->ne[0] + j*16 + i]); + // } + // printf("\n"); + // } + // printf("\n"); + // } + // printf("\n"); + // exit(0); + //} +} + +// ggml_compute_forward_get_rows_back + +static void ggml_compute_forward_get_rows_back_f32_f16( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + const ggml_tensor * src1 = dst->src[1]; + + if (params->ith != 0) { + return; + } + + GGML_ASSERT(ggml_is_contiguous(dst)); + + // ggml_compute_forward_dup_same_cont(params, opt0, dst); + + memset(dst->data, 0, ggml_nbytes(dst)); + + const int nc = src0->ne[0]; + const int nr = ggml_nelements(src1); + + GGML_ASSERT( dst->ne[0] == nc); + GGML_ASSERT(src0->nb[0] == sizeof(ggml_fp16_t)); + + for (int i = 0; i < nr; ++i) { + const int r = ((int32_t *) src1->data)[i]; + + for (int j = 0; j < nc; ++j) { + ggml_fp16_t v = ((ggml_fp16_t *) ((char *) src0->data + i*src0->nb[1]))[j]; + ((float *) ((char *) dst->data + r*dst->nb[1]))[j] += GGML_FP16_TO_FP32(v); + } + } +} + +static void ggml_compute_forward_get_rows_back_f32( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + const ggml_tensor * src1 = dst->src[1]; + + if (params->ith != 0) { + return; + } + + GGML_ASSERT(ggml_is_contiguous(dst)); + + // ggml_compute_forward_dup_same_cont(params, opt0, dst); + + memset(dst->data, 0, ggml_nbytes(dst)); + + const int nc = src0->ne[0]; + const int nr = ggml_nelements(src1); + + GGML_ASSERT( dst->ne[0] == nc); + GGML_ASSERT(src0->nb[0] == sizeof(float)); + + for (int i = 0; i < nr; ++i) { + const int r = ((int32_t *) src1->data)[i]; + + ggml_vec_add_f32(nc, + (float *) ((char *) dst->data + r*dst->nb[1]), + (float *) ((char *) dst->data + r*dst->nb[1]), + (float *) ((char *) src0->data + i*src0->nb[1])); + } +} + +void ggml_compute_forward_get_rows_back( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F16: + { + ggml_compute_forward_get_rows_back_f32_f16(params, dst); + } break; + case GGML_TYPE_F32: + { + ggml_compute_forward_get_rows_back_f32(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } + + //static bool first = true; + //printf("ne0 = %d, ne1 = %d, ne2 = %d\n", dst->ne[0], dst->ne[1], dst->ne[2]); + //if (first) { + // first = false; + //} else { + // for (int k = 0; k < dst->ne[1]; ++k) { + // for (int j = 0; j < dst->ne[0]/16; ++j) { + // for (int i = 0; i < 16; ++i) { + // printf("%8.4f ", ((float *) dst->data)[k*dst->ne[0] + j*16 + i]); + // } + // printf("\n"); + // } + // printf("\n"); + // } + // printf("\n"); + // exit(0); + //} +} + +// ggml_compute_forward_diag + +static void ggml_compute_forward_diag_f32( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + + if (params->ith != 0) { + return; + } + + // TODO: handle transposed/permuted matrices + + GGML_TENSOR_UNARY_OP_LOCALS + + GGML_ASSERT(ne00 == ne0); + GGML_ASSERT(ne00 == ne1); + GGML_ASSERT(ne01 == 1); + GGML_ASSERT(ne02 == ne2); + GGML_ASSERT(ne03 == ne3); + + GGML_ASSERT(nb00 == sizeof(float)); + GGML_ASSERT(nb0 == sizeof(float)); + + for (int i3 = 0; i3 < ne3; i3++) { + for (int i2 = 0; i2 < ne2; i2++) { + for (int i1 = 0; i1 < ne1; i1++) { + float * d = (float *)((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1); + float * s = (float *)((char *) src0->data + i3*nb03 + i2*nb02); + for (int i0 = 0; i0 < i1; i0++) { + d[i0] = 0; + } + d[i1] = s[i1]; + for (int i0 = i1+1; i0 < ne0; i0++) { + d[i0] = 0; + } + } + } + } +} + +void ggml_compute_forward_diag( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_diag_f32(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +// ggml_compute_forward_diag_mask_inf + +static void ggml_compute_forward_diag_mask_f32( + const ggml_compute_params * params, + ggml_tensor * dst, + const float value) { + + const ggml_tensor * src0 = dst->src[0]; + + const int ith = params->ith; + const int nth = params->nth; + + const int n_past = ((int32_t *) dst->op_params)[0]; + const bool inplace = src0->data == dst->data; + + GGML_ASSERT(n_past >= 0); + + if (!inplace) { + if (ith == 0) { + // memcpy needs to be synchronized across threads to avoid race conditions. + // => do it in INIT phase + GGML_ASSERT(ggml_nelements(dst) == ggml_nelements(src0)); + GGML_ASSERT(ggml_is_contiguous(dst) && ggml_is_contiguous(src0)); + memcpy( + ((char *) dst->data), + ((char *) src0->data), + ggml_nbytes(dst)); + } + ggml_barrier(params->threadpool); + } + + // TODO: handle transposed/permuted matrices + + const int n = ggml_nrows(src0); + const int nc = src0->ne[0]; + const int nr = src0->ne[1]; + const int nz = n/nr; + + GGML_ASSERT( dst->nb[0] == sizeof(float)); + GGML_ASSERT(src0->nb[0] == sizeof(float)); + + for (int k = 0; k < nz; k++) { + for (int j = ith; j < nr; j += nth) { + for (int i = n_past; i < nc; i++) { + if (i > n_past + j) { + *(float *)((char *) dst->data + k*dst->nb[2] + j*dst->nb[1] + i*dst->nb[0]) = value; + } + } + } + } +} + +void ggml_compute_forward_diag_mask_inf( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_diag_mask_f32(params, dst, -INFINITY); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +void ggml_compute_forward_diag_mask_zero( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_diag_mask_f32(params, dst, 0); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +// ggml_compute_forward_soft_max + +static void ggml_compute_forward_soft_max_f32( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + const ggml_tensor * src1 = dst->src[1]; + + assert(ggml_is_contiguous(dst)); + assert(ggml_are_same_shape(src0, dst)); + + float scale = 1.0f; + float max_bias = 0.0f; + + memcpy(&scale, (float *) dst->op_params + 0, sizeof(float)); + memcpy(&max_bias, (float *) dst->op_params + 1, sizeof(float)); + + // TODO: handle transposed/permuted matrices + + const int ith = params->ith; + const int nth = params->nth; + + GGML_TENSOR_UNARY_OP_LOCALS + + //const int64_t ne11 = src1 ? src1->ne[1] : 1; + + // TODO: is this supposed to be ceil instead of floor? + // https://huggingface.co/mosaicml/mpt-7b/blob/main/attention.py#L370 + const uint32_t n_head = ne02; + const uint32_t n_head_log2 = 1u << (uint32_t) floor(log2(n_head)); + + const float m0 = powf(2.0f, -(max_bias ) / n_head_log2); + const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2); + + const int nc = src0->ne[0]; + const int nr = ggml_nrows(src0); + + // rows per thread + const int dr = (nr + nth - 1)/nth; + + // row range for this thread + const int ir0 = dr*ith; + const int ir1 = MIN(ir0 + dr, nr); + + float * wp = (float *) params->wdata + (nc + CACHE_LINE_SIZE_F32) * ith; + + const bool use_f16 = (src1 && src1->type == GGML_TYPE_F16); + + for (int i1 = ir0; i1 < ir1; i1++) { + // ALiBi + const uint32_t h = (i1/ne01)%ne02; // head + const float slope = (max_bias > 0.0f) ? h < n_head_log2 ? powf(m0, h + 1) : powf(m1, 2*(h - n_head_log2) + 1) : 1.0f; + + float * sp = (float *)((char *) src0->data + i1*src0->nb[1]); + float * dp = (float *)((char *) dst->data + i1*dst->nb[1]); + + // broadcast the mask across rows + ggml_fp16_t * mp_f16 = src1 ? (ggml_fp16_t *)((char *) src1->data) + (i1%ne01)*ne00 : NULL; + float * mp_f32 = src1 ? (float *)((char *) src1->data) + (i1%ne01)*ne00 : NULL; + + ggml_vec_cpy_f32 (nc, wp, sp); + ggml_vec_scale_f32(nc, wp, scale); + if (mp_f32) { + if (use_f16) { + for (int i = 0; i < nc; ++i) { + wp[i] += slope*GGML_FP16_TO_FP32(mp_f16[i]); + } + } else { + for (int i = 0; i < nc; ++i) { + wp[i] += slope*mp_f32[i]; + } + } + } + +#ifndef NDEBUG + for (int i = 0; i < nc; ++i) { + //printf("p[%d] = %f\n", i, p[i]); + assert(!isnan(wp[i])); + } +#endif + + float max = -INFINITY; + ggml_vec_max_f32(nc, &max, wp); + + ggml_float sum = ggml_vec_soft_max_f32(nc, dp, wp, max); + assert(sum > 0.0); + + sum = 1.0/sum; + ggml_vec_scale_f32(nc, dp, sum); + +#ifndef NDEBUG + for (int i = 0; i < nc; ++i) { + assert(!isnan(dp[i])); + assert(!isinf(dp[i])); + } +#endif + } +} + +void ggml_compute_forward_soft_max( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_soft_max_f32(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + + +// ggml_compute_forward_soft_max_ext_back + +static void ggml_compute_forward_soft_max_ext_back_f32( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + const ggml_tensor * src1 = dst->src[1]; + + GGML_ASSERT(ggml_is_contiguous(src0)); + GGML_ASSERT(ggml_is_contiguous(src1)); + GGML_ASSERT(ggml_is_contiguous(dst)); + GGML_ASSERT(ggml_are_same_shape(src0, dst)); + GGML_ASSERT(ggml_are_same_shape(src1, dst)); + + float scale = 1.0f; + float max_bias = 0.0f; + + memcpy(&scale, (const float *) dst->op_params + 0, sizeof(float)); + memcpy(&max_bias, (const float *) dst->op_params + 1, sizeof(float)); + + GGML_ASSERT(max_bias == 0.0f); + + // TODO: handle transposed/permuted matrices + + const int ith = params->ith; + const int nth = params->nth; + + const int nc = src0->ne[0]; + const int nr = ggml_nrows(src0); + + // rows per thread + const int dr = (nr + nth - 1)/nth; + + // row range for this thread + const int ir0 = dr*ith; + const int ir1 = MIN(ir0 + dr, nr); + + for (int i1 = ir0; i1 < ir1; i1++) { + float *dy = (float *)((char *) src0->data + i1*src0->nb[1]); + float *y = (float *)((char *) src1->data + i1*src1->nb[1]); + float *dx = (float *)((char *) dst->data + i1*dst->nb[1]); + +#ifndef NDEBUG + for (int i = 0; i < nc; ++i) { + //printf("p[%d] = %f\n", i, p[i]); + assert(!isnan(dy[i])); + assert(!isnan(y[i])); + } +#endif + // Jii = yi - yi*yi + // Jij = -yi*yj + // J = diag(y)-y.T*y + // dx = J * dy + // dxk = sum_i(Jki * dyi) + // dxk = sum_i(-yk*yi * dyi) - (-yk*yk)*dyk + (yk - yk*yk)*dyk + // dxk = sum_i(-yk*yi * dyi) + yk*yk*dyk + yk*dyk - yk*yk*dyk + // dxk = sum_i(-yk*yi * dyi) + yk*dyk + // dxk = -yk * sum_i(yi * dyi) + yk*dyk + // dxk = -yk * dot(y, dy) + yk*dyk + // dxk = yk * (- dot(y, dy) + dyk) + // dxk = yk * (dyk - dot(y, dy)) + // + // post-order: + // dot_y_dy := dot(y, dy) + // dx := dy + // dx := dx - dot_y_dy + // dx := dx * y + + // linear runtime, no additional memory + float dot_y_dy = 0; + ggml_vec_dot_f32 (nc, &dot_y_dy, 0, y, 0, dy, 0, 1); + ggml_vec_cpy_f32 (nc, dx, dy); + ggml_vec_acc1_f32 (nc, dx, -dot_y_dy); + ggml_vec_mul_f32 (nc, dx, dx, y); + ggml_vec_scale_f32(nc, dx, scale); + +#ifndef NDEBUG + for (int i = 0; i < nc; ++i) { + assert(!isnan(dx[i])); + assert(!isinf(dx[i])); + } +#endif + } +} + +void ggml_compute_forward_soft_max_ext_back( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_soft_max_ext_back_f32(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +// ggml_compute_forward_clamp + +static void ggml_compute_forward_clamp_f32( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + + float min; + float max; + memcpy(&min, (float *) dst->op_params + 0, sizeof(float)); + memcpy(&max, (float *) dst->op_params + 1, sizeof(float)); + + const int ith = params->ith; + const int nth = params->nth; + + const int n = ggml_nrows(src0); + const int nc = src0->ne[0]; + + const size_t nb00 = src0->nb[0]; + const size_t nb01 = src0->nb[1]; + + const size_t nb0 = dst->nb[0]; + const size_t nb1 = dst->nb[1]; + + GGML_ASSERT( nb0 == sizeof(float)); + GGML_ASSERT(nb00 == sizeof(float)); + + for (int j = ith; j < n; j += nth) { + float * dst_ptr = (float *) ((char *) dst->data + j*nb1); + float * src0_ptr = (float *) ((char *) src0->data + j*nb01); + + for (int i = 0; i < nc; i++) { + dst_ptr[i] = MAX(MIN(src0_ptr[i], max), min); + } + } +} + +static void ggml_compute_forward_clamp_f16( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + + float min; + float max; + memcpy(&min, (float *) dst->op_params + 0, sizeof(float)); + memcpy(&max, (float *) dst->op_params + 1, sizeof(float)); + + const int ith = params->ith; + const int nth = params->nth; + + const int n = ggml_nrows(src0); + const int nc = src0->ne[0]; + + const size_t nb00 = src0->nb[0]; + const size_t nb01 = src0->nb[1]; + + const size_t nb0 = dst->nb[0]; + const size_t nb1 = dst->nb[1]; + + GGML_ASSERT( nb0 == sizeof(ggml_fp16_t)); + GGML_ASSERT(nb00 == sizeof(ggml_fp16_t)); + + for (int j = ith; j < n; j += nth) { + ggml_fp16_t * dst_ptr = (ggml_fp16_t *) ((char *) dst->data + j*nb1); + ggml_fp16_t * src0_ptr = (ggml_fp16_t *) ((char *) src0->data + j*nb01); + + for (int i = 0; i < nc; i++) { + float v = GGML_FP16_TO_FP32(src0_ptr[i]); + dst_ptr[i] = GGML_FP32_TO_FP16(MAX(MIN(v, max), min)); + } + } +} + +void ggml_compute_forward_clamp( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_clamp_f32(params, dst); + } break; + case GGML_TYPE_F16: + { + ggml_compute_forward_clamp_f16(params, dst); + } break; + case GGML_TYPE_BF16: + case GGML_TYPE_Q4_0: + case GGML_TYPE_Q4_1: + case GGML_TYPE_Q5_0: + case GGML_TYPE_Q5_1: + case GGML_TYPE_Q8_0: + case GGML_TYPE_Q8_1: + case GGML_TYPE_Q2_K: + case GGML_TYPE_Q3_K: + case GGML_TYPE_Q4_K: + case GGML_TYPE_Q5_K: + case GGML_TYPE_Q6_K: + case GGML_TYPE_TQ1_0: + case GGML_TYPE_TQ2_0: + case GGML_TYPE_IQ2_XXS: + case GGML_TYPE_IQ2_XS: + case GGML_TYPE_IQ3_XXS: + case GGML_TYPE_IQ1_S: + case GGML_TYPE_IQ1_M: + case GGML_TYPE_IQ4_NL: + case GGML_TYPE_IQ4_XS: + case GGML_TYPE_IQ3_S: + case GGML_TYPE_IQ2_S: + case GGML_TYPE_Q8_K: + case GGML_TYPE_I8: + case GGML_TYPE_I16: + case GGML_TYPE_I32: + case GGML_TYPE_I64: + case GGML_TYPE_F64: + case GGML_TYPE_COUNT: + { + GGML_ABORT("fatal error"); + } + } +} + +// ggml_compute_forward_rope + +static float rope_yarn_ramp(const float low, const float high, const int i0) { + const float y = (i0 / 2 - low) / MAX(0.001f, high - low); + return 1 - MIN(1, MAX(0, y)); +} + +// YaRN algorithm based on LlamaYaRNScaledRotaryEmbedding.py from https://github.com/jquesnelle/yarn +// MIT licensed. Copyright (c) 2023 Jeffrey Quesnelle and Bowen Peng. +static void rope_yarn( + float theta_extrap, float freq_scale, float corr_dims[2], int64_t i0, float ext_factor, float mscale, + float * cos_theta, float * sin_theta) { + // Get n-d rotational scaling corrected for extrapolation + float theta_interp = freq_scale * theta_extrap; + float theta = theta_interp; + if (ext_factor != 0.0f) { + float ramp_mix = rope_yarn_ramp(corr_dims[0], corr_dims[1], i0) * ext_factor; + theta = theta_interp * (1 - ramp_mix) + theta_extrap * ramp_mix; + + // Get n-d magnitude scaling corrected for interpolation + mscale *= 1.0f + 0.1f * logf(1.0f / freq_scale); + } + *cos_theta = cosf(theta) * mscale; + *sin_theta = sinf(theta) * mscale; +} + +static void ggml_rope_cache_init( + float theta_base, float freq_scale, const float * freq_factors, float corr_dims[2], int64_t ne0, float ext_factor, float mscale, + float * cache, float sin_sign, float theta_scale) { + // ref: https://github.com/jquesnelle/yarn/blob/master/scaled_rope/LlamaYaRNScaledRotaryEmbedding.py + float theta = theta_base; + for (int64_t i0 = 0; i0 < ne0; i0 += 2) { + const float ff = freq_factors ? freq_factors[i0/2] : 1.0f; + rope_yarn( + theta/ff, freq_scale, corr_dims, i0, ext_factor, mscale, &cache[i0 + 0], &cache[i0 + 1] + ); + cache[i0 + 1] *= sin_sign; + + theta *= theta_scale; + } +} + +static void ggml_mrope_cache_init( + float theta_base_t, float theta_base_h, float theta_base_w, float theta_base_e, int sections[4], bool indep_sects, + float freq_scale, const float * freq_factors, float corr_dims[2], int64_t ne0, float ext_factor, float mscale, + float * cache, float sin_sign, float theta_scale) { + // ref: https://github.com/jquesnelle/yarn/blob/master/scaled_rope/LlamaYaRNScaledRotaryEmbedding.py + float theta_t = theta_base_t; + float theta_h = theta_base_h; + float theta_w = theta_base_w; + float theta_e = theta_base_e; // extra position id for vision encoder + int sect_dims = sections[0] + sections[1] + sections[2] + sections[3]; + int sec_w = sections[1] + sections[0]; + int sec_e = sections[2] + sec_w; + GGML_ASSERT(sect_dims <= ne0); + + for (int64_t i0 = 0; i0 < ne0; i0 += 2) { + const float ff = freq_factors ? freq_factors[i0/2] : 1.0f; + + int sector = (i0 / 2) % sect_dims; + if (indep_sects) { + // compute theta independently for each dim sections + // (i.e. reset corresponding theta when `i0` go from one section to another) + if (sector == 0) { + theta_t = theta_base_t; + } + else if (sector == sections[0]) { + theta_h = theta_base_h;; + } + else if (sector == sec_w) { + theta_w = theta_base_w; + } + else if (sector == sec_e) { + theta_e = theta_base_e; + } + } + + float theta = theta_t; + if (sector >= sections[0] && sector < sec_w) { + theta = theta_h; + } + else if (sector >= sec_w && sector < sec_w + sections[2]) { + theta = theta_w; + } + else if (sector >= sec_w + sections[2]) { + theta = theta_e; + } + + rope_yarn( + theta/ff, freq_scale, corr_dims, i0, ext_factor, mscale, &cache[i0 + 0], &cache[i0 + 1] + ); + cache[i0 + 1] *= sin_sign; + + theta_t *= theta_scale; + theta_w *= theta_scale; + theta_h *= theta_scale; + theta_e *= theta_scale; + } +} + +static void ggml_compute_forward_rope_f32( + const ggml_compute_params * params, + ggml_tensor * dst, + const bool forward) { + + const ggml_tensor * src0 = dst->src[0]; + const ggml_tensor * src1 = dst->src[1]; + const ggml_tensor * src2 = dst->src[2]; + + float freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow; + int sections[4]; + + //const int n_past = ((int32_t *) dst->op_params)[0]; + const int n_dims = ((int32_t *) dst->op_params)[1]; + const int mode = ((int32_t *) dst->op_params)[2]; + //const int n_ctx = ((int32_t *) dst->op_params)[3]; + const int n_ctx_orig = ((int32_t *) dst->op_params)[4]; + + memcpy(&freq_base, (int32_t *) dst->op_params + 5, sizeof(float)); + memcpy(&freq_scale, (int32_t *) dst->op_params + 6, sizeof(float)); + memcpy(&ext_factor, (int32_t *) dst->op_params + 7, sizeof(float)); + memcpy(&attn_factor, (int32_t *) dst->op_params + 8, sizeof(float)); + memcpy(&beta_fast, (int32_t *) dst->op_params + 9, sizeof(float)); + memcpy(&beta_slow, (int32_t *) dst->op_params + 10, sizeof(float)); + memcpy(§ions, (int32_t *) dst->op_params + 11, sizeof(int)*4); + + GGML_TENSOR_UNARY_OP_LOCALS + + //printf("ne0: %d, ne1: %d, ne2: %d, ne3: %d\n", ne0, ne1, ne2, ne3); + //printf("n_past = %d, ne2 = %d\n", n_past, ne2); + + GGML_ASSERT(nb00 == sizeof(float)); + + const int ith = params->ith; + const int nth = params->nth; + + const int nr = ggml_nrows(dst); + + GGML_ASSERT(n_dims <= ne0); + GGML_ASSERT(n_dims % 2 == 0); + + // rows per thread + const int dr = (nr + nth - 1)/nth; + + // row range for this thread + const int ir0 = dr*ith; + const int ir1 = MIN(ir0 + dr, nr); + + // row index used to determine which thread to use + int ir = 0; + + const float theta_scale = powf(freq_base, -2.0f/n_dims); + + float corr_dims[2]; + ggml_rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow, corr_dims); + + const bool is_neox = mode & GGML_ROPE_TYPE_NEOX; + const bool is_mrope = mode & GGML_ROPE_TYPE_MROPE; // ggml_rope_multi, multimodal rotary position embedding + const bool is_vision = mode == GGML_ROPE_TYPE_VISION; + + if (is_mrope) { + GGML_ASSERT(sections[0] > 0 || sections[1] > 0 || sections[2] > 0); + } + + if (is_vision) { + GGML_ASSERT(n_dims == ne0/2); + } + + const float * freq_factors = NULL; + if (src2 != NULL) { + GGML_ASSERT(src2->type == GGML_TYPE_F32); + GGML_ASSERT(src2->ne[0] >= n_dims / 2); + freq_factors = (const float *) src2->data; + } + + // backward process uses inverse rotation by cos and sin. + // cos and sin build a rotation matrix, where the inverse is the transpose. + // this essentially just switches the sign of sin. + const float sin_sign = forward ? 1.0f : -1.0f; + + const int32_t * pos = (const int32_t *) src1->data; + + for (int64_t i3 = 0; i3 < ne3; i3++) { // batch + for (int64_t i2 = 0; i2 < ne2; i2++) { // seq-len + + float * cache = (float *) params->wdata + (ne0 + CACHE_LINE_SIZE_F32)*ith; + if (!is_mrope) { + const int64_t p = pos[i2]; + ggml_rope_cache_init(p, freq_scale, freq_factors, corr_dims, ne0, ext_factor, attn_factor, cache, sin_sign, theta_scale); + } + else { + const int64_t p_t = pos[i2]; + const int64_t p_h = pos[i2 + ne2]; + const int64_t p_w = pos[i2 + ne2 * 2]; + const int64_t p_e = pos[i2 + ne2 * 3]; + ggml_mrope_cache_init( + p_t, p_h, p_w, p_e, sections, is_vision, + freq_scale, freq_factors, corr_dims, ne0, ext_factor, attn_factor, cache, sin_sign, theta_scale); + } + + for (int64_t i1 = 0; i1 < ne1; i1++) { // attn-heads + if (ir++ < ir0) continue; + if (ir > ir1) break; + + if (is_neox || is_mrope) { + if (is_vision){ + for (int64_t i0 = 0; i0 < n_dims; i0 += 2) { + const int64_t ic = i0/2; + + const float cos_theta = cache[i0 + 0]; + const float sin_theta = cache[i0 + 1]; + + const float * const src = (float *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + ic*nb00); + float * dst_data = (float *)((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + ic*nb0); + + const float x0 = src[0]; + const float x1 = src[n_dims]; + + dst_data[0] = x0*cos_theta - x1*sin_theta; + dst_data[n_dims] = x0*sin_theta + x1*cos_theta; + } + } else { + for (int64_t i0 = 0; i0 < n_dims; i0 += 2) { + const int64_t ic = i0/2; + + const float cos_theta = cache[i0 + 0]; + const float sin_theta = cache[i0 + 1]; + + const float * const src = (float *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + ic*nb00); + float * dst_data = (float *)((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + ic*nb0); + + const float x0 = src[0]; + const float x1 = src[n_dims/2]; + + dst_data[0] = x0*cos_theta - x1*sin_theta; + dst_data[n_dims/2] = x0*sin_theta + x1*cos_theta; + } + } + } else { + for (int64_t i0 = 0; i0 < n_dims; i0 += 2) { + const float cos_theta = cache[i0 + 0]; + const float sin_theta = cache[i0 + 1]; + + const float * const src = (float *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00); + float * dst_data = (float *)((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0); + + const float x0 = src[0]; + const float x1 = src[1]; + + dst_data[0] = x0*cos_theta - x1*sin_theta; + dst_data[1] = x0*sin_theta + x1*cos_theta; + } + } + + if (is_vision) { + for (int64_t i0 = n_dims; i0 < ne0; i0 += 2) { + const int64_t ic = i0/2; + + const float cos_theta = cache[i0 + 0]; + const float sin_theta = cache[i0 + 1]; + + const float * const src = (float *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + ic*nb00); + float * dst_data = (float *)((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + ic*nb0); + + const float x0 = src[0]; + const float x1 = src[n_dims]; + + dst_data[0] = x0*cos_theta - x1*sin_theta; + dst_data[n_dims] = x0*sin_theta + x1*cos_theta; + } + } else { + // fill the remain channels with data from src tensor + for (int64_t i0 = n_dims; i0 < ne0; i0 += 2) { + const float * const src = (float *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00); + float * dst_data = (float *)((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0); + + dst_data[0] = src[0]; + dst_data[1] = src[1]; + } + } + } + } + } +} + +// TODO: deduplicate f16/f32 code +static void ggml_compute_forward_rope_f16( + const ggml_compute_params * params, + ggml_tensor * dst, + const bool forward) { + + const ggml_tensor * src0 = dst->src[0]; + const ggml_tensor * src1 = dst->src[1]; + const ggml_tensor * src2 = dst->src[2]; + + float freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow; + int sections[4]; + + //const int n_past = ((int32_t *) dst->op_params)[0]; + const int n_dims = ((int32_t *) dst->op_params)[1]; + const int mode = ((int32_t *) dst->op_params)[2]; + //const int n_ctx = ((int32_t *) dst->op_params)[3]; + const int n_ctx_orig = ((int32_t *) dst->op_params)[4]; + memcpy(&freq_base, (int32_t *) dst->op_params + 5, sizeof(float)); + memcpy(&freq_scale, (int32_t *) dst->op_params + 6, sizeof(float)); + memcpy(&ext_factor, (int32_t *) dst->op_params + 7, sizeof(float)); + memcpy(&attn_factor, (int32_t *) dst->op_params + 8, sizeof(float)); + memcpy(&beta_fast, (int32_t *) dst->op_params + 9, sizeof(float)); + memcpy(&beta_slow, (int32_t *) dst->op_params + 10, sizeof(float)); + memcpy(§ions, (int32_t *) dst->op_params + 11, sizeof(int)*4); + + + GGML_TENSOR_UNARY_OP_LOCALS + + //printf("ne0: %d, ne1: %d, ne2: %d, ne3: %d\n", ne0, ne1, ne2, ne3); + //printf("n_past = %d, ne2 = %d\n", n_past, ne2); + + GGML_ASSERT(nb0 == sizeof(ggml_fp16_t)); + + const int ith = params->ith; + const int nth = params->nth; + + const int nr = ggml_nrows(dst); + + GGML_ASSERT(n_dims <= ne0); + GGML_ASSERT(n_dims % 2 == 0); + + // rows per thread + const int dr = (nr + nth - 1)/nth; + + // row range for this thread + const int ir0 = dr*ith; + const int ir1 = MIN(ir0 + dr, nr); + + // row index used to determine which thread to use + int ir = 0; + + const float theta_scale = powf(freq_base, -2.0f/n_dims); + + float corr_dims[2]; + ggml_rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow, corr_dims); + + const bool is_neox = mode & GGML_ROPE_TYPE_NEOX; + const bool is_mrope = mode & GGML_ROPE_TYPE_MROPE; + const bool is_vision = mode == GGML_ROPE_TYPE_VISION; + + if (is_mrope) { + GGML_ASSERT(sections[0] > 0 || sections[1] > 0 || sections[2] > 0); + } + + if (is_vision) { + GGML_ASSERT(n_dims == ne0/2); + } + + const float * freq_factors = NULL; + if (src2 != NULL) { + GGML_ASSERT(src2->type == GGML_TYPE_F32); + GGML_ASSERT(src2->ne[0] >= n_dims / 2); + freq_factors = (const float *) src2->data; + } + + // backward process uses inverse rotation by cos and sin. + // cos and sin build a rotation matrix, where the inverse is the transpose. + // this essentially just switches the sign of sin. + const float sin_sign = forward ? 1.0f : -1.0f; + + const int32_t * pos = (const int32_t *) src1->data; + + for (int64_t i3 = 0; i3 < ne3; i3++) { + for (int64_t i2 = 0; i2 < ne2; i2++) { + + float * cache = (float *) params->wdata + (ne0 + CACHE_LINE_SIZE_F32)*ith; + if (!is_mrope) { + const int64_t p = pos[i2]; + ggml_rope_cache_init(p, freq_scale, freq_factors, corr_dims, ne0, ext_factor, attn_factor, cache, sin_sign, theta_scale); + } + else { + const int64_t p_t = pos[i2]; + const int64_t p_h = pos[i2 + ne2]; + const int64_t p_w = pos[i2 + ne2 * 2]; + const int64_t p_e = pos[i2 + ne2 * 3]; + ggml_mrope_cache_init( + p_t, p_h, p_w, p_e, sections, is_vision, + freq_scale, freq_factors, corr_dims, ne0, ext_factor, attn_factor, cache, sin_sign, theta_scale); + } + + for (int64_t i1 = 0; i1 < ne1; i1++) { + if (ir++ < ir0) continue; + if (ir > ir1) break; + + if (is_neox || is_mrope) { + if (is_vision) { + for (int64_t i0 = 0; i0 < n_dims; i0 += 2) { + const int64_t ic = i0/2; + + const float cos_theta = cache[i0 + 0]; + const float sin_theta = cache[i0 + 1]; + + const ggml_fp16_t * const src = (ggml_fp16_t *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + ic*nb00); + ggml_fp16_t * dst_data = (ggml_fp16_t *)((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + ic*nb0); + + const float x0 = GGML_FP16_TO_FP32(src[0]); + const float x1 = GGML_FP16_TO_FP32(src[n_dims]); + + dst_data[0] = GGML_FP32_TO_FP16(x0*cos_theta - x1*sin_theta); + dst_data[n_dims] = GGML_FP32_TO_FP16(x0*sin_theta + x1*cos_theta); + } + } else { + for (int64_t i0 = 0; i0 < n_dims; i0 += 2) { + const int64_t ic = i0/2; + + const float cos_theta = cache[i0 + 0]; + const float sin_theta = cache[i0 + 1]; + + const ggml_fp16_t * const src = (ggml_fp16_t *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + ic*nb00); + ggml_fp16_t * dst_data = (ggml_fp16_t *)((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + ic*nb0); + + const float x0 = GGML_FP16_TO_FP32(src[0]); + const float x1 = GGML_FP16_TO_FP32(src[n_dims/2]); + + dst_data[0] = GGML_FP32_TO_FP16(x0*cos_theta - x1*sin_theta); + dst_data[n_dims/2] = GGML_FP32_TO_FP16(x0*sin_theta + x1*cos_theta); + } + } + } else { + for (int64_t i0 = 0; i0 < n_dims; i0 += 2) { + const float cos_theta = cache[i0 + 0]; + const float sin_theta = cache[i0 + 1]; + + const ggml_fp16_t * const src = (ggml_fp16_t *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00); + ggml_fp16_t * dst_data = (ggml_fp16_t *)((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0); + + const float x0 = GGML_FP16_TO_FP32(src[0]); + const float x1 = GGML_FP16_TO_FP32(src[1]); + + dst_data[0] = GGML_FP32_TO_FP16(x0*cos_theta - x1*sin_theta); + dst_data[1] = GGML_FP32_TO_FP16(x0*sin_theta + x1*cos_theta); + } + } + + if (is_vision) { + for (int64_t i0 = n_dims; i0 < ne0; i0 += 2) { + const int64_t ic = i0/2; + + const float cos_theta = cache[i0 + 0]; + const float sin_theta = cache[i0 + 1]; + + const ggml_fp16_t * const src = (ggml_fp16_t *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + ic*nb00); + ggml_fp16_t * dst_data = (ggml_fp16_t *)((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + ic*nb0); + + const float x0 = GGML_FP16_TO_FP32(src[0]); + const float x1 = GGML_FP16_TO_FP32(src[n_dims]); + + dst_data[0] = GGML_FP32_TO_FP16(x0*cos_theta - x1*sin_theta); + dst_data[n_dims] = GGML_FP32_TO_FP16(x0*sin_theta + x1*cos_theta); + } + } else { + for (int64_t i0 = n_dims; i0 < ne0; i0 += 2) { + const ggml_fp16_t * const src = (ggml_fp16_t *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00); + ggml_fp16_t * dst_data = (ggml_fp16_t *)((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0); + + dst_data[0] = src[0]; + dst_data[1] = src[1]; + } + } + } + } + } +} + +void ggml_compute_forward_rope( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F16: + { + ggml_compute_forward_rope_f16(params, dst, true); + } break; + case GGML_TYPE_F32: + { + ggml_compute_forward_rope_f32(params, dst, true); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +// ggml_compute_forward_rope_back + +void ggml_compute_forward_rope_back( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F16: + { + ggml_compute_forward_rope_f16(params, dst, false); + } break; + case GGML_TYPE_F32: + { + ggml_compute_forward_rope_f32(params, dst, false); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +// ggml_compute_forward_conv_transpose_1d + +static void ggml_compute_forward_conv_transpose_1d_f16_f32( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + const ggml_tensor * src1 = dst->src[1]; + + GGML_ASSERT(src0->type == GGML_TYPE_F16); + GGML_ASSERT(src1->type == GGML_TYPE_F32); + GGML_ASSERT( dst->type == GGML_TYPE_F32); + + GGML_TENSOR_BINARY_OP_LOCALS + + const int ith = params->ith; + const int nth = params->nth; + + const int nk = ne00*ne01*ne02; + + GGML_ASSERT(nb00 == sizeof(ggml_fp16_t)); + GGML_ASSERT(nb10 == sizeof(float)); + + if (ith == 0) { + memset(params->wdata, 0, params->wsize); + + // permute kernel data (src0) from (K x Cout x Cin) to (Cin x K x Cout) + { + ggml_fp16_t * const wdata = (ggml_fp16_t *) params->wdata + 0; + + for (int64_t i02 = 0; i02 < ne02; i02++) { + for (int64_t i01 = 0; i01 < ne01; i01++) { + const ggml_fp16_t * const src = (ggml_fp16_t *)((char *) src0->data + i02*nb02 + i01*nb01); + ggml_fp16_t * dst_data = wdata + i01*ne00*ne02; + for (int64_t i00 = 0; i00 < ne00; i00++) { + dst_data[i00*ne02 + i02] = src[i00]; + } + } + } + } + + // permute source data (src1) from (L x Cin) to (Cin x L) + { + ggml_fp16_t * const wdata = (ggml_fp16_t *) params->wdata + nk; + ggml_fp16_t * dst_data = wdata; + + for (int64_t i11 = 0; i11 < ne11; i11++) { + const float * const src = (float *)((char *) src1->data + i11*nb11); + for (int64_t i10 = 0; i10 < ne10; i10++) { + dst_data[i10*ne11 + i11] = GGML_FP32_TO_FP16(src[i10]); + } + } + } + + // need to zero dst since we are accumulating into it + memset(dst->data, 0, ggml_nbytes(dst)); + } + ggml_barrier(params->threadpool); + + const int32_t s0 = ((const int32_t*)(dst->op_params))[0]; + + // total rows in dst + const int nr = ne1; + + // rows per thread + const int dr = (nr + nth - 1)/nth; + + // row range for this thread + const int ir0 = dr*ith; + const int ir1 = MIN(ir0 + dr, nr); + + ggml_fp16_t * const wdata = (ggml_fp16_t *) params->wdata + 0; + ggml_fp16_t * const wdata_src = wdata + nk; + + for (int i1 = ir0; i1 < ir1; i1++) { + float * dst_data = (float *)((char *) dst->data + i1*nb1); + ggml_fp16_t * wdata_kernel = wdata + i1*ne02*ne00; + for (int i10 = 0; i10 < ne10; i10++) { + const int i1n = i10*ne11; + for (int i00 = 0; i00 < ne00; i00++) { + float v = 0; + ggml_vec_dot_f16(ne02, &v, 0, + (ggml_fp16_t *) wdata_src + i1n, 0, + (ggml_fp16_t *) wdata_kernel + i00*ne02, 0, 1); + dst_data[i10*s0 + i00] += v; + } + } + } +} + +static void ggml_compute_forward_conv_transpose_1d_f32( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + const ggml_tensor * src1 = dst->src[1]; + + GGML_ASSERT(src0->type == GGML_TYPE_F32); + GGML_ASSERT(src1->type == GGML_TYPE_F32); + GGML_ASSERT( dst->type == GGML_TYPE_F32); + + GGML_TENSOR_BINARY_OP_LOCALS + + const int ith = params->ith; + const int nth = params->nth; + + const int nk = ne00*ne01*ne02; + + GGML_ASSERT(nb00 == sizeof(float)); + GGML_ASSERT(nb10 == sizeof(float)); + + if (ith == 0) { + memset(params->wdata, 0, params->wsize); + + // prepare kernel data (src0) from (K x Cout x Cin) to (Cin x K x Cout) + { + float * const wdata = (float *) params->wdata + 0; + + for (int64_t i02 = 0; i02 < ne02; i02++) { + for (int64_t i01 = 0; i01 < ne01; i01++) { + const float * const src = (float *)((char *) src0->data + i02*nb02 + i01*nb01); + float * dst_data = wdata + i01*ne00*ne02; + for (int64_t i00 = 0; i00 < ne00; i00++) { + dst_data[i00*ne02 + i02] = src[i00]; + } + } + } + } + + // prepare source data (src1) + { + float * const wdata = (float *) params->wdata + nk; + float * dst_data = wdata; + + for (int64_t i11 = 0; i11 < ne11; i11++) { + const float * const src = (float *)((char *) src1->data + i11*nb11); + for (int64_t i10 = 0; i10 < ne10; i10++) { + dst_data[i10*ne11 + i11] = src[i10]; + } + } + } + + // need to zero dst since we are accumulating into it + memset(dst->data, 0, ggml_nbytes(dst)); + } + ggml_barrier(params->threadpool); + + const int32_t s0 = ((const int32_t*)(dst->op_params))[0]; + + // total rows in dst + const int nr = ne1; + + // rows per thread + const int dr = (nr + nth - 1)/nth; + + // row range for this thread + const int ir0 = dr*ith; + const int ir1 = MIN(ir0 + dr, nr); + + float * const wdata = (float *) params->wdata + 0; + float * const wdata_src = wdata + nk; + + for (int i1 = ir0; i1 < ir1; i1++) { + float * dst_data = (float *)((char *) dst->data + i1*nb1); + float * wdata_kernel = wdata + i1*ne02*ne00; + for (int i10 = 0; i10 < ne10; i10++) { + const int i1n = i10*ne11; + for (int i00 = 0; i00 < ne00; i00++) { + float v = 0; + ggml_vec_dot_f32(ne02, &v, 0, + wdata_src + i1n, 0, + wdata_kernel + i00*ne02, 0, 1); + dst_data[i10*s0 + i00] += v; + } + } + } +} + +void ggml_compute_forward_conv_transpose_1d( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F16: + { + ggml_compute_forward_conv_transpose_1d_f16_f32(params, dst); + } break; + case GGML_TYPE_F32: + { + ggml_compute_forward_conv_transpose_1d_f32(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +// ggml_compute_forward_im2col_f32 +// src0: kernel [OC, IC, KH, KW] +// src1: image [N, IC, IH, IW] +// dst: result [N, OH, OW, IC*KH*KW] +static void ggml_compute_forward_im2col_f32( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + const ggml_tensor * src1 = dst->src[1]; + + GGML_ASSERT(src1->type == GGML_TYPE_F32); + GGML_ASSERT( dst->type == GGML_TYPE_F32); + + GGML_TENSOR_BINARY_OP_LOCALS; + + const int32_t s0 = ((const int32_t *)(dst->op_params))[0]; + const int32_t s1 = ((const int32_t *)(dst->op_params))[1]; + const int32_t p0 = ((const int32_t *)(dst->op_params))[2]; + const int32_t p1 = ((const int32_t *)(dst->op_params))[3]; + const int32_t d0 = ((const int32_t *)(dst->op_params))[4]; + const int32_t d1 = ((const int32_t *)(dst->op_params))[5]; + const bool is_2D = ((const int32_t *)(dst->op_params))[6] == 1; + + const int ith = params->ith; + const int nth = params->nth; + + const int64_t N = is_2D ? ne13 : ne12; + const int64_t IC = is_2D ? ne12 : ne11; + const int64_t IH = is_2D ? ne11 : 1; + const int64_t IW = ne10; + + const int64_t KH = is_2D ? ne01 : 1; + const int64_t KW = ne00; + + const int64_t OH = is_2D ? ne2 : 1; + const int64_t OW = ne1; + + int ofs0 = is_2D ? nb13 : nb12; + int ofs1 = is_2D ? nb12 : nb11; + + GGML_ASSERT(nb10 == sizeof(float)); + + // im2col: [N, IC, IH, IW] => [N, OH, OW, IC*KH*KW] + { + float * const wdata = (float *) dst->data; + + for (int64_t in = 0; in < N; in++) { + for (int64_t ioh = 0; ioh < OH; ioh++) { // 1 + for (int64_t iow = 0; iow < OW; iow++) { + for (int64_t iic = ith; iic < IC; iic += nth) { + + // micro kernel + float * dst_data = wdata + (in*OH*OW + ioh*OW + iow)*(IC*KH*KW); // [IC, KH, KW] + const float * const src_data = (float *)((char *) src1->data + in*ofs0 + iic*ofs1); // [IH, IW] + + for (int64_t ikh = 0; ikh < KH; ikh++) { // 1 + for (int64_t ikw = 0; ikw < KW; ikw++) { + const int64_t iiw = iow*s0 + ikw*d0 - p0; + const int64_t iih = ioh*s1 + ikh*d1 - p1; + + if (iih < 0 || iih >= IH || iiw < 0 || iiw >= IW) { + dst_data[iic*(KH*KW) + ikh*KW + ikw] = 0; + } else { + dst_data[iic*(KH*KW) + ikh*KW + ikw] = (src_data[iih*IW + iiw]); + } + } + } + } + } + } + } + } +} + + +// ggml_compute_forward_im2col_f16 +// src0: kernel [OC, IC, KH, KW] +// src1: image [N, IC, IH, IW] +// dst: result [N, OH, OW, IC*KH*KW] +static void ggml_compute_forward_im2col_f16( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + const ggml_tensor * src1 = dst->src[1]; + + GGML_ASSERT(src0->type == GGML_TYPE_F16); + GGML_ASSERT(src1->type == GGML_TYPE_F32); + GGML_ASSERT( dst->type == GGML_TYPE_F16); + + GGML_TENSOR_BINARY_OP_LOCALS; + + const int32_t s0 = ((const int32_t *)(dst->op_params))[0]; + const int32_t s1 = ((const int32_t *)(dst->op_params))[1]; + const int32_t p0 = ((const int32_t *)(dst->op_params))[2]; + const int32_t p1 = ((const int32_t *)(dst->op_params))[3]; + const int32_t d0 = ((const int32_t *)(dst->op_params))[4]; + const int32_t d1 = ((const int32_t *)(dst->op_params))[5]; + const bool is_2D = ((const int32_t *)(dst->op_params))[6] == 1; + + const int ith = params->ith; + const int nth = params->nth; + + const int64_t N = is_2D ? ne13 : ne12; + const int64_t IC = is_2D ? ne12 : ne11; + const int64_t IH = is_2D ? ne11 : 1; + const int64_t IW = ne10; + + const int64_t KH = is_2D ? ne01 : 1; + const int64_t KW = ne00; + + const int64_t OH = is_2D ? ne2 : 1; + const int64_t OW = ne1; + + int ofs0 = is_2D ? nb13 : nb12; + int ofs1 = is_2D ? nb12 : nb11; + + GGML_ASSERT(nb00 == sizeof(ggml_fp16_t)); + GGML_ASSERT(nb10 == sizeof(float)); + + // im2col: [N, IC, IH, IW] => [N, OH, OW, IC*KH*KW] + { + ggml_fp16_t * const wdata = (ggml_fp16_t *) dst->data; + + for (int64_t in = 0; in < N; in++) { + for (int64_t ioh = 0; ioh < OH; ioh++) { // 1 + for (int64_t iow = 0; iow < OW; iow++) { + for (int64_t iic = ith; iic < IC; iic += nth) { + + // micro kernel + ggml_fp16_t * dst_data = wdata + (in*OH*OW + ioh*OW + iow)*(IC*KH*KW); // [IC, KH, KW] + const float * const src_data = (float *)((char *) src1->data + in*ofs0 + iic*ofs1); // [IH, IW] + + for (int64_t ikh = 0; ikh < KH; ikh++) { // 1 + for (int64_t ikw = 0; ikw < KW; ikw++) { + const int64_t iiw = iow*s0 + ikw*d0 - p0; + const int64_t iih = ioh*s1 + ikh*d1 - p1; + + if (iih < 0 || iih >= IH || iiw < 0 || iiw >= IW) { + dst_data[iic*(KH*KW) + ikh*KW + ikw] = 0; + } else { + dst_data[iic*(KH*KW) + ikh*KW + ikw] = GGML_FP32_TO_FP16(src_data[iih*IW + iiw]); + } + } + } + } + } + } + } + } +} + +void ggml_compute_forward_im2col( + const ggml_compute_params * params, + ggml_tensor * dst) { + switch (dst->type) { + case GGML_TYPE_F16: + { + ggml_compute_forward_im2col_f16(params, dst); + } break; + case GGML_TYPE_F32: + { + ggml_compute_forward_im2col_f32(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +// ggml_compute_forward_im2col_back_f32 + +void ggml_compute_forward_im2col_back_f32( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; // gradients of forward pass output + const ggml_tensor * src1 = dst->src[1]; // convolution kernel + + GGML_ASSERT(src0->type == GGML_TYPE_F32); + GGML_ASSERT(src1->type == GGML_TYPE_F32); + GGML_ASSERT( dst->type == GGML_TYPE_F32); + + GGML_TENSOR_BINARY_OP_LOCALS; + + const int32_t s0 = ((const int32_t *)(dst->op_params))[0]; + const int32_t s1 = ((const int32_t *)(dst->op_params))[1]; + const int32_t p0 = ((const int32_t *)(dst->op_params))[2]; + const int32_t p1 = ((const int32_t *)(dst->op_params))[3]; + const int32_t d0 = ((const int32_t *)(dst->op_params))[4]; + const int32_t d1 = ((const int32_t *)(dst->op_params))[5]; + const bool is_2D = ((const int32_t *)(dst->op_params))[6] == 1; + + const int ith = params->ith; + const int nth = params->nth; + + const int64_t N = is_2D ? ne3 : ne2; + const int64_t IC = is_2D ? ne2 : ne1; + const int64_t IH = is_2D ? ne1 : 1; + const int64_t IW = ne0; + + const int64_t KH = is_2D ? ne11 : 1; + const int64_t KW = ne10; + + const int64_t OH = is_2D ? ne02 : 1; + const int64_t OW = ne01; + + int ofs0 = is_2D ? nb3 : nb2; + int ofs1 = is_2D ? nb2 : nb1; + + GGML_ASSERT(nb0 == sizeof(float)); + + // im2col: [N, IC, IH, IW] => [N, OH, OW, IC*KH*KW] + { + float * const wdata = (float *) dst->data; + + for (int64_t in = 0; in < N; in++) { + for (int64_t iic = ith; iic < IC; iic += nth) { + for (int64_t iih = 0; iih < IH; iih++) { + for (int64_t iiw = 0; iiw < IW; iiw++) { + + // micro kernel + float grad = 0.0f; + for (int64_t ikh = 0; ikh < KH; ikh++) { + for (int64_t ikw = 0; ikw < KW; ikw++) { + // For s0 > 1 some values were skipped over in the forward pass. + // These values have tmpw % s0 != 0 and need to be skipped in the backwards pass as well. + const int64_t tmpw = (iiw + p0 - ikw*d0); + if (tmpw % s0 != 0) { + continue; + } + const int64_t iow = tmpw / s0; + + // Equivalent logic as above except for s1. + int64_t ioh; + if (is_2D) { + const int64_t tmph = iih + p1 - ikh*d1; + + if (tmph % s1 != 0) { + continue; + } + + ioh = tmph / s1; + } else { + ioh = 0; + } + + if (iow < 0 || iow >= OW || ioh < 0 || ioh >= OH) { + continue; + } + + const float * const grad_in = (const float *) src0->data + + (in*OH*OW + ioh*OW + iow)*(IC*KH*KW); // [IC, KH, KW] + grad += grad_in[iic*(KH*KW) + ikh*KW + ikw]; + } + } + float * dst_data = (float *)((char *) wdata + (in*ofs0 + iic*ofs1)); // [IH, IW] + dst_data[iih*IW + iiw] = grad; + } + } + } + } + } +} + +// ggml_compute_forward_conv_transpose_2d + +void ggml_compute_forward_conv_transpose_2d( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + const ggml_tensor * src1 = dst->src[1]; + + GGML_ASSERT(src0->type == GGML_TYPE_F16); + GGML_ASSERT(src1->type == GGML_TYPE_F32); + GGML_ASSERT( dst->type == GGML_TYPE_F32); + + GGML_TENSOR_BINARY_OP_LOCALS + + const int ith = params->ith; + const int nth = params->nth; + + const int nk = ne00*ne01*ne02*ne03; + + GGML_ASSERT(nb00 == sizeof(ggml_fp16_t)); + GGML_ASSERT(nb10 == sizeof(float)); + + if (ith == 0) { + memset(params->wdata, 0, params->wsize); + + // permute kernel data (src0) from (Kw x Kh x Cout x Cin) to (Cin x Kw x Kh x Cout) + { + ggml_fp16_t * const wdata = (ggml_fp16_t *) params->wdata + 0; + + for (int64_t i03 = 0; i03 < ne03; i03++) { + for (int64_t i02 = 0; i02 < ne02; i02++) { + const ggml_fp16_t * const src = (ggml_fp16_t *)((char *) src0->data + i03*nb03 + i02*nb02); + ggml_fp16_t * dst_data = wdata + i02*ne01*ne00*ne03; + for (int64_t i01 = 0; i01 < ne01; i01++) { + for (int64_t i00 = 0; i00 < ne00; i00++) { + dst_data[i01*ne00*ne03 + i00*ne03 + i03] = src[i01 * ne00 + i00]; + } + } + } + } + } + + // permute source data (src1) from (Sw x Sh x Cin) to (Cin x Sw x Sh) + { + ggml_fp16_t * const wdata = (ggml_fp16_t *) params->wdata + nk; + for (int i12 = 0; i12 < ne12; i12++) { + for (int i11 = 0; i11 < ne11; i11++) { + const float * const src = (float *)((char *) src1->data + i12*nb12 + i11*nb11); + ggml_fp16_t * dst_data = wdata + i11*ne10*ne12; + for (int i10 = 0; i10 < ne10; i10++) { + dst_data[i10*ne12 + i12] = GGML_FP32_TO_FP16(src[i10]); + } + } + } + } + + memset(dst->data, 0, ggml_nbytes(dst)); + } + ggml_barrier(params->threadpool); + + const int32_t stride = ggml_get_op_params_i32(dst, 0); + + // total patches in dst + const int np = ne2; + + // patches per thread + const int dp = (np + nth - 1)/nth; + + // patch range for this thread + const int ip0 = dp*ith; + const int ip1 = MIN(ip0 + dp, np); + + ggml_fp16_t * const wdata = (ggml_fp16_t *) params->wdata + 0; + ggml_fp16_t * const wdata_src = wdata + nk; + + for (int i2 = ip0; i2 < ip1; i2++) { // Cout + float * dst_data = (float *)((char *) dst->data + i2*nb2); + ggml_fp16_t * wdata_kernel = wdata + i2*ne01*ne00*ne03; + for (int i11 = 0; i11 < ne11; i11++) { + for (int i10 = 0; i10 < ne10; i10++) { + const int i1n = i11*ne10*ne12 + i10*ne12; + for (int i01 = 0; i01 < ne01; i01++) { + for (int i00 = 0; i00 < ne00; i00++) { + float v = 0; + ggml_vec_dot_f16(ne03, &v, 0, + wdata_src + i1n, 0, + wdata_kernel + i01*ne00*ne03 + i00*ne03, 0, 1); + dst_data[(i11*stride + i01)*ne0 + i10*stride + i00] += v; + } + } + } + } + } +} + +// ggml_compute_forward_pool_1d_sk_p0 + +static void ggml_compute_forward_pool_1d_sk_p0( + const ggml_compute_params * params, + const ggml_op_pool op, + const int k, + ggml_tensor * dst) { + + const ggml_tensor * src = dst->src[0]; + + assert(src->type == GGML_TYPE_F32 || src->type == GGML_TYPE_F16); + + if (params->ith != 0) { + return; + } + + const char * cdata = (const char *)src->data; + const char * const data_end = cdata + ggml_nbytes(src); + float * drow = (float *)dst->data; + + const int64_t rs = dst->ne[0]; + + while (cdata < data_end) { + const void * srow = (const void *)cdata; + int j = 0; + for (int64_t i = 0; i < rs; ++i) { + switch (op) { + case GGML_OP_POOL_AVG: drow[i] = 0; break; + case GGML_OP_POOL_MAX: drow[i] = -FLT_MAX; break; + case GGML_OP_POOL_COUNT: GGML_ABORT("fatal error"); + } + for (int ki = 0; ki < k; ++ki) { + const float srow_j = (src->type == GGML_TYPE_F32) ? ((const float*)srow)[j] : GGML_FP16_TO_FP32(((const ggml_fp16_t*)srow)[j]); + switch (op) { + case GGML_OP_POOL_AVG: drow[i] += srow_j; break; + case GGML_OP_POOL_MAX: if (srow_j > drow[i]) drow[i] = srow_j; break; + case GGML_OP_POOL_COUNT: GGML_ABORT("fatal error"); + } + ++j; + } + switch (op) { + case GGML_OP_POOL_AVG: drow[i] /= k; break; + case GGML_OP_POOL_MAX: break; + case GGML_OP_POOL_COUNT: GGML_ABORT("fatal error"); + } + } + + cdata += src->nb[1]; + drow += rs; + } +} + +// ggml_compute_forward_pool_1d + +void ggml_compute_forward_pool_1d( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const int32_t * opts = (const int32_t *)dst->op_params; + ggml_op_pool op = static_cast(opts[0]); + const int k0 = opts[1]; + const int s0 = opts[2]; + const int p0 = opts[3]; + GGML_ASSERT(p0 == 0); // padding not supported + GGML_ASSERT(k0 == s0); // only s = k supported + + ggml_compute_forward_pool_1d_sk_p0(params, op, k0, dst); +} + +// ggml_compute_forward_pool_2d + +void ggml_compute_forward_pool_2d( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src = dst->src[0]; + + assert(src->type == GGML_TYPE_F32 || src->type == GGML_TYPE_F16); + + if (params->ith != 0) { + return; + } + + const int32_t * opts = (const int32_t *)dst->op_params; + ggml_op_pool op = static_cast(opts[0]); + const int k0 = opts[1]; + const int k1 = opts[2]; + const int s0 = opts[3]; + const int s1 = opts[4]; + const int p0 = opts[5]; + const int p1 = opts[6]; + const char * cdata = (const char*)src->data; + const char * const data_end = cdata + ggml_nbytes(src); + + const int64_t px = dst->ne[0]; + const int64_t py = dst->ne[1]; + const int64_t pa = px * py; + + float * dplane = (float *)dst->data; + + const int ka = k0 * k1; + const int offset0 = -p0; + const int offset1 = -p1; + + while (cdata < data_end) { + for (int oy = 0; oy < py; ++oy) { + float * const drow = dplane + oy * px; + for (int ox = 0; ox < px; ++ox) { + float * const out = drow + ox; + switch (op) { + case GGML_OP_POOL_AVG: *out = 0; break; + case GGML_OP_POOL_MAX: *out = -FLT_MAX; break; + case GGML_OP_POOL_COUNT: GGML_ABORT("fatal error"); + } + + const int ix = offset0 + ox * s0; + const int iy = offset1 + oy * s1; + + for (int ky = 0; ky < k1; ++ky) { + if (iy + ky < 0 || iy + ky >= src->ne[1]) continue; + const void * srow = (const void *)(cdata + src->nb[1] * (iy + ky)); + for (int kx = 0; kx < k0; ++kx) { + int j = ix + kx; + if (j < 0 || j >= src->ne[0]) continue; + const float srow_j = (src->type == GGML_TYPE_F32) ? ((const float*)srow)[j] : GGML_FP16_TO_FP32(((const ggml_fp16_t*)srow)[j]); + switch (op) { + case GGML_OP_POOL_AVG: *out += srow_j; break; + case GGML_OP_POOL_MAX: if (srow_j > *out) *out = srow_j; break; + case GGML_OP_POOL_COUNT: GGML_ABORT("fatal error"); + } + } + } + switch (op) { + case GGML_OP_POOL_AVG: *out /= ka; break; + case GGML_OP_POOL_MAX: break; + case GGML_OP_POOL_COUNT: GGML_ABORT("fatal error"); + } + } + } + + cdata += src->nb[2]; + dplane += pa; + } +} + +// ggml_compute_forward_pool_2d_back + +void ggml_compute_forward_pool_2d_back( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src = dst->src[0]; + const ggml_tensor * dstf = dst->src[1]; // forward tensor of dst + + assert(dst->type == GGML_TYPE_F32 || dst->type == GGML_TYPE_F16); + + if (params->ith != 0) { + return; + } + + const int32_t * opts = (const int32_t *)dst->op_params; + ggml_op_pool op = static_cast(opts[0]); + const int k0 = opts[1]; + const int k1 = opts[2]; + const int s0 = opts[3]; + const int s1 = opts[4]; + const int p0 = opts[5]; + const int p1 = opts[6]; + + char * cdata = (char *) dst->data; + const char * cdataf = (const char *) dstf->data; + const char * const data_end = cdata + ggml_nbytes(dst); + + GGML_ASSERT(params->ith == 0); + memset(cdata, 0, ggml_nbytes(dst)); + + const int64_t px = src->ne[0]; + const int64_t py = src->ne[1]; + const int64_t pa = px * py; + + const float * splane = (const float *) src->data; + + const int ka = k0 * k1; + const int offset0 = -p0; + const int offset1 = -p1; + + while (cdata < data_end) { + for (int oy = 0; oy < py; ++oy) { + const float * const srow = splane + oy * px; + for (int ox = 0; ox < px; ++ox) { + const float grad0 = srow[ox]; + + const int ix = offset0 + ox * s0; + const int iy = offset1 + oy * s1; + + if (op == GGML_OP_POOL_MAX) { + float maxval = -FLT_MAX; + int kxmax = -1; + int kymax = -1; + + for (int ky = 0; ky < k1; ++ky) { + if (iy + ky < 0 || iy + ky >= dst->ne[1]) { + continue; + } + const void * drowf = (const void *)(cdataf + dst->nb[1] * (iy + ky)); + for (int kx = 0; kx < k0; ++kx) { + int j = ix + kx; + if (j < 0 || j >= dst->ne[0]) { + continue; + } + + const float val = dst->type == GGML_TYPE_F32 ? + ((const float *) drowf)[j] : GGML_FP16_TO_FP32(((const ggml_fp16_t *) drowf)[j]); + if (val <= maxval) { + continue; + } + + maxval = val; + kxmax = kx; + kymax = ky; + } + } + + if (kxmax == -1 || kymax == -1) { + continue; + } + + void * drow = (void *)(cdata + dst->nb[1] * (iy + kymax)); + const int j = ix + kxmax; + if (dst->type == GGML_TYPE_F32) { + ((float *) drow)[j] += grad0; + } else { + ((ggml_fp16_t *) drow)[j] = GGML_FP32_TO_FP16(grad0 + GGML_FP16_TO_FP32(((const ggml_fp16_t *) drow)[j])); + } + } else if (op == GGML_OP_POOL_AVG) { + const float grad = grad0 / ka; + + for (int ky = 0; ky < k1; ++ky) { + if (iy + ky < 0 || iy + ky >= dst->ne[1]) { + continue; + } + void * drow = (void *)(cdata + dst->nb[1] * (iy + ky)); + for (int kx = 0; kx < k0; ++kx) { + int j = ix + kx; + if (j < 0 || j >= dst->ne[0]) { + continue; + } + + if (dst->type == GGML_TYPE_F32) { + ((float *) drow)[j] += grad; + } else { + ((ggml_fp16_t *) drow)[j] += GGML_FP32_TO_FP16(grad); + } + } + } + } else { + GGML_ASSERT(false); + } + } + } + + cdata += dst->nb[2]; + cdataf += dst->nb[2]; + splane += pa; + } +} + +// ggml_compute_forward_upscale + +static void ggml_compute_forward_upscale_f32( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + + GGML_ASSERT(src0->type == GGML_TYPE_F32); + + const int ith = params->ith; + const int nth = params->nth; + + GGML_TENSOR_UNARY_OP_LOCALS + + const float sf0 = (float)ne0/src0->ne[0]; + const float sf1 = (float)ne1/src0->ne[1]; + const float sf2 = (float)ne2/src0->ne[2]; + const float sf3 = (float)ne3/src0->ne[3]; + + const ggml_scale_mode mode = (ggml_scale_mode) ggml_get_op_params_i32(dst, 0); + + if (mode == GGML_SCALE_MODE_NEAREST) { + for (int64_t i3 = 0; i3 < ne3; i3++) { + const int64_t i03 = i3 / sf3; + for (int64_t i2 = ith; i2 < ne2; i2 += nth) { + const int64_t i02 = i2 / sf2; + for (int64_t i1 = 0; i1 < ne1; i1++) { + const int64_t i01 = i1 / sf1; + for (int64_t i0 = 0; i0 < ne0; i0++) { + const int64_t i00 = i0 / sf0; + + const float * x = (float *)((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03); + float * y = (float *)((char *) dst->data + i0*nb0 + i1*nb1 + i2*nb2 + i3*nb3); + + *y = *x; + } + } + } + } + } else if (mode == GGML_SCALE_MODE_BILINEAR) { + // setting a pixel offset of 0 would replicate the behavior of pytorch interpolate with align_corners=True + const float pixel_offset = 0.5f; + + for (int64_t i3 = 0; i3 < ne3; i3++) { + const int64_t i03 = i3 / sf3; + for (int64_t i2 = ith; i2 < ne2; i2 += nth) { + const int64_t i02 = i2 / sf2; + for (int64_t i1 = 0; i1 < ne1; i1++) { + const float y = ((float)i1 + pixel_offset) / sf1 - pixel_offset; + int64_t y0 = (int64_t)floorf(y); + int64_t y1 = y0 + 1; + + y0 = std::max(int64_t(0), std::min(y0, ne01 - 1)); + y1 = std::max(int64_t(0), std::min(y1, ne01 - 1)); + + float dy = y - (float)y0; + dy = std::max(0.0f, std::min(dy, 1.0f)); + + for (int64_t i0 = 0; i0 < ne0; i0++) { + const float x = ((float)i0 + pixel_offset) / sf0 - pixel_offset; + int64_t x0 = (int64_t)floorf(x); + int64_t x1 = x0 + 1; + + x0 = std::max(int64_t(0), std::min(x0, ne00 - 1)); + x1 = std::max(int64_t(0), std::min(x1, ne00 - 1)); + + float dx = x - (float)x0; + dx = std::max(0.0f, std::min(dx, 1.0f)); + + // fetch the four surrounding pixel values and interpolate + const float a = *(const float *)((const char *)src0->data + x0*nb00 + y0*nb01 + i02*nb02 + i03*nb03); + const float b = *(const float *)((const char *)src0->data + x1*nb00 + y0*nb01 + i02*nb02 + i03*nb03); + const float c = *(const float *)((const char *)src0->data + x0*nb00 + y1*nb01 + i02*nb02 + i03*nb03); + const float d = *(const float *)((const char *)src0->data + x1*nb00 + y1*nb01 + i02*nb02 + i03*nb03); + + const float val = a*(1 - dx)*(1 - dy) + b*dx*(1 - dy) + c*(1 - dx)*dy + d*dx*dy; + + float * y_dst = (float *)((char *)dst->data + i0*nb0 + i1*nb1 + i2*nb2 + i3*nb3); + *y_dst = val; + } + } + } + } + } else { + GGML_ABORT("unsupported upscale mode"); + } +} + +void ggml_compute_forward_upscale( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_upscale_f32(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + + +// ggml_compute_forward_pad + +static void ggml_compute_forward_pad_f32( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + + GGML_ASSERT(src0->nb[0] == sizeof(float)); + GGML_ASSERT( dst->nb[0] == sizeof(float)); + + const int ith = params->ith; + const int nth = params->nth; + + GGML_TENSOR_UNARY_OP_LOCALS + + float * dst_ptr = (float *) dst->data; + + // TODO: optimize + + for (int64_t i2 = 0; i2 < ne2; ++i2) { + for (int64_t i1 = ith; i1 < ne1; i1 += nth) { + for (int64_t i0 = 0; i0 < ne0; ++i0) { + for (int64_t i3 = 0; i3 < ne3; ++i3) { + const int64_t dst_idx = i3*(ne0*ne1*ne2) + i2*(ne0*ne1) + i1*ne0 + i0; + + const float * src_ptr = (const float *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00); + + if (i0 < ne00 && i1 < ne01 && i2 < ne02 && i3 < ne03) { + dst_ptr[dst_idx] = *src_ptr; + } else { + dst_ptr[dst_idx] = 0; + } + } + } + } + } +} + +void ggml_compute_forward_pad( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_pad_f32(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +// ggml_compute_forward_pad_reflect_1d + +void ggml_compute_forward_pad_reflect_1d( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + + GGML_ASSERT(src0->type == GGML_TYPE_F32); + GGML_ASSERT( dst->type == GGML_TYPE_F32); + + const int ith = params->ith; + const int nth = params->nth; + + const int32_t * opts = (const int32_t *) dst->op_params; + const int p0 = opts[0]; + const int p1 = opts[1]; + + GGML_TENSOR_UNARY_OP_LOCALS + + for (int64_t i3 = 0; i3 < ne3; i3++) { + for (int64_t i2 = 0; i2 < ne2; i2++) { + for (int64_t i1 = ith; i1 < ne1; i1 += nth) { + float * left = (float *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + p0*nb0); + float * right = (float *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + (ne0-p1-1)*nb0); + + ggml_vec_cpy_f32(ne00, left, (float *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01)); + + for (int i0 = 1; i0 <= p0; i0++) { left[-i0] = left[i0]; } + for (int i0 = 1; i0 <= p1; i0++) { right[i0] = right[-i0]; } + } + } + } +} + +// ggml_compute_forward_arange + +static void ggml_compute_forward_arange_f32( + const ggml_compute_params * params, + ggml_tensor * dst) { + + GGML_ASSERT(dst->nb[0] == sizeof(float)); + + const int ith = params->ith; + const int nth = params->nth; + + const float start = ggml_get_op_params_f32(dst, 0); + const float stop = ggml_get_op_params_f32(dst, 1); + const float step = ggml_get_op_params_f32(dst, 2); + + const int64_t steps = (int64_t) ceilf((stop - start) / step); + + GGML_ASSERT(ggml_nelements(dst) == steps); + + for (int64_t i = ith; i < steps; i+= nth) { + float value = start + step * i; + ((float *)dst->data)[i] = value; + } +} + +void ggml_compute_forward_arange( + const ggml_compute_params * params, + ggml_tensor * dst) { + switch (dst->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_arange_f32(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +static void ggml_compute_forward_timestep_embedding_f32( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + + GGML_ASSERT(src0->nb[0] == sizeof(float)); + + const int ith = params->ith; + const int nth = params->nth; + + GGML_TENSOR_UNARY_OP_LOCALS + + const int dim = ggml_get_op_params_i32(dst, 0); + const int max_period = ggml_get_op_params_i32(dst, 1); + + int half = dim / 2; + + for (int64_t i = 0; i < ne00; i++) { + float * embed_data = (float *)((char *) dst->data + i*nb1); + for (int64_t j = ith; j < half; j += nth) { + float timestep = ((float *)src0->data)[i]; + float freq = (float)expf(-logf(max_period) * j / half); + float arg = timestep * freq; + embed_data[j] = cosf(arg); + embed_data[j + half] = sinf(arg); + } + if (dim % 2 != 0 && ith == 0) { + embed_data[dim] = 0.f; + } + } +} + +void ggml_compute_forward_timestep_embedding( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_timestep_embedding_f32(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +// ggml_compute_forward_argsort + +static void ggml_compute_forward_argsort_f32( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + + GGML_TENSOR_UNARY_OP_LOCALS + + GGML_ASSERT(nb0 == sizeof(float)); + + const int ith = params->ith; + const int nth = params->nth; + + const int64_t nr = ggml_nrows(src0); + + ggml_sort_order order = (ggml_sort_order) ggml_get_op_params_i32(dst, 0); + + for (int64_t i = ith; i < nr; i += nth) { + int32_t * dst_data = (int32_t *)((char *) dst->data + i*nb1); + const float * src_data = (float *)((char *) src0->data + i*nb01); + + for (int64_t j = 0; j < ne0; j++) { + dst_data[j] = j; + } + + // C doesn't have a functional sort, so we do a bubble sort instead + for (int64_t j = 0; j < ne0; j++) { + for (int64_t k = j + 1; k < ne0; k++) { + if ((order == GGML_SORT_ORDER_ASC && src_data[dst_data[j]] > src_data[dst_data[k]]) || + (order == GGML_SORT_ORDER_DESC && src_data[dst_data[j]] < src_data[dst_data[k]])) { + int32_t tmp = dst_data[j]; + dst_data[j] = dst_data[k]; + dst_data[k] = tmp; + } + } + } + } +} + +void ggml_compute_forward_argsort( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_argsort_f32(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +// ggml_compute_forward_flash_attn_ext + +static void ggml_compute_forward_flash_attn_ext_f16( + const ggml_compute_params * params, + const ggml_tensor * q, + const ggml_tensor * k, + const ggml_tensor * v, + const ggml_tensor * mask, + ggml_tensor * dst) { + + GGML_TENSOR_LOCALS(int64_t, neq, q, ne) + GGML_TENSOR_LOCALS(size_t, nbq, q, nb) + GGML_TENSOR_LOCALS(int64_t, nek, k, ne) + GGML_TENSOR_LOCALS(size_t, nbk, k, nb) + GGML_TENSOR_LOCALS(int64_t, nev, v, ne) + GGML_TENSOR_LOCALS(size_t, nbv, v, nb) + GGML_TENSOR_LOCALS(int64_t, ne, dst, ne) + GGML_TENSOR_LOCALS(size_t, nb, dst, nb) + + const int ith = params->ith; + const int nth = params->nth; + + const int64_t DK = nek0; + const int64_t DV = nev0; + const int64_t N = neq1; + + GGML_ASSERT(ne0 == DV); + GGML_ASSERT(ne2 == N); + + // input tensor rows must be contiguous + GGML_ASSERT(nbq0 == ggml_type_size(q->type)); + GGML_ASSERT(nbk0 == ggml_type_size(k->type)); + GGML_ASSERT(nbv0 == ggml_type_size(v->type)); + + GGML_ASSERT(neq0 == DK); + GGML_ASSERT(nek0 == DK); + GGML_ASSERT(nev0 == DV); + + GGML_ASSERT(neq1 == N); + + // dst cannot be transposed or permuted + GGML_ASSERT(nb0 == sizeof(float)); + GGML_ASSERT(nb0 <= nb1); + GGML_ASSERT(nb1 <= nb2); + GGML_ASSERT(nb2 <= nb3); + + // broadcast factors + const int64_t rk2 = neq2/nek2; + const int64_t rk3 = neq3/nek3; + + const int64_t rv2 = neq2/nev2; + const int64_t rv3 = neq3/nev3; + + // parallelize by q rows using ggml_vec_dot_f32 + + // total rows in q + const int nr = neq1*neq2*neq3; + + // rows per thread + const int dr = (nr + nth - 1)/nth; + + // row range for this thread + const int ir0 = dr*ith; + const int ir1 = MIN(ir0 + dr, nr); + + float scale = 1.0f; + float max_bias = 0.0f; + float logit_softcap = 0.0f; + + memcpy(&scale, (float *) dst->op_params + 0, sizeof(float)); + memcpy(&max_bias, (float *) dst->op_params + 1, sizeof(float)); + memcpy(&logit_softcap, (float *) dst->op_params + 2, sizeof(float)); + + if (logit_softcap != 0) { + scale /= logit_softcap; + } + + const uint32_t n_head = neq2; + const uint32_t n_head_log2 = 1u << (uint32_t) floor(log2(n_head)); + + const float m0 = powf(2.0f, -(max_bias ) / n_head_log2); + const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2); + + ggml_type const k_vec_dot_type = ggml_get_type_traits_cpu(k->type)->vec_dot_type; + ggml_from_float_t const q_to_vec_dot = ggml_get_type_traits_cpu(k_vec_dot_type)->from_float; + ggml_vec_dot_t const kq_vec_dot = ggml_get_type_traits_cpu(k->type)->vec_dot; + ggml_to_float_t const v_to_float = ggml_get_type_traits(v->type)->to_float; + + GGML_ASSERT(( q_to_vec_dot) && "fattn: unsupported K-type"); + GGML_ASSERT((v->type == GGML_TYPE_F32 || v_to_float ) && "fattn: unsupported V-type"); + + // loop over n_batch and n_head + for (int ir = ir0; ir < ir1; ++ir) { + // q indices + const int iq3 = ir/(neq2*neq1); + const int iq2 = (ir - iq3*neq2*neq1)/neq1; + const int iq1 = (ir - iq3*neq2*neq1 - iq2*neq1); + + const uint32_t h = iq2; // head index + const float slope = (max_bias > 0.0f) ? h < n_head_log2 ? powf(m0, h + 1) : powf(m1, 2*(h - n_head_log2) + 1) : 1.0f; + + float S = 0.0f; // sum + float M = -INFINITY; // maximum KQ value + + float * VKQ32 = (float *) params->wdata + ith*(1*DK + 2*DV + CACHE_LINE_SIZE_F32); // FP32 VKQ accumulator + float * V32 = (VKQ32 + 1*DV); // (temporary) FP32 V buffer + ggml_fp16_t * VKQ16 = (ggml_fp16_t *) (VKQ32 + 1*DV); // (temporary) FP16 VKQ accumulator + ggml_fp16_t * Q_q = (ggml_fp16_t *) (VKQ32 + 2*DV); // (temporary) buffer for Q converted to quantized/FP16 + + if (v->type == GGML_TYPE_F16) { + memset(VKQ16, 0, DV*sizeof(ggml_fp16_t)); + } else { + memset(VKQ32, 0, DV*sizeof(float)); + } + + const ggml_fp16_t * mp = mask ? (ggml_fp16_t *)((char *) mask->data + iq1*mask->nb[1]) : NULL; + + // k indices + const int ik3 = iq3 / rk3; + const int ik2 = iq2 / rk2; + + // v indices + const int iv3 = iq3 / rv3; + const int iv2 = iq2 / rv2; + + const float * pq = (const float *) ((char *) q->data + (iq1*nbq1 + iq2*nbq2 + iq3*nbq3)); + q_to_vec_dot(pq, Q_q, DK); + + // online softmax / attention + // loop over n_kv and n_head_kv + // ref: https://arxiv.org/pdf/2112.05682.pdf + for (int64_t ic = 0; ic < nek1; ++ic) { + const float mv = mp ? slope*GGML_FP16_TO_FP32(mp[ic]) : 0.0f; + if (mv == -INFINITY) { + continue; + } + + float s; // KQ value + + const char * k_data = (const char *) k->data + ( ic*nbk1 + ik2*nbk2 + ik3*nbk3); + kq_vec_dot(DK, &s, 0, k_data, 0, Q_q, 0, 1); + + s = s*scale; // scale KQ value + + if (logit_softcap != 0.0f) { + s = logit_softcap*tanhf(s); + } + + s += mv; // apply mask + + const float Mold = M; + + float ms = 1.0f; // upon new higher max val, scale VKQ and KQ sum with this value + float vs = 1.0f; // post-softmax KQ value, expf(s - M) + + const char * v_data = ((const char *) v->data + (ic*nbv1 + iv2*nbv2 + iv3*nbv3)); + + if (v->type == GGML_TYPE_F16) { + if (s > M) { + // s is new maximum, ms < 1.0f, vs == expf(s - s) == 1.0f + M = s; + ms = expf(Mold - M); + + // V = V*expf(Mold - M) + ggml_vec_scale_f16(DV, VKQ16, ms); + } else { + // no new maximum, ms == 1.0f, vs != 1.0f + vs = expf(s - M); + } + + // V += v*expf(s - M) + ggml_vec_mad_f16(DV, VKQ16, (const ggml_fp16_t *) v_data, vs); + } else { + if (s > M) { + // s is new maximum, ms < 1.0f, vs == expf(s - s) == 1.0f + M = s; + ms = expf(Mold - M); + + // V = V*expf(Mold - M) + ggml_vec_scale_f32(DV, VKQ32, ms); + } else { + // no new maximum, ms == 1.0f, vs != 1.0f + vs = expf(s - M); + } + + // V += v*expf(s - M) + if (v_to_float) { + v_to_float(v_data, V32, DV); + ggml_vec_mad_f32(DV, VKQ32, V32, vs); + } else { + // V is F32 + ggml_vec_mad_f32(DV, VKQ32, (const float *) v_data, vs); + } + } + + S = S*ms + vs; // scale and increment sum with partial sum + } + + if (v->type == GGML_TYPE_F16) { + for (int64_t d = 0; d < DV; ++d) { + VKQ32[d] = GGML_FP16_TO_FP32(VKQ16[d]); + } + } + + // V /= S + const float S_inv = 1.0f/S; + ggml_vec_scale_f32(DV, VKQ32, S_inv); + + // dst indices + const int i1 = iq1; + const int i2 = iq2; + const int i3 = iq3; + + // original + //memcpy((char *) dst->data + (i1*nb1 + i2*nb2 + i3*nb3), V, nev0*sizeof(float)); + + // permute(0, 2, 1, 3) + memcpy((char *) dst->data + (i3*ne2*ne1 + i2 + i1*ne1)*nb1, VKQ32, nb1); + } +} + +void ggml_compute_forward_flash_attn_ext( + const ggml_compute_params * params, + const ggml_tensor * q, + const ggml_tensor * k, + const ggml_tensor * v, + const ggml_tensor * mask, + ggml_tensor * dst) { + switch (dst->op_params[3]) { + case GGML_PREC_DEFAULT: + case GGML_PREC_F32: + { + // uses F32 accumulators + ggml_compute_forward_flash_attn_ext_f16(params, q, k, v, mask, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +// ggml_compute_forward_flash_attn_back + +static void ggml_compute_forward_flash_attn_back_f32( + const ggml_compute_params * params, + const bool masked, + ggml_tensor * dst) { + + const ggml_tensor * q = dst->src[0]; + const ggml_tensor * k = dst->src[1]; + const ggml_tensor * v = dst->src[2]; + const ggml_tensor * d = dst->src[3]; + + GGML_TENSOR_LOCALS(int64_t, neq, q, ne) + GGML_TENSOR_LOCALS(size_t, nbq, q, nb) + GGML_TENSOR_LOCALS(int64_t, nek, k, ne) + GGML_TENSOR_LOCALS(size_t, nbk, k, nb) + GGML_TENSOR_LOCALS(int64_t, nev, v, ne) + GGML_TENSOR_LOCALS(size_t, nbv, v, nb) + GGML_TENSOR_LOCALS(int64_t, ned, d, ne) + GGML_TENSOR_LOCALS(size_t, nbd, d, nb) + GGML_TENSOR_LOCALS(int64_t, ne, dst, ne) + GGML_TENSOR_LOCALS(size_t, nb, dst, nb) + + const int ith = params->ith; + const int nth = params->nth; + + const int64_t D = neq0; + const int64_t N = neq1; + const int64_t P = nek1 - N; + const int64_t M = P + N; + + const int Mup = ggml_up(M, GGML_SOFT_MAX_UNROLL); + const int mxDM = MAX(D, Mup); + + // GGML_ASSERT(ne0 == D); + // GGML_ASSERT(ne1 == N); + GGML_ASSERT(P >= 0); + + GGML_ASSERT(nbq0 == sizeof(float)); + GGML_ASSERT(nbk0 == sizeof(float)); + GGML_ASSERT(nbv0 == sizeof(float)); + + GGML_ASSERT(neq0 == D); + GGML_ASSERT(nek0 == D); + GGML_ASSERT(nev1 == D); + GGML_ASSERT(ned0 == D); + + GGML_ASSERT(neq1 == N); + GGML_ASSERT(nek1 == N + P); + GGML_ASSERT(nev1 == D); + GGML_ASSERT(ned1 == N); + + // dst cannot be transposed or permuted + GGML_ASSERT(nb0 == sizeof(float)); + GGML_ASSERT(nb0 <= nb1); + GGML_ASSERT(nb1 <= nb2); + GGML_ASSERT(nb2 <= nb3); + + if (ith == 0) { + memset(dst->data, 0, nb0*ne0*ne1*ne2*ne3); + } + ggml_barrier(params->threadpool); + + const int64_t elem_q = ggml_nelements(q); + const int64_t elem_k = ggml_nelements(k); + + ggml_type result_type = dst->type; + GGML_ASSERT(ggml_blck_size(result_type) == 1); + const size_t tsize = ggml_type_size(result_type); + + const size_t offs_q = 0; + const size_t offs_k = offs_q + GGML_PAD(elem_q * tsize, GGML_MEM_ALIGN); + const size_t offs_v = offs_k + GGML_PAD(elem_k * tsize, GGML_MEM_ALIGN); + + void * grad_q = (char *) dst->data; + void * grad_k = (char *) dst->data + offs_k; + void * grad_v = (char *) dst->data + offs_v; + + const size_t nbgq1 = nb0*neq0; + const size_t nbgq2 = nb0*neq0*neq1; + const size_t nbgq3 = nb0*neq0*neq1*neq2; + + const size_t nbgk1 = nb0*nek0; + const size_t nbgk2 = nb0*nek0*nek1; + const size_t nbgk3 = nb0*nek0*nek1*neq2; + + const size_t nbgv1 = nb0*nev0; + const size_t nbgv2 = nb0*nev0*nev1; + const size_t nbgv3 = nb0*nev0*nev1*neq2; + + // parallelize by k rows using ggml_vec_dot_f32 + + // total rows in k + const int nr = nek2*nek3; + + // rows per thread + const int dr = (nr + nth - 1)/nth; + + // row range for this thread + const int ir0 = dr*ith; + const int ir1 = MIN(ir0 + dr, nr); + + const float scale = 1.0f/sqrtf(D); + + //printf("P=%d N=%d D=%d ir0=%d ir1=%d scale = %f\n", P, N, D, ir0, ir1, scale); + + // how often k2 (and v2) is repeated in q2 + int nrep = neq2/nek2; + + for (int ir = ir0; ir < ir1; ++ir) { + // q indices + const int ik3 = ir/(nek2); + const int ik2 = ir - ik3*nek2; + + const int iq3 = ik3; + const int id3 = ik3; + const int iv3 = ik3; + const int iv2 = ik2; + + for (int irep = 0; irep < nrep; ++irep) { + const int iq2 = ik2 + irep*nek2; + const int id2 = iq2; + + // (ik2 + irep*nek2) % nek2 == ik2 + for (int iq1 = 0; iq1 < neq1; ++iq1) { + const int id1 = iq1; + + // not sure about CACHE_LINE_SIZE_F32.. + // - maybe it must not be multiplied by 2 and excluded from .. in SM 1*(..) offset? + float * S = (float *) params->wdata + ith*2*(mxDM + CACHE_LINE_SIZE_F32) + 0*(mxDM+CACHE_LINE_SIZE_F32); + float * SM = (float *) params->wdata + ith*2*(mxDM + CACHE_LINE_SIZE_F32) + 1*(mxDM+CACHE_LINE_SIZE_F32); + + for (int i = M; i < Mup; ++i) { + S[i] = -INFINITY; + } + + const int64_t masked_begin = masked ? (P + iq1 + 1) : M; + for (int64_t ic = 0; ic < masked_begin; ++ic) { + // k indices + const int ik1 = ic; + + // S indices + const int i1 = ik1; + + ggml_vec_dot_f32(neq0, + S + i1, 0, + (float *) ((char *) k->data + (ik1*nbk1 + ik2*nbk2 + ik3*nbk3)), 0, + (float *) ((char *) q->data + (iq1*nbq1 + iq2*nbq2 + iq3*nbq3)), 0, 1); + } + + // scale + ggml_vec_scale_f32(masked_begin, S, scale); + + for (int64_t i = masked_begin; i < M; i++) { + S[i] = -INFINITY; + } + + // softmax + // exclude known -INF S[..] values from max and loop + // dont forget to set their SM values to zero + { + float max = -INFINITY; + ggml_vec_max_f32(masked_begin, &max, S); + + ggml_float sum = 0.0; + { +#ifdef GGML_SOFT_MAX_ACCELERATE + max = -max; + vDSP_vsadd(SM, 1, &max, SM, 1, Mup); + vvexpf(SM, SM, &Mup); + ggml_vec_sum_f32(Mup, &sum, SM); +#else + sum = ggml_vec_soft_max_f32(Mup, SM, S, max); +#endif + } + + assert(sum > 0.0); + + sum = 1.0/sum; + ggml_vec_scale_f32(masked_begin, SM, sum); + + } + + // step-by-step explanation + { + // forward-process shape grads from backward process + // parallel_for ik2,ik3: + // for irep: + // iq2 = ik2 + irep*nek2 + // k[:D,:M,:,:] [D,M,:,:] grad[k][:D,:M,ik2,ik3] += grad[kcur] + // q[:D,:N,:,:] [D,N,:,:] grad[q][:D,iq1,iq2,iq3] += grad[qcur] + // v[:M,:D,:,:] [M,D,:,:] grad[v][:M,:D,iv2,iv3] += grad[vcur] + // for iq1: + // kcur = k[:D,:M,ik2,ik3] [D,M,1,1] grad[kcur] = grad[S1].T @ qcur + // qcur = q[:D,iq1,iq2,iq3] [D,1,1,1] grad[qcur] = grad[S1] @ kcur + // vcur = v[:M,:D,iv2,iv3] [M,D,1,1] grad[vcur] = grad[S5].T @ S4 + // S0 = -Inf [D,1,1,1] + // ~S1[i] = dot(kcur[:D,i], qcur) + // S1 = qcur @ kcur.T [M,1,1,1] grad[S1] = grad[S2] * scale + // S2 = S1 * scale [M,1,1,1] grad[S2] = diag_mask_zero(grad[S3], P) + // S3 = diag_mask_inf(S2, P) [M,1,1,1] grad[S3] = S4 * (grad[S4] - dot(S4, grad[S4])) + // S4 = softmax(S3) [M,1,1,1] grad[S4] = grad[S5] @ vcur + // ~S5[i] = dot(vcur[:,i], S4) + // S5 = S4 @ vcur.T [D,1,1,1] grad[S5] = d[:D,id1,id2,id3] + // ~dst[i,iq1,iq2,iq3] = S5[i] ^ + // dst[:D,iq1,iq2,iq3] = S5 | grad[dst[:D,iq1,iq2,iq3]] = d[:D,id1,id2,id3] + // dst backward-/ grad[dst] = d + // + // output gradients with their dependencies: + // + // grad[kcur] = grad[S1].T @ qcur + // grad[S1] = diag_mask_zero(grad[S3], P) * scale + // grad[S3] = S4 * (grad[S4] - dot(S4, grad[S4])) + // grad[S4] = grad[S5] @ vcur + // grad[S4] = d[:D,id1,id2,id3] @ vcur + // grad[qcur] = grad[S1] @ kcur + // grad[vcur] = grad[S5].T @ S4 + // grad[vcur] = d[:D,id1,id2,id3].T @ S4 + // + // in post-order: + // + // S1 = qcur @ kcur.T + // S2 = S1 * scale + // S3 = diag_mask_inf(S2, P) + // S4 = softmax(S3) + // grad[S4] = d[:D,id1,id2,id3] @ vcur + // grad[S3] = S4 * (grad[S4] - dot(S4, grad[S4])) + // grad[S1] = diag_mask_zero(grad[S3], P) * scale + // grad[qcur] = grad[S1] @ kcur + // grad[kcur] = grad[S1].T @ qcur + // grad[vcur] = d[:D,id1,id2,id3].T @ S4 + // + // using less variables (SM=S4): + // + // S = diag_mask_inf(qcur @ kcur.T * scale, P) + // SM = softmax(S) + // S = d[:D,iq1,iq2,iq3] @ vcur + // dot_SM_gradSM = dot(SM, S) + // S = SM * (S - dot(SM, S)) + // S = diag_mask_zero(S, P) * scale + // + // grad[q][:D,iq1,iq2,iq3] += S @ kcur + // grad[k][:D,:M,ik2,ik3] += S.T @ qcur + // grad[v][:M,:D,iv2,iv3] += d[:D,id1,id2,id3].T @ SM + } + + // S = gradSM = d[:D,id1,id2,id3] @ vcur[:,:,iv2,iv3] + // S = d[:D,id1,id2,id3] @ vcur[:,:,iv2,iv3] + // for ic: + // S[:M] += vcur[:M,ic,iv2,iv3] * d[ic,id1,id2,id3] + // exclude known future zero S[..] values from operation + ggml_vec_set_f32(masked_begin, S, 0); + for (int64_t ic = 0; ic < D; ++ic) { + ggml_vec_mad_f32(masked_begin, + S, + (float *) ((char *) v->data + ( ic*nbv1 + iv2*nbv2 + iv3*nbv3)), + *(float *) ((char *) d->data + (ic*nbd0 + id1*nbd1 + id2*nbd2 + id3*nbd3))); + } + + // S = SM * (S - dot(SM, S)) + float dot_SM_gradSM = 0; + ggml_vec_dot_f32 (masked_begin, &dot_SM_gradSM, 0, SM, 0, S, 0, 1); + ggml_vec_acc1_f32(M, S, -dot_SM_gradSM); + ggml_vec_mul_f32 (masked_begin, S, S, SM); + + // S = diag_mask_zero(S, P) * scale + // already done by above ggml_vec_set_f32 + + // exclude known zero S[..] values from operation + ggml_vec_scale_f32(masked_begin, S, scale); + + // S shape [M,1] + // SM shape [M,1] + // kcur shape [D,M] + // qcur shape [D,1] + // vcur shape [M,D] + + // grad[q][:D,iq1,iq2,iq3] += S @ kcur + // grad[q][:D,iq1,iq2,iq3] += shape[M,1] @ shape[D,M] + // for ic: + // grad[q][:D,iq1,iq2,iq3] += S[ic] * kcur[:D,ic,ik2,ik3] + // exclude known zero S[..] values from loop + for (int64_t ic = 0; ic < masked_begin; ++ic) { + ggml_vec_mad_f32(D, + (float *) ((char *) grad_q + (iq1*nbgq1 + iq2*nbgq2 + iq3*nbgq3)), + (float *) ((char *) k->data + (ic*nbk1 + ik2*nbk2 + ik3*nbk3)), + S[ic]); + } + + // grad[k][:D,:M,iq2,iq3] += S.T @ qcur + // for ic: + // grad[k][:D,ic,iq2,iq3] += S.T[0,ic] * qcur[:D,0] + // grad[k][:D,ic,iq2,iq3] += S[ic] * qcur[:D,0] + // exclude known zero S[..] values from loop + for (int64_t ic = 0; ic < masked_begin; ++ic) { + ggml_vec_mad_f32(D, + (float *) ((char *) grad_k + (ic*nbgk1 + ik2*nbgk2 + ik3*nbgk3)), + (float *) ((char *) q->data + (iq1*nbq1 + iq2*nbq2 + iq3*nbq3)), + S[ic]); + } + + // grad[v][:M,:D,iv2,iv3] += d[:D,id1,id2,id3].T @ SM + // for ic: + // grad[v][:M,ic,iv2,iv3] += d[:D,id1,id2,id3].T[0,ic] * SM[:M] + // grad[v][:M,ic,iv2,iv3] += d[ic,id1,id2,id3] * SM[:M] + // exclude known zero SM[..] values from mad + for (int64_t ic = 0; ic < D; ++ic) { + ggml_vec_mad_f32(masked_begin, + (float *) ((char *) grad_v + ( ic*nbgv1 + iv2*nbgv2 + iv3*nbgv3)), + SM, + *(float *) ((char *) d->data + (ic*nbd0 + id1*nbd1 + id2*nbd2 + id3*nbd3))); + } + } + } + } +} + +void ggml_compute_forward_flash_attn_back( + const ggml_compute_params * params, + const bool masked, + ggml_tensor * dst) { + + const ggml_tensor * q = dst->src[0]; + + switch (q->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_flash_attn_back_f32(params, masked, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +// ggml_compute_forward_ssm_conv + +static void ggml_compute_forward_ssm_conv_f32( + const ggml_compute_params * params, + ggml_tensor * dst) { + const ggml_tensor * src0 = dst->src[0]; // conv_x + const ggml_tensor * src1 = dst->src[1]; // conv1d.weight + + const int ith = params->ith; + const int nth = params->nth; + + const int nc = src1->ne[0]; // d_conv + const int ncs = src0->ne[0]; // d_conv - 1 + n_t + const int nr = src0->ne[1]; // d_inner + const int n_t = dst->ne[1]; // tokens per sequence + const int n_s = dst->ne[2]; // number of sequences in the batch + + GGML_ASSERT( dst->ne[0] == nr); + GGML_ASSERT(src0->nb[0] == sizeof(float)); + GGML_ASSERT(src1->nb[0] == sizeof(float)); + GGML_ASSERT(src0->nb[1] == src0->ne[0]*sizeof(float)); + + // rows per thread + const int dr = (nr + nth - 1)/nth; + + // row range for this thread + const int ir0 = dr*ith; + const int ir1 = MIN(ir0 + dr, nr); + const int ir = ir1 - ir0; + + for (int i3 = 0; i3 < n_s; ++i3) { + for (int i2 = 0; i2 < n_t; ++i2) { + // {d_conv - 1 + n_t, d_inner, n_seqs} + // sliding window + const float * s = (const float *) ((const char *) src0->data + ir0*(src0->nb[1]) + i2*(src0->nb[0]) + i3*(src0->nb[2])); // {d_conv, d_inner, n_s} + const float * c = (const float *) ((const char *) src1->data + ir0*(src1->nb[1])); // {d_conv, d_inner} + float * x = (float *) ((char *) dst->data + ir0*(dst->nb[0]) + i2*(dst->nb[1]) + i3*(dst->nb[2])); // {d_inner, n_t, n_s} + + // TODO: transpose the output for smaller strides for big batches? + // d_inner + for (int i1 = 0; i1 < ir; ++i1) { + // rowwise dot product + // NOTE: not using ggml_vec_dot_f32, because its sum is in double precision + float sumf = 0.0f; + + // d_conv + for (int i0 = 0; i0 < nc; ++i0) { + sumf += s[i0 + i1*ncs] * c[i0 + i1*nc]; + } + x[i1] = sumf; + } + } + } +} + +void ggml_compute_forward_ssm_conv( + const ggml_compute_params * params, + ggml_tensor * dst) { + switch (dst->src[0]->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_ssm_conv_f32(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +// ggml_compute_forward_ssm_scan + +static void ggml_compute_forward_ssm_scan_f32( + const ggml_compute_params * params, + ggml_tensor * dst) { + const ggml_tensor * src0 = dst->src[0]; // s + const ggml_tensor * src1 = dst->src[1]; // x + const ggml_tensor * src2 = dst->src[2]; // dt + const ggml_tensor * src3 = dst->src[3]; // A + const ggml_tensor * src4 = dst->src[4]; // B + const ggml_tensor * src5 = dst->src[5]; // C + + const int ith = params->ith; + const int nth = params->nth; + + const int64_t nc = src0->ne[0]; // d_state + const int64_t nr = src0->ne[1]; // d_inner + const int64_t n_t = src1->ne[1]; // number of tokens per sequence + const int64_t n_s = src0->ne[2]; // number of sequences in the batch + + GGML_ASSERT(ggml_nelements(src1) + ggml_nelements(src0) == ggml_nelements(dst)); + GGML_ASSERT(src0->nb[0] == sizeof(float)); + GGML_ASSERT(src1->nb[0] == sizeof(float)); + GGML_ASSERT(src2->nb[0] == sizeof(float)); + GGML_ASSERT(src3->nb[0] == sizeof(float)); + GGML_ASSERT(src4->nb[0] == sizeof(float)); + GGML_ASSERT(src5->nb[0] == sizeof(float)); + // required for the dot product between s and C + GGML_ASSERT(src0->nb[1] == src0->ne[0]*sizeof(float)); + // required for per-sequence offsets for states + GGML_ASSERT(src0->nb[2] == src0->ne[0]*src0->ne[1]*sizeof(float)); + // required to get correct offset for state destination (i.e. src1->nb[3]) + GGML_ASSERT(src1->nb[3] == src1->ne[0]*src1->ne[1]*src1->ne[2]*sizeof(float)); + + // rows per thread + const int dr = (nr + nth - 1)/nth; + + // row range for this thread + const int ir0 = dr*ith; + const int ir1 = MIN(ir0 + dr, nr); + const int ir = ir1 - ir0; + + for (int i3 = 0; i3 < n_s; ++i3) { + for (int i2 = 0; i2 < n_t; ++i2) { + const float * s0 = (const float *) ((const char *) src0->data + ir0*(src0->nb[1]) + i3*(src0->nb[2])); // {d_state, d_inner, n_s} + const float * x = (const float *) ((const char *) src1->data + ir0*(src1->nb[0]) + i2*(src1->nb[1]) + i3*(src1->nb[2])); // {d_inner, n_t, n_s} + const float * dt = (const float *) ((const char *) src2->data + ir0*(src2->nb[0]) + i2*(src2->nb[1]) + i3*(src2->nb[2])); // {d_inner, n_t, n_s} + const float * A = (const float *) ((const char *) src3->data + ir0*(src3->nb[1])); // {d_state, d_inner} + const float * B = (const float *) ((const char *) src4->data + i2*(src4->nb[1]) + i3*(src4->nb[2])); // {d_state, n_t, n_s} + const float * C = (const float *) ((const char *) src5->data + i2*(src5->nb[1]) + i3*(src5->nb[2])); // {d_state, n_t, n_s} + float * y = ( float *) (( char *) dst->data + ir0*(src1->nb[0]) + i2*(src1->nb[1]) + i3*(src1->nb[2])); // {d_inner, n_t, n_s} + float * s = ( float *) (( char *) dst->data + ir0*(src0->nb[1]) + i3*(src0->nb[2]) + src1->nb[3]); // {d_state, d_inner, n_s} + + // use the output as the source for the next token-wise iterations + if (i2 > 0) { s0 = s; } + + // d_inner + for (int i1 = 0; i1 < ir; ++i1) { + // ref: https://github.com/state-spaces/mamba/blob/34076d664838588a3c97727b263478ab9f621a07/mamba_ssm/ops/triton/selective_state_update.py#L78 + float dt_soft_plus = dt[i1] <= 20.0f ? log1pf(expf(dt[i1])) : dt[i1]; + float x_dt = x[i1] * dt_soft_plus; + float sumf = 0.0f; + // d_state + for (int i0 = 0; i0 < nc; ++i0) { + int i = i0 + i1*nc; + // state = prev_state * dA + dB * x + float state = (s0[i] * expf(dt_soft_plus * A[i])) + (B[i0] * x_dt); + // y = rowwise_dotprod(state, C) + sumf += state * C[i0]; + s[i] = state; + } + y[i1] = sumf; + } + } + } +} + +void ggml_compute_forward_ssm_scan( + const ggml_compute_params * params, + ggml_tensor * dst) { + switch (dst->src[0]->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_ssm_scan_f32(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +// ggml_compute_forward_win_part + +static void ggml_compute_forward_win_part_f32( + const ggml_compute_params * params, + ggml_tensor * dst) { + GGML_UNUSED(params); + + const ggml_tensor * src0 = dst->src[0]; + + GGML_TENSOR_LOCALS(int64_t, ne0, src0, ne) + GGML_TENSOR_LOCALS(int64_t, ne, dst, ne) + + const int32_t nep0 = ((const int32_t *)(dst->op_params))[0]; + const int32_t nep1 = ((const int32_t *)(dst->op_params))[1]; + const int32_t w = ((const int32_t *)(dst->op_params))[2]; + + assert(ne00 == ne0); + assert(ne3 == nep0*nep1); + + // TODO: optimize / multi-thread + for (int py = 0; py < nep1; ++py) { + for (int px = 0; px < nep0; ++px) { + const int64_t i3 = py*nep0 + px; + for (int64_t i2 = 0; i2 < ne2; ++i2) { + for (int64_t i1 = 0; i1 < ne1; ++i1) { + for (int64_t i0 = 0; i0 < ne0; ++i0) { + const int64_t i02 = py*w + i2; + const int64_t i01 = px*w + i1; + const int64_t i00 = i0; + + const int64_t i = i3*ne2*ne1*ne0 + i2*ne1*ne0 + i1*ne0 + i0; + const int64_t j = i02*ne01*ne00 + i01*ne00 + i00; + + if (py*w + i2 >= ne02 || px*w + i1 >= ne01) { + ((float *) dst->data)[i] = 0.0f; + } else { + ((float *) dst->data)[i] = ((float *) src0->data)[j]; + } + } + } + } + } + } +} + +void ggml_compute_forward_win_part( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_win_part_f32(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +// ggml_compute_forward_win_unpart + +static void ggml_compute_forward_win_unpart_f32( + const ggml_compute_params * params, + ggml_tensor * dst) { + GGML_UNUSED(params); + + const ggml_tensor * src0 = dst->src[0]; + + GGML_TENSOR_LOCALS(int64_t, ne0, src0, ne) + GGML_TENSOR_LOCALS(int64_t, ne, dst, ne) + + const int32_t w = ((const int32_t *)(dst->op_params))[0]; + + // padding + const int px = (w - ne1%w)%w; + //const int py = (w - ne2%w)%w; + + const int npx = (px + ne1)/w; + //const int npy = (py + ne2)/w; + + assert(ne0 == ne00); + + // TODO: optimize / multi-thread + for (int64_t i2 = 0; i2 < ne2; ++i2) { + for (int64_t i1 = 0; i1 < ne1; ++i1) { + for (int64_t i0 = 0; i0 < ne0; ++i0) { + const int ip2 = i2/w; + const int ip1 = i1/w; + + const int64_t i02 = i2%w; + const int64_t i01 = i1%w; + const int64_t i00 = i0; + + const int64_t i = (ip2*npx + ip1)*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00 + i00; + const int64_t j = i2*ne1*ne0 + i1*ne0 + i0; + + ((float *) dst->data)[j] = ((float *) src0->data)[i]; + } + } + } +} + +void ggml_compute_forward_win_unpart( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_win_unpart_f32(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +//gmml_compute_forward_unary + +void ggml_compute_forward_unary( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_unary_op op = ggml_get_unary_op(dst); + + switch (op) { + case GGML_UNARY_OP_ABS: + { + ggml_compute_forward_abs(params, dst); + } break; + case GGML_UNARY_OP_SGN: + { + ggml_compute_forward_sgn(params, dst); + } break; + case GGML_UNARY_OP_NEG: + { + ggml_compute_forward_neg(params, dst); + } break; + case GGML_UNARY_OP_STEP: + { + ggml_compute_forward_step(params, dst); + } break; + case GGML_UNARY_OP_TANH: + { + ggml_compute_forward_tanh(params, dst); + } break; + case GGML_UNARY_OP_ELU: + { + ggml_compute_forward_elu(params, dst); + } break; + case GGML_UNARY_OP_RELU: + { + ggml_compute_forward_relu(params, dst); + } break; + case GGML_UNARY_OP_SIGMOID: + { + ggml_compute_forward_sigmoid(params, dst); + } break; + case GGML_UNARY_OP_GELU: + { + ggml_compute_forward_gelu(params, dst); + } break; + case GGML_UNARY_OP_GELU_QUICK: + { + ggml_compute_forward_gelu_quick(params, dst); + } break; + case GGML_UNARY_OP_SILU: + { + ggml_compute_forward_silu(params, dst); + } break; + case GGML_UNARY_OP_HARDSWISH: + { + ggml_compute_forward_hardswish(params, dst); + } break; + case GGML_UNARY_OP_HARDSIGMOID: + { + ggml_compute_forward_hardsigmoid(params, dst); + } break; + case GGML_UNARY_OP_EXP: + { + ggml_compute_forward_exp(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +// ggml_compute_forward_get_rel_pos + +static void ggml_compute_forward_get_rel_pos_f16( + const ggml_compute_params * params, + ggml_tensor * dst) { + GGML_UNUSED(params); + + const ggml_tensor * src0 = dst->src[0]; + + // ref: https://github.com/facebookresearch/segment-anything/blob/main/segment_anything/modeling/image_encoder.py#L292-L322 + + GGML_TENSOR_UNARY_OP_LOCALS + + const int64_t w = ne1; + + ggml_fp16_t * src0_data = (ggml_fp16_t *) src0->data; + ggml_fp16_t * dst_data = (ggml_fp16_t *) dst->data; + + for (int64_t i2 = 0; i2 < ne2; ++i2) { + for (int64_t i1 = 0; i1 < ne1; ++i1) { + const int64_t pos = (w - i1 - 1) + i2; + for (int64_t i0 = 0; i0 < ne0; ++i0) { + dst_data[i2*ne1*ne0 + i1*ne0 + i0] = src0_data[pos*ne00 + i0]; + } + } + } +} + +void ggml_compute_forward_get_rel_pos( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F16: + case GGML_TYPE_BF16: + { + ggml_compute_forward_get_rel_pos_f16(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +// ggml_compute_forward_add_rel_pos + +static void ggml_compute_forward_add_rel_pos_f32( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + const ggml_tensor * src1 = dst->src[1]; + const ggml_tensor * src2 = dst->src[2]; + + const bool inplace = (bool) ((int32_t *) dst->op_params)[0]; + if (!inplace) { + if (params->ith == 0) { + memcpy((char *) dst->data, (char *) src0->data, ggml_nbytes(dst)); + } + ggml_barrier(params->threadpool); + } + // ref: https://github.com/facebookresearch/segment-anything/blob/main/segment_anything/modeling/image_encoder.py#L357-L359 + + float * src1_data = (float *) src1->data; + float * src2_data = (float *) src2->data; + float * dst_data = (float *) dst->data; + + const int64_t ne10 = src1->ne[0]; + const int64_t ne11 = src1->ne[1]; + const int64_t ne12 = src1->ne[2]; + const int64_t ne13 = src1->ne[3]; + + const int ith = params->ith; + const int nth = params->nth; + + // total patches in dst + const int np = ne13; + + // patches per thread + const int dp = (np + nth - 1)/nth; + + // patch range for this thread + const int ip0 = dp*ith; + const int ip1 = MIN(ip0 + dp, np); + + for (int64_t i13 = ip0; i13 < ip1; ++i13) { + for (int64_t i12 = 0; i12 < ne12; ++i12) { + for (int64_t i11 = 0; i11 < ne11; ++i11) { + const int64_t jp1 = i13*ne12*ne11*ne10 + i12*ne11*ne10 + i11*ne10; + for (int64_t i10 = 0; i10 < ne10; ++i10) { + const int64_t jp0 = jp1 + i10; + const float src1_e = src1_data[jp0]; + const float src2_e = src2_data[jp0]; + + const int64_t jdh = jp0 * ne10; + const int64_t jdw = jdh - (ne10 - 1) * i10; + + for (int64_t j = 0; j < ne10; ++j) { + dst_data[jdh + j ] += src2_e; + dst_data[jdw + j*ne10] += src1_e; + } + } + } + } + } +} + +void ggml_compute_forward_add_rel_pos( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_add_rel_pos_f32(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +// ggml_compute_forward_rwkv_wkv6 + +static void ggml_compute_forward_rwkv_wkv6_f32( + const ggml_compute_params * params, + ggml_tensor * dst) { + const int64_t T = dst->src[1]->ne[2]; + const int64_t C = dst->ne[0]; + const int64_t HEADS = dst->src[1]->ne[1]; + const int64_t n_seqs = dst->src[5]->ne[1]; + const int64_t head_size = C / HEADS; + + float * dst_data = (float *) dst->data; + float * state = ((float *) dst->data) + C * T; + + const int ith = params->ith; + const int nth = params->nth; + + if (ith >= HEADS) { + return; + } + + const int h_start = (HEADS * ith) / nth; + const int h_end = ((HEADS * (ith + 1)) / nth < HEADS) ? + (HEADS * (ith + 1)) / nth : HEADS; + + float * k = (float *) dst->src[0]->data; + float * v = (float *) dst->src[1]->data; + float * r = (float *) dst->src[2]->data; + float * time_faaaa = (float *) dst->src[3]->data; + float * time_decay = (float *) dst->src[4]->data; + + size_t t_stride = HEADS * head_size; // Same to C + + size_t h_stride = C / HEADS; + GGML_ASSERT(C % HEADS == 0); // C must be divisible by HEADS + size_t h_stride_2d = head_size * head_size; + + if (ith == 0) { + memset(dst_data, 0, T * C * sizeof(float)); + } + ggml_barrier(params->threadpool); + + + #if defined(__AVX__) && !defined(__AVX512F__) + #define GGML_F32X GGML_F32x8 + #define GGML_F32X_SET1 GGML_F32x8_SET1 + #define GGML_F32X_LOAD GGML_F32x8_LOAD + #define GGML_F32X_STORE GGML_F32x8_STORE + #define GGML_F32X_MUL GGML_F32x8_MUL + #define GGML_F32X_FMA GGML_F32x8_FMA + #define WKV_VECTOR_SIZE 8 + #elif defined(__AVX512F__) + #define GGML_F32X GGML_F32x16 + #define GGML_F32X_SET1 GGML_F32x16_SET1 + #define GGML_F32X_LOAD GGML_F32x16_LOAD + #define GGML_F32X_STORE GGML_F32x16_STORE + #define GGML_F32X_MUL GGML_F32x16_MUL + #define GGML_F32X_FMA GGML_F32x16_FMA + #define WKV_VECTOR_SIZE 16 + #elif defined(__ARM_NEON) && defined(__aarch64__) + #define GGML_F32X GGML_F32x4 + #define GGML_F32X_SET1 GGML_F32x4_SET1 + #define GGML_F32X_LOAD GGML_F32x4_LOAD + #define GGML_F32X_STORE GGML_F32x4_STORE + #define GGML_F32X_MUL GGML_F32x4_MUL + #define GGML_F32X_FMA GGML_F32x4_FMA + #define WKV_VECTOR_SIZE 4 + #endif + + #ifdef WKV_VECTOR_SIZE + const int64_t vec_count = head_size / WKV_VECTOR_SIZE; + + for (int64_t t = 0; t < T; t++) { + size_t t_offset = t * t_stride; + size_t state_offset = head_size * C * (t / (T / n_seqs)); + float * state_cur = state + state_offset; + float * state_prev = t % (T / n_seqs) ? state_cur : (float*)dst->src[5]->data + state_offset; + + for (int64_t h = h_start; h < h_end; h++) { + size_t h_offset = h * h_stride; + size_t t_h_offset = t_offset + h_offset; + size_t h_2d_offset = h * h_stride_2d; + + for (int64_t i = 0; i < head_size; i++) { + size_t t_h_i_offset = t_h_offset + i; + size_t h_i_offset = h_offset + i; + size_t h_2d_i_offset = h_2d_offset + i * h_stride; + + float k_val = k[t_h_i_offset]; + float r_val = r[t_h_i_offset]; + float time_faaaa_val = time_faaaa[h_i_offset]; + float time_decay_val = time_decay[t_h_i_offset]; + + // Broadcast scalar values to vectors + GGML_F32X k_vec = GGML_F32X_SET1(k_val); + GGML_F32X r_vec = GGML_F32X_SET1(r_val); + GGML_F32X time_faaaa_vec = GGML_F32X_SET1(time_faaaa_val); + GGML_F32X time_decay_vec = GGML_F32X_SET1(time_decay_val); + + for (int64_t j = 0; j < vec_count; j++) { + size_t base_j = j * WKV_VECTOR_SIZE; + size_t t_h_j_offset = t_h_offset + base_j; + size_t h_2d_i_j_offset = h_2d_i_offset + base_j; + + // Load x elements at once + GGML_F32X v_vec = GGML_F32X_LOAD(&v[t_h_j_offset]); + GGML_F32X prev_state_vec = GGML_F32X_LOAD(&state_prev[h_2d_i_j_offset]); + GGML_F32X dst_vec = GGML_F32X_LOAD(&dst_data[t_h_j_offset]); + + // Compute kv = v * k + GGML_F32X kv_vec = GGML_F32X_MUL(v_vec, k_vec); + + // Compute temp = kv * time_faaaa + prev_state + GGML_F32X temp_vec = GGML_F32X_FMA(prev_state_vec, kv_vec, time_faaaa_vec); + + // Update dst: dst += temp * r + dst_vec = GGML_F32X_FMA(dst_vec, temp_vec, r_vec); + GGML_F32X_STORE(&dst_data[t_h_j_offset], dst_vec); + + // Update state: state = prev_state * time_decay + kv + GGML_F32X new_state_vec = GGML_F32X_FMA(kv_vec, prev_state_vec, time_decay_vec); + GGML_F32X_STORE(&state_cur[h_2d_i_j_offset], new_state_vec); + } + + // Handle remaining elements, this will not be used. + for (int64_t j = vec_count * WKV_VECTOR_SIZE; j < head_size; j++) { + size_t t_h_j_offset = t_h_offset + j; + size_t h_2d_i_j_offset = h_2d_i_offset + j; + float v_val = v[t_h_j_offset]; + float kv_val = v_val * k_val; + float prev_state_val = state_prev[h_2d_i_j_offset]; + float temp_val = kv_val * time_faaaa_val + prev_state_val; + dst_data[t_h_j_offset] += temp_val * r_val; + state_cur[h_2d_i_j_offset] = prev_state_val * time_decay_val + kv_val; + } + } + } + } + + #else + // basically fused operations: + // dst = r @ (time_faaaa * (k @ v) + state), + // state = time_decay * state + (k @ v), + // recursive through each token + for (int64_t t = 0; t < T; t++) { + size_t t_offset = t * t_stride; + size_t state_offset = head_size * C * (t / (T / n_seqs)); + float * state_cur = state + state_offset; + float * state_prev = t % (T / n_seqs) ? state_cur : (float*)dst->src[5]->data + state_offset; + + for (int64_t h = h_start; h < h_end; h++) { + size_t h_offset = h * h_stride; + size_t t_h_offset = t_offset + h_offset; + size_t h_2d_offset = h * h_stride_2d; + + for (int64_t i = 0; i < head_size; i++) { + size_t t_h_i_offset = t_h_offset + i; + size_t h_i_offset = h_offset + i; + size_t h_2d_i_offset = h_2d_offset + i * h_stride; + + float k_val = k[t_h_i_offset]; + float r_val = r[t_h_i_offset]; + float time_faaaa_val = time_faaaa[h_i_offset]; + // RWKV v6: different time_decay for each token. + float time_decay_val = time_decay[t_h_i_offset]; + + for (int64_t j = 0; j < head_size; j++) { + size_t t_h_j_offset = t_h_offset + j; + size_t h_2d_i_j_offset = h_2d_i_offset + j; + + float v_val = v[t_h_j_offset]; + float kv_val = v_val * k_val; + float prev_state_val = state_prev[h_2d_i_j_offset]; + float temp_val = kv_val * time_faaaa_val + prev_state_val; + dst_data[t_h_j_offset] += temp_val * r_val; + state_cur[h_2d_i_j_offset] = prev_state_val * time_decay_val + kv_val; + } + } + } + } + #endif +} + + +void ggml_compute_forward_rwkv_wkv6( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_rwkv_wkv6_f32(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +// ggml_compute_forward_gla + +static void ggml_compute_forward_gla_f32( + const ggml_compute_params * params, + ggml_tensor * dst) { + const int64_t T = dst->src[1]->ne[2]; + const int64_t C = dst->ne[0]; + const int64_t HEADS = dst->src[1]->ne[1]; + const int64_t n_seqs = dst->src[4]->ne[1]; + const int64_t head_size = C / HEADS; + const float scale = ggml_get_op_params_f32(dst, 0); + + float * dst_data = (float *) dst->data; + float * state = ((float *) dst->data) + C * T; + + const int ith = params->ith; + const int nth = params->nth; + + if (ith >= HEADS) { + return; + } + + const int h_start = (HEADS * ith) / nth; + const int h_end = ((HEADS * (ith + 1)) / nth < HEADS) ? + (HEADS * (ith + 1)) / nth : HEADS; + + float * k = (float *) dst->src[0]->data; + float * v = (float *) dst->src[1]->data; + float * q = (float *) dst->src[2]->data; + float * g = (float *) dst->src[3]->data; + + size_t t_stride = HEADS * head_size; // Same to C + + size_t h_stride = C / HEADS; + GGML_ASSERT(C % HEADS == 0); // C must be divisible by HEADS + size_t h_stride_2d = head_size * head_size; + + if (ith == 0) { + memset(dst_data, 0, T * C * sizeof(float)); + } + ggml_barrier(params->threadpool); + + + #if defined(__AVX__) && !defined(__AVX512F__) + #define GGML_F32X GGML_F32x8 + #define GGML_F32X_SET1 GGML_F32x8_SET1 + #define GGML_F32X_LOAD GGML_F32x8_LOAD + #define GGML_F32X_STORE GGML_F32x8_STORE + #define GGML_F32X_MUL GGML_F32x8_MUL + #define GGML_F32X_FMA GGML_F32x8_FMA + #define GLA_VECTOR_SIZE 8 + #elif defined(__AVX512F__) + #define GGML_F32X GGML_F32x16 + #define GGML_F32X_SET1 GGML_F32x16_SET1 + #define GGML_F32X_LOAD GGML_F32x16_LOAD + #define GGML_F32X_STORE GGML_F32x16_STORE + #define GGML_F32X_MUL GGML_F32x16_MUL + #define GGML_F32X_FMA GGML_F32x16_FMA + #define GLA_VECTOR_SIZE 16 + #elif defined(__ARM_NEON) && defined(__aarch64__) + #define GGML_F32X GGML_F32x4 + #define GGML_F32X_SET1 GGML_F32x4_SET1 + #define GGML_F32X_LOAD GGML_F32x4_LOAD + #define GGML_F32X_STORE GGML_F32x4_STORE + #define GGML_F32X_MUL GGML_F32x4_MUL + #define GGML_F32X_FMA GGML_F32x4_FMA + #define GLA_VECTOR_SIZE 4 + #endif + + #ifdef GLA_VECTOR_SIZE + const int64_t vec_count = head_size / GLA_VECTOR_SIZE; + + for (int64_t t = 0; t < T; t++) { + size_t t_offset = t * t_stride; + size_t state_offset = head_size * C * (t / (T / n_seqs)); + float * state_cur = state + state_offset; + float * state_prev = t % (T / n_seqs) ? state_cur : (float*)dst->src[4]->data + state_offset; + + for (int64_t h = h_start; h < h_end; h++) { + size_t h_offset = h * h_stride; + size_t t_h_offset = t_offset + h_offset; + size_t h_2d_offset = h * h_stride_2d; + + for (int64_t i = 0; i < head_size; i++) { + size_t t_h_i_offset = t_h_offset + i; + size_t h_2d_i_offset = h_2d_offset + i * h_stride; + + float k_val = k[t_h_i_offset]; + float q_val = q[t_h_i_offset] * scale; + float g_val = g[t_h_i_offset]; + + // Broadcast scalar values to vectors + GGML_F32X k_vec = GGML_F32X_SET1(k_val); + GGML_F32X q_vec = GGML_F32X_SET1(q_val); + GGML_F32X g_vec = GGML_F32X_SET1(g_val); + + for (int64_t j = 0; j < vec_count; j++) { + size_t base_j = j * GLA_VECTOR_SIZE; + size_t t_h_j_offset = t_h_offset + base_j; + size_t h_2d_i_j_offset = h_2d_i_offset + base_j; + + // Load x elements at once + GGML_F32X v_vec = GGML_F32X_LOAD(&v[t_h_j_offset]); + GGML_F32X prev_state_vec = GGML_F32X_LOAD(&state_prev[h_2d_i_j_offset]); + GGML_F32X dst_vec = GGML_F32X_LOAD(&dst_data[t_h_j_offset]); + + // Compute kv = v * k + GGML_F32X kv_vec = GGML_F32X_MUL(v_vec, k_vec); + + // Compute temp = prev_state * g + kv + GGML_F32X temp_vec = GGML_F32X_FMA(kv_vec, prev_state_vec, g_vec); + + // Update dst: dst += temp * q + dst_vec = GGML_F32X_FMA(dst_vec, temp_vec, q_vec); + GGML_F32X_STORE(&dst_data[t_h_j_offset], dst_vec); + + // Update state + GGML_F32X_STORE(&state_cur[h_2d_i_j_offset], temp_vec); + } + + // Handle remaining elements, this will not be used. + for (int64_t j = vec_count * GLA_VECTOR_SIZE; j < head_size; j++) { + size_t t_h_j_offset = t_h_offset + j; + size_t h_2d_i_j_offset = h_2d_i_offset + j; + float v_val = v[t_h_j_offset]; + float kv_val = v_val * k_val; + float prev_state_val = state_prev[h_2d_i_j_offset]; + float temp_val = kv_val + prev_state_val * g_val; + dst_data[t_h_j_offset] += temp_val * q_val; + state_cur[h_2d_i_j_offset] = temp_val; + } + } + } + } + + #else + for (int64_t t = 0; t < T; t++) { + size_t t_offset = t * t_stride; + size_t state_offset = head_size * C * (t / (T / n_seqs)); + float * state_cur = state + state_offset; + float * state_prev = t % (T / n_seqs) ? state_cur : (float*)dst->src[4]->data + state_offset; + + for (int64_t h = h_start; h < h_end; h++) { + size_t h_offset = h * h_stride; + size_t t_h_offset = t_offset + h_offset; + size_t h_2d_offset = h * h_stride_2d; + + for (int64_t i = 0; i < head_size; i++) { + size_t t_h_i_offset = t_h_offset + i; + size_t h_2d_i_offset = h_2d_offset + i * h_stride; + + float k_val = k[t_h_i_offset]; + float q_val = q[t_h_i_offset] * scale; + float g_val = g[t_h_i_offset]; + + for (int64_t j = 0; j < head_size; j++) { + size_t t_h_j_offset = t_h_offset + j; + size_t h_2d_i_j_offset = h_2d_i_offset + j; + + float v_val = v[t_h_j_offset]; + float kv_val = v_val * k_val; + float prev_state_val = state_prev[h_2d_i_j_offset]; + float temp_val = prev_state_val * g_val + kv_val; + dst_data[t_h_j_offset] += temp_val * q_val; + state_cur[h_2d_i_j_offset] = temp_val; + } + } + } + } + #endif +} + + +void ggml_compute_forward_gla( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_gla_f32(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +// ggml_compute_forward_rwkv_wkv7 + +static void ggml_compute_forward_rwkv_wkv7_f32( + const ggml_compute_params * params, + ggml_tensor * dst) { + const int64_t T = dst->src[1]->ne[2]; + const int64_t C = dst->ne[0]; + const int64_t HEADS = dst->src[1]->ne[1]; + const int64_t n_seqs = dst->src[6]->ne[1]; + const int64_t head_size = C / HEADS; + + float * dst_data = (float *) dst->data; + float * state = ((float *) dst->data) + C * T; + + const int ith = params->ith; + const int nth = params->nth; + + if (ith >= HEADS) { + return; + } + + const int h_start = (HEADS * ith) / nth; + const int h_end = ((HEADS * (ith + 1)) / nth < HEADS) ? + (HEADS * (ith + 1)) / nth : HEADS; + + float * r = (float *) dst->src[0]->data; + float * w = (float *) dst->src[1]->data; + float * k = (float *) dst->src[2]->data; + float * v = (float *) dst->src[3]->data; + float * a = (float *) dst->src[4]->data; + float * b = (float *) dst->src[5]->data; + + int64_t t_stride = HEADS * head_size; // Same to C + + int64_t h_stride = C / HEADS; + GGML_ASSERT(C % HEADS == 0); // C must be divisible by HEADS + int64_t h_stride_2d = head_size * head_size; + + #if defined(GGML_SIMD) + for (int64_t t = 0; t < T; t++) { + int64_t t_offset = t * t_stride; + int64_t state_offset = head_size * C * (t / (T / n_seqs)); + float * state_cur = state + state_offset; + float * state_prev = t % (T / n_seqs) ? state_cur : (float*)dst->src[6]->data + state_offset; + + for (int64_t h = h_start; h < h_end; h++) { + int64_t h_offset = h * h_stride; + int64_t t_h_offset = t_offset + h_offset; + int64_t h_2d_offset = h * h_stride_2d; + + for (int64_t ii = 0; ii < head_size; ii++) { + int64_t t_h_i_offset = t_h_offset + ii; + int64_t h_2d_i_offset = h_2d_offset + ii * h_stride; + + GGML_F32_VEC v_vec = GGML_F32_VEC_SET1(v[t_h_i_offset]); + + float sa = 0; + { + GGML_F32_VEC sum[GGML_F32_ARR] = { GGML_F32_VEC_ZERO }; + GGML_F32_VEC ax[GGML_F32_ARR]; + GGML_F32_VEC ay[GGML_F32_ARR]; + for (int64_t j = 0; j < head_size; j += GGML_F32_STEP) { + for (int64_t kk = 0; kk < GGML_F32_ARR; kk++) { + ax[kk] = GGML_F32_VEC_LOAD(&a[t_h_offset + j + kk * GGML_F32_EPR]); + ay[kk] = GGML_F32_VEC_LOAD(&state_prev[h_2d_i_offset + j + kk * GGML_F32_EPR]); + sum[kk] = GGML_F32_VEC_FMA(sum[kk], ax[kk], ay[kk]); + } + } + GGML_F32_VEC_REDUCE(sa, sum); + } + + GGML_F32_VEC sa_vec = GGML_F32_VEC_SET1(sa); + + int64_t j = 0; + GGML_F32_VEC result_vec[GGML_F32_ARR] = { GGML_F32_VEC_ZERO }; + for (; j < head_size; j += GGML_F32_STEP) { + for (int64_t kk = 0; kk < GGML_F32_ARR; kk++) { + int64_t t_h_j_offset = t_h_offset + j + kk * GGML_F32_EPR; + int64_t h_2d_i_j_offset = h_2d_i_offset + j + kk * GGML_F32_EPR; + + GGML_F32_VEC r_vec = GGML_F32_VEC_LOAD(&r[t_h_j_offset]); + GGML_F32_VEC w_vec = GGML_F32_VEC_LOAD(&w[t_h_j_offset]); + GGML_F32_VEC k_vec = GGML_F32_VEC_LOAD(&k[t_h_j_offset]); + GGML_F32_VEC b_vec = GGML_F32_VEC_LOAD(&b[t_h_j_offset]); + + k_vec = GGML_F32_VEC_MUL(v_vec, k_vec); + + GGML_F32_VEC state_vec = GGML_F32_VEC_LOAD(&state_prev[h_2d_i_j_offset]); + // kv + s * decay + sa * b + state_vec = GGML_F32_VEC_FMA(k_vec, state_vec, w_vec); + state_vec = GGML_F32_VEC_FMA(state_vec, sa_vec, b_vec); + GGML_F32_VEC_STORE(&state_cur[h_2d_i_j_offset], state_vec); + + result_vec[kk] = GGML_F32_VEC_FMA(result_vec[kk], state_vec, r_vec); + } + } + GGML_F32_VEC_REDUCE(dst_data[t_h_i_offset], result_vec); + + // There shouldn't be left-overs though. + for (; j < head_size; j++) { + int64_t t_h_j_offset = t_h_offset + j; + int64_t h_2d_i_j_offset = h_2d_i_offset + j; + + float r_val = r[t_h_j_offset]; + float w_val = w[t_h_j_offset]; + float k_val = k[t_h_j_offset]; + float b_val = b[t_h_j_offset]; + float kv_val = v[t_h_i_offset] * k_val; + + float prev_state_val = state_prev[h_2d_i_j_offset]; + state_cur[h_2d_i_j_offset] = prev_state_val * w_val + kv_val + sa * b_val; + dst_data[t_h_i_offset] += state_cur[h_2d_i_j_offset] * r_val; + } + } + } + } + #else + for (int64_t t = 0; t < T; t++) { + int64_t t_offset = t * t_stride; + int64_t state_offset = head_size * C * (t / (T / n_seqs)); + float * state_cur = state + state_offset; + float * state_prev = t % (T / n_seqs) ? state_cur : (float*)dst->src[6]->data + state_offset; + + for (int64_t h = h_start; h < h_end; h++) { + int64_t h_offset = h * h_stride; + int64_t t_h_offset = t_offset + h_offset; + int64_t h_2d_offset = h * h_stride_2d; + + for (int64_t i = 0; i < head_size; i++) { + int64_t t_h_i_offset = t_h_offset + i; + int64_t h_2d_i_offset = h_2d_offset + i * h_stride; + + float v_val = v[t_h_i_offset]; + + float sa = 0, result = 0; + for (int64_t j = 0; j < head_size; j++) { + sa += a[t_h_offset + j] * state_prev[h_2d_i_offset + j]; + } + + for (int64_t j = 0; j < head_size; j++) { + int64_t t_h_j_offset = t_h_offset + j; + int64_t h_2d_i_j_offset = h_2d_i_offset + j; + + float r_val = r[t_h_j_offset]; + float w_val = w[t_h_j_offset]; + float k_val = k[t_h_j_offset]; + float b_val = b[t_h_j_offset]; + float kv_val = v_val * k_val; + float prev_state_val = state_prev[h_2d_i_j_offset]; + state_cur[h_2d_i_j_offset] = prev_state_val * w_val + kv_val + sa * b_val; + result += state_cur[h_2d_i_j_offset] * r_val; + } + dst_data[t_h_i_offset] = result; + } + } + } + #endif +} + + +void ggml_compute_forward_rwkv_wkv7( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_rwkv_wkv7_f32(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +// ggml_compute_forward_map_custom1 + +void ggml_compute_forward_map_custom1( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * a = dst->src[0]; + + struct ggml_map_custom1_op_params p; + memcpy(&p, dst->op_params, sizeof(p)); + + p.fun(dst, a, params->ith, params->nth, p.userdata); +} + +// ggml_compute_forward_map_custom2 + +void ggml_compute_forward_map_custom2( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * a = dst->src[0]; + const ggml_tensor * b = dst->src[1]; + + struct ggml_map_custom2_op_params p; + memcpy(&p, dst->op_params, sizeof(p)); + + p.fun(dst, a, b, params->ith, params->nth, p.userdata); +} + +// ggml_compute_forward_map_custom3 + +void ggml_compute_forward_map_custom3( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * a = dst->src[0]; + const ggml_tensor * b = dst->src[1]; + const ggml_tensor * c = dst->src[2]; + + struct ggml_map_custom3_op_params p; + memcpy(&p, dst->op_params, sizeof(p)); + + p.fun(dst, a, b, c, params->ith, params->nth, p.userdata); +} + +// ggml_compute_forward_custom + +void ggml_compute_forward_custom( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + + struct ggml_custom_op_params p; + memcpy(&p, dst->op_params, sizeof(p)); + + p.fun(dst, params->ith, params->nth, p.userdata); +} + +// ggml_compute_forward_cross_entropy_loss + +static void ggml_compute_forward_cross_entropy_loss_f32( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + const ggml_tensor * src1 = dst->src[1]; + + GGML_ASSERT(src0->type == GGML_TYPE_F32); + GGML_ASSERT(src1->type == GGML_TYPE_F32); + GGML_ASSERT(src0->nb[0] == ggml_type_size(src0->type)); + GGML_ASSERT(src1->nb[0] == ggml_type_size(src1->type)); + GGML_ASSERT(ggml_are_same_shape(src0, src1)); + GGML_ASSERT(ggml_is_scalar(dst)); + GGML_ASSERT(dst->type == GGML_TYPE_F32); + + // TODO: handle transposed/permuted matrices + const int64_t nc = src0->ne[0]; + const int64_t nr = ggml_nrows(src0); + + const int ith = params->ith; + const int nth = params->nth; + + float * sums = (float *) params->wdata; + float * st = ((float *) params->wdata) + nth + ith*nc; + float sum_thread = 0.0f; + + GGML_ASSERT(params->wsize >= sizeof(float) * (nth + nth * nc)); + + // rows per thread + const int64_t dr = (nr + nth - 1)/nth; + + // row range for this thread + const int64_t ir0 = dr*ith; + const int64_t ir1 = MIN(ir0 + dr, nr); + + for (int64_t i1 = ir0; i1 < ir1; ++i1) { + const float * s0 = (const float *)((const char *) src0->data + i1*src0->nb[1]); + const float * s1 = (const float *)((const char *) src1->data + i1*src1->nb[1]); + +#ifndef NDEBUG + for (int64_t i = 0; i < nc; ++i) { + //printf("p[%d] = %f\n", i, p[i]); + assert(!isnan(s0[i])); + assert(!isnan(s1[i])); + } +#endif + + float max = -INFINITY; + ggml_vec_max_f32(nc, &max, s0); + const ggml_float sum_softmax = ggml_vec_log_soft_max_f32(nc, st, s0, max); + assert(sum_softmax >= 0.0); + + ggml_vec_add1_f32(nc, st, st, -sum_softmax); + ggml_vec_mul_f32(nc, st, st, s1); + + float sum_st = 0.0f; + ggml_vec_sum_f32(nc, &sum_st, st); + sum_thread += sum_st; + +#ifndef NDEBUG + for (int64_t i = 0; i < nc; ++i) { + assert(!isnan(st[i])); + assert(!isinf(st[i])); + } +#endif + } + sums[ith] = sum_thread; + ggml_barrier(params->threadpool); + + if (ith == 0) { + float * dp = (float *) dst->data; + ggml_vec_sum_f32(nth, dp, sums); + dp[0] *= -1.0f / (float) nr; + } +} + +void ggml_compute_forward_cross_entropy_loss( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_cross_entropy_loss_f32(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +// ggml_compute_forward_cross_entropy_loss_back + +static void ggml_compute_forward_cross_entropy_loss_back_f32( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * grad = dst->src[0]; // gradient of forward pass output + const ggml_tensor * src0f = dst->src[1]; // src0 of forward pass + const ggml_tensor * src1f = dst->src[2]; // src1 of forward pass + + GGML_ASSERT(ggml_is_contiguous(dst)); + GGML_ASSERT(ggml_is_contiguous(src0f)); + GGML_ASSERT(ggml_is_contiguous(src1f)); + GGML_ASSERT(ggml_is_contiguous(grad)); + GGML_ASSERT(ggml_are_same_shape(src0f, src1f) && ggml_are_same_shape(src0f, dst)); + + const int64_t ith = params->ith; + const int64_t nth = params->nth; + + // TODO: handle transposed/permuted matrices + const int64_t nc = src0f->ne[0]; + const int64_t nr = ggml_nrows(src0f); + + // rows per thread + const int64_t dr = (nr + nth - 1)/nth; + + // row range for this thread + const int64_t ir0 = dr*ith; + const int64_t ir1 = MIN(ir0 + dr, nr); + + const float d_by_nr = ((const float *) grad->data)[0] / (float) nr; + + for (int64_t i1 = ir0; i1 < ir1; i1++) { + float * ds0 = (float *)((char *) dst->data + i1*dst->nb[1]); + const float * s0 = (const float *)((const char *) src0f->data + i1*src0f->nb[1]); + const float * s1 = (const float *)((const char *) src1f->data + i1*src1f->nb[1]); + +#ifndef NDEBUG + for (int64_t i = 0; i < nc; ++i) { + //printf("p[%d] = %f\n", i, p[i]); + assert(!isnan(s0[i])); + assert(!isnan(s1[i])); + } +#endif + + // soft_max + float max = -INFINITY; + ggml_vec_max_f32(nc, &max, s0); + const ggml_float sum = ggml_vec_soft_max_f32(nc, ds0, s0, max); + assert(sum > 0.0); + ggml_vec_scale_f32(nc, ds0, 1.0/sum); + + // grad(src0f) = (softmax(src0f) - src1f) * grad(cross_entropy_loss(src0f, src1f)) / nr + ggml_vec_sub_f32(nc, ds0, ds0, s1); + ggml_vec_scale_f32(nc, ds0, d_by_nr); + +#ifndef NDEBUG + for (int64_t i = 0; i < nc; ++i) { + assert(!isnan(ds0[i])); + assert(!isinf(ds0[i])); + } +#endif + } +} + +void ggml_compute_forward_cross_entropy_loss_back( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_cross_entropy_loss_back_f32(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} + +static void ggml_compute_forward_opt_step_adamw_f32( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + const ggml_tensor * src0_grad = dst->src[1]; + const ggml_tensor * src0_grad_m = dst->src[2]; + const ggml_tensor * src0_grad_v = dst->src[3]; + const ggml_tensor * adamw_params = dst->src[4]; + + GGML_ASSERT(ggml_are_same_shape(src0, src0_grad)); + GGML_ASSERT(ggml_are_same_shape(src0, src0_grad_m)); + GGML_ASSERT(ggml_are_same_shape(src0, src0_grad_v)); + GGML_ASSERT(ggml_nelements(adamw_params) == 7); + + const int ith = params->ith; + const int nth = params->nth; + + const int nr = ggml_nrows(src0); + + GGML_TENSOR_UNARY_OP_LOCALS + GGML_ASSERT(nb00 == sizeof(float)); + + // rows per thread + const int dr = (nr + nth - 1)/nth; + + // row range for this thread + const int ir0 = dr*ith; + const int ir1 = MIN(ir0 + dr, nr); + + const float * adamw_params_ptr = ggml_get_data_f32(adamw_params); + const float alpha = adamw_params_ptr[0]; + const float beta1 = adamw_params_ptr[1]; + const float beta2 = adamw_params_ptr[2]; + const float eps = adamw_params_ptr[3]; + const float wd = adamw_params_ptr[4]; + const float beta1h = adamw_params_ptr[5]; + const float beta2h = adamw_params_ptr[6]; + + for (int ir = ir0; ir < ir1; ++ir) { + const int64_t i03 = ir/(ne02*ne01); + const int64_t i02 = (ir - i03*ne02*ne01)/ne01; + const int64_t i01 = (ir - i03*ne02*ne01 - i02*ne01); + + const size_t offset = i03*nb03 + i02*nb02 + i01*nb01; + + float * w = (float *) ((char *) src0->data + offset); // weight + const float * g = (const float *) ((const char *) src0_grad->data + offset); // grad + float * m = (float *) ((char *) src0_grad_m->data + offset); + float * v = (float *) ((char *) src0_grad_v->data + offset); + + for (int i00 = 0; i00 < ne00; ++i00) { + m[i00] = m[i00]*beta1 + g[i00]*(1.0f - beta1); + v[i00] = v[i00]*beta2 + g[i00]*g[i00]*(1.0f - beta2); + + const float mh = m[i00]*beta1h; + const float vh = sqrtf(v[i00]*beta2h) + eps; + + // The weight decay is applied independently of the Adam momenta m and v. + // This is NOT equivalent to l2 regularization that adds w[i00]*w[i00] to the loss. + // See: https://arxiv.org/pdf/1711.05101v3.pdf + w[i00] = w[i00]*(1.0f - alpha*wd) - alpha*mh/vh; + } + } +} + +void ggml_compute_forward_opt_step_adamw( + const ggml_compute_params * params, + ggml_tensor * dst) { + + const ggml_tensor * src0 = dst->src[0]; + + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_opt_step_adamw_f32(params, dst); + } break; + default: + { + GGML_ABORT("fatal error"); + } + } +} diff --git a/ggml/src/ggml-cpu/ops.h b/ggml/src/ggml-cpu/ops.h new file mode 100644 index 000000000..410a37204 --- /dev/null +++ b/ggml/src/ggml-cpu/ops.h @@ -0,0 +1,109 @@ +#pragma once + +#include "ggml.h" + +// +// cache line +// + +#if defined(__cpp_lib_hardware_interference_size) +#define CACHE_LINE_SIZE std::hardware_destructive_interference_size +#else +#if defined(__POWER9_VECTOR__) +#define CACHE_LINE_SIZE 128 +#elif defined(__VXE__) || defined(__VXE2__) +#define CACHE_LINE_SIZE 256 +#else +#define CACHE_LINE_SIZE 64 +#endif +#endif + +static const size_t CACHE_LINE_SIZE_F32 = CACHE_LINE_SIZE/sizeof(float); + +#ifdef __cplusplus +extern "C" { +#endif + +void ggml_compute_forward_dup(const struct ggml_compute_params * params, struct ggml_tensor * dst); +void ggml_compute_forward_add(const struct ggml_compute_params * params, struct ggml_tensor * dst); +void ggml_compute_forward_add1(const struct ggml_compute_params * params, struct ggml_tensor * dst); +void ggml_compute_forward_acc(const struct ggml_compute_params * params, struct ggml_tensor * dst); +void ggml_compute_forward_sum(const struct ggml_compute_params * params, struct ggml_tensor * dst); +void ggml_compute_forward_sum_rows(const struct ggml_compute_params * params, struct ggml_tensor * dst); +void ggml_compute_forward_mean(const struct ggml_compute_params * params, struct ggml_tensor * dst); +void ggml_compute_forward_argmax(const struct ggml_compute_params * params, struct ggml_tensor * dst); +void ggml_compute_forward_count_equal(const struct ggml_compute_params * params, struct ggml_tensor * dst); +void ggml_compute_forward_repeat(const struct ggml_compute_params * params, struct ggml_tensor * dst); +void ggml_compute_forward_repeat_back(const struct ggml_compute_params * params, struct ggml_tensor * dst); +void ggml_compute_forward_concat(const struct ggml_compute_params * params, struct ggml_tensor * dst); +void ggml_compute_forward_silu_back(const struct ggml_compute_params * params, struct ggml_tensor * dst); +void ggml_compute_forward_norm(const struct ggml_compute_params * params, struct ggml_tensor * dst); +void ggml_compute_forward_rms_norm(const struct ggml_compute_params * params, struct ggml_tensor * dst); +void ggml_compute_forward_rms_norm_back(const struct ggml_compute_params * params, struct ggml_tensor * dst); +void ggml_compute_forward_group_norm(const struct ggml_compute_params * params, struct ggml_tensor * dst); +void ggml_compute_forward_l2_norm(const struct ggml_compute_params * params, struct ggml_tensor * dst); +void ggml_compute_forward_out_prod(const struct ggml_compute_params * params, struct ggml_tensor * dst); +void ggml_compute_forward_scale(const struct ggml_compute_params * params, struct ggml_tensor * dst); +void ggml_compute_forward_set(const struct ggml_compute_params * params, struct ggml_tensor * dst); +void ggml_compute_forward_cpy(const struct ggml_compute_params * params, struct ggml_tensor * dst); +void ggml_compute_forward_cont(const struct ggml_compute_params * params, struct ggml_tensor * dst); +void ggml_compute_forward_reshape(const struct ggml_compute_params * params, struct ggml_tensor * dst); +void ggml_compute_forward_view(const struct ggml_compute_params * params, struct ggml_tensor * dst); +void ggml_compute_forward_permute(const struct ggml_compute_params * params, struct ggml_tensor * dst); +void ggml_compute_forward_transpose(const struct ggml_compute_params * params, struct ggml_tensor * dst); +void ggml_compute_forward_get_rows(const struct ggml_compute_params * params, struct ggml_tensor * dst); +void ggml_compute_forward_get_rows_back(const struct ggml_compute_params * params, struct ggml_tensor * dst); +void ggml_compute_forward_diag(const struct ggml_compute_params * params, struct ggml_tensor * dst); +void ggml_compute_forward_diag_mask_inf(const struct ggml_compute_params * params, struct ggml_tensor * dst); +void ggml_compute_forward_diag_mask_zero(const struct ggml_compute_params * params, struct ggml_tensor * dst); +void ggml_compute_forward_soft_max(const struct ggml_compute_params * params, struct ggml_tensor * dst); +void ggml_compute_forward_soft_max_ext_back(const struct ggml_compute_params * params, struct ggml_tensor * dst); +void ggml_compute_forward_rope(const struct ggml_compute_params * params, struct ggml_tensor * dst); +void ggml_compute_forward_rope_back(const struct ggml_compute_params * params, struct ggml_tensor * dst); +void ggml_compute_forward_clamp(const struct ggml_compute_params * params, struct ggml_tensor * dst); +void ggml_compute_forward_conv_transpose_1d(const struct ggml_compute_params * params, struct ggml_tensor * dst); +void ggml_compute_forward_im2col(const struct ggml_compute_params * params, struct ggml_tensor * dst); +void ggml_compute_forward_im2col_back_f32(const struct ggml_compute_params * params, struct ggml_tensor * dst); +void ggml_compute_forward_conv_transpose_2d(const struct ggml_compute_params * params, struct ggml_tensor * dst); +void ggml_compute_forward_pool_1d(const struct ggml_compute_params * params, struct ggml_tensor * dst); +void ggml_compute_forward_pool_2d(const struct ggml_compute_params * params, struct ggml_tensor * dst); +void ggml_compute_forward_pool_2d_back(const struct ggml_compute_params * params, struct ggml_tensor * dst); +void ggml_compute_forward_upscale(const struct ggml_compute_params * params, struct ggml_tensor * dst); +void ggml_compute_forward_pad(const struct ggml_compute_params * params, struct ggml_tensor * dst); +void ggml_compute_forward_pad_reflect_1d(const struct ggml_compute_params * params, struct ggml_tensor * dst); +void ggml_compute_forward_arange(const struct ggml_compute_params * params, struct ggml_tensor * dst); +void ggml_compute_forward_timestep_embedding(const struct ggml_compute_params * params, struct ggml_tensor * dst); +void ggml_compute_forward_argsort(const struct ggml_compute_params * params, struct ggml_tensor * dst); +void ggml_compute_forward_leaky_relu(const struct ggml_compute_params * params, struct ggml_tensor * dst); +void ggml_compute_forward_flash_attn_ext( + const struct ggml_compute_params * params, + const struct ggml_tensor * q, + const struct ggml_tensor * k, + const struct ggml_tensor * v, + const struct ggml_tensor * mask, + struct ggml_tensor * dst); +void ggml_compute_forward_flash_attn_back( + const struct ggml_compute_params * params, + const bool masked, + struct ggml_tensor * dst); +void ggml_compute_forward_ssm_conv(const struct ggml_compute_params * params, struct ggml_tensor * dst); +void ggml_compute_forward_ssm_scan(const struct ggml_compute_params * params, struct ggml_tensor * dst); +void ggml_compute_forward_win_part(const struct ggml_compute_params * params, struct ggml_tensor * dst); +void ggml_compute_forward_win_unpart(const struct ggml_compute_params * params, struct ggml_tensor * dst); +void ggml_compute_forward_unary(const struct ggml_compute_params * params, struct ggml_tensor * dst); +void ggml_compute_forward_get_rel_pos(const struct ggml_compute_params * params, struct ggml_tensor * dst); +void ggml_compute_forward_add_rel_pos(const struct ggml_compute_params * params, struct ggml_tensor * dst); +void ggml_compute_forward_rwkv_wkv6(const struct ggml_compute_params * params, struct ggml_tensor * dst); +void ggml_compute_forward_rwkv_wkv7(const struct ggml_compute_params * params, struct ggml_tensor * dst); +void ggml_compute_forward_gla(const struct ggml_compute_params * params, struct ggml_tensor * dst); +void ggml_compute_forward_map_custom1(const struct ggml_compute_params * params, struct ggml_tensor * dst); +void ggml_compute_forward_map_custom2(const struct ggml_compute_params * params, struct ggml_tensor * dst); +void ggml_compute_forward_map_custom3(const struct ggml_compute_params * params, struct ggml_tensor * dst); +void ggml_compute_forward_custom(const struct ggml_compute_params * params, struct ggml_tensor * dst); +void ggml_compute_forward_cross_entropy_loss(const struct ggml_compute_params * params, struct ggml_tensor * dst); +void ggml_compute_forward_cross_entropy_loss_back(const struct ggml_compute_params * params, struct ggml_tensor * dst); +void ggml_compute_forward_opt_step_adamw(const struct ggml_compute_params * params, struct ggml_tensor * dst); + +#ifdef __cplusplus +} +#endif diff --git a/ggml/src/ggml-cpu/simd-mappings.h b/ggml/src/ggml-cpu/simd-mappings.h new file mode 100644 index 000000000..04d10cec2 --- /dev/null +++ b/ggml/src/ggml-cpu/simd-mappings.h @@ -0,0 +1,892 @@ +#pragma once + +#include "ggml-cpu-impl.h" + +// +// simd mappings +// + +// we define a common set of C macros which map to specific intrinsics based on the current architecture +// we then implement the fundamental computation operations below using only these macros +// adding support for new architectures requires to define the corresponding SIMD macros +// +// GGML_F32_STEP / GGML_F16_STEP +// number of elements to process in a single step +// +// GGML_F32_EPR / GGML_F16_EPR +// number of elements to fit in a single register +// + +#if defined(__ARM_NEON) && defined(__ARM_FEATURE_FMA) + +#define GGML_SIMD + +// F32 NEON + +#define GGML_F32_STEP 16 +#define GGML_F32_EPR 4 + +#define GGML_F32x4 float32x4_t +#define GGML_F32x4_ZERO vdupq_n_f32(0.0f) +#define GGML_F32x4_SET1(x) vdupq_n_f32(x) +#define GGML_F32x4_LOAD vld1q_f32 +#define GGML_F32x4_STORE vst1q_f32 +#define GGML_F32x4_FMA(a, b, c) vfmaq_f32(a, b, c) +#define GGML_F32x4_ADD vaddq_f32 +#define GGML_F32x4_MUL vmulq_f32 +#define GGML_F32x4_REDUCE_ONE(x) vaddvq_f32(x) +#define GGML_F32x4_REDUCE(res, x) \ +{ \ + int offset = GGML_F32_ARR >> 1; \ + for (int i = 0; i < offset; ++i) { \ + (x)[i] = vaddq_f32((x)[i], (x)[offset+i]); \ + } \ + offset >>= 1; \ + for (int i = 0; i < offset; ++i) { \ + (x)[i] = vaddq_f32((x)[i], (x)[offset+i]); \ + } \ + offset >>= 1; \ + for (int i = 0; i < offset; ++i) { \ + (x)[i] = vaddq_f32((x)[i], (x)[offset+i]); \ + } \ + (res) = (ggml_float) GGML_F32x4_REDUCE_ONE((x)[0]); \ +} + +#define GGML_F32_VEC GGML_F32x4 +#define GGML_F32_VEC_ZERO GGML_F32x4_ZERO +#define GGML_F32_VEC_SET1 GGML_F32x4_SET1 +#define GGML_F32_VEC_LOAD GGML_F32x4_LOAD +#define GGML_F32_VEC_STORE GGML_F32x4_STORE +#define GGML_F32_VEC_FMA GGML_F32x4_FMA +#define GGML_F32_VEC_ADD GGML_F32x4_ADD +#define GGML_F32_VEC_MUL GGML_F32x4_MUL +#define GGML_F32_VEC_REDUCE GGML_F32x4_REDUCE + +// F16 NEON + +#if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) + #define GGML_F16_STEP 32 + #define GGML_F16_EPR 8 + + #define GGML_F16x8 float16x8_t + #define GGML_F16x8_ZERO vdupq_n_f16(0.0f) + #define GGML_F16x8_SET1(x) vdupq_n_f16(x) + #define GGML_F16x8_LOAD(x) vld1q_f16((const __fp16 *)(x)) + #define GGML_F16x8_STORE vst1q_f16 + #define GGML_F16x8_FMA(a, b, c) vfmaq_f16(a, b, c) + #define GGML_F16x8_ADD vaddq_f16 + #define GGML_F16x8_MUL vmulq_f16 + #define GGML_F16x8_REDUCE(res, x) \ + do { \ + int offset = GGML_F16_ARR >> 1; \ + for (int i = 0; i < offset; ++i) { \ + (x)[i] = vaddq_f16((x)[i], (x)[offset+i]); \ + } \ + offset >>= 1; \ + for (int i = 0; i < offset; ++i) { \ + (x)[i] = vaddq_f16((x)[i], (x)[offset+i]); \ + } \ + offset >>= 1; \ + for (int i = 0; i < offset; ++i) { \ + (x)[i] = vaddq_f16((x)[i], (x)[offset+i]); \ + } \ + const float32x4_t t0 = vcvt_f32_f16(vget_low_f16 ((x)[0])); \ + const float32x4_t t1 = vcvt_f32_f16(vget_high_f16((x)[0])); \ + (res) = (ggml_float) vaddvq_f32(vaddq_f32(t0, t1)); \ + } while (0) + + #define GGML_F16_VEC GGML_F16x8 + #define GGML_F16_VEC_ZERO GGML_F16x8_ZERO + #define GGML_F16_VEC_SET1 GGML_F16x8_SET1 + #define GGML_F16_VEC_LOAD(p, i) GGML_F16x8_LOAD(p) + #define GGML_F16_VEC_STORE(p, r, i) GGML_F16x8_STORE((__fp16 *)(p), (r)[i]) + #define GGML_F16_VEC_FMA GGML_F16x8_FMA + #define GGML_F16_VEC_ADD GGML_F16x8_ADD + #define GGML_F16_VEC_MUL GGML_F16x8_MUL + #define GGML_F16_VEC_REDUCE GGML_F16x8_REDUCE +#else + // if FP16 vector arithmetic is not supported, we use FP32 instead + // and take advantage of the vcvt_ functions to convert to/from FP16 + + #define GGML_F16_STEP 16 + #define GGML_F16_EPR 4 + + #define GGML_F32Cx4 float32x4_t + #define GGML_F32Cx4_ZERO vdupq_n_f32(0.0f) + #define GGML_F32Cx4_SET1(x) vdupq_n_f32(x) + #define GGML_F32Cx4_LOAD(x) vcvt_f32_f16(vld1_f16((const __fp16 *)(x))) + #define GGML_F32Cx4_STORE(x, y) vst1_f16(x, vcvt_f16_f32(y)) + #define GGML_F32Cx4_FMA(a, b, c) vfmaq_f32(a, b, c) + #define GGML_F32Cx4_ADD vaddq_f32 + #define GGML_F32Cx4_MUL vmulq_f32 + #define GGML_F32Cx4_REDUCE GGML_F32x4_REDUCE + + #define GGML_F16_VEC GGML_F32Cx4 + #define GGML_F16_VEC_ZERO GGML_F32Cx4_ZERO + #define GGML_F16_VEC_SET1 GGML_F32Cx4_SET1 + #define GGML_F16_VEC_LOAD(p, i) GGML_F32Cx4_LOAD(p) + #define GGML_F16_VEC_STORE(p, r, i) GGML_F32Cx4_STORE((__fp16 *)(p), r[i]) + #define GGML_F16_VEC_FMA GGML_F32Cx4_FMA + #define GGML_F16_VEC_ADD GGML_F32Cx4_ADD + #define GGML_F16_VEC_MUL GGML_F32Cx4_MUL + #define GGML_F16_VEC_REDUCE GGML_F32Cx4_REDUCE +#endif + +#elif defined(__AVX512F__) + +#define GGML_SIMD + +// F32 AVX512 + +#define GGML_F32_STEP 64 +#define GGML_F32_EPR 16 + +#define GGML_F32x16 __m512 +#define GGML_F32x16_ZERO _mm512_setzero_ps() +#define GGML_F32x16_SET1(x) _mm512_set1_ps(x) +#define GGML_F32x16_LOAD _mm512_loadu_ps +#define GGML_F32x16_STORE _mm512_storeu_ps +// _mm512_fmadd_ps is defined in AVX512F so no guard is required +#define GGML_F32x16_FMA(a, b, c) _mm512_fmadd_ps(b, c, a) +#define GGML_F32x16_ADD _mm512_add_ps +#define GGML_F32x16_MUL _mm512_mul_ps +#define GGML_F32x16_REDUCE(res, x) \ +do { \ + int offset = GGML_F32_ARR >> 1; \ + for (int i = 0; i < offset; ++i) { \ + x[i] = _mm512_add_ps(x[i], x[offset+i]); \ + } \ + offset >>= 1; \ + for (int i = 0; i < offset; ++i) { \ + x[i] = _mm512_add_ps(x[i], x[offset+i]); \ + } \ + offset >>= 1; \ + for (int i = 0; i < offset; ++i) { \ + x[i] = _mm512_add_ps(x[i], x[offset+i]); \ + } \ + res = (ggml_float) _mm512_reduce_add_ps(x[0]); \ +} while (0) + +// TODO: is this optimal ? + +#define GGML_F32_VEC GGML_F32x16 +#define GGML_F32_VEC_ZERO GGML_F32x16_ZERO +#define GGML_F32_VEC_SET1 GGML_F32x16_SET1 +#define GGML_F32_VEC_LOAD GGML_F32x16_LOAD +#define GGML_F32_VEC_STORE GGML_F32x16_STORE +#define GGML_F32_VEC_FMA GGML_F32x16_FMA +#define GGML_F32_VEC_ADD GGML_F32x16_ADD +#define GGML_F32_VEC_MUL GGML_F32x16_MUL +#define GGML_F32_VEC_REDUCE GGML_F32x16_REDUCE + +// F16 AVX512 + +// F16 AVX + +#define GGML_F16_STEP 64 +#define GGML_F16_EPR 16 + +// AVX512 has FP16 extension (AVX512_FP16) but I don't have it on my machine so I use FP32 instead + +#define GGML_F32Cx16 __m512 +#define GGML_F32Cx16_ZERO _mm512_setzero_ps() +#define GGML_F32Cx16_SET1(x) _mm512_set1_ps(x) + +// unlike _mm256_cvt intrinsics that require F16C, _mm512_cvt is defined in AVX512F +// so F16C guard isn't required +#define GGML_F32Cx16_LOAD(x) _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)(x))) +#define GGML_F32Cx16_STORE(x, y) _mm256_storeu_si256((__m256i *)(x), _mm512_cvtps_ph(y, 0)) + +#define GGML_F32Cx16_FMA(a, b, c) _mm512_fmadd_ps(b, c, a) +#define GGML_F32Cx16_ADD _mm512_add_ps +#define GGML_F32Cx16_MUL _mm512_mul_ps +#define GGML_F32Cx16_REDUCE(res, x) \ +do { \ + int offset = GGML_F32_ARR >> 1; \ + for (int i = 0; i < offset; ++i) { \ + x[i] = _mm512_add_ps(x[i], x[offset+i]); \ + } \ + offset >>= 1; \ + for (int i = 0; i < offset; ++i) { \ + x[i] = _mm512_add_ps(x[i], x[offset+i]); \ + } \ + offset >>= 1; \ + for (int i = 0; i < offset; ++i) { \ + x[i] = _mm512_add_ps(x[i], x[offset+i]); \ + } \ + res = (ggml_float) _mm512_reduce_add_ps(x[0]); \ +} while (0) + +#define GGML_F16_VEC GGML_F32Cx16 +#define GGML_F16_VEC_ZERO GGML_F32Cx16_ZERO +#define GGML_F16_VEC_SET1 GGML_F32Cx16_SET1 +#define GGML_F16_VEC_LOAD(p, i) GGML_F32Cx16_LOAD(p) +#define GGML_F16_VEC_STORE(p, r, i) GGML_F32Cx16_STORE(p, r[i]) +#define GGML_F16_VEC_FMA GGML_F32Cx16_FMA +#define GGML_F16_VEC_ADD GGML_F32Cx16_ADD +#define GGML_F16_VEC_MUL GGML_F32Cx16_MUL + +#define GGML_F16_VEC_REDUCE GGML_F32Cx16_REDUCE +#elif defined(__AVX__) + +#define GGML_SIMD + +// F32 AVX + +#define GGML_F32_STEP 32 +#define GGML_F32_EPR 8 + +#define GGML_F32x8 __m256 +#define GGML_F32x8_ZERO _mm256_setzero_ps() +#define GGML_F32x8_SET1(x) _mm256_set1_ps(x) +#define GGML_F32x8_LOAD _mm256_loadu_ps +#define GGML_F32x8_STORE _mm256_storeu_ps +#if defined(__FMA__) + #define GGML_F32x8_FMA(a, b, c) _mm256_fmadd_ps(b, c, a) +#else + #define GGML_F32x8_FMA(a, b, c) _mm256_add_ps(_mm256_mul_ps(b, c), a) +#endif +#define GGML_F32x8_ADD _mm256_add_ps +#define GGML_F32x8_MUL _mm256_mul_ps +#define GGML_F32x8_REDUCE(res, x) \ +do { \ + int offset = GGML_F32_ARR >> 1; \ + for (int i = 0; i < offset; ++i) { \ + x[i] = _mm256_add_ps(x[i], x[offset+i]); \ + } \ + offset >>= 1; \ + for (int i = 0; i < offset; ++i) { \ + x[i] = _mm256_add_ps(x[i], x[offset+i]); \ + } \ + offset >>= 1; \ + for (int i = 0; i < offset; ++i) { \ + x[i] = _mm256_add_ps(x[i], x[offset+i]); \ + } \ + const __m128 t0 = _mm_add_ps(_mm256_castps256_ps128(x[0]), \ + _mm256_extractf128_ps(x[0], 1)); \ + const __m128 t1 = _mm_hadd_ps(t0, t0); \ + res = (ggml_float) _mm_cvtss_f32(_mm_hadd_ps(t1, t1)); \ +} while (0) +// TODO: is this optimal ? + +#define GGML_F32_VEC GGML_F32x8 +#define GGML_F32_VEC_ZERO GGML_F32x8_ZERO +#define GGML_F32_VEC_SET1 GGML_F32x8_SET1 +#define GGML_F32_VEC_LOAD GGML_F32x8_LOAD +#define GGML_F32_VEC_STORE GGML_F32x8_STORE +#define GGML_F32_VEC_FMA GGML_F32x8_FMA +#define GGML_F32_VEC_ADD GGML_F32x8_ADD +#define GGML_F32_VEC_MUL GGML_F32x8_MUL +#define GGML_F32_VEC_REDUCE GGML_F32x8_REDUCE + +// F16 AVX + +#define GGML_F16_STEP 32 +#define GGML_F16_EPR 8 + +// F16 arithmetic is not supported by AVX, so we use F32 instead + +#define GGML_F32Cx8 __m256 +#define GGML_F32Cx8_ZERO _mm256_setzero_ps() +#define GGML_F32Cx8_SET1(x) _mm256_set1_ps(x) + +#if defined(__F16C__) +// the _mm256_cvt intrinsics require F16C +#define GGML_F32Cx8_LOAD(x) _mm256_cvtph_ps(_mm_loadu_si128((const __m128i *)(x))) +#define GGML_F32Cx8_STORE(x, y) _mm_storeu_si128((__m128i *)(x), _mm256_cvtps_ph(y, 0)) +#else +static inline __m256 __avx_f32cx8_load(const ggml_fp16_t * x) { + float tmp[8]; + + for (int i = 0; i < 8; i++) { + tmp[i] = GGML_FP16_TO_FP32(x[i]); + } + + return _mm256_loadu_ps(tmp); +} +static inline void __avx_f32cx8_store(ggml_fp16_t *x, __m256 y) { + float arr[8]; + + _mm256_storeu_ps(arr, y); + + for (int i = 0; i < 8; i++) + x[i] = GGML_FP32_TO_FP16(arr[i]); +} +#define GGML_F32Cx8_LOAD(x) __avx_f32cx8_load(x) +#define GGML_F32Cx8_STORE(x, y) __avx_f32cx8_store(x, y) +#endif + +#define GGML_F32Cx8_FMA GGML_F32x8_FMA +#define GGML_F32Cx8_ADD _mm256_add_ps +#define GGML_F32Cx8_MUL _mm256_mul_ps +#define GGML_F32Cx8_REDUCE GGML_F32x8_REDUCE + +#define GGML_F16_VEC GGML_F32Cx8 +#define GGML_F16_VEC_ZERO GGML_F32Cx8_ZERO +#define GGML_F16_VEC_SET1 GGML_F32Cx8_SET1 +#define GGML_F16_VEC_LOAD(p, i) GGML_F32Cx8_LOAD(p) +#define GGML_F16_VEC_STORE(p, r, i) GGML_F32Cx8_STORE(p, r[i]) +#define GGML_F16_VEC_FMA GGML_F32Cx8_FMA +#define GGML_F16_VEC_ADD GGML_F32Cx8_ADD +#define GGML_F16_VEC_MUL GGML_F32Cx8_MUL +#define GGML_F16_VEC_REDUCE GGML_F32Cx8_REDUCE + +#elif defined(__POWER9_VECTOR__) + +#define GGML_SIMD + +// F32 POWER9 + +#define GGML_F32_STEP 32 +#define GGML_F32_EPR 4 + +#define GGML_F32x4 vector float +#define GGML_F32x4_ZERO 0.0f +#define GGML_F32x4_SET1 vec_splats +#define GGML_F32x4_LOAD(p) vec_xl(0, p) +#define GGML_F32x4_STORE(p, r) vec_xst(r, 0, p) +#define GGML_F32x4_FMA(a, b, c) vec_madd(b, c, a) +#define GGML_F32x4_ADD vec_add +#define GGML_F32x4_MUL vec_mul +#define GGML_F32x4_REDUCE(res, x) \ +{ \ + int offset = GGML_F32_ARR >> 1; \ + for (int i = 0; i < offset; ++i) { \ + x[i] = vec_add(x[i], x[offset+i]); \ + } \ + offset >>= 1; \ + for (int i = 0; i < offset; ++i) { \ + x[i] = vec_add(x[i], x[offset+i]); \ + } \ + offset >>= 1; \ + for (int i = 0; i < offset; ++i) { \ + x[i] = vec_add(x[i], x[offset+i]); \ + } \ + res = vec_extract(x[0], 0) + \ + vec_extract(x[0], 1) + \ + vec_extract(x[0], 2) + \ + vec_extract(x[0], 3); \ +} + +#define GGML_F32_VEC GGML_F32x4 +#define GGML_F32_VEC_ZERO GGML_F32x4_ZERO +#define GGML_F32_VEC_SET1 GGML_F32x4_SET1 +#define GGML_F32_VEC_LOAD GGML_F32x4_LOAD +#define GGML_F32_VEC_STORE GGML_F32x4_STORE +#define GGML_F32_VEC_FMA GGML_F32x4_FMA +#define GGML_F32_VEC_ADD GGML_F32x4_ADD +#define GGML_F32_VEC_MUL GGML_F32x4_MUL +#define GGML_F32_VEC_REDUCE GGML_F32x4_REDUCE + +// F16 POWER9 +#define GGML_F16_STEP GGML_F32_STEP +#define GGML_F16_EPR GGML_F32_EPR +#define GGML_F16_VEC GGML_F32x4 +#define GGML_F16_VEC_ZERO GGML_F32x4_ZERO +#define GGML_F16_VEC_SET1 GGML_F32x4_SET1 +#define GGML_F16_VEC_FMA GGML_F32x4_FMA +#define GGML_F16_VEC_ADD GGML_F32x4_ADD +#define GGML_F16_VEC_MUL GGML_F32x4_MUL +#define GGML_F16_VEC_REDUCE GGML_F32x4_REDUCE +// Use vec_xl, not vec_ld, in case the load address is not aligned. +#define GGML_F16_VEC_LOAD(p, i) (i & 0x1) ? \ + vec_extract_fp32_from_shorth(vec_xl(0, p - GGML_F16_EPR)) : \ + vec_extract_fp32_from_shortl(vec_xl(0, p)) +static inline unsigned char ggml_endian_byte(int i) { + uint16_t tmp_val = 1; + return ((unsigned char *)&tmp_val)[i]; +} +#define GGML_ENDIAN_BYTE(i) ggml_endian_byte(i) +#define GGML_F16_VEC_STORE(p, r, i) \ + if (i & 0x1) \ + vec_xst(vec_pack_to_short_fp32(r[i - GGML_ENDIAN_BYTE(1)], \ + r[i - GGML_ENDIAN_BYTE(0)]), \ + 0, p - GGML_F16_EPR) + +#elif defined(__wasm_simd128__) + +#define GGML_SIMD + +// F32 WASM + +#define GGML_F32_STEP 16 +#define GGML_F32_EPR 4 + +#define GGML_F32x4 v128_t +#define GGML_F32x4_ZERO wasm_f32x4_splat(0.0f) +#define GGML_F32x4_SET1(x) wasm_f32x4_splat(x) +#define GGML_F32x4_LOAD wasm_v128_load +#define GGML_F32x4_STORE wasm_v128_store +#define GGML_F32x4_FMA(a, b, c) wasm_f32x4_add(wasm_f32x4_mul(b, c), a) +#define GGML_F32x4_ADD wasm_f32x4_add +#define GGML_F32x4_MUL wasm_f32x4_mul +#define GGML_F32x4_REDUCE(res, x) \ +{ \ + int offset = GGML_F32_ARR >> 1; \ + for (int i = 0; i < offset; ++i) { \ + x[i] = wasm_f32x4_add(x[i], x[offset+i]); \ + } \ + offset >>= 1; \ + for (int i = 0; i < offset; ++i) { \ + x[i] = wasm_f32x4_add(x[i], x[offset+i]); \ + } \ + offset >>= 1; \ + for (int i = 0; i < offset; ++i) { \ + x[i] = wasm_f32x4_add(x[i], x[offset+i]); \ + } \ + res = wasm_f32x4_extract_lane(x[0], 0) + \ + wasm_f32x4_extract_lane(x[0], 1) + \ + wasm_f32x4_extract_lane(x[0], 2) + \ + wasm_f32x4_extract_lane(x[0], 3); \ +} + +#define GGML_F32_VEC GGML_F32x4 +#define GGML_F32_VEC_ZERO GGML_F32x4_ZERO +#define GGML_F32_VEC_SET1 GGML_F32x4_SET1 +#define GGML_F32_VEC_LOAD GGML_F32x4_LOAD +#define GGML_F32_VEC_STORE GGML_F32x4_STORE +#define GGML_F32_VEC_FMA GGML_F32x4_FMA +#define GGML_F32_VEC_ADD GGML_F32x4_ADD +#define GGML_F32_VEC_MUL GGML_F32x4_MUL +#define GGML_F32_VEC_REDUCE GGML_F32x4_REDUCE + +// F16 WASM + +#define GGML_F16_STEP 16 +#define GGML_F16_EPR 4 + +inline static v128_t __wasm_f16x4_load(const ggml_fp16_t * p) { + float tmp[4]; + + tmp[0] = GGML_FP16_TO_FP32(p[0]); + tmp[1] = GGML_FP16_TO_FP32(p[1]); + tmp[2] = GGML_FP16_TO_FP32(p[2]); + tmp[3] = GGML_FP16_TO_FP32(p[3]); + + return wasm_v128_load(tmp); +} + +inline static void __wasm_f16x4_store(ggml_fp16_t * p, v128_t x) { + float tmp[4]; + + wasm_v128_store(tmp, x); + + p[0] = GGML_FP32_TO_FP16(tmp[0]); + p[1] = GGML_FP32_TO_FP16(tmp[1]); + p[2] = GGML_FP32_TO_FP16(tmp[2]); + p[3] = GGML_FP32_TO_FP16(tmp[3]); +} + +#define GGML_F16x4 v128_t +#define GGML_F16x4_ZERO wasm_f32x4_splat(0.0f) +#define GGML_F16x4_SET1(x) wasm_f32x4_splat(x) +#define GGML_F16x4_LOAD(x) __wasm_f16x4_load(x) +#define GGML_F16x4_STORE(x, y) __wasm_f16x4_store(x, y) +#define GGML_F16x4_FMA GGML_F32x4_FMA +#define GGML_F16x4_ADD wasm_f32x4_add +#define GGML_F16x4_MUL wasm_f32x4_mul +#define GGML_F16x4_REDUCE(res, x) \ +{ \ + int offset = GGML_F16_ARR >> 1; \ + for (int i = 0; i < offset; ++i) { \ + x[i] = wasm_f32x4_add(x[i], x[offset+i]); \ + } \ + offset >>= 1; \ + for (int i = 0; i < offset; ++i) { \ + x[i] = wasm_f32x4_add(x[i], x[offset+i]); \ + } \ + offset >>= 1; \ + for (int i = 0; i < offset; ++i) { \ + x[i] = wasm_f32x4_add(x[i], x[offset+i]); \ + } \ + res = (ggml_float) (wasm_f32x4_extract_lane(x[0], 0) + \ + wasm_f32x4_extract_lane(x[0], 1) + \ + wasm_f32x4_extract_lane(x[0], 2) + \ + wasm_f32x4_extract_lane(x[0], 3)); \ +} + +#define GGML_F16_VEC GGML_F16x4 +#define GGML_F16_VEC_ZERO GGML_F16x4_ZERO +#define GGML_F16_VEC_SET1 GGML_F16x4_SET1 +#define GGML_F16_VEC_LOAD(p, i) GGML_F16x4_LOAD(p) +#define GGML_F16_VEC_STORE(p, r, i) GGML_F16x4_STORE(p, r[i]) +#define GGML_F16_VEC_FMA GGML_F16x4_FMA +#define GGML_F16_VEC_ADD GGML_F16x4_ADD +#define GGML_F16_VEC_MUL GGML_F16x4_MUL +#define GGML_F16_VEC_REDUCE GGML_F16x4_REDUCE + +#elif defined(__SSE3__) + +#define GGML_SIMD + +// F32 SSE + +#define GGML_F32_STEP 32 +#define GGML_F32_EPR 4 + +#define GGML_F32x4 __m128 +#define GGML_F32x4_ZERO _mm_setzero_ps() +#define GGML_F32x4_SET1(x) _mm_set1_ps(x) +#define GGML_F32x4_LOAD _mm_loadu_ps +#define GGML_F32x4_STORE _mm_storeu_ps +#if defined(__FMA__) + // TODO: Does this work? + #define GGML_F32x4_FMA(a, b, c) _mm_fmadd_ps(b, c, a) +#else + #define GGML_F32x4_FMA(a, b, c) _mm_add_ps(_mm_mul_ps(b, c), a) +#endif +#define GGML_F32x4_ADD _mm_add_ps +#define GGML_F32x4_MUL _mm_mul_ps +#define GGML_F32x4_REDUCE(res, x) \ +{ \ + int offset = GGML_F32_ARR >> 1; \ + for (int i = 0; i < offset; ++i) { \ + x[i] = _mm_add_ps(x[i], x[offset+i]); \ + } \ + offset >>= 1; \ + for (int i = 0; i < offset; ++i) { \ + x[i] = _mm_add_ps(x[i], x[offset+i]); \ + } \ + offset >>= 1; \ + for (int i = 0; i < offset; ++i) { \ + x[i] = _mm_add_ps(x[i], x[offset+i]); \ + } \ + const __m128 t0 = _mm_hadd_ps(x[0], x[0]); \ + res = (ggml_float) _mm_cvtss_f32(_mm_hadd_ps(t0, t0)); \ +} +// TODO: is this optimal ? + +#define GGML_F32_VEC GGML_F32x4 +#define GGML_F32_VEC_ZERO GGML_F32x4_ZERO +#define GGML_F32_VEC_SET1 GGML_F32x4_SET1 +#define GGML_F32_VEC_LOAD GGML_F32x4_LOAD +#define GGML_F32_VEC_STORE GGML_F32x4_STORE +#define GGML_F32_VEC_FMA GGML_F32x4_FMA +#define GGML_F32_VEC_ADD GGML_F32x4_ADD +#define GGML_F32_VEC_MUL GGML_F32x4_MUL +#define GGML_F32_VEC_REDUCE GGML_F32x4_REDUCE + +// F16 SSE + +#define GGML_F16_STEP 32 +#define GGML_F16_EPR 4 + +static inline __m128 __sse_f16x4_load(const ggml_fp16_t * x) { + float tmp[4]; + + tmp[0] = GGML_FP16_TO_FP32(x[0]); + tmp[1] = GGML_FP16_TO_FP32(x[1]); + tmp[2] = GGML_FP16_TO_FP32(x[2]); + tmp[3] = GGML_FP16_TO_FP32(x[3]); + + return _mm_loadu_ps(tmp); +} + +static inline void __sse_f16x4_store(ggml_fp16_t * x, __m128 y) { + float arr[4]; + + _mm_storeu_ps(arr, y); + + x[0] = GGML_FP32_TO_FP16(arr[0]); + x[1] = GGML_FP32_TO_FP16(arr[1]); + x[2] = GGML_FP32_TO_FP16(arr[2]); + x[3] = GGML_FP32_TO_FP16(arr[3]); +} + +#define GGML_F32Cx4 __m128 +#define GGML_F32Cx4_ZERO _mm_setzero_ps() +#define GGML_F32Cx4_SET1(x) _mm_set1_ps(x) +#define GGML_F32Cx4_LOAD(x) __sse_f16x4_load(x) +#define GGML_F32Cx4_STORE(x, y) __sse_f16x4_store(x, y) +#define GGML_F32Cx4_FMA GGML_F32x4_FMA +#define GGML_F32Cx4_ADD _mm_add_ps +#define GGML_F32Cx4_MUL _mm_mul_ps +#define GGML_F32Cx4_REDUCE GGML_F32x4_REDUCE + +#define GGML_F16_VEC GGML_F32Cx4 +#define GGML_F16_VEC_ZERO GGML_F32Cx4_ZERO +#define GGML_F16_VEC_SET1 GGML_F32Cx4_SET1 +#define GGML_F16_VEC_LOAD(p, i) GGML_F32Cx4_LOAD(p) +#define GGML_F16_VEC_STORE(p, r, i) GGML_F32Cx4_STORE(p, r[i]) +#define GGML_F16_VEC_FMA GGML_F32Cx4_FMA +#define GGML_F16_VEC_ADD GGML_F32Cx4_ADD +#define GGML_F16_VEC_MUL GGML_F32Cx4_MUL +#define GGML_F16_VEC_REDUCE GGML_F32Cx4_REDUCE + +#elif defined(__loongarch_asx) + +#define GGML_SIMD + +// F32 LASX +#define GGML_F32_STEP 32 +#define GGML_F32_EPR 8 + +#define GGML_F32x8 __m256 +#define GGML_F32x8_ZERO (__m256)__lasx_xvldi(0) +#define GGML_F32x8_SET1(x) (__m256)__lasx_xvreplfr2vr_s((x)) +#define GGML_F32x8_LOAD(x) (__m256)__lasx_xvld((x), 0) +#define GGML_F32x8_STORE(x,y) __lasx_xvst((y), (x), 0) +#define GGML_F32x8_FMA(a, b, c) __lasx_xvfmadd_s(b, c, a) +#define GGML_F32x8_ADD __lasx_xvfadd_s +#define GGML_F32x8_MUL __lasx_xvfmul_s +#define GGML_F32x8_REDUCE(res, x) \ +do { \ + int offset = GGML_F32_ARR >> 1; \ + for (int i = 0; i < offset; ++i) { \ + x[i] = __lasx_xvfadd_s(x[i], x[offset+i]); \ + } \ + offset >>= 1; \ + for (int i = 0; i < offset; ++i) { \ + x[i] = __lasx_xvfadd_s(x[i], x[offset+i]); \ + } \ + offset >>= 1; \ + for (int i = 0; i < offset; ++i) { \ + x[i] = __lasx_xvfadd_s(x[i], x[offset+i]); \ + } \ + float *tmp_p = (float *)&x[0]; \ + res = tmp_p[0] + tmp_p[1] + tmp_p[2] + tmp_p[3] + tmp_p[4] + tmp_p[5] + tmp_p[6] + tmp_p[7]; \ +} while (0) +// TODO: is this optimal ? + +#define GGML_F32_VEC GGML_F32x8 +#define GGML_F32_VEC_ZERO GGML_F32x8_ZERO +#define GGML_F32_VEC_SET1 GGML_F32x8_SET1 +#define GGML_F32_VEC_LOAD GGML_F32x8_LOAD +#define GGML_F32_VEC_STORE GGML_F32x8_STORE +#define GGML_F32_VEC_FMA GGML_F32x8_FMA +#define GGML_F32_VEC_ADD GGML_F32x8_ADD +#define GGML_F32_VEC_MUL GGML_F32x8_MUL +#define GGML_F32_VEC_REDUCE GGML_F32x8_REDUCE + +// F16 LASX + +#define GGML_F16_STEP 32 +#define GGML_F16_EPR 8 + +// F16 arithmetic is not supported by LASX, so we use F32 instead + +#define GGML_F32Cx8 __m256 +#define GGML_F32Cx8_ZERO (__m256)__lasx_xvldi(0) +#define GGML_F32Cx8_SET1(x) (__m256)__lasx_xvreplgr2vr_w((x)) + +static inline __m256 __lasx_f32cx8_load(const ggml_fp16_t * x) { + __m256i a; + memcpy(&a, x, sizeof(ggml_fp16_t) * 8); + a = __lasx_xvpermi_d(a, 0 | (1 << 4)); + return __lasx_xvfcvtl_s_h(a); +} + +static inline void __lasx_f32cx8_store(ggml_fp16_t * x, __m256 y) { + __m256i a = __lasx_xvfcvt_h_s(y, y); + a = __lasx_xvpermi_d(a, 0 | (2 << 2)); + memcpy(x, &a, sizeof(ggml_fp16_t) * 8); +} +#define GGML_F32Cx8_LOAD(x) __lasx_f32cx8_load(x) +#define GGML_F32Cx8_STORE(x, y) __lasx_f32cx8_store(x, y) + +#define GGML_F32Cx8_FMA GGML_F32x8_FMA +#define GGML_F32Cx8_ADD __lasx_xvfadd_s +#define GGML_F32Cx8_MUL __lasx_xvfmul_s +#define GGML_F32Cx8_REDUCE GGML_F32x8_REDUCE + +#define GGML_F16_VEC GGML_F32Cx8 +#define GGML_F16_VEC_ZERO GGML_F32Cx8_ZERO +#define GGML_F16_VEC_SET1 GGML_F32Cx8_SET1 +#define GGML_F16_VEC_LOAD(p, i) GGML_F32Cx8_LOAD(p) +#define GGML_F16_VEC_STORE(p, r, i) GGML_F32Cx8_STORE(p, r[i]) +#define GGML_F16_VEC_FMA GGML_F32Cx8_FMA +#define GGML_F16_VEC_ADD GGML_F32Cx8_ADD +#define GGML_F16_VEC_MUL GGML_F32Cx8_MUL +#define GGML_F16_VEC_REDUCE GGML_F32Cx8_REDUCE + +#elif defined(__loongarch_sx) + +#define GGML_SIMD + +// F32 LSX + +#define GGML_F32_STEP 32 +#define GGML_F32_EPR 4 + +#define GGML_F32x4 __m128 +#define GGML_F32x4_ZERO __lsx_vldi(0) +#define GGML_F32x4_SET1(x) __lsx_vinsgr2vr_w(__lsx_vldi(0),(x), 0) +#define GGML_F32x4_LOAD(x) __lsx_vld((x), 0) +#define GGML_F32x4_STORE((x),(y)) __lsx_vst((y), (x), 0) +#define GGML_F32x4_FMA(a, b, c) __lsx_vfmadd_s(b, c, a) +#define GGML_F32x4_ADD __lsx_vfadd_s +#define GGML_F32x4_MUL __lsx_vfmul_s +#define GGML_F32x4_REDUCE(res, x) \ +{ \ + int offset = GGML_F32_ARR >> 1; \ + for (int i = 0; i < offset; ++i) { \ + x[i] = __lsx_vfadd_s(x[i], x[offset + i]); \ + } \ + offset >>= 1; \ + for (int i = 0; i < offset; ++i) { \ + x[i] = __lsx_vfadd_s(x[i], x[offset + i]); \ + } \ + offset >>= 1; \ + for (int i = 0; i < offset; ++i) { \ + x[i] = __lsx_vfadd_s(x[i], x[offset + i]); \ + } \ + __m128i tmp = __lsx_vsrli_d((__m128i) x[0], 32); \ + tmp = (__m128i) __lsx_vfadd_s((__m128) tmp, x[0]); \ + tmp = __lsx_vpickev_w(__lsx_vldi(0), tmp); \ + const __m128 t0 = __lsx_vshuf4i_w(tmp, 0x88); \ + tmp = __lsx_vsrli_d((__m128i) t0, 32); \ + tmp = (__m128i) __lsx_vfadd_s((__m128) tmp, t0); \ + tmp = __lsx_vpickev_w(__lsx_vldi(0), tmp); \ + res = (ggml_float) __lsx_vpickve2gr_w(__lsx_vshuf4i_w(tmp, 0x88), 0); \ +} + +#define GGML_F32_VEC GGML_F32x4 +#define GGML_F32_VEC_ZERO GGML_F32x4_ZERO +#define GGML_F32_VEC_SET1 GGML_F32x4_SET1 +#define GGML_F32_VEC_LOAD GGML_F32x4_LOAD +#define GGML_F32_VEC_STORE GGML_F32x4_STORE +#define GGML_F32_VEC_FMA GGML_F32x4_FMA +#define GGML_F32_VEC_ADD GGML_F32x4_ADD +#define GGML_F32_VEC_MUL GGML_F32x4_MUL +#define GGML_F32_VEC_REDUCE GGML_F32x4_REDUCE + +// F16 LSX + +#define GGML_F16_STEP 32 +#define GGML_F16_EPR 4 + +static inline __m128 __lsx_f16x4_load(const ggml_fp16_t * x) { + float tmp[4]; + + tmp[0] = GGML_FP16_TO_FP32(x[0]); + tmp[1] = GGML_FP16_TO_FP32(x[1]); + tmp[2] = GGML_FP16_TO_FP32(x[2]); + tmp[3] = GGML_FP16_TO_FP32(x[3]); + + return __lsx_vld(tmp, 0); +} + +static inline void __lsx_f16x4_store(ggml_fp16_t * x, __m128 y) { + float arr[4]; + + __lsx_vst(y, arr, 0); + + x[0] = GGML_FP32_TO_FP16(arr[0]); + x[1] = GGML_FP32_TO_FP16(arr[1]); + x[2] = GGML_FP32_TO_FP16(arr[2]); + x[3] = GGML_FP32_TO_FP16(arr[3]); +} + +#define GGML_F32Cx4 __m128 +#define GGML_F32Cx4_ZERO __lsx_vldi(0) +#define GGML_F32Cx4_SET1(x) __lsx_vinsgr2vr_w(__lsx_vldi(0),(x), 0) +#define GGML_F32Cx4_LOAD(x) __lsx_f16x4_load(x) +#define GGML_F32Cx4_STORE(x, y) __lsx_f16x4_store(x, y) +#define GGML_F32Cx4_FMA GGML_F32x4_FMA +#define GGML_F32Cx4_ADD __lsx_vfadd_s +#define GGML_F32Cx4_MUL __lsx_vfmul_s +#define GGML_F32Cx4_REDUCE GGML_F32x4_REDUCE + +#define GGML_F16_VEC GGML_F32Cx4 +#define GGML_F16_VEC_ZERO GGML_F32Cx4_ZERO +#define GGML_F16_VEC_SET1 GGML_F32Cx4_SET1 +#define GGML_F16_VEC_LOAD(p, i) GGML_F32Cx4_LOAD(p) +#define GGML_F16_VEC_STORE(p, r, i) GGML_F32Cx4_STORE(p, r[i]) +#define GGML_F16_VEC_FMA GGML_F32Cx4_FMA +#define GGML_F16_VEC_ADD GGML_F32Cx4_ADD +#define GGML_F16_VEC_MUL GGML_F32Cx4_MUL +#define GGML_F16_VEC_REDUCE GGML_F32Cx4_REDUCE + +#elif defined(__VXE__) || defined(__VXE2__) + +#define GGML_SIMD + +// F32 s390x + +#define GGML_F32_STEP 32 +#define GGML_F32_EPR 4 + +#define GGML_F32x4 __vector float +#define GGML_F32x4_ZERO vec_splats(0.0f) +#define GGML_F32x4_SET1 vec_splats +#define GGML_F32x4_LOAD(p) vec_xl(0, p) +#define GGML_F32x4_STORE(p, r) vec_xst(r, 0, p) +#define GGML_F32x4_FMA(a, b, c) vec_madd(b, c, a) +#define GGML_F32x4_ADD vec_add +#define GGML_F32x4_MUL vec_mul +#define GGML_F32x4_REDUCE(res, x) \ +{ \ + int offset = GGML_F32_ARR >> 1; \ + for (int i = 0; i < offset; ++i) { \ + x[i] = vec_add(x[i], x[offset + i]); \ + } \ + offset >>= 1; \ + for (int i = 0; i < offset; ++i) { \ + x[i] = vec_add(x[i], x[offset + i]); \ + } \ + offset >>= 1; \ + for (int i = 0; i < offset; ++i) { \ + x[i] = vec_add(x[i], x[offset + i]); \ + } \ + res = vec_extract(x[0], 0) + \ + vec_extract(x[0], 1) + \ + vec_extract(x[0], 2) + \ + vec_extract(x[0], 3); \ +} + +#define GGML_F32_VEC GGML_F32x4 +#define GGML_F32_VEC_ZERO GGML_F32x4_ZERO +#define GGML_F32_VEC_SET1 GGML_F32x4_SET1 +#define GGML_F32_VEC_LOAD GGML_F32x4_LOAD +#define GGML_F32_VEC_STORE GGML_F32x4_STORE +#define GGML_F32_VEC_FMA GGML_F32x4_FMA +#define GGML_F32_VEC_ADD GGML_F32x4_ADD +#define GGML_F32_VEC_MUL GGML_F32x4_MUL +#define GGML_F32_VEC_REDUCE GGML_F32x4_REDUCE + +// F16 s390x +#define GGML_F16_STEP GGML_F32_STEP +#define GGML_F16_EPR GGML_F32_EPR + +static inline __vector float __lzs_f16cx4_load(const ggml_fp16_t * x) { + float tmp[4]; + + for (int i = 0; i < 4; i++) { + tmp[i] = GGML_FP16_TO_FP32(x[i]); + } + + // note: keep type-cast here to prevent compiler bugs + // see: https://github.com/ggml-org/llama.cpp/issues/12846 + return vec_xl(0, (const float *)(tmp)); +} + +static inline void __lzs_f16cx4_store(ggml_fp16_t * x, __vector float y) { + float arr[4]; + + // note: keep type-cast here to prevent compiler bugs + // see: https://github.com/ggml-org/llama.cpp/issues/12846 + vec_xst(y, 0, (float *)(arr)); + + for (int i = 0; i < 4; i++) { + x[i] = GGML_FP32_TO_FP16(arr[i]); + } +} + +#define GGML_F16_VEC GGML_F32x4 +#define GGML_F16_VEC_ZERO GGML_F32x4_ZERO +#define GGML_F16_VEC_SET1 GGML_F32x4_SET1 +#define GGML_F16_VEC_LOAD(p, i) __lzs_f16cx4_load(p) +#define GGML_F16_VEC_STORE(p, r, i) __lzs_f16cx4_store(p, r[i]) +#define GGML_F16_VEC_FMA GGML_F32x4_FMA +#define GGML_F16_VEC_ADD GGML_F32x4_ADD +#define GGML_F16_VEC_MUL GGML_F32x4_MUL +#define GGML_F16_VEC_REDUCE GGML_F32x4_REDUCE + +#endif + +// GGML_F32_ARR / GGML_F16_ARR +// number of registers to use per step +#ifdef GGML_SIMD +#define GGML_F32_ARR (GGML_F32_STEP/GGML_F32_EPR) +#define GGML_F16_ARR (GGML_F16_STEP/GGML_F16_EPR) +#endif diff --git a/ggml/src/ggml-cpu/unary-ops.cpp b/ggml/src/ggml-cpu/unary-ops.cpp new file mode 100644 index 000000000..4fce569b3 --- /dev/null +++ b/ggml/src/ggml-cpu/unary-ops.cpp @@ -0,0 +1,186 @@ +#include "unary-ops.h" + +static inline float op_abs(float x) { + return fabsf(x); +} + +static inline float op_sgn(float x) { + return (x > 0.f) ? 1.f : ((x < 0.f) ? -1.f : 0.f); +} + +static inline float op_neg(float x) { + return -x; +} + +static inline float op_step(float x) { + return (x > 0.f) ? 1.f : 0.f; +} + +static inline float op_tanh(float x) { + return tanhf(x); +} + +static inline float op_elu(float x) { + return (x > 0.f) ? x : expm1f(x); +} + +static inline float op_relu(float x) { + return (x > 0.f) ? x : 0.f; +} + +static inline float op_sigmoid(float x) { + return 1.f / (1.f + expf(-x)); +} + +static inline float op_hardsigmoid(float x) { + return fminf(1.0f, fmaxf(0.0f, (x + 3.0f) / 6.0f)); +} + +static inline float op_exp(float x) { + return expf(x); +} + +static inline float op_hardswish(float x) { + return x * fminf(1.0f, fmaxf(0.0f, (x + 3.0f) / 6.0f)); +} + +static inline float op_sqr(float x) { + return x * x; +} + +static inline float op_sqrt(float x) { + return sqrtf(x); +} + +static inline float op_sin(float x) { + return sinf(x); +} + +static inline float op_cos(float x) { + return cosf(x); +} + +static inline float op_log(float x) { + return logf(x); +} + +template +static inline void vec_unary_op(int64_t n, dst_t * y, const src0_t * x) { + constexpr auto src0_to_f32 = type_conversion_table::to_f32; + constexpr auto f32_to_dst = type_conversion_table::from_f32; + + for (int i = 0; i < n; i++) { + y[i] = f32_to_dst(op(src0_to_f32(x[i]))); + } +} + +template +static void apply_unary_op(const ggml_compute_params * params, ggml_tensor * dst) { + const ggml_tensor * src0 = dst->src[0]; + + GGML_ASSERT(ggml_is_contiguous_1(src0) && ggml_is_contiguous_1(dst) && ggml_are_same_shape(src0, dst)); + + GGML_TENSOR_UNARY_OP_LOCALS + + GGML_ASSERT( nb0 == sizeof(dst_t)); + GGML_ASSERT(nb00 == sizeof(src0_t)); + + const auto [ir0, ir1] = get_thread_range(params, src0); + + for (int64_t ir = ir0; ir < ir1; ++ir) { + const int64_t i03 = ir/(ne02*ne01); + const int64_t i02 = (ir - i03*ne02*ne01)/ne01; + const int64_t i01 = (ir - i03*ne02*ne01 - i02*ne01); + + dst_t * dst_ptr = (dst_t *) ((char *) dst->data + i03*nb3 + i02*nb2 + i01*nb1 ); + const src0_t * src0_ptr = (const src0_t *) ((const char *) src0->data + i03*nb03 + i02*nb02 + i01*nb01); + + vec_unary_op(ne0, dst_ptr, src0_ptr); + } +} + +// TODO: Use the 'traits' lookup table (for type conversion fns), instead of a mass of 'if' conditions with long templates +template +static void unary_op(const ggml_compute_params * params, ggml_tensor * dst) { + const ggml_tensor * src0 = dst->src[0]; + + /* */ if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) { // all f32 + apply_unary_op(params, dst); + } else if (src0->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F16) { // all f16 + apply_unary_op(params, dst); + } else if (src0->type == GGML_TYPE_BF16 && dst->type == GGML_TYPE_BF16) { // all bf16 + apply_unary_op(params, dst); + } else if (src0->type == GGML_TYPE_BF16 && dst->type == GGML_TYPE_F32) { + apply_unary_op(params, dst); + } else if (src0->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F32) { + apply_unary_op(params, dst); + } else { + fprintf(stderr, "%s: unsupported types: dst: %s, src0: %s\n", __func__, + ggml_type_name(dst->type), ggml_type_name(src0->type)); + GGML_ABORT("fatal error"); + } +} + +void ggml_compute_forward_abs(const ggml_compute_params * params, ggml_tensor * dst) { + unary_op(params, dst); +} + +void ggml_compute_forward_sgn(const ggml_compute_params * params, ggml_tensor * dst) { + unary_op(params, dst); +} + +void ggml_compute_forward_neg(const ggml_compute_params * params, ggml_tensor * dst) { + unary_op(params, dst); +} + +void ggml_compute_forward_step(const ggml_compute_params * params, ggml_tensor * dst) { + unary_op(params, dst); +} + +void ggml_compute_forward_tanh(const ggml_compute_params * params, ggml_tensor * dst) { + unary_op(params, dst); +} + +void ggml_compute_forward_elu(const ggml_compute_params * params, ggml_tensor * dst) { + unary_op(params, dst); +} + +void ggml_compute_forward_relu(const ggml_compute_params * params, ggml_tensor * dst) { + unary_op(params, dst); +} + +void ggml_compute_forward_sigmoid(const ggml_compute_params * params, ggml_tensor * dst) { + unary_op(params, dst); +} + +void ggml_compute_forward_hardsigmoid(const ggml_compute_params * params, ggml_tensor * dst) { + unary_op(params, dst); +} + +void ggml_compute_forward_exp(const ggml_compute_params * params, ggml_tensor * dst) { + unary_op(params, dst); +} + +void ggml_compute_forward_hardswish(const ggml_compute_params * params, ggml_tensor * dst) { + unary_op(params, dst); +} + +void ggml_compute_forward_sqr(const ggml_compute_params * params, ggml_tensor * dst) { + unary_op(params, dst); +} + +void ggml_compute_forward_sqrt(const ggml_compute_params * params, ggml_tensor * dst) { + unary_op(params, dst); +} + +void ggml_compute_forward_sin(const ggml_compute_params * params, ggml_tensor * dst) { + unary_op(params, dst); +} + +void ggml_compute_forward_cos(const ggml_compute_params * params, ggml_tensor * dst) { + unary_op(params, dst); +} + +void ggml_compute_forward_log(const ggml_compute_params * params, ggml_tensor * dst) { + unary_op(params, dst); +} diff --git a/ggml/src/ggml-cpu/unary-ops.h b/ggml/src/ggml-cpu/unary-ops.h new file mode 100644 index 000000000..b1ade2c8e --- /dev/null +++ b/ggml/src/ggml-cpu/unary-ops.h @@ -0,0 +1,28 @@ +#pragma once + +#include "common.h" + +#ifdef __cplusplus +extern "C" { +#endif + +void ggml_compute_forward_abs(const struct ggml_compute_params * params, struct ggml_tensor * dst); +void ggml_compute_forward_sgn(const struct ggml_compute_params * params, struct ggml_tensor * dst); +void ggml_compute_forward_neg(const struct ggml_compute_params * params, struct ggml_tensor * dst); +void ggml_compute_forward_step(const struct ggml_compute_params * params, struct ggml_tensor * dst); +void ggml_compute_forward_tanh(const struct ggml_compute_params * params, struct ggml_tensor * dst); +void ggml_compute_forward_elu(const struct ggml_compute_params * params, struct ggml_tensor * dst); +void ggml_compute_forward_relu(const struct ggml_compute_params * params, struct ggml_tensor * dst); +void ggml_compute_forward_sigmoid(const struct ggml_compute_params * params, struct ggml_tensor * dst); +void ggml_compute_forward_hardsigmoid(const struct ggml_compute_params * params, struct ggml_tensor * dst); +void ggml_compute_forward_exp(const struct ggml_compute_params * params, struct ggml_tensor * dst); +void ggml_compute_forward_hardswish(const struct ggml_compute_params * params, struct ggml_tensor * dst); +void ggml_compute_forward_sqr(const struct ggml_compute_params * params, struct ggml_tensor * dst); +void ggml_compute_forward_sqrt(const struct ggml_compute_params * params, struct ggml_tensor * dst); +void ggml_compute_forward_sin(const struct ggml_compute_params * params, struct ggml_tensor * dst); +void ggml_compute_forward_cos(const struct ggml_compute_params * params, struct ggml_tensor * dst); +void ggml_compute_forward_log(const struct ggml_compute_params * params, struct ggml_tensor * dst); + +#ifdef __cplusplus +} +#endif diff --git a/ggml/src/ggml-cpu/vec.cpp b/ggml/src/ggml-cpu/vec.cpp new file mode 100644 index 000000000..dfe2218e3 --- /dev/null +++ b/ggml/src/ggml-cpu/vec.cpp @@ -0,0 +1,258 @@ +#include "vec.h" + +#include + +#if defined(_MSC_VER) +// disable "possible loss of data" to avoid hundreds of casts +// we should just be careful :) +#pragma warning(disable: 4244 4267) +#endif + +// precomputed gelu table for f16 (128 KB) +ggml_fp16_t ggml_table_gelu_f16[1 << 16]; + +// precomputed quick gelu table for f16 (128 KB) +ggml_fp16_t ggml_table_gelu_quick_f16[1 << 16]; + +void ggml_vec_dot_f32(int n, float * GGML_RESTRICT s, size_t bs, const float * GGML_RESTRICT x, size_t bx, const float * GGML_RESTRICT y, size_t by, int nrc) { + assert(nrc == 1); + GGML_UNUSED(nrc); + GGML_UNUSED(bx); + GGML_UNUSED(by); + GGML_UNUSED(bs); + +#if defined(GGML_SIMD) + float sumf = 0.0f; + const int np = (n & ~(GGML_F32_STEP - 1)); + + GGML_F32_VEC sum[GGML_F32_ARR] = { GGML_F32_VEC_ZERO }; + + GGML_F32_VEC ax[GGML_F32_ARR]; + GGML_F32_VEC ay[GGML_F32_ARR]; + + for (int i = 0; i < np; i += GGML_F32_STEP) { + for (int j = 0; j < GGML_F32_ARR; j++) { + ax[j] = GGML_F32_VEC_LOAD(x + i + j*GGML_F32_EPR); + ay[j] = GGML_F32_VEC_LOAD(y + i + j*GGML_F32_EPR); + + sum[j] = GGML_F32_VEC_FMA(sum[j], ax[j], ay[j]); + } + } + + // reduce sum0..sum3 to sum0 + GGML_F32_VEC_REDUCE(sumf, sum); + + // leftovers + for (int i = np; i < n; ++i) { + sumf += x[i]*y[i]; + } +#else + // scalar + ggml_float sumf = 0.0; + for (int i = 0; i < n; ++i) { + sumf += (ggml_float)(x[i]*y[i]); + } +#endif + + *s = sumf; +} + +void ggml_vec_dot_bf16(int n, float * GGML_RESTRICT s, size_t bs, ggml_bf16_t * GGML_RESTRICT x, size_t bx, ggml_bf16_t * GGML_RESTRICT y, size_t by, int nrc) { + assert(nrc == 1); + GGML_UNUSED(nrc); + GGML_UNUSED(bx); + GGML_UNUSED(by); + GGML_UNUSED(bs); + int i = 0; + ggml_float sumf = 0; + +#if defined(__AVX512BF16__) + __m512 c1 = _mm512_setzero_ps(); + __m512 c2 = _mm512_setzero_ps(); + for (; i + 64 <= n; i += 64) { + c1 = _mm512_dpbf16_ps(c1, m512bh(_mm512_loadu_si512((x + i))), + m512bh(_mm512_loadu_si512((y + i)))); + c2 = _mm512_dpbf16_ps(c2, m512bh(_mm512_loadu_si512((x + i + 32))), + m512bh(_mm512_loadu_si512((y + i + 32)))); + } + sumf += (ggml_float)_mm512_reduce_add_ps(c1); + sumf += (ggml_float)_mm512_reduce_add_ps(c2); + +#elif defined(__AVX512F__) +#define LOAD(p) _mm512_castsi512_ps(_mm512_slli_epi32(_mm512_cvtepu16_epi32(_mm256_loadu_si256((const __m256i *)(p))), 16)) + __m512 c1 = _mm512_setzero_ps(); + __m512 c2 = _mm512_setzero_ps(); + for (; i + 32 <= n; i += 32) { + c1 = _mm512_add_ps(_mm512_mul_ps(LOAD(x + i), LOAD(y + i)), c1); + c2 = _mm512_add_ps(_mm512_mul_ps(LOAD(x + i + 16), LOAD(y + i + 16)), c2); + } + sumf += (ggml_float)_mm512_reduce_add_ps(c1); + sumf += (ggml_float)_mm512_reduce_add_ps(c2); + +#undef LOAD +#elif defined(__AVX2__) || defined(__AVX__) +#if defined(__AVX2__) +#define LOAD(p) _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_cvtepu16_epi32(_mm_loadu_si128((const __m128i *)(p))), 16)) +#else +#define LOAD(p) _mm256_castsi256_ps(_mm256_insertf128_si256(_mm256_castsi128_si256(_mm_slli_epi32(_mm_cvtepu16_epi32(_mm_loadu_si128((const __m128i *)(p))), 16)), (_mm_slli_epi32(_mm_cvtepu16_epi32(_mm_bsrli_si128(_mm_loadu_si128((const __m128i *)(p)), 8)), 16)), 1)) +#endif + __m256 c1 = _mm256_setzero_ps(); + __m256 c2 = _mm256_setzero_ps(); + __m256 c3 = _mm256_setzero_ps(); + __m256 c4 = _mm256_setzero_ps(); + for (; i + 32 <= n; i += 32) { + c1 = _mm256_add_ps(_mm256_mul_ps(LOAD(x + i), LOAD(y + i)), c1); + c2 = _mm256_add_ps(_mm256_mul_ps(LOAD(x + i + 8), LOAD(y + i + 8)), c2); + c3 = _mm256_add_ps(_mm256_mul_ps(LOAD(x + i + 16), LOAD(y + i + 16)), c3); + c4 = _mm256_add_ps(_mm256_mul_ps(LOAD(x + i + 24), LOAD(y + i + 24)), c4); + } + __m128 g; + c1 = _mm256_add_ps(_mm256_add_ps(c1, c3), + _mm256_add_ps(c2, c4)); + g = _mm_add_ps(_mm256_extractf128_ps(c1, 1), + _mm256_castps256_ps128(c1)); + g = _mm_add_ps(g, _mm_movehl_ps(g, g)); + g = _mm_add_ss(g, _mm_movehdup_ps(g)); + sumf += (ggml_float)_mm_cvtss_f32(g); + +#undef LOAD +#endif + + for (; i < n; ++i) { + sumf += (ggml_float)(GGML_BF16_TO_FP32(x[i]) * + GGML_BF16_TO_FP32(y[i])); + } + *s = sumf; +} + +void ggml_vec_dot_f16(int n, float * GGML_RESTRICT s, size_t bs, ggml_fp16_t * GGML_RESTRICT x, size_t bx, ggml_fp16_t * GGML_RESTRICT y, size_t by, int nrc) { + assert(nrc == 1); + GGML_UNUSED(nrc); + GGML_UNUSED(bx); + GGML_UNUSED(by); + GGML_UNUSED(bs); + + ggml_float sumf = 0.0; + +#if defined(GGML_SIMD) + const int np = (n & ~(GGML_F16_STEP - 1)); + + GGML_F16_VEC sum[GGML_F16_ARR] = { GGML_F16_VEC_ZERO }; + + GGML_F16_VEC ax[GGML_F16_ARR]; + GGML_F16_VEC ay[GGML_F16_ARR]; + + for (int i = 0; i < np; i += GGML_F16_STEP) { + for (int j = 0; j < GGML_F16_ARR; j++) { + ax[j] = GGML_F16_VEC_LOAD(x + i + j*GGML_F16_EPR, j); + ay[j] = GGML_F16_VEC_LOAD(y + i + j*GGML_F16_EPR, j); + + sum[j] = GGML_F16_VEC_FMA(sum[j], ax[j], ay[j]); + } + } + + // reduce sum0..sum3 to sum0 + GGML_F16_VEC_REDUCE(sumf, sum); + + // leftovers + for (int i = np; i < n; ++i) { + sumf += (ggml_float)(GGML_FP16_TO_FP32(x[i])*GGML_FP16_TO_FP32(y[i])); + } +#else + for (int i = 0; i < n; ++i) { + sumf += (ggml_float)(GGML_FP16_TO_FP32(x[i])*GGML_FP16_TO_FP32(y[i])); + } +#endif + + *s = sumf; +} + +void ggml_vec_silu_f32(const int n, float * y, const float * x) { + int i = 0; +#if defined(__AVX512F__) && defined(__AVX512DQ__) + for (; i + 15 < n; i += 16) { + _mm512_storeu_ps(y + i, ggml_v_silu(_mm512_loadu_ps(x + i))); + } +#elif defined(__AVX2__) && defined(__FMA__) + for (; i + 7 < n; i += 8) { + _mm256_storeu_ps(y + i, ggml_v_silu(_mm256_loadu_ps(x + i))); + } +#elif defined(__SSE2__) + for (; i + 3 < n; i += 4) { + _mm_storeu_ps(y + i, ggml_v_silu(_mm_loadu_ps(x + i))); + } +#elif defined(__ARM_NEON) && defined(__aarch64__) + for (; i + 3 < n; i += 4) { + vst1q_f32(y + i, ggml_v_silu(vld1q_f32(x + i))); + } +#endif + for (; i < n; ++i) { + y[i] = ggml_silu_f32(x[i]); + } +} + +ggml_float ggml_vec_soft_max_f32(const int n, float * y, const float * x, float max) { + int i = 0; + ggml_float sum = 0; +#if defined(__AVX512F__) && defined(__AVX512DQ__) + for (; i + 15 < n; i += 16) { + __m512 val = ggml_v_expf(_mm512_sub_ps(_mm512_loadu_ps(x + i), + _mm512_set1_ps(max))); + _mm512_storeu_ps(y + i, val); + sum += (ggml_float)_mm512_reduce_add_ps(val); + } +#elif defined(__AVX2__) && defined(__FMA__) + for (; i + 7 < n; i += 8) { + __m256 val = ggml_v_expf(_mm256_sub_ps(_mm256_loadu_ps(x + i), + _mm256_set1_ps(max))); + _mm256_storeu_ps(y + i, val); + __m128 val2 = _mm_add_ps(_mm256_extractf128_ps(val, 1), + _mm256_castps256_ps128(val)); + val2 = _mm_add_ps(val2, _mm_movehl_ps(val2, val2)); + val2 = _mm_add_ss(val2, _mm_movehdup_ps(val2)); + sum += (ggml_float)_mm_cvtss_f32(val2); + } +#elif defined(__SSE2__) + for (; i + 3 < n; i += 4) { + __m128 val = ggml_v_expf(_mm_sub_ps(_mm_loadu_ps(x + i), + _mm_set1_ps(max))); + _mm_storeu_ps(y + i, val); +#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__) + val = _mm_add_ps(val, _mm_movehl_ps(val, val)); + val = _mm_add_ss(val, _mm_movehdup_ps(val)); +#else + __m128 tmp = _mm_shuffle_ps(val, val, _MM_SHUFFLE(2, 3, 0, 1)); + val = _mm_add_ps(val, tmp); + tmp = _mm_movehl_ps(tmp, val); + val = _mm_add_ss(val, tmp); +#endif + sum += (ggml_float)_mm_cvtss_f32(val); + } +#elif defined(__ARM_NEON) && defined(__aarch64__) + for (; i + 3 < n; i += 4) { + float32x4_t val = ggml_v_expf(vsubq_f32(vld1q_f32(x + i), + vdupq_n_f32(max))); + vst1q_f32(y + i, val); + sum += (ggml_float)vaddvq_f32(val); + } +#endif + for (; i < n; ++i) { + float val = expf(x[i] - max); + sum += (ggml_float)val; + y[i] = val; + } + return sum; +} + +ggml_float ggml_vec_log_soft_max_f32(const int n, float * y, const float * x, float max) { + // log(soft_max) = log(soft_max_i / soft_max_sum) = log(soft_max_i) - log(soft_max_sum) = (logit_i - max) - log(soft_max_i) + + int i = 0; + ggml_float sum = 0; + for (; i < n; ++i) { + float val = x[i] - max; + y[i] = val; + sum += (ggml_float)expf(val); + } + return sum = (ggml_float)logf(sum); +} diff --git a/ggml/src/ggml-cpu/vec.h b/ggml/src/ggml-cpu/vec.h new file mode 100644 index 000000000..23cbb3051 --- /dev/null +++ b/ggml/src/ggml-cpu/vec.h @@ -0,0 +1,802 @@ +// Vectorized functions for fundamental operations + +#pragma once + +#include "ggml-impl.h" +#include "simd-mappings.h" +#include "ggml.h" + +#if defined(GGML_USE_ACCELERATE) +#include +#endif + +// floating point type used to accumulate sums +typedef double ggml_float; + +#define GGML_GELU_FP16 +#define GGML_GELU_QUICK_FP16 + +#define GGML_SOFT_MAX_UNROLL 4 +#define GGML_VEC_DOT_UNROLL 2 +#define GGML_VEC_MAD_UNROLL 32 + +#ifdef __cplusplus +extern "C" { +#endif + +// +// global data +// + +// precomputed gelu table for f16 (128 KB) +extern ggml_fp16_t ggml_table_gelu_f16[1 << 16]; + +// precomputed quick gelu table for f16 (128 KB) +extern ggml_fp16_t ggml_table_gelu_quick_f16[1 << 16]; + +// +// fundamental operations +// + +void ggml_vec_dot_f32(int n, float * GGML_RESTRICT s, size_t bs, const float * GGML_RESTRICT x, size_t bx, const float * GGML_RESTRICT y, size_t by, int nrc); +void ggml_vec_dot_bf16(int n, float * GGML_RESTRICT s, size_t bs, ggml_bf16_t * GGML_RESTRICT x, size_t bx, ggml_bf16_t * GGML_RESTRICT y, size_t by, int nrc); +void ggml_vec_dot_f16(int n, float * GGML_RESTRICT s, size_t bs, ggml_fp16_t * GGML_RESTRICT x, size_t bx, ggml_fp16_t * GGML_RESTRICT y, size_t by, int nrc); + +void ggml_vec_silu_f32(const int n, float * y, const float * x); +ggml_float ggml_vec_soft_max_f32(const int n, float * y, const float * x, float max); +ggml_float ggml_vec_log_soft_max_f32(const int n, float * y, const float * x, float max); + +inline static void ggml_vec_set_i8(const int n, int8_t * x, const int8_t v) { for (int i = 0; i < n; ++i) x[i] = v; } +inline static void ggml_vec_set_i16(const int n, int16_t * x, const int16_t v) { for (int i = 0; i < n; ++i) x[i] = v; } + +inline static void ggml_vec_set_i32(const int n, int32_t * x, const int32_t v) { for (int i = 0; i < n; ++i) x[i] = v; } +inline static void ggml_vec_cpy_i32(const int n, int32_t * y, const int32_t * x) { for (int i = 0; i < n; ++i) y[i] = x[i]; } + +inline static void ggml_vec_set_f16(const int n, ggml_fp16_t * x, const ggml_fp16_t v) { for (int i = 0; i < n; ++i) x[i] = v; } +inline static void ggml_vec_set_bf16(const int n, ggml_bf16_t * x, const ggml_bf16_t v) { for (int i = 0; i < n; ++i) x[i] = v; } +inline static void ggml_vec_add_f32 (const int n, float * z, const float * x, const float * y) { for (int i = 0; i < n; ++i) z[i] = x[i] + y[i]; } +inline static void ggml_vec_add_f16 (const int n, ggml_fp16_t * z, const ggml_fp16_t * x, const ggml_fp16_t * y) { + for (int i = 0; i < n; ++i) { + z[i] = GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(x[i]) + GGML_FP16_TO_FP32(y[i])); + } +} +inline static void ggml_vec_add1_f32(const int n, float * z, const float * x, const float v) { for (int i = 0; i < n; ++i) z[i] = x[i] + v; } +inline static void ggml_vec_acc_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] += x[i]; } +inline static void ggml_vec_acc1_f32(const int n, float * y, const float v) { for (int i = 0; i < n; ++i) y[i] += v; } +inline static void ggml_vec_sub_f32 (const int n, float * z, const float * x, const float * y) { for (int i = 0; i < n; ++i) z[i] = x[i] - y[i]; } +inline static void ggml_vec_sub_f16 (const int n, ggml_fp16_t * z, const ggml_fp16_t * x, const ggml_fp16_t * y) { + for (int i = 0; i < n; ++i) { + z[i] = GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(x[i]) - GGML_FP16_TO_FP32(y[i])); + } +} +inline static void ggml_vec_set_f32 (const int n, float * x, const float v) { for (int i = 0; i < n; ++i) x[i] = v; } +inline static void ggml_vec_cpy_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = x[i]; } +inline static void ggml_vec_neg_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = -x[i]; } +inline static void ggml_vec_neg_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) { + for (int i = 0; i < n; ++i) { + y[i] = GGML_FP32_TO_FP16(-GGML_FP16_TO_FP32(x[i])); + } +} + +inline static void ggml_vec_mul_f32 (const int n, float * z, const float * x, const float * y) { for (int i = 0; i < n; ++i) z[i] = x[i]*y[i]; } +inline static void ggml_vec_mul_f16 (const int n, ggml_fp16_t * z, const ggml_fp16_t * x, const ggml_fp16_t * y) { + for (int i = 0; i < n; ++i) { + z[i] = GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(x[i]) * GGML_FP16_TO_FP32(y[i])); + } +} +inline static void ggml_vec_div_f32 (const int n, float * z, const float * x, const float * y) { for (int i = 0; i < n; ++i) z[i] = x[i]/y[i]; } +inline static void ggml_vec_div_f16 (const int n, ggml_fp16_t * z, const ggml_fp16_t * x, const ggml_fp16_t * y) { + for (int i = 0; i < n; ++i) { + z[i] = GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(x[i]) / GGML_FP16_TO_FP32(y[i])); + } +} + +// compute GGML_VEC_DOT_UNROLL dot products at once +// xs - x row stride in bytes +inline static void ggml_vec_dot_f16_unroll(const int n, const int xs, float * GGML_RESTRICT s, void * GGML_RESTRICT xv, ggml_fp16_t * GGML_RESTRICT y) { + ggml_float sumf[GGML_VEC_DOT_UNROLL] = { 0.0 }; + + ggml_fp16_t * GGML_RESTRICT x[GGML_VEC_DOT_UNROLL]; + + for (int i = 0; i < GGML_VEC_DOT_UNROLL; ++i) { + x[i] = (ggml_fp16_t *) ((char *) xv + i*xs); + } + +#if defined(GGML_SIMD) + const int np = (n & ~(GGML_F16_STEP - 1)); + + GGML_F16_VEC sum[GGML_VEC_DOT_UNROLL][GGML_F16_ARR] = { { GGML_F16_VEC_ZERO } }; + + GGML_F16_VEC ax[GGML_F16_ARR]; + GGML_F16_VEC ay[GGML_F16_ARR]; + + for (int i = 0; i < np; i += GGML_F16_STEP) { + for (int j = 0; j < GGML_F16_ARR; j++) { + ay[j] = GGML_F16_VEC_LOAD(y + i + j*GGML_F16_EPR, j); + + for (int k = 0; k < GGML_VEC_DOT_UNROLL; ++k) { + ax[j] = GGML_F16_VEC_LOAD(x[k] + i + j*GGML_F16_EPR, j); + + sum[k][j] = GGML_F16_VEC_FMA(sum[k][j], ax[j], ay[j]); + } + } + } + + // reduce sum0..sum3 to sum0 + for (int k = 0; k < GGML_VEC_DOT_UNROLL; ++k) { + GGML_F16_VEC_REDUCE(sumf[k], sum[k]); + } + + // leftovers + for (int i = np; i < n; ++i) { + for (int j = 0; j < GGML_VEC_DOT_UNROLL; ++j) { + sumf[j] += (ggml_float)(GGML_FP16_TO_FP32(x[j][i])*GGML_FP16_TO_FP32(y[i])); + } + } +#else + for (int i = 0; i < n; ++i) { + for (int j = 0; j < GGML_VEC_DOT_UNROLL; ++j) { + sumf[j] += (ggml_float)(GGML_FP16_TO_FP32(x[j][i])*GGML_FP16_TO_FP32(y[i])); + } + } +#endif + + for (int i = 0; i < GGML_VEC_DOT_UNROLL; ++i) { + s[i] = (float)sumf[i]; + } +} + +inline static void ggml_vec_mad_f32(const int n, float * GGML_RESTRICT y, const float * GGML_RESTRICT x, const float v) { +#if defined(GGML_SIMD) + const int np = (n & ~(GGML_F32_STEP - 1)); + + GGML_F32_VEC vx = GGML_F32_VEC_SET1(v); + + GGML_F32_VEC ax[GGML_F32_ARR]; + GGML_F32_VEC ay[GGML_F32_ARR]; + + for (int i = 0; i < np; i += GGML_F32_STEP) { + for (int j = 0; j < GGML_F32_ARR; j++) { + ax[j] = GGML_F32_VEC_LOAD(x + i + j*GGML_F32_EPR); + ay[j] = GGML_F32_VEC_LOAD(y + i + j*GGML_F32_EPR); + ay[j] = GGML_F32_VEC_FMA(ay[j], ax[j], vx); + + GGML_F32_VEC_STORE(y + i + j*GGML_F32_EPR, ay[j]); + } + } + + // leftovers + for (int i = np; i < n; ++i) { + y[i] += x[i]*v; + } +#else + // scalar + for (int i = 0; i < n; ++i) { + y[i] += x[i]*v; + } +#endif +} + +inline static void ggml_vec_mad_f16(const int n, ggml_fp16_t * GGML_RESTRICT y, const ggml_fp16_t * GGML_RESTRICT x, const float v) { +#if defined(GGML_SIMD) + const int np = (n & ~(GGML_F16_STEP - 1)); + + GGML_F16_VEC vx = GGML_F16_VEC_SET1(v); + + GGML_F16_VEC ax[GGML_F16_ARR]; + GGML_F16_VEC ay[GGML_F16_ARR]; + + for (int i = 0; i < np; i += GGML_F16_STEP) { + for (int j = 0; j < GGML_F16_ARR; j++) { + ax[j] = GGML_F16_VEC_LOAD(x + i + j*GGML_F16_EPR, j); + ay[j] = GGML_F16_VEC_LOAD(y + i + j*GGML_F16_EPR, j); + ay[j] = GGML_F16_VEC_FMA(ay[j], ax[j], vx); + + GGML_F16_VEC_STORE(y + i + j*GGML_F16_EPR, ay, j); + } + } + + // leftovers + for (int i = np; i < n; ++i) { + y[i] = GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(y[i]) + GGML_FP16_TO_FP32(x[i])*v); + } +#else + // scalar + for (int i = 0; i < n; ++i) { + y[i] = GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(y[i]) + GGML_FP16_TO_FP32(x[i])*v); + } +#endif +} + +// xs and vs are byte strides of x and v +inline static void ggml_vec_mad_f32_unroll(const int n, const int xs, const int vs, float * GGML_RESTRICT y, const float * GGML_RESTRICT xv, const float * GGML_RESTRICT vv) { + + const float * GGML_RESTRICT x[GGML_VEC_MAD_UNROLL]; + const float * GGML_RESTRICT v[GGML_VEC_MAD_UNROLL]; + + for (int i = 0; i < GGML_VEC_MAD_UNROLL; ++i) { + x[i] = (const float *) ((const char *) xv + i*xs); + v[i] = (const float *) ((const char *) vv + i*vs); + } + +#if defined(GGML_SIMD) + const int np = (n & ~(GGML_F32_STEP - 1)); + + GGML_F32_VEC vx[GGML_VEC_MAD_UNROLL]; + + for (int k = 0; k < GGML_VEC_MAD_UNROLL; ++k) { + vx[k] = GGML_F32_VEC_SET1(v[k][0]); + } + + GGML_F32_VEC ax[GGML_VEC_MAD_UNROLL][GGML_F32_ARR]; + GGML_F32_VEC ay[GGML_F32_ARR]; + + for (int i = 0; i < np; i += GGML_F32_STEP) { + for (int j = 0; j < GGML_F32_ARR; j++) { + ay[j] = GGML_F32_VEC_LOAD(y + i + j*GGML_F32_EPR); + + for (int k = 0; k < GGML_VEC_MAD_UNROLL; ++k) { + ax[k][j] = GGML_F32_VEC_LOAD(x[k] + i + j*GGML_F32_EPR); + ay[j] = GGML_F32_VEC_FMA(ay[j], ax[k][j], vx[k]); + } + + GGML_F32_VEC_STORE(y + i + j*GGML_F32_EPR, ay[j]); + } + } + + // leftovers + for (int k = 0; k < GGML_VEC_MAD_UNROLL; ++k) { + for (int i = np; i < n; ++i) { + y[i] += x[k][i]*v[k][0]; + } + } +#else + // scalar + for (int k = 0; k < GGML_VEC_MAD_UNROLL; ++k) { + for (int i = 0; i < n; ++i) { + y[i] += x[k][i]*v[k][0]; + } + } +#endif +} + +//inline static void ggml_vec_scale_f32(const int n, float * y, const float v) { for (int i = 0; i < n; ++i) y[i] *= v; } +inline static void ggml_vec_scale_f32(const int n, float * y, const float v) { +#if defined(GGML_USE_ACCELERATE) + vDSP_vsmul(y, 1, &v, y, 1, n); +#elif defined(GGML_SIMD) + const int np = (n & ~(GGML_F32_STEP - 1)); + + GGML_F32_VEC vx = GGML_F32_VEC_SET1(v); + + GGML_F32_VEC ay[GGML_F32_ARR]; + + for (int i = 0; i < np; i += GGML_F32_STEP) { + for (int j = 0; j < GGML_F32_ARR; j++) { + ay[j] = GGML_F32_VEC_LOAD(y + i + j*GGML_F32_EPR); + ay[j] = GGML_F32_VEC_MUL(ay[j], vx); + + GGML_F32_VEC_STORE(y + i + j*GGML_F32_EPR, ay[j]); + } + } + + // leftovers + for (int i = np; i < n; ++i) { + y[i] *= v; + } +#else + // scalar + for (int i = 0; i < n; ++i) { + y[i] *= v; + } +#endif +} + +inline static void ggml_vec_scale_f16(const int n, ggml_fp16_t * y, const float v) { +#if defined(GGML_SIMD) + const int np = (n & ~(GGML_F16_STEP - 1)); + + GGML_F16_VEC vx = GGML_F16_VEC_SET1(v); + + GGML_F16_VEC ay[GGML_F16_ARR]; + + for (int i = 0; i < np; i += GGML_F16_STEP) { + for (int j = 0; j < GGML_F16_ARR; j++) { + ay[j] = GGML_F16_VEC_LOAD(y + i + j*GGML_F16_EPR, j); + ay[j] = GGML_F16_VEC_MUL(ay[j], vx); + + GGML_F16_VEC_STORE(y + i + j*GGML_F16_EPR, ay, j); + } + } + + // leftovers + for (int i = np; i < n; ++i) { + y[i] = GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(y[i])*v); + } +#else + // scalar + for (int i = 0; i < n; ++i) { + y[i] = GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(y[i])*v); + } +#endif +} + +inline static void ggml_vec_norm_f32 (const int n, float * s, const float * x) { ggml_vec_dot_f32(n, s, 0, x, 0, x, 0, 1); *s = sqrtf(*s); } +inline static void ggml_vec_sqr_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = x[i]*x[i]; } +inline static void ggml_vec_sqr_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) { + for (int i = 0; i < n; ++i) { + float v = GGML_FP16_TO_FP32(x[i]); + y[i] = GGML_FP32_TO_FP16(v*v); + } +} +inline static void ggml_vec_sqrt_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = sqrtf(x[i]); } +inline static void ggml_vec_sqrt_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) { + for (int i = 0; i < n; ++i) { + y[i] = GGML_FP32_TO_FP16(sqrtf(GGML_FP16_TO_FP32(x[i]))); + } +} +inline static void ggml_vec_log_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = logf(x[i]); } +inline static void ggml_vec_log_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) { + for (int i = 0; i < n; ++i) { + y[i] = GGML_FP32_TO_FP16(logf(GGML_FP16_TO_FP32(x[i]))); + } +} +inline static void ggml_vec_sin_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = sinf(x[i]); } +inline static void ggml_vec_sin_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) { + for (int i = 0; i < n; ++i) { + y[i] = GGML_FP32_TO_FP16(sinf(GGML_FP16_TO_FP32(x[i]))); + } +} +inline static void ggml_vec_cos_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = cosf(x[i]); } +inline static void ggml_vec_cos_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) { + for (int i = 0; i < n; ++i) { + y[i] = GGML_FP32_TO_FP16(cosf(GGML_FP16_TO_FP32(x[i]))); + } +} +inline static void ggml_vec_abs_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = fabsf(x[i]); } +inline static void ggml_vec_abs_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) { + for (int i = 0; i < n; ++i) { + y[i] = GGML_FP32_TO_FP16(fabsf(GGML_FP16_TO_FP32(x[i]))); + } +} +inline static void ggml_vec_sgn_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? 1.f : ((x[i] < 0.f) ? -1.f : 0.f); } +inline static void ggml_vec_sgn_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) { + for (int i = 0; i < n; ++i) { + float v = GGML_FP16_TO_FP32(x[i]); + y[i] = GGML_FP32_TO_FP16((v > 0.f) ? 1.f : ((v < 0.f) ? -1.f : 0.f)); + } +} +inline static void ggml_vec_step_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? 1.f : 0.f; } +inline static void ggml_vec_step_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) { + for (int i = 0; i < n; ++i) { + y[i] = GGML_FP32_TO_FP16((GGML_FP16_TO_FP32(x[i]) > 0.f) ? 1.f : 0.f); + } +} +inline static void ggml_vec_tanh_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = tanhf(x[i]); } +inline static void ggml_vec_tanh_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) { + for (int i = 0; i < n; ++i) { + y[i] = GGML_FP32_TO_FP16(tanhf(GGML_FP16_TO_FP32(x[i]))); + } +} +inline static void ggml_vec_elu_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? x[i] : expm1f(x[i]); } +inline static void ggml_vec_elu_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) { + for (int i = 0; i < n; ++i) { + y[i] = GGML_FP32_TO_FP16(expm1f(GGML_FP16_TO_FP32(x[i]))); + } +} +inline static void ggml_vec_relu_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? x[i] : 0.f; } +inline static void ggml_vec_relu_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) { + for (int i = 0; i < n; ++i) { + float v = GGML_FP16_TO_FP32(x[i]); + y[i] = GGML_FP32_TO_FP16((v > 0.f) ? v : 0.f); + } +} +inline static void ggml_vec_leaky_relu_f32 (const int n, float * y, const float * x, const float ns) { for (int i = 0; i < n; ++i) y[i] = ((x[i] > 0.f) ? x[i] : 0.f) + ns * ((x[i] < 0.0f) ? x[i] : 0.f); } +inline static void ggml_vec_leaky_relu_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x, const float ns) { + for (int i = 0; i < n; ++i) { + float v = GGML_FP16_TO_FP32(x[i]); + y[i] = GGML_FP32_TO_FP16(((v > 0.f) ? v : 0.f) + ns * ((v < 0.0f) ? v : 0.f)); + } +} +inline static void ggml_vec_sigmoid_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = 1.f / (1.f + expf(-x[i])); } +inline static void ggml_vec_sigmoid_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) { + for (int i = 0; i < n; ++i) { + y[i] = GGML_FP32_TO_FP16(1.f / (1.f + expf(-GGML_FP16_TO_FP32(x[i])))); + } +} +// TODO: optimize performance +inline static void ggml_vec_hardswish_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = x[i] * fminf(1.0f, fmaxf(0.0f, (x[i] + 3.0f) / 6.0f)); } +inline static void ggml_vec_hardswish_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) { + for (int i = 0; i < n; ++i) { + float v = GGML_FP16_TO_FP32(x[i]); + y[i] = GGML_FP32_TO_FP16(v * fminf(1.0f, fmaxf(0.0f, (v + 3.0f) / 6.0f))); + } +} +inline static void ggml_vec_hardsigmoid_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = fminf(1.0f, fmaxf(0.0f, (x[i] + 3.0f) / 6.0f)); } +inline static void ggml_vec_hardsigmoid_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) { + for (int i = 0; i < n; ++i) { + y[i] = GGML_FP32_TO_FP16(fminf(1.0f, fmaxf(0.0f, (GGML_FP16_TO_FP32(x[i]) + 3.0f) / 6.0f))); + } +} +inline static void ggml_vec_exp_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = expf(x[i]); } +inline static void ggml_vec_exp_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) { + for (int i = 0; i < n; ++i) { + y[i] = GGML_FP32_TO_FP16(expf(GGML_FP16_TO_FP32(x[i]))); + } +} + +static const float GELU_COEF_A = 0.044715f; +static const float GELU_QUICK_COEF = -1.702f; +static const float SQRT_2_OVER_PI = 0.79788456080286535587989211986876f; + +inline static float ggml_gelu_f32(float x) { + return 0.5f*x*(1.0f + tanhf(SQRT_2_OVER_PI*x*(1.0f + GELU_COEF_A*x*x))); +} + +inline static void ggml_vec_gelu_f16(const int n, ggml_fp16_t * y, const ggml_fp16_t * x) { + const uint16_t * i16 = (const uint16_t *) x; + for (int i = 0; i < n; ++i) { + y[i] = ggml_table_gelu_f16[i16[i]]; + } +} + +#ifdef GGML_GELU_FP16 +inline static void ggml_vec_gelu_f32(const int n, float * y, const float * x) { + uint16_t t; + for (int i = 0; i < n; ++i) { + if (x[i] <= -10.0f) { + y[i] = 0.0f; + } else if (x[i] >= 10.0f) { + y[i] = x[i]; + } else { + ggml_fp16_t fp16 = GGML_FP32_TO_FP16(x[i]); + memcpy(&t, &fp16, sizeof(uint16_t)); + y[i] = GGML_FP16_TO_FP32(ggml_table_gelu_f16[t]); + } + } +} +#else +inline static void ggml_vec_gelu_f32(const int n, float * y, const float * x) { + for (int i = 0; i < n; ++i) { + y[i] = ggml_gelu_f32(x[i]); + } +} +#endif + +inline static float ggml_gelu_quick_f32(float x) { + return x*(1.0f/(1.0f+expf(GELU_QUICK_COEF*x))); +} + +//inline static void ggml_vec_gelu_quick_f16(const int n, ggml_fp16_t * y, const ggml_fp16_t * x) { +// const uint16_t * i16 = (const uint16_t *) x; +// for (int i = 0; i < n; ++i) { +// y[i] = ggml_table_gelu_quick_f16[i16[i]]; +// } +//} + +#ifdef GGML_GELU_QUICK_FP16 +inline static void ggml_vec_gelu_quick_f32(const int n, float * y, const float * x) { + uint16_t t; + for (int i = 0; i < n; ++i) { + ggml_fp16_t fp16 = GGML_FP32_TO_FP16(x[i]); + memcpy(&t, &fp16, sizeof(uint16_t)); + y[i] = GGML_FP16_TO_FP32(ggml_table_gelu_quick_f16[t]); + } +} +#else +inline static void ggml_vec_gelu_quick_f32(const int n, float * y, const float * x) { + for (int i = 0; i < n; ++i) { + y[i] = ggml_gelu_quick_f32(x[i]); + } +} +#endif + +inline static void ggml_vec_gelu_quick_f16(const int n, ggml_fp16_t * y, const ggml_fp16_t * x) { + for (int i = 0; i < n; ++i) { + float v = GGML_FP16_TO_FP32(x[i]); + y[i] = GGML_FP32_TO_FP16(v*(1.0f/(1.0f+expf(GELU_QUICK_COEF*v)))); + } +} + +// Sigmoid Linear Unit (SiLU) function +inline static float ggml_silu_f32(float x) { + return x/(1.0f + expf(-x)); +} +inline static ggml_fp16_t ggml_silu_f16(ggml_fp16_t x) { + float v = GGML_FP16_TO_FP32(x); + return GGML_FP32_TO_FP16(v/(1.0f + expf(-v))); +} + +#if __FINITE_MATH_ONLY__ +#error "some routines in ggml.c require non-finite math arithmetics -- pass -fno-finite-math-only to the compiler to fix" +#error "ref: https://github.com/ggml-org/llama.cpp/pull/7154#issuecomment-2143844461" +#endif + +#if defined(__ARM_NEON) && defined(__aarch64__) + +// adapted from arm limited optimized routine +// the maximum error is 1.45358 plus 0.5 ulps +// numbers above 88.38 will flush to infinity +// numbers beneath -103.97 will flush to zero +inline static float32x4_t ggml_v_expf(float32x4_t x) { + const float32x4_t r = vdupq_n_f32(0x1.8p23f); + const float32x4_t z = vfmaq_f32(r, x, vdupq_n_f32(0x1.715476p+0f)); + const float32x4_t n = vsubq_f32(z, r); + const float32x4_t b = vfmsq_f32(vfmsq_f32(x, n, vdupq_n_f32(0x1.62e4p-1f)), n, + vdupq_n_f32(0x1.7f7d1cp-20f)); + const uint32x4_t e = vshlq_n_u32(vreinterpretq_u32_f32(z), 23); + const float32x4_t k = vreinterpretq_f32_u32(vaddq_u32(e, vreinterpretq_u32_f32(vdupq_n_f32(1)))); + const uint32x4_t c = vcagtq_f32(n, vdupq_n_f32(126)); + const float32x4_t u = vmulq_f32(b, b); + const float32x4_t j = vfmaq_f32( + vmulq_f32(vdupq_n_f32(0x1.ffffecp-1f), b), + vfmaq_f32(vfmaq_f32(vdupq_n_f32(0x1.fffdb6p-2f), vdupq_n_f32(0x1.555e66p-3f), b), + vfmaq_f32(vdupq_n_f32(0x1.573e2ep-5f), vdupq_n_f32(0x1.0e4020p-7f), b), u), u); + if (!vpaddd_u64(vreinterpretq_u64_u32(c))) + return vfmaq_f32(k, j, k); + const uint32x4_t d = vandq_u32(vclezq_f32(n), vdupq_n_u32(0x82000000)); + const float32x4_t s1 = vreinterpretq_f32_u32(vaddq_u32(d, vdupq_n_u32(0x7f000000))); + const float32x4_t s2 = vreinterpretq_f32_u32(vsubq_u32(e, d)); + return vbslq_f32(vcagtq_f32(n, vdupq_n_f32(192)), vmulq_f32(s1, s1), + vbslq_f32(c, vmulq_f32(vfmaq_f32(s2, s2, j), s1), vfmaq_f32(k, k, j))); +} + +// computes silu x/(1+exp(-x)) in single precision vector +inline static float32x4_t ggml_v_silu(float32x4_t x) { + const float32x4_t one = vdupq_n_f32(1.0f); + const float32x4_t zero = vdupq_n_f32(0.0f); + const float32x4_t neg_x = vsubq_f32(zero, x); + const float32x4_t exp_neg_x = ggml_v_expf(neg_x); + const float32x4_t one_plus_exp_neg_x = vaddq_f32(one, exp_neg_x); + return vdivq_f32(x, one_plus_exp_neg_x); +} + +#elif defined(__AVX512F__) && defined(__AVX512DQ__) + +// adapted from arm limited optimized routine +// the maximum error is 1.45358 plus 0.5 ulps +// numbers above 88.38 will flush to infinity +// numbers beneath -103.97 will flush to zero +inline static __m512 ggml_v_expf(__m512 x) { + const __m512 r = _mm512_set1_ps(0x1.8p23f); + const __m512 z = _mm512_fmadd_ps(x, _mm512_set1_ps(0x1.715476p+0f), r); + const __m512 n = _mm512_sub_ps(z, r); + const __m512 b = + _mm512_fnmadd_ps(n, _mm512_set1_ps(0x1.7f7d1cp-20f), + _mm512_fnmadd_ps(n, _mm512_set1_ps(0x1.62e4p-1f), x)); + const __mmask16 d = + _mm512_cmp_ps_mask(_mm512_abs_ps(n), _mm512_set1_ps(192), _CMP_GT_OQ); + const __m512 u = _mm512_mul_ps(b, b); + const __m512 j = _mm512_fmadd_ps( + _mm512_fmadd_ps(_mm512_fmadd_ps(_mm512_set1_ps(0x1.0e4020p-7f), b, + _mm512_set1_ps(0x1.573e2ep-5f)), + u, + _mm512_fmadd_ps(_mm512_set1_ps(0x1.555e66p-3f), b, + _mm512_set1_ps(0x1.fffdb6p-2f))), + u, + _mm512_fmadd_ps(_mm512_set1_ps(0x1.ffffecp-1f), b, _mm512_set1_ps(1.0F))); + const __m512 res = _mm512_scalef_ps(j, n); + if (_mm512_kortestz(d, d)) + return res; + const __m512 zero = _mm512_setzero_ps(); + const __m512 alt = _mm512_mask_blend_ps( + _mm512_cmp_ps_mask(n, zero, _CMP_LE_OQ), _mm512_set1_ps(INFINITY), zero); + return _mm512_mask_blend_ps(d, res, alt); +} + +// computes silu x/(1+exp(-x)) in single precision vector +inline static __m512 ggml_v_silu(__m512 x) { + const __m512 one = _mm512_set1_ps(1); + const __m512 zero = _mm512_setzero_ps(); + const __m512 neg_x = _mm512_sub_ps(zero, x); + const __m512 exp_neg_x = ggml_v_expf(neg_x); + const __m512 one_plus_exp_neg_x = _mm512_add_ps(one, exp_neg_x); + return _mm512_div_ps(x, one_plus_exp_neg_x); +} + +#elif defined(__AVX2__) && defined(__FMA__) + +// adapted from arm limited optimized routine +// the maximum error is 1.45358 plus 0.5 ulps +// numbers above 88.38 will flush to infinity +// numbers beneath -103.97 will flush to zero +inline static __m256 ggml_v_expf(__m256 x) { + const __m256 r = _mm256_set1_ps(0x1.8p23f); + const __m256 z = _mm256_fmadd_ps(x, _mm256_set1_ps(0x1.715476p+0f), r); + const __m256 n = _mm256_sub_ps(z, r); + const __m256 b = _mm256_fnmadd_ps(n, _mm256_set1_ps(0x1.7f7d1cp-20f), + _mm256_fnmadd_ps(n, _mm256_set1_ps(0x1.62e4p-1f), x)); + const __m256i e = _mm256_slli_epi32(_mm256_castps_si256(z), 23); + const __m256 k = _mm256_castsi256_ps( + _mm256_add_epi32(e, _mm256_castps_si256(_mm256_set1_ps(1)))); + const __m256i c = _mm256_castps_si256( + _mm256_cmp_ps(_mm256_andnot_ps(_mm256_set1_ps(-0.f), n), + _mm256_set1_ps(126), _CMP_GT_OQ)); + const __m256 u = _mm256_mul_ps(b, b); + const __m256 j = _mm256_fmadd_ps(_mm256_fmadd_ps(_mm256_fmadd_ps(_mm256_set1_ps(0x1.0e4020p-7f), b, + _mm256_set1_ps(0x1.573e2ep-5f)), u, + _mm256_fmadd_ps(_mm256_set1_ps(0x1.555e66p-3f), b, + _mm256_set1_ps(0x1.fffdb6p-2f))), + u, _mm256_mul_ps(_mm256_set1_ps(0x1.ffffecp-1f), b)); + if (!_mm256_movemask_ps(_mm256_castsi256_ps(c))) + return _mm256_fmadd_ps(j, k, k); + const __m256i g = _mm256_and_si256( + _mm256_castps_si256(_mm256_cmp_ps(n, _mm256_setzero_ps(), _CMP_LE_OQ)), + _mm256_set1_epi32(0x82000000u)); + const __m256 s1 = + _mm256_castsi256_ps(_mm256_add_epi32(g, _mm256_set1_epi32(0x7f000000u))); + const __m256 s2 = _mm256_castsi256_ps(_mm256_sub_epi32(e, g)); + const __m256i d = _mm256_castps_si256( + _mm256_cmp_ps(_mm256_andnot_ps(_mm256_set1_ps(-0.f), n), + _mm256_set1_ps(192), _CMP_GT_OQ)); + return _mm256_or_ps( + _mm256_and_ps(_mm256_castsi256_ps(d), _mm256_mul_ps(s1, s1)), + _mm256_andnot_ps( + _mm256_castsi256_ps(d), + _mm256_or_ps( + _mm256_and_ps(_mm256_castsi256_ps(c), + _mm256_mul_ps(_mm256_fmadd_ps(s2, j, s2), s1)), + _mm256_andnot_ps(_mm256_castsi256_ps(c), _mm256_fmadd_ps(k, j, k))))); +} + +// computes silu x/(1+exp(-x)) in single precision vector +inline static __m256 ggml_v_silu(__m256 x) { + const __m256 one = _mm256_set1_ps(1); + const __m256 zero = _mm256_setzero_ps(); + const __m256 neg_x = _mm256_sub_ps(zero, x); + const __m256 exp_neg_x = ggml_v_expf(neg_x); + const __m256 one_plus_exp_neg_x = _mm256_add_ps(one, exp_neg_x); + return _mm256_div_ps(x, one_plus_exp_neg_x); +} + +#elif defined(__SSE2__) // __AVX2__ / __ARM_NEON + +#if defined(__FMA__) +#define MADD128(x, y, z) _mm_fmadd_ps(x, y, z) +#define NMADD128(x, y, z) _mm_fnmadd_ps(x, y, z) +#else +#define MADD128(x, y, z) _mm_add_ps(_mm_mul_ps(x, y), z) +#define NMADD128(x, y, z) _mm_sub_ps(z, _mm_mul_ps(x, y)) +#endif + +// adapted from arm limited optimized routine +// the maximum error is 1.45358 plus 0.5 ulps +// numbers above 88.38 will flush to infinity +// numbers beneath -103.97 will flush to zero +inline static __m128 ggml_v_expf(__m128 x) { + const __m128 r = _mm_set1_ps(0x1.8p23f); + const __m128 z = MADD128(x, _mm_set1_ps(0x1.715476p+0f), r); + const __m128 n = _mm_sub_ps(z, r); + const __m128 b = + NMADD128(n, _mm_set1_ps(0x1.7f7d1cp-20f), NMADD128(n, _mm_set1_ps(0x1.62e4p-1f), x)); + const __m128i e = _mm_slli_epi32(_mm_castps_si128(z), 23); + const __m128 k = _mm_castsi128_ps(_mm_add_epi32(e, _mm_castps_si128(_mm_set1_ps(1)))); + const __m128i c = + _mm_castps_si128(_mm_cmpgt_ps(_mm_andnot_ps(_mm_set1_ps(-0.f), n), _mm_set1_ps(126))); + const __m128 u = _mm_mul_ps(b, b); + const __m128 j = + MADD128(MADD128(MADD128(_mm_set1_ps(0x1.0e4020p-7f), b, _mm_set1_ps(0x1.573e2ep-5f)), u, + MADD128(_mm_set1_ps(0x1.555e66p-3f), b, _mm_set1_ps(0x1.fffdb6p-2f))), + u, _mm_mul_ps(_mm_set1_ps(0x1.ffffecp-1f), b)); + if (!_mm_movemask_epi8(c)) + return MADD128(j, k, k); + const __m128i g = _mm_and_si128(_mm_castps_si128(_mm_cmple_ps(n, _mm_setzero_ps())), + _mm_set1_epi32(0x82000000u)); + const __m128 s1 = _mm_castsi128_ps(_mm_add_epi32(g, _mm_set1_epi32(0x7f000000u))); + const __m128 s2 = _mm_castsi128_ps(_mm_sub_epi32(e, g)); + const __m128i d = + _mm_castps_si128(_mm_cmpgt_ps(_mm_andnot_ps(_mm_set1_ps(-0.f), n), _mm_set1_ps(192))); + return _mm_or_ps( + _mm_and_ps(_mm_castsi128_ps(d), _mm_mul_ps(s1, s1)), + _mm_andnot_ps(_mm_castsi128_ps(d), + _mm_or_ps(_mm_and_ps(_mm_castsi128_ps(c), _mm_mul_ps(MADD128(s2, j, s2), s1)), + _mm_andnot_ps(_mm_castsi128_ps(c), MADD128(k, j, k))))); +} + +// computes silu x/(1+exp(-x)) in single precision vector +inline static __m128 ggml_v_silu(__m128 x) { + const __m128 one = _mm_set1_ps(1); + const __m128 zero = _mm_setzero_ps(); + const __m128 neg_x = _mm_sub_ps(zero, x); + const __m128 exp_neg_x = ggml_v_expf(neg_x); + const __m128 one_plus_exp_neg_x = _mm_add_ps(one, exp_neg_x); + return _mm_div_ps(x, one_plus_exp_neg_x); +} + +#endif // __ARM_NEON / __AVX2__ / __SSE2__ + +inline static void ggml_vec_silu_f16(const int n, ggml_fp16_t * y, const ggml_fp16_t * x) { + for (int i = 0; i < n; ++i) { + y[i] = ggml_silu_f16(x[i]); + } +} + +inline static float ggml_silu_backward_f32(float x, float dy) { + const float s = 1.0f/(1.0f + expf(-x)); + return dy*s*(1.0f + x*(1.0f - s)); +} + +inline static ggml_fp16_t ggml_silu_backward_f16(ggml_fp16_t x, ggml_fp16_t dy) { + const float v = GGML_FP16_TO_FP32(x); + const float s = 1.0f/(1.0f + expf(-v)); + return GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(dy)*s*(1.0f + v*(1.0f - s))); +} + +inline static void ggml_vec_silu_backward_f32(const int n, float * dx, const float * x, const float * dy) { + for (int i = 0; i < n; ++i) { + dx[i] = ggml_silu_backward_f32(x[i], dy[i]); + } +} + +inline static void ggml_vec_silu_backward_f16(const int n, ggml_fp16_t * dx, const ggml_fp16_t * x, const ggml_fp16_t * dy) { + for (int i = 0; i < n; ++i) { + dx[i] = ggml_silu_backward_f16(x[i], dy[i]); + } +} + +inline static void ggml_vec_sum_f32(const int n, float * s, const float * x) { +#ifndef GGML_USE_ACCELERATE + ggml_float sum = 0.0; + for (int i = 0; i < n; ++i) { + sum += (ggml_float)x[i]; + } + *s = (float)sum; +#else + vDSP_sve(x, 1, s, n); +#endif +} + +inline static void ggml_vec_sum_f32_ggf(const int n, ggml_float * s, const float * x) { + ggml_float sum = 0.0; + for (int i = 0; i < n; ++i) { + sum += (ggml_float)x[i]; + } + *s = sum; +} + +inline static void ggml_vec_sum_f16_ggf(const int n, float * s, const ggml_fp16_t * x) { + float sum = 0.0f; + for (int i = 0; i < n; ++i) { + sum += GGML_FP16_TO_FP32(x[i]); + } + *s = sum; +} + +inline static void ggml_vec_sum_bf16_ggf(const int n, float * s, const ggml_bf16_t * x) { + float sum = 0.0f; + for (int i = 0; i < n; ++i) { + sum += GGML_BF16_TO_FP32(x[i]); + } + *s = sum; +} + +inline static void ggml_vec_max_f32(const int n, float * s, const float * x) { +#ifndef GGML_USE_ACCELERATE + float max = -INFINITY; + for (int i = 0; i < n; ++i) { + max = MAX(max, x[i]); + } + *s = max; +#else + vDSP_maxv(x, 1, s, n); +#endif +} + +inline static void ggml_vec_norm_inv_f32(const int n, float * s, const float * x) { + ggml_vec_norm_f32(n, s, x); + *s = 1.f/(*s); +} + +inline static void ggml_vec_argmax_f32(const int n, int * s, const float * x) { + float max = -INFINITY; + int idx = 0; + for (int i = 0; i < n; ++i) { + max = MAX(max, x[i]); + if (max == x[i]) { idx = i; } + } + *s = idx; +} + +#ifdef __cplusplus +} +#endif diff --git a/ggml/src/ggml-cuda/CMakeLists.txt b/ggml/src/ggml-cuda/CMakeLists.txt index 119fd39b8..8623214c7 100644 --- a/ggml/src/ggml-cuda/CMakeLists.txt +++ b/ggml/src/ggml-cuda/CMakeLists.txt @@ -7,7 +7,7 @@ if (CUDAToolkit_FOUND) if (NOT DEFINED CMAKE_CUDA_ARCHITECTURES) # native == GPUs available at build time - # 52 == Maxwell, lowest CUDA 12 standard + # 50 == Maxwell, lowest CUDA 12 standard # 60 == P100, FP16 CUDA intrinsics # 61 == Pascal, __dp4a instruction (per-byte integer dot product) # 70 == V100, FP16 tensor cores @@ -15,9 +15,9 @@ if (CUDAToolkit_FOUND) if (GGML_NATIVE AND CUDAToolkit_VERSION VERSION_GREATER_EQUAL "11.6" AND CMAKE_VERSION VERSION_GREATER_EQUAL "3.24") set(CMAKE_CUDA_ARCHITECTURES "native") elseif(GGML_CUDA_F16 OR GGML_CUDA_DMMV_F16) - set(CMAKE_CUDA_ARCHITECTURES "60;61;70;75") + set(CMAKE_CUDA_ARCHITECTURES "60;61;70;75;80") else() - set(CMAKE_CUDA_ARCHITECTURES "52;61;70;75") + set(CMAKE_CUDA_ARCHITECTURES "50;61;70;75;80") endif() endif() message(STATUS "Using CUDA architectures: ${CMAKE_CUDA_ARCHITECTURES}") @@ -69,6 +69,10 @@ if (CUDAToolkit_FOUND) add_compile_definitions(GGML_CUDA_NO_VMM) endif() + if (NOT GGML_CUDA_FA) + add_compile_definitions(GGML_CUDA_NO_FA) + endif() + if (GGML_CUDA_F16 OR GGML_CUDA_DMMV_F16) add_compile_definitions(GGML_CUDA_F16) endif() @@ -98,6 +102,15 @@ if (CUDAToolkit_FOUND) set(CUDA_FLAGS -use_fast_math) + if (CUDAToolkit_VERSION VERSION_GREATER_EQUAL "12.8") + # Options are: + # - none (not recommended) + # - speed (nvcc's default) + # - balance + # - size + list(APPEND CUDA_FLAGS -compress-mode=${GGML_CUDA_COMPRESSION_MODE}) + endif() + if (GGML_FATAL_WARNINGS) list(APPEND CUDA_FLAGS -Werror all-warnings) endif() diff --git a/ggml/src/ggml-cuda/binbcast.cu b/ggml/src/ggml-cuda/binbcast.cu index ce4b9cfb5..e1fbf0e13 100644 --- a/ggml/src/ggml-cuda/binbcast.cu +++ b/ggml/src/ggml-cuda/binbcast.cu @@ -294,11 +294,13 @@ static void ggml_cuda_op_bin_bcast( const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const void * src0_dd, const void * src1_dd, void * dst_dd, cudaStream_t stream) { - GGML_ASSERT(src1->type == GGML_TYPE_F32); + GGML_ASSERT(src1->type == GGML_TYPE_F32 || src1->type == GGML_TYPE_F16); if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) { op()(src0, src1, dst, (const float *)src0_dd, (const float *)src1_dd, (float *)dst_dd, stream); - } else if (src0->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F16) { + } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F16) { + op()(src0, src1, dst, (const half *) src0_dd, (const half *)src1_dd, (half *) dst_dd, stream); + } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F16) { op()(src0, src1, dst, (const half *) src0_dd, (const float *)src1_dd, (half *) dst_dd, stream); } else if (src0->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F32) { op()(src0, src1, dst, (const half *) src0_dd, (const float *)src1_dd, (float *)dst_dd, stream); diff --git a/ggml/src/ggml-cuda/clamp.cu b/ggml/src/ggml-cuda/clamp.cu index 8009a3e3d..fe415e7f7 100644 --- a/ggml/src/ggml-cuda/clamp.cu +++ b/ggml/src/ggml-cuda/clamp.cu @@ -1,34 +1,45 @@ #include "clamp.cuh" -static __global__ void clamp_f32(const float * x, float * dst, const float min, const float max, const int k) { +static __device__ __forceinline__ float op_clamp(float x, float min, float max) { + return fminf(fmaxf(x, min), max); +} + +template +static __global__ void op_clamp_kernel(const T * x, T * dst, const T min, const T max, const int k) { const int i = blockDim.x*blockIdx.x + threadIdx.x; if (i >= k) { return; } - dst[i] = x[i] < min ? min : (x[i] > max ? max : x[i]); + dst[i] = (T)op_clamp((float)x[i], (float)min, (float)max); } -static void clamp_f32_cuda(const float * x, float * dst, const float min, const float max, const int k, cudaStream_t stream) { +template +static void clamp_cuda(const T * x, T * dst, const T min, const T max, const int k, cudaStream_t stream) { const int num_blocks = (k + CUDA_CLAMP_BLOCK_SIZE - 1) / CUDA_CLAMP_BLOCK_SIZE; - clamp_f32<<>>(x, dst, min, max, k); + op_clamp_kernel<<>>(x, dst, min, max, k); } void ggml_cuda_op_clamp(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { const ggml_tensor * src0 = dst->src[0]; - const float * src0_d = (const float *)src0->data; - float * dst_d = (float *)dst->data; + const void * src0_d = src0->data; + void * dst_d = dst->data; cudaStream_t stream = ctx.stream(); - GGML_ASSERT(src0->type == GGML_TYPE_F32); - GGML_ASSERT( dst->type == GGML_TYPE_F32); + GGML_ASSERT(src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16); + GGML_ASSERT( dst->type == GGML_TYPE_F32 || dst->type == GGML_TYPE_F16); + GGML_ASSERT(src0->type == dst->type); float min; float max; memcpy(&min, dst->op_params, sizeof(float)); memcpy(&max, (float *) dst->op_params + 1, sizeof(float)); - clamp_f32_cuda(src0_d, dst_d, min, max, ggml_nelements(src0), stream); + if (src0->type == GGML_TYPE_F16) { + clamp_cuda((const half *)src0_d, (half *)dst_d, (half)min, (half)max, ggml_nelements(src0), stream); + } else { + clamp_cuda((const float *)src0_d, (float *)dst_d, (float)min, (float)max, ggml_nelements(src0), stream); + } } diff --git a/ggml/src/ggml-cuda/common.cuh b/ggml/src/ggml-cuda/common.cuh index 174916bc9..8284a0017 100644 --- a/ggml/src/ggml-cuda/common.cuh +++ b/ggml/src/ggml-cuda/common.cuh @@ -41,14 +41,18 @@ #define CUDART_HMAX 11070 // CUDA 11.7, min. ver. for which __hmax and __hmax2 are known to work (may be higher than needed) #define CUDART_HMASK 12000 // CUDA 12.0, min. ver. for half2 -> uint mask comparisons -#define GGML_CUDA_CC_PASCAL 600 -#define GGML_CUDA_CC_DP4A 610 // minimum compute capability for __dp4a, an intrinsic for byte-wise dot products -#define GGML_CUDA_CC_VOLTA 700 -#define GGML_CUDA_CC_TURING 750 -#define GGML_CUDA_CC_AMPERE 800 -#define GGML_CUDA_CC_OFFSET_AMD 0x1000000 +#define GGML_CUDA_CC_PASCAL 600 +#define GGML_CUDA_CC_DP4A 610 // minimum compute capability for __dp4a, an intrinsic for byte-wise dot products +#define GGML_CUDA_CC_VOLTA 700 +#define GGML_CUDA_CC_TURING 750 +#define GGML_CUDA_CC_AMPERE 800 +#define GGML_CUDA_CC_ADA_LOVELACE 890 +#define GGML_CUDA_CC_OFFSET_AMD 0x1000000 +#define GGML_CUDA_CC_OFFSET_MTHREADS 0x0100000 +#define GGML_CUDA_CC_IS_NVIDIA(cc) (cc < GGML_CUDA_CC_OFFSET_MTHREADS) -// GCN/CNDA, wave size is 64 +// AMD +// GCN/CDNA, wave size is 64 #define GGML_CUDA_CC_GCN4 (GGML_CUDA_CC_OFFSET_AMD + 0x803) // Tonga, Fiji, Polaris, minimum for fast fp16 #define GGML_CUDA_CC_VEGA (GGML_CUDA_CC_OFFSET_AMD + 0x900) // Vega56/64, minimum for fp16 dual issue #define GGML_CUDA_CC_VEGA20 (GGML_CUDA_CC_OFFSET_AMD + 0x906) // MI50/Radeon VII, minimum for dp4a @@ -56,20 +60,73 @@ #define GGML_CUDA_CC_CDNA2 (GGML_CUDA_CC_OFFSET_AMD + 0x910) // MI210, minimum acc register renameing #define GGML_CUDA_CC_CDNA3 (GGML_CUDA_CC_OFFSET_AMD + 0x942) // MI300 -// RNDA removes MFMA, dp4a, xnack, acc registers, wave size is 32 +// RDNA removes MFMA, dp4a, xnack, acc registers, wave size is 32 #define GGML_CUDA_CC_RDNA1 (GGML_CUDA_CC_OFFSET_AMD + 0x1010) // RX 5000 #define GGML_CUDA_CC_RDNA2 (GGML_CUDA_CC_OFFSET_AMD + 0x1030) // RX 6000, minimum for dp4a #define GGML_CUDA_CC_RDNA3 (GGML_CUDA_CC_OFFSET_AMD + 0x1100) // RX 7000, minimum for WMMA +#define GGML_CUDA_CC_RDNA4 (GGML_CUDA_CC_OFFSET_AMD + 0x1200) // RX 9000 +#define GGML_CUDA_CC_IS_AMD(cc) (cc >= GGML_CUDA_CC_OFFSET_AMD) #define GGML_CUDA_CC_IS_RDNA(cc) (cc >= GGML_CUDA_CC_RDNA1) #define GGML_CUDA_CC_IS_RDNA1(cc) (cc >= GGML_CUDA_CC_RDNA1 && cc < GGML_CUDA_CC_RDNA2) #define GGML_CUDA_CC_IS_RDNA2(cc) (cc >= GGML_CUDA_CC_RDNA2 && cc < GGML_CUDA_CC_RDNA3) -#define GGML_CUDA_CC_IS_RDNA3(cc) (cc >= GGML_CUDA_CC_RDNA3) +#define GGML_CUDA_CC_IS_RDNA3(cc) (cc >= GGML_CUDA_CC_RDNA3 && cc < GGML_CUDA_CC_RDNA4) +#define GGML_CUDA_CC_IS_RDNA4(cc) (cc >= GGML_CUDA_CC_RDNA4) #define GGML_CUDA_CC_IS_GCN(cc) (cc > GGML_CUDA_CC_OFFSET_AMD && cc < GGML_CUDA_CC_CDNA) #define GGML_CUDA_CC_IS_CDNA(cc) (cc >= GGML_CUDA_CC_CDNA && cc < GGML_CUDA_CC_RDNA1) -#define GGML_CUDA_CC_QY1 210 -#define GGML_CUDA_CC_QY2 220 +// Moore Threads +#define GGML_CUDA_MUSA_ARCH_IS_QY1 (__MUSA_ARCH__ <= 210) + +#define GGML_CUDA_CC_QY1 (GGML_MUSA_CC_OFFSET_MTHREADS + 0x210) // MTT S80, MTT S3000 +#define GGML_CUDA_CC_QY2 (GGML_MUSA_CC_OFFSET_MTHREADS + 0x220) // MTT S4000 +#define GGML_CUDA_CC_NG (GGML_MUSA_CC_OFFSET_MTHREADS + 0x310) // TBD + +#define GGML_CUDA_CC_IS_MTHREADS(cc) (cc >= GGML_CUDA_CC_OFFSET_MTHREADS && cc < GGML_CUDA_CC_OFFSET_AMD) +#define GGML_CUDA_CC_IS_QY1(cc) (cc >= GGML_CUDA_CC_QY1 && cc < GGML_CUDA_CC_QY2) +#define GGML_CUDA_CC_IS_QY2(cc) (cc >= GGML_CUDA_CC_QY2 && cc < GGML_CUDA_CC_NEXT) +#define GGML_CUDA_CC_IS_NG(cc) (cc >= GGML_CUDA_CC_NG) + +#ifdef __CUDA_ARCH_LIST__ +constexpr bool ggml_cuda_has_arch_impl(int) { + return false; +} + +template +constexpr bool ggml_cuda_has_arch_impl(const int arch, const int first, Archs... rest) { + return arch == first || ggml_cuda_has_arch_impl(arch, rest...); +} + +constexpr bool ggml_cuda_has_arch(const int arch) { + return ggml_cuda_has_arch_impl(arch, __CUDA_ARCH_LIST__); +} + +constexpr int ggml_cuda_highest_compiled_arch_impl(const int arch, const int cur) { + if (cur == 0) { + GGML_ABORT("ggml was not compiled with any CUDA arch <= %d", arch); + } + return cur; +} + +template +constexpr int ggml_cuda_highest_compiled_arch_impl(const int arch, const int cur, const int first, Archs... rest) { + if (first <= arch && first > cur) { + return ggml_cuda_highest_compiled_arch_impl(arch, first, rest...); + } else { + return ggml_cuda_highest_compiled_arch_impl(arch, cur, rest...); + } +} + +constexpr int ggml_cuda_highest_compiled_arch(const int arch) { + return ggml_cuda_highest_compiled_arch_impl(arch, 0, __CUDA_ARCH_LIST__); +} +#else +static int ggml_cuda_highest_compiled_arch(const int arch) { + return arch; +} +#endif // __CUDA_ARCH_LIST__ + +// --------------------------------------------------------------------------------------------------------- #define MATRIX_ROW_PADDING 512 // last row of quant. matrices is a multiple of this to avoid out-of-bounds memory accesses @@ -124,11 +181,11 @@ static const char * cu_get_error_str(CUresult err) { #define CU_CHECK(err) CUDA_CHECK_GEN(err, CUDA_SUCCESS, cu_get_error_str) #endif -#if CUDART_VERSION >= 11100 || defined(GGML_USE_MUSA) +#if CUDART_VERSION >= 11010 || defined(GGML_USE_MUSA) #define GGML_CUDA_ASSUME(x) __builtin_assume(x) #else #define GGML_CUDA_ASSUME(x) -#endif // CUDART_VERSION >= 11100 +#endif // CUDART_VERSION >= 11010 #ifdef GGML_CUDA_F16 typedef half dfloat; // dequantize float @@ -154,26 +211,58 @@ typedef float2 dfloat2; #define FP16_MMA_AVAILABLE #endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA +#if defined(GGML_HIP_ROCWMMA_FATTN) && (defined(CDNA) || defined(RDNA3) || defined(RDNA4)) +#define FP16_MMA_AVAILABLE +#endif // defined(GGML_HIP_ROCWMMA_FATTN) && (defined(CDNA) || defined(RDNA3) || defined(RDNA4)) + #if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= GGML_CUDA_CC_TURING #define NEW_MMA_AVAILABLE #endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= GGML_CUDA_CC_TURING -#if !(defined(GGML_USE_MUSA) && __MUSA_ARCH__ <= GGML_CUDA_CC_QY1) -#define FLASH_ATTN_AVAILABLE -#endif // !(defined(GGML_USE_MUSA) && __MUSA_ARCH__ <= GGML_CUDA_CC_QY1) +#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE +#define CP_ASYNC_AVAILABLE +#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE -static constexpr bool fast_fp16_available(const int cc) { - return cc >= GGML_CUDA_CC_PASCAL && cc != 610; +#if !defined(GGML_CUDA_NO_FA) && !(defined(GGML_USE_MUSA) && GGML_CUDA_MUSA_ARCH_IS_QY1) +#define FLASH_ATTN_AVAILABLE +#endif // !defined(GGML_CUDA_NO_FA) && !(defined(GGML_USE_MUSA) && GGML_CUDA_MUSA_ARCH_IS_QY1) + +static bool fp16_available(const int cc) { + return ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_PASCAL; } -// Any FP16 tensor cores are available. -static constexpr bool fp16_mma_available(const int cc) { - return cc < GGML_CUDA_CC_OFFSET_AMD && cc >= GGML_CUDA_CC_VOLTA; +static bool fast_fp16_available(const int cc) { + return (GGML_CUDA_CC_IS_NVIDIA(cc) && fp16_available(cc) && cc != 610) || GGML_CUDA_CC_IS_AMD(cc); +} + +// To be used for feature selection of external libraries, e.g. cuBLAS. +static bool fast_fp16_hardware_available(const int cc) { + return (GGML_CUDA_CC_IS_NVIDIA(cc) && cc >= GGML_CUDA_CC_PASCAL && cc != 610) || GGML_CUDA_CC_IS_AMD(cc); +} + +// Any FP16 tensor core instructions are available for ggml code. +static bool fp16_mma_available(const int cc) { +#if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__) && !defined(GGML_HIP_ROCWMMA_FATTN) + return false; +#else + return (GGML_CUDA_CC_IS_NVIDIA(cc) && ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_VOLTA) || + GGML_CUDA_CC_IS_CDNA(cc) || GGML_CUDA_CC_IS_RDNA3(cc) || GGML_CUDA_CC_IS_RDNA4(cc); +#endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__) && !defined(GGML_HIP_ROCWMMA_FATTN) +} + +// To be used for feature selection of external libraries, e.g. cuBLAS. +static bool fp16_mma_hardware_available(const int cc) { + return (GGML_CUDA_CC_IS_NVIDIA(cc) && cc >= GGML_CUDA_CC_VOLTA) || + GGML_CUDA_CC_IS_CDNA(cc) || GGML_CUDA_CC_IS_RDNA3(cc) || GGML_CUDA_CC_IS_RDNA4(cc); } // Volta technically had FP16 tensor cores but they work very differently compared to Turing and later. -static constexpr bool new_mma_available(const int cc) { - return cc < GGML_CUDA_CC_OFFSET_AMD && cc >= GGML_CUDA_CC_TURING; +static bool new_mma_available(const int cc) { + return GGML_CUDA_CC_IS_NVIDIA(cc) && ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_TURING; +} + +static bool cp_async_available(const int cc) { + return cc < GGML_CUDA_CC_OFFSET_AMD && ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_AMPERE; } static constexpr __device__ int ggml_cuda_get_physical_warp_size() { @@ -199,6 +288,10 @@ static __device__ void no_device_code( __trap(); GGML_UNUSED(no_device_code); // suppress unused function warning + +#if defined(GGML_USE_MUSA) + __builtin_unreachable(); +#endif // defined(GGML_USE_MUSA) } #ifdef __CUDA_ARCH__ @@ -320,11 +413,11 @@ static __device__ __forceinline__ uint32_t __hgt2_mask(const half2 a, const half static __device__ __forceinline__ int ggml_cuda_dp4a(const int a, const int b, int c) { #if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__) -#if defined(__gfx906__) || defined(__gfx908__) || defined(__gfx90a__) || defined(RDNA2) +#if defined(CDNA) || defined(RDNA2) || defined(__gfx906__) c = __builtin_amdgcn_sdot4(a, b, c, false); -#elif defined(RDNA3) +#elif defined(RDNA3) || defined(RDNA4) c = __builtin_amdgcn_sudot4( true, a, true, b, c, false); -#elif defined(__gfx1010__) || defined(__gfx900__) +#elif defined(RDNA1) || defined(__gfx900__) int tmp1; int tmp2; asm("\n \ @@ -347,13 +440,13 @@ static __device__ __forceinline__ int ggml_cuda_dp4a(const int a, const int b, i #else // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__) -#if __CUDA_ARCH__ >= GGML_CUDA_CC_DP4A +#if __CUDA_ARCH__ >= GGML_CUDA_CC_DP4A || defined(GGML_USE_MUSA) return __dp4a(a, b, c); -#else // __CUDA_ARCH__ >= GGML_CUDA_CC_DP4A +#else // __CUDA_ARCH__ >= GGML_CUDA_CC_DP4A || defined(GGML_USE_MUSA) const int8_t * a8 = (const int8_t *) &a; const int8_t * b8 = (const int8_t *) &b; return c + a8[0]*b8[0] + a8[1]*b8[1] + a8[2]*b8[2] + a8[3]*b8[3]; -#endif // __CUDA_ARCH__ >= GGML_CUDA_CC_DP4A +#endif // __CUDA_ARCH__ >= GGML_CUDA_CC_DP4A || defined(GGML_USE_MUSA) #endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__) } @@ -603,7 +696,7 @@ struct ggml_tensor_extra_gpu { }; -#if ((CUDART_VERSION >= 12000) && defined(GGML_CUDA_USE_GRAPHS)) || defined(GGML_HIP_GRAPHS) +#if (defined(GGML_CUDA_USE_GRAPHS) || defined(GGML_HIP_GRAPHS)) #define USE_CUDA_GRAPH #endif @@ -636,7 +729,13 @@ struct ggml_cuda_graph { bool disable_due_to_failed_graph_capture = false; int number_consecutive_updates = 0; std::vector ggml_graph_properties; - std::vector updated_kernel_arg; + bool use_cpy_indirection = false; + std::vector cpy_dest_ptrs; + char ** dest_ptrs_d; + int dest_ptrs_size = 0; + // Index to allow each cpy kernel to be aware of it's position within the graph + // relative to other cpy nodes. + int graph_cpynode_index = -1; #endif }; diff --git a/ggml/src/ggml-cuda/concat.cu b/ggml/src/ggml-cuda/concat.cu index aafbaf803..e9ffd274b 100644 --- a/ggml/src/ggml-cuda/concat.cu +++ b/ggml/src/ggml-cuda/concat.cu @@ -38,7 +38,7 @@ static __global__ void concat_f32_dim1(const float * x, const float * y, float * blockIdx.y * ne0 + blockIdx.z * ne0 * gridDim.y; - if (blockIdx.y < ne01) { // src0 + if (blockIdx.y < (unsigned)ne01) { // src0 int offset_src = nidx + blockIdx.y * ne0 + @@ -64,7 +64,7 @@ static __global__ void concat_f32_dim2(const float * x, const float * y, float * blockIdx.y * ne0 + blockIdx.z * ne0 * gridDim.y; - if (blockIdx.z < ne02) { // src0 + if (blockIdx.z < (unsigned)ne02) { // src0 int offset_src = nidx + blockIdx.y * ne0 + diff --git a/ggml/src/ggml-cuda/conv-transpose-1d.cu b/ggml/src/ggml-cuda/conv-transpose-1d.cu index b1e94d6f7..fe4caf674 100644 --- a/ggml/src/ggml-cuda/conv-transpose-1d.cu +++ b/ggml/src/ggml-cuda/conv-transpose-1d.cu @@ -34,6 +34,10 @@ static __global__ void conv_transpose_1d_kernel( } } dst[global_index] = accumulator; + GGML_UNUSED(p0); GGML_UNUSED(d0); GGML_UNUSED(src0_ne3); + GGML_UNUSED(src1_ne3); GGML_UNUSED(dst_ne3); + GGML_UNUSED(src1_ne1); GGML_UNUSED(dst_ne1); + GGML_UNUSED(src1_ne2); GGML_UNUSED(dst_ne2); } static void conv_transpose_1d_f32_f32_cuda( @@ -75,8 +79,6 @@ void ggml_cuda_op_conv_transpose_1d(ggml_backend_cuda_context & ctx, ggml_tensor const int p0 = 0;//opts[3]; const int d0 = 1;//opts[4]; - const int64_t kernel_size = ggml_nelements(src0); - const int64_t input_size = ggml_nelements(src1); const int64_t output_size = ggml_nelements(dst); conv_transpose_1d_f32_f32_cuda(s0, p0, d0, output_size, diff --git a/ggml/src/ggml-cuda/convert.cu b/ggml/src/ggml-cuda/convert.cu index 5b0dfacef..a224ec0e1 100644 --- a/ggml/src/ggml-cuda/convert.cu +++ b/ggml/src/ggml-cuda/convert.cu @@ -577,9 +577,9 @@ static __global__ void convert_unary(const void * __restrict__ vx, dst_t * __res return; } - const src_t * x = (src_t *) vx; + const src_t * x = (const src_t *) vx; - y[i] = x[i]; + y[i] = float(x[i]); } template @@ -588,6 +588,17 @@ static void convert_unary_cuda(const void * __restrict__ vx, dst_t * __restrict_ convert_unary<<>>(vx, y, k); } +to_bf16_cuda_t ggml_get_to_bf16_cuda(ggml_type type) { + switch (type) { + case GGML_TYPE_F32: + return convert_unary_cuda; + case GGML_TYPE_F16: + return convert_unary_cuda; + default: + return nullptr; + } +} + to_fp16_cuda_t ggml_get_to_fp16_cuda(ggml_type type) { switch (type) { case GGML_TYPE_Q4_0: @@ -599,7 +610,7 @@ to_fp16_cuda_t ggml_get_to_fp16_cuda(ggml_type type) { case GGML_TYPE_Q5_1: return dequantize_block_cuda; case GGML_TYPE_Q8_0: - if (ggml_cuda_info().devices[ggml_cuda_get_device()].cc >= GGML_CUDA_CC_PASCAL) { + if (fp16_available(ggml_cuda_info().devices[ggml_cuda_get_device()].cc)) { return dequantize_block_q8_0_f16_cuda; } return dequantize_block_cuda; @@ -633,6 +644,8 @@ to_fp16_cuda_t ggml_get_to_fp16_cuda(ggml_type type) { return dequantize_row_iq3_s_cuda; case GGML_TYPE_F32: return convert_unary_cuda; + case GGML_TYPE_BF16: + return convert_unary_cuda; default: return nullptr; } diff --git a/ggml/src/ggml-cuda/convert.cuh b/ggml/src/ggml-cuda/convert.cuh index 5394be9f1..411a13cf1 100644 --- a/ggml/src/ggml-cuda/convert.cuh +++ b/ggml/src/ggml-cuda/convert.cuh @@ -7,7 +7,10 @@ using to_t_cuda_t = void (*)(const void * __restrict__ x, T * __restrict__ y, in typedef to_t_cuda_t to_fp32_cuda_t; typedef to_t_cuda_t to_fp16_cuda_t; +typedef to_t_cuda_t to_bf16_cuda_t; to_fp16_cuda_t ggml_get_to_fp16_cuda(ggml_type type); +to_bf16_cuda_t ggml_get_to_bf16_cuda(ggml_type type); + to_fp32_cuda_t ggml_get_to_fp32_cuda(ggml_type type); diff --git a/ggml/src/ggml-cuda/cp-async.cuh b/ggml/src/ggml-cuda/cp-async.cuh new file mode 100644 index 000000000..ecb659997 --- /dev/null +++ b/ggml/src/ggml-cuda/cp-async.cuh @@ -0,0 +1,46 @@ +// Simplified API for asynchronous data loading. + +#include "common.cuh" + +// Copies data from global to shared memory, cg == cache global. +// Both the src and dst pointers must be aligned to 16 bit. +// Shared memory uses 32 bit addressing, the pointer is passed as unsigned int. +// Generic pointers can be converted to 32 bit shared memory pointers using __cvta_generic_to_shared. +// Only the 16 bit copy is exposed because 4 and 8 bit copies did not yield performance improvements. +template +static __device__ __forceinline__ void cp_async_cg_16(const unsigned int dst, const void * src) { + static_assert(preload == 0 || preload == 64 || preload == 128 || preload == 256, "bad preload"); +#ifdef CP_ASYNC_AVAILABLE +#if CUDART_VERSION >= 11040 + if (preload == 256) { + asm volatile("cp.async.cg.shared.global.L2::256B [%0], [%1], 16;" + : : "r"(dst), "l"(src)); + } else if (preload == 128) { + asm volatile("cp.async.cg.shared.global.L2::128B [%0], [%1], 16;" + : : "r"(dst), "l"(src)); + } else if (preload == 64) { + asm volatile("cp.async.cg.shared.global.L2::64B [%0], [%1], 16;" + : : "r"(dst), "l"(src)); + } else +#endif // CUDART_VERSION >= 11040 + { + asm volatile("cp.async.cg.shared.global [%0], [%1], 16;" + : : "r"(dst), "l"(src)); + } +#else + GGML_UNUSED(dst); + GGML_UNUSED(src); + NO_DEVICE_CODE; +#endif // CP_ASYNC_AVAILABLE +} + +// Makes each thread wait until its asynchronous data copies are done. +// This does NOT provide any additional synchronization. +// In particular, when copying data with multiple warps a call to __syncthreads will be needed. +static __device__ __forceinline__ void cp_async_wait_all() { +#ifdef CP_ASYNC_AVAILABLE + asm volatile("cp.async.wait_all;"); +#else + NO_DEVICE_CODE; +#endif // CP_ASYNC_AVAILABLE +} diff --git a/ggml/src/ggml-cuda/cpy.cu b/ggml/src/ggml-cuda/cpy.cu index 54c0f66d2..4f4faa3e6 100644 --- a/ggml/src/ggml-cuda/cpy.cu +++ b/ggml/src/ggml-cuda/cpy.cu @@ -1,4 +1,5 @@ #include "cpy.cuh" +#include "dequantize.cuh" typedef void (*cpy_kernel_t)(const char * cx, char * cdst); @@ -9,6 +10,13 @@ static __device__ void cpy_1_f32_f32(const char * cxi, char * cdsti) { *dsti = *xi; } +static __device__ void cpy_1_f32_bf16(const char * cxi, char * cdsti) { + const float * xi = (const float *) cxi; + nv_bfloat16 * dsti = (nv_bfloat16 *) cdsti; + + *dsti = *xi; +} + static __device__ void cpy_1_f32_f16(const char * cxi, char * cdsti) { const float * xi = (const float *) cxi; half * dsti = (half *) cdsti; @@ -31,16 +39,18 @@ static __device__ void cpy_1_f16_f32(const char * cxi, char * cdsti) { } template -static __global__ void cpy_f32_f16(const char * cx, char * cdst, const int ne, +static __global__ void cpy_f32_f16(const char * cx, char * cdst_direct, const int ne, const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02, const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, - const int nb12, const int nb13) { + const int nb12, const int nb13, char ** cdst_indirect, int graph_cpynode_index) { const int64_t i = blockDim.x*blockIdx.x + threadIdx.x; if (i >= ne) { return; } + char * cdst = (cdst_indirect != nullptr) ? cdst_indirect[graph_cpynode_index]: cdst_direct; + // determine indices i03/i13, i02/i12, i01/i11, i00/i10 as a function of index i of flattened tensor // then combine those indices with the corresponding byte offsets to get the total offsets const int64_t i03 = i/(ne00 * ne01 * ne02); @@ -82,13 +92,14 @@ static __device__ void cpy_blck_f32_q8_0(const char * cxi, char * cdsti) { } static __device__ void cpy_blck_q8_0_f32(const char * cxi, char * cdsti) { - const block_q8_0 * xi = (const block_q8_0 *) cxi; - float * dsti = (float *) cdsti; + float * cdstf = (float *)(cdsti); - const float d = (float)xi->d; - - for (int j = 0; j < QK8_0; j++) { - dsti[j] = xi->qs[j] * d; +#pragma unroll + for (int j = 0; j < QK8_0; j += 2) { + dfloat2 dq; + dequantize_q8_0(cxi, 0, j, dq); + *(cdstf + j) = dq.x; + *(cdstf + j + 1) = dq.y; } } @@ -225,6 +236,18 @@ static __device__ void cpy_blck_f32_q5_1(const char * cxi, char * cdsti) { memcpy(dsti->qh, &qh, sizeof(qh)); } +template +static __device__ void cpy_blck_q_f32(const char * cxi, char * cdsti) { + float * cdstf = (float *)(cdsti); + +#pragma unroll + for (int j = 0; j < qk/2; j++) { + dfloat2 dq; + dequant(cxi, 0, j, dq); + *(cdstf + j) = dq.x; + *(cdstf + j + qk/2) = dq.y; + } +} static __device__ __forceinline__ int best_index_int8(int n, const int8_t * val, float x) { if (x <= val[0]) return 0; @@ -274,16 +297,18 @@ static __device__ void cpy_blck_f32_iq4_nl(const char * cxi, char * cdsti) { } template -static __global__ void cpy_f32_q(const char * cx, char * cdst, const int ne, +static __global__ void cpy_f32_q(const char * cx, char * cdst_direct, const int ne, const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02, const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, - const int nb12, const int nb13) { + const int nb12, const int nb13, char ** cdst_indirect, int graph_cpynode_index) { const int i = (blockDim.x*blockIdx.x + threadIdx.x)*qk; if (i >= ne) { return; } + char * cdst = (cdst_indirect != nullptr) ? cdst_indirect[graph_cpynode_index]: cdst_direct; + const int i03 = i/(ne00 * ne01 * ne02); const int i02 = (i - i03*ne00*ne01*ne02 )/ (ne00*ne01); const int i01 = (i - i03*ne00*ne01*ne02 - i02*ne01*ne00) / ne00; @@ -300,16 +325,18 @@ static __global__ void cpy_f32_q(const char * cx, char * cdst, const int ne, } template -static __global__ void cpy_q_f32(const char * cx, char * cdst, const int ne, +static __global__ void cpy_q_f32(const char * cx, char * cdst_direct, const int ne, const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02, const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, - const int nb12, const int nb13) { + const int nb12, const int nb13, char ** cdst_indirect, int graph_cpynode_index) { const int i = (blockDim.x*blockIdx.x + threadIdx.x)*qk; if (i >= ne) { return; } + char * cdst = (cdst_indirect != nullptr) ? cdst_indirect[graph_cpynode_index]: cdst_direct; + const int i03 = i/(ne00 * ne01 * ne02); const int i02 = (i - i03*ne00*ne01*ne02 )/ (ne00*ne01); const int i01 = (i - i03*ne00*ne01*ne02 - i02*ne01*ne00) / ne00; @@ -325,120 +352,203 @@ static __global__ void cpy_q_f32(const char * cx, char * cdst, const int ne, cpy_blck(cx + x_offset, cdst + dst_offset); } +// Copy destination pointers to GPU to be available when pointer indirection is in use + +void ggml_cuda_cpy_dest_ptrs_copy(ggml_cuda_graph * cuda_graph, char ** host_dest_ptrs, const int host_dest_ptrs_size, cudaStream_t stream) { +#if defined(GGML_CUDA_USE_GRAPHS) || defined(GGML_HIP_GRAPHS) + if (cuda_graph->dest_ptrs_size < host_dest_ptrs_size) { // (re-)allocate GPU memory for destination pointers + CUDA_CHECK(cudaStreamSynchronize(stream)); + if (cuda_graph->dest_ptrs_d != nullptr) { + CUDA_CHECK(cudaFree(cuda_graph->dest_ptrs_d)); + } + CUDA_CHECK(cudaMalloc(&cuda_graph->dest_ptrs_d, host_dest_ptrs_size*sizeof(char *))); + cuda_graph->dest_ptrs_size = host_dest_ptrs_size; + } + // copy destination pointers to GPU + CUDA_CHECK(cudaMemcpyAsync(cuda_graph->dest_ptrs_d, host_dest_ptrs, host_dest_ptrs_size*sizeof(char *), cudaMemcpyHostToDevice, stream)); + cuda_graph->graph_cpynode_index = 0; // reset index +#else + GGML_UNUSED(cuda_graph); GGML_UNUSED(host_dest_ptrs); + GGML_UNUSED(host_dest_ptrs_size); GGML_UNUSED(stream); +#endif +} + static void ggml_cpy_f16_f32_cuda( const char * cx, char * cdst, const int ne, const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02, - const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream) { + const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream, char ** cdst_indirect, int & graph_cpynode_index) { const int num_blocks = (ne + CUDA_CPY_BLOCK_SIZE - 1) / CUDA_CPY_BLOCK_SIZE; cpy_f32_f16<<>> - (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13); + (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++); } static void ggml_cpy_f32_f32_cuda( const char * cx, char * cdst, const int ne, const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02, - const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream) { + const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream, char ** cdst_indirect, int & graph_cpynode_index) { const int num_blocks = (ne + CUDA_CPY_BLOCK_SIZE - 1) / CUDA_CPY_BLOCK_SIZE; cpy_f32_f16<<>> - (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13); + (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++); +} + +static void ggml_cpy_f32_bf16_cuda( + const char * cx, char * cdst, const int ne, + const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02, + const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream, char ** cdst_indirect, int & graph_cpynode_index) { + + const int num_blocks = (ne + CUDA_CPY_BLOCK_SIZE - 1) / CUDA_CPY_BLOCK_SIZE; + cpy_f32_f16<<>> + (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++); } static void ggml_cpy_f32_f16_cuda( const char * cx, char * cdst, const int ne, const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02, - const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream) { + const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream, char ** cdst_indirect, int & graph_cpynode_index) { const int num_blocks = (ne + CUDA_CPY_BLOCK_SIZE - 1) / CUDA_CPY_BLOCK_SIZE; cpy_f32_f16<<>> - (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13); + (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++); } static void ggml_cpy_f32_q8_0_cuda( const char * cx, char * cdst, const int ne, const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02, - const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream) { + const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream, char ** cdst_indirect, int & graph_cpynode_index) { GGML_ASSERT(ne % QK8_0 == 0); const int num_blocks = ne / QK8_0; cpy_f32_q<<>> - (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13); + (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++); } static void ggml_cpy_q8_0_f32_cuda( const char * cx, char * cdst, const int ne, const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02, - const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream) { + const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream, char ** cdst_indirect, int & graph_cpynode_index) { const int num_blocks = ne; cpy_q_f32<<>> - (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13); + (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++); } static void ggml_cpy_f32_q4_0_cuda( const char * cx, char * cdst, const int ne, const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02, - const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream) { + const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream, char ** cdst_indirect, int & graph_cpynode_index) { GGML_ASSERT(ne % QK4_0 == 0); const int num_blocks = ne / QK4_0; cpy_f32_q<<>> - (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13); + (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++); +} + +static void ggml_cpy_q4_0_f32_cuda( + const char * cx, char * cdst, const int ne, + const int ne00, const int ne01, const int ne02, + const int nb00, const int nb01, const int nb02, + const int nb03, const int ne10, const int ne11, const int ne12, + const int nb10, const int nb11, const int nb12, const int nb13, + cudaStream_t stream, char ** cdst_indirect, int & graph_cpynode_index) { + const int num_blocks = ne; + cpy_q_f32, QK4_0><<>>( + cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, + ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++); } static void ggml_cpy_f32_q4_1_cuda( const char * cx, char * cdst, const int ne, const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02, - const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream) { + const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream, char ** cdst_indirect, int & graph_cpynode_index) { GGML_ASSERT(ne % QK4_1 == 0); const int num_blocks = ne / QK4_1; cpy_f32_q<<>> - (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13); + (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++); +} + +static void ggml_cpy_q4_1_f32_cuda( + const char * cx, char * cdst, const int ne, + const int ne00, const int ne01, const int ne02, + const int nb00, const int nb01, const int nb02, + const int nb03, const int ne10, const int ne11, const int ne12, + const int nb10, const int nb11, const int nb12, const int nb13, + cudaStream_t stream, char ** cdst_indirect, int & graph_cpynode_index) { + const int num_blocks = ne; + cpy_q_f32, QK4_1><<>>( + cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, + ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++); } static void ggml_cpy_f32_q5_0_cuda( const char * cx, char * cdst, const int ne, const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02, - const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream) { + const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream, char ** cdst_indirect, int & graph_cpynode_index) { GGML_ASSERT(ne % QK5_0 == 0); const int num_blocks = ne / QK5_0; cpy_f32_q<<>> - (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13); + (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++); +} + +static void ggml_cpy_q5_0_f32_cuda( + const char * cx, char * cdst, const int ne, + const int ne00, const int ne01, const int ne02, + const int nb00, const int nb01, const int nb02, + const int nb03, const int ne10, const int ne11, const int ne12, + const int nb10, const int nb11, const int nb12, const int nb13, + cudaStream_t stream, char ** cdst_indirect, int & graph_cpynode_index) { + const int num_blocks = ne; + cpy_q_f32, QK5_0><<>>( + cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, + ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++); } static void ggml_cpy_f32_q5_1_cuda( const char * cx, char * cdst, const int ne, const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02, - const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream) { + const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream, char ** cdst_indirect, int & graph_cpynode_index) { GGML_ASSERT(ne % QK5_1 == 0); const int num_blocks = ne / QK5_1; cpy_f32_q<<>> - (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13); + (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++); +} + +static void ggml_cpy_q5_1_f32_cuda( + const char * cx, char * cdst, const int ne, + const int ne00, const int ne01, const int ne02, + const int nb00, const int nb01, const int nb02, + const int nb03, const int ne10, const int ne11, const int ne12, + const int nb10, const int nb11, const int nb12, const int nb13, + cudaStream_t stream, char ** cdst_indirect, int & graph_cpynode_index) { + const int num_blocks = ne; + cpy_q_f32, QK5_1><<>>( + cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, + ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++); } static void ggml_cpy_f32_iq4_nl_cuda( const char * cx, char * cdst, const int ne, const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02, - const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream) { + const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream, char ** cdst_indirect, int & graph_cpynode_index) { GGML_ASSERT(ne % QK4_NL == 0); const int num_blocks = ne / QK4_NL; cpy_f32_q<<>> - (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13); + (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++); } static void ggml_cpy_f16_f16_cuda( const char * cx, char * cdst, const int ne, const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02, - const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream) { + const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream, char ** cdst_indirect, int & graph_cpynode_index) { const int num_blocks = (ne + CUDA_CPY_BLOCK_SIZE - 1) / CUDA_CPY_BLOCK_SIZE; cpy_f32_f16<<>> - (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13); + (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, cdst_indirect, graph_cpynode_index++); } void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, ggml_tensor * src1) { @@ -475,35 +585,62 @@ void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, gg char * src0_ddc = (char *) src0->data; char * src1_ddc = (char *) src1->data; + char ** dest_ptrs_d = nullptr; + int graph_cpynode_index = -1; +#if defined(GGML_CUDA_USE_GRAPHS) || defined(GGML_HIP_GRAPHS) + if(ctx.cuda_graph->use_cpy_indirection) { + dest_ptrs_d = ctx.cuda_graph->dest_ptrs_d; + graph_cpynode_index = ctx.cuda_graph->graph_cpynode_index; + } +#endif if (src0->type == src1->type && ggml_is_contiguous(src0) && ggml_is_contiguous(src1)) { GGML_ASSERT(ggml_nbytes(src0) == ggml_nbytes(src1)); CUDA_CHECK(cudaMemcpyAsync(src1_ddc, src0_ddc, ggml_nbytes(src0), cudaMemcpyDeviceToDevice, main_stream)); } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32) { - ggml_cpy_f32_f32_cuda (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream); + ggml_cpy_f32_f32_cuda (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index); + } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_BF16) { + ggml_cpy_f32_bf16_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index); } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F16) { - ggml_cpy_f32_f16_cuda (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream); + ggml_cpy_f32_f16_cuda (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index); } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q8_0) { - ggml_cpy_f32_q8_0_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream); + ggml_cpy_f32_q8_0_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index); } else if (src0->type == GGML_TYPE_Q8_0 && src1->type == GGML_TYPE_F32) { - ggml_cpy_q8_0_f32_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream); + ggml_cpy_q8_0_f32_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index); } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q4_0) { - ggml_cpy_f32_q4_0_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream); + ggml_cpy_f32_q4_0_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index); + } else if (src0->type == GGML_TYPE_Q4_0 && src1->type == GGML_TYPE_F32) { + ggml_cpy_q4_0_f32_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, + nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index); } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q4_1) { - ggml_cpy_f32_q4_1_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream); + ggml_cpy_f32_q4_1_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index); + } else if (src0->type == GGML_TYPE_Q4_1 && src1->type == GGML_TYPE_F32) { + ggml_cpy_q4_1_f32_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, + nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index); } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q5_0) { - ggml_cpy_f32_q5_0_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream); + ggml_cpy_f32_q5_0_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index); + } else if (src0->type == GGML_TYPE_Q5_0 && src1->type == GGML_TYPE_F32) { + ggml_cpy_q5_0_f32_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, + nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index); } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_IQ4_NL) { - ggml_cpy_f32_iq4_nl_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream); + ggml_cpy_f32_iq4_nl_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index); } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q5_1) { - ggml_cpy_f32_q5_1_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream); + ggml_cpy_f32_q5_1_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index); + } else if (src0->type == GGML_TYPE_Q5_1 && src1->type == GGML_TYPE_F32) { + ggml_cpy_q5_1_f32_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index); } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F16) { - ggml_cpy_f16_f16_cuda (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream); + ggml_cpy_f16_f16_cuda (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index); } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F32) { - ggml_cpy_f16_f32_cuda (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream); + ggml_cpy_f16_f32_cuda (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index); } else { GGML_ABORT("%s: unsupported type combination (%s to %s)\n", __func__, ggml_type_name(src0->type), ggml_type_name(src1->type)); } +#if defined(GGML_CUDA_USE_GRAPHS) || defined(GGML_HIP_GRAPHS) + if(ctx.cuda_graph->use_cpy_indirection) { + ctx.cuda_graph->graph_cpynode_index = graph_cpynode_index; + } +#endif + } void ggml_cuda_dup(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { @@ -516,6 +653,8 @@ void* ggml_cuda_cpy_fn(const ggml_tensor * src0, ggml_tensor * src1) { return nullptr; } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32) { return (void*) cpy_f32_f16; + } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_BF16) { + return (void*) cpy_f32_f16; } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F16) { return (void*) cpy_f32_f16; } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q8_0) { @@ -524,14 +663,22 @@ void* ggml_cuda_cpy_fn(const ggml_tensor * src0, ggml_tensor * src1) { return (void*) cpy_q_f32; } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q4_0) { return (void*) cpy_f32_q; + } else if (src0->type == GGML_TYPE_Q4_0 && src1->type == GGML_TYPE_F32) { + return (void*) cpy_q_f32, QK4_0>; } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q4_1) { return (void*) cpy_f32_q; + } else if (src0->type == GGML_TYPE_Q4_1 && src1->type == GGML_TYPE_F32) { + return (void*) cpy_q_f32, QK4_1>; } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q5_0) { return (void*) cpy_f32_q; + } else if (src0->type == GGML_TYPE_Q5_0 && src1->type == GGML_TYPE_F32) { + return (void*) cpy_q_f32, QK5_0>; } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_IQ4_NL) { return (void*) cpy_f32_q; } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q5_1) { return (void*) cpy_f32_q; + } else if (src0->type == GGML_TYPE_Q5_1 && src1->type == GGML_TYPE_F32) { + return (void*) cpy_q_f32, QK5_1>; } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F16) { return (void*) cpy_f32_f16; } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F32) { diff --git a/ggml/src/ggml-cuda/cpy.cuh b/ggml/src/ggml-cuda/cpy.cuh index 28b06cdda..6bed0564d 100644 --- a/ggml/src/ggml-cuda/cpy.cuh +++ b/ggml/src/ggml-cuda/cpy.cuh @@ -7,3 +7,5 @@ void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, gg void ggml_cuda_dup(ggml_backend_cuda_context & ctx, ggml_tensor * dst); void* ggml_cuda_cpy_fn(const ggml_tensor * src0, ggml_tensor * src1); + +void ggml_cuda_cpy_dest_ptrs_copy(ggml_cuda_graph * cuda_graph, char ** host_dest_ptrs, const int host_dest_ptrs_size, cudaStream_t stream); diff --git a/ggml/src/ggml-cuda/cross-entropy-loss.cu b/ggml/src/ggml-cuda/cross-entropy-loss.cu index 27599a2b0..0ce4afbb2 100644 --- a/ggml/src/ggml-cuda/cross-entropy-loss.cu +++ b/ggml/src/ggml-cuda/cross-entropy-loss.cu @@ -123,13 +123,13 @@ void ggml_cuda_cross_entropy_loss(ggml_backend_cuda_context & ctx, ggml_tensor * ggml_cuda_pool_alloc dst_tmp(pool, blocks_num.x); if (nbytes_shared <= smpbo) { -#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) +#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && !defined(GGML_USE_MUSA) static bool shared_memory_limit_raised[GGML_CUDA_MAX_DEVICES] = {false}; if (!shared_memory_limit_raised[id]) { - CUDA_CHECK(cudaFuncSetAttribute(cross_entropy_loss_back_f32, cudaFuncAttributeMaxDynamicSharedMemorySize, smpbo)); + CUDA_CHECK(cudaFuncSetAttribute(cross_entropy_loss_f32, cudaFuncAttributeMaxDynamicSharedMemorySize, smpbo)); shared_memory_limit_raised[id] = true; } -#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) +#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && !defined(GGML_USE_MUSA) cross_entropy_loss_f32<<>>(src0_d, src1_d, dst_tmp.ptr, ne00, nrows); } else { cross_entropy_loss_f32<<>>(src0_d, src1_d, dst_tmp.ptr, ne00, nrows); @@ -175,13 +175,13 @@ void ggml_cuda_cross_entropy_loss_back(ggml_backend_cuda_context & ctx, ggml_ten const size_t smpbo = ggml_cuda_info().devices[id].smpbo; if (nbytes_shared <= smpbo) { -#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) +#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && !defined(GGML_USE_MUSA) static bool shared_memory_limit_raised[GGML_CUDA_MAX_DEVICES] = {false}; if (!shared_memory_limit_raised[id]) { CUDA_CHECK(cudaFuncSetAttribute(cross_entropy_loss_back_f32, cudaFuncAttributeMaxDynamicSharedMemorySize, smpbo)); shared_memory_limit_raised[id] = true; } -#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) +#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && !defined(GGML_USE_MUSA) cross_entropy_loss_back_f32<<>>(grad_d, src0f_d, src1f_d, dst_d, ne00); } else { cross_entropy_loss_back_f32<<>>(grad_d, src0f_d, src1f_d, dst_d, ne00); diff --git a/ggml/src/ggml-cuda/fattn-common.cuh b/ggml/src/ggml-cuda/fattn-common.cuh index d40ee2da4..56121705b 100644 --- a/ggml/src/ggml-cuda/fattn-common.cuh +++ b/ggml/src/ggml-cuda/fattn-common.cuh @@ -52,7 +52,7 @@ typedef half (*vec_dot_KQ_f16_t)( typedef float (*vec_dot_KQ_f32_t)( const char * __restrict__ K_c, const void * __restrict__ Q_v, const int * __restrict__ Q_q8 , const void * __restrict__ Q_ds); -template +template static __device__ __forceinline__ T vec_dot_fattn_vec_KQ_q4_0( const char * __restrict__ K_c, const void * __restrict__ Q_v, const int * __restrict__ Q_q8, const void * __restrict__ Q_ds_v) { @@ -62,7 +62,7 @@ static __device__ __forceinline__ T vec_dot_fattn_vec_KQ_q4_0( T sum = 0.0f; #pragma unroll - for (int k_KQ_0 = 0; k_KQ_0 < D/sizeof(int); k_KQ_0 += WARP_SIZE) { + for (int k_KQ_0 = 0; k_KQ_0 < int(D/sizeof(int)); k_KQ_0 += warp_size) { const int k_KQ = k_KQ_0 + threadIdx.x; const int ib = k_KQ / QI8_1; @@ -70,7 +70,7 @@ static __device__ __forceinline__ T vec_dot_fattn_vec_KQ_q4_0( const int shift = k_KQ & (QI8_1/2); const int v = (get_int_b2(K_q4_0[ib].qs, iqs4) >> shift) & 0x0F0F0F0F; - const int u = Q_q8[k_KQ_0/WARP_SIZE]; + const int u = Q_q8[k_KQ_0/warp_size]; const int sumi = ggml_cuda_dp4a(v, u, 0); @@ -78,21 +78,21 @@ static __device__ __forceinline__ T vec_dot_fattn_vec_KQ_q4_0( if (std::is_same::value) { const half2 * Q_ds = (const half2 *) Q_ds_v; - const half2 sum2 = __half2half2(K_q4_0[ib].d) * Q_ds[k_KQ_0/WARP_SIZE]; + const half2 sum2 = __half2half2(K_q4_0[ib].d) * Q_ds[k_KQ_0/warp_size]; sum += (T) (((half) sumi)*__low2half(sum2) - __high2half(sum2) /* *8/QI8_1 == 1 */); } else #endif // FP16_AVAILABLE { const float2 * Q_ds = (const float2 *) Q_ds_v; - sum += (T) (__half2float(K_q4_0[ib].d) * (sumi*Q_ds[k_KQ_0/WARP_SIZE].x - (8/QI8_1)*Q_ds[k_KQ_0/WARP_SIZE].y)); + sum += (T) (__half2float(K_q4_0[ib].d) * (sumi*Q_ds[k_KQ_0/warp_size].x - (8/QI8_1)*Q_ds[k_KQ_0/warp_size].y)); } } return sum; } -template +template static __device__ __forceinline__ T vec_dot_fattn_vec_KQ_q4_1( const char * __restrict__ K_c, const void * __restrict__ Q_v, const int * __restrict__ Q_q8, const void * __restrict__ Q_ds_v) { @@ -102,7 +102,7 @@ static __device__ __forceinline__ T vec_dot_fattn_vec_KQ_q4_1( T sum = 0.0f; #pragma unroll - for (int k_KQ_0 = 0; k_KQ_0 < D/sizeof(int); k_KQ_0 += WARP_SIZE) { + for (int k_KQ_0 = 0; k_KQ_0 < int(D/sizeof(int)); k_KQ_0 += warp_size) { const int k_KQ = k_KQ_0 + threadIdx.x; const int ib = k_KQ / QI8_1; @@ -110,7 +110,7 @@ static __device__ __forceinline__ T vec_dot_fattn_vec_KQ_q4_1( const int shift = k_KQ & (QI8_1/2); const int v = (get_int_b4(K_q4_1[ib].qs, iqs4) >> shift) & 0x0F0F0F0F; - const int u = Q_q8[k_KQ_0/WARP_SIZE]; + const int u = Q_q8[k_KQ_0/warp_size]; const int sumi = ggml_cuda_dp4a(v, u, 0); @@ -118,7 +118,7 @@ static __device__ __forceinline__ T vec_dot_fattn_vec_KQ_q4_1( if (std::is_same::value) { const half2 * Q_ds = (const half2 *) Q_ds_v; - const half2 d4d8_m4s8 = K_q4_1[ib].dm * Q_ds[k_KQ_0/WARP_SIZE]; + const half2 d4d8_m4s8 = K_q4_1[ib].dm * Q_ds[k_KQ_0/warp_size]; const half2 sumid4d8_m4s8scaled = d4d8_m4s8 * make_half2(sumi, 1.0f/QI8_1); sum += (T) (__low2half(sumid4d8_m4s8scaled) + __high2half(sumid4d8_m4s8scaled)); } else @@ -126,8 +126,8 @@ static __device__ __forceinline__ T vec_dot_fattn_vec_KQ_q4_1( { const float2 * Q_ds = (const float2 *) Q_ds_v; - const float sumid4d8 = __low2float(K_q4_1[ib].dm)*Q_ds[k_KQ_0/WARP_SIZE].x * sumi; - const float m4s8scaled = __high2float(K_q4_1[ib].dm)*Q_ds[k_KQ_0/WARP_SIZE].y / QI8_1; + const float sumid4d8 = __low2float(K_q4_1[ib].dm)*Q_ds[k_KQ_0/warp_size].x * sumi; + const float m4s8scaled = __high2float(K_q4_1[ib].dm)*Q_ds[k_KQ_0/warp_size].y / QI8_1; sum += (T) (sumid4d8 + m4s8scaled); } @@ -136,7 +136,7 @@ static __device__ __forceinline__ T vec_dot_fattn_vec_KQ_q4_1( return sum; } -template +template static __device__ __forceinline__ T vec_dot_fattn_vec_KQ_q5_0( const char * __restrict__ K_c, const void * __restrict__ Q_v, const int * __restrict__ Q_q8, const void * __restrict__ Q_ds_v) { @@ -146,7 +146,7 @@ static __device__ __forceinline__ T vec_dot_fattn_vec_KQ_q5_0( T sum = 0.0f; #pragma unroll - for (int k_KQ_0 = 0; k_KQ_0 < D/sizeof(int); k_KQ_0 += WARP_SIZE) { + for (int k_KQ_0 = 0; k_KQ_0 < int(D/sizeof(int)); k_KQ_0 += warp_size) { const int k_KQ = k_KQ_0 + threadIdx.x; const int ib = k_KQ / QI8_1; @@ -161,7 +161,7 @@ static __device__ __forceinline__ T vec_dot_fattn_vec_KQ_q5_0( v |= (vh << 18) & 0x00100000; // 2 -> 20 v |= (vh << 25) & 0x10000000; // 3 -> 28 - const int u = Q_q8[k_KQ_0/WARP_SIZE]; + const int u = Q_q8[k_KQ_0/warp_size]; const int sumi = ggml_cuda_dp4a(v, u, 0); @@ -169,21 +169,21 @@ static __device__ __forceinline__ T vec_dot_fattn_vec_KQ_q5_0( if (std::is_same::value) { const half2 * Q_ds = (const half2 *) Q_ds_v; - const half2 sum2 = __half2half2(K_q5_0[ib].d) * Q_ds[k_KQ_0/WARP_SIZE]; + const half2 sum2 = __half2half2(K_q5_0[ib].d) * Q_ds[k_KQ_0/warp_size]; sum += (T) (((half) sumi)*__low2half(sum2) - __high2half(sum2)*__float2half(2.0f)) /* *16/QI8_1 == 2 */; } else #endif // FP16_AVAILABLE { const float2 * Q_ds = (const float2 *) Q_ds_v; - sum += (T) (__half2float(K_q5_0[ib].d) * (sumi*Q_ds[k_KQ_0/WARP_SIZE].x - (16/QI8_1)*Q_ds[k_KQ_0/WARP_SIZE].y)); + sum += (T) (__half2float(K_q5_0[ib].d) * (sumi*Q_ds[k_KQ_0/warp_size].x - (16/QI8_1)*Q_ds[k_KQ_0/warp_size].y)); } } return sum; } -template +template static __device__ __forceinline__ T vec_dot_fattn_vec_KQ_q5_1( const char * __restrict__ K_c, const void * __restrict__ Q_v, const int * __restrict__ Q_q8, const void * __restrict__ Q_ds_v) { @@ -193,7 +193,7 @@ static __device__ __forceinline__ T vec_dot_fattn_vec_KQ_q5_1( T sum = 0.0f; #pragma unroll - for (int k_KQ_0 = 0; k_KQ_0 < D/sizeof(int); k_KQ_0 += WARP_SIZE) { + for (int k_KQ_0 = 0; k_KQ_0 < int(D/sizeof(int)); k_KQ_0 += warp_size) { const int k_KQ = k_KQ_0 + threadIdx.x; const int ib = k_KQ / QI8_1; @@ -208,7 +208,7 @@ static __device__ __forceinline__ T vec_dot_fattn_vec_KQ_q5_1( v |= (vh << 18) & 0x00100000; // 2 -> 20 v |= (vh << 25) & 0x10000000; // 3 -> 28 - const int u = Q_q8[k_KQ_0/WARP_SIZE]; + const int u = Q_q8[k_KQ_0/warp_size]; const int sumi = ggml_cuda_dp4a(v, u, 0); @@ -216,7 +216,7 @@ static __device__ __forceinline__ T vec_dot_fattn_vec_KQ_q5_1( if (std::is_same::value) { const half2 * Q_ds = (const half2 *) Q_ds_v; - const half2 d5d8_m5s8 = K_q5_1[ib].dm * Q_ds[k_KQ_0/WARP_SIZE]; + const half2 d5d8_m5s8 = K_q5_1[ib].dm * Q_ds[k_KQ_0/warp_size]; const half2 sumid5d8_m5s8scaled = d5d8_m5s8 * make_half2(sumi, 1.0f/QI8_1); sum += (T) (__low2half(sumid5d8_m5s8scaled) + __high2half(sumid5d8_m5s8scaled)); } else @@ -224,8 +224,8 @@ static __device__ __forceinline__ T vec_dot_fattn_vec_KQ_q5_1( { const float2 * Q_ds = (const float2 *) Q_ds_v; - const float sumid5d8 = __low2float(K_q5_1[ib].dm)*Q_ds[k_KQ_0/WARP_SIZE].x * sumi; - const float m5s8scaled = __high2float(K_q5_1[ib].dm)*Q_ds[k_KQ_0/WARP_SIZE].y / QI8_1; + const float sumid5d8 = __low2float(K_q5_1[ib].dm)*Q_ds[k_KQ_0/warp_size].x * sumi; + const float m5s8scaled = __high2float(K_q5_1[ib].dm)*Q_ds[k_KQ_0/warp_size].y / QI8_1; sum += (T) (sumid5d8 + m5s8scaled); } @@ -234,7 +234,7 @@ static __device__ __forceinline__ T vec_dot_fattn_vec_KQ_q5_1( return sum; } -template +template static __device__ __forceinline__ T vec_dot_fattn_vec_KQ_q8_0( const char * __restrict__ K_c, const void * __restrict__ Q_v, const int * __restrict__ Q_q8, const void * __restrict__ Q_ds_v) { @@ -244,7 +244,7 @@ static __device__ __forceinline__ T vec_dot_fattn_vec_KQ_q8_0( T sum = 0.0f; #pragma unroll - for (int k_KQ_0 = 0; k_KQ_0 < D/sizeof(int); k_KQ_0 += WARP_SIZE) { + for (int k_KQ_0 = 0; k_KQ_0 < int(D/sizeof(int)); k_KQ_0 += warp_size) { const int k_KQ = k_KQ_0 + threadIdx.x; const int ib = k_KQ / QI8_0; @@ -255,19 +255,19 @@ static __device__ __forceinline__ T vec_dot_fattn_vec_KQ_q8_0( T Q_d; if (std::is_same::value) { const half2 * Q_ds = (const half2 *) Q_ds_v; - Q_d = __low2half(Q_ds[k_KQ_0/WARP_SIZE]); + Q_d = __low2half(Q_ds[k_KQ_0/warp_size]); } else { const float2 * Q_ds = (const float2 *) Q_ds_v; - Q_d = Q_ds[k_KQ_0/WARP_SIZE].x; + Q_d = Q_ds[k_KQ_0/warp_size].x; } - sum += vec_dot_q8_0_q8_1_impl(&v, &Q_q8[k_KQ_0/WARP_SIZE], K_q8_0[ib].d, Q_d); + sum += vec_dot_q8_0_q8_1_impl(&v, &Q_q8[k_KQ_0/warp_size], K_q8_0[ib].d, Q_d); } return sum; } -template +template static __device__ __forceinline__ T vec_dot_fattn_vec_KQ_f16( const char * __restrict__ K_c, const void * __restrict__ Q_v, const int * __restrict__ Q_q8 , const void * __restrict__ Q_ds_v) { @@ -282,11 +282,11 @@ static __device__ __forceinline__ T vec_dot_fattn_vec_KQ_f16( half2 sum2 = make_half2(0.0f, 0.0f); #pragma unroll - for (int k_KQ_0 = 0; k_KQ_0 < D/2; k_KQ_0 += WARP_SIZE) { + for (int k_KQ_0 = 0; k_KQ_0 < D/2; k_KQ_0 += warp_size) { const int k_KQ = k_KQ_0 + threadIdx.x; const half2 K_ik = K_h2[k_KQ]; - sum2 += K_ik * Q_h2[k_KQ_0/WARP_SIZE]; + sum2 += K_ik * Q_h2[k_KQ_0/warp_size]; } return __low2half(sum2) + __high2half(sum2); @@ -298,12 +298,12 @@ static __device__ __forceinline__ T vec_dot_fattn_vec_KQ_f16( float sum = 0.0f; #pragma unroll - for (int k_KQ_0 = 0; k_KQ_0 < D/2; k_KQ_0 += WARP_SIZE) { + for (int k_KQ_0 = 0; k_KQ_0 < D/2; k_KQ_0 += warp_size) { const int k_KQ = k_KQ_0 + threadIdx.x; const half2 K_ik = K_h2[k_KQ]; - sum += __low2float(K_ik) * Q_f2[k_KQ_0/WARP_SIZE].x; - sum += __high2float(K_ik) * Q_f2[k_KQ_0/WARP_SIZE].y; + sum += __low2float(K_ik) * Q_f2[k_KQ_0/warp_size].x; + sum += __high2float(K_ik) * Q_f2[k_KQ_0/warp_size].y; } return sum; @@ -315,14 +315,14 @@ static __device__ __forceinline__ void quantize_q8_1_to_shared( float vals[sizeof(int)] = {0.0f}; #pragma unroll - for (int l = 0; l < sizeof(int); ++l) { + for (int l = 0; l < int(sizeof(int)); ++l) { vals[l] = scale * x[4*threadIdx.x + l]; } float amax = fabsf(vals[0]); float sum = vals[0]; #pragma unroll - for (int l = 1; l < sizeof(int); ++l) { + for (int l = 1; l < int(sizeof(int)); ++l) { amax = fmaxf(amax, fabsf(vals[l])); sum += vals[l]; } @@ -338,7 +338,7 @@ static __device__ __forceinline__ void quantize_q8_1_to_shared( if (d != 0.0f) { #pragma unroll - for (int l = 0; l < sizeof(int); ++l) { + for (int l = 0; l < int(sizeof(int)); ++l) { q8[l] = roundf(vals[l] / d); } } @@ -474,25 +474,25 @@ static __device__ __forceinline__ T dequantize_1_f16(const void * __restrict__ v return x[i]; } -template +template constexpr __device__ vec_dot_KQ_f16_t get_vec_dot_KQ_f16(ggml_type type_K) { - return type_K == GGML_TYPE_Q4_0 ? vec_dot_fattn_vec_KQ_q4_0 : - type_K == GGML_TYPE_Q4_1 ? vec_dot_fattn_vec_KQ_q4_1 : - type_K == GGML_TYPE_Q5_0 ? vec_dot_fattn_vec_KQ_q5_0 : - type_K == GGML_TYPE_Q5_1 ? vec_dot_fattn_vec_KQ_q5_1 : - type_K == GGML_TYPE_Q8_0 ? vec_dot_fattn_vec_KQ_q8_0 : - type_K == GGML_TYPE_F16 ? vec_dot_fattn_vec_KQ_f16 : + return type_K == GGML_TYPE_Q4_0 ? vec_dot_fattn_vec_KQ_q4_0 : + type_K == GGML_TYPE_Q4_1 ? vec_dot_fattn_vec_KQ_q4_1 : + type_K == GGML_TYPE_Q5_0 ? vec_dot_fattn_vec_KQ_q5_0 : + type_K == GGML_TYPE_Q5_1 ? vec_dot_fattn_vec_KQ_q5_1 : + type_K == GGML_TYPE_Q8_0 ? vec_dot_fattn_vec_KQ_q8_0 : + type_K == GGML_TYPE_F16 ? vec_dot_fattn_vec_KQ_f16 : nullptr; } -template +template constexpr __device__ vec_dot_KQ_f32_t get_vec_dot_KQ_f32(ggml_type type_K) { - return type_K == GGML_TYPE_Q4_0 ? vec_dot_fattn_vec_KQ_q4_0 : - type_K == GGML_TYPE_Q4_1 ? vec_dot_fattn_vec_KQ_q4_1 : - type_K == GGML_TYPE_Q5_0 ? vec_dot_fattn_vec_KQ_q5_0 : - type_K == GGML_TYPE_Q5_1 ? vec_dot_fattn_vec_KQ_q5_1 : - type_K == GGML_TYPE_Q8_0 ? vec_dot_fattn_vec_KQ_q8_0 : - type_K == GGML_TYPE_F16 ? vec_dot_fattn_vec_KQ_f16 : + return type_K == GGML_TYPE_Q4_0 ? vec_dot_fattn_vec_KQ_q4_0 : + type_K == GGML_TYPE_Q4_1 ? vec_dot_fattn_vec_KQ_q4_1 : + type_K == GGML_TYPE_Q5_0 ? vec_dot_fattn_vec_KQ_q5_0 : + type_K == GGML_TYPE_Q5_1 ? vec_dot_fattn_vec_KQ_q5_1 : + type_K == GGML_TYPE_Q8_0 ? vec_dot_fattn_vec_KQ_q8_0 : + type_K == GGML_TYPE_F16 ? vec_dot_fattn_vec_KQ_f16 : nullptr; } @@ -516,27 +516,25 @@ constexpr __device__ dequantize_1_f32_t get_dequantize_1_f32(ggml_type type_V) { nullptr; } -// The HIP compiler for some reason complains that it can't unroll a loop because of the jt*ncols + j >= ne01 conditional. -#ifdef __clang__ -#pragma clang diagnostic push -#pragma clang diagnostic ignored "-Wpass-failed" -#endif // __clang__ - -template // D == head size -#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) +template // D == head size __launch_bounds__(D, 1) -#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) static __global__ void flash_attn_stream_k_fixup( float * __restrict__ dst, const float2 * __restrict__ dst_fixup, const int ne01, const int ne02, const int ne11) { - const float * dst_fixup_data = ((const float *) dst_fixup) + gridDim.x*(2*2*ncols); - - const int iter_k = ne11 / KQ_stride; - const int iter_j = (ne01 + (ncols - 1)) / ncols; + constexpr int ncols = ncols1*ncols2; const int bidx0 = blockIdx.x; + const int j = blockIdx.y; + const int c = blockIdx.z; + const int jc = j*ncols2 + c; + const int tid = threadIdx.x; - const int kbc0 = (bidx0 + 0)*iter_k*iter_j*ne02 / gridDim.x; - const int kbc0_stop = (bidx0 + 1)*iter_k*iter_j*ne02 / gridDim.x; + const float * dst_fixup_data = ((const float *) dst_fixup) + gridDim.x*(2*2*ncols); + + const int iter_k = ne11 / FATTN_KQ_STRIDE; + const int iter_j = (ne01 + (ncols1 - 1)) / ncols1; + + const int kbc0 = (bidx0 + 0)*iter_k*iter_j*(ne02/ncols2) / gridDim.x; + const int kbc0_stop = (bidx0 + 1)*iter_k*iter_j*(ne02/ncols2) / gridDim.x; const bool did_not_have_any_data = kbc0 == kbc0_stop; const bool wrote_beginning_of_tile = kbc0 % iter_k == 0; @@ -548,22 +546,22 @@ static __global__ void flash_attn_stream_k_fixup( const int channel = kbc0 / (iter_k*iter_j); const int jt = (kbc0 - channel*iter_k*iter_j) / iter_k; - dst += jt*ncols*ne02*D + channel*D; + if (jt*ncols1 + j >= ne01) { + return; + } + + dst += jt*ne02*(ncols1*D) + channel*(ncols2*D) + (j*ne02 + c)*D + tid; // Load the partial result that needs a fixup: - float dst_val[ncols] = {0.0f}; - float max_val[ncols] = {0.0f}; - float rowsum[ncols] = {0.0f}; -#pragma unroll - for (int j = 0; j < ncols; ++j) { - if (jt*ncols + j >= ne01) { - break; - } - dst_val[j] = dst[j*ne02*D + threadIdx.x]; + float dst_val = 0.0f; + float max_val = 0.0f; + float rowsum = 0.0f; + { + dst_val = *dst; - const float2 tmp = dst_fixup[bidx0*ncols + j]; - max_val[j] = tmp.x; - rowsum[j] = tmp.y; + const float2 tmp = dst_fixup[bidx0*ncols + jc]; + max_val = tmp.x; + rowsum = tmp.y; } // Iterate over previous blocks and compute the combined results. @@ -571,36 +569,30 @@ static __global__ void flash_attn_stream_k_fixup( int bidx = bidx0 - 1; int kbc_stop = kbc0; while(true) { - const int kbc = bidx*iter_k*iter_j*ne02 / gridDim.x; + const int kbc = bidx*iter_k*iter_j*(ne02/ncols2) / gridDim.x; if (kbc == kbc_stop) { // Did not have any data. bidx--; kbc_stop = kbc; continue; } -#pragma unroll - for (int j = 0; j < ncols; ++j) { - if (jt*ncols + j >= ne01) { - break; - } - const float dst_add = dst_fixup_data[bidx*ncols*D + j*D + threadIdx.x]; + const float dst_add = dst_fixup_data[bidx*ncols*D + jc*D + tid]; - const float2 tmp = dst_fixup[(gridDim.x + bidx)*ncols + j]; + const float2 tmp = dst_fixup[(gridDim.x + bidx)*ncols + jc]; - // Scale the current and new value accumulators depending on the max. values. - const float max_val_new = fmaxf(max_val[j], tmp.x); + // Scale the current and new value accumulators depending on the max. values. + const float max_val_new = fmaxf(max_val, tmp.x); - const float diff_val = max_val[j] - max_val_new; - const float diff_add = tmp.x - max_val_new; + const float diff_val = max_val - max_val_new; + const float diff_add = tmp.x - max_val_new; - const float scale_val = diff_val >= SOFTMAX_FTZ_THRESHOLD ? expf(diff_val) : 0.0f; - const float scale_add = diff_add >= SOFTMAX_FTZ_THRESHOLD ? expf(diff_add) : 0.0f; + const float scale_val = diff_val >= SOFTMAX_FTZ_THRESHOLD ? expf(diff_val) : 0.0f; + const float scale_add = diff_add >= SOFTMAX_FTZ_THRESHOLD ? expf(diff_add) : 0.0f; - dst_val[j] = scale_val*dst_val[j] + scale_add*dst_add; - rowsum[j] = scale_val*rowsum[j] + scale_add*tmp.y; + dst_val = scale_val*dst_val + scale_add*dst_add; + rowsum = scale_val*rowsum + scale_add*tmp.y; - max_val[j] = max_val_new; - } + max_val = max_val_new; // If this block started in a previous tile we are done and don't need to combine additional partial results. if (kbc % iter_k == 0 || kbc/iter_k < kbc0/iter_k) { @@ -611,63 +603,53 @@ static __global__ void flash_attn_stream_k_fixup( } // Write back final result: -#pragma unroll - for (int j = 0; j < ncols; ++j) { - if (jt*ncols + j >= ne01) { - return; - } - dst[j*ne02*D + threadIdx.x] = dst_val[j] / rowsum[j]; - } + *dst = dst_val / rowsum; } -#ifdef __clang__ -#pragma clang diagnostic pop -#endif // __clang__ - -template // D == head size +template // D == head size #if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) __launch_bounds__(D, 1) #endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) static __global__ void flash_attn_combine_results( const float * __restrict__ VKQ_parts, const float2 * __restrict__ VKQ_meta, - float * __restrict__ dst) { - VKQ_parts += parallel_blocks*D * gridDim.y*blockIdx.x; - VKQ_meta += parallel_blocks * gridDim.y*blockIdx.x; - dst += D * gridDim.y*blockIdx.x; + float * __restrict__ dst, + const int parallel_blocks) { + VKQ_parts += parallel_blocks*D * gridDim.z*blockIdx.x; + VKQ_meta += parallel_blocks * gridDim.z*blockIdx.x; + dst += D * gridDim.z*blockIdx.x; const int tid = threadIdx.x; __builtin_assume(tid < D); - __shared__ float2 meta[parallel_blocks]; + extern __shared__ float2 meta[]; if (tid < 2*parallel_blocks) { - ((float *) meta)[threadIdx.x] = ((const float *)VKQ_meta) [blockIdx.y*(2*parallel_blocks) + tid]; + ((float *) meta)[threadIdx.x] = ((const float *)VKQ_meta) [blockIdx.z*(2*parallel_blocks) + tid]; } __syncthreads(); float kqmax = meta[0].x; -#pragma unroll for (int l = 1; l < parallel_blocks; ++l) { kqmax = max(kqmax, meta[l].x); } float VKQ_numerator = 0.0f; float VKQ_denominator = 0.0f; -#pragma unroll for (int l = 0; l < parallel_blocks; ++l) { const float diff = meta[l].x - kqmax; - const float KQ_max_scale = expf(diff); + float KQ_max_scale = expf(diff); const uint32_t ftz_mask = 0xFFFFFFFF * (diff > SOFTMAX_FTZ_THRESHOLD); *((uint32_t *) &KQ_max_scale) &= ftz_mask; - VKQ_numerator += KQ_max_scale * VKQ_parts[l*gridDim.y*D + blockIdx.y*D + tid]; + VKQ_numerator += KQ_max_scale * VKQ_parts[l*gridDim.z*D + blockIdx.z*D + tid]; VKQ_denominator += KQ_max_scale * meta[l].y; } - dst[blockIdx.y*D + tid] = VKQ_numerator / VKQ_denominator; + dst[blockIdx.z*D + tid] = VKQ_numerator / VKQ_denominator; } +[[noreturn]] static void on_no_fattn_vec_case(const int D) { if (D == 64) { fprintf(stderr, "Unsupported KV type combination for head_size 64.\n"); @@ -689,12 +671,13 @@ static void on_no_fattn_vec_case(const int D) { } } -// parallel_blocks == 0 is stream-k decomposition -template +template void launch_fattn( - ggml_backend_cuda_context & ctx, ggml_tensor * dst, fattn_kernel_t fattn_kernel, - const int nwarps, const size_t nbytes_shared, const bool need_f16_K, const bool need_f16_V + ggml_backend_cuda_context & ctx, ggml_tensor * dst, fattn_kernel_t fattn_kernel, const int nwarps, const size_t nbytes_shared, + const int KQ_row_granularity, const bool need_f16_K, const bool need_f16_V, const bool stream_k, const int warp_size = WARP_SIZE ) { + constexpr int ncols = ncols1 * ncols2; + const ggml_tensor * Q = dst->src[0]; const ggml_tensor * K = dst->src[1]; const ggml_tensor * V = dst->src[2]; @@ -716,7 +699,9 @@ void launch_fattn( ggml_cuda_pool & pool = ctx.pool(); cudaStream_t main_stream = ctx.stream(); - const int nsm = ggml_cuda_info().devices[ggml_cuda_get_device()].nsm; + const int id = ggml_cuda_get_device(); + const int cc = ggml_cuda_info().devices[id].cc; + const int nsm = ggml_cuda_info().devices[id].nsm; ggml_cuda_pool_alloc K_f16(pool); ggml_cuda_pool_alloc V_f16(pool); @@ -761,28 +746,66 @@ void launch_fattn( nb23 = nb23*bs*sizeof(half)/ts; } - const int ntiles_x = ((Q->ne[1] + cols_per_block - 1) / cols_per_block); - const int ntiles_total = ntiles_x*Q->ne[2]*Q->ne[3]; + int parallel_blocks = 1; - const dim3 block_dim(WARP_SIZE, nwarps, 1); + const int ntiles_x = ((Q->ne[1] + ncols1 - 1) / ncols1); + const int ntiles_total = ntiles_x * (Q->ne[2] / ncols2) * Q->ne[3]; + + const dim3 block_dim(warp_size, nwarps, 1); dim3 blocks_num; - if (parallel_blocks == 0) { + if (stream_k) { // For short contexts it can be faster to have the SMs work on whole tiles because this lets us skip the fixup. - const int tiles_nwaves = (ntiles_total - nsm - 1) / nsm; - const bool tiles_inefficient = 3*nsm < 2*tiles_nwaves*ntiles_total; - const bool short_context = K->ne[1] < 4096; + const int max_blocks = 2*nsm; + const int tiles_nwaves = (ntiles_total + max_blocks - 1) / max_blocks; + const int tiles_efficiency_percent = 100 * ntiles_total / (max_blocks*tiles_nwaves); - const int nblocks_stream_k = 2*nsm; + const int nblocks_stream_k = max_blocks; - blocks_num.x = short_context && !tiles_inefficient ? ntiles_total : nblocks_stream_k; + const bool use_stream_k = cc >= GGML_CUDA_CC_ADA_LOVELACE || tiles_efficiency_percent < 75; + + blocks_num.x = use_stream_k ? nblocks_stream_k : ntiles_total; blocks_num.y = 1; blocks_num.z = 1; - dst_tmp_meta.alloc(blocks_num.x*cols_per_block * (2*2 + D) * sizeof(float)); + dst_tmp_meta.alloc(blocks_num.x*ncols * (2*2 + D) * sizeof(float)); } else { - blocks_num.x = parallel_blocks*ntiles_x; - blocks_num.y = Q->ne[2]; - blocks_num.z = Q->ne[3]; + GGML_ASSERT(K->ne[1] % KQ_row_granularity == 0); + const int ntiles_KQ = K->ne[1] / KQ_row_granularity; // Max. number of parallel blocks limited by tensor size. + + int max_blocks_per_sm = 1; // Max. number of active blocks limited by occupancy. + CUDA_CHECK(cudaOccupancyMaxActiveBlocksPerMultiprocessor(&max_blocks_per_sm, fattn_kernel, block_dim.x * block_dim.y * block_dim.z, nbytes_shared)); + + // parallel_blocks should be at least large enough to achieve max. occupancy for a single wave: + parallel_blocks = std::max((nsm * max_blocks_per_sm) / ntiles_total, 1); + + // parallel_blocks must not be larger than what the tensor size allows: + parallel_blocks = std::min(parallel_blocks, ntiles_KQ); + + // If ntiles_total % blocks_per_wave != 0 then some efficiency is lost due to tail effects. + // Test whether parallel_blocks can be set to a higher value for better efficiency. + const int blocks_per_wave = nsm * max_blocks_per_sm; + int nwaves_best = 0; + int efficiency_percent_best = 0; + for (int parallel_blocks_test = parallel_blocks; parallel_blocks_test <= ntiles_KQ; ++parallel_blocks_test) { + const int nblocks_total = ntiles_total * parallel_blocks_test; + const int nwaves = (nblocks_total + blocks_per_wave - 1) / blocks_per_wave; + const int efficiency_percent = 100 * nblocks_total / (nwaves*blocks_per_wave); + + // Stop trying configurations with more waves if we already have good efficiency to avoid excessive overhead. + if (efficiency_percent_best >= 90 && nwaves > nwaves_best) { + break; + } + + if (efficiency_percent > efficiency_percent_best) { + nwaves_best = nwaves; + efficiency_percent_best = efficiency_percent; + parallel_blocks = parallel_blocks_test; + } + } + + blocks_num.x = ntiles_x; + blocks_num.y = parallel_blocks; + blocks_num.z = Q->ne[2]*Q->ne[3]; if (parallel_blocks > 1) { dst_tmp.alloc(parallel_blocks*ggml_nelements(KQV)); @@ -790,7 +813,6 @@ void launch_fattn( } } - float scale = 1.0f; float max_bias = 0.0f; float logit_softcap = 0.0f; @@ -809,12 +831,13 @@ void launch_fattn( const float m0 = powf(2.0f, -(max_bias ) / n_head_log2); const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2); + GGML_ASSERT(block_dim.x % warp_size == 0); fattn_kernel<<>>( (const char *) Q->data, K_data, V_data, mask ? ((const char *) mask->data) : nullptr, - (parallel_blocks) > 1 ? dst_tmp.ptr : (float *) KQV->data, dst_tmp_meta.ptr, + !stream_k && parallel_blocks > 1 ? dst_tmp.ptr : (float *) KQV->data, dst_tmp_meta.ptr, scale, max_bias, m0, m1, n_head_log2, logit_softcap, Q->ne[0], Q->ne[1], Q->ne[2], Q->ne[3], K->ne[0], K->ne[1], K->ne[2], K->ne[3], @@ -826,22 +849,23 @@ void launch_fattn( ); CUDA_CHECK(cudaGetLastError()); - if constexpr (parallel_blocks == 0) { - if (blocks_num.x % ntiles_total != 0) { // Fixup is only needed if the SMs work on fractional tiles. + if (stream_k) { + if (ntiles_total % blocks_num.x != 0) { // Fixup is only needed if the SMs work on fractional tiles. const dim3 block_dim_combine(D, 1, 1); - const dim3 blocks_num_combine = blocks_num; + const dim3 blocks_num_combine = {blocks_num.x, ncols1, ncols2}; - flash_attn_stream_k_fixup + flash_attn_stream_k_fixup <<>> ((float *) KQV->data, dst_tmp_meta.ptr, Q->ne[1], Q->ne[2], K->ne[1]); } - } else if constexpr (parallel_blocks > 1) { + } else if (parallel_blocks > 1) { const dim3 block_dim_combine(D, 1, 1); - const dim3 blocks_num_combine(Q->ne[1], blocks_num.y, blocks_num.z); + const dim3 blocks_num_combine(Q->ne[1], 1, blocks_num.z); + const size_t nbytes_shared_combine = parallel_blocks*sizeof(float2); - flash_attn_combine_results - <<>> - (dst_tmp.ptr, dst_tmp_meta.ptr, (float *) KQV->data); + flash_attn_combine_results + <<>> + (dst_tmp.ptr, dst_tmp_meta.ptr, (float *) KQV->data, parallel_blocks); } CUDA_CHECK(cudaGetLastError()); } diff --git a/ggml/src/ggml-cuda/fattn-mma-f16.cuh b/ggml/src/ggml-cuda/fattn-mma-f16.cuh index 05bc91a3b..04804a15c 100644 --- a/ggml/src/ggml-cuda/fattn-mma-f16.cuh +++ b/ggml/src/ggml-cuda/fattn-mma-f16.cuh @@ -1,103 +1,519 @@ #include "common.cuh" +#include "cp-async.cuh" #include "mma.cuh" #include "fattn-common.cuh" -template -static __device__ __forceinline__ void flash_attn_ext_f16_process_tile( +using namespace ggml_cuda_mma; + +typedef tile<16, 8, half2> tile_A; +typedef tile< 8, 8, half2> tile_B; +typedef tile<16, 8, half2> tile_B_16; +typedef tile<16, 8, float> tile_C_KQ; +typedef tile<16, 16, float> tile_C_KQ_16; +typedef tile<16, 4, half2> tile_C_VKQ; +typedef tile<16, 8, half2> tile_C_VKQ_16; + +template +static __device__ __forceinline__ void flash_attn_ext_f16_load_tile( + const half2 * const __restrict__ KV, half2 * const __restrict__ tile_KV, const int stride_KV) { + constexpr int D2_padded = D/2 + 4; // Size of D in half2, padded to avoid shared memory bank conflicts. + + // If cp.async is available, load up to the highest power of 2 in D asynchronously: +#ifdef CP_ASYNC_AVAILABLE + static_assert(D >= 64 && D < 512, "bad D"); + constexpr int k0_sync_start = D/2 < 64 ? 32 : (D/2 < 128 ? 64 : 128); + + const unsigned int tile_KV_32 = __cvta_generic_to_shared(tile_KV); + + constexpr int preload = 64; + constexpr int h2_per_chunk = 16/sizeof(half2); + constexpr int chunks_per_row = k0_sync_start / h2_per_chunk; + constexpr int stride_i = WARP_SIZE / chunks_per_row; +#pragma unroll + for (int i0 = 0; i0 < KQ_per_iter; i0 += nwarps*stride_i) { + const int i = i0 + threadIdx.y*stride_i + (chunks_per_row == WARP_SIZE ? 0 : threadIdx.x / chunks_per_row); + const int k = (chunks_per_row == WARP_SIZE ? threadIdx.x : threadIdx.x % chunks_per_row)*h2_per_chunk; + + cp_async_cg_16(tile_KV_32 + (i*D2_padded + k)*sizeof(half2), KV + i*stride_KV + k); + } +#else + constexpr int k0_sync_start = 0; +#endif // CP_ASYNC_AVAILABLE + static_assert(k0_sync_start % WARP_SIZE == 0, "bad k0_sync_start"); + + // If D is not a power of 2, the rest is loaded synchronously. + // K/V data is loaded with decreasing granularity for D for better memory bandwidth. + static_assert(KQ_per_iter % (4*nwarps) == 0, "out of bounds"); +#pragma unroll + for (int stride_k : {WARP_SIZE, WARP_SIZE/2, WARP_SIZE/4}) { + const int k0_start = stride_k == WARP_SIZE ? k0_sync_start : D/2 - (D/2) % (2*stride_k); + const int k0_stop = D/2 - (D/2) % (1*stride_k); + const int stride_i = WARP_SIZE / stride_k; + + if (k0_start == k0_stop || k0_stop <= k0_sync_start) { + continue; + } + +#pragma unroll + for (int i0 = 0; i0 < KQ_per_iter; i0 += nwarps*stride_i) { + const int i = i0 + threadIdx.y*stride_i + (stride_k == WARP_SIZE ? 0 : threadIdx.x / stride_k); + +#pragma unroll + for (int k0 = k0_start; k0 < k0_stop; k0 += stride_k) { + const int k = k0 + (stride_k == WARP_SIZE ? threadIdx.x : threadIdx.x % stride_k); + + tile_KV[i*D2_padded + k] = KV[i*stride_KV + k]; + } + } + } +} + +template +static __device__ __forceinline__ void flash_attn_ext_f16_load_mask( + const half2 * const __restrict__ mask_h2, half2 * const __restrict__ tile_mask, const int stride_mask) { + static_assert(KQ_per_iter == 2*WARP_SIZE || KQ_per_iter == WARP_SIZE, "bad KQ_per_iter"); +#ifdef CP_ASYNC_AVAILABLE + constexpr int preload = KQ_per_iter * sizeof(half); + constexpr int cols_per_warp = 8*WARP_SIZE/KQ_per_iter; + constexpr int stride_j = nwarps * cols_per_warp; + + const unsigned int tile_mask_32 = __cvta_generic_to_shared(tile_mask); + +#pragma unroll + for (int j0 = 0; j0 < ncols1; j0 += stride_j) { + const int j = j0 + threadIdx.y*cols_per_warp + + (KQ_per_iter == 2*WARP_SIZE ? threadIdx.x / (WARP_SIZE/4) : threadIdx.x / (WARP_SIZE/8)); + + if (j0 + stride_j > ncols1 && j >= ncols1) { + break; + } + + const int i = 4 * (KQ_per_iter == 2*WARP_SIZE ? threadIdx.x % (WARP_SIZE/4) : threadIdx.x % (WARP_SIZE/8)); + + cp_async_cg_16(tile_mask_32 + j*(KQ_per_iter*sizeof(half) + 16) + i*sizeof(half2), mask_h2 + j*stride_mask + i); + } +#else + constexpr int cols_per_warp = 2*WARP_SIZE/KQ_per_iter; + constexpr int stride_j = nwarps * cols_per_warp; +#pragma unroll + for (int j0 = 0; j0 < ncols1; j0 += stride_j) { + const int j = j0 + threadIdx.y*cols_per_warp + (KQ_per_iter == 2*WARP_SIZE ? 0 : threadIdx.x / (WARP_SIZE/2)); + + if (j0 + stride_j > ncols1 && j >= ncols1) { + break; + } + + const int i = KQ_per_iter == 2*WARP_SIZE ? threadIdx.x : threadIdx.x % (WARP_SIZE/2); + + tile_mask[j*(KQ_per_iter/2 + 4) + i] = mask_h2[j*stride_mask + i]; + } +#endif // CP_ASYNC_AVAILABLE +} + +template +static __device__ __forceinline__ void flash_attn_ext_f16_iter( const float2 * const __restrict__ Q_f2, const half2 * const __restrict__ K_h2, const half2 * const __restrict__ V_h2, - const half * const __restrict__ maskh, + const half2 * const __restrict__ mask_h2, float2 * const __restrict__ dstk, float2 * const __restrict__ dstk_fixup, const float scale, const float slope, const float logit_softcap, - const int ne00, const int ne01, const int ne02, - const int ne03, - const int ne10, - const int ne11, - const int ne12, - const int ne13, - const int ne31, - const int nb31, - const int nb01, - const int nb02, - const int nb03, - const int nb11, - const int nb12, - const int nb13, - const int nb21, - const int nb22, - const int nb23, - const int ne0, - const int ne1, - const int ne2, - const int ne3, + const int stride_KV, + const int stride_mask, + const int jt, + half2 * const __restrict__ tile_K, + half2 * const __restrict__ tile_V, + half2 * const __restrict__ tile_mask, + const tile_B * const __restrict__ Q_B, + tile_C_VKQ * const __restrict__ VKQ_C, + float * const __restrict__ KQ_max, + float * const __restrict__ KQ_rowsum, + const int kb0) { +#ifdef NEW_MMA_AVAILABLE + constexpr int cols_per_warp = ntiles * tile_B::I; + constexpr int cols_per_thread = ntiles == 1 ? 2 : ntiles; + constexpr int np = nwarps * (cols_per_warp/ncols2) / ncols1; // Number of parallel CUDA warps per Q column. + constexpr int D2_padded = D/2 + 4; // Size of D in half2, padded to avoid shared memory bank conflicts. + + const int k_VKQ_0 = kb0 * KQ_per_iter; + tile_C_KQ KQ_C[KQ_per_iter/(np*tile_C_KQ::I) * ntiles]; + + // Use wide variants of tiles if ntiles >= 2. + tile_B_16 * Q_B_16 = (tile_B_16 *) Q_B; + tile_C_VKQ_16 * VKQ_C_16 = (tile_C_VKQ_16 *) VKQ_C; + tile_C_KQ_16 * KQ_C_16 = (tile_C_KQ_16 *) KQ_C; + +#ifdef CP_ASYNC_AVAILABLE + cp_async_wait_all(); + __syncthreads(); + flash_attn_ext_f16_load_tile(V_h2 + k_VKQ_0*stride_KV, tile_V, stride_KV); +#else + if (ncols2 > 1 || mask_h2) { + flash_attn_ext_f16_load_mask(mask_h2 + k_VKQ_0/2, tile_mask, stride_mask); + } + flash_attn_ext_f16_load_tile(K_h2 + k_VKQ_0*stride_KV, tile_K, stride_KV); + __syncthreads(); +#endif // CP_ASYNC_AVAILABLE + + // Calculate tile of KQ: +#pragma unroll + for (int i_KQ_00 = 0; i_KQ_00 < KQ_per_iter; i_KQ_00 += np*tile_A::I) { + const int i_KQ_0 = i_KQ_00 + (threadIdx.y % np)*tile_A::I; +#pragma unroll + for (int k_KQ_0 = 0; k_KQ_0 < D/2; k_KQ_0 += tile_A::J) { + tile_A K_A; + load_ldmatrix(K_A, tile_K + i_KQ_0*D2_padded + k_KQ_0, D2_padded); + if (ntiles == 1) { + mma(KQ_C[i_KQ_00/(np*tile_A::I)], K_A, Q_B[k_KQ_0/tile_A::J]); + } else { +#pragma unroll + for (int t = 0; t < ntiles/2; ++t) { + // Wide version of KQ_C is column-major => swap A and B. + mma(KQ_C_16[i_KQ_00/(np*tile_A::I) * ntiles/2 + t], Q_B_16[k_KQ_0/tile_A::J * ntiles/2 + t], K_A); + } + } + } + } + +#ifndef CP_ASYNC_AVAILABLE + __syncthreads(); // Only needed if tile_K == tile_V. +#endif // CP_ASYNC_AVAILABLE + + if (use_logit_softcap) { + static_assert(KQ_per_iter % (np*tile_C_KQ::I) == 0, "bad loop size"); +#pragma unroll + for (int i = 0; i < KQ_per_iter/(np*tile_C_KQ::I) * ntiles; ++i) { +#pragma unroll + for (int l = 0; l < tile_C_KQ::ne; ++l) { + KQ_C[i].x[l] = logit_softcap*tanhf(KQ_C[i].x[l]); + } + } + } + + float KQ_max_new[cols_per_thread]; +#pragma unroll + for (int col = 0; col < cols_per_thread; ++col) { + KQ_max_new[col] = KQ_max[col]; + } + float KQ_rowsum_add[cols_per_thread] = {0.0f}; + + if (ntiles == 1) { + if (ncols2 > 1 || mask_h2) { +#pragma unroll + for (int i00 = 0; i00 < KQ_per_iter; i00 += np*tile_C_KQ::I) { + const int i0 = i00 + (threadIdx.y % np)*tile_C_KQ::I; +#pragma unroll + for (int l = 0; l < tile_C_KQ::ne; ++l) { + const int i = i0 + tile_C_KQ::get_i(l); + const int j = ((threadIdx.y / np)*tile_C_KQ::J + tile_C_KQ::get_j(l)) / ncols2; + + KQ_C[i00/(np*tile_C_KQ::I)].x[l] += slope * + __half2float(((const half *) tile_mask)[j*(KQ_per_iter + 8) + i]); + } + } + } + + // Calculate softmax for each KQ column using the current max. value. + // The divisor is stored in KQ_rowsum and will be applied at the end. + static_assert(KQ_per_iter % (np*tile_C_KQ::I) == 0, "bad loop size"); +#pragma unroll + for (int k = 0; k < KQ_per_iter/(np*tile_C_KQ::I); ++k) { +#pragma unroll + for (int l = 0; l < tile_C_KQ::ne; ++l) { + KQ_max_new[l % 2] = fmaxf(KQ_max_new[l % 2], KQ_C[k].x[l]); + } + } + + // Values per KQ column are spread across 8 threads, does not need full warp reduce: +#pragma unroll + for (int col = 0; col < cols_per_thread; ++col) { +#pragma unroll + for (int offset = 16; offset >= 4; offset >>= 1) { + KQ_max_new[col] = fmaxf(KQ_max_new[col], __shfl_xor_sync(0xFFFFFFFF, KQ_max_new[col], offset, WARP_SIZE)); + } + } + + static_assert(KQ_per_iter % (np*tile_C_KQ::I) == 0, "bad loop size"); + +#pragma unroll + for (int k = 0; k < KQ_per_iter/(np*tile_C_KQ::I); ++k) { +#pragma unroll + for (int l = 0; l < tile_C_KQ::ne; ++l) { + KQ_C[k].x[l] = expf(KQ_C[k].x[l] - KQ_max_new[l % 2]); + + KQ_rowsum_add[l % 2] += KQ_C[k].x[l]; + } + } + } else { // ntiles > 1 + if (ncols2 > 1 || mask_h2) { +#pragma unroll + for (int i00 = 0; i00 < KQ_per_iter; i00 += np*tile_C_KQ_16::J) { + const int i0 = i00 + (threadIdx.y % np)*tile_C_KQ_16::J; +#pragma unroll + for (int t = 0; t < ntiles/2; ++t) { +#pragma unroll + for (int l0 = 0; l0 < tile_C_KQ_16::ne; l0 += 2) { + const int i = (i0 + tile_C_KQ_16::get_j(l0)) / 2; + const int j = ((threadIdx.y / np)*cols_per_warp + t*tile_C_KQ_16::I + tile_C_KQ_16::get_i(l0)) / ncols2; + + const float2 tmp = __half22float2(tile_mask[j*(KQ_per_iter/2 + 4) + i]); + const int KQ_index = i00/(np*tile_C_KQ_16::J) * ntiles/2 + t; + KQ_C_16[KQ_index].x[l0 + 0] += slope*tmp.x; + KQ_C_16[KQ_index].x[l0 + 1] += slope*tmp.y; + } + } + } + } + + // Calculate softmax for each KQ column using the current max. value. + // The divisor is stored in KQ_rowsum and will be applied at the end. + static_assert(KQ_per_iter % (np*tile_C_KQ::I) == 0, "bad loop size"); +#pragma unroll + for (int k = 0; k < KQ_per_iter/(np*tile_C_KQ_16::J); ++k) { +#pragma unroll + for (int t = 0; t < ntiles/2; ++t) { +#pragma unroll + for (int l = 0; l < tile_C_KQ_16::ne; ++l) { + const int KQ_index = 2*t + (l/2) % 2; + KQ_max_new[KQ_index] = fmaxf(KQ_max_new[KQ_index], KQ_C_16[k*ntiles/2 + t].x[l]); + } + } + } + + // Values per KQ column are spread across 4 threads, does not need full warp reduce: +#pragma unroll + for (int col = 0; col < cols_per_thread; ++col) { +#pragma unroll + for (int offset = 2; offset >= 1; offset >>= 1) { + KQ_max_new[col] = fmaxf(KQ_max_new[col], __shfl_xor_sync(0xFFFFFFFF, KQ_max_new[col], offset, WARP_SIZE)); + } + } + + static_assert(KQ_per_iter % (np*tile_C_KQ_16::J) == 0, "bad loop size"); +#pragma unroll + for (int k = 0; k < KQ_per_iter/(np*tile_C_KQ_16::J); ++k) { +#pragma unroll + for (int t = 0; t < ntiles/2; ++t) { +#pragma unroll + for (int l = 0; l < tile_C_KQ_16::ne; ++l) { + const int KQ_index = 2*t + (l/2) % 2; + + KQ_C_16[k*ntiles/2 + t].x[l] = expf(KQ_C_16[k*ntiles/2 + t].x[l] - KQ_max_new[KQ_index]); + + KQ_rowsum_add[KQ_index] += KQ_C_16[k*ntiles/2 + t].x[l]; + } + } + } + } + + { + float KQ_max_scale[cols_per_thread]; +#pragma unroll + for (int col = 0; col < cols_per_thread; ++col) { + KQ_max_scale[col] = expf(KQ_max[col] - KQ_max_new[col]); + KQ_max[col] = KQ_max_new[col]; + + // Scale previous KQ_rowsum to account for a potential increase in KQ_max: + KQ_rowsum[col] = KQ_max_scale[col]*KQ_rowsum[col] + KQ_rowsum_add[col]; + } + + if (ntiles == 1) { + const half2 KQ_max_scale_h2 = make_half2(KQ_max_scale[0], KQ_max_scale[1]); +#pragma unroll + for (int i = 0; i < D/tile_C_VKQ::I; ++i) { +#pragma unroll + for (int l = 0; l < tile_C_VKQ::ne; ++l) { + VKQ_C[i].x[l] *= KQ_max_scale_h2; + } + } + } else { +#pragma unroll + for (int col = 0; col < cols_per_thread; ++col) { + const half2 KQ_max_scale_h2 = make_half2(KQ_max_scale[col], KQ_max_scale[col]); +#pragma unroll + for (int i = 0; i < D/tile_C_VKQ_16::J; ++i) { +#pragma unroll + for (int l0 = 0; l0 < tile_C_VKQ_16::ne; l0 += 2) { + VKQ_C_16[i*ntiles/2 + col/2].x[l0 + col % 2] *= KQ_max_scale_h2; + } + } + } + } + } + + // Convert KQ C tiles into B tiles for VKQ calculation: + tile_B B[KQ_per_iter/(np*2*tile_B::J) * ntiles]; + tile_B_16 * B_16 = (tile_B_16 *) B; + static_assert(KQ_per_iter % (np*2*tile_B::J) == 0, "bad loop size"); + if (ntiles == 1) { +#pragma unroll + for (int k = 0; k < KQ_per_iter/(np*2*tile_B::J); ++k) { + B[k] = get_transposed(get_half2(KQ_C[k])); + } + } else { + for (int k = 0; k < KQ_per_iter/(np*2*tile_B_16::J); ++k) { +#pragma unroll + for (int t = 0; t < ntiles/2; ++t) { + B_16[k*ntiles/2 + t] = get_half2(KQ_C_16[k*ntiles/2 + t]); + } + } + } + +#ifdef CP_ASYNC_AVAILABLE + // Preload K tile for next iteration: + cp_async_wait_all(); + __syncthreads(); + if (!last_iter) { + if (ncols2 > 1 || mask_h2) { + flash_attn_ext_f16_load_mask(mask_h2 + (k_VKQ_0 + KQ_per_iter)/2, tile_mask, stride_mask); + } + flash_attn_ext_f16_load_tile(K_h2 + (k_VKQ_0 + KQ_per_iter)*stride_KV, tile_K, stride_KV); + } +#else + flash_attn_ext_f16_load_tile(V_h2 + k_VKQ_0*stride_KV, tile_V, stride_KV); + __syncthreads(); +#endif // CP_ASYNC_AVAILABLE + + // Calculate VKQ tile: +#pragma unroll + for (int i_VKQ_0 = 0; i_VKQ_0 < D; i_VKQ_0 += tile_C_VKQ::I) { + static_assert((KQ_per_iter/2) % (np*tile_A::J) == 0, "bad loop size"); +#pragma unroll + for (int k00 = 0; k00 < KQ_per_iter/2; k00 += np*tile_A::J) { + const int k0 = k00 + (threadIdx.y % np)*tile_A::J; + + tile_A A; + load_ldmatrix_trans(A, tile_V + 2*k0*D2_padded + i_VKQ_0/2, D2_padded); + if (ntiles == 1) { + mma(VKQ_C[i_VKQ_0/tile_C_VKQ::I], A, B[k00/(np*tile_A::J)]); + } else { +#pragma unroll + for (int t = 0; t < ntiles/2; ++t) { + // Wide version of VKQ_C is column-major => swap A and B. + mma(VKQ_C_16[i_VKQ_0/tile_C_VKQ::I * ntiles/2 + t], B_16[k00/(np*tile_A::J) * ntiles/2 + t], A); + } + } + } + } + +#ifndef CP_ASYNC_AVAILABLE + __syncthreads(); // Only needed if tile_K == tile_V. +#endif // CP_ASYNC_AVAILABLE + +#else + GGML_UNUSED(Q_f2); GGML_UNUSED(K_h2); GGML_UNUSED(V_h2); + GGML_UNUSED(mask_h2); GGML_UNUSED(dstk); GGML_UNUSED(dstk_fixup); + GGML_UNUSED(scale); GGML_UNUSED(slope); GGML_UNUSED(logit_softcap); + GGML_UNUSED(ne01); GGML_UNUSED(ne02); GGML_UNUSED(stride_KV); + GGML_UNUSED(stride_mask); GGML_UNUSED(jt); GGML_UNUSED(tile_K); + GGML_UNUSED(stride_mask); GGML_UNUSED(jt); GGML_UNUSED(tile_K); + GGML_UNUSED(tile_V); GGML_UNUSED(tile_mask); GGML_UNUSED(Q_B); + GGML_UNUSED(VKQ_C); GGML_UNUSED(KQ_max); GGML_UNUSED(KQ_rowsum); + GGML_UNUSED(kb0); + NO_DEVICE_CODE; +#endif // NEW_MMA_AVAILABLE +} + +template +static __device__ __forceinline__ void flash_attn_ext_f16_process_tile( + const float2 * const __restrict__ Q_f2, + const half2 * const __restrict__ K_h2, + const half2 * const __restrict__ V_h2, + const half2 * const __restrict__ mask_h2, + float2 * const __restrict__ dstk, + float2 * const __restrict__ dstk_fixup, + const float scale, + const float slope, + const float logit_softcap, + const int ne01, + const int ne02, + const int stride_Q1, + const int stride_Q2, + const int stride_KV, + const int stride_mask, const int jt, const int kb0_start, const int kb0_stop) { #ifdef NEW_MMA_AVAILABLE //In this kernel Q, K, V are matrices while i, j, k are matrix indices. - typedef mma_A_I16K8 mma_A; - typedef mma_B_J8K8 mma_B; - typedef mma_C_I16J8 mma_C_KQ; - typedef mma_C_I16J8 mma_C_VKQ; + constexpr int ncols = ncols1 * ncols2; + constexpr int cols_per_warp = ntiles * tile_B::I; + constexpr int cols_per_thread = ntiles == 1 ? 2 : ntiles; + constexpr int np = nwarps * (cols_per_warp/ncols2) / ncols1; // Number of parallel CUDA warps per Q column. - static_assert(nwarps*mma_B::J % ncols == 0, "bad nwarps"); - constexpr int np = nwarps*mma_B::J / ncols; // Number of parallel CUDA warps per Q column. + static_assert(nwarps * (cols_per_warp/ncols2) % ncols1 == 0, "bad nwarps"); - static_assert(D % nwarps == 0, "bad D"); - static_assert(KQ_stride % nwarps == 0, "bad KQ_stride"); + static_assert(D % nwarps == 0, "bad D"); + static_assert(KQ_per_iter % nwarps == 0, "bad KQ_per_iter"); constexpr int D2_padded = D/2 + 4; // Size of D in half2, padded to avoid shared memory bank conflicts. - extern __shared__ half2 tile_KV[]; // Temporary shared buffer for loading K/V data with KQ_stride*D logical elements. - const int stride_Q = nb01 / sizeof(float2); - const int stride_KV = nb11 / sizeof(half2); - const int stride_mask = nb31 / sizeof(half); + // Temporary shared buffer for loading K/V data with KQ_per_iter*D logical elements: + extern __shared__ half2 tile_K[]; +#ifdef CP_ASYNC_AVAILABLE + half2 * tile_V = tile_K + KQ_per_iter*D2_padded; +#else + half2 * tile_V = tile_K; +#endif // CP_ASYNC_AVAILABLE + half2 * tile_mask = tile_V + KQ_per_iter*D2_padded; - mma_B Q_B[D/(2*mma_B::K)]; - mma_C_VKQ VKQ_C[D/mma_C_VKQ::I]; + tile_B Q_B[D/(2*tile_B::J) * ntiles]; + tile_C_VKQ VKQ_C[D/tile_C_VKQ::I * ntiles]; - float2 KQ_rowsum = {0.0f, 0.0f}; - float2 KQ_max = {-FLT_MAX/2.0f, -FLT_MAX/2.0f}; - float2 KQ_max_scale = {0.0f, 0.0f}; + tile_B_16 * Q_B_16 = (tile_B_16 *) Q_B; + tile_C_VKQ_16 * VKQ_C_16 = (tile_C_VKQ_16 *) VKQ_C; - // Temporarily load Q data into tile_KV, will be loaded into registers afterwards. + float KQ_rowsum[cols_per_thread] = {0.0f}; + float KQ_max[cols_per_thread]; +#pragma unroll + for (int col = 0; col < cols_per_thread; ++col) { + KQ_max[col] = -FLT_MAX/2.0f; + } + + // Temporarily load Q data into tile_K, will be loaded into registers afterwards. // The loading is done with decreasing granularity for D for better memory bandwidth. const half2 scale_h2 = make_half2(scale, scale); #pragma unroll for (int stride_k : {WARP_SIZE, WARP_SIZE/2, WARP_SIZE/4}) { - const int k0_start = stride_k == WARP_SIZE ? 0 : D/2 - (D/2) % (2*stride_k); - const int k0_stop = D/2 - (D/2) % (1*stride_k); - const int stride_j = WARP_SIZE / stride_k; + const int k0_start = stride_k == WARP_SIZE ? 0 : D/2 - (D/2) % (2*stride_k); + const int k0_stop = D/2 - (D/2) % (1*stride_k); + const int stride_jc = WARP_SIZE / stride_k; - if (nwarps*stride_j > ncols && threadIdx.y*stride_j >= ncols) { - break; + if (k0_start == k0_stop) { + continue; } #pragma unroll - for (int j0 = 0; j0 < ncols; j0 += nwarps*stride_j) { - const int j = j0 + threadIdx.y*stride_j + (stride_k == WARP_SIZE ? 0 : threadIdx.x / stride_k); + for (int jc0 = 0; jc0 < ncols; jc0 += nwarps*stride_jc) { + const int jc = jc0 + threadIdx.y*stride_jc + (stride_k == WARP_SIZE ? 0 : threadIdx.x / stride_k); - if (jt*ncols + j < ne01) { + if (jc0 + nwarps*stride_jc > ncols && jc >= ncols) { + break; + } + + const int j = jc / ncols2; + const int c = jc % ncols2; + + if (jt*ncols1 + j < ne01) { #pragma unroll for (int k0 = k0_start; k0 < k0_stop; k0 += stride_k) { const int k = k0 + (stride_k == WARP_SIZE ? threadIdx.x : threadIdx.x % stride_k); - const float2 tmp = Q_f2[(jt*ncols + j)*stride_Q + k]; - tile_KV[j*D2_padded + k] = scale_h2 * make_half2(tmp.x, tmp.y); + const float2 tmp = Q_f2[(jt*ncols1 + j)*stride_Q1 + c*stride_Q2 + k]; + tile_K[jc*D2_padded + k] = scale_h2 * make_half2(tmp.x, tmp.y); } } else { #pragma unroll for (int k0 = k0_start; k0 < k0_stop; k0 += stride_k) { const int k = k0 + (stride_k == WARP_SIZE ? threadIdx.x : threadIdx.x % stride_k); - tile_KV[j*D2_padded + k] = make_half2(0.0f, 0.0f); + tile_K[jc*D2_padded + k] = make_half2(0.0f, 0.0f); } } } @@ -106,286 +522,217 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile( __syncthreads(); { - const int j0 = (threadIdx.y / np) * mma_B::J; + const int j0 = (threadIdx.y / np) * cols_per_warp; #pragma unroll - for (int k0 = 0; k0 < D/2; k0 += mma_B::K) { - Q_B[k0/mma_B::K].load_ldmatrix(tile_KV + j0*D2_padded + k0, D2_padded); + for (int k0 = 0; k0 < D/2; k0 += tile_B::J) { + if (ntiles == 1) { + load_ldmatrix(Q_B[k0/tile_B::J], tile_K + j0*D2_padded + k0, D2_padded); + } else { +#pragma unroll + for (int t = 0; t < ntiles/2; ++t) { + load_ldmatrix(Q_B_16[k0/tile_B_16::J * ntiles/2 + t], + tile_K + (j0 + t*tile_B_16::I)*D2_padded + k0, D2_padded); + } + } } } __syncthreads(); + // Preload mask and K data for first iteration when using cp_async: +#ifdef CP_ASYNC_AVAILABLE + if (ncols2 > 1 || mask_h2) { + flash_attn_ext_f16_load_mask(mask_h2 + kb0_start*KQ_per_iter/2, tile_mask, stride_mask); + } + flash_attn_ext_f16_load_tile(K_h2 + kb0_start*KQ_per_iter*stride_KV, tile_K, stride_KV); +#endif // CP_ASYNC_AVAILABLE + // Iterate over ne11 == previous tokens: - for (int kb0 = kb0_start; kb0 < kb0_stop; ++kb0) { - const int k_VKQ_0 = kb0*KQ_stride; - mma_C_KQ KQ_C[KQ_stride/(np*mma_C_KQ::I)]; - - // Load K data into tile with decreasing granularity for D for better memory bandwidth: - static_assert(KQ_stride % (4*nwarps) == 0, "out of bounds"); -#pragma unroll - for (int stride_k : {WARP_SIZE, WARP_SIZE/2, WARP_SIZE/4}) { - const int k0_start = stride_k == WARP_SIZE ? 0 : D/2 - (D/2) % (2*stride_k); - const int k0_stop = D/2 - (D/2) % (1*stride_k); - const int stride_i = WARP_SIZE / stride_k; - -#pragma unroll - for (int i_KQ_0 = 0; i_KQ_0 < KQ_stride; i_KQ_0 += nwarps*stride_i) { - const int i_KQ = i_KQ_0 + threadIdx.y*stride_i + (stride_k == WARP_SIZE ? 0 : threadIdx.x / stride_k); - -#pragma unroll - for (int k_KQ_0 = k0_start; k_KQ_0 < k0_stop; k_KQ_0 += stride_k) { - const int k_KQ = k_KQ_0 + (stride_k == WARP_SIZE ? threadIdx.x : threadIdx.x % stride_k); - - tile_KV[i_KQ*D2_padded + k_KQ] = K_h2[(k_VKQ_0 + i_KQ)*stride_KV + k_KQ]; - } - } - } - - __syncthreads(); - - // Calculate tile of KQ: -#pragma unroll - for (int i_KQ_00 = 0; i_KQ_00 < KQ_stride; i_KQ_00 += np*mma_A::I) { - const int i_KQ_0 = i_KQ_00 + (threadIdx.y % np)*mma_A::I; -#pragma unroll - for (int k_KQ_0 = 0; k_KQ_0 < D/2; k_KQ_0 += mma_A::K) { - mma_A K_A; - K_A.load_ldmatrix(tile_KV + i_KQ_0*D2_padded + k_KQ_0, D2_padded); - KQ_C[i_KQ_00/(np*mma_A::I)].mma(K_A, Q_B[k_KQ_0/mma_A::K]); - } - } - - __syncthreads(); - - if (use_logit_softcap) { - static_assert(KQ_stride % (np*mma_C_KQ::I) == 0, "bad loop size"); -#pragma unroll - for (int i = 0; i < KQ_stride/(np*mma_C_KQ::I); ++i) { -#pragma unroll - for (int l = 0; l < mma_C_KQ::ne; ++l) { - KQ_C[i].x[l] = logit_softcap*tanhf(KQ_C[i].x[l]); - } - } - } - - if (maskh) { - static_assert(KQ_stride % (np *mma_C_KQ::I) == 0, "bad loop size"); - static_assert(ncols % (nwarps/np*mma_C_KQ::J) == 0, "bad loop size"); -#pragma unroll - for (int i00 = 0; i00 < KQ_stride; i00 += np*mma_C_KQ::I) { - const int i0 = i00 + (threadIdx.y % np)*mma_C_KQ::I; -#pragma unroll - for (int l = 0; l < mma_C_KQ::ne; ++l) { - const int i = i0 + mma_C_KQ::get_i(l); - const int j = (threadIdx.y / np)*mma_C_KQ::J + mma_C_KQ::get_j(l); - - KQ_C[i00/(np*mma_C_KQ::I)].x[l] += slope*__half2float(maskh[j*stride_mask + k_VKQ_0 + i]); - } - } - } - - // Calculate softmax for each KQ column using the current max. value. - // The divisor is stored in KQ_rowsum and will be applied at the end. - float2 KQ_max_new = KQ_max; - static_assert(KQ_stride % (np*mma_C_KQ::I) == 0, "bad loop size"); -#pragma unroll - for (int k = 0; k < KQ_stride/(np*mma_C_KQ::I); ++k) { -#pragma unroll - for (int l0 = 0; l0 < mma_C_KQ::ne; l0 += 2) { - KQ_max_new.x = fmaxf(KQ_max_new.x, KQ_C[k].x[l0 + 0]); - KQ_max_new.y = fmaxf(KQ_max_new.y, KQ_C[k].x[l0 + 1]); - } - } - - // Values per KQ column are spread across 8 threads, does not need full warp reduce: -#pragma unroll - for (int offset = 16; offset > 2; offset >>= 1) { - KQ_max_new.x = fmaxf(KQ_max_new.x, __shfl_xor_sync(0xFFFFFFFF, KQ_max_new.x, offset, WARP_SIZE)); - KQ_max_new.y = fmaxf(KQ_max_new.y, __shfl_xor_sync(0xFFFFFFFF, KQ_max_new.y, offset, WARP_SIZE)); - } - - { - const float2 diff = make_float2(KQ_max.x - KQ_max_new.x, KQ_max.y - KQ_max_new.y); - KQ_max_scale = make_float2(expf(diff.x), expf(diff.y)); - if (diff.x <= SOFTMAX_FTZ_THRESHOLD) { - KQ_max_scale.x = 0.0f; - } - if (diff.y <= SOFTMAX_FTZ_THRESHOLD) { - KQ_max_scale.y = 0.0f; - } - KQ_max = KQ_max_new; - } - - float2 KQ_rowsum_add = make_float2(0.0f, 0.0f); - static_assert(KQ_stride % (np*mma_C_KQ::I) == 0, "bad loop size"); -#pragma unroll - for (int k = 0; k < KQ_stride/(np*mma_C_KQ::I); ++k) { -#pragma unroll - for (int l = 0; l < mma_C_KQ::ne; ++l) { - const float KQ_max_l = l % 2 == 0 ? KQ_max.x : KQ_max.y; - const float diff = KQ_C[k].x[l] - KQ_max_l; - KQ_C[k].x[l] = expf(diff); - if (diff <= SOFTMAX_FTZ_THRESHOLD) { - KQ_C[k].x[l] = 0.0f; - } - - if (l % 2 == 0) { - KQ_rowsum_add.x += KQ_C[k].x[l]; - } else { - KQ_rowsum_add.y += KQ_C[k].x[l]; - } - } - } - - // Scale previous KQ_rowsum to account for a potential increase in KQ_max: - KQ_rowsum.x = KQ_max_scale.x*KQ_rowsum.x + KQ_rowsum_add.x; - KQ_rowsum.y = KQ_max_scale.y*KQ_rowsum.y + KQ_rowsum_add.y; - - const half2 KQ_max_scale_h2 = make_half2(KQ_max_scale.x, KQ_max_scale.y); -#pragma unroll - for (int i = 0; i < D/mma_C_VKQ::I; ++i) { -#pragma unroll - for (int l = 0; l < mma_C_VKQ::ne; ++l) { - VKQ_C[i].x[l] *= KQ_max_scale_h2; - } - } - - // Convert KQ C tiles into B tiles for VKQ calculation: - mma_B B[KQ_stride/(np*2*mma_B::K)]; - static_assert(KQ_stride % (np*2*mma_B::K) == 0, "bad loop size"); -#pragma unroll - for (int k = 0; k < KQ_stride/(np*2*mma_B::K); ++k) { - B[k] = KQ_C[k].to_mma_B(); - } - - // Load V data into tile with decreasing granularity for D for better memory bandwidth: - static_assert(KQ_stride % (4*nwarps) == 0, "out of bounds"); -#pragma unroll - for (int stride_i : {WARP_SIZE, WARP_SIZE/2, WARP_SIZE/4}) { - const int i0_start = stride_i == WARP_SIZE ? 0 : D/2 - (D/2) % (2*stride_i); - const int i0_stop = D/2 - (D/2) % (1*stride_i); - const int stride_k = WARP_SIZE / stride_i; - -#pragma unroll - for (int k_V_0 = 0; k_V_0 < KQ_stride; k_V_0 += nwarps*stride_k) { - const int k_V = k_V_0 + threadIdx.y*stride_k + (stride_i == WARP_SIZE ? 0 : threadIdx.x / stride_i); - -#pragma unroll - for (int i_V_0 = i0_start; i_V_0 < i0_stop; i_V_0 += stride_i) { - const int i_V = i_V_0 + (stride_i == WARP_SIZE ? threadIdx.x : threadIdx.x % stride_i); - - tile_KV[k_V*D2_padded + i_V] = V_h2[(k_VKQ_0 + k_V)*stride_KV + i_V]; - } - } - } - - __syncthreads(); - - // Calculate VKQ tile: -#pragma unroll - for (int i_VKQ_0 = 0; i_VKQ_0 < D; i_VKQ_0 += mma_C_VKQ::I) { - static_assert((KQ_stride/2) % (np*mma_A::K) == 0, "bad loop size"); -#pragma unroll - for (int k00 = 0; k00 < KQ_stride/2; k00 += np*mma_A::K) { - const int k0 = k00 + (threadIdx.y % np)*mma_A::K; - - mma_A A; - A.load_ldmatrix_trans(tile_KV + 2*k0*D2_padded + i_VKQ_0/2, D2_padded); - VKQ_C[i_VKQ_0/mma_C_VKQ::I].mma(A, B[k00/(np*mma_A::K)]); - } - } - - __syncthreads(); + for (int kb0 = kb0_start; kb0 < kb0_stop-1; ++kb0) { + constexpr bool last_iter = false; + flash_attn_ext_f16_iter + (Q_f2, K_h2, V_h2, mask_h2, dstk, dstk_fixup, scale, slope, logit_softcap, + ne01, ne02, stride_KV, stride_mask, jt, tile_K, tile_V, tile_mask, Q_B, VKQ_C, KQ_max, KQ_rowsum, kb0); + } + { // kb0_start is always < kb0_stop so the last iter can be executed unconditionally. + constexpr bool last_iter = true; + flash_attn_ext_f16_iter + (Q_f2, K_h2, V_h2, mask_h2, dstk, dstk_fixup, scale, slope, logit_softcap, + ne01, ne02, stride_KV, stride_mask, jt, tile_K, tile_V, tile_mask, Q_B, VKQ_C, KQ_max, KQ_rowsum, kb0_stop-1); } + // With cp_async there is no __syncthreads at the end of the iter, + // there can be a race condition on shared memory access for combining/writing back results. +#ifdef CP_ASYNC_AVAILABLE + if (nwarps*cols_per_warp > KQ_per_iter) { + __syncthreads(); + } +#endif // CP_ASYNC_AVAILABLE + // Finally, sum up partial KQ rowsums. - // The partial sums are spread across 8 threads each, does not need full reduce. + // The partial sums are spread across 8/4 threads each, does not need full reduce. + { + constexpr int offset_first = ntiles == 1 ? 16 : 2; + constexpr int offset_last = ntiles == 1 ? 4 : 1; #pragma unroll - for (int offset = 16; offset > 2; offset >>= 1) { - KQ_rowsum.x += __shfl_xor_sync(0xFFFFFFFF, KQ_rowsum.x, offset, WARP_SIZE); - KQ_rowsum.y += __shfl_xor_sync(0xFFFFFFFF, KQ_rowsum.y, offset, WARP_SIZE); + for (int col = 0; col < cols_per_thread; ++col) { +#pragma unroll + for (int offset = offset_first; offset >= offset_last; offset >>= 1) { + KQ_rowsum[col] += __shfl_xor_sync(0xFFFFFFFF, KQ_rowsum[col], offset, WARP_SIZE); + } + } } // Write VKQ accumulators to shared memory in column-major format. // It's faster to do small writes to shared memory, then large write to VRAM than to do small writes to VRAM. // Also for np > 1 the combination is done via these values in shared memory. - const int j_cwd = threadIdx.y*mma_B::J + mma_B::get_j(-1); // j combine write data + if (ntiles == 1) { + const int jc_cwd = threadIdx.y*tile_B::I + tile_B::get_i(-1); // jc combine write data #pragma unroll - for (int k0 = 0; k0 < D/2; k0 += mma_B::K) { - const mma_B B = VKQ_C[k0/mma_B::K].to_mma_B(); // Conversion of C to B matrix puts it in column-major format. + for (int k0 = 0; k0 < D/2; k0 += tile_B::J) { + const tile_B B = get_transposed(VKQ_C[k0/tile_B::J]); // Conversion of C to B matrix puts it in column-major format. #pragma unroll - for (int l = 0; l < mma_B::ne; ++l) { - const int k = k0 + mma_B::get_k(l); + for (int l = 0; l < tile_B::ne; ++l) { + const int k = k0 + tile_B::get_j(l); - tile_KV[j_cwd*D2_padded + k] = B.x[l]; + tile_K[jc_cwd*D2_padded + k] = B.x[l]; + } + } + } else { +#pragma unroll + for (int t = 0; t < ntiles/2; ++t) { + const int j0 = threadIdx.y*cols_per_warp + t*tile_C_VKQ_16::I; +#pragma unroll + for (int k0 = 0; k0 < D/2; k0 += tile_C_VKQ_16::J) { +#pragma unroll + for (int l = 0; l < tile_C_VKQ_16::ne; ++l) { + const int j = j0 + tile_C_VKQ_16::get_i(l); + const int k = k0 + tile_C_VKQ_16::get_j(l); + + tile_K[j*D2_padded + k] = VKQ_C_16[k0/tile_C_VKQ_16::J * ntiles/2 + t].x[l]; + } + } } } - const int j_cwmo = (threadIdx.x % (2*mma_C_VKQ::J)) / mma_C_VKQ::J; // j combine write meta offset - const int j_cwm = threadIdx.y*(2*mma_C_VKQ::J) + 2*mma_C_VKQ::get_j(-1) + j_cwmo; // j combine write meta - const float2 KQ_cmr = make_float2(((const float *) &KQ_max)[j_cwmo], ((const float *) &KQ_rowsum)[j_cwmo]); // KQ combine max rowsum + if constexpr (ntiles == 1) { + const int jc_cwmo = (threadIdx.x % (2*tile_C_VKQ::J)) / tile_C_VKQ::J; // jc combine write meta offset + const int jc_cwm = threadIdx.y*(2*tile_C_VKQ::J) + 2*tile_C_VKQ::get_j(-1) + jc_cwmo; // jc combine write meta + const float2 KQ_cmr = make_float2(KQ_max[jc_cwmo], KQ_rowsum[jc_cwmo]); // KQ combine max rowsum - if (((!needs_fixup && !is_fixup) || np > 1) && threadIdx.x < 2*mma_C_VKQ::J) { - // Use the 16 bytes of padding in each row to store the meta data: KQ max, KQ rowsum, KQ max scale. - ((float2 *) tile_KV)[j_cwm*(D2_padded/2) + D/4] = KQ_cmr; + if (((!needs_fixup && !is_fixup) || np > 1) && threadIdx.x < 2*tile_C_VKQ::J) { + // Use the 16 bytes of padding in each row to store the meta data: KQ max, KQ rowsum, KQ max scale. + ((float2 *) tile_K)[jc_cwm*(D2_padded/2) + D/4] = KQ_cmr; + } + + __syncthreads(); + + if (np == 1) { + // No combination is needed, the meta data can be directly written from registers to VRAM. + if (needs_fixup && threadIdx.x < tile_B::I) { + float2 * dstk_fixup_meta = dstk_fixup + blockIdx.x*ncols; + dstk_fixup_meta[jc_cwm] = KQ_cmr; + } + if (is_fixup && threadIdx.x < tile_B::I) { + float2 * dstk_fixup_meta = dstk_fixup + (gridDim.x + blockIdx.x)*ncols; + dstk_fixup_meta[jc_cwm] = KQ_cmr; + } + } + } else { + static_assert(ntiles == 2 || ntiles == 4, "bad ntiles"); + const int jc_cwm = threadIdx.y*cols_per_warp // jc combine write meta + + (ntiles == 4 ? ((threadIdx.x % 4) / 2) * tile_C_VKQ_16::I : 0) + + tile_C_VKQ_16::get_i(threadIdx.x % 4); + const float2 KQ_cmr = make_float2(KQ_max[threadIdx.x % cols_per_thread], KQ_rowsum[threadIdx.x % cols_per_thread]); // KQ combine max rowsum + + if (((!needs_fixup && !is_fixup) || np > 1) && (ntiles == 4 || threadIdx.x % 4 < cols_per_thread)) { + // Use the 16 bytes of padding in each row to store the meta data: KQ max, KQ rowsum, KQ max scale. + ((float2 *) tile_K)[jc_cwm*(D2_padded/2) + D/4] = KQ_cmr; + } + + __syncthreads(); + + if (np == 1) { + // No combination is needed, the meta data can be directly written from registers to VRAM. + if (needs_fixup && (ntiles == 4 || threadIdx.x % 4 < ntiles)) { + float2 * dstk_fixup_meta = dstk_fixup + blockIdx.x*ncols; + dstk_fixup_meta[jc_cwm] = KQ_cmr; + } + if (is_fixup && (ntiles == 4 || threadIdx.x % 4 < ntiles)) { + float2 * dstk_fixup_meta = dstk_fixup + (gridDim.x + blockIdx.x)*ncols; + dstk_fixup_meta[jc_cwm] = KQ_cmr; + } + } } - __syncthreads(); - - static_assert(np == 1 || np == 2 || np == 4, "bad np"); - if (np == 1) { - // No combination is needed, the meta data can be directly written from registers to VRAM. - if (needs_fixup && threadIdx.x < mma_B::J) { - float2 * dstk_fixup_meta = dstk_fixup + blockIdx.x*ncols; - dstk_fixup_meta[j_cwm] = KQ_cmr; - } - if (is_fixup && threadIdx.x < mma_B::J) { - float2 * dstk_fixup_meta = dstk_fixup + (gridDim.x + blockIdx.x)*ncols; - dstk_fixup_meta[j_cwm] = KQ_cmr; - } - } else if (threadIdx.y % np == 0) { + static_assert(np == 1 || ntiles == 1 || ntiles == 2, "bad ntiles"); + if (np > 1 && threadIdx.y % np == 0) { // Combine the meta data for parallel warps via shared memory. // Warps with threadIdx.y % np != 0 must NOT return early. // All threads must return simultaneously to avoid race conditions with work on the next tile. - float * meta_j = (float *) tile_KV + (threadIdx.y*mma_B::J + threadIdx.x)*D2_padded + D/2; + constexpr int nmeta = np*cols_per_warp >= WARP_SIZE ? np*cols_per_warp/WARP_SIZE : 1; - float KQ_cm = -FLT_MAX/2; // KQ combine max per parallel warp. - if (np*mma_B::J == WARP_SIZE || threadIdx.x < np*mma_B::J) { - KQ_cm = meta_j[0]; + const int jc_meta = threadIdx.y*cols_per_warp + (np*cols_per_warp < WARP_SIZE ? threadIdx.x % (np*cols_per_warp) : threadIdx.x); + float2 * const meta_ptr = ((float2 *) tile_K) + jc_meta*(D2_padded/2) + D/4; + float2 meta[nmeta]; +#pragma unroll + for (int imeta = 0; imeta < nmeta; ++imeta) { + meta[imeta] = meta_ptr[imeta * WARP_SIZE * D2_padded/2]; } - float KQ_cmn = KQ_cm; // KQ combine max new, max between all parallel warps. + float KQ_cmn = meta[0].x; // KQ combine max new, max between all parallel warps. #pragma unroll - for (int offset = np*mma_B::J/2; offset >= mma_B::J; offset >>= 1) { + for (int imeta = 1; imeta < nmeta; ++imeta) { + KQ_cmn = fmaxf(KQ_cmn, meta[imeta].x); + } +#pragma unroll + for (int offset = np*cols_per_warp/2; offset >= cols_per_warp; offset >>= 1) { + if (offset >= WARP_SIZE) { + continue; + } KQ_cmn = fmaxf(KQ_cmn, __shfl_xor_sync(0xFFFFFFFF, KQ_cmn, offset, WARP_SIZE)); } - const float KQ_cms = expf(KQ_cm - KQ_cmn); // KQ combine max scale per warp. - float KQ_crs = 0.0f; // KQ combine rowsum, scaled sum of all parallel warps. - if (np*mma_B::J == WARP_SIZE || threadIdx.x < np*mma_B::J) { - KQ_crs = KQ_cms*meta_j[1]; + float KQ_cms[nmeta]; // KQ combine max scale per warp. +#pragma unroll + for (int imeta = 0; imeta < nmeta; ++imeta) { + KQ_cms[imeta] = expf(meta[imeta].x - KQ_cmn); + } + + float KQ_crs = KQ_cms[0]*meta[0].y; // KQ combine rowsum, scaled sum of all parallel warps. +#pragma unroll + for (int imeta = 1; imeta < nmeta; ++imeta) { + KQ_crs += KQ_cms[imeta]*meta[imeta].y; } #pragma unroll - for (int offset = np*mma_B::J/2; offset >= mma_B::J; offset >>= 1) { + for (int offset = np*cols_per_warp/2; offset >= cols_per_warp; offset >>= 1) { + if (offset >= WARP_SIZE) { + continue; + } KQ_crs += __shfl_xor_sync(0xFFFFFFFF, KQ_crs, offset, WARP_SIZE); } // Write back combined meta data: - if (np*mma_B::J == WARP_SIZE || threadIdx.x < np*mma_B::J) { - meta_j[0] = KQ_cmn; // Combined max. KQ values. - meta_j[1] = KQ_crs; // Combined KQ rowsums. - meta_j[2] = KQ_cms; // KQ max scales per parallel warp. +#pragma unroll + for (int imeta = 0; imeta < nmeta; ++imeta) { + if (np*cols_per_warp >= WARP_SIZE || threadIdx.x < np*cols_per_warp) { + // Combined KQ max scale + rowsum. + meta_ptr[imeta * WARP_SIZE * D2_padded/2] = make_float2(KQ_cms[imeta], KQ_crs); + } } - if (needs_fixup && threadIdx.x < mma_B::J) { + + // Combined KQ max + rowsum. + static_assert(cols_per_warp <= WARP_SIZE); + if (needs_fixup && (cols_per_warp == WARP_SIZE || threadIdx.x < cols_per_warp)) { float2 * dstk_fixup_meta = dstk_fixup + blockIdx.x*ncols; - dstk_fixup_meta[(threadIdx.y/np)*mma_B::J + threadIdx.x] = make_float2(KQ_cmn, KQ_crs); + dstk_fixup_meta[(threadIdx.y/np)*cols_per_warp + threadIdx.x] = make_float2(KQ_cmn, KQ_crs); } - if (is_fixup && threadIdx.x < mma_B::J) { + if (is_fixup && (cols_per_warp == WARP_SIZE || threadIdx.x < cols_per_warp)) { float2 * dstk_fixup_meta = dstk_fixup + (gridDim.x + blockIdx.x)*ncols; - dstk_fixup_meta[(threadIdx.y/np)*mma_B::J + threadIdx.x] = make_float2(KQ_cmn, KQ_crs); + dstk_fixup_meta[(threadIdx.y/np)*cols_per_warp + threadIdx.x] = make_float2(KQ_cmn, KQ_crs); } } @@ -400,23 +747,32 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile( #pragma unroll for (int stride_k : {WARP_SIZE, WARP_SIZE/2, WARP_SIZE/4}) { - const int k0_start = stride_k == WARP_SIZE ? 0 : D/2 - (D/2) % (2*stride_k); - const int k0_stop = D/2 - (D/2) % (1*stride_k); - const int stride_j = WARP_SIZE / stride_k; + const int k0_start = stride_k == WARP_SIZE ? 0 : D/2 - (D/2) % (2*stride_k); + const int k0_stop = D/2 - (D/2) % (1*stride_k); + const int stride_jc = WARP_SIZE / stride_k; - if (nwarps*stride_j > ncols && threadIdx.y*stride_j >= ncols) { - break; + if (k0_start == k0_stop) { + continue; } #pragma unroll - for (int j0_dst = 0; j0_dst < ncols; j0_dst += (nwarps/np)*stride_j) { - const int j_dst = j0_dst + (threadIdx.y/np)*stride_j + (stride_k == WARP_SIZE ? 0 : threadIdx.x / stride_k); - const int j_tile_KV = (j_dst/mma_B::J)*(np*mma_B::J) + j_dst % mma_B::J; + for (int jc0_dst = 0; jc0_dst < ncols; jc0_dst += (nwarps/np)*stride_jc) { + const int jc_dst = jc0_dst + (threadIdx.y/np)*stride_jc + (stride_k == WARP_SIZE ? 0 : threadIdx.x / stride_k); - if (!is_fixup && jt*ncols + j_dst >= ne01) { + if (jc0_dst + (nwarps/np)*stride_jc > ncols && jc_dst >= ncols) { + break; + } + + const int jc_tile_K = (jc_dst/cols_per_warp)*(np*cols_per_warp) + jc_dst % cols_per_warp; + + const int j_dst = jc_dst / ncols2; + const int c_dst = jc_dst % ncols2; + + if (!is_fixup && jt*ncols1 + j_dst >= ne01) { continue; } - const float * meta_j = (const float *) tile_KV + j_tile_KV*D2_padded + D/2; + + const float * meta_j = (const float *) tile_K + jc_tile_K*D2_padded + D/2; #pragma unroll for (int k0 = k0_start; k0 < k0_stop; k0 += stride_k) { const int k = k0 + (stride_k == WARP_SIZE ? threadIdx.x : threadIdx.x % stride_k); @@ -424,8 +780,8 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile( float2 dstk_val = make_float2(0.0f, 0.0f); #pragma unroll for (int ip = 0; ip < np; ++ip) { - const float KQ_crs = np == 1 ? 1.0f : meta_j[ip*mma_B::J*D2_padded + 2]; - const float2 dstk_val_add = __half22float2(tile_KV[(j_tile_KV + ip*mma_B::J)*D2_padded + k]); + const float KQ_crs = np == 1 ? 1.0f : meta_j[ip*cols_per_warp * D2_padded + 0]; + const float2 dstk_val_add = __half22float2(tile_K[(jc_tile_K + ip*cols_per_warp) * D2_padded + k]); dstk_val.x += dstk_val_add.x*KQ_crs; dstk_val.y += dstk_val_add.y*KQ_crs; } @@ -437,9 +793,9 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile( } if (is_fixup) { - dstk_fixup_data[j_dst*(D/2) + k] = dstk_val; + dstk_fixup_data[jc_dst*(D/2) + k] = dstk_val; } else { - dstk[(jt*ncols + j_dst)*ne02*(D/2) + k] = dstk_val; + dstk[((jt*ncols1 + j_dst)*ne02 + c_dst)*(D/2) + k] = dstk_val; } } } @@ -450,14 +806,18 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile( __syncthreads(); } #else - NO_DEVICE_CODE; + GGML_UNUSED(Q_f2); GGML_UNUSED(K_h2); GGML_UNUSED(V_h2); + GGML_UNUSED(mask_h2); GGML_UNUSED(dstk); GGML_UNUSED(dstk_fixup); + GGML_UNUSED(scale); GGML_UNUSED(slope); GGML_UNUSED(logit_softcap); + GGML_UNUSED(ne01); GGML_UNUSED(ne02); GGML_UNUSED(stride_Q1); + GGML_UNUSED(stride_Q2); GGML_UNUSED(stride_KV); GGML_UNUSED(stride_mask); + GGML_UNUSED(jt); GGML_UNUSED(kb0_start); GGML_UNUSED(kb0_stop); + NO_DEVICE_CODE; #endif // NEW_MMA_AVAILABLE } -template -#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) +template __launch_bounds__(nwarps*WARP_SIZE, 2) -#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) static __global__ void flash_attn_ext_f16( const char * __restrict__ Q, const char * __restrict__ K, @@ -494,22 +854,31 @@ static __global__ void flash_attn_ext_f16( const int ne1, const int ne2, const int ne3) { +#if defined(FLASH_ATTN_AVAILABLE) && defined(NEW_MMA_AVAILABLE) + // Skip unused kernel variants for faster compilation: if (use_logit_softcap && !(D == 128 || D == 256)) { NO_DEVICE_CODE; return; } - static_assert(FATTN_KQ_STRIDE % KQ_stride == 0, "bad KQ_stride"); + static_assert(FATTN_KQ_STRIDE % KQ_per_iter == 0, "bad KQ_per_iter"); const int gqa_ratio = ne02 / ne12; // With grouped query attention there are > 1 Q matrices per K, V matrix. - const int iter_k = ne11 / KQ_stride; - const int iter_j = (ne01 + (ncols - 1)) / ncols; + const int stride_Q1 = nb01 / sizeof(float2); + const int stride_Q2 = nb02 / sizeof(float2); + const int stride_KV = nb11 / sizeof(half2); + const int stride_mask = nb31 / sizeof(half2); + + const int iter_k = ne11 / FATTN_KQ_STRIDE; + const int iter_j = (ne01 + (ncols1 - 1)) / ncols1; + + constexpr int kb_niter = FATTN_KQ_STRIDE / KQ_per_iter; // Number of kernel iterations per assigned KQ slice. // kbc == k block continuous, current index in continuous ijk space. - int kbc = (blockIdx.x + 0)*iter_k*iter_j*ne02 / gridDim.x; - const int kbc_stop = (blockIdx.x + 1)*iter_k*iter_j*ne02 / gridDim.x; + int kbc = (blockIdx.x + 0)*iter_k*iter_j*(ne02/ncols2) / gridDim.x; + const int kbc_stop = (blockIdx.x + 1)*iter_k*iter_j*(ne02/ncols2) / gridDim.x; // If the seams of 2 CUDA blocks fall within an output tile their results need to be combined. // For this we need to track both the block that starts the tile (needs_fixup) and the block that finishes the tile (is_fixup). @@ -522,27 +891,28 @@ static __global__ void flash_attn_ext_f16( const int channel = kbc / (iter_k*iter_j); const int jt = (kbc - channel*iter_k*iter_j) / iter_k; // j index of current tile. - const float2 * Q_f2 = (const float2 *) (Q + nb02* channel); - const half2 * K_h2 = (const half2 *) (K + nb12*(channel / gqa_ratio)); - const half2 * V_h2 = (const half2 *) (V + nb12*(channel / gqa_ratio)); // K and V have same shape - const half * maskh = mask ? (const half *) mask + (nb31/sizeof(half))*jt*ncols : nullptr; - float2 * dstk = ((float2 *) dst) + channel*(D/2); + const float2 * Q_f2 = (const float2 *) (Q + nb02* channel*ncols2); + const half2 * K_h2 = (const half2 *) (K + nb12*(channel*ncols2 / gqa_ratio)); + const half2 * V_h2 = (const half2 *) (V + nb12*(channel*ncols2 / gqa_ratio)); // K and V have same shape + const half2 * mask_h2 = ncols2 > 1 || mask ? (const half2 *) mask + (nb31/sizeof(half2))*jt*ncols1 : nullptr; + float2 * dstk = ((float2 *) dst) + channel*(ncols2 * D/2); - const float slope = get_alibi_slope(max_bias, channel, n_head_log2, m0, m1); + const float slope = ncols2 == 1 ? get_alibi_slope(max_bias, channel, n_head_log2, m0, m1) : 1.0f; + + const int kb0_start_kernel = kb0_start * kb_niter; + const int kb0_stop_kernel = kb0_stop * kb_niter; constexpr bool is_fixup = false; // All but (potentially) the last iterations write their data to dst rather than the fixup buffer. if (kb0_start == 0) { constexpr bool needs_fixup = false; // CUDA block is working on an entire tile. - flash_attn_ext_f16_process_tile - (Q_f2, K_h2, V_h2, maskh, dstk, dst_meta, scale, slope, logit_softcap, - ne00, ne01, ne02, ne03, ne10, ne11, ne12, ne13, ne31, nb31, nb01, nb02, nb03, nb11, nb12, nb13, nb21, nb22, nb23, ne0, ne1, ne2, ne3, - jt, kb0_start, kb0_stop); + flash_attn_ext_f16_process_tile + (Q_f2, K_h2, V_h2, mask_h2, dstk, dst_meta, scale, slope, logit_softcap, + ne01, ne02, stride_Q1, stride_Q2, stride_KV, stride_mask, jt, kb0_start_kernel, kb0_stop_kernel); } else { constexpr bool needs_fixup = true; // CUDA block is working on the beginning of a tile. - flash_attn_ext_f16_process_tile - (Q_f2, K_h2, V_h2, maskh, dstk, dst_meta, scale, slope, logit_softcap, - ne00, ne01, ne02, ne03, ne10, ne11, ne12, ne13, ne31, nb31, nb01, nb02, nb03, nb11, nb12, nb13, nb21, nb22, nb23, ne0, ne1, ne2, ne3, - jt, kb0_start, kb0_stop); + flash_attn_ext_f16_process_tile + (Q_f2, K_h2, V_h2, mask_h2, dstk, dst_meta, scale, slope, logit_softcap, + ne01, ne02, stride_Q1, stride_Q2, stride_KV, stride_mask, jt, kb0_start_kernel, kb0_stop_kernel); } kbc += iter_k; @@ -559,36 +929,59 @@ static __global__ void flash_attn_ext_f16( const int channel = kbc / (iter_k*iter_j); const int jt = (kbc - channel*iter_k*iter_j) / iter_k; // j index of current tile. - const float2 * Q_f2 = (const float2 *) (Q + nb02* channel); - const half2 * K_h2 = (const half2 *) (K + nb12*(channel / gqa_ratio)); - const half2 * V_h2 = (const half2 *) (V + nb12*(channel / gqa_ratio)); // K and V have same shape - const half * maskh = mask ? (const half *) mask + (nb31/sizeof(half))*jt*ncols : nullptr; - float2 * dstk = ((float2 *) dst) + channel*(D/2); + const float2 * Q_f2 = (const float2 *) (Q + nb02* channel*ncols2); + const half2 * K_h2 = (const half2 *) (K + nb12*(channel*ncols2 / gqa_ratio)); + const half2 * V_h2 = (const half2 *) (V + nb12*(channel*ncols2 / gqa_ratio)); // K and V have same shape + const half2 * mask_h2 = ncols2 > 1 || mask ? (const half2 *) mask + (nb31/sizeof(half2))*jt*ncols1 : nullptr; + float2 * dstk = ((float2 *) dst) + channel*(ncols2 * D/2); - const float slope = get_alibi_slope(max_bias, channel, n_head_log2, m0, m1); + const float slope = ncols2 == 1 ? get_alibi_slope(max_bias, channel, n_head_log2, m0, m1) : 1.0f; + + const int kb0_start_kernel = kb0_start * kb_niter; + const int kb0_stop_kernel = kb0_stop * kb_niter; constexpr bool is_fixup = true; // Last index writes its data to fixup buffer to avoid data races with other blocks. constexpr bool needs_fixup = false; - flash_attn_ext_f16_process_tile - (Q_f2, K_h2, V_h2, maskh, dstk, dst_meta, scale, slope, logit_softcap, - ne00, ne01, ne02, ne03, ne10, ne11, ne12, ne13, ne31, nb31, nb01, nb02, nb03, nb11, nb12, nb13, nb21, nb22, nb23, ne0, ne1, ne2, ne3, - jt, kb0_start, kb0_stop); + flash_attn_ext_f16_process_tile + (Q_f2, K_h2, V_h2, mask_h2, dstk, dst_meta, scale, slope, logit_softcap, + ne01, ne02, stride_Q1, stride_Q2, stride_KV, stride_mask, jt, kb0_start_kernel, kb0_stop_kernel); +#else + GGML_UNUSED(Q); GGML_UNUSED(K); GGML_UNUSED(V); GGML_UNUSED(mask); + GGML_UNUSED(dst); GGML_UNUSED(dst_meta); GGML_UNUSED(scale); + GGML_UNUSED(max_bias); GGML_UNUSED(m0); GGML_UNUSED(m1); + GGML_UNUSED(n_head_log2); GGML_UNUSED(logit_softcap); GGML_UNUSED(ne00); + GGML_UNUSED(ne01); GGML_UNUSED(ne02); GGML_UNUSED(ne03); GGML_UNUSED(ne10); + GGML_UNUSED(ne11); GGML_UNUSED(ne12); GGML_UNUSED(ne13); GGML_UNUSED(ne31); + GGML_UNUSED(nb31); GGML_UNUSED(nb01); GGML_UNUSED(nb02); GGML_UNUSED(nb03); + GGML_UNUSED(nb11); GGML_UNUSED(nb12); GGML_UNUSED(nb13); GGML_UNUSED(nb21); + GGML_UNUSED(nb22); GGML_UNUSED(nb23); GGML_UNUSED(ne0); GGML_UNUSED(ne1); + GGML_UNUSED(ne2); GGML_UNUSED(ne3); + NO_DEVICE_CODE; +#endif // defined(FLASH_ATTN_AVAILABLE) && defined(NEW_MMA_AVAILABLE) } -template +template void ggml_cuda_flash_attn_ext_mma_f16_case(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { - typedef mma_A_I16K8 mma_A; - typedef mma_B_J8K8 mma_B; + constexpr int ncols = ncols1 * ncols2; + constexpr int KQ_per_iter = D <= 128 && ncols1 <= 64 ? 64 : 32; + constexpr int nwarps = (KQ_per_iter == 32 && ncols <= 16) ? 2 : 4; + constexpr int ntiles = ncols <= 8 ? 1 : (ncols <= 64 ? 2 : 4); + constexpr int cols_per_warp = ntiles * tile_B::I; - static_assert(D % mma_B::K == 0, "bad D"); - static_assert(cols_per_block % mma_B::J == 0, "bad cols_per_block"); + static_assert(D % tile_B::J == 0, "bad D"); + static_assert(ncols % cols_per_warp == 0, "bad ncols"); const ggml_tensor * KQV = dst; + const int id = ggml_cuda_get_device(); + const int cc = ggml_cuda_info().devices[id].cc; - constexpr int KQ_stride = D <= 128 ? 64 : 32; - constexpr int nwarps = (KQ_stride == 32 && cols_per_block <= 16) ? - cols_per_block/mma_B::J * KQ_stride/mma_A::I : (cols_per_block <= 8 ? 4 : 8); - constexpr size_t nbytes_shared = std::max(KQ_stride, nwarps*mma_B::J) * (D + 8) * sizeof(half); + const int KQ_shared_rows = cp_async_available(cc) ? 2*KQ_per_iter : KQ_per_iter; + + const size_t nbytes_shared_KV = KQ_shared_rows * (D + 8) * sizeof(half); + const size_t nbytes_shared_mask = ncols1 * (KQ_per_iter + 8) * sizeof(half); + const size_t nbytes_shared_combine = nwarps*cols_per_warp * (D + 8) * sizeof(half); + + const size_t nbytes_shared_total = std::max(nbytes_shared_KV + nbytes_shared_mask, nbytes_shared_combine); float logit_softcap; memcpy(&logit_softcap, (const float *) KQV->op_params + 2, sizeof(float)); @@ -596,42 +989,59 @@ void ggml_cuda_flash_attn_ext_mma_f16_case(ggml_backend_cuda_context & ctx, ggml fattn_kernel_t fattn_kernel; if (logit_softcap == 0.0f) { constexpr bool use_logit_softcap = false; - fattn_kernel = flash_attn_ext_f16; + fattn_kernel = flash_attn_ext_f16; } else { constexpr bool use_logit_softcap = true; - fattn_kernel = flash_attn_ext_f16; + fattn_kernel = flash_attn_ext_f16; } - launch_fattn(ctx, dst, fattn_kernel, nwarps, nbytes_shared, true, true); + + launch_fattn + (ctx, dst, fattn_kernel, nwarps, nbytes_shared_total, FATTN_KQ_STRIDE, true, true, true); } -#define DECL_FATTN_MMA_F16_CASE(D, cols_per_block) \ + +#define DECL_FATTN_MMA_F16_CASE(D, ncols1, ncols2) \ template void ggml_cuda_flash_attn_ext_mma_f16_case \ - (ggml_backend_cuda_context & ctx, ggml_tensor * dst) \ + (ggml_backend_cuda_context & ctx, ggml_tensor * dst) \ -extern DECL_FATTN_MMA_F16_CASE( 64, 8); -extern DECL_FATTN_MMA_F16_CASE( 80, 8); -extern DECL_FATTN_MMA_F16_CASE( 96, 8); -extern DECL_FATTN_MMA_F16_CASE(112, 8); -extern DECL_FATTN_MMA_F16_CASE(128, 8); -extern DECL_FATTN_MMA_F16_CASE(256, 8); +#define DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(D, ncols) \ + extern DECL_FATTN_MMA_F16_CASE(D, (ncols)/1, 1); \ + extern DECL_FATTN_MMA_F16_CASE(D, (ncols)/2, 2); \ + extern DECL_FATTN_MMA_F16_CASE(D, (ncols)/4, 4); \ + extern DECL_FATTN_MMA_F16_CASE(D, (ncols)/8, 8); \ -extern DECL_FATTN_MMA_F16_CASE( 64, 16); -extern DECL_FATTN_MMA_F16_CASE( 80, 16); -extern DECL_FATTN_MMA_F16_CASE( 96, 16); -extern DECL_FATTN_MMA_F16_CASE(112, 16); -extern DECL_FATTN_MMA_F16_CASE(128, 16); -extern DECL_FATTN_MMA_F16_CASE(256, 16); +DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 64, 8) +DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 80, 8) +DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 96, 8) +DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(112, 8) +DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(128, 8) +DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(256, 8) -extern DECL_FATTN_MMA_F16_CASE( 64, 32); -extern DECL_FATTN_MMA_F16_CASE( 80, 32); -extern DECL_FATTN_MMA_F16_CASE( 96, 32); -extern DECL_FATTN_MMA_F16_CASE(112, 32); -extern DECL_FATTN_MMA_F16_CASE(128, 32); -extern DECL_FATTN_MMA_F16_CASE(256, 32); +DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 64, 16) +DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 80, 16) +DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 96, 16) +DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(112, 16) +DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(128, 16) +DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(256, 16) -extern DECL_FATTN_MMA_F16_CASE( 64, 64); -extern DECL_FATTN_MMA_F16_CASE( 80, 64); -extern DECL_FATTN_MMA_F16_CASE( 96, 64); -extern DECL_FATTN_MMA_F16_CASE(112, 64); -extern DECL_FATTN_MMA_F16_CASE(128, 64); -extern DECL_FATTN_MMA_F16_CASE(256, 64); +DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 64, 32) +DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 80, 32) +DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 96, 32) +DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(112, 32) +DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(128, 32) +DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(256, 32) + +DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 64, 64) +DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 80, 64) +DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 96, 64) +DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(112, 64) +DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(128, 64) +DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(256, 64) + +// Kernels with ncols == 128 are only 4% faster due to register pressure. +// DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 64, 128) +// DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 80, 128) +// DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 96, 128) +// DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(112, 128) +// DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(128, 128) +// DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(256, 128) // Needs too much shared memory. diff --git a/ggml/src/ggml-cuda/fattn-tile-f16.cu b/ggml/src/ggml-cuda/fattn-tile-f16.cu index d4edbad07..e0039e175 100644 --- a/ggml/src/ggml-cuda/fattn-tile-f16.cu +++ b/ggml/src/ggml-cuda/fattn-tile-f16.cu @@ -4,7 +4,7 @@ #define FATTN_KQ_STRIDE_TILE_F16 64 -template // D == head size +template // D == head size #if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) __launch_bounds__(nwarps*WARP_SIZE, 1) #endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) @@ -44,12 +44,7 @@ static __global__ void flash_attn_tile_ext_f16( const int ne1, const int ne2, const int ne3) { -#ifdef FP16_AVAILABLE - -#ifndef FLASH_ATTN_AVAILABLE - NO_DEVICE_CODE; - return; -#endif // FLASH_ATTN_AVAILABLE +#if defined(FLASH_ATTN_AVAILABLE) && defined(FP16_AVAILABLE) // Skip unused kernel variants for faster compilation: #ifdef FP16_MMA_AVAILABLE @@ -63,18 +58,17 @@ static __global__ void flash_attn_tile_ext_f16( //In this kernel Q, K, V are matrices while i, j, k are matrix indices. - const int ic0 = (blockIdx.x / parallel_blocks) * ncols; // Index of the Q/QKV column to work on. - const int ip = blockIdx.x % parallel_blocks; // Index in group of blocks running for the same column in parallel. + const int ic0 = blockIdx.x * ncols; // Index of the Q/QKV column to work on. const int gqa_ratio = ne02 / ne12; // With grouped query attention there are > 1 Q matrices per K, V matrix. - const float2 * Q_f2 = (const float2 *) (Q + nb02* blockIdx.y + nb01*ic0); - const half2 * K_h2 = (const half2 *) (K + nb12*(blockIdx.y / gqa_ratio)); - const half2 * V_h2 = (const half2 *) (V + nb12*(blockIdx.y / gqa_ratio)); // K and V have same shape + const float2 * Q_f2 = (const float2 *) (Q + nb02* blockIdx.z + nb01*ic0); + const half2 * K_h2 = (const half2 *) (K + nb12*(blockIdx.z / gqa_ratio)); + const half2 * V_h2 = (const half2 *) (V + nb12*(blockIdx.z / gqa_ratio)); // K and V have same shape const half * maskh = (const half *) mask + ne11*ic0; const int stride_KV2 = nb11 / sizeof(half2); - const float slopef = get_alibi_slope(max_bias, blockIdx.y, n_head_log2, m0, m1); + const float slopef = get_alibi_slope(max_bias, blockIdx.z, n_head_log2, m0, m1); const half slopeh = __float2half(slopef); static_assert(D % (2*WARP_SIZE) == 0, "D not divisible by 2*WARP_SIZE == 64."); @@ -110,8 +104,7 @@ static __global__ void flash_attn_tile_ext_f16( __syncthreads(); - const int k_start = parallel_blocks == 1 ? 0 : ip*FATTN_KQ_STRIDE_TILE_F16; - for (int k_VKQ_0 = k_start; k_VKQ_0 < ne11; k_VKQ_0 += parallel_blocks*FATTN_KQ_STRIDE_TILE_F16) { + for (int k_VKQ_0 = blockIdx.y*FATTN_KQ_STRIDE_TILE_F16; k_VKQ_0 < ne11; k_VKQ_0 += gridDim.y*FATTN_KQ_STRIDE_TILE_F16) { // Calculate KQ tile and keep track of new maximum KQ values: half kqmax_new[ncols/nwarps]; @@ -276,24 +269,36 @@ static __global__ void flash_attn_tile_ext_f16( const int i0 = i00 + 2*threadIdx.x; half2 dst_val = VKQ[j_VKQ_0/nwarps][i0/(2*WARP_SIZE)]; - if (parallel_blocks == 1) { + if (gridDim.y == 1) { dst_val /= __half2half2(kqsum_j); } - const int j_dst = (ic0 + j_VKQ)*parallel_blocks + ip; - dst[j_dst*D*gridDim.y + D*blockIdx.y + i0 + 0] = __low2float(dst_val); - dst[j_dst*D*gridDim.y + D*blockIdx.y + i0 + 1] = __high2float(dst_val); + const int j_dst = (ic0 + j_VKQ)*gridDim.y + blockIdx.y; + dst[j_dst*D*gridDim.z + D*blockIdx.z + i0 + 0] = __low2float(dst_val); + dst[j_dst*D*gridDim.z + D*blockIdx.z + i0 + 1] = __high2float(dst_val); } - if (parallel_blocks != 1 && threadIdx.x == 0) { - dst_meta[(ic0 + j_VKQ)*gridDim.y*parallel_blocks + blockIdx.y*parallel_blocks + ip] = make_float2(kqmax[j_VKQ_0/nwarps], kqsum_j); + if (gridDim.y != 1 && threadIdx.x == 0) { + dst_meta[((ic0 + j_VKQ)*gridDim.z + blockIdx.z) * gridDim.y + blockIdx.y] = make_float2(kqmax[j_VKQ_0/nwarps], kqsum_j); } } #else - NO_DEVICE_CODE; -#endif // FP16_AVAILABLE + GGML_UNUSED(Q); GGML_UNUSED(K); GGML_UNUSED(V); GGML_UNUSED(mask); + GGML_UNUSED(dst); GGML_UNUSED(dst_meta); GGML_UNUSED(scale); + GGML_UNUSED(max_bias); GGML_UNUSED(m0); GGML_UNUSED(m1); + GGML_UNUSED(n_head_log2); GGML_UNUSED(logit_softcap); + GGML_UNUSED(ne00); GGML_UNUSED(ne01); GGML_UNUSED(ne02); + GGML_UNUSED(ne03); GGML_UNUSED(ne10); GGML_UNUSED(ne11); + GGML_UNUSED(ne12); GGML_UNUSED(ne13); GGML_UNUSED(ne31); + GGML_UNUSED(nb31); GGML_UNUSED(nb01); GGML_UNUSED(nb02); + GGML_UNUSED(nb03); GGML_UNUSED(nb11); GGML_UNUSED(nb12); + GGML_UNUSED(nb13); GGML_UNUSED(nb21); GGML_UNUSED(nb22); + GGML_UNUSED(nb23); GGML_UNUSED(ne0); GGML_UNUSED(ne1); + GGML_UNUSED(ne2); GGML_UNUSED(ne3); + NO_DEVICE_CODE; +#endif // defined(FLASH_ATTN_AVAILABLE) && defined(FP16_AVAILABLE) } -template +template void launch_fattn_tile_f16_64_128(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { const ggml_tensor * Q = dst->src[0]; switch (Q->ne[0]) { @@ -301,15 +306,17 @@ void launch_fattn_tile_f16_64_128(ggml_backend_cuda_context & ctx, ggml_tensor * constexpr int D = 64; constexpr int nwarps = 8; constexpr size_t nbytes_shared = 0; - fattn_kernel_t fattn_kernel = flash_attn_tile_ext_f16; - launch_fattn(ctx, dst, fattn_kernel, nwarps, nbytes_shared, true, true); + fattn_kernel_t fattn_kernel = flash_attn_tile_ext_f16; + launch_fattn + (ctx, dst, fattn_kernel, nwarps, nbytes_shared, FATTN_KQ_STRIDE_TILE_F16, true, true, false); } break; case 128: { constexpr int D = 128; constexpr int nwarps = 8; constexpr size_t nbytes_shared = 0; - fattn_kernel_t fattn_kernel = flash_attn_tile_ext_f16; - launch_fattn(ctx, dst, fattn_kernel, nwarps, nbytes_shared, true, true); + fattn_kernel_t fattn_kernel = flash_attn_tile_ext_f16; + launch_fattn + (ctx, dst, fattn_kernel, nwarps, nbytes_shared, FATTN_KQ_STRIDE_TILE_F16, true, true, false); } break; default: { GGML_ABORT("FlashAttention without tensor cores only supports head sizes 64 and 128."); @@ -329,37 +336,22 @@ void ggml_cuda_flash_attn_ext_tile_f16(ggml_backend_cuda_context & ctx, ggml_ten if (Q->ne[1] <= 16) { constexpr int cols_per_block = 16; - constexpr int parallel_blocks = 4; if (logit_softcap == 0.0f) { constexpr bool use_logit_softcap = false; - launch_fattn_tile_f16_64_128(ctx, dst); + launch_fattn_tile_f16_64_128(ctx, dst); } else { constexpr bool use_logit_softcap = true; - launch_fattn_tile_f16_64_128(ctx, dst); - } - return; - } - - if (Q->ne[1] <= 32) { - constexpr int cols_per_block = 32; - constexpr int parallel_blocks = 4; - if (logit_softcap == 0.0f) { - constexpr bool use_logit_softcap = false; - launch_fattn_tile_f16_64_128(ctx, dst); - } else { - constexpr bool use_logit_softcap = true; - launch_fattn_tile_f16_64_128(ctx, dst); + launch_fattn_tile_f16_64_128(ctx, dst); } return; } constexpr int cols_per_block = 32; - constexpr int parallel_blocks = 1; if (logit_softcap == 0.0f) { constexpr bool use_logit_softcap = false; - launch_fattn_tile_f16_64_128(ctx, dst); + launch_fattn_tile_f16_64_128(ctx, dst); } else { constexpr bool use_logit_softcap = true; - launch_fattn_tile_f16_64_128(ctx, dst); + launch_fattn_tile_f16_64_128(ctx, dst); } } diff --git a/ggml/src/ggml-cuda/fattn-tile-f32.cu b/ggml/src/ggml-cuda/fattn-tile-f32.cu index 0d274f332..fcb6f848f 100644 --- a/ggml/src/ggml-cuda/fattn-tile-f32.cu +++ b/ggml/src/ggml-cuda/fattn-tile-f32.cu @@ -4,7 +4,7 @@ #define FATTN_KQ_STRIDE_TILE_F32 32 -template // D == head size +template // D == head size #if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) __launch_bounds__(nwarps*WARP_SIZE, 1) #endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) @@ -44,10 +44,7 @@ static __global__ void flash_attn_tile_ext_f32( const int ne1, const int ne2, const int ne3) { -#ifndef FLASH_ATTN_AVAILABLE - NO_DEVICE_CODE; - return; -#endif // FLASH_ATTN_AVAILABLE +#ifdef FLASH_ATTN_AVAILABLE // Skip unused kernel variants for faster compilation: #ifdef FP16_MMA_AVAILABLE @@ -55,24 +52,35 @@ static __global__ void flash_attn_tile_ext_f32( return; #endif // FP16_MMA_AVAILABLE if (use_logit_softcap && !(D == 128 || D == 256)) { + GGML_UNUSED(Q); GGML_UNUSED(K); GGML_UNUSED(V); GGML_UNUSED(mask); + GGML_UNUSED(dst); GGML_UNUSED(dst_meta); GGML_UNUSED(scale); + GGML_UNUSED(max_bias); GGML_UNUSED(m0); GGML_UNUSED(m1); + GGML_UNUSED(n_head_log2); GGML_UNUSED(logit_softcap); + GGML_UNUSED(ne00); GGML_UNUSED(ne01); GGML_UNUSED(ne02); + GGML_UNUSED(ne03); GGML_UNUSED(ne10); GGML_UNUSED(ne11); + GGML_UNUSED(ne12); GGML_UNUSED(ne13); GGML_UNUSED(ne31); + GGML_UNUSED(nb31); GGML_UNUSED(nb01); GGML_UNUSED(nb02); + GGML_UNUSED(nb03); GGML_UNUSED(nb11); GGML_UNUSED(nb12); + GGML_UNUSED(nb13); GGML_UNUSED(nb21); GGML_UNUSED(nb22); + GGML_UNUSED(nb23); GGML_UNUSED(ne0); GGML_UNUSED(ne1); + GGML_UNUSED(ne2); GGML_UNUSED(ne3); NO_DEVICE_CODE; return; } // In this kernel Q, K, V are matrices while i, j, k are matrix indices. - const int ic0 = (blockIdx.x / parallel_blocks) * ncols; // Index of the Q/QKV column to work on. - const int ip = blockIdx.x % parallel_blocks; // Index in group of blocks running for the same column in parallel. + const int ic0 = blockIdx.x * ncols; // Index of the Q/QKV column to work on. const int gqa_ratio = ne02 / ne12; // With grouped query attention there are > 1 Q matrices per K, V matrix. - const float2 * Q_f2 = (const float2 *) (Q + nb02* blockIdx.y + nb01*ic0); - const half2 * K_h2 = (const half2 *) (K + nb12*(blockIdx.y / gqa_ratio)); - const half2 * V_h2 = (const half2 *) (V + nb12*(blockIdx.y / gqa_ratio)); // K and V have same shape + const float2 * Q_f2 = (const float2 *) (Q + nb02* blockIdx.z + nb01*ic0); + const half2 * K_h2 = (const half2 *) (K + nb12*(blockIdx.z / gqa_ratio)); + const half2 * V_h2 = (const half2 *) (V + nb12*(blockIdx.z / gqa_ratio)); // K and V have same shape const half * maskh = (const half *) mask + ne11*ic0; const int stride_KV2 = nb11 / sizeof(half2); - const float slope = get_alibi_slope(max_bias, blockIdx.y, n_head_log2, m0, m1); + const float slope = get_alibi_slope(max_bias, blockIdx.z, n_head_log2, m0, m1); static_assert(D % (2*WARP_SIZE) == 0, "D not divisible by 2*WARP_SIZE == 64."); @@ -106,8 +114,7 @@ static __global__ void flash_attn_tile_ext_f32( __syncthreads(); - const int k_start = parallel_blocks == 1 ? 0 : ip*FATTN_KQ_STRIDE_TILE_F32; - for (int k_VKQ_0 = k_start; k_VKQ_0 < ne11; k_VKQ_0 += parallel_blocks*FATTN_KQ_STRIDE_TILE_F32) { + for (int k_VKQ_0 = blockIdx.y*FATTN_KQ_STRIDE_TILE_F32; k_VKQ_0 < ne11; k_VKQ_0 += gridDim.y*FATTN_KQ_STRIDE_TILE_F32) { // Calculate KQ tile and keep track of new maximum KQ values: float kqmax_new[ncols/nwarps]; @@ -272,22 +279,37 @@ static __global__ void flash_attn_tile_ext_f32( const int i0 = i00 + 2*threadIdx.x; float2 dst_val = VKQ[j_VKQ_0/nwarps][i0/(2*WARP_SIZE)]; - if (parallel_blocks == 1) { + if (gridDim.y == 1) { dst_val.x /= kqsum_j; dst_val.y /= kqsum_j; } - const int j_dst = (ic0 + j_VKQ)*parallel_blocks + ip; - dst[j_dst*D*gridDim.y + D*blockIdx.y + i0 + 0] = dst_val.x; - dst[j_dst*D*gridDim.y + D*blockIdx.y + i0 + 1] = dst_val.y; + const int j_dst = (ic0 + j_VKQ)*gridDim.y + blockIdx.y; + dst[j_dst*D*gridDim.z + D*blockIdx.z + i0 + 0] = dst_val.x; + dst[j_dst*D*gridDim.z + D*blockIdx.z + i0 + 1] = dst_val.y; } - if (parallel_blocks != 1 && threadIdx.x == 0) { - dst_meta[(ic0 + j_VKQ)*gridDim.y*parallel_blocks + blockIdx.y*parallel_blocks + ip] = make_float2(kqmax[j_VKQ_0/nwarps], kqsum_j); + if (gridDim.y != 1 && threadIdx.x == 0) { + dst_meta[((ic0 + j_VKQ)*gridDim.z + blockIdx.z) * gridDim.y + blockIdx.y] = make_float2(kqmax[j_VKQ_0/nwarps], kqsum_j); } } +#else + GGML_UNUSED(Q); GGML_UNUSED(K); GGML_UNUSED(V); GGML_UNUSED(mask); + GGML_UNUSED(dst); GGML_UNUSED(dst_meta); GGML_UNUSED(scale); + GGML_UNUSED(max_bias); GGML_UNUSED(m0); GGML_UNUSED(m1); + GGML_UNUSED(n_head_log2); GGML_UNUSED(logit_softcap); + GGML_UNUSED(ne00); GGML_UNUSED(ne01); GGML_UNUSED(ne02); + GGML_UNUSED(ne03); GGML_UNUSED(ne10); GGML_UNUSED(ne11); + GGML_UNUSED(ne12); GGML_UNUSED(ne13); GGML_UNUSED(ne31); + GGML_UNUSED(nb31); GGML_UNUSED(nb01); GGML_UNUSED(nb02); + GGML_UNUSED(nb03); GGML_UNUSED(nb11); GGML_UNUSED(nb12); + GGML_UNUSED(nb13); GGML_UNUSED(nb21); GGML_UNUSED(nb22); + GGML_UNUSED(nb23); GGML_UNUSED(ne0); GGML_UNUSED(ne1); + GGML_UNUSED(ne2); GGML_UNUSED(ne3); + NO_DEVICE_CODE; +#endif // FLASH_ATTN_AVAILABLE } -template +template void launch_fattn_tile_f32_64_128(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { const ggml_tensor * Q = dst->src[0]; switch (Q->ne[0]) { @@ -295,15 +317,17 @@ void launch_fattn_tile_f32_64_128(ggml_backend_cuda_context & ctx, ggml_tensor * constexpr int D = 64; constexpr int nwarps = 8; constexpr size_t nbytes_shared = 0; - fattn_kernel_t fattn_kernel = flash_attn_tile_ext_f32; - launch_fattn(ctx, dst, fattn_kernel, nwarps, nbytes_shared, true, true); + fattn_kernel_t fattn_kernel = flash_attn_tile_ext_f32; + launch_fattn + (ctx, dst, fattn_kernel, nwarps, nbytes_shared, FATTN_KQ_STRIDE_TILE_F32, true, true, false); } break; case 128: { constexpr int D = 128; constexpr int nwarps = 8; constexpr size_t nbytes_shared = 0; - fattn_kernel_t fattn_kernel = flash_attn_tile_ext_f32; - launch_fattn(ctx, dst, fattn_kernel, nwarps, nbytes_shared, true, true); + fattn_kernel_t fattn_kernel = flash_attn_tile_ext_f32; + launch_fattn + (ctx, dst, fattn_kernel, nwarps, nbytes_shared, FATTN_KQ_STRIDE_TILE_F32, true, true, false); } break; default: { GGML_ABORT("FlashAttention without tensor cores only supports head sizes 64 and 128."); @@ -320,37 +344,22 @@ void ggml_cuda_flash_attn_ext_tile_f32(ggml_backend_cuda_context & ctx, ggml_ten if (Q->ne[1] <= 16) { constexpr int cols_per_block = 16; - constexpr int parallel_blocks = 4; if (logit_softcap == 0.0f) { constexpr bool use_logit_softcap = false; - launch_fattn_tile_f32_64_128(ctx, dst); + launch_fattn_tile_f32_64_128(ctx, dst); } else { constexpr bool use_logit_softcap = true; - launch_fattn_tile_f32_64_128(ctx, dst); - } - return; - } - - if (Q->ne[1] <= 32) { - constexpr int cols_per_block = 32; - constexpr int parallel_blocks = 4; - if (logit_softcap == 0.0f) { - constexpr bool use_logit_softcap = false; - launch_fattn_tile_f32_64_128(ctx, dst); - } else { - constexpr bool use_logit_softcap = true; - launch_fattn_tile_f32_64_128(ctx, dst); + launch_fattn_tile_f32_64_128(ctx, dst); } return; } constexpr int cols_per_block = 32; - constexpr int parallel_blocks = 1; if (logit_softcap == 0.0f) { constexpr bool use_logit_softcap = false; - launch_fattn_tile_f32_64_128(ctx, dst); + launch_fattn_tile_f32_64_128(ctx, dst); } else { constexpr bool use_logit_softcap = true; - launch_fattn_tile_f32_64_128(ctx, dst); + launch_fattn_tile_f32_64_128(ctx, dst); } } diff --git a/ggml/src/ggml-cuda/fattn-vec-f16.cuh b/ggml/src/ggml-cuda/fattn-vec-f16.cuh index d9ac44246..e17d2d0e4 100644 --- a/ggml/src/ggml-cuda/fattn-vec-f16.cuh +++ b/ggml/src/ggml-cuda/fattn-vec-f16.cuh @@ -1,7 +1,7 @@ #include "common.cuh" #include "fattn-common.cuh" -template // D == head size +template // D == head size #if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) __launch_bounds__(D, 1) #endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) @@ -41,12 +41,7 @@ static __global__ void flash_attn_vec_ext_f16( const int ne1, const int ne2, const int ne3) { -#ifdef FP16_AVAILABLE - -#ifndef FLASH_ATTN_AVAILABLE - NO_DEVICE_CODE; - return; -#endif // FLASH_ATTN_AVAILABLE +#if defined(FLASH_ATTN_AVAILABLE) && defined(FP16_AVAILABLE) // Skip unused kernel variants for faster compilation: if (use_logit_softcap && !(D == 128 || D == 256)) { @@ -60,17 +55,16 @@ static __global__ void flash_attn_vec_ext_f16( constexpr bool Q_q8_1 = type_K != GGML_TYPE_F16; constexpr dequantize_1_f16_t dequantize_1_v = get_dequantize_1_f16(type_V); - const int ic0 = (blockIdx.x / parallel_blocks) * ncols; // Index of the Q/QKV column to work on. - const int ip = blockIdx.x % parallel_blocks; // Index in group of blocks running for the same column in parallel. + const int ic0 = blockIdx.x * ncols; // Index of the Q/QKV column to work on. const int gqa_ratio = ne02 / ne12; // With grouped query attention there are > 1 Q matrices per K, V matrix. - Q += nb02* blockIdx.y + nb01*ic0; - K += nb12*(blockIdx.y / gqa_ratio); - V += nb22*(blockIdx.y / gqa_ratio); + Q += nb02* blockIdx.z + nb01*ic0; + K += nb12*(blockIdx.z / gqa_ratio); + V += nb22*(blockIdx.z / gqa_ratio); const half * maskh = (const half *) mask + ne11*ic0; - const float slopef = get_alibi_slope(max_bias, blockIdx.y, n_head_log2, m0, m1); + const float slopef = get_alibi_slope(max_bias, blockIdx.z, n_head_log2, m0, m1); const half slopeh = __float2half(slopef); static_assert(D % (2*WARP_SIZE) == 0, "D not divisible by 2*WARP_SIZE == 64."); @@ -177,8 +171,7 @@ static __global__ void flash_attn_vec_ext_f16( half2 VKQ[ncols] = {{0.0f, 0.0f}}; - const int k_start = parallel_blocks == 1 ? 0 : ip*D; - for (int k_VKQ_0 = k_start; k_VKQ_0 < ne11; k_VKQ_0 += parallel_blocks*D) { + for (int k_VKQ_0 = blockIdx.y*D; k_VKQ_0 < ne11; k_VKQ_0 += gridDim.y*D) { // Calculate KQ tile and keep track of new maximum KQ values: // For unknown reasons using a half array of size 1 for kqmax_new causes a performance regression, @@ -288,29 +281,41 @@ static __global__ void flash_attn_vec_ext_f16( kqsum[j_VKQ] = warp_reduce_sum((float)kqsum[j_VKQ]); half dst_val = (__low2half(VKQ[j_VKQ]) + __high2half(VKQ[j_VKQ])); - if (parallel_blocks == 1) { + if (gridDim.y == 1) { dst_val /= kqsum[j_VKQ]; } - const int j_dst = (ic0 + j_VKQ)*parallel_blocks + ip; - dst[j_dst*D*gridDim.y + D*blockIdx.y + tid] = dst_val; + const int j_dst = (ic0 + j_VKQ)*gridDim.y + blockIdx.y; + dst[j_dst*D*gridDim.z + D*blockIdx.z + tid] = dst_val; } - if (parallel_blocks != 1 && tid < ncols && (ncols <= 2 || ic0 + tid < ne01)) { - dst_meta[(ic0 + tid)*gridDim.y*parallel_blocks + blockIdx.y*parallel_blocks + ip] = make_float2(kqmax[tid], kqsum[tid]); + if (gridDim.y != 1 && tid < ncols && (ncols <= 2 || ic0 + tid < ne01)) { + dst_meta[((ic0 + tid)*gridDim.z + blockIdx.z) * gridDim.y + blockIdx.y] = make_float2(kqmax[tid], kqsum[tid]); } #else - NO_DEVICE_CODE; -#endif // FP16_AVAILABLE + GGML_UNUSED(Q); GGML_UNUSED(K); GGML_UNUSED(V); GGML_UNUSED(mask); + GGML_UNUSED(dst); GGML_UNUSED(dst_meta); GGML_UNUSED(scale); + GGML_UNUSED(max_bias); GGML_UNUSED(m0); GGML_UNUSED(m1); + GGML_UNUSED(n_head_log2); GGML_UNUSED(logit_softcap); + GGML_UNUSED(ne00); GGML_UNUSED(ne01); GGML_UNUSED(ne02); + GGML_UNUSED(ne03); GGML_UNUSED(ne10); GGML_UNUSED(ne11); + GGML_UNUSED(ne12); GGML_UNUSED(ne13); GGML_UNUSED(ne31); + GGML_UNUSED(nb31); GGML_UNUSED(nb01); GGML_UNUSED(nb02); + GGML_UNUSED(nb03); GGML_UNUSED(nb11); GGML_UNUSED(nb12); + GGML_UNUSED(nb13); GGML_UNUSED(nb21); GGML_UNUSED(nb22); + GGML_UNUSED(nb23); GGML_UNUSED(ne0); GGML_UNUSED(ne1); + GGML_UNUSED(ne2); GGML_UNUSED(ne3); + NO_DEVICE_CODE; +#endif // defined(FLASH_ATTN_AVAILABLE) && defined(FP16_AVAILABLE) } -template +template void ggml_cuda_flash_attn_ext_vec_f16_case_impl(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { constexpr int nwarps = D/WARP_SIZE; - fattn_kernel_t fattn_kernel = flash_attn_vec_ext_f16; + fattn_kernel_t fattn_kernel = flash_attn_vec_ext_f16; constexpr bool need_f16_K = D != 128; constexpr bool need_f16_V = D != 128 && D != 64; constexpr size_t nbytes_shared = 0; - launch_fattn(ctx, dst, fattn_kernel, nwarps, nbytes_shared, need_f16_K, need_f16_V); + launch_fattn(ctx, dst, fattn_kernel, nwarps, nbytes_shared, D, need_f16_K, need_f16_V, false); } template @@ -330,65 +335,48 @@ void ggml_cuda_flash_attn_ext_vec_f16_case(ggml_backend_cuda_context & ctx, ggml memcpy(&logit_softcap, (const float *) KQV->op_params + 2, sizeof(float)); if (Q->ne[1] == 1) { - constexpr int cols_per_block = 1; - constexpr int parallel_blocks = 4; + constexpr int cols_per_block = 1; if (logit_softcap == 0.0f) { constexpr bool use_logit_softcap = false; - ggml_cuda_flash_attn_ext_vec_f16_case_impl(ctx, dst); + ggml_cuda_flash_attn_ext_vec_f16_case_impl(ctx, dst); } else { constexpr bool use_logit_softcap = true; - ggml_cuda_flash_attn_ext_vec_f16_case_impl(ctx, dst); + ggml_cuda_flash_attn_ext_vec_f16_case_impl(ctx, dst); } return; } if (Q->ne[1] == 2) { - constexpr int cols_per_block = 2; - constexpr int parallel_blocks = 4; + constexpr int cols_per_block = 2; if (logit_softcap == 0.0f) { constexpr bool use_logit_softcap = false; - ggml_cuda_flash_attn_ext_vec_f16_case_impl(ctx, dst); + ggml_cuda_flash_attn_ext_vec_f16_case_impl(ctx, dst); } else { constexpr bool use_logit_softcap = true; - ggml_cuda_flash_attn_ext_vec_f16_case_impl(ctx, dst); + ggml_cuda_flash_attn_ext_vec_f16_case_impl(ctx, dst); } return; } if (Q->ne[1] <= 4) { - constexpr int cols_per_block = 4; - constexpr int parallel_blocks = 4; + constexpr int cols_per_block = 4; if (logit_softcap == 0.0f) { constexpr bool use_logit_softcap = false; - ggml_cuda_flash_attn_ext_vec_f16_case_impl(ctx, dst); + ggml_cuda_flash_attn_ext_vec_f16_case_impl(ctx, dst); } else { constexpr bool use_logit_softcap = true; - ggml_cuda_flash_attn_ext_vec_f16_case_impl(ctx, dst); + ggml_cuda_flash_attn_ext_vec_f16_case_impl(ctx, dst); } return; } - if (Q->ne[1] <= 8) { - constexpr int cols_per_block = 8; - constexpr int parallel_blocks = 4; - if (logit_softcap == 0.0f) { - constexpr bool use_logit_softcap = false; - ggml_cuda_flash_attn_ext_vec_f16_case_impl(ctx, dst); - } else { - constexpr bool use_logit_softcap = true; - ggml_cuda_flash_attn_ext_vec_f16_case_impl(ctx, dst); - } - return; - } - - constexpr int cols_per_block = 8; - constexpr int parallel_blocks = 1; + constexpr int cols_per_block = 8; if (logit_softcap == 0.0f) { constexpr bool use_logit_softcap = false; - ggml_cuda_flash_attn_ext_vec_f16_case_impl(ctx, dst); + ggml_cuda_flash_attn_ext_vec_f16_case_impl(ctx, dst); } else { constexpr bool use_logit_softcap = true; - ggml_cuda_flash_attn_ext_vec_f16_case_impl(ctx, dst); + ggml_cuda_flash_attn_ext_vec_f16_case_impl(ctx, dst); } } diff --git a/ggml/src/ggml-cuda/fattn-vec-f32.cuh b/ggml/src/ggml-cuda/fattn-vec-f32.cuh index 6ef8f9dcc..d42ddca49 100644 --- a/ggml/src/ggml-cuda/fattn-vec-f32.cuh +++ b/ggml/src/ggml-cuda/fattn-vec-f32.cuh @@ -1,7 +1,7 @@ #include "common.cuh" #include "fattn-common.cuh" -template // D == head size +template // D == head size #if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) __launch_bounds__(D, 1) #endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) @@ -41,13 +41,22 @@ static __global__ void flash_attn_vec_ext_f32( const int ne1, const int ne2, const int ne3) { -#ifndef FLASH_ATTN_AVAILABLE - NO_DEVICE_CODE; - return; -#endif // FLASH_ATTN_AVAILABLE +#ifdef FLASH_ATTN_AVAILABLE // Skip unused kernel variants for faster compilation: if (use_logit_softcap && !(D == 128 || D == 256)) { + GGML_UNUSED(Q); GGML_UNUSED(K); GGML_UNUSED(V); GGML_UNUSED(mask); + GGML_UNUSED(dst); GGML_UNUSED(dst_meta); GGML_UNUSED(scale); + GGML_UNUSED(max_bias); GGML_UNUSED(m0); GGML_UNUSED(m1); + GGML_UNUSED(n_head_log2); GGML_UNUSED(logit_softcap); + GGML_UNUSED(ne00); GGML_UNUSED(ne01); GGML_UNUSED(ne02); + GGML_UNUSED(ne03); GGML_UNUSED(ne10); GGML_UNUSED(ne11); + GGML_UNUSED(ne12); GGML_UNUSED(ne13); GGML_UNUSED(ne31); + GGML_UNUSED(nb31); GGML_UNUSED(nb01); GGML_UNUSED(nb02); + GGML_UNUSED(nb03); GGML_UNUSED(nb11); GGML_UNUSED(nb12); + GGML_UNUSED(nb13); GGML_UNUSED(nb21); GGML_UNUSED(nb22); + GGML_UNUSED(nb23); GGML_UNUSED(ne0); GGML_UNUSED(ne1); + GGML_UNUSED(ne2); GGML_UNUSED(ne3); NO_DEVICE_CODE; return; } @@ -58,16 +67,15 @@ static __global__ void flash_attn_vec_ext_f32( constexpr bool Q_q8_1 = type_K != GGML_TYPE_F16; constexpr dequantize_1_f32_t dequantize_1_v = get_dequantize_1_f32(type_V); - const int ic0 = (blockIdx.x / parallel_blocks) * ncols; // Index of the Q/QKV column to work on. - const int ip = blockIdx.x % parallel_blocks; // Index in group of blocks running for the same column in parallel. + const int ic0 = blockIdx.x * ncols; // Index of the Q/QKV column to work on. const int gqa_ratio = ne02 / ne12; // With grouped query attention there are > 1 Q matrices per K, V matrix. - Q += nb02* blockIdx.y + nb01*ic0; - K += nb12*(blockIdx.y / gqa_ratio); - V += nb22*(blockIdx.y / gqa_ratio); // K and V have same shape + Q += nb02* blockIdx.z + nb01*ic0; + K += nb12*(blockIdx.z / gqa_ratio); + V += nb22*(blockIdx.z / gqa_ratio); // K and V have same shape const half * maskh = (const half *) mask + ne11*ic0; - const float slope = get_alibi_slope(max_bias, blockIdx.y, n_head_log2, m0, m1); + const float slope = get_alibi_slope(max_bias, blockIdx.z, n_head_log2, m0, m1); static_assert(D % (2*WARP_SIZE) == 0, "D not divisible by 2*WARP_SIZE == 64."); constexpr int nwarps = D / WARP_SIZE; @@ -118,7 +126,7 @@ static __global__ void flash_attn_vec_ext_f32( // Set memory to zero if out of bounds: if (ncols > 2 && ic0 + j >= ne01) { #pragma unroll - for (int i0 = 0; i0 < D/sizeof(int); i0 += WARP_SIZE) { + for (int i0 = 0; i0 < int(D/sizeof(int)); i0 += WARP_SIZE) { const int i = i0 + threadIdx.x; tmp_q_i32[i] = 0; @@ -131,7 +139,7 @@ static __global__ void flash_attn_vec_ext_f32( const float * Q_f = (const float *) (Q + j*nb01); #pragma unroll - for (int i0 = 0; i0 < D/sizeof(int); i0 += WARP_SIZE) { + for (int i0 = 0; i0 < int(D/sizeof(int)); i0 += WARP_SIZE) { quantize_q8_1_to_shared(Q_f + 4*i0, scale, tmp_q_i32, tmp_q_ds); } } @@ -144,7 +152,7 @@ static __global__ void flash_attn_vec_ext_f32( float2 * tmp_q_ds = (float2 *) (tmp_q_i32 + D/sizeof(int)); #pragma unroll - for (int i0 = 0; i0 < D/sizeof(int); i0 += WARP_SIZE) { + for (int i0 = 0; i0 < int(D/sizeof(int)); i0 += WARP_SIZE) { const int i = i0 + threadIdx.x; Q_i32[j][i0/WARP_SIZE] = tmp_q_i32[i]; @@ -170,8 +178,7 @@ static __global__ void flash_attn_vec_ext_f32( float VKQ[ncols] = {0.0f}; - const int k_start = parallel_blocks == 1 ? 0 : ip*D; - for (int k_VKQ_0 = k_start; k_VKQ_0 < ne11; k_VKQ_0 += parallel_blocks*D) { + for (int k_VKQ_0 = blockIdx.y*D; k_VKQ_0 < ne11; k_VKQ_0 += gridDim.y*D) { // Calculate KQ tile and keep track of new maximum KQ values: float kqmax_new_arr[ncols]; @@ -271,26 +278,39 @@ static __global__ void flash_attn_vec_ext_f32( kqsum[j_VKQ] = warp_reduce_sum(kqsum[j_VKQ]); float dst_val = VKQ[j_VKQ]; - if (parallel_blocks == 1) { + if (gridDim.y == 1) { dst_val /= kqsum[j_VKQ]; } - const int j_dst = (ic0 + j_VKQ)*parallel_blocks + ip; - dst[j_dst*D*gridDim.y + D*blockIdx.y + tid] = dst_val; + const int j_dst = (ic0 + j_VKQ)*gridDim.y + blockIdx.y; + dst[j_dst*D*gridDim.z + D*blockIdx.z + tid] = dst_val; } - if (parallel_blocks != 1 && tid < ncols && (ncols <= 2 || ic0 + tid < ne01)) { - dst_meta[(ic0 + tid)*gridDim.y*parallel_blocks + blockIdx.y*parallel_blocks + ip] = make_float2(kqmax[tid], kqsum[tid]); + if (gridDim.y != 1 && tid < ncols && (ncols <= 2 || ic0 + tid < ne01)) { + dst_meta[((ic0 + tid)*gridDim.z + blockIdx.z) * gridDim.y + blockIdx.y] = make_float2(kqmax[tid], kqsum[tid]); } +#else + GGML_UNUSED(Q); GGML_UNUSED(K); GGML_UNUSED(V); GGML_UNUSED(mask); + GGML_UNUSED(dst); GGML_UNUSED(dst_meta); GGML_UNUSED(scale); + GGML_UNUSED(max_bias); GGML_UNUSED(m0); GGML_UNUSED(m1); + GGML_UNUSED(n_head_log2); GGML_UNUSED(logit_softcap); GGML_UNUSED(ne00); + GGML_UNUSED(ne01); GGML_UNUSED(ne02); GGML_UNUSED(ne03); GGML_UNUSED(ne10); + GGML_UNUSED(ne11); GGML_UNUSED(ne12); GGML_UNUSED(ne13); GGML_UNUSED(ne31); + GGML_UNUSED(nb31); GGML_UNUSED(nb01); GGML_UNUSED(nb02); GGML_UNUSED(nb03); + GGML_UNUSED(nb11); GGML_UNUSED(nb12); GGML_UNUSED(nb13); GGML_UNUSED(nb21); + GGML_UNUSED(nb22); GGML_UNUSED(nb23); GGML_UNUSED(ne0); GGML_UNUSED(ne1); + GGML_UNUSED(ne2); GGML_UNUSED(ne3); + NO_DEVICE_CODE; +#endif // FLASH_ATTN_AVAILABLE } -template +template void ggml_cuda_flash_attn_ext_vec_f32_case_impl(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { constexpr int nwarps = D/WARP_SIZE; - fattn_kernel_t fattn_kernel = flash_attn_vec_ext_f32; + fattn_kernel_t fattn_kernel = flash_attn_vec_ext_f32; constexpr bool need_f16_K = D != 128; constexpr bool need_f16_V = D != 128 && D != 64; constexpr size_t nbytes_shared = 0; - launch_fattn(ctx, dst, fattn_kernel, nwarps, nbytes_shared, need_f16_K, need_f16_V); + launch_fattn(ctx, dst, fattn_kernel, nwarps, nbytes_shared, D, need_f16_K, need_f16_V, false); } template @@ -307,65 +327,48 @@ void ggml_cuda_flash_attn_ext_vec_f32_case(ggml_backend_cuda_context & ctx, ggml memcpy(&logit_softcap, (const float *) KQV->op_params + 2, sizeof(float)); if (Q->ne[1] == 1) { - constexpr int cols_per_block = 1; - constexpr int parallel_blocks = 4; + constexpr int cols_per_block = 1; if (logit_softcap == 0.0f) { constexpr bool use_logit_softcap = false; - ggml_cuda_flash_attn_ext_vec_f32_case_impl(ctx, dst); + ggml_cuda_flash_attn_ext_vec_f32_case_impl(ctx, dst); } else { constexpr bool use_logit_softcap = true; - ggml_cuda_flash_attn_ext_vec_f32_case_impl(ctx, dst); + ggml_cuda_flash_attn_ext_vec_f32_case_impl(ctx, dst); } return; } if (Q->ne[1] == 2) { - constexpr int cols_per_block = 2; - constexpr int parallel_blocks = 4; + constexpr int cols_per_block = 2; if (logit_softcap == 0.0f) { constexpr bool use_logit_softcap = false; - ggml_cuda_flash_attn_ext_vec_f32_case_impl(ctx, dst); + ggml_cuda_flash_attn_ext_vec_f32_case_impl(ctx, dst); } else { constexpr bool use_logit_softcap = true; - ggml_cuda_flash_attn_ext_vec_f32_case_impl(ctx, dst); + ggml_cuda_flash_attn_ext_vec_f32_case_impl(ctx, dst); } return; } if (Q->ne[1] <= 4) { - constexpr int cols_per_block = 4; - constexpr int parallel_blocks = 4; + constexpr int cols_per_block = 4; if (logit_softcap == 0.0f) { constexpr bool use_logit_softcap = false; - ggml_cuda_flash_attn_ext_vec_f32_case_impl(ctx, dst); + ggml_cuda_flash_attn_ext_vec_f32_case_impl(ctx, dst); } else { constexpr bool use_logit_softcap = true; - ggml_cuda_flash_attn_ext_vec_f32_case_impl(ctx, dst); + ggml_cuda_flash_attn_ext_vec_f32_case_impl(ctx, dst); } return; } - if (Q->ne[1] <= 8) { - constexpr int cols_per_block = 8; - constexpr int parallel_blocks = 4; - if (logit_softcap == 0.0f) { - constexpr bool use_logit_softcap = false; - ggml_cuda_flash_attn_ext_vec_f32_case_impl(ctx, dst); - } else { - constexpr bool use_logit_softcap = true; - ggml_cuda_flash_attn_ext_vec_f32_case_impl(ctx, dst); - } - return; - } - - constexpr int cols_per_block = 8; - constexpr int parallel_blocks = 1; + constexpr int cols_per_block = 8; if (logit_softcap == 0.0f) { constexpr bool use_logit_softcap = false; - ggml_cuda_flash_attn_ext_vec_f32_case_impl(ctx, dst); + ggml_cuda_flash_attn_ext_vec_f32_case_impl(ctx, dst); } else { constexpr bool use_logit_softcap = true; - ggml_cuda_flash_attn_ext_vec_f32_case_impl(ctx, dst); + ggml_cuda_flash_attn_ext_vec_f32_case_impl(ctx, dst); } } diff --git a/ggml/src/ggml-cuda/fattn-wmma-f16.cu b/ggml/src/ggml-cuda/fattn-wmma-f16.cu index 45702ad65..bc21b27a0 100644 --- a/ggml/src/ggml-cuda/fattn-wmma-f16.cu +++ b/ggml/src/ggml-cuda/fattn-wmma-f16.cu @@ -7,14 +7,19 @@ #include "fattn-wmma-f16.cuh" #ifdef FP16_MMA_AVAILABLE +#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) #include +namespace wmma = nvcuda::wmma; +#elif defined(GGML_HIP_ROCWMMA_FATTN) && defined(FP16_MMA_AVAILABLE) +#undef HIP_ENABLE_WARP_SYNC_BUILTINS // conflicts with rocWMMA headers +#include +namespace wmma = rocwmma; +#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) #endif // FP16_MMA_AVAILABLE // D == head size, VKQ_stride == num VKQ rows calculated in parallel: -template -#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) -__launch_bounds__(nwarps*WARP_SIZE, 1) -#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) +template +__launch_bounds__(nwarps*ggml_cuda_get_physical_warp_size(), 1) static __global__ void flash_attn_ext_f16( const char * __restrict__ Q, const char * __restrict__ K, @@ -51,7 +56,7 @@ static __global__ void flash_attn_ext_f16( const int ne1, const int ne2, const int ne3) { -#if __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA +#if defined(FLASH_ATTN_AVAILABLE) && (__CUDA_ARCH__ == GGML_CUDA_CC_VOLTA || (defined(GGML_HIP_ROCWMMA_FATTN) && defined(FP16_MMA_AVAILABLE))) // Skip unused kernel variants for faster compilation: if (use_logit_softcap && !(D == 128 || D == 256)) { NO_DEVICE_CODE; @@ -60,19 +65,20 @@ static __global__ void flash_attn_ext_f16( //In this kernel Q, K, V are matrices while i, j, k are matrix indices. - const int ic0 = ncols*(blockIdx.x / parallel_blocks); // Index of the first Q/QKV column to work on. - const int ip = blockIdx.x % parallel_blocks; // Index in group of blocks running for the same column in parallel. + constexpr int warp_size = ggml_cuda_get_physical_warp_size(); + + const int ic0 = ncols*blockIdx.x; // Index of the first Q/QKV column to work on. static_assert(D <= FATTN_KQ_STRIDE, "D must be <= FATTN_KQ_STRIDE."); static_assert(ncols == 8 || ncols % 16 == 0, "ncols must be 8 or a multiple of 16."); constexpr int frag_m = ncols == 8 ? 32 : 16; constexpr int frag_n = ncols == 8 ? 8 : 16; static_assert(D % frag_m == 0, "If ncols == 8 then D % frag_m must be 0."); - typedef nvcuda::wmma::fragment frag_a_K; - typedef nvcuda::wmma::fragment frag_a_V; - typedef nvcuda::wmma::fragment frag_b; - typedef nvcuda::wmma::fragment frag_c_KQ; - typedef nvcuda::wmma::fragment frag_c_VKQ; + typedef wmma::fragment frag_a_K; + typedef wmma::fragment frag_a_V; + typedef wmma::fragment frag_b; + typedef wmma::fragment frag_c_KQ; + typedef wmma::fragment frag_c_VKQ; constexpr int KQ_stride_tc = nwarps*frag_m; // Number of KQ rows calculated in parallel. constexpr int VKQ_ratio = KQ_stride_tc/VKQ_stride; // Number of parallel VKQ accumulators needed to keep all warps busy. @@ -84,16 +90,16 @@ static __global__ void flash_attn_ext_f16( constexpr int kqar = sizeof(KQ_acc_t)/sizeof(half); const int gqa_ratio = ne02 / ne12; // With grouped query attention there are > 1 Q matrices per K, V matrix. - const float * Q_f = (const float *) (Q + nb02* blockIdx.y + nb01*ic0); - const half * K_h = (const half *) (K + nb12*(blockIdx.y / gqa_ratio)); - const half * V_h = (const half *) (V + nb12*(blockIdx.y / gqa_ratio)); // K and V have same shape + const float * Q_f = (const float *) (Q + nb02* blockIdx.z + nb01*ic0); + const half * K_h = (const half *) (K + nb12*(blockIdx.z / gqa_ratio)); + const half * V_h = (const half *) (V + nb12*(blockIdx.z / gqa_ratio)); // K and V have same shape const half * maskh = (const half *) mask + (nb31/sizeof(half))* ic0; const half2 * mask2 = (const half2 *) mask + (nb31/sizeof(half))*(ic0/2); const int stride_Q = nb01 / sizeof(float); const int stride_KV = nb11 / sizeof(half); - const float slopef = get_alibi_slope(max_bias, blockIdx.y, n_head_log2, m0, m1); + const float slopef = get_alibi_slope(max_bias, blockIdx.z, n_head_log2, m0, m1); const half slopeh = __float2half(slopef); const half2 slope2 = make_half2(slopef, slopef); @@ -132,9 +138,9 @@ static __global__ void flash_attn_ext_f16( for (int j0 = 0; j0 < ncols; j0 += nwarps) { const int j = j0 + threadIdx.y; #pragma unroll - for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) { + for (int i0 = 0; i0 < D/2; i0 += warp_size) { const int i = i0 + threadIdx.x; - if (i0 + WARP_SIZE > D/2 && i >= D/2) { + if (i0 + warp_size > D/2 && i >= D/2) { break; } VKQ2[j*(D_padded/2) + i] = make_half2(0.0f, 0.0f); @@ -146,9 +152,9 @@ static __global__ void flash_attn_ext_f16( for (int j0 = 0; j0 < ncols; j0 += nwarps) { const int j = j0 + threadIdx.y; #pragma unroll - for (int i0 = 0; i0 < D; i0 += WARP_SIZE) { + for (int i0 = 0; i0 < D; i0 += warp_size) { const int i = i0 + threadIdx.x; - if (i0 + WARP_SIZE > D && i >= D) { + if (i0 + warp_size > D && i >= D) { break; } KQ[j*D_padded + i] = ic0 + j < ne01 ? Q_f[j*stride_Q + i] * scale : 0.0f; @@ -162,34 +168,34 @@ static __global__ void flash_attn_ext_f16( for (int i0 = 0; i0 < D; i0 += 16) { #pragma unroll for (int j0 = 0; j0 < ncols; j0 += frag_n) { - nvcuda::wmma::load_matrix_sync(Q_b[i0/16][j0/frag_n], KQ + j0*D_padded + i0, D_padded); + wmma::load_matrix_sync(Q_b[i0/16][j0/frag_n], KQ + j0*D_padded + i0, D_padded); } } __syncthreads(); // Iterate over ne11 == previous tokens: - for (int k_VKQ_0 = ip*FATTN_KQ_STRIDE; k_VKQ_0 < ne11; k_VKQ_0 += parallel_blocks*FATTN_KQ_STRIDE) { + for (int k_VKQ_0 = blockIdx.y*FATTN_KQ_STRIDE; k_VKQ_0 < ne11; k_VKQ_0 += gridDim.y*FATTN_KQ_STRIDE) { // Calculate tile of KQ: #pragma unroll for (int i_KQ_0 = 0; i_KQ_0 < FATTN_KQ_STRIDE; i_KQ_0 += KQ_stride_tc) { frag_c_KQ KQ_c[ncols/frag_n]; #pragma unroll for (int j = 0; j < ncols/frag_n; ++j) { - nvcuda::wmma::fill_fragment(KQ_c[j], 0.0f); + wmma::fill_fragment(KQ_c[j], static_cast(0.0f)); } #pragma unroll for (int k_KQ_0 = 0; k_KQ_0 < D; k_KQ_0 += 16) { frag_a_K K_a; - nvcuda::wmma::load_matrix_sync(K_a, K_h + (k_VKQ_0 + i_KQ_0 + frag_m*threadIdx.y)*stride_KV + k_KQ_0, stride_KV); + wmma::load_matrix_sync(K_a, K_h + (k_VKQ_0 + i_KQ_0 + frag_m*threadIdx.y)*stride_KV + k_KQ_0, stride_KV); #pragma unroll for (int j = 0; j < ncols/frag_n; ++j) { - nvcuda::wmma::mma_sync(KQ_c[j], K_a, Q_b[k_KQ_0/16][j], KQ_c[j]); + wmma::mma_sync(KQ_c[j], K_a, Q_b[k_KQ_0/16][j], KQ_c[j]); } } #pragma unroll for (int j0 = 0; j0 < ncols; j0 += frag_n) { - nvcuda::wmma::store_matrix_sync((KQ_acc_t *) KQ + j0*kqs_padded + i_KQ_0 + frag_m*threadIdx.y, KQ_c[j0/frag_n], kqs_padded, nvcuda::wmma::mem_col_major); + wmma::store_matrix_sync((KQ_acc_t *) KQ + j0*kqs_padded + i_KQ_0 + frag_m*threadIdx.y, KQ_c[j0/frag_n], kqs_padded, wmma::mem_col_major); } } @@ -202,27 +208,27 @@ static __global__ void flash_attn_ext_f16( const int j = j0 + threadIdx.y; if (std::is_same::value) { - float KQ_f_tmp[FATTN_KQ_STRIDE / WARP_SIZE]; + float KQ_f_tmp[FATTN_KQ_STRIDE / warp_size]; #pragma unroll - for (int k0 = 0; k0 < FATTN_KQ_STRIDE; k0 += WARP_SIZE) { + for (int k0 = 0; k0 < FATTN_KQ_STRIDE; k0 += warp_size) { const int k = k0 + threadIdx.x; - KQ_f_tmp[k0/WARP_SIZE] = KQ_f[j*kqs_padded + k]; + KQ_f_tmp[k0/warp_size] = KQ_f[j*kqs_padded + k]; if (use_logit_softcap) { - KQ_f_tmp[k0/WARP_SIZE] = logit_softcap*tanhf(KQ_f_tmp[k0/WARP_SIZE]); + KQ_f_tmp[k0/warp_size] = logit_softcap*tanhf(KQ_f_tmp[k0/warp_size]); } } float KQ_max_new = KQ_max_f[j0/nwarps]; #pragma unroll - for (int k0 = 0; k0 < FATTN_KQ_STRIDE; k0 += WARP_SIZE) { + for (int k0 = 0; k0 < FATTN_KQ_STRIDE; k0 += warp_size) { const int k = k0 + threadIdx.x; - KQ_f_tmp[k0/WARP_SIZE] += mask ? __half2float(slopeh*maskh[j*(nb31/sizeof(half)) + k_VKQ_0 + k]) : 0.0f; - KQ_max_new = max(KQ_max_new, KQ_f_tmp[k0/WARP_SIZE]); + KQ_f_tmp[k0/warp_size] += mask ? __half2float(slopeh*maskh[j*(nb31/sizeof(half)) + k_VKQ_0 + k]) : 0.0f; + KQ_max_new = max(KQ_max_new, KQ_f_tmp[k0/warp_size]); } - KQ_max_new = warp_reduce_max(KQ_max_new); + KQ_max_new = warp_reduce_max(KQ_max_new); const float diff = KQ_max_f[j0/nwarps] - KQ_max_new; KQ_max_scale_f[j0/nwarps] = expf(diff); @@ -233,48 +239,48 @@ static __global__ void flash_attn_ext_f16( float KQ_rowsum_add = 0.0f; #pragma unroll - for (int k0 = 0; k0 < FATTN_KQ_STRIDE; k0 += WARP_SIZE) { + for (int k0 = 0; k0 < FATTN_KQ_STRIDE; k0 += warp_size) { const int k = k0 + threadIdx.x; - const float diff = KQ_f_tmp[k0/WARP_SIZE] - KQ_max_f[j0/nwarps]; - KQ_f_tmp[k0/WARP_SIZE] = expf(diff); + const float diff = KQ_f_tmp[k0/warp_size] - KQ_max_f[j0/nwarps]; + KQ_f_tmp[k0/warp_size] = expf(diff); if (diff <= SOFTMAX_FTZ_THRESHOLD) { - KQ_f_tmp[k0/WARP_SIZE] = 0.0f; + KQ_f_tmp[k0/warp_size] = 0.0f; } - KQ_rowsum_add += KQ_f_tmp[k0/WARP_SIZE]; - KQ[j*(kqar*kqs_padded) + k] = KQ_f_tmp[k0/WARP_SIZE]; + KQ_rowsum_add += KQ_f_tmp[k0/warp_size]; + KQ[j*(kqar*kqs_padded) + k] = KQ_f_tmp[k0/warp_size]; } - KQ_rowsum_add = warp_reduce_sum(KQ_rowsum_add); + KQ_rowsum_add = warp_reduce_sum(KQ_rowsum_add); // Scale previous KQ_rowsum to account for a potential increase in KQ_max: KQ_rowsum_f[j0/nwarps] = KQ_max_scale_f[j0/nwarps]*KQ_rowsum_f[j0/nwarps] + KQ_rowsum_add; } else { - half2 KQ2_tmp[FATTN_KQ_STRIDE/(2*WARP_SIZE)]; + half2 KQ2_tmp[FATTN_KQ_STRIDE/(2*warp_size)]; #pragma unroll - for (int k0 = 0; k0 < FATTN_KQ_STRIDE/2; k0 += WARP_SIZE) { + for (int k0 = 0; k0 < FATTN_KQ_STRIDE/2; k0 += warp_size) { const int k = k0 + threadIdx.x; - KQ2_tmp[k0/WARP_SIZE] = KQ2[j*(kqs_padded/2) + k]; + KQ2_tmp[k0/warp_size] = KQ2[j*(kqs_padded/2) + k]; if (use_logit_softcap) { // There is no dedicated tangens hyperbolicus function for half2. - KQ2_tmp[k0/WARP_SIZE] = h2exp(KQ2_tmp[k0/WARP_SIZE]*make_half2(2.0f, 2.0f)); - KQ2_tmp[k0/WARP_SIZE] = (KQ2_tmp[k0/WARP_SIZE] - make_half2(1.0f, 1.0f)) - /(KQ2_tmp[k0/WARP_SIZE] + make_half2(1.0f, 1.0f)); + KQ2_tmp[k0/warp_size] = h2exp(KQ2_tmp[k0/warp_size]*make_half2(2.0f, 2.0f)); + KQ2_tmp[k0/warp_size] = (KQ2_tmp[k0/warp_size] - make_half2(1.0f, 1.0f)) + /(KQ2_tmp[k0/warp_size] + make_half2(1.0f, 1.0f)); - KQ2_tmp[k0/WARP_SIZE] *= logit_softcap_2; + KQ2_tmp[k0/warp_size] *= logit_softcap_2; } } half2 KQ_max_new = KQ_max_h2[j0/nwarps]; #pragma unroll - for (int k0 = 0; k0 < FATTN_KQ_STRIDE/2; k0 += WARP_SIZE) { + for (int k0 = 0; k0 < FATTN_KQ_STRIDE/2; k0 += warp_size) { const int k = k0 + threadIdx.x; - KQ2_tmp[k0/WARP_SIZE] += mask ? slope2*mask2[(j*ne11 + k_VKQ_0)/2 + k] : make_half2(0.0f, 0.0f); - KQ_max_new = ggml_cuda_hmax2(KQ_max_new, KQ2_tmp[k0/WARP_SIZE]); + KQ2_tmp[k0/warp_size] += mask ? slope2*mask2[(j*ne11 + k_VKQ_0)/2 + k] : make_half2(0.0f, 0.0f); + KQ_max_new = ggml_cuda_hmax2(KQ_max_new, KQ2_tmp[k0/warp_size]); } - KQ_max_new = __half2half2(warp_reduce_max(ggml_cuda_hmax(__low2half(KQ_max_new), __high2half(KQ_max_new)))); + KQ_max_new = __half2half2(warp_reduce_max(ggml_cuda_hmax(__low2half(KQ_max_new), __high2half(KQ_max_new)))); const half2 diff = KQ_max_h2[j0/nwarps] - KQ_max_new; KQ_max_scale_h2[j0/nwarps] = h2exp(diff); const uint32_t ftz_mask = __hgt2_mask(diff, make_half2(SOFTMAX_FTZ_THRESHOLD, SOFTMAX_FTZ_THRESHOLD)); @@ -283,17 +289,17 @@ static __global__ void flash_attn_ext_f16( half2 KQ_rowsum_add = make_half2(0.0f, 0.0f); #pragma unroll - for (int k0 = 0; k0 < FATTN_KQ_STRIDE/2; k0 += WARP_SIZE) { + for (int k0 = 0; k0 < FATTN_KQ_STRIDE/2; k0 += warp_size) { const int k = k0 + threadIdx.x; - const half2 diff = KQ2_tmp[k0/WARP_SIZE] - KQ_max_h2[j0/nwarps]; - KQ2_tmp[k0/WARP_SIZE] = h2exp(diff); + const half2 diff = KQ2_tmp[k0/warp_size] - KQ_max_h2[j0/nwarps]; + KQ2_tmp[k0/warp_size] = h2exp(diff); const uint32_t ftz_mask = __hgt2_mask(diff, make_half2(SOFTMAX_FTZ_THRESHOLD, SOFTMAX_FTZ_THRESHOLD)); - *((uint32_t *) &KQ2_tmp[k0/WARP_SIZE]) &= ftz_mask; - KQ_rowsum_add += KQ2_tmp[k0/WARP_SIZE]; - KQ2[j*(kqs_padded/2) + k] = KQ2_tmp[k0/WARP_SIZE]; + *((uint32_t *) &KQ2_tmp[k0/warp_size]) &= ftz_mask; + KQ_rowsum_add += KQ2_tmp[k0/warp_size]; + KQ2[j*(kqs_padded/2) + k] = KQ2_tmp[k0/warp_size]; } - KQ_rowsum_add = warp_reduce_sum(KQ_rowsum_add); + KQ_rowsum_add = warp_reduce_sum(KQ_rowsum_add); // Scale previous KQ_rowsum to account for a potential increase in KQ_max: KQ_rowsum_h2[j0/nwarps] = KQ_max_scale_h2[j0/nwarps]*KQ_rowsum_h2[j0/nwarps] + KQ_rowsum_add; @@ -308,7 +314,7 @@ static __global__ void flash_attn_ext_f16( #pragma unroll for (int k0 = 0; k0 < FATTN_KQ_STRIDE; k0 += VKQ_ratio*16) { const int k = k0 + (threadIdx.y % VKQ_ratio)*16; - nvcuda::wmma::load_matrix_sync( + wmma::load_matrix_sync( KQ_b[k0/(VKQ_ratio*16)][j0/frag_n], KQ + j0*(kqar*kqs_padded) + k, kqar*kqs_padded); @@ -320,7 +326,7 @@ static __global__ void flash_attn_ext_f16( for (int i_VKQ_0 = 0; i_VKQ_0 < D; i_VKQ_0 += VKQ_stride) { #pragma unroll for (int j = 0; j < ncols/frag_n; ++j) { - nvcuda::wmma::fill_fragment(VKQ_c[i_VKQ_0/VKQ_stride][j], 0.0f); + wmma::fill_fragment(VKQ_c[i_VKQ_0/VKQ_stride][j], static_cast(0.0f)); } #pragma unroll @@ -328,10 +334,10 @@ static __global__ void flash_attn_ext_f16( const int k = k0 + (threadIdx.y % VKQ_ratio)*16; frag_a_V v_a; - nvcuda::wmma::load_matrix_sync(v_a, V_h + (k_VKQ_0 + k)*stride_KV + i_VKQ_0 + frag_m*(threadIdx.y/VKQ_ratio), stride_KV); + wmma::load_matrix_sync(v_a, V_h + (k_VKQ_0 + k)*stride_KV + i_VKQ_0 + frag_m*(threadIdx.y/VKQ_ratio), stride_KV); #pragma unroll for (int j = 0; j < ncols/frag_n; ++j) { - nvcuda::wmma::mma_sync(VKQ_c[i_VKQ_0/VKQ_stride][j], v_a, KQ_b[k0/(VKQ_ratio*16)][j], VKQ_c[i_VKQ_0/VKQ_stride][j]); + wmma::mma_sync(VKQ_c[i_VKQ_0/VKQ_stride][j], v_a, KQ_b[k0/(VKQ_ratio*16)][j], VKQ_c[i_VKQ_0/VKQ_stride][j]); } } } @@ -343,10 +349,10 @@ static __global__ void flash_attn_ext_f16( for (int i_KQ_0 = 0; i_KQ_0 < D; i_KQ_0 += VKQ_stride) { #pragma unroll for (int j0 = 0; j0 < ncols; j0 += frag_n) { - nvcuda::wmma::store_matrix_sync( + wmma::store_matrix_sync( KQ + offset_k + j0*D_padded + i_KQ_0 + frag_m*(threadIdx.y/VKQ_ratio), VKQ_c[i_KQ_0/VKQ_stride][j0/frag_n], - D_padded, nvcuda::wmma::mem_col_major); + D_padded, wmma::mem_col_major); } } @@ -364,9 +370,9 @@ static __global__ void flash_attn_ext_f16( } #pragma unroll - for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) { + for (int i0 = 0; i0 < D/2; i0 += warp_size) { const int i = i0 + threadIdx.x; - if (i0 + WARP_SIZE > D/2 && i >= D/2) { + if (i0 + warp_size > D/2 && i >= D/2) { break; } @@ -388,7 +394,7 @@ static __global__ void flash_attn_ext_f16( if (ic0 + j_VKQ >= ne01) { return; } - const int j_dst = (ic0 + j_VKQ)*parallel_blocks + ip; + const int j_dst = (ic0 + j_VKQ)*gridDim.y + blockIdx.y; float KQ_rowsum_j; if (std::is_same::value) { @@ -398,19 +404,19 @@ static __global__ void flash_attn_ext_f16( } #pragma unroll - for (int i0 = 0; i0 < D; i0 += WARP_SIZE) { + for (int i0 = 0; i0 < D; i0 += warp_size) { const int i = i0 + threadIdx.x; - if (i0 + WARP_SIZE > D && i >= D) { + if (i0 + warp_size > D && i >= D) { break; } float dst_val = VKQ[j_VKQ*D_padded + i]; - if (parallel_blocks == 1) { + if (gridDim.y == 1) { dst_val /= KQ_rowsum_j; } - dst[j_dst*gridDim.y*D + blockIdx.y*D + i] = dst_val; + dst[j_dst*gridDim.z*D + blockIdx.z*D + i] = dst_val; } - if (parallel_blocks == 1 || threadIdx.x != 0) { + if (gridDim.y == 1 || threadIdx.x != 0) { continue; } @@ -421,11 +427,21 @@ static __global__ void flash_attn_ext_f16( dst_meta_val.x = __low2float(KQ_max_h2[j0/nwarps]); } dst_meta_val.y = KQ_rowsum_j; - dst_meta[(ic0 + j_VKQ)*gridDim.y*parallel_blocks + blockIdx.y*parallel_blocks + ip] = dst_meta_val; + dst_meta[((ic0 + j_VKQ)*gridDim.z + blockIdx.z) * gridDim.y + blockIdx.y] = dst_meta_val; } #else - NO_DEVICE_CODE; -#endif // __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA + GGML_UNUSED(Q); GGML_UNUSED(K); GGML_UNUSED(V); GGML_UNUSED(mask); + GGML_UNUSED(dst); GGML_UNUSED(dst_meta); GGML_UNUSED(scale); + GGML_UNUSED(max_bias); GGML_UNUSED(m0); GGML_UNUSED(m1); + GGML_UNUSED(n_head_log2); GGML_UNUSED(logit_softcap); + GGML_UNUSED(ne00); GGML_UNUSED(ne01); GGML_UNUSED(ne02); GGML_UNUSED(ne03); + GGML_UNUSED(ne10); GGML_UNUSED(ne11); GGML_UNUSED(ne12); GGML_UNUSED(ne13); + GGML_UNUSED(ne31); GGML_UNUSED(nb31); GGML_UNUSED(nb01); GGML_UNUSED(nb02); + GGML_UNUSED(nb03); GGML_UNUSED(nb11); GGML_UNUSED(nb12); GGML_UNUSED(nb13); + GGML_UNUSED(nb21); GGML_UNUSED(nb22); GGML_UNUSED(nb23); + GGML_UNUSED(ne0); GGML_UNUSED(ne1); GGML_UNUSED(ne2); GGML_UNUSED(ne3); + NO_DEVICE_CODE; +#endif // defined(FLASH_ATTN_AVAILABLE) && (__CUDA_ARCH__ == GGML_CUDA_CC_VOLTA || (defined(GGML_HIP_ROCWMMA_FATTN) && defined(FP16_MMA_AVAILABLE))) } constexpr int get_max_power_of_2(int x) { @@ -455,59 +471,26 @@ static_assert(get_VKQ_stride( 80, 4, 16) == 16, "Test failed."); template void ggml_cuda_flash_attn_ext_wmma_f16_case(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { const ggml_tensor * KQV = dst; - const ggml_tensor * Q = dst->src[0]; constexpr int nwarps = 4; constexpr int frag_m = cols_per_block == 8 && D % 32 == 0 ? 32 : 16; - const int blocks_num_pb1 = ((Q->ne[1] + cols_per_block - 1) / cols_per_block)*Q->ne[2]*Q->ne[3]; - const int nsm = ggml_cuda_info().devices[ggml_cuda_get_device()].nsm; + const int warp_size = ggml_cuda_info().devices[ggml_cuda_get_device()].warp_size; float logit_softcap; memcpy(&logit_softcap, (const float *) KQV->op_params + 2, sizeof(float)); - if (4*blocks_num_pb1 < 2*nsm) { - constexpr int parallel_blocks = 4; - fattn_kernel_t fattn_kernel; - if (logit_softcap == 0.0f) { - constexpr bool use_logit_softcap = false; - fattn_kernel = flash_attn_ext_f16< - D, cols_per_block, nwarps, get_VKQ_stride(D, nwarps, frag_m), parallel_blocks, KQ_acc_t, use_logit_softcap>; - } else { - constexpr bool use_logit_softcap = true; - fattn_kernel = flash_attn_ext_f16< - D, cols_per_block, nwarps, get_VKQ_stride(D, nwarps, frag_m), parallel_blocks, KQ_acc_t, use_logit_softcap>; - } - launch_fattn(ctx, dst, fattn_kernel, nwarps, 0, true, true); - return; - } - if (2*blocks_num_pb1 < 2*nsm) { - constexpr int parallel_blocks = 2; - fattn_kernel_t fattn_kernel; - if (logit_softcap == 0.0f) { - constexpr bool use_logit_softcap = false; - fattn_kernel = flash_attn_ext_f16< - D, cols_per_block, nwarps, get_VKQ_stride(D, nwarps, frag_m), parallel_blocks, KQ_acc_t, use_logit_softcap>; - } else { - constexpr bool use_logit_softcap = true; - fattn_kernel = flash_attn_ext_f16< - D, cols_per_block, nwarps, get_VKQ_stride(D, nwarps, frag_m), parallel_blocks, KQ_acc_t, use_logit_softcap>; - } - launch_fattn(ctx, dst, fattn_kernel, nwarps, 0, true, true); - return; - } - constexpr int parallel_blocks = 1; fattn_kernel_t fattn_kernel; if (logit_softcap == 0.0f) { constexpr bool use_logit_softcap = false; fattn_kernel = flash_attn_ext_f16< - D, cols_per_block, nwarps, get_VKQ_stride(D, nwarps, frag_m), parallel_blocks, KQ_acc_t, use_logit_softcap>; + D, cols_per_block, nwarps, get_VKQ_stride(D, nwarps, frag_m), KQ_acc_t, use_logit_softcap>; } else { constexpr bool use_logit_softcap = true; fattn_kernel = flash_attn_ext_f16< - D, cols_per_block, nwarps, get_VKQ_stride(D, nwarps, frag_m), parallel_blocks, KQ_acc_t, use_logit_softcap>; + D, cols_per_block, nwarps, get_VKQ_stride(D, nwarps, frag_m), KQ_acc_t, use_logit_softcap>; } - launch_fattn(ctx, dst, fattn_kernel, nwarps, 0, true, true); + launch_fattn(ctx, dst, fattn_kernel, nwarps, 0, FATTN_KQ_STRIDE, true, true, false, warp_size); } void ggml_cuda_flash_attn_ext_wmma_f16(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { @@ -515,6 +498,7 @@ void ggml_cuda_flash_attn_ext_wmma_f16(ggml_backend_cuda_context & ctx, ggml_ten const ggml_tensor * Q = dst->src[0]; const enum ggml_prec prec = ggml_flash_attn_ext_get_prec(KQV); + const int warp_size = ggml_cuda_info().devices[ctx.device].warp_size; if (prec != GGML_PREC_DEFAULT) { if (Q->ne[1] <= 32 || Q->ne[0] > 128) { @@ -571,7 +555,8 @@ void ggml_cuda_flash_attn_ext_wmma_f16(ggml_backend_cuda_context & ctx, ggml_ten return; } - if (Q->ne[1] <= 8 && Q->ne[0] % WARP_SIZE == 0) { +#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) + if (Q->ne[1] <= 8 && Q->ne[0] % warp_size == 0) { constexpr int cols_per_block = 8; switch (Q->ne[0]) { case 64: @@ -592,6 +577,7 @@ void ggml_cuda_flash_attn_ext_wmma_f16(ggml_backend_cuda_context & ctx, ggml_ten } return; } +#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) if (Q->ne[1] <= 32) { constexpr int cols_per_block = 16; diff --git a/ggml/src/ggml-cuda/fattn.cu b/ggml/src/ggml-cuda/fattn.cu index b0cf152f5..7a2d1e453 100644 --- a/ggml/src/ggml-cuda/fattn.cu +++ b/ggml/src/ggml-cuda/fattn.cu @@ -8,28 +8,50 @@ #include "fattn-wmma-f16.cuh" #include "fattn.cuh" -template +template +static void ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { + const ggml_tensor * Q = dst->src[0]; + + if (Q->ne[1] <= 8/ncols2) { + ggml_cuda_flash_attn_ext_mma_f16_case(ctx, dst); + return; + } + + if (Q->ne[1] <= 16/ncols2) { + ggml_cuda_flash_attn_ext_mma_f16_case(ctx, dst); + return; + } + + if (Q->ne[1] <= 32/ncols2) { + ggml_cuda_flash_attn_ext_mma_f16_case(ctx, dst); + return; + } + + ggml_cuda_flash_attn_ext_mma_f16_case(ctx, dst); +} + +template static void ggml_cuda_flash_attn_ext_mma_f16_switch_hs(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { const ggml_tensor * Q = dst->src[0]; switch (Q->ne[0]) { case 64: - ggml_cuda_flash_attn_ext_mma_f16_case< 64, cols_per_block>(ctx, dst); + ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1< 64, ncols2>(ctx, dst); break; case 80: - ggml_cuda_flash_attn_ext_mma_f16_case< 80, cols_per_block>(ctx, dst); + ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1< 80, ncols2>(ctx, dst); break; case 96: - ggml_cuda_flash_attn_ext_mma_f16_case< 96, cols_per_block>(ctx, dst); + ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1< 96, ncols2>(ctx, dst); break; case 112: - ggml_cuda_flash_attn_ext_mma_f16_case<112, cols_per_block>(ctx, dst); + ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<112, ncols2>(ctx, dst); break; case 128: - ggml_cuda_flash_attn_ext_mma_f16_case<128, cols_per_block>(ctx, dst); + ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<128, ncols2>(ctx, dst); break; case 256: - ggml_cuda_flash_attn_ext_mma_f16_case<256, cols_per_block>(ctx, dst); + ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<256, ncols2>(ctx, dst); break; default: GGML_ABORT("fatal error"); @@ -38,24 +60,35 @@ static void ggml_cuda_flash_attn_ext_mma_f16_switch_hs(ggml_backend_cuda_context } static void ggml_cuda_flash_attn_ext_mma_f16(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { - const ggml_tensor * Q = dst->src[0]; + const ggml_tensor * KQV = dst; + const ggml_tensor * Q = dst->src[0]; + const ggml_tensor * K = dst->src[1]; + const ggml_tensor * mask = dst->src[3]; - if (Q->ne[1] <= 8) { + float max_bias = 0.0f; + memcpy(&max_bias, (const float *) KQV->op_params + 1, sizeof(float)); + + const float use_gqa_opt = mask && max_bias == 0.0f; + + GGML_ASSERT(Q->ne[2] % K->ne[2] == 0); + const int gqa_ratio = Q->ne[2] / K->ne[2]; + + if (use_gqa_opt && gqa_ratio % 8 == 0) { ggml_cuda_flash_attn_ext_mma_f16_switch_hs<8>(ctx, dst); return; } - if (Q->ne[1] <= 16) { - ggml_cuda_flash_attn_ext_mma_f16_switch_hs<16>(ctx, dst); + if (use_gqa_opt && gqa_ratio == 4) { + ggml_cuda_flash_attn_ext_mma_f16_switch_hs<4>(ctx, dst); return; } - if (Q->ne[1] <= 32) { - ggml_cuda_flash_attn_ext_mma_f16_switch_hs<32>(ctx, dst); + if (use_gqa_opt && gqa_ratio == 2) { + ggml_cuda_flash_attn_ext_mma_f16_switch_hs<2>(ctx, dst); return; } - ggml_cuda_flash_attn_ext_mma_f16_switch_hs<64>(ctx, dst); + ggml_cuda_flash_attn_ext_mma_f16_switch_hs<1>(ctx, dst); } #define FATTN_VEC_F16_CASE(D, type_K, type_V) \ @@ -209,15 +242,26 @@ static void ggml_cuda_flash_attn_ext_vec_f32(ggml_backend_cuda_context & ctx, gg } void ggml_cuda_flash_attn_ext(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { - const ggml_tensor * KQV = dst; - const ggml_tensor * Q = dst->src[0]; + const ggml_tensor * KQV = dst; + const ggml_tensor * Q = dst->src[0]; + const ggml_tensor * K = dst->src[1]; + const ggml_tensor * V = dst->src[2]; + const ggml_tensor * mask = dst->src[3]; ggml_cuda_set_device(ctx.device); const int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc; + const int warp_size = ggml_cuda_info().devices[ggml_cuda_get_device()].warp_size; const enum ggml_prec prec = ggml_flash_attn_ext_get_prec(KQV); - // On AMD the tile kernels perform poorly, use the vec kernel instead: - if (cc >= GGML_CUDA_CC_OFFSET_AMD) { + if (GGML_CUDA_CC_IS_AMD(cc)) { +#if defined(GGML_HIP_ROCWMMA_FATTN) + if (fp16_mma_available(cc)) { + ggml_cuda_flash_attn_ext_wmma_f16(ctx, dst); + return; + } +#endif // defined(GGML_HIP_ROCWMMA_FATTN) + + // On AMD the tile kernels perform poorly, use the vec kernel instead: if (prec == GGML_PREC_DEFAULT && fast_fp16_available(cc)) { ggml_cuda_flash_attn_ext_vec_f16(ctx, dst); } else { @@ -237,13 +281,13 @@ void ggml_cuda_flash_attn_ext(ggml_backend_cuda_context & ctx, ggml_tensor * dst if (!fp16_mma_available(cc)) { if (prec == GGML_PREC_DEFAULT) { - if (Q->ne[1] <= 8) { + if (Q->ne[1] <= 8 || Q->ne[0] == 256) { ggml_cuda_flash_attn_ext_vec_f16(ctx, dst); } else { ggml_cuda_flash_attn_ext_tile_f16(ctx, dst); } } else { - if (Q->ne[1] <= 8) { + if (Q->ne[1] <= 8 || Q->ne[0] == 256) { ggml_cuda_flash_attn_ext_vec_f32(ctx, dst); } else { ggml_cuda_flash_attn_ext_tile_f32(ctx, dst); @@ -252,18 +296,21 @@ void ggml_cuda_flash_attn_ext(ggml_backend_cuda_context & ctx, ggml_tensor * dst return; } - if (Q->ne[1] == 1 && Q->ne[0] % (2*WARP_SIZE) == 0) { + const bool gqa_opt_applies = ((Q->ne[2] / K->ne[2]) % 2 == 0) && mask; // The mma-based kernels have GQA-specific optimizations + const bool mma_needs_data_conversion = K->type != GGML_TYPE_F16 || V->type != GGML_TYPE_F16; + const bool mma_faster_for_bs1 = new_mma_available(cc) && gqa_opt_applies && cc < GGML_CUDA_CC_ADA_LOVELACE && !mma_needs_data_conversion; + const bool can_use_vector_kernel = Q->ne[0] % (2*warp_size) == 0; + if (Q->ne[1] == 1 && can_use_vector_kernel && !mma_faster_for_bs1) { if (prec == GGML_PREC_DEFAULT) { ggml_cuda_flash_attn_ext_vec_f16(ctx, dst); - return; - } else if(Q->ne[0] <= 128) { + } else { ggml_cuda_flash_attn_ext_vec_f32(ctx, dst); - return; } + return; } // The MMA implementation needs Turing or newer, use the old WMMA code for Volta: - if (cc == GGML_CUDA_CC_VOLTA) { + if (fp16_mma_available(cc) && !new_mma_available(cc)) { ggml_cuda_flash_attn_ext_wmma_f16(ctx, dst); return; } diff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cu index 4dbaefdba..fafe9633e 100644 --- a/ggml/src/ggml-cuda/ggml-cuda.cu +++ b/ggml/src/ggml-cuda/ggml-cuda.cu @@ -31,12 +31,14 @@ #include "ggml-cuda/rope.cuh" #include "ggml-cuda/scale.cuh" #include "ggml-cuda/softmax.cuh" +#include "ggml-cuda/ssm-conv.cuh" +#include "ggml-cuda/ssm-scan.cuh" #include "ggml-cuda/sum.cuh" #include "ggml-cuda/sumrows.cuh" #include "ggml-cuda/tsembd.cuh" #include "ggml-cuda/unary.cuh" #include "ggml-cuda/upscale.cuh" -#include "ggml-cuda/wkv6.cuh" +#include "ggml-cuda/wkv.cuh" #include "ggml-cuda/gla.cuh" #include "ggml.h" @@ -178,11 +180,11 @@ static ggml_cuda_device_info ggml_cuda_init() { int major_version = 0; size_t version_length = 0; if (rocblas_get_version_string_size(&version_length) == rocblas_status_success) { - std::string version(version_length, '\0'); + std::vector version(version_length+1, '\0'); if (rocblas_get_version_string(version.data(), version.size()) == rocblas_status_success) { - version.resize(::strlen(version.c_str())); + version.resize(::strlen(version.data())); int parsed_value = 0; - if (std::from_chars(version.c_str(), version.c_str() + version.length(), parsed_value).ec == std::errc()) { + if (std::from_chars(version.data(), version.data() + version.size(), parsed_value).ec == std::errc()) { major_version = parsed_value; } } @@ -261,6 +263,14 @@ static ggml_cuda_device_info ggml_cuda_init() { GGML_LOG_INFO(" Device %d: %s, %s (0x%x), VMM: %s, Wave Size: %d\n", id, prop.name, prop.gcnArchName, info.devices[id].cc & 0xffff, device_vmm ? "yes" : "no", prop.warpSize); +#elif defined(GGML_USE_MUSA) + // FIXME: Ensure compatibility with varying warp sizes across different MUSA archs. + info.devices[id].warp_size = 32; + info.devices[id].smpbo = prop.sharedMemPerBlockOptin; + info.devices[id].cc = GGML_CUDA_CC_OFFSET_MTHREADS + prop.major * 0x100; + info.devices[id].cc += prop.minor * 0x10; + GGML_LOG_INFO(" Device %d: %s, compute capability %d.%d, VMM: %s\n", + id, prop.name, prop.major, prop.minor, device_vmm ? "yes" : "no"); #else info.devices[id].smpbo = prop.sharedMemPerBlockOptin; info.devices[id].cc = 100*prop.major + 10*prop.minor; @@ -534,12 +544,12 @@ static void * ggml_backend_cuda_buffer_get_base(ggml_backend_buffer_t buffer) { return ctx->dev_ptr; } -static void ggml_backend_cuda_buffer_init_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor) { +static enum ggml_status ggml_backend_cuda_buffer_init_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor) { ggml_backend_cuda_buffer_context * ctx = (ggml_backend_cuda_buffer_context *)buffer->context; if (tensor->view_src != NULL) { assert(tensor->view_src->buffer->buft == buffer->buft); - return; + return GGML_STATUS_SUCCESS; } if (ggml_is_quantized(tensor->type) && tensor->view_src == nullptr && ggml_backend_buffer_get_usage(buffer) != GGML_BACKEND_BUFFER_USAGE_COMPUTE) { @@ -552,6 +562,7 @@ static void ggml_backend_cuda_buffer_init_tensor(ggml_backend_buffer_t buffer, g CUDA_CHECK(cudaMemset((char *)tensor->data + original_size, 0, padded_size - original_size)); } } + return GGML_STATUS_SUCCESS; } static void ggml_backend_cuda_buffer_memset_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor, uint8_t value, size_t offset, size_t size) { @@ -786,7 +797,7 @@ static void * ggml_backend_cuda_split_buffer_get_base(ggml_backend_buffer_t buff GGML_UNUSED(buffer); } -static void ggml_backend_cuda_split_buffer_init_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor) { +static enum ggml_status ggml_backend_cuda_split_buffer_init_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor) { GGML_ASSERT(tensor->view_src == nullptr); // views of split tensors are not supported ggml_backend_cuda_split_buffer_context * ctx = (ggml_backend_cuda_split_buffer_context *)buffer->context; @@ -832,6 +843,7 @@ static void ggml_backend_cuda_split_buffer_init_tensor(ggml_backend_buffer_t buf } } tensor->extra = extra; + return GGML_STATUS_SUCCESS; } static void ggml_backend_cuda_split_buffer_set_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor, const void * data, size_t offset, size_t size) { @@ -1178,11 +1190,39 @@ static void ggml_cuda_op_mul_mat_cublas( // ldc == nrows of the matrix that cuBLAS writes into int64_t ldc = id == ctx.device ? ne0 : row_diff; - const int compute_capability = ggml_cuda_info().devices[id].cc; + const int cc = ggml_cuda_info().devices[id].cc; const bool use_fp16 = (src0->type == GGML_TYPE_F16 || ggml_is_quantized(src0->type)) && ggml_is_contiguous(src0) && row_diff == src0->ne[1] && dst->op_params[0] == GGML_PREC_DEFAULT; - if (compute_capability >= GGML_CUDA_CC_VOLTA && use_fp16) { + if (src0->type == GGML_TYPE_BF16 && ggml_is_contiguous(src0) && row_diff == src0->ne[1]) { + ggml_cuda_pool_alloc src1_as_bf16(ctx.pool(id)); + if (src1->type != GGML_TYPE_BF16) { + const to_bf16_cuda_t to_bf16_cuda = ggml_get_to_bf16_cuda(src1->type); + GGML_ASSERT(to_bf16_cuda != nullptr); + size_t ne = src1_ncols*ne10; + src1_as_bf16.alloc(ne); + to_bf16_cuda(src1_ddf_i, src1_as_bf16.get(), ne, stream); + } + const nv_bfloat16 * src1_ptr = src1->type == GGML_TYPE_BF16 ? (const nv_bfloat16 *) src1_ddf_i : src1_as_bf16.get(); + const nv_bfloat16 * src0_ptr = (const nv_bfloat16 *)src0_dd_i; + ggml_cuda_pool_alloc dst_bf16(ctx.pool(id), row_diff*src1_ncols); + + const float alpha_f32 = 1.0f; + const float beta_f32 = 0.0f; + + CUBLAS_CHECK(cublasSetStream(ctx.cublas_handle(id), stream)); + CUBLAS_CHECK( + cublasGemmEx(ctx.cublas_handle(id), CUBLAS_OP_T, CUBLAS_OP_N, + row_diff, src1_ncols, ne10, + &alpha_f32, src0_ptr, CUDA_R_16BF, ne00, + src1_ptr, CUDA_R_16BF, ne10, + &beta_f32, dst_bf16.get(), CUDA_R_16BF, ldc, + CUBLAS_COMPUTE_32F, + CUBLAS_GEMM_DEFAULT_TENSOR_OP)); + + const to_fp32_cuda_t to_fp32_cuda = ggml_get_to_fp32_cuda(GGML_TYPE_BF16); + to_fp32_cuda(dst_bf16.get(), dst_dd_i, row_diff*src1_ncols, stream); + } else if (((GGML_CUDA_CC_IS_NVIDIA(cc) && cc >= GGML_CUDA_CC_VOLTA) || GGML_CUDA_CC_IS_AMD(cc)) && use_fp16) { // convert src0 and src1 to fp16, multiply as fp16, convert dst to fp32 ggml_cuda_pool_alloc src0_as_f16(ctx.pool(id)); if (src0->type != GGML_TYPE_F16) { @@ -1206,7 +1246,7 @@ static void ggml_cuda_op_mul_mat_cublas( CUBLAS_CHECK(cublasSetStream(ctx.cublas_handle(id), stream)); - if (GGML_CUDA_CC_IS_CDNA(compute_capability)) { + if (GGML_CUDA_CC_IS_CDNA(cc) || GGML_CUDA_CC_IS_RDNA4(cc)) { const float alpha = 1.0f; const float beta = 0.0f; CUBLAS_CHECK( @@ -1480,12 +1520,7 @@ static void ggml_cuda_op_mul_mat( const size_t nbytes_data = ggml_nbytes(src0); const size_t nbytes_padding = ggml_row_size(src0->type, MATRIX_ROW_PADDING - ne00 % MATRIX_ROW_PADDING); dev[id].src0_dd = dev[id].src0_dd_alloc.alloc(ctx.pool(id), nbytes_data + nbytes_padding); - // TODO: remove this for MUSA once the Guilty Lockup issue is resolved -#ifndef GGML_USE_MUSA CUDA_CHECK(cudaMemsetAsync(dev[id].src0_dd, 0, nbytes_data + nbytes_padding, stream)); -#else // GGML_USE_MUSA - CUDA_CHECK(cudaMemsetAsync(dev[id].src0_dd + nbytes_data, 0, nbytes_padding, stream)); -#endif // !GGML_USE_MUSA } // If src0 is on a temporary compute buffer (partial offloading) there may be some padding that needs to be cleared: @@ -1754,7 +1789,9 @@ static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, co beta = &beta_f32; } - if (GGML_CUDA_CC_IS_CDNA(ggml_cuda_info().devices[ctx.device].cc)) { + int id = ggml_cuda_get_device(); + const int cc = ggml_cuda_info().devices[id].cc; + if (GGML_CUDA_CC_IS_CDNA(cc) || GGML_CUDA_CC_IS_RDNA4(cc)) { cu_compute_type = CUBLAS_COMPUTE_32F; alpha = &alpha_f32; beta = &beta_f32; @@ -1787,9 +1824,6 @@ static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, co } } #else -#ifdef GGML_USE_MUSA - GGML_ASSERT(false); -#else // !GGML_USE_MUSA if (r2 == 1 && r3 == 1 && ggml_is_contiguous_2(src0) && ggml_is_contiguous_2(src1)) { // there is no broadcast and src0, src1 are contiguous across dims 2, 3 // use cublasGemmStridedBatchedEx @@ -1832,10 +1866,9 @@ static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, co cu_compute_type, CUBLAS_GEMM_DEFAULT_TENSOR_OP)); } -#endif // GGML_USE_MUSA #endif - if (dst->op_params[0] == GGML_PREC_DEFAULT) { + if (dst->op_params[0] == GGML_PREC_DEFAULT && cu_data_type == CUDA_R_16F) { const to_fp32_cuda_t to_fp32_cuda = ggml_get_to_fp32_cuda(GGML_TYPE_F16); to_fp32_cuda(dst_f16.get(), dst_ddf, ne_dst, main_stream); } @@ -1867,14 +1900,14 @@ static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor const int cc = ggml_cuda_info().devices[id].cc; use_mul_mat_q = use_mul_mat_q && ggml_cuda_should_use_mmq(src0->type, cc, src1->ne[1]); - any_gpus_with_slow_fp16 = any_gpus_with_slow_fp16 || !fast_fp16_available(cc); - any_gpus_without_fp16_mma = any_gpus_without_fp16_mma || !fp16_mma_available(cc); + any_gpus_with_slow_fp16 = any_gpus_with_slow_fp16 || !fast_fp16_hardware_available(cc); + any_gpus_without_fp16_mma = any_gpus_without_fp16_mma || !fp16_mma_hardware_available(cc); } } else { const int cc = ggml_cuda_info().devices[ctx.device].cc; use_mul_mat_q = use_mul_mat_q && ggml_cuda_should_use_mmq(src0->type, cc, src1->ne[1]); - any_gpus_with_slow_fp16 = any_gpus_with_slow_fp16 || !fast_fp16_available(cc); - any_gpus_without_fp16_mma = any_gpus_without_fp16_mma || !fp16_mma_available(cc); + any_gpus_with_slow_fp16 = any_gpus_with_slow_fp16 || !fast_fp16_hardware_available(cc); + any_gpus_without_fp16_mma = any_gpus_without_fp16_mma || !fp16_mma_hardware_available(cc); } // debug helpers @@ -2148,6 +2181,12 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg break; case GGML_OP_UNARY: switch (ggml_get_unary_op(dst)) { + case GGML_UNARY_OP_ABS: + ggml_cuda_op_abs(ctx, dst); + break; + case GGML_UNARY_OP_SGN: + ggml_cuda_op_sgn(ctx, dst); + break; case GGML_UNARY_OP_NEG: ggml_cuda_op_neg(ctx, dst); break; @@ -2191,6 +2230,9 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg case GGML_OP_GROUP_NORM: ggml_cuda_op_group_norm(ctx, dst); break; + case GGML_OP_L2_NORM: + ggml_cuda_op_l2_norm(ctx, dst); + break; case GGML_OP_CONCAT: ggml_cuda_op_concat(ctx, dst); break; @@ -2245,6 +2287,9 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg case GGML_OP_CLAMP: ggml_cuda_op_clamp(ctx, dst); break; + case GGML_OP_LOG: + ggml_cuda_op_log(ctx, dst); + break; case GGML_OP_NONE: case GGML_OP_RESHAPE: case GGML_OP_VIEW: @@ -2281,6 +2326,12 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg case GGML_OP_SUM_ROWS: ggml_cuda_op_sum_rows(ctx, dst); break; + case GGML_OP_SSM_CONV: + ggml_cuda_op_ssm_conv(ctx, dst); + break; + case GGML_OP_SSM_SCAN: + ggml_cuda_op_ssm_scan(ctx, dst); + break; case GGML_OP_ARGSORT: ggml_cuda_op_argsort(ctx, dst); break; @@ -2296,6 +2347,9 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg case GGML_OP_GATED_LINEAR_ATTN: ggml_cuda_op_gated_linear_attn(ctx, dst); break; + case GGML_OP_RWKV_WKV7: + ggml_cuda_op_rwkv_wkv7(ctx, dst); + break; case GGML_OP_CROSS_ENTROPY_LOSS_BACK: ggml_cuda_cross_entropy_loss_back(ctx, dst); break; @@ -2415,10 +2469,11 @@ static void ggml_backend_cuda_synchronize(ggml_backend_t backend) { #ifdef USE_CUDA_GRAPH static bool check_node_graph_compatibility_and_refresh_copy_ops(ggml_backend_cuda_context * cuda_ctx, ggml_cgraph * cgraph, - std::vector & ggml_cuda_cpy_fn_ptrs, bool use_cuda_graph) { + bool use_cuda_graph) { // Loop over nodes in GGML graph to obtain info needed for CUDA graph - cuda_ctx->cuda_graph->updated_kernel_arg.clear(); + cuda_ctx->cuda_graph->cpy_dest_ptrs.clear(); + for (int i = 0; i < cgraph->n_nodes; i++) { ggml_tensor * node = cgraph->nodes[i]; @@ -2450,8 +2505,11 @@ static bool check_node_graph_compatibility_and_refresh_copy_ops(ggml_backend_cud } if (node->op == GGML_OP_CPY) { - // store the copy op parameter which changes with each token. - cuda_ctx->cuda_graph->updated_kernel_arg.push_back((char **) &(node->src[1]->data)); + + // Store the pointers which are updated for each token, such that these can be sent + // to the device and accessed using indirection from CUDA graph + cuda_ctx->cuda_graph->cpy_dest_ptrs.push_back((char *) node->src[1]->data); + // store a pointer to each copy op CUDA kernel to identify it later void * ptr = ggml_cuda_cpy_fn(node->src[0], node->src[1]); if (!ptr) { @@ -2459,10 +2517,6 @@ static bool check_node_graph_compatibility_and_refresh_copy_ops(ggml_backend_cud #ifndef NDEBUG GGML_LOG_DEBUG("%s: disabling CUDA graphs due to unsupported copy op\n", __func__); #endif - } else { - if (std::find(ggml_cuda_cpy_fn_ptrs.begin(), ggml_cuda_cpy_fn_ptrs.end(), ptr) == ggml_cuda_cpy_fn_ptrs.end()) { - ggml_cuda_cpy_fn_ptrs.push_back(ptr); - } } } @@ -2471,6 +2525,12 @@ static bool check_node_graph_compatibility_and_refresh_copy_ops(ggml_backend_cud } } + if (use_cuda_graph) { + cuda_ctx->cuda_graph->use_cpy_indirection = true; + // copy pointers to GPU so they can be accessed via indirection within CUDA graph + ggml_cuda_cpy_dest_ptrs_copy(cuda_ctx->cuda_graph.get(), cuda_ctx->cuda_graph->cpy_dest_ptrs.data(), cuda_ctx->cuda_graph->cpy_dest_ptrs.size(), cuda_ctx->stream()); + } + return use_cuda_graph; } @@ -2525,51 +2585,6 @@ static bool ggml_graph_node_has_matching_properties(ggml_tensor * node, ggml_gra return true; } -static void maintain_cuda_graph(ggml_backend_cuda_context * cuda_ctx, std::vector & ggml_cuda_cpy_fn_ptrs, bool cuda_graph_update_required) { - - if (cuda_graph_update_required) { - // Extract nodes from graph - // First call with null argument gets number of nodes in graph - CUDA_CHECK(cudaGraphGetNodes(cuda_ctx->cuda_graph->graph, nullptr, &cuda_ctx->cuda_graph->num_nodes)); - // Subsequent call with non-null argument gets nodes - cuda_ctx->cuda_graph->nodes.clear(); - cuda_ctx->cuda_graph->nodes.resize(cuda_ctx->cuda_graph->num_nodes); - cuda_ctx->cuda_graph->params.clear(); - cuda_ctx->cuda_graph->params.resize(cuda_ctx->cuda_graph->num_nodes); - if (cuda_ctx->cuda_graph->num_nodes > 0) { - CUDA_CHECK(cudaGraphGetNodes(cuda_ctx->cuda_graph->graph, cuda_ctx->cuda_graph->nodes.data(), &cuda_ctx->cuda_graph->num_nodes)); - - // Loop over nodes, and extract kernel parameters from each node - for (size_t i = 0; i < cuda_ctx->cuda_graph->num_nodes; i++) { - cudaGraphNodeType node_type; - CUDA_CHECK(cudaGraphNodeGetType(cuda_ctx->cuda_graph->nodes[i], &node_type)); - if (node_type == cudaGraphNodeTypeKernel) { - cudaError_t stat = cudaGraphKernelNodeGetParams(cuda_ctx->cuda_graph->nodes[i], &cuda_ctx->cuda_graph->params[i]); // Get params using runtime - if (stat == cudaErrorInvalidDeviceFunction) { - // Fails due to incorrect handling by CUDA runtime of CUDA BLAS node. - // We don't need to update blas nodes, so clear error and move on. - (void)cudaGetLastError(); - } else { - GGML_ASSERT(stat == cudaSuccess); - } - } - } - } - } else { - // One of the arguments to the copy kernel is updated for each token, hence we need to - // replace that argument with the updated value in the CUDA graph - // on update steps, the live parameters will already be captured - int k = 0; - for (size_t i = 0; i < cuda_ctx->cuda_graph->num_nodes; i++) { - if(count(ggml_cuda_cpy_fn_ptrs.begin(), ggml_cuda_cpy_fn_ptrs.end(), cuda_ctx->cuda_graph->params[i].func) > 0) { - char ** updated_kernel_arg_ptr = cuda_ctx->cuda_graph->updated_kernel_arg.at(k++); - cuda_ctx->cuda_graph->params[i].kernelParams[1] = updated_kernel_arg_ptr; - CUDA_CHECK(cudaGraphKernelNodeSetParams(cuda_ctx->cuda_graph->nodes[i], &cuda_ctx->cuda_graph->params[i])); - } - } - } -} - static bool is_cuda_graph_update_required(ggml_backend_cuda_context * cuda_ctx, ggml_cgraph * cgraph) { bool cuda_graph_update_required = false; @@ -2602,13 +2617,15 @@ static bool is_cuda_graph_update_required(ggml_backend_cuda_context * cuda_ctx, static void update_cuda_graph_executable(ggml_backend_cuda_context * cuda_ctx) { +#if CUDART_VERSION >= 12000 cudaGraphExecUpdateResultInfo result_info; -#ifdef __HIP_PLATFORM_AMD__ - hipGraphNode_t errorNode; - hipError_t stat = hipGraphExecUpdate(cuda_ctx->cuda_graph->instance, cuda_ctx->cuda_graph->graph, &errorNode, &result_info); -#else cudaError_t stat = cudaGraphExecUpdate(cuda_ctx->cuda_graph->instance, cuda_ctx->cuda_graph->graph, &result_info); -#endif +#else + cudaGraphNode_t errorNode; + cudaGraphExecUpdateResult result_info; + cudaError_t stat = cudaGraphExecUpdate(cuda_ctx->cuda_graph->instance, cuda_ctx->cuda_graph->graph, &errorNode, &result_info); +#endif // CUDART_VERSION >= 12000 + if (stat == cudaErrorGraphExecUpdateFailure) { #ifndef NDEBUG GGML_LOG_DEBUG("%s: CUDA graph update failed\n", __func__); @@ -2627,8 +2644,7 @@ static void update_cuda_graph_executable(ggml_backend_cuda_context * cuda_ctx) { #endif static void evaluate_and_capture_cuda_graph(ggml_backend_cuda_context * cuda_ctx, ggml_cgraph * cgraph, - [[maybe_unused]] std::vector & ggml_cuda_cpy_fn_ptrs, bool & graph_evaluated_or_captured, bool & use_cuda_graph, - bool & cuda_graph_update_required) { + bool & graph_evaluated_or_captured, bool & use_cuda_graph, bool & cuda_graph_update_required) { while (!graph_evaluated_or_captured) { // Only perform the graph execution if CUDA graphs are not enabled, or we are capturing the graph. @@ -2678,13 +2694,9 @@ static void evaluate_and_capture_cuda_graph(ggml_backend_cuda_context * cuda_ctx if (cuda_ctx->cuda_graph->instance == nullptr) { // Create executable graph from captured graph. CUDA_CHECK(cudaGraphInstantiate(&cuda_ctx->cuda_graph->instance, cuda_ctx->cuda_graph->graph, NULL, NULL, 0)); } - - // Perform update to graph (if required for this token), and change copy parameter (required for every token) - maintain_cuda_graph(cuda_ctx, ggml_cuda_cpy_fn_ptrs, cuda_graph_update_required); - - // Update graph executable - update_cuda_graph_executable(cuda_ctx); - + if (cuda_graph_update_required) { // Update graph executable + update_cuda_graph_executable(cuda_ctx); + } // Launch graph CUDA_CHECK(cudaGraphLaunch(cuda_ctx->cuda_graph->instance, cuda_ctx->stream())); #else @@ -2698,10 +2710,6 @@ static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t backend, ggml_cuda_set_device(cuda_ctx->device); - // vector of pointers to CUDA cpy kernels, which are required to identify - // kernel parameters which need updated in the graph for each token - std::vector ggml_cuda_cpy_fn_ptrs; - #ifdef USE_CUDA_GRAPH static const bool disable_cuda_graphs_due_to_env = (getenv("GGML_CUDA_DISABLE_GRAPHS") != nullptr); @@ -2735,8 +2743,7 @@ static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t backend, if (use_cuda_graph) { cuda_graph_update_required = is_cuda_graph_update_required(cuda_ctx, cgraph); - use_cuda_graph = check_node_graph_compatibility_and_refresh_copy_ops(cuda_ctx, cgraph, - ggml_cuda_cpy_fn_ptrs, use_cuda_graph); + use_cuda_graph = check_node_graph_compatibility_and_refresh_copy_ops(cuda_ctx, cgraph, use_cuda_graph); // Disable CUDA graphs (from the next token) if the use-case is demanding too many consecutive graph updates. if (use_cuda_graph && cuda_graph_update_required) { @@ -2757,6 +2764,10 @@ static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t backend, CUDA_CHECK(cudaStreamBeginCapture(cuda_ctx->stream(), cudaStreamCaptureModeRelaxed)); } + if (!use_cuda_graph) { + cuda_ctx->cuda_graph->use_cpy_indirection = false; + } + #else bool use_cuda_graph = false; bool cuda_graph_update_required = false; @@ -2764,7 +2775,7 @@ static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t backend, bool graph_evaluated_or_captured = false; - evaluate_and_capture_cuda_graph(cuda_ctx, cgraph, ggml_cuda_cpy_fn_ptrs, graph_evaluated_or_captured, use_cuda_graph, cuda_graph_update_required); + evaluate_and_capture_cuda_graph(cuda_ctx, cgraph, graph_evaluated_or_captured, use_cuda_graph, cuda_graph_update_required); return GGML_STATUS_SUCCESS; } @@ -2840,7 +2851,7 @@ bool ggml_backend_cuda_register_host_buffer(void * buffer, size_t size) { return false; } -#if CUDART_VERSION >= 11100 || defined(GGML_USE_MUSA) +#if CUDART_VERSION >= 11010 || defined(GGML_USE_MUSA) cudaError_t err = cudaHostRegister(buffer, size, cudaHostRegisterPortable | cudaHostRegisterReadOnly); if (err != cudaSuccess) { // clear the error @@ -2852,8 +2863,10 @@ bool ggml_backend_cuda_register_host_buffer(void * buffer, size_t size) { } return true; #else + GGML_UNUSED(buffer); + GGML_UNUSED(size); return false; -#endif +#endif // CUDART_VERSION >= 11010 || defined(GGML_USE_MUSA) } void ggml_backend_cuda_unregister_host_buffer(void * buffer) { @@ -2961,6 +2974,8 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g switch (op->op) { case GGML_OP_UNARY: switch (ggml_get_unary_op(op)) { + case GGML_UNARY_OP_ABS: + case GGML_UNARY_OP_SGN: case GGML_UNARY_OP_NEG: case GGML_UNARY_OP_STEP: case GGML_UNARY_OP_GELU: @@ -3064,6 +3079,9 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_F32) { return true; } + if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_BF16) { + return true; + } if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_F16) { return true; } @@ -3076,15 +3094,27 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_Q4_0) { return true; } + if (src0_type == GGML_TYPE_Q4_0 && src1_type == GGML_TYPE_F32) { + return true; + } if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_Q4_1) { return true; } + if (src0_type == GGML_TYPE_Q4_1 && src1_type == GGML_TYPE_F32) { + return true; + } if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_Q5_0) { return true; } + if (src0_type == GGML_TYPE_Q5_0 && src1_type == GGML_TYPE_F32) { + return true; + } if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_Q5_1) { return true; } + if (src0_type == GGML_TYPE_Q5_1 && src1_type == GGML_TYPE_F32) { + return true; + } if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_IQ4_NL) { return true; } @@ -3131,10 +3161,11 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g return false; } break; case GGML_OP_SILU_BACK: - return ggml_is_contiguous(op->src[0]); + return ggml_is_contiguous(op->src[0]) && op->src[0]->type == GGML_TYPE_F32; break; case GGML_OP_NORM: case GGML_OP_RMS_NORM: + case GGML_OP_L2_NORM: return true; case GGML_OP_RMS_NORM_BACK: return ggml_is_contiguous(op->src[0]) && op->ne[0] % WARP_SIZE == 0; @@ -3155,6 +3186,9 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g case GGML_OP_SIN: case GGML_OP_COS: case GGML_OP_CLAMP: + case GGML_OP_LOG: + case GGML_OP_SSM_SCAN: + case GGML_OP_SSM_CONV: return true; case GGML_OP_CONT: return op->src[0]->type != GGML_TYPE_BF16; @@ -3182,17 +3216,29 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g case GGML_OP_GROUP_NORM: return ggml_is_contiguous(op->src[0]); case GGML_OP_UPSCALE: + return op->src[0]->type == GGML_TYPE_F32 && op->op_params[0] == GGML_SCALE_MODE_NEAREST; case GGML_OP_PAD: case GGML_OP_ARANGE: case GGML_OP_TIMESTEP_EMBEDDING: case GGML_OP_LEAKY_RELU: case GGML_OP_RWKV_WKV6: case GGML_OP_GATED_LINEAR_ATTN: + case GGML_OP_RWKV_WKV7: return true; case GGML_OP_FLASH_ATTN_EXT: { #ifndef FLASH_ATTN_AVAILABLE return false; -#endif +#endif // FLASH_ATTN_AVAILABLE + if (op->src[1]->ne[0] != op->src[2]->ne[0]) { + // different head sizes of K and V are not supported yet + return false; + } + if (op->src[0]->ne[0] == 192) { + return false; + } + if (op->src[0]->ne[3] != 1) { + return false; + } if (op->src[1]->type == GGML_TYPE_BF16 || op->src[2]->type == GGML_TYPE_BF16) { return false; } @@ -3205,8 +3251,8 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g if (op->src[0]->ne[0] == 256 && op->src[1]->type == GGML_TYPE_F16 && op->src[2]->type == GGML_TYPE_F16) { return true; } - const int cc = ggml_cuda_info().devices[dev_ctx->device].cc; - return cc >= GGML_CUDA_CC_VOLTA && cc < GGML_CUDA_CC_OFFSET_AMD && op->src[1]->type == GGML_TYPE_F16 && op->src[2]->type == GGML_TYPE_F16; + return fp16_mma_available(ggml_cuda_info().devices[dev_ctx->device].cc) && + op->src[1]->type == GGML_TYPE_F16 && op->src[2]->type == GGML_TYPE_F16; } case GGML_OP_CROSS_ENTROPY_LOSS: case GGML_OP_CROSS_ENTROPY_LOSS_BACK: diff --git a/ggml/src/ggml-cuda/mma.cuh b/ggml/src/ggml-cuda/mma.cuh index bbc0a35ae..2af63355a 100644 --- a/ggml/src/ggml-cuda/mma.cuh +++ b/ggml/src/ggml-cuda/mma.cuh @@ -4,11 +4,12 @@ // https://docs.nvidia.com/cuda/parallel-thread-execution/index.html#matrix-multiply-accumulate-operation-using-mma-instruction // // Like with nvcuda::wmma there are three types of matrix tiles: A, B, and C with A @ B = C. -// A is a row-major matrix with shape I x K. -// B is a column-major matrix with shape K x J. -// C is a column-major matrix with shape I x J. -// Note that along their lowest dimension I, J, and K are measured in physical 32 bit elements instead of logical elements. -// The functions get_i, get_j, and get_k can be used to get the physical 32 bit index of the lth element of a thread within a tile. +// A is a row-major matrix with shape M x K. +// B is a column-major matrix with shape K x N. +// C is a column-major matrix with shape M x N. +// A, B, and C are represented using the same fundamental data type: a row-major matrix with I rows and J columns. +// Note that J is measured in physical 32 bit elements instead of logical elements. +// The methods get_i and get_j can be used to get the physical 32 bit index of the lth element of a thread within a tile. // All matrix tiles have ne physical 32 bit elements per warp. // // As described in the documentation, all pointers for load_ldmatrix must be to shared memory and aligned to 16 bytes. @@ -23,8 +24,9 @@ static __device__ __forceinline__ int ggml_cuda_movmatrix(const int x) { #ifdef NEW_MMA_AVAILABLE asm("movmatrix.sync.aligned.m8n8.trans.b16 %0, %1;" - : "+r"(ret) : "r"(x)); + : "=r"(ret) : "r"(x)); #else + GGML_UNUSED(x); NO_DEVICE_CODE; #endif // defined(NEW_MMA_AVAILABLE) return ret; @@ -52,407 +54,343 @@ static __device__ __forceinline__ int ggml_cuda_movmatrix(const int x) { #endif // CUDART_VERSION >= 11080 +static __device__ __forceinline__ half2 ggml_cuda_movmatrix(const half2 x) { + half2 ret; + *((int *) &ret) = ggml_cuda_movmatrix(*((const int *) &x)); + return ret; +} -template -struct mma_A_I16K4 { - static_assert(sizeof(T) == 4, "bad type size"); +namespace ggml_cuda_mma { - static constexpr int I = 16; - static constexpr int K = 4; - static constexpr int ne = 2; + template + struct tile { + static constexpr int I = I_; + static constexpr int J = J_; + static constexpr int ne = I * J / WARP_SIZE; + T x[ne] = {0}; - T x[ne]; + static __device__ __forceinline__ int get_i(const int l) { + if constexpr (I == 8 && (J == 4 || J == 8)) { + return threadIdx.x / 4; + } else if constexpr (I == 16 && J == 8) { + return (l / 2) * 8 + threadIdx.x / 4; + } else if constexpr (I == 16 && J == 16) { + return ((l / 2) % 2) * 8 + threadIdx.x / 4; + } else { + static_assert(I == -1 && J == -1, "template specialization not implemented"); + } + } - static __device__ __forceinline__ int get_i(const int l) { - const int ret = (l%2) * (I/2) + threadIdx.x / K; - GGML_CUDA_ASSUME(ret >= 0); - GGML_CUDA_ASSUME(ret < I); - return ret; - } + static __device__ __forceinline__ int get_j(const int l) { + if constexpr (I == 8 && J == 4) { + return threadIdx.x % 4; + } else if constexpr (I == 8 && J == 8) { + return 4 * l + threadIdx.x % 4; + } else if constexpr (I == 16 && J == 8) { + return 2 * (threadIdx.x % 4) + l % 2; + } else if constexpr (I == 16 && J == 16) { + return 8 * (l / 4) + 2 * (threadIdx.x % 4) + l % 2; + } else { + static_assert(I == -1 && J == -1, "template specialization not implemented"); + } + } + }; - static __device__ __forceinline__ int get_k(const int /* l */) { - const int ret = threadIdx.x % K; - GGML_CUDA_ASSUME(ret >= 0); - GGML_CUDA_ASSUME(ret < K); - return ret; - } + template + struct tile { + static constexpr int I = I_; + static constexpr int J = J_; + static constexpr int ne = I * J / WARP_SIZE; + half2 x[ne] = {{0.0f, 0.0f}}; - __device__ __forceinline__ void load_generic(const T * __restrict__ xs0, const int & stride) { + static __device__ __forceinline__ int get_i(const int l) { + if constexpr (I == 8 && J == 8) { + return threadIdx.x / 4; + } else if constexpr (I == 16 && J == 4) { + return l * 8 + threadIdx.x / 4; + } else if constexpr (I == 16 && J == 8) { + return (l % 2) * 8 + threadIdx.x / 4; + } else { + static_assert(I == -1 && J == -1, "template specialization not implemented"); + } + } + + static __device__ __forceinline__ int get_j(const int l) { + if constexpr (I == 8 && J == 8) { + return l * 4 + threadIdx.x % 4; + } else if constexpr (I == 16 && J == 4) { + return threadIdx.x % 4; + } else if constexpr (I == 16 && J == 8) { + return (l / 2) * 4 + threadIdx.x % 4; + } else { + static_assert(I == -1 && J == -1, "template specialization not implemented"); + } + } + }; + + template + static __device__ __forceinline__ tile get_half2(const tile & tile_float) { + tile ret; #pragma unroll - for (int l = 0; l < ne; ++l) { - x[l] = xs0[get_i(l)*stride + get_k(l)]; + for (int l0 = 0; l0 < tile_float.ne; l0 += 2) { + ret.x[l0/2] = make_half2(tile_float.x[l0 + 0], tile_float.x[l0 + 1]); + } + return ret; + } + + static __device__ __forceinline__ tile<8, 8, half2> get_transposed(const tile<16, 4, half2> & t) { + tile<8, 8, half2> ret; + ret.x[0] = ggml_cuda_movmatrix(t.x[0]); + ret.x[1] = ggml_cuda_movmatrix(t.x[1]); + + return ret; + } + + template + static __device__ __forceinline__ void load_generic(tile & t, const T * __restrict__ xs0, const int stride) { +#pragma unroll + for (int l = 0; l < t.ne; ++l) { + t.x[l] = xs0[t.get_i(l)*stride + t.get_j(l)]; } } - __device__ __forceinline__ void load_ldmatrix(const T * __restrict__ xs0, const int & stride) { + template + static __device__ __forceinline__ void load_ldmatrix( + tile<8, 8, T> & t, const T * __restrict__ xs0, const int stride) { #ifdef NEW_MMA_AVAILABLE - int * xi = (int *) x; - const int * xs = (const int *) xs0 + (threadIdx.x%I)*stride; - asm("ldmatrix.sync.aligned.m8n8.x2.b16 {%0, %1}, [%2];" - : "+r"(xi[0]), "+r"(xi[1]) + int * xi = (int *) t.x; + const int * xs = (const int *) xs0 + (threadIdx.x % t.I) * stride + ((threadIdx.x / t.I) * (t.J / 2)) % t.J; + asm volatile("ldmatrix.sync.aligned.m8n8.x2.b16 {%0, %1}, [%2];" + : "=r"(xi[0]), "=r"(xi[1]) + : "l"(xs)); +#else + load_generic(t, xs0, stride); +#endif // NEW_MMA_AVAILABLE + } + + template + static __device__ __forceinline__ void load_ldmatrix( + tile<16, 4, T> & t, const T * __restrict__ xs0, const int stride) { +#ifdef NEW_MMA_AVAILABLE + int * xi = (int *) t.x; + const int * xs = (const int *) xs0 + (threadIdx.x % t.I) * stride; + asm volatile("ldmatrix.sync.aligned.m8n8.x2.b16 {%0, %1}, [%2];" + : "=r"(xi[0]), "=r"(xi[1]) : "l"(xs)); #else load_generic(xs0, stride); + GGML_UNUSED(t); #endif // NEW_MMA_AVAILABLE } -}; -template -struct mma_A_I16K8 { - static_assert(sizeof(T) == 4, "bad type size"); - - static constexpr int I = 16; - static constexpr int K = 8; - static constexpr int ne = 4; - - T x[ne]; - - static __device__ __forceinline__ int get_i(const int l) { - const int ret = (l%2) * (I/2) + threadIdx.x / (K/2); - GGML_CUDA_ASSUME(ret >= 0); - GGML_CUDA_ASSUME(ret < I); - return ret; - } - - static __device__ __forceinline__ int get_k(const int l) { - const int ret = (l/2) * (K/2) + threadIdx.x % (K/2); - GGML_CUDA_ASSUME(ret >= 0); - GGML_CUDA_ASSUME(ret < K); - return ret; - } - - __device__ __forceinline__ void load_generic(const T * __restrict__ xs0, const int & stride) { -#pragma unroll - for (int l = 0; l < ne; ++l) { - x[l] = xs0[get_i(l)*stride + get_k(l)]; - } - } - - __device__ __forceinline__ void load_ldmatrix(const T * __restrict__ xs0, const int & stride) { + template + static __device__ __forceinline__ void load_ldmatrix( + tile<16, 8, T> & t, const T * __restrict__ xs0, const int stride) { #ifdef NEW_MMA_AVAILABLE - int * xi = (int * ) x; - const int * xs = (const int *) xs0 + (threadIdx.x%I)*stride + (threadIdx.x/I)*(K/2); - asm("ldmatrix.sync.aligned.m8n8.x4.b16 {%0, %1, %2, %3}, [%4];" - : "+r"(xi[0]), "+r"(xi[1]), "+r"(xi[2]), "+r"(xi[3]) + int * xi = (int * ) t.x; + const int * xs = (const int *) xs0 + (threadIdx.x % t.I) * stride + (threadIdx.x / t.I) * (t.J / 2); + asm volatile("ldmatrix.sync.aligned.m8n8.x4.b16 {%0, %1, %2, %3}, [%4];" + : "=r"(xi[0]), "=r"(xi[1]), "=r"(xi[2]), "=r"(xi[3]) : "l"(xs)); #else + load_generic(t, xs0, stride); +#endif // NEW_MMA_AVAILABLE + } + + template + static __device__ __forceinline__ void load_ldmatrix_trans( + tile<16, 8, T> & t, const T * __restrict__ xs0, const int stride) { +#ifdef NEW_MMA_AVAILABLE + int * xi = (int * ) t.x; + const int * xs = (const int *) xs0 + (threadIdx.x % t.I) * stride + (threadIdx.x / t.I) * (t.J / 2); + asm volatile("ldmatrix.sync.aligned.m8n8.x4.trans.b16 {%0, %1, %2, %3}, [%4];" + : "=r"(xi[0]), "=r"(xi[2]), "=r"(xi[1]), "=r"(xi[3]) + : "l"(xs)); +#else + GGML_UNUSED(t); GGML_UNUSED(xs0); GGML_UNUSED(stride); NO_DEVICE_CODE; #endif // NEW_MMA_AVAILABLE } - __device__ __forceinline__ void load_ldmatrix_trans(const T * __restrict__ xs0, const int & stride) { -#ifdef NEW_MMA_AVAILABLE - int * xi = (int * ) x; - const int * xs = (const int *) xs0 + (threadIdx.x%I)*stride + (threadIdx.x/I)*(K/2); - asm("ldmatrix.sync.aligned.m8n8.x4.trans.b16 {%0, %1, %2, %3}, [%4];" - : "+r"(xi[0]), "+r"(xi[2]), "+r"(xi[1]), "+r"(xi[3]) - : "l"(xs)); -#else - GGML_UNUSED(xs0); - GGML_UNUSED(stride); - NO_DEVICE_CODE; -#endif // NEW_MMA_AVAILABLE - } - - __device__ __forceinline__ void transpose() { - int * xi = (int *) x; - xi[0] = ggml_cuda_movmatrix(xi[0]); - - const int tmp = ggml_cuda_movmatrix(xi[1]); - xi[1] = ggml_cuda_movmatrix(xi[2]); - xi[2] = tmp; - - xi[3] = ggml_cuda_movmatrix(xi[3]); - } -}; - -template -struct mma_B_J8K4 { - static_assert(sizeof(T) == 4, "bad type size"); - - static constexpr int J = 8; - static constexpr int K = 4; - static constexpr int ne = 1; - - T x[ne]; - - static __device__ __forceinline__ int get_j(const int /* l */) { - const int ret = threadIdx.x / K; - GGML_CUDA_ASSUME(ret >= 0); - GGML_CUDA_ASSUME(ret < J); - return ret; - } - - static __device__ __forceinline__ int get_k(const int /* l */) { - const int ret = threadIdx.x % K; - GGML_CUDA_ASSUME(ret >= 0); - GGML_CUDA_ASSUME(ret < K); - return ret; - } - - __device__ __forceinline__ void load_generic(const T * __restrict__ xs0, const int & stride) { -#pragma unroll - for (int l = 0; l < ne; ++l) { - x[l] = xs0[get_j(l)*stride + get_k(l)]; - } - } - - __device__ __forceinline__ void load_ldmatrix(const T * __restrict__ xs0, const int & stride) { -#ifdef NEW_MMA_AVAILABLE - int * xi = (int *) x; - const int * xs = (const int *) xs0 + (threadIdx.x%J)*stride; - asm("ldmatrix.sync.aligned.m8n8.x1.b16 {%0}, [%1];" - : "+r"(xi[0]) : "l"(xs)); -#else - load_generic(xs0, stride); -#endif // NEW_MMA_AVAILABLE - } -}; - -template -struct mma_B_J8K8 { - static_assert(sizeof(T) == 4, "bad type size"); - - static constexpr int J = 8; - static constexpr int K = 8; - static constexpr int ne = 2; - - T x[ne]; - - static __device__ __forceinline__ int get_j(const int /* l */) { - const int ret = threadIdx.x / (K/2); - GGML_CUDA_ASSUME(ret >= 0); - GGML_CUDA_ASSUME(ret < J); - return ret; - } - - static __device__ __forceinline__ int get_k(const int l) { - const int ret = l * (K/2) + threadIdx.x % (K/2); - GGML_CUDA_ASSUME(ret >= 0); - GGML_CUDA_ASSUME(ret < K); - return ret; - } - - __device__ __forceinline__ void load_generic(const T * __restrict__ xs0, const int & stride) { -#pragma unroll - for (int l = 0; l < ne; ++l) { - x[l] = xs0[get_j(l)*stride + get_k(l)]; - } - } - - __device__ __forceinline__ void load_ldmatrix(const T * __restrict__ xs0, const int & stride) { -#ifdef NEW_MMA_AVAILABLE - int * xi = (int *) x; - const int * xs = (const int *) xs0 + (threadIdx.x%J)*stride + ((threadIdx.x/J)*(K/2)) % K; - asm("ldmatrix.sync.aligned.m8n8.x2.b16 {%0, %1}, [%2];" - : "+r"(xi[0]), "+r"(xi[1]) - : "l"(xs)); -#else - load_generic(xs0, stride); -#endif // NEW_MMA_AVAILABLE - } -}; - -template -struct mma_C_I16J8 {}; - -template <> -struct mma_C_I16J8 { - static constexpr int I = 16; - static constexpr int J = 8; - static constexpr int ne = 4; - - int x[ne] = {0}; - - static __device__ __forceinline__ int get_i(const int l) { - const int ret = (l/2) * (I/2) + threadIdx.x / (J/2); - GGML_CUDA_ASSUME(ret >= 0); - GGML_CUDA_ASSUME(ret < I); - return ret; - } - - static __device__ __forceinline__ int get_j(const int l) { - const int ret = 2 * (threadIdx.x % (J/2)) + l%2; - GGML_CUDA_ASSUME(ret >= 0); - GGML_CUDA_ASSUME(ret < J); - return ret; - } - - __device__ __forceinline__ void mma(const mma_A_I16K4 & mma_A, const mma_B_J8K4 & mma_B) { + static __device__ __forceinline__ void mma( + tile<16, 8, int> & D, const tile<16, 4, int> & A, const tile<8, 4, int> & B) { #ifdef NEW_MMA_AVAILABLE #if __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE asm("mma.sync.aligned.m16n8k16.row.col.s32.s8.s8.s32 {%0, %1, %2, %3}, {%4, %5}, {%6}, {%0, %1, %2, %3};" - : "+r"(x[0]), "+r"(x[1]), "+r"(x[2]), "+r"(x[3]) - : "r"(mma_A.x[0]), "r"(mma_A.x[1]), "r"(mma_B.x[0])); + : "+r"(D.x[0]), "+r"(D.x[1]), "+r"(D.x[2]), "+r"(D.x[3]) + : "r"(A.x[0]), "r"(A.x[1]), "r"(B.x[0])); #else // On Turing m16n8k16 mma is not available, use 2x m8n8k16 mma instead: asm("mma.sync.aligned.m8n8k16.row.col.s32.s8.s8.s32 {%0, %1}, {%2}, {%3}, {%0, %1};" - : "+r"(x[0]), "+r"(x[1]) - : "r"(mma_A.x[0]), "r"(mma_B.x[0])); + : "+r"(D.x[0]), "+r"(D.x[1]) + : "r"(A.x[0]), "r"(B.x[0])); asm("mma.sync.aligned.m8n8k16.row.col.s32.s8.s8.s32 {%0, %1}, {%2}, {%3}, {%0, %1};" - : "+r"(x[2]), "+r"(x[3]) - : "r"(mma_A.x[1]), "r"(mma_B.x[0])); + : "+r"(D.x[2]), "+r"(D.x[3]) + : "r"(A.x[1]), "r"(B.x[0])); #endif // __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE #else - GGML_UNUSED(mma_A); - GGML_UNUSED(mma_B); + GGML_UNUSED(D); + GGML_UNUSED(A); + GGML_UNUSED(B); NO_DEVICE_CODE; #endif // NEW_MMA_AVAILABLE } - __device__ __forceinline__ void mma(const mma_A_I16K8 & mma_A, const mma_B_J8K8 & mma_B) { + static __device__ __forceinline__ void mma( + tile<16, 8, int> & D, const tile<16, 8, int> & A, const tile<8, 8, int> & B) { #ifdef NEW_MMA_AVAILABLE #if __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE asm("mma.sync.aligned.m16n8k32.row.col.s32.s8.s8.s32 {%0, %1, %2, %3}, {%4, %5, %6, %7}, {%8, %9}, {%0, %1, %2, %3};" - : "+r"(x[0]), "+r"(x[1]), "+r"(x[2]), "+r"(x[3]) - : "r"(mma_A.x[0]), "r"(mma_A.x[1]), "r"(mma_A.x[2]), "r"(mma_A.x[3]), "r"(mma_B.x[0]), "r"(mma_B.x[1])); + : "+r"(D.x[0]), "+r"(D.x[1]), "+r"(D.x[2]), "+r"(D.x[3]) + : "r"(A.x[0]), "r"(A.x[1]), "r"(A.x[2]), "r"(A.x[3]), "r"(B.x[0]), "r"(B.x[1])); #else // On Turing m16n8k32 mma is not available, use 4x m8n8k16 mma instead: asm("mma.sync.aligned.m8n8k16.row.col.s32.s8.s8.s32 {%0, %1}, {%2}, {%3}, {%0, %1};" - : "+r"(x[0]), "+r"(x[1]) - : "r"(mma_A.x[0]), "r"(mma_B.x[0])); + : "+r"(D.x[0]), "+r"(D.x[1]) + : "r"(A.x[0]), "r"(B.x[0])); asm("mma.sync.aligned.m8n8k16.row.col.s32.s8.s8.s32 {%0, %1}, {%2}, {%3}, {%0, %1};" - : "+r"(x[2]), "+r"(x[3]) - : "r"(mma_A.x[1]), "r"(mma_B.x[0])); + : "+r"(D.x[2]), "+r"(D.x[3]) + : "r"(A.x[1]), "r"(B.x[0])); asm("mma.sync.aligned.m8n8k16.row.col.s32.s8.s8.s32 {%0, %1}, {%2}, {%3}, {%0, %1};" - : "+r"(x[0]), "+r"(x[1]) - : "r"(mma_A.x[2]), "r"(mma_B.x[1])); + : "+r"(D.x[0]), "+r"(D.x[1]) + : "r"(A.x[2]), "r"(B.x[1])); asm("mma.sync.aligned.m8n8k16.row.col.s32.s8.s8.s32 {%0, %1}, {%2}, {%3}, {%0, %1};" - : "+r"(x[2]), "+r"(x[3]) - : "r"(mma_A.x[3]), "r"(mma_B.x[1])); + : "+r"(D.x[2]), "+r"(D.x[3]) + : "r"(A.x[3]), "r"(B.x[1])); #endif // __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE #else - GGML_UNUSED(mma_A); - GGML_UNUSED(mma_B); + GGML_UNUSED(D); + GGML_UNUSED(A); + GGML_UNUSED(B); NO_DEVICE_CODE; #endif // NEW_MMA_AVAILABLE } -}; -template <> -struct mma_C_I16J8 { - static constexpr int I = 16; - static constexpr int J = 4; - static constexpr int ne = 2; - - half2 x[ne] = {{0.0f, 0.0f}, {0.0f, 0.0f}}; - - static __device__ __forceinline__ int get_i(const int l) { - const int ret = l * (I/2) + threadIdx.x / J; - GGML_CUDA_ASSUME(ret >= 0); - GGML_CUDA_ASSUME(ret < I); - return ret; - } - - static __device__ __forceinline__ int get_j(const int /* l */) { - const int ret = threadIdx.x % J; - GGML_CUDA_ASSUME(ret >= 0); - GGML_CUDA_ASSUME(ret < J); - return ret; - } - - __device__ __forceinline__ void mma(const mma_A_I16K8 & mma_A, const mma_B_J8K8 & mma_B) { + static __device__ __forceinline__ void mma( + tile<16, 4, half2> & D, const tile<16, 8, half2> & A, const tile<8, 8, half2> & B) { #ifdef NEW_MMA_AVAILABLE - int * Axi = (int *) mma_A.x; - int * Bxi = (int *) mma_B.x; - int * xi = (int *) x; + const int * Axi = (const int *) A.x; + const int * Bxi = (const int *) B.x; + int * Dxi = (int *) D.x; #if __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE asm("mma.sync.aligned.m16n8k16.row.col.f16.f16.f16.f16 {%0, %1}, {%2, %3, %4, %5}, {%6, %7}, {%0, %1};" - : "+r"(xi[0]), "+r"(xi[1]) + : "+r"(Dxi[0]), "+r"(Dxi[1]) : "r"(Axi[0]), "r"(Axi[1]), "r"(Axi[2]), "r"(Axi[3]), "r"(Bxi[0]), "r"(Bxi[1])); #else // On Turing m16n8k16 mma is not available, use 2x m8n8k8 mma instead: asm("mma.sync.aligned.m16n8k8.row.col.f16.f16.f16.f16 {%0, %1}, {%2, %3}, {%4}, {%0, %1};" - : "+r"(xi[0]), "+r"(xi[1]) + : "+r"(Dxi[0]), "+r"(Dxi[1]) : "r"(Axi[0]), "r"(Axi[1]), "r"(Bxi[0])); asm("mma.sync.aligned.m16n8k8.row.col.f16.f16.f16.f16 {%0, %1}, {%2, %3}, {%4}, {%0, %1};" - : "+r"(xi[0]), "+r"(xi[1]) + : "+r"(Dxi[0]), "+r"(Dxi[1]) : "r"(Axi[2]), "r"(Axi[3]), "r"(Bxi[1])); #endif // __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE #else - GGML_UNUSED(mma_A); - GGML_UNUSED(mma_B); + GGML_UNUSED(D); + GGML_UNUSED(A); + GGML_UNUSED(B); NO_DEVICE_CODE; #endif // NEW_MMA_AVAILABLE } - __device__ __forceinline__ mma_B_J8K8 to_mma_B() { - mma_B_J8K8 mma_B; - - int * xi = (int *) x; - int * Bxi = (int *) mma_B.x; - Bxi[0] = ggml_cuda_movmatrix(xi[0]); - Bxi[1] = ggml_cuda_movmatrix(xi[1]); - - return mma_B; - } -}; - -template <> -struct mma_C_I16J8 { - static constexpr int I = 16; - static constexpr int J = 8; - static constexpr int ne = 4; - - float x[ne] = {0.0f, 0.0f, 0.0f, 0.0f}; - - static __device__ __forceinline__ int get_i(const int l) { - const int ret = (l/2) * (I/2) + threadIdx.x / (J/2); - GGML_CUDA_ASSUME(ret >= 0); - GGML_CUDA_ASSUME(ret < I); - return ret; - } - - static __device__ __forceinline__ int get_j(const int l) { - const int ret = 2 * (threadIdx.x % (J/2)) + l%2; - GGML_CUDA_ASSUME(ret >= 0); - GGML_CUDA_ASSUME(ret < J); - return ret; - } - - __device__ __forceinline__ void mma(const mma_A_I16K8 & mma_A, const mma_B_J8K8 & mma_B) { + static __device__ __forceinline__ void mma( + tile<16, 8, half2> & D, const tile<16, 8, half2> & A, const tile<16, 8, half2> & B) { #ifdef NEW_MMA_AVAILABLE - int * Axi = (int *) mma_A.x; - int * Bxi = (int *) mma_B.x; - int * xi = (int *) x; + const int * Axi = (const int *) A.x; + const int * Bxi = (const int *) B.x; + int * Dxi = (int *) D.x; +#if __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE + asm("mma.sync.aligned.m16n8k16.row.col.f16.f16.f16.f16 {%0, %1}, {%2, %3, %4, %5}, {%6, %7}, {%0, %1};" + : "+r"(Dxi[0]), "+r"(Dxi[1]) + : "r"(Axi[0]), "r"(Axi[1]), "r"(Axi[2]), "r"(Axi[3]), "r"(Bxi[0]), "r"(Bxi[2])); + asm("mma.sync.aligned.m16n8k16.row.col.f16.f16.f16.f16 {%0, %1}, {%2, %3, %4, %5}, {%6, %7}, {%0, %1};" + : "+r"(Dxi[2]), "+r"(Dxi[3]) + : "r"(Axi[0]), "r"(Axi[1]), "r"(Axi[2]), "r"(Axi[3]), "r"(Bxi[1]), "r"(Bxi[3])); +#else + // On Turing m16n8k16 mma is not available, use 4x m8n8k8 mma instead: + asm("mma.sync.aligned.m16n8k8.row.col.f16.f16.f16.f16 {%0, %1}, {%2, %3}, {%4}, {%0, %1};" + : "+r"(Dxi[0]), "+r"(Dxi[1]) + : "r"(Axi[0]), "r"(Axi[1]), "r"(Bxi[0])); + asm("mma.sync.aligned.m16n8k8.row.col.f16.f16.f16.f16 {%0, %1}, {%2, %3}, {%4}, {%0, %1};" + : "+r"(Dxi[0]), "+r"(Dxi[1]) + : "r"(Axi[2]), "r"(Axi[3]), "r"(Bxi[2])); + asm("mma.sync.aligned.m16n8k8.row.col.f16.f16.f16.f16 {%0, %1}, {%2, %3}, {%4}, {%0, %1};" + : "+r"(Dxi[2]), "+r"(Dxi[3]) + : "r"(Axi[0]), "r"(Axi[1]), "r"(Bxi[1])); + asm("mma.sync.aligned.m16n8k8.row.col.f16.f16.f16.f16 {%0, %1}, {%2, %3}, {%4}, {%0, %1};" + : "+r"(Dxi[2]), "+r"(Dxi[3]) + : "r"(Axi[2]), "r"(Axi[3]), "r"(Bxi[3])); +#endif // __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE +#else + GGML_UNUSED(D); + GGML_UNUSED(A); + GGML_UNUSED(B); + NO_DEVICE_CODE; +#endif // NEW_MMA_AVAILABLE + } + + static __device__ __forceinline__ void mma( + tile<16, 8, float> & D, const tile<16, 8, half2> & A, const tile<8, 8, half2> & B) { +#ifdef NEW_MMA_AVAILABLE + const int * Axi = (const int *) A.x; + const int * Bxi = (const int *) B.x; + int * Dxi = (int *) D.x; #if __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE asm("mma.sync.aligned.m16n8k16.row.col.f32.f16.f16.f32 {%0, %1, %2, %3}, {%4, %5, %6, %7}, {%8, %9}, {%0, %1, %2, %3};" - : "+r"(xi[0]), "+r"(xi[1]), "+r"(xi[2]), "+r"(xi[3]) + : "+r"(Dxi[0]), "+r"(Dxi[1]), "+r"(Dxi[2]), "+r"(Dxi[3]) : "r"(Axi[0]), "r"(Axi[1]), "r"(Axi[2]), "r"(Axi[3]), "r"(Bxi[0]), "r"(Bxi[1])); #else // On Turing m16n8k16 mma is not available, use 2x m8n8k8 mma instead: asm("mma.sync.aligned.m16n8k8.row.col.f32.f16.f16.f32 {%0, %1, %2, %3}, {%4, %5}, {%6}, {%0, %1, %2, %3};" - : "+r"(xi[0]), "+r"(xi[1]), "+r"(xi[2]), "+r"(xi[3]) + : "+r"(Dxi[0]), "+r"(Dxi[1]), "+r"(Dxi[2]), "+r"(Dxi[3]) : "r"(Axi[0]), "r"(Axi[1]), "r"(Bxi[0])); asm("mma.sync.aligned.m16n8k8.row.col.f32.f16.f16.f32 {%0, %1, %2, %3}, {%4, %5}, {%6}, {%0, %1, %2, %3};" - : "+r"(xi[0]), "+r"(xi[1]), "+r"(xi[2]), "+r"(xi[3]) + : "+r"(Dxi[0]), "+r"(Dxi[1]), "+r"(Dxi[2]), "+r"(Dxi[3]) : "r"(Axi[2]), "r"(Axi[3]), "r"(Bxi[1])); #endif // __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE #else - GGML_UNUSED(mma_A); - GGML_UNUSED(mma_B); + GGML_UNUSED(D); + GGML_UNUSED(A); + GGML_UNUSED(B); NO_DEVICE_CODE; #endif // NEW_MMA_AVAILABLE } - __device__ __forceinline__ mma_B_J8K8 to_mma_B() { - mma_B_J8K8 mma_B; - mma_B.x[0] = make_half2(x[0], x[1]); - mma_B.x[1] = make_half2(x[2], x[3]); - - int * Bxi = (int *) mma_B.x; - Bxi[0] = ggml_cuda_movmatrix(Bxi[0]); - Bxi[1] = ggml_cuda_movmatrix(Bxi[1]); - - return mma_B; + static __device__ __forceinline__ void mma( + tile<16, 16, float> & D, const tile<16, 8, half2> & A, const tile<16, 8, half2> & B) { +#ifdef NEW_MMA_AVAILABLE + const int * Axi = (const int *) A.x; + const int * Bxi = (const int *) B.x; + int * Dxi = (int *) D.x; +#if __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE + asm("mma.sync.aligned.m16n8k16.row.col.f32.f16.f16.f32 {%0, %1, %2, %3}, {%4, %5, %6, %7}, {%8, %9}, {%0, %1, %2, %3};" + : "+r"(Dxi[0]), "+r"(Dxi[1]), "+r"(Dxi[2]), "+r"(Dxi[3]) + : "r"(Axi[0]), "r"(Axi[1]), "r"(Axi[2]), "r"(Axi[3]), "r"(Bxi[0]), "r"(Bxi[2])); + asm("mma.sync.aligned.m16n8k16.row.col.f32.f16.f16.f32 {%0, %1, %2, %3}, {%4, %5, %6, %7}, {%8, %9}, {%0, %1, %2, %3};" + : "+r"(Dxi[4]), "+r"(Dxi[5]), "+r"(Dxi[6]), "+r"(Dxi[7]) + : "r"(Axi[0]), "r"(Axi[1]), "r"(Axi[2]), "r"(Axi[3]), "r"(Bxi[1]), "r"(Bxi[3])); +#else + // On Turing m16n8k16 mma is not available, use 4x m8n8k8 mma instead: + asm("mma.sync.aligned.m16n8k8.row.col.f32.f16.f16.f32 {%0, %1, %2, %3}, {%4, %5}, {%6}, {%0, %1, %2, %3};" + : "+r"(Dxi[0]), "+r"(Dxi[1]), "+r"(Dxi[2]), "+r"(Dxi[3]) + : "r"(Axi[0]), "r"(Axi[1]), "r"(Bxi[0])); + asm("mma.sync.aligned.m16n8k8.row.col.f32.f16.f16.f32 {%0, %1, %2, %3}, {%4, %5}, {%6}, {%0, %1, %2, %3};" + : "+r"(Dxi[0]), "+r"(Dxi[1]), "+r"(Dxi[2]), "+r"(Dxi[3]) + : "r"(Axi[2]), "r"(Axi[3]), "r"(Bxi[2])); + asm("mma.sync.aligned.m16n8k8.row.col.f32.f16.f16.f32 {%0, %1, %2, %3}, {%4, %5}, {%6}, {%0, %1, %2, %3};" + : "+r"(Dxi[4]), "+r"(Dxi[5]), "+r"(Dxi[6]), "+r"(Dxi[7]) + : "r"(Axi[0]), "r"(Axi[1]), "r"(Bxi[1])); + asm("mma.sync.aligned.m16n8k8.row.col.f32.f16.f16.f32 {%0, %1, %2, %3}, {%4, %5}, {%6}, {%0, %1, %2, %3};" + : "+r"(Dxi[4]), "+r"(Dxi[5]), "+r"(Dxi[6]), "+r"(Dxi[7]) + : "r"(Axi[2]), "r"(Axi[3]), "r"(Bxi[3])); +#endif // __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE +#else + GGML_UNUSED(D); + GGML_UNUSED(A); + GGML_UNUSED(B); + NO_DEVICE_CODE; +#endif // NEW_MMA_AVAILABLE } - - __device__ __forceinline__ void load_generic(const float * __restrict__ xs0, const int & stride) { -#pragma unroll - for (int l = 0; l < ne; ++l) { - x[l] = xs0[get_j(l)*stride + get_i(l)]; - } - } -}; +} diff --git a/ggml/src/ggml-cuda/mmq.cu b/ggml/src/ggml-cuda/mmq.cu index 45212f66c..b36b43d54 100644 --- a/ggml/src/ggml-cuda/mmq.cu +++ b/ggml/src/ggml-cuda/mmq.cu @@ -18,7 +18,7 @@ void ggml_cuda_op_mul_mat_q( const int64_t stride00 = ne00 / ggml_blck_size(src0->type); int id = ggml_cuda_get_device(); - const int compute_capability = ggml_cuda_info().devices[id].cc; + const int cc = ggml_cuda_info().devices[id].cc; // the main device has a larger memory buffer to hold the results from all GPUs // nrows_dst == nrows of the matrix that the kernel writes into @@ -27,7 +27,8 @@ void ggml_cuda_op_mul_mat_q( // The stream-k decomposition is only faster for recent NVIDIA GPUs. // Also its fixup needs to allocate a temporary buffer in the memory pool. // There are multiple parallel CUDA streams for src1_ncols != ne11 which would introduce a race condition for this buffer. - const bool use_stream_k = compute_capability >= GGML_CUDA_CC_VOLTA && compute_capability < GGML_CUDA_CC_OFFSET_AMD && src1_ncols == ne11; + const bool use_stream_k = GGML_CUDA_CC_IS_NVIDIA(cc) && + ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_VOLTA && src1_ncols == ne11; const mmq_args args = {src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stride00, src1_padded_row_size, src1_ncols, ne11, nrows_dst, use_stream_k}; switch (src0->type) { @@ -136,7 +137,7 @@ bool ggml_cuda_should_use_mmq(enum ggml_type type, int cc, int64_t ne11) { return true; } - if (cc < GGML_CUDA_CC_DP4A) { + if (ggml_cuda_highest_compiled_arch(cc) < GGML_CUDA_CC_DP4A) { return false; } @@ -144,9 +145,9 @@ bool ggml_cuda_should_use_mmq(enum ggml_type type, int cc, int64_t ne11) { return true; #endif //GGML_CUDA_FORCE_MMQ - if (cc < GGML_CUDA_CC_OFFSET_AMD) { - return cc < GGML_CUDA_CC_VOLTA || ne11 < MMQ_DP4A_MAX_BATCH_SIZE; + if (GGML_CUDA_CC_IS_NVIDIA(cc)) { + return !fp16_mma_hardware_available(cc) || ne11 < MMQ_DP4A_MAX_BATCH_SIZE; } - return (!GGML_CUDA_CC_IS_RDNA3(cc) && !GGML_CUDA_CC_IS_CDNA(cc) && !GGML_CUDA_CC_IS_GCN(cc)) || ne11 < MMQ_DP4A_MAX_BATCH_SIZE; + return (!GGML_CUDA_CC_IS_RDNA4(cc) && !GGML_CUDA_CC_IS_RDNA3(cc) && !GGML_CUDA_CC_IS_CDNA(cc)) || ne11 < MMQ_DP4A_MAX_BATCH_SIZE; } diff --git a/ggml/src/ggml-cuda/mmq.cuh b/ggml/src/ggml-cuda/mmq.cuh index 7a2c4d85b..532358018 100644 --- a/ggml/src/ggml-cuda/mmq.cuh +++ b/ggml/src/ggml-cuda/mmq.cuh @@ -7,6 +7,8 @@ #include #include +using namespace ggml_cuda_mma; + #define MMQ_DP4A_MAX_BATCH_SIZE 64 // Max. batch size to use for dp4a MMQ kernels when FP16 tensor cores are available. #define MMQ_ITER_K 256 #define MMQ_NWARPS 8 @@ -86,12 +88,13 @@ struct tile_x_sizes { int sc; }; -static constexpr int get_mmq_x_max_host(const int cc) { +static int get_mmq_x_max_host(const int cc) { return new_mma_available(cc) ? 128 : + GGML_CUDA_CC_IS_NVIDIA(cc) && ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_VOLTA ? #ifdef GGML_CUDA_FORCE_MMQ - cc >= GGML_CUDA_CC_VOLTA && cc < GGML_CUDA_CC_OFFSET_AMD ? 128 : 64; + 128 : 64; #else - cc >= GGML_CUDA_CC_VOLTA && cc < GGML_CUDA_CC_OFFSET_AMD ? MMQ_DP4A_MAX_BATCH_SIZE : 64; + MMQ_DP4A_MAX_BATCH_SIZE : 64; #endif // GGML_CUDA_FORCE_MMQ } @@ -106,9 +109,9 @@ static constexpr __device__ int get_mmq_x_max_device() { #if __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA #ifdef GGML_CUDA_FORCE_MMQ - return MMQ_DP4A_MAX_BATCH_SIZE; -#else // GGML_CUDA_FORCE_MMQ return 128; +#else // GGML_CUDA_FORCE_MMQ + return MMQ_DP4A_MAX_BATCH_SIZE; #endif // GGML_CUDA_FORCE_MMQ #else // __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA @@ -119,8 +122,9 @@ static constexpr __device__ int get_mmq_x_max_device() { #endif // NEW_MMA_AVAILABLE } -static constexpr int get_mmq_y_host(const int cc) { - return cc >= GGML_CUDA_CC_OFFSET_AMD ? (GGML_CUDA_CC_IS_RDNA1(cc) ? 64 : 128) : (cc >= GGML_CUDA_CC_VOLTA ? 128 : 64); +static int get_mmq_y_host(const int cc) { + return GGML_CUDA_CC_IS_AMD(cc) ? (GGML_CUDA_CC_IS_RDNA1(cc) ? 64 : 128) : + ((GGML_CUDA_CC_IS_NVIDIA(cc) && ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_VOLTA) ? 128 : 64); } static constexpr __device__ int get_mmq_y_device() { @@ -645,15 +649,15 @@ template static __device__ __forceinline__ void vec_dot_q8_0_q8_1_mma( const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k00) { - typedef mma_A_I16K8 mma_A; - typedef mma_B_J8K8 mma_B; - typedef mma_C_I16J8 mma_C; + typedef tile<16, 8, int> tile_A; + typedef tile< 8, 8, int> tile_B; + typedef tile<16, 8, int> tile_C; constexpr int granularity = mmq_get_granularity_device(mmq_x); constexpr int rows_per_warp = 2 * granularity; - constexpr int ntx = rows_per_warp/mma_C::I; // Number of x minitiles per warp. + constexpr int ntx = rows_per_warp/tile_C::I; // Number of x minitiles per warp. - y += (threadIdx.y % ntx) * (mma_B::J*MMQ_TILE_Y_K); + y += (threadIdx.y % ntx) * (tile_B::I*MMQ_TILE_Y_K); const int * x_qs = (const int *) x; const float * x_df = (const float *) x_qs + 2*WARP_SIZE; @@ -661,8 +665,8 @@ static __device__ __forceinline__ void vec_dot_q8_0_q8_1_mma( const float * y_df = (const float *) y; const half2 * y_ds = (const half2 *) y; - mma_A A[ntx][WARP_SIZE/QI8_0]; - float dA[ntx][mma_C::ne/2][WARP_SIZE/QI8_0]; + tile_A A[ntx][WARP_SIZE/QI8_0]; + float dA[ntx][tile_C::ne/2][WARP_SIZE/QI8_0]; const int i0 = (threadIdx.y/ntx)*rows_per_warp; @@ -672,12 +676,12 @@ static __device__ __forceinline__ void vec_dot_q8_0_q8_1_mma( for (int k01 = 0; k01 < WARP_SIZE; k01 += QI8_0) { const int k0 = k00 + k01; - A[n][k01/QI8_0].load_ldmatrix(x_qs + (i0 + n*mma_A::I)*MMQ_MMA_TILE_X_K_Q8_0 + k0, MMQ_MMA_TILE_X_K_Q8_0); + load_ldmatrix(A[n][k01/QI8_0], x_qs + (i0 + n*tile_A::I)*MMQ_MMA_TILE_X_K_Q8_0 + k0, MMQ_MMA_TILE_X_K_Q8_0); } #pragma unroll - for (int l = 0; l < mma_C::ne/2; ++l) { - const int i = i0 + n*mma_A::I + mma_C::get_i(2*l); + for (int l = 0; l < tile_C::ne/2; ++l) { + const int i = i0 + n*tile_A::I + tile_C::get_i(2*l); #pragma unroll for (int k01 = 0; k01 < WARP_SIZE; k01 += QI8_0) { @@ -689,17 +693,17 @@ static __device__ __forceinline__ void vec_dot_q8_0_q8_1_mma( } #pragma unroll - for (int j0 = 0; j0 < mmq_x; j0 += ntx*mma_C::J) { + for (int j0 = 0; j0 < mmq_x; j0 += ntx*tile_C::J) { #pragma unroll for (int k01 = 0; k01 < WARP_SIZE; k01 += QI8_0) { - mma_B B; - float dB[mma_C::ne/2]; + tile_B B; + float dB[tile_C::ne/2]; - B.load_generic(y_qs + j0*MMQ_TILE_Y_K + k01, MMQ_TILE_Y_K); // faster than load_ldmatrix + load_generic(B, y_qs + j0*MMQ_TILE_Y_K + k01, MMQ_TILE_Y_K); // faster than load_ldmatrix #pragma unroll - for (int l = 0; l < mma_C::ne/2; ++l) { - const int j = j0 + mma_C::get_j(l); + for (int l = 0; l < tile_C::ne/2; ++l) { + const int j = j0 + tile_C::get_j(l); if (ds_layout == MMQ_Q8_1_DS_LAYOUT_D4) { dB[l] = y_df[j*MMQ_TILE_Y_K + k01/QI8_1]; @@ -710,12 +714,12 @@ static __device__ __forceinline__ void vec_dot_q8_0_q8_1_mma( #pragma unroll for (int n = 0; n < ntx; ++n) { - mma_C C; - C.mma(A[n][k01/QI8_0], B); + tile_C C; + mma(C, A[n][k01/QI8_0], B); #pragma unroll - for (int l = 0; l < mma_C::ne; ++l) { - sum[(j0/mma_C::J + n)*mma_C::ne + l] += C.x[l]*dA[n][l/2][k01/QI8_0]*dB[l%2]; + for (int l = 0; l < tile_C::ne; ++l) { + sum[(j0/tile_C::J + n)*tile_C::ne + l] += C.x[l]*dA[n][l/2][k01/QI8_0]*dB[l%2]; } } } @@ -756,23 +760,23 @@ template static __device__ __forceinline__ void vec_dot_q8_1_q8_1_mma( const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k00) { - typedef mma_A_I16K8 mma_A; - typedef mma_B_J8K8 mma_B; - typedef mma_C_I16J8 mma_C; + typedef tile<16, 8, int> tile_A; + typedef tile< 8, 8, int> tile_B; + typedef tile<16, 8, int> tile_C; constexpr int granularity = mmq_get_granularity_device(mmq_x); constexpr int rows_per_warp = 2 * granularity; - constexpr int ntx = rows_per_warp/mma_C::I; // Number of x minitiles per warp. + constexpr int ntx = rows_per_warp/tile_C::I; // Number of x minitiles per warp. - y += (threadIdx.y % ntx) * (mma_B::J*MMQ_TILE_Y_K); + y += (threadIdx.y % ntx) * (tile_B::J*MMQ_TILE_Y_K); const int * x_qs = (const int *) x; const half2 * x_dm = (const half2 *) x_qs + 2*WARP_SIZE; const int * y_qs = (const int *) y + 4; const half2 * y_dm = (const half2 *) y; - mma_A A[ntx][WARP_SIZE/QI8_1]; - float2 dmA[ntx][mma_C::ne/2][WARP_SIZE/QI8_1]; + tile_A A[ntx][WARP_SIZE/QI8_1]; + float2 dmA[ntx][tile_C::ne/2][WARP_SIZE/QI8_1]; const int i0 = (threadIdx.y/ntx)*rows_per_warp; @@ -782,12 +786,12 @@ static __device__ __forceinline__ void vec_dot_q8_1_q8_1_mma( for (int k01 = 0; k01 < WARP_SIZE; k01 += QI8_1) { const int k0 = k00 + k01; - A[n][k01/QI8_1].load_ldmatrix(x_qs + (i0 + n*mma_A::I)*MMQ_MMA_TILE_X_K_Q8_1 + k0, MMQ_MMA_TILE_X_K_Q8_1); + load_ldmatrix(A[n][k01/QI8_1], x_qs + (i0 + n*tile_A::I)*MMQ_MMA_TILE_X_K_Q8_1 + k0, MMQ_MMA_TILE_X_K_Q8_1); } #pragma unroll - for (int l = 0; l < mma_C::ne/2; ++l) { - const int i = i0 + n*mma_A::I + mma_C::get_i(2*l); + for (int l = 0; l < tile_C::ne/2; ++l) { + const int i = i0 + n*tile_A::I + tile_C::get_i(2*l); #pragma unroll for (int k01 = 0; k01 < WARP_SIZE; k01 += QI8_1) { @@ -799,30 +803,30 @@ static __device__ __forceinline__ void vec_dot_q8_1_q8_1_mma( } #pragma unroll - for (int j0 = 0; j0 < mmq_x; j0 += ntx*mma_C::J) { + for (int j0 = 0; j0 < mmq_x; j0 += ntx*tile_C::J) { #pragma unroll for (int k01 = 0; k01 < WARP_SIZE; k01 += QI8_1) { - mma_B B; - float2 dsB[mma_C::ne/2]; + tile_B B; + float2 dsB[tile_C::ne/2]; - B.load_generic(y_qs + j0*MMQ_TILE_Y_K + k01, MMQ_TILE_Y_K); // faster than load_ldmatrix + load_generic(B, y_qs + j0*MMQ_TILE_Y_K + k01, MMQ_TILE_Y_K); // faster than load_ldmatrix #pragma unroll - for (int l = 0; l < mma_C::ne/2; ++l) { - const int j = j0 + mma_C::get_j(l); + for (int l = 0; l < tile_C::ne/2; ++l) { + const int j = j0 + tile_C::get_j(l); dsB[l] = __half22float2(y_dm[j*MMQ_TILE_Y_K + k01/QI8_1]); } #pragma unroll for (int n = 0; n < ntx; ++n) { - mma_C C; - C.mma(A[n][k01/QI8_1], B); + tile_C C; + mma(C, A[n][k01/QI8_1], B); #pragma unroll - for (int l = 0; l < mma_C::ne; ++l) { - sum[(j0/mma_C::J + n)*mma_C::ne + l] += dmA[n][l/2][k01/QI8_1].x*dsB[l%2].x*C.x[l]; - sum[(j0/mma_C::J + n)*mma_C::ne + l] += dmA[n][l/2][k01/QI8_1].y*dsB[l%2].y; + for (int l = 0; l < tile_C::ne; ++l) { + sum[(j0/tile_C::J + n)*tile_C::ne + l] += dmA[n][l/2][k01/QI8_1].x*dsB[l%2].x*C.x[l]; + sum[(j0/tile_C::J + n)*tile_C::ne + l] += dmA[n][l/2][k01/QI8_1].y*dsB[l%2].y; } } } @@ -866,26 +870,26 @@ static __device__ __forceinline__ void vec_dot_q8_0_16_q8_1_mma( const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k00) { #ifdef NEW_MMA_AVAILABLE - typedef mma_A_I16K4 mma_A; - typedef mma_A_I16K8 mma_A_K8; - typedef mma_B_J8K4 mma_B; - typedef mma_C_I16J8 mma_C; + typedef tile<16, 4, int> tile_A; + typedef tile<16, 8, int> tile_A_8; + typedef tile< 8, 4, int> tile_B; + typedef tile<16, 8, int> tile_C; constexpr int granularity = mmq_get_granularity_device(mmq_x); constexpr int rows_per_warp = 2 * granularity; - constexpr int ntx = rows_per_warp/mma_C::I; // Number of x minitiles per warp. + constexpr int ntx = rows_per_warp/tile_C::I; // Number of x minitiles per warp. - y += (threadIdx.y % ntx) * (mma_B::J*MMQ_TILE_Y_K); + y += (threadIdx.y % ntx) * (tile_B::I*MMQ_TILE_Y_K); const int * x_qs = (const int *) x; const float * x_df = (const float *) x_qs + WARP_SIZE*2; const int * y_qs = (const int *) y + 4; const float * y_df = (const float *) y; - const int i0 = (threadIdx.y / ntx) * (ntx*mma_A::I); + const int i0 = (threadIdx.y / ntx) * (ntx*tile_A::I); - mma_A A[ntx][8]; - float dA[ntx][mma_C::ne/2][8]; + tile_A A[ntx][8]; + float dA[ntx][tile_C::ne/2][8]; #pragma unroll for (int n = 0; n < ntx; ++n) { @@ -893,12 +897,12 @@ static __device__ __forceinline__ void vec_dot_q8_0_16_q8_1_mma( for (int k01 = 0; k01 < WARP_SIZE; k01 += 8) { const int k0 = k00 + k01; - ((mma_A_K8 *) A[n])[k01/8].load_ldmatrix(x_qs + (i0 + n*mma_A::I)*MMQ_MMA_TILE_X_K_Q3_K + k0, MMQ_MMA_TILE_X_K_Q3_K); + load_ldmatrix(((tile_A_8 *) A[n])[k01/8], x_qs + (i0 + n*tile_A::I)*MMQ_MMA_TILE_X_K_Q3_K + k0, MMQ_MMA_TILE_X_K_Q3_K); } #pragma unroll - for (int l = 0; l < mma_C::ne/2; ++l) { - const int i = i0 + n*mma_C::I + mma_C::get_i(2*l); + for (int l = 0; l < tile_C::ne/2; ++l) { + const int i = i0 + n*tile_C::I + tile_C::get_i(2*l); #pragma unroll for (int k01 = 0; k01 < WARP_SIZE; k01 += 4) { @@ -910,38 +914,38 @@ static __device__ __forceinline__ void vec_dot_q8_0_16_q8_1_mma( } #pragma unroll - for (int j0 = 0; j0 < mmq_x; j0 += ntx*mma_C::J) { + for (int j0 = 0; j0 < mmq_x; j0 += ntx*tile_C::J) { #pragma unroll for (int k01 = 0; k01 < WARP_SIZE; k01 += QR3_K*VDR_Q3_K_Q8_1_MMQ) { - mma_B B[2]; - float dB[mma_C::ne/2]; + tile_B B[2]; + float dB[tile_C::ne/2]; // Here load_generic is faster than load_ldmatrix. - B[0].load_generic(y_qs + j0*MMQ_TILE_Y_K + (k01 + 0), MMQ_TILE_Y_K); - B[1].load_generic(y_qs + j0*MMQ_TILE_Y_K + (k01 + mma_B::K), MMQ_TILE_Y_K); + load_generic(B[0], y_qs + j0*MMQ_TILE_Y_K + (k01 + 0), MMQ_TILE_Y_K); + load_generic(B[1], y_qs + j0*MMQ_TILE_Y_K + (k01 + tile_B::J), MMQ_TILE_Y_K); #pragma unroll - for (int l = 0; l < mma_C::ne/2; ++l) { - const int j = j0 + mma_C::get_j(l); + for (int l = 0; l < tile_C::ne/2; ++l) { + const int j = j0 + tile_C::get_j(l); dB[l] = y_df[j*MMQ_TILE_Y_K + k01/QI8_1]; } #pragma unroll for (int n = 0; n < ntx; ++n) { - mma_C C[2]; - C[0].mma(A[n][k01/4 + 0], B[0]); - C[1].mma(A[n][k01/4 + 1], B[1]); + tile_C C[2]; + mma(C[0], A[n][k01/4 + 0], B[0]); + mma(C[1], A[n][k01/4 + 1], B[1]); #pragma unroll - for (int l = 0; l < mma_C::ne; ++l) { - sum[(j0/mma_C::J + n)*mma_C::ne + l] += dB[l%2]*(C[0].x[l]*dA[n][l/2][k01/4 + 0] + C[1].x[l]*dA[n][l/2][k01/4 + 1]); + for (int l = 0; l < tile_C::ne; ++l) { + sum[(j0/tile_C::J + n)*tile_C::ne + l] += dB[l%2]*(C[0].x[l]*dA[n][l/2][k01/4 + 0] + C[1].x[l]*dA[n][l/2][k01/4 + 1]); } } } } #else - GGML_UNUSED(x); GGML_UNUSED(y); GGML_UNUSED(sum); + GGML_UNUSED(x); GGML_UNUSED(y); GGML_UNUSED(sum); GGML_UNUSED(k00); NO_DEVICE_CODE; #endif // NEW_MMA_AVAILABLE } @@ -1020,7 +1024,7 @@ static __device__ __forceinline__ void vec_dot_q2_K_q8_1_dp4a( } #pragma unroll - for (int k01 = 0; k01 < WARP_SIZE; k01 += QR2_K*VDR_Q2_K_Q8_1_MMQ) { + for (int k01 = 0; k01 < WARP_SIZE/2; k01 += QR2_K*VDR_Q2_K_Q8_1_MMQ) { const int k0 = k00 + k01; #pragma unroll @@ -1031,19 +1035,34 @@ static __device__ __forceinline__ void vec_dot_q2_K_q8_1_dp4a( for (int i0 = 0; i0 < mmq_y; i0 += WARP_SIZE) { const int i = i0 + threadIdx.x; - if (k01 < WARP_SIZE/2) { - constexpr int ns = 2; - sum[j0/nwarps*mmq_y/WARP_SIZE + i0/WARP_SIZE] += vec_dot_q2_K_q8_1_impl_mmq( - &x_qs[i*(2*WARP_SIZE + 1) + k0], &y_qs[j*MMQ_TILE_Y_K + k01], - &x_dm[i*(WARP_SIZE + 1) + k0/4], k01 < WARP_SIZE/2 ? y_df[j0/nwarps].x : y_df[j0/nwarps].y, - &y_ds[j*MMQ_TILE_Y_K + (1 + k01/QI8_1)]); - } else { - constexpr int ns = 1; - sum[j0/nwarps*mmq_y/WARP_SIZE + i0/WARP_SIZE] += vec_dot_q2_K_q8_1_impl_mmq( - &x_qs[i*(2*WARP_SIZE + 1) + k0], &y_qs[j*MMQ_TILE_Y_K + k01], - &x_dm[i*(WARP_SIZE + 1) + k0/4], k01 < WARP_SIZE/2 ? y_df[j0/nwarps].x : y_df[j0/nwarps].y, - &y_ds[j*MMQ_TILE_Y_K + (1 + k01/QI8_1)]); - } + constexpr int ns = 2; + sum[j0/nwarps*mmq_y/WARP_SIZE + i0/WARP_SIZE] += vec_dot_q2_K_q8_1_impl_mmq( + &x_qs[i*(2*WARP_SIZE + 1) + k0], &y_qs[j*MMQ_TILE_Y_K + k01], + &x_dm[i*(WARP_SIZE + 1) + k0/4], k01 < WARP_SIZE/2 ? y_df[j0/nwarps].x : y_df[j0/nwarps].y, + &y_ds[j*MMQ_TILE_Y_K + (1 + k01/QI8_1)]); + } + } + } + + // Some compilers fail to unroll the loop over k01 if there is a conditional statement for ns in the inner loop. + // As a workaround 2 separate loops are used instead. +#pragma unroll + for (int k01 = WARP_SIZE/2; k01 < WARP_SIZE; k01 += QR2_K*VDR_Q2_K_Q8_1_MMQ) { + const int k0 = k00 + k01; + +#pragma unroll + for (int j0 = 0; j0 < mmq_x; j0 += nwarps) { + const int j = j0 + threadIdx.y; + +#pragma unroll + for (int i0 = 0; i0 < mmq_y; i0 += WARP_SIZE) { + const int i = i0 + threadIdx.x; + + constexpr int ns = 1; + sum[j0/nwarps*mmq_y/WARP_SIZE + i0/WARP_SIZE] += vec_dot_q2_K_q8_1_impl_mmq( + &x_qs[i*(2*WARP_SIZE + 1) + k0], &y_qs[j*MMQ_TILE_Y_K + k01], + &x_dm[i*(WARP_SIZE + 1) + k0/4], k01 < WARP_SIZE/2 ? y_df[j0/nwarps].x : y_df[j0/nwarps].y, + &y_ds[j*MMQ_TILE_Y_K + (1 + k01/QI8_1)]); } } } @@ -1054,27 +1073,27 @@ static __device__ __forceinline__ void vec_dot_q2_K_q8_1_mma( const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k00) { #ifdef NEW_MMA_AVAILABLE - typedef mma_A_I16K4 mma_A; - typedef mma_A_I16K8 mma_A_K8; - typedef mma_B_J8K4 mma_B; - typedef mma_C_I16J8 mma_C; + typedef tile<16, 4, int> tile_A; + typedef tile<16, 8, int> tile_A_8; + typedef tile< 8, 4, int> tile_B; + typedef tile<16, 8, int> tile_C; constexpr int granularity = mmq_get_granularity_device(mmq_x); constexpr int rows_per_warp = 2 * granularity; - constexpr int ntx = rows_per_warp/mma_C::I; // Number of x minitiles per warp. + constexpr int ntx = rows_per_warp/tile_C::I; // Number of x minitiles per warp. - y += (threadIdx.y % ntx) * (mma_B::J*MMQ_TILE_Y_K); + y += (threadIdx.y % ntx) * (tile_B::I*MMQ_TILE_Y_K); const int * x_qs = (const int *) x; const half2 * x_dm = (const half2 *) x_qs + WARP_SIZE*2; const int * y_qs = (const int *) y + 4; const half2 * y_ds = (const half2 *) y; - const int i0 = (threadIdx.y / ntx) * (ntx*mma_A::I); + const int i0 = (threadIdx.y / ntx) * (ntx*tile_A::I); - mma_A A[ntx][8]; - float dA[ntx][mma_C::ne/2][8]; - float mA[ntx][mma_C::ne/2][8]; + tile_A A[ntx][8]; + float dA[ntx][tile_C::ne/2][8]; + float mA[ntx][tile_C::ne/2][8]; #pragma unroll for (int n = 0; n < ntx; ++n) { @@ -1082,15 +1101,15 @@ static __device__ __forceinline__ void vec_dot_q2_K_q8_1_mma( for (int k01 = 0; k01 < WARP_SIZE; k01 += QI8_1) { const int k0 = k00 + k01; - ((mma_A_K8 *) A[n])[k01/QI8_1].load_ldmatrix(x_qs + (i0 + n*mma_A::I)*MMQ_MMA_TILE_X_K_Q2_K + k0, MMQ_MMA_TILE_X_K_Q2_K); + load_ldmatrix(((tile_A_8 *) A[n])[k01/QI8_1], x_qs + (i0 + n*tile_A::I)*MMQ_MMA_TILE_X_K_Q2_K + k0, MMQ_MMA_TILE_X_K_Q2_K); } } #pragma unroll for (int n = 0; n < ntx; ++n) { #pragma unroll - for (int l = 0; l < mma_C::ne/2; ++l) { - const int i = i0 + n*mma_C::I + mma_C::get_i(2*l); + for (int l = 0; l < tile_C::ne/2; ++l) { + const int i = i0 + n*tile_C::I + tile_C::get_i(2*l); #pragma unroll for (int k01 = 0; k01 < WARP_SIZE; k01 += QI8_1/2) { @@ -1105,58 +1124,58 @@ static __device__ __forceinline__ void vec_dot_q2_K_q8_1_mma( } #pragma unroll - for (int j0 = 0; j0 < mmq_x; j0 += ntx*mma_C::J) { - float2 dB[mma_C::ne/2]; + for (int j0 = 0; j0 < mmq_x; j0 += ntx*tile_C::J) { + float2 dB[tile_C::ne/2]; #pragma unroll - for (int l = 0; l < mma_C::ne/2; ++l) { - const int j = j0 + mma_C::get_j(l); + for (int l = 0; l < tile_C::ne/2; ++l) { + const int j = j0 + tile_C::get_j(l); dB[l] = __half22float2(y_ds[j*MMQ_TILE_Y_K]); } #pragma unroll for (int k01 = 0; k01 < WARP_SIZE; k01 += QI8_1) { - mma_B B[2]; + tile_B B[2]; // Here load_generic is faster than load_ldmatrix. - B[0].load_generic(y_qs + j0*MMQ_TILE_Y_K + (k01 + 0), MMQ_TILE_Y_K); - B[1].load_generic(y_qs + j0*MMQ_TILE_Y_K + (k01 + mma_B::K), MMQ_TILE_Y_K); + load_generic(B[0], y_qs + j0*MMQ_TILE_Y_K + (k01 + 0), MMQ_TILE_Y_K); + load_generic(B[1], y_qs + j0*MMQ_TILE_Y_K + (k01 + tile_B::J), MMQ_TILE_Y_K); - mma_C Cm[2]; + tile_C Cm[2]; if (k01 >= WARP_SIZE * 3/4) { - mma_A A1; + tile_A A1; A1.x[0] = 0x01010101; A1.x[1] = 0x01010101; - Cm[0].mma(A1, B[0]); - Cm[1].mma(A1, B[1]); + mma(Cm[0], A1, B[0]); + mma(Cm[1], A1, B[1]); } #pragma unroll for (int n = 0; n < ntx; ++n) { - mma_C Cd[2]; + tile_C Cd[2]; - Cd[0].mma(A[n][k01/4 + 0], B[0]); - Cd[1].mma(A[n][k01/4 + 1], B[1]); + mma(Cd[0], A[n][k01/4 + 0], B[0]); + mma(Cd[1], A[n][k01/4 + 1], B[1]); #pragma unroll - for (int l = 0; l < mma_C::ne; ++l) { + for (int l = 0; l < tile_C::ne; ++l) { float tmp = Cd[0].x[l]*dA[n][l/2][k01/4 + 0] + Cd[1].x[l]*dA[n][l/2][k01/4 + 1]; if (k01 >= WARP_SIZE * 3/4) { tmp -= Cm[0].x[l]*mA[n][l/2][k01/4 + 0] + Cm[1].x[l]*mA[n][l/2][k01/4 + 1]; } - sum[(j0/mma_C::J + n)*mma_C::ne + l] += tmp*(k01 < WARP_SIZE/2 ? dB[l%2].x : dB[l%2].y); + sum[(j0/tile_C::J + n)*tile_C::ne + l] += tmp*(k01 < WARP_SIZE/2 ? dB[l%2].x : dB[l%2].y); } } } #pragma unroll for (int k01 = 0; k01 < WARP_SIZE * 3/4; k01 += QI8_1) { - float2 sB[mma_C::ne/2]; + float2 sB[tile_C::ne/2]; #pragma unroll - for (int l = 0; l < mma_C::ne/2; ++l) { - const int j = j0 + mma_C::get_j(l); + for (int l = 0; l < tile_C::ne/2; ++l) { + const int j = j0 + tile_C::get_j(l); sB[l] = __half22float2(y_ds[j*MMQ_TILE_Y_K + (1 + k01/QI8_1)]); } @@ -1164,15 +1183,15 @@ static __device__ __forceinline__ void vec_dot_q2_K_q8_1_mma( #pragma unroll for (int n = 0; n < ntx; ++n) { #pragma unroll - for (int l = 0; l < mma_C::ne; ++l) { - sum[(j0/mma_C::J + n)*mma_C::ne + l] -= mA[n][l/2][k01/4 + 0]*sB[l%2].x; - sum[(j0/mma_C::J + n)*mma_C::ne + l] -= mA[n][l/2][k01/4 + 1]*sB[l%2].y; + for (int l = 0; l < tile_C::ne; ++l) { + sum[(j0/tile_C::J + n)*tile_C::ne + l] -= mA[n][l/2][k01/4 + 0]*sB[l%2].x; + sum[(j0/tile_C::J + n)*tile_C::ne + l] -= mA[n][l/2][k01/4 + 1]*sB[l%2].y; } } } } #else - GGML_UNUSED(x); GGML_UNUSED(y); GGML_UNUSED(sum); + GGML_UNUSED(x); GGML_UNUSED(y); GGML_UNUSED(sum); GGML_UNUSED(k00); NO_DEVICE_CODE; #endif // NEW_MMA_AVAILABLE } @@ -1249,7 +1268,7 @@ template static __device__ __forceinlin const float d = bxi->d; #pragma unroll - for (int l = 0; l < sizeof(int); ++l) { + for (int l = 0; l < int(sizeof(int)); ++l) { x_df[i*MMQ_MMA_TILE_X_K_Q3_K + sizeof(int)*(threadIdx.x % (WARP_SIZE/8)) + l] = d*sc8[l]; } #else @@ -1372,7 +1391,7 @@ template static __device__ __forceinlin const half2 dm = bxi->dm * make_half2(1.0f, -1.0f); #pragma unroll - for (int l = 0; l < sizeof(int); ++l) { + for (int l = 0; l < int(sizeof(int)); ++l) { x_dm[i*MMQ_MMA_TILE_X_K_Q8_1 + sizeof(int)*ksc + l] = dm*make_half2(sc8[l], m8[l]); } } @@ -1513,7 +1532,7 @@ template static __device__ __forceinlin const half2 dm = bxi->dm * make_half2(1.0f, -1.0f); #pragma unroll - for (int l = 0; l < sizeof(int); ++l) { + for (int l = 0; l < int(sizeof(int)); ++l) { x_dm[i*MMQ_MMA_TILE_X_K_Q8_1 + sizeof(int)*ksc + l] = dm*make_half2(sc8[l], m8[l]); } } @@ -1706,15 +1725,15 @@ static __device__ __forceinline__ void vec_dot_q6_K_q8_1_mma( const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k00) { #ifdef NEW_MMA_AVAILABLE - typedef mma_A_I16K4 mma_A; - typedef mma_B_J8K4 mma_B; - typedef mma_C_I16J8 mma_C; + typedef tile<16, 4, int> tile_A; + typedef tile< 8, 4, int> tile_B; + typedef tile<16, 8, int> tile_C; constexpr int granularity = mmq_get_granularity_device(mmq_x); constexpr int rows_per_warp = 2 * granularity; - constexpr int ntx = rows_per_warp/mma_C::I; // Number of x minitiles per warp. + constexpr int ntx = rows_per_warp/tile_C::I; // Number of x minitiles per warp. - y += (threadIdx.y % ntx) * (mma_B::J*MMQ_TILE_Y_K); + y += (threadIdx.y % ntx) * (tile_B::I*MMQ_TILE_Y_K); const int * x_qs = (const int *) x; const float * x_df = (const float *) x_qs + WARP_SIZE*2; @@ -1722,11 +1741,11 @@ static __device__ __forceinline__ void vec_dot_q6_K_q8_1_mma( const int * y_qs = (const int *) y + 4; const float * y_df = (const float *) y; - const int i0 = (threadIdx.y / ntx) * (ntx*mma_A::I); + const int i0 = (threadIdx.y / ntx) * (ntx*tile_A::I); - mma_A A[ntx][8]; - int scA[ntx][mma_C::ne/2][8]; - float dA[ntx][mma_C::ne/2]; + tile_A A[ntx][8]; + int scA[ntx][tile_C::ne/2][8]; + float dA[ntx][tile_C::ne/2]; #pragma unroll for (int n = 0; n < ntx; ++n) { @@ -1734,8 +1753,8 @@ static __device__ __forceinline__ void vec_dot_q6_K_q8_1_mma( for (int k01 = 0; k01 < WARP_SIZE; k01 += 8) { const int k0 = k00 + k01; - A[n][k01/4 + 0].load_ldmatrix(x_qs + (i0 + n*mma_A::I)*MMQ_MMA_TILE_X_K_Q6_K + (k0 + 0), MMQ_MMA_TILE_X_K_Q6_K); - A[n][k01/4 + 1].load_ldmatrix(x_qs + (i0 + n*mma_A::I)*MMQ_MMA_TILE_X_K_Q6_K + (k0 + mma_A::K), MMQ_MMA_TILE_X_K_Q6_K); + load_ldmatrix(A[n][k01/4 + 0], x_qs + (i0 + n*tile_A::I)*MMQ_MMA_TILE_X_K_Q6_K + (k0 + 0), MMQ_MMA_TILE_X_K_Q6_K); + load_ldmatrix(A[n][k01/4 + 1], x_qs + (i0 + n*tile_A::I)*MMQ_MMA_TILE_X_K_Q6_K + (k0 + tile_A::J), MMQ_MMA_TILE_X_K_Q6_K); } #pragma unroll @@ -1743,8 +1762,8 @@ static __device__ __forceinline__ void vec_dot_q6_K_q8_1_mma( const int k0 = k00 + k01; #pragma unroll - for (int l = 0; l < mma_C::ne/2; ++l) { - const int i = i0 + n*mma_C::I + mma_C::get_i(2*l); + for (int l = 0; l < tile_C::ne/2; ++l) { + const int i = i0 + n*tile_C::I + tile_C::get_i(2*l); const int sc_packed = x_sc[i*MMQ_MMA_TILE_X_K_Q6_K + k0/16]; const int8_t * sc = (const int8_t *) &sc_packed; @@ -1757,41 +1776,41 @@ static __device__ __forceinline__ void vec_dot_q6_K_q8_1_mma( } #pragma unroll - for (int l = 0; l < mma_C::ne/2; ++l) { - const int i = i0 + n*mma_C::I + mma_C::get_i(2*l); + for (int l = 0; l < tile_C::ne/2; ++l) { + const int i = i0 + n*tile_C::I + tile_C::get_i(2*l); dA[n][l] = x_df[i*MMQ_MMA_TILE_X_K_Q6_K]; } } #pragma unroll - for (int j0 = 0; j0 < mmq_x; j0 += ntx*mma_C::J) { - float tmp[ntx][mma_C::ne] = {{0.0f}}; + for (int j0 = 0; j0 < mmq_x; j0 += ntx*tile_C::J) { + float tmp[ntx][tile_C::ne] = {{0.0f}}; #pragma unroll for (int k01 = 0; k01 < WARP_SIZE; k01 += 8) { - mma_B B[2]; - float dB[mma_C::ne/2]; + tile_B B[2]; + float dB[tile_C::ne/2]; // Here load_generic is faster than load_ldmatrix. - B[0].load_generic(y_qs + j0*MMQ_TILE_Y_K + 0 + k01, MMQ_TILE_Y_K); - B[1].load_generic(y_qs + j0*MMQ_TILE_Y_K + mma_B::K + k01, MMQ_TILE_Y_K); + load_generic(B[0], y_qs + j0*MMQ_TILE_Y_K + 0 + k01, MMQ_TILE_Y_K); + load_generic(B[1], y_qs + j0*MMQ_TILE_Y_K + tile_B::J + k01, MMQ_TILE_Y_K); #pragma unroll - for (int l = 0; l < mma_C::ne/2; ++l) { - const int j = j0 + mma_C::get_j(l); + for (int l = 0; l < tile_C::ne/2; ++l) { + const int j = j0 + tile_C::get_j(l); dB[l] = y_df[j*MMQ_TILE_Y_K + k01/QI8_1]; } #pragma unroll for (int n = 0; n < ntx; ++n) { - mma_C C[2]; - C[0].mma(A[n][k01/4 + 0], B[0]); - C[1].mma(A[n][k01/4 + 1], B[1]); + tile_C C[2]; + mma(C[0], A[n][k01/4 + 0], B[0]); + mma(C[1], A[n][k01/4 + 1], B[1]); #pragma unroll - for (int l = 0; l < mma_C::ne; ++l) { + for (int l = 0; l < tile_C::ne; ++l) { tmp[n][l] += (C[0].x[l]*scA[n][l/2][k01/4 + 0] + C[1].x[l]*scA[n][l/2][k01/4 + 1])*dB[l%2]; } } @@ -1800,13 +1819,13 @@ static __device__ __forceinline__ void vec_dot_q6_K_q8_1_mma( #pragma unroll for (int n = 0; n < ntx; ++n) { #pragma unroll - for (int l = 0; l < mma_C::ne; ++l) { - sum[(j0/mma_C::J + n)*mma_C::ne + l] += tmp[n][l]*dA[n][l/2]; + for (int l = 0; l < tile_C::ne; ++l) { + sum[(j0/tile_C::J + n)*tile_C::ne + l] += tmp[n][l]*dA[n][l/2]; } } } #else - GGML_UNUSED(x); GGML_UNUSED(y); GGML_UNUSED(sum); + GGML_UNUSED(x); GGML_UNUSED(y); GGML_UNUSED(sum); GGML_UNUSED(k00); NO_DEVICE_CODE; #endif // NEW_MMA_AVAILABLE } @@ -2310,36 +2329,36 @@ template static __device__ __forceinline__ void mmq_write_back_mma( const float * __restrict__ sum, float * __restrict__ dst, const int & stride, const int & i_max, const int & j_max) { - typedef mma_C_I16J8 mma_C; + typedef tile<16, 8, int> tile_C; constexpr int granularity = mmq_get_granularity_device(mmq_x); constexpr int rows_per_warp = 2 * granularity; - constexpr int ntx = rows_per_warp/mma_C::I; // Number of x minitiles per warp. + constexpr int ntx = rows_per_warp/tile_C::I; // Number of x minitiles per warp. - const int i0 = (threadIdx.y / ntx) * (ntx*mma_C::I); + const int i0 = (threadIdx.y / ntx) * (ntx*tile_C::I); #ifdef NEW_MMA_AVAILABLE - static_assert(nwarps*mma_C::I == mmq_y, "nwarps*mma_C::I != mmq_y"); + static_assert(nwarps*tile_C::I == mmq_y, "nwarps*tile_C::I != mmq_y"); #endif // NEW_MMA_AVAILABLE #pragma unroll - for (int j0 = 0; j0 < mmq_x; j0 += ntx*mma_C::J) { + for (int j0 = 0; j0 < mmq_x; j0 += ntx*tile_C::J) { #pragma unroll for (int n = 0; n < ntx; ++n) { #pragma unroll - for (int l = 0; l < mma_C::ne; ++l) { - const int j = j0 + (threadIdx.y % ntx) * mma_C::J + mma_C::get_j(l); + for (int l = 0; l < tile_C::ne; ++l) { + const int j = j0 + (threadIdx.y % ntx) * tile_C::J + tile_C::get_j(l); if (j > j_max) { continue; } - const int i = i0 + n*mma_C::I + mma_C::get_i(l); + const int i = i0 + n*tile_C::I + tile_C::get_i(l); if (need_check && i > i_max) { continue; } - dst[j*stride + i] = sum[(j0/mma_C::J + n)*mma_C::ne + l]; + dst[j*stride + i] = sum[(j0/tile_C::J + n)*tile_C::ne + l]; } } } @@ -2566,6 +2585,8 @@ static __device__ void mul_mat_q_process_tile( } else { write_back(sum, dst + jt*mmq_x*ne0 + it*mmq_y, ne0, tile_x_max_i, tile_y_max_j); } + + GGML_UNUSED(ne00); GGML_UNUSED(ne10); } @@ -2573,9 +2594,9 @@ static __device__ void mul_mat_q_process_tile( template #if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__) -#if defined(RDNA3) || defined(RDNA2) || defined(CDNA) || defined(GCN) +#if defined(RDNA4) || defined(RDNA3) || defined(RDNA2) || defined(CDNA) || defined(GCN) __launch_bounds__(WARP_SIZE*nwarps, 2) -#endif // defined(RDNA3) || defined(RDNA2) || defined(CDNA) || defined(GCN) +#endif // defined(RDNA4) || defined(RDNA3) || defined(RDNA2) || defined(CDNA) || defined(GCN) #else #if __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA __launch_bounds__(WARP_SIZE*nwarps, 1) @@ -2691,7 +2712,7 @@ static __global__ void mul_mat_q_stream_k_fixup( const int it = (kbc_stop - jt*(blocks_per_ne00*nty)) / blocks_per_ne00; // Skip fixup tile if it's unrelated to the output tile assigned to this CUDA block: - if (it != blockIdx.x || jt != blockIdx.y) { + if ((unsigned)it != blockIdx.x || (unsigned)jt != blockIdx.y) { continue; } @@ -2768,14 +2789,14 @@ static void launch_mul_mat_q(ggml_backend_cuda_context & ctx, const mmq_args & a const int shmem = mmq_get_shmem(mmq_x, mmq_y, cc); -#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) +#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && !defined(GGML_USE_MUSA) static bool shmem_limit_raised[GGML_CUDA_MAX_DEVICES] = {false}; if (!shmem_limit_raised[id]) { CUDA_CHECK(cudaFuncSetAttribute(mul_mat_q, cudaFuncAttributeMaxDynamicSharedMemorySize, shmem)); CUDA_CHECK(cudaFuncSetAttribute(mul_mat_q, cudaFuncAttributeMaxDynamicSharedMemorySize, shmem)); shmem_limit_raised[id] = true; } -#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) +#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && !defined(GGML_USE_MUSA) const int nty = (args.ne01 + mmq_y - 1) / mmq_y; const int ntx = (args.ne11 + mmq_x - 1) / mmq_x; @@ -2821,14 +2842,13 @@ static void launch_mul_mat_q(ggml_backend_cuda_context & ctx, const mmq_args & a template void mul_mat_q_case(ggml_backend_cuda_context & ctx, const mmq_args & args, cudaStream_t stream) { const int id = ggml_cuda_get_device(); - const int nsm = ggml_cuda_info().devices[id].nsm; const int cc = ggml_cuda_info().devices[id].cc; const int smpbo = ggml_cuda_info().devices[id].smpbo; const int mmq_x_max = get_mmq_x_max_host(cc); const int mmq_y = get_mmq_y_host(cc); const int block_num_y = (args.ne01 + mmq_y - 1) / mmq_y; - const bool use_stream_k = cc >= GGML_CUDA_CC_VOLTA && cc < GGML_CUDA_CC_OFFSET_AMD; + const bool use_stream_k = GGML_CUDA_CC_IS_NVIDIA(cc) && ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_VOLTA; int mmq_x_best = 0; int nparts_best = INT_MAX; diff --git a/ggml/src/ggml-cuda/mmv.cu b/ggml/src/ggml-cuda/mmv.cu index f89ed03b5..b39961cd1 100644 --- a/ggml/src/ggml-cuda/mmv.cu +++ b/ggml/src/ggml-cuda/mmv.cu @@ -29,7 +29,7 @@ static __global__ void mul_mat_vec( __syncthreads(); } - float sumf; + float sumf = 0.0f; if constexpr (std::is_same::value) { const half2 * x2 = (const half2 *) x; diff --git a/ggml/src/ggml-cuda/mmvq.cu b/ggml/src/ggml-cuda/mmvq.cu index 4fb466ca0..eef8585a7 100644 --- a/ggml/src/ggml-cuda/mmvq.cu +++ b/ggml/src/ggml-cuda/mmvq.cu @@ -47,11 +47,89 @@ static constexpr __device__ int get_vdr_mmvq(ggml_type type) { 1; } +enum mmvq_parameter_table_id { + MMVQ_PARAMETERS_GENERIC = 0, + MMVQ_PARAMETERS_GCN, + MMVQ_PARAMETERS_RDNA2 +}; + +static constexpr __device__ mmvq_parameter_table_id get_device_table_id() { +#if defined(RDNA2) || defined(RDNA3) || defined(RDNA4) + return MMVQ_PARAMETERS_RDNA2; +#elif defined(GCN) || defined(CDNA) + return MMVQ_PARAMETERS_GCN; +#else + return MMVQ_PARAMETERS_GENERIC; +#endif +} + +static __host__ mmvq_parameter_table_id get_device_table_id(int cc) { + if (GGML_CUDA_CC_IS_RDNA2(cc) || GGML_CUDA_CC_IS_RDNA3(cc) || GGML_CUDA_CC_IS_RDNA4(cc)) { + return MMVQ_PARAMETERS_RDNA2; + } + if (GGML_CUDA_CC_IS_GCN(cc) || GGML_CUDA_CC_IS_CDNA(cc)) { + return MMVQ_PARAMETERS_GCN; + } + return MMVQ_PARAMETERS_GENERIC; +} + +static constexpr __host__ __device__ int calc_nwarps(int ncols_y, mmvq_parameter_table_id table_id) { + if (table_id == MMVQ_PARAMETERS_GENERIC) { + switch (ncols_y) { + case 1: + case 2: + case 3: + case 4: + return 4; + case 5: + case 6: + case 7: + case 8: + return 2; + default: + return 1; + } + } else if (table_id == MMVQ_PARAMETERS_GCN) { + switch (ncols_y) { + case 1: + case 2: + case 3: + case 4: + return 2; + case 5: + case 6: + case 7: + case 8: + default: + return 1; + } + } + return 1; +} + +static constexpr __host__ __device__ int calc_rows_per_block(int ncols_y, int table_id) { + if (table_id == MMVQ_PARAMETERS_GENERIC || table_id == MMVQ_PARAMETERS_GCN) { + switch (ncols_y) { + case 1: + return 1; + case 2: + case 3: + case 4: + case 5: + case 6: + case 7: + case 8: + return 2; + default: + return 1; + } + } + return 1; +} + template -#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) // tell the compiler to use as many registers as it wants, see nwarps definition below -__launch_bounds__((ncols_y <= 4 ? 4 : 2)*WARP_SIZE, 1) -#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) +__launch_bounds__(calc_nwarps(ncols_y, get_device_table_id())*ggml_cuda_get_physical_warp_size(), 1) static __global__ void mul_mat_vec_q( const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst, const int ncols_x, const int nrows_x, const int nrows_y, const int nrows_dst) { @@ -59,25 +137,21 @@ static __global__ void mul_mat_vec_q( constexpr int qk = ggml_cuda_type_traits::qk; constexpr int qi = ggml_cuda_type_traits::qi; constexpr int vdr = get_vdr_mmvq(type); + constexpr mmvq_parameter_table_id table_id = get_device_table_id(); + constexpr int nwarps = calc_nwarps(ncols_y, table_id); + constexpr int rows_per_cuda_block = calc_rows_per_block(ncols_y, table_id); + constexpr int warp_size = ggml_cuda_get_physical_warp_size(); constexpr vec_dot_q_cuda_t vec_dot_q_cuda = get_vec_dot_q_cuda(type); -#if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__) && (defined(RDNA2) || defined(RDNA3)) - constexpr int nwarps = 1; - constexpr int rows_per_cuda_block = 1; -#else - constexpr int nwarps = ncols_y <= 4 ? 4 : 2; - constexpr int rows_per_cuda_block = ncols_y == 1 ? 1 : 2; -#endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__) && !defined(RDNA2) && !defined(RDNA3) - - const int tid = WARP_SIZE*threadIdx.y + threadIdx.x; + const int tid = warp_size*threadIdx.y + threadIdx.x; const int row0 = rows_per_cuda_block*blockIdx.x; const int blocks_per_row_x = ncols_x / qk; const int blocks_per_col_y = nrows_y / QK8_1; - constexpr int blocks_per_iter = vdr * nwarps*WARP_SIZE / qi; + constexpr int blocks_per_iter = vdr * nwarps*warp_size / qi; -// partial sum for each thread - float tmp[ncols_y][rows_per_cuda_block] = {0.0f}; + // partial sum for each thread + float tmp[ncols_y][rows_per_cuda_block] = {{0.0f}}; const block_q8_1 * y = (const block_q8_1 *) vy; @@ -96,7 +170,7 @@ static __global__ void mul_mat_vec_q( } } - __shared__ float tmp_shared[nwarps-1 > 0 ? nwarps-1 : 1][ncols_y][rows_per_cuda_block][WARP_SIZE]; + __shared__ float tmp_shared[nwarps-1 > 0 ? nwarps-1 : 1][ncols_y][rows_per_cuda_block][warp_size]; if (threadIdx.y > 0) { #pragma unroll for (int j = 0; j < ncols_y; ++j) { @@ -120,13 +194,22 @@ static __global__ void mul_mat_vec_q( for (int l = 0; l < nwarps-1; ++l) { tmp[j][i] += tmp_shared[l][j][i][threadIdx.x]; } - tmp[j][i] = warp_reduce_sum(tmp[j][i]); + tmp[j][i] = warp_reduce_sum(tmp[j][i]); } - if (threadIdx.x < rows_per_cuda_block && (rows_per_cuda_block == 1 || row0 + threadIdx.x < nrows_dst)) { + if (threadIdx.x < rows_per_cuda_block && (rows_per_cuda_block == 1 || row0 + threadIdx.x < (unsigned)nrows_dst)) { dst[j*nrows_dst + row0 + threadIdx.x] = tmp[j][threadIdx.x]; } } + + GGML_UNUSED(nrows_x); +} + +static std::pair calc_launch_params(const int ncols_y, const int nrows_x, const int warp_size, const mmvq_parameter_table_id table_id) { + const int64_t nblocks = (nrows_x + calc_rows_per_block(ncols_y, table_id) - 1) / calc_rows_per_block(ncols_y, table_id); + const dim3 block_nums(nblocks, 1, 1); + const dim3 block_dims(warp_size, calc_nwarps(ncols_y, table_id), 1); + return {block_nums, block_dims}; } template @@ -137,65 +220,67 @@ static void mul_mat_vec_q_cuda( GGML_ASSERT(ncols_x % ggml_blck_size(type) == 0); GGML_ASSERT(ncols_y <= MMVQ_MAX_BATCH_SIZE); - int id = ggml_cuda_get_device(); - - int64_t nwarps = 1; - int64_t rows_per_cuda_block = 1; - - if (ggml_cuda_info().devices[id].cc < GGML_CUDA_CC_RDNA2) { // NVIDIA and AMD older than RDNA2 - switch(ncols_y) { - case 1: - nwarps = 4; - rows_per_cuda_block = 1; - break; - case 2: - case 3: - case 4: - nwarps = 4; - rows_per_cuda_block = 2; - break; - case 5: - case 6: - case 7: - case 8: - nwarps = 2; - rows_per_cuda_block = 2; - break; - default: - GGML_ABORT("fatal error"); - break; - } - } - - const int64_t nblocks = (nrows_x + rows_per_cuda_block - 1) / rows_per_cuda_block; - const dim3 block_nums(nblocks, 1, 1); - const dim3 block_dims(WARP_SIZE, nwarps, 1); + const int device = ggml_cuda_get_device(); + const int warp_size = ggml_cuda_info().devices[device].warp_size; + const mmvq_parameter_table_id table_id = get_device_table_id(ggml_cuda_info().devices[device].cc); switch (ncols_y) { case 1: - mul_mat_vec_q<<>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst); + { + constexpr int c_ncols_y = 1; + std::pair dims = calc_launch_params(c_ncols_y, nrows_x, warp_size, table_id); + mul_mat_vec_q<<>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst); break; + } case 2: - mul_mat_vec_q<<>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst); + { + constexpr int c_ncols_y = 2; + std::pair dims = calc_launch_params(c_ncols_y, nrows_x, warp_size, table_id); + mul_mat_vec_q<<>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst); break; + } case 3: - mul_mat_vec_q<<>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst); + { + constexpr int c_ncols_y = 3; + std::pair dims = calc_launch_params(c_ncols_y, nrows_x, warp_size, table_id); + mul_mat_vec_q<<>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst); break; + } case 4: - mul_mat_vec_q<<>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst); + { + constexpr int c_ncols_y = 4; + std::pair dims = calc_launch_params(c_ncols_y, nrows_x, warp_size, table_id); + mul_mat_vec_q<<>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst); break; + } case 5: - mul_mat_vec_q<<>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst); + { + constexpr int c_ncols_y = 5; + std::pair dims = calc_launch_params(c_ncols_y, nrows_x, warp_size, table_id); + mul_mat_vec_q<<>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst); break; + } case 6: - mul_mat_vec_q<<>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst); + { + constexpr int c_ncols_y = 6; + std::pair dims = calc_launch_params(c_ncols_y, nrows_x, warp_size, table_id); + mul_mat_vec_q<<>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst); break; + } case 7: - mul_mat_vec_q<<>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst); + { + constexpr int c_ncols_y = 7; + std::pair dims = calc_launch_params(c_ncols_y, nrows_x, warp_size, table_id); + mul_mat_vec_q<<>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst); break; + } case 8: - mul_mat_vec_q<<>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst); + { + constexpr int c_ncols_y = 8; + std::pair dims = calc_launch_params(c_ncols_y, nrows_x, warp_size, table_id); + mul_mat_vec_q<<>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst); break; + } default: GGML_ABORT("fatal error"); break; diff --git a/ggml/src/ggml-cuda/norm.cu b/ggml/src/ggml-cuda/norm.cu index f127616ed..0020dbcec 100644 --- a/ggml/src/ggml-cuda/norm.cu +++ b/ggml/src/ggml-cuda/norm.cu @@ -201,6 +201,85 @@ static __global__ void rms_norm_back_f32( } } +// template +// static __global__ void l2_norm_f32(const float * x, float * dst, const int ncols, const float eps) { +// const int row = blockIdx.x*blockDim.y + threadIdx.y; +// const int tid = threadIdx.x; + +// float tmp = 0.0f; // partial sum for thread in warp + +// for (int col = tid; col < ncols; col += block_size) { +// const float xi = x[row*ncols + col]; +// tmp += xi * xi; +// } + +// // sum up partial sums +// tmp = warp_reduce_sum(tmp); +// if (block_size > WARP_SIZE) { +// __shared__ float s_sum[32]; +// int warp_id = threadIdx.x / WARP_SIZE; +// int lane_id = threadIdx.x % WARP_SIZE; +// if (lane_id == 0) { +// s_sum[warp_id] = tmp; +// } +// __syncthreads(); +// tmp = s_sum[lane_id]; +// tmp = warp_reduce_sum(tmp); +// } + +// // from https://pytorch.org/docs/stable/generated/torch.nn.functional.normalize.html +// const float scale = rsqrtf(fmaxf(tmp, eps * eps)); + +// for (int col = tid; col < ncols; col += block_size) { +// dst[row*ncols + col] = scale * x[row*ncols + col]; +// } +// } + +template +static __global__ void l2_norm_f32( + const float * x, float * dst, const int ncols, const int64_t stride_row, const int64_t stride_channel, + const int64_t stride_sample, const float eps) { + const int nrows = gridDim.x; + const int nchannels = gridDim.y; + + const int row = blockIdx.x; + const int channel = blockIdx.y; + const int sample = blockIdx.z; + const int tid = threadIdx.x; + + x += sample*stride_sample + channel*stride_channel + row*stride_row; + dst += ((sample*nchannels + channel)*nrows + row)*ncols; + + float tmp = 0.0f; // partial sum for thread in warp + + for (int col = tid; col < ncols; col += block_size) { + const float xi = x[col]; + tmp += xi * xi; + } + + // sum up partial sums + tmp = warp_reduce_sum(tmp); + if constexpr (block_size > WARP_SIZE) { + static_assert(block_size == 1024, "unexpected block_size"); + __shared__ float s_sum[32]; + const int warp_id = threadIdx.x / WARP_SIZE; + const int lane_id = threadIdx.x % WARP_SIZE; + if (lane_id == 0) { + s_sum[warp_id] = tmp; + } + __syncthreads(); + tmp = s_sum[lane_id]; + tmp = warp_reduce_sum(tmp); + } + + // from https://pytorch.org/docs/stable/generated/torch.nn.functional.normalize.html + const float scale = rsqrtf(fmaxf(tmp, eps * eps)); + + for (int col = tid; col < ncols; col += block_size) { + dst[col] = scale * x[col]; + } +} + static void norm_f32_cuda( const float * x, float * dst, const int ncols, const int nrows, const int nchannels, const int nsamples, const int64_t stride_row, const int64_t stride_channel, const int64_t stride_sample, const float eps, cudaStream_t stream) { @@ -248,6 +327,19 @@ static void rms_norm_back_f32_cuda(const float * grad, const float * xf, float * } } +static void l2_norm_f32_cuda( + const float * x, float * dst, const int ncols, const int nrows, const int nchannels, const int nsamples, + const int64_t stride_row, const int64_t stride_channel, const int64_t stride_sample, const float eps, cudaStream_t stream) { + const dim3 blocks_num(nrows, nchannels, nsamples); + if (ncols < 1024) { + const dim3 block_dims(WARP_SIZE, 1, 1); + l2_norm_f32<<>>(x, dst, ncols, stride_row, stride_channel, stride_sample, eps); + } else { + const dim3 block_dims(1024, 1, 1); + l2_norm_f32<1024><<>>(x, dst, ncols, stride_row, stride_channel, stride_sample, eps); + } +} + void ggml_cuda_op_norm(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { const ggml_tensor * src0 = dst->src[0]; const float * src0_d = (const float *) src0->data; @@ -340,3 +432,27 @@ void ggml_cuda_op_rms_norm_back(ggml_backend_cuda_context & ctx, ggml_tensor * d rms_norm_back_f32_cuda(grad_d, src0f_d, dst_d, ne00, nrows, eps, stream); } + +void ggml_cuda_op_l2_norm(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { + const ggml_tensor * src0 = dst->src[0]; + const float * src0_d = (const float *) src0->data; + float * dst_d = (float *) dst->data; + cudaStream_t stream = ctx.stream(); + + GGML_ASSERT(src0->type == GGML_TYPE_F32); + GGML_ASSERT( dst->type == GGML_TYPE_F32); + + GGML_TENSOR_UNARY_OP_LOCALS; + + float eps; + memcpy(&eps, dst->op_params, sizeof(float)); + GGML_ASSERT(eps >= 0.0f); + + const size_t ts0 = ggml_type_size(src0->type); + GGML_ASSERT(nb00 == ts0); + const int64_t s01 = nb01 / ts0; + const int64_t s02 = nb02 / ts0; + const int64_t s03 = nb03 / ts0; + + l2_norm_f32_cuda(src0_d, dst_d, ne00, ne01, ne02, ne03, s01, s02, s03, eps, stream); +} diff --git a/ggml/src/ggml-cuda/norm.cuh b/ggml/src/ggml-cuda/norm.cuh index d63d34380..706a5660a 100644 --- a/ggml/src/ggml-cuda/norm.cuh +++ b/ggml/src/ggml-cuda/norm.cuh @@ -7,3 +7,5 @@ void ggml_cuda_op_group_norm(ggml_backend_cuda_context & ctx, ggml_tensor * dst) void ggml_cuda_op_rms_norm(ggml_backend_cuda_context & ctx, ggml_tensor * dst); void ggml_cuda_op_rms_norm_back(ggml_backend_cuda_context & ctx, ggml_tensor * dst); + +void ggml_cuda_op_l2_norm(ggml_backend_cuda_context & ctx, ggml_tensor * dst); diff --git a/ggml/src/ggml-cuda/pad.cu b/ggml/src/ggml-cuda/pad.cu index aba539e8d..77432b046 100644 --- a/ggml/src/ggml-cuda/pad.cu +++ b/ggml/src/ggml-cuda/pad.cu @@ -14,7 +14,7 @@ static __global__ void pad_f32(const float * x, float * dst, const int ne0, cons nidx + blockIdx.y * ne0 + blockIdx.z * ne0 * gridDim.y; - if (nidx < ne00 && blockIdx.y < ne01 && blockIdx.z < ne02*ne03) { + if (nidx < ne00 && blockIdx.y < (unsigned)ne01 && blockIdx.z < (unsigned)(ne02*ne03)) { int offset_src = nidx + blockIdx.y * ne00 + diff --git a/ggml/src/ggml-cuda/ssm-conv.cu b/ggml/src/ggml-cuda/ssm-conv.cu new file mode 100644 index 000000000..f63757196 --- /dev/null +++ b/ggml/src/ggml-cuda/ssm-conv.cu @@ -0,0 +1,148 @@ +#include "ssm-conv.cuh" + +template +static __global__ void ssm_conv_f32(const float * __restrict__ src0, const float * __restrict__ src1, + const int src0_nb0, const int src0_nb1, const int src0_nb2, const int src1_nb1, + float * __restrict__ dst, const int dst_nb0, const int dst_nb1, const int dst_nb2, + const int64_t n_t) { + GGML_UNUSED(src0_nb0); + const int tid = threadIdx.x; + const int bidx = blockIdx.x; + const int bidy = blockIdx.y; + + const float * x_block = (const float *) ((const char *) src0 + bidx * src0_nb2 + bidy * split_d_inner * src0_nb1); + const float * w_block = (const float *) ((const char *) src1 + bidy * split_d_inner * src1_nb1); + float * y_block = (float *) ((char *) dst + bidx * dst_nb2 + bidy * split_d_inner * dst_nb0); + + const int stride_x = src0_nb1 / sizeof(float); + const int stride_w = src1_nb1 / sizeof(float); + const int stride_y = dst_nb1 / sizeof(float); + + float x[d_conv] = { 0.0f }; + float w[d_conv] = { 0.0f }; + +#pragma unroll + for (size_t j = 0; j < d_conv; j++) { + w[j] = w_block[tid * stride_w + j]; + } + + for (int64_t i = 0; i < n_t; i++) { + float sumf = 0.0f; + + if (i == 0) { + for (size_t j = 0; j < d_conv; j++) { + x[j] = x_block[tid * stride_x + j]; + } + } else { + x[(i - 1) % d_conv] = x_block[tid * stride_x + i + d_conv - 1]; + } + +#pragma unroll + for (size_t j = 0; j < d_conv; j++) { + sumf += x[(i + j) % d_conv] * w[j]; + } + y_block[i * stride_y + tid] = sumf; + } +} + +template +static __global__ void ssm_conv_long_token_f32(const float * __restrict__ src0, const float * __restrict__ src1, + const int src0_nb0, const int src0_nb1, const int src0_nb2, + const int src1_nb1, float * __restrict__ dst, const int dst_nb0, + const int dst_nb1, const int dst_nb2, const int64_t n_t) { + const int tid = threadIdx.x; + const int bidx = blockIdx.x; + const int bidy = blockIdx.y; + const int bidz = blockIdx.z; + + const float * x_block = (const float *) ((const char *) src0 + bidx * src0_nb2 + bidy * split_d_inner * src0_nb1 + + bidz * split_n_t * src0_nb0); + const float * w_block = (const float *) ((const char *) src1 + bidy * split_d_inner * src1_nb1); + float * y_block = + (float *) ((char *) dst + bidx * dst_nb2 + bidz * split_n_t * dst_nb1 + bidy * split_d_inner * dst_nb0); + + const int stride_x = src0_nb1 / sizeof(float); + const int stride_w = src1_nb1 / sizeof(float); + const int stride_y = dst_nb1 / sizeof(float); + + float x[d_conv] = { 0.0f }; + float w[d_conv] = { 0.0f }; + +#pragma unroll + for (size_t j = 0; j < d_conv; j++) { + w[j] = w_block[tid * stride_w + j]; + } + +#pragma unroll + for (int64_t i = 0; i < split_n_t; i++) { + if (bidz * split_n_t + i < n_t) { + float sumf = 0.0f; + + if (i == 0) { + for (size_t j = 0; j < d_conv; j++) { + x[j] = x_block[tid * stride_x + j]; + } + } else { + x[(i - 1) % d_conv] = x_block[tid * stride_x + i + d_conv - 1]; + } + +#pragma unroll + for (size_t j = 0; j < d_conv; j++) { + sumf += x[(i + j) % d_conv] * w[j]; + } + y_block[i * stride_y + tid] = sumf; + } + } +} + +static void ssm_conv_f32_cuda(const float * src0, const float * src1, const int src0_nb0, const int src0_nb1, + const int src0_nb2, const int src1_nb1, float * dst, const int dst_nb0, const int dst_nb1, + const int dst_nb2, const int64_t nc, const int64_t nr, const int64_t n_t, + const int64_t n_s, cudaStream_t stream) { + const int threads = 128; + GGML_ASSERT(nr % threads == 0); + + if (n_t <= 32) { + const dim3 blocks(n_s, (nr + threads - 1) / threads, 1); + if (nc == 4) { + ssm_conv_f32<<>>(src0, src1, src0_nb0, src0_nb1, src0_nb2, src1_nb1, + dst, dst_nb0, dst_nb1, dst_nb2, n_t); + } else { + GGML_ABORT("Only support kernel size = 4 now."); + } + } else { + if (nc == 4) { + const int64_t split_n_t = 32; + dim3 blocks(n_s, (nr + threads - 1) / threads, (n_t + split_n_t - 1) / split_n_t); + ssm_conv_long_token_f32<<>>( + src0, src1, src0_nb0, src0_nb1, src0_nb2, src1_nb1, dst, dst_nb0, dst_nb1, dst_nb2, n_t); + } else { + GGML_ABORT("Only support kernel size = 4 right now."); + } + } +} + +void ggml_cuda_op_ssm_conv(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { + const struct ggml_tensor * src0 = dst->src[0]; // conv_x + const struct ggml_tensor * src1 = dst->src[1]; // conv1d.weight + + const int64_t nc = src1->ne[0]; // d_conv + const int64_t nr = src0->ne[1]; // d_inner + const int64_t n_t = dst->ne[1]; // tokens per sequence + const int64_t n_s = dst->ne[2]; // number of sequences in the batch + + GGML_ASSERT(dst->ne[0] == nr); + GGML_ASSERT(src0->nb[0] == sizeof(float)); + GGML_ASSERT(src1->nb[0] == sizeof(float)); + GGML_ASSERT(src0->nb[1] == src0->ne[0] * sizeof(float)); + + const float * src0_d = (const float *) src0->data; + const float * src1_d = (const float *) src1->data; + float * dst_d = (float *) dst->data; + cudaStream_t stream = ctx.stream(); + + GGML_ASSERT(src0->type == GGML_TYPE_F32); + GGML_ASSERT(dst->type == GGML_TYPE_F32); + ssm_conv_f32_cuda(src0_d, src1_d, src0->nb[0], src0->nb[1], src0->nb[2], src1->nb[1], dst_d, dst->nb[0], dst->nb[1], + dst->nb[2], nc, nr, n_t, n_s, stream); +} diff --git a/ggml/src/ggml-cuda/ssm-conv.cuh b/ggml/src/ggml-cuda/ssm-conv.cuh new file mode 100644 index 000000000..8e6c1f00b --- /dev/null +++ b/ggml/src/ggml-cuda/ssm-conv.cuh @@ -0,0 +1,3 @@ +#include "common.cuh" + +void ggml_cuda_op_ssm_conv(ggml_backend_cuda_context & ctx, ggml_tensor * dst); diff --git a/ggml/src/ggml-cuda/ssm-scan.cu b/ggml/src/ggml-cuda/ssm-scan.cu new file mode 100644 index 000000000..37ee208c0 --- /dev/null +++ b/ggml/src/ggml-cuda/ssm-scan.cu @@ -0,0 +1,153 @@ +#include "ssm-scan.cuh" + +template +__global__ void __launch_bounds__(splitD, 2) + ssm_scan_f32(const float * __restrict__ src0, const float * __restrict__ src1, const float * __restrict__ src2, + const float * __restrict__ src3, const float * __restrict__ src4, const float * __restrict__ src5, + const int src0_nb1, const int src0_nb2, const int src1_nb0, const int src1_nb1, const int src1_nb2, + const int src1_nb3, const int src2_nb0, const int src2_nb1, const int src2_nb2, const int src3_nb1, + const int src4_nb1, const int src4_nb2, const int src5_nb1, const int src5_nb2, + float * __restrict__ dst, const int64_t L) { + GGML_UNUSED(src1_nb0); + GGML_UNUSED(src2_nb0); + const int bidx = blockIdx.x; // split along B + const int bidy = blockIdx.y; // split along D + const int tid = threadIdx.x; + const int wid = tid / 32; + const int wtid = tid % 32; + + extern __shared__ float smem[]; + const int stride_sA = N + 1; + const int stride_ss0 = N + 1; + float * smem_A = smem; + float * smem_s0 = smem_A + splitD * stride_sA; + + const float * s0_block = (const float *) ((const char *) src0 + bidx * src0_nb2 + bidy * splitD * src0_nb1); + const float * x_block = (const float *) ((const char *) src1 + (bidx * src1_nb2) + bidy * splitD * sizeof(float)); + const float * dt_block = (const float *) ((const char *) src2 + (bidx * src2_nb2) + bidy * splitD * sizeof(float)); + const float * A_block = (const float *) ((const char *) src3 + bidy * splitD * src3_nb1); + const float * B_block = (const float *) ((const char *) src4 + (bidx * src4_nb2)); + const float * C_block = (const float *) ((const char *) src5 + (bidx * src5_nb2)); + float * y_block = (float *) ((char *) dst + (bidx * src1_nb2) + bidy * splitD * sizeof(float)); + float * s_block = (float *) ((char *) dst + src1_nb3 + bidx * src0_nb2 + bidy * splitD * src0_nb1); + + const int stride_s0 = src0_nb1 / sizeof(float); + const int stride_x = src1_nb1 / sizeof(float); + const int stride_dt = src2_nb1 / sizeof(float); + const int stride_A = src3_nb1 / sizeof(float); + const int stride_B = src4_nb1 / sizeof(float); + const int stride_C = src5_nb1 / sizeof(float); + const int stride_s = stride_s0; + const int stride_y = stride_x; + + // can N not be 16? for example 32? + if (N == 16) { +#pragma unroll + for (size_t i = 0; i < splitD / 4; i += 2) { + float value = A_block[(wid * warpSize + i) * stride_A + wtid]; + // todo: bank conflict + // I am always confused with how to use the swizzling method to solve + // bank conflit. Hoping somebody can tell me. + smem_A[(wid * warpSize + i) * stride_sA + wtid + ((wtid / 16) > 0 ? 1 : 0)] = value; + } +#pragma unroll + for (size_t i = 0; i < splitD / 4; i += 2) { + float value = s0_block[(wid * warpSize + i) * stride_s0 + wtid]; + smem_s0[(wid * warpSize + i) * stride_ss0 + wtid + ((wtid / 16) > 0 ? 1 : 0)] = value; + } + } + + __syncthreads(); + + for (int64_t i = 0; i < L; i++) { + float dt_soft_plus = dt_block[i * stride_dt + tid]; + if (dt_soft_plus <= 20.0f) { + dt_soft_plus = log1pf(exp(dt_soft_plus)); + } + float x_dt = x_block[i * stride_x + tid] * dt_soft_plus; + float sumf = 0.0f; +#pragma unroll + for (size_t j = 0; j < N; j++) { + float state = (smem_s0[tid * stride_ss0 + j] * expf(dt_soft_plus * smem_A[tid * stride_sA + j])) + + (B_block[i * stride_B + j] * x_dt); + sumf += state * C_block[i * stride_C + j]; + if (i == L - 1) { + s_block[tid * stride_s + j] = state; + } else { + smem_s0[tid * stride_ss0 + j] = state; + } + } + __syncthreads(); + y_block[i * stride_y + tid] = sumf; + } +} + +static void ssm_scan_f32_cuda(const float * src0, const float * src1, const float * src2, const float * src3, + const float * src4, const float * src5, const int src0_nb1, const int src0_nb2, + const int src1_nb0, const int src1_nb1, const int src1_nb2, const int src1_nb3, + const int src2_nb0, const int src2_nb1, const int src2_nb2, const int src3_nb1, + const int src4_nb1, const int src4_nb2, const int src5_nb1, const int src5_nb2, + float * dst, const int64_t N, const int64_t D, const int64_t L, const int64_t B, + cudaStream_t stream) { + const int threads = 128; + // todo: consider D cannot be divided,does this situation exist? + GGML_ASSERT(D % threads == 0); + const dim3 blocks(B, (D + threads - 1) / threads, 1); + const int smem_size = (threads * (N + 1) * 2) * sizeof(float); + if (N == 16) { + ssm_scan_f32<128, 16><<>>( + src0, src1, src2, src3, src4, src5, src0_nb1, src0_nb2, src1_nb0, src1_nb1, src1_nb2, src1_nb3, src2_nb0, + src2_nb1, src2_nb2, src3_nb1, src4_nb1, src4_nb2, src5_nb1, src5_nb2, dst, L); + } else { + GGML_ABORT("doesn't support N!=16."); + } +} + +void ggml_cuda_op_ssm_scan(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { + const struct ggml_tensor * src0 = dst->src[0]; // s + const struct ggml_tensor * src1 = dst->src[1]; // x + const struct ggml_tensor * src2 = dst->src[2]; // dt + const struct ggml_tensor * src3 = dst->src[3]; // A + const struct ggml_tensor * src4 = dst->src[4]; // B + const struct ggml_tensor * src5 = dst->src[5]; // C + + // const int64_t d_state = src0->ne[0]; + // const int64_t d_inner = src0->ne[1]; + // const int64_t l = src1->ne[1]; + // const int64_t b = src0->ne[2]; + + const int64_t nc = src0->ne[0]; // d_state + const int64_t nr = src0->ne[1]; // d_inner + const int64_t n_t = src1->ne[1]; // number of tokens per sequence + const int64_t n_s = src0->ne[2]; // number of sequences in the batch + + GGML_ASSERT(ggml_nelements(src1) + ggml_nelements(src0) == ggml_nelements(dst)); + GGML_ASSERT(src0->nb[0] == sizeof(float)); + GGML_ASSERT(src1->nb[0] == sizeof(float)); + GGML_ASSERT(src2->nb[0] == sizeof(float)); + GGML_ASSERT(src3->nb[0] == sizeof(float)); + GGML_ASSERT(src4->nb[0] == sizeof(float)); + GGML_ASSERT(src5->nb[0] == sizeof(float)); + // required for the dot product between s and C + GGML_ASSERT(src0->nb[1] == src0->ne[0] * sizeof(float)); + // required for per-sequence offsets for states + GGML_ASSERT(src0->nb[2] == src0->ne[0] * src0->ne[1] * sizeof(float)); + // required to get correct offset for state destination (i.e. src1->nb[3]) + GGML_ASSERT(src1->nb[3] == src1->ne[0] * src1->ne[1] * src1->ne[2] * sizeof(float)); + + const float * src0_d = (const float *) src0->data; + const float * src1_d = (const float *) src1->data; + const float * src2_d = (const float *) src2->data; + const float * src3_d = (const float *) src3->data; + const float * src4_d = (const float *) src4->data; + const float * src5_d = (const float *) src5->data; + float * dst_d = (float *) dst->data; + cudaStream_t stream = ctx.stream(); + + GGML_ASSERT(src0->type == GGML_TYPE_F32); + GGML_ASSERT(dst->type == GGML_TYPE_F32); + + ssm_scan_f32_cuda(src0_d, src1_d, src2_d, src3_d, src4_d, src5_d, src0->nb[1], src0->nb[2], src1->nb[0], + src1->nb[1], src1->nb[2], src1->nb[3], src2->nb[0], src2->nb[1], src2->nb[2], src3->nb[1], + src4->nb[1], src4->nb[2], src5->nb[1], src5->nb[2], dst_d, nc, nr, n_t, n_s, stream); +} diff --git a/ggml/src/ggml-cuda/ssm-scan.cuh b/ggml/src/ggml-cuda/ssm-scan.cuh new file mode 100644 index 000000000..ee078f5eb --- /dev/null +++ b/ggml/src/ggml-cuda/ssm-scan.cuh @@ -0,0 +1,3 @@ +#include "common.cuh" + +void ggml_cuda_op_ssm_scan(ggml_backend_cuda_context & ctx, ggml_tensor * dst); diff --git a/ggml/src/ggml-cuda/sum.cu b/ggml/src/ggml-cuda/sum.cu index e0dafc1d2..f9589080a 100644 --- a/ggml/src/ggml-cuda/sum.cu +++ b/ggml/src/ggml-cuda/sum.cu @@ -1,6 +1,6 @@ -#if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA) && CUDART_VERSION >= 11700 +#if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA) && CUDART_VERSION >= 11070 #define USE_CUB -#endif // !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA) && CUDART_VERSION >= 11700 +#endif // !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA) && CUDART_VERSION >= 11070 #ifdef USE_CUB #include diff --git a/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-cpb16.cu b/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-cpb16.cu deleted file mode 100644 index f09bdeff7..000000000 --- a/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-cpb16.cu +++ /dev/null @@ -1,10 +0,0 @@ -// This file has been autogenerated by generate_cu_files.py, do not edit manually. - -#include "../fattn-mma-f16.cuh" - -DECL_FATTN_MMA_F16_CASE(64, 16); -DECL_FATTN_MMA_F16_CASE(80, 16); -DECL_FATTN_MMA_F16_CASE(96, 16); -DECL_FATTN_MMA_F16_CASE(112, 16); -DECL_FATTN_MMA_F16_CASE(128, 16); -DECL_FATTN_MMA_F16_CASE(256, 16); diff --git a/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-cpb32.cu b/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-cpb32.cu deleted file mode 100644 index 221108873..000000000 --- a/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-cpb32.cu +++ /dev/null @@ -1,10 +0,0 @@ -// This file has been autogenerated by generate_cu_files.py, do not edit manually. - -#include "../fattn-mma-f16.cuh" - -DECL_FATTN_MMA_F16_CASE(64, 32); -DECL_FATTN_MMA_F16_CASE(80, 32); -DECL_FATTN_MMA_F16_CASE(96, 32); -DECL_FATTN_MMA_F16_CASE(112, 32); -DECL_FATTN_MMA_F16_CASE(128, 32); -DECL_FATTN_MMA_F16_CASE(256, 32); diff --git a/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-cpb64.cu b/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-cpb64.cu deleted file mode 100644 index d24b08575..000000000 --- a/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-cpb64.cu +++ /dev/null @@ -1,10 +0,0 @@ -// This file has been autogenerated by generate_cu_files.py, do not edit manually. - -#include "../fattn-mma-f16.cuh" - -DECL_FATTN_MMA_F16_CASE(64, 64); -DECL_FATTN_MMA_F16_CASE(80, 64); -DECL_FATTN_MMA_F16_CASE(96, 64); -DECL_FATTN_MMA_F16_CASE(112, 64); -DECL_FATTN_MMA_F16_CASE(128, 64); -DECL_FATTN_MMA_F16_CASE(256, 64); diff --git a/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-cpb8.cu b/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-cpb8.cu deleted file mode 100644 index bdf86c0ea..000000000 --- a/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-cpb8.cu +++ /dev/null @@ -1,10 +0,0 @@ -// This file has been autogenerated by generate_cu_files.py, do not edit manually. - -#include "../fattn-mma-f16.cuh" - -DECL_FATTN_MMA_F16_CASE(64, 8); -DECL_FATTN_MMA_F16_CASE(80, 8); -DECL_FATTN_MMA_F16_CASE(96, 8); -DECL_FATTN_MMA_F16_CASE(112, 8); -DECL_FATTN_MMA_F16_CASE(128, 8); -DECL_FATTN_MMA_F16_CASE(256, 8); diff --git a/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_8.cu b/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_8.cu new file mode 100644 index 000000000..80108615a --- /dev/null +++ b/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_8.cu @@ -0,0 +1,10 @@ +// This file has been autogenerated by generate_cu_files.py, do not edit manually. + +#include "../fattn-mma-f16.cuh" + +DECL_FATTN_MMA_F16_CASE(64, 1, 8); +DECL_FATTN_MMA_F16_CASE(80, 1, 8); +DECL_FATTN_MMA_F16_CASE(96, 1, 8); +DECL_FATTN_MMA_F16_CASE(112, 1, 8); +DECL_FATTN_MMA_F16_CASE(128, 1, 8); +DECL_FATTN_MMA_F16_CASE(256, 1, 8); diff --git a/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_1.cu b/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_1.cu new file mode 100644 index 000000000..66161c0ab --- /dev/null +++ b/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_1.cu @@ -0,0 +1,10 @@ +// This file has been autogenerated by generate_cu_files.py, do not edit manually. + +#include "../fattn-mma-f16.cuh" + +DECL_FATTN_MMA_F16_CASE(64, 16, 1); +DECL_FATTN_MMA_F16_CASE(80, 16, 1); +DECL_FATTN_MMA_F16_CASE(96, 16, 1); +DECL_FATTN_MMA_F16_CASE(112, 16, 1); +DECL_FATTN_MMA_F16_CASE(128, 16, 1); +DECL_FATTN_MMA_F16_CASE(256, 16, 1); diff --git a/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_2.cu b/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_2.cu new file mode 100644 index 000000000..ee88c72aa --- /dev/null +++ b/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_2.cu @@ -0,0 +1,10 @@ +// This file has been autogenerated by generate_cu_files.py, do not edit manually. + +#include "../fattn-mma-f16.cuh" + +DECL_FATTN_MMA_F16_CASE(64, 16, 2); +DECL_FATTN_MMA_F16_CASE(80, 16, 2); +DECL_FATTN_MMA_F16_CASE(96, 16, 2); +DECL_FATTN_MMA_F16_CASE(112, 16, 2); +DECL_FATTN_MMA_F16_CASE(128, 16, 2); +DECL_FATTN_MMA_F16_CASE(256, 16, 2); diff --git a/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_4.cu b/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_4.cu new file mode 100644 index 000000000..d888a5a42 --- /dev/null +++ b/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_4.cu @@ -0,0 +1,10 @@ +// This file has been autogenerated by generate_cu_files.py, do not edit manually. + +#include "../fattn-mma-f16.cuh" + +DECL_FATTN_MMA_F16_CASE(64, 16, 4); +DECL_FATTN_MMA_F16_CASE(80, 16, 4); +DECL_FATTN_MMA_F16_CASE(96, 16, 4); +DECL_FATTN_MMA_F16_CASE(112, 16, 4); +DECL_FATTN_MMA_F16_CASE(128, 16, 4); +DECL_FATTN_MMA_F16_CASE(256, 16, 4); diff --git a/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_4.cu b/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_4.cu new file mode 100644 index 000000000..d93a2d08e --- /dev/null +++ b/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_4.cu @@ -0,0 +1,10 @@ +// This file has been autogenerated by generate_cu_files.py, do not edit manually. + +#include "../fattn-mma-f16.cuh" + +DECL_FATTN_MMA_F16_CASE(64, 2, 4); +DECL_FATTN_MMA_F16_CASE(80, 2, 4); +DECL_FATTN_MMA_F16_CASE(96, 2, 4); +DECL_FATTN_MMA_F16_CASE(112, 2, 4); +DECL_FATTN_MMA_F16_CASE(128, 2, 4); +DECL_FATTN_MMA_F16_CASE(256, 2, 4); diff --git a/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_8.cu b/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_8.cu new file mode 100644 index 000000000..617464c94 --- /dev/null +++ b/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_8.cu @@ -0,0 +1,10 @@ +// This file has been autogenerated by generate_cu_files.py, do not edit manually. + +#include "../fattn-mma-f16.cuh" + +DECL_FATTN_MMA_F16_CASE(64, 2, 8); +DECL_FATTN_MMA_F16_CASE(80, 2, 8); +DECL_FATTN_MMA_F16_CASE(96, 2, 8); +DECL_FATTN_MMA_F16_CASE(112, 2, 8); +DECL_FATTN_MMA_F16_CASE(128, 2, 8); +DECL_FATTN_MMA_F16_CASE(256, 2, 8); diff --git a/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_1.cu b/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_1.cu new file mode 100644 index 000000000..970d2b686 --- /dev/null +++ b/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_1.cu @@ -0,0 +1,10 @@ +// This file has been autogenerated by generate_cu_files.py, do not edit manually. + +#include "../fattn-mma-f16.cuh" + +DECL_FATTN_MMA_F16_CASE(64, 32, 1); +DECL_FATTN_MMA_F16_CASE(80, 32, 1); +DECL_FATTN_MMA_F16_CASE(96, 32, 1); +DECL_FATTN_MMA_F16_CASE(112, 32, 1); +DECL_FATTN_MMA_F16_CASE(128, 32, 1); +DECL_FATTN_MMA_F16_CASE(256, 32, 1); diff --git a/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_2.cu b/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_2.cu new file mode 100644 index 000000000..65cd377c3 --- /dev/null +++ b/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_2.cu @@ -0,0 +1,10 @@ +// This file has been autogenerated by generate_cu_files.py, do not edit manually. + +#include "../fattn-mma-f16.cuh" + +DECL_FATTN_MMA_F16_CASE(64, 32, 2); +DECL_FATTN_MMA_F16_CASE(80, 32, 2); +DECL_FATTN_MMA_F16_CASE(96, 32, 2); +DECL_FATTN_MMA_F16_CASE(112, 32, 2); +DECL_FATTN_MMA_F16_CASE(128, 32, 2); +DECL_FATTN_MMA_F16_CASE(256, 32, 2); diff --git a/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_2.cu b/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_2.cu new file mode 100644 index 000000000..f4a8bf348 --- /dev/null +++ b/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_2.cu @@ -0,0 +1,10 @@ +// This file has been autogenerated by generate_cu_files.py, do not edit manually. + +#include "../fattn-mma-f16.cuh" + +DECL_FATTN_MMA_F16_CASE(64, 4, 2); +DECL_FATTN_MMA_F16_CASE(80, 4, 2); +DECL_FATTN_MMA_F16_CASE(96, 4, 2); +DECL_FATTN_MMA_F16_CASE(112, 4, 2); +DECL_FATTN_MMA_F16_CASE(128, 4, 2); +DECL_FATTN_MMA_F16_CASE(256, 4, 2); diff --git a/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_4.cu b/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_4.cu new file mode 100644 index 000000000..de191a8ab --- /dev/null +++ b/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_4.cu @@ -0,0 +1,10 @@ +// This file has been autogenerated by generate_cu_files.py, do not edit manually. + +#include "../fattn-mma-f16.cuh" + +DECL_FATTN_MMA_F16_CASE(64, 4, 4); +DECL_FATTN_MMA_F16_CASE(80, 4, 4); +DECL_FATTN_MMA_F16_CASE(96, 4, 4); +DECL_FATTN_MMA_F16_CASE(112, 4, 4); +DECL_FATTN_MMA_F16_CASE(128, 4, 4); +DECL_FATTN_MMA_F16_CASE(256, 4, 4); diff --git a/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_8.cu b/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_8.cu new file mode 100644 index 000000000..e8cb0e1b3 --- /dev/null +++ b/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_8.cu @@ -0,0 +1,10 @@ +// This file has been autogenerated by generate_cu_files.py, do not edit manually. + +#include "../fattn-mma-f16.cuh" + +DECL_FATTN_MMA_F16_CASE(64, 4, 8); +DECL_FATTN_MMA_F16_CASE(80, 4, 8); +DECL_FATTN_MMA_F16_CASE(96, 4, 8); +DECL_FATTN_MMA_F16_CASE(112, 4, 8); +DECL_FATTN_MMA_F16_CASE(128, 4, 8); +DECL_FATTN_MMA_F16_CASE(256, 4, 8); diff --git a/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_64-ncols2_1.cu b/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_64-ncols2_1.cu new file mode 100644 index 000000000..a532e9629 --- /dev/null +++ b/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_64-ncols2_1.cu @@ -0,0 +1,10 @@ +// This file has been autogenerated by generate_cu_files.py, do not edit manually. + +#include "../fattn-mma-f16.cuh" + +DECL_FATTN_MMA_F16_CASE(64, 64, 1); +DECL_FATTN_MMA_F16_CASE(80, 64, 1); +DECL_FATTN_MMA_F16_CASE(96, 64, 1); +DECL_FATTN_MMA_F16_CASE(112, 64, 1); +DECL_FATTN_MMA_F16_CASE(128, 64, 1); +DECL_FATTN_MMA_F16_CASE(256, 64, 1); diff --git a/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_1.cu b/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_1.cu new file mode 100644 index 000000000..bf25181aa --- /dev/null +++ b/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_1.cu @@ -0,0 +1,10 @@ +// This file has been autogenerated by generate_cu_files.py, do not edit manually. + +#include "../fattn-mma-f16.cuh" + +DECL_FATTN_MMA_F16_CASE(64, 8, 1); +DECL_FATTN_MMA_F16_CASE(80, 8, 1); +DECL_FATTN_MMA_F16_CASE(96, 8, 1); +DECL_FATTN_MMA_F16_CASE(112, 8, 1); +DECL_FATTN_MMA_F16_CASE(128, 8, 1); +DECL_FATTN_MMA_F16_CASE(256, 8, 1); diff --git a/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_2.cu b/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_2.cu new file mode 100644 index 000000000..378c132e6 --- /dev/null +++ b/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_2.cu @@ -0,0 +1,10 @@ +// This file has been autogenerated by generate_cu_files.py, do not edit manually. + +#include "../fattn-mma-f16.cuh" + +DECL_FATTN_MMA_F16_CASE(64, 8, 2); +DECL_FATTN_MMA_F16_CASE(80, 8, 2); +DECL_FATTN_MMA_F16_CASE(96, 8, 2); +DECL_FATTN_MMA_F16_CASE(112, 8, 2); +DECL_FATTN_MMA_F16_CASE(128, 8, 2); +DECL_FATTN_MMA_F16_CASE(256, 8, 2); diff --git a/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_4.cu b/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_4.cu new file mode 100644 index 000000000..372641be9 --- /dev/null +++ b/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_4.cu @@ -0,0 +1,10 @@ +// This file has been autogenerated by generate_cu_files.py, do not edit manually. + +#include "../fattn-mma-f16.cuh" + +DECL_FATTN_MMA_F16_CASE(64, 8, 4); +DECL_FATTN_MMA_F16_CASE(80, 8, 4); +DECL_FATTN_MMA_F16_CASE(96, 8, 4); +DECL_FATTN_MMA_F16_CASE(112, 8, 4); +DECL_FATTN_MMA_F16_CASE(128, 8, 4); +DECL_FATTN_MMA_F16_CASE(256, 8, 4); diff --git a/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_8.cu b/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_8.cu new file mode 100644 index 000000000..9ff5968b6 --- /dev/null +++ b/ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_8.cu @@ -0,0 +1,10 @@ +// This file has been autogenerated by generate_cu_files.py, do not edit manually. + +#include "../fattn-mma-f16.cuh" + +DECL_FATTN_MMA_F16_CASE(64, 8, 8); +DECL_FATTN_MMA_F16_CASE(80, 8, 8); +DECL_FATTN_MMA_F16_CASE(96, 8, 8); +DECL_FATTN_MMA_F16_CASE(112, 8, 8); +DECL_FATTN_MMA_F16_CASE(128, 8, 8); +DECL_FATTN_MMA_F16_CASE(256, 8, 8); diff --git a/ggml/src/ggml-cuda/template-instances/generate_cu_files.py b/ggml/src/ggml-cuda/template-instances/generate_cu_files.py index a2628f16e..dd373a09d 100755 --- a/ggml/src/ggml-cuda/template-instances/generate_cu_files.py +++ b/ggml/src/ggml-cuda/template-instances/generate_cu_files.py @@ -18,7 +18,7 @@ SOURCE_FATTN_MMA_START = """// This file has been autogenerated by generate_cu_f """ -SOURCE_FATTN_MMA_CASE = "DECL_FATTN_MMA_F16_CASE({head_size}, {cols_per_block});\n" +SOURCE_FATTN_MMA_CASE = "DECL_FATTN_MMA_F16_CASE({head_size}, {ncols1}, {ncols2});\n" TYPES_MMQ = [ "GGML_TYPE_Q4_0", "GGML_TYPE_Q4_1", "GGML_TYPE_Q5_0", "GGML_TYPE_Q5_1", "GGML_TYPE_Q8_0", @@ -57,12 +57,18 @@ for vkq_size in [16, 32]: with open(f"fattn-vec-f{vkq_size}-instance-hs{head_size}-{get_short_name(type_k)}-{get_short_name(type_v)}.cu", "w") as f: f.write(SOURCE_FATTN_VEC.format(vkq_size=vkq_size, head_size=head_size, type_k=type_k, type_v=type_v)) -for cols_per_block in [8, 16, 32, 64]: - with open(f"fattn-mma-f16-instance-cpb{cols_per_block}.cu", "w") as f: - f.write(SOURCE_FATTN_MMA_START) +for ncols in [8, 16, 32, 64, 128]: + for ncols2 in [1, 2, 4, 8]: + ncols1 = ncols // ncols2 + if ncols == 128: + continue # Too much register pressure. + with open(f"fattn-mma-f16-instance-ncols1_{ncols1}-ncols2_{ncols2}.cu", "w") as f: + f.write(SOURCE_FATTN_MMA_START) - for head_size in [64, 80, 96, 112, 128, 256]: - f.write(SOURCE_FATTN_MMA_CASE.format(cols_per_block=cols_per_block, head_size=head_size)) + for head_size in [64, 80, 96, 112, 128, 256]: + if ncols == 128 and head_size == 256: + continue # Needs too much shared memory. + f.write(SOURCE_FATTN_MMA_CASE.format(ncols1=ncols1, ncols2=ncols2, head_size=head_size)) for type in TYPES_MMQ: with open(f"mmq-instance-{get_short_name(type)}.cu", "w") as f: diff --git a/ggml/src/ggml-cuda/unary.cu b/ggml/src/ggml-cuda/unary.cu index 6b21f407d..ec5773e01 100644 --- a/ggml/src/ggml-cuda/unary.cu +++ b/ggml/src/ggml-cuda/unary.cu @@ -1,305 +1,213 @@ #include "unary.cuh" -static __global__ void neg_f32(const float * x, float * dst, const int k) { - const int i = blockDim.x*blockIdx.x + threadIdx.x; - - if (i >= k) { - return; - } - - dst[i] = -x[i]; +static __device__ __forceinline__ float op_abs(float x) { + return fabsf(x); } -static __global__ void step_f32(const float * x, float * dst, const int k) { - const int i = blockDim.x*blockIdx.x + threadIdx.x; - - if (i >= k) { - return; - } - - dst[i] = x[i] > 0.0f; +static __device__ __forceinline__ float op_sgn(float x) { + return (x > 0.f ? 1.f : ((x < 0.f ? -1.f : 0.f))); } -static __global__ void gelu_f32(const float * x, float * dst, const int k) { +static __device__ __forceinline__ float op_neg(float x) { + return -x; +} + +static __device__ __forceinline__ float op_step(float x) { + return x > 0.0f; +} + +static __device__ __forceinline__ float op_gelu(float x) { const float GELU_COEF_A = 0.044715f; const float SQRT_2_OVER_PI = 0.79788456080286535587989211986876f; - const int i = blockDim.x*blockIdx.x + threadIdx.x; - if (i >= k) { - return; - } - - float xi = x[i]; - dst[i] = 0.5f*xi*(1.0f + tanhf(SQRT_2_OVER_PI*xi*(1.0f + GELU_COEF_A*xi*xi))); + return 0.5f*x*(1.0f + tanhf(SQRT_2_OVER_PI*x*(1.0f + GELU_COEF_A*x*x))); } -static __global__ void gelu_quick_f32(const float * x, float * dst, int k) { +static __device__ __forceinline__ float op_gelu_quick(float x) { const float GELU_QUICK_COEF = -1.702f; - const int i = blockDim.x*blockIdx.x + threadIdx.x; - if (i >= k) { - return; - } - dst[i] = x[i] * (1.0f / (1.0f + expf(GELU_QUICK_COEF * x[i]))); + + return x * (1.0f / (1.0f + expf(GELU_QUICK_COEF * x))); } -static __global__ void silu_f32(const float * x, float * dst, const int k) { - const int i = blockDim.x*blockIdx.x + threadIdx.x; - - if (i >= k) { - return; - } - dst[i] = x[i] / (1.0f + expf(-x[i])); +static __device__ __forceinline__ float op_silu(float x) { + return x / (1.0f + expf(-x)); } -static __global__ void silu_back_f32( - const float * grad, const float * xf, float * dst, const int k) { +static __device__ __forceinline__ float op_tanh(float x) { + return tanhf(x); +} + +static __device__ __forceinline__ float op_relu(float x) { + return fmaxf(x, 0); +} + +static __device__ __forceinline__ float op_sigmoid(float x) { + return 1.0f / (1.0f + expf(-x)); +} + +static __device__ __forceinline__ float op_hardsigmoid(float x) { + return fminf(1.0f, fmaxf(0.0f, (x + 3.0f) / 6.0f)); +} + +static __device__ __forceinline__ float op_hardswish(float x) { + return x * fminf(1.0f, fmaxf(0.0f, (x + 3.0f) / 6.0f)); +} + +static __device__ __forceinline__ float op_exp(float x) { + return expf(x); +} + +static __device__ __forceinline__ float op_sqr(float x) { + return x * x; +} + +static __device__ __forceinline__ float op_sqrt(float x) { + return sqrtf(x); +} + +static __device__ __forceinline__ float op_sin(float x) { + return sinf(x); +} + +static __device__ __forceinline__ float op_cos(float x) { + return cosf(x); +} + +static __device__ __forceinline__ float op_log(float x) { + return logf(x); +} + +template +static __global__ void unary_op_kernel(const T * x, T * dst, const int k) { const int i = blockDim.x*blockIdx.x + threadIdx.x; if (i >= k) { return; } - const float xfi = xf[i]; - const float s = 1.0f / (1.0f + expf(-xfi)); - dst[i] = grad[i] * s * (1.0f + xfi * (1.0f - s)); + dst[i] = (T)op((float)x[i]); } -static __global__ void tanh_f32(const float * x, float * dst, int k) { - const int i = blockDim.x*blockIdx.x + threadIdx.x; - if (i >= k) { - return; - } - dst[i] = tanhf(x[i]); -} - -static __global__ void relu_f32(const float * x, float * dst, const int k) { - const int i = blockDim.x*blockIdx.x + threadIdx.x; - - if (i >= k) { - return; - } - dst[i] = fmaxf(x[i], 0); -} - -static __global__ void sigmoid_f32(const float * x, float * dst, const int k) { - const int i = blockDim.x*blockIdx.x + threadIdx.x; - - if (i >= k) { - return; - } - dst[i] = 1.0f / (1.0f + expf(-x[i])); -} - -static __global__ void hardsigmoid_f32(const float * x, float * dst, const int k) { - const int i = blockDim.x*blockIdx.x + threadIdx.x; - - if (i >= k) { - return; - } - dst[i] = fminf(1.0f, fmaxf(0.0f, (x[i] + 3.0f) / 6.0f)); -} - -static __global__ void hardswish_f32(const float * x, float * dst, const int k) { - const int i = blockDim.x*blockIdx.x + threadIdx.x; - - if (i >= k) { - return; - } - dst[i] = x[i] * fminf(1.0f, fmaxf(0.0f, (x[i] + 3.0f) / 6.0f)); -} - -static __global__ void exp_f32(const float * x, float * dst, const int k) { - const int i = blockDim.x*blockIdx.x + threadIdx.x; - - if (i >= k) { - return; - } - dst[i] = expf(x[i]); -} - -static __global__ void leaky_relu_f32(const float * x, float * dst, const int k, const float negative_slope) { - const int i = blockDim.x*blockIdx.x + threadIdx.x; - if (i >= k) { - return; - } - dst[i] = fmaxf(x[i], 0) + fminf(x[i], 0.0f) * negative_slope; -} - -static __global__ void sqr_f32(const float * x, float * dst, const int k) { - const int i = blockDim.x*blockIdx.x + threadIdx.x; - - if (i >= k) { - return; - } - dst[i] = x[i] * x[i]; -} - -static __global__ void sqrt_f32(const float * x, float * dst, const int k) { - const int i = blockDim.x*blockIdx.x + threadIdx.x; - - if (i >= k) { - return; - } - dst[i] = sqrtf(x[i]); -} - -static __global__ void sin_f32(const float * x, float * dst, const int k) { - const int i = blockDim.x*blockIdx.x + threadIdx.x; - - if (i >= k) { - return; - } - dst[i] = sinf(x[i]); -} - -static __global__ void cos_f32(const float * x, float * dst, const int k) { - const int i = blockDim.x*blockIdx.x + threadIdx.x; - - if (i >= k) { - return; - } - dst[i] = cosf(x[i]); -} - -static void neg_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) { +template +static void unary_cuda(const T * x, T * dst, const int k, cudaStream_t stream) { const int num_blocks = (k + CUDA_NEG_BLOCK_SIZE - 1) / CUDA_NEG_BLOCK_SIZE; - neg_f32<<>>(x, dst, k); + unary_op_kernel<<>>(x, dst, k); } -static void step_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) { - const int num_blocks = (k + CUDA_STEP_BLOCK_SIZE - 1) / CUDA_STEP_BLOCK_SIZE; - step_f32<<>>(x, dst, k); +template +void ggml_cuda_op_unary(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { + const ggml_tensor * src0 = dst->src[0]; + const void * src0_d = src0->data; + void * dst_d = dst->data; + cudaStream_t stream = ctx.stream(); + + GGML_ASSERT(ggml_is_contiguous(src0)); + + GGML_ASSERT(src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16); + GGML_ASSERT( dst->type == GGML_TYPE_F32 || dst->type == GGML_TYPE_F16); + GGML_ASSERT(src0->type == dst->type); + + if (src0->type == GGML_TYPE_F16) { + unary_cuda((const half *)src0_d, (half *)dst_d, ggml_nelements(src0), stream); + } else { + unary_cuda((const float *)src0_d, (float *)dst_d, ggml_nelements(src0), stream); + } } -static void gelu_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) { - const int num_blocks = (k + CUDA_GELU_BLOCK_SIZE - 1) / CUDA_GELU_BLOCK_SIZE; - gelu_f32<<>>(x, dst, k); +void ggml_cuda_op_abs(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { + ggml_cuda_op_unary(ctx, dst); } -static void gelu_quick_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) { - const int num_blocks = (k + CUDA_GELU_BLOCK_SIZE - 1) / CUDA_GELU_BLOCK_SIZE; - gelu_quick_f32<<>>(x, dst, k); -} - -static void silu_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) { - const int num_blocks = (k + CUDA_SILU_BLOCK_SIZE - 1) / CUDA_SILU_BLOCK_SIZE; - silu_f32<<>>(x, dst, k); -} - -static void silu_back_f32_cuda(const float * grad, const float * x, float * dst, const int k, cudaStream_t stream) { - const int num_blocks = (k + CUDA_SILU_BACK_BLOCK_SIZE - 1) / CUDA_SILU_BLOCK_SIZE; - silu_back_f32<<>>(grad, x, dst, k); -} - -static void tanh_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) { - const int num_blocks = (k + CUDA_TANH_BLOCK_SIZE - 1) / CUDA_TANH_BLOCK_SIZE; - tanh_f32<<>>(x, dst, k); -} - -static void relu_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) { - const int num_blocks = (k + CUDA_RELU_BLOCK_SIZE - 1) / CUDA_RELU_BLOCK_SIZE; - relu_f32<<>>(x, dst, k); -} - -static void sigmoid_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) { - const int num_blocks = (k + CUDA_SIGMOID_BLOCK_SIZE - 1) / CUDA_SIGMOID_BLOCK_SIZE; - sigmoid_f32<<>>(x, dst, k); -} - -static void hardsigmoid_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) { - const int num_blocks = (k + CUDA_HARDSIGMOID_BLOCK_SIZE - 1) / CUDA_HARDSIGMOID_BLOCK_SIZE; - hardsigmoid_f32<<>>(x, dst, k); -} - -static void hardswish_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) { - const int num_blocks = (k + CUDA_HARDSWISH_BLOCK_SIZE - 1) / CUDA_HARDSWISH_BLOCK_SIZE; - hardswish_f32<<>>(x, dst, k); -} - -static void exp_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) { - const int num_blocks = (k + CUDA_EXP_BLOCK_SIZE - 1) / CUDA_EXP_BLOCK_SIZE; - exp_f32<<>>(x, dst, k); -} - -static void leaky_relu_f32_cuda(const float * x, float * dst, const int k, const float negative_slope, cudaStream_t stream) { - const int num_blocks = (k + CUDA_RELU_BLOCK_SIZE - 1) / CUDA_RELU_BLOCK_SIZE; - leaky_relu_f32<<>>(x, dst, k, negative_slope); -} - -static void sqr_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) { - const int num_blocks = (k + CUDA_SQR_BLOCK_SIZE - 1) / CUDA_SQR_BLOCK_SIZE; - sqr_f32<<>>(x, dst, k); -} - -static void sqrt_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) { - const int num_blocks = (k + CUDA_SQRT_BLOCK_SIZE - 1) / CUDA_SQRT_BLOCK_SIZE; - sqrt_f32<<>>(x, dst, k); -} - -static void sin_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) { - const int num_blocks = (k + CUDA_SIN_BLOCK_SIZE - 1) / CUDA_SIN_BLOCK_SIZE; - sin_f32<<>>(x, dst, k); -} - -static void cos_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) { - const int num_blocks = (k + CUDA_COS_BLOCK_SIZE - 1) / CUDA_COS_BLOCK_SIZE; - cos_f32<<>>(x, dst, k); +void ggml_cuda_op_sgn(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { + ggml_cuda_op_unary(ctx, dst); } void ggml_cuda_op_neg(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { - const ggml_tensor * src0 = dst->src[0]; - const float * src0_d = (const float *)src0->data; - float * dst_d = (float *)dst->data; - cudaStream_t stream = ctx.stream(); - - GGML_ASSERT(ggml_is_contiguous(src0)); - - GGML_ASSERT(src0->type == GGML_TYPE_F32); - GGML_ASSERT( dst->type == GGML_TYPE_F32); - - neg_f32_cuda(src0_d, dst_d, ggml_nelements(src0), stream); + ggml_cuda_op_unary(ctx, dst); } void ggml_cuda_op_step(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { - const ggml_tensor * src0 = dst->src[0]; - const float * src0_d = (const float *)src0->data; - float * dst_d = (float *)dst->data; - cudaStream_t stream = ctx.stream(); - - GGML_ASSERT(ggml_is_contiguous(src0)); - - GGML_ASSERT(src0->type == GGML_TYPE_F32); - GGML_ASSERT( dst->type == GGML_TYPE_F32); - - step_f32_cuda(src0_d, dst_d, ggml_nelements(src0), stream); + ggml_cuda_op_unary(ctx, dst); } void ggml_cuda_op_gelu(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { - const ggml_tensor * src0 = dst->src[0]; - const float * src0_d = (const float *)src0->data; - float * dst_d = (float *)dst->data; - cudaStream_t stream = ctx.stream(); + ggml_cuda_op_unary(ctx, dst); +} - GGML_ASSERT(ggml_is_contiguous(src0)); - - GGML_ASSERT(src0->type == GGML_TYPE_F32); - GGML_ASSERT( dst->type == GGML_TYPE_F32); - - gelu_f32_cuda(src0_d, dst_d, ggml_nelements(src0), stream); +void ggml_cuda_op_gelu_quick(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { + ggml_cuda_op_unary(ctx, dst); } void ggml_cuda_op_silu(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { - const ggml_tensor * src0 = dst->src[0]; - const float * src0_d = (const float *)src0->data; - float * dst_d = (float *)dst->data; - cudaStream_t stream = ctx.stream(); + ggml_cuda_op_unary(ctx, dst); +} - GGML_ASSERT(ggml_is_contiguous(src0)); +void ggml_cuda_op_tanh(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { + ggml_cuda_op_unary(ctx, dst); +} - GGML_ASSERT(src0->type == GGML_TYPE_F32); - GGML_ASSERT( dst->type == GGML_TYPE_F32); +void ggml_cuda_op_relu(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { + ggml_cuda_op_unary(ctx, dst); +} - silu_f32_cuda(src0_d, dst_d, ggml_nelements(src0), stream); +void ggml_cuda_op_sigmoid(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { + ggml_cuda_op_unary(ctx, dst); +} + +void ggml_cuda_op_hardsigmoid(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { + ggml_cuda_op_unary(ctx, dst); +} + +void ggml_cuda_op_hardswish(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { + ggml_cuda_op_unary(ctx, dst); +} + +void ggml_cuda_op_exp(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { + ggml_cuda_op_unary(ctx, dst); +} + +void ggml_cuda_op_sqr(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { + ggml_cuda_op_unary(ctx, dst); +} + +void ggml_cuda_op_sqrt(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { + ggml_cuda_op_unary(ctx, dst); +} + +void ggml_cuda_op_sin(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { + ggml_cuda_op_unary(ctx, dst); +} + +void ggml_cuda_op_cos(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { + ggml_cuda_op_unary(ctx, dst); +} + +void ggml_cuda_op_log(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { + ggml_cuda_op_unary(ctx, dst); +} + +/* silu_back */ + +static __device__ __forceinline__ float op_silu_back(float grad, float x) { + const float s = 1.0f / (1.0f + expf(-x)); + return grad * s * (1.0f + x * (1.0f - s)); +} + +template +static __global__ void silu_back_kernel(const T * grad, const T * xf, T * dst, const int k) { + const int i = blockDim.x*blockIdx.x + threadIdx.x; + + if (i >= k) { + return; + } + + dst[i] = (T)op_silu_back((float)grad[i], (float)xf[i]); +} + +template +static void silu_back_cuda(const T * grad, const T * x, T * dst, const int k, cudaStream_t stream) { + const int num_blocks = (k + CUDA_SILU_BACK_BLOCK_SIZE - 1) / CUDA_SILU_BLOCK_SIZE; + silu_back_kernel<<>>(grad, x, dst, k); } void ggml_cuda_op_silu_back(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { @@ -314,179 +222,58 @@ void ggml_cuda_op_silu_back(ggml_backend_cuda_context & ctx, ggml_tensor * dst) GGML_ASSERT(ggml_is_contiguous(src0)); - GGML_ASSERT(src0->type == GGML_TYPE_F32); - GGML_ASSERT( dst->type == GGML_TYPE_F32); + GGML_ASSERT(src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16); + GGML_ASSERT( dst->type == GGML_TYPE_F32 || dst->type == GGML_TYPE_F16); + GGML_ASSERT(src0->type == dst->type); - silu_back_f32_cuda(src0_d, src1_d, dst_d, ggml_nelements(src0), stream); + if (src0->type == GGML_TYPE_F16) { + silu_back_cuda((const half *)src0_d, (const half *)src1_d, (half *)dst_d, ggml_nelements(src0), stream); + } else { + silu_back_cuda((const float*)src0_d, (const float*)src1_d, (float *)dst_d, ggml_nelements(src0), stream); + } } -void ggml_cuda_op_gelu_quick(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { - const ggml_tensor * src0 = dst->src[0]; - const float * src0_d = (const float *)src0->data; - float * dst_d = (float *)dst->data; - cudaStream_t stream = ctx.stream(); +/* leaky relu */ - GGML_ASSERT(ggml_is_contiguous(src0)); - - GGML_ASSERT(src0->type == GGML_TYPE_F32); - GGML_ASSERT( dst->type == GGML_TYPE_F32); - - gelu_quick_f32_cuda(src0_d, dst_d, ggml_nelements(src0), stream); +static __device__ __forceinline__ float op_leaky_relu(float x, const float negative_slope) { + return fmaxf(x, 0) + fminf(x, 0.0f) * negative_slope; } -void ggml_cuda_op_tanh(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { - const ggml_tensor * src0 = dst->src[0]; - const float * src0_d = (const float *)src0->data; - float * dst_d = (float *)dst->data; - cudaStream_t stream = ctx.stream(); +template +static __global__ void leaky_relu_kernel(const T * x, T * dst, const int k, const float negative_slope) { + const int i = blockDim.x*blockIdx.x + threadIdx.x; - GGML_ASSERT(ggml_is_contiguous(src0)); + if (i >= k) { + return; + } - GGML_ASSERT(src0->type == GGML_TYPE_F32); - GGML_ASSERT( dst->type == GGML_TYPE_F32); - - tanh_f32_cuda(src0_d, dst_d, ggml_nelements(src0), stream); + dst[i] = (T)op_leaky_relu((float)x[i], negative_slope); } -void ggml_cuda_op_relu(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { - const ggml_tensor * src0 = dst->src[0]; - const float * src0_d = (const float *)src0->data; - float * dst_d = (float *)dst->data; - cudaStream_t stream = ctx.stream(); - - GGML_ASSERT(ggml_is_contiguous(src0)); - - GGML_ASSERT(src0->type == GGML_TYPE_F32); - GGML_ASSERT( dst->type == GGML_TYPE_F32); - - relu_f32_cuda(src0_d, dst_d, ggml_nelements(src0), stream); -} - -void ggml_cuda_op_sigmoid(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { - const ggml_tensor * src0 = dst->src[0]; - const float * src0_d = (const float *)src0->data; - float * dst_d = (float *)dst->data; - cudaStream_t stream = ctx.stream(); - - GGML_ASSERT(ggml_is_contiguous(src0)); - - GGML_ASSERT(src0->type == GGML_TYPE_F32); - GGML_ASSERT( dst->type == GGML_TYPE_F32); - - sigmoid_f32_cuda(src0_d, dst_d, ggml_nelements(src0), stream); -} - -void ggml_cuda_op_hardsigmoid(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { - const ggml_tensor * src0 = dst->src[0]; - const float * src0_d = (const float *)src0->data; - float * dst_d = (float *)dst->data; - cudaStream_t stream = ctx.stream(); - - GGML_ASSERT(ggml_is_contiguous(src0)); - - GGML_ASSERT(src0->type == GGML_TYPE_F32); - GGML_ASSERT( dst->type == GGML_TYPE_F32); - - hardsigmoid_f32_cuda(src0_d, dst_d, ggml_nelements(src0), stream); -} - -void ggml_cuda_op_hardswish(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { - const ggml_tensor * src0 = dst->src[0]; - const float * src0_d = (const float *)src0->data; - float * dst_d = (float *)dst->data; - cudaStream_t stream = ctx.stream(); - - GGML_ASSERT(ggml_is_contiguous(src0)); - - GGML_ASSERT(src0->type == GGML_TYPE_F32); - GGML_ASSERT( dst->type == GGML_TYPE_F32); - - hardswish_f32_cuda(src0_d, dst_d, ggml_nelements(src0), stream); -} - -void ggml_cuda_op_exp(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { - const ggml_tensor * src0 = dst->src[0]; - const float * src0_d = (const float *)src0->data; - float * dst_d = (float *)dst->data; - cudaStream_t stream = ctx.stream(); - - GGML_ASSERT(ggml_is_contiguous(src0)); - - GGML_ASSERT(src0->type == GGML_TYPE_F32); - GGML_ASSERT( dst->type == GGML_TYPE_F32); - - exp_f32_cuda(src0_d, dst_d, ggml_nelements(src0), stream); +template +static void leaky_relu_cuda(const T * x, T * dst, const int k, const float negative_slope, cudaStream_t stream) { + const int num_blocks = (k + CUDA_RELU_BLOCK_SIZE - 1) / CUDA_RELU_BLOCK_SIZE; + leaky_relu_kernel<<>>(x, dst, k, negative_slope); } void ggml_cuda_op_leaky_relu(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { const ggml_tensor * src0 = dst->src[0]; - const float * src0_d = (const float *)src0->data; - float * dst_d = (float *)dst->data; + const void * src0_d = src0->data; + void * dst_d = dst->data; cudaStream_t stream = ctx.stream(); GGML_ASSERT(ggml_is_contiguous(src0)); - GGML_ASSERT(src0->type == GGML_TYPE_F32); - GGML_ASSERT( dst->type == GGML_TYPE_F32); + GGML_ASSERT(src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16); + GGML_ASSERT( dst->type == GGML_TYPE_F32 || dst->type == GGML_TYPE_F16); + GGML_ASSERT(src0->type == dst->type); float negative_slope; memcpy(&negative_slope, dst->op_params, sizeof(float)); - leaky_relu_f32_cuda(src0_d, dst_d, ggml_nelements(src0), negative_slope, stream); -} - -void ggml_cuda_op_sqr(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { - const ggml_tensor * src0 = dst->src[0]; - const float * src0_d = (const float *)src0->data; - float * dst_d = (float *)dst->data; - cudaStream_t stream = ctx.stream(); - - GGML_ASSERT(ggml_is_contiguous(src0)); - - GGML_ASSERT(src0->type == GGML_TYPE_F32); - GGML_ASSERT( dst->type == GGML_TYPE_F32); - - sqr_f32_cuda(src0_d, dst_d, ggml_nelements(src0), stream); -} - -void ggml_cuda_op_sqrt(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { - const ggml_tensor * src0 = dst->src[0]; - const float * src0_d = (const float *)src0->data; - float * dst_d = (float *)dst->data; - cudaStream_t stream = ctx.stream(); - - GGML_ASSERT(ggml_is_contiguous(src0)); - - GGML_ASSERT(src0->type == GGML_TYPE_F32); - GGML_ASSERT( dst->type == GGML_TYPE_F32); - - sqrt_f32_cuda(src0_d, dst_d, ggml_nelements(src0), stream); -} - -void ggml_cuda_op_sin(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { - const ggml_tensor * src0 = dst->src[0]; - const float * src0_d = (const float *)src0->data; - float * dst_d = (float *)dst->data; - cudaStream_t stream = ctx.stream(); - - GGML_ASSERT(ggml_is_contiguous(src0)); - - GGML_ASSERT(src0->type == GGML_TYPE_F32); - GGML_ASSERT( dst->type == GGML_TYPE_F32); - - sin_f32_cuda(src0_d, dst_d, ggml_nelements(src0), stream); -} - -void ggml_cuda_op_cos(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { - const ggml_tensor * src0 = dst->src[0]; - const float * src0_d = (const float *)src0->data; - float * dst_d = (float *)dst->data; - cudaStream_t stream = ctx.stream(); - - GGML_ASSERT(ggml_is_contiguous(src0)); - - GGML_ASSERT(src0->type == GGML_TYPE_F32); - GGML_ASSERT( dst->type == GGML_TYPE_F32); - - cos_f32_cuda(src0_d, dst_d, ggml_nelements(src0), stream); + if (src0->type == GGML_TYPE_F16) { + leaky_relu_cuda((const half *)src0_d, (half *)dst_d, ggml_nelements(src0), negative_slope, stream); + } else { + leaky_relu_cuda((const float *)src0_d, (float *)dst_d, ggml_nelements(src0), negative_slope, stream); + } } diff --git a/ggml/src/ggml-cuda/unary.cuh b/ggml/src/ggml-cuda/unary.cuh index e7f62643a..940a1feed 100644 --- a/ggml/src/ggml-cuda/unary.cuh +++ b/ggml/src/ggml-cuda/unary.cuh @@ -16,6 +16,10 @@ #define CUDA_SIN_BLOCK_SIZE 256 #define CUDA_COS_BLOCK_SIZE 256 +void ggml_cuda_op_abs(ggml_backend_cuda_context & ctx, ggml_tensor * dst); + +void ggml_cuda_op_sgn(ggml_backend_cuda_context & ctx, ggml_tensor * dst); + void ggml_cuda_op_neg(ggml_backend_cuda_context & ctx, ggml_tensor * dst); void ggml_cuda_op_step(ggml_backend_cuda_context & ctx, ggml_tensor * dst); @@ -49,3 +53,5 @@ void ggml_cuda_op_sqrt(ggml_backend_cuda_context & ctx, ggml_tensor * dst); void ggml_cuda_op_sin(ggml_backend_cuda_context & ctx, ggml_tensor * dst); void ggml_cuda_op_cos(ggml_backend_cuda_context & ctx, ggml_tensor * dst); + +void ggml_cuda_op_log(ggml_backend_cuda_context & ctx, ggml_tensor * dst); diff --git a/ggml/src/ggml-cuda/upscale.cu b/ggml/src/ggml-cuda/upscale.cu index cf513c3ad..524e97957 100644 --- a/ggml/src/ggml-cuda/upscale.cu +++ b/ggml/src/ggml-cuda/upscale.cu @@ -19,7 +19,7 @@ static __global__ void upscale_f32(const float * x, float * dst, int i02 = i12 / sf2; int i03 = i13 / sf3; - dst[index] = *(float *)((char *)x + i03 * nb03 + i02 * nb02 + i01 * nb01 + i00 * nb00); + dst[index] = *( (const float *)((const char *)x + i03 * nb03 + i02 * nb02 + i01 * nb01 + i00 * nb00) ); } static void upscale_f32_cuda(const float * x, float * dst, diff --git a/ggml/src/ggml-cuda/vendors/hip.h b/ggml/src/ggml-cuda/vendors/hip.h index 81964611c..420b41b8d 100644 --- a/ggml/src/ggml-cuda/vendors/hip.h +++ b/ggml/src/ggml-cuda/vendors/hip.h @@ -20,6 +20,7 @@ #define CUBLAS_STATUS_SUCCESS HIPBLAS_STATUS_SUCCESS #define CUBLAS_TF32_TENSOR_OP_MATH 0 #define CUDA_R_16F HIPBLAS_R_16F +#define CUDA_R_16BF HIPBLAS_R_16B #define CUDA_R_32F HIPBLAS_R_32F #define CU_DEVICE_ATTRIBUTE_VIRTUAL_MEMORY_MANAGEMENT_SUPPORTED hipDeviceAttributeVirtualMemoryManagementSupported #define CU_MEM_ALLOC_GRANULARITY_RECOMMENDED hipMemAllocationGranularityRecommended @@ -112,7 +113,7 @@ #define cudaGraphExecDestroy hipGraphExecDestroy #define cudaGraphLaunch hipGraphLaunch #define cudaErrorGraphExecUpdateFailure hipErrorGraphExecUpdateFailure -#define cudaGraphExecUpdateResultInfo hipGraphExecUpdateResult +#define cudaGraphExecUpdateResult hipGraphExecUpdateResult #define cudaGraphNodeType hipGraphNodeType #define cudaGraphNodeTypeKernel hipGraphNodeTypeKernel #define cudaGraphInstantiate hipGraphInstantiate @@ -129,6 +130,7 @@ #define cudaGraph_t hipGraph_t #define cudaStream_t hipStream_t #define cudaSuccess hipSuccess +#define cudaOccupancyMaxActiveBlocksPerMultiprocessor hipOccupancyMaxActiveBlocksPerMultiprocessor #define __trap() do { abort(); __builtin_unreachable(); } while(0) #define CUBLAS_STATUS_SUCCESS HIPBLAS_STATUS_SUCCESS #define CUBLAS_STATUS_NOT_INITIALIZED HIPBLAS_STATUS_NOT_INITIALIZED @@ -150,6 +152,10 @@ #define CDNA #endif +#if defined(__GFX12__) +#define RDNA4 +#endif + #if defined(__gfx1100__) || defined(__gfx1101__) || defined(__gfx1102__) || defined(__gfx1103__) || \ defined(__gfx1150__) || defined(__gfx1151__) #define RDNA3 diff --git a/ggml/src/ggml-cuda/vendors/musa.h b/ggml/src/ggml-cuda/vendors/musa.h index 6cc1b69ee..937779a90 100644 --- a/ggml/src/ggml-cuda/vendors/musa.h +++ b/ggml/src/ggml-cuda/vendors/musa.h @@ -15,6 +15,7 @@ #define CUBLAS_STATUS_SUCCESS MUBLAS_STATUS_SUCCESS #define CUBLAS_TF32_TENSOR_OP_MATH MUBLAS_MATH_MODE_DEFAULT #define CUDA_R_16F MUSA_R_16F +#define CUDA_R_16BF MUSA_R_16BF #define CUDA_R_32F MUSA_R_32F #define cublasComputeType_t cudaDataType_t #define cublasCreate mublasCreate @@ -119,7 +120,7 @@ #define cudaGraphExecDestroy musaGraphExecDestroy #define cudaGraphExec_t musaGraphExec_t #define cudaGraphExecUpdate musaGraphExecUpdate -#define cudaGraphExecUpdateResultInfo musaGraphExecUpdateResult +#define cudaGraphExecUpdateResult musaGraphExecUpdateResult #define cudaGraphGetNodes musaGraphGetNodes #define cudaGraphInstantiate musaGraphInstantiate #define cudaGraphKernelNodeGetParams musaGraphKernelNodeGetParams @@ -132,6 +133,8 @@ #define cudaGraph_t musaGraph_t #define cudaKernelNodeParams musaKernelNodeParams #define cudaStreamCaptureModeRelaxed musaStreamCaptureModeRelaxed +#define cudaStreamBeginCapture musaStreamBeginCapture #define cudaStreamEndCapture musaStreamEndCapture +#define cudaOccupancyMaxActiveBlocksPerMultiprocessor musaOccupancyMaxActiveBlocksPerMultiprocessor typedef mt_bfloat16 nv_bfloat16; diff --git a/ggml/src/ggml-cuda/wkv.cu b/ggml/src/ggml-cuda/wkv.cu new file mode 100644 index 000000000..d2fced705 --- /dev/null +++ b/ggml/src/ggml-cuda/wkv.cu @@ -0,0 +1,199 @@ +#include "common.cuh" +#include "wkv.cuh" + +template +static __global__ void rwkv_wkv_f32(const int B, const int T, const int C, const int H, const float * k, const float * v, const float * r, const float * tf, const float * td, const float * s, float * dst) { + const int tid = threadIdx.x; + const int bid = blockIdx.x; + + const int head_size = block_size; + const int batch_i = bid / H; + const int head_i = bid % H; + const int state_size = C * head_size; + const int n_seq_tokens = T / B; + + float state[head_size]; + __shared__ float _k[head_size], _r[head_size], _tf[head_size], _td[head_size]; + + #pragma unroll + for (int i = 0; i < head_size; i++) { + state[i] = s[batch_i * state_size + head_i * head_size * head_size + i * head_size + tid]; + } + + __syncthreads(); + _tf[tid] = tf[head_i * head_size + tid]; + __syncthreads(); + + for (int t = batch_i * n_seq_tokens * C + head_i * head_size + tid; t < (batch_i + 1) * n_seq_tokens * C + head_i * head_size + tid; t += C) { + __syncthreads(); + _k[tid] = k[t]; + _r[tid] = r[t]; + _td[tid] = td[t]; + __syncthreads(); + + const float _v = v[t]; + float y = 0; + for (int j = 0; j < head_size; j += 4) { + const float4& k = (float4&)(_k[j]); + const float4& r = (float4&)(_r[j]); + const float4& tf = (float4&)(_tf[j]); + const float4& td = (float4&)(_td[j]); + float4& s = (float4&)(state[j]); + float4 kv; + + kv.x = k.x * _v; + kv.y = k.y * _v; + kv.z = k.z * _v; + kv.w = k.w * _v; + + y += r.x * (tf.x * kv.x + s.x); + y += r.y * (tf.y * kv.y + s.y); + y += r.z * (tf.z * kv.z + s.z); + y += r.w * (tf.w * kv.w + s.w); + + s.x = s.x * td.x + kv.x; + s.y = s.y * td.y + kv.y; + s.z = s.z * td.z + kv.z; + s.w = s.w * td.w + kv.w; + } + dst[t] = y; + } + + #pragma unroll + for (int i = 0; i < head_size; i++) { + dst[T * C + batch_i * state_size + head_i * head_size * head_size + i * head_size + tid] = state[i]; + } +} + +template +static __global__ void rwkv_wkv7_f32(const int B, const int T, const int C, const int H, const float * r, const float * w, const float * k, const float * v, const float * a, const float * b, const float * s, float * dst) { + const int tid = threadIdx.x; + const int bid = blockIdx.x; + + const int head_size = block_size; + const int batch_i = bid / H; + const int head_i = bid % H; + const int state_size = C * head_size; + const int n_seq_tokens = T / B; + + float state[head_size]; + __shared__ float _r[head_size], _w[head_size], _k[head_size], _a[head_size], _b[head_size]; + +#ifndef GGML_USE_MUSA + #pragma unroll +#endif + for (int i = 0; i < head_size; i++) { + state[i] = s[batch_i * state_size + head_i * head_size * head_size + tid * head_size + i]; + } + + for (int t = batch_i * n_seq_tokens * C + head_i * head_size + tid; t < (batch_i + 1) * n_seq_tokens * C + head_i * head_size + tid; t += C) { + __syncthreads(); + _r[tid] = r[t]; + _w[tid] = w[t]; + _k[tid] = k[t]; + _a[tid] = a[t]; + _b[tid] = b[t]; + __syncthreads(); + + float sa = 0; + #pragma unroll + for (int j = 0; j < head_size; j += 4) + { + const float4& a = (float4&)(_a[j]); + const float4& s = (float4&)(state[j]); + sa += a.x * s.x; + sa += a.y * s.y; + sa += a.z * s.z; + sa += a.w * s.w; + } + + const float _v = v[t]; + float y = 0; + for (int j = 0; j < head_size; j += 4) { + const float4& r = (float4&)(_r[j]); + const float4& w = (float4&)(_w[j]); + const float4& k = (float4&)(_k[j]); + const float4& b = (float4&)(_b[j]); + float4& s = (float4&)(state[j]); + float4 kv; + + kv.x = k.x * _v; + kv.y = k.y * _v; + kv.z = k.z * _v; + kv.w = k.w * _v; + + s.x = s.x * w.x + kv.x + sa * b.x; + s.y = s.y * w.y + kv.y + sa * b.y; + s.z = s.z * w.z + kv.z + sa * b.z; + s.w = s.w * w.w + kv.w + sa * b.w; + + y += s.x * r.x; + y += s.y * r.y; + y += s.z * r.z; + y += s.w * r.w; + } + dst[t] = y; + } + + #pragma unroll + for (int i = 0; i < head_size; i++) { + dst[T * C + batch_i * state_size + head_i * head_size * head_size + tid * head_size + i] = state[i]; + } +} + +void ggml_cuda_op_rwkv_wkv6(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { + const float * k_d = (const float *)dst->src[0]->data; + const float * v_d = (const float *)dst->src[1]->data; + const float * r_d = (const float *)dst->src[2]->data; + const float * tf_d = (const float *)dst->src[3]->data; + const float * td_d = (const float *)dst->src[4]->data; + const float * s_d = (const float *)dst->src[5]->data; + + const int64_t B = dst->src[5]->ne[1]; + const int64_t T = dst->src[0]->ne[2]; + const int64_t C = dst->ne[0]; + const int64_t H = dst->src[0]->ne[1]; + + float * dst_d = (float *)dst->data; + + cudaStream_t stream = ctx.stream(); + + GGML_ASSERT(dst->src[5]->type == GGML_TYPE_F32); + GGML_ASSERT(C % H == 0); + GGML_ASSERT(C / H == CUDA_WKV_BLOCK_SIZE || C / H == CUDA_WKV_BLOCK_SIZE * 2); + + if (C / H == CUDA_WKV_BLOCK_SIZE) { + rwkv_wkv_f32<<>>(B, T, C, H, k_d, v_d, r_d, tf_d, td_d, s_d, dst_d); + } else { + rwkv_wkv_f32<<>>(B, T, C, H, k_d, v_d, r_d, tf_d, td_d, s_d, dst_d); + } +} + +void ggml_cuda_op_rwkv_wkv7(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { + const float * r_d = (const float *)dst->src[0]->data; + const float * w_d = (const float *)dst->src[1]->data; + const float * k_d = (const float *)dst->src[2]->data; + const float * v_d = (const float *)dst->src[3]->data; + const float * a_d = (const float *)dst->src[4]->data; + const float * b_d = (const float *)dst->src[5]->data; + const float * s_d = (const float *)dst->src[6]->data; + + const int64_t B = dst->src[6]->ne[1]; + const int64_t T = dst->src[0]->ne[2]; + const int64_t C = dst->ne[0]; + const int64_t H = dst->src[0]->ne[1]; + + float * dst_d = (float *)dst->data; + + cudaStream_t stream = ctx.stream(); + + GGML_ASSERT(dst->src[6]->type == GGML_TYPE_F32); + GGML_ASSERT(C % H == 0); + GGML_ASSERT(C / H == CUDA_WKV_BLOCK_SIZE || C / H == CUDA_WKV_BLOCK_SIZE * 2); + + if (C / H == CUDA_WKV_BLOCK_SIZE) { + rwkv_wkv7_f32<<>>(B, T, C, H, r_d, w_d, k_d, v_d, a_d, b_d, s_d, dst_d); + } else { + rwkv_wkv7_f32<<>>(B, T, C, H, r_d, w_d, k_d, v_d, a_d, b_d, s_d, dst_d); + } +} diff --git a/ggml/src/ggml-cuda/wkv6.cuh b/ggml/src/ggml-cuda/wkv.cuh similarity index 62% rename from ggml/src/ggml-cuda/wkv6.cuh rename to ggml/src/ggml-cuda/wkv.cuh index a7124ee51..9623dd7f8 100644 --- a/ggml/src/ggml-cuda/wkv6.cuh +++ b/ggml/src/ggml-cuda/wkv.cuh @@ -3,3 +3,5 @@ #define CUDA_WKV_BLOCK_SIZE 64 void ggml_cuda_op_rwkv_wkv6(ggml_backend_cuda_context & ctx, ggml_tensor * dst); + +void ggml_cuda_op_rwkv_wkv7(ggml_backend_cuda_context & ctx, ggml_tensor * dst); diff --git a/ggml/src/ggml-cuda/wkv6.cu b/ggml/src/ggml-cuda/wkv6.cu deleted file mode 100644 index bbdafbee5..000000000 --- a/ggml/src/ggml-cuda/wkv6.cu +++ /dev/null @@ -1,89 +0,0 @@ -#include "common.cuh" -#include "wkv6.cuh" - -static __global__ void rwkv_wkv_f32(const int B, const int T, const int C, const int H, const float * k, const float * v, const float * r, const float * tf, const float * td, const float * s, float * dst) { - const int tid = threadIdx.x; - const int bid = blockIdx.x; - - const int head_size = CUDA_WKV_BLOCK_SIZE; - const int batch_i = bid / H; - const int head_i = bid % H; - const int state_size = C * head_size; - const int n_seq_tokens = T / B; - - float state[head_size]; - __shared__ float _k[head_size], _r[head_size], _tf[head_size], _td[head_size]; - - #pragma unroll - for (int i = 0; i < head_size; i++) { - state[i] = s[batch_i * state_size + head_i * head_size * head_size + i * head_size + tid]; - } - - __syncthreads(); - _tf[tid] = tf[head_i * head_size + tid]; - __syncthreads(); - - for (int t = batch_i * n_seq_tokens * C + head_i * head_size + tid; t < (batch_i + 1) * n_seq_tokens * C + head_i * head_size + tid; t += C) { - __syncthreads(); - _k[tid] = k[t]; - _r[tid] = r[t]; - _td[tid] = td[t]; - __syncthreads(); - - const float _v = v[t]; - float y = 0; - for (int j = 0; j < head_size; j += 4) { - const float4& k = (float4&)(_k[j]); - const float4& r = (float4&)(_r[j]); - const float4& tf = (float4&)(_tf[j]); - const float4& td = (float4&)(_td[j]); - float4& s = (float4&)(state[j]); - float4 kv; - - kv.x = k.x * _v; - kv.y = k.y * _v; - kv.z = k.z * _v; - kv.w = k.w * _v; - - y += r.x * (tf.x * kv.x + s.x); - y += r.y * (tf.y * kv.y + s.y); - y += r.z * (tf.z * kv.z + s.z); - y += r.w * (tf.w * kv.w + s.w); - - s.x = s.x * td.x + kv.x; - s.y = s.y * td.y + kv.y; - s.z = s.z * td.z + kv.z; - s.w = s.w * td.w + kv.w; - } - dst[t] = y; - } - - #pragma unroll - for (int i = 0; i < head_size; i++) { - dst[T * C + batch_i * state_size + head_i * head_size * head_size + i * head_size + tid] = state[i]; - } -} - -void ggml_cuda_op_rwkv_wkv6(ggml_backend_cuda_context & ctx, ggml_tensor * dst) { - const float * k_d = (const float *)dst->src[0]->data; - const float * v_d = (const float *)dst->src[1]->data; - const float * r_d = (const float *)dst->src[2]->data; - const float * tf_d = (const float *)dst->src[3]->data; - const float * td_d = (const float *)dst->src[4]->data; - const float * s_d = (const float *)dst->src[5]->data; - - const int64_t B = dst->src[5]->ne[1]; - const int64_t T = dst->src[0]->ne[2]; - const int64_t C = dst->ne[0]; - const int64_t H = dst->src[0]->ne[1]; - - float * dst_d = (float *)dst->data; - - cudaStream_t stream = ctx.stream(); - - GGML_ASSERT(dst->src[5]->type == GGML_TYPE_F32); - GGML_ASSERT(C % H == 0); - GGML_ASSERT(C / H == CUDA_WKV_BLOCK_SIZE); // The current cuda kernel is designed for RWKV6, HEAD_SIZE == 64 - - rwkv_wkv_f32<<>>(B, T, C, H, k_d, v_d, r_d, tf_d, td_d, s_d, dst_d); -} diff --git a/ggml/src/ggml-hip/CMakeLists.txt b/ggml/src/ggml-hip/CMakeLists.txt index f4a468363..e3762649f 100644 --- a/ggml/src/ggml-hip/CMakeLists.txt +++ b/ggml/src/ggml-hip/CMakeLists.txt @@ -39,6 +39,12 @@ endif() find_package(hip REQUIRED) find_package(hipblas REQUIRED) find_package(rocblas REQUIRED) +if (GGML_HIP_ROCWMMA_FATTN) + CHECK_INCLUDE_FILE_CXX("rocwmma/rocwmma.hpp" FOUND_ROCWMMA) + if (NOT ${FOUND_ROCWMMA}) + message(FATAL_ERROR "rocwmma has not been found") + endif() +endif() if (${hip_VERSION} VERSION_LESS 5.5) message(FATAL_ERROR "At least ROCM/HIP V5.5 is required") @@ -107,6 +113,14 @@ if (GGML_HIP_NO_VMM) add_compile_definitions(GGML_HIP_NO_VMM) endif() +if (GGML_HIP_ROCWMMA_FATTN) + add_compile_definitions(GGML_HIP_ROCWMMA_FATTN) +endif() + +if (NOT GGML_CUDA_FA) + add_compile_definitions(GGML_CUDA_NO_FA) +endif() + if (CXX_IS_HIPCC) set_source_files_properties(${GGML_SOURCES_ROCM} PROPERTIES LANGUAGE CXX) target_link_libraries(ggml-hip PRIVATE hip::device) diff --git a/ggml/src/ggml-impl.h b/ggml/src/ggml-impl.h index eab017889..a19cfb14e 100644 --- a/ggml/src/ggml-impl.h +++ b/ggml/src/ggml-impl.h @@ -16,7 +16,7 @@ #include #endif // __ARM_FEATURE_SVE -#if defined(__ARM_NEON) && !defined(__CUDACC__) +#if defined(__ARM_NEON) && !defined(__CUDACC__) && !defined(__MUSACC__) // if YCM cannot find , make a symbolic link to it, for example: // // $ ln -sfn /Library/Developer/CommandLineTools/usr/lib/clang/13.1.6/include/arm_neon.h ./src/ @@ -148,8 +148,14 @@ struct ggml_map_custom2_op_params { struct ggml_map_custom3_op_params { ggml_custom3_op_t fun; - int n_tasks; - void * userdata; + int n_tasks; + void * userdata; +}; + +struct ggml_custom_op_params { + ggml_custom_op_t fun; + int n_tasks; + void * userdata; }; // bitset @@ -311,29 +317,28 @@ GGML_API void ggml_aligned_free(void * ptr, size_t size); // FP16 to FP32 conversion -#if defined(__ARM_NEON) - #if defined(_MSC_VER) || (defined(__CUDACC__) && __CUDACC_VER_MAJOR__ <= 11) - typedef uint16_t ggml_fp16_internal_t; - #else - typedef __fp16 ggml_fp16_internal_t; - #endif -#endif - -#if defined(__ARM_NEON) && !defined(_MSC_VER) && !(defined(__CUDACC__) && __CUDACC_VER_MAJOR__ <= 11) +// 16-bit float +// on Arm, we use __fp16 +// on x86, we use uint16_t +// +// for old CUDA compilers (<= 11), we use uint16_t: ref https://github.com/ggml-org/llama.cpp/pull/10616 +// for MUSA compilers , we use uint16_t: ref https://github.com/ggml-org/llama.cpp/pull/11843 +// +#if defined(__ARM_NEON) && !(defined(__CUDACC__) && __CUDACC_VER_MAJOR__ <= 11) && !defined(__MUSACC__) #define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x) #define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x) #define GGML_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x) static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) { - ggml_fp16_internal_t tmp; + __fp16 tmp; memcpy(&tmp, &h, sizeof(ggml_fp16_t)); return (float)tmp; } static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) { ggml_fp16_t res; - ggml_fp16_internal_t tmp = f; + __fp16 tmp = f; memcpy(&res, &tmp, sizeof(ggml_fp16_t)); return res; } @@ -357,8 +362,8 @@ GGML_API void ggml_aligned_free(void * ptr, size_t size); #define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x) static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) { - register float f; - register double d; + float f; + double d; __asm__( "mtfprd %0,%2\n" "xscvhpdp %0,%0\n" @@ -370,8 +375,8 @@ GGML_API void ggml_aligned_free(void * ptr, size_t size); } static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) { - register double d; - register ggml_fp16_t r; + double d; + ggml_fp16_t r; __asm__( /* xscvdphp can work on double or single precision */ "xscvdphp %0,%2\n" "mffprd %1,%0\n" : @@ -381,6 +386,35 @@ GGML_API void ggml_aligned_free(void * ptr, size_t size); return r; } +#elif defined(__riscv) && defined(GGML_RV_ZFH) + + static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) { + float f; + __asm__( + "fmv.h.x %[f], %[h]\n\t" + "fcvt.s.h %[f], %[f]" + : [f] "=&f" (f) + : [h] "r" (h) + ); + return f; + } + + static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) { + ggml_fp16_t res; + __asm__( + "fcvt.h.s %[f], %[f]\n\t" + "fmv.x.h %[h], %[f]" + : [h] "=&r" (res) + : [f] "f" (f) + ); + return res; + } + + #define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x) + #define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x) + #define GGML_FP16_TO_FP32(x) GGML_COMPUTE_FP16_TO_FP32(x) + #define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x) + #else // FP16 <-> FP32 @@ -456,7 +490,7 @@ GGML_API void ggml_aligned_free(void * ptr, size_t size); #define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x) #define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x) -#endif // defined(__ARM_NEON) && (!defined(__MSC_VER) +#endif // defined(__ARM_NEON) && !(defined(__CUDACC__) && __CUDACC_VER_MAJOR__ <= 11) && !defined(__MUSACC__) // precomputed f32 table for f16 (256 KB) // defined in ggml.c, initialized in ggml_init() diff --git a/ggml/src/ggml-metal/CMakeLists.txt b/ggml/src/ggml-metal/CMakeLists.txt index 89fcde2fa..e22232780 100644 --- a/ggml/src/ggml-metal/CMakeLists.txt +++ b/ggml/src/ggml-metal/CMakeLists.txt @@ -27,12 +27,12 @@ configure_file(../ggml-common.h ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-common.h configure_file(ggml-metal.metal ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.metal COPYONLY) configure_file(ggml-metal-impl.h ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal-impl.h COPYONLY) +set(METALLIB_COMMON "${CMAKE_CURRENT_SOURCE_DIR}/../ggml-common.h") if (GGML_METAL_EMBED_LIBRARY) enable_language(ASM) add_compile_definitions(GGML_METAL_EMBED_LIBRARY) - set(METALLIB_COMMON "${CMAKE_CURRENT_SOURCE_DIR}/../ggml-common.h") set(METALLIB_SOURCE "${CMAKE_CURRENT_SOURCE_DIR}/ggml-metal.metal") set(METALLIB_IMPL "${CMAKE_CURRENT_SOURCE_DIR}/ggml-metal-impl.h") @@ -88,12 +88,11 @@ else() add_custom_command( OUTPUT ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib - COMMAND xcrun -sdk macosx metal ${XC_FLAGS} -c ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.metal -o ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.air - COMMAND xcrun -sdk macosx metallib ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.air -o ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib - COMMAND rm -f ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.air + COMMAND xcrun -sdk macosx metal ${XC_FLAGS} -c ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.metal -o - | + xcrun -sdk macosx metallib - -o ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib COMMAND rm -f ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-common.h COMMAND rm -f ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.metal - DEPENDS ggml-metal.metal ggml-common.h + DEPENDS ggml-metal.metal ${METALLIB_COMMON} COMMENT "Compiling Metal kernels" ) diff --git a/ggml/src/ggml-metal/ggml-metal-impl.h b/ggml/src/ggml-metal/ggml-metal-impl.h index e3dc25f16..8721b272d 100644 --- a/ggml/src/ggml-metal/ggml-metal-impl.h +++ b/ggml/src/ggml-metal/ggml-metal-impl.h @@ -1,6 +1,70 @@ #ifndef GGML_METAL_IMPL #define GGML_METAL_IMPL +// kernel parameters for mat-vec threadgroups +// +// N_R0: number of src0 rows to process per simdgroup +// N_SG: number of simdgroups per threadgroup +// +// TODO: for optimal performance, become function of the device and work size + +#define N_R0_Q4_0 4 +#define N_SG_Q4_0 2 + +#define N_R0_Q4_1 4 +#define N_SG_Q4_1 2 + +#define N_R0_Q5_0 4 +#define N_SG_Q5_0 2 + +#define N_R0_Q5_1 4 +#define N_SG_Q5_1 2 + +#define N_R0_Q8_0 4 +#define N_SG_Q8_0 2 + +#define N_R0_Q2_K 4 +#define N_SG_Q2_K 2 + +#define N_R0_Q3_K 2 +#define N_SG_Q3_K 2 + +#define N_R0_Q4_K 4 +#define N_SG_Q4_K 2 + +#define N_R0_Q5_K 2 +#define N_SG_Q5_K 2 + +#define N_R0_Q6_K 1 +#define N_SG_Q6_K 2 + +#define N_R0_IQ1_S 4 +#define N_SG_IQ1_S 2 + +#define N_R0_IQ1_M 4 +#define N_SG_IQ1_M 2 + +#define N_R0_IQ2_XXS 4 +#define N_SG_IQ2_XXS 2 + +#define N_R0_IQ2_XS 4 +#define N_SG_IQ2_XS 2 + +#define N_R0_IQ2_S 4 +#define N_SG_IQ2_S 2 + +#define N_R0_IQ3_XXS 4 +#define N_SG_IQ3_XXS 2 + +#define N_R0_IQ3_S 4 +#define N_SG_IQ3_S 2 + +#define N_R0_IQ4_NL 2 +#define N_SG_IQ4_NL 2 + +#define N_R0_IQ4_XS 2 +#define N_SG_IQ4_XS 2 + // kernel argument structs // // - element counters (e.g. ne00) typically use int32_t to reduce register usage @@ -155,9 +219,12 @@ typedef struct { int32_t ne11; int32_t ne_12_2; // assume K and V are same shape int32_t ne_12_3; - uint64_t nb_12_1; - uint64_t nb_12_2; - uint64_t nb_12_3; + uint64_t nb11; + uint64_t nb12; + uint64_t nb13; + uint64_t nb21; + uint64_t nb22; + uint64_t nb23; uint64_t nb31; int32_t ne1; int32_t ne2; @@ -285,4 +352,246 @@ typedef struct { float eps; } ggml_metal_kargs_rms_norm; +typedef struct { + int32_t ne00; + int32_t ne00_4; + uint64_t nb01; + float eps; +} ggml_metal_kargs_l2_norm; + +typedef struct { + int64_t ne00; + int64_t ne01; + int64_t ne02; + uint64_t nb00; + uint64_t nb01; + uint64_t nb02; + int32_t n_groups; + float eps; +} ggml_metal_kargs_group_norm; + +typedef struct { + int32_t IC; + int32_t IL; + int32_t K; + int32_t s0; + uint64_t nb0; + uint64_t nb1; +} ggml_metal_kargs_conv_transpose_1d; + +typedef struct { + uint64_t ofs0; + uint64_t ofs1; + int32_t IW; + int32_t IH; + int32_t CHW; + int32_t s0; + int32_t s1; + int32_t p0; + int32_t p1; + int32_t d0; + int32_t d1; + int32_t N; + int32_t KH; + int32_t KW; + int32_t KHW; // KH * KW, pre-computed on CPU to save GPU resources +} ggml_metal_kargs_im2col; + +typedef struct { + int64_t ne00; + int64_t ne01; + int64_t ne02; + int64_t ne03; + uint64_t nb00; + uint64_t nb01; + uint64_t nb02; + uint64_t nb03; + int64_t ne10; + int64_t ne11; + int64_t ne12; + int64_t ne13; + uint64_t nb10; + uint64_t nb11; + uint64_t nb12; + uint64_t nb13; + int64_t ne0; + int64_t ne1; + int64_t ne2; + int64_t ne3; + uint64_t nb0; + uint64_t nb1; + uint64_t nb2; + uint64_t nb3; +} ggml_metal_kargs_sum_rows; + +typedef struct { + int64_t ne00; + int64_t ne01; + int64_t ne02; + float scale; + float max_bias; + float m0; + float m1; + uint32_t n_head_log2; +} ggml_metal_kargs_soft_max; + +typedef struct { + int64_t ne00; + int64_t ne01; + int n_past; +} ggml_metal_kargs_diag_mask_inf; + +typedef struct { + int64_t ne00; + int64_t ne01; + int64_t ne02; + uint64_t nb00; + uint64_t nb01; + uint64_t nb02; + int64_t ne10; + int64_t ne11; + uint64_t nb10; + uint64_t nb11; + int64_t ne0; + int64_t ne1; + int64_t ne2; + uint64_t nb0; + uint64_t nb1; + uint64_t nb2; +} ggml_metal_kargs_ssm_conv; + +typedef struct { + int64_t d_state; + int64_t d_inner; + int64_t n_seq_tokens; + int64_t n_seqs; + uint64_t nb00; + uint64_t nb01; + uint64_t nb02; + uint64_t nb10; + uint64_t nb11; + uint64_t nb12; + uint64_t nb13; + uint64_t nb20; + uint64_t nb21; + uint64_t nb22; + uint64_t nb30; + uint64_t nb31; + uint64_t nb40; + uint64_t nb41; + uint64_t nb42; + uint64_t nb50; + uint64_t nb51; + uint64_t nb52; +} ggml_metal_kargs_ssm_scan; + +typedef struct { + int64_t ne00; + uint64_t nb01; + uint64_t nb02; + int64_t ne10; + uint64_t nb10; + uint64_t nb11; + uint64_t nb1; + uint64_t nb2; +} ggml_metal_kargs_get_rows; + +typedef struct { + int64_t ne00; + int64_t ne01; + int64_t ne02; + int64_t ne03; + uint64_t nb00; + uint64_t nb01; + uint64_t nb02; + uint64_t nb03; + int64_t ne0; + int64_t ne1; + int64_t ne2; + int64_t ne3; + uint64_t nb0; + uint64_t nb1; + uint64_t nb2; + uint64_t nb3; + float sf0; + float sf1; + float sf2; + float sf3; +} ggml_metal_kargs_upscale; + +typedef struct { + int64_t ne00; + int64_t ne01; + int64_t ne02; + int64_t ne03; + uint64_t nb00; + uint64_t nb01; + uint64_t nb02; + uint64_t nb03; + int64_t ne0; + int64_t ne1; + int64_t ne2; + int64_t ne3; + uint64_t nb0; + uint64_t nb1; + uint64_t nb2; + uint64_t nb3; +} ggml_metal_kargs_pad; + +typedef struct { + int64_t ne00; + int64_t ne01; + int64_t ne02; + int64_t ne03; + uint64_t nb00; + uint64_t nb01; + uint64_t nb02; + uint64_t nb03; + int64_t ne0; + int64_t ne1; + int64_t ne2; + int64_t ne3; + uint64_t nb0; + uint64_t nb1; + uint64_t nb2; + uint64_t nb3; + int32_t p0; + int32_t p1; +} ggml_metal_kargs_pad_reflect_1d; + +typedef struct { + uint64_t nb1; + int dim; + int max_period; +} ggml_metal_kargs_timestep_embedding; + +typedef struct { + float slope; +} ggml_metal_kargs_leaky_relu; + +typedef struct { + int64_t ncols; + int64_t ncols_pad; +} ggml_metal_kargs_argsort; + +typedef struct { + int64_t ne0; + float start; + float step; +} ggml_metal_kargs_arange; + +typedef struct { + int32_t k0; + int32_t k1; + int32_t s0; + int32_t s1; + int32_t p0; + int32_t p1; + int64_t IH; + int64_t IW; + int64_t OH; + int64_t OW; + int64_t parallel_elements; +} ggml_metal_kargs_pool_2d; + #endif // GGML_METAL_IMPL diff --git a/ggml/src/ggml-metal/ggml-metal.m b/ggml/src/ggml-metal/ggml-metal.m index 944d90af3..9f1c6c6cc 100644 --- a/ggml/src/ggml-metal/ggml-metal.m +++ b/ggml/src/ggml-metal/ggml-metal.m @@ -24,7 +24,7 @@ #endif // create residency sets only on macOS >= 15.0 -#if TARGET_OS_OSX && __MAC_OS_X_VERSION_MAX_ALLOWED >= 150000 || \ +#if !TARGET_CPU_X86_64 && TARGET_OS_OSX && __MAC_OS_X_VERSION_MAX_ALLOWED >= 150000 || \ TARGET_OS_IOS && __IPHONE_OS_VERSION_MAX_ALLOWED >= 180000 || \ TARGET_OS_TV && __TV_OS_VERSION_MAX_ALLOWED >= 180000 || \ TARGET_OS_VISION && __VISION_OS_VERSION_MAX_ALLOWED >= 200000 @@ -46,6 +46,7 @@ static struct ggml_backend_device g_ggml_backend_metal_device; static struct ggml_backend_metal_device_context { id mtl_device; int mtl_device_ref_count; + id mtl_library; bool has_simdgroup_reduction; bool has_simdgroup_mm; @@ -57,6 +58,7 @@ static struct ggml_backend_metal_device_context { } g_ggml_ctx_dev_main = { /*.mtl_device =*/ nil, /*.mtl_device_ref_count =*/ 0, + /*.mtl_library =*/ nil, /*.has_simdgroup_reduction =*/ false, /*.has_simdgroup_mm =*/ false, /*.has_residency_sets =*/ false, @@ -108,6 +110,11 @@ static void ggml_backend_metal_device_rel(struct ggml_backend_metal_device_conte ctx->mtl_device_ref_count--; if (ctx->mtl_device_ref_count == 0) { + if (ctx->mtl_library) { + [ctx->mtl_library release]; + ctx->mtl_library = nil; + } + if (ctx->mtl_device) { [ctx->mtl_device release]; ctx->mtl_device = nil; @@ -177,10 +184,13 @@ enum ggml_metal_kernel_type { GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ4_XS, GGML_METAL_KERNEL_TYPE_GET_ROWS_I32, GGML_METAL_KERNEL_TYPE_RMS_NORM, + GGML_METAL_KERNEL_TYPE_L2_NORM, GGML_METAL_KERNEL_TYPE_GROUP_NORM, GGML_METAL_KERNEL_TYPE_NORM, GGML_METAL_KERNEL_TYPE_SSM_CONV_F32, GGML_METAL_KERNEL_TYPE_SSM_SCAN_F32, + GGML_METAL_KERNEL_TYPE_RWKV_WKV6_F32, + GGML_METAL_KERNEL_TYPE_RWKV_WKV7_F32, GGML_METAL_KERNEL_TYPE_MUL_MV_F32_F32, GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32, GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_1ROW, @@ -341,42 +351,56 @@ enum ggml_metal_kernel_type { GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H96, GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H112, GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H128, + GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H192, + GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_HK192_HV128, GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H256, GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H64, GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H80, GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H96, GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H112, GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H128, + GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H192, + GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_HK192_HV128, GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H256, GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_H64, GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_H80, GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_H96, GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_H112, GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_H128, + GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_H192, + GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_HK192_HV128, GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_H256, GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_H64, GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_H80, GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_H96, GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_H112, GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_H128, + GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_H192, + GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_HK192_HV128, GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_H256, GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_H64, GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_H80, GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_H96, GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_H112, GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_H128, + GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_H192, + GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_HK192_HV128, GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_H256, GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_H64, GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_H80, GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_H96, GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_H112, GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_H128, + GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_H192, + GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_HK192_HV128, GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_H256, GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H64, GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H80, GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H96, GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H112, GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H128, + GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H192, + GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_HK192_HV128, GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H256, GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H128, GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_BF16_H128, @@ -385,6 +409,20 @@ enum ggml_metal_kernel_type { GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_0_H128, GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_1_H128, GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q8_0_H128, + GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H192, + GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_BF16_H192, + GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_0_H192, + GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_1_H192, + GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_0_H192, + GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_1_H192, + GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q8_0_H192, + GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_HK192_HV128, + GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_BF16_HK192_HV128, + GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_0_HK192_HV128, + GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_1_HK192_HV128, + GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_0_HK192_HV128, + GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_1_HK192_HV128, + GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q8_0_HK192_HV128, GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H256, GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_BF16_H256, GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_0_H256, @@ -407,6 +445,16 @@ enum ggml_metal_kernel_type { GGML_METAL_KERNEL_TYPE_CPY_F32_Q5_0, GGML_METAL_KERNEL_TYPE_CPY_F32_Q5_1, GGML_METAL_KERNEL_TYPE_CPY_F32_IQ4_NL, + GGML_METAL_KERNEL_TYPE_CPY_Q4_0_F32, + GGML_METAL_KERNEL_TYPE_CPY_Q4_0_F16, + GGML_METAL_KERNEL_TYPE_CPY_Q4_1_F32, + GGML_METAL_KERNEL_TYPE_CPY_Q4_1_F16, + GGML_METAL_KERNEL_TYPE_CPY_Q5_0_F32, + GGML_METAL_KERNEL_TYPE_CPY_Q5_0_F16, + GGML_METAL_KERNEL_TYPE_CPY_Q5_1_F32, + GGML_METAL_KERNEL_TYPE_CPY_Q5_1_F16, + GGML_METAL_KERNEL_TYPE_CPY_Q8_0_F32, + GGML_METAL_KERNEL_TYPE_CPY_Q8_0_F16, GGML_METAL_KERNEL_TYPE_CONCAT, GGML_METAL_KERNEL_TYPE_SQR, GGML_METAL_KERNEL_TYPE_SQRT, @@ -457,11 +505,13 @@ struct ggml_backend_metal_context { // for now it is easier to work in a separate file // static NSString * const msl_library_source = @"see metal.metal"; +#if !GGML_METAL_EMBED_LIBRARY // Here to assist with NSBundle Path Hack @interface GGMLMetalClass : NSObject @end @implementation GGMLMetalClass @end +#endif static void * ggml_metal_host_malloc(size_t n) { void * data = NULL; @@ -483,6 +533,139 @@ static void * ggml_metal_host_malloc(size_t n) { return data; } +// load library +// +// - first check if the library is embedded +// - then check if the library is in the bundle +// - if not found, load the source and compile it +// - if that fails, return NULL +static id ggml_metal_load_library(id device, bool use_bfloat) { + id metal_library = nil; + NSError * error = nil; + NSString * src = nil; + +#if GGML_METAL_EMBED_LIBRARY + GGML_LOG_INFO("%s: using embedded metal library\n", __func__); + + extern const char ggml_metallib_start[]; + extern const char ggml_metallib_end[]; + + src = [[NSString alloc] initWithBytes:ggml_metallib_start length:(ggml_metallib_end-ggml_metallib_start) encoding:NSUTF8StringEncoding]; + +#else + +#ifdef SWIFT_PACKAGE + NSBundle * bundle = SWIFTPM_MODULE_BUNDLE; +#else + NSBundle * bundle = [NSBundle bundleForClass:[GGMLMetalClass class]]; +#endif + + NSString * path_lib = [bundle pathForResource:@"default" ofType:@"metallib"]; + if (path_lib == nil) { + // Try to find the resource in the directory where the current binary located. + NSString * current_binary = [[NSProcessInfo processInfo] arguments][0]; + NSString * bin_dir = [current_binary stringByDeletingLastPathComponent]; + NSString * default_metallib_path = [NSString pathWithComponents:@[bin_dir, @"default.metallib"]]; + if ([[NSFileManager defaultManager] isReadableFileAtPath:default_metallib_path]) { + GGML_LOG_INFO("%s: found '%s'\n", __func__, [default_metallib_path UTF8String]); + NSDictionary * atts = [[NSFileManager defaultManager] attributesOfItemAtPath:default_metallib_path error:&error]; + if (atts && atts[NSFileType] == NSFileTypeSymbolicLink) { + // Optionally, if this is a symlink, try to resolve it. + default_metallib_path = [[NSFileManager defaultManager] destinationOfSymbolicLinkAtPath:default_metallib_path error:&error]; + if (default_metallib_path && [default_metallib_path length] > 0 && ![[default_metallib_path substringToIndex:1] isEqualToString:@"/"]) { + // It is a relative path, adding the binary directory as directory prefix. + default_metallib_path = [NSString pathWithComponents:@[bin_dir, default_metallib_path]]; + } + if (!default_metallib_path || ![[NSFileManager defaultManager] isReadableFileAtPath:default_metallib_path]) { + // Link to the resource could not be resolved. + default_metallib_path = nil; + } else { + GGML_LOG_INFO("%s: symlink resolved '%s'\n", __func__, [default_metallib_path UTF8String]); + } + } + } else { + // The resource couldn't be found in the binary's directory. + default_metallib_path = nil; + } + path_lib = default_metallib_path; + } + + if (path_lib != nil) { + // pre-compiled library found + NSURL * libURL = [NSURL fileURLWithPath:path_lib]; + GGML_LOG_INFO("%s: loading '%s'\n", __func__, [path_lib UTF8String]); + + metal_library = [device newLibraryWithURL:libURL error:&error]; + if (error) { + GGML_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]); + return NULL; + } + } else { + GGML_LOG_INFO("%s: default.metallib not found, loading from source\n", __func__); + + NSString * path_source; + NSString * path_resource = [[NSProcessInfo processInfo].environment objectForKey:@"GGML_METAL_PATH_RESOURCES"]; + + GGML_LOG_INFO("%s: GGML_METAL_PATH_RESOURCES = %s\n", __func__, path_resource ? [path_resource UTF8String] : "nil"); + + if (path_resource) { + path_source = [path_resource stringByAppendingPathComponent:@"ggml-metal.metal"]; + } else { + path_source = [bundle pathForResource:@"ggml-metal" ofType:@"metal"]; + } + + if (path_source == nil) { + GGML_LOG_WARN("%s: error: could not use bundle path to find ggml-metal.metal, falling back to trying cwd\n", __func__); + path_source = @"ggml-metal.metal"; + } + + GGML_LOG_INFO("%s: loading '%s'\n", __func__, [path_source UTF8String]); + + src = [NSString stringWithContentsOfFile:path_source encoding:NSUTF8StringEncoding error:&error]; + if (error) { + GGML_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]); + return NULL; + } + } +#endif + + if (!metal_library) { + @autoreleasepool { + // dictionary of preprocessor macros + NSMutableDictionary * prep = [NSMutableDictionary dictionary]; + + if (use_bfloat) { + [prep setObject:@"1" forKey:@"GGML_METAL_USE_BF16"]; + } + +#if GGML_METAL_EMBED_LIBRARY + [prep setObject:@"1" forKey:@"GGML_METAL_EMBED_LIBRARY"]; +#endif + + MTLCompileOptions * options = [MTLCompileOptions new]; + options.preprocessorMacros = prep; + + //[options setFastMathEnabled:false]; + + metal_library = [device newLibraryWithSource:src options:options error:&error]; + if (error) { + GGML_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]); + return NULL; + } + +#if !__has_feature(objc_arc) + [options release]; +#endif + } + } + +#if GGML_METAL_EMBED_LIBRARY + [src release]; +#endif // GGML_METAL_EMBED_LIBRARY + + return metal_library; +} + static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t dev) { GGML_LOG_INFO("%s: allocating\n", __func__); @@ -510,137 +693,14 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de ctx->d_queue = dispatch_queue_create("ggml-metal", DISPATCH_QUEUE_CONCURRENT); - id metal_library; - // load library - // - // - first check if the library is embedded - // - then check if the library is in the bundle - // - if not found, load the source and compile it - // - if that fails, return NULL - { - NSBundle * bundle = nil; -#ifdef SWIFT_PACKAGE - bundle = SWIFTPM_MODULE_BUNDLE; -#else - bundle = [NSBundle bundleForClass:[GGMLMetalClass class]]; -#endif - - NSError * error = nil; - -#if GGML_METAL_EMBED_LIBRARY - const bool try_metallib = false; -#else - const bool try_metallib = true; -#endif - - NSString * path_lib = [bundle pathForResource:@"default" ofType:@"metallib"]; - if (path_lib == nil) { - // Try to find the resource in the directory where the current binary located. - NSString * current_binary = [[NSProcessInfo processInfo] arguments][0]; - NSString * bin_dir = [current_binary stringByDeletingLastPathComponent]; - NSString * default_metallib_path = [NSString pathWithComponents:@[bin_dir, @"default.metallib"]]; - if ([[NSFileManager defaultManager] isReadableFileAtPath:default_metallib_path]) { - GGML_LOG_INFO("%s: found '%s'\n", __func__, [default_metallib_path UTF8String]); - NSDictionary * atts = [[NSFileManager defaultManager] attributesOfItemAtPath:default_metallib_path error:&error]; - if (atts && atts[NSFileType] == NSFileTypeSymbolicLink) { - // Optionally, if this is a symlink, try to resolve it. - default_metallib_path = [[NSFileManager defaultManager] destinationOfSymbolicLinkAtPath:default_metallib_path error:&error]; - if (default_metallib_path && [default_metallib_path length] > 0 && ![[default_metallib_path substringToIndex:1] isEqualToString:@"/"]) { - // It is a relative path, adding the binary directory as directory prefix. - default_metallib_path = [NSString pathWithComponents:@[bin_dir, default_metallib_path]]; - } - if (!default_metallib_path || ![[NSFileManager defaultManager] isReadableFileAtPath:default_metallib_path]) { - // Link to the resource could not be resolved. - default_metallib_path = nil; - } else { - GGML_LOG_INFO("%s: symlink resolved '%s'\n", __func__, [default_metallib_path UTF8String]); - } - } - } else { - // The resource couldn't be found in the binary's directory. - default_metallib_path = nil; - } - path_lib = default_metallib_path; - } - - if (try_metallib && path_lib != nil) { - // pre-compiled library found - NSURL * libURL = [NSURL fileURLWithPath:path_lib]; - GGML_LOG_INFO("%s: loading '%s'\n", __func__, [path_lib UTF8String]); - - metal_library = [device newLibraryWithURL:libURL error:&error]; - if (error) { - GGML_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]); - return NULL; - } - } else { -#if GGML_METAL_EMBED_LIBRARY - GGML_LOG_INFO("%s: using embedded metal library\n", __func__); - - extern const char ggml_metallib_start[]; - extern const char ggml_metallib_end[]; - - NSString * src = [[NSString alloc] initWithBytes:ggml_metallib_start length:(ggml_metallib_end-ggml_metallib_start) encoding:NSUTF8StringEncoding]; -#else - GGML_LOG_INFO("%s: default.metallib not found, loading from source\n", __func__); - - NSString * path_source; - NSString * path_resource = [[NSProcessInfo processInfo].environment objectForKey:@"GGML_METAL_PATH_RESOURCES"]; - - GGML_LOG_INFO("%s: GGML_METAL_PATH_RESOURCES = %s\n", __func__, path_resource ? [path_resource UTF8String] : "nil"); - - if (path_resource) { - path_source = [path_resource stringByAppendingPathComponent:@"ggml-metal.metal"]; - } else { - path_source = [bundle pathForResource:@"ggml-metal" ofType:@"metal"]; - } - - if (path_source == nil) { - GGML_LOG_WARN("%s: error: could not use bundle path to find ggml-metal.metal, falling back to trying cwd\n", __func__); - path_source = @"ggml-metal.metal"; - } - - GGML_LOG_INFO("%s: loading '%s'\n", __func__, [path_source UTF8String]); - - NSString * src = [NSString stringWithContentsOfFile:path_source encoding:NSUTF8StringEncoding error:&error]; - if (error) { - GGML_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]); - return NULL; - } -#endif // GGML_METAL_EMBED_LIBRARY - - @autoreleasepool { - // dictionary of preprocessor macros - NSMutableDictionary * prep = [NSMutableDictionary dictionary]; - - if (ctx_dev->use_bfloat) { - [prep setObject:@"1" forKey:@"GGML_METAL_USE_BF16"]; - } - -#if GGML_METAL_EMBED_LIBRARY - [prep setObject:@"1" forKey:@"GGML_METAL_EMBED_LIBRARY"]; -#endif - - MTLCompileOptions * options = [MTLCompileOptions new]; - options.preprocessorMacros = prep; - - //[options setFastMathEnabled:false]; - - metal_library = [device newLibraryWithSource:src options:options error:&error]; - if (error) { - GGML_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]); - return NULL; - } - -#if !__has_feature(objc_arc) - [options release]; -#endif - } -#if GGML_METAL_EMBED_LIBRARY - [src release]; -#endif // GGML_METAL_EMBED_LIBRARY - } + if (ctx_dev->mtl_library == nil) { + ctx_dev->mtl_library = ggml_metal_load_library(device, ctx_dev->use_bfloat); + } + id metal_library = ctx_dev->mtl_library; + if (metal_library == nil) { + GGML_LOG_ERROR("%s: error: metal library is nil\n", __func__); + return NULL; } // print MTL GPU family: @@ -714,7 +774,6 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de [metal_function release]; \ if (error) { \ GGML_LOG_ERROR("%s: error: load pipeline error: %s\n", __func__, [[error description] UTF8String]); \ - [metal_library release]; \ return NULL; \ } \ } else { \ @@ -727,304 +786,343 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de // simd_sum and simd_max requires MTLGPUFamilyApple7 - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ADD, add, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ADD_ROW, add_row, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SUB, sub, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SUB_ROW, sub_row, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL, mul, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_ROW, mul_row, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_DIV, div, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_DIV_ROW, div_row, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_REPEAT_F32, repeat_f32, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_REPEAT_F16, repeat_f16, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_REPEAT_I32, repeat_i32, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_REPEAT_I16, repeat_i16, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SCALE, scale, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SCALE_4, scale_4, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CLAMP, clamp, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_TANH, tanh, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_RELU, relu, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SIGMOID, sigmoid, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GELU, gelu, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GELU_4, gelu_4, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GELU_QUICK, gelu_quick, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GELU_QUICK_4, gelu_quick_4, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SILU, silu, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SILU_4, silu_4, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ELU, elu, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SOFT_MAX_F16, soft_max_f16, has_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SOFT_MAX_F16_4, soft_max_f16_4, has_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SOFT_MAX_F32, soft_max_f32, has_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SOFT_MAX_F32_4, soft_max_f32_4, has_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_DIAG_MASK_INF, diag_mask_inf, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_DIAG_MASK_INF_8, diag_mask_inf_8, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_F32, get_rows_f32, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_F16, get_rows_f16, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_BF16, get_rows_bf16, use_bfloat); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q4_0, get_rows_q4_0, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q4_1, get_rows_q4_1, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q5_0, get_rows_q5_0, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q5_1, get_rows_q5_1, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q8_0, get_rows_q8_0, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q2_K, get_rows_q2_K, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q3_K, get_rows_q3_K, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q4_K, get_rows_q4_K, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q5_K, get_rows_q5_K, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q6_K, get_rows_q6_K, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_XXS, get_rows_iq2_xxs, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_XS, get_rows_iq2_xs, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ3_XXS, get_rows_iq3_xxs, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ3_S, get_rows_iq3_s, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_S, get_rows_iq2_s, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ1_S, get_rows_iq1_s, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ1_M, get_rows_iq1_m, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ4_NL, get_rows_iq4_nl, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ4_XS, get_rows_iq4_xs, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_I32, get_rows_i32, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_RMS_NORM, rms_norm, has_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GROUP_NORM, group_norm, has_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_NORM, norm, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SSM_CONV_F32, ssm_conv_f32, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SSM_SCAN_F32, ssm_scan_f32, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F32_F32, mul_mv_f32_f32, has_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_F32, mul_mv_bf16_f32, has_simdgroup_reduction && use_bfloat); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_F32_1ROW, mul_mv_bf16_f32_1row, has_simdgroup_reduction && use_bfloat); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_F32_L4, mul_mv_bf16_f32_l4, has_simdgroup_reduction && use_bfloat); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_BF16, mul_mv_bf16_bf16, has_simdgroup_reduction && use_bfloat); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32, mul_mv_f16_f32, has_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_1ROW, mul_mv_f16_f32_1row, has_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_L4, mul_mv_f16_f32_l4, has_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F16, mul_mv_f16_f16, has_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_0_F32, mul_mv_q4_0_f32, has_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_1_F32, mul_mv_q4_1_f32, has_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_0_F32, mul_mv_q5_0_f32, has_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_1_F32, mul_mv_q5_1_f32, has_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q8_0_F32, mul_mv_q8_0_f32, has_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_F16_F32_R1_2, mul_mv_ext_f16_f32_r1_2, has_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_F16_F32_R1_3, mul_mv_ext_f16_f32_r1_3, has_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_F16_F32_R1_4, mul_mv_ext_f16_f32_r1_4, has_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_F16_F32_R1_5, mul_mv_ext_f16_f32_r1_5, has_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_0_F32_R1_2, mul_mv_ext_q4_0_f32_r1_2, has_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_0_F32_R1_3, mul_mv_ext_q4_0_f32_r1_3, has_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_0_F32_R1_4, mul_mv_ext_q4_0_f32_r1_4, has_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_0_F32_R1_5, mul_mv_ext_q4_0_f32_r1_5, has_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_1_F32_R1_2, mul_mv_ext_q4_1_f32_r1_2, has_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_1_F32_R1_3, mul_mv_ext_q4_1_f32_r1_3, has_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_1_F32_R1_4, mul_mv_ext_q4_1_f32_r1_4, has_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_1_F32_R1_5, mul_mv_ext_q4_1_f32_r1_5, has_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q5_0_F32_R1_2, mul_mv_ext_q5_0_f32_r1_2, has_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q5_0_F32_R1_3, mul_mv_ext_q5_0_f32_r1_3, has_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q5_0_F32_R1_4, mul_mv_ext_q5_0_f32_r1_4, has_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q5_0_F32_R1_5, mul_mv_ext_q5_0_f32_r1_5, has_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q5_1_F32_R1_2, mul_mv_ext_q5_1_f32_r1_2, has_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q5_1_F32_R1_3, mul_mv_ext_q5_1_f32_r1_3, has_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q5_1_F32_R1_4, mul_mv_ext_q5_1_f32_r1_4, has_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q5_1_F32_R1_5, mul_mv_ext_q5_1_f32_r1_5, has_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q8_0_F32_R1_2, mul_mv_ext_q8_0_f32_r1_2, has_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q8_0_F32_R1_3, mul_mv_ext_q8_0_f32_r1_3, has_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q8_0_F32_R1_4, mul_mv_ext_q8_0_f32_r1_4, has_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q8_0_F32_R1_5, mul_mv_ext_q8_0_f32_r1_5, has_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_K_F32_R1_2, mul_mv_ext_q4_K_f32_r1_2, has_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_K_F32_R1_3, mul_mv_ext_q4_K_f32_r1_3, has_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_K_F32_R1_4, mul_mv_ext_q4_K_f32_r1_4, has_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_K_F32_R1_5, mul_mv_ext_q4_K_f32_r1_5, has_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q5_K_F32_R1_2, mul_mv_ext_q5_K_f32_r1_2, has_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q5_K_F32_R1_3, mul_mv_ext_q5_K_f32_r1_3, has_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q5_K_F32_R1_4, mul_mv_ext_q5_K_f32_r1_4, has_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q5_K_F32_R1_5, mul_mv_ext_q5_K_f32_r1_5, has_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q6_K_F32_R1_2, mul_mv_ext_q6_K_f32_r1_2, has_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q6_K_F32_R1_3, mul_mv_ext_q6_K_f32_r1_3, has_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q6_K_F32_R1_4, mul_mv_ext_q6_K_f32_r1_4, has_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q6_K_F32_R1_5, mul_mv_ext_q6_K_f32_r1_5, has_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_IQ4_NL_F32_R1_2, mul_mv_ext_iq4_nl_f32_r1_2, has_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_IQ4_NL_F32_R1_3, mul_mv_ext_iq4_nl_f32_r1_3, has_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_IQ4_NL_F32_R1_4, mul_mv_ext_iq4_nl_f32_r1_4, has_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_IQ4_NL_F32_R1_5, mul_mv_ext_iq4_nl_f32_r1_5, has_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q2_K_F32, mul_mv_q2_K_f32, has_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q3_K_F32, mul_mv_q3_K_f32, has_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_K_F32, mul_mv_q4_K_f32, has_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_K_F32, mul_mv_q5_K_f32, has_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q6_K_F32, mul_mv_q6_K_f32, has_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_XXS_F32, mul_mv_iq2_xxs_f32, has_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_XS_F32, mul_mv_iq2_xs_f32, has_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ3_XXS_F32, mul_mv_iq3_xxs_f32, has_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ3_S_F32, mul_mv_iq3_s_f32, has_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_S_F32, mul_mv_iq2_s_f32, has_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ1_S_F32, mul_mv_iq1_s_f32, has_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ1_M_F32, mul_mv_iq1_m_f32, has_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ4_NL_F32, mul_mv_iq4_nl_f32, has_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ4_XS_F32, mul_mv_iq4_xs_f32, has_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F32_F32, mul_mv_id_f32_f32, has_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F32, mul_mv_id_f16_f32, has_simdgroup_reduction); - //GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F32_1ROW, mul_mv_id_f16_f32_1row, has_simdgroup_reduction); - //GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F32_L4, mul_mv_id_f16_f32_l4, has_simdgroup_reduction); - //GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F16, mul_mv_id_f16_f16, has_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_BF16_F32, mul_mv_id_bf16_f32, has_simdgroup_reduction && use_bfloat); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q4_0_F32, mul_mv_id_q4_0_f32, has_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q4_1_F32, mul_mv_id_q4_1_f32, has_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q5_0_F32, mul_mv_id_q5_0_f32, has_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q5_1_F32, mul_mv_id_q5_1_f32, has_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q8_0_F32, mul_mv_id_q8_0_f32, has_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q2_K_F32, mul_mv_id_q2_K_f32, has_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q3_K_F32, mul_mv_id_q3_K_f32, has_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q4_K_F32, mul_mv_id_q4_K_f32, has_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q5_K_F32, mul_mv_id_q5_K_f32, has_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q6_K_F32, mul_mv_id_q6_K_f32, has_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_XXS_F32, mul_mv_id_iq2_xxs_f32, has_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_XS_F32, mul_mv_id_iq2_xs_f32, has_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ3_XXS_F32, mul_mv_id_iq3_xxs_f32, has_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ3_S_F32, mul_mv_id_iq3_s_f32, has_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_S_F32, mul_mv_id_iq2_s_f32, has_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ1_S_F32, mul_mv_id_iq1_s_f32, has_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ1_M_F32, mul_mv_id_iq1_m_f32, has_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ4_NL_F32, mul_mv_id_iq4_nl_f32, has_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ4_XS_F32, mul_mv_id_iq4_xs_f32, has_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_F32_F32, mul_mm_f32_f32, has_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_F16_F32, mul_mm_f16_f32, has_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_BF16_F32, mul_mm_bf16_f32, has_simdgroup_mm && use_bfloat); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_0_F32, mul_mm_q4_0_f32, has_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_1_F32, mul_mm_q4_1_f32, has_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_0_F32, mul_mm_q5_0_f32, has_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_1_F32, mul_mm_q5_1_f32, has_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q8_0_F32, mul_mm_q8_0_f32, has_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q2_K_F32, mul_mm_q2_K_f32, has_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q3_K_F32, mul_mm_q3_K_f32, has_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_K_F32, mul_mm_q4_K_f32, has_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_K_F32, mul_mm_q5_K_f32, has_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q6_K_F32, mul_mm_q6_K_f32, has_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_XXS_F32, mul_mm_iq2_xxs_f32, has_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_XS_F32, mul_mm_iq2_xs_f32, has_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ3_XXS_F32, mul_mm_iq3_xxs_f32, has_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ3_S_F32, mul_mm_iq3_s_f32, has_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_S_F32, mul_mm_iq2_s_f32, has_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ1_S_F32, mul_mm_iq1_s_f32, has_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ1_M_F32, mul_mm_iq1_m_f32, has_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_NL_F32, mul_mm_iq4_nl_f32, has_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_XS_F32, mul_mm_iq4_xs_f32, has_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F32_F32, mul_mm_id_f32_f32, has_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F16_F32, mul_mm_id_f16_f32, has_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_BF16_F32, mul_mm_id_bf16_f32, has_simdgroup_mm && use_bfloat); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_0_F32, mul_mm_id_q4_0_f32, has_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_1_F32, mul_mm_id_q4_1_f32, has_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_0_F32, mul_mm_id_q5_0_f32, has_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_1_F32, mul_mm_id_q5_1_f32, has_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q8_0_F32, mul_mm_id_q8_0_f32, has_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q2_K_F32, mul_mm_id_q2_K_f32, has_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q3_K_F32, mul_mm_id_q3_K_f32, has_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_K_F32, mul_mm_id_q4_K_f32, has_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_K_F32, mul_mm_id_q5_K_f32, has_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q6_K_F32, mul_mm_id_q6_K_f32, has_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XXS_F32, mul_mm_id_iq2_xxs_f32, has_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XS_F32, mul_mm_id_iq2_xs_f32, has_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_XXS_F32, mul_mm_id_iq3_xxs_f32, has_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_S_F32, mul_mm_id_iq3_s_f32, has_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_S_F32, mul_mm_id_iq2_s_f32, has_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ1_S_F32, mul_mm_id_iq1_s_f32, has_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ1_M_F32, mul_mm_id_iq1_m_f32, has_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_NL_F32, mul_mm_id_iq4_nl_f32, has_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_XS_F32, mul_mm_id_iq4_xs_f32, has_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ROPE_NORM_F32, rope_norm_f32, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ROPE_NORM_F16, rope_norm_f16, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ROPE_NEOX_F32, rope_neox_f32, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ROPE_NEOX_F16, rope_neox_f16, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_IM2COL_F16, im2col_f16, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_IM2COL_F32, im2col_f32, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_IM2COL_EXT_F16, im2col_ext_f16, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_IM2COL_EXT_F32, im2col_ext_f32, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CONV_TRANSPOSE_1D_F32_F32, conv_transpose_1d_f32_f32, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CONV_TRANSPOSE_1D_F16_F32, conv_transpose_1d_f16_f32, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_UPSCALE_F32, upscale_f32, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_PAD_F32, pad_f32, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_PAD_REFLECT_1D_F32, pad_reflect_1d_f32, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_TIMESTEP_EMBEDDING_F32, timestep_embedding_f32, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ARANGE_F32, arange_f32, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_ASC, argsort_f32_i32_asc, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_DESC, argsort_f32_i32_desc, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_LEAKY_RELU_F32, leaky_relu_f32, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H64, flash_attn_ext_f16_h64, has_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H80, flash_attn_ext_f16_h80, has_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H96, flash_attn_ext_f16_h96, has_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H112, flash_attn_ext_f16_h112, has_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H128, flash_attn_ext_f16_h128, has_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H256, flash_attn_ext_f16_h256, has_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H64, flash_attn_ext_bf16_h64, has_simdgroup_mm && use_bfloat); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H80, flash_attn_ext_bf16_h80, has_simdgroup_mm && use_bfloat); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H96, flash_attn_ext_bf16_h96, has_simdgroup_mm && use_bfloat); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H112, flash_attn_ext_bf16_h112, has_simdgroup_mm && use_bfloat); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H128, flash_attn_ext_bf16_h128, has_simdgroup_mm && use_bfloat); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H256, flash_attn_ext_bf16_h256, has_simdgroup_mm && use_bfloat); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_H64, flash_attn_ext_q4_0_h64, has_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_H80, flash_attn_ext_q4_0_h80, has_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_H96, flash_attn_ext_q4_0_h96, has_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_H112, flash_attn_ext_q4_0_h112, has_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_H128, flash_attn_ext_q4_0_h128, has_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_H256, flash_attn_ext_q4_0_h256, has_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_H64, flash_attn_ext_q4_1_h64, has_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_H80, flash_attn_ext_q4_1_h80, has_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_H96, flash_attn_ext_q4_1_h96, has_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_H112, flash_attn_ext_q4_1_h112, has_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_H128, flash_attn_ext_q4_1_h128, has_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_H256, flash_attn_ext_q4_1_h256, has_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_H64, flash_attn_ext_q5_0_h64, has_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_H80, flash_attn_ext_q5_0_h80, has_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_H96, flash_attn_ext_q5_0_h96, has_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_H112, flash_attn_ext_q5_0_h112, has_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_H128, flash_attn_ext_q5_0_h128, has_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_H256, flash_attn_ext_q5_0_h256, has_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_H64, flash_attn_ext_q5_1_h64, has_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_H80, flash_attn_ext_q5_1_h80, has_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_H96, flash_attn_ext_q5_1_h96, has_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_H112, flash_attn_ext_q5_1_h112, has_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_H128, flash_attn_ext_q5_1_h128, has_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_H256, flash_attn_ext_q5_1_h256, has_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H64, flash_attn_ext_q8_0_h64, has_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H80, flash_attn_ext_q8_0_h80, has_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H96, flash_attn_ext_q8_0_h96, has_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H112, flash_attn_ext_q8_0_h112, has_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H128, flash_attn_ext_q8_0_h128, has_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H256, flash_attn_ext_q8_0_h256, has_simdgroup_mm); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H128, flash_attn_ext_vec_f16_h128, has_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_BF16_H128, flash_attn_ext_vec_bf16_h128, has_simdgroup_reduction && use_bfloat); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_0_H128, flash_attn_ext_vec_q4_0_h128, has_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_1_H128, flash_attn_ext_vec_q4_1_h128, has_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_0_H128, flash_attn_ext_vec_q5_0_h128, has_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_1_H128, flash_attn_ext_vec_q5_1_h128, has_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q8_0_H128, flash_attn_ext_vec_q8_0_h128, has_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H256, flash_attn_ext_vec_f16_h256, has_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_BF16_H256, flash_attn_ext_vec_bf16_h256, has_simdgroup_reduction && use_bfloat); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_0_H256, flash_attn_ext_vec_q4_0_h256, has_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_1_H256, flash_attn_ext_vec_q4_1_h256, has_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_0_H256, flash_attn_ext_vec_q5_0_h256, has_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_1_H256, flash_attn_ext_vec_q5_1_h256, has_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q8_0_H256, flash_attn_ext_vec_q8_0_h256, has_simdgroup_reduction); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SET_F32, set_f32, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SET_I32, set_i32, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_F32, cpy_f32_f32, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_F16, cpy_f32_f16, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_BF16, cpy_f32_bf16, use_bfloat); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F16_F32, cpy_f16_f32, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F16_F16, cpy_f16_f16, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_BF16_F32, cpy_bf16_f32, use_bfloat); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_BF16_BF16, cpy_bf16_bf16, use_bfloat); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q8_0, cpy_f32_q8_0, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q4_0, cpy_f32_q4_0, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q4_1, cpy_f32_q4_1, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q5_0, cpy_f32_q5_0, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q5_1, cpy_f32_q5_1, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_IQ4_NL, cpy_f32_iq4_nl, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CONCAT, concat, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SQR, sqr, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SQRT, sqrt, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SIN, sin, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_COS, cos, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SUM_ROWS, sum_rows, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ARGMAX, argmax, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_POOL_2D_AVG_F32, pool_2d_avg_f32, true); - GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_POOL_2D_MAX_F32, pool_2d_max_f32, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ADD, add, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ADD_ROW, add_row, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SUB, sub, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SUB_ROW, sub_row, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL, mul, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_ROW, mul_row, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_DIV, div, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_DIV_ROW, div_row, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_REPEAT_F32, repeat_f32, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_REPEAT_F16, repeat_f16, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_REPEAT_I32, repeat_i32, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_REPEAT_I16, repeat_i16, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SCALE, scale, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SCALE_4, scale_4, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CLAMP, clamp, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_TANH, tanh, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_RELU, relu, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SIGMOID, sigmoid, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GELU, gelu, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GELU_4, gelu_4, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GELU_QUICK, gelu_quick, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GELU_QUICK_4, gelu_quick_4, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SILU, silu, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SILU_4, silu_4, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ELU, elu, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SOFT_MAX_F16, soft_max_f16, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SOFT_MAX_F16_4, soft_max_f16_4, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SOFT_MAX_F32, soft_max_f32, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SOFT_MAX_F32_4, soft_max_f32_4, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_DIAG_MASK_INF, diag_mask_inf, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_DIAG_MASK_INF_8, diag_mask_inf_8, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_F32, get_rows_f32, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_F16, get_rows_f16, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_BF16, get_rows_bf16, use_bfloat); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q4_0, get_rows_q4_0, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q4_1, get_rows_q4_1, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q5_0, get_rows_q5_0, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q5_1, get_rows_q5_1, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q8_0, get_rows_q8_0, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q2_K, get_rows_q2_K, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q3_K, get_rows_q3_K, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q4_K, get_rows_q4_K, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q5_K, get_rows_q5_K, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q6_K, get_rows_q6_K, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_XXS, get_rows_iq2_xxs, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_XS, get_rows_iq2_xs, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ3_XXS, get_rows_iq3_xxs, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ3_S, get_rows_iq3_s, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_S, get_rows_iq2_s, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ1_S, get_rows_iq1_s, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ1_M, get_rows_iq1_m, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ4_NL, get_rows_iq4_nl, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ4_XS, get_rows_iq4_xs, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_I32, get_rows_i32, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_RMS_NORM, rms_norm, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_L2_NORM, l2_norm, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GROUP_NORM, group_norm, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_NORM, norm, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SSM_CONV_F32, ssm_conv_f32, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SSM_SCAN_F32, ssm_scan_f32, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_RWKV_WKV6_F32, rwkv_wkv6_f32, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_RWKV_WKV7_F32, rwkv_wkv7_f32, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F32_F32, mul_mv_f32_f32, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_F32, mul_mv_bf16_f32, has_simdgroup_reduction && use_bfloat); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_F32_1ROW, mul_mv_bf16_f32_1row, has_simdgroup_reduction && use_bfloat); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_F32_L4, mul_mv_bf16_f32_l4, has_simdgroup_reduction && use_bfloat); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_BF16, mul_mv_bf16_bf16, has_simdgroup_reduction && use_bfloat); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32, mul_mv_f16_f32, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_1ROW, mul_mv_f16_f32_1row, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_L4, mul_mv_f16_f32_l4, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F16, mul_mv_f16_f16, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_0_F32, mul_mv_q4_0_f32, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_1_F32, mul_mv_q4_1_f32, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_0_F32, mul_mv_q5_0_f32, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_1_F32, mul_mv_q5_1_f32, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q8_0_F32, mul_mv_q8_0_f32, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_F16_F32_R1_2, mul_mv_ext_f16_f32_r1_2, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_F16_F32_R1_3, mul_mv_ext_f16_f32_r1_3, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_F16_F32_R1_4, mul_mv_ext_f16_f32_r1_4, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_F16_F32_R1_5, mul_mv_ext_f16_f32_r1_5, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_0_F32_R1_2, mul_mv_ext_q4_0_f32_r1_2, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_0_F32_R1_3, mul_mv_ext_q4_0_f32_r1_3, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_0_F32_R1_4, mul_mv_ext_q4_0_f32_r1_4, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_0_F32_R1_5, mul_mv_ext_q4_0_f32_r1_5, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_1_F32_R1_2, mul_mv_ext_q4_1_f32_r1_2, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_1_F32_R1_3, mul_mv_ext_q4_1_f32_r1_3, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_1_F32_R1_4, mul_mv_ext_q4_1_f32_r1_4, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_1_F32_R1_5, mul_mv_ext_q4_1_f32_r1_5, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q5_0_F32_R1_2, mul_mv_ext_q5_0_f32_r1_2, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q5_0_F32_R1_3, mul_mv_ext_q5_0_f32_r1_3, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q5_0_F32_R1_4, mul_mv_ext_q5_0_f32_r1_4, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q5_0_F32_R1_5, mul_mv_ext_q5_0_f32_r1_5, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q5_1_F32_R1_2, mul_mv_ext_q5_1_f32_r1_2, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q5_1_F32_R1_3, mul_mv_ext_q5_1_f32_r1_3, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q5_1_F32_R1_4, mul_mv_ext_q5_1_f32_r1_4, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q5_1_F32_R1_5, mul_mv_ext_q5_1_f32_r1_5, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q8_0_F32_R1_2, mul_mv_ext_q8_0_f32_r1_2, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q8_0_F32_R1_3, mul_mv_ext_q8_0_f32_r1_3, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q8_0_F32_R1_4, mul_mv_ext_q8_0_f32_r1_4, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q8_0_F32_R1_5, mul_mv_ext_q8_0_f32_r1_5, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_K_F32_R1_2, mul_mv_ext_q4_K_f32_r1_2, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_K_F32_R1_3, mul_mv_ext_q4_K_f32_r1_3, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_K_F32_R1_4, mul_mv_ext_q4_K_f32_r1_4, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_K_F32_R1_5, mul_mv_ext_q4_K_f32_r1_5, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q5_K_F32_R1_2, mul_mv_ext_q5_K_f32_r1_2, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q5_K_F32_R1_3, mul_mv_ext_q5_K_f32_r1_3, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q5_K_F32_R1_4, mul_mv_ext_q5_K_f32_r1_4, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q5_K_F32_R1_5, mul_mv_ext_q5_K_f32_r1_5, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q6_K_F32_R1_2, mul_mv_ext_q6_K_f32_r1_2, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q6_K_F32_R1_3, mul_mv_ext_q6_K_f32_r1_3, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q6_K_F32_R1_4, mul_mv_ext_q6_K_f32_r1_4, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q6_K_F32_R1_5, mul_mv_ext_q6_K_f32_r1_5, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_IQ4_NL_F32_R1_2, mul_mv_ext_iq4_nl_f32_r1_2, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_IQ4_NL_F32_R1_3, mul_mv_ext_iq4_nl_f32_r1_3, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_IQ4_NL_F32_R1_4, mul_mv_ext_iq4_nl_f32_r1_4, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_IQ4_NL_F32_R1_5, mul_mv_ext_iq4_nl_f32_r1_5, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q2_K_F32, mul_mv_q2_K_f32, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q3_K_F32, mul_mv_q3_K_f32, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_K_F32, mul_mv_q4_K_f32, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_K_F32, mul_mv_q5_K_f32, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q6_K_F32, mul_mv_q6_K_f32, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_XXS_F32, mul_mv_iq2_xxs_f32, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_XS_F32, mul_mv_iq2_xs_f32, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ3_XXS_F32, mul_mv_iq3_xxs_f32, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ3_S_F32, mul_mv_iq3_s_f32, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_S_F32, mul_mv_iq2_s_f32, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ1_S_F32, mul_mv_iq1_s_f32, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ1_M_F32, mul_mv_iq1_m_f32, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ4_NL_F32, mul_mv_iq4_nl_f32, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ4_XS_F32, mul_mv_iq4_xs_f32, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F32_F32, mul_mv_id_f32_f32, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F32, mul_mv_id_f16_f32, has_simdgroup_reduction); + //GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F32_1ROW, mul_mv_id_f16_f32_1row, has_simdgroup_reduction); + //GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F32_L4, mul_mv_id_f16_f32_l4, has_simdgroup_reduction); + //GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F16, mul_mv_id_f16_f16, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_BF16_F32, mul_mv_id_bf16_f32, has_simdgroup_reduction && use_bfloat); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q4_0_F32, mul_mv_id_q4_0_f32, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q4_1_F32, mul_mv_id_q4_1_f32, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q5_0_F32, mul_mv_id_q5_0_f32, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q5_1_F32, mul_mv_id_q5_1_f32, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q8_0_F32, mul_mv_id_q8_0_f32, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q2_K_F32, mul_mv_id_q2_K_f32, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q3_K_F32, mul_mv_id_q3_K_f32, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q4_K_F32, mul_mv_id_q4_K_f32, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q5_K_F32, mul_mv_id_q5_K_f32, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q6_K_F32, mul_mv_id_q6_K_f32, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_XXS_F32, mul_mv_id_iq2_xxs_f32, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_XS_F32, mul_mv_id_iq2_xs_f32, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ3_XXS_F32, mul_mv_id_iq3_xxs_f32, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ3_S_F32, mul_mv_id_iq3_s_f32, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_S_F32, mul_mv_id_iq2_s_f32, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ1_S_F32, mul_mv_id_iq1_s_f32, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ1_M_F32, mul_mv_id_iq1_m_f32, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ4_NL_F32, mul_mv_id_iq4_nl_f32, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ4_XS_F32, mul_mv_id_iq4_xs_f32, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_F32_F32, mul_mm_f32_f32, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_F16_F32, mul_mm_f16_f32, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_BF16_F32, mul_mm_bf16_f32, has_simdgroup_mm && use_bfloat); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_0_F32, mul_mm_q4_0_f32, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_1_F32, mul_mm_q4_1_f32, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_0_F32, mul_mm_q5_0_f32, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_1_F32, mul_mm_q5_1_f32, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q8_0_F32, mul_mm_q8_0_f32, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q2_K_F32, mul_mm_q2_K_f32, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q3_K_F32, mul_mm_q3_K_f32, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_K_F32, mul_mm_q4_K_f32, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_K_F32, mul_mm_q5_K_f32, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q6_K_F32, mul_mm_q6_K_f32, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_XXS_F32, mul_mm_iq2_xxs_f32, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_XS_F32, mul_mm_iq2_xs_f32, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ3_XXS_F32, mul_mm_iq3_xxs_f32, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ3_S_F32, mul_mm_iq3_s_f32, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_S_F32, mul_mm_iq2_s_f32, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ1_S_F32, mul_mm_iq1_s_f32, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ1_M_F32, mul_mm_iq1_m_f32, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_NL_F32, mul_mm_iq4_nl_f32, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_XS_F32, mul_mm_iq4_xs_f32, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F32_F32, mul_mm_id_f32_f32, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F16_F32, mul_mm_id_f16_f32, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_BF16_F32, mul_mm_id_bf16_f32, has_simdgroup_mm && use_bfloat); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_0_F32, mul_mm_id_q4_0_f32, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_1_F32, mul_mm_id_q4_1_f32, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_0_F32, mul_mm_id_q5_0_f32, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_1_F32, mul_mm_id_q5_1_f32, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q8_0_F32, mul_mm_id_q8_0_f32, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q2_K_F32, mul_mm_id_q2_K_f32, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q3_K_F32, mul_mm_id_q3_K_f32, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_K_F32, mul_mm_id_q4_K_f32, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_K_F32, mul_mm_id_q5_K_f32, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q6_K_F32, mul_mm_id_q6_K_f32, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XXS_F32, mul_mm_id_iq2_xxs_f32, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XS_F32, mul_mm_id_iq2_xs_f32, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_XXS_F32, mul_mm_id_iq3_xxs_f32, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_S_F32, mul_mm_id_iq3_s_f32, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_S_F32, mul_mm_id_iq2_s_f32, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ1_S_F32, mul_mm_id_iq1_s_f32, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ1_M_F32, mul_mm_id_iq1_m_f32, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_NL_F32, mul_mm_id_iq4_nl_f32, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_XS_F32, mul_mm_id_iq4_xs_f32, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ROPE_NORM_F32, rope_norm_f32, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ROPE_NORM_F16, rope_norm_f16, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ROPE_NEOX_F32, rope_neox_f32, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ROPE_NEOX_F16, rope_neox_f16, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_IM2COL_F16, im2col_f16, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_IM2COL_F32, im2col_f32, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_IM2COL_EXT_F16, im2col_ext_f16, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_IM2COL_EXT_F32, im2col_ext_f32, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CONV_TRANSPOSE_1D_F32_F32, conv_transpose_1d_f32_f32, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CONV_TRANSPOSE_1D_F16_F32, conv_transpose_1d_f16_f32, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_UPSCALE_F32, upscale_f32, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_PAD_F32, pad_f32, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_PAD_REFLECT_1D_F32, pad_reflect_1d_f32, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_TIMESTEP_EMBEDDING_F32, timestep_embedding_f32, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ARANGE_F32, arange_f32, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_ASC, argsort_f32_i32_asc, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_DESC, argsort_f32_i32_desc, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_LEAKY_RELU_F32, leaky_relu_f32, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H64, flash_attn_ext_f16_h64, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H80, flash_attn_ext_f16_h80, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H96, flash_attn_ext_f16_h96, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H112, flash_attn_ext_f16_h112, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H128, flash_attn_ext_f16_h128, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H192, flash_attn_ext_f16_h192, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_HK192_HV128, flash_attn_ext_f16_hk192_hv128, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H256, flash_attn_ext_f16_h256, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H64, flash_attn_ext_bf16_h64, has_simdgroup_mm && use_bfloat); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H80, flash_attn_ext_bf16_h80, has_simdgroup_mm && use_bfloat); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H96, flash_attn_ext_bf16_h96, has_simdgroup_mm && use_bfloat); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H112, flash_attn_ext_bf16_h112, has_simdgroup_mm && use_bfloat); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H128, flash_attn_ext_bf16_h128, has_simdgroup_mm && use_bfloat); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H192, flash_attn_ext_bf16_h192, has_simdgroup_mm && use_bfloat); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_HK192_HV128, flash_attn_ext_bf16_hk192_hv128, has_simdgroup_mm && use_bfloat); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H256, flash_attn_ext_bf16_h256, has_simdgroup_mm && use_bfloat); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_H64, flash_attn_ext_q4_0_h64, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_H80, flash_attn_ext_q4_0_h80, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_H96, flash_attn_ext_q4_0_h96, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_H112, flash_attn_ext_q4_0_h112, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_H128, flash_attn_ext_q4_0_h128, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_H192, flash_attn_ext_q4_0_h192, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_HK192_HV128, flash_attn_ext_q4_0_hk192_hv128, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_H256, flash_attn_ext_q4_0_h256, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_H64, flash_attn_ext_q4_1_h64, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_H80, flash_attn_ext_q4_1_h80, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_H96, flash_attn_ext_q4_1_h96, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_H112, flash_attn_ext_q4_1_h112, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_H128, flash_attn_ext_q4_1_h128, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_H192, flash_attn_ext_q4_1_h192, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_HK192_HV128, flash_attn_ext_q4_1_hk192_hv128, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_H256, flash_attn_ext_q4_1_h256, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_H64, flash_attn_ext_q5_0_h64, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_H80, flash_attn_ext_q5_0_h80, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_H96, flash_attn_ext_q5_0_h96, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_H112, flash_attn_ext_q5_0_h112, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_H128, flash_attn_ext_q5_0_h128, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_H192, flash_attn_ext_q5_0_h192, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_HK192_HV128, flash_attn_ext_q5_0_hk192_hv128, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_H256, flash_attn_ext_q5_0_h256, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_H64, flash_attn_ext_q5_1_h64, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_H80, flash_attn_ext_q5_1_h80, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_H96, flash_attn_ext_q5_1_h96, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_H112, flash_attn_ext_q5_1_h112, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_H128, flash_attn_ext_q5_1_h128, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_H192, flash_attn_ext_q5_1_h192, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_HK192_HV128, flash_attn_ext_q5_1_hk192_hv128, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_H256, flash_attn_ext_q5_1_h256, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H64, flash_attn_ext_q8_0_h64, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H80, flash_attn_ext_q8_0_h80, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H96, flash_attn_ext_q8_0_h96, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H112, flash_attn_ext_q8_0_h112, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H128, flash_attn_ext_q8_0_h128, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H192, flash_attn_ext_q8_0_h192, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_HK192_HV128, flash_attn_ext_q8_0_hk192_hv128, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H256, flash_attn_ext_q8_0_h256, has_simdgroup_mm); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H128, flash_attn_ext_vec_f16_h128, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_BF16_H128, flash_attn_ext_vec_bf16_h128, has_simdgroup_reduction && use_bfloat); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_0_H128, flash_attn_ext_vec_q4_0_h128, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_1_H128, flash_attn_ext_vec_q4_1_h128, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_0_H128, flash_attn_ext_vec_q5_0_h128, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_1_H128, flash_attn_ext_vec_q5_1_h128, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q8_0_H128, flash_attn_ext_vec_q8_0_h128, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H192, flash_attn_ext_vec_f16_h192, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_BF16_H192, flash_attn_ext_vec_bf16_h192, has_simdgroup_reduction && use_bfloat); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_0_H192, flash_attn_ext_vec_q4_0_h192, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_1_H192, flash_attn_ext_vec_q4_1_h192, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_0_H192, flash_attn_ext_vec_q5_0_h192, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_1_H192, flash_attn_ext_vec_q5_1_h192, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q8_0_H192, flash_attn_ext_vec_q8_0_h192, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_HK192_HV128, flash_attn_ext_vec_f16_hk192_hv128, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_BF16_HK192_HV128, flash_attn_ext_vec_bf16_hk192_hv128, has_simdgroup_reduction && use_bfloat); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_0_HK192_HV128, flash_attn_ext_vec_q4_0_hk192_hv128, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_1_HK192_HV128, flash_attn_ext_vec_q4_1_hk192_hv128, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_0_HK192_HV128, flash_attn_ext_vec_q5_0_hk192_hv128, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_1_HK192_HV128, flash_attn_ext_vec_q5_1_hk192_hv128, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q8_0_HK192_HV128, flash_attn_ext_vec_q8_0_hk192_hv128, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H256, flash_attn_ext_vec_f16_h256, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_BF16_H256, flash_attn_ext_vec_bf16_h256, has_simdgroup_reduction && use_bfloat); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_0_H256, flash_attn_ext_vec_q4_0_h256, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_1_H256, flash_attn_ext_vec_q4_1_h256, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_0_H256, flash_attn_ext_vec_q5_0_h256, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_1_H256, flash_attn_ext_vec_q5_1_h256, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q8_0_H256, flash_attn_ext_vec_q8_0_h256, has_simdgroup_reduction); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SET_F32, set_f32, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SET_I32, set_i32, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_F32, cpy_f32_f32, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_F16, cpy_f32_f16, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_BF16, cpy_f32_bf16, use_bfloat); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F16_F32, cpy_f16_f32, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F16_F16, cpy_f16_f16, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_BF16_F32, cpy_bf16_f32, use_bfloat); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_BF16_BF16, cpy_bf16_bf16, use_bfloat); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q8_0, cpy_f32_q8_0, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q4_0, cpy_f32_q4_0, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q4_1, cpy_f32_q4_1, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q5_0, cpy_f32_q5_0, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q5_1, cpy_f32_q5_1, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_IQ4_NL, cpy_f32_iq4_nl, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_Q4_0_F32, cpy_q4_0_f32, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_Q4_0_F16, cpy_q4_0_f16, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_Q4_1_F32, cpy_q4_1_f32, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_Q4_1_F16, cpy_q4_1_f16, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_Q5_0_F32, cpy_q5_0_f32, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_Q5_0_F16, cpy_q5_0_f16, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_Q5_1_F32, cpy_q5_1_f32, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_Q5_1_F16, cpy_q5_1_f16, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_Q8_0_F32, cpy_q8_0_f32, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_Q8_0_F16, cpy_q8_0_f16, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CONCAT, concat, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SQR, sqr, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SQRT, sqrt, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SIN, sin, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_COS, cos, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SUM_ROWS, sum_rows, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ARGMAX, argmax, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_POOL_2D_AVG_F32, pool_2d_avg_f32, true); + GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_POOL_2D_MAX_F32, pool_2d_max_f32, true); } - [metal_library release]; - return ctx; } @@ -1180,7 +1278,7 @@ static bool ggml_metal_supports_op(const struct ggml_backend_metal_device_contex case GGML_UNARY_OP_GELU_QUICK: case GGML_UNARY_OP_SILU: case GGML_UNARY_OP_ELU: - return ggml_is_contiguous(op->src[0]); + return ggml_is_contiguous(op->src[0]) && op->src[0]->type == GGML_TYPE_F32; default: return false; } @@ -1190,26 +1288,32 @@ static bool ggml_metal_supports_op(const struct ggml_backend_metal_device_contex case GGML_OP_TRANSPOSE: case GGML_OP_PERMUTE: case GGML_OP_CONCAT: + return true; case GGML_OP_ADD: case GGML_OP_SUB: - case GGML_OP_ACC: case GGML_OP_MUL: case GGML_OP_DIV: + return op->src[0]->type == GGML_TYPE_F32; + case GGML_OP_ACC: case GGML_OP_REPEAT: case GGML_OP_SCALE: - case GGML_OP_CLAMP: case GGML_OP_CONV_TRANSPOSE_1D: return true; + case GGML_OP_CLAMP: + return op->src[0]->type == GGML_TYPE_F32; case GGML_OP_SQR: case GGML_OP_SQRT: case GGML_OP_SIN: case GGML_OP_COS: - return ggml_is_contiguous(op->src[0]); + return ggml_is_contiguous(op->src[0]) && op->src[0]->type == GGML_TYPE_F32; + case GGML_OP_LOG: + return false; // TODO: implement case GGML_OP_SUM_ROWS: case GGML_OP_SOFT_MAX: case GGML_OP_GROUP_NORM: return has_simdgroup_reduction && ggml_is_contiguous(op->src[0]); case GGML_OP_RMS_NORM: + case GGML_OP_L2_NORM: return has_simdgroup_reduction && (op->ne[0] % 4 == 0 && ggml_is_contiguous_1(op->src[0])); case GGML_OP_ARGMAX: return true; @@ -1230,22 +1334,31 @@ static bool ggml_metal_supports_op(const struct ggml_backend_metal_device_contex return op->src[0]->type == GGML_TYPE_F16; case GGML_OP_POOL_1D: return false; - case GGML_OP_POOL_2D: case GGML_OP_UPSCALE: + return op->src[0]->type == GGML_TYPE_F32 && op->op_params[0] == GGML_SCALE_MODE_NEAREST; + case GGML_OP_POOL_2D: case GGML_OP_PAD: case GGML_OP_PAD_REFLECT_1D: - case GGML_OP_ARANGE: case GGML_OP_TIMESTEP_EMBEDDING: case GGML_OP_ARGSORT: case GGML_OP_LEAKY_RELU: + return op->src[0]->type == GGML_TYPE_F32; + case GGML_OP_ARANGE: return true; case GGML_OP_FLASH_ATTN_EXT: + if (op->src[0]->ne[0] == 32) { + // head size == 32 (e.g. bert-bge-small) + // TODO: not sure if it is worth adding kernels for this size + return false; + } if (op->src[1]->type != op->src[2]->type) { return false; } return has_simdgroup_mm; // TODO: over-restricted for vec-kernels case GGML_OP_SSM_CONV: case GGML_OP_SSM_SCAN: + case GGML_OP_RWKV_WKV6: + case GGML_OP_RWKV_WKV7: return true; case GGML_OP_MUL_MAT: case GGML_OP_MUL_MAT_ID: @@ -1287,6 +1400,18 @@ static bool ggml_metal_supports_op(const struct ggml_backend_metal_device_contex default: return false; } + case GGML_TYPE_Q4_0: + case GGML_TYPE_Q4_1: + case GGML_TYPE_Q5_0: + case GGML_TYPE_Q5_1: + case GGML_TYPE_Q8_0: + switch (op->type) { + case GGML_TYPE_F32: + case GGML_TYPE_F16: + return true; + default: + return false; + } default: return false; }; @@ -1906,34 +2031,38 @@ static void ggml_metal_encode_node( id pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SUM_ROWS].pipeline; - // TODO: add ggml_metal_kargs struct + + ggml_metal_kargs_sum_rows args = { + /*.ne00 =*/ ne00, + /*.ne01 =*/ ne01, + /*.ne02 =*/ ne02, + /*.ne03 =*/ ne03, + /*.nb00 =*/ nb00, + /*.nb01 =*/ nb01, + /*.nb02 =*/ nb02, + /*.nb03 =*/ nb03, + /*.ne10 =*/ ne10, + /*.ne11 =*/ ne11, + /*.ne12 =*/ ne12, + /*.ne13 =*/ ne13, + /*.nb10 =*/ nb10, + /*.nb11 =*/ nb11, + /*.nb12 =*/ nb12, + /*.nb13 =*/ nb13, + /*.ne0 =*/ ne0, + /*.ne1 =*/ ne1, + /*.ne2 =*/ ne2, + /*.ne3 =*/ ne3, + /*.nb0 =*/ nb0, + /*.nb1 =*/ nb1, + /*.nb2 =*/ nb2, + /*.nb3 =*/ nb3, + }; + [encoder setComputePipelineState:pipeline]; [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; [encoder setBuffer:id_dst offset:offs_dst atIndex:1]; - [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:2]; - [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:3]; - [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:4]; - [encoder setBytes:&ne03 length:sizeof(ne03) atIndex:5]; - [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:6]; - [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:7]; - [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:8]; - [encoder setBytes:&nb03 length:sizeof(nb03) atIndex:9]; - [encoder setBytes:&ne10 length:sizeof(ne10) atIndex:10]; - [encoder setBytes:&ne11 length:sizeof(ne11) atIndex:11]; - [encoder setBytes:&ne12 length:sizeof(ne12) atIndex:12]; - [encoder setBytes:&ne13 length:sizeof(ne13) atIndex:13]; - [encoder setBytes:&nb10 length:sizeof(nb10) atIndex:14]; - [encoder setBytes:&nb11 length:sizeof(nb11) atIndex:15]; - [encoder setBytes:&nb12 length:sizeof(nb12) atIndex:16]; - [encoder setBytes:&nb13 length:sizeof(nb13) atIndex:17]; - [encoder setBytes:&ne0 length:sizeof(ne0) atIndex:18]; - [encoder setBytes:&ne1 length:sizeof(ne1) atIndex:19]; - [encoder setBytes:&ne2 length:sizeof(ne2) atIndex:20]; - [encoder setBytes:&ne3 length:sizeof(ne3) atIndex:21]; - [encoder setBytes:&nb0 length:sizeof(nb0) atIndex:22]; - [encoder setBytes:&nb1 length:sizeof(nb1) atIndex:23]; - [encoder setBytes:&nb2 length:sizeof(nb2) atIndex:24]; - [encoder setBytes:&nb3 length:sizeof(nb3) atIndex:25]; + [encoder setBytes:&args length:sizeof(args) atIndex:2]; [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)]; } break; @@ -1982,8 +2111,17 @@ static void ggml_metal_encode_node( const float m0 = powf(2.0f, -(max_bias ) / n_head_log2); const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2); - // TODO: add ggml_metal_kargs struct - // TODO: optimize (see https://github.com/ggerganov/llama.cpp/pull/10238/commits/7941b6b9ec29a2866fec6fa6c51612515ca509f6) + ggml_metal_kargs_soft_max args = { + /*.ne00 =*/ ne00, + /*.ne01 =*/ ne01, + /*.ne02 =*/ ne02, + /*.scale =*/ scale, + /*.max_bias =*/ max_bias, + /*.m0 =*/ m0, + /*.m1 =*/ m1, + /*.n_head_log2 =*/ n_head_log2, + }; + [encoder setComputePipelineState:pipeline]; [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; if (id_src1) { @@ -1992,14 +2130,7 @@ static void ggml_metal_encode_node( [encoder setBuffer:id_src0 offset:offs_src0 atIndex:1]; } [encoder setBuffer:id_dst offset:offs_dst atIndex:2]; - [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:3]; - [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:4]; - [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:5]; - [encoder setBytes:&scale length:sizeof(scale) atIndex:6]; - [encoder setBytes:&max_bias length:sizeof(max_bias) atIndex:7]; - [encoder setBytes:&m0 length:sizeof(m0) atIndex:8]; - [encoder setBytes:&m1 length:sizeof(m1) atIndex:9]; - [encoder setBytes:&n_head_log2 length:sizeof(n_head_log2) atIndex:10]; + [encoder setBytes:&args length:sizeof(args) atIndex:3]; [encoder setThreadgroupMemoryLength:32*sizeof(float) atIndex:0]; @@ -2017,13 +2148,16 @@ static void ggml_metal_encode_node( pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_DIAG_MASK_INF].pipeline; } - // TODO: add ggml_metal_kargs struct + ggml_metal_kargs_diag_mask_inf args = { + /*.ne00 =*/ ne00, + /*.ne01 =*/ ne01, + /*.n_past =*/ n_past, + }; + [encoder setComputePipelineState:pipeline]; [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; [encoder setBuffer:id_dst offset:offs_dst atIndex:1]; - [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:2]; - [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:3]; - [encoder setBytes:&n_past length:sizeof(int) atIndex:4]; + [encoder setBytes:&args length:sizeof(args) atIndex:2]; if (ne00%8 == 0) { [encoder dispatchThreadgroups:MTLSizeMake(ne00*ne01*ne02/8, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)]; @@ -2042,27 +2176,30 @@ static void ggml_metal_encode_node( id pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SSM_CONV_F32].pipeline; - // TODO: add ggml_metal_kargs struct + ggml_metal_kargs_ssm_conv args = { + /*.ne00 =*/ ne00, + /*.ne01 =*/ ne01, + /*.ne02 =*/ ne02, + /*.nb00 =*/ nb00, + /*.nb01 =*/ nb01, + /*.nb02 =*/ nb02, + /*.ne10 =*/ ne10, + /*.ne11 =*/ ne11, + /*.nb10 =*/ nb10, + /*.nb11 =*/ nb11, + /*.ne0 =*/ ne0, + /*.ne1 =*/ ne1, + /*.ne2 =*/ ne2, + /*.nb0 =*/ nb0, + /*.nb1 =*/ nb1, + /*.nb2 =*/ nb2, + }; + [encoder setComputePipelineState:pipeline]; [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1]; [encoder setBuffer:id_dst offset:offs_dst atIndex:2]; - [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:3]; - [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:4]; - [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:5]; - [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:6]; - [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:7]; - [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:8]; - [encoder setBytes:&ne10 length:sizeof(ne10) atIndex:9]; - [encoder setBytes:&ne11 length:sizeof(ne11) atIndex:10]; - [encoder setBytes:&nb10 length:sizeof(nb10) atIndex:11]; - [encoder setBytes:&nb11 length:sizeof(nb11) atIndex:12]; - [encoder setBytes:&ne0 length:sizeof(ne0) atIndex:13]; - [encoder setBytes:&ne1 length:sizeof(ne1) atIndex:14]; - [encoder setBytes:&ne2 length:sizeof(ne2) atIndex:15]; - [encoder setBytes:&nb0 length:sizeof(nb0) atIndex:16]; - [encoder setBytes:&nb1 length:sizeof(nb1) atIndex:17]; - [encoder setBytes:&nb2 length:sizeof(nb2) atIndex:18]; + [encoder setBytes:&args length:sizeof(args) atIndex:3]; [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne1, ne02) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)]; } break; @@ -2113,7 +2250,64 @@ static void ggml_metal_encode_node( id pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SSM_SCAN_F32].pipeline; - // TODO: add ggml_metal_kargs struct + ggml_metal_kargs_ssm_scan args = { + /*.d_state =*/ d_state, + /*.d_inner =*/ d_inner, + /*.n_seq_tokens =*/ n_seq_tokens, + /*.n_seqs =*/ n_seqs, + /*.nb00 =*/ nb00, + /*.nb01 =*/ nb01, + /*.nb02 =*/ nb02, + /*.nb10 =*/ nb10, + /*.nb11 =*/ nb11, + /*.nb12 =*/ nb12, + /*.nb13 =*/ nb13, + /*.nb20 =*/ nb20, + /*.nb21 =*/ nb21, + /*.nb22 =*/ nb22, + /*.nb30 =*/ nb30, + /*.nb31 =*/ nb31, + /*.nb40 =*/ nb40, + /*.nb41 =*/ nb41, + /*.nb42 =*/ nb42, + /*.nb50 =*/ nb50, + /*.nb51 =*/ nb51, + /*.nb52 =*/ nb52, + }; + + [encoder setComputePipelineState:pipeline]; + [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; + [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1]; + [encoder setBuffer:id_src2 offset:offs_src2 atIndex:2]; + [encoder setBuffer:id_src3 offset:offs_src3 atIndex:3]; + [encoder setBuffer:id_src4 offset:offs_src4 atIndex:4]; + [encoder setBuffer:id_src5 offset:offs_src5 atIndex:5]; + [encoder setBuffer:id_dst offset:offs_dst atIndex:6]; + [encoder setBytes:&args length:sizeof(args) atIndex:7]; + + [encoder dispatchThreadgroups:MTLSizeMake(d_inner, n_seqs, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)]; + } break; + case GGML_OP_RWKV_WKV6: + { + const int64_t B = dst->src[5]->ne[1]; + const int64_t T = dst->src[0]->ne[2]; + const int64_t C = dst->ne[0]; + const int64_t H = dst->src[0]->ne[1]; + + GGML_ASSERT(dst->src[5]->type == GGML_TYPE_F32); + GGML_ASSERT(C % H == 0); + GGML_ASSERT(C / H == 64); + + size_t offs_src3 = 0; + size_t offs_src4 = 0; + size_t offs_src5 = 0; + + id id_src3 = dst->src[3] ? ggml_metal_get_buffer(dst->src[3], &offs_src3) : nil; + id id_src4 = dst->src[4] ? ggml_metal_get_buffer(dst->src[4], &offs_src4) : nil; + id id_src5 = dst->src[5] ? ggml_metal_get_buffer(dst->src[5], &offs_src5) : nil; + + id pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_RWKV_WKV6_F32].pipeline; + [encoder setComputePipelineState:pipeline]; [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1]; @@ -2123,31 +2317,52 @@ static void ggml_metal_encode_node( [encoder setBuffer:id_src5 offset:offs_src5 atIndex:5]; [encoder setBuffer:id_dst offset:offs_dst atIndex:6]; - [encoder setBytes:&d_state length:sizeof(d_state) atIndex:7]; - [encoder setBytes:&d_inner length:sizeof(d_inner) atIndex:8]; - [encoder setBytes:&n_seq_tokens length:sizeof(n_seq_tokens) atIndex:9]; - [encoder setBytes:&n_seqs length:sizeof(n_seqs) atIndex:10]; + [encoder setBytes:&B length:sizeof(B) atIndex:7]; + [encoder setBytes:&T length:sizeof(T) atIndex:8]; + [encoder setBytes:&C length:sizeof(C) atIndex:9]; + [encoder setBytes:&H length:sizeof(H) atIndex:10]; - [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:11]; - [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:12]; - [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:13]; - [encoder setBytes:&nb10 length:sizeof(nb10) atIndex:14]; - [encoder setBytes:&nb11 length:sizeof(nb11) atIndex:15]; - [encoder setBytes:&nb12 length:sizeof(nb12) atIndex:16]; - [encoder setBytes:&nb13 length:sizeof(nb13) atIndex:17]; - [encoder setBytes:&nb20 length:sizeof(nb20) atIndex:18]; - [encoder setBytes:&nb21 length:sizeof(nb21) atIndex:19]; - [encoder setBytes:&nb22 length:sizeof(nb22) atIndex:20]; - [encoder setBytes:&nb30 length:sizeof(nb30) atIndex:21]; - [encoder setBytes:&nb31 length:sizeof(nb31) atIndex:22]; - [encoder setBytes:&nb40 length:sizeof(nb40) atIndex:23]; - [encoder setBytes:&nb41 length:sizeof(nb41) atIndex:24]; - [encoder setBytes:&nb42 length:sizeof(nb42) atIndex:25]; - [encoder setBytes:&nb50 length:sizeof(nb50) atIndex:26]; - [encoder setBytes:&nb51 length:sizeof(nb51) atIndex:27]; - [encoder setBytes:&nb52 length:sizeof(nb52) atIndex:28]; + [encoder dispatchThreadgroups:MTLSizeMake(B * H, 1, 1) threadsPerThreadgroup:MTLSizeMake(C/ H, 1, 1)]; + } break; + case GGML_OP_RWKV_WKV7: + { + const int64_t B = dst->src[6]->ne[1]; + const int64_t T = dst->src[0]->ne[2]; + const int64_t C = dst->ne[0]; + const int64_t H = dst->src[0]->ne[1]; - [encoder dispatchThreadgroups:MTLSizeMake(d_inner, n_seqs, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)]; + GGML_ASSERT(dst->src[6]->type == GGML_TYPE_F32); + GGML_ASSERT(C % H == 0); + GGML_ASSERT(C / H == 64); + + size_t offs_src3 = 0; + size_t offs_src4 = 0; + size_t offs_src5 = 0; + size_t offs_src6 = 0; + + id id_src3 = dst->src[3] ? ggml_metal_get_buffer(dst->src[3], &offs_src3) : nil; + id id_src4 = dst->src[4] ? ggml_metal_get_buffer(dst->src[4], &offs_src4) : nil; + id id_src5 = dst->src[5] ? ggml_metal_get_buffer(dst->src[5], &offs_src5) : nil; + id id_src6 = dst->src[6] ? ggml_metal_get_buffer(dst->src[6], &offs_src6) : nil; + + id pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_RWKV_WKV7_F32].pipeline; + + [encoder setComputePipelineState:pipeline]; + [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; + [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1]; + [encoder setBuffer:id_src2 offset:offs_src2 atIndex:2]; + [encoder setBuffer:id_src3 offset:offs_src3 atIndex:3]; + [encoder setBuffer:id_src4 offset:offs_src4 atIndex:4]; + [encoder setBuffer:id_src5 offset:offs_src5 atIndex:5]; + [encoder setBuffer:id_src6 offset:offs_src6 atIndex:6]; + [encoder setBuffer:id_dst offset:offs_dst atIndex:7]; + + [encoder setBytes:&B length:sizeof(B) atIndex:8]; + [encoder setBytes:&T length:sizeof(T) atIndex:9]; + [encoder setBytes:&C length:sizeof(C) atIndex:10]; + [encoder setBytes:&H length:sizeof(H) atIndex:11]; + + [encoder dispatchThreadgroups:MTLSizeMake(B * H, 1, 1) threadsPerThreadgroup:MTLSizeMake(C/ H, 1, 1)]; } break; case GGML_OP_MUL_MAT: { @@ -2408,171 +2623,180 @@ static void ggml_metal_encode_node( [encoder setThreadgroupMemoryLength:8192 atIndex:0]; [encoder dispatchThreadgroups:MTLSizeMake( (ne11 + 31)/32, (ne01 + 63)/64, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(128, 1, 1)]; } else { - int nth0 = 32; - int nth1 = 1; - int nrows = 1; - //printf("vector: ne00 = %6d, ne01 = %6d, ne02 = %6d, ne11 = %6d, ne12 = %6d\n", ne00, ne01, ne02, ne11, ne12); - id pipeline = nil; + int nsg = 0; // number of simdgroups + int nr0 = 0; // number of src0 rows per simdgroup + int nr1 = 1; // number of src1 rows per threadgroup + + size_t smem = 0; // shared memory + // use custom matrix x vector kernel switch (src0t) { case GGML_TYPE_F32: { GGML_ASSERT(src1t == GGML_TYPE_F32); + nsg = 1; + nr0 = 1; + nr1 = 4; pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_F32_F32].pipeline; - nrows = 4; } break; case GGML_TYPE_F16: { - nth0 = 32; - nth1 = 1; + nsg = 1; + nr0 = 1; if (src1t == GGML_TYPE_F32) { if (ne11 * ne12 < 4) { pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_1ROW].pipeline; } else if (ne00 >= 128 && ne01 >= 8 && ne00%4 == 0) { pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_L4].pipeline; - nrows = ne11; + nr1 = ne11; } else { pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32].pipeline; - nrows = 4; + nr1 = 4; } } else { pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F16].pipeline; - nrows = 4; + nr1 = 4; } } break; case GGML_TYPE_BF16: { - nth0 = 32; - nth1 = 1; + nsg = 1; + nr0 = 1; if (src1t == GGML_TYPE_F32) { if (ne11 * ne12 < 4) { pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_F32_1ROW].pipeline; } else if (ne00 >= 128 && ne01 >= 8 && ne00%4 == 0) { pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_F32_L4].pipeline; - nrows = ne11; + nr1 = ne11; } else { pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_F32].pipeline; - nrows = 4; + nr1 = 4; } } else { pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_BF16].pipeline; - nrows = 4; + nr1 = 4; } } break; case GGML_TYPE_Q4_0: { - nth0 = 8; - nth1 = 8; + nsg = N_SG_Q4_0; + nr0 = N_R0_Q4_0; pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_0_F32].pipeline; } break; case GGML_TYPE_Q4_1: { - nth0 = 8; - nth1 = 8; + nsg = N_SG_Q4_1; + nr0 = N_R0_Q4_1; pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_1_F32].pipeline; } break; case GGML_TYPE_Q5_0: { - nth0 = 8; - nth1 = 8; + nsg = N_SG_Q5_0; + nr0 = N_R0_Q5_0; pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_0_F32].pipeline; } break; case GGML_TYPE_Q5_1: { - nth0 = 8; - nth1 = 8; + nsg = N_SG_Q5_1; + nr0 = N_R0_Q5_1; pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_1_F32].pipeline; } break; case GGML_TYPE_Q8_0: { - nth0 = 8; - nth1 = 8; + nsg = N_SG_Q8_0; + nr0 = N_R0_Q8_0; pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q8_0_F32].pipeline; } break; case GGML_TYPE_Q2_K: { - nth0 = 2; - nth1 = 32; + nsg = N_SG_Q2_K; + nr0 = N_R0_Q2_K; pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q2_K_F32].pipeline; } break; case GGML_TYPE_Q3_K: { - nth0 = 2; - nth1 = 32; + nsg = N_SG_Q3_K; + nr0 = N_R0_Q3_K; pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q3_K_F32].pipeline; } break; case GGML_TYPE_Q4_K: { - nth0 = 4; //1; - nth1 = 8; //32; + nsg = N_SG_Q4_K; + nr0 = N_R0_Q4_K; pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_K_F32].pipeline; } break; case GGML_TYPE_Q5_K: { - nth0 = 2; - nth1 = 32; + nsg = N_SG_Q5_K; + nr0 = N_R0_Q5_K; pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_K_F32].pipeline; } break; case GGML_TYPE_Q6_K: { - nth0 = 2; - nth1 = 32; + nsg = N_SG_Q6_K; + nr0 = N_R0_Q6_K; pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q6_K_F32].pipeline; } break; case GGML_TYPE_IQ2_XXS: { - nth0 = 4; - nth1 = 16; + nsg = N_SG_IQ2_XXS; + nr0 = N_R0_IQ2_XXS; + smem = 256*8+128; pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_XXS_F32].pipeline; } break; case GGML_TYPE_IQ2_XS: { - nth0 = 4; - nth1 = 16; + nsg = N_SG_IQ2_XS; + nr0 = N_R0_IQ2_XS; + smem = 512*8+128; pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_XS_F32].pipeline; } break; case GGML_TYPE_IQ3_XXS: { - nth0 = 4; - nth1 = 16; + nsg = N_SG_IQ3_XXS; + nr0 = N_R0_IQ3_XXS; + smem = 256*4+128; pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ3_XXS_F32].pipeline; } break; case GGML_TYPE_IQ3_S: { - nth0 = 4; - nth1 = 16; + nsg = N_SG_IQ3_S; + nr0 = N_R0_IQ3_S; + smem = 512*4; pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ3_S_F32].pipeline; } break; case GGML_TYPE_IQ2_S: { - nth0 = 4; - nth1 = 16; + nsg = N_SG_IQ2_S; + nr0 = N_R0_IQ2_S; pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_S_F32].pipeline; } break; case GGML_TYPE_IQ1_S: { - nth0 = 4; - nth1 = 16; + nsg = N_SG_IQ1_S; + nr0 = N_R0_IQ1_S; pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ1_S_F32].pipeline; } break; case GGML_TYPE_IQ1_M: { - nth0 = 4; - nth1 = 16; + nsg = N_SG_IQ1_M; + nr0 = N_R0_IQ1_M; pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ1_M_F32].pipeline; } break; case GGML_TYPE_IQ4_NL: { - nth0 = 4; - nth1 = 16; + nsg = N_SG_IQ4_NL; + nr0 = N_R0_IQ4_NL; + smem = 32*sizeof(float); pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ4_NL_F32].pipeline; } break; case GGML_TYPE_IQ4_XS: { - nth0 = 4; - nth1 = 16; + nsg = N_SG_IQ4_XS; + nr0 = N_R0_IQ4_XS; + smem = 32*sizeof(float); pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ4_XS_F32].pipeline; } break; default: @@ -2609,41 +2833,10 @@ static void ggml_metal_encode_node( [encoder setBuffer:id_src1 offset:offs_src1 atIndex:2]; [encoder setBuffer:id_dst offset:offs_dst atIndex:3]; - if (src0t == GGML_TYPE_Q4_0 || src0t == GGML_TYPE_Q4_1 || src0t == GGML_TYPE_Q5_0 || - src0t == GGML_TYPE_Q5_1 || src0t == GGML_TYPE_Q8_0 || src0t == GGML_TYPE_Q2_K || - src0t == GGML_TYPE_IQ1_S || src0t == GGML_TYPE_IQ1_M || src0t == GGML_TYPE_IQ2_S) { - [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)]; - } - else if (src0t == GGML_TYPE_IQ2_XXS || src0t == GGML_TYPE_IQ2_XS) { - const int mem_size = src0t == GGML_TYPE_IQ2_XXS ? 256*8+128 : 512*8+128; - [encoder setThreadgroupMemoryLength:mem_size atIndex:0]; - [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)]; - } - else if (src0t == GGML_TYPE_IQ3_XXS || src0t == GGML_TYPE_IQ3_S) { - const int mem_size = src0t == GGML_TYPE_IQ3_XXS ? 256*4+128 : 512*4; - [encoder setThreadgroupMemoryLength:mem_size atIndex:0]; - [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)]; - } - else if (src0t == GGML_TYPE_IQ4_NL || src0t == GGML_TYPE_IQ4_XS) { - const int mem_size = 32*sizeof(float); - [encoder setThreadgroupMemoryLength:mem_size atIndex:0]; - [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)]; - } - else if (src0t == GGML_TYPE_Q4_K) { - [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)]; - } - else if (src0t == GGML_TYPE_Q3_K) { - [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)]; - } - else if (src0t == GGML_TYPE_Q5_K) { - [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)]; - } - else if (src0t == GGML_TYPE_Q6_K) { - [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 1)/2, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)]; - } else { - const int64_t ny = (ne11 + nrows - 1)/nrows; - [encoder dispatchThreadgroups:MTLSizeMake(ne01, ny, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)]; + if (smem > 0) { + [encoder setThreadgroupMemoryLength:smem atIndex:0]; } + [encoder dispatchThreadgroups:MTLSizeMake((ne01 + nr0*nsg - 1)/(nr0*nsg), (ne11 + nr1 - 1)/nr1, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(32, nsg, 1)]; } } break; case GGML_OP_MUL_MAT_ID: @@ -2669,20 +2862,19 @@ static void ggml_metal_encode_node( // ne21 = n_rows const int dst_rows = ne20*ne21; const int dst_rows_min = n_as; - const int dst_rows_max = (device.maxThreadgroupMemoryLength - 32 - 8192)/4; + const int dst_rows_max = (device.maxThreadgroupMemoryLength/2 - 8192)/4; // max size of the rowids array in the kernel shared buffer - GGML_ASSERT(dst_rows <= dst_rows_max); + //GGML_ASSERT(dst_rows <= dst_rows_max); // for now the matrix-matrix multiplication kernel only works on A14+/M1+ SoCs // AMD GPU and older A-chips will reuse matrix-vector multiplication kernel - // !!! - // TODO: for now, always use mat-vec kernels until we figure out how to improve the - // indirect matrix multiplication - // !!! if ([device supportsFamily:MTLGPUFamilyApple7] && ne00 % 32 == 0 && ne00 >= 64 && - dst_rows > dst_rows_min) { + //ne01 / ne02 >= 512 && // NOTE: this is based on Mixtral shapes, might need adjustments + dst_rows > dst_rows_min && + dst_rows <= dst_rows_max) { + // some Metal matrix data types require aligned pointers // ref: https://developer.apple.com/metal/Metal-Shading-Language-Specification.pdf (Table 2.5) switch (src0->type) { @@ -2749,146 +2941,155 @@ static void ggml_metal_encode_node( [encoder dispatchThreadgroups:MTLSizeMake((ne21 + 31)/32, (ne01 + 63)/64, n_as) threadsPerThreadgroup:MTLSizeMake(128, 1, 1)]; } else { - int nth0 = 32; - int nth1 = 1; - int nrows = 1; - //printf("vector: ne00 = %6d, ne01 = %6d, ne02 = %6d, ne11 = %6d, ne12 = %6d\n", ne00, ne01, ne02, ne11, ne12); - id pipeline = nil; + int nsg = 0; // number of simdgroups + int nr0 = 0; // number of src0 rows per simdgroup + int nr1 = 1; // number of src1 rows per threadgroup + + size_t smem = 0; // shared memory + // use custom matrix x vector kernel switch (src0t) { case GGML_TYPE_F32: { GGML_ASSERT(src1t == GGML_TYPE_F32); + nsg = 1; + nr0 = 1; pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F32_F32].pipeline; } break; case GGML_TYPE_F16: { GGML_ASSERT(src1t == GGML_TYPE_F32); - nth0 = 32; - nth1 = 1; + nsg = 1; + nr0 = 1; pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F32].pipeline; } break; case GGML_TYPE_BF16: { GGML_ASSERT(src1t == GGML_TYPE_F32); - nth0 = 32; - nth1 = 1; + nsg = 1; + nr0 = 1; pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_BF16_F32].pipeline; } break; case GGML_TYPE_Q4_0: { - nth0 = 8; - nth1 = 8; + nsg = N_SG_Q4_0; + nr0 = N_R0_Q4_0; pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q4_0_F32].pipeline; } break; case GGML_TYPE_Q4_1: { - nth0 = 8; - nth1 = 8; + nsg = N_SG_Q4_1; + nr0 = N_R0_Q4_1; pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q4_1_F32].pipeline; } break; case GGML_TYPE_Q5_0: { - nth0 = 8; - nth1 = 8; + nsg = N_SG_Q5_0; + nr0 = N_R0_Q5_0; pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q5_0_F32].pipeline; } break; case GGML_TYPE_Q5_1: { - nth0 = 8; - nth1 = 8; + nsg = N_SG_Q5_1; + nr0 = N_R0_Q5_1; pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q5_1_F32].pipeline; } break; case GGML_TYPE_Q8_0: { - nth0 = 8; - nth1 = 8; + nsg = N_SG_Q8_0; + nr0 = N_R0_Q8_0; pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q8_0_F32].pipeline; } break; case GGML_TYPE_Q2_K: { - nth0 = 2; - nth1 = 32; + nsg = N_SG_Q2_K; + nr0 = N_R0_Q2_K; pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q2_K_F32].pipeline; } break; case GGML_TYPE_Q3_K: { - nth0 = 2; - nth1 = 32; + nsg = N_SG_Q3_K; + nr0 = N_R0_Q3_K; pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q3_K_F32].pipeline; } break; case GGML_TYPE_Q4_K: { - nth0 = 4; //1; - nth1 = 8; //32; + nsg = N_SG_Q4_K; + nr0 = N_R0_Q4_K; pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q4_K_F32].pipeline; } break; case GGML_TYPE_Q5_K: { - nth0 = 2; - nth1 = 32; + nsg = N_SG_Q5_K; + nr0 = N_R0_Q5_K; pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q5_K_F32].pipeline; } break; case GGML_TYPE_Q6_K: { - nth0 = 2; - nth1 = 32; + nsg = N_SG_Q6_K; + nr0 = N_R0_Q6_K; pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q6_K_F32].pipeline; } break; case GGML_TYPE_IQ2_XXS: { - nth0 = 4; - nth1 = 16; + nsg = N_SG_IQ2_XXS; + nr0 = N_R0_IQ2_XXS; + smem = 256*8+128; pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_XXS_F32].pipeline; } break; case GGML_TYPE_IQ2_XS: { - nth0 = 4; - nth1 = 16; + nsg = N_SG_IQ2_XS; + nr0 = N_R0_IQ2_XS; + smem = 512*8+128; pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_XS_F32].pipeline; } break; case GGML_TYPE_IQ3_XXS: { - nth0 = 4; - nth1 = 16; + nsg = N_SG_IQ3_XXS; + nr0 = N_R0_IQ3_XXS; + smem = 256*4+128; pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ3_XXS_F32].pipeline; } break; case GGML_TYPE_IQ3_S: { - nth0 = 4; - nth1 = 16; + nsg = N_SG_IQ3_S; + nr0 = N_R0_IQ3_S; + smem = 512*4; pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ3_S_F32].pipeline; } break; case GGML_TYPE_IQ2_S: { - nth0 = 4; - nth1 = 16; + nsg = N_SG_IQ2_S; + nr0 = N_R0_IQ2_S; pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_S_F32].pipeline; } break; case GGML_TYPE_IQ1_S: { - nth0 = 4; - nth1 = 16; + nsg = N_SG_IQ1_S; + nr0 = N_R0_IQ1_S; pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ1_S_F32].pipeline; } break; case GGML_TYPE_IQ1_M: { - nth0 = 4; - nth1 = 16; + nsg = N_SG_IQ1_M; + nr0 = N_R0_IQ1_M; pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ1_M_F32].pipeline; } break; case GGML_TYPE_IQ4_NL: { - nth0 = 4; - nth1 = 16; + nsg = N_SG_IQ4_NL; + nr0 = N_R0_IQ4_NL; + smem = 32*sizeof(float); pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ4_NL_F32].pipeline; } break; case GGML_TYPE_IQ4_XS: { - nth0 = 4; - nth1 = 16; + nsg = N_SG_IQ4_XS; + nr0 = N_R0_IQ4_XS; + smem = 32*sizeof(float); pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ4_XS_F32].pipeline; } break; default: @@ -2899,7 +3100,7 @@ static void ggml_metal_encode_node( }; if (ggml_is_quantized(src0t)) { - GGML_ASSERT(ne00 >= nth0*nth1); + GGML_ASSERT(ne00 >= nsg*nr0); } ggml_metal_kargs_mul_mv_id args = { @@ -2932,43 +3133,12 @@ static void ggml_metal_encode_node( [encoder setBuffer:id_src2 offset:offs_src2 atIndex:4]; const int64_t _ne1 = 1; - const int tgz = dst_rows; + const int64_t ne123 = dst_rows; - if (src0t == GGML_TYPE_Q4_0 || src0t == GGML_TYPE_Q4_1 || src0t == GGML_TYPE_Q5_0 || - src0t == GGML_TYPE_Q5_1 || src0t == GGML_TYPE_Q8_0 || src0t == GGML_TYPE_Q2_K || - src0t == GGML_TYPE_IQ1_S || src0t == GGML_TYPE_IQ1_M || src0t == GGML_TYPE_IQ2_S) { - [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, _ne1, tgz) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)]; - } - else if (src0t == GGML_TYPE_IQ2_XXS || src0t == GGML_TYPE_IQ2_XS) { - const int mem_size = src0t == GGML_TYPE_IQ2_XXS ? 256*8+128 : 512*8+128; - [encoder setThreadgroupMemoryLength:mem_size atIndex:0]; - [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, _ne1, tgz) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)]; - } - else if (src0t == GGML_TYPE_IQ3_XXS || src0t == GGML_TYPE_IQ3_S) { - const int mem_size = src0t == GGML_TYPE_IQ3_XXS ? 256*4+128 : 512*4; - [encoder setThreadgroupMemoryLength:mem_size atIndex:0]; - [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, _ne1, tgz) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)]; - } - else if (src0t == GGML_TYPE_IQ4_NL || src0t == GGML_TYPE_IQ4_XS) { - const int mem_size = 32*sizeof(float); - [encoder setThreadgroupMemoryLength:mem_size atIndex:0]; - [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, _ne1, tgz) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)]; - } - else if (src0t == GGML_TYPE_Q4_K) { - [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, _ne1, tgz) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)]; - } - else if (src0t == GGML_TYPE_Q3_K) { - [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, _ne1, tgz) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)]; - } - else if (src0t == GGML_TYPE_Q5_K) { - [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, _ne1, tgz) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)]; - } - else if (src0t == GGML_TYPE_Q6_K) { - [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 1)/2, _ne1, tgz) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)]; - } else { - const int64_t ny = (_ne1 + nrows - 1)/nrows; // = _ne1 - [encoder dispatchThreadgroups:MTLSizeMake(ne01, ny, tgz) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)]; + if (smem > 0) { + [encoder setThreadgroupMemoryLength:smem atIndex:0]; } + [encoder dispatchThreadgroups:MTLSizeMake((ne01 + nr0*nsg - 1)/(nr0*nsg), (_ne1 + nr1 - 1)/nr1, ne123) threadsPerThreadgroup:MTLSizeMake(32, nsg, 1)]; } } break; case GGML_OP_GET_ROWS: @@ -3002,19 +3172,22 @@ static void ggml_metal_encode_node( default: GGML_ABORT("not implemented"); } - // TODO: add ggml_metal_kargs struct + ggml_metal_kargs_get_rows args = { + /*.ne00 =*/ ne00, + /*.nb01 =*/ nb01, + /*.nb02 =*/ nb02, + /*.ne10 =*/ ne10, + /*.nb10 =*/ nb10, + /*.nb11 =*/ nb11, + /*.nb1 =*/ nb1, + /*.nb2 =*/ nb2, + }; + [encoder setComputePipelineState:pipeline]; [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1]; [encoder setBuffer:id_dst offset:offs_dst atIndex:2]; - [encoder setBytes:&ne00 length:sizeof( int64_t) atIndex:3]; - [encoder setBytes:&nb01 length:sizeof(uint64_t) atIndex:4]; - [encoder setBytes:&nb02 length:sizeof(uint64_t) atIndex:5]; - [encoder setBytes:&ne10 length:sizeof( int64_t) atIndex:6]; - [encoder setBytes:&nb10 length:sizeof( int64_t) atIndex:7]; - [encoder setBytes:&nb11 length:sizeof( int64_t) atIndex:8]; - [encoder setBytes:&nb1 length:sizeof(uint64_t) atIndex:9]; - [encoder setBytes:&nb2 length:sizeof(uint64_t) atIndex:10]; + [encoder setBytes:&args length:sizeof(args) atIndex:3]; [encoder dispatchThreadgroups:MTLSizeMake(ne10, ne11, 1) threadsPerThreadgroup:MTLSizeMake(32, 1, 1)]; } break; @@ -3052,6 +3225,42 @@ static void ggml_metal_encode_node( const int64_t nrows = ggml_nrows(src0); + [encoder dispatchThreadgroups:MTLSizeMake(nrows, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)]; + } break; + case GGML_OP_L2_NORM: + { + GGML_ASSERT(ne00 % 4 == 0); + GGML_ASSERT(ggml_is_contiguous_1(src0)); + + float eps; + memcpy(&eps, dst->op_params, sizeof(float)); + + id pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_L2_NORM].pipeline; + + int nth = 32; // SIMD width + + while (nth < ne00/4 && nth < (int) pipeline.maxTotalThreadsPerThreadgroup) { + nth *= 2; + } + + nth = MIN(nth, ne00/4); + + ggml_metal_kargs_l2_norm args = { + /*.ne00 =*/ ne00, + /*.ne00_4 =*/ ne00/4, + /*.nb01 =*/ nb01, + /*.eps =*/ eps, + }; + + [encoder setComputePipelineState:pipeline]; + [encoder setBytes:&args length:sizeof(args) atIndex:0]; + [encoder setBuffer:id_src0 offset:offs_src0 atIndex:1]; + [encoder setBuffer:id_dst offset:offs_dst atIndex:2]; + + [encoder setThreadgroupMemoryLength:32*sizeof(float) atIndex:0]; + + const int64_t nrows = ggml_nrows(src0); + [encoder dispatchThreadgroups:MTLSizeMake(nrows, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)]; } break; case GGML_OP_GROUP_NORM: @@ -3071,18 +3280,21 @@ static void ggml_metal_encode_node( id pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GROUP_NORM].pipeline; - // TODO: add ggml_metal_kargs struct + ggml_metal_kargs_group_norm args = { + /*.ne00 =*/ ne00, + /*.ne01 =*/ ne01, + /*.ne02 =*/ ne02, + /*.nb00 =*/ nb00, + /*.nb01 =*/ nb01, + /*.nb02 =*/ nb02, + /*.n_groups =*/ n_groups, + /*.eps =*/ eps, + }; + [encoder setComputePipelineState:pipeline]; [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; [encoder setBuffer:id_dst offset:offs_dst atIndex:1]; - [encoder setBytes:&ne00 length:sizeof( int64_t) atIndex:2]; - [encoder setBytes:&ne01 length:sizeof( int64_t) atIndex:3]; - [encoder setBytes:&ne02 length:sizeof( int64_t) atIndex:4]; - [encoder setBytes:&nb00 length:sizeof(uint64_t) atIndex:5]; - [encoder setBytes:&nb01 length:sizeof(uint64_t) atIndex:6]; - [encoder setBytes:&nb02 length:sizeof(uint64_t) atIndex:7]; - [encoder setBytes:&n_groups length:sizeof( int32_t) atIndex:8]; - [encoder setBytes:&eps length:sizeof( float) atIndex:9]; + [encoder setBytes:&args length:sizeof(args) atIndex:2]; [encoder setThreadgroupMemoryLength:32*sizeof(float) atIndex:0]; [encoder dispatchThreadgroups:MTLSizeMake(n_groups, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)]; @@ -3240,8 +3452,8 @@ static void ggml_metal_encode_node( const int32_t CHW = IC * KH * KW; - const int32_t ofs0 = src1->nb[is_2D ? 3 : 2] / 4; - const int32_t ofs1 = src1->nb[is_2D ? 2 : 1] / 4; + const uint64_t ofs0 = src1->nb[is_2D ? 3 : 2] / 4; + const uint64_t ofs1 = src1->nb[is_2D ? 2 : 1] / 4; id pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_IM2COL_F32].pipeline; @@ -3263,27 +3475,30 @@ static void ggml_metal_encode_node( default: GGML_ABORT("fatal error"); }; - // TODO: add ggml_metal_kargs struct + ggml_metal_kargs_im2col args = { + /*.ofs0 =*/ ofs0, + /*.ofs1 =*/ ofs1, + /*.IW =*/ IW, + /*.IH =*/ IH, + /*.CHW =*/ CHW, + /*.s0 =*/ s0, + /*.s1 =*/ s1, + /*.p0 =*/ p0, + /*.p1 =*/ p1, + /*.d0 =*/ d0, + /*.d1 =*/ d1, + /*.N =*/ N, + /*.KH =*/ KH, + /*.KW =*/ KW, + /*.KHW =*/ KH * KW, + }; + [encoder setComputePipelineState:pipeline]; [encoder setBuffer:id_src1 offset:offs_src1 atIndex:0]; [encoder setBuffer:id_dst offset:offs_dst atIndex:1]; - [encoder setBytes:&ofs0 length:sizeof(int32_t) atIndex:2]; - [encoder setBytes:&ofs1 length:sizeof(int32_t) atIndex:3]; - [encoder setBytes:&IW length:sizeof(int32_t) atIndex:4]; - [encoder setBytes:&IH length:sizeof(int32_t) atIndex:5]; - [encoder setBytes:&CHW length:sizeof(int32_t) atIndex:6]; - [encoder setBytes:&s0 length:sizeof(int32_t) atIndex:7]; - [encoder setBytes:&s1 length:sizeof(int32_t) atIndex:8]; - [encoder setBytes:&p0 length:sizeof(int32_t) atIndex:9]; - [encoder setBytes:&p1 length:sizeof(int32_t) atIndex:10]; - [encoder setBytes:&d0 length:sizeof(int32_t) atIndex:11]; - [encoder setBytes:&d1 length:sizeof(int32_t) atIndex:12]; + [encoder setBytes:&args length:sizeof(args) atIndex:2]; if (is_gt_mttpt) { - [encoder setBytes:&N length:sizeof(int32_t) atIndex:13]; - [encoder setBytes:&KH length:sizeof(int32_t) atIndex:14]; - [encoder setBytes:&KW length:sizeof(int32_t) atIndex:15]; - const uint64_t n_threads = MIN(pipeline.maxTotalThreadsPerThreadgroup, (uint64_t)N); const int64_t quotient = N / n_threads + (N % n_threads > 0 ? 1 : 0); @@ -3323,16 +3538,20 @@ static void ggml_metal_encode_node( default: GGML_ABORT("fatal error"); }; + ggml_metal_kargs_conv_transpose_1d args = { + /*.IC =*/ IC, + /*.IL =*/ IL, + /*.K =*/ K, + /*.s0 =*/ s0, + /*.nb0 =*/ nb0, + /*.nb1 =*/ nb1, + }; + [encoder setComputePipelineState:pipeline]; [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1]; [encoder setBuffer:id_dst offset:offs_dst atIndex:2]; - [encoder setBytes:&IC length:sizeof( int32_t) atIndex:3]; - [encoder setBytes:&IL length:sizeof( int32_t) atIndex:4]; - [encoder setBytes:&K length:sizeof( int32_t) atIndex:5]; - [encoder setBytes:&s0 length:sizeof( int32_t) atIndex:6]; - [encoder setBytes:&nb0 length:sizeof(uint64_t) atIndex:7]; - [encoder setBytes:&nb1 length:sizeof(uint64_t) atIndex:8]; + [encoder setBytes:&args length:sizeof(args) atIndex:3]; [encoder dispatchThreadgroups:MTLSizeMake(OL, OC, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)]; } break; @@ -3347,30 +3566,33 @@ static void ggml_metal_encode_node( const id pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_UPSCALE_F32].pipeline; - // TODO: add ggml_metal_kargs struct + ggml_metal_kargs_upscale args = { + /*.ne00 =*/ ne00, + /*.ne01 =*/ ne01, + /*.ne02 =*/ ne02, + /*.ne03 =*/ ne03, + /*.nb00 =*/ nb00, + /*.nb01 =*/ nb01, + /*.nb02 =*/ nb02, + /*.nb03 =*/ nb03, + /*.ne0 =*/ ne0, + /*.ne1 =*/ ne1, + /*.ne2 =*/ ne2, + /*.ne3 =*/ ne3, + /*.nb0 =*/ nb0, + /*.nb1 =*/ nb1, + /*.nb2 =*/ nb2, + /*.nb3 =*/ nb3, + /*.sf0 =*/ sf0, + /*.sf1 =*/ sf1, + /*.sf2 =*/ sf2, + /*.sf3 =*/ sf3 + }; + [encoder setComputePipelineState:pipeline]; [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; [encoder setBuffer:id_dst offset:offs_dst atIndex:1]; - [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:2]; - [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:3]; - [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:4]; - [encoder setBytes:&ne03 length:sizeof(ne03) atIndex:5]; - [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:6]; - [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:7]; - [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:8]; - [encoder setBytes:&nb03 length:sizeof(nb03) atIndex:9]; - [encoder setBytes:&ne0 length:sizeof(ne0) atIndex:10]; - [encoder setBytes:&ne1 length:sizeof(ne1) atIndex:11]; - [encoder setBytes:&ne2 length:sizeof(ne2) atIndex:12]; - [encoder setBytes:&ne3 length:sizeof(ne3) atIndex:13]; - [encoder setBytes:&nb0 length:sizeof(nb0) atIndex:14]; - [encoder setBytes:&nb1 length:sizeof(nb1) atIndex:15]; - [encoder setBytes:&nb2 length:sizeof(nb2) atIndex:16]; - [encoder setBytes:&nb3 length:sizeof(nb3) atIndex:17]; - [encoder setBytes:&sf0 length:sizeof(sf0) atIndex:18]; - [encoder setBytes:&sf1 length:sizeof(sf1) atIndex:19]; - [encoder setBytes:&sf2 length:sizeof(sf2) atIndex:20]; - [encoder setBytes:&sf3 length:sizeof(sf3) atIndex:21]; + [encoder setBytes:&args length:sizeof(args) atIndex:2]; const int nth = MIN((int) pipeline.maxTotalThreadsPerThreadgroup, ne0); @@ -3382,26 +3604,29 @@ static void ggml_metal_encode_node( id pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_PAD_F32].pipeline; - // TODO: add ggml_metal_kargs struct + ggml_metal_kargs_pad args = { + /*.ne00 =*/ ne00, + /*.ne01 =*/ ne01, + /*.ne02 =*/ ne02, + /*.ne03 =*/ ne03, + /*.nb00 =*/ nb00, + /*.nb01 =*/ nb01, + /*.nb02 =*/ nb02, + /*.nb03 =*/ nb03, + /*.ne0 =*/ ne0, + /*.ne1 =*/ ne1, + /*.ne2 =*/ ne2, + /*.ne3 =*/ ne3, + /*.nb0 =*/ nb0, + /*.nb1 =*/ nb1, + /*.nb2 =*/ nb2, + /*.nb3 =*/ nb3 + }; + [encoder setComputePipelineState:pipeline]; [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; [encoder setBuffer:id_dst offset:offs_dst atIndex:1]; - [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:2]; - [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:3]; - [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:4]; - [encoder setBytes:&ne03 length:sizeof(ne03) atIndex:5]; - [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:6]; - [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:7]; - [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:8]; - [encoder setBytes:&nb03 length:sizeof(nb03) atIndex:9]; - [encoder setBytes:&ne0 length:sizeof(ne0) atIndex:10]; - [encoder setBytes:&ne1 length:sizeof(ne1) atIndex:11]; - [encoder setBytes:&ne2 length:sizeof(ne2) atIndex:12]; - [encoder setBytes:&ne3 length:sizeof(ne3) atIndex:13]; - [encoder setBytes:&nb0 length:sizeof(nb0) atIndex:14]; - [encoder setBytes:&nb1 length:sizeof(nb1) atIndex:15]; - [encoder setBytes:&nb2 length:sizeof(nb2) atIndex:16]; - [encoder setBytes:&nb3 length:sizeof(nb3) atIndex:17]; + [encoder setBytes:&args length:sizeof(args) atIndex:2]; const int nth = MIN(1024, ne0); @@ -3416,24 +3641,31 @@ static void ggml_metal_encode_node( id pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_PAD_REFLECT_1D_F32].pipeline; + ggml_metal_kargs_pad_reflect_1d args = { + /*.ne00 =*/ ne00, + /*.ne01 =*/ ne01, + /*.ne02 =*/ ne02, + /*.ne03 =*/ ne03, + /*.nb00 =*/ nb00, + /*.nb01 =*/ nb01, + /*.nb02 =*/ nb02, + /*.nb03 =*/ nb03, + /*.ne0 =*/ ne0, + /*.ne1 =*/ ne1, + /*.ne2 =*/ ne2, + /*.ne3 =*/ ne3, + /*.nb0 =*/ nb0, + /*.nb1 =*/ nb1, + /*.nb2 =*/ nb2, + /*.nb3 =*/ nb3, + /*.p0 =*/ p0, + /*.p1 =*/ p1 + }; + [encoder setComputePipelineState:pipeline]; [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; [encoder setBuffer:id_dst offset:offs_dst atIndex:1]; - [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:2]; - [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:3]; - [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:4]; - [encoder setBytes:&ne03 length:sizeof(ne03) atIndex:5]; - [encoder setBytes:&ne0 length:sizeof(ne0) atIndex:6]; - [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:7]; - [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:8]; - [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:9]; - [encoder setBytes:&nb03 length:sizeof(nb03) atIndex:10]; - [encoder setBytes:&nb0 length:sizeof(nb0) atIndex:11]; - [encoder setBytes:&nb1 length:sizeof(nb1) atIndex:12]; - [encoder setBytes:&nb2 length:sizeof(nb2) atIndex:13]; - [encoder setBytes:&nb3 length:sizeof(nb3) atIndex:14]; - [encoder setBytes:&p0 length:sizeof(p0) atIndex:15]; - [encoder setBytes:&p1 length:sizeof(p1) atIndex:16]; + [encoder setBytes:&args length:sizeof(args) atIndex:2]; const int nth = MIN(1024, ne0); @@ -3451,12 +3683,15 @@ static void ggml_metal_encode_node( id pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ARANGE_F32].pipeline; - // TODO: add ggml_metal_kargs struct + ggml_metal_kargs_arange args = { + /*.ne0 =*/ ne0, + /*.start =*/ start, + /*.step =*/ step + }; + [encoder setComputePipelineState:pipeline]; - [encoder setBuffer:id_dst offset:offs_dst atIndex:0]; - [encoder setBytes:&ne0 length:sizeof(ne0) atIndex:1]; - [encoder setBytes:&start length:sizeof(start) atIndex:2]; - [encoder setBytes:&step length:sizeof(step) atIndex:3]; + [encoder setBuffer:id_dst offset:offs_dst atIndex:0]; + [encoder setBytes:&args length:sizeof(args) atIndex:1]; const int nth = MIN(1024, ne0); @@ -3473,13 +3708,16 @@ static void ggml_metal_encode_node( id pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_TIMESTEP_EMBEDDING_F32].pipeline; - // TODO: add ggml_metal_kargs struct + ggml_metal_kargs_timestep_embedding args = { + /*.nb1 =*/ nb1, + /*.dim =*/ dim, + /*.max_period =*/ max_period + }; + [encoder setComputePipelineState:pipeline]; [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; [encoder setBuffer:id_dst offset:offs_dst atIndex:1]; - [encoder setBytes:&nb1 length:sizeof(nb1) atIndex:2]; - [encoder setBytes:&dim length:sizeof(dim) atIndex:3]; - [encoder setBytes:&max_period length:sizeof(max_period) atIndex:4]; + [encoder setBytes:&args length:sizeof(args) atIndex:2]; const int nth = MIN(1024, half); @@ -3512,12 +3750,15 @@ static void ggml_metal_encode_node( default: GGML_ABORT("fatal error"); }; - // TODO: add ggml_metal_kargs struct + ggml_metal_kargs_argsort args = { + /*.ncols =*/ ne00, + /*.ncols_pad =*/ ne00_padded + }; + [encoder setComputePipelineState:pipeline]; - [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; - [encoder setBuffer:id_dst offset:offs_dst atIndex:1]; - [encoder setBytes:&ne00 length:sizeof( int64_t) atIndex:2]; - [encoder setBytes:&ne00_padded length:sizeof( int64_t) atIndex:3]; + [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; + [encoder setBuffer:id_dst offset:offs_dst atIndex:1]; + [encoder setBytes:&args length:sizeof(args) atIndex:2]; [encoder setThreadgroupMemoryLength:mem_size atIndex:0]; [encoder dispatchThreadgroups:MTLSizeMake(1, nrows, 1) threadsPerThreadgroup:MTLSizeMake(ne00_padded, 1, 1)]; @@ -3531,11 +3772,14 @@ static void ggml_metal_encode_node( id pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_LEAKY_RELU_F32].pipeline; - // TODO: add ggml_metal_kargs struct + ggml_metal_kargs_leaky_relu args = { + /*.slope =*/ slope + }; + [encoder setComputePipelineState:pipeline]; [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; [encoder setBuffer:id_dst offset:offs_dst atIndex:1]; - [encoder setBytes:&slope length:sizeof(slope) atIndex:2]; + [encoder setBytes:&args length:sizeof(args) atIndex:2]; const int64_t n = ggml_nelements(dst); @@ -3549,7 +3793,9 @@ static void ggml_metal_encode_node( GGML_ASSERT(src0->type == GGML_TYPE_F32); GGML_ASSERT(src1->type == src2->type); - GGML_ASSERT(ggml_are_same_shape (src1, src2)); + //GGML_ASSERT(ggml_are_same_shape (src1, src2)); + GGML_ASSERT(ne11 == ne21); + GGML_ASSERT(ne12 == ne22); struct ggml_tensor * src3 = node->src[3]; @@ -3596,125 +3842,161 @@ static void ggml_metal_encode_node( // TODO: add vec kernels for (ne00%64 == 0) and maybe also for (ne00%32 == 0) // for now avoiding mainly to keep the number of templates/kernels a bit lower - if (ne01 >= 4 || (ne00%128 != 0)) { + // these are now trivial to add after: https://github.com/ggml-org/llama.cpp/pull/12612 + if (ne01 >= 4 || (ne00%128 != 0 && ne00 != 192)) { switch (src1->type) { case GGML_TYPE_F16: { - switch (ne00) { - case 64: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H64 ].pipeline; break; - case 80: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H80 ].pipeline; break; - case 96: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H96 ].pipeline; break; - case 112: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H112].pipeline; break; - case 128: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H128].pipeline; break; - case 256: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H256].pipeline; break; - default: - { - GGML_LOG_ERROR("unsupported size: %lld\n", ne00); - GGML_LOG_ERROR("add template specialization for this size\n"); - GGML_ABORT("add template specialization for this size"); - } + if (ne00 == 192 && ne20 == 128) { + pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_HK192_HV128].pipeline; + } else { + switch (ne00) { + case 64: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H64 ].pipeline; break; + case 80: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H80 ].pipeline; break; + case 96: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H96 ].pipeline; break; + case 112: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H112].pipeline; break; + case 128: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H128].pipeline; break; + case 192: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H192].pipeline; break; + case 256: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H256].pipeline; break; + default: + { + GGML_LOG_ERROR("unsupported size: %lld\n", ne00); + GGML_LOG_ERROR("add template specialization for this size\n"); + GGML_ABORT("add template specialization for this size"); + } + } } } break; case GGML_TYPE_BF16: { - switch (ne00) { - case 64: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H64 ].pipeline; break; - case 80: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H80 ].pipeline; break; - case 96: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H96 ].pipeline; break; - case 112: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H112].pipeline; break; - case 128: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H128].pipeline; break; - case 256: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H256].pipeline; break; - default: - { - GGML_LOG_ERROR("unsupported size: %lld\n", ne00); - GGML_LOG_ERROR("add template specialization for this size\n"); - GGML_ABORT("add template specialization for this size"); - } + if (ne00 == 192 && ne20 == 128) { + pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_HK192_HV128].pipeline; + } else { + switch (ne00) { + case 64: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H64 ].pipeline; break; + case 80: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H80 ].pipeline; break; + case 96: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H96 ].pipeline; break; + case 112: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H112].pipeline; break; + case 128: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H128].pipeline; break; + case 192: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H192].pipeline; break; + case 256: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H256].pipeline; break; + default: + { + GGML_LOG_ERROR("unsupported size: %lld\n", ne00); + GGML_LOG_ERROR("add template specialization for this size\n"); + GGML_ABORT("add template specialization for this size"); + } + } } } break; case GGML_TYPE_Q4_0: { - switch (ne00) { - case 64: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_H64 ].pipeline; break; - case 80: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_H80 ].pipeline; break; - case 96: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_H96 ].pipeline; break; - case 112: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_H112].pipeline; break; - case 128: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_H128].pipeline; break; - case 256: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_H256].pipeline; break; - default: - { - GGML_LOG_ERROR("unsupported size: %lld\n", ne00); - GGML_LOG_ERROR("add template specialization for this size\n"); - GGML_ABORT("add template specialization for this size"); - } + if (ne00 == 192 && ne20 == 128) { + pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_HK192_HV128].pipeline; + } else { + switch (ne00) { + case 64: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_H64 ].pipeline; break; + case 80: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_H80 ].pipeline; break; + case 96: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_H96 ].pipeline; break; + case 112: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_H112].pipeline; break; + case 128: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_H128].pipeline; break; + case 192: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_H192].pipeline; break; + case 256: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_H256].pipeline; break; + default: + { + GGML_LOG_ERROR("unsupported size: %lld\n", ne00); + GGML_LOG_ERROR("add template specialization for this size\n"); + GGML_ABORT("add template specialization for this size"); + } + } } } break; case GGML_TYPE_Q4_1: { - switch (ne00) { - case 64: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_H64 ].pipeline; break; - case 80: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_H80 ].pipeline; break; - case 96: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_H96 ].pipeline; break; - case 112: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_H112].pipeline; break; - case 128: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_H128].pipeline; break; - case 256: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_H256].pipeline; break; - default: - { - GGML_LOG_ERROR("unsupported size: %lld\n", ne00); - GGML_LOG_ERROR("add template specialization for this size\n"); - GGML_ABORT("add template specialization for this size"); - } + if (ne00 == 192 && ne20 == 128) { + pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_HK192_HV128].pipeline; + } else { + switch (ne00) { + case 64: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_H64 ].pipeline; break; + case 80: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_H80 ].pipeline; break; + case 96: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_H96 ].pipeline; break; + case 112: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_H112].pipeline; break; + case 128: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_H128].pipeline; break; + case 192: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_H192].pipeline; break; + case 256: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_1_H256].pipeline; break; + default: + { + GGML_LOG_ERROR("unsupported size: %lld\n", ne00); + GGML_LOG_ERROR("add template specialization for this size\n"); + GGML_ABORT("add template specialization for this size"); + } + } } } break; case GGML_TYPE_Q5_0: { - switch (ne00) { - case 64: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_H64 ].pipeline; break; - case 80: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_H80 ].pipeline; break; - case 96: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_H96 ].pipeline; break; - case 112: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_H112].pipeline; break; - case 128: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_H128].pipeline; break; - case 256: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_H256].pipeline; break; - default: - { - GGML_LOG_ERROR("unsupported size: %lld\n", ne00); - GGML_LOG_ERROR("add template specialization for this size\n"); - GGML_ABORT("add template specialization for this size"); - } + if (ne00 == 192 && ne20 == 128) { + pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_HK192_HV128].pipeline; + } else { + switch (ne00) { + case 64: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_H64 ].pipeline; break; + case 80: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_H80 ].pipeline; break; + case 96: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_H96 ].pipeline; break; + case 112: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_H112].pipeline; break; + case 128: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_H128].pipeline; break; + case 192: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_H192].pipeline; break; + case 256: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_0_H256].pipeline; break; + default: + { + GGML_LOG_ERROR("unsupported size: %lld\n", ne00); + GGML_LOG_ERROR("add template specialization for this size\n"); + GGML_ABORT("add template specialization for this size"); + } + } } } break; case GGML_TYPE_Q5_1: { - switch (ne00) { - case 64: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_H64 ].pipeline; break; - case 80: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_H80 ].pipeline; break; - case 96: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_H96 ].pipeline; break; - case 112: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_H112].pipeline; break; - case 128: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_H128].pipeline; break; - case 256: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_H256].pipeline; break; - default: - { - GGML_LOG_ERROR("unsupported size: %lld\n", ne00); - GGML_LOG_ERROR("add template specialization for this size\n"); - GGML_ABORT("add template specialization for this size"); - } + if (ne00 == 192 && ne20 == 128) { + pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_HK192_HV128].pipeline; + } else { + switch (ne00) { + case 64: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_H64 ].pipeline; break; + case 80: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_H80 ].pipeline; break; + case 96: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_H96 ].pipeline; break; + case 112: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_H112].pipeline; break; + case 128: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_H128].pipeline; break; + case 192: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_H192].pipeline; break; + case 256: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q5_1_H256].pipeline; break; + default: + { + GGML_LOG_ERROR("unsupported size: %lld\n", ne00); + GGML_LOG_ERROR("add template specialization for this size\n"); + GGML_ABORT("add template specialization for this size"); + } + } } } break; case GGML_TYPE_Q8_0: { - switch (ne00) { - case 64: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H64 ].pipeline; break; - case 80: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H80 ].pipeline; break; - case 96: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H96 ].pipeline; break; - case 112: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H112].pipeline; break; - case 128: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H128].pipeline; break; - case 256: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H256].pipeline; break; - default: - { - GGML_LOG_ERROR("unsupported size: %lld\n", ne00); - GGML_LOG_ERROR("add template specialization for this size\n"); - GGML_ABORT("add template specialization for this size"); - } + if (ne00 == 192 && ne20 == 128) { + pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_HK192_HV128].pipeline; + } else { + switch (ne00) { + case 64: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H64 ].pipeline; break; + case 80: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H80 ].pipeline; break; + case 96: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H96 ].pipeline; break; + case 112: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H112].pipeline; break; + case 128: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H128].pipeline; break; + case 192: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H192].pipeline; break; + case 256: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H256].pipeline; break; + default: + { + GGML_LOG_ERROR("unsupported size: %lld\n", ne00); + GGML_LOG_ERROR("add template specialization for this size\n"); + GGML_ABORT("add template specialization for this size"); + } + } } } break; default: @@ -3746,6 +4028,42 @@ static void ggml_metal_encode_node( } } } break; + case 192: + { + if (ne20 == 128) { + switch (src1->type) { + case GGML_TYPE_F16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_HK192_HV128].pipeline; break; + case GGML_TYPE_BF16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_BF16_HK192_HV128].pipeline; break; + case GGML_TYPE_Q4_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_0_HK192_HV128].pipeline; break; + case GGML_TYPE_Q4_1: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_1_HK192_HV128].pipeline; break; + case GGML_TYPE_Q5_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_0_HK192_HV128].pipeline; break; + case GGML_TYPE_Q5_1: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_1_HK192_HV128].pipeline; break; + case GGML_TYPE_Q8_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q8_0_HK192_HV128].pipeline; break; + default: + { + GGML_LOG_ERROR("unsupported type: %d\n", src1->type); + GGML_LOG_ERROR("add template specialization for this type\n"); + GGML_ABORT("add template specialization for this type"); + } + } + } else { + switch (src1->type) { + case GGML_TYPE_F16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H192].pipeline; break; + case GGML_TYPE_BF16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_BF16_H192].pipeline; break; + case GGML_TYPE_Q4_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_0_H192].pipeline; break; + case GGML_TYPE_Q4_1: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_1_H192].pipeline; break; + case GGML_TYPE_Q5_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_0_H192].pipeline; break; + case GGML_TYPE_Q5_1: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_1_H192].pipeline; break; + case GGML_TYPE_Q8_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q8_0_H192].pipeline; break; + default: + { + GGML_LOG_ERROR("unsupported type: %d\n", src1->type); + GGML_LOG_ERROR("add template specialization for this type\n"); + GGML_ABORT("add template specialization for this type"); + } + } + } + } break; case 256: { switch (src1->type) { @@ -3783,9 +4101,12 @@ static void ggml_metal_encode_node( /*.ne11 =*/ ne11, /*.ne_12_2 =*/ ne12, /*.ne_12_3 =*/ ne13, - /*.nb_12_1 =*/ nb11, - /*.nb_12_2 =*/ nb12, - /*.nb_12_3 =*/ nb13, + /*.nb11 =*/ nb11, + /*.nb12 =*/ nb12, + /*.nb13 =*/ nb13, + /*.nb21 =*/ nb21, + /*.nb22 =*/ nb22, + /*.nb23 =*/ nb23, /*.nb31 =*/ nb31, /*.ne1 =*/ ne1, /*.ne2 =*/ ne2, @@ -3864,10 +4185,9 @@ static void ggml_metal_encode_node( // ne00*(nsg) // each simdgroup has a full f16 head vector in shared mem to accumulate results // -#define FATTN_SMEM(nsg) (GGML_PAD((nqptg*(ne00 + 2*ncpsg*(nsg)) + ne00*(nsg))*(sizeof(float)/2), 16)) +#define FATTN_SMEM(nsg) (GGML_PAD((nqptg*(GGML_PAD(ne00, 128) + 4*ncpsg*(nsg)) + ne20*(nsg))*(sizeof(float)/2), 16)) int64_t nsgmax = 2; - while (true) { const size_t smem = FATTN_SMEM(nsgmax); if (smem > device.maxThreadgroupMemoryLength) { @@ -3899,10 +4219,6 @@ static void ggml_metal_encode_node( case GGML_OP_CPY: case GGML_OP_CONT: { - GGML_ASSERT(ne00 % ggml_blck_size(src0->type) == 0); - - int nth = MIN(1024, ne00/ggml_blck_size(src0->type)); - id pipeline = nil; switch (src0t) { @@ -3936,7 +4252,47 @@ static void ggml_metal_encode_node( switch (dstt) { case GGML_TYPE_F32: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_BF16_F32].pipeline; break; case GGML_TYPE_BF16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_BF16_BF16].pipeline; break; - default: GGML_ASSERT(false && "not implemented"); + default: GGML_ABORT("not implemented"); + }; + } break; + case GGML_TYPE_Q4_0: + { + switch (dstt) { + case GGML_TYPE_F32: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_Q4_0_F32].pipeline; break; + case GGML_TYPE_F16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_Q4_0_F16].pipeline; break; + default: GGML_ABORT("not implemented"); + }; + } break; + case GGML_TYPE_Q4_1: + { + switch (dstt) { + case GGML_TYPE_F32: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_Q4_1_F32].pipeline; break; + case GGML_TYPE_F16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_Q4_1_F16].pipeline; break; + default: GGML_ABORT("not implemented"); + }; + } break; + case GGML_TYPE_Q5_0: + { + switch (dstt) { + case GGML_TYPE_F32: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_Q5_0_F32].pipeline; break; + case GGML_TYPE_F16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_Q5_0_F16].pipeline; break; + default: GGML_ABORT("not implemented"); + }; + } break; + case GGML_TYPE_Q5_1: + { + switch (dstt) { + case GGML_TYPE_F32: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_Q5_1_F32].pipeline; break; + case GGML_TYPE_F16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_Q5_1_F16].pipeline; break; + default: GGML_ABORT("not implemented"); + }; + } break; + case GGML_TYPE_Q8_0: + { + switch (dstt) { + case GGML_TYPE_F32: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_Q8_0_F32].pipeline; break; + case GGML_TYPE_F16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_Q8_0_F16].pipeline; break; + default: GGML_ABORT("not implemented"); }; } break; default: GGML_ABORT("not implemented"); @@ -3966,7 +4322,11 @@ static void ggml_metal_encode_node( [encoder setBuffer:id_src0 offset:offs_src0 atIndex:1]; [encoder setBuffer:id_dst offset:offs_dst atIndex:2]; + GGML_ASSERT(ne00 % ggml_blck_size(src0->type) == 0); + int nth = MIN(1024, ne00/ggml_blck_size(src0->type)); + [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)]; + } break; case GGML_OP_SET: { @@ -4071,21 +4431,24 @@ static void ggml_metal_encode_node( const int64_t n_threads = MIN((int64_t)[pipeline maxTotalThreadsPerThreadgroup], parallel_elements); const int64_t n_tg = (parallel_elements + n_threads - 1) / n_threads; - // TODO: add ggml_metal_kargs struct + ggml_metal_kargs_pool_2d args_pool_2d = { + /* .k0 = */ k0, + /* .k1 = */ k1, + /* .s0 = */ s0, + /* .s1 = */ s1, + /* .p0 = */ p0, + /* .p1 = */ p1, + /* .IH = */ IH, + /* .IW = */ IW, + /* .OH = */ OH, + /* .OW = */ OW, + /* .parallel_elements = */ parallel_elements + }; + [encoder setComputePipelineState:pipeline]; - [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; - [encoder setBuffer:id_dst offset:offs_dst atIndex:1]; - [encoder setBytes:&k0 length:sizeof(int32_t) atIndex:2]; - [encoder setBytes:&k1 length:sizeof(int32_t) atIndex:3]; - [encoder setBytes:&s0 length:sizeof(int32_t) atIndex:4]; - [encoder setBytes:&s1 length:sizeof(int32_t) atIndex:5]; - [encoder setBytes:&p0 length:sizeof(int32_t) atIndex:6]; - [encoder setBytes:&p1 length:sizeof(int32_t) atIndex:7]; - [encoder setBytes:&IH length:sizeof(int64_t) atIndex:8]; - [encoder setBytes:&IW length:sizeof(int64_t) atIndex:9]; - [encoder setBytes:&OH length:sizeof(int64_t) atIndex:10]; - [encoder setBytes:&OW length:sizeof(int64_t) atIndex:11]; - [encoder setBytes:¶llel_elements length:sizeof(int64_t) atIndex:12]; + [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; + [encoder setBuffer:id_dst offset:offs_dst atIndex:1]; + [encoder setBytes:&args_pool_2d length:sizeof(args_pool_2d) atIndex:2]; [encoder dispatchThreadgroups:MTLSizeMake(n_tg, 1, 1) threadsPerThreadgroup:MTLSizeMake(n_threads, 1, 1)]; } break; diff --git a/ggml/src/ggml-metal/ggml-metal.metal b/ggml/src/ggml-metal/ggml-metal.metal index 44f04c909..b08666e27 100644 --- a/ggml/src/ggml-metal/ggml-metal.metal +++ b/ggml/src/ggml-metal/ggml-metal.metal @@ -3,8 +3,7 @@ #if defined(GGML_METAL_EMBED_LIBRARY) __embed_ggml-common.h__ #else -// TODO: this should not be a relative path, but can't figure out how to set Metal include paths in Package.swift -#include "../ggml-common.h" +#include "ggml-common.h" #endif #include "ggml-metal-impl.h" @@ -49,7 +48,7 @@ void dequantize_f16(device const half4x4 * src, short il, thread type4x4 & reg) template void dequantize_f16_t4(device const half4 * src, short il, thread type4 & reg) { - reg = (type4)(*(src + il)); + reg = (type4)(*(src)); } #if defined(GGML_METAL_USE_BF16) @@ -57,6 +56,11 @@ template void dequantize_bf16(device const bfloat4x4 * src, short il, thread type4x4 & reg) { reg = (type4x4)(*src); } + +template +void dequantize_bf16_t4(device const bfloat4 * src, short il, thread type4 & reg) { + reg = (type4)(*(src)); +} #endif template @@ -373,24 +377,33 @@ void dequantize_q5_K(device const block_q5_K *xb, short il, thread type4x4 & reg template void dequantize_q6_K(device const block_q6_K *xb, short il, thread type4x4 & reg) { const half d_all = xb->d; - device const uint8_t * ql = (device const uint8_t *)xb->ql; - device const uint8_t * qh = (device const uint8_t *)xb->qh; + device const uint16_t * ql = (device const uint16_t *)xb->ql; + device const uint16_t * qh = (device const uint16_t *)xb->qh; device const int8_t * scales = (device const int8_t *)xb->scales; - ql = ql + 64*(il/8) + 32*((il/2)&1) + 16*(il&1); - qh = qh + 32*(il/8) + 16*(il&1); + ql = ql + 32*(il/8) + 16*((il/2)&1) + 8*(il&1); + qh = qh + 16*(il/8) + 8*(il&1); float sc = scales[(il%2) + 2 * ((il/2))]; il = (il/2) & 3; - const uint16_t kmask1 = il>1 ? (il>2 ? 192 : 48) : (il>0 ? 12 : 3); - const uint16_t kmask2 = il>1 ? 0xF0 : 0x0F; - const float coef = il>1 ? 1.f/16.f : 1.f; + const uint32_t kmask1 = il>1 ? (il>2 ? 0xC0C0C0C0 : 0x30303030) : (il>0 ? 0x0C0C0C0C : 0x03030303); + const uint32_t kmask2 = il>1 ? 0xF0F0F0F0 : 0x0F0F0F0F; const float ml = d_all * sc * 32.f; - const float dl = d_all * sc * coef; - for (int i = 0; i < 16; ++i) { - const half q = il&1 ? ((ql[i] & kmask2) | ((qh[i] & kmask1) << 2)) - : ((ql[i] & kmask2) | ((qh[i] & kmask1) << 4)); - reg[i/4][i%4] = dl * q - ml; + const float dl0 = d_all * sc; + const float dl1 = dl0 / 256.f; + const float dl2 = dl0 / (256.f * 256.f); + const float dl3 = dl0 / (256.f * 256.f * 256.f); + const uint8_t shr_h = il>2 ? 2 : 0; + const uint8_t shl_h = il>1 ? 0 : (il>0 ? 2 : 4); + const uint8_t shr_l = il>1 ? 4 : 0; + for (int i = 0; i < 4; ++i) { + const uint32_t low = (ql[2*i] | (uint32_t)(ql[2*i+1] << 16)) & kmask2; + const uint32_t high = (qh[2*i] | (uint32_t)(qh[2*i+1] << 16)) & kmask1; + const uint32_t q = ((high << shl_h) >> shr_h) | (low >> shr_l); + reg[i][0] = dl0 * ((half)(q & 0xFF)) - ml; + reg[i][1] = dl1 * ((float)(q & 0xFF00)) - ml; + reg[i][2] = dl2 * ((float)(q & 0xFF0000)) - ml; + reg[i][3] = dl3 * ((float)(q & 0xFF000000)) - ml; } } @@ -939,45 +952,22 @@ kernel void kernel_cos( kernel void kernel_sum_rows( device const float * src0, device float * dst, - constant int64_t & ne00, - constant int64_t & ne01, - constant int64_t & ne02, - constant int64_t & ne03, - constant uint64_t & nb00, - constant uint64_t & nb01, - constant uint64_t & nb02, - constant uint64_t & nb03, - constant int64_t & ne10, - constant int64_t & ne11, - constant int64_t & ne12, - constant int64_t & ne13, - constant uint64_t & nb10, - constant uint64_t & nb11, - constant uint64_t & nb12, - constant uint64_t & nb13, - constant int64_t & ne0, - constant int64_t & ne1, - constant int64_t & ne2, - constant int64_t & ne3, - constant uint64_t & nb0, - constant uint64_t & nb1, - constant uint64_t & nb2, - constant uint64_t & nb3, + constant ggml_metal_kargs_sum_rows & args, uint3 tpig[[thread_position_in_grid]]) { int64_t i3 = tpig.z; int64_t i2 = tpig.y; int64_t i1 = tpig.x; - if (i3 >= ne03 || i2 >= ne02 || i1 >= ne01) { + if (i3 >= args.ne03 || i2 >= args.ne02 || i1 >= args.ne01) { return; } - device const float * src_row = (device const float *) ((device const char *) src0 + i1*nb01 + i2*nb02 + i3*nb03); - device float * dst_row = (device float *) ((device char *) dst + i1*nb1 + i2*nb2 + i3*nb3); + device const float * src_row = (device const float *) ((device const char *) src0 + i1*args.nb01 + i2*args.nb02 + i3*args.nb03); + device float * dst_row = (device float *) ((device char *) dst + i1*args.nb1 + i2*args.nb2 + i3*args.nb3); float row_sum = 0; - for (int64_t i0 = 0; i0 < ne00; i0++) { + for (int64_t i0 = 0; i0 < args.ne00; i0++) { row_sum += src_row[i0]; } @@ -989,36 +979,29 @@ kernel void kernel_soft_max( device const char * src0, device const char * src1, device char * dst, - constant int64_t & ne00, - constant int64_t & ne01, - constant int64_t & ne02, - constant float & scale, - constant float & max_bias, - constant float & m0, - constant float & m1, - constant uint32_t & n_head_log2, + constant ggml_metal_kargs_soft_max & args, threadgroup float * buf [[threadgroup(0)]], uint tgpig[[threadgroup_position_in_grid]], uint tpitg[[thread_position_in_threadgroup]], uint sgitg[[simdgroup_index_in_threadgroup]], uint tiisg[[thread_index_in_simdgroup]], uint ntg[[threads_per_threadgroup]]) { - const int64_t i03 = (tgpig) / (ne02*ne01); - const int64_t i02 = (tgpig - i03*ne02*ne01) / ne01; - const int64_t i01 = (tgpig - i03*ne02*ne01 - i02*ne01); + const int64_t i03 = (tgpig) / (args.ne02*args.ne01); + const int64_t i02 = (tgpig - i03*args.ne02*args.ne01) / args.ne01; + const int64_t i01 = (tgpig - i03*args.ne02*args.ne01 - i02*args.ne01); - device const float * psrc0 = (device const float *) src0 + (i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00); - device const T * pmask = src1 != src0 ? (device const T *) src1 + i01*ne00 : nullptr; - device float * pdst = (device float *) dst + (i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00); + device const float * psrc0 = (device const float *) src0 + (i03*args.ne02*args.ne01*args.ne00 + i02*args.ne01*args.ne00 + i01*args.ne00); + device const T * pmask = src1 != src0 ? (device const T *) src1 + i01*args.ne00 : nullptr; + device float * pdst = (device float *) dst + (i03*args.ne02*args.ne01*args.ne00 + i02*args.ne01*args.ne00 + i01*args.ne00); float slope = 1.0f; // ALiBi - if (max_bias > 0.0f) { + if (args.max_bias > 0.0f) { const int64_t h = i02; - const float base = h < n_head_log2 ? m0 : m1; - const int exp = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1; + const float base = h < args.n_head_log2 ? args.m0 : args.m1; + const int exp = h < args.n_head_log2 ? h + 1 : 2*(h - args.n_head_log2) + 1; slope = pow(base, exp); } @@ -1026,8 +1009,8 @@ kernel void kernel_soft_max( // parallel max float lmax = -INFINITY; - for (int i00 = tpitg; i00 < ne00; i00 += ntg) { - lmax = MAX(lmax, psrc0[i00]*scale + (pmask ? slope*pmask[i00] : 0.0f)); + for (int i00 = tpitg; i00 < args.ne00; i00 += ntg) { + lmax = MAX(lmax, psrc0[i00]*args.scale + (pmask ? slope*pmask[i00] : 0.0f)); } // find the max value in the block @@ -1051,14 +1034,14 @@ kernel void kernel_soft_max( // parallel sum float lsum = 0.0f; - for (int i00 = tpitg; i00 < ne00; i00 += ntg) { - const float exp_psrc0 = exp((psrc0[i00]*scale + (pmask ? slope*pmask[i00] : 0.0f)) - max_val); + for (int i00 = tpitg; i00 < args.ne00; i00 += ntg) { + const float exp_psrc0 = exp((psrc0[i00]*args.scale + (pmask ? slope*pmask[i00] : 0.0f)) - max_val); lsum += exp_psrc0; pdst[i00] = exp_psrc0; } // This barrier fixes a failing test - // ref: https://github.com/ggerganov/ggml/pull/621#discussion_r1425156335 + // ref: https://github.com/ggml-org/ggml/pull/621#discussion_r1425156335 threadgroup_barrier(mem_flags::mem_none); float sum = simd_sum(lsum); @@ -1082,7 +1065,7 @@ kernel void kernel_soft_max( const float inv_sum = 1.0f/sum; - for (int i00 = tpitg; i00 < ne00; i00 += ntg) { + for (int i00 = tpitg; i00 < args.ne00; i00 += ntg) { pdst[i00] *= inv_sum; } } @@ -1092,35 +1075,28 @@ kernel void kernel_soft_max_4( device const char * src0, device const char * src1, device char * dst, - constant int64_t & ne00, - constant int64_t & ne01, - constant int64_t & ne02, - constant float & scale, - constant float & max_bias, - constant float & m0, - constant float & m1, - constant uint32_t & n_head_log2, + constant ggml_metal_kargs_soft_max & args, threadgroup float * buf [[threadgroup(0)]], uint tgpig[[threadgroup_position_in_grid]], uint tpitg[[thread_position_in_threadgroup]], uint sgitg[[simdgroup_index_in_threadgroup]], uint tiisg[[thread_index_in_simdgroup]], uint ntg[[threads_per_threadgroup]]) { - const int64_t i03 = (tgpig) / (ne02*ne01); - const int64_t i02 = (tgpig - i03*ne02*ne01) / ne01; - const int64_t i01 = (tgpig - i03*ne02*ne01 - i02*ne01); + const int64_t i03 = (tgpig) / (args.ne02*args.ne01); + const int64_t i02 = (tgpig - i03*args.ne02*args.ne01) / args.ne01; + const int64_t i01 = (tgpig - i03*args.ne02*args.ne01 - i02*args.ne01); - device const float4 * psrc4 = (device const float4 *) src0 + (i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00)/4; - device const T * pmask = src1 != src0 ? (device const T *) src1 + i01*ne00/4 : nullptr; - device float4 * pdst4 = (device float4 *) dst + (i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00)/4; + device const float4 * psrc4 = (device const float4 *) src0 + (i03*args.ne02*args.ne01*args.ne00 + i02*args.ne01*args.ne00 + i01*args.ne00)/4; + device const T * pmask = src1 != src0 ? (device const T *) src1 + i01*args.ne00/4 : nullptr; + device float4 * pdst4 = (device float4 *) dst + (i03*args.ne02*args.ne01*args.ne00 + i02*args.ne01*args.ne00 + i01*args.ne00)/4; float slope = 1.0f; - if (max_bias > 0.0f) { + if (args.max_bias > 0.0f) { const int64_t h = i02; - const float base = h < n_head_log2 ? m0 : m1; - const int exp = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1; + const float base = h < args.n_head_log2 ? args.m0 : args.m1; + const int exp = h < args.n_head_log2 ? h + 1 : 2*(h - args.n_head_log2) + 1; slope = pow(base, exp); } @@ -1128,8 +1104,8 @@ kernel void kernel_soft_max_4( // parallel max float4 lmax4 = -INFINITY; - for (int i00 = tpitg; i00 < ne00/4; i00 += ntg) { - lmax4 = fmax(lmax4, psrc4[i00]*scale + (float4)((pmask ? slope*pmask[i00] : 0.0f))); + for (int i00 = tpitg; i00 < args.ne00/4; i00 += ntg) { + lmax4 = fmax(lmax4, psrc4[i00]*args.scale + (float4)((pmask ? slope*pmask[i00] : 0.0f))); } const float lmax = MAX(MAX(lmax4[0], lmax4[1]), MAX(lmax4[2], lmax4[3])); @@ -1154,8 +1130,8 @@ kernel void kernel_soft_max_4( // parallel sum float4 lsum4 = 0.0f; - for (int i00 = tpitg; i00 < ne00/4; i00 += ntg) { - const float4 exp_psrc4 = exp((psrc4[i00]*scale + (float4)((pmask ? slope*pmask[i00] : 0.0f))) - max_val); + for (int i00 = tpitg; i00 < args.ne00/4; i00 += ntg) { + const float4 exp_psrc4 = exp((psrc4[i00]*args.scale + (float4)((pmask ? slope*pmask[i00] : 0.0f))) - max_val); lsum4 += exp_psrc4; pdst4[i00] = exp_psrc4; } @@ -1163,7 +1139,7 @@ kernel void kernel_soft_max_4( const float lsum = lsum4[0] + lsum4[1] + lsum4[2] + lsum4[3]; // This barrier fixes a failing test - // ref: https://github.com/ggerganov/ggml/pull/621#discussion_r1425156335 + // ref: https://github.com/ggml-org/ggml/pull/621#discussion_r1425156335 threadgroup_barrier(mem_flags::mem_none); float sum = simd_sum(lsum); @@ -1187,7 +1163,7 @@ kernel void kernel_soft_max_4( const float inv_sum = 1.0f/sum; - for (int i00 = tpitg; i00 < ne00/4; i00 += ntg) { + for (int i00 = tpitg; i00 < args.ne00/4; i00 += ntg) { pdst4[i00] *= inv_sum; } } @@ -1203,27 +1179,23 @@ template [[host_name("kernel_soft_max_f32_4")]] kernel kernel_soft_max_4_t kerne kernel void kernel_diag_mask_inf( device const float * src0, device float * dst, - constant int64_t & ne00, - constant int64_t & ne01, - constant int & n_past, + constant ggml_metal_kargs_diag_mask_inf & args, uint3 tpig[[thread_position_in_grid]]) { const int64_t i02 = tpig[2]; const int64_t i01 = tpig[1]; const int64_t i00 = tpig[0]; - if (i00 > n_past + i01) { - dst[i02*ne01*ne00 + i01*ne00 + i00] = -INFINITY; + if (i00 > args.n_past + i01) { + dst[i02*args.ne01*args.ne00 + i01*args.ne00 + i00] = -INFINITY; } else { - dst[i02*ne01*ne00 + i01*ne00 + i00] = src0[i02*ne01*ne00 + i01*ne00 + i00]; + dst[i02*args.ne01*args.ne00 + i01*args.ne00 + i00] = src0[i02*args.ne01*args.ne00 + i01*args.ne00 + i00]; } } kernel void kernel_diag_mask_inf_8( device const float4 * src0, device float4 * dst, - constant int64_t & ne00, - constant int64_t & ne01, - constant int & n_past, + constant ggml_metal_kargs_diag_mask_inf & args, uint3 tpig[[thread_position_in_grid]]) { const int64_t i = 2*tpig[0]; @@ -1231,42 +1203,26 @@ kernel void kernel_diag_mask_inf_8( dst[i+0] = src0[i+0]; dst[i+1] = src0[i+1]; int64_t i4 = 4*i; - const int64_t i02 = i4/(ne00*ne01); i4 -= i02*ne00*ne01; - const int64_t i01 = i4/(ne00); i4 -= i01*ne00; + const int64_t i02 = i4/(args.ne00*args.ne01); i4 -= i02*args.ne00*args.ne01; + const int64_t i01 = i4/(args.ne00); i4 -= i01*args.ne00; const int64_t i00 = i4; for (int k = 3; k >= 0; --k) { - if (i00 + 4 + k <= n_past + i01) { + if (i00 + 4 + k <= args.n_past + i01) { break; } dst[i+1][k] = -INFINITY; - if (i00 + k > n_past + i01) { + if (i00 + k > args.n_past + i01) { dst[i][k] = -INFINITY; } } } // ref: ggml.c:ggml_compute_forward_ssm_conv_f32 -// TODO: optimize kernel void kernel_ssm_conv_f32( device const void * src0, device const void * src1, device float * dst, - constant int64_t & ne00, - constant int64_t & ne01, - constant int64_t & ne02, - constant uint64_t & nb00, - constant uint64_t & nb01, - constant uint64_t & nb02, - constant int64_t & ne10, - constant int64_t & ne11, - constant uint64_t & nb10, - constant uint64_t & nb11, - constant int64_t & ne0, - constant int64_t & ne1, - constant int64_t & ne2, - constant uint64_t & nb0, - constant uint64_t & nb1, - constant uint64_t & nb2, + constant ggml_metal_kargs_ssm_conv & args, uint3 tgpig[[threadgroup_position_in_grid]], uint3 tpitg[[thread_position_in_threadgroup]], uint3 ntg[[threads_per_threadgroup]]) { @@ -1274,15 +1230,15 @@ kernel void kernel_ssm_conv_f32( const int64_t i2 = tgpig.y; const int64_t i3 = tgpig.z; - const int64_t nc = ne10; - //const int64_t ncs = ne00; - //const int64_t nr = ne01; - //const int64_t n_t = ne1; - //const int64_t n_s = ne2; + const int64_t nc = args.ne10; + //const int64_t ncs = args.ne00; + //const int64_t nr = args.ne01; + //const int64_t n_t = args.ne1; + //const int64_t n_s = args.ne2; - device const float * s = (device const float *) ((device const char *) src0 + ir*nb01 + i2*nb00 + i3*nb02); - device const float * c = (device const float *) ((device const char *) src1 + ir*nb11); - device float * x = (device float *) ((device char *) dst + ir*nb0 + i2*nb1 + i3*nb2); + device const float * s = (device const float *) ((device const char *) src0 + ir*args.nb01 + i2*args.nb00 + i3*args.nb02); + device const float * c = (device const float *) ((device const char *) src1 + ir*args.nb11); + device float * x = (device float *) ((device char *) dst + ir*args.nb0 + i2*args.nb1 + i3*args.nb2); float sumf = 0.0f; @@ -1294,7 +1250,6 @@ kernel void kernel_ssm_conv_f32( } // ref: ggml.c:ggml_compute_forward_ssm_scan_f32 -// TODO: optimize kernel void kernel_ssm_scan_f32( device const void * src0, device const void * src1, @@ -1303,48 +1258,27 @@ kernel void kernel_ssm_scan_f32( device const void * src4, device const void * src5, device float * dst, - constant int64_t & d_state, - constant int64_t & d_inner, - constant int64_t & n_seq_tokens, - constant int64_t & n_seqs, - constant uint64_t & nb00, - constant uint64_t & nb01, - constant uint64_t & nb02, - constant uint64_t & nb10, - constant uint64_t & nb11, - constant uint64_t & nb12, - constant uint64_t & nb13, - constant uint64_t & nb20, - constant uint64_t & nb21, - constant uint64_t & nb22, - constant uint64_t & nb30, - constant uint64_t & nb31, - constant uint64_t & nb40, - constant uint64_t & nb41, - constant uint64_t & nb42, - constant uint64_t & nb50, - constant uint64_t & nb51, - constant uint64_t & nb52, + constant ggml_metal_kargs_ssm_scan & args, uint3 tgpig[[threadgroup_position_in_grid]], uint3 tpitg[[thread_position_in_threadgroup]], uint3 ntg[[threads_per_threadgroup]]) { const int64_t ir = tgpig.x; const int64_t i3 = tgpig.y; - const int64_t nc = d_state; - //const int64_t nr = d_inner; - const int64_t n_t = n_seq_tokens; - //const int64_t n_s = n_seqs; + const int64_t nc = args.d_state; + // const int64_t nr = args.d_inner; + const int64_t n_t = args.n_seq_tokens; + // const int64_t n_s = args.n_seqs; for (int64_t i2 = 0; i2 < n_t; ++i2) { - device const float * s0 = (device const float *) ((device const char *) src0 + ir*nb01 + i3*nb02); - device const float * x = (device const float *) ((device const char *) src1 + ir*nb10 + i2*nb11 + i3*nb12); - device const float * dt = (device const float *) ((device const char *) src2 + ir*nb20 + i2*nb21 + i3*nb22); - device const float * A = (device const float *) ((device const char *) src3 + ir*nb31); - device const float * B = (device const float *) ((device const char *) src4 + i2*nb41 + i3*nb42); - device const float * C = (device const float *) ((device const char *) src5 + i2*nb51 + i3*nb52); - device float * y = (device float *) ((device char *) dst + ir*nb10 + i2*nb11 + i3*nb12); // TODO: do not use src1 strides - device float * s = (device float *) ((device char *) dst + ir*nb01 + i3*nb02 + nb13); + device const float * s0 = (device const float *) ((device const char *) src0 + ir*args.nb01 + i3*args.nb02); + device const float * x = (device const float *) ((device const char *) src1 + ir*args.nb10 + i2*args.nb11 + i3*args.nb12); + device const float * dt = (device const float *) ((device const char *) src2 + ir*args.nb20 + i2*args.nb21 + i3*args.nb22); + device const float * A = (device const float *) ((device const char *) src3 + ir*args.nb31); + device const float * B = (device const float *) ((device const char *) src4 + i2*args.nb41 + i3*args.nb42); + device const float * C = (device const float *) ((device const char *) src5 + i2*args.nb51 + i3*args.nb52); + device float * y = (device float *) ((device char *) dst + ir*args.nb10 + i2*args.nb11 + i3*args.nb12); // TODO: do not use src1 strides + device float * s = (device float *) ((device char *) dst + ir*args.nb01 + i3*args.nb02 + args.nb13); if (i2 > 0) { s0 = s; @@ -1366,6 +1300,184 @@ kernel void kernel_ssm_scan_f32( } } +kernel void kernel_rwkv_wkv6_f32( + device const float * k, + device const float * v, + device const float * r, + device const float * tf, + device const float * td, + device const float * state_in, + device float * dst, + constant uint & B, + constant uint & T, + constant uint & C, + constant uint & H, + uint3 tgpig[[threadgroup_position_in_grid]], + uint3 tpitg[[thread_position_in_threadgroup]], + uint3 ntg[[threads_per_threadgroup]]) { + + const uint head_size = 64; // TODO: support head_size = 128 + const uint batch_id = tgpig.x / H; + const uint head_id = tgpig.x % H; + const uint tid = tpitg.x; + + if (batch_id >= B || head_id >= H) { + return; + } + + const uint state_size = C * head_size; + const uint n_seq_tokens = T / B; + + threadgroup float _k[head_size]; + threadgroup float _r[head_size]; + threadgroup float _tf[head_size]; + threadgroup float _td[head_size]; + + float state[head_size]; + + for (uint i = 0; i < head_size; i++) { + state[i] = state_in[batch_id * state_size + head_id * head_size * head_size + + i * head_size + tid]; + } + + threadgroup_barrier(mem_flags::mem_threadgroup); + _tf[tid] = tf[head_id * head_size + tid]; + threadgroup_barrier(mem_flags::mem_threadgroup); + + const uint start_t = batch_id * n_seq_tokens * C + head_id * head_size + tid; + const uint end_t = (batch_id + 1) * n_seq_tokens * C + head_id * head_size + tid; + + for (uint t = start_t; t < end_t; t += C) { + threadgroup_barrier(mem_flags::mem_threadgroup); + _k[tid] = k[t]; + _r[tid] = r[t]; + _td[tid] = td[t]; + threadgroup_barrier(mem_flags::mem_threadgroup); + + const float v_val = v[t]; + float y = 0.0; + + for (uint j = 0; j < head_size; j += 4) { + float4 k_vec = float4(_k[j], _k[j+1], _k[j+2], _k[j+3]); + float4 r_vec = float4(_r[j], _r[j+1], _r[j+2], _r[j+3]); + float4 tf_vec = float4(_tf[j], _tf[j+1], _tf[j+2], _tf[j+3]); + float4 td_vec = float4(_td[j], _td[j+1], _td[j+2], _td[j+3]); + float4 s_vec = float4(state[j], state[j+1], state[j+2], state[j+3]); + + float4 kv = k_vec * v_val; + + float4 temp = tf_vec * kv + s_vec; + y += dot(r_vec, temp); + + s_vec = s_vec * td_vec + kv; + state[j] = s_vec[0]; + state[j+1] = s_vec[1]; + state[j+2] = s_vec[2]; + state[j+3] = s_vec[3]; + } + + dst[t] = y; + } + + for (uint i = 0; i < head_size; i++) { + dst[T * C + batch_id * state_size + head_id * head_size * head_size + + i * head_size + tid] = state[i]; + } +} + +kernel void kernel_rwkv_wkv7_f32( + device const float * r, + device const float * w, + device const float * k, + device const float * v, + device const float * a, + device const float * b, + device const float * state_in, + device float * dst, + constant uint & B, + constant uint & T, + constant uint & C, + constant uint & H, + uint3 tgpig[[threadgroup_position_in_grid]], + uint3 tpitg[[thread_position_in_threadgroup]], + uint3 ntg[[threads_per_threadgroup]]) { + + const uint head_size = 64; // TODO: support head_size = 128 + const uint batch_id = tgpig.x / H; + const uint head_id = tgpig.x % H; + const uint tid = tpitg.x; + + if (batch_id >= B || head_id >= H) { + return; + } + + const uint state_size = C * head_size; + const uint n_seq_tokens = T / B; + + threadgroup float _r[head_size]; + threadgroup float _w[head_size]; + threadgroup float _k[head_size]; + threadgroup float _a[head_size]; + threadgroup float _b[head_size]; + + float state[head_size]; + + for (uint i = 0; i < head_size; i++) { + state[i] = state_in[batch_id * state_size + head_id * head_size * head_size + + tid * head_size + i]; + } + + const uint start_t = batch_id * n_seq_tokens * C + head_id * head_size + tid; + const uint end_t = (batch_id + 1) * n_seq_tokens * C + head_id * head_size + tid; + + for (uint t = start_t; t < end_t; t += C) { + threadgroup_barrier(mem_flags::mem_threadgroup); + _r[tid] = r[t]; + _w[tid] = w[t]; + _k[tid] = k[t]; + _a[tid] = a[t]; + _b[tid] = b[t]; + threadgroup_barrier(mem_flags::mem_threadgroup); + + const float v_val = v[t]; + float y = 0.0, sa = 0.0; + + float4 sa_vec(0.0); + + for (uint j = 0; j < head_size; j += 4) { + float4 a_vec = float4(_a[j], _a[j+1], _a[j+2], _a[j+3]); + float4 s_vec = float4(state[j], state[j+1], state[j+2], state[j+3]); + sa_vec += a_vec * s_vec; + } + sa = sa_vec[0] + sa_vec[1] + sa_vec[2] + sa_vec[3]; + + for (uint j = 0; j < head_size; j += 4) { + float4 r_vec = float4(_r[j], _r[j+1], _r[j+2], _r[j+3]); + float4 w_vec = float4(_w[j], _w[j+1], _w[j+2], _w[j+3]); + float4 k_vec = float4(_k[j], _k[j+1], _k[j+2], _k[j+3]); + float4 b_vec = float4(_b[j], _b[j+1], _b[j+2], _b[j+3]); + float4 s_vec = float4(state[j], state[j+1], state[j+2], state[j+3]); + + float4 kv = k_vec * v_val; + + s_vec = s_vec * w_vec + kv + sa * b_vec; + y += dot(s_vec, r_vec); + + state[j] = s_vec[0]; + state[j+1] = s_vec[1]; + state[j+2] = s_vec[2]; + state[j+3] = s_vec[3]; + } + + dst[t] = y; + } + + for (uint i = 0; i < head_size; i++) { + dst[T * C + batch_id * state_size + head_id * head_size * head_size + + tid * head_size + i] = state[i]; + } +} + kernel void kernel_argmax( device const void * x, device int32_t * dst, @@ -1534,25 +1646,61 @@ kernel void kernel_rms_norm( } } +kernel void kernel_l2_norm( + constant ggml_metal_kargs_l2_norm & args, + device const char * src0, + device char * dst, + threadgroup float * shmem_f32 [[threadgroup(0)]], + uint tgpig[[threadgroup_position_in_grid]], + ushort tpitg[[thread_position_in_threadgroup]], + ushort sgitg[[simdgroup_index_in_threadgroup]], + ushort tiisg[[thread_index_in_simdgroup]], + ushort ntg[[threads_per_threadgroup]]) { + if (sgitg == 0) { + shmem_f32[tiisg] = 0.0f; + } + + device const float4 * x = (device const float4 *) (src0 + tgpig*args.nb01); + + float sumf = 0.0f; + + // parallel sum + for (int i00 = tpitg; i00 < args.ne00_4; i00 += ntg) { + sumf += dot(x[i00], x[i00]); + } + sumf = simd_sum(sumf); + + threadgroup_barrier(mem_flags::mem_threadgroup); + + if (tiisg == 0) { + shmem_f32[sgitg] = sumf; + } + + threadgroup_barrier(mem_flags::mem_threadgroup); + + sumf = shmem_f32[tiisg]; + sumf = simd_sum(sumf); + + const float scale = 1.0f/sqrt(max(sumf, args.eps)); + + device float4 * y = (device float4 *) dst + tgpig*args.ne00_4; + for (int i00 = tpitg; i00 < args.ne00_4; i00 += ntg) { + y[i00] = x[i00] * scale; + } +} + kernel void kernel_group_norm( device const float * src0, device float * dst, - constant int64_t & ne00, - constant int64_t & ne01, - constant int64_t & ne02, - constant uint64_t & nb00, - constant uint64_t & nb01, - constant uint64_t & nb02, - constant int32_t & n_groups, - constant float & eps, + constant ggml_metal_kargs_group_norm & args, threadgroup float * buf [[threadgroup(0)]], uint tgpig[[threadgroup_position_in_grid]], uint tpitg[[thread_position_in_threadgroup]], uint sgitg[[simdgroup_index_in_threadgroup]], uint tiisg[[thread_index_in_simdgroup]], uint ntg[[threads_per_threadgroup]]) { - const int64_t ne = ne00*ne01*ne02; - const int64_t gs = ne00*ne01*((ne02 + n_groups - 1) / n_groups); + const int64_t ne = args.ne00*args.ne01*args.ne02; + const int64_t gs = args.ne00*args.ne01*((args.ne02 + args.n_groups - 1) / args.n_groups); int start = tgpig * gs; int end = start + gs; @@ -1616,7 +1764,7 @@ kernel void kernel_group_norm( } const float variance = tmp / gs; - const float scale = 1.0f/sqrt(variance + eps); + const float scale = 1.0f/sqrt(variance + args.eps); for (int j = start; j < end; j += ntg) { dst[j] *= scale; } @@ -1710,14 +1858,7 @@ inline float block_q_n_dot_y(device const block_q5_1 * qb_curr, float sumy, thre return d * (acc[0] + acc[1] + acc[2] + acc[3]) + sumy * m; } -// putting them in the kernel cause a significant performance penalty -#define N_DST 4 // each SIMD group works on 4 rows -#define N_SIMDGROUP 2 // number of SIMD groups in a thread group -//Note: This is a template, but strictly speaking it only applies to -// quantizations where the block size is 32. It also does not -// guard against the number of rows not being divisible by -// N_DST, so this is another explicit assumption of the implementation. -template +template void mul_vec_q_n_f32_impl( args_t args, device const char * src0, @@ -1733,7 +1874,7 @@ void mul_vec_q_n_f32_impl( const int r1 = tgpig.y; const int im = tgpig.z; - const int first_row = (r0 * nsg + sgitg) * nr; + const int first_row = (r0 * nsg + sgitg) * nr0; const uint i12 = im%args.ne12; const uint i13 = im/args.ne12; @@ -1745,15 +1886,15 @@ void mul_vec_q_n_f32_impl( device const float * y = (device const float *) (src1 + offset1); // pointers to src0 rows - device const block_q_type * ax[nr]; - for (int row = 0; row < nr; ++row) { + device const block_q_type * ax[nr0]; + for (int row = 0; row < nr0; ++row) { const uint64_t offset0 = (first_row + row)*args.nb01 + (i12/args.r2)*args.nb02 + (i13/args.r3)*args.nb03; ax[row] = (device const block_q_type *) ((device char *) src0 + offset0); } float yl[16]; // src1 vector cache - float sumf[nr] = {0.f}; + float sumf[nr0] = {0.f}; const short ix = (tiisg/2); const short il = (tiisg%2)*8; @@ -1765,7 +1906,7 @@ void mul_vec_q_n_f32_impl( float sumy[2] = { 0.f, 0.f }; #pragma unroll - for (int i = 0; i < 8; i += 2) { + for (short i = 0; i < 8; i += 2) { sumy[0] += yb[i + 0] + yb[i + 1]; yl[i + 0] = yb[i + 0]; yl[i + 1] = yb[i + 1]/256.f; @@ -1776,7 +1917,7 @@ void mul_vec_q_n_f32_impl( } #pragma unroll - for (int row = 0; row < nr; row++) { + for (short row = 0; row < nr0; row++) { sumf[row] += block_q_n_dot_y(ax[row] + ib, sumy[0] + sumy[1], yl, il); } @@ -1785,7 +1926,7 @@ void mul_vec_q_n_f32_impl( device float * dst_f32 = (device float *) dst + im*args.ne0*args.ne1 + r1*args.ne0; - for (int row = 0; row < nr; ++row) { + for (int row = 0; row < nr0; ++row) { const float tot = simd_sum(sumf[row]); if (tiisg == 0 && first_row + row < args.ne01) { @@ -1802,7 +1943,7 @@ kernel void kernel_mul_mv_q4_0_f32( uint3 tgpig[[threadgroup_position_in_grid]], ushort tiisg[[thread_index_in_simdgroup]], ushort sgitg[[simdgroup_index_in_threadgroup]]) { - mul_vec_q_n_f32_impl(args, src0, src1, dst, nullptr, tgpig, tiisg, sgitg); + mul_vec_q_n_f32_impl(args, src0, src1, dst, nullptr, tgpig, tiisg, sgitg); } kernel void kernel_mul_mv_q4_1_f32( @@ -1813,7 +1954,7 @@ kernel void kernel_mul_mv_q4_1_f32( uint3 tgpig[[threadgroup_position_in_grid]], ushort tiisg[[thread_index_in_simdgroup]], ushort sgitg[[simdgroup_index_in_threadgroup]]) { - mul_vec_q_n_f32_impl(args, src0, src1, dst, nullptr, tgpig, tiisg, sgitg); + mul_vec_q_n_f32_impl(args, src0, src1, dst, nullptr, tgpig, tiisg, sgitg); } kernel void kernel_mul_mv_q5_0_f32( @@ -1824,7 +1965,7 @@ kernel void kernel_mul_mv_q5_0_f32( uint3 tgpig[[threadgroup_position_in_grid]], ushort tiisg[[thread_index_in_simdgroup]], ushort sgitg[[simdgroup_index_in_threadgroup]]) { - mul_vec_q_n_f32_impl(args, src0, src1, dst, nullptr, tgpig, tiisg, sgitg); + mul_vec_q_n_f32_impl(args, src0, src1, dst, nullptr, tgpig, tiisg, sgitg); } kernel void kernel_mul_mv_q5_1_f32( @@ -1835,12 +1976,12 @@ kernel void kernel_mul_mv_q5_1_f32( uint3 tgpig[[threadgroup_position_in_grid]], ushort tiisg[[thread_index_in_simdgroup]], ushort sgitg[[simdgroup_index_in_threadgroup]]) { - mul_vec_q_n_f32_impl(args, src0, src1, dst, nullptr, tgpig, tiisg, sgitg); + mul_vec_q_n_f32_impl(args, src0, src1, dst, nullptr, tgpig, tiisg, sgitg); } #define NB_Q8_0 8 -template +template void kernel_mul_mv_q8_0_f32_impl( args_t args, device const char * src0, @@ -1850,16 +1991,13 @@ void kernel_mul_mv_q8_0_f32_impl( uint3 tgpig, ushort tiisg, ushort sgitg) { - const int nr = N_DST; - const int nsg = N_SIMDGROUP; - const int nw = N_SIMDWIDTH; - const int nb = args.ne00/QK8_0; + const int r0 = tgpig.x; const int r1 = tgpig.y; const int im = tgpig.z; - const int first_row = (r0*nsg + sgitg)*nr; + const int first_row = (r0 * nsg + sgitg) * nr0; const uint i12 = im%args.ne12; const uint i13 = im/args.ne12; @@ -1871,15 +2009,15 @@ void kernel_mul_mv_q8_0_f32_impl( device const float * y = (device const float *) (src1 + offset1); // pointers to src0 rows - device const block_q8_0 * ax[nr]; - for (int row = 0; row < nr; ++row) { + device const block_q8_0 * ax[nr0]; + for (int row = 0; row < nr0; ++row) { const uint64_t offset0 = (first_row + row)*args.nb01 + (i12/args.r2)*args.nb02 + (i13/args.r3)*args.nb03; ax[row] = (device const block_q8_0 *) ((device char *) src0 + offset0); } float yl[NB_Q8_0]; - float sumf[nr] = { 0.f }; + float sumf[nr0] = { 0.f }; const short ix = tiisg/4; const short il = tiisg%4; @@ -1892,7 +2030,7 @@ void kernel_mul_mv_q8_0_f32_impl( yl[i] = yb[i]; } - for (int row = 0; row < nr; row++) { + for (short row = 0; row < nr0; row++) { device const int8_t * qs = ax[row][ib].qs + il*NB_Q8_0; float sumq = 0.f; for (short iq = 0; iq < NB_Q8_0; ++iq) { @@ -1906,7 +2044,7 @@ void kernel_mul_mv_q8_0_f32_impl( device float * dst_f32 = (device float *) dst + (uint64_t)im*args.ne0*args.ne1 + (uint64_t)r1*args.ne0; - for (int row = 0; row < nr; ++row) { + for (int row = 0; row < nr0; ++row) { const float tot = simd_sum(sumf[row]); if (tiisg == 0 && first_row + row < args.ne01) { @@ -1924,7 +2062,7 @@ kernel void kernel_mul_mv_q8_0_f32( uint3 tgpig[[threadgroup_position_in_grid]], ushort tiisg[[thread_index_in_simdgroup]], ushort sgitg[[simdgroup_index_in_threadgroup]]) { - kernel_mul_mv_q8_0_f32_impl(args, src0, src1, dst, nullptr, tgpig, tiisg, sgitg); + kernel_mul_mv_q8_0_f32_impl(args, src0, src1, dst, nullptr, tgpig, tiisg, sgitg); } // mat-vec kernel processing in chunks of float4 @@ -2261,9 +2399,9 @@ void kernel_mul_mv_impl( sumf += (T0) x[i] * (T1) y[i]; } - float all_sum = simd_sum(sumf); + float sum_all = simd_sum(sumf); if (tiisg == 0) { - dst_f32[(uint64_t)r1*args.ne0 + r0] = all_sum; + dst_f32[(uint64_t)r1*args.ne0 + r0] = sum_all; } } } else { @@ -2284,10 +2422,10 @@ void kernel_mul_mv_impl( sumf += dot((float4) x4[i], (float4) y4[i]); } - float all_sum = simd_sum(sumf); + float sum_all = simd_sum(sumf); if (tiisg == 0) { - for (int i = 4*(args.ne00/4); i < args.ne00; ++i) all_sum += (float) (x[i] * y[i]); - dst_f32[(uint64_t)r1*args.ne0 + r0] = all_sum; + for (int i = 4*(args.ne00/4); i < args.ne00; ++i) sum_all += (float) (x[i] * y[i]); + dst_f32[(uint64_t)r1*args.ne0 + r0] = sum_all; } } } @@ -2349,9 +2487,9 @@ kernel void kernel_mul_mv_1row( for (int i = tiisg; i < args.ne00; i += 32) { sumf += (float) x[i] * (float) y[i]; } - float all_sum = simd_sum(sumf); + float sum_all = simd_sum(sumf); if (tiisg == 0) { - dst_f32[r0] = all_sum; + dst_f32[r0] = sum_all; } } else { device const T4 * x4 = (device const T4 *) x; @@ -2361,11 +2499,11 @@ kernel void kernel_mul_mv_1row( sumf += dot((float4) x4[i], y4[i]); } - float all_sum = simd_sum(sumf); + float sum_all = simd_sum(sumf); if (tiisg == 0) { - for (int i = 4*(args.ne00/4); i < args.ne00; ++i) all_sum += (float) (x[i] * y[i]); - dst_f32[r0] = all_sum; + for (int i = 4*(args.ne00/4); i < args.ne00; ++i) sum_all += (float) (x[i] * y[i]); + dst_f32[r0] = sum_all; } } } @@ -2410,9 +2548,9 @@ kernel void kernel_mul_mv_l4( sumf += dot((float4) x4[i], y4[i]); } - float all_sum = simd_sum(sumf); + float sum_all = simd_sum(sumf); if (tiisg == 0) { - dst_f32[(uint64_t)r1*args.ne0 + r0] = all_sum; + dst_f32[(uint64_t)r1*args.ne0 + r0] = sum_all; } } } @@ -2580,17 +2718,7 @@ template [[host_name("kernel_rope_neox_f16")]] kernel kernel_rope_neox_t kernel_ typedef void (im2col_t)( device const float * x, device char * dst, - constant int32_t & ofs0, - constant int32_t & ofs1, - constant int32_t & IW, - constant int32_t & IH, - constant int32_t & CHW, - constant int32_t & s0, - constant int32_t & s1, - constant int32_t & p0, - constant int32_t & p1, - constant int32_t & d0, - constant int32_t & d1, + constant ggml_metal_kargs_im2col & args, uint3 tgpig[[threadgroup_position_in_grid]], uint3 tgpg[[threadgroups_per_grid]], uint3 tpitg[[thread_position_in_threadgroup]], @@ -2600,17 +2728,7 @@ template kernel void kernel_im2col( device const float * x, device char * dst, - constant int32_t & ofs0, - constant int32_t & ofs1, - constant int32_t & IW, - constant int32_t & IH, - constant int32_t & CHW, - constant int32_t & s0, - constant int32_t & s1, - constant int32_t & p0, - constant int32_t & p1, - constant int32_t & d0, - constant int32_t & d1, + constant ggml_metal_kargs_im2col & args, uint3 tgpig[[threadgroup_position_in_grid]], uint3 tgpg[[threadgroups_per_grid]], uint3 tpitg[[thread_position_in_threadgroup]], @@ -2631,17 +2749,17 @@ kernel void kernel_im2col( const int64_t ioh = tgpig[1]; const int64_t iow = tgpig[2]; - const int64_t iiw = iow*s0 + ikw*d0 - p0; - const int64_t iih = ioh*s1 + ikh*d1 - p1; + const int64_t iiw = iow*args.s0 + ikw*args.d0 - args.p0; + const int64_t iih = ioh*args.s1 + ikh*args.d1 - args.p1; - const int64_t offset_dst = (in*OH*OW + ioh*OW + iow)*CHW + (iic*(KH*KW) + ikh*KW + ikw); + const int64_t offset_dst = (in*OH*OW + ioh*OW + iow)*args.CHW + (iic*(KH*KW) + ikh*KW + ikw); device T * pdst = (device T *) (dst); - if (iih < 0 || iih >= IH || iiw < 0 || iiw >= IW) { + if (iih < 0 || iih >= args.IH || iiw < 0 || iiw >= args.IW) { pdst[offset_dst] = 0.0f; } else { - const int64_t offset_src = in*ofs0 + iic*ofs1 + iih*IW + iiw; + const int64_t offset_src = in*args.ofs0 + iic*args.ofs1 + iih*args.IW + iiw; pdst[offset_dst] = x[offset_src]; } } @@ -2652,20 +2770,7 @@ template [[host_name("kernel_im2col_f16")]] kernel im2col_t kernel_im2col; typedef void (im2col_ext_t)( device const float * x, device char * dst, - constant int32_t & ofs0, - constant int32_t & ofs1, - constant int32_t & IW, - constant int32_t & IH, - constant int32_t & CHW, - constant int32_t & s0, - constant int32_t & s1, - constant int32_t & p0, - constant int32_t & p1, - constant int32_t & d0, - constant int32_t & d1, - constant int32_t & N, - constant int32_t & KH, - constant int32_t & KW, + constant ggml_metal_kargs_im2col & args, uint3 tgpig[[threadgroup_position_in_grid]], uint3 tgpg[[threadgroups_per_grid]], uint3 tpitg[[thread_position_in_threadgroup]], @@ -2675,53 +2780,40 @@ template kernel void kernel_im2col_ext( device const float * x, device char * dst, - constant int32_t & ofs0, - constant int32_t & ofs1, - constant int32_t & IW, - constant int32_t & IH, - constant int32_t & CHW, - constant int32_t & s0, - constant int32_t & s1, - constant int32_t & p0, - constant int32_t & p1, - constant int32_t & d0, - constant int32_t & d1, - constant int32_t & N, - constant int32_t & KH, - constant int32_t & KW, + constant ggml_metal_kargs_im2col & args, uint3 tgpig[[threadgroup_position_in_grid]], uint3 tgpg[[threadgroups_per_grid]], // tgpg[0] = D x IC x KH x KW, CHW = IC x KH x KW uint3 tpitg[[thread_position_in_threadgroup]], uint3 ntg[[threads_per_threadgroup]]) { // [M, 1, 1] - const int64_t KHW = KH * KW; // KHW == ntg[1] * ntg[2], KW == ntg[2] + const int64_t KHW = (int64_t)args.KHW; - const int64_t d = tgpig[0] / CHW; - const int64_t chw = tgpig[0] % CHW; + const int64_t d = tgpig[0] / args.CHW; + const int64_t chw = tgpig[0] % args.CHW; const int64_t tgpig_0 = chw / KHW; // 0 ~ (IC - 1) const int64_t HW = tgpig[0] % KHW; const int64_t tpitg_0 = (d * ntg[0]) + tpitg[0]; - if (tpitg_0 >= N) { + if (tpitg_0 >= args.N) { return; } - const int64_t tpitg_1 = HW / KW; - const int64_t tpitg_2 = HW % KW; + const int64_t tpitg_1 = HW / args.KW; + const int64_t tpitg_2 = HW % args.KW; - const int64_t iiw = tgpig[2] * s0 + tpitg_2 * d0 - p0; - const int64_t iih = tgpig[1] * s1 + tpitg_1 * d1 - p1; + const int64_t iiw = tgpig[2] * args.s0 + tpitg_2 * args.d0 - args.p0; + const int64_t iih = tgpig[1] * args.s1 + tpitg_1 * args.d1 - args.p1; const int64_t offset_dst = - (tpitg_0 * tgpg[1] * tgpg[2] + tgpig[1] * tgpg[2] + tgpig[2]) * CHW + - (tgpig_0 * KHW + tpitg_1 * KW + tpitg_2); + (tpitg_0 * tgpg[1] * tgpg[2] + tgpig[1] * tgpg[2] + tgpig[2]) * args.CHW + + (tgpig_0 * KHW + tpitg_1 * args.KW + tpitg_2); device T * pdst = (device T *) (dst); - if (iih < 0 || iih >= IH || iiw < 0 || iiw >= IW) { + if (iih < 0 || iih >= args.IH || iiw < 0 || iiw >= args.IW) { pdst[offset_dst] = 0.0f; } else { - const int64_t offset_src = tpitg_0 * ofs0 + tgpig_0 * ofs1; - pdst[offset_dst] = x[offset_src + iih * IW + iiw]; + const int64_t offset_src = tpitg_0 * args.ofs0 + tgpig_0 * args.ofs1; + pdst[offset_dst] = x[offset_src + iih * args.IW + iiw]; } } @@ -2732,12 +2824,7 @@ typedef void (conv_transpose_1d_t)( device const float * src0, device const float * src1, device char * dst, - constant int32_t & IC, - constant int32_t & IL, - constant int32_t & K, - constant int32_t & s0, - constant uint64_t & nb0, - constant uint64_t & nb1, + constant ggml_metal_kargs_conv_transpose_1d & args, uint3 tgpig[[threadgroup_position_in_grid]], uint3 tgpg[[threadgroups_per_grid]]); @@ -2746,29 +2833,24 @@ kernel void kernel_conv_transpose_1d( device const T * src0, device const float * src1, device char * dst, - constant int32_t & IC, - constant int32_t & IL, - constant int32_t & K, - constant int32_t & s0, - constant uint64_t & nb0, - constant uint64_t & nb1, + constant ggml_metal_kargs_conv_transpose_1d & args, uint3 tgpig[[threadgroup_position_in_grid]], uint3 tgpg[[threadgroups_per_grid]]) { float v = 0.0f; - for (int64_t c = 0; c < IC; c++) { - const int32_t kernel_offset = c * tgpg[1] * K + K * tgpig[1]; - const int32_t input_offset = c * IL; + for (int64_t c = 0; c < args.IC; c++) { + const int32_t kernel_offset = c * tgpg[1] * args.K + args.K * tgpig[1]; + const int32_t input_offset = c * args.IL; - for (int64_t i = 0; i < IL; i++) { - if (tgpig[0] >= i * s0 && tgpig[0] < i * s0 + K) { - v += src0[kernel_offset + tgpig[0] - i * s0] * src1[input_offset + i]; + for (int64_t i = 0; i < args.IL; i++) { + if (tgpig[0] >= i * args.s0 && tgpig[0] < i * args.s0 + args.K) { + v += src0[kernel_offset + tgpig[0] - i * args.s0] * src1[input_offset + i]; } } } - device float * dst_ptr = (device float *) (dst + tgpig[0] * nb0 + tgpig[1] * nb1); + device float * dst_ptr = (device float *) (dst + tgpig[0] * args.nb0 + tgpig[1] * args.nb1); dst_ptr[0] = v; } @@ -2778,12 +2860,7 @@ kernel void kernel_conv_transpose_1d( device const float * src0, device const float * src1, device char * dst, - constant int32_t & IC, - constant int32_t & IL, - constant int32_t & K, - constant int32_t & s0, - constant uint64_t & nb0, - constant uint64_t & nb1, + constant ggml_metal_kargs_conv_transpose_1d & args, uint3 tgpig[[threadgroup_position_in_grid]], uint3 tgpg[[threadgroups_per_grid]]); @@ -2792,38 +2869,14 @@ kernel void kernel_conv_transpose_1d( device const half * src0, device const float * src1, device char * dst, - constant int32_t & IC, - constant int32_t & IL, - constant int32_t & K, - constant int32_t & s0, - constant uint64_t & nb0, - constant uint64_t & nb1, + constant ggml_metal_kargs_conv_transpose_1d & args, uint3 tgpig[[threadgroup_position_in_grid]], uint3 tgpg[[threadgroups_per_grid]]); kernel void kernel_upscale_f32( device const char * src0, device char * dst, - constant int64_t & ne00, - constant int64_t & ne01, - constant int64_t & ne02, - constant int64_t & ne03, - constant uint64_t & nb00, - constant uint64_t & nb01, - constant uint64_t & nb02, - constant uint64_t & nb03, - constant int64_t & ne0, - constant int64_t & ne1, - constant int64_t & ne2, - constant int64_t & ne3, - constant uint64_t & nb0, - constant uint64_t & nb1, - constant uint64_t & nb2, - constant uint64_t & nb3, - constant float & sf0, - constant float & sf1, - constant float & sf2, - constant float & sf3, + constant ggml_metal_kargs_upscale & args, uint3 tgpig[[threadgroup_position_in_grid]], uint3 tpitg[[thread_position_in_threadgroup]], uint3 ntg[[threads_per_threadgroup]]) { @@ -2832,15 +2885,15 @@ kernel void kernel_upscale_f32( const int64_t i2 = tgpig.y; const int64_t i1 = tgpig.x; - const int64_t i03 = i3/sf3; - const int64_t i02 = i2/sf2; - const int64_t i01 = i1/sf1; + const int64_t i03 = i3/args.sf3; + const int64_t i02 = i2/args.sf2; + const int64_t i01 = i1/args.sf1; - for (int i0 = tpitg.x; i0 < ne0; i0 += ntg.x) { - const int64_t i00 = i0/sf0; + for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) { + const int64_t i00 = i0/args.sf0; - device const float * src0_ptr = (device const float *) (src0 + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00); - device float * dst_ptr = (device float *) (dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0); + device const float * src0_ptr = (device const float *) (src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01 + i00*args.nb00); + device float * dst_ptr = (device float *) (dst + i3*args.nb3 + i2*args.nb2 + i1*args.nb1 + i0*args.nb0); dst_ptr[0] = src0_ptr[0]; } @@ -2849,22 +2902,7 @@ kernel void kernel_upscale_f32( kernel void kernel_pad_f32( device const char * src0, device char * dst, - constant int64_t & ne00, - constant int64_t & ne01, - constant int64_t & ne02, - constant int64_t & ne03, - constant uint64_t & nb00, - constant uint64_t & nb01, - constant uint64_t & nb02, - constant uint64_t & nb03, - constant int64_t & ne0, - constant int64_t & ne1, - constant int64_t & ne2, - constant int64_t & ne3, - constant uint64_t & nb0, - constant uint64_t & nb1, - constant uint64_t & nb2, - constant uint64_t & nb3, + constant ggml_metal_kargs_pad & args, uint3 tgpig[[threadgroup_position_in_grid]], uint3 tpitg[[thread_position_in_threadgroup]], uint3 ntg[[threads_per_threadgroup]]) { @@ -2877,12 +2915,12 @@ kernel void kernel_pad_f32( const int64_t i02 = i2; const int64_t i01 = i1; - device const float * src0_ptr = (device const float *) (src0 + i03*nb03 + i02*nb02 + i01*nb01); - device float * dst_ptr = (device float *) (dst + i3*nb3 + i2*nb2 + i1*nb1); + device const float * src0_ptr = (device const float *) (src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01); + device float * dst_ptr = (device float *) (dst + i3*args.nb3 + i2*args.nb2 + i1*args.nb1); - if (i1 < ne01 && i2 < ne02 && i3 < ne03) { - for (int i0 = tpitg.x; i0 < ne0; i0 += ntg.x) { - if (i0 < ne00) { + if (i1 < args.ne01 && i2 < args.ne02 && i3 < args.ne03) { + for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) { + if (i0 < args.ne00) { dst_ptr[i0] = src0_ptr[i0]; } else { dst_ptr[i0] = 0.0f; @@ -2892,7 +2930,7 @@ kernel void kernel_pad_f32( return; } - for (int i0 = tpitg.x; i0 < ne0; i0 += ntg.x) { + for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) { dst_ptr[i0] = 0.0f; } } @@ -2900,21 +2938,7 @@ kernel void kernel_pad_f32( kernel void kernel_pad_reflect_1d_f32( device const char * src0, device char * dst, - constant int64_t & ne00, - constant int64_t & ne01, - constant int64_t & ne02, - constant int64_t & ne03, - constant int64_t & ne0, - constant uint64_t & nb00, - constant uint64_t & nb01, - constant uint64_t & nb02, - constant uint64_t & nb03, - constant uint64_t & nb0, - constant uint64_t & nb1, - constant uint64_t & nb2, - constant uint64_t & nb3, - constant int32_t & p0, - constant int32_t & p1, + constant ggml_metal_kargs_pad_reflect_1d & args, uint3 tgpig[[threadgroup_position_in_grid]], uint3 tgpg[[threadgroups_per_grid]], uint3 tpitg[[thread_position_in_threadgroup]], @@ -2928,17 +2952,17 @@ kernel void kernel_pad_reflect_1d_f32( const int64_t i02 = i2; const int64_t i01 = i1; - device const float * src0_ptr = (device const float *) (src0 + i03*nb03 + i02*nb02 + i01*nb01); - device float * dst_ptr = (device float *) (dst + i3*nb3 + i2*nb2 + i1*nb1); + device const float * src0_ptr = (device const float *) (src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01); + device float * dst_ptr = (device float *) (dst + i3*args.nb3 + i2*args.nb2 + i1*args.nb1); - if (i1 < ne01 && i2 < ne02 && i3 < ne03) { - for (int i0 = tpitg.x; i0 < ne0; i0 += ntg.x) { - if (i0 < p0) { - dst_ptr[i0] = src0_ptr[p0 - i0]; - } else if (i0 < ne0 - p1) { - dst_ptr[i0] = src0_ptr[i0 - p0]; + if (i1 < args.ne01 && i2 < args.ne02 && i3 < args.ne03) { + for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) { + if (i0 < args.p0) { + dst_ptr[i0] = src0_ptr[args.p0 - i0]; + } else if (i0 < args.ne0 - args.p1) { + dst_ptr[i0] = src0_ptr[i0 - args.p0]; } else { - dst_ptr[i0] = src0_ptr[(ne0 - p1 - p0) - (p1 + 1 - (ne0 - i0)) - 1]; + dst_ptr[i0] = src0_ptr[(args.ne0 - args.p1 - args.p0) - (args.p1 + 1 - (args.ne0 - i0)) - 1]; } } } @@ -2946,44 +2970,40 @@ kernel void kernel_pad_reflect_1d_f32( kernel void kernel_arange_f32( device char * dst, - constant int64_t & ne0, - constant float & start, - constant float & step, + constant ggml_metal_kargs_arange & args, uint3 tgpig[[threadgroup_position_in_grid]], uint3 tpitg[[thread_position_in_threadgroup]], uint3 ntg[[threads_per_threadgroup]]) { device float * dst_ptr = (device float *) dst; - for (int i0 = tpitg.x; i0 < ne0; i0 += ntg.x) { - dst_ptr[i0] = start + step * i0; + for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) { + dst_ptr[i0] = args.start + args.step * i0; } } kernel void kernel_timestep_embedding_f32( device const char * src0, device char * dst, - constant uint64_t & nb1, - constant int & dim, - constant int & max_period, + constant ggml_metal_kargs_timestep_embedding & args, uint3 tgpig[[threadgroup_position_in_grid]], uint3 tpitg[[thread_position_in_threadgroup]], uint3 ntg[[threads_per_threadgroup]]) { int i = tgpig.x; - device float * embed_data = (device float *)(dst + i*nb1); + device float * embed_data = (device float *)(dst + i*args.nb1); - int half_ = dim / 2; + int half_ = args.dim / 2; for (int j = tpitg.x; j < half_; j += ntg.x) { float timestep = ((device float *)src0)[i]; - float freq = (float)exp(-log((float)max_period) * j / half_); + float freq = (float)exp(-log((float)args.max_period) * j / half_); float arg = timestep * freq; embed_data[j ] = cos(arg); embed_data[j + half_] = sin(arg); } - if (dim % 2 != 0 && tpitg.x == 0) { - embed_data[dim] = 0.f; + if (args.dim % 2 != 0 && tpitg.x == 0) { + embed_data[args.dim] = 0.f; } } @@ -2991,8 +3011,7 @@ kernel void kernel_timestep_embedding_f32( typedef void (argsort_t)( device const float * x, device int32_t * dst, - constant int64_t & ncols, - constant int64_t & ncols_pad, + constant ggml_metal_kargs_argsort & args, threadgroup int32_t * shared_values [[threadgroup(0)]], uint3 tgpig[[threadgroup_position_in_grid]], uint3 tpitg[[thread_position_in_threadgroup]]); @@ -3001,8 +3020,7 @@ template kernel void kernel_argsort_f32_i32( device const float * x, device int32_t * dst, - constant int64_t & ncols, - constant int64_t & ncols_pad, + constant ggml_metal_kargs_argsort & args, threadgroup int32_t * shared_values [[threadgroup(0)]], uint3 tgpig[[threadgroup_position_in_grid]], uint3 tpitg[[thread_position_in_threadgroup]]) { @@ -3010,9 +3028,9 @@ kernel void kernel_argsort_f32_i32( int col = tpitg[0]; int row = tgpig[1]; - if (col >= ncols_pad) return; + if (col >= args.ncols_pad) return; - device const float * x_row = x + row * ncols; + device const float * x_row = x + row * args.ncols; threadgroup int32_t * dst_row = shared_values; // initialize indices @@ -3020,21 +3038,21 @@ kernel void kernel_argsort_f32_i32( threadgroup_barrier(mem_flags::mem_threadgroup); - for (int k = 2; k <= ncols_pad; k *= 2) { + for (int k = 2; k <= args.ncols_pad; k *= 2) { for (int j = k / 2; j > 0; j /= 2) { int ixj = col ^ j; if (ixj > col) { if ((col & k) == 0) { - if (dst_row[col] >= ncols || - (dst_row[ixj] < ncols && (order == GGML_SORT_ORDER_ASC ? + if (dst_row[col] >= args.ncols || + (dst_row[ixj] < args.ncols && (order == GGML_SORT_ORDER_ASC ? x_row[dst_row[col]] > x_row[dst_row[ixj]] : x_row[dst_row[col]] < x_row[dst_row[ixj]])) ) { SWAP(dst_row[col], dst_row[ixj]); } } else { - if (dst_row[ixj] >= ncols || - (dst_row[col] < ncols && (order == GGML_SORT_ORDER_ASC ? + if (dst_row[ixj] >= args.ncols || + (dst_row[col] < args.ncols && (order == GGML_SORT_ORDER_ASC ? x_row[dst_row[col]] < x_row[dst_row[ixj]] : x_row[dst_row[col]] > x_row[dst_row[ixj]])) ) { @@ -3047,8 +3065,8 @@ kernel void kernel_argsort_f32_i32( } // copy the result to dst without the padding - if (col < ncols) { - dst[row * ncols + col] = dst_row[col]; + if (col < args.ncols) { + dst[row * args.ncols + col] = dst_row[col]; } } @@ -3058,9 +3076,9 @@ template [[host_name("kernel_argsort_f32_i32_desc")]] kernel argsort_t kernel_ar kernel void kernel_leaky_relu_f32( device const float * src0, device float * dst, - constant float & slope, + constant ggml_metal_kargs_leaky_relu & args, uint tpig[[thread_position_in_grid]]) { - dst[tpig] = src0[tpig] > 0.0f ? src0[tpig] : src0[tpig] * slope; + dst[tpig] = src0[tpig] > 0.0f ? src0[tpig] : src0[tpig] * args.slope; } // ref: https://arxiv.org/pdf/2307.08691.pdf @@ -3087,7 +3105,8 @@ template< typename vd4x4_t, // key type in device memory short nl_v, void (*deq_v)(device const vd4x4_t *, short, thread v4x4_t &), - short D, // head size + short DK, // K head size + short DV, // V head size short Q = 8, // queries per threadgroup short KV = 8, // key/value processed per each simdgroup short C = 32> // cache items per threadgroup @@ -3109,20 +3128,24 @@ kernel void kernel_flash_attn_ext( const int iq2 = tgpig[1]; const int iq1 = tgpig[0]*Q; - const short D4 = D/4; - const short D8 = D/8; - const short D16 = D/16; - const short NW = N_SIMDWIDTH; - const short SH = (2*C + Q); // shared memory per simdgroup (s_t == float) + constexpr short DK4 = DK/4; + constexpr short DK8 = DK/8; + constexpr short DK16 = DK/16; + constexpr short DV4 = DV/4; + constexpr short DV8 = DV/8; + constexpr short DV16 = DV/16; + + constexpr short NW = N_SIMDWIDTH; + constexpr short SH = (2*C + Q); // shared memory per simdgroup (s_t == float) const short TS = nsg*SH; // shared memory size per query in (s_t == float) - const short T = D + 2*TS; // shared memory size per query in (half) + const short T = DK + 2*TS; // shared memory size per query in (half) - threadgroup q_t * sq = (threadgroup q_t *) (shmem_f16 + 0*D); // holds the query data - threadgroup q4_t * sq4 = (threadgroup q4_t *) (shmem_f16 + 0*D); // same as above but in q4_t - threadgroup o_t * so = (threadgroup o_t *) (shmem_f16 + 0*D); // reuse query data for accumulation - threadgroup o4_t * so4 = (threadgroup o4_t *) (shmem_f16 + 0*D); // same as above but in o4_t - threadgroup s_t * ss = (threadgroup s_t *) (shmem_f16 + 2*sgitg*SH + Q*D); // scratch buffer for attention, mask and diagonal matrix + threadgroup q_t * sq = (threadgroup q_t *) (shmem_f16 + 0*DK); // holds the query data + threadgroup q4_t * sq4 = (threadgroup q4_t *) (shmem_f16 + 0*DK); // same as above but in q4_t + threadgroup o_t * so = (threadgroup o_t *) (shmem_f16 + 0*DK); // reuse query data for accumulation + threadgroup o4_t * so4 = (threadgroup o4_t *) (shmem_f16 + 0*DK); // same as above but in o4_t + threadgroup s_t * ss = (threadgroup s_t *) (shmem_f16 + 2*sgitg*SH + Q*DK); // scratch buffer for attention, mask and diagonal matrix threadgroup k_t * sk = (threadgroup k_t *) (shmem_f16 + sgitg*(4*16*KV) + Q*T); // scratch buffer to load K in shared memory threadgroup k4x4_t * sk4x4 = (threadgroup k4x4_t *) (shmem_f16 + sgitg*(4*16*KV) + Q*T); // same as above but in k4x4_t @@ -3131,23 +3154,23 @@ kernel void kernel_flash_attn_ext( threadgroup v4x4_t * sv4x4 = (threadgroup v4x4_t *) (shmem_f16 + sgitg*(4*16*KV) + Q*T); // same as above but in v4x4_t // store the result for all queries in local memory in 8x8 matrices (the O matrix from the paper) - o8x8_t lo[D8]; + o8x8_t lo[DV8]; // load heads from Q to shared memory for (short j = sgitg; j < Q; j += nsg) { device const float4 * q4 = (device const float4 *) ((device const char *) q + ((iq1 + j)*args.nb01 + iq2*args.nb02 + iq3*args.nb03)); - for (short i = tiisg; i < D4; i += NW) { + for (short i = tiisg; i < DK4; i += NW) { if (iq1 + j < args.ne01) { - sq4[j*D4 + i] = (q4_t) q4[i]; + sq4[j*DK4 + i] = (q4_t) q4[i]; } else { - sq4[j*D4 + i] = (q4_t) 0.0f; + sq4[j*DK4 + i] = (q4_t) 0.0f; } } } // zero out lo - for (short i = 0; i < D8; ++i) { + for (short i = 0; i < DV8; ++i) { lo[i] = make_filled_simdgroup_matrix((o_t) 0.0f); } @@ -3161,8 +3184,8 @@ kernel void kernel_flash_attn_ext( threadgroup_barrier(mem_flags::mem_threadgroup); { - half S[Q] = { [0 ... Q-1] = 0.0f }; - half M[Q] = { [0 ... Q-1] = -__FLT16_MAX__/2 }; + float S[Q] = { [0 ... Q-1] = 0.0f }; + float M[Q] = { [0 ... Q-1] = -__FLT16_MAX__/2 }; // thread indices inside the simdgroup // TODO: see if we can utilize quad-group functions for better performance @@ -3177,22 +3200,15 @@ kernel void kernel_flash_attn_ext( const short ikv2 = iq2/(args.ne02/args.ne_12_2); const short ikv3 = iq3/(args.ne03/args.ne_12_3); - // load the queries from shared memory into local memory - q8x8_t mq[D8]; - - for (short i = 0; i < D8; ++i) { - simdgroup_load(mq[i], sq + i*8, D); - } - const bool has_mask = mask != q; - half slope = 1.0f; + float slope = 1.0f; // ALiBi if (args.max_bias > 0.0f) { const short h = iq2; - const half base = h < args.n_head_log2 ? args.m0 : args.m1; + const float base = h < args.n_head_log2 ? args.m0 : args.m1; const short exph = h < args.n_head_log2 ? h + 1 : 2*(h - args.n_head_log2) + 1; slope = pow(base, exph); @@ -3208,14 +3224,14 @@ kernel void kernel_flash_attn_ext( if (has_mask) { // used to detect blocks full of -INF - half smax = -INFINITY; + float smax = -INFINITY; // load the mask in shared memory #pragma unroll(Q) for (short j = 0; j < Q; ++j) { device const half * pm = (device const half *) ((device const char *) mask + (iq1 + j)*args.nb31); - const half m = pm[ic + tiisg]; + const float m = pm[ic + tiisg]; ss[j*TS + C + tiisg] = m; smax = max(smax, m); @@ -3236,20 +3252,22 @@ kernel void kernel_flash_attn_ext( // this is compile-time check, so it does not have runtime overhead if (is_same::value) { // we can read directly from global memory - device const k_t * pk = (device const k_t *) ((device const char *) k + ((ic + 8*cc)*args.nb_12_1 + ikv2*args.nb_12_2 + ikv3*args.nb_12_3)); + device const k_t * pk = (device const k_t *) ((device const char *) k + ((ic + 8*cc)*args.nb11 + ikv2*args.nb12 + ikv3*args.nb13)); - #pragma unroll(D8) - for (short i = 0; i < D8; ++i) { + #pragma unroll(DK8) + for (short i = 0; i < DK8; ++i) { k8x8_t mk; - simdgroup_load(mk, pk + i*8, args.nb_12_1/sizeof(k_t), 0, true); // transpose // TODO: use ne10 + simdgroup_load(mk, pk + i*8, args.nb11/sizeof(k_t), 0, true); // transpose // TODO: use ne10 - simdgroup_multiply_accumulate(mqk, mq[i], mk, mqk); + q8x8_t mq; + simdgroup_load(mq, sq + i*8, DK); + simdgroup_multiply_accumulate(mqk, mq, mk, mqk); } } else { - for (short ii = 0; ii < D16; ii += 4) { - device const kd4x4_t * pk4x4 = (device const kd4x4_t *) ((device const char *) k + ((ic + 8*cc + ty)*args.nb_12_1 + ikv2*args.nb_12_2 + ikv3*args.nb_12_3)); + for (short ii = 0; ii < DK16; ii += 4) { + device const kd4x4_t * pk4x4 = (device const kd4x4_t *) ((device const char *) k + ((ic + 8*cc + ty)*args.nb11 + ikv2*args.nb12 + ikv3*args.nb13)); - if (D16%4 == 0) { + if (DK16%4 == 0) { // the head is evenly divisible by 4*16 = 64, so no need for bound checks { k4x4_t tmp; @@ -3262,15 +3280,18 @@ kernel void kernel_flash_attn_ext( #pragma unroll(4) for (short k = 0; k < 4; ++k) { k8x8_t mk; + q8x8_t mq; simdgroup_load(mk, sk + 16*k + 0*8, 4*16, 0, true); // transpose - simdgroup_multiply_accumulate(mqk, mq[2*(ii + k) + 0], mk, mqk); + simdgroup_load(mq, sq + (2*(ii + k) + 0)*8, DK); + simdgroup_multiply_accumulate(mqk, mq, mk, mqk); simdgroup_load(mk, sk + 16*k + 1*8, 4*16, 0, true); // transpose - simdgroup_multiply_accumulate(mqk, mq[2*(ii + k) + 1], mk, mqk); + simdgroup_load(mq, sq + (2*(ii + k) + 1)*8, DK); + simdgroup_multiply_accumulate(mqk, mq, mk, mqk); } } else { - if (ii + tx < D16) { + if (ii + tx < DK16) { k4x4_t tmp; deq_k(pk4x4 + (ii + tx)/nl_k, (ii + tx)%nl_k, tmp); sk4x4[4*ty + tx] = tmp; @@ -3278,14 +3299,17 @@ kernel void kernel_flash_attn_ext( simdgroup_barrier(mem_flags::mem_threadgroup); - for (short k = 0; k < 4 && ii + k < D16; ++k) { + for (short k = 0; k < 4 && ii + k < DK16; ++k) { k8x8_t mk; + q8x8_t mq; simdgroup_load(mk, sk + 16*k + 0*8, 4*16, 0, true); // transpose - simdgroup_multiply_accumulate(mqk, mq[2*(ii + k) + 0], mk, mqk); + simdgroup_load(mq, sq + (2*(ii + k) + 0)*8, DK); + simdgroup_multiply_accumulate(mqk, mq, mk, mqk); simdgroup_load(mk, sk + 16*k + 1*8, 4*16, 0, true); // transpose - simdgroup_multiply_accumulate(mqk, mq[2*(ii + k) + 1], mk, mqk); + simdgroup_load(mq, sq + (2*(ii + k) + 1)*8, DK); + simdgroup_multiply_accumulate(mqk, mq, mk, mqk); } } } @@ -3303,10 +3327,10 @@ kernel void kernel_flash_attn_ext( // online softmax { for (ushort j = 0; j < Q; ++j) { - const half m = M[j]; + const float m = M[j]; // scale and apply the logitcap / mask - half s = ss[j*TS + tiisg]*args.scale; + float s = ss[j*TS + tiisg]*args.scale; if (args.logit_softcap != 0.0f) { s = args.logit_softcap*precise::tanh(s); @@ -3317,8 +3341,8 @@ kernel void kernel_flash_attn_ext( M[j] = simd_max(max(M[j], s)); - const half ms = exp(m - M[j]); - const half vs = exp(s - M[j]); + const float ms = exp(m - M[j]); + const float vs = exp(s - M[j]); S[j] = S[j]*ms + simd_sum(vs); @@ -3337,8 +3361,8 @@ kernel void kernel_flash_attn_ext( s8x8_t mm; simdgroup_load(mm, ss + 2*C, TS, 0, false); - #pragma unroll(D8) - for (short i = 0; i < D8; ++i) { + #pragma unroll(DV8) + for (short i = 0; i < DV8; ++i) { simdgroup_multiply(lo[i], mm, lo[i]); } } @@ -3351,20 +3375,20 @@ kernel void kernel_flash_attn_ext( if (is_same::value) { // we can read directly from global memory - device const v_t * pv = (device const v_t *) ((device const char *) v + ((ic + 8*cc)*args.nb_12_1 + ikv2*args.nb_12_2 + ikv3*args.nb_12_3)); + device const v_t * pv = (device const v_t *) ((device const char *) v + ((ic + 8*cc)*args.nb21 + ikv2*args.nb22 + ikv3*args.nb23)); - #pragma unroll(D8) - for (short i = 0; i < D8; ++i) { + #pragma unroll(DV8) + for (short i = 0; i < DV8; ++i) { v8x8_t mv; - simdgroup_load(mv, pv + i*8, args.nb_12_1/sizeof(v_t), 0, false); // TODO: use ne20 + simdgroup_load(mv, pv + i*8, args.nb21/sizeof(v_t), 0, false); // TODO: use ne20 simdgroup_multiply_accumulate(lo[i], ms, mv, lo[i]); } } else { - for (short ii = 0; ii < D16; ii += 4) { - device const vd4x4_t * pv4x4 = (device const vd4x4_t *) ((device const char *) v + ((ic + 8*cc + ty)*args.nb_12_1 + ikv2*args.nb_12_2 + ikv3*args.nb_12_3)); + for (short ii = 0; ii < DV16; ii += 4) { + device const vd4x4_t * pv4x4 = (device const vd4x4_t *) ((device const char *) v + ((ic + 8*cc + ty)*args.nb21 + ikv2*args.nb22 + ikv3*args.nb23)); - if (D16%4 == 0) { + if (DV16%4 == 0) { // no need for bound checks { v4x4_t tmp; @@ -3385,7 +3409,7 @@ kernel void kernel_flash_attn_ext( simdgroup_multiply_accumulate(lo[2*(ii + k) + 1], ms, mv, lo[2*(ii + k) + 1]); } } else { - if (ii + tx < D16) { + if (ii + tx < DV16) { v4x4_t tmp; deq_v(pv4x4 + (ii + tx)/nl_v, (ii + tx)%nl_v, tmp); sv4x4[4*ty + tx] = tmp; @@ -3393,7 +3417,7 @@ kernel void kernel_flash_attn_ext( simdgroup_barrier(mem_flags::mem_threadgroup); - for (short k = 0; k < 4 && ii + k < D16; ++k) { + for (short k = 0; k < 4 && ii + k < DV16; ++k) { v8x8_t mv; simdgroup_load(mv, sv + 16*k + 0*8, 4*16, 0, false); @@ -3420,15 +3444,15 @@ kernel void kernel_flash_attn_ext( // reduce the warps sequentially for (ushort sg = 1; sg < nsg; ++sg) { - half S = { 0.0f }; - half M = { -__FLT16_MAX__/2 }; + float S = { 0.0f }; + float M = { -__FLT16_MAX__/2 }; threadgroup_barrier(mem_flags::mem_threadgroup); // each simdgroup stores its output to shared memory, reusing sq if (sgitg == sg) { - for (short i = 0; i < D8; ++i) { - simdgroup_store(lo[i], so + i*8, D, 0, false); + for (short i = 0; i < DV8; ++i) { + simdgroup_store(lo[i], so + i*8, DV, 0, false); } } @@ -3437,16 +3461,16 @@ kernel void kernel_flash_attn_ext( // the first simdgroup accumulates the results from the other simdgroups if (sgitg == 0) { for (short j = 0; j < Q; ++j) { - const half S0 = ss[j*TS + 0]; - const half S1 = ss[j*TS + sg*SH + 0]; + const float S0 = ss[j*TS + 0]; + const float S1 = ss[j*TS + sg*SH + 0]; - const half M0 = ss[j*TS + 1]; - const half M1 = ss[j*TS + sg*SH + 1]; + const float M0 = ss[j*TS + 1]; + const float M1 = ss[j*TS + sg*SH + 1]; M = max(M0, M1); - const half ms0 = exp(M0 - M); - const half ms1 = exp(M1 - M); + const float ms0 = exp(M0 - M); + const float ms1 = exp(M1 - M); S = S0*ms0 + S1*ms1; @@ -3467,11 +3491,11 @@ kernel void kernel_flash_attn_ext( simdgroup_load(ms0, ss + 2*C, TS, 0, false); simdgroup_load(ms1, ss + 2*C + sg*SH, TS, 0, false); - #pragma unroll(D8) - for (short i = 0; i < D8; ++i) { + #pragma unroll(DV8) + for (short i = 0; i < DV8; ++i) { o8x8_t t; - simdgroup_load (t, so + i*8, D, 0, false); + simdgroup_load (t, so + i*8, DV, 0, false); simdgroup_multiply(t, ms1, t); simdgroup_multiply_accumulate(lo[i], ms0, lo[i], t); @@ -3482,8 +3506,8 @@ kernel void kernel_flash_attn_ext( // store result to shared memory (reuse sq) if (sgitg == 0) { - for (short i = 0; i < D8; ++i) { - simdgroup_store(lo[i], so + i*8, D, 0, false); + for (short i = 0; i < DV8; ++i) { + simdgroup_store(lo[i], so + i*8, DV, 0, false); } } @@ -3494,8 +3518,8 @@ kernel void kernel_flash_attn_ext( for (short j = 0; j < Q && iq1 + j < args.ne01; ++j) { const float S = ss[j*TS + 0]; - for (short i = tiisg; i < D4; i += NW) { - dst4[((uint64_t)iq3*args.ne2*args.ne1 + iq2 + (uint64_t)(iq1 + j)*args.ne1)*D4 + i] = (float4) so4[j*D4 + i]/S; + for (short i = tiisg; i < DV4; i += NW) { + dst4[((uint64_t)iq3*args.ne2*args.ne1 + iq2 + (uint64_t)(iq1 + j)*args.ne1)*DV4 + i] = (float4) so4[j*DV4 + i]/S; } } } @@ -3512,80 +3536,94 @@ kernel void kernel_flash_attn_ext( float, simdgroup_float8x8, \ half, half4, simdgroup_half8x8 -typedef decltype(kernel_flash_attn_ext) flash_attn_ext_t; +typedef decltype(kernel_flash_attn_ext) flash_attn_ext_t; -template [[host_name("kernel_flash_attn_ext_f16_h64" )]] kernel flash_attn_ext_t kernel_flash_attn_ext; -template [[host_name("kernel_flash_attn_ext_f16_h80" )]] kernel flash_attn_ext_t kernel_flash_attn_ext; -template [[host_name("kernel_flash_attn_ext_f16_h96" )]] kernel flash_attn_ext_t kernel_flash_attn_ext; -template [[host_name("kernel_flash_attn_ext_f16_h112")]] kernel flash_attn_ext_t kernel_flash_attn_ext; -template [[host_name("kernel_flash_attn_ext_f16_h128")]] kernel flash_attn_ext_t kernel_flash_attn_ext; -template [[host_name("kernel_flash_attn_ext_f16_h256")]] kernel flash_attn_ext_t kernel_flash_attn_ext; +template [[host_name("kernel_flash_attn_ext_f16_h64" )]] kernel flash_attn_ext_t kernel_flash_attn_ext; +template [[host_name("kernel_flash_attn_ext_f16_h80" )]] kernel flash_attn_ext_t kernel_flash_attn_ext; +template [[host_name("kernel_flash_attn_ext_f16_h96" )]] kernel flash_attn_ext_t kernel_flash_attn_ext; +template [[host_name("kernel_flash_attn_ext_f16_h112")]] kernel flash_attn_ext_t kernel_flash_attn_ext; +template [[host_name("kernel_flash_attn_ext_f16_h128")]] kernel flash_attn_ext_t kernel_flash_attn_ext; +template [[host_name("kernel_flash_attn_ext_f16_h192")]] kernel flash_attn_ext_t kernel_flash_attn_ext; +template [[host_name("kernel_flash_attn_ext_f16_hk192_hv128")]] kernel flash_attn_ext_t kernel_flash_attn_ext; +template [[host_name("kernel_flash_attn_ext_f16_h256")]] kernel flash_attn_ext_t kernel_flash_attn_ext; #if defined(GGML_METAL_USE_BF16) -template [[host_name("kernel_flash_attn_ext_bf16_h64" )]] kernel flash_attn_ext_t kernel_flash_attn_ext; -template [[host_name("kernel_flash_attn_ext_bf16_h80" )]] kernel flash_attn_ext_t kernel_flash_attn_ext; -template [[host_name("kernel_flash_attn_ext_bf16_h96" )]] kernel flash_attn_ext_t kernel_flash_attn_ext; -template [[host_name("kernel_flash_attn_ext_bf16_h112")]] kernel flash_attn_ext_t kernel_flash_attn_ext; -template [[host_name("kernel_flash_attn_ext_bf16_h128")]] kernel flash_attn_ext_t kernel_flash_attn_ext; -template [[host_name("kernel_flash_attn_ext_bf16_h256")]] kernel flash_attn_ext_t kernel_flash_attn_ext; +template [[host_name("kernel_flash_attn_ext_bf16_h64" )]] kernel flash_attn_ext_t kernel_flash_attn_ext; +template [[host_name("kernel_flash_attn_ext_bf16_h80" )]] kernel flash_attn_ext_t kernel_flash_attn_ext; +template [[host_name("kernel_flash_attn_ext_bf16_h96" )]] kernel flash_attn_ext_t kernel_flash_attn_ext; +template [[host_name("kernel_flash_attn_ext_bf16_h112")]] kernel flash_attn_ext_t kernel_flash_attn_ext; +template [[host_name("kernel_flash_attn_ext_bf16_h128")]] kernel flash_attn_ext_t kernel_flash_attn_ext; +template [[host_name("kernel_flash_attn_ext_bf16_h192")]] kernel flash_attn_ext_t kernel_flash_attn_ext; +template [[host_name("kernel_flash_attn_ext_bf16_hk192_hv128")]] kernel flash_attn_ext_t kernel_flash_attn_ext; +template [[host_name("kernel_flash_attn_ext_bf16_h256")]] kernel flash_attn_ext_t kernel_flash_attn_ext; #endif -template [[host_name("kernel_flash_attn_ext_q4_0_h64" )]] kernel flash_attn_ext_t kernel_flash_attn_ext; -template [[host_name("kernel_flash_attn_ext_q4_0_h80" )]] kernel flash_attn_ext_t kernel_flash_attn_ext; -template [[host_name("kernel_flash_attn_ext_q4_0_h96" )]] kernel flash_attn_ext_t kernel_flash_attn_ext; -template [[host_name("kernel_flash_attn_ext_q4_0_h112")]] kernel flash_attn_ext_t kernel_flash_attn_ext; -template [[host_name("kernel_flash_attn_ext_q4_0_h128")]] kernel flash_attn_ext_t kernel_flash_attn_ext; -template [[host_name("kernel_flash_attn_ext_q4_0_h256")]] kernel flash_attn_ext_t kernel_flash_attn_ext; +template [[host_name("kernel_flash_attn_ext_q4_0_h64" )]] kernel flash_attn_ext_t kernel_flash_attn_ext; +template [[host_name("kernel_flash_attn_ext_q4_0_h80" )]] kernel flash_attn_ext_t kernel_flash_attn_ext; +template [[host_name("kernel_flash_attn_ext_q4_0_h96" )]] kernel flash_attn_ext_t kernel_flash_attn_ext; +template [[host_name("kernel_flash_attn_ext_q4_0_h112")]] kernel flash_attn_ext_t kernel_flash_attn_ext; +template [[host_name("kernel_flash_attn_ext_q4_0_h128")]] kernel flash_attn_ext_t kernel_flash_attn_ext; +template [[host_name("kernel_flash_attn_ext_q4_0_h192")]] kernel flash_attn_ext_t kernel_flash_attn_ext; +template [[host_name("kernel_flash_attn_ext_q4_0_hk192_hv128")]] kernel flash_attn_ext_t kernel_flash_attn_ext; +template [[host_name("kernel_flash_attn_ext_q4_0_h256")]] kernel flash_attn_ext_t kernel_flash_attn_ext; -template [[host_name("kernel_flash_attn_ext_q4_1_h64" )]] kernel flash_attn_ext_t kernel_flash_attn_ext; -template [[host_name("kernel_flash_attn_ext_q4_1_h80" )]] kernel flash_attn_ext_t kernel_flash_attn_ext; -template [[host_name("kernel_flash_attn_ext_q4_1_h96" )]] kernel flash_attn_ext_t kernel_flash_attn_ext; -template [[host_name("kernel_flash_attn_ext_q4_1_h112")]] kernel flash_attn_ext_t kernel_flash_attn_ext; -template [[host_name("kernel_flash_attn_ext_q4_1_h128")]] kernel flash_attn_ext_t kernel_flash_attn_ext; -template [[host_name("kernel_flash_attn_ext_q4_1_h256")]] kernel flash_attn_ext_t kernel_flash_attn_ext; +template [[host_name("kernel_flash_attn_ext_q4_1_h64" )]] kernel flash_attn_ext_t kernel_flash_attn_ext; +template [[host_name("kernel_flash_attn_ext_q4_1_h80" )]] kernel flash_attn_ext_t kernel_flash_attn_ext; +template [[host_name("kernel_flash_attn_ext_q4_1_h96" )]] kernel flash_attn_ext_t kernel_flash_attn_ext; +template [[host_name("kernel_flash_attn_ext_q4_1_h112")]] kernel flash_attn_ext_t kernel_flash_attn_ext; +template [[host_name("kernel_flash_attn_ext_q4_1_h128")]] kernel flash_attn_ext_t kernel_flash_attn_ext; +template [[host_name("kernel_flash_attn_ext_q4_1_h192")]] kernel flash_attn_ext_t kernel_flash_attn_ext; +template [[host_name("kernel_flash_attn_ext_q4_1_hk192_hv128")]] kernel flash_attn_ext_t kernel_flash_attn_ext; +template [[host_name("kernel_flash_attn_ext_q4_1_h256")]] kernel flash_attn_ext_t kernel_flash_attn_ext; -template [[host_name("kernel_flash_attn_ext_q5_0_h64" )]] kernel flash_attn_ext_t kernel_flash_attn_ext; -template [[host_name("kernel_flash_attn_ext_q5_0_h80" )]] kernel flash_attn_ext_t kernel_flash_attn_ext; -template [[host_name("kernel_flash_attn_ext_q5_0_h96" )]] kernel flash_attn_ext_t kernel_flash_attn_ext; -template [[host_name("kernel_flash_attn_ext_q5_0_h112")]] kernel flash_attn_ext_t kernel_flash_attn_ext; -template [[host_name("kernel_flash_attn_ext_q5_0_h128")]] kernel flash_attn_ext_t kernel_flash_attn_ext; -template [[host_name("kernel_flash_attn_ext_q5_0_h256")]] kernel flash_attn_ext_t kernel_flash_attn_ext; +template [[host_name("kernel_flash_attn_ext_q5_0_h64" )]] kernel flash_attn_ext_t kernel_flash_attn_ext; +template [[host_name("kernel_flash_attn_ext_q5_0_h80" )]] kernel flash_attn_ext_t kernel_flash_attn_ext; +template [[host_name("kernel_flash_attn_ext_q5_0_h96" )]] kernel flash_attn_ext_t kernel_flash_attn_ext; +template [[host_name("kernel_flash_attn_ext_q5_0_h112")]] kernel flash_attn_ext_t kernel_flash_attn_ext; +template [[host_name("kernel_flash_attn_ext_q5_0_h128")]] kernel flash_attn_ext_t kernel_flash_attn_ext; +template [[host_name("kernel_flash_attn_ext_q5_0_h192")]] kernel flash_attn_ext_t kernel_flash_attn_ext; +template [[host_name("kernel_flash_attn_ext_q5_0_hk192_hv128")]] kernel flash_attn_ext_t kernel_flash_attn_ext; +template [[host_name("kernel_flash_attn_ext_q5_0_h256")]] kernel flash_attn_ext_t kernel_flash_attn_ext; -template [[host_name("kernel_flash_attn_ext_q5_1_h64" )]] kernel flash_attn_ext_t kernel_flash_attn_ext; -template [[host_name("kernel_flash_attn_ext_q5_1_h80" )]] kernel flash_attn_ext_t kernel_flash_attn_ext; -template [[host_name("kernel_flash_attn_ext_q5_1_h96" )]] kernel flash_attn_ext_t kernel_flash_attn_ext; -template [[host_name("kernel_flash_attn_ext_q5_1_h112")]] kernel flash_attn_ext_t kernel_flash_attn_ext; -template [[host_name("kernel_flash_attn_ext_q5_1_h128")]] kernel flash_attn_ext_t kernel_flash_attn_ext; -template [[host_name("kernel_flash_attn_ext_q5_1_h256")]] kernel flash_attn_ext_t kernel_flash_attn_ext; +template [[host_name("kernel_flash_attn_ext_q5_1_h64" )]] kernel flash_attn_ext_t kernel_flash_attn_ext; +template [[host_name("kernel_flash_attn_ext_q5_1_h80" )]] kernel flash_attn_ext_t kernel_flash_attn_ext; +template [[host_name("kernel_flash_attn_ext_q5_1_h96" )]] kernel flash_attn_ext_t kernel_flash_attn_ext; +template [[host_name("kernel_flash_attn_ext_q5_1_h112")]] kernel flash_attn_ext_t kernel_flash_attn_ext; +template [[host_name("kernel_flash_attn_ext_q5_1_h128")]] kernel flash_attn_ext_t kernel_flash_attn_ext; +template [[host_name("kernel_flash_attn_ext_q5_1_h192")]] kernel flash_attn_ext_t kernel_flash_attn_ext; +template [[host_name("kernel_flash_attn_ext_q5_1_hk192_hv128")]] kernel flash_attn_ext_t kernel_flash_attn_ext; +template [[host_name("kernel_flash_attn_ext_q5_1_h256")]] kernel flash_attn_ext_t kernel_flash_attn_ext; -template [[host_name("kernel_flash_attn_ext_q8_0_h64" )]] kernel flash_attn_ext_t kernel_flash_attn_ext; -template [[host_name("kernel_flash_attn_ext_q8_0_h80" )]] kernel flash_attn_ext_t kernel_flash_attn_ext; -template [[host_name("kernel_flash_attn_ext_q8_0_h96" )]] kernel flash_attn_ext_t kernel_flash_attn_ext; -template [[host_name("kernel_flash_attn_ext_q8_0_h112")]] kernel flash_attn_ext_t kernel_flash_attn_ext; -template [[host_name("kernel_flash_attn_ext_q8_0_h128")]] kernel flash_attn_ext_t kernel_flash_attn_ext; -template [[host_name("kernel_flash_attn_ext_q8_0_h256")]] kernel flash_attn_ext_t kernel_flash_attn_ext; +template [[host_name("kernel_flash_attn_ext_q8_0_h64" )]] kernel flash_attn_ext_t kernel_flash_attn_ext; +template [[host_name("kernel_flash_attn_ext_q8_0_h80" )]] kernel flash_attn_ext_t kernel_flash_attn_ext; +template [[host_name("kernel_flash_attn_ext_q8_0_h96" )]] kernel flash_attn_ext_t kernel_flash_attn_ext; +template [[host_name("kernel_flash_attn_ext_q8_0_h112")]] kernel flash_attn_ext_t kernel_flash_attn_ext; +template [[host_name("kernel_flash_attn_ext_q8_0_h128")]] kernel flash_attn_ext_t kernel_flash_attn_ext; +template [[host_name("kernel_flash_attn_ext_q8_0_h192")]] kernel flash_attn_ext_t kernel_flash_attn_ext; +template [[host_name("kernel_flash_attn_ext_q8_0_hk192_hv128")]] kernel flash_attn_ext_t kernel_flash_attn_ext; +template [[host_name("kernel_flash_attn_ext_q8_0_h256")]] kernel flash_attn_ext_t kernel_flash_attn_ext; #undef FA_TYPES template< - typename q4_t, // query types in shared memory - typename q4x4_t, - typename k4x4_t, // key types in shared memory - typename v4x4_t, // value types in shared memory - typename qk_t, // Q*K types - typename s_t, // soft-max types + typename q4_t, // query types in shared memory + typename k4_t, // key types in shared memory + typename v4_t, // value types in shared memory + typename qk_t, // Q*K types + typename s_t, // soft-max types typename s4_t, - typename s4x4_t, - typename o4x4_t, // attention accumulation types - typename kd4x4_t, // key type in device memory + typename o4_t, // attention accumulation types + typename kd4_t, // key type in device memory short nl_k, - void (*deq_k)(device const kd4x4_t *, short, thread k4x4_t &), - typename vd4x4_t, // key type in device memory + void (*deq_k_t4)(device const kd4_t *, short, thread k4_t &), + typename vd4_t, // key type in device memory short nl_v, - void (*deq_v)(device const vd4x4_t *, short, thread v4x4_t &), - short D, // head size - short Q = 1, // queries per threadgroup - short C = 32> // cache items per threadgroup + void (*deq_v_t4)(device const vd4_t *, short, thread v4_t &), + short DK, // K head size + short DV, // V head size + short NE = 4, // head elements per thread + short Q = 1, // queries per threadgroup + short C = 32> // cache items per threadgroup kernel void kernel_flash_attn_ext_vec( constant ggml_metal_kargs_flash_attn_ext & args, device const char * q, @@ -3604,29 +3642,28 @@ kernel void kernel_flash_attn_ext_vec( const int iq2 = tgpig[1]; const int iq1 = tgpig[0]; - const short D4 = D/4; - const short D16 = D/16; - const short NW = N_SIMDWIDTH; - const short NL = NW/4; // note: this can be adjusted to support D%64 == 0 and D%32 == 0 - const short SH = 2*C; // shared memory per simdgroup + constexpr short DK4 = DK/4; + constexpr short DV4 = DV/4; + constexpr short NW = N_SIMDWIDTH; + constexpr short NL = NW/NE; // note: this can be adjusted to support different head sizes and simdgroup work loads + constexpr short SH = 4*C; // shared memory per simdgroup - const short T = D + nsg*SH; // shared memory size per query in (half) + const short T = DK + nsg*SH; // shared memory size per query in (half) - //threadgroup q_t * sq = (threadgroup q_t *) (shmem_f16 + 0*D); // holds the query data - threadgroup q4_t * sq4 = (threadgroup q4_t *) (shmem_f16 + 0*D); // same as above but in q4_t - threadgroup q4x4_t * sq4x4 = (threadgroup q4x4_t *) (shmem_f16 + 0*D); // same as above but in q4x4_t - threadgroup s_t * ss = (threadgroup s_t *) (shmem_f16 + sgitg*SH + Q*D); // scratch buffer for attention - threadgroup s4_t * ss4 = (threadgroup s4_t *) (shmem_f16 + sgitg*SH + Q*D); // same as above but in s4_t - threadgroup half * sm = (threadgroup half *) (shmem_f16 + sgitg*SH + C + Q*D); // scratch buffer for mask - threadgroup o4x4_t * sr4x4 = (threadgroup o4x4_t *) (shmem_f16 + sgitg*D + Q*T); // scratch buffer for the results + //threadgroup q_t * sq = (threadgroup q_t *) (shmem_f16 + 0*DK); // holds the query data + threadgroup q4_t * sq4 = (threadgroup q4_t *) (shmem_f16 + 0*DK); // same as above but in q4_t + threadgroup s_t * ss = (threadgroup s_t *) (shmem_f16 + sgitg*SH + Q*DK); // scratch buffer for attention + threadgroup s4_t * ss4 = (threadgroup s4_t *) (shmem_f16 + sgitg*SH + Q*DK); // same as above but in s4_t + threadgroup float * sm = (threadgroup float *) (shmem_f16 + sgitg*SH + 2*C + Q*DK); // scratch buffer for mask + threadgroup o4_t * sr4 = (threadgroup o4_t *) (shmem_f16 + sgitg*DV + Q*T); // scratch buffer for the results - // store the result for all queries in local memory in 8x8 matrices (the O matrix from the paper) - o4x4_t lo[D16/NL]; + // store the result for all queries in local memory (the O matrix from the paper) + o4_t lo[DV4/NL]; // load heads from Q to shared memory device const float4 * q4 = (device const float4 *) ((device const char *) q + (iq1*args.nb01 + iq2*args.nb02 + iq3*args.nb03)); - for (short i = tiisg; i < D4; i += NW) { + for (short i = tiisg; i < DK4; i += NW) { if (iq1 < args.ne01) { sq4[i] = (q4_t) q4[i]; } else { @@ -3635,8 +3672,8 @@ kernel void kernel_flash_attn_ext_vec( } // zero out lo - for (short i = 0; i < D16/NL; ++i) { - lo[i] = (o4x4_t) 0.0f; + for (short i = 0; i < DV4/NL; ++i) { + lo[i] = (o4_t) 0.0f; } // zero out shared memory SH @@ -3647,8 +3684,8 @@ kernel void kernel_flash_attn_ext_vec( threadgroup_barrier(mem_flags::mem_threadgroup); { - half S = 0.0f; - half M = -__FLT16_MAX__/2; + float S = 0.0f; + float M = -__FLT16_MAX__/2; // thread indices inside the simdgroup const short tx = tiisg%NL; @@ -3661,26 +3698,18 @@ kernel void kernel_flash_attn_ext_vec( const short ikv2 = iq2/(args.ne02/args.ne_12_2); const short ikv3 = iq3/(args.ne03/args.ne_12_3); - // load the queries from shared memory into local memory - q4x4_t mq[D16/NL]; - - #pragma unroll(D16/NL) - for (short ii = 0; ii < D16; ii += NL) { - mq[ii/NL] = sq4x4[ii + tx]; - } - const bool has_mask = mask != q; // pointer to the mask device const half * pm = (device const half *) (mask + iq1*args.nb31); - half slope = 1.0f; + float slope = 1.0f; // ALiBi if (args.max_bias > 0.0f) { const short h = iq2; - const half base = h < args.n_head_log2 ? args.m0 : args.m1; + const float base = h < args.n_head_log2 ? args.m0 : args.m1; const short exph = h < args.n_head_log2 ? h + 1 : 2*(h - args.n_head_log2) + 1; slope = pow(base, exph); @@ -3700,43 +3729,56 @@ kernel void kernel_flash_attn_ext_vec( // Q*K^T { - // each simdgroup processes 1 query and 4 (NW/NL) keys - for (short cc = 0; cc < C/4; ++cc) { - qk_t mqka[4] = { 0.0, 0.0, 0.0, 0.0 }; + // each simdgroup processes 1 query and NE (NW/NL) head elements + for (short cc = 0; cc < C/NE; ++cc) { + qk_t mqk = 0.0f; - device const kd4x4_t * pk = (device const kd4x4_t *) ((device const char *) k + ((ic + 4*cc + ty)*args.nb_12_1 + ikv2*args.nb_12_2 + ikv3*args.nb_12_3)); + device const kd4_t * pk = (device const kd4_t *) ((device const char *) k + ((ic + NE*cc + ty)*args.nb11 + ikv2*args.nb12 + ikv3*args.nb13)); - #pragma unroll(D16/NL) - for (short ii = 0; ii < D16; ii += NL) { + #pragma unroll(DK4/NL) + for (short ii = 0; ii < DK4; ii += NL) { const short i = ii + tx; - k4x4_t mk; - deq_k(pk + i/nl_k, i%nl_k, mk); + k4_t mk; + deq_k_t4(pk + i/nl_k, i%nl_k, mk); // note: this is less precise than the version below - //mqka[0] += dot(mq[ii/NL][0], mk[0]); - //mqka[1] += dot(mq[ii/NL][1], mk[1]); - //mqka[2] += dot(mq[ii/NL][2], mk[2]); - //mqka[3] += dot(mq[ii/NL][3], mk[3]); + //mqka[0] += dot(mq[0], mk[0]); + //mqka[1] += dot(mq[1], mk[1]); + //mqka[2] += dot(mq[2], mk[2]); + //mqka[3] += dot(mq[3], mk[3]); - mqka[0] += dot((float4) mq[ii/NL][0], (float4) mk[0]); - mqka[1] += dot((float4) mq[ii/NL][1], (float4) mk[1]); - mqka[2] += dot((float4) mq[ii/NL][2], (float4) mk[2]); - mqka[3] += dot((float4) mq[ii/NL][3], (float4) mk[3]); + //q4x4_t mq = sq4x4[i]; + //mqka[0] += dot((float4) mq[0], (float4) mk[0]); + //mqka[1] += dot((float4) mq[1], (float4) mk[1]); + //mqka[2] += dot((float4) mq[2], (float4) mk[2]); + //mqka[3] += dot((float4) mq[3], (float4) mk[3]); + + mqk += dot((float4) mk, (float4) sq4[i]); } - qk_t mqk = mqka[0] + mqka[1] + mqka[2] + mqka[3]; + static_assert(NE > 1, "NE must be > 1"); // note: not sure why NE == 1 fails - // simdgroup reduce + // simdgroup reduce (NE = 4) // [ 0 .. 7] -> [ 0] // [ 8 .. 15] -> [ 8] // [16 .. 23] -> [16] // [24 .. 31] -> [24] - //mqk += simd_shuffle_down(mqk, 16); - //mqk += simd_shuffle_down(mqk, 8); - mqk += simd_shuffle_down(mqk, 4); - mqk += simd_shuffle_down(mqk, 2); - mqk += simd_shuffle_down(mqk, 1); + if (NE <= 1) { + mqk += simd_shuffle_down(mqk, 16); + } + if (NE <= 2) { + mqk += simd_shuffle_down(mqk, 8); + } + if (NE <= 4) { + mqk += simd_shuffle_down(mqk, 4); + } + if (NE <= 8) { + mqk += simd_shuffle_down(mqk, 2); + } + if (NE <= 16) { + mqk += simd_shuffle_down(mqk, 1); + } // mqk = mqk*scale + mask*slope if (tx == 0) { @@ -3746,9 +3788,9 @@ kernel void kernel_flash_attn_ext_vec( mqk = args.logit_softcap*precise::tanh(mqk); } - mqk += sm[4*cc + ty]*slope; + mqk += sm[NE*cc + ty]*slope; - ss[4*cc + ty] = mqk; + ss[NE*cc + ty] = mqk; } } } @@ -3757,13 +3799,13 @@ kernel void kernel_flash_attn_ext_vec( // online softmax { - const half m = M; - const half s = ss[tiisg]; + const float m = M; + const float s = ss[tiisg]; M = simd_max(max(M, s)); - const half ms = exp(m - M); - const half vs = exp(s - M); + const float ms = exp(m - M); + const float vs = exp(s - M); S = S*ms + simd_sum(vs); @@ -3771,8 +3813,8 @@ kernel void kernel_flash_attn_ext_vec( ss[tiisg] = vs; // O = diag(ms)*O - #pragma unroll(D16/NL) - for (short ii = 0; ii < D16; ii += NL) { + #pragma unroll(DV4/NL) + for (short ii = 0; ii < DV4; ii += NL) { lo[ii/NL] *= ms; } } @@ -3781,19 +3823,20 @@ kernel void kernel_flash_attn_ext_vec( // O = O + (Q*K^T)*V { - for (short cc = 0; cc < C/4; ++cc) { - device const vd4x4_t * pv4 = (device const vd4x4_t *) ((device const char *) v + ((ic + 4*cc + ty)*args.nb_12_1 + ikv2*args.nb_12_2 + ikv3*args.nb_12_3)); + //#pragma unroll(C/NE) + for (short cc = 0; cc < C/NE; ++cc) { + device const vd4_t * pv4 = (device const vd4_t *) ((device const char *) v + ((ic + NE*cc + ty)*args.nb21 + ikv2*args.nb22 + ikv3*args.nb23)); - const s4x4_t ms(ss[4*cc + ty]); + const s4_t ms(ss[NE*cc + ty]); - #pragma unroll(D16/NL) - for (short ii = 0; ii < D16; ii += NL) { + #pragma unroll(DV4/NL) + for (short ii = 0; ii < DV4; ii += NL) { const short i = ii + tx; - v4x4_t mv; - deq_v(pv4 + i/nl_v, i%nl_v, mv); + v4_t mv; + deq_v_t4(pv4 + i/nl_v, i%nl_v, mv); - lo[ii/NL] += mv*ms; + lo[ii/NL] += o4_t(float4(mv)*float4(ms)); } } } @@ -3806,7 +3849,7 @@ kernel void kernel_flash_attn_ext_vec( } } - // simdgroup reduce + // simdgroup reduce (NE = 4) // [ 0, 8, 16, 24] -> [ 0] // [ 1, 9, 17, 25] -> [ 1] // [ 2, 10, 18, 26] -> [ 2] @@ -3815,37 +3858,48 @@ kernel void kernel_flash_attn_ext_vec( // [ 5, 13, 21, 29] -> [ 5] // [ 6, 14, 22, 30] -> [ 6] // [ 7, 15, 23, 31] -> [ 7] - for (short ii = 0; ii < D16; ii += NL) { - lo[ii/NL][0] += simd_shuffle_down(lo[ii/NL][0], 16); - lo[ii/NL][0] += simd_shuffle_down(lo[ii/NL][0], 8); - //lo[ii/NL][0] += simd_shuffle_down(lo[ii/NL][0], 4); - //lo[ii/NL][0] += simd_shuffle_down(lo[ii/NL][0], 2); - //lo[ii/NL][0] += simd_shuffle_down(lo[ii/NL][0], 1); + for (short ii = 0; ii < DV4; ii += NL) { + if (NE > 1) { + lo[ii/NL][0] += simd_shuffle_down(lo[ii/NL][0], 16); + lo[ii/NL][1] += simd_shuffle_down(lo[ii/NL][1], 16); + lo[ii/NL][2] += simd_shuffle_down(lo[ii/NL][2], 16); + lo[ii/NL][3] += simd_shuffle_down(lo[ii/NL][3], 16); + } - lo[ii/NL][1] += simd_shuffle_down(lo[ii/NL][1], 16); - lo[ii/NL][1] += simd_shuffle_down(lo[ii/NL][1], 8); - //lo[ii/NL][1] += simd_shuffle_down(lo[ii/NL][1], 4); - //lo[ii/NL][1] += simd_shuffle_down(lo[ii/NL][1], 2); - //lo[ii/NL][1] += simd_shuffle_down(lo[ii/NL][1], 1); + if (NE > 2) { + lo[ii/NL][0] += simd_shuffle_down(lo[ii/NL][0], 8); + lo[ii/NL][1] += simd_shuffle_down(lo[ii/NL][1], 8); + lo[ii/NL][2] += simd_shuffle_down(lo[ii/NL][2], 8); + lo[ii/NL][3] += simd_shuffle_down(lo[ii/NL][3], 8); + } - lo[ii/NL][2] += simd_shuffle_down(lo[ii/NL][2], 16); - lo[ii/NL][2] += simd_shuffle_down(lo[ii/NL][2], 8); - //lo[ii/NL][2] += simd_shuffle_down(lo[ii/NL][2], 4); - //lo[ii/NL][2] += simd_shuffle_down(lo[ii/NL][2], 2); - //lo[ii/NL][2] += simd_shuffle_down(lo[ii/NL][2], 1); + if (NE > 4) { + lo[ii/NL][0] += simd_shuffle_down(lo[ii/NL][0], 4); + lo[ii/NL][1] += simd_shuffle_down(lo[ii/NL][1], 4); + lo[ii/NL][2] += simd_shuffle_down(lo[ii/NL][2], 4); + lo[ii/NL][3] += simd_shuffle_down(lo[ii/NL][3], 4); + } - lo[ii/NL][3] += simd_shuffle_down(lo[ii/NL][3], 16); - lo[ii/NL][3] += simd_shuffle_down(lo[ii/NL][3], 8); - //lo[ii/NL][3] += simd_shuffle_down(lo[ii/NL][3], 4); - //lo[ii/NL][3] += simd_shuffle_down(lo[ii/NL][3], 2); - //lo[ii/NL][3] += simd_shuffle_down(lo[ii/NL][3], 1); + if (NE > 8) { + lo[ii/NL][0] += simd_shuffle_down(lo[ii/NL][0], 2); + lo[ii/NL][1] += simd_shuffle_down(lo[ii/NL][1], 2); + lo[ii/NL][2] += simd_shuffle_down(lo[ii/NL][2], 2); + lo[ii/NL][3] += simd_shuffle_down(lo[ii/NL][3], 2); + } + + if (NE > 16) { + lo[ii/NL][0] += simd_shuffle_down(lo[ii/NL][0], 1); + lo[ii/NL][1] += simd_shuffle_down(lo[ii/NL][1], 1); + lo[ii/NL][2] += simd_shuffle_down(lo[ii/NL][2], 1); + lo[ii/NL][3] += simd_shuffle_down(lo[ii/NL][3], 1); + } } threadgroup_barrier(mem_flags::mem_threadgroup); // store results to shared memory - for (short i = tiisg; i < D16; i += NL) { - sr4x4[i] = lo[i/NL]; + for (short i = tiisg; i < DV4; i += NL) { + sr4[i] = lo[i/NL]; } threadgroup_barrier(mem_flags::mem_threadgroup); @@ -3853,18 +3907,18 @@ kernel void kernel_flash_attn_ext_vec( // parallel reduce for (short r = nsg/2; r > 0; r >>= 1) { if (sgitg < r) { - const half S0 = ss[ 0]; - const half S1 = ss[r*SH + 0]; + const float S0 = ss[ 0]; + const float S1 = ss[r*(SH/2) + 0]; - const half M0 = ss[ 1]; - const half M1 = ss[r*SH + 1]; + const float M0 = ss[ 1]; + const float M1 = ss[r*(SH/2) + 1]; - const half M = max(M0, M1); + const float M = max(M0, M1); - const half ms0 = exp(M0 - M); - const half ms1 = exp(M1 - M); + const float ms0 = exp(M0 - M); + const float ms1 = exp(M1 - M); - const half S = S0*ms0 + S1*ms1; + const float S = S0*ms0 + S1*ms1; if (tiisg == 0) { ss[0] = S; @@ -3872,22 +3926,22 @@ kernel void kernel_flash_attn_ext_vec( } // O_0 = diag(ms0)*O_0 + diag(ms1)*O_1 - for (short i = tiisg; i < D16; i += NW) { - sr4x4[i] = sr4x4[i]*ms0 + sr4x4[i + r*D16]*ms1; + for (short i = tiisg; i < DV4; i += NW) { + sr4[i] = sr4[i]*ms0 + sr4[i + r*DV4]*ms1; } } threadgroup_barrier(mem_flags::mem_threadgroup); } - device float4x4 * dst44 = (device float4x4 *) dst; + device float4 * dst4 = (device float4 *) dst; // final rescale with 1/S and store to global memory if (sgitg == 0) { const float S = ss[0]; - for (short i = tiisg; i < D16; i += NW) { - dst44[((uint64_t)iq3*args.ne2*args.ne1 + iq2 + (uint64_t)iq1*args.ne1)*D16 + i] = (float4x4) sr4x4[i]/S; + for (short i = tiisg; i < DV4; i += NW) { + dst4[((uint64_t)iq3*args.ne2*args.ne1 + iq2 + (uint64_t)iq1*args.ne1)*DV4 + i] = (float4) sr4[i]/S; } } } @@ -3896,34 +3950,54 @@ kernel void kernel_flash_attn_ext_vec( // in the other (non-vec) kernel, we need s_t to also be float because we scale during the soft_max // #define FA_TYPES \ - half4, half4x4, \ - half4x4, \ - half4x4, \ - float, \ - half, half4, half4x4, \ - half4x4 + half4, \ + half4, \ + half4, \ + float, \ + float, float4, \ + half4 -typedef decltype(kernel_flash_attn_ext_vec) flash_attn_ext_vec_t; +typedef decltype(kernel_flash_attn_ext_vec) flash_attn_ext_vec_t; -template [[host_name("kernel_flash_attn_ext_vec_f16_h128")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec; +template [[host_name("kernel_flash_attn_ext_vec_f16_h128")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec; #if defined(GGML_METAL_USE_BF16) -template [[host_name("kernel_flash_attn_ext_vec_bf16_h128")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec; +template [[host_name("kernel_flash_attn_ext_vec_bf16_h128")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec; #endif -template [[host_name("kernel_flash_attn_ext_vec_q4_0_h128")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec; -template [[host_name("kernel_flash_attn_ext_vec_q4_1_h128")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec; -template [[host_name("kernel_flash_attn_ext_vec_q5_0_h128")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec; -template [[host_name("kernel_flash_attn_ext_vec_q5_1_h128")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec; -template [[host_name("kernel_flash_attn_ext_vec_q8_0_h128")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec; +template [[host_name("kernel_flash_attn_ext_vec_q4_0_h128")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec; +template [[host_name("kernel_flash_attn_ext_vec_q4_1_h128")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec; +template [[host_name("kernel_flash_attn_ext_vec_q5_0_h128")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec; +template [[host_name("kernel_flash_attn_ext_vec_q5_1_h128")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec; +template [[host_name("kernel_flash_attn_ext_vec_q8_0_h128")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec; -template [[host_name("kernel_flash_attn_ext_vec_f16_h256")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec; +template [[host_name("kernel_flash_attn_ext_vec_f16_h192")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec; #if defined(GGML_METAL_USE_BF16) -template [[host_name("kernel_flash_attn_ext_vec_bf16_h256")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec; +template [[host_name("kernel_flash_attn_ext_vec_bf16_h192")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec; #endif -template [[host_name("kernel_flash_attn_ext_vec_q4_0_h256")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec; -template [[host_name("kernel_flash_attn_ext_vec_q4_1_h256")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec; -template [[host_name("kernel_flash_attn_ext_vec_q5_0_h256")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec; -template [[host_name("kernel_flash_attn_ext_vec_q5_1_h256")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec; -template [[host_name("kernel_flash_attn_ext_vec_q8_0_h256")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec; +template [[host_name("kernel_flash_attn_ext_vec_q4_0_h192")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec; +template [[host_name("kernel_flash_attn_ext_vec_q4_1_h192")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec; +template [[host_name("kernel_flash_attn_ext_vec_q5_0_h192")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec; +template [[host_name("kernel_flash_attn_ext_vec_q5_1_h192")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec; +template [[host_name("kernel_flash_attn_ext_vec_q8_0_h192")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec; + +template [[host_name("kernel_flash_attn_ext_vec_f16_hk192_hv128")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec; +#if defined(GGML_METAL_USE_BF16) +template [[host_name("kernel_flash_attn_ext_vec_bf16_hk192_hv128")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec; +#endif +template [[host_name("kernel_flash_attn_ext_vec_q4_0_hk192_hv128")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec; +template [[host_name("kernel_flash_attn_ext_vec_q4_1_hk192_hv128")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec; +template [[host_name("kernel_flash_attn_ext_vec_q5_0_hk192_hv128")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec; +template [[host_name("kernel_flash_attn_ext_vec_q5_1_hk192_hv128")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec; +template [[host_name("kernel_flash_attn_ext_vec_q8_0_hk192_hv128")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec; + +template [[host_name("kernel_flash_attn_ext_vec_f16_h256")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec; +#if defined(GGML_METAL_USE_BF16) +template [[host_name("kernel_flash_attn_ext_vec_bf16_h256")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec; +#endif +template [[host_name("kernel_flash_attn_ext_vec_q4_0_h256")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec; +template [[host_name("kernel_flash_attn_ext_vec_q4_1_h256")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec; +template [[host_name("kernel_flash_attn_ext_vec_q5_0_h256")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec; +template [[host_name("kernel_flash_attn_ext_vec_q5_1_h256")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec; +template [[host_name("kernel_flash_attn_ext_vec_q8_0_h256")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec; #undef FA_TYPES @@ -4298,7 +4372,7 @@ kernel void kernel_cpy_f32_iq4_nl( float amax = 0.0f; // absolute max float max = 0.0f; - for (int j = 0; j < QK4_0; j++) { + for (int j = 0; j < QK4_NL; j++) { const float v = src[j]; if (amax < fabs(v)) { amax = fabs(v); @@ -4332,6 +4406,49 @@ kernel void kernel_cpy_f32_iq4_nl( } } +template +kernel void kernel_cpy_q_f32( + constant ggml_metal_kargs_cpy & args, + device const char * src0, + device char * dst, + uint3 tgpig[[threadgroup_position_in_grid]], + ushort3 tpitg[[thread_position_in_threadgroup]], + ushort3 ntg[[threads_per_threadgroup]]) { + const int i03 = tgpig[2]; + const int i02 = tgpig[1]; + const int i01 = tgpig[0]; + + const int64_t n = i03*args.ne02*args.ne01*args.ne00 + i02*args.ne01*args.ne00 + i01*args.ne00; + + const int64_t i3 = n/(args.ne2*args.ne1*args.ne0); + const int64_t i2 = (n - i3*args.ne2*args.ne1*args.ne0)/(args.ne1*args.ne0); + const int64_t i1 = (n - i3*args.ne2*args.ne1*args.ne0 - i2*args.ne1*args.ne0)/args.ne0; + const int64_t i0 = (n - i3*args.ne2*args.ne1*args.ne0 - i2*args.ne1*args.ne0 - i1*args.ne0); + + device const block_q * src_data = (device const block_q *)(src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01); + device T4x4 * dst_data = (device T4x4 *)(dst + i3*args.nb3 + i2*args.nb2 + i1*args.nb1 + i0*args.nb0); + + for (int64_t i00 = tpitg.x; i00 < args.ne00/16; i00 += ntg.x) { + T4x4 temp; + dequantize_func(src_data + i00/nl, i00%nl, temp); + dst_data[i00] = temp; + } +} + +typedef decltype(kernel_cpy_q_f32) cpy_q_f_t; + +template [[host_name("kernel_cpy_q4_0_f32")]] kernel cpy_q_f_t kernel_cpy_q_f32; +template [[host_name("kernel_cpy_q4_1_f32")]] kernel cpy_q_f_t kernel_cpy_q_f32; +template [[host_name("kernel_cpy_q5_0_f32")]] kernel cpy_q_f_t kernel_cpy_q_f32; +template [[host_name("kernel_cpy_q5_1_f32")]] kernel cpy_q_f_t kernel_cpy_q_f32; +template [[host_name("kernel_cpy_q8_0_f32")]] kernel cpy_q_f_t kernel_cpy_q_f32; + +template [[host_name("kernel_cpy_q4_0_f16")]] kernel cpy_q_f_t kernel_cpy_q_f32; +template [[host_name("kernel_cpy_q4_1_f16")]] kernel cpy_q_f_t kernel_cpy_q_f32; +template [[host_name("kernel_cpy_q5_0_f16")]] kernel cpy_q_f_t kernel_cpy_q_f32; +template [[host_name("kernel_cpy_q5_1_f16")]] kernel cpy_q_f_t kernel_cpy_q_f32; +template [[host_name("kernel_cpy_q8_0_f16")]] kernel cpy_q_f_t kernel_cpy_q_f32; + kernel void kernel_concat( constant ggml_metal_kargs_concat & args, device const char * src0, @@ -4363,7 +4480,7 @@ kernel void kernel_concat( } } -template +template void kernel_mul_mv_q2_K_f32_impl( args_t args, device const char * src0, @@ -4379,7 +4496,7 @@ void kernel_mul_mv_q2_K_f32_impl( const int r1 = tgpig.y; const int im = tgpig.z; - const int first_row = (r0 * N_SIMDGROUP + sgitg) * N_DST; + const int first_row = (r0 * nsg + sgitg) * nr0; const uint i12 = im%args.ne12; const uint i13 = im/args.ne12; @@ -4391,20 +4508,19 @@ void kernel_mul_mv_q2_K_f32_impl( device const float * y = (device const float *) (src1 + offset1); float yl[32]; - float sumf[N_DST]={0.f}, all_sum; + float sumf[nr0]={0.f}; - const int ix = tiisg/8; // 0...3 - const int it = tiisg%8; // 0...7 - const int iq = it/4; // 0 or 1 - const int ir = it%4; // 0...3 - const int is = (8*ir)/16;// 0 or 1 + const short ix = tiisg/8; // 0...3 + const short it = tiisg%8; // 0...7 + const short iq = it/4; // 0 or 1 + const short ir = it%4; // 0...3 + const short is = (8*ir)/16;// 0 or 1 device const float * y4 = y + ix * QK_K + 128 * iq + 8 * ir; for (int ib = ix; ib < nb; ib += 4) { - float4 sumy = {0.f, 0.f, 0.f, 0.f}; - for (int i = 0; i < 8; ++i) { + for (short i = 0; i < 8; ++i) { yl[i+ 0] = y4[i+ 0]; sumy[0] += yl[i+ 0]; yl[i+ 8] = y4[i+32]; sumy[1] += yl[i+ 8]; yl[i+16] = y4[i+64]; sumy[2] += yl[i+16]; @@ -4415,7 +4531,7 @@ void kernel_mul_mv_q2_K_f32_impl( device const uint16_t * qs = (device const uint16_t *)x[ib].qs + 16 * iq + 4 * ir; device const half * dh = &x[ib].d; - for (int row = 0; row < N_DST; row++) { + for (short row = 0; row < nr0; row++) { float4 acc1 = {0.f, 0.f, 0.f, 0.f}; float4 acc2 = {0.f, 0.f, 0.f, 0.f}; for (int i = 0; i < 8; i += 2) { @@ -4446,10 +4562,10 @@ void kernel_mul_mv_q2_K_f32_impl( device float * dst_f32 = (device float *) dst + (uint64_t)im*args.ne0*args.ne1 + (uint64_t)r1*args.ne0; - for (int row = 0; row < N_DST && first_row + row < args.ne0; ++row) { - all_sum = simd_sum(sumf[row]); + for (int row = 0; row < nr0 && first_row + row < args.ne0; ++row) { + float sum_all = simd_sum(sumf[row]); if (tiisg == 0) { - dst_f32[first_row + row] = all_sum; + dst_f32[first_row + row] = sum_all; } } } @@ -4464,10 +4580,10 @@ kernel void kernel_mul_mv_q2_K_f32( ushort tiisg[[thread_index_in_simdgroup]], ushort sgitg[[simdgroup_index_in_threadgroup]]) { - kernel_mul_mv_q2_K_f32_impl(args, src0, src1, dst, nullptr, tgpig, tiisg, sgitg); + kernel_mul_mv_q2_K_f32_impl(args, src0, src1, dst, nullptr, tgpig, tiisg, sgitg); } -template +template void kernel_mul_mv_q3_K_f32_impl( args_t args, device const char * src0, @@ -4484,7 +4600,7 @@ void kernel_mul_mv_q3_K_f32_impl( const int r1 = tgpig.y; const int im = tgpig.z; - const int first_row = (r0 * N_SIMDGROUP + sgitg) * 2; + const int first_row = (r0 * nsg + sgitg) * nr0; const uint i12 = im%args.ne12; const uint i13 = im/args.ne12; @@ -4500,13 +4616,12 @@ void kernel_mul_mv_q3_K_f32_impl( //const uint16_t kmask1 = 0x3030; //const uint16_t kmask2 = 0x0f0f; - const int tid = tiisg/4; - const int ix = tiisg%4; - const int ip = tid/4; // 0 or 1 - const int il = 2*((tid%4)/2); // 0 or 2 - const int ir = tid%2; - const int n = 8; - const int l0 = n*ir; + const short tid = tiisg/4; + const short ix = tiisg%4; + const short ip = tid/4; // 0 or 1 + const short il = 2*((tid%4)/2); // 0 or 2 + const short ir = tid%2; + const short l0 = 8*ir; // One would think that the Metal compiler would figure out that ip and il can only have // 4 possible states, and optimize accordingly. Well, no. It needs help, and we do it @@ -4531,8 +4646,8 @@ void kernel_mul_mv_q3_K_f32_impl( const uint16_t s_shift1 = 4*ip; const uint16_t s_shift2 = s_shift1 + il; - const int q_offset = 32*ip + l0; - const int y_offset = 128*ip + 32*il + l0; + const short q_offset = 32*ip + l0; + const short y_offset = 128*ip + 32*il + l0; device const float * y1 = yy + ix*QK_K + y_offset; @@ -4540,10 +4655,11 @@ void kernel_mul_mv_q3_K_f32_impl( thread uint16_t * scales16 = (thread uint16_t *)&scales32; thread const int8_t * scales = (thread const int8_t *)&scales32; - float sumf1[2] = {0.f}; - float sumf2[2] = {0.f}; + float sumf1[nr0] = {0.f}; + float sumf2[nr0] = {0.f}; + for (int i = ix; i < nb; i += 4) { - for (int l = 0; l < 8; ++l) { + for (short l = 0; l < 8; ++l) { yl[l+ 0] = y1[l+ 0]; yl[l+ 8] = y1[l+16]; yl[l+16] = y1[l+32]; @@ -4555,7 +4671,7 @@ void kernel_mul_mv_q3_K_f32_impl( device const uint16_t * a = (device const uint16_t *)(x[i].scales); device const half * dh = &x[i].d; - for (int row = 0; row < 2; ++row) { + for (short row = 0; row < nr0; ++row) { const float d_all = (float)dh[0]; scales16[0] = a[4]; @@ -4566,7 +4682,7 @@ void kernel_mul_mv_q3_K_f32_impl( scales32 = ((scales32 >> s_shift1) & 0x0f0f0f0f) | aux32; float s1 = 0, s2 = 0, s3 = 0, s4 = 0, s5 = 0, s6 = 0; - for (int l = 0; l < n; l += 2) { + for (short l = 0; l < 8; l += 2) { const int32_t qs = q[l/2]; s1 += yl[l+0] * (qs & qm[il/2][0]); s2 += yl[l+1] * (qs & qm[il/2][1]); @@ -4581,7 +4697,7 @@ void kernel_mul_mv_q3_K_f32_impl( sumf2[row] += d2 * (scales[2] - 32); s1 = s2 = s3 = s4 = s5 = s6 = 0; - for (int l = 0; l < n; l += 2) { + for (short l = 0; l < 8; l += 2) { const int32_t qs = q[l/2+8]; s1 += yl[l+8] * (qs & qm[il/2][0]); s2 += yl[l+9] * (qs & qm[il/2][1]); @@ -4604,7 +4720,7 @@ void kernel_mul_mv_q3_K_f32_impl( y1 += 4 * QK_K; } - for (int row = 0; row < 2; ++row) { + for (int row = 0; row < nr0; ++row) { const float sumf = (sumf1[row] + 0.25f * sumf2[row]) / (1 << shift); sumf1[row] = simd_sum(sumf); } @@ -4612,7 +4728,7 @@ void kernel_mul_mv_q3_K_f32_impl( device float * dst_f32 = (device float *) dst + (uint64_t)im*args.ne0*args.ne1 + (uint64_t)r1*args.ne0; if (tiisg == 0) { - for (int row = 0; row < 2 && first_row + row < args.ne0; ++row) { + for (int row = 0; row < nr0 && first_row + row < args.ne0; ++row) { dst_f32[first_row + row] = sumf1[row]; } } @@ -4628,10 +4744,10 @@ kernel void kernel_mul_mv_q3_K_f32( ushort tiisg[[thread_index_in_simdgroup]], ushort sgitg[[simdgroup_index_in_threadgroup]]) { - kernel_mul_mv_q3_K_f32_impl(args, src0, src1, dst, nullptr, tgpig, tiisg, sgitg); + kernel_mul_mv_q3_K_f32_impl(args, src0, src1, dst, nullptr, tgpig, tiisg, sgitg); } -template +template void kernel_mul_mv_q4_K_f32_impl( args_t args, device const char * src0, @@ -4641,22 +4757,22 @@ void kernel_mul_mv_q4_K_f32_impl( uint3 tgpig, ushort tiisg, ushort sgitg) { - const uint16_t kmask1 = 0x3f3f; const uint16_t kmask2 = 0x0f0f; const uint16_t kmask3 = 0xc0c0; - const int ix = tiisg/8; // 0...3 - const int it = tiisg%8; // 0...7 - const int iq = it/4; // 0 or 1 - const int ir = it%4; // 0...3 + const short ix = tiisg/8; // 0...3 + const short it = tiisg%8; // 0...7 + const short iq = it/4; // 0 or 1 + const short ir = it%4; // 0...3 const int nb = args.ne00/QK_K; + const int r0 = tgpig.x; const int r1 = tgpig.y; const int im = tgpig.z; - //const int first_row = (r0 * N_SIMDGROUP + sgitg) * N_DST; - const int first_row = r0 * N_DST; + + const int first_row = (r0 * nsg + sgitg) * nr0; const uint i12 = im%args.ne12; const uint i13 = im/args.ne12; @@ -4669,7 +4785,8 @@ void kernel_mul_mv_q4_K_f32_impl( float yl[16]; float yh[16]; - float sumf[N_DST]={0.f}, all_sum; + + float sumf[nr0]={0.f}; device const float * y4 = y + ix * QK_K + 64 * iq + 8 * ir; @@ -4678,7 +4795,8 @@ void kernel_mul_mv_q4_K_f32_impl( for (int ib = ix; ib < nb; ib += 4) { float4 sumy = {0.f, 0.f, 0.f, 0.f}; - for (int i = 0; i < 8; ++i) { + + for (short i = 0; i < 8; ++i) { yl[i+0] = y4[i+ 0]; sumy[0] += yl[i+0]; yl[i+8] = y4[i+ 32]; sumy[1] += yl[i+8]; yh[i+0] = y4[i+128]; sumy[2] += yh[i+0]; @@ -4689,7 +4807,7 @@ void kernel_mul_mv_q4_K_f32_impl( device const uint16_t * q1 = (device const uint16_t *)x[ib].qs + 16 * iq + 4 * ir; device const half * dh = &x[ib].d; - for (int row = 0; row < N_DST; row++) { + for (short row = 0; row < nr0; row++) { sc16[0] = sc[0] & kmask1; sc16[1] = sc[2] & kmask1; sc16[2] = ((sc[4] >> 0) & kmask2) | ((sc[0] & kmask3) >> 2); @@ -4699,19 +4817,21 @@ void kernel_mul_mv_q4_K_f32_impl( float4 acc1 = {0.f, 0.f, 0.f, 0.f}; float4 acc2 = {0.f, 0.f, 0.f, 0.f}; - for (int i = 0; i < 8; i += 2) { - acc1[0] += yl[i+0] * (q1[i/2] & 0x000F); - acc1[1] += yl[i+1] * (q1[i/2] & 0x0F00); - acc1[2] += yl[i+8] * (q1[i/2] & 0x00F0); - acc1[3] += yl[i+9] * (q1[i/2] & 0xF000); - acc2[0] += yh[i+0] * (q2[i/2] & 0x000F); - acc2[1] += yh[i+1] * (q2[i/2] & 0x0F00); - acc2[2] += yh[i+8] * (q2[i/2] & 0x00F0); - acc2[3] += yh[i+9] * (q2[i/2] & 0xF000); + + for (short i = 0; i < 4; ++i) { + acc1[0] += yl[2*i + 0] * (q1[i] & 0x000F); + acc1[1] += yl[2*i + 1] * (q1[i] & 0x0F00); + acc1[2] += yl[2*i + 8] * (q1[i] & 0x00F0); + acc1[3] += yl[2*i + 9] * (q1[i] & 0xF000); + acc2[0] += yh[2*i + 0] * (q2[i] & 0x000F); + acc2[1] += yh[2*i + 1] * (q2[i] & 0x0F00); + acc2[2] += yh[2*i + 8] * (q2[i] & 0x00F0); + acc2[3] += yh[2*i + 9] * (q2[i] & 0xF000); } float dall = dh[0]; float dmin = dh[1]; + sumf[row] += dall * ((acc1[0] + 1.f/256.f * acc1[1]) * sc8[0] + (acc1[2] + 1.f/256.f * acc1[3]) * sc8[1] * 1.f/16.f + (acc2[0] + 1.f/256.f * acc2[1]) * sc8[4] + @@ -4728,10 +4848,10 @@ void kernel_mul_mv_q4_K_f32_impl( device float * dst_f32 = (device float *) dst + (int64_t)im*args.ne0*args.ne1 + (int64_t)r1*args.ne0; - for (int row = 0; row < N_DST && first_row + row < args.ne0; ++row) { - all_sum = simd_sum(sumf[row]); + for (int row = 0; row < nr0 && first_row + row < args.ne0; ++row) { + float sum_all = simd_sum(sumf[row]); if (tiisg == 0) { - dst_f32[first_row + row] = all_sum; + dst_f32[first_row + row] = sum_all; } } } @@ -4746,10 +4866,10 @@ kernel void kernel_mul_mv_q4_K_f32( ushort tiisg[[thread_index_in_simdgroup]], ushort sgitg[[simdgroup_index_in_threadgroup]]) { - kernel_mul_mv_q4_K_f32_impl(args, src0, src1, dst, nullptr, tgpig, tiisg, sgitg); + kernel_mul_mv_q4_K_f32_impl(args, src0, src1, dst, nullptr, tgpig, tiisg, sgitg); } -template +template void kernel_mul_mv_q5_K_f32_impl( args_t args, device const char * src0, @@ -4766,7 +4886,7 @@ void kernel_mul_mv_q5_K_f32_impl( const int r1 = tgpig.y; const int im = tgpig.z; - const int first_row = (r0 * N_SIMDGROUP + sgitg) * 2; + const int first_row = (r0 * nsg + sgitg) * nr0; const uint i12 = im%args.ne12; const uint i13 = im/args.ne12; @@ -4777,7 +4897,7 @@ void kernel_mul_mv_q5_K_f32_impl( device const block_q5_K * x = (device const block_q5_K *) (src0 + offset0); device const float * yy = (device const float *) (src1 + offset1); - float sumf[2]={0.f}; + float sumf[nr0]={0.f}; float yl[16], yh[16]; @@ -4785,15 +4905,14 @@ void kernel_mul_mv_q5_K_f32_impl( const uint16_t kmask2 = 0x0f0f; const uint16_t kmask3 = 0xc0c0; - const int tid = tiisg/4; - const int ix = tiisg%4; - const int iq = tid/4; - const int ir = tid%4; - const int n = 8; + const short tid = tiisg/4; + const short ix = tiisg%4; + const short iq = tid/4; + const short ir = tid%4; - const int l0 = n*ir; - const int q_offset = 32*iq + l0; - const int y_offset = 64*iq + l0; + const short l0 = 8*ir; + const short q_offset = 32*iq + l0; + const short y_offset = 64*iq + l0; const uint8_t hm1 = 1u << (2*iq); const uint8_t hm2 = hm1 << 1; @@ -4813,14 +4932,14 @@ void kernel_mul_mv_q5_K_f32_impl( device const float * y2 = y1 + 128; float4 sumy = {0.f, 0.f, 0.f, 0.f}; - for (int l = 0; l < 8; ++l) { + for (short l = 0; l < 8; ++l) { yl[l+0] = y1[l+ 0]; sumy[0] += yl[l+0]; yl[l+8] = y1[l+32]; sumy[1] += yl[l+8]; yh[l+0] = y2[l+ 0]; sumy[2] += yh[l+0]; yh[l+8] = y2[l+32]; sumy[3] += yh[l+8]; } - for (int row = 0; row < 2; ++row) { + for (short row = 0; row < nr0; ++row) { device const uint8_t * q2 = q1 + 64; sc16[0] = a[0] & kmask1; @@ -4830,7 +4949,7 @@ void kernel_mul_mv_q5_K_f32_impl( float4 acc1 = {0.f}; float4 acc2 = {0.f}; - for (int l = 0; l < n; ++l) { + for (short l = 0; l < 8; ++l) { uint8_t h = qh[l]; acc1[0] += yl[l+0] * (q1[l] & 0x0F); acc1[1] += yl[l+8] * (q1[l] & 0xF0); @@ -4860,7 +4979,7 @@ void kernel_mul_mv_q5_K_f32_impl( device float * dst_f32 = (device float *) dst + (uint64_t)im*args.ne0*args.ne1 + (uint64_t)r1*args.ne0; - for (int row = 0; row < 2 && first_row + row < args.ne0; ++row) { + for (int row = 0; row < nr0 && first_row + row < args.ne0; ++row) { const float tot = simd_sum(sumf[row]); if (tiisg == 0) { dst_f32[first_row + row] = tot; @@ -4878,10 +4997,10 @@ kernel void kernel_mul_mv_q5_K_f32( ushort tiisg[[thread_index_in_simdgroup]], ushort sgitg[[simdgroup_index_in_threadgroup]]) { - kernel_mul_mv_q5_K_f32_impl(args, src0, src1, dst, nullptr, tgpig, tiisg, sgitg); + kernel_mul_mv_q5_K_f32_impl(args, src0, src1, dst, nullptr, tgpig, tiisg, sgitg); } -template +template void kernel_mul_mv_q6_K_f32_impl( args_t args, device const char * src0, @@ -4903,62 +5022,77 @@ void kernel_mul_mv_q6_K_f32_impl( const int r1 = tgpig.y; const int im = tgpig.z; - const int row = 2*r0 + sgitg; - - if (row >= args.ne0) { - return; - } + const int first_row = (r0 * nsg + sgitg) * nr0; const uint i12 = im%args.ne12; const uint i13 = im/args.ne12; - const uint64_t offset0 = row*args.nb01 + (i12/args.r2)*args.nb02 + (i13/args.r3)*args.nb03; - const uint64_t offset1 = r1*args.nb11 + (i12 )*args.nb12 + (i13 )*args.nb13; + const uint64_t offset0 = first_row*args.nb01 + (i12/args.r2)*args.nb02 + (i13/args.r3)*args.nb03; + const uint64_t offset1 = r1*args.nb11 + (i12 )*args.nb12 + (i13 )*args.nb13; device const block_q6_K * x = (device const block_q6_K *) (src0 + offset0); device const float * yy = (device const float *) (src1 + offset1); - float sumf = 0; + float sumf[nr0] = { 0.f }; - const int tid = tiisg/2; - const int ix = tiisg%2; - const int ip = tid/8; // 0 or 1 - const int il = tid%8; - const int n = 4; - const int l0 = n*il; - const int is = 8*ip + l0/16; + float yl[16]; - const int y_offset = 128*ip + l0; - const int q_offset_l = 64*ip + l0; - const int q_offset_h = 32*ip + l0; + const short tid = tiisg/2; + const short ix = tiisg%2; + const short ip = tid/8; // 0 or 1 + const short il = tid%8; + const short l0 = 4*il; + const short is = 8*ip + l0/16; + + const short y_offset = 128*ip + l0; + const short q_offset_l = 64*ip + l0; + const short q_offset_h = 32*ip + l0; for (int i = ix; i < nb; i += 2) { device const uint8_t * q1 = x[i].ql + q_offset_l; device const uint8_t * q2 = q1 + 32; device const uint8_t * qh = x[i].qh + q_offset_h; device const int8_t * sc = x[i].scales + is; + device const half * dh = &x[i].d; device const float * y = yy + i * QK_K + y_offset; - const float dall = x[i].d; - - float4 sums = {0.f, 0.f, 0.f, 0.f}; - for (int l = 0; l < n; ++l) { - sums[0] += y[l+ 0] * ((int8_t)((q1[l] & 0xF) | ((qh[l] & kmask1) << 4)) - 32); - sums[1] += y[l+32] * ((int8_t)((q2[l] & 0xF) | ((qh[l] & kmask2) << 2)) - 32); - sums[2] += y[l+64] * ((int8_t)((q1[l] >> 4) | ((qh[l] & kmask3) << 0)) - 32); - sums[3] += y[l+96] * ((int8_t)((q2[l] >> 4) | ((qh[l] & kmask4) >> 2)) - 32); + for (short l = 0; l < 4; ++l) { + yl[4*l + 0] = y[l + 0]; + yl[4*l + 1] = y[l + 32]; + yl[4*l + 2] = y[l + 64]; + yl[4*l + 3] = y[l + 96]; } - sumf += dall * (sums[0] * sc[0] + sums[1] * sc[2] + sums[2] * sc[4] + sums[3] * sc[6]); + for (short row = 0; row < nr0; ++row) { + const float dall = dh[0]; + float4 sums = {0.f, 0.f, 0.f, 0.f}; + + for (short l = 0; l < 4; ++l) { + sums[0] += yl[4*l + 0] * ((int8_t)((q1[l] & 0xF) | ((qh[l] & kmask1) << 4)) - 32); + sums[1] += yl[4*l + 1] * ((int8_t)((q2[l] & 0xF) | ((qh[l] & kmask2) << 2)) - 32); + sums[2] += yl[4*l + 2] * ((int8_t)((q1[l] >> 4) | ((qh[l] & kmask3) << 0)) - 32); + sums[3] += yl[4*l + 3] * ((int8_t)((q2[l] >> 4) | ((qh[l] & kmask4) >> 2)) - 32); + } + + sumf[row] += dall * (sums[0] * sc[0] + sums[1] * sc[2] + sums[2] * sc[4] + sums[3] * sc[6]); + + q1 += args.nb01; + q2 += args.nb01; + qh += args.nb01; + sc += args.nb01; + dh += args.nb01/2; + } } device float * dst_f32 = (device float *) dst + (uint64_t)im*args.ne0*args.ne1 + (uint64_t)r1*args.ne0; - const float tot = simd_sum(sumf); - if (tiisg == 0) { - dst_f32[row] = tot; + for (int row = 0; row < nr0 && first_row + row < args.ne0; ++row) { + float sum_all = simd_sum(sumf[row]); + if (tiisg == 0) { + dst_f32[first_row + row] = sum_all; + } } } @@ -4972,12 +5106,12 @@ kernel void kernel_mul_mv_q6_K_f32( ushort tiisg[[thread_index_in_simdgroup]], ushort sgitg[[simdgroup_index_in_threadgroup]]) { - kernel_mul_mv_q6_K_f32_impl(args, src0, src1, dst, nullptr, tgpig, tiisg, sgitg); + kernel_mul_mv_q6_K_f32_impl(args, src0, src1, dst, nullptr, tgpig, tiisg, sgitg); } // ======================= "True" 2-bit -template +template void kernel_mul_mv_iq2_xxs_f32_impl( args_t args, device const char * src0, @@ -4993,7 +5127,7 @@ void kernel_mul_mv_iq2_xxs_f32_impl( const int r1 = tgpig.y; const int im = tgpig.z; - const int first_row = (r0 * N_SIMDGROUP + sgitg) * N_DST; + const int first_row = (r0 * nsg + sgitg) * nr0; const uint i12 = im%args.ne12; const uint i13 = im/args.ne12; @@ -5005,7 +5139,7 @@ void kernel_mul_mv_iq2_xxs_f32_impl( device const float * y = (device const float *) (src1 + offset1); float yl[32]; - float sumf[N_DST]={0.f}, all_sum; + float sumf[nr0]={0.f}; const int nb32 = nb * (QK_K / 32); @@ -5026,8 +5160,7 @@ void kernel_mul_mv_iq2_xxs_f32_impl( device const float * y4 = y + 32 * ix; for (int ib32 = ix; ib32 < nb32; ib32 += 32) { - - for (int i = 0; i < 32; ++i) { + for (short i = 0; i < 32; ++i) { yl[i] = y4[i]; } @@ -5038,18 +5171,17 @@ void kernel_mul_mv_iq2_xxs_f32_impl( device const uint16_t * q2 = xr->qs + 4 * ib; device const half * dh = &xr->d; - for (int row = 0; row < N_DST; row++) { - + for (short row = 0; row < nr0; row++) { const float db = dh[0]; device const uint8_t * aux8 = (device const uint8_t *)q2; const uint32_t aux32 = q2[2] | (q2[3] << 16); const float d = db * (0.5f + (aux32 >> 28)); float sum = 0; - for (int l = 0; l < 4; ++l) { + for (short l = 0; l < 4; ++l) { const threadgroup uint8_t * grid = (const threadgroup uint8_t *)(svalues + aux8[l]); const uint8_t signs = ssigns[(aux32 >> 7*l) & 127]; - for (int j = 0; j < 8; ++j) { + for (short j = 0; j < 8; ++j) { sum += yl[8*l + j] * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f); } } @@ -5064,10 +5196,10 @@ void kernel_mul_mv_iq2_xxs_f32_impl( device float * dst_f32 = (device float *) dst + (uint64_t)im*args.ne0*args.ne1 + (uint64_t)r1*args.ne0; - for (int row = 0; row < N_DST && first_row + row < args.ne0; ++row) { - all_sum = simd_sum(sumf[row]); + for (int row = 0; row < nr0 && first_row + row < args.ne0; ++row) { + float sum_all = simd_sum(sumf[row]); if (tiisg == 0) { - dst_f32[first_row + row] = all_sum * 0.25f; + dst_f32[first_row + row] = sum_all * 0.25f; } } } @@ -5082,10 +5214,10 @@ kernel void kernel_mul_mv_iq2_xxs_f32( uint3 tgpig[[threadgroup_position_in_grid]], ushort tiisg[[thread_index_in_simdgroup]], ushort sgitg[[simdgroup_index_in_threadgroup]]) { - kernel_mul_mv_iq2_xxs_f32_impl(args, src0, src1, dst, shmem, tgpig, tiisg, sgitg); + kernel_mul_mv_iq2_xxs_f32_impl(args, src0, src1, dst, shmem, tgpig, tiisg, sgitg); } -template +template void kernel_mul_mv_iq2_xs_f32_impl( args_t args, device const char * src0, @@ -5101,7 +5233,7 @@ void kernel_mul_mv_iq2_xs_f32_impl( const int r1 = tgpig.y; const int im = tgpig.z; - const int first_row = (r0 * N_SIMDGROUP + sgitg) * N_DST; + const int first_row = (r0 * nsg + sgitg) * nr0; const uint i12 = im%args.ne12; const uint i13 = im/args.ne12; @@ -5113,7 +5245,7 @@ void kernel_mul_mv_iq2_xs_f32_impl( device const float * y = (device const float *) (src1 + offset1); float yl[32]; - float sumf[N_DST]={0.f}, all_sum; + float sumf[nr0]={0.f}; const int nb32 = nb * (QK_K / 32); @@ -5134,8 +5266,7 @@ void kernel_mul_mv_iq2_xs_f32_impl( device const float * y4 = y + 32 * ix; for (int ib32 = ix; ib32 < nb32; ib32 += 32) { - - for (int i = 0; i < 32; ++i) { + for (short i = 0; i < 32; ++i) { yl[i] = y4[i]; } @@ -5147,8 +5278,7 @@ void kernel_mul_mv_iq2_xs_f32_impl( device const uint8_t * sc = xr->scales + ib; device const half * dh = &xr->d; - for (int row = 0; row < N_DST; row++) { - + for (short row = 0; row < nr0; row++) { const float db = dh[0]; const uint8_t ls1 = sc[0] & 0xf; const uint8_t ls2 = sc[0] >> 4; @@ -5156,17 +5286,17 @@ void kernel_mul_mv_iq2_xs_f32_impl( const float d2 = db * (0.5f + ls2); float sum1 = 0, sum2 = 0; - for (int l = 0; l < 2; ++l) { + for (short l = 0; l < 2; ++l) { const threadgroup uint8_t * grid = (const threadgroup uint8_t *)(svalues + (q2[l] & 511)); const uint8_t signs = ssigns[(q2[l] >> 9)]; - for (int j = 0; j < 8; ++j) { + for (short j = 0; j < 8; ++j) { sum1 += yl[8*l + j] * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f); } } - for (int l = 2; l < 4; ++l) { + for (short l = 2; l < 4; ++l) { const threadgroup uint8_t * grid = (const threadgroup uint8_t *)(svalues + (q2[l] & 511)); const uint8_t signs = ssigns[(q2[l] >> 9)]; - for (int j = 0; j < 8; ++j) { + for (short j = 0; j < 8; ++j) { sum2 += yl[8*l + j] * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f); } } @@ -5182,10 +5312,10 @@ void kernel_mul_mv_iq2_xs_f32_impl( device float * dst_f32 = (device float *) dst + (uint64_t)im*args.ne0*args.ne1 + (uint64_t)r1*args.ne0; - for (int row = 0; row < N_DST && first_row + row < args.ne0; ++row) { - all_sum = simd_sum(sumf[row]); + for (int row = 0; row < nr0 && first_row + row < args.ne0; ++row) { + float sum_all = simd_sum(sumf[row]); if (tiisg == 0) { - dst_f32[first_row + row] = all_sum * 0.25f; + dst_f32[first_row + row] = sum_all * 0.25f; } } } @@ -5201,10 +5331,10 @@ kernel void kernel_mul_mv_iq2_xs_f32( ushort tiisg[[thread_index_in_simdgroup]], ushort sgitg[[simdgroup_index_in_threadgroup]]) { - kernel_mul_mv_iq2_xs_f32_impl(args, src0, src1, dst, shmem, tgpig, tiisg, sgitg); + kernel_mul_mv_iq2_xs_f32_impl(args, src0, src1, dst, shmem, tgpig, tiisg, sgitg); } -template +template void kernel_mul_mv_iq3_xxs_f32_impl( args_t args, device const char * src0, @@ -5220,7 +5350,7 @@ void kernel_mul_mv_iq3_xxs_f32_impl( const int r1 = tgpig.y; const int im = tgpig.z; - const int first_row = (r0 * N_SIMDGROUP + sgitg) * N_DST; + const int first_row = (r0 * nsg + sgitg) * nr0; const uint i12 = im%args.ne12; const uint i13 = im/args.ne12; @@ -5232,7 +5362,7 @@ void kernel_mul_mv_iq3_xxs_f32_impl( device const float * y = (device const float *) (src1 + offset1); float yl[32]; - float sumf[N_DST]={0.f}, all_sum; + float sumf[nr0]={0.f}; const int nb32 = nb * (QK_K / 32); @@ -5253,7 +5383,7 @@ void kernel_mul_mv_iq3_xxs_f32_impl( device const float * y4 = y + 32 * ix; for (int ib32 = ix; ib32 < nb32; ib32 += 32) { - for (int i = 0; i < 32; ++i) { + for (short i = 0; i < 32; ++i) { yl[i] = y4[i]; } @@ -5265,17 +5395,17 @@ void kernel_mul_mv_iq3_xxs_f32_impl( device const uint16_t * gas = (device const uint16_t *)(xr->qs + QK_K/4) + 2 * ib; device const half * dh = &xr->d; - for (int row = 0; row < N_DST; row++) { + for (short row = 0; row < nr0; row++) { const float db = dh[0]; const uint32_t aux32 = gas[0] | (gas[1] << 16); const float d = db * (0.5f + (aux32 >> 28)); float2 sum = {0}; - for (int l = 0; l < 4; ++l) { + for (short l = 0; l < 4; ++l) { const threadgroup uint8_t * grid1 = (const threadgroup uint8_t *)(svalues + q3[2*l+0]); const threadgroup uint8_t * grid2 = (const threadgroup uint8_t *)(svalues + q3[2*l+1]); const uint8_t signs = ssigns[(aux32 >> 7*l) & 127]; - for (int j = 0; j < 4; ++j) { + for (short j = 0; j < 4; ++j) { sum[0] += yl[8*l + j + 0] * grid1[j] * (signs & kmask_iq2xs[j+0] ? -1.f : 1.f); sum[1] += yl[8*l + j + 4] * grid2[j] * (signs & kmask_iq2xs[j+4] ? -1.f : 1.f); } @@ -5292,10 +5422,10 @@ void kernel_mul_mv_iq3_xxs_f32_impl( device float * dst_f32 = (device float *) dst + (uint64_t)im*args.ne0*args.ne1 + (uint64_t)r1*args.ne0; - for (int row = 0; row < N_DST && first_row + row < args.ne0; ++row) { - all_sum = simd_sum(sumf[row]); + for (int row = 0; row < nr0 && first_row + row < args.ne0; ++row) { + float sum_all = simd_sum(sumf[row]); if (tiisg == 0) { - dst_f32[first_row + row] = all_sum * 0.5f; + dst_f32[first_row + row] = sum_all * 0.5f; } } } @@ -5311,10 +5441,10 @@ kernel void kernel_mul_mv_iq3_xxs_f32( ushort tiisg[[thread_index_in_simdgroup]], ushort sgitg[[simdgroup_index_in_threadgroup]]) { - kernel_mul_mv_iq3_xxs_f32_impl(args, src0, src1, dst, shmem, tgpig, tiisg, sgitg); + kernel_mul_mv_iq3_xxs_f32_impl(args, src0, src1, dst, shmem, tgpig, tiisg, sgitg); } -template +template void kernel_mul_mv_iq3_s_f32_impl( args_t args, device const char * src0, @@ -5330,7 +5460,7 @@ void kernel_mul_mv_iq3_s_f32_impl( const int r1 = tgpig.y; const int im = tgpig.z; - const int first_row = (r0 * N_SIMDGROUP + sgitg) * N_DST; + const int first_row = (r0 * nsg + sgitg) * nr0; const uint i12 = im%args.ne12; const uint i13 = im/args.ne12; @@ -5342,7 +5472,7 @@ void kernel_mul_mv_iq3_s_f32_impl( device const float * y = (device const float *) (src1 + offset1); float yl[32]; - float sumf[N_DST]={0.f}, all_sum; + float sumf[nr0]={0.f}; const int nb32 = nb * (QK_K / 32); @@ -5359,8 +5489,7 @@ void kernel_mul_mv_iq3_s_f32_impl( device const float * y4 = y + 32 * ix; for (int ib32 = ix; ib32 < nb32; ib32 += 32) { - - for (int i = 0; i < 32; ++i) { + for (short i = 0; i < 32; ++i) { yl[i] = y4[i]; } @@ -5374,18 +5503,17 @@ void kernel_mul_mv_iq3_s_f32_impl( device const uint8_t * signs = xr->signs + 4 * ib; device const half * dh = &xr->d; - for (int row = 0; row < N_DST; row++) { - + for (short row = 0; row < nr0; row++) { const float db = dh[0]; const float d = db * (1 + 2*((sc[0] >> 4*(ib%2)) & 0xf)); float2 sum = {0}; - for (int l = 0; l < 4; ++l) { + for (short l = 0; l < 4; ++l) { const threadgroup uint32_t * table1 = qh[0] & kmask_iq2xs[2*l+0] ? svalues + 256 : svalues; const threadgroup uint32_t * table2 = qh[0] & kmask_iq2xs[2*l+1] ? svalues + 256 : svalues; const threadgroup uint8_t * grid1 = (const threadgroup uint8_t *)(table1 + qs[2*l+0]); const threadgroup uint8_t * grid2 = (const threadgroup uint8_t *)(table2 + qs[2*l+1]); - for (int j = 0; j < 4; ++j) { + for (short j = 0; j < 4; ++j) { sum[0] += yl[8*l + j + 0] * grid1[j] * select(1, -1, signs[l] & kmask_iq2xs[j+0]); sum[1] += yl[8*l + j + 4] * grid2[j] * select(1, -1, signs[l] & kmask_iq2xs[j+4]); } @@ -5404,10 +5532,10 @@ void kernel_mul_mv_iq3_s_f32_impl( device float * dst_f32 = (device float *) dst + (uint64_t)im*args.ne0*args.ne1 + (uint64_t)r1*args.ne0; - for (int row = 0; row < N_DST && first_row + row < args.ne0; ++row) { - all_sum = simd_sum(sumf[row]); + for (int row = 0; row < nr0 && first_row + row < args.ne0; ++row) { + float sum_all = simd_sum(sumf[row]); if (tiisg == 0) { - dst_f32[first_row + row] = all_sum; + dst_f32[first_row + row] = sum_all; } } } @@ -5423,10 +5551,10 @@ kernel void kernel_mul_mv_iq3_s_f32( ushort tiisg[[thread_index_in_simdgroup]], ushort sgitg[[simdgroup_index_in_threadgroup]]) { - kernel_mul_mv_iq3_s_f32_impl(args, src0, src1, dst, shmem, tgpig, tiisg, sgitg); + kernel_mul_mv_iq3_s_f32_impl(args, src0, src1, dst, shmem, tgpig, tiisg, sgitg); } -template +template void kernel_mul_mv_iq2_s_f32_impl( args_t args, device const char * src0, @@ -5442,7 +5570,7 @@ void kernel_mul_mv_iq2_s_f32_impl( const int r1 = tgpig.y; const int im = tgpig.z; - const int first_row = (r0 * N_SIMDGROUP + sgitg) * N_DST; + const int first_row = (r0 * nsg + sgitg) * nr0; const uint i12 = im%args.ne12; const uint i13 = im/args.ne12; @@ -5454,7 +5582,7 @@ void kernel_mul_mv_iq2_s_f32_impl( device const float * y = (device const float *) (src1 + offset1); float yl[32]; - float sumf[N_DST]={0.f}, all_sum; + float sumf[nr0]={0.f}; const int nb32 = nb * (QK_K / 32); @@ -5466,13 +5594,12 @@ void kernel_mul_mv_iq2_s_f32_impl( // threadgroup_barrier(mem_flags::mem_threadgroup); //} - const int ix = tiisg; + const short ix = tiisg; device const float * y4 = y + 32 * ix; for (int ib32 = ix; ib32 < nb32; ib32 += 32) { - - for (int i = 0; i < 32; ++i) { + for (short i = 0; i < 32; ++i) { yl[i] = y4[i]; } @@ -5486,19 +5613,18 @@ void kernel_mul_mv_iq2_s_f32_impl( device const uint8_t * signs = qs + QK_K/8; device const half * dh = &xr->d; - for (int row = 0; row < N_DST; row++) { - + for (short row = 0; row < nr0; row++) { const float db = dh[0]; const float d1 = db * (0.5f + (sc[0] & 0xf)); const float d2 = db * (0.5f + (sc[0] >> 4)); float2 sum = {0}; - for (int l = 0; l < 2; ++l) { + for (short l = 0; l < 2; ++l) { //const threadgroup uint8_t * grid1 = (const threadgroup uint8_t *)(svalues + (qs[l+0] | ((qh[0] << (8-2*l)) & 0x300))); //const threadgroup uint8_t * grid2 = (const threadgroup uint8_t *)(svalues + (qs[l+2] | ((qh[0] << (4-2*l)) & 0x300))); constant uint8_t * grid1 = (constant uint8_t *)(iq2s_grid + (qs[l+0] | ((qh[0] << (8-2*l)) & 0x300))); constant uint8_t * grid2 = (constant uint8_t *)(iq2s_grid + (qs[l+2] | ((qh[0] << (4-2*l)) & 0x300))); - for (int j = 0; j < 8; ++j) { + for (short j = 0; j < 8; ++j) { sum[0] += yl[8*l + j + 0] * grid1[j] * select(1, -1, signs[l+0] & kmask_iq2xs[j]); sum[1] += yl[8*l + j + 16] * grid2[j] * select(1, -1, signs[l+2] & kmask_iq2xs[j]); } @@ -5517,10 +5643,10 @@ void kernel_mul_mv_iq2_s_f32_impl( device float * dst_f32 = (device float *) dst + (uint64_t)im*args.ne0*args.ne1 + (uint64_t)r1*args.ne0; - for (int row = 0; row < N_DST && first_row + row < args.ne0; ++row) { - all_sum = simd_sum(sumf[row]); + for (int row = 0; row < nr0 && first_row + row < args.ne0; ++row) { + float sum_all = simd_sum(sumf[row]); if (tiisg == 0) { - dst_f32[first_row + row] = all_sum * 0.25f; + dst_f32[first_row + row] = sum_all * 0.25f; } } } @@ -5536,10 +5662,10 @@ kernel void kernel_mul_mv_iq2_s_f32( ushort tiisg[[thread_index_in_simdgroup]], ushort sgitg[[simdgroup_index_in_threadgroup]]) { - kernel_mul_mv_iq2_s_f32_impl(args, src0, src1, dst, shmem, tgpig, tiisg, sgitg); + kernel_mul_mv_iq2_s_f32_impl(args, src0, src1, dst, shmem, tgpig, tiisg, sgitg); } -template +template void kernel_mul_mv_iq1_s_f32_impl( args_t args, device const char * src0, @@ -5555,7 +5681,7 @@ void kernel_mul_mv_iq1_s_f32_impl( const int r1 = tgpig.y; const int im = tgpig.z; - const int first_row = (r0 * N_SIMDGROUP + sgitg) * N_DST; + const int first_row = (r0 * nsg + sgitg) * nr0; const uint i12 = im%args.ne12; const uint i13 = im/args.ne12; @@ -5567,18 +5693,17 @@ void kernel_mul_mv_iq1_s_f32_impl( device const float * y = (device const float *) (src1 + offset1); float yl[32]; - float sumf[N_DST]={0.f}, all_sum; + float sumf[nr0]={0.f}; const int nb32 = nb * (QK_K / 32); - const int ix = tiisg; + const short ix = tiisg; device const float * y4 = y + 32 * ix; for (int ib32 = ix; ib32 < nb32; ib32 += 32) { - float sumy = 0; - for (int i = 0; i < 32; ++i) { + for (short i = 0; i < 32; ++i) { yl[i] = y4[i]; sumy += yl[i]; } @@ -5591,15 +5716,14 @@ void kernel_mul_mv_iq1_s_f32_impl( device const uint16_t * qh = xr->qh + ib; device const half * dh = &xr->d; - for (int row = 0; row < N_DST; row++) { - + for (short row = 0; row < nr0; row++) { constant uint8_t * grid1 = (constant uint8_t *)(iq1s_grid_gpu + (qs[0] | ((qh[0] << 8) & 0x700))); constant uint8_t * grid2 = (constant uint8_t *)(iq1s_grid_gpu + (qs[1] | ((qh[0] << 5) & 0x700))); constant uint8_t * grid3 = (constant uint8_t *)(iq1s_grid_gpu + (qs[2] | ((qh[0] << 2) & 0x700))); constant uint8_t * grid4 = (constant uint8_t *)(iq1s_grid_gpu + (qs[3] | ((qh[0] >> 1) & 0x700))); float sum = 0; - for (int j = 0; j < 4; ++j) { + for (short j = 0; j < 4; ++j) { sum += yl[j+ 0] * (grid1[j] & 0xf) + yl[j+ 4] * (grid1[j] >> 4) + yl[j+ 8] * (grid2[j] & 0xf) + yl[j+12] * (grid2[j] >> 4) + yl[j+16] * (grid3[j] & 0xf) + yl[j+20] * (grid3[j] >> 4) @@ -5617,15 +5741,28 @@ void kernel_mul_mv_iq1_s_f32_impl( device float * dst_f32 = (device float *) dst + (uint64_t)im*args.ne0*args.ne1 + (uint64_t)r1*args.ne0; - for (int row = 0; row < N_DST && first_row + row < args.ne0; ++row) { - all_sum = simd_sum(sumf[row]); + for (int row = 0; row < nr0 && first_row + row < args.ne0; ++row) { + float sum_all = simd_sum(sumf[row]); if (tiisg == 0) { - dst_f32[first_row + row] = all_sum; + dst_f32[first_row + row] = sum_all; } } } -template +[[host_name("kernel_mul_mv_iq1_s_f32")]] +kernel void kernel_mul_mv_iq1_s_f32( + constant ggml_metal_kargs_mul_mv & args, + device const char * src0, + device const char * src1, + device char * dst, + uint3 tgpig[[threadgroup_position_in_grid]], + ushort tiisg[[thread_index_in_simdgroup]], + ushort sgitg[[simdgroup_index_in_threadgroup]]) { + + kernel_mul_mv_iq1_s_f32_impl(args, src0, src1, dst, nullptr, tgpig, tiisg, sgitg); +} + +template void kernel_mul_mv_iq1_m_f32_impl( args_t args, device const char * src0, @@ -5637,11 +5774,12 @@ void kernel_mul_mv_iq1_m_f32_impl( ushort sgitg) { const int nb = args.ne00/QK_K; + const int r0 = tgpig.x; const int r1 = tgpig.y; const int im = tgpig.z; - const int first_row = (r0 * N_SIMDGROUP + sgitg) * N_DST; + const int first_row = (r0 * nsg + sgitg) * nr0; const uint i12 = im%args.ne12; const uint i13 = im/args.ne12; @@ -5653,20 +5791,19 @@ void kernel_mul_mv_iq1_m_f32_impl( device const float * y = (device const float *) (src1 + offset1); float yl[32]; - float sumf[N_DST]={0.f}, all_sum; + float sumf[nr0]={0.f}; const int nb32 = nb * (QK_K / 32); - const int ix = tiisg; + const short ix = tiisg; device const float * y4 = y + 32 * ix; iq1m_scale_t scale; for (int ib32 = ix; ib32 < nb32; ib32 += 32) { - float4 sumy = {0.f}; - for (int i = 0; i < 8; ++i) { + for (short i = 0; i < 8; ++i) { yl[i+ 0] = y4[i+ 0]; sumy[0] += yl[i+ 0]; yl[i+ 8] = y4[i+ 8]; sumy[1] += yl[i+ 8]; yl[i+16] = y4[i+16]; sumy[2] += yl[i+16]; @@ -5681,7 +5818,7 @@ void kernel_mul_mv_iq1_m_f32_impl( device const uint8_t * qh = xr->qh + 2 * ib; device const uint16_t * sc = (device const uint16_t *)xr->scales; - for (int row = 0; row < N_DST; row++) { + for (short row = 0; row < nr0; row++) { scale.u16 = (sc[0] >> 12) | ((sc[1] >> 8) & 0x00f0) | ((sc[2] >> 4) & 0x0f00) | (sc[3] & 0xf000); constant uint8_t * grid1 = (constant uint8_t *)(iq1s_grid_gpu + (qs[0] | ((qh[0] << 8) & 0x700))); @@ -5690,7 +5827,7 @@ void kernel_mul_mv_iq1_m_f32_impl( constant uint8_t * grid4 = (constant uint8_t *)(iq1s_grid_gpu + (qs[3] | ((qh[1] << 4) & 0x700))); float2 sum = {0.f}; - for (int j = 0; j < 4; ++j) { + for (short j = 0; j < 4; ++j) { sum[0] += yl[j+ 0] * (grid1[j] & 0xf) + yl[j+ 4] * (grid1[j] >> 4) + yl[j+ 8] * (grid2[j] & 0xf) + yl[j+12] * (grid2[j] >> 4); sum[1] += yl[j+16] * (grid3[j] & 0xf) + yl[j+20] * (grid3[j] >> 4) @@ -5712,15 +5849,28 @@ void kernel_mul_mv_iq1_m_f32_impl( device float * dst_f32 = (device float *) dst + (uint64_t)im*args.ne0*args.ne1 + (uint64_t)r1*args.ne0; - for (int row = 0; row < N_DST && first_row + row < args.ne0; ++row) { - all_sum = simd_sum(sumf[row]); + for (int row = 0; row < nr0 && first_row + row < args.ne0; ++row) { + float sum_all = simd_sum(sumf[row]); if (tiisg == 0) { - dst_f32[first_row + row] = all_sum; + dst_f32[first_row + row] = sum_all; } } } -template +[[host_name("kernel_mul_mv_iq1_m_f32")]] +kernel void kernel_mul_mv_iq1_m_f32( + constant ggml_metal_kargs_mul_mv & args, + device const char * src0, + device const char * src1, + device char * dst, + uint3 tgpig[[threadgroup_position_in_grid]], + ushort tiisg[[thread_index_in_simdgroup]], + ushort sgitg[[simdgroup_index_in_threadgroup]]) { + + kernel_mul_mv_iq1_m_f32_impl(args, src0, src1, dst, nullptr, tgpig, tiisg, sgitg); +} + +template void kernel_mul_mv_iq4_nl_f32_impl( args_t args, device const char * src0, @@ -5733,10 +5883,12 @@ void kernel_mul_mv_iq4_nl_f32_impl( threadgroup float * shmem_f32 = (threadgroup float *) shmem; const int nb = args.ne00/QK4_NL; + const int r0 = tgpig.x; const int r1 = tgpig.y; const int im = tgpig.z; - const int first_row = (r0 * 2 + sgitg) * 2; + + const int first_row = (r0 * nsg + sgitg) * nr0; const uint i12 = im%args.ne12; const uint i13 = im/args.ne12; @@ -5747,14 +5899,14 @@ void kernel_mul_mv_iq4_nl_f32_impl( device const block_iq4_nl * x = (device const block_iq4_nl *) (src0 + offset0); device const float * y = (device const float *) (src1 + offset1); - const int ix = tiisg/2; // 0...15 - const int it = tiisg%2; // 0 or 1 + const short ix = tiisg/2; // 0...15 + const short it = tiisg%2; // 0 or 1 shmem_f32[tiisg] = kvalues_iq4nl_f[tiisg%16]; threadgroup_barrier(mem_flags::mem_threadgroup); float4 yl[4]; - float sumf[2]={0.f}, all_sum; + float sumf[nr0]={0.f}; device const float * yb = y + ix * QK4_NL + it * 8; @@ -5764,12 +5916,13 @@ void kernel_mul_mv_iq4_nl_f32_impl( float4 qf1, qf2; for (int ib = ix; ib < nb; ib += 16) { - device const float4 * y4 = (device const float4 *)yb; - yl[0] = y4[0]; yl[1] = y4[4]; yl[2] = y4[1]; yl[3] = y4[5]; - - for (int row = 0; row < 2 && first_row + row < args.ne01; ++row) { + yl[0] = y4[0]; + yl[1] = y4[4]; + yl[2] = y4[1]; + yl[3] = y4[5]; + for (short row = 0; row < nr0; row++) { device const block_iq4_nl & xb = x[row*nb + ib]; device const uint16_t * q4 = (device const uint16_t *)(xb.qs + 8*it); @@ -5794,7 +5947,6 @@ void kernel_mul_mv_iq4_nl_f32_impl( acc1 += acc2; sumf[row] += (float)xb.d * (acc1[0] + acc1[1] + acc1[2] + acc1[3]); - } yb += 16 * QK4_NL; @@ -5802,15 +5954,29 @@ void kernel_mul_mv_iq4_nl_f32_impl( device float * dst_f32 = (device float *) dst + (uint64_t)im*args.ne0*args.ne1 + (uint64_t)r1*args.ne0; - for (int row = 0; row < 2 && first_row + row < args.ne0; ++row) { - all_sum = simd_sum(sumf[row]); + for (int row = 0; row < nr0 && first_row + row < args.ne0; ++row) { + float sum_all = simd_sum(sumf[row]); if (tiisg == 0) { - dst_f32[first_row + row] = all_sum; + dst_f32[first_row + row] = sum_all; } } } -template +[[host_name("kernel_mul_mv_iq4_nl_f32")]] +kernel void kernel_mul_mv_iq4_nl_f32( + constant ggml_metal_kargs_mul_mv & args, + device const char * src0, + device const char * src1, + device char * dst, + threadgroup char * shmem [[threadgroup(0)]], + uint3 tgpig[[threadgroup_position_in_grid]], + ushort tiisg[[thread_index_in_simdgroup]], + ushort sgitg[[simdgroup_index_in_threadgroup]]) { + + kernel_mul_mv_iq4_nl_f32_impl(args, src0, src1, dst, shmem, tgpig, tiisg, sgitg); +} + +template void kernel_mul_mv_iq4_xs_f32_impl( args_t args, device const char * src0, @@ -5826,7 +5992,7 @@ void kernel_mul_mv_iq4_xs_f32_impl( const int r0 = tgpig.x; const int r1 = tgpig.y; const int im = tgpig.z; - const int first_row = (r0 * 2 + sgitg) * 2; + const int first_row = (r0 * nsg + sgitg) * nr0; const uint i12 = im%args.ne12; const uint i13 = im/args.ne12; @@ -5837,16 +6003,16 @@ void kernel_mul_mv_iq4_xs_f32_impl( device const block_iq4_xs * x = (device const block_iq4_xs *) (src0 + offset0); device const float * y = (device const float *) (src1 + offset1); - const int ix = tiisg/16; // 0 or 1 - const int it = tiisg%16; // 0...15 - const int ib = it/2; - const int il = it%2; + const short ix = tiisg/16; // 0 or 1 + const short it = tiisg%16; // 0...15 + const short ib = it/2; + const short il = it%2; shmem_f32[tiisg] = kvalues_iq4nl_f[tiisg%16]; threadgroup_barrier(mem_flags::mem_threadgroup); float4 yl[4]; - float sumf[2]={0.f}, all_sum; + float sumf[nr0]={0.f}; device const float * yb = y + ix * QK_K + ib * 32 + il * 8; @@ -5857,9 +6023,12 @@ void kernel_mul_mv_iq4_xs_f32_impl( for (int ibl = ix; ibl < nb; ibl += 2) { device const float4 * y4 = (device const float4 *)yb; - yl[0] = y4[0]; yl[1] = y4[4]; yl[2] = y4[1]; yl[3] = y4[5]; + yl[0] = y4[0]; + yl[1] = y4[4]; + yl[2] = y4[1]; + yl[3] = y4[5]; - for (int row = 0; row < 2; ++row) { + for (short row = 0; row < nr0; ++row) { device const block_iq4_xs & xb = x[row*nb + ibl]; device const uint32_t * q4 = (device const uint32_t *)(xb.qs + 16*ib + 8*il); @@ -5883,7 +6052,6 @@ void kernel_mul_mv_iq4_xs_f32_impl( const int ls = (((xb.scales_l[ib/2] >> 4*(ib%2)) & 0xf) | (((xb.scales_h >> 2*ib) & 3) << 4)) - 32; sumf[row] += (float)xb.d * ls * (acc1[0] + acc1[1] + acc1[2] + acc1[3]); - } yb += 2 * QK_K; @@ -5891,54 +6059,14 @@ void kernel_mul_mv_iq4_xs_f32_impl( device float * dst_f32 = (device float *) dst + (uint64_t)im*args.ne0*args.ne1 + (uint64_t)r1*args.ne0; - for (int row = 0; row < 2 && first_row + row < args.ne0; ++row) { - all_sum = simd_sum(sumf[row]); + for (int row = 0; row < nr0 && first_row + row < args.ne0; ++row) { + float sum_all = simd_sum(sumf[row]); if (tiisg == 0) { - dst_f32[first_row + row] = all_sum; + dst_f32[first_row + row] = sum_all; } } } -[[host_name("kernel_mul_mv_iq1_s_f32")]] -kernel void kernel_mul_mv_iq1_s_f32( - constant ggml_metal_kargs_mul_mv & args, - device const char * src0, - device const char * src1, - device char * dst, - uint3 tgpig[[threadgroup_position_in_grid]], - ushort tiisg[[thread_index_in_simdgroup]], - ushort sgitg[[simdgroup_index_in_threadgroup]]) { - - kernel_mul_mv_iq1_s_f32_impl(args, src0, src1, dst, nullptr, tgpig, tiisg, sgitg); -} - -[[host_name("kernel_mul_mv_iq1_m_f32")]] -kernel void kernel_mul_mv_iq1_m_f32( - constant ggml_metal_kargs_mul_mv & args, - device const char * src0, - device const char * src1, - device char * dst, - uint3 tgpig[[threadgroup_position_in_grid]], - ushort tiisg[[thread_index_in_simdgroup]], - ushort sgitg[[simdgroup_index_in_threadgroup]]) { - - kernel_mul_mv_iq1_m_f32_impl(args, src0, src1, dst, nullptr, tgpig, tiisg, sgitg); -} - -[[host_name("kernel_mul_mv_iq4_nl_f32")]] -kernel void kernel_mul_mv_iq4_nl_f32( - constant ggml_metal_kargs_mul_mv & args, - device const char * src0, - device const char * src1, - device char * dst, - threadgroup char * shmem [[threadgroup(0)]], - uint3 tgpig[[threadgroup_position_in_grid]], - ushort tiisg[[thread_index_in_simdgroup]], - ushort sgitg[[simdgroup_index_in_threadgroup]]) { - - kernel_mul_mv_iq4_nl_f32_impl(args, src0, src1, dst, shmem, tgpig, tiisg, sgitg); -} - [[host_name("kernel_mul_mv_iq4_xs_f32")]] kernel void kernel_mul_mv_iq4_xs_f32( constant ggml_metal_kargs_mul_mv & args, @@ -5950,7 +6078,7 @@ kernel void kernel_mul_mv_iq4_xs_f32( ushort tiisg[[thread_index_in_simdgroup]], ushort sgitg[[simdgroup_index_in_threadgroup]]) { - kernel_mul_mv_iq4_xs_f32_impl(args, src0, src1, dst, shmem, tgpig, tiisg, sgitg); + kernel_mul_mv_iq4_xs_f32_impl(args, src0, src1, dst, shmem, tgpig, tiisg, sgitg); } template @@ -5958,28 +6086,21 @@ kernel void kernel_get_rows_q( device const void * src0, device const void * src1, device float * dst, - constant int64_t & ne00, - constant uint64_t & nb01, - constant uint64_t & nb02, - constant int64_t & ne10, - constant uint64_t & nb10, - constant uint64_t & nb11, - constant uint64_t & nb1, - constant uint64_t & nb2, + constant ggml_metal_kargs_get_rows & args, uint3 tgpig[[threadgroup_position_in_grid]], uint tiitg[[thread_index_in_threadgroup]], uint3 tptg [[threads_per_threadgroup]]) { const int64_t i10 = tgpig.x; const int64_t i11 = tgpig.y; - const int64_t r = ((const device int32_t *) ((const device char *) src1 + i11*nb11 + i10*nb10))[0]; + const int64_t r = ((const device int32_t *) ((const device char *) src1 + i11*args.nb11 + i10*args.nb10))[0]; const int64_t i02 = i11; - for (int64_t ind = tiitg; ind < ne00/16; ind += tptg.x) { + for (int64_t ind = tiitg; ind < args.ne00/16; ind += tptg.x) { float4x4 temp; - dequantize_func(((device const block_q *) ((const device char *) src0 + r*nb01 + i02*nb02)) + ind/nl, ind%nl, temp); - *(((device float4x4 *) ((device char *) dst + i11*nb2 + i10*nb1)) + ind) = temp; + dequantize_func(((device const block_q *) ((const device char *) src0 + r*args.nb01 + i02*args.nb02)) + ind/nl, ind%nl, temp); + *(((device float4x4 *) ((device char *) dst + i11*args.nb2 + i10*args.nb1)) + ind) = temp; } } @@ -5988,27 +6109,20 @@ kernel void kernel_get_rows_f( device const void * src0, device const void * src1, device float * dst, - constant int64_t & ne00, - constant uint64_t & nb01, - constant uint64_t & nb02, - constant int64_t & ne10, - constant uint64_t & nb10, - constant uint64_t & nb11, - constant uint64_t & nb1, - constant uint64_t & nb2, + constant ggml_metal_kargs_get_rows & args, uint3 tgpig[[threadgroup_position_in_grid]], uint tiitg[[thread_index_in_threadgroup]], uint3 tptg [[threads_per_threadgroup]]) { const int64_t i10 = tgpig.x; const int64_t i11 = tgpig.y; - const int64_t r = ((const device int32_t *) ((const device char *) src1 + i11*nb11 + i10*nb10))[0]; + const int64_t r = ((const device int32_t *) ((const device char *) src1 + i11*args.nb11 + i10*args.nb10))[0]; const int64_t i02 = i11; - for (int ind = tiitg; ind < ne00; ind += tptg.x) { - (( device float *) (( device char *) dst + i11*nb2 + i10*nb1))[ind] = - ((const device T *) ((const device char *) src0 + i02*nb02 + r*nb01))[ind]; + for (int ind = tiitg; ind < args.ne00; ind += tptg.x) { + (( device float *) (( device char *) dst + i11*args.nb2 + i10*args.nb1))[ind] = + ((const device T *) ((const device char *) src0 + i02*args.nb02 + r*args.nb01))[ind]; } } @@ -6016,27 +6130,20 @@ kernel void kernel_get_rows_i32( device const void * src0, device const void * src1, device int32_t * dst, - constant int64_t & ne00, - constant uint64_t & nb01, - constant uint64_t & nb02, - constant int64_t & ne10, - constant uint64_t & nb10, - constant uint64_t & nb11, - constant uint64_t & nb1, - constant uint64_t & nb2, + constant ggml_metal_kargs_get_rows & args, uint3 tgpig[[threadgroup_position_in_grid]], uint tiitg[[thread_index_in_threadgroup]], uint3 tptg [[threads_per_threadgroup]]) { const int64_t i10 = tgpig.x; const int64_t i11 = tgpig.y; - const int64_t r = ((const device int32_t *) ((const device char *) src1 + i11*nb11 + i10*nb10))[0]; + const int64_t r = ((const device int32_t *) ((const device char *) src1 + i11*args.nb11 + i10*args.nb10))[0]; const int64_t i02 = i11; - for (int ind = tiitg; ind < ne00; ind += tptg.x) { - (( device int32_t *) (( device char *) dst + i11*nb2 + i10*nb1))[ind] = - ((const device int32_t *) ((const device char *) src0 + i02*nb02 + r*nb01))[ind]; + for (int ind = tiitg; ind < args.ne00; ind += tptg.x) { + (( device int32_t *) (( device char *) dst + i11*args.nb2 + i10*args.nb1))[ind] = + ((const device int32_t *) ((const device char *) src0 + i02*args.nb02 + r*args.nb01))[ind]; } } @@ -6615,121 +6722,103 @@ template [[host_name("kernel_mul_mv_id_f16_f32")]] kernel kernel_mul_mv_id_t #if defined(GGML_METAL_USE_BF16) template [[host_name("kernel_mul_mv_id_bf16_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id>>; #endif -template [[host_name("kernel_mul_mv_id_q8_0_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id>; -template [[host_name("kernel_mul_mv_id_q4_0_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id>>; -template [[host_name("kernel_mul_mv_id_q4_1_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id>>; -template [[host_name("kernel_mul_mv_id_q5_0_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id>>; -template [[host_name("kernel_mul_mv_id_q5_1_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id>>; -template [[host_name("kernel_mul_mv_id_q2_K_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id>; -template [[host_name("kernel_mul_mv_id_q3_K_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id>; -template [[host_name("kernel_mul_mv_id_q4_K_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id>; -template [[host_name("kernel_mul_mv_id_q5_K_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id>; -template [[host_name("kernel_mul_mv_id_q6_K_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id>; -template [[host_name("kernel_mul_mv_id_iq1_s_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id>; -template [[host_name("kernel_mul_mv_id_iq1_m_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id>; -template [[host_name("kernel_mul_mv_id_iq2_xxs_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id>; -template [[host_name("kernel_mul_mv_id_iq2_xs_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id>; -template [[host_name("kernel_mul_mv_id_iq3_xxs_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id>; -template [[host_name("kernel_mul_mv_id_iq3_s_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id>; -template [[host_name("kernel_mul_mv_id_iq2_s_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id>; -template [[host_name("kernel_mul_mv_id_iq4_nl_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id>; -template [[host_name("kernel_mul_mv_id_iq4_xs_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id>; +template [[host_name("kernel_mul_mv_id_q8_0_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id>>; + +template [[host_name("kernel_mul_mv_id_q4_0_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id>>; +template [[host_name("kernel_mul_mv_id_q4_1_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id>>; +template [[host_name("kernel_mul_mv_id_q5_0_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id>>; +template [[host_name("kernel_mul_mv_id_q5_1_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id>>; + +template [[host_name("kernel_mul_mv_id_q2_K_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id>>; +template [[host_name("kernel_mul_mv_id_q3_K_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id>>; +template [[host_name("kernel_mul_mv_id_q4_K_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id>>; +template [[host_name("kernel_mul_mv_id_q5_K_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id>>; +template [[host_name("kernel_mul_mv_id_q6_K_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id>>; +template [[host_name("kernel_mul_mv_id_iq1_s_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id>>; +template [[host_name("kernel_mul_mv_id_iq1_m_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id>>; +template [[host_name("kernel_mul_mv_id_iq2_xxs_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id>>; +template [[host_name("kernel_mul_mv_id_iq2_xs_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id>>; +template [[host_name("kernel_mul_mv_id_iq3_xxs_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id>>; +template [[host_name("kernel_mul_mv_id_iq3_s_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id>>; +template [[host_name("kernel_mul_mv_id_iq2_s_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id>>; +template [[host_name("kernel_mul_mv_id_iq4_nl_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id>>; +template [[host_name("kernel_mul_mv_id_iq4_xs_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id>>; kernel void kernel_pool_2d_max_f32( device const float * src0, device float * dst, - constant int32_t & k0, - constant int32_t & k1, - constant int32_t & s0, - constant int32_t & s1, - constant int32_t & p0, - constant int32_t & p1, - constant int64_t & IH, - constant int64_t & IW, - constant int64_t & OH, - constant int64_t & OW, - constant int64_t & parallel_elements, + constant ggml_metal_kargs_pool_2d & args, uint gid[[thread_position_in_grid]]) { - if (gid >= parallel_elements) { + if (gid >= args.parallel_elements) { return; } const int idx = gid; - const int I_HW = IH * IW; - const int O_HW = OH * OW; + const int I_HW = args.IH * args.IW; + const int O_HW = args.OH * args.OW; const int nc = idx / O_HW; - const int cur_oh = idx % O_HW / OW; - const int cur_ow = idx % O_HW % OW; + const int cur_oh = idx % O_HW / args.OW; + const int cur_ow = idx % O_HW % args.OW; device const float * i_ptr = src0 + nc * I_HW; device float * o_ptr = dst + nc * O_HW; - const int start_h = cur_oh * s1 - p1; + const int start_h = cur_oh * args.s1 - args.p1; const int bh = MAX(0, start_h); - const int eh = MIN(IH, start_h + k1); - const int start_w = cur_ow * s0 - p0; + const int eh = MIN(args.IH, start_h + args.k1); + const int start_w = cur_ow * args.s0 - args.p0; const int bw = MAX(0, start_w); - const int ew = MIN(IW, start_w + k0); + const int ew = MIN(args.IW, start_w + args.k0); float res = -INFINITY; for (int i = bh; i < eh; i += 1) { for (int j = bw; j < ew; j += 1) { - res = MAX(res, i_ptr[i * IW + j]); + res = MAX(res, i_ptr[i * args.IW + j]); } } - o_ptr[cur_oh * OW + cur_ow] = res; + o_ptr[cur_oh * args.OW + cur_ow] = res; } kernel void kernel_pool_2d_avg_f32( device const float * src0, device float * dst, - constant int32_t & k0, - constant int32_t & k1, - constant int32_t & s0, - constant int32_t & s1, - constant int32_t & p0, - constant int32_t & p1, - constant int64_t & IH, - constant int64_t & IW, - constant int64_t & OH, - constant int64_t & OW, - constant int64_t & parallel_elements, + constant ggml_metal_kargs_pool_2d & args, uint gid[[thread_position_in_grid]]) { - if (gid >= parallel_elements) { + if (gid >= args.parallel_elements) { return; } const int idx = gid; - const int I_HW = IH * IW; - const int O_HW = OH * OW; + const int I_HW = args.IH * args.IW; + const int O_HW = args.OH * args.OW; const int nc = idx / O_HW; - const int cur_oh = idx % O_HW / OW; - const int cur_ow = idx % O_HW % OW; + const int cur_oh = idx % O_HW / args.OW; + const int cur_ow = idx % O_HW % args.OW; device const float * i_ptr = src0 + nc * I_HW; device float * o_ptr = dst + nc * O_HW; - const int start_h = cur_oh * s1 - p1; + const int start_h = cur_oh * args.s1 - args.p1; const int bh = MAX(0, start_h); - const int eh = MIN(IH, start_h + k1); - const int start_w = cur_ow * s0 - p0; + const int eh = MIN(args.IH, start_h + args.k1); + const int start_w = cur_ow * args.s0 - args.p0; const int bw = MAX(0, start_w); - const int ew = MIN(IW, start_w + k0); + const int ew = MIN(args.IW, start_w + args.k0); // const float scale = 1. / ((eh - bh) * (ew - bw)); - const float scale = 1. / (k0 * k1); + const float scale = 1. / (args.k0 * args.k1); float res = 0; for (int i = bh; i < eh; i += 1) { for (int j = bw; j < ew; j += 1) { - float cur = i_ptr[i * IW + j]; + float cur = i_ptr[i * args.IW + j]; res += cur * scale; } } - o_ptr[cur_oh * OW + cur_ow] = res; + o_ptr[cur_oh * args.OW + cur_ow] = res; } diff --git a/ggml/src/ggml-musa/CMakeLists.txt b/ggml/src/ggml-musa/CMakeLists.txt index 2f555416e..92f05d555 100644 --- a/ggml/src/ggml-musa/CMakeLists.txt +++ b/ggml/src/ggml-musa/CMakeLists.txt @@ -21,7 +21,7 @@ if (MUSAToolkit_FOUND) message(STATUS "MUSA Toolkit found") if (NOT DEFINED MUSA_ARCHITECTURES) - set(MUSA_ARCHITECTURES "21;22") + set(MUSA_ARCHITECTURES "21;22;31") endif() message(STATUS "Using MUSA architectures: ${MUSA_ARCHITECTURES}") @@ -49,7 +49,7 @@ if (MUSAToolkit_FOUND) set_source_files_properties(${GGML_SOURCES_MUSA} PROPERTIES LANGUAGE CXX) foreach(SOURCE ${GGML_SOURCES_MUSA}) - set(COMPILE_FLAGS "-x musa -mtgpu") + set(COMPILE_FLAGS "-fsigned-char -x musa -mtgpu") foreach(ARCH ${MUSA_ARCHITECTURES}) set(COMPILE_FLAGS "${COMPILE_FLAGS} --cuda-gpu-arch=mp_${ARCH}") endforeach() @@ -67,10 +67,6 @@ if (MUSAToolkit_FOUND) add_compile_definitions(GGML_USE_MUSA) add_compile_definitions(GGML_CUDA_PEER_MAX_BATCH_SIZE=${GGML_CUDA_PEER_MAX_BATCH_SIZE}) - if (GGML_CUDA_GRAPHS) - add_compile_definitions(GGML_CUDA_USE_GRAPHS) - endif() - if (GGML_CUDA_FORCE_MMQ) add_compile_definitions(GGML_CUDA_FORCE_MMQ) endif() @@ -83,6 +79,10 @@ if (MUSAToolkit_FOUND) add_compile_definitions(GGML_CUDA_NO_VMM) endif() + if (NOT GGML_CUDA_FA) + add_compile_definitions(GGML_CUDA_NO_FA) + endif() + if (GGML_CUDA_F16 OR GGML_CUDA_DMMV_F16) add_compile_definitions(GGML_CUDA_F16) endif() diff --git a/ggml/src/ggml-opencl/CMakeLists.txt b/ggml/src/ggml-opencl/CMakeLists.txt index 45328a657..624cb1b9d 100644 --- a/ggml/src/ggml-opencl/CMakeLists.txt +++ b/ggml/src/ggml-opencl/CMakeLists.txt @@ -15,6 +15,7 @@ if (GGML_OPENCL_PROFILING) endif () add_compile_definitions(GGML_OPENCL_SOA_Q) +add_compile_definitions(GGML_OPENCL_TARGET_VERSION=${GGML_OPENCL_TARGET_VERSION}) if (GGML_OPENCL_USE_ADRENO_KERNELS) message(STATUS "OpenCL will use matmul kernels optimized for Adreno") @@ -24,124 +25,47 @@ endif () if (GGML_OPENCL_EMBED_KERNELS) add_compile_definitions(GGML_OPENCL_EMBED_KERNELS) - set(OPENCL_CL_SOURCE_EMBED "${CMAKE_BINARY_DIR}/autogenerated/ggml-opencl.cl.h") - set(OPENCL_MM_CL_SOURCE_EMBED "${CMAKE_BINARY_DIR}/autogenerated/ggml-opencl_mm.cl.h") - set(OPENCL_CVT_CL_SOURCE_EMBED "${CMAKE_BINARY_DIR}/autogenerated/ggml-opencl_cvt.cl.h") + set(EMBED_KERNEL_SCRIPT "${CMAKE_CURRENT_SOURCE_DIR}/kernels/embed_kernel.py") + file(MAKE_DIRECTORY "${CMAKE_CURRENT_BINARY_DIR}/autogenerated") - set(OPENCL_GEMV_NOSHUFFLE_SOURCE_EMBED "${CMAKE_BINARY_DIR}/autogenerated/ggml-opencl_gemv_noshuffle.cl.h") - set(OPENCL_GEMV_NOSHUFFLE_GENERAL_SOURCE_EMBED "${CMAKE_BINARY_DIR}/autogenerated/ggml-opencl_gemv_noshuffle_general.cl.h") - set(OPENCL_MUL_MAT_Ab_Bi_8x4_SOURCE_EMBED "${CMAKE_BINARY_DIR}/autogenerated/ggml-opencl_mul_mat_Ab_Bi_8x4.cl.h") - set(OPENCL_TRANSPOSE_16_SOURCE_EMBED "${CMAKE_BINARY_DIR}/autogenerated/ggml-opencl_transpose_16.cl.h") - set(OPENCL_TRANSPOSE_32_SOURCE_EMBED "${CMAKE_BINARY_DIR}/autogenerated/ggml-opencl_transpose_32.cl.h") - set(OPENCL_TRANSPOSE_32_16_SOURCE_EMBED "${CMAKE_BINARY_DIR}/autogenerated/ggml-opencl_transpose_32_16.cl.h") - - set(EMBED_KERNEL_SCRIPT "${CMAKE_CURRENT_SOURCE_DIR}/kernels/embed_kernel.py") - file(MAKE_DIRECTORY "${CMAKE_BINARY_DIR}/autogenerated") - - include_directories("${CMAKE_BINARY_DIR}/autogenerated") - - # Python must be accessible from command line - add_custom_command( - OUTPUT ${OPENCL_CL_SOURCE_EMBED} - COMMAND ${Python3_EXECUTABLE} ${EMBED_KERNEL_SCRIPT} - ${CMAKE_CURRENT_SOURCE_DIR}/kernels/ggml-opencl.cl - ${OPENCL_CL_SOURCE_EMBED} - DEPENDS kernels/ggml-opencl.cl ${EMBED_KERNEL_SCRIPT} - COMMENT "Generate ggml-opencl.cl.h" - ) - - add_custom_command( - OUTPUT ${OPENCL_MM_CL_SOURCE_EMBED} - COMMAND ${Python3_EXECUTABLE} ${EMBED_KERNEL_SCRIPT} - ${CMAKE_CURRENT_SOURCE_DIR}/kernels/ggml-opencl_mm.cl - ${OPENCL_MM_CL_SOURCE_EMBED} - DEPENDS kernels/ggml-opencl_mm.cl ${EMBED_KERNEL_SCRIPT} - COMMENT "Generate ggml-opencl_mm.cl.h" - ) - - add_custom_command( - OUTPUT ${OPENCL_CVT_CL_SOURCE_EMBED} - COMMAND ${Python3_EXECUTABLE} ${EMBED_KERNEL_SCRIPT} - ${CMAKE_CURRENT_SOURCE_DIR}/kernels/ggml-opencl_cvt.cl - ${OPENCL_CVT_CL_SOURCE_EMBED} - DEPENDS kernels/ggml-opencl_cvt.cl ${EMBED_KERNEL_SCRIPT} - COMMENT "Generate ggml-opencl_cvt.cl.h" - ) - - add_custom_command( - OUTPUT ${OPENCL_GEMV_NOSHUFFLE_SOURCE_EMBED} - COMMAND ${Python3_EXECUTABLE} ${EMBED_KERNEL_SCRIPT} - ${CMAKE_CURRENT_SOURCE_DIR}/kernels/ggml-opencl_gemv_noshuffle.cl - ${OPENCL_GEMV_NOSHUFFLE_SOURCE_EMBED} - DEPENDS kernels/ggml-opencl_gemv_noshuffle.cl ${EMBED_KERNEL_SCRIPT} - COMMENT "Generate ggml-opencl_gemv_noshuffle.cl.h" - ) - - add_custom_command( - OUTPUT ${OPENCL_GEMV_NOSHUFFLE_GENERAL_SOURCE_EMBED} - COMMAND ${Python3_EXECUTABLE} ${EMBED_KERNEL_SCRIPT} - ${CMAKE_CURRENT_SOURCE_DIR}/kernels/ggml-opencl_gemv_noshuffle_general.cl - ${OPENCL_GEMV_NOSHUFFLE_GENERAL_SOURCE_EMBED} - DEPENDS kernels/ggml-opencl_gemv_noshuffle_general.cl ${EMBED_KERNEL_SCRIPT} - COMMENT "Generate ggml-opencl_gemv_noshuffle_general.cl.h" - ) - - add_custom_command( - OUTPUT ${OPENCL_MUL_MAT_Ab_Bi_8x4_SOURCE_EMBED} - COMMAND ${Python3_EXECUTABLE} ${EMBED_KERNEL_SCRIPT} - ${CMAKE_CURRENT_SOURCE_DIR}/kernels/ggml-opencl_mul_mat_Ab_Bi_8x4.cl - ${OPENCL_MUL_MAT_Ab_Bi_8x4_SOURCE_EMBED} - DEPENDS kernels/ggml-opencl_mul_mat_Ab_Bi_8x4.cl ${EMBED_KERNEL_SCRIPT} - COMMENT "Generate ggml-opencl_mul_mat_Ab_Bi_8x4.cl.cl.h" - ) - - add_custom_command( - OUTPUT ${OPENCL_TRANSPOSE_16_SOURCE_EMBED} - COMMAND ${Python3_EXECUTABLE} ${EMBED_KERNEL_SCRIPT} - ${CMAKE_CURRENT_SOURCE_DIR}/kernels/ggml-opencl_transpose_16.cl - ${OPENCL_TRANSPOSE_16_SOURCE_EMBED} - DEPENDS kernels/ggml-opencl_transpose_16.cl ${EMBED_KERNEL_SCRIPT} - COMMENT "Generate ggml-opencl_transpose_16.cl.h" - ) - - add_custom_command( - OUTPUT ${OPENCL_TRANSPOSE_32_SOURCE_EMBED} - COMMAND ${Python3_EXECUTABLE} ${EMBED_KERNEL_SCRIPT} - ${CMAKE_CURRENT_SOURCE_DIR}/kernels/ggml-opencl_transpose_32.cl - ${OPENCL_TRANSPOSE_32_SOURCE_EMBED} - DEPENDS kernels/ggml-opencl_transpose_32.cl ${EMBED_KERNEL_SCRIPT} - COMMENT "Generate ggml-opencl_transpose_32.cl.h" - ) - - add_custom_command( - OUTPUT ${OPENCL_TRANSPOSE_32_16_SOURCE_EMBED} - COMMAND ${Python3_EXECUTABLE} ${EMBED_KERNEL_SCRIPT} - ${CMAKE_CURRENT_SOURCE_DIR}/kernels/ggml-opencl_transpose_32_16.cl - ${OPENCL_TRANSPOSE_32_16_SOURCE_EMBED} - DEPENDS kernels/ggml-opencl_transpose_32_16.cl ${EMBED_KERNEL_SCRIPT} - COMMENT "Generate ggml-opencl_transpose_32_16.cl.h" - ) - - target_sources(${TARGET_NAME} PRIVATE - ${OPENCL_CL_SOURCE_EMBED} - ${OPENCL_MM_CL_SOURCE_EMBED} - ${OPENCL_CVT_CL_SOURCE_EMBED} - ${OPENCL_GEMV_NOSHUFFLE_SOURCE_EMBED} - ${OPENCL_GEMV_NOSHUFFLE_GENERAL_SOURCE_EMBED} - ${OPENCL_MUL_MAT_Ab_Bi_8x4_SOURCE_EMBED} - ${OPENCL_TRANSPOSE_16_SOURCE_EMBED} - ${OPENCL_TRANSPOSE_32_SOURCE_EMBED} - ${OPENCL_TRANSPOSE_32_16_SOURCE_EMBED}) -else () - # copy ggml-opencl.cl to bin directory - configure_file(kernels/ggml-opencl.cl ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-opencl.cl COPYONLY) - configure_file(kernels/ggml-opencl_mm.cl ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-opencl_mm.cl COPYONLY) - configure_file(kernels/ggml-opencl_cvt.cl ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-opencl_cvt.cl COPYONLY) - - configure_file(kernels/ggml-opencl_gemv_noshuffle.cl ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-opencl_gemv_noshuffle.cl COPYONLY) - configure_file(kernels/ggml-opencl_gemv_noshuffle_general.cl ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-opencl_gemv_noshuffle_general.cl COPYONLY) - configure_file(kernels/ggml-opencl_mul_mat_Ab_Bi_8x4.cl ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-opencl_mul_mat_Ab_Bi_8x4.cl COPYONLY) - configure_file(kernels/ggml-opencl_transpose_16.cl ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-opencl_transpose_16.cl COPYONLY) - configure_file(kernels/ggml-opencl_transpose_32.cl ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-opencl_transpose_32.cl COPYONLY) - configure_file(kernels/ggml-opencl_transpose_32_16.cl ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-opencl_transpose_32_16.cl COPYONLY) + target_include_directories(${TARGET_NAME} PRIVATE "${CMAKE_CURRENT_BINARY_DIR}/autogenerated") endif () + +function(ggml_opencl_add_kernel KNAME) + set(KERN_HDR ${CMAKE_CURRENT_BINARY_DIR}/autogenerated/${KNAME}.cl.h) + set(KERN_SRC ${CMAKE_CURRENT_SOURCE_DIR}/kernels/${KNAME}.cl) + + if (GGML_OPENCL_EMBED_KERNELS) + message(STATUS "opencl: embedding kernel ${KNAME}") + + # Python must be accessible from command line + add_custom_command( + OUTPUT ${KERN_HDR} + COMMAND ${Python3_EXECUTABLE} ${EMBED_KERNEL_SCRIPT} ${KERN_SRC} ${KERN_HDR} + DEPENDS ${KERN_SRC} ${EMBED_KERNEL_SCRIPT} + COMMENT "Generate ${KERN_HDR}" + ) + + target_sources(${TARGET_NAME} PRIVATE ${KERN_HDR}) + else () + message(STATUS "opencl: adding kernel ${KNAME}") + configure_file(${KERN_SRC} ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/${KNAME}.cl COPYONLY) + endif () +endfunction() + +set(GGML_OPENCL_KERNELS + ggml-opencl + ggml-opencl_mm + ggml-opencl_cvt + ggml-opencl_gemv_noshuffle + ggml-opencl_gemv_noshuffle_general + ggml-opencl_mul_mat_Ab_Bi_8x4 + ggml-opencl_transpose_16 + ggml-opencl_transpose_32 + ggml-opencl_transpose_32_16 + ggml-opencl_im2col +) + +foreach (K ${GGML_OPENCL_KERNELS}) + ggml_opencl_add_kernel(${K}) +endforeach() diff --git a/ggml/src/ggml-opencl/ggml-opencl.cpp b/ggml/src/ggml-opencl/ggml-opencl.cpp index ed90e471a..b8b5cbd3e 100644 --- a/ggml/src/ggml-opencl/ggml-opencl.cpp +++ b/ggml/src/ggml-opencl/ggml-opencl.cpp @@ -1,4 +1,4 @@ -#define CL_TARGET_OPENCL_VERSION 220 +#define CL_TARGET_OPENCL_VERSION GGML_OPENCL_TARGET_VERSION #define CL_USE_DEPRECATED_OPENCL_1_2_APIS // suppress warnings in CL headers for GCC and Clang @@ -25,6 +25,8 @@ #include #include #include +#include +#include #undef MIN #undef MAX @@ -62,6 +64,97 @@ enum ADRENO_GPU_GEN { X1E, }; +struct ggml_cl_version { + cl_uint major = 0; + cl_uint minor = 0; +}; + +// Parses a version string of form "XX.YY ". On an error returns ggml_cl_version with all zeroes. +static ggml_cl_version parse_cl_version(std::string_view str) { + size_t major_str_begin = 0; + size_t major_str_end = str.find(".", major_str_begin); + if (major_str_end == std::string::npos) { + return {}; + } + + size_t minor_str_begin = major_str_end + 1; + size_t minor_str_end = str.find(" ", minor_str_begin); + if (minor_str_end == std::string::npos) { + return {}; + } + + cl_uint version_major; + if (std::from_chars(str.data() + major_str_begin, str.data() + major_str_end, version_major).ec != std::errc{}) { + return {}; + } + + cl_uint version_minor; + if (std::from_chars(str.data() + minor_str_begin, str.data() + minor_str_end, version_minor).ec != std::errc{}) { + return {}; + } + return { version_major, version_minor }; +} + +// Returns OpenCL platform's version. On an error returns ggml_cl_version with all zeroes. +static ggml_cl_version get_opencl_platform_version(cl_platform_id platform) { + size_t param_size; + CL_CHECK(clGetPlatformInfo(platform, CL_PLATFORM_VERSION, 0, nullptr, ¶m_size)); + std::unique_ptr param_storage(new char[param_size]); + CL_CHECK(clGetPlatformInfo(platform, CL_PLATFORM_VERSION, param_size, param_storage.get(), nullptr)); + + auto param_value = std::string_view(param_storage.get(), param_size); + const std::string version_prefix = "OpenCL "; // Suffix: "XX.YY " + if (param_value.find(version_prefix) != 0) { + return {}; + } + param_value.remove_prefix(version_prefix.length()); + return parse_cl_version(param_value); +} + +// Return a version to use in OpenCL C compilation. On an error returns ggml_cl_version with all zeroes. +static ggml_cl_version get_opencl_c_version(ggml_cl_version platform_version, cl_device_id device) { + size_t param_size; + +#if CL_TARGET_OPENCL_VERSION >= 300 + if (platform_version.major >= 3) { + CL_CHECK(clGetDeviceInfo(device, CL_DEVICE_OPENCL_C_ALL_VERSIONS, 0, nullptr, ¶m_size)); + if (!param_size) { + return {}; + } + + std::unique_ptr versions(new cl_name_version[param_size]); + CL_CHECK(clGetDeviceInfo(device, CL_DEVICE_OPENCL_C_ALL_VERSIONS, param_size, versions.get(), nullptr)); + unsigned versions_count = param_size / sizeof(cl_name_version); + + cl_version version_max = 0; + for (unsigned i = 0; i < versions_count; i++) { + version_max = std::max(versions[i].version, version_max); + } + + return { CL_VERSION_MAJOR(version_max), CL_VERSION_MINOR(version_max) }; + } +#else + GGML_UNUSED(platform_version); +#endif // CL_TARGET_OPENCL_VERSION >= 300 + + CL_CHECK(clGetDeviceInfo(device, CL_DEVICE_OPENCL_C_VERSION, 0, nullptr, ¶m_size)); + if (!param_size) { + return {}; + } + + std::unique_ptr param_storage(new char[param_size]); + CL_CHECK(clGetDeviceInfo(device, CL_DEVICE_OPENCL_C_VERSION, param_size, param_storage.get(), nullptr)); + auto param_value = std::string_view(param_storage.get(), param_size); + + const std::string version_prefix = "OpenCL C "; // Suffix: "XX.YY " + if (param_value.find(version_prefix) != 0) { + return {}; + } + param_value.remove_prefix(version_prefix.length()); + + return parse_cl_version(param_value); +} + static ADRENO_GPU_GEN get_adreno_gpu_gen(const char *device_name) { if (strstr(device_name, "730") || strstr(device_name, "740") || @@ -131,20 +224,24 @@ struct ggml_backend_opencl_context { cl_program program; cl_program program_1; cl_program program_2; + cl_program program_im2col; cl_kernel kernel_add, kernel_add_row; cl_kernel kernel_mul, kernel_mul_row; cl_kernel kernel_scale; cl_kernel kernel_silu, kernel_silu_4; cl_kernel kernel_gelu, kernel_gelu_4; + cl_kernel kernel_gelu_quick, kernel_gelu_quick_4; cl_kernel kernel_relu; cl_kernel kernel_clamp; cl_kernel kernel_norm; cl_kernel kernel_rms_norm; cl_kernel kernel_diag_mask_inf, kernel_diag_mask_inf_8; cl_kernel kernel_soft_max, kernel_soft_max_4; + cl_kernel kernel_soft_max_f16, kernel_soft_max_4_f16; cl_kernel kernel_get_rows_f32, kernel_get_rows_f16, kernel_get_rows_q4_0; cl_kernel kernel_rope_norm_f32, kernel_rope_norm_f16, kernel_rope_neox_f32, kernel_rope_neox_f16; + cl_kernel kernel_rope_multi_f32, kernel_rope_multi_f16, kernel_rope_vision_f32, kernel_rope_vision_f16; cl_kernel kernel_cpy_f16_f16, kernel_cpy_f16_f32, kernel_cpy_f32_f16, kernel_cpy_f32_f32; cl_kernel kernel_mul_mat_f32_f32; cl_kernel kernel_mul_mat_f16_f16; @@ -158,6 +255,7 @@ struct ggml_backend_opencl_context { kernel_mul_mat_q4_0_f32_flat_img_v0; cl_kernel kernel_mul_mat_q4_0_f32_1d_8x_flat, kernel_mul_mat_q4_0_f32_1d_16x_flat; cl_kernel kernel_mul_mv_q6_K_f32; + cl_kernel kernel_im2col_f32, kernel_im2col_f16; #ifdef GGML_OPENCL_USE_ADRENO_KERNELS // Transpose kernels @@ -203,8 +301,27 @@ static int ggml_backend_opencl_n_devices = 0; struct ProfilingInfo { std::string op_name; std::string kernel_name; - // Kernel execution time in nanoseconds. - cl_ulong duration_ns; + + cl_kernel kernel; + cl_event evt; + + cl_ulong cmd_queued; + cl_ulong cmd_submit; + cl_ulong cmd_start; + cl_ulong cmd_end; + cl_ulong overhead_start; + cl_ulong overhead_end; + // For the times below, see spec for clGetEventProfilingInfo + // The time kernel spent in cmd queue - SUBMIT - QUEUED + cl_ulong cmd_queued_duration_ns; + // The time kernel spent for submission - START - SUBMIT + cl_ulong cmd_submit_duration_ns; + // Kernel execution time in nanoseconds - END - START + cl_ulong cmd_duration_ns; + // The time for the kernel to complete - COMPLETE - END + cl_ulong cmd_complete_duration_ns; + // Total time to finish the kernel - COMPELTE - QUEUED + cl_ulong cmd_total_duration_ns; // Global and local work sizes. size_t global_size[3]; size_t local_size[3]; @@ -277,7 +394,7 @@ static ggml_backend_opencl_context * ggml_cl2_init(ggml_backend_dev_t dev) { cl_int err; -#ifdef GGML_PROFILE_OPENCL +#ifdef GGML_OPENCL_PROFILING GGML_LOG_INFO("ggml_opencl: OpenCL profiling enabled\n"); #endif @@ -298,6 +415,7 @@ static ggml_backend_opencl_context * ggml_cl2_init(ggml_backend_dev_t dev) { unsigned number; cl_device_type type; char name[128]; + char version[128]; }; enum { NPLAT = 16, NDEV = 16 }; @@ -338,6 +456,7 @@ static ggml_backend_opencl_context * ggml_cl2_init(ggml_backend_dev_t dev) { d->platform = p; CL_CHECK(clGetDeviceInfo(d->id, CL_DEVICE_NAME, sizeof(d->name), &d->name, NULL)); CL_CHECK(clGetDeviceInfo(d->id, CL_DEVICE_TYPE, sizeof(d->type), &d->type, NULL)); + CL_CHECK(clGetDeviceInfo(d->id, CL_DEVICE_VERSION, sizeof(d->version), &d->version, NULL)); if (p->default_device == NULL && d->type == CL_DEVICE_TYPE_GPU) { p->default_device = d; @@ -430,7 +549,7 @@ static ggml_backend_opencl_context * ggml_cl2_init(ggml_backend_dev_t dev) { } GGML_LOG_INFO("ggml_opencl: selecting platform: '%s'\n", default_device->platform->name); - GGML_LOG_INFO("ggml_opencl: selecting device: '%s'\n", default_device->name); + GGML_LOG_INFO("ggml_opencl: selecting device: '%s (%s)'\n", default_device->name, default_device->version); if (default_device->type != CL_DEVICE_TYPE_GPU) { GGML_LOG_WARN("ggml_opencl: warning, not a GPU: '%s'.\n", default_device->name); } @@ -439,23 +558,18 @@ static ggml_backend_opencl_context * ggml_cl2_init(ggml_backend_dev_t dev) { dev_ctx->device = default_device->id; backend_ctx->device = default_device->id; - if (strstr(default_device->name, "Adreno")) { + if (strstr(default_device->name, "Adreno") || + strstr(default_device->name, "Qualcomm") || + strstr(default_device->version, "Adreno")) { backend_ctx->gpu_family = GPU_FAMILY::ADRENO; - backend_ctx->adreno_gen = get_adreno_gpu_gen(default_device->name); - - // Default wave size is 128, A8x uses 64. - if (backend_ctx->adreno_gen == ADRENO_GPU_GEN::A8X) { - backend_ctx->adreno_wave_size = 64; - } else if (backend_ctx->adreno_gen == ADRENO_GPU_GEN::A7X || - backend_ctx->adreno_gen == ADRENO_GPU_GEN::X1E) { - backend_ctx->adreno_wave_size = 128; - } else { - backend_ctx->adreno_wave_size = 128; - GGML_LOG_WARN("ggml_opencl: Unsupported Adreno GPU: %s, " - "using wave size %d, " - "may not work as expected\n", - backend_ctx->device_name.c_str(), backend_ctx->adreno_wave_size); + // Usually device version contains the detailed device name + backend_ctx->adreno_gen = get_adreno_gpu_gen(default_device->version); + if (backend_ctx->adreno_gen == ADRENO_GPU_GEN::ADRENO_UNKNOWN) { + backend_ctx->adreno_gen = get_adreno_gpu_gen(default_device->name); } + + // Use wave size of 64 for all Adreno GPUs. + backend_ctx->adreno_wave_size = 64; } else if (strstr(default_device->name, "Intel")) { backend_ctx->gpu_family = GPU_FAMILY::INTEL; } else { @@ -480,16 +594,11 @@ static ggml_backend_opencl_context * ggml_cl2_init(ggml_backend_dev_t dev) { // A local ref of cl_device_id for convenience cl_device_id device = backend_ctx->device; - // Check device OpenCL version, OpenCL 2.0 or above is required - size_t device_ver_str_size; - clGetDeviceInfo(device, CL_DEVICE_VERSION, 0, NULL, &device_ver_str_size); - char *device_ver_buffer = (char *)alloca(device_ver_str_size + 1); - clGetDeviceInfo(device, CL_DEVICE_VERSION, device_ver_str_size, device_ver_buffer, NULL); - device_ver_buffer[device_ver_str_size] = '\0'; - GGML_LOG_INFO("ggml_opencl: device OpenCL version: %s\n", device_ver_buffer); + ggml_cl_version platform_version = get_opencl_platform_version(default_device->platform->id); - if (strstr(device_ver_buffer, "OpenCL 2") == NULL && - strstr(device_ver_buffer, "OpenCL 3") == NULL) { + // Check device OpenCL version, OpenCL 2.0 or above is required + ggml_cl_version opencl_c_version = get_opencl_c_version(platform_version, device); + if (opencl_c_version.major < 2) { GGML_LOG_ERROR("ggml_opencl: OpenCL 2.0 or above is required\n"); return backend_ctx; } @@ -526,15 +635,17 @@ static ggml_backend_opencl_context * ggml_cl2_init(ggml_backend_dev_t dev) { // If OpenCL 3.0 is supported, then check for cl_khr_subgroups, which becomes // optional in OpenCL 3.0 (cl_khr_subgroup is mandatory in OpenCL 2.x) - if (strstr(device_ver_buffer, "OpenCL 3") && - strstr(ext_buffer, "cl_khr_subgroups") == NULL && + if (opencl_c_version.major == 3 && strstr(ext_buffer, "cl_khr_subgroups") == NULL && strstr(ext_buffer, "cl_intel_subgroups") == NULL) { GGML_LOG_ERROR("ggml_opencl: device does not support subgroups (cl_khr_subgroups or cl_intel_subgroups) " "(note that subgroups is an optional feature in OpenCL 3.0)\n"); return backend_ctx; } - CL_CHECK(clGetDeviceInfo(device, CL_DEVICE_MEM_BASE_ADDR_ALIGN, sizeof(cl_uint), &backend_ctx->alignment, NULL)); + cl_uint base_align_in_bits; + CL_CHECK(clGetDeviceInfo(device, CL_DEVICE_MEM_BASE_ADDR_ALIGN, sizeof(cl_uint), &base_align_in_bits, NULL)); + GGML_ASSERT(base_align_in_bits % 8u == 0); + backend_ctx->alignment = base_align_in_bits / 8u; GGML_LOG_INFO("ggml_opencl: mem base addr align: %u\n", backend_ctx->alignment); clGetDeviceInfo(device, CL_DEVICE_MAX_MEM_ALLOC_SIZE, sizeof(size_t), &backend_ctx->max_alloc_size, NULL); @@ -588,9 +699,12 @@ static ggml_backend_opencl_context * ggml_cl2_init(ggml_backend_dev_t dev) { const std::string kernel_src = read_file("ggml-opencl.cl"); #endif - std::string compile_opts = - "-cl-std=CL2.0 -cl-mad-enable -cl-unsafe-math-optimizations " - "-cl-finite-math-only -cl-fast-relaxed-math "; + auto opencl_c_std = + std::string("CL") + std::to_string(opencl_c_version.major) + "." + std::to_string(opencl_c_version.minor); + + std::string compile_opts = std::string("-cl-std=") + opencl_c_std + + " -cl-mad-enable -cl-unsafe-math-optimizations" + " -cl-finite-math-only -cl-fast-relaxed-math"; backend_ctx->program = build_program_from_source(context, device, kernel_src.c_str(), compile_opts); // Non matmul kernels. @@ -606,6 +720,8 @@ static ggml_backend_opencl_context * ggml_cl2_init(ggml_backend_dev_t dev) { CL_CHECK((backend_ctx->kernel_silu_4 = clCreateKernel(backend_ctx->program, "kernel_silu_4", &err), err)); CL_CHECK((backend_ctx->kernel_gelu = clCreateKernel(backend_ctx->program, "kernel_gelu", &err), err)); CL_CHECK((backend_ctx->kernel_gelu_4 = clCreateKernel(backend_ctx->program, "kernel_gelu_4", &err), err)); + CL_CHECK((backend_ctx->kernel_gelu_quick = clCreateKernel(backend_ctx->program, "kernel_gelu_quick", &err), err)); + CL_CHECK((backend_ctx->kernel_gelu_quick_4 = clCreateKernel(backend_ctx->program, "kernel_gelu_quick_4", &err), err)); CL_CHECK((backend_ctx->kernel_relu = clCreateKernel(backend_ctx->program, "kernel_relu", &err), err)); CL_CHECK((backend_ctx->kernel_clamp = clCreateKernel(backend_ctx->program, "kernel_clamp", &err), err)); CL_CHECK((backend_ctx->kernel_norm = clCreateKernel(backend_ctx->program, "kernel_norm", &err), err)); @@ -614,10 +730,16 @@ static ggml_backend_opencl_context * ggml_cl2_init(ggml_backend_dev_t dev) { CL_CHECK((backend_ctx->kernel_diag_mask_inf_8 = clCreateKernel(backend_ctx->program, "kernel_diag_mask_inf_8", &err), err)); CL_CHECK((backend_ctx->kernel_soft_max = clCreateKernel(backend_ctx->program, "kernel_soft_max", &err), err)); CL_CHECK((backend_ctx->kernel_soft_max_4 = clCreateKernel(backend_ctx->program, "kernel_soft_max_4", &err), err)); + CL_CHECK((backend_ctx->kernel_soft_max_f16 = clCreateKernel(backend_ctx->program, "kernel_soft_max_f16", &err), err)); + CL_CHECK((backend_ctx->kernel_soft_max_4_f16 = clCreateKernel(backend_ctx->program, "kernel_soft_max_4_f16", &err), err)); CL_CHECK((backend_ctx->kernel_rope_norm_f32 = clCreateKernel(backend_ctx->program, "kernel_rope_norm_f32", &err), err)); CL_CHECK((backend_ctx->kernel_rope_norm_f16 = clCreateKernel(backend_ctx->program, "kernel_rope_norm_f16", &err), err)); CL_CHECK((backend_ctx->kernel_rope_neox_f32 = clCreateKernel(backend_ctx->program, "kernel_rope_neox_f32", &err), err)); CL_CHECK((backend_ctx->kernel_rope_neox_f16 = clCreateKernel(backend_ctx->program, "kernel_rope_neox_f16", &err), err)); + CL_CHECK((backend_ctx->kernel_rope_multi_f32 = clCreateKernel(backend_ctx->program, "kernel_rope_multi_f32", &err), err)); + CL_CHECK((backend_ctx->kernel_rope_multi_f16 = clCreateKernel(backend_ctx->program, "kernel_rope_multi_f16", &err), err)); + CL_CHECK((backend_ctx->kernel_rope_vision_f32 = clCreateKernel(backend_ctx->program, "kernel_rope_vision_f32", &err), err)); + CL_CHECK((backend_ctx->kernel_rope_vision_f16 = clCreateKernel(backend_ctx->program, "kernel_rope_vision_f16", &err), err)); CL_CHECK((backend_ctx->kernel_cpy_f16_f16 = clCreateKernel(backend_ctx->program, "kernel_cpy_f16_f16", &err), err)); CL_CHECK((backend_ctx->kernel_cpy_f16_f32 = clCreateKernel(backend_ctx->program, "kernel_cpy_f16_f32", &err), err)); CL_CHECK((backend_ctx->kernel_cpy_f32_f16 = clCreateKernel(backend_ctx->program, "kernel_cpy_f32_f16", &err), err)); @@ -665,6 +787,19 @@ static ggml_backend_opencl_context * ggml_cl2_init(ggml_backend_dev_t dev) { CL_CHECK((backend_ctx->kernel_convert_block_q4_0_noshuffle = clCreateKernel(backend_ctx->program_2, "kernel_convert_block_q4_0_noshuffle", &err), err)); + // im2col kernels +#ifdef GGML_OPENCL_EMBED_KERNELS + const std::string kernel_src_im2col { + #include "ggml-opencl_im2col.cl.h" + }; +#else + const std::string kernel_src_im2col = read_file("ggml-opencl_im2col.cl"); +#endif + backend_ctx->program_im2col = build_program_from_source(context, device, kernel_src_im2col.c_str(), compile_opts); + + CL_CHECK((backend_ctx->kernel_im2col_f32 = clCreateKernel(backend_ctx->program_im2col, "kernel_im2col_f32", &err), err)); + CL_CHECK((backend_ctx->kernel_im2col_f16 = clCreateKernel(backend_ctx->program_im2col, "kernel_im2col_f16", &err), err)); + // Kernels for Adreno #ifdef GGML_OPENCL_USE_ADRENO_KERNELS #ifdef GGML_OPENCL_EMBED_KERNELS @@ -698,10 +833,10 @@ static ggml_backend_opencl_context * ggml_cl2_init(ggml_backend_dev_t dev) { CL_CHECK((backend_ctx->kernel_transpose_16 = clCreateKernel(backend_ctx->program_transpose_16, "kernel_transpose_16", &err), err)); // Gemv general - std::string CL_gemv_compile_opts = - " -cl-std=CL2.0 " - " -cl-mad-enable " - " -DSIMDGROUP_WIDTH=" + std::to_string(backend_ctx->adreno_wave_size); + std::string CL_gemv_compile_opts = std::string("-cl-std=") + opencl_c_std + + " -cl-mad-enable " + " -DSIMDGROUP_WIDTH=" + + std::to_string(backend_ctx->adreno_wave_size); if (has_vector_subgroup_broadcast) { CL_gemv_compile_opts += " -DVECTOR_SUB_GROUP_BROADCAT "; } @@ -718,12 +853,12 @@ static ggml_backend_opencl_context * ggml_cl2_init(ggml_backend_dev_t dev) { CL_CHECK((backend_ctx->CL_mul_mat_vec_q4_0_f32_1d_4x_flat_general = clCreateKernel(backend_ctx->program_CL_gemv_general, "kernel_gemv_noshuffle", &err), err)); // Gemv 2048, 16384 - CL_gemv_compile_opts = - " -cl-std=CL2.0 " - " -cl-mad-enable " - " -DLINE_STRIDE_A=2048 " - " -DBLOCK_STRIDE_A=16384 " - " -DSIMDGROUP_WIDTH=" + std::to_string(backend_ctx->adreno_wave_size); + CL_gemv_compile_opts = std::string("-cl-std=") + opencl_c_std + + " -cl-mad-enable " + " -DLINE_STRIDE_A=2048 " + " -DBLOCK_STRIDE_A=16384 " + " -DSIMDGROUP_WIDTH=" + + std::to_string(backend_ctx->adreno_wave_size); if (has_vector_subgroup_broadcast) { CL_gemv_compile_opts += " -DVECTOR_SUB_GROUP_BROADCAT "; } @@ -740,12 +875,12 @@ static ggml_backend_opencl_context * ggml_cl2_init(ggml_backend_dev_t dev) { CL_CHECK((backend_ctx->CL_mul_mat_vec_q4_0_f32_1d_4x_flat_4096_1_4096 = clCreateKernel(backend_ctx->program_CL_gemv_4096_1_4096, "kernel_gemv_noshuffle", &err), err)); // Gemv 2048, 16384 - CL_gemv_compile_opts = - " -cl-std=CL2.0 " - " -cl-mad-enable " - " -DLINE_STRIDE_A=2048 " - " -DBLOCK_STRIDE_A=16384 " - " -DSIMDGROUP_WIDTH=" + std::to_string(backend_ctx->adreno_wave_size); + CL_gemv_compile_opts = std::string("-cl-std=") + opencl_c_std + + " -cl-mad-enable " + " -DLINE_STRIDE_A=2048 " + " -DBLOCK_STRIDE_A=16384 " + " -DSIMDGROUP_WIDTH=" + + std::to_string(backend_ctx->adreno_wave_size); if (has_vector_subgroup_broadcast) { CL_gemv_compile_opts += " -DVECTOR_SUB_GROUP_BROADCAT "; } @@ -755,12 +890,12 @@ static ggml_backend_opencl_context * ggml_cl2_init(ggml_backend_dev_t dev) { CL_CHECK((backend_ctx->CL_mul_mat_vec_q4_0_f32_1d_4x_flat_4096_1_11008 = clCreateKernel(backend_ctx->program_CL_gemv_4096_1_11008, "kernel_gemv_noshuffle", &err), err)); // Gemv 5504, 44032 - CL_gemv_compile_opts = - " -cl-std=CL2.0 " - " -cl-mad-enable " - " -DLINE_STRIDE_A=5504 " - " -DBLOCK_STRIDE_A=44032 " - " -DSIMDGROUP_WIDTH=" + std::to_string(backend_ctx->adreno_wave_size); + CL_gemv_compile_opts = std::string("-cl-std=") + opencl_c_std + + " -cl-mad-enable " + " -DLINE_STRIDE_A=5504 " + " -DBLOCK_STRIDE_A=44032 " + " -DSIMDGROUP_WIDTH=" + + std::to_string(backend_ctx->adreno_wave_size); if (has_vector_subgroup_broadcast) { CL_gemv_compile_opts += " -DVECTOR_SUB_GROUP_BROADCAT "; } @@ -770,12 +905,12 @@ static ggml_backend_opencl_context * ggml_cl2_init(ggml_backend_dev_t dev) { CL_CHECK((backend_ctx->CL_mul_mat_vec_q4_0_f32_1d_4x_flat_11008_1_4096 = clCreateKernel(backend_ctx->program_CL_gemv_11008_1_4096, "kernel_gemv_noshuffle", &err), err)); // Gemv 16000, 128000 - CL_gemv_compile_opts = - " -cl-std=CL2.0 " - " -cl-mad-enable " - " -DLINE_STRIDE_A=16000 " - " -DBLOCK_STRIDE_A=128000 " - " -DSIMDGROUP_WIDTH=" + std::to_string(backend_ctx->adreno_wave_size); + CL_gemv_compile_opts = std::string("-cl-std=") + opencl_c_std + + " -cl-mad-enable " + " -DLINE_STRIDE_A=16000 " + " -DBLOCK_STRIDE_A=128000 " + " -DSIMDGROUP_WIDTH=" + + std::to_string(backend_ctx->adreno_wave_size); if (has_vector_subgroup_broadcast) { CL_gemv_compile_opts += " -DVECTOR_SUB_GROUP_BROADCAT "; } @@ -794,10 +929,30 @@ static ggml_backend_opencl_context * ggml_cl2_init(ggml_backend_dev_t dev) { backend_ctx->program_CL_gemm = build_program_from_source(context, device, kernel_src_CL_gemm.c_str(), compile_opts); CL_CHECK((backend_ctx->CL_mul_mat_Ab_Bi_8x4 = clCreateKernel(backend_ctx->program_CL_gemm, "kernel_mul_mat_Ab_Bi_8x4", &err), err)); + // TODO: fixme: these sizes are hardcoded for now. + // they should be allocated based on the model's size + // and the device's max alloc size // Allocate intermediate buffers and images - size_t max_A_q_d_bytes = 311164928; - size_t max_A_s_d_bytes = 38895616; - size_t max_B_d_bytes = 45088768; + size_t required_A_q_d_bytes = 311164928; + size_t required_A_s_d_bytes = 38895616; + size_t required_B_d_bytes = 45088768; + + // Ensure buffer sizes do not exceed the maximum allocation size + size_t max_A_q_d_bytes = MIN(required_A_q_d_bytes, backend_ctx->max_alloc_size); + size_t max_A_s_d_bytes = MIN(required_A_s_d_bytes, backend_ctx->max_alloc_size); + size_t max_B_d_bytes = MIN(required_B_d_bytes, backend_ctx->max_alloc_size); + if (required_A_q_d_bytes > backend_ctx->max_alloc_size) { + GGML_LOG_WARN("ggml_opencl: A_q_d buffer size reduced from %zu to %zu due to device limitations.\n", + required_A_q_d_bytes, max_A_q_d_bytes); + } + if (required_A_s_d_bytes > backend_ctx->max_alloc_size) { + GGML_LOG_WARN("ggml_opencl: A_s_d buffer size reduced from %zu to %zu due to device limitations.\n", + required_A_s_d_bytes, max_A_s_d_bytes); + } + if (required_B_d_bytes > backend_ctx->max_alloc_size) { + GGML_LOG_WARN("ggml_opencl: B_d buffer size reduced from %zu to %zu due to device limitations.\n", + required_B_d_bytes, max_B_d_bytes); + } CL_CHECK((backend_ctx->A_q_d_max = clCreateBuffer(context, 0, max_A_q_d_bytes, NULL, &err), err)); CL_CHECK((backend_ctx->A_s_d_max = clCreateBuffer(context, 0, max_A_s_d_bytes, NULL, &err), err)); @@ -818,12 +973,56 @@ static void ggml_cl2_free(void) { return; } + // Populate profiling info + for (ProfilingInfo & info : g_profiling_info) { + cl_ulong cmd_queued; + cl_ulong cmd_submit; + cl_ulong cmd_start; + cl_ulong cmd_end; + cl_ulong cmd_complete; + + CL_CHECK(clWaitForEvents(1, &info.evt)); + CL_CHECK(clGetEventProfilingInfo( + info.evt, CL_PROFILING_COMMAND_QUEUED, sizeof(cl_ulong), &cmd_queued, NULL)); + CL_CHECK(clGetEventProfilingInfo( + info.evt, CL_PROFILING_COMMAND_SUBMIT, sizeof(cl_ulong), &cmd_submit, NULL)); + CL_CHECK(clGetEventProfilingInfo( + info.evt, CL_PROFILING_COMMAND_START, sizeof(cl_ulong), &cmd_start, NULL)); + CL_CHECK(clGetEventProfilingInfo( + info.evt, CL_PROFILING_COMMAND_END, sizeof(cl_ulong), &cmd_end, NULL)); + CL_CHECK(clGetEventProfilingInfo( + info.evt, CL_PROFILING_COMMAND_COMPLETE, sizeof(cl_ulong), &cmd_complete, NULL)); + CL_CHECK(clReleaseEvent(info.evt)); + + char kernel_name[512]; + CL_CHECK(clGetKernelInfo(info.kernel, CL_KERNEL_FUNCTION_NAME, + sizeof(kernel_name), kernel_name, NULL)); + info.kernel_name = kernel_name; + + info.cmd_queued = cmd_queued; + info.cmd_submit = cmd_submit; + info.cmd_start = cmd_start; + info.cmd_end = cmd_end; + + info.cmd_queued_duration_ns = cmd_submit - cmd_queued; + info.cmd_submit_duration_ns = cmd_start - cmd_submit; + info.cmd_duration_ns = cmd_end - cmd_start; + info.cmd_complete_duration_ns = cmd_complete - cmd_end; + info.cmd_total_duration_ns = cmd_complete - cmd_queued; + } + + // Dump a csv float total_kernel_time = 0; - fprintf(fperf, "op name, kernel name, duration (ms), global size, local size, output size\n"); + fprintf(fperf, "op name, kernel name, queued duration (ms), submit duration(ms), exec duration (ms), complete duration (ms), total duration (ms), global size, local size, output size\n"); for (const ProfilingInfo & info : g_profiling_info) { - total_kernel_time += info.duration_ns/1.e6f; - fprintf(fperf, "%s,%s,%f,%zux%zux%zu,%zux%zux%zu,%zux%zux%zux%zu\n", - info.op_name.c_str(), info.kernel_name.c_str(), info.duration_ns/1.e6f, + total_kernel_time += info.cmd_duration_ns/1.e6f; + fprintf(fperf, "%s,%s,%f,%f,%f,%f,%f,%zux%zux%zu,%zux%zux%zu,%zux%zux%zux%zu\n", + info.op_name.c_str(), info.kernel_name.c_str(), + info.cmd_queued_duration_ns/1.e6f, + info.cmd_submit_duration_ns/1.e6f, + info.cmd_duration_ns/1.e6f, + info.cmd_complete_duration_ns/1.e6f, + info.cmd_total_duration_ns/1.e6f, info.global_size[0], info.global_size[1], info.global_size[2], info.local_size[0], info.local_size[2], info.local_size[2], info.output_size[0], info.output_size[1], info.output_size[2], info.output_size[3]); @@ -831,6 +1030,27 @@ static void ggml_cl2_free(void) { fclose(fperf); GGML_LOG_INFO("ggml_opencl: total kernel time: %f\n", total_kernel_time); + + // Dump a simple chrome trace + FILE* ftrace = fopen("cl_trace.json", "w"); + if (!ftrace) { + GGML_LOG_ERROR("Failed to open cl_trace.json\n"); + return; + } + + fprintf(ftrace, "[\n"); + for (const ProfilingInfo & info : g_profiling_info) { + fprintf(ftrace, "{\"name\": \"%s\", \"cat\": \"OpenCL\", \"ph\": \"B\", \"ts\": %lu, \"pid\": \"\", \"tid\": \"Host\"},\n", + info.kernel_name.c_str(), info.cmd_queued/1000); + fprintf(ftrace, "{\"name\": \"%s\", \"cat\": \"OpenCL\", \"ph\": \"E\", \"ts\": %lu, \"pid\": \"\", \"tid\": \"Host\"},\n", + info.kernel_name.c_str(), info.cmd_submit/1000); + + fprintf(ftrace, "{\"name\": \"%s\", \"cat\": \"OpenCL\", \"ph\": \"B\", \"ts\": %lu, \"pid\": \"\", \"tid\": \"Device\"},\n", + info.kernel_name.c_str(), info.cmd_start/1000); + fprintf(ftrace, "{\"name\": \"%s\", \"cat\": \"OpenCL\", \"ph\": \"E\", \"ts\": %lu, \"pid\": \"\", \"tid\": \"Device\"},\n", + info.kernel_name.c_str(), info.cmd_end/1000); + } + fclose(ftrace); #endif } @@ -1012,17 +1232,19 @@ static bool ggml_opencl_supports_op(ggml_backend_dev_t dev, const struct ggml_te case GGML_OP_ADD: case GGML_OP_SCALE: case GGML_OP_MUL: - return true; + return op->src[0]->type == GGML_TYPE_F32; case GGML_OP_UNARY: switch (ggml_get_unary_op(op)) { case GGML_UNARY_OP_GELU: case GGML_UNARY_OP_SILU: case GGML_UNARY_OP_RELU: - return ggml_is_contiguous(op->src[0]); + case GGML_UNARY_OP_GELU_QUICK: + return ggml_is_contiguous(op->src[0]) && op->src[0]->type == GGML_TYPE_F32; default: return false; } case GGML_OP_CLAMP: + return op->src[0]->type == GGML_TYPE_F32; case GGML_OP_SOFT_MAX: case GGML_OP_NORM: case GGML_OP_RMS_NORM: @@ -1044,7 +1266,27 @@ static bool ggml_opencl_supports_op(ggml_backend_dev_t dev, const struct ggml_te return true; case GGML_OP_DIAG_MASK_INF: return op->ne[3] == 1; - case GGML_OP_ROPE: + case GGML_OP_ROPE: { + const int mode = ((const int32_t *) op->op_params)[2]; + const bool is_mrope = mode & GGML_ROPE_TYPE_MROPE; + const bool is_vision = mode == GGML_ROPE_TYPE_VISION; + if (is_mrope && !is_vision) { + if (op->src[0]->type == GGML_TYPE_F32 || + op->src[0]->type == GGML_TYPE_F16) { + return true; + } + return false; + } + if (is_vision) { + if (op->src[0]->type == GGML_TYPE_F32 || + op->src[0]->type == GGML_TYPE_F16) { + return true; + } + return false; + } + return true; + } + case GGML_OP_IM2COL: return true; default: return false; @@ -1198,20 +1440,17 @@ struct ggml_backend_opencl_buffer_context { std::string name; }; -static void * const cl_ptr_base = (void *)(uintptr_t) 0x1000; - static void ggml_backend_opencl_buffer_free_buffer(ggml_backend_buffer_t buffer) { ggml_backend_opencl_buffer_context * ctx = (ggml_backend_opencl_buffer_context *) buffer->context; delete ctx; } static void * ggml_backend_opencl_buffer_get_base(ggml_backend_buffer_t buffer) { - return cl_ptr_base; - - GGML_UNUSED(buffer); + ggml_backend_opencl_context * backend_ctx = ggml_cl2_init(buffer->buft->device); + return (void *) (uintptr_t) backend_ctx->alignment; } -static void ggml_backend_opencl_buffer_init_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor) { +static enum ggml_status ggml_backend_opencl_buffer_init_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor) { ggml_backend_opencl_buffer_context * ctx = (ggml_backend_opencl_buffer_context *) buffer->context; ggml_cl2_init(buffer->buft->device); @@ -1241,7 +1480,7 @@ static void ggml_backend_opencl_buffer_init_tensor(ggml_backend_buffer_t buffer, tensor->extra = view_extra; } else { { - size_t offset = (char *)tensor->data - (char *)cl_ptr_base; + size_t offset = (char *) tensor->data - (char *) ggml_backend_opencl_buffer_get_base(buffer); ggml_tensor_extra_cl * extra = ctx->ggml_opencl_alloc_temp_tensor_extra(); extra->offset = offset; @@ -1251,6 +1490,7 @@ static void ggml_backend_opencl_buffer_init_tensor(ggml_backend_buffer_t buffer, tensor->extra = extra; } } + return GGML_STATUS_SUCCESS; } // The optimized gemm and gemv kernels are used for large matrices without batch. @@ -1365,6 +1605,11 @@ static void ggml_backend_opencl_buffer_set_tensor(ggml_backend_buffer_t buffer, int M = tensor->ne[1]; // ne01 int K = tensor->ne[0]; // ne00 + //For matrix-vector multiplication kernel, we assume K is a multiple of 32 + GGML_ASSERT(K % 32 == 0); + //For transpose kernels, we assume K is a multiple of 4 (satisfied by prior assert), and M is a multiple of 4 + GGML_ASSERT(M % 4 == 0); + // transpose is out of place, so we need to allocate transposed buffers // <----------------------------------------------------------------------------------> // // use sub_buffer of max buffer size instead @@ -1405,36 +1650,36 @@ static void ggml_backend_opencl_buffer_set_tensor(ggml_backend_buffer_t buffer, cl_mem qT_d_image1D; cl_mem dT_d_image1D; - cl_image_format img_fmt_1d = { CL_RGBA, CL_FLOAT }; + cl_image_format img_fmt_1d = { CL_RGBA, CL_HALF_FLOAT }; cl_image_desc img_desc_1d; memset(&img_desc_1d, 0, sizeof(img_desc_1d)); img_desc_1d.image_type = CL_MEM_OBJECT_IMAGE1D_BUFFER; - img_desc_1d.image_width = M * K / 8 / 4; + img_desc_1d.image_width = M * K / 4 / 4; img_desc_1d.buffer = extra->q; q_d_image1D = clCreateImage(context, 0, &img_fmt_1d, &img_desc_1d, NULL, &err); CL_CHECK(err); - img_fmt_1d = { CL_RGBA, CL_FLOAT }; + img_fmt_1d = { CL_RGBA, CL_HALF_FLOAT }; memset(&img_desc_1d, 0, sizeof(img_desc_1d)); img_desc_1d.image_type = CL_MEM_OBJECT_IMAGE1D_BUFFER; - img_desc_1d.image_width = M * K / 8 / 4; + img_desc_1d.image_width = M * K / 4 / 4; img_desc_1d.buffer = qT_d; qT_d_image1D = clCreateImage(context, 0, &img_fmt_1d, &img_desc_1d, NULL, &err); CL_CHECK(err); - img_fmt_1d = { CL_RGBA, CL_FLOAT }; + img_fmt_1d = { CL_RGBA, CL_HALF_FLOAT }; memset(&img_desc_1d, 0, sizeof(img_desc_1d)); img_desc_1d.image_type = CL_MEM_OBJECT_IMAGE1D_BUFFER; - img_desc_1d.image_width = M * K / 32 / 4 / 2; + img_desc_1d.image_width = M * K / 32 / 4; img_desc_1d.buffer = extra->d; d_d_image1D = clCreateImage(context, 0, &img_fmt_1d, &img_desc_1d, NULL, &err); CL_CHECK(err); - img_fmt_1d = { CL_RGBA, CL_FLOAT }; + img_fmt_1d = { CL_RGBA, CL_HALF_FLOAT }; memset(&img_desc_1d, 0, sizeof(img_desc_1d)); img_desc_1d.image_type = CL_MEM_OBJECT_IMAGE1D_BUFFER; - img_desc_1d.image_width = M * K / 32 / 4 / 2; + img_desc_1d.image_width = M * K / 32 / 4; img_desc_1d.buffer = dT_d; dT_d_image1D = clCreateImage(context, 0, &img_fmt_1d, &img_desc_1d, NULL, &err); CL_CHECK(err); @@ -1443,8 +1688,8 @@ static void ggml_backend_opencl_buffer_set_tensor(ggml_backend_buffer_t buffer, // set up and call the transpose kernels // <----------------------------------------------------------------------------------> // // weights - int height_q = M / 8; - int width_q = K / 8 / 4; + int height_q = M / 4; + int width_q = K / 4 / 4; kernel = backend_ctx->kernel_transpose_16; CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &q_d_image1D)); @@ -1458,8 +1703,8 @@ static void ggml_backend_opencl_buffer_set_tensor(ggml_backend_buffer_t buffer, CL_CHECK(clWaitForEvents(1, &evt)); // scales - int height_s = M / 8; - int width_s = K / 32 / 8; + int height_s = M / 4; + int width_s = K / 32 / 4; kernel = backend_ctx->kernel_transpose_16; CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &d_d_image1D)); @@ -1853,7 +2098,6 @@ static void dump_tensor(ggml_backend_t backend, const struct ggml_tensor * tenso void * buf_d; #endif -#ifdef GGML_USE_OPENCL // Make sure everything is done. CL_CHECK(clFinish(queue)); @@ -1889,7 +2133,6 @@ static void dump_tensor(ggml_backend_t backend, const struct ggml_tensor * tenso extra->offset, ggml_nbytes(tensor), buf, 0, NULL, NULL)); CL_CHECK(clFinish(queue)); #endif // GGML_OPENCL_SOA_Q -#endif // GGML_USE_OPENCL // Open file and dump. char fname[512]; @@ -1967,25 +2210,14 @@ static void dump_tensor(ggml_backend_t backend, const struct ggml_tensor * tenso // Profiling utility //------------------------------------------------------------------------------ #ifdef GGML_OPENCL_PROFILING -void populateProfilingInfo( +static void populateProfilingInfo( ProfilingInfo& info, cl_event evt, cl_kernel kernel, size_t global_size[3], size_t local_size[3], const ggml_tensor * tensor) { - cl_ulong start; - cl_ulong end; - CL_CHECK(clWaitForEvents(1, &evt)); - CL_CHECK(clGetEventProfilingInfo( - evt, CL_PROFILING_COMMAND_START, sizeof(cl_ulong), &start, NULL)); - CL_CHECK(clGetEventProfilingInfo( - evt, CL_PROFILING_COMMAND_END, sizeof(cl_ulong), &end, NULL)); + info.op_name = tensor->name; + info.kernel = kernel; + info.evt = evt; - char kernel_name[512]; - CL_CHECK(clGetKernelInfo(kernel, CL_KERNEL_FUNCTION_NAME, - sizeof(kernel_name), kernel_name, NULL)); - - info.duration_ns = end - start; - info.op_name = tensor->name; - info.kernel_name = kernel_name; info.local_size[0] = local_size[0]; info.local_size[1] = local_size[1]; info.local_size[2] = local_size[2]; @@ -2414,6 +2646,53 @@ static void ggml_cl_gelu(ggml_backend_t backend, const ggml_tensor * src0, const #endif } +static void ggml_cl_gelu_quick(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { + GGML_ASSERT(src0); + GGML_ASSERT(src0->extra); + GGML_ASSERT(dst); + GGML_ASSERT(dst->extra); + + UNUSED(src1); + + ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context; + cl_command_queue queue = backend_ctx->queue; + + ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra; + ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra; + + cl_ulong offset0 = extra0->offset + src0->view_offs; + cl_ulong offsetd = extrad->offset + dst->view_offs; + + cl_kernel kernel; + + int n = ggml_nelements(dst); + + if (n % 4 == 0) { + kernel = backend_ctx->kernel_gelu_quick_4; + n /= 4; + } else { + kernel = backend_ctx->kernel_gelu_quick; + } + + CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &extra0->data_device)); + CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong), &offset0)); + CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem), &extrad->data_device)); + CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong), &offsetd)); + + size_t global_work_size[] = {(size_t)n, 1, 1}; + size_t local_work_size[] = {64, 1, 1}; + +#ifdef GGML_OPENCL_PROFILING + cl_event evt; + clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, &evt); + + g_profiling_info.emplace_back(); + populateProfilingInfo(g_profiling_info.back(), evt, kernel, global_work_size, local_work_size, dst); +#else + clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, NULL); +#endif +} + static void ggml_cl_silu(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { GGML_ASSERT(src0); GGML_ASSERT(src0->extra); @@ -2569,26 +2848,33 @@ static void ggml_cl_norm(ggml_backend_t backend, const ggml_tensor * src0, const memcpy(&eps, dst->op_params, sizeof(float)); const int ne00 = src0 ? src0->ne[0] : 0; - const cl_ulong nb01 = src0 ? src0->nb[1] : 0; + const int ne01 = src0 ? src0->ne[1] : 0; + const int ne02 = src0 ? src0->ne[2] : 0; + const int ne03 = src0 ? src0->ne[3] : 0; - GGML_ASSERT(ggml_is_contiguous_1(src0)); + const cl_ulong nb01 = src0 ? src0->nb[1] : 0; + const cl_ulong nb02 = src0 ? src0->nb[2] : 0; + const cl_ulong nb03 = src0 ? src0->nb[3] : 0; const int nth = MIN(64, ne00); cl_kernel kernel = backend_ctx->kernel_norm; - CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &extra0->data_device)); - CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong), &offset0)); - CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem), &extrad->data_device)); - CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong), &offsetd)); - CL_CHECK(clSetKernelArg(kernel, 4, sizeof(int), &ne00)); - CL_CHECK(clSetKernelArg(kernel, 5, sizeof(cl_ulong), &nb01)); - CL_CHECK(clSetKernelArg(kernel, 6, sizeof(float), &eps)); - CL_CHECK(clSetKernelArg(kernel, 7, sizeof(float)*nth, NULL)); + CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &extra0->data_device)); + CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong), &offset0)); + CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem), &extrad->data_device)); + CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong), &offsetd)); + CL_CHECK(clSetKernelArg(kernel, 4, sizeof(int), &ne00)); + CL_CHECK(clSetKernelArg(kernel, 5, sizeof(int), &ne01)); + CL_CHECK(clSetKernelArg(kernel, 6, sizeof(int), &ne02)); + CL_CHECK(clSetKernelArg(kernel, 7, sizeof(int), &ne03)); + CL_CHECK(clSetKernelArg(kernel, 8, sizeof(cl_ulong), &nb01)); + CL_CHECK(clSetKernelArg(kernel, 9, sizeof(cl_ulong), &nb02)); + CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_ulong), &nb03)); + CL_CHECK(clSetKernelArg(kernel, 11, sizeof(float), &eps)); + CL_CHECK(clSetKernelArg(kernel, 12, sizeof(float)*nth, NULL)); - const int64_t nrows = ggml_nrows(src0); - - size_t global_work_size[] = {(size_t)nrows*nth, 1, 1}; + size_t global_work_size[] = {(size_t)ne01*nth, (size_t)ne02, (size_t)ne03}; size_t local_work_size[] = {(size_t)nth, 1, 1}; #ifdef GGML_OPENCL_PROFILING @@ -2626,16 +2912,19 @@ static void ggml_cl_rms_norm(ggml_backend_t backend, const ggml_tensor * src0, c memcpy(&eps, dst->op_params, sizeof(float)); const int ne00 = src0 ? src0->ne[0] : 0; + const int ne01 = src0 ? src0->ne[1] : 0; + const int ne02 = src0 ? src0->ne[2] : 0; + const int ne03 = src0 ? src0->ne[3] : 0; + const cl_ulong nb01 = src0 ? src0->nb[1] : 0; + const cl_ulong nb02 = src0 ? src0->nb[2] : 0; + const cl_ulong nb03 = src0 ? src0->nb[3] : 0; GGML_ASSERT(ne00 % 4 == 0); - GGML_ASSERT(ggml_is_contiguous_1(src0)); const int nth = MIN(64, ne00); - const int64_t nrows = ggml_nrows(src0); - - size_t global_work_size[] = {(size_t)nrows*nth, 1, 1}; + size_t global_work_size[] = {(size_t)ne01*nth, (size_t)ne02, (size_t)ne03}; size_t local_work_size[] = {(size_t)nth, 1, 1}; cl_kernel kernel = backend_ctx->kernel_rms_norm; @@ -2650,15 +2939,20 @@ static void ggml_cl_rms_norm(ggml_backend_t backend, const ggml_tensor * src0, c sizeof(local_work_size), local_work_size, sizeof(size_t), &sgs, NULL)); - CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &extra0->data_device)); - CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong), &offset0)); - CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem), &extrad->data_device)); - CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong), &offsetd)); - CL_CHECK(clSetKernelArg(kernel, 4, sizeof(int), &ne00)); - CL_CHECK(clSetKernelArg(kernel, 5, sizeof(cl_ulong), &nb01)); - CL_CHECK(clSetKernelArg(kernel, 6, sizeof(float), &eps)); + CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &extra0->data_device)); + CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong), &offset0)); + CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem), &extrad->data_device)); + CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong), &offsetd)); + CL_CHECK(clSetKernelArg(kernel, 4, sizeof(int), &ne00)); + CL_CHECK(clSetKernelArg(kernel, 5, sizeof(int), &ne01)); + CL_CHECK(clSetKernelArg(kernel, 6, sizeof(int), &ne02)); + CL_CHECK(clSetKernelArg(kernel, 7, sizeof(int), &ne03)); + CL_CHECK(clSetKernelArg(kernel, 8, sizeof(cl_ulong), &nb01)); + CL_CHECK(clSetKernelArg(kernel, 9, sizeof(cl_ulong), &nb02)); + CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_ulong), &nb03)); + CL_CHECK(clSetKernelArg(kernel, 11, sizeof(float), &eps)); // This is local memory - the size depends on subgroup size. - CL_CHECK(clSetKernelArg(kernel, 7, sizeof(float)*nth/sgs, NULL)); + CL_CHECK(clSetKernelArg(kernel, 12, sizeof(float)*nth/sgs, NULL)); #ifdef GGML_OPENCL_PROFILING cl_event evt; @@ -2854,6 +3148,9 @@ static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, co CL_CHECK(status); int height_B = N/4; + if (height_B == 0) { + height_B = 1; + } int width_B = K/4; int padded_height_B = (N + padding)/4; @@ -3002,11 +3299,12 @@ static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, co } if (N == 1) { - local_work_size[0] = backend_ctx->adreno_wave_size; // localsize + size_t wavesize = backend_ctx->adreno_wave_size; + local_work_size[0] = wavesize; // localsize local_work_size[1] = 4; // reduce factor local_work_size[2] = 1; - global_work_size[0] = M / 2; + global_work_size[0] = (((M / 2) + wavesize - 1) / wavesize) * wavesize; global_work_size[1] = 4; // reduce factor global_work_size[2] = 1; } @@ -3015,6 +3313,7 @@ static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, co // enqueue kernel with profiling // <--------------------------------------------> // #ifdef GGML_OPENCL_PROFILING + cl_event evt; CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, &evt)); g_profiling_info.emplace_back(); @@ -3666,6 +3965,8 @@ static void ggml_cl_soft_max(ggml_backend_t backend, const ggml_tensor * src0, c const float m0 = powf(2.0f, -(max_bias ) / n_head_log2); const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2); + const bool use_f16 = (src1 && src1->type == GGML_TYPE_F16); + // Local size must be wave size. Each workgroup is a wave, working on a row, // where a row corresponds to leading dimension. int nth = MIN(32, ne00); @@ -3683,9 +3984,17 @@ static void ggml_cl_soft_max(ggml_backend_t backend, const ggml_tensor * src0, c cl_kernel kernel; if (ne00%4 == 0) { - kernel = backend_ctx->kernel_soft_max_4; + if (use_f16) { + kernel = backend_ctx->kernel_soft_max_4_f16; + } else { + kernel = backend_ctx->kernel_soft_max_4; + } } else { - kernel = backend_ctx->kernel_soft_max; + if (use_f16) { + kernel = backend_ctx->kernel_soft_max_f16; + } else { + kernel = backend_ctx->kernel_soft_max; + } } CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &extra0->data_device)); @@ -3746,10 +4055,10 @@ static void ggml_cl_rope(ggml_backend_t backend, const ggml_tensor * src0, const const int ne02 = src0 ? src0->ne[2] : 0; const int ne03 = src0 ? src0->ne[3] : 0; - const int nb00 = src0 ? src0->nb[0] : 0; - const int nb01 = src0 ? src0->nb[1] : 0; - const int nb02 = src0 ? src0->nb[2] : 0; - const int nb03 = src0 ? src0->nb[3] : 0; + const cl_ulong nb00 = src0 ? src0->nb[0] : 0; + const cl_ulong nb01 = src0 ? src0->nb[1] : 0; + const cl_ulong nb02 = src0 ? src0->nb[2] : 0; + const cl_ulong nb03 = src0 ? src0->nb[3] : 0; const int ne10 = src1 ? src1->ne[0] : 0; const int ne11 = src1 ? src1->ne[1] : 0; UNUSED(ne11); @@ -3761,12 +4070,13 @@ static void ggml_cl_rope(ggml_backend_t backend, const ggml_tensor * src0, const const int ne2 = dst ? dst->ne[2] : 0; const int ne3 = dst ? dst->ne[3] : 0; - const int nb0 = dst ? dst->nb[0] : 0; - const int nb1 = dst ? dst->nb[1] : 0; - const int nb2 = dst ? dst->nb[2] : 0; - const int nb3 = dst ? dst->nb[3] : 0; + const cl_ulong nb0 = dst ? dst->nb[0] : 0; + const cl_ulong nb1 = dst ? dst->nb[1] : 0; + const cl_ulong nb2 = dst ? dst->nb[2] : 0; + const cl_ulong nb3 = dst ? dst->nb[3] : 0; - GGML_ASSERT(ne10 == ne02); + GGML_ASSERT(ne10 % ne02 == 0); + GGML_ASSERT(ne10 >= ne02); int nth = MIN(64, ne00); @@ -3781,6 +4091,7 @@ static void ggml_cl_rope(ggml_backend_t backend, const ggml_tensor * src0, const float attn_factor; float beta_fast; float beta_slow; + int32_t sections[4]; memcpy(&freq_base, (int32_t *) dst->op_params + 5, sizeof(float)); memcpy(&freq_scale, (int32_t *) dst->op_params + 6, sizeof(float)); @@ -3788,23 +4099,23 @@ static void ggml_cl_rope(ggml_backend_t backend, const ggml_tensor * src0, const memcpy(&attn_factor, (int32_t *) dst->op_params + 8, sizeof(float)); memcpy(&beta_fast, (int32_t *) dst->op_params + 9, sizeof(float)); memcpy(&beta_slow, (int32_t *) dst->op_params + 10, sizeof(float)); + memcpy(§ions, (int32_t *) dst->op_params + 11, sizeof(int32_t)*4); const bool is_neox = mode & 2; + const bool is_mrope = mode & GGML_ROPE_TYPE_MROPE; + const bool is_vision = mode == GGML_ROPE_TYPE_VISION; + + if (is_mrope) { + GGML_ASSERT(sections[0] > 0 || sections[1] > 0 || sections[2] > 0); + } + + if (is_vision) { + GGML_ASSERT(n_dims == ne00/2); + } cl_kernel kernel; - if (!is_neox) { - switch (src0->type) { - case GGML_TYPE_F32: - kernel = backend_ctx->kernel_rope_norm_f32; - break; - case GGML_TYPE_F16: - kernel = backend_ctx->kernel_rope_norm_f16; - break; - default: - GGML_ASSERT(false); - }; - } else { + if (is_neox) { switch (src0->type) { case GGML_TYPE_F32: kernel = backend_ctx->kernel_rope_neox_f32; @@ -3815,6 +4126,39 @@ static void ggml_cl_rope(ggml_backend_t backend, const ggml_tensor * src0, const default: GGML_ASSERT(false); }; + } else if (is_mrope && !is_vision) { + switch (src0->type) { + case GGML_TYPE_F32: + kernel = backend_ctx->kernel_rope_multi_f32; + break; + case GGML_TYPE_F16: + kernel = backend_ctx->kernel_rope_multi_f16; + break; + default: + GGML_ASSERT(false); + }; + } else if (is_vision) { + switch (src0->type) { + case GGML_TYPE_F32: + kernel = backend_ctx->kernel_rope_vision_f32; + break; + case GGML_TYPE_F16: + kernel = backend_ctx->kernel_rope_vision_f16; + break; + default: + GGML_ASSERT(false); + } + } else { + switch (src0->type) { + case GGML_TYPE_F32: + kernel = backend_ctx->kernel_rope_norm_f32; + break; + case GGML_TYPE_F16: + kernel = backend_ctx->kernel_rope_norm_f16; + break; + default: + GGML_ASSERT(false); + }; } CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &extra0->data_device)); @@ -3850,6 +4194,9 @@ static void ggml_cl_rope(ggml_backend_t backend, const ggml_tensor * src0, const CL_CHECK(clSetKernelArg(kernel, 30, sizeof(float), &attn_factor)); CL_CHECK(clSetKernelArg(kernel, 31, sizeof(float), &beta_fast)); CL_CHECK(clSetKernelArg(kernel, 32, sizeof(float), &beta_slow)); + if (is_mrope || is_vision) { + CL_CHECK(clSetKernelArg(kernel, 33, sizeof(int32_t)*4, §ions)); + } size_t global_work_size[] = {(size_t)ne01*nth, (size_t)ne02, (size_t)ne03}; size_t local_work_size[] = {(size_t)nth, 1, 1}; @@ -3865,6 +4212,98 @@ static void ggml_cl_rope(ggml_backend_t backend, const ggml_tensor * src0, const #endif } +static void ggml_cl_im2col(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { + GGML_ASSERT(src0); + GGML_ASSERT(src1); + GGML_ASSERT(src1->extra); + GGML_ASSERT(dst); + GGML_ASSERT(dst->extra); + + // src0 - filter, src1 - input + GGML_ASSERT(src1->type == GGML_TYPE_F32); + GGML_ASSERT(dst->type == GGML_TYPE_F16 || dst->type == GGML_TYPE_F32); + + ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context; + cl_command_queue queue = backend_ctx->queue; + + ggml_tensor_extra_cl * extra1 = (ggml_tensor_extra_cl *)src1->extra; + ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra; + + cl_ulong offset1 = extra1->offset + src1->view_offs; + cl_ulong offsetd = extrad->offset + dst->view_offs; + + const int32_t s0 = ((const int32_t*)(dst->op_params))[0]; + const int32_t s1 = ((const int32_t*)(dst->op_params))[1]; + const int32_t p0 = ((const int32_t*)(dst->op_params))[2]; + const int32_t p1 = ((const int32_t*)(dst->op_params))[3]; + const int32_t d0 = ((const int32_t*)(dst->op_params))[4]; + const int32_t d1 = ((const int32_t*)(dst->op_params))[5]; + + const bool is_2D = ((const int32_t*)(dst->op_params))[6] == 1; + + const cl_long IC = src1->ne[is_2D ? 2 : 1]; + const cl_long IH = is_2D ? src1->ne[1] : 1; + const cl_long IW = src1->ne[0]; + + const cl_long KH = is_2D ? src0->ne[1] : 1; + const cl_long KW = src0->ne[0]; + + const cl_long OH = is_2D ? dst->ne[2] : 1; + const cl_long OW = dst->ne[1]; + + // nb is byte offset, src is type float32 + const cl_ulong delta_offset = src1->nb[is_2D ? 2 : 1]/4; + const cl_long batch = src1->ne[is_2D ? 3 : 2]; + const cl_ulong batch_offset = src1->nb[is_2D ? 3 : 2]/4; + + const cl_long pelements = OW*KW*KH; + const cl_long CHW = IC*KH*KW; + + cl_kernel kernel; + + if(dst->type == GGML_TYPE_F16) { + kernel = backend_ctx->kernel_im2col_f16; + } else { + kernel = backend_ctx->kernel_im2col_f32; + } + + CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &extra1->data_device)); + CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong), &offset1)); + CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem), &extrad->data_device)); + CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong), &offsetd)); + CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_ulong), &batch_offset)); + CL_CHECK(clSetKernelArg(kernel, 5, sizeof(cl_ulong), &delta_offset)); + CL_CHECK(clSetKernelArg(kernel, 6, sizeof(cl_long), &IW)); + CL_CHECK(clSetKernelArg(kernel, 7, sizeof(cl_long), &IH)); + CL_CHECK(clSetKernelArg(kernel, 8, sizeof(cl_long), &IC)); + CL_CHECK(clSetKernelArg(kernel, 9, sizeof(cl_long), &OW)); + CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_long), &OH)); + CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_long), &KW)); + CL_CHECK(clSetKernelArg(kernel, 12, sizeof(cl_long), &KH)); + CL_CHECK(clSetKernelArg(kernel, 13, sizeof(cl_long), &pelements)); + CL_CHECK(clSetKernelArg(kernel, 14, sizeof(cl_long), &CHW)); + CL_CHECK(clSetKernelArg(kernel, 15, sizeof(int), &s0)); + CL_CHECK(clSetKernelArg(kernel, 16, sizeof(int), &s1)); + CL_CHECK(clSetKernelArg(kernel, 17, sizeof(int), &p0)); + CL_CHECK(clSetKernelArg(kernel, 18, sizeof(int), &p1)); + CL_CHECK(clSetKernelArg(kernel, 19, sizeof(int), &d0)); + CL_CHECK(clSetKernelArg(kernel, 20, sizeof(int), &d1)); + + const int num_blocks = (pelements + 256 - 1) / 256; + size_t global_work_size[] = {(size_t)num_blocks*256, (size_t)OH, (size_t)batch*IC}; + size_t local_work_size[] = {256, 1, 1}; + +#ifdef GGML_OPENCL_PROFILING + cl_event evt; + CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, &evt)); + + g_profiling_info.emplace_back(); + populateProfilingInfo(g_profiling_info.back(), evt, kernel, global_work_size, local_work_size, dst); +#else + CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, NULL)); +#endif +} + //------------------------------------------------------------------------------ // Op offloading //------------------------------------------------------------------------------ @@ -3923,6 +4362,12 @@ bool ggml_cl_compute_forward(ggml_backend_t backend, struct ggml_tensor * tensor } func = ggml_cl_gelu; break; + case GGML_UNARY_OP_GELU_QUICK: + if (!any_on_device) { + return false; + } + func = ggml_cl_gelu_quick; + break; case GGML_UNARY_OP_SILU: if (!any_on_device) { return false; @@ -3995,6 +4440,12 @@ bool ggml_cl_compute_forward(ggml_backend_t backend, struct ggml_tensor * tensor } func = ggml_cl_rope; break; + case GGML_OP_IM2COL: + if (!any_on_device) { + return false; + } + func = ggml_cl_im2col; + break; default: return false; } diff --git a/ggml/src/ggml-opencl/kernels/ggml-opencl.cl b/ggml/src/ggml-opencl/kernels/ggml-opencl.cl index d1cdf709b..b88792887 100644 --- a/ggml/src/ggml-opencl/kernels/ggml-opencl.cl +++ b/ggml/src/ggml-opencl/kernels/ggml-opencl.cl @@ -404,6 +404,7 @@ kernel void kernel_scale( // gelu //------------------------------------------------------------------------------ #define GELU_COEF_A 0.044715f +#define GELU_QUICK_COEF -1.702f #define SQRT_2_OVER_PI 0.79788456080286535587989211986876f kernel void kernel_gelu( @@ -434,6 +435,32 @@ kernel void kernel_gelu_4( dst[get_global_id(0)] = 0.5f*x*(1.0f + tanh(SQRT_2_OVER_PI*x*(1.0f + GELU_COEF_A*x*x))); } +kernel void kernel_gelu_quick( + global float * src0, + ulong offset0, + global float * dst, + ulong offsetd +) { + src0 = (global float*)((global char*)src0 + offset0); + dst = (global float*)((global char*)dst + offsetd); + + float x = src0[get_global_id(0)]; + dst[get_global_id(0)] = x*(1.0f/(1.0f+exp(GELU_QUICK_COEF*x))); +} + +kernel void kernel_gelu_quick_4( + global float4 * src0, + ulong offset0, + global float4 * dst, + ulong offsetd +) { + src0 = (global float4*)((global char*)src0 + offset0); + dst = (global float4*)((global char*)dst + offsetd); + + float4 x = src0[get_global_id(0)]; + dst[get_global_id(0)] = x*(1.0f/(1.0f+exp(GELU_QUICK_COEF*x))); +} + //------------------------------------------------------------------------------ // silu //------------------------------------------------------------------------------ @@ -506,14 +533,23 @@ kernel void kernel_norm( global float * dst, ulong offsetd, int ne00, + int ne01, + int ne02, + int ne03, ulong nb01, + ulong nb02, + ulong nb03, float eps, local float * sum ) { src0 = (global void*)((global char*)src0 + offset0); dst = (global void*)((global char*)dst + offsetd); - global float * x = (global float *) ((global char *) src0 + get_group_id(0)*nb01); + int i03 = get_group_id(2); + int i02 = get_group_id(1); + int i01 = get_group_id(0); + + global float * x = (global float *) ((global char *) src0 + i03*nb03 + i02*nb02 + i01*nb01); // MEAN // parallel sum @@ -533,7 +569,7 @@ kernel void kernel_norm( // recenter and VARIANCE barrier(CLK_LOCAL_MEM_FENCE); - global float * y = dst + get_group_id(0)*ne00; + global float * y = dst + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00; sum[get_local_id(0)] = 0.0f; for (int i00 = get_local_id(0); i00 < ne00; i00 += get_local_size(0)) { y[i00] = x[i00] - mean; @@ -566,14 +602,23 @@ kernel void kernel_rms_norm( global float * dst, ulong offsetd, int ne00, + int ne01, + int ne02, + int ne03, ulong nb01, + ulong nb02, + ulong nb03, float eps, local float * sum // Note, the size depends on number of subgroups ) { src0 = (global void*)((global char*)src0 + offset0); dst = (global float*)((global char*)dst + offsetd); - global float4 * x = (global float4 *) ((global char *) src0 + get_group_id(0)*nb01); + int i03 = get_group_id(2); + int i02 = get_group_id(1); + int i01 = get_group_id(0); + + global float4 * x = (global float4 *) ((global char *) src0 + i03*nb03 + i02*nb02 + i01*nb01); global float * x_scalar = (global float *) x; float4 sumf = 0; float all_sum = 0; @@ -607,7 +652,7 @@ kernel void kernel_rms_norm( const float mean = sum[0]; const float scale = 1.0f/sqrt(mean + eps); - global float4 * y = (global float4 *) (dst + get_group_id(0)*ne00); + global float4 * y = (global float4 *) (dst + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00); global float * y_scalar = (global float *) y; for (int i00 = get_local_id(0); i00 < ne00/4; i00 += get_local_size(0)) { y[i00] = x[i00] * scale; @@ -679,6 +724,9 @@ kernel void kernel_diag_mask_inf_8( //------------------------------------------------------------------------------ // softmax //------------------------------------------------------------------------------ +#ifdef ADRENO_GPU +REQD_SUBGROUP_SIZE_64 +#endif kernel void kernel_soft_max( global float * src0, ulong offset0, @@ -811,6 +859,141 @@ kernel void kernel_soft_max_4( } } +#ifdef ADRENO_GPU +REQD_SUBGROUP_SIZE_64 +#endif +kernel void kernel_soft_max_f16( + global float * src0, + ulong offset0, + global half * src1, + ulong offset1, + global float * dst, + ulong offsetd, + int ne00, + int ne01, + int ne02, + float scale, + float max_bias, + float m0, + float m1, + int n_head_log2 +) { + src0 = (global float *)((global char *)src0 + offset0); + src1 = (global half *)((global char *)src1 + offset1); + dst = (global float *)((global char *)dst + offsetd); + + int i03 = get_group_id(2); + int i02 = get_group_id(1); + int i01 = get_group_id(0); + + global float * psrc0 = src0 + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00; + global half * pmask = (global char *)src1 != (global char *)src0 ? src1 + i01*ne00 : 0; + global float * pdst = dst + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00; + + float slope = 1.0f; + + // ALiBi + if (max_bias > 0.0f) { + int h = i02; + + float base = h < n_head_log2 ? m0 : m1; + int exp = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1; + + slope = pow(base, exp); + } + + // parallel max + float lmax = -INFINITY; + for (int i00 = get_local_id(0); i00 < ne00; i00 += get_local_size(0)) { + lmax = fmax(lmax, psrc0[i00]*scale + (pmask ? slope*pmask[i00] : 0.0f)); + } + float max = sub_group_reduce_max(lmax); + + // parallel sum + float lsum = 0.0f; + for (int i00 = get_local_id(0); i00 < ne00; i00 += get_local_size(0)) { + float exp_psrc0 = exp((psrc0[i00]*scale + (pmask ? slope*pmask[i00] : 0.0f)) - max); + lsum += exp_psrc0; + // Remember the result of exp here. exp is expensive, so we really do not + // wish to compute it twice. + pdst[i00] = exp_psrc0; + } + + const float sum = sub_group_reduce_add(lsum); + + for (int i00 = get_local_id(0); i00 < ne00; i00 += get_local_size(0)) { + pdst[i00] /= sum; + } +} + +#ifdef ADRENO_GPU +REQD_SUBGROUP_SIZE_64 +#endif +kernel void kernel_soft_max_4_f16( + global float * src0, + ulong offset0, + global half * src1, + ulong offset1, + global float * dst, + ulong offsetd, + int ne00, + int ne01, + int ne02, + float scale, + float max_bias, + float m0, + float m1, + int n_head_log2 +) { + src0 = (global float *)((global char *)src0 + offset0); + src1 = (global half *)((global char *)src1 + offset1); + dst = (global float *)((global char *)dst + offsetd); + + int i03 = get_group_id(2); + int i02 = get_group_id(1); + int i01 = get_group_id(0); + + global float4 * psrc4 = (global float4 *)(src0 + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00); + global half4 * pmask = (global char *)src1 != (global char *)src0 ? (global half4 *)(src1 + i01*ne00) : 0; + global float4 * pdst4 = (global float4 *)(dst + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00); + + float slope = 1.0f; + + // ALiBi + if (max_bias > 0.0f) { + int h = i02; + + float base = h < n_head_log2 ? m0 : m1; + int exp = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1; + + slope = pow(base, exp); + } + + // parallel max + float4 lmax4 = -INFINITY; + for (int i00 = get_local_id(0); i00 < ne00/4; i00 += get_local_size(0)) { + lmax4 = fmax(lmax4, psrc4[i00]*scale + slope*(pmask ? convert_float4(pmask[i00]) : 0.0f)); + } + float lmax = fmax(fmax(lmax4.s0, lmax4.s1), fmax(lmax4.s2, lmax4.s3)); + + const float max = sub_group_reduce_max(lmax); + + // parallel sum + float4 lsum4 = 0.0f; + for (int i00 = get_local_id(0); i00 < ne00/4; i00 += get_local_size(0)) { + const float4 exp_psrc4 = exp((psrc4[i00]*scale + slope*(pmask ? convert_float4(pmask[i00]) : 0.0f)) - max); + lsum4 += exp_psrc4; + pdst4[i00] = exp_psrc4; + } + float lsum = lsum4.s0 + lsum4.s1 + lsum4.s2 + lsum4.s3; + + const float sum = sub_group_reduce_add(lsum); + + for (int i00 = get_local_id(0); i00 < ne00/4; i00 += get_local_size(0)) { + pdst4[i00] /= sum; + } +} + //------------------------------------------------------------------------------ // kernel_rope //------------------------------------------------------------------------------ @@ -1169,6 +1352,368 @@ kernel void kernel_rope_neox_f16( } } +kernel void kernel_rope_multi_f32( + global void * src0, + ulong offset0, + global int * src1, + ulong offset1, + global float * src2, + ulong offset2, + global float * dst, + ulong offsetd, + int ne00, + int ne01, + int ne02, + int ne03, + ulong nb00, + ulong nb01, + ulong nb02, + ulong nb03, + int ne0, + int ne1, + int ne2, + int ne3, + ulong nb0, + ulong nb1, + ulong nb2, + ulong nb3, + int n_past, + int n_dims, + int n_ctx_orig, + float freq_base, + float freq_scale, + float ext_factor, + float attn_factor, + float beta_fast, + float beta_slow, + int4 sections +) { + src0 = (global void*)((global char*)src0 + offset0); + src1 = (global int*)((global char*)src1 + offset1); + src2 = (global float*)((global char*)src2 + offset2); + dst = (global float*)((global char*)dst + offsetd); + + int i3 = get_group_id(2); + int i2 = get_group_id(1); + int i1 = get_group_id(0); + + float2 corr_dims = rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow); + + global int * pos = src1; + + const int sect_dims = sections.s0 + sections.s1 + sections.s2 + sections.s3; + const int sec_w = sections.s1 + sections.s0; + + float inv_ndims = -1.f/n_dims; + + for (int i0 = 2*get_local_id(0); i0 < ne0; i0 += 2*get_local_size(0)) { + if (i0 < n_dims) { + int ic = i0/2; + + const int sector = (i0 / 2) % sect_dims; + float theta_base = 0.0f; + + if (sector < sections.s0) { + theta_base = pos[i2]; + } + else if (sector >= sections.s0 && sector < sec_w) { + theta_base = pos[i2 + ne2 * 1]; + } + else if (sector >= sec_w && sector < sec_w + sections.s2) { + theta_base = pos[i2 + ne2 * 2]; + } + else if (sector >= sec_w + sections.s2) { + theta_base = pos[i2 + ne2 * 3]; + } + + const float theta = theta_base * pow(freq_base, inv_ndims*i0); + + const float freq_factor = src2 != src0 ? src2[ic] : 1.0f; + + float2 cos_sin_theta = rope_yarn(theta/freq_factor, freq_scale, corr_dims, i0, ext_factor, attn_factor); + + global float * src = (global float *)((global char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + ic*nb00); + global float * dst_data = (global float *)((global char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + ic*nb0); + + const float x0 = src[0]; + const float x1 = src[n_dims/2]; + + dst_data[0] = x0*cos_sin_theta.s0 - x1*cos_sin_theta.s1; + dst_data[n_dims/2] = x0*cos_sin_theta.s1 + x1*cos_sin_theta.s0; + } else { + global float * const src = (global float *)((global char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00); + global float * dst_data = (global float *)((global char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0); + + dst_data[0] = src[0]; + dst_data[1] = src[1]; + } + } +} + +kernel void kernel_rope_multi_f16( + global void * src0, + ulong offset0, + global int * src1, + ulong offset1, + global float * src2, + ulong offset2, + global half * dst, + ulong offsetd, + int ne00, + int ne01, + int ne02, + int ne03, + ulong nb00, + ulong nb01, + ulong nb02, + ulong nb03, + int ne0, + int ne1, + int ne2, + int ne3, + ulong nb0, + ulong nb1, + ulong nb2, + ulong nb3, + int n_past, + int n_dims, + int n_ctx_orig, + float freq_base, + float freq_scale, + float ext_factor, + float attn_factor, + float beta_fast, + float beta_slow, + int4 sections +) { + src0 = (global void*)((global char*)src0 + offset0); + src1 = (global int*)((global char*)src1 + offset1); + src2 = (global float*)((global char*)src2 + offset2); + dst = (global float*)((global char*)dst + offsetd); + + int i3 = get_group_id(2); + int i2 = get_group_id(1); + int i1 = get_group_id(0); + + float2 corr_dims = rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow); + + global int * pos = src1; + + const int sect_dims = sections.s0 + sections.s1 + sections.s2 + sections.s3; + const int sec_w = sections.s1 + sections.s0; + + float inv_ndims = -1.f/n_dims; + + for (int i0 = 2*get_local_id(0); i0 < ne0; i0 += 2*get_local_size(0)) { + if (i0 < n_dims) { + int ic = i0/2; + + const int sector = (i0 / 2) % sect_dims; + float theta_base = 0.0f; + + if (sector < sections.s0) { + theta_base = pos[i2]; + } + else if (sector >= sections.s0 && sector < sec_w) { + theta_base = pos[i2 + ne2 * 1]; + } + else if (sector >= sec_w && sector < sec_w + sections.s2) { + theta_base = pos[i2 + ne2 * 2]; + } + else if (sector >= sec_w + sections.s2) { + theta_base = pos[i2 + ne2 * 3]; + } + + const float theta = theta_base * pow(freq_base, inv_ndims*i0); + + const float freq_factor = src2 != src0 ? src2[ic] : 1.0f; + + float2 cos_sin_theta = rope_yarn(theta/freq_factor, freq_scale, corr_dims, i0, ext_factor, attn_factor); + + global half * src = (global half *)((global char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + ic*nb00); + global half * dst_data = (global half *)((global char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + ic*nb0); + + const float x0 = src[0]; + const float x1 = src[n_dims/2]; + + dst_data[0] = x0*cos_sin_theta.s0 - x1*cos_sin_theta.s1; + dst_data[n_dims/2] = x0*cos_sin_theta.s1 + x1*cos_sin_theta.s0; + } else { + global half * const src = (global half *)((global char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00); + global half * dst_data = (global half *)((global char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0); + + dst_data[0] = src[0]; + dst_data[1] = src[1]; + } + } +} + +kernel void kernel_rope_vision_f32( + global void * src0, + ulong offset0, + global int * src1, + ulong offset1, + global float * src2, + ulong offset2, + global float * dst, + ulong offsetd, + int ne00, + int ne01, + int ne02, + int ne03, + ulong nb00, + ulong nb01, + ulong nb02, + ulong nb03, + int ne0, + int ne1, + int ne2, + int ne3, + ulong nb0, + ulong nb1, + ulong nb2, + ulong nb3, + int n_past, + int n_dims, + int n_ctx_orig, + float freq_base, + float freq_scale, + float ext_factor, + float attn_factor, + float beta_fast, + float beta_slow, + int4 sections +) { + src0 = (global void*)((global char*)src0 + offset0); + src1 = (global int*)((global char*)src1 + offset1); + src2 = (global float*)((global char*)src2 + offset2); + dst = (global float*)((global char*)dst + offsetd); + + int i3 = get_group_id(2); + int i2 = get_group_id(1); + int i1 = get_group_id(0); + + float2 corr_dims = rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow); + + global int * pos = src1; + + const int sect_dims = sections.s0 + sections.s1; + const int sec_w = sections.s1 + sections.s0; + + float inv_ndims = -1.f/n_dims; + + for (int i0 = 2*get_local_id(0); i0 < ne0; i0 += 2*get_local_size(0)) { + int ic = i0/2; + + const int sector = (i0/2) % sect_dims; + float theta_base = 0.0f; + + if (sector < sections.s0) { + const int p = sector; + theta_base = pos[i2] * pow(freq_base, inv_ndims*2.0f*p); + } else if (sector >= sections.s0 && sector < sec_w) { + const int p = sector - sections.s0; + theta_base = pos[i2 + ne2] * pow(freq_base, inv_ndims*2.0f*p); + } + + const float freq_factor = src2 != src0 ? src2[ic] : 1.0f; + + float2 cos_sin_theta = rope_yarn(theta_base/freq_factor, freq_scale, corr_dims, i0, ext_factor, attn_factor); + + global float * src = (global float *)((global char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + ic*nb00); + global float * dst_data = (global float *)((global char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + ic*nb0); + + const float x0 = src[0]; + const float x1 = src[n_dims]; + + dst_data[0] = x0*cos_sin_theta.s0 - x1*cos_sin_theta.s1; + dst_data[n_dims] = x0*cos_sin_theta.s1 + x1*cos_sin_theta.s0; + } +} + +kernel void kernel_rope_vision_f16( + global void * src0, + ulong offset0, + global int * src1, + ulong offset1, + global float * src2, + ulong offset2, + global half * dst, + ulong offsetd, + int ne00, + int ne01, + int ne02, + int ne03, + ulong nb00, + ulong nb01, + ulong nb02, + ulong nb03, + int ne0, + int ne1, + int ne2, + int ne3, + ulong nb0, + ulong nb1, + ulong nb2, + ulong nb3, + int n_past, + int n_dims, + int n_ctx_orig, + float freq_base, + float freq_scale, + float ext_factor, + float attn_factor, + float beta_fast, + float beta_slow, + int4 sections +) { + src0 = (global void*)((global char*)src0 + offset0); + src1 = (global int*)((global char*)src1 + offset1); + src2 = (global float*)((global char*)src2 + offset2); + dst = (global float*)((global char*)dst + offsetd); + + int i3 = get_group_id(2); + int i2 = get_group_id(1); + int i1 = get_group_id(0); + + float2 corr_dims = rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow); + + global int * pos = src1; + + const int sect_dims = sections.s0 + sections.s1; + const int sec_w = sections.s1 + sections.s0; + + float inv_ndims = -1.f/n_dims; + + for (int i0 = 2*get_local_id(0); i0 < ne0; i0 += 2*get_local_size(0)) { + int ic = i0/2; + + const int sector = (i0/2) % sect_dims; + float theta_base = 0.0f; + + if (sector < sections.s0) { + const int p = sector; + theta_base = pos[i2] * pow(freq_base, inv_ndims*2.0f*p); + } else if (sector >= sections.s0 && sector < sec_w) { + const int p = sector - sections.s0; + theta_base = pos[i2 + ne2] * pow(freq_base, inv_ndims*2.0f*p); + } + + const float freq_factor = src2 != src0 ? src2[ic] : 1.0f; + + float2 cos_sin_theta = rope_yarn(theta_base/freq_factor, freq_scale, corr_dims, i0, ext_factor, attn_factor); + + global half * src = (global half *)((global char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + ic*nb00); + global half * dst_data = (global half *)((global char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + ic*nb0); + + const float x0 = src[0]; + const float x1 = src[n_dims]; + + dst_data[0] = x0*cos_sin_theta.s0 - x1*cos_sin_theta.s1; + dst_data[n_dims] = x0*cos_sin_theta.s1 + x1*cos_sin_theta.s0; + } +} + //------------------------------------------------------------------------------ // cpy //------------------------------------------------------------------------------ @@ -1659,6 +2204,9 @@ kernel void kernel_mul_mat_f16_f16( //------------------------------------------------------------------------------ // mul_mat_f16_f32_1row //------------------------------------------------------------------------------ +#ifdef ADRENO_GPU +REQD_SUBGROUP_SIZE_64 +#endif kernel void kernel_mul_mat_f16_f32_1row( global char * src0, ulong offset0, diff --git a/ggml/src/ggml-opencl/kernels/ggml-opencl_gemv_noshuffle.cl b/ggml/src/ggml-opencl/kernels/ggml-opencl_gemv_noshuffle.cl index 5e195411d..ee5c79f00 100644 --- a/ggml/src/ggml-opencl/kernels/ggml-opencl_gemv_noshuffle.cl +++ b/ggml/src/ggml-opencl/kernels/ggml-opencl_gemv_noshuffle.cl @@ -1,9 +1,11 @@ #pragma OPENCL EXTENSION cl_khr_fp16 : enable #pragma OPENCL EXTENSION cl_khr_subgroups : enable -#pragma OPENCL EXTENSION cl_qcom_subgroup_uniform_load: enable -#pragma OPENCL EXTENSION cl_qcom_subgroup_constant_load: enable -#pragma OPENCL EXTENSION cl_qcom_extra_vector_types : enable + +#ifdef cl_qcom_reqd_sub_group_size #pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable +#define ADRENO_GPU 1 +#define REQD_SUBGROUP_SIZE_64 __attribute__((qcom_reqd_sub_group_size("half"))) +#endif // assume #define QK4_0 32 @@ -186,8 +188,9 @@ total_sums.s1 += (((bits4.s7 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y.s6; \ total_sums.s1 += (((bits4.s7 & 0xF000) >> 12) - 8) * scale.s1 * shared_y.s7; \ - -__attribute__((qcom_reqd_sub_group_size("full"))) +#ifdef ADRENO_GPU +REQD_SUBGROUP_SIZE_64 +#endif __kernel void kernel_gemv_noshuffle( __read_only image1d_buffer_t src0_q, // quantized A global half2 * src0_d, // A scales diff --git a/ggml/src/ggml-opencl/kernels/ggml-opencl_gemv_noshuffle_general.cl b/ggml/src/ggml-opencl/kernels/ggml-opencl_gemv_noshuffle_general.cl index 5bdd4d067..469d3edef 100644 --- a/ggml/src/ggml-opencl/kernels/ggml-opencl_gemv_noshuffle_general.cl +++ b/ggml/src/ggml-opencl/kernels/ggml-opencl_gemv_noshuffle_general.cl @@ -1,9 +1,11 @@ #pragma OPENCL EXTENSION cl_khr_fp16 : enable #pragma OPENCL EXTENSION cl_khr_subgroups : enable -#pragma OPENCL EXTENSION cl_qcom_subgroup_uniform_load: enable -#pragma OPENCL EXTENSION cl_qcom_subgroup_constant_load: enable -#pragma OPENCL EXTENSION cl_qcom_extra_vector_types : enable + +#ifdef cl_qcom_reqd_sub_group_size #pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable +#define ADRENO_GPU 1 +#define REQD_SUBGROUP_SIZE_64 __attribute__((qcom_reqd_sub_group_size("half"))) +#endif // assume #define QK4_0 32 @@ -186,8 +188,9 @@ total_sums.s1 += (((bits4.s7 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y.s6; \ total_sums.s1 += (((bits4.s7 & 0xF000) >> 12) - 8) * scale.s1 * shared_y.s7; \ - -__attribute__((qcom_reqd_sub_group_size("full"))) +#ifdef ADRENO_GPU +REQD_SUBGROUP_SIZE_64 +#endif __kernel void kernel_gemv_noshuffle( __read_only image1d_buffer_t src0_q, // quantized A global half2 * src0_d, // A scales diff --git a/ggml/src/ggml-opencl/kernels/ggml-opencl_im2col.cl b/ggml/src/ggml-opencl/kernels/ggml-opencl_im2col.cl new file mode 100644 index 000000000..9b41dfb25 --- /dev/null +++ b/ggml/src/ggml-opencl/kernels/ggml-opencl_im2col.cl @@ -0,0 +1,146 @@ +#ifdef cl_khr_fp16 +#pragma OPENCL EXTENSION cl_khr_fp16 : enable +#elif defined(cl_amd_fp16) +#pragma OPENCL EXTENSION cl_amd_fp16 : enable +#else +#error "Half precision floating point not supportedby OpenCL implementation on your device." +#endif + +#ifdef cl_khr_subgroups +#pragma OPENCL EXTENSION cl_khr_subgroups : enable +#elif defined(cl_intel_subgroups) +#pragma OPENCL EXTENSION cl_intel_subgroups : enable +#else +#error "Subgroup not supported on your device." +#endif + +#ifdef cl_intel_required_subgroup_size +// Always use subgroup size of 32 on Intel. +#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable +#define INTEL_GPU 1 +#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16))) +#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32))) +#elif defined(cl_qcom_reqd_sub_group_size) +// Always use subgroups size of 64 on Adreno. +#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable +#define ADRENO_GPU 1 +#define REQD_SUBGROUP_SIZE_64 __attribute__((qcom_reqd_sub_group_size("half"))) +#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full"))) +#else +// TODO: do not know how to choose subgroup size on other GPUs. +#error "Selecting subgroup size is not supported on your device." +#endif + +kernel void kernel_im2col_f32( + global float * src1, + ulong offset1, + global float * dst, + ulong offsetd, + ulong batch_offset, + ulong delta_offset, + long IW, + long IH, + long IC, + long OW, + long OH, + long KW, + long KH, + long pelements, + long CHW, + int s0, + int s1, + int p0, + int p1, + int d0, + int d1 +) { + // threadIdx.x + blockIdx.x * blockDim.x + long i = get_global_id(0); + if (i >= pelements) { + return; + } + + src1 = (global float*)((global char*)src1 + offset1); + dst = (global float*)((global char*)dst + offsetd); + + long ksize = OW * (KH > 1 ? KW : 1); + long kx = i / ksize; + long kd = kx * ksize; + long ky = (i - kd) / OW; + long ix = i % OW; + + long oh = get_group_id(1); + long batch = get_group_id(2) / IC; + long ic = get_group_id(2) % IC; + + long iiw = ix * s0 + kx * d0 - p0; + long iih = oh * s1 + ky * d1 - p1; + + long offset_dst = + ((batch * OH + oh) * OW + ix) * CHW + + (ic * (KW * KH) + ky * KW + kx); + + if (iih < 0 || iih >= IH || iiw < 0 || iiw >= IW) { + dst[offset_dst] = 0.0f; + } else { + long offset_src = ic * delta_offset + batch * batch_offset; + dst[offset_dst] = src1[offset_src + iih * IW + iiw]; + } +} + +kernel void kernel_im2col_f16( + global float * src1, + ulong offset1, + global half * dst, + ulong offsetd, + ulong batch_offset, + ulong delta_offset, + long IW, + long IH, + long IC, + long OW, + long OH, + long KW, + long KH, + long pelements, + long CHW, + int s0, + int s1, + int p0, + int p1, + int d0, + int d1 +) { + long i = get_global_id(0); + + if (i >= pelements) { + return; + } + + src1 = (global float*)((global char*)src1 + offset1); + dst = (global half*)((global char*)dst + offsetd); + + long ksize = OW * (KH > 1 ? KW : 1); + long kx = i / ksize; + long kd = kx * ksize; + long ky = (i - kd) / OW; + long ix = i % OW; + + long oh = get_group_id(1); + long batch = get_group_id(2) / IC; + long ic = get_group_id(2) % IC; + + long iiw = ix * s0 + kx * d0 - p0; + long iih = oh * s1 + ky * d1 - p1; + + long offset_dst = + ((batch * OH + oh) * OW + ix) * CHW + + (ic * (KW * KH) + ky * KW + kx); + + if (iih < 0 || iih >= IH || iiw < 0 || iiw >= IW) { + dst[offset_dst] = 0.0f; + } else { + long offset_src = ic * delta_offset + batch * batch_offset; + dst[offset_dst] = src1[offset_src + iih * IW + iiw]; + } +} diff --git a/ggml/src/ggml-opencl/kernels/ggml-opencl_mul_mat_Ab_Bi_8x4.cl b/ggml/src/ggml-opencl/kernels/ggml-opencl_mul_mat_Ab_Bi_8x4.cl index 57768c803..ecb577b99 100644 --- a/ggml/src/ggml-opencl/kernels/ggml-opencl_mul_mat_Ab_Bi_8x4.cl +++ b/ggml/src/ggml-opencl/kernels/ggml-opencl_mul_mat_Ab_Bi_8x4.cl @@ -7,7 +7,16 @@ #pragma OPENCL EXTENSION cl_khr_fp16 : enable #pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable -__attribute__((qcom_reqd_sub_group_size("full"))) +#ifdef cl_qcom_reqd_sub_group_size +#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable +#define ADRENO_GPU 1 +#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full"))) +#endif + +#ifdef ADRENO_GPU +REQD_SUBGROUP_SIZE_128 +#endif + kernel void kernel_mul_mat_Ab_Bi_8x4( global const ushort * src0_q, // quantized A global const half * src0_d, // A scales diff --git a/ggml/src/ggml-opencl/kernels/ggml-opencl_transpose_16.cl b/ggml/src/ggml-opencl/kernels/ggml-opencl_transpose_16.cl index d59a0c05d..cd4e0afba 100644 --- a/ggml/src/ggml-opencl/kernels/ggml-opencl_transpose_16.cl +++ b/ggml/src/ggml-opencl/kernels/ggml-opencl_transpose_16.cl @@ -1,4 +1,6 @@ -// 16-bit transpose, loading/storing an 8x8 tile of elements +// 16-bit transpose, loading/storing a 4x4 tile of elements + +#pragma OPENCL EXTENSION cl_khr_fp16 : enable kernel void kernel_transpose_16( __read_only image1d_buffer_t input, @@ -9,24 +11,16 @@ kernel void kernel_transpose_16( const int i = get_global_id(0); const int j = get_global_id(1); - const int i_3 = i<<3; - const int j_3 = j<<3; + const int i_2 = i<<2; + const int j_2 = j<<2; - ushort8 temp0 = as_ushort8(read_imagef(input, (j_3+0)*cols+i)); - ushort8 temp1 = as_ushort8(read_imagef(input, (j_3+1)*cols+i)); - ushort8 temp2 = as_ushort8(read_imagef(input, (j_3+2)*cols+i)); - ushort8 temp3 = as_ushort8(read_imagef(input, (j_3+3)*cols+i)); - ushort8 temp4 = as_ushort8(read_imagef(input, (j_3+4)*cols+i)); - ushort8 temp5 = as_ushort8(read_imagef(input, (j_3+5)*cols+i)); - ushort8 temp6 = as_ushort8(read_imagef(input, (j_3+6)*cols+i)); - ushort8 temp7 = as_ushort8(read_imagef(input, (j_3+7)*cols+i)); + half4 temp0 = read_imageh(input, (j_2+0)*cols+i); + half4 temp1 = read_imageh(input, (j_2+1)*cols+i); + half4 temp2 = read_imageh(input, (j_2+2)*cols+i); + half4 temp3 = read_imageh(input, (j_2+3)*cols+i); - write_imagef(output, (i_3+0)*rows+j, as_float4((ushort8)(temp0.s0, temp1.s0, temp2.s0, temp3.s0, temp4.s0, temp5.s0, temp6.s0, temp7.s0))); - write_imagef(output, (i_3+1)*rows+j, as_float4((ushort8)(temp0.s1, temp1.s1, temp2.s1, temp3.s1, temp4.s1, temp5.s1, temp6.s1, temp7.s1))); - write_imagef(output, (i_3+2)*rows+j, as_float4((ushort8)(temp0.s2, temp1.s2, temp2.s2, temp3.s2, temp4.s2, temp5.s2, temp6.s2, temp7.s2))); - write_imagef(output, (i_3+3)*rows+j, as_float4((ushort8)(temp0.s3, temp1.s3, temp2.s3, temp3.s3, temp4.s3, temp5.s3, temp6.s3, temp7.s3))); - write_imagef(output, (i_3+4)*rows+j, as_float4((ushort8)(temp0.s4, temp1.s4, temp2.s4, temp3.s4, temp4.s4, temp5.s4, temp6.s4, temp7.s4))); - write_imagef(output, (i_3+5)*rows+j, as_float4((ushort8)(temp0.s5, temp1.s5, temp2.s5, temp3.s5, temp4.s5, temp5.s5, temp6.s5, temp7.s5))); - write_imagef(output, (i_3+6)*rows+j, as_float4((ushort8)(temp0.s6, temp1.s6, temp2.s6, temp3.s6, temp4.s6, temp5.s6, temp6.s6, temp7.s6))); - write_imagef(output, (i_3+7)*rows+j, as_float4((ushort8)(temp0.s7, temp1.s7, temp2.s7, temp3.s7, temp4.s7, temp5.s7, temp6.s7, temp7.s7))); + write_imageh(output, (i_2+0)*rows+j, (half4)(temp0.s0, temp1.s0, temp2.s0, temp3.s0)); + write_imageh(output, (i_2+1)*rows+j, (half4)(temp0.s1, temp1.s1, temp2.s1, temp3.s1)); + write_imageh(output, (i_2+2)*rows+j, (half4)(temp0.s2, temp1.s2, temp2.s2, temp3.s2)); + write_imageh(output, (i_2+3)*rows+j, (half4)(temp0.s3, temp1.s3, temp2.s3, temp3.s3)); } diff --git a/ggml/src/ggml-quants.c b/ggml/src/ggml-quants.c index 7918388ae..ac918a60d 100644 --- a/ggml/src/ggml-quants.c +++ b/ggml/src/ggml-quants.c @@ -28,7 +28,7 @@ #define UNUSED GGML_UNUSED // reference implementation for deterministic creation of model files -void quantize_row_q4_0_ref(const float * restrict x, block_q4_0 * restrict y, int64_t k) { +void quantize_row_q4_0_ref(const float * GGML_RESTRICT x, block_q4_0 * GGML_RESTRICT y, int64_t k) { static const int qk = QK4_0; assert(k % qk == 0); @@ -65,7 +65,7 @@ void quantize_row_q4_0_ref(const float * restrict x, block_q4_0 * restrict y, in } } -void quantize_row_q4_1_ref(const float * restrict x, block_q4_1 * restrict y, int64_t k) { +void quantize_row_q4_1_ref(const float * GGML_RESTRICT x, block_q4_1 * GGML_RESTRICT y, int64_t k) { const int qk = QK4_1; assert(k % qk == 0); @@ -102,7 +102,7 @@ void quantize_row_q4_1_ref(const float * restrict x, block_q4_1 * restrict y, in } } -void quantize_row_q5_0_ref(const float * restrict x, block_q5_0 * restrict y, int64_t k) { +void quantize_row_q5_0_ref(const float * GGML_RESTRICT x, block_q5_0 * GGML_RESTRICT y, int64_t k) { static const int qk = QK5_0; assert(k % qk == 0); @@ -146,7 +146,7 @@ void quantize_row_q5_0_ref(const float * restrict x, block_q5_0 * restrict y, in } } -void quantize_row_q5_1_ref(const float * restrict x, block_q5_1 * restrict y, int64_t k) { +void quantize_row_q5_1_ref(const float * GGML_RESTRICT x, block_q5_1 * GGML_RESTRICT y, int64_t k) { const int qk = QK5_1; assert(k % qk == 0); @@ -191,7 +191,7 @@ void quantize_row_q5_1_ref(const float * restrict x, block_q5_1 * restrict y, in } // reference implementation for deterministic creation of model files -void quantize_row_q8_0_ref(const float * restrict x, block_q8_0 * restrict y, int64_t k) { +void quantize_row_q8_0_ref(const float * GGML_RESTRICT x, block_q8_0 * GGML_RESTRICT y, int64_t k) { assert(k % QK8_0 == 0); const int nb = k / QK8_0; @@ -217,7 +217,7 @@ void quantize_row_q8_0_ref(const float * restrict x, block_q8_0 * restrict y, in } // reference implementation for deterministic creation of model files -void quantize_row_q8_1_ref(const float * restrict x, block_q8_1 * restrict y, int64_t k) { +void quantize_row_q8_1_ref(const float * GGML_RESTRICT x, block_q8_1 * GGML_RESTRICT y, int64_t k) { assert(QK8_1 == 32); assert(k % QK8_1 == 0); const int nb = k / QK8_1; @@ -252,7 +252,7 @@ void quantize_row_q8_1_ref(const float * restrict x, block_q8_1 * restrict y, in } } -void dequantize_row_q4_0(const block_q4_0 * restrict x, float * restrict y, int64_t k) { +void dequantize_row_q4_0(const block_q4_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k) { static const int qk = QK4_0; assert(k % qk == 0); @@ -272,7 +272,7 @@ void dequantize_row_q4_0(const block_q4_0 * restrict x, float * restrict y, int6 } } -void dequantize_row_q4_1(const block_q4_1 * restrict x, float * restrict y, int64_t k) { +void dequantize_row_q4_1(const block_q4_1 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k) { static const int qk = QK4_1; assert(k % qk == 0); @@ -293,7 +293,7 @@ void dequantize_row_q4_1(const block_q4_1 * restrict x, float * restrict y, int6 } } -void dequantize_row_q5_0(const block_q5_0 * restrict x, float * restrict y, int64_t k) { +void dequantize_row_q5_0(const block_q5_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k) { static const int qk = QK5_0; assert(k % qk == 0); @@ -319,7 +319,7 @@ void dequantize_row_q5_0(const block_q5_0 * restrict x, float * restrict y, int6 } } -void dequantize_row_q5_1(const block_q5_1 * restrict x, float * restrict y, int64_t k) { +void dequantize_row_q5_1(const block_q5_1 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k) { static const int qk = QK5_1; assert(k % qk == 0); @@ -346,7 +346,7 @@ void dequantize_row_q5_1(const block_q5_1 * restrict x, float * restrict y, int6 } } -void dequantize_row_q8_0(const block_q8_0 * restrict x, float * restrict y, int64_t k) { +void dequantize_row_q8_0(const block_q8_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k) { static const int qk = QK8_0; assert(k % qk == 0); @@ -376,8 +376,8 @@ static inline int nearest_int(float fval) { return (i & 0x007fffff) - 0x00400000; } -static float make_qx_quants(int n, int nmax, const float * restrict x, int8_t * restrict L, int rmse_type, - const float * restrict qw) { +static float make_qx_quants(int n, int nmax, const float * GGML_RESTRICT x, int8_t * GGML_RESTRICT L, int rmse_type, + const float * GGML_RESTRICT qw) { float max = 0; float amax = 0; for (int i = 0; i < n; ++i) { @@ -445,7 +445,7 @@ static float make_qx_quants(int n, int nmax, const float * restrict x, int8_t * return scale; } -static float make_q3_quants(int n, int nmax, const float * restrict x, int8_t * restrict L, bool do_rmse) { +static float make_q3_quants(int n, int nmax, const float * GGML_RESTRICT x, int8_t * GGML_RESTRICT L, bool do_rmse) { float max = 0; float amax = 0; for (int i = 0; i < n; ++i) { @@ -504,7 +504,7 @@ static float make_q3_quants(int n, int nmax, const float * restrict x, int8_t * return 1/iscale; } -static float make_qkx1_quants(int n, int nmax, const float * restrict x, uint8_t * restrict L, float * restrict the_min, +static float make_qkx1_quants(int n, int nmax, const float * GGML_RESTRICT x, uint8_t * GGML_RESTRICT L, float * GGML_RESTRICT the_min, int ntry, float alpha) { float min = x[0]; float max = x[0]; @@ -547,8 +547,8 @@ static float make_qkx1_quants(int n, int nmax, const float * restrict x, uint8_t return scale; } -static float make_qkx2_quants(int n, int nmax, const float * restrict x, const float * restrict weights, - uint8_t * restrict L, float * restrict the_min, uint8_t * restrict Laux, +static float make_qkx2_quants(int n, int nmax, const float * GGML_RESTRICT x, const float * GGML_RESTRICT weights, + uint8_t * GGML_RESTRICT L, float * GGML_RESTRICT the_min, uint8_t * GGML_RESTRICT Laux, float rmin, float rdelta, int nstep, bool use_mad) { float min = x[0]; float max = x[0]; @@ -628,7 +628,7 @@ static float make_qkx2_quants(int n, int nmax, const float * restrict x, const f return scale; } -static inline void get_scale_min_k4(int j, const uint8_t * restrict q, uint8_t * restrict d, uint8_t * restrict m) { +static inline void get_scale_min_k4(int j, const uint8_t * GGML_RESTRICT q, uint8_t * GGML_RESTRICT d, uint8_t * GGML_RESTRICT m) { if (j < 4) { *d = q[j] & 63; *m = q[j + 4] & 63; } else { @@ -639,7 +639,7 @@ static inline void get_scale_min_k4(int j, const uint8_t * restrict q, uint8_t * //========================- 2-bit (de)-quantization -void quantize_row_q2_K_ref(const float * restrict x, block_q2_K * restrict y, int64_t k) { +void quantize_row_q2_K_ref(const float * GGML_RESTRICT x, block_q2_K * GGML_RESTRICT y, int64_t k) { assert(k % QK_K == 0); const int nb = k / QK_K; @@ -709,7 +709,7 @@ void quantize_row_q2_K_ref(const float * restrict x, block_q2_K * restrict y, in } } -void dequantize_row_q2_K(const block_q2_K * restrict x, float * restrict y, int64_t k) { +void dequantize_row_q2_K(const block_q2_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k) { assert(k % QK_K == 0); const int nb = k / QK_K; @@ -741,8 +741,8 @@ void dequantize_row_q2_K(const block_q2_K * restrict x, float * restrict y, int6 } } -static float make_qkx3_quants(int n, int nmax, const float * restrict x, const float * restrict weights, - uint8_t * restrict L, float * restrict the_min, uint8_t * restrict Laux, +static float make_qkx3_quants(int n, int nmax, const float * GGML_RESTRICT x, const float * GGML_RESTRICT weights, + uint8_t * GGML_RESTRICT L, float * GGML_RESTRICT the_min, uint8_t * GGML_RESTRICT Laux, float rmin, float rdelta, int nstep, bool use_mad) { float min = x[0]; float max = x[0]; @@ -824,7 +824,7 @@ static float make_qkx3_quants(int n, int nmax, const float * restrict x, const f return scale; } -static float make_qp_quants(int n, int nmax, const float * restrict x, uint8_t * restrict L, const float * quant_weights) { +static float make_qp_quants(int n, int nmax, const float * GGML_RESTRICT x, uint8_t * GGML_RESTRICT L, const float * quant_weights) { float max = 0; for (int i = 0; i < n; ++i) { max = MAX(max, x[i]); @@ -897,7 +897,7 @@ static float make_qp_quants(int n, int nmax, const float * restrict x, uint8_t * return sumlx/suml2; } -static void quantize_row_q2_K_impl(const float * restrict x, block_q2_K * restrict y, int k, const float * restrict quant_weights) { +static void quantize_row_q2_K_impl(const float * GGML_RESTRICT x, block_q2_K * GGML_RESTRICT y, int k, const float * GGML_RESTRICT quant_weights) { GGML_ASSERT(quant_weights); assert(k % QK_K == 0); const int nb = k / QK_K; @@ -917,7 +917,7 @@ static void quantize_row_q2_K_impl(const float * restrict x, block_q2_K * restri for (int j = 0; j < QK_K; ++j) sumx2 += x[j]*x[j]; float sigma2 = sumx2/QK_K; for (int j = 0; j < QK_K/16; ++j) { - const float * restrict qw = quant_weights + QK_K * i + 16*j; + const float * GGML_RESTRICT qw = quant_weights + QK_K * i + 16*j; for (int l = 0; l < 16; ++l) weight[l] = qw[l] * sqrtf(sigma2 + x[16*j + l]*x[16*j + l]); for (int l = 0; l < QK_K/16; ++l) sw[j] += weight[l]; scales[j] = make_qkx3_quants(16, 3, x + 16*j, weight, L + 16*j, &mins[j], Laux, -0.9f, 0.05f, 36, false); @@ -959,7 +959,7 @@ static void quantize_row_q2_K_impl(const float * restrict x, block_q2_K * restri } } -size_t quantize_q2_K(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) { +size_t quantize_q2_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) { size_t row_size = ggml_row_size(GGML_TYPE_Q2_K, n_per_row); if (!quant_weights) { quantize_row_q2_K_ref(src, dst, (int64_t)nrow*n_per_row); @@ -977,7 +977,7 @@ size_t quantize_q2_K(const float * restrict src, void * restrict dst, int64_t nr //========================= 3-bit (de)-quantization -void quantize_row_q3_K_ref(const float * restrict x, block_q3_K * restrict y, int64_t k) { +void quantize_row_q3_K_ref(const float * GGML_RESTRICT x, block_q3_K * GGML_RESTRICT y, int64_t k) { assert(k % QK_K == 0); const int nb = k / QK_K; @@ -1053,7 +1053,7 @@ void quantize_row_q3_K_ref(const float * restrict x, block_q3_K * restrict y, in } } -void dequantize_row_q3_K(const block_q3_K * restrict x, float * restrict y, int64_t k) { +void dequantize_row_q3_K(const block_q3_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k) { assert(k % QK_K == 0); const int nb = k / QK_K; @@ -1067,8 +1067,8 @@ void dequantize_row_q3_K(const block_q3_K * restrict x, float * restrict y, int6 const float d_all = GGML_FP16_TO_FP32(x[i].d); - const uint8_t * restrict q = x[i].qs; - const uint8_t * restrict hm = x[i].hmask; + const uint8_t * GGML_RESTRICT q = x[i].qs; + const uint8_t * GGML_RESTRICT hm = x[i].hmask; uint8_t m = 1; memcpy(aux, x[i].scales, 12); @@ -1103,7 +1103,7 @@ void dequantize_row_q3_K(const block_q3_K * restrict x, float * restrict y, int6 } } -static void quantize_row_q3_K_impl(const float * restrict x, block_q3_K * restrict y, int64_t n_per_row, const float * restrict quant_weights) { +static void quantize_row_q3_K_impl(const float * GGML_RESTRICT x, block_q3_K * GGML_RESTRICT y, int64_t n_per_row, const float * GGML_RESTRICT quant_weights) { assert(n_per_row % QK_K == 0); const int nb = n_per_row / QK_K; @@ -1187,7 +1187,7 @@ static void quantize_row_q3_K_impl(const float * restrict x, block_q3_K * restri } } -size_t quantize_q3_K(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) { +size_t quantize_q3_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) { size_t row_size = ggml_row_size(GGML_TYPE_Q3_K, n_per_row); if (!quant_weights) { quantize_row_q3_K_ref(src, dst, (int64_t)nrow*n_per_row); @@ -1205,7 +1205,7 @@ size_t quantize_q3_K(const float * restrict src, void * restrict dst, int64_t nr // ====================== 4-bit (de)-quantization -void quantize_row_q4_K_ref(const float * restrict x, block_q4_K * restrict y, int64_t k) { +void quantize_row_q4_K_ref(const float * GGML_RESTRICT x, block_q4_K * GGML_RESTRICT y, int64_t k) { assert(k % QK_K == 0); const int nb = k / QK_K; @@ -1277,7 +1277,7 @@ void quantize_row_q4_K_ref(const float * restrict x, block_q4_K * restrict y, in } } -void dequantize_row_q4_K(const block_q4_K * restrict x, float * restrict y, int64_t k) { +void dequantize_row_q4_K(const block_q4_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k) { assert(k % QK_K == 0); const int nb = k / QK_K; @@ -1301,7 +1301,7 @@ void dequantize_row_q4_K(const block_q4_K * restrict x, float * restrict y, int6 } } -static void quantize_row_q4_K_impl(const float * restrict x, block_q4_K * restrict y, int64_t n_per_row, const float * quant_weights) { +static void quantize_row_q4_K_impl(const float * GGML_RESTRICT x, block_q4_K * GGML_RESTRICT y, int64_t n_per_row, const float * quant_weights) { assert(n_per_row % QK_K == 0); const int64_t nb = n_per_row / QK_K; @@ -1374,7 +1374,7 @@ static void quantize_row_q4_K_impl(const float * restrict x, block_q4_K * restri } } -size_t quantize_q4_K(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) { +size_t quantize_q4_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) { size_t row_size = ggml_row_size(GGML_TYPE_Q4_K, n_per_row); if (!quant_weights) { quantize_row_q4_K_ref(src, dst, (int64_t)nrow*n_per_row); @@ -1392,7 +1392,7 @@ size_t quantize_q4_K(const float * restrict src, void * restrict dst, int64_t nr // ====================== 5-bit (de)-quantization -void quantize_row_q5_K_ref(const float * restrict x, block_q5_K * restrict y, int64_t k) { +void quantize_row_q5_K_ref(const float * GGML_RESTRICT x, block_q5_K * GGML_RESTRICT y, int64_t k) { assert(k % QK_K == 0); const int64_t nb = k / QK_K; @@ -1454,8 +1454,8 @@ void quantize_row_q5_K_ref(const float * restrict x, block_q5_K * restrict y, in } } - uint8_t * restrict qh = y[i].qh; - uint8_t * restrict ql = y[i].qs; + uint8_t * GGML_RESTRICT qh = y[i].qh; + uint8_t * GGML_RESTRICT ql = y[i].qs; memset(qh, 0, QK_K/8); uint8_t m1 = 1, m2 = 2; @@ -1479,7 +1479,7 @@ void quantize_row_q5_K_ref(const float * restrict x, block_q5_K * restrict y, in } } -void dequantize_row_q5_K(const block_q5_K * restrict x, float * restrict y, int64_t k) { +void dequantize_row_q5_K(const block_q5_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k) { assert(k % QK_K == 0); const int64_t nb = k / QK_K; @@ -1506,7 +1506,7 @@ void dequantize_row_q5_K(const block_q5_K * restrict x, float * restrict y, int6 } } -static void quantize_row_q5_K_impl(const float * restrict x, block_q5_K * restrict y, int64_t n_per_row, const float * quant_weights) { +static void quantize_row_q5_K_impl(const float * GGML_RESTRICT x, block_q5_K * GGML_RESTRICT y, int64_t n_per_row, const float * quant_weights) { assert(n_per_row % QK_K == 0); const int64_t nb = n_per_row / QK_K; @@ -1573,8 +1573,8 @@ static void quantize_row_q5_K_impl(const float * restrict x, block_q5_K * restri } } - uint8_t * restrict qh = y[i].qh; - uint8_t * restrict ql = y[i].qs; + uint8_t * GGML_RESTRICT qh = y[i].qh; + uint8_t * GGML_RESTRICT ql = y[i].qs; memset(qh, 0, QK_K/8); uint8_t m1 = 1, m2 = 2; @@ -1599,7 +1599,7 @@ static void quantize_row_q5_K_impl(const float * restrict x, block_q5_K * restri } } -size_t quantize_q5_K(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) { +size_t quantize_q5_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) { size_t row_size = ggml_row_size(GGML_TYPE_Q5_K, n_per_row); if (!quant_weights) { quantize_row_q5_K_ref(src, dst, (int64_t)nrow*n_per_row); @@ -1617,7 +1617,7 @@ size_t quantize_q5_K(const float * restrict src, void * restrict dst, int64_t nr // ====================== 6-bit (de)-quantization -void quantize_row_q6_K_ref(const float * restrict x, block_q6_K * restrict y, int64_t k) { +void quantize_row_q6_K_ref(const float * GGML_RESTRICT x, block_q6_K * GGML_RESTRICT y, int64_t k) { assert(k % QK_K == 0); const int64_t nb = k / QK_K; @@ -1667,8 +1667,8 @@ void quantize_row_q6_K_ref(const float * restrict x, block_q6_K * restrict y, in } } - uint8_t * restrict ql = y[i].ql; - uint8_t * restrict qh = y[i].qh; + uint8_t * GGML_RESTRICT ql = y[i].ql; + uint8_t * GGML_RESTRICT qh = y[i].qh; for (int j = 0; j < QK_K; j += 128) { for (int l = 0; l < 32; ++l) { const uint8_t q1 = L[j + l + 0] & 0xF; @@ -1687,16 +1687,16 @@ void quantize_row_q6_K_ref(const float * restrict x, block_q6_K * restrict y, in } } -void dequantize_row_q6_K(const block_q6_K * restrict x, float * restrict y, int64_t k) { +void dequantize_row_q6_K(const block_q6_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k) { assert(k % QK_K == 0); const int64_t nb = k / QK_K; for (int i = 0; i < nb; i++) { const float d = GGML_FP16_TO_FP32(x[i].d); - const uint8_t * restrict ql = x[i].ql; - const uint8_t * restrict qh = x[i].qh; - const int8_t * restrict sc = x[i].scales; + const uint8_t * GGML_RESTRICT ql = x[i].ql; + const uint8_t * GGML_RESTRICT qh = x[i].qh; + const int8_t * GGML_RESTRICT sc = x[i].scales; for (int n = 0; n < QK_K; n += 128) { for (int l = 0; l < 32; ++l) { @@ -1718,7 +1718,7 @@ void dequantize_row_q6_K(const block_q6_K * restrict x, float * restrict y, int6 } } -static void quantize_row_q6_K_impl(const float * restrict x, block_q6_K * restrict y, int64_t n_per_row, const float * quant_weights) { +static void quantize_row_q6_K_impl(const float * GGML_RESTRICT x, block_q6_K * GGML_RESTRICT y, int64_t n_per_row, const float * quant_weights) { assert(n_per_row % QK_K == 0); const int64_t nb = n_per_row / QK_K; @@ -1781,8 +1781,8 @@ static void quantize_row_q6_K_impl(const float * restrict x, block_q6_K * restri } } - uint8_t * restrict ql = y[i].ql; - uint8_t * restrict qh = y[i].qh; + uint8_t * GGML_RESTRICT ql = y[i].ql; + uint8_t * GGML_RESTRICT qh = y[i].qh; for (int j = 0; j < QK_K; j += 128) { for (int l = 0; l < 32; ++l) { const uint8_t q1 = L[j + l + 0] & 0xF; @@ -1802,7 +1802,7 @@ static void quantize_row_q6_K_impl(const float * restrict x, block_q6_K * restri } } -size_t quantize_q6_K(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) { +size_t quantize_q6_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) { size_t row_size = ggml_row_size(GGML_TYPE_Q6_K, n_per_row); if (!quant_weights) { quantize_row_q6_K_ref(src, dst, (int64_t)nrow*n_per_row); @@ -1818,7 +1818,7 @@ size_t quantize_q6_K(const float * restrict src, void * restrict dst, int64_t nr return nrow * row_size; } -static void quantize_row_q4_0_impl(const float * restrict x, block_q4_0 * restrict y, int64_t n_per_row, const float * quant_weights) { +static void quantize_row_q4_0_impl(const float * GGML_RESTRICT x, block_q4_0 * GGML_RESTRICT y, int64_t n_per_row, const float * quant_weights) { static_assert(QK4_0 == 32, "QK4_0 must be 32"); if (!quant_weights) { @@ -1846,7 +1846,7 @@ static void quantize_row_q4_0_impl(const float * restrict x, block_q4_0 * restri } } -size_t quantize_q4_0(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) { +size_t quantize_q4_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) { if (!quant_weights) { quantize_row_q4_0_ref(src, dst, (int64_t)nrow*n_per_row); return nrow * ggml_row_size(GGML_TYPE_Q4_0, n_per_row); @@ -1861,7 +1861,7 @@ size_t quantize_q4_0(const float * restrict src, void * restrict dst, int64_t nr return nrow * row_size; } -static void quantize_row_q4_1_impl(const float * restrict x, block_q4_1 * restrict y, int64_t n_per_row, const float * quant_weights) { +static void quantize_row_q4_1_impl(const float * GGML_RESTRICT x, block_q4_1 * GGML_RESTRICT y, int64_t n_per_row, const float * quant_weights) { static_assert(QK4_1 == 32, "QK4_1 must be 32"); if (!quant_weights) { @@ -1891,7 +1891,7 @@ static void quantize_row_q4_1_impl(const float * restrict x, block_q4_1 * restri } } -size_t quantize_q4_1(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) { +size_t quantize_q4_1(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) { if (!quant_weights) { quantize_row_q4_1_ref(src, dst, (int64_t)nrow*n_per_row); return nrow * ggml_row_size(GGML_TYPE_Q4_1, n_per_row); @@ -1906,7 +1906,7 @@ size_t quantize_q4_1(const float * restrict src, void * restrict dst, int64_t nr return nrow * row_size; } -static void quantize_row_q5_0_impl(const float * restrict x, block_q5_0 * restrict y, int64_t n_per_row, const float * quant_weights) { +static void quantize_row_q5_0_impl(const float * GGML_RESTRICT x, block_q5_0 * GGML_RESTRICT y, int64_t n_per_row, const float * quant_weights) { static_assert(QK5_0 == 32, "QK5_0 must be 32"); if (!quant_weights) { @@ -1945,7 +1945,7 @@ static void quantize_row_q5_0_impl(const float * restrict x, block_q5_0 * restri } } -size_t quantize_q5_0(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) { +size_t quantize_q5_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) { if (!quant_weights) { quantize_row_q5_0_ref(src, dst, (int64_t)nrow*n_per_row); return nrow * ggml_row_size(GGML_TYPE_Q5_0, n_per_row); @@ -1960,7 +1960,7 @@ size_t quantize_q5_0(const float * restrict src, void * restrict dst, int64_t nr return nrow * row_size; } -static void quantize_row_q5_1_impl(const float * restrict x, block_q5_1 * restrict y, int64_t n_per_row, const float * quant_weights) { +static void quantize_row_q5_1_impl(const float * GGML_RESTRICT x, block_q5_1 * GGML_RESTRICT y, int64_t n_per_row, const float * quant_weights) { static_assert(QK5_1 == 32, "QK5_1 must be 32"); if (!quant_weights) { @@ -1998,7 +1998,7 @@ static void quantize_row_q5_1_impl(const float * restrict x, block_q5_1 * restri } } -size_t quantize_q5_1(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) { +size_t quantize_q5_1(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) { if (!quant_weights) { quantize_row_q5_1_ref(src, dst, (int64_t)nrow*n_per_row); return nrow * ggml_row_size(GGML_TYPE_Q5_1, n_per_row); @@ -2013,7 +2013,7 @@ size_t quantize_q5_1(const float * restrict src, void * restrict dst, int64_t nr return nrow * row_size; } -size_t quantize_q8_0(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) { +size_t quantize_q8_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) { (void)quant_weights; // not used const size_t row_size = ggml_row_size(GGML_TYPE_Q8_0, n_per_row); quantize_row_q8_0_ref(src, dst, (int64_t)nrow*n_per_row); @@ -2022,7 +2022,7 @@ size_t quantize_q8_0(const float * restrict src, void * restrict dst, int64_t nr // ====================== Ternary (de)-quantization (BitNet b1.58 and TriLMs) -void quantize_row_tq1_0_ref(const float * restrict x, block_tq1_0 * restrict y, int64_t k) { +void quantize_row_tq1_0_ref(const float * GGML_RESTRICT x, block_tq1_0 * GGML_RESTRICT y, int64_t k) { assert(k % QK_K == 0); const int64_t nb = k / QK_K; @@ -2088,7 +2088,7 @@ void quantize_row_tq1_0_ref(const float * restrict x, block_tq1_0 * restrict y, } } -void quantize_row_tq2_0_ref(const float * restrict x, block_tq2_0 * restrict y, int64_t k) { +void quantize_row_tq2_0_ref(const float * GGML_RESTRICT x, block_tq2_0 * GGML_RESTRICT y, int64_t k) { assert(k % QK_K == 0); const int64_t nb = k / QK_K; @@ -2120,21 +2120,21 @@ void quantize_row_tq2_0_ref(const float * restrict x, block_tq2_0 * restrict y, } } -size_t quantize_tq1_0(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) { +size_t quantize_tq1_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) { (void)quant_weights; // not used const size_t row_size = ggml_row_size(GGML_TYPE_TQ1_0, n_per_row); quantize_row_tq1_0_ref(src, dst, (int64_t)nrow*n_per_row); return nrow * row_size; } -size_t quantize_tq2_0(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) { +size_t quantize_tq2_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) { (void)quant_weights; // not used const size_t row_size = ggml_row_size(GGML_TYPE_TQ2_0, n_per_row); quantize_row_tq2_0_ref(src, dst, (int64_t)nrow*n_per_row); return nrow * row_size; } -void dequantize_row_tq1_0(const block_tq1_0 * restrict x, float * restrict y, int64_t k) { +void dequantize_row_tq1_0(const block_tq1_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k) { assert(k % QK_K == 0); const int64_t nb = k / QK_K; @@ -2173,7 +2173,7 @@ void dequantize_row_tq1_0(const block_tq1_0 * restrict x, float * restrict y, in } } -void dequantize_row_tq2_0(const block_tq2_0 * restrict x, float * restrict y, int64_t k) { +void dequantize_row_tq2_0(const block_tq2_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k) { assert(k % QK_K == 0); const int64_t nb = k / QK_K; @@ -2194,7 +2194,7 @@ void dequantize_row_tq2_0(const block_tq2_0 * restrict x, float * restrict y, in // ====================== "True" 2-bit (de)-quantization -void dequantize_row_iq2_xxs(const block_iq2_xxs * restrict x, float * restrict y, int64_t k) { +void dequantize_row_iq2_xxs(const block_iq2_xxs * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k) { assert(k % QK_K == 0); const int64_t nb = k / QK_K; @@ -2222,7 +2222,7 @@ void dequantize_row_iq2_xxs(const block_iq2_xxs * restrict x, float * restrict y // ====================== 2.3125 bpw (de)-quantization -void dequantize_row_iq2_xs(const block_iq2_xs * restrict x, float * restrict y, int64_t k) { +void dequantize_row_iq2_xs(const block_iq2_xs * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k) { assert(k % QK_K == 0); const int64_t nb = k / QK_K; @@ -2249,7 +2249,7 @@ void dequantize_row_iq2_xs(const block_iq2_xs * restrict x, float * restrict y, // ====================== 2.5625 bpw (de)-quantization -void dequantize_row_iq2_s(const block_iq2_s * restrict x, float * restrict y, int64_t k) { +void dequantize_row_iq2_s(const block_iq2_s * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k) { assert(k % QK_K == 0); const int64_t nb = k / QK_K; @@ -2281,7 +2281,7 @@ void dequantize_row_iq2_s(const block_iq2_s * restrict x, float * restrict y, in // ====================== 3.0625 bpw (de)-quantization -void dequantize_row_iq3_xxs(const block_iq3_xxs * restrict x, float * restrict y, int64_t k) { +void dequantize_row_iq3_xxs(const block_iq3_xxs * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k) { assert(k % QK_K == 0); const int64_t nb = k / QK_K; @@ -2313,7 +2313,7 @@ void dequantize_row_iq3_xxs(const block_iq3_xxs * restrict x, float * restrict y // ====================== 3.3125 bpw (de)-quantization -void dequantize_row_iq3_s(const block_iq3_s * restrict x, float * restrict y, int64_t k) { +void dequantize_row_iq3_s(const block_iq3_s * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k) { assert(k % QK_K == 0); const int64_t nb = k / QK_K; @@ -2356,7 +2356,7 @@ void dequantize_row_iq3_s(const block_iq3_s * restrict x, float * restrict y, in // ====================== 1.5625 bpw (de)-quantization -void dequantize_row_iq1_s(const block_iq1_s * restrict x, float * restrict y, int64_t k) { +void dequantize_row_iq1_s(const block_iq1_s * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k) { assert(k % QK_K == 0); const int64_t nb = k / QK_K; @@ -2381,7 +2381,7 @@ void dequantize_row_iq1_s(const block_iq1_s * restrict x, float * restrict y, in } } -void dequantize_row_iq1_m(const block_iq1_m * restrict x, float * restrict y, int64_t k) { +void dequantize_row_iq1_m(const block_iq1_m * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k) { assert(k % QK_K == 0); const int64_t nb = k / QK_K; @@ -2433,7 +2433,7 @@ void dequantize_row_iq1_m(const block_iq1_m * restrict x, float * restrict y, in static const int8_t kvalues_iq4nl[16] = {-127, -104, -83, -65, -49, -35, -22, -10, 1, 13, 25, 38, 53, 69, 89, 113}; -void dequantize_row_iq4_nl(const block_iq4_nl * restrict x, float * restrict y, int64_t k) { +void dequantize_row_iq4_nl(const block_iq4_nl * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k) { assert(k % QK4_NL == 0); const int64_t nb = k / QK4_NL; @@ -2451,7 +2451,7 @@ void dequantize_row_iq4_nl(const block_iq4_nl * restrict x, float * restrict y, } } -void dequantize_row_iq4_xs(const block_iq4_xs * restrict x, float * restrict y, int64_t k) { +void dequantize_row_iq4_xs(const block_iq4_xs * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k) { assert(k % QK_K == 0); const int64_t nb = k / QK_K; @@ -2476,7 +2476,7 @@ void dequantize_row_iq4_xs(const block_iq4_xs * restrict x, float * restrict y, //===================================== Q8_K ============================================== -void quantize_row_q8_K_ref(const float * restrict x, block_q8_K * restrict y, int64_t k) { +void quantize_row_q8_K_ref(const float * GGML_RESTRICT x, block_q8_K * GGML_RESTRICT y, int64_t k) { assert(k % QK_K == 0); const int64_t nb = k / QK_K; @@ -2515,7 +2515,7 @@ void quantize_row_q8_K_ref(const float * restrict x, block_q8_K * restrict y, in } } -void dequantize_row_q8_K(const block_q8_K * restrict x, float * restrict y, int64_t k) { +void dequantize_row_q8_K(const block_q8_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k) { assert(k % QK_K == 0); const int64_t nb = k / QK_K; @@ -2927,8 +2927,8 @@ void iq2xs_free_impl(enum ggml_type type) { } } -static int iq2_find_best_neighbour(const uint16_t * restrict neighbours, const uint64_t * restrict grid, - const float * restrict xval, const float * restrict weight, float scale, int8_t * restrict L) { +static int iq2_find_best_neighbour(const uint16_t * GGML_RESTRICT neighbours, const uint64_t * GGML_RESTRICT grid, + const float * GGML_RESTRICT xval, const float * GGML_RESTRICT weight, float scale, int8_t * GGML_RESTRICT L) { int num_neighbors = neighbours[0]; GGML_ASSERT(num_neighbors > 0); float best_d2 = FLT_MAX; @@ -2951,7 +2951,7 @@ static int iq2_find_best_neighbour(const uint16_t * restrict neighbours, const u return grid_index; } -static void quantize_row_iq2_xxs_impl(const float * restrict x, void * restrict vy, int64_t n, const float * restrict quant_weights) { +static void quantize_row_iq2_xxs_impl(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t n, const float * GGML_RESTRICT quant_weights) { const int gindex = iq2_data_index(GGML_TYPE_IQ2_XXS); @@ -3124,7 +3124,7 @@ static void quantize_row_iq2_xxs_impl(const float * restrict x, void * restrict } } -static void quantize_row_iq2_xs_impl(const float * restrict x, void * restrict vy, int64_t n, const float * restrict quant_weights) { +static void quantize_row_iq2_xs_impl(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t n, const float * GGML_RESTRICT quant_weights) { const int gindex = iq2_data_index(GGML_TYPE_IQ2_XS); @@ -3304,7 +3304,7 @@ static void quantize_row_iq2_xs_impl(const float * restrict x, void * restrict v } } -size_t quantize_iq2_xxs(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) { +size_t quantize_iq2_xxs(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) { GGML_ASSERT(n_per_row%QK_K == 0); int64_t nblock = n_per_row/QK_K; char * qrow = (char *)dst; @@ -3316,7 +3316,7 @@ size_t quantize_iq2_xxs(const float * restrict src, void * restrict dst, int64_t return nrow * nblock * sizeof(block_iq2_xxs); } -size_t quantize_iq2_xs(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) { +size_t quantize_iq2_xs(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) { GGML_ASSERT(n_per_row%QK_K == 0); int64_t nblock = n_per_row/QK_K; char * qrow = (char *)dst; @@ -3521,8 +3521,8 @@ void iq3xs_free_impl(int grid_size) { } } -static int iq3_find_best_neighbour(const uint16_t * restrict neighbours, const uint32_t * restrict grid, - const float * restrict xval, const float * restrict weight, float scale, int8_t * restrict L) { +static int iq3_find_best_neighbour(const uint16_t * GGML_RESTRICT neighbours, const uint32_t * GGML_RESTRICT grid, + const float * GGML_RESTRICT xval, const float * GGML_RESTRICT weight, float scale, int8_t * GGML_RESTRICT L) { int num_neighbors = neighbours[0]; GGML_ASSERT(num_neighbors > 0); float best_d2 = FLT_MAX; @@ -3545,8 +3545,8 @@ static int iq3_find_best_neighbour(const uint16_t * restrict neighbours, const u return grid_index; } -static void quantize_row_iq3_xxs_impl(int grid_size, const float * restrict x, void * restrict vy, int64_t n, - const float * restrict quant_weights) { +static void quantize_row_iq3_xxs_impl(int grid_size, const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t n, + const float * GGML_RESTRICT quant_weights) { const int gindex = iq3_data_index(grid_size); @@ -3758,7 +3758,7 @@ static void quantize_row_iq3_xxs_impl(int grid_size, const float * restrict x, v } } -size_t quantize_iq3_xxs(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) { +size_t quantize_iq3_xxs(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) { GGML_ASSERT(n_per_row%QK_K == 0); int64_t nblock = n_per_row/QK_K; char * qrow = (char *)dst; @@ -3770,13 +3770,13 @@ size_t quantize_iq3_xxs(const float * restrict src, void * restrict dst, int64_t return nrow * nblock * sizeof(block_iq3_xxs); } -void quantize_row_iq3_xxs_ref(const float * restrict x, block_iq3_xxs * restrict y, int64_t k) { +void quantize_row_iq3_xxs_ref(const float * GGML_RESTRICT x, block_iq3_xxs * GGML_RESTRICT y, int64_t k) { assert(k % QK_K == 0); quantize_row_iq3_xxs_impl(256, x, y, k, NULL); } -static void quantize_row_iq3_s_impl(int block_size, const float * restrict x, void * restrict vy, int n, - const float * restrict quant_weights, +static void quantize_row_iq3_s_impl(int block_size, const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int n, + const float * GGML_RESTRICT quant_weights, float * scales, float * weight, float * xval, @@ -3958,7 +3958,7 @@ static void quantize_row_iq3_s_impl(int block_size, const float * restrict x, vo } #define IQ3S_BLOCK_SIZE 32 -size_t quantize_iq3_s(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) { +size_t quantize_iq3_s(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) { GGML_ASSERT(n_per_row%QK_K == 0); int64_t nblock = n_per_row/QK_K; float scales[QK_K/IQ3S_BLOCK_SIZE]; @@ -3980,7 +3980,7 @@ size_t quantize_iq3_s(const float * restrict src, void * restrict dst, int64_t n return nrow * nblock * sizeof(block_iq3_s); } -void quantize_row_iq3_s_ref(const float * restrict x, block_iq3_s * restrict y, int64_t k) { +void quantize_row_iq3_s_ref(const float * GGML_RESTRICT x, block_iq3_s * GGML_RESTRICT y, int64_t k) { assert(k % QK_K == 0); quantize_iq3_s(x, y, 1, k, NULL); } @@ -3988,8 +3988,8 @@ void quantize_row_iq3_s_ref(const float * restrict x, block_iq3_s * restrict y, // =================================== 1.5 bpw =================================================== -static int iq1_find_best_neighbour(const uint16_t * restrict neighbours, const uint64_t * restrict grid, - const float * restrict xval, const float * restrict weight, float * scale, int8_t * restrict L, int ngrid) { +static int iq1_find_best_neighbour(const uint16_t * GGML_RESTRICT neighbours, const uint64_t * GGML_RESTRICT grid, + const float * GGML_RESTRICT xval, const float * GGML_RESTRICT weight, float * scale, int8_t * GGML_RESTRICT L, int ngrid) { int num_neighbors = neighbours[0]; GGML_ASSERT(num_neighbors > 0); float best_score = -FLT_MAX; @@ -4048,8 +4048,8 @@ static int iq1_find_best_neighbour(const uint16_t * restrict neighbours, const u return grid_index; } -static int iq1_find_best_neighbour2(const uint16_t * restrict neighbours, const uint64_t * restrict grid, - const float * restrict xval, const float * restrict weight, float scale, const float * restrict xg, int8_t * restrict L, int ngrid) { +static int iq1_find_best_neighbour2(const uint16_t * GGML_RESTRICT neighbours, const uint64_t * GGML_RESTRICT grid, + const float * GGML_RESTRICT xval, const float * GGML_RESTRICT weight, float scale, const float * GGML_RESTRICT xg, int8_t * GGML_RESTRICT L, int ngrid) { int num_neighbors = neighbours[0]; GGML_ASSERT(num_neighbors > 0); float best_score = FLT_MAX; @@ -4113,7 +4113,7 @@ static int iq1_sort_helper(const void * left, const void * right) { #define IQ1S_BLOCK_SIZE 32 #define IQ1M_BLOCK_SIZE 16 -static void quantize_row_iq1_s_impl(const float * restrict x, void * restrict vy, int64_t n, const float * restrict quant_weights, +static void quantize_row_iq1_s_impl(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t n, const float * GGML_RESTRICT quant_weights, float * scales, float * weight, float * sumx, @@ -4271,7 +4271,7 @@ static void quantize_row_iq1_s_impl(const float * restrict x, void * restrict vy } } -size_t quantize_iq1_s(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) { +size_t quantize_iq1_s(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) { GGML_ASSERT(n_per_row%QK_K == 0); float scales[QK_K/IQ1S_BLOCK_SIZE]; float weight[IQ1S_BLOCK_SIZE]; @@ -4291,7 +4291,7 @@ size_t quantize_iq1_s(const float * restrict src, void * restrict dst, int64_t n return nrow * nblock * sizeof(block_iq1_s); } -static void quantize_row_iq1_m_impl(const float * restrict x, void * restrict vy, int64_t n, const float * restrict quant_weights, +static void quantize_row_iq1_m_impl(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t n, const float * GGML_RESTRICT quant_weights, float * scales, float * weight, float * pairs, @@ -4539,7 +4539,7 @@ static void quantize_row_iq1_m_impl(const float * restrict x, void * restrict vy } } -size_t quantize_iq1_m(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) { +size_t quantize_iq1_m(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) { GGML_ASSERT(n_per_row%QK_K == 0); float scales[QK_K/IQ1M_BLOCK_SIZE]; float weight[IQ1M_BLOCK_SIZE]; @@ -4570,7 +4570,7 @@ static inline int best_index_int8(int n, const int8_t * val, float x) { return x - val[mu-1] < val[mu] - x ? mu-1 : mu; } -static void quantize_row_iq4_nl_impl(const int super_block_size, const int block_size, const float * restrict x, +static void quantize_row_iq4_nl_impl(const int super_block_size, const int block_size, const float * GGML_RESTRICT x, ggml_fp16_t * dh, uint8_t * q4, uint16_t * scales_h, uint8_t * scales_l, float * scales, float * weight, uint8_t * L, const int8_t * values, @@ -4681,7 +4681,7 @@ static void quantize_row_iq4_nl_impl(const int super_block_size, const int block } } -size_t quantize_iq4_nl(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) { +size_t quantize_iq4_nl(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) { GGML_ASSERT(n_per_row%QK4_NL == 0); int64_t nblock = n_per_row/QK4_NL; char * qrow = (char *)dst; @@ -4703,8 +4703,8 @@ size_t quantize_iq4_nl(const float * restrict src, void * restrict dst, int64_t return nrow * nblock * sizeof(block_iq4_nl); } -//void quantize_row_iq4_nl_ref(const float * restrict x, void * restrict vy, int64_t k) { -void quantize_row_iq4_nl_ref(const float * restrict x, block_iq4_nl * restrict y, int64_t k) { +//void quantize_row_iq4_nl_ref(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) { +void quantize_row_iq4_nl_ref(const float * GGML_RESTRICT x, block_iq4_nl * GGML_RESTRICT y, int64_t k) { GGML_ASSERT(k%QK4_NL == 0); int64_t nblock = k/QK4_NL; uint8_t L[QK4_NL]; @@ -4719,7 +4719,7 @@ void quantize_row_iq4_nl_ref(const float * restrict x, block_iq4_nl * restrict y } } -size_t quantize_iq4_xs(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) { +size_t quantize_iq4_xs(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) { GGML_ASSERT(n_per_row%QK_K == 0); int64_t nblock = n_per_row/QK_K; char * qrow = (char *)dst; @@ -4739,14 +4739,14 @@ size_t quantize_iq4_xs(const float * restrict src, void * restrict dst, int64_t return nrow * nblock * sizeof(block_iq4_xs); } -void quantize_row_iq4_xs_ref(const float * restrict x, block_iq4_xs * restrict y, int64_t k) { +void quantize_row_iq4_xs_ref(const float * GGML_RESTRICT x, block_iq4_xs * GGML_RESTRICT y, int64_t k) { assert(k % QK_K == 0); quantize_iq4_xs(x, y, 1, k, NULL); } // =============================== 2.5625 bpw -static void quantize_row_iq2_s_impl(const float * restrict x, void * restrict vy, int64_t n, const float * restrict quant_weights) { +static void quantize_row_iq2_s_impl(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t n, const float * GGML_RESTRICT quant_weights) { const int gindex = iq2_data_index(GGML_TYPE_IQ2_S); @@ -4914,7 +4914,7 @@ static void quantize_row_iq2_s_impl(const float * restrict x, void * restrict vy } } -size_t quantize_iq2_s(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) { +size_t quantize_iq2_s(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) { GGML_ASSERT(n_per_row%QK_K == 0); int64_t nblock = n_per_row/QK_K; char * qrow = (char *)dst; @@ -4926,7 +4926,7 @@ size_t quantize_iq2_s(const float * restrict src, void * restrict dst, int64_t n return nrow * nblock * sizeof(block_iq2_s); } -void quantize_row_iq2_s_ref(const float * restrict x, block_iq2_s * restrict y, int64_t k) { +void quantize_row_iq2_s_ref(const float * GGML_RESTRICT x, block_iq2_s * GGML_RESTRICT y, int64_t k) { assert(k % QK_K == 0); quantize_iq2_s(x, y, 1, k, NULL); } diff --git a/ggml/src/ggml-rpc/ggml-rpc.cpp b/ggml/src/ggml-rpc/ggml-rpc.cpp index 97873acc7..862b9b666 100644 --- a/ggml/src/ggml-rpc/ggml-rpc.cpp +++ b/ggml/src/ggml-rpc/ggml-rpc.cpp @@ -26,6 +26,10 @@ # include #endif #include +#include +#include + +namespace fs = std::filesystem; #ifdef _WIN32 typedef SOCKET sockfd_t; @@ -80,6 +84,7 @@ enum rpc_cmd { RPC_CMD_FREE_BUFFER, RPC_CMD_BUFFER_CLEAR, RPC_CMD_SET_TENSOR, + RPC_CMD_SET_TENSOR_HASH, RPC_CMD_GET_TENSOR, RPC_CMD_COPY_TENSOR, RPC_CMD_GRAPH_COMPUTE, @@ -89,6 +94,9 @@ enum rpc_cmd { RPC_CMD_COUNT, }; +// Try RPC_CMD_SET_TENSOR_HASH first when data size is larger than this threshold +const size_t HASH_THRESHOLD = 10 * 1024 * 1024; + struct rpc_msg_get_alloc_size_req { rpc_tensor tensor; }; @@ -135,6 +143,10 @@ struct rpc_msg_buffer_clear_req { uint8_t value; }; +struct rpc_msg_set_tensor_hash_rsp { + uint8_t result; +}; + struct rpc_msg_get_tensor_req { rpc_tensor tensor; uint64_t offset; @@ -187,6 +199,18 @@ struct ggml_backend_rpc_buffer_context { // RPC helper functions +// Computes FNV-1a hash of the data +static uint64_t fnv_hash(const uint8_t * data, size_t len) { + const uint64_t fnv_prime = 0x100000001b3ULL; + uint64_t hash = 0xcbf29ce484222325ULL; + + for (size_t i = 0; i < len; ++i) { + hash ^= data[i]; + hash *= fnv_prime; + } + return hash; +} + static std::shared_ptr make_socket(sockfd_t fd) { #ifdef _WIN32 if (fd == INVALID_SOCKET) { @@ -464,7 +488,7 @@ static rpc_tensor serialize_tensor(const ggml_tensor * tensor) { return result; } -static void ggml_backend_rpc_buffer_init_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor) { +static enum ggml_status ggml_backend_rpc_buffer_init_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor) { ggml_backend_rpc_buffer_context * ctx = (ggml_backend_rpc_buffer_context *)buffer->context; // CUDA backend on the server pads everything to 512 due to CUDA limitations. @@ -478,14 +502,31 @@ static void ggml_backend_rpc_buffer_init_tensor(ggml_backend_buffer_t buffer, gg bool status = send_rpc_cmd(ctx->sock, RPC_CMD_INIT_TENSOR, &request, sizeof(request), nullptr, 0); GGML_ASSERT(status); } + return GGML_STATUS_SUCCESS; } static void ggml_backend_rpc_buffer_set_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor, const void * data, size_t offset, size_t size) { ggml_backend_rpc_buffer_context * ctx = (ggml_backend_rpc_buffer_context *)buffer->context; - // input serialization format: | rpc_tensor | offset (8 bytes) | data (size bytes) | + rpc_tensor rpc_tensor = serialize_tensor(tensor); + if (size > HASH_THRESHOLD) { + // input serialization format: | rpc_tensor | offset (8 bytes) | hash (8 bytes) + size_t input_size = sizeof(rpc_tensor) + sizeof(uint64_t) + sizeof(uint64_t); + std::vector input(input_size, 0); + uint64_t hash = fnv_hash((const uint8_t*)data, size); + memcpy(input.data(), &rpc_tensor, sizeof(rpc_tensor)); + memcpy(input.data() + sizeof(rpc_tensor), &offset, sizeof(offset)); + memcpy(input.data() + sizeof(rpc_tensor) + sizeof(offset), &hash, sizeof(hash)); + rpc_msg_set_tensor_hash_rsp response; + bool status = send_rpc_cmd(ctx->sock, RPC_CMD_SET_TENSOR_HASH, input.data(), input.size(), &response, sizeof(response)); + GGML_ASSERT(status); + if (response.result) { + // the server has the same data, no need to send it + return; + } + } + // input serialization format: | rpc_tensor | offset (8 bytes) | data (size bytes) size_t input_size = sizeof(rpc_tensor) + sizeof(uint64_t) + size; std::vector input(input_size, 0); - rpc_tensor rpc_tensor = serialize_tensor(tensor); memcpy(input.data(), &rpc_tensor, sizeof(rpc_tensor)); memcpy(input.data() + sizeof(rpc_tensor), &offset, sizeof(offset)); memcpy(input.data() + sizeof(rpc_tensor) + sizeof(offset), data, size); @@ -771,7 +812,9 @@ void ggml_backend_rpc_get_device_memory(const char * endpoint, size_t * free, si class rpc_server { public: - rpc_server(ggml_backend_t backend) : backend(backend) {} + rpc_server(ggml_backend_t backend, const char * cache_dir) + : backend(backend), cache_dir(cache_dir) { + } ~rpc_server(); void alloc_buffer(const rpc_msg_alloc_buffer_req & request, rpc_msg_alloc_buffer_rsp & response); @@ -781,6 +824,7 @@ public: bool free_buffer(const rpc_msg_free_buffer_req & request); bool buffer_clear(const rpc_msg_buffer_clear_req & request); bool set_tensor(const std::vector & input); + bool set_tensor_hash(const std::vector & input, rpc_msg_set_tensor_hash_rsp & response); bool get_tensor(const rpc_msg_get_tensor_req & request, std::vector & response); bool copy_tensor(const rpc_msg_copy_tensor_req & request, rpc_msg_copy_tensor_rsp & response); bool graph_compute(const std::vector & input, rpc_msg_graph_compute_rsp & response); @@ -788,6 +832,7 @@ public: bool get_alloc_size(const rpc_msg_get_alloc_size_req & request, rpc_msg_get_alloc_size_rsp & response); private: + bool get_cached_file(uint64_t hash, std::vector & data); ggml_tensor * deserialize_tensor(struct ggml_context * ctx, const rpc_tensor * tensor); ggml_tensor * create_node(uint64_t id, struct ggml_context * ctx, @@ -796,6 +841,7 @@ private: ggml_backend_t backend; + const char * cache_dir; std::unordered_set buffers; }; @@ -959,11 +1005,85 @@ bool rpc_server::set_tensor(const std::vector & input) { } const void * data = input.data() + sizeof(rpc_tensor) + sizeof(offset); + if (cache_dir && size > HASH_THRESHOLD) { + uint64_t hash = fnv_hash((const uint8_t*)data, size); + char hash_str[17]; + snprintf(hash_str, sizeof(hash_str), "%016" PRIx64, hash); + // save to cache_dir/hash_str + fs::path cache_file = fs::path(cache_dir) / hash_str; + std::ofstream ofs(cache_file, std::ios::binary); + ofs.write((const char *)data, size); + printf("[%s] saved to '%s'\n", __func__, cache_file.c_str()); + } ggml_backend_tensor_set(tensor, data, offset, size); ggml_free(ctx); return true; } +bool rpc_server::get_cached_file(uint64_t hash, std::vector & data) { + if (!cache_dir) { + return false; + } + char hash_str[17]; + snprintf(hash_str, sizeof(hash_str), "%016" PRIx64, hash); + fs::path cache_file = fs::path(cache_dir) / hash_str; + if (!fs::exists(cache_file)) { + return false; + } + std::ifstream ifs(cache_file, std::ios::binary); + ifs.seekg(0, std::ios::end); + size_t size = ifs.tellg(); + ifs.seekg(0, std::ios::beg); + data.resize(size); + ifs.read((char *)data.data(), size); + return true; +} + +bool rpc_server::set_tensor_hash(const std::vector & input, rpc_msg_set_tensor_hash_rsp & response) +{ + // serialization format: | rpc_tensor | offset (8 bytes) | hash (8 bytes) | + if (input.size() != sizeof(rpc_tensor) + 16) { + return false; + } + const rpc_tensor * in_tensor = (const rpc_tensor *)input.data(); + uint64_t offset; + memcpy(&offset, input.data() + sizeof(rpc_tensor), sizeof(offset)); + const uint64_t * hash = (const uint64_t *)(input.data() + sizeof(rpc_tensor) + sizeof(offset)); + std::vector cached_file; + if (!get_cached_file(*hash, cached_file)) { + response.result = 0; + return true; + } + size_t size = cached_file.size(); + struct ggml_init_params params { + /*.mem_size =*/ ggml_tensor_overhead(), + /*.mem_buffer =*/ NULL, + /*.no_alloc =*/ true, + }; + struct ggml_context * ctx = ggml_init(params); + ggml_tensor * tensor = deserialize_tensor(ctx, in_tensor); + if (tensor == nullptr) { + GGML_LOG_ERROR("[%s] error deserializing tensor\n", __func__); + ggml_free(ctx); + return false; + } + GGML_PRINT_DEBUG("[%s] buffer: %p, data: %p, offset: %" PRIu64 ", size: %zu, hash: %" PRIx64 "\n", __func__, (void*)tensor->buffer, tensor->data, offset, size, *hash); + + // sanitize tensor->data + { + const size_t p0 = (size_t) ggml_backend_buffer_get_base(tensor->buffer); + const size_t p1 = p0 + ggml_backend_buffer_get_size(tensor->buffer); + + if (in_tensor->data + offset < p0 || in_tensor->data + offset >= p1 || size > (p1 - in_tensor->data - offset)) { + GGML_ABORT("[%s] tensor->data out of bounds\n", __func__); + } + } + ggml_backend_tensor_set(tensor, cached_file.data(), offset, size); + response.result = 1; + ggml_free(ctx); + return true; +} + bool rpc_server::init_tensor(const rpc_msg_init_tensor_req & request) { struct ggml_init_params params { /*.mem_size =*/ ggml_tensor_overhead(), @@ -1147,8 +1267,9 @@ rpc_server::~rpc_server() { } } -static void rpc_serve_client(ggml_backend_t backend, sockfd_t sockfd, size_t free_mem, size_t total_mem) { - rpc_server server(backend); +static void rpc_serve_client(ggml_backend_t backend, const char * cache_dir, + sockfd_t sockfd, size_t free_mem, size_t total_mem) { + rpc_server server(backend, cache_dir); while (true) { uint8_t cmd; if (!recv_data(sockfd, &cmd, 1)) { @@ -1259,6 +1380,20 @@ static void rpc_serve_client(ggml_backend_t backend, sockfd_t sockfd, size_t fre } break; } + case RPC_CMD_SET_TENSOR_HASH: { + std::vector input; + if (!recv_msg(sockfd, input)) { + return; + } + rpc_msg_set_tensor_hash_rsp response; + if (!server.set_tensor_hash(input, response)) { + return; + } + if (!send_msg(sockfd, &response, sizeof(response))) { + return; + } + break; + } case RPC_CMD_INIT_TENSOR: { rpc_msg_init_tensor_req request; if (!recv_msg(sockfd, &request,sizeof(request))) { @@ -1334,7 +1469,9 @@ static void rpc_serve_client(ggml_backend_t backend, sockfd_t sockfd, size_t fre } } -void ggml_backend_rpc_start_server(ggml_backend_t backend, const char * endpoint, size_t free_mem, size_t total_mem) { +void ggml_backend_rpc_start_server(ggml_backend_t backend, const char * endpoint, + const char * cache_dir, + size_t free_mem, size_t total_mem) { std::string host; int port; if (!parse_endpoint(endpoint, host, port)) { @@ -1363,7 +1500,7 @@ void ggml_backend_rpc_start_server(ggml_backend_t backend, const char * endpoint } printf("Accepted client connection, free_mem=%zu, total_mem=%zu\n", free_mem, total_mem); fflush(stdout); - rpc_serve_client(backend, client_socket->fd, free_mem, total_mem); + rpc_serve_client(backend, cache_dir, client_socket->fd, free_mem, total_mem); printf("Client connection closed\n"); fflush(stdout); } diff --git a/ggml/src/ggml-sycl/CMakeLists.txt b/ggml/src/ggml-sycl/CMakeLists.txt index 3579a311a..6699b70ba 100644 --- a/ggml/src/ggml-sycl/CMakeLists.txt +++ b/ggml/src/ggml-sycl/CMakeLists.txt @@ -1,3 +1,5 @@ +message(STATUS "GGML_SYCL_TARGET=${GGML_SYCL_TARGET}") + if (NOT GGML_SYCL_TARGET MATCHES "^(INTEL|NVIDIA|AMD)$") message(FATAL_ERROR "Invalid backend chosen, supported options are INTEL, NVIDIA, or AMD") endif() @@ -21,6 +23,64 @@ ggml_add_backend_library(ggml-sycl ../../include/ggml-sycl.h ) +file(GLOB GGML_HEADERS_SYCL "*.hpp") +file(GLOB GGML_SOURCES_SYCL "*.cpp") +target_sources(ggml-sycl PRIVATE ${GGML_HEADERS_SYCL} ${GGML_SOURCES_SYCL}) + +if (WIN32) + # To generate a Visual Studio solution, using Intel C++ Compiler for ggml-sycl is mandatory + if( ${CMAKE_GENERATOR} MATCHES "Visual Studio" AND NOT (${CMAKE_GENERATOR_TOOLSET} MATCHES "Intel C")) + set_target_properties(ggml-sycl PROPERTIES VS_PLATFORM_TOOLSET "Intel C++ Compiler 2025") + set(CMAKE_CXX_COMPILER "icx") + set(CMAKE_CXX_COMPILER_ID "IntelLLVM") + endif() +endif() + +find_package(IntelSYCL) +if (IntelSYCL_FOUND) + # Use oneAPI CMake when possible + target_link_libraries(ggml-sycl PRIVATE IntelSYCL::SYCL_CXX) +else() + # Fallback to the simplest way of enabling SYCL when using intel/llvm nightly for instance + target_compile_options(ggml-sycl PRIVATE "-fsycl") + target_link_options(ggml-sycl PRIVATE "-fsycl") +endif() + +target_compile_options(ggml-sycl PRIVATE "-Wno-narrowing") + +# Link against oneDNN +find_package(DNNL) +set(GGML_SYCL_DNNL 0) +if(DNNL_FOUND) + if (DEFINED ENV{ONEAPI_ROOT} AND NOT DEFINED DNNL_GPU_VENDOR) + # Assuming oneDNN packaged with oneapi release is used which + # supports only intel target + set(DNNL_GPU_VENDOR "INTEL") + if(NOT "${GGML_SYCL_TARGET}" STREQUAL "INTEL") + message(WARNING "oneDNN builds bundled with oneapi release only support INTEL target") + endif() + endif() + + # Verify oneDNN was compiled for the same target as llama + if("${GGML_SYCL_TARGET}" STREQUAL "${DNNL_GPU_VENDOR}") + target_link_libraries(ggml-sycl PRIVATE DNNL::dnnl) + set(GGML_SYCL_DNNL 1) + get_target_property(CONFIGS DNNL::dnnl IMPORTED_CONFIGURATIONS) + foreach(CONFIG ${CONFIGS}) + get_target_property(DNNL_LIB DNNL::dnnl IMPORTED_LOCATION_${CONFIG}) + message(STATUS "Found oneDNN: ${DNNL_LIB}") + endforeach() + else() + message(WARNING + "oneDNN must be compiled for the same target as llama.cpp. + llama.cpp: ${GGML_SYCL_TARGET}, oneDNN: ${DNNL_GPU_VENDOR}. + Disabling oneDNN support.") + endif() +else() + message(STATUS "oneDNN not found, disabling oneDNN support") +endif() +target_compile_definitions(ggml-sycl PRIVATE GGML_SYCL_DNNL=${GGML_SYCL_DNNL}) + if (GGML_SYCL_F16) if (GGML_SYCL_TARGET STREQUAL "AMD") message(WARNING "AMD target does not entirely support FP16 in the SYCL backend.") @@ -28,8 +88,6 @@ if (GGML_SYCL_F16) add_compile_definitions(GGML_SYCL_F16) endif() -set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wno-narrowing -fsycl") - if (GGML_SYCL_TARGET STREQUAL "NVIDIA") add_compile_definitions(GGML_SYCL_WARP_SIZE=32) elseif (GGML_SYCL_TARGET STREQUAL "AMD") @@ -42,43 +100,84 @@ else() add_compile_definitions(GGML_SYCL_WARP_SIZE=16) endif() -file(GLOB GGML_HEADERS_SYCL "*.hpp") -file(GLOB GGML_SOURCES_SYCL "*.cpp") -target_sources(ggml-sycl PRIVATE ${GGML_HEADERS_SYCL} ${GGML_SOURCES_SYCL}) - -find_package(DNNL) -message("-- DNNL found:" ${DNNL_FOUND}) +if (GGML_SYCL_GRAPH) + target_compile_definitions(ggml-sycl PRIVATE GGML_SYCL_GRAPH) +endif() +# Link against Intel oneMKL or oneMath if (GGML_SYCL_TARGET STREQUAL "INTEL") - add_compile_definitions(GGML_SYCL_DNNL=${DNNL_FOUND}) -else() - add_compile_definitions(GGML_SYCL_DNNL=0) -endif() - -if (${DNNL_FOUND} AND GGML_SYCL_TARGET STREQUAL "INTEL") - target_link_libraries(ggml-sycl PRIVATE DNNL::dnnl) -endif() - -if (WIN32) - find_package(IntelSYCL REQUIRED) + # Intel devices use Intel oneMKL directly instead of oneMath to avoid the limitation of linking Intel oneMKL statically + # See https://github.com/uxlfoundation/oneMath/issues/654 find_package(MKL REQUIRED) - target_link_libraries(ggml-sycl PRIVATE IntelSYCL::SYCL_CXX MKL::MKL MKL::MKL_SYCL) + target_link_libraries(ggml-sycl PRIVATE MKL::MKL_SYCL::BLAS) + target_compile_definitions(ggml-sycl PRIVATE GGML_SYCL_USE_INTEL_ONEMKL) else() - if (GGML_SYCL_TARGET STREQUAL "INTEL") - target_link_libraries(ggml-sycl PRIVATE sycl OpenCL mkl_core pthread m dl mkl_sycl_blas mkl_intel_ilp64 mkl_tbb_thread) - elseif (GGML_SYCL_TARGET STREQUAL "NVIDIA") - set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fsycl-targets=nvptx64-nvidia-cuda") - add_compile_definitions(GGML_SYCL_NVIDIA) - target_link_libraries(ggml-sycl PRIVATE sycl pthread m dl onemkl_blas_cublas) + find_package(oneMath QUIET) + if (NOT oneMath_FOUND) + message(STATUS "oneMath not found: oneMath will be automatically downloaded") + # Use FetchContent to automatically pull and build oneMath + include(FetchContent) + set(BUILD_FUNCTIONAL_TESTS False) + set(BUILD_EXAMPLES False) + set(TARGET_DOMAINS blas) + if (GGML_SYCL_TARGET STREQUAL "NVIDIA") + set(ENABLE_MKLCPU_BACKEND False) + set(ENABLE_MKLGPU_BACKEND False) + set(ENABLE_CUBLAS_BACKEND True) + elseif (GGML_SYCL_TARGET STREQUAL "AMD") + set(ENABLE_MKLCPU_BACKEND False) + set(ENABLE_MKLGPU_BACKEND False) + set(ENABLE_ROCBLAS_BACKEND True) + # Ensure setting a string variable here is not overriden by oneMath CACHE variables + cmake_policy(SET CMP0126 NEW) + # Setting the device architecture is only needed and useful for AMD devices in oneMath + set(HIP_TARGETS ${GGML_SYCL_DEVICE_ARCH} CACHE STRING "oneMath HIP target" FORCE) + endif() + FetchContent_Declare( + ONEMATH + GIT_REPOSITORY https://github.com/uxlfoundation/oneMath.git + GIT_TAG c255b1b4c41e2ee3059455c1f96a965d6a62568a + ) + FetchContent_MakeAvailable(ONEMATH) + # Create alias to match with find_package targets name + function(onemath_alias target) + if (TARGET ${target}_obj) + # Silence verbose warnings from external libraries + target_compile_options(${target}_obj PRIVATE -w) + endif() + if (TARGET ${target}) + add_library(ONEMATH::${target} ALIAS ${target}) + endif() + endfunction() + onemath_alias(onemath) + onemath_alias(onemath_blas_mklcpu) + onemath_alias(onemath_blas_mklgpu) + onemath_alias(onemath_blas_cublas) + onemath_alias(onemath_blas_rocblas) + endif() + + # Below oneMath compile-time dispatching is used for better performance + if (GGML_SYCL_TARGET STREQUAL "NVIDIA") + target_link_libraries(ggml-sycl PRIVATE ONEMATH::onemath_blas_cublas) + target_compile_options(ggml-sycl PRIVATE "-fsycl-targets=nvptx64-nvidia-cuda") + target_link_options(ggml-sycl PRIVATE "-fsycl-targets=nvptx64-nvidia-cuda") + target_compile_definitions(ggml-sycl PRIVATE GGML_SYCL_NVIDIA) elseif (GGML_SYCL_TARGET STREQUAL "AMD") if (NOT GGML_SYCL_DEVICE_ARCH) message(ERROR "Can't enable SYCL hip backend, GGML_SYCL_DEVICE_ARCH has not been set.") endif() - set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fsycl-targets=amdgcn-amd-amdhsa") - target_link_libraries(ggml-sycl PRIVATE sycl pthread m dl onemkl) + target_link_libraries(ggml-sycl PRIVATE ONEMATH::onemath_blas_rocblas) + target_compile_options(ggml-sycl PRIVATE "-fsycl-targets=amdgcn-amd-amdhsa") + target_link_options(ggml-sycl PRIVATE "-fsycl-targets=amdgcn-amd-amdhsa") + target_compile_definitions(ggml-sycl PRIVATE GGML_SYCL_AMD) + else() + # Fallback to oneMath runtime dispatcher + target_link_libraries(ggml-sycl PRIVATE ONEMATH::onemath) + target_compile_definitions(ggml-sycl PRIVATE GGML_SYCL_GENERIC) endif() - - if (GGML_SYCL_DEVICE_ARCH) - set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Xsycl-target-backend --offload-arch=${GGML_SYCL_DEVICE_ARCH}") - endif() +endif() + +if (GGML_SYCL_DEVICE_ARCH) + target_compile_options(ggml-sycl PRIVATE -Xsycl-target-backend --offload-arch=${GGML_SYCL_DEVICE_ARCH}) + target_link_options(ggml-sycl PRIVATE -Xsycl-target-backend --offload-arch=${GGML_SYCL_DEVICE_ARCH}) endif() diff --git a/ggml/src/ggml-sycl/backend.hpp b/ggml/src/ggml-sycl/backend.hpp index b1df4e5db..73d807cab 100644 --- a/ggml/src/ggml-sycl/backend.hpp +++ b/ggml/src/ggml-sycl/backend.hpp @@ -26,9 +26,10 @@ #include "softmax.hpp" #include "tsembd.hpp" #include "im2col.hpp" -#include "wkv6.hpp" +#include "wkv.hpp" #include "outprod.hpp" #include "element_wise.hpp" +#include "cpy.hpp" #include "gla.hpp" #endif // GGML_SYCL_BACKEND_HPP diff --git a/ggml/src/ggml-sycl/common.cpp b/ggml/src/ggml-sycl/common.cpp index 022e7b763..05fd5ef46 100644 --- a/ggml/src/ggml-sycl/common.cpp +++ b/ggml/src/ggml-sycl/common.cpp @@ -66,36 +66,18 @@ int64_t downsample_sycl_global_range(int64_t accumulate_block_num, int64_t block return sycl_down_blk_size; } -void ggml_sycl_op_flatten(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, - const ggml_tensor *src1, ggml_tensor *dst, - const ggml_sycl_op_flatten_t op) try { - - const bool use_src1 = src1 != nullptr; - if(use_src1) - GGML_ASSERT(strcmp(src1->buffer->buft->iface.get_name(src1->buffer->buft), GGML_SYCL_NAME "_Split") != 0); - GGML_ASSERT(strcmp(dst->buffer->buft->iface.get_name(dst->buffer->buft), GGML_SYCL_NAME "_Split") != 0); - - // dd = data device - float * src0_ddf = (float *) src0->data; - float * src1_ddf = use_src1 ? (float *) src1->data : nullptr; - float * dst_ddf = (float *) dst->data; - - ggml_sycl_pool_alloc src0_f(ctx.pool()); - ggml_sycl_pool_alloc src1_f(ctx.pool()); - ggml_sycl_pool_alloc dst_f(ctx.pool()); - - ggml_sycl_set_device(ctx.device); - queue_ptr main_stream = ctx.stream(); - // GGML_SYCL_DEBUG("ctx.device=%d, main_stream=%p src0_on_device=%d, src1_on_device=%d, dst_on_device=%d\n", - // ctx.device, main_stream, src0_on_device, src1_on_device, dst_on_device); - - // do the computation - op(ctx, src0, src1, dst, src0_ddf, src1_ddf, dst_ddf, main_stream); - // print_ggml_tensor("tensor", dst); -} -catch (sycl::exception const &exc) { - - std::cerr << exc.what() << "Exception caught at file:" << __FILE__ - << ", line:" << __LINE__ << std::endl; - std::exit(1); +void release_extra_gpu(ggml_tensor_extra_gpu * extra, std::vector streams) { + for (int i = 0; i < ggml_sycl_info().device_count; ++i) { + for (int64_t is = 0; is < GGML_SYCL_MAX_STREAMS; ++is) { + if (extra->events[i][is] != nullptr) { + SYCL_CHECK(CHECK_TRY_ERROR(dpct::destroy_event(extra->events[i][is]))); + } + } + if (extra->data_device[i] != nullptr && streams.size()>0) { + ggml_sycl_set_device(i); + SYCL_CHECK( + CHECK_TRY_ERROR(sycl::free(extra->data_device[i], *(streams[i])))); + } + } + delete extra; } diff --git a/ggml/src/ggml-sycl/common.hpp b/ggml/src/ggml-sycl/common.hpp index abad847ca..3e1ceeaa4 100644 --- a/ggml/src/ggml-sycl/common.hpp +++ b/ggml/src/ggml-sycl/common.hpp @@ -19,6 +19,9 @@ #include "dpct/helper.hpp" #include "ggml-sycl.h" #include "presets.hpp" +#include "sycl_hw.hpp" + + #if GGML_SYCL_DNNL #include "dnnl.hpp" #include "dnnl_sycl.hpp" @@ -31,11 +34,15 @@ #pragma clang diagnostic ignored "-Wnested-anon-types" #include "ggml-common.h" #pragma clang diagnostic pop +#include "ggml-impl.h" void* ggml_sycl_host_malloc(size_t size); void ggml_sycl_host_free(void* ptr); -static int g_ggml_sycl_debug = 0; + +extern int g_ggml_sycl_debug; +extern int g_ggml_sycl_disable_optimize; + #define GGML_SYCL_DEBUG(...) \ do { \ if (g_ggml_sycl_debug) \ @@ -163,7 +170,6 @@ static size_t g_scratch_offset = 0; int get_current_device_id(); inline dpct::err0 ggml_sycl_set_device(const int device) try { - int current_device_id; SYCL_CHECK(CHECK_TRY_ERROR(current_device_id = get_current_device_id())); @@ -182,18 +188,24 @@ inline dpct::err0 ggml_sycl_set_device(const int device) try { } ////////////////////// +struct optimize_feature { + bool reorder=false; +}; + +struct sycl_device_info { + int cc; // compute capability + // int nsm; // number of streaming multiprocessors + // size_t smpb; // max. shared memory per block + bool vmm; // virtual memory support + size_t total_vram; + sycl_hw_info hw_info; + optimize_feature opt_feature; +}; + struct ggml_sycl_device_info { int device_count; - struct sycl_device_info { - int cc; // compute capability - // int nsm; // number of streaming multiprocessors - // size_t smpb; // max. shared memory per block - bool vmm; // virtual memory support - size_t total_vram; - }; - sycl_device_info devices[GGML_SYCL_MAX_DEVICES] = {}; std::array default_tensor_split = {}; @@ -229,6 +241,14 @@ struct ggml_sycl_pool_alloc { } } + T * realloc(size_t size) { + GGML_ASSERT(pool != nullptr); + if (ptr) + pool->free(ptr, actual_size); + ptr = (T *) pool->alloc(size * sizeof(T), &this->actual_size); + return ptr; + } + // size is in number of elements T * alloc(size_t size) { GGML_ASSERT(pool != nullptr); @@ -260,17 +280,47 @@ struct ggml_tensor_extra_gpu { // tensors dpct::event_ptr events[GGML_SYCL_MAX_DEVICES] [GGML_SYCL_MAX_STREAMS]; // events for synchronizing multiple GPUs + optimize_feature optimized_feature; }; +void release_extra_gpu(ggml_tensor_extra_gpu * extra, std::vector streams={}); + +inline optimize_feature check_gpu_optimize_feature(syclex::architecture &arch) { + optimize_feature opt; + + opt.reorder = + (arch == syclex::architecture::intel_gpu_dg1 || + arch == syclex::architecture::intel_gpu_acm_g10 || + arch == syclex::architecture::intel_gpu_acm_g11 || + arch == syclex::architecture::intel_gpu_acm_g12 || + arch == syclex::architecture::intel_gpu_pvc || + arch == syclex::architecture::intel_gpu_pvc_vg || + arch == syclex::architecture::intel_gpu_mtl_u || + arch == syclex::architecture::intel_gpu_mtl_s || + arch == syclex::architecture::intel_gpu_mtl_h || + arch == syclex::architecture::intel_gpu_arl_u || + arch == syclex::architecture::intel_gpu_arl_s || + arch == syclex::architecture::intel_gpu_arl_h || + arch == syclex::architecture::intel_gpu_bmg_g21 || + arch == syclex::architecture::intel_gpu_lnl_m + ); + + return opt; +} + +namespace sycl_ex = sycl::ext::oneapi::experimental; struct ggml_backend_sycl_context { int device; std::string name; + optimize_feature opt_feature; + bool optimized_graph=false; queue_ptr qptrs[GGML_SYCL_MAX_DEVICES][GGML_SYCL_MAX_STREAMS] = { { nullptr } }; explicit ggml_backend_sycl_context(int device) : device(device), name(GGML_SYCL_NAME + std::to_string(device)) { + opt_feature = ggml_sycl_info().devices[device].opt_feature; } queue_ptr stream(int device, int stream) { @@ -328,10 +378,29 @@ struct ggml_backend_sycl_context { dnnl::stream stream_dnnl() { return stream_dnnl(device, 0); } + dnnl::memory get_scratchpad_mem(const dnnl::memory::desc & scratchpad_md, + const dnnl::engine & eng, const queue_ptr q) { + ggml_sycl_pool_alloc * pool; + auto it = scratchpad_map.find(q); + if (it == scratchpad_map.end()) { + scratchpad_map[q] = std::make_unique>(this->pool()); + pool = scratchpad_map[q].get(); + } else { + pool = it->second.get(); + } + + size_t scratchpad_size = scratchpad_md.get_size(); + if (scratchpad_size > pool->actual_size) { + pool->realloc(scratchpad_size); + } + void * mem_ptr = pool->get(); + return dnnl::memory(scratchpad_md, eng, mem_ptr); + } #endif // pool std::unique_ptr pools[GGML_SYCL_MAX_DEVICES]; + std::unordered_map>> scratchpad_map; std::unique_ptr host_pools[GGML_SYCL_MAX_DEVICES]; @@ -350,6 +419,10 @@ struct ggml_backend_sycl_context { return pool(device); } +#ifdef GGML_SYCL_GRAPH + std::unique_ptr> exec_graph = nullptr; +#endif + ggml_sycl_pool & host_pool(int device) { if (host_pools[device] == nullptr) { host_pools[device] = new_pool_for_host(stream(device, 0), device); @@ -421,17 +494,12 @@ static __dpct_inline__ Tp* get_pointer(sycl::local_accessor acc) { int64_t downsample_sycl_global_range(int64_t accumulate_block_num, int64_t block_size); -typedef void (*ggml_sycl_op_flatten_t)(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, - const ggml_tensor *src1, - ggml_tensor *dst, const float *src0_dd, - const float *src1_dd, float *dst_dd, - const queue_ptr &main_stream); - template static void k_bin_bcast(const src0_t * src0, const src1_t * src1, dst_t * dst, int ne0, int ne1, int ne2, int ne3, int ne10, int ne11, int ne12, int ne13, /*int s0, */ int s1, int s2, int s3, + /*int s00,*/ int s01, int s02, int s03, /*int s10,*/ int s11, int s12, int s13, const sycl::nd_item<3> &item_ct1) { const int i0s = item_ct1.get_local_range(2) * item_ct1.get_group(2) + @@ -453,9 +521,9 @@ static void k_bin_bcast(const src0_t * src0, const src1_t * src1, dst_t * dst, const int i12 = i2 % ne12; const int i13 = i3 % ne13; - const size_t i_src0 = i3*s3 + i2*s2 + i1*s1; + const size_t i_src0 = i3*s03 + i2*s02 + i1*s01; const size_t i_src1 = i13*s13 + i12*s12 + i11*s11; - const size_t i_dst = i_src0; + const size_t i_dst = i3*s3 + i2*s2 + i1*s1; const src0_t * src0_row = src0 + i_src0; const src1_t * src1_row = src1 + i_src1; @@ -473,6 +541,7 @@ static void k_bin_bcast_unravel(const src0_t * src0, const src1_t * src1, dst_t int ne0, int ne1, int ne2, int ne3, int ne10, int ne11, int ne12, int ne13, /*int s0, */ int s1, int s2, int s3, + /*int s00,*/ int s01, int s02, int s03, /*int s10,*/ int s11, int s12, int s13, const sycl::nd_item<3> &item_ct1) { @@ -492,9 +561,9 @@ static void k_bin_bcast_unravel(const src0_t * src0, const src1_t * src1, dst_t const int i12 = i2 % ne12; const int i13 = i3 % ne13; - const size_t i_src0 = i3*s3 + i2*s2 + i1*s1; + const size_t i_src0 = i3*s03 + i2*s02 + i1*s01; const size_t i_src1 = i13*s13 + i12*s12 + i11*s11; - const size_t i_dst = i_src0; + const size_t i_dst = i3*s3 + i2*s2 + i1*s1; const src0_t * src0_row = src0 + i_src0; const src1_t * src1_row = src1 + i_src1; @@ -524,9 +593,11 @@ struct bin_bcast_sycl { int nr[4] = { nr0, nr1, nr2, nr3 }; // collapse dimensions until first broadcast dimension - int64_t cne0[] = {ne0, ne1, ne2, ne3}; + int64_t cne[] = {ne0, ne1, ne2, ne3}; + int64_t cne0[] = {ne00, ne01, ne02, ne03}; int64_t cne1[] = {ne10, ne11, ne12, ne13}; - size_t cnb0[] = {nb0, nb1, nb2, nb3}; + size_t cnb[] = {nb0, nb1, nb2, nb3}; + size_t cnb0[] = {nb00, nb01, nb02, nb03}; size_t cnb1[] = {nb10, nb11, nb12, nb13}; auto collapse = [](int64_t cne[]) { cne[0] *= cne[1]; @@ -541,32 +612,41 @@ struct bin_bcast_sycl { cnb[3] *= cne[3]; }; - for (int i = 0; i < 4; i++) { - if (nr[i] != 1) { - break; - } - if (i > 0) { - collapse_nb(cnb0, cne0); - collapse_nb(cnb1, cne1); - collapse(cne0); - collapse(cne1); + if (ggml_is_contiguous(src0) && ggml_is_contiguous(src1) && ggml_is_contiguous(dst)) { + for (int i = 0; i < 4; i++) { + if (nr[i] != 1) { + break; + } + if (i > 0) { + collapse_nb(cnb, cne); + collapse_nb(cnb0, cne0); + collapse_nb(cnb1, cne1); + collapse(cne); + collapse(cne0); + collapse(cne1); + } } } { - int64_t ne0 = cne0[0]; - int64_t ne1 = cne0[1]; - int64_t ne2 = cne0[2]; - int64_t ne3 = cne0[3]; + int64_t ne0 = cne[0]; + int64_t ne1 = cne[1]; + int64_t ne2 = cne[2]; + int64_t ne3 = cne[3]; int64_t ne10 = cne1[0]; int64_t ne11 = cne1[1]; int64_t ne12 = cne1[2]; int64_t ne13 = cne1[3]; - size_t nb0 = cnb0[0]; - size_t nb1 = cnb0[1]; - size_t nb2 = cnb0[2]; - size_t nb3 = cnb0[3]; + size_t nb0 = cnb[0]; + size_t nb1 = cnb[1]; + size_t nb2 = cnb[2]; + size_t nb3 = cnb[3]; + + size_t nb00 = cnb0[0]; + size_t nb01 = cnb0[1]; + size_t nb02 = cnb0[2]; + size_t nb03 = cnb0[3]; size_t nb10 = cnb1[0]; size_t nb11 = cnb1[1]; @@ -583,6 +663,28 @@ struct bin_bcast_sycl { size_t s12 = nb12 / sizeof(src1_t); size_t s13 = nb13 / sizeof(src1_t); + size_t s00 = nb00 / sizeof(src0_t); + size_t s01 = nb01 / sizeof(src0_t); + size_t s02 = nb02 / sizeof(src0_t); + size_t s03 = nb03 / sizeof(src0_t); + + GGML_UNUSED(s00); + + GGML_ASSERT(nb0 % sizeof(dst_t) == 0); + GGML_ASSERT(nb1 % sizeof(dst_t) == 0); + GGML_ASSERT(nb2 % sizeof(dst_t) == 0); + GGML_ASSERT(nb3 % sizeof(dst_t) == 0); + + GGML_ASSERT(nb00 % sizeof(src0_t) == 0); + GGML_ASSERT(nb01 % sizeof(src0_t) == 0); + GGML_ASSERT(nb02 % sizeof(src0_t) == 0); + GGML_ASSERT(nb03 % sizeof(src0_t) == 0); + + GGML_ASSERT(nb10 % sizeof(src1_t) == 0); + GGML_ASSERT(nb11 % sizeof(src1_t) == 0); + GGML_ASSERT(nb12 % sizeof(src1_t) == 0); + GGML_ASSERT(nb13 % sizeof(src1_t) == 0); + GGML_ASSERT(s0 == 1); GGML_ASSERT(s10 == 1); @@ -619,8 +721,8 @@ struct bin_bcast_sycl { [=](sycl::nd_item<3> item_ct1) { k_bin_bcast_unravel( src0_dd, src1_dd, dst_dd, ne0, ne1, ne2, ne3, - ne10, ne11, ne12, ne13, s1, s2, s3, s11, s12, - s13, item_ct1); + ne10, ne11, ne12, ne13, s1, s2, s3, s01, s02, + s03, s11, s12, s13, item_ct1); }); } } else { @@ -638,7 +740,7 @@ struct bin_bcast_sycl { [=](sycl::nd_item<3> item_ct1) { k_bin_bcast(src0_dd, src1_dd, dst_dd, ne0, ne1, ne2, ne3, ne10, ne11, ne12, ne13, - s1, s2, s3, s11, s12, s13, + s1, s2, s3, s01, s02, s03, s11, s12, s13, item_ct1); }); } @@ -649,24 +751,22 @@ struct bin_bcast_sycl { template inline void ggml_sycl_op_bin_bcast(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, - const ggml_tensor *src1, ggml_tensor *dst, - const float *src0_dd, const float *src1_dd, - float *dst_dd, - const queue_ptr &main_stream) { + const ggml_tensor *src1, ggml_tensor *dst) { + dpct::queue_ptr main_stream = ctx.stream(); if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) { - op()(ctx, src0, src1, dst, src0_dd, src1_dd, dst_dd, main_stream); + op()(ctx, src0, src1, dst, (const float *)src0->data, (const float *)src1->data, (float *)dst->data, main_stream); } else if (src0->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F16) { - op()(ctx, src0, src1, dst, (const sycl::half *)src0_dd, src1_dd, - (sycl::half *)dst_dd, main_stream); + op()(ctx, src0, src1, dst, (const sycl::half *)src0->data, (const float *)src1->data, + (sycl::half *)dst->data, main_stream); } else if (src0->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F32) { - op()(ctx, src0, src1, dst, (const sycl::half *)src0_dd, src1_dd, dst_dd, + op()(ctx, src0, src1, dst, (const sycl::half *)src0->data, (const float *)src1->data, (float *)dst->data, main_stream); } else if (src0->type == GGML_TYPE_I32 && dst->type == GGML_TYPE_I32) { - op()(ctx, src0, src1, dst, (const int32_t *)src0_dd, (const int32_t *)src1_dd, (int32_t *)dst_dd, + op()(ctx, src0, src1, dst, (const int32_t *)src0->data, (const int32_t *)src1->data, (int32_t *)dst->data, main_stream); } else if (src0->type == GGML_TYPE_I16 && dst->type == GGML_TYPE_I16) { - op()(ctx, src0, src1, dst, (const int16_t *)src0_dd, (const int16_t *)src1_dd, (int16_t *)dst_dd, + op()(ctx, src0, src1, dst, (const int16_t *)src0->data, (const int16_t *)src1->data, (int16_t *)dst->data, main_stream); } else { fprintf(stderr, "%s: unsupported types: dst: %s, src0: %s, src1: %s\n", __func__, @@ -676,9 +776,4 @@ inline void ggml_sycl_op_bin_bcast(ggml_backend_sycl_context & ctx, const ggml_t } bool gpu_has_xmx(sycl::device &dev); - -void ggml_sycl_op_flatten(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, - const ggml_tensor *src1, ggml_tensor *dst, - const ggml_sycl_op_flatten_t op); - #endif // GGML_SYCL_COMMON_HPP diff --git a/ggml/src/ggml-sycl/convert.cpp b/ggml/src/ggml-sycl/convert.cpp index 05b01db2d..76ac6a4dd 100644 --- a/ggml/src/ggml-sycl/convert.cpp +++ b/ggml/src/ggml-sycl/convert.cpp @@ -125,6 +125,25 @@ static void dequantize_row_q4_0_sycl(const void *vx, dst_t *y, const int64_t k, } } +template +static void dequantize_row_q4_0_sycl_reorder(const void *vx, dst_t *y, const int64_t k, + dpct::queue_ptr stream) { + + dpct::has_capability_or_fail(stream->get_device(), + {sycl::aspect::fp16}); + + int constexpr WARP_K = WARP_SIZE * QK4_0; + const int n_warp = (k + WARP_K - 1) / WARP_K; + GGML_ASSERT(k % 2 == 0); + stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, n_warp) * + sycl::range<3>(1, 1, WARP_SIZE), + sycl::range<3>(1, 1, WARP_SIZE)), + [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]]{ + dequantize_block_q4_0_reorder(vx, y, k, item_ct1); + }); + +} + template static void dequantize_row_q4_1_sycl(const void *vx, dst_t *y, const int64_t k, dpct::queue_ptr stream) { @@ -452,10 +471,15 @@ static void convert_unary_sycl(const void *__restrict__ vx, } } -to_fp16_sycl_t ggml_get_to_fp16_sycl(ggml_type type) { +to_fp16_sycl_t ggml_get_to_fp16_sycl(ggml_type type, ggml_tensor *dst) { switch (type) { case GGML_TYPE_Q4_0: - return dequantize_block_sycl; + if (dst->src[0]->extra && + ((ggml_tensor_extra_gpu*)dst->src[0]->extra)->optimized_feature.reorder) { + return dequantize_row_q4_0_sycl_reorder; + } else { + return dequantize_block_sycl; + } case GGML_TYPE_Q4_1: return dequantize_block_sycl; case GGML_TYPE_Q5_0: @@ -499,10 +523,15 @@ to_fp16_sycl_t ggml_get_to_fp16_sycl(ggml_type type) { } } -to_fp32_sycl_t ggml_get_to_fp32_sycl(ggml_type type) { +to_fp32_sycl_t ggml_get_to_fp32_sycl(ggml_type type, ggml_tensor *dst) { switch (type) { case GGML_TYPE_Q4_0: - return dequantize_row_q4_0_sycl; + if (dst->src[0]->extra && + ((ggml_tensor_extra_gpu*)dst->src[0]->extra)->optimized_feature.reorder) { + return dequantize_row_q4_0_sycl_reorder; + } else { + return dequantize_row_q4_0_sycl; + } case GGML_TYPE_Q4_1: return dequantize_row_q4_1_sycl; case GGML_TYPE_Q5_0: diff --git a/ggml/src/ggml-sycl/convert.hpp b/ggml/src/ggml-sycl/convert.hpp index 0ce2874aa..355dae22b 100644 --- a/ggml/src/ggml-sycl/convert.hpp +++ b/ggml/src/ggml-sycl/convert.hpp @@ -21,7 +21,7 @@ using to_t_sycl_t = void (*)(const void *__restrict__ x, T *__restrict__ y, typedef to_t_sycl_t to_fp32_sycl_t; typedef to_t_sycl_t to_fp16_sycl_t; -to_fp16_sycl_t ggml_get_to_fp16_sycl(ggml_type type); -to_fp32_sycl_t ggml_get_to_fp32_sycl(ggml_type type); +to_fp16_sycl_t ggml_get_to_fp16_sycl(ggml_type type, ggml_tensor *dst); +to_fp32_sycl_t ggml_get_to_fp32_sycl(ggml_type type, ggml_tensor *dst); #endif // GGML_SYCL_CONVERT_HPP diff --git a/ggml/src/ggml-sycl/cpy.cpp b/ggml/src/ggml-sycl/cpy.cpp new file mode 100644 index 000000000..5a2314589 --- /dev/null +++ b/ggml/src/ggml-sycl/cpy.cpp @@ -0,0 +1,701 @@ +#include "cpy.hpp" + +#include + +#include "dequantize.hpp" + +static __dpct_inline__ int best_index_int8(int n, const int8_t * val, float x) { + if (x <= val[0]) { + return 0; + } + if (x >= val[n - 1]) { + return n - 1; + } + int ml = 0, mu = n - 1; + while (mu - ml > 1) { + int mav = (ml + mu) / 2; + if (x < val[mav]) { + mu = mav; + } else { + ml = mav; + } + } + return x - val[mu - 1] < val[mu] - x ? mu - 1 : mu; +} + +static void cpy_1_f32_f32(const char * cxi, char * cdsti) { + const float * xi = (const float *) cxi; + float * dsti = (float *) cdsti; + + *dsti = *xi; +} + +static void cpy_1_f32_f16(const char * cxi, char * cdsti) { + const float * xi = (const float *) cxi; + sycl::half * dsti = (sycl::half *) cdsti; + + *dsti = sycl::vec(*xi).convert()[0]; +} + +static void cpy_1_f16_f16(const char * cxi, char * cdsti) { + const sycl::half * xi = (const sycl::half *) cxi; + sycl::half * dsti = (sycl::half *) cdsti; + + *dsti = *xi; +} + +static void cpy_1_f16_f32(const char * cxi, char * cdsti) { + const sycl::half * xi = (const sycl::half *) cxi; + float * dsti = (float *) cdsti; + + *dsti = *xi; +} + +static void cpy_1_i16_i16(const char * cxi, char * cdsti) { + const int16_t * xi = (const int16_t *) cxi; + int16_t * dsti = (int16_t *) cdsti; + + *dsti = *xi; +} + +static void cpy_1_i32_i32(const char * cxi, char * cdsti) { + const int32_t * xi = (const int32_t *) cxi; + int32_t * dsti = (int32_t *) cdsti; + + *dsti = *xi; +} + +template +static void cpy_f32_f16(const char * cx, char * cdst, const int ne, const int ne00, const int ne01, const int ne02, + const int nb00, const int nb01, const int nb02, const int nb03, const int ne10, const int ne11, + const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, + const sycl::nd_item<3> & item_ct1) { + const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) + item_ct1.get_local_id(2); + + if (i >= ne) { + return; + } + + // determine indices i02/i12, i01/i11, i00/i10 as a function of index i of flattened tensor + // then combine those indices with the corresponding byte offsets to get the total offsets + const int i03 = i / (ne00 * ne01 * ne02); + const int i02 = (i - i03 * ne00 * ne01 * ne02) / (ne00 * ne01); + const int i01 = (i - i03 * ne00 * ne01 * ne02 - i02 * ne01 * ne00) / ne00; + const int i00 = i - i03 * ne00 * ne01 * ne02 - i02 * ne01 * ne00 - i01 * ne00; + const int x_offset = i00 * nb00 + i01 * nb01 + i02 * nb02 + i03 * nb03; + + const int i13 = i / (ne10 * ne11 * ne12); + const int i12 = (i - i13 * ne10 * ne11 * ne12) / (ne10 * ne11); + const int i11 = (i - i13 * ne10 * ne11 * ne12 - i12 * ne10 * ne11) / ne10; + const int i10 = i - i13 * ne10 * ne11 * ne12 - i12 * ne10 * ne11 - i11 * ne10; + const int dst_offset = i10 * nb10 + i11 * nb11 + i12 * nb12 + i13 * nb13; + + cpy_1(cx + x_offset, cdst + dst_offset); +} + +static void cpy_blck_f32_q8_0(const char * cxi, char * cdsti) { + const float * xi = (const float *) cxi; + block_q8_0 * dsti = (block_q8_0 *) cdsti; + + float amax = 0.0f; // absolute max + + for (int j = 0; j < QK8_0; j++) { + const float v = xi[j]; + amax = sycl::fmax(amax, sycl::fabs((float) v)); + } + + const float d = amax / ((1 << 7) - 1); + const float id = d ? 1.0f / d : 0.0f; + + dsti->d = d; + + for (int j = 0; j < QK8_0; ++j) { + const float x0 = xi[j] * id; + + dsti->qs[j] = sycl::round((float) x0); + } +} + +static void cpy_blck_q8_0_f32(const char * cxi, char * cdsti) { + float * cdstf = (float *) (cdsti); + + for (int j = 0; j < QK8_0; j += 2) { + dfloat2 dq; + dequantize_q8_0(cxi, 0, j, dq); + *(cdstf + j) = dq.x(); + *(cdstf + j + 1) = dq.y(); + } +} + +static void cpy_blck_f32_q4_0(const char * cxi, char * cdsti) { + const float * xi = (const float *) cxi; + block_q4_0 * dsti = (block_q4_0 *) cdsti; + + float amax = 0.0f; + float vmax = 0.0f; + + for (int j = 0; j < QK4_0; ++j) { + const float v = xi[j]; + if (amax < sycl::fabs((float) v)) { + amax = sycl::fabs((float) v); + vmax = v; + } + } + + const float d = vmax / -8; + const float id = d ? 1.0f / d : 0.0f; + + dsti->d = d; + + for (int j = 0; j < QK4_0 / 2; ++j) { + const float x0 = xi[0 + j] * id; + const float x1 = xi[QK4_0 / 2 + j] * id; + + const uint8_t xi0 = dpct::min(15, (int8_t) (x0 + 8.5f)); + const uint8_t xi1 = dpct::min(15, (int8_t) (x1 + 8.5f)); + + dsti->qs[j] = xi0; + dsti->qs[j] |= xi1 << 4; + } +} + +static void cpy_blck_f32_q4_1(const char * cxi, char * cdsti) { + const float * xi = (const float *) cxi; + block_q4_1 * dsti = (block_q4_1 *) cdsti; + + float vmin = FLT_MAX; + float vmax = -FLT_MAX; + + for (int j = 0; j < QK4_1; ++j) { + const float v = xi[j]; + + if (v < vmin) { + vmin = v; + } + if (v > vmax) { + vmax = v; + } + } + + const float d = (vmax - vmin) / ((1 << 4) - 1); + const float id = d ? 1.0f / d : 0.0f; + + dsti->dm.x() = d; + dsti->dm.y() = vmin; + + for (int j = 0; j < QK4_1 / 2; ++j) { + const float x0 = (xi[0 + j] - vmin) * id; + const float x1 = (xi[QK4_1 / 2 + j] - vmin) * id; + + const uint8_t xi0 = dpct::min(15, (int8_t) (x0 + 0.5f)); + const uint8_t xi1 = dpct::min(15, (int8_t) (x1 + 0.5f)); + + dsti->qs[j] = xi0; + dsti->qs[j] |= xi1 << 4; + } +} + +static void cpy_blck_f32_q5_0(const char * cxi, char * cdsti) { + const float * xi = (const float *) cxi; + block_q5_0 * dsti = (block_q5_0 *) cdsti; + + float amax = 0.0f; + float vmax = 0.0f; + + for (int j = 0; j < QK5_0; ++j) { + const float v = xi[j]; + if (amax < sycl::fabs((float) v)) { + amax = sycl::fabs((float) v); + vmax = v; + } + } + + const float d = vmax / -16; + const float id = d ? 1.0f / d : 0.0f; + + dsti->d = d; + + uint32_t qh = 0; + for (int j = 0; j < QK5_0 / 2; ++j) { + const float x0 = xi[0 + j] * id; + const float x1 = xi[QK5_0 / 2 + j] * id; + + const uint8_t xi0 = dpct::min(31, (int8_t) (x0 + 16.5f)); + const uint8_t xi1 = dpct::min(31, (int8_t) (x1 + 16.5f)); + + dsti->qs[j] = (xi0 & 0xf) | ((xi1 & 0xf) << 4); + qh |= ((xi0 & 0x10u) >> 4) << (j + 0); + qh |= ((xi1 & 0x10u) >> 4) << (j + QK5_0 / 2); + } + memcpy(dsti->qh, &qh, sizeof(qh)); +} + +static void cpy_blck_f32_q5_1(const char * cxi, char * cdsti) { + const float * xi = (const float *) cxi; + block_q5_1 * dsti = (block_q5_1 *) cdsti; + + float min = xi[0]; + float max = xi[0]; + + for (int j = 1; j < QK5_1; ++j) { + const float v = xi[j]; + min = v < min ? v : min; + max = v > max ? v : max; + } + + const float d = (max - min) / 31; + const float id = d ? 1.0f / d : 0.0f; + + dsti->dm.x() = d; + dsti->dm.y() = min; + + uint32_t qh = 0; + for (int j = 0; j < QK5_1 / 2; ++j) { + const float x0 = (xi[0 + j] - min) * id; + const float x1 = (xi[QK5_1 / 2 + j] - min) * id; + + const uint8_t xi0 = (uint8_t) (x0 + 0.5f); + const uint8_t xi1 = (uint8_t) (x1 + 0.5f); + + dsti->qs[j] = (xi0 & 0xf) | ((xi1 & 0xf) << 4); + qh |= ((xi0 & 0x10u) >> 4) << (j + 0); + qh |= ((xi1 & 0x10u) >> 4) << (j + QK5_1 / 2); + } + memcpy(dsti->qh, &qh, sizeof(qh)); +} + +static void cpy_blck_f32_iq4_nl(const char * cxi, char * cdsti) { + const float * xi = (const float *) cxi; + block_iq4_nl * dsti = (block_iq4_nl *) cdsti; + + float amax = 0.0f; + float vmax = 0.0f; + + for (int j = 0; j < QK4_NL; ++j) { + const float v = xi[j]; + if (amax < sycl::fabs((float) v)) { + amax = sycl::fabs((float) v); + vmax = v; + } + } + + float d = vmax / kvalues_iq4nl[0]; + const float id = d ? 1.0f / d : 0.0f; + + float sumqx = 0, sumq2 = 0; + for (int j = 0; j < QK4_NL / 2; ++j) { + const float x0 = xi[0 + j] * id; + const float x1 = xi[QK4_NL / 2 + j] * id; + const uint8_t xi0 = best_index_int8(16, kvalues_iq4nl, x0); + const uint8_t xi1 = best_index_int8(16, kvalues_iq4nl, x1); + dsti->qs[j] = xi0 | (xi1 << 4); + const float v0 = kvalues_iq4nl[xi0]; + const float v1 = kvalues_iq4nl[xi1]; + const float w0 = xi[0 + j] * xi[0 + j]; + const float w1 = xi[QK4_NL / 2 + j] * xi[QK4_NL / 2 + j]; + sumqx += w0 * v0 * xi[j] + w1 * v1 * xi[QK4_NL / 2 + j]; + sumq2 += w0 * v0 * v0 + w1 * v1 * v1; + } + + dsti->d = sumq2 > 0 ? sumqx / sumq2 : d; +} + +template static void cpy_blck_q_f32(const char * cxi, char * cdsti) { + float * cdstf = (float *) (cdsti); + + for (int j = 0; j < qk / 2; j++) { + dfloat2 dq; + dequant(cxi, 0, j, dq); + *(cdstf + j) = dq.x(); + *(cdstf + j + qk / 2) = dq.y(); + } +} + +template +static void cpy_f32_q(const char * cx, char * cdst, const int ne, const int ne00, const int ne01, const int ne02, + const int nb00, const int nb01, const int nb02, const int nb03, const int ne10, const int ne11, + const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, + const sycl::nd_item<3> & item_ct1) { + const int i = (item_ct1.get_local_range(2) * item_ct1.get_group(2) + item_ct1.get_local_id(2)) * qk; + + if (i >= ne) { + return; + } + + const int i03 = i / (ne00 * ne01 * ne02); + const int i02 = (i - i03 * ne00 * ne01 * ne02) / (ne00 * ne01); + const int i01 = (i - i03 * ne00 * ne01 * ne02 - i02 * ne01 * ne00) / ne00; + const int i00 = i - i03 * ne00 * ne01 * ne02 - i02 * ne01 * ne00 - i01 * ne00; + const int x_offset = i00 * nb00 + i01 * nb01 + i02 * nb02 + i03 * nb03; + + const int i13 = i / (ne10 * ne11 * ne12); + const int i12 = (i - i13 * ne10 * ne11 * ne12) / (ne10 * ne11); + const int i11 = (i - i13 * ne10 * ne11 * ne12 - i12 * ne10 * ne11) / ne10; + const int i10 = i - i13 * ne10 * ne11 * ne12 - i12 * ne10 * ne11 - i11 * ne10; + const int dst_offset = (i10 / qk) * nb10 + i11 * nb11 + i12 * nb12 + i13 * nb13; + + cpy_blck(cx + x_offset, cdst + dst_offset); +} + +template +static void cpy_q_f32(const char * cx, char * cdst, const int ne, const int ne00, const int ne01, const int ne02, + const int nb00, const int nb01, const int nb02, const int nb03, const int ne10, const int ne11, + const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, + const sycl::nd_item<3> & item_ct1) { + const int i = (item_ct1.get_local_range(2) * item_ct1.get_group(2) + item_ct1.get_local_id(2)) * qk; + + if (i >= ne) { + return; + } + + const int i03 = i / (ne00 * ne01 * ne02); + const int i02 = (i - i03 * ne00 * ne01 * ne02) / (ne00 * ne01); + const int i01 = (i - i03 * ne00 * ne01 * ne02 - i02 * ne01 * ne00) / ne00; + const int i00 = i - i03 * ne00 * ne01 * ne02 - i02 * ne01 * ne00 - i01 * ne00; + const int x_offset = (i00 / qk) * nb00 + i01 * nb01 + i02 * nb02 + i03 * nb03; + + const int i13 = i / (ne10 * ne11 * ne12); + const int i12 = (i - i13 * ne10 * ne11 * ne12) / (ne10 * ne11); + const int i11 = (i - i13 * ne10 * ne11 * ne12 - i12 * ne10 * ne11) / ne10; + const int i10 = i - i13 * ne10 * ne11 * ne12 - i12 * ne10 * ne11 - i11 * ne10; + const int dst_offset = i10 * nb10 + i11 * nb11 + i12 * nb12 + i13 * nb13; + + cpy_blck(cx + x_offset, cdst + dst_offset); +} + +static void ggml_cpy_f16_f32_sycl(const char * cx, char * cdst, const int ne, const int ne00, const int ne01, + const int ne02, const int nb00, const int nb01, const int nb02, const int nb03, + const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, + const int nb12, const int nb13, queue_ptr stream) { + const int num_blocks = (ne + SYCL_CPY_BLOCK_SIZE - 1) / SYCL_CPY_BLOCK_SIZE; + { + dpct::has_capability_or_fail(stream->get_device(), { sycl::aspect::fp16 }); + + stream->parallel_for( + sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), + sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), + [=](sycl::nd_item<3> item_ct1) { + cpy_f32_f16(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, + nb10, nb11, nb12, nb13, item_ct1); + }); + } +} + +static void ggml_cpy_f32_f32_sycl(const char * cx, char * cdst, const int ne, const int ne00, const int ne01, + const int ne02, const int nb00, const int nb01, const int nb02, const int nb03, + const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, + const int nb12, const int nb13, queue_ptr stream) { + const int num_blocks = (ne + SYCL_CPY_BLOCK_SIZE - 1) / SYCL_CPY_BLOCK_SIZE; + { + dpct::has_capability_or_fail(stream->get_device(), { sycl::aspect::fp16 }); + + stream->parallel_for( + sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), + sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), + [=](sycl::nd_item<3> item_ct1) { + cpy_f32_f16(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, + nb10, nb11, nb12, nb13, item_ct1); + }); + } +} + +static void ggml_cpy_f32_f16_sycl(const char * cx, char * cdst, const int ne, const int ne00, const int ne01, + const int ne02, const int nb00, const int nb01, const int nb02, const int nb03, + const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, + const int nb12, const int nb13, queue_ptr stream) { + const int num_blocks = (ne + SYCL_CPY_BLOCK_SIZE - 1) / SYCL_CPY_BLOCK_SIZE; + { + dpct::has_capability_or_fail(stream->get_device(), { sycl::aspect::fp16 }); + + stream->parallel_for( + sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), + sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), + [=](sycl::nd_item<3> item_ct1) { + cpy_f32_f16(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, + nb10, nb11, nb12, nb13, item_ct1); + }); + } +} + +static void ggml_cpy_f32_q8_0_sycl(const char * cx, char * cdst, const int ne, const int ne00, const int ne01, + const int ne02, const int nb00, const int nb01, const int nb02, const int nb03, + const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, + const int nb12, const int nb13, queue_ptr stream) { + GGML_ASSERT(ne % QK8_0 == 0); + const int num_blocks = ne / QK8_0; + stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)), + [=](sycl::nd_item<3> item_ct1) { + cpy_f32_q(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, + ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1); + }); +} + +static void ggml_cpy_q8_0_f32_sycl(const char * cx, char * cdst, const int ne, const int ne00, const int ne01, + const int ne02, const int nb00, const int nb01, const int nb02, const int nb03, + const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, + const int nb12, const int nb13, queue_ptr stream) { + const int num_blocks = ne; + stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)), + [=](sycl::nd_item<3> item_ct1) { + cpy_q_f32(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, + ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1); + }); +} + +static void ggml_cpy_f32_q4_0_sycl(const char * cx, char * cdst, const int ne, const int ne00, const int ne01, + const int ne02, const int nb00, const int nb01, const int nb02, const int nb03, + const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, + const int nb12, const int nb13, queue_ptr stream) { + GGML_ASSERT(ne % QK4_0 == 0); + const int num_blocks = ne / QK4_0; + stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)), + [=](sycl::nd_item<3> item_ct1) { + cpy_f32_q(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, + ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1); + }); +} + +static void ggml_cpy_q4_0_f32_sycl(const char * cx, char * cdst, const int ne, const int ne00, const int ne01, + const int ne02, const int nb00, const int nb01, const int nb02, const int nb03, + const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, + const int nb12, const int nb13, queue_ptr stream) { + const int num_blocks = ne; + stream->parallel_for( + sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)), [=](sycl::nd_item<3> item_ct1) { + cpy_q_f32, QK4_0>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, + nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, + item_ct1); + }); +} + +static void ggml_cpy_f32_q4_1_sycl(const char * cx, char * cdst, const int ne, const int ne00, const int ne01, + const int ne02, const int nb00, const int nb01, const int nb02, const int nb03, + const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, + const int nb12, const int nb13, queue_ptr stream) { + GGML_ASSERT(ne % QK4_1 == 0); + const int num_blocks = ne / QK4_1; + stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)), + [=](sycl::nd_item<3> item_ct1) { + cpy_f32_q(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, + ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1); + }); +} + +static void ggml_cpy_q4_1_f32_sycl(const char * cx, char * cdst, const int ne, const int ne00, const int ne01, + const int ne02, const int nb00, const int nb01, const int nb02, const int nb03, + const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, + const int nb12, const int nb13, queue_ptr stream) { + const int num_blocks = ne; + stream->parallel_for( + sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)), [=](sycl::nd_item<3> item_ct1) { + cpy_q_f32, QK4_1>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, + nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, + item_ct1); + }); +} + +static void ggml_cpy_f32_q5_0_sycl(const char * cx, char * cdst, const int ne, const int ne00, const int ne01, + const int ne02, const int nb00, const int nb01, const int nb02, const int nb03, + const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, + const int nb12, const int nb13, queue_ptr stream) { + GGML_ASSERT(ne % QK5_0 == 0); + const int num_blocks = ne / QK5_0; + stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)), + [=](sycl::nd_item<3> item_ct1) { + cpy_f32_q(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, + ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1); + }); +} + +static void ggml_cpy_q5_0_f32_sycl(const char * cx, char * cdst, const int ne, const int ne00, const int ne01, + const int ne02, const int nb00, const int nb01, const int nb02, const int nb03, + const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, + const int nb12, const int nb13, queue_ptr stream) { + const int num_blocks = ne; + stream->parallel_for( + sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)), [=](sycl::nd_item<3> item_ct1) { + cpy_q_f32, QK5_0>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, + nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, + item_ct1); + }); +} + +static void ggml_cpy_f32_q5_1_sycl(const char * cx, char * cdst, const int ne, const int ne00, const int ne01, + const int ne02, const int nb00, const int nb01, const int nb02, const int nb03, + const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, + const int nb12, const int nb13, queue_ptr stream) { + GGML_ASSERT(ne % QK5_1 == 0); + const int num_blocks = ne / QK5_1; + stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)), + [=](sycl::nd_item<3> item_ct1) { + cpy_f32_q(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, + ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1); + }); +} + +static void ggml_cpy_q5_1_f32_sycl(const char * cx, char * cdst, const int ne, const int ne00, const int ne01, + const int ne02, const int nb00, const int nb01, const int nb02, const int nb03, + const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, + const int nb12, const int nb13, queue_ptr stream) { + const int num_blocks = ne; + stream->parallel_for( + sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)), [=](sycl::nd_item<3> item_ct1) { + cpy_q_f32, QK5_1>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, + nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, + item_ct1); + }); +} + +static void ggml_cpy_f32_iq4_nl_sycl(const char * cx, char * cdst, const int ne, const int ne00, const int ne01, + const int ne02, const int nb00, const int nb01, const int nb02, const int nb03, + const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, + const int nb12, const int nb13, queue_ptr stream) { + GGML_ASSERT(ne % QK4_NL == 0); + const int num_blocks = ne / QK4_NL; + stream->parallel_for( + sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)), [=](sycl::nd_item<3> item_ct1) { + cpy_f32_q(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, + ne12, nb10, nb11, nb12, nb13, item_ct1); + }); +} + +static void ggml_cpy_f16_f16_sycl(const char * cx, char * cdst, const int ne, const int ne00, const int ne01, + const int ne02, const int nb00, const int nb01, const int nb02, const int nb03, + const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, + const int nb12, const int nb13, queue_ptr stream) { + const int num_blocks = (ne + SYCL_CPY_BLOCK_SIZE - 1) / SYCL_CPY_BLOCK_SIZE; + { + dpct::has_capability_or_fail(stream->get_device(), { sycl::aspect::fp16 }); + + stream->parallel_for( + sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), + sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), + [=](sycl::nd_item<3> item_ct1) { + cpy_f32_f16(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, + nb10, nb11, nb12, nb13, item_ct1); + }); + } +} + +static void ggml_cpy_i16_i16_sycl(const char * cx, char * cdst, const int ne, const int ne00, const int ne01, + const int ne02, const int nb00, const int nb01, const int nb02, const int nb03, + const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, + const int nb12, const int nb13, queue_ptr stream) { + const int num_blocks = (ne + SYCL_CPY_BLOCK_SIZE - 1) / SYCL_CPY_BLOCK_SIZE; + { + // dpct::has_capability_or_fail(stream->get_device(), + // {sycl::aspect::fp16}); + + stream->parallel_for( + sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), + sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), + [=](sycl::nd_item<3> item_ct1) { + cpy_f32_f16(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, + nb10, nb11, nb12, nb13, item_ct1); + }); + } +} + +static void ggml_cpy_i32_i32_sycl(const char * cx, char * cdst, const int ne, const int ne00, const int ne01, + const int ne02, const int nb00, const int nb01, const int nb02, const int nb03, + const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, + const int nb12, const int nb13, queue_ptr stream) { + const int num_blocks = (ne + SYCL_CPY_BLOCK_SIZE - 1) / SYCL_CPY_BLOCK_SIZE; + { + // dpct::has_capability_or_fail(stream->get_device(), + // {sycl::aspect::fp16}); + + stream->parallel_for( + sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), + sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), + [=](sycl::nd_item<3> item_ct1) { + cpy_f32_f16(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, + nb10, nb11, nb12, nb13, item_ct1); + }); + } +} + +void ggml_sycl_cpy(ggml_backend_sycl_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1) try { + const int64_t ne = ggml_nelements(src0); + GGML_ASSERT(ne == ggml_nelements(src1)); + + GGML_ASSERT(ggml_nbytes(src0) <= INT_MAX); + GGML_ASSERT(ggml_nbytes(src1) <= INT_MAX); + + GGML_TENSOR_BINARY_OP_LOCALS01; + + SYCL_CHECK(ggml_sycl_set_device(ctx.device)); + queue_ptr main_stream = ctx.stream(); + + char * src0_ddc = (char *) src0->data; + char * src1_ddc = (char *) src1->data; + GGML_SYCL_DEBUG("[SYCL] %s: Tensor supplied: %s to %s\n", __func__, ggml_type_name(src0->type), + ggml_type_name(src1->type)); + + if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32) { + ggml_cpy_f32_f32_sycl(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, + nb11, nb12, nb13, main_stream); + } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F16) { + ggml_cpy_f32_f16_sycl(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, + nb11, nb12, nb13, main_stream); + } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q8_0) { + ggml_cpy_f32_q8_0_sycl(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, + nb11, nb12, nb13, main_stream); + } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q4_0) { + ggml_cpy_f32_q4_0_sycl(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, + nb11, nb12, nb13, main_stream); + } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q4_1) { + ggml_cpy_f32_q4_1_sycl(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, + nb11, nb12, nb13, main_stream); + } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F32) { + ggml_cpy_f16_f32_sycl(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, + nb11, nb12, nb13, main_stream); + } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F16) { + ggml_cpy_f16_f16_sycl(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, + nb11, nb12, nb13, main_stream); + } else if (src0->type == GGML_TYPE_I16 && src1->type == GGML_TYPE_I16) { + ggml_cpy_i16_i16_sycl(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, + nb11, nb12, nb13, main_stream); + } else if (src0->type == GGML_TYPE_I32 && src1->type == GGML_TYPE_I32) { + ggml_cpy_i32_i32_sycl(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, + nb11, nb12, nb13, main_stream); + } else if (src0->type == GGML_TYPE_Q4_0 && src1->type == GGML_TYPE_F32) { + ggml_cpy_q4_0_f32_sycl(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, + nb11, nb12, nb13, main_stream); + } else if (src0->type == GGML_TYPE_Q4_1 && src1->type == GGML_TYPE_F32) { + ggml_cpy_q4_1_f32_sycl(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, + nb11, nb12, nb13, main_stream); + } else if (src0->type == GGML_TYPE_Q8_0 && src1->type == GGML_TYPE_F32) { + ggml_cpy_q8_0_f32_sycl(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, + nb11, nb12, nb13, main_stream); + } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q5_0) { + ggml_cpy_f32_q5_0_sycl(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, + nb11, nb12, nb13, main_stream); + } else if (src0->type == GGML_TYPE_Q5_0 && src1->type == GGML_TYPE_F32) { + ggml_cpy_q5_0_f32_sycl(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, + nb11, nb12, nb13, main_stream); + } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q5_1) { + ggml_cpy_f32_q5_1_sycl(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, + nb11, nb12, nb13, main_stream); + } else if (src0->type == GGML_TYPE_Q5_1 && src1->type == GGML_TYPE_F32) { + ggml_cpy_q5_1_f32_sycl(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, + nb11, nb12, nb13, main_stream); + } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_IQ4_NL) { + ggml_cpy_f32_iq4_nl_sycl(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, + nb10, nb11, nb12, nb13, main_stream); + } else { + GGML_LOG_ERROR("%s: unsupported type combination (%s to %s)\n", __func__, ggml_type_name(src0->type), + ggml_type_name(src1->type)); + GGML_ABORT("fatal error"); + } +} catch (const sycl::exception & exc) { + std::cerr << exc.what() << "Exception caught at file:" << __FILE__ << ", line:" << __LINE__ << std::endl; + std::exit(1); +} + +void ggml_sycl_dup(ggml_backend_sycl_context & ctx, ggml_tensor * dst) { + // TODO: why do we pass dst as src1 here? + GGML_SYCL_DEBUG("[SYCL] call %s\n", __func__); + ggml_sycl_cpy(ctx, dst->src[0], dst); + GGML_SYCL_DEBUG("[SYCL] call %s done\n", __func__); +} diff --git a/ggml/src/ggml-sycl/cpy.hpp b/ggml/src/ggml-sycl/cpy.hpp new file mode 100644 index 000000000..0a0f561d2 --- /dev/null +++ b/ggml/src/ggml-sycl/cpy.hpp @@ -0,0 +1,11 @@ +#ifndef GGML_SYCL_CPY_HPP +#define GGML_SYCL_CPY_HPP + +#include "common.hpp" + +typedef void (*cpy_kernel_t)(const char * cx, char * cdst); + +void ggml_sycl_cpy(ggml_backend_sycl_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1); +void ggml_sycl_dup(ggml_backend_sycl_context & ctx, ggml_tensor * dst); + +#endif // GGML_SYCL_CPY_HPP diff --git a/ggml/src/ggml-sycl/dequantize.hpp b/ggml/src/ggml-sycl/dequantize.hpp index b8304c3a2..651c2160d 100644 --- a/ggml/src/ggml-sycl/dequantize.hpp +++ b/ggml/src/ggml-sycl/dequantize.hpp @@ -16,6 +16,8 @@ #include "common.hpp" typedef void (*dequantize_kernel_t)(const void * vx, const int64_t ib, const int iqs, dfloat2 & v); +typedef void (*dequantize_kernel_t_reorder)(const void *d, const int64_t ib, const void *qs, + const int iqs, dfloat2 &v); static __dpct_inline__ void dequantize_q4_0(const void *vx, const int64_t ib, const int iqs, dfloat2 &v) { @@ -40,6 +42,29 @@ static __dpct_inline__ void dequantize_q4_0(const void *vx, const int64_t ib, #endif // GGML_SYCL_F16 } +static __dpct_inline__ void dequantize_q4_0_reorder(const void *d_ptr, const int64_t ib, const void *qs, + const int iqs, dfloat2 &v) { + // const block_q4_0 * x = (const block_q4_0 *) vx; + + const dfloat d = (const dfloat)*((const sycl::half*)d_ptr+ib); + + const int vui = *((const uint8_t *)qs+iqs); + + v.x() = vui & 0xF; + v.y() = vui >> 4; + +#ifdef GGML_SYCL_F16 + // v = v - {8.0f, 8.0f}; + // v = v * {d, d}; + v.s0() = (v.s0() - 8.0f) * d; + v.s1() = (v.s1() - 8.0f) * d; + +#else + v.x() = (v.x() - 8.0f) * d; + v.y() = (v.y() - 8.0f) * d; +#endif // GGML_SYCL_F16 +} + static __dpct_inline__ void dequantize_q4_1(const void *vx, const int64_t ib, const int iqs, dfloat2 &v) { const block_q4_1 * x = (const block_q4_1 *) vx; @@ -167,6 +192,36 @@ static void dequantize_block_q4_0(const void * __restrict__ vx, dst_t * __restri } } +template +static void dequantize_block_q4_0_reorder(const void * __restrict__ vx, dst_t * __restrict__ yy, int64_t nb32, + const sycl::nd_item<3> &item_ct1) { + + const int64_t i = item_ct1.get_group(2); + auto k=nb32; + // assume 32 threads + const int64_t tid = item_ct1.get_local_id(2); + const int lane_ib = i * WARP_SIZE + tid; + + if (lane_ib >= k / QK4_0) { + return; + } + + dst_t * y_ptr = yy + lane_ib * QK4_0; + + auto qs = (const uint8_t*)vx + lane_ib * QK4_0 / 2; + auto s_ptr = (const sycl::half*)((const uint8_t*)vx + k / 2) + lane_ib; + + const float d = float(*s_ptr); + +#pragma unroll + for (int l = 0; l < QK4_0 / 2; ++l) { + int vq = qs[l]; + y_ptr[l + 0] = d * ((vq & 0xF) - 8); + y_ptr[l + 16] = d * ((vq >> 4) - 8); + } + +} + template static void dequantize_block_q4_1(const void * __restrict__ vx, dst_t * __restrict__ yy, int64_t nb32, const sycl::nd_item<3> &item_ct1) { diff --git a/ggml/src/ggml-sycl/dmmv.cpp b/ggml/src/ggml-sycl/dmmv.cpp index 0d097357c..04a85fa35 100644 --- a/ggml/src/ggml-sycl/dmmv.cpp +++ b/ggml/src/ggml-sycl/dmmv.cpp @@ -3,7 +3,6 @@ #include "dequantize.hpp" #include "presets.hpp" - static void convert_f16(const void * vx, const int64_t ib, const int iqs, dfloat2 & v){ const sycl::half *x = (const sycl::half *)vx; @@ -91,6 +90,112 @@ static void dequantize_mul_mat_vec(const void * __restrict__ vx, const dfloat * } } +template +static void dequantize_mul_mat_vec_reorder(const void * __restrict__ vx, const dfloat * __restrict__ y, float * __restrict__ dst, const int ncols, const int nrows, + const sycl::nd_item<3> &item_ct1) { + // qk = quantized weights per x block + // qr = number of quantized weights per data value in x block + const int row = item_ct1.get_group(2) * item_ct1.get_local_range(1) + + item_ct1.get_local_id(1); + + if (row >= nrows) { + return; + } + + const int tid = item_ct1.get_local_id(2); + + + const int ncols_left = ncols % (QK4_0*WARP_SIZE); + const int ncols_align = ncols - ncols_left; + const int iter_stride = 8*2*GGML_SYCL_DMMV_X; + const int vals_per_iter = iter_stride / WARP_SIZE; // num quantized vals per thread and i iter //64/16=4, 512/16/2= 16 + const int y_offset = qr == 1 ? 1 : qk/2; + +// partial sum for each thread +#ifdef GGML_SYCL_F16 + sycl::half2 tmp = {0.0f, 0.0f}; // two sums for f16 to take advantage of half2 intrinsics +#else + float tmp = 0.0f; +#endif // GGML_SYCL_F16 + const char *d_ptr = (const char*)vx+ncols*nrows/2; + int i=0; + for (i = 0; i < ncols_align; i += iter_stride) { + const int col = i + vals_per_iter*tid; + const int ib = (row*ncols + col)/qk; // x block index + const int iqs = (col%qk)/qr; // x quant index + const int iybs = col - col%qk; // y block start index + +// processing >2 values per i iter is faster for fast GPUs +#pragma unroll + for (int j = 0; j < vals_per_iter; j += 2) { + // process 2 vals per j iter + + // dequantize + // for qr = 2 the iqs needs to increase by 1 per j iter because 2 weights per data val + dfloat2 v; + dequantize_kernel_reorder((const void *)d_ptr, ib, (const void *)vx, ib * QK4_0 / 2 +iqs+j/qr, v); + + // matrix multiplication + // for qr = 2 the y index needs to increase by 1 per j iter because of y_offset = qk/2 +#ifdef GGML_SYCL_F16 + dfloat2 t1{y[iybs + iqs + j / qr + 0], + y[iybs + iqs + j / qr + y_offset]}; + + tmp += v * t1; +#else + tmp += v.x() * y[iybs + iqs + j / qr + 0]; + tmp += v.y() * y[iybs + iqs + j / qr + y_offset]; +#endif // GGML_SYCL_F16 + } + } + + for (; i < ncols; i += iter_stride) { + if (tid>=ncols_left/QK4_0) continue; + const int col = i + vals_per_iter*tid; + const int ib = (row*ncols + col)/qk; // x block index + const int iqs = (col%qk)/qr; // x quant index + const int iybs = col - col%qk; // y block start index + +// processing >2 values per i iter is faster for fast GPUs +#pragma unroll + for (int j = 0; j < vals_per_iter; j += 2) { + // process 2 vals per j iter + + // dequantize + // for qr = 2 the iqs needs to increase by 1 per j iter because 2 weights per data val + dfloat2 v; + dequantize_kernel_reorder((const void *)d_ptr, ib, (const void *)vx, ib * QK4_0 / 2 +iqs+j/qr, v); + + // matrix multiplication + // for qr = 2 the y index needs to increase by 1 per j iter because of y_offset = qk/2 +#ifdef GGML_SYCL_F16 + dfloat2 t1{y[iybs + iqs + j / qr + 0], + y[iybs + iqs + j / qr + y_offset]}; + + tmp += v * t1; +#else + tmp += v.x() * y[iybs + iqs + j / qr + 0]; + tmp += v.y() * y[iybs + iqs + j / qr + y_offset]; +#endif // GGML_SYCL_F16 + } + } + + // sum up partial sums and write back result + const int mask_start = ncols > GGML_SYCL_DMMV_X ? WARP_SIZE >> 1 : WARP_SIZE >> 2; + for (int mask = mask_start; mask > 0; mask >>= 1) { + tmp += + dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); + } + + if (tid == 0) { +#ifdef GGML_SYCL_F16 + dst[row] = tmp.x() + tmp.y(); +#else + dst[row] = tmp; +#endif // GGML_SYCL_F16 + } +} + static void convert_mul_mat_vec_f16_sycl(const void *vx, const dfloat *y, float *dst, const int ncols, const int nrows, @@ -105,7 +210,7 @@ static void convert_mul_mat_vec_f16_sycl(const void *vx, const dfloat *y, stream->parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(WARP_SIZE)]] { + [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { dequantize_mul_mat_vec<1, 1, convert_f16>(vx, y, dst, ncols, nrows, item_ct1); }); @@ -759,6 +864,28 @@ static void dequantize_mul_mat_vec_q6_k(const void * __restrict__ vx, const floa } } +static void dequantize_mul_mat_vec_q4_0_sycl_reorder(const void *vx, const dfloat *y, + float *dst, const int ncols, + const int nrows, + dpct::queue_ptr stream) { + GGML_ASSERT(ncols % GGML_SYCL_DMMV_X == 0); + const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; + // the number of rows may exceed maximum grid size in the y or z dimensions, use the x dimension instead + const sycl::range<3> block_nums(1, 1, block_num_y); + const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); + { + dpct::has_capability_or_fail(stream->get_device(), + {sycl::aspect::fp16}); + + stream->parallel_for( + sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { + dequantize_mul_mat_vec_reorder( + vx, y, dst, ncols, nrows, item_ct1); + }); + } +} + static void dequantize_mul_mat_vec_q4_0_sycl(const void *vx, const dfloat *y, float *dst, const int ncols, @@ -775,7 +902,7 @@ static void dequantize_mul_mat_vec_q4_0_sycl(const void *vx, const dfloat *y, stream->parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(WARP_SIZE)]] { + [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { dequantize_mul_mat_vec( vx, y, dst, ncols, nrows, item_ct1); }); @@ -796,7 +923,7 @@ static void dequantize_mul_mat_vec_q4_1_sycl(const void *vx, const dfloat *y, stream->parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(WARP_SIZE)]] { + [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { dequantize_mul_mat_vec( vx, y, dst, ncols, nrows, item_ct1); }); @@ -817,7 +944,7 @@ static void dequantize_mul_mat_vec_q5_0_sycl(const void *vx, const dfloat *y, stream->parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(WARP_SIZE)]] { + [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { dequantize_mul_mat_vec( vx, y, dst, ncols, nrows, item_ct1); }); @@ -838,7 +965,7 @@ static void dequantize_mul_mat_vec_q5_1_sycl(const void *vx, const dfloat *y, stream->parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(WARP_SIZE)]] { + [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { dequantize_mul_mat_vec( vx, y, dst, ncols, nrows, item_ct1); }); @@ -859,7 +986,7 @@ static void dequantize_mul_mat_vec_q8_0_sycl(const void *vx, const dfloat *y, stream->parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(WARP_SIZE)]] { + [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { dequantize_mul_mat_vec( vx, y, dst, ncols, nrows, item_ct1); }); @@ -877,7 +1004,7 @@ static void dequantize_mul_mat_vec_q2_K_sycl(const void *vx, const float *y, const sycl::range<3> block_dims(1, ny, QK_WARP_SIZE); stream->parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(QK_WARP_SIZE)]] { + [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(QK_WARP_SIZE)]] { dequantize_mul_mat_vec_q2_k(vx, y, dst, ncols, nrows, item_ct1); }); } @@ -893,7 +1020,7 @@ static void dequantize_mul_mat_vec_q3_K_sycl(const void *vx, const float *y, const sycl::range<3> block_dims(1, ny, QK_WARP_SIZE); stream->parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(QK_WARP_SIZE)]] { + [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(QK_WARP_SIZE)]] { dequantize_mul_mat_vec_q3_k(vx, y, dst, ncols, nrows, item_ct1); }); } @@ -909,7 +1036,7 @@ static void dequantize_mul_mat_vec_q4_K_sycl(const void *vx, const float *y, const sycl::range<3> block_dims(1, ny, QK_WARP_SIZE); stream->parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(QK_WARP_SIZE)]] { + [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(QK_WARP_SIZE)]] { dequantize_mul_mat_vec_q4_k(vx, y, dst, ncols, nrows, item_ct1); }); } @@ -922,7 +1049,7 @@ static void dequantize_mul_mat_vec_q5_K_sycl(const void *vx, const float *y, const sycl::range<3> block_dims(1, 1, QK_WARP_SIZE); stream->parallel_for( sycl::nd_range<3>(sycl::range<3>(1, 1, nrows) * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(QK_WARP_SIZE)]] { + [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(QK_WARP_SIZE)]] { dequantize_mul_mat_vec_q5_k(vx, y, dst, ncols, item_ct1); }); } @@ -938,7 +1065,7 @@ static void dequantize_mul_mat_vec_q6_K_sycl(const void *vx, const float *y, const sycl::range<3> block_dims(1, ny, QK_WARP_SIZE); stream->parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(QK_WARP_SIZE)]] { + [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(QK_WARP_SIZE)]] { dequantize_mul_mat_vec_q6_k(vx, y, dst, ncols, nrows, item_ct1); }); } @@ -953,7 +1080,6 @@ void ggml_sycl_op_dequantize_mul_mat_vec( const int64_t ne00 = src0->ne[0]; const int64_t row_diff = row_high - row_low; - GGML_ASSERT(src1->type == GGML_TYPE_F32); // on some GPUs it is faster to convert src1 to half and to use half precision intrinsics #ifdef GGML_SYCL_F16 @@ -967,7 +1093,7 @@ void ggml_sycl_op_dequantize_mul_mat_vec( if (src1_convert_f16) { src1_dfloat = src1_dfloat_a.alloc(ne00); - const to_fp16_sycl_t to_fp16_sycl = ggml_get_to_fp16_sycl(src1->type); + const to_fp16_sycl_t to_fp16_sycl = ggml_get_to_fp16_sycl(src1->type, dst); GGML_ASSERT(to_fp16_sycl != nullptr); to_fp16_sycl(src1_ddf_i, src1_dfloat, ne00, stream); } @@ -977,7 +1103,12 @@ void ggml_sycl_op_dequantize_mul_mat_vec( switch (src0->type) { case GGML_TYPE_Q4_0: - dequantize_mul_mat_vec_q4_0_sycl(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream); + if ((ggml_tensor_extra_gpu*)dst->src[0]->extra && + ((ggml_tensor_extra_gpu*)dst->src[0]->extra)->optimized_feature.reorder) { + dequantize_mul_mat_vec_q4_0_sycl_reorder(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream); + } else { + dequantize_mul_mat_vec_q4_0_sycl(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream); + } break; case GGML_TYPE_Q4_1: dequantize_mul_mat_vec_q4_1_sycl(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream); @@ -1012,7 +1143,6 @@ void ggml_sycl_op_dequantize_mul_mat_vec( default: printf("ggml_sycl_op_dequantize_mul_mat_vec unsupported GGML_TYPE %d\n", src0->type); GGML_ABORT("fatal error"); - break; } GGML_UNUSED(src1); @@ -1020,4 +1150,5 @@ void ggml_sycl_op_dequantize_mul_mat_vec( GGML_UNUSED(src1_ddq_i); GGML_UNUSED(src1_ncols); GGML_UNUSED(src1_padded_row_size); + GGML_UNUSED(ctx); } diff --git a/ggml/src/ggml-sycl/dpct/helper.hpp b/ggml/src/ggml-sycl/dpct/helper.hpp index c96395be6..d538965b0 100644 --- a/ggml/src/ggml-sycl/dpct/helper.hpp +++ b/ggml/src/ggml-sycl/dpct/helper.hpp @@ -16,9 +16,18 @@ #include #include #include -#include #include +#ifdef GGML_SYCL_USE_INTEL_ONEMKL +#include +// Allow to use the same namespace for Intel oneMKL and oneMath +namespace oneapi { + namespace math = mkl; +} +#else +#include +#endif + #include "ggml.h" #if defined(__linux__) @@ -83,13 +92,32 @@ inline std::string get_device_backend_and_type(const sycl::device &device) { } template struct matrix_info_t { - oneapi::mkl::transpose transpose_info[2]; + oneapi::math::transpose transpose_info[2]; Ts value_info[2]; std::int64_t size_info[3]; std::int64_t ld_info[3]; std::int64_t groupsize_info; }; +inline auto get_onemath_backend(sycl::queue& queue) +#if defined(GGML_SYCL_GENERIC) || defined(GGML_SYCL_USE_INTEL_ONEMKL) + -> sycl::queue& +#endif +{ +// If the backend is known at compile-time, use oneMath backend_selector to use +// compile-time dispatching and avoid the need to dlopen libraries. Otherwise +// fallback to runtime dispatching. +#if defined(GGML_SYCL_NVIDIA) + return oneapi::math::backend_selector{ queue }; +#elif defined(GGML_SYCL_AMD) + return oneapi::math::backend_selector{ queue }; +#elif defined(GGML_SYCL_GENERIC) || defined(GGML_SYCL_USE_INTEL_ONEMKL) + return queue; +#else + static_assert(false, "Unsupported backend"); +#endif +} + namespace dpct { typedef sycl::queue *queue_ptr; @@ -1686,26 +1714,18 @@ namespace dpct namespace detail { - template - inline void gemm_impl(sycl::queue &q, oneapi::mkl::transpose a_trans, - oneapi::mkl::transpose b_trans, int m, int n, int k, - const void *alpha, const void *a, int lda, const void *b, - int ldb, const void *beta, void *c, int ldc) - { - Ts alpha_value = dpct::get_value(reinterpret_cast(alpha), q); - Ts beta_value = dpct::get_value(reinterpret_cast(beta), q); - auto data_a = get_memory(a); - auto data_b = get_memory(b); - auto data_c = get_memory(c); -#ifdef GGML_SYCL_NVIDIA - oneapi::mkl::blas::column_major::gemm(oneapi::mkl::backend_selector{ q }, - a_trans, b_trans, m, n, k, alpha_value, data_a, lda, data_b, ldb, - beta_value, data_c, ldc); -#else - oneapi::mkl::blas::column_major::gemm(q, a_trans, b_trans, m, n, k, alpha_value, data_a, lda, data_b, ldb, - beta_value, data_c, ldc); -#endif - } + template + inline void gemm_impl(sycl::queue & q, oneapi::math::transpose a_trans, oneapi::math::transpose b_trans, int m, + int n, int k, const void * alpha, const void * a, int lda, const void * b, int ldb, + const void * beta, void * c, int ldc) { + Ts alpha_value = dpct::get_value(reinterpret_cast(alpha), q); + Ts beta_value = dpct::get_value(reinterpret_cast(beta), q); + auto data_a = get_memory(a); + auto data_b = get_memory(b); + auto data_c = get_memory(c); + oneapi::math::blas::column_major::gemm(get_onemath_backend(q), a_trans, b_trans, m, n, k, alpha_value, data_a, + lda, data_b, ldb, beta_value, data_c, ldc); + } template class vectorized_binary @@ -1735,7 +1755,7 @@ namespace dpct }; template - inline void gemm_batch_impl(sycl::queue & q, oneapi::mkl::transpose a_trans, oneapi::mkl::transpose b_trans, + inline void gemm_batch_impl(sycl::queue & q, oneapi::math::transpose a_trans, oneapi::math::transpose b_trans, int m, int n, int k, const void * alpha, const void ** a, int lda, const void ** b, int ldb, const void * beta, void ** c, int ldc, int batch_size, matrix_info_t * matrix_info) { @@ -1754,48 +1774,28 @@ namespace dpct matrix_info->ld_info[2] = ldc; matrix_info->groupsize_info = batch_size; -#ifdef GGML_SYCL_NVIDIA - sycl::event e = oneapi::mkl::blas::column_major::gemm_batch( - oneapi::mkl::backend_selector{ q }, matrix_info->transpose_info, - matrix_info->transpose_info + 1, matrix_info->size_info, matrix_info->size_info + 1, - matrix_info->size_info + 2, reinterpret_cast(matrix_info->value_info), - reinterpret_cast(a), matrix_info->ld_info, reinterpret_cast(b), - matrix_info->ld_info + 1, reinterpret_cast(matrix_info->value_info + 1), - reinterpret_cast(c), matrix_info->ld_info + 2, 1, &(matrix_info->groupsize_info)); -#else - sycl::event e = oneapi::mkl::blas::column_major::gemm_batch( - q, matrix_info->transpose_info, matrix_info->transpose_info + 1, matrix_info->size_info, - matrix_info->size_info + 1, matrix_info->size_info + 2, reinterpret_cast(matrix_info->value_info), - reinterpret_cast(a), matrix_info->ld_info, reinterpret_cast(b), - matrix_info->ld_info + 1, reinterpret_cast(matrix_info->value_info + 1), - reinterpret_cast(c), matrix_info->ld_info + 2, 1, &(matrix_info->groupsize_info)); -#endif + sycl::event e = oneapi::math::blas::column_major::gemm_batch( + get_onemath_backend(q), matrix_info->transpose_info, matrix_info->transpose_info + 1, + matrix_info->size_info, matrix_info->size_info + 1, matrix_info->size_info + 2, + reinterpret_cast(matrix_info->value_info), reinterpret_cast(a), matrix_info->ld_info, + reinterpret_cast(b), matrix_info->ld_info + 1, + reinterpret_cast(matrix_info->value_info + 1), reinterpret_cast(c), + matrix_info->ld_info + 2, 1, &(matrix_info->groupsize_info)); } template - inline void - gemm_batch_impl(sycl::queue &q, oneapi::mkl::transpose a_trans, - oneapi::mkl::transpose b_trans, int m, int n, - int k, const void *alpha, const void *a, int lda, - long long int stride_a, const void *b, int ldb, - long long int stride_b, const void *beta, void *c, - int ldc, long long int stride_c, int batch_size) - { + inline void gemm_batch_impl(sycl::queue & q, oneapi::math::transpose a_trans, oneapi::math::transpose b_trans, + int m, int n, int k, const void * alpha, const void * a, int lda, + long long int stride_a, const void * b, int ldb, long long int stride_b, + const void * beta, void * c, int ldc, long long int stride_c, int batch_size) { Ts alpha_value = dpct::get_value(reinterpret_cast(alpha), q); Ts beta_value = dpct::get_value(reinterpret_cast(beta), q); auto data_a = get_memory(a); auto data_b = get_memory(b); auto data_c = get_memory(c); -#ifdef GGML_SYCL_NVIDIA - oneapi::mkl::blas::column_major::gemm_batch( - oneapi::mkl::backend_selector{ q }, a_trans, b_trans, m, n, k, - alpha_value, data_a, lda, stride_a, data_b, ldb, stride_b, beta_value, data_c, ldc, stride_c, - batch_size); -#else - oneapi::mkl::blas::column_major::gemm_batch(q, a_trans, b_trans, m, n, k, alpha_value, data_a, lda, - stride_a, data_b, ldb, stride_b, beta_value, data_c, ldc, - stride_c, batch_size); -#endif + oneapi::math::blas::column_major::gemm_batch(get_onemath_backend(q), a_trans, b_trans, m, n, k, alpha_value, + data_a, lda, stride_a, data_b, ldb, stride_b, beta_value, + data_c, ldc, stride_c, batch_size); } } // namespace detail @@ -2259,13 +2259,10 @@ namespace dpct sycl::range<3>(x, y, 1), direction); } - inline void gemm(sycl::queue &q, oneapi::mkl::transpose a_trans, - oneapi::mkl::transpose b_trans, int m, int n, int k, - const void *alpha, const void *a, library_data_t a_type, - int lda, const void *b, library_data_t b_type, int ldb, - const void *beta, void *c, library_data_t c_type, int ldc, - library_data_t scaling_type) - { + inline void gemm(sycl::queue & q, oneapi::math::transpose a_trans, oneapi::math::transpose b_trans, int m, int n, + int k, const void * alpha, const void * a, library_data_t a_type, int lda, const void * b, + library_data_t b_type, int ldb, const void * beta, void * c, library_data_t c_type, int ldc, + library_data_t scaling_type) { if (scaling_type == library_data_t::real_float && c_type == library_data_t::complex_float) { @@ -2329,9 +2326,8 @@ namespace dpct library_data_t::real_bfloat16, library_data_t::real_bfloat16, library_data_t::real_float, library_data_t::real_float): { - detail::gemm_impl(q, a_trans, b_trans, m, n, k, alpha, a, lda, b, - ldb, beta, c, ldc); + detail::gemm_impl( + q, a_trans, b_trans, m, n, k, alpha, a, lda, b, ldb, beta, c, ldc); break; } case detail::get_type_combination_id( @@ -2369,8 +2365,7 @@ namespace dpct library_data_t::real_bfloat16, library_data_t::real_bfloat16, library_data_t::real_bfloat16, library_data_t::real_float): { - detail::gemm_impl( + detail::gemm_impl( q, a_trans, b_trans, m, n, k, alpha, a, lda, b, ldb, beta, c, ldc); break; } @@ -2390,7 +2385,7 @@ namespace dpct default: throw std::runtime_error("the combination of data type is unsupported"); } - } // gemm() + } // gemm() /// Computes a batch of matrix-matrix product with general matrices. /// \param [in] q The queue where the routine should be executed. @@ -2412,7 +2407,7 @@ namespace dpct /// \param [in] ldc Leading dimension of C. /// \param [in] batch_size Specifies the number of matrix multiply operations to perform. /// \param [in] scaling_type Data type of the scaling factors. - inline void gemm_batch(sycl::queue & q, oneapi::mkl::transpose a_trans, oneapi::mkl::transpose b_trans, int m, + inline void gemm_batch(sycl::queue & q, oneapi::math::transpose a_trans, oneapi::math::transpose b_trans, int m, int n, int k, const void * alpha, const void * a[], library_data_t a_type, int lda, const void * b[], library_data_t b_type, int ldb, const void * beta, void * c[], library_data_t c_type, int ldc, int batch_size, library_data_t scaling_type, @@ -2450,7 +2445,7 @@ namespace dpct library_data_t::real_bfloat16, library_data_t::real_bfloat16, library_data_t::real_bfloat16, library_data_t::real_float): { - detail::gemm_batch_impl( + detail::gemm_batch_impl( q, a_trans, b_trans, m, n, k, alpha, a, lda, b, ldb, beta, c, ldc, batch_size, matrix_info); break; } @@ -2458,7 +2453,7 @@ namespace dpct library_data_t::real_bfloat16, library_data_t::real_bfloat16, library_data_t::real_float, library_data_t::real_float): { - detail::gemm_batch_impl( + detail::gemm_batch_impl( q, a_trans, b_trans, m, n, k, alpha, a, lda, b, ldb, beta, c, ldc, batch_size, matrix_info); break; } @@ -2534,15 +2529,11 @@ namespace dpct /// \param [in] stride_c Stride between the different C matrices. /// \param [in] batch_size Specifies the number of matrix multiply operations to perform. /// \param [in] scaling_type Data type of the scaling factors. - inline void gemm_batch(sycl::queue &q, oneapi::mkl::transpose a_trans, - oneapi::mkl::transpose b_trans, int m, int n, int k, - const void *alpha, const void *a, library_data_t a_type, - int lda, long long int stride_a, const void *b, - library_data_t b_type, int ldb, long long int stride_b, - const void *beta, void *c, library_data_t c_type, - int ldc, long long int stride_c, int batch_size, - library_data_t scaling_type) - { + inline void gemm_batch(sycl::queue & q, oneapi::math::transpose a_trans, oneapi::math::transpose b_trans, int m, + int n, int k, const void * alpha, const void * a, library_data_t a_type, int lda, + long long int stride_a, const void * b, library_data_t b_type, int ldb, + long long int stride_b, const void * beta, void * c, library_data_t c_type, int ldc, + long long int stride_c, int batch_size, library_data_t scaling_type) { if (scaling_type == library_data_t::real_float && c_type == library_data_t::complex_float) { @@ -2611,20 +2602,18 @@ namespace dpct library_data_t::real_bfloat16, library_data_t::real_bfloat16, library_data_t::real_bfloat16, library_data_t::real_float): { - detail::gemm_batch_impl( - q, a_trans, b_trans, m, n, k, alpha, a, lda, stride_a, b, ldb, stride_b, - beta, c, ldc, stride_c, batch_size); + detail::gemm_batch_impl( + q, a_trans, b_trans, m, n, k, alpha, a, lda, stride_a, b, ldb, stride_b, beta, c, ldc, stride_c, + batch_size); break; } case detail::get_type_combination_id( library_data_t::real_bfloat16, library_data_t::real_bfloat16, library_data_t::real_float, library_data_t::real_float): { - detail::gemm_batch_impl(q, a_trans, b_trans, m, n, k, alpha, a, lda, - stride_a, b, ldb, stride_b, beta, c, ldc, - stride_c, batch_size); + detail::gemm_batch_impl( + q, a_trans, b_trans, m, n, k, alpha, a, lda, stride_a, b, ldb, stride_b, beta, c, ldc, stride_c, + batch_size); break; } #endif diff --git a/ggml/src/ggml-sycl/element_wise.cpp b/ggml/src/ggml-sycl/element_wise.cpp index 4bcd74376..b36aa7a9d 100644 --- a/ggml/src/ggml-sycl/element_wise.cpp +++ b/ggml/src/ggml-sycl/element_wise.cpp @@ -1,7 +1,8 @@ #include "common.hpp" +#include "ggml.h" #include "element_wise.hpp" -void acc_f32(const float * x, const float * y, float * dst, const int ne, +static void acc_f32(const float * x, const float * y, float * dst, const int ne, const int ne10, const int ne11, const int ne12, const int nb1, const int nb2, int offset, const sycl::nd_item<3> &item_ct1) { const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) + @@ -20,10 +21,11 @@ void acc_f32(const float * x, const float * y, float * dst, const int ne, } } -void gelu_f32(const float * x, float * dst, const int k, +template +static void gelu(const T * x, T * dst, const int k, const sycl::nd_item<3> &item_ct1) { - const float GELU_COEF_A = 0.044715f; - const float SQRT_2_OVER_PI = 0.79788456080286535587989211986876f; + const T GELU_COEF_A = static_cast(0.044715f); + const T SQRT_2_OVER_PI = static_cast(0.79788456080286535587989211986876f); const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) + item_ct1.get_local_id(2); @@ -32,12 +34,13 @@ void gelu_f32(const float * x, float * dst, const int k, } float xi = x[i]; - dst[i] = 0.5f * xi * - (1.0f + - sycl::tanh(SQRT_2_OVER_PI * xi * (1.0f + GELU_COEF_A * xi * xi))); + dst[i] = static_cast(0.5f) * xi * + (static_cast(1.0f) + + sycl::tanh(SQRT_2_OVER_PI * xi * (static_cast(1.0f) + GELU_COEF_A * xi * xi))); } -void silu_f32(const float * x, float * dst, const int k, +template +static void silu(const T * x, T * dst, const int k, const sycl::nd_item<3> &item_ct1) { const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) + item_ct1.get_local_id(2); @@ -45,10 +48,11 @@ void silu_f32(const float * x, float * dst, const int k, if (i >= k) { return; } - dst[i] = x[i] / (1.0f + sycl::native::exp(-x[i])); + dst[i] = x[i] / (static_cast(1.0f) + sycl::native::exp(-x[i])); } -void gelu_quick_f32(const float *x, float *dst, int k, +template +static void gelu_quick(const T *x, T *dst, int k, const sycl::nd_item<3> &item_ct1) { const float GELU_QUICK_COEF = -1.702f; const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) + @@ -56,20 +60,22 @@ void gelu_quick_f32(const float *x, float *dst, int k, if (i >= k) { return; } - dst[i] = x[i] * (1.0f / (1.0f + sycl::native::exp(GELU_QUICK_COEF * x[i]))); + dst[i] = x[i] * (static_cast(1.0f) / (static_cast(1.0f) + sycl::native::exp(GELU_QUICK_COEF * x[i]))); } -void tanh_f32(const float *x, float *dst, int k, +template +static void tanh(const T *x, T *dst, int k, const sycl::nd_item<3> &item_ct1) { const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) + item_ct1.get_local_id(2); if (i >= k) { return; } - dst[i] = sycl::tanh((float)(x[i])); + dst[i] = sycl::tanh((x[i])); } -void relu_f32(const float * x, float * dst, const int k, +template +static void relu(const T * x, T * dst, const int k, const sycl::nd_item<3> &item_ct1) { const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) + item_ct1.get_local_id(2); @@ -77,10 +83,11 @@ void relu_f32(const float * x, float * dst, const int k, if (i >= k) { return; } - dst[i] = sycl::fmax((float)(x[i]), (float)0); + dst[i] = sycl::fmax((x[i]), static_cast(0)); } -void sigmoid_f32(const float * x, float * dst, const int k, +template +static void sigmoid(const T * x, T * dst, const int k, const sycl::nd_item<3> &item_ct1) { const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) + item_ct1.get_local_id(2); @@ -88,10 +95,11 @@ void sigmoid_f32(const float * x, float * dst, const int k, if (i >= k) { return; } - dst[i] = 1.0f / (1.0f + sycl::native::exp(-x[i])); + dst[i] = 1.0f / (static_cast(1.0f) + sycl::native::exp(-x[i])); } -void sqrt_f32(const float * x, float * dst, const int k, +template +static void sqrt(const T * x, T * dst, const int k, const sycl::nd_item<3> &item_ct1) { const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) + item_ct1.get_local_id(2); @@ -102,7 +110,8 @@ void sqrt_f32(const float * x, float * dst, const int k, dst[i] = sycl::sqrt(x[i]); } -void sin_f32(const float * x, float * dst, const int k, +template +static void sin(const T * x, T * dst, const int k, const sycl::nd_item<3> &item_ct1) { const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) + item_ct1.get_local_id(2); @@ -113,7 +122,8 @@ void sin_f32(const float * x, float * dst, const int k, dst[i] = sycl::sin(x[i]); } -void cos_f32(const float * x, float * dst, const int k, +template +static void cos(const T * x, T * dst, const int k, const sycl::nd_item<3> &item_ct1) { const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) + item_ct1.get_local_id(2); @@ -124,7 +134,8 @@ void cos_f32(const float * x, float * dst, const int k, dst[i] = sycl::cos(x[i]); } -void hardsigmoid_f32(const float * x, float * dst, const int k, +template +static void hardsigmoid(const T * x, T * dst, const int k, const sycl::nd_item<3> &item_ct1) { const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) + item_ct1.get_local_id(2); @@ -132,10 +143,11 @@ void hardsigmoid_f32(const float * x, float * dst, const int k, if (i >= k) { return; } - dst[i] = sycl::fmin(1.0f, sycl::fmax(0.0f, (x[i] + 3.0f) / 6.0f)); + dst[i] = sycl::fmin(static_cast(1.0f), sycl::fmax(static_cast(0.0f), (x[i] + static_cast(3.0f)) / static_cast(6.0f))); } -void hardswish_f32(const float * x, float * dst, const int k, +template +static void hardswish(const T * x, T * dst, const int k, const sycl::nd_item<3> &item_ct1) { const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) + item_ct1.get_local_id(2); @@ -143,10 +155,11 @@ void hardswish_f32(const float * x, float * dst, const int k, if (i >= k) { return; } - dst[i] = x[i] * sycl::fmin(1.0f, sycl::fmax(0.0f, (x[i] + 3.0f) / 6.0f)); + dst[i] = x[i] * sycl::fmin(static_cast(1.0f), sycl::fmax(static_cast(0.0f), (x[i] + static_cast(3.0f)) / static_cast(6.0f))); } -void exp_f32(const float * x, float * dst, const int k, +template +static void exp(const T * x, T * dst, const int k, const sycl::nd_item<3> &item_ct1) { const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) + item_ct1.get_local_id(2); @@ -157,7 +170,8 @@ void exp_f32(const float * x, float * dst, const int k, dst[i] = sycl::exp(x[i]); } -void log_f32(const float * x, float * dst, const int k, +template +static void log(const T * x, T * dst, const int k, const sycl::nd_item<3> &item_ct1) { const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) + item_ct1.get_local_id(2); @@ -165,15 +179,16 @@ void log_f32(const float * x, float * dst, const int k, if (i >= k) { return; } - float xi = x[i]; + T xi = x[i]; if (xi <= 0) { - dst[i] = -INFINITY; + dst[i] = neg_infinity(); } else { dst[i] = sycl::log(xi); } } -void neg_f32(const float * x, float * dst, const int k, +template +static void neg(const T * x, T * dst, const int k, const sycl::nd_item<3> &item_ct1) { const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) + item_ct1.get_local_id(2); @@ -184,7 +199,8 @@ void neg_f32(const float * x, float * dst, const int k, dst[i] = -x[i]; } -void step_f32(const float * x, float * dst, const int k, +template +static void step(const T * x, T * dst, const int k, const sycl::nd_item<3> &item_ct1) { const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) + item_ct1.get_local_id(2); @@ -192,21 +208,23 @@ void step_f32(const float * x, float * dst, const int k, if (i >= k) { return; } - dst[i] = x[i] > 0.0f; + dst[i] = x[i] > static_cast(0.0f); } -void leaky_relu_f32(const float *x, float *dst, const int k, const float negative_slope, +template +static void leaky_relu(const T *x, T *dst, const int k, const float negative_slope, const sycl::nd_item<3> &item_ct1) { const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) + item_ct1.get_local_id(2); if (i >= k) { return; } - dst[i] = sycl::fmax((float)(x[i]), (float)0) + - sycl::fmin((float)(x[i]), 0.0f) * negative_slope; + dst[i] = sycl::fmax((x[i]), static_cast(0)) + + sycl::fmin((x[i]), static_cast(0.0f)) * negative_slope; } -void sqr_f32(const float * x, float * dst, const int k, +template +static void sqr(const T * x, T * dst, const int k, const sycl::nd_item<3> &item_ct1) { const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) + item_ct1.get_local_id(2); @@ -217,7 +235,8 @@ void sqr_f32(const float * x, float * dst, const int k, dst[i] = x[i] * x[i]; } -void upscale_f32(const float *x, float *dst, const int nb00, const int nb01, +template +static void upscale(const T *x, T *dst, const int nb00, const int nb01, const int nb02, const int nb03, const int ne10, const int ne11, const int ne12, const int ne13, const float sf0, const float sf1, const float sf2, const float sf3, const sycl::nd_item<1> &item_ct1) { @@ -237,10 +256,11 @@ void upscale_f32(const float *x, float *dst, const int nb00, const int nb01, int i02 = i12 / sf2; int i03 = i13 / sf3; - dst[index] = *(const float *)((const char *)x + i03 * nb03 + i02 * nb02 + i01 * nb01 + i00 * nb00); + dst[index] = *(const T *)((const char *)x + i03 * nb03 + i02 * nb02 + i01 * nb01 + i00 * nb00); } -void pad_f32(const float *x, float *dst, const int ne0, const int ne00, const int ne01, const int ne02, +template +static void pad(const T *x, T *dst, const int ne0, const int ne00, const int ne01, const int ne02, const sycl::nd_item<3> &item_ct1) { int nidx = item_ct1.get_local_id(2) + item_ct1.get_group(2) * item_ct1.get_local_range(2); @@ -256,13 +276,25 @@ void pad_f32(const float *x, float *dst, const int ne0, const int ne00, const i item_ct1.get_group(0) * ne00 * ne01; dst[offset_dst] = x[offset_src]; } else { - dst[offset_dst] = 0.0f; + dst[offset_dst] = static_cast(0.0f); } } +template +static void clamp(const T * x, T * dst, const float min, const float max, const int k, + const sycl::nd_item<3> &item_ct1) { + const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) + + item_ct1.get_local_id(2); -void acc_f32_sycl(const float *x, const float *y, float *dst, + if (i >= k) { + return; + } + + dst[i] = x[i] < static_cast(min) ? static_cast(min) : (x[i] > static_cast(max) ? static_cast(max) : x[i]); +} + +static void acc_f32_sycl(const float *x, const float *y, float *dst, const int n_elements, const int ne10, const int ne11, const int ne12, const int nb1, const int nb2, const int offset, queue_ptr stream) { @@ -277,7 +309,8 @@ void acc_f32_sycl(const float *x, const float *y, float *dst, }); } -void gelu_f32_sycl(const float *x, float *dst, const int k, +template +static void gelu_sycl(const T *x, T *dst, const int k, queue_ptr stream) { const int num_blocks = (k + SYCL_GELU_BLOCK_SIZE - 1) / SYCL_GELU_BLOCK_SIZE; stream->parallel_for( @@ -285,11 +318,12 @@ void gelu_f32_sycl(const float *x, float *dst, const int k, sycl::range<3>(1, 1, SYCL_GELU_BLOCK_SIZE), sycl::range<3>(1, 1, SYCL_GELU_BLOCK_SIZE)), [=](sycl::nd_item<3> item_ct1) { - gelu_f32(x, dst, k, item_ct1); + gelu(x, dst, k, item_ct1); }); } -void silu_f32_sycl(const float *x, float *dst, const int k, +template +static void silu_sycl(const T *x, T *dst, const int k, queue_ptr stream) { const int num_blocks = (k + SYCL_SILU_BLOCK_SIZE - 1) / SYCL_SILU_BLOCK_SIZE; stream->parallel_for( @@ -297,11 +331,12 @@ void silu_f32_sycl(const float *x, float *dst, const int k, sycl::range<3>(1, 1, SYCL_SILU_BLOCK_SIZE), sycl::range<3>(1, 1, SYCL_SILU_BLOCK_SIZE)), [=](sycl::nd_item<3> item_ct1) { - silu_f32(x, dst, k, item_ct1); + silu(x, dst, k, item_ct1); }); } -void gelu_quick_f32_sycl(const float *x, float *dst, const int k, +template +static void gelu_quick_sycl(const T *x, T *dst, const int k, queue_ptr stream) { const int num_blocks = (k + SYCL_GELU_BLOCK_SIZE - 1) / SYCL_GELU_BLOCK_SIZE; stream->parallel_for( @@ -309,11 +344,12 @@ void gelu_quick_f32_sycl(const float *x, float *dst, const int k, sycl::range<3>(1, 1, SYCL_GELU_BLOCK_SIZE), sycl::range<3>(1, 1, SYCL_GELU_BLOCK_SIZE)), [=](sycl::nd_item<3> item_ct1) { - gelu_quick_f32(x, dst, k, item_ct1); + gelu_quick(x, dst, k, item_ct1); }); } -void tanh_f32_sycl(const float *x, float *dst, const int k, +template +static void tanh_sycl(const T *x, T *dst, const int k, queue_ptr stream) { const int num_blocks = (k + SYCL_TANH_BLOCK_SIZE - 1) / SYCL_TANH_BLOCK_SIZE; stream->parallel_for( @@ -321,11 +357,12 @@ void tanh_f32_sycl(const float *x, float *dst, const int k, sycl::range<3>(1, 1, SYCL_TANH_BLOCK_SIZE), sycl::range<3>(1, 1, SYCL_TANH_BLOCK_SIZE)), [=](sycl::nd_item<3> item_ct1) { - tanh_f32(x, dst, k, item_ct1); + tanh(x, dst, k, item_ct1); }); } -void relu_f32_sycl(const float *x, float *dst, const int k, +template +static void relu_sycl(const T *x, T *dst, const int k, queue_ptr stream) { const int num_blocks = (k + SYCL_RELU_BLOCK_SIZE - 1) / SYCL_RELU_BLOCK_SIZE; stream->parallel_for( @@ -333,11 +370,12 @@ void relu_f32_sycl(const float *x, float *dst, const int k, sycl::range<3>(1, 1, SYCL_RELU_BLOCK_SIZE), sycl::range<3>(1, 1, SYCL_RELU_BLOCK_SIZE)), [=](sycl::nd_item<3> item_ct1) { - relu_f32(x, dst, k, item_ct1); + relu(x, dst, k, item_ct1); }); } -void hardsigmoid_f32_sycl(const float *x, float *dst, const int k, +template +static void hardsigmoid_sycl(const T *x, T *dst, const int k, queue_ptr stream) { const int num_blocks = (k + SYCL_HARDSIGMOID_BLOCK_SIZE - 1) / SYCL_HARDSIGMOID_BLOCK_SIZE; stream->parallel_for( @@ -345,11 +383,12 @@ void hardsigmoid_f32_sycl(const float *x, float *dst, const int k, sycl::range<3>(1, 1, SYCL_HARDSIGMOID_BLOCK_SIZE), sycl::range<3>(1, 1, SYCL_HARDSIGMOID_BLOCK_SIZE)), [=](sycl::nd_item<3> item_ct1) { - hardsigmoid_f32(x, dst, k, item_ct1); + hardsigmoid(x, dst, k, item_ct1); }); } -void hardswish_f32_sycl(const float *x, float *dst, const int k, +template +static void hardswish_sycl(const T *x, T *dst, const int k, queue_ptr stream) { const int num_blocks = (k + SYCL_HARDSWISH_BLOCK_SIZE - 1) / SYCL_HARDSWISH_BLOCK_SIZE; stream->parallel_for( @@ -357,11 +396,12 @@ void hardswish_f32_sycl(const float *x, float *dst, const int k, sycl::range<3>(1, 1, SYCL_HARDSWISH_BLOCK_SIZE), sycl::range<3>(1, 1, SYCL_HARDSWISH_BLOCK_SIZE)), [=](sycl::nd_item<3> item_ct1) { - hardswish_f32(x, dst, k, item_ct1); + hardswish(x, dst, k, item_ct1); }); } -void exp_f32_sycl(const float *x, float *dst, const int k, +template +static void exp_sycl(const T *x, T *dst, const int k, queue_ptr stream) { const int num_blocks = (k + SYCL_EXP_BLOCK_SIZE - 1) / SYCL_EXP_BLOCK_SIZE; stream->parallel_for( @@ -369,11 +409,12 @@ void exp_f32_sycl(const float *x, float *dst, const int k, sycl::range<3>(1, 1, SYCL_EXP_BLOCK_SIZE), sycl::range<3>(1, 1, SYCL_EXP_BLOCK_SIZE)), [=](sycl::nd_item<3> item_ct1) { - exp_f32(x, dst, k, item_ct1); + exp(x, dst, k, item_ct1); }); } -void log_f32_sycl(const float *x, float *dst, const int k, +template +static void log_sycl(const T *x, T *dst, const int k, queue_ptr stream) { const int num_blocks = (k + SYCL_EXP_BLOCK_SIZE - 1) / SYCL_EXP_BLOCK_SIZE; stream->parallel_for( @@ -381,11 +422,12 @@ void log_f32_sycl(const float *x, float *dst, const int k, sycl::range<3>(1, 1, SYCL_EXP_BLOCK_SIZE), sycl::range<3>(1, 1, SYCL_EXP_BLOCK_SIZE)), [=](sycl::nd_item<3> item_ct1) { - log_f32(x, dst, k, item_ct1); + log(x, dst, k, item_ct1); }); } -void neg_f32_sycl(const float *x, float *dst, const int k, +template +static void neg_sycl(const T *x, T *dst, const int k, queue_ptr stream) { const int num_blocks = (k + SYCL_NEG_BLOCK_SIZE - 1) / SYCL_NEG_BLOCK_SIZE; stream->parallel_for( @@ -393,11 +435,12 @@ void neg_f32_sycl(const float *x, float *dst, const int k, sycl::range<3>(1, 1, SYCL_NEG_BLOCK_SIZE), sycl::range<3>(1, 1, SYCL_NEG_BLOCK_SIZE)), [=](sycl::nd_item<3> item_ct1) { - neg_f32(x, dst, k, item_ct1); + neg(x, dst, k, item_ct1); }); } -void step_f32_sycl(const float *x, float *dst, const int k, +template +static void step_sycl(const T *x, T *dst, const int k, queue_ptr stream) { const int num_blocks = (k + SYCL_NEG_BLOCK_SIZE - 1) / SYCL_NEG_BLOCK_SIZE; stream->parallel_for( @@ -405,11 +448,12 @@ void step_f32_sycl(const float *x, float *dst, const int k, sycl::range<3>(1, 1, SYCL_NEG_BLOCK_SIZE), sycl::range<3>(1, 1, SYCL_NEG_BLOCK_SIZE)), [=](sycl::nd_item<3> item_ct1) { - step_f32(x, dst, k, item_ct1); + step(x, dst, k, item_ct1); }); } -void sigmoid_f32_sycl(const float *x, float *dst, const int k, +template +static void sigmoid_sycl(const T *x, T *dst, const int k, queue_ptr stream) { const int num_blocks = (k + SYCL_SIGMOID_BLOCK_SIZE - 1) / SYCL_SIGMOID_BLOCK_SIZE; stream->parallel_for( @@ -417,11 +461,12 @@ void sigmoid_f32_sycl(const float *x, float *dst, const int k, sycl::range<3>(1, 1, SYCL_SIGMOID_BLOCK_SIZE), sycl::range<3>(1, 1, SYCL_SIGMOID_BLOCK_SIZE)), [=](sycl::nd_item<3> item_ct1) { - sigmoid_f32(x, dst, k, item_ct1); + sigmoid(x, dst, k, item_ct1); }); } -void sqrt_f32_sycl(const float *x, float *dst, const int k, +template +static void sqrt_sycl(const T *x, T *dst, const int k, queue_ptr stream) { const int num_blocks = (k + SYCL_SQRT_BLOCK_SIZE - 1) / SYCL_SQRT_BLOCK_SIZE; stream->parallel_for( @@ -429,11 +474,12 @@ void sqrt_f32_sycl(const float *x, float *dst, const int k, sycl::range<3>(1, 1, SYCL_SQRT_BLOCK_SIZE), sycl::range<3>(1, 1, SYCL_SQRT_BLOCK_SIZE)), [=](sycl::nd_item<3> item_ct1) { - sqrt_f32(x, dst, k, item_ct1); + sqrt(x, dst, k, item_ct1); }); } -void sin_f32_sycl(const float *x, float *dst, const int k, +template +static void sin_sycl(const T *x, T *dst, const int k, queue_ptr stream) { const int num_blocks = (k + SYCL_SIN_BLOCK_SIZE - 1) / SYCL_SIN_BLOCK_SIZE; stream->parallel_for( @@ -441,11 +487,12 @@ void sin_f32_sycl(const float *x, float *dst, const int k, sycl::range<3>(1, 1, SYCL_SIN_BLOCK_SIZE), sycl::range<3>(1, 1, SYCL_SIN_BLOCK_SIZE)), [=](sycl::nd_item<3> item_ct1) { - sin_f32(x, dst, k, item_ct1); + sin(x, dst, k, item_ct1); }); } -void cos_f32_sycl(const float *x, float *dst, const int k, +template +static void cos_sycl(const T *x, T *dst, const int k, queue_ptr stream) { const int num_blocks = (k + SYCL_SIN_BLOCK_SIZE - 1) / SYCL_SIN_BLOCK_SIZE; stream->parallel_for( @@ -453,11 +500,12 @@ void cos_f32_sycl(const float *x, float *dst, const int k, sycl::range<3>(1, 1, SYCL_SIN_BLOCK_SIZE), sycl::range<3>(1, 1, SYCL_SIN_BLOCK_SIZE)), [=](sycl::nd_item<3> item_ct1) { - cos_f32(x, dst, k, item_ct1); + cos(x, dst, k, item_ct1); }); } -void leaky_relu_f32_sycl(const float *x, float *dst, const int k, +template +static void leaky_relu_sycl(const T *x, T *dst, const int k, const float negative_slope, queue_ptr stream) { const int num_blocks = (k + SYCL_RELU_BLOCK_SIZE - 1) / SYCL_RELU_BLOCK_SIZE; @@ -466,11 +514,12 @@ void leaky_relu_f32_sycl(const float *x, float *dst, const int k, sycl::range<3>(1, 1, SYCL_RELU_BLOCK_SIZE), sycl::range<3>(1, 1, SYCL_RELU_BLOCK_SIZE)), [=](sycl::nd_item<3> item_ct1) { - leaky_relu_f32(x, dst, k, negative_slope, item_ct1); + leaky_relu(x, dst, k, negative_slope, item_ct1); }); } -void sqr_f32_sycl(const float *x, float *dst, const int k, +template +static void sqr_sycl(const T *x, T *dst, const int k, queue_ptr stream) { const int num_blocks = (k + SYCL_SQR_BLOCK_SIZE - 1) / SYCL_SQR_BLOCK_SIZE; stream->parallel_for( @@ -478,11 +527,12 @@ void sqr_f32_sycl(const float *x, float *dst, const int k, sycl::range<3>(1, 1, SYCL_SQR_BLOCK_SIZE), sycl::range<3>(1, 1, SYCL_SQR_BLOCK_SIZE)), [=](sycl::nd_item<3> item_ct1) { - sqr_f32(x, dst, k, item_ct1); + sqr(x, dst, k, item_ct1); }); } -void upscale_f32_sycl(const float *x, float *dst, const int nb00, const int nb01, +template +static void upscale_sycl(const T *x, T *dst, const int nb00, const int nb01, const int nb02, const int nb03, const int ne10, const int ne11, const int ne12, const int ne13, const float sf0, const float sf1, const float sf2, const float sf3, queue_ptr stream) { @@ -492,11 +542,12 @@ void upscale_f32_sycl(const float *x, float *dst, const int nb00, const int nb01 stream->parallel_for( sycl::nd_range<1>(gridDim, sycl::range<1>(SYCL_UPSCALE_BLOCK_SIZE)), [=](sycl::nd_item<1> item_ct1) { - upscale_f32(x, dst, nb00, nb01, nb02, nb03, ne10, ne11, ne12, ne13, sf0, sf1, sf2, sf3, item_ct1); + upscale(x, dst, nb00, nb01, nb02, nb03, ne10, ne11, ne12, ne13, sf0, sf1, sf2, sf3, item_ct1); }); } -void pad_f32_sycl(const float *x, float *dst, const int ne00, +template +static void pad_sycl(const T *x, T *dst, const int ne00, const int ne01, const int ne02, const int ne0, const int ne1, const int ne2, queue_ptr stream) { int num_blocks = (ne0 + SYCL_PAD_BLOCK_SIZE - 1) / SYCL_PAD_BLOCK_SIZE; @@ -505,526 +556,882 @@ void pad_f32_sycl(const float *x, float *dst, const int ne00, sycl::nd_range<3>(gridDim * sycl::range<3>(1, 1, SYCL_PAD_BLOCK_SIZE), sycl::range<3>(1, 1, SYCL_PAD_BLOCK_SIZE)), [=](sycl::nd_item<3> item_ct1) { - pad_f32(x, dst, ne0, ne00, ne01, ne02, item_ct1); + pad(x, dst, ne0, ne00, ne01, ne02, item_ct1); }); } -inline void ggml_sycl_op_silu(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1, - ggml_tensor *dst, const float *src0_dd, - const float *src1_dd, float *dst_dd, - const queue_ptr &main_stream) { - - GGML_ASSERT(src0->type == GGML_TYPE_F32); - GGML_ASSERT( dst->type == GGML_TYPE_F32); - - silu_f32_sycl(src0_dd, dst_dd, ggml_nelements(src0), main_stream); - - GGML_UNUSED(src1); - GGML_UNUSED(dst); - GGML_UNUSED(src1_dd); - GGML_UNUSED(ctx); +template +static void clamp_sycl(const T *x, T *dst, const float min, + const float max, const int k, + queue_ptr stream) { + const int num_blocks = (k + SYCL_CLAMP_BLOCK_SIZE - 1) / SYCL_CLAMP_BLOCK_SIZE; + stream->parallel_for( + sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * + sycl::range<3>(1, 1, SYCL_CLAMP_BLOCK_SIZE), + sycl::range<3>(1, 1, SYCL_CLAMP_BLOCK_SIZE)), + [=](sycl::nd_item<3> item_ct1) { + clamp(x, dst, min, max, k, item_ct1); + }); } -inline void ggml_sycl_op_gelu(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1, - ggml_tensor *dst, const float *src0_dd, - const float *src1_dd, float *dst_dd, - const queue_ptr &main_stream) { - - GGML_ASSERT(src0->type == GGML_TYPE_F32); - GGML_ASSERT( dst->type == GGML_TYPE_F32); - - gelu_f32_sycl(src0_dd, dst_dd, ggml_nelements(src0), main_stream); - - GGML_UNUSED(src1); - GGML_UNUSED(dst); - GGML_UNUSED(src1_dd); - GGML_UNUSED(ctx); -} -inline void ggml_sycl_op_gelu_quick(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, - const ggml_tensor *src1, ggml_tensor *dst, - const float *src0_dd, const float *src1_dd, - float *dst_dd, - const queue_ptr &main_stream) { - - GGML_ASSERT(src0->type == GGML_TYPE_F32); - GGML_ASSERT( dst->type == GGML_TYPE_F32); - - gelu_quick_f32_sycl(src0_dd, dst_dd, ggml_nelements(src0), main_stream); - - GGML_UNUSED(src1); - GGML_UNUSED(dst); - GGML_UNUSED(src1_dd); - GGML_UNUSED(ctx); +inline void ggml_sycl_op_silu(ggml_backend_sycl_context & ctx, ggml_tensor * dst) { +#if defined (GGML_SYCL_F16) + GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32 || dst->src[0]->type == GGML_TYPE_F16); + GGML_ASSERT(dst->type == GGML_TYPE_F32 || dst->type == GGML_TYPE_F16); +#else + GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32); + GGML_ASSERT(dst->type == GGML_TYPE_F32); +#endif + GGML_ASSERT(dst->src[0]->type == dst->type); + dpct::queue_ptr main_stream = ctx.stream(); + SYCL_CHECK(ggml_sycl_set_device(ctx.device)); + switch (dst->type) { +#if defined (GGML_SYCL_F16) + case GGML_TYPE_F16: + { + auto data_pts = cast_data(dst); + silu_sycl(data_pts.src, data_pts.dst, ggml_nelements(dst->src[0]), main_stream); + break; + } +#endif + case GGML_TYPE_F32: + { + auto data_pts = cast_data(dst); + silu_sycl(data_pts.src, data_pts.dst, ggml_nelements(dst->src[0]), main_stream); + break; + } + default: + GGML_ABORT("GGML tensor type not supported!\n"); + break; + } } -inline void ggml_sycl_op_tanh(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1, - ggml_tensor *dst, const float *src0_dd, - const float *src1_dd, float *dst_dd, - const queue_ptr &main_stream) { - - GGML_ASSERT(src0->type == GGML_TYPE_F32); - GGML_ASSERT( dst->type == GGML_TYPE_F32); - tanh_f32_sycl(src0_dd, dst_dd, ggml_nelements(src0), main_stream); - - GGML_UNUSED(src1); - GGML_UNUSED(dst); - GGML_UNUSED(src1_dd); - GGML_UNUSED(ctx); +inline void ggml_sycl_op_gelu(ggml_backend_sycl_context & ctx, ggml_tensor * dst) { +#if defined (GGML_SYCL_F16) + GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32 || dst->src[0]->type == GGML_TYPE_F16); + GGML_ASSERT(dst->type == GGML_TYPE_F32 || dst->type == GGML_TYPE_F16); +#else + GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32); + GGML_ASSERT(dst->type == GGML_TYPE_F32); +#endif + GGML_ASSERT(dst->src[0]->type == dst->type); + dpct::queue_ptr main_stream = ctx.stream(); + SYCL_CHECK(ggml_sycl_set_device(ctx.device)); + switch (dst->type) { +#if defined (GGML_SYCL_F16) + case GGML_TYPE_F16: + { + auto data_pts = cast_data(dst); + gelu_sycl(data_pts.src, data_pts.dst, ggml_nelements(dst->src[0]), main_stream); + break; + } +#endif + case GGML_TYPE_F32: + { + auto data_pts = cast_data(dst); + gelu_sycl(data_pts.src, data_pts.dst, ggml_nelements(dst->src[0]), main_stream); + break; + } + default: + GGML_ABORT("GGML tensor type not supported!\n"); + break; + } } -inline void ggml_sycl_op_relu(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1, - ggml_tensor *dst, const float *src0_dd, - const float *src1_dd, float *dst_dd, - const queue_ptr &main_stream) { - - GGML_ASSERT(src0->type == GGML_TYPE_F32); - GGML_ASSERT( dst->type == GGML_TYPE_F32); - - relu_f32_sycl(src0_dd, dst_dd, ggml_nelements(src0), main_stream); - - GGML_UNUSED(src1); - GGML_UNUSED(dst); - GGML_UNUSED(src1_dd); - GGML_UNUSED(ctx); +inline void ggml_sycl_op_gelu_quick(ggml_backend_sycl_context & ctx, ggml_tensor *dst) { +#if defined (GGML_SYCL_F16) + GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32 || dst->src[0]->type == GGML_TYPE_F16); + GGML_ASSERT(dst->type == GGML_TYPE_F32 || dst->type == GGML_TYPE_F16); +#else + GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32); + GGML_ASSERT(dst->type == GGML_TYPE_F32); +#endif + GGML_ASSERT(dst->src[0]->type == dst->type); + dpct::queue_ptr main_stream = ctx.stream(); + SYCL_CHECK(ggml_sycl_set_device(ctx.device)); + switch (dst->type) { +#if defined (GGML_SYCL_F16) + case GGML_TYPE_F16: + { + auto data_pts = cast_data(dst); + gelu_quick_sycl(data_pts.src, data_pts.dst, ggml_nelements(dst->src[0]), main_stream); + break; + } +#endif + case GGML_TYPE_F32: + { + auto data_pts = cast_data(dst); + gelu_quick_sycl(data_pts.src, data_pts.dst, ggml_nelements(dst->src[0]), main_stream); + break; + } + default: + GGML_ABORT("GGML tensor type not supported!\n"); + break; + } } -inline void ggml_sycl_op_hardsigmoid(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, - const ggml_tensor *src1, ggml_tensor *dst, - const float *src0_dd, const float *src1_dd, - float *dst_dd, - const queue_ptr &main_stream) { - - GGML_ASSERT(src0->type == GGML_TYPE_F32); - GGML_ASSERT( dst->type == GGML_TYPE_F32); - - hardsigmoid_f32_sycl(src0_dd, dst_dd, ggml_nelements(src0), main_stream); - - GGML_UNUSED(src1); - GGML_UNUSED(dst); - GGML_UNUSED(src1_dd); - GGML_UNUSED(ctx); +inline void ggml_sycl_op_tanh(ggml_backend_sycl_context & ctx, ggml_tensor * dst) { +#if defined (GGML_SYCL_F16) + GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32 || dst->src[0]->type == GGML_TYPE_F16); + GGML_ASSERT(dst->type == GGML_TYPE_F32 || dst->type == GGML_TYPE_F16); +#else + GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32); + GGML_ASSERT(dst->type == GGML_TYPE_F32); +#endif + GGML_ASSERT(dst->src[0]->type == dst->type); + dpct::queue_ptr main_stream = ctx.stream(); + SYCL_CHECK(ggml_sycl_set_device(ctx.device)); + switch (dst->type) { +#if defined (GGML_SYCL_F16) + case GGML_TYPE_F16: + { + auto data_pts = cast_data(dst); + tanh_sycl(data_pts.src, data_pts.dst, ggml_nelements(dst->src[0]), main_stream); + break; + } +#endif + case GGML_TYPE_F32: + { + auto data_pts = cast_data(dst); + tanh_sycl(data_pts.src, data_pts.dst, ggml_nelements(dst->src[0]), main_stream); + break; + } + default: + GGML_ABORT("GGML tensor type not supported!\n"); + break; + } } -inline void ggml_sycl_op_hardswish(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, - const ggml_tensor *src1, ggml_tensor *dst, - const float *src0_dd, const float *src1_dd, - float *dst_dd, const queue_ptr &main_stream) { +inline void ggml_sycl_op_relu(ggml_backend_sycl_context & ctx, ggml_tensor * dst) { +#if defined (GGML_SYCL_F16) + GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32 || dst->src[0]->type == GGML_TYPE_F16); + GGML_ASSERT(dst->type == GGML_TYPE_F32 || dst->type == GGML_TYPE_F16); +#else + GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32); + GGML_ASSERT(dst->type == GGML_TYPE_F32); +#endif + GGML_ASSERT(dst->src[0]->type == dst->type); + dpct::queue_ptr main_stream = ctx.stream(); + SYCL_CHECK(ggml_sycl_set_device(ctx.device)); - GGML_ASSERT(src0->type == GGML_TYPE_F32); - GGML_ASSERT( dst->type == GGML_TYPE_F32); - - hardswish_f32_sycl(src0_dd, dst_dd, ggml_nelements(src0), main_stream); - - GGML_UNUSED(src1); - GGML_UNUSED(dst); - GGML_UNUSED(src1_dd); - GGML_UNUSED(ctx); + switch (dst->type) { +#if defined (GGML_SYCL_F16) + case GGML_TYPE_F16: + { + auto data_pts = cast_data(dst); + relu_sycl(data_pts.src, data_pts.dst, ggml_nelements(dst->src[0]), main_stream); + break; + } +#endif + case GGML_TYPE_F32: + { + auto data_pts = cast_data(dst); + relu_sycl(data_pts.src, data_pts.dst, ggml_nelements(dst->src[0]), main_stream); + break; + } + default: + GGML_ABORT("GGML tensor type not supported!\n"); + break; + } } -inline void ggml_sycl_op_exp(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, - const ggml_tensor *src1, ggml_tensor *dst, - const float *src0_dd, const float *src1_dd, - float *dst_dd, const queue_ptr &main_stream) { +inline void ggml_sycl_op_hardsigmoid(ggml_backend_sycl_context & ctx, ggml_tensor * dst) { +#if defined (GGML_SYCL_F16) + GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32 || dst->src[0]->type == GGML_TYPE_F16); + GGML_ASSERT(dst->type == GGML_TYPE_F32 || dst->type == GGML_TYPE_F16); +#else + GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32); + GGML_ASSERT(dst->type == GGML_TYPE_F32); +#endif + GGML_ASSERT(dst->src[0]->type == dst->type); - GGML_ASSERT(src0->type == GGML_TYPE_F32); - GGML_ASSERT( dst->type == GGML_TYPE_F32); + dpct::queue_ptr main_stream = ctx.stream(); + SYCL_CHECK(ggml_sycl_set_device(ctx.device)); - exp_f32_sycl(src0_dd, dst_dd, ggml_nelements(src0), main_stream); - - GGML_UNUSED(src1); - GGML_UNUSED(dst); - GGML_UNUSED(src1_dd); - GGML_UNUSED(ctx); + switch (dst->type) { +#if defined (GGML_SYCL_F16) + case GGML_TYPE_F16: + { + auto data_pts = cast_data(dst); + hardsigmoid_sycl(data_pts.src, data_pts.dst, ggml_nelements(dst->src[0]), main_stream); + break; + } +#endif + case GGML_TYPE_F32: + { + auto data_pts = cast_data(dst); + hardsigmoid_sycl(data_pts.src, data_pts.dst, ggml_nelements(dst->src[0]), main_stream); + break; + } + default: + GGML_ABORT("GGML tensor type not supported!\n"); + break; + } } -inline void ggml_sycl_op_log(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, - const ggml_tensor *src1, ggml_tensor *dst, - const float *src0_dd, const float *src1_dd, - float *dst_dd, const queue_ptr &main_stream) { - - GGML_ASSERT(src0->type == GGML_TYPE_F32); - GGML_ASSERT( dst->type == GGML_TYPE_F32); - - log_f32_sycl(src0_dd, dst_dd, ggml_nelements(src0), main_stream); - - GGML_UNUSED(src1); - GGML_UNUSED(dst); - GGML_UNUSED(src1_dd); - GGML_UNUSED(ctx); +inline void ggml_sycl_op_hardswish(ggml_backend_sycl_context & ctx, ggml_tensor * dst) { +#if defined (GGML_SYCL_F16) + GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32 || dst->src[0]->type == GGML_TYPE_F16); + GGML_ASSERT(dst->type == GGML_TYPE_F32 || dst->type == GGML_TYPE_F16); +#else + GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32); + GGML_ASSERT(dst->type == GGML_TYPE_F32); +#endif + GGML_ASSERT(dst->src[0]->type == dst->type); + dpct::queue_ptr main_stream = ctx.stream(); + SYCL_CHECK(ggml_sycl_set_device(ctx.device)); + switch (dst->type) { +#if defined (GGML_SYCL_F16) + case GGML_TYPE_F16: + { + auto data_pts = cast_data(dst); + hardswish_sycl(data_pts.src, data_pts.dst, ggml_nelements(dst->src[0]), main_stream); + break; + } +#endif + case GGML_TYPE_F32: + { + auto data_pts = cast_data(dst); + hardswish_sycl(data_pts.src, data_pts.dst, ggml_nelements(dst->src[0]), main_stream); + break; + } + default: + GGML_ABORT("GGML tensor type not supported!\n"); + break; + } } -inline void ggml_sycl_op_sigmoid(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, - const ggml_tensor *src1, ggml_tensor *dst, - const float *src0_dd, const float *src1_dd, - float *dst_dd, const queue_ptr &main_stream) { - - GGML_ASSERT(src0->type == GGML_TYPE_F32); - GGML_ASSERT( dst->type == GGML_TYPE_F32); - - sigmoid_f32_sycl(src0_dd, dst_dd, ggml_nelements(src0), main_stream); - - GGML_UNUSED(src1); - GGML_UNUSED(dst); - GGML_UNUSED(src1_dd); - GGML_UNUSED(ctx); +inline void ggml_sycl_op_exp(ggml_backend_sycl_context & ctx, ggml_tensor * dst) { +#if defined (GGML_SYCL_F16) + GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32 || dst->src[0]->type == GGML_TYPE_F16); + GGML_ASSERT(dst->type == GGML_TYPE_F32 || dst->type == GGML_TYPE_F16); +#else + GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32); + GGML_ASSERT(dst->type == GGML_TYPE_F32); +#endif + GGML_ASSERT(dst->src[0]->type == dst->type); + dpct::queue_ptr main_stream = ctx.stream(); + SYCL_CHECK(ggml_sycl_set_device(ctx.device)); + switch (dst->type) { +#if defined (GGML_SYCL_F16) + case GGML_TYPE_F16: + { + auto data_pts = cast_data(dst); + exp_sycl(data_pts.src, data_pts.dst, ggml_nelements(dst->src[0]), main_stream); + break; + } +#endif + case GGML_TYPE_F32: + { + auto data_pts = cast_data(dst); + exp_sycl(data_pts.src, data_pts.dst, ggml_nelements(dst->src[0]), main_stream); + break; + } + default: + GGML_ABORT("GGML tensor type not supported!\n"); + break; + } } -inline void ggml_sycl_op_sqrt(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, - const ggml_tensor *src1, ggml_tensor *dst, - const float *src0_dd, const float *src1_dd, - float *dst_dd, const queue_ptr &main_stream) { - - GGML_ASSERT(src0->type == GGML_TYPE_F32); - GGML_ASSERT( dst->type == GGML_TYPE_F32); - - sqrt_f32_sycl(src0_dd, dst_dd, ggml_nelements(src0), main_stream); - - GGML_UNUSED(src1); - GGML_UNUSED(dst); - GGML_UNUSED(src1_dd); - GGML_UNUSED(ctx); +inline void ggml_sycl_op_log(ggml_backend_sycl_context & ctx, ggml_tensor * dst) { +#if defined (GGML_SYCL_F16) + GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32 || dst->src[0]->type == GGML_TYPE_F16); + GGML_ASSERT(dst->type == GGML_TYPE_F32 || dst->type == GGML_TYPE_F16); +#else + GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32); + GGML_ASSERT(dst->type == GGML_TYPE_F32); +#endif + GGML_ASSERT(dst->src[0]->type == dst->type); + dpct::queue_ptr main_stream = ctx.stream(); + SYCL_CHECK(ggml_sycl_set_device(ctx.device)); + switch (dst->type) { +#if defined (GGML_SYCL_F16) + case GGML_TYPE_F16: + { + auto data_pts = cast_data(dst); + log_sycl(data_pts.src, data_pts.dst, ggml_nelements(dst->src[0]), main_stream); + break; + } +#endif + case GGML_TYPE_F32: + { + auto data_pts = cast_data(dst); + log_sycl(data_pts.src, data_pts.dst, ggml_nelements(dst->src[0]), main_stream); + break; + } + default: + GGML_ABORT("GGML tensor type not supported!\n"); + break; + } } -inline void ggml_sycl_op_sin(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, - const ggml_tensor *src1, ggml_tensor *dst, - const float *src0_dd, const float *src1_dd, - float *dst_dd, const queue_ptr &main_stream) { - - GGML_ASSERT(src0->type == GGML_TYPE_F32); - GGML_ASSERT( dst->type == GGML_TYPE_F32); - - sin_f32_sycl(src0_dd, dst_dd, ggml_nelements(src0), main_stream); - - GGML_UNUSED(src1); - GGML_UNUSED(dst); - GGML_UNUSED(src1_dd); - GGML_UNUSED(ctx); +inline void ggml_sycl_op_sigmoid(ggml_backend_sycl_context & ctx, ggml_tensor * dst) { +#if defined (GGML_SYCL_F16) + GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32 || dst->src[0]->type == GGML_TYPE_F16); + GGML_ASSERT(dst->type == GGML_TYPE_F32 || dst->type == GGML_TYPE_F16); +#else + GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32); + GGML_ASSERT(dst->type == GGML_TYPE_F32); +#endif + GGML_ASSERT(dst->src[0]->type == dst->type); + dpct::queue_ptr main_stream = ctx.stream(); + SYCL_CHECK(ggml_sycl_set_device(ctx.device)); + switch (dst->type) { +#if defined (GGML_SYCL_F16) + case GGML_TYPE_F16: + { + auto data_pts = cast_data(dst); + sigmoid_sycl(data_pts.src, data_pts.dst, ggml_nelements(dst->src[0]), main_stream); + break; + } +#endif + case GGML_TYPE_F32: + { + auto data_pts = cast_data(dst); + sigmoid_sycl(data_pts.src, data_pts.dst, ggml_nelements(dst->src[0]), main_stream); + break; + } + default: + GGML_ABORT("GGML tensor type not supported!\n"); + break; + } } -inline void ggml_sycl_op_cos(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, - const ggml_tensor *src1, ggml_tensor *dst, - const float *src0_dd, const float *src1_dd, - float *dst_dd, const queue_ptr &main_stream) { +inline void ggml_sycl_op_sqrt(ggml_backend_sycl_context & ctx, ggml_tensor * dst) { +#if defined (GGML_SYCL_F16) + GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32 || dst->src[0]->type == GGML_TYPE_F16); + GGML_ASSERT(dst->type == GGML_TYPE_F32 || dst->type == GGML_TYPE_F16); +#else + GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32); + GGML_ASSERT(dst->type == GGML_TYPE_F32); +#endif + GGML_ASSERT(dst->src[0]->type == dst->type); - GGML_ASSERT(src0->type == GGML_TYPE_F32); - GGML_ASSERT( dst->type == GGML_TYPE_F32); - - cos_f32_sycl(src0_dd, dst_dd, ggml_nelements(src0), main_stream); - - GGML_UNUSED(src1); - GGML_UNUSED(dst); - GGML_UNUSED(src1_dd); - GGML_UNUSED(ctx); + dpct::queue_ptr main_stream = ctx.stream(); + SYCL_CHECK(ggml_sycl_set_device(ctx.device)); + switch (dst->type) { +#if defined (GGML_SYCL_F16) + case GGML_TYPE_F16: + { + auto data_pts = cast_data(dst); + sqrt_sycl(data_pts.src, data_pts.dst, ggml_nelements(dst->src[0]), main_stream); + break; + } +#endif + case GGML_TYPE_F32: + { + auto data_pts = cast_data(dst); + sqrt_sycl(data_pts.src, data_pts.dst, ggml_nelements(dst->src[0]), main_stream); + break; + } + default: + GGML_ABORT("GGML tensor type not supported!\n"); + break; + } } -inline void ggml_sycl_op_step(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, - const ggml_tensor *src1, ggml_tensor *dst, - const float *src0_dd, const float *src1_dd, - float *dst_dd, const queue_ptr &main_stream) { - - GGML_ASSERT(src0->type == GGML_TYPE_F32); - GGML_ASSERT( dst->type == GGML_TYPE_F32); - - step_f32_sycl(src0_dd, dst_dd, ggml_nelements(src0), main_stream); - - GGML_UNUSED(src1); - GGML_UNUSED(dst); - GGML_UNUSED(src1_dd); - GGML_UNUSED(ctx); +inline void ggml_sycl_op_sin(ggml_backend_sycl_context & ctx, ggml_tensor * dst) { +#if defined (GGML_SYCL_F16) + GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32 || dst->src[0]->type == GGML_TYPE_F16); + GGML_ASSERT(dst->type == GGML_TYPE_F32 || dst->type == GGML_TYPE_F16); +#else + GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32); + GGML_ASSERT(dst->type == GGML_TYPE_F32); +#endif + GGML_ASSERT(dst->src[0]->type == dst->type); + dpct::queue_ptr main_stream = ctx.stream(); + SYCL_CHECK(ggml_sycl_set_device(ctx.device)); + switch (dst->type) { +#if defined (GGML_SYCL_F16) + case GGML_TYPE_F16: + { + auto data_pts = cast_data(dst); + sin_sycl(data_pts.src, data_pts.dst, ggml_nelements(dst->src[0]), main_stream); + break; + } +#endif + case GGML_TYPE_F32: + { + auto data_pts = cast_data(dst); + sin_sycl(data_pts.src, data_pts.dst, ggml_nelements(dst->src[0]), main_stream); + break; + } + default: + GGML_ABORT("GGML tensor type not supported!\n"); + break; + } } -inline void ggml_sycl_op_neg(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, - const ggml_tensor *src1, ggml_tensor *dst, - const float *src0_dd, const float *src1_dd, - float *dst_dd, const queue_ptr &main_stream) { - - GGML_ASSERT(src0->type == GGML_TYPE_F32); - GGML_ASSERT( dst->type == GGML_TYPE_F32); - - neg_f32_sycl(src0_dd, dst_dd, ggml_nelements(src0), main_stream); - - GGML_UNUSED(src1); - GGML_UNUSED(dst); - GGML_UNUSED(src1_dd); - GGML_UNUSED(ctx); +inline void ggml_sycl_op_cos(ggml_backend_sycl_context & ctx, ggml_tensor * dst) { +#if defined (GGML_SYCL_F16) + GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32 || dst->src[0]->type == GGML_TYPE_F16); + GGML_ASSERT(dst->type == GGML_TYPE_F32 || dst->type == GGML_TYPE_F16); +#else + GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32); + GGML_ASSERT(dst->type == GGML_TYPE_F32); +#endif + GGML_ASSERT(dst->src[0]->type == dst->type); + dpct::queue_ptr main_stream = ctx.stream(); + SYCL_CHECK(ggml_sycl_set_device(ctx.device)); + switch (dst->type) { +#if defined (GGML_SYCL_F16) + case GGML_TYPE_F16: + { + auto data_pts = cast_data(dst); + cos_sycl(data_pts.src, data_pts.dst, ggml_nelements(dst->src[0]), main_stream); + break; + } +#endif + case GGML_TYPE_F32: + { + auto data_pts = cast_data(dst); + cos_sycl(data_pts.src, data_pts.dst, ggml_nelements(dst->src[0]), main_stream); + break; + } + default: + GGML_ABORT("GGML tensor type not supported!\n"); + break; + } } -inline void ggml_sycl_op_leaky_relu(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, - const ggml_tensor *src1, ggml_tensor *dst, - const float *src0_dd, const float *src1_dd, - float *dst_dd, - const queue_ptr &main_stream) { +inline void ggml_sycl_op_step(ggml_backend_sycl_context & ctx, ggml_tensor * dst) { +#if defined (GGML_SYCL_F16) + GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32 || dst->src[0]->type == GGML_TYPE_F16); + GGML_ASSERT(dst->type == GGML_TYPE_F32 || dst->type == GGML_TYPE_F16); +#else + GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32); + GGML_ASSERT(dst->type == GGML_TYPE_F32); +#endif + GGML_ASSERT(dst->src[0]->type == dst->type); + dpct::queue_ptr main_stream = ctx.stream(); + SYCL_CHECK(ggml_sycl_set_device(ctx.device)); + switch (dst->type) { +#if defined (GGML_SYCL_F16) + case GGML_TYPE_F16: + { + auto data_pts = cast_data(dst); + step_sycl(data_pts.src, data_pts.dst, ggml_nelements(dst->src[0]), main_stream); + break; + } +#endif + case GGML_TYPE_F32: + { + auto data_pts = cast_data(dst); + step_sycl(data_pts.src, data_pts.dst, ggml_nelements(dst->src[0]), main_stream); + break; + } + default: + GGML_ABORT("GGML tensor type not supported!\n"); + break; + } +} - GGML_ASSERT(src0->type == GGML_TYPE_F32); - GGML_ASSERT( dst->type == GGML_TYPE_F32); +inline void ggml_sycl_op_neg(ggml_backend_sycl_context & ctx, ggml_tensor * dst) { +#if defined (GGML_SYCL_F16) + GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32 || dst->src[0]->type == GGML_TYPE_F16); + GGML_ASSERT(dst->type == GGML_TYPE_F32 || dst->type == GGML_TYPE_F16); +#else + GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32); + GGML_ASSERT(dst->type == GGML_TYPE_F32); +#endif + GGML_ASSERT(dst->src[0]->type == dst->type); + dpct::queue_ptr main_stream = ctx.stream(); + SYCL_CHECK(ggml_sycl_set_device(ctx.device)); + switch (dst->type) { +#if defined (GGML_SYCL_F16) + case GGML_TYPE_F16: + { + auto data_pts = cast_data(dst); + neg_sycl(data_pts.src, data_pts.dst, ggml_nelements(dst->src[0]), main_stream); + break; + } +#endif + case GGML_TYPE_F32: + { + auto data_pts = cast_data(dst); + neg_sycl(data_pts.src, data_pts.dst, ggml_nelements(dst->src[0]), main_stream); + break; + } + default: + GGML_ABORT("GGML tensor type not supported!\n"); + break; + } +} +inline void ggml_sycl_op_leaky_relu(ggml_backend_sycl_context & ctx, ggml_tensor * dst) { +#if defined (GGML_SYCL_F16) + GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32 || dst->src[0]->type == GGML_TYPE_F16); + GGML_ASSERT(dst->type == GGML_TYPE_F32 || dst->type == GGML_TYPE_F16); +#else + GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32); + GGML_ASSERT(dst->type == GGML_TYPE_F32); +#endif + + GGML_ASSERT(dst->src[0]->type == dst->type); float negative_slope; memcpy(&negative_slope, dst->op_params, sizeof(float)); - - leaky_relu_f32_sycl(src0_dd, dst_dd, ggml_nelements(src0), negative_slope, main_stream); - - GGML_UNUSED(src1); - GGML_UNUSED(dst); - GGML_UNUSED(src1_dd); - GGML_UNUSED(ctx); + dpct::queue_ptr main_stream = ctx.stream(); + SYCL_CHECK(ggml_sycl_set_device(ctx.device)); + switch (dst->type) { +#if defined (GGML_SYCL_F16) + case GGML_TYPE_F16: + { + auto data_pts = cast_data(dst); + leaky_relu_sycl(data_pts.src, data_pts.dst, ggml_nelements(dst->src[0]), negative_slope, main_stream); + break; + } +#endif + case GGML_TYPE_F32: + { + auto data_pts = cast_data(dst); + leaky_relu_sycl(data_pts.src, data_pts.dst, ggml_nelements(dst->src[0]), negative_slope, main_stream); + break; + } + default: + GGML_ABORT("GGML tensor type not supported!\n"); + break; + } } -inline void ggml_sycl_op_sqr(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1, - ggml_tensor *dst, const float *src0_dd, - const float *src1_dd, float *dst_dd, - const queue_ptr &main_stream) { - - GGML_ASSERT(src0->type == GGML_TYPE_F32); - GGML_ASSERT( dst->type == GGML_TYPE_F32); - - sqr_f32_sycl(src0_dd, dst_dd, ggml_nelements(src0), main_stream); - - GGML_UNUSED(src1); - GGML_UNUSED(dst); - GGML_UNUSED(src1_dd); - GGML_UNUSED(ctx); -} - -inline void ggml_sycl_op_upscale(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, - const ggml_tensor *src1, ggml_tensor *dst, - const float *src0_dd, const float *src1_dd, - float *dst_dd, - const queue_ptr &main_stream) { - - GGML_ASSERT(src0->type == GGML_TYPE_F32); +inline void ggml_sycl_op_sqr(ggml_backend_sycl_context & ctx, ggml_tensor * dst) { + #if defined (GGML_SYCL_F16) + GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32 || dst->src[0]->type == GGML_TYPE_F16); + GGML_ASSERT(dst->type == GGML_TYPE_F32 || dst->type == GGML_TYPE_F16); +#else + GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32); GGML_ASSERT(dst->type == GGML_TYPE_F32); - - const float sf0 = (float)dst->ne[0]/src0->ne[0]; - const float sf1 = (float)dst->ne[1]/src0->ne[1]; - const float sf2 = (float)dst->ne[2]/src0->ne[2]; - const float sf3 = (float)dst->ne[3]/src0->ne[3]; - - upscale_f32_sycl(src0_dd, dst_dd, src0->nb[0], src0->nb[1], src0->nb[2], src0->nb[3], - dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3], sf0, sf1, sf2, sf3, - main_stream); - - GGML_UNUSED(src1); - GGML_UNUSED(dst); - GGML_UNUSED(src1_dd); - GGML_UNUSED(ctx); +#endif + GGML_ASSERT(dst->src[0]->type == dst->type); + dpct::queue_ptr main_stream = ctx.stream(); + SYCL_CHECK(ggml_sycl_set_device(ctx.device)); + switch (dst->type) { +#if defined (GGML_SYCL_F16) + case GGML_TYPE_F16: + { + auto data_pts = cast_data(dst); + sqr_sycl(data_pts.src, data_pts.dst, ggml_nelements(dst->src[0]), main_stream); + break; + } +#endif + case GGML_TYPE_F32: + { + auto data_pts = cast_data(dst); + sqr_sycl(data_pts.src, data_pts.dst, ggml_nelements(dst->src[0]), main_stream); + break; + } + default: + GGML_ABORT("GGML tensor type not supported!\n"); + break; + } } -inline void ggml_sycl_op_pad(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1, - ggml_tensor *dst, const float *src0_dd, - const float *src1_dd, float *dst_dd, - const queue_ptr &main_stream) { - - GGML_ASSERT(src0->type == GGML_TYPE_F32); +inline void ggml_sycl_op_upscale(ggml_backend_sycl_context & ctx, ggml_tensor * dst) { +#if defined (GGML_SYCL_F16) + GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32 || dst->src[0]->type == GGML_TYPE_F16); + GGML_ASSERT(dst->type == GGML_TYPE_F32 || dst->type == GGML_TYPE_F16); +#else + GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32); GGML_ASSERT(dst->type == GGML_TYPE_F32); - GGML_ASSERT(src0->ne[3] == 1 && dst->ne[3] == 1); // just 3D tensors +#endif + GGML_ASSERT(dst->src[0]->type == dst->type); - pad_f32_sycl(src0_dd, dst_dd, - src0->ne[0], src0->ne[1], src0->ne[2], - dst->ne[0], dst->ne[1], dst->ne[2], main_stream); + dpct::queue_ptr main_stream = ctx.stream(); + SYCL_CHECK(ggml_sycl_set_device(ctx.device)); - GGML_UNUSED(src1); - GGML_UNUSED(dst); - GGML_UNUSED(src1_dd); - GGML_UNUSED(ctx); + const float sf0 = (float) dst->ne[0] / dst->src[0]->ne[0]; + const float sf1 = (float) dst->ne[1] / dst->src[0]->ne[1]; + const float sf2 = (float) dst->ne[2] / dst->src[0]->ne[2]; + const float sf3 = (float) dst->ne[3] / dst->src[0]->ne[3]; + switch (dst->type) { +#if defined (GGML_SYCL_F16) + case GGML_TYPE_F16: + { + auto data_pts = cast_data(dst); + upscale_sycl(data_pts.src, data_pts.dst, dst->src[0]->nb[0], dst->src[0]->nb[1], dst->src[0]->nb[2], + dst->src[0]->nb[3], dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3], sf0, sf1, sf2, sf3, + main_stream); + break; + } +#endif + case GGML_TYPE_F32: + { + auto data_pts = cast_data(dst); + upscale_sycl(data_pts.src, data_pts.dst, dst->src[0]->nb[0], dst->src[0]->nb[1], dst->src[0]->nb[2], + dst->src[0]->nb[3], dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3], sf0, sf1, sf2, sf3, + main_stream); + break; + } + default: + GGML_ABORT("GGML tensor type not supported!\n"); + break; + } } -inline void ggml_sycl_op_acc(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1, - ggml_tensor *dst, const float *src0_dd, - const float *src1_dd, float *dst_dd, - const queue_ptr &main_stream) { +inline void ggml_sycl_op_pad(ggml_backend_sycl_context & ctx, ggml_tensor * dst) { +#if defined (GGML_SYCL_F16) + GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32 || dst->src[0]->type == GGML_TYPE_F16); + GGML_ASSERT(dst->type == GGML_TYPE_F32 || dst->type == GGML_TYPE_F16); +#else + GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32); + GGML_ASSERT(dst->type == GGML_TYPE_F32); +#endif + GGML_ASSERT(dst->src[0]->type == dst->type); + GGML_ASSERT(dst->src[0]->ne[3] == 1 && dst->ne[3] == 1); // just 3D tensors + dpct::queue_ptr main_stream = ctx.stream(); + SYCL_CHECK(ggml_sycl_set_device(ctx.device)); + switch (dst->type) { +#if defined (GGML_SYCL_F16) + case GGML_TYPE_F16: + { + auto data_pts = cast_data(dst); + pad_sycl(data_pts.src, data_pts.dst, dst->src[0]->ne[0], dst->src[0]->ne[1], dst->src[0]->ne[2], dst->ne[0], + dst->ne[1], dst->ne[2], main_stream); + break; + } +#endif + case GGML_TYPE_F32: + { + auto data_pts = cast_data(dst); + pad_sycl(data_pts.src, data_pts.dst, dst->src[0]->ne[0], dst->src[0]->ne[1], dst->src[0]->ne[2], dst->ne[0], + dst->ne[1], dst->ne[2], main_stream); + break; + } + default: + GGML_ABORT("GGML tensor type not supported!\n"); + break; + } +} - GGML_ASSERT(src0->type == GGML_TYPE_F32); - GGML_ASSERT(src1->type == GGML_TYPE_F32); +inline void ggml_sycl_op_clamp(ggml_backend_sycl_context & ctx, ggml_tensor * dst) { +#if defined(GGML_SYCL_F16) + GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32 || dst->src[0]->type == GGML_TYPE_F16); + GGML_ASSERT(dst->type == GGML_TYPE_F32 || dst->type == GGML_TYPE_F16); +#else + + GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32); + GGML_ASSERT(dst->type == GGML_TYPE_F32); +#endif + GGML_ASSERT(dst->src[0]->type == dst->type); + dpct::queue_ptr main_stream = ctx.stream(); + SYCL_CHECK(ggml_sycl_set_device(ctx.device)); + float min; + float max; + memcpy(&min, dst->op_params, sizeof(float)); + memcpy(&max, (float *) dst->op_params + 1, sizeof(float)); + + switch (dst->type) { +#if defined(GGML_SYCL_F16) + case GGML_TYPE_F16: + { + auto data_pts = cast_data(dst); + clamp_sycl(data_pts.src, data_pts.dst, min, max, ggml_nelements(dst->src[0]), main_stream); + break; + } +#endif + case GGML_TYPE_F32: + { + auto data_pts = cast_data(dst); + clamp_sycl(data_pts.src, data_pts.dst, min, max, ggml_nelements(dst->src[0]), main_stream); + break; + } + default: + GGML_ABORT("GGML tensor type not supported!\n"); + break; + } +} + +inline void ggml_sycl_op_acc(ggml_backend_sycl_context & ctx, ggml_tensor *dst) { + + GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32); + GGML_ASSERT(dst->src[1]->type == GGML_TYPE_F32); GGML_ASSERT( dst->type == GGML_TYPE_F32); GGML_ASSERT(dst->ne[3] == 1); // just 3D tensors supported + dpct::queue_ptr main_stream = ctx.stream(); + SYCL_CHECK(ggml_sycl_set_device(ctx.device)); + const float * src0_dd = static_cast(dst->src[0]->data); + const float * src1_dd = static_cast(dst->src[1]->data); + float * dst_dd = static_cast(dst->data); int nb1 = dst->op_params[0] / 4; // 4 bytes of float32 int nb2 = dst->op_params[1] / 4; // 4 bytes of float32 // int nb3 = dst->op_params[2] / 4; // 4 bytes of float32 - unused int offset = dst->op_params[3] / 4; // offset in bytes - acc_f32_sycl(src0_dd, src1_dd, dst_dd, ggml_nelements(dst), src1->ne[0], src1->ne[1], src1->ne[2], nb1, nb2, offset, main_stream); - - GGML_UNUSED(dst); - GGML_UNUSED(ctx); + acc_f32_sycl(src0_dd, src1_dd, dst_dd, ggml_nelements(dst), dst->src[1]->ne[0], dst->src[1]->ne[1], dst->src[1]->ne[2], nb1, nb2, offset, main_stream); } -inline void ggml_sycl_op_add(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1, - ggml_tensor *dst, const float *src0_dd, - const float *src1_dd, float *dst_dd, - const queue_ptr &main_stream) { - ggml_sycl_op_bin_bcast>(ctx, src0, src1, dst, src0_dd, src1_dd, dst_dd, main_stream); +inline void ggml_sycl_op_add(ggml_backend_sycl_context & ctx, ggml_tensor *dst) { + + ggml_sycl_op_bin_bcast>(ctx, dst->src[0], dst->src[1], dst); } -inline void ggml_sycl_op_sub(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1, - ggml_tensor *dst, const float *src0_dd, - const float *src1_dd, float *dst_dd, - const queue_ptr &main_stream) { +inline void ggml_sycl_op_sub(ggml_backend_sycl_context & ctx, ggml_tensor *dst) { - ggml_sycl_op_bin_bcast>(ctx, src0, src1, dst, src0_dd, src1_dd, dst_dd, main_stream); + ggml_sycl_op_bin_bcast>(ctx, dst->src[0], dst->src[1], dst); } -inline void ggml_sycl_op_mul(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1, - ggml_tensor *dst, const float *src0_dd, - const float *src1_dd, float *dst_dd, - const queue_ptr &main_stream) { +inline void ggml_sycl_op_mul(ggml_backend_sycl_context & ctx, ggml_tensor *dst) { - ggml_sycl_op_bin_bcast>(ctx, src0, src1, dst, src0_dd, src1_dd, dst_dd, main_stream); + ggml_sycl_op_bin_bcast>(ctx, dst->src[0], dst->src[1], dst); } -inline void ggml_sycl_op_div(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1, - ggml_tensor *dst, const float *src0_dd, - const float *src1_dd, float *dst_dd, - const queue_ptr &main_stream) { +inline void ggml_sycl_op_div(ggml_backend_sycl_context & ctx, ggml_tensor *dst) { - ggml_sycl_op_bin_bcast>(ctx, src0, src1, dst, src0_dd, src1_dd, dst_dd, main_stream); + ggml_sycl_op_bin_bcast>(ctx, dst->src[0], dst->src[1], dst); } void ggml_sycl_sqrt(ggml_backend_sycl_context & ctx, ggml_tensor * dst) { - GGML_SYCL_DEBUG("call %s\n", __func__); - ggml_sycl_op_flatten(ctx, dst->src[0], dst->src[1], dst, ggml_sycl_op_sqrt); + GGML_SYCL_DEBUG("call %s: DST Tensor type: %s\n", __func__, ggml_type_name(dst->type)); + ggml_sycl_op_sqrt(ctx, dst); GGML_SYCL_DEBUG("call %s done\n", __func__); } void ggml_sycl_sin(ggml_backend_sycl_context & ctx, ggml_tensor * dst) { - GGML_SYCL_DEBUG("call %s\n", __func__); - ggml_sycl_op_flatten(ctx, dst->src[0], dst->src[1], dst, ggml_sycl_op_sin); + GGML_SYCL_DEBUG("call %s: DST Tensor type: %s\n", __func__, ggml_type_name(dst->type)); + ggml_sycl_op_sin(ctx, dst); GGML_SYCL_DEBUG("call %s done\n", __func__); } void ggml_sycl_cos(ggml_backend_sycl_context & ctx, ggml_tensor * dst) { - GGML_SYCL_DEBUG("call %s\n", __func__); - ggml_sycl_op_flatten(ctx, dst->src[0], dst->src[1], dst, ggml_sycl_op_cos); + GGML_SYCL_DEBUG("call %s: DST Tensor type: %s\n", __func__, ggml_type_name(dst->type)); + ggml_sycl_op_cos(ctx, dst); GGML_SYCL_DEBUG("call %s done\n", __func__); } void ggml_sycl_acc(ggml_backend_sycl_context & ctx, ggml_tensor * dst) { - GGML_SYCL_DEBUG("call %s\n", __func__); - ggml_sycl_op_flatten(ctx, dst->src[0], dst->src[1], dst, ggml_sycl_op_acc); + GGML_SYCL_DEBUG("call %s: DST Tensor type: %s\n", __func__, ggml_type_name(dst->type)); + ggml_sycl_op_acc(ctx, dst); GGML_SYCL_DEBUG("call %s done\n", __func__); } void ggml_sycl_gelu(ggml_backend_sycl_context & ctx, ggml_tensor * dst) { - GGML_SYCL_DEBUG("call %s\n", __func__); - ggml_sycl_op_flatten(ctx, dst->src[0], dst->src[1], dst, ggml_sycl_op_gelu); + GGML_SYCL_DEBUG("call %s: DST Tensor type: %s\n", __func__, ggml_type_name(dst->type)); + ggml_sycl_op_gelu(ctx, dst); GGML_SYCL_DEBUG("call %s done\n", __func__); } void ggml_sycl_silu(ggml_backend_sycl_context & ctx, ggml_tensor * dst) { - GGML_SYCL_DEBUG("call %s\n", __func__); - ggml_sycl_op_flatten(ctx, dst->src[0], dst->src[1], dst, ggml_sycl_op_silu); + GGML_SYCL_DEBUG("call %s: DST Tensor type: %s\n", __func__, ggml_type_name(dst->type)); + ggml_sycl_op_silu(ctx, dst); GGML_SYCL_DEBUG("call %s done\n", __func__); } void ggml_sycl_gelu_quick(ggml_backend_sycl_context & ctx, ggml_tensor * dst) { - GGML_SYCL_DEBUG("call %s\n", __func__); - ggml_sycl_op_flatten(ctx, dst->src[0], dst->src[1], dst, ggml_sycl_op_gelu_quick); + GGML_SYCL_DEBUG("call %s: DST Tensor type: %s\n", __func__, ggml_type_name(dst->type)); + ggml_sycl_op_gelu_quick(ctx, dst); GGML_SYCL_DEBUG("call %s done\n", __func__); } void ggml_sycl_tanh(ggml_backend_sycl_context & ctx, ggml_tensor * dst) { - GGML_SYCL_DEBUG("call %s\n", __func__); - ggml_sycl_op_flatten(ctx, dst->src[0], dst->src[1], dst, ggml_sycl_op_tanh); + GGML_SYCL_DEBUG("call %s: DST Tensor type: %s\n", __func__, ggml_type_name(dst->type)); + ggml_sycl_op_tanh(ctx, dst); GGML_SYCL_DEBUG("call %s done\n", __func__); } void ggml_sycl_relu(ggml_backend_sycl_context & ctx, ggml_tensor * dst) { - GGML_SYCL_DEBUG("call %s\n", __func__); - ggml_sycl_op_flatten(ctx, dst->src[0], dst->src[1], dst, ggml_sycl_op_relu); + GGML_SYCL_DEBUG("call %s: DST Tensor type: %s\n", __func__, ggml_type_name(dst->type)); + ggml_sycl_op_relu(ctx, dst); GGML_SYCL_DEBUG("call %s done\n", __func__); } void ggml_sycl_sigmoid(ggml_backend_sycl_context & ctx, ggml_tensor * dst) { - GGML_SYCL_DEBUG("call %s\n", __func__); - ggml_sycl_op_flatten(ctx, dst->src[0], dst->src[1], dst, ggml_sycl_op_sigmoid); + GGML_SYCL_DEBUG("call %s: DST Tensor type: %s\n", __func__, ggml_type_name(dst->type)); + ggml_sycl_op_sigmoid(ctx, dst); GGML_SYCL_DEBUG("call %s done\n", __func__); } void ggml_sycl_hardsigmoid(ggml_backend_sycl_context & ctx, ggml_tensor * dst) { - GGML_SYCL_DEBUG("call %s\n", __func__); - ggml_sycl_op_flatten(ctx, dst->src[0], dst->src[1], dst, ggml_sycl_op_hardsigmoid); + GGML_SYCL_DEBUG("call %s: DST Tensor type: %s\n", __func__, ggml_type_name(dst->type)); + ggml_sycl_op_hardsigmoid(ctx, dst); GGML_SYCL_DEBUG("call %s done\n", __func__); } void ggml_sycl_hardswish(ggml_backend_sycl_context & ctx, ggml_tensor * dst) { - GGML_SYCL_DEBUG("call %s\n", __func__); - ggml_sycl_op_flatten(ctx, dst->src[0], dst->src[1], dst, ggml_sycl_op_hardswish); + GGML_SYCL_DEBUG("call %s: DST Tensor type: %s\n", __func__, ggml_type_name(dst->type)); + ggml_sycl_op_hardswish(ctx, dst); GGML_SYCL_DEBUG("call %s done\n", __func__); } void ggml_sycl_exp(ggml_backend_sycl_context & ctx, ggml_tensor * dst) { - GGML_SYCL_DEBUG("call %s\n", __func__); - ggml_sycl_op_flatten(ctx, dst->src[0], dst->src[1], dst, ggml_sycl_op_exp); + GGML_SYCL_DEBUG("call %s: DST Tensor type: %s\n", __func__, ggml_type_name(dst->type)); + ggml_sycl_op_exp(ctx, dst); GGML_SYCL_DEBUG("call %s done\n", __func__); } void ggml_sycl_log(ggml_backend_sycl_context & ctx, ggml_tensor * dst) { - GGML_SYCL_DEBUG("call %s\n", __func__); - ggml_sycl_op_flatten(ctx, dst->src[0], dst->src[1], dst, ggml_sycl_op_log); + GGML_SYCL_DEBUG("call %s: DST Tensor type: %s\n", __func__, ggml_type_name(dst->type)); + ggml_sycl_op_log(ctx, dst); GGML_SYCL_DEBUG("call %s done\n", __func__); } void ggml_sycl_neg(ggml_backend_sycl_context & ctx, ggml_tensor * dst) { - GGML_SYCL_DEBUG("call %s\n", __func__); - ggml_sycl_op_flatten(ctx, dst->src[0], dst->src[1], dst, ggml_sycl_op_neg); + GGML_SYCL_DEBUG("call %s: DST Tensor type: %s\n", __func__, ggml_type_name(dst->type)); + ggml_sycl_op_neg(ctx, dst); GGML_SYCL_DEBUG("call %s done\n", __func__); } void ggml_sycl_step(ggml_backend_sycl_context & ctx, ggml_tensor * dst) { - GGML_SYCL_DEBUG("call %s\n", __func__); - ggml_sycl_op_flatten(ctx, dst->src[0], dst->src[1], dst, ggml_sycl_op_step); + GGML_SYCL_DEBUG("call %s: DST Tensor type: %s\n", __func__, ggml_type_name(dst->type)); + ggml_sycl_op_step(ctx, dst); GGML_SYCL_DEBUG("call %s done\n", __func__); } void ggml_sycl_leaky_relu(ggml_backend_sycl_context & ctx, ggml_tensor * dst) { - GGML_SYCL_DEBUG("call %s\n", __func__); - ggml_sycl_op_flatten(ctx, dst->src[0], dst->src[1], dst, ggml_sycl_op_leaky_relu); + GGML_SYCL_DEBUG("call %s: DST Tensor type: %s\n", __func__, ggml_type_name(dst->type)); + ggml_sycl_op_leaky_relu(ctx, dst); GGML_SYCL_DEBUG("call %s done\n", __func__); } void ggml_sycl_sqr(ggml_backend_sycl_context & ctx, ggml_tensor * dst) { - GGML_SYCL_DEBUG("call %s\n", __func__); - ggml_sycl_op_flatten(ctx, dst->src[0], dst->src[1], dst, ggml_sycl_op_sqr); + GGML_SYCL_DEBUG("call %s: DST Tensor type: %s\n", __func__, ggml_type_name(dst->type)); + ggml_sycl_op_sqr(ctx, dst); GGML_SYCL_DEBUG("call %s done\n", __func__); } void ggml_sycl_upscale(ggml_backend_sycl_context & ctx, ggml_tensor * dst) { - GGML_SYCL_DEBUG("call %s\n", __func__); - ggml_sycl_op_flatten(ctx, dst->src[0], dst->src[1], dst, ggml_sycl_op_upscale); + GGML_SYCL_DEBUG("call %s: DST Tensor type: %s\n", __func__, ggml_type_name(dst->type)); + ggml_sycl_op_upscale(ctx, dst); GGML_SYCL_DEBUG("call %s done\n", __func__); } void ggml_sycl_pad(ggml_backend_sycl_context & ctx, ggml_tensor * dst) { - GGML_SYCL_DEBUG("call %s\n", __func__); - ggml_sycl_op_flatten(ctx, dst->src[0], dst->src[1], dst, ggml_sycl_op_pad); + GGML_SYCL_DEBUG("call %s: DST Tensor type: %s\n", __func__, ggml_type_name(dst->type)); + ggml_sycl_op_pad(ctx, dst); GGML_SYCL_DEBUG("call %s done\n", __func__); } +void ggml_sycl_clamp(ggml_backend_sycl_context & ctx, ggml_tensor * dst) { + GGML_SYCL_DEBUG("call %s: DST Tensor type: %s\n", __func__, ggml_type_name(dst->type)); + ggml_sycl_op_clamp(ctx, dst); + GGML_SYCL_DEBUG("call %s done\n", __func__); +} + + void ggml_sycl_add(ggml_backend_sycl_context & ctx, ggml_tensor * dst) { GGML_SYCL_DEBUG("call %s\n", __func__); - ggml_sycl_op_flatten(ctx, dst->src[0], dst->src[1], dst, ggml_sycl_op_add); + ggml_sycl_op_add(ctx, dst); GGML_SYCL_DEBUG("call %s done\n", __func__); } void ggml_sycl_sub(ggml_backend_sycl_context & ctx, ggml_tensor * dst) { GGML_SYCL_DEBUG("call %s\n", __func__); - ggml_sycl_op_flatten(ctx, dst->src[0], dst->src[1], dst, ggml_sycl_op_sub); + ggml_sycl_op_sub(ctx, dst); GGML_SYCL_DEBUG("call %s done\n", __func__); } void ggml_sycl_mul(ggml_backend_sycl_context & ctx, ggml_tensor * dst) { GGML_SYCL_DEBUG("call %s\n", __func__); - ggml_sycl_op_flatten(ctx, dst->src[0], dst->src[1], dst, ggml_sycl_op_mul); + ggml_sycl_op_mul(ctx, dst); GGML_SYCL_DEBUG("call %s done\n", __func__); } void ggml_sycl_div(ggml_backend_sycl_context & ctx, ggml_tensor * dst) { GGML_SYCL_DEBUG("call %s\n", __func__); - ggml_sycl_op_flatten(ctx, dst->src[0], dst->src[1], dst, ggml_sycl_op_div); + ggml_sycl_op_div(ctx, dst); GGML_SYCL_DEBUG("call %s done\n", __func__); } diff --git a/ggml/src/ggml-sycl/element_wise.hpp b/ggml/src/ggml-sycl/element_wise.hpp index 464432645..76d91ba10 100644 --- a/ggml/src/ggml-sycl/element_wise.hpp +++ b/ggml/src/ggml-sycl/element_wise.hpp @@ -2,6 +2,13 @@ #define GGML_SYCL_ELEMENTWISE_HPP #include "common.hpp" +#include "ggml.h" +#include + +template +T neg_infinity() { + return -std::numeric_limits::infinity(); +} static __dpct_inline__ float op_repeat(const float a, const float b) { return b; @@ -24,6 +31,19 @@ static __dpct_inline__ float op_div(const float a, const float b) { return a / b; } +template +struct typed_data { + const T * src; + T * dst; +}; + +template +typed_data cast_data(ggml_tensor * dst) { + return { + /* .src = */ static_cast(dst->src[0]->data), + /* .dst = */ static_cast(dst->data) + }; +} void ggml_sycl_sqrt(ggml_backend_sycl_context & ctx, ggml_tensor * dst); @@ -65,6 +85,10 @@ void ggml_sycl_upscale(ggml_backend_sycl_context & ctx, ggml_tensor * dst); void ggml_sycl_pad(ggml_backend_sycl_context & ctx, ggml_tensor * dst); +void ggml_sycl_clamp(ggml_backend_sycl_context & ctx, ggml_tensor * dst); + +// --------- + void ggml_sycl_add(ggml_backend_sycl_context & ctx, ggml_tensor * dst); void ggml_sycl_sub(ggml_backend_sycl_context & ctx, ggml_tensor * dst); diff --git a/ggml/src/ggml-sycl/gemm.hpp b/ggml/src/ggml-sycl/gemm.hpp index 3f0f34ad6..4ebbb5b66 100644 --- a/ggml/src/ggml-sycl/gemm.hpp +++ b/ggml/src/ggml-sycl/gemm.hpp @@ -13,9 +13,6 @@ #ifndef GGML_SYCL_GEMM_HPP #define GGML_SYCL_GEMM_HPP -#include -#include - #include "ggml-sycl.h" #if GGML_SYCL_DNNL @@ -35,62 +32,34 @@ public: else static_assert(0); } - static inline void row_gemm(sycl::queue& q, bool a_trans, - bool b_trans, int m, int n, int k, - const void* a, dt at, const void* b, dt bt, void* c, dt ct) - { - // Get the device associated with the queue - sycl::device dev = q.get_device(); - // Get the context associated with the queue - sycl::context ctx = q.get_context(); - const dnnl::engine eng = dnnl::sycl_interop::make_engine(dev, ctx); - const dnnl::stream stream = dnnl::sycl_interop::make_stream(eng, q); + static inline void row_gemm(ggml_backend_sycl_context & ctx, bool a_trans, bool b_trans, int m, int n, int k, + const void * a, dt at, const void * b, dt bt, void * c, dt ct, const queue_ptr & q) { + auto stream = ctx.stream_dnnl(q); + auto eng = ctx.engine_dnnl(q); dnnl::memory::dims a_dims = { m, k }; dnnl::memory::dims b_dims = { k, n }; dnnl::memory::dims c_dims = { m, n }; const auto a_in_md = dnnl::memory::desc(a_dims, at, a_trans ? tag::ba : tag::ab); const auto b_in_md = dnnl::memory::desc(b_dims, bt, b_trans ? tag::ba : tag::ab); - const auto c_md = dnnl::memory::desc(c_dims, ct, tag::ab); + const auto c_md = dnnl::memory::desc(c_dims, ct, tag::ab); + + dnnl::primitive_attr primitive_attr; + primitive_attr.set_scratchpad_mode(dnnl::scratchpad_mode::user); + auto a_mem = dnnl::memory(a_in_md, eng, const_cast(a)); auto b_mem = dnnl::memory(b_in_md, eng, const_cast(b)); - auto matmul_pd = dnnl::matmul::primitive_desc(eng, a_in_md, b_in_md, c_md); + auto matmul_pd = dnnl::matmul::primitive_desc(eng, a_in_md, b_in_md, c_md, primitive_attr); auto c_mem = dnnl::memory(matmul_pd.dst_desc(), eng, c); - // Create the primitive. + auto scratchpad_md = matmul_pd.scratchpad_desc(); + auto scratchpad_mem = ctx.get_scratchpad_mem(scratchpad_md, eng, q); auto matmul_prim = dnnl::matmul(matmul_pd); - // Primitive arguments. - std::unordered_map matmul_args; - matmul_args.insert({ DNNL_ARG_SRC, a_mem }); - matmul_args.insert({ DNNL_ARG_WEIGHTS, b_mem }); - matmul_args.insert({ DNNL_ARG_DST, c_mem }); - - matmul_prim.execute(stream, matmul_args); - } - - - static inline void row_gemm(const dnnl::stream& stream, bool a_trans, - bool b_trans, int m, int n, int k, - const void* a, dt at, const void* b, dt bt, void* c, dt ct) - { - auto const eng = stream.get_engine(); - dnnl::memory::dims a_dims = { m, k }; - dnnl::memory::dims b_dims = { k, n }; - dnnl::memory::dims c_dims = { m, n }; - const auto a_in_md = dnnl::memory::desc(a_dims, at, a_trans ? tag::ba : tag::ab); - const auto b_in_md = dnnl::memory::desc(b_dims, bt, b_trans ? tag::ba : tag::ab); - const auto c_md = dnnl::memory::desc(c_dims, ct, tag::ab); - auto a_mem = dnnl::memory(a_in_md, eng, const_cast(a)); - auto b_mem = dnnl::memory(b_in_md, eng, const_cast(b)); - auto matmul_pd = dnnl::matmul::primitive_desc(eng, a_in_md, b_in_md, c_md); - auto c_mem = dnnl::memory(matmul_pd.dst_desc(), eng, c); - - // Create the primitive. - auto matmul_prim = dnnl::matmul(matmul_pd); - // Primitive arguments. + std::unordered_map matmul_args; matmul_args.insert({ DNNL_ARG_SRC, a_mem }); matmul_args.insert({ DNNL_ARG_WEIGHTS, b_mem }); matmul_args.insert({ DNNL_ARG_DST, c_mem }); + matmul_args.insert({ DNNL_ARG_SCRATCHPAD, scratchpad_mem }); matmul_prim.execute(stream, matmul_args); } diff --git a/ggml/src/ggml-sycl/getrows.cpp b/ggml/src/ggml-sycl/getrows.cpp new file mode 100644 index 000000000..64665be46 --- /dev/null +++ b/ggml/src/ggml-sycl/getrows.cpp @@ -0,0 +1,311 @@ +// +// MIT license +// Copyright (C) 2024 Intel Corporation +// SPDX-License-Identifier: MIT +// + +// +// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. +// See https://llvm.org/LICENSE.txt for license information. +// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception +// + +#include "ggml-impl.h" +#include "common.hpp" +#include "dequantize.hpp" +#include "getrows.hpp" + + +template +static void k_get_rows( + const void * src0, const int32_t * src1, dst_t * dst, + int64_t ne00, /*int64_t ne01, int64_t ne02, int64_t ne03,*/ + /*int64_t ne10, int64_t ne11,*/ int64_t ne12, /*int64_t ne13,*/ + /*size_t s0,*/ size_t s1, size_t s2, size_t s3, + /*size_t nb00,*/ size_t nb01, size_t nb02, size_t nb03, + size_t s10, size_t s11, size_t s12, + const sycl::nd_item<3> &item_ct1/*, size_t s13*/) { + + const int i00 = (item_ct1.get_group(2) * item_ct1.get_local_range(2) + + item_ct1.get_local_id(2)) * + 2; + const int i10 = item_ct1.get_local_range(1) * item_ct1.get_group(1) + + item_ct1.get_local_id(1); + const int i11 = (item_ct1.get_group(0) * item_ct1.get_local_range(0) + + item_ct1.get_local_id(0)) / + ne12; + const int i12 = (item_ct1.get_group(0) * item_ct1.get_local_range(0) + + item_ct1.get_local_id(0)) % + ne12; + + if (i00 >= ne00) { + return; + } + + const int i01 = src1[i10*s10 + i11*s11 + i12*s12]; + + dst_t * dst_row = dst + i10*s1 + i11*s2 + i12*s3; + const void * src0_row = (const char *)src0 + i01*nb01 + i11*nb02 + i12*nb03; + + const int ib = i00/qk; // block index + const int iqs = (i00%qk)/qr; // quant index + const int iybs = i00 - i00%qk; // dst block start index + const int y_offset = qr == 1 ? 1 : qk/2; + + // dequantize + dfloat2 v; + dequantize_kernel(src0_row, ib, iqs, v); + + dst_row[iybs + iqs + 0] = v.x(); + dst_row[iybs + iqs + y_offset] = v.y(); +} + +template +static void k_get_rows_reorder( + const void * src0, const void *src0_dq, const int32_t * src1, dst_t * dst, + int64_t ne00, /*int64_t ne01, int64_t ne02, int64_t ne03,*/ + /*int64_t ne10, int64_t ne11,*/ int64_t ne12, /*int64_t ne13,*/ + /*size_t s0,*/ size_t s1, size_t s2, size_t s3, + /*size_t nb00,*/ size_t nb01, size_t nb02, size_t nb03, + size_t s10, size_t s11, size_t s12, + const sycl::nd_item<3> &item_ct1/*, size_t s13*/) { + + const int i00 = (item_ct1.get_group(2) * item_ct1.get_local_range(2) + + item_ct1.get_local_id(2)) * + 2; + const int i10 = item_ct1.get_local_range(1) * item_ct1.get_group(1) + + item_ct1.get_local_id(1); + const int i11 = (item_ct1.get_group(0) * item_ct1.get_local_range(0) + + item_ct1.get_local_id(0)) / + ne12; + const int i12 = (item_ct1.get_group(0) * item_ct1.get_local_range(0) + + item_ct1.get_local_id(0)) % + ne12; + + if (i00 >= ne00) { + return; + } + auto ncols = ne00; + const int i01 = src1[i10*s10 + i11*s11 + i12*s12]; + + dst_t * dst_row = dst + i10*s1 + i11*s2 + i12*s3; + + const int src0_off = i01 * ncols + i00; + const int ib = src0_off / QK4_0; // block index + const int iqs = (i00%qk)/qr; // x quant index + const int iybs = i00 - i00%qk; // dst block start index + const int y_offset = qr == 1 ? 1 : qk/2; + + // dequantize + dfloat2 v; + dequantize_kernel_recorder((const void *)src0_dq, ib, (const void *)src0, src0_off/2, v); + + dst_row[iybs + iqs + 0] = v.x(); + dst_row[iybs + iqs + y_offset] = v.y(); + + GGML_UNUSED(nb01); + GGML_UNUSED(nb02); + GGML_UNUSED(nb03); +} + +template +static void k_get_rows_float( + const src0_t * src0, const int32_t * src1, dst_t * dst, + int64_t ne00, /*int64_t ne01, int64_t ne02, int64_t ne03,*/ + /*int64_t ne10, int64_t ne11,*/ int64_t ne12, /*int64_t ne13,*/ + /*size_t s0,*/ size_t s1, size_t s2, size_t s3, + /*size_t nb00,*/ size_t nb01, size_t nb02, size_t nb03, + size_t s10, size_t s11, size_t s12, + const sycl::nd_item<3> &item_ct1/*, size_t s13*/) { + + const int i00 = item_ct1.get_group(2) * item_ct1.get_local_range(2) + + item_ct1.get_local_id(2); + const int i10 = item_ct1.get_local_range(1) * item_ct1.get_group(1) + + item_ct1.get_local_id(1); + const int i11 = (item_ct1.get_group(0) * item_ct1.get_local_range(0) + + item_ct1.get_local_id(0)) / + ne12; + const int i12 = (item_ct1.get_group(0) * item_ct1.get_local_range(0) + + item_ct1.get_local_id(0)) % + ne12; + + if (i00 >= ne00) { + return; + } + + const int i01 = src1[i10*s10 + i11*s11 + i12*s12]; + + dst_t * dst_row = dst + i10*s1 + i11*s2 + i12*s3; + const src0_t * src0_row = (const src0_t *)((const char *)src0 + i01*nb01 + i11*nb02 + i12*nb03); + + dst_row[i00] = src0_row[i00]; +} + +template +static void get_rows_sycl(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1, + ggml_tensor *dst, const void *src0_dd, + const int32_t *src1_dd, float *dst_dd, + queue_ptr stream) { + + GGML_TENSOR_BINARY_OP_LOCALS + + const sycl::range<3> block_dims(1, 1, SYCL_GET_ROWS_BLOCK_SIZE); + const int block_num_x = (ne00 + 2*SYCL_GET_ROWS_BLOCK_SIZE - 1) / (2*SYCL_GET_ROWS_BLOCK_SIZE); + const sycl::range<3> block_nums(ne11 * ne12, ne10, block_num_x); + + // strides in elements + //const size_t s0 = nb0 / ggml_element_size(dst); + const size_t s1 = nb1 / ggml_element_size(dst); + const size_t s2 = nb2 / ggml_element_size(dst); + const size_t s3 = nb3 / ggml_element_size(dst); + + const size_t s10 = nb10 / ggml_element_size(src1); + const size_t s11 = nb11 / ggml_element_size(src1); + const size_t s12 = nb12 / ggml_element_size(src1); + //const size_t s13 = nb13 / ggml_element_size(src1); + + GGML_ASSERT(ne00 % 2 == 0); + + stream->parallel_for(sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) { + k_get_rows( + src0_dd, src1_dd, dst_dd, ne00, ne12, s1, s2, + s3, nb01, nb02, nb03, s10, s11, s12, item_ct1); + }); + + GGML_UNUSED(dst); + GGML_UNUSED(ctx); +} + +template +static void get_rows_sycl_reorder(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1, + ggml_tensor *dst, const void *src0_dd, + const int32_t *src1_dd, float *dst_dd, + queue_ptr stream) { + + GGML_TENSOR_BINARY_OP_LOCALS + + const sycl::range<3> block_dims(1, 1, SYCL_GET_ROWS_BLOCK_SIZE); + const int block_num_x = (ne00 + 2*SYCL_GET_ROWS_BLOCK_SIZE - 1) / (2*SYCL_GET_ROWS_BLOCK_SIZE); + const sycl::range<3> block_nums(ne11 * ne12, ne10, block_num_x); + + // strides in elements + //const size_t s0 = nb0 / ggml_element_size(dst); + const size_t s1 = nb1 / ggml_element_size(dst); + const size_t s2 = nb2 / ggml_element_size(dst); + const size_t s3 = nb3 / ggml_element_size(dst); + + const size_t s10 = nb10 / ggml_element_size(src1); + const size_t s11 = nb11 / ggml_element_size(src1); + const size_t s12 = nb12 / ggml_element_size(src1); + //const size_t s13 = nb13 / ggml_element_size(src1); + + GGML_ASSERT(ne00 % 2 == 0); + + const uint8_t* src0_q = (const uint8_t*)src0_dd; + const size_t ncols = ne00; + const size_t nrows = ne01; + const sycl::half* src0_dq = (const sycl::half*)(src0_q + nrows * ncols / 2); + stream->parallel_for(sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]]{ + k_get_rows_reorder( + src0_dd, src0_dq, src1_dd, dst_dd, ne00, ne12, s1, s2, + s3, nb01, nb02, nb03, s10, s11, s12, item_ct1); + }); + + GGML_UNUSED(dst); + GGML_UNUSED(ctx); +} + + +template +static void get_rows_sycl_float(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, + const ggml_tensor *src1, ggml_tensor *dst, + const src0_t *src0_dd, const int32_t *src1_dd, + float *dst_dd, queue_ptr stream) { + + GGML_TENSOR_BINARY_OP_LOCALS + + const sycl::range<3> block_dims(1, 1, SYCL_GET_ROWS_BLOCK_SIZE); + const int block_num_x = (ne00 + SYCL_GET_ROWS_BLOCK_SIZE - 1) / SYCL_GET_ROWS_BLOCK_SIZE; + const sycl::range<3> block_nums(ne11 * ne12, ne10, block_num_x); + + // strides in elements + //const size_t s0 = nb0 / ggml_element_size(dst); + const size_t s1 = nb1 / ggml_element_size(dst); + const size_t s2 = nb2 / ggml_element_size(dst); + const size_t s3 = nb3 / ggml_element_size(dst); + + const size_t s10 = nb10 / ggml_element_size(src1); + const size_t s11 = nb11 / ggml_element_size(src1); + const size_t s12 = nb12 / ggml_element_size(src1); + //const size_t s13 = nb13 / ggml_element_size(src1); + + { + dpct::has_capability_or_fail(stream->get_device(), + {sycl::aspect::fp16}); + + stream->parallel_for( + sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) { + k_get_rows_float(src0_dd, src1_dd, dst_dd, ne00, ne12, s1, s2, + s3, nb01, nb02, nb03, s10, s11, s12, item_ct1); + }); + } + + GGML_UNUSED(dst); + GGML_UNUSED(ctx); +} + +void ggml_sycl_op_get_rows(ggml_backend_sycl_context & ctx, ggml_tensor *dst) { + + GGML_ASSERT(dst->src[1]->type == GGML_TYPE_I32); + GGML_ASSERT(dst->type == GGML_TYPE_F32); + + GGML_ASSERT(dst->src[0]->nb[0] == ggml_type_size(dst->src[0]->type)); + GGML_ASSERT(dst->src[1]->nb[0] == ggml_type_size(dst->src[1]->type)); + GGML_ASSERT(dst->nb[0] == ggml_type_size(dst->type)); + + const int32_t * src1_i32 = (const int32_t *) dst->src[1]->data; + /* TODO: Refactor and remove duplicates */ + switch (dst->src[0]->type) { + case GGML_TYPE_F16: + get_rows_sycl_float(ctx, dst->src[0], dst->src[1], dst, (const sycl::half *)dst->src[0]->data, + src1_i32, (float *)dst->data, ctx.stream()); + break; + case GGML_TYPE_F32: + get_rows_sycl_float(ctx, dst->src[0], dst->src[1], dst, (const float *)dst->src[0]->data, + src1_i32, (float *)dst->data, ctx.stream()); + break; + case GGML_TYPE_Q4_0: + if (ctx.opt_feature.reorder && dst->op == GGML_OP_MUL_MAT) { + get_rows_sycl_reorder(ctx, dst->src[0], dst->src[1], dst, (const float *)dst->src[0]->data, + src1_i32, (float *)dst->data, ctx.stream()); + } else { + get_rows_sycl(ctx, dst->src[0], dst->src[1], dst, (const float *)dst->src[0]->data, + src1_i32, (float *)dst->data, ctx.stream()); + } + break; + case GGML_TYPE_Q4_1: + get_rows_sycl(ctx, dst->src[0], dst->src[1], dst, (const float *)dst->src[0]->data, + src1_i32, (float *)dst->data, ctx.stream()); + break; + case GGML_TYPE_Q5_0: + get_rows_sycl(ctx, dst->src[0], dst->src[1], dst, (const float *)dst->src[0]->data, + src1_i32, (float *)dst->data, ctx.stream()); + break; + case GGML_TYPE_Q5_1: + get_rows_sycl(ctx, dst->src[0], dst->src[1], dst, (const float *)dst->src[0]->data, + src1_i32, (float *)dst->data, ctx.stream()); + break; + case GGML_TYPE_Q8_0: + get_rows_sycl(ctx, dst->src[0], dst->src[1], dst, (const float *)dst->src[0]->data, + src1_i32, (float *)dst->data, ctx.stream()); + break; + default: + // TODO: k-quants + GGML_LOG_ERROR("%s: unsupported type: %s\n", __func__, ggml_type_name(dst->src[0]->type)); + GGML_ABORT("fatal error"); + } +} + diff --git a/ggml/src/ggml-sycl/getrows.hpp b/ggml/src/ggml-sycl/getrows.hpp new file mode 100644 index 000000000..1c560cd9f --- /dev/null +++ b/ggml/src/ggml-sycl/getrows.hpp @@ -0,0 +1,20 @@ +// +// MIT license +// Copyright (C) 2024 Intel Corporation +// SPDX-License-Identifier: MIT +// + +// +// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. +// See https://llvm.org/LICENSE.txt for license information. +// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception +// + +#ifndef GGML_SYCL_GETROWS_HPP +#define GGML_SYCL_GETROWS_HPP + +#include "common.hpp" + +void ggml_sycl_op_get_rows(ggml_backend_sycl_context & ctx, ggml_tensor *dst); + +#endif // GGML_SYCL_GETROWS_HPP diff --git a/ggml/src/ggml-sycl/ggml-sycl.cpp b/ggml/src/ggml-sycl/ggml-sycl.cpp index 3d24d2165..3e48a9244 100644 --- a/ggml/src/ggml-sycl/ggml-sycl.cpp +++ b/ggml/src/ggml-sycl/ggml-sycl.cpp @@ -37,10 +37,17 @@ #include "ggml-backend-impl.h" #include "ggml-sycl/backend.hpp" +#include "ggml-sycl/common.hpp" #include "ggml-sycl/presets.hpp" #include "ggml-sycl/gemm.hpp" +#include "ggml-sycl/sycl_hw.hpp" +#include "ggml-sycl/getrows.hpp" +#include "ggml.h" static bool g_sycl_loaded = false; +int g_ggml_sycl_debug = 0; +int g_ggml_sycl_disable_optimize = 0; +int g_ggml_sycl_disable_graph = 0; static ggml_sycl_device_info ggml_sycl_init() { ggml_sycl_device_info info = {}; @@ -63,14 +70,18 @@ static ggml_sycl_device_info ggml_sycl_init() { for (int i = 0; i < info.device_count; ++i) { info.devices[i].vmm = 0; dpct::device_info prop; + sycl::device device = dpct::dev_mgr::instance().get_device(i); + SYCL_CHECK(CHECK_TRY_ERROR(dpct::get_device_info( - prop, dpct::dev_mgr::instance().get_device(i)))); + prop, device))); info.default_tensor_split[i] = total_vram; total_vram += prop.get_global_mem_size(); info.devices[i].cc = 100 * prop.get_major_version() + 10 * prop.get_minor_version(); + info.devices[i].hw_info = get_device_hw_info(&device); + info.devices[i].opt_feature = check_gpu_optimize_feature(info.devices[i].hw_info.arch); info.max_work_group_sizes[i] = prop.get_max_work_group_size(); } @@ -86,7 +97,7 @@ const ggml_sycl_device_info & ggml_sycl_info() { return info; } -void print_device_detail(int id, sycl::device &device, std::string device_type) { +static void print_device_detail(int id, sycl::device &device, std::string device_type) { dpct::device_info prop; SYCL_CHECK(CHECK_TRY_ERROR( @@ -109,6 +120,27 @@ void print_device_detail(int id, sycl::device &device, std::string device_type) global_mem_size, device.get_info().c_str()); } +static void print_device_opt_feature(int device_count) { + GGML_LOG_INFO("SYCL Optimization Feature:\n"); + GGML_LOG_INFO( + "|ID| Device Type|Reorder|\n"); + GGML_LOG_INFO( + "|--|-------------------|-------|\n"); + std::map DeviceNums; + for (int id = 0; id < device_count; ++id) { + sycl::device device = dpct::dev_mgr::instance().get_device(id); + std::string backend_type = get_device_backend_and_type(device); + int type_id = DeviceNums[backend_type]++; + std::stringstream device_type; + device_type << "[" << backend_type << ":" << std::to_string(type_id) + << "]"; + std::string device_type_s = device_type.str(); + device_type_s = std::regex_replace(device_type_s, std::regex("ext_oneapi_"), ""); + GGML_LOG_INFO("|%2d|%19s|%7s|\n", id, device_type_s.c_str(), + ggml_sycl_info().devices[id].opt_feature.reorder ? "Y": "N"); + } + +} void ggml_backend_sycl_print_sycl_devices() { GGML_SYCL_DEBUG("[SYCL] call ggml_backend_sycl_print_sycl_devices\n"); int device_count = dpct::dev_mgr::instance().device_count(); @@ -137,6 +169,8 @@ void ggml_backend_sycl_print_sycl_devices() { << "]"; print_device_detail(id, device, device_type.str()); } + + print_device_opt_feature(device_count); } static inline int get_sycl_env(const char *env_name, int default_val) { @@ -157,18 +191,24 @@ static void ggml_check_sycl() try { static bool initialized = false; if (!initialized) { - GGML_SYCL_DEBUG("[SYCL] call ggml_check_sycl\n"); g_ggml_sycl_debug = get_sycl_env("GGML_SYCL_DEBUG", 0); - GGML_LOG_INFO("GGML_SYCL_DEBUG: %d\n", g_ggml_sycl_debug); + g_ggml_sycl_disable_optimize= get_sycl_env("GGML_SYCL_DISABLE_OPT", 1); + g_ggml_sycl_disable_graph = get_sycl_env("GGML_SYCL_DISABLE_GRAPH", 1); + GGML_SYCL_DEBUG("[SYCL] call ggml_check_sycl\n"); + GGML_LOG_INFO("Running with Environment Variables:\n"); + GGML_LOG_INFO(" GGML_SYCL_DEBUG: %d\n", g_ggml_sycl_debug); + GGML_LOG_INFO(" GGML_SYCL_DISABLE_OPT: %d\n", g_ggml_sycl_disable_optimize); + GGML_LOG_INFO(" GGML_SYCL_DISABLE_GRAPH: %d\n", g_ggml_sycl_disable_graph); + GGML_LOG_INFO("Build with Macros:\n"); #if defined(GGML_SYCL_FORCE_MMQ) - GGML_LOG_INFO("GGML_SYCL_FORCE_MMQ: yes\n"); + GGML_LOG_INFO(" GGML_SYCL_FORCE_MMQ: yes\n"); #else - GGML_LOG_INFO("GGML_SYCL_FORCE_MMQ: no\n"); + GGML_LOG_INFO(" GGML_SYCL_FORCE_MMQ: no\n"); #endif #if defined(GGML_SYCL_F16) - GGML_LOG_INFO("GGML_SYCL_F16: yes\n"); + GGML_LOG_INFO(" GGML_SYCL_F16: yes\n"); #else - GGML_LOG_INFO("GGML_SYCL_F16: no\n"); + GGML_LOG_INFO(" GGML_SYCL_F16: no\n"); #endif /* NOT REMOVE, keep it for next optimize for XMX. @@ -240,19 +280,27 @@ struct ggml_backend_sycl_buffer_context { void * dev_ptr = nullptr; queue_ptr stream; std::string name; + optimize_feature opt_feature; + std::vector tensor_extras; - ggml_backend_sycl_buffer_context(int device, void * dev_ptr, queue_ptr stream) : + ggml_backend_sycl_buffer_context(int device, void * dev_ptr, queue_ptr stream) : device(device), dev_ptr(dev_ptr), stream(stream) { check_allow_gpu_index(device); name = (GGML_SYCL_NAME + std::to_string(device)); + opt_feature = ggml_sycl_info().devices[device].opt_feature; } - ~ggml_backend_sycl_buffer_context() { if (dev_ptr != nullptr) { ggml_sycl_set_device(device); SYCL_CHECK(CHECK_TRY_ERROR(sycl::free(dev_ptr, *stream))); } + + //release extra used by tensors + for (ggml_tensor_extra_gpu * extra : tensor_extras) { + release_extra_gpu(extra); + } + } }; @@ -280,16 +328,20 @@ static void * ggml_backend_sycl_buffer_get_base(ggml_backend_buffer_t buffer) { return ctx->dev_ptr; } -static void +static enum ggml_status ggml_backend_sycl_buffer_init_tensor(ggml_backend_buffer_t buffer, ggml_tensor *tensor) try { ggml_backend_sycl_buffer_context * ctx = (ggml_backend_sycl_buffer_context *)buffer->context; if (tensor->view_src != NULL) { assert(tensor->view_src->buffer->buft == buffer->buft); - return; + return GGML_STATUS_SUCCESS; + } + if (tensor->type == GGML_TYPE_Q4_0) { + ggml_tensor_extra_gpu * extra = new ggml_tensor_extra_gpu{}; + tensor->extra = extra; + ctx->tensor_extras.push_back(extra); //used to release it when destroy ctx. } - if (ggml_is_quantized(tensor->type)) { // initialize padding to 0 to avoid possible NaN values @@ -302,6 +354,7 @@ ggml_backend_sycl_buffer_init_tensor(ggml_backend_buffer_t buffer, padded_size - original_size).wait())); } } + return GGML_STATUS_SUCCESS; } catch (sycl::exception const &exc) { std::cerr << exc.what() << "Exception caught at file:" << __FILE__ @@ -315,11 +368,12 @@ static void ggml_backend_sycl_buffer_set_tensor(ggml_backend_buffer_t buffer, size_t size) try { ggml_backend_sycl_buffer_context * ctx = ( ggml_backend_sycl_buffer_context *)buffer->context; - ggml_sycl_set_device(ctx->device); auto stream = &(dpct::dev_mgr::instance().get_device(ctx->device).default_queue()); SYCL_CHECK( CHECK_TRY_ERROR(dpct::dev_mgr::instance().get_device(ctx->device).queues_wait_and_throw())); + // Note: Use host buffer to save the data from mmap(), then copy to device. It's workaround for mmap() issue on PVC GPU. + // This function will be called during load model from disk. Use memory buffer replace dynamic won't save more time and brings potential memory leak risk here. char* host_buf = (char*)malloc(size); memcpy(host_buf, data, size); SYCL_CHECK( @@ -353,7 +407,7 @@ catch (sycl::exception const &exc) { std::exit(1); } -void dev2dev_memcpy(sycl::queue &q_dst, sycl::queue &q_src, void *ptr_dst, +static void dev2dev_memcpy(sycl::queue &q_dst, sycl::queue &q_src, void *ptr_dst, const void *ptr_src, size_t size) { char *host_buf = (char *)malloc(size); q_src.memcpy(host_buf, (const char *)ptr_src, size).wait(); @@ -439,16 +493,49 @@ catch (sycl::exception const &exc) { std::exit(1); } +static void ggml_backend_sycl_buffer_memset_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor, uint8_t value, + size_t offset, size_t size) { + GGML_SYCL_DEBUG(" [SYCL] call %s\n", __func__); + ggml_backend_sycl_buffer_context * ctx = (ggml_backend_sycl_buffer_context *) buffer->context; + SYCL_CHECK(ggml_sycl_set_device(ctx->device)); + auto stream = &(dpct::dev_mgr::instance().get_device(ctx->device).default_queue()); + if (size == 0) { + return; // Nothing to do + } + if (tensor->data == nullptr) { + GGML_ABORT("Error: Tensor data pointer is null.\n"); + } + void * target_ptr = static_cast(tensor->data) + offset; + SYCL_CHECK(CHECK_TRY_ERROR((*stream).memset(target_ptr, value, size))); + SYCL_CHECK(CHECK_TRY_ERROR((*stream).wait())); +} + +static void ggml_backend_sycl_buffer_reset(ggml_backend_buffer_t buffer) { + GGML_SYCL_DEBUG("[SYCL] call %s\n", __func__); + if (buffer == nullptr) { + return; + } + + ggml_backend_sycl_buffer_context * ctx = (ggml_backend_sycl_buffer_context *) buffer->context; + + if (ctx != nullptr) { + for (ggml_tensor_extra_gpu * extra : ctx->tensor_extras) { + release_extra_gpu(extra); + } + ctx->tensor_extras.clear(); // reset the tensor_extras vector + } +} + static const ggml_backend_buffer_i ggml_backend_sycl_buffer_interface = { /* .free_buffer = */ ggml_backend_sycl_buffer_free_buffer, /* .get_base = */ ggml_backend_sycl_buffer_get_base, /* .init_tensor = */ ggml_backend_sycl_buffer_init_tensor, - /* .memset_tensor = */ NULL, + /* .memset_tensor = */ ggml_backend_sycl_buffer_memset_tensor, /* .set_tensor = */ ggml_backend_sycl_buffer_set_tensor, /* .get_tensor = */ ggml_backend_sycl_buffer_get_tensor, /* .cpy_tensor = */ ggml_backend_sycl_buffer_cpy_tensor, /* .clear = */ ggml_backend_sycl_buffer_clear, - /* .reset = */ NULL, + /* .reset = */ ggml_backend_sycl_buffer_reset, }; // sycl buffer type @@ -529,7 +616,6 @@ ggml_backend_buffer_type_t ggml_backend_sycl_buffer_type(int device) { static std::mutex mutex; std::lock_guard lock(mutex); - GGML_SYCL_DEBUG("[SYCL] call ggml_backend_sycl_buffer_type\n"); auto dev_count = ggml_backend_sycl_get_device_count(); @@ -557,7 +643,7 @@ ggml_backend_buffer_type_t ggml_backend_sycl_buffer_type(int device) { return &ggml_backend_sycl_buffer_types[device]; } -ggml_backend_buffer_type_t ggml_backend_sycl_buffer_type(ggml_backend_sycl_context * ctx) { +static ggml_backend_buffer_type_t ggml_backend_sycl_buffer_type(ggml_backend_sycl_context * ctx) { GGML_SYCL_DEBUG("[SYCL] call ggml_backend_sycl_buffer_type\n"); int device = ctx->device; @@ -659,32 +745,7 @@ struct ggml_backend_sycl_split_buffer_type_context { struct ggml_backend_sycl_split_buffer_context { ~ggml_backend_sycl_split_buffer_context() try { for (ggml_tensor_extra_gpu * extra : tensor_extras) { - for (int i = 0; i < ggml_sycl_info().device_count; ++i) { - for (int64_t is = 0; is < GGML_SYCL_MAX_STREAMS; ++is) { - if (extra->events[i][is] != nullptr) { - /* - DPCT1009:206: SYCL uses exceptions to report errors and - does not use the error codes. The original code was - commented out and a warning string was inserted. You - need to rewrite this code. - */ - SYCL_CHECK(CHECK_TRY_ERROR( - dpct::destroy_event(extra->events[i][is]))); - } - } - if (extra->data_device[i] != nullptr) { - /* - DPCT1009:207: SYCL uses exceptions to report errors and does - not use the error codes. The original code was commented out - and a warning string was inserted. You need to rewrite this - code. - */ - ggml_sycl_set_device(i); - SYCL_CHECK(CHECK_TRY_ERROR(sycl::free( - extra->data_device[i], *(streams[i])))); - } - } - delete extra; + release_extra_gpu(extra, streams); } } catch (sycl::exception const &exc) { @@ -709,7 +770,7 @@ static void * ggml_backend_sycl_split_buffer_get_base(ggml_backend_buffer_t buff GGML_UNUSED(buffer); } -static void +static enum ggml_status ggml_backend_sycl_split_buffer_init_tensor(ggml_backend_buffer_t buffer, ggml_tensor *tensor) try { GGML_ASSERT(tensor->view_src == nullptr); // views of split tensors are not supported @@ -722,7 +783,7 @@ ggml_backend_sycl_split_buffer_init_tensor(ggml_backend_buffer_t buffer, ggml_tensor_extra_gpu * extra = new ggml_tensor_extra_gpu{}; ctx->tensor_extras.push_back(extra); - ctx->streams.push_back(&(dpct::get_current_device().default_queue())); + ctx->streams.push_back(&(dpct::get_current_device().default_queue())); for (int i = 0; i < ggml_sycl_info().device_count; ++i) { int64_t row_low, row_high; @@ -784,6 +845,7 @@ ggml_backend_sycl_split_buffer_init_tensor(ggml_backend_buffer_t buffer, } } tensor->extra = extra; + return GGML_STATUS_SUCCESS; } catch (sycl::exception const &exc) { std::cerr << exc.what() << "Exception caught at file:" << __FILE__ @@ -1263,8 +1325,6 @@ std::unique_ptr ggml_backend_sycl_context::new_pool_for_device(q // struct ggml_sycl_pool_vmm : public ggml_sycl_pool /// kernels - -typedef void (*cpy_kernel_t)(const char * cx, char * cdst); typedef void (*ggml_sycl_op_mul_mat_t)( ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1, ggml_tensor *dst, @@ -1336,83 +1396,6 @@ static void quantize_q8_1(const float * __restrict__ x, void * __restrict__ vy, reinterpret_cast(y[ib].ds.y()) = sum; } -template -static void k_get_rows( - const void * src0, const int32_t * src1, dst_t * dst, - int64_t ne00, /*int64_t ne01, int64_t ne02, int64_t ne03,*/ - /*int64_t ne10, int64_t ne11,*/ int64_t ne12, /*int64_t ne13,*/ - /*size_t s0,*/ size_t s1, size_t s2, size_t s3, - /*size_t nb00,*/ size_t nb01, size_t nb02, size_t nb03, - size_t s10, size_t s11, size_t s12, - const sycl::nd_item<3> &item_ct1/*, size_t s13*/) { - - const int i00 = (item_ct1.get_group(2) * item_ct1.get_local_range(2) + - item_ct1.get_local_id(2)) * - 2; - const int i10 = item_ct1.get_local_range(1) * item_ct1.get_group(1) + - item_ct1.get_local_id(1); - const int i11 = (item_ct1.get_group(0) * item_ct1.get_local_range(0) + - item_ct1.get_local_id(0)) / - ne12; - const int i12 = (item_ct1.get_group(0) * item_ct1.get_local_range(0) + - item_ct1.get_local_id(0)) % - ne12; - - if (i00 >= ne00) { - return; - } - - const int i01 = src1[i10*s10 + i11*s11 + i12*s12]; - - dst_t * dst_row = dst + i10*s1 + i11*s2 + i12*s3; - const void * src0_row = (const char *)src0 + i01*nb01 + i11*nb02 + i12*nb03; - - const int ib = i00/qk; // block index - const int iqs = (i00%qk)/qr; // quant index - const int iybs = i00 - i00%qk; // dst block start index - const int y_offset = qr == 1 ? 1 : qk/2; - - // dequantize - dfloat2 v; - dequantize_kernel(src0_row, ib, iqs, v); - - dst_row[iybs + iqs + 0] = v.x(); - dst_row[iybs + iqs + y_offset] = v.y(); -} - -template -static void k_get_rows_float( - const src0_t * src0, const int32_t * src1, dst_t * dst, - int64_t ne00, /*int64_t ne01, int64_t ne02, int64_t ne03,*/ - /*int64_t ne10, int64_t ne11,*/ int64_t ne12, /*int64_t ne13,*/ - /*size_t s0,*/ size_t s1, size_t s2, size_t s3, - /*size_t nb00,*/ size_t nb01, size_t nb02, size_t nb03, - size_t s10, size_t s11, size_t s12, - const sycl::nd_item<3> &item_ct1/*, size_t s13*/) { - - const int i00 = item_ct1.get_group(2) * item_ct1.get_local_range(2) + - item_ct1.get_local_id(2); - const int i10 = item_ct1.get_local_range(1) * item_ct1.get_group(1) + - item_ct1.get_local_id(1); - const int i11 = (item_ct1.get_group(0) * item_ct1.get_local_range(0) + - item_ct1.get_local_id(0)) / - ne12; - const int i12 = (item_ct1.get_group(0) * item_ct1.get_local_range(0) + - item_ct1.get_local_id(0)) % - ne12; - - if (i00 >= ne00) { - return; - } - - const int i01 = src1[i10*s10 + i11*s11 + i12*s12]; - - dst_t * dst_row = dst + i10*s1 + i11*s2 + i12*s3; - const src0_t * src0_row = (const src0_t *)((const char *)src0 + i01*nb01 + i11*nb02 + i12*nb03); - - dst_row[i00] = src0_row[i00]; -} - static void mul_mat_p021_f16_f32( const void * __restrict__ vx, const float * __restrict__ y, float * __restrict__ dst, const int ncols_x, const int nrows_x, const int nchannels_x, const int nchannels_y, @@ -1523,193 +1506,6 @@ static void mul_mat_vec_nc_f16_f32( // nc == non-contiguous } } -static void cpy_1_f32_f32(const char * cxi, char * cdsti) { - const float * xi = (const float *) cxi; - float * dsti = (float *) cdsti; - - *dsti = *xi; -} - -static void cpy_1_f32_f16(const char * cxi, char * cdsti) { - const float * xi = (const float *) cxi; - sycl::half *dsti = (sycl::half *)cdsti; - - *dsti = sycl::vec(*xi) - .convert()[0]; -} - -static void cpy_1_f16_f16(const char * cxi, char * cdsti) { - const sycl::half *xi = (const sycl::half *)cxi; - sycl::half *dsti = (sycl::half *)cdsti; - - *dsti = *xi; -} - -static void cpy_1_f16_f32(const char * cxi, char * cdsti) { - const sycl::half *xi = (const sycl::half *)cxi; - float * dsti = (float *) cdsti; - - *dsti = *xi; -} - -static void cpy_1_i16_i16(const char * cxi, char * cdsti) { - const int16_t *xi = (const int16_t *)cxi; - int16_t *dsti = (int16_t *)cdsti; - - *dsti = *xi; -} - -static void cpy_1_i32_i32(const char * cxi, char * cdsti) { - const int32_t *xi = (const int32_t *)cxi; - int32_t *dsti = (int32_t *)cdsti; - - *dsti = *xi; -} - -template -static void cpy_f32_f16(const char * cx, char * cdst, const int ne, - const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02, - const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, - const int nb12, const int nb13, const sycl::nd_item<3> &item_ct1) { - const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) + - item_ct1.get_local_id(2); - - if (i >= ne) { - return; - } - - // determine indices i02/i12, i01/i11, i00/i10 as a function of index i of flattened tensor - // then combine those indices with the corresponding byte offsets to get the total offsets - const int i03 = i/(ne00 * ne01 * ne02); - const int i02 = (i - i03*ne00*ne01*ne02 )/ (ne00*ne01); - const int i01 = (i - i03*ne00*ne01*ne02 - i02*ne01*ne00) / ne00; - const int i00 = i - i03*ne00*ne01*ne02 - i02*ne01*ne00 - i01*ne00; - const int x_offset = i00*nb00 + i01*nb01 + i02*nb02 + i03 * nb03; - - const int i13 = i/(ne10 * ne11 * ne12); - const int i12 = (i - i13*ne10*ne11*ne12) / (ne10*ne11); - const int i11 = (i - i13*ne10*ne11*ne12 - i12*ne10*ne11) / ne10; - const int i10 = i - i13*ne10*ne11*ne12 - i12*ne10*ne11 - i11*ne10; - const int dst_offset = i10*nb10 + i11*nb11 + i12*nb12 + i13 * nb13; - - cpy_1(cx + x_offset, cdst + dst_offset); -} - -static void cpy_blck_f32_q8_0(const char * cxi, char * cdsti) { - const float * xi = (const float *) cxi; - block_q8_0 * dsti = (block_q8_0 *) cdsti; - - float amax = 0.0f; // absolute max - - for (int j = 0; j < QK8_0; j++) { - const float v = xi[j]; - amax = sycl::fmax(amax, sycl::fabs((float)v)); - } - - const float d = amax / ((1 << 7) - 1); - const float id = d ? 1.0f/d : 0.0f; - - dsti->d = d; - - for (int j = 0; j < QK8_0; ++j) { - const float x0 = xi[j]*id; - - dsti->qs[j] = sycl::round((float)x0); - } -} - -static void cpy_blck_f32_q4_0(const char * cxi, char * cdsti) { - const float * xi = (const float *) cxi; - block_q4_0 * dsti = (block_q4_0 *) cdsti; - - float amax = 0.0f; - float vmax = 0.0f; - - for (int j = 0; j < QK4_0; ++j) { - const float v = xi[j]; - if (amax < sycl::fabs((float)v)) { - amax = sycl::fabs((float)v); - vmax = v; - } - } - - const float d = vmax / -8; - const float id = d ? 1.0f/d : 0.0f; - - dsti->d = d; - - for (int j = 0; j < QK4_0/2; ++j) { - const float x0 = xi[0 + j]*id; - const float x1 = xi[QK4_0/2 + j]*id; - - const uint8_t xi0 = dpct::min(15, (int8_t)(x0 + 8.5f)); - const uint8_t xi1 = dpct::min(15, (int8_t)(x1 + 8.5f)); - - dsti->qs[j] = xi0; - dsti->qs[j] |= xi1 << 4; - } -} - -static void cpy_blck_f32_q4_1(const char * cxi, char * cdsti) { - const float * xi = (const float *) cxi; - block_q4_1 * dsti = (block_q4_1 *) cdsti; - - float vmin = FLT_MAX; - float vmax = -FLT_MAX; - - for (int j = 0; j < QK4_1; ++j) { - const float v = xi[j]; - - if (v < vmin) vmin = v; - if (v > vmax) vmax = v; - } - - const float d = (vmax - vmin) / ((1 << 4) - 1); - const float id = d ? 1.0f/d : 0.0f; - - dsti->dm.x() = d; - dsti->dm.y() = vmin; - - for (int j = 0; j < QK4_1/2; ++j) { - const float x0 = (xi[0 + j] - vmin)*id; - const float x1 = (xi[QK4_1/2 + j] - vmin)*id; - - const uint8_t xi0 = dpct::min(15, (int8_t)(x0 + 0.5f)); - const uint8_t xi1 = dpct::min(15, (int8_t)(x1 + 0.5f)); - - dsti->qs[j] = xi0; - dsti->qs[j] |= xi1 << 4; - } -} - -template -static void cpy_f32_q(const char * cx, char * cdst, const int ne, - const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02, - const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, - const int nb12, const int nb13, const sycl::nd_item<3> &item_ct1) { - const int i = (item_ct1.get_local_range(2) * item_ct1.get_group(2) + - item_ct1.get_local_id(2)) * - qk; - - if (i >= ne) { - return; - } - - const int i03 = i/(ne00 * ne01 * ne02); - const int i02 = (i - i03*ne00*ne01*ne02 )/ (ne00*ne01); - const int i01 = (i - i03*ne00*ne01*ne02 - i02*ne01*ne00) / ne00; - const int i00 = i - i03*ne00*ne01*ne02 - i02*ne01*ne00 - i01*ne00; - const int x_offset = i00*nb00 + i01*nb01 + i02*nb02 + i03 * nb03; - - const int i13 = i/(ne10 * ne11 * ne12); - const int i12 = (i - i13*ne10*ne11*ne12) / (ne10*ne11); - const int i11 = (i - i13*ne10*ne11*ne12 - i12*ne10*ne11) / ne10; - const int i10 = i - i13*ne10*ne11*ne12 - i12*ne10*ne11 - i11*ne10; - const int dst_offset = (i10/qk)*nb10 + i11*nb11 + i12*nb12 + i13*nb13; - - cpy_blck(cx + x_offset, cdst + dst_offset); -} - static void k_sum_rows_f32(const float * x, float * dst, const int ncols, const sycl::nd_item<3> &item_ct1) { const int row = item_ct1.get_group(1); @@ -1821,17 +1617,6 @@ static void scale_f32(const float * x, float * dst, const float scale, const int dst[i] = scale * x[i]; } -static void clamp_f32(const float * x, float * dst, const float min, const float max, const int k, - const sycl::nd_item<3> &item_ct1) { - const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) + - item_ct1.get_local_id(2); - - if (i >= k) { - return; - } - - dst[i] = x[i] < min ? min : (x[i] > max ? max : x[i]); -} template static void pool2d_nchw_kernel( @@ -1895,81 +1680,6 @@ static void pool2d_nchw_kernel( o_ptr[cur_oh * ow + cur_ow] = res; } -template -static void get_rows_sycl(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1, - ggml_tensor *dst, const void *src0_dd, - const int32_t *src1_dd, float *dst_dd, - queue_ptr stream) { - - GGML_TENSOR_BINARY_OP_LOCALS - - const sycl::range<3> block_dims(1, 1, SYCL_GET_ROWS_BLOCK_SIZE); - const int block_num_x = (ne00 + 2*SYCL_GET_ROWS_BLOCK_SIZE - 1) / (2*SYCL_GET_ROWS_BLOCK_SIZE); - const sycl::range<3> block_nums(ne11 * ne12, ne10, block_num_x); - - // strides in elements - //const size_t s0 = nb0 / ggml_element_size(dst); - const size_t s1 = nb1 / ggml_element_size(dst); - const size_t s2 = nb2 / ggml_element_size(dst); - const size_t s3 = nb3 / ggml_element_size(dst); - - const size_t s10 = nb10 / ggml_element_size(src1); - const size_t s11 = nb11 / ggml_element_size(src1); - const size_t s12 = nb12 / ggml_element_size(src1); - //const size_t s13 = nb13 / ggml_element_size(src1); - - GGML_ASSERT(ne00 % 2 == 0); - - stream->parallel_for(sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) { - k_get_rows( - src0_dd, src1_dd, dst_dd, ne00, ne12, s1, s2, - s3, nb01, nb02, nb03, s10, s11, s12, item_ct1); - }); - - GGML_UNUSED(dst); - GGML_UNUSED(ctx); -} - -template -static void get_rows_sycl_float(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, - const ggml_tensor *src1, ggml_tensor *dst, - const src0_t *src0_dd, const int32_t *src1_dd, - float *dst_dd, queue_ptr stream) { - - GGML_TENSOR_BINARY_OP_LOCALS - - const sycl::range<3> block_dims(1, 1, SYCL_GET_ROWS_BLOCK_SIZE); - const int block_num_x = (ne00 + SYCL_GET_ROWS_BLOCK_SIZE - 1) / SYCL_GET_ROWS_BLOCK_SIZE; - const sycl::range<3> block_nums(ne11 * ne12, ne10, block_num_x); - - // strides in elements - //const size_t s0 = nb0 / ggml_element_size(dst); - const size_t s1 = nb1 / ggml_element_size(dst); - const size_t s2 = nb2 / ggml_element_size(dst); - const size_t s3 = nb3 / ggml_element_size(dst); - - const size_t s10 = nb10 / ggml_element_size(src1); - const size_t s11 = nb11 / ggml_element_size(src1); - const size_t s12 = nb12 / ggml_element_size(src1); - //const size_t s13 = nb13 / ggml_element_size(src1); - - { - dpct::has_capability_or_fail(stream->get_device(), - {sycl::aspect::fp16}); - - stream->parallel_for( - sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) { - k_get_rows_float(src0_dd, src1_dd, dst_dd, ne00, ne12, s1, s2, - s3, nb01, nb02, nb03, s10, s11, s12, item_ct1); - }); - } - - GGML_UNUSED(dst); - GGML_UNUSED(ctx); -} - static void quantize_row_q8_1_sycl(const float *x, void *vy, const int kx, const int ky, const int kx_padded, queue_ptr stream) { @@ -1984,7 +1694,7 @@ static void quantize_row_q8_1_sycl(const float *x, void *vy, const int kx, stream->parallel_for( sycl::nd_range<3>(num_blocks * block_size, block_size), - [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(WARP_SIZE)]] { + [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { quantize_q8_1(x, vy, kx, kx_padded, item_ct1); }); } @@ -2005,7 +1715,7 @@ static void ggml_mul_mat_p021_f16_f32_sycl(const void *vx, const float *y, stream->parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(WARP_SIZE)]] { + [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { mul_mat_p021_f16_f32(vx, y, dst, ncols_x, nrows_x, nchannels_x, nchannels_y, item_ct1); }); @@ -2025,7 +1735,7 @@ static void ggml_mul_mat_vec_nc_f16_f32_sycl( stream->parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(WARP_SIZE)]] { + [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { mul_mat_vec_nc_f16_f32(vx, y, dst, ncols_x, nrows_x, row_stride_x, channel_stride_x, nchannels_y / nchannels_x, item_ct1); @@ -2033,231 +1743,7 @@ static void ggml_mul_mat_vec_nc_f16_f32_sycl( } } -static void -ggml_cpy_f16_f32_sycl(const char *cx, char *cdst, const int ne, const int ne00, - const int ne01, const int ne02, const int nb00, - const int nb01, const int nb02, const int nb03, - const int ne10, const int ne11, const int ne12, - const int nb10, const int nb11, const int nb12, - const int nb13, queue_ptr stream) { - const int num_blocks = (ne + SYCL_CPY_BLOCK_SIZE - 1) / SYCL_CPY_BLOCK_SIZE; - { - dpct::has_capability_or_fail(stream->get_device(), - {sycl::aspect::fp16}); - - stream->parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * - sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), - sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), - [=](sycl::nd_item<3> item_ct1) { - cpy_f32_f16(cx, cdst, ne, ne00, ne01, ne02, nb00, - nb01, nb02, nb03, ne10, ne11, ne12, - nb10, nb11, nb12, nb13, item_ct1); - }); - } -} - -static void ggml_cpy_f32_f32_sycl(const char *cx, char *cdst, const int ne, - const int ne00, const int ne01, - const int ne02, const int nb00, - const int nb01, const int nb02, - const int nb03, const int ne10, - const int ne11, const int ne12, - const int nb10, const int nb11, - const int nb12, const int nb13, - queue_ptr stream) { - - const int num_blocks = (ne + SYCL_CPY_BLOCK_SIZE - 1) / SYCL_CPY_BLOCK_SIZE; - { - dpct::has_capability_or_fail(stream->get_device(), - {sycl::aspect::fp16}); - - stream->parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * - sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), - sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), - [=](sycl::nd_item<3> item_ct1) { - cpy_f32_f16(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, - nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, - item_ct1); - }); - } -} - -static void ggml_cpy_f32_f16_sycl(const char *cx, char *cdst, const int ne, - const int ne00, const int ne01, - const int ne02, const int nb00, - const int nb01, const int nb02, - const int nb03, const int ne10, - const int ne11, const int ne12, - const int nb10, const int nb11, - const int nb12, const int nb13, - queue_ptr stream) { - - const int num_blocks = (ne + SYCL_CPY_BLOCK_SIZE - 1) / SYCL_CPY_BLOCK_SIZE; - { - dpct::has_capability_or_fail(stream->get_device(), - {sycl::aspect::fp16}); - - stream->parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * - sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), - sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), - [=](sycl::nd_item<3> item_ct1) { - cpy_f32_f16(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, - nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, - item_ct1); - }); - } -} - -static void ggml_cpy_f32_q8_0_sycl(const char *cx, char *cdst, const int ne, - const int ne00, const int ne01, - const int ne02, const int nb00, - const int nb01, const int nb02, - const int nb03, const int ne10, - const int ne11, const int ne12, - const int nb10, const int nb11, - const int nb12, const int nb13, - queue_ptr stream) { - - GGML_ASSERT(ne % QK8_0 == 0); - const int num_blocks = ne / QK8_0; - stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), - sycl::range<3>(1, 1, 1)), - [=](sycl::nd_item<3> item_ct1) { - cpy_f32_q( - cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, - nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, - item_ct1); - }); -} - -static void ggml_cpy_f32_q4_0_sycl(const char *cx, char *cdst, const int ne, - const int ne00, const int ne01, - const int ne02, const int nb00, - const int nb01, const int nb02, - const int nb03, const int ne10, - const int ne11, const int ne12, - const int nb10, const int nb11, - const int nb12, const int nb13, - queue_ptr stream) { - - GGML_ASSERT(ne % QK4_0 == 0); - const int num_blocks = ne / QK4_0; - stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), - sycl::range<3>(1, 1, 1)), - [=](sycl::nd_item<3> item_ct1) { - cpy_f32_q( - cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, - nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, - item_ct1); - }); -} - -static void ggml_cpy_f32_q4_1_sycl(const char *cx, char *cdst, const int ne, - const int ne00, const int ne01, - const int ne02, const int nb00, - const int nb01, const int nb02, - const int nb03, const int ne10, - const int ne11, const int ne12, - const int nb10, const int nb11, - const int nb12, const int nb13, - queue_ptr stream) { - - GGML_ASSERT(ne % QK4_1 == 0); - const int num_blocks = ne / QK4_1; - stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), - sycl::range<3>(1, 1, 1)), - [=](sycl::nd_item<3> item_ct1) { - cpy_f32_q( - cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, - nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, - item_ct1); - }); -} - -static void ggml_cpy_f16_f16_sycl(const char *cx, char *cdst, const int ne, - const int ne00, const int ne01, - const int ne02, const int nb00, - const int nb01, const int nb02, - const int nb03, const int ne10, - const int ne11, const int ne12, - const int nb10, const int nb11, - const int nb12, const int nb13, - queue_ptr stream) { - - const int num_blocks = (ne + SYCL_CPY_BLOCK_SIZE - 1) / SYCL_CPY_BLOCK_SIZE; - { - dpct::has_capability_or_fail(stream->get_device(), - {sycl::aspect::fp16}); - - stream->parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * - sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), - sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), - [=](sycl::nd_item<3> item_ct1) { - cpy_f32_f16(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, - nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, - item_ct1); - }); - } -} - -static void ggml_cpy_i16_i16_sycl(const char *cx, char *cdst, const int ne, - const int ne00, const int ne01, - const int ne02, const int nb00, - const int nb01, const int nb02, - const int nb03, const int ne10, - const int ne11, const int ne12, - const int nb10, const int nb11, - const int nb12, const int nb13, - queue_ptr stream) { - - const int num_blocks = (ne + SYCL_CPY_BLOCK_SIZE - 1) / SYCL_CPY_BLOCK_SIZE; - { - // dpct::has_capability_or_fail(stream->get_device(), - // {sycl::aspect::fp16}); - - stream->parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * - sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), - sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), - [=](sycl::nd_item<3> item_ct1) { - cpy_f32_f16(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, - nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, - item_ct1); - }); - } -} - -static void ggml_cpy_i32_i32_sycl(const char *cx, char *cdst, const int ne, - const int ne00, const int ne01, - const int ne02, const int nb00, - const int nb01, const int nb02, - const int nb03, const int ne10, - const int ne11, const int ne12, - const int nb10, const int nb11, - const int nb12, const int nb13, - queue_ptr stream) { - - const int num_blocks = (ne + SYCL_CPY_BLOCK_SIZE - 1) / SYCL_CPY_BLOCK_SIZE; - { - // dpct::has_capability_or_fail(stream->get_device(), - // {sycl::aspect::fp16}); - - stream->parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * - sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), - sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), - [=](sycl::nd_item<3> item_ct1) { - cpy_f32_f16(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, - nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, - item_ct1); - }); - } -} static void scale_f32_sycl(const float *x, float *dst, const float scale, const int k, queue_ptr stream) { @@ -2271,18 +1757,6 @@ static void scale_f32_sycl(const float *x, float *dst, const float scale, }); } -static void clamp_f32_sycl(const float *x, float *dst, const float min, - const float max, const int k, - queue_ptr stream) { - const int num_blocks = (k + SYCL_CLAMP_BLOCK_SIZE - 1) / SYCL_CLAMP_BLOCK_SIZE; - stream->parallel_for( - sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * - sycl::range<3>(1, 1, SYCL_CLAMP_BLOCK_SIZE), - sycl::range<3>(1, 1, SYCL_CLAMP_BLOCK_SIZE)), - [=](sycl::nd_item<3> item_ct1) { - clamp_f32(x, dst, min, max, k, item_ct1); - }); -} static void sum_rows_f32_sycl(const float *x, float *dst, const int ncols, const int nrows, queue_ptr stream) { @@ -2290,7 +1764,7 @@ static void sum_rows_f32_sycl(const float *x, float *dst, const int ncols, const sycl::range<3> block_nums(1, nrows, 1); stream->parallel_for(sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) - [[intel::reqd_sub_group_size(WARP_SIZE)]] { + [[sycl::reqd_sub_group_size(WARP_SIZE)]] { k_sum_rows_f32(x, dst, ncols, item_ct1); }); } @@ -2493,62 +1967,8 @@ catch (sycl::exception const &exc) { std::exit(1); } -static void ggml_sycl_op_get_rows(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, - const ggml_tensor *src1, ggml_tensor *dst, - const float *src0_d, const float *src1_d, - float *dst_d, const queue_ptr &stream) { - - GGML_ASSERT(src1->type == GGML_TYPE_I32); - GGML_ASSERT(dst->type == GGML_TYPE_F32); - - GGML_ASSERT(src0->nb[0] == ggml_type_size(src0->type)); - GGML_ASSERT(src1->nb[0] == ggml_type_size(src1->type)); - GGML_ASSERT(dst->nb[0] == ggml_type_size(dst->type)); - - const int32_t * src1_i32 = (const int32_t *) src1_d; - - switch (src0->type) { - case GGML_TYPE_F16: - get_rows_sycl_float(ctx, src0, src1, dst, (const sycl::half *)src0_d, - src1_i32, dst_d, stream); - break; - case GGML_TYPE_F32: - get_rows_sycl_float(ctx, src0, src1, dst, src0_d, src1_i32, dst_d, stream); - break; - case GGML_TYPE_Q4_0: - get_rows_sycl(ctx, src0, src1, dst, src0_d, src1_i32, dst_d, stream); - break; - case GGML_TYPE_Q4_1: - get_rows_sycl(ctx, src0, src1, dst, src0_d, src1_i32, dst_d, stream); - break; - case GGML_TYPE_Q5_0: - get_rows_sycl(ctx, src0, src1, dst, src0_d, src1_i32, dst_d, stream); - break; - case GGML_TYPE_Q5_1: - get_rows_sycl(ctx, src0, src1, dst, src0_d, src1_i32, dst_d, stream); - break; - case GGML_TYPE_Q8_0: - get_rows_sycl(ctx, src0, src1, dst, src0_d, src1_i32, dst_d, stream); - break; - default: - // TODO: k-quants - GGML_LOG_ERROR("%s: unsupported type: %s\n", __func__, ggml_type_name(src0->type)); - GGML_ABORT("fatal error"); - break; - } -} - - -static void ggml_sycl_op_repeat(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, - const ggml_tensor *src1, ggml_tensor *dst, - const float *src0_d, const float *src1_d, - float *dst_d, - const queue_ptr &main_stream) { - - ggml_sycl_op_bin_bcast>(ctx, dst, src0, dst, nullptr, src0_d, dst_d, main_stream); - - GGML_UNUSED(src1); - GGML_UNUSED(src1_d); +static void ggml_sycl_op_repeat(ggml_backend_sycl_context & ctx, ggml_tensor *dst) { + ggml_sycl_op_bin_bcast>(ctx, dst, dst->src[0], dst); } @@ -2588,11 +2008,10 @@ inline void ggml_sycl_op_mul_mat_sycl( if ((src0->type == GGML_TYPE_F16 || ggml_is_quantized(src0->type)) && use_fp16 && ggml_is_contiguous(src0) && row_diff == src0->ne[1] && dst->op_params[0] == GGML_PREC_DEFAULT) { - // GGML_SYCL_DEBUG("ggml_sycl_op_mul_mat_sycl - fp16 path\n"); ggml_sycl_pool_alloc src0_as_f16(ctx.pool()); if (src0->type != GGML_TYPE_F16) { - const to_fp16_sycl_t to_fp16_sycl = ggml_get_to_fp16_sycl(src0->type); + const to_fp16_sycl_t to_fp16_sycl = ggml_get_to_fp16_sycl(src0->type, dst); GGML_ASSERT(to_fp16_sycl != nullptr); size_t ne = row_diff*ne00; src0_as_f16.alloc(ne); @@ -2604,7 +2023,7 @@ inline void ggml_sycl_op_mul_mat_sycl( ggml_sycl_pool_alloc src1_as_f16(ctx.pool()); if (src1->type != GGML_TYPE_F16) { - const to_fp16_sycl_t to_fp16_sycl = ggml_get_to_fp16_sycl(src1->type); + const to_fp16_sycl_t to_fp16_sycl = ggml_get_to_fp16_sycl(src1->type, dst); GGML_ASSERT(to_fp16_sycl != nullptr); size_t ne = src1_ncols*ne10; src1_as_f16.alloc(ne); @@ -2619,19 +2038,19 @@ inline void ggml_sycl_op_mul_mat_sycl( const sycl::half alpha_f16 = 1.0f; const sycl::half beta_f16 = 0.0f; SYCL_CHECK(CHECK_TRY_ERROR(dpct::gemm( - *stream, oneapi::mkl::transpose::trans, - oneapi::mkl::transpose::nontrans, row_diff, src1_ncols, ne10, + *stream, oneapi::math::transpose::trans, + oneapi::math::transpose::nontrans, row_diff, src1_ncols, ne10, &alpha_f16, src0_ptr, dpct::library_data_t::real_half, ne00, src1_ptr, dpct::library_data_t::real_half, ne10, &beta_f16, dst_f16.get(), dpct::library_data_t::real_half, ldc, dpct::library_data_t::real_half))); - const to_fp32_sycl_t to_fp32_sycl = ggml_get_to_fp32_sycl(GGML_TYPE_F16); + const to_fp32_sycl_t to_fp32_sycl = ggml_get_to_fp32_sycl(GGML_TYPE_F16, dst); to_fp32_sycl(dst_f16.get(), dst_dd_i, row_diff*src1_ncols, stream); #else - auto dnnl_stream = ctx.stream_dnnl(stream); - DnnlGemmWrapper::row_gemm(dnnl_stream, false, true, src1_ncols, row_diff, ne10, src1_ptr, DnnlGemmWrapper::to_dt(), - src0_ptr, DnnlGemmWrapper::to_dt(), dst_f16.get(), DnnlGemmWrapper::to_dt()); - const to_fp32_sycl_t to_fp32_sycl = ggml_get_to_fp32_sycl(GGML_TYPE_F16); + DnnlGemmWrapper::row_gemm(ctx, false, true, src1_ncols, row_diff, ne10, src1_ptr, + DnnlGemmWrapper::to_dt(), src0_ptr, DnnlGemmWrapper::to_dt(), + dst_f16.get(), DnnlGemmWrapper::to_dt(), stream); + const to_fp32_sycl_t to_fp32_sycl = ggml_get_to_fp32_sycl(GGML_TYPE_F16, dst); to_fp32_sycl(dst_f16.get(), dst_dd_i, row_diff* src1_ncols, stream); #endif } @@ -2640,13 +2059,13 @@ inline void ggml_sycl_op_mul_mat_sycl( ggml_sycl_pool_alloc src0_ddq_as_f32(ctx.pool()); ggml_sycl_pool_alloc src1_ddq_as_f32(ctx.pool()); if (src0->type != GGML_TYPE_F32) { - const to_fp32_sycl_t to_fp32_sycl = ggml_get_to_fp32_sycl(src0->type); + const to_fp32_sycl_t to_fp32_sycl = ggml_get_to_fp32_sycl(src0->type, dst); GGML_ASSERT(to_fp32_sycl != nullptr); src0_ddq_as_f32.alloc(row_diff*ne00); to_fp32_sycl(src0_dd_i, src0_ddq_as_f32.get(), row_diff*ne00, stream); } if (src1->type != GGML_TYPE_F32) { - const to_fp32_sycl_t to_fp32_sycl = ggml_get_to_fp32_sycl(src1->type); + const to_fp32_sycl_t to_fp32_sycl = ggml_get_to_fp32_sycl(src1->type, dst); GGML_ASSERT(to_fp32_sycl != nullptr); src1_ddq_as_f32.alloc(src1_ncols*ne10); to_fp32_sycl(src1_ddf_i, src1_ddq_as_f32.get(), src1_ncols*ne10, stream); @@ -2657,21 +2076,14 @@ inline void ggml_sycl_op_mul_mat_sycl( #if !GGML_SYCL_DNNL const float alpha = 1.0f; const float beta = 0.0f; -# ifdef GGML_SYCL_NVIDIA - SYCL_CHECK(CHECK_TRY_ERROR(oneapi::mkl::blas::column_major::gemm( - oneapi::mkl::backend_selector{ *stream }, oneapi::mkl::transpose::trans, - oneapi::mkl::transpose::nontrans, row_diff, src1_ncols, ne10, dpct::get_value(&alpha, *stream), src0_ddf_i, - ne00, src1_ddf1_i, ne10, dpct::get_value(&beta, *stream), dst_dd_i, ldc))); -# else - SYCL_CHECK(CHECK_TRY_ERROR(oneapi::mkl::blas::column_major::gemm( - *stream, oneapi::mkl::transpose::trans, oneapi::mkl::transpose::nontrans, row_diff, src1_ncols, ne10, - dpct::get_value(&alpha, *stream), src0_ddf_i, ne00, src1_ddf1_i, ne10, dpct::get_value(&beta, *stream), - dst_dd_i, ldc))); -# endif + SYCL_CHECK(CHECK_TRY_ERROR(oneapi::math::blas::column_major::gemm( + get_onemath_backend(*stream), oneapi::math::transpose::trans, oneapi::math::transpose::nontrans, row_diff, + src1_ncols, ne10, dpct::get_value(&alpha, *stream), src0_ddf_i, ne00, src1_ddf1_i, ne10, + dpct::get_value(&beta, *stream), dst_dd_i, ldc))); #else - auto dnnl_stream = ctx.stream_dnnl(stream); - DnnlGemmWrapper::row_gemm(dnnl_stream, false, true, src1_ncols, row_diff, ne10, src1_ddf1_i, DnnlGemmWrapper::to_dt(), - src0_ddf_i, DnnlGemmWrapper::to_dt(), dst_dd_i, DnnlGemmWrapper::to_dt()); + DnnlGemmWrapper::row_gemm(ctx, false, true, src1_ncols, row_diff, ne10, src1_ddf1_i, + DnnlGemmWrapper::to_dt(), src0_ddf_i, DnnlGemmWrapper::to_dt(), + dst_dd_i, DnnlGemmWrapper::to_dt(), stream); #endif } GGML_UNUSED(dst); @@ -2684,13 +2096,14 @@ catch (sycl::exception const &exc) { std::exit(1); } -static void ggml_sycl_op_pool2d(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, - const ggml_tensor *src1, ggml_tensor *dst, - const float *src0_dd, const float *src1_dd, - float *dst_dd, const queue_ptr &main_stream) { +static void ggml_sycl_op_pool2d(ggml_backend_sycl_context & ctx, ggml_tensor *dst) { - GGML_ASSERT(src0->type == GGML_TYPE_F32); + GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32); GGML_ASSERT( dst->type == GGML_TYPE_F32); + dpct::queue_ptr main_stream = ctx.stream(); + SYCL_CHECK(ggml_sycl_set_device(ctx.device)); + const float * src0_dd = static_cast(dst->src[0]->data); + float * dst_dd = static_cast(dst->data); const int32_t * opts = (const int32_t *)dst->op_params; enum ggml_op_pool op = static_cast(opts[0]); @@ -2701,8 +2114,8 @@ static void ggml_sycl_op_pool2d(ggml_backend_sycl_context & ctx, const ggml_tens const int p0 = opts[5]; const int p1 = opts[6]; - const int64_t IH = src0->ne[1]; - const int64_t IW = src0->ne[0]; + const int64_t IH = dst->src[0]->ne[1]; + const int64_t IW = dst->src[0]->ne[0]; const int64_t N = dst->ne[3]; const int64_t OC = dst->ne[2]; @@ -2721,163 +2134,105 @@ static void ggml_sycl_op_pool2d(ggml_backend_sycl_context & ctx, const ggml_tens parallel_elements, src0_dd, dst_dd, op, item_ct1); }); - - GGML_UNUSED(src1); - GGML_UNUSED(src1_dd); - GGML_UNUSED(ctx); } -inline void ggml_sycl_op_sum(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, - const ggml_tensor *src1, ggml_tensor *dst, - const float *src0_dd, const float *src1_dd, - float *dst_dd, - const queue_ptr &main_stream) { - GGML_ASSERT(src0->type == GGML_TYPE_F32); +inline void ggml_sycl_op_sum(ggml_backend_sycl_context & ctx, ggml_tensor *dst) { + GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32); GGML_ASSERT( dst->type == GGML_TYPE_F32); + dpct::queue_ptr main_stream = ctx.stream(); + SYCL_CHECK(ggml_sycl_set_device(ctx.device)); + const float * src0_dd = static_cast(dst->src[0]->data); + float * dst_dd = static_cast(dst->data); - const int64_t ne = ggml_nelements(src0); + const int64_t ne = ggml_nelements(dst->src[0]); sum_rows_f32_sycl(src0_dd, dst_dd, ne, 1, main_stream); - - GGML_UNUSED(src1); - GGML_UNUSED(dst); - GGML_UNUSED(src1_dd); - GGML_UNUSED(ctx); } -inline void ggml_sycl_op_sum_rows(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, - const ggml_tensor *src1, ggml_tensor *dst, - const float *src0_dd, const float *src1_dd, - float *dst_dd, - const queue_ptr &main_stream) { +inline void ggml_sycl_op_sum_rows(ggml_backend_sycl_context & ctx, ggml_tensor *dst) { - GGML_ASSERT(src0->type == GGML_TYPE_F32); + GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32); GGML_ASSERT( dst->type == GGML_TYPE_F32); + dpct::queue_ptr main_stream = ctx.stream(); + SYCL_CHECK(ggml_sycl_set_device(ctx.device)); + const float * src0_dd = static_cast(dst->src[0]->data); + float * dst_dd = static_cast(dst->data); - const int64_t ncols = src0->ne[0]; - const int64_t nrows = ggml_nrows(src0); + const int64_t ncols = dst->src[0]->ne[0]; + const int64_t nrows = ggml_nrows(dst->src[0]); sum_rows_f32_sycl(src0_dd, dst_dd, ncols, nrows, main_stream); - - GGML_UNUSED(src1); - GGML_UNUSED(dst); - GGML_UNUSED(src1_dd); - GGML_UNUSED(ctx); } -inline void ggml_sycl_op_argsort(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, - const ggml_tensor *src1, ggml_tensor *dst, - const float *src0_dd, const float *src1_dd, - float *dst_dd, - const queue_ptr &main_stream) { +inline void ggml_sycl_op_argsort(ggml_backend_sycl_context & ctx, ggml_tensor * dst) { + GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32); + GGML_ASSERT(dst->type == GGML_TYPE_I32); + dpct::queue_ptr main_stream = ctx.stream(); + SYCL_CHECK(ggml_sycl_set_device(ctx.device)); + const float * src0_dd = static_cast(dst->src[0]->data); + int32_t * dst_dd = static_cast(dst->data); - GGML_ASSERT(src0->type == GGML_TYPE_F32); - GGML_ASSERT( dst->type == GGML_TYPE_I32); - const int64_t ncols = src0->ne[0]; - const int64_t nrows = ggml_nrows(src0); + const int64_t ncols = dst->src[0]->ne[0]; + const int64_t nrows = ggml_nrows(dst->src[0]); enum ggml_sort_order order = (enum ggml_sort_order) dst->op_params[0]; - argsort_f32_i32_sycl(src0_dd, (int *)dst_dd, ncols, nrows, order, main_stream); - - GGML_UNUSED(src1); - GGML_UNUSED(dst); - GGML_UNUSED(src1_dd); - GGML_UNUSED(ctx); + argsort_f32_i32_sycl(src0_dd, (int *) dst_dd, ncols, nrows, order, main_stream); } -inline void ggml_sycl_op_argmax(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, - const ggml_tensor *src1, ggml_tensor *dst, - const float *src0_dd, const float *src1_dd, - float *dst_dd, - const queue_ptr &main_stream) { +inline void ggml_sycl_op_argmax(ggml_backend_sycl_context & ctx, ggml_tensor *dst) { - GGML_ASSERT(src0->type == GGML_TYPE_F32); + GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32); GGML_ASSERT( dst->type == GGML_TYPE_I32); - const int64_t ncols = src0->ne[0]; - const int64_t nrows = ggml_nrows(src0); + dpct::queue_ptr main_stream = ctx.stream(); + SYCL_CHECK(ggml_sycl_set_device(ctx.device)); + const float * src0_dd = static_cast(dst->src[0]->data); + int32_t * dst_dd = static_cast(dst->data); - argmax_f32_i32_sycl(src0_dd, (int *)dst_dd, ncols, nrows, main_stream); + const int64_t ncols = dst->src[0]->ne[0]; + const int64_t nrows = ggml_nrows(dst->src[0]); - GGML_UNUSED(src1); - GGML_UNUSED(dst); - GGML_UNUSED(src1_dd); - GGML_UNUSED(ctx); + argmax_f32_i32_sycl(src0_dd, dst_dd, ncols, nrows, main_stream); } -inline void ggml_sycl_op_diag_mask_inf(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, - const ggml_tensor *src1, - ggml_tensor *dst, const float *src0_dd, - const float *src1_dd, float *dst_dd, - const queue_ptr &main_stream) { +inline void ggml_sycl_op_diag_mask_inf(ggml_backend_sycl_context & ctx,ggml_tensor *dst) { - GGML_ASSERT(src0->type == GGML_TYPE_F32); + GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32); GGML_ASSERT( dst->type == GGML_TYPE_F32); + dpct::queue_ptr main_stream = ctx.stream(); + SYCL_CHECK(ggml_sycl_set_device(ctx.device)); + const float * src0_dd = static_cast(dst->src[0]->data); + float * dst_dd = static_cast(dst->data); - const int64_t ne00 = src0->ne[0]; - const int64_t ne01 = src0->ne[1]; - const int nrows0 = ggml_nrows(src0); + const int64_t ne00 = dst->src[0]->ne[0]; + const int64_t ne01 = dst->src[0]->ne[1]; + const int nrows0 = ggml_nrows(dst->src[0]); const int n_past = ((int32_t *) dst->op_params)[0]; diag_mask_inf_f32_sycl(src0_dd, dst_dd, ne00, nrows0, ne01, n_past, main_stream); - - GGML_UNUSED(src1); - GGML_UNUSED(dst); - GGML_UNUSED(src1_dd); - GGML_UNUSED(ctx); } -inline void ggml_sycl_op_scale(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1, - ggml_tensor *dst, const float *src0_dd, - const float *src1_dd, float *dst_dd, - const queue_ptr &main_stream) { +inline void ggml_sycl_op_scale(ggml_backend_sycl_context & ctx, ggml_tensor *dst) { - GGML_ASSERT(src0->type == GGML_TYPE_F32); + GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32); GGML_ASSERT( dst->type == GGML_TYPE_F32); + dpct::queue_ptr main_stream = ctx.stream(); + SYCL_CHECK(ggml_sycl_set_device(ctx.device)); + const float * src0_dd = static_cast(dst->src[0]->data); + float * dst_dd = static_cast(dst->data); float scale; memcpy(&scale, dst->op_params, sizeof(float)); - scale_f32_sycl(src0_dd, dst_dd, scale, ggml_nelements(src0), main_stream); + scale_f32_sycl(src0_dd, dst_dd, scale, ggml_nelements(dst->src[0]), main_stream); /* DPCT1010:87: SYCL uses exceptions to report errors and does not use the error codes. The call was replaced with 0. You need to rewrite this code. */ SYCL_CHECK(0); - - GGML_UNUSED(src1); - GGML_UNUSED(dst); - GGML_UNUSED(src1_dd); - GGML_UNUSED(ctx); -} - -inline void ggml_sycl_op_clamp(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1, - ggml_tensor *dst, const float *src0_dd, - const float *src1_dd, float *dst_dd, - const queue_ptr &main_stream) { - - GGML_ASSERT(src0->type == GGML_TYPE_F32); - GGML_ASSERT( dst->type == GGML_TYPE_F32); - - float min; - float max; - memcpy(&min, dst->op_params, sizeof(float)); - memcpy(&max, (float *) dst->op_params + 1, sizeof(float)); - - clamp_f32_sycl(src0_dd, dst_dd, min, max, ggml_nelements(src0), main_stream); - /* - DPCT1010:88: SYCL uses exceptions to report errors and does not use the - error codes. The call was replaced with 0. You need to rewrite this code. - */ - SYCL_CHECK(0); - - GGML_UNUSED(src1); - GGML_UNUSED(dst); - GGML_UNUSED(src1_dd); - GGML_UNUSED(ctx); } static void ggml_sycl_set_peer_access(const int n_tokens, int main_device) { @@ -3084,7 +2439,6 @@ static void ggml_sycl_op_mul_mat(ggml_backend_sycl_context & ctx, const ggml_ten for (int64_t src1_col_0 = 0; src1_col_0 < ne11; src1_col_0 += src1_col_stride) { const int64_t is = split ? (src1_col_0/src1_col_stride) % GGML_SYCL_MAX_STREAMS : 0; const int64_t src1_ncols = src1_col_0 + src1_col_stride > ne11 ? ne11 - src1_col_0 : src1_col_stride; - for (int i = 0; i < ggml_sycl_info().device_count; ++i) { if ((!split && i != ctx.device) || dev[i].row_low == dev[i].row_high) { continue; @@ -3248,31 +2602,37 @@ catch (sycl::exception const &exc) { static void ggml_sycl_repeat(ggml_backend_sycl_context & ctx, ggml_tensor * dst) { GGML_SYCL_DEBUG("call %s\n", __func__); - ggml_sycl_op_flatten(ctx, dst->src[0], dst->src[1], dst, ggml_sycl_op_repeat); + ggml_sycl_op_repeat(ctx, dst); GGML_SYCL_DEBUG("call %s done\n", __func__); } static void ggml_sycl_get_rows(ggml_backend_sycl_context & ctx, ggml_tensor * dst) { GGML_SYCL_DEBUG("call %s\n", __func__); - ggml_sycl_op_flatten(ctx, dst->src[0], dst->src[1], dst, ggml_sycl_op_get_rows); + ggml_sycl_op_get_rows(ctx, dst); GGML_SYCL_DEBUG("call %s done\n", __func__); } static void ggml_sycl_norm(ggml_backend_sycl_context & ctx, ggml_tensor * dst) { GGML_SYCL_DEBUG("call %s\n", __func__); - ggml_sycl_op_flatten(ctx, dst->src[0], dst->src[1], dst, ggml_sycl_op_norm); + ggml_sycl_op_norm(ctx, dst); GGML_SYCL_DEBUG("call %s done\n", __func__); } static void ggml_sycl_rms_norm(ggml_backend_sycl_context & ctx, ggml_tensor * dst) { GGML_SYCL_DEBUG("call %s\n", __func__); - ggml_sycl_op_flatten(ctx, dst->src[0], dst->src[1], dst, ggml_sycl_op_rms_norm); + ggml_sycl_op_rms_norm(ctx, dst); + GGML_SYCL_DEBUG("call %s done\n", __func__); +} + +static void ggml_sycl_l2_norm(ggml_backend_sycl_context & ctx, ggml_tensor * dst) { + GGML_SYCL_DEBUG("call %s\n", __func__); + ggml_sycl_op_l2_norm(ctx, dst); GGML_SYCL_DEBUG("call %s done\n", __func__); } static void ggml_sycl_group_norm(ggml_backend_sycl_context & ctx, ggml_tensor * dst) { GGML_SYCL_DEBUG("call %s\n", __func__); - ggml_sycl_op_flatten(ctx, dst->src[0], dst->src[1], dst, ggml_sycl_op_group_norm); + ggml_sycl_op_group_norm(ctx, dst); GGML_SYCL_DEBUG("call %s done\n", __func__); } @@ -3392,7 +2752,7 @@ static void ggml_sycl_mul_mat_batched_sycl(ggml_backend_sycl_context & ctx, // convert src1 to fp16 ggml_sycl_pool_alloc src1_f16_alloc(ctx.pool()); if (src1->type != GGML_TYPE_F16) { - const to_fp16_sycl_t to_fp16_sycl = ggml_get_to_fp16_sycl(src1->type); + const to_fp16_sycl_t to_fp16_sycl = ggml_get_to_fp16_sycl(src1->type, dst); const int64_t ne_src1 = ggml_nelements(src1); src1_f16_alloc.alloc(ne_src1); GGML_ASSERT(to_fp16_sycl != nullptr); @@ -3428,14 +2788,10 @@ static void ggml_sycl_mul_mat_batched_sycl(ggml_backend_sycl_context & ctx, if (r2 == 1 && r3 == 1 && ggml_is_contiguous_2(src0) && ggml_is_contiguous_2(src1)) { // there is no broadcast and src0, src1 are contiguous across dims 2, 3 SYCL_CHECK(CHECK_TRY_ERROR(dpct::gemm_batch( - *main_stream, oneapi::mkl::transpose::trans, - oneapi::mkl::transpose::nontrans, ne01, ne11, ne10, alpha, - (const char *)src0_as_f16, dpct::library_data_t::real_half, - nb01 / nb00, nb02 / nb00, - (const char *)src1_f16, dpct::library_data_t::real_half, - nb11 / nb10, nb12 / nb10, beta, - (char *)dst_t, cu_data_type, ne01, nb2 / nb0, - ne12 * ne13, cu_compute_type))); + *main_stream, oneapi::math::transpose::trans, oneapi::math::transpose::nontrans, ne01, ne11, ne10, alpha, + (const char *) src0_as_f16, dpct::library_data_t::real_half, nb01 / nb00, nb02 / nb00, + (const char *) src1_f16, dpct::library_data_t::real_half, nb11 / nb10, nb12 / nb10, beta, (char *) dst_t, + cu_data_type, ne01, nb2 / nb0, ne12 * ne13, cu_compute_type))); } else { const int ne23 = ne12*ne13; @@ -3470,7 +2826,7 @@ static void ggml_sycl_mul_mat_batched_sycl(ggml_backend_sycl_context & ctx, }); } SYCL_CHECK(CHECK_TRY_ERROR(dpct::gemm_batch( - *main_stream, oneapi::mkl::transpose::trans, oneapi::mkl::transpose::nontrans, ne01, ne11, ne10, alpha, + *main_stream, oneapi::math::transpose::trans, oneapi::math::transpose::nontrans, ne01, ne11, ne10, alpha, (const void **) (ptrs_src.get() + 0 * ne23), dpct::library_data_t::real_half, nb01 / nb00, (const void **) (ptrs_src.get() + 1 * ne23), dpct::library_data_t::real_half, nb11 / nb10, beta, (void **) (ptrs_dst.get() + 0 * ne23), cu_data_type, ne01, ne23, cu_compute_type, matrix_info.get()))); @@ -3488,7 +2844,7 @@ inline bool ggml_sycl_supports_mmq(enum ggml_type type) { return false; } -bool ggml_sycl_supports_dmmv(enum ggml_type type) { +static bool ggml_sycl_supports_dmmv(enum ggml_type type) { switch (type) { case GGML_TYPE_Q4_0: case GGML_TYPE_Q4_1: @@ -3508,6 +2864,7 @@ bool ggml_sycl_supports_dmmv(enum ggml_type type) { } static void ggml_sycl_mul_mat(ggml_backend_sycl_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { + const bool split = ggml_backend_buffer_is_sycl_split(src0->buffer); int64_t min_compute_capability = INT_MAX; @@ -3569,6 +2926,7 @@ static void ggml_sycl_mul_mat(ggml_backend_sycl_context & ctx, const ggml_tensor ggml_sycl_mul_mat_batched_sycl(ctx, src0, src1, dst); } else if (use_dequantize_mul_mat_vec) { ggml_sycl_op_mul_mat(ctx, src0, src1, dst, ggml_sycl_op_dequantize_mul_mat_vec, false); + // save_tensor_txt("1/dst_1.txt", (float*) dst->data, src0->ne[1], sizeof(float), ctx.stream()); } else if (use_mul_mat_vec_q) { ggml_sycl_op_mul_mat(ctx, src0, src1, dst, ggml_sycl_op_mul_mat_vec_q, true); } else if (use_mul_mat_q) { @@ -3701,8 +3059,8 @@ static void ggml_sycl_mul_mat_id(ggml_backend_sycl_context & ctx, const int64_t i2 = i12; src0_row.data = src0_original + i02*nb02; - src1_row.data = src1_original + + i11*nb11 + i12*nb12; - dst_row.data = dst_original + i1*nb1 + i2*nb2; + src1_row.data = src1_original + i11*nb11 + i12*nb12; + dst_row.data = dst_original + i1*nb1 + i2*nb2; ggml_sycl_mul_mat(ctx, &src0_row, &src1_row, &dst_row); } @@ -3814,104 +3172,48 @@ catch (sycl::exception const &exc) { } static void ggml_sycl_scale(ggml_backend_sycl_context & ctx, ggml_tensor * dst) { - ggml_sycl_op_flatten(ctx, dst->src[0], dst->src[1], dst, ggml_sycl_op_scale); -} - -static void ggml_sycl_clamp(ggml_backend_sycl_context & ctx, ggml_tensor * dst) { - ggml_sycl_op_flatten(ctx, dst->src[0], dst->src[1], dst, ggml_sycl_op_clamp); -} - -static void ggml_sycl_cpy(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1, - ggml_tensor *dst) try { - const int64_t ne = ggml_nelements(src0); - GGML_ASSERT(ne == ggml_nelements(src1)); - - GGML_ASSERT(ggml_nbytes(src0) <= INT_MAX); - GGML_ASSERT(ggml_nbytes(src1) <= INT_MAX); - - GGML_TENSOR_BINARY_OP_LOCALS01; - - SYCL_CHECK(ggml_sycl_set_device(ctx.device)); - queue_ptr main_stream = ctx.stream(); - - char * src0_ddc = (char *) src0->data; - char * src1_ddc = (char *) src1->data; - - if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32) { - ggml_cpy_f32_f32_sycl (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream); - } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F16) { - ggml_cpy_f32_f16_sycl (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream); - } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q8_0) { - ggml_cpy_f32_q8_0_sycl(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream); - } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q4_0) { - ggml_cpy_f32_q4_0_sycl(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream); - } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q4_1) { - ggml_cpy_f32_q4_1_sycl(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream); - } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F32) { - ggml_cpy_f16_f32_sycl (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream); - } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F16) { - ggml_cpy_f16_f16_sycl (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream); - } else if (src0->type == GGML_TYPE_I16 && src1->type == GGML_TYPE_I16) { - ggml_cpy_i16_i16_sycl (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream); - } else if (src0->type == GGML_TYPE_I32 && src1->type == GGML_TYPE_I32) { - ggml_cpy_i32_i32_sycl (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream); - } else { - GGML_LOG_ERROR("%s: unsupported type combination (%s to %s)\n", __func__, - ggml_type_name(src0->type), ggml_type_name(src1->type)); - GGML_ABORT("fatal error"); - } - GGML_UNUSED(dst); -} -catch (sycl::exception const &exc) { - std::cerr << exc.what() << "Exception caught at file:" << __FILE__ - << ", line:" << __LINE__ << std::endl; - std::exit(1); -} - -static void ggml_sycl_dup(ggml_backend_sycl_context & ctx, ggml_tensor * dst) { - // TODO: why do we pass dst as src1 here? - ggml_sycl_cpy(ctx, dst->src[0], dst, nullptr); + ggml_sycl_op_scale(ctx, dst); } static void ggml_sycl_diag_mask_inf(ggml_backend_sycl_context & ctx, ggml_tensor * dst) { - ggml_sycl_op_flatten(ctx, dst->src[0], dst->src[1], dst, ggml_sycl_op_diag_mask_inf); + ggml_sycl_op_diag_mask_inf(ctx, dst); } static void ggml_sycl_rope(ggml_backend_sycl_context & ctx, ggml_tensor * dst) { GGML_ASSERT(ggml_is_contiguous(dst->src[0])); // TODO: this restriction is temporary until non-cont support is implemented - ggml_sycl_op_flatten(ctx, dst->src[0], dst->src[1], dst, ggml_sycl_op_rope); + ggml_sycl_op_rope(ctx, dst); } static void ggml_sycl_pool2d(ggml_backend_sycl_context & ctx, ggml_tensor * dst) { - ggml_sycl_op_flatten(ctx, dst->src[0], dst->src[1], dst, ggml_sycl_op_pool2d); + ggml_sycl_op_pool2d(ctx, dst); } static void ggml_sycl_im2col(ggml_backend_sycl_context & ctx, ggml_tensor * dst) { - ggml_sycl_op_flatten(ctx, dst->src[0], dst->src[1], dst, ggml_sycl_op_im2col); + ggml_sycl_op_im2col(ctx, dst); } static void ggml_sycl_sum(ggml_backend_sycl_context & ctx, ggml_tensor * dst) { GGML_ASSERT(ggml_is_contiguous(dst->src[0])); - ggml_sycl_op_flatten(ctx, dst->src[0], dst->src[1], dst, ggml_sycl_op_sum); + ggml_sycl_op_sum(ctx, dst); } static void ggml_sycl_sum_rows(ggml_backend_sycl_context & ctx, ggml_tensor * dst) { GGML_ASSERT(ggml_is_contiguous(dst->src[0])); - ggml_sycl_op_flatten(ctx, dst->src[0], dst->src[1], dst, ggml_sycl_op_sum_rows); + ggml_sycl_op_sum_rows(ctx, dst); } static void ggml_sycl_argsort(ggml_backend_sycl_context & ctx, ggml_tensor * dst) { GGML_ASSERT(ggml_is_contiguous(dst->src[0])); - ggml_sycl_op_flatten(ctx, dst->src[0], dst->src[1], dst, ggml_sycl_op_argsort); + ggml_sycl_op_argsort(ctx, dst); } static void ggml_sycl_argmax(ggml_backend_sycl_context & ctx, ggml_tensor * dst) { GGML_ASSERT(ggml_is_contiguous(dst->src[0])); - ggml_sycl_op_flatten(ctx, dst->src[0], dst->src[1], dst, ggml_sycl_op_argmax); + ggml_sycl_op_argmax(ctx, dst); } -void ggml_sycl_set_main_device(const int main_device) try { +static void ggml_sycl_set_main_device(const int main_device) try { if (dpct::get_current_device_id() == static_cast (main_device)) { return; } @@ -3932,7 +3234,7 @@ catch (sycl::exception const &exc) { std::exit(1); } -bool ggml_sycl_compute_forward(ggml_backend_sycl_context & ctx, struct ggml_tensor * dst) { +static bool ggml_sycl_compute_forward(ggml_backend_sycl_context & ctx, struct ggml_tensor * dst) try { if (!g_sycl_loaded) return false; if (dst->src[0] != nullptr && ggml_backend_buffer_is_sycl_split(dst->src[0]->buffer)) { @@ -4034,6 +3336,9 @@ bool ggml_sycl_compute_forward(ggml_backend_sycl_context & ctx, struct ggml_tens case GGML_OP_RMS_NORM: ggml_sycl_rms_norm(ctx, dst); break; + case GGML_OP_L2_NORM: + ggml_sycl_l2_norm(ctx, dst); + break; case GGML_OP_MUL_MAT: if (dst->src[0]->ne[3] != dst->src[1]->ne[3]) { return false; @@ -4069,7 +3374,7 @@ bool ggml_sycl_compute_forward(ggml_backend_sycl_context & ctx, struct ggml_tens ggml_sycl_clamp(ctx, dst); break; case GGML_OP_CPY: - ggml_sycl_cpy(ctx, dst->src[0], dst->src[1], dst); + ggml_sycl_cpy(ctx, dst->src[0], dst->src[1]); break; case GGML_OP_CONT: ggml_sycl_dup(ctx, dst); @@ -4111,6 +3416,9 @@ bool ggml_sycl_compute_forward(ggml_backend_sycl_context & ctx, struct ggml_tens case GGML_OP_RWKV_WKV6: ggml_sycl_op_rwkv_wkv6(ctx, dst); break; + case GGML_OP_RWKV_WKV7: + ggml_sycl_op_rwkv_wkv7(ctx, dst); + break; case GGML_OP_GATED_LINEAR_ATTN: ggml_sycl_op_gated_linear_attn(ctx, dst); break; @@ -4119,6 +3427,9 @@ bool ggml_sycl_compute_forward(ggml_backend_sycl_context & ctx, struct ggml_tens } return true; +} catch (sycl::exception & e) { + std::cerr << e.what() << "Exception caught at file:" << __FILE__ << ", line:" << __LINE__ << std::endl; + std::exit(1); } GGML_API void ggml_backend_sycl_get_device_description(int device, char *description, @@ -4250,10 +3561,71 @@ catch (sycl::exception const &exc) { std::exit(1); } -static ggml_status ggml_backend_sycl_graph_compute(ggml_backend_t backend, ggml_cgraph * cgraph) { - ggml_backend_sycl_context * sycl_ctx = (ggml_backend_sycl_context *)backend->context; - ggml_sycl_set_main_device(sycl_ctx->device); +static void reorder_qw(char *data_device, const int ncols, const int nrows, + size_t size, size_t offset, dpct::queue_ptr stream) { + auto tmp_buf = sycl::malloc_shared(size, *stream); + SYCL_CHECK( + CHECK_TRY_ERROR((*stream).memcpy(tmp_buf, data_device, size) + .wait())); + GGML_ASSERT((size % sizeof(block_q4_0) == 0)); + GGML_ASSERT((offset % sizeof(block_q4_0) == 0)); + int offset_blks = offset / sizeof(block_q4_0); + auto qs_ptr = (uint8_t*)data_device + offset_blks * QK4_0 / 2;; + auto d_ptr = (sycl::half*)(qs_ptr + ncols * nrows / 2) + offset_blks; + stream->parallel_for( + size / sizeof(block_q4_0), + [=](auto i) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { + const block_q4_0* x = (const block_q4_0*)tmp_buf; + const int ib = i; + + for (int j = 0; j < QK4_0/2; j ++) + { + *(qs_ptr + ib * QK4_0 / 2 + j) = x[ib].qs[j]; + } + *(d_ptr + ib) = x[ib].d; + }); + + sycl::free(tmp_buf, *stream); +} + +static void reorder_qw(ggml_tensor * src0, dpct::queue_ptr stream) { + char*data_device = (char*)src0->data; + size_t ncols = src0->ne[0]; + size_t nrows = src0->ne[1]; + size_t size = ggml_nbytes(src0); + + reorder_qw(data_device, ncols, nrows, size, 0, stream); +} + +static void opt_for_reorder(ggml_tensor * dst, dpct::queue_ptr stream) { + ggml_tensor *src0 = dst->src[0]; + ggml_tensor *src1 = dst->src[1]; + + if (dst->op == GGML_OP_MUL_MAT && src0->type == GGML_TYPE_Q4_0 && + src1->ne[2]==1 && src1->ne[3]==1) { + reorder_qw(src0, stream); + ggml_tensor_extra_gpu* extra = (ggml_tensor_extra_gpu*)src0->extra; + GGML_ASSERT(extra); + extra->optimized_feature.reorder = true; //used to decode/dequan in next steps. + } +} + +static void optimize_graph_once(ggml_cgraph * cgraph, ggml_backend_sycl_context * ctx) { + dpct::queue_ptr stream = ctx->stream(); + if (ctx->optimized_graph) { + return; + } + ctx->optimized_graph = true; + + for (int i = 0; i < cgraph->n_nodes; i++) { + if (ctx->opt_feature.reorder) opt_for_reorder(cgraph->nodes[i], stream); + } +} + +static void ggml_backend_sycl_graph_compute_impl(ggml_backend_sycl_context * sycl_ctx, ggml_cgraph * cgraph) { + ggml_sycl_set_main_device(sycl_ctx->device); + if (!g_ggml_sycl_disable_optimize) optimize_graph_once(cgraph, sycl_ctx); for (int i = 0; i < cgraph->n_nodes; i++) { ggml_tensor * node = cgraph->nodes[i]; @@ -4274,7 +3646,49 @@ static ggml_status ggml_backend_sycl_graph_compute(ggml_backend_t backend, ggml_ } GGML_ASSERT(ok); } +} +static ggml_status ggml_backend_sycl_graph_compute(ggml_backend_t backend, ggml_cgraph * cgraph) { + auto * sycl_ctx = static_cast(backend->context); + +#ifdef GGML_SYCL_GRAPH + if (!g_ggml_sycl_disable_graph) { + const bool graph_support = dpct::get_device(sycl_ctx->device).has(sycl::aspect::ext_oneapi_limited_graph); + if (!graph_support) { + GGML_SYCL_DEBUG("[SYCL-GRAPH] can not use graphs on device:%d\n", sycl_ctx->device); + ggml_backend_sycl_graph_compute_impl(sycl_ctx, cgraph); + return GGML_STATUS_SUCCESS; + } + + sycl_ex::command_graph model_sycl_graph(*(sycl_ctx->stream())); + model_sycl_graph.begin_recording(*(sycl_ctx->stream())); + ggml_backend_sycl_graph_compute_impl(sycl_ctx, cgraph); + model_sycl_graph.end_recording(); + + const bool graph_update_support = dpct::get_device(sycl_ctx->device).has(sycl::aspect::ext_oneapi_graph); + if (!sycl_ctx->exec_graph || !graph_update_support) { + auto exec_graph = graph_update_support ? model_sycl_graph.finalize(sycl_ex::property::graph::updatable{}) : + model_sycl_graph.finalize(); + sycl_ctx->exec_graph = std::make_unique< + sycl_ex::command_graph>(exec_graph); + } else { + try { + sycl_ctx->exec_graph->update(model_sycl_graph); + GGML_SYCL_DEBUG("[SYCL-GRAPH] update success\n"); + } catch (sycl::exception const & e) { + GGML_SYCL_DEBUG("[SYCL-GRAPH] Exception when updating graph, %s\n", e.what()); + auto exec_graph = model_sycl_graph.finalize({sycl_ex::property::graph::updatable{}}); + sycl_ctx->exec_graph = std::make_unique< + sycl_ex::command_graph>(exec_graph); + } + } + + sycl_ctx->stream()->ext_oneapi_graph(*(sycl_ctx->exec_graph)); + } else +#endif + { + ggml_backend_sycl_graph_compute_impl(sycl_ctx, cgraph); + } return GGML_STATUS_SUCCESS; } @@ -4339,7 +3753,6 @@ bool ggml_backend_is_sycl(ggml_backend_t backend) { } int ggml_backend_sycl_get_device_count() { - GGML_SYCL_DEBUG("[SYCL] call ggml_backend_sycl_get_device_count\n"); return ggml_sycl_info().device_count; } @@ -4429,7 +3842,7 @@ static bool ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, const g return true; } return false; - } break; + } case GGML_OP_UNARY: switch (ggml_get_unary_op(op)) { case GGML_UNARY_OP_NEG: @@ -4443,11 +3856,14 @@ static bool ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, const g case GGML_UNARY_OP_GELU_QUICK: case GGML_UNARY_OP_TANH: case GGML_UNARY_OP_EXP: - return ggml_is_contiguous(op->src[0]); +#if defined (GGML_SYCL_F16) + return ggml_is_contiguous(op->src[0]) && (op->type == op->src[0]->type); +#else + return ggml_is_contiguous(op->src[0]) && (op->src[0]->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32) && (op->type == op->src[0]->type); +#endif default: return false; } - break; case GGML_OP_MUL_MAT: case GGML_OP_MUL_MAT_ID: { @@ -4478,7 +3894,7 @@ static bool ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, const g return false; } return true; - } break; + } case GGML_OP_OUT_PROD: return op->type == GGML_TYPE_F32 && op->src[0]->type == GGML_TYPE_F32 && op->src[1]->type == GGML_TYPE_F32 && op->ne[2] == 1 && op->ne[3] == 1; case GGML_OP_GET_ROWS: @@ -4495,7 +3911,7 @@ static bool ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, const g default: return false; } - } break; + } case GGML_OP_CPY: { ggml_type src0_type = op->src[0]->type; @@ -4521,13 +3937,37 @@ static bool ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, const g if (src0_type == GGML_TYPE_F16 && src1_type == GGML_TYPE_F32) { return true; } + if (src0_type == GGML_TYPE_Q8_0 && src1_type == GGML_TYPE_F32) { + return true; + } + if (src0_type == GGML_TYPE_Q4_0 && src1_type == GGML_TYPE_F32) { + return true; + } + if (src0_type == GGML_TYPE_Q4_1 && src1_type == GGML_TYPE_F32) { + return true; + } + if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_Q5_0) { + return true; + } + if (src0_type == GGML_TYPE_Q5_0 && src1_type == GGML_TYPE_F32) { + return true; + } + if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_Q5_1) { + return true; + } + if (src0_type == GGML_TYPE_Q5_1 && src1_type == GGML_TYPE_F32) { + return true; + } + if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_IQ4_NL) { + return true; + } return false; - } break; + } case GGML_OP_CONCAT: { ggml_type src0_type = op->src[0]->type; return src0_type != GGML_TYPE_I32 && src0_type != GGML_TYPE_I16; - } break; + } case GGML_OP_DUP: case GGML_OP_ARGMAX: case GGML_OP_NONE: @@ -4536,23 +3976,30 @@ static bool ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, const g case GGML_OP_VIEW: case GGML_OP_PERMUTE: case GGML_OP_TRANSPOSE: + return true; case GGML_OP_ADD: case GGML_OP_ADD1: - case GGML_OP_LOG: case GGML_OP_SUB: case GGML_OP_MUL: case GGML_OP_DIV: - return true; - case GGML_OP_NORM: - case GGML_OP_RMS_NORM: - case GGML_OP_GROUP_NORM: - return ggml_is_contiguous(op->src[0]); - case GGML_OP_SCALE: + return (op->src[0]->type == GGML_TYPE_F32); case GGML_OP_SQR: case GGML_OP_SQRT: case GGML_OP_SIN: case GGML_OP_COS: case GGML_OP_CLAMP: + case GGML_OP_LOG: +#if defined (GGML_SYCL_F16) + return ((op->type == GGML_TYPE_F32 || op->type == GGML_SYCL_F16) && (op->src[0]->type == GGML_TYPE_F32 || op->src[0]->type == GGML_SYCL_F16) && (op->type == op->src[0]->type)); +#else + return (op->type == GGML_TYPE_F32 && op->src[0]->type == GGML_TYPE_F32) && (op->type == op->src[0]->type); +#endif + case GGML_OP_NORM: + case GGML_OP_RMS_NORM: + case GGML_OP_L2_NORM: + case GGML_OP_GROUP_NORM: + return ggml_is_contiguous(op->src[0]); + case GGML_OP_SCALE: return true; case GGML_OP_CONT: return op->src[0]->type != GGML_TYPE_BF16; @@ -4573,16 +4020,18 @@ static bool ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, const g case GGML_OP_IM2COL: // TODO: add support for the new F32 operations return op->src[0]->type == GGML_TYPE_F16; + case GGML_OP_UPSCALE: + return op->src[0]->type == GGML_TYPE_F32 && op->op_params[0] == GGML_SCALE_MODE_NEAREST; case GGML_OP_POOL_2D: case GGML_OP_SUM: case GGML_OP_SUM_ROWS: case GGML_OP_ARGSORT: case GGML_OP_ACC: - case GGML_OP_UPSCALE: case GGML_OP_PAD: case GGML_OP_LEAKY_RELU: case GGML_OP_TIMESTEP_EMBEDDING: case GGML_OP_RWKV_WKV6: + case GGML_OP_RWKV_WKV7: case GGML_OP_GATED_LINEAR_ATTN: return true; default: diff --git a/ggml/src/ggml-sycl/im2col.cpp b/ggml/src/ggml-sycl/im2col.cpp index 6146a99ed..009b42035 100644 --- a/ggml/src/ggml-sycl/im2col.cpp +++ b/ggml/src/ggml-sycl/im2col.cpp @@ -82,10 +82,9 @@ static void im2col_sycl( } } -void ggml_sycl_op_im2col( - ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1, - ggml_tensor *dst, const float *src0_dd, const float *src1_dd, float *dst_dd, - const queue_ptr &main_stream) { +void ggml_sycl_op_im2col(ggml_backend_sycl_context & ctx, ggml_tensor *dst) { + const ggml_tensor * src0 = dst->src[0]; + const ggml_tensor * src1 = dst->src[1]; GGML_ASSERT(src0->type == GGML_TYPE_F16); GGML_ASSERT(src1->type == GGML_TYPE_F32); @@ -115,12 +114,8 @@ void ggml_sycl_op_im2col( const size_t batch_offset = src1->nb[3] / 4; // nb is byte offset, src is type float32 if (dst->type == GGML_TYPE_F16) { - im2col_sycl(src1_dd, (sycl::half *)dst_dd, IW, IH, OW, OH, KW, KH, IC, batch, batch_offset, delta_offset, s0, s1, p0, p1, d0, d1, main_stream); + im2col_sycl((const float *) src1->data, (sycl::half *)dst->data, IW, IH, OW, OH, KW, KH, IC, batch, batch_offset, delta_offset, s0, s1, p0, p1, d0, d1, ctx.stream()); } else { - im2col_sycl(src1_dd, (float *)dst_dd, IW, IH, OW, OH, KW, KH, IC, batch, batch_offset, delta_offset, s0, s1, p0, p1, d0, d1, main_stream); + im2col_sycl((const float *) src1->data, (float *)dst->data, IW, IH, OW, OH, KW, KH, IC, batch, batch_offset, delta_offset, s0, s1, p0, p1, d0, d1, ctx.stream()); } - - GGML_UNUSED(src0); - GGML_UNUSED(src0_dd); - GGML_UNUSED(ctx); } diff --git a/ggml/src/ggml-sycl/im2col.hpp b/ggml/src/ggml-sycl/im2col.hpp index 7db144fbb..dbbb248dd 100644 --- a/ggml/src/ggml-sycl/im2col.hpp +++ b/ggml/src/ggml-sycl/im2col.hpp @@ -16,8 +16,6 @@ #include "common.hpp" void ggml_sycl_op_im2col( - ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1, - ggml_tensor *dst, const float *src0_dd, const float *src1_dd, float *dst_dd, - const queue_ptr &main_stream); + ggml_backend_sycl_context & ctx, ggml_tensor *dst); #endif // GGML_SYCL_IM2COL_HPP diff --git a/ggml/src/ggml-sycl/mmq.cpp b/ggml/src/ggml-sycl/mmq.cpp index 8ea82c940..ffb272aa2 100644 --- a/ggml/src/ggml-sycl/mmq.cpp +++ b/ggml/src/ggml-sycl/mmq.cpp @@ -3017,7 +3017,6 @@ void ggml_sycl_op_mul_mat_q( break; default: GGML_ABORT("fatal error"); - break; } GGML_UNUSED(src1); diff --git a/ggml/src/ggml-sycl/mmvq.cpp b/ggml/src/ggml-sycl/mmvq.cpp index 221f65c21..1b92ba2d6 100644 --- a/ggml/src/ggml-sycl/mmvq.cpp +++ b/ggml/src/ggml-sycl/mmvq.cpp @@ -3,44 +3,42 @@ #include template -static void mul_mat_vec_q(const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst, const int ncols, const int nrows, - const sycl::nd_item<3> &item_ct1) { - const int row = item_ct1.get_group(2) * item_ct1.get_local_range(1) + - item_ct1.get_local_id(1); +static void mul_mat_vec_q(const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst, + const int ncols, const int nrows, const sycl::nd_item<3> & item_ct1) { + const int row = item_ct1.get_group(2) * item_ct1.get_local_range(1) + item_ct1.get_local_id(1); if (row >= nrows) { return; } - const int blocks_per_row = ncols / qk; - const int blocks_per_warp = vdr * QK_WARP_SIZE / qi; - assert(blocks_per_warp>0); + const int blocks_per_row = ncols / qk; + constexpr int blocks_per_warp = (vdr * WARP_SIZE + qi - 1) / qi; // Ensuring blocks_per_warp > 0 -// partial sum for each thread + assert(blocks_per_warp > 0); + + // partial sum for each thread float tmp = 0.0f; - const block_q_t * x = (const block_q_t *) vx; + const block_q_t * x = (const block_q_t *) vx; const block_q8_1 * y = (const block_q8_1 *) vy; - for (int i = item_ct1.get_local_id(2) / (qi / vdr); i < blocks_per_row; - i += blocks_per_warp) { - const int ibx = row*blocks_per_row + i; // x block index + for (int i = item_ct1.get_local_id(2) / (qi / vdr); i < blocks_per_row; i += blocks_per_warp) { + const int ibx = row * blocks_per_row + i; // x block index - const int iby = i * (qk/QK8_1); // y block index that aligns with ibx + const int iby = i * (qk / QK8_1); // y block index that aligns with ibx - const int iqs = - vdr * - (item_ct1.get_local_id(2) % - (qi / vdr)); // x block quant index when casting the quants to int + for (size_t elem = 0; elem < qi / vdr; elem += WARP_SIZE) { + const int iqs = elem + vdr * (item_ct1.get_local_id(2) % + (qi / vdr)); // x block quant index when casting the quants to int - tmp += vec_dot_q_sycl(&x[ibx], &y[iby], iqs); + tmp += vec_dot_q_sycl(&x[ibx], &y[iby], iqs); + } } // sum up partial sums and write back result #pragma unroll - for (int mask = QK_WARP_SIZE / 2; mask > 0; mask >>= 1) { - tmp += - dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); + for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) { + tmp += dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); } if (item_ct1.get_local_id(2) == 0) { @@ -62,7 +60,7 @@ static void mul_mat_vec_q_iq2_xxs_q8_1(const void *__restrict__ vx, } const int blocks_per_row = ncols / qk; - const int blocks_per_warp = vdr * QK_WARP_SIZE / qi; + const int blocks_per_warp = vdr * WARP_SIZE / qi; assert(blocks_per_warp>0); // partial sum for each thread @@ -87,7 +85,7 @@ static void mul_mat_vec_q_iq2_xxs_q8_1(const void *__restrict__ vx, // sum up partial sums and write back result #pragma unroll - for (int mask = QK_WARP_SIZE / 2; mask > 0; mask >>= 1) { + for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) { tmp += dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); } @@ -111,7 +109,7 @@ static void mul_mat_vec_q_iq2_xs_q8_1(const void *__restrict__ vx, } const int blocks_per_row = ncols / qk; - const int blocks_per_warp = vdr * QK_WARP_SIZE / qi; + const int blocks_per_warp = vdr * WARP_SIZE / qi; assert(blocks_per_warp>0); // partial sum for each thread float tmp = 0.0f; @@ -135,7 +133,7 @@ static void mul_mat_vec_q_iq2_xs_q8_1(const void *__restrict__ vx, // sum up partial sums and write back result #pragma unroll - for (int mask = QK_WARP_SIZE / 2; mask > 0; mask >>= 1) { + for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) { tmp += dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); } @@ -159,7 +157,7 @@ static void mul_mat_vec_q_iq2_s_q8_1(const void *__restrict__ vx, } const int blocks_per_row = ncols / qk; - const int blocks_per_warp = vdr * QK_WARP_SIZE / qi; + const int blocks_per_warp = vdr * WARP_SIZE / qi; assert(blocks_per_warp>0); // partial sum for each thread float tmp = 0.0f; @@ -183,7 +181,7 @@ static void mul_mat_vec_q_iq2_s_q8_1(const void *__restrict__ vx, // sum up partial sums and write back result #pragma unroll - for (int mask = QK_WARP_SIZE / 2; mask > 0; mask >>= 1) { + for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) { tmp += dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); } @@ -207,7 +205,7 @@ static void mul_mat_vec_q_iq3_xxs_q8_1(const void *__restrict__ vx, } const int blocks_per_row = ncols / qk; - const int blocks_per_warp = vdr * QK_WARP_SIZE / qi; + const int blocks_per_warp = vdr * WARP_SIZE / qi; assert(blocks_per_warp>0); // partial sum for each thread float tmp = 0.0f; @@ -231,7 +229,7 @@ static void mul_mat_vec_q_iq3_xxs_q8_1(const void *__restrict__ vx, // sum up partial sums and write back result #pragma unroll - for (int mask = QK_WARP_SIZE / 2; mask > 0; mask >>= 1) { + for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) { tmp += dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); } @@ -255,7 +253,7 @@ static void mul_mat_vec_q_iq3_s_q8_1(const void *__restrict__ vx, } const int blocks_per_row = ncols / qk; - const int blocks_per_warp = vdr * QK_WARP_SIZE / qi; + const int blocks_per_warp = vdr * WARP_SIZE / qi; assert(blocks_per_warp>0); // partial sum for each thread float tmp = 0.0f; @@ -279,7 +277,7 @@ static void mul_mat_vec_q_iq3_s_q8_1(const void *__restrict__ vx, // sum up partial sums and write back result #pragma unroll - for (int mask = QK_WARP_SIZE / 2; mask > 0; mask >>= 1) { + for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) { tmp += dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); } @@ -303,7 +301,7 @@ static void mul_mat_vec_q_iq1_s_q8_1(const void *__restrict__ vx, } const int blocks_per_row = ncols / qk; - const int blocks_per_warp = vdr * QK_WARP_SIZE / qi; + const int blocks_per_warp = vdr * WARP_SIZE / qi; assert(blocks_per_warp>0); // partial sum for each thread float tmp = 0.0f; @@ -327,7 +325,7 @@ static void mul_mat_vec_q_iq1_s_q8_1(const void *__restrict__ vx, // sum up partial sums and write back result #pragma unroll - for (int mask = QK_WARP_SIZE / 2; mask > 0; mask >>= 1) { + for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) { tmp += dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); } @@ -351,7 +349,7 @@ static void mul_mat_vec_q_iq1_m_q8_1(const void *__restrict__ vx, } const int blocks_per_row = ncols / qk; - const int blocks_per_warp = vdr * QK_WARP_SIZE / qi; + const int blocks_per_warp = vdr * WARP_SIZE / qi; assert(blocks_per_warp>0); // partial sum for each thread float tmp = 0.0f; @@ -375,7 +373,7 @@ static void mul_mat_vec_q_iq1_m_q8_1(const void *__restrict__ vx, // sum up partial sums and write back result #pragma unroll - for (int mask = QK_WARP_SIZE / 2; mask > 0; mask >>= 1) { + for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) { tmp += dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); } @@ -399,7 +397,7 @@ static void mul_mat_vec_q_iq4_nl_q8_1(const void *__restrict__ vx, } const int blocks_per_row = ncols / qk; - const int blocks_per_warp = vdr * QK_WARP_SIZE / qi; + const int blocks_per_warp = vdr * WARP_SIZE / qi; assert(blocks_per_warp>0); // partial sum for each thread float tmp = 0.0f; @@ -423,7 +421,7 @@ static void mul_mat_vec_q_iq4_nl_q8_1(const void *__restrict__ vx, // sum up partial sums and write back result #pragma unroll - for (int mask = QK_WARP_SIZE / 2; mask > 0; mask >>= 1) { + for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) { tmp += dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); } @@ -448,7 +446,7 @@ static void mul_mat_vec_q_iq4_xs_q8_1(const void *__restrict__ vx, } const int blocks_per_row = ncols / qk; - const int blocks_per_warp = vdr * QK_WARP_SIZE / qi; + const int blocks_per_warp = vdr * WARP_SIZE / qi; assert(blocks_per_warp>0); // partial sum for each thread float tmp = 0.0f; @@ -472,7 +470,7 @@ static void mul_mat_vec_q_iq4_xs_q8_1(const void *__restrict__ vx, // sum up partial sums and write back result #pragma unroll - for (int mask = QK_WARP_SIZE / 2; mask > 0; mask >>= 1) { + for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) { tmp += dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); } @@ -489,7 +487,7 @@ static void mul_mat_vec_q4_0_q8_1_sycl(const void *vx, const void *vy, GGML_ASSERT(ncols % QK4_0 == 0); const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; const sycl::range<3> block_nums(1, 1, block_num_y); - const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, QK_WARP_SIZE); + const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); { stream->submit([&](sycl::handler &cgh) { @@ -497,7 +495,7 @@ static void mul_mat_vec_q4_0_q8_1_sycl(const void *vx, const void *vy, cgh.parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) - [[intel::reqd_sub_group_size(QK_WARP_SIZE)]] { + [[sycl::reqd_sub_group_size(WARP_SIZE)]] { mul_mat_vec_q( vx, vy, dst, ncols, nrows, item_ct1); @@ -513,7 +511,7 @@ static void mul_mat_vec_q4_1_q8_1_sycl(const void *vx, const void *vy, GGML_ASSERT(ncols % QK4_1 == 0); const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; const sycl::range<3> block_nums(1, 1, block_num_y); - const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, QK_WARP_SIZE); + const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); { stream->submit([&](sycl::handler &cgh) { @@ -521,7 +519,7 @@ static void mul_mat_vec_q4_1_q8_1_sycl(const void *vx, const void *vy, cgh.parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) - [[intel::reqd_sub_group_size(QK_WARP_SIZE)]] { + [[sycl::reqd_sub_group_size(WARP_SIZE)]] { mul_mat_vec_q( vx, vy, dst, ncols, nrows, item_ct1); @@ -537,7 +535,7 @@ static void mul_mat_vec_q5_0_q8_1_sycl(const void *vx, const void *vy, GGML_ASSERT(ncols % QK5_0 == 0); const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; const sycl::range<3> block_nums(1, 1, block_num_y); - const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, QK_WARP_SIZE); + const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); { stream->submit([&](sycl::handler &cgh) { @@ -545,7 +543,7 @@ static void mul_mat_vec_q5_0_q8_1_sycl(const void *vx, const void *vy, cgh.parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) - [[intel::reqd_sub_group_size(QK_WARP_SIZE)]] { + [[sycl::reqd_sub_group_size(WARP_SIZE)]] { mul_mat_vec_q( vx, vy, dst, ncols, nrows, item_ct1); @@ -561,7 +559,7 @@ static void mul_mat_vec_q5_1_q8_1_sycl(const void *vx, const void *vy, GGML_ASSERT(ncols % QK5_1 == 0); const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; const sycl::range<3> block_nums(1, 1, block_num_y); - const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, QK_WARP_SIZE); + const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); { stream->submit([&](sycl::handler &cgh) { @@ -569,7 +567,7 @@ static void mul_mat_vec_q5_1_q8_1_sycl(const void *vx, const void *vy, cgh.parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) - [[intel::reqd_sub_group_size(QK_WARP_SIZE)]] { + [[sycl::reqd_sub_group_size(WARP_SIZE)]] { mul_mat_vec_q( vx, vy, dst, ncols, nrows, item_ct1); @@ -585,7 +583,7 @@ static void mul_mat_vec_q8_0_q8_1_sycl(const void *vx, const void *vy, GGML_ASSERT(ncols % QK8_0 == 0); const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; const sycl::range<3> block_nums(1, 1, block_num_y); - const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, QK_WARP_SIZE); + const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); { stream->submit([&](sycl::handler &cgh) { @@ -593,7 +591,7 @@ static void mul_mat_vec_q8_0_q8_1_sycl(const void *vx, const void *vy, cgh.parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) - [[intel::reqd_sub_group_size(QK_WARP_SIZE)]] { + [[sycl::reqd_sub_group_size(WARP_SIZE)]] { mul_mat_vec_q( vx, vy, dst, ncols, nrows, item_ct1); @@ -609,7 +607,7 @@ static void mul_mat_vec_q2_K_q8_1_sycl(const void *vx, const void *vy, GGML_ASSERT(ncols % QK_K == 0); const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; const sycl::range<3> block_nums(1, 1, block_num_y); - const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, QK_WARP_SIZE); + const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); { stream->submit([&](sycl::handler &cgh) { @@ -617,7 +615,7 @@ static void mul_mat_vec_q2_K_q8_1_sycl(const void *vx, const void *vy, cgh.parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) - [[intel::reqd_sub_group_size(QK_WARP_SIZE)]] { + [[sycl::reqd_sub_group_size(WARP_SIZE)]] { mul_mat_vec_q( vx, vy, dst, ncols, nrows, item_ct1); @@ -633,7 +631,7 @@ static void mul_mat_vec_q3_K_q8_1_sycl(const void *vx, const void *vy, GGML_ASSERT(ncols % QK_K == 0); const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; const sycl::range<3> block_nums(1, 1, block_num_y); - const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, QK_WARP_SIZE); + const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); { stream->submit([&](sycl::handler &cgh) { @@ -641,7 +639,7 @@ static void mul_mat_vec_q3_K_q8_1_sycl(const void *vx, const void *vy, cgh.parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) - [[intel::reqd_sub_group_size(QK_WARP_SIZE)]] { + [[sycl::reqd_sub_group_size(WARP_SIZE)]] { mul_mat_vec_q( vx, vy, dst, ncols, nrows, item_ct1); @@ -657,7 +655,7 @@ static void mul_mat_vec_q4_K_q8_1_sycl(const void *vx, const void *vy, GGML_ASSERT(ncols % QK_K == 0); const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; const sycl::range<3> block_nums(1, 1, block_num_y); - const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, QK_WARP_SIZE); + const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); { stream->submit([&](sycl::handler &cgh) { @@ -665,7 +663,7 @@ static void mul_mat_vec_q4_K_q8_1_sycl(const void *vx, const void *vy, cgh.parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) - [[intel::reqd_sub_group_size(QK_WARP_SIZE)]] { + [[sycl::reqd_sub_group_size(WARP_SIZE)]] { mul_mat_vec_q( vx, vy, dst, ncols, nrows, item_ct1); @@ -681,7 +679,7 @@ static void mul_mat_vec_q5_K_q8_1_sycl(const void *vx, const void *vy, GGML_ASSERT(ncols % QK_K == 0); const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; const sycl::range<3> block_nums(1, 1, block_num_y); - const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, QK_WARP_SIZE); + const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); { stream->submit([&](sycl::handler &cgh) { @@ -689,7 +687,7 @@ static void mul_mat_vec_q5_K_q8_1_sycl(const void *vx, const void *vy, cgh.parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) - [[intel::reqd_sub_group_size(QK_WARP_SIZE)]] { + [[sycl::reqd_sub_group_size(WARP_SIZE)]] { mul_mat_vec_q( vx, vy, dst, ncols, nrows, item_ct1); @@ -705,7 +703,7 @@ static void mul_mat_vec_q6_K_q8_1_sycl(const void *vx, const void *vy, GGML_ASSERT(ncols % QK_K == 0); const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; const sycl::range<3> block_nums(1, 1, block_num_y); - const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, QK_WARP_SIZE); + const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); { stream->submit([&](sycl::handler &cgh) { @@ -713,7 +711,7 @@ static void mul_mat_vec_q6_K_q8_1_sycl(const void *vx, const void *vy, cgh.parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) - [[intel::reqd_sub_group_size(QK_WARP_SIZE)]] { + [[sycl::reqd_sub_group_size(WARP_SIZE)]] { mul_mat_vec_q( vx, vy, dst, ncols, nrows, item_ct1); @@ -730,13 +728,13 @@ static void mul_mat_vec_iq2_xxs_q8_1_sycl(const void *vx, const void *vy, GGML_ASSERT(ncols % QK_K == 0); const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; const sycl::range<3> block_nums(1, 1, block_num_y); - const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, QK_WARP_SIZE); + const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); { stream->submit([&](sycl::handler &cgh) { cgh.parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) - [[intel::reqd_sub_group_size(QK_WARP_SIZE)]] { + [[sycl::reqd_sub_group_size(WARP_SIZE)]] { mul_mat_vec_q_iq2_xxs_q8_1( vx, vy, dst, ncols, nrows, item_ct1); }); @@ -751,13 +749,13 @@ static void mul_mat_vec_iq2_xs_q8_1_sycl(const void *vx, const void *vy, GGML_ASSERT(ncols % QK_K == 0); const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; const sycl::range<3> block_nums(1, 1, block_num_y); - const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, QK_WARP_SIZE); + const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); { stream->submit([&](sycl::handler & cgh) { cgh.parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) - [[intel::reqd_sub_group_size(QK_WARP_SIZE)]] { + [[sycl::reqd_sub_group_size(WARP_SIZE)]] { mul_mat_vec_q_iq2_xs_q8_1( vx, vy, dst, ncols, nrows, item_ct1); }); @@ -772,14 +770,14 @@ static void mul_mat_vec_iq2_s_q8_1_sycl(const void *vx, const void *vy, GGML_ASSERT(ncols % QK_K == 0); const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; const sycl::range<3> block_nums(1, 1, block_num_y); - const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, QK_WARP_SIZE); + const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); { stream->submit([&](sycl::handler &cgh) { cgh.parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) - [[intel::reqd_sub_group_size(QK_WARP_SIZE)]] { + [[sycl::reqd_sub_group_size(WARP_SIZE)]] { mul_mat_vec_q_iq2_s_q8_1( vx, vy, dst, ncols, nrows, item_ct1); }); @@ -794,14 +792,14 @@ static void mul_mat_vec_iq3_xxs_q8_1_sycl(const void *vx, const void *vy, GGML_ASSERT(ncols % QK_K == 0); const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; const sycl::range<3> block_nums(1, 1, block_num_y); - const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, QK_WARP_SIZE); + const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); { stream->submit([&](sycl::handler &cgh) { cgh.parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) - [[intel::reqd_sub_group_size(QK_WARP_SIZE)]] { + [[sycl::reqd_sub_group_size(WARP_SIZE)]] { mul_mat_vec_q_iq3_xxs_q8_1( vx, vy, dst, ncols, nrows, item_ct1); }); @@ -816,14 +814,14 @@ static void mul_mat_vec_iq3_s_q8_1_sycl(const void *vx, const void *vy, GGML_ASSERT(ncols % QK_K == 0); const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; const sycl::range<3> block_nums(1, 1, block_num_y); - const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, QK_WARP_SIZE); + const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); { stream->submit([&](sycl::handler &cgh) { cgh.parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) - [[intel::reqd_sub_group_size(QK_WARP_SIZE)]] { + [[sycl::reqd_sub_group_size(WARP_SIZE)]] { mul_mat_vec_q_iq3_s_q8_1( vx, vy, dst, ncols, nrows, item_ct1); }); @@ -838,14 +836,14 @@ static void mul_mat_vec_iq1_s_q8_1_sycl(const void *vx, const void *vy, GGML_ASSERT(ncols % QK_K == 0); const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; const sycl::range<3> block_nums(1, 1, block_num_y); - const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, QK_WARP_SIZE); + const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); { stream->submit([&](sycl::handler &cgh) { cgh.parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) - [[intel::reqd_sub_group_size(QK_WARP_SIZE)]] { + [[sycl::reqd_sub_group_size(WARP_SIZE)]] { mul_mat_vec_q_iq1_s_q8_1( vx, vy, dst, ncols, nrows, item_ct1); }); @@ -860,13 +858,13 @@ static void mul_mat_vec_iq1_m_q8_1_sycl(const void *vx, const void *vy, GGML_ASSERT(ncols % QK_K == 0); const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; const sycl::range<3> block_nums(1, 1, block_num_y); - const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, QK_WARP_SIZE); + const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); { stream->submit([&](sycl::handler &cgh) { cgh.parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) - [[intel::reqd_sub_group_size(QK_WARP_SIZE)]] { + [[sycl::reqd_sub_group_size(WARP_SIZE)]] { mul_mat_vec_q_iq1_m_q8_1( vx, vy, dst, ncols, nrows, item_ct1); }); @@ -881,14 +879,14 @@ static void mul_mat_vec_iq4_nl_q8_1_sycl(const void *vx, const void *vy, GGML_ASSERT(ncols % QK4_NL == 0); const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; const sycl::range<3> block_nums(1, 1, block_num_y); - const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, QK_WARP_SIZE); + const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); { stream->submit([&](sycl::handler &cgh) { cgh.parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) - [[intel::reqd_sub_group_size(QK_WARP_SIZE)]] { + [[sycl::reqd_sub_group_size(WARP_SIZE)]] { mul_mat_vec_q_iq4_nl_q8_1( vx, vy, dst, ncols, nrows, item_ct1); }); @@ -903,14 +901,14 @@ static void mul_mat_vec_iq4_xs_q8_1_sycl(const void *vx, const void *vy, GGML_ASSERT(ncols % QK_K == 0); const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; const sycl::range<3> block_nums(1, 1, block_num_y); - const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, QK_WARP_SIZE); + const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); { stream->submit([&](sycl::handler &cgh) { cgh.parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) - [[intel::reqd_sub_group_size(QK_WARP_SIZE)]] { + [[sycl::reqd_sub_group_size(WARP_SIZE)]] { mul_mat_vec_q_iq4_xs_q8_1( vx, vy, dst, ncols, nrows, item_ct1); }); @@ -1005,7 +1003,6 @@ void ggml_sycl_op_mul_mat_vec_q( break; default: GGML_ABORT("fatal error"); - break; } } GGML_UNUSED(src1); diff --git a/ggml/src/ggml-sycl/norm.cpp b/ggml/src/ggml-sycl/norm.cpp index 9cf2be155..4e9f438b4 100644 --- a/ggml/src/ggml-sycl/norm.cpp +++ b/ggml/src/ggml-sycl/norm.cpp @@ -180,6 +180,50 @@ static void rms_norm_f32(const float* x, float* dst, const int ncols, const floa } } +static void l2_norm_f32(const float* x, float* dst, const int ncols, const float eps, + const sycl::nd_item<3>& item_ct1, float* s_sum, int block_size) { + const int row = item_ct1.get_group(2) * item_ct1.get_local_range(1) + + item_ct1.get_local_id(1); + const int tid = item_ct1.get_local_id(2); + const int nthreads = item_ct1.get_local_range(2); + const int nwarps = nthreads / WARP_SIZE; + float tmp = 0.0f; // partial sum for thread in warp + + for (int col = tid; col < ncols; col += block_size) { + const float xi = x[row * ncols + col]; + tmp += xi * xi; + } + + // sum up partial sums + tmp = warp_reduce_sum(tmp, item_ct1); + if (block_size > WARP_SIZE) { + + int warp_id = item_ct1.get_local_id(2) / WARP_SIZE; + int lane_id = item_ct1.get_local_id(2) % WARP_SIZE; + if (lane_id == 0) { + s_sum[warp_id] = tmp; + } + /* + DPCT1118:3: SYCL group functions and algorithms must be encountered in + converged control flow. You may need to adjust the code. + */ + item_ct1.barrier(sycl::access::fence_space::local_space); + size_t nreduce = nwarps / WARP_SIZE; + tmp = 0.f; + for (size_t i = 0; i < nreduce; i += 1) + { + tmp += s_sum[lane_id + i * WARP_SIZE]; + } + tmp = warp_reduce_sum(tmp, item_ct1); + } + + const float scale = sycl::rsqrt(sycl::max(tmp, eps * eps)); + + for (int col = tid; col < ncols; col += block_size) { + dst[row * ncols + col] = scale * x[row * ncols + col]; + } +} + static void norm_f32_sycl(const float* x, float* dst, const int ncols, const int nrows, const float eps, queue_ptr stream, int device) { @@ -191,7 +235,7 @@ static void norm_f32_sycl(const float* x, float* dst, const int ncols, sycl::nd_range<3>(sycl::range<3>(1, 1, nrows) * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) - [[intel::reqd_sub_group_size(WARP_SIZE)]] { + [[sycl::reqd_sub_group_size(WARP_SIZE)]] { norm_f32(x, dst, ncols, eps, item_ct1, nullptr, WARP_SIZE); }); @@ -214,7 +258,7 @@ static void norm_f32_sycl(const float* x, float* dst, const int ncols, sycl::nd_range<3>(sycl::range<3>(1, 1, nrows) * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) - [[intel::reqd_sub_group_size(WARP_SIZE)]] { + [[sycl::reqd_sub_group_size(WARP_SIZE)]] { norm_f32(x, dst, ncols, eps, item_ct1, get_pointer(s_sum_acc_ct1), work_group_size); }); @@ -233,7 +277,7 @@ static void group_norm_f32_sycl(const float* x, float* dst, sycl::nd_range<3>(sycl::range<3>(1, 1, num_groups) * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) - [[intel::reqd_sub_group_size(WARP_SIZE)]] { + [[sycl::reqd_sub_group_size(WARP_SIZE)]] { group_norm_f32( x, dst, group_size, ne_elements, eps_ct4, item_ct1, nullptr, WARP_SIZE); @@ -260,7 +304,7 @@ static void group_norm_f32_sycl(const float* x, float* dst, sycl::nd_range<3>(sycl::range<3>(1, 1, num_groups) * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) - [[intel::reqd_sub_group_size(WARP_SIZE)]] { + [[sycl::reqd_sub_group_size(WARP_SIZE)]] { group_norm_f32(x, dst, group_size, ne_elements, eps_ct4, item_ct1, get_pointer(s_sum_acc_ct1), work_group_size); @@ -281,7 +325,7 @@ static void rms_norm_f32_sycl(const float* x, float* dst, const int ncols, sycl::nd_range<3>(sycl::range<3>(1, 1, nrows) * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) - [[intel::reqd_sub_group_size(WARP_SIZE)]] { + [[sycl::reqd_sub_group_size(WARP_SIZE)]] { rms_norm_f32(x, dst, ncols, eps, item_ct1, nullptr, WARP_SIZE); }); @@ -303,7 +347,7 @@ static void rms_norm_f32_sycl(const float* x, float* dst, const int ncols, sycl::nd_range<3>(sycl::range<3>(1, 1, nrows) * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) - [[intel::reqd_sub_group_size(WARP_SIZE)]] { + [[sycl::reqd_sub_group_size(WARP_SIZE)]] { rms_norm_f32(x, dst, ncols, eps, item_ct1, get_pointer(s_sum_acc_ct1), work_group_size); }); @@ -311,68 +355,120 @@ static void rms_norm_f32_sycl(const float* x, float* dst, const int ncols, } } -void ggml_sycl_op_norm(ggml_backend_sycl_context& ctx, const ggml_tensor* src0, const ggml_tensor* src1, - ggml_tensor* dst, const float* src0_dd, - const float* src1_dd, float* dst_dd, - const queue_ptr& main_stream) { +static void l2_norm_f32_sycl(const float* x, float* dst, const int ncols, + const int nrows, const float eps, + queue_ptr stream, int device) { + GGML_ASSERT(ncols % WARP_SIZE == 0); + // printf("%s ncols=%d, nrows=%d, WARP_SIZE=%d\n", __func__, ncols, nrows, WARP_SIZE); + if (ncols < 1024) { + const sycl::range<3> block_dims(1, 1, WARP_SIZE); + stream->submit([&](sycl::handler& cgh) { + cgh.parallel_for( + sycl::nd_range<3>(sycl::range<3>(1, 1, nrows) * block_dims, + block_dims), + [=](sycl::nd_item<3> item_ct1) + [[sycl::reqd_sub_group_size(WARP_SIZE)]] { + l2_norm_f32(x, dst, ncols, eps, item_ct1, + nullptr, WARP_SIZE); + }); + }); + } + else { + const int work_group_size = ggml_sycl_info().max_work_group_sizes[device]; + assert(work_group_size % (WARP_SIZE * WARP_SIZE) == 0); + const sycl::range<3> block_dims(1, 1, work_group_size); + /* + DPCT1049:19: The work-group size passed to the SYCL kernel may exceed + the limit. To get the device limit, query + info::device::max_work_group_size. Adjust the work-group size if needed. + */ + stream->submit([&](sycl::handler& cgh) { + sycl::local_accessor s_sum_acc_ct1(sycl::range<1>(work_group_size / WARP_SIZE), + cgh); + cgh.parallel_for( + sycl::nd_range<3>(sycl::range<3>(1, 1, nrows) * block_dims, + block_dims), + [=](sycl::nd_item<3> item_ct1) + [[sycl::reqd_sub_group_size(WARP_SIZE)]] { + l2_norm_f32(x, dst, ncols, eps, item_ct1, + get_pointer(s_sum_acc_ct1), work_group_size); + }); + }); + } +} - GGML_ASSERT(src0->type == GGML_TYPE_F32); +void ggml_sycl_op_norm(ggml_backend_sycl_context& ctx, ggml_tensor* dst) { + + GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32); GGML_ASSERT(dst->type == GGML_TYPE_F32); - const int64_t ne00 = src0->ne[0]; - const int64_t nrows = ggml_nrows(src0); + const int64_t ne00 = dst->src[0]->ne[0]; + const int64_t nrows = ggml_nrows(dst->src[0]); + dpct::queue_ptr main_stream = ctx.stream(); + SYCL_CHECK(ggml_sycl_set_device(ctx.device)); + const float * src0_dd = static_cast(dst->src[0]->data); + float * dst_dd = static_cast(dst->data); float eps; memcpy(&eps, dst->op_params, sizeof(float)); norm_f32_sycl(src0_dd, dst_dd, ne00, nrows, eps, main_stream, ctx.device); - - (void)src1; - (void)dst; - (void)src1_dd; } -void ggml_sycl_op_group_norm(ggml_backend_sycl_context& ctx, const ggml_tensor* src0, - const ggml_tensor* src1, ggml_tensor* dst, - const float* src0_dd, const float* src1_dd, - float* dst_dd, - const queue_ptr& main_stream) { +void ggml_sycl_op_group_norm(ggml_backend_sycl_context& ctx, ggml_tensor* dst) { - GGML_ASSERT(src0->type == GGML_TYPE_F32); + GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32); GGML_ASSERT(dst->type == GGML_TYPE_F32); int num_groups = dst->op_params[0]; + dpct::queue_ptr main_stream = ctx.stream(); + SYCL_CHECK(ggml_sycl_set_device(ctx.device)); + + const float * src0_dd = static_cast(dst->src[0]->data); + float * dst_dd = static_cast(dst->data); float eps; memcpy(&eps, dst->op_params + 1, sizeof(float)); - int group_size = src0->ne[0] * src0->ne[1] * ((src0->ne[2] + num_groups - 1) / num_groups); - group_norm_f32_sycl(src0_dd, dst_dd, num_groups, eps, group_size, src0->ne[0] * src0->ne[1] * src0->ne[2], main_stream, ctx.device); - - (void)src1; - (void)dst; - (void)src1_dd; - GGML_UNUSED(ctx); + int group_size = dst->src[0]->ne[0] * dst->src[0]->ne[1] * ((dst->src[0]->ne[2] + num_groups - 1) / num_groups); + group_norm_f32_sycl(src0_dd, dst_dd, num_groups, eps, group_size, dst->src[0]->ne[0] * dst->src[0]->ne[1] * dst->src[0]->ne[2], main_stream, ctx.device); } -void ggml_sycl_op_rms_norm(ggml_backend_sycl_context& ctx, const ggml_tensor* src0, - const ggml_tensor* src1, ggml_tensor* dst, - const float* src0_dd, const float* src1_dd, - float* dst_dd, - const queue_ptr& main_stream) { +void ggml_sycl_op_rms_norm(ggml_backend_sycl_context & ctx, ggml_tensor * dst) { - GGML_ASSERT(src0->type == GGML_TYPE_F32); + GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32); GGML_ASSERT(dst->type == GGML_TYPE_F32); - const int64_t ne00 = src0->ne[0]; - const int64_t nrows = ggml_nrows(src0); + const int64_t ne00 = dst->src[0]->ne[0]; + const int64_t nrows = ggml_nrows(dst->src[0]); + dpct::queue_ptr main_stream = ctx.stream(); + SYCL_CHECK(ggml_sycl_set_device(ctx.device)); + + const float * src0_dd = static_cast(dst->src[0]->data); + float * dst_dd = static_cast(dst->data); float eps; memcpy(&eps, dst->op_params, sizeof(float)); rms_norm_f32_sycl(src0_dd, dst_dd, ne00, nrows, eps, main_stream, ctx.device); - - (void)src1; - (void)dst; - (void)src1_dd; +} + +void ggml_sycl_op_l2_norm(ggml_backend_sycl_context& ctx, ggml_tensor* dst) { + + GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32); + GGML_ASSERT(dst->type == GGML_TYPE_F32); + + dpct::queue_ptr main_stream = ctx.stream(); + SYCL_CHECK(ggml_sycl_set_device(ctx.device)); + + const int64_t ne00 = dst->src[0]->ne[0]; + const int64_t nrows = ggml_nrows(dst->src[0]); + const float * src0_dd = static_cast(dst->src[0]->data); + float * dst_dd = static_cast(dst->data); + + float eps; + memcpy(&eps, dst->op_params, sizeof(float)); + + l2_norm_f32_sycl(src0_dd, dst_dd, ne00, nrows, eps, main_stream, ctx.device); + } diff --git a/ggml/src/ggml-sycl/norm.hpp b/ggml/src/ggml-sycl/norm.hpp index a9ad9156f..612cd67cf 100644 --- a/ggml/src/ggml-sycl/norm.hpp +++ b/ggml/src/ggml-sycl/norm.hpp @@ -15,21 +15,12 @@ #include "common.hpp" -void ggml_sycl_op_norm(ggml_backend_sycl_context& ctx, const ggml_tensor* src0, const ggml_tensor* src1, - ggml_tensor* dst, const float* src0_dd, - const float* src1_dd, float* dst_dd, - const queue_ptr& main_stream); +void ggml_sycl_op_norm(ggml_backend_sycl_context& ctx, ggml_tensor* dst); -void ggml_sycl_op_rms_norm(ggml_backend_sycl_context& ctx, const ggml_tensor* src0, - const ggml_tensor* src1, ggml_tensor* dst, - const float* src0_dd, const float* src1_dd, - float* dst_dd, - const queue_ptr& main_stream); +void ggml_sycl_op_rms_norm(ggml_backend_sycl_context& ctx, ggml_tensor* dst); -void ggml_sycl_op_group_norm(ggml_backend_sycl_context& ctx, const ggml_tensor* src0, - const ggml_tensor* src1, ggml_tensor* dst, - const float* src0_dd, const float* src1_dd, - float* dst_dd, - const queue_ptr& main_stream); +void ggml_sycl_op_group_norm(ggml_backend_sycl_context& ctx, ggml_tensor* dst); + +void ggml_sycl_op_l2_norm(ggml_backend_sycl_context& ctx, ggml_tensor* dst); #endif // GGML_SYCL_NORM_HPP diff --git a/ggml/src/ggml-sycl/outprod.cpp b/ggml/src/ggml-sycl/outprod.cpp index 8e8347ff4..b60415784 100644 --- a/ggml/src/ggml-sycl/outprod.cpp +++ b/ggml/src/ggml-sycl/outprod.cpp @@ -1,8 +1,5 @@ -#include -#include #include "outprod.hpp" - void ggml_sycl_op_out_prod(ggml_backend_sycl_context& ctx, ggml_tensor* dst) { const ggml_tensor *src0 = dst->src[0]; const ggml_tensor *src1 = dst->src[1]; @@ -34,20 +31,13 @@ void ggml_sycl_op_out_prod(ggml_backend_sycl_context& ctx, ggml_tensor* dst) { // Handle transposition of src1 const bool src1_T = ggml_is_transposed(src1); - const oneapi::mkl::transpose src1_op = - src1_T ? oneapi::mkl::transpose::nontrans : oneapi::mkl::transpose::trans; + const oneapi::math::transpose src1_op = src1_T ? oneapi::math::transpose::nontrans : oneapi::math::transpose::trans; const int64_t ldb = (src1_T ? nb10 : nb11) / sizeof(float); try { - // Perform matrix multiplication using oneMKL GEMM -#ifdef GGML_SYCL_NVIDIA - oneapi::mkl::blas::column_major::gemm(oneapi::mkl::backend_selector{ *stream }, - oneapi::mkl::transpose::nontrans, src1_op, ne0, ne1, ne01, alpha, src0_d, - ne00, src1_d, ldb, beta, dst_d, ne0); -#else - oneapi::mkl::blas::column_major::gemm(*stream, oneapi::mkl::transpose::nontrans, src1_op, ne0, ne1, ne01, alpha, - src0_d, ne00, src1_d, ldb, beta, dst_d, ne0); -#endif + // Perform matrix multiplication using oneMath GEMM + oneapi::math::blas::column_major::gemm(get_onemath_backend(*stream), oneapi::math::transpose::nontrans, src1_op, + ne0, ne1, ne01, alpha, src0_d, ne00, src1_d, ldb, beta, dst_d, ne0); } catch (sycl::exception const& exc) { std::cerr << exc.what() << std::endl; diff --git a/ggml/src/ggml-sycl/rope.cpp b/ggml/src/ggml-sycl/rope.cpp index 1244b231a..bbcb356e9 100644 --- a/ggml/src/ggml-sycl/rope.cpp +++ b/ggml/src/ggml-sycl/rope.cpp @@ -192,18 +192,15 @@ static void rope_neox_sycl( } } -void ggml_sycl_op_rope( - ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1, ggml_tensor *dst, - const float *src0_dd, const float *src1_dd, float *dst_dd, const queue_ptr &main_stream) { - const ggml_tensor * src2 = dst->src[2]; +void ggml_sycl_op_rope(ggml_backend_sycl_context & ctx, ggml_tensor *dst) { - GGML_ASSERT(src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16); + GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32 || dst->src[0]->type == GGML_TYPE_F16); GGML_ASSERT( dst->type == GGML_TYPE_F32 || dst->type == GGML_TYPE_F16); - GGML_ASSERT(src0->type == dst->type); + GGML_ASSERT(dst->src[0]->type == dst->type); - const int64_t ne00 = src0->ne[0]; - const int64_t ne01 = src0->ne[1]; - const int64_t nr = ggml_nrows(src0); + const int64_t ne00 = dst->src[0]->ne[0]; + const int64_t ne01 = dst->src[0]->ne[1]; + const int64_t nr = ggml_nrows(dst->src[0]); //const int n_past = ((int32_t *) dst->op_params)[0]; const int n_dims = ((int32_t *) dst->op_params)[1]; @@ -228,49 +225,47 @@ void ggml_sycl_op_rope( const bool is_neox = mode & GGML_ROPE_TYPE_NEOX; - const int32_t * pos = (const int32_t *) src1_dd; + const int32_t * pos = (const int32_t *) dst->src[1]->data; const float * freq_factors = nullptr; - if (src2 != nullptr) { - freq_factors = (const float *) src2->data; + if (dst->src[2] != nullptr) { + freq_factors = (const float *) dst->src[2]->data; } rope_corr_dims corr_dims; ggml_rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow, corr_dims.v); + dpct::queue_ptr main_stream = ctx.stream(); + SYCL_CHECK(ggml_sycl_set_device(ctx.device)); + // compute if (is_neox) { - if (src0->type == GGML_TYPE_F32) { + if (dst->src[0]->type == GGML_TYPE_F32) { rope_neox_sycl( - (const float *)src0_dd, (float *)dst_dd, ne00, n_dims, nr, pos, freq_scale, ne01, freq_base, ext_factor, + (const float *)dst->src[0]->data, (float *)dst->data, ne00, n_dims, nr, pos, freq_scale, ne01, freq_base, ext_factor, attn_factor, corr_dims, freq_factors, main_stream ); - } else if (src0->type == GGML_TYPE_F16) { + } else if (dst->src[0]->type == GGML_TYPE_F16) { rope_neox_sycl( - (const sycl::half *)src0_dd, (sycl::half *)dst_dd, ne00, n_dims, nr, pos, freq_scale, ne01, freq_base, ext_factor, + (const sycl::half *)dst->src[0]->data, (sycl::half *)dst->data, ne00, n_dims, nr, pos, freq_scale, ne01, freq_base, ext_factor, attn_factor, corr_dims, freq_factors, main_stream ); } else { GGML_ABORT("fatal error"); } } else { - if (src0->type == GGML_TYPE_F32) { + if (dst->src[0]->type == GGML_TYPE_F32) { rope_norm_sycl( - (const float *)src0_dd, (float *)dst_dd, ne00, n_dims, nr, pos, freq_scale, ne01, freq_base, ext_factor, + (const float *)dst->src[0]->data, (float *)dst->data, ne00, n_dims, nr, pos, freq_scale, ne01, freq_base, ext_factor, attn_factor, corr_dims, freq_factors, main_stream ); - } else if (src0->type == GGML_TYPE_F16) { + } else if (dst->src[0]->type == GGML_TYPE_F16) { rope_norm_sycl( - (const sycl::half *)src0_dd, (sycl::half *)dst_dd, ne00, n_dims, nr, pos, freq_scale, ne01, freq_base, ext_factor, + (const sycl::half *)dst->src[0]->data, (sycl::half *)dst->data, ne00, n_dims, nr, pos, freq_scale, ne01, freq_base, ext_factor, attn_factor, corr_dims, freq_factors, main_stream ); } else { GGML_ABORT("fatal error"); } } - - GGML_UNUSED(src1); - GGML_UNUSED(dst); - GGML_UNUSED(src1_dd); - GGML_UNUSED(ctx); } diff --git a/ggml/src/ggml-sycl/rope.hpp b/ggml/src/ggml-sycl/rope.hpp index 00354c313..a399bddb8 100644 --- a/ggml/src/ggml-sycl/rope.hpp +++ b/ggml/src/ggml-sycl/rope.hpp @@ -15,8 +15,6 @@ #include "common.hpp" -void ggml_sycl_op_rope( - ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1, ggml_tensor *dst, - const float *src0_dd, const float *src1_dd, float *dst_dd, const queue_ptr &main_stream); +void ggml_sycl_op_rope(ggml_backend_sycl_context & ctx, ggml_tensor *dst); #endif // GGML_SYCL_ROPE_HPP diff --git a/ggml/src/ggml-sycl/softmax.cpp b/ggml/src/ggml-sycl/softmax.cpp index 563e0655f..7563d9ced 100644 --- a/ggml/src/ggml-sycl/softmax.cpp +++ b/ggml/src/ggml-sycl/softmax.cpp @@ -132,7 +132,7 @@ static void soft_max_f32_submitter(const float * x, const T * mask, float * dst, cgh.parallel_for( sycl::nd_range<3>(block_nums * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(WARP_SIZE)]] { + [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] { soft_max_f32(x, mask, dst, ncols_par, nrows_y, scale, max_bias, m0, m1, n_head_log2, item_ct1, @@ -249,13 +249,16 @@ void ggml_sycl_op_soft_max(ggml_backend_sycl_context & ctx, ggml_tensor * dst) { if (dst->src[1] && dst->src[1]->type == GGML_TYPE_F16) { const sycl::half * src1_dd = static_cast(dst->src[1]->data); + GGML_SYCL_DEBUG("%s: F16 mask\n", __func__); soft_max_f32_sycl(src0_dd, src1_dd, dst_dd, ne00, nrows_x, nrows_y, scale, max_bias, main_stream, ctx.device); } else if (dst->src[1] && dst->src[1]->type == GGML_TYPE_F32) { const float * src1_dd = static_cast(dst->src[1]->data); + GGML_SYCL_DEBUG("%s: F32 mask\n", __func__); soft_max_f32_sycl(src0_dd, src1_dd, dst_dd, ne00, nrows_x, nrows_y, scale, max_bias, main_stream, ctx.device); } else { /* mask unavailable */ + GGML_SYCL_DEBUG("%s: No mask\n", __func__); soft_max_f32_sycl(src0_dd, nullptr, dst_dd, ne00, nrows_x, nrows_y, scale, max_bias, main_stream, ctx.device); } } diff --git a/ggml/src/ggml-sycl/sycl_hw.cpp b/ggml/src/ggml-sycl/sycl_hw.cpp new file mode 100644 index 000000000..da121ffc2 --- /dev/null +++ b/ggml/src/ggml-sycl/sycl_hw.cpp @@ -0,0 +1,13 @@ +#include "sycl_hw.hpp" + + +sycl_hw_info get_device_hw_info(sycl::device *device_ptr) { + sycl_hw_info res; + int32_t id = device_ptr->get_info(); + res.device_id = id; + + syclex::architecture arch = device_ptr->get_info(); + res.arch = arch; + + return res; +} diff --git a/ggml/src/ggml-sycl/sycl_hw.hpp b/ggml/src/ggml-sycl/sycl_hw.hpp new file mode 100644 index 000000000..bf689450c --- /dev/null +++ b/ggml/src/ggml-sycl/sycl_hw.hpp @@ -0,0 +1,23 @@ +#ifndef SYCL_HW_HPP +#define SYCL_HW_HPP + +#include +#include +#include +#include + +#include + +namespace syclex = sycl::ext::oneapi::experimental; + +struct sycl_hw_info { + syclex::architecture arch; + int32_t device_id; +}; + +bool is_in_vector(std::vector &vec, int item); + +sycl_hw_info get_device_hw_info(sycl::device *device_ptr); + + +#endif // SYCL_HW_HPP diff --git a/ggml/src/ggml-sycl/wkv.cpp b/ggml/src/ggml-sycl/wkv.cpp new file mode 100644 index 000000000..540f6fbf5 --- /dev/null +++ b/ggml/src/ggml-sycl/wkv.cpp @@ -0,0 +1,305 @@ +#include +#include "wkv.hpp" + +constexpr int WKV_BLOCK_SIZE = 64; // Matching CUDA_WKV_BLOCK_SIZE + +// Helper function for the main kernel +template +static void rwkv_wkv6_f32_kernel( + const int B, const int T, const int C, const int H, + const float* k, const float* v, const float* r, + const float* tf, const float* td, const float* s, + float* dst, const sycl::nd_item<3>& item_ct1, float* shared_mem) { + + const int tid = item_ct1.get_local_id(2); + const int bid = item_ct1.get_group(2); + + const int head_size = block_size; + const int batch_i = bid / H; + const int head_i = bid % H; + const int state_size = C * head_size; + const int n_seq_tokens = T / B; + + // Set up shared memory pointers + float* _k = shared_mem; + float* _r = _k + head_size; + float* _tf = _r + head_size; + float* _td = _tf + head_size; + + // Local state array + float state[block_size]; + + // Load initial state + #pragma unroll + for (int i = 0; i < head_size; i++) { + state[i] = s[batch_i * state_size + head_i * head_size * head_size + i * head_size + tid]; + } + + // Sync threads before shared memory operations + item_ct1.barrier(sycl::access::fence_space::local_space); + + // Load time-mixing parameters + _tf[tid] = tf[head_i * head_size + tid]; + item_ct1.barrier(sycl::access::fence_space::local_space); + + // Main sequence processing loop + for (int t = batch_i * n_seq_tokens * C + head_i * head_size + tid; + t < (batch_i + 1) * n_seq_tokens * C + head_i * head_size + tid; + t += C) { + + item_ct1.barrier(sycl::access::fence_space::local_space); + + // Load current timestep data to shared memory + _k[tid] = k[t]; + _r[tid] = r[t]; + _td[tid] = td[t]; + + item_ct1.barrier(sycl::access::fence_space::local_space); + + const float _v = v[t]; + float y = 0; + + // Process in chunks of 4 for better vectorization + sycl::float4 k4, r4, tf4, td4, s4; + #pragma unroll + for (int j = 0; j < head_size; j += 4) { + // Load data in vec4 chunks + k4 = sycl::float4(_k[j], _k[j+1], _k[j+2], _k[j+3]); + r4 = sycl::float4(_r[j], _r[j+1], _r[j+2], _r[j+3]); + tf4 = sycl::float4(_tf[j], _tf[j+1], _tf[j+2], _tf[j+3]); + td4 = sycl::float4(_td[j], _td[j+1], _td[j+2], _td[j+3]); + s4 = sycl::float4(state[j], state[j+1], state[j+2], state[j+3]); + + // Compute key-value product + sycl::float4 kv4 = k4 * _v; + + // Accumulate weighted sum + y += sycl::dot(r4, tf4 * kv4 + s4); + + // Update state + s4 = s4 * td4 + kv4; + + // Store updated state + state[j] = s4.x(); + state[j+1] = s4.y(); + state[j+2] = s4.z(); + state[j+3] = s4.w(); + } + + dst[t] = y; + } + + // Save final state + #pragma unroll + for (int i = 0; i < head_size; i++) { + dst[T * C + batch_i * state_size + head_i * head_size * head_size + i * head_size + tid] = state[i]; + } +} + +template +static void rwkv_wkv7_f32_kernel( + const int B, const int T, const int C, const int H, + const float* r, const float* w, const float* k, const float* v, + const float* a, const float* b, const float* s, + float* dst, const sycl::nd_item<3>& item_ct1, float* shared_mem) { + + const int tid = item_ct1.get_local_id(2); + const int bid = item_ct1.get_group(2); + + const int head_size = block_size; + const int batch_i = bid / H; + const int head_i = bid % H; + const int state_size = C * head_size; + const int n_seq_tokens = T / B; + + float* _r = shared_mem; + float* _w = _r + head_size; + float* _k = _w + head_size; + float* _a = _k + head_size; + float* _b = _a + head_size; + + float state[block_size]; + + #pragma unroll + for (int i = 0; i < head_size; i++) { + state[i] = s[batch_i * state_size + head_i * head_size * head_size + tid * head_size + i]; + } + + for (int t = batch_i * n_seq_tokens * C + head_i * head_size + tid; + t < (batch_i + 1) * n_seq_tokens * C + head_i * head_size + tid; + t += C) { + + item_ct1.barrier(sycl::access::fence_space::local_space); + + _r[tid] = r[t]; + _w[tid] = w[t]; + _k[tid] = k[t]; + _a[tid] = a[t]; + _b[tid] = b[t]; + + item_ct1.barrier(sycl::access::fence_space::local_space); + + const float _v = v[t]; + float y = 0, sa = 0; + sycl::float4 a4, s4; + + #pragma unroll + for (int j = 0; j < head_size; j += 4) { + a4 = sycl::float4(_a[j], _a[j+1], _a[j+2], _a[j+3]); + s4 = sycl::float4(state[j], state[j+1], state[j+2], state[j+3]); + sa += sycl::dot(a4, s4); + } + + sycl::float4 r4, w4, k4, b4; + #pragma unroll + for (int j = 0; j < head_size; j += 4) { + r4 = sycl::float4(_r[j], _r[j+1], _r[j+2], _r[j+3]); + w4 = sycl::float4(_w[j], _w[j+1], _w[j+2], _w[j+3]); + k4 = sycl::float4(_k[j], _k[j+1], _k[j+2], _k[j+3]); + b4 = sycl::float4(_b[j], _b[j+1], _b[j+2], _b[j+3]); + s4 = sycl::float4(state[j], state[j+1], state[j+2], state[j+3]); + + sycl::float4 kv4 = k4 * _v; + + s4 = s4 * w4 + kv4 + sa * b4; + y += sycl::dot(r4, s4); + + state[j] = s4.x(); + state[j+1] = s4.y(); + state[j+2] = s4.z(); + state[j+3] = s4.w(); + } + + dst[t] = y; + } + + #pragma unroll + for (int i = 0; i < head_size; i++) { + dst[T * C + batch_i * state_size + head_i * head_size * head_size + tid * head_size + i] = state[i]; + } +} + +void ggml_sycl_op_rwkv_wkv6(ggml_backend_sycl_context& ctx, ggml_tensor* dst) { + + const ggml_tensor *src0 = dst->src[0]; + const ggml_tensor *src1 = dst->src[1]; + + const float* k_d = (const float*)dst->src[0]->data; + const float* v_d = (const float*)dst->src[1]->data; + const float* r_d = (const float*)dst->src[2]->data; + const float* tf_d = (const float*)dst->src[3]->data; + const float* td_d = (const float*)dst->src[4]->data; + const float* s_d = (const float*)dst->src[5]->data; + float* dst_d = (float*)dst->data; + + const int64_t B = dst->src[5]->ne[1]; + const int64_t T = dst->src[0]->ne[2]; + const int64_t C = dst->ne[0]; + const int64_t H = dst->src[0]->ne[1]; + + GGML_ASSERT(dst->src[5]->type == GGML_TYPE_F32); + GGML_ASSERT(C % H == 0); + GGML_ASSERT(C / H == WKV_BLOCK_SIZE || C / H == WKV_BLOCK_SIZE * 2); // The current sycl kernel is designed for RWKV6, HEAD_SIZE == 64 + + dpct::queue_ptr stream = ctx.stream(); + + // Calculate execution configuration + const size_t shared_mem_size = C / H * 4 * sizeof(float); // For k, r, tf, td + sycl::range<3> block_dims(1, 1, C / H); + sycl::range<3> grid_dims(1, 1, B * H); + + // Submit kernel + if (C / H == WKV_BLOCK_SIZE) { + stream->submit([&](sycl::handler& cgh) { + sycl::local_accessor shared_mem_acc(shared_mem_size, cgh); + + cgh.parallel_for( + sycl::nd_range<3>(grid_dims * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) { + rwkv_wkv6_f32_kernel( + B, T, C, H, k_d, v_d, r_d, tf_d, td_d, s_d, dst_d, + item_ct1, (float*)shared_mem_acc.get_multi_ptr().get() + ); + }); + }); + } else { + stream->submit([&](sycl::handler& cgh) { + sycl::local_accessor shared_mem_acc(shared_mem_size, cgh); + + cgh.parallel_for( + sycl::nd_range<3>(grid_dims * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) { + rwkv_wkv6_f32_kernel( + B, T, C, H, k_d, v_d, r_d, tf_d, td_d, s_d, dst_d, + item_ct1, (float*)shared_mem_acc.get_multi_ptr().get() + ); + }); + }); + } + + GGML_UNUSED(src0); + GGML_UNUSED(src1); +} + +void ggml_sycl_op_rwkv_wkv7(ggml_backend_sycl_context& ctx, ggml_tensor* dst) { + + const ggml_tensor *src0 = dst->src[0]; + const ggml_tensor *src1 = dst->src[1]; + + const float* r_d = (const float*)dst->src[0]->data; + const float* w_d = (const float*)dst->src[1]->data; + const float* k_d = (const float*)dst->src[2]->data; + const float* v_d = (const float*)dst->src[3]->data; + const float* a_d = (const float*)dst->src[4]->data; + const float* b_d = (const float*)dst->src[5]->data; + const float* s_d = (const float*)dst->src[6]->data; + float* dst_d = (float*)dst->data; + + const int64_t B = dst->src[6]->ne[1]; + const int64_t T = dst->src[0]->ne[2]; + const int64_t C = dst->ne[0]; + const int64_t H = dst->src[0]->ne[1]; + + GGML_ASSERT(dst->src[6]->type == GGML_TYPE_F32); + GGML_ASSERT(C % H == 0); + GGML_ASSERT(C / H == WKV_BLOCK_SIZE || C / H == WKV_BLOCK_SIZE * 2); + + dpct::queue_ptr stream = ctx.stream(); + + // Calculate execution configuration + const size_t shared_mem_size = C / H * 5 * sizeof(float); // For r, w, k, a, b + sycl::range<3> block_dims(1, 1, C / H); + sycl::range<3> grid_dims(1, 1, B * H); + + // Submit kernel + if (C / H == WKV_BLOCK_SIZE) { + stream->submit([&](sycl::handler& cgh) { + sycl::local_accessor shared_mem_acc(shared_mem_size, cgh); + + cgh.parallel_for( + sycl::nd_range<3>(grid_dims * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) { + rwkv_wkv7_f32_kernel( + B, T, C, H, r_d, w_d, k_d, v_d, a_d, b_d, s_d, dst_d, + item_ct1, (float*)shared_mem_acc.get_multi_ptr().get() + ); + }); + }); + } else { + stream->submit([&](sycl::handler& cgh) { + sycl::local_accessor shared_mem_acc(shared_mem_size, cgh); + + cgh.parallel_for( + sycl::nd_range<3>(grid_dims * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) { + rwkv_wkv7_f32_kernel( + B, T, C, H, r_d, w_d, k_d, v_d, a_d, b_d, s_d, dst_d, + item_ct1, (float*)shared_mem_acc.get_multi_ptr().get() + ); + }); + }); + } + + GGML_UNUSED(src0); + GGML_UNUSED(src1); +} diff --git a/ggml/src/ggml-sycl/wkv.hpp b/ggml/src/ggml-sycl/wkv.hpp new file mode 100644 index 000000000..9f34a1001 --- /dev/null +++ b/ggml/src/ggml-sycl/wkv.hpp @@ -0,0 +1,10 @@ +#ifndef GGML_SYCL_WKV_HPP +#define GGML_SYCL_WKV_HPP + +#include "common.hpp" + +void ggml_sycl_op_rwkv_wkv6(ggml_backend_sycl_context & ctx, ggml_tensor * dst); + +void ggml_sycl_op_rwkv_wkv7(ggml_backend_sycl_context & ctx, ggml_tensor * dst); + +#endif // GGML_SYCL_WKV_HPP diff --git a/ggml/src/ggml-sycl/wkv6.cpp b/ggml/src/ggml-sycl/wkv6.cpp deleted file mode 100644 index b54c20964..000000000 --- a/ggml/src/ggml-sycl/wkv6.cpp +++ /dev/null @@ -1,143 +0,0 @@ -#include -#include "wkv6.hpp" - -constexpr int WKV_BLOCK_SIZE = 64; // Matching CUDA_WKV_BLOCK_SIZE - -// Helper function for the main kernel -static void rwkv_wkv_f32_kernel( - const int B, const int T, const int C, const int H, - const float* k, const float* v, const float* r, - const float* tf, const float* td, const float* s, - float* dst, const sycl::nd_item<3>& item_ct1, float* shared_mem) { - - const int tid = item_ct1.get_local_id(2); - const int bid = item_ct1.get_group(2); - - const int head_size = WKV_BLOCK_SIZE; - const int batch_i = bid / H; - const int head_i = bid % H; - const int state_size = C * head_size; - const int n_seq_tokens = T / B; - - // Set up shared memory pointers - float* _k = shared_mem; - float* _r = _k + head_size; - float* _tf = _r + head_size; - float* _td = _tf + head_size; - - // Local state array - float state[WKV_BLOCK_SIZE]; - - // Load initial state - #pragma unroll - for (int i = 0; i < head_size; i++) { - state[i] = s[batch_i * state_size + head_i * head_size * head_size + i * head_size + tid]; - } - - // Sync threads before shared memory operations - item_ct1.barrier(sycl::access::fence_space::local_space); - - // Load time-mixing parameters - _tf[tid] = tf[head_i * head_size + tid]; - item_ct1.barrier(sycl::access::fence_space::local_space); - - // Main sequence processing loop - for (int t = batch_i * n_seq_tokens * C + head_i * head_size + tid; - t < (batch_i + 1) * n_seq_tokens * C + head_i * head_size + tid; - t += C) { - - item_ct1.barrier(sycl::access::fence_space::local_space); - - // Load current timestep data to shared memory - _k[tid] = k[t]; - _r[tid] = r[t]; - _td[tid] = td[t]; - - item_ct1.barrier(sycl::access::fence_space::local_space); - - const float _v = v[t]; - float y = 0; - - // Process in chunks of 4 for better vectorization - sycl::float4 k4, r4, tf4, td4, s4; - #pragma unroll - for (int j = 0; j < head_size; j += 4) { - // Load data in vec4 chunks - k4 = sycl::float4(_k[j], _k[j+1], _k[j+2], _k[j+3]); - r4 = sycl::float4(_r[j], _r[j+1], _r[j+2], _r[j+3]); - tf4 = sycl::float4(_tf[j], _tf[j+1], _tf[j+2], _tf[j+3]); - td4 = sycl::float4(_td[j], _td[j+1], _td[j+2], _td[j+3]); - s4 = sycl::float4(state[j], state[j+1], state[j+2], state[j+3]); - - // Compute key-value product - sycl::float4 kv4 = k4 * _v; - - // Accumulate weighted sum - y += sycl::dot(r4, tf4 * kv4 + s4); - - // Update state - s4 = s4 * td4 + kv4; - - // Store updated state - state[j] = s4.x(); - state[j+1] = s4.y(); - state[j+2] = s4.z(); - state[j+3] = s4.w(); - } - - dst[t] = y; - } - - // Save final state - #pragma unroll - for (int i = 0; i < head_size; i++) { - dst[T * C + batch_i * state_size + head_i * head_size * head_size + i * head_size + tid] = state[i]; - } -} - -void ggml_sycl_op_rwkv_wkv6(ggml_backend_sycl_context& ctx, ggml_tensor* dst) { - - const ggml_tensor *src0 = dst->src[0]; - const ggml_tensor *src1 = dst->src[1]; - - const float* k_d = (const float*)dst->src[0]->data; - const float* v_d = (const float*)dst->src[1]->data; - const float* r_d = (const float*)dst->src[2]->data; - const float* tf_d = (const float*)dst->src[3]->data; - const float* td_d = (const float*)dst->src[4]->data; - const float* s_d = (const float*)dst->src[5]->data; - float* dst_d = (float*)dst->data; - - const int64_t B = dst->src[5]->ne[1]; - const int64_t T = dst->src[0]->ne[2]; - const int64_t C = dst->ne[0]; - const int64_t H = dst->src[0]->ne[1]; - - GGML_ASSERT(dst->src[5]->type == GGML_TYPE_F32); - GGML_ASSERT(C % H == 0); - GGML_ASSERT(C / H == WKV_BLOCK_SIZE); // The current sycl kernel is designed for RWKV6, HEAD_SIZE == 64 - - dpct::queue_ptr stream = ctx.stream(); - - // Calculate execution configuration - const size_t shared_mem_size = WKV_BLOCK_SIZE * 4 * sizeof(float); // For k, r, tf, td - sycl::range<3> block_dims(1, 1, C / H); - sycl::range<3> grid_dims(1, 1, B * H); - - // Submit kernel - stream->submit([&](sycl::handler& cgh) { - sycl::local_accessor shared_mem_acc(shared_mem_size, cgh); - - cgh.parallel_for( - sycl::nd_range<3>(grid_dims * block_dims, block_dims), - [=](sycl::nd_item<3> item_ct1) { - rwkv_wkv_f32_kernel( - B, T, C, H, k_d, v_d, r_d, tf_d, td_d, s_d, dst_d, - item_ct1, (float*)shared_mem_acc.get_multi_ptr().get() - ); - }); - }); - - GGML_UNUSED(src0); - GGML_UNUSED(src1); -} diff --git a/ggml/src/ggml-sycl/wkv6.hpp b/ggml/src/ggml-sycl/wkv6.hpp deleted file mode 100644 index 8c596a997..000000000 --- a/ggml/src/ggml-sycl/wkv6.hpp +++ /dev/null @@ -1,9 +0,0 @@ -#ifndef GGML_SYCL_WKV6_HPP -#define GGML_SYCL_WKV6_HPP - -#include "common.hpp" - -void ggml_sycl_op_rwkv_wkv6(ggml_backend_sycl_context & ctx, ggml_tensor * dst); - - -#endif // GGML_SYCL_WKV6_HPP diff --git a/ggml/src/ggml-vulkan/CMakeLists.txt b/ggml/src/ggml-vulkan/CMakeLists.txt index d970f7e20..9d028f718 100644 --- a/ggml/src/ggml-vulkan/CMakeLists.txt +++ b/ggml/src/ggml-vulkan/CMakeLists.txt @@ -32,8 +32,10 @@ if (Vulkan_FOUND) if (${glslc_error} MATCHES ".*extension not supported: GL_KHR_cooperative_matrix.*") message(STATUS "GL_KHR_cooperative_matrix not supported by glslc") + set(GGML_VULKAN_COOPMAT_GLSLC_SUPPORT OFF) else() message(STATUS "GL_KHR_cooperative_matrix supported by glslc") + set(GGML_VULKAN_COOPMAT_GLSLC_SUPPORT ON) add_compile_definitions(GGML_VULKAN_COOPMAT_GLSLC_SUPPORT) endif() @@ -46,11 +48,29 @@ if (Vulkan_FOUND) if (${glslc_error} MATCHES ".*extension not supported: GL_NV_cooperative_matrix2.*") message(STATUS "GL_NV_cooperative_matrix2 not supported by glslc") + set(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT OFF) else() message(STATUS "GL_NV_cooperative_matrix2 supported by glslc") + set(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT ON) add_compile_definitions(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT) endif() + # Compile a test shader to determine whether GL_EXT_integer_dot_product is supported. + # If it's not, there will be an error to stderr. + # If it's supported, set a define to indicate that we should compile those shaders + execute_process(COMMAND ${Vulkan_GLSLC_EXECUTABLE} -o - -fshader-stage=compute --target-env=vulkan1.3 "${CMAKE_CURRENT_SOURCE_DIR}/vulkan-shaders/test_integer_dot_support.comp" + OUTPUT_VARIABLE glslc_output + ERROR_VARIABLE glslc_error) + + if (${glslc_error} MATCHES ".*extension not supported: GL_EXT_integer_dot_product.*") + message(STATUS "GL_EXT_integer_dot_product not supported by glslc") + set(GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT OFF) + else() + message(STATUS "GL_EXT_integer_dot_product supported by glslc") + set(GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT ON) + add_compile_definitions(GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT) + endif() + target_link_libraries(ggml-vulkan PRIVATE Vulkan::Vulkan) target_include_directories(ggml-vulkan PRIVATE ${CMAKE_CURRENT_BINARY_DIR}) @@ -119,6 +139,9 @@ if (Vulkan_FOUND) SOURCE_DIR ${CMAKE_CURRENT_SOURCE_DIR}/vulkan-shaders CMAKE_ARGS -DCMAKE_TOOLCHAIN_FILE=${HOST_CMAKE_TOOLCHAIN_FILE} -DCMAKE_INSTALL_PREFIX=${CMAKE_BINARY_DIR} + -DGGML_VULKAN_COOPMAT_GLSLC_SUPPORT=${GGML_VULKAN_COOPMAT_GLSLC_SUPPORT} + -DGGML_VULKAN_COOPMAT2_GLSLC_SUPPORT=${GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT} + -DGGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT=${GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT} BUILD_COMMAND ${CMAKE_COMMAND} --build . INSTALL_COMMAND ${CMAKE_COMMAND} --install . INSTALL_DIR ${CMAKE_BINARY_DIR} diff --git a/ggml/src/ggml-vulkan/cmake/host-toolchain.cmake.in b/ggml/src/ggml-vulkan/cmake/host-toolchain.cmake.in index b6af747a5..2d8a85696 100644 --- a/ggml/src/ggml-vulkan/cmake/host-toolchain.cmake.in +++ b/ggml/src/ggml-vulkan/cmake/host-toolchain.cmake.in @@ -4,8 +4,8 @@ set(CMAKE_CXX_FLAGS -O2) set(CMAKE_FIND_ROOT_PATH_MODE_PROGRAM NEVER) set(CMAKE_FIND_ROOT_PATH_MODE_LIBRARY NEVER) set(CMAKE_FIND_ROOT_PATH_MODE_INCLUDE NEVER) -set(CMAKE_C_COMPILER @HOST_C_COMPILER@) -set(CMAKE_CXX_COMPILER @HOST_CXX_COMPILER@) +set(CMAKE_C_COMPILER "@HOST_C_COMPILER@") +set(CMAKE_CXX_COMPILER "@HOST_CXX_COMPILER@") set(CMAKE_RUNTIME_OUTPUT_DIRECTORY @CMAKE_RUNTIME_OUTPUT_DIRECTORY@) if("@CMAKE_C_COMPILER_ID@" STREQUAL "MSVC") diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp index d32ba4efb..783a0ff86 100644 --- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp +++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp @@ -24,12 +24,36 @@ #include #include +#if defined(_MSC_VER) +# define NOMINMAX 1 +# include +# define YIELD() YieldProcessor() +#elif defined(__clang__) || defined(__GNUC__) +# if defined(__x86_64__) ||defined(__i386__) +# include +# define YIELD() _mm_pause() +# elif defined(__arm__) || defined(__aarch64__) +# if defined(__clang__) +# include +# define YIELD() __yield() +# else +# define YIELD() asm volatile("yield") +# endif +# endif +#endif + +#if !defined(YIELD) +#define YIELD() +#endif + #include "ggml-impl.h" #include "ggml-backend-impl.h" #include "ggml-vulkan-shaders.hpp" +#define ROUNDUP_POW2(M, N) (((M) + (N) - 1) & ~((N) - 1)) #define CEIL_DIV(M, N) (((M) + (N)-1) / (N)) +static bool is_pow2(uint32_t x) { return x > 1 && (x & (x-1)) == 0; } #define VK_VENDOR_ID_AMD 0x1002 #define VK_VENDOR_ID_APPLE 0x106b @@ -148,6 +172,67 @@ class vk_perf_logger; static void ggml_vk_destroy_buffer(vk_buffer& buf); static constexpr uint32_t mul_mat_vec_max_cols = 8; +static constexpr uint32_t p021_max_gqa_ratio = 8; + +enum vk_device_architecture { + OTHER, + AMD_GCN, + AMD_RDNA1, + AMD_RDNA2, + AMD_RDNA3, +}; + +static vk_device_architecture get_device_architecture(const vk::PhysicalDevice& device) { + vk::PhysicalDeviceProperties props = device.getProperties(); + + if (props.vendorID == VK_VENDOR_ID_AMD) { + const std::vector ext_props = device.enumerateDeviceExtensionProperties(); + + bool amd_shader_core_properties = false; + bool integer_dot_product = false; + bool subgroup_size_control = false; + + for (const auto& properties : ext_props) { + if (strcmp("VK_AMD_shader_core_properties", properties.extensionName) == 0) { + amd_shader_core_properties = true; + } else if (strcmp("VK_KHR_shader_integer_dot_product", properties.extensionName) == 0) { + integer_dot_product = true; + } else if (strcmp("VK_EXT_subgroup_size_control", properties.extensionName) == 0) { + subgroup_size_control = true; + } + } + + if (!amd_shader_core_properties || !integer_dot_product || !subgroup_size_control) { + return vk_device_architecture::OTHER; + } + + vk::PhysicalDeviceProperties2 props2; + vk::PhysicalDeviceShaderCorePropertiesAMD shader_core_props_amd; + vk::PhysicalDeviceShaderIntegerDotProductPropertiesKHR integer_dot_props; + vk::PhysicalDeviceSubgroupSizeControlPropertiesEXT subgroup_size_control_props; + + props2.pNext = &shader_core_props_amd; + shader_core_props_amd.pNext = &integer_dot_props; + integer_dot_props.pNext = &subgroup_size_control_props; + + device.getProperties2(&props2); + + if (subgroup_size_control_props.maxSubgroupSize == 64 && subgroup_size_control_props.minSubgroupSize == 64) { + return vk_device_architecture::AMD_GCN; + } + if (subgroup_size_control_props.maxSubgroupSize == 64 && subgroup_size_control_props.minSubgroupSize == 32) { + // RDNA + if (shader_core_props_amd.wavefrontsPerSimd == 20) { + return vk_device_architecture::AMD_RDNA1; + } + if (integer_dot_props.integerDotProduct4x8BitPackedMixedSignednessAccelerated) { + return vk_device_architecture::AMD_RDNA3; + } + return vk_device_architecture::AMD_RDNA2; + } + } + return vk_device_architecture::OTHER; +} struct vk_device_struct { std::mutex mutex; @@ -161,13 +246,18 @@ struct vk_device_struct { bool pipeline_robustness; vk::Device device; uint32_t vendor_id; + vk_device_architecture architecture; vk_queue compute_queue; vk_queue transfer_queue; bool single_queue; uint32_t subgroup_size; uint32_t shader_core_count; bool uma; + bool prefer_host_memory; bool float_controls_rte_fp16; + bool subgroup_add; + + bool integer_dot_product; bool subgroup_size_control; uint32_t subgroup_min_size; @@ -180,6 +270,12 @@ struct vk_device_struct { uint32_t coopmat_m; uint32_t coopmat_n; uint32_t coopmat_k; + + bool coopmat_int_support; + uint32_t coopmat_int_m; + uint32_t coopmat_int_n; + uint32_t coopmat_int_k; + bool coopmat2; size_t idx; @@ -198,10 +294,10 @@ struct vk_device_struct { vk_matmul_pipeline pipeline_matmul_f32_f16 {}; vk_matmul_pipeline2 pipeline_matmul_f16; vk_matmul_pipeline2 pipeline_matmul_f16_f32; - vk_pipeline pipeline_matmul_split_k_reduce; - vk_matmul_pipeline2 pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_COUNT]; vk_matmul_pipeline2 pipeline_dequant_mul_mat_mat[GGML_TYPE_COUNT]; + vk_matmul_pipeline2 pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_COUNT]; + vk_matmul_pipeline2 pipeline_dequant_mul_mat_mat_q8_1[GGML_TYPE_COUNT]; vk_matmul_pipeline pipeline_matmul_id_f32 {}; vk_matmul_pipeline2 pipeline_matmul_id_f16; @@ -209,18 +305,22 @@ struct vk_device_struct { vk_matmul_pipeline2 pipeline_dequant_mul_mat_mat_id[GGML_TYPE_COUNT]; + vk_pipeline pipeline_matmul_split_k_reduce; + vk_pipeline pipeline_quantize_q8_1; + vk_pipeline pipeline_dequant[GGML_TYPE_COUNT]; vk_pipeline pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_COUNT][mul_mat_vec_max_cols]; vk_pipeline pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_COUNT][mul_mat_vec_max_cols]; vk_pipeline pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_COUNT]; - vk_pipeline pipeline_mul_mat_vec_p021_f16_f32; + vk_pipeline pipeline_mul_mat_vec_p021_f16_f32[p021_max_gqa_ratio]; vk_pipeline pipeline_mul_mat_vec_nc_f16_f32; vk_pipeline pipeline_get_rows[GGML_TYPE_COUNT]; vk_pipeline pipeline_get_rows_f32[GGML_TYPE_COUNT]; vk_pipeline pipeline_acc_f32; vk_pipeline pipeline_add_f32, pipeline_add_f32_norepeat; vk_pipeline pipeline_add_f16_f32_f16, pipeline_add_f16_f32_f16_norepeat; + vk_pipeline pipeline_sub_f32, pipeline_sub_f32_norepeat; vk_pipeline pipeline_mul_f32, pipeline_mul_f32_norepeat; vk_pipeline pipeline_div_f32, pipeline_div_f32_norepeat; vk_pipeline pipeline_concat_f32, pipeline_concat_f16, pipeline_concat_i32; @@ -231,7 +331,7 @@ struct vk_device_struct { vk_pipeline pipeline_cos_f32; vk_pipeline pipeline_clamp_f32; vk_pipeline pipeline_pad_f32; - vk_pipeline pipeline_repeat_f32; + vk_pipeline pipeline_repeat_f32, pipeline_repeat_back_f32; vk_pipeline pipeline_cpy_f32_f32, pipeline_cpy_f32_f16, pipeline_cpy_f16_f16; vk_pipeline pipeline_contig_cpy_f32_f32, pipeline_contig_cpy_f32_f16, pipeline_contig_cpy_f16_f16; vk_pipeline pipeline_cpy_f32_quant[GGML_TYPE_COUNT]; @@ -239,23 +339,34 @@ struct vk_device_struct { vk_pipeline pipeline_norm_f32; vk_pipeline pipeline_group_norm_f32; vk_pipeline pipeline_rms_norm_f32; + vk_pipeline pipeline_rms_norm_back_f32; + vk_pipeline pipeline_l2_norm_f32; vk_pipeline pipeline_gelu_f32; vk_pipeline pipeline_gelu_quick_f32; vk_pipeline pipeline_silu_f32; + vk_pipeline pipeline_silu_back_f32; vk_pipeline pipeline_relu_f32; vk_pipeline pipeline_leaky_relu_f32; vk_pipeline pipeline_tanh_f32; + vk_pipeline pipeline_sigmoid_f32; vk_pipeline pipeline_diag_mask_inf_f32; vk_pipeline pipeline_soft_max_f32, pipeline_soft_max_f32_f16; vk_pipeline pipeline_soft_max_f32_wg512, pipeline_soft_max_f32_f16_wg512; + vk_pipeline pipeline_soft_max_back_f32; vk_pipeline pipeline_rope_norm_f32, pipeline_rope_norm_f16; vk_pipeline pipeline_rope_neox_f32, pipeline_rope_neox_f16; + vk_pipeline pipeline_rope_multi_f32, pipeline_rope_multi_f16; + vk_pipeline pipeline_rope_vision_f32, pipeline_rope_vision_f16; vk_pipeline pipeline_argsort_f32; vk_pipeline pipeline_sum_rows_f32; + vk_pipeline pipeline_argmax_f32; + vk_pipeline pipeline_count_equal_i32; vk_pipeline pipeline_im2col_f32, pipeline_im2col_f32_f16; vk_pipeline pipeline_timestep_embedding_f32; vk_pipeline pipeline_pool2d_f32; vk_pipeline pipeline_rwkv_wkv6_f32; + vk_pipeline pipeline_rwkv_wkv7_f32; + vk_pipeline pipeline_opt_step_adamw_f32; // [2][2][2] is for {f16acc,f32acc}x{large,small_rows}x{unaligned, aligned} vk_pipeline pipeline_flash_attn_f32_f16_D64[GGML_TYPE_COUNT][2][2][2]; @@ -264,6 +375,7 @@ struct vk_device_struct { vk_pipeline pipeline_flash_attn_f32_f16_D112[GGML_TYPE_COUNT][2][2][2]; vk_pipeline pipeline_flash_attn_f32_f16_D128[GGML_TYPE_COUNT][2][2][2]; vk_pipeline pipeline_flash_attn_f32_f16_D256[GGML_TYPE_COUNT][2][2][2]; + vk_pipeline pipeline_flash_attn_split_k_reduce; std::unordered_map pipelines; std::unordered_map pipeline_descriptor_set_requirements; @@ -357,6 +469,7 @@ struct vk_mat_mat_push_constants { uint32_t batch_stride_a; uint32_t batch_stride_b; uint32_t batch_stride_d; uint32_t k_split; uint32_t ne02; uint32_t ne12; uint32_t broadcast2; uint32_t broadcast3; + uint32_t padded_N; }; struct vk_mat_vec_push_constants { uint32_t ncols; uint32_t stride_a; uint32_t stride_b; uint32_t stride_d; @@ -369,6 +482,7 @@ struct vk_mat_mat_id_push_constants { uint32_t stride_a; uint32_t stride_b; uint32_t stride_d; uint32_t batch_stride_a; uint32_t batch_stride_b; uint32_t batch_stride_d; uint32_t nei0; uint32_t nei1; uint32_t nbi1; uint32_t ne11; + uint32_t padded_N; }; struct vk_mat_vec_id_push_constants { uint32_t ncols; uint32_t stride_a; uint32_t stride_b; uint32_t stride_d; @@ -411,6 +525,10 @@ struct vk_flash_attn_push_constants { uint32_t n_head_log2; float m0; float m1; + + uint32_t gqa_ratio; + uint32_t split_kv; + uint32_t k_num; }; struct vk_op_push_constants { @@ -493,6 +611,11 @@ struct vk_op_rope_push_constants { float corr_dims[2]; float theta_scale; uint32_t has_ff; + uint32_t ne02; + uint32_t s1; + uint32_t s2; + int32_t sections[4]; + uint32_t is_back; }; struct vk_op_soft_max_push_constants { @@ -549,13 +672,11 @@ struct vk_op_rwkv_wkv6_push_constants { uint32_t H; }; -// Allow pre-recording command buffers -struct vk_staging_memcpy { - vk_staging_memcpy(void * _dst, const void * _src, size_t _n) : dst(_dst), src(_src), n(_n) {} - - void * dst; - const void * src; - size_t n; +struct vk_op_rwkv_wkv7_push_constants { + uint32_t B; + uint32_t T; + uint32_t C; + uint32_t H; }; struct vk_op_upscale_push_constants { @@ -565,6 +686,15 @@ struct vk_op_upscale_push_constants { float sf0; float sf1; float sf2; float sf3; }; +// Allow pre-recording command buffers +struct vk_staging_memcpy { + vk_staging_memcpy(void * _dst, const void * _src, size_t _n) : dst(_dst), src(_src), n(_n) {} + + void * dst; + const void * src; + size_t n; +}; + struct vk_context_struct { vk_submission * s; std::vector seqs; @@ -679,7 +809,8 @@ struct ggml_backend_vk_context { ggml_vk_garbage_collector gc; size_t prealloc_size_x, prealloc_size_y, prealloc_size_split_k; vk_buffer prealloc_x, prealloc_y, prealloc_split_k; - vk::Fence fence; + vk::Fence fence, almost_ready_fence; + bool almost_ready_fence_pending {}; vk_buffer buffer_pool[MAX_VK_BUFFERS]; @@ -770,6 +901,39 @@ typedef void (*ggml_vk_func_t)(ggml_backend_vk_context * ctx, vk_context& subctx static void ggml_backend_vk_free(ggml_backend_t backend); +// Wait for ctx->fence to be signaled. +static void ggml_vk_wait_for_fence(ggml_backend_vk_context * ctx) { + // Use waitForFences while most of the graph executes. Hopefully the CPU can sleep + // during this wait. + if (ctx->almost_ready_fence_pending) { + VK_CHECK(ctx->device->device.waitForFences({ ctx->almost_ready_fence }, true, UINT64_MAX), "almost_ready_fence"); + ctx->device->device.resetFences({ ctx->almost_ready_fence }); + ctx->almost_ready_fence_pending = false; + } + + // Spin (w/pause) waiting for the graph to finish executing. + vk::Result result; + while ((result = ctx->device->device.getFenceStatus(ctx->fence)) != vk::Result::eSuccess) { + if (result != vk::Result::eNotReady) { + fprintf(stderr, "ggml_vulkan: error %s at %s:%d\n", to_string(result).c_str(), __FILE__, __LINE__); + exit(1); + } + for (uint32_t i = 0; i < 100; ++i) { + YIELD(); + YIELD(); + YIELD(); + YIELD(); + YIELD(); + YIELD(); + YIELD(); + YIELD(); + YIELD(); + YIELD(); + } + } + ctx->device->device.resetFences({ ctx->fence }); +} + // variables to track number of compiles in progress static uint32_t compile_count = 0; static std::mutex compile_count_mutex; @@ -1294,7 +1458,9 @@ static vk_buffer ggml_vk_create_buffer_check(vk_device& device, size_t size, vk: static vk_buffer ggml_vk_create_buffer_device(vk_device& device, size_t size) { vk_buffer buf; try { - if (device->uma) { + if (device->prefer_host_memory) { + buf = ggml_vk_create_buffer(device, size, vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent, vk::MemoryPropertyFlagBits::eDeviceLocal); + } else if (device->uma) { // Fall back to host memory type buf = ggml_vk_create_buffer(device, size, vk::MemoryPropertyFlagBits::eDeviceLocal, vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent); } else { @@ -1369,7 +1535,7 @@ static std::array fa_rows_cols(uint32_t D, uint32_t clamp, ggml_typ // small rows, large cols if (small_rows) { - return {flash_attention_num_small_rows, 128}; + return {flash_attention_num_small_rows, 64}; } // small cols to reduce register count if (ggml_is_quantized(type) || D == 256) { @@ -1382,6 +1548,10 @@ static bool ggml_vk_matmul_shmem_support(const vk_device& device, const std::vec uint32_t lut_size = 0; switch (src0_type) { + case GGML_TYPE_IQ1_S: + case GGML_TYPE_IQ1_M: + lut_size = 2*2048; + break; case GGML_TYPE_IQ2_XXS: lut_size = 8*256; break; @@ -1423,16 +1593,85 @@ static bool ggml_vk_matmul_shmem_support(const vk_device& device, const std::vec return supported; } +struct GpuPipelineConfig { + // GPU architecture identifier. + // Example: vk_device_architecture::AMD_GCN + vk_device_architecture arch; + + // Mapping of pipeline names to their specific subgroup sizes. + // Example: {"soft_max_f32", 64} + std::unordered_map pipelines; + + // Default subgroup size for this GPU. + // Defaults to 0 if not explicitly provided. + uint32_t default_subgroup_size = 0; +}; + +// Pipeline configuration for RDNA1 GPUs. +static const std::unordered_map rdna1_pipelines = { + {"soft_max", 64}, {"im2col", 64}, + {"argmax", 64}, {"mul_mat_vec", 64}, + {"mul_mat_vec_f16", 32}, {"mul_mat_vec_f32_f16", 32} +}; + +// Pipeline configuration for RDNA2 GPUs. +static const std::unordered_map rdna2_pipelines = { + {"soft_max", 64}, {"im2col", 64}, +}; + +static constexpr uint32_t RDNA_DEFAULT_SUBGROUP_SIZE = 32; + +// Define configurations for different GPUs. +static std::vector gpu_pipeline_configs = { + { + vk_device_architecture::AMD_RDNA1, + { + rdna1_pipelines, + }, + RDNA_DEFAULT_SUBGROUP_SIZE + }, + { + vk_device_architecture::AMD_RDNA2, + { + rdna2_pipelines, + }, + RDNA_DEFAULT_SUBGROUP_SIZE + }, +}; + +static uint32_t get_subgroup_size(const std::string &pipeline_name, const vk_device_architecture &arch) { + for (const auto &config : gpu_pipeline_configs) { + if (config.arch == arch) { + auto pipIt = config.pipelines.find(pipeline_name); + if (pipIt != config.pipelines.end()) { + return pipIt->second; + } + std::vector> sorted_pipelines(config.pipelines.begin(), config.pipelines.end()); + std::sort(sorted_pipelines.begin(), sorted_pipelines.end(), + [](const auto &a, const auto &b) { return a.first.size() > b.first.size(); }); + for (const auto &entry : sorted_pipelines) { + if (pipeline_name.find(entry.first) != std::string::npos) { + return entry.second; + } + } + return config.default_subgroup_size; + } + } + return 0; // If no matching configuration is found +} + static void ggml_vk_load_shaders(vk_device& device) { VK_LOG_DEBUG("ggml_vk_load_shaders(" << device->name << ")"); // some shaders have a minimum subgroup size + const uint32_t subgroup_size_8 = std::max(device->subgroup_size, 8u); const uint32_t subgroup_size_16 = std::max(device->subgroup_size, 16u); const uint32_t subgroup_size_32 = std::max(device->subgroup_size, 32u); // mulmat std::vector l_warptile, m_warptile, s_warptile, l_warptile_mmq, m_warptile_mmq, s_warptile_mmq, + l_warptile_mmq_int, m_warptile_mmq_int, s_warptile_mmq_int, l_warptile_mmq_k, m_warptile_mmq_k, s_warptile_mmq_k, l_warptile_mmqid, m_warptile_mmqid, s_warptile_mmqid; std::array l_wg_denoms, m_wg_denoms, s_wg_denoms, @@ -1443,36 +1682,36 @@ static void ggml_vk_load_shaders(vk_device& device) { uint32_t l_align, m_align, s_align; if (device->coopmat2) { // spec constants and tile sizes for non-quant matmul/matmul_id - l_warptile = { 256, 128, 256, 64 }; - m_warptile = { 256, 128, 128, 64 }; - s_warptile = { 128, 64, 64, 64 }; + l_warptile = { 256, 128, 256, 64, 1 }; + m_warptile = { 256, 128, 128, 64, 0 }; + s_warptile = { 128, 64, 64, 64, 0 }; l_wg_denoms = {128, 256, 1 }; m_wg_denoms = {128, 128, 1 }; s_wg_denoms = { 64, 64, 1 }; // spec constants and tile sizes for quant matmul (non-Qi_K) - l_warptile_mmq = { 256, 128, 256, 64 }; - m_warptile_mmq = { 256, 128, 128, 64 }; - s_warptile_mmq = { 256, 128, 128, 64 }; + l_warptile_mmq = { 256, 128, 256, 64, 1 }; + m_warptile_mmq = { 256, 128, 128, 64, 1 }; + s_warptile_mmq = { 256, 32, 64, 128, 0 }; l_mmq_wg_denoms = { 128, 256, 1 }; m_mmq_wg_denoms = { 128, 128, 1 }; - s_mmq_wg_denoms = { 128, 128, 1 }; + s_mmq_wg_denoms = { 32, 64, 1 }; // spec constants and tile sizes for quant matmul (Qi_K) - l_warptile_mmq_k = { 256, 128, 512, 16 }; - m_warptile_mmq_k = { 256, 128, 256, 16 }; - s_warptile_mmq_k = { 256, 32, 128, 64 }; - l_mmq_wg_denoms_k = { 128, 512, 1 }; - m_mmq_wg_denoms_k = { 128, 256, 1 }; - s_mmq_wg_denoms_k = { 32, 128, 1 }; + l_warptile_mmq_k = { 256, 64, 128, 64, 1 }; + m_warptile_mmq_k = { 256, 32, 64, 64, 0 }; + s_warptile_mmq_k = { 256, 32, 32, 128, 0 }; + l_mmq_wg_denoms_k = { 64, 128, 1 }; + m_mmq_wg_denoms_k = { 32, 64, 1 }; + s_mmq_wg_denoms_k = { 32, 32, 1 }; // spec constants and tile sizes for quant matmul_id - l_warptile_mmqid = { 256, 128, 128, 16 }; - m_warptile_mmqid = { 256, 128, 64, 16 }; - s_warptile_mmqid = { 256, 64, 64, 16 }; - l_mmqid_wg_denoms = { 128, 128, 1 }; + l_warptile_mmqid = { 256, 128, 64, 16, 0 }; + m_warptile_mmqid = { 256, 128, 64, 16, 0 }; + s_warptile_mmqid = { 256, 128, 64, 16, 0 }; + l_mmqid_wg_denoms = { 128, 64, 1 }; m_mmqid_wg_denoms = { 128, 64, 1 }; - s_mmqid_wg_denoms = { 64, 64, 1 }; + s_mmqid_wg_denoms = { 128, 64, 1 }; l_align = 128; m_align = 64; @@ -1489,13 +1728,17 @@ static void ggml_vk_load_shaders(vk_device& device) { const uint32_t tk_m = device->coopmat_support ? device->coopmat_k : 1; const uint32_t tk_s = device->coopmat_support ? device->coopmat_k : 1; - l_warptile = { 128, 128, 128, 16, device->subgroup_size * 2, 64, 2, tm_l, tn_l, tk_l, device->subgroup_size }; - m_warptile = { 128, 64, 64, 16, device->subgroup_size, 32, 2, tm_m, tn_m, tk_m, device->subgroup_size }; - s_warptile = { subgroup_size_16, 32, 32, 16, 32, 32, 2, tm_s, tn_s, tk_s, device->subgroup_size }; + l_warptile = { 128, 128, 128, 16, subgroup_size_8 * 2, 64, 2, tm_l, tn_l, tk_l, subgroup_size_8 }; + m_warptile = { 128, 64, 64, 16, subgroup_size_8, 32, 2, tm_m, tn_m, tk_m, subgroup_size_8 }; + s_warptile = { subgroup_size_16, 32, 32, 16, 32, 32, 2, tm_s, tn_s, tk_s, subgroup_size_8 }; - l_warptile_mmq = { 128, 128, 128, 32, device->subgroup_size * 2, 64, 2, tm_l, tn_l, tk_l, device->subgroup_size }; - m_warptile_mmq = { 128, 64, 64, 32, device->subgroup_size, 32, 2, tm_m, tn_m, tk_m, device->subgroup_size }; - s_warptile_mmq = { subgroup_size_32, 32, 32, 32, 32, 32, 2, tm_s, tn_s, tk_s, device->subgroup_size }; + l_warptile_mmq = { 128, 128, 128, 32, subgroup_size_8 * 2, 64, 2, tm_l, tn_l, tk_l, subgroup_size_8 }; + m_warptile_mmq = { 128, 64, 64, 32, subgroup_size_8, 32, 2, tm_m, tn_m, tk_m, subgroup_size_8 }; + s_warptile_mmq = { subgroup_size_32, 32, 32, 32, 32, 32, 2, tm_s, tn_s, tk_s, subgroup_size_8 }; + + l_warptile_mmq_int = { 128, 128, 128, 32, subgroup_size_8 * 2, 64, 2, 4, 4, 1, subgroup_size_8 }; + m_warptile_mmq_int = { 128, 64, 64, 32, subgroup_size_8, 32, 2, 2, 2, 1, subgroup_size_8 }; + s_warptile_mmq_int = { subgroup_size_32, 32, 32, 32, 32, 32, 2, 2, 1, 1, subgroup_size_8 }; l_mmq_wg_denoms = l_wg_denoms = {128, 128, 1 }; m_mmq_wg_denoms = m_wg_denoms = { 64, 64, 1 }; @@ -1548,6 +1791,10 @@ static void ggml_vk_load_shaders(vk_device& device) { uint32_t parameter_count, uint32_t push_constant_size, std::array wg_denoms, const std::vector& specialization_constants, uint32_t align, bool disable_robustness = false, bool require_full_subgroups = false, uint32_t required_subgroup_size = 0) { + if (!require_full_subgroups && required_subgroup_size == 0) { + required_subgroup_size = get_subgroup_size(name, device->architecture); + } + if (!pipeline) { pipeline = std::make_shared(); pipeline->name = name; @@ -1586,6 +1833,8 @@ static void ggml_vk_load_shaders(vk_device& device) { // can't use 256 for D==80. uint32_t wg_size = (small_rows && (D % 32) == 0) ? 256 : 128; auto rows_cols = fa_rows_cols(D, clamp, type, small_rows); + // mask dim1 is padded to 64, we rely on this to avoid clamping mask loads + GGML_ASSERT((GGML_KQ_MASK_PAD % rows_cols[0]) == 0); return {wg_size, rows_cols[0], rows_cols[1], (D), clamp}; }; @@ -1619,6 +1868,8 @@ static void ggml_vk_load_shaders(vk_device& device) { //CREATE_FA(GGML_TYPE_Q4_K, q4_k) //CREATE_FA(GGML_TYPE_Q5_K, q5_k) //CREATE_FA(GGML_TYPE_Q6_K, q6_k) + //CREATE_FA(GGML_TYPE_IQ1_S, iq1_s) + //CREATE_FA(GGML_TYPE_IQ1_M, iq1_m) //CREATE_FA(GGML_TYPE_IQ2_XXS, iq2_xxs) //CREATE_FA(GGML_TYPE_IQ2_XS, iq2_xs) //CREATE_FA(GGML_TYPE_IQ2_S, iq2_s) @@ -1653,6 +1904,8 @@ static void ggml_vk_load_shaders(vk_device& device) { CREATE_MM(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_Q4_K].f16acc, matmul_q4_k_f16, _f16acc, mmq_wg_denoms_k, warptile_mmq_k, vk_mat_mat_push_constants, 3) CREATE_MM(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_Q5_K].f16acc, matmul_q5_k_f16, _f16acc, mmq_wg_denoms_k, warptile_mmq_k, vk_mat_mat_push_constants, 3) CREATE_MM(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_Q6_K].f16acc, matmul_q6_k_f16, _f16acc, mmq_wg_denoms_k, warptile_mmq_k, vk_mat_mat_push_constants, 3) + CREATE_MM(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_IQ1_S].f16acc, matmul_iq1_s_f16, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3) + CREATE_MM(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_IQ1_M].f16acc, matmul_iq1_m_f16, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3) CREATE_MM(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_IQ2_XXS].f16acc, matmul_iq2_xxs_f16, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3) CREATE_MM(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_IQ2_XS].f16acc, matmul_iq2_xs_f16, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3) CREATE_MM(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_IQ2_S].f16acc, matmul_iq2_s_f16, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3) @@ -1672,6 +1925,8 @@ static void ggml_vk_load_shaders(vk_device& device) { CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K].f16acc, matmul_id_q4_k_f16, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4) CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K].f16acc, matmul_id_q5_k_f16, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4) CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K].f16acc, matmul_id_q6_k_f16, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4) + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_S].f16acc, matmul_id_iq1_s_f16, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4) + CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_M].f16acc, matmul_id_iq1_m_f16, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4) CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XXS].f16acc, matmul_id_iq2_xxs_f16, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4) CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XS].f16acc, matmul_id_iq2_xs_f16, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4) CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_S].f16acc, matmul_id_iq2_s_f16, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4) @@ -1726,6 +1981,8 @@ static void ggml_vk_load_shaders(vk_device& device) { CREATE_MM(GGML_TYPE_Q4_K, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K].f16acc, matmul_q4_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); CREATE_MM(GGML_TYPE_Q5_K, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K].f16acc, matmul_q5_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); CREATE_MM(GGML_TYPE_Q6_K, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K].f16acc, matmul_q6_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM(GGML_TYPE_IQ1_S, pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ1_S].f16acc, matmul_iq1_s_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM(GGML_TYPE_IQ1_M, pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ1_M].f16acc, matmul_iq1_m_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); CREATE_MM(GGML_TYPE_IQ2_XXS, pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ2_XXS].f16acc, matmul_iq2_xxs_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); CREATE_MM(GGML_TYPE_IQ2_XS, pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ2_XS].f16acc, matmul_iq2_xs_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); CREATE_MM(GGML_TYPE_IQ2_S, pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ2_S].f16acc, matmul_iq2_s_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); @@ -1745,6 +2002,8 @@ static void ggml_vk_load_shaders(vk_device& device) { CREATE_MM(GGML_TYPE_Q4_K, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K].f16acc, matmul_q4_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); CREATE_MM(GGML_TYPE_Q5_K, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K].f16acc, matmul_q5_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); CREATE_MM(GGML_TYPE_Q6_K, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K].f16acc, matmul_q6_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM(GGML_TYPE_IQ1_S, pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ1_S].f16acc, matmul_iq1_s_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM(GGML_TYPE_IQ1_M, pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ1_M].f16acc, matmul_iq1_m_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); CREATE_MM(GGML_TYPE_IQ2_XXS, pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ2_XXS].f16acc, matmul_iq2_xxs_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); CREATE_MM(GGML_TYPE_IQ2_XS, pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ2_XS].f16acc, matmul_iq2_xs_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); CREATE_MM(GGML_TYPE_IQ2_S, pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ2_S].f16acc, matmul_iq2_s_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); @@ -1770,13 +2029,15 @@ static void ggml_vk_load_shaders(vk_device& device) { CREATE_MM(GGML_TYPE_Q4_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K].f16acc, matmul_id_q4_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); CREATE_MM(GGML_TYPE_Q5_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K].f16acc, matmul_id_q5_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); CREATE_MM(GGML_TYPE_Q6_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K].f16acc, matmul_id_q6_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(GGML_TYPE_IQ1_S, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_S].f16acc, matmul_id_iq1_s_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(GGML_TYPE_IQ1_M, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_M].f16acc, matmul_id_iq1_m_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); CREATE_MM(GGML_TYPE_IQ2_XXS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XXS].f16acc, matmul_id_iq2_xxs_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); - CREATE_MM(GGML_TYPE_IQ2_XS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XS].f16acc, matmul_id_iq2_xs_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); - CREATE_MM(GGML_TYPE_IQ2_S, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_S].f16acc, matmul_id_iq2_s_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(GGML_TYPE_IQ2_XS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XS].f16acc, matmul_id_iq2_xs_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(GGML_TYPE_IQ2_S, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_S].f16acc, matmul_id_iq2_s_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); CREATE_MM(GGML_TYPE_IQ3_XXS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_XXS].f16acc, matmul_id_iq3_xxs_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); - CREATE_MM(GGML_TYPE_IQ3_S, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_S].f16acc, matmul_id_iq3_s_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); - CREATE_MM(GGML_TYPE_IQ4_XS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_XS].f16acc, matmul_id_iq4_xs_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); - CREATE_MM(GGML_TYPE_IQ4_NL, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL].f16acc, matmul_id_iq4_nl_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(GGML_TYPE_IQ3_S, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_S].f16acc, matmul_id_iq3_s_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(GGML_TYPE_IQ4_XS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_XS].f16acc, matmul_id_iq4_xs_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(GGML_TYPE_IQ4_NL, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL].f16acc, matmul_id_iq4_nl_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); } else { CREATE_MM(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0].f16acc, matmul_id_q4_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); CREATE_MM(GGML_TYPE_Q4_1, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1].f16acc, matmul_id_q4_1_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); @@ -1789,13 +2050,15 @@ static void ggml_vk_load_shaders(vk_device& device) { CREATE_MM(GGML_TYPE_Q4_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K].f16acc, matmul_id_q4_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); CREATE_MM(GGML_TYPE_Q5_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K].f16acc, matmul_id_q5_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); CREATE_MM(GGML_TYPE_Q6_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K].f16acc, matmul_id_q6_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(GGML_TYPE_IQ1_S, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_S].f16acc, matmul_id_iq1_s_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(GGML_TYPE_IQ1_M, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_M].f16acc, matmul_id_iq1_m_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); CREATE_MM(GGML_TYPE_IQ2_XXS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XXS].f16acc, matmul_id_iq2_xxs_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); - CREATE_MM(GGML_TYPE_IQ2_XS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XS].f16acc, matmul_id_iq2_xs_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); - CREATE_MM(GGML_TYPE_IQ2_S, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_S].f16acc, matmul_id_iq2_s_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(GGML_TYPE_IQ2_XS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XS].f16acc, matmul_id_iq2_xs_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(GGML_TYPE_IQ2_S, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_S].f16acc, matmul_id_iq2_s_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); CREATE_MM(GGML_TYPE_IQ3_XXS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_XXS].f16acc, matmul_id_iq3_xxs_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); - CREATE_MM(GGML_TYPE_IQ3_S, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_S].f16acc, matmul_id_iq3_s_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); - CREATE_MM(GGML_TYPE_IQ4_XS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_XS].f16acc, matmul_id_iq4_xs_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); - CREATE_MM(GGML_TYPE_IQ4_NL, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL].f16acc, matmul_id_iq4_nl_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(GGML_TYPE_IQ3_S, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_S].f16acc, matmul_id_iq3_s_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(GGML_TYPE_IQ4_XS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_XS].f16acc, matmul_id_iq4_xs_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(GGML_TYPE_IQ4_NL, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL].f16acc, matmul_id_iq4_nl_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); } #undef CREATE_MM2 #undef CREATE_MM @@ -1817,6 +2080,14 @@ static void ggml_vk_load_shaders(vk_device& device) { if (device->mul_mat ## ID ## _s[TYPE]) \ ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_s, #NAMELC #F16ACC "_aligned_s", NAMELC ## _aligned ## F16ACC ## _len, NAMELC ## _aligned ## F16ACC ## _data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, s_align); \ +#define CREATE_MMQ(TYPE, PIPELINE_NAME, NAMELC, F16ACC, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT, ID) \ + if (device->mul_mat ## ID ## _l[TYPE]) \ + ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->l, #NAMELC #F16ACC "_l", NAMELC ## F16ACC ## _len, NAMELC ## F16ACC ## _data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, 1); \ + if (device->mul_mat ## ID ## _m[TYPE]) \ + ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->m, #NAMELC #F16ACC "_m", NAMELC ## F16ACC ## _len, NAMELC ## F16ACC ## _data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, 1); \ + if (device->mul_mat ## ID ## _s[TYPE]) \ + ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->s, #NAMELC #F16ACC "_s", NAMELC ## F16ACC ## _len, NAMELC ## F16ACC ## _data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, 1); \ + // Create 2 variants, {f16,f32} accumulator #define CREATE_MM2(TYPE, PIPELINE_NAME, NAMELC, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT, ID) \ CREATE_MM(TYPE, PIPELINE_NAME . f16acc, NAMELC, _f16acc, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT, ID) \ @@ -1838,6 +2109,8 @@ static void ggml_vk_load_shaders(vk_device& device) { CREATE_MM(GGML_TYPE_Q4_K, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K].f16acc, matmul_q4_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); CREATE_MM(GGML_TYPE_Q5_K, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K].f16acc, matmul_q5_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); CREATE_MM(GGML_TYPE_Q6_K, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K].f16acc, matmul_q6_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM(GGML_TYPE_IQ1_S, pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ1_S].f16acc, matmul_iq1_s_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM(GGML_TYPE_IQ1_M, pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ1_M].f16acc, matmul_iq1_m_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); CREATE_MM(GGML_TYPE_IQ2_XXS, pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ2_XXS].f16acc, matmul_iq2_xxs_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); CREATE_MM(GGML_TYPE_IQ2_XS, pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ2_XS].f16acc, matmul_iq2_xs_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); CREATE_MM(GGML_TYPE_IQ2_S, pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ2_S].f16acc, matmul_iq2_s_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); @@ -1846,6 +2119,16 @@ static void ggml_vk_load_shaders(vk_device& device) { CREATE_MM(GGML_TYPE_IQ4_XS, pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_XS].f16acc, matmul_iq4_xs_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); CREATE_MM(GGML_TYPE_IQ4_NL, pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL].f16acc, matmul_iq4_nl_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); +#if defined(GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT) + if (device->integer_dot_product) { + CREATE_MMQ(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat_q8_1[GGML_TYPE_Q4_0].f16acc, matmul_q4_0_q8_1, _f16acc, mmq_wg_denoms, warptile_mmq_int, vk_mat_mat_push_constants, 3, ); + CREATE_MMQ(GGML_TYPE_Q4_1, pipeline_dequant_mul_mat_mat_q8_1[GGML_TYPE_Q4_1].f16acc, matmul_q4_1_q8_1, _f16acc, mmq_wg_denoms, warptile_mmq_int, vk_mat_mat_push_constants, 3, ); + CREATE_MMQ(GGML_TYPE_Q5_0, pipeline_dequant_mul_mat_mat_q8_1[GGML_TYPE_Q5_0].f16acc, matmul_q5_0_q8_1, _f16acc, mmq_wg_denoms, warptile_mmq_int, vk_mat_mat_push_constants, 3, ); + CREATE_MMQ(GGML_TYPE_Q5_1, pipeline_dequant_mul_mat_mat_q8_1[GGML_TYPE_Q5_1].f16acc, matmul_q5_1_q8_1, _f16acc, mmq_wg_denoms, warptile_mmq_int, vk_mat_mat_push_constants, 3, ); + CREATE_MMQ(GGML_TYPE_Q8_0, pipeline_dequant_mul_mat_mat_q8_1[GGML_TYPE_Q8_0].f16acc, matmul_q8_0_q8_1, _f16acc, mmq_wg_denoms, warptile_mmq_int, vk_mat_mat_push_constants, 3, ); + } +#endif + CREATE_MM(GGML_TYPE_F32, pipeline_matmul_id_f32, matmul_id_f32_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id); CREATE_MM2(GGML_TYPE_F16, pipeline_matmul_id_f16, matmul_id_f16, wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id); CREATE_MM2(GGML_TYPE_F16, pipeline_matmul_id_f16_f32, matmul_id_f16_f32, wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id); @@ -1861,6 +2144,8 @@ static void ggml_vk_load_shaders(vk_device& device) { CREATE_MM(GGML_TYPE_Q4_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K].f16acc, matmul_id_q4_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); CREATE_MM(GGML_TYPE_Q5_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K].f16acc, matmul_id_q5_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); CREATE_MM(GGML_TYPE_Q6_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K].f16acc, matmul_id_q6_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(GGML_TYPE_IQ1_S, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_S].f16acc, matmul_id_iq1_s_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(GGML_TYPE_IQ1_M, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_M].f16acc, matmul_id_iq1_m_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); CREATE_MM(GGML_TYPE_IQ2_XXS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XXS].f16acc, matmul_id_iq2_xxs_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); CREATE_MM(GGML_TYPE_IQ2_XS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XS].f16acc, matmul_id_iq2_xs_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); CREATE_MM(GGML_TYPE_IQ2_S, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_S].f16acc, matmul_id_iq2_s_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); @@ -1869,6 +2154,7 @@ static void ggml_vk_load_shaders(vk_device& device) { CREATE_MM(GGML_TYPE_IQ4_XS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_XS].f16acc, matmul_id_iq4_xs_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); CREATE_MM(GGML_TYPE_IQ4_NL, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL].f16acc, matmul_id_iq4_nl_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); #undef CREATE_MM2 +#undef CREATE_MMQ #undef CREATE_MM } else { // Create 6 variants, {s,m,l}x{unaligned,aligned} @@ -1886,6 +2172,14 @@ static void ggml_vk_load_shaders(vk_device& device) { if (device->mul_mat ## ID ## _s[TYPE]) \ ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_s, #NAMELC #F16ACC "_aligned_s", NAMELC ## _aligned ## F16ACC ## _fp32_len, NAMELC ## _aligned ## F16ACC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, s_align); \ +#define CREATE_MMQ(TYPE, PIPELINE_NAME, NAMELC, F16ACC, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT, ID) \ + if (device->mul_mat ## ID ## _l[TYPE]) \ + ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->l, #NAMELC #F16ACC "_l", NAMELC ## F16ACC ## _fp32_len, NAMELC ## F16ACC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, 1); \ + if (device->mul_mat ## ID ## _m[TYPE]) \ + ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->m, #NAMELC #F16ACC "_m", NAMELC ## F16ACC ## _fp32_len, NAMELC ## F16ACC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, 1); \ + if (device->mul_mat ## ID ## _s[TYPE]) \ + ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->s, #NAMELC #F16ACC "_s", NAMELC ## F16ACC ## _fp32_len, NAMELC ## F16ACC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, 1); \ + CREATE_MM(GGML_TYPE_F32, pipeline_matmul_f32, matmul_f32_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 3, ); CREATE_MM(GGML_TYPE_F32, pipeline_matmul_f32_f16, matmul_f32_f16, , wg_denoms, warptile, vk_mat_mat_push_constants, 3, ); CREATE_MM(GGML_TYPE_F16, pipeline_matmul_f16.f32acc, matmul_f16, , wg_denoms, warptile, vk_mat_mat_push_constants, 3, ); @@ -1902,13 +2196,25 @@ static void ggml_vk_load_shaders(vk_device& device) { CREATE_MM(GGML_TYPE_Q4_K, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K].f32acc, matmul_q4_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); CREATE_MM(GGML_TYPE_Q5_K, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K].f32acc, matmul_q5_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); CREATE_MM(GGML_TYPE_Q6_K, pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K].f32acc, matmul_q6_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM(GGML_TYPE_IQ1_S, pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ1_S].f32acc, matmul_iq1_s_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM(GGML_TYPE_IQ1_M, pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ1_M].f32acc, matmul_iq1_m_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); CREATE_MM(GGML_TYPE_IQ2_XXS, pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ2_XXS].f32acc, matmul_iq2_xxs_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); - CREATE_MM(GGML_TYPE_IQ2_XS, pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ2_XS].f32acc, matmul_iq2_xs_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); - CREATE_MM(GGML_TYPE_IQ2_S, pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ2_S].f32acc, matmul_iq2_s_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM(GGML_TYPE_IQ2_XS, pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ2_XS].f32acc, matmul_iq2_xs_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM(GGML_TYPE_IQ2_S, pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ2_S].f32acc, matmul_iq2_s_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); CREATE_MM(GGML_TYPE_IQ3_XXS, pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ3_XXS].f32acc, matmul_iq3_xxs_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); - CREATE_MM(GGML_TYPE_IQ3_S, pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ3_S].f32acc, matmul_iq3_s_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); - CREATE_MM(GGML_TYPE_IQ4_XS, pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_XS].f32acc, matmul_iq4_xs_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); - CREATE_MM(GGML_TYPE_IQ4_NL, pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL].f32acc, matmul_iq4_nl_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM(GGML_TYPE_IQ3_S, pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ3_S].f32acc, matmul_iq3_s_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM(GGML_TYPE_IQ4_XS, pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_XS].f32acc, matmul_iq4_xs_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + CREATE_MM(GGML_TYPE_IQ4_NL, pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL].f32acc, matmul_iq4_nl_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, ); + +#if defined(GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT) + if (device->integer_dot_product) { + CREATE_MMQ(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat_q8_1[GGML_TYPE_Q4_0].f32acc, matmul_q4_0_q8_1, , mmq_wg_denoms, warptile_mmq_int, vk_mat_mat_push_constants, 3, ); + CREATE_MMQ(GGML_TYPE_Q4_1, pipeline_dequant_mul_mat_mat_q8_1[GGML_TYPE_Q4_1].f32acc, matmul_q4_1_q8_1, , mmq_wg_denoms, warptile_mmq_int, vk_mat_mat_push_constants, 3, ); + CREATE_MMQ(GGML_TYPE_Q5_0, pipeline_dequant_mul_mat_mat_q8_1[GGML_TYPE_Q5_0].f32acc, matmul_q5_0_q8_1, , mmq_wg_denoms, warptile_mmq_int, vk_mat_mat_push_constants, 3, ); + CREATE_MMQ(GGML_TYPE_Q5_1, pipeline_dequant_mul_mat_mat_q8_1[GGML_TYPE_Q5_1].f32acc, matmul_q5_1_q8_1, , mmq_wg_denoms, warptile_mmq_int, vk_mat_mat_push_constants, 3, ); + CREATE_MMQ(GGML_TYPE_Q8_0, pipeline_dequant_mul_mat_mat_q8_1[GGML_TYPE_Q8_0].f32acc, matmul_q8_0_q8_1, , mmq_wg_denoms, warptile_mmq_int, vk_mat_mat_push_constants, 3, ); + } +#endif CREATE_MM(GGML_TYPE_F32, pipeline_matmul_id_f32, matmul_id_f32_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id); CREATE_MM(GGML_TYPE_F16, pipeline_matmul_id_f16.f32acc, matmul_id_f16, , wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id); @@ -1925,13 +2231,15 @@ static void ggml_vk_load_shaders(vk_device& device) { CREATE_MM(GGML_TYPE_Q4_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K].f32acc, matmul_id_q4_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); CREATE_MM(GGML_TYPE_Q5_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K].f32acc, matmul_id_q5_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); CREATE_MM(GGML_TYPE_Q6_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K].f32acc, matmul_id_q6_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(GGML_TYPE_IQ1_S, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_S].f32acc, matmul_id_iq1_s_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(GGML_TYPE_IQ1_M, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_M].f32acc, matmul_id_iq1_m_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); CREATE_MM(GGML_TYPE_IQ2_XXS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XXS].f32acc, matmul_id_iq2_xxs_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); - CREATE_MM(GGML_TYPE_IQ2_XS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XS].f32acc, matmul_id_iq2_xs_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); - CREATE_MM(GGML_TYPE_IQ2_S, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_S].f32acc, matmul_id_iq2_s_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(GGML_TYPE_IQ2_XS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XS].f32acc, matmul_id_iq2_xs_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(GGML_TYPE_IQ2_S, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_S].f32acc, matmul_id_iq2_s_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); CREATE_MM(GGML_TYPE_IQ3_XXS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_XXS].f32acc, matmul_id_iq3_xxs_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); - CREATE_MM(GGML_TYPE_IQ3_S, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_S].f32acc, matmul_id_iq3_s_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); - CREATE_MM(GGML_TYPE_IQ4_XS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_XS].f32acc, matmul_id_iq4_xs_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); - CREATE_MM(GGML_TYPE_IQ4_NL, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL].f32acc, matmul_id_iq4_nl_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(GGML_TYPE_IQ3_S, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_S].f32acc, matmul_id_iq3_s_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(GGML_TYPE_IQ4_XS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_XS].f32acc, matmul_id_iq4_xs_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); + CREATE_MM(GGML_TYPE_IQ4_NL, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL].f32acc, matmul_id_iq4_nl_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id); #undef CREATE_MM } @@ -1941,12 +2249,13 @@ static void ggml_vk_load_shaders(vk_device& device) { uint32_t rm_stdq = 1; uint32_t rm_kq = 2; if (device->vendor_id == VK_VENDOR_ID_AMD) { - if (device->subgroup_min_size == 64 && device->subgroup_max_size == 64) { // GCN + if (device->architecture == AMD_GCN) { rm_stdq = 2; rm_kq = 4; } } else if (device->vendor_id == VK_VENDOR_ID_INTEL) rm_stdq = 2; + uint32_t rm_iq = 2 * rm_kq; for (uint32_t i = 0; i < mul_mat_vec_max_cols; ++i) { ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_F32 ][i], "mul_mat_vec_f32_f32_f32_"+std::to_string(i+1), mul_mat_vec_f32_f32_f32_len, mul_mat_vec_f32_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2, i+1}, 1); @@ -1961,13 +2270,15 @@ static void ggml_vk_load_shaders(vk_device& device) { ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q4_K][i], "mul_mat_vec_q4_k_f32_f32_"+std::to_string(i+1), mul_mat_vec_q4_k_f32_f32_len, mul_mat_vec_q4_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true); ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q5_K][i], "mul_mat_vec_q5_k_f32_f32_"+std::to_string(i+1), mul_mat_vec_q5_k_f32_f32_len, mul_mat_vec_q5_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true); ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q6_K][i], "mul_mat_vec_q6_k_f32_f32_"+std::to_string(i+1), mul_mat_vec_q6_k_f32_f32_len, mul_mat_vec_q6_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_IQ2_XXS][i], "mul_mat_vec_iq2_xxs_f32_f32_"+std::to_string(i+1), mul_mat_vec_iq2_xxs_f32_f32_len, mul_mat_vec_iq2_xxs_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_IQ2_XS][i], "mul_mat_vec_iq2_xs_f32_f32_"+std::to_string(i+1), mul_mat_vec_iq2_xs_f32_f32_len, mul_mat_vec_iq2_xs_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_IQ2_S][i], "mul_mat_vec_iq2_s_f32_f32_"+std::to_string(i+1), mul_mat_vec_iq2_s_f32_f32_len, mul_mat_vec_iq2_s_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_IQ3_XXS][i], "mul_mat_vec_iq3_xxs_f32_f32_"+std::to_string(i+1), mul_mat_vec_iq3_xxs_f32_f32_len, mul_mat_vec_iq3_xxs_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_IQ3_S][i], "mul_mat_vec_iq3_s_f32_f32_"+std::to_string(i+1), mul_mat_vec_iq3_s_f32_f32_len, mul_mat_vec_iq3_s_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_IQ4_XS][i], "mul_mat_vec_iq4_xs_f32_f32_"+std::to_string(i+1), mul_mat_vec_iq4_xs_f32_f32_len, mul_mat_vec_iq4_xs_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_IQ4_NL][i], "mul_mat_vec_iq4_nl_f32_f32_"+std::to_string(i+1), mul_mat_vec_iq4_nl_f32_f32_len, mul_mat_vec_iq4_nl_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {subgroup_size_16, 2*rm_stdq, i+1}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_IQ1_S][i], "mul_mat_vec_iq1_s_f32_f32_"+std::to_string(i+1), mul_mat_vec_iq1_s_f32_f32_len, mul_mat_vec_iq1_s_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_IQ1_M][i], "mul_mat_vec_iq1_m_f32_f32_"+std::to_string(i+1), mul_mat_vec_iq1_m_f32_f32_len, mul_mat_vec_iq1_m_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_IQ2_XXS][i], "mul_mat_vec_iq2_xxs_f32_f32_"+std::to_string(i+1), mul_mat_vec_iq2_xxs_f32_f32_len, mul_mat_vec_iq2_xxs_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_IQ2_XS][i], "mul_mat_vec_iq2_xs_f32_f32_"+std::to_string(i+1), mul_mat_vec_iq2_xs_f32_f32_len, mul_mat_vec_iq2_xs_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_IQ2_S][i], "mul_mat_vec_iq2_s_f32_f32_"+std::to_string(i+1), mul_mat_vec_iq2_s_f32_f32_len, mul_mat_vec_iq2_s_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_IQ3_XXS][i], "mul_mat_vec_iq3_xxs_f32_f32_"+std::to_string(i+1), mul_mat_vec_iq3_xxs_f32_f32_len, mul_mat_vec_iq3_xxs_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_IQ3_S][i], "mul_mat_vec_iq3_s_f32_f32_"+std::to_string(i+1), mul_mat_vec_iq3_s_f32_f32_len, mul_mat_vec_iq3_s_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_IQ4_XS][i], "mul_mat_vec_iq4_xs_f32_f32_"+std::to_string(i+1), mul_mat_vec_iq4_xs_f32_f32_len, mul_mat_vec_iq4_xs_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_IQ4_NL][i], "mul_mat_vec_iq4_nl_f32_f32_"+std::to_string(i+1), mul_mat_vec_iq4_nl_f32_f32_len, mul_mat_vec_iq4_nl_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true); ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_F32 ][i], "mul_mat_vec_f32_f16_f32_"+std::to_string(i+1), mul_mat_vec_f32_f16_f32_len, mul_mat_vec_f32_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2, i+1}, 1); ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_F16 ][i], "mul_mat_vec_f16_f16_f32_"+std::to_string(i+1), mul_mat_vec_f16_f16_f32_len, mul_mat_vec_f16_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2, i+1}, 1); @@ -1981,13 +2292,15 @@ static void ggml_vk_load_shaders(vk_device& device) { ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q4_K][i], "mul_mat_vec_q4_k_f16_f32_"+std::to_string(i+1), mul_mat_vec_q4_k_f16_f32_len, mul_mat_vec_q4_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true); ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q5_K][i], "mul_mat_vec_q5_k_f16_f32_"+std::to_string(i+1), mul_mat_vec_q5_k_f16_f32_len, mul_mat_vec_q5_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true); ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q6_K][i], "mul_mat_vec_q6_k_f16_f32_"+std::to_string(i+1), mul_mat_vec_q6_k_f16_f32_len, mul_mat_vec_q6_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ2_XXS][i], "mul_mat_vec_iq2_xxs_f16_f32_"+std::to_string(i+1), mul_mat_vec_iq2_xxs_f16_f32_len, mul_mat_vec_iq2_xxs_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ2_XS][i], "mul_mat_vec_iq2_xs_f16_f32_"+std::to_string(i+1), mul_mat_vec_iq2_xs_f16_f32_len, mul_mat_vec_iq2_xs_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ2_S][i], "mul_mat_vec_iq2_s_f16_f32_"+std::to_string(i+1), mul_mat_vec_iq2_s_f16_f32_len, mul_mat_vec_iq2_s_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ3_XXS][i], "mul_mat_vec_iq3_xxs_f16_f32_"+std::to_string(i+1), mul_mat_vec_iq3_xxs_f16_f32_len, mul_mat_vec_iq3_xxs_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ3_S][i], "mul_mat_vec_iq3_s_f16_f32_"+std::to_string(i+1), mul_mat_vec_iq3_s_f16_f32_len, mul_mat_vec_iq3_s_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ4_XS][i], "mul_mat_vec_iq4_xs_f16_f32_"+std::to_string(i+1), mul_mat_vec_iq4_xs_f16_f32_len, mul_mat_vec_iq4_xs_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ4_NL][i], "mul_mat_vec_iq4_nl_f16_f32_"+std::to_string(i+1), mul_mat_vec_iq4_nl_f16_f32_len, mul_mat_vec_iq4_nl_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {subgroup_size_16, 2*rm_stdq, i+1}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ1_S][i], "mul_mat_vec_iq1_s_f16_f32_"+std::to_string(i+1), mul_mat_vec_iq1_s_f16_f32_len, mul_mat_vec_iq1_s_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ1_M][i], "mul_mat_vec_iq1_m_f16_f32_"+std::to_string(i+1), mul_mat_vec_iq1_m_f16_f32_len, mul_mat_vec_iq1_m_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ2_XXS][i], "mul_mat_vec_iq2_xxs_f16_f32_"+std::to_string(i+1), mul_mat_vec_iq2_xxs_f16_f32_len, mul_mat_vec_iq2_xxs_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ2_XS][i], "mul_mat_vec_iq2_xs_f16_f32_"+std::to_string(i+1), mul_mat_vec_iq2_xs_f16_f32_len, mul_mat_vec_iq2_xs_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ2_S][i], "mul_mat_vec_iq2_s_f16_f32_"+std::to_string(i+1), mul_mat_vec_iq2_s_f16_f32_len, mul_mat_vec_iq2_s_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ3_XXS][i], "mul_mat_vec_iq3_xxs_f16_f32_"+std::to_string(i+1), mul_mat_vec_iq3_xxs_f16_f32_len, mul_mat_vec_iq3_xxs_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ3_S][i], "mul_mat_vec_iq3_s_f16_f32_"+std::to_string(i+1), mul_mat_vec_iq3_s_f16_f32_len, mul_mat_vec_iq3_s_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ4_XS][i], "mul_mat_vec_iq4_xs_f16_f32_"+std::to_string(i+1), mul_mat_vec_iq4_xs_f16_f32_len, mul_mat_vec_iq4_xs_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ4_NL][i], "mul_mat_vec_iq4_nl_f16_f32_"+std::to_string(i+1), mul_mat_vec_iq4_nl_f16_f32_len, mul_mat_vec_iq4_nl_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq, i+1}, 1, true); } ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_F32 ], "mul_mat_vec_id_f32_f32", mul_mat_vec_id_f32_f32_len, mul_mat_vec_id_f32_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1); @@ -2002,13 +2315,15 @@ static void ggml_vk_load_shaders(vk_device& device) { ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q4_K], "mul_mat_vec_id_q4_k_f32", mul_mat_vec_id_q4_k_f32_len, mul_mat_vec_id_q4_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq}, 1, true); ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q5_K], "mul_mat_vec_id_q5_k_f32", mul_mat_vec_id_q5_k_f32_len, mul_mat_vec_id_q5_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq}, 1, true); ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q6_K], "mul_mat_vec_id_q6_k_f32", mul_mat_vec_id_q6_k_f32_len, mul_mat_vec_id_q6_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq}, 1, true); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_IQ2_XXS], "mul_mat_vec_id_iq2_xxs_f32", mul_mat_vec_id_iq2_xxs_f32_len, mul_mat_vec_id_iq2_xxs_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq}, 1, true); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_IQ2_XS], "mul_mat_vec_id_iq2_xs_f32", mul_mat_vec_id_iq2_xs_f32_len, mul_mat_vec_id_iq2_xs_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq}, 1, true); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_IQ2_S], "mul_mat_vec_id_iq2_s_f32", mul_mat_vec_id_iq2_s_f32_len, mul_mat_vec_id_iq2_s_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq}, 1, true); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_IQ3_XXS], "mul_mat_vec_id_iq3_xxs_f32", mul_mat_vec_id_iq3_xxs_f32_len, mul_mat_vec_id_iq3_xxs_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq}, 1, true); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_IQ3_S], "mul_mat_vec_id_iq3_s_f32", mul_mat_vec_id_iq3_s_f32_len, mul_mat_vec_id_iq3_s_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq}, 1, true); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_IQ4_XS], "mul_mat_vec_id_iq4_xs_f32", mul_mat_vec_id_iq4_xs_f32_len, mul_mat_vec_id_iq4_xs_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq}, 1, true); - ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_IQ4_NL], "mul_mat_vec_id_iq4_nl_f32", mul_mat_vec_id_iq4_nl_f32_len, mul_mat_vec_id_iq4_nl_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2*rm_stdq, 1, 1}, {subgroup_size_16, 2*rm_stdq}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_IQ1_S], "mul_mat_vec_id_iq1_s_f32", mul_mat_vec_id_iq1_s_f32_len, mul_mat_vec_id_iq1_s_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_IQ1_M], "mul_mat_vec_id_iq1_m_f32", mul_mat_vec_id_iq1_m_f32_len, mul_mat_vec_id_iq1_m_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_IQ2_XXS], "mul_mat_vec_id_iq2_xxs_f32", mul_mat_vec_id_iq2_xxs_f32_len, mul_mat_vec_id_iq2_xxs_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_IQ2_XS], "mul_mat_vec_id_iq2_xs_f32", mul_mat_vec_id_iq2_xs_f32_len, mul_mat_vec_id_iq2_xs_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_IQ2_S], "mul_mat_vec_id_iq2_s_f32", mul_mat_vec_id_iq2_s_f32_len, mul_mat_vec_id_iq2_s_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_IQ3_XXS], "mul_mat_vec_id_iq3_xxs_f32", mul_mat_vec_id_iq3_xxs_f32_len, mul_mat_vec_id_iq3_xxs_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_IQ3_S], "mul_mat_vec_id_iq3_s_f32", mul_mat_vec_id_iq3_s_f32_len, mul_mat_vec_id_iq3_s_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_IQ4_XS], "mul_mat_vec_id_iq4_xs_f32", mul_mat_vec_id_iq4_xs_f32_len, mul_mat_vec_id_iq4_xs_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_IQ4_NL], "mul_mat_vec_id_iq4_nl_f32", mul_mat_vec_id_iq4_nl_f32_len, mul_mat_vec_id_iq4_nl_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {rm_iq, 1, 1}, {subgroup_size_16, rm_iq}, 1, true); // dequant shaders ggml_vk_create_pipeline(device, device->pipeline_dequant[GGML_TYPE_F32 ], "f32_to_f16", dequant_f32_len, dequant_f32_data, "main", 2, 5 * sizeof(uint32_t), {256 * 16, 1, 1}, {}, 1); @@ -2022,6 +2337,8 @@ static void ggml_vk_load_shaders(vk_device& device) { ggml_vk_create_pipeline(device, device->pipeline_dequant[GGML_TYPE_Q4_K], "dequant_q4_k", dequant_q4_k_len, dequant_q4_k_data, "main", 2, 5 * sizeof(uint32_t), {256 * 32, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_dequant[GGML_TYPE_Q5_K], "dequant_q5_k", dequant_q5_k_len, dequant_q5_k_data, "main", 2, 5 * sizeof(uint32_t), {256 * 64, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_dequant[GGML_TYPE_Q6_K], "dequant_q6_k", dequant_q6_k_len, dequant_q6_k_data, "main", 2, 5 * sizeof(uint32_t), {256 * 64, 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_dequant[GGML_TYPE_IQ1_S], "dequant_iq1_s", dequant_iq1_s_len, dequant_iq1_s_data, "main", 2, 5 * sizeof(uint32_t), {256 * 32, 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_dequant[GGML_TYPE_IQ1_M], "dequant_iq1_m", dequant_iq1_m_len, dequant_iq1_m_data, "main", 2, 5 * sizeof(uint32_t), {256 * 32, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_dequant[GGML_TYPE_IQ2_XXS], "dequant_iq2_xxs", dequant_iq2_xxs_len, dequant_iq2_xxs_data, "main", 2, 5 * sizeof(uint32_t), {256 * 32, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_dequant[GGML_TYPE_IQ2_XS], "dequant_iq2_xs", dequant_iq2_xs_len, dequant_iq2_xs_data, "main", 2, 5 * sizeof(uint32_t), {256 * 32, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_dequant[GGML_TYPE_IQ2_S], "dequant_iq2_s", dequant_iq2_s_len, dequant_iq2_s_data, "main", 2, 5 * sizeof(uint32_t), {256 * 32, 1, 1}, {}, 1); @@ -2038,6 +2355,8 @@ static void ggml_vk_load_shaders(vk_device& device) { ggml_vk_create_pipeline(device, device->pipeline_get_rows[GGML_TYPE_Q5_0], "get_rows_q5_0", get_rows_q5_0_len, get_rows_q5_0_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_get_rows[GGML_TYPE_Q5_1], "get_rows_q5_1", get_rows_q5_1_len, get_rows_q5_1_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_get_rows[GGML_TYPE_Q8_0], "get_rows_q8_0", get_rows_q8_0_len, get_rows_q8_0_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_get_rows[GGML_TYPE_IQ1_S], "get_rows_iq1_s", get_rows_iq1_s_len, get_rows_iq1_s_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_get_rows[GGML_TYPE_IQ1_M], "get_rows_iq1_m", get_rows_iq1_m_len, get_rows_iq1_m_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_get_rows[GGML_TYPE_IQ2_XXS], "get_rows_iq2_xxs", get_rows_iq2_xxs_len, get_rows_iq2_xxs_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_get_rows[GGML_TYPE_IQ2_XS], "get_rows_iq2_xs", get_rows_iq2_xs_len, get_rows_iq2_xs_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_get_rows[GGML_TYPE_IQ2_S], "get_rows_iq2_s", get_rows_iq2_s_len, get_rows_iq2_s_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1); @@ -2053,6 +2372,8 @@ static void ggml_vk_load_shaders(vk_device& device) { ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_Q5_0], "get_rows_q5_0_f32", get_rows_q5_0_f32_len, get_rows_q5_0_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_Q5_1], "get_rows_q5_1_f32", get_rows_q5_1_f32_len, get_rows_q5_1_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_Q8_0], "get_rows_q8_0_f32", get_rows_q8_0_f32_len, get_rows_q8_0_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_IQ1_S], "get_rows_iq1_s_f32", get_rows_iq1_s_f32_len, get_rows_iq1_s_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_IQ1_M], "get_rows_iq1_m_f32", get_rows_iq1_m_f32_len, get_rows_iq1_m_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_IQ2_XXS], "get_rows_iq2_xxs_f32", get_rows_iq2_xxs_f32_len, get_rows_iq2_xxs_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_IQ2_XS], "get_rows_iq2_xs_f32", get_rows_iq2_xs_f32_len, get_rows_iq2_xs_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_IQ2_S], "get_rows_iq2_s_f32", get_rows_iq2_s_f32_len, get_rows_iq2_s_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1); @@ -2062,13 +2383,23 @@ static void ggml_vk_load_shaders(vk_device& device) { ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_IQ4_NL], "get_rows_iq4_nl_f32", get_rows_iq4_nl_f32_len, get_rows_iq4_nl_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_matmul_split_k_reduce, "split_k_reduce", split_k_reduce_len, split_k_reduce_data, "main", 2, 2 * sizeof(uint32_t), {256 * 4, 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_flash_attn_split_k_reduce, "fa_split_k_reduce", fa_split_k_reduce_len, fa_split_k_reduce_data, "main", 2, 3 * sizeof(uint32_t), {1, 1, 1}, {}, 1, true); + ggml_vk_create_pipeline(device, device->pipeline_quantize_q8_1, "quantize_q8_1", quantize_q8_1_len, quantize_q8_1_data, "main", 2, 1 * sizeof(uint32_t), {32 * device->subgroup_size / 8, 1, 1}, { device->subgroup_size }, 1); - ggml_vk_create_pipeline(device, device->pipeline_mul_mat_vec_p021_f16_f32, "mul_mat_vec_p021_f16_f32", mul_mat_vec_p021_f16_f32_len, mul_mat_vec_p021_f16_f32_data, "main", 3, 6 * sizeof(uint32_t), {1, 1, 1}, {}, 1); + for (uint32_t i = 0; i < p021_max_gqa_ratio; ++i) { + if (device->subgroup_add && device->subgroup_require_full_support) { + ggml_vk_create_pipeline(device, device->pipeline_mul_mat_vec_p021_f16_f32[i], "mul_mat_vec_p021_f16_f32"+std::to_string(i+1), mul_mat_vec_p021_f16_f32_subgroup_add_len, mul_mat_vec_p021_f16_f32_subgroup_add_data, "main", 3, 6 * sizeof(uint32_t), {1, 1, 1}, {device->subgroup_size, i + 1}, 1, true, true); + } else { + ggml_vk_create_pipeline(device, device->pipeline_mul_mat_vec_p021_f16_f32[i], "mul_mat_vec_p021_f16_f32"+std::to_string(i+1), mul_mat_vec_p021_f16_f32_len, mul_mat_vec_p021_f16_f32_data, "main", 3, 6 * sizeof(uint32_t), {1, 1, 1}, {device->subgroup_size, i + 1}, 1, true); + } + } ggml_vk_create_pipeline(device, device->pipeline_mul_mat_vec_nc_f16_f32, "mul_mat_vec_nc_f16_f32", mul_mat_vec_nc_f16_f32_len, mul_mat_vec_nc_f16_f32_data, "main", 3, 7 * sizeof(uint32_t), {1, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_norm_f32, "norm_f32", norm_f32_len, norm_f32_data, "main", 2, sizeof(vk_op_push_constants), {1, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_group_norm_f32, "group_norm_f32", group_norm_f32_len, group_norm_f32_data, "main", 2, sizeof(vk_op_push_constants), {1, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_rms_norm_f32, "rms_norm_f32", rms_norm_f32_len, rms_norm_f32_data, "main", 2, sizeof(vk_op_push_constants), {1, 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_rms_norm_back_f32, "rms_norm_back_f32", rms_norm_back_f32_len, rms_norm_back_f32_data, "main", 3, sizeof(vk_op_push_constants), {1, 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_l2_norm_f32, "l2_norm_f32", l2_norm_f32_len, l2_norm_f32_data, "main", 2, sizeof(vk_op_push_constants), {1, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_f32, "cpy_f32_f32", cpy_f32_f32_len, cpy_f32_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_f16, "cpy_f32_f16", cpy_f32_f16_len, cpy_f32_f16_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1); @@ -2077,13 +2408,21 @@ static void ggml_vk_load_shaders(vk_device& device) { ggml_vk_create_pipeline(device, device->pipeline_contig_cpy_f32_f32, "contig_cpy_f32_f32", contig_cpy_f32_f32_len, contig_cpy_f32_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_contig_cpy_f32_f16, "contig_cpy_f32_f16", contig_cpy_f32_f16_len, contig_cpy_f32_f16_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_contig_cpy_f16_f16, "contig_cpy_f16_f16", contig_cpy_f16_f16_len, contig_cpy_f16_f16_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1); - - ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q4_0], "cpy_f32_q4_0", cpy_f32_q4_0_len, cpy_f32_q4_0_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q4_0), 1, 1}, {}, 1); - ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q4_1], "cpy_f32_q4_1", cpy_f32_q4_1_len, cpy_f32_q4_1_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q4_1), 1, 1}, {}, 1); - ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q5_0], "cpy_f32_q5_0", cpy_f32_q5_0_len, cpy_f32_q5_0_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q5_0), 1, 1}, {}, 1); - ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q5_1], "cpy_f32_q5_1", cpy_f32_q5_1_len, cpy_f32_q5_1_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q5_1), 1, 1}, {}, 1); - ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q8_0], "cpy_f32_q8_0", cpy_f32_q8_0_len, cpy_f32_q8_0_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q8_0), 1, 1}, {}, 1); - ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_IQ4_NL], "cpy_f32_iq4_nl", cpy_f32_iq4_nl_len, cpy_f32_iq4_nl_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_IQ4_NL), 1, 1}, {}, 1); + if (device->float_controls_rte_fp16) { + ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q4_0], "cpy_f32_q4_0", cpy_f32_q4_0_rte_len, cpy_f32_q4_0_rte_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q4_0), 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q4_1], "cpy_f32_q4_1", cpy_f32_q4_1_rte_len, cpy_f32_q4_1_rte_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q4_1), 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q5_0], "cpy_f32_q5_0", cpy_f32_q5_0_rte_len, cpy_f32_q5_0_rte_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q5_0), 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q5_1], "cpy_f32_q5_1", cpy_f32_q5_1_rte_len, cpy_f32_q5_1_rte_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q5_1), 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q8_0], "cpy_f32_q8_0", cpy_f32_q8_0_rte_len, cpy_f32_q8_0_rte_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q8_0), 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_IQ4_NL], "cpy_f32_iq4_nl", cpy_f32_iq4_nl_rte_len, cpy_f32_iq4_nl_rte_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_IQ4_NL), 1, 1}, {}, 1); + } else { + ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q4_0], "cpy_f32_q4_0", cpy_f32_q4_0_len, cpy_f32_q4_0_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q4_0), 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q4_1], "cpy_f32_q4_1", cpy_f32_q4_1_len, cpy_f32_q4_1_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q4_1), 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q5_0], "cpy_f32_q5_0", cpy_f32_q5_0_len, cpy_f32_q5_0_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q5_0), 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q5_1], "cpy_f32_q5_1", cpy_f32_q5_1_len, cpy_f32_q5_1_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q5_1), 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q8_0], "cpy_f32_q8_0", cpy_f32_q8_0_len, cpy_f32_q8_0_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q8_0), 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_IQ4_NL], "cpy_f32_iq4_nl", cpy_f32_iq4_nl_len, cpy_f32_iq4_nl_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_IQ4_NL), 1, 1}, {}, 1); + } ggml_vk_create_pipeline(device, device->pipeline_cpy_quant_f32[GGML_TYPE_Q4_0], "cpy_q4_0_f32", cpy_q4_0_f32_len, cpy_q4_0_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q4_0), 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_cpy_quant_f32[GGML_TYPE_Q4_1], "cpy_q4_1_f32", cpy_q4_1_f32_len, cpy_q4_1_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q4_1), 1, 1}, {}, 1); @@ -2099,6 +2438,8 @@ static void ggml_vk_load_shaders(vk_device& device) { ggml_vk_create_pipeline(device, device->pipeline_acc_f32, "acc_f32", acc_f32_len, acc_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_sub_f32, "sub_f32", sub_f32_len, sub_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {0}, 1); + ggml_vk_create_pipeline(device, device->pipeline_sub_f32_norepeat, "sub_f32_norepeat", sub_f32_len, sub_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {1}, 1); ggml_vk_create_pipeline(device, device->pipeline_mul_f32, "mul_f32", mul_f32_len, mul_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {0}, 1); ggml_vk_create_pipeline(device, device->pipeline_mul_f32_norepeat, "mul_f32_norepeat", mul_f32_len, mul_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {1}, 1); ggml_vk_create_pipeline(device, device->pipeline_div_f32, "div_f32", div_f32_len, div_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {0}, 1); @@ -2121,13 +2462,16 @@ static void ggml_vk_load_shaders(vk_device& device) { ggml_vk_create_pipeline(device, device->pipeline_pad_f32, "pad_f32", pad_f32_len, pad_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_repeat_f32, "repeat_f32", repeat_f32_len, repeat_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_repeat_back_f32, "repeat_back_f32", repeat_back_f32_len, repeat_back_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_gelu_f32, "gelu_f32", gelu_f32_len, gelu_f32_data, "main", 2, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_gelu_quick_f32, "gelu_quick_f32", gelu_quick_f32_len, gelu_quick_f32_data, "main", 2, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_silu_f32, "silu_f32", silu_f32_len, silu_f32_data, "main", 2, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_silu_back_f32, "silu_back_f32", silu_back_f32_len, silu_back_f32_data, "main", 3, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_relu_f32, "relu_f32", relu_f32_len, relu_f32_data, "main", 2, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_leaky_relu_f32, "leaky_relu_f32", leaky_relu_f32_len, leaky_relu_f32_data, "main", 2, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_tanh_f32, "tanh_f32", tanh_f32_len, tanh_f32_data, "main", 2, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_sigmoid_f32, "sigmoid_f32", sigmoid_f32_len, sigmoid_f32_data, "main", 2, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_diag_mask_inf_f32, "diag_mask_inf_f32", diag_mask_inf_f32_len, diag_mask_inf_f32_data, "main", 2, sizeof(vk_op_diag_mask_push_constants), {1, 512, 1}, {}, 1, true); @@ -2135,22 +2479,33 @@ static void ggml_vk_load_shaders(vk_device& device) { ggml_vk_create_pipeline(device, device->pipeline_soft_max_f32_wg512, "soft_max_f32_wg512", soft_max_f32_len, soft_max_f32_data, "main", 3, sizeof(vk_op_soft_max_push_constants), {1, 1, 1}, { 512 }, 1); ggml_vk_create_pipeline(device, device->pipeline_soft_max_f32_f16, "soft_max_f32_f16", soft_max_f32_f16_len, soft_max_f32_f16_data, "main", 3, sizeof(vk_op_soft_max_push_constants), {1, 1, 1}, { device->subgroup_size }, 1); ggml_vk_create_pipeline(device, device->pipeline_soft_max_f32_f16_wg512, "soft_max_f32_f16_wg512", soft_max_f32_f16_len, soft_max_f32_f16_data, "main", 3, sizeof(vk_op_soft_max_push_constants), {1, 1, 1}, { 512 }, 1); + ggml_vk_create_pipeline(device, device->pipeline_soft_max_back_f32, "soft_max_back_f32", soft_max_back_f32_len, soft_max_back_f32_data, "main", 3, sizeof(vk_op_push_constants), {1, 1, 1}, { device->subgroup_size }, 1); ggml_vk_create_pipeline(device, device->pipeline_rope_norm_f32, "rope_norm_f32", rope_norm_f32_len, rope_norm_f32_data, "main", 4, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_rope_neox_f32, "rope_neox_f32", rope_neox_f32_len, rope_neox_f32_data, "main", 4, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_rope_multi_f32, "rope_multi_f32", rope_multi_f32_len, rope_multi_f32_data, "main", 4, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_rope_vision_f32, "rope_vision_f32", rope_vision_f32_len, rope_vision_f32_data, "main", 4, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1); if (device->float_controls_rte_fp16) { ggml_vk_create_pipeline(device, device->pipeline_rope_norm_f16, "rope_norm_f16", rope_norm_f16_rte_len, rope_norm_f16_rte_data, "main", 4, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_rope_neox_f16, "rope_neox_f16", rope_neox_f16_rte_len, rope_neox_f16_rte_data, "main", 4, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_rope_multi_f16, "rope_multi_f16", rope_multi_f16_rte_len, rope_multi_f16_rte_data, "main", 4, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_rope_vision_f16, "rope_vision_f16", rope_vision_f16_rte_len, rope_vision_f16_rte_data, "main", 4, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1); } else { ggml_vk_create_pipeline(device, device->pipeline_rope_norm_f16, "rope_norm_f16", rope_norm_f16_len, rope_norm_f16_data, "main", 4, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_rope_neox_f16, "rope_neox_f16", rope_neox_f16_len, rope_neox_f16_data, "main", 4, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_rope_multi_f16, "rope_multi_f16", rope_multi_f16_len, rope_multi_f16_data, "main", 4, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_rope_vision_f16, "rope_vision_f16", rope_vision_f16_len, rope_vision_f16_data, "main", 4, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1); } ggml_vk_create_pipeline(device, device->pipeline_argsort_f32, "argsort_f32", argsort_f32_len, argsort_f32_data, "main", 2, sizeof(vk_op_argsort_push_constants), {1024, 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_argmax_f32, "argmax_f32", argmax_f32_len, argmax_f32_data, "main", 2, sizeof(vk_op_push_constants), {1, 1, 1}, { device->subgroup_size }, 1); + ggml_vk_create_pipeline(device, device->pipeline_sum_rows_f32, "sum_rows_f32", sum_rows_f32_len, sum_rows_f32_data, "main", 2, sizeof(vk_op_push_constants), {1, 1, 1}, { device->subgroup_size }, 1); + ggml_vk_create_pipeline(device, device->pipeline_count_equal_i32, "count_equal_i32", count_equal_i32_len, count_equal_i32_data, "main", 3, sizeof(vk_op_push_constants), {512, 1, 1}, { device->subgroup_size }, 1); + ggml_vk_create_pipeline(device, device->pipeline_im2col_f32, "im2col_f32", im2col_f32_len, im2col_f32_data, "main", 2, sizeof(vk_op_im2col_push_constants), {512, 1, 1}, { device->subgroup_size }, 1, true); if (device->float_controls_rte_fp16) { ggml_vk_create_pipeline(device, device->pipeline_im2col_f32_f16, "im2col_f32_f16", im2col_f32_f16_rte_len, im2col_f32_f16_rte_data, "main", 2, sizeof(vk_op_im2col_push_constants), {512, 1, 1}, { device->subgroup_size }, 1, true); @@ -2164,13 +2519,17 @@ static void ggml_vk_load_shaders(vk_device& device) { ggml_vk_create_pipeline(device, device->pipeline_rwkv_wkv6_f32, "rwkv_wkv6_f32", rwkv_wkv6_f32_len, rwkv_wkv6_f32_data, "main", 7, sizeof(vk_op_rwkv_wkv6_push_constants), {1, 1, 1}, {device->subgroup_size}, 1); + ggml_vk_create_pipeline(device, device->pipeline_rwkv_wkv7_f32, "rwkv_wkv7_f32", rwkv_wkv7_f32_len, rwkv_wkv7_f32_data, "main", 8, sizeof(vk_op_rwkv_wkv7_push_constants), {1, 1, 1}, {device->subgroup_size}, 1); + + ggml_vk_create_pipeline(device, device->pipeline_opt_step_adamw_f32, "opt_step_adamw_f32", opt_step_adamw_f32_len, opt_step_adamw_f32_data, "main", 5, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1); + for (auto &c : compiles) { c.wait(); } device->need_compiles = false; } -static bool ggml_vk_khr_cooperative_matrix_support(const vk::PhysicalDeviceProperties& props, const vk::PhysicalDeviceDriverProperties& driver_props); +static bool ggml_vk_khr_cooperative_matrix_support(const vk::PhysicalDeviceProperties& props, const vk::PhysicalDeviceDriverProperties& driver_props, vk_device_architecture arch); static vk_device ggml_vk_get_device(size_t idx) { VK_LOG_DEBUG("ggml_vk_get_device(" << idx << ")"); @@ -2199,6 +2558,11 @@ static vk_device ggml_vk_get_device(size_t idx) { device->physical_device = physical_devices[dev_num]; const std::vector ext_props = device->physical_device.enumerateDeviceExtensionProperties(); + device->architecture = get_device_architecture(device->physical_device); + + const char* GGML_VK_PREFER_HOST_MEMORY = getenv("GGML_VK_PREFER_HOST_MEMORY"); + device->prefer_host_memory = GGML_VK_PREFER_HOST_MEMORY != nullptr; + bool fp16_storage = false; bool fp16_compute = false; bool maintenance4_support = false; @@ -2207,8 +2571,8 @@ static vk_device ggml_vk_get_device(size_t idx) { bool pipeline_robustness = false; bool coopmat2_support = false; device->coopmat_support = false; + device->integer_dot_product = false; - // Check if maintenance4 is supported for (const auto& properties : ext_props) { if (strcmp("VK_KHR_maintenance4", properties.extensionName) == 0) { maintenance4_support = true; @@ -2233,6 +2597,11 @@ static vk_device ggml_vk_get_device(size_t idx) { } else if (strcmp("VK_NV_cooperative_matrix2", properties.extensionName) == 0 && !getenv("GGML_VK_DISABLE_COOPMAT2")) { coopmat2_support = true; +#if defined(GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT) + } else if (strcmp("VK_KHR_shader_integer_dot_product", properties.extensionName) == 0 && + !getenv("GGML_VK_DISABLE_INTEGER_DOT_PRODUCT")) { + device->integer_dot_product = true; +#endif } } @@ -2243,13 +2612,16 @@ static vk_device ggml_vk_get_device(size_t idx) { vk::PhysicalDeviceDriverProperties driver_props; vk::PhysicalDeviceShaderSMBuiltinsPropertiesNV sm_props; vk::PhysicalDeviceShaderCoreProperties2AMD amd_shader_core_properties2_props; + vk::PhysicalDeviceVulkan11Properties vk11_props; vk::PhysicalDeviceVulkan12Properties vk12_props; vk::PhysicalDeviceSubgroupSizeControlPropertiesEXT subgroup_size_control_props; + vk::PhysicalDeviceShaderIntegerDotProductPropertiesKHR shader_integer_dot_product_props; props2.pNext = &props3; props3.pNext = &subgroup_props; subgroup_props.pNext = &driver_props; - driver_props.pNext = &vk12_props; + driver_props.pNext = &vk11_props; + vk11_props.pNext = &vk12_props; VkBaseOutStructure * last_struct = (VkBaseOutStructure *)&vk12_props; @@ -2278,6 +2650,11 @@ static vk_device ggml_vk_get_device(size_t idx) { } #endif + if (device->integer_dot_product) { + last_struct->pNext = (VkBaseOutStructure *)&shader_integer_dot_product_props; + last_struct = (VkBaseOutStructure *)&shader_integer_dot_product_props; + } + device->physical_device.getProperties2(&props2); device->properties = props2.properties; device->vendor_id = device->properties.vendorID; @@ -2296,13 +2673,9 @@ static vk_device ggml_vk_get_device(size_t idx) { if (GGML_VK_SUBALLOCATION_BLOCK_SIZE != nullptr) { device->suballocation_block_size = std::stoul(GGML_VK_SUBALLOCATION_BLOCK_SIZE); -#if defined(_WIN32) - } else if (device->vendor_id == VK_VENDOR_ID_NVIDIA) { + } else { // Limit batching of allocations to 1GB by default to avoid fragmentation issues device->suballocation_block_size = 1024*1024*1024; -#endif - } else { - device->suballocation_block_size = device->max_memory_allocation_size; } device->suballocation_block_size = std::min(device->suballocation_block_size, device->max_memory_allocation_size); @@ -2317,14 +2690,19 @@ static vk_device ggml_vk_get_device(size_t idx) { } device->float_controls_rte_fp16 = vk12_props.shaderRoundingModeRTEFloat16; + device->subgroup_add = (vk11_props.subgroupSupportedStages & vk::ShaderStageFlagBits::eCompute) && + (vk11_props.subgroupSupportedOperations & vk::SubgroupFeatureFlagBits::eArithmetic); + const bool force_disable_f16 = getenv("GGML_VK_DISABLE_F16") != nullptr; device->fp16 = !force_disable_f16 && fp16_storage && fp16_compute; - if (!ggml_vk_khr_cooperative_matrix_support(device->properties, driver_props)) { + if (!ggml_vk_khr_cooperative_matrix_support(device->properties, driver_props, device->architecture)) { device->coopmat_support = false; } + device->integer_dot_product = device->integer_dot_product && shader_integer_dot_product_props.integerDotProduct4x8BitPackedSignedAccelerated; + std::vector queue_family_props = device->physical_device.getQueueFamilyProperties(); // Try to find a non-graphics compute queue and transfer-focused queues @@ -2417,6 +2795,14 @@ static vk_device ggml_vk_get_device(size_t idx) { device_extensions.push_back("VK_KHR_maintenance4"); } + VkPhysicalDeviceShaderIntegerDotProductFeaturesKHR shader_integer_dot_product_features {}; + shader_integer_dot_product_features.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_INTEGER_DOT_PRODUCT_FEATURES_KHR; + if (device->integer_dot_product) { + last_struct->pNext = (VkBaseOutStructure *)&shader_integer_dot_product_features; + last_struct = (VkBaseOutStructure *)&shader_integer_dot_product_features; + device_extensions.push_back("VK_KHR_shader_integer_dot_product"); + } + vkGetPhysicalDeviceFeatures2(device->physical_device, &device_features2); device->fp16 = device->fp16 && vk12_features.shaderFloat16; @@ -2586,6 +2972,17 @@ static vk_device ggml_vk_get_device(size_t idx) { device->coopmat_acc_f16_support = true; } } + } else if ((vk::ComponentTypeKHR)prop.AType == vk::ComponentTypeKHR::eSint8 && + (vk::ComponentTypeKHR)prop.BType == vk::ComponentTypeKHR::eSint8 && + (vk::ComponentTypeKHR)prop.CType == vk::ComponentTypeKHR::eSint32 && + (vk::ComponentTypeKHR)prop.ResultType == vk::ComponentTypeKHR::eSint32 && + (vk::ScopeKHR)prop.scope == vk::ScopeKHR::eSubgroup && + device->coopmat_int_m == 0 + ) { + device->coopmat_int_support = true; + device->coopmat_int_m = prop.MSize; + device->coopmat_int_n = prop.NSize; + device->coopmat_int_k = prop.KSize; } } @@ -2690,22 +3087,11 @@ static void ggml_vk_print_gpu_info(size_t idx) { vk::PhysicalDevice physical_device = devices[dev_num]; std::vector ext_props = physical_device.enumerateDeviceExtensionProperties(); - vk::PhysicalDeviceProperties2 props2; - vk::PhysicalDeviceMaintenance3Properties props3; - vk::PhysicalDeviceSubgroupProperties subgroup_props; - vk::PhysicalDeviceDriverProperties driver_props; - props2.pNext = &props3; - props3.pNext = &subgroup_props; - subgroup_props.pNext = &driver_props; - physical_device.getProperties2(&props2); - - const size_t subgroup_size = subgroup_props.subgroupSize; - const bool uma = props2.properties.deviceType == vk::PhysicalDeviceType::eIntegratedGpu; - bool fp16_storage = false; bool fp16_compute = false; bool coopmat_support = false; bool coopmat2_support = false; + bool integer_dot_product = false; for (auto properties : ext_props) { if (strcmp("VK_KHR_16bit_storage", properties.extensionName) == 0) { @@ -2721,25 +3107,44 @@ static void ggml_vk_print_gpu_info(size_t idx) { } else if (strcmp("VK_NV_cooperative_matrix2", properties.extensionName) == 0 && !getenv("GGML_VK_DISABLE_COOPMAT2")) { coopmat2_support = true; +#endif +#if defined(GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT) + } else if (strcmp("VK_KHR_shader_integer_dot_product", properties.extensionName) == 0 && + !getenv("GGML_VK_DISABLE_INTEGER_DOT_PRODUCT")) { + integer_dot_product = true; #endif } } - if (!ggml_vk_khr_cooperative_matrix_support(props2.properties, driver_props)) { - coopmat_support = false; - } + const vk_device_architecture device_architecture = get_device_architecture(physical_device); const char* GGML_VK_DISABLE_F16 = getenv("GGML_VK_DISABLE_F16"); bool force_disable_f16 = GGML_VK_DISABLE_F16 != nullptr; bool fp16 = !force_disable_f16 && fp16_storage && fp16_compute; - vk::PhysicalDeviceFeatures device_features = physical_device.getFeatures(); + vk::PhysicalDeviceProperties2 props2; + vk::PhysicalDeviceMaintenance3Properties props3; + vk::PhysicalDeviceSubgroupProperties subgroup_props; + vk::PhysicalDeviceDriverProperties driver_props; + vk::PhysicalDeviceShaderIntegerDotProductPropertiesKHR shader_integer_dot_product_props; + props2.pNext = &props3; + props3.pNext = &subgroup_props; + subgroup_props.pNext = &driver_props; + + // Pointer to the last chain element + VkBaseOutStructure * last_struct = (VkBaseOutStructure *)&driver_props; + + if (integer_dot_product) { + last_struct->pNext = (VkBaseOutStructure *)&shader_integer_dot_product_props; + last_struct = (VkBaseOutStructure *)&shader_integer_dot_product_props; + } + + physical_device.getProperties2(&props2); VkPhysicalDeviceFeatures2 device_features2; device_features2.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FEATURES_2; device_features2.pNext = nullptr; - device_features2.features = (VkPhysicalDeviceFeatures)device_features; VkPhysicalDeviceVulkan11Features vk11_features; vk11_features.pNext = nullptr; @@ -2752,7 +3157,7 @@ static void ggml_vk_print_gpu_info(size_t idx) { vk11_features.pNext = &vk12_features; // Pointer to the last chain element - VkBaseOutStructure * last_struct = (VkBaseOutStructure *)&vk12_features; + last_struct = (VkBaseOutStructure *)&vk12_features; #if defined(GGML_VULKAN_COOPMAT_GLSLC_SUPPORT) VkPhysicalDeviceCooperativeMatrixFeaturesKHR coopmat_features; @@ -2764,20 +3169,39 @@ static void ggml_vk_print_gpu_info(size_t idx) { last_struct->pNext = (VkBaseOutStructure *)&coopmat_features; last_struct = (VkBaseOutStructure *)&coopmat_features; } +#endif + + VkPhysicalDeviceShaderIntegerDotProductFeaturesKHR shader_integer_dot_product_features {}; + shader_integer_dot_product_features.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_INTEGER_DOT_PRODUCT_FEATURES_KHR; + if (integer_dot_product) { + last_struct->pNext = (VkBaseOutStructure *)&shader_integer_dot_product_features; + last_struct = (VkBaseOutStructure *)&shader_integer_dot_product_features; + } vkGetPhysicalDeviceFeatures2(physical_device, &device_features2); fp16 = fp16 && vk12_features.shaderFloat16; - coopmat_support = coopmat_support && coopmat_features.cooperativeMatrix; + uint32_t default_subgroup_size = get_subgroup_size("", device_architecture); + const size_t subgroup_size = (default_subgroup_size != 0) ? default_subgroup_size : subgroup_props.subgroupSize; + const bool uma = props2.properties.deviceType == vk::PhysicalDeviceType::eIntegratedGpu; + + integer_dot_product = integer_dot_product + && shader_integer_dot_product_props.integerDotProduct4x8BitPackedSignedAccelerated + && shader_integer_dot_product_features.shaderIntegerDotProduct; + + coopmat_support = coopmat_support +#if defined(GGML_VULKAN_COOPMAT_GLSLC_SUPPORT) + && coopmat_features.cooperativeMatrix #endif + && ggml_vk_khr_cooperative_matrix_support(props2.properties, driver_props, device_architecture); std::string matrix_cores = coopmat2_support ? "NV_coopmat2" : coopmat_support ? "KHR_coopmat" : "none"; std::string device_name = props2.properties.deviceName.data(); - GGML_LOG_DEBUG("ggml_vulkan: %zu = %s (%s) | uma: %d | fp16: %d | warp size: %zu | shared memory: %d | matrix cores: %s\n", + GGML_LOG_DEBUG("ggml_vulkan: %zu = %s (%s) | uma: %d | fp16: %d | warp size: %zu | shared memory: %d | int dot: %d | matrix cores: %s\n", idx, device_name.c_str(), driver_props.driverName.data(), uma, fp16, subgroup_size, - props2.properties.limits.maxComputeSharedMemorySize, matrix_cores.c_str()); + props2.properties.limits.maxComputeSharedMemorySize, integer_dot_product, matrix_cores.c_str()); if (props2.properties.deviceType == vk::PhysicalDeviceType::eCpu) { GGML_LOG_DEBUG("ggml_vulkan: Warning: Device type is CPU. This is probably not the device you want.\n"); @@ -2787,14 +3211,12 @@ static void ggml_vk_print_gpu_info(size_t idx) { static bool ggml_vk_instance_validation_ext_available(const std::vector& instance_extensions); static bool ggml_vk_instance_portability_enumeration_ext_available(const std::vector& instance_extensions); -void ggml_vk_instance_init() { +static void ggml_vk_instance_init() { if (vk_instance_initialized) { return; } VK_LOG_DEBUG("ggml_vk_instance_init()"); - vk_instance_initialized = true; - uint32_t api_version = vk::enumerateInstanceVersion(); if (api_version < VK_API_VERSION_1_2) { @@ -2845,6 +3267,7 @@ void ggml_vk_instance_init() { GGML_LOG_DEBUG("ggml_vulkan: Validation layers enabled\n"); } vk_instance.instance = vk::createInstance(instance_create_info); + vk_instance_initialized = true; size_t num_available_devices = vk_instance.instance.enumeratePhysicalDevices().size(); @@ -2869,7 +3292,7 @@ void ggml_vk_instance_init() { // Make sure at least one device exists if (devices.empty()) { std::cerr << "ggml_vulkan: Error: No devices found." << std::endl; - GGML_ABORT("fatal error"); + return; } // Default to using all dedicated GPUs @@ -2980,6 +3403,7 @@ static void ggml_vk_init(ggml_backend_vk_context * ctx, size_t idx) { ctx->prealloc_size_split_k = 0; ctx->fence = ctx->device->device.createFence({}); + ctx->almost_ready_fence = ctx->device->device.createFence({}); #ifdef GGML_VULKAN_CHECK_RESULTS const char* skip_checks = getenv("GGML_VULKAN_SKIP_CHECKS"); @@ -3003,6 +3427,8 @@ static vk_pipeline ggml_vk_get_to_fp16(ggml_backend_vk_context * ctx, ggml_type case GGML_TYPE_Q4_K: case GGML_TYPE_Q5_K: case GGML_TYPE_Q6_K: + case GGML_TYPE_IQ1_S: + case GGML_TYPE_IQ1_M: case GGML_TYPE_IQ2_XXS: case GGML_TYPE_IQ2_XS: case GGML_TYPE_IQ2_S: @@ -3042,6 +3468,17 @@ static vk_matmul_pipeline ggml_vk_get_mul_mat_mat_pipeline(ggml_backend_vk_conte } } + // MMQ + if (src1_type == GGML_TYPE_Q8_1) { + vk_matmul_pipeline pipelines = ctx->device->pipeline_dequant_mul_mat_mat_q8_1[src0_type].f16acc; + + if (pipelines->s == nullptr && pipelines->m == nullptr && pipelines->l == nullptr) { + return nullptr; + } + + return pipelines; + } + if (src1_type != GGML_TYPE_F32 && !ctx->device->coopmat2) { return nullptr; } @@ -3057,6 +3494,8 @@ static vk_matmul_pipeline ggml_vk_get_mul_mat_mat_pipeline(ggml_backend_vk_conte case GGML_TYPE_Q4_K: case GGML_TYPE_Q5_K: case GGML_TYPE_Q6_K: + case GGML_TYPE_IQ1_S: + case GGML_TYPE_IQ1_M: case GGML_TYPE_IQ2_XXS: case GGML_TYPE_IQ2_XS: case GGML_TYPE_IQ2_S: @@ -3094,6 +3533,8 @@ static vk_pipeline ggml_vk_get_dequantize_mul_mat_vec(ggml_backend_vk_context * case GGML_TYPE_Q4_K: case GGML_TYPE_Q5_K: case GGML_TYPE_Q6_K: + case GGML_TYPE_IQ1_S: + case GGML_TYPE_IQ1_M: case GGML_TYPE_IQ2_XXS: case GGML_TYPE_IQ2_XS: case GGML_TYPE_IQ2_S: @@ -3143,6 +3584,8 @@ static vk_matmul_pipeline ggml_vk_get_mul_mat_mat_id_pipeline(ggml_backend_vk_co case GGML_TYPE_Q4_K: case GGML_TYPE_Q5_K: case GGML_TYPE_Q6_K: + case GGML_TYPE_IQ1_S: + case GGML_TYPE_IQ1_M: case GGML_TYPE_IQ2_XXS: case GGML_TYPE_IQ2_XS: case GGML_TYPE_IQ2_S: @@ -3175,6 +3618,8 @@ static vk_pipeline ggml_vk_get_dequantize_mul_mat_vec_id(ggml_backend_vk_context case GGML_TYPE_Q4_K: case GGML_TYPE_Q5_K: case GGML_TYPE_Q6_K: + case GGML_TYPE_IQ1_S: + case GGML_TYPE_IQ1_M: case GGML_TYPE_IQ2_XXS: case GGML_TYPE_IQ2_XS: case GGML_TYPE_IQ2_S: @@ -3326,8 +3771,6 @@ static vk_submission ggml_vk_begin_submission(vk_device& device, vk_queue& q, bo return s; } - - static void ggml_vk_dispatch_pipeline(ggml_backend_vk_context* ctx, vk_context& subctx, vk_pipeline& pipeline, std::initializer_list const& descriptor_buffer_infos, size_t push_constant_size, const void* push_constants, std::array elements) { const uint32_t wg0 = CEIL_DIV(elements[0], pipeline->wg_denoms[0]); const uint32_t wg1 = CEIL_DIV(elements[1], pipeline->wg_denoms[1]); @@ -3717,6 +4160,12 @@ static void ggml_vk_buffer_copy(vk_buffer& dst, size_t dst_offset, vk_buffer& sr } } +static void ggml_vk_buffer_memset_async(vk_context& ctx, vk_buffer& dst, size_t offset, uint32_t c, size_t size) { + VK_LOG_DEBUG("ggml_vk_buffer_memset_async(" << offset << ", " << c << ", " << size << ")"); + + ctx->s->buffer.fillBuffer(dst->buffer, offset, size, c); +} + static void ggml_vk_buffer_memset(vk_buffer& dst, size_t offset, uint32_t c, size_t size) { VK_LOG_DEBUG("ggml_vk_buffer_memset(" << offset << ", " << c << ", " << size << ")"); @@ -3745,20 +4194,30 @@ static uint32_t ggml_vk_guess_split_k(ggml_backend_vk_context * ctx, int m, int if (split_k == 3) { split_k = 2; } + if (ctx->device->coopmat2) { + // coopmat2 shader expects splits to be aligned to 256 + while (split_k > 1 && ((k / split_k) % 256) != 0) { + split_k /= 2; + } + } } } return split_k; } -static vk_pipeline ggml_vk_guess_matmul_pipeline(ggml_backend_vk_context * ctx, vk_matmul_pipeline& mmp, int m, int n, bool aligned, ggml_type src0_type) { - VK_LOG_DEBUG("ggml_vk_guess_matmul_pipeline(" << m << ", " << n << ", " << aligned << ", " << ggml_type_name(src0_type) << ")"); +static vk_pipeline ggml_vk_guess_matmul_pipeline(ggml_backend_vk_context * ctx, vk_matmul_pipeline& mmp, uint32_t m, uint32_t n, bool aligned, ggml_type src0_type, ggml_type src1_type) { + VK_LOG_DEBUG("ggml_vk_guess_matmul_pipeline(" << m << ", " << n << ", " << aligned << ", " << ggml_type_name(src0_type) << ", " << ggml_type_name(src1_type) << ")"); if (ctx->device->coopmat2) { - if ((ctx->device->mul_mat_l[src0_type] && (m % mmp->l->wg_denoms[0]) == 0 && (n % mmp->l->wg_denoms[1]) == 0) || (!ctx->device->mul_mat_m[src0_type] && !ctx->device->mul_mat_s[src0_type])) { + // Use large shader when the N dimension is greater than the medium shader's tile size + uint32_t crossover_large = mmp->m->wg_denoms[1]; + if ((ctx->device->mul_mat_l[src0_type] && (n > crossover_large)) || (!ctx->device->mul_mat_m[src0_type] && !ctx->device->mul_mat_s[src0_type])) { return aligned ? mmp->a_l : mmp->l; } - if ((ctx->device->mul_mat_m[src0_type] && (m % mmp->m->wg_denoms[0]) == 0 && (n % mmp->m->wg_denoms[1]) == 0) || !ctx->device->mul_mat_s[src0_type]) { + // Use medium shader when the N dimension is greater than the small shader's tile size + uint32_t crossover_medium = mmp->s->wg_denoms[1]; + if ((ctx->device->mul_mat_m[src0_type] && (n > crossover_medium)) || !ctx->device->mul_mat_s[src0_type]) { return aligned ? mmp->a_m : mmp->m; } return aligned ? mmp->a_s : mmp->s; @@ -3773,9 +4232,9 @@ static vk_pipeline ggml_vk_guess_matmul_pipeline(ggml_backend_vk_context * ctx, return aligned ? mmp->a_l : mmp->l; } -static uint32_t ggml_vk_guess_matmul_pipeline_align(ggml_backend_vk_context * ctx, vk_matmul_pipeline& mmp, int m, int n, ggml_type src0_type) { - VK_LOG_DEBUG("ggml_vk_guess_matmul_pipeline_align(" << m << ", " << n << ", " << ggml_type_name(src0_type) << ")"); - return ggml_vk_guess_matmul_pipeline(ctx, mmp, m, n, true, src0_type)->align; +static uint32_t ggml_vk_guess_matmul_pipeline_align(ggml_backend_vk_context * ctx, vk_matmul_pipeline& mmp, int m, int n, ggml_type src0_type, ggml_type src1_type) { + VK_LOG_DEBUG("ggml_vk_guess_matmul_pipeline_align(" << m << ", " << n << ", " << ggml_type_name(src0_type) << ", " << ggml_type_name(src1_type) << ")"); + return ggml_vk_guess_matmul_pipeline(ctx, mmp, m, n, true, src0_type, src1_type)->align; } static void ggml_vk_matmul( @@ -3783,18 +4242,19 @@ static void ggml_vk_matmul( vk_subbuffer&& a, vk_subbuffer&& b, vk_subbuffer&& d, vk_subbuffer&& split_k_buffer, uint32_t m, uint32_t n, uint32_t k, uint32_t stride_a, uint32_t stride_b, uint32_t stride_d, uint32_t batch_stride_a, uint32_t batch_stride_b, uint32_t batch_stride_d, - uint32_t split_k, uint32_t batch, uint32_t ne02, uint32_t ne12, uint32_t broadcast2, uint32_t broadcast3) { - VK_LOG_DEBUG("ggml_vk_matmul(a: (" << a.buffer->buffer << ", " << a.offset << ", " << a.size << "), b: (" << b.buffer->buffer << ", " << b.offset << ", " << b.size << "), d: (" << d.buffer->buffer << ", " << d.offset << ", " << d.size << "), split_k: (" << (split_k_buffer.buffer != nullptr ? split_k_buffer.buffer->buffer : VK_NULL_HANDLE) << ", " << split_k_buffer.offset << ", " << split_k_buffer.size << "), m: " << m << ", n: " << n << ", k: " << k << ", stride_a: " << stride_a << ", stride_b: " << stride_b << ", stride_d: " << stride_d << ", batch_stride_a: " << batch_stride_a << ", batch_stride_b: " << batch_stride_b << ", batch_stride_d: " << batch_stride_d << ", split_k: " << split_k << ", batch: " << batch << ", ne02: " << ne02 << ", ne12: " << ne12 << ", broadcast2: " << broadcast2 << ", broadcast3: " << broadcast3 << ")"); + uint32_t split_k, uint32_t batch, uint32_t ne02, uint32_t ne12, uint32_t broadcast2, uint32_t broadcast3, + uint32_t padded_n) { + VK_LOG_DEBUG("ggml_vk_matmul(a: (" << a.buffer->buffer << ", " << a.offset << ", " << a.size << "), b: (" << b.buffer->buffer << ", " << b.offset << ", " << b.size << "), d: (" << d.buffer->buffer << ", " << d.offset << ", " << d.size << "), split_k: (" << (split_k_buffer.buffer != nullptr ? split_k_buffer.buffer->buffer : VK_NULL_HANDLE) << ", " << split_k_buffer.offset << ", " << split_k_buffer.size << "), m: " << m << ", n: " << n << ", k: " << k << ", stride_a: " << stride_a << ", stride_b: " << stride_b << ", stride_d: " << stride_d << ", batch_stride_a: " << batch_stride_a << ", batch_stride_b: " << batch_stride_b << ", batch_stride_d: " << batch_stride_d << ", split_k: " << split_k << ", batch: " << batch << ", ne02: " << ne02 << ", ne12: " << ne12 << ", broadcast2: " << broadcast2 << ", broadcast3: " << broadcast3 << ", padded_n: " << padded_n << ")"); ggml_vk_sync_buffers(subctx); if (split_k == 1) { - const vk_mat_mat_push_constants pc = { m, n, k, stride_a, stride_b, stride_d, batch_stride_a, batch_stride_b, batch_stride_d, k, ne02, ne12, broadcast2, broadcast3 }; + const vk_mat_mat_push_constants pc = { m, n, k, stride_a, stride_b, stride_d, batch_stride_a, batch_stride_b, batch_stride_d, k, ne02, ne12, broadcast2, broadcast3, padded_n }; ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { a, b, d }, sizeof(vk_mat_mat_push_constants), &pc, { m, n, batch }); return; } GGML_ASSERT(batch_stride_d == m * n); - const vk_mat_mat_push_constants pc1 = { m, n, k, stride_a, stride_b, stride_d, batch_stride_a, batch_stride_b, batch_stride_d, CEIL_DIV(k, split_k), ne02, ne12, broadcast2, broadcast3 }; + const vk_mat_mat_push_constants pc1 = { m, n, k, stride_a, stride_b, stride_d, batch_stride_a, batch_stride_b, batch_stride_d, CEIL_DIV(k, split_k), ne02, ne12, broadcast2, broadcast3, padded_n }; // Make sure enough workgroups get assigned for split k to work ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { a, b, split_k_buffer }, sizeof(vk_mat_mat_push_constants), &pc1, { (CEIL_DIV(m, pipeline->wg_denoms[0]) * pipeline->wg_denoms[0]) * split_k, n, batch }); ggml_vk_sync_buffers(subctx); @@ -3802,14 +4262,18 @@ static void ggml_vk_matmul( ggml_vk_dispatch_pipeline(ctx, subctx, ctx->device->pipeline_matmul_split_k_reduce, { split_k_buffer, d }, pc2.size() * sizeof(uint32_t), pc2.data(), { m * n * batch, 1, 1 }); } -static vk_pipeline ggml_vk_guess_matmul_id_pipeline(ggml_backend_vk_context * ctx, vk_matmul_pipeline& mmp, int m, int n, bool aligned, ggml_type src0_type) { - VK_LOG_DEBUG("ggml_vk_guess_matmul_pipeline(" << m << ", " << n << ", " << aligned << ", " << ggml_type_name(src0_type) << ")"); +static vk_pipeline ggml_vk_guess_matmul_id_pipeline(ggml_backend_vk_context * ctx, vk_matmul_pipeline& mmp, uint32_t m, uint32_t n, bool aligned, ggml_type src0_type) { + VK_LOG_DEBUG("ggml_vk_guess_matmul_id_pipeline(" << m << ", " << n << ", " << aligned << ", " << ggml_type_name(src0_type) << ")"); if (ctx->device->coopmat2) { - if ((ctx->device->mul_mat_id_l[src0_type] && (m % mmp->l->wg_denoms[0]) == 0 && (n % mmp->l->wg_denoms[1]) == 0) || (!ctx->device->mul_mat_id_m[src0_type] && !ctx->device->mul_mat_id_s[src0_type])) { + // Use large shader when the N dimension is greater than the medium shader's tile size + uint32_t crossover_large = mmp->m->wg_denoms[1]; + if ((ctx->device->mul_mat_id_l[src0_type] && (n > crossover_large)) || (!ctx->device->mul_mat_id_m[src0_type] && !ctx->device->mul_mat_id_s[src0_type])) { return aligned ? mmp->a_l : mmp->l; } - if ((ctx->device->mul_mat_id_m[src0_type] && (m % mmp->m->wg_denoms[0]) == 0 && (n % mmp->m->wg_denoms[1]) == 0) || !ctx->device->mul_mat_id_s[src0_type]) { + // Use medium shader when the N dimension is greater than the small shader's tile size + uint32_t crossover_medium = mmp->s->wg_denoms[1]; + if ((ctx->device->mul_mat_id_m[src0_type] && (n > crossover_medium)) || !ctx->device->mul_mat_id_s[src0_type]) { return aligned ? mmp->a_m : mmp->m; } return aligned ? mmp->a_s : mmp->s; @@ -3834,14 +4298,15 @@ static void ggml_vk_matmul_id( vk_subbuffer&& a, vk_subbuffer&& b, vk_subbuffer&& d, vk_subbuffer&& ids, uint32_t m, uint32_t n, uint32_t k, uint32_t stride_a, uint32_t stride_b, uint32_t stride_d, uint32_t batch_stride_a, uint32_t batch_stride_b, uint32_t batch_stride_d, - uint32_t n_as, uint32_t nei0, uint32_t nei1, uint32_t nbi1, uint32_t ne11) { + uint32_t n_as, uint32_t nei0, uint32_t nei1, uint32_t nbi1, uint32_t ne11, + uint32_t padded_n) { VK_LOG_DEBUG("ggml_vk_matmul_id(a: (" << a.buffer->buffer << ", " << a.offset << ", " << a.size << "), b: (" << b.buffer->buffer << ", " << b.offset << ", " << b.size << "), d: (" << d.buffer->buffer << ", " << d.offset << ", " << d.size << "), ids: (" << ids.buffer->buffer << ", " << ids.offset << ", " << ids.size << "), " << "m: " << m << ", n: " << n << ", k: " << k << ", stride_a: " << stride_a << ", stride_b: " << stride_b << ", stride_d: " << stride_d << ", " << "batch_stride_a: " << batch_stride_a << ", batch_stride_b: " << batch_stride_b << ", batch_stride_d: " << batch_stride_d << ", " << "n_as: " << n_as << ", nei0: " << nei0 << ", nei1: " << nei1 << ", nbi1: " << nbi1 << ", ne11: " << ne11 << ")"); ggml_vk_sync_buffers(subctx); const vk_mat_mat_id_push_constants pc = { m, n, k, stride_a, stride_b, stride_d, batch_stride_a, batch_stride_b, batch_stride_d, - nei0, nei1, nbi1, ne11 }; + nei0, nei1, nbi1, ne11, padded_n }; ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { a, b, d, ids }, sizeof(vk_mat_mat_id_push_constants), &pc, { m, nei1, n_as }); } @@ -3939,6 +4404,25 @@ static void ggml_vk_cpy_to_contiguous(ggml_backend_vk_context * ctx, vk_context& ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { in, out }, sizeof(vk_op_unary_push_constants), &pc, elements); } +static vk_pipeline ggml_vk_get_quantize_pipeline(ggml_backend_vk_context * ctx, ggml_type type) { + switch(type) { + case GGML_TYPE_Q8_1: + return ctx->device->pipeline_quantize_q8_1; + default: + std::cerr << "Missing quantize pipeline for type: " << ggml_type_name(type) << std::endl; + GGML_ABORT("fatal error"); + } +} + +static void ggml_vk_quantize_q8_1(ggml_backend_vk_context * ctx, vk_context& subctx, vk_subbuffer&& in, vk_subbuffer&& out, uint32_t ne) { + VK_LOG_DEBUG("ggml_vk_quantize_q8_1(" << "buffer in size=" << in.buffer->size << ", buffer out size=" << out.buffer->size << ", " << ne << ")"); + + vk_pipeline pipeline = ggml_vk_get_quantize_pipeline(ctx, GGML_TYPE_Q8_1); + + ggml_vk_sync_buffers(subctx); + ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { in, out }, sizeof(uint32_t), &ne, { ne, 1, 1 }); +} + static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) { VK_LOG_DEBUG("ggml_vk_mul_mat_q_f16((" << src0 << ", name=" << src0->name << ", type=" << src0->type << ", ne0=" << src0->ne[0] << ", ne1=" << src0->ne[1] << ", ne2=" << src0->ne[2] << ", ne3=" << src0->ne[3] << ", nb0=" << src0->nb[0] << ", nb1=" << src0->nb[1] << ", nb2=" << src0->nb[2] << ", nb3=" << src0->nb[3]; std::cerr << "), (" << src1 << ", name=" << src1->name << ", type=" << src1->type << ", ne0=" << src1->ne[0] << ", ne1=" << src1->ne[1] << ", ne2=" << src1->ne[2] << ", ne3=" << src1->ne[3] << ", nb0=" << src1->nb[0] << ", nb1=" << src1->nb[1] << ", nb2=" << src1->nb[2] << ", nb3=" << src1->nb[3]; @@ -3990,10 +4474,19 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& sub const bool y_f32_kernel = src1->type == GGML_TYPE_F32 && !y_non_contig; - vk_matmul_pipeline mmp = ggml_vk_get_mul_mat_mat_pipeline(ctx, src0->type, y_non_contig ? GGML_TYPE_F16 : src1->type, (ggml_prec)dst->op_params[0]); + bool quantize_y = ctx->device->integer_dot_product && src1->type == GGML_TYPE_F32 && ggml_is_contiguous(src1) && (ne11 * ne10) % 4 == 0; + + // Check for mmq first + vk_matmul_pipeline mmp = quantize_y ? ggml_vk_get_mul_mat_mat_pipeline(ctx, src0->type, GGML_TYPE_Q8_1, (ggml_prec)dst->op_params[0]) : nullptr; + + if (mmp == nullptr) { + // Fall back to f16 dequant mul mat + mmp = ggml_vk_get_mul_mat_mat_pipeline(ctx, src0->type, y_non_contig ? GGML_TYPE_F16 : src1->type, (ggml_prec)dst->op_params[0]); + quantize_y = false; + } const bool qx_needs_dequant = mmp == nullptr || x_non_contig; - const bool qy_needs_dequant = (src1->type != GGML_TYPE_F16 && !y_f32_kernel) || y_non_contig; + const bool qy_needs_dequant = !quantize_y && ((src1->type != GGML_TYPE_F16 && !y_f32_kernel) || y_non_contig); if (qx_needs_dequant) { // Fall back to dequant + f16 mulmat @@ -4003,25 +4496,28 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& sub // Not implemented GGML_ASSERT(y_non_contig || !qy_needs_dequant); // NOLINT + const uint32_t kpad = quantize_y ? 0 : ggml_vk_align_size(ne10, ggml_vk_guess_matmul_pipeline_align(ctx, mmp, ne01, ne11, qx_needs_dequant ? GGML_TYPE_F16 : src0->type, quantize_y ? GGML_TYPE_Q8_1 : (y_f32_kernel ? GGML_TYPE_F32 : src1->type))); + const bool aligned = !quantize_y && ne10 == kpad && ne01 > 8 && ne11 > 8; + + vk_pipeline pipeline = ggml_vk_guess_matmul_pipeline(ctx, mmp, ne01, ne11, aligned, qx_needs_dequant ? GGML_TYPE_F16 : src0->type, quantize_y ? GGML_TYPE_Q8_1 : (y_f32_kernel ? GGML_TYPE_F32 : src1->type)); + + // Reserve extra storage in the N dimension for the Y matrix, so we can avoid bounds-checking + uint32_t padded_n = qy_needs_dequant ? ROUNDUP_POW2(ne11, pipeline->wg_denoms[1]) : ne11; const int x_ne = ne01 * ne00; - const int y_ne = ne11 * ne10; + const int y_ne = padded_n * ne10; const int d_ne = ne11 * ne01; - const uint32_t kpad = ggml_vk_align_size(ne10, ggml_vk_guess_matmul_pipeline_align(ctx, mmp, ne01, ne11, qx_needs_dequant ? GGML_TYPE_F16 : src0->type)); - const bool aligned = ne10 == kpad && ne01 > 8 && ne11 > 8; - - vk_pipeline pipeline = ggml_vk_guess_matmul_pipeline(ctx, mmp, ne01, ne11, aligned, qx_needs_dequant ? GGML_TYPE_F16 : src0->type); - const uint32_t split_k = ggml_vk_guess_split_k(ctx, ne01, ne11, ne10, pipeline); const uint64_t qx_sz = ggml_type_size(src0->type) * x_ne / ggml_blck_size(src0->type); const uint64_t qy_sz = ggml_type_size(src1->type) * y_ne / ggml_blck_size(src1->type); const uint64_t x_sz = !qx_needs_dequant ? qx_sz : sizeof(ggml_fp16_t) * x_ne; - const uint64_t y_sz = y_f32_kernel ? sizeof(float) * y_ne : sizeof(ggml_fp16_t) * y_ne; + const uint64_t y_sz = quantize_y ? (y_ne * ggml_type_size(GGML_TYPE_Q8_1) / ggml_blck_size(GGML_TYPE_Q8_1)) : (y_f32_kernel ? sizeof(float) * y_ne : sizeof(ggml_fp16_t) * y_ne); const uint64_t d_sz = sizeof(float) * d_ne; vk_pipeline to_fp16_vk_0 = nullptr; vk_pipeline to_fp16_vk_1 = nullptr; + vk_pipeline to_q8_1 = nullptr; if (x_non_contig) { to_fp16_vk_0 = ggml_vk_get_cpy_pipeline(ctx, src0, nullptr, GGML_TYPE_F16); @@ -4036,6 +4532,10 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& sub GGML_ASSERT(!qx_needs_dequant || to_fp16_vk_0 != nullptr); // NOLINT GGML_ASSERT(!qy_needs_dequant || to_fp16_vk_1 != nullptr); // NOLINT + if (quantize_y) { + to_q8_1 = ggml_vk_get_quantize_pipeline(ctx, GGML_TYPE_Q8_1); + } + if (dryrun) { const uint64_t x_sz_upd = x_sz * ne02 * ne03; const uint64_t y_sz_upd = y_sz * ne12 * ne13; @@ -4049,7 +4549,7 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& sub if (qx_needs_dequant && ctx->prealloc_size_x < x_sz_upd) { ctx->prealloc_size_x = x_sz_upd; } - if (qy_needs_dequant && ctx->prealloc_size_y < y_sz_upd) { + if ((qy_needs_dequant || quantize_y) && ctx->prealloc_size_y < y_sz_upd) { ctx->prealloc_size_y = y_sz_upd; } if (split_k > 1 && ctx->prealloc_size_split_k < split_k_size) { @@ -4064,6 +4564,9 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& sub if (qy_needs_dequant) { ggml_pipeline_request_descriptor_sets(ctx->device, to_fp16_vk_1, 1); } + if (quantize_y) { + ggml_pipeline_request_descriptor_sets(ctx->device, to_q8_1, 1); + } if (split_k > 1) { ggml_pipeline_request_descriptor_sets(ctx->device, ctx->device->pipeline_matmul_split_k_reduce, 1); } @@ -4098,7 +4601,10 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& sub } if (qy_needs_dequant) { d_Y = ctx->prealloc_y; - GGML_ASSERT(d_Y->size >= y_sz * ne02 * ne03); + GGML_ASSERT(d_Y->size >= y_sz * ne12 * ne13); + } else if (quantize_y) { + d_Y = ctx->prealloc_y; + GGML_ASSERT(d_Y->size >= y_ne * ggml_type_size(GGML_TYPE_Q8_1) / ggml_blck_size(GGML_TYPE_Q8_1)); } else { d_Y = d_Qy; y_buf_offset = qy_buf_offset; @@ -4115,6 +4621,9 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& sub if (y_non_contig) { ggml_vk_cpy_to_contiguous(ctx, subctx, to_fp16_vk_1, src1, { d_Qy, qy_buf_offset, VK_WHOLE_SIZE }, { d_Y, 0, VK_WHOLE_SIZE }); } + if (quantize_y) { + ggml_vk_quantize_q8_1(ctx, subctx, { d_Qy, qy_buf_offset, VK_WHOLE_SIZE }, { d_Y, 0, VK_WHOLE_SIZE }, y_ne * ne12 * ne13); + } uint32_t stride_batch_x = ne00*ne01; uint32_t stride_batch_y = ne10*ne11; @@ -4123,7 +4632,7 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& sub stride_batch_x = src0->nb[0] / ggml_type_size(src0->type); } - if (!ggml_vk_dim01_contiguous(src1) && !qy_needs_dequant) { + if (!ggml_vk_dim01_contiguous(src1) && !qy_needs_dequant && !quantize_y) { stride_batch_y = src1->nb[0] / ggml_type_size(src1->type); } @@ -4134,7 +4643,7 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& sub { d_D, d_buf_offset, d_sz * ne12 * ne13 }, { ctx->prealloc_split_k, 0, d_sz * ne12 * ne13 * split_k }, ne01, ne11, ne10, ne10, ne10, ne01, stride_batch_x, stride_batch_y, ne20*ne21, - split_k, ne12*ne13, ne02, ne12, r2, r3 + split_k, ne12*ne13, ne02, ne12, r2, r3, padded_n ); // NOLINT } @@ -4371,9 +4880,15 @@ static void ggml_vk_mul_mat_vec_p021_f16_f32(ggml_backend_vk_context * ctx, vk_c const uint64_t qy_sz = ggml_type_size(src1->type) * y_ne / ggml_blck_size(src1->type); const uint64_t d_sz = sizeof(float) * d_ne; + // With grouped query attention there are > 1 Q matrices per K, V matrix. + uint32_t gqa_ratio = (uint32_t)ne12 / (uint32_t)ne02; + if (gqa_ratio > 8 || gqa_ratio == 0 || ne12 != ne02 * gqa_ratio) { + gqa_ratio = 1; + } + if (dryrun) { // Request descriptor sets - ggml_pipeline_request_descriptor_sets(ctx->device, ctx->device->pipeline_mul_mat_vec_p021_f16_f32, 1); + ggml_pipeline_request_descriptor_sets(ctx->device, ctx->device->pipeline_mul_mat_vec_p021_f16_f32[gqa_ratio - 1], 1); return; } @@ -4397,8 +4912,15 @@ static void ggml_vk_mul_mat_vec_p021_f16_f32(ggml_backend_vk_context * ctx, vk_c // compute const std::array pc = { (uint32_t)ne00, (uint32_t)ne01, (uint32_t)ne02, (uint32_t)ne12, (uint32_t)(qy_shader_offset / ggml_type_size(src1->type)), (uint32_t)(d_shader_offset / ggml_type_size(dst->type)) }; + + uint32_t workgroups_z = (uint32_t)ne12; + // When gqa_ratio > 1, each invocation does multiple rows and we can launch fewer workgroups + if (gqa_ratio > 1) { + workgroups_z /= gqa_ratio; + } + ggml_vk_sync_buffers(subctx); - ggml_vk_dispatch_pipeline(ctx, subctx, ctx->device->pipeline_mul_mat_vec_p021_f16_f32, { vk_subbuffer{ d_Qx, qx_buf_offset, qx_sz }, vk_subbuffer{ d_Qy, qy_buffer_offset, qy_sz + qy_shader_offset }, vk_subbuffer{ d_D, d_buffer_offset, d_sz + d_shader_offset } }, 6 * sizeof(uint32_t), &pc, { 1, (uint32_t)ne01, (uint32_t)ne12 }); + ggml_vk_dispatch_pipeline(ctx, subctx, ctx->device->pipeline_mul_mat_vec_p021_f16_f32[gqa_ratio - 1], { vk_subbuffer{ d_Qx, qx_buf_offset, qx_sz }, vk_subbuffer{ d_Qy, qy_buffer_offset, qy_sz + qy_shader_offset }, vk_subbuffer{ d_D, d_buffer_offset, d_sz + d_shader_offset } }, 6 * sizeof(uint32_t), &pc, { 1, (uint32_t)ne01, workgroups_z }); } static void ggml_vk_mul_mat_vec_nc_f16_f32(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) { @@ -4585,15 +5107,17 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context& // Not implemented GGML_ASSERT(y_non_contig || !qy_needs_dequant); // NOLINT - const uint64_t x_ne = ne01 * ne00; - const uint64_t y_ne = ne11 * ne10; - const uint64_t d_ne = ne21 * ne20; - const uint32_t kpad = ggml_vk_align_size(ne10, ggml_vk_guess_matmul_id_pipeline_align(ctx, mmp, ne01, nei1, qx_needs_dequant ? GGML_TYPE_F16 : src0->type)); const bool aligned = ne10 == kpad && ne01 > 8 && nei1 > 8; vk_pipeline pipeline = ggml_vk_guess_matmul_id_pipeline(ctx, mmp, ne01, nei1, aligned, qx_needs_dequant ? GGML_TYPE_F16 : src0->type); + // Reserve extra storage in the N dimension for the Y matrix, so we can avoid bounds-checking + uint32_t padded_n = qy_needs_dequant ? ROUNDUP_POW2(ne11, pipeline->wg_denoms[1]) :ne11; + const uint64_t x_ne = ne01 * ne00; + const uint64_t y_ne = padded_n * ne10; + const uint64_t d_ne = ne21 * ne20; + const uint64_t qx_sz = ggml_type_size(src0->type) * x_ne / ggml_blck_size(src0->type); const uint64_t qy_sz = ggml_type_size(src1->type) * y_ne / ggml_blck_size(src1->type); const uint64_t x_sz = !qx_needs_dequant ? qx_sz : sizeof(ggml_fp16_t) * x_ne; @@ -4675,7 +5199,7 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context& } if (qy_needs_dequant) { d_Y = ctx->prealloc_y; - GGML_ASSERT(d_Y->size >= y_sz * ne02 * ne03); + GGML_ASSERT(d_Y->size >= y_sz * ne12 * ne13); } else { d_Y = d_Qy; y_buf_offset = qy_buf_offset; @@ -4712,7 +5236,7 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context& { d_D, d_buf_offset, d_sz * ne22 * ne23 }, { d_ids, ids_buf_offset, ids_sz }, ne01, ne21, ne10, ne10, ne10, ne01, stride_batch_x, stride_batch_y, ne20*ne21, - n_as, nei0, nei1, nbi1 / ggml_type_size(ids->type), ne11 + n_as, nei0, nei1, nbi1 / ggml_type_size(ids->type), ne11, padded_n ); // NOLINT } @@ -4940,7 +5464,7 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx const uint32_t nbm1 = mask ? mask->nb[1] : 0; const uint32_t D = neq0; - const uint32_t N = neq1; + uint32_t N = neq1; const uint32_t KV = nek1; GGML_ASSERT(ne0 == D); @@ -4995,12 +5519,60 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx // the "aligned" shader variant will forcibly align strides, for performance (q_stride & 7) == 0 && (k_stride & 7) == 0 && (v_stride & 7) == 0; + // mask dim1 is padded to 64, we rely on this to avoid clamping mask loads + GGML_ASSERT((nem1 % GGML_KQ_MASK_PAD) == 0); + vk_pipeline pipeline = pipelines[aligned]; assert(pipeline); + uint32_t gqa_ratio = 1; + uint32_t qk_ratio = neq2 / nek2; + uint32_t workgroups_x = (uint32_t)neq1; + uint32_t workgroups_y = (uint32_t)neq2; + uint32_t workgroups_z = (uint32_t)neq3; + + if (N == 1 && qk_ratio > 1 && is_pow2(qk_ratio) && gqa_ratio <= flash_attention_num_small_rows && + qk_ratio * nek2 == neq2 && nek2 == nev2 && neq3 == 1 && nek3 == 1 && nev3 == 1) { + // grouped query attention - make the N dimension equal to gqa_ratio, reduce + // workgroups proportionally in y dimension. The shader will detect gqa_ratio > 1 + // and change addressing calculations to index Q's dimension 2. + gqa_ratio = qk_ratio; + N = gqa_ratio; + workgroups_y /= N; + } + + uint32_t split_kv = KV; + uint32_t split_k = 1; + + if (gqa_ratio > 1 && ctx->device->shader_core_count > 0) { + GGML_ASSERT(workgroups_x == 1); + // Try to run two workgroups per SM. + split_k = ctx->device->shader_core_count * 2 / workgroups_y; + if (split_k > 1) { + // Try to evenly split KV into split_k chunks, but it needs to be a multiple + // of "align", so recompute split_k based on that. + split_kv = ROUNDUP_POW2(KV / split_k, pipelines[1]->align); + split_k = CEIL_DIV(KV, split_kv); + workgroups_x = split_k; + } + } + + // Reserve space for split_k temporaries. For each split, we need to store the O matrix (D x ne1) + // and the per-row m and L values (ne1 rows). + const uint64_t split_k_size = split_k > 1 ? (D * ne1 * sizeof(float) + ne1 * sizeof(float) * 2) * split_k : 0; + if (split_k_size > ctx->device->max_memory_allocation_size) { + GGML_ABORT("Requested preallocation size is too large"); + } + if (ctx->prealloc_size_split_k < split_k_size) { + ctx->prealloc_size_split_k = split_k_size; + } + if (dryrun) { // Request descriptor sets ggml_pipeline_request_descriptor_sets(ctx->device, pipeline, 1); + if (split_k > 1) { + ggml_pipeline_request_descriptor_sets(ctx->device, ctx->device->pipeline_flash_attn_split_k_reduce, 1); + } return; } @@ -5021,8 +5593,6 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx const float m0 = powf(2.0f, -(max_bias ) / n_head_log2); const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2); - ggml_vk_sync_buffers(subctx); - vk_buffer d_Q = nullptr, d_K = nullptr, d_V = nullptr, d_D = nullptr, d_M = nullptr; size_t q_buf_offset = 0, k_buf_offset = 0, v_buf_offset = 0, d_buf_offset = 0, m_buf_offset = 0; @@ -5087,16 +5657,45 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx v_stride, (uint32_t)nbv2, (uint32_t)nbv3, nbm1, scale, max_bias, logit_softcap, - mask != nullptr, n_head_log2, m0, m1 }; - ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, - { - vk_subbuffer{d_Q, q_buf_offset, VK_WHOLE_SIZE}, - vk_subbuffer{d_K, k_buf_offset, VK_WHOLE_SIZE}, - vk_subbuffer{d_V, v_buf_offset, VK_WHOLE_SIZE}, - vk_subbuffer{d_M, m_buf_offset, VK_WHOLE_SIZE}, - vk_subbuffer{d_D, d_buf_offset, VK_WHOLE_SIZE}, - }, - sizeof(vk_flash_attn_push_constants), &pc, { (uint32_t)neq1, (uint32_t)neq2, (uint32_t)neq3 }); + mask != nullptr, n_head_log2, m0, m1, + gqa_ratio, split_kv, split_k }; + + ggml_vk_sync_buffers(subctx); + + if (split_k > 1) { + ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, + { + vk_subbuffer{d_Q, q_buf_offset, VK_WHOLE_SIZE}, + vk_subbuffer{d_K, k_buf_offset, VK_WHOLE_SIZE}, + vk_subbuffer{d_V, v_buf_offset, VK_WHOLE_SIZE}, + vk_subbuffer{d_M, m_buf_offset, VK_WHOLE_SIZE}, + vk_subbuffer{ctx->prealloc_split_k, 0, VK_WHOLE_SIZE}, + }, + // We only use split_k when group query attention is enabled, which means + // there's no more than one tile of rows (i.e. workgroups_x would have been + // one). We reuse workgroups_x to mean the number of splits, so we need to + // cancel out the divide by wg_denoms[0]. + sizeof(vk_flash_attn_push_constants), &pc, { workgroups_x * pipeline->wg_denoms[0], workgroups_y, workgroups_z }); + + ggml_vk_sync_buffers(subctx); + const std::array pc2 = { D, (uint32_t)ne1, split_k }; + ggml_vk_dispatch_pipeline(ctx, subctx, ctx->device->pipeline_flash_attn_split_k_reduce, + { + vk_subbuffer{ctx->prealloc_split_k, 0, VK_WHOLE_SIZE}, + vk_subbuffer{d_D, d_buf_offset, VK_WHOLE_SIZE}, + }, + pc2.size() * uint32_t{sizeof(uint32_t)}, pc2.data(), { (uint32_t)ne1, 1, 1 }); + } else { + ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, + { + vk_subbuffer{d_Q, q_buf_offset, VK_WHOLE_SIZE}, + vk_subbuffer{d_K, k_buf_offset, VK_WHOLE_SIZE}, + vk_subbuffer{d_V, v_buf_offset, VK_WHOLE_SIZE}, + vk_subbuffer{d_M, m_buf_offset, VK_WHOLE_SIZE}, + vk_subbuffer{d_D, d_buf_offset, VK_WHOLE_SIZE}, + }, + sizeof(vk_flash_attn_push_constants), &pc, { workgroups_x, workgroups_y, workgroups_z }); + } } static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst, ggml_op op) { @@ -5123,6 +5722,11 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const return ggml_are_same_shape(src0, src1) ? ctx->device->pipeline_add_f16_f32_f16_norepeat : ctx->device->pipeline_add_f16_f32_f16; } return nullptr; + case GGML_OP_SUB: + if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) { + return ggml_are_same_shape(src0, src1) ? ctx->device->pipeline_sub_f32_norepeat : ctx->device->pipeline_sub_f32; + } + return nullptr; case GGML_OP_MUL: if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) { return ggml_are_same_shape(src0, src1) ? ctx->device->pipeline_mul_f32_norepeat : ctx->device->pipeline_mul_f32; @@ -5145,7 +5749,7 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const } return nullptr; case GGML_OP_UPSCALE: - if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) { + if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32 && dst->op_params[0] == GGML_SCALE_MODE_NEAREST) { return ctx->device->pipeline_upscale_f32; } return nullptr; @@ -5184,10 +5788,20 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const return ctx->device->pipeline_repeat_f32; } return nullptr; + case GGML_OP_REPEAT_BACK: + if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) { + return ctx->device->pipeline_repeat_back_f32; + } + return nullptr; case GGML_OP_CPY: case GGML_OP_CONT: case GGML_OP_DUP: return ggml_vk_get_cpy_pipeline(ctx, src0, dst, dst->type); + case GGML_OP_SILU_BACK: + if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) { + return ctx->device->pipeline_silu_back_f32; + } + return nullptr; case GGML_OP_NORM: if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) { return ctx->device->pipeline_norm_f32; @@ -5203,6 +5817,16 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const return ctx->device->pipeline_rms_norm_f32; } return nullptr; + case GGML_OP_RMS_NORM_BACK: + if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) { + return ctx->device->pipeline_rms_norm_back_f32; + } + return nullptr; + case GGML_OP_L2_NORM: + if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) { + return ctx->device->pipeline_l2_norm_f32; + } + return nullptr; case GGML_OP_UNARY: switch (ggml_get_unary_op(dst)) { case GGML_UNARY_OP_SILU: @@ -5230,6 +5854,11 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const return ctx->device->pipeline_tanh_f32; } break; + case GGML_UNARY_OP_SIGMOID: + if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) { + return ctx->device->pipeline_sigmoid_f32; + } + break; default: break; } @@ -5249,10 +5878,18 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const return src0->ne[0] > 1024 ? ctx->device->pipeline_soft_max_f32_f16_wg512 : ctx->device->pipeline_soft_max_f32_f16; } return nullptr; + case GGML_OP_SOFT_MAX_BACK: + if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) { + return ctx->device->pipeline_soft_max_back_f32; + } + return nullptr; case GGML_OP_ROPE: + case GGML_OP_ROPE_BACK: { const int mode = ((const int32_t *) dst->op_params)[2]; const bool is_neox = mode & GGML_ROPE_TYPE_NEOX; + const bool is_mrope = mode & GGML_ROPE_TYPE_MROPE; + const bool is_vision = mode == GGML_ROPE_TYPE_VISION; if (is_neox) { if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) { @@ -5261,6 +5898,20 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const if (src0->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F16) { return ctx->device->pipeline_rope_neox_f16; } + } else if (is_mrope && !is_vision) { + if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) { + return ctx->device->pipeline_rope_multi_f32; + } + if (src0->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F16) { + return ctx->device->pipeline_rope_multi_f16; + } + } else if (is_vision) { + if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) { + return ctx->device->pipeline_rope_vision_f32; + } + if (src0->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F16) { + return ctx->device->pipeline_rope_vision_f16; + } } else { if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) { return ctx->device->pipeline_rope_norm_f32; @@ -5276,11 +5927,22 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const return ctx->device->pipeline_argsort_f32; } return nullptr; + case GGML_OP_SUM: case GGML_OP_SUM_ROWS: if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) { return ctx->device->pipeline_sum_rows_f32; } return nullptr; + case GGML_OP_ARGMAX: + if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_I32) { + return ctx->device->pipeline_argmax_f32; + } + return nullptr; + case GGML_OP_COUNT_EQUAL: + if (src0->type == GGML_TYPE_I32 && src1->type == GGML_TYPE_I32 && dst->type == GGML_TYPE_I64) { + return ctx->device->pipeline_count_equal_i32; + } + return nullptr; case GGML_OP_IM2COL: if (src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) { return ctx->device->pipeline_im2col_f32; @@ -5304,6 +5966,16 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const return ctx->device->pipeline_rwkv_wkv6_f32; } return nullptr; + case GGML_OP_RWKV_WKV7: + if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) { + return ctx->device->pipeline_rwkv_wkv7_f32; + } + return nullptr; + case GGML_OP_OPT_STEP_ADAMW: + if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) { + return ctx->device->pipeline_opt_step_adamw_f32; + } + return nullptr; case GGML_OP_LEAKY_RELU: if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) { return ctx->device->pipeline_leaky_relu_f32; @@ -5321,6 +5993,7 @@ static bool ggml_vk_op_supports_incontiguous(ggml_op op) { case GGML_OP_CPY: case GGML_OP_GET_ROWS: case GGML_OP_ADD: + case GGML_OP_SUB: case GGML_OP_MUL: case GGML_OP_DIV: case GGML_OP_CONCAT: @@ -5331,6 +6004,8 @@ static bool ggml_vk_op_supports_incontiguous(ggml_op op) { case GGML_OP_CLAMP: case GGML_OP_PAD: case GGML_OP_REPEAT: + case GGML_OP_REPEAT_BACK: + case GGML_OP_ROPE: return true; default: return false; @@ -5542,8 +6217,12 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co switch (op) { case GGML_OP_NORM: case GGML_OP_RMS_NORM: + case GGML_OP_RMS_NORM_BACK: + case GGML_OP_L2_NORM: case GGML_OP_SOFT_MAX: + case GGML_OP_SOFT_MAX_BACK: case GGML_OP_SUM_ROWS: + case GGML_OP_ARGMAX: { const uint32_t nr = ggml_nrows(src0); if (nr > 262144) { @@ -5554,6 +6233,10 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co elements = { nr, 1, 1 }; } } break; + case GGML_OP_SUM: + // We use GGML_OP_SUM_ROWS with 1 row. + elements = { 1, 1, 1 }; + break; case GGML_OP_GROUP_NORM: { const uint32_t num_groups = dst->op_params[0]; @@ -5561,6 +6244,7 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co } break; case GGML_OP_DIAG_MASK_INF: case GGML_OP_ROPE: + case GGML_OP_ROPE_BACK: elements = { (uint32_t)ggml_nrows(src0), (uint32_t)ne00, 1 }; break; case GGML_OP_GET_ROWS: @@ -5600,6 +6284,7 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co elements = { N * OC * OH * OW, 1, 1}; } break; case GGML_OP_ADD: + case GGML_OP_SUB: case GGML_OP_DIV: case GGML_OP_MUL: case GGML_OP_SCALE: @@ -5609,6 +6294,7 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co case GGML_OP_CLAMP: case GGML_OP_PAD: case GGML_OP_REPEAT: + case GGML_OP_REPEAT_BACK: case GGML_OP_CPY: case GGML_OP_CONCAT: case GGML_OP_UPSCALE: @@ -5654,7 +6340,7 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co ggml_vk_sync_buffers(subctx); ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_X, x_buf_offset, x_sz }, subbuf_y, vk_subbuffer{ d_D, d_buf_offset, d_sz } }, sizeof(PC), &pc, elements); - } else if (op == GGML_OP_ROPE) { + } else if (op == GGML_OP_ROPE || op == GGML_OP_ROPE_BACK) { // Empty src2 is possible in rope, but the shader needs a buffer vk_subbuffer subbuf_z; if (use_src2) { @@ -5669,6 +6355,12 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co // im2col uses only src1 and dst buffers ggml_vk_sync_buffers(subctx); ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_Y, y_buf_offset, y_sz }, vk_subbuffer{ d_D, d_buf_offset, d_sz } }, sizeof(PC), &pc, elements); + } else if (op == GGML_OP_COUNT_EQUAL) { + ggml_vk_sync_buffers(subctx); + // count_equal assumes that destination buffer is initialized with zeroes + ggml_vk_buffer_memset_async(subctx, d_D, d_buf_offset, 0, d_sz); + ggml_vk_sync_buffers(subctx); + ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_X, x_buf_offset, x_sz }, vk_subbuffer{ d_Y, y_buf_offset, y_sz }, vk_subbuffer{ d_D, d_buf_offset, d_sz } }, sizeof(PC), &pc, elements); } else if (use_src2) { ggml_vk_sync_buffers(subctx); ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_X, x_buf_offset, x_sz }, vk_subbuffer{ d_Y, y_buf_offset, y_sz }, vk_subbuffer{ d_Z, z_buf_offset, z_sz }, vk_subbuffer{ d_D, d_buf_offset, d_sz } }, sizeof(PC), &pc, elements); @@ -5731,6 +6423,21 @@ static void ggml_vk_add(ggml_backend_vk_context * ctx, vk_context& subctx, const }, dryrun); } +static void ggml_vk_sub(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) { + const uint32_t src0_type_size = ggml_type_size(src0->type); + const uint32_t src1_type_size = ggml_type_size(src1->type); + const uint32_t dst_type_size = ggml_type_size(dst->type); + + ggml_vk_op_f32(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_SUB, { + (uint32_t)ggml_nelements(src0), + (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2],(uint32_t)src0->ne[3], (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)src0->nb[1] / src0_type_size, (uint32_t)src0->nb[2] / src0_type_size, (uint32_t)src0->nb[3] / src0_type_size, + (uint32_t)src1->ne[0], (uint32_t)src1->ne[1], (uint32_t)src1->ne[2],(uint32_t)src1->ne[3], (uint32_t)src1->nb[0] / src1_type_size, (uint32_t)src1->nb[1] / src1_type_size, (uint32_t)src1->nb[2] / src1_type_size, (uint32_t)src1->nb[3] / src1_type_size, + (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2],(uint32_t) dst->ne[3], (uint32_t) dst->nb[0] / dst_type_size, (uint32_t) dst->nb[1] / dst_type_size, (uint32_t) dst->nb[2] / dst_type_size, (uint32_t) dst->nb[3] / dst_type_size, + 0, + 0.0f, 0.0f, 0, + }, dryrun); +} + static void ggml_vk_mul(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) { const uint32_t src0_type_size = ggml_type_size(src0->type); const uint32_t src1_type_size = ggml_type_size(src1->type); @@ -5761,23 +6468,17 @@ static void ggml_vk_div(ggml_backend_vk_context * ctx, vk_context& subctx, const }, dryrun); } -static void ggml_vk_op_f32_rwkv6(ggml_backend_vk_context * ctx, vk_context& subctx, ggml_tensor * dst, const vk_op_rwkv_wkv6_push_constants&& pc, bool dryrun = false) { - const ggml_tensor * k = dst->src[0]; - const ggml_tensor * v = dst->src[1]; - const ggml_tensor * r = dst->src[2]; - const ggml_tensor * tf = dst->src[3]; - const ggml_tensor * td = dst->src[4]; - const ggml_tensor * state = dst->src[5]; +static void ggml_vk_op_f32_wkv(ggml_backend_vk_context * ctx, vk_context& subctx, ggml_tensor * dst, const vk_op_rwkv_wkv6_push_constants&& pc, int version, bool dryrun = false) { + GGML_ASSERT(version == 6 || version == 7); + int num_srcs = version == 6 ? 6 : 7; + + for (int i = 0; i < num_srcs; i++) { + GGML_ASSERT(!ggml_is_quantized(dst->src[i]->type)); + } - GGML_ASSERT(!ggml_is_quantized(k->type)); - GGML_ASSERT(!ggml_is_quantized(v->type)); - GGML_ASSERT(!ggml_is_quantized(r->type)); - GGML_ASSERT(!ggml_is_quantized(tf->type)); - GGML_ASSERT(!ggml_is_quantized(td->type)); - GGML_ASSERT(!ggml_is_quantized(state->type)); GGML_ASSERT(dst->buffer != nullptr); - vk_pipeline pipeline = ggml_vk_op_get_pipeline(ctx, k, v, r, dst, GGML_OP_RWKV_WKV6); + vk_pipeline pipeline = ggml_vk_op_get_pipeline(ctx, dst->src[0], dst->src[1], dst->src[2], dst, dst->op); GGML_ASSERT(pipeline != nullptr); if (dryrun) { @@ -5786,89 +6487,73 @@ static void ggml_vk_op_f32_rwkv6(ggml_backend_vk_context * ctx, vk_context& subc } ggml_backend_vk_buffer_context * dst_buf_ctx = (ggml_backend_vk_buffer_context *)dst->buffer->context; - ggml_backend_vk_buffer_context * k_buf_ctx = (ggml_backend_vk_buffer_context *)k->buffer->context; - ggml_backend_vk_buffer_context * v_buf_ctx = (ggml_backend_vk_buffer_context *)v->buffer->context; - ggml_backend_vk_buffer_context * r_buf_ctx = (ggml_backend_vk_buffer_context *)r->buffer->context; - ggml_backend_vk_buffer_context * tf_buf_ctx = (ggml_backend_vk_buffer_context *)tf->buffer->context; - ggml_backend_vk_buffer_context * td_buf_ctx = (ggml_backend_vk_buffer_context *)td->buffer->context; - ggml_backend_vk_buffer_context * state_buf_ctx = (ggml_backend_vk_buffer_context *)state->buffer->context; + ggml_backend_vk_buffer_context * src_buf_ctxs[7] = { nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr }; + for (int i = 0; i < num_srcs; i++) { + src_buf_ctxs[i] = (ggml_backend_vk_buffer_context *)dst->src[i]->buffer->context; + } ggml_vk_sync_buffers(subctx); - vk_buffer d_D = nullptr, d_K = nullptr, d_V = nullptr, d_R = nullptr, d_TF = nullptr, d_TD = nullptr, d_State = nullptr; - size_t k_offset = 0, v_offset = 0, r_offset = 0, tf_offset = 0, td_offset = 0, state_offset = 0, dst_offset = 0; - bool K_uma = false, V_uma = false, R_uma = false, TF_uma = false, TD_uma = false, STATE_uma = false, DST_uma = false; + vk_buffer d_D = nullptr, d_srcs[7] = { nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, nullptr }; + size_t dst_offset = 0, src_offsets[7] = { 0, 0, 0, 0, 0, 0, 0 }; + bool dst_uma = false, srcs_uma[7] = { false, false, false, false, false, false, false }; if (ctx->device->uma) { - ggml_vk_host_get(ctx->device, k->data, d_K, k_offset); - ggml_vk_host_get(ctx->device, v->data, d_V, v_offset); - ggml_vk_host_get(ctx->device, r->data, d_R, r_offset); - ggml_vk_host_get(ctx->device, tf->data, d_TF, tf_offset); - ggml_vk_host_get(ctx->device, td->data, d_TD, td_offset); - ggml_vk_host_get(ctx->device, state->data, d_State, state_offset); + for (int i = 0; i < num_srcs; i++) { + ggml_vk_host_get(ctx->device, dst->src[i]->data, d_srcs[i], src_offsets[i]); + srcs_uma[i] = d_srcs[i] != nullptr; + } + ggml_vk_host_get(ctx->device, dst->data, d_D, dst_offset); - - K_uma = d_K != nullptr; - V_uma = d_V != nullptr; - R_uma = d_R != nullptr; - TF_uma = d_TF != nullptr; - TD_uma = d_TD != nullptr; - STATE_uma = d_State != nullptr; - DST_uma = d_D != nullptr; + dst_uma = d_D != nullptr; } - if (!K_uma) { - d_K = k_buf_ctx->dev_buffer; - k_offset = vk_tensor_offset(k) + k->view_offs; + uint64_t src_sizes[7] = { 0, 0, 0, 0, 0, 0, 0 }; + for (int i = 0; i < num_srcs; i++) { + src_sizes[i] = ggml_nbytes(dst->src[i]); + if (!srcs_uma[i]) { + d_srcs[i] = src_buf_ctxs[i]->dev_buffer; + src_offsets[i] = vk_tensor_offset(dst->src[i]) + dst->src[i]->view_offs; + } } - if (!V_uma) { - d_V = v_buf_ctx->dev_buffer; - v_offset = vk_tensor_offset(v) + v->view_offs; - } - if (!R_uma) { - d_R = r_buf_ctx->dev_buffer; - r_offset = vk_tensor_offset(r) + r->view_offs; - } - if (!TF_uma) { - d_TF = tf_buf_ctx->dev_buffer; - tf_offset = vk_tensor_offset(tf) + tf->view_offs; - } - if (!TD_uma) { - d_TD = td_buf_ctx->dev_buffer; - td_offset = vk_tensor_offset(td) + td->view_offs; - } - if (!STATE_uma) { - d_State = state_buf_ctx->dev_buffer; - state_offset = vk_tensor_offset(state) + state->view_offs; - } - if (!DST_uma) { + + const uint64_t dst_size = ggml_nbytes(dst); + if (!dst_uma) { d_D = dst_buf_ctx->dev_buffer; dst_offset = vk_tensor_offset(dst) + dst->view_offs; } - const uint64_t k_size = ggml_nbytes(k); - const uint64_t v_size = ggml_nbytes(v); - const uint64_t r_size = ggml_nbytes(r); - const uint64_t tf_size = ggml_nbytes(tf); - const uint64_t td_size = ggml_nbytes(td); - const uint64_t state_size = ggml_nbytes(state); - const uint64_t dst_size = ggml_nbytes(dst); - std::array elements = { (uint32_t)(pc.B * pc.H), 1, 1 }; - ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { - vk_subbuffer{ d_K, k_offset, k_size }, - vk_subbuffer{ d_V, v_offset, v_size }, - vk_subbuffer{ d_R, r_offset, r_size }, - vk_subbuffer{ d_TF, tf_offset, tf_size }, - vk_subbuffer{ d_TD, td_offset, td_size }, - vk_subbuffer{ d_State, state_offset, state_size }, - vk_subbuffer{ d_D, dst_offset, dst_size } - }, sizeof(vk_op_rwkv_wkv6_push_constants), &pc, elements); + if (version == 6) { + ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { + vk_subbuffer{ d_srcs[0], src_offsets[0], src_sizes[0] }, + vk_subbuffer{ d_srcs[1], src_offsets[1], src_sizes[1] }, + vk_subbuffer{ d_srcs[2], src_offsets[2], src_sizes[2] }, + vk_subbuffer{ d_srcs[3], src_offsets[3], src_sizes[3] }, + vk_subbuffer{ d_srcs[4], src_offsets[4], src_sizes[4] }, + vk_subbuffer{ d_srcs[5], src_offsets[5], src_sizes[5] }, + vk_subbuffer{ d_D, dst_offset, dst_size } + }, sizeof(vk_op_rwkv_wkv6_push_constants), &pc, elements); + } else if (version == 7) { + ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { + vk_subbuffer{ d_srcs[0], src_offsets[0], src_sizes[0] }, + vk_subbuffer{ d_srcs[1], src_offsets[1], src_sizes[1] }, + vk_subbuffer{ d_srcs[2], src_offsets[2], src_sizes[2] }, + vk_subbuffer{ d_srcs[3], src_offsets[3], src_sizes[3] }, + vk_subbuffer{ d_srcs[4], src_offsets[4], src_sizes[4] }, + vk_subbuffer{ d_srcs[5], src_offsets[5], src_sizes[5] }, + vk_subbuffer{ d_srcs[6], src_offsets[6], src_sizes[6] }, + vk_subbuffer{ d_D, dst_offset, dst_size } + }, sizeof(vk_op_rwkv_wkv7_push_constants), &pc, elements); + } else { + // shouldn't happen + GGML_ASSERT(false); + } } static void ggml_vk_rwkv_wkv6(ggml_backend_vk_context * ctx, vk_context& subctx, ggml_tensor * dst, bool dryrun = false) { @@ -5877,7 +6562,7 @@ static void ggml_vk_rwkv_wkv6(ggml_backend_vk_context * ctx, vk_context& subctx, const size_t n_heads = dst->src[0]->ne[1]; const size_t n_seqs = dst->src[5]->ne[1]; - ggml_vk_op_f32_rwkv6( + ggml_vk_op_f32_wkv( ctx, subctx, dst, { (uint32_t)n_seqs, @@ -5885,6 +6570,131 @@ static void ggml_vk_rwkv_wkv6(ggml_backend_vk_context * ctx, vk_context& subctx, (uint32_t)n_embed, (uint32_t)n_heads, }, + 6, + dryrun + ); +} + +static void ggml_vk_rwkv_wkv7(ggml_backend_vk_context * ctx, vk_context& subctx, ggml_tensor * dst, bool dryrun = false) { + const size_t seq_length = dst->src[0]->ne[2]; + const size_t n_embed = dst->ne[0]; + const size_t n_heads = dst->src[0]->ne[1]; + const size_t n_seqs = dst->src[6]->ne[1]; + + ggml_vk_op_f32_wkv( + ctx, subctx, dst, + { + (uint32_t)n_seqs, + (uint32_t)seq_length, + (uint32_t)n_embed, + (uint32_t)n_heads, + }, + 7, + dryrun + ); +} + +static void ggml_vk_op_f32_opt_step_adamw(ggml_backend_vk_context * ctx, vk_context& subctx, ggml_tensor * dst, const vk_op_push_constants&& pc, bool dryrun = false) { + const ggml_tensor * x = dst->src[0]; + const ggml_tensor * g = dst->src[1]; + const ggml_tensor * gm = dst->src[2]; + const ggml_tensor * gv = dst->src[3]; + const ggml_tensor * p = dst->src[4]; + + GGML_ASSERT(x->type == GGML_TYPE_F32); + GGML_ASSERT(g->type == GGML_TYPE_F32); + GGML_ASSERT(gm->type == GGML_TYPE_F32); + GGML_ASSERT(gv->type == GGML_TYPE_F32); + GGML_ASSERT(p->type == GGML_TYPE_F32); + GGML_ASSERT(dst->buffer != nullptr); + GGML_ASSERT(ggml_is_contiguous(x)); + GGML_ASSERT(ggml_is_contiguous(g)); + GGML_ASSERT(ggml_is_contiguous(gm)); + GGML_ASSERT(ggml_is_contiguous(gv)); + GGML_ASSERT(ggml_is_contiguous(p)); + GGML_ASSERT(ggml_are_same_shape(x, g)); + GGML_ASSERT(ggml_are_same_shape(x, gm)); + GGML_ASSERT(ggml_are_same_shape(x, gv)); + GGML_ASSERT(ggml_nelements(p) == 7); + + vk_pipeline pipeline = ggml_vk_op_get_pipeline(ctx, g, gm, gv, dst, GGML_OP_OPT_STEP_ADAMW); + GGML_ASSERT(pipeline != nullptr); + + if (dryrun) { + ggml_pipeline_request_descriptor_sets(ctx->device, pipeline, 1); + return; + } + + ggml_backend_vk_buffer_context * x_buf_ctx = (ggml_backend_vk_buffer_context *)x->buffer->context; + ggml_backend_vk_buffer_context * g_buf_ctx = (ggml_backend_vk_buffer_context *)g->buffer->context; + ggml_backend_vk_buffer_context * gm_buf_ctx = (ggml_backend_vk_buffer_context *)gm->buffer->context; + ggml_backend_vk_buffer_context * gv_buf_ctx = (ggml_backend_vk_buffer_context *)gv->buffer->context; + ggml_backend_vk_buffer_context * p_buf_ctx = (ggml_backend_vk_buffer_context *)p->buffer->context; + + ggml_vk_sync_buffers(subctx); + + vk_buffer d_X = nullptr, d_G = nullptr, d_GM = nullptr, d_GV = nullptr, d_P = nullptr; + size_t x_offset = 0, g_offset = 0, gm_offset = 0, gv_offset = 0, p_offset = 0; + bool X_uma = false, G_uma = false, GM_uma = false, GV_uma = false, P_uma = false; + + if (ctx->device->uma) { + ggml_vk_host_get(ctx->device, x->data, d_X, x_offset); + ggml_vk_host_get(ctx->device, g->data, d_G, g_offset); + ggml_vk_host_get(ctx->device, gm->data, d_GM, gm_offset); + ggml_vk_host_get(ctx->device, gv->data, d_GV, gv_offset); + ggml_vk_host_get(ctx->device, p->data, d_P, p_offset); + + X_uma = d_X != nullptr; + G_uma = d_G != nullptr; + GM_uma = d_GM != nullptr; + GV_uma = d_GV != nullptr; + P_uma = d_P != nullptr; + } + + if (!X_uma) { + d_X = x_buf_ctx->dev_buffer; + x_offset = vk_tensor_offset(x) + x->view_offs; + } + if (!G_uma) { + d_G = g_buf_ctx->dev_buffer; + g_offset = vk_tensor_offset(g) + g->view_offs; + } + if (!GM_uma) { + d_GM = gm_buf_ctx->dev_buffer; + gm_offset = vk_tensor_offset(gm) + gm->view_offs; + } + if (!GV_uma) { + d_GV = gv_buf_ctx->dev_buffer; + gv_offset = vk_tensor_offset(gv) + gv->view_offs; + } + if (!P_uma) { + d_P = p_buf_ctx->dev_buffer; + p_offset = vk_tensor_offset(p) + p->view_offs; + } + + const uint64_t x_size = ggml_nbytes(x); + const uint64_t g_size = ggml_nbytes(g); + const uint64_t gm_size = ggml_nbytes(gm); + const uint64_t gv_size = ggml_nbytes(gv); + const uint64_t p_size = ggml_nbytes(p); + + std::array elements = { (uint32_t)ggml_nelements(x), 1, 1 }; + + ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { + vk_subbuffer{ d_X, x_offset, x_size }, + vk_subbuffer{ d_G, g_offset, g_size }, + vk_subbuffer{ d_GM, gm_offset, gm_size }, + vk_subbuffer{ d_GV, gv_offset, gv_size }, + vk_subbuffer{ d_P, p_offset, p_size }, + }, sizeof(vk_op_push_constants), &pc, elements); +} + +static void ggml_vk_opt_step_adamw(ggml_backend_vk_context * ctx, vk_context& subctx, ggml_tensor * dst, bool dryrun = false) { + const size_t n = ggml_nelements(dst->src[0]); + + ggml_vk_op_f32_opt_step_adamw( + ctx, subctx, dst, + { (uint32_t)n, 0, 0.0f, 0.0f }, dryrun ); } @@ -6022,6 +6832,20 @@ static void ggml_vk_repeat(ggml_backend_vk_context * ctx, vk_context& subctx, co }, dryrun); } +static void ggml_vk_repeat_back(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) { + const uint32_t src0_type_size = ggml_type_size(src0->type); + const uint32_t dst_type_size = ggml_type_size(dst->type); + + ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_REPEAT_BACK, { + (uint32_t)ggml_nelements(dst), + (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2], (uint32_t)src0->ne[3], (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)src0->nb[1] / src0_type_size, (uint32_t)src0->nb[2] / src0_type_size, (uint32_t)src0->nb[3] / src0_type_size, + (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2], (uint32_t) dst->ne[3], (uint32_t) dst->nb[0] / dst_type_size, (uint32_t) dst->nb[1] / dst_type_size, (uint32_t) dst->nb[2] / dst_type_size, (uint32_t) dst->nb[3] / dst_type_size, + 0, + 0.0f, 0.0f, + 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, + }, dryrun); +} + static void ggml_vk_cpy(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) { const uint32_t src0_type_size = ggml_type_size(src0->type); const uint32_t dst_type_size = ggml_type_size(dst->type); @@ -6036,6 +6860,10 @@ static void ggml_vk_cpy(ggml_backend_vk_context * ctx, vk_context& subctx, const }, dryrun); } +static void ggml_vk_silu_back(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) { + ggml_vk_op_f32(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_SILU_BACK, { (uint32_t)ggml_nelements(src0), 0, 0.0f, 0.0f }, dryrun); +} + static void ggml_vk_norm(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) { float * op_params = (float *)dst->op_params; @@ -6058,6 +6886,16 @@ static void ggml_vk_rms_norm(ggml_backend_vk_context * ctx, vk_context& subctx, ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_RMS_NORM, { (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], op_params[0], 0.0f }, dryrun); } +static void ggml_vk_rms_norm_back(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) { + float * op_params = (float *)dst->op_params; + ggml_vk_op_f32(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_RMS_NORM_BACK, { (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], op_params[0], 0.0f }, dryrun); +} + +static void ggml_vk_l2_norm(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) { + float * op_params = (float *)dst->op_params; + ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_L2_NORM, { (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], op_params[0], 0.0f }, dryrun); +} + static void ggml_vk_unary(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) { ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_UNARY, { (uint32_t)ggml_nelements(src0), 0, 0.0f, 0.0f }, dryrun); } @@ -6093,9 +6931,14 @@ static void ggml_vk_soft_max(ggml_backend_vk_context * ctx, vk_context& subctx, }, dryrun); } -static void ggml_vk_rope(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst, bool dryrun = false) { +static void ggml_vk_soft_max_back(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) { + float * op_params = (float *)dst->op_params; + ggml_vk_op_f32(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_SOFT_MAX_BACK, { (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], op_params[0], op_params[1] }, dryrun); +} + +static void ggml_vk_rope(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst, bool backprop, bool dryrun = false) { const int n_dims = ((int32_t *) dst->op_params)[1]; - // const int mode = ((int32_t *) dst->op_params)[2]; + const int mode = ((int32_t *) dst->op_params)[2]; // const int n_ctx = ((int32_t *) dst->op_params)[3]; const int n_ctx_orig = ((int32_t *) dst->op_params)[4]; const float freq_base = ((float *) dst->op_params)[5]; @@ -6104,16 +6947,24 @@ static void ggml_vk_rope(ggml_backend_vk_context * ctx, vk_context& subctx, cons const float attn_factor = ((float *) dst->op_params)[8]; const float beta_fast = ((float *) dst->op_params)[9]; const float beta_slow = ((float *) dst->op_params)[10]; + int sections[4] {}; + if (mode & GGML_ROPE_TYPE_MROPE) { + memcpy(sections, (int32_t *) dst->op_params + 11, sizeof(int)*4); + } float corr_dims[2]; ggml_rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow, corr_dims); const float theta_scale = powf(freq_base, -2.0f/n_dims); + uint32_t s1 = src0->nb[1] / ggml_type_size(src0->type); + uint32_t s2 = src0->nb[2] / ggml_type_size(src0->type); + ggml_vk_op_f32(ctx, subctx, src0, src1, src2, dst, GGML_OP_ROPE, { (uint32_t)src0->ne[0], (uint32_t)n_dims, freq_scale, (uint32_t)src0->ne[1], freq_base, ext_factor, attn_factor, {corr_dims[0], corr_dims[1]}, theta_scale, - src2 != nullptr, + src2 != nullptr, (uint32_t)src0->ne[2], s1, s2, + sections[0], sections[1], sections[2], sections[3], backprop }, dryrun); } @@ -6136,10 +6987,22 @@ static void ggml_vk_argsort(ggml_backend_vk_context * ctx, vk_context& subctx, c }, dryrun); } +static void ggml_vk_sum(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) { + ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_SUM, { (uint32_t)ggml_nelements(src0), 0, 0.0f, 0.0f }, dryrun); +} + static void ggml_vk_sum_rows(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) { ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_SUM_ROWS, { (uint32_t)src0->ne[0], 0, 0.0f, 0.0f }, dryrun); } +static void ggml_vk_argmax(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) { + ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_ARGMAX, { (uint32_t)src0->ne[0], 0, 0.0f, 0.0f }, dryrun); +} + +static void ggml_vk_count_equal(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) { + ggml_vk_op_f32(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_COUNT_EQUAL, { (uint32_t)ggml_nelements(src0), 0, 0.0f, 0.0f }, dryrun); +} + static void ggml_vk_im2col(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) { const int32_t s0 = dst->op_params[0]; const int32_t s1 = dst->op_params[1]; @@ -6373,6 +7236,10 @@ static void ggml_vk_test_matmul(ggml_backend_vk_context * ctx, size_t m, size_t } } + if (ctx->device->need_compiles) { + ggml_vk_load_shaders(ctx->device); + } + ggml_pipeline_allocate_descriptor_sets(ctx->device); vk_buffer d_X = ggml_vk_create_buffer_check(ctx->device, sizeof(X_TYPE) * x_ne, vk::MemoryPropertyFlagBits::eDeviceLocal); @@ -6422,7 +7289,7 @@ static void ggml_vk_test_matmul(ggml_backend_vk_context * ctx, size_t m, size_t ctx, subctx, p, ggml_vk_subbuffer(d_X), ggml_vk_subbuffer(d_Y), ggml_vk_subbuffer(d_D), ggml_vk_subbuffer(ctx->prealloc_split_k), m, n, k, k, k, m, k*m, k*n, m*n, - split_k, batch, batch, batch, 1, 1 + split_k, batch, batch, batch, 1, 1, n ); } ggml_vk_ctx_end(subctx); @@ -6621,6 +7488,10 @@ static void ggml_vk_test_dequant(ggml_backend_vk_context * ctx, size_t ne, ggml_ ggml_pipeline_request_descriptor_sets(ctx->device, p, 1); + if (ctx->device->need_compiles) { + ggml_vk_load_shaders(ctx->device); + } + ggml_pipeline_allocate_descriptor_sets(ctx->device); ggml_vk_buffer_write(qx_buf, 0, qx, qx_sz); @@ -6680,66 +7551,198 @@ static void ggml_vk_test_dequant(ggml_backend_vk_context * ctx, size_t ne, ggml_ free(x_chk); } -static void ggml_vk_test_dequant_matmul(ggml_backend_vk_context * ctx, size_t m, size_t n, size_t k, size_t batch, size_t num_it, size_t split_k, size_t shader_size, ggml_type quant) { +// This does not work without ggml q8_1 quantization support +// +// typedef uint16_t ggml_half; +// typedef uint32_t ggml_half2; +// +// #define QK8_1 32 +// typedef struct { +// union { +// struct { +// ggml_half d; // delta +// ggml_half s; // d * sum(qs[i]) +// } GGML_COMMON_AGGR_S; +// ggml_half2 ds; +// } GGML_COMMON_AGGR_U; +// int8_t qs[QK8_1]; // quants +// } block_q8_1; +// +// static void ggml_vk_test_quantize(ggml_backend_vk_context * ctx, size_t ne, ggml_type quant) { +// VK_LOG_DEBUG("ggml_vk_test_quantize(" << ne << ")"); +// GGML_ASSERT(quant == GGML_TYPE_Q8_1); +// +// const size_t x_sz = sizeof(float) * ne; +// const size_t qx_sz = ne * ggml_type_size(quant)/ggml_blck_size(quant); +// float * x = (float *) malloc(x_sz); +// block_q8_1 * qx = (block_q8_1 *)malloc(qx_sz); +// block_q8_1 * qx_res = (block_q8_1 *)malloc(qx_sz); +// vk_buffer x_buf = ggml_vk_create_buffer_check(ctx->device, x_sz, vk::MemoryPropertyFlagBits::eDeviceLocal); +// vk_buffer qx_buf = ggml_vk_create_buffer_check(ctx->device, qx_sz, vk::MemoryPropertyFlagBits::eDeviceLocal); +// +// for (size_t i = 0; i < ne; i++) { +// x[i] = rand() / (float)RAND_MAX; +// } +// +// vk_pipeline p = ggml_vk_get_quantize_pipeline(ctx, quant); +// +// ggml_pipeline_request_descriptor_sets(ctx->device, p, 1); +// +// if (ctx->device->need_compiles) { +// ggml_vk_load_shaders(ctx->device); +// } +// +// ggml_pipeline_allocate_descriptor_sets(ctx->device); +// +// ggml_vk_buffer_write(x_buf, 0, x, x_sz); +// +// vk_context subctx = ggml_vk_create_context(ctx, ctx->device->compute_queue); +// ggml_vk_ctx_begin(ctx->device, subctx); +// ggml_vk_quantize_q8_1(ctx, subctx, ggml_vk_subbuffer(x_buf), ggml_vk_subbuffer(qx_buf), ne); +// ggml_vk_ctx_end(subctx); +// +// auto begin = std::chrono::high_resolution_clock::now(); +// +// ggml_vk_submit(subctx, ctx->fence); +// VK_CHECK(ctx->device->device.waitForFences({ ctx->fence }, true, UINT64_MAX), "ggml_vk_test_quantize waitForFences"); +// ctx->device->device.resetFences({ ctx->fence }); +// +// auto end = std::chrono::high_resolution_clock::now(); +// +// double ms_quant = std::chrono::duration_cast(end-begin).count() / 1000.0; +// ggml_vk_buffer_read(qx_buf, 0, qx, qx_sz); +// +// ggml_vk_quantize_data(x, qx_res, ne, quant); +// +// int first_err = -1; +// +// for (size_t i = 0; i < ne / 32; i++) { +// double error = std::fabs(ggml_fp16_to_fp32(qx_res[i].GGML_COMMON_AGGR_U.GGML_COMMON_AGGR_S.d) - ggml_fp16_to_fp32(qx[i].GGML_COMMON_AGGR_U.GGML_COMMON_AGGR_S.d)); +// +// if (first_err < 0 && error > 0.1) { +// first_err = i; +// } +// +// error = std::fabs(ggml_fp16_to_fp32(qx_res[i].GGML_COMMON_AGGR_U.GGML_COMMON_AGGR_S.s) - ggml_fp16_to_fp32(qx[i].GGML_COMMON_AGGR_U.GGML_COMMON_AGGR_S.s)); +// +// if (first_err < 0 && error > 0.1) { +// first_err = i; +// } +// +// for (size_t j = 0; j < 32; j++) { +// uint64_t error = std::abs(qx_res[i].qs[j] - qx[i].qs[j]); +// +// if (first_err < 0 && error > 1) { +// first_err = i; +// } +// } +// } +// +// std::cerr << "TEST QUANTIZE " << ggml_type_name(quant) << " time=" << ms_quant << "ms " << (first_err == -1 ? "CORRECT" : "INCORRECT") << std::endl; +// +// if (first_err != -1) { +// std::cerr << "first_error = " << first_err << std::endl; +// std::cerr << "Actual result: " << std::endl << std::endl; +// std::cout << "d=" << ggml_fp16_to_fp32(qx[first_err].GGML_COMMON_AGGR_U.GGML_COMMON_AGGR_S.d) << " s=" << ggml_fp16_to_fp32(qx[first_err].GGML_COMMON_AGGR_U.GGML_COMMON_AGGR_S.s) << " "; +// for (size_t j = 0; j < 32; j++) { +// std::cout << " qs" << j << "=" << (uint32_t)qx[first_err].qs[j] << " "; +// } +// std::cerr << std::endl << std::endl << "Expected result: " << std::endl << std::endl; +// std::cout << "d=" << ggml_fp16_to_fp32(qx_res[first_err].GGML_COMMON_AGGR_U.GGML_COMMON_AGGR_S.d) << " s=" << ggml_fp16_to_fp32(qx_res[first_err].GGML_COMMON_AGGR_U.GGML_COMMON_AGGR_S.s) << " "; +// for (size_t j = 0; j < 32; j++) { +// std::cout << " qs" << j << "=" << (uint32_t)qx_res[first_err].qs[j] << " "; +// } +// std::cerr << std::endl; +// } +// +// ggml_vk_destroy_buffer(x_buf); +// ggml_vk_destroy_buffer(qx_buf); +// +// free(x); +// free(qx); +// free(qx_res); +// } + +static void ggml_vk_test_dequant_matmul(ggml_backend_vk_context * ctx, size_t m, size_t n, size_t k, size_t batch, size_t num_it, size_t split_k, size_t shader_size, ggml_type quant, bool mmq = false) { VK_LOG_DEBUG("ggml_vk_test_dequant_matmul(" << m << ", " << n << ", " << k << ", " << batch << ", " << num_it << ", " << split_k << ", " << ggml_type_name(quant) << ")"); const size_t x_ne = m * k * batch; const size_t y_ne = k * n * batch; const size_t d_ne = m * n * batch; + vk_matmul_pipeline2 * pipelines; + + if (mmq) { + pipelines = ctx->device->pipeline_dequant_mul_mat_mat_q8_1; + } else { + pipelines = ctx->device->pipeline_dequant_mul_mat_mat; + } + + const bool fp16acc = ctx->device->fp16; + vk_pipeline p; std::string shname; if (shader_size == 0) { - p = ctx->device->fp16 ? ctx->device->pipeline_dequant_mul_mat_mat[quant].f16acc->a_s : ctx->device->pipeline_dequant_mul_mat_mat[quant].f32acc->a_s; + p = fp16acc ? pipelines[quant].f16acc->a_s : pipelines[quant].f32acc->a_s; shname = std::string(ggml_type_name(quant)) + "_ALIGNED_S"; } else if (shader_size == 1) { - p = ctx->device->fp16 ? ctx->device->pipeline_dequant_mul_mat_mat[quant].f16acc->a_m : ctx->device->pipeline_dequant_mul_mat_mat[quant].f32acc->a_m; + p = fp16acc ? pipelines[quant].f16acc->a_m : pipelines[quant].f32acc->a_m; shname = std::string(ggml_type_name(quant)) + "_ALIGNED_M"; } else if (shader_size == 2) { - p = ctx->device->fp16 ? ctx->device->pipeline_dequant_mul_mat_mat[quant].f16acc->a_l : ctx->device->pipeline_dequant_mul_mat_mat[quant].f32acc->a_l; + p = fp16acc ? pipelines[quant].f16acc->a_l : pipelines[quant].f32acc->a_l; shname = std::string(ggml_type_name(quant)) + "_ALIGNED_L"; } else { GGML_ASSERT(0); } - const size_t kpad = ggml_vk_align_size(k, p->align); + const size_t kpad = mmq ? 0 : ggml_vk_align_size(k, p->align); - if (k != kpad) { + if (mmq || k != kpad) { if (shader_size == 0) { - p = ctx->device->fp16 ? ctx->device->pipeline_dequant_mul_mat_mat[quant].f16acc->s : ctx->device->pipeline_dequant_mul_mat_mat[quant].f32acc->s; + p = fp16acc ? pipelines[quant].f16acc->s : pipelines[quant].f32acc->s; shname = std::string(ggml_type_name(quant)) + "_S"; } else if (shader_size == 1) { - p = ctx->device->fp16 ? ctx->device->pipeline_dequant_mul_mat_mat[quant].f16acc->m : ctx->device->pipeline_dequant_mul_mat_mat[quant].f32acc->m; + p = fp16acc ? pipelines[quant].f16acc->m : pipelines[quant].f32acc->m; shname = std::string(ggml_type_name(quant)) + "_M"; } else if (shader_size == 2) { - p = ctx->device->fp16 ? ctx->device->pipeline_dequant_mul_mat_mat[quant].f16acc->l : ctx->device->pipeline_dequant_mul_mat_mat[quant].f32acc->l; + p = fp16acc ? pipelines[quant].f16acc->l : pipelines[quant].f32acc->l; shname = std::string(ggml_type_name(quant)) + "_L"; } else { GGML_ASSERT(0); } } + if (p == nullptr) { + std::cerr << "error: no pipeline for ggml_vk_test_dequant_matmul " << ggml_type_name(quant) << std::endl; + return; + } + const size_t x_sz = sizeof(float) * x_ne; const size_t y_sz = sizeof(float) * y_ne; const size_t qx_sz = x_ne * ggml_type_size(quant)/ggml_blck_size(quant); + const size_t qy_sz = mmq ? y_ne * ggml_type_size(GGML_TYPE_Q8_1)/ggml_blck_size(GGML_TYPE_Q8_1) : y_sz; const size_t d_sz = sizeof(float) * d_ne; float * x = (float *) malloc(x_sz); float * y = (float *) malloc(y_sz); void * qx = malloc(qx_sz); vk_buffer qx_buf = ggml_vk_create_buffer_check(ctx->device, qx_sz, vk::MemoryPropertyFlagBits::eDeviceLocal); vk_buffer y_buf = ggml_vk_create_buffer_check(ctx->device, y_sz, vk::MemoryPropertyFlagBits::eDeviceLocal); + vk_buffer qy_buf = ggml_vk_create_buffer_check(ctx->device, qy_sz, vk::MemoryPropertyFlagBits::eDeviceLocal); vk_buffer d_buf = ggml_vk_create_buffer_check(ctx->device, d_sz, vk::MemoryPropertyFlagBits::eDeviceLocal); float * d = (float *) malloc(d_sz); float * d_chk = (float *) malloc(d_sz); for (size_t i = 0; i < x_ne; i++) { x[i] = (rand() / (float)RAND_MAX) * 2.0f - 1.0f; + // x[i] = (i % k == i / k) ? 1.0f : 0.0f; + // x[i] = i % k; } ggml_vk_quantize_data(x, qx, x_ne, quant); for (size_t i = 0; i < y_ne; i++) { - // y[i] = rand() / (float)RAND_MAX; - y[i] = (i % k == i / k) ? 1.0f : 0.0f; + y[i] = (rand() / (float)RAND_MAX) * 2.0f - 1.0f; + // y[i] = (i % k == i / k) ? 1.0f : 0.0f; + // y[i] = i % k; } ggml_pipeline_request_descriptor_sets(ctx->device, p, num_it); @@ -6754,6 +7757,13 @@ static void ggml_vk_test_dequant_matmul(ggml_backend_vk_context * ctx, size_t m, ctx->prealloc_split_k = ggml_vk_create_buffer_check(ctx->device, sizeof(float) * d_ne * split_k, vk::MemoryPropertyFlagBits::eDeviceLocal); } } + if (mmq) { + ggml_pipeline_request_descriptor_sets(ctx->device, ctx->device->pipeline_quantize_q8_1, num_it); + } + + if (ctx->device->need_compiles) { + ggml_vk_load_shaders(ctx->device); + } ggml_pipeline_allocate_descriptor_sets(ctx->device); @@ -6762,13 +7772,25 @@ static void ggml_vk_test_dequant_matmul(ggml_backend_vk_context * ctx, size_t m, vk_context subctx = ggml_vk_create_context(ctx, ctx->device->compute_queue); ggml_vk_ctx_begin(ctx->device, subctx); - for (size_t i = 0; i < num_it; i++) { - ggml_vk_matmul( - ctx, subctx, p, ggml_vk_subbuffer(qx_buf), ggml_vk_subbuffer(y_buf), ggml_vk_subbuffer(d_buf), ggml_vk_subbuffer(ctx->prealloc_split_k), - m, n, k, - k, k, m, k*m, k*n, m*n, - split_k, batch, batch, batch, 1, 1 - ); + if (mmq) { + for (size_t i = 0; i < num_it; i++) { + ggml_vk_quantize_q8_1(ctx, subctx, { y_buf, 0, y_sz }, { qy_buf, 0, qy_sz }, y_ne); + ggml_vk_matmul( + ctx, subctx, p, { qx_buf, 0, qx_sz }, { qy_buf, 0, qy_sz }, { d_buf, 0, d_sz }, { ctx->prealloc_split_k, 0, ctx->prealloc_size_split_k }, + m, n, k, + k, k, m, k*m, k*n, m*n, + split_k, batch, batch, batch, 1, 1, n + ); + } + } else { + for (size_t i = 0; i < num_it; i++) { + ggml_vk_matmul( + ctx, subctx, p, { qx_buf, 0, qx_sz }, { y_buf, 0, y_sz }, { d_buf, 0, d_sz }, { ctx->prealloc_split_k, 0, ctx->prealloc_size_split_k }, + m, n, k, + k, k, m, k*m, k*n, m*n, + split_k, batch, batch, batch, 1, 1, n + ); + } } ggml_vk_ctx_end(subctx); @@ -6826,7 +7848,11 @@ static void ggml_vk_test_dequant_matmul(ggml_backend_vk_context * ctx, size_t m, double tflops = 2.0*m*n*k*batch*num_it / (time_ms / 1000.0) / (1000.0*1000.0*1000.0*1000.0); - std::cerr << "TEST MMQ " << shname << " m=" << m << " n=" << n << " k=" << k << " batch=" << batch << " split_k=" << split_k << " matmul " << time_ms / num_it << "ms " << tflops << " TFLOPS avg_err=" << avg_err << std::endl; + std::cerr << "TEST dequant matmul " << shname; + if (mmq) { + std::cerr << " mmq"; + } + std::cerr << " m=" << m << " n=" << n << " k=" << k << " batch=" << batch << " split_k=" << split_k << " matmul " << time_ms / num_it << "ms " << tflops << " TFLOPS avg_err=" << avg_err << std::endl; if (avg_err > 0.01 || std::isnan(avg_err)) { std::cerr << "m = " << first_err_m << " n = " << first_err_n << " b = " << first_err_b << std::endl; @@ -6836,6 +7862,12 @@ static void ggml_vk_test_dequant_matmul(ggml_backend_vk_context * ctx, size_t m, std::cerr << "Expected result: " << std::endl << std::endl; ggml_vk_print_matrix_area(d_chk, GGML_TYPE_F32, m, n, first_err_m, first_err_n, first_err_b); + std::cerr << "src0: " << std::endl << std::endl; + ggml_vk_print_matrix_area(x, GGML_TYPE_F32, k, m, first_err_m, first_err_n, first_err_b); + std::cerr << std::endl; + std::cerr << "src1: " << std::endl << std::endl; + ggml_vk_print_matrix_area(y, GGML_TYPE_F32, k, n, first_err_m, first_err_n, first_err_b); + if (split_k > 1) { float * split_k_buf = (float *) malloc(sizeof(float) * d_ne * split_k); ggml_vk_buffer_read(ctx->prealloc_split_k, 0, split_k_buf, sizeof(float) * d_ne * split_k); @@ -6858,6 +7890,7 @@ static void ggml_vk_test_dequant_matmul(ggml_backend_vk_context * ctx, size_t m, ggml_vk_destroy_buffer(qx_buf); ggml_vk_destroy_buffer(y_buf); + ggml_vk_destroy_buffer(qy_buf); ggml_vk_destroy_buffer(d_buf); free(x); @@ -6892,6 +7925,24 @@ static void ggml_vk_preallocate_buffers(ggml_backend_vk_context * ctx) { }; const size_t num_it = 100; + ggml_vk_test_dequant_matmul(ctx, 4096, 512, 4096, 2, num_it, 1, 0, GGML_TYPE_Q4_0); + ggml_vk_test_dequant_matmul(ctx, 4096, 512, 4096, 2, num_it, 1, 1, GGML_TYPE_Q4_0); + ggml_vk_test_dequant_matmul(ctx, 4096, 512, 4096, 2, num_it, 1, 2, GGML_TYPE_Q4_0); + + ggml_vk_test_dequant_matmul(ctx, 4096, 512, 4096, 2, num_it, 1, 0, GGML_TYPE_Q4_0, true); + ggml_vk_test_dequant_matmul(ctx, 4096, 512, 4096, 2, num_it, 1, 1, GGML_TYPE_Q4_0, true); + ggml_vk_test_dequant_matmul(ctx, 4096, 512, 4096, 2, num_it, 1, 2, GGML_TYPE_Q4_0, true); + + ggml_vk_test_dequant_matmul(ctx, 4096, 512, 4096, 2, num_it, 1, 0, GGML_TYPE_Q8_0); + ggml_vk_test_dequant_matmul(ctx, 4096, 512, 4096, 2, num_it, 1, 1, GGML_TYPE_Q8_0); + ggml_vk_test_dequant_matmul(ctx, 4096, 512, 4096, 2, num_it, 1, 2, GGML_TYPE_Q8_0); + + ggml_vk_test_dequant_matmul(ctx, 4096, 512, 4096, 2, num_it, 1, 0, GGML_TYPE_Q8_0, true); + ggml_vk_test_dequant_matmul(ctx, 4096, 512, 4096, 2, num_it, 1, 1, GGML_TYPE_Q8_0, true); + ggml_vk_test_dequant_matmul(ctx, 4096, 512, 4096, 2, num_it, 1, 2, GGML_TYPE_Q8_0, true); + + abort(); + for (size_t i = 0; i < vals.size(); i += 3) { ggml_vk_test_matmul(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 1, 0); ggml_vk_test_matmul(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 1, 1); @@ -6966,11 +8017,11 @@ static void ggml_vk_preallocate_buffers(ggml_backend_vk_context * ctx) { } } -static bool ggml_vk_compute_forward(ggml_backend_vk_context* ctx, ggml_tensor* tensor, int tensor_idx, bool use_fence); +static bool ggml_vk_compute_forward(ggml_backend_vk_context* ctx, ggml_tensor* tensor, int tensor_idx, bool use_fence, bool almost_ready); // Returns true if node has enqueued work into the queue, false otherwise // If submit is true the current all operations queued so far are being submitted to Vulkan to overlap cmdlist creation and GPU execution. -static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * node, int node_idx, ggml_tensor *node_begin, int node_idx_begin, bool dryrun, bool last_node, bool submit){ +static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * node, int node_idx, ggml_tensor *node_begin, int node_idx_begin, bool dryrun, bool last_node, bool almost_ready, bool submit){ if (ggml_is_empty(node) || !node->buffer) { return false; } @@ -6998,15 +8049,18 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod case GGML_UNARY_OP_GELU_QUICK: case GGML_UNARY_OP_RELU: case GGML_UNARY_OP_TANH: + case GGML_UNARY_OP_SIGMOID: break; default: return false; } break; case GGML_OP_REPEAT: + case GGML_OP_REPEAT_BACK: case GGML_OP_GET_ROWS: case GGML_OP_ADD: case GGML_OP_ACC: + case GGML_OP_SUB: case GGML_OP_MUL: case GGML_OP_DIV: case GGML_OP_CONCAT: @@ -7020,22 +8074,32 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod case GGML_OP_CPY: case GGML_OP_CONT: case GGML_OP_DUP: + case GGML_OP_SILU_BACK: case GGML_OP_NORM: case GGML_OP_GROUP_NORM: case GGML_OP_RMS_NORM: + case GGML_OP_RMS_NORM_BACK: + case GGML_OP_L2_NORM: case GGML_OP_DIAG_MASK_INF: case GGML_OP_SOFT_MAX: + case GGML_OP_SOFT_MAX_BACK: case GGML_OP_ROPE: + case GGML_OP_ROPE_BACK: case GGML_OP_MUL_MAT: case GGML_OP_MUL_MAT_ID: case GGML_OP_ARGSORT: + case GGML_OP_SUM: case GGML_OP_SUM_ROWS: + case GGML_OP_ARGMAX: + case GGML_OP_COUNT_EQUAL: case GGML_OP_IM2COL: case GGML_OP_TIMESTEP_EMBEDDING: case GGML_OP_POOL_2D: case GGML_OP_RWKV_WKV6: + case GGML_OP_RWKV_WKV7: case GGML_OP_LEAKY_RELU: case GGML_OP_FLASH_ATTN_EXT: + case GGML_OP_OPT_STEP_ADAMW: break; default: std::cerr << "ggml_vulkan: Error: Missing op: " << ggml_op_name(node->op) << std::endl; @@ -7056,9 +8120,11 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod } else { switch (node->op) { case GGML_OP_REPEAT: + case GGML_OP_REPEAT_BACK: case GGML_OP_ACC: case GGML_OP_GET_ROWS: case GGML_OP_ADD: + case GGML_OP_SUB: case GGML_OP_MUL: case GGML_OP_DIV: case GGML_OP_CONCAT: @@ -7072,15 +8138,23 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod case GGML_OP_CPY: case GGML_OP_CONT: case GGML_OP_DUP: + case GGML_OP_SILU_BACK: case GGML_OP_NORM: case GGML_OP_GROUP_NORM: case GGML_OP_RMS_NORM: + case GGML_OP_RMS_NORM_BACK: + case GGML_OP_L2_NORM: case GGML_OP_UNARY: case GGML_OP_DIAG_MASK_INF: case GGML_OP_SOFT_MAX: + case GGML_OP_SOFT_MAX_BACK: case GGML_OP_ROPE: + case GGML_OP_ROPE_BACK: case GGML_OP_ARGSORT: + case GGML_OP_SUM: case GGML_OP_SUM_ROWS: + case GGML_OP_ARGMAX: + case GGML_OP_COUNT_EQUAL: case GGML_OP_IM2COL: case GGML_OP_TIMESTEP_EMBEDDING: case GGML_OP_POOL_2D: @@ -7101,6 +8175,10 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod case GGML_OP_REPEAT: ggml_vk_repeat(ctx, compute_ctx, src0, node, dryrun); + break; + case GGML_OP_REPEAT_BACK: + ggml_vk_repeat_back(ctx, compute_ctx, src0, node, dryrun); + break; case GGML_OP_ACC: ggml_vk_acc(ctx, compute_ctx, src0, src1, node, dryrun); @@ -7113,6 +8191,10 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod case GGML_OP_ADD: ggml_vk_add(ctx, compute_ctx, src0, src1, node, dryrun); + break; + case GGML_OP_SUB: + ggml_vk_sub(ctx, compute_ctx, src0, src1, node, dryrun); + break; case GGML_OP_MUL: ggml_vk_mul(ctx, compute_ctx, src0, src1, node, dryrun); @@ -7159,6 +8241,10 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod case GGML_OP_DUP: ggml_vk_cpy(ctx, compute_ctx, src0, node, dryrun); + break; + case GGML_OP_SILU_BACK: + ggml_vk_silu_back(ctx, compute_ctx, src0, src1, node, dryrun); + break; case GGML_OP_NORM: ggml_vk_norm(ctx, compute_ctx, src0, node, dryrun); @@ -7171,6 +8257,14 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod case GGML_OP_RMS_NORM: ggml_vk_rms_norm(ctx, compute_ctx, src0, node, dryrun); + break; + case GGML_OP_RMS_NORM_BACK: + ggml_vk_rms_norm_back(ctx, compute_ctx, src0, src1, node, dryrun); + + break; + case GGML_OP_L2_NORM: + ggml_vk_l2_norm(ctx, compute_ctx, src0, node, dryrun); + break; case GGML_OP_UNARY: switch (ggml_get_unary_op(node)) { @@ -7179,6 +8273,7 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod case GGML_UNARY_OP_GELU_QUICK: case GGML_UNARY_OP_RELU: case GGML_UNARY_OP_TANH: + case GGML_UNARY_OP_SIGMOID: ggml_vk_unary(ctx, compute_ctx, src0, node, dryrun); break; default: @@ -7192,18 +8287,38 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod case GGML_OP_SOFT_MAX: ggml_vk_soft_max(ctx, compute_ctx, src0, src1, node, dryrun); + break; + case GGML_OP_SOFT_MAX_BACK: + ggml_vk_soft_max_back(ctx, compute_ctx, src0, src1, node, dryrun); + break; case GGML_OP_ROPE: - ggml_vk_rope(ctx, compute_ctx, src0, src1, src2, node, dryrun); + ggml_vk_rope(ctx, compute_ctx, src0, src1, src2, node, false, dryrun); + + break; + case GGML_OP_ROPE_BACK: + ggml_vk_rope(ctx, compute_ctx, src0, src1, src2, node, true, dryrun); break; case GGML_OP_ARGSORT: ggml_vk_argsort(ctx, compute_ctx, src0, node, dryrun); + break; + case GGML_OP_SUM: + ggml_vk_sum(ctx, compute_ctx, src0, node, dryrun); + break; case GGML_OP_SUM_ROWS: ggml_vk_sum_rows(ctx, compute_ctx, src0, node, dryrun); + break; + case GGML_OP_ARGMAX: + ggml_vk_argmax(ctx, compute_ctx, src0, node, dryrun); + + break; + case GGML_OP_COUNT_EQUAL: + ggml_vk_count_equal(ctx, compute_ctx, src0, src1, node, dryrun); + break; case GGML_OP_IM2COL: ggml_vk_im2col(ctx, compute_ctx, src0, src1, node, dryrun); @@ -7238,6 +8353,16 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod case GGML_OP_RWKV_WKV6: ggml_vk_rwkv_wkv6(ctx, compute_ctx, node, dryrun); + break; + + case GGML_OP_RWKV_WKV7: + ggml_vk_rwkv_wkv7(ctx, compute_ctx, node, dryrun); + + break; + + case GGML_OP_OPT_STEP_ADAMW: + ggml_vk_opt_step_adamw(ctx, compute_ctx, node, dryrun); + break; default: return false; @@ -7268,7 +8393,7 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod ctx->compute_ctx.reset(); - bool ok = ggml_vk_compute_forward(ctx, node_begin, node_idx_begin, false); + bool ok = ggml_vk_compute_forward(ctx, node_begin, node_idx_begin, false, almost_ready); if (!ok) { if (node->op == GGML_OP_UNARY) { std::cerr << __func__ << ": error: op not supported UNARY " << node->name << " (" << ggml_unary_op_name(static_cast(node->op_params[0])) << ")" << std::endl; @@ -7282,13 +8407,14 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod return true; } -static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_tensor * tensor, int tensor_idx, bool use_fence = true){ +static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_tensor * tensor, int tensor_idx, bool use_fence = true, bool almost_ready = false) { ggml_backend_buffer * buf = nullptr; switch (tensor->op) { case GGML_OP_ADD: case GGML_OP_ACC: case GGML_OP_GET_ROWS: + case GGML_OP_SUB: case GGML_OP_MUL: case GGML_OP_DIV: case GGML_OP_CONCAT: @@ -7302,25 +8428,36 @@ static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_tensor * case GGML_OP_CPY: case GGML_OP_CONT: case GGML_OP_DUP: + case GGML_OP_SILU_BACK: case GGML_OP_NORM: case GGML_OP_GROUP_NORM: case GGML_OP_RMS_NORM: + case GGML_OP_RMS_NORM_BACK: + case GGML_OP_L2_NORM: case GGML_OP_DIAG_MASK_INF: case GGML_OP_SOFT_MAX: + case GGML_OP_SOFT_MAX_BACK: case GGML_OP_ROPE: + case GGML_OP_ROPE_BACK: case GGML_OP_RESHAPE: case GGML_OP_VIEW: case GGML_OP_PERMUTE: case GGML_OP_TRANSPOSE: case GGML_OP_NONE: case GGML_OP_ARGSORT: + case GGML_OP_SUM: case GGML_OP_SUM_ROWS: + case GGML_OP_ARGMAX: + case GGML_OP_COUNT_EQUAL: case GGML_OP_IM2COL: case GGML_OP_TIMESTEP_EMBEDDING: case GGML_OP_POOL_2D: case GGML_OP_RWKV_WKV6: + case GGML_OP_RWKV_WKV7: case GGML_OP_LEAKY_RELU: case GGML_OP_REPEAT: + case GGML_OP_REPEAT_BACK: + case GGML_OP_OPT_STEP_ADAMW: buf = tensor->buffer; break; @@ -7331,6 +8468,7 @@ static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_tensor * case GGML_UNARY_OP_GELU_QUICK: case GGML_UNARY_OP_RELU: case GGML_UNARY_OP_TANH: + case GGML_UNARY_OP_SIGMOID: buf = tensor->buffer; break; default: @@ -7372,12 +8510,15 @@ static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_tensor * memcpy(cpy.dst, cpy.src, cpy.n); } - ggml_vk_submit(subctx, use_fence ? ctx->fence : vk::Fence{}); + if (almost_ready && !ctx->almost_ready_fence_pending && !use_fence) { + ggml_vk_submit(subctx, ctx->almost_ready_fence); + ctx->almost_ready_fence_pending = true; + } else { + ggml_vk_submit(subctx, use_fence ? ctx->fence : vk::Fence{}); + } if (use_fence) { - VK_CHECK(ctx->device->device.waitForFences({ ctx->fence }, true, UINT64_MAX), "ggml_vk_compute_forward waitForFences"); - - ctx->device->device.resetFences({ ctx->fence }); + ggml_vk_wait_for_fence(ctx); } #ifdef GGML_VULKAN_CHECK_RESULTS ggml_vk_check_results_1(tensor); @@ -7463,6 +8604,7 @@ static void ggml_vk_cleanup(ggml_backend_vk_context * ctx) { ctx->gc.events.clear(); ctx->device->device.destroyFence(ctx->fence); + ctx->device->device.destroyFence(ctx->almost_ready_fence); } static int ggml_vk_get_device_count() { @@ -7505,11 +8647,21 @@ static void * ggml_backend_vk_buffer_get_base(ggml_backend_buffer_t buffer) { UNUSED(buffer); } -static void ggml_backend_vk_buffer_init_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor) { +static enum ggml_status ggml_backend_vk_buffer_init_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor) { VK_LOG_DEBUG("ggml_backend_vk_buffer_init_tensor(" << buffer << " (" << buffer->context << "), " << tensor << ")"); if (tensor->view_src != nullptr) { GGML_ASSERT(tensor->view_src->buffer->buft == buffer->buft); } + return GGML_STATUS_SUCCESS; +} + +static void ggml_backend_vk_buffer_memset_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor, uint8_t value, size_t offset, size_t size) { + VK_LOG_DEBUG("ggml_backend_vk_buffer_memset_tensor(" << buffer << ", " << tensor << ", " << value << ", " << offset << ", " << size << ")"); + ggml_backend_vk_buffer_context * buf_ctx = (ggml_backend_vk_buffer_context *)buffer->context; + vk_buffer buf = buf_ctx->dev_buffer; + + uint32_t val32 = (uint32_t)value * 0x01010101; + ggml_vk_buffer_memset(buf, vk_tensor_offset(tensor) + tensor->view_offs + offset, val32, size); } static void ggml_backend_vk_buffer_set_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor, const void * data, size_t offset, size_t size) { @@ -7556,7 +8708,7 @@ static ggml_backend_buffer_i ggml_backend_vk_buffer_interface = { /* .free_buffer = */ ggml_backend_vk_buffer_free_buffer, /* .get_base = */ ggml_backend_vk_buffer_get_base, /* .init_tensor = */ ggml_backend_vk_buffer_init_tensor, - /* .memset_tensor = */ NULL, + /* .memset_tensor = */ ggml_backend_vk_buffer_memset_tensor, /* .set_tensor = */ ggml_backend_vk_buffer_set_tensor, /* .get_tensor = */ ggml_backend_vk_buffer_get_tensor, /* .cpy_tensor = */ ggml_backend_vk_buffer_cpy_tensor, @@ -7799,8 +8951,7 @@ static void ggml_backend_vk_synchronize(ggml_backend_t backend) { } ggml_vk_submit(transfer_ctx, ctx->fence); - VK_CHECK(ctx->device->device.waitForFences({ ctx->fence }, true, UINT64_MAX), "ggml_backend_vk_synchronize waitForFences"); - ctx->device->device.resetFences({ ctx->fence }); + ggml_vk_wait_for_fence(ctx); for (auto& cpy : transfer_ctx->out_memcpys) { memcpy(cpy.dst, cpy.src, cpy.n); @@ -7817,8 +8968,12 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg VK_LOG_DEBUG("ggml_backend_vk_graph_compute(" << cgraph->n_nodes << " nodes)"); ggml_backend_vk_context * ctx = (ggml_backend_vk_context *)backend->context; + uint64_t total_mat_mul_bytes = 0; for (int i = 0; i < cgraph->n_nodes; i++) { - ggml_vk_build_graph(ctx, cgraph->nodes[i], i, nullptr, 0, true, false, false); + ggml_vk_build_graph(ctx, cgraph->nodes[i], i, nullptr, 0, true, false, false, false); + if (cgraph->nodes[i]->op == GGML_OP_MUL_MAT || cgraph->nodes[i]->op == GGML_OP_MUL_MAT_ID) { + total_mat_mul_bytes += ggml_nbytes(cgraph->nodes[i]->src[0]); + } } if (ctx->device->need_compiles) { ggml_vk_load_shaders(ctx->device); @@ -7839,19 +8994,32 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg bool first_node_in_batch = true; // true if next node will be first node in a batch int submit_node_idx = 0; // index to first node in a batch - // Submit work every nodes_per_submit nodes to overlap CPU cmdbuffer generation with GPU execution. - // Start with a smaller count to get work submitted right away, and increase it after each submit. - int nodes_per_submit = 20; + // Submit after enough work has accumulated, to overlap CPU cmdbuffer generation with GPU execution. + // Estimate the amount of matmul work by looking at the weight matrix size, and submit every 100MB + // (and scaled down based on model size, so smaller models submit earlier). + // Also submit at least every 100 nodes, in case there are workloads without as much matmul. + int nodes_per_submit = 100; int submitted_nodes = 0; int submit_count = 0; + uint64_t mul_mat_bytes = 0; + uint64_t mul_mat_bytes_per_submit = std::min(uint64_t(100*1000*1000), total_mat_mul_bytes / 40u); for (int i = 0; i < cgraph->n_nodes; i++) { if (first_node_in_batch) { submit_node_idx = i; } - bool submit = (submitted_nodes >= nodes_per_submit) || (i == last_node); + if (cgraph->nodes[i]->op == GGML_OP_MUL_MAT || cgraph->nodes[i]->op == GGML_OP_MUL_MAT_ID) { + mul_mat_bytes += ggml_nbytes(cgraph->nodes[i]->src[0]); + } - bool enqueued = ggml_vk_build_graph(ctx, cgraph->nodes[i], i, cgraph->nodes[submit_node_idx], submit_node_idx, false, i == last_node, submit); + // Signal the almost_ready fence when the graph is mostly complete (< 20% remaining) + bool almost_ready = (cgraph->n_nodes - i) < cgraph->n_nodes / 5; + bool submit = (submitted_nodes >= nodes_per_submit) || + (mul_mat_bytes >= mul_mat_bytes_per_submit) || + (i == last_node) || + (almost_ready && !ctx->almost_ready_fence_pending); + + bool enqueued = ggml_vk_build_graph(ctx, cgraph->nodes[i], i, cgraph->nodes[submit_node_idx], submit_node_idx, false, i == last_node, almost_ready, submit); if (enqueued) { ++submitted_nodes; @@ -7863,16 +9031,12 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg #endif } - if (submit) { + if (submit && enqueued) { first_node_in_batch = true; submitted_nodes = 0; - switch (submit_count) { - case 0: - nodes_per_submit = 50; - break; - default: - nodes_per_submit = 100; - break; + mul_mat_bytes = 0; + if (submit_count < 3) { + mul_mat_bytes_per_submit *= 2; } submit_count++; } @@ -8023,7 +9187,8 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm case GGML_UNARY_OP_SILU: case GGML_UNARY_OP_RELU: case GGML_UNARY_OP_TANH: - return ggml_is_contiguous(op->src[0]); + case GGML_UNARY_OP_SIGMOID: + return ggml_is_contiguous(op->src[0]) && op->src[0]->type == GGML_TYPE_F32; default: return false; } @@ -8051,6 +9216,8 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm case GGML_TYPE_Q4_K: case GGML_TYPE_Q5_K: case GGML_TYPE_Q6_K: + case GGML_TYPE_IQ1_S: + case GGML_TYPE_IQ1_M: case GGML_TYPE_IQ2_XXS: case GGML_TYPE_IQ2_XS: case GGML_TYPE_IQ2_S: @@ -8098,6 +9265,10 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm default: return false; } + if (op->src[1]->ne[0] != op->src[2]->ne[0]) { + // different head sizes of K and V are not supported yet + return false; + } if (op->src[0]->type != GGML_TYPE_F32) { return false; } @@ -8125,6 +9296,8 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm //case GGML_TYPE_Q4_K: //case GGML_TYPE_Q5_K: //case GGML_TYPE_Q6_K: + //case GGML_TYPE_IQ1_S: + //case GGML_TYPE_IQ1_M: //case GGML_TYPE_IQ2_XXS: //case GGML_TYPE_IQ2_XS: //case GGML_TYPE_IQ2_S: @@ -8148,6 +9321,8 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm case GGML_TYPE_Q5_0: case GGML_TYPE_Q5_1: case GGML_TYPE_Q8_0: + case GGML_TYPE_IQ1_S: + case GGML_TYPE_IQ1_M: case GGML_TYPE_IQ2_XXS: case GGML_TYPE_IQ2_XS: case GGML_TYPE_IQ2_S: @@ -8203,17 +9378,10 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm } break; case GGML_OP_REPEAT: return ggml_type_size(op->type) == sizeof(float) && ggml_type_size(op->src[0]->type) == sizeof(float); + case GGML_OP_REPEAT_BACK: + return op->type == GGML_TYPE_F32 && op->src[0]->type == GGML_TYPE_F32; case GGML_OP_ROPE: - { - const int mode = ((const int32_t *) op->op_params)[2]; - if (mode & GGML_ROPE_TYPE_MROPE) { - return false; - } - if (mode & GGML_ROPE_TYPE_VISION) { - return false; - } - return ggml_is_contiguous(op->src[0]); - } + case GGML_OP_ROPE_BACK: case GGML_OP_NONE: case GGML_OP_RESHAPE: case GGML_OP_VIEW: @@ -8223,28 +9391,40 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm case GGML_OP_NORM: case GGML_OP_GROUP_NORM: case GGML_OP_RMS_NORM: + case GGML_OP_L2_NORM: return ggml_is_contiguous(op->src[0]); case GGML_OP_ADD: - case GGML_OP_ACC: + case GGML_OP_SUB: case GGML_OP_MUL: case GGML_OP_DIV: - case GGML_OP_CONCAT: - case GGML_OP_UPSCALE: - case GGML_OP_SCALE: + case GGML_OP_SILU_BACK: + case GGML_OP_RMS_NORM_BACK: case GGML_OP_SQR: case GGML_OP_SIN: case GGML_OP_COS: case GGML_OP_CLAMP: + return op->src[0]->type == GGML_TYPE_F32; + case GGML_OP_UPSCALE: + return op->op_params[0] == GGML_SCALE_MODE_NEAREST; + case GGML_OP_ACC: + case GGML_OP_CONCAT: + case GGML_OP_SCALE: case GGML_OP_PAD: case GGML_OP_DIAG_MASK_INF: case GGML_OP_SOFT_MAX: + case GGML_OP_SOFT_MAX_BACK: case GGML_OP_ARGSORT: + case GGML_OP_SUM: case GGML_OP_SUM_ROWS: + case GGML_OP_ARGMAX: + case GGML_OP_COUNT_EQUAL: case GGML_OP_IM2COL: case GGML_OP_TIMESTEP_EMBEDDING: case GGML_OP_POOL_2D: case GGML_OP_RWKV_WKV6: + case GGML_OP_RWKV_WKV7: case GGML_OP_LEAKY_RELU: + case GGML_OP_OPT_STEP_ADAMW: return true; default: return false; @@ -8344,8 +9524,13 @@ ggml_backend_reg_t ggml_backend_vk_reg() { /* .iface = */ ggml_backend_vk_reg_i, /* .context = */ nullptr, }; - - return ® + try { + ggml_vk_instance_init(); + return ® + } catch (const vk::SystemError& e) { + VK_LOG_DEBUG("ggml_backend_vk_reg() -> Error: System error: " << e.what()); + return nullptr; + } } // Extension availability @@ -8384,7 +9569,7 @@ static bool ggml_vk_instance_portability_enumeration_ext_available(const std::ve UNUSED(instance_extensions); } -static bool ggml_vk_khr_cooperative_matrix_support(const vk::PhysicalDeviceProperties& props, const vk::PhysicalDeviceDriverProperties& driver_props) { +static bool ggml_vk_khr_cooperative_matrix_support(const vk::PhysicalDeviceProperties& props, const vk::PhysicalDeviceDriverProperties& driver_props, vk_device_architecture arch) { switch (props.vendorID) { case VK_VENDOR_ID_INTEL: // Intel drivers don't support coopmat properly yet @@ -8392,10 +9577,7 @@ static bool ggml_vk_khr_cooperative_matrix_support(const vk::PhysicalDevicePrope case VK_VENDOR_ID_AMD: if (driver_props.driverID == vk::DriverId::eAmdProprietary || driver_props.driverID == vk::DriverId::eAmdOpenSource) { // Workaround for AMD proprietary driver reporting support on all GPUs - const std::string name = props.deviceName; - return name.rfind("AMD Radeon RX 7", 0) == 0 || name.rfind("AMD Radeon(TM) RX 7", 0) == 0 || // RDNA 3 consumer GPUs - name.rfind("AMD Radeon PRO W7", 0) == 0 || name.rfind("AMD Radeon(TM) PRO W7", 0) == 0 || // RDNA 3 workstation GPUs - name.rfind("AMD Radeon 7", 0) == 0 || name.rfind("AMD Radeon(TM) 7", 0) == 0; // RDNA 3 APUs + return arch == vk_device_architecture::AMD_RDNA3; } return true; default: @@ -8512,8 +9694,6 @@ static void ggml_vk_check_results_0(ggml_tensor * tensor) { ggml_tensor * src0 = tensor->src[0]; ggml_tensor * src1 = tensor->src[1]; - ggml_tensor * src2 = tensor->src[2]; - ggml_tensor * src3 = tensor->src[3]; struct ggml_init_params iparams = { /*.mem_size =*/ 2ul*1024ul*1024ul*1024ul, @@ -8523,239 +9703,128 @@ static void ggml_vk_check_results_0(ggml_tensor * tensor) { struct ggml_context * ggml_ctx = ggml_init(iparams); - struct ggml_tensor * src0_clone = nullptr; - struct ggml_tensor * src1_clone = nullptr; - struct ggml_tensor * src2_clone = nullptr; - struct ggml_tensor * src3_clone = nullptr; + std::array src_clone = {nullptr, nullptr, nullptr, nullptr, nullptr, nullptr}; + std::array src_size = {0, 0, 0, 0, 0, 0}; + std::array src_buffer = {nullptr, nullptr, nullptr, nullptr, nullptr, nullptr}; + const char * srci_name[6] = {"src0", "src1", "src2", "src3", "src4", "src5"}; + struct ggml_tensor * tensor_clone = nullptr; - size_t src0_size; - size_t src1_size; - size_t src2_size; - size_t src3_size; + for (int i = 0; i < 6; i++) { + ggml_tensor * srci = tensor->src[i]; + if (srci == nullptr) { + continue; + } + ggml_tensor * srci_clone = ggml_dup_tensor(ggml_ctx, srci); + size_t srci_size = ggml_nbytes(srci); - void * src0_buffer = nullptr; - void * src1_buffer = nullptr; - void * src2_buffer = nullptr; - void * src3_buffer = nullptr; + src_clone[i] = srci_clone; + src_size[i] = ggml_nbytes(srci); + src_buffer[i] = malloc(srci_size); - if (src0 != nullptr) { - src0_clone = ggml_dup_tensor(ggml_ctx, src0); - - src0_size = ggml_nbytes(src0); - - src0_buffer = malloc(src0_size); - src0_clone->data = src0_buffer; - if (ggml_backend_buffer_is_host(src0->buffer)) { - memcpy(src0_clone->data, src0->data, src0_size); - memcpy(src0_clone->nb, src0->nb, sizeof(size_t) * GGML_MAX_DIMS); - } else if (ggml_backend_buffer_is_vk(src0->buffer)) { - ggml_backend_vk_buffer_context * buf_ctx = (ggml_backend_vk_buffer_context *)src0->buffer->context; + srci_clone->data = src_buffer[i]; + if (ggml_backend_buffer_is_host(srci->buffer)) { + memcpy(srci_clone->data, srci->data, srci_size); + memcpy(srci_clone->nb, srci->nb, sizeof(size_t) * GGML_MAX_DIMS); + } else if (ggml_backend_buffer_is_vk(srci->buffer)) { + ggml_backend_vk_buffer_context * buf_ctx = (ggml_backend_vk_buffer_context *)srci->buffer->context; vk_buffer& buffer_gpu = buf_ctx->dev_buffer; - uint64_t offset = vk_tensor_offset(src0) + src0->view_offs; - if (!ggml_is_contiguous(src0) && ggml_vk_dim01_contiguous(src0)) { - for (int i3 = 0; i3 < src0->ne[3]; i3++) { - for (int i2 = 0; i2 < src0->ne[2]; i2++) { - const int idx = i3*src0->ne[2] + i2; - ggml_vk_buffer_read(buffer_gpu, offset + idx * src0->nb[2], ((char *)src0_clone->data + idx * src0_clone->nb[2]), src0->ne[1] * src0->nb[1]); + uint64_t offset = vk_tensor_offset(srci) + srci->view_offs; + if (!ggml_is_contiguous(srci) && ggml_vk_dim01_contiguous(srci)) { + for (int i3 = 0; i3 < srci->ne[3]; i3++) { + for (int i2 = 0; i2 < srci->ne[2]; i2++) { + const int idx = i3*srci->ne[2] + i2; + ggml_vk_buffer_read(buffer_gpu, offset + idx * srci->nb[2], ((char *)srci_clone->data + idx * srci_clone->nb[2]), srci->ne[1] * srci->nb[1]); } } - src0_clone->nb[0] = src0->nb[0]; - src0_clone->nb[1] = src0->nb[1]; + srci_clone->nb[0] = srci->nb[0]; + srci_clone->nb[1] = srci->nb[1]; for (int i = 2; i < GGML_MAX_DIMS; i++) { - src0_clone->nb[i] = src0_clone->nb[i - 1]*src0_clone->ne[i - 1]; + srci_clone->nb[i] = srci_clone->nb[i - 1]*srci_clone->ne[i - 1]; } } else { - if (offset + src0_size >= buffer_gpu->size) { - src0_size = buffer_gpu->size - offset; + if (offset + srci_size >= buffer_gpu->size) { + srci_size = buffer_gpu->size - offset; } - ggml_vk_buffer_read(buffer_gpu, offset, src0_clone->data, src0_size); - memcpy(src0_clone->nb, src0->nb, sizeof(size_t) * GGML_MAX_DIMS); + ggml_vk_buffer_read(buffer_gpu, offset, srci_clone->data, srci_size); + memcpy(srci_clone->nb, srci->nb, sizeof(size_t) * GGML_MAX_DIMS); } } else { GGML_ABORT("fatal error"); } if (vk_output_tensor > 0 && vk_output_tensor == check_counter) { - ggml_vk_print_tensor(src0, "src0"); - } - } - if (src1 != nullptr) { - src1_clone = ggml_dup_tensor(ggml_ctx, src1); - - src1_size = ggml_nbytes(src1); - - src1_buffer = malloc(src1_size); - src1_clone->data = src1_buffer; - if (ggml_backend_buffer_is_host(src1->buffer)) { - memcpy(src1_clone->data, src1->data, src1_size); - memcpy(src1_clone->nb, src1->nb, sizeof(size_t) * GGML_MAX_DIMS); - } else if (ggml_backend_buffer_is_vk(src1->buffer)) { - ggml_backend_vk_buffer_context * buf_ctx = (ggml_backend_vk_buffer_context *)src1->buffer->context; - vk_buffer& buffer_gpu = buf_ctx->dev_buffer; - uint64_t offset = vk_tensor_offset(src1) + src1->view_offs; - if (!ggml_is_contiguous(src1) && ggml_vk_dim01_contiguous(src1)) { - for (int i3 = 0; i3 < src1->ne[3]; i3++) { - for (int i2 = 0; i2 < src1->ne[2]; i2++) { - const int idx = i3*src1->ne[2] + i2; - ggml_vk_buffer_read(buffer_gpu, offset + idx * src1->nb[2], ((char *)src1_clone->data + idx * src1_clone->nb[2]), src1->ne[1] * src1->nb[1]); - } - } - - src1_clone->nb[0] = src1->nb[0]; - src1_clone->nb[1] = src1->nb[1]; - for (int i = 2; i < GGML_MAX_DIMS; i++) { - src1_clone->nb[i] = src1_clone->nb[i - 1]*src1_clone->ne[i - 1]; - } - } else { - if (offset + src1_size >= buffer_gpu->size) { - src1_size = buffer_gpu->size - offset; - } - ggml_vk_buffer_read(buffer_gpu, offset, src1_clone->data, src1_size); - memcpy(src1_clone->nb, src1->nb, sizeof(size_t) * GGML_MAX_DIMS); - } - } else { - GGML_ABORT("fatal error"); - } - - if (vk_output_tensor > 0 && vk_output_tensor == check_counter) { - ggml_vk_print_tensor(src1, "src1"); - } - } - if (src2 != nullptr) { - src2_clone = ggml_dup_tensor(ggml_ctx, src2); - - src2_size = ggml_nbytes(src2); - - src2_buffer = malloc(src2_size); - src2_clone->data = src2_buffer; - if (ggml_backend_buffer_is_host(src2->buffer)) { - memcpy(src2_clone->data, src2->data, src2_size); - memcpy(src2_clone->nb, src2->nb, sizeof(size_t) * GGML_MAX_DIMS); - } else if (ggml_backend_buffer_is_vk(src2->buffer)) { - ggml_backend_vk_buffer_context * buf_ctx = (ggml_backend_vk_buffer_context *)src2->buffer->context; - vk_buffer& buffer_gpu = buf_ctx->dev_buffer; - uint64_t offset = vk_tensor_offset(src2) + src2->view_offs; - if (!ggml_is_contiguous(src2) && ggml_vk_dim01_contiguous(src2)) { - for (int i3 = 0; i3 < src2->ne[3]; i3++) { - for (int i2 = 0; i2 < src2->ne[2]; i2++) { - const int idx = i3*src2->ne[2] + i2; - ggml_vk_buffer_read(buffer_gpu, offset + idx * src2->nb[2], ((char *)src2_clone->data + idx * src2_clone->nb[2]), src2->ne[1] * src2->nb[1]); - } - } - - src2_clone->nb[0] = src2->nb[0]; - src2_clone->nb[1] = src2->nb[1]; - for (int i = 2; i < GGML_MAX_DIMS; i++) { - src2_clone->nb[i] = src2_clone->nb[i - 1]*src2_clone->ne[i - 1]; - } - } else { - if (offset + src2_size >= buffer_gpu->size) { - src2_size = buffer_gpu->size - offset; - } - ggml_vk_buffer_read(buffer_gpu, offset, src2_clone->data, src2_size); - memcpy(src2_clone->nb, src2->nb, sizeof(size_t) * GGML_MAX_DIMS); - } - } else { - GGML_ABORT("fatal error"); - } - - if (vk_output_tensor > 0 && vk_output_tensor == check_counter) { - ggml_vk_print_tensor(src2, "src2"); - } - } - if (src3 != nullptr) { - src3_clone = ggml_dup_tensor(ggml_ctx, src3); - - src3_size = ggml_nbytes(src3); - - src3_buffer = malloc(src3_size); - src3_clone->data = src3_buffer; - if (ggml_backend_buffer_is_host(src3->buffer)) { - memcpy(src3_clone->data, src3->data, src3_size); - memcpy(src3_clone->nb, src3->nb, sizeof(size_t) * GGML_MAX_DIMS); - } else if (ggml_backend_buffer_is_vk(src3->buffer)) { - ggml_backend_vk_buffer_context * buf_ctx = (ggml_backend_vk_buffer_context *)src3->buffer->context; - vk_buffer& buffer_gpu = buf_ctx->dev_buffer; - uint64_t offset = vk_tensor_offset(src3) + src3->view_offs; - if (!ggml_is_contiguous(src3) && ggml_vk_dim01_contiguous(src3)) { - for (int i3 = 0; i3 < src3->ne[3]; i3++) { - for (int i2 = 0; i2 < src3->ne[2]; i2++) { - const int idx = i3*src3->ne[2] + i2; - ggml_vk_buffer_read(buffer_gpu, offset + idx * src3->nb[2], ((char *)src3_clone->data + idx * src3_clone->nb[2]), src3->ne[1] * src3->nb[1]); - } - } - - src3_clone->nb[0] = src3->nb[0]; - src3_clone->nb[1] = src3->nb[1]; - for (int i = 2; i < GGML_MAX_DIMS; i++) { - src3_clone->nb[i] = src3_clone->nb[i - 1]*src3_clone->ne[i - 1]; - } - } else { - if (offset + src3_size >= buffer_gpu->size) { - src3_size = buffer_gpu->size - offset; - } - ggml_vk_buffer_read(buffer_gpu, offset, src3_clone->data, src3_size); - memcpy(src3_clone->nb, src3->nb, sizeof(size_t) * GGML_MAX_DIMS); - } - } else { - GGML_ABORT("fatal error"); - } - - if (vk_output_tensor > 0 && vk_output_tensor == check_counter) { - ggml_vk_print_tensor(src3, "src3"); + ggml_vk_print_tensor(srci, srci_name[i]); } } if (tensor->op == GGML_OP_FLASH_ATTN_EXT) { - const float *params = (const float *)tensor->op_params; - tensor_clone = ggml_flash_attn_ext(ggml_ctx, src0_clone, src1_clone, src2_clone, src3_clone, params[0], params[1], params[2]); + const float * params = (const float *)tensor->op_params; + tensor_clone = ggml_flash_attn_ext(ggml_ctx, src_clone[0], src_clone[1], src_clone[2], src_clone[3], params[0], params[1], params[2]); } else if (tensor->op == GGML_OP_MUL_MAT) { - tensor_clone = ggml_mul_mat(ggml_ctx, src0_clone, src1_clone); + tensor_clone = ggml_mul_mat(ggml_ctx, src_clone[0], src_clone[1]); } else if (tensor->op == GGML_OP_MUL_MAT_ID) { - tensor_clone = ggml_mul_mat_id(ggml_ctx, src0_clone, src1_clone, src2_clone); + tensor_clone = ggml_mul_mat_id(ggml_ctx, src_clone[0], src_clone[1], src_clone[2]); + } else if (tensor->op == GGML_OP_SUB) { + tensor_clone = ggml_sub(ggml_ctx, src_clone[0], src_clone[1]); } else if (tensor->op == GGML_OP_MUL) { - tensor_clone = ggml_mul(ggml_ctx, src0_clone, src1_clone); + tensor_clone = ggml_mul(ggml_ctx, src_clone[0], src_clone[1]); } else if (tensor->op == GGML_OP_DIV) { - tensor_clone = ggml_div(ggml_ctx, src0_clone, src1_clone); + tensor_clone = ggml_div(ggml_ctx, src_clone[0], src_clone[1]); } else if (tensor->op == GGML_OP_CONCAT) { - tensor_clone = ggml_concat(ggml_ctx, src0_clone, src1_clone, *(int *)tensor->op_params); + tensor_clone = ggml_concat(ggml_ctx, src_clone[0], src_clone[1], *(int *)tensor->op_params); } else if (tensor->op == GGML_OP_UPSCALE) { - tensor_clone = ggml_upscale_ext(ggml_ctx, src0_clone, tensor->ne[0], tensor->ne[1], tensor->ne[2], tensor->ne[3]); + tensor_clone = ggml_upscale_ext(ggml_ctx, src_clone[0], tensor->ne[0], tensor->ne[1], tensor->ne[2], tensor->ne[3], tensor->op_params[0], tensor->op_params[1], (ggml_scale_mode) tensor->op_params[0]); } else if (tensor->op == GGML_OP_SCALE) { - tensor_clone = ggml_scale(ggml_ctx, src0_clone, ((float *)tensor->op_params)[0]); + const float * params = (const float *)tensor->op_params; + tensor_clone = ggml_scale(ggml_ctx, src_clone[0], params[0]); } else if (tensor->op == GGML_OP_SQR) { - tensor_clone = ggml_sqr(ggml_ctx, src0_clone); + tensor_clone = ggml_sqr(ggml_ctx, src_clone[0]); } else if (tensor->op == GGML_OP_SIN) { - tensor_clone = ggml_sin(ggml_ctx, src0_clone); + tensor_clone = ggml_sin(ggml_ctx, src_clone[0]); } else if (tensor->op == GGML_OP_COS) { - tensor_clone = ggml_cos(ggml_ctx, src0_clone); + tensor_clone = ggml_cos(ggml_ctx, src_clone[0]); } else if (tensor->op == GGML_OP_CLAMP) { - tensor_clone = ggml_clamp(ggml_ctx, src0_clone, ((float *)tensor->op_params)[0], ((float *)tensor->op_params)[1]); + const float * params = (const float *)tensor->op_params; + tensor_clone = ggml_clamp(ggml_ctx, src_clone[0], params[0], params[1]); } else if (tensor->op == GGML_OP_PAD) { - tensor_clone = ggml_pad(ggml_ctx, src0_clone, tensor->ne[0] - src0_clone->ne[0], tensor->ne[1] - src0_clone->ne[1], tensor->ne[2] - src0_clone->ne[2], tensor->ne[3] - src0_clone->ne[3]); + tensor_clone = ggml_pad(ggml_ctx, src_clone[0], tensor->ne[0] - src_clone[0]->ne[0], tensor->ne[1] - src_clone[0]->ne[1], tensor->ne[2] - src_clone[0]->ne[2], tensor->ne[3] - src_clone[0]->ne[3]); } else if (tensor->op == GGML_OP_REPEAT) { - tensor_clone = ggml_repeat(ggml_ctx, src0_clone, tensor); + tensor_clone = ggml_repeat(ggml_ctx, src_clone[0], tensor); + } else if (tensor->op == GGML_OP_REPEAT_BACK) { + tensor_clone = ggml_repeat_back(ggml_ctx, src_clone[0], tensor); } else if (tensor->op == GGML_OP_ADD) { - tensor_clone = ggml_add(ggml_ctx, src0_clone, src1_clone); + tensor_clone = ggml_add(ggml_ctx, src_clone[0], src_clone[1]); } else if (tensor->op == GGML_OP_ACC) { - tensor_clone = ggml_acc(ggml_ctx, src0_clone, src1_clone, tensor->op_params[0], tensor->op_params[1], tensor->op_params[2], tensor->op_params[3]); + tensor_clone = ggml_acc(ggml_ctx, src_clone[0], src_clone[1], tensor->op_params[0], tensor->op_params[1], tensor->op_params[2], tensor->op_params[3]); } else if (tensor->op == GGML_OP_NORM) { - tensor_clone = ggml_norm(ggml_ctx, src0_clone, *(float *)tensor->op_params); + tensor_clone = ggml_norm(ggml_ctx, src_clone[0], *(float *)tensor->op_params); } else if (tensor->op == GGML_OP_GROUP_NORM) { - tensor_clone = ggml_group_norm(ggml_ctx, src0_clone, *(int *)tensor->op_params, ((float *)tensor->op_params)[1]); + const float * float_params = (const float *)tensor->op_params; + tensor_clone = ggml_group_norm(ggml_ctx, src_clone[0], tensor->op_params[0], float_params[1]); } else if (tensor->op == GGML_OP_RMS_NORM) { - tensor_clone = ggml_rms_norm(ggml_ctx, src0_clone, *(float *)tensor->op_params); + tensor_clone = ggml_rms_norm(ggml_ctx, src_clone[0], *(float *)tensor->op_params); + } else if (tensor->op == GGML_OP_RMS_NORM_BACK) { + const float eps = ((float *) tensor->op_params)[0]; + tensor_clone = ggml_rms_norm_back(ggml_ctx, src_clone[0], src_clone[1], eps); + } else if (tensor->op == GGML_OP_SILU_BACK) { + tensor_clone = ggml_silu_back(ggml_ctx, src_clone[0], src_clone[1]); + } else if (tensor->op == GGML_OP_L2_NORM) { + const float eps = ((float *) tensor->op_params)[0]; + tensor_clone = ggml_l2_norm(ggml_ctx, src_clone[0], eps); } else if (tensor->op == GGML_OP_SOFT_MAX) { if (src1 != nullptr) { - tensor_clone = ggml_soft_max_ext(ggml_ctx, src0_clone, src1_clone, ((float *)tensor->op_params)[0], ((float *)tensor->op_params)[1]); + const float * params = (const float *)tensor->op_params; + tensor_clone = ggml_soft_max_ext(ggml_ctx, src_clone[0], src_clone[1], params[0], params[1]); } else { - tensor_clone = ggml_soft_max(ggml_ctx, src0_clone); + tensor_clone = ggml_soft_max(ggml_ctx, src_clone[0]); } + } else if (tensor->op == GGML_OP_SOFT_MAX_BACK) { + tensor_clone = ggml_soft_max_ext_back(ggml_ctx, src_clone[0], src_clone[1], ((float *)tensor->op_params)[0], ((float *)tensor->op_params)[1]); } else if (tensor->op == GGML_OP_DIAG_MASK_INF) { - tensor_clone = ggml_diag_mask_inf(ggml_ctx, src0_clone, *(int *)tensor->op_params); - } else if (tensor->op == GGML_OP_ROPE) { + tensor_clone = ggml_diag_mask_inf(ggml_ctx, src_clone[0], tensor->op_params[0]); + } else if (tensor->op == GGML_OP_ROPE || tensor->op == GGML_OP_ROPE_BACK) { const int n_dims = ((int32_t *) tensor->op_params)[1]; const int mode = ((int32_t *) tensor->op_params)[2]; //const int n_ctx_ggml = ((int32_t *) tensor->op_params)[3]; @@ -8766,23 +9835,39 @@ static void ggml_vk_check_results_0(ggml_tensor * tensor) { const float attn_factor = ((float *) tensor->op_params)[8]; const float beta_fast = ((float *) tensor->op_params)[9]; const float beta_slow = ((float *) tensor->op_params)[10]; - tensor_clone = ggml_rope_ext(ggml_ctx, src0_clone, src1_clone, src2_clone, n_dims, mode, n_ctx_orig_ggml, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow); + if (mode & GGML_ROPE_TYPE_MROPE) { + int32_t *sections = ((int32_t *) tensor->op_params) + 11; + if (tensor->op == GGML_OP_ROPE) { + tensor_clone = ggml_rope_multi(ggml_ctx, src_clone[0], src_clone[1], src_clone[2], n_dims, sections, mode, n_ctx_orig_ggml, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow); + } else { + tensor_clone = ggml_rope_multi_back(ggml_ctx, src_clone[0], src_clone[1], src_clone[2], n_dims, sections, mode, n_ctx_orig_ggml, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow); + } + } else { + if (tensor->op == GGML_OP_ROPE) { + tensor_clone = ggml_rope_ext(ggml_ctx, src_clone[0], src_clone[1], src_clone[2], n_dims, mode, n_ctx_orig_ggml, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow); + } else { + tensor_clone = ggml_rope_ext_back(ggml_ctx, src_clone[0], src_clone[1], src_clone[2], n_dims, mode, n_ctx_orig_ggml, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow); + } + } } else if (tensor->op == GGML_OP_UNARY) { switch (ggml_get_unary_op(tensor)) { case GGML_UNARY_OP_SILU: - tensor_clone = ggml_silu(ggml_ctx, src0_clone); + tensor_clone = ggml_silu(ggml_ctx, src_clone[0]); break; case GGML_UNARY_OP_GELU: - tensor_clone = ggml_gelu(ggml_ctx, src0_clone); + tensor_clone = ggml_gelu(ggml_ctx, src_clone[0]); break; case GGML_UNARY_OP_GELU_QUICK: - tensor_clone = ggml_gelu_quick(ggml_ctx, src0_clone); + tensor_clone = ggml_gelu_quick(ggml_ctx, src_clone[0]); break; case GGML_UNARY_OP_RELU: - tensor_clone = ggml_relu(ggml_ctx, src0_clone); + tensor_clone = ggml_relu(ggml_ctx, src_clone[0]); break; case GGML_UNARY_OP_TANH: - tensor_clone = ggml_tanh(ggml_ctx, src0_clone); + tensor_clone = ggml_tanh(ggml_ctx, src_clone[0]); + break; + case GGML_UNARY_OP_SIGMOID: + tensor_clone = ggml_sigmoid(ggml_ctx, src_clone[0]); break; default: std::cerr << "Missing vk_check_results OP: " << ggml_op_name(tensor->op) << std::endl; @@ -8790,28 +9875,34 @@ static void ggml_vk_check_results_0(ggml_tensor * tensor) { } } else if (tensor->op == GGML_OP_CPY || tensor->op == GGML_OP_DUP) { if (src1 == nullptr) { - tensor_clone = ggml_dup(ggml_ctx, src0_clone); + tensor_clone = ggml_dup(ggml_ctx, src_clone[0]); tensor_clone->type = tensor->type; } else { - tensor_clone = ggml_cpy(ggml_ctx, src0_clone, src1_clone); + tensor_clone = ggml_cpy(ggml_ctx, src_clone[0], src_clone[1]); } } else if (tensor->op == GGML_OP_CONT) { - tensor_clone = ggml_cont_4d(ggml_ctx, src0_clone, tensor->ne[0], tensor->ne[1], tensor->ne[2], tensor->ne[3]); + tensor_clone = ggml_cont_4d(ggml_ctx, src_clone[0], tensor->ne[0], tensor->ne[1], tensor->ne[2], tensor->ne[3]); } else if (tensor->op == GGML_OP_RESHAPE) { - tensor_clone = ggml_reshape_4d(ggml_ctx, src0_clone, tensor->ne[0], tensor->ne[1], tensor->ne[2], tensor->ne[3]); + tensor_clone = ggml_reshape_4d(ggml_ctx, src_clone[0], tensor->ne[0], tensor->ne[1], tensor->ne[2], tensor->ne[3]); } else if (tensor->op == GGML_OP_VIEW) { - tensor_clone = ggml_view_4d(ggml_ctx, src0_clone, tensor->ne[0], tensor->ne[1], tensor->ne[2], tensor->ne[3], tensor->nb[1], tensor->nb[2], tensor->nb[3], ((int32_t *) tensor->op_params)[0]); + tensor_clone = ggml_view_4d(ggml_ctx, src_clone[0], tensor->ne[0], tensor->ne[1], tensor->ne[2], tensor->ne[3], tensor->nb[1], tensor->nb[2], tensor->nb[3], ((int32_t *) tensor->op_params)[0]); } else if (tensor->op == GGML_OP_PERMUTE) { int32_t * params = (int32_t *)tensor->op_params; - tensor_clone = ggml_permute(ggml_ctx, src0_clone, params[0], params[1], params[2], params[3]); + tensor_clone = ggml_permute(ggml_ctx, src_clone[0], params[0], params[1], params[2], params[3]); } else if (tensor->op == GGML_OP_TRANSPOSE) { - tensor_clone = ggml_transpose(ggml_ctx, src0_clone); + tensor_clone = ggml_transpose(ggml_ctx, src_clone[0]); } else if (tensor->op == GGML_OP_GET_ROWS) { - tensor_clone = ggml_get_rows(ggml_ctx, src0_clone, src1_clone); + tensor_clone = ggml_get_rows(ggml_ctx, src_clone[0], src_clone[1]); } else if (tensor->op == GGML_OP_ARGSORT) { - tensor_clone = ggml_argsort(ggml_ctx, src0_clone, (ggml_sort_order) *(int *)tensor->op_params); + tensor_clone = ggml_argsort(ggml_ctx, src_clone[0], (ggml_sort_order) *(int *)tensor->op_params); + } else if (tensor->op == GGML_OP_SUM) { + tensor_clone = ggml_sum(ggml_ctx, src_clone[0]); } else if (tensor->op == GGML_OP_SUM_ROWS) { - tensor_clone = ggml_sum_rows(ggml_ctx, src0_clone); + tensor_clone = ggml_sum_rows(ggml_ctx, src_clone[0]); + } else if (tensor->op == GGML_OP_ARGMAX) { + tensor_clone = ggml_argmax(ggml_ctx, src_clone[0]); + } else if (tensor->op == GGML_OP_COUNT_EQUAL) { + tensor_clone = ggml_count_equal(ggml_ctx, src_clone[0], src_clone[1]); } else if (tensor->op == GGML_OP_IM2COL) { const int32_t s0 = tensor->op_params[0]; const int32_t s1 = tensor->op_params[1]; @@ -8821,11 +9912,11 @@ static void ggml_vk_check_results_0(ggml_tensor * tensor) { const int32_t d1 = tensor->op_params[5]; const bool is_2D = tensor->op_params[6] == 1; - tensor_clone = ggml_im2col(ggml_ctx, src0_clone, src1_clone, s0, s1, p0, p1, d0, d1, is_2D, tensor->type); + tensor_clone = ggml_im2col(ggml_ctx, src_clone[0], src_clone[1], s0, s1, p0, p1, d0, d1, is_2D, tensor->type); } else if (tensor->op == GGML_OP_TIMESTEP_EMBEDDING) { const int32_t dim = tensor->op_params[0]; const int32_t max_period = tensor->op_params[1]; - tensor_clone = ggml_timestep_embedding(ggml_ctx, src0_clone, dim, max_period); + tensor_clone = ggml_timestep_embedding(ggml_ctx, src_clone[0], dim, max_period); } else if (tensor->op == GGML_OP_POOL_2D) { enum ggml_op_pool op = static_cast(tensor->op_params[0]); const int32_t k0 = tensor->op_params[1]; @@ -8835,13 +9926,20 @@ static void ggml_vk_check_results_0(ggml_tensor * tensor) { const int32_t p0 = tensor->op_params[5]; const int32_t p1 = tensor->op_params[6]; - tensor_clone = ggml_pool_2d(ggml_ctx, src0_clone, op, k0, k1, s0, s1, p0, p1); + tensor_clone = ggml_pool_2d(ggml_ctx, src_clone[0], op, k0, k1, s0, s1, p0, p1); } else if (tensor->op == GGML_OP_LEAKY_RELU) { const float * op_params = (const float *)tensor->op_params; - tensor_clone = ggml_leaky_relu(ggml_ctx, src0_clone, op_params[0], false); + tensor_clone = ggml_leaky_relu(ggml_ctx, src_clone[0], op_params[0], false); } else if (tensor->op == GGML_OP_RWKV_WKV6) { - tensor_clone = ggml_rwkv_wkv6(ggml_ctx, tensor->src[0], tensor->src[1], tensor->src[2], tensor->src[3], - tensor->src[4], tensor->src[5]); + tensor_clone = ggml_rwkv_wkv6(ggml_ctx, src_clone[0], src_clone[1], + src_clone[2], src_clone[3], src_clone[4], src_clone[5]); + } else if (tensor->op == GGML_OP_RWKV_WKV7) { + tensor_clone = ggml_rwkv_wkv7(ggml_ctx, src_clone[0], src_clone[1], src_clone[2], src_clone[3], + src_clone[4], src_clone[5], src_clone[6]); + } else if (tensor->op == GGML_OP_OPT_STEP_ADAMW) { + src_clone[0]->flags = src0->flags; + tensor_clone = ggml_opt_step_adamw(ggml_ctx, src_clone[0], src_clone[1], + src_clone[2], src_clone[3], src_clone[4]); } else { std::cerr << "Missing vk_check_results OP: " << ggml_op_name(tensor->op) << std::endl; @@ -8863,11 +9961,10 @@ static void ggml_vk_check_results_0(ggml_tensor * tensor) { memcpy(comp_result, tensor_clone->data, comp_size); memcpy(comp_nb, tensor_clone->nb, sizeof(size_t) * GGML_MAX_DIMS); - if (src0 != nullptr) { - free(src0_buffer); - } - if (src1 != nullptr) { - free(src1_buffer); + for (int i = 0; i < 6; i++) { + if (src_buffer[i] != nullptr) { + free(src_buffer[i]); + } } ggml_free(ggml_ctx); @@ -8931,6 +10028,9 @@ static void ggml_vk_check_results_1(ggml_tensor * tensor) { } else if (tensor->type == GGML_TYPE_I32) { correct = *(int32_t *) ((char *) comp_result + i3*comp_nb[3] + i2*comp_nb[2] + i1*comp_nb[1] + i0*comp_nb[0]); result = *(int32_t *) ((char *) tensor_data + i3*tensor->nb[3] + i2*tensor->nb[2] + i1*tensor->nb[1] + i0*tensor->nb[0]); + } else if (tensor->type == GGML_TYPE_I64) { + correct = *(int64_t *) ((char *) comp_result + i3*comp_nb[3] + i2*comp_nb[2] + i1*comp_nb[1] + i0*comp_nb[0]); + result = *(int64_t *) ((char *) tensor_data + i3*tensor->nb[3] + i2*tensor->nb[2] + i1*tensor->nb[1] + i0*tensor->nb[0]); } else { std::cerr << "Results check not implemented for type " << ggml_type_name(tensor->type) << std::endl; } diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/CMakeLists.txt b/ggml/src/ggml-vulkan/vulkan-shaders/CMakeLists.txt index 074031087..d6e0b2a5a 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/CMakeLists.txt +++ b/ggml/src/ggml-vulkan/vulkan-shaders/CMakeLists.txt @@ -1,9 +1,17 @@ -find_package (Threads REQUIRED) -find_program(GLSLC_EXECUTABLE glslc) -if(NOT GLSLC_EXECUTABLE) - message(FATAL_ERROR "glslc not found.") -endif() +cmake_minimum_required(VERSION 3.19) +project("vulkan-shaders-gen" C CXX) +find_package (Threads REQUIRED) + +if (GGML_VULKAN_COOPMAT_GLSLC_SUPPORT) + add_compile_definitions(GGML_VULKAN_COOPMAT_GLSLC_SUPPORT) +endif() +if (GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT) + add_compile_definitions(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT) +endif() +if (GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT) + add_compile_definitions(GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT) +endif() set(TARGET vulkan-shaders-gen) add_executable(${TARGET} vulkan-shaders-gen.cpp) install(TARGETS ${TARGET} RUNTIME) diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/argmax.comp b/ggml/src/ggml-vulkan/vulkan-shaders/argmax.comp new file mode 100644 index 000000000..eaf4da341 --- /dev/null +++ b/ggml/src/ggml-vulkan/vulkan-shaders/argmax.comp @@ -0,0 +1,51 @@ +#version 450 + +#include "generic_head.comp" +#include "types.comp" + +#extension GL_EXT_control_flow_attributes : enable + +layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in; + +layout (binding = 0) readonly buffer A {A_TYPE data_a[];}; +layout (binding = 1) writeonly buffer D {D_TYPE data_d[];}; + +layout (constant_id = 0) const uint BLOCK_SIZE = 32; + +shared FLOAT_TYPE tmpmax[BLOCK_SIZE]; +shared uint tmp[BLOCK_SIZE]; + +void main() { + const uint row = gl_WorkGroupID.z * 262144 + gl_WorkGroupID.y * 512 + gl_WorkGroupID.x; + const uint col = gl_LocalInvocationID.x; + + if (col >= p.KX) { + return; + } + A_TYPE amax = data_a[row*p.KX + col]; + tmp[col] = col; + + for (uint i = col + BLOCK_SIZE; i < p.KX; i += BLOCK_SIZE) { + A_TYPE val = data_a[row*p.KX + i]; + if (val > amax) { + amax = val; + tmp[col] = i; + } + } + tmpmax[col] = amax; + + barrier(); + [[unroll]] for (int s = int(BLOCK_SIZE) / 2; s > 0; s >>= 1) { + if (col < s && col + s < p.KX) { + if (tmpmax[col] < tmpmax[col + s]) { + tmpmax[col] = tmpmax[col + s]; + tmp[col] = tmp[col + s]; + } + } + barrier(); + } + + if (col == 0) { + data_d[row] = D_TYPE(tmp[0]); + } +} diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/copy_from_quant.comp b/ggml/src/ggml-vulkan/vulkan-shaders/copy_from_quant.comp index 9c9fe9626..dbc7daa33 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/copy_from_quant.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/copy_from_quant.comp @@ -12,7 +12,7 @@ layout(local_size_x = 1, local_size_y = 1, local_size_z = 1) in; #endif void main() { -#if defined(DATA_A_IQ2_XXS) || defined(DATA_A_IQ2_XS) || defined(DATA_A_IQ2_S) || defined(DATA_A_IQ3_XXS) || defined(DATA_A_IQ3_S) || defined(DATA_A_IQ4_XS) || defined(DATA_A_IQ4_NL) +#ifdef NEEDS_INIT_IQ_SHMEM init_iq_shmem(gl_WorkGroupSize); if (gl_LocalInvocationIndex.x != 0) { return; diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/copy_to_quant.comp b/ggml/src/ggml-vulkan/vulkan-shaders/copy_to_quant.comp index 660811086..9c76437d9 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/copy_to_quant.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/copy_to_quant.comp @@ -1,5 +1,10 @@ #version 450 +#if RTE16 +#extension GL_EXT_spirv_intrinsics : enable +spirv_execution_mode(capabilities = [4467], 4462, 16); // RoundingModeRTE, 16 bits +#endif // RTE16 + #include "types.comp" #include "generic_unary_head.comp" @@ -217,7 +222,7 @@ void quantize(uint dst_idx, uint src_idx) #endif void main() { -#if defined(DATA_A_IQ2_XXS) || defined(DATA_A_IQ2_XS) || defined(DATA_A_IQ2_S) || defined(DATA_A_IQ3_XXS) || defined(DATA_A_IQ3_S) || defined(DATA_A_IQ4_XS) || defined(DATA_A_IQ4_NL) +#ifdef NEEDS_INIT_IQ_SHMEM init_iq_shmem(gl_WorkGroupSize); if (gl_LocalInvocationIndex.x != 0) { return; diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/count_equal.comp b/ggml/src/ggml-vulkan/vulkan-shaders/count_equal.comp new file mode 100644 index 000000000..d9345497c --- /dev/null +++ b/ggml/src/ggml-vulkan/vulkan-shaders/count_equal.comp @@ -0,0 +1,31 @@ +#version 450 + +#extension GL_EXT_control_flow_attributes : enable + +#include "types.comp" +#include "generic_head.comp" + +layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in; + +layout (binding = 0) readonly buffer X {A_TYPE data_a[];}; +layout (binding = 1) readonly buffer Y {B_TYPE data_b[];}; +layout (binding = 2) buffer D {D_TYPE data_d[];}; + +const uint CHUNK_SIZE = 512; + +void main() { + const uint base = gl_WorkGroupID.x * CHUNK_SIZE; + const uint col = gl_LocalInvocationID.x; + + uint count = 0; + [[unroll]] + for (uint i = 0; i < CHUNK_SIZE; i += gl_WorkGroupSize.x) { + const uint idx = base + i + col; + if (idx >= p.KX) { + break; + } + count += uint(data_a[idx] == data_b[idx]); + } + + atomicAdd(data_d[0], D_TYPE(count)); +} diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs.comp b/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs.comp index ecfdbfaa8..2a162a2c8 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs.comp @@ -82,9 +82,86 @@ vec2 dequantize(uint ib, uint iqs, uint a_offset) { return vec2(int(data_a[a_offset + ib].qs[iqs]), int(data_a[a_offset + ib].qs[iqs + 1])); } vec4 dequantize4(uint ib, uint iqs, uint a_offset) { - uint32_t v0 = data_a_packed16[a_offset + ib].qs[iqs/2]; - uint32_t v1 = data_a_packed16[a_offset + ib].qs[iqs/2 + 1]; - return vec4(int8_t(v0 & 0xFF), int8_t(v0 >> 8), int8_t(v1 & 0xFF), int8_t(v1 >> 8)); + const i8vec2 v0 = unpack8(int32_t(data_a_packed16[a_offset + ib].qs[iqs/2])).xy; // vec4 used due to #12147 + const i8vec2 v1 = unpack8(int32_t(data_a_packed16[a_offset + ib].qs[iqs/2 + 1])).xy; + return vec4(v0.x, v0.y, v1.x, v1.y); +} +#endif + +#if defined(DATA_A_IQ1_S) +vec2 dequantize(uint ib, uint iqs, uint a_offset) { + const uint ib32 = iqs / 32; + const uint ib8 = iqs / 8; + const int i8 = int(iqs % 8); + const uint qh = data_a[a_offset + ib].qh[ib32]; + const uint qs = data_a[a_offset + ib].qs[ib8]; + const float dl = float(2 * bitfieldExtract(qh, 12, 3) + 1); + const float delta = ((qh & 0x8000) != 0) ? -IQ1S_DELTA : IQ1S_DELTA; + const uint idxhi = bitfieldExtract(qh, 3 * int(ib8 & 3), 3); + const int16_t grid = int16_t(iq1s_grid[qs | (idxhi << 8)]); + // Signed bitfield extract. + const ivec2 gvec = ivec2( + bitfieldExtract(grid, 2 * (i8), 2), + bitfieldExtract(grid, 2 * (i8 + 1), 2) + ); + return dl * (vec2(gvec) + delta); +} +vec4 dequantize4(uint ib, uint iqs, uint a_offset) { + const uint ib32 = iqs / 32; + const uint ib8 = iqs / 8; + const int i8 = int(iqs % 8); + const uint qh = data_a[a_offset + ib].qh[ib32]; + const uint qs = data_a[a_offset + ib].qs[ib8]; + const float dl = 2 * bitfieldExtract(qh, 12, 3) + 1; + const float delta = ((qh & 0x8000) != 0) ? -IQ1S_DELTA : IQ1S_DELTA; + const int16_t grid = int16_t(iq1s_grid[qs | (bitfieldExtract(qh, 3 * int(ib8 & 3), 3) << 8)]); + // Signed bitfield extract. + const ivec4 gvec = ivec4( + bitfieldExtract(grid, 2 * (i8), 2), + bitfieldExtract(grid, 2 * (i8 + 1), 2), + bitfieldExtract(grid, 2 * (i8 + 2), 2), + bitfieldExtract(grid, 2 * (i8 + 3), 2) + ); + return dl * (vec4(gvec) + delta); +} +#endif + +#if defined(DATA_A_IQ1_M) +vec2 dequantize(uint ib, uint iqs, uint a_offset) { + const uint ib8 = iqs / 8; + const uint ib16 = iqs / 16; + const int i8 = int(iqs % 8); + const uint sc = data_a[a_offset + ib].scales[iqs / 64]; + const uint qs = data_a[a_offset + ib].qs[ib8]; + const uint qh = data_a[a_offset + ib].qh[ib16] >> (4 * (ib8 & 1)); + const float dl = 2 * bitfieldExtract(sc, 3 * int(ib16 & 3), 3) + 1; + const float delta = ((qh & 8) != 0) ? -IQ1M_DELTA : IQ1M_DELTA; + const int16_t grid = int16_t(iq1s_grid[qs | ((qh & 7) << 8)]); + // Signed bitfield extract. + const ivec2 gvec = ivec2( + bitfieldExtract(grid, 2 * (i8), 2), + bitfieldExtract(grid, 2 * (i8 + 1), 2) + ); + return dl * (vec2(gvec) + delta); +} +vec4 dequantize4(uint ib, uint iqs, uint a_offset) { + const uint ib8 = iqs / 8; + const uint ib16 = iqs / 16; + const int i8 = int(iqs % 8); + const uint sc = data_a[a_offset + ib].scales[iqs / 64]; + const uint qs = data_a[a_offset + ib].qs[ib8]; + const uint qh = data_a[a_offset + ib].qh[ib16] >> (4 * (ib8 & 1)); + const float dl = 2 * bitfieldExtract(sc, 3 * int(ib16 & 3), 3) + 1; + const float delta = ((qh & 8) != 0) ? -IQ1M_DELTA : IQ1M_DELTA; + const int16_t grid = int16_t(iq1s_grid[qs | ((qh & 7) << 8)]); + // Signed bitfield extract. + const ivec4 gvec = ivec4( + bitfieldExtract(grid, 2 * (i8), 2), + bitfieldExtract(grid, 2 * (i8 + 1), 2), + bitfieldExtract(grid, 2 * (i8 + 2), 2), + bitfieldExtract(grid, 2 * (i8 + 3), 2) + ); + return dl * (vec4(gvec) + delta); } #endif @@ -357,7 +434,16 @@ vec2 get_dm(uint ib, uint a_offset) { } #endif -#if defined(DATA_A_Q4_0) || defined(DATA_A_Q5_0) || defined(DATA_A_Q8_0) || defined(DATA_A_IQ2_XXS) || defined(DATA_A_IQ2_XS) || defined(DATA_A_IQ2_S) || defined(DATA_A_IQ3_XXS) || defined(DATA_A_IQ3_S) || defined(DATA_A_IQ4_XS) || defined(DATA_A_IQ4_NL) +#if defined(DATA_A_IQ1_M) +vec2 get_dm(uint ib, uint a_offset) { + const uint16_t[4] scales = data_a[a_offset + ib].scales; + const u16vec4 s = u16vec4(scales[0], scales[1], scales[2], scales[3]) >> 12; + const float d = float(unpackHalf2x16(s.x | (s.y << 4) | (s.z << 8) | (s.w << 12)).x); + return vec2(d, 0); +} +#endif + +#if defined(DATA_A_Q4_0) || defined(DATA_A_Q5_0) || defined(DATA_A_Q8_0) || defined(DATA_A_IQ1_S) || defined(DATA_A_IQ2_XXS) || defined(DATA_A_IQ2_XS) || defined(DATA_A_IQ2_S) || defined(DATA_A_IQ3_XXS) || defined(DATA_A_IQ3_S) || defined(DATA_A_IQ4_XS) || defined(DATA_A_IQ4_NL) vec2 get_dm(uint ib, uint a_offset) { return vec2(float(data_a[a_offset + ib].d), 0); } diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs_cm2.comp b/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs_cm2.comp index 0eba37420..962d2353f 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs_cm2.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs_cm2.comp @@ -92,7 +92,7 @@ float16_t dequantFuncQ8_0(const in decodeBufQ8_0 bl, const in uint blockCoords[2 const uint iqs = idx; // Load 16b and select the byte for this element - int32_t qs = unpack8(int32_t(bl.block.qs[(iqs & 0x1E) >> 1]))[iqs & 1]; + int32_t qs = unpack8(bl.block.qs[(iqs & 0x1E) >> 1])[iqs & 1]; float16_t ret = float16_t(qs) * d; return ret; } @@ -167,6 +167,101 @@ layout(buffer_reference, std430, buffer_reference_align = 16) buffer decodeBufQ4 block_q4_K_packed128 block; }; +#if defined(IS_MUL_MM2) + +// For Q4_K and Q5_K in the mat-mul shader, we decode a tile's worth of scales +// into shared memory and then process the whole tile using those scales. +// There is a fetch function that loads into private variables and then a store +// function that stores into shared memory. +// Q4_K and Q5_K have the same encoding of scales, so everything is shared except +// the part that fetches from the structure (which has a different block layout). +#if defined(DATA_A_Q4_K) || defined(DATA_A_Q5_K) +const uint shAscales_stride = (BM + 2); +// 1 scale per 32 elements -> 8 scales per block, per row +shared vec2 shAscales[8 * shAscales_stride]; +uvec4 row_v; +#endif + +#if defined(DATA_A_Q4_K) +layout (binding = 0) readonly buffer A_Q4_K_128 {block_q4_K_packed128 data_a_q4_k_packed128[];}; + +void fetch_scalesQ4_K(uint ir_BM, uint pos_a, uint stride_a, uint block_k, uint tid, bool in_bounds) +{ + uint tids_per_row = BLOCK_SIZE / BM; + uint is_per_tid = 8 / tids_per_row; + uint is_start = is_per_tid * (tid % tids_per_row); + uint tid_row = tid / tids_per_row; + + uint row = ir_BM + tid_row; + uint block_index = pos_a + row * stride_a + (block_k / QUANT_K); + if (in_bounds || row < p.M) { + row_v = data_a_q4_k_packed128[block_index].q4k[0]; + } +} +#endif +#if defined(DATA_A_Q5_K) +layout (binding = 0) readonly buffer A_Q5_K_128 {block_q5_K_packed128 data_a_q5_k_packed128[];}; + +void fetch_scalesQ5_K(uint ir_BM, uint pos_a, uint stride_a, uint block_k, uint tid, bool in_bounds) +{ + uint tids_per_row = BLOCK_SIZE / BM; + uint is_per_tid = 8 / tids_per_row; + uint is_start = is_per_tid * (tid % tids_per_row); + uint tid_row = tid / tids_per_row; + + uint row = ir_BM + tid_row; + uint block_index = pos_a + row * stride_a + (block_k / QUANT_K); + if (in_bounds || row < p.M) { + row_v = data_a_q5_k_packed128[block_index].q5k[0]; + } +} +#endif + +#if defined(DATA_A_Q4_K) || defined(DATA_A_Q5_K) +void store_scalesQ4_K(uint tid) +{ + barrier(); + + uint tids_per_row = BLOCK_SIZE / BM; + uint is_per_tid = 8 / tids_per_row; + uint is_start = is_per_tid * (tid % tids_per_row); + uint tid_row = tid / tids_per_row; + + [[unroll]] for (uint idx = 0; idx < is_per_tid; ++idx) { + uint is = idx + is_start; + uvec4 v = row_v; + const vec2 loadd = vec2(unpackFloat2x16(v.x)); + + uint32_t sc; + uint32_t mbyte; + + uint32_t scale0 = v.y; + uint32_t scale4 = v.z; + uint32_t scale8 = v.w; + + uint32_t sc_lo = scale0; + uint32_t mb_lo = scale4; + uint32_t sc_hi = (scale8 & 0x0F0F0F0F) | ((scale0 & 0xC0C0C0C0) >> 2); + uint32_t mb_hi = ((scale8 & 0xF0F0F0F0) >> 4) | ((scale4 & 0xC0C0C0C0) >> 2); + + sc = is < 4 ? sc_lo : sc_hi; + mbyte = is < 4 ? mb_lo : mb_hi; + sc = sc >> (8 * (is & 3)); + mbyte = mbyte >> (8 * (is & 3)); + sc &= 0x3F; + mbyte &= 0x3F; + + const float d = loadd.x * float(sc); + const float m = loadd.y * float(mbyte); + shAscales[is * shAscales_stride + tid_row] = vec2(d,m); + } + + barrier(); +} +#endif + +#endif + float16_t dequantFuncQ4_K(const in decodeBufQ4_K bl, const in uint blockCoords[2], const in uint coordInBlock[2]) { decodeBufQ4_K_packed16 bl16 = decodeBufQ4_K_packed16(bl); @@ -176,9 +271,13 @@ float16_t dequantFuncQ4_K(const in decodeBufQ4_K bl, const in uint blockCoords[2 const uint b = (idx & 0x20) >> 5; // 0,1 const uint is = (idx & 0xE0) >> 5; // 0..7 +#if defined(IS_MUL_MM2) && defined(DATA_A_Q4_K) + vec2 v = shAscales[is * shAscales_stride + (blockCoords[0] % BM)]; + float d = v.x; + float m = v.y; +#else uvec4 v = bl128.block.q4k[0]; - - const f16vec2 loadd = unpackFloat2x16(v.x); + const vec2 loadd = vec2(unpackFloat2x16(v.x)); uint32_t sc; uint32_t mbyte; @@ -199,15 +298,16 @@ float16_t dequantFuncQ4_K(const in decodeBufQ4_K bl, const in uint blockCoords[2 sc &= 0x3F; mbyte &= 0x3F; - const float16_t d = loadd.x * float16_t(sc); - const float16_t m = loadd.y * float16_t(mbyte); + const float d = loadd.x * float(sc); + const float m = loadd.y * float(mbyte); +#endif uint qs = uint32_t(bl16.block.qs[((idx & 0xC0) >> 2) + ((idx & 0x1E) >> 1)]); qs = (qs >> (b * 4 + 8 * (idx & 1))) & 0xF; - float16_t ret = d * float16_t(qs) - m; + float ret = d * float(qs) - m; - return ret; + return float16_t(ret); } layout(buffer_reference, std430, buffer_reference_align = 16) buffer decodeBufQ5_K { @@ -231,6 +331,11 @@ float16_t dequantFuncQ5_K(const in decodeBufQ5_K bl, const in uint blockCoords[2 const uint b = (idx & 0x20) >> 5; // 0,1 const uint is = (idx & 0xE0) >> 5; // 0..7 +#if defined(IS_MUL_MM2) && defined(DATA_A_Q5_K) + vec2 v = shAscales[is * shAscales_stride + (blockCoords[0] % BM)]; + float d = v.x; + float m = v.y; +#else uvec4 v = bl128.block.q5k[0]; const f16vec2 loadd = unpackFloat2x16(v.x); @@ -256,6 +361,7 @@ float16_t dequantFuncQ5_K(const in decodeBufQ5_K bl, const in uint blockCoords[2 const float16_t d = loadd.x * float16_t(sc); const float16_t m = loadd.y * float16_t(mbyte); +#endif uint qh = uint32_t(bl16.block.qh[(idx & 0x1E) >> 1]); qh = ((qh >> is) & 0x101) << 4; @@ -264,9 +370,9 @@ float16_t dequantFuncQ5_K(const in decodeBufQ5_K bl, const in uint blockCoords[2 qs = (qs >> (b * 4)) & 0x0F0F; qs = unpack8(qs | qh)[idx & 1]; - float16_t ret = d * (float16_t(qs)) - m; + float ret = d * float(qs) - m; - return ret; + return float16_t(ret); } layout(buffer_reference, std430, buffer_reference_align = 2) buffer decodeBufQ6_K { @@ -301,6 +407,62 @@ float16_t dequantFuncQ6_K(const in decodeBufQ6_K bl, const in uint blockCoords[2 return ret; } +#if defined(DATA_A_IQ1_S) +layout(buffer_reference, std430, buffer_reference_align = 2) buffer decodeBufIQ1_S { + block_iq1_s block; +}; + +float16_t dequantFuncIQ1_S(const in decodeBufIQ1_S bl, const in uint blockCoords[2], const in uint coordInBlock[2]) +{ + const float16_t d = bl.block.d; + const uint idx = coordInBlock[1]; + + const uint ib32 = (idx & 0xE0) >> 5; + const uint ib8 = (idx & 0xF8) >> 3; + + const uint qh = bl.block.qh[ib32]; + const uint qs = bl.block.qs[ib8]; + const float dl = d * float(2 * bitfieldExtract(qh, 12, 3) + 1); + const float delta = ((qh & 0x8000) != 0) ? -IQ1S_DELTA : IQ1S_DELTA; + const uint grid = iq1s_grid[qs | (bitfieldExtract(qh, 3 * int(ib8 & 3), 3) << 8)]; + + float16_t ret = float16_t(dl) * (float16_t(bitfieldExtract(int(grid), 2 * int(idx % 8), 2)) + float16_t(delta)); + return ret; +} +#endif + +#if defined(DATA_A_IQ1_M) +layout(buffer_reference, std430, buffer_reference_align = 2) buffer decodeBufIQ1_M { + block_iq1_m block; +}; + +layout(buffer_reference, std430, buffer_reference_align = 8) buffer decodeBufIQ1_M_packed64 { + block_iq1_m_packed64 block; +}; + +float16_t dequantFuncIQ1_M(const in decodeBufIQ1_M bl, const in uint blockCoords[2], const in uint coordInBlock[2]) +{ + decodeBufIQ1_M_packed64 bl64 = decodeBufIQ1_M_packed64(bl); + const uint idx = coordInBlock[1]; + + uvec2 scales = unpack32(bl64.block.scales); + const float16_t d = uint16BitsToHalf(uint16_t(((scales.x & 0xF000) >> 12) | ((scales.x & 0xF0000000) >> 24) | ((scales.y & 0xF000) >> 4) | ((scales.y & 0xF0000000) >> 16))); + + const uint ib8 = (idx & 0xF8) >> 3; + const uint ib16 = (idx & 0xF0) >> 4; + const int i8 = int(idx % 8); + const uint sc = bl.block.scales[ib8 / 8]; + const uint qs = bl.block.qs[ib8]; + const uint qh = bl.block.qh[ib16] >> (4 * (ib8 & 1)); + const float dl = 2 * bitfieldExtract(sc, 3 * int(ib16 & 3), 3) + 1; + const float delta = ((qh & 8) != 0) ? -IQ1S_DELTA : IQ1S_DELTA; + const uint grid = iq1s_grid[qs | ((qh & 7) << 8)]; + + float16_t ret = d * float16_t(dl) * (float16_t(bitfieldExtract(int(grid), 2 * i8, 2)) + float16_t(delta)); + return ret; +} +#endif + #if defined(DATA_A_IQ2_XXS) layout(buffer_reference, std430, buffer_reference_align = 2) buffer decodeBufIQ2_XXS { block_iq2_xxs block; @@ -508,10 +670,18 @@ float16_t dequantFuncIQ4_NL(const in decodeBufIQ4_NL bl, const in uint blockCoor #define dequantFuncA dequantFuncQ3_K #elif defined(DATA_A_Q4_K) #define dequantFuncA dequantFuncQ4_K +#define fetch_scales fetch_scalesQ4_K +#define store_scales store_scalesQ4_K #elif defined(DATA_A_Q5_K) #define dequantFuncA dequantFuncQ5_K +#define fetch_scales fetch_scalesQ5_K +#define store_scales store_scalesQ4_K #elif defined(DATA_A_Q6_K) #define dequantFuncA dequantFuncQ6_K +#elif defined(DATA_A_IQ1_S) +#define dequantFuncA dequantFuncIQ1_S +#elif defined(DATA_A_IQ1_M) +#define dequantFuncA dequantFuncIQ1_M #elif defined(DATA_A_IQ2_XXS) #define dequantFuncA dequantFuncIQ2_XXS #elif defined(DATA_A_IQ2_XS) diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq1_m.comp b/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq1_m.comp new file mode 100644 index 000000000..39184ef58 --- /dev/null +++ b/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq1_m.comp @@ -0,0 +1,42 @@ +#version 450 + +#extension GL_EXT_shader_explicit_arithmetic_types_float16 : require + +#include "dequant_head.comp" + +layout(local_size_x = 256, local_size_y = 1, local_size_z = 1) in; + +layout (binding = 0) readonly buffer A {block_iq1_m data_a[];}; +layout (binding = 1) writeonly buffer D {D_TYPE data_b[];}; + +void main() { + // Each thread handles 1 subblock (32 values with 2 scales) + const uint ib = gl_WorkGroupID.x * 32 + gl_LocalInvocationID.x / 8; + + init_iq_shmem(gl_WorkGroupSize); + + if (ib >= p.nel / 256) { + return; + } + + const uint ib32 = gl_LocalInvocationID.x % 8; + const uint ib64 = ib32 / 2; + const uint b_idx = 256 * ib + 32 * ib32; + + const uint16_t[4] scales = data_a[ib].scales; + const u16vec4 s = u16vec4(scales[0], scales[1], scales[2], scales[3]) >> 12; + const float d = float(unpackHalf2x16(s.x | (s.y << 4) | (s.z << 8) | (s.w << 12)).x); + + const uint sc = data_a[ib].scales[ib64]; + [[unroll]] for (int l = 0; l < 4; ++l) { + const uint ib16 = 2 * ib32 + l / 2; + const float dl = d * (2 * bitfieldExtract(sc, 3 * int(ib16 & 3), 3) + 1); + const uint qh = data_a[ib].qh[ib16] >> (4 * (l & 1)); + const uint qs = data_a[ib].qs[4 * ib32 + l]; + const float delta = ((qh & 8) != 0) ? -IQ1M_DELTA : IQ1M_DELTA; + const int16_t grid = int16_t(iq1s_grid[qs | ((qh & 7) << 8)]); + [[unroll]] for (int j = 0; j < 8; ++j) { + data_b[b_idx + 8 * l + j] = D_TYPE(dl * (bitfieldExtract(grid, 2*j, 2) + delta)); + } + } +} diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq1_s.comp b/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq1_s.comp new file mode 100644 index 000000000..fd1e4e30d --- /dev/null +++ b/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq1_s.comp @@ -0,0 +1,35 @@ +#version 450 + +#include "dequant_head.comp" + +layout(local_size_x = 256, local_size_y = 1, local_size_z = 1) in; + +layout (binding = 0) readonly buffer A {block_iq1_s data_a[];}; +layout (binding = 1) writeonly buffer D {D_TYPE data_b[];}; + +void main() { + // Each thread handles 1 subblock (32 values with 2 scales) + const uint ib = gl_WorkGroupID.x * 32 + gl_LocalInvocationID.x / 8; + + init_iq_shmem(gl_WorkGroupSize); + + if (ib >= p.nel / 256) { + return; + } + + const uint ib32 = gl_LocalInvocationID.x % 8; + const uint b_idx = 256 * ib + 32 * ib32; + + uint qh = data_a[ib].qh[ib32]; + const float d = float(data_a[ib].d); + const float dl = d * float(2 * bitfieldExtract(qh, 12, 3) + 1); + const float delta = ((qh & 0x8000) != 0) ? -IQ1S_DELTA : IQ1S_DELTA; + [[unroll]] for (uint l = 0; l < 4; ++l) { + const uint qs = data_a[ib].qs[4 * ib32 + l]; + const uint hi = bitfieldExtract(qh, 3 * int(l), 3); + const int16_t grid = int16_t(iq1s_grid[qs | (hi << 8)]); + [[unroll]] for (int j = 0; j < 8; ++j) { + data_b[b_idx + 8 * l + j] = D_TYPE(dl * (bitfieldExtract(grid, 2*j, 2) + delta)); + } + } +} diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm2.comp b/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm2.comp index ba88ce79a..e1baa85f9 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm2.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm2.comp @@ -61,6 +61,10 @@ layout (push_constant) uniform parameter { uint32_t n_head_log2; float m0; float m1; + + uint32_t gqa_ratio; + uint32_t split_kv; + uint32_t k_num; } p; layout (binding = 0) readonly buffer Q {uint8_t data_q[];}; @@ -103,20 +107,62 @@ ACC_TYPE Max(const in uint32_t row, const in uint32_t col, const in ACC_TYPE ele #define DECODEFUNC #endif +// Store the output when doing grouped query attention. +// Rows index by Q's dimension 2, and the first N rows are valid. +D_TYPE perElemOpGqaStore(const in uint32_t r, const in uint32_t c, const in D_TYPE elem, const in uint32_t o_offset, const in uint32_t iq2, const in uint32_t N) +{ + if (r < N && c < D) { + uint32_t offset = (iq2 + r) * D + c; + data_o[o_offset + offset] = D_TYPE(elem); + } + return elem; +} + +// Store column zero. This is used to save per-row m and L values for split_k. +ACC_TYPE perElemOpStoreCol0(const in uint32_t r, const in uint32_t c, const in ACC_TYPE elem, const in uint32_t o_offset, const in uint32_t iq2, const in uint32_t N) +{ + if (r < N && c == 0) { + uint32_t offset = iq2 + r; + data_o[o_offset + offset] = D_TYPE(elem); + } + return elem; +} + +// Load the slope matrix, indexed by Q's dimension 2. +ACC_TYPE perElemOpComputeSlope(const in uint32_t r, const in uint32_t c, const in ACC_TYPE elem, const in uint32_t iq2) +{ + const uint32_t h = iq2 + (r & (p.gqa_ratio - 1)); + + const ACC_TYPE base = ACC_TYPE(h < p.n_head_log2 ? p.m0 : p.m1); + const int exph = int(h < p.n_head_log2 ? h + 1 : 2*(h - p.n_head_log2) + 1); + + return ACC_TYPE(pow(base, ACC_TYPE(exph))); +} + void main() { -#if defined(DATA_A_IQ2_XXS) || defined(DATA_A_IQ2_XS) || defined(DATA_A_IQ2_S) || defined(DATA_A_IQ3_XXS) || defined(DATA_A_IQ3_S) || defined(DATA_A_IQ4_XS) || defined(DATA_A_IQ4_NL) +#ifdef NEEDS_INIT_IQ_SHMEM init_iq_shmem(gl_WorkGroupSize); #endif const uint32_t N = p.N; const uint32_t KV = p.KV; + uint32_t i = gl_WorkGroupID.x; + uint32_t split_k_index = 0; + + if (p.k_num > 1) { + i = 0; + split_k_index = gl_WorkGroupID.x; + } + const uint32_t Tr = CEIL_DIV(N, Br); - const uint32_t Tc = CEIL_DIV(KV, Bc); - const uint32_t i = gl_WorkGroupID.x; + const uint32_t start_j = split_k_index * p.split_kv / Bc; + const uint32_t end_j = CEIL_DIV(min(KV, (split_k_index + 1) * p.split_kv), Bc); - const uint32_t iq2 = gl_WorkGroupID.y; + // When not using grouped query attention, all rows share the same iq2, equal to gl_WorkGroupID.y. + // When using grouped query attention, each workgroup does gqa_ratio consecutive values of iq2. + const uint32_t iq2 = gl_WorkGroupID.y * p.gqa_ratio; const uint32_t iq3 = gl_WorkGroupID.z; // broadcast factors @@ -149,10 +195,17 @@ void main() { tensorLayoutK = setTensorLayoutDimensionNV(tensorLayoutK, KV, D); tensorLayoutV = setTensorLayoutDimensionNV(tensorLayoutV, KV, D); - // nb?1 are already divided by the type size and are in units of elements - uint32_t q_stride = p.nb01; + // nb?1 are already divided by the type size and are in units of elements. + // When using grouped query attention, Q is indexed by iq2, so the stride + // should be nb02 (which is in bytes). + uint32_t q_stride = p.gqa_ratio > 1 ? (p.nb02 / 4) : p.nb01; uint32_t k_stride = p.nb11; uint32_t v_stride = p.nb21; + // When using grouped query attention, all rows use the same mask (stride 0). + // "p.gqa_ratio >> 16" is just a roundabout way of writing zero + // that prevents the compiler from folding the "&" through the select + // and breaking the alignment detection. + uint32_t m_stride = (p.gqa_ratio > 1) ? (p.gqa_ratio >> 16) : KV; // hint to the compiler that strides are aligned for the aligned variant of the shader if (Clamp != gl_CooperativeMatrixClampModeConstantNV) { @@ -161,6 +214,7 @@ void main() { k_stride &= ~7; v_stride &= ~7; #endif + m_stride &= ~7; } tensorLayoutQ = setTensorLayoutStrideNV(tensorLayoutQ, q_stride, 1); tensorLayoutK = setTensorLayoutStrideNV(tensorLayoutK, k_stride, 1); @@ -179,23 +233,21 @@ void main() { coopmat L, M; - L = coopmat(0); - M = coopmat(-1.0/0.0); + // Use -FLT_MAX/2 rather than -inf to reduce the possibility of NaNs, e.g. when computing Mold-M. + const float NEG_FLT_MAX_OVER_2 = uintBitsToFloat(0xFEFFFFFF); - ACC_TYPE slope = ACC_TYPE(1.0); + L = coopmat(0); + M = coopmat(NEG_FLT_MAX_OVER_2); + + coopmat slopeMat = coopmat(1.0); // ALiBi if (p.max_bias > 0.0f) { - const uint32_t h = iq2; - - const ACC_TYPE base = ACC_TYPE(h < p.n_head_log2 ? p.m0 : p.m1); - const int exph = int(h < p.n_head_log2 ? h + 1 : 2*(h - p.n_head_log2) + 1); - - slope = pow(base, ACC_TYPE(exph)); + coopMatPerElementNV(slopeMat, slopeMat, perElemOpComputeSlope, iq2); } [[dont_unroll]] - for (uint32_t j = 0; j < Tc; ++j) { + for (uint32_t j = start_j; j < end_j; ++j) { coopmat S = coopmat(0); @@ -213,14 +265,15 @@ void main() { } if (p.mask != 0) { - tensorLayoutNV<2, gl_CooperativeMatrixClampModeConstantNV> tensorLayoutM = createTensorLayoutNV(2, gl_CooperativeMatrixClampModeConstantNV); + tensorLayoutNV<2, Clamp> tensorLayoutM = createTensorLayoutNV(2, Clamp); tensorLayoutM = setTensorLayoutDimensionNV(tensorLayoutM, p.nem1, KV); + tensorLayoutM = setTensorLayoutStrideNV(tensorLayoutM, m_stride, 1); coopmat mv; coopMatLoadTensorNV(mv, data_m, 0, sliceTensorLayoutNV(tensorLayoutM, i * Br, Br, j * Bc, Bc)); - S += slope*coopmat(mv); + S += slopeMat*coopmat(mv); } // Clear padding elements to -inf, so they don't contribute to rowmax @@ -231,7 +284,7 @@ void main() { uint R = ((i + 1) * Br > N) ? (N % Br) : Br; uint C = ((j + 1) * Bc > KV) ? (KV % Bc) : Bc; - coopMatPerElementNV(S, S, replacePadding, ACC_TYPE(-1.0/0.0), R, C); + coopMatPerElementNV(S, S, replacePadding, ACC_TYPE(NEG_FLT_MAX_OVER_2), R, C); } coopmat rowmax, P, rowsum, eM; @@ -280,9 +333,25 @@ void main() { // resize eM by using smear/reduce coopMatReduceNV(eMdiag, eM, gl_CooperativeMatrixReduceRowNV, smearReduce); - O = eMdiag * O; + // multiply with fp16 accumulation, then add to O. + coopmat PV = coopmat(0); + PV = coopMatMulAdd(P_A, V, PV); - O = coopMatMulAdd(P_A, V, O); + O = eMdiag * O + coopmat(PV); + } + + // If there is split_k, then the split_k resolve shader does the final + // division by L. Store the intermediate O value and per-row m and L values. + if (p.k_num > 1) { + coopmat O_D = coopmat(O); + + uint32_t o_offset = D * p.ne1 * split_k_index; + coopMatPerElementNV(O_D, O_D, perElemOpGqaStore, o_offset, iq2, N); + + o_offset = D * p.ne1 * p.k_num + p.ne1 * split_k_index * 2; + coopMatPerElementNV(L, L, perElemOpStoreCol0, o_offset, iq2, N); + coopMatPerElementNV(M, M, perElemOpStoreCol0, o_offset + p.ne1, iq2, N); + return; } coopmat Ldiag; @@ -297,13 +366,18 @@ void main() { O = Ldiag*O; - tensorLayoutNV<3, gl_CooperativeMatrixClampModeConstantNV> tensorLayoutD = createTensorLayoutNV(3, gl_CooperativeMatrixClampModeConstantNV); - tensorLayoutD = setTensorLayoutDimensionNV(tensorLayoutD, p.ne2, p.ne1, D); - - // permute dimensions - tensorViewNV<3, false, 1, 0, 2> tensorViewPermute = createTensorViewNV(3, false, 1, 0, 2); uint32_t o_offset = iq3*p.ne2*p.ne1; coopmat O_D = coopmat(O); - coopMatStoreTensorNV(O_D, data_o, o_offset, sliceTensorLayoutNV(tensorLayoutD, i * Br, Br, iq2, 1, 0, D), tensorViewPermute); + if (p.gqa_ratio > 1) { + coopMatPerElementNV(O_D, O_D, perElemOpGqaStore, o_offset, iq2, N); + } else { + tensorLayoutNV<3, gl_CooperativeMatrixClampModeConstantNV> tensorLayoutD = createTensorLayoutNV(3, gl_CooperativeMatrixClampModeConstantNV); + tensorLayoutD = setTensorLayoutDimensionNV(tensorLayoutD, p.ne2, p.ne1, D); + + // permute dimensions + tensorViewNV<3, false, 1, 0, 2> tensorViewPermute = createTensorViewNV(3, false, 1, 0, 2); + + coopMatStoreTensorNV(O_D, data_o, o_offset, sliceTensorLayoutNV(tensorLayoutD, i * Br, Br, iq2, N, 0, D), tensorViewPermute); + } } diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_split_k_reduce.comp b/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_split_k_reduce.comp new file mode 100644 index 000000000..a7e395685 --- /dev/null +++ b/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_split_k_reduce.comp @@ -0,0 +1,59 @@ +#version 450 + +#extension GL_EXT_control_flow_attributes : enable + +#define BLOCK_SIZE 32 + +layout(local_size_x = BLOCK_SIZE, local_size_y = 1, local_size_z = 1) in; + +layout (binding = 0) readonly buffer A {float data_a[];}; +layout (binding = 1) writeonly buffer D {float data_d[];}; + +layout (push_constant) uniform parameter { + uint D; + uint N; + uint k_num; +} p; + +void main() { + // Each workgroup handles a row + const uint n = gl_WorkGroupID.x; + const uint tid = gl_LocalInvocationID.x; + + uint D = p.D; + uint N = p.N; + uint k_num = p.k_num; + + uint l_offset = D * N * k_num + n; + uint m_offset = D * N * k_num + N + n; + uint lm_stride = N * 2; + + // Compute the max m value for the row + float m_max = -1.0/0.0; + [[unroll]] for (uint k = 0; k < k_num; ++k) { + float m = data_a[m_offset + k * lm_stride]; + m_max = max(m_max, m); + } + + // Compute L based on m_max + float L = 0; + [[unroll]] for (uint k = 0; k < k_num; ++k) { + float l = data_a[l_offset + k * lm_stride]; + float m = data_a[m_offset + k * lm_stride]; + L += exp(m - m_max) * l; + } + + L = 1.0 / L; + + // Scale and sum the O contributions based on m_max and store the result to memory + for (uint d = tid; d < D; d += BLOCK_SIZE) { + float O = 0.0; + [[unroll]] for (uint k = 0; k < k_num; ++k) { + uint o_offset = D * N * k + D * n + d; + float m = data_a[m_offset + k * lm_stride]; + O += exp(m - m_max) * data_a[o_offset]; + } + O *= L; + data_d[D * n + d] = O; + } +} diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/get_rows_quant.comp b/ggml/src/ggml-vulkan/vulkan-shaders/get_rows_quant.comp index c16a2a9f6..cfd645a38 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/get_rows_quant.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/get_rows_quant.comp @@ -1,5 +1,7 @@ #version 450 +#extension GL_EXT_control_flow_attributes : enable + #include "types.comp" #include "generic_binary_head.comp" #include "dequant_funcs.comp" @@ -12,7 +14,7 @@ void main() { const uint i11 = (gl_GlobalInvocationID.z)/p.ne12; const uint i12 = (gl_GlobalInvocationID.z)%p.ne12; -#if defined(DATA_A_IQ2_XXS) || defined(DATA_A_IQ2_XS) || defined(DATA_A_IQ2_S) || defined(DATA_A_IQ3_XXS) || defined(DATA_A_IQ3_S) || defined(DATA_A_IQ4_XS) || defined(DATA_A_IQ4_NL) +#ifdef NEEDS_INIT_IQ_SHMEM init_iq_shmem(gl_WorkGroupSize); #endif diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/im2col.comp b/ggml/src/ggml-vulkan/vulkan-shaders/im2col.comp index 122b1e93f..09aa849e8 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/im2col.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/im2col.comp @@ -40,6 +40,20 @@ void main() { const uint batch = gl_GlobalInvocationID.z / p.IC; const uint ic = gl_GlobalInvocationID.z % p.IC; + const uint src_base = ic * p.offset_delta + batch * p.batch_offset; + const uint dst_base = ((batch * p.OH + oh) * p.OW) * p.CHW + ic * (p.KW * p.KH); + const int oh_s1 = int(oh) * p.s1; + const uint ksize = p.OW * (p.KH > 1 ? p.KW : 1); + + const uint base_linear_idx = gidx * NUM_ITER; + + const uint max_ky = ksize / p.OW; + + uint current_kx = base_linear_idx / ksize; + const uint rem = base_linear_idx - (current_kx * ksize); + uint current_ky = rem / p.OW; + uint current_ix = rem % p.OW; + A_TYPE values[NUM_ITER]; uint offset_dst[NUM_ITER]; [[unroll]] for (uint idx = 0; idx < NUM_ITER; ++idx) { @@ -48,36 +62,35 @@ void main() { [[unroll]] for (uint idx = 0; idx < NUM_ITER; ++idx) { - const uint i = gidx * NUM_ITER + idx; + const uint linear_idx = base_linear_idx + idx; - const uint ksize = p.OW * (p.KH > 1 ? p.KW : 1); - const uint kx = i / ksize; - const uint kd = kx * ksize; - const uint ky = (i - kd) / p.OW; - const uint ix = i % p.OW; - - const uint iiw = ix * p.s0 + kx * p.d0 - p.p0; - const uint iih = oh * p.s1 + ky * p.d1 - p.p1; - - offset_dst[idx] = - ((batch * p.OH + oh) * p.OW + ix) * p.CHW + - (ic * (p.KW * p.KH) + ky * p.KW + kx); - - if (i >= p.pelements) { + if (linear_idx >= p.pelements) { continue; } - if (iih < p.IH && iiw < p.IW) { - const uint offset_src = ic * p.offset_delta + batch * p.batch_offset; - values[idx] = data_a[offset_src + iih * p.IW + iiw]; + const uint iiw = current_ix * p.s0 + current_kx * p.d0 - p.p0; + const uint iih = oh_s1 + current_ky * p.d1 - p.p1; + + offset_dst[idx] = dst_base + current_ix * p.CHW + current_ky * p.KW + current_kx; + + if ((iih < p.IH) && (iiw < p.IW)) { + values[idx] = data_a[src_base + iih * p.IW + iiw]; + } + + if (++current_ix == p.OW) { + current_ix = 0; + if (++current_ky == max_ky) { + current_ky = 0; + current_kx++; + } } } [[unroll]] for (uint idx = 0; idx < NUM_ITER; ++idx) { - const uint i = gidx * NUM_ITER + idx; + const uint linear_idx = base_linear_idx + idx; - if (i >= p.pelements) { + if (linear_idx >= p.pelements) { continue; } diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/l2_norm.comp b/ggml/src/ggml-vulkan/vulkan-shaders/l2_norm.comp new file mode 100644 index 000000000..deba8c398 --- /dev/null +++ b/ggml/src/ggml-vulkan/vulkan-shaders/l2_norm.comp @@ -0,0 +1,41 @@ +#version 450 + +#include "generic_head.comp" +#include "types.comp" + +#extension GL_EXT_control_flow_attributes : enable +#define BLOCK_SIZE 512 + +layout(local_size_x = BLOCK_SIZE, local_size_y = 1, local_size_z = 1) in; + +layout (binding = 0) readonly buffer X {A_TYPE data_a[];}; +layout (binding = 1) writeonly buffer D {D_TYPE data_d[];}; + +shared FLOAT_TYPE sum[BLOCK_SIZE]; + +void main() { + const uint row = gl_WorkGroupID.z * 262144 + gl_WorkGroupID.y * 512 + gl_WorkGroupID.x; + const uint tid = gl_LocalInvocationID.x; + + sum[tid] = FLOAT_TYPE(0.0f); // partial sum for thread in warp + + [[unroll]] for (uint col = tid; col < p.KX; col += BLOCK_SIZE) { + const FLOAT_TYPE xi = FLOAT_TYPE(data_a[row*p.KX + col]); + sum[tid] += xi * xi; + } + + // sum up partial sums and write back result + barrier(); + [[unroll]] for (int s = BLOCK_SIZE / 2; s > 0; s >>= 1) { + if (tid < s) { + sum[tid] += sum[tid + s]; + } + barrier(); + } + + const FLOAT_TYPE scale = inversesqrt(max(sum[0], FLOAT_TYPE(p.param1))); + + [[unroll]] for (uint col = tid; col < p.KX; col += BLOCK_SIZE) { + data_d[row*p.KX + col] = D_TYPE(scale * FLOAT_TYPE(data_a[row*p.KX + col])); + } +} diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec.comp index d7e99727d..775b48cd0 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec.comp @@ -105,6 +105,16 @@ void compute_outputs(const uint32_t first_row, const uint32_t num_rows) { int unroll_count = 4; uint unrolled_iters = num_iters & ~(unroll_count - 1); +#if K_PER_ITER == 2 + // If the K dimension is odd, we need lastiter==true on the last iteration + // so OOB is computed correctly. Skip some unrolling to make that happen. + if ((p.ncols & 1) != 0 && + unrolled_iters == num_iters && + unrolled_iters > 0) { + unrolled_iters -= unroll_count; + } +#endif + uint i = 0; while (i < unrolled_iters) { // Manually partially unroll the loop @@ -113,8 +123,18 @@ void compute_outputs(const uint32_t first_row, const uint32_t num_rows) { i++; } } + unroll_count = 2; unrolled_iters = num_iters & ~(unroll_count - 1); + +#if K_PER_ITER == 2 + if ((p.ncols & 1) != 0 && + unrolled_iters == num_iters && + unrolled_iters > 0) { + unrolled_iters -= unroll_count; + } +#endif + while (i < unrolled_iters) { // Manually partially unroll the loop [[unroll]] for (uint k = 0; k < unroll_count; ++k) { @@ -133,7 +153,7 @@ void compute_outputs(const uint32_t first_row, const uint32_t num_rows) { void main() { const uint first_row = NUM_ROWS * (gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z); -#if defined(DATA_A_IQ2_XXS) || defined(DATA_A_IQ2_XS) || defined(DATA_A_IQ2_S) || defined(DATA_A_IQ3_XXS) || defined(DATA_A_IQ3_S) || defined(DATA_A_IQ4_XS) || defined(DATA_A_IQ4_NL) +#ifdef NEEDS_INIT_IQ_SHMEM init_iq_shmem(gl_WorkGroupSize); #endif diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq1_m.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq1_m.comp new file mode 100644 index 000000000..e4acbd4f9 --- /dev/null +++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq1_m.comp @@ -0,0 +1,82 @@ +#version 450 +#extension GL_EXT_shader_explicit_arithmetic_types_int32 : require + +#include "mul_mat_vec_base.comp" + +layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in; + +FLOAT_TYPE temp[NUM_COLS][NUM_ROWS]; + +void calc_superblock(const uint a_offset, const uint b_offset, const uint ib32, const uint i, const uint num_blocks_per_row, const uint first_row, const uint num_rows) { + const uint y_idx = i * QUANT_K + 32 * ib32; + + uint ibi = a_offset / QUANT_K + first_row * num_blocks_per_row + i; + [[unroll]] for (uint n = 0; n < num_rows; ++n) { + const uint16_t[4] scales = data_a[ibi].scales; + const u16vec4 s = u16vec4(scales[0], scales[1], scales[2], scales[3]) >> 12; + const float d = float(unpackHalf2x16(s.x | (s.y << 4) | (s.z << 8) | (s.w << 12)).x); + + const uint sc = data_a[ibi].scales[ib32 / 2] >> (6 * (ib32 & 1)); + [[unroll]] for (uint l = 0; l < 4; ++l) { + const uint qh = data_a[ibi].qh[2 * ib32 + l / 2] >> (4 * (l&1)); + const uint qs = data_a[ibi].qs[4 * ib32 + l]; + const float delta = ((qh & 8) != 0) ? -IQ1M_DELTA : IQ1M_DELTA; + const float dl = d * (2 * bitfieldExtract(sc, 3 * int(l / 2), 3) + 1); + + const int16_t grid = int16_t(iq1s_grid[qs | ((qh & 7) << 8)]); + + [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) { + vec4 b0 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 + 2*l + 0]); + vec4 b4 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 + 2*l + 1]); + + FLOAT_TYPE sum = FLOAT_TYPE(0.0); + [[unroll]] for (int k = 0; k < 4; ++k) { + sum = fma(FLOAT_TYPE(b0[k]), bitfieldExtract(grid, 2 * k, 2) + delta, + fma(FLOAT_TYPE(b4[k]), bitfieldExtract(grid, 8 + 2 * k, 2) + delta, sum)); + } + temp[j][n] = fma(dl, sum, temp[j][n]); + } + } + ibi += num_blocks_per_row; + } +} + +void compute_outputs(const uint32_t first_row, const uint32_t num_rows) { + uint a_offset, b_offset, d_offset; + get_offsets(a_offset, b_offset, d_offset); + + const uint num_blocks_per_row = p.ncols / QUANT_K; + + // 8 threads are used to process each block + const uint blocks_per_wg = gl_WorkGroupSize.x/8; + const uint tid = gl_LocalInvocationID.x; + const uint itid = tid % 8; // 0...7 + const uint ix = tid / 8; + + [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) { + [[unroll]] for (uint i = 0; i < NUM_ROWS; ++i) { + temp[j][i] = FLOAT_TYPE(0); + } + } + + [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += blocks_per_wg) + calc_superblock(a_offset, b_offset, itid, i, num_blocks_per_row, first_row, num_rows); + + reduce_result(temp, d_offset, first_row, num_rows, tid); +} + +void main() { + const uint first_row = NUM_ROWS * (gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z); + + init_iq_shmem(gl_WorkGroupSize); + + // do NUM_ROWS at a time, unless there aren't enough remaining rows + if (first_row + NUM_ROWS <= p.stride_d) { + compute_outputs(first_row, NUM_ROWS); + } else { + if (first_row >= p.stride_d) { + return; + } + compute_outputs(first_row, p.stride_d - first_row); + } +} diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq1_s.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq1_s.comp new file mode 100644 index 000000000..309da0991 --- /dev/null +++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq1_s.comp @@ -0,0 +1,79 @@ +#version 450 +#extension GL_EXT_shader_explicit_arithmetic_types_int32 : require + +#include "mul_mat_vec_base.comp" + +layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in; + +FLOAT_TYPE temp[NUM_COLS][NUM_ROWS]; + +void calc_superblock(const uint a_offset, const uint b_offset, const uint ib32, const uint i, const uint num_blocks_per_row, const uint first_row, const uint num_rows) { + const uint y_idx = i * QUANT_K + 32 * ib32; + + uint ibi = a_offset / QUANT_K + first_row * num_blocks_per_row + i; + [[unroll]] for (uint n = 0; n < num_rows; ++n) { + const float d = float(data_a[ibi].d); + const uint qh = data_a[ibi].qh[ib32]; + const float dl = d * float(2 * bitfieldExtract(qh, 12, 3) + 1); + const float delta = ((qh & 0x8000) != 0) ? -IQ1S_DELTA : IQ1S_DELTA; + + [[unroll]] for (uint l = 0; l < 4; ++l) { + const uint qs = data_a[ibi].qs[4 * ib32 + l]; + const uint idxhi = bitfieldExtract(qh, 3 * int(l), 3); + const int16_t grid = int16_t(iq1s_grid[qs | (idxhi << 8)]); + + [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) { + vec4 b0 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 + 2*l + 0]); + vec4 b4 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 + 2*l + 1]); + + FLOAT_TYPE sum = FLOAT_TYPE(0.0); + [[unroll]] for (int k = 0; k < 4; ++k) { + sum = fma(FLOAT_TYPE(b0[k]), bitfieldExtract(grid, 2 * k, 2) + delta, + fma(FLOAT_TYPE(b4[k]), bitfieldExtract(grid, 8 + 2 * k, 2) + delta, sum)); + } + temp[j][n] = fma(dl, sum, temp[j][n]); + } + } + ibi += num_blocks_per_row; + } +} + +void compute_outputs(const uint32_t first_row, const uint32_t num_rows) { + uint a_offset, b_offset, d_offset; + get_offsets(a_offset, b_offset, d_offset); + + const uint num_blocks_per_row = p.ncols / QUANT_K; + + // 8 threads are used to process each block + const uint blocks_per_wg = gl_WorkGroupSize.x/8; + const uint tid = gl_LocalInvocationID.x; + const uint itid = tid % 8; // 0...7 + const uint ix = tid / 8; + + [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) { + [[unroll]] for (uint i = 0; i < NUM_ROWS; ++i) { + temp[j][i] = FLOAT_TYPE(0); + } + } + + [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += blocks_per_wg) + calc_superblock(a_offset, b_offset, itid, i, num_blocks_per_row, first_row, num_rows); + + reduce_result(temp, d_offset, first_row, num_rows, tid); +} + +void main() { + const uint first_row = NUM_ROWS * (gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z); + + init_iq_shmem(gl_WorkGroupSize); + + // do NUM_ROWS at a time, unless there aren't enough remaining rows + if (first_row + NUM_ROWS <= p.stride_d) { + compute_outputs(first_row, NUM_ROWS); + } else { + if (first_row >= p.stride_d) { + return; + } + compute_outputs(first_row, p.stride_d - first_row); + } +} diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq2_s.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq2_s.comp new file mode 100644 index 000000000..8d01536fa --- /dev/null +++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq2_s.comp @@ -0,0 +1,90 @@ +#version 450 +#extension GL_EXT_shader_explicit_arithmetic_types_int32 : require + +#include "mul_mat_vec_base.comp" + +layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in; + +FLOAT_TYPE temp[NUM_COLS][NUM_ROWS]; + +void calc_superblock(const uint a_offset, const uint b_offset, const uint itid, const uint i, const uint num_blocks_per_row, const uint first_row, const uint num_rows) { + const uint y_idx = i * QUANT_K + 16 * itid; + const uint nibble_shift = 4 * (itid & 1); + const uint ib32 = itid / 2; // 0..7 + + uint ibi = a_offset / QUANT_K + first_row * num_blocks_per_row + i; + [[unroll]] for (uint n = 0; n < num_rows; ++n) { + const float d = float(data_a[ibi].d); + const uint scale = (data_a[ibi].scales[ib32] >> nibble_shift) & 0xF; + const float db = d * (0.5 + scale) * 0.25; + + const uint qh = data_a[ibi].qh[ib32]; + const u8vec2 qs16 = unpack8(uint32_t(data_a_packed16[ibi].qs[itid])).xy; // vec4 used due to #12147 + const u8vec2 sign16 = unpack8(uint32_t(data_a_packed16[ibi].qs[QUANT_K / 16 + itid])).xy; + [[unroll]] for (uint l = 0; l < 2; ++l) { + const uint8_t sign = sign16[l]; + const uint qs = qs16[l] | ((qh << (8 - nibble_shift - 2 * l)) & 0x300); + const uvec2 grid = iq2s_grid[qs]; + const vec4 grid0 = vec4(unpack8(grid.x)); + const vec4 grid1 = vec4(unpack8(grid.y)); + + [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) { + vec4 b0 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 + 2*l + 0]); + vec4 b4 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 + 2*l + 1]); + + FLOAT_TYPE sum = + fma(FLOAT_TYPE(b0.x), FLOAT_TYPE((sign & 1) != 0 ? -grid0.x : grid0.x), + fma(FLOAT_TYPE(b0.y), FLOAT_TYPE((sign & 2) != 0 ? -grid0.y : grid0.y), + fma(FLOAT_TYPE(b0.z), FLOAT_TYPE((sign & 4) != 0 ? -grid0.z : grid0.z), + fma(FLOAT_TYPE(b0.w), FLOAT_TYPE((sign & 8) != 0 ? -grid0.w : grid0.w), + fma(FLOAT_TYPE(b4.x), FLOAT_TYPE((sign & 16) != 0 ? -grid1.x : grid1.x), + fma(FLOAT_TYPE(b4.y), FLOAT_TYPE((sign & 32) != 0 ? -grid1.y : grid1.y), + fma(FLOAT_TYPE(b4.z), FLOAT_TYPE((sign & 64) != 0 ? -grid1.z : grid1.z), + fma(FLOAT_TYPE(b4.w), FLOAT_TYPE((sign & 128) != 0 ? -grid1.w : grid1.w), + FLOAT_TYPE(0.0))))))))); + temp[j][n] = fma(db, sum, temp[j][n]); + } + } + ibi += num_blocks_per_row; + } +} + +void compute_outputs(const uint32_t first_row, const uint32_t num_rows) { + uint a_offset, b_offset, d_offset; + get_offsets(a_offset, b_offset, d_offset); + + const uint num_blocks_per_row = p.ncols / QUANT_K; + + // 16 threads are used to process each block + const uint blocks_per_wg = gl_WorkGroupSize.x/16; + const uint tid = gl_LocalInvocationID.x; + const uint itid = tid % 16; // 0...15 + const uint ix = tid / 16; + + [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) { + [[unroll]] for (uint i = 0; i < NUM_ROWS; ++i) { + temp[j][i] = FLOAT_TYPE(0); + } + } + + [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += blocks_per_wg) + calc_superblock(a_offset, b_offset, itid, i, num_blocks_per_row, first_row, num_rows); + + reduce_result(temp, d_offset, first_row, num_rows, tid); +} + +void main() { + const uint first_row = NUM_ROWS * (gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z); + + init_iq_shmem(gl_WorkGroupSize); + + // do NUM_ROWS at a time, unless there aren't enough remaining rows + if (first_row + NUM_ROWS <= p.stride_d) { + compute_outputs(first_row, NUM_ROWS); + } else { + if (first_row >= p.stride_d) { + return; + } + compute_outputs(first_row, p.stride_d - first_row); + } +} diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq2_xs.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq2_xs.comp new file mode 100644 index 000000000..c49604324 --- /dev/null +++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq2_xs.comp @@ -0,0 +1,87 @@ +#version 450 +#extension GL_EXT_shader_explicit_arithmetic_types_int32 : require + +#include "mul_mat_vec_base.comp" + +layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in; + +FLOAT_TYPE temp[NUM_COLS][NUM_ROWS]; + +void calc_superblock(const uint a_offset, const uint b_offset, const uint itid, const uint i, const uint num_blocks_per_row, const uint first_row, const uint num_rows) { + const uint y_idx = i * QUANT_K + 16 * itid; + const uint nibble_shift = 4 * (itid & 1); + const uint ib32 = itid / 2; // 0..7 + + uint ibi = a_offset / QUANT_K + first_row * num_blocks_per_row + i; + [[unroll]] for (uint n = 0; n < num_rows; ++n) { + const float d = float(data_a[ibi].d); + const uint scale = (data_a[ibi].scales[ib32] >> nibble_shift) & 0xF; + const float db = d * (0.5 + scale) * 0.25; + + [[unroll]] for (uint l = 0; l < 2; ++l) { + const uint qs = data_a[ibi].qs[2 * itid + l]; + const uint sign = qs >> 9; + const uint sign7 = bitCount(sign); + const vec4 grid0 = vec4(unpack8(iq2xs_grid[qs & 511].x)); + const vec4 grid1 = vec4(unpack8(iq2xs_grid[qs & 511].y)); + + [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) { + vec4 b0 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 + 2*l + 0]); + vec4 b4 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 + 2*l + 1]); + + FLOAT_TYPE sum = + fma(FLOAT_TYPE(b0.x), FLOAT_TYPE((sign & 1) != 0 ? -grid0.x : grid0.x), + fma(FLOAT_TYPE(b0.y), FLOAT_TYPE((sign & 2) != 0 ? -grid0.y : grid0.y), + fma(FLOAT_TYPE(b0.z), FLOAT_TYPE((sign & 4) != 0 ? -grid0.z : grid0.z), + fma(FLOAT_TYPE(b0.w), FLOAT_TYPE((sign & 8) != 0 ? -grid0.w : grid0.w), + fma(FLOAT_TYPE(b4.x), FLOAT_TYPE((sign & 16) != 0 ? -grid1.x : grid1.x), + fma(FLOAT_TYPE(b4.y), FLOAT_TYPE((sign & 32) != 0 ? -grid1.y : grid1.y), + fma(FLOAT_TYPE(b4.z), FLOAT_TYPE((sign & 64) != 0 ? -grid1.z : grid1.z), + fma(FLOAT_TYPE(b4.w), FLOAT_TYPE((sign7 & 1) != 0 ? -grid1.w : grid1.w), + FLOAT_TYPE(0.0))))))))); + temp[j][n] = fma(db, sum, temp[j][n]); + } + } + ibi += num_blocks_per_row; + } +} + +void compute_outputs(const uint32_t first_row, const uint32_t num_rows) { + uint a_offset, b_offset, d_offset; + get_offsets(a_offset, b_offset, d_offset); + + const uint num_blocks_per_row = p.ncols / QUANT_K; + + // 16 threads are used to process each block + const uint blocks_per_wg = gl_WorkGroupSize.x/16; + const uint tid = gl_LocalInvocationID.x; + const uint itid = tid % 16; // 0...15 + const uint ix = tid / 16; + + [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) { + [[unroll]] for (uint i = 0; i < NUM_ROWS; ++i) { + temp[j][i] = FLOAT_TYPE(0); + } + } + + [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += blocks_per_wg) + calc_superblock(a_offset, b_offset, itid, i, num_blocks_per_row, first_row, num_rows); + + reduce_result(temp, d_offset, first_row, num_rows, tid); +} + +void main() { + const uint first_row = NUM_ROWS * (gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z); + + init_iq_shmem(gl_WorkGroupSize); + + // do NUM_ROWS at a time, unless there aren't enough remaining rows + if (first_row + NUM_ROWS <= p.stride_d) { + compute_outputs(first_row, NUM_ROWS); + } else { + if (first_row >= p.stride_d) { + return; + } + compute_outputs(first_row, p.stride_d - first_row); + } +} diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq2_xxs.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq2_xxs.comp new file mode 100644 index 000000000..94d4b92e1 --- /dev/null +++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq2_xxs.comp @@ -0,0 +1,87 @@ +#version 450 +#extension GL_EXT_shader_explicit_arithmetic_types_int32 : require + +#include "mul_mat_vec_base.comp" + +layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in; + +FLOAT_TYPE temp[NUM_COLS][NUM_ROWS]; + +void calc_superblock(const uint a_offset, const uint b_offset, const uint itid, const uint i, const uint num_blocks_per_row, const uint first_row, const uint num_rows) { + const uint y_idx = i * QUANT_K + 16 * itid; + const uint ib32 = itid / 2; // 0..7 + + uint ibi = a_offset / QUANT_K + first_row * num_blocks_per_row + i; + [[unroll]] for (uint n = 0; n < num_rows; ++n) { + const float d = float(data_a[ibi].d); + const uint signscale = pack32(u16vec2( + data_a_packed16[ibi].qs[4 * ib32 + 2], + data_a_packed16[ibi].qs[4 * ib32 + 3])); + const float db = d * 0.25 * (0.5 + (signscale >> 28)); + [[unroll]] for (uint l = 0; l < 2; ++l) { + const uint qs = data_a[ibi].qs[8 * ib32 + 2 * (itid & 1) + l]; + const uint sign = bitfieldExtract(signscale, 7 * int(2 * (itid & 1) + l), 7); + const uint sign7 = bitCount(sign); + const vec4 grid0 = vec4(unpack8(iq2xxs_grid[qs].x)); + const vec4 grid1 = vec4(unpack8(iq2xxs_grid[qs].y)); + + [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) { + const vec4 b0 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 + 2*l + 0]); + const vec4 b4 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 + 2*l + 1]); + + FLOAT_TYPE sum = + fma(FLOAT_TYPE(b0.x), FLOAT_TYPE((sign & 1) != 0 ? -grid0.x : grid0.x), + fma(FLOAT_TYPE(b0.y), FLOAT_TYPE((sign & 2) != 0 ? -grid0.y : grid0.y), + fma(FLOAT_TYPE(b0.z), FLOAT_TYPE((sign & 4) != 0 ? -grid0.z : grid0.z), + fma(FLOAT_TYPE(b0.w), FLOAT_TYPE((sign & 8) != 0 ? -grid0.w : grid0.w), + fma(FLOAT_TYPE(b4.x), FLOAT_TYPE((sign & 16) != 0 ? -grid1.x : grid1.x), + fma(FLOAT_TYPE(b4.y), FLOAT_TYPE((sign & 32) != 0 ? -grid1.y : grid1.y), + fma(FLOAT_TYPE(b4.z), FLOAT_TYPE((sign & 64) != 0 ? -grid1.z : grid1.z), + fma(FLOAT_TYPE(b4.w), FLOAT_TYPE((sign7 & 1) != 0 ? -grid1.w : grid1.w), + FLOAT_TYPE(0.0))))))))); + temp[j][n] = fma(db, sum, temp[j][n]); + } + } + ibi += num_blocks_per_row; + } +} + +void compute_outputs(const uint32_t first_row, const uint32_t num_rows) { + uint a_offset, b_offset, d_offset; + get_offsets(a_offset, b_offset, d_offset); + + const uint num_blocks_per_row = p.ncols / QUANT_K; + + // 16 threads are used to process each block + const uint blocks_per_wg = gl_WorkGroupSize.x/16; + const uint tid = gl_LocalInvocationID.x; + const uint itid = tid % 16; // 0...15 + const uint ix = tid / 16; + + [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) { + [[unroll]] for (uint i = 0; i < NUM_ROWS; ++i) { + temp[j][i] = FLOAT_TYPE(0); + } + } + + [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += blocks_per_wg) + calc_superblock(a_offset, b_offset, itid, i, num_blocks_per_row, first_row, num_rows); + + reduce_result(temp, d_offset, first_row, num_rows, tid); +} + +void main() { + const uint first_row = NUM_ROWS * (gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z); + + init_iq_shmem(gl_WorkGroupSize); + + // do NUM_ROWS at a time, unless there aren't enough remaining rows + if (first_row + NUM_ROWS <= p.stride_d) { + compute_outputs(first_row, NUM_ROWS); + } else { + if (first_row >= p.stride_d) { + return; + } + compute_outputs(first_row, p.stride_d - first_row); + } +} diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq3_s.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq3_s.comp new file mode 100644 index 000000000..f021e4047 --- /dev/null +++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq3_s.comp @@ -0,0 +1,90 @@ +#version 450 +#extension GL_EXT_shader_explicit_arithmetic_types_int32 : require + +#include "mul_mat_vec_base.comp" + +layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in; + +FLOAT_TYPE temp[NUM_COLS][NUM_ROWS]; + +void calc_superblock(const uint a_offset, const uint b_offset, const uint ib32, const uint i, const uint num_blocks_per_row, const uint first_row, const uint num_rows) { + const uint y_idx = i * QUANT_K + 32 * ib32; + + uint ibi = a_offset / QUANT_K + first_row * num_blocks_per_row + i; + [[unroll]] for (uint n = 0; n < num_rows; ++n) { + const float d = float(data_a[ibi].d); + const uint scale = (data_a[ibi].scales[ib32/2] >> (4 * (ib32 & 1))) & 0xF; + const float dscale = d * (1 + 2 * scale); + const uint qh = data_a[ibi].qh[ib32]; + FLOAT_TYPE sum[NUM_COLS]; + [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) { + sum[j] = 0.0; + } + [[unroll]] for (uint l = 0; l < 4; ++l) { + const u8vec2 qs = unpack8(uint32_t(data_a_packed16[ibi].qs[4 * ib32 + l])).xy; // vec4 used due to #12147 + const uint sign = data_a[ibi].signs[4 * ib32 + l]; + const vec4 grid0 = vec4(unpack8(iq3s_grid[qs.x | ((qh << (8 - 2*l)) & 0x100)])); + const vec4 grid1 = vec4(unpack8(iq3s_grid[qs.y | ((qh << (7 - 2*l)) & 0x100)])); + + [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) { + const vec4 b0 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 + 2*l + 0]); + const vec4 b4 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 + 2*l + 1]); + + sum[j] = + fma(FLOAT_TYPE(b0.x), FLOAT_TYPE((sign & 1) != 0 ? -grid0.x : grid0.x), + fma(FLOAT_TYPE(b0.y), FLOAT_TYPE((sign & 2) != 0 ? -grid0.y : grid0.y), + fma(FLOAT_TYPE(b0.z), FLOAT_TYPE((sign & 4) != 0 ? -grid0.z : grid0.z), + fma(FLOAT_TYPE(b0.w), FLOAT_TYPE((sign & 8) != 0 ? -grid0.w : grid0.w), + fma(FLOAT_TYPE(b4.x), FLOAT_TYPE((sign & 16) != 0 ? -grid1.x : grid1.x), + fma(FLOAT_TYPE(b4.y), FLOAT_TYPE((sign & 32) != 0 ? -grid1.y : grid1.y), + fma(FLOAT_TYPE(b4.z), FLOAT_TYPE((sign & 64) != 0 ? -grid1.z : grid1.z), + fma(FLOAT_TYPE(b4.w), FLOAT_TYPE((sign & 128) != 0 ? -grid1.w : grid1.w), + sum[j])))))))); + } + } + [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) { + temp[j][n] = fma(dscale, sum[j], temp[j][n]); + } + ibi += num_blocks_per_row; + } +} + +void compute_outputs(const uint32_t first_row, const uint32_t num_rows) { + uint a_offset, b_offset, d_offset; + get_offsets(a_offset, b_offset, d_offset); + + const uint num_blocks_per_row = p.ncols / QUANT_K; + + // 8 threads are used to process each block + const uint blocks_per_wg = gl_WorkGroupSize.x/8; + const uint tid = gl_LocalInvocationID.x; + const uint itid = tid % 8; // 0...7 + const uint ix = tid / 8; + + [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) { + [[unroll]] for (uint i = 0; i < NUM_ROWS; ++i) { + temp[j][i] = FLOAT_TYPE(0); + } + } + + [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += blocks_per_wg) + calc_superblock(a_offset, b_offset, itid, i, num_blocks_per_row, first_row, num_rows); + + reduce_result(temp, d_offset, first_row, num_rows, tid); +} + +void main() { + const uint first_row = NUM_ROWS * (gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z); + + init_iq_shmem(gl_WorkGroupSize); + + // do NUM_ROWS at a time, unless there aren't enough remaining rows + if (first_row + NUM_ROWS <= p.stride_d) { + compute_outputs(first_row, NUM_ROWS); + } else { + if (first_row >= p.stride_d) { + return; + } + compute_outputs(first_row, p.stride_d - first_row); + } +} diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq3_xxs.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq3_xxs.comp new file mode 100644 index 000000000..3fe9dc3a4 --- /dev/null +++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq3_xxs.comp @@ -0,0 +1,88 @@ +#version 450 +#extension GL_EXT_shader_explicit_arithmetic_types_int32 : require + +#include "mul_mat_vec_base.comp" + +layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in; + +FLOAT_TYPE temp[NUM_COLS][NUM_ROWS]; + +void calc_superblock(const uint a_offset, const uint b_offset, const uint itid, const uint i, const uint num_blocks_per_row, const uint first_row, const uint num_rows) { + const uint y_idx = i * QUANT_K + 16 * itid; + const uint ib32 = itid / 2; // 0..7 + + uint ibi = a_offset / QUANT_K + first_row * num_blocks_per_row + i; + [[unroll]] for (uint n = 0; n < num_rows; ++n) { + const float d = float(data_a[ibi].d); + const uint signscale = pack32(u16vec2( + data_a_packed16[ibi].qs[QUANT_K / 8 + 2 * ib32], + data_a_packed16[ibi].qs[QUANT_K / 8 + 2 * ib32 + 1])); + const float db = d * 0.5 * (0.5 + (signscale >> 28)); + [[unroll]] for (uint l = 0; l < 2; ++l) { + const uint qs0 = data_a[ibi].qs[8 * ib32 + 4 * (itid & 1) + 2 * l]; + const uint qs1 = data_a[ibi].qs[8 * ib32 + 4 * (itid & 1) + 2 * l + 1]; + const uint sign = bitfieldExtract(signscale, 7 * int(2 * (itid & 1) + l), 7); + const uint sign7 = bitCount(sign); + const vec4 grid0 = vec4(unpack8(iq3xxs_grid[qs0])); + const vec4 grid1 = vec4(unpack8(iq3xxs_grid[qs1])); + + [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) { + const vec4 b0 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 + 2*l + 0]); + const vec4 b4 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 + 2*l + 1]); + + FLOAT_TYPE sum = + fma(FLOAT_TYPE(b0.x), FLOAT_TYPE((sign & 1) != 0 ? -grid0.x : grid0.x), + fma(FLOAT_TYPE(b0.y), FLOAT_TYPE((sign & 2) != 0 ? -grid0.y : grid0.y), + fma(FLOAT_TYPE(b0.z), FLOAT_TYPE((sign & 4) != 0 ? -grid0.z : grid0.z), + fma(FLOAT_TYPE(b0.w), FLOAT_TYPE((sign & 8) != 0 ? -grid0.w : grid0.w), + fma(FLOAT_TYPE(b4.x), FLOAT_TYPE((sign & 16) != 0 ? -grid1.x : grid1.x), + fma(FLOAT_TYPE(b4.y), FLOAT_TYPE((sign & 32) != 0 ? -grid1.y : grid1.y), + fma(FLOAT_TYPE(b4.z), FLOAT_TYPE((sign & 64) != 0 ? -grid1.z : grid1.z), + fma(FLOAT_TYPE(b4.w), FLOAT_TYPE((sign7 & 1) != 0 ? -grid1.w : grid1.w), + FLOAT_TYPE(0.0))))))))); + temp[j][n] = fma(db, sum, temp[j][n]); + } + } + ibi += num_blocks_per_row; + } +} + +void compute_outputs(const uint32_t first_row, const uint32_t num_rows) { + uint a_offset, b_offset, d_offset; + get_offsets(a_offset, b_offset, d_offset); + + const uint num_blocks_per_row = p.ncols / QUANT_K; + + // 16 threads are used to process each block + const uint blocks_per_wg = gl_WorkGroupSize.x/16; + const uint tid = gl_LocalInvocationID.x; + const uint itid = tid % 16; // 0...15 + const uint ix = tid / 16; + + [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) { + [[unroll]] for (uint i = 0; i < NUM_ROWS; ++i) { + temp[j][i] = FLOAT_TYPE(0); + } + } + + [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += blocks_per_wg) + calc_superblock(a_offset, b_offset, itid, i, num_blocks_per_row, first_row, num_rows); + + reduce_result(temp, d_offset, first_row, num_rows, tid); +} + +void main() { + const uint first_row = NUM_ROWS * (gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z); + + init_iq_shmem(gl_WorkGroupSize); + + // do NUM_ROWS at a time, unless there aren't enough remaining rows + if (first_row + NUM_ROWS <= p.stride_d) { + compute_outputs(first_row, NUM_ROWS); + } else { + if (first_row >= p.stride_d) { + return; + } + compute_outputs(first_row, p.stride_d - first_row); + } +} diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_nc.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_nc.comp index 1cc4996d3..48376637f 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_nc.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_nc.comp @@ -12,6 +12,9 @@ layout (binding = 0) readonly buffer A {A_TYPE data_a[];}; layout (binding = 1) readonly buffer B {B_TYPE data_b[];}; layout (binding = 2) writeonly buffer D {D_TYPE dst[];}; +layout (binding = 0) readonly buffer AV4 {A_TYPE_VEC4 data_a_v4[];}; +layout (binding = 1) readonly buffer BV4 {B_TYPE_VEC4 data_b_v4[];}; + layout (push_constant) uniform parameter { uint ncols_x; @@ -37,25 +40,66 @@ void main() { const uint idst = channel*nrows_dst + row_dst; - tmp[tid] = 0.0f; + FLOAT_TYPE temp = 0.0f; - for (uint col_x0 = 0; col_x0 < p.ncols_x; col_x0 += BLOCK_SIZE) { - const uint col_x = col_x0 + tid; + // Detect alignment for vector loads + bool is_aligned = (p.ncols_x % 4) == 0 && (p.row_stride_x % 4) == 0 && (p.channel_stride_x % 4) == 0; - if (col_x >= p.ncols_x) { - break; + for (uint col_x0 = 0; col_x0 < p.ncols_x;) { + + // Unroll 2x and do vec4 loads if aligned + const uint unroll_count = 2; + if (col_x0 + unroll_count * 4 * BLOCK_SIZE <= p.ncols_x && is_aligned) { + [[unroll]] for (uint i = 0; i < unroll_count; ++i) { + const uint col_x = col_x0 + 4*tid; + + const uint row_y = col_x; + + const uint ix = channel_x*p.channel_stride_x + row_x*p.row_stride_x + col_x; + const uint iy = channel*nrows_y + row_y; + + const vec4 av4 = vec4(data_a_v4[ix / 4]); + const vec4 bv4 = vec4(data_b_v4[iy / 4]); + + temp += dot(av4, bv4); + + col_x0 += 4*BLOCK_SIZE; + } + // do vec4 loads if aligned + } else if (col_x0 + 4*BLOCK_SIZE <= p.ncols_x && is_aligned) { + const uint col_x = col_x0 + 4*tid; + + const uint row_y = col_x; + + const uint ix = channel_x*p.channel_stride_x + row_x*p.row_stride_x + col_x; + const uint iy = channel*nrows_y + row_y; + + const vec4 av4 = vec4(data_a_v4[ix / 4]); + const vec4 bv4 = vec4(data_b_v4[iy / 4]); + + temp += dot(av4, bv4); + + col_x0 += 4*BLOCK_SIZE; + } else { + const uint col_x = col_x0 + tid; + if (col_x >= p.ncols_x) { + break; + } + + const uint row_y = col_x; + + const uint ix = channel_x*p.channel_stride_x + row_x*p.row_stride_x + col_x; + const uint iy = channel*nrows_y + row_y; + + const FLOAT_TYPE xi = FLOAT_TYPE(data_a[ix]); + + temp = fma(xi, FLOAT_TYPE(data_b[iy]), temp); + col_x0 += BLOCK_SIZE; } - - const uint row_y = col_x; - - const uint ix = channel_x*p.channel_stride_x + row_x*p.row_stride_x + col_x; - const uint iy = channel*nrows_y + row_y; - - const FLOAT_TYPE xi = FLOAT_TYPE(data_a[ix]); - - tmp[tid] = fma(xi, FLOAT_TYPE(data_b[iy]), tmp[tid]); } + tmp[tid] = temp; + // sum up partial sums and write back result barrier(); [[unroll]] for (int s = BLOCK_SIZE / 2; s > 0; s >>= 1) { diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_p021.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_p021.comp index 9b443807d..7aa070eeb 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_p021.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_p021.comp @@ -2,16 +2,25 @@ #extension GL_EXT_control_flow_attributes : enable #extension GL_EXT_shader_16bit_storage : require +#if USE_SUBGROUP_ADD +#extension GL_KHR_shader_subgroup_arithmetic : enable +#endif -#define BLOCK_SIZE 32 #define FLOAT_TYPE float -layout(local_size_x = BLOCK_SIZE, local_size_y = 1, local_size_z = 1) in; +layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in; layout (binding = 0) readonly buffer A {A_TYPE data_a[];}; layout (binding = 1) readonly buffer B {B_TYPE data_b[];}; layout (binding = 2) writeonly buffer D {D_TYPE dst[];}; +layout (binding = 0) readonly buffer AV4 {A_TYPE_VEC4 data_a_v4[];}; +layout (binding = 1) readonly buffer BV4 {B_TYPE_VEC4 data_b_v4[];}; + +layout(constant_id = 0) const int BLOCK_SIZE = 32; +// gqa_ratio is in the range [1,8] +layout(constant_id = 1) const uint gqa_ratio = 1; + layout (push_constant) uniform parameter { uint ncols_x; @@ -22,52 +31,124 @@ layout (push_constant) uniform parameter uint d_offset; } p; -shared FLOAT_TYPE tmp[BLOCK_SIZE]; +#if !USE_SUBGROUP_ADD +shared FLOAT_TYPE tmp[8][BLOCK_SIZE]; +#endif void main() { const uint tid = gl_LocalInvocationID.x; const uint row_x = gl_GlobalInvocationID.y; - const uint channel = gl_GlobalInvocationID.z; - const uint channel_x = channel / (p.nchannels_y / p.nchannels_x); + + uint channel, channel_x; + + // When gqa_ratio > 1, each invocation does multiple rows. + // The row in the A matrix is starting from channel / gqa_ratio and the + // rows in the B matrix are [channel, channel+gqa_ratio). + // When gpa_ratio is 1, each invocation does one row. + if (gqa_ratio > 1) { + channel_x = gl_GlobalInvocationID.z; + channel = channel_x * gqa_ratio; + } else { + channel = gl_GlobalInvocationID.z; + channel_x = channel / (p.nchannels_y / p.nchannels_x);; + } const uint nrows_y = p.ncols_x; const uint nrows_dst = p.nrows_x; const uint row_dst = row_x; - tmp[tid] = FLOAT_TYPE(0.0f); - - for (uint col_x0 = 0; col_x0 < p.ncols_x; col_x0 += BLOCK_SIZE) { - const uint col_x = col_x0 + tid; - - if (col_x >= p.ncols_x) { - break; - } - - // x is transposed and permuted - const uint ix = row_x*p.nchannels_x*p.ncols_x + channel_x*p.ncols_x + col_x; - const FLOAT_TYPE xi = FLOAT_TYPE(data_a[ix]); - - const uint row_y = col_x; - - // y is not transposed but permuted - const uint iy = channel*nrows_y + row_y; - - tmp[tid] = fma(xi, FLOAT_TYPE(data_b[iy]), tmp[tid]); + FLOAT_TYPE temp[8]; + [[unroll]] for (uint i = 0; i < 8; ++i) { + temp[i] = FLOAT_TYPE(0.0f); } - // dst is not transposed and not permuted - const uint idst = channel*nrows_dst + row_dst; + // Detect alignment for vector loads + bool is_aligned = (p.ncols_x % 4) == 0 && (p.nchannels_x % 4) == 0 && (nrows_y % 4) == 0; + for (uint col_x0 = 0; col_x0 < p.ncols_x; col_x0 += BLOCK_SIZE) { + + // Use vec4 loads if aligned + if (col_x0 + 4*BLOCK_SIZE <= p.ncols_x && is_aligned) { + + uint col_x = col_x0 + 4*tid; + const uint row_y = col_x; + + // x is transposed and permuted + const uint ix = row_x*p.nchannels_x*p.ncols_x + channel_x*p.ncols_x + col_x; + const vec4 av4 = vec4(data_a_v4[ix / 4]); + + [[unroll]] for (uint c = 0; c < gqa_ratio; ++c) { + // y is not transposed but permuted + const uint iy = (channel + c)*nrows_y + row_y; + + vec4 bv4 = data_b_v4[iy / 4]; + temp[c] += dot(av4, bv4); + } + + col_x0 += 3*BLOCK_SIZE; + } else { + const uint col_x = col_x0 + tid; + + if (col_x >= p.ncols_x) { + break; + } + + // x is transposed and permuted + const uint ix = row_x*p.nchannels_x*p.ncols_x + channel_x*p.ncols_x + col_x; + const FLOAT_TYPE xi = FLOAT_TYPE(data_a[ix]); + + const uint row_y = col_x; + + [[unroll]] for (uint c = 0; c < gqa_ratio; ++c) { + // y is not transposed but permuted + const uint iy = (channel + c)*nrows_y + row_y; + + temp[c] = fma(xi, FLOAT_TYPE(data_b[iy]), temp[c]); + } + } + } + +#if USE_SUBGROUP_ADD + // reduce vec4 at a time + vec4 t = vec4(temp[0], temp[1], temp[2], temp[3]); + t = subgroupAdd(t); + temp[0] = t[0]; + temp[1] = t[1]; + temp[2] = t[2]; + temp[3] = t[3]; + if (gqa_ratio > 4) { + t = vec4(temp[4], temp[5], temp[6], temp[7]); + t = subgroupAdd(t); + temp[4] = t[0]; + temp[5] = t[1]; + temp[6] = t[2]; + temp[7] = t[3]; + } +#else + [[unroll]] for (uint c = 0; c < gqa_ratio; ++c) { + tmp[c][tid] = temp[c]; + } // sum up partial sums and write back result barrier(); [[unroll]] for (int s = BLOCK_SIZE / 2; s > 0; s >>= 1) { if (tid < s) { - tmp[tid] += tmp[tid + s]; + [[unroll]] for (uint c = 0; c < gqa_ratio; ++c) { + temp[c] += tmp[c][tid + s]; + tmp[c][tid] = temp[c]; + } } barrier(); } + [[unroll]] for (uint c = 0; c < gqa_ratio; ++c) { + temp[c] = tmp[c][tid]; + } +#endif if (tid == 0) { - dst[idst] = tmp[0]; + [[unroll]] for (uint c = 0; c < gqa_ratio; ++c) { + // dst is not transposed and not permuted + const uint idst = (channel + c)*nrows_dst + row_dst; + dst[idst] = temp[c]; + } } } diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q2_k.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q2_k.comp index 8cdc640e8..423ceb8a3 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q2_k.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q2_k.comp @@ -5,23 +5,24 @@ layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in; -shared FLOAT_TYPE sccache1[BLOCK_SIZE/16][16]; -shared FLOAT_TYPE sccache2[BLOCK_SIZE/16][16]; +shared FLOAT_TYPE sccache1[2][BLOCK_SIZE/16][16]; +shared FLOAT_TYPE sccache2[2][BLOCK_SIZE/16][16]; FLOAT_TYPE temp[NUM_COLS][NUM_ROWS]; +uint csel = 0; void calc_superblock(const uint a_offset, const uint b_offset, const uint itid, const uint v_im, const uint ix, const uint q_offset, const uint y_offset, const uint i, const uint num_blocks_per_row, const uint first_row, const uint num_rows, const bool all_threads) { const uint y_idx = i * QUANT_K + y_offset; [[unroll]] for (uint n = 0; n < num_rows; ++n) { const uint ib0 = a_offset / QUANT_K + (first_row+n)*num_blocks_per_row; + csel ^= 1; - barrier(); if (!all_threads) { // when we don't have enough blocks to use all threads if (i < num_blocks_per_row) { const uint32_t scale = uint32_t(data_a[ib0 + i].scales[itid]); - sccache1[ix][itid] = FLOAT_TYPE(scale & 0xF); - sccache2[ix][itid] = FLOAT_TYPE((scale >> 4) & 0xF); + sccache1[csel][ix][itid] = FLOAT_TYPE(scale & 0xF); + sccache2[csel][ix][itid] = FLOAT_TYPE((scale >> 4) & 0xF); } barrier(); @@ -29,8 +30,8 @@ void calc_superblock(const uint a_offset, const uint b_offset, const uint itid, continue; } else { const uint32_t scale = uint32_t(data_a[ib0 + i].scales[itid]); - sccache1[ix][itid] = FLOAT_TYPE(scale & 0xF); - sccache2[ix][itid] = FLOAT_TYPE((scale >> 4) & 0xF); + sccache1[csel][ix][itid] = FLOAT_TYPE(scale & 0xF); + sccache2[csel][ix][itid] = FLOAT_TYPE((scale >> 4) & 0xF); barrier(); } @@ -57,22 +58,22 @@ void calc_superblock(const uint a_offset, const uint b_offset, const uint itid, FLOAT_TYPE sum1 = FLOAT_TYPE(0.0); FLOAT_TYPE sum2 = FLOAT_TYPE(0.0); [[unroll]] for (int l = 0; l < 2; ++l) { - sum1 = fma(FLOAT_TYPE(b0[l]), sccache1[ix][ 8*v_im] * qs_u32_0[l ], - fma(FLOAT_TYPE(b16[l]), sccache1[ix][1 + 8*v_im] * qs_u32_0[l+2], - fma(FLOAT_TYPE(b32[l]), sccache1[ix][2 + 8*v_im] * qs_u32_2[l ], - fma(FLOAT_TYPE(b48[l]), sccache1[ix][3 + 8*v_im] * qs_u32_2[l+2], - fma(FLOAT_TYPE(b64[l]), sccache1[ix][4 + 8*v_im] * qs_u32_4[l ], - fma(FLOAT_TYPE(b80[l]), sccache1[ix][5 + 8*v_im] * qs_u32_4[l+2], - fma(FLOAT_TYPE(b96[l]), sccache1[ix][6 + 8*v_im] * qs_u32_6[l ], - fma(FLOAT_TYPE(b112[l]), sccache1[ix][7 + 8*v_im] * qs_u32_6[l+2], sum1)))))))); - sum2 = fma(FLOAT_TYPE(b0[l]), sccache2[ix][ 8*v_im], - fma(FLOAT_TYPE(b16[l]), sccache2[ix][1 + 8*v_im], - fma(FLOAT_TYPE(b32[l]), sccache2[ix][2 + 8*v_im], - fma(FLOAT_TYPE(b48[l]), sccache2[ix][3 + 8*v_im], - fma(FLOAT_TYPE(b64[l]), sccache2[ix][4 + 8*v_im], - fma(FLOAT_TYPE(b80[l]), sccache2[ix][5 + 8*v_im], - fma(FLOAT_TYPE(b96[l]), sccache2[ix][6 + 8*v_im], - fma(FLOAT_TYPE(b112[l]), sccache2[ix][7 + 8*v_im], sum2)))))))); + sum1 = fma(FLOAT_TYPE(b0[l]), sccache1[csel][ix][ 8*v_im] * qs_u32_0[l ], + fma(FLOAT_TYPE(b16[l]), sccache1[csel][ix][1 + 8*v_im] * qs_u32_0[l+2], + fma(FLOAT_TYPE(b32[l]), sccache1[csel][ix][2 + 8*v_im] * qs_u32_2[l ], + fma(FLOAT_TYPE(b48[l]), sccache1[csel][ix][3 + 8*v_im] * qs_u32_2[l+2], + fma(FLOAT_TYPE(b64[l]), sccache1[csel][ix][4 + 8*v_im] * qs_u32_4[l ], + fma(FLOAT_TYPE(b80[l]), sccache1[csel][ix][5 + 8*v_im] * qs_u32_4[l+2], + fma(FLOAT_TYPE(b96[l]), sccache1[csel][ix][6 + 8*v_im] * qs_u32_6[l ], + fma(FLOAT_TYPE(b112[l]), sccache1[csel][ix][7 + 8*v_im] * qs_u32_6[l+2], sum1)))))))); + sum2 = fma(FLOAT_TYPE(b0[l]), sccache2[csel][ix][ 8*v_im], + fma(FLOAT_TYPE(b16[l]), sccache2[csel][ix][1 + 8*v_im], + fma(FLOAT_TYPE(b32[l]), sccache2[csel][ix][2 + 8*v_im], + fma(FLOAT_TYPE(b48[l]), sccache2[csel][ix][3 + 8*v_im], + fma(FLOAT_TYPE(b64[l]), sccache2[csel][ix][4 + 8*v_im], + fma(FLOAT_TYPE(b80[l]), sccache2[csel][ix][5 + 8*v_im], + fma(FLOAT_TYPE(b96[l]), sccache2[csel][ix][6 + 8*v_im], + fma(FLOAT_TYPE(b112[l]), sccache2[csel][ix][7 + 8*v_im], sum2)))))))); } temp[j][n] = fma(dall, sum1, fma(-dmin, sum2, temp[j][n])); } diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q3_k.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q3_k.comp index 3116fad16..e91724a28 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q3_k.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q3_k.comp @@ -5,20 +5,21 @@ layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in; -shared FLOAT_TYPE sccache[BLOCK_SIZE/16][2][8]; +shared FLOAT_TYPE sccache[2][BLOCK_SIZE/16][2][8]; FLOAT_TYPE temp[NUM_COLS][NUM_ROWS]; +uint csel = 0; void calc_superblock(const uint a_offset, const uint b_offset, const uint ix, const uint itid8, const uint v_im, const uint v_im4, const uint v_in, const uint32_t hm_m[4], const uint q_offset, const uint y_offset, const uint s_shift, const uint i, const uint num_blocks_per_row, const uint first_row, const uint num_rows, const bool all_threads) { const uint y_idx = i * QUANT_K + y_offset; [[unroll]] for (uint n = 0; n < num_rows; ++n) { const uint ib0 = a_offset / QUANT_K + (first_row+n)*num_blocks_per_row; + csel ^= 1; if (!all_threads) { // when we don't have enough blocks to use all threads - barrier(); if (i < num_blocks_per_row) - sccache[ix][v_im][itid8] = FLOAT_TYPE(int8_t(((data_a[ib0+i].scales[itid8] >> v_im4) & 0xF) | (((data_a[ib0+i].scales[itid8%4+8] >> s_shift) & 3) << 4)) - 32); + sccache[csel][ix][v_im][itid8] = FLOAT_TYPE(int8_t(((data_a[ib0+i].scales[itid8] >> v_im4) & 0xF) | (((data_a[ib0+i].scales[itid8%4+8] >> s_shift) & 3) << 4)) - 32); barrier(); if (i >= num_blocks_per_row) @@ -40,8 +41,7 @@ void calc_superblock(const uint a_offset, const uint b_offset, const uint ix, co const vec4 qs_u32_6 = vec4(unpack8((qs_u32 >> 6) & 0x03030303)); if (all_threads) { - barrier(); - sccache[ix][v_im][itid8] = FLOAT_TYPE(int8_t(((data_a[ib0+i].scales[itid8] >> v_im4) & 0xF) | (((data_a[ib0+i].scales[itid8%4+8] >> s_shift) & 3) << 4)) - 32); + sccache[csel][ix][v_im][itid8] = FLOAT_TYPE(int8_t(((data_a[ib0+i].scales[itid8] >> v_im4) & 0xF) | (((data_a[ib0+i].scales[itid8%4+8] >> s_shift) & 3) << 4)) - 32); barrier(); } @@ -59,14 +59,14 @@ void calc_superblock(const uint a_offset, const uint b_offset, const uint ix, co FLOAT_TYPE sum = FLOAT_TYPE(0.0); [[unroll]] for (int l = 0; l < 2; ++l) { - sum = fma(FLOAT_TYPE( b0[l]) * sccache[ix][v_im][0], qs_u32_0[l ] - hmk_0[l ], - fma(FLOAT_TYPE( b16[l]) * sccache[ix][v_im][1], qs_u32_0[l+2] - hmk_0[l+2], - fma(FLOAT_TYPE( b32[l]) * sccache[ix][v_im][2], qs_u32_2[l ] - hmk_1[l ], - fma(FLOAT_TYPE( b48[l]) * sccache[ix][v_im][3], qs_u32_2[l+2] - hmk_1[l+2], - fma(FLOAT_TYPE( b64[l]) * sccache[ix][v_im][4], qs_u32_4[l ] - hmk_2[l ], - fma(FLOAT_TYPE( b80[l]) * sccache[ix][v_im][5], qs_u32_4[l+2] - hmk_2[l+2], - fma(FLOAT_TYPE( b96[l]) * sccache[ix][v_im][6], qs_u32_6[l ] - hmk_3[l ], - fma(FLOAT_TYPE(b112[l]) * sccache[ix][v_im][7], qs_u32_6[l+2] - hmk_3[l+2], sum)))))))); + sum = fma(FLOAT_TYPE( b0[l]) * sccache[csel][ix][v_im][0], qs_u32_0[l ] - hmk_0[l ], + fma(FLOAT_TYPE( b16[l]) * sccache[csel][ix][v_im][1], qs_u32_0[l+2] - hmk_0[l+2], + fma(FLOAT_TYPE( b32[l]) * sccache[csel][ix][v_im][2], qs_u32_2[l ] - hmk_1[l ], + fma(FLOAT_TYPE( b48[l]) * sccache[csel][ix][v_im][3], qs_u32_2[l+2] - hmk_1[l+2], + fma(FLOAT_TYPE( b64[l]) * sccache[csel][ix][v_im][4], qs_u32_4[l ] - hmk_2[l ], + fma(FLOAT_TYPE( b80[l]) * sccache[csel][ix][v_im][5], qs_u32_4[l+2] - hmk_2[l+2], + fma(FLOAT_TYPE( b96[l]) * sccache[csel][ix][v_im][6], qs_u32_6[l ] - hmk_3[l ], + fma(FLOAT_TYPE(b112[l]) * sccache[csel][ix][v_im][7], qs_u32_6[l+2] - hmk_3[l+2], sum)))))))); } temp[j][n] = fma(d, sum, temp[j][n]); } diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q6_k.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q6_k.comp index f05f96b5e..d53d9ee0a 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q6_k.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q6_k.comp @@ -6,20 +6,21 @@ layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in; -shared FLOAT_TYPE sccache[BLOCK_SIZE/16][16]; +shared FLOAT_TYPE sccache[2][BLOCK_SIZE/16][16]; FLOAT_TYPE temp[NUM_COLS][NUM_ROWS]; +uint csel = 0; void calc_superblock(const uint a_offset, const uint b_offset, const uint itid, const uint ix, const uint ql_offset, const uint qh_offset, const uint s_offset, const uint y_offset, const uint i, const uint num_blocks_per_row, const uint first_row, const uint num_rows, const bool all_threads) { const uint y_idx = i * QUANT_K + y_offset; [[unroll]] for (uint n = 0; n < num_rows; ++n) { const uint ib0 = a_offset / QUANT_K + (first_row+n)*num_blocks_per_row; + csel ^= 1; if (!all_threads) { // when we don't have enough blocks to use all threads - barrier(); if (i < num_blocks_per_row) - sccache[ix][itid] = FLOAT_TYPE(data_a[ib0 + i].scales[itid]); + sccache[csel][ix][itid] = FLOAT_TYPE(data_a[ib0 + i].scales[itid]); barrier(); if (i >= num_blocks_per_row) @@ -51,8 +52,7 @@ void calc_superblock(const uint a_offset, const uint b_offset, const uint itid, const vec4 q3 = vec4(unpack8(q3_u32)) - 32; if (all_threads) { - barrier(); - sccache[ix][itid] = FLOAT_TYPE(data_a[ib0 + i].scales[itid]); + sccache[csel][ix][itid] = FLOAT_TYPE(data_a[ib0 + i].scales[itid]); barrier(); } @@ -71,7 +71,7 @@ void calc_superblock(const uint a_offset, const uint b_offset, const uint itid, sum[2] = fma(FLOAT_TYPE(by64[l]), q2[l], sum[2]); sum[3] = fma(FLOAT_TYPE(by96[l]), q3[l], sum[3]); } - temp[j][n] = fma(fma(sum[0], sccache[ix][s_offset], fma(sum[1], sccache[ix][s_offset + 2], fma(sum[2], sccache[ix][s_offset + 4], sum[3] * sccache[ix][s_offset + 6]))), d, temp[j][n]); + temp[j][n] = fma(fma(sum[0], sccache[csel][ix][s_offset], fma(sum[1], sccache[csel][ix][s_offset + 2], fma(sum[2], sccache[csel][ix][s_offset + 4], sum[3] * sccache[csel][ix][s_offset + 6]))), d, temp[j][n]); } } } diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp index 33b2234e7..23ce8ceec 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp @@ -6,6 +6,9 @@ #ifdef FLOAT16 #extension GL_EXT_shader_explicit_arithmetic_types_float16 : require #endif +#if defined(DATA_A_IQ1_M) +#extension GL_EXT_shader_explicit_arithmetic_types_float16 : require +#endif #ifdef COOPMAT #extension GL_KHR_cooperative_matrix : enable @@ -29,6 +32,13 @@ layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in; layout (binding = 0) readonly buffer A {A_TYPE data_a[];}; +#if defined(A_TYPE_PACKED16) +layout (binding = 0) readonly buffer A_PACKED16 {A_TYPE_PACKED16 data_a_packed16[];}; +#endif +#if defined(A_TYPE_PACKED32) +layout (binding = 0) readonly buffer A_PACKED32 {A_TYPE_PACKED32 data_a_packed32[];}; +#endif + layout (binding = 1) readonly buffer B {B_TYPE data_b[];}; layout (binding = 2) writeonly buffer D {D_TYPE data_d[];}; @@ -95,7 +105,7 @@ shared ACC_TYPE coopmat_stage[TM * TN * NUM_WARPS]; #endif void main() { -#if defined(DATA_A_IQ2_XXS) || defined(DATA_A_IQ2_XS) || defined(DATA_A_IQ2_S) || defined(DATA_A_IQ3_XXS) || defined(DATA_A_IQ3_S) || defined(DATA_A_IQ4_XS) || defined(DATA_A_IQ4_NL) +#ifdef NEEDS_INIT_IQ_SHMEM init_iq_shmem(gl_WorkGroupSize); #endif @@ -202,7 +212,7 @@ void main() { #else ACC_TYPE sums[WMITER * TM * WNITER * TN]; FLOAT_TYPE cache_a[WMITER * TM]; - FLOAT_TYPE cache_b[WNITER * TN]; + FLOAT_TYPE cache_b[TN]; [[unroll]] for (uint i = 0; i < WMITER*TM*WNITER*TN; i++) { sums[i] = ACC_TYPE(0.0f); @@ -240,74 +250,100 @@ void main() { #endif #elif defined(DATA_A_Q4_0) const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a; - const uint buf_idx = (loadc_a + l) * SHMEM_STRIDE + loadr_a; + const uint buf_idx = (loadc_a + l) * SHMEM_STRIDE + 4 * loadr_a; - const uint ib = idx / 16; - const uint iqs = idx & 0xF; + const uint ib = idx / 4; + const uint iqs = idx & 0x03; - const float d = float(data_a[ib].d); - const uint vui = uint(data_a[ib].qs[iqs]); - const vec2 v = (vec2(vui & 0xF, vui >> 4) - 8.0f) * d; + const float d = float(data_a_packed16[ib].d); + const uint vui = uint(data_a_packed16[ib].qs[2*iqs]) | (uint(data_a_packed16[ib].qs[2*iqs + 1]) << 16); + const vec4 v0 = (vec4(unpack8(vui & 0x0F0F0F0F)) - 8.0f) * d; + const vec4 v1 = (vec4(unpack8((vui >> 4) & 0x0F0F0F0F)) - 8.0f) * d; - buf_a[buf_idx ] = FLOAT_TYPE(v.x); - buf_a[buf_idx + 16] = FLOAT_TYPE(v.y); + buf_a[buf_idx ] = FLOAT_TYPE(v0.x); + buf_a[buf_idx + 1 ] = FLOAT_TYPE(v0.y); + buf_a[buf_idx + 2 ] = FLOAT_TYPE(v0.z); + buf_a[buf_idx + 3 ] = FLOAT_TYPE(v0.w); + buf_a[buf_idx + 16] = FLOAT_TYPE(v1.x); + buf_a[buf_idx + 17] = FLOAT_TYPE(v1.y); + buf_a[buf_idx + 18] = FLOAT_TYPE(v1.z); + buf_a[buf_idx + 19] = FLOAT_TYPE(v1.w); #elif defined(DATA_A_Q4_1) const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a; - const uint buf_idx = (loadc_a + l) * SHMEM_STRIDE + loadr_a; + const uint buf_idx = (loadc_a + l) * SHMEM_STRIDE + 4 * loadr_a; - const uint ib = idx / 16; - const uint iqs = idx & 0xF; + const uint ib = idx / 4; + const uint iqs = idx & 0x03; - const float d = float(data_a[ib].d); - const float m = float(data_a[ib].m); - const uint vui = uint(data_a[ib].qs[iqs]); - const vec2 v = vec2(vui & 0xF, vui >> 4) * d + m; + const float d = float(data_a_packed16[ib].d); + const float m = float(data_a_packed16[ib].m); + const uint vui = uint(data_a_packed16[ib].qs[2*iqs]) | (uint(data_a_packed16[ib].qs[2*iqs + 1]) << 16); + const vec4 v0 = vec4(unpack8(vui & 0x0F0F0F0F)) * d + m; + const vec4 v1 = vec4(unpack8((vui >> 4) & 0x0F0F0F0F)) * d + m; - buf_a[buf_idx ] = FLOAT_TYPE(v.x); - buf_a[buf_idx + 16] = FLOAT_TYPE(v.y); + buf_a[buf_idx ] = FLOAT_TYPE(v0.x); + buf_a[buf_idx + 1 ] = FLOAT_TYPE(v0.y); + buf_a[buf_idx + 2 ] = FLOAT_TYPE(v0.z); + buf_a[buf_idx + 3 ] = FLOAT_TYPE(v0.w); + buf_a[buf_idx + 16] = FLOAT_TYPE(v1.x); + buf_a[buf_idx + 17] = FLOAT_TYPE(v1.y); + buf_a[buf_idx + 18] = FLOAT_TYPE(v1.z); + buf_a[buf_idx + 19] = FLOAT_TYPE(v1.w); #elif defined(DATA_A_Q5_0) const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a; - const uint buf_idx = (loadc_a + l) * SHMEM_STRIDE + loadr_a; + const uint buf_idx = (loadc_a + l) * SHMEM_STRIDE + 2 * loadr_a; - const uint ib = idx / 16; - const uint iqs = idx & 0xF; + const uint ib = idx / 8; + const uint iqs = idx & 0x07; - const float d = float(data_a[ib].d); - const uint uint_qh = uint(data_a[ib].qh[1]) << 16 | data_a[ib].qh[0]; - const ivec2 qh = ivec2(((uint_qh >> iqs) << 4) & 0x10, (uint_qh >> (iqs + 12)) & 0x10); - const uint vui = uint(data_a[ib].qs[iqs]); - const vec2 v = (vec2((vui & 0xF) | qh.x, (vui >> 4) | qh.y) - 16.0f) * d; + const float d = float(data_a_packed16[ib].d); + const uint uint_qh = uint(data_a_packed16[ib].qh[1]) << 16 | uint(data_a_packed16[ib].qh[0]); + const ivec2 qh0 = ivec2(((uint_qh >> 2*iqs) << 4) & 0x10, (uint_qh >> (2*iqs + 12)) & 0x10); + const ivec2 qh1 = ivec2(((uint_qh >> (2*iqs + 1)) << 4) & 0x10, (uint_qh >> (2*iqs + 13)) & 0x10); + + const uint vui = uint(data_a_packed16[ib].qs[iqs]); + const vec4 v = (vec4((vui & 0xF) | qh0.x, ((vui >> 4) & 0xF) | qh0.y, ((vui >> 8) & 0xF) | qh1.x, (vui >> 12) | qh1.y) - 16.0f) * d; buf_a[buf_idx ] = FLOAT_TYPE(v.x); + buf_a[buf_idx + 1 ] = FLOAT_TYPE(v.z); buf_a[buf_idx + 16] = FLOAT_TYPE(v.y); + buf_a[buf_idx + 17] = FLOAT_TYPE(v.w); #elif defined(DATA_A_Q5_1) const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a; - const uint buf_idx = (loadc_a + l) * SHMEM_STRIDE + loadr_a; + const uint buf_idx = (loadc_a + l) * SHMEM_STRIDE + 2 * loadr_a; - const uint ib = idx / 16; - const uint iqs = idx & 0xF; + const uint ib = idx / 8; + const uint iqs = idx & 0x07; - const float d = float(data_a[ib].d); - const float m = float(data_a[ib].m); - const uint uint_qh = data_a[ib].qh; - const ivec2 qh = ivec2(((uint_qh >> iqs) << 4) & 0x10, (uint_qh >> (iqs + 12)) & 0x10); - const uint vui = uint(data_a[ib].qs[iqs]); - const vec2 v = vec2((vui & 0xF) | qh.x, (vui >> 4) | qh.y) * d + m; + const float d = float(data_a_packed16[ib].d); + const float m = float(data_a_packed16[ib].m); + const uint uint_qh = data_a_packed16[ib].qh; + const ivec2 qh0 = ivec2(((uint_qh >> 2*iqs) << 4) & 0x10, (uint_qh >> (2*iqs + 12)) & 0x10); + const ivec2 qh1 = ivec2(((uint_qh >> (2*iqs + 1)) << 4) & 0x10, (uint_qh >> (2*iqs + 13)) & 0x10); + + const uint vui = uint(data_a_packed16[ib].qs[iqs]); + const vec4 v = vec4((vui & 0xF) | qh0.x, ((vui >> 4) & 0xF) | qh0.y, ((vui >> 8) & 0xF) | qh1.x, (vui >> 12) | qh1.y) * d + m; buf_a[buf_idx ] = FLOAT_TYPE(v.x); + buf_a[buf_idx + 1 ] = FLOAT_TYPE(v.z); buf_a[buf_idx + 16] = FLOAT_TYPE(v.y); + buf_a[buf_idx + 17] = FLOAT_TYPE(v.w); #elif defined(DATA_A_Q8_0) const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a; const uint buf_idx = (loadc_a + l) * SHMEM_STRIDE + loadr_a * LOAD_VEC_A; - const uint ib = idx / 16; - const uint iqs = (idx & 0xF) * 2; + const uint ib = idx / 8; + const uint iqs = idx & 0x07; - const float d = float(data_a[ib].d); - const vec2 v = vec2(int(data_a[ib].qs[iqs]), int(data_a[ib].qs[iqs + 1])) * d; + const float d = float(data_a_packed16[ib].d); + const i8vec2 v0 = unpack8(int32_t(data_a_packed16[ib].qs[2*iqs])).xy; // vec4 used due to #12147 + const i8vec2 v1 = unpack8(int32_t(data_a_packed16[ib].qs[2*iqs + 1])).xy; + const vec4 v = vec4(v0.x, v0.y, v1.x, v1.y) * d; buf_a[buf_idx ] = FLOAT_TYPE(v.x); buf_a[buf_idx + 1] = FLOAT_TYPE(v.y); + buf_a[buf_idx + 2] = FLOAT_TYPE(v.z); + buf_a[buf_idx + 3] = FLOAT_TYPE(v.w); #elif defined(DATA_A_Q2_K) const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a; const uint buf_idx = (loadc_a + l) * SHMEM_STRIDE + loadr_a * LOAD_VEC_A; @@ -437,6 +473,56 @@ void main() { buf_a[buf_idx ] = FLOAT_TYPE(dscale * float(int8_t(((data_a[ib].ql[qsi ] >> (b * 4)) & 0xF) | (((data_a[ib].qh[qhi ] >> qhshift) & 3) << 4)) - 32)); buf_a[buf_idx + 1] = FLOAT_TYPE(dscale * float(int8_t(((data_a[ib].ql[qsi + 1] >> (b * 4)) & 0xF) | (((data_a[ib].qh[qhi + 1] >> qhshift) & 3) << 4)) - 32)); +#elif defined(DATA_A_IQ1_S) + const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a; + const uint buf_idx = (loadc_a + l) * SHMEM_STRIDE + loadr_a * LOAD_VEC_A; + + const uint ib = idx / 128; // 2 values per idx + const uint ib32 = (idx % 128) / 16; // 0..7 + const uint ib8 = (idx % 128) / 4; + const int i8 = 2 * int(idx % 4); + + const float d = float(data_a[ib].d); + const uint qh = data_a[ib].qh[ib32]; + const uint qs = data_a[ib].qs[ib8]; + const float dl = d * (2 * bitfieldExtract(qh, 12, 3) + 1); + const float delta = ((qh & 0x8000) != 0) ? -IQ1S_DELTA : IQ1S_DELTA; + const int16_t grid = int16_t(iq1s_grid[qs | (bitfieldExtract(qh, 3 * int(ib8 & 3), 3) << 8)]); + + const ivec2 gvec = ivec2( + bitfieldExtract(grid, 2 * (i8), 2), + bitfieldExtract(grid, 2 * (i8 + 1), 2) + ); + const vec2 v = dl * (vec2(gvec) + delta); + + buf_a[buf_idx ] = FLOAT_TYPE(v.x); + buf_a[buf_idx + 1] = FLOAT_TYPE(v.y); +#elif defined(DATA_A_IQ1_M) + const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a; + const uint buf_idx = (loadc_a + l) * SHMEM_STRIDE + loadr_a * LOAD_VEC_A; + + const uint ib = idx / 128; // 2 values per idx + const uint ib8 = (idx % 128) / 4; + const uint ib16 = ib8 / 2; + const int i8 = 2 * int(idx % 4); + + const uint16_t[4] scales = data_a[ib].scales; + const u16vec4 s = u16vec4(scales[0], scales[1], scales[2], scales[3]) >> 12; + const float d = float(unpackHalf2x16(s.x | (s.y << 4) | (s.z << 8) | (s.w << 12)).x); + const uint sc = scales[ib8 / 8]; + const uint qs = data_a[ib].qs[ib8]; + const uint qh = data_a[ib].qh[ib16] >> (4 * (ib8 & 1)); + const float dl = d * (2 * bitfieldExtract(sc, 3 * int(ib16 & 3), 3) + 1); + const float delta = ((qh & 8) != 0) ? -IQ1M_DELTA : IQ1M_DELTA; + const int16_t grid = int16_t(iq1s_grid[qs | ((qh & 7) << 8)]); + const ivec2 gvec = ivec2( + bitfieldExtract(grid, 2 * (i8), 2), + bitfieldExtract(grid, 2 * (i8 + 1), 2) + ); + const vec2 v = dl * (vec2(gvec) + delta); + + buf_a[buf_idx ] = FLOAT_TYPE(v.x); + buf_a[buf_idx + 1] = FLOAT_TYPE(v.y); #elif defined(DATA_A_IQ2_XXS) const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a; const uint buf_idx = (loadc_a + l) * SHMEM_STRIDE + loadr_a * LOAD_VEC_A; @@ -458,7 +544,7 @@ void main() { const uint sign = (sign7 | (bitCount(sign7) << 7)) >> (2 * (idx % 4)); const i8vec2 sign01 = i8vec2(1 - (2 & i8vec2(int8_t(sign << 1), int8_t(sign)))); const uint grid = iq2xxs_grid[qs][(idx % 4) / 2] >> (16 * (idx & 1)); - const vec2 v = db * vec2(sign01) * vec2(unpack8(grid).xy); + const vec2 v = db * vec2(sign01) * vec2(unpack8(grid).xy); // vec4 used due to #12147 buf_a[buf_idx ] = FLOAT_TYPE(v.x); buf_a[buf_idx + 1] = FLOAT_TYPE(v.y); @@ -478,7 +564,7 @@ void main() { const uint sign = (sign7 | (bitCount(sign7) << 7)) >> (2 * (idx % 4)); const i8vec2 sign01 = i8vec2(1 - (2 & i8vec2(int8_t(sign << 1), int8_t(sign)))); const uint grid = iq2xs_grid[qs & 511][(idx % 4) / 2] >> (16 * (idx & 1)); - const vec2 v = db * vec2(sign01) * vec2(unpack8(grid).xy); + const vec2 v = db * vec2(sign01) * vec2(unpack8(grid).xy); // vec4 used due to #12147 buf_a[buf_idx ] = FLOAT_TYPE(v.x); buf_a[buf_idx + 1] = FLOAT_TYPE(v.y); @@ -500,7 +586,7 @@ void main() { const float db = d * 0.25 * (0.5 + scale); const i8vec2 sign01 = i8vec2(1 - (2 & i8vec2(int8_t(sign << 1), int8_t(sign)))); const uint16_t grid = unpack16(iq2s_grid[qs | ((qh << (8 - qhshift)) & 0x300)][(idx & 2) >> 1])[idx & 1]; - const vec2 v = db * vec2(sign01) * vec2(unpack8(grid)); + const vec2 v = db * vec2(sign01) * vec2(unpack8(uint32_t(grid)).xy); // vec4 used due to #12147 buf_a[buf_idx ] = FLOAT_TYPE(v.x); buf_a[buf_idx + 1] = FLOAT_TYPE(v.y); @@ -525,7 +611,7 @@ void main() { const uint sign = (sign7 | (bitCount(sign7) << 7)) >> (2 * (idx % 4)); const i8vec2 sign01 = i8vec2(1 - (2 & i8vec2(int8_t(sign << 1), int8_t(sign)))); const uint grid = iq3xxs_grid[qs] >> (16 * (idx & 1)); - const vec2 v = db * vec2(sign01) * vec2(unpack8(grid).xy); + const vec2 v = db * vec2(sign01) * vec2(unpack8(grid).xy); // vec4 used due to #12147 buf_a[buf_idx ] = FLOAT_TYPE(v.x); buf_a[buf_idx + 1] = FLOAT_TYPE(v.y); @@ -545,7 +631,7 @@ void main() { const i8vec2 sign01 = i8vec2(1 - (2 & i8vec2(sign << 1, sign))); const float db = d * (1 + 2 * ((scale >> (4 * (iqh & 1))) & 0xf)); const uint32_t grid = iq3s_grid[qs | ((qh << (8 - (iqs % 8))) & 256)] >> (16 * (idx % 2)); - const vec2 v = db * vec2(sign01) * vec2(unpack8(grid).xy); + const vec2 v = db * vec2(sign01) * vec2(unpack8(grid).xy); // vec4 used due to #12147 buf_a[buf_idx ] = FLOAT_TYPE(v.x); buf_a[buf_idx + 1] = FLOAT_TYPE(v.y); @@ -570,17 +656,18 @@ void main() { buf_a[buf_idx + 1] = FLOAT_TYPE(v.y); #elif defined(DATA_A_IQ4_NL) const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a; - const uint buf_idx = (loadc_a + l) * SHMEM_STRIDE + loadr_a; + const uint buf_idx = (loadc_a + l) * SHMEM_STRIDE + 2 * loadr_a; - const uint ib = idx / 16; - const uint iqs = idx & 0xF; + const uint ib = idx / 8; + const uint iqs = idx & 0x07; - const float d = float(data_a[ib].d); - const uint vui = uint(data_a[ib].qs[iqs]); - const vec2 v = vec2(kvalues_iq4nl[vui & 0xF], kvalues_iq4nl[vui >> 4]) * d; + const FLOAT_TYPE d = FLOAT_TYPE(data_a_packed16[ib].d); + const uint vui = uint(data_a_packed16[ib].qs[iqs]); - buf_a[buf_idx ] = FLOAT_TYPE(v.x); - buf_a[buf_idx + 16] = FLOAT_TYPE(v.y); + buf_a[buf_idx ] = FLOAT_TYPE(kvalues_iq4nl[vui & 0xF]) * d; + buf_a[buf_idx + 1 ] = FLOAT_TYPE(kvalues_iq4nl[bitfieldExtract(vui, 8, 4)]) * d; + buf_a[buf_idx + 16] = FLOAT_TYPE(kvalues_iq4nl[bitfieldExtract(vui, 4, 4)]) * d; + buf_a[buf_idx + 17] = FLOAT_TYPE(kvalues_iq4nl[vui >> 12]) * d; #endif } [[unroll]] for (uint l = 0; l < BN; l += loadstride_b) { @@ -657,16 +744,14 @@ void main() { } [[unroll]] for (uint wsic = 0; wsic < WNITER; wsic++) { [[unroll]] for (uint j = 0; j < TN; j++) { - cache_b[wsic * TN + j] = buf_b[(warp_c * WN + wsic * WSUBN + tiwc * TN + j) * SHMEM_STRIDE + i]; + cache_b[j] = buf_b[(warp_c * WN + wsic * WSUBN + tiwc * TN + j) * SHMEM_STRIDE + i]; } - } - [[unroll]] for (uint wsic = 0; wsic < WNITER; wsic++) { [[unroll]] for (uint wsir = 0; wsir < WMITER; wsir++) { [[unroll]] for (uint cc = 0; cc < TN; cc++) { [[unroll]] for (uint cr = 0; cr < TM; cr++) { const uint sums_idx = (wsic * TN + cc) * (WMITER * TM) + wsir * TM + cr; - sums[sums_idx] = fma(ACC_TYPE(cache_a[wsir * TM + cr]), ACC_TYPE(cache_b[wsic * TN + cc]), sums[sums_idx]); + sums[sums_idx] = fma(ACC_TYPE(cache_a[wsir * TM + cr]), ACC_TYPE(cache_b[cc]), sums[sums_idx]); } } } @@ -690,7 +775,7 @@ void main() { [[unroll]] for (uint cm_col = 0; cm_col < cms_per_col; cm_col++) { coopMatStore(sums[cm_col * cms_per_row + cm_row], coopmat_stage, warp_i * TM * TN, TM, gl_CooperativeMatrixLayoutColumnMajor); - [[unroll]] for (uint col = 0; col < BN; col += storestride) { + [[unroll]] for (uint col = 0; col < TN; col += storestride) { const uint row_i = dc + cm_col * TN + col + store_c; if (row_i >= _ne1) break; diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_cm2.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_cm2.comp index 7e29bbfec..06b7ab09e 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_cm2.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_cm2.comp @@ -19,10 +19,17 @@ layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in; +#define IS_MUL_MM2 1 + +layout (constant_id = 0) const uint BLOCK_SIZE = 256; layout (constant_id = 1) const uint BM = 64; layout (constant_id = 2) const uint BN = 64; layout (constant_id = 3) const uint BK = 16; // Assumed to be 32 if working with a quant +layout (constant_id = 4) const bool enable_smaller_matrices = false; +const uint BNover2 = enable_smaller_matrices ? (BN / 2) : BN; +const uint BNover4 = enable_smaller_matrices ? (BN / 4) : BN; + layout (push_constant) uniform parameter { uint M; @@ -48,6 +55,8 @@ layout (push_constant) uniform parameter uint broadcast2; uint broadcast3; #endif + // N dimension for the B matrix can be >= p.N + uint padded_N; } p; @@ -64,6 +73,13 @@ layout (binding = 2) writeonly buffer D {D_TYPE data_d[];}; #define DECODEFUNCA #endif +#if !defined(fetch_scales) +#define fetch_scales(a, b, c, d, e, f) +#endif +#if !defined(store_scales) +#define store_scales(a) +#endif + #ifdef MUL_MAT_ID layout (binding = 3) readonly buffer IDS {int data_ids[];}; @@ -106,10 +122,12 @@ D_TYPE perElemOpD(const in uint32_t r, const in uint32_t c, const in D_TYPE elem #endif void main() { -#if defined(DATA_A_IQ2_XXS) || defined(DATA_A_IQ2_XS) || defined(DATA_A_IQ2_S) || defined(DATA_A_IQ3_XXS) || defined(DATA_A_IQ3_S) || defined(DATA_A_IQ4_XS) || defined(DATA_A_IQ4_NL) +#ifdef NEEDS_INIT_IQ_SHMEM init_iq_shmem(gl_WorkGroupSize); #endif + const uint tid = gl_LocalInvocationIndex; + #ifdef MUL_MAT_ID const uint expert_idx = gl_GlobalInvocationID.z; #else @@ -166,15 +184,13 @@ void main() { const uint end_k = min(p.K, (ik + 1) * p.k_split); #endif - coopmat sum; - sum = coopmat(0.0); - #ifdef MUL_MAT_ID uint pos_a = (expert_idx * p.batch_stride_a) / QUANT_K; uint pos_b = 0; #else uint pos_a = (batch_idx_a * p.batch_stride_a) / QUANT_K; uint pos_b = batch_idx * p.batch_stride_b; + uint pos_d = batch_idx * p.batch_stride_d + ik * p.batch_stride_d * gl_NumWorkGroups.z; #endif uint stride_a = p.stride_a / QUANT_K; @@ -195,6 +211,7 @@ void main() { tensorLayoutNV<2> tensorLayoutB = createTensorLayoutNV(2); tensorLayoutNV<2, gl_CooperativeMatrixClampModeConstantNV> tensorLayoutBClamp = createTensorLayoutNV(2, gl_CooperativeMatrixClampModeConstantNV); tensorLayoutNV<2, gl_CooperativeMatrixClampModeConstantNV> tensorLayoutD = createTensorLayoutNV(2, gl_CooperativeMatrixClampModeConstantNV); + tensorLayoutD = setTensorLayoutStrideNV(tensorLayoutD, p.stride_d, 1); #if QUANT_K > 1 tensorLayoutA = setTensorLayoutBlockSizeNV(tensorLayoutA, 1, QUANT_K); @@ -202,24 +219,32 @@ void main() { #endif // Use end_k rather than p.K as the dimension because that's what - // we need to bound check against when using split_k + // we need to bound check against when using split_k. + // Bounds check B against padded_N, but bounds check D against N. tensorLayoutA = setTensorLayoutDimensionNV(tensorLayoutA, p.M, end_k); - tensorLayoutB = setTensorLayoutDimensionNV(tensorLayoutB, p.N, end_k); + tensorLayoutB = setTensorLayoutDimensionNV(tensorLayoutB, p.padded_N, end_k); tensorLayoutD = setTensorLayoutDimensionNV(tensorLayoutD, p.N, p.M); tensorLayoutAClamp = setTensorLayoutDimensionNV(tensorLayoutAClamp, p.M, end_k); - tensorLayoutBClamp = setTensorLayoutDimensionNV(tensorLayoutBClamp, p.N, end_k); + tensorLayoutBClamp = setTensorLayoutDimensionNV(tensorLayoutBClamp, p.padded_N, end_k); tensorViewNV<2, false, 1, 0> tensorViewTranspose = createTensorViewNV(2, false, 1, 0); #if !defined(MUL_MAT_ID) + + const uint START_ALIGN_K = 256; + // For Qi_K (block size 256), unroll whole 256 element tiles. + // For legacy quants (block size 32), unroll 8x. + const uint UNROLL_K = (QUANT_K == 256) ? 256 : (BK * 8); + const uint unroll_count = UNROLL_K / BK; + // Detect a fast path where all loads are entirely in bounds and no clamping is required - if ((ir + 1) * BM <= p.M && (ic + 1) * BN <= p.N && (start_k % BK) == 0 && (end_k % BK) == 0 && + if ((ir + 1) * BM <= p.M && (ic + 1) * BN <= p.padded_N && (start_k % START_ALIGN_K) == 0 && (end_k % BK) == 0 && #if QUANT_K == 1 (stride_a % 8) == 0 && #endif - (stride_b % 8) == 0 && (start_k % 8) == 0) { + (stride_b % 8) == 0) { // Hint to the compiler that values are aligned (want 16B alignment) - start_k &= ~7; + start_k &= ~(START_ALIGN_K-1); stride_b &= ~7; #if QUANT_K == 1 stride_a &= ~7; @@ -228,17 +253,131 @@ void main() { tensorLayoutA = setTensorLayoutStrideNV(tensorLayoutA, stride_a, 1); tensorLayoutB = setTensorLayoutStrideNV(tensorLayoutB, stride_b, 1); - uint k_iters = (end_k - start_k + BK - 1) / BK; + uint k_iters = (end_k - start_k) / UNROLL_K; + uint block_k = start_k; - for (uint block_k = start_k, i = 0; i < k_iters; block_k += BK, ++i) { + // fetch scale values for a tile of quants. These will be copied into shared memory. + // The fetches and stores are pipelined to hide the latency. + fetch_scales(ir * BM, pos_a, stride_a, start_k, tid, true); - coopmat mat_a; - coopmat mat_b; + if (enable_smaller_matrices && ic * BN + BNover4 >= p.N) { + coopmat sum = coopmat(0.0); + for (uint i = 0; i < k_iters; ++i) { - coopMatLoadTensorNV(mat_a, data_a, pos_a, sliceTensorLayoutNV(tensorLayoutA, ir * BM, BM, block_k, BK) DECODEFUNCA); - coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, ic * BN, BN, block_k, BK), tensorViewTranspose); + store_scales(tid); + if (block_k + UNROLL_K < end_k) { + fetch_scales(ir * BM, pos_a, stride_a, block_k + UNROLL_K, tid, true); + } - sum = coopMatMulAdd(mat_a, mat_b, sum); + // Manually partial unroll + [[unroll]] for (uint j = 0; j < unroll_count; ++j) { + coopmat mat_a; + coopmat mat_b; + + coopMatLoadTensorNV(mat_a, data_a, pos_a, sliceTensorLayoutNV(tensorLayoutA, ir * BM, BM, block_k, BK) DECODEFUNCA); + coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, ic * BN, BNover4, block_k, BK), tensorViewTranspose); + + sum = coopMatMulAdd(mat_a, mat_b, sum); + block_k += BK; + } + } + // Do any remaining iterations that were not unrolled + if (block_k < end_k) { + store_scales(tid); + } + while (block_k < end_k) { + coopmat mat_a; + coopmat mat_b; + + coopMatLoadTensorNV(mat_a, data_a, pos_a, sliceTensorLayoutNV(tensorLayoutA, ir * BM, BM, block_k, BK) DECODEFUNCA); + coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, ic * BN, BNover4, block_k, BK), tensorViewTranspose); + + sum = coopMatMulAdd(mat_a, mat_b, sum); + block_k += BK; + } + coopmat mat_d = coopmat(sum); + + coopMatStoreTensorNV(mat_d, data_d, pos_d, sliceTensorLayoutNV(tensorLayoutD, ic * BN, BNover4, ir * BM, BM), tensorViewTranspose); + return; + } else if (enable_smaller_matrices && ic * BN + BNover2 >= p.N) { + coopmat sum = coopmat(0.0); + for (uint i = 0; i < k_iters; ++i) { + + store_scales(tid); + if (block_k + UNROLL_K < end_k) { + fetch_scales(ir * BM, pos_a, stride_a, block_k + UNROLL_K, tid, true); + } + + // Manually partial unroll + [[unroll]] for (uint j = 0; j < unroll_count; ++j) { + coopmat mat_a; + coopmat mat_b; + + coopMatLoadTensorNV(mat_a, data_a, pos_a, sliceTensorLayoutNV(tensorLayoutA, ir * BM, BM, block_k, BK) DECODEFUNCA); + coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, ic * BN, BNover2, block_k, BK), tensorViewTranspose); + + sum = coopMatMulAdd(mat_a, mat_b, sum); + block_k += BK; + } + } + // Do any remaining iterations that were not unrolled + if (block_k < end_k) { + store_scales(tid); + } + while (block_k < end_k) { + coopmat mat_a; + coopmat mat_b; + + coopMatLoadTensorNV(mat_a, data_a, pos_a, sliceTensorLayoutNV(tensorLayoutA, ir * BM, BM, block_k, BK) DECODEFUNCA); + coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, ic * BN, BNover2, block_k, BK), tensorViewTranspose); + + sum = coopMatMulAdd(mat_a, mat_b, sum); + block_k += BK; + } + coopmat mat_d = coopmat(sum); + + coopMatStoreTensorNV(mat_d, data_d, pos_d, sliceTensorLayoutNV(tensorLayoutD, ic * BN, BNover2, ir * BM, BM), tensorViewTranspose); + return; + } else { + coopmat sum = coopmat(0.0); + + for (uint i = 0; i < k_iters; ++i) { + + store_scales(tid); + if (block_k + UNROLL_K < end_k) { + fetch_scales(ir * BM, pos_a, stride_a, block_k + UNROLL_K, tid, true); + } + + // Manually partial unroll + [[unroll]] for (uint j = 0; j < unroll_count; ++j) { + coopmat mat_a; + coopmat mat_b; + + coopMatLoadTensorNV(mat_a, data_a, pos_a, sliceTensorLayoutNV(tensorLayoutA, ir * BM, BM, block_k, BK) DECODEFUNCA); + coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, ic * BN, BN, block_k, BK), tensorViewTranspose); + + sum = coopMatMulAdd(mat_a, mat_b, sum); + block_k += BK; + } + } + // Do any remaining iterations that were not unrolled + if (block_k < end_k) { + store_scales(tid); + } + while (block_k < end_k) { + coopmat mat_a; + coopmat mat_b; + + coopMatLoadTensorNV(mat_a, data_a, pos_a, sliceTensorLayoutNV(tensorLayoutA, ir * BM, BM, block_k, BK) DECODEFUNCA); + coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, ic * BN, BN, block_k, BK), tensorViewTranspose); + + sum = coopMatMulAdd(mat_a, mat_b, sum); + block_k += BK; + } + coopmat mat_d = coopmat(sum); + + coopMatStoreTensorNV(mat_d, data_d, pos_d, sliceTensorLayoutNV(tensorLayoutD, ic * BN, BN, ir * BM, BM), tensorViewTranspose); + return; } } else #endif // !defined(MUL_MAT_ID) @@ -251,61 +390,43 @@ void main() { tensorLayoutBClamp = setTensorLayoutStrideNV(tensorLayoutBClamp, stride_b, 1); + coopmat sum; + sum = coopmat(0.0); + + uint k_iters = (end_k - start_k + BK - 1) / BK; + + fetch_scales(ir * BM, pos_a, stride_a, start_k, tid, false); + [[dont_unroll]] - for (uint block_k = start_k; block_k < end_k; block_k += BK) { + for (uint block_k = start_k, i = 0; i < k_iters; block_k += BK, ++i) { + + store_scales(tid); + if (block_k + BK < end_k) { + fetch_scales(ir * BM, pos_a, stride_a, block_k + BK, tid, false); + } coopmat mat_a; coopmat mat_b; - // Clamping is expensive, so detect different code paths for each combination - // of A and B needing clamping. - bool unclampedA = (ir + 1) * BM <= p.M && block_k + BK <= end_k && (block_k % 8) == 0; + coopMatLoadTensorNV(mat_a, data_a, pos_a, sliceTensorLayoutNV(tensorLayoutAClamp, ir * BM, BM, block_k, BK) DECODEFUNCA); #ifdef MUL_MAT_ID - bool unclampedB = true; + coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, ic * BN, BN, block_k, BK), tensorViewTranspose, decodeFuncB); #else - bool unclampedB = (ic + 1) * BN <= p.N && block_k + BK <= end_k && (block_k % 8) == 0; + coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutBClamp, ic * BN, BN, block_k, BK), tensorViewTranspose); #endif - if (unclampedA && unclampedB) { - coopMatLoadTensorNV(mat_a, data_a, pos_a, sliceTensorLayoutNV(tensorLayoutA, ir * BM, BM, (block_k & ~7), BK) DECODEFUNCA); -#ifdef MUL_MAT_ID - coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, ic * BN, BN, block_k, BK), tensorViewTranspose, decodeFuncB); -#else - coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, ic * BN, BN, (block_k & ~7), BK), tensorViewTranspose); -#endif - sum = coopMatMulAdd(mat_a, mat_b, sum); - } else if (unclampedA && !unclampedB) { - coopMatLoadTensorNV(mat_a, data_a, pos_a, sliceTensorLayoutNV(tensorLayoutA, ir * BM, BM, (block_k & ~7), BK) DECODEFUNCA); - coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutBClamp, ic * BN, BN, block_k, BK), tensorViewTranspose); - sum = coopMatMulAdd(mat_a, mat_b, sum); - } else if (!unclampedA && unclampedB) { - coopMatLoadTensorNV(mat_a, data_a, pos_a, sliceTensorLayoutNV(tensorLayoutAClamp, ir * BM, BM, block_k, BK) DECODEFUNCA); -#ifdef MUL_MAT_ID - coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, ic * BN, BN, block_k, BK), tensorViewTranspose, decodeFuncB); -#else - coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, ic * BN, BN, (block_k & ~7), BK), tensorViewTranspose); -#endif - sum = coopMatMulAdd(mat_a, mat_b, sum); - } else if (!unclampedA && !unclampedB) { - coopMatLoadTensorNV(mat_a, data_a, pos_a, sliceTensorLayoutNV(tensorLayoutAClamp, ir * BM, BM, block_k, BK) DECODEFUNCA); - coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutBClamp, ic * BN, BN, block_k, BK), tensorViewTranspose); - - sum = coopMatMulAdd(mat_a, mat_b, sum); - } + sum = coopMatMulAdd(mat_a, mat_b, sum); } - } - // Convert from ACC_TYPE to D_TYPE - coopmat mat_d; - mat_d = coopmat(sum); + // Convert from ACC_TYPE to D_TYPE + coopmat mat_d; + mat_d = coopmat(sum); #ifdef MUL_MAT_ID - // Call callback to store each element, remapping row through shared memory - coopMatPerElementNV(mat_d, mat_d, perElemOpD, ir, ic); + // Call callback to store each element, remapping row through shared memory + coopMatPerElementNV(mat_d, mat_d, perElemOpD, ir, ic); #else - tensorLayoutD = setTensorLayoutStrideNV(tensorLayoutD, p.stride_d, 1); - - uint pos_d = batch_idx * p.batch_stride_d + ik * p.batch_stride_d * gl_NumWorkGroups.z; - coopMatStoreTensorNV(mat_d, data_d, pos_d, sliceTensorLayoutNV(tensorLayoutD, ic * BN, BN, ir * BM, BM), tensorViewTranspose); + coopMatStoreTensorNV(mat_d, data_d, pos_d, sliceTensorLayoutNV(tensorLayoutD, ic * BN, BN, ir * BM, BM), tensorViewTranspose); #endif + } } diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq.comp new file mode 100644 index 000000000..284a35caa --- /dev/null +++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq.comp @@ -0,0 +1,442 @@ +#version 450 + +#extension GL_EXT_control_flow_attributes : enable +#extension GL_EXT_shader_16bit_storage : require +#extension GL_EXT_shader_explicit_arithmetic_types_int8 : require + +#extension GL_EXT_integer_dot_product : require + +#ifdef FLOAT16 +#extension GL_EXT_shader_explicit_arithmetic_types_float16 : require +#endif + +#ifdef COOPMAT +#extension GL_KHR_cooperative_matrix : enable +#extension GL_KHR_memory_scope_semantics : enable +#extension GL_KHR_shader_subgroup_basic : enable +#endif + +#ifdef MUL_MAT_ID +#extension GL_EXT_shader_explicit_arithmetic_types_int16 : require +#endif + +#include "types.comp" + +layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in; + +layout (binding = 0) readonly buffer A {A_TYPE_PACKED16 data_a[];}; +#if defined(A_TYPE_PACKED32) +layout (binding = 0) readonly buffer A_PACKED32 {A_TYPE_PACKED32 data_a_packed32[];}; +#endif +layout (binding = 1) readonly buffer B {block_q8_1_packed32 data_b[];}; +layout (binding = 2) writeonly buffer D {D_TYPE data_d[];}; + +#ifdef MUL_MAT_ID +layout (binding = 3) readonly buffer IDS {int data_ids[];}; +#endif + +layout (push_constant) uniform parameter +{ + uint M; + uint N; + uint K; + uint stride_a; + uint stride_b; + uint stride_d; + + uint batch_stride_a; + uint batch_stride_b; + uint batch_stride_d; + +#ifdef MUL_MAT_ID + uint nei0; + uint nei1; + uint nbi1; + uint ne11; +#else + uint k_split; + uint ne02; + uint ne12; + uint broadcast2; + uint broadcast3; +#endif +} p; + +layout (constant_id = 0) const uint BLOCK_SIZE = 64; +layout (constant_id = 1) const uint BM = 64; +layout (constant_id = 2) const uint BN = 64; +// layout (constant_id = 3) const uint BK = 32; +layout (constant_id = 4) const uint WM = 32; +layout (constant_id = 5) const uint WN = 32; +layout (constant_id = 6) const uint WMITER = 2; +layout (constant_id = 7) const uint TM = 4; +layout (constant_id = 8) const uint TN = 2; +layout (constant_id = 9) const uint TK = 1; // Only needed for coopmat +layout (constant_id = 10) const uint WARP = 32; + +#define BK 32 + +#ifdef COOPMAT +#define SHMEM_STRIDE (BK / 4 + 4) +#else +#define SHMEM_STRIDE (BK / 4 + 1) +#endif + +shared int32_t buf_a_qs[BM * SHMEM_STRIDE]; + +#ifndef COOPMAT +#if QUANT_AUXF == 1 +shared FLOAT_TYPE buf_a_dm[BM]; +#else +shared FLOAT_TYPE_VEC2 buf_a_dm[BM]; +#endif +#endif + +shared int32_t buf_b_qs[BN * SHMEM_STRIDE]; +#ifndef COOPMAT +shared FLOAT_TYPE_VEC2 buf_b_ds[BN]; +#endif + +#define LOAD_VEC_A (4 * QUANT_R) +#define LOAD_VEC_B 4 + +#ifdef MUL_MAT_ID +shared u16vec2 row_ids[3072]; +#endif // MUL_MAT_ID + +#define NUM_WARPS (BLOCK_SIZE / WARP) + +#ifdef COOPMAT +shared ACC_TYPE coopmat_stage[TM * TN * NUM_WARPS]; +#endif + +#include "mul_mmq_funcs.comp" + +void main() { +#ifdef NEEDS_INIT_IQ_SHMEM + init_iq_shmem(gl_WorkGroupSize); +#endif + +#ifdef MUL_MAT_ID + const uint expert_idx = gl_GlobalInvocationID.z; +#else + const uint batch_idx = gl_GlobalInvocationID.z; + + const uint i13 = batch_idx / p.ne12; + const uint i12 = batch_idx % p.ne12; + + const uint i03 = i13 / p.broadcast3; + const uint i02 = i12 / p.broadcast2; + + const uint batch_idx_a = i03 * p.ne02 + i02; +#endif + + const uint blocks_m = (p.M + BM - 1) / BM; + const uint ir = gl_WorkGroupID.x % blocks_m; + const uint ik = gl_WorkGroupID.x / blocks_m; + const uint ic = gl_WorkGroupID.y; + + const uint WNITER = (WM * WN) / (WARP * TM * TN * WMITER); + const uint WSUBM = WM / WMITER; + const uint WSUBN = WN / WNITER; + +#ifdef COOPMAT + const uint warp_i = gl_SubgroupID; + + const uint tiw = gl_SubgroupInvocationID; + + const uint cms_per_row = WM / TM; + const uint cms_per_col = WN / TN; + + const uint storestride = WARP / TM; + const uint store_r = tiw % TM; + const uint store_c = tiw / TM; +#else + const uint warp_i = gl_LocalInvocationID.x / WARP; + + const uint tiw = gl_LocalInvocationID.x % WARP; + + const uint tiwr = tiw % (WSUBM / TM); + const uint tiwc = tiw / (WSUBM / TM); +#endif + + const uint warp_r = warp_i % (BM / WM); + const uint warp_c = warp_i / (BM / WM); + + const uint loadr_a = gl_LocalInvocationID.x % (BK / LOAD_VEC_A); + const uint loadc_a = gl_LocalInvocationID.x / (BK / LOAD_VEC_A); + const uint loadr_b = gl_LocalInvocationID.x % (BK / LOAD_VEC_B); + const uint loadc_b = gl_LocalInvocationID.x / (BK / LOAD_VEC_B); + + const uint loadstride_a = BLOCK_SIZE * LOAD_VEC_A / BK; + const uint loadstride_b = BLOCK_SIZE * LOAD_VEC_B / BK; + +#ifdef MUL_MAT_ID + uint _ne1 = 0; + for (uint ii1 = 0; ii1 < p.nei1; ii1++) { + for (uint ii0 = 0; ii0 < p.nei0; ii0++) { + if (data_ids[ii1*p.nbi1 + ii0] == expert_idx) { + row_ids[_ne1] = u16vec2(ii0, ii1); + _ne1++; + } + } + } + + barrier(); + + // Workgroup has no work + if (ic * BN >= _ne1) return; +#endif + +#ifdef MUL_MAT_ID + const uint start_k = 0; + const uint end_k = p.K; +#else + const uint start_k = ik * p.k_split; + const uint end_k = min(p.K, (ik + 1) * p.k_split); +#endif + + uint pos_a_ib = ( +#ifdef MUL_MAT_ID + expert_idx * p.batch_stride_a + +#else + batch_idx_a * p.batch_stride_a + +#endif + ir * BM * p.stride_a + start_k) / BK; +#ifdef MUL_MAT_ID + uint pos_b_ib = 0; +#else + uint pos_b_ib = (batch_idx * p.batch_stride_b + ic * BN * p.stride_b + start_k) / BK; +#endif + +#ifdef COOPMAT + coopmat cache_a; + coopmat cache_b; + coopmat cm_result; + + coopmat factors[cms_per_row * cms_per_col]; + + coopmat sums[cms_per_row * cms_per_col]; + + [[unroll]] for (uint i = 0; i < cms_per_row * cms_per_col; i++) { + sums[i] = coopmat(0.0f); + } +#else + int32_t cache_a_qs[WMITER * TM * BK / 4]; + + int32_t cache_b_qs[TN * BK / 4]; + + ACC_TYPE sums[WMITER * TM * WNITER * TN]; + + [[unroll]] for (uint i = 0; i < WMITER*TM*WNITER*TN; i++) { + sums[i] = ACC_TYPE(0.0f); + } +#endif + +#if QUANT_AUXF == 1 + FLOAT_TYPE cache_a_dm[WMITER * TM]; +#else + FLOAT_TYPE_VEC2 cache_a_dm[WMITER * TM]; +#endif + + FLOAT_TYPE_VEC2 cache_b_ds[TN]; + + for (uint block = start_k; block < end_k; block += BK) { + [[unroll]] for (uint l = 0; loadc_a + l < BM; l += loadstride_a) { + const uint ib = pos_a_ib + (loadc_a + l) * p.stride_a / BK; + const uint iqs = loadr_a; + const uint buf_ib = loadc_a + l; + + if (iqs == 0) { +#if QUANT_AUXF == 1 + buf_a_dm[buf_ib] = get_d(ib); +#else + buf_a_dm[buf_ib] = get_dm(ib); +#endif + } +#if QUANT_R == 1 + buf_a_qs[buf_ib * SHMEM_STRIDE + iqs] = repack(ib, iqs); +#else + const i32vec2 vals = repack(ib, iqs); + buf_a_qs[buf_ib * SHMEM_STRIDE + iqs ] = vals.x; + buf_a_qs[buf_ib * SHMEM_STRIDE + iqs + 4] = vals.y; +#endif + } + [[unroll]] for (uint l = 0; loadc_b + l < BN; l += loadstride_b) { +#ifdef MUL_MAT_ID + const u16vec2 row_idx = row_ids[ic * BN + loadc_b + l]; + const uint idx = pos_b_ib + row_idx.y * p.batch_stride_b / LOAD_VEC_B + (row_idx.x % p.ne11) * p.stride_b / LOAD_VEC_B + loadr_b; + const uint ib = idx / 8; + const uint iqs = idx & 0x7; +#else + const uint ib = pos_b_ib + (loadc_b + l) * p.stride_b / BK; + const uint iqs = loadr_b; +#endif + + const uint buf_ib = loadc_b + l; + + if (iqs == 0) { + buf_b_ds[buf_ib] = FLOAT_TYPE_VEC2(data_b[ib].ds); + } + buf_b_qs[buf_ib * SHMEM_STRIDE + iqs] = data_b[ib].qs[iqs]; + } + + barrier(); + + pos_a_ib += 1; + pos_b_ib += 1; + +#ifdef COOPMAT + [[unroll]] for (uint cm_row = 0; cm_row < cms_per_row; cm_row++) { + const uint ib_a = warp_r * WM + cm_row * TM; + // Load from shared into cache + coopMatLoad(cache_a, buf_a_qs, ib_a * SHMEM_STRIDE, SHMEM_STRIDE, gl_CooperativeMatrixLayoutRowMajor); + + // TODO: only cache values that are actually needed + [[unroll]] for (uint t_idx = 0; t_idx < TM; t_idx++) { + cache_a_dm[t_idx] = buf_a_dm[ib_a + t_idx]; + } + + [[unroll]] for (uint cm_col = 0; cm_col < cms_per_col; cm_col++) { + const uint ib_b = warp_c * WN + cm_col * TN; + coopMatLoad(cache_b, buf_b_qs, ib_b * SHMEM_STRIDE, SHMEM_STRIDE, gl_CooperativeMatrixLayoutColumnMajor); + + // TODO: only cache values that are actually needed + [[unroll]] for (uint t_idx = 0; t_idx < TN; t_idx++) { + cache_b_dm[t_idx] = buf_b_d[ib_b + t_idx]; + } + + cm_result = coopmat(0); + cm_result = coopMatMulAdd(cache_a, cache_b, cm_result); + + [[unroll]] for (uint col = 0; col < TN; col += storestride) { + coopmat_stage[warp_i * TM * TN + (store_c + col) * TM + store_r] = ACC_TYPE(float(cache_a_d[store_r]) * float(cache_b_d[store_c + col])); + } + + coopMatLoad(factors, coopmat_stage, warp_i * TM * TN, TM, gl_CooperativeMatrixLayoutColumnMajor); + sums[cm_col * cms_per_row + cm_row] += factors * coopmat(cm_result); + } + } +#else + // Load from shared into cache + [[unroll]] for (uint wsir = 0; wsir < WMITER; wsir++) { + [[unroll]] for (uint cr = 0; cr < TM; cr++) { + const uint ib = warp_r * WM + wsir * WSUBM + tiwr * TM + cr; + cache_a_dm[wsir * TM + cr] = buf_a_dm[ib]; + [[unroll]] for (uint idx_k = 0; idx_k < BK / 4; idx_k++) { + cache_a_qs[(wsir * TM + cr) * (BK / 4) + idx_k] = buf_a_qs[ib * SHMEM_STRIDE + idx_k]; + } + } + } + + [[unroll]] for (uint wsic = 0; wsic < WNITER; wsic++) { + [[unroll]] for (uint cc = 0; cc < TN; cc++) { + const uint ib = warp_c * WN + wsic * WSUBN + tiwc * TN + cc; + cache_b_ds[cc] = buf_b_ds[ib]; + [[unroll]] for (uint idx_k = 0; idx_k < BK / 4; idx_k++) { + cache_b_qs[cc * (BK / 4) + idx_k] = buf_b_qs[ib * SHMEM_STRIDE + idx_k]; + } + } + + [[unroll]] for (uint wsir = 0; wsir < WMITER; wsir++) { + [[unroll]] for (uint cc = 0; cc < TN; cc++) { + [[unroll]] for (uint cr = 0; cr < TM; cr++) { + const uint cache_a_idx = wsir * TM + cr; + const uint sums_idx = (wsic * TN + cc) * (WMITER * TM) + wsir * TM + cr; + int32_t q_sum = 0; + [[unroll]] for (uint idx_k = 0; idx_k < BK / 4; idx_k++) { + q_sum += dotPacked4x8EXT(cache_a_qs[cache_a_idx * (BK / 4) + idx_k], + cache_b_qs[cc * (BK / 4) + idx_k]); + } + + sums[sums_idx] += mul_q8_1(q_sum, cache_a_dm[cache_a_idx], cache_b_ds[cc]); + } + } + } + } +#endif + + barrier(); + } + + const uint dr = ir * BM + warp_r * WM; + const uint dc = ic * BN + warp_c * WN; + +#ifndef MUL_MAT_ID + const uint offsets = batch_idx * p.batch_stride_d + ik * p.batch_stride_d * gl_NumWorkGroups.z; +#endif + +#ifdef COOPMAT +#ifdef MUL_MAT_ID + [[unroll]] for (uint cm_row = 0; cm_row < cms_per_row; cm_row++) { + [[unroll]] for (uint cm_col = 0; cm_col < cms_per_col; cm_col++) { + coopMatStore(sums[cm_col * cms_per_row + cm_row], coopmat_stage, warp_i * TM * TN, TM, gl_CooperativeMatrixLayoutColumnMajor); + + [[unroll]] for (uint col = 0; col < BN; col += storestride) { + const uint row_i = dc + cm_col * TN + col + store_c; + if (row_i >= _ne1) break; + + const u16vec2 row_idx = row_ids[row_i]; + + data_d[row_idx.y * p.batch_stride_d + row_idx.x * p.stride_d + dr + cm_row * TM + store_r] = D_TYPE(coopmat_stage[warp_i * TM * TN + (col + store_c) * TM + store_r]); + } + } + } +#else + const bool is_aligned = p.stride_d % 4 == 0; // Assumption: D_TYPE == float + + [[unroll]] for (uint cm_row = 0; cm_row < cms_per_row; cm_row++) { + [[unroll]] for (uint cm_col = 0; cm_col < cms_per_col; cm_col++) { + const bool is_in_bounds = dr + (cm_row + 1) * TM <= p.M && dc + (cm_col + 1) * TN <= p.N; + + if (is_aligned && is_in_bounds) { + // Full coopMat is within bounds and stride_d is aligned with 16B + coopmat cm_dtype = coopmat(sums[cm_col * cms_per_row + cm_row]); + coopMatStore(cm_dtype, data_d, offsets + (dc + cm_col * TN) * p.stride_d + dr + cm_row * TM, p.stride_d, gl_CooperativeMatrixLayoutColumnMajor); + } else if (is_in_bounds) { + // Full coopMat is within bounds, but stride_d is not aligned + coopMatStore(sums[cm_col * cms_per_row + cm_row], coopmat_stage, warp_i * TM * TN, TM, gl_CooperativeMatrixLayoutColumnMajor); + + [[unroll]] for (uint col = 0; col < TN; col += storestride) { + data_d[offsets + (dc + cm_col * TN + col + store_c) * p.stride_d + dr + cm_row * TM + store_r] = D_TYPE(coopmat_stage[warp_i * TM * TN + (col + store_c) * TM + store_r]); + } + } else if (dr + cm_row * TM < p.M && dc + cm_col * TN < p.N) { + // Partial coopMat is within bounds + coopMatStore(sums[cm_col * cms_per_row + cm_row], coopmat_stage, warp_i * TM * TN, TM, gl_CooperativeMatrixLayoutColumnMajor); + + [[unroll]] for (uint col = 0; col < TN; col += storestride) { + if (dr + cm_row * TM + store_r < p.M && dc + cm_col * TN + col + store_c < p.N) { + data_d[offsets + (dc + cm_col * TN + col + store_c) * p.stride_d + dr + cm_row * TM + store_r] = D_TYPE(coopmat_stage[warp_i * TM * TN + (col + store_c) * TM + store_r]); + } + } + } + } + } +#endif // MUL_MAT_ID +#else + [[unroll]] for (uint wsic = 0; wsic < WNITER; wsic++) { + [[unroll]] for (uint wsir = 0; wsir < WMITER; wsir++) { + + const uint dr_warp = dr + wsir * WSUBM + tiwr * TM; + const uint dc_warp = dc + wsic * WSUBN + tiwc * TN; + [[unroll]] for (uint cc = 0; cc < TN; cc++) { +#ifdef MUL_MAT_ID + const uint row_i = dc_warp + cc; + if (row_i >= _ne1) break; + + const u16vec2 row_idx = row_ids[row_i]; +#endif // MUL_MAT_ID + [[unroll]] for (uint cr = 0; cr < TM; cr++) { +#ifdef MUL_MAT_ID + data_d[row_idx.y * p.batch_stride_d + row_idx.x * p.stride_d + dr_warp + cr] = D_TYPE(sums[(wsic * TN + cc) * (WMITER * TM) + wsir * TM + cr]); +#else + if (dr_warp + cr < p.M && dc_warp + cc < p.N) { + data_d[offsets + (dc_warp + cc) * p.stride_d + dr_warp + cr] = D_TYPE(sums[(wsic * TN + cc) * (WMITER * TM) + wsir * TM + cr]); + } +#endif // MUL_MAT_ID + } + } + } + } +#endif // COOPMAT +} diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq_funcs.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq_funcs.comp new file mode 100644 index 000000000..63b15471b --- /dev/null +++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mmq_funcs.comp @@ -0,0 +1,99 @@ +#extension GL_EXT_shader_explicit_arithmetic_types_int32 : require +#extension GL_EXT_shader_explicit_arithmetic_types_int16 : require +#extension GL_EXT_shader_explicit_arithmetic_types_int8 : require + +#include "types.comp" + +// Each iqs value maps to a 32-bit integer + +#if defined(DATA_A_Q4_0) +i32vec2 repack(uint ib, uint iqs) { + // Use 2-byte loads since a q4_0 block (18 bytes) is not divisible by 4 + const u16vec2 quants = u16vec2(data_a[ib].qs[iqs * 2 ], + data_a[ib].qs[iqs * 2 + 1]); + const uint32_t vui = pack32(quants); + return i32vec2( vui & 0x0F0F0F0F, + (vui >> 4) & 0x0F0F0F0F); +} + +ACC_TYPE mul_q8_1(int32_t q_sum, float da, vec2 dsb) { + return ACC_TYPE(da * (float(q_sum) * dsb.x - 8.0f * dsb.y)); +} +#endif + +#if defined(DATA_A_Q4_1) +i32vec2 repack(uint ib, uint iqs) { + // Use 4-byte loads since a q4_1 block (20 bytes) is divisible by 4 + const uint32_t vui = data_a_packed32[ib].qs[iqs]; + return i32vec2( vui & 0x0F0F0F0F, + (vui >> 4) & 0x0F0F0F0F); +} + +ACC_TYPE mul_q8_1(int32_t q_sum, vec2 dma, vec2 dsb) { + return ACC_TYPE(float(q_sum) * dma.x * dsb.x + dma.y * dsb.y); +} +#endif + +#if defined(DATA_A_Q5_0) +i32vec2 repack(uint ib, uint iqs) { + // Use 2-byte loads since a q5_0 block (22 bytes) is not divisible by 4 + const u16vec2 quants = u16vec2(data_a[ib].qs[iqs * 2 ], + data_a[ib].qs[iqs * 2 + 1]); + const uint32_t vui = pack32(quants); + const int32_t qh = int32_t((uint32_t(data_a[ib].qh[1]) << 16 | data_a[ib].qh[0]) >> (4 * iqs)); + const int32_t v0 = int32_t(vui & 0x0F0F0F0F) + | ((qh & 0xF) * 0x02040810) & 0x10101010; // (0,1,2,3) -> (4,12,20,28) + + const int32_t v1 = int32_t((vui >> 4) & 0x0F0F0F0F) + | (((qh >> 16) & 0xF) * 0x02040810) & 0x10101010; // (16,17,18,19) -> (4,12,20,28) + + return i32vec2(v0, v1); +} + +ACC_TYPE mul_q8_1(int32_t q_sum, float da, vec2 dsb) { + return ACC_TYPE(da * (float(q_sum) * dsb.x - 16.0f * dsb.y)); +} +#endif + +#if defined(DATA_A_Q5_1) +i32vec2 repack(uint ib, uint iqs) { + // Use 4-byte loads since a q5_1 block (24 bytes) is divisible by 4 + const uint32_t vui = data_a_packed32[ib].qs[iqs]; + const int32_t qh = int32_t(data_a_packed32[ib].qh >> (4 * iqs)); + const int32_t v0 = int32_t(vui & 0x0F0F0F0F) + | ((qh & 0xF) * 0x02040810) & 0x10101010; // (0,1,2,3) -> (4,12,20,28) + + const int32_t v1 = int32_t((vui >> 4) & 0x0F0F0F0F) + | (((qh >> 16) & 0xF) * 0x02040810) & 0x10101010; // (16,17,18,19) -> (4,12,20,28) + + return i32vec2(v0, v1); +} + +ACC_TYPE mul_q8_1(int32_t q_sum, vec2 dma, vec2 dsb) { + return ACC_TYPE(float(q_sum) * dma.x * dsb.x + dma.y * dsb.y); +} +#endif + +#if defined(DATA_A_Q8_0) +int32_t repack(uint ib, uint iqs) { + // Use 2-byte loads since a q8_0 block (34 bytes) is not divisible by 4 + return pack32(i16vec2(data_a[ib].qs[iqs * 2 ], + data_a[ib].qs[iqs * 2 + 1])); +} + +ACC_TYPE mul_q8_1(int32_t q_sum, float da, vec2 dsb) { + return ACC_TYPE(float(q_sum) * da * dsb.x); +} +#endif + +#if defined(DATA_A_Q4_0) || defined(DATA_A_Q5_0) || defined(DATA_A_Q8_0) || defined(DATA_A_IQ1_S) || defined(DATA_A_IQ2_XXS) || defined(DATA_A_IQ2_XS) || defined(DATA_A_IQ2_S) || defined(DATA_A_IQ3_XXS) || defined(DATA_A_IQ3_S) || defined(DATA_A_IQ4_XS) || defined(DATA_A_IQ4_NL) +FLOAT_TYPE get_d(uint ib) { + return FLOAT_TYPE(data_a[ib].d); +} +#endif + +#if defined(DATA_A_Q4_1) || defined(DATA_A_Q5_1) +FLOAT_TYPE_VEC2 get_dm(uint ib) { + return FLOAT_TYPE_VEC2(data_a_packed32[ib].dm); +} +#endif diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/opt_step_adamw.comp b/ggml/src/ggml-vulkan/vulkan-shaders/opt_step_adamw.comp new file mode 100644 index 000000000..e0214fe76 --- /dev/null +++ b/ggml/src/ggml-vulkan/vulkan-shaders/opt_step_adamw.comp @@ -0,0 +1,42 @@ +#version 450 + +#include "generic_head.comp" +#include "types.comp" + +#extension GL_EXT_control_flow_attributes : enable + +layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in; + +layout (binding = 0) buffer X {A_TYPE x[];}; +layout (binding = 1) readonly buffer G {A_TYPE grad[];}; +layout (binding = 2) buffer GM {A_TYPE gradm[];}; +layout (binding = 3) buffer GV {A_TYPE gradv[];}; +layout (binding = 4) readonly buffer P {float params[7];}; + +void main() { + const uint i = gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x; + + if (i >= p.KX) { + return; + } + + const float alpha = params[0]; + const float beta1 = params[1]; + const float beta2 = params[2]; + const float eps = params[3]; + const float wd = params[4]; + const float beta1h = params[5]; + const float beta2h = params[6]; + + const float gi = grad[i]; + const float gmi = gradm[i]*beta1 + gi*(1.0f - beta1); + const float gvi = gradv[i]*beta2 + gi*gi*(1.0f - beta2); + + gradm[i] = gmi; + gradv[i] = gvi; + + const float mh = gmi*beta1h; + const float vh = sqrt(gvi*beta2h) + eps; + + x[i] = x[i]*(1.0f - alpha*wd) - alpha*mh/vh; +} diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/quantize_q8_1.comp b/ggml/src/ggml-vulkan/vulkan-shaders/quantize_q8_1.comp new file mode 100644 index 000000000..e2e020fec --- /dev/null +++ b/ggml/src/ggml-vulkan/vulkan-shaders/quantize_q8_1.comp @@ -0,0 +1,77 @@ +#version 450 + +#extension GL_EXT_control_flow_attributes : require +#extension GL_EXT_shader_16bit_storage : require + +layout (push_constant) uniform parameter +{ + uint ne; +} p; + +#include "types.comp" + +layout(constant_id = 0) const uint GROUP_SIZE = 32; +layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in; + +layout (binding = 0) readonly buffer A {vec4 data_a[];}; +layout (binding = 1) writeonly buffer D {block_q8_1_packed32 data_b[];}; + +shared float shmem[GROUP_SIZE]; + +void quantize() { + const uint wgid = gl_WorkGroupID.x; + const uint tid = gl_LocalInvocationID.x; + + // Each thread handles a vec4, so 8 threads handle a block + const uint blocks_per_group = GROUP_SIZE / 8; + + const uint block_in_wg = tid / 8; + + const uint ib = wgid * blocks_per_group + block_in_wg; + const uint iqs = tid % 8; + + if (ib >= gl_NumWorkGroups.x * blocks_per_group) { + return; + } + + const uint a_idx = ib * 8 + iqs; + + vec4 vals = a_idx < p.ne ? data_a[a_idx] : vec4(0.0f); + const vec4 abs_vals = abs(vals); + + // Find absolute max for each block + shmem[tid] = max(max(abs_vals.x, abs_vals.y), max(abs_vals.z, abs_vals.w)); + barrier(); + [[unroll]] for (uint s = 4; s > 0; s >>= 1) { + if (iqs < s) { + shmem[tid] = max(shmem[tid], shmem[tid + s]); + } + barrier(); + } + + const float amax = shmem[block_in_wg * 8]; + const float d = amax / 127.0; + const float d_inv = d != 0.0 ? 1.0 / d : 0.0; + vals = round(vals * d_inv); + data_b[ib].qs[iqs] = pack32(i8vec4(round(vals))); + barrier(); + + // Calculate the sum for each block + shmem[tid] = vals.x + vals.y + vals.z + vals.w; + barrier(); + [[unroll]] for (uint s = 4; s > 0; s >>= 1) { + if (iqs < s) { + shmem[tid] += shmem[tid + s]; + } + barrier(); + } + if (iqs == 0) { + const float sum = shmem[tid]; + + data_b[ib].ds = f16vec2(vec2(d, sum * d)); + } +} + +void main() { + quantize(); +} diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/repeat_back.comp b/ggml/src/ggml-vulkan/vulkan-shaders/repeat_back.comp new file mode 100644 index 000000000..d86279934 --- /dev/null +++ b/ggml/src/ggml-vulkan/vulkan-shaders/repeat_back.comp @@ -0,0 +1,37 @@ +#version 450 + +#include "types.comp" +#include "generic_unary_head.comp" + +layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in; + +void main() { + const uint idx = get_idx(); + + if (idx >= p.ne) { + return; + } + + // Destination multi-index (inlined dst_idx) + const uint i13 = fastdiv(idx, p.ne1_012mp, p.ne1_012L); + const uint i13_offset = i13 * p.ne12*p.ne11*p.ne10; + const uint i12 = fastdiv(idx - i13_offset, p.ne1_01mp, p.ne1_01L); + const uint i12_offset = i12*p.ne11*p.ne10; + const uint i11 = fastdiv(idx - i13_offset - i12_offset, p.ne1_0mp, p.ne1_0L); + const uint i10 = idx - i13_offset - i12_offset - i11*p.ne10; + const uint d_idx = i13*p.nb13 + i12*p.nb12 + i11*p.nb11 + i10*p.nb10; + + // Accumulate from sources + A_TYPE acc = A_TYPE(0); + for (uint i3 = i13; i3 < p.ne03; i3 += p.ne13) { + for (uint i2 = i12; i2 < p.ne02; i2 += p.ne12) { + for (uint i1 = i11; i1 < p.ne01; i1 += p.ne11) { + for (uint i0 = i10; i0 < p.ne00; i0 += p.ne10) { + acc += data_a[i3*p.nb03 + i2*p.nb02 + i1*p.nb01 + i0*p.nb00]; + } + } + } + } + + data_d[get_doffset() + d_idx] = D_TYPE(acc); +} diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/rms_norm_back.comp b/ggml/src/ggml-vulkan/vulkan-shaders/rms_norm_back.comp new file mode 100644 index 000000000..76009f3df --- /dev/null +++ b/ggml/src/ggml-vulkan/vulkan-shaders/rms_norm_back.comp @@ -0,0 +1,55 @@ +#version 450 + +#include "generic_head.comp" +#include "types.comp" + +#extension GL_EXT_control_flow_attributes : enable +#define BLOCK_SIZE 512 + +layout(local_size_x = BLOCK_SIZE, local_size_y = 1, local_size_z = 1) in; + +layout (binding = 0) readonly buffer G {A_TYPE data_a[];}; +layout (binding = 1) readonly buffer X {B_TYPE data_b[];}; +layout (binding = 2) writeonly buffer D {D_TYPE data_d[];}; + +shared FLOAT_TYPE sum_xx[BLOCK_SIZE]; +shared FLOAT_TYPE sum_xg[BLOCK_SIZE]; + +void main() { + const uint row = gl_WorkGroupID.z * 262144 + gl_WorkGroupID.y * 512 + gl_WorkGroupID.x; + const uint tid = gl_LocalInvocationID.x; + + // Compute derivative of x[i]/norm(x) = g[i]/norm(x) - x[i] dot(x,g)/KX / norm(x)^1.5 + + // partial sums for thread in warp + sum_xx[tid] = FLOAT_TYPE(0.0f); + sum_xg[tid] = FLOAT_TYPE(0.0f); + + [[unroll]] for (uint col = tid; col < p.KX; col += BLOCK_SIZE) { + const FLOAT_TYPE gi = FLOAT_TYPE(data_a[row*p.KX + col]); + const FLOAT_TYPE xi = FLOAT_TYPE(data_b[row*p.KX + col]); + sum_xx[tid] += xi * xi; + sum_xg[tid] += xi * gi; + } + + // sum up partial sums and write back result + barrier(); + [[unroll]] for (int s = BLOCK_SIZE / 2; s > 0; s >>= 1) { + if (tid < s) { + sum_xx[tid] += sum_xx[tid + s]; + sum_xg[tid] += sum_xg[tid + s]; + } + barrier(); + } + + const FLOAT_TYPE eps = FLOAT_TYPE(p.param1); + const FLOAT_TYPE mean = sum_xx[0] / FLOAT_TYPE(p.KX); + const FLOAT_TYPE scale_g = inversesqrt(mean + eps); + const FLOAT_TYPE scale_x = -scale_g * sum_xg[0] / (sum_xx[0] + FLOAT_TYPE(p.KX) * eps); + + [[unroll]] for (uint col = tid; col < p.KX; col += BLOCK_SIZE) { + data_d[row*p.KX + col] = D_TYPE( + scale_g * FLOAT_TYPE(data_a[row*p.KX + col]) + + scale_x * FLOAT_TYPE(data_b[row*p.KX + col])); + } +} diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/rope_head.comp b/ggml/src/ggml-vulkan/vulkan-shaders/rope_head.comp index 574b51ca5..96c9c4cbd 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/rope_head.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/rope_head.comp @@ -25,6 +25,11 @@ layout (push_constant) uniform parameter { float corr_dims[2]; float theta_scale; uint has_ff; + uint ne02; + uint s1; + uint s2; + int sections[4]; + uint is_back; } p; float rope_yarn_ramp(const float low, const float high, const uint i0) { @@ -44,6 +49,10 @@ void rope_yarn(const float theta_extrap, const uint i0, out float cos_theta, out // Get n-d magnitude scaling corrected for interpolation mscale *= 1.0f + 0.1f * log(1.0f / p.freq_scale); } + // Backprogagation uses inverted rotation + if (p.is_back != 0) { + theta = -theta; + } cos_theta = cos(theta) * mscale; sin_theta = sin(theta) * mscale; } diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/rope_multi.comp b/ggml/src/ggml-vulkan/vulkan-shaders/rope_multi.comp new file mode 100644 index 000000000..4f5b1a0ec --- /dev/null +++ b/ggml/src/ggml-vulkan/vulkan-shaders/rope_multi.comp @@ -0,0 +1,60 @@ +#version 450 + +#include "rope_head.comp" + +void main() { + const uint i0 = 2*gl_GlobalInvocationID.y; + uint ne0 = p.ncols; + uint ne1 = p.p_delta_rows; + uint ne2 = p.ne02; + + if (i0 >= ne0) { + return; + } + + const uint row_dst = gl_GlobalInvocationID.x; + + if (i0 >= p.n_dims) { + const uint i = row_dst*ne0 + i0; + + data_d[i + 0] = data_a[i + 0]; + data_d[i + 1] = data_a[i + 1]; + + return; + } + + const uint row_x = row_dst % ne1; + const uint channel_x = row_dst / ne1; + + const uint idst = row_dst*ne0 + i0/2; + const uint ix = channel_x*p.s2 + row_x*p.s1 + i0/2; + + const int sect_dims = p.sections[0] + p.sections[1] + p.sections[2] + p.sections[3]; + const int sec_w = p.sections[1] + p.sections[0]; + const uint sector = (i0 / 2) % sect_dims; + + float theta_base = 0.0; + if (sector < p.sections[0]) { + theta_base = data_pos[channel_x]*pow(p.theta_scale, i0/2.0f); + } + else if (sector >= p.sections[0] && sector < sec_w) { + theta_base = data_pos[channel_x + ne2 * 1]*pow(p.theta_scale, i0/2.0f); + } + else if (sector >= sec_w && sector < sec_w + p.sections[2]) { + theta_base = data_pos[channel_x + ne2 * 2]*pow(p.theta_scale, i0/2.0f); + } + else if (sector >= sec_w + p.sections[2]) { + theta_base = data_pos[channel_x + ne2 * 3]*pow(p.theta_scale, i0/2.0f); + } + + const float freq_factor = p.has_ff != 0 ? data_ff[i0/2] : 1.0f; + + float cos_theta, sin_theta; + rope_yarn(theta_base / freq_factor, i0, cos_theta, sin_theta); + + const float x0 = float(data_a[ix + 0]); + const float x1 = float(data_a[ix + p.n_dims/2]); + + data_d[idst + 0] = D_TYPE(x0*cos_theta - x1*sin_theta); + data_d[idst + p.n_dims/2] = D_TYPE(x0*sin_theta + x1*cos_theta); +} diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/rope_neox.comp b/ggml/src/ggml-vulkan/vulkan-shaders/rope_neox.comp index 83b46b69b..db775c456 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/rope_neox.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/rope_neox.comp @@ -3,15 +3,18 @@ #include "rope_head.comp" void main() { - const uint col = gl_GlobalInvocationID.y * 2; - const uint row = gl_GlobalInvocationID.x; + const uint i0 = 2*gl_GlobalInvocationID.y; + uint ne0 = p.ncols; + uint ne1 = p.p_delta_rows; - if (col >= p.ncols) { + if (i0 >= ne0) { return; } - if (col >= p.n_dims) { - const uint i = row*p.ncols + col; + const uint row_dst = gl_GlobalInvocationID.x; + + if (i0 >= p.n_dims) { + const uint i = row_dst*ne0 + i0; data_d[i + 0] = data_a[i + 0]; data_d[i + 1] = data_a[i + 1]; @@ -19,19 +22,22 @@ void main() { return; } - const uint i = row*p.ncols + col/2; - const uint i2 = row/p.p_delta_rows; + const uint row_x = row_dst % ne1; + const uint channel_x = row_dst / ne1; - const float theta_base = data_pos[i2] * pow(p.theta_scale, col/2.0f); + const uint idst = row_dst*ne0 + i0/2; + const uint ix = channel_x*p.s2 + row_x*p.s1 + i0/2; - const float freq_factor = p.has_ff != 0 ? data_ff[col/2] : 1.0f; + const float theta_base = data_pos[channel_x] * pow(p.theta_scale, i0/2.0f); + + const float freq_factor = p.has_ff != 0 ? data_ff[i0/2] : 1.0f; float cos_theta, sin_theta; - rope_yarn(theta_base / freq_factor, col, cos_theta, sin_theta); + rope_yarn(theta_base / freq_factor, i0, cos_theta, sin_theta); - const float x0 = float(data_a[i + 0]); - const float x1 = float(data_a[i + p.n_dims/2]); + const float x0 = float(data_a[ix + 0]); + const float x1 = float(data_a[ix + p.n_dims/2]); - data_d[i + 0] = D_TYPE(x0*cos_theta - x1*sin_theta); - data_d[i + p.n_dims/2] = D_TYPE(x0*sin_theta + x1*cos_theta); + data_d[idst + 0] = D_TYPE(x0*cos_theta - x1*sin_theta); + data_d[idst + p.n_dims/2] = D_TYPE(x0*sin_theta + x1*cos_theta); } diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/rope_norm.comp b/ggml/src/ggml-vulkan/vulkan-shaders/rope_norm.comp index e416ad938..4ad35e549 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/rope_norm.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/rope_norm.comp @@ -3,15 +3,18 @@ #include "rope_head.comp" void main() { - const uint col = gl_GlobalInvocationID.y * 2; - const uint row = gl_GlobalInvocationID.x; + const uint i0 = 2*gl_GlobalInvocationID.y; + uint ne0 = p.ncols; + uint ne1 = p.p_delta_rows; - if (col >= p.ncols) { + if (i0 >= ne0) { return; } - if (col >= p.n_dims) { - const uint i = row*p.ncols + col; + const uint row_dst = gl_GlobalInvocationID.x; + + if (i0 >= p.n_dims) { + const uint i = row_dst*ne0 + i0; data_d[i + 0] = data_a[i + 0]; data_d[i + 1] = data_a[i + 1]; @@ -19,19 +22,22 @@ void main() { return; } - const uint i = row*p.ncols + col; - const uint i2 = row/p.p_delta_rows; + const uint row_x = row_dst % ne1; + const uint channel_x = row_dst / ne1; - const float theta_base = data_pos[i2] * pow(p.theta_scale, col/2.0f); + const uint idst = row_dst*ne0 + i0; + const uint ix = channel_x*p.s2 + row_x*p.s1 + i0; - const float freq_factor = p.has_ff != 0 ? data_ff[col/2] : 1.0f; + const float theta_base = data_pos[channel_x] * pow(p.theta_scale, i0/2.0f); + + const float freq_factor = p.has_ff != 0 ? data_ff[i0/2] : 1.0f; float cos_theta, sin_theta; - rope_yarn(theta_base / freq_factor, col, cos_theta, sin_theta); + rope_yarn(theta_base / freq_factor, i0, cos_theta, sin_theta); - const float x0 = float(data_a[i + 0]); - const float x1 = float(data_a[i + 1]); + const float x0 = float(data_a[ix + 0]); + const float x1 = float(data_a[ix + 1]); - data_d[i + 0] = D_TYPE(x0*cos_theta - x1*sin_theta); - data_d[i + 1] = D_TYPE(x0*sin_theta + x1*cos_theta); + data_d[idst + 0] = D_TYPE(x0*cos_theta - x1*sin_theta); + data_d[idst + 1] = D_TYPE(x0*sin_theta + x1*cos_theta); } diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/rope_vision.comp b/ggml/src/ggml-vulkan/vulkan-shaders/rope_vision.comp new file mode 100644 index 000000000..cedacc4d1 --- /dev/null +++ b/ggml/src/ggml-vulkan/vulkan-shaders/rope_vision.comp @@ -0,0 +1,47 @@ +#version 450 + +#include "rope_head.comp" + +void main() { + const uint i0 = 2*gl_GlobalInvocationID.y; + uint ne0 = p.ncols; + uint ne1 = p.p_delta_rows; + uint ne2 = p.ne02; + + if (i0 >= ne0) { + return; + } + + const uint row_dst = gl_GlobalInvocationID.x; + + const uint row_x = row_dst % ne1; + const uint channel_x = row_dst / ne1; + + const uint idst = row_dst*ne0 + i0/2; + const uint ix = channel_x*p.s2 + row_x*p.s1 + i0/2; + + const int sect_dims = p.sections[0] + p.sections[1]; + const int sec_w = p.sections[1] + p.sections[0]; + const uint sector = (i0 / 2) % sect_dims; + + float theta_base = 0.0; + if (sector < p.sections[0]) { + const uint p0 = sector; + theta_base = data_pos[channel_x]*pow(p.theta_scale, p0); + } + else if (sector >= p.sections[0] && sector < sec_w) { + const uint p0 = sector - p.sections[0]; + theta_base = data_pos[channel_x + ne2]*pow(p.theta_scale, p0); + } + + const float freq_factor = p.has_ff != 0 ? data_ff[i0/2] : 1.0f; + + float cos_theta, sin_theta; + rope_yarn(theta_base / freq_factor, i0, cos_theta, sin_theta); + + const float x0 = float(data_a[ix + 0]); + const float x1 = float(data_a[ix + p.n_dims]); + + data_d[idst + 0] = D_TYPE(x0*cos_theta - x1*sin_theta); + data_d[idst + p.n_dims] = D_TYPE(x0*sin_theta + x1*cos_theta); +} diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/sigmoid.comp b/ggml/src/ggml-vulkan/vulkan-shaders/sigmoid.comp new file mode 100644 index 000000000..776581e2c --- /dev/null +++ b/ggml/src/ggml-vulkan/vulkan-shaders/sigmoid.comp @@ -0,0 +1,20 @@ +#version 450 + +#include "generic_head.comp" +#include "types.comp" + +#extension GL_EXT_control_flow_attributes : enable + +layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in; + +layout (binding = 0) readonly buffer X {A_TYPE data_a[];}; +layout (binding = 1) writeonly buffer D {D_TYPE data_d[];}; + +void main() { + const uint i = gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x; + + if (i >= p.KX) { + return; + } + data_d[i] = D_TYPE(1. / (1 + exp(-1. *data_a[i]))); +} diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/silu_back.comp b/ggml/src/ggml-vulkan/vulkan-shaders/silu_back.comp new file mode 100644 index 000000000..f9afa9b13 --- /dev/null +++ b/ggml/src/ggml-vulkan/vulkan-shaders/silu_back.comp @@ -0,0 +1,26 @@ +#version 450 + +#include "generic_head.comp" +#include "types.comp" + +#extension GL_EXT_control_flow_attributes : enable + +layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in; + +layout (binding = 0) readonly buffer G {A_TYPE data_g[];}; +layout (binding = 1) readonly buffer X {B_TYPE data_x[];}; +layout (binding = 2) writeonly buffer D {D_TYPE data_d[];}; + +void main() { + const uint i = gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x; + + if (i >= p.KX) { + return; + } + + // Compute derivative of SiLU(x): 1/(1+exp(-x)) - x*exp(-x)/(1+exp(-x))^2 + + const float xi = float(data_x[i]); + const float s = 1.0f / (1.0f + exp(-xi)); + data_d[i] = D_TYPE(data_g[i] * (s + xi * s * (1 - s))); +} diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/soft_max_back.comp b/ggml/src/ggml-vulkan/vulkan-shaders/soft_max_back.comp new file mode 100644 index 000000000..29bd77d7e --- /dev/null +++ b/ggml/src/ggml-vulkan/vulkan-shaders/soft_max_back.comp @@ -0,0 +1,50 @@ +#version 450 + +#extension GL_EXT_control_flow_attributes : enable + +#include "generic_head.comp" +#include "types.comp" + +layout(constant_id = 0) const uint BLOCK_SIZE = 32; +layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in; + +// In this shader Y = softmax(X) and X is not provided as input. + +layout (binding = 0) readonly buffer G {A_TYPE data_g[];}; +layout (binding = 1) readonly buffer Y {B_TYPE data_y[];}; +layout (binding = 2) buffer D {D_TYPE data_d[];}; + +shared FLOAT_TYPE sum_yg[BLOCK_SIZE]; + +void main() { + const uint row = gl_WorkGroupID.z * 262144 + gl_WorkGroupID.y * 512 + gl_WorkGroupID.x; + const uint tid = gl_LocalInvocationID.x; + + FLOAT_TYPE scale = p.param1; + + // partial sums for thread in warp + sum_yg[tid] = FLOAT_TYPE(0.0f); + + [[unroll]] for (uint col = tid; col < p.KX; col += BLOCK_SIZE) { + const FLOAT_TYPE gi = FLOAT_TYPE(data_g[row*p.KX + col]); + const FLOAT_TYPE yi = FLOAT_TYPE(data_y[row*p.KX + col]); + sum_yg[tid] += yi * gi; + } + + // sum up partial sums and write back result + barrier(); + [[unroll]] for (uint s = BLOCK_SIZE / 2; s > 0; s >>= 1) { + if (tid < s) { + sum_yg[tid] += sum_yg[tid + s]; + } + barrier(); + } + + const FLOAT_TYPE dot_yg = sum_yg[0]; + + [[unroll]] for (uint col = tid; col < p.KX; col += BLOCK_SIZE) { + data_d[row*p.KX + col] = D_TYPE(scale + * (FLOAT_TYPE(data_g[row*p.KX + col]) - dot_yg) + * FLOAT_TYPE(data_y[row*p.KX + col])); + } +} diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/sub.comp b/ggml/src/ggml-vulkan/vulkan-shaders/sub.comp new file mode 100644 index 000000000..72353cc32 --- /dev/null +++ b/ggml/src/ggml-vulkan/vulkan-shaders/sub.comp @@ -0,0 +1,29 @@ +#version 450 + +#extension GL_EXT_shader_16bit_storage : require + +#include "types.comp" +#include "generic_binary_head.comp" + +const uint num_threads = 256; + +layout(local_size_x = num_threads, local_size_y = 1, local_size_z = 1) in; + +void main() { + uint idx = get_idx(); + + // num_threads * num_iter must equal 512, to match the wg_denoms and get_idx calculation + const uint num_iter = 2; + + [[unroll]] for (uint i = 0; i < num_iter; ++i) { + if (idx >= p.ne) { + continue; + } + uint i00, i01, i02, i03; + get_indices(idx, i00, i01, i02, i03); + + data_d[get_doffset() + dst_idx(i00, i01, i02, i03)] = D_TYPE(FLOAT_TYPE(data_a[get_aoffset() + src0_idx(i00, i01, i02, i03)]) - FLOAT_TYPE(data_b[get_boffset() + src1_idx(i00, i01, i02, i03)])); + + idx += num_threads; + } +} diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/test_integer_dot_support.comp b/ggml/src/ggml-vulkan/vulkan-shaders/test_integer_dot_support.comp new file mode 100644 index 000000000..470e3074d --- /dev/null +++ b/ggml/src/ggml-vulkan/vulkan-shaders/test_integer_dot_support.comp @@ -0,0 +1,7 @@ +#version 460 + +#extension GL_EXT_integer_dot_product : require + +void main() +{ +} diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/types.comp b/ggml/src/ggml-vulkan/vulkan-shaders/types.comp index db643a54c..f5b29bfb1 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/types.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/types.comp @@ -1,7 +1,7 @@ - #if !defined(GGML_TYPES_COMP) #define GGML_TYPES_COMP +#extension GL_EXT_shader_explicit_arithmetic_types_int64 : require #extension GL_EXT_shader_explicit_arithmetic_types_int32 : require #extension GL_EXT_shader_explicit_arithmetic_types_int16 : require #extension GL_EXT_shader_explicit_arithmetic_types_int8 : require @@ -50,6 +50,7 @@ struct block_q4_0_packed16 #if defined(DATA_A_Q4_0) #define QUANT_K QUANT_K_Q4_0 #define QUANT_R QUANT_R_Q4_0 +#define QUANT_AUXF 1 #define A_TYPE block_q4_0 #define A_TYPE_PACKED16 block_q4_0_packed16 #endif @@ -71,11 +72,19 @@ struct block_q4_1_packed16 uint16_t qs[16/2]; }; +struct block_q4_1_packed32 +{ + f16vec2 dm; + uint32_t qs[16/4]; +}; + #if defined(DATA_A_Q4_1) #define QUANT_K QUANT_K_Q4_1 #define QUANT_R QUANT_R_Q4_1 +#define QUANT_AUXF 2 #define A_TYPE block_q4_1 #define A_TYPE_PACKED16 block_q4_1_packed16 +#define A_TYPE_PACKED32 block_q4_1_packed32 #endif #define QUANT_K_Q5_0 32 @@ -98,6 +107,7 @@ struct block_q5_0_packed16 #if defined(DATA_A_Q5_0) #define QUANT_K QUANT_K_Q5_0 #define QUANT_R QUANT_R_Q5_0 +#define QUANT_AUXF 1 #define A_TYPE block_q5_0 #define A_TYPE_PACKED16 block_q5_0_packed16 #endif @@ -121,11 +131,20 @@ struct block_q5_1_packed16 uint16_t qs[16/2]; }; +struct block_q5_1_packed32 +{ + f16vec2 dm; + uint qh; + uint32_t qs[16/4]; +}; + #if defined(DATA_A_Q5_1) #define QUANT_K QUANT_K_Q5_1 #define QUANT_R QUANT_R_Q5_1 +#define QUANT_AUXF 2 #define A_TYPE block_q5_1 #define A_TYPE_PACKED16 block_q5_1_packed16 +#define A_TYPE_PACKED32 block_q5_1_packed32 #endif #define QUANT_K_Q8_0 32 @@ -139,16 +158,42 @@ struct block_q8_0 struct block_q8_0_packed16 { float16_t d; - uint16_t qs[32/2]; + int16_t qs[32/2]; +}; +struct block_q8_0_packed32 +{ + float16_t d; + int32_t qs[32/4]; }; #if defined(DATA_A_Q8_0) #define QUANT_K QUANT_K_Q8_0 #define QUANT_R QUANT_R_Q8_0 +#define QUANT_AUXF 1 #define A_TYPE block_q8_0 #define A_TYPE_PACKED16 block_q8_0_packed16 +#define A_TYPE_PACKED32 block_q8_0_packed32 #endif +#define QUANT_K_Q8_1 32 +#define QUANT_R_Q8_1 1 + +struct block_q8_1 +{ + f16vec2 ds; + int8_t qs[32]; +}; +struct block_q8_1_packed16 +{ + f16vec2 ds; + int16_t qs[16]; +}; +struct block_q8_1_packed32 +{ + f16vec2 ds; + int32_t qs[8]; +}; + // K-quants #define QUANT_K_Q2_K 256 @@ -294,6 +339,196 @@ struct block_q6_K_packed16 // IQuants +#define QUANT_K_IQ1_S 256 +#define QUANT_R_IQ1_S 1 + +struct block_iq1_s { + float16_t d; + uint8_t qs[QUANT_K_IQ1_S/8]; + uint16_t qh[QUANT_K_IQ1_S/32]; +}; + +#define QUANT_K_IQ1_M 256 +#define QUANT_R_IQ1_M 1 + +struct block_iq1_m { + uint8_t qs[QUANT_K_IQ1_M/8]; + uint8_t qh[QUANT_K_IQ1_M/16]; + uint16_t scales[QUANT_K_IQ1_M/64]; +}; + +struct block_iq1_m_packed64 { + uint64_t qs[QUANT_K_IQ1_M/8/8]; + uint64_t qh[QUANT_K_IQ1_M/16/8]; + uint64_t scales; +}; + +#if defined(DATA_A_IQ1_S) +#define QUANT_K QUANT_K_IQ1_S +#define QUANT_R QUANT_R_IQ1_S +#define A_TYPE block_iq1_s +#endif + +#if defined(DATA_A_IQ1_M) +#define QUANT_K QUANT_K_IQ1_M +#define QUANT_R QUANT_R_IQ1_M +#define A_TYPE block_iq1_m +#endif + +#if defined(DATA_A_IQ1_S) || defined(DATA_A_IQ1_M) +#define IQ1S_DELTA 0.125f +#define IQ1M_DELTA 0.125f + +// Packed IQ1S grid where every 2 vec8 are encoded on 32 bits (2 bits per coordinate). +const uint[1024] iq1s_grid_const = { + 0xfffdffff, 0xfff7fff0, 0xffccfff5, 0xffdfffc0, 0xffd7ffdd, 0xff30ffd5, 0xff03ff0c, 0xff10ff01, + 0xff7dff7f, 0xff75ff77, 0xff5fff40, 0xff57ff5d, 0xfcf3ff55, 0xfcccfcf0, 0xfcc1fcc3, 0xfcc5fcc4, + 0xfc3cfcd0, 0xfc34fc31, 0xfc00fc0d, 0xfc1cfc05, 0xfc11fc13, 0xfc70fc17, 0xfc43fc4c, 0xfc50fc41, + 0xfdfdfdff, 0xfdf5fdf7, 0xfddffdc0, 0xfdd7fddd, 0xfd30fdd5, 0xfd04fd0c, 0xfd14fd13, 0xfd7dfd7f, + 0xfd75fd77, 0xfd40fd4c, 0xfd5ffd44, 0xfd57fd5d, 0xf3ccfd55, 0xf3c1f3c3, 0xf33cf3d0, 0xf300f334, + 0xf313f305, 0xf34cf310, 0xf350f344, 0xf0f3f0fc, 0xf0f1f0f0, 0xf0c7f0c0, 0xf0d4f0c5, 0xf030f03f, + 0xf00ff035, 0xf003f00c, 0xf001f000, 0xf01ff004, 0xf010f01d, 0xf015f017, 0xf04cf07c, 0xf047f040, + 0xf05cf045, 0xf050f053, 0xf054f051, 0xf1c4f1c3, 0xf133f13c, 0xf10df10f, 0xf107f100, 0xf11cf11f, + 0xf114f111, 0xf14cf170, 0xf144f143, 0xf7fdf7ff, 0xf7f5f7f7, 0xf7dff7c0, 0xf7d7f7dd, 0xf730f7d5, + 0xf701f70c, 0xf77ff710, 0xf777f77d, 0xf740f775, 0xf75df75f, 0xf755f757, 0xf4ccf4f0, 0xf4c4f4c3, + 0xf4d0f4d3, 0xf40ff43c, 0xf400f40c, 0xf413f41c, 0xf44cf414, 0xf441f443, 0xf450f444, 0xf5fdf5ff, + 0xf5f5f5f7, 0xf5dff5c0, 0xf5d7f5dd, 0xf530f5d5, 0xf504f50c, 0xf510f51c, 0xf57df57f, 0xf577f570, + 0xf540f575, 0xf55df55f, 0xf555f557, 0xcfcccfcf, 0xcfc4cfc3, 0xcfd0cfd3, 0xcf33cf3c, 0xcf00cf0f, + 0xcf1ccf07, 0xcf10cf13, 0xcf4ccf14, 0xcf41cf43, 0xcf50cf5c, 0xccf3ccfc, 0xccf4ccf1, 0xcccdcccf, + 0xccc7ccc0, 0xccd3ccdc, 0xcc30ccd4, 0xcc0fcc35, 0xcc0dcc0c, 0xcc00cc03, 0xcc04cc01, 0xcc10cc1f, + 0xcc4dcc73, 0xcc5ccc40, 0xcdcccc53, 0xcdc1cdc3, 0xcd3fcdd0, 0xcd34cd31, 0xcd00cd0d, 0xcd05cd07, + 0xcd11cd13, 0xcd4ccd70, 0xcd41cd43, 0xc3fccd50, 0xc3f4c3f1, 0xc3c0c3c3, 0xc3c4c3c7, 0xc3d1c3dc, + 0xc330c33c, 0xc337c331, 0xc30cc335, 0xc300c303, 0xc304c301, 0xc310c31d, 0xc373c317, 0xc34fc374, + 0xc340c343, 0xc344c347, 0xc35cc345, 0xc350c353, 0xc0fdc354, 0xc0f5c0f0, 0xc0c3c0cc, 0xc0c1c0c0, + 0xc0dfc0c4, 0xc0d0c0dd, 0xc0d5c0d7, 0xc033c03c, 0xc031c030, 0xc00dc00c, 0xc000c003, 0xc004c001, + 0xc01cc005, 0xc010c013, 0xc014c011, 0xc07dc07f, 0xc070c073, 0xc075c077, 0xc04cc04f, 0xc040c043, + 0xc044c041, 0xc05fc045, 0xc050c05d, 0xc1f3c1fc, 0xc1f1c1f0, 0xc1c1c1c0, 0xc1c5c1c7, 0xc1d1c1dc, + 0xc13dc13f, 0xc130c133, 0xc135c137, 0xc100c10c, 0xc107c101, 0xc11cc104, 0xc110c113, 0xc114c117, + 0xc171c115, 0xc14dc175, 0xc153c140, 0xc7ccc154, 0xc7d0c7c1, 0xc733c73c, 0xc734c731, 0xc700c70f, + 0xc705c707, 0xc71cc71f, 0xc711c713, 0xc770c714, 0xc743c74c, 0xc4cfc750, 0xc4c0c4cd, 0xc4dcc4c5, + 0xc43dc4d0, 0xc430c433, 0xc40cc437, 0xc400c403, 0xc404c401, 0xc41fc405, 0xc415c410, 0xc44cc474, + 0xc440c44d, 0xc45cc447, 0xc454c451, 0xc5c1c5f4, 0xc5d1c5d3, 0xc531c533, 0xc50fc534, 0xc500c50d, + 0xc51cc507, 0xc514c511, 0xc54cc570, 0xc545c541, 0xdffddfff, 0xdff5dff7, 0xdfdfdfc0, 0xdfd0dfdd, + 0xdfd5dfd7, 0xdf0cdf30, 0xdf1cdf04, 0xdf7fdf10, 0xdf77df7d, 0xdf40df75, 0xdf5ddf5f, 0xdf57df50, + 0xdcf0df55, 0xdcc3dccc, 0xdcd0dcc4, 0xdc33dc3d, 0xdc00dc34, 0xdc05dc07, 0xdc13dc1c, 0xdc11dc10, + 0xdc4fdc70, 0xdc44dc41, 0xddfcdc50, 0xddf5ddf7, 0xddc0ddcc, 0xdddddddf, 0xddd5ddd7, 0xdd0cdd30, + 0xdd04dd01, 0xdd7cdd10, 0xdd75dd77, 0xdd40dd4c, 0xdd5ddd5f, 0xdd55dd57, 0xd3c3d3f0, 0xd3c4d3c1, + 0xd333d3d0, 0xd331d330, 0xd30dd334, 0xd307d300, 0xd311d305, 0xd34cd370, 0xd344d343, 0xd350d35c, + 0xd0c0d0f4, 0xd0d4d0dc, 0xd030d03f, 0xd00cd037, 0xd000d003, 0xd01dd004, 0xd017d010, 0xd04fd074, + 0xd040d043, 0xd045d047, 0xd053d05c, 0xd054d051, 0xd1cfd1f0, 0xd1c4d1cd, 0xd13cd1d0, 0xd100d134, + 0xd11cd11f, 0xd173d114, 0xd14fd171, 0xd7ffd145, 0xd7f7d7fd, 0xd7c0d7f5, 0xd7ddd7df, 0xd7d5d7d7, + 0xd70cd730, 0xd710d703, 0xd77dd77f, 0xd775d777, 0xd75dd75f, 0xd755d757, 0xd4ccd4f4, 0xd4c4d4c3, + 0xd431d4d0, 0xd40dd434, 0xd41cd400, 0xd411d413, 0xd470d414, 0xd441d44f, 0xd453d444, 0xd5ffd450, + 0xd5f7d5fd, 0xd5dfd5f5, 0xd5d7d5dd, 0xd530d5d5, 0xd501d50c, 0xd510d504, 0xd57dd57f, 0xd575d577, + 0xd55fd540, 0xd557d55d, 0x3ff0d555, 0x3fc13fcc, 0x3f343fd0, 0x3f003f0d, 0x3f053f07, 0x3f133f1c, + 0x3f433f11, 0x3f5c3f44, 0x3cff3f51, 0x3cf33cfc, 0x3cf43cf1, 0x3cc03ccd, 0x3cc73cc1, 0x3cdc3cc5, + 0x3cd43cd1, 0x3c373c30, 0x3c0c3c35, 0x3c003c03, 0x3c043c01, 0x3c103c05, 0x3c153c17, 0x3c733c7c, + 0x3c4f3c71, 0x3c403c4d, 0x3c5c3c5f, 0x3df03c5d, 0x3dc33dcc, 0x3dd03dc1, 0x3d0d3d3c, 0x3d053d00, + 0x3d143d13, 0x3d433d74, 0x33fc3d50, 0x33c433c0, 0x333033d4, 0x33353337, 0x3303330c, 0x33013300, + 0x331d331c, 0x33173310, 0x337c3315, 0x33743371, 0x334d334f, 0x335f3340, 0x3354335c, 0x30fd30fc, + 0x30f530f0, 0x30c330cc, 0x30c130c0, 0x30df30c4, 0x30d530d0, 0x3033303c, 0x30313030, 0x300f3034, + 0x3003300c, 0x30013000, 0x30043007, 0x3013301c, 0x30113010, 0x307d3014, 0x30703073, 0x304c3077, + 0x30403043, 0x30443041, 0x30503045, 0x30553057, 0x31f031fc, 0x31c331f4, 0x31c731c0, 0x31dc31c5, + 0x31d431d3, 0x313d313f, 0x31373130, 0x310c310f, 0x3100310d, 0x31043101, 0x3110311d, 0x317c3117, + 0x31753170, 0x31403143, 0x3153315c, 0x37f03151, 0x37c037cc, 0x37d037c5, 0x3734373d, 0x3700370f, + 0x371c3707, 0x37113713, 0x37703714, 0x3743374c, 0x37443741, 0x34fc3750, 0x34f134f0, 0x34cf34f5, + 0x34c034c3, 0x34dc34c7, 0x34d134d3, 0x3430343f, 0x340c3435, 0x3403340d, 0x34013400, 0x341f3404, + 0x3410341d, 0x34153411, 0x34743471, 0x3440344d, 0x34473441, 0x3453345c, 0x34543451, 0x353335c1, + 0x35343531, 0x35073500, 0x35133505, 0x35433514, 0x0ffc3550, 0x0ff00ff3, 0x0ff40ff1, 0x0fc00fcd, + 0x0fdc0fc5, 0x0fd40fd3, 0x0f300f3f, 0x0f0c0f37, 0x0f000f03, 0x0f040f01, 0x0f170f10, 0x0f740f71, + 0x0f470f40, 0x0f5c0f5f, 0x0f540f51, 0x0cf70cf0, 0x0cf50cf4, 0x0cc30ccc, 0x0cc10cc0, 0x0cc40cc7, + 0x0cd00cdf, 0x0cd70cd1, 0x0c3c0cd5, 0x0c300c33, 0x0c340c31, 0x0c0c0c0f, 0x0c030c0d, 0x0c010c00, + 0x0c040c07, 0x0c1c0c05, 0x0c100c13, 0x0c140c11, 0x0c700c7d, 0x0c430c4c, 0x0c410c40, 0x0c5f0c44, + 0x0c550c50, 0x0df10dfc, 0x0dc00dcd, 0x0ddc0dc5, 0x0d3d0dd3, 0x0d350d30, 0x0d030d0c, 0x0d010d00, + 0x0d1d0d04, 0x0d700d10, 0x0d4d0d4f, 0x0d440d40, 0x0d530d45, 0x03f003f3, 0x03c303cc, 0x03c103c0, + 0x03c403c7, 0x03d003dc, 0x03d503d7, 0x0333033c, 0x03310330, 0x03350334, 0x030c030f, 0x03000303, + 0x03070301, 0x03050304, 0x031d031c, 0x03100313, 0x03140311, 0x0377037f, 0x034c0375, 0x03400343, + 0x03440341, 0x0353035c, 0x03550350, 0x00fd00fc, 0x00f000f3, 0x00f400f1, 0x00cc00cf, 0x00c300cd, + 0x00c100c0, 0x00c500c4, 0x00d300dc, 0x00d100d0, 0x003f00d4, 0x003d003c, 0x00300033, 0x00370031, + 0x000f0034, 0x000d000c, 0x00000003, 0x00070001, 0x00050004, 0x001c001f, 0x00100013, 0x00170011, + 0x00150014, 0x0073007c, 0x00740070, 0x004f0075, 0x0043004c, 0x00410040, 0x00440047, 0x0053005c, + 0x00510050, 0x01ff0054, 0x01fd01fc, 0x01f101f3, 0x01f401f7, 0x01c301cc, 0x01c701c0, 0x01df01c4, + 0x01dd01dc, 0x01d001d3, 0x01d701d1, 0x013c01d4, 0x01310130, 0x01340137, 0x010f0135, 0x010d010c, + 0x01000103, 0x01070101, 0x01050104, 0x0113011c, 0x01140110, 0x0170017d, 0x01770171, 0x01750174, + 0x0140014c, 0x015d0145, 0x01510150, 0x01540157, 0x07f007f3, 0x07f407f1, 0x07c007cf, 0x07dc07c7, + 0x073007d5, 0x07350737, 0x0703070c, 0x07010700, 0x07040707, 0x071d071f, 0x07100713, 0x0774077d, + 0x074d074f, 0x07470740, 0x0754075c, 0x04fd04fc, 0x04f504f0, 0x04c304cc, 0x04c104c0, 0x04d004c4, + 0x0433043c, 0x04310430, 0x040f0434, 0x040d040c, 0x04000403, 0x04070401, 0x04050404, 0x0413041c, + 0x04110410, 0x047c0414, 0x04740470, 0x0443044c, 0x04410440, 0x04440447, 0x05f30450, 0x05c005f7, + 0x05df05c5, 0x05d105d0, 0x053005d4, 0x05340537, 0x0500050c, 0x05070501, 0x051d0504, 0x05170510, + 0x057c0515, 0x054d0575, 0x05410540, 0x05450547, 0x1ff0055c, 0x1fc11fc3, 0x1fd01fc4, 0x1f0f1f33, + 0x1f011f00, 0x1f051f07, 0x1f131f1c, 0x1f141f11, 0x1f411f7c, 0x1cfc1f50, 0x1cf11cf3, 0x1ccd1cf4, + 0x1cdc1cc0, 0x1cd11cdd, 0x1c301cd4, 0x1c0c1c34, 0x1c011c00, 0x1c101c04, 0x1c151c11, 0x1c751c73, + 0x1c401c4d, 0x1c511c5c, 0x1dcc1c54, 0x1dc41dc1, 0x1d3c1d3f, 0x1d001d31, 0x1d071d01, 0x1d701d1f, + 0x1d411d4c, 0x13cc1d50, 0x13c013cd, 0x13c513c1, 0x13d113dc, 0x133f13d4, 0x1330133d, 0x13351337, + 0x1303130c, 0x13011300, 0x13051304, 0x131d131f, 0x13731310, 0x13741370, 0x134d134f, 0x13401343, + 0x13471341, 0x135c1345, 0x13541353, 0x10f710f0, 0x10cc10f5, 0x10c110c0, 0x103310c4, 0x10311030, + 0x100f1034, 0x1003100c, 0x10011000, 0x101c1004, 0x10101013, 0x10141011, 0x10741071, 0x104c1075, + 0x10411040, 0x10451044, 0x1050105d, 0x10571051, 0x11f411fd, 0x11df11c0, 0x11d711d1, 0x113f11d4, + 0x11371130, 0x110c1135, 0x11001103, 0x11071101, 0x111f1105, 0x11171110, 0x117d117f, 0x11751170, + 0x11411143, 0x11441147, 0x1153115f, 0x11551151, 0x17c417c1, 0x173c17d0, 0x1700170d, 0x171c1705, + 0x17701714, 0x1747174c, 0x14fc1751, 0x14cf14f3, 0x14dc14c0, 0x14d114d3, 0x143f14d4, 0x1430143c, + 0x14371431, 0x1403140c, 0x14011400, 0x141f1404, 0x14151410, 0x1473147d, 0x14401475, 0x1453145c, + 0x14541450, 0x15c115cc, 0x153c15c7, 0x15341533, 0x1500150f, 0x15051507, 0x15101513, 0x15711514, + 0x15471543, 0x15511545, 0x7ffd7fff, 0x7ff57ff7, 0x7fdd7fdf, 0x7fd57fd7, 0x7f0f7f30, 0x7f037f0c, + 0x7f047f01, 0x7f7f7f10, 0x7f777f7d, 0x7f407f75, 0x7f5d7f5f, 0x7f557f57, 0x7ccc7cf0, 0x7cc17cc3, + 0x7cd07cc4, 0x7c337c3c, 0x7c0f7c34, 0x7c007c0d, 0x7c077c01, 0x7c137c04, 0x7c147c11, 0x7c747c70, + 0x7c417c43, 0x7c507c44, 0x7dfd7dff, 0x7df57df7, 0x7ddf7dc0, 0x7dd77ddd, 0x7d0c7dd5, 0x7d047d03, + 0x7d7f7d10, 0x7d777d7d, 0x7d407d75, 0x7d5d7d5f, 0x7d557d57, 0x73c473c3, 0x7333733c, 0x7300730c, + 0x731c7305, 0x73147313, 0x73447343, 0x70f470fc, 0x70c070cd, 0x70d170c5, 0x703f70d4, 0x7030703c, + 0x700c7037, 0x70007003, 0x70047001, 0x70107005, 0x70177011, 0x707c7015, 0x70717073, 0x704f7074, + 0x7040704d, 0x70517047, 0x71c171cc, 0x71d071c4, 0x7133713c, 0x71357134, 0x7100710f, 0x71057104, + 0x7111711c, 0x71707115, 0x7145714c, 0x77ff7153, 0x77f777fd, 0x77c077f5, 0x77dd77df, 0x77d577d7, + 0x7730773c, 0x7703770c, 0x77107704, 0x777f7714, 0x7777777d, 0x77407775, 0x775d775f, 0x77557757, + 0x74f174f0, 0x74c374cc, 0x74d074c1, 0x7433743c, 0x74347431, 0x740d740f, 0x74057400, 0x7413741c, + 0x74417470, 0x74507444, 0x75fd75ff, 0x75f575f7, 0x75df75c0, 0x75d775dd, 0x753075d5, 0x7503750c, + 0x757f7501, 0x7577757d, 0x75407575, 0x755d755f, 0x75557557, 0x4fcc4ff0, 0x4fc74fc1, 0x4fd04fc4, + 0x4f314f3c, 0x4f004f34, 0x4f054f07, 0x4f154f14, 0x4f4c4f70, 0x4f414f43, 0x4f504f44, 0x4cf34cfc, + 0x4cf44cf1, 0x4cc04ccf, 0x4cc54cc7, 0x4cd34cdc, 0x4cd44cd1, 0x4c304c3f, 0x4c0c4c0f, 0x4c004c03, + 0x4c044c01, 0x4c104c1d, 0x4c714c73, 0x4c404c4d, 0x4c5c4c47, 0x4c514c53, 0x4df04c54, 0x4dc34dcc, + 0x4dd04dc4, 0x4d314d33, 0x4d0f4d34, 0x4d004d0d, 0x4d114d07, 0x4d704d14, 0x4d414d43, 0x43fc4d54, + 0x43f143f3, 0x43c043cf, 0x43d143c7, 0x4335433f, 0x4303430c, 0x43014300, 0x43044307, 0x431c431f, + 0x4310431d, 0x43714373, 0x4343434d, 0x43474340, 0x4354435c, 0x40f040ff, 0x40f540f7, 0x40cc40cf, + 0x40c040c3, 0x40c440c1, 0x40d040dc, 0x40d540d4, 0x4033403c, 0x40314030, 0x400f4034, 0x400d400c, + 0x40004003, 0x40074001, 0x40054004, 0x4013401c, 0x40114010, 0x407c4014, 0x40774070, 0x404d404c, + 0x40404043, 0x40444041, 0x405f4045, 0x4050405d, 0x40554057, 0x41f341fc, 0x41c041cf, 0x41df41c4, + 0x41d441d1, 0x41374130, 0x410c4134, 0x4100410d, 0x41044101, 0x41174110, 0x4173417d, 0x41754174, + 0x4143414d, 0x41534140, 0x41544151, 0x47c147f0, 0x47d047c4, 0x4731473c, 0x470d470f, 0x47014700, + 0x47134705, 0x47704710, 0x4741474c, 0x47504744, 0x44f144f3, 0x44cf44f4, 0x44c044cd, 0x44c544c7, + 0x44dc44df, 0x44d144d3, 0x443d443f, 0x44374430, 0x440c4435, 0x44004403, 0x44044401, 0x4410441d, + 0x44154411, 0x4473447c, 0x444d444f, 0x44454440, 0x4451445c, 0x45c045f0, 0x453345d0, 0x45344531, + 0x4500450f, 0x451c4507, 0x454c4570, 0x45404543, 0x5fff4541, 0x5ff75ffd, 0x5fc05ff5, 0x5fdd5fdf, + 0x5fd55fd7, 0x5f0c5f30, 0x5f015f03, 0x5f7f5f04, 0x5f775f7d, 0x5f405f75, 0x5f5d5f5f, 0x5f555f57, + 0x5cf45cf0, 0x5cc35ccc, 0x5cc45cc1, 0x5c315cc5, 0x5c0c5c34, 0x5c075c00, 0x5c1c5c05, 0x5c705c13, + 0x5c4d5c4f, 0x5c445c41, 0x5df75dfd, 0x5dcf5df5, 0x5ddd5dc4, 0x5dd55dd7, 0x5d0c5d30, 0x5d045d01, + 0x5d7f5d10, 0x5d775d7d, 0x5d405d75, 0x5d5d5d5f, 0x5d555d57, 0x53d053c4, 0x5333533c, 0x5303530f, + 0x53075300, 0x531c5305, 0x53115310, 0x53145317, 0x50f15370, 0x50cf50f4, 0x50c050cd, 0x50d150c7, + 0x503d50d4, 0x500c5030, 0x50005003, 0x50045001, 0x50155010, 0x5073507c, 0x50715070, 0x504d5074, + 0x50475040, 0x51cc51f0, 0x51c551c1, 0x51d051dc, 0x51315133, 0x510d5135, 0x51015100, 0x511f5107, + 0x5171511d, 0x5140514f, 0x51445141, 0x5153515c, 0x57ff5151, 0x57f757fd, 0x57df57f5, 0x57d757dd, + 0x570c57d5, 0x57015703, 0x577f5704, 0x5777577d, 0x57405775, 0x575d575f, 0x57555757, 0x54c354f0, + 0x54dc54c4, 0x543c54d0, 0x5400540f, 0x541c5405, 0x54145411, 0x5441544f, 0x55fd55ff, 0x55f555f7, + 0x55dd55df, 0x55d555d7, 0x5503550c, 0x557f5501, 0x5577557d, 0x55405575, 0x555d555f, 0x55555557 +}; + +shared uint16_t iq1s_grid[2048]; + +#define NEEDS_INIT_IQ_SHMEM +void init_iq_shmem(uvec3 wgsize) +{ + // copy the table into shared memory and sync + [[unroll]] for (uint i = 0; i < iq1s_grid_const.length(); i += wgsize.x) { + uint idx = i + gl_LocalInvocationIndex.x; + if (iq1s_grid_const.length() % wgsize.x == 0 || idx < iq1s_grid_const.length()) { + u16vec2 g = unpack16(iq1s_grid_const[idx]); + iq1s_grid[2*idx+0] = g.x; + iq1s_grid[2*idx+1] = g.y; + } + } + barrier(); +} +#endif + #define QUANT_K_IQ2_XXS 256 #define QUANT_R_IQ2_XXS 1 @@ -380,11 +615,14 @@ const uvec2[256] iq2xxs_grid_const = { shared uvec2 iq2xxs_grid[256]; +#define NEEDS_INIT_IQ_SHMEM void init_iq_shmem(uvec3 wgsize) { // copy the table into shared memory and sync - for (uint i = gl_LocalInvocationIndex.x; i < iq2xxs_grid.length(); i += wgsize.x) { - iq2xxs_grid[i] = iq2xxs_grid_const[i]; + [[unroll]] for (uint i = 0; i < iq2xxs_grid.length(); i += wgsize.x) { + if (iq2xxs_grid_const.length() % wgsize.x == 0 || i + gl_LocalInvocationIndex.x < iq2xxs_grid_const.length()) { + iq2xxs_grid[i + gl_LocalInvocationIndex.x] = iq2xxs_grid_const[i + gl_LocalInvocationIndex.x]; + } } barrier(); } @@ -547,11 +785,14 @@ const uvec2 iq2xs_grid_const[512] = { shared uvec2 iq2xs_grid[512]; +#define NEEDS_INIT_IQ_SHMEM void init_iq_shmem(uvec3 wgsize) { // copy the table into shared memory and sync - for (uint i = gl_LocalInvocationIndex.x; i < iq2xs_grid.length(); i += wgsize.x) { - iq2xs_grid[i] = iq2xs_grid_const[i]; + [[unroll]] for (uint i = 0; i < iq2xs_grid.length(); i += wgsize.x) { + if (iq2xs_grid.length() % wgsize.x == 0 || i + gl_LocalInvocationIndex.x < iq2xs_grid_const.length()) { + iq2xs_grid[i + gl_LocalInvocationIndex.x] = iq2xs_grid_const[i + gl_LocalInvocationIndex.x]; + } } barrier(); } @@ -573,6 +814,14 @@ struct block_iq2_s uint8_t scales[QUANT_K_IQ2_S/32]; }; +struct block_iq2_s_packed16 +{ + float16_t d; + uint16_t qs[QUANT_K_IQ2_S/8]; + uint16_t qh[QUANT_K_IQ2_S/64]; + uint16_t scales[QUANT_K_IQ2_S/64]; +}; + #if defined(DATA_A_IQ2_S) const uvec2 iq2s_grid_const[1024] = { @@ -836,11 +1085,14 @@ const uvec2 iq2s_grid_const[1024] = { shared uvec2 iq2s_grid[1024]; +#define NEEDS_INIT_IQ_SHMEM void init_iq_shmem(uvec3 wgsize) { // copy the table into shared memory and sync - for (uint i = gl_LocalInvocationIndex.x; i < iq2s_grid.length(); i += wgsize.x) { - iq2s_grid[i] = iq2s_grid_const[i]; + [[unroll]] for (uint i = 0; i < iq2s_grid.length(); i += wgsize.x) { + if (iq2s_grid.length() % wgsize.x == 0 || i + gl_LocalInvocationIndex.x < iq2s_grid_const.length()) { + iq2s_grid[i + gl_LocalInvocationIndex.x] = iq2s_grid_const[i + gl_LocalInvocationIndex.x]; + } } barrier(); } @@ -848,6 +1100,7 @@ void init_iq_shmem(uvec3 wgsize) #define QUANT_K QUANT_K_IQ2_S #define QUANT_R QUANT_R_IQ2_S #define A_TYPE block_iq2_s +#define A_TYPE_PACKED16 block_iq2_s_packed16 #endif #define QUANT_K_IQ3_XXS 256 @@ -904,11 +1157,14 @@ const uint32_t iq3xxs_grid_const[256] = { shared uint32_t iq3xxs_grid[256]; +#define NEEDS_INIT_IQ_SHMEM void init_iq_shmem(uvec3 wgsize) { // copy the table into shared memory and sync - for (uint i = gl_LocalInvocationIndex.x; i < iq3xxs_grid.length(); i += wgsize.x) { - iq3xxs_grid[i] = iq3xxs_grid_const[i]; + [[unroll]] for (uint i = 0; i < iq3xxs_grid.length(); i += wgsize.x) { + if (iq3xxs_grid.length() % wgsize.x == 0 || i + gl_LocalInvocationIndex.x < iq3xxs_grid.length()) { + iq3xxs_grid[i + gl_LocalInvocationIndex.x] = iq3xxs_grid_const[i + gl_LocalInvocationIndex.x]; + } } barrier(); } @@ -1011,11 +1267,14 @@ const uint32_t iq3s_grid_const[512] = { shared uint32_t iq3s_grid[512]; +#define NEEDS_INIT_IQ_SHMEM void init_iq_shmem(uvec3 wgsize) { // copy the table into shared memory and sync - for (uint i = gl_LocalInvocationIndex.x; i < iq3s_grid.length(); i += wgsize.x) { - iq3s_grid[i] = iq3s_grid_const[i]; + [[unroll]] for (uint i = 0; i < iq3s_grid.length(); i += wgsize.x) { + if (iq3s_grid.length() % wgsize.x == 0 || i + gl_LocalInvocationIndex.x < iq3s_grid.length()) { + iq3s_grid[i + gl_LocalInvocationIndex.x] = iq3s_grid_const[i + gl_LocalInvocationIndex.x]; + } } barrier(); } @@ -1073,6 +1332,7 @@ const int8_t kvalues_iq4nl_const[16] = { shared FLOAT_TYPE kvalues_iq4nl[16]; +#define NEEDS_INIT_IQ_SHMEM void init_iq_shmem(uvec3 wgsize) { // copy the table into shared memory and sync diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp b/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp index 77e7e1148..cf74625cc 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp @@ -55,6 +55,8 @@ const std::vector type_names = { "q4_k", "q5_k", "q6_k", + "iq1_s", + "iq1_m", "iq2_xxs", "iq2_xs", "iq2_s", @@ -182,6 +184,13 @@ std::string to_uppercase(const std::string& input) { return result; } +bool string_starts_with(const std::string& str, const std::string& prefix) { + if (prefix.size() > str.size()) { + return false; + } + return std::equal(prefix.begin(), prefix.end(), str.begin()); +} + bool string_ends_with(const std::string& str, const std::string& suffix) { if (suffix.size() > str.size()) { return false; @@ -286,7 +295,10 @@ void matmul_shaders(bool fp16, bool matmul_id, bool coopmat, bool coopmat2, bool std::string aligned_b_type_f32 = coopmat2 ? "float" : fp16 ? "mat2x4" : "vec4"; std::string aligned_b_type_f16 = coopmat2 ? "float16_t" : fp16 ? "f16mat2x4" : "f16vec4"; - std::map base_dict = {{"FLOAT_TYPE", (coopmat2 || fp16) ? "float16_t" : "float"}}; + std::map base_dict = { + {"FLOAT_TYPE", (coopmat2 || fp16) ? "float16_t" : "float"}, + {"FLOAT_TYPE_VEC2", (coopmat2 || fp16) ? "f16vec2" : "vec2"}, + }; std::string shader_name = "matmul"; if (matmul_id) { @@ -304,9 +316,7 @@ void matmul_shaders(bool fp16, bool matmul_id, bool coopmat, bool coopmat2, bool base_dict["COOPMAT"] = "1"; } - base_dict["ACC_TYPE"] = f16acc ? "float16_t" : "float"; - - std::string source_name = coopmat2 ? "mul_mm_cm2.comp" : "mul_mm.comp"; + const std::string source_name = coopmat2 ? "mul_mm_cm2.comp" : "mul_mm.comp"; // Shaders with f16 B_TYPE string_to_spv(shader_name + "_f32_f16", source_name, merge_maps(base_dict, {{"DATA_A_F32", "1"}, {"B_TYPE", "float16_t"}, {"D_TYPE", "float"}, }), fp16, coopmat, coopmat2, f16acc); @@ -316,22 +326,34 @@ void matmul_shaders(bool fp16, bool matmul_id, bool coopmat, bool coopmat2, bool string_to_spv(shader_name + "_f16", source_name, merge_maps(base_dict, {{"DATA_A_F16", "1"}, {"B_TYPE", "float16_t"}, {"D_TYPE", "float"}}), fp16, coopmat, coopmat2, f16acc); for (const auto& tname : type_names) { + std::string load_vec_quant = "2"; + if ((tname == "q4_0") || (tname == "q4_1")) + load_vec_quant = "8"; + else if ((tname == "q5_0") || (tname == "q5_1") || (tname == "q8_0") || (tname == "iq4_nl")) + load_vec_quant = "4"; + std::string data_a_key = "DATA_A_" + to_uppercase(tname); // For unaligned, load one at a time for f32/f16, or two at a time for quants - std::string load_vec_a_unaligned = (coopmat2 || tname == "f32" || tname == "f16") ? "1" : "2"; + std::string load_vec_a_unaligned = (coopmat2 || tname == "f32" || tname == "f16") ? "1" : load_vec_quant; // For aligned matmul loads - std::string load_vec_a = (coopmat2 || tname == "f32" || tname == "f16") ? load_vec : "2"; + std::string load_vec_a = (coopmat2 || tname == "f32" || tname == "f16") ? load_vec : load_vec_quant; // don't generate f32 variants for coopmat2 if (!coopmat2) { - string_to_spv(shader_name + "_" + tname + "_f32", source_name, merge_maps(base_dict, {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a_unaligned}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"B_IS_FLOAT", "1"}}), fp16, coopmat, coopmat2, f16acc); - string_to_spv(shader_name + "_" + tname + "_f32_aligned", source_name, merge_maps(base_dict, {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f32}, {"D_TYPE", "float"}, {"B_IS_FLOAT", "1"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc); + string_to_spv(shader_name + "_" + tname + "_f32", source_name, merge_maps(base_dict, {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a_unaligned}, {"B_TYPE", "float"}, {"D_TYPE", "float"}}), fp16, coopmat, coopmat2, f16acc); + string_to_spv(shader_name + "_" + tname + "_f32_aligned", source_name, merge_maps(base_dict, {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f32}, {"D_TYPE", "float"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc); } if (tname != "f16" && tname != "f32") { - string_to_spv(shader_name + "_" + tname + "_f16", source_name, merge_maps(base_dict, {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a_unaligned}, {"B_TYPE", "float16_t"}, {"D_TYPE", "float"}, {"B_IS_FLOAT", "1"}}), fp16, coopmat, coopmat2, f16acc); - string_to_spv(shader_name + "_" + tname + "_f16_aligned", source_name, merge_maps(base_dict, {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f16}, {"D_TYPE", "float"}, {"B_IS_FLOAT", "1"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc); + string_to_spv(shader_name + "_" + tname + "_f16", source_name, merge_maps(base_dict, {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a_unaligned}, {"B_TYPE", "float16_t"}, {"D_TYPE", "float"}}), fp16, coopmat, coopmat2, f16acc); + string_to_spv(shader_name + "_" + tname + "_f16_aligned", source_name, merge_maps(base_dict, {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f16}, {"D_TYPE", "float"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc); } + +#if defined(GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT) + if (!coopmat && !coopmat2 && !matmul_id && (tname == "q4_0" || tname == "q4_1" || tname == "q5_0" || tname == "q5_1" || tname == "q8_0")) { + string_to_spv(shader_name + "_" + tname + "_q8_1", "mul_mmq.comp", merge_maps(base_dict, {{data_a_key, "1"}, {"D_TYPE", "float"},}), fp16, coopmat, coopmat2, f16acc); + } +#endif } } @@ -387,7 +409,7 @@ void process_shaders() { for (const auto& tname : type_names) { // mul mat vec std::string data_a_key = "DATA_A_" + to_uppercase(tname); - std::string shader = (string_ends_with(tname, "_k")) ? "mul_mat_vec_" + tname + ".comp" : "mul_mat_vec.comp"; + std::string shader = (string_ends_with(tname, "_k") || string_starts_with(tname, "iq1_") || string_starts_with(tname, "iq2_") || string_starts_with(tname, "iq3_")) ? "mul_mat_vec_" + tname + ".comp" : "mul_mat_vec.comp"; string_to_spv("mul_mat_vec_" + tname + "_f32_f32", shader, merge_maps(base_dict, {{data_a_key, "1"}, {"B_TYPE", "float"}, {"B_TYPE_VEC2", "vec2"}, {"B_TYPE_VEC4", "vec4"}, {"D_TYPE", "float"}})); string_to_spv("mul_mat_vec_" + tname + "_f16_f32", shader, merge_maps(base_dict, {{data_a_key, "1"}, {"B_TYPE", "float16_t"}, {"B_TYPE_VEC2", "f16vec2"}, {"B_TYPE_VEC4", "f16vec4"}, {"D_TYPE", "float"}})); @@ -411,13 +433,16 @@ void process_shaders() { } } - string_to_spv("mul_mat_vec_p021_f16_f32", "mul_mat_vec_p021.comp", {{"A_TYPE", "float16_t"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}}); - string_to_spv("mul_mat_vec_nc_f16_f32", "mul_mat_vec_nc.comp", {{"A_TYPE", "float16_t"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}}); + string_to_spv("mul_mat_vec_p021_f16_f32_subgroup_add", "mul_mat_vec_p021.comp", {{"A_TYPE", "float16_t"}, {"A_TYPE_VEC4", "f16vec4"}, {"B_TYPE", "float"}, {"B_TYPE_VEC4", "vec4"}, {"D_TYPE", "float"}, {"USE_SUBGROUP_ADD", "1"}}); + string_to_spv("mul_mat_vec_p021_f16_f32", "mul_mat_vec_p021.comp", {{"A_TYPE", "float16_t"}, {"A_TYPE_VEC4", "f16vec4"}, {"B_TYPE", "float"}, {"B_TYPE_VEC4", "vec4"}, {"D_TYPE", "float"}}); + string_to_spv("mul_mat_vec_nc_f16_f32", "mul_mat_vec_nc.comp", {{"A_TYPE", "float16_t"}, {"A_TYPE_VEC4", "f16vec4"}, {"B_TYPE", "float"}, {"B_TYPE_VEC4", "vec4"}, {"D_TYPE", "float"}}); // Norms string_to_spv("norm_f32", "norm.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}})); string_to_spv("group_norm_f32", "group_norm.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}})); string_to_spv("rms_norm_f32", "rms_norm.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}})); + string_to_spv("rms_norm_back_f32", "rms_norm_back.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}})); + string_to_spv("l2_norm_f32", "l2_norm.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}})); string_to_spv("cpy_f32_f32", "copy.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}}); string_to_spv("cpy_f32_f16", "copy.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float16_t"}}); @@ -428,21 +453,27 @@ void process_shaders() { for (std::string t : {"q4_0", "q4_1", "q5_0", "q5_1", "q8_0", "iq4_nl"}) { string_to_spv("cpy_f32_" + t, "copy_to_quant.comp", {{"DATA_A_" + to_uppercase(t), "1"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}}); + string_to_spv("cpy_f32_" + t + "_rte", "copy_to_quant.comp", {{"DATA_A_" + to_uppercase(t), "1"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}, {"RTE16", "1"}}); string_to_spv("cpy_" + t + "_f32", "copy_from_quant.comp", {{"DATA_A_" + to_uppercase(t), "1"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}}); } string_to_spv("add_f32", "add.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}}); string_to_spv("add_f16_f32_f16", "add.comp", {{"A_TYPE", "float16_t"}, {"B_TYPE", "float"}, {"D_TYPE", "float16_t"}, {"FLOAT_TYPE", "float"}}); + string_to_spv("sub_f32", "sub.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}}); + string_to_spv("acc_f32", "acc.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}}); string_to_spv("split_k_reduce", "mul_mat_split_k_reduce.comp", {}); + string_to_spv("fa_split_k_reduce", "flash_attn_split_k_reduce.comp", {}); + string_to_spv("quantize_q8_1", "quantize_q8_1.comp", {}); string_to_spv("mul_f32", "mul.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}}); string_to_spv("div_f32", "div.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}}); string_to_spv("repeat_f32", "repeat.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}}); + string_to_spv("repeat_back_f32", "repeat_back.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}}); string_to_spv("scale_f32", "scale.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}}); @@ -465,14 +496,17 @@ void process_shaders() { string_to_spv("gelu_f32", "gelu.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}}); string_to_spv("gelu_quick_f32", "gelu_quick.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}}); string_to_spv("silu_f32", "silu.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}}); + string_to_spv("silu_back_f32", "silu_back.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}}); string_to_spv("relu_f32", "relu.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}}); string_to_spv("leaky_relu_f32", "leaky_relu.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}}); string_to_spv("tanh_f32", "tanh.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}}); + string_to_spv("sigmoid_f32", "sigmoid.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}}); string_to_spv("diag_mask_inf_f32", "diag_mask_inf.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}}); string_to_spv("soft_max_f32", "soft_max.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}})); string_to_spv("soft_max_f32_f16", "soft_max.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float16_t"}, {"D_TYPE", "float"}})); + string_to_spv("soft_max_back_f32", "soft_max_back.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}})); string_to_spv("rope_norm_f32", "rope_norm.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}}); string_to_spv("rope_norm_f16", "rope_norm.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}}); @@ -482,9 +516,19 @@ void process_shaders() { string_to_spv("rope_neox_f16", "rope_neox.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}}); string_to_spv("rope_neox_f16_rte", "rope_neox.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"RTE16", "1"}}); + string_to_spv("rope_multi_f32", "rope_multi.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}}); + string_to_spv("rope_multi_f16", "rope_multi.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}}); + string_to_spv("rope_multi_f16_rte", "rope_multi.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"RTE16", "1"}}); + + string_to_spv("rope_vision_f32", "rope_vision.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}}); + string_to_spv("rope_vision_f16", "rope_vision.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}}); + string_to_spv("rope_vision_f16_rte", "rope_vision.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"RTE16", "1"}}); + string_to_spv("argsort_f32", "argsort.comp", {{"A_TYPE", "float"}}); + string_to_spv("argmax_f32", "argmax.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "int"}})); string_to_spv("sum_rows_f32", "sum_rows.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}})); + string_to_spv("count_equal_i32", "count_equal.comp", merge_maps(base_dict, {{"A_TYPE", "int"}, {"B_TYPE", "int"}, {"D_TYPE", "int"}})); string_to_spv("im2col_f32", "im2col.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}})); string_to_spv("im2col_f32_f16", "im2col.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float16_t"}})); @@ -496,6 +540,10 @@ void process_shaders() { string_to_spv("rwkv_wkv6_f32", "wkv6.comp", merge_maps(base_dict, {{"A_TYPE", "float"}})); + string_to_spv("rwkv_wkv7_f32", "wkv7.comp", merge_maps(base_dict, {{"A_TYPE", "float"}})); + + string_to_spv("opt_step_adamw_f32", "opt_step_adamw.comp", merge_maps(base_dict, {{"A_TYPE", "float"}})); + for (auto &c : compiles) { c.wait(); } diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/wkv7.comp b/ggml/src/ggml-vulkan/vulkan-shaders/wkv7.comp new file mode 100644 index 000000000..88c1c02b3 --- /dev/null +++ b/ggml/src/ggml-vulkan/vulkan-shaders/wkv7.comp @@ -0,0 +1,91 @@ +#version 450 + +#extension GL_EXT_control_flow_attributes : require + +#define BLOCK_SIZE 64 +layout(local_size_x = BLOCK_SIZE, local_size_y = 1, local_size_z = 1) in; + +layout(push_constant) uniform Parameters { + uint B; + uint T; + uint C; + uint H; +}; + +layout(binding = 0) readonly buffer RBuf { A_TYPE r[]; }; +layout(binding = 1) readonly buffer WBuf { A_TYPE w[]; }; +layout(binding = 2) readonly buffer KBuf { A_TYPE k[]; }; +layout(binding = 3) readonly buffer VBuf { A_TYPE v[]; }; +layout(binding = 4) readonly buffer ABuf { A_TYPE a[]; }; +layout(binding = 5) readonly buffer BBuf { A_TYPE b[]; }; +layout(binding = 6) readonly buffer StateBuf { A_TYPE state_in[]; }; +layout(binding = 7) buffer DstBuf { A_TYPE dst[]; }; + +shared A_TYPE _r[BLOCK_SIZE], _w[BLOCK_SIZE], _k[BLOCK_SIZE], _a[BLOCK_SIZE], _b[BLOCK_SIZE]; + +void main() { + const uint head_size = BLOCK_SIZE; + const uint batch_id = gl_WorkGroupID.x / H; + const uint head_id = gl_WorkGroupID.x % H; + const uint tid = gl_LocalInvocationID.x; + + const uint state_size = C * head_size; + const uint n_seq_tokens = T / B; + + if (batch_id >= B || head_id >= H) { + return; + } + + A_TYPE state[BLOCK_SIZE]; + [[unroll]] for (uint i = 0; i < head_size; i++) { + state[i] = state_in[batch_id * state_size + head_id * head_size * head_size + + tid * head_size + i]; + } + + const uint start_t = batch_id * n_seq_tokens * C + head_id * head_size + tid; + const uint end_t = (batch_id + 1) * n_seq_tokens * C + head_id * head_size + tid; + + for (uint t = start_t; t < end_t; t += C) { + barrier(); + _r[tid] = r[t]; + _w[tid] = w[t]; + _k[tid] = k[t]; + _a[tid] = a[t]; + _b[tid] = b[t]; + barrier(); + + A_TYPE sa = 0.0; + [[unroll]] for (uint j = 0; j < head_size; j += 4) { + vec4 s_vec = vec4(state[j], state[j+1], state[j+2], state[j+3]); + vec4 a_vec = vec4(_a[j], _a[j+1], _a[j+2], _a[j+3]); + sa += dot(s_vec, a_vec); + } + + const A_TYPE v_val = v[t]; + A_TYPE y = 0.0; + + [[unroll]] for (uint j = 0; j < head_size; j += 4) { + vec4 r_vec = vec4(_r[j], _r[j+1], _r[j+2], _r[j+3]); + vec4 w_vec = vec4(_w[j], _w[j+1], _w[j+2], _w[j+3]); + vec4 k_vec = vec4(_k[j], _k[j+1], _k[j+2], _k[j+3]); + vec4 b_vec = vec4(_b[j], _b[j+1], _b[j+2], _b[j+3]); + vec4 s_vec = vec4(state[j], state[j+1], state[j+2], state[j+3]); + + vec4 kv = k_vec * v_val; + s_vec = s_vec * w_vec + kv + sa * b_vec; + y += dot(r_vec, s_vec); + + state[j] = s_vec.x; + state[j+1] = s_vec.y; + state[j+2] = s_vec.z; + state[j+3] = s_vec.w; + } + + dst[t] = y; + } + + [[unroll]] for (uint i = 0; i < head_size; i++) { + dst[T * C + batch_id * state_size + head_id * head_size * head_size + + tid * head_size + i] = state[i]; + } +} diff --git a/ggml/src/ggml.c b/ggml/src/ggml.c index 3b4861542..950772c75 100644 --- a/ggml/src/ggml.c +++ b/ggml/src/ggml.c @@ -240,7 +240,11 @@ void ggml_log_callback_default(enum ggml_log_level level, const char * text, voi void * ggml_aligned_malloc(size_t size) { +#if defined(__s390x__) + const int alignment = 256; +#else const int alignment = 64; +#endif #if defined(_MSC_VER) || defined(__MINGW32__) return _aligned_malloc(size, alignment); @@ -561,9 +565,9 @@ FILE * ggml_fopen(const char * fname, const char * mode) { #endif } -static void ggml_vec_dot_f32(int n, float * restrict s, size_t bs, const float * restrict x, size_t bx, const float * restrict y, size_t by, int nrc); -static void ggml_vec_dot_f16(int n, float * restrict s, size_t bs, ggml_fp16_t * restrict x, size_t bx, ggml_fp16_t * restrict y, size_t by, int nrc); -static void ggml_vec_dot_bf16(int n, float * restrict s, size_t bs, ggml_bf16_t * restrict x, size_t bx, ggml_bf16_t * restrict y, size_t by, int nrc); +static void ggml_vec_dot_f32(int n, float * GGML_RESTRICT s, size_t bs, const float * GGML_RESTRICT x, size_t bx, const float * GGML_RESTRICT y, size_t by, int nrc); +static void ggml_vec_dot_f16(int n, float * GGML_RESTRICT s, size_t bs, ggml_fp16_t * GGML_RESTRICT x, size_t bx, ggml_fp16_t * GGML_RESTRICT y, size_t by, int nrc); +static void ggml_vec_dot_bf16(int n, float * GGML_RESTRICT s, size_t bs, ggml_bf16_t * GGML_RESTRICT x, size_t bx, ggml_bf16_t * GGML_RESTRICT y, size_t by, int nrc); static const struct ggml_type_traits type_traits[GGML_TYPE_COUNT] = { [GGML_TYPE_I8] = { @@ -925,6 +929,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = { "RMS_NORM", "RMS_NORM_BACK", "GROUP_NORM", + "L2_NORM", "MUL_MAT", "MUL_MAT_ID", @@ -973,26 +978,22 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = { "ADD_REL_POS", "RWKV_WKV6", "GATED_LINEAR_ATTN", + "RWKV_WKV7", "UNARY", - "MAP_UNARY", - "MAP_BINARY", - - "MAP_CUSTOM1_F32", - "MAP_CUSTOM2_F32", - "MAP_CUSTOM3_F32", - "MAP_CUSTOM1", "MAP_CUSTOM2", "MAP_CUSTOM3", + "CUSTOM", + "CROSS_ENTROPY_LOSS", "CROSS_ENTROPY_LOSS_BACK", "OPT_STEP_ADAMW", }; -static_assert(GGML_OP_COUNT == 83, "GGML_OP_COUNT != 83"); +static_assert(GGML_OP_COUNT == 81, "GGML_OP_COUNT != 81"); static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = { "none", @@ -1022,6 +1023,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = { "rms_norm(x)", "rms_norm_back(x)", "group_norm(x)", + "l2_norm(x)", "X*Y", "X[i]*Y", @@ -1070,26 +1072,22 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = { "add_rel_pos(x)", "rwkv_wkv6(k, v, r, tf, td, s)", "gated_linear_attn(k, v, q, gate, s)", + "rwkv_wkv7(r, w, k, v, a, b, s)", "unary(x)", - "f(x)", - "f(x,y)", - - "custom_f32(x)", - "custom_f32(x,y)", - "custom_f32(x,y,z)", + "map_custom(x)", + "map_custom(x,y)", + "map_custom(x,y,z)", "custom(x)", - "custom(x,y)", - "custom(x,y,z)", "cross_entropy_loss(x,y)", "cross_entropy_loss_back(x,y)", "adamw(x)", }; -static_assert(GGML_OP_COUNT == 83, "GGML_OP_COUNT != 83"); +static_assert(GGML_OP_COUNT == 81, "GGML_OP_COUNT != 81"); static_assert(GGML_OP_POOL_COUNT == 2, "GGML_OP_POOL_COUNT != 2"); @@ -1151,6 +1149,12 @@ int64_t ggml_nrows(const struct ggml_tensor * tensor) { } size_t ggml_nbytes(const struct ggml_tensor * tensor) { + for (int i = 0; i < GGML_MAX_DIMS; ++i) { + if (tensor->ne[i] <= 0) { + return 0; + } + } + size_t nbytes; const size_t blck_size = ggml_blck_size(tensor->type); if (blck_size == 1) { @@ -1379,7 +1383,7 @@ bool ggml_are_same_stride(const struct ggml_tensor * t0, const struct ggml_tenso (t0->nb[3] == t1->nb[3]); } -// check if t1 can be represented as a repeatition of t0 +// check if t1 can be represented as a repetition of t0 bool ggml_can_repeat(const struct ggml_tensor * t0, const struct ggml_tensor * t1) { static_assert(GGML_MAX_DIMS == 4, "GGML_MAX_DIMS is not 4 - update this function"); @@ -2328,6 +2332,7 @@ struct ggml_tensor * ggml_concat( struct ggml_tensor * b, int dim) { GGML_ASSERT(dim >= 0 && dim < GGML_MAX_DIMS); + GGML_ASSERT(a->type == b->type); int64_t ne[GGML_MAX_DIMS]; for (int d = 0; d < GGML_MAX_DIMS; ++d) { @@ -2681,6 +2686,37 @@ struct ggml_tensor * ggml_group_norm_inplace( return ggml_group_norm_impl(ctx, a, n_groups, eps, true); } +// ggml_l2_norm + +static struct ggml_tensor * ggml_l2_norm_impl( + struct ggml_context * ctx, + struct ggml_tensor * a, + float eps, + bool inplace) { + struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); + + ggml_set_op_params_f32(result, 0, eps); + + result->op = GGML_OP_L2_NORM; + result->src[0] = a; + + return result; +} + +struct ggml_tensor * ggml_l2_norm( + struct ggml_context * ctx, + struct ggml_tensor * a, + float eps) { + return ggml_l2_norm_impl(ctx, a, eps, false); +} + +struct ggml_tensor * ggml_l2_norm_inplace( + struct ggml_context * ctx, + struct ggml_tensor * a, + float eps) { + return ggml_l2_norm_impl(ctx, a, eps, true); +} + // ggml_mul_mat static inline bool ggml_can_mul_mat(const struct ggml_tensor * t0, const struct ggml_tensor * t1) { @@ -4138,7 +4174,8 @@ static struct ggml_tensor * ggml_upscale_impl( int ne0, int ne1, int ne2, - int ne3) { + int ne3, + enum ggml_scale_mode mode) { GGML_ASSERT(a->ne[0] <= ne0); GGML_ASSERT(a->ne[1] <= ne1); GGML_ASSERT(a->ne[2] <= ne2); @@ -4146,6 +4183,8 @@ static struct ggml_tensor * ggml_upscale_impl( struct ggml_tensor * result = ggml_new_tensor_4d(ctx, a->type, ne0, ne1, ne2, ne3); + ggml_set_op_params_i32(result, 0, mode); + result->op = GGML_OP_UPSCALE; result->src[0] = a; @@ -4155,8 +4194,9 @@ static struct ggml_tensor * ggml_upscale_impl( struct ggml_tensor * ggml_upscale( struct ggml_context * ctx, struct ggml_tensor * a, - int scale_factor) { - return ggml_upscale_impl(ctx, a, a->ne[0] * scale_factor, a->ne[1] * scale_factor, a->ne[2], a->ne[3]); + int scale_factor, + enum ggml_scale_mode mode) { + return ggml_upscale_impl(ctx, a, a->ne[0] * scale_factor, a->ne[1] * scale_factor, a->ne[2], a->ne[3], mode); } struct ggml_tensor * ggml_upscale_ext( @@ -4165,8 +4205,9 @@ struct ggml_tensor * ggml_upscale_ext( int ne0, int ne1, int ne2, - int ne3) { - return ggml_upscale_impl(ctx, a, ne0, ne1, ne2, ne3); + int ne3, + enum ggml_scale_mode mode) { + return ggml_upscale_impl(ctx, a, ne0, ne1, ne2, ne3, mode); } // ggml_pad @@ -4329,7 +4370,7 @@ struct ggml_tensor * ggml_flash_attn_ext( } // permute(0, 2, 1, 3) - int64_t ne[4] = { q->ne[0], q->ne[2], q->ne[1], q->ne[3] }; + int64_t ne[4] = { v->ne[0], q->ne[2], q->ne[1], q->ne[3] }; struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne); float params[] = { scale, max_bias, logit_softcap }; @@ -4715,6 +4756,54 @@ struct ggml_tensor * ggml_gated_linear_attn( return result; } +// ggml_rwkv_wkv7 + +struct ggml_tensor * ggml_rwkv_wkv7( + struct ggml_context * ctx, + struct ggml_tensor * r, + struct ggml_tensor * w, + struct ggml_tensor * k, + struct ggml_tensor * v, + struct ggml_tensor * a, + struct ggml_tensor * b, + struct ggml_tensor * state) { + GGML_ASSERT(ggml_is_contiguous(r)); + GGML_ASSERT(ggml_is_contiguous(w)); + GGML_ASSERT(ggml_is_contiguous(k)); + GGML_ASSERT(ggml_is_contiguous(v)); + GGML_ASSERT(ggml_is_contiguous(a)); + GGML_ASSERT(ggml_is_contiguous(b)); + GGML_ASSERT(ggml_is_contiguous(state)); + + const int64_t S = k->ne[0]; + const int64_t H = k->ne[1]; + const int64_t n_tokens = k->ne[2]; + const int64_t n_seqs = state->ne[1]; + { + GGML_ASSERT(w->ne[0] == S && w->ne[1] == H && w->ne[2] == n_tokens); + GGML_ASSERT(k->ne[0] == S && k->ne[1] == H && k->ne[2] == n_tokens); + GGML_ASSERT(v->ne[0] == S && v->ne[1] == H && v->ne[2] == n_tokens); + GGML_ASSERT(a->ne[0] == S && a->ne[1] == H && a->ne[2] == n_tokens); + GGML_ASSERT(b->ne[0] == S && b->ne[1] == H && b->ne[2] == n_tokens); + GGML_ASSERT(ggml_nelements(state) == S * S * H * n_seqs); + } + + // concat output and new_state + const int64_t ne[4] = { S * H, n_tokens + S * n_seqs, 1, 1 }; + struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne); + + result->op = GGML_OP_RWKV_WKV7; + result->src[0] = r; + result->src[1] = w; + result->src[2] = k; + result->src[3] = v; + result->src[4] = a; + result->src[5] = b; + result->src[6] = state; + + return result; +} + // ggml_unary static struct ggml_tensor * ggml_unary_impl( @@ -4748,179 +4837,6 @@ struct ggml_tensor * ggml_unary_inplace( return ggml_unary_impl(ctx, a, op, true); } -// ggml_map_unary - -static struct ggml_tensor * ggml_map_unary_impl_f32( - struct ggml_context * ctx, - struct ggml_tensor * a, - const ggml_unary_op_f32_t fun, - bool inplace) { - struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); - - ggml_set_op_params(result, (const void *) &fun, sizeof(fun)); - - result->op = GGML_OP_MAP_UNARY; - result->src[0] = a; - - return result; -} - -struct ggml_tensor * ggml_map_unary_f32( - struct ggml_context * ctx, - struct ggml_tensor * a, - const ggml_unary_op_f32_t fun) { - return ggml_map_unary_impl_f32(ctx, a, fun, false); -} - -struct ggml_tensor * ggml_map_unary_inplace_f32( - struct ggml_context * ctx, - struct ggml_tensor * a, - const ggml_unary_op_f32_t fun) { - return ggml_map_unary_impl_f32(ctx, a, fun, true); -} - -// ggml_map_binary - -static struct ggml_tensor * ggml_map_binary_impl_f32( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b, - const ggml_binary_op_f32_t fun, - bool inplace) { - GGML_ASSERT(ggml_are_same_shape(a, b)); - - struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); - - ggml_set_op_params(result, (const void *) &fun, sizeof(fun)); - - result->op = GGML_OP_MAP_BINARY; - result->src[0] = a; - result->src[1] = b; - - return result; -} - -struct ggml_tensor * ggml_map_binary_f32( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b, - const ggml_binary_op_f32_t fun) { - return ggml_map_binary_impl_f32(ctx, a, b, fun, false); -} - -struct ggml_tensor * ggml_map_binary_inplace_f32( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b, - const ggml_binary_op_f32_t fun) { - return ggml_map_binary_impl_f32(ctx, a, b, fun, true); -} - -// ggml_map_custom1_f32 - -static struct ggml_tensor * ggml_map_custom1_impl_f32( - struct ggml_context * ctx, - struct ggml_tensor * a, - const ggml_custom1_op_f32_t fun, - bool inplace) { - struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); - - ggml_set_op_params(result, (const void *) &fun, sizeof(fun)); - - result->op = GGML_OP_MAP_CUSTOM1_F32; - result->src[0] = a; - - return result; -} - -struct ggml_tensor * ggml_map_custom1_f32( - struct ggml_context * ctx, - struct ggml_tensor * a, - const ggml_custom1_op_f32_t fun) { - return ggml_map_custom1_impl_f32(ctx, a, fun, false); -} - -struct ggml_tensor * ggml_map_custom1_inplace_f32( - struct ggml_context * ctx, - struct ggml_tensor * a, - const ggml_custom1_op_f32_t fun) { - return ggml_map_custom1_impl_f32(ctx, a, fun, true); -} - -// ggml_map_custom2_f32 - -static struct ggml_tensor * ggml_map_custom2_impl_f32( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b, - const ggml_custom2_op_f32_t fun, - bool inplace) { - struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); - - ggml_set_op_params(result, (const void *) &fun, sizeof(fun)); - - result->op = GGML_OP_MAP_CUSTOM2_F32; - result->src[0] = a; - result->src[1] = b; - - return result; -} - -struct ggml_tensor * ggml_map_custom2_f32( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b, - const ggml_custom2_op_f32_t fun) { - return ggml_map_custom2_impl_f32(ctx, a, b, fun, false); -} - -struct ggml_tensor * ggml_map_custom2_inplace_f32( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b, - const ggml_custom2_op_f32_t fun) { - return ggml_map_custom2_impl_f32(ctx, a, b, fun, true); -} - -// ggml_map_custom3_f32 - -static struct ggml_tensor * ggml_map_custom3_impl_f32( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b, - struct ggml_tensor * c, - const ggml_custom3_op_f32_t fun, - bool inplace) { - struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); - - ggml_set_op_params(result, (const void *) &fun, sizeof(fun)); - - result->op = GGML_OP_MAP_CUSTOM3_F32; - result->src[0] = a; - result->src[1] = b; - result->src[2] = c; - - return result; -} - -struct ggml_tensor * ggml_map_custom3_f32( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b, - struct ggml_tensor * c, - const ggml_custom3_op_f32_t fun) { - return ggml_map_custom3_impl_f32(ctx, a, b, c, fun, false); -} - -struct ggml_tensor * ggml_map_custom3_inplace_f32( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b, - struct ggml_tensor * c, - const ggml_custom3_op_f32_t fun) { - return ggml_map_custom3_impl_f32(ctx, a, b, c, fun, true); -} - // ggml_map_custom1 static struct ggml_tensor * ggml_map_custom1_impl( @@ -4939,7 +4855,7 @@ static struct ggml_tensor * ggml_map_custom1_impl( /*.n_tasks =*/ n_tasks, /*.userdata =*/ userdata }; - ggml_set_op_params(result, (const void *) ¶ms, sizeof(params)); + ggml_set_op_params(result, ¶ms, sizeof(params)); result->op = GGML_OP_MAP_CUSTOM1; result->src[0] = a; @@ -4984,7 +4900,7 @@ static struct ggml_tensor * ggml_map_custom2_impl( /*.n_tasks =*/ n_tasks, /*.userdata =*/ userdata }; - ggml_set_op_params(result, (const void *) ¶ms, sizeof(params)); + ggml_set_op_params(result, ¶ms, sizeof(params)); result->op = GGML_OP_MAP_CUSTOM2; result->src[0] = a; @@ -5033,7 +4949,7 @@ static struct ggml_tensor * ggml_map_custom3_impl( /*.n_tasks =*/ n_tasks, /*.userdata =*/ userdata }; - ggml_set_op_params(result, (const void *) ¶ms, sizeof(params)); + ggml_set_op_params(result, ¶ms, sizeof(params)); result->op = GGML_OP_MAP_CUSTOM3; result->src[0] = a; @@ -5065,6 +4981,66 @@ struct ggml_tensor * ggml_map_custom3_inplace( return ggml_map_custom3_impl(ctx, a, b, c, fun, n_tasks, userdata, true); } +struct ggml_tensor * ggml_custom_4d( + struct ggml_context * ctx, + enum ggml_type type, + int64_t ne0, + int64_t ne1, + int64_t ne2, + int64_t ne3, + struct ggml_tensor ** args, + int n_args, + ggml_custom_op_t fun, + int n_tasks, + void * userdata) { + + GGML_ASSERT(n_args < GGML_MAX_SRC); + + struct ggml_tensor * result = ggml_new_tensor_4d(ctx, type, ne0, ne1, ne2, ne3); + + struct ggml_custom_op_params params = { + /*.fun =*/ fun, + /*.n_tasks =*/ n_tasks, + /*.userdata =*/ userdata + }; + ggml_set_op_params(result, ¶ms, sizeof(params)); + + result->op = GGML_OP_CUSTOM; + for (int i = 0; i < n_args; i++) { + result->src[i] = args[i]; + } + + return result; +} + +struct ggml_tensor * ggml_custom_inplace( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor ** args, + int n_args, + ggml_custom_op_t fun, + int n_tasks, + void * userdata) { + + GGML_ASSERT(n_args < GGML_MAX_SRC - 1); + + struct ggml_tensor * result = ggml_view_tensor(ctx, a); + + struct ggml_custom_op_params params = { + /*.fun =*/ fun, + /*.n_tasks =*/ n_tasks, + /*.userdata =*/ userdata + }; + ggml_set_op_params(result, ¶ms, sizeof(params)); + + result->op = GGML_OP_CUSTOM; + result->src[0] = a; + for (int i = 0; i < n_args; i++) { + result->src[i + 1] = args[i]; + } + + return result; +} // ggml_cross_entropy_loss struct ggml_tensor * ggml_cross_entropy_loss( diff --git a/ggml/src/gguf.cpp b/ggml/src/gguf.cpp index ab13669c5..381a9c7dc 100644 --- a/ggml/src/gguf.cpp +++ b/ggml/src/gguf.cpp @@ -932,6 +932,7 @@ static void gguf_check_reserved_keys(const std::string & key, const T val) { if constexpr (std::is_same::value) { GGML_ASSERT(val > 0 && (val & (val - 1)) == 0 && GGUF_KEY_GENERAL_ALIGNMENT " must be power of 2"); } else { + GGML_UNUSED(val); GGML_ABORT(GGUF_KEY_GENERAL_ALIGNMENT " must be type u32"); } } diff --git a/gguf-py/README.md b/gguf-py/README.md index 2e513633d..dd4ab7bde 100644 --- a/gguf-py/README.md +++ b/gguf-py/README.md @@ -1,9 +1,9 @@ ## gguf -This is a Python package for writing binary files in the [GGUF](https://github.com/ggerganov/ggml/pull/302) +This is a Python package for writing binary files in the [GGUF](https://github.com/ggml-org/ggml/pull/302) (GGML Universal File) format. -See [convert_hf_to_gguf.py](https://github.com/ggerganov/llama.cpp/blob/master/convert_hf_to_gguf.py) +See [convert_hf_to_gguf.py](https://github.com/ggml-org/llama.cpp/blob/master/convert_hf_to_gguf.py) as an example for its usage. ## Installation @@ -13,17 +13,17 @@ pip install gguf ## API Examples/Simple Tools -[examples/writer.py](https://github.com/ggerganov/llama.cpp/blob/master/gguf-py/examples/writer.py) — Generates `example.gguf` in the current directory to demonstrate generating a GGUF file. Note that this file cannot be used as a model. +[examples/writer.py](https://github.com/ggml-org/llama.cpp/blob/master/gguf-py/examples/writer.py) — Generates `example.gguf` in the current directory to demonstrate generating a GGUF file. Note that this file cannot be used as a model. -[examples/reader.py](https://github.com/ggerganov/llama.cpp/blob/master/gguf-py/examples/reader.py) — Extracts and displays key-value pairs and tensor details from a GGUF file in a readable format. +[examples/reader.py](https://github.com/ggml-org/llama.cpp/blob/master/gguf-py/examples/reader.py) — Extracts and displays key-value pairs and tensor details from a GGUF file in a readable format. -[gguf/scripts/gguf_dump.py](https://github.com/ggerganov/llama.cpp/blob/master/gguf-py/gguf/scripts/gguf_dump.py) — Dumps a GGUF file's metadata to the console. +[gguf/scripts/gguf_dump.py](https://github.com/ggml-org/llama.cpp/blob/master/gguf-py/gguf/scripts/gguf_dump.py) — Dumps a GGUF file's metadata to the console. -[gguf/scripts/gguf_set_metadata.py](https://github.com/ggerganov/llama.cpp/blob/master/gguf-py/gguf/scripts/gguf_set_metadata.py) — Allows changing simple metadata values in a GGUF file by key. +[gguf/scripts/gguf_set_metadata.py](https://github.com/ggml-org/llama.cpp/blob/master/gguf-py/gguf/scripts/gguf_set_metadata.py) — Allows changing simple metadata values in a GGUF file by key. -[gguf/scripts/gguf_convert_endian.py](https://github.com/ggerganov/llama.cpp/blob/master/gguf-py/gguf/scripts/gguf_convert_endian.py) — Allows converting the endianness of GGUF files. +[gguf/scripts/gguf_convert_endian.py](https://github.com/ggml-org/llama.cpp/blob/master/gguf-py/gguf/scripts/gguf_convert_endian.py) — Allows converting the endianness of GGUF files. -[gguf/scripts/gguf_new_metadata.py](https://github.com/ggerganov/llama.cpp/blob/master/gguf-py/gguf/scripts/gguf_new_metadata.py) — Copies a GGUF file with added/modified/removed metadata values. +[gguf/scripts/gguf_new_metadata.py](https://github.com/ggml-org/llama.cpp/blob/master/gguf-py/gguf/scripts/gguf_new_metadata.py) — Copies a GGUF file with added/modified/removed metadata values. ## Development Maintainers who participate in development of this package are advised to install it in editable mode: diff --git a/gguf-py/examples/reader.py b/gguf-py/examples/reader.py index d841048c6..703b782b5 100644 --- a/gguf-py/examples/reader.py +++ b/gguf-py/examples/reader.py @@ -2,12 +2,14 @@ import logging import sys from pathlib import Path -from gguf.gguf_reader import GGUFReader logger = logging.getLogger("reader") +# Necessary to load the local gguf package sys.path.insert(0, str(Path(__file__).parent.parent)) +from gguf.gguf_reader import GGUFReader + def read_gguf_file(gguf_file_path): """ diff --git a/gguf-py/gguf/constants.py b/gguf-py/gguf/constants.py index 0b1f21fc7..5b4b0e9ed 100644 --- a/gguf-py/gguf/constants.py +++ b/gguf-py/gguf/constants.py @@ -116,24 +116,29 @@ class Keys: RESIDUAL_SCALE = "{arch}.residual_scale" EMBEDDING_SCALE = "{arch}.embedding_scale" TOKEN_SHIFT_COUNT = "{arch}.token_shift_count" + INTERLEAVE_MOE_LAYER_STEP = "{arch}.interleave_moe_layer_step" class Attention: - HEAD_COUNT = "{arch}.attention.head_count" - HEAD_COUNT_KV = "{arch}.attention.head_count_kv" - MAX_ALIBI_BIAS = "{arch}.attention.max_alibi_bias" - CLAMP_KQV = "{arch}.attention.clamp_kqv" - KEY_LENGTH = "{arch}.attention.key_length" - VALUE_LENGTH = "{arch}.attention.value_length" - LAYERNORM_EPS = "{arch}.attention.layer_norm_epsilon" - LAYERNORM_RMS_EPS = "{arch}.attention.layer_norm_rms_epsilon" - GROUPNORM_EPS = "{arch}.attention.group_norm_epsilon" - GROUPNORM_GROUPS = "{arch}.attention.group_norm_groups" - CAUSAL = "{arch}.attention.causal" - Q_LORA_RANK = "{arch}.attention.q_lora_rank" - KV_LORA_RANK = "{arch}.attention.kv_lora_rank" - REL_BUCKETS_COUNT = "{arch}.attention.relative_buckets_count" - SLIDING_WINDOW = "{arch}.attention.sliding_window" - SCALE = "{arch}.attention.scale" + HEAD_COUNT = "{arch}.attention.head_count" + HEAD_COUNT_KV = "{arch}.attention.head_count_kv" + MAX_ALIBI_BIAS = "{arch}.attention.max_alibi_bias" + CLAMP_KQV = "{arch}.attention.clamp_kqv" + KEY_LENGTH = "{arch}.attention.key_length" + VALUE_LENGTH = "{arch}.attention.value_length" + LAYERNORM_EPS = "{arch}.attention.layer_norm_epsilon" + LAYERNORM_RMS_EPS = "{arch}.attention.layer_norm_rms_epsilon" + GROUPNORM_EPS = "{arch}.attention.group_norm_epsilon" + GROUPNORM_GROUPS = "{arch}.attention.group_norm_groups" + CAUSAL = "{arch}.attention.causal" + Q_LORA_RANK = "{arch}.attention.q_lora_rank" + KV_LORA_RANK = "{arch}.attention.kv_lora_rank" + DECAY_LORA_RANK = "{arch}.attention.decay_lora_rank" + ICLR_LORA_RANK = "{arch}.attention.iclr_lora_rank" + VALUE_RESIDUAL_MIX_LORA_RANK = "{arch}.attention.value_residual_mix_lora_rank" + GATE_LORA_RANK = "{arch}.attention.gate_lora_rank" + REL_BUCKETS_COUNT = "{arch}.attention.relative_buckets_count" + SLIDING_WINDOW = "{arch}.attention.sliding_window" + SCALE = "{arch}.attention.scale" class Rope: DIMENSION_COUNT = "{arch}.rope.dimension_count" @@ -230,6 +235,7 @@ class GGUFType: class MODEL_ARCH(IntEnum): LLAMA = auto() + LLAMA4 = auto() DECI = auto() FALCON = auto() BAICHUAN = auto() @@ -249,6 +255,8 @@ class MODEL_ARCH(IntEnum): QWEN2 = auto() QWEN2MOE = auto() QWEN2VL = auto() + QWEN3 = auto() + QWEN3MOE = auto() PHI2 = auto() PHI3 = auto() PHIMOE = auto() @@ -260,9 +268,12 @@ class MODEL_ARCH(IntEnum): MINICPM3 = auto() GEMMA = auto() GEMMA2 = auto() + GEMMA3 = auto() STARCODER2 = auto() RWKV6 = auto() RWKV6QWEN2 = auto() + RWKV7 = auto() + ARWKV7 = auto() MAMBA = auto() XVERSE = auto() COMMAND_R = auto() @@ -276,6 +287,7 @@ class MODEL_ARCH(IntEnum): DEEPSEEK = auto() DEEPSEEK2 = auto() CHATGLM = auto() + GLM4 = auto() BITNET = auto() T5 = auto() T5ENCODER = auto() @@ -286,6 +298,8 @@ class MODEL_ARCH(IntEnum): GRANITE_MOE = auto() CHAMELEON = auto() WAVTOKENIZER_DEC = auto() + PLM = auto() + BAILINGMOE = auto() class MODEL_TENSOR(IntEnum): @@ -335,8 +349,20 @@ class MODEL_TENSOR(IntEnum): SSM_A = auto() SSM_D = auto() SSM_OUT = auto() + TIME_MIX_W0 = auto() TIME_MIX_W1 = auto() TIME_MIX_W2 = auto() + TIME_MIX_A0 = auto() + TIME_MIX_A1 = auto() + TIME_MIX_A2 = auto() + TIME_MIX_V0 = auto() + TIME_MIX_V1 = auto() + TIME_MIX_V2 = auto() + TIME_MIX_G1 = auto() + TIME_MIX_G2 = auto() + TIME_MIX_K_K = auto() + TIME_MIX_K_A = auto() + TIME_MIX_R_K = auto() TIME_MIX_LERP_X = auto() TIME_MIX_LERP_K = auto() TIME_MIX_LERP_V = auto() @@ -417,6 +443,7 @@ class MODEL_TENSOR(IntEnum): MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = { MODEL_ARCH.LLAMA: "llama", + MODEL_ARCH.LLAMA4: "llama4", MODEL_ARCH.DECI: "deci", MODEL_ARCH.FALCON: "falcon", MODEL_ARCH.BAICHUAN: "baichuan", @@ -436,6 +463,8 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = { MODEL_ARCH.QWEN2: "qwen2", MODEL_ARCH.QWEN2MOE: "qwen2moe", MODEL_ARCH.QWEN2VL: "qwen2vl", + MODEL_ARCH.QWEN3: "qwen3", + MODEL_ARCH.QWEN3MOE: "qwen3moe", MODEL_ARCH.PHI2: "phi2", MODEL_ARCH.PHI3: "phi3", MODEL_ARCH.PHIMOE: "phimoe", @@ -447,9 +476,12 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = { MODEL_ARCH.MINICPM3: "minicpm3", MODEL_ARCH.GEMMA: "gemma", MODEL_ARCH.GEMMA2: "gemma2", + MODEL_ARCH.GEMMA3: "gemma3", MODEL_ARCH.STARCODER2: "starcoder2", MODEL_ARCH.RWKV6: "rwkv6", MODEL_ARCH.RWKV6QWEN2: "rwkv6qwen2", + MODEL_ARCH.RWKV7: "rwkv7", + MODEL_ARCH.ARWKV7: "arwkv7", MODEL_ARCH.MAMBA: "mamba", MODEL_ARCH.XVERSE: "xverse", MODEL_ARCH.COMMAND_R: "command-r", @@ -463,6 +495,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = { MODEL_ARCH.DEEPSEEK: "deepseek", MODEL_ARCH.DEEPSEEK2: "deepseek2", MODEL_ARCH.CHATGLM: "chatglm", + MODEL_ARCH.GLM4: "glm4", MODEL_ARCH.BITNET: "bitnet", MODEL_ARCH.T5: "t5", MODEL_ARCH.T5ENCODER: "t5encoder", @@ -473,6 +506,8 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = { MODEL_ARCH.GRANITE_MOE: "granitemoe", MODEL_ARCH.CHAMELEON: "chameleon", MODEL_ARCH.WAVTOKENIZER_DEC: "wavtokenizer-dec", + MODEL_ARCH.PLM: "plm", + MODEL_ARCH.BAILINGMOE: "bailingmoe", } TENSOR_NAMES: dict[MODEL_TENSOR, str] = { @@ -522,8 +557,20 @@ TENSOR_NAMES: dict[MODEL_TENSOR, str] = { MODEL_TENSOR.SSM_A: "blk.{bid}.ssm_a", MODEL_TENSOR.SSM_D: "blk.{bid}.ssm_d", MODEL_TENSOR.SSM_OUT: "blk.{bid}.ssm_out", + MODEL_TENSOR.TIME_MIX_W0: "blk.{bid}.time_mix_w0", MODEL_TENSOR.TIME_MIX_W1: "blk.{bid}.time_mix_w1", MODEL_TENSOR.TIME_MIX_W2: "blk.{bid}.time_mix_w2", + MODEL_TENSOR.TIME_MIX_A0: "blk.{bid}.time_mix_a0", + MODEL_TENSOR.TIME_MIX_A1: "blk.{bid}.time_mix_a1", + MODEL_TENSOR.TIME_MIX_A2: "blk.{bid}.time_mix_a2", + MODEL_TENSOR.TIME_MIX_V0: "blk.{bid}.time_mix_v0", + MODEL_TENSOR.TIME_MIX_V1: "blk.{bid}.time_mix_v1", + MODEL_TENSOR.TIME_MIX_V2: "blk.{bid}.time_mix_v2", + MODEL_TENSOR.TIME_MIX_G1: "blk.{bid}.time_mix_g1", + MODEL_TENSOR.TIME_MIX_G2: "blk.{bid}.time_mix_g2", + MODEL_TENSOR.TIME_MIX_K_K: "blk.{bid}.time_mix_k_k", + MODEL_TENSOR.TIME_MIX_K_A: "blk.{bid}.time_mix_k_a", + MODEL_TENSOR.TIME_MIX_R_K: "blk.{bid}.time_mix_r_k", MODEL_TENSOR.TIME_MIX_LERP_X: "blk.{bid}.time_mix_lerp_x", MODEL_TENSOR.TIME_MIX_LERP_K: "blk.{bid}.time_mix_lerp_k", MODEL_TENSOR.TIME_MIX_LERP_V: "blk.{bid}.time_mix_lerp_v", @@ -623,6 +670,29 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = { MODEL_TENSOR.FFN_DOWN_EXP, MODEL_TENSOR.FFN_UP_EXP, ], + MODEL_ARCH.LLAMA4: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ROPE_FREQS, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_Q, + MODEL_TENSOR.ATTN_K, + MODEL_TENSOR.ATTN_V, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.ATTN_ROT_EMBD, + MODEL_TENSOR.FFN_GATE_INP, + MODEL_TENSOR.FFN_NORM, + MODEL_TENSOR.FFN_GATE, + MODEL_TENSOR.FFN_DOWN, + MODEL_TENSOR.FFN_UP, + MODEL_TENSOR.FFN_GATE_EXP, + MODEL_TENSOR.FFN_DOWN_EXP, + MODEL_TENSOR.FFN_UP_EXP, + MODEL_TENSOR.FFN_GATE_SHEXP, + MODEL_TENSOR.FFN_DOWN_SHEXP, + MODEL_TENSOR.FFN_UP_SHEXP, + ], MODEL_ARCH.DECI: [ MODEL_TENSOR.TOKEN_EMBD, MODEL_TENSOR.OUTPUT_NORM, @@ -896,6 +966,40 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = { MODEL_TENSOR.FFN_DOWN_SHEXP, MODEL_TENSOR.FFN_UP_SHEXP, ], + MODEL_ARCH.QWEN3: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ROPE_FREQS, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_Q, + MODEL_TENSOR.ATTN_Q_NORM, + MODEL_TENSOR.ATTN_K, + MODEL_TENSOR.ATTN_K_NORM, + MODEL_TENSOR.ATTN_V, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.FFN_NORM, + MODEL_TENSOR.FFN_GATE, + MODEL_TENSOR.FFN_DOWN, + MODEL_TENSOR.FFN_UP, + ], + MODEL_ARCH.QWEN3MOE: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_Q, + MODEL_TENSOR.ATTN_Q_NORM, + MODEL_TENSOR.ATTN_K, + MODEL_TENSOR.ATTN_K_NORM, + MODEL_TENSOR.ATTN_V, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.FFN_NORM, + MODEL_TENSOR.FFN_GATE_INP, + MODEL_TENSOR.FFN_GATE_EXP, + MODEL_TENSOR.FFN_DOWN_EXP, + MODEL_TENSOR.FFN_UP_EXP, + ], MODEL_ARCH.PLAMO: [ MODEL_TENSOR.TOKEN_EMBD, MODEL_TENSOR.OUTPUT_NORM, @@ -1084,6 +1188,24 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = { MODEL_TENSOR.FFN_PRE_NORM, MODEL_TENSOR.FFN_POST_NORM, ], + MODEL_ARCH.GEMMA3: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.ATTN_Q, + MODEL_TENSOR.ATTN_Q_NORM, + MODEL_TENSOR.ATTN_K, + MODEL_TENSOR.ATTN_K_NORM, + MODEL_TENSOR.ATTN_V, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.FFN_GATE, + MODEL_TENSOR.FFN_DOWN, + MODEL_TENSOR.FFN_UP, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_POST_NORM, + MODEL_TENSOR.FFN_PRE_NORM, + MODEL_TENSOR.FFN_POST_NORM, + ], MODEL_ARCH.STARCODER2: [ MODEL_TENSOR.TOKEN_EMBD, MODEL_TENSOR.OUTPUT_NORM, @@ -1160,6 +1282,68 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = { MODEL_TENSOR.FFN_DOWN, MODEL_TENSOR.FFN_UP, ], + MODEL_ARCH.RWKV7: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.TOKEN_EMBD_NORM, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_NORM_2, + MODEL_TENSOR.TIME_MIX_LERP_FUSED, + MODEL_TENSOR.TIME_MIX_W0, + MODEL_TENSOR.TIME_MIX_W1, + MODEL_TENSOR.TIME_MIX_W2, + MODEL_TENSOR.TIME_MIX_A0, + MODEL_TENSOR.TIME_MIX_A1, + MODEL_TENSOR.TIME_MIX_A2, + MODEL_TENSOR.TIME_MIX_V0, + MODEL_TENSOR.TIME_MIX_V1, + MODEL_TENSOR.TIME_MIX_V2, + MODEL_TENSOR.TIME_MIX_G1, + MODEL_TENSOR.TIME_MIX_G2, + MODEL_TENSOR.TIME_MIX_K_K, + MODEL_TENSOR.TIME_MIX_K_A, + MODEL_TENSOR.TIME_MIX_R_K, + MODEL_TENSOR.TIME_MIX_KEY, + MODEL_TENSOR.TIME_MIX_VALUE, + MODEL_TENSOR.TIME_MIX_RECEPTANCE, + MODEL_TENSOR.TIME_MIX_LN, + MODEL_TENSOR.TIME_MIX_OUTPUT, + MODEL_TENSOR.CHANNEL_MIX_LERP_K, + MODEL_TENSOR.CHANNEL_MIX_KEY, + MODEL_TENSOR.CHANNEL_MIX_VALUE, + ], + MODEL_ARCH.ARWKV7: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.TOKEN_EMBD_NORM, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.TIME_MIX_LERP_FUSED, + MODEL_TENSOR.TIME_MIX_W0, + MODEL_TENSOR.TIME_MIX_W1, + MODEL_TENSOR.TIME_MIX_W2, + MODEL_TENSOR.TIME_MIX_A0, + MODEL_TENSOR.TIME_MIX_A1, + MODEL_TENSOR.TIME_MIX_A2, + MODEL_TENSOR.TIME_MIX_V0, + MODEL_TENSOR.TIME_MIX_V1, + MODEL_TENSOR.TIME_MIX_V2, + MODEL_TENSOR.TIME_MIX_G1, + MODEL_TENSOR.TIME_MIX_G2, + MODEL_TENSOR.TIME_MIX_K_K, + MODEL_TENSOR.TIME_MIX_K_A, + MODEL_TENSOR.TIME_MIX_R_K, + MODEL_TENSOR.TIME_MIX_KEY, + MODEL_TENSOR.TIME_MIX_VALUE, + MODEL_TENSOR.TIME_MIX_RECEPTANCE, + MODEL_TENSOR.TIME_MIX_LN, + MODEL_TENSOR.TIME_MIX_OUTPUT, + MODEL_TENSOR.FFN_NORM, + MODEL_TENSOR.FFN_GATE, + MODEL_TENSOR.FFN_DOWN, + MODEL_TENSOR.FFN_UP, + ], MODEL_ARCH.MAMBA: [ MODEL_TENSOR.TOKEN_EMBD, MODEL_TENSOR.OUTPUT_NORM, @@ -1357,6 +1541,20 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = { MODEL_TENSOR.FFN_UP_SHEXP, MODEL_TENSOR.FFN_EXP_PROBS_B, ], + MODEL_ARCH.PLM: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_Q, + MODEL_TENSOR.ATTN_KV_A_MQA, + MODEL_TENSOR.ATTN_KV_A_NORM, + MODEL_TENSOR.ATTN_KV_B, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.FFN_NORM, + MODEL_TENSOR.FFN_UP, + MODEL_TENSOR.FFN_DOWN, + ], MODEL_ARCH.CHATGLM : [ MODEL_TENSOR.TOKEN_EMBD, MODEL_TENSOR.ROPE_FREQS, @@ -1372,6 +1570,23 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = { MODEL_TENSOR.FFN_DOWN, MODEL_TENSOR.FFN_UP, ], + MODEL_ARCH.GLM4 : [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.ROPE_FREQS, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_QKV, + MODEL_TENSOR.ATTN_Q, + MODEL_TENSOR.ATTN_K, + MODEL_TENSOR.ATTN_V, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.FFN_NORM, + MODEL_TENSOR.FFN_DOWN, + MODEL_TENSOR.FFN_UP, + MODEL_TENSOR.ATTN_POST_NORM, + MODEL_TENSOR.FFN_POST_NORM, + ], MODEL_ARCH.BITNET: [ MODEL_TENSOR.ATTN_Q, MODEL_TENSOR.ATTN_K, @@ -1544,6 +1759,25 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = { MODEL_TENSOR.POSNET_ATTN_V, MODEL_TENSOR.POSNET_ATTN_OUT, ], + MODEL_ARCH.BAILINGMOE: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ROPE_FREQS, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_Q, + MODEL_TENSOR.ATTN_K, + MODEL_TENSOR.ATTN_V, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.FFN_GATE_INP, + MODEL_TENSOR.FFN_NORM, + MODEL_TENSOR.FFN_GATE_EXP, + MODEL_TENSOR.FFN_DOWN_EXP, + MODEL_TENSOR.FFN_UP_EXP, + MODEL_TENSOR.FFN_GATE_SHEXP, + MODEL_TENSOR.FFN_DOWN_SHEXP, + MODEL_TENSOR.FFN_UP_SHEXP, + ], # TODO } @@ -1596,6 +1830,9 @@ MODEL_TENSOR_SKIP: dict[MODEL_ARCH, list[MODEL_TENSOR]] = { MODEL_TENSOR.ROPE_FREQS, MODEL_TENSOR.ATTN_ROT_EMBD, ], + MODEL_ARCH.BAILINGMOE: [ + MODEL_TENSOR.ROPE_FREQS, + ], } # diff --git a/gguf-py/gguf/gguf_reader.py b/gguf-py/gguf/gguf_reader.py index e17a4e831..5991cdb76 100644 --- a/gguf-py/gguf/gguf_reader.py +++ b/gguf-py/gguf/gguf_reader.py @@ -6,6 +6,7 @@ from __future__ import annotations import logging import os +import sys from collections import OrderedDict from typing import Any, Literal, NamedTuple, TypeVar, Union @@ -15,7 +16,6 @@ import numpy.typing as npt from .quants import quant_shape_to_byte_shape if __name__ == "__main__": - import sys from pathlib import Path # Allow running file in package as a script. @@ -28,6 +28,7 @@ from gguf.constants import ( GGUF_VERSION, GGMLQuantizationType, GGUFValueType, + GGUFEndian, ) logger = logging.getLogger(__name__) @@ -53,6 +54,48 @@ class ReaderField(NamedTuple): types: list[GGUFValueType] = [] + def contents(self, index_or_slice: int | slice = slice(None)) -> Any: + if self.types: + to_string = lambda x: str(x.tobytes(), encoding='utf-8') # noqa: E731 + main_type = self.types[0] + + if main_type == GGUFValueType.ARRAY: + sub_type = self.types[-1] + + if sub_type == GGUFValueType.STRING: + indices = self.data[index_or_slice] + + if isinstance(index_or_slice, int): + return to_string(self.parts[indices]) # type: ignore + else: + return [to_string(self.parts[idx]) for idx in indices] # type: ignore + else: + # FIXME: When/if _get_field_parts() support multi-dimensional arrays, this must do so too + + # Check if it's unsafe to perform slice optimization on data + # if any(True for idx in self.data if len(self.parts[idx]) != 1): + # optim_slice = slice(None) + # else: + # optim_slice = index_or_slice + # index_or_slice = slice(None) + + # if isinstance(optim_slice, int): + # return self.parts[self.data[optim_slice]].tolist()[0] + # else: + # return [pv for idx in self.data[optim_slice] for pv in self.parts[idx].tolist()][index_or_slice] + + if isinstance(index_or_slice, int): + return self.parts[self.data[index_or_slice]].tolist()[0] + else: + return [pv for idx in self.data[index_or_slice] for pv in self.parts[idx].tolist()] + + if main_type == GGUFValueType.STRING: + return to_string(self.parts[-1]) + else: + return self.parts[-1].tolist()[0] + + return None + class ReaderTensor(NamedTuple): name: str @@ -101,10 +144,19 @@ class GGUFReader: # If we get 0 here that means it's (probably) a GGUF file created for # the opposite byte order of the machine this script is running on. self.byte_order = 'S' - temp_version = temp_version.newbyteorder(self.byte_order) + temp_version = temp_version.view(temp_version.dtype.newbyteorder(self.byte_order)) version = temp_version[0] if version not in READER_SUPPORTED_VERSIONS: raise ValueError(f'Sorry, file appears to be version {version} which we cannot handle') + if sys.byteorder == "little": + # Host is little endian + host_endian = GGUFEndian.LITTLE + swapped_endian = GGUFEndian.BIG + else: + # Sorry PDP or other weird systems that don't use BE or LE. + host_endian = GGUFEndian.BIG + swapped_endian = GGUFEndian.LITTLE + self.endianess = swapped_endian if self.byte_order == "S" else host_endian self.fields: OrderedDict[str, ReaderField] = OrderedDict() self.tensors: list[ReaderTensor] = [] offs += self._push_field(ReaderField(offs, 'GGUF.version', [temp_version], [0], [GGUFValueType.UINT32])) @@ -146,9 +198,7 @@ class GGUFReader: itemsize = int(np.empty([], dtype = dtype).itemsize) end_offs = offset + itemsize * count arr = self.data[offset:end_offs].view(dtype=dtype)[:count] - if override_order is None: - return arr - return arr.view(arr.dtype.newbyteorder(override_order)) + return arr.view(arr.dtype.newbyteorder(self.byte_order if override_order is None else override_order)) def _push_field(self, field: ReaderField, skip_sum: bool = False) -> int: if field.name in self.fields: @@ -190,6 +240,7 @@ class GGUFReader: offs += int(alen.nbytes) aparts: list[npt.NDArray[Any]] = [raw_itype, alen] data_idxs: list[int] = [] + # FIXME: Handle multi-dimensional arrays properly instead of flattening for idx in range(alen[0]): curr_size, curr_parts, curr_idxs, curr_types = self._get_field_parts(offs, raw_itype[0]) if idx == 0: diff --git a/gguf-py/gguf/gguf_writer.py b/gguf-py/gguf/gguf_writer.py index 080d2b9dc..485550aad 100644 --- a/gguf-py/gguf/gguf_writer.py +++ b/gguf-py/gguf/gguf_writer.py @@ -746,6 +746,9 @@ class GGUFWriter: def add_token_shift_count(self, count: int) -> None: self.add_uint32(Keys.LLM.TOKEN_SHIFT_COUNT.format(arch=self.arch), count) + def add_interleave_moe_layer_step(self, value: int) -> None: + self.add_uint32(Keys.LLM.INTERLEAVE_MOE_LAYER_STEP.format(arch=self.arch), value) + def add_layer_norm_eps(self, value: float) -> None: self.add_float32(Keys.Attention.LAYERNORM_EPS.format(arch=self.arch), value) @@ -767,6 +770,18 @@ class GGUFWriter: def add_kv_lora_rank(self, length: int) -> None: self.add_uint32(Keys.Attention.KV_LORA_RANK.format(arch=self.arch), length) + def add_decay_lora_rank(self, length: int) -> None: + self.add_uint32(Keys.Attention.DECAY_LORA_RANK.format(arch=self.arch), length) + + def add_iclr_lora_rank(self, length: int) -> None: + self.add_uint32(Keys.Attention.ICLR_LORA_RANK.format(arch=self.arch), length) + + def add_value_residual_mix_lora_rank(self, length: int) -> None: + self.add_uint32(Keys.Attention.VALUE_RESIDUAL_MIX_LORA_RANK.format(arch=self.arch), length) + + def add_gate_lora_rank(self, length: int) -> None: + self.add_uint32(Keys.Attention.GATE_LORA_RANK.format(arch=self.arch), length) + def add_relative_attn_buckets_count(self, value: int) -> None: self.add_uint32(Keys.Attention.REL_BUCKETS_COUNT.format(arch=self.arch), value) diff --git a/gguf-py/gguf/lazy.py b/gguf-py/gguf/lazy.py index 8d4fece2d..f9bcadae0 100644 --- a/gguf-py/gguf/lazy.py +++ b/gguf-py/gguf/lazy.py @@ -139,6 +139,16 @@ class LazyBase(ABC, metaclass=LazyMeta): if isinstance(res, cls._tensor_type): return cls(meta=cls.eager_to_meta(res), args=args, kwargs=kwargs, func=fn) + elif isinstance(res, tuple) and all(isinstance(t, cls._tensor_type) for t in res): + # share the evaluation between lazy tuple elements + shared_args: list = [args, None] + + def eager_tuple_element(a: list[Any], i: int = 0, /, **kw) -> LazyBase: + assert len(a) == 2 + if a[1] is None: + a[1] = fn(*a[0], **kw) + return a[1][i] + return tuple(cls(meta=cls.eager_to_meta(res[i]), args=(shared_args, i), kwargs=kwargs, func=eager_tuple_element) for i in range(len(res))) else: del res # not needed # non-tensor return likely relies on the contents of the args diff --git a/gguf-py/gguf/metadata.py b/gguf-py/gguf/metadata.py index 962c27b20..e807f4346 100644 --- a/gguf-py/gguf/metadata.py +++ b/gguf-py/gguf/metadata.py @@ -121,19 +121,39 @@ class Metadata: if not model_card_path.is_file(): return {} - # The model card metadata is assumed to always be in YAML + # The model card metadata is assumed to always be in YAML (frontmatter) # ref: https://github.com/huggingface/transformers/blob/a5c642fe7a1f25d3bdcd76991443ba6ff7ee34b2/src/transformers/modelcard.py#L468-L473 + yaml_content: str = "" with open(model_card_path, "r", encoding="utf-8") as f: - if f.readline() == "---\n": - raw = f.read().partition("---\n")[0] - data = yaml.safe_load(raw) - if isinstance(data, dict): - return data - else: - logger.error(f"while reading YAML model card frontmatter, data is {type(data)} instead of dict") - return {} - else: + content = f.read() + lines = content.splitlines() + lines_yaml = [] + if len(lines) == 0: + # Empty file return {} + if len(lines) > 0 and lines[0] != "---": + # No frontmatter + return {} + for line in lines[1:]: + if line == "---": + break # End of frontmatter + else: + lines_yaml.append(line) + yaml_content = "\n".join(lines_yaml) + "\n" + + # Quick hack to fix the Norway problem + # https://hitchdev.com/strictyaml/why/implicit-typing-removed/ + yaml_content = yaml_content.replace("- no\n", "- \"no\"\n") + + if yaml_content: + data = yaml.safe_load(yaml_content) + if isinstance(data, dict): + return data + else: + logger.error(f"while reading YAML model card frontmatter, data is {type(data)} instead of dict") + return {} + else: + return {} @staticmethod def load_hf_parameters(model_path: Optional[Path] = None) -> dict[str, Any]: diff --git a/gguf-py/gguf/scripts/gguf_convert_endian.py b/gguf-py/gguf/scripts/gguf_convert_endian.py index f97e91bd4..0e0febaa7 100755 --- a/gguf-py/gguf/scripts/gguf_convert_endian.py +++ b/gguf-py/gguf/scripts/gguf_convert_endian.py @@ -20,22 +20,15 @@ logger = logging.getLogger("gguf-convert-endian") def convert_byteorder(reader: gguf.GGUFReader, args: argparse.Namespace) -> None: - if np.uint32(1) == np.uint32(1).newbyteorder("<"): - # Host is little endian - host_endian = "little" - swapped_endian = "big" + file_endian = reader.endianess.name + if reader.byte_order == 'S': + host_endian = 'BIG' if file_endian == 'LITTLE' else 'LITTLE' else: - # Sorry PDP or other weird systems that don't use BE or LE. - host_endian = "big" - swapped_endian = "little" - if reader.byte_order == "S": - file_endian = swapped_endian - else: - file_endian = host_endian - order = host_endian if args.order == "native" else args.order - logger.info(f"* Host is {host_endian.upper()} endian, GGUF file seems to be {file_endian.upper()} endian") + host_endian = file_endian + order = host_endian if args.order == "native" else args.order.upper() + logger.info(f"* Host is {host_endian} endian, GGUF file seems to be {file_endian} endian") if file_endian == order: - logger.info(f"* File is already {order.upper()} endian. Nothing to do.") + logger.info(f"* File is already {order} endian. Nothing to do.") sys.exit(0) logger.info("* Checking tensors for conversion compatibility") for tensor in reader.tensors: @@ -43,9 +36,11 @@ def convert_byteorder(reader: gguf.GGUFReader, args: argparse.Namespace) -> None gguf.GGMLQuantizationType.F32, gguf.GGMLQuantizationType.F16, gguf.GGMLQuantizationType.Q8_0, + gguf.GGMLQuantizationType.Q4_K, + gguf.GGMLQuantizationType.Q6_K, ): raise ValueError(f"Cannot handle type {tensor.tensor_type.name} for tensor {repr(tensor.name)}") - logger.info(f"* Preparing to convert from {file_endian.upper()} to {order.upper()}") + logger.info(f"* Preparing to convert from {file_endian} to {order}") if args.dry_run: return logger.warning("*** Warning *** Warning *** Warning **") @@ -96,6 +91,59 @@ def convert_byteorder(reader: gguf.GGUFReader, args: argparse.Namespace) -> None if block_num % 100000 == 0: inner_pbar.set_description(f"Byte-swapping Blocks [{(n_blocks - block_num) // n_blocks}]") + elif tensor.tensor_type == gguf.GGMLQuantizationType.Q4_K: + # Handle Q4_K tensor blocks (block_q4_k) + # Specific handling of block_q4_k is required. + # Each block_q4_k consists of 2 f16 values followed by 140 int8 values. + + # first flatten structure + newshape = 1 + for i in tensor.data.shape: + newshape *= i + + tensor.data.resize(newshape) + + block_size = 144 + n_blocks = len(tensor.data) // block_size + for block_num in (inner_pbar := tqdm(range(n_blocks), desc="Byte-swapping Blocks", leave=False)): + block_offs = block_num * block_size + + # Byte-Swap f16 sized fields + delta = tensor.data[block_offs:block_offs + 2].view(dtype=np.uint16) + delta.byteswap(inplace=True) + + delta = tensor.data[block_offs + 2:block_offs + 4].view(dtype=np.uint16) + delta.byteswap(inplace=True) + + # Byte-Swap + if block_num % 100000 == 0: + inner_pbar.set_description(f"Byte-swapping Blocks [{(n_blocks - block_num) // n_blocks}]") + + elif tensor.tensor_type == gguf.GGMLQuantizationType.Q6_K: + # Handle Q6_K tensor blocks (block_q6_k) + # Specific handling of block_q6_k is required. + # Each block_q6_k consists of 208 int8 values followed by 1 f16 value. + + # first flatten structure + newshape = 1 + for i in tensor.data.shape: + newshape *= i + + tensor.data.resize(newshape) + + block_size = 210 + n_blocks = len(tensor.data) // block_size + for block_num in (inner_pbar := tqdm(range(n_blocks), desc="Byte-swapping Blocks", leave=False)): + block_offs = block_num * block_size + + # Byte-Swap f16 sized field + delta = tensor.data[block_offs + 208:block_offs + 210].view(dtype=np.uint16) + delta.byteswap(inplace=True) + + # Byte-Swap + if block_num % 100000 == 0: + inner_pbar.set_description(f"Byte-swapping Blocks [{(n_blocks - block_num) // n_blocks}]") + else: # Handle other tensor types tensor.data.byteswap(inplace=True) diff --git a/gguf-py/gguf/scripts/gguf_dump.py b/gguf-py/gguf/scripts/gguf_dump.py index f95b4fd48..e282892d6 100755 --- a/gguf-py/gguf/scripts/gguf_dump.py +++ b/gguf-py/gguf/scripts/gguf_dump.py @@ -9,8 +9,6 @@ import sys from pathlib import Path from typing import Any -import numpy as np - # Necessary to load the local gguf package if "NO_LOCAL_GGUF" not in os.environ and (Path(__file__).parent.parent.parent.parent / 'gguf-py').exists(): sys.path.insert(0, str(Path(__file__).parent.parent.parent)) @@ -21,11 +19,11 @@ logger = logging.getLogger("gguf-dump") def get_file_host_endian(reader: GGUFReader) -> tuple[str, str]: - host_endian = 'LITTLE' if np.uint32(1) == np.uint32(1).newbyteorder("<") else 'BIG' + file_endian = reader.endianess.name if reader.byte_order == 'S': - file_endian = 'BIG' if host_endian == 'LITTLE' else 'LITTLE' + host_endian = 'BIG' if file_endian == 'LITTLE' else 'LITTLE' else: - file_endian = host_endian + host_endian = file_endian return (host_endian, file_endian) @@ -45,12 +43,20 @@ def dump_metadata(reader: GGUFReader, args: argparse.Namespace) -> None: pretty_type = str(field.types[-1].name) log_message = f' {n:5}: {pretty_type:10} | {len(field.data):8} | {field.name}' - if len(field.types) == 1: + if field.types: curr_type = field.types[0] if curr_type == GGUFValueType.STRING: - log_message += ' = {0}'.format(repr(str(bytes(field.parts[-1]), encoding='utf-8')[:60])) - elif field.types[0] in reader.gguf_scalar_to_np: - log_message += ' = {0}'.format(field.parts[-1][0]) + content = field.contents() + if len(content) > 60: + content = content[:57] + '...' + log_message += ' = {0}'.format(repr(content)) + elif curr_type in reader.gguf_scalar_to_np: + log_message += ' = {0}'.format(field.contents()) + else: + content = repr(field.contents(slice(6))) + if len(field.data) > 6: + content = content[:-1] + ', ...]' + log_message += ' = {0}'.format(content) print(log_message) # noqa: NP100 if args.no_tensors: return @@ -82,15 +88,9 @@ def dump_metadata_json(reader: GGUFReader, args: argparse.Namespace) -> None: curr["array_types"] = [t.name for t in field.types][1:] if not args.json_array: continue - itype = field.types[-1] - if itype == GGUFValueType.STRING: - curr["value"] = [str(bytes(field.parts[idx]), encoding="utf-8") for idx in field.data] - else: - curr["value"] = [pv for idx in field.data for pv in field.parts[idx].tolist()] - elif field.types[0] == GGUFValueType.STRING: - curr["value"] = str(bytes(field.parts[-1]), encoding="utf-8") + curr["value"] = field.contents() else: - curr["value"] = field.parts[-1].tolist()[0] + curr["value"] = field.contents() if not args.no_tensors: for idx, tensor in enumerate(reader.tensors): tensors[tensor.name] = { @@ -181,7 +181,7 @@ def element_count_rounded_notation(count: int) -> str: def translate_tensor_name(name): words = name.split(".") - # Source: https://github.com/ggerganov/ggml/blob/master/docs/gguf.md#standardized-tensor-names + # Source: https://github.com/ggml-org/ggml/blob/master/docs/gguf.md#standardized-tensor-names abbreviation_dictionary = { 'token_embd': 'Token embedding', 'pos_embd': 'Position embedding', diff --git a/gguf-py/gguf/scripts/gguf_new_metadata.py b/gguf-py/gguf/scripts/gguf_new_metadata.py index a8cfc9d58..7aff6c925 100755 --- a/gguf-py/gguf/scripts/gguf_new_metadata.py +++ b/gguf-py/gguf/scripts/gguf_new_metadata.py @@ -8,7 +8,6 @@ import sys import json from pathlib import Path -import numpy as np from tqdm import tqdm from typing import Any, Sequence, NamedTuple @@ -27,45 +26,10 @@ class MetadataDetails(NamedTuple): description: str = '' -def get_byteorder(reader: gguf.GGUFReader) -> gguf.GGUFEndian: - if np.uint32(1) == np.uint32(1).newbyteorder("<"): - # Host is little endian - host_endian = gguf.GGUFEndian.LITTLE - swapped_endian = gguf.GGUFEndian.BIG - else: - # Sorry PDP or other weird systems that don't use BE or LE. - host_endian = gguf.GGUFEndian.BIG - swapped_endian = gguf.GGUFEndian.LITTLE - - if reader.byte_order == "S": - return swapped_endian - else: - return host_endian - - -def decode_field(field: gguf.ReaderField | None) -> Any: - if field and field.types: - main_type = field.types[0] - - if main_type == gguf.GGUFValueType.ARRAY: - sub_type = field.types[-1] - - if sub_type == gguf.GGUFValueType.STRING: - return [str(bytes(field.parts[idx]), encoding='utf-8') for idx in field.data] - else: - return [pv for idx in field.data for pv in field.parts[idx].tolist()] - if main_type == gguf.GGUFValueType.STRING: - return str(bytes(field.parts[-1]), encoding='utf-8') - else: - return field.parts[-1][0] - - return None - - def get_field_data(reader: gguf.GGUFReader, key: str) -> Any: field = reader.get_field(key) - return decode_field(field) + return field.contents() if field else None def find_token(token_list: Sequence[int], token: str) -> Sequence[int]: @@ -93,7 +57,7 @@ def copy_with_new_metadata(reader: gguf.GGUFReader, writer: gguf.GGUFWriter, new logger.debug(f'Removing {field.name}') continue - old_val = MetadataDetails(field.types[0], decode_field(field)) + old_val = MetadataDetails(field.types[0], field.contents()) val = new_metadata.get(field.name, old_val) if field.name in new_metadata: @@ -192,7 +156,6 @@ def main() -> None: reader = gguf.GGUFReader(args.input, 'r') arch = get_field_data(reader, gguf.Keys.General.ARCHITECTURE) - endianess = get_byteorder(reader) token_list = get_field_data(reader, gguf.Keys.Tokenizer.LIST) or [] @@ -230,7 +193,7 @@ def main() -> None: sys.exit(0) logger.info(f'* Writing: {args.output}') - writer = gguf.GGUFWriter(args.output, arch=arch, endianess=endianess) + writer = gguf.GGUFWriter(args.output, arch=arch, endianess=reader.endianess) alignment = get_field_data(reader, gguf.Keys.General.ALIGNMENT) if alignment is not None: diff --git a/gguf-py/gguf/tensor_mapping.py b/gguf-py/gguf/tensor_mapping.py index 617791e24..35154e9b5 100644 --- a/gguf-py/gguf/tensor_mapping.py +++ b/gguf-py/gguf/tensor_mapping.py @@ -13,7 +13,7 @@ class TensorNameMap: "transformer.wte", # gpt2 gpt-j mpt refact qwen dbrx jais exaone "transformer.word_embeddings", # falcon "word_embeddings", # bloom - "model.embed_tokens", # llama-hf nemotron olmoe olmo2 rwkv6qwen2 + "model.embed_tokens", # llama-hf nemotron olmoe olmo2 rwkv6qwen2 glm4-0414 "tok_embeddings", # llama-pth "embeddings.word_embeddings", # bert nomic-bert "language_model.embedding.word_embeddings", # persimmon @@ -27,7 +27,10 @@ class TensorNameMap: "embedding.word_embeddings", # chatglm "transformer.token_embeddings", # openelm "shared", # t5 - "rwkv.embeddings", # rwkv + "rwkv.embeddings", # rwkv6 + "model.embeddings", # rwkv7 + "model.word_embeddings", # bailingmoe + "language_model.model.embed_tokens", # llama4 ), # Token type embeddings @@ -42,6 +45,9 @@ class TensorNameMap: "emb_ln", # nomic-bert "transformer.norm", # openelm "rwkv.blocks.0.pre_ln", # rwkv + "rwkv.blocks.0.pre_ln", # rwkv6 + "model.pre_ln", # rwkv7 + "model.layers.0.pre_norm", # rwkv7 "backbone.norm", # wavtokenizer ), @@ -62,6 +68,7 @@ class TensorNameMap: "output_layer", # chatglm "head", # rwkv "head.out", # wavtokenizer + "language_model.lm_head", # llama4 ), # Output norm @@ -81,8 +88,10 @@ class TensorNameMap: "encoder.final_layernorm", # chatglm "transformer.norm", # openelm "model.norm", # nemotron - "rwkv.ln_out", # rwkv + "rwkv.ln_out", # rwkv6 + "model.ln_out", # rwkv7 "backbone.final_layer_norm", # wavtokenizer + "language_model.model.norm", # llama4 ), # Rope frequencies @@ -122,14 +131,17 @@ class TensorNameMap: "transformer.blocks.{bid}.norm_attn_norm.norm_1", # dbrx "encoder.layers.{bid}.input_layernorm", # chatglm "transformer.layers.{bid}.attn_norm", # openelm - "rwkv.blocks.{bid}.ln1", # rwkv + "rwkv.blocks.{bid}.ln1", # rwkv6 + "model.layers.{bid}.ln1", # rwkv7 + "language_model.model.layers.{bid}.input_layernorm", # llama4 ), # Attention norm 2 MODEL_TENSOR.ATTN_NORM_2: ( "transformer.h.{bid}.ln_attn", # falcon40b "encoder.layer.{bid}.layer_norm_1", # jina-v2-code - "rwkv.blocks.{bid}.ln2", # rwkv + "rwkv.blocks.{bid}.ln2", # rwkv6 + "model.layers.{bid}.ln2", # rwkv7 ), # Attention query-key-value @@ -161,6 +173,7 @@ class TensorNameMap: "model.layers.{bid}.attention.wq", # internlm2 "transformer.decoder_layer.{bid}.multi_head_attention.query",# Grok "transformer.h.{bid}.attn.attention.q_proj", # exaone + "language_model.model.layers.{bid}.self_attn.q_proj", # llama4 ), # Attention key @@ -175,6 +188,7 @@ class TensorNameMap: "model.layers.{bid}.attention.wk", # internlm2 "transformer.decoder_layer.{bid}.multi_head_attention.key",# Grok "transformer.h.{bid}.attn.attention.k_proj", # exaone + "language_model.model.layers.{bid}.self_attn.k_proj", # llama4 ), # Attention value @@ -188,6 +202,7 @@ class TensorNameMap: "model.layers.{bid}.attention.wv", # internlm2 "transformer.decoder_layer.{bid}.multi_head_attention.value",# Grok "transformer.h.{bid}.attn.attention.v_proj", # exaone + "language_model.model.layers.{bid}.self_attn.v_proj", # llama4 ), # Attention output @@ -214,6 +229,7 @@ class TensorNameMap: "encoder.layers.{bid}.self_attention.dense", # chatglm "transformer.layers.{bid}.attn.out_proj", # openelm "transformer.h.{bid}.attn.attention.out_proj", # exaone + "language_model.model.layers.{bid}.self_attn.o_proj", # llama4 ), # Attention output norm @@ -225,7 +241,8 @@ class TensorNameMap: ), MODEL_TENSOR.ATTN_POST_NORM: ( - "model.layers.{bid}.post_attention_layernorm", # gemma2 olmo2 + "model.layers.{bid}.post_attention_layernorm", # gemma2 olmo2 # ge + "model.layers.{bid}.post_self_attn_layernorm", # glm-4-0414 ), # Rotary embeddings @@ -251,6 +268,7 @@ class TensorNameMap: "transformer.decoder_layer.{bid}.rms_norm_2", # Grok "encoder.layers.{bid}.post_attention_layernorm", # chatglm "transformer.layers.{bid}.ffn_norm", # openelm + "language_model.model.layers.{bid}.post_attention_layernorm", # llama4 ), # Post feed-forward norm @@ -261,6 +279,7 @@ class TensorNameMap: # Post feed-forward norm MODEL_TENSOR.FFN_POST_NORM: ( "model.layers.{bid}.post_feedforward_layernorm", # gemma2 olmo2 + "model.layers.{bid}.post_mlp_layernorm", # glm-4-0414 ), MODEL_TENSOR.FFN_GATE_INP: ( @@ -270,6 +289,7 @@ class TensorNameMap: "transformer.decoder_layer.{bid}.router", # Grok "transformer.blocks.{bid}.ffn.router.layer", # dbrx "model.layers.{bid}.block_sparse_moe.router.layer", # granitemoe + "language_model.model.layers.{bid}.feed_forward.router", # llama4 ), MODEL_TENSOR.FFN_GATE_INP_SHEXP: ( @@ -298,7 +318,7 @@ class TensorNameMap: "h.{bid}.mlp.c_fc", # gpt2 "transformer.h.{bid}.mlp.fc1", # phi2 "model.layers.{bid}.mlp.fc1", # phi2 - "model.layers.{bid}.mlp.gate_up_proj", # phi3 + "model.layers.{bid}.mlp.gate_up_proj", # phi3 glm-4-0414 "model.layers.layers.{bid}.mlp.up_proj", # plamo "model.layers.{bid}.feed_forward.w3", # internlm2 "encoder.layers.{bid}.mlp.fc11", # nomic-bert @@ -307,6 +327,7 @@ class TensorNameMap: "model.layers.{bid}.residual_mlp.w3", # arctic "encoder.layers.{bid}.mlp.dense_h_to_4h", # chatglm "transformer.h.{bid}.mlp.c_fc_1", # exaone + "language_model.model.layers.{bid}.feed_forward.up_proj", # llama4 ), MODEL_TENSOR.FFN_UP_EXP: ( @@ -315,11 +336,13 @@ class TensorNameMap: "transformer.blocks.{bid}.ffn.experts.mlp.v1", # dbrx "model.layers.{bid}.mlp.experts.up_proj", # qwen2moe olmoe (merged) "model.layers.{bid}.block_sparse_moe.experts.w3", # phimoe (merged) + "language_model.model.layers.{bid}.feed_forward.experts.up_proj", # llama4 ), MODEL_TENSOR.FFN_UP_SHEXP: ( "model.layers.{bid}.mlp.shared_expert.up_proj", # qwen2moe "model.layers.{bid}.mlp.shared_experts.up_proj", # deepseek deepseek2 + "language_model.model.layers.{bid}.feed_forward.shared_expert.up_proj", # llama4 ), # AWQ-activation gate @@ -340,6 +363,7 @@ class TensorNameMap: "transformer.h.{bid}.mlp.linear_1", # refact "model.layers.{bid}.residual_mlp.w1", # arctic "transformer.h.{bid}.mlp.c_fc_0", # exaone + "language_model.model.layers.{bid}.feed_forward.gate_proj", # llama4 ), MODEL_TENSOR.FFN_GATE_EXP: ( @@ -348,11 +372,13 @@ class TensorNameMap: "transformer.blocks.{bid}.ffn.experts.mlp.w1", # dbrx "model.layers.{bid}.mlp.experts.gate_proj", # qwen2moe olmoe (merged) "model.layers.{bid}.block_sparse_moe.experts.w1", # phimoe (merged) + "language_model.model.layers.{bid}.feed_forward.experts.gate_proj", # llama4 ), MODEL_TENSOR.FFN_GATE_SHEXP: ( "model.layers.{bid}.mlp.shared_expert.gate_proj", # qwen2moe "model.layers.{bid}.mlp.shared_experts.gate_proj", # deepseek deepseek2 + "language_model.model.layers.{bid}.feed_forward.shared_expert.gate_proj", # llama4 ), # Feed-forward down @@ -381,6 +407,7 @@ class TensorNameMap: "encoder.layer.{bid}.mlp.down_layer", # jina-bert-v2 "encoder.layers.{bid}.mlp.dense_4h_to_h", # chatglm "model.layers.h.{bid}.mlp.c_proj", # exaone + "language_model.model.layers.{bid}.feed_forward.down_proj", # llama4 ), MODEL_TENSOR.FFN_DOWN_EXP: ( @@ -390,11 +417,13 @@ class TensorNameMap: "model.layers.{bid}.mlp.experts.down_proj", # qwen2moe olmoe (merged) "model.layers.{bid}.block_sparse_moe.output_linear", # granitemoe "model.layers.{bid}.block_sparse_moe.experts.w2", # phimoe (merged) + "language_model.model.layers.{bid}.feed_forward.experts.down_proj", # llama4 ), MODEL_TENSOR.FFN_DOWN_SHEXP: ( "model.layers.{bid}.mlp.shared_expert.down_proj", # qwen2moe "model.layers.{bid}.mlp.shared_experts.down_proj", # deepseek deepseek2 + "language_model.model.layers.{bid}.feed_forward.shared_expert.down_proj", # llama4 ), MODEL_TENSOR.ATTN_Q_NORM: ( @@ -462,112 +491,174 @@ class TensorNameMap: "backbone.layers.{bid}.mixer.out_proj", ), + MODEL_TENSOR.TIME_MIX_W0: ( + "model.layers.{bid}.attention.w0", # rwkv7 + ), + MODEL_TENSOR.TIME_MIX_W1: ( - "rwkv.blocks.{bid}.attention.time_maa_w1", # rwkv v6 - "model.layers.{bid}.self_attn.time_maa_w1", # rwkv6qwen2 + "rwkv.blocks.{bid}.attention.time_maa_w1", # rwkv6 + "model.layers.{bid}.self_attn.time_maa_w1", # rwkv6qwen2 + "model.layers.{bid}.attention.w1", # rwkv7 ), MODEL_TENSOR.TIME_MIX_W2: ( - "rwkv.blocks.{bid}.attention.time_maa_w2", # rwkv v6 - "model.layers.{bid}.self_attn.time_maa_w2", # rwkv6qwen2 + "rwkv.blocks.{bid}.attention.time_maa_w2", # rwkv6 + "model.layers.{bid}.self_attn.time_maa_w2", # rwkv6qwen2 + "model.layers.{bid}.attention.w2", # rwkv7 + ), + + MODEL_TENSOR.TIME_MIX_A0: ( + "model.layers.{bid}.attention.a0", # rwkv7 + ), + + MODEL_TENSOR.TIME_MIX_A1: ( + "model.layers.{bid}.attention.a1", # rwkv7 + ), + + MODEL_TENSOR.TIME_MIX_A2: ( + "model.layers.{bid}.attention.a2", # rwkv7 + ), + + MODEL_TENSOR.TIME_MIX_V0: ( + "model.layers.{bid}.attention.v0", # rwkv7 + ), + + MODEL_TENSOR.TIME_MIX_V1: ( + "model.layers.{bid}.attention.v1", # rwkv7 + ), + + MODEL_TENSOR.TIME_MIX_V2: ( + "model.layers.{bid}.attention.v2", # rwkv7 + ), + + MODEL_TENSOR.TIME_MIX_G1: ( + "model.layers.{bid}.attention.g1", # rwkv7 + ), + + MODEL_TENSOR.TIME_MIX_G2: ( + "model.layers.{bid}.attention.g2", # rwkv7 + ), + + MODEL_TENSOR.TIME_MIX_K_K: ( + "model.layers.{bid}.attention.k_k", # rwkv7 + ), + + MODEL_TENSOR.TIME_MIX_K_A: ( + "model.layers.{bid}.attention.k_a", # rwkv7 + ), + + MODEL_TENSOR.TIME_MIX_R_K: ( + "model.layers.{bid}.attention.r_k", # rwkv7 ), MODEL_TENSOR.TIME_MIX_LERP_X: ( - "rwkv.blocks.{bid}.attention.time_maa_x", # rwkv v6 + "rwkv.blocks.{bid}.attention.time_maa_x", # rwkv6 "model.layers.{bid}.self_attn.time_maa_x", # rwkv6qwen2 ), MODEL_TENSOR.TIME_MIX_LERP_K: ( - "rwkv.blocks.{bid}.attention.time_maa_k", # rwkv v6 + "rwkv.blocks.{bid}.attention.time_maa_k", # rwkv6 "model.layers.{bid}.self_attn.time_maa_k", # rwkv6qwen2 ), MODEL_TENSOR.TIME_MIX_LERP_V: ( - "rwkv.blocks.{bid}.attention.time_maa_v", # rwkv v6 + "rwkv.blocks.{bid}.attention.time_maa_v", # rwkv6 "model.layers.{bid}.self_attn.time_maa_v", # rwkv6qwen2 ), MODEL_TENSOR.TIME_MIX_LERP_R: ( - "rwkv.blocks.{bid}.attention.time_maa_r", # rwkv v6 + "rwkv.blocks.{bid}.attention.time_maa_r", # rwkv6 "model.layers.{bid}.self_attn.time_maa_r", # rwkv6qwen2 ), MODEL_TENSOR.TIME_MIX_LERP_G: ( - "rwkv.blocks.{bid}.attention.time_maa_g", # rwkv v6 + "rwkv.blocks.{bid}.attention.time_maa_g", # rwkv6 "model.layers.{bid}.self_attn.time_maa_g", # rwkv6qwen2 ), MODEL_TENSOR.TIME_MIX_LERP_W: ( - "rwkv.blocks.{bid}.attention.time_maa_w", # rwkv v6 + "rwkv.blocks.{bid}.attention.time_maa_w", # rwkv6 "model.layers.{bid}.self_attn.time_maa_w", # rwkv6qwen2 ), MODEL_TENSOR.TIME_MIX_FIRST: ( - "rwkv.blocks.{bid}.attention.time_faaaa", # rwkv v6 + "rwkv.blocks.{bid}.attention.time_faaaa", # rwkv6 ), MODEL_TENSOR.TIME_MIX_DECAY: ( - "rwkv.blocks.{bid}.attention.time_decay", # rwkv v6 + "rwkv.blocks.{bid}.attention.time_decay", # rwkv6 "model.layers.{bid}.self_attn.time_decay", # rwkv6qwen2 ), MODEL_TENSOR.TIME_MIX_DECAY_W1: ( - "rwkv.blocks.{bid}.attention.time_decay_w1", # rwkv v6 + "rwkv.blocks.{bid}.attention.time_decay_w1", # rwkv6 "model.layers.{bid}.self_attn.time_decay_w1", # rwkv6qwen2 ), MODEL_TENSOR.TIME_MIX_DECAY_W2: ( - "rwkv.blocks.{bid}.attention.time_decay_w2", # rwkv v6 + "rwkv.blocks.{bid}.attention.time_decay_w2", # rwkv6 "model.layers.{bid}.self_attn.time_decay_w2", # rwkv6qwen2 ), MODEL_TENSOR.TIME_MIX_KEY: ( - "rwkv.blocks.{bid}.attention.key", # rwkv + "rwkv.blocks.{bid}.attention.key", # rwkv6 "model.layers.{bid}.self_attn.k_proj", # rwkv6qwen2 + "model.layers.{bid}.attention.key", # rwkv7 + "model.layers.{bid}.attention.k_proj", # rwkv7 ), MODEL_TENSOR.TIME_MIX_VALUE: ( - "rwkv.blocks.{bid}.attention.value", # rwkv + "rwkv.blocks.{bid}.attention.value", # rwkv6 "model.layers.{bid}.self_attn.v_proj", # rwkv6qwen2 + "model.layers.{bid}.attention.value", # rwkv7 + "model.layers.{bid}.attention.v_proj", # rwkv7 ), MODEL_TENSOR.TIME_MIX_RECEPTANCE: ( - "rwkv.blocks.{bid}.attention.receptance", # rwkv - "model.layers.{bid}.self_attn.q_proj", # rwkv6qwen2 + "rwkv.blocks.{bid}.attention.receptance", # rwkv6 + "model.layers.{bid}.self_attn.q_proj", # rwkv6qwen2 + "model.layers.{bid}.attention.receptance", # rwkv7 + "model.layers.{bid}.attention.r_proj", # rwkv7 ), MODEL_TENSOR.TIME_MIX_GATE: ( - "rwkv.blocks.{bid}.attention.gate", # rwkv - "model.layers.{bid}.self_attn.gate", # rwkv6qwen2 + "rwkv.blocks.{bid}.attention.gate", # rwkv6 + "model.layers.{bid}.self_attn.gate", # rwkv6qwen2 ), MODEL_TENSOR.TIME_MIX_LN: ( - "rwkv.blocks.{bid}.attention.ln_x", # rwkv + "rwkv.blocks.{bid}.attention.ln_x", # rwkv6 + "model.layers.{bid}.attention.ln_x" # rwkv7 ), MODEL_TENSOR.TIME_MIX_OUTPUT: ( - "rwkv.blocks.{bid}.attention.output", # rwkv + "rwkv.blocks.{bid}.attention.output", # rwkv6 "model.layers.{bid}.self_attn.o_proj", # rwkv6qwen2 + "model.layers.{bid}.attention.output", # rwkv7 + "model.layers.{bid}.attention.o_proj", # rwkv7 ), MODEL_TENSOR.CHANNEL_MIX_LERP_K: ( - "rwkv.blocks.{bid}.feed_forward.time_maa_k", # rwkv v6 + "rwkv.blocks.{bid}.feed_forward.time_maa_k", # rwkv6 + "model.layers.{bid}.feed_forward.x_k", # rwkv7 ), MODEL_TENSOR.CHANNEL_MIX_LERP_R: ( - "rwkv.blocks.{bid}.feed_forward.time_maa_r", # rwkv v6 + "rwkv.blocks.{bid}.feed_forward.time_maa_r", # rwkv6 ), MODEL_TENSOR.CHANNEL_MIX_KEY: ( - "rwkv.blocks.{bid}.feed_forward.key", # rwkv + "rwkv.blocks.{bid}.feed_forward.key", # rwkv6 + "model.layers.{bid}.feed_forward.key", # rwkv7 ), MODEL_TENSOR.CHANNEL_MIX_RECEPTANCE: ( - "rwkv.blocks.{bid}.feed_forward.receptance", # rwkv + "rwkv.blocks.{bid}.feed_forward.receptance", # rwkv6 ), MODEL_TENSOR.CHANNEL_MIX_VALUE: ( - "rwkv.blocks.{bid}.feed_forward.value", # rwkv + "rwkv.blocks.{bid}.feed_forward.value", # rwkv6 + "model.layers.{bid}.feed_forward.value", # rwkv7 ), MODEL_TENSOR.ATTN_Q_A: ( diff --git a/gguf-py/gguf/utility.py b/gguf-py/gguf/utility.py index 40d59b75e..e5251aef8 100644 --- a/gguf-py/gguf/utility.py +++ b/gguf-py/gguf/utility.py @@ -1,7 +1,11 @@ from __future__ import annotations +from dataclasses import dataclass from typing import Literal +import os +import json + def fill_templated_filename(filename: str, output_type: str | None) -> str: # Given a file name fill in any type templates e.g. 'some-model-name.{ftype}.gguf' @@ -47,7 +51,7 @@ def size_label(total_params: int, shared_params: int, expert_params: int, expert def naming_convention(model_name: str | None, base_name: str | None, finetune_string: str | None, version_string: str | None, size_label: str | None, output_type: str | None, model_type: Literal['vocab', 'LoRA'] | None = None) -> str: - # Reference: https://github.com/ggerganov/ggml/blob/master/docs/gguf.md#gguf-naming-convention + # Reference: https://github.com/ggml-org/ggml/blob/master/docs/gguf.md#gguf-naming-convention if base_name is not None: name = base_name.strip().replace(' ', '-').replace('/', '-') @@ -67,3 +71,194 @@ def naming_convention(model_name: str | None, base_name: str | None, finetune_st kind = f"-{model_type.strip().replace(' ', '-')}" if model_type is not None else "" return f"{name}{parameters}{finetune}{version}{encoding}{kind}" + + +@dataclass +class RemoteTensor: + dtype: str + shape: tuple[int, ...] + offset_start: int + size: int + url: str + + def data(self) -> bytearray: + # TODO: handle request errors (maybe with limited retries?) + # NOTE: using a bytearray, otherwise PyTorch complains the buffer is not writeable + data = bytearray(SafetensorRemote.get_data_by_range(url=self.url, start=self.offset_start, size=self.size)) + return data + + +class SafetensorRemote: + """ + Uility class to handle remote safetensor files. + This class is designed to work with Hugging Face model repositories. + + Example (one model has single safetensor file, the other has multiple): + for model_id in ["ngxson/TEST-Tiny-Llama4", "Qwen/Qwen2.5-7B-Instruct"]: + tensors = SafetensorRemote.get_list_tensors_hf_model(model_id) + print(tensors) + + Example reading tensor data: + tensors = SafetensorRemote.get_list_tensors_hf_model(model_id) + for name, meta in tensors.items(): + dtype, shape, offset_start, size, remote_safetensor_url = meta + # read the tensor data + data = SafetensorRemote.get_data_by_range(remote_safetensor_url, offset_start, size) + print(data) + """ + + BASE_DOMAIN = "https://huggingface.co" + ALIGNMENT = 8 # bytes + + @classmethod + def get_list_tensors_hf_model(cls, model_id: str) -> dict[str, RemoteTensor]: + """ + Get list of tensors from a Hugging Face model repository. + + Returns a dictionary of tensor names and their metadata. + Each tensor is represented as a tuple of (dtype, shape, offset_start, size, remote_safetensor_url) + """ + # case 1: model has only one single model.safetensor file + is_single_file = cls.check_file_exist(f"{cls.BASE_DOMAIN}/{model_id}/resolve/main/model.safetensors") + if is_single_file: + url = f"{cls.BASE_DOMAIN}/{model_id}/resolve/main/model.safetensors" + return cls.get_list_tensors(url) + + # case 2: model has multiple files + index_url = f"{cls.BASE_DOMAIN}/{model_id}/resolve/main/model.safetensors.index.json" + is_multiple_files = cls.check_file_exist(index_url) + if is_multiple_files: + # read the index file + index_data = cls.get_data_by_range(index_url, 0) + index_str = index_data.decode('utf-8') + index_json = json.loads(index_str) + assert index_json.get("weight_map") is not None, "weight_map not found in index file" + weight_map = index_json["weight_map"] + # get the list of files + all_files = list(set(weight_map.values())) + all_files.sort() # make sure we load shard files in order + # get the list of tensors + tensors: dict[str, RemoteTensor] = {} + for file in all_files: + url = f"{cls.BASE_DOMAIN}/{model_id}/resolve/main/{file}" + for key, val in cls.get_list_tensors(url).items(): + tensors[key] = val + return tensors + + raise ValueError(f"Model {model_id} does not have any safetensor files") + + @classmethod + def get_list_tensors(cls, url: str) -> dict[str, RemoteTensor]: + """ + Get list of tensors from a remote safetensor file. + + Returns a dictionary of tensor names and their metadata. + Each tensor is represented as a tuple of (dtype, shape, offset_start, size) + """ + metadata, data_start_offset = cls.get_metadata(url) + res: dict[str, RemoteTensor] = {} + + for name, meta in metadata.items(): + if name == "__metadata__": + continue + if not isinstance(meta, dict): + raise ValueError(f"Invalid metadata for tensor '{name}': {meta}") + try: + dtype = meta["dtype"] + shape = meta["shape"] + offset_start_relative, offset_end_relative = meta["data_offsets"] + size = offset_end_relative - offset_start_relative + offset_start = data_start_offset + offset_start_relative + res[name] = RemoteTensor(dtype=dtype, shape=tuple(shape), offset_start=offset_start, size=size, url=url) + except KeyError as e: + raise ValueError(f"Missing key in metadata for tensor '{name}': {e}, meta = {meta}") + + return res + + @classmethod + def get_metadata(cls, url: str) -> tuple[dict, int]: + """ + Get JSON metadata from a remote safetensor file. + + Returns tuple of (metadata, data_start_offset) + """ + # Request first 5MB of the file (hopefully enough for metadata) + read_size = 5 * 1024 * 1024 + raw_data = cls.get_data_by_range(url, 0, read_size) + + # Parse header + # First 8 bytes contain the metadata length as u64 little-endian + if len(raw_data) < 8: + raise ValueError("Not enough data to read metadata size") + metadata_length = int.from_bytes(raw_data[:8], byteorder='little') + + # Calculate the data start offset + data_start_offset = 8 + metadata_length + alignment = SafetensorRemote.ALIGNMENT + if data_start_offset % alignment != 0: + data_start_offset += alignment - (data_start_offset % alignment) + + # Check if we have enough data to read the metadata + if len(raw_data) < 8 + metadata_length: + raise ValueError(f"Could not read complete metadata. Need {8 + metadata_length} bytes, got {len(raw_data)}") + + # Extract metadata bytes and parse as JSON + metadata_bytes = raw_data[8:8 + metadata_length] + metadata_str = metadata_bytes.decode('utf-8') + try: + metadata = json.loads(metadata_str) + return metadata, data_start_offset + except json.JSONDecodeError as e: + raise ValueError(f"Failed to parse safetensor metadata as JSON: {e}") + + @classmethod + def get_data_by_range(cls, url: str, start: int, size: int = -1) -> bytes: + """ + Get raw byte data from a remote file by range. + If size is not specified, it will read the entire file. + """ + import requests + from urllib.parse import urlparse + + parsed_url = urlparse(url) + if not parsed_url.scheme or not parsed_url.netloc: + raise ValueError(f"Invalid URL: {url}") + + headers = cls._get_request_headers() + if size > -1: + headers["Range"] = f"bytes={start}-{start + size}" + response = requests.get(url, allow_redirects=True, headers=headers) + response.raise_for_status() + + # Get raw byte data + return response.content[:size] + + @classmethod + def check_file_exist(cls, url: str) -> bool: + """ + Check if a file exists at the given URL. + Returns True if the file exists, False otherwise. + """ + import requests + from urllib.parse import urlparse + + parsed_url = urlparse(url) + if not parsed_url.scheme or not parsed_url.netloc: + raise ValueError(f"Invalid URL: {url}") + + try: + headers = cls._get_request_headers() + headers["Range"] = "bytes=0-0" + response = requests.head(url, allow_redirects=True, headers=headers) + # Success (2xx) or redirect (3xx) + return 200 <= response.status_code < 400 + except requests.RequestException: + return False + + @classmethod + def _get_request_headers(cls) -> dict[str, str]: + """Prepare common headers for requests.""" + headers = {"User-Agent": "convert_hf_to_gguf"} + if os.environ.get("HF_TOKEN"): + headers["Authorization"] = f"Bearer {os.environ['HF_TOKEN']}" + return headers diff --git a/gguf-py/gguf/vocab.py b/gguf-py/gguf/vocab.py index f2645f921..cca097986 100644 --- a/gguf-py/gguf/vocab.py +++ b/gguf-py/gguf/vocab.py @@ -127,7 +127,7 @@ class SpecialVocab: self.merges = merges elif isinstance(merges[0], list) and len(merges[0]) == 2 and isinstance(merges[0][0], str): # New format since transformers 4.45 to support spaces in merges - # ref: https://github.com/ggerganov/llama.cpp/issues/9692 + # ref: https://github.com/ggml-org/llama.cpp/issues/9692 # TODO: internally store as the new format instead of converting to old if any(' ' in s for pair in merges for s in pair): logger.warning(f'Spaces in merges detected, encoding as {chr(ord(" ") + 256)!r}') @@ -154,7 +154,12 @@ class SpecialVocab: return True with open(tokenizer_config_file, encoding = 'utf-8') as f: tokenizer_config = json.load(f) - chat_template = tokenizer_config.get('chat_template') + chat_template_alt = None + chat_template_file = path / 'chat_template.json' + if chat_template_file.is_file(): + with open(chat_template_file, encoding = 'utf-8') as f: + chat_template_alt = json.load(f).get('chat_template') + chat_template = tokenizer_config.get('chat_template', chat_template_alt) if chat_template is None or isinstance(chat_template, (str, list)): self.chat_template = chat_template else: diff --git a/gguf-py/pyproject.toml b/gguf-py/pyproject.toml index 78c6baa64..d214e8720 100644 --- a/gguf-py/pyproject.toml +++ b/gguf-py/pyproject.toml @@ -1,6 +1,6 @@ [tool.poetry] name = "gguf" -version = "0.15.0" +version = "0.16.0" description = "Read and write ML models in GGUF for GGML" authors = ["GGML "] packages = [ @@ -9,7 +9,7 @@ packages = [ ] readme = "README.md" homepage = "https://ggml.ai" -repository = "https://github.com/ggerganov/llama.cpp" +repository = "https://github.com/ggml-org/llama.cpp" keywords = ["ggml", "gguf", "llama.cpp"] classifiers = [ "Programming Language :: Python :: 3", diff --git a/grammars/README.md b/grammars/README.md index 976954091..935213f5c 100644 --- a/grammars/README.md +++ b/grammars/README.md @@ -98,7 +98,7 @@ This guide provides a brief overview. Check out the GBNF files in this directory ## Troubleshooting -Grammars currently have performance gotchas (see https://github.com/ggerganov/llama.cpp/issues/4218). +Grammars currently have performance gotchas (see https://github.com/ggml-org/llama.cpp/issues/4218). ### Efficient optional repetitions @@ -126,7 +126,7 @@ You can use GBNF grammars: - in CLI, with [examples/json_schema_to_grammar.py](../examples/json_schema_to_grammar.py) - in JavaScript with [json-schema-to-grammar.mjs](../examples/server/public_legacy/json-schema-to-grammar.mjs) (this is used by the [server](../examples/server)'s Web UI) -Take a look at [tests](../tests/test-json-schema-to-grammar.cpp) to see which features are likely supported (you'll also find usage examples in https://github.com/ggerganov/llama.cpp/pull/5978, https://github.com/ggerganov/llama.cpp/pull/6659 & https://github.com/ggerganov/llama.cpp/pull/6555). +Take a look at [tests](../tests/test-json-schema-to-grammar.cpp) to see which features are likely supported (you'll also find usage examples in https://github.com/ggml-org/llama.cpp/pull/5978, https://github.com/ggml-org/llama.cpp/pull/6659 & https://github.com/ggml-org/llama.cpp/pull/6555). ```bash llama-cli \ @@ -185,10 +185,10 @@ Here is also a list of known limitations (contributions welcome): - `additionalProperties` defaults to `false` (produces faster grammars + reduces hallucinations). - `"additionalProperties": true` may produce keys that contain unescaped newlines. - Unsupported features are skipped silently. It is currently advised to use the command-line Python converter (see above) to see any warnings, and to inspect the resulting grammar / test it w/ [llama-gbnf-validator](../examples/gbnf-validator/gbnf-validator.cpp). -- Can't mix `properties` w/ `anyOf` / `oneOf` in the same type (https://github.com/ggerganov/llama.cpp/issues/7703) +- Can't mix `properties` w/ `anyOf` / `oneOf` in the same type (https://github.com/ggml-org/llama.cpp/issues/7703) - [prefixItems](https://json-schema.org/draft/2020-12/json-schema-core#name-prefixitems) is broken (but [items](https://json-schema.org/draft/2020-12/json-schema-core#name-items) works) - `minimum`, `exclusiveMinimum`, `maximum`, `exclusiveMaximum`: only supported for `"type": "integer"` for now, not `number` -- Nested `$ref`s are broken (https://github.com/ggerganov/llama.cpp/issues/8073) +- Nested `$ref`s are broken (https://github.com/ggml-org/llama.cpp/issues/8073) - [pattern](https://json-schema.org/draft/2020-12/json-schema-validation#name-pattern)s must start with `^` and end with `$` - Remote `$ref`s not supported in the C++ version (Python & JavaScript versions fetch https refs) - `string` [formats](https://json-schema.org/draft/2020-12/json-schema-validation#name-defined-formats) lack `uri`, `email` diff --git a/include/llama.h b/include/llama.h index 3784f7d39..5fd99f4c8 100644 --- a/include/llama.h +++ b/include/llama.h @@ -60,6 +60,7 @@ extern "C" { struct llama_model; struct llama_context; struct llama_sampler; + struct llama_kv_cache; typedef int32_t llama_pos; typedef int32_t llama_token; @@ -105,6 +106,11 @@ extern "C" { LLAMA_VOCAB_PRE_TYPE_CHAMELEON = 26, LLAMA_VOCAB_PRE_TYPE_MINERVA = 27, LLAMA_VOCAB_PRE_TYPE_DEEPSEEK3_LLM = 28, + LLAMA_VOCAB_PRE_TYPE_GPT4O = 29, + LLAMA_VOCAB_PRE_TYPE_SUPERBPE = 30, + LLAMA_VOCAB_PRE_TYPE_TRILLION = 31, + LLAMA_VOCAB_PRE_TYPE_BAILINGMOE = 32, + LLAMA_VOCAB_PRE_TYPE_LLAMA4 = 33, }; enum llama_rope_type { @@ -213,7 +219,7 @@ extern "C" { LLAMA_SPLIT_MODE_ROW = 2, // split layers and KV across GPUs, use tensor parallelism if supported }; - // TODO: simplify (https://github.com/ggerganov/llama.cpp/pull/9294#pullrequestreview-2286561979) + // TODO: simplify (https://github.com/ggml-org/llama.cpp/pull/9294#pullrequestreview-2286561979) typedef struct llama_token_data { llama_token id; // token id float logit; // log-odds of the token @@ -275,10 +281,18 @@ extern "C" { }; }; + struct llama_model_tensor_buft_override { + const char * pattern; + ggml_backend_buffer_type_t buft; + }; + struct llama_model_params { // NULL-terminated list of devices to use for offloading (if NULL, all available devices are used) ggml_backend_dev_t * devices; + // NULL-terminated list of buffer types to use for tensors that match a pattern + const struct llama_model_tensor_buft_override * tensor_buft_overrides; + int32_t n_gpu_layers; // number of layers to store in VRAM enum llama_split_mode split_mode; // how to split the model across multiple GPUs @@ -307,7 +321,7 @@ extern "C" { }; // NOTE: changing the default values of parameters marked as [EXPERIMENTAL] may cause crashes or incorrect results in certain configurations - // https://github.com/ggerganov/llama.cpp/pull/7544 + // https://github.com/ggml-org/llama.cpp/pull/7544 struct llama_context_params { uint32_t n_ctx; // text context, 0 = from model uint32_t n_batch; // logical maximum batch size that can be submitted to llama_decode @@ -320,7 +334,7 @@ extern "C" { enum llama_pooling_type pooling_type; // whether to pool (sum) embedding results by sequence id enum llama_attention_type attention_type; // attention type to use for embeddings - // ref: https://github.com/ggerganov/llama.cpp/pull/2054 + // ref: https://github.com/ggml-org/llama.cpp/pull/2054 float rope_freq_base; // RoPE base frequency, 0 = from model float rope_freq_scale; // RoPE frequency scaling factor, 0 = from model float yarn_ext_factor; // YaRN extrapolation mix factor, negative = from model @@ -385,7 +399,7 @@ extern "C" { struct llama_adapter_lora; // Helpers for getting default parameters - // TODO: update API to start accepting pointers to params structs (https://github.com/ggerganov/llama.cpp/discussions/9172) + // TODO: update API to start accepting pointers to params structs (https://github.com/ggml-org/llama.cpp/discussions/9172) LLAMA_API struct llama_model_params llama_model_default_params(void); LLAMA_API struct llama_context_params llama_context_default_params(void); LLAMA_API struct llama_sampler_chain_params llama_sampler_chain_default_params(void); @@ -468,7 +482,8 @@ extern "C" { DEPRECATED(LLAMA_API int32_t llama_n_vocab (const struct llama_vocab * vocab), "use llama_vocab_n_tokens instead"); LLAMA_API const struct llama_model * llama_get_model (const struct llama_context * ctx); - LLAMA_API enum llama_pooling_type llama_pooling_type(const struct llama_context * ctx); + LLAMA_API struct llama_kv_cache * llama_get_kv_self ( struct llama_context * ctx); + LLAMA_API enum llama_pooling_type llama_pooling_type(const struct llama_context * ctx); // TODO: rename to llama_get_pooling_type LLAMA_API const struct llama_vocab * llama_model_get_vocab(const struct llama_model * model); LLAMA_API enum llama_rope_type llama_model_rope_type(const struct llama_model * model); @@ -477,6 +492,7 @@ extern "C" { LLAMA_API int32_t llama_model_n_embd (const struct llama_model * model); LLAMA_API int32_t llama_model_n_layer (const struct llama_model * model); LLAMA_API int32_t llama_model_n_head (const struct llama_model * model); + LLAMA_API int32_t llama_model_n_head_kv (const struct llama_model * model); // Get the model's RoPE frequency scaling factor LLAMA_API float llama_model_rope_freq_scale_train(const struct llama_model * model); @@ -584,7 +600,7 @@ extern "C" { // KV cache // - // TODO: remove llama_kv_cache_view_* API + // TODO: start using struct llama_kv_cache // Information associated with an individual cell in the KV cache view. struct llama_kv_cache_view_cell { @@ -639,13 +655,19 @@ extern "C" { // Returns the number of tokens in the KV cache (slow, use only for debug) // If a KV cell has multiple sequences assigned to it, it will be counted multiple times - LLAMA_API int32_t llama_get_kv_cache_token_count(const struct llama_context * ctx); + LLAMA_API int32_t llama_kv_self_n_tokens(const struct llama_context * ctx); + + DEPRECATED(LLAMA_API int32_t llama_get_kv_cache_token_count(const struct llama_context * ctx), + "use llama_kv_self_n_tokens instead"); // Returns the number of used KV cells (i.e. have at least one sequence assigned to them) - LLAMA_API int32_t llama_get_kv_cache_used_cells(const struct llama_context * ctx); + LLAMA_API int32_t llama_kv_self_used_cells(const struct llama_context * ctx); + + DEPRECATED(LLAMA_API int32_t llama_get_kv_cache_used_cells(const struct llama_context * ctx), + "use llama_kv_self_used_cells instead"); // Clear the KV cache - both cell info is erased and KV data is zeroed - LLAMA_API void llama_kv_cache_clear( + LLAMA_API void llama_kv_self_clear( struct llama_context * ctx); // Removes all tokens that belong to the specified sequence and have positions in [p0, p1) @@ -653,7 +675,7 @@ extern "C" { // seq_id < 0 : match any sequence // p0 < 0 : [0, p1] // p1 < 0 : [p0, inf) - LLAMA_API bool llama_kv_cache_seq_rm( + LLAMA_API bool llama_kv_self_seq_rm( struct llama_context * ctx, llama_seq_id seq_id, llama_pos p0, @@ -663,7 +685,7 @@ extern "C" { // Note that this does not allocate extra KV cache memory - it simply assigns the tokens to the new sequence // p0 < 0 : [0, p1] // p1 < 0 : [p0, inf) - LLAMA_API void llama_kv_cache_seq_cp( + LLAMA_API void llama_kv_self_seq_cp( struct llama_context * ctx, llama_seq_id seq_id_src, llama_seq_id seq_id_dst, @@ -671,17 +693,17 @@ extern "C" { llama_pos p1); // Removes all tokens that do not belong to the specified sequence - LLAMA_API void llama_kv_cache_seq_keep( + LLAMA_API void llama_kv_self_seq_keep( struct llama_context * ctx, llama_seq_id seq_id); // Adds relative position "delta" to all tokens that belong to the specified sequence and have positions in [p0, p1) // If the KV cache is RoPEd, the KV data is updated accordingly: // - lazily on next llama_decode() - // - explicitly with llama_kv_cache_update() + // - explicitly with llama_kv_self_update() // p0 < 0 : [0, p1] // p1 < 0 : [p0, inf) - LLAMA_API void llama_kv_cache_seq_add( + LLAMA_API void llama_kv_self_seq_add( struct llama_context * ctx, llama_seq_id seq_id, llama_pos p0, @@ -691,10 +713,10 @@ extern "C" { // Integer division of the positions by factor of `d > 1` // If the KV cache is RoPEd, the KV data is updated accordingly: // - lazily on next llama_decode() - // - explicitly with llama_kv_cache_update() + // - explicitly with llama_kv_self_update() // p0 < 0 : [0, p1] // p1 < 0 : [p0, inf) - LLAMA_API void llama_kv_cache_seq_div( + LLAMA_API void llama_kv_self_seq_div( struct llama_context * ctx, llama_seq_id seq_id, llama_pos p0, @@ -702,24 +724,76 @@ extern "C" { int d); // Returns the largest position present in the KV cache for the specified sequence - LLAMA_API llama_pos llama_kv_cache_seq_pos_max( + LLAMA_API llama_pos llama_kv_self_seq_pos_max( struct llama_context * ctx, - llama_seq_id seq_id); - - // TODO: the llama_kv_cache_defrag and llama_kv_cache_update API tightly couples llama_context with llama_kv_cache - // how to avoid this? + llama_seq_id seq_id); // Defragment the KV cache // This will be applied: // - lazily on next llama_decode() - // - explicitly with llama_kv_cache_update() - LLAMA_API void llama_kv_cache_defrag(struct llama_context * ctx); - - // Apply the KV cache updates (such as K-shifts, defragmentation, etc.) - LLAMA_API void llama_kv_cache_update(struct llama_context * ctx); + // - explicitly with llama_kv_self_update() + LLAMA_API void llama_kv_self_defrag(struct llama_context * ctx); // Check if the context supports KV cache shifting - LLAMA_API bool llama_kv_cache_can_shift(struct llama_context * ctx); + LLAMA_API bool llama_kv_self_can_shift(const struct llama_context * ctx); + + // Apply the KV cache updates (such as K-shifts, defragmentation, etc.) + LLAMA_API void llama_kv_self_update(struct llama_context * ctx); + + DEPRECATED(LLAMA_API void llama_kv_cache_clear( + struct llama_context * ctx), + "use llama_kv_self_clear instead"); + + DEPRECATED(LLAMA_API bool llama_kv_cache_seq_rm( + struct llama_context * ctx, + llama_seq_id seq_id, + llama_pos p0, + llama_pos p1), + "use llama_kv_self_seq_rm instead"); + + DEPRECATED(LLAMA_API void llama_kv_cache_seq_cp( + struct llama_context * ctx, + llama_seq_id seq_id_src, + llama_seq_id seq_id_dst, + llama_pos p0, + llama_pos p1), + "use llama_kv_self_seq_cp instead"); + + DEPRECATED(LLAMA_API void llama_kv_cache_seq_keep( + struct llama_context * ctx, + llama_seq_id seq_id), + "use llama_kv_self_seq_keep instead"); + + DEPRECATED(LLAMA_API void llama_kv_cache_seq_add( + struct llama_context * ctx, + llama_seq_id seq_id, + llama_pos p0, + llama_pos p1, + llama_pos delta), + "use llama_kv_self_seq_add instead"); + + DEPRECATED(LLAMA_API void llama_kv_cache_seq_div( + struct llama_context * ctx, + llama_seq_id seq_id, + llama_pos p0, + llama_pos p1, + int d), + "use llama_kv_self_seq_div instead"); + + DEPRECATED(LLAMA_API llama_pos llama_kv_cache_seq_pos_max( + struct llama_context * ctx, + llama_seq_id seq_id), + "use llama_kv_self_seq_pos_max instead"); + + DEPRECATED(LLAMA_API void llama_kv_cache_defrag(struct llama_context * ctx), + "use llama_kv_self_defrag instead"); + + DEPRECATED(LLAMA_API bool llama_kv_cache_can_shift(const struct llama_context * ctx), + "use llama_kv_self_can_shift instead"); + + DEPRECATED(LLAMA_API void llama_kv_cache_update(struct llama_context * ctx), + "use llama_kv_self_update instead"); + // // State / sessions @@ -883,6 +957,10 @@ extern "C" { // If set to true, the model will only attend to the past tokens LLAMA_API void llama_set_causal_attn(struct llama_context * ctx, bool causal_attn); + // Set whether the model is in warmup mode or not + // If true, all model tensors are activated during llama_decode() to load and cache their weights. + LLAMA_API void llama_set_warmup(struct llama_context * ctx, bool warmup); + // Set abort callback LLAMA_API void llama_set_abort_callback(struct llama_context * ctx, ggml_abort_callback abort_callback, void * abort_callback_data); @@ -1040,7 +1118,7 @@ extern "C" { /// Apply chat template. Inspired by hf apply_chat_template() on python. /// Both "model" and "custom_template" are optional, but at least one is required. "custom_template" has higher precedence than "model" - /// NOTE: This function does not use a jinja parser. It only support a pre-defined list of template. See more: https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template + /// NOTE: This function does not use a jinja parser. It only support a pre-defined list of template. See more: https://github.com/ggml-org/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template /// @param tmpl A Jinja template to use for this chat. If this is nullptr, the model’s default chat template will be used instead. /// @param chat Pointer to a list of multiple llama_chat_message /// @param n_msg Number of llama_chat_message in this chat @@ -1149,7 +1227,7 @@ extern "C" { /// @details Sorts candidate tokens by their logits in descending order and calculate probabilities based on logits. /// NOTE: Avoid using on the full vocabulary as the sorting can become slow. For example, apply top-k or top-p sampling first. DEPRECATED(LLAMA_API struct llama_sampler * llama_sampler_init_softmax (void), - "will be removed in the future (see https://github.com/ggerganov/llama.cpp/pull/9896#discussion_r1800920915)"); + "will be removed in the future (see https://github.com/ggml-org/llama.cpp/pull/9896#discussion_r1800920915)"); /// @details Top-K sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751 LLAMA_API struct llama_sampler * llama_sampler_init_top_k (int32_t k); @@ -1157,7 +1235,7 @@ extern "C" { /// @details Nucleus sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751 LLAMA_API struct llama_sampler * llama_sampler_init_top_p (float p, size_t min_keep); - /// @details Minimum P sampling as described in https://github.com/ggerganov/llama.cpp/pull/3841 + /// @details Minimum P sampling as described in https://github.com/ggml-org/llama.cpp/pull/3841 LLAMA_API struct llama_sampler * llama_sampler_init_min_p (float p, size_t min_keep); /// @details Locally Typical Sampling implementation described in the paper https://arxiv.org/abs/2202.00666. @@ -1172,6 +1250,9 @@ extern "C" { /// @details XTC sampler as described in https://github.com/oobabooga/text-generation-webui/pull/6335 LLAMA_API struct llama_sampler * llama_sampler_init_xtc (float p, float t, size_t min_keep, uint32_t seed); + /// @details Top n sigma sampling as described in academic paper "Top-nσ: Not All Logits Are You Need" https://arxiv.org/pdf/2411.07641 + LLAMA_API struct llama_sampler * llama_sampler_init_top_n_sigma(float n); + /// @details Mirostat 1.0 algorithm described in the paper https://arxiv.org/abs/2007.14966. Uses tokens instead of words. /// @param candidates A vector of `llama_token_data` containing the candidate tokens, their probabilities (p), and log-odds (logit) for the current position in the generated text. /// @param tau The target cross-entropy (or surprise) value you want to achieve for the generated text. A higher value corresponds to more surprising or less predictable text, while a lower value corresponds to less surprising or more predictable text. @@ -1195,22 +1276,38 @@ extern "C" { float tau, float eta); + /// @details Intializes a GBNF grammar, see grammars/README.md for details. + /// @param vocab The vocabulary that this grammar will be used with. + /// @param grammar_str The production rules for the grammar, encoded as a string. Returns an empty grammar if empty. Returns NULL if parsing of grammar_str fails. + /// @param grammar_root The name of the start symbol for the grammar. LLAMA_API struct llama_sampler * llama_sampler_init_grammar( const struct llama_vocab * vocab, const char * grammar_str, const char * grammar_root); - /// @details Lazy grammar sampler, introduced in https://github.com/ggerganov/llama.cpp/pull/9639 - /// @param trigger_words A list of words that will trigger the grammar sampler. This may be updated to a loose regex syntax (w/ ^) in a near future. - /// @param trigger_tokens A list of tokens that will trigger the grammar sampler. - LLAMA_API struct llama_sampler * llama_sampler_init_grammar_lazy( + DEPRECATED(LLAMA_API struct llama_sampler * llama_sampler_init_grammar_lazy( const struct llama_vocab * vocab, const char * grammar_str, const char * grammar_root, const char ** trigger_words, size_t num_trigger_words, const llama_token * trigger_tokens, - size_t num_trigger_tokens); + size_t num_trigger_tokens), + "use llama_sampler_init_grammar_lazy_patterns instead"); + + + /// @details Lazy grammar sampler, introduced in https://github.com/ggml-org/llama.cpp/pull/9639 + /// @param trigger_patterns A list of patterns that will trigger the grammar sampler. Pattern will be matched from the start of the generation output, and grammar sampler will be fed content starting from its first match group. + /// @param trigger_tokens A list of tokens that will trigger the grammar sampler. Grammar sampler will be fed content starting from the trigger token included. + LLAMA_API struct llama_sampler * llama_sampler_init_grammar_lazy_patterns( + const struct llama_vocab * vocab, + const char * grammar_str, + const char * grammar_root, + const char ** trigger_patterns, + size_t num_trigger_patterns, + const llama_token * trigger_tokens, + size_t num_trigger_tokens); + /// NOTE: Avoid using on the full vocabulary as searching for repeated tokens can become slow. For example, apply top-k or top-p sampling first. LLAMA_API struct llama_sampler * llama_sampler_init_penalties( diff --git a/licenses/LICENSE-curl b/licenses/LICENSE-curl new file mode 100644 index 000000000..da9c03825 --- /dev/null +++ b/licenses/LICENSE-curl @@ -0,0 +1,9 @@ +Copyright (c) 1996 - 2025, Daniel Stenberg, daniel@haxx.se, and many contributors, see the THANKS file. + +All rights reserved. + +Permission to use, copy, modify, and distribute this software for any purpose with or without fee is hereby granted, provided that the above copyright notice and this permission notice appear in all copies. + +THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT OF THIRD PARTY RIGHTS. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + +Except as contained in this notice, the name of a copyright holder shall not be used in advertising or otherwise to promote the sale, use or other dealings in this Software without prior written authorization of the copyright holder. diff --git a/licenses/LICENSE-httplib b/licenses/LICENSE-httplib new file mode 100644 index 000000000..47c418e07 --- /dev/null +++ b/licenses/LICENSE-httplib @@ -0,0 +1,21 @@ +The MIT License (MIT) + +Copyright (c) 2017 yhirose + +Permission is hereby granted, free of charge, to any person obtaining a copy +of this software and associated documentation files (the "Software"), to deal +in the Software without restriction, including without limitation the rights +to use, copy, modify, merge, publish, distribute, sublicense, and/or sell +copies of the Software, and to permit persons to whom the Software is +furnished to do so, subject to the following conditions: + +The above copyright notice and this permission notice shall be included in all +copies or substantial portions of the Software. + +THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, +OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE +SOFTWARE. diff --git a/licenses/LICENSE-jsonhpp b/licenses/LICENSE-jsonhpp new file mode 100644 index 000000000..b5a10275c --- /dev/null +++ b/licenses/LICENSE-jsonhpp @@ -0,0 +1,21 @@ +MIT License + +Copyright (c) 2013-2025 Niels Lohmann + +Permission is hereby granted, free of charge, to any person obtaining a copy +of this software and associated documentation files (the "Software"), to deal +in the Software without restriction, including without limitation the rights +to use, copy, modify, merge, publish, distribute, sublicense, and/or sell +copies of the Software, and to permit persons to whom the Software is +furnished to do so, subject to the following conditions: + +The above copyright notice and this permission notice shall be included in all +copies or substantial portions of the Software. + +THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, +OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE +SOFTWARE. diff --git a/examples/run/linenoise.cpp/LICENSE b/licenses/LICENSE-linenoise similarity index 100% rename from examples/run/linenoise.cpp/LICENSE rename to licenses/LICENSE-linenoise diff --git a/media/llama1-logo.svg b/media/llama1-logo.svg new file mode 100644 index 000000000..e080481fa --- /dev/null +++ b/media/llama1-logo.svg @@ -0,0 +1,34 @@ + + + + + + + + + + + + + + + + + + + + + + \ No newline at end of file diff --git a/models/ggml-vocab-gpt-4o.gguf.inp b/models/ggml-vocab-gpt-4o.gguf.inp new file mode 100644 index 000000000..9baf7d77a --- /dev/null +++ b/models/ggml-vocab-gpt-4o.gguf.inp @@ -0,0 +1,112 @@ +ied 4 ½ months +__ggml_vocab_test__ +Führer +__ggml_vocab_test__ + +__ggml_vocab_test__ + +__ggml_vocab_test__ + +__ggml_vocab_test__ + +__ggml_vocab_test__ + +__ggml_vocab_test__ + + +__ggml_vocab_test__ + + + +__ggml_vocab_test__ + + + + +__ggml_vocab_test__ + + +__ggml_vocab_test__ +Hello world +__ggml_vocab_test__ + Hello world +__ggml_vocab_test__ +Hello World +__ggml_vocab_test__ + Hello World +__ggml_vocab_test__ + Hello World! +__ggml_vocab_test__ +Hello, world! +__ggml_vocab_test__ + Hello, world! +__ggml_vocab_test__ + this is 🦙.cpp +__ggml_vocab_test__ +w048 7tuijk dsdfhu +__ggml_vocab_test__ +нещо на Български +__ggml_vocab_test__ +កាន់តែពិសេសអាចខលចេញ +__ggml_vocab_test__ +🚀 (normal) 😶‍🌫️ (multiple emojis concatenated) ✅ (only emoji that has its own token) +__ggml_vocab_test__ +Hello +__ggml_vocab_test__ + Hello +__ggml_vocab_test__ + Hello +__ggml_vocab_test__ + Hello +__ggml_vocab_test__ + Hello +__ggml_vocab_test__ + Hello + Hello +__ggml_vocab_test__ + ( +__ggml_vocab_test__ + + = +__ggml_vocab_test__ +' era +__ggml_vocab_test__ +Hello, y'all! How are you 😁 ?我想在apple工作1314151天～ +__ggml_vocab_test__ +!!!!!! +__ggml_vocab_test__ +3 +__ggml_vocab_test__ +33 +__ggml_vocab_test__ +333 +__ggml_vocab_test__ +3333 +__ggml_vocab_test__ +33333 +__ggml_vocab_test__ +333333 +__ggml_vocab_test__ +3333333 +__ggml_vocab_test__ +33333333 +__ggml_vocab_test__ +333333333 +__ggml_vocab_test__ +Cửa Việt +__ggml_vocab_test__ + discards +__ggml_vocab_test__ + + + + + + + + + + + +🚀 (normal) 😶‍🌫️ (multiple emojis concatenated) ✅ 🦙🦙 3 33 333 3333 33333 333333 3333333 33333333 3.3 3..3 3...3 កាន់តែពិសេសអាច😁 ?我想在apple工作1314151天～ ------======= нещо на Български ''''''```````""""......!!!!!!?????? I've been 'told he's there, 'RE you sure? 'M not sure I'll make it, 'D you like some tea? We'Ve a'lL +__ggml_vocab_test__ diff --git a/models/ggml-vocab-gpt-4o.gguf.out b/models/ggml-vocab-gpt-4o.gguf.out new file mode 100644 index 000000000..478df726f --- /dev/null +++ b/models/ggml-vocab-gpt-4o.gguf.out @@ -0,0 +1,46 @@ + 1165 220 19 220 27124 5503 + 37 19194 259 + + 220 + 256 + 271 + 197 + 198 + 279 + 2499 + 2775 + 13225 2375 + 32949 2375 + 13225 5922 + 32949 5922 + 32949 5922 0 + 13225 11 2375 0 + 32949 11 2375 0 + 495 382 9552 99 247 13 17159 + 86 45404 220 22 10191 2852 22924 4750 6916 + 3907 53641 1235 185386 8118 + 11400 107516 15867 20804 22851 134178 77431 32010 104312 37984 16329 27751 89335 + 112927 222 350 14559 8 22861 114 2524 64364 104 15148 350 76466 166700 121942 780 8 91349 350 7393 74471 484 853 1617 2316 6602 8 + 13225 + 32949 + 220 32949 + 256 32949 + 271 32949 + 271 32949 198 271 32949 + 350 + 198 314 + 6 6837 + 13225 11 342 70653 0 3253 553 481 22861 223 1423 7522 18165 2178 34058 22369 16412 32999 16 867 8208 + 147475 + 18 + 2546 + 15517 + 15517 18 + 15517 2546 + 15517 15517 + 15517 15517 18 + 15517 15517 2546 + 15517 15517 15517 + 34 60213 53904 + 2960 3098 + 126470 25980 160432 16609 2775 4066 172261 19432 112927 222 350 14559 8 22861 114 2524 64364 104 15148 350 76466 166700 121942 780 8 91349 9552 99 247 4103 99 247 220 18 220 2546 220 15517 220 15517 18 220 15517 2546 220 15517 15517 220 15517 15517 18 220 15517 15517 2546 220 18 13 18 220 18 485 18 220 18 1008 18 44735 107516 15867 20804 22851 134178 77431 32010 104312 156437 1423 7522 18165 2178 34058 22369 16412 32999 16 867 8208 105024 106657 1967 53641 1235 185386 8118 22434 39336 26178 26178 168394 194663 27271 147475 25883 6961 9790 1339 461 83 1280 19016 1354 11 461 1099 481 3239 30 461 44 625 3239 17291 1520 480 11 461 35 481 1299 1236 17966 30 1416 6 27493 261 54602 43 diff --git a/models/ggml-vocab-llama4.gguf.inp b/models/ggml-vocab-llama4.gguf.inp new file mode 100644 index 000000000..9baf7d77a --- /dev/null +++ b/models/ggml-vocab-llama4.gguf.inp @@ -0,0 +1,112 @@ +ied 4 ½ months +__ggml_vocab_test__ +Führer +__ggml_vocab_test__ + +__ggml_vocab_test__ + +__ggml_vocab_test__ + +__ggml_vocab_test__ + +__ggml_vocab_test__ + +__ggml_vocab_test__ + + +__ggml_vocab_test__ + + + +__ggml_vocab_test__ + + + + +__ggml_vocab_test__ + + +__ggml_vocab_test__ +Hello world +__ggml_vocab_test__ + Hello world +__ggml_vocab_test__ +Hello World +__ggml_vocab_test__ + Hello World +__ggml_vocab_test__ + Hello World! +__ggml_vocab_test__ +Hello, world! +__ggml_vocab_test__ + Hello, world! +__ggml_vocab_test__ + this is 🦙.cpp +__ggml_vocab_test__ +w048 7tuijk dsdfhu +__ggml_vocab_test__ +нещо на Български +__ggml_vocab_test__ +កាន់តែពិសេសអាចខលចេញ +__ggml_vocab_test__ +🚀 (normal) 😶‍🌫️ (multiple emojis concatenated) ✅ (only emoji that has its own token) +__ggml_vocab_test__ +Hello +__ggml_vocab_test__ + Hello +__ggml_vocab_test__ + Hello +__ggml_vocab_test__ + Hello +__ggml_vocab_test__ + Hello +__ggml_vocab_test__ + Hello + Hello +__ggml_vocab_test__ + ( +__ggml_vocab_test__ + + = +__ggml_vocab_test__ +' era +__ggml_vocab_test__ +Hello, y'all! How are you 😁 ?我想在apple工作1314151天～ +__ggml_vocab_test__ +!!!!!! +__ggml_vocab_test__ +3 +__ggml_vocab_test__ +33 +__ggml_vocab_test__ +333 +__ggml_vocab_test__ +3333 +__ggml_vocab_test__ +33333 +__ggml_vocab_test__ +333333 +__ggml_vocab_test__ +3333333 +__ggml_vocab_test__ +33333333 +__ggml_vocab_test__ +333333333 +__ggml_vocab_test__ +Cửa Việt +__ggml_vocab_test__ + discards +__ggml_vocab_test__ + + + + + + + + + + + +🚀 (normal) 😶‍🌫️ (multiple emojis concatenated) ✅ 🦙🦙 3 33 333 3333 33333 333333 3333333 33333333 3.3 3..3 3...3 កាន់តែពិសេសអាច😁 ?我想在apple工作1314151天～ ------======= нещо на Български ''''''```````""""......!!!!!!?????? I've been 'told he's there, 'RE you sure? 'M not sure I'll make it, 'D you like some tea? We'Ve a'lL +__ggml_vocab_test__ diff --git a/models/ggml-vocab-llama4.gguf.out b/models/ggml-vocab-llama4.gguf.out new file mode 100644 index 000000000..7ca46ce59 --- /dev/null +++ b/models/ggml-vocab-llama4.gguf.out @@ -0,0 +1,46 @@ + 1190 220 32 220 18215 7112 + 50 16800 258 + + 220 + 256 + 277 + 197 + 198 + 368 + 2946 + 3271 + 19873 3817 + 39715 3817 + 19873 7353 + 39715 7353 + 39715 7353 13 + 19873 24 3817 13 + 39715 24 3817 13 + 544 373 9522 112 247 26 36315 + 99 39923 220 35 9607 21498 21470 3679 9433 + 1595 7653 633 79829 34051 1636 + 8755 102595 115960 21125 148305 96819 102816 39048 14105 22528 160234 + 114590 222 330 14879 21 51358 127 12817 93293 117 24204 330 68239 881 120327 170428 21 89101 330 7384 88230 511 947 1492 3742 7233 21 + 19873 + 39715 + 220 39715 + 256 39715 + 277 39715 + 277 39715 198 277 39715 + 330 + 198 319 + 19 7359 + 19873 24 386 87799 13 2403 583 650 51358 223 1663 155736 1522 42056 7544 13336 28785 29 4412 20645 + 17931 4959 + 31 + 1922 + 12325 + 12325 31 + 12325 1922 + 12325 12325 + 12325 12325 31 + 12325 12325 1922 + 12325 12325 12325 + 47 19811 12077 + 3260 3579 + 198 7283 51499 191231 20192 3271 3322 9287 2143 17860 114590 222 330 14879 21 51358 127 12817 93293 117 24204 330 68239 881 120327 170428 21 89101 9522 112 247 172394 247 220 31 220 1922 220 12325 220 12325 31 220 12325 1922 220 12325 12325 220 12325 12325 31 220 12325 12325 1922 220 31 26 31 220 31 396 31 220 31 1043 31 117131 102595 115960 21125 148305 96819 102816 80883 223 1663 155736 1522 42056 7544 13336 28785 29 4412 20645 79745 150278 117079 633 79829 34051 1636 25611 41990 109428 1488 91054 24072 17931 4959 29795 9296 16517 1806 481 96 1386 36633 1609 24 481 1109 650 5074 43 481 57 702 5074 27088 2170 536 24 481 48 650 1933 1696 30262 43 1665 19 32818 262 27236 56 diff --git a/models/templates/README.md b/models/templates/README.md new file mode 100644 index 000000000..e4fd104fc --- /dev/null +++ b/models/templates/README.md @@ -0,0 +1,22 @@ +These templates can be updated with the following commands: + +```bash +./scripts/get_chat_template.py CohereForAI/c4ai-command-r-plus tool_use > models/templates/CohereForAI-c4ai-command-r-plus-tool_use.jinja +./scripts/get_chat_template.py CohereForAI/c4ai-command-r7b-12-2024 default > models/templates/CohereForAI-c4ai-command-r7b-12-2024-default.jinja +./scripts/get_chat_template.py CohereForAI/c4ai-command-r7b-12-2024 rag > models/templates/CohereForAI-c4ai-command-r7b-12-2024-rag.jinja +./scripts/get_chat_template.py CohereForAI/c4ai-command-r7b-12-2024 tool_use > models/templates/CohereForAI-c4ai-command-r7b-12-2024-tool_use.jinja +./scripts/get_chat_template.py deepseek-ai/DeepSeek-R1-Distill-Llama-8B > models/templates/deepseek-ai-DeepSeek-R1-Distill-Llama-8B.jinja +./scripts/get_chat_template.py deepseek-ai/DeepSeek-R1-Distill-Qwen-32B > models/templates/deepseek-ai-DeepSeek-R1-Distill-Qwen-32B.jinja +./scripts/get_chat_template.py fireworks-ai/llama-3-firefunction-v2 > models/templates/fireworks-ai-llama-3-firefunction-v2.jinja +./scripts/get_chat_template.py google/gemma-2-2b-it > models/templates/google-gemma-2-2b-it.jinja +./scripts/get_chat_template.py meetkai/functionary-medium-v3.1 > models/templates/meetkai-functionary-medium-v3.1.jinja +./scripts/get_chat_template.py meetkai/functionary-medium-v3.2 > models/templates/meetkai-functionary-medium-v3.2.jinja +./scripts/get_chat_template.py meta-llama/Llama-3.1-8B-Instruct > models/templates/meta-llama-Llama-3.1-8B-Instruct.jinja +./scripts/get_chat_template.py meta-llama/Llama-3.2-3B-Instruct > models/templates/meta-llama-Llama-3.2-3B-Instruct.jinja +./scripts/get_chat_template.py meta-llama/Llama-3.3-70B-Instruct > models/templates/meta-llama-Llama-3.3-70B-Instruct.jinja +./scripts/get_chat_template.py microsoft/Phi-3.5-mini-instruct > models/templates/microsoft-Phi-3.5-mini-instruct.jinja +./scripts/get_chat_template.py mistralai/Mistral-Nemo-Instruct-2407 > models/templates/mistralai-Mistral-Nemo-Instruct-2407.jinja +./scripts/get_chat_template.py NousResearch/Hermes-2-Pro-Llama-3-8B tool_use > models/templates/NousResearch-Hermes-2-Pro-Llama-3-8B-tool_use.jinja +./scripts/get_chat_template.py NousResearch/Hermes-3-Llama-3.1-8B tool_use > models/templates/NousResearch-Hermes-3-Llama-3.1-8B-tool_use.jinja +./scripts/get_chat_template.py Qwen/Qwen2.5-7B-Instruct > models/templates/Qwen-Qwen2.5-7B-Instruct.jinja +``` \ No newline at end of file diff --git a/models/templates/deepseek-ai-DeepSeek-R1-Distill-Llama-8B.jinja b/models/templates/deepseek-ai-DeepSeek-R1-Distill-Llama-8B.jinja index 02a1c3bce..c2066bd73 100644 --- a/models/templates/deepseek-ai-DeepSeek-R1-Distill-Llama-8B.jinja +++ b/models/templates/deepseek-ai-DeepSeek-R1-Distill-Llama-8B.jinja @@ -1 +1 @@ -{% if not add_generation_prompt is defined %}{% set add_generation_prompt = false %}{% endif %}{% set ns = namespace(is_first=false, is_tool=false, is_output_first=true, system_prompt='') %}{%- for message in messages %}{%- if message['role'] == 'system' %}{% set ns.system_prompt = message['content'] %}{%- endif %}{%- endfor %}{{bos_token}}{{ns.system_prompt}}{%- for message in messages %}{%- if message['role'] == 'user' %}{%- set ns.is_tool = false -%}{{'<｜User｜>' + message['content']}}{%- endif %}{%- if message['role'] == 'assistant' and message['content'] is none %}{%- set ns.is_tool = false -%}{%- for tool in message['tool_calls']%}{%- if not ns.is_first %}{{'<｜Assistant｜><｜tool▁calls▁begin｜><｜tool▁call▁begin｜>' + tool['type'] + '<｜tool▁sep｜>' + tool['function']['name'] + '\n' + '```json' + '\n' + tool['function']['arguments'] + '\n' + '```' + '<｜tool▁call▁end｜>'}}{%- set ns.is_first = true -%}{%- else %}{{'\n' + '<｜tool▁call▁begin｜>' + tool['type'] + '<｜tool▁sep｜>' + tool['function']['name'] + '\n' + '```json' + '\n' + tool['function']['arguments'] + '\n' + '```' + '<｜tool▁call▁end｜>'}}{{'<｜tool▁calls▁end｜><｜end▁of▁sentence｜>'}}{%- endif %}{%- endfor %}{%- endif %}{%- if message['role'] == 'assistant' and message['content'] is not none %}{%- if ns.is_tool %}{{'<｜tool▁outputs▁end｜>' + message['content'] + '<｜end▁of▁sentence｜>'}}{%- set ns.is_tool = false -%}{%- else %}{% set content = message['content'] %}{% if '' in content %}{% set content = content.split('')[-1] %}{% endif %}{{'<｜Assistant｜>' + content + '<｜end▁of▁sentence｜>'}}{%- endif %}{%- endif %}{%- if message['role'] == 'tool' %}{%- set ns.is_tool = true -%}{%- if ns.is_output_first %}{{'<｜tool▁outputs▁begin｜><｜tool▁output▁begin｜>' + message['content'] + '<｜tool▁output▁end｜>'}}{%- set ns.is_output_first = false %}{%- else %}{{'\n<｜tool▁output▁begin｜>' + message['content'] + '<｜tool▁output▁end｜>'}}{%- endif %}{%- endif %}{%- endfor -%}{% if ns.is_tool %}{{'<｜tool▁outputs▁end｜>'}}{% endif %}{% if add_generation_prompt and not ns.is_tool %}{{'<｜Assistant｜>'}}{% endif %} \ No newline at end of file +{% if not add_generation_prompt is defined %}{% set add_generation_prompt = false %}{% endif %}{% set ns = namespace(is_first=false, is_tool=false, is_output_first=true, system_prompt='') %}{%- for message in messages %}{%- if message['role'] == 'system' %}{% set ns.system_prompt = message['content'] %}{%- endif %}{%- endfor %}{{bos_token}}{{ns.system_prompt}}{%- for message in messages %}{%- if message['role'] == 'user' %}{%- set ns.is_tool = false -%}{{'<｜User｜>' + message['content']}}{%- endif %}{%- if message['role'] == 'assistant' and message['content'] is none %}{%- set ns.is_tool = false -%}{%- for tool in message['tool_calls']%}{%- if not ns.is_first %}{{'<｜Assistant｜><｜tool▁calls▁begin｜><｜tool▁call▁begin｜>' + tool['type'] + '<｜tool▁sep｜>' + tool['function']['name'] + '\n' + '```json' + '\n' + tool['function']['arguments'] + '\n' + '```' + '<｜tool▁call▁end｜>'}}{%- set ns.is_first = true -%}{%- else %}{{'\n' + '<｜tool▁call▁begin｜>' + tool['type'] + '<｜tool▁sep｜>' + tool['function']['name'] + '\n' + '```json' + '\n' + tool['function']['arguments'] + '\n' + '```' + '<｜tool▁call▁end｜>'}}{{'<｜tool▁calls▁end｜><｜end▁of▁sentence｜>'}}{%- endif %}{%- endfor %}{%- endif %}{%- if message['role'] == 'assistant' and message['content'] is not none %}{%- if ns.is_tool %}{{'<｜tool▁outputs▁end｜>' + message['content'] + '<｜end▁of▁sentence｜>'}}{%- set ns.is_tool = false -%}{%- else %}{% set content = message['content'] %}{% if '' in content %}{% set content = content.split('')[-1] %}{% endif %}{{'<｜Assistant｜>' + content + '<｜end▁of▁sentence｜>'}}{%- endif %}{%- endif %}{%- if message['role'] == 'tool' %}{%- set ns.is_tool = true -%}{%- if ns.is_output_first %}{{'<｜tool▁outputs▁begin｜><｜tool▁output▁begin｜>' + message['content'] + '<｜tool▁output▁end｜>'}}{%- set ns.is_output_first = false %}{%- else %}{{'\n<｜tool▁output▁begin｜>' + message['content'] + '<｜tool▁output▁end｜>'}}{%- endif %}{%- endif %}{%- endfor -%}{% if ns.is_tool %}{{'<｜tool▁outputs▁end｜>'}}{% endif %}{% if add_generation_prompt and not ns.is_tool %}{{'<｜Assistant｜>\n'}}{% endif %} \ No newline at end of file diff --git a/models/templates/deepseek-ai-DeepSeek-R1-Distill-Qwen-32B.jinja b/models/templates/deepseek-ai-DeepSeek-R1-Distill-Qwen-32B.jinja index 2ebfe7c1e..c2066bd73 100644 --- a/models/templates/deepseek-ai-DeepSeek-R1-Distill-Qwen-32B.jinja +++ b/models/templates/deepseek-ai-DeepSeek-R1-Distill-Qwen-32B.jinja @@ -1,56 +1 @@ -{% if not add_generation_prompt is defined %} -{% set add_generation_prompt = false %} -{% endif %} -{% set ns = namespace(is_first=false, is_tool=false, is_output_first=true, system_prompt='') %} -{%- for message in messages %} -{%- if message['role'] == 'system' %} -{% set ns.system_prompt = message['content'] %} -{%- endif %} -{%- endfor %} -{{bos_token}} -{{ns.system_prompt}} -{%- for message in messages %} -{%- if message['role'] == 'user' %} -{%- set ns.is_tool = false -%} -{{'<｜User｜>' + message['content']}} -{%- endif %} -{%- if message['role'] == 'assistant' and message['content'] is none %} -{%- set ns.is_tool = false -%} -{%- for tool in message['tool_calls']%} -{%- if not ns.is_first %} -{{'<｜Assistant｜><｜tool▁calls▁begin｜><｜tool▁call▁begin｜>' + tool['type'] + '<｜tool▁sep｜>' + tool['function']['name'] + '\n' + '```json' + '\n' + tool['function']['arguments'] + '\n' + '```' + '<｜tool▁call▁end｜>'}} -{%- set ns.is_first = true -%} -{%- else %} -{{'\n' + '<｜tool▁call▁begin｜>' + tool['type'] + '<｜tool▁sep｜>' + tool['function']['name'] + '\n' + '```json' + '\n' + tool['function']['arguments'] + '\n' + '```' + '<｜tool▁call▁end｜>'}} -{{'<｜tool▁calls▁end｜><｜end▁of▁sentence｜>'}} -{%- endif %} -{%- endfor %} -{%- endif %} -{%- if message['role'] == 'assistant' and message['content'] is not none %} -{%- if ns.is_tool %} -{{'<｜tool▁outputs▁end｜>' + message['content'] + '<｜end▁of▁sentence｜>'}} -{%- set ns.is_tool = false -%} -{%- else %} -{% set content = message['content'] %} -{% if '' in content %} -{% set content = content.split('')[-1] %} -{% endif %} -{{'<｜Assistant｜>' + content + '<｜end▁of▁sentence｜>'}} -{%- endif %} -{%- endif %} -{%- if message['role'] == 'tool' %} -{%- set ns.is_tool = true -%} -{%- if ns.is_output_first %} -{{'<｜tool▁outputs▁begin｜><｜tool▁output▁begin｜>' + message['content'] + '<｜tool▁output▁end｜>'}} -{%- set ns.is_output_first = false %} -{%- else %} -{{'\n<｜tool▁output▁begin｜>' + message['content'] + '<｜tool▁output▁end｜>'}} -{%- endif %} -{%- endif %} -{%- endfor -%} -{% if ns.is_tool %} -{{'<｜tool▁outputs▁end｜>'}} -{% endif %} -{% if add_generation_prompt and not ns.is_tool %} -{{'<｜Assistant｜>'}} -{% endif %} \ No newline at end of file +{% if not add_generation_prompt is defined %}{% set add_generation_prompt = false %}{% endif %}{% set ns = namespace(is_first=false, is_tool=false, is_output_first=true, system_prompt='') %}{%- for message in messages %}{%- if message['role'] == 'system' %}{% set ns.system_prompt = message['content'] %}{%- endif %}{%- endfor %}{{bos_token}}{{ns.system_prompt}}{%- for message in messages %}{%- if message['role'] == 'user' %}{%- set ns.is_tool = false -%}{{'<｜User｜>' + message['content']}}{%- endif %}{%- if message['role'] == 'assistant' and message['content'] is none %}{%- set ns.is_tool = false -%}{%- for tool in message['tool_calls']%}{%- if not ns.is_first %}{{'<｜Assistant｜><｜tool▁calls▁begin｜><｜tool▁call▁begin｜>' + tool['type'] + '<｜tool▁sep｜>' + tool['function']['name'] + '\n' + '```json' + '\n' + tool['function']['arguments'] + '\n' + '```' + '<｜tool▁call▁end｜>'}}{%- set ns.is_first = true -%}{%- else %}{{'\n' + '<｜tool▁call▁begin｜>' + tool['type'] + '<｜tool▁sep｜>' + tool['function']['name'] + '\n' + '```json' + '\n' + tool['function']['arguments'] + '\n' + '```' + '<｜tool▁call▁end｜>'}}{{'<｜tool▁calls▁end｜><｜end▁of▁sentence｜>'}}{%- endif %}{%- endfor %}{%- endif %}{%- if message['role'] == 'assistant' and message['content'] is not none %}{%- if ns.is_tool %}{{'<｜tool▁outputs▁end｜>' + message['content'] + '<｜end▁of▁sentence｜>'}}{%- set ns.is_tool = false -%}{%- else %}{% set content = message['content'] %}{% if '' in content %}{% set content = content.split('')[-1] %}{% endif %}{{'<｜Assistant｜>' + content + '<｜end▁of▁sentence｜>'}}{%- endif %}{%- endif %}{%- if message['role'] == 'tool' %}{%- set ns.is_tool = true -%}{%- if ns.is_output_first %}{{'<｜tool▁outputs▁begin｜><｜tool▁output▁begin｜>' + message['content'] + '<｜tool▁output▁end｜>'}}{%- set ns.is_output_first = false %}{%- else %}{{'\n<｜tool▁output▁begin｜>' + message['content'] + '<｜tool▁output▁end｜>'}}{%- endif %}{%- endif %}{%- endfor -%}{% if ns.is_tool %}{{'<｜tool▁outputs▁end｜>'}}{% endif %}{% if add_generation_prompt and not ns.is_tool %}{{'<｜Assistant｜>\n'}}{% endif %} \ No newline at end of file diff --git a/models/templates/llama-cpp-deepseek-r1.jinja b/models/templates/llama-cpp-deepseek-r1.jinja new file mode 100644 index 000000000..fcb1732eb --- /dev/null +++ b/models/templates/llama-cpp-deepseek-r1.jinja @@ -0,0 +1,76 @@ +{%- if not add_generation_prompt is defined -%} + {%- set add_generation_prompt = false -%} +{%- endif -%} +{%- set ns = namespace(is_first=false, is_tool_outputs=false, is_output_first=true, system_prompt='') -%} +{%- for message in messages -%} + {%- if message['role'] == 'system' -%} + {%- set ns.system_prompt = message['content'] -%} + {%- endif -%} +{%- endfor -%} +{{bos_token}} +{%- if tools %} +You can call any of the following function tools to satisfy the user's requests: {{tools | map(attribute='function') | tojson(indent=2)}} + +Example function tool call syntax: + +<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>function<｜tool▁sep｜>example_function_name +```json +{ + "arg1": "some_value" + ... +} +``` +<｜tool▁call▁end｜><｜tool▁calls▁end｜> + +{% endif -%} +{{ns.system_prompt}} +{%- macro flush_tool_outputs() -%} + {%- if ns.is_tool_outputs -%} + {{- '<｜tool▁outputs▁end｜><｜end▁of▁sentence｜>' -}} + {%- set ns.is_tool_outputs = false -%} + {%- endif -%} +{%- endmacro -%} +{{- flush_tool_outputs() -}} +{%- for message in messages -%} + {%- if message['role'] != 'tool' -%} + {{- flush_tool_outputs() -}} + {%- endif -%} + {%- if message['role'] == 'user' -%} + {{- '<｜User｜>' + message['content'] + '<｜end▁of▁sentence｜>' -}} + {%- endif -%} + {%- if message['role'] == 'assistant' and message['content'] is none -%} + {{- '<｜Assistant｜><｜tool▁calls▁begin｜>' -}} + {%- set ns.is_first = true -%} + {%- for tc in message['tool_calls'] -%} + {%- if ns.is_first -%} + {%- set ns.is_first = false -%} + {%- else -%} + {{- '\n' -}} + {%- endif -%} + {%- set tool_name = tc['function']['name'] -%} + {%- set tool_args = tc['function']['arguments'] -%} + {{- '<｜tool▁call▁begin｜>' + tc['type'] + '<｜tool▁sep｜>' + tool_name + '\n' + '```json' + '\n' + tool_args + '\n' + '```' + '<｜tool▁call▁end｜>' -}} + {%- endfor -%} + {{- '<｜tool▁calls▁end｜><｜end▁of▁sentence｜>' -}} + {%- endif -%} + {%- if message['role'] == 'assistant' and message['content'] is not none -%} + {{- flush_tool_outputs() -}} + {%- set content = message['content'] -%} + {%- if '' in content -%} + {%- set content = content.split('')[-1] -%} + {%- endif -%} + {{- '<｜Assistant｜>' + content + '<｜end▁of▁sentence｜>' -}} + {%- endif -%} + {%- if message['role'] == 'tool' -%} + {%- set ns.is_tool_outputs = true -%} + {%- if ns.is_output_first -%} + {{- '<｜tool▁outputs▁begin｜>' -}} + {%- set ns.is_output_first = false -%} + {%- endif -%} + {{- '\n<｜tool▁output▁begin｜>' + message['content'] + '<｜tool▁output▁end｜>' -}} + {%- endif -%} +{%- endfor -%} +{{- flush_tool_outputs() -}} +{%- if add_generation_prompt and not ns.is_tool_outputs -%} + {{- '<｜Assistant｜>\n' -}} +{%- endif -%} \ No newline at end of file diff --git a/pyproject.toml b/pyproject.toml index 84e71de6d..ed62264ba 100644 --- a/pyproject.toml +++ b/pyproject.toml @@ -5,7 +5,7 @@ description = "Scripts that ship with llama.cpp" authors = ["GGML "] readme = "README.md" homepage = "https://ggml.ai" -repository = "https://github.com/ggerganov/llama.cpp" +repository = "https://github.com/ggml-org/llama.cpp" keywords = ["ggml", "gguf", "llama.cpp"] packages = [{ include = "*.py", from = "." }] classifiers = [ diff --git a/requirements.txt b/requirements.txt index 98c53db81..cf3116c6c 100644 --- a/requirements.txt +++ b/requirements.txt @@ -11,3 +11,4 @@ -r ./requirements/requirements-convert_legacy_imatrix_to_gguf.txt -r ./requirements/requirements-convert_llama_ggml_to_gguf.txt -r ./requirements/requirements-convert_lora_to_gguf.txt +-r ./requirements/requirements-tool_bench.txt diff --git a/requirements/requirements-all.txt b/requirements/requirements-all.txt index 94de59d7e..439db8886 100644 --- a/requirements/requirements-all.txt +++ b/requirements/requirements-all.txt @@ -10,3 +10,4 @@ -r ./requirements-convert_hf_to_gguf_update.txt -r ./requirements-convert_legacy_llama.txt -r ./requirements-convert_llama_ggml_to_gguf.txt +-r ./requirements-tool_bench.txt diff --git a/requirements/requirements-tool_bench.txt b/requirements/requirements-tool_bench.txt new file mode 100644 index 000000000..b94521fc7 --- /dev/null +++ b/requirements/requirements-tool_bench.txt @@ -0,0 +1,12 @@ +aiohttp~=3.9.3 +pytest~=8.3.3 +huggingface_hub~=0.23.2 +matplotlib~=3.10.0 +numpy~=1.26.4 +openai~=1.55.3 +pandas~=2.2.3 +prometheus-client~=0.20.0 +requests~=2.32.3 +wget~=3.2 +typer~=0.15.1 +seaborn~=0.13.2 diff --git a/scripts/apple/validate-apps.sh b/scripts/apple/validate-apps.sh new file mode 100755 index 000000000..a571aa6fc --- /dev/null +++ b/scripts/apple/validate-apps.sh @@ -0,0 +1,5 @@ +#!/bin/bash +./scripts/apple/validate-ios.sh +./scripts/apple/validate-macos.sh +./scripts/apple/validate-visionos.sh +./scripts/apple/validate-tvos.sh diff --git a/scripts/apple/validate-ios.sh b/scripts/apple/validate-ios.sh new file mode 100755 index 000000000..7bda1b972 --- /dev/null +++ b/scripts/apple/validate-ios.sh @@ -0,0 +1,820 @@ +#!/bin/bash +# validate-ios.sh - Validate iOS Application with embedded llama.xcframework using SwiftUI + +# Authentication options (optional) (can be set via environment variables) +# To use: export APPLE_ID=your.email@example.com +# export APPLE_PASSWORD=your-app-specific-password +# ./validate-ios.sh +APPLE_ID=${APPLE_ID:-""} +APPLE_PASSWORD=${APPLE_PASSWORD:-""} + +# Ensure the script exits on error +set -e + +# Function to print usage instructions +print_usage() { + echo "Usage: ./validate-ios.sh [OPTIONS]" + echo "" + echo "Options:" + echo " --help Show this help message" + echo " --apple-id EMAIL Apple ID email for validation" + echo " --apple-password PWD App-specific password for Apple ID" + echo "" + echo "Environment variables:" + echo " APPLE_ID Apple ID email for validation" + echo " APPLE_PASSWORD App-specific password for Apple ID" + echo "" + echo "Notes:" + echo " - Command line options take precedence over environment variables" + echo " - Authentication is optional. If not provided, alternative validation will be performed" + echo " - For APPLE_PASSWORD, use an app-specific password generated at https://appleid.apple.com/account/manage" +} + +# Parse command line arguments +while [[ $# -gt 0 ]]; do + case $1 in + --help) + print_usage + exit 0 + ;; + --apple-id) + APPLE_ID="$2" + shift 2 + ;; + --apple-password) + APPLE_PASSWORD="$2" + shift 2 + ;; + *) + echo "Unknown option: $1" + print_usage + exit 1 + ;; + esac +done + +# Function to clean up in case of error +cleanup() { + # Don't clean up temp files on error to help with debugging + echo "===== iOS Validation Process Failed =====" + exit 1 +} + +# Set up trap to call cleanup function on error +trap cleanup ERR + +set -e # Exit on any error + +ROOT_DIR="$( cd "$( dirname "${BASH_SOURCE[0]}" )/../.." && pwd )" +BUILD_DIR="${ROOT_DIR}/validation-builds/ios" + +# Configuration +APP_NAME="iOSLlamaTest" +BUNDLE_ID="org.ggml.iOSLlamaTest" +XCFRAMEWORK_PATH="${ROOT_DIR}/build-apple/llama.xcframework" +TEMP_DIR="${BUILD_DIR}/temp" +ARCHIVE_PATH="${BUILD_DIR}/${APP_NAME}.xcarchive" +IPA_PATH="${BUILD_DIR}/${APP_NAME}.ipa" +VALIDATION_DIR="${BUILD_DIR}/validation" + +# Create necessary directories +mkdir -p "${BUILD_DIR}" +mkdir -p "${TEMP_DIR}" +mkdir -p "${VALIDATION_DIR}" + +echo "===== iOS Validation Process Started =====" + +# 1. Create a simple test app project +echo "Creating test iOS app project..." +mkdir -p "${TEMP_DIR}/${APP_NAME}/${APP_NAME}" +cat > "${TEMP_DIR}/${APP_NAME}/${APP_NAME}/Info.plist" << EOF + + + + + CFBundleDevelopmentRegion + en + CFBundleExecutable + ${APP_NAME} + CFBundleIdentifier + ${BUNDLE_ID} + CFBundleInfoDictionaryVersion + 6.0 + CFBundleName + ${APP_NAME} + CFBundlePackageType + APPL + CFBundleShortVersionString + 1.0 + CFBundleVersion + 1 + LSRequiresIPhoneOS + + UILaunchScreen + + UIRequiredDeviceCapabilities + + armv7 + + UISupportedInterfaceOrientations + + UIInterfaceOrientationPortrait + + + +EOF + +# Create SwiftUI app files +mkdir -p "${TEMP_DIR}/${APP_NAME}/${APP_NAME}/Sources" + +# Create App.swift +cat > "${TEMP_DIR}/${APP_NAME}/${APP_NAME}/Sources/App.swift" << EOF +import SwiftUI +import llama + +@main +struct LlamaTestApp: App { + var body: some Scene { + WindowGroup { + ContentView() + } + } +} +EOF + +# Create ContentView.swift +cat > "${TEMP_DIR}/${APP_NAME}/${APP_NAME}/Sources/ContentView.swift" << EOF +import SwiftUI +import llama + +struct ContentView: View { + // Test that we can initialize a llama context params struct + let params = llama_context_default_params() + + var body: some View { + VStack(spacing: 20) { + Text("Llama Framework Test") + .font(.largeTitle) + .padding() + + Text("llama_context_default_params() created successfully") + .font(.headline) + .multilineTextAlignment(.center) + .padding() + + // Display some param values to confirm the framework is working + Text("n_ctx: \(params.n_ctx)") + .font(.body) + + Text("n_batch: \(params.n_batch)") + .font(.body) + + Spacer() + } + .padding() + } +} + +struct ContentView_Previews: PreviewProvider { + static var previews: some View { + ContentView() + } +} +EOF + +# Create project.pbxproj, fixing the framework search paths issues +mkdir -p "${TEMP_DIR}/${APP_NAME}/${APP_NAME}.xcodeproj" +cat > "${TEMP_DIR}/${APP_NAME}/${APP_NAME}.xcodeproj/project.pbxproj" << 'EOF' +// !$*UTF8*$! +{ + archiveVersion = 1; + classes = { + }; + objectVersion = 54; + objects = { + +/* Begin PBXBuildFile section */ + 11111111111111111111111 /* App.swift in Sources */ = {isa = PBXBuildFile; fileRef = 22222222222222222222222; }; + 33333333333333333333333 /* ContentView.swift in Sources */ = {isa = PBXBuildFile; fileRef = 44444444444444444444444; }; + 55555555555555555555555 /* llama.xcframework in Frameworks */ = {isa = PBXBuildFile; fileRef = 66666666666666666666666; }; + 77777777777777777777777 /* llama.xcframework in Embed Frameworks */ = {isa = PBXBuildFile; fileRef = 66666666666666666666666; }; +/* End PBXBuildFile section */ + +/* Begin PBXCopyFilesBuildPhase section */ + 88888888888888888888888 /* Embed Frameworks */ = { + isa = PBXCopyFilesBuildPhase; + buildActionMask = 2147483647; + dstPath = ""; + dstSubfolderSpec = 10; + files = ( + 77777777777777777777777 /* llama.xcframework in Embed Frameworks */, + ); + name = "Embed Frameworks"; + runOnlyForDeploymentPostprocessing = 0; + }; +/* End PBXCopyFilesBuildPhase section */ + +/* Begin PBXFileReference section */ +EOF + +# Continue with the project.pbxproj file, using the APP_NAME variable appropriately +cat >> "${TEMP_DIR}/${APP_NAME}/${APP_NAME}.xcodeproj/project.pbxproj" << EOF + 99999999999999999999999 /* ${APP_NAME}.app */ = {isa = PBXFileReference; explicitFileType = wrapper.application; includeInIndex = 0; path = "${APP_NAME}.app"; sourceTree = BUILT_PRODUCTS_DIR; }; + 22222222222222222222222 /* App.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = App.swift; sourceTree = ""; }; + 44444444444444444444444 /* ContentView.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = ContentView.swift; sourceTree = ""; }; + AAAAAAAAAAAAAAAAAAAAAAA /* Info.plist */ = {isa = PBXFileReference; lastKnownFileType = text.plist.xml; path = Info.plist; sourceTree = ""; }; + 66666666666666666666666 /* llama.xcframework */ = {isa = PBXFileReference; lastKnownFileType = wrapper.xcframework; path = llama.xcframework; sourceTree = ""; }; +/* End PBXFileReference section */ +EOF + +# Add the rest of the project file with fixed framework search paths +cat >> "${TEMP_DIR}/${APP_NAME}/${APP_NAME}.xcodeproj/project.pbxproj" << 'EOF' +/* Begin PBXFrameworksBuildPhase section */ + BBBBBBBBBBBBBBBBBBBBBBBB /* Frameworks */ = { + isa = PBXFrameworksBuildPhase; + buildActionMask = 2147483647; + files = ( + 55555555555555555555555 /* llama.xcframework in Frameworks */, + ); + runOnlyForDeploymentPostprocessing = 0; + }; +/* End PBXFrameworksBuildPhase section */ + +/* Begin PBXGroup section */ +EOF + +# Continue with the project.pbxproj file, using the APP_NAME variable appropriately +cat >> "${TEMP_DIR}/${APP_NAME}/${APP_NAME}.xcodeproj/project.pbxproj" << EOF + CCCCCCCCCCCCCCCCCCCCCCCC /* Products */ = { + isa = PBXGroup; + children = ( + 99999999999999999999999 /* ${APP_NAME}.app */, + ); + name = Products; + sourceTree = ""; + }; +EOF + +cat >> "${TEMP_DIR}/${APP_NAME}/${APP_NAME}.xcodeproj/project.pbxproj" << 'EOF' + DDDDDDDDDDDDDDDDDDDDDDDD /* Frameworks */ = { + isa = PBXGroup; + children = ( + 66666666666666666666666 /* llama.xcframework */, + ); + name = Frameworks; + sourceTree = ""; + }; + EEEEEEEEEEEEEEEEEEEEEEEE = { + isa = PBXGroup; + children = ( + FFFFFFFFFFFFFFFFFFFFFFFF /* iOSLlamaTest */, + CCCCCCCCCCCCCCCCCCCCCCCC /* Products */, + DDDDDDDDDDDDDDDDDDDDDDDD /* Frameworks */, + ); + sourceTree = ""; + }; + FFFFFFFFFFFFFFFFFFFFFFFF /* iOSLlamaTest */ = { + isa = PBXGroup; + children = ( + 1111111111111111111111AA /* Sources */, + AAAAAAAAAAAAAAAAAAAAAAA /* Info.plist */, + ); + path = "iOSLlamaTest"; + sourceTree = ""; + }; + 1111111111111111111111AA /* Sources */ = { + isa = PBXGroup; + children = ( + 22222222222222222222222 /* App.swift */, + 44444444444444444444444 /* ContentView.swift */, + ); + path = Sources; + sourceTree = ""; + }; +/* End PBXGroup section */ +EOF + +# Continue with the project.pbxproj file, using the APP_NAME variable appropriately +cat >> "${TEMP_DIR}/${APP_NAME}/${APP_NAME}.xcodeproj/project.pbxproj" << EOF +/* Begin PBXNativeTarget section */ + 3333333333333333333333AA /* ${APP_NAME} */ = { + isa = PBXNativeTarget; + buildConfigurationList = 4444444444444444444444AA /* Build configuration list for PBXNativeTarget "${APP_NAME}" */; + buildPhases = ( + 5555555555555555555555AA /* Sources */, + BBBBBBBBBBBBBBBBBBBBBBBB /* Frameworks */, + 6666666666666666666666AA /* Resources */, + 88888888888888888888888 /* Embed Frameworks */, + ); + buildRules = ( + ); + dependencies = ( + ); + name = "${APP_NAME}"; + productName = "${APP_NAME}"; + productReference = 99999999999999999999999 /* ${APP_NAME}.app */; + productType = "com.apple.product-type.application"; + }; +/* End PBXNativeTarget section */ + +/* Begin PBXProject section */ + 7777777777777777777777AA /* Project object */ = { + isa = PBXProject; + attributes = { + LastSwiftUpdateCheck = 1240; + LastUpgradeCheck = 1240; + TargetAttributes = { + 3333333333333333333333AA = { + CreatedOnToolsVersion = 12.4; + }; + }; + }; + buildConfigurationList = 8888888888888888888888AA /* Build configuration list for PBXProject "${APP_NAME}" */; + compatibilityVersion = "Xcode 12.0"; + developmentRegion = en; + hasScannedForEncodings = 0; + knownRegions = ( + en, + Base, + ); + mainGroup = EEEEEEEEEEEEEEEEEEEEEEEE; + productRefGroup = CCCCCCCCCCCCCCCCCCCCCCCC /* Products */; + projectDirPath = ""; + projectRoot = ""; + targets = ( + 3333333333333333333333AA /* ${APP_NAME} */, + ); + }; +/* End PBXProject section */ +EOF + +# Add the rest of the file with correct FRAMEWORK_SEARCH_PATHS +cat >> "${TEMP_DIR}/${APP_NAME}/${APP_NAME}.xcodeproj/project.pbxproj" << 'EOF' +/* Begin PBXResourcesBuildPhase section */ + 6666666666666666666666AA /* Resources */ = { + isa = PBXResourcesBuildPhase; + buildActionMask = 2147483647; + files = ( + ); + runOnlyForDeploymentPostprocessing = 0; + }; +/* End PBXResourcesBuildPhase section */ + +/* Begin PBXSourcesBuildPhase section */ + 5555555555555555555555AA /* Sources */ = { + isa = PBXSourcesBuildPhase; + buildActionMask = 2147483647; + files = ( + 33333333333333333333333 /* ContentView.swift in Sources */, + 11111111111111111111111 /* App.swift in Sources */, + ); + runOnlyForDeploymentPostprocessing = 0; + }; +/* End PBXSourcesBuildPhase section */ + +/* Begin XCBuildConfiguration section */ + 9999999999999999999999AA /* Debug */ = { + isa = XCBuildConfiguration; + buildSettings = { + ALWAYS_SEARCH_USER_PATHS = NO; + CLANG_ANALYZER_NONNULL = YES; + CLANG_ANALYZER_NUMBER_OBJECT_CONVERSION = YES_AGGRESSIVE; + CLANG_CXX_LANGUAGE_STANDARD = "gnu++14"; + CLANG_CXX_LIBRARY = "libc++"; + CLANG_ENABLE_MODULES = YES; + CLANG_ENABLE_OBJC_ARC = YES; + CLANG_ENABLE_OBJC_WEAK = YES; + CLANG_WARN_BLOCK_CAPTURE_AUTORELEASING = YES; + CLANG_WARN_BOOL_CONVERSION = YES; + CLANG_WARN_COMMA = YES; + CLANG_WARN_CONSTANT_CONVERSION = YES; + CLANG_WARN_DEPRECATED_OBJC_IMPLEMENTATIONS = YES; + CLANG_WARN_DIRECT_OBJC_ISA_USAGE = YES_ERROR; + CLANG_WARN_DOCUMENTATION_COMMENTS = YES; + CLANG_WARN_EMPTY_BODY = YES; + CLANG_WARN_ENUM_CONVERSION = YES; + CLANG_WARN_INFINITE_RECURSION = YES; + CLANG_WARN_INT_CONVERSION = YES; + CLANG_WARN_NON_LITERAL_NULL_CONVERSION = YES; + CLANG_WARN_OBJC_IMPLICIT_RETAIN_SELF = YES; + CLANG_WARN_OBJC_LITERAL_CONVERSION = YES; + CLANG_WARN_OBJC_ROOT_CLASS = YES_ERROR; + CLANG_WARN_QUOTED_INCLUDE_IN_FRAMEWORK_HEADER = YES; + CLANG_WARN_RANGE_LOOP_ANALYSIS = YES; + CLANG_WARN_STRICT_PROTOTYPES = YES; + CLANG_WARN_SUSPICIOUS_MOVE = YES; + CLANG_WARN_UNGUARDED_AVAILABILITY = YES_AGGRESSIVE; + CLANG_WARN_UNREACHABLE_CODE = YES; + CLANG_WARN__DUPLICATE_METHOD_MATCH = YES; + COPY_PHASE_STRIP = NO; + DEBUG_INFORMATION_FORMAT = dwarf; + ENABLE_STRICT_OBJC_MSGSEND = YES; + ENABLE_TESTABILITY = YES; + GCC_C_LANGUAGE_STANDARD = gnu11; + GCC_DYNAMIC_NO_PIC = NO; + GCC_NO_COMMON_BLOCKS = YES; + GCC_OPTIMIZATION_LEVEL = 0; + GCC_PREPROCESSOR_DEFINITIONS = ( + "DEBUG=1", + "$(inherited)", + ); + GCC_WARN_64_TO_32_BIT_CONVERSION = YES; + GCC_WARN_ABOUT_RETURN_TYPE = YES_ERROR; + GCC_WARN_UNDECLARED_SELECTOR = YES; + GCC_WARN_UNINITIALIZED_AUTOS = YES_AGGRESSIVE; + GCC_WARN_UNUSED_FUNCTION = YES; + GCC_WARN_UNUSED_VARIABLE = YES; + IPHONEOS_DEPLOYMENT_TARGET = 16.4; + MTL_ENABLE_DEBUG_INFO = INCLUDE_SOURCE; + MTL_FAST_MATH = YES; + ONLY_ACTIVE_ARCH = YES; + SDKROOT = iphoneos; + SWIFT_ACTIVE_COMPILATION_CONDITIONS = DEBUG; + SWIFT_OPTIMIZATION_LEVEL = "-Onone"; + }; + name = Debug; + }; + AAAAAAAAAAAAAAAAAAAAABBB /* Release */ = { + isa = XCBuildConfiguration; + buildSettings = { + ALWAYS_SEARCH_USER_PATHS = NO; + CLANG_ANALYZER_NONNULL = YES; + CLANG_ANALYZER_NUMBER_OBJECT_CONVERSION = YES_AGGRESSIVE; + CLANG_CXX_LANGUAGE_STANDARD = "gnu++14"; + CLANG_CXX_LIBRARY = "libc++"; + CLANG_ENABLE_MODULES = YES; + CLANG_ENABLE_OBJC_ARC = YES; + CLANG_ENABLE_OBJC_WEAK = YES; + CLANG_WARN_BLOCK_CAPTURE_AUTORELEASING = YES; + CLANG_WARN_BOOL_CONVERSION = YES; + CLANG_WARN_COMMA = YES; + CLANG_WARN_CONSTANT_CONVERSION = YES; + CLANG_WARN_DEPRECATED_OBJC_IMPLEMENTATIONS = YES; + CLANG_WARN_DIRECT_OBJC_ISA_USAGE = YES_ERROR; + CLANG_WARN_DOCUMENTATION_COMMENTS = YES; + CLANG_WARN_EMPTY_BODY = YES; + CLANG_WARN_ENUM_CONVERSION = YES; + CLANG_WARN_INFINITE_RECURSION = YES; + CLANG_WARN_INT_CONVERSION = YES; + CLANG_WARN_NON_LITERAL_NULL_CONVERSION = YES; + CLANG_WARN_OBJC_IMPLICIT_RETAIN_SELF = YES; + CLANG_WARN_OBJC_LITERAL_CONVERSION = YES; + CLANG_WARN_OBJC_ROOT_CLASS = YES_ERROR; + CLANG_WARN_QUOTED_INCLUDE_IN_FRAMEWORK_HEADER = YES; + CLANG_WARN_RANGE_LOOP_ANALYSIS = YES; + CLANG_WARN_STRICT_PROTOTYPES = YES; + CLANG_WARN_SUSPICIOUS_MOVE = YES; + CLANG_WARN_UNGUARDED_AVAILABILITY = YES_AGGRESSIVE; + CLANG_WARN_UNREACHABLE_CODE = YES; + CLANG_WARN__DUPLICATE_METHOD_MATCH = YES; + COPY_PHASE_STRIP = NO; + DEBUG_INFORMATION_FORMAT = "dwarf-with-dsym"; + ENABLE_NS_ASSERTIONS = NO; + ENABLE_STRICT_OBJC_MSGSEND = YES; + GCC_C_LANGUAGE_STANDARD = gnu11; + GCC_NO_COMMON_BLOCKS = YES; + GCC_WARN_64_TO_32_BIT_CONVERSION = YES; + GCC_WARN_ABOUT_RETURN_TYPE = YES_ERROR; + GCC_WARN_UNDECLARED_SELECTOR = YES; + GCC_WARN_UNINITIALIZED_AUTOS = YES_AGGRESSIVE; + GCC_WARN_UNUSED_FUNCTION = YES; + GCC_WARN_UNUSED_VARIABLE = YES; + IPHONEOS_DEPLOYMENT_TARGET = 16.4; + MTL_ENABLE_DEBUG_INFO = NO; + MTL_FAST_MATH = YES; + SDKROOT = iphoneos; + SWIFT_COMPILATION_MODE = wholemodule; + SWIFT_OPTIMIZATION_LEVEL = "-O"; + VALIDATE_PRODUCT = YES; + }; + name = Release; + }; + BBBBBBBBBBBBBBBBBBBBBBCCC /* Debug */ = { + isa = XCBuildConfiguration; + buildSettings = { + ASSETCATALOG_COMPILER_APPICON_NAME = AppIcon; + ASSETCATALOG_COMPILER_GLOBAL_ACCENT_COLOR_NAME = AccentColor; + CODE_SIGN_STYLE = Manual; + DEVELOPMENT_TEAM = ""; + ENABLE_PREVIEWS = YES; + FRAMEWORK_SEARCH_PATHS = "$(PROJECT_DIR)"; + INFOPLIST_FILE = "iOSLlamaTest/Info.plist"; + LD_RUNPATH_SEARCH_PATHS = ( + "$(inherited)", + "@executable_path/Frameworks", + ); + PRODUCT_BUNDLE_IDENTIFIER = "org.ggml.iOSLlamaTest"; + PRODUCT_NAME = "$(TARGET_NAME)"; + PROVISIONING_PROFILE_SPECIFIER = ""; + SWIFT_VERSION = 5.0; + TARGETED_DEVICE_FAMILY = "1,2"; + }; + name = Debug; + }; + CCCCCCCCCCCCCCCCCCCCCCDDD /* Release */ = { + isa = XCBuildConfiguration; + buildSettings = { + ASSETCATALOG_COMPILER_APPICON_NAME = AppIcon; + ASSETCATALOG_COMPILER_GLOBAL_ACCENT_COLOR_NAME = AccentColor; + CODE_SIGN_STYLE = Manual; + DEVELOPMENT_TEAM = ""; + ENABLE_PREVIEWS = YES; + FRAMEWORK_SEARCH_PATHS = ( + "$(inherited)", + "$(PROJECT_DIR)", + ); + INFOPLIST_FILE = "iOSLlamaTest/Info.plist"; + LD_RUNPATH_SEARCH_PATHS = ( + "$(inherited)", + "@executable_path/Frameworks", + ); + PRODUCT_BUNDLE_IDENTIFIER = "org.ggml.iOSLlamaTest"; + PRODUCT_NAME = "$(TARGET_NAME)"; + PROVISIONING_PROFILE_SPECIFIER = ""; + SWIFT_VERSION = 5.0; + TARGETED_DEVICE_FAMILY = "1,2"; + }; + name = Release; + }; +/* End XCBuildConfiguration section */ +EOF + +# Finish the project.pbxproj file +cat >> "${TEMP_DIR}/${APP_NAME}/${APP_NAME}.xcodeproj/project.pbxproj" << EOF +/* Begin XCConfigurationList section */ + 8888888888888888888888AA /* Build configuration list for PBXProject "${APP_NAME}" */ = { + isa = XCConfigurationList; + buildConfigurations = ( + 9999999999999999999999AA /* Debug */, + AAAAAAAAAAAAAAAAAAAAABBB /* Release */, + ); + defaultConfigurationIsVisible = 0; + defaultConfigurationName = Release; + }; + 4444444444444444444444AA /* Build configuration list for PBXNativeTarget "${APP_NAME}" */ = { + isa = XCConfigurationList; + buildConfigurations = ( + BBBBBBBBBBBBBBBBBBBBBBCCC /* Debug */, + CCCCCCCCCCCCCCCCCCCCCCDDD /* Release */, + ); + defaultConfigurationIsVisible = 0; + defaultConfigurationName = Release; + }; +/* End XCConfigurationList section */ + }; + rootObject = 7777777777777777777777AA /* Project object */; +} +EOF + +# 2. Copy XCFramework to test project +echo "Copying XCFramework to test project..." +cp -R "${XCFRAMEWORK_PATH}" "${TEMP_DIR}/${APP_NAME}/" + +# 3. Build and archive the app +echo "Building and archiving test app..." +cd "${TEMP_DIR}/${APP_NAME}" + +# Create a simple xcscheme file to avoid xcodebuild scheme issues +mkdir -p "${APP_NAME}.xcodeproj/xcshareddata/xcschemes" +cat > "${APP_NAME}.xcodeproj/xcshareddata/xcschemes/${APP_NAME}.xcscheme" << EOF + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + +EOF + +# Now use xcodebuild with an explicitly defined product name +xcodebuild -project "${APP_NAME}.xcodeproj" -scheme "${APP_NAME}" -sdk iphoneos -configuration Release archive -archivePath "${ARCHIVE_PATH}" CODE_SIGN_IDENTITY="-" CODE_SIGNING_REQUIRED=NO CODE_SIGNING_ALLOWED=NO PRODUCT_NAME="${APP_NAME}" SWIFT_OPTIMIZATION_LEVEL="-Onone" -quiet + +# 4. Create IPA from archive +echo "Creating IPA from archive..." +mkdir -p "${TEMP_DIR}/Payload" +cp -R "${ARCHIVE_PATH}/Products/Applications/${APP_NAME}.app" "${TEMP_DIR}/Payload/" + +# Check and log app structure before zipping +echo "App structure:" +ls -la "${TEMP_DIR}/Payload/${APP_NAME}.app/" +echo "Frameworks:" +ls -la "${TEMP_DIR}/Payload/${APP_NAME}.app/Frameworks/" 2>/dev/null || echo "No Frameworks directory found" + +cd "${TEMP_DIR}" +zip -r "${IPA_PATH}" Payload + +# Check embedded provisioning profile +echo "Checking provisioning profile (if any)..." +PROVISIONING_PROFILE=$(find "${ARCHIVE_PATH}/Products/Applications/${APP_NAME}.app" -name "embedded.mobileprovision" 2>/dev/null) +if [ -n "$PROVISIONING_PROFILE" ]; then + echo "Found embedded provisioning profile:" + security cms -D -i "$PROVISIONING_PROFILE" || echo "Unable to decode provisioning profile" +else + echo "No embedded provisioning profile found (expected for ad-hoc builds)" +fi + +# 5. Validate the IPA +echo "Validating IPA..." +VALIDATION_OUTPUT="${VALIDATION_DIR}/validation_output.txt" + +# Check if authentication credentials are provided +AUTH_ARGS="" +if [ -n "$APPLE_ID" ] && [ -n "$APPLE_PASSWORD" ]; then + echo "Using Apple ID authentication for validation..." + AUTH_ARGS="--username \"$APPLE_ID\" --password \"$APPLE_PASSWORD\"" +else + echo "No authentication credentials provided. Will perform basic validation." + echo "To use your personal developer account, you can run the script with:" + echo " APPLE_ID='your.email@example.com' APPLE_PASSWORD='your-app-specific-password' ./validate-ios.sh" + echo "Note: You need to create an app-specific password at https://appleid.apple.com/account/manage" +fi + +# Run validation with detailed output +echo "Running validation with altool..." +if [ -n "$AUTH_ARGS" ]; then + # Use eval to properly handle the quoted arguments + eval "xcrun altool --validate-app -f \"${IPA_PATH}\" --type ios --output-format xml $AUTH_ARGS" 2>&1 | tee "${VALIDATION_OUTPUT}" +else + xcrun altool --validate-app -f "${IPA_PATH}" --type ios --output-format xml 2>&1 | tee "${VALIDATION_OUTPUT}" +fi +VALIDATION_RESULT=$? + +# Final validation result +FINAL_VALIDATION_RESULT=0 + +# Check if validation failed because the app isn't in App Store Connect +if grep -q "No suitable application records were found" "${VALIDATION_OUTPUT}"; then + echo "⚠️ App Store Connect Warning: The app bundle identifier is not found in App Store Connect" + echo "This is expected for apps that haven't been registered in App Store Connect yet." + echo "This doesn't indicate a problem with the build or framework." + + # Perform alternative validation + echo "Performing alternative validation checks..." + + # Check if IPA was created successfully + if [ -f "${IPA_PATH}" ] && [ -s "${IPA_PATH}" ]; then + echo "✅ IPA file created successfully" + else + echo "❌ IPA file not created or empty" + FINAL_VALIDATION_RESULT=1 + fi + + # Check if app binary exists and is executable + if [ -f "${TEMP_DIR}/Payload/${APP_NAME}.app/${APP_NAME}" ] && [ -x "${TEMP_DIR}/Payload/${APP_NAME}.app/${APP_NAME}" ]; then + echo "✅ App binary exists and is executable" + else + echo "❌ App binary missing or not executable" + FINAL_VALIDATION_RESULT=1 + fi + + # Check if framework was properly embedded + if [ -d "${TEMP_DIR}/Payload/${APP_NAME}.app/Frameworks/llama.framework" ]; then + echo "✅ llama.framework properly embedded" + else + echo "❌ llama.framework not properly embedded" + FINAL_VALIDATION_RESULT=1 + fi + + # Check if framework binary exists + if [ -f "${TEMP_DIR}/Payload/${APP_NAME}.app/Frameworks/llama.framework/llama" ]; then + echo "✅ Framework binary exists" + + # Further validate framework by checking architecture + ARCHS=$(lipo -info "${TEMP_DIR}/Payload/${APP_NAME}.app/Frameworks/llama.framework/llama" 2>/dev/null | grep -o "arm64\\|armv7\\|x86_64" | tr '\n' ' ') + if [ -n "$ARCHS" ]; then + echo "✅ Framework architecture(s): $ARCHS" + else + echo "⚠️ Could not determine framework architecture" + fi + else + echo "❌ Framework binary missing" + FINAL_VALIDATION_RESULT=1 + fi + + if [ $FINAL_VALIDATION_RESULT -eq 0 ]; then + echo "✅ Alternative validation PASSED: App built successfully with embedded framework" + else + echo "❌ Alternative validation FAILED: Issues found with the app or framework" + fi +elif grep -q "You must specify authentication credentials" "${VALIDATION_OUTPUT}" && [ -z "$AUTH_ARGS" ]; then + echo "✅ iOS Validation PASSED: IPA successfully validated" + echo "Results saved to ${VALIDATION_OUTPUT}" +else + echo "❌ iOS Validation FAILED: IPA validation found issues" + echo "See validation output at ${VALIDATION_OUTPUT}" + echo "" + echo "==== VALIDATION ERRORS ====" + + # Try to extract specific errors from the output + if grep -q "Error" "${VALIDATION_OUTPUT}"; then + grep -A 5 "Error" "${VALIDATION_OUTPUT}" + else + # If no specific error found, show the whole log + cat "${VALIDATION_OUTPUT}" + fi + + # Additional debugging: check IPA contents + echo "" + echo "==== IPA CONTENTS ====" + mkdir -p "${TEMP_DIR}/ipa_contents" + unzip -q "${IPA_PATH}" -d "${TEMP_DIR}/ipa_contents" + ls -la "${TEMP_DIR}/ipa_contents/Payload/${APP_NAME}.app/" + + # Check for code signing issues + echo "" + echo "==== CODE SIGNING INFO ====" + codesign -vv -d "${TEMP_DIR}/ipa_contents/Payload/${APP_NAME}.app" 2>&1 || echo "Code signing verification failed" + + # Check embedded frameworks + echo "" + echo "==== FRAMEWORK INFO ====" + ls -la "${TEMP_DIR}/ipa_contents/Payload/${APP_NAME}.app/Frameworks/" 2>/dev/null || echo "No Frameworks directory found" +fi + +# Don't clean up on error to allow inspection +if [ $FINAL_VALIDATION_RESULT -ne 0 ]; then + echo "" + echo "Temporary files kept for inspection at: ${TEMP_DIR}" + echo "===== iOS Validation Process Failed =====" + exit 1 +fi + +# Clean up temporary files but keep build artifacts +if [ $FINAL_VALIDATION_RESULT -eq 0 ]; then + echo "Cleaning up temporary files..." + #rm -rf "${TEMP_DIR}" +fi + +echo "===== iOS Validation Process Completed =====" +exit $FINAL_VALIDATION_RESULT diff --git a/scripts/apple/validate-macos.sh b/scripts/apple/validate-macos.sh new file mode 100755 index 000000000..6dc28e694 --- /dev/null +++ b/scripts/apple/validate-macos.sh @@ -0,0 +1,781 @@ +#!/bin/bash +# validate-macos.sh - Validate macOS Application with embedded llama.xcframework using SwiftUI + +# Authentication options (optional) (can be set via environment variables) +# To use: export APPLE_ID=your.email@example.com +# export APPLE_PASSWORD=your-app-specific-password +# ./validate-macos.sh +APPLE_ID=${APPLE_ID:-""} +APPLE_PASSWORD=${APPLE_PASSWORD:-""} + +# Ensure the script exits on error +set -e + +# Function to print usage instructions +print_usage() { + echo "Usage: ./validate-macos.sh [OPTIONS]" + echo "" + echo "Options:" + echo " --help Show this help message" + echo " --apple-id EMAIL Apple ID email for validation" + echo " --apple-password PWD App-specific password for Apple ID" + echo "" + echo "Environment variables:" + echo " APPLE_ID Apple ID email for validation" + echo " APPLE_PASSWORD App-specific password for Apple ID" + echo "" + echo "Notes:" + echo " - Command line options take precedence over environment variables" + echo " - Authentication is optional. If not provided, alternative validation will be performed" + echo " - For APPLE_PASSWORD, use an app-specific password generated at https://appleid.apple.com/account/manage" +} + +# Parse command line arguments +while [[ $# -gt 0 ]]; do + case $1 in + --help) + print_usage + exit 0 + ;; + --apple-id) + APPLE_ID="$2" + shift 2 + ;; + --apple-password) + APPLE_PASSWORD="$2" + shift 2 + ;; + *) + echo "Unknown option: $1" + print_usage + exit 1 + ;; + esac +done + +# Function to clean up in case of error +cleanup() { + # Don't clean up temp files on error to help with debugging + echo "===== macOS Validation Process Failed =====" + exit 1 +} + +# Set up trap to call cleanup function on error +trap cleanup ERR + +set -e # Exit on any error + +ROOT_DIR="$( cd "$( dirname "${BASH_SOURCE[0]}" )/../.." && pwd )" +BUILD_DIR="${ROOT_DIR}/validation-builds/ios" + +# Configuration +APP_NAME="MacOSLlamaTest" +BUNDLE_ID="org.ggml.MacOSLlamaTest" +XCFRAMEWORK_PATH="${ROOT_DIR}/build-apple/llama.xcframework" +TEMP_DIR="${BUILD_DIR}/temp" +ARCHIVE_PATH="${BUILD_DIR}/${APP_NAME}.xcarchive" +APP_PATH="${BUILD_DIR}/${APP_NAME}.app" +ZIP_PATH="${BUILD_DIR}/${APP_NAME}.zip" +VALIDATION_DIR="${BUILD_DIR}/validation" + +# Create necessary directories +mkdir -p "${BUILD_DIR}" +mkdir -p "${TEMP_DIR}" +mkdir -p "${VALIDATION_DIR}" + +echo "===== macOS Validation Process Started =====" + +# 1. Create a simple test app project +echo "Creating test macOS app project..." +mkdir -p "${TEMP_DIR}/${APP_NAME}/${APP_NAME}" +cat > "${TEMP_DIR}/${APP_NAME}/${APP_NAME}/Info.plist" << EOF + + + + + CFBundleDevelopmentRegion + en + CFBundleExecutable + ${APP_NAME} + CFBundleIdentifier + ${BUNDLE_ID} + CFBundleInfoDictionaryVersion + 6.0 + CFBundleName + ${APP_NAME} + CFBundlePackageType + APPL + CFBundleShortVersionString + 1.0 + CFBundleVersion + 1 + LSMinimumSystemVersion + 12.0 + NSHumanReadableCopyright + Copyright © 2025 GGML. All rights reserved. + NSPrincipalClass + NSApplication + + +EOF + +# Create SwiftUI app files +mkdir -p "${TEMP_DIR}/${APP_NAME}/${APP_NAME}/Sources" + +# Create App.swift +cat > "${TEMP_DIR}/${APP_NAME}/${APP_NAME}/Sources/App.swift" << EOF +import SwiftUI +import llama + +@main +struct LlamaTestApp: App { + var body: some Scene { + WindowGroup { + ContentView() + } + } +} +EOF + +# Create ContentView.swift with macOS specific elements +cat > "${TEMP_DIR}/${APP_NAME}/${APP_NAME}/Sources/ContentView.swift" << EOF +import SwiftUI +import llama + +struct ContentView: View { + // Test that we can initialize a llama context params struct + let params = llama_context_default_params() + + var body: some View { + VStack(spacing: 20) { + Text("Llama Framework Test on macOS") + .font(.largeTitle) + .padding() + + Text("llama_context_default_params() created successfully") + .font(.headline) + .multilineTextAlignment(.center) + .padding() + + // Display some param values to confirm the framework is working + Text("n_ctx: \(params.n_ctx)") + .font(.body) + + Text("n_batch: \(params.n_batch)") + .font(.body) + + Spacer() + } + .padding() + .frame(width: 600, height: 400) + } +} + +struct ContentView_Previews: PreviewProvider { + static var previews: some View { + ContentView() + } +} +EOF + +# Create project.pbxproj, fixing the framework search paths issues +mkdir -p "${TEMP_DIR}/${APP_NAME}/${APP_NAME}.xcodeproj" +cat > "${TEMP_DIR}/${APP_NAME}/${APP_NAME}.xcodeproj/project.pbxproj" << 'EOF' +// !$*UTF8*$! +{ + archiveVersion = 1; + classes = { + }; + objectVersion = 54; + objects = { + +/* Begin PBXBuildFile section */ + 11111111111111111111111 /* App.swift in Sources */ = {isa = PBXBuildFile; fileRef = 22222222222222222222222; }; + 33333333333333333333333 /* ContentView.swift in Sources */ = {isa = PBXBuildFile; fileRef = 44444444444444444444444; }; + 55555555555555555555555 /* llama.xcframework in Frameworks */ = {isa = PBXBuildFile; fileRef = 66666666666666666666666; }; + 77777777777777777777777 /* llama.xcframework in Embed Frameworks */ = {isa = PBXBuildFile; fileRef = 66666666666666666666666; }; +/* End PBXBuildFile section */ + +/* Begin PBXCopyFilesBuildPhase section */ + 88888888888888888888888 /* Embed Frameworks */ = { + isa = PBXCopyFilesBuildPhase; + buildActionMask = 2147483647; + dstPath = ""; + dstSubfolderSpec = 10; + files = ( + 77777777777777777777777 /* llama.xcframework in Embed Frameworks */, + ); + name = "Embed Frameworks"; + runOnlyForDeploymentPostprocessing = 0; + }; +/* End PBXCopyFilesBuildPhase section */ + +/* Begin PBXFileReference section */ +EOF + +# Continue with the project.pbxproj file, using the APP_NAME variable appropriately +cat >> "${TEMP_DIR}/${APP_NAME}/${APP_NAME}.xcodeproj/project.pbxproj" << EOF + 99999999999999999999999 /* ${APP_NAME}.app */ = {isa = PBXFileReference; explicitFileType = wrapper.application; includeInIndex = 0; path = "${APP_NAME}.app"; sourceTree = BUILT_PRODUCTS_DIR; }; + 22222222222222222222222 /* App.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = App.swift; sourceTree = ""; }; + 44444444444444444444444 /* ContentView.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = ContentView.swift; sourceTree = ""; }; + AAAAAAAAAAAAAAAAAAAAAAA /* Info.plist */ = {isa = PBXFileReference; lastKnownFileType = text.plist.xml; path = Info.plist; sourceTree = ""; }; + 66666666666666666666666 /* llama.xcframework */ = {isa = PBXFileReference; lastKnownFileType = wrapper.xcframework; path = llama.xcframework; sourceTree = ""; }; +/* End PBXFileReference section */ +EOF + +# Add the rest of the project file with fixed framework search paths +cat >> "${TEMP_DIR}/${APP_NAME}/${APP_NAME}.xcodeproj/project.pbxproj" << 'EOF' +/* Begin PBXFrameworksBuildPhase section */ + BBBBBBBBBBBBBBBBBBBBBBBB /* Frameworks */ = { + isa = PBXFrameworksBuildPhase; + buildActionMask = 2147483647; + files = ( + 55555555555555555555555 /* llama.xcframework in Frameworks */, + ); + runOnlyForDeploymentPostprocessing = 0; + }; +/* End PBXFrameworksBuildPhase section */ + +/* Begin PBXGroup section */ +EOF + +# Continue with the project.pbxproj file, using the APP_NAME variable appropriately +cat >> "${TEMP_DIR}/${APP_NAME}/${APP_NAME}.xcodeproj/project.pbxproj" << EOF + CCCCCCCCCCCCCCCCCCCCCCCC /* Products */ = { + isa = PBXGroup; + children = ( + 99999999999999999999999 /* ${APP_NAME}.app */, + ); + name = Products; + sourceTree = ""; + }; +EOF + +cat >> "${TEMP_DIR}/${APP_NAME}/${APP_NAME}.xcodeproj/project.pbxproj" << 'EOF' + DDDDDDDDDDDDDDDDDDDDDDDD /* Frameworks */ = { + isa = PBXGroup; + children = ( + 66666666666666666666666 /* llama.xcframework */, + ); + name = Frameworks; + sourceTree = ""; + }; + EEEEEEEEEEEEEEEEEEEEEEEE = { + isa = PBXGroup; + children = ( + FFFFFFFFFFFFFFFFFFFFFFFF /* MacOSLlamaTest */, + CCCCCCCCCCCCCCCCCCCCCCCC /* Products */, + DDDDDDDDDDDDDDDDDDDDDDDD /* Frameworks */, + ); + sourceTree = ""; + }; + FFFFFFFFFFFFFFFFFFFFFFFF /* MacOSLlamaTest */ = { + isa = PBXGroup; + children = ( + 1111111111111111111111AA /* Sources */, + AAAAAAAAAAAAAAAAAAAAAAA /* Info.plist */, + ); + path = "MacOSLlamaTest"; + sourceTree = ""; + }; + 1111111111111111111111AA /* Sources */ = { + isa = PBXGroup; + children = ( + 22222222222222222222222 /* App.swift */, + 44444444444444444444444 /* ContentView.swift */, + ); + path = Sources; + sourceTree = ""; + }; +/* End PBXGroup section */ +EOF + +# Continue with the project.pbxproj file, using the APP_NAME variable appropriately +cat >> "${TEMP_DIR}/${APP_NAME}/${APP_NAME}.xcodeproj/project.pbxproj" << EOF +/* Begin PBXNativeTarget section */ + 3333333333333333333333AA /* ${APP_NAME} */ = { + isa = PBXNativeTarget; + buildConfigurationList = 4444444444444444444444AA /* Build configuration list for PBXNativeTarget "${APP_NAME}" */; + buildPhases = ( + 5555555555555555555555AA /* Sources */, + BBBBBBBBBBBBBBBBBBBBBBBB /* Frameworks */, + 6666666666666666666666AA /* Resources */, + 88888888888888888888888 /* Embed Frameworks */, + ); + buildRules = ( + ); + dependencies = ( + ); + name = "${APP_NAME}"; + productName = "${APP_NAME}"; + productReference = 99999999999999999999999 /* ${APP_NAME}.app */; + productType = "com.apple.product-type.application"; + }; +/* End PBXNativeTarget section */ + +/* Begin PBXProject section */ + 7777777777777777777777AA /* Project object */ = { + isa = PBXProject; + attributes = { + LastSwiftUpdateCheck = 1240; + LastUpgradeCheck = 1240; + TargetAttributes = { + 3333333333333333333333AA = { + CreatedOnToolsVersion = 12.4; + }; + }; + }; + buildConfigurationList = 8888888888888888888888AA /* Build configuration list for PBXProject "${APP_NAME}" */; + compatibilityVersion = "Xcode 12.0"; + developmentRegion = en; + hasScannedForEncodings = 0; + knownRegions = ( + en, + Base, + ); + mainGroup = EEEEEEEEEEEEEEEEEEEEEEEE; + productRefGroup = CCCCCCCCCCCCCCCCCCCCCCCC /* Products */; + projectDirPath = ""; + projectRoot = ""; + targets = ( + 3333333333333333333333AA /* ${APP_NAME} */, + ); + }; +/* End PBXProject section */ +EOF + +# Add the rest of the file with correct FRAMEWORK_SEARCH_PATHS and macOS settings +cat >> "${TEMP_DIR}/${APP_NAME}/${APP_NAME}.xcodeproj/project.pbxproj" << 'EOF' +/* Begin PBXResourcesBuildPhase section */ + 6666666666666666666666AA /* Resources */ = { + isa = PBXResourcesBuildPhase; + buildActionMask = 2147483647; + files = ( + ); + runOnlyForDeploymentPostprocessing = 0; + }; +/* End PBXResourcesBuildPhase section */ + +/* Begin PBXSourcesBuildPhase section */ + 5555555555555555555555AA /* Sources */ = { + isa = PBXSourcesBuildPhase; + buildActionMask = 2147483647; + files = ( + 33333333333333333333333 /* ContentView.swift in Sources */, + 11111111111111111111111 /* App.swift in Sources */, + ); + runOnlyForDeploymentPostprocessing = 0; + }; +/* End PBXSourcesBuildPhase section */ + +/* Begin XCBuildConfiguration section */ + 9999999999999999999999AA /* Debug */ = { + isa = XCBuildConfiguration; + buildSettings = { + ALWAYS_SEARCH_USER_PATHS = NO; + CLANG_ANALYZER_NONNULL = YES; + CLANG_ANALYZER_NUMBER_OBJECT_CONVERSION = YES_AGGRESSIVE; + CLANG_CXX_LANGUAGE_STANDARD = "gnu++14"; + CLANG_CXX_LIBRARY = "libc++"; + CLANG_ENABLE_MODULES = YES; + CLANG_ENABLE_OBJC_ARC = YES; + CLANG_ENABLE_OBJC_WEAK = YES; + CLANG_WARN_BLOCK_CAPTURE_AUTORELEASING = YES; + CLANG_WARN_BOOL_CONVERSION = YES; + CLANG_WARN_COMMA = YES; + CLANG_WARN_CONSTANT_CONVERSION = YES; + CLANG_WARN_DEPRECATED_OBJC_IMPLEMENTATIONS = YES; + CLANG_WARN_DIRECT_OBJC_ISA_USAGE = YES_ERROR; + CLANG_WARN_DOCUMENTATION_COMMENTS = YES; + CLANG_WARN_EMPTY_BODY = YES; + CLANG_WARN_ENUM_CONVERSION = YES; + CLANG_WARN_INFINITE_RECURSION = YES; + CLANG_WARN_INT_CONVERSION = YES; + CLANG_WARN_NON_LITERAL_NULL_CONVERSION = YES; + CLANG_WARN_OBJC_IMPLICIT_RETAIN_SELF = YES; + CLANG_WARN_OBJC_LITERAL_CONVERSION = YES; + CLANG_WARN_OBJC_ROOT_CLASS = YES_ERROR; + CLANG_WARN_QUOTED_INCLUDE_IN_FRAMEWORK_HEADER = YES; + CLANG_WARN_RANGE_LOOP_ANALYSIS = YES; + CLANG_WARN_STRICT_PROTOTYPES = YES; + CLANG_WARN_SUSPICIOUS_MOVE = YES; + CLANG_WARN_UNGUARDED_AVAILABILITY = YES_AGGRESSIVE; + CLANG_WARN_UNREACHABLE_CODE = YES; + CLANG_WARN__DUPLICATE_METHOD_MATCH = YES; + COPY_PHASE_STRIP = NO; + DEBUG_INFORMATION_FORMAT = dwarf; + ENABLE_STRICT_OBJC_MSGSEND = YES; + ENABLE_TESTABILITY = YES; + GCC_C_LANGUAGE_STANDARD = gnu11; + GCC_DYNAMIC_NO_PIC = NO; + GCC_NO_COMMON_BLOCKS = YES; + GCC_OPTIMIZATION_LEVEL = 0; + GCC_PREPROCESSOR_DEFINITIONS = ( + "DEBUG=1", + "$(inherited)", + ); + GCC_WARN_64_TO_32_BIT_CONVERSION = YES; + GCC_WARN_ABOUT_RETURN_TYPE = YES_ERROR; + GCC_WARN_UNDECLARED_SELECTOR = YES; + GCC_WARN_UNINITIALIZED_AUTOS = YES_AGGRESSIVE; + GCC_WARN_UNUSED_FUNCTION = YES; + GCC_WARN_UNUSED_VARIABLE = YES; + MACOSX_DEPLOYMENT_TARGET = 12.0; + MTL_ENABLE_DEBUG_INFO = INCLUDE_SOURCE; + MTL_FAST_MATH = YES; + ONLY_ACTIVE_ARCH = YES; + SDKROOT = macosx; + SWIFT_ACTIVE_COMPILATION_CONDITIONS = DEBUG; + SWIFT_OPTIMIZATION_LEVEL = "-Onone"; + }; + name = Debug; + }; + AAAAAAAAAAAAAAAAAAAAABBB /* Release */ = { + isa = XCBuildConfiguration; + buildSettings = { + ALWAYS_SEARCH_USER_PATHS = NO; + CLANG_ANALYZER_NONNULL = YES; + CLANG_ANALYZER_NUMBER_OBJECT_CONVERSION = YES_AGGRESSIVE; + CLANG_CXX_LANGUAGE_STANDARD = "gnu++14"; + CLANG_CXX_LIBRARY = "libc++"; + CLANG_ENABLE_MODULES = YES; + CLANG_ENABLE_OBJC_ARC = YES; + CLANG_ENABLE_OBJC_WEAK = YES; + CLANG_WARN_BLOCK_CAPTURE_AUTORELEASING = YES; + CLANG_WARN_BOOL_CONVERSION = YES; + CLANG_WARN_COMMA = YES; + CLANG_WARN_CONSTANT_CONVERSION = YES; + CLANG_WARN_DEPRECATED_OBJC_IMPLEMENTATIONS = YES; + CLANG_WARN_DIRECT_OBJC_ISA_USAGE = YES_ERROR; + CLANG_WARN_DOCUMENTATION_COMMENTS = YES; + CLANG_WARN_EMPTY_BODY = YES; + CLANG_WARN_ENUM_CONVERSION = YES; + CLANG_WARN_INFINITE_RECURSION = YES; + CLANG_WARN_INT_CONVERSION = YES; + CLANG_WARN_NON_LITERAL_NULL_CONVERSION = YES; + CLANG_WARN_OBJC_IMPLICIT_RETAIN_SELF = YES; + CLANG_WARN_OBJC_LITERAL_CONVERSION = YES; + CLANG_WARN_OBJC_ROOT_CLASS = YES_ERROR; + CLANG_WARN_QUOTED_INCLUDE_IN_FRAMEWORK_HEADER = YES; + CLANG_WARN_RANGE_LOOP_ANALYSIS = YES; + CLANG_WARN_STRICT_PROTOTYPES = YES; + CLANG_WARN_SUSPICIOUS_MOVE = YES; + CLANG_WARN_UNGUARDED_AVAILABILITY = YES_AGGRESSIVE; + CLANG_WARN_UNREACHABLE_CODE = YES; + CLANG_WARN__DUPLICATE_METHOD_MATCH = YES; + COPY_PHASE_STRIP = NO; + DEBUG_INFORMATION_FORMAT = "dwarf-with-dsym"; + ENABLE_NS_ASSERTIONS = NO; + ENABLE_STRICT_OBJC_MSGSEND = YES; + GCC_C_LANGUAGE_STANDARD = gnu11; + GCC_NO_COMMON_BLOCKS = YES; + GCC_WARN_64_TO_32_BIT_CONVERSION = YES; + GCC_WARN_ABOUT_RETURN_TYPE = YES_ERROR; + GCC_WARN_UNDECLARED_SELECTOR = YES; + GCC_WARN_UNINITIALIZED_AUTOS = YES_AGGRESSIVE; + GCC_WARN_UNUSED_FUNCTION = YES; + GCC_WARN_UNUSED_VARIABLE = YES; + MACOSX_DEPLOYMENT_TARGET = 12.0; + MTL_ENABLE_DEBUG_INFO = NO; + MTL_FAST_MATH = YES; + SDKROOT = macosx; + SWIFT_COMPILATION_MODE = wholemodule; + SWIFT_OPTIMIZATION_LEVEL = "-O"; + }; + name = Release; + }; + BBBBBBBBBBBBBBBBBBBBBBCCC /* Debug */ = { + isa = XCBuildConfiguration; + buildSettings = { + ASSETCATALOG_COMPILER_APPICON_NAME = AppIcon; + ASSETCATALOG_COMPILER_GLOBAL_ACCENT_COLOR_NAME = AccentColor; + CODE_SIGN_STYLE = Manual; + COMBINE_HIDPI_IMAGES = YES; + DEVELOPMENT_TEAM = ""; + ENABLE_HARDENED_RUNTIME = YES; + ENABLE_PREVIEWS = YES; + FRAMEWORK_SEARCH_PATHS = "$(PROJECT_DIR)"; + INFOPLIST_FILE = "MacOSLlamaTest/Info.plist"; + LD_RUNPATH_SEARCH_PATHS = ( + "$(inherited)", + "@executable_path/../Frameworks", + ); + PRODUCT_BUNDLE_IDENTIFIER = "org.ggml.MacOSLlamaTest"; + PRODUCT_NAME = "$(TARGET_NAME)"; + PROVISIONING_PROFILE_SPECIFIER = ""; + SWIFT_VERSION = 5.0; + }; + name = Debug; + }; + CCCCCCCCCCCCCCCCCCCCCCDDD /* Release */ = { + isa = XCBuildConfiguration; + buildSettings = { + ASSETCATALOG_COMPILER_APPICON_NAME = AppIcon; + ASSETCATALOG_COMPILER_GLOBAL_ACCENT_COLOR_NAME = AccentColor; + CODE_SIGN_STYLE = Manual; + COMBINE_HIDPI_IMAGES = YES; + DEVELOPMENT_TEAM = ""; + ENABLE_HARDENED_RUNTIME = YES; + ENABLE_PREVIEWS = YES; + FRAMEWORK_SEARCH_PATHS = ( + "$(inherited)", + "$(PROJECT_DIR)", + ); + INFOPLIST_FILE = "MacOSLlamaTest/Info.plist"; + LD_RUNPATH_SEARCH_PATHS = ( + "$(inherited)", + "@executable_path/../Frameworks", + ); + PRODUCT_BUNDLE_IDENTIFIER = "org.ggml.MacOSLlamaTest"; + PRODUCT_NAME = "$(TARGET_NAME)"; + PROVISIONING_PROFILE_SPECIFIER = ""; + SWIFT_VERSION = 5.0; + }; + name = Release; + }; +/* End XCBuildConfiguration section */ +EOF + +# Finish the project.pbxproj file +cat >> "${TEMP_DIR}/${APP_NAME}/${APP_NAME}.xcodeproj/project.pbxproj" << EOF +/* Begin XCConfigurationList section */ + 8888888888888888888888AA /* Build configuration list for PBXProject "${APP_NAME}" */ = { + isa = XCConfigurationList; + buildConfigurations = ( + 9999999999999999999999AA /* Debug */, + AAAAAAAAAAAAAAAAAAAAABBB /* Release */, + ); + defaultConfigurationIsVisible = 0; + defaultConfigurationName = Release; + }; + 4444444444444444444444AA /* Build configuration list for PBXNativeTarget "${APP_NAME}" */ = { + isa = XCConfigurationList; + buildConfigurations = ( + BBBBBBBBBBBBBBBBBBBBBBCCC /* Debug */, + CCCCCCCCCCCCCCCCCCCCCCDDD /* Release */, + ); + defaultConfigurationIsVisible = 0; + defaultConfigurationName = Release; + }; +/* End XCConfigurationList section */ + }; + rootObject = 7777777777777777777777AA /* Project object */; +} +EOF + +# 2. Copy XCFramework to test project +echo "Copying XCFramework to test project..." +cp -R "${XCFRAMEWORK_PATH}" "${TEMP_DIR}/${APP_NAME}/" + +# 3. Build and archive the app +echo "Building and archiving test app..." +cd "${TEMP_DIR}/${APP_NAME}" + +# Create a simple xcscheme file to avoid xcodebuild scheme issues +mkdir -p "${APP_NAME}.xcodeproj/xcshareddata/xcschemes" +cat > "${APP_NAME}.xcodeproj/xcshareddata/xcschemes/${APP_NAME}.xcscheme" << EOF + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + +EOF + +# Now use xcodebuild with an explicitly defined product name for macOS +xcodebuild -project "${APP_NAME}.xcodeproj" -scheme "${APP_NAME}" -sdk macosx -configuration Release archive -archivePath "${ARCHIVE_PATH}" CODE_SIGN_IDENTITY="-" CODE_SIGNING_REQUIRED=NO CODE_SIGNING_ALLOWED=NO PRODUCT_NAME="${APP_NAME}" SWIFT_OPTIMIZATION_LEVEL="-Onone" -quiet + +# 4. Create a package for distribution +echo "Creating distributable package from archive..." +cp -R "${ARCHIVE_PATH}/Products/Applications/${APP_NAME}.app" "${APP_PATH}" + +# Check and log app structure +echo "App structure:" +ls -la "${APP_PATH}" +echo "Frameworks:" +ls -la "${APP_PATH}/Contents/Frameworks/" 2>/dev/null || echo "No Frameworks directory found" + +# Create a zip file for potential distribution +cd "${BUILD_DIR}" +zip -r "${ZIP_PATH}" "${APP_NAME}.app" + +# Check embedded provisioning profile +echo "Checking provisioning profile (if any)..." +PROVISIONING_PROFILE=$(find "${APP_PATH}/Contents" -name "embedded.provisionprofile" 2>/dev/null) +if [ -n "$PROVISIONING_PROFILE" ]; then + echo "Found embedded provisioning profile:" + security cms -D -i "$PROVISIONING_PROFILE" || echo "Unable to decode provisioning profile" +else + echo "No embedded provisioning profile found (expected for ad-hoc builds)" +fi + +# 5. Validate the app +echo "Validating macOS app..." +VALIDATION_OUTPUT="${VALIDATION_DIR}/validation_output.txt" + +# Check if authentication credentials are provided +AUTH_ARGS="" +if [ -n "$APPLE_ID" ] && [ -n "$APPLE_PASSWORD" ]; then + echo "Using Apple ID authentication for validation..." + AUTH_ARGS="--username \"$APPLE_ID\" --password \"$APPLE_PASSWORD\"" +else + echo "No authentication credentials provided. Will perform basic validation." + echo "To use your personal developer account, you can run the script with:" + echo " APPLE_ID='your.email@example.com' APPLE_PASSWORD='your-app-specific-password' ./validate-macos.sh" + echo "Note: You need to create an app-specific password at https://appleid.apple.com/account/manage" +fi + +# For macOS we need to use notarytool or alternative checks because altool doesn't support macOS apps in the same way +echo "Note: For macOS, formal notarization process would require Apple Developer credentials." +echo "Performing alternative validation checks..." + +# Final validation result +FINAL_VALIDATION_RESULT=0 + +# Check if app was created successfully +if [ -d "${APP_PATH}" ] && [ -s "${APP_PATH}/Contents/MacOS/${APP_NAME}" ]; then + echo "✅ App package created successfully" +else + echo "❌ App package not created or binary missing" + FINAL_VALIDATION_RESULT=1 +fi + +# Check if app binary exists and is executable +if [ -f "${APP_PATH}/Contents/MacOS/${APP_NAME}" ] && [ -x "${APP_PATH}/Contents/MacOS/${APP_NAME}" ]; then + echo "✅ App binary exists and is executable" +else + echo "❌ App binary missing or not executable" + FINAL_VALIDATION_RESULT=1 +fi + +# Check if framework was properly embedded +if [ -d "${APP_PATH}/Contents/Frameworks/llama.framework" ]; then + echo "✅ llama.framework properly embedded" +else + echo "❌ llama.framework not properly embedded" + FINAL_VALIDATION_RESULT=1 +fi + +# Check if framework binary exists +if [ -f "${APP_PATH}/Contents/Frameworks/llama.framework/Versions/A/llama" ]; then + echo "✅ Framework binary exists" + + # Further validate framework by checking architecture + ARCHS=$(lipo -info "${APP_PATH}/Contents/Frameworks/llama.framework/Versions/A/llama" 2>/dev/null | grep -o "arm64\\|x86_64" | tr '\n' ' ') + if [ -n "$ARCHS" ]; then + echo "✅ Framework architecture(s): $ARCHS" + else + echo "⚠️ Could not determine framework architecture" + fi +else + echo "❌ Framework binary missing" + FINAL_VALIDATION_RESULT=1 +fi + +# Check code signing +echo "" +echo "==== CODE SIGNING INFO ====" +codesign -vv -d "${APP_PATH}" 2>&1 || echo "Code signing verification not available (expected for ad-hoc builds)" + +if [ $FINAL_VALIDATION_RESULT -eq 0 ]; then + if [ -n "$AUTH_ARGS" ]; then + echo "" + echo "To notarize this app with Apple (requires Apple Developer account):" + echo "xcrun notarytool submit \"${ZIP_PATH}\" --apple-id \"your-apple-id\" --password \"your-app-specific-password\" --team-id \"your-team-id\" --wait" + echo "" + fi + echo "✅ Validation PASSED: macOS app built successfully with embedded framework" +else + echo "❌ Validation FAILED: Issues found with the app or framework" +fi + +# Don't clean up on error to allow inspection +if [ $FINAL_VALIDATION_RESULT -ne 0 ]; then + echo "" + echo "Temporary files kept for inspection at: ${TEMP_DIR}" + echo "===== macOS Validation Process Failed =====" + exit 1 +fi + +# Clean up temporary files but keep build artifacts +if [ $FINAL_VALIDATION_RESULT -eq 0 ]; then + echo "Cleaning up temporary files..." + #rm -rf "${TEMP_DIR}" +fi + +echo "===== macOS Validation Process Completed =====" +echo "App package available at: ${APP_PATH}" +echo "Zipped app available at: ${ZIP_PATH}" +exit $FINAL_VALIDATION_RESULT diff --git a/scripts/apple/validate-tvos.sh b/scripts/apple/validate-tvos.sh new file mode 100755 index 000000000..6120189e8 --- /dev/null +++ b/scripts/apple/validate-tvos.sh @@ -0,0 +1,813 @@ +#!/bin/bash +# validate-tvos.sh - Validate tvOS Application with embedded llama.xcframework using SwiftUI + +# Authentication options (optional) (can be set via environment variables) +# To use: export APPLE_ID=your.email@example.com +# export APPLE_PASSWORD=your-app-specific-password +# ./validate-tvos.sh +APPLE_ID=${APPLE_ID:-""} +APPLE_PASSWORD=${APPLE_PASSWORD:-""} + +# Ensure the script exits on error +set -e + +# Function to print usage instructions +print_usage() { + echo "Usage: ./validate-tvos.sh [OPTIONS]" + echo "" + echo "Options:" + echo " --help Show this help message" + echo " --apple-id EMAIL Apple ID email for validation" + echo " --apple-password PWD App-specific password for Apple ID" + echo "" + echo "Environment variables:" + echo " APPLE_ID Apple ID email for validation" + echo " APPLE_PASSWORD App-specific password for Apple ID" + echo "" + echo "Notes:" + echo " - Command line options take precedence over environment variables" + echo " - Authentication is optional. If not provided, alternative validation will be performed" + echo " - For APPLE_PASSWORD, use an app-specific password generated at https://appleid.apple.com/account/manage" +} + +# Parse command line arguments +while [[ $# -gt 0 ]]; do + case $1 in + --help) + print_usage + exit 0 + ;; + --apple-id) + APPLE_ID="$2" + shift 2 + ;; + --apple-password) + APPLE_PASSWORD="$2" + shift 2 + ;; + *) + echo "Unknown option: $1" + print_usage + exit 1 + ;; + esac +done + +# Function to clean up in case of error +cleanup() { + # Don't clean up temp files on error to help with debugging + echo "===== tvOS Validation Process Failed =====" + exit 1 +} + +# Set up trap to call cleanup function on error +trap cleanup ERR + +set -e # Exit on any error + +ROOT_DIR="$( cd "$( dirname "${BASH_SOURCE[0]}" )/../.." && pwd )" +BUILD_DIR="${ROOT_DIR}/validation-builds/ios" + +# Configuration +APP_NAME="TVOSLlamaTest" +BUNDLE_ID="org.ggml.TVOSLlamaTest" +XCFRAMEWORK_PATH="${ROOT_DIR}/build-apple/llama.xcframework" +TEMP_DIR="${BUILD_DIR}/temp" +ARCHIVE_PATH="${BUILD_DIR}/${APP_NAME}.xcarchive" +IPA_PATH="${BUILD_DIR}/${APP_NAME}.ipa" +VALIDATION_DIR="${BUILD_DIR}/validation" + +# Create necessary directories +mkdir -p "${BUILD_DIR}" +mkdir -p "${TEMP_DIR}" +mkdir -p "${VALIDATION_DIR}" + +echo "===== tvOS Validation Process Started =====" + +# 1. Create a simple test app project +echo "Creating test tvOS app project..." +mkdir -p "${TEMP_DIR}/${APP_NAME}/${APP_NAME}" +cat > "${TEMP_DIR}/${APP_NAME}/${APP_NAME}/Info.plist" << EOF + + + + + CFBundleDevelopmentRegion + en + CFBundleExecutable + ${APP_NAME} + CFBundleIdentifier + ${BUNDLE_ID} + CFBundleInfoDictionaryVersion + 6.0 + CFBundleName + ${APP_NAME} + CFBundlePackageType + APPL + CFBundleShortVersionString + 1.0 + CFBundleVersion + 1 + UIRequiredDeviceCapabilities + + arm64 + + + +EOF + +# Create SwiftUI app files +mkdir -p "${TEMP_DIR}/${APP_NAME}/${APP_NAME}/Sources" + +# Create App.swift +cat > "${TEMP_DIR}/${APP_NAME}/${APP_NAME}/Sources/App.swift" << EOF +import SwiftUI +import llama + +@main +struct LlamaTestApp: App { + var body: some Scene { + WindowGroup { + ContentView() + } + } +} +EOF + +# Create ContentView.swift with tvOS specific elements +cat > "${TEMP_DIR}/${APP_NAME}/${APP_NAME}/Sources/ContentView.swift" << EOF +import SwiftUI +import llama + +struct ContentView: View { + // Test that we can initialize a llama context params struct + let params = llama_context_default_params() + + var body: some View { + VStack(spacing: 40) { + Text("Llama Framework Test on tvOS") + .font(.largeTitle) + .padding() + + Text("llama_context_default_params() created successfully") + .font(.headline) + .multilineTextAlignment(.center) + .padding() + + // Display some param values to confirm the framework is working + Text("n_ctx: \(params.n_ctx)") + .font(.title2) + + Text("n_batch: \(params.n_batch)") + .font(.title2) + + Spacer() + } + .padding(50) + // Larger size suitable for TV display + } +} + +struct ContentView_Previews: PreviewProvider { + static var previews: some View { + ContentView() + } +} +EOF + +# Create project.pbxproj, fixing the framework search paths issues +mkdir -p "${TEMP_DIR}/${APP_NAME}/${APP_NAME}.xcodeproj" +cat > "${TEMP_DIR}/${APP_NAME}/${APP_NAME}.xcodeproj/project.pbxproj" << 'EOF' +// !$*UTF8*$! +{ + archiveVersion = 1; + classes = { + }; + objectVersion = 54; + objects = { + +/* Begin PBXBuildFile section */ + 11111111111111111111111 /* App.swift in Sources */ = {isa = PBXBuildFile; fileRef = 22222222222222222222222; }; + 33333333333333333333333 /* ContentView.swift in Sources */ = {isa = PBXBuildFile; fileRef = 44444444444444444444444; }; + 55555555555555555555555 /* llama.xcframework in Frameworks */ = {isa = PBXBuildFile; fileRef = 66666666666666666666666; }; + 77777777777777777777777 /* llama.xcframework in Embed Frameworks */ = {isa = PBXBuildFile; fileRef = 66666666666666666666666; }; +/* End PBXBuildFile section */ + +/* Begin PBXCopyFilesBuildPhase section */ + 88888888888888888888888 /* Embed Frameworks */ = { + isa = PBXCopyFilesBuildPhase; + buildActionMask = 2147483647; + dstPath = ""; + dstSubfolderSpec = 10; + files = ( + 77777777777777777777777 /* llama.xcframework in Embed Frameworks */, + ); + name = "Embed Frameworks"; + runOnlyForDeploymentPostprocessing = 0; + }; +/* End PBXCopyFilesBuildPhase section */ + +/* Begin PBXFileReference section */ +EOF + +# Continue with the project.pbxproj file, using the APP_NAME variable appropriately +cat >> "${TEMP_DIR}/${APP_NAME}/${APP_NAME}.xcodeproj/project.pbxproj" << EOF + 99999999999999999999999 /* ${APP_NAME}.app */ = {isa = PBXFileReference; explicitFileType = wrapper.application; includeInIndex = 0; path = "${APP_NAME}.app"; sourceTree = BUILT_PRODUCTS_DIR; }; + 22222222222222222222222 /* App.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = App.swift; sourceTree = ""; }; + 44444444444444444444444 /* ContentView.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = ContentView.swift; sourceTree = ""; }; + AAAAAAAAAAAAAAAAAAAAAAA /* Info.plist */ = {isa = PBXFileReference; lastKnownFileType = text.plist.xml; path = Info.plist; sourceTree = ""; }; + 66666666666666666666666 /* llama.xcframework */ = {isa = PBXFileReference; lastKnownFileType = wrapper.xcframework; path = llama.xcframework; sourceTree = ""; }; +/* End PBXFileReference section */ +EOF + +# Add the rest of the project file with fixed framework search paths +cat >> "${TEMP_DIR}/${APP_NAME}/${APP_NAME}.xcodeproj/project.pbxproj" << 'EOF' +/* Begin PBXFrameworksBuildPhase section */ + BBBBBBBBBBBBBBBBBBBBBBBB /* Frameworks */ = { + isa = PBXFrameworksBuildPhase; + buildActionMask = 2147483647; + files = ( + 55555555555555555555555 /* llama.xcframework in Frameworks */, + ); + runOnlyForDeploymentPostprocessing = 0; + }; +/* End PBXFrameworksBuildPhase section */ + +/* Begin PBXGroup section */ +EOF + +# Continue with the project.pbxproj file, using the APP_NAME variable appropriately +cat >> "${TEMP_DIR}/${APP_NAME}/${APP_NAME}.xcodeproj/project.pbxproj" << EOF + CCCCCCCCCCCCCCCCCCCCCCCC /* Products */ = { + isa = PBXGroup; + children = ( + 99999999999999999999999 /* ${APP_NAME}.app */, + ); + name = Products; + sourceTree = ""; + }; +EOF + +cat >> "${TEMP_DIR}/${APP_NAME}/${APP_NAME}.xcodeproj/project.pbxproj" << 'EOF' + DDDDDDDDDDDDDDDDDDDDDDDD /* Frameworks */ = { + isa = PBXGroup; + children = ( + 66666666666666666666666 /* llama.xcframework */, + ); + name = Frameworks; + sourceTree = ""; + }; + EEEEEEEEEEEEEEEEEEEEEEEE = { + isa = PBXGroup; + children = ( + FFFFFFFFFFFFFFFFFFFFFFFF /* TVOSLlamaTest */, + CCCCCCCCCCCCCCCCCCCCCCCC /* Products */, + DDDDDDDDDDDDDDDDDDDDDDDD /* Frameworks */, + ); + sourceTree = ""; + }; + FFFFFFFFFFFFFFFFFFFFFFFF /* TVOSLlamaTest */ = { + isa = PBXGroup; + children = ( + 1111111111111111111111AA /* Sources */, + AAAAAAAAAAAAAAAAAAAAAAA /* Info.plist */, + ); + path = "TVOSLlamaTest"; + sourceTree = ""; + }; + 1111111111111111111111AA /* Sources */ = { + isa = PBXGroup; + children = ( + 22222222222222222222222 /* App.swift */, + 44444444444444444444444 /* ContentView.swift */, + ); + path = Sources; + sourceTree = ""; + }; +/* End PBXGroup section */ +EOF + +# Continue with the project.pbxproj file, using the APP_NAME variable appropriately +cat >> "${TEMP_DIR}/${APP_NAME}/${APP_NAME}.xcodeproj/project.pbxproj" << EOF +/* Begin PBXNativeTarget section */ + 3333333333333333333333AA /* ${APP_NAME} */ = { + isa = PBXNativeTarget; + buildConfigurationList = 4444444444444444444444AA /* Build configuration list for PBXNativeTarget "${APP_NAME}" */; + buildPhases = ( + 5555555555555555555555AA /* Sources */, + BBBBBBBBBBBBBBBBBBBBBBBB /* Frameworks */, + 6666666666666666666666AA /* Resources */, + 88888888888888888888888 /* Embed Frameworks */, + ); + buildRules = ( + ); + dependencies = ( + ); + name = "${APP_NAME}"; + productName = "${APP_NAME}"; + productReference = 99999999999999999999999 /* ${APP_NAME}.app */; + productType = "com.apple.product-type.application"; + }; +/* End PBXNativeTarget section */ + +/* Begin PBXProject section */ + 7777777777777777777777AA /* Project object */ = { + isa = PBXProject; + attributes = { + LastSwiftUpdateCheck = 1240; + LastUpgradeCheck = 1240; + TargetAttributes = { + 3333333333333333333333AA = { + CreatedOnToolsVersion = 12.4; + }; + }; + }; + buildConfigurationList = 8888888888888888888888AA /* Build configuration list for PBXProject "${APP_NAME}" */; + compatibilityVersion = "Xcode 12.0"; + developmentRegion = en; + hasScannedForEncodings = 0; + knownRegions = ( + en, + Base, + ); + mainGroup = EEEEEEEEEEEEEEEEEEEEEEEE; + productRefGroup = CCCCCCCCCCCCCCCCCCCCCCCC /* Products */; + projectDirPath = ""; + projectRoot = ""; + targets = ( + 3333333333333333333333AA /* ${APP_NAME} */, + ); + }; +/* End PBXProject section */ +EOF + +# Add the rest of the file with correct FRAMEWORK_SEARCH_PATHS and tvOS settings +cat >> "${TEMP_DIR}/${APP_NAME}/${APP_NAME}.xcodeproj/project.pbxproj" << 'EOF' +/* Begin PBXResourcesBuildPhase section */ + 6666666666666666666666AA /* Resources */ = { + isa = PBXResourcesBuildPhase; + buildActionMask = 2147483647; + files = ( + ); + runOnlyForDeploymentPostprocessing = 0; + }; +/* End PBXResourcesBuildPhase section */ + +/* Begin PBXSourcesBuildPhase section */ + 5555555555555555555555AA /* Sources */ = { + isa = PBXSourcesBuildPhase; + buildActionMask = 2147483647; + files = ( + 33333333333333333333333 /* ContentView.swift in Sources */, + 11111111111111111111111 /* App.swift in Sources */, + ); + runOnlyForDeploymentPostprocessing = 0; + }; +/* End PBXSourcesBuildPhase section */ + +/* Begin XCBuildConfiguration section */ + 9999999999999999999999AA /* Debug */ = { + isa = XCBuildConfiguration; + buildSettings = { + ALWAYS_SEARCH_USER_PATHS = NO; + CLANG_ANALYZER_NONNULL = YES; + CLANG_ANALYZER_NUMBER_OBJECT_CONVERSION = YES_AGGRESSIVE; + CLANG_CXX_LANGUAGE_STANDARD = "gnu++14"; + CLANG_CXX_LIBRARY = "libc++"; + CLANG_ENABLE_MODULES = YES; + CLANG_ENABLE_OBJC_ARC = YES; + CLANG_ENABLE_OBJC_WEAK = YES; + CLANG_WARN_BLOCK_CAPTURE_AUTORELEASING = YES; + CLANG_WARN_BOOL_CONVERSION = YES; + CLANG_WARN_COMMA = YES; + CLANG_WARN_CONSTANT_CONVERSION = YES; + CLANG_WARN_DEPRECATED_OBJC_IMPLEMENTATIONS = YES; + CLANG_WARN_DIRECT_OBJC_ISA_USAGE = YES_ERROR; + CLANG_WARN_DOCUMENTATION_COMMENTS = YES; + CLANG_WARN_EMPTY_BODY = YES; + CLANG_WARN_ENUM_CONVERSION = YES; + CLANG_WARN_INFINITE_RECURSION = YES; + CLANG_WARN_INT_CONVERSION = YES; + CLANG_WARN_NON_LITERAL_NULL_CONVERSION = YES; + CLANG_WARN_OBJC_IMPLICIT_RETAIN_SELF = YES; + CLANG_WARN_OBJC_LITERAL_CONVERSION = YES; + CLANG_WARN_OBJC_ROOT_CLASS = YES_ERROR; + CLANG_WARN_QUOTED_INCLUDE_IN_FRAMEWORK_HEADER = YES; + CLANG_WARN_RANGE_LOOP_ANALYSIS = YES; + CLANG_WARN_STRICT_PROTOTYPES = YES; + CLANG_WARN_SUSPICIOUS_MOVE = YES; + CLANG_WARN_UNGUARDED_AVAILABILITY = YES_AGGRESSIVE; + CLANG_WARN_UNREACHABLE_CODE = YES; + CLANG_WARN__DUPLICATE_METHOD_MATCH = YES; + COPY_PHASE_STRIP = NO; + DEBUG_INFORMATION_FORMAT = dwarf; + ENABLE_STRICT_OBJC_MSGSEND = YES; + ENABLE_TESTABILITY = YES; + GCC_C_LANGUAGE_STANDARD = gnu11; + GCC_DYNAMIC_NO_PIC = NO; + GCC_NO_COMMON_BLOCKS = YES; + GCC_OPTIMIZATION_LEVEL = 0; + GCC_PREPROCESSOR_DEFINITIONS = ( + "DEBUG=1", + "$(inherited)", + ); + GCC_WARN_64_TO_32_BIT_CONVERSION = YES; + GCC_WARN_ABOUT_RETURN_TYPE = YES_ERROR; + GCC_WARN_UNDECLARED_SELECTOR = YES; + GCC_WARN_UNINITIALIZED_AUTOS = YES_AGGRESSIVE; + GCC_WARN_UNUSED_FUNCTION = YES; + GCC_WARN_UNUSED_VARIABLE = YES; + TVOS_DEPLOYMENT_TARGET = 15.0; + MTL_ENABLE_DEBUG_INFO = INCLUDE_SOURCE; + MTL_FAST_MATH = YES; + ONLY_ACTIVE_ARCH = YES; + SDKROOT = appletvos; + SWIFT_ACTIVE_COMPILATION_CONDITIONS = DEBUG; + SWIFT_OPTIMIZATION_LEVEL = "-Onone"; + }; + name = Debug; + }; + AAAAAAAAAAAAAAAAAAAAABBB /* Release */ = { + isa = XCBuildConfiguration; + buildSettings = { + ALWAYS_SEARCH_USER_PATHS = NO; + CLANG_ANALYZER_NONNULL = YES; + CLANG_ANALYZER_NUMBER_OBJECT_CONVERSION = YES_AGGRESSIVE; + CLANG_CXX_LANGUAGE_STANDARD = "gnu++14"; + CLANG_CXX_LIBRARY = "libc++"; + CLANG_ENABLE_MODULES = YES; + CLANG_ENABLE_OBJC_ARC = YES; + CLANG_ENABLE_OBJC_WEAK = YES; + CLANG_WARN_BLOCK_CAPTURE_AUTORELEASING = YES; + CLANG_WARN_BOOL_CONVERSION = YES; + CLANG_WARN_COMMA = YES; + CLANG_WARN_CONSTANT_CONVERSION = YES; + CLANG_WARN_DEPRECATED_OBJC_IMPLEMENTATIONS = YES; + CLANG_WARN_DIRECT_OBJC_ISA_USAGE = YES_ERROR; + CLANG_WARN_DOCUMENTATION_COMMENTS = YES; + CLANG_WARN_EMPTY_BODY = YES; + CLANG_WARN_ENUM_CONVERSION = YES; + CLANG_WARN_INFINITE_RECURSION = YES; + CLANG_WARN_INT_CONVERSION = YES; + CLANG_WARN_NON_LITERAL_NULL_CONVERSION = YES; + CLANG_WARN_OBJC_IMPLICIT_RETAIN_SELF = YES; + CLANG_WARN_OBJC_LITERAL_CONVERSION = YES; + CLANG_WARN_OBJC_ROOT_CLASS = YES_ERROR; + CLANG_WARN_QUOTED_INCLUDE_IN_FRAMEWORK_HEADER = YES; + CLANG_WARN_RANGE_LOOP_ANALYSIS = YES; + CLANG_WARN_STRICT_PROTOTYPES = YES; + CLANG_WARN_SUSPICIOUS_MOVE = YES; + CLANG_WARN_UNGUARDED_AVAILABILITY = YES_AGGRESSIVE; + CLANG_WARN_UNREACHABLE_CODE = YES; + CLANG_WARN__DUPLICATE_METHOD_MATCH = YES; + COPY_PHASE_STRIP = NO; + DEBUG_INFORMATION_FORMAT = "dwarf-with-dsym"; + ENABLE_NS_ASSERTIONS = NO; + ENABLE_STRICT_OBJC_MSGSEND = YES; + GCC_C_LANGUAGE_STANDARD = gnu11; + GCC_NO_COMMON_BLOCKS = YES; + GCC_WARN_64_TO_32_BIT_CONVERSION = YES; + GCC_WARN_ABOUT_RETURN_TYPE = YES_ERROR; + GCC_WARN_UNDECLARED_SELECTOR = YES; + GCC_WARN_UNINITIALIZED_AUTOS = YES_AGGRESSIVE; + GCC_WARN_UNUSED_FUNCTION = YES; + GCC_WARN_UNUSED_VARIABLE = YES; + TVOS_DEPLOYMENT_TARGET = 15.0; + MTL_ENABLE_DEBUG_INFO = NO; + MTL_FAST_MATH = YES; + SDKROOT = appletvos; + SWIFT_COMPILATION_MODE = wholemodule; + SWIFT_OPTIMIZATION_LEVEL = "-O"; + VALIDATE_PRODUCT = YES; + }; + name = Release; + }; + BBBBBBBBBBBBBBBBBBBBBBCCC /* Debug */ = { + isa = XCBuildConfiguration; + buildSettings = { + ASSETCATALOG_COMPILER_APPICON_NAME = AppIcon; + ASSETCATALOG_COMPILER_GLOBAL_ACCENT_COLOR_NAME = AccentColor; + CODE_SIGN_STYLE = Manual; + DEVELOPMENT_TEAM = ""; + ENABLE_PREVIEWS = YES; + FRAMEWORK_SEARCH_PATHS = "$(PROJECT_DIR)"; + INFOPLIST_FILE = "TVOSLlamaTest/Info.plist"; + LD_RUNPATH_SEARCH_PATHS = ( + "$(inherited)", + "@executable_path/Frameworks", + ); + PRODUCT_BUNDLE_IDENTIFIER = "org.ggml.TVOSLlamaTest"; + PRODUCT_NAME = "$(TARGET_NAME)"; + PROVISIONING_PROFILE_SPECIFIER = ""; + SWIFT_VERSION = 5.0; + TARGETED_DEVICE_FAMILY = 3; + }; + name = Debug; + }; + CCCCCCCCCCCCCCCCCCCCCCDDD /* Release */ = { + isa = XCBuildConfiguration; + buildSettings = { + ASSETCATALOG_COMPILER_APPICON_NAME = AppIcon; + ASSETCATALOG_COMPILER_GLOBAL_ACCENT_COLOR_NAME = AccentColor; + CODE_SIGN_STYLE = Manual; + DEVELOPMENT_TEAM = ""; + ENABLE_PREVIEWS = YES; + FRAMEWORK_SEARCH_PATHS = ( + "$(inherited)", + "$(PROJECT_DIR)", + ); + INFOPLIST_FILE = "TVOSLlamaTest/Info.plist"; + LD_RUNPATH_SEARCH_PATHS = ( + "$(inherited)", + "@executable_path/Frameworks", + ); + PRODUCT_BUNDLE_IDENTIFIER = "org.ggml.TVOSLlamaTest"; + PRODUCT_NAME = "$(TARGET_NAME)"; + PROVISIONING_PROFILE_SPECIFIER = ""; + SWIFT_VERSION = 5.0; + TARGETED_DEVICE_FAMILY = 3; + }; + name = Release; + }; +/* End XCBuildConfiguration section */ +EOF + +# Finish the project.pbxproj file +cat >> "${TEMP_DIR}/${APP_NAME}/${APP_NAME}.xcodeproj/project.pbxproj" << EOF +/* Begin XCConfigurationList section */ + 8888888888888888888888AA /* Build configuration list for PBXProject "${APP_NAME}" */ = { + isa = XCConfigurationList; + buildConfigurations = ( + 9999999999999999999999AA /* Debug */, + AAAAAAAAAAAAAAAAAAAAABBB /* Release */, + ); + defaultConfigurationIsVisible = 0; + defaultConfigurationName = Release; + }; + 4444444444444444444444AA /* Build configuration list for PBXNativeTarget "${APP_NAME}" */ = { + isa = XCConfigurationList; + buildConfigurations = ( + BBBBBBBBBBBBBBBBBBBBBBCCC /* Debug */, + CCCCCCCCCCCCCCCCCCCCCCDDD /* Release */, + ); + defaultConfigurationIsVisible = 0; + defaultConfigurationName = Release; + }; +/* End XCConfigurationList section */ + }; + rootObject = 7777777777777777777777AA /* Project object */; +} +EOF + +# 2. Copy XCFramework to test project +echo "Copying XCFramework to test project..." +cp -R "${XCFRAMEWORK_PATH}" "${TEMP_DIR}/${APP_NAME}/" + +# 3. Build and archive the app +echo "Building and archiving test app..." +cd "${TEMP_DIR}/${APP_NAME}" + +# Create a simple xcscheme file to avoid xcodebuild scheme issues +mkdir -p "${APP_NAME}.xcodeproj/xcshareddata/xcschemes" +cat > "${APP_NAME}.xcodeproj/xcshareddata/xcschemes/${APP_NAME}.xcscheme" << EOF + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + +EOF + +# Now use xcodebuild with an explicitly defined product name for tvOS +xcodebuild -project "${APP_NAME}.xcodeproj" -scheme "${APP_NAME}" -sdk appletvos -configuration Release archive -archivePath "${ARCHIVE_PATH}" CODE_SIGN_IDENTITY="-" CODE_SIGNING_REQUIRED=NO CODE_SIGNING_ALLOWED=NO PRODUCT_NAME="${APP_NAME}" SWIFT_OPTIMIZATION_LEVEL="-Onone" -quiet + +# 4. Create IPA from archive +echo "Creating IPA from archive..." +mkdir -p "${TEMP_DIR}/Payload" +cp -R "${ARCHIVE_PATH}/Products/Applications/${APP_NAME}.app" "${TEMP_DIR}/Payload/" + +# Check and log app structure before zipping +echo "App structure:" +ls -la "${TEMP_DIR}/Payload/${APP_NAME}.app/" +echo "Frameworks:" +ls -la "${TEMP_DIR}/Payload/${APP_NAME}.app/Frameworks/" 2>/dev/null || echo "No Frameworks directory found" + +cd "${TEMP_DIR}" +zip -r "${IPA_PATH}" Payload + +# Check embedded provisioning profile +echo "Checking provisioning profile (if any)..." +PROVISIONING_PROFILE=$(find "${ARCHIVE_PATH}/Products/Applications/${APP_NAME}.app" -name "embedded.mobileprovision" 2>/dev/null) +if [ -n "$PROVISIONING_PROFILE" ]; then + echo "Found embedded provisioning profile:" + security cms -D -i "$PROVISIONING_PROFILE" || echo "Unable to decode provisioning profile" +else + echo "No embedded provisioning profile found (expected for ad-hoc builds)" +fi + +# 5. Validate the IPA +echo "Validating IPA..." +VALIDATION_OUTPUT="${VALIDATION_DIR}/validation_output.txt" + +# Check if authentication credentials are provided +AUTH_ARGS="" +if [ -n "$APPLE_ID" ] && [ -n "$APPLE_PASSWORD" ]; then + echo "Using Apple ID authentication for validation..." + AUTH_ARGS="--username \"$APPLE_ID\" --password \"$APPLE_PASSWORD\"" +else + echo "No authentication credentials provided. Will perform basic validation." + echo "To use your personal developer account, you can run the script with:" + echo " APPLE_ID='your.email@example.com' APPLE_PASSWORD='your-app-specific-password' ./validate-tvos.sh" + echo "Note: You need to create an app-specific password at https://appleid.apple.com/account/manage" +fi + +# Run validation with detailed output +echo "Running validation with altool..." +if [ -n "$AUTH_ARGS" ]; then + # Use eval to properly handle the quoted arguments + eval "xcrun altool --validate-app -f \"${IPA_PATH}\" --type tvos --output-format xml $AUTH_ARGS" 2>&1 | tee "${VALIDATION_OUTPUT}" +else + xcrun altool --validate-app -f "${IPA_PATH}" --type tvos --output-format xml 2>&1 | tee "${VALIDATION_OUTPUT}" +fi +VALIDATION_RESULT=$? + +# Final validation result +FINAL_VALIDATION_RESULT=0 + +# Check if validation failed because the app isn't in App Store Connect +if grep -q "No suitable application records were found" "${VALIDATION_OUTPUT}"; then + echo "⚠️ App Store Connect Warning: The app bundle identifier is not found in App Store Connect" + echo "This is expected for apps that haven't been registered in App Store Connect yet." + echo "This doesn't indicate a problem with the build or framework." + + # Perform alternative validation + echo "Performing alternative validation checks..." + + # Check if IPA was created successfully + if [ -f "${IPA_PATH}" ] && [ -s "${IPA_PATH}" ]; then + echo "✅ IPA file created successfully" + else + echo "❌ IPA file not created or empty" + FINAL_VALIDATION_RESULT=1 + fi + + # Check if app binary exists and is executable + if [ -f "${TEMP_DIR}/Payload/${APP_NAME}.app/${APP_NAME}" ] && [ -x "${TEMP_DIR}/Payload/${APP_NAME}.app/${APP_NAME}" ]; then + echo "✅ App binary exists and is executable" + else + echo "❌ App binary missing or not executable" + FINAL_VALIDATION_RESULT=1 + fi + + # Check if framework was properly embedded + if [ -d "${TEMP_DIR}/Payload/${APP_NAME}.app/Frameworks/llama.framework" ]; then + echo "✅ llama.framework properly embedded" + else + echo "❌ llama.framework not properly embedded" + FINAL_VALIDATION_RESULT=1 + fi + + # Check if framework binary exists + if [ -f "${TEMP_DIR}/Payload/${APP_NAME}.app/Frameworks/llama.framework/llama" ]; then + echo "✅ Framework binary exists" + + # Further validate framework by checking architecture + ARCHS=$(lipo -info "${TEMP_DIR}/Payload/${APP_NAME}.app/Frameworks/llama.framework/llama" 2>/dev/null | grep -o "arm64\\|x86_64" | tr '\n' ' ') + if [ -n "$ARCHS" ]; then + echo "✅ Framework architecture(s): $ARCHS" + else + echo "⚠️ Could not determine framework architecture" + fi + else + echo "❌ Framework binary missing" + FINAL_VALIDATION_RESULT=1 + fi + + if [ $FINAL_VALIDATION_RESULT -eq 0 ]; then + echo "✅ Alternative validation PASSED: App built successfully with embedded framework" + else + echo "❌ Alternative validation FAILED: Issues found with the app or framework" + fi +elif grep -q "You must specify authentication credentials" "${VALIDATION_OUTPUT}" && [ -z "$AUTH_ARGS" ]; then + echo "✅ tvOS Validation PASSED: IPA successfully validated" + echo "Results saved to ${VALIDATION_OUTPUT}" +else + echo "❌ tvOS Validation FAILED: IPA validation found issues" + echo "See validation output at ${VALIDATION_OUTPUT}" + echo "" + echo "==== VALIDATION ERRORS ====" + + # Try to extract specific errors from the output + if grep -q "Error" "${VALIDATION_OUTPUT}"; then + grep -A 5 "Error" "${VALIDATION_OUTPUT}" + else + # If no specific error found, show the whole log + cat "${VALIDATION_OUTPUT}" + fi + + # Additional debugging: check IPA contents + echo "" + echo "==== IPA CONTENTS ====" + mkdir -p "${TEMP_DIR}/ipa_contents" + unzip -q "${IPA_PATH}" -d "${TEMP_DIR}/ipa_contents" + ls -la "${TEMP_DIR}/ipa_contents/Payload/${APP_NAME}.app/" + + # Check for code signing issues + echo "" + echo "==== CODE SIGNING INFO ====" + codesign -vv -d "${TEMP_DIR}/ipa_contents/Payload/${APP_NAME}.app" 2>&1 || echo "Code signing verification failed" + + # Check embedded frameworks + echo "" + echo "==== FRAMEWORK INFO ====" + ls -la "${TEMP_DIR}/ipa_contents/Payload/${APP_NAME}.app/Frameworks/" 2>/dev/null || echo "No Frameworks directory found" +fi + +# Don't clean up on error to allow inspection +if [ $FINAL_VALIDATION_RESULT -ne 0 ]; then + echo "" + echo "Temporary files kept for inspection at: ${TEMP_DIR}" + echo "===== tvOS Validation Process Failed =====" + exit 1 +fi + +# Clean up temporary files but keep build artifacts +if [ $FINAL_VALIDATION_RESULT -eq 0 ]; then + echo "Cleaning up temporary files..." + #rm -rf "${TEMP_DIR}" +fi + +echo "===== tvOS Validation Process Completed =====" +exit $FINAL_VALIDATION_RESULT diff --git a/scripts/apple/validate-visionos.sh b/scripts/apple/validate-visionos.sh new file mode 100755 index 000000000..a18ddcce4 --- /dev/null +++ b/scripts/apple/validate-visionos.sh @@ -0,0 +1,811 @@ +#!/bin/bash +# validate-visionos.sh - Validate visionOS Application with embedded llama.xcframework using SwiftUI + +# Authentication options (optional) (can be set via environment variables) +# To use: export APPLE_ID=your.email@example.com +# export APPLE_PASSWORD=your-app-specific-password +# ./validate-visionos.sh +APPLE_ID=${APPLE_ID:-""} +APPLE_PASSWORD=${APPLE_PASSWORD:-""} + +# Ensure the script exits on error +set -e + +# Function to print usage instructions +print_usage() { + echo "Usage: ./validate-visionos.sh [OPTIONS]" + echo "" + echo "Options:" + echo " --help Show this help message" + echo " --apple-id EMAIL Apple ID email for validation" + echo " --apple-password PWD App-specific password for Apple ID" + echo "" + echo "Environment variables:" + echo " APPLE_ID Apple ID email for validation" + echo " APPLE_PASSWORD App-specific password for Apple ID" + echo "" + echo "Notes:" + echo " - Command line options take precedence over environment variables" + echo " - Authentication is optional. If not provided, alternative validation will be performed" + echo " - For APPLE_PASSWORD, use an app-specific password generated at https://appleid.apple.com/account/manage" +} + +# Parse command line arguments +while [[ $# -gt 0 ]]; do + case $1 in + --help) + print_usage + exit 0 + ;; + --apple-id) + APPLE_ID="$2" + shift 2 + ;; + --apple-password) + APPLE_PASSWORD="$2" + shift 2 + ;; + *) + echo "Unknown option: $1" + print_usage + exit 1 + ;; + esac +done + +# Function to clean up in case of error +cleanup() { + # Don't clean up temp files on error to help with debugging + echo "===== visionOS Validation Process Failed =====" + exit 1 +} + +# Set up trap to call cleanup function on error +trap cleanup ERR + +set -e # Exit on any error + +ROOT_DIR="$( cd "$( dirname "${BASH_SOURCE[0]}" )/../.." && pwd )" +BUILD_DIR="${ROOT_DIR}/validation-builds/visionos" + +# Configuration +APP_NAME="VisionOSLlamaTest" +BUNDLE_ID="org.ggml.VisionOSLlamaTest" +XCFRAMEWORK_PATH="${ROOT_DIR}/build-apple/llama.xcframework" +TEMP_DIR="${BUILD_DIR}/temp" +ARCHIVE_PATH="${BUILD_DIR}/${APP_NAME}.xcarchive" +IPA_PATH="${BUILD_DIR}/${APP_NAME}.ipa" +VALIDATION_DIR="${BUILD_DIR}/validation" + +# Create necessary directories +mkdir -p "${BUILD_DIR}" +mkdir -p "${TEMP_DIR}" +mkdir -p "${VALIDATION_DIR}" + +echo "===== visionOS Validation Process Started =====" + +# 1. Create a simple test app project +echo "Creating test visionOS app project..." +mkdir -p "${TEMP_DIR}/${APP_NAME}/${APP_NAME}" +cat > "${TEMP_DIR}/${APP_NAME}/${APP_NAME}/Info.plist" << EOF + + + + + CFBundleDevelopmentRegion + en + CFBundleExecutable + ${APP_NAME} + CFBundleIdentifier + ${BUNDLE_ID} + CFBundleInfoDictionaryVersion + 6.0 + CFBundleName + ${APP_NAME} + CFBundlePackageType + APPL + CFBundleShortVersionString + 1.0 + CFBundleVersion + 1 + + +EOF + +# Create SwiftUI app files +mkdir -p "${TEMP_DIR}/${APP_NAME}/${APP_NAME}/Sources" + +# Create App.swift +cat > "${TEMP_DIR}/${APP_NAME}/${APP_NAME}/Sources/App.swift" << EOF +import SwiftUI +import llama + +@main +struct LlamaTestApp: App { + var body: some Scene { + WindowGroup { + ContentView() + } + } +} +EOF + +# Create ContentView.swift with visionOS specific elements +cat > "${TEMP_DIR}/${APP_NAME}/${APP_NAME}/Sources/ContentView.swift" << EOF +import SwiftUI +import llama + +struct ContentView: View { + // Test that we can initialize a llama context params struct + let params = llama_context_default_params() + + var body: some View { + VStack(spacing: 20) { + Text("Llama Framework Test on visionOS") + .font(.largeTitle) + .padding() + + Text("llama_context_default_params() created successfully") + .font(.headline) + .multilineTextAlignment(.center) + .padding() + + // Display some param values to confirm the framework is working + Text("n_ctx: \(params.n_ctx)") + .font(.body) + + Text("n_batch: \(params.n_batch)") + .font(.body) + + Spacer() + } + .padding() + .frame(width: 500, height: 400) + } +} + +struct ContentView_Previews: PreviewProvider { + static var previews: some View { + ContentView() + } +} +EOF + +# Create project.pbxproj, fixing the framework search paths issues +mkdir -p "${TEMP_DIR}/${APP_NAME}/${APP_NAME}.xcodeproj" +cat > "${TEMP_DIR}/${APP_NAME}/${APP_NAME}.xcodeproj/project.pbxproj" << 'EOF' +// !$*UTF8*$! +{ + archiveVersion = 1; + classes = { + }; + objectVersion = 54; + objects = { + +/* Begin PBXBuildFile section */ + 11111111111111111111111 /* App.swift in Sources */ = {isa = PBXBuildFile; fileRef = 22222222222222222222222; }; + 33333333333333333333333 /* ContentView.swift in Sources */ = {isa = PBXBuildFile; fileRef = 44444444444444444444444; }; + 55555555555555555555555 /* llama.xcframework in Frameworks */ = {isa = PBXBuildFile; fileRef = 66666666666666666666666; }; + 77777777777777777777777 /* llama.xcframework in Embed Frameworks */ = {isa = PBXBuildFile; fileRef = 66666666666666666666666; }; +/* End PBXBuildFile section */ + +/* Begin PBXCopyFilesBuildPhase section */ + 88888888888888888888888 /* Embed Frameworks */ = { + isa = PBXCopyFilesBuildPhase; + buildActionMask = 2147483647; + dstPath = ""; + dstSubfolderSpec = 10; + files = ( + 77777777777777777777777 /* llama.xcframework in Embed Frameworks */, + ); + name = "Embed Frameworks"; + runOnlyForDeploymentPostprocessing = 0; + }; +/* End PBXCopyFilesBuildPhase section */ + +/* Begin PBXFileReference section */ +EOF + +# Continue with the project.pbxproj file, using the APP_NAME variable appropriately +cat >> "${TEMP_DIR}/${APP_NAME}/${APP_NAME}.xcodeproj/project.pbxproj" << EOF + 99999999999999999999999 /* ${APP_NAME}.app */ = {isa = PBXFileReference; explicitFileType = wrapper.application; includeInIndex = 0; path = "${APP_NAME}.app"; sourceTree = BUILT_PRODUCTS_DIR; }; + 22222222222222222222222 /* App.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = App.swift; sourceTree = ""; }; + 44444444444444444444444 /* ContentView.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = ContentView.swift; sourceTree = ""; }; + AAAAAAAAAAAAAAAAAAAAAAA /* Info.plist */ = {isa = PBXFileReference; lastKnownFileType = text.plist.xml; path = Info.plist; sourceTree = ""; }; + 66666666666666666666666 /* llama.xcframework */ = {isa = PBXFileReference; lastKnownFileType = wrapper.xcframework; path = llama.xcframework; sourceTree = ""; }; +/* End PBXFileReference section */ +EOF + +# Add the rest of the project file with fixed framework search paths +cat >> "${TEMP_DIR}/${APP_NAME}/${APP_NAME}.xcodeproj/project.pbxproj" << 'EOF' +/* Begin PBXFrameworksBuildPhase section */ + BBBBBBBBBBBBBBBBBBBBBBBB /* Frameworks */ = { + isa = PBXFrameworksBuildPhase; + buildActionMask = 2147483647; + files = ( + 55555555555555555555555 /* llama.xcframework in Frameworks */, + ); + runOnlyForDeploymentPostprocessing = 0; + }; +/* End PBXFrameworksBuildPhase section */ + +/* Begin PBXGroup section */ +EOF + +# Continue with the project.pbxproj file, using the APP_NAME variable appropriately +cat >> "${TEMP_DIR}/${APP_NAME}/${APP_NAME}.xcodeproj/project.pbxproj" << EOF + CCCCCCCCCCCCCCCCCCCCCCCC /* Products */ = { + isa = PBXGroup; + children = ( + 99999999999999999999999 /* ${APP_NAME}.app */, + ); + name = Products; + sourceTree = ""; + }; +EOF + +cat >> "${TEMP_DIR}/${APP_NAME}/${APP_NAME}.xcodeproj/project.pbxproj" << 'EOF' + DDDDDDDDDDDDDDDDDDDDDDDD /* Frameworks */ = { + isa = PBXGroup; + children = ( + 66666666666666666666666 /* llama.xcframework */, + ); + name = Frameworks; + sourceTree = ""; + }; + EEEEEEEEEEEEEEEEEEEEEEEE = { + isa = PBXGroup; + children = ( + FFFFFFFFFFFFFFFFFFFFFFFF /* VisionOSLlamaTest */, + CCCCCCCCCCCCCCCCCCCCCCCC /* Products */, + DDDDDDDDDDDDDDDDDDDDDDDD /* Frameworks */, + ); + sourceTree = ""; + }; + FFFFFFFFFFFFFFFFFFFFFFFF /* VisionOSLlamaTest */ = { + isa = PBXGroup; + children = ( + 1111111111111111111111AA /* Sources */, + AAAAAAAAAAAAAAAAAAAAAAA /* Info.plist */, + ); + path = "VisionOSLlamaTest"; + sourceTree = ""; + }; + 1111111111111111111111AA /* Sources */ = { + isa = PBXGroup; + children = ( + 22222222222222222222222 /* App.swift */, + 44444444444444444444444 /* ContentView.swift */, + ); + path = Sources; + sourceTree = ""; + }; +/* End PBXGroup section */ +EOF + +# Continue with the project.pbxproj file, using the APP_NAME variable appropriately +cat >> "${TEMP_DIR}/${APP_NAME}/${APP_NAME}.xcodeproj/project.pbxproj" << EOF +/* Begin PBXNativeTarget section */ + 3333333333333333333333AA /* ${APP_NAME} */ = { + isa = PBXNativeTarget; + buildConfigurationList = 4444444444444444444444AA /* Build configuration list for PBXNativeTarget "${APP_NAME}" */; + buildPhases = ( + 5555555555555555555555AA /* Sources */, + BBBBBBBBBBBBBBBBBBBBBBBB /* Frameworks */, + 6666666666666666666666AA /* Resources */, + 88888888888888888888888 /* Embed Frameworks */, + ); + buildRules = ( + ); + dependencies = ( + ); + name = "${APP_NAME}"; + productName = "${APP_NAME}"; + productReference = 99999999999999999999999 /* ${APP_NAME}.app */; + productType = "com.apple.product-type.application"; + }; +/* End PBXNativeTarget section */ + +/* Begin PBXProject section */ + 7777777777777777777777AA /* Project object */ = { + isa = PBXProject; + attributes = { + LastSwiftUpdateCheck = 1510; + LastUpgradeCheck = 1510; + TargetAttributes = { + 3333333333333333333333AA = { + CreatedOnToolsVersion = 15.1; + }; + }; + }; + buildConfigurationList = 8888888888888888888888AA /* Build configuration list for PBXProject "${APP_NAME}" */; + compatibilityVersion = "Xcode 15.0"; + developmentRegion = en; + hasScannedForEncodings = 0; + knownRegions = ( + en, + Base, + ); + mainGroup = EEEEEEEEEEEEEEEEEEEEEEEE; + productRefGroup = CCCCCCCCCCCCCCCCCCCCCCCC /* Products */; + projectDirPath = ""; + projectRoot = ""; + targets = ( + 3333333333333333333333AA /* ${APP_NAME} */, + ); + }; +/* End PBXProject section */ +EOF + +# Add the rest of the file with correct FRAMEWORK_SEARCH_PATHS +cat >> "${TEMP_DIR}/${APP_NAME}/${APP_NAME}.xcodeproj/project.pbxproj" << 'EOF' +/* Begin PBXResourcesBuildPhase section */ + 6666666666666666666666AA /* Resources */ = { + isa = PBXResourcesBuildPhase; + buildActionMask = 2147483647; + files = ( + ); + runOnlyForDeploymentPostprocessing = 0; + }; +/* End PBXResourcesBuildPhase section */ + +/* Begin PBXSourcesBuildPhase section */ + 5555555555555555555555AA /* Sources */ = { + isa = PBXSourcesBuildPhase; + buildActionMask = 2147483647; + files = ( + 33333333333333333333333 /* ContentView.swift in Sources */, + 11111111111111111111111 /* App.swift in Sources */, + ); + runOnlyForDeploymentPostprocessing = 0; + }; +/* End PBXSourcesBuildPhase section */ + +/* Begin XCBuildConfiguration section */ + 9999999999999999999999AA /* Debug */ = { + isa = XCBuildConfiguration; + buildSettings = { + ALWAYS_SEARCH_USER_PATHS = NO; + CLANG_ANALYZER_NONNULL = YES; + CLANG_ANALYZER_NUMBER_OBJECT_CONVERSION = YES_AGGRESSIVE; + CLANG_CXX_LANGUAGE_STANDARD = "gnu++14"; + CLANG_CXX_LIBRARY = "libc++"; + CLANG_ENABLE_MODULES = YES; + CLANG_ENABLE_OBJC_ARC = YES; + CLANG_ENABLE_OBJC_WEAK = YES; + CLANG_WARN_BLOCK_CAPTURE_AUTORELEASING = YES; + CLANG_WARN_BOOL_CONVERSION = YES; + CLANG_WARN_COMMA = YES; + CLANG_WARN_CONSTANT_CONVERSION = YES; + CLANG_WARN_DEPRECATED_OBJC_IMPLEMENTATIONS = YES; + CLANG_WARN_DIRECT_OBJC_ISA_USAGE = YES_ERROR; + CLANG_WARN_DOCUMENTATION_COMMENTS = YES; + CLANG_WARN_EMPTY_BODY = YES; + CLANG_WARN_ENUM_CONVERSION = YES; + CLANG_WARN_INFINITE_RECURSION = YES; + CLANG_WARN_INT_CONVERSION = YES; + CLANG_WARN_NON_LITERAL_NULL_CONVERSION = YES; + CLANG_WARN_OBJC_IMPLICIT_RETAIN_SELF = YES; + CLANG_WARN_OBJC_LITERAL_CONVERSION = YES; + CLANG_WARN_OBJC_ROOT_CLASS = YES_ERROR; + CLANG_WARN_QUOTED_INCLUDE_IN_FRAMEWORK_HEADER = YES; + CLANG_WARN_RANGE_LOOP_ANALYSIS = YES; + CLANG_WARN_STRICT_PROTOTYPES = YES; + CLANG_WARN_SUSPICIOUS_MOVE = YES; + CLANG_WARN_UNGUARDED_AVAILABILITY = YES_AGGRESSIVE; + CLANG_WARN_UNREACHABLE_CODE = YES; + CLANG_WARN__DUPLICATE_METHOD_MATCH = YES; + COPY_PHASE_STRIP = NO; + DEBUG_INFORMATION_FORMAT = dwarf; + ENABLE_STRICT_OBJC_MSGSEND = YES; + ENABLE_TESTABILITY = YES; + GCC_C_LANGUAGE_STANDARD = gnu11; + GCC_DYNAMIC_NO_PIC = NO; + GCC_NO_COMMON_BLOCKS = YES; + GCC_OPTIMIZATION_LEVEL = 0; + GCC_PREPROCESSOR_DEFINITIONS = ( + "DEBUG=1", + "$(inherited)", + ); + GCC_WARN_64_TO_32_BIT_CONVERSION = YES; + GCC_WARN_ABOUT_RETURN_TYPE = YES_ERROR; + GCC_WARN_UNDECLARED_SELECTOR = YES; + GCC_WARN_UNINITIALIZED_AUTOS = YES_AGGRESSIVE; + GCC_WARN_UNUSED_FUNCTION = YES; + GCC_WARN_UNUSED_VARIABLE = YES; + MTL_ENABLE_DEBUG_INFO = INCLUDE_SOURCE; + MTL_FAST_MATH = YES; + ONLY_ACTIVE_ARCH = YES; + SDKROOT = xros; + SWIFT_ACTIVE_COMPILATION_CONDITIONS = DEBUG; + SWIFT_OPTIMIZATION_LEVEL = "-Onone"; + XROS_DEPLOYMENT_TARGET = 1.0; + }; + name = Debug; + }; + AAAAAAAAAAAAAAAAAAAAABBB /* Release */ = { + isa = XCBuildConfiguration; + buildSettings = { + ALWAYS_SEARCH_USER_PATHS = NO; + CLANG_ANALYZER_NONNULL = YES; + CLANG_ANALYZER_NUMBER_OBJECT_CONVERSION = YES_AGGRESSIVE; + CLANG_CXX_LANGUAGE_STANDARD = "gnu++14"; + CLANG_CXX_LIBRARY = "libc++"; + CLANG_ENABLE_MODULES = YES; + CLANG_ENABLE_OBJC_ARC = YES; + CLANG_ENABLE_OBJC_WEAK = YES; + CLANG_WARN_BLOCK_CAPTURE_AUTORELEASING = YES; + CLANG_WARN_BOOL_CONVERSION = YES; + CLANG_WARN_COMMA = YES; + CLANG_WARN_CONSTANT_CONVERSION = YES; + CLANG_WARN_DEPRECATED_OBJC_IMPLEMENTATIONS = YES; + CLANG_WARN_DIRECT_OBJC_ISA_USAGE = YES_ERROR; + CLANG_WARN_DOCUMENTATION_COMMENTS = YES; + CLANG_WARN_EMPTY_BODY = YES; + CLANG_WARN_ENUM_CONVERSION = YES; + CLANG_WARN_INFINITE_RECURSION = YES; + CLANG_WARN_INT_CONVERSION = YES; + CLANG_WARN_NON_LITERAL_NULL_CONVERSION = YES; + CLANG_WARN_OBJC_IMPLICIT_RETAIN_SELF = YES; + CLANG_WARN_OBJC_LITERAL_CONVERSION = YES; + CLANG_WARN_OBJC_ROOT_CLASS = YES_ERROR; + CLANG_WARN_QUOTED_INCLUDE_IN_FRAMEWORK_HEADER = YES; + CLANG_WARN_RANGE_LOOP_ANALYSIS = YES; + CLANG_WARN_STRICT_PROTOTYPES = YES; + CLANG_WARN_SUSPICIOUS_MOVE = YES; + CLANG_WARN_UNGUARDED_AVAILABILITY = YES_AGGRESSIVE; + CLANG_WARN_UNREACHABLE_CODE = YES; + CLANG_WARN__DUPLICATE_METHOD_MATCH = YES; + COPY_PHASE_STRIP = NO; + DEBUG_INFORMATION_FORMAT = "dwarf-with-dsym"; + ENABLE_NS_ASSERTIONS = NO; + ENABLE_STRICT_OBJC_MSGSEND = YES; + GCC_C_LANGUAGE_STANDARD = gnu11; + GCC_NO_COMMON_BLOCKS = YES; + GCC_WARN_64_TO_32_BIT_CONVERSION = YES; + GCC_WARN_ABOUT_RETURN_TYPE = YES_ERROR; + GCC_WARN_UNDECLARED_SELECTOR = YES; + GCC_WARN_UNINITIALIZED_AUTOS = YES_AGGRESSIVE; + GCC_WARN_UNUSED_FUNCTION = YES; + GCC_WARN_UNUSED_VARIABLE = YES; + MTL_ENABLE_DEBUG_INFO = NO; + MTL_FAST_MATH = YES; + SDKROOT = xros; + SWIFT_COMPILATION_MODE = wholemodule; + SWIFT_OPTIMIZATION_LEVEL = "-O"; + VALIDATE_PRODUCT = YES; + XROS_DEPLOYMENT_TARGET = 1.0; + }; + name = Release; + }; + BBBBBBBBBBBBBBBBBBBBBBCCC /* Debug */ = { + isa = XCBuildConfiguration; + buildSettings = { + ASSETCATALOG_COMPILER_APPICON_NAME = AppIcon; + ASSETCATALOG_COMPILER_GLOBAL_ACCENT_COLOR_NAME = AccentColor; + CODE_SIGN_STYLE = Manual; + DEVELOPMENT_TEAM = ""; + ENABLE_PREVIEWS = YES; + FRAMEWORK_SEARCH_PATHS = "$(PROJECT_DIR)"; + INFOPLIST_FILE = "VisionOSLlamaTest/Info.plist"; + LD_RUNPATH_SEARCH_PATHS = ( + "$(inherited)", + "@executable_path/Frameworks", + ); + PRODUCT_BUNDLE_IDENTIFIER = "org.ggml.VisionOSLlamaTest"; + PRODUCT_NAME = "$(TARGET_NAME)"; + PROVISIONING_PROFILE_SPECIFIER = ""; + SUPPORTED_PLATFORMS = "xros xrsimulator"; + SWIFT_VERSION = 5.0; + TARGETED_DEVICE_FAMILY = "1,2,7"; + }; + name = Debug; + }; + CCCCCCCCCCCCCCCCCCCCCCDDD /* Release */ = { + isa = XCBuildConfiguration; + buildSettings = { + ASSETCATALOG_COMPILER_APPICON_NAME = AppIcon; + ASSETCATALOG_COMPILER_GLOBAL_ACCENT_COLOR_NAME = AccentColor; + CODE_SIGN_STYLE = Manual; + DEVELOPMENT_TEAM = ""; + ENABLE_PREVIEWS = YES; + FRAMEWORK_SEARCH_PATHS = ( + "$(inherited)", + "$(PROJECT_DIR)", + ); + INFOPLIST_FILE = "VisionOSLlamaTest/Info.plist"; + LD_RUNPATH_SEARCH_PATHS = ( + "$(inherited)", + "@executable_path/Frameworks", + ); + PRODUCT_BUNDLE_IDENTIFIER = "org.ggml.VisionOSLlamaTest"; + PRODUCT_NAME = "$(TARGET_NAME)"; + PROVISIONING_PROFILE_SPECIFIER = ""; + SUPPORTED_PLATFORMS = "xros xrsimulator"; + SWIFT_VERSION = 5.0; + TARGETED_DEVICE_FAMILY = "1,2,7"; + }; + name = Release; + }; +/* End XCBuildConfiguration section */ +EOF + +# Finish the project.pbxproj file +cat >> "${TEMP_DIR}/${APP_NAME}/${APP_NAME}.xcodeproj/project.pbxproj" << EOF +/* Begin XCConfigurationList section */ + 8888888888888888888888AA /* Build configuration list for PBXProject "${APP_NAME}" */ = { + isa = XCConfigurationList; + buildConfigurations = ( + 9999999999999999999999AA /* Debug */, + AAAAAAAAAAAAAAAAAAAAABBB /* Release */, + ); + defaultConfigurationIsVisible = 0; + defaultConfigurationName = Release; + }; + 4444444444444444444444AA /* Build configuration list for PBXNativeTarget "${APP_NAME}" */ = { + isa = XCConfigurationList; + buildConfigurations = ( + BBBBBBBBBBBBBBBBBBBBBBCCC /* Debug */, + CCCCCCCCCCCCCCCCCCCCCCDDD /* Release */, + ); + defaultConfigurationIsVisible = 0; + defaultConfigurationName = Release; + }; +/* End XCConfigurationList section */ + }; + rootObject = 7777777777777777777777AA /* Project object */; +} +EOF + +# 2. Copy XCFramework to test project +echo "Copying XCFramework to test project..." +cp -R "${XCFRAMEWORK_PATH}" "${TEMP_DIR}/${APP_NAME}/" + +# 3. Build and archive the app +echo "Building and archiving test app..." +cd "${TEMP_DIR}/${APP_NAME}" + +# Create a simple xcscheme file to avoid xcodebuild scheme issues +mkdir -p "${APP_NAME}.xcodeproj/xcshareddata/xcschemes" +cat > "${APP_NAME}.xcodeproj/xcshareddata/xcschemes/${APP_NAME}.xcscheme" << EOF + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + +EOF + +# Now use xcodebuild with an explicitly defined product name for visionOS +xcodebuild -project "${APP_NAME}.xcodeproj" -scheme "${APP_NAME}" -sdk xros -configuration Release archive -archivePath "${ARCHIVE_PATH}" CODE_SIGN_IDENTITY="-" CODE_SIGNING_REQUIRED=NO CODE_SIGNING_ALLOWED=NO PRODUCT_NAME="${APP_NAME}" SWIFT_OPTIMIZATION_LEVEL="-Onone" -quiet + +# 4. Create IPA from archive +echo "Creating IPA from archive..." +mkdir -p "${TEMP_DIR}/Payload" +cp -R "${ARCHIVE_PATH}/Products/Applications/${APP_NAME}.app" "${TEMP_DIR}/Payload/" + +# Check and log app structure before zipping +echo "App structure:" +ls -la "${TEMP_DIR}/Payload/${APP_NAME}.app/" +echo "Frameworks:" +ls -la "${TEMP_DIR}/Payload/${APP_NAME}.app/Frameworks/" 2>/dev/null || echo "No Frameworks directory found" + +cd "${TEMP_DIR}" +zip -r "${IPA_PATH}" Payload + +# Check embedded provisioning profile +echo "Checking provisioning profile (if any)..." +PROVISIONING_PROFILE=$(find "${ARCHIVE_PATH}/Products/Applications/${APP_NAME}.app" -name "embedded.mobileprovision" 2>/dev/null) +if [ -n "$PROVISIONING_PROFILE" ]; then + echo "Found embedded provisioning profile:" + security cms -D -i "$PROVISIONING_PROFILE" || echo "Unable to decode provisioning profile" +else + echo "No embedded provisioning profile found (expected for ad-hoc builds)" +fi + +# 5. Validate the IPA +echo "Validating IPA..." +VALIDATION_OUTPUT="${VALIDATION_DIR}/validation_output.txt" + +# Check if authentication credentials are provided +AUTH_ARGS="" +if [ -n "$APPLE_ID" ] && [ -n "$APPLE_PASSWORD" ]; then + echo "Using Apple ID authentication for validation..." + AUTH_ARGS="--username \"$APPLE_ID\" --password \"$APPLE_PASSWORD\"" +else + echo "No authentication credentials provided. Will perform basic validation." + echo "To use your personal developer account, you can run the script with:" + echo " APPLE_ID='your.email@example.com' APPLE_PASSWORD='your-app-specific-password' ./validate-visionos.sh" + echo "Note: You need to create an app-specific password at https://appleid.apple.com/account/manage" +fi + +# Run validation with detailed output +echo "Running validation with altool..." +if [ -n "$AUTH_ARGS" ]; then + # Use eval to properly handle the quoted arguments + eval "xcrun altool --validate-app -f \"${IPA_PATH}\" --type visionos --output-format xml $AUTH_ARGS" 2>&1 | tee "${VALIDATION_OUTPUT}" +else + xcrun altool --validate-app -f "${IPA_PATH}" --type visionos --output-format xml 2>&1 | tee "${VALIDATION_OUTPUT}" +fi +VALIDATION_RESULT=$? + +# Final validation result +FINAL_VALIDATION_RESULT=0 + +# Check if validation failed because the app isn't in App Store Connect +if grep -q "No suitable application records were found" "${VALIDATION_OUTPUT}"; then + echo "⚠️ App Store Connect Warning: The app bundle identifier is not found in App Store Connect" + echo "This is expected for apps that haven't been registered in App Store Connect yet." + echo "This doesn't indicate a problem with the build or framework." + + # Perform alternative validation + echo "Performing alternative validation checks..." + + # Check if IPA was created successfully + if [ -f "${IPA_PATH}" ] && [ -s "${IPA_PATH}" ]; then + echo "✅ IPA file created successfully" + else + echo "❌ IPA file not created or empty" + FINAL_VALIDATION_RESULT=1 + fi + + # Check if app binary exists and is executable + if [ -f "${TEMP_DIR}/Payload/${APP_NAME}.app/${APP_NAME}" ] && [ -x "${TEMP_DIR}/Payload/${APP_NAME}.app/${APP_NAME}" ]; then + echo "✅ App binary exists and is executable" + else + echo "❌ App binary missing or not executable" + FINAL_VALIDATION_RESULT=1 + fi + + # Check if framework was properly embedded + if [ -d "${TEMP_DIR}/Payload/${APP_NAME}.app/Frameworks/llama.framework" ]; then + echo "✅ llama.framework properly embedded" + else + echo "❌ llama.framework not properly embedded" + FINAL_VALIDATION_RESULT=1 + fi + + # Check if framework binary exists + if [ -f "${TEMP_DIR}/Payload/${APP_NAME}.app/Frameworks/llama.framework/llama" ]; then + echo "✅ Framework binary exists" + + # Further validate framework by checking architecture + ARCHS=$(lipo -info "${TEMP_DIR}/Payload/${APP_NAME}.app/Frameworks/llama.framework/llama" 2>/dev/null | grep -o "arm64\\|x86_64" | tr '\n' ' ') + if [ -n "$ARCHS" ]; then + echo "✅ Framework architecture(s): $ARCHS" + else + echo "⚠️ Could not determine framework architecture" + fi + else + echo "❌ Framework binary missing" + FINAL_VALIDATION_RESULT=1 + fi + + if [ $FINAL_VALIDATION_RESULT -eq 0 ]; then + echo "✅ Alternative validation PASSED: App built successfully with embedded framework" + else + echo "❌ Alternative validation FAILED: Issues found with the app or framework" + fi +elif grep -q "You must specify authentication credentials" "${VALIDATION_OUTPUT}" && [ -z "$AUTH_ARGS" ]; then + echo "✅ visionOS Validation PASSED: IPA successfully validated" + echo "Results saved to ${VALIDATION_OUTPUT}" +else + echo "❌ visionOS Validation FAILED: IPA validation found issues" + echo "See validation output at ${VALIDATION_OUTPUT}" + echo "" + echo "==== VALIDATION ERRORS ====" + + # Try to extract specific errors from the output + if grep -q "Error" "${VALIDATION_OUTPUT}"; then + grep -A 5 "Error" "${VALIDATION_OUTPUT}" + else + # If no specific error found, show the whole log + cat "${VALIDATION_OUTPUT}" + fi + + # Additional debugging: check IPA contents + echo "" + echo "==== IPA CONTENTS ====" + mkdir -p "${TEMP_DIR}/ipa_contents" + unzip -q "${IPA_PATH}" -d "${TEMP_DIR}/ipa_contents" + ls -la "${TEMP_DIR}/ipa_contents/Payload/${APP_NAME}.app/" + + # Check for code signing issues + echo "" + echo "==== CODE SIGNING INFO ====" + codesign -vv -d "${TEMP_DIR}/ipa_contents/Payload/${APP_NAME}.app" 2>&1 || echo "Code signing verification failed" + + # Check embedded frameworks + echo "" + echo "==== FRAMEWORK INFO ====" + ls -la "${TEMP_DIR}/ipa_contents/Payload/${APP_NAME}.app/Frameworks/" 2>/dev/null || echo "No Frameworks directory found" +fi + +# Don't clean up on error to allow inspection +if [ $FINAL_VALIDATION_RESULT -ne 0 ]; then + echo "" + echo "Temporary files kept for inspection at: ${TEMP_DIR}" + echo "===== visionOS Validation Process Failed =====" + exit 1 +fi + +# Clean up temporary files but keep build artifacts +if [ $FINAL_VALIDATION_RESULT -eq 0 ]; then + echo "Cleaning up temporary files..." + #rm -rf "${TEMP_DIR}" +fi + +echo "===== visionOS Validation Process Completed =====" +exit $FINAL_VALIDATION_RESULT diff --git a/scripts/check-requirements.sh b/scripts/check-requirements.sh index d3bbded13..4c3b05f68 100755 --- a/scripts/check-requirements.sh +++ b/scripts/check-requirements.sh @@ -170,7 +170,7 @@ check_convert_script examples/convert_legacy_llama.py for py in convert_*.py; do # skip convert_hf_to_gguf_update.py # TODO: the check is failing for some reason: - # https://github.com/ggerganov/llama.cpp/actions/runs/8875330981/job/24364557177?pr=6920 + # https://github.com/ggml-org/llama.cpp/actions/runs/8875330981/job/24364557177?pr=6920 [[ $py == convert_hf_to_gguf_update.py ]] && continue check_convert_script "$py" diff --git a/scripts/compare-llama-bench.py b/scripts/compare-llama-bench.py index 239c458d8..6205fe88d 100755 --- a/scripts/compare-llama-bench.py +++ b/scripts/compare-llama-bench.py @@ -124,9 +124,22 @@ if input_file is None: connection = sqlite3.connect(input_file) cursor = connection.cursor() -builds = cursor.execute("SELECT DISTINCT build_commit FROM test;").fetchall() -commit_short_len = len(builds[0][0]) +build_len_min: int = cursor.execute("SELECT MIN(LENGTH(build_commit)) from test;").fetchone()[0] +build_len_max: int = cursor.execute("SELECT MAX(LENGTH(build_commit)) from test;").fetchone()[0] + +if build_len_min != build_len_max: + logger.warning(f"{input_file} contains commit hashes of differing lengths. It's possible that the wrong commits will be compared. " + "Try purging the the database of old commits.") + cursor.execute(f"UPDATE test SET build_commit = SUBSTRING(build_commit, 1, {build_len_min});") + +build_len: int = build_len_min + +builds = cursor.execute("SELECT DISTINCT build_commit FROM test;").fetchall() +builds = list(map(lambda b: b[0], builds)) # list[tuple[str]] -> list[str] + +if not builds: + raise RuntimeError(f"{input_file} does not contain any builds.") try: repo = git.Repo(".", search_parent_directories=True) @@ -140,11 +153,11 @@ def find_parent_in_data(commit: git.Commit): seen_hexsha8 = set() while heap: depth, current_commit = heapq.heappop(heap) - current_hexsha8 = commit.hexsha[:commit_short_len] - if (current_hexsha8,) in builds: + current_hexsha8 = commit.hexsha[:build_len] + if current_hexsha8 in builds: return current_hexsha8 for parent in commit.parents: - parent_hexsha8 = parent.hexsha[:commit_short_len] + parent_hexsha8 = parent.hexsha[:build_len] if parent_hexsha8 not in seen_hexsha8: seen_hexsha8.add(parent_hexsha8) heapq.heappush(heap, (depth + 1, parent)) @@ -158,40 +171,40 @@ def get_all_parent_hexsha8s(commit: git.Commit): while unvisited: current_commit = unvisited.pop(0) - visited.append(current_commit.hexsha[:commit_short_len]) + visited.append(current_commit.hexsha[:build_len]) for parent in current_commit.parents: - if parent.hexsha[:commit_short_len] not in visited: + if parent.hexsha[:build_len] not in visited: unvisited.append(parent) return visited -def get_commit_name(hexsha8): +def get_commit_name(hexsha8: str): """Helper function to find a human-readable name for a commit if possible.""" if repo is None: return hexsha8 for h in repo.heads: - if h.commit.hexsha[:commit_short_len] == hexsha8: + if h.commit.hexsha[:build_len] == hexsha8: return h.name for t in repo.tags: - if t.commit.hexsha[:commit_short_len] == hexsha8: + if t.commit.hexsha[:build_len] == hexsha8: return t.name return hexsha8 -def get_commit_hexsha8(name): +def get_commit_hexsha8(name: str): """Helper function to search for a commit given a human-readable name.""" if repo is None: return None for h in repo.heads: if h.name == name: - return h.commit.hexsha[:commit_short_len] + return h.commit.hexsha[:build_len] for t in repo.tags: if t.name == name: - return t.commit.hexsha[:commit_short_len] + return t.commit.hexsha[:build_len] for c in repo.iter_commits("--all"): - if c.hexsha[:commit_short_len] == name[:commit_short_len]: - return c.hexsha[:commit_short_len] + if c.hexsha[:build_len] == name[:build_len]: + return c.hexsha[:build_len] return None @@ -199,7 +212,7 @@ hexsha8_baseline = name_baseline = None # If the user specified a baseline, try to find a commit for it: if known_args.baseline is not None: - if (known_args.baseline,) in builds: + if known_args.baseline in builds: hexsha8_baseline = known_args.baseline if hexsha8_baseline is None: hexsha8_baseline = get_commit_hexsha8(known_args.baseline) @@ -228,7 +241,7 @@ hexsha8_compare = name_compare = None # If the user has specified a compare value, try to find a corresponding commit: if known_args.compare is not None: - if (known_args.compare,) in builds: + if known_args.compare in builds: hexsha8_compare = known_args.compare if hexsha8_compare is None: hexsha8_compare = get_commit_hexsha8(known_args.compare) diff --git a/scripts/fetch_server_test_models.py b/scripts/fetch_server_test_models.py index 05690b138..e6775bfc5 100755 --- a/scripts/fetch_server_test_models.py +++ b/scripts/fetch_server_test_models.py @@ -75,7 +75,7 @@ if __name__ == '__main__': logging.info(f' - {m.hf_repo} / {m.hf_file}') cli_path = os.environ.get( - 'LLAMA_SERVER_BIN_PATH', + 'LLAMA_CLI_BIN_PATH', os.path.join( os.path.dirname(__file__), '../build/bin/Release/llama-cli.exe' if os.name == 'nt' else '../build/bin/llama-cli')) diff --git a/scripts/get_chat_template.py b/scripts/get_chat_template.py old mode 100644 new mode 100755 index e8982d11a..b4827b317 --- a/scripts/get_chat_template.py +++ b/scripts/get_chat_template.py @@ -7,9 +7,8 @@ ./scripts/get_chat_template.py model_id [variant] Examples: - ./scripts/get_chat_template.py NousResearch/Meta-Llama-3-8B-Instruct - ./scripts/get_chat_template.py NousResearch/Hermes-3-Llama-3.1-8B tool_use - ./scripts/get_chat_template.py meta-llama/Llama-3.2-3B-Instruct + ./scripts/get_chat_template.py CohereForAI/c4ai-command-r-plus tool_use + ./scripts/get_chat_template.py microsoft/Phi-3.5-mini-instruct ''' import json @@ -22,7 +21,7 @@ def get_chat_template(model_id, variant=None): # Use huggingface_hub library if available. # Allows access to gated models if the user has access and ran `huggingface-cli login`. from huggingface_hub import hf_hub_download - with open(hf_hub_download(repo_id=model_id, filename="tokenizer_config.json")) as f: + with open(hf_hub_download(repo_id=model_id, filename="tokenizer_config.json"), encoding="utf-8") as f: config_str = f.read() except ImportError: import requests diff --git a/scripts/sync-ggml-am.sh b/scripts/sync-ggml-am.sh index ec4f4b0a2..204354209 100755 --- a/scripts/sync-ggml-am.sh +++ b/scripts/sync-ggml-am.sh @@ -69,7 +69,11 @@ while read c; do git format-patch -U${ctx} -k $c~1..$c --stdout -- \ CMakeLists.txt \ src/CMakeLists.txt \ - cmake/FindSIMD.cmake \ + cmake/BuildTypes.cmake \ + cmake/GitVars.cmake \ + cmake/common.cmake \ + cmake/ggml-config.cmake.in \ + src/ggml-cpu/cmake/FindSIMD.cmake \ src/ggml*.h \ src/ggml*.c \ src/ggml*.cpp \ @@ -121,7 +125,12 @@ if [ -f $SRC_LLAMA/ggml-src.patch ]; then # # CMakelists.txt -> ggml/CMakeLists.txt # src/CMakeLists.txt -> ggml/src/CMakeLists.txt - # cmake/FindSIMD.cmake -> ggml/cmake/FindSIMD.cmake + + # cmake/BuildTypes.cmake -> ggml/cmake/BuildTypes.cmake + # cmake/GitVars.cmake -> ggml/cmake/GitVars.cmake + # cmake/common.cmake -> ggml/cmake/common.cmake + # cmake/ggml-config.cmake.in -> ggml/cmake/ggml-config.cmake.in + # src/ggml-cpu/cmake/FindSIMD.cmake -> ggml/src/ggml-cpu/cmake/FindSIMD.cmake # # src/ggml*.c -> ggml/src/ggml*.c # src/ggml*.cpp -> ggml/src/ggml*.cpp @@ -149,29 +158,33 @@ if [ -f $SRC_LLAMA/ggml-src.patch ]; then # scripts/gen-authors.sh -> scripts/gen-authors.sh cat ggml-src.patch | sed -E \ - -e 's/([[:space:]]|[ab]\/)CMakeLists.txt/\1ggml\/CMakeLists.txt/g' \ - -e 's/([[:space:]]|[ab]\/)src\/CMakeLists.txt/\1ggml\/src\/CMakeLists.txt/g' \ - -e 's/([[:space:]]|[ab]\/)cmake\/FindSIMD.cmake/\1ggml\/cmake\/FindSIMD.cmake/g' \ - -e 's/([[:space:]]|[ab]\/)src\/ggml(.*)\.c/\1ggml\/src\/ggml\2.c/g' \ - -e 's/([[:space:]]|[ab]\/)src\/ggml(.*)\.cpp/\1ggml\/src\/ggml\2.cpp/g' \ - -e 's/([[:space:]]|[ab]\/)src\/ggml(.*)\.h/\1ggml\/src\/ggml\2.h/g' \ - -e 's/([[:space:]]|[ab]\/)src\/gguf(.*)\.cpp/\1ggml\/src\/gguf\2.cpp/g' \ - -e 's/([[:space:]]|[ab]\/)src\/ggml-blas\//\1ggml\/src\/ggml-blas\//g' \ - -e 's/([[:space:]]|[ab]\/)src\/ggml-cann\//\1ggml\/src\/ggml-cann\//g' \ - -e 's/([[:space:]]|[ab]\/)src\/ggml-cpu\//\1ggml\/src\/ggml-cpu\//g' \ - -e 's/([[:space:]]|[ab]\/)src\/ggml-cuda\//\1ggml\/src\/ggml-cuda\//g' \ - -e 's/([[:space:]]|[ab]\/)src\/ggml-hip\//\1ggml\/src\/ggml-hip\//g' \ - -e 's/([[:space:]]|[ab]\/)src\/ggml-kompute\//\1ggml\/src\/ggml-kompute\//g' \ - -e 's/([[:space:]]|[ab]\/)src\/ggml-metal\//\1ggml\/src\/ggml-metal\//g' \ - -e 's/([[:space:]]|[ab]\/)src\/ggml-opencl\//\1ggml\/src\/ggml-opencl\//g' \ - -e 's/([[:space:]]|[ab]\/)src\/ggml-rpc\//\1ggml\/src\/ggml-rpc\//g' \ - -e 's/([[:space:]]|[ab]\/)src\/ggml-sycl\//\1ggml\/src\/ggml-sycl\//g' \ - -e 's/([[:space:]]|[ab]\/)src\/ggml-vulkan\//\1ggml\/src\/ggml-vulkan\//g' \ - -e 's/([[:space:]]|[ab]\/)include\/ggml(.*)\.h/\1ggml\/include\/ggml\2.h/g' \ - -e 's/([[:space:]]|[ab]\/)include\/gguf(.*)\.h/\1ggml\/include\/gguf\2.h/g' \ - -e 's/([[:space:]]|[ab]\/)tests\/(.*)\.cpp/\1tests\/\2.cpp/g' \ - -e 's/([[:space:]]|[ab]\/)LICENSE/\1LICENSE/g' \ - -e 's/([[:space:]]|[ab]\/)scripts\/gen-authors\.sh/\1scripts\/gen-authors.sh/g' \ + -e 's/([[:space:]]| [ab]\/)CMakeLists.txt/\1ggml\/CMakeLists.txt/g' \ + -e 's/([[:space:]]| [ab]\/)src\/CMakeLists.txt/\1ggml\/src\/CMakeLists.txt/g' \ + -e 's/([[:space:]]| [ab]\/)cmake\/BuildTypes.cmake/\1ggml\/cmake\/BuildTypes.cmake/g' \ + -e 's/([[:space:]]| [ab]\/)cmake\/GitVars.cmake/\1ggml\/cmake\/GitVars.cmake/g' \ + -e 's/([[:space:]]| [ab]\/)cmake\/common.cmake/\1ggml\/cmake\/common.cmake/g' \ + -e 's/([[:space:]]| [ab]\/)cmake\/ggml-config.cmake.in/\1ggml\/cmake\/ggml-config.cmake.in/g' \ + -e 's/([[:space:]]| [ab]\/)src\/ggml-cpu\/cmake\/FindSIMD.cmake/\1ggml\/src\/ggml-cpu\/cmake\/FindSIMD.cmake/g' \ + -e 's/([[:space:]]| [ab]\/)src\/ggml(.*)\.c/\1ggml\/src\/ggml\2.c/g' \ + -e 's/([[:space:]]| [ab]\/)src\/ggml(.*)\.cpp/\1ggml\/src\/ggml\2.cpp/g' \ + -e 's/([[:space:]]| [ab]\/)src\/ggml(.*)\.h/\1ggml\/src\/ggml\2.h/g' \ + -e 's/([[:space:]]| [ab]\/)src\/gguf(.*)\.cpp/\1ggml\/src\/gguf\2.cpp/g' \ + -e 's/([[:space:]]| [ab]\/)src\/ggml-blas\//\1ggml\/src\/ggml-blas\//g' \ + -e 's/([[:space:]]| [ab]\/)src\/ggml-cann\//\1ggml\/src\/ggml-cann\//g' \ + -e 's/([[:space:]]| [ab]\/)src\/ggml-cpu\//\1ggml\/src\/ggml-cpu\//g' \ + -e 's/([[:space:]]| [ab]\/)src\/ggml-cuda\//\1ggml\/src\/ggml-cuda\//g' \ + -e 's/([[:space:]]| [ab]\/)src\/ggml-hip\//\1ggml\/src\/ggml-hip\//g' \ + -e 's/([[:space:]]| [ab]\/)src\/ggml-kompute\//\1ggml\/src\/ggml-kompute\//g' \ + -e 's/([[:space:]]| [ab]\/)src\/ggml-metal\//\1ggml\/src\/ggml-metal\//g' \ + -e 's/([[:space:]]| [ab]\/)src\/ggml-opencl\//\1ggml\/src\/ggml-opencl\//g' \ + -e 's/([[:space:]]| [ab]\/)src\/ggml-rpc\//\1ggml\/src\/ggml-rpc\//g' \ + -e 's/([[:space:]]| [ab]\/)src\/ggml-sycl\//\1ggml\/src\/ggml-sycl\//g' \ + -e 's/([[:space:]]| [ab]\/)src\/ggml-vulkan\//\1ggml\/src\/ggml-vulkan\//g' \ + -e 's/([[:space:]]| [ab]\/)include\/ggml(.*)\.h/\1ggml\/include\/ggml\2.h/g' \ + -e 's/([[:space:]]| [ab]\/)include\/gguf(.*)\.h/\1ggml\/include\/gguf\2.h/g' \ + -e 's/([[:space:]]| [ab]\/)tests\/(.*)\.cpp/\1tests\/\2.cpp/g' \ + -e 's/([[:space:]]| [ab]\/)LICENSE/\1LICENSE/g' \ + -e 's/([[:space:]]| [ab]\/)scripts\/gen-authors\.sh/\1scripts\/gen-authors.sh/g' \ > ggml-src.patch.tmp mv ggml-src.patch.tmp ggml-src.patch diff --git a/scripts/sync-ggml.last b/scripts/sync-ggml.last index 255109ab7..7111936ba 100644 --- a/scripts/sync-ggml.last +++ b/scripts/sync-ggml.last @@ -1 +1 @@ -08b538031f7f944e84f472483ef5d26bf5190ead +2abf606f098844faebee578996cae9c6d63a40e2 diff --git a/scripts/sync-ggml.sh b/scripts/sync-ggml.sh index e83d415c0..aa1a46b4b 100755 --- a/scripts/sync-ggml.sh +++ b/scripts/sync-ggml.sh @@ -2,7 +2,9 @@ cp -rpv ../ggml/CMakeLists.txt ./ggml/CMakeLists.txt cp -rpv ../ggml/src/CMakeLists.txt ./ggml/src/CMakeLists.txt -cp -rpv ../ggml/cmake/FindSIMD.cmake ./ggml/cmake/FindSIMD.cmake + +cp -rpv ../ggml/cmake/* ./ggml/cmake/ +cp -rpv ../ggml/src/ggml-cpu/cmake/* ./ggml/src/ggml-cpu/cmake/ cp -rpv ../ggml/src/ggml*.c ./ggml/src/ cp -rpv ../ggml/src/ggml*.cpp ./ggml/src/ diff --git a/scripts/tool_bench.py b/scripts/tool_bench.py new file mode 100755 index 000000000..0f406bc42 --- /dev/null +++ b/scripts/tool_bench.py @@ -0,0 +1,368 @@ +#!/usr/bin/env uv run +''' + Simplistic tool call benchmarks for llama-server and ollama. + + Essentially runs the tests at server/examples/server/tests/unit/test_tool_call.py N times, at different temperatures and on different backends (current llama-server, baseline llama-server and ollama), + and plots the results of multiple runs (from same .jsonl file or multiple ones) as a success rate heatmap. + + Simple usage example: + + cmake -B build -DLLAMA_CURL=1 && cmake --build build --config Release -j -t llama-server + + export LLAMA_SERVER_BIN_PATH=$PWD/build/bin/llama-server + export LLAMA_CACHE=${LLAMA_CACHE:-$HOME/Library/Caches/llama.cpp} + + ./scripts/tool_bench.py run --n 30 --temp -1 --temp 0 --temp 1 --model "Qwen 2.5 1.5B Q4_K_M" --output qwen1.5b.jsonl --hf bartowski/Qwen2.5-1.5B-Instruct-GGUF --ollama qwen2.5:1.5b-instruct-q4_K_M + ./scripts/tool_bench.py run --n 30 --temp -1 --temp 0 --temp 1 --model "Qwen 2.5 Coder 7B Q4_K_M" --output qwenc7b.jsonl --hf bartowski/Qwen2.5-Coder-7B-Instruct-GGUF --ollama qwen2.5-coder:7b + + ./scripts/tool_bench.py plot *.jsonl # Opens window w/ heatmap + ./scripts/tool_bench.py plot qwen*.jsonl --output qwen.png # Saves heatmap to qwen.png + + (please see ./scripts/tool_bench.sh for a more complete example) +''' +# /// script +# requires-python = ">=3.10" +# dependencies = [ +# "pytest", +# "pandas", +# "matplotlib", +# "seaborn", +# "requests", +# "wget", +# "typer", +# ] +# /// +from contextlib import contextmanager +from pathlib import Path +import re +from statistics import mean, median +from typing import Annotated, Dict, List, Optional, Tuple +import atexit +import json +import logging +import matplotlib.pyplot as plt +import numpy as np +import pandas as pd +import seaborn as sns +import subprocess +import sys +import time +import typer + +sys.path.insert(0, Path(__file__).parent.parent.as_posix()) +if True: + from examples.server.tests.utils import ServerProcess + from examples.server.tests.unit.test_tool_call import TIMEOUT_SERVER_START, do_test_calc_result, do_test_hello_world, do_test_weather + + +@contextmanager +def scoped_server(sp: ServerProcess): + def stop(): + nonlocal sp + if sp is not None: + sp.stop() + sp = None # type: ignore + atexit.register(stop) + yield sp + stop() + + +logging.basicConfig( + level=logging.INFO, + format='%(asctime)s - %(levelname)s - %(message)s' +) +logger = logging.getLogger(__name__) + +app = typer.Typer() + + +@app.command() +def plot(files: List[Path], output: Optional[Path] = None, test_regex: Optional[str] = None, server_regex: Optional[str] = None): + + lines: List[Dict] = [] + for file in files: + if not file.exists(): + logger.error(f"File not found: {file}") + continue + + try: + with file.open() as f: + raw_data = f.read() + logger.info(f"Reading {file} ({len(raw_data)} bytes)") + + for line_num, line in enumerate(raw_data.split('\n'), 1): + line = line.strip() + if not line: + continue + try: + record = json.loads(line) + lines.append(record) + except json.JSONDecodeError as e: + logger.warning(f"Invalid JSON at {file}:{line_num} - {e}") + except Exception as e: + logger.error(f"Error processing {file}: {e}") + + if not lines: + raise Exception("No valid data was loaded") + + data_dict: Dict[Tuple, float] = {} + models: List[str] = [] + temps = set() + tests = set() + server_names = set() + total_counts = set() + for rec in lines: + try: + model = rec["model"] + temp = rec["temp"] + server_name = rec["server_name"] + test = rec["test"] + success = rec["success_ratio"] + success_count = rec["success_count"] + failure_count = rec["failure_count"] + total_count = success_count + failure_count + total_counts.add(total_count) + + if test_regex and not re.search(test_regex, test): + continue + + if server_regex and not re.search(server_regex, server_name): + continue + + data_dict[(model, temp, server_name, test)] = success + + if model not in models: + models.append(model) + temps.add(temp) + tests.add(test) + server_names.add(server_name) + + except KeyError as e: + logger.warning(f"Missing required field in record: {e}") + + if len(total_counts) > 1: + logger.warning(f"Total counts are not consistent: {total_counts}") + + # Sort the collected values + temps = list(sorted(temps, key=lambda x: x if x is not None else -1)) + tests = list(sorted(tests)) + server_names = list(sorted(server_names)) + + logger.info(f"Processed {len(lines)} lines") + logger.info(f"Found {len(data_dict)} valid data points") + logger.info(f"Models: {models}") + logger.info(f"Temperatures: {temps}") + logger.info(f"Tests: {tests}") + logger.info(f"Servers: {server_names}") + + matrix: list[list[float]] = [] + index: list[str] = [] + + all_cols = [ + (server_name, test) + for server_name in server_names + for test in tests + ] + for model in models: + for temp in temps: + index.append(f"{model} @ {temp}") + row_vals = [ + data_dict.get((model, temp, server_name, test), np.nan) + for server_name, test in all_cols + ] + matrix.append(row_vals) + + columns: list[str] = [f"{server_name}\n{test}" for server_name, test in all_cols] + + df = pd.DataFrame(matrix, index=np.array(index), columns=np.array(columns)) + + plt.figure(figsize=(12, 6)) + + sns.heatmap( + df, annot=True, cmap="RdYlGn", vmin=0.0, vmax=1.0, cbar=True, fmt=".2f", center=0.5, square=True, linewidths=0.5, + cbar_kws={"label": "Success Ratio"}, + ) + + plt.title(f"Tool Call Bench (n = {str(min(total_counts)) if len(total_counts) == 1 else f'{min(total_counts)}-{max(total_counts)}'})\nSuccess Ratios by Server & Test", pad=20) + plt.xlabel("Server & Test", labelpad=10) + plt.ylabel("Model @ Temperature", labelpad=10) + + plt.xticks(rotation=45, ha='right') + plt.yticks(rotation=0) + + plt.tight_layout() + + if output: + plt.savefig(output, dpi=300, bbox_inches='tight') + logger.info(f"Plot saved to {output}") + else: + plt.show() + + +@app.command() +def run( + output: Annotated[Path, typer.Option(help="Output JSON file")], + model: Annotated[Optional[str], typer.Option(help="Name of the model to test (server agnostic)")] = None, + hf: Annotated[Optional[str], typer.Option(help="GGUF huggingface model repo id (+ optional quant) to test w/ llama-server")] = None, + chat_template: Annotated[Optional[str], typer.Option(help="Chat template override for llama-server")] = None, + ollama: Annotated[Optional[str], typer.Option(help="Ollama model tag to test")] = None, + llama_baseline: Annotated[Optional[str], typer.Option(help="llama-server baseline binary path to use as baseline")] = None, + n: Annotated[int, typer.Option(help="Number of times to run each test")] = 10, + temp: Annotated[Optional[List[float]], typer.Option(help="Set of temperatures to test")] = None, + top_p: Annotated[Optional[float], typer.Option(help="top_p")] = None, + top_k: Annotated[Optional[int], typer.Option(help="top_k")] = None, + ctk: Annotated[Optional[str], typer.Option(help="ctk")] = None, + ctv: Annotated[Optional[str], typer.Option(help="ctv")] = None, + fa: Annotated[Optional[bool], typer.Option(help="fa")] = None, + seed: Annotated[Optional[int], typer.Option(help="Random seed")] = None, + port: Annotated[int, typer.Option(help="llama-server port")] = 8084, + force: Annotated[bool, typer.Option(help="Force overwrite of output file")] = False, + append: Annotated[bool, typer.Option(help="Append to output file")] = False, + + test_hello_world: Annotated[bool, typer.Option(help="Whether to run the hello world test")] = True, + test_weather: Annotated[bool, typer.Option(help="Whether to run the weather test")] = True, + test_calc_result: Annotated[bool, typer.Option(help="Whether to run the calc result test")] = False, +): + # Check only one of output and append + + n_predict = 512 # High because of DeepSeek R1 + # n_ctx = 8192 + n_ctx = 2048 + + assert force or append or not output.exists(), f"Output file already exists: {output}; use --force to overwrite" + + with output.open('a' if append else 'w') as output_file: + + def run(server: ServerProcess, *, server_name: str, model_id: str, temp: Optional[float] = None, output_kwargs={}, request_kwargs={}): + request_kwargs = {**request_kwargs} + if temp is not None: + request_kwargs['temperature'] = temp + if top_p is not None: + request_kwargs['top_p'] = top_p + if top_k is not None: + request_kwargs['top_k'] = top_k + if seed is not None: + request_kwargs['seed'] = seed + + request_kwargs['cache_prompt'] = False + + tests = {} + if test_hello_world: + tests["hello world"] = lambda server: do_test_hello_world(server, **request_kwargs) + if test_weather: + tests["weather"] = lambda server: do_test_weather(server, **request_kwargs) + if test_calc_result: + tests["calc result"] = lambda server: do_test_calc_result(server, None, 512, **request_kwargs) + + for test_name, test in tests.items(): + success_count = 0 + failure_count = 0 + failures = [] + success_times = [] + failure_times = [] + logger.info(f"Running {test_name} ({server_name}, {model}): ") + for i in range(n): + start_time = time.time() + + def elapsed(): + return time.time() - start_time + + try: + test(server) + success_times.append(elapsed()) + success_count += 1 + logger.info('success') + except Exception as e: + logger.error(f'failure: {e}') + failure_count += 1 + failure_times.append(elapsed()) + failures.append(str(e)) + # import traceback + # traceback.print_exc() + output_file.write(json.dumps({**output_kwargs, **dict( + model=model, + server_name=server_name, + model_id=model_id, + test=test_name, + temp=t, + top_p=top_p, + top_k=top_k, + ctk=ctk, + ctv=ctv, + seed=seed, + success_ratio=float(success_count) / n, + avg_time=mean(success_times + failure_times), + median_time=median(success_times + failure_times), + success_count=success_count, + success_times=success_times, + failure_count=failure_count, + failure_times=failure_times, + failures=list(set(failures)), + )}) + '\n') + output_file.flush() + + for t in [None] if temp is None else [t if t >= 0 else None for t in temp]: + if hf is not None: + + servers: list[Tuple[str, Optional[str]]] = [('llama-server', None)] + if llama_baseline is not None: + servers.append(('llama-server (baseline)', llama_baseline)) + + for server_name, server_path in servers: + server = ServerProcess() + server.n_ctx = n_ctx + server.n_slots = 1 + server.jinja = True + server.ctk = ctk + server.ctv = ctv + server.fa = fa + server.n_predict = n_predict + server.model_hf_repo = hf + server.model_hf_file = None + server.chat_template = chat_template + server.server_path = server_path + if port is not None: + server.server_port = port + # server.debug = True + + with scoped_server(server): + server.start(timeout_seconds=TIMEOUT_SERVER_START) + for ignore_chat_grammar in [False]: + run( + server, + server_name=server_name, + model_id=hf, + temp=t, + output_kwargs=dict( + chat_template=chat_template, + ), + request_kwargs=dict( + ignore_chat_grammar=ignore_chat_grammar, + ), + ) + + if ollama is not None: + server = ServerProcess() + server.server_port = 11434 + server.server_host = "localhost" + subprocess.check_call(["ollama", "pull", ollama]) + + with scoped_server(server): + run( + server, + server_name="ollama", + model_id=ollama, + temp=t, + output_kwargs=dict( + chat_template=None, + ), + request_kwargs=dict( + model=ollama, + max_tokens=n_predict, + num_ctx = n_ctx, + ), + ) + + +if __name__ == "__main__": + app() diff --git a/scripts/tool_bench.sh b/scripts/tool_bench.sh new file mode 100755 index 000000000..6c7616a88 --- /dev/null +++ b/scripts/tool_bench.sh @@ -0,0 +1,66 @@ +#!/bin/bash +set -euo pipefail + +cmake --build build -j + +export LLAMA_CACHE=${LLAMA_CACHE:-$HOME/Library/Caches/llama.cpp} +export LLAMA_SERVER_BIN_PATH=$PWD/build/bin/llama-server + +if [ ! -x "$LLAMA_SERVER_BIN_PATH" ]; then + echo "Could not find llama-server binary at $LLAMA_SERVER_BIN_PATH" + exit 1 +fi +if [ ! -d "$LLAMA_CACHE" ]; then + echo "Could not find llama cache at $LLAMA_CACHE, please set LLAMA_CACHE explicitly." + exit 1 +fi + +export ARGS=( + --llama-baseline="$(which llama-server)" + --n 30 + --temp -1 # Leaves temperature parameter unset (use the server's default, e.g. 0.6 for ollama) + --temp 0 + --temp 0.5 + --temp 0.75 + --temp 1 + --temp 1.5 + --temp 2 + --temp 5 + "$@" +) + +./scripts/tool_bench.py run ${ARGS[@]} --model "Qwen 2.5 Coder 0.5B Q4_K_M" --output ../qwenc0.5b.jsonl --hf bartowski/Qwen2.5-Coder-0.5B-Instruct-GGUF:Q4_K_M --ollama qwen2.5-coder:0.5b-instruct-q4_K_M +./scripts/tool_bench.py run ${ARGS[@]} --model "Qwen 2.5 Coder 1.5B Q4_K_M" --output ../qwenc1.5b.jsonl --hf bartowski/Qwen2.5-Coder-1.5B-Instruct-GGUF:Q4_K_M --ollama qwen2.5-coder:1.5b-instruct-q4_K_M +./scripts/tool_bench.py run ${ARGS[@]} --model "Qwen 2.5 Coder 3B Q4_K_M" --output ../qwenc3b.jsonl --hf bartowski/Qwen2.5-Coder-3B-Instruct-GGUF:Q4_K_M --ollama qwen2.5-coder:3b-instruct-q4_K_M +./scripts/tool_bench.py run ${ARGS[@]} --model "Qwen 2.5 Coder 7B Q4_K_M" --output ../qwenc7b.jsonl --hf bartowski/Qwen2.5-Coder-7B-Instruct-GGUF:Q4_K_M --ollama qwen2.5-coder:7b-instruct-q4_K_M +./scripts/tool_bench.py run ${ARGS[@]} --model "Qwen 2.5 Coder 32B Q4_K_M" --output ../qwenc32b.jsonl --hf bartowski/Qwen2.5-Coder-32B-Instruct-GGUF:Q4_K_M --ollama qwen2.5-coder:32B-instruct-q4_K_M +./scripts/tool_bench.py run ${ARGS[@]} --model "Qwen 2.5 1.5B Q4_K_M" --output ../qwen1.5b.jsonl --hf bartowski/Qwen2.5-1.5B-Instruct-GGUF:Q4_K_M --ollama qwen2.5:1.5b-instruct-q4_K_M +./scripts/tool_bench.py run ${ARGS[@]} --model "Qwen 2.5 3B Q4_K_M" --output ../qwen3b.jsonl --hf bartowski/Qwen2.5-3B-Instruct-GGUF:Q4_K_M --ollama qwen2.5:3b-instruct-q4_K_M +./scripts/tool_bench.py run ${ARGS[@]} --model "Qwen 2.5 7B Q4_K_M" --output ../qwen7b.jsonl --hf bartowski/Qwen2.5-7B-Instruct-GGUF:Q4_K_M --ollama qwen2.5:7b-instruct-q4_K_M + +./scripts/tool_bench.py run ${ARGS[@]} --model "Llama 3.2 Instruct 1B Q4_K_M" --output ../llama1b.jsonl --hf bartowski/Llama-3.2-1B-Instruct-GGUF:Q4_K_M --ollama llama3.2:1b-instruct-q4_K_M +./scripts/tool_bench.py run ${ARGS[@]} --model "Llama 3.2 Instruct 3B Q4_K_M" --output ../llama3b.jsonl --hf bartowski/Llama-3.2-3B-Instruct-GGUF:Q4_K_M --ollama llama3.2:3b-instruct-q4_K_M +./scripts/tool_bench.py run ${ARGS[@]} --model "Llama 3.1 Instruct 8B Q4_K_M" --output ../llama8b.jsonl --hf bartowski/Meta-Llama-3.1-8B-Instruct-GGUF:Q4_K_M --ollama llama3.1:8b-instruct-q4_K_M +./scripts/tool_bench.py run ${ARGS[@]} --model "Llama 3.3 70B Q4_K_M" --output ../llama70b.jsonl --hf bartowski/Llama-3.3-70B-Instruct-GGUF:Q4_K_M + +./scripts/tool_bench.py run ${ARGS[@]} --model "Mistral Nemo Q4_K_M" --output ../nemo.jsonl --hf bartowski/Mistral-Nemo-Instruct-2407-GGUF:Q4_K_M --ollama mistral-nemo:12b-instruct-2407-q4_K_M + +./scripts/tool_bench.py run ${ARGS[@]} --model "Hermes 3 Llama 3.1 8B Q4_K_M" --output ../hermes3.jsonl --hf bartowski/Hermes-3-Llama-3.1-8B-GGUF:Q4_K_M --ollama hermes3:8b-llama3.1-q4_K_M --chat-template-file <( python scripts/get_chat_template.py NousResearch/Hermes-3-Llama-3.1-8B tool_use ) +./scripts/tool_bench.py run ${ARGS[@]} --model "Hermes 2 Pro Llama 3 8B Q4_K_M" --output ../hermes2.jsonl --hf bartowski/Hermes-2-Pro-Llama-3-8B-GGUF:Q4_K_M --ollama hermes2:8b-llama3-q4_K_M --chat-template-file <( python scripts/get_chat_template.py NousResearch/Hermes-2-Pro-Llama-3-8B tool_use ) + +./scripts/tool_bench.py run ${ARGS[@]} --model "Functionary Small V3.2 Q4_K_M" --output ../funct3.2.jsonl --hf bartowski/functionary-small-v3.2-GGUF:Q4_K_M +./scripts/tool_bench.py run ${ARGS[@]} --model "FireFunction V2 IQ1_M" --output ../firef2.jsonl --hf bartowski/firefunction-v2-GGUF:IQ1_M --chat-template-file <( python scripts/get_chat_template.py fireworks-ai/llama-3-firefunction-v2 tool_use ) + +./scripts/tool_bench.py run ${ARGS[@]} --model "Command R7B 12-2024 Q6_K_L" --output ../c4ai.jsonl --hf bartowski/c4ai-command-r7b-12-2024-GGUF:Q6_K_L --chat-template-file <( python scripts/get_chat_template.py CohereForAI/c4ai-command-r7b-12-2024 tool_use ) + +./scripts/tool_bench.py run ${ARGS[@]} --model "Gemma 2 2B Q8_0" --output ../gemma2.jsonl --hf bartowski/gemma-2-2b-it-GGUF:Q8_0 +./scripts/tool_bench.py run ${ARGS[@]} --model "Phi 4 Instruct Q4_K_M" --output ../phi4.jsonl --hf bartowski/phi-4-GGUF:Q4_K_M # --ollama phi4 +./scripts/tool_bench.py run ${ARGS[@]} --model "Phi 3.5 Mini Instruct Q4_K_M" --output ../phi3.5.jsonl --hf bartowski/Phi-3.5-mini-instruct-GGUF:Q4_K_M # --ollama phi3.5:3.8b-mini-instruct-q4_K_M + +# ./scripts/tool_bench.py run ${ARGS[@]} --model "DeepSeek R1 Distill Qwen 7B Q6_K_L" --output ../dsqw7.jsonl --hf bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q6_K_L --chat-template-file <( python scripts/get_chat_template.py NousResearch/DeepSeek-R1-Distill-Qwen-7B tool_use ) +# ./scripts/tool_bench.py run ${ARGS[@]} --model "DeepSeek R1 Distill Qwen 32B Q4_K_M" --output ../dsqw32.jsonl --hf bartowski/DeepSeek-R1-Distill-Qwen-32B-GGUF:Q4_K_M --chat-template-file <( python scripts/get_chat_template.py NousResearch/DeepSeek-R1-Distill-Qwen-32B tool_use ) + + +for f in ../*.jsonl; do + ./scripts/tool_bench.py plot "$f" --output ${f%.jsonl}.png || true +done diff --git a/spm-headers/ggml-alloc.h b/spm-headers/ggml-alloc.h deleted file mode 120000 index 0361ffc38..000000000 --- a/spm-headers/ggml-alloc.h +++ /dev/null @@ -1 +0,0 @@ -../ggml/include/ggml-alloc.h \ No newline at end of file diff --git a/spm-headers/ggml-backend.h b/spm-headers/ggml-backend.h deleted file mode 120000 index 7295f0f0d..000000000 --- a/spm-headers/ggml-backend.h +++ /dev/null @@ -1 +0,0 @@ -../ggml/include/ggml-backend.h \ No newline at end of file diff --git a/spm-headers/ggml-cpp.h b/spm-headers/ggml-cpp.h deleted file mode 120000 index 8a8604cc2..000000000 --- a/spm-headers/ggml-cpp.h +++ /dev/null @@ -1 +0,0 @@ -../ggml/include/ggml-cpp.h \ No newline at end of file diff --git a/spm-headers/ggml-cpu.h b/spm-headers/ggml-cpu.h deleted file mode 120000 index 66e629607..000000000 --- a/spm-headers/ggml-cpu.h +++ /dev/null @@ -1 +0,0 @@ -../ggml/include/ggml-cpu.h \ No newline at end of file diff --git a/spm-headers/ggml-metal.h b/spm-headers/ggml-metal.h deleted file mode 120000 index aefad5fa0..000000000 --- a/spm-headers/ggml-metal.h +++ /dev/null @@ -1 +0,0 @@ -../ggml/include/ggml-metal.h \ No newline at end of file diff --git a/spm-headers/ggml.h b/spm-headers/ggml.h deleted file mode 120000 index 0bdfeacbd..000000000 --- a/spm-headers/ggml.h +++ /dev/null @@ -1 +0,0 @@ -../ggml/include/ggml.h \ No newline at end of file diff --git a/spm-headers/llama.h b/spm-headers/llama.h deleted file mode 120000 index b31388f0d..000000000 --- a/spm-headers/llama.h +++ /dev/null @@ -1 +0,0 @@ -../include/llama.h \ No newline at end of file diff --git a/src/CMakeLists.txt b/src/CMakeLists.txt index e1b02e4c0..9f7ab13f1 100644 --- a/src/CMakeLists.txt +++ b/src/CMakeLists.txt @@ -15,21 +15,24 @@ add_library(llama llama-chat.cpp llama-context.cpp llama-grammar.cpp + llama-graph.cpp llama-hparams.cpp llama-impl.cpp + llama-io.cpp llama-kv-cache.cpp + llama-memory.cpp llama-mmap.cpp llama-model-loader.cpp llama-model.cpp llama-quant.cpp llama-sampling.cpp llama-vocab.cpp - unicode.h - unicode.cpp unicode-data.cpp + unicode.cpp + unicode.h ) -target_include_directories(llama PUBLIC . ../include ../common) +target_include_directories(llama PUBLIC . ../include) target_compile_features (llama PUBLIC cxx_std_17) # don't bump target_link_libraries(llama PUBLIC ggml) diff --git a/src/llama-adapter.cpp b/src/llama-adapter.cpp index 8a0800463..7ac54d239 100644 --- a/src/llama-adapter.cpp +++ b/src/llama-adapter.cpp @@ -4,14 +4,13 @@ #include "llama-mmap.h" #include "llama-model.h" -#include #include #include #include // vec -struct ggml_tensor * llama_adapter_cvec::tensor_for(int il) const { +ggml_tensor * llama_adapter_cvec::tensor_for(int il) const { if (il < 0 || il < layer_start || il > layer_end || (size_t) il >= tensors.size()) { return nullptr; } @@ -19,7 +18,7 @@ struct ggml_tensor * llama_adapter_cvec::tensor_for(int il) const { return tensors[il]; } -struct ggml_tensor * llama_adapter_cvec::apply_to(struct ggml_context * ctx, struct ggml_tensor * cur, int il) const { +ggml_tensor * llama_adapter_cvec::apply_to(ggml_context * ctx, ggml_tensor * cur, int il) const { ggml_tensor * layer_dir = tensor_for(il); if (layer_dir != nullptr) { cur = ggml_add(ctx, cur, layer_dir); @@ -40,7 +39,7 @@ bool llama_adapter_cvec::init(const llama_model & model) { auto ctx_for_buft = [&](ggml_backend_buffer_type_t buft) -> ggml_context * { auto it = ctx_map.find(buft); if (it == ctx_map.end()) { - struct ggml_init_params params = { + ggml_init_params params = { /*.mem_size =*/ hparams.n_layer*ggml_tensor_overhead(), /*.mem_buffer =*/ NULL, /*.no_alloc =*/ true, @@ -91,7 +90,7 @@ bool llama_adapter_cvec::init(const llama_model & model) { return true; } -int32_t llama_adapter_cvec::apply( +bool llama_adapter_cvec::apply( const llama_model & model, const float * data, size_t len, @@ -104,17 +103,17 @@ int32_t llama_adapter_cvec::apply( // disable the current control vector (but leave allocated for later) layer_start = -1; layer_end = -1; - return 0; + return true; } if (n_embd != (int) hparams.n_embd) { LLAMA_LOG_ERROR("%s: control vector n_embd does not match model\n", __func__); - return 1; + return false; } if (tensors.empty()) { if (!init(model)) { - return 1; + return false; } } @@ -130,12 +129,12 @@ int32_t llama_adapter_cvec::apply( } } - return 0; + return true; } // lora -llama_adapter_lora_weight * llama_adapter_lora::get_weight(struct ggml_tensor * w) { +llama_adapter_lora_weight * llama_adapter_lora::get_weight(ggml_tensor * w) { const std::string name(w->name); const auto pos = ab_map.find(name); @@ -146,11 +145,11 @@ llama_adapter_lora_weight * llama_adapter_lora::get_weight(struct ggml_tensor * return nullptr; } -static void llama_adapter_lora_init_impl(struct llama_model & model, const char * path_lora, struct llama_adapter_lora & adapter) { +static void llama_adapter_lora_init_impl(llama_model & model, const char * path_lora, llama_adapter_lora & adapter) { LLAMA_LOG_INFO("%s: loading lora adapter from '%s' ...\n", __func__, path_lora); ggml_context * ctx_init; - struct gguf_init_params meta_gguf_params = { + gguf_init_params meta_gguf_params = { /* .no_alloc = */ true, /* .ctx = */ &ctx_init, }; @@ -201,7 +200,7 @@ static void llama_adapter_lora_init_impl(struct llama_model & model, const char auto it = ctx_map.find(buft); if (it == ctx_map.end()) { // add a new context - struct ggml_init_params params = { + ggml_init_params params = { /*.mem_size =*/ n_tensors*ggml_tensor_overhead(), /*.mem_buffer =*/ NULL, /*.no_alloc =*/ true, @@ -248,6 +247,26 @@ static void llama_adapter_lora_init_impl(struct llama_model & model, const char } } + // get extra buffer types of the CPU + // TODO: a more general solution for non-CPU extra buft should be imlpemented in the future + // ref: https://github.com/ggml-org/llama.cpp/pull/12593#pullrequestreview-2718659948 + std::vector buft_extra; + { + auto * cpu_dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU); + auto * cpu_reg = ggml_backend_dev_backend_reg(cpu_dev); + + auto ggml_backend_dev_get_extra_bufts_fn = (ggml_backend_dev_get_extra_bufts_t) + ggml_backend_reg_get_proc_address(cpu_reg, "ggml_backend_dev_get_extra_bufts"); + + if (ggml_backend_dev_get_extra_bufts_fn) { + ggml_backend_buffer_type_t * extra_bufts = ggml_backend_dev_get_extra_bufts_fn(cpu_dev); + while (extra_bufts && *extra_bufts) { + buft_extra.emplace_back(*extra_bufts); + ++extra_bufts; + } + } + } + // add tensors for (auto & it : ab_map) { const std::string & name = it.first; @@ -264,7 +283,23 @@ static void llama_adapter_lora_init_impl(struct llama_model & model, const char throw std::runtime_error("LoRA tensor '" + name + "' does not exist in base model (hint: maybe wrong base model?)"); } - struct ggml_context * dev_ctx = ctx_for_buft(ggml_backend_buffer_get_type(model_tensor->buffer)); + auto * buft = ggml_backend_buffer_get_type(model_tensor->buffer); + + // do not load loras to extra buffer types (i.e. bufts for repacking) -> use the CPU in that case + for (auto & ex : buft_extra) { + if (ex == buft) { + LLAMA_LOG_WARN("%s: lora for '%s' cannot use buft '%s', fallback to CPU\n", __func__, model_tensor->name, ggml_backend_buft_name(buft)); + + auto * cpu_dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU); + buft = ggml_backend_dev_buffer_type(cpu_dev); + + break; + } + } + + LLAMA_LOG_DEBUG("%s: lora for '%s' -> '%s'\n", __func__, model_tensor->name, ggml_backend_buft_name(buft)); + + ggml_context * dev_ctx = ctx_for_buft(buft); // validate tensor shape if (is_token_embd) { // expect B to be non-transposed, A and B are flipped; see llm_build_inp_embd() @@ -281,8 +316,8 @@ static void llama_adapter_lora_init_impl(struct llama_model & model, const char } // save tensor to adapter - struct ggml_tensor * tensor_a = ggml_dup_tensor(dev_ctx, w.a); - struct ggml_tensor * tensor_b = ggml_dup_tensor(dev_ctx, w.b); + ggml_tensor * tensor_a = ggml_dup_tensor(dev_ctx, w.a); + ggml_tensor * tensor_b = ggml_dup_tensor(dev_ctx, w.b); ggml_set_name(tensor_a, w.a->name); ggml_set_name(tensor_b, w.b->name); adapter.ab_map[name] = llama_adapter_lora_weight(tensor_a, tensor_b); @@ -308,7 +343,7 @@ static void llama_adapter_lora_init_impl(struct llama_model & model, const char { llama_file gguf_file(path_lora, "rb"); std::vector read_buf; - auto set_tensor = [&](struct ggml_tensor * orig, struct ggml_tensor * dev) { + auto set_tensor = [&](ggml_tensor * orig, ggml_tensor * dev) { size_t offs = gguf_get_data_offset(ctx_gguf.get()) + gguf_get_tensor_offset(ctx_gguf.get(), gguf_find_tensor(ctx_gguf.get(), orig->name)); size_t size = ggml_nbytes(orig); read_buf.resize(size); @@ -327,8 +362,8 @@ static void llama_adapter_lora_init_impl(struct llama_model & model, const char LLAMA_LOG_INFO("%s: loaded %zu tensors from lora file\n", __func__, adapter.ab_map.size()*2); } -struct llama_adapter_lora * llama_adapter_lora_init(struct llama_model * model, const char * path_lora) { - struct llama_adapter_lora * adapter = new llama_adapter_lora(); +llama_adapter_lora * llama_adapter_lora_init(llama_model * model, const char * path_lora) { + llama_adapter_lora * adapter = new llama_adapter_lora(); try { llama_adapter_lora_init_impl(*model, path_lora, *adapter); @@ -342,6 +377,6 @@ struct llama_adapter_lora * llama_adapter_lora_init(struct llama_model * model, return nullptr; } -void llama_adapter_lora_free(struct llama_adapter_lora * adapter) { +void llama_adapter_lora_free(llama_adapter_lora * adapter) { delete adapter; } diff --git a/src/llama-adapter.h b/src/llama-adapter.h index 603fa08f6..65824e972 100644 --- a/src/llama-adapter.h +++ b/src/llama-adapter.h @@ -15,11 +15,11 @@ // struct llama_adapter_cvec { - struct ggml_tensor * tensor_for(int il) const; + ggml_tensor * tensor_for(int il) const; - struct ggml_tensor * apply_to(struct ggml_context * ctx, struct ggml_tensor * cur, int il) const; + ggml_tensor * apply_to(ggml_context * ctx, ggml_tensor * cur, int il) const; - int32_t apply( + bool apply( const llama_model & model, const float * data, size_t len, @@ -36,7 +36,7 @@ private: std::vector ctxs; std::vector bufs; - std::vector tensors; // per layer + std::vector tensors; // per layer }; // @@ -44,8 +44,8 @@ private: // struct llama_adapter_lora_weight { - struct ggml_tensor * a = nullptr; - struct ggml_tensor * b = nullptr; + ggml_tensor * a = nullptr; + ggml_tensor * b = nullptr; // get actual scale based on rank and alpha float get_scale(float alpha, float adapter_scale) const { @@ -55,12 +55,12 @@ struct llama_adapter_lora_weight { } llama_adapter_lora_weight() = default; - llama_adapter_lora_weight(struct ggml_tensor * a, struct ggml_tensor * b) : a(a), b(b) {} + llama_adapter_lora_weight(ggml_tensor * a, ggml_tensor * b) : a(a), b(b) {} }; struct llama_adapter_lora { // map tensor name to lora_a_b - std::unordered_map ab_map; + std::unordered_map ab_map; std::vector ctxs; std::vector bufs; @@ -70,5 +70,7 @@ struct llama_adapter_lora { llama_adapter_lora() = default; ~llama_adapter_lora() = default; - llama_adapter_lora_weight * get_weight(struct ggml_tensor * w); + llama_adapter_lora_weight * get_weight(ggml_tensor * w); }; + +using llama_adapter_loras = std::unordered_map; diff --git a/src/llama-arch.cpp b/src/llama-arch.cpp index 97a1e7e5e..a6fddc7fd 100644 --- a/src/llama-arch.cpp +++ b/src/llama-arch.cpp @@ -6,6 +6,7 @@ static const std::map LLM_ARCH_NAMES = { { LLM_ARCH_LLAMA, "llama" }, + { LLM_ARCH_LLAMA4, "llama4" }, { LLM_ARCH_DECI, "deci" }, { LLM_ARCH_FALCON, "falcon" }, { LLM_ARCH_GROK, "grok" }, @@ -25,6 +26,8 @@ static const std::map LLM_ARCH_NAMES = { { LLM_ARCH_QWEN2, "qwen2" }, { LLM_ARCH_QWEN2MOE, "qwen2moe" }, { LLM_ARCH_QWEN2VL, "qwen2vl" }, + { LLM_ARCH_QWEN3, "qwen3" }, + { LLM_ARCH_QWEN3MOE, "qwen3moe" }, { LLM_ARCH_PHI2, "phi2" }, { LLM_ARCH_PHI3, "phi3" }, { LLM_ARCH_PHIMOE, "phimoe" }, @@ -36,6 +39,7 @@ static const std::map LLM_ARCH_NAMES = { { LLM_ARCH_MINICPM3, "minicpm3" }, { LLM_ARCH_GEMMA, "gemma" }, { LLM_ARCH_GEMMA2, "gemma2" }, + { LLM_ARCH_GEMMA3, "gemma3" }, { LLM_ARCH_STARCODER2, "starcoder2" }, { LLM_ARCH_MAMBA, "mamba" }, { LLM_ARCH_XVERSE, "xverse" }, @@ -50,6 +54,7 @@ static const std::map LLM_ARCH_NAMES = { { LLM_ARCH_DEEPSEEK, "deepseek" }, { LLM_ARCH_DEEPSEEK2, "deepseek2" }, { LLM_ARCH_CHATGLM, "chatglm" }, + { LLM_ARCH_GLM4, "glm4" }, { LLM_ARCH_BITNET, "bitnet" }, { LLM_ARCH_T5, "t5" }, { LLM_ARCH_T5ENCODER, "t5encoder" }, @@ -58,10 +63,14 @@ static const std::map LLM_ARCH_NAMES = { { LLM_ARCH_EXAONE, "exaone" }, { LLM_ARCH_RWKV6, "rwkv6" }, { LLM_ARCH_RWKV6QWEN2, "rwkv6qwen2" }, + { LLM_ARCH_RWKV7, "rwkv7" }, + { LLM_ARCH_ARWKV7, "arwkv7" }, { LLM_ARCH_GRANITE, "granite" }, { LLM_ARCH_GRANITE_MOE, "granitemoe" }, { LLM_ARCH_CHAMELEON, "chameleon" }, { LLM_ARCH_WAVTOKENIZER_DEC, "wavtokenizer-dec" }, + { LLM_ARCH_PLM, "plm" }, + { LLM_ARCH_BAILINGMOE, "bailingmoe" }, { LLM_ARCH_UNKNOWN, "(unknown)" }, }; @@ -70,6 +79,7 @@ static const std::map LLM_KV_NAMES = { { LLM_KV_GENERAL_ARCHITECTURE, "general.architecture" }, { LLM_KV_GENERAL_QUANTIZATION_VERSION, "general.quantization_version" }, { LLM_KV_GENERAL_ALIGNMENT, "general.alignment" }, + { LLM_KV_GENERAL_FILE_TYPE, "general.file_type" }, { LLM_KV_GENERAL_NAME, "general.name" }, { LLM_KV_GENERAL_AUTHOR, "general.author" }, { LLM_KV_GENERAL_VERSION, "general.version" }, @@ -108,23 +118,28 @@ static const std::map LLM_KV_NAMES = { { LLM_KV_RESIDUAL_SCALE, "%s.residual_scale" }, { LLM_KV_EMBEDDING_SCALE, "%s.embedding_scale" }, { LLM_KV_TOKEN_SHIFT_COUNT, "%s.token_shift_count" }, + { LLM_KV_INTERLEAVE_MOE_LAYER_STEP, "%s.interleave_moe_layer_step" }, - { LLM_KV_ATTENTION_HEAD_COUNT, "%s.attention.head_count" }, - { LLM_KV_ATTENTION_HEAD_COUNT_KV, "%s.attention.head_count_kv" }, - { LLM_KV_ATTENTION_MAX_ALIBI_BIAS, "%s.attention.max_alibi_bias" }, - { LLM_KV_ATTENTION_CLAMP_KQV, "%s.attention.clamp_kqv" }, - { LLM_KV_ATTENTION_KEY_LENGTH, "%s.attention.key_length" }, - { LLM_KV_ATTENTION_VALUE_LENGTH, "%s.attention.value_length" }, - { LLM_KV_ATTENTION_LAYERNORM_EPS, "%s.attention.layer_norm_epsilon" }, - { LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, "%s.attention.layer_norm_rms_epsilon" }, - { LLM_KV_ATTENTION_GROUPNORM_EPS, "%s.attention.group_norm_epsilon" }, - { LLM_KV_ATTENTION_GROUPNORM_GROUPS, "%s.attention.group_norm_groups" }, - { LLM_KV_ATTENTION_CAUSAL, "%s.attention.causal" }, - { LLM_KV_ATTENTION_Q_LORA_RANK, "%s.attention.q_lora_rank" }, - { LLM_KV_ATTENTION_KV_LORA_RANK, "%s.attention.kv_lora_rank" }, - { LLM_KV_ATTENTION_RELATIVE_BUCKETS_COUNT, "%s.attention.relative_buckets_count" }, - { LLM_KV_ATTENTION_SLIDING_WINDOW, "%s.attention.sliding_window" }, - { LLM_KV_ATTENTION_SCALE, "%s.attention.scale" }, + { LLM_KV_ATTENTION_HEAD_COUNT, "%s.attention.head_count" }, + { LLM_KV_ATTENTION_HEAD_COUNT_KV, "%s.attention.head_count_kv" }, + { LLM_KV_ATTENTION_MAX_ALIBI_BIAS, "%s.attention.max_alibi_bias" }, + { LLM_KV_ATTENTION_CLAMP_KQV, "%s.attention.clamp_kqv" }, + { LLM_KV_ATTENTION_KEY_LENGTH, "%s.attention.key_length" }, + { LLM_KV_ATTENTION_VALUE_LENGTH, "%s.attention.value_length" }, + { LLM_KV_ATTENTION_LAYERNORM_EPS, "%s.attention.layer_norm_epsilon" }, + { LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, "%s.attention.layer_norm_rms_epsilon" }, + { LLM_KV_ATTENTION_GROUPNORM_EPS, "%s.attention.group_norm_epsilon" }, + { LLM_KV_ATTENTION_GROUPNORM_GROUPS, "%s.attention.group_norm_groups" }, + { LLM_KV_ATTENTION_CAUSAL, "%s.attention.causal" }, + { LLM_KV_ATTENTION_Q_LORA_RANK, "%s.attention.q_lora_rank" }, + { LLM_KV_ATTENTION_KV_LORA_RANK, "%s.attention.kv_lora_rank" }, + { LLM_KV_ATTENTION_DECAY_LORA_RANK, "%s.attention.decay_lora_rank" }, + { LLM_KV_ATTENTION_ICLR_LORA_RANK, "%s.attention.iclr_lora_rank" }, + { LLM_KV_ATTENTION_VALUE_RESIDUAL_MIX_LORA_RANK, "%s.attention.value_residual_mix_lora_rank" }, + { LLM_KV_ATTENTION_GATE_LORA_RANK, "%s.attention.gate_lora_rank" }, + { LLM_KV_ATTENTION_RELATIVE_BUCKETS_COUNT, "%s.attention.relative_buckets_count" }, + { LLM_KV_ATTENTION_SLIDING_WINDOW, "%s.attention.sliding_window" }, + { LLM_KV_ATTENTION_SCALE, "%s.attention.scale" }, { LLM_KV_ROPE_DIMENSION_COUNT, "%s.rope.dimension_count" }, { LLM_KV_ROPE_DIMENSION_SECTIONS, "%s.rope.dimension_sections" }, @@ -223,6 +238,35 @@ static const std::map> LLM_TENSOR_N { LLM_TENSOR_FFN_UP_EXPS, "blk.%d.ffn_up_exps" }, }, }, + { + LLM_ARCH_LLAMA4, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, + { LLM_TENSOR_OUTPUT, "output" }, + { LLM_TENSOR_ROPE_FREQS, "rope_freqs" }, + { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, + { LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" }, + { LLM_TENSOR_ATTN_K, "blk.%d.attn_k" }, + { LLM_TENSOR_ATTN_V, "blk.%d.attn_v" }, + { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, + { LLM_TENSOR_ATTN_ROT_EMBD, "blk.%d.attn_rot_embd" }, + { LLM_TENSOR_FFN_GATE_INP, "blk.%d.ffn_gate_inp" }, + { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" }, + { LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" }, + { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" }, + { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, + { LLM_TENSOR_FFN_GATE_EXP, "blk.%d.ffn_gate.%d" }, + { LLM_TENSOR_FFN_DOWN_EXP, "blk.%d.ffn_down.%d" }, + { LLM_TENSOR_FFN_UP_EXP, "blk.%d.ffn_up.%d" }, + { LLM_TENSOR_FFN_GATE_EXPS, "blk.%d.ffn_gate_exps" }, + { LLM_TENSOR_FFN_DOWN_EXPS, "blk.%d.ffn_down_exps" }, + { LLM_TENSOR_FFN_UP_EXPS, "blk.%d.ffn_up_exps" }, + { LLM_TENSOR_FFN_GATE_SHEXP, "blk.%d.ffn_gate_shexp" }, + { LLM_TENSOR_FFN_DOWN_SHEXP, "blk.%d.ffn_down_shexp" }, + { LLM_TENSOR_FFN_UP_SHEXP, "blk.%d.ffn_up_shexp" }, + }, + }, { LLM_ARCH_DECI, { @@ -554,6 +598,45 @@ static const std::map> LLM_TENSOR_N { LLM_TENSOR_FFN_UP_SHEXP, "blk.%d.ffn_up_shexp" }, }, }, + { + LLM_ARCH_QWEN3, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, + { LLM_TENSOR_OUTPUT, "output" }, + { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, + { LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" }, + { LLM_TENSOR_ATTN_Q_NORM, "blk.%d.attn_q_norm" }, + { LLM_TENSOR_ATTN_K, "blk.%d.attn_k" }, + { LLM_TENSOR_ATTN_K_NORM, "blk.%d.attn_k_norm" }, + { LLM_TENSOR_ATTN_V, "blk.%d.attn_v" }, + { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, + { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" }, + { LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" }, + { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" }, + { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, + }, + }, + { + LLM_ARCH_QWEN3MOE, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, + { LLM_TENSOR_OUTPUT, "output" }, + { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, + { LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" }, + { LLM_TENSOR_ATTN_Q_NORM, "blk.%d.attn_q_norm" }, + { LLM_TENSOR_ATTN_K, "blk.%d.attn_k" }, + { LLM_TENSOR_ATTN_K_NORM, "blk.%d.attn_k_norm" }, + { LLM_TENSOR_ATTN_V, "blk.%d.attn_v" }, + { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, + { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" }, + { LLM_TENSOR_FFN_GATE_INP, "blk.%d.ffn_gate_inp" }, + { LLM_TENSOR_FFN_GATE_EXPS, "blk.%d.ffn_gate_exps" }, + { LLM_TENSOR_FFN_DOWN_EXPS, "blk.%d.ffn_down_exps" }, + { LLM_TENSOR_FFN_UP_EXPS, "blk.%d.ffn_up_exps" }, + }, + }, { LLM_ARCH_PHI2, { @@ -766,6 +849,27 @@ static const std::map> LLM_TENSOR_N { LLM_TENSOR_FFN_POST_NORM, "blk.%d.post_ffw_norm" }, }, }, + { + LLM_ARCH_GEMMA3, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, + { LLM_TENSOR_OUTPUT, "output" }, + { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, + { LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" }, + { LLM_TENSOR_ATTN_Q_NORM, "blk.%d.attn_q_norm" }, + { LLM_TENSOR_ATTN_K, "blk.%d.attn_k" }, + { LLM_TENSOR_ATTN_K_NORM, "blk.%d.attn_k_norm" }, + { LLM_TENSOR_ATTN_V, "blk.%d.attn_v" }, + { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, + { LLM_TENSOR_ATTN_POST_NORM, "blk.%d.post_attention_norm" }, + { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" }, + { LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" }, + { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" }, + { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, + { LLM_TENSOR_FFN_POST_NORM, "blk.%d.post_ffw_norm" }, + }, + }, { LLM_ARCH_STARCODER2, { @@ -1015,6 +1119,22 @@ static const std::map> LLM_TENSOR_N { LLM_TENSOR_FFN_EXP_PROBS_B, "blk.%d.exp_probs_b" }, }, }, + { + LLM_ARCH_PLM, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, + { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, + { LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" }, + { LLM_TENSOR_ATTN_KV_A_MQA, "blk.%d.attn_kv_a_mqa" }, + { LLM_TENSOR_ATTN_KV_A_NORM, "blk.%d.attn_kv_a_norm" }, + { LLM_TENSOR_ATTN_KV_B, "blk.%d.attn_kv_b" }, + { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, + { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" }, + { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" }, + { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, + }, + }, { LLM_ARCH_CHATGLM, { @@ -1033,6 +1153,25 @@ static const std::map> LLM_TENSOR_N { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" }, }, }, + { + LLM_ARCH_GLM4, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_ROPE_FREQS, "rope_freqs" }, + { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, + { LLM_TENSOR_OUTPUT, "output" }, + { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, + { LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" }, + { LLM_TENSOR_ATTN_K, "blk.%d.attn_k" }, + { LLM_TENSOR_ATTN_V, "blk.%d.attn_v" }, + { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, + { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" }, + { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, + { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" }, + { LLM_TENSOR_ATTN_POST_NORM, "blk.%d.post_attention_norm" }, + { LLM_TENSOR_FFN_POST_NORM, "blk.%d.post_ffw_norm" }, + }, + }, { LLM_ARCH_BITNET, { @@ -1217,6 +1356,74 @@ static const std::map> LLM_TENSOR_N { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, }, }, + { + LLM_ARCH_RWKV7, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_TOKEN_EMBD_NORM, "token_embd_norm" }, + { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, + { LLM_TENSOR_OUTPUT, "output" }, + { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, + { LLM_TENSOR_ATTN_NORM_2, "blk.%d.attn_norm_2" }, + { LLM_TENSOR_TIME_MIX_W0, "blk.%d.time_mix_w0" }, + { LLM_TENSOR_TIME_MIX_W1, "blk.%d.time_mix_w1" }, + { LLM_TENSOR_TIME_MIX_W2, "blk.%d.time_mix_w2" }, + { LLM_TENSOR_TIME_MIX_A0, "blk.%d.time_mix_a0" }, + { LLM_TENSOR_TIME_MIX_A1, "blk.%d.time_mix_a1" }, + { LLM_TENSOR_TIME_MIX_A2, "blk.%d.time_mix_a2" }, + { LLM_TENSOR_TIME_MIX_V0, "blk.%d.time_mix_v0" }, + { LLM_TENSOR_TIME_MIX_V1, "blk.%d.time_mix_v1" }, + { LLM_TENSOR_TIME_MIX_V2, "blk.%d.time_mix_v2" }, + { LLM_TENSOR_TIME_MIX_G1, "blk.%d.time_mix_g1" }, + { LLM_TENSOR_TIME_MIX_G2, "blk.%d.time_mix_g2" }, + { LLM_TENSOR_TIME_MIX_K_K, "blk.%d.time_mix_k_k" }, + { LLM_TENSOR_TIME_MIX_K_A, "blk.%d.time_mix_k_a" }, + { LLM_TENSOR_TIME_MIX_R_K, "blk.%d.time_mix_r_k" }, + { LLM_TENSOR_TIME_MIX_LERP_FUSED, "blk.%d.time_mix_lerp_fused" }, + { LLM_TENSOR_TIME_MIX_KEY, "blk.%d.time_mix_key" }, + { LLM_TENSOR_TIME_MIX_VALUE, "blk.%d.time_mix_value" }, + { LLM_TENSOR_TIME_MIX_RECEPTANCE, "blk.%d.time_mix_receptance" }, + { LLM_TENSOR_TIME_MIX_LN, "blk.%d.time_mix_ln" }, + { LLM_TENSOR_TIME_MIX_OUTPUT, "blk.%d.time_mix_output" }, + { LLM_TENSOR_CHANNEL_MIX_LERP_K, "blk.%d.channel_mix_lerp_k" }, + { LLM_TENSOR_CHANNEL_MIX_KEY, "blk.%d.channel_mix_key" }, + { LLM_TENSOR_CHANNEL_MIX_VALUE, "blk.%d.channel_mix_value" }, + }, + }, + { + LLM_ARCH_ARWKV7, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_TOKEN_EMBD_NORM, "token_embd_norm" }, + { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, + { LLM_TENSOR_OUTPUT, "output" }, + { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, + { LLM_TENSOR_TIME_MIX_W0, "blk.%d.time_mix_w0" }, + { LLM_TENSOR_TIME_MIX_W1, "blk.%d.time_mix_w1" }, + { LLM_TENSOR_TIME_MIX_W2, "blk.%d.time_mix_w2" }, + { LLM_TENSOR_TIME_MIX_A0, "blk.%d.time_mix_a0" }, + { LLM_TENSOR_TIME_MIX_A1, "blk.%d.time_mix_a1" }, + { LLM_TENSOR_TIME_MIX_A2, "blk.%d.time_mix_a2" }, + { LLM_TENSOR_TIME_MIX_V0, "blk.%d.time_mix_v0" }, + { LLM_TENSOR_TIME_MIX_V1, "blk.%d.time_mix_v1" }, + { LLM_TENSOR_TIME_MIX_V2, "blk.%d.time_mix_v2" }, + { LLM_TENSOR_TIME_MIX_G1, "blk.%d.time_mix_g1" }, + { LLM_TENSOR_TIME_MIX_G2, "blk.%d.time_mix_g2" }, + { LLM_TENSOR_TIME_MIX_K_K, "blk.%d.time_mix_k_k" }, + { LLM_TENSOR_TIME_MIX_K_A, "blk.%d.time_mix_k_a" }, + { LLM_TENSOR_TIME_MIX_R_K, "blk.%d.time_mix_r_k" }, + { LLM_TENSOR_TIME_MIX_LERP_FUSED, "blk.%d.time_mix_lerp_fused" }, + { LLM_TENSOR_TIME_MIX_KEY, "blk.%d.time_mix_key" }, + { LLM_TENSOR_TIME_MIX_VALUE, "blk.%d.time_mix_value" }, + { LLM_TENSOR_TIME_MIX_RECEPTANCE, "blk.%d.time_mix_receptance" }, + { LLM_TENSOR_TIME_MIX_LN, "blk.%d.time_mix_ln" }, + { LLM_TENSOR_TIME_MIX_OUTPUT, "blk.%d.time_mix_output" }, + { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" }, + { LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" }, + { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" }, + { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, + }, + }, { LLM_ARCH_GRANITE, { @@ -1296,6 +1503,29 @@ static const std::map> LLM_TENSOR_N { LLM_TENSOR_POS_NET_ATTN_OUT, "posnet.%d.attn_output" }, }, }, + { + LLM_ARCH_BAILINGMOE, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, + { LLM_TENSOR_OUTPUT, "output" }, + { LLM_TENSOR_ROPE_FREQS, "rope_freqs" }, + { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, + { LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" }, + { LLM_TENSOR_ATTN_K, "blk.%d.attn_k" }, + { LLM_TENSOR_ATTN_V, "blk.%d.attn_v" }, + { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, + { LLM_TENSOR_FFN_GATE_INP, "blk.%d.ffn_gate_inp" }, + { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" }, + { LLM_TENSOR_FFN_GATE_EXPS, "blk.%d.ffn_gate_exps" }, + { LLM_TENSOR_FFN_DOWN_EXPS, "blk.%d.ffn_down_exps" }, + { LLM_TENSOR_FFN_UP_EXPS, "blk.%d.ffn_up_exps" }, + { LLM_TENSOR_FFN_GATE_INP_SHEXP, "blk.%d.ffn_gate_inp_shexp" }, + { LLM_TENSOR_FFN_GATE_SHEXP, "blk.%d.ffn_gate_shexp" }, + { LLM_TENSOR_FFN_DOWN_SHEXP, "blk.%d.ffn_down_shexp" }, + { LLM_TENSOR_FFN_UP_SHEXP, "blk.%d.ffn_up_shexp" }, + }, + }, { LLM_ARCH_UNKNOWN, { @@ -1376,6 +1606,12 @@ static const std::map LLM_TENSOR_INFOS = { {LLM_TENSOR_SSM_OUT, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, {LLM_TENSOR_TIME_MIX_W1, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, {LLM_TENSOR_TIME_MIX_W2, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, + {LLM_TENSOR_TIME_MIX_A1, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, + {LLM_TENSOR_TIME_MIX_A2, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, + {LLM_TENSOR_TIME_MIX_V1, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, + {LLM_TENSOR_TIME_MIX_V2, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, + {LLM_TENSOR_TIME_MIX_G1, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, + {LLM_TENSOR_TIME_MIX_G2, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, {LLM_TENSOR_TIME_MIX_DECAY_W1, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, {LLM_TENSOR_TIME_MIX_DECAY_W2, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, {LLM_TENSOR_TIME_MIX_KEY, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, @@ -1394,6 +1630,9 @@ static const std::map LLM_TENSOR_INFOS = { {LLM_TENSOR_TIME_MIX_LN, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, {LLM_TENSOR_CHANNEL_MIX_LERP_K, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, {LLM_TENSOR_CHANNEL_MIX_LERP_R, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, + {LLM_TENSOR_TIME_MIX_K_K, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, + {LLM_TENSOR_TIME_MIX_K_A, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, + {LLM_TENSOR_TIME_MIX_R_K, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, {LLM_TENSOR_TIME_MIX_LERP_W, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}}, {LLM_TENSOR_TIME_MIX_LERP_K, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}}, {LLM_TENSOR_TIME_MIX_LERP_V, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}}, @@ -1401,6 +1640,9 @@ static const std::map LLM_TENSOR_INFOS = { {LLM_TENSOR_TIME_MIX_LERP_G, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}}, {LLM_TENSOR_TIME_MIX_LERP_FUSED, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}}, {LLM_TENSOR_TIME_MIX_DECAY, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}}, + {LLM_TENSOR_TIME_MIX_W0, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}}, + {LLM_TENSOR_TIME_MIX_A0, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}}, + {LLM_TENSOR_TIME_MIX_V0, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}}, {LLM_TENSOR_TIME_MIX_FIRST, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_RWKV_WKV6}}, {LLM_TENSOR_ATTN_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, {LLM_TENSOR_ATTN_NORM_2, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, diff --git a/src/llama-arch.h b/src/llama-arch.h index 122fdcebe..2c2099b3c 100644 --- a/src/llama-arch.h +++ b/src/llama-arch.h @@ -10,6 +10,7 @@ enum llm_arch { LLM_ARCH_LLAMA, + LLM_ARCH_LLAMA4, LLM_ARCH_DECI, LLM_ARCH_FALCON, LLM_ARCH_BAICHUAN, @@ -29,6 +30,8 @@ enum llm_arch { LLM_ARCH_QWEN2, LLM_ARCH_QWEN2MOE, LLM_ARCH_QWEN2VL, + LLM_ARCH_QWEN3, + LLM_ARCH_QWEN3MOE, LLM_ARCH_PHI2, LLM_ARCH_PHI3, LLM_ARCH_PHIMOE, @@ -40,6 +43,7 @@ enum llm_arch { LLM_ARCH_MINICPM3, LLM_ARCH_GEMMA, LLM_ARCH_GEMMA2, + LLM_ARCH_GEMMA3, LLM_ARCH_STARCODER2, LLM_ARCH_MAMBA, LLM_ARCH_XVERSE, @@ -54,6 +58,7 @@ enum llm_arch { LLM_ARCH_DEEPSEEK, LLM_ARCH_DEEPSEEK2, LLM_ARCH_CHATGLM, + LLM_ARCH_GLM4, LLM_ARCH_BITNET, LLM_ARCH_T5, LLM_ARCH_T5ENCODER, @@ -62,10 +67,14 @@ enum llm_arch { LLM_ARCH_EXAONE, LLM_ARCH_RWKV6, LLM_ARCH_RWKV6QWEN2, + LLM_ARCH_RWKV7, + LLM_ARCH_ARWKV7, LLM_ARCH_GRANITE, LLM_ARCH_GRANITE_MOE, LLM_ARCH_CHAMELEON, LLM_ARCH_WAVTOKENIZER_DEC, + LLM_ARCH_PLM, + LLM_ARCH_BAILINGMOE, LLM_ARCH_UNKNOWN, }; @@ -74,6 +83,7 @@ enum llm_kv { LLM_KV_GENERAL_ARCHITECTURE, LLM_KV_GENERAL_QUANTIZATION_VERSION, LLM_KV_GENERAL_ALIGNMENT, + LLM_KV_GENERAL_FILE_TYPE, LLM_KV_GENERAL_NAME, LLM_KV_GENERAL_AUTHOR, LLM_KV_GENERAL_VERSION, @@ -112,6 +122,7 @@ enum llm_kv { LLM_KV_RESIDUAL_SCALE, LLM_KV_EMBEDDING_SCALE, LLM_KV_TOKEN_SHIFT_COUNT, + LLM_KV_INTERLEAVE_MOE_LAYER_STEP, LLM_KV_ATTENTION_HEAD_COUNT, LLM_KV_ATTENTION_HEAD_COUNT_KV, @@ -126,6 +137,10 @@ enum llm_kv { LLM_KV_ATTENTION_CAUSAL, LLM_KV_ATTENTION_Q_LORA_RANK, LLM_KV_ATTENTION_KV_LORA_RANK, + LLM_KV_ATTENTION_DECAY_LORA_RANK, + LLM_KV_ATTENTION_ICLR_LORA_RANK, + LLM_KV_ATTENTION_VALUE_RESIDUAL_MIX_LORA_RANK, + LLM_KV_ATTENTION_GATE_LORA_RANK, LLM_KV_ATTENTION_RELATIVE_BUCKETS_COUNT, LLM_KV_ATTENTION_SLIDING_WINDOW, LLM_KV_ATTENTION_SCALE, @@ -242,6 +257,8 @@ enum llm_tensor { LLM_TENSOR_ATTN_Q_NORM, LLM_TENSOR_ATTN_K_NORM, LLM_TENSOR_LAYER_OUT_NORM, + LLM_TENSOR_POST_ATTN_NORM, + LLM_TENSOR_POST_MLP_NORM, LLM_TENSOR_SSM_IN, LLM_TENSOR_SSM_CONV1D, LLM_TENSOR_SSM_X, @@ -249,8 +266,20 @@ enum llm_tensor { LLM_TENSOR_SSM_A, LLM_TENSOR_SSM_D, LLM_TENSOR_SSM_OUT, + LLM_TENSOR_TIME_MIX_W0, LLM_TENSOR_TIME_MIX_W1, LLM_TENSOR_TIME_MIX_W2, + LLM_TENSOR_TIME_MIX_A0, + LLM_TENSOR_TIME_MIX_A1, + LLM_TENSOR_TIME_MIX_A2, + LLM_TENSOR_TIME_MIX_V0, + LLM_TENSOR_TIME_MIX_V1, + LLM_TENSOR_TIME_MIX_V2, + LLM_TENSOR_TIME_MIX_G1, + LLM_TENSOR_TIME_MIX_G2, + LLM_TENSOR_TIME_MIX_K_K, + LLM_TENSOR_TIME_MIX_K_A, + LLM_TENSOR_TIME_MIX_R_K, LLM_TENSOR_TIME_MIX_LERP_X, LLM_TENSOR_TIME_MIX_LERP_W, LLM_TENSOR_TIME_MIX_LERP_K, diff --git a/src/llama-batch.h b/src/llama-batch.h index 773c3808b..f1df40d27 100644 --- a/src/llama-batch.h +++ b/src/llama-batch.h @@ -42,9 +42,9 @@ struct llama_sbatch { bool logits_all; // TODO: remove once lctx.logits_all is removed too // sorted indices into the batch - std::vector ids; + std::vector ids; // batch indices of the output - std::vector out_ids; + std::vector out_ids; std::vector seq; const llama_batch * batch = nullptr; diff --git a/src/llama-chat.cpp b/src/llama-chat.cpp index 028a64794..721faa4e8 100644 --- a/src/llama-chat.cpp +++ b/src/llama-chat.cpp @@ -4,6 +4,7 @@ #include #include +#include #if __cplusplus >= 202000L #define LU8(x) (const char*)(u8##x) @@ -58,6 +59,9 @@ static const std::map LLM_CHAT_TEMPLATES = { { "granite", LLM_CHAT_TEMPLATE_GRANITE }, { "gigachat", LLM_CHAT_TEMPLATE_GIGACHAT }, { "megrez", LLM_CHAT_TEMPLATE_MEGREZ }, + { "yandex", LLM_CHAT_TEMPLATE_YANDEX }, + { "bailing", LLM_CHAT_TEMPLATE_BAILING }, + { "llama4", LLM_CHAT_TEMPLATE_LLAMA4 }, }; llm_chat_template llm_chat_template_from_str(const std::string & name) { @@ -167,6 +171,12 @@ llm_chat_template llm_chat_detect_template(const std::string & tmpl) { return LLM_CHAT_TEMPLATE_GIGACHAT; } else if (tmpl_contains("<|role_start|>")) { return LLM_CHAT_TEMPLATE_MEGREZ; + } else if (tmpl_contains(" Ассистент:")) { + return LLM_CHAT_TEMPLATE_YANDEX; + } else if (tmpl_contains("ASSISTANT") && tmpl_contains("'HUMAN'")) { + return LLM_CHAT_TEMPLATE_BAILING; + } else if (tmpl_contains("<|header_start|>") && tmpl_contains("<|header_end|>")) { + return LLM_CHAT_TEMPLATE_LLAMA4; } return LLM_CHAT_TEMPLATE_UNKNOWN; } @@ -566,7 +576,51 @@ int32_t llm_chat_apply_template( if (add_ass) { ss << "<|role_start|>assistant<|role_end|>"; } - } else { + } else if (tmpl == LLM_CHAT_TEMPLATE_YANDEX) { + // Yandex template ("\n\n" is defined as EOT token) + + ss << ""; + + for (size_t i = 0; i < chat.size(); i++) { + std::string role(chat[i]->role); + if (role == "user") { + ss << " Пользователь: " << chat[i]->content << "\n\n"; + } else if (role == "assistant") { + ss << " Ассистент: " << chat[i]->content << "\n\n"; + } + } + + // Add generation prompt if needed + if (add_ass) { + ss << " Ассистент:[SEP]"; + } + } else if (tmpl == LLM_CHAT_TEMPLATE_BAILING) { + // Bailing (Ling) template + for (auto message : chat) { + std::string role(message->role); + + if (role == "user") { + role = "HUMAN"; + } else { + std::transform(role.begin(), role.end(), role.begin(), ::toupper); + } + + ss << "" << role << "" << message->content; + } + + if (add_ass) { + ss << "ASSISTANT"; + } + } else if (tmpl == LLM_CHAT_TEMPLATE_LLAMA4) { + // Llama 4 + for (auto message : chat) { + std::string role(message->role); + ss << "<|header_start|>" << role << "<|header_end|>\n\n" << trim(message->content) << "<|eot|>"; + } + if (add_ass) { + ss << "<|header_start|>assistant<|header_end|>\n\n"; + } + } else { // template not supported return -1; } @@ -584,4 +638,3 @@ int32_t llama_chat_builtin_templates(const char ** output, size_t len) { } return (int32_t) LLM_CHAT_TEMPLATES.size(); } - diff --git a/src/llama-chat.h b/src/llama-chat.h index 2f6a0e3e2..34537ca21 100644 --- a/src/llama-chat.h +++ b/src/llama-chat.h @@ -38,6 +38,9 @@ enum llm_chat_template { LLM_CHAT_TEMPLATE_GRANITE, LLM_CHAT_TEMPLATE_GIGACHAT, LLM_CHAT_TEMPLATE_MEGREZ, + LLM_CHAT_TEMPLATE_YANDEX, + LLM_CHAT_TEMPLATE_BAILING, + LLM_CHAT_TEMPLATE_LLAMA4, LLM_CHAT_TEMPLATE_UNKNOWN, }; diff --git a/src/llama-context.cpp b/src/llama-context.cpp index 671d2a81a..4735e98ea 100644 --- a/src/llama-context.cpp +++ b/src/llama-context.cpp @@ -1,551 +1,1562 @@ #include "llama-context.h" #include "llama-impl.h" +#include "llama-io.h" #include "llama-mmap.h" +#include "llama-model.h" +#include "llama-kv-cache.h" #include -#include #include #include +#include -void llama_set_k_shift(struct llama_context & lctx) { - const int64_t kv_size = lctx.kv_self.size; +// +// llama_context +// - assert(ggml_backend_buffer_is_host(lctx.inp_K_shift->buffer)); +llama_context::llama_context( + const llama_model & model, + llama_context_params params) : + model(model) { + LLAMA_LOG_INFO("%s: constructing llama_context\n", __func__); - int32_t * data = (int32_t *) lctx.inp_K_shift->data; + t_start_us = model.t_start_us; + t_load_us = model.t_load_us; - for (int i = 0; i < kv_size; ++i) { - data[i] = lctx.kv_self.cells[i].delta; - } -} + const auto & hparams = model.hparams; -void llama_set_s_copy(struct llama_context & lctx) { - const int64_t kv_size = lctx.kv_self.size; + cparams.n_seq_max = std::max(1u, params.n_seq_max); + cparams.n_threads = params.n_threads; + cparams.n_threads_batch = params.n_threads_batch; + cparams.yarn_ext_factor = params.yarn_ext_factor; + cparams.yarn_attn_factor = params.yarn_attn_factor; + cparams.yarn_beta_fast = params.yarn_beta_fast; + cparams.yarn_beta_slow = params.yarn_beta_slow; + cparams.defrag_thold = params.defrag_thold; + cparams.embeddings = params.embeddings; + cparams.offload_kqv = params.offload_kqv; + cparams.flash_attn = params.flash_attn; + cparams.no_perf = params.no_perf; + cparams.pooling_type = params.pooling_type; + cparams.warmup = false; - assert(ggml_backend_buffer_is_host(lctx.inp_s_copy->buffer)); + cparams.n_ctx = params.n_ctx == 0 ? hparams.n_ctx_train : params.n_ctx; + cparams.rope_freq_base = params.rope_freq_base == 0.0f ? hparams.rope_freq_base_train : params.rope_freq_base; + cparams.rope_freq_scale = params.rope_freq_scale == 0.0f ? hparams.rope_freq_scale_train : params.rope_freq_scale; - int32_t * data = (int32_t *) lctx.inp_s_copy->data; + cparams.n_ctx_orig_yarn = params.yarn_orig_ctx != 0 ? params.yarn_orig_ctx : + hparams.n_ctx_orig_yarn != 0 ? hparams.n_ctx_orig_yarn : + hparams.n_ctx_train; - for (int i = 0; i < kv_size; ++i) { - data[i] = lctx.kv_self.cells[i].src; - } -} + cparams.cb_eval = params.cb_eval; + cparams.cb_eval_user_data = params.cb_eval_user_data; -// llama input - -static int32_t llama_relative_position_bucket(llama_pos x, llama_pos y, uint64_t n_buckets, bool bidirectional) { - // TODO move to hparams if a T5 variant appears that uses a different value - const int64_t max_distance = 128; - - if (bidirectional) { - n_buckets >>= 1; + auto rope_scaling_type = params.rope_scaling_type; + if (rope_scaling_type == LLAMA_ROPE_SCALING_TYPE_UNSPECIFIED) { + rope_scaling_type = hparams.rope_scaling_type_train; } - const int64_t max_exact = n_buckets >> 1; + if (rope_scaling_type == LLAMA_ROPE_SCALING_TYPE_NONE) { + cparams.rope_freq_scale = 1.0f; // never scale if scaling type is none + } - int32_t relative_position = x - y; - int32_t relative_bucket = 0; - if (bidirectional) { - relative_bucket += (relative_position > 0) * n_buckets; - relative_position = abs(relative_position); + if (cparams.yarn_ext_factor < 0.0f) { // negative indicates 'not set' + cparams.yarn_ext_factor = rope_scaling_type == LLAMA_ROPE_SCALING_TYPE_YARN ? 1.0f : 0.0f; + } + + cparams.yarn_attn_factor *= hparams.rope_attn_factor; + + if (cparams.pooling_type == LLAMA_POOLING_TYPE_UNSPECIFIED) { + if (hparams.pooling_type == LLAMA_POOLING_TYPE_UNSPECIFIED) { + cparams.pooling_type = LLAMA_POOLING_TYPE_NONE; + } else { + cparams.pooling_type = hparams.pooling_type; + } + } + + if (params.attention_type == LLAMA_ATTENTION_TYPE_UNSPECIFIED) { + cparams.causal_attn = hparams.causal_attn; } else { - relative_position = -std::min(relative_position, 0); + cparams.causal_attn = params.attention_type == LLAMA_ATTENTION_TYPE_CAUSAL; + } + + // with causal attention, the batch size is limited by the context size + cparams.n_batch = cparams.causal_attn ? std::min(cparams.n_ctx, params.n_batch) : params.n_batch; + + // the batch has to be at least GGML_KQ_MASK_PAD because we will be padding the KQ_mask + // this is required by GPU kernels in order to avoid out-of-bounds accesses (e.g. ggml_flash_attn_ext) + // ref: https://github.com/ggerganov/llama.cpp/pull/5021 + // TODO: this padding is not needed for the cache-less context so we should probably move it to llama_context_kv_self + if (cparams.n_batch < GGML_KQ_MASK_PAD) { + LLAMA_LOG_WARN("%s: n_batch is less than GGML_KQ_MASK_PAD - increasing to %d\n", __func__, GGML_KQ_MASK_PAD); + cparams.n_batch = GGML_KQ_MASK_PAD; + } + + cparams.n_ubatch = std::min(cparams.n_batch, params.n_ubatch == 0 ? params.n_batch : params.n_ubatch); + + const uint32_t n_ctx_per_seq = cparams.n_ctx / cparams.n_seq_max; + + LLAMA_LOG_INFO("%s: n_seq_max = %u\n", __func__, cparams.n_seq_max); + LLAMA_LOG_INFO("%s: n_ctx = %u\n", __func__, cparams.n_ctx); + LLAMA_LOG_INFO("%s: n_ctx_per_seq = %u\n", __func__, n_ctx_per_seq); + LLAMA_LOG_INFO("%s: n_batch = %u\n", __func__, cparams.n_batch); + LLAMA_LOG_INFO("%s: n_ubatch = %u\n", __func__, cparams.n_ubatch); + LLAMA_LOG_INFO("%s: causal_attn = %d\n", __func__, cparams.causal_attn); + LLAMA_LOG_INFO("%s: flash_attn = %d\n", __func__, cparams.flash_attn); + LLAMA_LOG_INFO("%s: freq_base = %.1f\n", __func__, cparams.rope_freq_base); + LLAMA_LOG_INFO("%s: freq_scale = %g\n", __func__, cparams.rope_freq_scale); + + if (n_ctx_per_seq < hparams.n_ctx_train) { + LLAMA_LOG_WARN("%s: n_ctx_per_seq (%u) < n_ctx_train (%u) -- the full capacity of the model will not be utilized\n", + __func__, n_ctx_per_seq, hparams.n_ctx_train); + } + + if (n_ctx_per_seq > hparams.n_ctx_train) { + LLAMA_LOG_WARN("%s: n_ctx_pre_seq (%u) > n_ctx_train (%u) -- possible training context overflow\n", + __func__, n_ctx_per_seq, hparams.n_ctx_train); + } + + logits_all = params.logits_all; + + if (!hparams.vocab_only) { + // GPU backends + for (auto * dev : model.devices) { + ggml_backend_t backend = ggml_backend_dev_init(dev, nullptr); + if (backend == nullptr) { + throw std::runtime_error(format("failed to initialize %s backend", ggml_backend_dev_name(dev))); + } + backends.emplace_back(backend); + } + + // add ACCEL backends (such as BLAS) + for (size_t i = 0; i < ggml_backend_dev_count(); ++i) { + ggml_backend_dev_t dev = ggml_backend_dev_get(i); + if (ggml_backend_dev_type(dev) == GGML_BACKEND_DEVICE_TYPE_ACCEL) { + ggml_backend_t backend = ggml_backend_dev_init(dev, nullptr); + if (backend == nullptr) { + throw std::runtime_error(format("failed to initialize %s backend", ggml_backend_dev_name(dev))); + } + backends.emplace_back(backend); + } + } + + // add CPU backend + backend_cpu = ggml_backend_init_by_type(GGML_BACKEND_DEVICE_TYPE_CPU, nullptr); + if (backend_cpu == nullptr) { + throw std::runtime_error("failed to initialize CPU backend"); + } + backends.emplace_back(backend_cpu); + + // create a list of the set_n_threads functions in the backends + for (auto & backend : backends) { + ggml_backend_dev_t dev = ggml_backend_get_device(backend.get()); + ggml_backend_reg_t reg = dev ? ggml_backend_dev_backend_reg(dev) : nullptr; + if (reg) { + auto ggml_backend_set_n_threads_fn = (ggml_backend_set_n_threads_t) ggml_backend_reg_get_proc_address(reg, "ggml_backend_set_n_threads"); + if (ggml_backend_set_n_threads_fn) { + set_n_threads_fns.emplace_back(backend.get(), ggml_backend_set_n_threads_fn); + } + } + } + + llama_set_abort_callback(this, params.abort_callback, params.abort_callback_data); + + // graph outputs buffer + { + // resized during inference when a batch uses more outputs + if ((uint32_t) output_reserve(params.n_seq_max) < params.n_seq_max) { + throw std::runtime_error("failed to reserve initial output buffer"); + } + + LLAMA_LOG_INFO("%s: %10s output buffer size = %8.2f MiB\n", __func__, + ggml_backend_buffer_name (buf_output.get()), + ggml_backend_buffer_get_size(buf_output.get()) / 1024.0 / 1024.0); + } + } + + // init the memory module + // TODO: for now, always create a unified KV cache + if (!hparams.vocab_only) { + kv_self.reset(static_cast(model.create_memory())); + + LLAMA_LOG_DEBUG("%s: n_ctx = %u\n", __func__, cparams.n_ctx); + + cparams.n_ctx = GGML_PAD(cparams.n_ctx, kv_self->get_padding(cparams)); + + LLAMA_LOG_DEBUG("%s: n_ctx = %u (padded)\n", __func__, cparams.n_ctx); + + uint32_t kv_size = cparams.n_ctx; + ggml_type type_k = params.type_k; + ggml_type type_v = params.type_v; + + if (llama_model_is_recurrent(&model)) { + // Mamba needs at least as many KV cells as there are sequences kept at any time + kv_size = std::max((uint32_t) 1, params.n_seq_max); + // it's probably best to keep as much precision as possible for the states + type_k = GGML_TYPE_F32; // required by ggml_ssm_conv for Mamba's conv_states + type_v = GGML_TYPE_F32; // required by ggml_ssm_scan for Mamba's ssm_states + } + + GGML_ASSERT(hparams.n_embd_head_k % ggml_blck_size(type_k) == 0); + GGML_ASSERT(hparams.n_embd_head_v % ggml_blck_size(type_v) == 0); + + if (!kv_self->init(model, cparams, type_k, type_v, kv_size, cparams.offload_kqv)) { + throw std::runtime_error("failed to initialize self-attention cache"); + } + + { + const size_t memory_size_k = kv_self->size_k_bytes(); + const size_t memory_size_v = kv_self->size_v_bytes(); + + LLAMA_LOG_INFO("%s: KV self size = %7.2f MiB, K (%s): %7.2f MiB, V (%s): %7.2f MiB\n", __func__, + (float)(memory_size_k + memory_size_v) / (1024.0f * 1024.0f), + ggml_type_name(type_k), (float)memory_size_k / (1024.0f * 1024.0f), + ggml_type_name(type_v), (float)memory_size_v / (1024.0f * 1024.0f)); + } + } + + // init backends + if (!hparams.vocab_only) { + LLAMA_LOG_DEBUG("%s: enumerating backends\n", __func__); + + backend_buft.clear(); + backend_ptrs.clear(); + + for (auto & backend : backends) { + auto * buft = ggml_backend_get_default_buffer_type(backend.get()); + auto backend_type = ggml_backend_dev_type(ggml_backend_get_device(backend.get())); + + if (backend_type == GGML_BACKEND_DEVICE_TYPE_CPU && !model.devices.empty()) { + // use the host buffer of the first device CPU for faster transfer of the intermediate state + auto * dev = model.devices[0]; + auto * host_buft = ggml_backend_dev_host_buffer_type(dev); + if (host_buft) { + buft = host_buft; + } + } + + backend_buft.push_back(buft); + backend_ptrs.push_back(backend.get()); + } + + LLAMA_LOG_DEBUG("%s: backend_ptrs.size() = %zu\n", __func__, backend_ptrs.size()); + + const size_t max_nodes = this->graph_max_nodes(); + + LLAMA_LOG_DEBUG("%s: max_nodes = %zu\n", __func__, max_nodes); + + // buffer used to store the computation graph and the tensor meta data + buf_compute_meta.resize(ggml_tensor_overhead()*max_nodes + ggml_graph_overhead_custom(max_nodes, false)); + + // TODO: move these checks to ggml_backend_sched + // enabling pipeline parallelism in the scheduler increases memory usage, so it is only done when necessary + bool pipeline_parallel = + model.n_devices() > 1 && + model.params.n_gpu_layers > (int) model.hparams.n_layer && + model.params.split_mode == LLAMA_SPLIT_MODE_LAYER && + cparams.offload_kqv && + !model.has_tensor_overrides(); + + // pipeline parallelism requires support for async compute and events in all devices + if (pipeline_parallel) { + for (auto & backend : backends) { + auto dev_type = ggml_backend_dev_type(ggml_backend_get_device(backend.get())); + if (dev_type == GGML_BACKEND_DEVICE_TYPE_CPU) { + // ignore CPU backend + continue; + } + auto * dev = ggml_backend_get_device(backend.get()); + ggml_backend_dev_props props; + ggml_backend_dev_get_props(dev, &props); + if (!props.caps.async || !props.caps.events) { + // device does not support async compute or events + pipeline_parallel = false; + break; + } + } + } + + sched.reset(ggml_backend_sched_new(backend_ptrs.data(), backend_buft.data(), backend_ptrs.size(), max_nodes, pipeline_parallel)); + + if (pipeline_parallel) { + LLAMA_LOG_INFO("%s: pipeline parallelism enabled (n_copies=%d)\n", __func__, ggml_backend_sched_get_n_copies(sched.get())); + } + } + + // reserve worst-case graph + if (!hparams.vocab_only) { + const uint32_t n_seqs = 1; // TODO: worst-case number of sequences + const uint32_t n_tokens = std::min(cparams.n_ctx, cparams.n_ubatch); + + llama_token token = model.vocab.token_bos(); // not actually used by llama_build_graph, but required to choose between token and embedding inputs graph + + // restore later + // TODO: something cleaner + const auto n_outputs_save = n_outputs; + + LLAMA_LOG_DEBUG("%s: worst-case: n_tokens = %d, n_seqs = %d, n_outputs = %d\n", __func__, n_tokens, n_seqs, n_outputs); + + int n_splits_pp = -1; + int n_nodes_pp = -1; + + int n_splits_tg = -1; + int n_nodes_tg = -1; + + // simulate full KV cache + kv_self->n = kv_self->size; + + cross.v_embd.clear(); + + // reserve pp graph first so that buffers are only allocated once + { + llama_ubatch ubatch_pp = { true, n_tokens, n_tokens / n_seqs, n_seqs, &token, nullptr, nullptr, nullptr, nullptr, nullptr}; + + // max number of outputs + n_outputs = ubatch_pp.n_tokens; + + LLAMA_LOG_DEBUG("%s: reserving graph for n_tokens = %d, n_seqs = %d\n", __func__, ubatch_pp.n_tokens, ubatch_pp.n_seqs); + + auto * gf = graph_init(); + graph_build(ctx_compute.get(), gf, ubatch_pp, LLM_GRAPH_TYPE_DEFAULT); + + if (!ggml_backend_sched_reserve(sched.get(), gf)) { + throw std::runtime_error("failed to allocate compute pp buffers"); + } + + n_splits_pp = ggml_backend_sched_get_n_splits(sched.get()); + n_nodes_pp = ggml_graph_n_nodes(gf); + } + + // reserve with tg graph to get the number of splits and nodes + { + llama_ubatch ubatch_tg = { true, 1, 1, n_seqs, &token, nullptr, nullptr, nullptr, nullptr, nullptr}; + + n_outputs = ubatch_tg.n_tokens; + + LLAMA_LOG_DEBUG("%s: reserving graph for n_tokens = %d, n_seqs = %d\n", __func__, ubatch_tg.n_tokens, ubatch_tg.n_seqs); + + auto * gf = graph_init(); + graph_build(ctx_compute.get(), gf, ubatch_tg, LLM_GRAPH_TYPE_DEFAULT); + + if (!ggml_backend_sched_reserve(sched.get(), gf)) { + throw std::runtime_error("failed to allocate compute tg buffers"); + } + + n_splits_tg = ggml_backend_sched_get_n_splits(sched.get()); + n_nodes_tg = ggml_graph_n_nodes(gf); + } + + // reserve again with pp graph to avoid ggml-alloc reallocations during inference + { + llama_ubatch ubatch_pp = { true, n_tokens, n_tokens / n_seqs, n_seqs, &token, nullptr, nullptr, nullptr, nullptr, nullptr}; + + n_outputs = ubatch_pp.n_tokens; + + LLAMA_LOG_DEBUG("%s: reserving graph for n_tokens = %d, n_seqs = %d\n", __func__, ubatch_pp.n_tokens, ubatch_pp.n_seqs); + + auto * gf = graph_init(); + graph_build(ctx_compute.get(), gf, ubatch_pp, LLM_GRAPH_TYPE_DEFAULT); + + if (!ggml_backend_sched_reserve(sched.get(), gf)) { + throw std::runtime_error("failed to allocate compute pp buffers"); + } + } + + n_outputs = n_outputs_save; + + for (size_t i = 0; i < backend_ptrs.size(); ++i) { + ggml_backend_t backend = backend_ptrs[i]; + ggml_backend_buffer_type_t buft = backend_buft[i]; + size_t size = ggml_backend_sched_get_buffer_size(sched.get(), backend); + if (size > 1) { + LLAMA_LOG_INFO("%s: %10s compute buffer size = %8.2f MiB\n", __func__, + ggml_backend_buft_name(buft), + size / 1024.0 / 1024.0); + } + } + + if (n_nodes_pp == n_nodes_tg) { + LLAMA_LOG_INFO("%s: graph nodes = %d\n", __func__, n_nodes_pp); + } else { + LLAMA_LOG_INFO("%s: graph nodes = %d (with bs=%d), %d (with bs=1)\n", __func__, n_nodes_pp, n_tokens, n_nodes_tg); + } + + if (n_splits_pp == n_splits_tg) { + LLAMA_LOG_INFO("%s: graph splits = %d\n", __func__, n_splits_pp); + } else { + LLAMA_LOG_INFO("%s: graph splits = %d (with bs=%d), %d (with bs=1)\n", __func__, n_splits_pp, n_tokens, n_splits_tg); + } } - int32_t relative_position_if_large = floorf(max_exact + logf(1.0 * relative_position / max_exact) * (n_buckets - max_exact) / log(1.0 * max_distance / max_exact)); - relative_position_if_large = std::min(relative_position_if_large, n_buckets - 1); - relative_bucket += (relative_position < max_exact ? relative_position : relative_position_if_large); - return relative_bucket; } -void llama_set_inputs(llama_context & lctx, const llama_ubatch & ubatch) { +llama_context::~llama_context() = default; + +void llama_context::synchronize() { + ggml_backend_sched_synchronize(sched.get()); + + // FIXME: if multiple single tokens are evaluated without a synchronization, + // the stats will be added to the prompt evaluation stats + // this should only happen when using batch size 1 to evaluate a batch + + // add the evaluation to the stats + if (n_queued_tokens == 1) { + if (!cparams.no_perf) { + t_eval_us += ggml_time_us() - t_compute_start_us; + } + n_eval++; + } else if (n_queued_tokens > 1) { + if (!cparams.no_perf) { + t_p_eval_us += ggml_time_us() - t_compute_start_us; + } + n_p_eval += n_queued_tokens; + } + + // get a more accurate load time, upon first eval + if (n_queued_tokens > 0 && !has_evaluated_once) { + t_load_us = ggml_time_us() - t_start_us; + has_evaluated_once = true; + } + + n_queued_tokens = 0; + t_compute_start_us = 0; +} + +const llama_model & llama_context::get_model() const { + return model; +} + +uint32_t llama_context::n_ctx() const { + return cparams.n_ctx; +} + +uint32_t llama_context::n_ctx_per_seq() const { + return cparams.n_ctx / cparams.n_seq_max; +} + +uint32_t llama_context::n_batch() const { + return cparams.n_batch; +} + +uint32_t llama_context::n_ubatch() const { + return cparams.n_ubatch; +} + +uint32_t llama_context::n_seq_max() const { + return cparams.n_seq_max; +} + +uint32_t llama_context::n_threads() const { + return cparams.n_threads; +} + +uint32_t llama_context::n_threads_batch() const { + return cparams.n_threads_batch; +} + +llama_kv_cache * llama_context::get_kv_self() { + return kv_self.get(); +} + +const llama_kv_cache * llama_context::get_kv_self() const { + return kv_self.get(); +} + +ggml_tensor * llama_context::build_rope_shift( + ggml_context * ctx0, + ggml_tensor * cur, + ggml_tensor * shift, + ggml_tensor * factors, + float freq_base, + float freq_scale, + ggml_backend_buffer * bbuf) const { + const auto & n_ctx_orig = cparams.n_ctx_orig_yarn; + + const auto & yarn_ext_factor = cparams.yarn_ext_factor; + const auto & yarn_attn_factor = cparams.yarn_attn_factor; + const auto & yarn_beta_fast = cparams.yarn_beta_fast; + const auto & yarn_beta_slow = cparams.yarn_beta_slow; + + const auto & hparams = model.hparams; + + const auto & n_rot = hparams.n_rot; + const auto & rope_type = hparams.rope_type; + + ggml_tensor * tmp; + + if (ggml_is_quantized(cur->type)) { + // dequantize to f32 -> RoPE -> quantize back + tmp = ggml_cast(ctx0, cur, GGML_TYPE_F32); + + if (bbuf) { + for (const auto & backend : backends) { + // Figure out which backend KV cache belongs to + if (ggml_backend_supports_buft(backend.get(), ggml_backend_buffer_get_type(bbuf))) { + ggml_backend_sched_set_tensor_backend(sched.get(), tmp, backend.get()); + break; + } + } + } + + tmp = ggml_rope_ext_inplace(ctx0, tmp, + shift, factors, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + yarn_ext_factor, yarn_attn_factor, yarn_beta_fast, yarn_beta_slow); + + tmp = ggml_cpy(ctx0, tmp, cur); + } else { + // we rotate only the first n_rot dimensions + tmp = ggml_rope_ext_inplace(ctx0, cur, + shift, factors, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + yarn_ext_factor, yarn_attn_factor, yarn_beta_fast, yarn_beta_slow); + } + + return tmp; +} + +class llm_graph_input_k_shift : public llm_graph_input_i { +public: + llm_graph_input_k_shift(const llama_kv_cache_unified * kv_self) : kv_self(kv_self) {} + virtual ~llm_graph_input_k_shift() = default; + + void set_input(const llama_ubatch * ubatch) override; + + ggml_tensor * k_shift; // I32 [kv_size] + + const llama_kv_cache_unified * kv_self; +}; + +void llm_graph_input_k_shift::set_input(const llama_ubatch * ubatch) { + GGML_UNUSED(ubatch); + + if (k_shift) { + assert(ggml_backend_buffer_is_host(k_shift->buffer)); + + int32_t * data = (int32_t *) k_shift->data; + + for (uint32_t i = 0; i < kv_self->size; ++i) { + data[i] = kv_self->cells[i].delta; + } + } +} + +llm_graph_result_ptr llama_context::build_kv_self_shift( + ggml_context * ctx0, + ggml_cgraph * gf) const { + auto res = std::make_unique(); + + const auto & hparams = model.hparams; + + const auto & n_layer = hparams.n_layer; + + const auto & n_embd_head_k = hparams.n_embd_head_k; + //const auto & n_embd_head_v = hparams.n_embd_head_v; + + //GGML_ASSERT(kv_self->size == n_ctx); + + auto inp = std::make_unique(kv_self.get()); + + inp->k_shift = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, cparams.n_ctx); + ggml_set_input(inp->k_shift); + + for (uint32_t il = 0; il < n_layer; ++il) { + const int64_t n_head_kv = hparams.n_head_kv(il); + const int64_t n_embd_k_gqa = hparams.n_embd_k_gqa(il); + + const bool is_swa = hparams.is_swa(il); + + // note: the swa rope params could become part of the cparams in the future + // if we decide to make them configurable, like the non-sliding ones + const float freq_base_l = is_swa ? hparams.rope_freq_base_train_swa : cparams.rope_freq_base; + const float freq_scale_l = is_swa ? hparams.rope_freq_scale_train_swa : cparams.rope_freq_scale; + + ggml_tensor * rope_factors = kv_self->cbs.get_rope_factors(n_ctx_per_seq(), il); + + ggml_tensor * k = + ggml_view_3d(ctx0, kv_self->k_l[il], + n_embd_head_k, n_head_kv, kv_self->size, + ggml_row_size(kv_self->k_l[il]->type, n_embd_head_k), + ggml_row_size(kv_self->k_l[il]->type, n_embd_k_gqa), + 0); + + ggml_tensor * cur = build_rope_shift(ctx0, k, inp->k_shift, rope_factors, freq_base_l, freq_scale_l, kv_self->k_l[il]->buffer); + + ggml_build_forward_expand(gf, cur); + } + + res->add_input(std::move(inp)); + + return res; +} + +llm_graph_result_ptr llama_context::build_kv_self_defrag( + ggml_context * ctx0, + ggml_cgraph * gf) const { + auto res = std::make_unique(); + + const auto & hparams = model.hparams; + + const auto & ids = kv_self->defrag_info.ids; + +#if 0 + // CPU defrag // - // set input data + // TODO: optimizations are possible: + // - multiple threads + // - avoid copying to the host memory when already there + // + // likely not worth the effort, as we have ggml_graph based defrag // - const auto & hparams = lctx.model.hparams; - const auto & cparams = lctx.cparams; - const auto & kv_self = lctx.kv_self; + const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa(); + const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(); - if (ubatch.token) { - const int64_t n_tokens = ubatch.n_tokens; + const uint32_t kv_size = size; - ggml_backend_tensor_set(lctx.inp_tokens, ubatch.token, 0, n_tokens*ggml_element_size(lctx.inp_tokens)); - } + std::vector buf_k; + std::vector buf_v; - if (ubatch.embd) { - const int64_t n_embd = hparams.n_embd; - const int64_t n_tokens = ubatch.n_tokens; + for (uint32_t il = 0; il < n_layer; ++il) { + const size_t k_size_row = ggml_row_size(k_l[il]->type, n_embd_k_gqa); + const size_t k_size = ggml_row_size(k_l[il]->type, n_embd_k_gqa*kv_size); - ggml_backend_tensor_set(lctx.inp_embd, ubatch.embd, 0, n_tokens*n_embd*ggml_element_size(lctx.inp_embd)); - } + const size_t v_size_el = ggml_type_size(v_l[il]->type); + const size_t v_size = ggml_row_size (v_l[il]->type, n_embd_v_gqa*kv_size); - if (ubatch.pos && lctx.inp_pos) { - const int64_t n_tokens = ubatch.n_tokens; - auto n_pos = lctx.n_pos_per_token; - ggml_backend_tensor_set(lctx.inp_pos, ubatch.pos, 0, n_tokens*n_pos*ggml_element_size(lctx.inp_pos)); - } + buf_k.resize(k_size); + buf_v.resize(v_size); - if (hparams.causal_attn || cparams.pooling_type == LLAMA_POOLING_TYPE_NONE) { - //GGML_ASSERT(lctx.inp_out_ids && "every model that can must skip unused outputs"); + ggml_backend_tensor_get(k_l[il], buf_k.data(), 0, buf_k.size()); + ggml_backend_tensor_get(v_l[il], buf_v.data(), 0, buf_v.size()); - if (!lctx.inp_out_ids) { - LLAMA_LOG_WARN("%s: 'lctx.inp_out_ids' is not created\n", __func__); - } else { - const int64_t n_tokens = ubatch.n_tokens; + // batch move [i, i+nm) to [id, id+nm) + // note: cells can move only to a lower index + for (uint32_t i = 0; i < n_kv; ++i) { + const uint32_t id = ids[i]; - GGML_ASSERT(ggml_backend_buffer_is_host(lctx.inp_out_ids->buffer)); - int32_t * data = (int32_t *) lctx.inp_out_ids->data; + if (i == id || id == n_kv) { + continue; + } - if (lctx.n_outputs == n_tokens) { - for (int i = 0; i < n_tokens; ++i) { - data[i] = i; + uint32_t nm = 1; + + while (i + nm < n_kv && ids[i + nm] == id + nm) { + nm++; + } + + // move keys + { + const int64_t os = i*k_size_row; + const int64_t od = id*k_size_row; + + memcpy(buf_k.data() + od, buf_k.data() + os, nm*k_size_row); + } + + // move values (note: they are transposed) + { + const int64_t os = i; + const int64_t od = id; + + for (uint32_t j = 0; j < n_embd_v_gqa; ++j) { + memcpy(buf_v.data() + (od + j*kv_size)*v_size_el, buf_v.data() + (os + j*kv_size)*v_size_el, nm*v_size_el); } - } else if (ubatch.output) { - int32_t n_outputs = 0; - for (int i = 0; i < n_tokens; ++i) { - if (ubatch.output[i]) { - data[n_outputs++] = i; - } - } - // the graph needs to have been passed the correct number of outputs - GGML_ASSERT(lctx.n_outputs == n_outputs); - } else if (lctx.n_outputs == 1) { - // only keep last output - data[0] = n_tokens - 1; + } + + i += nm - 1; + } + + ggml_backend_tensor_set(k_l[il], buf_k.data(), 0, buf_k.size()); + ggml_backend_tensor_set(v_l[il], buf_v.data(), 0, buf_v.size()); + } +#else + for (uint32_t i = 0; i < ids.size(); ++i) { + const uint32_t id = ids[i]; + + if (i == id || id == ids.size()) { + continue; + } + + uint32_t nm = 1; + + while (i + nm < ids.size() && ids[i + nm] == id + nm) { + nm++; + } + + for (uint32_t il = 0; il < hparams.n_layer; ++il) { // NOLINT + const int64_t n_embd_k_gqa = hparams.n_embd_k_gqa(il); + const int64_t n_embd_v_gqa = hparams.n_embd_v_gqa(il); + + ggml_tensor * view_k_src = ggml_view_2d(ctx0, kv_self->k_l[il], + n_embd_k_gqa, nm, + ggml_row_size(kv_self->k_l[il]->type, n_embd_k_gqa), + ggml_row_size(kv_self->k_l[il]->type, n_embd_k_gqa*i)); + + ggml_tensor * view_k_dst = ggml_view_2d(ctx0, kv_self->k_l[il], + n_embd_k_gqa, nm, + ggml_row_size(kv_self->k_l[il]->type, n_embd_k_gqa), + ggml_row_size(kv_self->k_l[il]->type, n_embd_k_gqa*id)); + + ggml_tensor * view_v_src; + ggml_tensor * view_v_dst; + + if (cparams.flash_attn) { + // NOTE: the V cache is not transposed when using flash attention + view_v_src = ggml_view_2d(ctx0, kv_self->v_l[il], + n_embd_v_gqa, nm, + ggml_row_size(kv_self->v_l[il]->type, n_embd_v_gqa), + ggml_row_size(kv_self->v_l[il]->type, n_embd_v_gqa*i)); + + view_v_dst = ggml_view_2d(ctx0, kv_self->v_l[il], + n_embd_v_gqa, nm, + ggml_row_size(kv_self->v_l[il]->type, n_embd_v_gqa), + ggml_row_size(kv_self->v_l[il]->type, n_embd_v_gqa*id)); } else { - GGML_ASSERT(lctx.n_outputs == 0); + view_v_src = ggml_view_2d(ctx0, kv_self->v_l[il], + nm, n_embd_v_gqa, + ggml_row_size(kv_self->v_l[il]->type, kv_self->size), + ggml_row_size(kv_self->v_l[il]->type, i)); + + view_v_dst = ggml_view_2d(ctx0, kv_self->v_l[il], + nm, n_embd_v_gqa, + ggml_row_size(kv_self->v_l[il]->type, kv_self->size), + ggml_row_size(kv_self->v_l[il]->type, id)); + } + + ggml_build_forward_expand(gf, ggml_cpy(ctx0, view_k_src, view_k_dst)); + ggml_build_forward_expand(gf, ggml_cpy(ctx0, view_v_src, view_v_dst)); + } + + i += nm - 1; + } + + //LLAMA_LOG_INFO("gf->n_nodes = %d\n", gf->n_nodes); +#endif + + return res; +} + +void llama_context::kv_self_update() { + auto & kv = kv_self; + + bool need_reserve = false; + + if (kv->has_shift) { + if (!kv->get_can_shift()) { + GGML_ABORT("The current context does not support K-shift"); + } + + LLAMA_LOG_DEBUG("%s: applying K-shift\n", __func__); + + // apply K-shift if needed + if (model.hparams.rope_type != LLAMA_ROPE_TYPE_NONE) { + ggml_backend_sched_reset(sched.get()); + + auto * gf = graph_init(); + + auto res = build_kv_self_shift(ctx_compute.get(), gf); + + ggml_backend_sched_alloc_graph(sched.get(), gf); + + res->set_inputs(nullptr); + + graph_compute(gf, false); + + need_reserve = true; + } + + { + kv->has_shift = false; + + for (uint32_t i = 0; i < kv->size; ++i) { + kv->cells[i].delta = 0; } } } - GGML_ASSERT( - // (!a || b) is a logical implication (a -> b) - // !hparams.causal_attn -> !cparams.causal_attn - (hparams.causal_attn || !cparams.causal_attn) && - "causal attention is not supported by this model" - ); + // defragment the KV cache if needed + if (kv->do_defrag) { + LLAMA_LOG_DEBUG("%s: defragmenting KV cache\n", __func__); - if (lctx.inp_KQ_mask || lctx.inp_KQ_mask_swa) { - // NOTE: hparams.causal_attn indicates the model is capable of generation and uses the kv cache. - if (cparams.causal_attn && !lctx.is_encoding) { - const int64_t n_kv = kv_self.n; - const int64_t n_tokens = ubatch.n_tokens; - const int64_t n_seq_tokens = ubatch.n_seq_tokens; - const int64_t n_seqs = ubatch.n_seqs; + if (kv->defrag_prepare(graph_max_nodes())) { + ggml_backend_sched_reset(sched.get()); + auto * gf = graph_init(); - float * data = nullptr; - float * data_swa = nullptr; + auto res = build_kv_self_defrag(ctx_compute.get(), gf); - if (lctx.inp_KQ_mask) { - GGML_ASSERT(ggml_backend_buffer_is_host(lctx.inp_KQ_mask->buffer)); - data = (float *) lctx.inp_KQ_mask->data; - } + ggml_backend_sched_alloc_graph(sched.get(), gf); - if (lctx.inp_KQ_mask_swa) { - GGML_ASSERT(ggml_backend_buffer_is_host(lctx.inp_KQ_mask_swa->buffer)); - data_swa = (float *) lctx.inp_KQ_mask_swa->data; - } + res->set_inputs(nullptr); - // For causal attention, use only the previous KV cells - // of the correct sequence for each token of the ubatch. - // It's assumed that if a token in the batch has multiple sequences, they are equivalent. - for (int h = 0; h < 1; ++h) { - for (int s = 0; s < n_seqs; ++s) { - const llama_seq_id seq_id = ubatch.seq_id[s][0]; + graph_compute(gf, false); - for (int j = 0; j < n_seq_tokens; ++j) { - const llama_pos pos = ubatch.pos[s*n_seq_tokens + j]; - - for (int i = 0; i < n_kv; ++i) { - float f; - if (!kv_self.cells[i].has_seq_id(seq_id) || kv_self.cells[i].pos > pos) { - f = -INFINITY; - } else { - if (hparams.use_alibi) { - f = -std::abs(kv_self.cells[i].pos - pos); - } else { - f = 0.0f; - } - } - - if (data) { - data[h*(n_kv*n_tokens) + s*(n_kv*n_seq_tokens) + j*n_kv + i] = f; - } - - // may need to cut off old tokens for sliding window - if (data_swa) { - if (pos - kv_self.cells[i].pos >= (int32_t)hparams.n_swa) { - f = -INFINITY; - } - data_swa[h*(n_kv*n_tokens) + s*(n_kv*n_seq_tokens) + j*n_kv + i] = f; - } - } - } - } - - if (data) { - for (int i = n_tokens; i < GGML_PAD(n_tokens, GGML_KQ_MASK_PAD); ++i) { - for (int j = 0; j < n_kv; ++j) { - data[h*(n_kv*n_tokens) + i*n_kv + j] = -INFINITY; - } - } - } - - if (data_swa) { - for (int i = n_tokens; i < GGML_PAD(n_tokens, GGML_KQ_MASK_PAD); ++i) { - for (int j = 0; j < n_kv; ++j) { - data_swa[h*(n_kv*n_tokens) + i*n_kv + j] = -INFINITY; - } - } - } - } - } else { - const int64_t n_tokens = ubatch.n_tokens; - const int64_t n_seq_tokens = ubatch.n_seq_tokens; - const int64_t n_seqs = ubatch.n_seqs; - // when using kv cache, the mask needs to match the kv cache size - const int64_t n_stride = hparams.causal_attn && !lctx.is_encoding ? kv_self.n : n_tokens; - - GGML_ASSERT(ggml_backend_buffer_is_host(lctx.inp_KQ_mask->buffer)); - - float * data = (float *) lctx.inp_KQ_mask->data; - - for (int h = 0; h < 1; ++h) { - for (int s1 = 0; s1 < n_seqs; ++s1) { - const llama_seq_id seq_id = ubatch.seq_id[s1][0]; - - for (int j = 0; j < n_seq_tokens; ++j) { - const int32_t tj = s1*n_seq_tokens + j; - - for (int s0 = 0; s0 < n_seqs; ++s0) { - for (int i = 0; i < n_seq_tokens; ++i) { - const int32_t ti = s0*n_seq_tokens + i; - float f = -INFINITY; - - for (int s = 0; s < ubatch.n_seq_id[s0]; ++s) { - if (ubatch.seq_id[s0][s] == seq_id) { - if (hparams.use_alibi) { - f = -std::abs(ubatch.pos[ti] - ubatch.pos[tj]); - } else { - f = 0.0f; - } - break; - } - } - - data[h*(n_tokens*n_tokens) + tj*n_stride + ti] = f; - } - } - - for (int i = n_tokens; i < n_stride; ++i) { - data[h*(n_tokens*n_tokens) + tj*n_stride + i] = -INFINITY; - } - } - } - } + need_reserve = true; } + + kv->do_defrag = false; } - if (cparams.embeddings && cparams.pooling_type == LLAMA_POOLING_TYPE_MEAN) { - const int64_t n_tokens = ubatch.n_tokens; - const int64_t n_seq_tokens = ubatch.n_seq_tokens; - const int64_t n_seqs = ubatch.n_seqs; + // reserve a worst case graph if needed + if (need_reserve) { + LLAMA_LOG_DEBUG("%s: reserving a worst case graph\n", __func__); - GGML_ASSERT(lctx.inp_mean); - GGML_ASSERT(ggml_backend_buffer_is_host(lctx.inp_mean->buffer)); + // build worst-case graph + uint32_t n_seqs = 1; // TODO: worst-case number of sequences + uint32_t n_tokens = std::min(cparams.n_ctx, cparams.n_ubatch); - float * data = (float *) lctx.inp_mean->data; - memset(lctx.inp_mean->data, 0, n_tokens * n_tokens * ggml_element_size(lctx.inp_mean)); + // simulate full KV cache + kv_self->n = kv_self->size; - std::vector sum(n_tokens, 0); + llama_token token = model.vocab.token_bos(); // not actually used by llama_build_graph, but required to choose between token and embedding inputs graph + llama_ubatch ubatch = { true, n_tokens, n_tokens / n_seqs, n_seqs, &token, nullptr, nullptr, nullptr, nullptr, nullptr}; - for (int s = 0; s < n_seqs; ++s) { - const llama_seq_id seq_id = ubatch.seq_id[s][0]; + auto * gf = graph_init(); + graph_build(ctx_compute.get(), gf, ubatch, LLM_GRAPH_TYPE_DEFAULT); - // TODO: adapt limits to n_seqs when ubatch.equal_seqs is true - GGML_ASSERT(seq_id < n_tokens && "seq_id cannot be larger than n_tokens with pooling_type == MEAN"); - - sum[seq_id] += ubatch.n_seq_tokens; - } - - std::vector div(n_tokens, 0.0f); - for (int i = 0; i < n_tokens; ++i) { - const uint64_t s = sum[i]; - if (s > 0) { - div[i] = 1.0f/float(s); - } - } - - for (int s = 0; s < n_seqs; ++s) { - const llama_seq_id seq_id = ubatch.seq_id[s][0]; - - for (int i = 0; i < n_seq_tokens; ++i) { - data[seq_id*n_tokens + s*n_seq_tokens + i] = div[seq_id]; - } - } - } - - if (cparams.embeddings && ( - cparams.pooling_type == LLAMA_POOLING_TYPE_CLS || - cparams.pooling_type == LLAMA_POOLING_TYPE_RANK)) { - const int64_t n_tokens = ubatch.n_tokens; - const int64_t n_seq_tokens = ubatch.n_seq_tokens; - const int64_t n_seqs = ubatch.n_seqs; - - GGML_ASSERT(lctx.inp_cls); - GGML_ASSERT(ggml_backend_buffer_is_host(lctx.inp_cls->buffer)); - - uint32_t * data = (uint32_t *) lctx.inp_cls->data; - memset(lctx.inp_cls->data, 0, n_tokens * ggml_element_size(lctx.inp_cls)); - - for (int s = 0; s < n_seqs; ++s) { - const llama_seq_id seq_id = ubatch.seq_id[s][0]; - - // TODO: adapt limits to n_seqs when ubatch.equal_seqs is true - GGML_ASSERT(seq_id < n_tokens && "seq_id cannot be larger than n_tokens with pooling_type == CLS or RANK"); - - for (int i = 0; i < n_seq_tokens; ++i) { - const llama_pos pos = ubatch.pos[s*n_seq_tokens + i]; - - if (pos == 0) { - data[seq_id] = s*n_seq_tokens + i; - } - } - } - } - - if (cparams.embeddings && cparams.pooling_type == LLAMA_POOLING_TYPE_LAST) { - const int64_t n_tokens = ubatch.n_tokens; - const int64_t n_seq_tokens = ubatch.n_seq_tokens; - const int64_t n_seqs = ubatch.n_seqs; - - GGML_ASSERT(lctx.inp_cls); - GGML_ASSERT(ggml_backend_buffer_is_host(lctx.inp_cls->buffer)); - - uint32_t * data = (uint32_t *) lctx.inp_cls->data; - memset(lctx.inp_cls->data, 0, n_tokens * ggml_element_size(lctx.inp_cls)); - - std::vector last_pos(n_tokens, -1); - std::vector last_row(n_tokens, -1); - - for (int s = 0; s < n_seqs; ++s) { - const llama_seq_id seq_id = ubatch.seq_id[s][0]; - - // TODO: adapt limits to n_seqs when ubatch.equal_seqs is true - GGML_ASSERT(seq_id < n_tokens && "seq_id cannot be larger than n_tokens with pooling_type == LAST"); - - for (int i = 0; i < n_seq_tokens; ++i) { - const llama_pos pos = ubatch.pos[s*n_seq_tokens + i]; - - if (pos >= last_pos[seq_id]) { - last_pos[seq_id] = pos; - last_row[seq_id] = s*n_seq_tokens + i; - } - } - } - - for (int i = 0; i < n_tokens; ++i) { - if (last_row[i] >= 0) { - data[i] = last_row[i]; - } - } - } - - if (kv_self.recurrent) { - const int64_t n_kv = kv_self.n; - - if (lctx.inp_s_mask) { - GGML_ASSERT(ggml_backend_buffer_is_host(lctx.inp_s_mask->buffer)); - float * data = (float *) lctx.inp_s_mask->data; - - // clear unused states - for (int i = 0; i < n_kv; ++i) { - const uint32_t cell_id = i + kv_self.head; - llama_kv_cell & kv_cell = lctx.kv_self.cells[cell_id]; - - data[i] = (float) (kv_cell.src >= 0); - - // only clear once - if (kv_cell.src < 0) { - kv_cell.src = cell_id; - } - } - } - - if (lctx.inp_s_copy) { - GGML_ASSERT(ggml_backend_buffer_is_host(lctx.inp_s_copy->buffer)); - int32_t * data = (int32_t *) lctx.inp_s_copy->data; - - // assuming copy destinations ALWAYS happen ONLY on the cells between head and head+n - for (uint32_t i = 0; i < n_kv; ++i) { - const uint32_t cell_id = i + kv_self.head; - llama_kv_cell & kv_cell = lctx.kv_self.cells[cell_id]; - - // prevent out-of-bound sources - if (kv_cell.src < 0 || (uint32_t) kv_cell.src >= kv_self.size) { - kv_cell.src = cell_id; - } - - data[i] = kv_cell.src; - - // ensure copy only happens once - if (kv_cell.src != (int32_t) cell_id) { - kv_cell.src = cell_id; - } - } - } - } - - if (lctx.inp_pos_bucket) { - const int64_t n_tokens = ubatch.n_tokens; - - GGML_ASSERT(ggml_backend_buffer_is_host(lctx.inp_pos_bucket->buffer)); - GGML_ASSERT(!ubatch.equal_seqs); // TODO: use ubatch.n_seqs instead of failing - - int32_t * data = (int32_t *) lctx.inp_pos_bucket->data; - - if (!lctx.is_encoding) { - const int64_t n_kv = kv_self.n; - for (int h = 0; h < 1; ++h) { - for (int j = 0; j < n_tokens; ++j) { - for (int i = 0; i < n_kv; ++i) { - data[h*(n_kv*n_tokens) + j*n_kv + i] = llama_relative_position_bucket(lctx.kv_self.cells[i].pos, ubatch.pos[j], hparams.n_rel_attn_bkts, lctx.is_encoding); - } - } - } - } else { - for (int h = 0; h < 1; ++h) { - for (int j = 0; j < n_tokens; ++j) { - for (int i = 0; i < n_tokens; ++i) { - data[h*(n_tokens*n_tokens) + j*n_tokens + i] = llama_relative_position_bucket(ubatch.pos[i], ubatch.pos[j], hparams.n_rel_attn_bkts, lctx.is_encoding); - } - } - } - } - } - - if (!lctx.is_encoding && lctx.inp_embd_enc) { - assert(lctx.inp_embd_enc->type == GGML_TYPE_F32); - assert((size_t) ggml_nelements(lctx.inp_embd_enc) == lctx.embd_enc.size()); - - ggml_backend_tensor_set(lctx.inp_embd_enc, lctx.embd_enc.data(), 0, ggml_nbytes(lctx.inp_embd_enc)); - } - - if (!lctx.is_encoding && lctx.inp_KQ_mask_cross) { - const int64_t n_output_enc = lctx.embd_enc.size() / hparams.n_embd; - const int64_t n_tokens = ubatch.n_tokens; - - GGML_ASSERT(ggml_backend_buffer_is_host(lctx.inp_KQ_mask_cross->buffer)); - GGML_ASSERT(!ubatch.equal_seqs); // TODO: use ubatch.n_seqs instead of failing - - float * data = (float *) lctx.inp_KQ_mask_cross->data; - - for (int h = 0; h < 1; ++h) { - for (int j = 0; j < n_tokens; ++j) { - for (int i = 0; i < n_output_enc; ++i) { - float f = -INFINITY; - for (int s = 0; s < ubatch.n_seq_id[j]; ++s) { - const llama_seq_id seq_id = ubatch.seq_id[j][s]; - if (lctx.seq_ids_enc[i].find(seq_id) != lctx.seq_ids_enc[i].end()) { - f = 0.0f; - } - } - data[h*(n_output_enc*n_tokens) + j*n_output_enc + i] = f; - } - } - - for (int i = n_tokens; i < GGML_PAD(n_tokens, GGML_KQ_MASK_PAD); ++i) { - for (int j = 0; j < n_output_enc; ++j) { - data[h*(n_output_enc*n_tokens) + i*n_output_enc + j] = -INFINITY; - } - } + // initialize scheduler with the worst-case graph + ggml_backend_sched_reset(sched.get()); + if (!ggml_backend_sched_reserve(sched.get(), gf)) { + LLAMA_LOG_ERROR("%s: failed to allocate compute buffers\n", __func__); } } } -// llama output +enum llama_pooling_type llama_context::pooling_type() const { + return cparams.pooling_type; +} -size_t llama_output_reserve(struct llama_context & lctx, size_t n_outputs) { - const auto & cparams = lctx.cparams; - const auto & hparams = lctx.model.hparams; - const auto & vocab = lctx.model.vocab; +float * llama_context::get_logits() { + // reorder logits for backward compatibility + output_reorder(); - const size_t n_outputs_max = std::max(n_outputs, (size_t) cparams.n_seq_max); + return logits; +} + +float * llama_context::get_logits_ith(int32_t i) { + int32_t j = -1; + + try { + if (logits == nullptr) { + throw std::runtime_error("no logits"); + } + + if (i < 0) { + j = n_outputs + i; + if (j < 0) { + throw std::runtime_error(format("negative index out of range [0, %d)", n_outputs)); + } + } else if ((size_t) i >= output_ids.size()) { + throw std::runtime_error(format("out of range [0, %zu)", output_ids.size())); + } else { + j = output_ids[i]; + } + + if (j < 0) { + throw std::runtime_error(format("batch.logits[%d] != true", i)); + } + if (j >= n_outputs) { + // This should not happen + throw std::runtime_error(format("corrupt output buffer (j=%d, n_outputs=%d)", j, n_outputs)); + } + + return logits + j*model.vocab.n_tokens(); + } catch (const std::exception & err) { + LLAMA_LOG_ERROR("%s: invalid logits id %d, reason: %s\n", __func__, i, err.what()); +#ifndef NDEBUG + GGML_ABORT("fatal error"); +#else + return nullptr; +#endif + } +} + +float * llama_context::get_embeddings() { + // reorder embeddings for backward compatibility + output_reorder(); + + return embd; +} + +float * llama_context::get_embeddings_ith(int32_t i) { + int32_t j = -1; + + try { + if (embd == nullptr) { + throw std::runtime_error("no embeddings"); + } + + if (i < 0) { + j = n_outputs + i; + if (j < 0) { + throw std::runtime_error(format("negative index out of range [0, %d)", n_outputs)); + } + } else if ((size_t) i >= output_ids.size()) { + throw std::runtime_error(format("out of range [0, %zu)", output_ids.size())); + } else { + j = output_ids[i]; + } + + if (j < 0) { + throw std::runtime_error(format("batch.logits[%d] != true", i)); + } + if (j >= n_outputs) { + // This should not happen + throw std::runtime_error(format("corrupt output buffer (j=%d, n_outputs=%d)", j, n_outputs)); + } + + return embd + j*model.hparams.n_embd; + } catch (const std::exception & err) { + LLAMA_LOG_ERROR("%s: invalid embeddings id %d, reason: %s\n", __func__, i, err.what()); +#ifndef NDEBUG + GGML_ABORT("fatal error"); +#else + return nullptr; +#endif + } +} + +float * llama_context::get_embeddings_seq(llama_seq_id seq_id) { + auto it = embd_seq.find(seq_id); + if (it == embd_seq.end()) { + return nullptr; + } + + return it->second.data(); +} + +void llama_context::attach_threadpool( + ggml_threadpool_t threadpool, + ggml_threadpool_t threadpool_batch) { + LLAMA_LOG_DEBUG("%s: call\n", __func__); + + this->threadpool = threadpool; + this->threadpool_batch = threadpool_batch ? threadpool_batch : threadpool; +} + +void llama_context::detach_threadpool() { + LLAMA_LOG_DEBUG("%s: call\n", __func__); + + this->threadpool = nullptr; + this->threadpool_batch = nullptr; +} + +void llama_context::set_n_threads(int32_t n_threads, int32_t n_threads_batch) { + LLAMA_LOG_DEBUG("%s: n_threads = %d, n_threads_batch = %d\n", __func__, n_threads, n_threads_batch); + + cparams.n_threads = n_threads; + cparams.n_threads_batch = n_threads_batch; +} + +void llama_context::set_abort_callback(bool (*abort_callback)(void * data), void * abort_callback_data) { + LLAMA_LOG_DEBUG("%s: call\n", __func__); + + this->abort_callback = abort_callback; + this->abort_callback_data = abort_callback_data; + + for (auto & backend : backends) { + auto * reg = ggml_backend_dev_backend_reg(ggml_backend_get_device(backend.get())); + auto * set_abort_callback_fn = (ggml_backend_set_abort_callback_t) ggml_backend_reg_get_proc_address(reg, "ggml_backend_set_abort_callback"); + if (set_abort_callback_fn) { + set_abort_callback_fn(backend.get(), this->abort_callback, this->abort_callback_data); + } + } +} + +void llama_context::set_embeddings(bool value) { + LLAMA_LOG_DEBUG("%s: value = %d\n", __func__, value); + + cparams.embeddings = value; +} + +void llama_context::set_causal_attn(bool value) { + LLAMA_LOG_DEBUG("%s: value = %d\n", __func__, value); + + cparams.causal_attn = value; +} + +void llama_context::set_warmup(bool value) { + LLAMA_LOG_DEBUG("%s: value = %d\n", __func__, value); + + cparams.warmup = value; +} + +void llama_context::set_adapter_lora( + llama_adapter_lora * adapter, + float scale) { + LLAMA_LOG_DEBUG("%s: adapter = %p, scale = %f\n", __func__, (void *) adapter, scale); + + loras[adapter] = scale; +} + +bool llama_context::rm_adapter_lora( + llama_adapter_lora * adapter) { + LLAMA_LOG_DEBUG("%s: adapter = %p\n", __func__, (void *) adapter); + + auto pos = loras.find(adapter); + if (pos != loras.end()) { + loras.erase(pos); + return true; + } + + return false; +} + +void llama_context::clear_adapter_lora() { + LLAMA_LOG_DEBUG("%s: call\n", __func__); + + loras.clear(); +} + +bool llama_context::apply_adapter_cvec( + const float * data, + size_t len, + int32_t n_embd, + int32_t il_start, + int32_t il_end) { + LLAMA_LOG_DEBUG("%s: il_start = %d, il_end = %d\n", __func__, il_start, il_end); + + return cvec.apply(model, data, len, n_embd, il_start, il_end); +} + +int llama_context::encode(llama_batch & inp_batch) { + if (inp_batch.n_tokens == 0) { + LLAMA_LOG_ERROR("%s: n_tokens == 0\n", __func__); + return -1; + } + + // temporary allocate memory for the input batch if needed + // TODO: this is incorrect for multiple sequences because pos_max() is the maximum across all sequences + llama_batch_allocr batch_allocr(inp_batch, inp_batch.pos ? -1 : kv_self->pos_max() + 1); + + const llama_batch & batch = batch_allocr.batch; + const int32_t n_tokens = batch.n_tokens; + + const auto & hparams = model.hparams; + + GGML_ASSERT((!batch.token && batch.embd) || (batch.token && !batch.embd)); // NOLINT + + if (batch.token) { + for (int32_t i = 0; i < n_tokens; ++i) { + if (batch.token[i] < 0 || (uint32_t) batch.token[i] >= model.vocab.n_tokens()) { + LLAMA_LOG_ERROR("%s: invalid token[%d] = %d\n", __func__, i, batch.token[i]); + return -1; + } + } + } + + // micro-batching is not possible for non-causal encoding, so we process the batch in a single shot + GGML_ASSERT(cparams.n_ubatch >= (uint32_t) n_tokens && "encoder requires n_ubatch >= n_tokens"); + + if (t_compute_start_us == 0) { + t_compute_start_us = ggml_time_us(); + } + + n_queued_tokens += n_tokens; + + const int64_t n_embd = hparams.n_embd; + + sbatch.from_batch(batch, n_embd, /* simple_split */ true, /* logits_all */ true); + + const llama_ubatch ubatch = sbatch.split_simple(n_tokens); + + // reserve output buffer + if (output_reserve(n_tokens) < n_tokens) { + LLAMA_LOG_ERROR("%s: could not reserve space for batch with %u outputs\n", __func__, n_tokens); + return -2; + }; + + for (int32_t i = 0; i < n_tokens; ++i) { + output_ids[i] = i; + } + + n_outputs = n_tokens; + + //batch_manager->prepare(ubatch); + + ggml_backend_sched_reset(sched.get()); + ggml_backend_sched_set_eval_callback(sched.get(), cparams.cb_eval, cparams.cb_eval_user_data); + + const auto causal_attn_org = cparams.causal_attn; + + // always use non-causal attention for encoder graphs + // TODO: this is a tmp solution until we have a proper way to support enc-dec models + // ref: https://github.com/ggml-org/llama.cpp/pull/12181#issuecomment-2730451223 + cparams.causal_attn = false; + + auto * gf = graph_init(); + auto res = graph_build(ctx_compute.get(), gf, ubatch, LLM_GRAPH_TYPE_ENCODER); + + ggml_backend_sched_alloc_graph(sched.get(), gf); + + res->set_inputs(&ubatch); + + cparams.causal_attn = causal_attn_org; + + const auto compute_status = graph_compute(gf, n_tokens > 1); + switch (compute_status) { + case GGML_STATUS_SUCCESS: + break; + case GGML_STATUS_ABORTED: + return 2; + case GGML_STATUS_ALLOC_FAILED: + return -2; + case GGML_STATUS_FAILED: + default: + return -3; + } + + auto * t_embd = res->get_embd_pooled() ? res->get_embd_pooled() : res->get_embd(); + + // extract embeddings + if (t_embd) { + ggml_backend_t backend_embd = ggml_backend_sched_get_tensor_backend(sched.get(), t_embd); + GGML_ASSERT(backend_embd != nullptr); + + GGML_ASSERT(embd != nullptr); + + switch (cparams.pooling_type) { + case LLAMA_POOLING_TYPE_NONE: + { + // extract token embeddings + GGML_ASSERT(n_tokens*n_embd <= (int64_t) embd_size); + ggml_backend_tensor_get_async(backend_embd, t_embd, embd, 0, n_tokens*n_embd*sizeof(float)); + } break; + case LLAMA_POOLING_TYPE_MEAN: + case LLAMA_POOLING_TYPE_CLS: + case LLAMA_POOLING_TYPE_LAST: + { + // extract sequence embeddings + auto & embd_seq_out = embd_seq; + embd_seq_out.clear(); + + GGML_ASSERT(!ubatch.equal_seqs); // TODO: handle equal splits + + for (int32_t i = 0; i < n_tokens; i++) { + const llama_seq_id seq_id = ubatch.seq_id[i][0]; + if (embd_seq_out.find(seq_id) != embd_seq_out.end()) { + continue; + } + embd_seq_out[seq_id].resize(n_embd); + ggml_backend_tensor_get_async(backend_embd, t_embd, embd_seq_out[seq_id].data(), (n_embd*seq_id)*sizeof(float), n_embd*sizeof(float)); + } + } break; + case LLAMA_POOLING_TYPE_RANK: + { + // TODO: this likely should be the same logic as in llama_decoder_internal, but better to + // wait for an encoder model that requires this pooling type in order to test it + // https://github.com/ggerganov/llama.cpp/pull/9510 + GGML_ABORT("RANK pooling not implemented yet"); + } + case LLAMA_POOLING_TYPE_UNSPECIFIED: + { + GGML_ABORT("unknown pooling type"); + } + } + } + + // Reset state for the next token before backend sync, to allow the CPU activities in the reset to + // overlap with device computation. + ggml_backend_sched_reset(sched.get()); + + // TODO: hacky solution + if (model.arch == LLM_ARCH_T5 && t_embd) { + //cross.t_embd = t_embd; + + synchronize(); + + cross.n_embd = t_embd->ne[0]; + cross.n_enc = t_embd->ne[1]; + cross.v_embd.resize(cross.n_embd*cross.n_enc); + memcpy(cross.v_embd.data(), embd, ggml_nbytes(t_embd)); + + // remember the sequence ids used during the encoding - needed for cross attention later + cross.seq_ids_enc.resize(n_tokens); + for (int32_t i = 0; i < n_tokens; i++) { + cross.seq_ids_enc[i].clear(); + for (int s = 0; s < ubatch.n_seq_id[i]; s++) { + llama_seq_id seq_id = ubatch.seq_id[i][s]; + cross.seq_ids_enc[i].insert(seq_id); + } + } + } + + return 0; +} + +int llama_context::decode(llama_batch & inp_batch) { + if (inp_batch.n_tokens == 0) { + LLAMA_LOG_ERROR("%s: n_tokens == 0\n", __func__); + return -1; + } + + // temporary allocate memory for the input batch if needed + // TODO: this is incorrect for multiple sequences because pos_max() is the maximum across all sequences + llama_batch_allocr batch_allocr(inp_batch, inp_batch.pos ? -1 : kv_self->pos_max() + 1); + + const llama_batch & batch = batch_allocr.batch; + + const auto & vocab = model.vocab; + const auto & hparams = model.hparams; + + const int32_t n_vocab = vocab.n_tokens(); + + const int64_t n_tokens_all = batch.n_tokens; + const int64_t n_embd = hparams.n_embd; + + llama_kv_cache_guard kv_guard(kv_self.get()); + + GGML_ASSERT((!batch.token && batch.embd) || (batch.token && !batch.embd)); // NOLINT + + if (batch.token) { + for (int64_t i = 0; i < n_tokens_all; ++i) { + if (batch.token[i] < 0 || (uint32_t) batch.token[i] >= model.vocab.n_tokens()) { + LLAMA_LOG_ERROR("%s: invalid token[%" PRId64 "] = %d\n", __func__, i, batch.token[i]); + throw std::runtime_error("invalid token"); + } + } + } + + GGML_ASSERT(n_tokens_all <= cparams.n_batch); + + GGML_ASSERT((cparams.causal_attn || cparams.n_ubatch >= n_tokens_all) && "non-causal attention requires n_ubatch >= n_tokens"); + + if (t_compute_start_us == 0) { + t_compute_start_us = ggml_time_us(); + } + n_queued_tokens += n_tokens_all; + + // this indicates we are doing pooled embedding, so we ignore batch.logits and output all tokens + const bool embd_pooled = cparams.embeddings && cparams.pooling_type != LLAMA_POOLING_TYPE_NONE; + + embd_seq.clear(); + + int64_t n_outputs_all = 0; + + // count outputs + if (batch.logits && !embd_pooled) { + for (uint32_t i = 0; i < n_tokens_all; ++i) { + n_outputs_all += batch.logits[i] != 0; + } + } else if (logits_all || embd_pooled) { + n_outputs_all = n_tokens_all; + } else { + // keep last output only + n_outputs_all = 1; + } + + const bool logits_all = n_outputs_all == n_tokens_all; + + sbatch.from_batch(batch, n_embd, + /* simple_split */ !kv_self->recurrent, + /* logits_all */ logits_all); + + // reserve output buffer + if (output_reserve(n_outputs_all) < n_outputs_all) { + LLAMA_LOG_ERROR("%s: could not reserve space for batch with %" PRId64 " outputs\n", __func__, n_outputs_all); + return -2; + }; + + // handle any pending defrags/shifts + kv_self_update(); + + int64_t n_outputs_prev = 0; + + while (sbatch.n_tokens > 0) { + llama_ubatch ubatch = llama_ubatch(); + + const auto & n_ubatch = cparams.n_ubatch; + + if (kv_self->recurrent) { + if (embd_pooled) { + // Pooled embeddings cannot be split across ubatches (yet) + ubatch = sbatch.split_seq(cparams.n_ubatch); + } else { + // recurrent model architectures are easier to implement + // with equal-length sequences + ubatch = sbatch.split_equal(cparams.n_ubatch); + } + } else { + ubatch = sbatch.split_simple(n_ubatch); + } + + // count the outputs in this u_batch + { + int32_t n_outputs_new = 0; + + if (n_outputs_all == n_tokens_all) { + n_outputs_new = ubatch.n_tokens; + } else { + GGML_ASSERT(ubatch.output); + for (uint32_t i = 0; i < ubatch.n_tokens; i++) { + n_outputs_new += (int32_t) (ubatch.output[i] != 0); + } + } + + // needs to happen before the graph is built + n_outputs = n_outputs_new; + } + + // find KV slot + { + if (!kv_self->find_slot(ubatch)) { + LLAMA_LOG_WARN("%s: failed to find KV cache slot for ubatch of size %d\n", __func__, ubatch.n_tokens); + + return 1; + } + + if (!kv_self->recurrent) { + // a heuristic, to avoid attending the full cache if it is not yet utilized + // after enough generations, the benefit from this heuristic disappears + // if we start defragmenting the cache, the benefit from this will be more important + const uint32_t pad = kv_self->get_padding(cparams); + kv_self->n = std::min(kv_self->size, std::max(pad, GGML_PAD(kv_self->cell_max(), pad))); + } + } + + //printf("kv_self.n = %5d, kv_self.used = %5d, kv_self.head = %5d\n", kv_self->n, kv_self->used, kv_self->head); + + ggml_backend_sched_reset(sched.get()); + ggml_backend_sched_set_eval_callback(sched.get(), cparams.cb_eval, cparams.cb_eval_user_data); + + auto * gf = graph_init(); + auto res = graph_build(ctx_compute.get(), gf, ubatch, LLM_GRAPH_TYPE_DECODER); + + // LLAMA_LOG_INFO("graph build time: %.3f ms (%d nodes, %d leafs)\n", (ggml_time_us() - t_start_us)/1000.0, gf->n_nodes, gf->n_leafs); + + ggml_backend_sched_alloc_graph(sched.get(), gf); + + res->set_inputs(&ubatch); + + const auto compute_status = graph_compute(gf, ubatch.n_tokens > 1); + if (compute_status != GGML_STATUS_SUCCESS) { + switch (compute_status) { + case GGML_STATUS_ABORTED: + return 2; + case GGML_STATUS_ALLOC_FAILED: + return -2; + case GGML_STATUS_FAILED: + default: + return -3; + } + } + + // plot the computation graph in dot format (for debugging purposes) + //if (n_past%100 == 0) { + // ggml_graph_dump_dot(gf, NULL, "llama.dot"); + //} + + auto * t_logits = cparams.embeddings ? nullptr : res->get_logits(); + auto * t_embd = cparams.embeddings ? res->get_embd() : nullptr; + + if (t_embd && res->get_embd_pooled()) { + t_embd = res->get_embd_pooled(); + } + + // extract logits + if (t_logits && n_outputs > 0) { + ggml_backend_t backend_res = ggml_backend_sched_get_tensor_backend(sched.get(), t_logits); + GGML_ASSERT(backend_res != nullptr); + GGML_ASSERT(logits != nullptr); + + float * logits_out = logits + n_outputs_prev*n_vocab; + + if (n_outputs) { + GGML_ASSERT( n_outputs_prev + n_outputs <= n_outputs_all); + GGML_ASSERT((n_outputs_prev + n_outputs)*n_vocab <= (int64_t) logits_size); + ggml_backend_tensor_get_async(backend_res, t_logits, logits_out, 0, n_outputs*n_vocab*sizeof(float)); + } + } + + // extract embeddings + if (t_embd && n_outputs > 0) { + ggml_backend_t backend_embd = ggml_backend_sched_get_tensor_backend(sched.get(), t_embd); + GGML_ASSERT(backend_embd != nullptr); + + switch (cparams.pooling_type) { + case LLAMA_POOLING_TYPE_NONE: + { + // extract token embeddings + GGML_ASSERT(embd != nullptr); + float * embd_out = embd + n_outputs_prev*n_embd; + + if (n_outputs) { + GGML_ASSERT( n_outputs_prev + n_outputs <= n_outputs_all); + GGML_ASSERT((n_outputs_prev + n_outputs)*n_embd <= (int64_t) embd_size); + ggml_backend_tensor_get_async(backend_embd, t_embd, embd_out, 0, n_outputs*n_embd*sizeof(float)); + } + } break; + case LLAMA_POOLING_TYPE_MEAN: + case LLAMA_POOLING_TYPE_CLS: + case LLAMA_POOLING_TYPE_LAST: + { + // extract sequence embeddings (cleared before processing each batch) + auto & embd_seq_out = embd_seq; + + for (uint32_t s = 0; s < ubatch.n_seqs; ++s) { + const llama_seq_id seq_id = ubatch.seq_id[s][0]; + if (embd_seq_out.find(seq_id) != embd_seq_out.end()) { + continue; + } + embd_seq_out[seq_id].resize(n_embd); + ggml_backend_tensor_get_async(backend_embd, t_embd, embd_seq_out[seq_id].data(), (n_embd*seq_id)*sizeof(float), n_embd*sizeof(float)); + } + } break; + case LLAMA_POOLING_TYPE_RANK: + { + // extract the rerank score - a single float per sequence + auto & embd_seq_out = embd_seq; + + for (uint32_t s = 0; s < ubatch.n_seqs; ++s) { + const llama_seq_id seq_id = ubatch.seq_id[s][0]; + if (embd_seq_out.find(seq_id) != embd_seq_out.end()) { + continue; + } + embd_seq_out[seq_id].resize(1); + ggml_backend_tensor_get_async(backend_embd, t_embd, embd_seq_out[seq_id].data(), (seq_id)*sizeof(float), sizeof(float)); + } + } break; + case LLAMA_POOLING_TYPE_UNSPECIFIED: + { + GGML_ABORT("unknown pooling type"); + } + } + } + + n_outputs_prev += n_outputs; + } + + // finalize the batch processing + kv_guard.commit(); + + // set output mappings + { + bool sorted_output = true; + + GGML_ASSERT(sbatch.out_ids.size() == (size_t) n_outputs_all); + + for (int64_t i = 0; i < n_outputs_all; ++i) { + int64_t out_id = sbatch.out_ids[i]; + output_ids[out_id] = i; + if (out_id != i) { + sorted_output = false; + } + } + + if (sorted_output) { + sbatch.out_ids.clear(); + } + } + + // set to total number of outputs in the batch, for use in llama_get_logits_ith + n_outputs = n_outputs_all; + + // wait for the computation to finish (automatically done when obtaining the model output) + //synchronize(); + + // decide if we need to defrag the kv cache + if (cparams.causal_attn && cparams.defrag_thold > 0.0f) { + // - do not defrag small contexts (i.e. < 2048 tokens) + // - count the padding towards the number of used tokens + const float fragmentation = kv_self->n >= 2048 ? std::max(0.0f, 1.0f - float(kv_self->used + kv_self->get_padding(cparams))/float(kv_self->n)) : 0.0f; + + // queue defragmentation for next llama_kv_cache_update + if (fragmentation > cparams.defrag_thold) { + LLAMA_LOG_DEBUG("%s: fragmentation: %.2f - requesting defrag\n", __func__, fragmentation); + + kv_self->defrag(); + } + } + + // Reset state for the next token before backend sync, to allow the CPU activities in the reset to + // overlap with device computation. + ggml_backend_sched_reset(sched.get()); + + return 0; +} + +// +// output +// + +int32_t llama_context::output_reserve(int32_t n_outputs) { + const auto & hparams = model.hparams; + const auto & vocab = model.vocab; + + const int64_t n_outputs_max = std::max(n_outputs, n_seq_max()); const auto n_batch = cparams.n_batch; const auto n_vocab = vocab.n_tokens(); const auto n_embd = hparams.n_embd; // TODO: use a per-batch flag for logits presence instead - const bool has_logits = !cparams.embeddings; - const bool has_embd = cparams.embeddings && (cparams.pooling_type == LLAMA_POOLING_TYPE_NONE); + bool has_logits = !cparams.embeddings; + bool has_embd = cparams.embeddings && (cparams.pooling_type == LLAMA_POOLING_TYPE_NONE); - const size_t logits_size = has_logits ? n_vocab*n_outputs_max : 0; - const size_t embd_size = has_embd ? n_embd*n_outputs_max : 0; - - if (lctx.output_ids.empty()) { - // init, never resized afterwards - lctx.output_ids.resize(n_batch); + // TODO: hacky enc-dec support + if (model.arch == LLM_ARCH_T5) { + has_logits = true; + has_embd = true; } - const size_t prev_size = lctx.buf_output ? ggml_backend_buffer_get_size(lctx.buf_output.get()) : 0; + logits_size = has_logits ? n_vocab*n_outputs_max : 0; + embd_size = has_embd ? n_embd*n_outputs_max : 0; + + if (output_ids.empty()) { + // init, never resized afterwards + output_ids.resize(n_batch); + } + + const size_t prev_size = buf_output ? ggml_backend_buffer_get_size(buf_output.get()) : 0; const size_t new_size = (logits_size + embd_size) * sizeof(float); // alloc only when more than the current capacity is required // TODO: also consider shrinking the buffer - if (!lctx.buf_output || prev_size < new_size) { - if (lctx.buf_output) { + if (!buf_output || prev_size < new_size) { + if (buf_output) { #ifndef NDEBUG // This doesn't happen often, but may be annoying in some cases (like the HellaSwag benchmark) LLAMA_LOG_INFO("%s: reallocating output buffer from size %.02f MiB to %.02f MiB\n", __func__, prev_size / 1024.0 / 1024.0, new_size / 1024.0 / 1024.0); #endif - lctx.buf_output = nullptr; - lctx.logits = nullptr; - lctx.embd = nullptr; + buf_output = nullptr; + logits = nullptr; + embd = nullptr; } auto * buft = ggml_backend_cpu_buffer_type(); // try to use the host buffer of the device where the output tensor is allocated for faster transfer to system memory - auto * output_dev = lctx.model.dev_output(); + auto * output_dev = model.dev_output(); auto * output_dev_host_buft = output_dev ? ggml_backend_dev_host_buffer_type(output_dev) : nullptr; if (output_dev_host_buft) { buft = output_dev_host_buft; } - lctx.buf_output.reset(ggml_backend_buft_alloc_buffer(buft, new_size)); - if (lctx.buf_output == nullptr) { + buf_output.reset(ggml_backend_buft_alloc_buffer(buft, new_size)); + if (buf_output == nullptr) { LLAMA_LOG_ERROR("%s: failed to allocate output buffer of size %.2f MiB\n", __func__, new_size / (1024.0 * 1024.0)); return 0; } } - float * output_base = (float *) ggml_backend_buffer_get_base(lctx.buf_output.get()); + float * output_base = (float *) ggml_backend_buffer_get_base(buf_output.get()); - lctx.logits = has_logits ? output_base : nullptr; - lctx.embd = has_embd ? output_base + logits_size : nullptr; - - lctx.output_size = n_outputs_max; - lctx.logits_size = logits_size; - lctx.embd_size = embd_size; + logits = has_logits ? output_base : nullptr; + embd = has_embd ? output_base + logits_size : nullptr; // set all ids as invalid (negative) - std::fill(lctx.output_ids.begin(), lctx.output_ids.end(), -1); + std::fill(output_ids.begin(), output_ids.end(), -1); - ggml_backend_buffer_clear(lctx.buf_output.get(), 0); + ggml_backend_buffer_clear(buf_output.get(), 0); - lctx.n_outputs = 0; + this->n_outputs = 0; + this->n_outputs_max = n_outputs_max; return n_outputs_max; } -void llama_output_reorder(struct llama_context & ctx) { - std::vector & out_ids = ctx.sbatch.out_ids; +void llama_context::output_reorder() { + auto & out_ids = sbatch.out_ids; if (!out_ids.empty()) { - const uint32_t n_vocab = ctx.model.vocab.n_tokens(); - const uint32_t n_embd = ctx.model.hparams.n_embd; + const uint32_t n_vocab = model.vocab.n_tokens(); + const uint32_t n_embd = model.hparams.n_embd; - const int32_t n_outputs = ctx.n_outputs; GGML_ASSERT((size_t) n_outputs == out_ids.size()); // TODO: is there something more efficient which also minimizes swaps? @@ -559,842 +1570,156 @@ void llama_output_reorder(struct llama_context & ctx) { } if (j_min == i) { continue; } std::swap(out_ids[i], out_ids[j_min]); - if (ctx.logits_size > 0) { + if (logits_size > 0) { for (uint32_t k = 0; k < n_vocab; k++) { - std::swap(ctx.logits[i*n_vocab + k], ctx.logits[j_min*n_vocab + k]); + std::swap(logits[i*n_vocab + k], logits[j_min*n_vocab + k]); } } - if (ctx.embd_size > 0) { + if (embd_size > 0) { for (uint32_t k = 0; k < n_embd; k++) { - std::swap(ctx.embd[i*n_embd + k], ctx.embd[j_min*n_embd + k]); + std::swap(embd[i*n_embd + k], embd[j_min*n_embd + k]); } } } - std::fill(ctx.output_ids.begin(), ctx.output_ids.end(), -1); + std::fill(output_ids.begin(), output_ids.end(), -1); for (int32_t i = 0; i < n_outputs; ++i) { - ctx.output_ids[out_ids[i]] = i; + output_ids[out_ids[i]] = i; } out_ids.clear(); } } // -// interface implementation +// graph // -void llama_free(struct llama_context * ctx) { - delete ctx; +int32_t llama_context::graph_max_nodes() const { + return std::max(65536, 5*model.n_tensors()); } -uint32_t llama_n_ctx(const struct llama_context * ctx) { - return ctx->cparams.n_ctx; +ggml_cgraph * llama_context::graph_init() { + ggml_init_params params = { + /*.mem_size =*/ buf_compute_meta.size(), + /*.mem_buffer =*/ buf_compute_meta.data(), + /*.no_alloc =*/ true, + }; + + ctx_compute.reset(ggml_init(params)); + + return ggml_new_graph_custom(ctx_compute.get(), graph_max_nodes(), false); } -uint32_t llama_n_batch(const struct llama_context * ctx) { - return ctx->cparams.n_batch; +llm_graph_result_ptr llama_context::graph_build( + ggml_context * ctx, + ggml_cgraph * gf, + const llama_ubatch & ubatch, + llm_graph_type gtype) { + return model.build_graph( + { + /*.ctx =*/ ctx, + /*.arch =*/ model.arch, + /*.hparams =*/ model.hparams, + /*.cparams =*/ cparams, + /*.ubatch =*/ ubatch, + /*.sched =*/ sched.get(), + /*.backend_cpu =*/ backend_cpu, + /*.cvec =*/ &cvec, + /*.loras =*/ &loras, + /*.memory =*/ kv_self.get(), + /*.cross =*/ &cross, + /*.n_outputs =*/ n_outputs, + /*.cb =*/ graph_get_cb(), + }, gf, gtype); } -uint32_t llama_n_ubatch(const struct llama_context * ctx) { - return ctx->cparams.n_ubatch; -} +ggml_status llama_context::graph_compute( + ggml_cgraph * gf, + bool batched) { + int n_threads = batched ? cparams.n_threads_batch : cparams.n_threads; + ggml_threadpool_t tp = batched ? threadpool_batch : threadpool; -uint32_t llama_n_seq_max(const struct llama_context * ctx) { - return ctx->kv_self.size; -} - -const struct llama_model * llama_get_model(const struct llama_context * ctx) { - return &ctx->model; -} - -enum llama_pooling_type llama_pooling_type(const struct llama_context * ctx) { - return ctx->cparams.pooling_type; -} - -void llama_attach_threadpool( - struct llama_context * ctx, - ggml_threadpool_t threadpool, - ggml_threadpool_t threadpool_batch) { - ctx->threadpool = threadpool; - ctx->threadpool_batch = threadpool_batch ? threadpool_batch : threadpool; -} - -void llama_detach_threadpool(struct llama_context * ctx) { - ctx->threadpool = nullptr; - ctx->threadpool_batch = nullptr; -} - -void llama_set_n_threads(struct llama_context * ctx, int32_t n_threads, int32_t n_threads_batch) { - ctx->cparams.n_threads = n_threads; - ctx->cparams.n_threads_batch = n_threads_batch; -} - -int32_t llama_n_threads(struct llama_context * ctx) { - return ctx->cparams.n_threads; -} - -int32_t llama_n_threads_batch(struct llama_context * ctx) { - return ctx->cparams.n_threads_batch; -} - -void llama_set_abort_callback(struct llama_context * ctx, bool (*abort_callback)(void * data), void * abort_callback_data) { - ctx->abort_callback = abort_callback; - ctx->abort_callback_data = abort_callback_data; - - for (auto & backend : ctx->backends) { - auto * reg = ggml_backend_dev_backend_reg(ggml_backend_get_device(backend.get())); - auto * set_abort_callback_fn = (ggml_backend_set_abort_callback_t) ggml_backend_reg_get_proc_address(reg, "ggml_backend_set_abort_callback"); - if (set_abort_callback_fn) { - set_abort_callback_fn(backend.get(), ctx->abort_callback, ctx->abort_callback_data); - } - } -} - -void llama_set_embeddings(struct llama_context * ctx, bool embeddings) { - ctx->cparams.embeddings = embeddings; -} - -void llama_set_causal_attn(struct llama_context * ctx, bool causal_attn) { - ctx->cparams.causal_attn = causal_attn; -} - -void llama_synchronize(struct llama_context * ctx) { - ggml_backend_sched_synchronize(ctx->sched.get()); - - // FIXME: if multiple single tokens are evaluated without a synchronization, - // the stats will be added to the prompt evaluation stats - // this should only happen when using batch size 1 to evaluate a batch - - // add the evaluation to the stats - if (ctx->n_queued_tokens == 1) { - if (!ctx->cparams.no_perf) { - ctx->t_eval_us += ggml_time_us() - ctx->t_compute_start_us; - } - ctx->n_eval++; - } else if (ctx->n_queued_tokens > 1) { - if (!ctx->cparams.no_perf) { - ctx->t_p_eval_us += ggml_time_us() - ctx->t_compute_start_us; - } - ctx->n_p_eval += ctx->n_queued_tokens; + if (backend_cpu != nullptr) { + auto * reg = ggml_backend_dev_backend_reg(ggml_backend_get_device(backend_cpu)); + auto * set_threadpool_fn = (decltype(ggml_backend_cpu_set_threadpool) *) ggml_backend_reg_get_proc_address(reg, "ggml_backend_cpu_set_threadpool"); + set_threadpool_fn(backend_cpu, tp); } - // get a more accurate load time, upon first eval - if (ctx->n_queued_tokens > 0 && !ctx->has_evaluated_once) { - ctx->t_load_us = ggml_time_us() - ctx->t_start_us; - ctx->has_evaluated_once = true; + // set the number of threads for all the backends + for (const auto & set_n_threads_fn : set_n_threads_fns) { + set_n_threads_fn.second(set_n_threads_fn.first, n_threads); } - ctx->n_queued_tokens = 0; - ctx->t_compute_start_us = 0; + auto status = ggml_backend_sched_graph_compute_async(sched.get(), gf); + if (status != GGML_STATUS_SUCCESS) { + LLAMA_LOG_ERROR("%s: ggml_backend_sched_graph_compute_async failed with error %d\n", __func__, status); + } + + // fprintf(stderr, "splits: %d\n", ggml_backend_sched_get_n_splits(sched)); + + return status; } -float * llama_get_logits(struct llama_context * ctx) { - llama_synchronize(ctx); - - // reorder logits for backward compatibility - // TODO: maybe deprecate this - llama_output_reorder(*ctx); - - return ctx->logits; -} - -float * llama_get_logits_ith(struct llama_context * ctx, int32_t i) { - int32_t j = -1; - - llama_synchronize(ctx); - - try { - if (ctx->logits == nullptr) { - throw std::runtime_error("no logits"); - } - - if (i < 0) { - j = ctx->n_outputs + i; - if (j < 0) { - throw std::runtime_error(format("negative index out of range [0, %d)", ctx->n_outputs)); - } - } else if ((size_t) i >= ctx->output_ids.size()) { - throw std::runtime_error(format("out of range [0, %zu)", ctx->output_ids.size())); +llm_graph_cb llama_context::graph_get_cb() const { + return [&](const llama_ubatch & ubatch, ggml_tensor * cur, const char * name, int il) { + if (il >= 0) { + ggml_format_name(cur, "%s-%d", name, il); } else { - j = ctx->output_ids[i]; + ggml_set_name(cur, name); } - if (j < 0) { - throw std::runtime_error(format("batch.logits[%d] != true", i)); - } - if (j >= ctx->n_outputs) { - // This should not happen - throw std::runtime_error(format("corrupt output buffer (j=%d, n_outputs=%d)", j, ctx->n_outputs)); - } - - return ctx->logits + j*ctx->model.vocab.n_tokens(); - } catch (const std::exception & err) { - LLAMA_LOG_ERROR("%s: invalid logits id %d, reason: %s\n", __func__, i, err.what()); -#ifndef NDEBUG - GGML_ABORT("fatal error"); -#else - return nullptr; -#endif - } -} - -float * llama_get_embeddings(struct llama_context * ctx) { - llama_synchronize(ctx); - - // reorder embeddings for backward compatibility - // TODO: maybe deprecate this - llama_output_reorder(*ctx); - - return ctx->embd; -} - -float * llama_get_embeddings_ith(struct llama_context * ctx, int32_t i) { - int32_t j = -1; - - llama_synchronize(ctx); - - try { - if (ctx->embd == nullptr) { - throw std::runtime_error("no embeddings"); - } - - if (i < 0) { - j = ctx->n_outputs + i; - if (j < 0) { - throw std::runtime_error(format("negative index out of range [0, %d)", ctx->n_outputs)); - } - } else if ((size_t) i >= ctx->output_ids.size()) { - throw std::runtime_error(format("out of range [0, %zu)", ctx->output_ids.size())); - } else { - j = ctx->output_ids[i]; - } - - if (j < 0) { - throw std::runtime_error(format("batch.logits[%d] != true", i)); - } - if (j >= ctx->n_outputs) { - // This should not happen - throw std::runtime_error(format("corrupt output buffer (j=%d, n_outputs=%d)", j, ctx->n_outputs)); - } - - return ctx->embd + j*ctx->model.hparams.n_embd; - } catch (const std::exception & err) { - LLAMA_LOG_ERROR("%s: invalid embeddings id %d, reason: %s\n", __func__, i, err.what()); -#ifndef NDEBUG - GGML_ABORT("fatal error"); -#else - return nullptr; -#endif - } -} - -float * llama_get_embeddings_seq(struct llama_context * ctx, llama_seq_id seq_id) { - llama_synchronize(ctx); - - auto it = ctx->embd_seq.find(seq_id); - if (it == ctx->embd_seq.end()) { - return nullptr; - } - - return it->second.data(); -} - -// llama state API - -// deprecated -size_t llama_get_state_size(struct llama_context * ctx) { - return llama_state_get_size(ctx); -} - -// deprecated -size_t llama_copy_state_data(struct llama_context * ctx, uint8_t * dst) { - return llama_state_get_data(ctx, dst, -1); -} - -// deprecated -size_t llama_set_state_data(struct llama_context * ctx, const uint8_t * src) { - return llama_state_set_data(ctx, src, -1); -} - -// deprecated -bool llama_load_session_file(struct llama_context * ctx, const char * path_session, llama_token * tokens_out, size_t n_token_capacity, size_t * n_token_count_out) { - return llama_state_load_file(ctx, path_session, tokens_out, n_token_capacity, n_token_count_out); -} - -// deprecated -bool llama_save_session_file(struct llama_context * ctx, const char * path_session, const llama_token * tokens, size_t n_token_count) { - return llama_state_save_file(ctx, path_session, tokens, n_token_count); -} - -// TODO: replace all non-fatal assertions with returned errors or exceptions -struct llama_data_write { - virtual void write(const void * src, size_t size) = 0; - virtual void write_tensor_data(const struct ggml_tensor * tensor, size_t offset, size_t size) = 0; - virtual size_t get_size_written() = 0; - virtual ~llama_data_write() = default; - - void write_string(const std::string & str) { - uint32_t str_size = str.size(); - - write(&str_size, sizeof(str_size)); - write(str.data(), str_size); - } - - void write_model_info(const struct llama_context * ctx) { - const std::string arch_str = llm_arch_name(ctx->model.arch); - write_string(arch_str); - // TODO: add more model-specific info which should prevent loading the session file if not identical - } - - //void write_rng(const std::mt19937 & rng) { - // std::ostringstream rng_ss; - // rng_ss << rng; - - // const std::string & rng_str = rng_ss.str(); - - // write_string(rng_str); - //} - - void write_output_ids(struct llama_context * ctx) { - llama_output_reorder(*ctx); - - const uint32_t n_outputs = ctx->n_outputs; - - std::vector output_pos; - - const size_t n_batch = ctx->cparams.n_batch; - const auto & output_ids = ctx->output_ids; - - GGML_ASSERT(n_outputs <= ctx->output_size); - - output_pos.resize(n_outputs); - - // build a more compact representation of the output ids - for (size_t i = 0; i < n_batch; ++i) { - // map an output id to a position in the batch - int32_t pos = output_ids[i]; - if (pos >= 0) { - GGML_ASSERT((uint32_t) pos < n_outputs); - output_pos[pos] = i; + if (!cparams.offload_kqv) { + if (strcmp(name, "kqv_merged_cont") == 0) { + // all nodes between the KV store and the attention output are run on the CPU + ggml_backend_sched_set_tensor_backend(sched.get(), cur, backend_cpu); } } - write(&n_outputs, sizeof(n_outputs)); - - if (n_outputs) { - write(output_pos.data(), n_outputs * sizeof(int32_t)); - } - } - - void write_logits(const struct llama_context * ctx) { - const uint64_t logits_size = std::min((uint64_t) ctx->logits_size, (uint64_t) ctx->n_outputs * ctx->model.vocab.n_tokens()); - - write(&logits_size, sizeof(logits_size)); - - if (logits_size) { - write(ctx->logits, logits_size * sizeof(float)); - } - } - - void write_embeddings(const struct llama_context * ctx) { - const uint64_t embeddings_size = std::min((uint64_t) ctx->embd_size, (uint64_t) ctx->n_outputs * ctx->model.hparams.n_embd); - - write(&embeddings_size, sizeof(embeddings_size)); - - if (embeddings_size) { - write(ctx->embd, embeddings_size * sizeof(float)); - } - } - - void write_kv_cache_meta(const llama_kv_cache & kv_self, const std::vector> & cell_ranges, llama_seq_id seq_id = -1) { - for (const auto & range : cell_ranges) { - for (uint32_t i = range.first; i < range.second; ++i) { - const auto & cell = kv_self.cells[i]; - const llama_pos pos = cell.pos; - const uint32_t n_seq_id = seq_id == -1 ? cell.seq_id.size() : 0; - - write(&pos, sizeof(pos)); - write(&n_seq_id, sizeof(n_seq_id)); - - if (n_seq_id) { - for (auto seq_id : cell.seq_id) { - write(&seq_id, sizeof(seq_id)); - } - } - } - } - } - - void write_kv_cache_data(const struct llama_context * ctx, const std::vector> & cell_ranges) { - const struct llama_kv_cache & kv_self = ctx->kv_self; - const struct llama_hparams & hparams = ctx->model.hparams; - - const uint32_t v_trans = kv_self.v_trans ? 1 : 0; - const uint32_t n_layer = hparams.n_layer; - - write(&v_trans, sizeof(v_trans)); - write(&n_layer, sizeof(n_layer)); - - std::vector tmp_buf; - - // Iterate and write all the keys first, each row is a cell - // Get whole range at a time - for (uint32_t il = 0; il < n_layer; ++il) { - const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa(il) + hparams.n_embd_k_s(); - - // Write key type - const int32_t k_type_i = (int32_t)kv_self.k_l[il]->type; - write(&k_type_i, sizeof(k_type_i)); - - // Write row size of key - const uint64_t k_size_row = ggml_row_size(kv_self.k_l[il]->type, n_embd_k_gqa); - write(&k_size_row, sizeof(k_size_row)); - - // Read each range of cells of k_size length each into tmp_buf and write out - for (const auto & range : cell_ranges) { - const size_t range_size = range.second - range.first; - const size_t buf_size = range_size * k_size_row; - write_tensor_data(kv_self.k_l[il], range.first * k_size_row, buf_size); - } - } - - if (!kv_self.v_trans) { - for (uint32_t il = 0; il < n_layer; ++il) { - const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il) + hparams.n_embd_v_s(); - - // Write value type - const int32_t v_type_i = (int32_t)kv_self.v_l[il]->type; - write(&v_type_i, sizeof(v_type_i)); - - // Write row size of value - const uint64_t v_size_row = ggml_row_size(kv_self.v_l[il]->type, n_embd_v_gqa); - write(&v_size_row, sizeof(v_size_row)); - - // Read each range of cells of v_size length each into tmp_buf and write out - for (const auto & range : cell_ranges) { - const size_t range_size = range.second - range.first; - const size_t buf_size = range_size * v_size_row; - write_tensor_data(kv_self.v_l[il], range.first * v_size_row, buf_size); - } - } - } else { - // When v is transposed, we also need the element size and get the element ranges from each row - const uint32_t kv_size = kv_self.size; - for (uint32_t il = 0; il < n_layer; ++il) { - const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il) + hparams.n_embd_v_s(); - - // Write value type - const int32_t v_type_i = (int32_t)kv_self.v_l[il]->type; - write(&v_type_i, sizeof(v_type_i)); - - // Write element size - const uint32_t v_size_el = ggml_type_size(kv_self.v_l[il]->type); - write(&v_size_el, sizeof(v_size_el)); - - // Write GQA embedding size - write(&n_embd_v_gqa, sizeof(n_embd_v_gqa)); - - // For each row, we get the element values of each cell - for (uint32_t j = 0; j < n_embd_v_gqa; ++j) { - // Read each range of cells of v_size_el length each into tmp_buf and write out - for (const auto & range : cell_ranges) { - const size_t range_size = range.second - range.first; - const size_t src_offset = (range.first + j * kv_size) * v_size_el; - const size_t buf_size = range_size * v_size_el; - write_tensor_data(kv_self.v_l[il], src_offset, buf_size); - } - } - } - } - } - - void write_kv_cache(const struct llama_context * ctx, llama_seq_id seq_id = -1) { - const struct llama_kv_cache & kv_self = ctx->kv_self; - std::vector> cell_ranges; // ranges, from inclusive, to exclusive - uint32_t cell_count = 0; - - // Count the number of cells with the specified seq_id - // Find all the ranges of cells with this seq id (or all, when -1) - uint32_t cell_range_begin = kv_self.size; - for (uint32_t i = 0; i < kv_self.size; ++i) { - const auto & cell = kv_self.cells[i]; - if ((seq_id == -1 && !cell.is_empty()) || cell.has_seq_id(seq_id)) { - ++cell_count; - if (cell_range_begin == kv_self.size) { - cell_range_begin = i; - } - } else { - if (cell_range_begin != kv_self.size) { - cell_ranges.emplace_back(cell_range_begin, i); - cell_range_begin = kv_self.size; - } - } - } - if (cell_range_begin != kv_self.size) { - cell_ranges.emplace_back(cell_range_begin, kv_self.size); - } - - // DEBUG CHECK: Sum of cell counts in ranges should equal the total cell count - uint32_t cell_count_check = 0; - for (const auto & range : cell_ranges) { - cell_count_check += range.second - range.first; - } - GGML_ASSERT(cell_count == cell_count_check); - - write(&cell_count, sizeof(cell_count)); - - write_kv_cache_meta(kv_self, cell_ranges, seq_id); - write_kv_cache_data(ctx, cell_ranges); - } -}; - -struct llama_data_read { - virtual const uint8_t * read(size_t size) = 0; - virtual void read_to(void * dst, size_t size) = 0; - virtual size_t get_size_read() = 0; - virtual ~llama_data_read() = default; - - void read_string(std::string & str) { - uint32_t str_size; - read_to(&str_size, sizeof(str_size)); - - str.assign((const char *) read(str_size), str_size); - } - - // validate model information - void read_model_info(const struct llama_context * ctx) { - const std::string cur_arch_str = llm_arch_name(ctx->model.arch); - - std::string arch_str; - read_string(arch_str); - if (cur_arch_str != arch_str) { - throw std::runtime_error(format("wrong model arch: '%s' instead of '%s'", arch_str.c_str(), cur_arch_str.c_str())); - } - // TODO: add more info which needs to be identical but which is not verified otherwise - } - - //void read_rng(std::mt19937 & rng) { - // std::string rng_str; - // read_string(rng_str); - - // std::istringstream rng_ss(rng_str); - // rng_ss >> rng; - - // if (rng_ss.fail()) { - // throw std::runtime_error("failed to load RNG state"); - // } - //} - - void read_output_ids(struct llama_context * ctx) { - std::vector output_pos; - - uint32_t n_outputs; - read_to(&n_outputs, sizeof(n_outputs)); - - if (n_outputs > llama_output_reserve(*ctx, n_outputs)) { - throw std::runtime_error("could not reserve outputs"); - } - - if (n_outputs) { - output_pos.resize(n_outputs); - read_to(output_pos.data(), n_outputs * sizeof(int32_t)); - - for (int32_t i = 0; i < (int32_t) output_pos.size(); ++i) { - int32_t id = output_pos[i]; - if ((uint32_t) id >= ctx->cparams.n_batch) { - throw std::runtime_error(format("invalid output id, %d does not fit in batch size of %u", id, ctx->cparams.n_batch)); - } - ctx->output_ids[id] = i; - } - - ctx->n_outputs = n_outputs; - } - } - - void read_logits(struct llama_context * ctx) { - uint64_t logits_size; - read_to(&logits_size, sizeof(logits_size)); - - if (ctx->logits_size < logits_size) { - throw std::runtime_error("logits buffer too small"); - } - - if (logits_size) { - read_to(ctx->logits, logits_size * sizeof(float)); - } - } - - void read_embeddings(struct llama_context * ctx) { - uint64_t embeddings_size; - read_to(&embeddings_size, sizeof(embeddings_size)); - - if (ctx->embd_size < embeddings_size) { - throw std::runtime_error("embeddings buffer too small"); - } - - if (embeddings_size) { - read_to(ctx->embd, embeddings_size * sizeof(float)); - } - } - - bool read_kv_cache_meta(struct llama_context * ctx, uint32_t cell_count, llama_seq_id dest_seq_id = -1) { - struct llama_kv_cache & kv_self = ctx->kv_self; - - if (dest_seq_id != -1) { - // single sequence - - llama_kv_cache_seq_rm(kv_self, dest_seq_id, -1, -1); - - llama_ubatch batch = ctx->sbatch.reserve_ubatch(cell_count, /* has_embd */ false); - batch.n_tokens = cell_count; - batch.n_seq_tokens = cell_count; - batch.n_seqs = 1; - - for (uint32_t i = 0; i < cell_count; ++i) { - llama_pos pos; - uint32_t n_seq_id; - - read_to(&pos, sizeof(pos)); - read_to(&n_seq_id, sizeof(n_seq_id)); - - if (n_seq_id != 0) { - LLAMA_LOG_ERROR("%s: invalid seq_id-agnostic kv cell\n", __func__); - return false; - } - - batch.pos[i] = pos; - } - batch.n_seq_id[0] = 1; - batch.seq_id[0] = &dest_seq_id; - if (!llama_kv_cache_find_slot(kv_self, batch)) { - LLAMA_LOG_ERROR("%s: failed to find available cells in kv cache\n", __func__); - return false; - } - - // DEBUG CHECK: kv_self.head should be our first cell, kv_self.head + cell_count - 1 should be our last cell (verify seq_id and pos values) - // Assume that this is one contiguous block of cells - GGML_ASSERT(kv_self.head + cell_count <= kv_self.size); - GGML_ASSERT(kv_self.cells[kv_self.head].pos == batch.pos[0]); - GGML_ASSERT(kv_self.cells[kv_self.head + cell_count - 1].pos == batch.pos[cell_count - 1]); - GGML_ASSERT(kv_self.cells[kv_self.head].has_seq_id(dest_seq_id)); - GGML_ASSERT(kv_self.cells[kv_self.head + cell_count - 1].has_seq_id(dest_seq_id)); - } else { - // whole KV cache restore - - if (cell_count > kv_self.size) { - LLAMA_LOG_ERROR("%s: not enough cells in kv cache\n", __func__); - return false; - } - - llama_kv_cache_clear(kv_self); - - for (uint32_t i = 0; i < cell_count; ++i) { - llama_kv_cell & cell = kv_self.cells[i]; - - llama_pos pos; - uint32_t n_seq_id; - - read_to(&pos, sizeof(pos)); - read_to(&n_seq_id, sizeof(n_seq_id)); - - cell.pos = pos; - - for (uint32_t j = 0; j < n_seq_id; ++j) { - llama_seq_id seq_id; - read_to(&seq_id, sizeof(seq_id)); - - if (seq_id < 0 || (uint32_t) seq_id >= llama_n_seq_max(ctx)) { - LLAMA_LOG_ERROR("%s: invalid seq_id, %d is out of range [0, %u)\n", __func__, seq_id, llama_n_seq_max(ctx)); - return false; - } - - cell.seq_id.insert(seq_id); - - if (kv_self.recurrent) { - int32_t & tail = kv_self.cells[seq_id].tail; - if (tail != -1) { - LLAMA_LOG_ERROR("%s: duplicate tail for seq_id %d in cell %d and %d\n", __func__, seq_id, i, tail); - return false; + // norm may be automatically assigned to the backend of the previous layer, increasing data transfer between backends + // FIXME: fix in ggml_backend_sched + const bool full_offload = model.params.n_gpu_layers > (int) model.hparams.n_layer; + if (ubatch.n_tokens < 32 || full_offload) { + if (il != -1 && strcmp(name, "norm") == 0) { + const auto & dev_layer = model.dev_layer(il); + for (const auto & backend : backends) { + if (ggml_backend_get_device(backend.get()) == dev_layer) { + if (ggml_backend_supports_op(backend.get(), cur)) { + ggml_backend_sched_set_tensor_backend(sched.get(), cur, backend.get()); } - tail = i; - } - } - } - - kv_self.head = 0; - kv_self.used = cell_count; - } - - if (kv_self.recurrent) { - for (uint32_t i = 0; i < cell_count; ++i) { - uint32_t cell_id = kv_self.head + i; - // make sure the recurrent states will keep their restored state - kv_self.cells[cell_id].src = cell_id; - } - } - - return true; - } - - bool read_kv_cache_data(struct llama_context * ctx, uint32_t cell_count) { - const struct llama_hparams & hparams = ctx->model.hparams; - struct llama_kv_cache & kv_self = ctx->kv_self; - uint32_t v_trans; - uint32_t n_layer; - read_to(&v_trans, sizeof(v_trans)); - read_to(&n_layer, sizeof(n_layer)); - - if (n_layer != hparams.n_layer) { - LLAMA_LOG_ERROR("%s: mismatched layer count (%u instead of %u)\n", __func__, n_layer, hparams.n_layer); - return false; - } - if (cell_count > kv_self.size) { - LLAMA_LOG_ERROR("%s: not enough cells in kv cache to restore state (%u > %u)\n", __func__, cell_count, kv_self.size); - return false; - } - if (kv_self.v_trans != (bool) v_trans) { - LLAMA_LOG_ERROR("%s: incompatible V transposition\n", __func__); - return false; - } - - // For each layer, read the keys for each cell, one row is one cell, read as one contiguous block - for (uint32_t il = 0; il < n_layer; ++il) { - const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa(il) + hparams.n_embd_k_s(); - - // Read type of key - int32_t k_type_i_ref; - read_to(&k_type_i_ref, sizeof(k_type_i_ref)); - const int32_t k_type_i = (int32_t)kv_self.k_l[il]->type; - if (k_type_i != k_type_i_ref) { - LLAMA_LOG_ERROR("%s: mismatched key type (%d != %d, layer %d)\n", __func__, k_type_i, k_type_i_ref, il); - return false; - } - - // Read row size of key - uint64_t k_size_row_ref; - read_to(&k_size_row_ref, sizeof(k_size_row_ref)); - const size_t k_size_row = ggml_row_size(kv_self.k_l[il]->type, n_embd_k_gqa); - if (k_size_row != k_size_row_ref) { - LLAMA_LOG_ERROR("%s: mismatched key row size (%zu != %zu, layer %d)\n", __func__, k_size_row, (size_t) k_size_row_ref, il); - return false; - } - - if (cell_count) { - // Read and set the keys for the whole cell range - ggml_backend_tensor_set(kv_self.k_l[il], read(cell_count * k_size_row), kv_self.head * k_size_row, cell_count * k_size_row); - } - } - - if (!kv_self.v_trans) { - for (uint32_t il = 0; il < n_layer; ++il) { - const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il) + hparams.n_embd_v_s(); - - // Read type of value - int32_t v_type_i_ref; - read_to(&v_type_i_ref, sizeof(v_type_i_ref)); - const int32_t v_type_i = (int32_t)kv_self.v_l[il]->type; - if (v_type_i != v_type_i_ref) { - LLAMA_LOG_ERROR("%s: mismatched value type (%d != %d, layer %d)\n", __func__, v_type_i, v_type_i_ref, il); - return false; - } - - // Read row size of value - uint64_t v_size_row_ref; - read_to(&v_size_row_ref, sizeof(v_size_row_ref)); - const size_t v_size_row = ggml_row_size(kv_self.v_l[il]->type, n_embd_v_gqa); - if (v_size_row != v_size_row_ref) { - LLAMA_LOG_ERROR("%s: mismatched value row size (%zu != %zu, layer %d)\n", __func__, v_size_row, (size_t) v_size_row_ref, il); - return false; - } - - if (cell_count) { - // Read and set the values for the whole cell range - ggml_backend_tensor_set(kv_self.v_l[il], read(cell_count * v_size_row), kv_self.head * v_size_row, cell_count * v_size_row); - } - } - } else { - // For each layer, read the values for each cell (transposed) - for (uint32_t il = 0; il < n_layer; ++il) { - const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il) + hparams.n_embd_v_s(); - - // Read type of value - int32_t v_type_i_ref; - read_to(&v_type_i_ref, sizeof(v_type_i_ref)); - const int32_t v_type_i = (int32_t)kv_self.v_l[il]->type; - if (v_type_i != v_type_i_ref) { - LLAMA_LOG_ERROR("%s: mismatched value type (%d != %d, layer %d)\n", __func__, v_type_i, v_type_i_ref, il); - return false; - } - - // Read element size of value - uint32_t v_size_el_ref; - read_to(&v_size_el_ref, sizeof(v_size_el_ref)); - const size_t v_size_el = ggml_type_size(kv_self.v_l[il]->type); - if (v_size_el != v_size_el_ref) { - LLAMA_LOG_ERROR("%s: mismatched value element size (%zu != %zu, layer %d)\n", __func__, v_size_el, (size_t) v_size_el_ref, il); - return false; - } - - // Read GQA embedding size - uint32_t n_embd_v_gqa_ref; - read_to(&n_embd_v_gqa_ref, sizeof(n_embd_v_gqa_ref)); - if (n_embd_v_gqa != n_embd_v_gqa_ref) { - LLAMA_LOG_ERROR("%s: mismatched GQA embedding size (%u != %u, layer %d)\n", __func__, n_embd_v_gqa, n_embd_v_gqa_ref, il); - return false; - } - - if (cell_count) { - // For each row in the transposed matrix, read the values for the whole cell range - for (uint32_t j = 0; j < n_embd_v_gqa; ++j) { - const size_t dst_offset = (kv_self.head + j * kv_self.size) * v_size_el; - ggml_backend_tensor_set(kv_self.v_l[il], read(cell_count * v_size_el), dst_offset, cell_count * v_size_el); } } } } - return true; - } + }; +} - void read_kv_cache(struct llama_context * ctx, llama_seq_id seq_id = -1) { - uint32_t cell_count; - read_to(&cell_count, sizeof(cell_count)); +// +// state save/load +// - bool res = read_kv_cache_meta(ctx, cell_count, seq_id) && read_kv_cache_data(ctx, cell_count); - - if (!res) { - if (seq_id == -1) { - llama_kv_cache_clear(ctx); - } else { - llama_kv_cache_seq_rm(ctx, seq_id, -1, -1); - } - throw std::runtime_error("failed to restore kv cache"); - } - } -}; - -struct llama_data_write_dummy : llama_data_write { - size_t size_written = 0; - - llama_data_write_dummy() {} +class llama_io_write_dummy : public llama_io_write_i { +public: + llama_io_write_dummy() = default; void write(const void * /* src */, size_t size) override { size_written += size; } - void write_tensor_data(const struct ggml_tensor * /* tensor */, size_t /* offset */, size_t size) override { + void write_tensor(const ggml_tensor * /* tensor */, size_t /* offset */, size_t size) override { size_written += size; } - size_t get_size_written() override { + size_t n_bytes() override { return size_written; } + +private: + size_t size_written = 0; }; -struct llama_data_write_buffer : llama_data_write { - uint8_t * ptr; - size_t buf_size = 0; - size_t size_written = 0; - - llama_data_write_buffer(uint8_t * p, size_t len) : ptr(p), buf_size(len) {} +class llama_io_write_buffer : public llama_io_write_i { +public: + llama_io_write_buffer( + uint8_t * p, size_t len) : ptr(p), buf_size(len) {} void write(const void * src, size_t size) override { if (size > buf_size) { @@ -1406,7 +1731,7 @@ struct llama_data_write_buffer : llama_data_write { buf_size -= size; } - void write_tensor_data(const struct ggml_tensor * tensor, size_t offset, size_t size) override { + void write_tensor(const ggml_tensor * tensor, size_t offset, size_t size) override { if (size > buf_size) { throw std::runtime_error("unexpectedly reached end of buffer"); } @@ -1416,17 +1741,19 @@ struct llama_data_write_buffer : llama_data_write { buf_size -= size; } - size_t get_size_written() override { + size_t n_bytes() override { return size_written; } + +private: + uint8_t * ptr; + size_t buf_size = 0; + size_t size_written = 0; }; -struct llama_data_read_buffer : llama_data_read { - const uint8_t * ptr; - size_t buf_size = 0; - size_t size_read = 0; - - llama_data_read_buffer(const uint8_t * p, size_t len) : ptr(p), buf_size(len) {} +class llama_io_read_buffer : public llama_io_read_i { +public: + llama_io_read_buffer(const uint8_t * p, size_t len) : ptr(p), buf_size(len) {} const uint8_t * read(size_t size) override { const uint8_t * base_ptr = ptr; @@ -1443,40 +1770,44 @@ struct llama_data_read_buffer : llama_data_read { memcpy(dst, read(size), size); } - size_t get_size_read() override { + size_t n_bytes() override { return size_read; } + +private: + const uint8_t * ptr; + size_t buf_size = 0; + size_t size_read = 0; }; -struct llama_data_write_file : llama_data_write { - llama_file * file; - size_t size_written = 0; - std::vector temp_buffer; - - llama_data_write_file(llama_file * f) : file(f) {} +class llama_io_write_file : public llama_io_write_i { +public: + llama_io_write_file(llama_file * f) : file(f) {} void write(const void * src, size_t size) override { file->write_raw(src, size); size_written += size; } - void write_tensor_data(const struct ggml_tensor * tensor, size_t offset, size_t size) override { + void write_tensor(const ggml_tensor * tensor, size_t offset, size_t size) override { temp_buffer.resize(size); ggml_backend_tensor_get(tensor, temp_buffer.data(), offset, size); write(temp_buffer.data(), temp_buffer.size()); } - size_t get_size_written() override { + size_t n_bytes() override { return size_written; } + +private: + llama_file * file; + size_t size_written = 0; + std::vector temp_buffer; }; -struct llama_data_read_file : llama_data_read { - llama_file * file; - size_t size_read = 0; - std::vector temp_buffer; - - llama_data_read_file(llama_file * f) : file(f) {} +class llama_io_read_file : public llama_io_read_i { +public: + llama_io_read_file(llama_file * f) : file(f) {} void read_to(void * dst, size_t size) override { file->read_raw(dst, size); @@ -1489,89 +1820,78 @@ struct llama_data_read_file : llama_data_read { return temp_buffer.data(); } - size_t get_size_read() override { + size_t n_bytes() override { return size_read; } + +private: + llama_file * file; + size_t size_read = 0; + std::vector temp_buffer; }; -/** copy state data into either a buffer or file depending on the passed in context - * - * file context: - * llama_file file("/path", "wb"); - * llama_data_write_file data_ctx(&file); - * llama_state_get_data_internal(ctx, data_ctx); - * - * buffer context: - * std::vector buf(max_size, 0); - * llama_data_write_buffer data_ctx(buf.data(), max_size); - * llama_state_get_data_internal(ctx, data_ctx); - * -*/ -static size_t llama_state_get_data_internal(struct llama_context * ctx, llama_data_write & data_ctx) { - llama_synchronize(ctx); - - data_ctx.write_model_info(ctx); - - // copy outputs - data_ctx.write_output_ids(ctx); - data_ctx.write_logits(ctx); - data_ctx.write_embeddings(ctx); - - data_ctx.write_kv_cache(ctx); - - return data_ctx.get_size_written(); -} - -size_t llama_state_get_data(struct llama_context * ctx, uint8_t * dst, size_t size) { - llama_data_write_buffer data_ctx(dst, size); +size_t llama_context::state_get_size() { + llama_io_write_dummy io; try { - return llama_state_get_data_internal(ctx, data_ctx); - } catch (const std::exception & err) { - LLAMA_LOG_ERROR("%s: error saving state: %s\n", __func__, err.what()); - return 0; - } -} - -// Returns the *actual* size of the state. -// Intended to be used when saving to state to a buffer. -size_t llama_state_get_size(struct llama_context * ctx) { - llama_data_write_dummy data_ctx; - try { - return llama_state_get_data_internal(ctx, data_ctx); + return state_write_data(io); } catch (const std::exception & err) { LLAMA_LOG_ERROR("%s: error getting state size: %s\n", __func__, err.what()); return 0; } } -static size_t llama_state_set_data_internal(struct llama_context * ctx, llama_data_read & data_ctx) { - llama_synchronize(ctx); - - data_ctx.read_model_info(ctx); - - // set outputs - data_ctx.read_output_ids(ctx); - data_ctx.read_logits(ctx); - data_ctx.read_embeddings(ctx); - - data_ctx.read_kv_cache(ctx); - - return data_ctx.get_size_read(); +size_t llama_context::state_get_data(uint8_t * dst, size_t size) { + llama_io_write_buffer io(dst, size); + try { + return state_write_data(io); + } catch (const std::exception & err) { + LLAMA_LOG_ERROR("%s: error saving state: %s\n", __func__, err.what()); + return 0; + } } -// Sets the state reading from the specified source address -size_t llama_state_set_data(struct llama_context * ctx, const uint8_t * src, size_t size) { - llama_data_read_buffer data_ctx(src, size); +size_t llama_context::state_set_data(const uint8_t * src, size_t size) { + llama_io_read_buffer io(src, size); try { - return llama_state_set_data_internal(ctx, data_ctx); + return state_read_data(io); } catch (const std::exception & err) { LLAMA_LOG_ERROR("%s: error loading state: %s\n", __func__, err.what()); return 0; } } -static bool llama_state_load_file_internal(struct llama_context * ctx, const char * path_session, llama_token * tokens_out, size_t n_token_capacity, size_t * n_token_count_out) { - llama_file file(path_session, "rb"); +size_t llama_context::state_seq_get_size(llama_seq_id seq_id) { + llama_io_write_dummy io; + try { + return state_seq_write_data(io, seq_id); + } catch (const std::exception & err) { + LLAMA_LOG_ERROR("%s: error getting state size: %s\n", __func__, err.what()); + return 0; + } +} + +size_t llama_context::state_seq_get_data(llama_seq_id seq_id, uint8_t * dst, size_t size) { + llama_io_write_buffer io(dst, size); + try { + return state_seq_write_data(io, seq_id); + } catch (const std::exception & err) { + LLAMA_LOG_ERROR("%s: error saving state: %s\n", __func__, err.what()); + return 0; + } +} + +size_t llama_context::state_seq_set_data(llama_seq_id seq_id, const uint8_t * src, size_t size) { + llama_io_read_buffer io(src, size); + try { + return state_seq_read_data(io, seq_id); + } catch (const std::exception & err) { + LLAMA_LOG_ERROR("%s: error loading state: %s\n", __func__, err.what()); + return 0; + } +} + +bool llama_context::state_load_file(const char * filepath, llama_token * tokens_out, size_t n_token_capacity, size_t * n_token_count_out) { + llama_file file(filepath, "rb"); // sanity checks { @@ -1601,28 +1921,20 @@ static bool llama_state_load_file_internal(struct llama_context * ctx, const cha { const size_t n_state_size_cur = file.size() - file.tell(); - llama_data_read_file data_ctx(&file); - const size_t n_read = llama_state_set_data_internal(ctx, data_ctx); + llama_io_read_file io( &file); + const size_t n_read = state_read_data(io); if (n_read != n_state_size_cur) { LLAMA_LOG_ERROR("%s: did not read all of the session file data! size %zu, got %zu\n", __func__, n_state_size_cur, n_read); return false; } } + return true; } -bool llama_state_load_file(struct llama_context * ctx, const char * path_session, llama_token * tokens_out, size_t n_token_capacity, size_t * n_token_count_out) { - try { - return llama_state_load_file_internal(ctx, path_session, tokens_out, n_token_capacity, n_token_count_out); - } catch (const std::exception & err) { - LLAMA_LOG_ERROR("%s: error loading session file: %s\n", __func__, err.what()); - return false; - } -} - -static bool llama_state_save_file_internal(struct llama_context * ctx, const char * path_session, const llama_token * tokens, size_t n_token_count) { - llama_file file(path_session, "wb"); +bool llama_context::state_save_file(const char * filepath, const llama_token * tokens, size_t n_token_count) { + llama_file file(filepath, "wb"); file.write_u32(LLAMA_SESSION_MAGIC); file.write_u32(LLAMA_SESSION_VERSION); @@ -1632,82 +1944,13 @@ static bool llama_state_save_file_internal(struct llama_context * ctx, const cha file.write_raw(tokens, sizeof(llama_token) * n_token_count); // save the context state using stream saving - llama_data_write_file data_ctx(&file); - llama_state_get_data_internal(ctx, data_ctx); + llama_io_write_file io(&file); + state_write_data(io); return true; } -bool llama_state_save_file(struct llama_context * ctx, const char * path_session, const llama_token * tokens, size_t n_token_count) { - try { - return llama_state_save_file_internal(ctx, path_session, tokens, n_token_count); - } catch (const std::exception & err) { - LLAMA_LOG_ERROR("%s: error saving session file: %s\n", __func__, err.what()); - return false; - } -} - -static size_t llama_state_seq_get_data_internal(struct llama_context * ctx, llama_data_write & data_ctx, llama_seq_id seq_id) { - llama_synchronize(ctx); - - data_ctx.write_kv_cache(ctx, seq_id); - - return data_ctx.get_size_written(); -} - -size_t llama_state_seq_get_size(struct llama_context * ctx, llama_seq_id seq_id) { - llama_data_write_dummy data_ctx; - return llama_state_seq_get_data_internal(ctx, data_ctx, seq_id); -} - -size_t llama_state_seq_get_data(struct llama_context * ctx, uint8_t * dst, size_t size, llama_seq_id seq_id) { - llama_data_write_buffer data_ctx(dst, size); - try { - return llama_state_seq_get_data_internal(ctx, data_ctx, seq_id); - } catch (const std::exception & err) { - LLAMA_LOG_ERROR("%s: error saving sequence state: %s\n", __func__, err.what()); - return 0; - } -} - -static size_t llama_state_seq_set_data_internal(struct llama_context * ctx, llama_data_read & data_ctx, llama_seq_id dest_seq_id) { - llama_synchronize(ctx); - - data_ctx.read_kv_cache(ctx, dest_seq_id); - - return data_ctx.get_size_read(); -} - -size_t llama_state_seq_set_data(struct llama_context * ctx, const uint8_t * src, size_t size, llama_seq_id dest_seq_id) { - llama_data_read_buffer data_ctx(src, size); - try { - return llama_state_seq_set_data_internal(ctx, data_ctx, dest_seq_id); - } catch (const std::exception & err) { - LLAMA_LOG_ERROR("%s: error loading sequence state: %s\n", __func__, err.what()); - return 0; - } -} - -static size_t llama_state_seq_save_file_internal(struct llama_context * ctx, const char * filepath, llama_seq_id seq_id, const llama_token * tokens, size_t n_token_count) { - llama_file file(filepath, "wb"); - - file.write_u32(LLAMA_STATE_SEQ_MAGIC); - file.write_u32(LLAMA_STATE_SEQ_VERSION); - - // save the prompt - file.write_u32((uint32_t) n_token_count); - file.write_raw(tokens, sizeof(llama_token) * n_token_count); - - // save the context state using stream saving - llama_data_write_file data_ctx(&file); - llama_state_seq_get_data_internal(ctx, data_ctx, seq_id); - - const size_t res = file.tell(); - GGML_ASSERT(res == sizeof(uint32_t) * 3 + sizeof(llama_token) * n_token_count + data_ctx.get_size_written()); - return res; -} - -static size_t llama_state_seq_load_file_internal(struct llama_context * ctx, const char * filepath, llama_seq_id dest_seq_id, llama_token * tokens_out, size_t n_token_capacity, size_t * n_token_count_out) { +size_t llama_context::state_seq_load_file(llama_seq_id seq_id, const char * filepath, llama_token * tokens_out, size_t n_token_capacity, size_t * n_token_count_out) { llama_file file(filepath, "rb"); // version checks @@ -1737,8 +1980,8 @@ static size_t llama_state_seq_load_file_internal(struct llama_context * ctx, con // restore the context state { const size_t state_size = file.size() - file.tell(); - llama_data_read_file data_ctx(&file); - const size_t nread = llama_state_seq_set_data_internal(ctx, data_ctx, dest_seq_id); + llama_io_read_file io(&file); + const size_t nread = state_seq_read_data(io, seq_id); if (!nread) { LLAMA_LOG_ERROR("%s: failed to restore sequence state\n", __func__); return 0; @@ -1750,26 +1993,839 @@ static size_t llama_state_seq_load_file_internal(struct llama_context * ctx, con return file.tell(); } -size_t llama_state_seq_save_file(struct llama_context * ctx, const char * filepath, llama_seq_id seq_id, const llama_token * tokens, size_t n_token_count) { +size_t llama_context::state_seq_save_file(llama_seq_id seq_id, const char * filepath, const llama_token * tokens, size_t n_token_count) { + llama_file file(filepath, "wb"); + + file.write_u32(LLAMA_STATE_SEQ_MAGIC); + file.write_u32(LLAMA_STATE_SEQ_VERSION); + + // save the prompt + file.write_u32((uint32_t) n_token_count); + file.write_raw(tokens, sizeof(llama_token) * n_token_count); + + // save the context state using stream saving + llama_io_write_file io(&file); + state_seq_write_data(io, seq_id); + + const size_t res = file.tell(); + GGML_ASSERT(res == sizeof(uint32_t) * 3 + sizeof(llama_token) * n_token_count + io.n_bytes()); + + return res; +} + +size_t llama_context::state_write_data(llama_io_write_i & io) { + LLAMA_LOG_DEBUG("%s: writing state\n", __func__); + + // write model info + { + LLAMA_LOG_DEBUG("%s: - writing model info\n", __func__); + + const std::string arch_str = llm_arch_name(model.arch); + io.write_string(arch_str); + // TODO: add more model-specific info which should prevent loading the session file if not identical + } + + // write output ids + { + LLAMA_LOG_DEBUG("%s: - writing output ids\n", __func__); + + output_reorder(); + + const auto n_outputs = this->n_outputs; + const auto & output_ids = this->output_ids; + + std::vector w_output_pos; + + GGML_ASSERT(n_outputs <= n_outputs_max); + + w_output_pos.resize(n_outputs); + + // build a more compact representation of the output ids + for (size_t i = 0; i < n_batch(); ++i) { + // map an output id to a position in the batch + int32_t pos = output_ids[i]; + if (pos >= 0) { + GGML_ASSERT(pos < n_outputs); + w_output_pos[pos] = i; + } + } + + io.write(&n_outputs, sizeof(n_outputs)); + + if (n_outputs) { + io.write(w_output_pos.data(), n_outputs * sizeof(int32_t)); + } + } + + // write logits + { + LLAMA_LOG_DEBUG("%s: - writing logits\n", __func__); + + const uint64_t logits_size = std::min((uint64_t) this->logits_size, (uint64_t) n_outputs * model.vocab.n_tokens()); + + io.write(&logits_size, sizeof(logits_size)); + + if (logits_size) { + io.write(logits, logits_size * sizeof(float)); + } + } + + // write embeddings + { + LLAMA_LOG_DEBUG("%s: - writing embeddings\n", __func__); + + const uint64_t embd_size = std::min((uint64_t) this->embd_size, (uint64_t) n_outputs * model.hparams.n_embd); + + io.write(&embd_size, sizeof(embd_size)); + + if (embd_size) { + io.write(embd, embd_size * sizeof(float)); + } + } + + LLAMA_LOG_DEBUG("%s: - writing KV self\n", __func__); + kv_self->state_write(io); + + return io.n_bytes(); +} + +size_t llama_context::state_read_data(llama_io_read_i & io) { + LLAMA_LOG_DEBUG("%s: reading state\n", __func__); + + // read model info + { + LLAMA_LOG_DEBUG("%s: - reading model info\n", __func__); + + const std::string cur_arch_str = llm_arch_name(model.arch); + + std::string arch_str; + io.read_string(arch_str); + if (cur_arch_str != arch_str) { + throw std::runtime_error(format("wrong model arch: '%s' instead of '%s'", arch_str.c_str(), cur_arch_str.c_str())); + } + // TODO: add more info which needs to be identical but which is not verified otherwise + } + + // read output ids + { + LLAMA_LOG_DEBUG("%s: - reading output ids\n", __func__); + + auto n_outputs = this->n_outputs; + io.read_to(&n_outputs, sizeof(n_outputs)); + + if (n_outputs > output_reserve(n_outputs)) { + throw std::runtime_error("could not reserve outputs"); + } + + std::vector output_pos; + + if (n_outputs) { + output_pos.resize(n_outputs); + io.read_to(output_pos.data(), n_outputs * sizeof(int32_t)); + + for (int32_t i = 0; i < (int32_t) output_pos.size(); ++i) { + int32_t id = output_pos[i]; + if ((uint32_t) id >= n_batch()) { + throw std::runtime_error(format("invalid output id, %d does not fit in batch size of %u", id, n_batch())); + } + this->output_ids[id] = i; + } + + this->n_outputs = n_outputs; + } + } + + // read logits + { + LLAMA_LOG_DEBUG("%s: - reading logits\n", __func__); + + uint64_t logits_size; + io.read_to(&logits_size, sizeof(logits_size)); + + if (this->logits_size < logits_size) { + throw std::runtime_error("logits buffer too small"); + } + + if (logits_size) { + io.read_to(this->logits, logits_size * sizeof(float)); + } + } + + // read embeddings + { + LLAMA_LOG_DEBUG("%s: - reading embeddings\n", __func__); + + uint64_t embd_size; + io.read_to(&embd_size, sizeof(embd_size)); + + if (this->embd_size < embd_size) { + throw std::runtime_error("embeddings buffer too small"); + } + + if (embd_size) { + io.read_to(this->embd, embd_size * sizeof(float)); + } + } + + LLAMA_LOG_DEBUG("%s: - reading KV self\n", __func__); + kv_self->state_read(io); + + return io.n_bytes(); +} + +size_t llama_context::state_seq_write_data(llama_io_write_i & io, llama_seq_id seq_id) { + GGML_UNUSED(seq_id); + + kv_self->state_write(io, seq_id); + + return io.n_bytes(); +} + +size_t llama_context::state_seq_read_data(llama_io_read_i & io, llama_seq_id seq_id) { + GGML_UNUSED(seq_id); + + kv_self->state_read(io, seq_id); + + return io.n_bytes(); +} + +// +// perf +// + +llama_perf_context_data llama_context::perf_get_data() const { + llama_perf_context_data data = {}; + + data.t_start_ms = 1e-3 * t_start_us; + data.t_load_ms = 1e-3 * t_load_us; + data.t_p_eval_ms = 1e-3 * t_p_eval_us; + data.t_eval_ms = 1e-3 * t_eval_us; + data.n_p_eval = std::max(1, n_p_eval); + data.n_eval = std::max(1, n_eval); + + return data; +} + +void llama_context::perf_reset() { + t_start_us = ggml_time_us(); + t_eval_us = n_eval = 0; + t_p_eval_us = n_p_eval = 0; +} + +// +// interface implementation +// + +llama_context_params llama_context_default_params() { + llama_context_params result = { + /*.n_ctx =*/ 512, + /*.n_batch =*/ 2048, + /*.n_ubatch =*/ 512, + /*.n_seq_max =*/ 1, + /*.n_threads =*/ GGML_DEFAULT_N_THREADS, // TODO: better default + /*.n_threads_batch =*/ GGML_DEFAULT_N_THREADS, + /*.rope_scaling_type =*/ LLAMA_ROPE_SCALING_TYPE_UNSPECIFIED, + /*.pooling_type =*/ LLAMA_POOLING_TYPE_UNSPECIFIED, + /*.attention_type =*/ LLAMA_ATTENTION_TYPE_UNSPECIFIED, + /*.rope_freq_base =*/ 0.0f, + /*.rope_freq_scale =*/ 0.0f, + /*.yarn_ext_factor =*/ -1.0f, + /*.yarn_attn_factor =*/ 1.0f, + /*.yarn_beta_fast =*/ 32.0f, + /*.yarn_beta_slow =*/ 1.0f, + /*.yarn_orig_ctx =*/ 0, + /*.defrag_thold =*/ -1.0f, + /*.cb_eval =*/ nullptr, + /*.cb_eval_user_data =*/ nullptr, + /*.type_k =*/ GGML_TYPE_F16, + /*.type_v =*/ GGML_TYPE_F16, + /*.logits_all =*/ false, + /*.embeddings =*/ false, + /*.offload_kqv =*/ true, + /*.flash_attn =*/ false, + /*.no_perf =*/ true, + /*.abort_callback =*/ nullptr, + /*.abort_callback_data =*/ nullptr, + }; + + return result; +} + +llama_context * llama_init_from_model( + llama_model * model, + llama_context_params params) { + if (!model) { + LLAMA_LOG_ERROR("%s: model cannot be NULL\n", __func__); + return nullptr; + } + + if (params.n_batch == 0 && params.n_ubatch == 0) { + LLAMA_LOG_ERROR("%s: n_batch and n_ubatch cannot both be zero\n", __func__); + return nullptr; + } + + if (params.n_ctx == 0 && model->hparams.n_ctx_train == 0) { + LLAMA_LOG_ERROR("%s: n_ctx and model->hparams.n_ctx_train cannot both be zero\n", __func__); + return nullptr; + } + + if (params.flash_attn && model->arch == LLM_ARCH_GROK) { + LLAMA_LOG_WARN("%s: flash_attn is not compatible with Grok - forcing off\n", __func__); + params.flash_attn = false; + } + + if (ggml_is_quantized(params.type_v) && !params.flash_attn) { + LLAMA_LOG_ERROR("%s: V cache quantization requires flash_attn\n", __func__); + return nullptr; + } + try { - return llama_state_seq_save_file_internal(ctx, filepath, seq_id, tokens, n_token_count); + auto * ctx = new llama_context(*model, params); + return ctx; + } catch (const std::exception & err) { + LLAMA_LOG_ERROR("%s: failed to initialize the context: %s\n", __func__, err.what()); + } + + return nullptr; +} + +// deprecated +llama_context * llama_new_context_with_model( + llama_model * model, + llama_context_params params) { + return llama_init_from_model(model, params); +} + +void llama_free(llama_context * ctx) { + delete ctx; +} + +uint32_t llama_n_ctx(const llama_context * ctx) { + return ctx->n_ctx(); +} + +uint32_t llama_n_batch(const llama_context * ctx) { + return ctx->n_batch(); +} + +uint32_t llama_n_ubatch(const llama_context * ctx) { + return ctx->n_ubatch(); +} + +uint32_t llama_n_seq_max(const llama_context * ctx) { + return ctx->n_seq_max(); +} + +const llama_model * llama_get_model(const llama_context * ctx) { + return &ctx->get_model(); +} + +llama_kv_cache * llama_get_kv_self(llama_context * ctx) { + return ctx->get_kv_self(); +} + +void llama_kv_self_update(llama_context * ctx) { + ctx->kv_self_update(); +} + +enum llama_pooling_type llama_pooling_type(const llama_context * ctx) { + return ctx->pooling_type(); +} + +void llama_attach_threadpool( + llama_context * ctx, + ggml_threadpool_t threadpool, + ggml_threadpool_t threadpool_batch) { + ctx->attach_threadpool(threadpool, threadpool_batch); +} + +void llama_detach_threadpool(llama_context * ctx) { + ctx->detach_threadpool(); +} + +void llama_set_n_threads(llama_context * ctx, int32_t n_threads, int32_t n_threads_batch) { + ctx->set_n_threads(n_threads, n_threads_batch); +} + +int32_t llama_n_threads(llama_context * ctx) { + return ctx->n_threads(); +} + +int32_t llama_n_threads_batch(llama_context * ctx) { + return ctx->n_threads_batch(); +} + +void llama_set_abort_callback(llama_context * ctx, bool (*abort_callback)(void * data), void * abort_callback_data) { + ctx->set_abort_callback(abort_callback, abort_callback_data); +} + +void llama_set_embeddings(llama_context * ctx, bool embeddings) { + ctx->set_embeddings(embeddings); +} + +void llama_set_causal_attn(llama_context * ctx, bool causal_attn) { + ctx->set_causal_attn(causal_attn); +} + +void llama_set_warmup(llama_context * ctx, bool warmup) { + ctx->set_warmup(warmup); +} + +void llama_synchronize(llama_context * ctx) { + ctx->synchronize(); +} + +float * llama_get_logits(llama_context * ctx) { + ctx->synchronize(); + + return ctx->get_logits(); +} + +float * llama_get_logits_ith(llama_context * ctx, int32_t i) { + ctx->synchronize(); + + return ctx->get_logits_ith(i); +} + +float * llama_get_embeddings(llama_context * ctx) { + ctx->synchronize(); + + return ctx->get_embeddings(); +} + +float * llama_get_embeddings_ith(llama_context * ctx, int32_t i) { + ctx->synchronize(); + + return ctx->get_embeddings_ith(i); +} + +float * llama_get_embeddings_seq(llama_context * ctx, llama_seq_id seq_id) { + ctx->synchronize(); + + return ctx->get_embeddings_seq(seq_id); +} + +// llama adapter API + +int32_t llama_set_adapter_lora( + llama_context * ctx, + llama_adapter_lora * adapter, + float scale) { + ctx->set_adapter_lora(adapter, scale); + + return 0; +} + +int32_t llama_rm_adapter_lora( + llama_context * ctx, + llama_adapter_lora * adapter) { + bool res = ctx->rm_adapter_lora(adapter); + + return res ? 0 : -1; +} + +void llama_clear_adapter_lora(llama_context * ctx) { + ctx->clear_adapter_lora(); +} + +int32_t llama_apply_adapter_cvec( + llama_context * ctx, + const float * data, + size_t len, + int32_t n_embd, + int32_t il_start, + int32_t il_end) { + bool res = ctx->apply_adapter_cvec(data, len, n_embd, il_start, il_end); + + return res ? 0 : -1; +} + +// +// kv cache view +// + +llama_kv_cache_view llama_kv_cache_view_init(const llama_context * ctx, int32_t n_seq_max) { + const auto * kv = ctx->get_kv_self(); + if (kv == nullptr) { + LLAMA_LOG_WARN("%s: the context does not have a KV cache\n", __func__); + return {}; + } + + return llama_kv_cache_view_init(*kv, n_seq_max); +} + +void llama_kv_cache_view_update(const llama_context * ctx, llama_kv_cache_view * view) { + const auto * kv = ctx->get_kv_self(); + if (kv == nullptr) { + LLAMA_LOG_WARN("%s: the context does not have a KV cache\n", __func__); + return; + } + + llama_kv_cache_view_update(view, kv); +} + +// +// kv cache +// + +// deprecated +int32_t llama_get_kv_cache_token_count(const llama_context * ctx) { + return llama_kv_self_n_tokens(ctx); +} + +int32_t llama_kv_self_n_tokens(const llama_context * ctx) { + const auto * kv = ctx->get_kv_self(); + if (!kv) { + return 0; + } + + return kv->get_n_tokens(); +} + +// deprecated +int32_t llama_get_kv_cache_used_cells(const llama_context * ctx) { + return llama_kv_self_used_cells(ctx); +} + +int32_t llama_kv_self_used_cells(const llama_context * ctx) { + const auto * kv = ctx->get_kv_self(); + if (!kv) { + return 0; + } + + return kv->get_used_cells(); +} + +// deprecated +void llama_kv_cache_clear(llama_context * ctx) { + llama_kv_self_clear(ctx); +} + +void llama_kv_self_clear(llama_context * ctx) { + auto * kv = ctx->get_kv_self(); + if (!kv) { + return; + } + + kv->clear(); +} + +// deprecated +bool llama_kv_cache_seq_rm( + llama_context * ctx, + llama_seq_id seq_id, + llama_pos p0, + llama_pos p1) { + return llama_kv_self_seq_rm(ctx, seq_id, p0, p1); +} + +bool llama_kv_self_seq_rm( + llama_context * ctx, + llama_seq_id seq_id, + llama_pos p0, + llama_pos p1) { + auto * kv = ctx->get_kv_self(); + if (!kv) { + return true; + } + + return kv->seq_rm(seq_id, p0, p1); +} + +// deprecated +void llama_kv_cache_seq_cp( + llama_context * ctx, + llama_seq_id seq_id_src, + llama_seq_id seq_id_dst, + llama_pos p0, + llama_pos p1) { + return llama_kv_self_seq_cp(ctx, seq_id_src, seq_id_dst, p0, p1); +} + +void llama_kv_self_seq_cp( + llama_context * ctx, + llama_seq_id seq_id_src, + llama_seq_id seq_id_dst, + llama_pos p0, + llama_pos p1) { + auto * kv = ctx->get_kv_self(); + if (!kv) { + return; + } + + return kv->seq_cp(seq_id_src, seq_id_dst, p0, p1); +} + +// deprecated +void llama_kv_cache_seq_keep( + llama_context * ctx, + llama_seq_id seq_id) { + return llama_kv_self_seq_keep(ctx, seq_id); +} + +void llama_kv_self_seq_keep(llama_context * ctx, llama_seq_id seq_id) { + auto * kv = ctx->get_kv_self(); + if (!kv) { + return; + } + + return kv->seq_keep(seq_id); +} + +// deprecated +void llama_kv_cache_seq_add( + llama_context * ctx, + llama_seq_id seq_id, + llama_pos p0, + llama_pos p1, + llama_pos delta) { + return llama_kv_self_seq_add(ctx, seq_id, p0, p1, delta); +} + +void llama_kv_self_seq_add( + llama_context * ctx, + llama_seq_id seq_id, + llama_pos p0, + llama_pos p1, + llama_pos delta) { + auto * kv = ctx->get_kv_self(); + if (!kv) { + return; + } + + return kv->seq_add(seq_id, p0, p1, delta); +} + +// deprecated +void llama_kv_cache_seq_div( + llama_context * ctx, + llama_seq_id seq_id, + llama_pos p0, + llama_pos p1, + int d) { + return llama_kv_self_seq_div(ctx, seq_id, p0, p1, d); +} + +void llama_kv_self_seq_div( + llama_context * ctx, + llama_seq_id seq_id, + llama_pos p0, + llama_pos p1, + int d) { + auto * kv = ctx->get_kv_self(); + if (!kv) { + return; + } + + return kv->seq_div(seq_id, p0, p1, d); +} + +// deprecated +llama_pos llama_kv_cache_seq_pos_max(llama_context * ctx, llama_seq_id seq_id) { + return llama_kv_self_seq_pos_max(ctx, seq_id); +} + +llama_pos llama_kv_self_seq_pos_max(llama_context * ctx, llama_seq_id seq_id) { + const auto * kv = ctx->get_kv_self(); + if (!kv) { + return 0; + } + + return kv->seq_pos_max(seq_id); +} + +// deprecated +void llama_kv_cache_defrag(llama_context * ctx) { + return llama_kv_self_defrag(ctx); +} + +void llama_kv_self_defrag(llama_context * ctx) { + auto * kv = ctx->get_kv_self(); + if (!kv) { + return; + } + + return kv->defrag(); +} + +// deprecated +bool llama_kv_cache_can_shift(const llama_context * ctx) { + return llama_kv_self_can_shift(ctx); +} + +bool llama_kv_self_can_shift(const llama_context * ctx) { + const auto * kv = ctx->get_kv_self(); + if (!kv) { + return false; + } + + return kv->get_can_shift(); +} + +// deprecated +void llama_kv_cache_update(llama_context * ctx) { + llama_kv_self_update(ctx); +} + +// llama state API + +// deprecated +size_t llama_get_state_size(llama_context * ctx) { + return llama_state_get_size(ctx); +} + +// deprecated +size_t llama_copy_state_data(llama_context * ctx, uint8_t * dst) { + return llama_state_get_data(ctx, dst, -1); +} + +// deprecated +size_t llama_set_state_data(llama_context * ctx, const uint8_t * src) { + return llama_state_set_data(ctx, src, -1); +} + +// deprecated +bool llama_load_session_file(llama_context * ctx, const char * path_session, llama_token * tokens_out, size_t n_token_capacity, size_t * n_token_count_out) { + return llama_state_load_file(ctx, path_session, tokens_out, n_token_capacity, n_token_count_out); +} + +// deprecated +bool llama_save_session_file(llama_context * ctx, const char * path_session, const llama_token * tokens, size_t n_token_count) { + return llama_state_save_file(ctx, path_session, tokens, n_token_count); +} + +// Returns the *actual* size of the state. +// Intended to be used when saving to state to a buffer. +size_t llama_state_get_size(llama_context * ctx) { + return ctx->state_get_size(); +} + +size_t llama_state_get_data(llama_context * ctx, uint8_t * dst, size_t size) { + ctx->synchronize(); + + return ctx->state_get_data(dst, size); +} + +// Sets the state reading from the specified source address +size_t llama_state_set_data(llama_context * ctx, const uint8_t * src, size_t size) { + ctx->synchronize(); + + return ctx->state_set_data(src, size); +} + +bool llama_state_load_file(llama_context * ctx, const char * path_session, llama_token * tokens_out, size_t n_token_capacity, size_t * n_token_count_out) { + ctx->synchronize(); + + try { + return ctx->state_load_file(path_session, tokens_out, n_token_capacity, n_token_count_out); + } catch (const std::exception & err) { + LLAMA_LOG_ERROR("%s: error loading session file: %s\n", __func__, err.what()); + return false; + } +} + +bool llama_state_save_file(llama_context * ctx, const char * path_session, const llama_token * tokens, size_t n_token_count) { + ctx->synchronize(); + + try { + return ctx->state_save_file(path_session, tokens, n_token_count); + } catch (const std::exception & err) { + LLAMA_LOG_ERROR("%s: error saving session file: %s\n", __func__, err.what()); + return false; + } +} + +size_t llama_state_seq_get_size(llama_context * ctx, llama_seq_id seq_id) { + return ctx->state_seq_get_size(seq_id); +} + +size_t llama_state_seq_get_data(llama_context * ctx, uint8_t * dst, size_t size, llama_seq_id seq_id) { + ctx->synchronize(); + + return ctx->state_seq_get_data(seq_id, dst, size); +} + +size_t llama_state_seq_set_data(llama_context * ctx, const uint8_t * src, size_t size, llama_seq_id seq_id) { + ctx->synchronize(); + + return ctx->state_seq_set_data(seq_id, src, size); +} + +size_t llama_state_seq_save_file(llama_context * ctx, const char * filepath, llama_seq_id seq_id, const llama_token * tokens, size_t n_token_count) { + ctx->synchronize(); + + try { + return ctx->state_seq_save_file(seq_id, filepath, tokens, n_token_count); } catch (const std::exception & err) { LLAMA_LOG_ERROR("%s: error saving sequence state file: %s\n", __func__, err.what()); return 0; } } -size_t llama_state_seq_load_file(struct llama_context * ctx, const char * filepath, llama_seq_id dest_seq_id, llama_token * tokens_out, size_t n_token_capacity, size_t * n_token_count_out) { +size_t llama_state_seq_load_file(llama_context * ctx, const char * filepath, llama_seq_id dest_seq_id, llama_token * tokens_out, size_t n_token_capacity, size_t * n_token_count_out) { + ctx->synchronize(); + try { - return llama_state_seq_load_file_internal(ctx, filepath, dest_seq_id, tokens_out, n_token_capacity, n_token_count_out); + return ctx->state_seq_load_file(dest_seq_id, filepath, tokens_out, n_token_capacity, n_token_count_out); } catch (const std::exception & err) { LLAMA_LOG_ERROR("%s: error loading sequence state file: %s\n", __func__, err.what()); return 0; } } -const std::vector> & llama_internal_get_tensor_map( - struct llama_context * ctx -) { - return ctx->model.tensors_by_name; +/// + +int32_t llama_encode( + llama_context * ctx, + llama_batch batch) { + const int ret = ctx->encode(batch); + if (ret != 0) { + LLAMA_LOG_ERROR("%s: failed to encode, ret = %d\n", __func__, ret); + } + + return ret; +} + +int32_t llama_decode( + llama_context * ctx, + llama_batch batch) { + const int ret = ctx->decode(batch); + if (ret != 0) { + LLAMA_LOG_ERROR("%s: failed to decode, ret = %d\n", __func__, ret); + } + + return ret; +} + +// +// perf +// + +llama_perf_context_data llama_perf_context(const llama_context * ctx) { + llama_perf_context_data data = {}; + + if (ctx == nullptr) { + return data; + } + + data = ctx->perf_get_data(); + + return data; +} + +void llama_perf_context_print(const llama_context * ctx) { + const auto data = llama_perf_context(ctx); + + const double t_end_ms = 1e-3 * ggml_time_us(); + + LLAMA_LOG_INFO("%s: load time = %10.2f ms\n", __func__, data.t_load_ms); + LLAMA_LOG_INFO("%s: prompt eval time = %10.2f ms / %5d tokens (%8.2f ms per token, %8.2f tokens per second)\n", + __func__, data.t_p_eval_ms, data.n_p_eval, data.t_p_eval_ms / data.n_p_eval, 1e3 / data.t_p_eval_ms * data.n_p_eval); + LLAMA_LOG_INFO("%s: eval time = %10.2f ms / %5d runs (%8.2f ms per token, %8.2f tokens per second)\n", + __func__, data.t_eval_ms, data.n_eval, data.t_eval_ms / data.n_eval, 1e3 / data.t_eval_ms * data.n_eval); + LLAMA_LOG_INFO("%s: total time = %10.2f ms / %5d tokens\n", __func__, (t_end_ms - data.t_start_ms), (data.n_p_eval + data.n_eval)); +} + +void llama_perf_context_reset(llama_context * ctx) { + ctx->perf_reset(); } diff --git a/src/llama-context.h b/src/llama-context.h index a9268b292..04facb544 100644 --- a/src/llama-context.h +++ b/src/llama-context.h @@ -3,66 +3,213 @@ #include "llama.h" #include "llama-batch.h" #include "llama-cparams.h" -#include "llama-model.h" -#include "llama-kv-cache.h" +#include "llama-graph.h" #include "llama-adapter.h" #include "ggml-cpp.h" #include -#include #include -#include + +struct llama_model; +struct llama_kv_cache; + +class llama_io_read_i; +class llama_io_write_i; struct llama_context { - llama_context(const llama_model & model) - : model(model) - , t_start_us(model.t_start_us) - , t_load_us(model.t_load_us) {} + // init scheduler and compute buffers, reserve worst-case graphs + llama_context( + const llama_model & model, + llama_context_params params); - const struct llama_model & model; + ~llama_context(); - struct llama_cparams cparams; - struct llama_sbatch sbatch; // TODO: revisit if needed - struct llama_kv_cache kv_self; - struct llama_adapter_cvec cvec; + void synchronize(); - std::unordered_map lora; + const llama_model & get_model() const; - std::vector backends; - std::vector> set_n_threads_fns; + uint32_t n_ctx() const; + uint32_t n_ctx_per_seq() const; + uint32_t n_batch() const; + uint32_t n_ubatch() const; + uint32_t n_seq_max() const; - ggml_backend_t backend_cpu = nullptr; + uint32_t n_threads() const; + uint32_t n_threads_batch() const; - ggml_threadpool_t threadpool = nullptr; - ggml_threadpool_t threadpool_batch = nullptr; + llama_kv_cache * get_kv_self(); + const llama_kv_cache * get_kv_self() const; - bool has_evaluated_once = false; + void kv_self_update(); - mutable int64_t t_start_us; - mutable int64_t t_load_us; - mutable int64_t t_p_eval_us = 0; - mutable int64_t t_eval_us = 0; + enum llama_pooling_type pooling_type() const; - mutable int64_t t_compute_start_us = 0; - mutable int64_t n_queued_tokens = 0; + float * get_logits(); + float * get_logits_ith(int32_t i); - mutable int32_t n_p_eval = 0; // number of tokens in eval calls for the prompt (with batch size > 1) - mutable int32_t n_eval = 0; // number of eval calls + float * get_embeddings(); + float * get_embeddings_ith(int32_t i); + float * get_embeddings_seq(llama_seq_id seq_id); - // host buffer for the model output (logits and embeddings) - ggml_backend_buffer_ptr buf_output; + void attach_threadpool( + ggml_threadpool_t threadpool, + ggml_threadpool_t threadpool_batch); + + void detach_threadpool(); + + void set_n_threads(int32_t n_threads, int32_t n_threads_batch); + + void set_abort_callback(bool (*abort_callback)(void * data), void * abort_callback_data); + + void set_embeddings (bool value); + void set_causal_attn(bool value); + void set_warmup(bool value); + + void set_adapter_lora( + llama_adapter_lora * adapter, + float scale); + + bool rm_adapter_lora( + llama_adapter_lora * adapter); + + void clear_adapter_lora(); + + bool apply_adapter_cvec( + const float * data, + size_t len, + int32_t n_embd, + int32_t il_start, + int32_t il_end); + + int encode(llama_batch & inp_batch); + int decode(llama_batch & inp_batch); + + // + // state save/load + // + + size_t state_get_size(); + size_t state_get_data( uint8_t * dst, size_t size); + size_t state_set_data(const uint8_t * src, size_t size); + + size_t state_seq_get_size(llama_seq_id seq_id); + size_t state_seq_get_data(llama_seq_id seq_id, uint8_t * dst, size_t size); + size_t state_seq_set_data(llama_seq_id seq_id, const uint8_t * src, size_t size); + + bool state_load_file( + const char * filepath, + llama_token * tokens_out, + size_t n_token_capacity, + size_t * n_token_count_out); + + bool state_save_file( + const char * filepath, + const llama_token * tokens, + size_t n_token_count); + + size_t state_seq_load_file( + llama_seq_id seq_id, + const char * filepath, + llama_token * tokens_out, + size_t n_token_capacity, + size_t * n_token_count_out); + + size_t state_seq_save_file( + llama_seq_id seq_id, + const char * filepath, + const llama_token * tokens, + size_t n_token_count); + + // + // perf + // + + llama_perf_context_data perf_get_data() const; + void perf_reset(); + +private: + // + // output + // + + // Make sure enough space is available for outputs. + // Returns max number of outputs for which space was reserved. + int32_t output_reserve(int32_t n_outputs); + + // make the outputs have the same order they had in the user-provided batch + // TODO: maybe remove this + void output_reorder(); + + // + // graph + // + + int32_t graph_max_nodes() const; + + // zero-out inputs and create the ctx_compute for the compute graph + ggml_cgraph * graph_init(); + + llm_graph_result_ptr graph_build( + ggml_context * ctx, + ggml_cgraph * gf, + const llama_ubatch & ubatch, + llm_graph_type gtype); + + // returns the result of ggml_backend_sched_graph_compute_async execution + ggml_status graph_compute( + ggml_cgraph * gf, + bool batched); + + llm_graph_cb graph_get_cb() const; + + // used by kv_self_update() + ggml_tensor * build_rope_shift( + ggml_context * ctx0, + ggml_tensor * cur, + ggml_tensor * shift, + ggml_tensor * factors, + float freq_base, + float freq_scale, + ggml_backend_buffer * bbuf) const; + + llm_graph_result_ptr build_kv_self_shift( + ggml_context * ctx0, + ggml_cgraph * gf) const; + + llm_graph_result_ptr build_kv_self_defrag( + ggml_context * ctx0, + ggml_cgraph * gf) const; + + // TODO: read/write lora adapters and cvec + size_t state_write_data(llama_io_write_i & io); + size_t state_read_data (llama_io_read_i & io); + + size_t state_seq_write_data(llama_io_write_i & io, llama_seq_id seq_id); + size_t state_seq_read_data (llama_io_read_i & io, llama_seq_id seq_id); + + // + // members + // + + const llama_model & model; + + llama_cparams cparams; + llama_adapter_cvec cvec; + llama_adapter_loras loras; + llama_sbatch sbatch; + + llama_cross cross; // TODO: tmp for handling cross-attention - need something better probably + + std::unique_ptr kv_self; + + // TODO: remove + bool logits_all = false; // decode output (2-dimensional array: [n_outputs][n_vocab]) size_t logits_size = 0; // capacity (of floats) for logits float * logits = nullptr; - std::vector output_ids; // map batch token positions to ids of the logits and embd buffers - size_t output_size = 0; // capacity (of tokens positions) for the output buffers - int32_t n_outputs = 0; // number of actually-used outputs in the current ubatch or last logical batch - - bool logits_all = false; - // embeddings output (2-dimensional array: [n_outputs][n_embd]) // populated only when pooling_type == LLAMA_POOLING_TYPE_NONE size_t embd_size = 0; // capacity (of floats) for embeddings @@ -72,57 +219,47 @@ struct llama_context { // populated only when pooling_type != LLAMA_POOLING_TYPE_NONE std::map> embd_seq; - // whether we are computing encoder output or decoder output - bool is_encoding = false; + int32_t n_outputs = 0; // number of actually-used outputs in the current ubatch or last logical batch + int32_t n_outputs_max = 0; // capacity (of tokens positions) for the output buffers - // TODO: find a better way to accommodate mutli-dimension position encoding methods - // number of position id each token get, 1 for each token in most cases. - // when using m-rope, it will be 3 position ids per token to representing 3 dimension coordinate. - int n_pos_per_token = 1; + std::vector output_ids; // map batch token positions to ids of the logits and embd buffers - // output of the encoder part of the encoder-decoder models - std::vector embd_enc; - std::vector> seq_ids_enc; - - // memory buffers used to evaluate the model - std::vector buf_compute_meta; ggml_backend_sched_ptr sched; + ggml_backend_t backend_cpu = nullptr; + std::vector backends; + + ggml_context_ptr ctx_compute; + + ggml_threadpool_t threadpool = nullptr; + ggml_threadpool_t threadpool_batch = nullptr; + ggml_abort_callback abort_callback = nullptr; void * abort_callback_data = nullptr; - // input tensors - struct ggml_tensor * inp_tokens; // I32 [n_batch] - struct ggml_tensor * inp_embd; // F32 [n_embd, n_batch] - struct ggml_tensor * inp_pos; // I32 [n_batch] - struct ggml_tensor * inp_out_ids; // I32 [n_outputs] - struct ggml_tensor * inp_KQ_mask; // F32 [kv_size, n_batch] - struct ggml_tensor * inp_KQ_mask_swa; // F32 [kv_size, n_batch] - struct ggml_tensor * inp_K_shift; // I32 [kv_size] - struct ggml_tensor * inp_mean; // F32 [n_batch, n_batch] - struct ggml_tensor * inp_cls; // I32 [n_batch] - struct ggml_tensor * inp_s_copy; // I32 [kv_size] - struct ggml_tensor * inp_s_mask; // F32 [1, n_kv] - struct ggml_tensor * inp_s_seq; // I32 [n_kv, n_batch] - struct ggml_tensor * inp_pos_bucket; // I32 [n_batch|n_kv, n_batch] - struct ggml_tensor * inp_embd_enc; // F32 [n_embd, n_outputs_enc] - struct ggml_tensor * inp_KQ_mask_cross; // F32 [n_outputs_enc, n_batch] + std::vector> set_n_threads_fns; + + // buffer types used for the compute buffer of each backend + std::vector backend_ptrs; + std::vector backend_buft; + + // memory buffers used to evaluate the model + std::vector buf_compute_meta; + + // host buffer for the model output (logits and embeddings) + ggml_backend_buffer_ptr buf_output; + + bool has_evaluated_once = false; + + // perf + mutable int64_t t_start_us = 0; + mutable int64_t t_load_us = 0; + mutable int64_t t_p_eval_us = 0; + mutable int64_t t_eval_us = 0; + + mutable int64_t t_compute_start_us = 0; + mutable int64_t n_queued_tokens = 0; + + mutable int32_t n_p_eval = 0; // number of tokens in eval calls for the prompt (with batch size > 1) + mutable int32_t n_eval = 0; // number of eval calls }; - -// TODO: make these methods of llama_context -void llama_set_k_shift(struct llama_context & lctx); - -void llama_set_s_copy(struct llama_context & lctx); - -void llama_set_inputs(llama_context & lctx, const llama_ubatch & ubatch); - -// Make sure enough space is available for outputs. -// Returns max number of outputs for which space was reserved. -size_t llama_output_reserve(struct llama_context & lctx, size_t n_outputs); - -// make the outputs have the same order they had in the user-provided batch -void llama_output_reorder(struct llama_context & ctx); - -// For internal test use -// TODO: remove -const std::vector> & llama_internal_get_tensor_map(struct llama_context * ctx); diff --git a/src/llama-cparams.h b/src/llama-cparams.h index 252012f3d..30e550f02 100644 --- a/src/llama-cparams.h +++ b/src/llama-cparams.h @@ -29,6 +29,7 @@ struct llama_cparams { bool offload_kqv; bool flash_attn; bool no_perf; + bool warmup; enum llama_pooling_type pooling_type; diff --git a/src/llama-grammar.cpp b/src/llama-grammar.cpp index 9b518d1ac..973b47ae0 100644 --- a/src/llama-grammar.cpp +++ b/src/llama-grammar.cpp @@ -345,194 +345,194 @@ const char * llama_grammar_parser::parse_sequence( size_t last_sym_start = rule.size(); const char * pos = src; - auto handle_repetitions = [&](int min_times, int max_times) { + auto handle_repetitions = [&](int min_times, int max_times) { - if (last_sym_start == rule.size()) { - throw std::runtime_error(std::string("expecting preceding item to */+/?/{ at ") + pos); - } + if (last_sym_start == rule.size()) { + throw std::runtime_error(std::string("expecting preceding item to */+/?/{ at ") + pos); + } - // apply transformation to previous symbol (last_sym_start to end) according to - // the following rewrite rules: - // S{m,n} --> S S S (m times) S'(n-m) - // S'(x) ::= S S'(x-1) | - // (... n-m definitions of these S' rules ...) - // S'(1) ::= S | - // S{m,} --> S S S (m times) S' - // S' ::= S S' | - // S* --> S{0,} - // --> S' ::= S S' | - // S+ --> S{1,} - // --> S S' - // S' ::= S S' | - // S? --> S{0,1} - // --> S' - // S' ::= S | + // apply transformation to previous symbol (last_sym_start to end) according to + // the following rewrite rules: + // S{m,n} --> S S S (m times) S'(n-m) + // S'(x) ::= S S'(x-1) | + // (... n-m definitions of these S' rules ...) + // S'(1) ::= S | + // S{m,} --> S S S (m times) S' + // S' ::= S S' | + // S* --> S{0,} + // --> S' ::= S S' | + // S+ --> S{1,} + // --> S S' + // S' ::= S S' | + // S? --> S{0,1} + // --> S' + // S' ::= S | - llama_grammar_rule prev_rule(rule.begin() + last_sym_start, rule.end()); - if (min_times == 0) { - rule.resize(last_sym_start); - } else { - // Repeat the previous elements (min_times - 1) times - for (int i = 1; i < min_times; i++) { - rule.insert(rule.end(), prev_rule.begin(), prev_rule.end()); - } - } - - uint32_t last_rec_rule_id = 0; - auto n_opt = max_times < 0 ? 1 : max_times - min_times; - - llama_grammar_rule rec_rule(prev_rule); - for (int i = 0; i < n_opt; i++) { - rec_rule.resize(prev_rule.size()); - uint32_t rec_rule_id = generate_symbol_id( rule_name); - if (i > 0 || max_times < 0) { - rec_rule.push_back({LLAMA_GRETYPE_RULE_REF, max_times < 0 ? rec_rule_id : last_rec_rule_id}); - } - rec_rule.push_back({LLAMA_GRETYPE_ALT, 0}); - rec_rule.push_back({LLAMA_GRETYPE_END, 0}); - add_rule( rec_rule_id, rec_rule); - last_rec_rule_id = rec_rule_id; - } - if (n_opt > 0) { - rule.push_back({LLAMA_GRETYPE_RULE_REF, last_rec_rule_id}); - } - }; - - while (*pos) { - if (*pos == '"') { // literal string - pos++; - last_sym_start = rule.size(); - while (*pos != '"') { - if (!*pos) { - throw std::runtime_error("unexpected end of input"); - } - auto char_pair = parse_char(pos); - pos = char_pair.second; - rule.push_back({LLAMA_GRETYPE_CHAR, char_pair.first}); - } - pos = parse_space(pos + 1, is_nested); - } else if (*pos == '[') { // char range(s) - pos++; - enum llama_gretype start_type = LLAMA_GRETYPE_CHAR; - if (*pos == '^') { - pos++; - start_type = LLAMA_GRETYPE_CHAR_NOT; - } - last_sym_start = rule.size(); - while (*pos != ']') { - if (!*pos) { - throw std::runtime_error("unexpected end of input"); - } - auto char_pair = parse_char(pos); - pos = char_pair.second; - enum llama_gretype type = last_sym_start < rule.size() - ? LLAMA_GRETYPE_CHAR_ALT - : start_type; - - rule.push_back({type, char_pair.first}); - if (pos[0] == '-' && pos[1] != ']') { - if (!pos[1]) { - throw std::runtime_error("unexpected end of input"); - } - auto endchar_pair = parse_char(pos + 1); - pos = endchar_pair.second; - rule.push_back({LLAMA_GRETYPE_CHAR_RNG_UPPER, endchar_pair.first}); - } - } - pos = parse_space(pos + 1, is_nested); - } else if (is_word_char(*pos)) { // rule reference - const char * name_end = parse_name(pos); - uint32_t ref_rule_id = get_symbol_id(pos, name_end - pos); - pos = parse_space(name_end, is_nested); - last_sym_start = rule.size(); - rule.push_back({LLAMA_GRETYPE_RULE_REF, ref_rule_id}); - } else if (*pos == '(') { // grouping - // parse nested alternates into synthesized rule - pos = parse_space(pos + 1, true); - uint32_t sub_rule_id = generate_symbol_id(rule_name); - pos = parse_alternates(pos, rule_name, sub_rule_id, true); - last_sym_start = rule.size(); - // output reference to synthesized rule - rule.push_back({LLAMA_GRETYPE_RULE_REF, sub_rule_id}); - if (*pos != ')') { - throw std::runtime_error(std::string("expecting ')' at ") + pos); - } - pos = parse_space(pos + 1, is_nested); - } else if (*pos == '.') { // any char - last_sym_start = rule.size(); - rule.push_back({LLAMA_GRETYPE_CHAR_ANY, 0}); - pos = parse_space(pos + 1, is_nested); - } else if (*pos == '*') { - pos = parse_space(pos + 1, is_nested); - handle_repetitions(0, -1); - } else if (*pos == '+') { - pos = parse_space(pos + 1, is_nested); - handle_repetitions(1, -1); - } else if (*pos == '?') { - pos = parse_space(pos + 1, is_nested); - handle_repetitions(0, 1); - } else if (*pos == '{') { - pos = parse_space(pos + 1, is_nested); - - if (!is_digit_char(*pos)) { - throw std::runtime_error(std::string("expecting an int at ") + pos); - } - const char * int_end = parse_int(pos); - int min_times = std::stoul(std::string(pos, int_end - pos)); - pos = parse_space(int_end, is_nested); - - int max_times = -1; - - if (*pos == '}') { - max_times = min_times; - pos = parse_space(pos + 1, is_nested); - } else if (*pos == ',') { - pos = parse_space(pos + 1, is_nested); - - if (is_digit_char(*pos)) { - const char * int_end = parse_int(pos); - max_times = std::stoul(std::string(pos, int_end - pos)); - pos = parse_space(int_end, is_nested); - } - - if (*pos != '}') { - throw std::runtime_error(std::string("expecting '}' at ") + pos); - } - pos = parse_space(pos + 1, is_nested); - } else { - throw std::runtime_error(std::string("expecting ',' at ") + pos); - } - handle_repetitions(min_times, max_times); - } else { - break; + llama_grammar_rule prev_rule(rule.begin() + last_sym_start, rule.end()); + if (min_times == 0) { + rule.resize(last_sym_start); + } else { + // Repeat the previous elements (min_times - 1) times + for (int i = 1; i < min_times; i++) { + rule.insert(rule.end(), prev_rule.begin(), prev_rule.end()); } } - return pos; + + uint32_t last_rec_rule_id = 0; + auto n_opt = max_times < 0 ? 1 : max_times - min_times; + + llama_grammar_rule rec_rule(prev_rule); + for (int i = 0; i < n_opt; i++) { + rec_rule.resize(prev_rule.size()); + uint32_t rec_rule_id = generate_symbol_id( rule_name); + if (i > 0 || max_times < 0) { + rec_rule.push_back({LLAMA_GRETYPE_RULE_REF, max_times < 0 ? rec_rule_id : last_rec_rule_id}); + } + rec_rule.push_back({LLAMA_GRETYPE_ALT, 0}); + rec_rule.push_back({LLAMA_GRETYPE_END, 0}); + add_rule( rec_rule_id, rec_rule); + last_rec_rule_id = rec_rule_id; + } + if (n_opt > 0) { + rule.push_back({LLAMA_GRETYPE_RULE_REF, last_rec_rule_id}); + } + }; + + while (*pos) { + if (*pos == '"') { // literal string + pos++; + last_sym_start = rule.size(); + while (*pos != '"') { + if (!*pos) { + throw std::runtime_error("unexpected end of input"); + } + auto char_pair = parse_char(pos); + pos = char_pair.second; + rule.push_back({LLAMA_GRETYPE_CHAR, char_pair.first}); + } + pos = parse_space(pos + 1, is_nested); + } else if (*pos == '[') { // char range(s) + pos++; + enum llama_gretype start_type = LLAMA_GRETYPE_CHAR; + if (*pos == '^') { + pos++; + start_type = LLAMA_GRETYPE_CHAR_NOT; + } + last_sym_start = rule.size(); + while (*pos != ']') { + if (!*pos) { + throw std::runtime_error("unexpected end of input"); + } + auto char_pair = parse_char(pos); + pos = char_pair.second; + enum llama_gretype type = last_sym_start < rule.size() + ? LLAMA_GRETYPE_CHAR_ALT + : start_type; + + rule.push_back({type, char_pair.first}); + if (pos[0] == '-' && pos[1] != ']') { + if (!pos[1]) { + throw std::runtime_error("unexpected end of input"); + } + auto endchar_pair = parse_char(pos + 1); + pos = endchar_pair.second; + rule.push_back({LLAMA_GRETYPE_CHAR_RNG_UPPER, endchar_pair.first}); + } + } + pos = parse_space(pos + 1, is_nested); + } else if (is_word_char(*pos)) { // rule reference + const char * name_end = parse_name(pos); + uint32_t ref_rule_id = get_symbol_id(pos, name_end - pos); + pos = parse_space(name_end, is_nested); + last_sym_start = rule.size(); + rule.push_back({LLAMA_GRETYPE_RULE_REF, ref_rule_id}); + } else if (*pos == '(') { // grouping + // parse nested alternates into synthesized rule + pos = parse_space(pos + 1, true); + uint32_t sub_rule_id = generate_symbol_id(rule_name); + pos = parse_alternates(pos, rule_name, sub_rule_id, true); + last_sym_start = rule.size(); + // output reference to synthesized rule + rule.push_back({LLAMA_GRETYPE_RULE_REF, sub_rule_id}); + if (*pos != ')') { + throw std::runtime_error(std::string("expecting ')' at ") + pos); + } + pos = parse_space(pos + 1, is_nested); + } else if (*pos == '.') { // any char + last_sym_start = rule.size(); + rule.push_back({LLAMA_GRETYPE_CHAR_ANY, 0}); + pos = parse_space(pos + 1, is_nested); + } else if (*pos == '*') { + pos = parse_space(pos + 1, is_nested); + handle_repetitions(0, -1); + } else if (*pos == '+') { + pos = parse_space(pos + 1, is_nested); + handle_repetitions(1, -1); + } else if (*pos == '?') { + pos = parse_space(pos + 1, is_nested); + handle_repetitions(0, 1); + } else if (*pos == '{') { + pos = parse_space(pos + 1, is_nested); + + if (!is_digit_char(*pos)) { + throw std::runtime_error(std::string("expecting an int at ") + pos); + } + const char * int_end = parse_int(pos); + int min_times = std::stoul(std::string(pos, int_end - pos)); + pos = parse_space(int_end, is_nested); + + int max_times = -1; + + if (*pos == '}') { + max_times = min_times; + pos = parse_space(pos + 1, is_nested); + } else if (*pos == ',') { + pos = parse_space(pos + 1, is_nested); + + if (is_digit_char(*pos)) { + const char * int_end = parse_int(pos); + max_times = std::stoul(std::string(pos, int_end - pos)); + pos = parse_space(int_end, is_nested); + } + + if (*pos != '}') { + throw std::runtime_error(std::string("expecting '}' at ") + pos); + } + pos = parse_space(pos + 1, is_nested); + } else { + throw std::runtime_error(std::string("expecting ',' at ") + pos); + } + handle_repetitions(min_times, max_times); + } else { + break; + } } + return pos; +} const char * llama_grammar_parser::parse_rule(const char * src) { - const char * name_end = parse_name(src); - const char * pos = parse_space(name_end, false); - size_t name_len = name_end - src; - uint32_t rule_id = get_symbol_id(src, name_len); - const std::string name(src, name_len); + const char * name_end = parse_name(src); + const char * pos = parse_space(name_end, false); + size_t name_len = name_end - src; + uint32_t rule_id = get_symbol_id(src, name_len); + const std::string name(src, name_len); - if (!(pos[0] == ':' && pos[1] == ':' && pos[2] == '=')) { - throw std::runtime_error(std::string("expecting ::= at ") + pos); - } - pos = parse_space(pos + 3, true); - - pos = parse_alternates(pos, name, rule_id, false); - - if (*pos == '\r') { - pos += pos[1] == '\n' ? 2 : 1; - } else if (*pos == '\n') { - pos++; - } else if (*pos) { - throw std::runtime_error(std::string("expecting newline or end at ") + pos); - } - return parse_space(pos, true); + if (!(pos[0] == ':' && pos[1] == ':' && pos[2] == '=')) { + throw std::runtime_error(std::string("expecting ::= at ") + pos); } + pos = parse_space(pos + 3, true); + + pos = parse_alternates(pos, name, rule_id, false); + + if (*pos == '\r') { + pos += pos[1] == '\n' ? 2 : 1; + } else if (*pos == '\n') { + pos++; + } else if (*pos) { + throw std::runtime_error(std::string("expecting newline or end at ") + pos); + } + return parse_space(pos, true); +} bool llama_grammar_parser::parse(const char * src) { try { @@ -969,7 +969,7 @@ struct llama_grammar * llama_grammar_init_impl( /* .awaiting_trigger = */ false, /* .trigger_buffer = */ "", /* .trigger_tokens = */ {}, - /* .trigger_words = */ {}, + /* .trigger_patterns = */ {}, }; } @@ -978,19 +978,15 @@ struct llama_grammar * llama_grammar_init_impl( const char * grammar_str, const char * grammar_root, bool lazy, - const char ** trigger_words, - size_t num_trigger_words, + const char ** trigger_patterns, + size_t num_trigger_patterns, const llama_token * trigger_tokens, size_t num_trigger_tokens) { llama_grammar_parser parser; // if there is a grammar, parse it - if (!parser.parse(grammar_str)) { - return nullptr; - } - - // will be empty (default) if there are parse errors - if (parser.rules.empty()) { + // rules will be empty (default) if there are parse errors + if (!parser.parse(grammar_str) || parser.rules.empty()) { fprintf(stderr, "%s: failed to parse grammar\n", __func__); return nullptr; } @@ -1054,14 +1050,16 @@ struct llama_grammar * llama_grammar_init_impl( } while (true); std::vector vec_trigger_tokens; - std::vector vec_trigger_words; + std::vector vec_trigger_patterns; for (size_t i = 0; i < num_trigger_tokens; i++) { GGML_ASSERT(trigger_tokens != nullptr); vec_trigger_tokens.push_back(trigger_tokens[i]); } - for (size_t i = 0; i < num_trigger_words; i++) { - GGML_ASSERT(trigger_words != nullptr); - vec_trigger_words.push_back(trigger_words[i]); + for (size_t i = 0; i < num_trigger_patterns; i++) { + GGML_ASSERT(trigger_patterns != nullptr); + auto & trigger = vec_trigger_patterns.emplace_back(); + trigger.pattern = trigger_patterns[i]; + trigger.regex = std::regex(trigger.pattern); } // Important: vec_rules has to be moved here, not copied, because stacks contains @@ -1076,7 +1074,7 @@ struct llama_grammar * llama_grammar_init_impl( /* .awaiting_trigger = */ lazy, /* .trigger_buffer = */ "", std::move(vec_trigger_tokens), - std::move(vec_trigger_words), + std::move(vec_trigger_patterns), }; } @@ -1089,7 +1087,7 @@ void llama_grammar_free_impl(struct llama_grammar * grammar) { } struct llama_grammar * llama_grammar_clone_impl(const struct llama_grammar & grammar) { - llama_grammar * result = new llama_grammar { + auto * result = new llama_grammar { grammar.vocab, grammar.rules, grammar.stacks, @@ -1098,7 +1096,7 @@ struct llama_grammar * llama_grammar_clone_impl(const struct llama_grammar & gra grammar.awaiting_trigger, grammar.trigger_buffer, grammar.trigger_tokens, - grammar.trigger_words, + grammar.trigger_patterns, }; // redirect elements in stacks to point to new rules @@ -1173,20 +1171,22 @@ void llama_grammar_accept_impl(struct llama_grammar & grammar, llama_token token LLAMA_LOG_DEBUG("Grammar triggered on token %u (`%s`)", token, piece.c_str()); return; } else { - // TODO: consider a smarter incremental substring search algorithm (store last position to search from). grammar.trigger_buffer += piece; - for (const auto & word : grammar.trigger_words) { - auto pos = grammar.trigger_buffer.find(word); - if (pos != std::string::npos) { + + std::smatch match; + for (const auto & trigger_pattern : grammar.trigger_patterns) { + if (std::regex_match(grammar.trigger_buffer, match, trigger_pattern.regex)) { grammar.awaiting_trigger = false; - auto constrained_str = grammar.trigger_buffer.substr(pos); + // get from the first match to the end of the string + auto constrained_str = grammar.trigger_buffer.substr(match.position(1)); + // std::string constrained_str(match[1].first, grammar.trigger_buffer.end()); grammar.trigger_buffer.clear(); llama_grammar_accept_str(grammar, constrained_str); - LLAMA_LOG_DEBUG("Grammar triggered on word `%s`", word.c_str()); + LLAMA_LOG_DEBUG("Grammar triggered on regex: '%s'\n", constrained_str.c_str()); return; } } - LLAMA_LOG_DEBUG("Grammar still awaiting trigger after token %d (`%s`) (buffer: `%s`)\n", token, piece.c_str(), grammar.trigger_buffer.c_str()); + LLAMA_LOG_DEBUG("Grammar still awaiting trigger after token %d (`%s`)\n", token, piece.c_str()); return; } } diff --git a/src/llama-grammar.h b/src/llama-grammar.h index 252d54d4c..f8c291de9 100644 --- a/src/llama-grammar.h +++ b/src/llama-grammar.h @@ -3,6 +3,7 @@ #include "llama.h" #include +#include #include #include @@ -105,6 +106,11 @@ struct llama_grammar_parser { void print(FILE * file); }; +struct llama_grammar_trigger_pattern { + std::string pattern; + std::regex regex; +}; + struct llama_grammar { // note: allow null vocab for testing (not great) const llama_vocab * vocab; @@ -116,13 +122,16 @@ struct llama_grammar { llama_partial_utf8 partial_utf8; // lazy grammars wait for trigger words or tokens before constraining the sampling. - // we still ahve trigger_tokens for non-lazy grammars to force printing of special trigger tokens. + // we still have trigger_tokens for non-lazy grammars to force printing of special trigger tokens. // (useful e.g. for tool_choice=required) bool lazy = false; bool awaiting_trigger = false; // Initialized to true for lazy grammars only std::string trigger_buffer; // Output buffered by lazy grammar. Will be cleared once trigger is found. std::vector trigger_tokens; // Tokens that trigger a lazy grammar, or tokens to force printing of (even if special). - std::vector trigger_words; + std::vector + trigger_patterns; // Regular expressions that trigger a lazy grammar. Must be a full match of the entire generated + // string, and the grammar will be given the string from the first match group onwards. + }; // @@ -141,8 +150,8 @@ struct llama_grammar * llama_grammar_init_impl( const char * grammar_str, const char * grammar_root, bool lazy, - const char ** trigger_words, - size_t num_trigger_words, + const char ** trigger_patterns, + size_t num_trigger_patterns, const llama_token * trigger_tokens, size_t num_trigger_tokens); diff --git a/src/llama-graph.cpp b/src/llama-graph.cpp new file mode 100644 index 000000000..cd955d63b --- /dev/null +++ b/src/llama-graph.cpp @@ -0,0 +1,1695 @@ +#include "llama-graph.h" + +#include "llama-impl.h" +#include "llama-batch.h" +#include "llama-cparams.h" +#include "llama-kv-cache.h" + +#include +#include +#include + +static int32_t llama_relative_position_bucket(llama_pos x, llama_pos y, uint64_t n_buckets, bool bidirectional) { + // TODO move to hparams if a T5 variant appears that uses a different value + const int64_t max_distance = 128; + + if (bidirectional) { + n_buckets >>= 1; + } + + const int64_t max_exact = n_buckets >> 1; + + int32_t relative_position = x - y; + int32_t relative_bucket = 0; + + if (bidirectional) { + relative_bucket += (relative_position > 0) * n_buckets; + relative_position = abs(relative_position); + } else { + relative_position = -std::min(relative_position, 0); + } + + int32_t relative_position_if_large = floorf(max_exact + logf(1.0 * relative_position / max_exact) * (n_buckets - max_exact) / log(1.0 * max_distance / max_exact)); + relative_position_if_large = std::min(relative_position_if_large, n_buckets - 1); + relative_bucket += (relative_position < max_exact ? relative_position : relative_position_if_large); + + return relative_bucket; +} + +void llm_graph_input_embd::set_input(const llama_ubatch * ubatch) { + if (ubatch->token) { + const int64_t n_tokens = ubatch->n_tokens; + + ggml_backend_tensor_set(tokens, ubatch->token, 0, n_tokens*ggml_element_size(tokens)); + } + + if (ubatch->embd) { + const int64_t n_embd = embd->ne[0]; + const int64_t n_tokens = ubatch->n_tokens; + + ggml_backend_tensor_set(embd, ubatch->embd, 0, n_tokens*n_embd*ggml_element_size(embd)); + } +} + +void llm_graph_input_pos::set_input(const llama_ubatch * ubatch) { + if (ubatch->pos && pos) { + const int64_t n_tokens = ubatch->n_tokens; + + ggml_backend_tensor_set(pos, ubatch->pos, 0, n_tokens*n_pos_per_token*ggml_element_size(pos)); + } +} + +void llm_graph_input_attn_temp::set_input(const llama_ubatch * ubatch) { + if (ubatch->pos && attn_scale) { + const int64_t n_tokens = ubatch->n_tokens; + + std::vector attn_scale_data(n_tokens, 0.0f); + for (int i = 0; i < n_tokens; ++i) { + const float pos = ubatch->pos[i]; + attn_scale_data[i] = std::log( + std::floor((pos + 1.0f) / n_attn_temp_floor_scale) + 1.0 + ) * f_attn_temp_scale + 1.0; + } + + ggml_backend_tensor_set(attn_scale, attn_scale_data.data(), 0, n_tokens*n_pos_per_token*ggml_element_size(attn_scale)); + } +} + +void llm_graph_input_pos_bucket::set_input(const llama_ubatch * ubatch) { + if (pos_bucket) { + const int64_t n_tokens = ubatch->n_tokens; + + GGML_ASSERT(ggml_backend_buffer_is_host(pos_bucket->buffer)); + GGML_ASSERT(!ubatch->equal_seqs); // TODO: use ubatch->n_seqs instead of failing + + int32_t * data = (int32_t *) pos_bucket->data; + + for (int h = 0; h < 1; ++h) { + for (int j = 0; j < n_tokens; ++j) { + for (int i = 0; i < n_tokens; ++i) { + data[h*(n_tokens*n_tokens) + j*n_tokens + i] = llama_relative_position_bucket(ubatch->pos[i], ubatch->pos[j], hparams.n_rel_attn_bkts, true); + } + } + } + } +} + +void llm_graph_input_pos_bucket_kv::set_input(const llama_ubatch * ubatch) { + if (pos_bucket) { + const int64_t n_tokens = ubatch->n_tokens; + + GGML_ASSERT(ggml_backend_buffer_is_host(pos_bucket->buffer)); + GGML_ASSERT(!ubatch->equal_seqs); // TODO: use ubatch->n_seqs instead of failing + + int32_t * data = (int32_t *) pos_bucket->data; + + const int64_t n_kv = kv_self->n; + + for (int h = 0; h < 1; ++h) { + for (int j = 0; j < n_tokens; ++j) { + for (int i = 0; i < n_kv; ++i) { + data[h*(n_kv*n_tokens) + j*n_kv + i] = llama_relative_position_bucket(kv_self->cells[i].pos, ubatch->pos[j], hparams.n_rel_attn_bkts, false); + } + } + } + } +} + +void llm_graph_input_out_ids::set_input(const llama_ubatch * ubatch) { + if (hparams.causal_attn || cparams.pooling_type == LLAMA_POOLING_TYPE_NONE) { + //GGML_ASSERT(out_ids && "every model that can must skip unused outputs"); + + if (!out_ids) { + LLAMA_LOG_WARN("%s: 'out_ids' is not created\n", __func__); + } else { + const int64_t n_tokens = ubatch->n_tokens; + + GGML_ASSERT(ggml_backend_buffer_is_host(out_ids->buffer)); + int32_t * data = (int32_t *) out_ids->data; + + if (n_outputs == n_tokens) { + for (int i = 0; i < n_tokens; ++i) { + data[i] = i; + } + } else if (ubatch->output) { + int32_t n_outputs = 0; + for (int i = 0; i < n_tokens; ++i) { + if (ubatch->output[i]) { + data[n_outputs++] = i; + } + } + // the graph needs to have been passed the correct number of outputs + GGML_ASSERT(n_outputs == n_outputs); + } else if (n_outputs == 1) { + // only keep last output + data[0] = n_tokens - 1; + } else { + GGML_ASSERT(n_outputs == 0); + } + } + } +} + +void llm_graph_input_mean::set_input(const llama_ubatch * ubatch) { + if (cparams.embeddings && cparams.pooling_type == LLAMA_POOLING_TYPE_MEAN) { + const int64_t n_tokens = ubatch->n_tokens; + const int64_t n_seq_tokens = ubatch->n_seq_tokens; + const int64_t n_seqs = ubatch->n_seqs; + + GGML_ASSERT(mean); + GGML_ASSERT(ggml_backend_buffer_is_host(mean->buffer)); + + float * data = (float *) mean->data; + memset(mean->data, 0, n_tokens * n_tokens * ggml_element_size(mean)); + + std::vector sum(n_tokens, 0); + + for (int s = 0; s < n_seqs; ++s) { + const llama_seq_id seq_id = ubatch->seq_id[s][0]; + + // TODO: adapt limits to n_seqs when ubatch->equal_seqs is true + GGML_ASSERT(seq_id < n_tokens && "seq_id cannot be larger than n_tokens with pooling_type == MEAN"); + + sum[seq_id] += ubatch->n_seq_tokens; + } + + std::vector div(n_tokens, 0.0f); + for (int i = 0; i < n_tokens; ++i) { + const uint64_t s = sum[i]; + if (s > 0) { + div[i] = 1.0f/float(s); + } + } + + for (int s = 0; s < n_seqs; ++s) { + const llama_seq_id seq_id = ubatch->seq_id[s][0]; + + for (int i = 0; i < n_seq_tokens; ++i) { + data[seq_id*n_tokens + s*n_seq_tokens + i] = div[seq_id]; + } + } + } +} + +void llm_graph_input_cls::set_input(const llama_ubatch * ubatch) { + if (cparams.embeddings && ( + cparams.pooling_type == LLAMA_POOLING_TYPE_CLS || + cparams.pooling_type == LLAMA_POOLING_TYPE_RANK)) { + const int64_t n_tokens = ubatch->n_tokens; + const int64_t n_seq_tokens = ubatch->n_seq_tokens; + const int64_t n_seqs = ubatch->n_seqs; + + GGML_ASSERT(cls); + GGML_ASSERT(ggml_backend_buffer_is_host(cls->buffer)); + + uint32_t * data = (uint32_t *) cls->data; + memset(cls->data, 0, n_tokens * ggml_element_size(cls)); + + for (int s = 0; s < n_seqs; ++s) { + const llama_seq_id seq_id = ubatch->seq_id[s][0]; + + // TODO: adapt limits to n_seqs when ubatch->equal_seqs is true + GGML_ASSERT(seq_id < n_tokens && "seq_id cannot be larger than n_tokens with pooling_type == CLS or RANK"); + + for (int i = 0; i < n_seq_tokens; ++i) { + const llama_pos pos = ubatch->pos[s*n_seq_tokens + i]; + + if (pos == 0) { + data[seq_id] = s*n_seq_tokens + i; + } + } + } + } + + if (cparams.embeddings && cparams.pooling_type == LLAMA_POOLING_TYPE_LAST) { + const int64_t n_tokens = ubatch->n_tokens; + const int64_t n_seq_tokens = ubatch->n_seq_tokens; + const int64_t n_seqs = ubatch->n_seqs; + + GGML_ASSERT(cls); + GGML_ASSERT(ggml_backend_buffer_is_host(cls->buffer)); + + uint32_t * data = (uint32_t *) cls->data; + memset(cls->data, 0, n_tokens * ggml_element_size(cls)); + + std::vector last_pos(n_tokens, -1); + std::vector last_row(n_tokens, -1); + + for (int s = 0; s < n_seqs; ++s) { + const llama_seq_id seq_id = ubatch->seq_id[s][0]; + + // TODO: adapt limits to n_seqs when ubatch->equal_seqs is true + GGML_ASSERT(seq_id < n_tokens && "seq_id cannot be larger than n_tokens with pooling_type == LAST"); + + for (int i = 0; i < n_seq_tokens; ++i) { + const llama_pos pos = ubatch->pos[s*n_seq_tokens + i]; + + if (pos >= last_pos[seq_id]) { + last_pos[seq_id] = pos; + last_row[seq_id] = s*n_seq_tokens + i; + } + } + } + + for (int i = 0; i < n_tokens; ++i) { + if (last_row[i] >= 0) { + data[i] = last_row[i]; + } + } + } +} + +void llm_graph_input_s_copy::set_input(const llama_ubatch * ubatch) { + GGML_UNUSED(ubatch); + + const int64_t n_kv = kv_self->n; + + if (s_copy) { + GGML_ASSERT(ggml_backend_buffer_is_host(s_copy->buffer)); + int32_t * data = (int32_t *) s_copy->data; + + // assuming copy destinations ALWAYS happen ONLY on the cells between head and head+n + for (uint32_t i = 0; i < n_kv; ++i) { + const uint32_t cell_id = i + kv_self->head; + + ////////////////////////////////////////////// + // TODO: this should not mutate the KV cache ! + llama_kv_cell & kv_cell = const_cast(kv_self)->cells[i]; + + // prevent out-of-bound sources + if (kv_cell.src < 0 || (uint32_t) kv_cell.src >= kv_self->size) { + kv_cell.src = cell_id; + } + + data[i] = kv_cell.src; + + // TODO: do not mutate the KV cache + // ensure copy only happens once + if (kv_cell.src != (int32_t) cell_id) { + kv_cell.src = cell_id; + } + } + } +} + +void llm_graph_input_s_mask::set_input(const llama_ubatch * ubatch) { + GGML_UNUSED(ubatch); + + const int64_t n_kv = kv_self->n; + + if (s_mask) { + GGML_ASSERT(ggml_backend_buffer_is_host(s_mask->buffer)); + float * data = (float *) s_mask->data; + + // clear unused states + for (int i = 0; i < n_kv; ++i) { + const uint32_t cell_id = i + kv_self->head; + + ////////////////////////////////////////////// + // TODO: this should not mutate the KV cache ! + llama_kv_cell & kv_cell = const_cast(kv_self)->cells[i]; + + data[i] = (float) (kv_cell.src >= 0); + + // only clear once + if (kv_cell.src < 0) { + kv_cell.src = cell_id; + } + } + } +} + +void llm_graph_input_cross_embd::set_input(const llama_ubatch * ubatch) { + GGML_UNUSED(ubatch); + + if (cross_embd && !cross->v_embd.empty()) { + assert(cross_embd->type == GGML_TYPE_F32); + + ggml_backend_tensor_set(cross_embd, cross->v_embd.data(), 0, ggml_nbytes(cross_embd)); + } +} + +void llm_graph_input_attn_no_cache::set_input(const llama_ubatch * ubatch) { + if (kq_mask) { + if (cparams.causal_attn) { + const int64_t n_kv = ubatch->n_tokens; + const int64_t n_tokens = ubatch->n_tokens; + const int64_t n_seq_tokens = ubatch->n_seq_tokens; + const int64_t n_seqs = ubatch->n_seqs; + + GGML_ASSERT(ggml_backend_buffer_is_host(kq_mask->buffer)); + float * data = (float *) kq_mask->data; + + for (int h = 0; h < 1; ++h) { + for (int s1 = 0; s1 < n_seqs; ++s1) { + const llama_seq_id seq_id = ubatch->seq_id[s1][0]; + + for (int j = 0; j < n_seq_tokens; ++j) { + const int32_t tj = s1*n_seq_tokens + j; + + for (int s0 = 0; s0 < n_seqs; ++s0) { + for (int i = 0; i < n_seq_tokens; ++i) { + const int32_t ti = s0*n_seq_tokens + i; + float f = -INFINITY; + + for (int s = 0; s < ubatch->n_seq_id[s0]; ++s) { + if (ubatch->seq_id[s0][s] == seq_id && ubatch->pos[ti] <= ubatch->pos[tj]) { + if (hparams.use_alibi) { + f = -std::abs(ubatch->pos[ti] - ubatch->pos[tj]); + } else { + f = 0.0f; + } + break; + } + } + + data[h*(n_kv*n_tokens) + tj*n_kv + ti] = f; + } + } + } + } + } + } else { + const int64_t n_tokens = ubatch->n_tokens; + const int64_t n_seq_tokens = ubatch->n_seq_tokens; + const int64_t n_seqs = ubatch->n_seqs; + const int64_t n_stride = ubatch->n_tokens; + + GGML_ASSERT(ggml_backend_buffer_is_host(kq_mask->buffer)); + + float * data = (float *) kq_mask->data; + + for (int h = 0; h < 1; ++h) { + for (int s1 = 0; s1 < n_seqs; ++s1) { + const llama_seq_id seq_id = ubatch->seq_id[s1][0]; + + for (int j = 0; j < n_seq_tokens; ++j) { + const int32_t tj = s1*n_seq_tokens + j; + + for (int s0 = 0; s0 < n_seqs; ++s0) { + for (int i = 0; i < n_seq_tokens; ++i) { + const int32_t ti = s0*n_seq_tokens + i; + float f = -INFINITY; + + for (int s = 0; s < ubatch->n_seq_id[s0]; ++s) { + if (ubatch->seq_id[s0][s] == seq_id) { + if (hparams.use_alibi) { + f = -std::abs(ubatch->pos[ti] - ubatch->pos[tj]); + } else { + f = 0.0f; + } + break; + } + } + + data[h*(n_tokens*n_tokens) + tj*n_stride + ti] = f; + } + } + + for (int i = n_tokens; i < n_stride; ++i) { + data[h*(n_tokens*n_tokens) + tj*n_stride + i] = -INFINITY; + } + } + } + } + } + } +} + +void llm_graph_input_attn_kv_unified::set_input(const llama_ubatch * ubatch) { + if (self_kq_mask || self_kq_mask_swa) { + const int64_t n_kv = kv_self->n; + const int64_t n_tokens = ubatch->n_tokens; + const int64_t n_seq_tokens = ubatch->n_seq_tokens; + const int64_t n_seqs = ubatch->n_seqs; + + float * data = nullptr; + float * data_swa = nullptr; + + if (self_kq_mask) { + GGML_ASSERT(ggml_backend_buffer_is_host(self_kq_mask->buffer)); + data = (float *) self_kq_mask->data; + } + + if (self_kq_mask_swa) { + GGML_ASSERT(ggml_backend_buffer_is_host(self_kq_mask_swa->buffer)); + data_swa = (float *) self_kq_mask_swa->data; + } + + // Use only the previous KV cells of the correct sequence for each token of the ubatch. + // It's assumed that if a token in the batch has multiple sequences, they are equivalent. + // Example with a cache of 10 tokens, 2 tokens populated in cache and 3 tokens in batch: + // Causal mask: + // xxx------- + // xxxx------ + // xxxxx----- + // Non-causal mask: + // xxxxx----- + // xxxxx----- + // xxxxx----- + // To visualize the mask, see https://github.com/ggml-org/llama.cpp/pull/12615 + for (int h = 0; h < 1; ++h) { + for (int s = 0; s < n_seqs; ++s) { + const llama_seq_id seq_id = ubatch->seq_id[s][0]; + + for (int j = 0; j < n_seq_tokens; ++j) { + const llama_pos pos = ubatch->pos[s*n_seq_tokens + j]; + for (int i = 0; i < n_kv; ++i) { + float f; + // mask the token if: + if (!kv_self->cells[i].has_seq_id(seq_id) // not the correct sequence + || (cparams.causal_attn && kv_self->cells[i].pos > pos) // for causal, mask future tokens + ) { + f = -INFINITY; + } else { + if (hparams.use_alibi) { + f = -std::abs(kv_self->cells[i].pos - pos); + } else { + f = 0.0f; + } + } + + if (data) { + data[h*(n_kv*n_tokens) + s*(n_kv*n_seq_tokens) + j*n_kv + i] = f; + } + + // may need to cut off old tokens for sliding window + // TODO @ngxson : we are currently re-using the swa logic to store the chunked mask, we should rename SWA to something more generic like "aux mask" + if (data_swa) { + if (hparams.n_attn_chunk) { + llama_pos pos_chunk_start = (pos / hparams.n_attn_chunk) * hparams.n_attn_chunk; + if (kv_self->cells[i].pos < pos_chunk_start || pos < pos_chunk_start) { + f = -INFINITY; + } + } else { + if (pos - kv_self->cells[i].pos >= (int32_t)hparams.n_swa) { + f = -INFINITY; + } + } + data_swa[h*(n_kv*n_tokens) + s*(n_kv*n_seq_tokens) + j*n_kv + i] = f; + } + } + } + } + + // mask padded tokens + if (data) { + for (int i = n_tokens; i < GGML_PAD(n_tokens, GGML_KQ_MASK_PAD); ++i) { + for (int j = 0; j < n_kv; ++j) { + data[h*(n_kv*n_tokens) + i*n_kv + j] = -INFINITY; + } + } + } + + // mask padded tokens + if (data_swa) { + for (int i = n_tokens; i < GGML_PAD(n_tokens, GGML_KQ_MASK_PAD); ++i) { + for (int j = 0; j < n_kv; ++j) { + data_swa[h*(n_kv*n_tokens) + i*n_kv + j] = -INFINITY; + } + } + } + } + } +} + +void llm_graph_input_attn_cross::set_input(const llama_ubatch * ubatch) { + if (cross_kq_mask) { + const int64_t n_enc = cross_kq_mask->ne[0]; + const int64_t n_tokens = ubatch->n_tokens; + + GGML_ASSERT(ggml_backend_buffer_is_host(cross_kq_mask->buffer)); + GGML_ASSERT(!ubatch->equal_seqs); // TODO: use ubatch->n_seqs instead of failing + + float * data = (float *) cross_kq_mask->data; + + for (int h = 0; h < 1; ++h) { + for (int j = 0; j < n_tokens; ++j) { + for (int i = 0; i < n_enc; ++i) { + float f = -INFINITY; + for (int s = 0; s < ubatch->n_seq_id[j]; ++s) { + const llama_seq_id seq_id = ubatch->seq_id[j][s]; + if (cross->seq_ids_enc[i].find(seq_id) != cross->seq_ids_enc[i].end()) { + f = 0.0f; + } + } + data[h*(n_enc*n_tokens) + j*n_enc + i] = f; + } + } + + for (int i = n_tokens; i < GGML_PAD(n_tokens, GGML_KQ_MASK_PAD); ++i) { + for (int j = 0; j < n_enc; ++j) { + data[h*(n_enc*n_tokens) + i*n_enc + j] = -INFINITY; + } + } + } + } +} + +// +// llm_graph_context +// + +llm_graph_context::llm_graph_context(const llm_graph_params & params) : + arch (params.arch), + hparams (params.hparams), + cparams (params.cparams), + ubatch (params.ubatch), + n_embd (hparams.n_embd), + n_layer (hparams.n_layer), + n_rot (hparams.n_rot), + n_ctx (cparams.n_ctx), + n_ctx_per_seq (cparams.n_ctx / cparams.n_seq_max), + n_head (hparams.n_head()), + n_head_kv (hparams.n_head_kv()), + n_embd_head_k (hparams.n_embd_head_k), + n_embd_k_gqa (hparams.n_embd_k_gqa()), + n_embd_head_v (hparams.n_embd_head_v), + n_embd_v_gqa (hparams.n_embd_v_gqa()), + n_expert (hparams.n_expert), + n_expert_used (cparams.warmup ? hparams.n_expert : hparams.n_expert_used), + freq_base (cparams.rope_freq_base), + freq_scale (cparams.rope_freq_scale), + ext_factor (cparams.yarn_ext_factor), + attn_factor (cparams.yarn_attn_factor), + beta_fast (cparams.yarn_beta_fast), + beta_slow (cparams.yarn_beta_slow), + norm_eps (hparams.f_norm_eps), + norm_rms_eps (hparams.f_norm_rms_eps), + n_tokens (ubatch.n_tokens), + n_outputs (params.n_outputs), + n_ctx_orig (cparams.n_ctx_orig_yarn), + pooling_type (cparams.pooling_type), + rope_type (hparams.rope_type), + ctx0 (params.ctx), + sched (params.sched), + backend_cpu (params.backend_cpu), + cvec (params.cvec), + loras (params.loras), + memory (params.memory), + cross (params.cross), + cb_func (params.cb), + res (std::make_unique()) { + } + +int64_t llm_graph_context::n_pos_per_token() const { + return arch == LLM_ARCH_QWEN2VL ? 4 : 1; +} + +void llm_graph_context::cb(ggml_tensor * cur, const char * name, int il) const { + if (cb_func) { + cb_func(ubatch, cur, name, il); + } +} + +ggml_tensor * llm_graph_context::build_cvec( + ggml_tensor * cur, + int il) const { + return cvec->apply_to(ctx0, cur, il); +} + +ggml_tensor * llm_graph_context::build_lora_mm( + ggml_tensor * w, + ggml_tensor * cur) const { + ggml_tensor * res = ggml_mul_mat(ctx0, w, cur); + + for (const auto & lora : *loras) { + llama_adapter_lora_weight * lw = lora.first->get_weight(w); + if (lw == nullptr) { + continue; + } + + const float adapter_scale = lora.second; + const float scale = lw->get_scale(lora.first->alpha, adapter_scale); + + ggml_tensor * ab_cur = ggml_mul_mat( + ctx0, lw->b, + ggml_mul_mat(ctx0, lw->a, cur) + ); + + ab_cur = ggml_scale(ctx0, ab_cur, scale); + res = ggml_add(ctx0, res, ab_cur); + } + + return res; +} + +ggml_tensor * llm_graph_context::build_lora_mm_id( + ggml_tensor * w, // ggml_tensor * as + ggml_tensor * cur, // ggml_tensor * b + ggml_tensor * ids) const { + ggml_tensor * res = ggml_mul_mat_id(ctx0, w, cur, ids); + for (const auto & lora : *loras) { + llama_adapter_lora_weight * lw = lora.first->get_weight(w); + if (lw == nullptr) { + continue; + } + + const float alpha = lora.first->alpha; + const float rank = (float) lw->b->ne[0]; + const float scale = alpha ? lora.second * alpha / rank : lora.second; + + ggml_tensor * ab_cur = ggml_mul_mat_id( + ctx0, lw->b, + ggml_mul_mat_id(ctx0, lw->a, cur, ids), + ids + ); + + ab_cur = ggml_scale(ctx0, ab_cur, scale); + res = ggml_add(ctx0, res, ab_cur); + } + + return res; +} + +ggml_tensor * llm_graph_context::build_norm( + ggml_tensor * cur, + ggml_tensor * mw, + ggml_tensor * mb, + llm_norm_type type, + int il) const { + switch (type) { + case LLM_NORM: cur = ggml_norm (ctx0, cur, hparams.f_norm_eps); break; + case LLM_NORM_RMS: cur = ggml_rms_norm(ctx0, cur, hparams.f_norm_rms_eps); break; + case LLM_NORM_GROUP: + { + cur = ggml_reshape_3d(ctx0, cur, cur->ne[0], 1, cur->ne[1]); + cur = ggml_group_norm(ctx0, cur, hparams.n_norm_groups, hparams.f_norm_group_eps); + cur = ggml_reshape_2d(ctx0, cur, cur->ne[0], cur->ne[2]); + } break; + } + + if (mw || mb) { + cb(cur, "norm", il); + } + + if (mw) { + cur = ggml_mul(ctx0, cur, mw); + if (mb) { + cb(cur, "norm_w", il); + } + } + + if (mb) { + cur = ggml_add(ctx0, cur, mb); + } + + return cur; +} + +ggml_tensor * llm_graph_context::build_ffn( + ggml_tensor * cur, + ggml_tensor * up, + ggml_tensor * up_b, + ggml_tensor * up_s, + ggml_tensor * gate, + ggml_tensor * gate_b, + ggml_tensor * gate_s, + ggml_tensor * down, + ggml_tensor * down_b, + ggml_tensor * down_s, + ggml_tensor * act_scales, + llm_ffn_op_type type_op, + llm_ffn_gate_type type_gate, + int il) const { + ggml_tensor * tmp = up ? build_lora_mm(up, cur) : cur; + cb(tmp, "ffn_up", il); + + if (up_b) { + tmp = ggml_add(ctx0, tmp, up_b); + cb(tmp, "ffn_up_b", il); + } + + if (up_s) { + tmp = ggml_mul(ctx0, tmp, up_s); + cb(tmp, "ffn_up_s", il); + } + + if (gate) { + switch (type_gate) { + case LLM_FFN_SEQ: + { + cur = build_lora_mm(gate, tmp); + cb(cur, "ffn_gate", il); + } break; + case LLM_FFN_PAR: + { + cur = build_lora_mm(gate, cur); + cb(cur, "ffn_gate", il); + } break; + } + + if (gate_b) { + cur = ggml_add(ctx0, cur, gate_b); + cb(cur, "ffn_gate_b", il); + } + + if (gate_s) { + cur = ggml_mul(ctx0, cur, gate_s); + cb(cur, "ffn_gate_s", il); + } + + } else { + cur = tmp; + } + + switch (type_op) { + case LLM_FFN_SILU: + { + cur = ggml_silu(ctx0, cur); + cb(cur, "ffn_silu", il); + } break; + case LLM_FFN_GELU: + { + cur = ggml_gelu(ctx0, cur); + cb(cur, "ffn_gelu", il); + if (act_scales != NULL) { + cur = ggml_div(ctx0, cur, act_scales); + cb(cur, "ffn_act", il); + } + } break; + case LLM_FFN_RELU: + { + cur = ggml_relu(ctx0, cur); + cb(cur, "ffn_relu", il); + } break; + case LLM_FFN_RELU_SQR: + { + cur = ggml_relu(ctx0, cur); + cb(cur, "ffn_relu", il); + + cur = ggml_sqr(ctx0, cur); + cb(cur, "ffn_sqr(relu)", il); + } break; + case LLM_FFN_SWIGLU: + { + // Project to 4h. If using swiglu double the output width, see https://arxiv.org/pdf/2002.05202.pdf + int64_t split_point = cur->ne[0] / 2; + ggml_tensor * x0 = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, split_point, cur->ne[1], cur->nb[1], 0)); + ggml_tensor * x1 = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, split_point, cur->ne[1], cur->nb[1], split_point * ggml_element_size(cur))); + + x0 = ggml_silu(ctx0, x0); + cb(cur, "ffn_silu", il); + + cur = ggml_mul(ctx0, x0, x1); + cb(cur, "ffn_mul", il); + } break; + } + + if (type_gate == LLM_FFN_PAR) { + cur = ggml_mul(ctx0, cur, tmp); + cb(cur, "ffn_gate_par", il); + } + + if (down) { + cur = build_lora_mm(down, cur); + } + + if (down_b) { + cb(cur, "ffn_down", il); + } + + if (down_b) { + cur = ggml_add(ctx0, cur, down_b); + } + + if (down_s) { + cur = ggml_mul(ctx0, cur, down_s); + cb(cur, "ffn_down_s", il); + } + + return cur; +} + +ggml_tensor * llm_graph_context::build_moe_ffn( + ggml_tensor * cur, + ggml_tensor * gate_inp, + ggml_tensor * up_exps, + ggml_tensor * gate_exps, + ggml_tensor * down_exps, + ggml_tensor * exp_probs_b, + int64_t n_expert, + int64_t n_expert_used, + llm_ffn_op_type type_op, + bool norm_w, + bool scale_w, + float w_scale, + llama_expert_gating_func_type gating_op, + int il) const { + const int64_t n_embd = cur->ne[0]; + const int64_t n_tokens = cur->ne[1]; + const bool weight_before_ffn = arch == LLM_ARCH_LLAMA4; // for llama4, we apply the sigmoid-ed weights before the FFN + + ggml_tensor * logits = build_lora_mm(gate_inp, cur); // [n_expert, n_tokens] + cb(logits, "ffn_moe_logits", il); + + ggml_tensor * probs = nullptr; + switch (gating_op) { + case LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX: + { + probs = ggml_soft_max(ctx0, logits); // [n_expert, n_tokens] + } break; + case LLAMA_EXPERT_GATING_FUNC_TYPE_SIGMOID: + { + probs = ggml_sigmoid(ctx0, logits); // [n_expert, n_tokens] + } break; + default: + GGML_ABORT("fatal error"); + } + cb(probs, "ffn_moe_probs", il); + + // add experts selection bias - introduced in DeepSeek V3 + // leave probs unbiased as it's later used to get expert weights + ggml_tensor * selection_probs = probs; + if (exp_probs_b != nullptr) { + selection_probs = ggml_add(ctx0, probs, exp_probs_b); + cb(selection_probs, "ffn_moe_probs_biased", il); + } + + // llama4 doesn't have exp_probs_b, and sigmoid is only used after top_k + // see: https://github.com/meta-llama/llama-models/blob/699a02993512fb36936b1b0741e13c06790bcf98/models/llama4/moe.py#L183-L198 + if (arch == LLM_ARCH_LLAMA4) { + selection_probs = logits; + } + + // select experts + ggml_tensor * selected_experts = ggml_top_k(ctx0, selection_probs, n_expert_used); // [n_expert_used, n_tokens] + cb(selected_experts->src[0], "ffn_moe_argsort", il); + cb(selected_experts, "ffn_moe_topk", il); + + ggml_tensor * weights = ggml_get_rows(ctx0, + ggml_reshape_3d(ctx0, probs, 1, n_expert, n_tokens), selected_experts); // [1, n_expert_used, n_tokens] + cb(weights, "ffn_moe_weights", il); + + if (norm_w) { + weights = ggml_reshape_2d(ctx0, weights, n_expert_used, n_tokens); + + ggml_tensor * weights_sum = ggml_sum_rows(ctx0, weights); // [1, n_tokens] + cb(weights_sum, "ffn_moe_weights_sum", il); + + weights = ggml_div(ctx0, weights, weights_sum); // [n_expert_used, n_tokens] + cb(weights, "ffn_moe_weights_norm", il); + + weights = ggml_reshape_3d(ctx0, weights, 1, n_expert_used, n_tokens); + } + if (scale_w) { + weights = ggml_scale(ctx0, weights, w_scale); + cb(weights, "ffn_moe_weights_scaled", il); + } + + cur = ggml_reshape_3d(ctx0, cur, n_embd, 1, n_tokens); + + if (weight_before_ffn) { + // TODO: this is a workaround as we don't yet have a repeat op that takes custom dim (ggml_repeat_4d) + ggml_tensor * repeated = ggml_new_tensor_3d(ctx0, cur->type, n_embd, n_expert_used, n_tokens); + repeated = ggml_repeat(ctx0, cur, repeated); // [n_embd, n_expert_used, n_tokens] + cur = ggml_mul(ctx0, repeated, weights); + cb(cur, "ffn_moe_weighted", il); + } + + ggml_tensor * up = build_lora_mm_id(up_exps, cur, selected_experts); // [n_ff, n_expert_used, n_tokens] + cb(up, "ffn_moe_up", il); + + ggml_tensor * gate = build_lora_mm_id(gate_exps, cur, selected_experts); // [n_ff, n_expert_used, n_tokens] + cb(gate, "ffn_moe_gate", il); + + switch (type_op) { + case LLM_FFN_SILU: + { + gate = ggml_silu(ctx0, gate); + cb(gate, "ffn_moe_silu", il); + } break; + case LLM_FFN_GELU: + { + gate = ggml_gelu(ctx0, gate); + cb(gate, "ffn_moe_gelu", il); + } break; + default: + GGML_ABORT("fatal error"); + } + + ggml_tensor * par = ggml_mul(ctx0, up, gate); // [n_ff, n_expert_used, n_tokens] + cb(par, "ffn_moe_gate_par", il); + + ggml_tensor * experts = build_lora_mm_id(down_exps, par, selected_experts); // [n_embd, n_expert_used, n_tokens] + cb(experts, "ffn_moe_down", il); + + if (!weight_before_ffn) { + experts = ggml_mul(ctx0, experts, weights); + cb(cur, "ffn_moe_weighted", il); + } + + // aggregate experts + ggml_tensor * moe_out = nullptr; + for (int i = 0; i < n_expert_used; ++i) { + ggml_tensor * cur_expert = ggml_view_2d(ctx0, experts, n_embd, n_tokens, + experts->nb[2], i*experts->nb[1]); + + if (i == 0) { + moe_out = cur_expert; + } else { + moe_out = ggml_add(ctx0, moe_out, cur_expert); + } + } + + if (n_expert_used == 1) { + // avoid returning a non-contiguous tensor + moe_out = ggml_cont(ctx0, moe_out); + } + + cb(moe_out, "ffn_moe_out", il); + + return moe_out; +} + +// input embeddings with optional lora +ggml_tensor * llm_graph_context::build_inp_embd(ggml_tensor * tok_embd) const { + const int64_t n_embd = hparams.n_embd; + + auto inp = std::make_unique(); + + ggml_tensor * cur = nullptr; + + if (ubatch.token) { + inp->tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, ubatch.n_tokens); + //cb(inp->tokens, "inp_tokens", -1); + ggml_set_input(inp->tokens); + + cur = ggml_get_rows(ctx0, tok_embd, inp->tokens); + + // apply lora for embedding tokens if needed + for (const auto & lora : *loras) { + llama_adapter_lora_weight * lw = lora.first->get_weight(tok_embd); + if (lw == nullptr) { + continue; + } + + const float adapter_scale = lora.second; + const float scale = lw->get_scale(lora.first->alpha, adapter_scale); + + ggml_tensor * inpL_delta = ggml_scale(ctx0, ggml_mul_mat( + ctx0, lw->b, // non-transposed lora_b + ggml_get_rows(ctx0, lw->a, inp->tokens) + ), scale); + + cur = ggml_add(ctx0, cur, inpL_delta); + } + } else { + inp->embd = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, ubatch.n_tokens); + ggml_set_input(inp->embd); + + cur = inp->embd; + } + + // For Granite architecture + if (hparams.f_embedding_scale != 0.0f) { + cur = ggml_scale(ctx0, cur, hparams.f_embedding_scale); + } + + cb(cur, "inp_embd", -1); + + res->add_input(std::move(inp)); + + return cur; +} + +ggml_tensor * llm_graph_context::build_inp_pos() const { + auto inp = std::make_unique(n_pos_per_token()); + + auto & cur = inp->pos; + + cur = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens*n_pos_per_token()); + ggml_set_input(cur); + + res->add_input(std::move(inp)); + + return cur; +} + +ggml_tensor * llm_graph_context::build_inp_attn_scale() const { + auto inp = std::make_unique(n_pos_per_token(), hparams.n_attn_temp_floor_scale, hparams.f_attn_temp_scale); + + auto & cur = inp->attn_scale; + + cur = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, 1, 1, n_tokens*n_pos_per_token()); + ggml_set_input(cur); + + res->add_input(std::move(inp)); + + return cur; +} + +ggml_tensor * llm_graph_context::build_inp_out_ids() const { + auto inp = std::make_unique(hparams, cparams, n_outputs); + + auto & cur = inp->out_ids; + + cur = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_outputs); + ggml_set_input(cur); + + res->add_input(std::move(inp)); + + return cur; +} + +ggml_tensor * llm_graph_context::build_inp_mean() const { + auto inp = std::make_unique(cparams); + + auto & cur = inp->mean; + + cur = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_tokens, n_tokens); + ggml_set_input(cur); + + res->add_input(std::move(inp)); + + return cur; +} + +ggml_tensor * llm_graph_context::build_inp_cls() const { + auto inp = std::make_unique(cparams); + + auto & cur = inp->cls; + + cur = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens); + ggml_set_input(cur); + + res->add_input(std::move(inp)); + + return cur; +} + +ggml_tensor * llm_graph_context::build_inp_s_copy() const { + const llama_kv_cache_unified * kv_self = static_cast(memory); + + auto inp = std::make_unique(kv_self); + + const auto n_kv = kv_self->n; + + auto & cur = inp->s_copy; + + cur = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_kv); + ggml_set_input(cur); + + res->add_input(std::move(inp)); + + return cur; +} + +ggml_tensor * llm_graph_context::build_inp_s_mask() const { + const llama_kv_cache_unified * kv_self = static_cast(memory); + + auto inp = std::make_unique(kv_self); + + const auto n_kv = kv_self->n; + + auto & cur = inp->s_mask; + + cur = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, 1, n_kv); + ggml_set_input(cur); + + res->add_input(std::move(inp)); + + return cur; +} + +ggml_tensor * llm_graph_context::build_inp_cross_embd() const { + auto inp = std::make_unique(cross); + + auto & cur = inp->cross_embd; + + // if we have the output embeddings from the encoder, use them directly + // TODO: needs more work to be correct, for now just use the tensor shape + //if (cross->t_embd) { + // cur = ggml_view_tensor(ctx0, cross->t_embd); + + // return cur; + //} + + const auto n_embd = !cross->v_embd.empty() ? cross->n_embd : hparams.n_embd; + const auto n_enc = !cross->v_embd.empty() ? cross->n_enc : hparams.n_ctx_train; + + cur = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, n_enc); + ggml_set_input(cur); + + res->add_input(std::move(inp)); + + return cur; +} + +ggml_tensor * llm_graph_context::build_inp_pos_bucket_enc() const { + auto inp = std::make_unique(hparams); + + auto & cur = inp->pos_bucket; + + cur = ggml_new_tensor_2d(ctx0, GGML_TYPE_I32, n_tokens, n_tokens); + ggml_set_input(cur); + + res->add_input(std::move(inp)); + + return cur; +} + +ggml_tensor * llm_graph_context::build_inp_pos_bucket_dec() const { + const llama_kv_cache_unified * kv_self = static_cast(memory); + + auto inp = std::make_unique(hparams, kv_self); + + const auto n_kv = kv_self->n; + + auto & cur = inp->pos_bucket; + + cur = ggml_new_tensor_2d(ctx0, GGML_TYPE_I32, n_kv, n_tokens); + ggml_set_input(cur); + + res->add_input(std::move(inp)); + + return cur; +} + +ggml_tensor * llm_graph_context::build_pos_bias(ggml_tensor * pos_bucket, ggml_tensor * attn_rel_b) const { + ggml_tensor * pos_bucket_1d = ggml_reshape_1d(ctx0, pos_bucket, pos_bucket->ne[0] * pos_bucket->ne[1]); + cb(pos_bucket_1d, "pos_bucket_1d", -1); + + ggml_tensor * pos_bias = ggml_get_rows(ctx0, attn_rel_b, pos_bucket_1d); + + pos_bias = ggml_reshape_3d(ctx0, pos_bias, pos_bias->ne[0], pos_bucket->ne[0], pos_bucket->ne[1]); + pos_bias = ggml_permute (ctx0, pos_bias, 2, 0, 1, 3); + pos_bias = ggml_cont (ctx0, pos_bias); + + cb(pos_bias, "pos_bias", -1); + + return pos_bias; +} + +ggml_tensor * llm_graph_context::build_attn_mha( + ggml_cgraph * gf, + ggml_tensor * q, + ggml_tensor * k, + ggml_tensor * v, + ggml_tensor * kq_b, + ggml_tensor * kq_mask, + bool v_trans, + float kq_scale) const { + //const int64_t n_embd_k_gqa = hparams.n_embd_k_gqa(il); + //const int64_t n_embd_v_gqa = hparams.n_embd_v_gqa(il); + + //const int64_t n_head = hparams.n_head(il); + //const int64_t n_head_kv = hparams.n_head_kv(il); + + //const auto & n_embd_head_k = hparams.n_embd_head_k; + //const auto & n_embd_head_v = hparams.n_embd_head_v; + + const auto n_embd_head_v = v_trans ? v->ne[1] : v->ne[0]; + + const auto n_tokens = q->ne[1]; + const auto n_head = q->ne[2]; + const auto n_kv = k->ne[1]; + + ggml_tensor * cur; + + // TODO: replace hardcoded padding with ggml-provided padding + if (cparams.flash_attn && (n_kv % 256 == 0) && kq_b == nullptr) { + GGML_ASSERT(kq_b == nullptr && "Flash attention does not support KQ bias yet"); + + if (v_trans) { + v = ggml_transpose(ctx0, v); + } + + // this can happen when KV cache is not used (e.g. an embedding model with non-causal attn) + if (k->type == GGML_TYPE_F32) { + k = ggml_cast(ctx0, k, GGML_TYPE_F16); + } + + if (v->type == GGML_TYPE_F32) { + v = ggml_cast(ctx0, v, GGML_TYPE_F16); + } + + cur = ggml_flash_attn_ext(ctx0, q, k, v, kq_mask, kq_scale, hparams.f_max_alibi_bias, + hparams.attn_soft_cap ? hparams.f_attn_logit_softcapping : 0.0f); + + ggml_flash_attn_ext_set_prec(cur, GGML_PREC_F32); + + cur = ggml_reshape_2d(ctx0, cur, n_embd_head_v*n_head, n_tokens); + } else { + ggml_tensor * kq = ggml_mul_mat(ctx0, k, q); + + // note: this op tends to require high floating point range + // while for some models F16 is enough, for others it is not, so we default to F32 here + ggml_mul_mat_set_prec(kq, GGML_PREC_F32); + + if (arch == LLM_ARCH_GROK) { + // need to do the following: + // multiply by attn_output_multiplyer of 0.08838834764831845 + // and then : + // kq = 30 * tanh(kq / 30) + // before the softmax below + + kq = ggml_tanh(ctx0, ggml_scale(ctx0, kq, 0.08838834764831845f/30.0f)); + kq = ggml_scale(ctx0, kq, 30); + } + + if (hparams.attn_soft_cap) { + kq = ggml_scale(ctx0, kq, 1.0f / hparams.f_attn_logit_softcapping); + kq = ggml_tanh (ctx0, kq); + kq = ggml_scale(ctx0, kq, hparams.f_attn_logit_softcapping); + } + + if (kq_b) { + kq = ggml_add(ctx0, kq, kq_b); + } + + kq = ggml_soft_max_ext(ctx0, kq, kq_mask, kq_scale, hparams.f_max_alibi_bias); + + if (!v_trans) { + // note: avoid this branch + v = ggml_cont(ctx0, ggml_transpose(ctx0, v)); + } + + ggml_tensor * kqv = ggml_mul_mat(ctx0, v, kq); + + ggml_tensor * kqv_merged = ggml_permute(ctx0, kqv, 0, 2, 1, 3); + + cur = ggml_cont_2d(ctx0, kqv_merged, n_embd_head_v*n_head, n_tokens); + + if (!cparams.offload_kqv) { + // all nodes between the KV store and the attention output are run on the CPU + ggml_backend_sched_set_tensor_backend(sched, cur, backend_cpu); + } + } + + ggml_build_forward_expand(gf, cur); + + return cur; +} + +llm_graph_input_attn_no_cache * llm_graph_context::build_attn_inp_no_cache() const { + auto inp = std::make_unique(hparams, cparams); + + // note: there is no KV cache, so the number of KV values is equal to the number of tokens in the batch + inp->kq_mask = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_tokens, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD)); + //cb(inp_kq_mask, "KQ_mask", -1); + ggml_set_input(inp->kq_mask); + + inp->kq_mask_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->kq_mask, GGML_TYPE_F16) : inp->kq_mask; + + return (llm_graph_input_attn_no_cache *) res->add_input(std::move(inp)); +} + +ggml_tensor * llm_graph_context::build_attn( + llm_graph_input_attn_no_cache * inp, + ggml_cgraph * gf, + ggml_tensor * wo, + ggml_tensor * wo_b, + ggml_tensor * q_cur, + ggml_tensor * k_cur, + ggml_tensor * v_cur, + ggml_tensor * kq_b, + float kq_scale, + int il) const { + GGML_UNUSED(n_tokens); + + // these nodes are added to the graph together so that they are not reordered + // by doing so, the number of splits in the graph is reduced + ggml_build_forward_expand(gf, q_cur); + ggml_build_forward_expand(gf, k_cur); + ggml_build_forward_expand(gf, v_cur); + + const auto & kq_mask = inp->get_kq_mask(); + + ggml_tensor * q = ggml_permute(ctx0, q_cur, 0, 2, 1, 3); + //cb(q, "q", il); + + ggml_tensor * k = ggml_permute(ctx0, k_cur, 0, 2, 1, 3); + //cb(k, "k", il); + + ggml_tensor * v = ggml_permute(ctx0, v_cur, 0, 2, 1, 3); + //cb(k, "v", il); + + ggml_tensor * cur = build_attn_mha(gf, q, k, v, kq_b, kq_mask, false, kq_scale); + + cb(cur, "kqv_out", il); + + if (wo) { + cur = build_lora_mm(wo, cur); + } + + if (wo_b) { + //cb(cur, "kqv_wo", il); + } + + if (wo_b) { + cur = ggml_add(ctx0, cur, wo_b); + } + + return cur; +} + +llm_graph_input_attn_kv_unified * llm_graph_context::build_attn_inp_kv_unified() const { + const llama_kv_cache_unified * kv_self = static_cast(memory); + + auto inp = std::make_unique(hparams, cparams, kv_self); + + const auto n_kv = kv_self->n; + + inp->self_kq_mask = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD)); + //cb(inp->self_kq_mask, "KQ_mask", -1); + ggml_set_input(inp->self_kq_mask); + + inp->self_kq_mask_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->self_kq_mask, GGML_TYPE_F16) : inp->self_kq_mask; + + if (hparams.n_swa_pattern > 1) { + GGML_ASSERT(hparams.n_swa > 0); + + inp->self_kq_mask_swa = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD)); + //cb(inp->self_kq_mask_swa, "KQ_mask_swa", -1); + ggml_set_input(inp->self_kq_mask_swa); + + inp->self_kq_mask_swa_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->self_kq_mask_swa, GGML_TYPE_F16) : inp->self_kq_mask_swa; + } + + return (llm_graph_input_attn_kv_unified *) res->add_input(std::move(inp)); +} + +ggml_tensor * llm_graph_context::build_attn( + llm_graph_input_attn_kv_unified * inp, + ggml_cgraph * gf, + ggml_tensor * wo, + ggml_tensor * wo_b, + ggml_tensor * q_cur, + ggml_tensor * k_cur, + ggml_tensor * v_cur, + ggml_tensor * kq_b, + float kq_scale, + int il) const { + // these nodes are added to the graph together so that they are not reordered + // by doing so, the number of splits in the graph is reduced + ggml_build_forward_expand(gf, q_cur); + ggml_build_forward_expand(gf, k_cur); + ggml_build_forward_expand(gf, v_cur); + + const llama_kv_cache_unified * kv_self = static_cast(memory); + const auto & n_ctx = cparams.n_ctx; + + const int64_t n_embd_k_gqa = hparams.n_embd_k_gqa(il); + const int64_t n_embd_v_gqa = hparams.n_embd_v_gqa(il); + + const auto n_tokens = q_cur->ne[2]; + + const bool v_trans = !cparams.flash_attn; + + // store to KV cache + { + GGML_ASSERT(!kv_self->recurrent); + + const auto kv_head = kv_self->head; + + GGML_ASSERT(kv_self->size == n_ctx); + + ggml_tensor * k_cache_view = ggml_view_1d(ctx0, kv_self->k_l[il], n_tokens*n_embd_k_gqa, ggml_row_size(kv_self->k_l[il]->type, n_embd_k_gqa)*kv_head); + //cb(k_cache_view, "k_cache_view", il); + + // note: storing RoPE-ed version of K in the KV cache + ggml_build_forward_expand(gf, ggml_cpy(ctx0, k_cur, k_cache_view)); + + v_cur = ggml_reshape_2d(ctx0, v_cur, n_embd_v_gqa, n_tokens); + + ggml_tensor * v_cache_view = nullptr; + + if (!v_trans) { + v_cache_view = ggml_view_1d(ctx0, kv_self->v_l[il], n_tokens*n_embd_v_gqa, ggml_row_size(kv_self->v_l[il]->type, n_embd_v_gqa)*kv_head); + } else { + // note: the V cache is transposed when not using flash attention + v_cache_view = ggml_view_2d(ctx0, kv_self->v_l[il], n_tokens, n_embd_v_gqa, + ( n_ctx)*ggml_element_size(kv_self->v_l[il]), + (kv_head)*ggml_element_size(kv_self->v_l[il])); + + v_cur = ggml_transpose(ctx0, v_cur); + } + //cb(v_cache_view, "v_cache_view", il); + + ggml_build_forward_expand(gf, ggml_cpy(ctx0, v_cur, v_cache_view)); + } + + const bool is_swa = hparams.is_swa(il); + + const auto & kq_mask = is_swa ? inp->get_kq_mask_swa() : inp->get_kq_mask(); + + const auto n_kv = kv_self->n; + + const int64_t n_head_kv = hparams.n_head_kv(il); + + const auto & n_embd_head_k = hparams.n_embd_head_k; + const auto & n_embd_head_v = hparams.n_embd_head_v; + + ggml_tensor * q = ggml_permute(ctx0, q_cur, 0, 2, 1, 3); + //cb(q, "q", il); + + ggml_tensor * k = + ggml_view_3d(ctx0, kv_self->k_l[il], + n_embd_head_k, n_kv, n_head_kv, + ggml_row_size(kv_self->k_l[il]->type, n_embd_k_gqa), + ggml_row_size(kv_self->k_l[il]->type, n_embd_head_k), + 0); + //cb(k, "k", il); + + ggml_tensor * v = !v_trans ? + ggml_view_3d(ctx0, kv_self->v_l[il], + n_embd_head_v, n_kv, n_head_kv, + ggml_row_size(kv_self->v_l[il]->type, n_embd_v_gqa), + ggml_row_size(kv_self->v_l[il]->type, n_embd_head_v), + 0) : + ggml_view_3d(ctx0, kv_self->v_l[il], + n_kv, n_embd_head_v, n_head_kv, + ggml_element_size(kv_self->v_l[il])*n_ctx, + ggml_element_size(kv_self->v_l[il])*n_ctx*n_embd_head_v, + 0); + + ggml_tensor * cur = build_attn_mha(gf, q, k, v, kq_b, kq_mask, v_trans, kq_scale); + cb(cur, "kqv_out", il); + + if (wo) { + cur = build_lora_mm(wo, cur); + } + + if (wo_b) { + //cb(cur, "kqv_wo", il); + } + + if (wo_b) { + cur = ggml_add(ctx0, cur, wo_b); + } + + return cur; +} + +llm_graph_input_attn_cross * llm_graph_context::build_attn_inp_cross() const { + auto inp = std::make_unique(cross); + + const int32_t n_enc = !cross->v_embd.empty() ? cross->n_enc : hparams.n_ctx_train; + + inp->cross_kq_mask = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_enc, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD)); + ggml_set_input(inp->cross_kq_mask); + + inp->cross_kq_mask_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->cross_kq_mask, GGML_TYPE_F16) : inp->cross_kq_mask; + + return (llm_graph_input_attn_cross *) res->add_input(std::move(inp)); +} + +ggml_tensor * llm_graph_context::build_attn( + llm_graph_input_attn_cross * inp, + ggml_cgraph * gf, + ggml_tensor * wo, + ggml_tensor * wo_b, + ggml_tensor * q_cur, + ggml_tensor * k_cur, + ggml_tensor * v_cur, + ggml_tensor * kq_b, + float kq_scale, + int il) const { + // these nodes are added to the graph together so that they are not reordered + // by doing so, the number of splits in the graph is reduced + ggml_build_forward_expand(gf, q_cur); + ggml_build_forward_expand(gf, k_cur); + ggml_build_forward_expand(gf, v_cur); + + const auto & kq_mask = inp->get_kq_mask_cross(); + + ggml_tensor * q = ggml_permute(ctx0, q_cur, 0, 2, 1, 3); + //cb(q, "q", il); + + ggml_tensor * k = ggml_permute(ctx0, k_cur, 0, 2, 1, 3); + //cb(k, "k", il); + + ggml_tensor * v = ggml_permute(ctx0, v_cur, 0, 2, 1, 3); + //cb(k, "v", il); + + ggml_tensor * cur = build_attn_mha(gf, q, k, v, kq_b, kq_mask, false, kq_scale); + + cb(cur, "kqv_out", il); + + if (wo) { + cur = build_lora_mm(wo, cur); + } + + if (wo_b) { + //cb(cur, "kqv_wo", il); + } + + if (wo_b) { + cur = ggml_add(ctx0, cur, wo_b); + } + + return cur; +} + +ggml_tensor * llm_graph_context::build_copy_mask_state( + ggml_cgraph * gf, + ggml_tensor * s, + ggml_tensor * state_copy, + ggml_tensor * state_mask, + int32_t n_state, + int32_t n_seqs) const { + const llama_kv_cache_unified * kv_self = static_cast(memory); + + const auto n_kv = kv_self->n; + const auto kv_head = kv_self->head; + + ggml_tensor * states = ggml_reshape_2d(ctx0, s, n_state, kv_self->size); + + // copy states + // NOTE: assuming the copy destinations are ALL contained between kv_head and kv_head + n_kv + // this shrinks the tensors's ne[1] to n_kv + states = ggml_get_rows(ctx0, states, state_copy); + + // clear states of sequences which are starting at the beginning of this batch + // FIXME: zero-out NANs? + states = ggml_mul(ctx0, states, state_mask); + + // copy states which won't be changed further (between n_seqs and n_kv) + ggml_build_forward_expand(gf, + ggml_cpy(ctx0, + ggml_view_1d(ctx0, states, n_state*(n_kv - n_seqs), (n_seqs )*n_state*ggml_element_size(states)), + ggml_view_1d(ctx0, s, n_state*(n_kv - n_seqs), (kv_head + n_seqs)*n_state*ggml_element_size(s)))); + + // the part of the states that will be used and modified + return ggml_view_2d(ctx0, states, n_state, n_seqs, states->nb[1], 0); +} + +ggml_tensor * llm_graph_context::build_rwkv_token_shift_load( + ggml_cgraph * gf, + ggml_tensor * state_copy, + ggml_tensor * state_mask, + const llama_ubatch & ubatch, + int il) const { + const llama_kv_cache_unified * kv_self = static_cast(memory); + + const auto token_shift_count = hparams.token_shift_count; + + const int64_t n_seqs = ubatch.n_seqs; + + ggml_tensor * token_shift_all = kv_self->k_l[il]; + + ggml_tensor * token_shift = build_copy_mask_state( + gf, token_shift_all, state_copy, state_mask, + hparams.n_embd_k_s(), n_seqs); + + token_shift = ggml_reshape_3d(ctx0, token_shift, hparams.n_embd, token_shift_count, n_seqs); + + return token_shift; +} + +ggml_tensor * llm_graph_context::build_rwkv_token_shift_store( + ggml_tensor * token_shift, + const llama_ubatch & ubatch, + int il) const { + const llama_kv_cache_unified * kv_self = static_cast(memory); + + const auto token_shift_count = hparams.token_shift_count; + const auto n_embd = hparams.n_embd; + + const int64_t n_seqs = ubatch.n_seqs; + + const auto kv_head = kv_self->head; + + return ggml_cpy( + ctx0, + ggml_view_1d(ctx0, token_shift, n_embd * n_seqs * token_shift_count, 0), + ggml_view_1d(ctx0, kv_self->k_l[il], hparams.n_embd_k_s() * n_seqs, hparams.n_embd_k_s() * kv_head * ggml_element_size(kv_self->k_l[il])) + ); +} + +void llm_graph_context::build_pooling( + ggml_cgraph * gf, + ggml_tensor * cls, + ggml_tensor * cls_b, + ggml_tensor * cls_out, + ggml_tensor * cls_out_b) const { + if (!cparams.embeddings) { + return; + } + + ggml_tensor * inp = res->t_embd; + + //// find result_norm tensor for input + //for (int i = ggml_graph_n_nodes(gf) - 1; i >= 0; --i) { + // inp = ggml_graph_node(gf, i); + // if (strcmp(inp->name, "result_norm") == 0 || strcmp(inp->name, "result_embd") == 0) { + // break; + // } + + // inp = nullptr; + //} + + GGML_ASSERT(inp != nullptr && "missing result_norm/result_embd tensor"); + + ggml_tensor * cur; + + switch (pooling_type) { + case LLAMA_POOLING_TYPE_NONE: + { + cur = inp; + } break; + case LLAMA_POOLING_TYPE_MEAN: + { + ggml_tensor * inp_mean = build_inp_mean(); + cur = ggml_mul_mat(ctx0, ggml_cont(ctx0, ggml_transpose(ctx0, inp)), inp_mean); + } break; + case LLAMA_POOLING_TYPE_CLS: + case LLAMA_POOLING_TYPE_LAST: + { + ggml_tensor * inp_cls = build_inp_cls(); + cur = ggml_get_rows(ctx0, inp, inp_cls); + } break; + case LLAMA_POOLING_TYPE_RANK: + { + ggml_tensor * inp_cls = build_inp_cls(); + inp = ggml_get_rows(ctx0, inp, inp_cls); + + // classification head + // https://github.com/huggingface/transformers/blob/5af7d41e49bbfc8319f462eb45253dcb3863dfb7/src/transformers/models/roberta/modeling_roberta.py#L1566 + GGML_ASSERT(cls != nullptr); + GGML_ASSERT(cls_b != nullptr); + + cur = ggml_add (ctx0, ggml_mul_mat(ctx0, cls, inp), cls_b); + cur = ggml_tanh(ctx0, cur); + + // some models don't have `cls_out`, for example: https://huggingface.co/jinaai/jina-reranker-v1-tiny-en + // https://huggingface.co/jinaai/jina-reranker-v1-tiny-en/blob/cb5347e43979c3084a890e3f99491952603ae1b7/modeling_bert.py#L884-L896 + if (cls_out) { + GGML_ASSERT(cls_out_b != nullptr); + + cur = ggml_add (ctx0, ggml_mul_mat(ctx0, cls_out, cur), cls_out_b); + } + } break; + default: + { + GGML_ABORT("unknown pooling type"); + } + } + + cb(cur, "result_embd_pooled", -1); + res->t_embd_pooled = cur; + + ggml_build_forward_expand(gf, cur); +} + diff --git a/src/llama-graph.h b/src/llama-graph.h new file mode 100644 index 000000000..5b6618f9e --- /dev/null +++ b/src/llama-graph.h @@ -0,0 +1,592 @@ +#pragma once + +#include "llama-arch.h" +#include "llama-hparams.h" +#include "llama-adapter.h" + +#include +#include +#include +#include +#include + +struct ggml_cgraph; +struct ggml_context; +struct ggml_tensor; + +struct llama_ubatch; +struct llama_cparams; + +class llama_memory_i; +class llama_kv_cache_unified; + +// certain models (typically multi-modal) can produce different types of graphs +enum llm_graph_type { + LLM_GRAPH_TYPE_DEFAULT, + LLM_GRAPH_TYPE_ENCODER, + LLM_GRAPH_TYPE_DECODER, +}; + +enum llm_ffn_op_type { + LLM_FFN_SILU, + LLM_FFN_GELU, + LLM_FFN_RELU, + LLM_FFN_RELU_SQR, + LLM_FFN_SWIGLU, +}; + +enum llm_ffn_gate_type { + LLM_FFN_SEQ, + LLM_FFN_PAR, // ffn_gate is parallel to ffn_up +}; + +enum llm_norm_type { + LLM_NORM, + LLM_NORM_RMS, + LLM_NORM_GROUP, +}; + +// TODO: tmp - need something better to pass the data from the encoder to the decoder +struct llama_cross { + // the output embeddings from the encoder as a ggml tensor + // TODO: this needs more work to be correct, for now copy the embeddings data to host memory + // ref: https://github.com/ggml-org/llama.cpp/pull/11213#discussion_r1969892524 + //ggml_tensor * t_embd = nullptr; + + int64_t n_embd = 0; + int64_t n_enc = 0; + + // embeddings data copied to host memory (tmp) + std::vector v_embd; + + // needed to construct the cross-attention mask in the decoder + std::vector> seq_ids_enc; +}; + +// +// llm_graph_input +// + +class llm_graph_input_i { +public: + virtual ~llm_graph_input_i() = default; + + virtual void set_input(const llama_ubatch * ubatch) = 0; +}; + +using llm_graph_input_ptr = std::unique_ptr; + + +class llm_graph_input_embd : public llm_graph_input_i { +public: + llm_graph_input_embd() = default; + virtual ~llm_graph_input_embd() = default; + + void set_input(const llama_ubatch * ubatch) override; + + ggml_tensor * tokens = nullptr; // I32 [n_batch] + ggml_tensor * embd = nullptr; // F32 [n_embd, n_batch] +}; + +class llm_graph_input_pos : public llm_graph_input_i { +public: + llm_graph_input_pos(int64_t n_pos_per_token) : n_pos_per_token(n_pos_per_token) {} + virtual ~llm_graph_input_pos() = default; + + void set_input(const llama_ubatch * ubatch) override; + + ggml_tensor * pos = nullptr; // I32 [n_batch] + + const int64_t n_pos_per_token = 1; +}; + +// temperature tuning, used by llama4 +class llm_graph_input_attn_temp : public llm_graph_input_i { +public: + llm_graph_input_attn_temp(int64_t n_pos_per_token, uint32_t n_attn_temp_floor_scale, float f_attn_temp_scale) + : n_pos_per_token(n_pos_per_token), n_attn_temp_floor_scale(n_attn_temp_floor_scale), f_attn_temp_scale(f_attn_temp_scale) {} + virtual ~llm_graph_input_attn_temp() = default; + + void set_input(const llama_ubatch * ubatch) override; + + ggml_tensor * attn_scale = nullptr; // F32 [n_batch] + + const int64_t n_pos_per_token = 1; + + const uint32_t n_attn_temp_floor_scale; + const float f_attn_temp_scale; +}; + +class llm_graph_input_pos_bucket : public llm_graph_input_i { +public: + llm_graph_input_pos_bucket(const llama_hparams & hparams) : hparams(hparams) {} + virtual ~llm_graph_input_pos_bucket() = default; + + void set_input(const llama_ubatch * ubatch) override; + + ggml_tensor * pos_bucket = nullptr; // I32 [n_batch, n_batch] + + const llama_hparams & hparams; +}; + +class llm_graph_input_pos_bucket_kv : public llm_graph_input_i { +public: + llm_graph_input_pos_bucket_kv( + const llama_hparams & hparams, + const llama_kv_cache_unified * kv_self) : hparams(hparams), kv_self(kv_self) {} + virtual ~llm_graph_input_pos_bucket_kv() = default; + + void set_input(const llama_ubatch * ubatch) override; + + ggml_tensor * pos_bucket = nullptr; // I32 [n_kv, n_batch] + + const llama_hparams & hparams; + const llama_kv_cache_unified * kv_self; +}; + +class llm_graph_input_out_ids : public llm_graph_input_i { +public: + llm_graph_input_out_ids( + const llama_hparams & hparams, + const llama_cparams & cparams, + int32_t n_outputs) : hparams(hparams), cparams(cparams), n_outputs(n_outputs) {} + virtual ~llm_graph_input_out_ids() = default; + + void set_input(const llama_ubatch * ubatch) override; + + ggml_tensor * out_ids; // I32 [n_outputs] + + const llama_hparams & hparams; + const llama_cparams & cparams; + + const int32_t n_outputs; +}; + +class llm_graph_input_mean : public llm_graph_input_i { +public: + llm_graph_input_mean(const llama_cparams & cparams) : cparams(cparams) {} + virtual ~llm_graph_input_mean() = default; + + void set_input(const llama_ubatch * ubatch) override; + + ggml_tensor * mean; // F32 [n_batch, n_batch] + + const llama_cparams & cparams; +}; + +class llm_graph_input_cls : public llm_graph_input_i { +public: + llm_graph_input_cls(const llama_cparams & cparams) : cparams(cparams) {} + virtual ~llm_graph_input_cls() = default; + + void set_input(const llama_ubatch * ubatch) override; + + ggml_tensor * cls; // I32 [n_batch] + + const llama_cparams & cparams; +}; + +class llm_graph_input_s_copy : public llm_graph_input_i { +public: + llm_graph_input_s_copy(const llama_kv_cache_unified * kv_self) : kv_self(kv_self) {} + virtual ~llm_graph_input_s_copy() = default; + + void set_input(const llama_ubatch * ubatch) override; + + ggml_tensor * s_copy; // I32 [kv_size] + + const llama_kv_cache_unified * kv_self; +}; + +class llm_graph_input_s_mask : public llm_graph_input_i { +public: + llm_graph_input_s_mask(const llama_kv_cache_unified * kv_self) : kv_self(kv_self) {} + virtual ~llm_graph_input_s_mask() = default; + + void set_input(const llama_ubatch * ubatch) override; + + ggml_tensor * s_mask; // F32 [1, n_kv] + + const llama_kv_cache_unified * kv_self; +}; + +class llm_graph_input_cross_embd : public llm_graph_input_i { +public: + llm_graph_input_cross_embd( + const llama_cross * cross) : cross(cross) {} + virtual ~llm_graph_input_cross_embd() = default; + + void set_input(const llama_ubatch * ubatch) override; + + ggml_tensor * cross_embd; // F32 [n_embd, n_outputs_enc] + + const llama_cross * cross; +}; + +class llm_graph_input_attn_no_cache : public llm_graph_input_i { +public: + llm_graph_input_attn_no_cache(const llama_hparams & hparams, const llama_cparams & cparams) : + hparams(hparams), + cparams(cparams) { + } + ~llm_graph_input_attn_no_cache() = default; + + void set_input(const llama_ubatch * ubatch) override; + + ggml_tensor * get_kq_mask() const { return kq_mask_cnv; } + + ggml_tensor * kq_mask = nullptr; // F32 [n_tokens, n_batch] + ggml_tensor * kq_mask_cnv = nullptr; // [n_tokens, n_batch] + + const llama_hparams & hparams; + const llama_cparams & cparams; +}; + +class llm_graph_input_attn_kv_unified : public llm_graph_input_i { +public: + llm_graph_input_attn_kv_unified( + const llama_hparams & hparams, + const llama_cparams & cparams, + const llama_kv_cache_unified * kv_self) : + hparams(hparams), + cparams(cparams), + kv_self(kv_self) { + } + ~llm_graph_input_attn_kv_unified() = default; + + void set_input(const llama_ubatch * ubatch) override; + + ggml_tensor * get_kq_mask() const { return self_kq_mask_cnv; } + ggml_tensor * get_kq_mask_swa() const { return self_kq_mask_swa_cnv; } + + ggml_tensor * self_kq_mask = nullptr; // F32 [n_kv, n_batch] + ggml_tensor * self_kq_mask_cnv = nullptr; // [n_kv, n_batch] + ggml_tensor * self_kq_mask_swa = nullptr; // F32 [n_kv, n_batch] + ggml_tensor * self_kq_mask_swa_cnv = nullptr; // [n_kv, n_batch] + + const llama_hparams & hparams; + const llama_cparams & cparams; + + const llama_kv_cache_unified * kv_self; +}; + +class llm_graph_input_attn_cross : public llm_graph_input_i { +public: + llm_graph_input_attn_cross(const llama_cross * cross) : cross(cross) {} + ~llm_graph_input_attn_cross() = default; + + void set_input(const llama_ubatch * ubatch) override; + + ggml_tensor * get_kq_mask_cross() const { return cross_kq_mask_cnv; } + + ggml_tensor * cross_kq_mask = nullptr; // F32 [n_outputs_enc, n_batch] + ggml_tensor * cross_kq_mask_cnv = nullptr; // F32 [n_outputs_enc, n_batch] + + const llama_cross * cross = nullptr; +}; + +// +// llm_graph_result +// + +// these objects deliver the result from the graph build process back to the llama_context +// note that the input tensors created for the graph are referenced here - the goal is to be able to populate their +// specific data, by calling the set_inputs() method +// along with the input tensors, the object also provides commonly used outputs tensors, such as logits, embeddings, etc. +// these are used by the llama_context to extact the relevant data, based on the compute parameters + +class llm_graph_result_i { +public: + virtual ~llm_graph_result_i() = default; + + virtual ggml_tensor * get_logits() = 0; + virtual ggml_tensor * get_embd() = 0; + virtual ggml_tensor * get_embd_pooled() = 0; + + virtual void set_inputs(const llama_ubatch * ubatch) = 0; +}; + +using llm_graph_result_ptr = std::unique_ptr; + + +class llm_graph_result : public llm_graph_result_i { +public: + virtual ~llm_graph_result() = default; + + ggml_tensor * get_logits() override { return t_logits; } + ggml_tensor * get_embd() override { return t_embd; } + ggml_tensor * get_embd_pooled() override { return t_embd_pooled; } + + void set_inputs(const llama_ubatch * ubatch) override { + for (auto & input : inputs) { + input->set_input(ubatch); + } + } + + llm_graph_input_i * add_input(llm_graph_input_ptr input) { + inputs.emplace_back(std::move(input)); + return inputs.back().get(); + } + + // important graph nodes + ggml_tensor * t_logits = nullptr; + ggml_tensor * t_embd = nullptr; + ggml_tensor * t_embd_pooled = nullptr; + + std::vector inputs; +}; + +// +// llm_graph_context +// + +// callback that allows us to apply custom logic to each tensor (e.g. ggml-alloc, offloading, etc.) +using llm_graph_cb = std::function; + +struct llm_graph_params { + ggml_context * ctx; + + const llm_arch arch; + + const llama_hparams & hparams; + const llama_cparams & cparams; + const llama_ubatch & ubatch; + + ggml_backend_sched * sched; + ggml_backend * backend_cpu; + + const llama_adapter_cvec * cvec; + const llama_adapter_loras * loras; + const llama_memory_i * memory; + const llama_cross * cross; + + int32_t n_outputs; + + const llm_graph_cb & cb; +}; + +struct llm_graph_context { + const llm_arch arch; + + const llama_hparams & hparams; + const llama_cparams & cparams; + const llama_ubatch & ubatch; + + const int64_t n_embd; + const int64_t n_layer; + const int64_t n_rot; + const int64_t n_ctx; // user-specified context size (can be different from n_ctx_train) + const int64_t n_ctx_per_seq; + const int64_t n_head; + const int64_t n_head_kv; + const int64_t n_embd_head_k; + const int64_t n_embd_k_gqa; + const int64_t n_embd_head_v; + const int64_t n_embd_v_gqa; + const int64_t n_expert; + const int64_t n_expert_used; + + const float freq_base; + const float freq_scale; + const float ext_factor; + const float attn_factor; + const float beta_fast; + const float beta_slow; + const float norm_eps; + const float norm_rms_eps; + + const int32_t n_tokens; + const int32_t n_outputs; + const int32_t n_ctx_orig; // yarn + + const enum llama_pooling_type pooling_type; + const enum llama_rope_type rope_type; + + ggml_context * ctx0 = nullptr; + + ggml_backend_sched * sched; + + ggml_backend * backend_cpu; // TODO: needed by build_attn_mha, figure out a way to remove? + + const llama_adapter_cvec * cvec; + const llama_adapter_loras * loras; + const llama_memory_i * memory; + const llama_cross * cross; + + const llm_graph_cb & cb_func; + + std::unique_ptr res; + + llm_graph_context(const llm_graph_params & params); + + int64_t n_pos_per_token() const; + + void cb(ggml_tensor * cur, const char * name, int il) const; + + // + // common + // + + ggml_tensor * build_cvec( + ggml_tensor * cur, + int il) const; + + // do mat_mul, while optionally apply lora + ggml_tensor * build_lora_mm( + ggml_tensor * w, + ggml_tensor * cur) const; + + // do mat_mul_id, while optionally apply lora + ggml_tensor * build_lora_mm_id( + ggml_tensor * w, // ggml_tensor * as + ggml_tensor * cur, // ggml_tensor * b + ggml_tensor * ids) const; + + ggml_tensor * build_norm( + ggml_tensor * cur, + ggml_tensor * mw, + ggml_tensor * mb, + llm_norm_type type, + int il) const; + + ggml_tensor * build_ffn( + ggml_tensor * cur, + ggml_tensor * up, + ggml_tensor * up_b, + ggml_tensor * up_s, + ggml_tensor * gate, + ggml_tensor * gate_b, + ggml_tensor * gate_s, + ggml_tensor * down, + ggml_tensor * down_b, + ggml_tensor * down_s, + ggml_tensor * act_scales, + llm_ffn_op_type type_op, + llm_ffn_gate_type type_gate, + int il) const; + + ggml_tensor * build_moe_ffn( + ggml_tensor * cur, + ggml_tensor * gate_inp, + ggml_tensor * up_exps, + ggml_tensor * gate_exps, + ggml_tensor * down_exps, + ggml_tensor * exp_probs_b, + int64_t n_expert, + int64_t n_expert_used, + llm_ffn_op_type type_op, + bool norm_w, + bool scale_w, + float w_scale, + llama_expert_gating_func_type gating_op, + int il) const; + + // + // inputs + // + + ggml_tensor * build_inp_embd(ggml_tensor * tok_embd) const; + ggml_tensor * build_inp_pos() const; + ggml_tensor * build_inp_attn_scale() const; + ggml_tensor * build_inp_out_ids() const; + ggml_tensor * build_inp_mean() const; + ggml_tensor * build_inp_cls() const; + ggml_tensor * build_inp_s_copy() const; + ggml_tensor * build_inp_s_mask() const; + + ggml_tensor * build_inp_cross_embd() const; + ggml_tensor * build_inp_pos_bucket_enc() const; + ggml_tensor * build_inp_pos_bucket_dec() const; + ggml_tensor * build_pos_bias(ggml_tensor * pos_bucket, ggml_tensor * attn_rel_b) const; + + // + // attention + // + + ggml_tensor * build_attn_mha( + ggml_cgraph * gf, + ggml_tensor * q, // [n_embd_head_q, n_tokens, n_head_q] + ggml_tensor * k, // [n_embd_head_k, n_tokens, n_head_k] + ggml_tensor * v, // [n_embd_head_v, n_tokens, n_head_v] (v_trans == false) + ggml_tensor * kq_b, + ggml_tensor * kq_mask, + bool v_trans, + float kq_scale) const; + + llm_graph_input_attn_no_cache * build_attn_inp_no_cache() const; + + ggml_tensor * build_attn( + llm_graph_input_attn_no_cache * inp, + ggml_cgraph * gf, + ggml_tensor * wo, + ggml_tensor * wo_b, + ggml_tensor * q_cur, // [n_embd_head_q, n_head_q, n_tokens] + ggml_tensor * k_cur, // [n_embd_head_k, n_head_k, n_tokens] + ggml_tensor * v_cur, // [n_embd_head_v, n_head_v, n_tokens] + ggml_tensor * kq_b, + float kq_scale, + int il) const; + + llm_graph_input_attn_kv_unified * build_attn_inp_kv_unified() const; + + ggml_tensor * build_attn( + llm_graph_input_attn_kv_unified * inp, + ggml_cgraph * gf, + ggml_tensor * wo, + ggml_tensor * wo_b, + ggml_tensor * q_cur, // [n_embd_head_q, n_head_q, n_tokens] + ggml_tensor * k_cur, // [n_embd_head_k, n_head_k, n_tokens] + ggml_tensor * v_cur, // [n_embd_head_v, n_head_v, n_tokens] + ggml_tensor * kq_b, + float kq_scale, + int il) const; + + llm_graph_input_attn_cross * build_attn_inp_cross() const; + + ggml_tensor * build_attn( + llm_graph_input_attn_cross * inp, + ggml_cgraph * gf, + ggml_tensor * wo, + ggml_tensor * wo_b, + ggml_tensor * q_cur, // [n_embd_head_q, n_head_q, n_tokens] + ggml_tensor * k_cur, // [n_embd_head_k, n_head_k, n_tokens] + ggml_tensor * v_cur, // [n_embd_head_v, n_head_v, n_tokens] + ggml_tensor * kq_b, + float kq_scale, + int il) const; + + // + // recurrent + // + + ggml_tensor * build_copy_mask_state( + ggml_cgraph * gf, + ggml_tensor * s, + ggml_tensor * state_copy, + ggml_tensor * state_mask, + int32_t n_state, + int32_t n_seqs) const; + + ggml_tensor * build_rwkv_token_shift_load( + ggml_cgraph * gf, + ggml_tensor * state_copy, + ggml_tensor * state_mask, + const llama_ubatch & ubatch, + int il) const; + + ggml_tensor * build_rwkv_token_shift_store( + ggml_tensor * token_shift, + const llama_ubatch & ubatch, + int il) const; + + // + // pooling + // + + void build_pooling( + ggml_cgraph * gf, + ggml_tensor * cls, + ggml_tensor * cls_b, + ggml_tensor * cls_out, + ggml_tensor * cls_out_b) const; +}; diff --git a/src/llama-hparams.cpp b/src/llama-hparams.cpp index ea87b2953..90dfe7a7f 100644 --- a/src/llama-hparams.cpp +++ b/src/llama-hparams.cpp @@ -69,3 +69,11 @@ uint32_t llama_hparams::n_embd_v_s() const { // corresponds to Mamba's ssm_states size return ssm_d_state * ssm_d_inner; } + +bool llama_hparams::is_swa(uint32_t il) const { + if (il < n_layer) { + return n_swa > 0 && n_swa_pattern > 0 && il % n_swa_pattern < (n_swa_pattern - 1); + } + + GGML_ABORT("fatal error"); +} diff --git a/src/llama-hparams.h b/src/llama-hparams.h index 1fe454103..4e0b57190 100644 --- a/src/llama-hparams.h +++ b/src/llama-hparams.h @@ -36,6 +36,7 @@ struct llama_hparams { uint32_t n_layer; uint32_t n_rot; uint32_t n_swa = 0; // sliding window attention (SWA) + uint32_t n_swa_pattern = 1; // by default, all layers use non-sliding-window attention uint32_t n_embd_head_k; // dimension of keys (d_k). d_q is assumed to be the same, but there are n_head q heads, and only n_head_kv k-v heads uint32_t n_embd_head_v; // dimension of values (d_v) aka n_embd_head uint32_t n_expert = 0; @@ -75,10 +76,16 @@ struct llama_hparams { uint32_t time_decay_extra_dim = 0; uint32_t wkv_head_size = 0; uint32_t token_shift_count = 2; + uint32_t n_lora_decay = 0; + uint32_t n_lora_iclr = 0; + uint32_t n_lora_value_res_mix = 0; + uint32_t n_lora_gate = 0; float rope_attn_factor = 1.0f; float rope_freq_base_train; + float rope_freq_base_train_swa; float rope_freq_scale_train; + float rope_freq_scale_train_swa; uint32_t n_ctx_orig_yarn; float rope_yarn_log_mul; @@ -105,6 +112,14 @@ struct llama_hparams { bool use_alibi = false; bool attn_soft_cap = false; + uint32_t n_moe_layer_step = 0; + bool use_kq_norm = true; + uint32_t n_attn_chunk = 0; + // values below seems to be fixed on llama4 + uint32_t n_no_rope_layer_step = 4; + uint32_t n_attn_temp_floor_scale = 8192; + float f_attn_temp_scale = 0.1; + // needed by encoder-decoder models (e.g. T5, FLAN-T5) // ref: https://github.com/ggerganov/llama.cpp/pull/8141 llama_token dec_start_token_id = LLAMA_TOKEN_NULL; @@ -133,6 +148,8 @@ struct llama_hparams { // dimension of the recurrent state embeddings uint32_t n_embd_v_s() const; + + bool is_swa(uint32_t il) const; }; static_assert(std::is_trivially_copyable::value, "llama_hparams must be trivially copyable"); diff --git a/src/llama-impl.h b/src/llama-impl.h index 12d1fb082..02b1d07f8 100644 --- a/src/llama-impl.h +++ b/src/llama-impl.h @@ -6,13 +6,13 @@ #include #ifdef __GNUC__ -#ifdef __MINGW32__ -#define LLAMA_ATTRIBUTE_FORMAT(...) __attribute__((format(gnu_printf, __VA_ARGS__))) +# if defined(__MINGW32__) && !defined(__clang__) +# define LLAMA_ATTRIBUTE_FORMAT(...) __attribute__((format(gnu_printf, __VA_ARGS__))) +# else +# define LLAMA_ATTRIBUTE_FORMAT(...) __attribute__((format(printf, __VA_ARGS__))) +# endif #else -#define LLAMA_ATTRIBUTE_FORMAT(...) __attribute__((format(printf, __VA_ARGS__))) -#endif -#else -#define LLAMA_ATTRIBUTE_FORMAT(...) +# define LLAMA_ATTRIBUTE_FORMAT(...) #endif // diff --git a/src/llama-io.cpp b/src/llama-io.cpp new file mode 100644 index 000000000..7ad70d163 --- /dev/null +++ b/src/llama-io.cpp @@ -0,0 +1,15 @@ +#include "llama-io.h" + +void llama_io_write_i::write_string(const std::string & str) { + uint32_t str_size = str.size(); + + write(&str_size, sizeof(str_size)); + write(str.data(), str_size); +} + +void llama_io_read_i::read_string(std::string & str) { + uint32_t str_size; + read_to(&str_size, sizeof(str_size)); + + str.assign((const char *) read(str_size), str_size); +} diff --git a/src/llama-io.h b/src/llama-io.h new file mode 100644 index 000000000..ce9216b83 --- /dev/null +++ b/src/llama-io.h @@ -0,0 +1,35 @@ +#pragma once + +#include +#include +#include + +struct ggml_tensor; + +class llama_io_write_i { +public: + llama_io_write_i() = default; + virtual ~llama_io_write_i() = default; + + virtual void write(const void * src, size_t size) = 0; + virtual void write_tensor(const ggml_tensor * tensor, size_t offset, size_t size) = 0; + + // bytes written so far + virtual size_t n_bytes() = 0; + + void write_string(const std::string & str); +}; + +class llama_io_read_i { +public: + llama_io_read_i() = default; + virtual ~llama_io_read_i() = default; + + virtual const uint8_t * read(size_t size) = 0; + virtual void read_to(void * dst, size_t size) = 0; + + // bytes read so far + virtual size_t n_bytes() = 0; + + void read_string(std::string & str); +}; diff --git a/src/llama-kv-cache.cpp b/src/llama-kv-cache.cpp index feffdf0de..dbf5f1187 100644 --- a/src/llama-kv-cache.cpp +++ b/src/llama-kv-cache.cpp @@ -6,86 +6,90 @@ #include "llama-model.h" #include +#include #include #include +#include -static const llama_kv_cache_slot_info llama_kv_cache_slot_info_failed{false}; - -uint32_t llama_kv_cache_get_padding(const struct llama_cparams & cparams) { - // the FA kernels require padding to avoid extra runtime boundary checks - return cparams.flash_attn ? 256u : 32u; +llama_kv_cache_unified::llama_kv_cache_unified(const llama_hparams & hparams, callbacks cbs) : hparams(hparams), cbs(std::move(cbs)) { } -bool llama_kv_cache_init( - struct llama_kv_cache & cache, - const llama_model & model, - const llama_cparams & cparams, - ggml_type type_k, - ggml_type type_v, - uint32_t kv_size, - bool offload) { - const struct llama_hparams & hparams = model.hparams; - +bool llama_kv_cache_unified::init( + const llama_model & model, + const llama_cparams & cparams, + ggml_type type_k, + ggml_type type_v, + uint32_t kv_size, + bool offload) { const int32_t n_layer = hparams.n_layer; - cache.has_shift = false; + has_shift = false; - cache.recurrent = llama_model_is_recurrent(&model); - cache.v_trans = !cache.recurrent && !cparams.flash_attn; - cache.can_shift = !cache.recurrent && model.arch != LLM_ARCH_DEEPSEEK2; // not supported due to MLA + recurrent = llama_model_is_recurrent(&model); + v_trans = !recurrent && !cparams.flash_attn; + can_shift = !recurrent && model.arch != LLM_ARCH_DEEPSEEK2; // not supported due to MLA LLAMA_LOG_INFO("%s: kv_size = %d, offload = %d, type_k = '%s', type_v = '%s', n_layer = %d, can_shift = %d\n", - __func__, kv_size, offload, ggml_type_name(type_k), ggml_type_name(type_v), n_layer, cache.can_shift); + __func__, kv_size, offload, ggml_type_name(type_k), ggml_type_name(type_v), n_layer, can_shift); - cache.head = 0; - cache.size = kv_size; - cache.used = 0; + head = 0; + size = kv_size; + used = 0; - cache.type_k = type_k; - cache.type_v = type_v; + this->type_k = type_k; + this->type_v = type_v; - cache.cells.clear(); - cache.cells.resize(kv_size); + cells.clear(); + cells.resize(kv_size); // create a context for each buffer type std::map ctx_map; auto ctx_for_buft = [&](ggml_backend_buffer_type_t buft) -> ggml_context * { auto it = ctx_map.find(buft); if (it == ctx_map.end()) { - struct ggml_init_params params = { + ggml_init_params params = { /*.mem_size =*/ size_t(2u*n_layer*ggml_tensor_overhead()), /*.mem_buffer =*/ NULL, /*.no_alloc =*/ true, }; + ggml_context * ctx = ggml_init(params); if (!ctx) { return nullptr; } + ctx_map[buft] = ctx; - cache.ctxs.emplace_back(ctx); + ctxs.emplace_back(ctx); + return ctx; } + return it->second; }; - cache.k_l.reserve(n_layer); - cache.v_l.reserve(n_layer); + k_l.reserve(n_layer); + v_l.reserve(n_layer); for (int i = 0; i < n_layer; i++) { const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa(i) + hparams.n_embd_k_s(); const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(i) + hparams.n_embd_v_s(); - LLAMA_LOG_DEBUG("%s: layer %d: n_embd_k_gqa = %d, n_embd_v_gqa = %d\n", __func__, i, n_embd_k_gqa, n_embd_v_gqa); + const char * dev_name = "CPU"; ggml_backend_buffer_type_t buft; if (offload) { auto * dev = model.dev_layer(i); buft = ggml_backend_dev_buffer_type(dev); + + dev_name = ggml_backend_dev_name(dev); } else { buft = ggml_backend_cpu_buffer_type(); } - ggml_context * ctx = ctx_for_buft(buft); + LLAMA_LOG_DEBUG("%s: layer %3d: n_embd_k_gqa = %d, n_embd_v_gqa = %d, dev = %s\n", __func__, + i, n_embd_k_gqa, n_embd_v_gqa, dev_name); + + ggml_context * ctx = ctx_for_buft(buft); if (!ctx) { LLAMA_LOG_ERROR("%s: failed to create ggml context for kv cache\n", __func__); return false; @@ -95,8 +99,8 @@ bool llama_kv_cache_init( ggml_tensor * v = ggml_new_tensor_1d(ctx, type_v, n_embd_v_gqa*kv_size); ggml_format_name(k, "cache_k_l%d", i); ggml_format_name(v, "cache_v_l%d", i); - cache.k_l.push_back(k); - cache.v_l.push_back(v); + k_l.push_back(k); + v_l.push_back(v); } // allocate tensors and initialize the buffers to avoid NaNs in the padding @@ -111,280 +115,80 @@ bool llama_kv_cache_init( } ggml_backend_buffer_clear(buf, 0); LLAMA_LOG_INFO("%s: %10s KV buffer size = %8.2f MiB\n", __func__, ggml_backend_buffer_name(buf), ggml_backend_buffer_get_size(buf)/1024.0/1024.0); - cache.bufs.emplace_back(buf); + bufs.emplace_back(buf); } return true; } -struct llama_kv_cache_slot_info llama_kv_cache_find_slot( - struct llama_kv_cache & cache, - const struct llama_ubatch & ubatch) { - const uint32_t n_tokens = ubatch.n_tokens; - const uint32_t n_seqs = ubatch.n_seqs; - const uint32_t n_seq_tokens = ubatch.n_seq_tokens; +int32_t llama_kv_cache_unified::get_n_tokens() const { + int32_t result = 0; - if (cache.recurrent) { - // For recurrent state architectures (like Mamba or RWKV), - // each cache cell can store the state for a whole sequence. - // A slot should be always be contiguous. - - // can only process batches with an equal number of new tokens in each sequence - GGML_ASSERT(ubatch.equal_seqs); - - int32_t min = cache.size - 1; - int32_t max = 0; - - // everything should fit if all seq_ids are smaller than the max - for (uint32_t s = 0; s < n_seqs; ++s) { - const uint32_t n_seq_id = ubatch.n_seq_id[s]; - for (uint32_t j = 0; j < n_seq_id; ++j) { - const llama_seq_id seq_id = ubatch.seq_id[s][j]; - - if (seq_id < 0 || (uint32_t) seq_id >= cache.size) { - // too big seq_id - // TODO: would it be possible to resize the cache instead? - LLAMA_LOG_ERROR("%s: seq_id=%d >= n_seq_max=%d Try using a bigger --parallel value\n", __func__, seq_id, cache.size); - return llama_kv_cache_slot_info_failed; - } - if (j > 0) { - llama_kv_cell & seq = cache.cells[seq_id]; - if (seq.tail >= 0) { - llama_kv_cell & cell = cache.cells[seq.tail]; - // clear cells from seq_ids that become shared - // (should not normally happen, but let's handle it anyway) - cell.seq_id.erase(seq_id); - seq.tail = -1; - if (cell.seq_id.empty()) { - cell.pos = -1; - cell.src = -1; - cache.used -= 1; - } - } - } - } - } - -#ifndef NDEBUG - { - std::vector tails_verif; - tails_verif.assign(cache.size, -1); - for (uint32_t i = 0; i < cache.size; ++i) { - llama_kv_cell & cell = cache.cells[i]; - for (llama_seq_id seq_id : cell.seq_id) { - if (tails_verif[seq_id] != -1) { - LLAMA_LOG_ERROR("%s: duplicate tail for seq_id %d in cell %d and %d\n", __func__, seq_id, i, tails_verif[seq_id]); - } - tails_verif[seq_id] = i; - } - } - for (uint32_t i = 0; i < cache.size; ++i) { - if (tails_verif[i] != cache.cells[i].tail) { - LLAMA_LOG_ERROR("%s: wrong tail for seq_id %d, (%d instead of %d)\n", __func__, i, cache.cells[i].tail, tails_verif[i]); - } - } - } -#endif - - // find next empty cell - uint32_t next_empty_cell = cache.head; - - for (uint32_t i = 0; i < cache.size; ++i) { - if (next_empty_cell >= cache.size) { next_empty_cell -= cache.size; } - llama_kv_cell & cell = cache.cells[next_empty_cell]; - if (cell.is_empty()) { break; } - next_empty_cell += 1; - } - - // find usable cell range - for (uint32_t s = 0; s < n_seqs; ++s) { - const llama_seq_id seq_id = ubatch.seq_id[s][0]; - llama_kv_cell & seq_meta = cache.cells[seq_id]; - bool has_cell = false; - if (seq_meta.tail >= 0) { - llama_kv_cell & cell = cache.cells[seq_meta.tail]; - GGML_ASSERT(cell.has_seq_id(seq_id)); - // does this seq_id "own" the cell? - if (cell.seq_id.size() == 1) { has_cell = true; } - } - if (!has_cell) { - llama_kv_cell & empty_cell = cache.cells[next_empty_cell]; - GGML_ASSERT(empty_cell.is_empty()); - // copy old tail into the empty cell - if (seq_meta.tail >= 0) { - llama_kv_cell & orig_cell = cache.cells[seq_meta.tail]; - empty_cell.pos = orig_cell.pos; - empty_cell.src = orig_cell.src; - orig_cell.seq_id.erase(seq_id); - empty_cell.seq_id.insert(seq_id); // will be overwritten - } - seq_meta.tail = next_empty_cell; - // find next empty cell - if (s + 1 < n_seqs) { - next_empty_cell += 1; - for (uint32_t i = 0; i < cache.size; ++i) { - if (next_empty_cell >= cache.size) { next_empty_cell -= cache.size; } - llama_kv_cell & cell = cache.cells[next_empty_cell]; - if (cell.is_empty()) { break; } - next_empty_cell += 1; - } - } - } - if (min > seq_meta.tail) { min = seq_meta.tail; } - if (max < seq_meta.tail) { max = seq_meta.tail; } - } - - // gather and re-order - for (uint32_t s = 0; s < n_seqs; ++s) { - int32_t dst_id = s + min; - int32_t src_id = cache.cells[ubatch.seq_id[s][0]].tail; - if (dst_id != src_id) { - llama_kv_cell & dst_cell = cache.cells[dst_id]; - llama_kv_cell & src_cell = cache.cells[src_id]; - - std::swap(dst_cell.pos, src_cell.pos); - std::swap(dst_cell.src, src_cell.src); - std::swap(dst_cell.seq_id, src_cell.seq_id); - - // swap tails (assuming they NEVER overlap) - for (const llama_seq_id seq_id : src_cell.seq_id) { - cache.cells[seq_id].tail = src_id; - } - for (const llama_seq_id seq_id : dst_cell.seq_id) { - cache.cells[seq_id].tail = dst_id; - } - } - } - - // update the pos of the used seqs - for (uint32_t s = 0; s < n_seqs; ++s) { - const llama_pos last_pos = ubatch.pos[n_seq_tokens * s + n_seq_tokens - 1]; - int32_t cell_id = s + min; - llama_kv_cell & cell = cache.cells[cell_id]; - - if (cell.pos >= 0 && last_pos != cell.pos + (llama_pos) n_seq_tokens) { - // What should happen when the pos backtracks or skips a value? - // Clearing the state mid-batch would require special-casing which isn't done. - LLAMA_LOG_WARN("%s: non-consecutive token position %d after %d for sequence %d with %u new tokens\n", - __func__, last_pos, cell.pos, ubatch.seq_id[s][0], n_seq_tokens); - } - cell.pos = last_pos; - cell.seq_id.clear(); - for (int32_t j = 0; j < ubatch.n_seq_id[s]; ++j) { - const llama_seq_id seq_id = ubatch.seq_id[s][j]; - cell.seq_id.insert(seq_id); - cache.cells[seq_id].tail = cell_id; - } - } - - // allow getting the range of used cells, from head to head + n - cache.head = min; - cache.n = max - min + 1; - cache.used = std::count_if(cache.cells.begin(), cache.cells.end(), - [](const llama_kv_cell& cell){ return !cell.is_empty(); }); - - // sanity check - return llama_kv_cache_slot_info(cache.n >= n_seqs); - } - // otherwise, one cell per token. - - if (n_tokens > cache.size) { - LLAMA_LOG_ERROR("%s: n_tokens=%d > cache.size=%d\n", __func__, n_tokens, cache.size); - return llama_kv_cache_slot_info_failed; + for (uint32_t i = 0; i < size; i++) { + result += cells[i].seq_id.size(); } - uint32_t n_tested = 0; - - while (true) { - if (cache.head + n_tokens > cache.size) { - n_tested += cache.size - cache.head; - cache.head = 0; - continue; - } - - bool found = true; - for (uint32_t i = 0; i < n_tokens; i++) { - if (cache.cells[cache.head + i].pos >= 0) { - found = false; - cache.head += i + 1; - n_tested += i + 1; - break; - } - } - - if (found) { - break; - } - - if (n_tested >= cache.size) { - //LLAMA_LOG_ERROR("%s: failed to find a slot for %d tokens\n", __func__, n_tokens); - return llama_kv_cache_slot_info_failed; - } - } - - for (uint32_t s = 0; s < n_seqs; s++) { - for (uint32_t i = 0; i < n_seq_tokens; ++i) { - uint32_t k = s*n_seq_tokens + i; - cache.cells[cache.head + k].pos = ubatch.pos[k]; - - for (int32_t j = 0; j < ubatch.n_seq_id[s]; j++) { - cache.cells[cache.head + k].seq_id.insert(ubatch.seq_id[s][j]); - } - } - } - - cache.used += n_tokens; - - return llama_kv_cache_slot_info(cache.head, cache.head + n_tokens); + return result; } -uint32_t llama_kv_cache_cell_max(const struct llama_kv_cache & cache) { - for (uint32_t i = cache.size; i > 0; --i) { - const llama_kv_cell & cell = cache.cells[i - 1]; - - if (cell.pos >= 0 && !cell.is_empty()) { - return i; - } - } - - return 0; +int32_t llama_kv_cache_unified::get_used_cells() const { + return used; } -void llama_kv_cache_clear(struct llama_kv_cache & cache) { - for (int32_t i = 0; i < (int32_t) cache.size; ++i) { - cache.cells[i].pos = -1; - cache.cells[i].seq_id.clear(); - cache.cells[i].src = -1; - cache.cells[i].tail = -1; +size_t llama_kv_cache_unified::total_size() const { + size_t size = 0; + for (const auto & buf : bufs) { + size += ggml_backend_buffer_get_size(buf.get()); } - cache.head = 0; - cache.used = 0; - for (auto & buf : cache.bufs) { + return size; +} + +llama_pos llama_kv_cache_unified::pos_max() const { + llama_pos pos_max = -1; + for (const auto & cell : cells) { + pos_max = std::max(pos_max, cell.pos); + } + + return pos_max; +} + +void llama_kv_cache_unified::clear() { + for (int32_t i = 0; i < (int32_t) size; ++i) { + cells[i].pos = -1; + cells[i].seq_id.clear(); + cells[i].src = -1; + cells[i].tail = -1; + } + head = 0; + used = 0; + + for (auto & buf : bufs) { ggml_backend_buffer_clear(buf.get(), 0); } } -bool llama_kv_cache_seq_rm( - struct llama_kv_cache & cache, - llama_seq_id seq_id, - llama_pos p0, - llama_pos p1) { - uint32_t new_head = cache.size; +bool llama_kv_cache_unified::seq_rm(llama_seq_id seq_id, llama_pos p0, llama_pos p1) { + uint32_t new_head = size; - if (p0 < 0) p0 = 0; - if (p1 < 0) p1 = std::numeric_limits::max(); + if (p0 < 0) { + p0 = 0; + } + + if (p1 < 0) { + p1 = std::numeric_limits::max(); + } // models like Mamba or RWKV can't have a state partially erased - if (cache.recurrent) { - if (seq_id >= (int64_t) cache.size) { + if (recurrent) { + if (seq_id >= (int64_t) size) { // could be fatal return false; } if (0 <= seq_id) { - int32_t & tail_id = cache.cells[seq_id].tail; + int32_t & tail_id = cells[seq_id].tail; if (tail_id >= 0) { - const llama_kv_cell & cell = cache.cells[tail_id]; + const llama_kv_cell & cell = cells[tail_id]; // partial intersection is invalid if ((0 < p0 && p0 <= cell.pos) || (0 < p1 && p1 <= cell.pos)) { return false; @@ -400,50 +204,63 @@ bool llama_kv_cache_seq_rm( return false; } } + + return true; } - for (uint32_t i = 0; i < cache.size; ++i) { - if (cache.cells[i].pos >= p0 && cache.cells[i].pos < p1) { + for (uint32_t i = 0; i < size; ++i) { + if (cells[i].pos >= p0 && cells[i].pos < p1) { if (seq_id < 0) { - cache.cells[i].seq_id.clear(); - } else if (cache.cells[i].has_seq_id(seq_id)) { - cache.cells[i].seq_id.erase(seq_id); + cells[i].seq_id.clear(); + } else if (cells[i].has_seq_id(seq_id)) { + cells[i].seq_id.erase(seq_id); } else { continue; } - if (cache.cells[i].is_empty()) { + if (cells[i].is_empty()) { // keep count of the number of used cells - if (cache.cells[i].pos >= 0) cache.used--; + if (cells[i].pos >= 0) { + used--; + } - cache.cells[i].pos = -1; - cache.cells[i].src = -1; - if (new_head == cache.size) new_head = i; + cells[i].pos = -1; + cells[i].src = -1; + + if (new_head == size) { + new_head = i; + } } } } // If we freed up a slot, set head to it so searching can start there. - if (new_head != cache.size && new_head < cache.head) cache.head = new_head; + if (new_head != size && new_head < head) { + head = new_head; + } return true; } -void llama_kv_cache_seq_cp( - struct llama_kv_cache & cache, - llama_seq_id seq_id_src, - llama_seq_id seq_id_dst, - llama_pos p0, - llama_pos p1) { - if (p0 < 0) p0 = 0; - if (p1 < 0) p1 = std::numeric_limits::max(); +void llama_kv_cache_unified::seq_cp(llama_seq_id seq_id_src, llama_seq_id seq_id_dst, llama_pos p0, llama_pos p1) { + if (seq_id_src == seq_id_dst) { + return; + } - if (cache.recurrent) { - if ((uint32_t) seq_id_dst < cache.size && (uint32_t) seq_id_src < cache.size) { - llama_kv_cell & tail_src = cache.cells[seq_id_src]; - llama_kv_cell & tail_dst = cache.cells[seq_id_dst]; + if (p0 < 0) { + p0 = 0; + } + + if (p1 < 0) { + p1 = std::numeric_limits::max(); + } + + if (recurrent) { + if ((uint32_t) seq_id_dst < size && (uint32_t) seq_id_src < size) { + llama_kv_cell & tail_src = cells[seq_id_src]; + llama_kv_cell & tail_dst = cells[seq_id_dst]; if (tail_dst.tail >= 0) { // clear destination seq_id if it wasn't empty - llama_kv_cell & cell_dst = cache.cells[tail_dst.tail]; + llama_kv_cell & cell_dst = cells[tail_dst.tail]; cell_dst.seq_id.erase(seq_id_dst); tail_dst.tail = -1; @@ -451,11 +268,11 @@ void llama_kv_cache_seq_cp( cell_dst.pos = -1; cell_dst.delta = -1; cell_dst.src = -1; - cache.used -= 1; + used -= 1; } } if (tail_src.tail >= 0) { - llama_kv_cell & cell_src = cache.cells[tail_src.tail]; + llama_kv_cell & cell_src = cells[tail_src.tail]; cell_src.seq_id.insert(seq_id_dst); tail_dst.tail = tail_src.tail; @@ -464,59 +281,75 @@ void llama_kv_cache_seq_cp( return; } - // otherwise, this is the KV cache of a Transformer-like model - cache.head = 0; + // otherwise, this is the KV of a Transformer-like model + head = 0; - for (uint32_t i = 0; i < cache.size; ++i) { - if (cache.cells[i].has_seq_id(seq_id_src) && cache.cells[i].pos >= p0 && cache.cells[i].pos < p1) { - cache.cells[i].seq_id.insert(seq_id_dst); + for (uint32_t i = 0; i < size; ++i) { + if (cells[i].has_seq_id(seq_id_src) && cells[i].pos >= p0 && cells[i].pos < p1) { + cells[i].seq_id.insert(seq_id_dst); } } } -void llama_kv_cache_seq_keep(struct llama_kv_cache & cache, llama_seq_id seq_id) { - uint32_t new_head = cache.size; +void llama_kv_cache_unified::seq_keep(llama_seq_id seq_id) { + uint32_t new_head = size; - for (uint32_t i = 0; i < cache.size; ++i) { - if (cache.recurrent && (llama_seq_id) i != seq_id) { - cache.cells[i].tail = -1; + for (uint32_t i = 0; i < size; ++i) { + if (recurrent && (llama_seq_id) i != seq_id) { + cells[i].tail = -1; } - if (!cache.cells[i].has_seq_id(seq_id)) { - if (cache.cells[i].pos >= 0) cache.used--; - cache.cells[i].pos = -1; - cache.cells[i].src = -1; - cache.cells[i].seq_id.clear(); - if (new_head == cache.size) new_head = i; + + if (!cells[i].has_seq_id(seq_id)) { + if (cells[i].pos >= 0) { + used--; + } + + cells[i].pos = -1; + cells[i].src = -1; + cells[i].seq_id.clear(); + + if (new_head == size){ + new_head = i; + } } else { - cache.cells[i].seq_id.clear(); - cache.cells[i].seq_id.insert(seq_id); + cells[i].seq_id.clear(); + cells[i].seq_id.insert(seq_id); } } // If we freed up a slot, set head to it so searching can start there. - if (new_head != cache.size && new_head < cache.head) cache.head = new_head; + if (new_head != size && new_head < head) { + head = new_head; + } } -void llama_kv_cache_seq_add( - struct llama_kv_cache & cache, - llama_seq_id seq_id, - llama_pos p0, - llama_pos p1, - llama_pos delta) { - uint32_t new_head = cache.size; +void llama_kv_cache_unified::seq_add(llama_seq_id seq_id, llama_pos p0, llama_pos p1, llama_pos delta) { + if (delta == 0) { + return; + } - if (p0 < 0) p0 = 0; - if (p1 < 0) p1 = std::numeric_limits::max(); - // If there is no range then return early to avoid looping over the cache. - if (p0 == p1) return; + uint32_t new_head = size; - if (cache.recurrent) { + if (p0 < 0) { + p0 = 0; + } + + if (p1 < 0) { + p1 = std::numeric_limits::max(); + } + + // If there is no range then return early to avoid looping over the + if (p0 == p1) { + return; + } + + if (recurrent) { // for Mamba-like or RWKV models, only the pos needs to be shifted - if (0 <= seq_id && seq_id < (int64_t) cache.size) { - const int32_t tail_id = cache.cells[seq_id].tail; + if (0 <= seq_id && seq_id < (int64_t) size) { + const int32_t tail_id = cells[seq_id].tail; if (tail_id >= 0) { - llama_kv_cell & cell = cache.cells[tail_id]; + llama_kv_cell & cell = cells[tail_id]; if (cell.has_seq_id(seq_id) && p0 <= cell.pos && cell.pos < p1) { cell.pos += delta; } @@ -525,19 +358,19 @@ void llama_kv_cache_seq_add( return; } - for (uint32_t i = 0; i < cache.size; ++i) { - if (cache.cells[i].has_seq_id(seq_id) && cache.cells[i].pos >= p0 && cache.cells[i].pos < p1) { - cache.has_shift = true; - cache.cells[i].pos += delta; - cache.cells[i].delta += delta; + for (uint32_t i = 0; i < size; ++i) { + if (cells[i].has_seq_id(seq_id) && cells[i].pos >= p0 && cells[i].pos < p1) { + has_shift = true; + cells[i].pos += delta; + cells[i].delta += delta; - if (cache.cells[i].pos < 0) { - if (!cache.cells[i].is_empty()) { - cache.used--; + if (cells[i].pos < 0) { + if (!cells[i].is_empty()) { + used--; } - cache.cells[i].pos = -1; - cache.cells[i].seq_id.clear(); - if (new_head == cache.size) { + cells[i].pos = -1; + cells[i].seq_id.clear(); + if (new_head == size) { new_head = i; } } @@ -546,93 +379,915 @@ void llama_kv_cache_seq_add( // If we freed up a slot, set head to it so searching can start there. // Otherwise we just start the next search from the beginning. - cache.head = new_head != cache.size ? new_head : 0; + head = new_head != size ? new_head : 0; } -void llama_kv_cache_seq_div( - struct llama_kv_cache & cache, - llama_seq_id seq_id, - llama_pos p0, - llama_pos p1, - int d) { - if (p0 < 0) p0 = 0; - if (p1 < 0) p1 = std::numeric_limits::max(); - // If there is no range then return early to avoid looping over the cache. - if (p0 == p1) return; +void llama_kv_cache_unified::seq_div(llama_seq_id seq_id, llama_pos p0, llama_pos p1, int d) { + if (d == 1) { + return; + } - if (cache.recurrent) { + if (p0 < 0) { + p0 = 0; + } + + if (p1 < 0) { + p1 = std::numeric_limits::max(); + } + + // If there is no range then return early to avoid looping over the cache. + if (p0 == p1) { + return; + } + + if (recurrent) { // for Mamba-like or RWKV models, only the pos needs to be changed - if (0 <= seq_id && seq_id < (int64_t) cache.size) { - const int32_t tail_id = cache.cells[seq_id].tail; + if (0 <= seq_id && seq_id < (int64_t) size) { + const int32_t tail_id = cells[seq_id].tail; if (tail_id >= 0) { - llama_kv_cell & cell = cache.cells[tail_id]; + llama_kv_cell & cell = cells[tail_id]; if (cell.has_seq_id(seq_id) && p0 <= cell.pos && cell.pos < p1) { cell.pos /= d; } } } + return; } - for (uint32_t i = 0; i < cache.size; ++i) { - if (cache.cells[i].has_seq_id(seq_id) && cache.cells[i].pos >= p0 && cache.cells[i].pos < p1) { - cache.has_shift = true; + for (uint32_t i = 0; i < size; ++i) { + if (cells[i].has_seq_id(seq_id) && cells[i].pos >= p0 && cells[i].pos < p1) { + has_shift = true; { - llama_pos p_old = cache.cells[i].pos; - cache.cells[i].pos /= d; - cache.cells[i].delta += cache.cells[i].pos - p_old; + llama_pos p_old = cells[i].pos; + cells[i].pos /= d; + cells[i].delta += cells[i].pos - p_old; } } } } -llama_pos llama_kv_cache_seq_pos_max(struct llama_kv_cache & cache, llama_seq_id seq_id) { +llama_pos llama_kv_cache_unified::seq_pos_max(llama_seq_id seq_id) const { llama_pos result = 0; - for (uint32_t i = 0; i < cache.size; ++i) { - if (cache.cells[i].has_seq_id(seq_id)) { - result = std::max(result, cache.cells[i].pos); + for (uint32_t i = 0; i < size; ++i) { + if (cells[i].has_seq_id(seq_id)) { + result = std::max(result, cells[i].pos); } } return result; } -void llama_kv_cache_defrag(struct llama_kv_cache & cache) { - if (!cache.recurrent) { - cache.do_defrag = true; +void llama_kv_cache_unified::defrag() { + if (!recurrent) { + do_defrag = true; } } -int32_t llama_get_kv_cache_token_count(const struct llama_kv_cache & kv) { - int result = 0; - - for (uint32_t i = 0; i < kv.size; i++) { - result += kv.cells[i].seq_id.size(); +void llama_kv_cache_unified::restore() { + if (pending.ranges.empty()) { + return; } - return result; + // TODO: tmp - move to llama_kv_cache_recurrent + if (recurrent) { + seq_rm(-1, -1, -1); + return; + } + + uint32_t new_head = size; + + for (auto & range : pending.ranges) { + for (uint32_t i = range.c0; i < range.c1; ++i) { + cells[i].seq_id.clear(); + + // keep count of the number of used cells + if (cells[i].pos >= 0) { + used--; + } + + cells[i].pos = -1; + cells[i].src = -1; + } + + new_head = std::min(new_head, range.c0); + } + + if (new_head != size && new_head < head) { + head = new_head; + } } -int32_t llama_get_kv_cache_used_cells(const struct llama_kv_cache & kv) { - return kv.used; +void llama_kv_cache_unified::commit() { + // TODO: tmp - move to llama_kv_cache_recurrent + if (recurrent) { + return; + } + + if (pending.ranges.empty()) { + LLAMA_LOG_WARN("%s: no pending KV cache updates to commit - might indicate a bug (ref: %s)\n", + __func__, "https://github.com/ggml-org/llama.cpp/pull/12695"); + return; + } + + pending.ranges.clear(); } -bool llama_kv_cache_can_shift(const struct llama_kv_cache & kv) { - return kv.can_shift; +bool llama_kv_cache_unified::get_can_shift() const { + return can_shift; +} + +bool llama_kv_cache_unified::find_slot( + const llama_ubatch & ubatch) { + const uint32_t n_tokens = ubatch.n_tokens; + const uint32_t n_seqs = ubatch.n_seqs; + const uint32_t n_seq_tokens = ubatch.n_seq_tokens; + + // if we have enough unused cells before the current head -> + // better to start searching from the beginning of the cache, hoping to fill it + if (head > used + 2*ubatch.n_tokens) { + head = 0; + } + + if (recurrent) { + // For recurrent state architectures (like Mamba or RWKV), + // each cache cell can store the state for a whole sequence. + // A slot should be always be contiguous. + + // can only process batches with an equal number of new tokens in each sequence + GGML_ASSERT(ubatch.equal_seqs); + + int32_t min = size - 1; + int32_t max = 0; + + // everything should fit if all seq_ids are smaller than the max + for (uint32_t s = 0; s < n_seqs; ++s) { + const uint32_t n_seq_id = ubatch.n_seq_id[s]; + for (uint32_t j = 0; j < n_seq_id; ++j) { + const llama_seq_id seq_id = ubatch.seq_id[s][j]; + + if (seq_id < 0 || (uint32_t) seq_id >= size) { + // too big seq_id + // TODO: would it be possible to resize the cache instead? + LLAMA_LOG_ERROR("%s: seq_id=%d >= n_seq_max=%d Try using a bigger --parallel value\n", __func__, seq_id, size); + return false; + } + if (j > 0) { + llama_kv_cell & seq = cells[seq_id]; + if (seq.tail >= 0) { + llama_kv_cell & cell = cells[seq.tail]; + // clear cells from seq_ids that become shared + // (should not normally happen, but let's handle it anyway) + cell.seq_id.erase(seq_id); + seq.tail = -1; + if (cell.seq_id.empty()) { + cell.pos = -1; + cell.src = -1; + used -= 1; + } + } + } + } + } + +#ifndef NDEBUG + { + std::vector tails_verif; + tails_verif.assign(size, -1); + for (uint32_t i = 0; i < size; ++i) { + llama_kv_cell & cell = cells[i]; + for (llama_seq_id seq_id : cell.seq_id) { + if (tails_verif[seq_id] != -1) { + LLAMA_LOG_ERROR("%s: duplicate tail for seq_id %d in cell %d and %d\n", __func__, seq_id, i, tails_verif[seq_id]); + } + tails_verif[seq_id] = i; + } + } + for (uint32_t i = 0; i < size; ++i) { + if (tails_verif[i] != cells[i].tail) { + LLAMA_LOG_ERROR("%s: wrong tail for seq_id %d, (%d instead of %d)\n", __func__, i, cells[i].tail, tails_verif[i]); + } + } + } +#endif + + // find next empty cell + uint32_t next_empty_cell = head; + + for (uint32_t i = 0; i < size; ++i) { + if (next_empty_cell >= size) { next_empty_cell -= size; } + llama_kv_cell & cell = cells[next_empty_cell]; + if (cell.is_empty()) { break; } + next_empty_cell += 1; + } + + // find usable cell range + for (uint32_t s = 0; s < n_seqs; ++s) { + const llama_seq_id seq_id = ubatch.seq_id[s][0]; + llama_kv_cell & seq_meta = cells[seq_id]; + bool has_cell = false; + if (seq_meta.tail >= 0) { + llama_kv_cell & cell = cells[seq_meta.tail]; + GGML_ASSERT(cell.has_seq_id(seq_id)); + // does this seq_id "own" the cell? + if (cell.seq_id.size() == 1) { has_cell = true; } + } + if (!has_cell) { + llama_kv_cell & empty_cell = cells[next_empty_cell]; + GGML_ASSERT(empty_cell.is_empty()); + // copy old tail into the empty cell + if (seq_meta.tail >= 0) { + llama_kv_cell & orig_cell = cells[seq_meta.tail]; + empty_cell.pos = orig_cell.pos; + empty_cell.src = orig_cell.src; + orig_cell.seq_id.erase(seq_id); + empty_cell.seq_id.insert(seq_id); // will be overwritten + } + seq_meta.tail = next_empty_cell; + // find next empty cell + if (s + 1 < n_seqs) { + next_empty_cell += 1; + for (uint32_t i = 0; i < size; ++i) { + if (next_empty_cell >= size) { next_empty_cell -= size; } + llama_kv_cell & cell = cells[next_empty_cell]; + if (cell.is_empty()) { break; } + next_empty_cell += 1; + } + } + } + if (min > seq_meta.tail) { min = seq_meta.tail; } + if (max < seq_meta.tail) { max = seq_meta.tail; } + } + + // gather and re-order + for (uint32_t s = 0; s < n_seqs; ++s) { + int32_t dst_id = s + min; + int32_t src_id = cells[ubatch.seq_id[s][0]].tail; + if (dst_id != src_id) { + llama_kv_cell & dst_cell = cells[dst_id]; + llama_kv_cell & src_cell = cells[src_id]; + + std::swap(dst_cell.pos, src_cell.pos); + std::swap(dst_cell.src, src_cell.src); + std::swap(dst_cell.seq_id, src_cell.seq_id); + + // swap tails (assuming they NEVER overlap) + for (const llama_seq_id seq_id : src_cell.seq_id) { + cells[seq_id].tail = src_id; + } + for (const llama_seq_id seq_id : dst_cell.seq_id) { + cells[seq_id].tail = dst_id; + } + } + } + + // update the pos of the used seqs + for (uint32_t s = 0; s < n_seqs; ++s) { + const llama_pos last_pos = ubatch.pos[n_seq_tokens * s + n_seq_tokens - 1]; + int32_t cell_id = s + min; + llama_kv_cell & cell = cells[cell_id]; + + if (cell.pos >= 0 && last_pos != cell.pos + (llama_pos) n_seq_tokens) { + // What should happen when the pos backtracks or skips a value? + // Clearing the state mid-batch would require special-casing which isn't done. + LLAMA_LOG_WARN("%s: non-consecutive token position %d after %d for sequence %d with %u new tokens\n", + __func__, last_pos, cell.pos, ubatch.seq_id[s][0], n_seq_tokens); + } + cell.pos = last_pos; + cell.seq_id.clear(); + for (int32_t j = 0; j < ubatch.n_seq_id[s]; ++j) { + const llama_seq_id seq_id = ubatch.seq_id[s][j]; + cell.seq_id.insert(seq_id); + cells[seq_id].tail = cell_id; + } + } + + // allow getting the range of used cells, from head to head + n + head = min; + n = max - min + 1; + used = std::count_if(cells.begin(), cells.end(), + [](const llama_kv_cell& cell){ return !cell.is_empty(); }); + + // sanity check + return n >= n_seqs; + } + + // otherwise, one cell per token. + + if (n_tokens > size) { + LLAMA_LOG_ERROR("%s: n_tokens = %d > size = %d\n", __func__, n_tokens, size); + return false; + } + + uint32_t n_tested = 0; + + while (true) { + if (head + n_tokens > size) { + n_tested += size - head; + head = 0; + continue; + } + + bool found = true; + for (uint32_t i = 0; i < n_tokens; i++) { + if (cells[head + i].pos >= 0) { + found = false; + head += i + 1; + n_tested += i + 1; + break; + } + } + + if (found) { + break; + } + + if (n_tested >= size) { + //LLAMA_LOG_ERROR("%s: failed to find a slot for %d tokens\n", __func__, n_tokens); + return false; + } + } + + for (uint32_t s = 0; s < n_seqs; s++) { + for (uint32_t i = 0; i < n_seq_tokens; ++i) { + uint32_t k = s*n_seq_tokens + i; + cells[head + k].pos = ubatch.pos[k]; + + for (int32_t j = 0; j < ubatch.n_seq_id[s]; j++) { + cells[head + k].seq_id.insert(ubatch.seq_id[s][j]); + } + } + } + + used += n_tokens; + + pending.ranges.push_back({head, head + n_tokens}); + + return true; +} + +uint32_t llama_kv_cache_unified::get_padding(const llama_cparams & cparams) const { + // the FA kernels require padding to avoid extra runtime boundary checks + return cparams.flash_attn ? 256u : 32u; +} + +uint32_t llama_kv_cache_unified::cell_max() const { + for (uint32_t i = size; i > 0; --i) { + const llama_kv_cell & cell = cells[i - 1]; + + if (cell.pos >= 0 && !cell.is_empty()) { + return i; + } + } + + return 0; +} + +size_t llama_kv_cache_unified::size_k_bytes() const { + size_t size_k_bytes = 0; + + for (const auto & k : k_l) { + size_k_bytes += ggml_nbytes(k); + } + + return size_k_bytes; +} + +size_t llama_kv_cache_unified::size_v_bytes() const { + size_t size_v_bytes = 0; + + for (const auto & v : v_l) { + size_v_bytes += ggml_nbytes(v); + } + + return size_v_bytes; +} + +bool llama_kv_cache_unified::defrag_prepare(int32_t n_max_nodes) { + const uint32_t n_layer = hparams.n_layer; + + const uint32_t n_kv = cell_max(); + const uint32_t n_used = used; + + assert(n_used <= n_kv); + + //const int64_t t_start = ggml_time_us(); + + // number of cells moved + uint32_t n_moves = 0; + + // each move requires 6*n_layer tensors (see graph_build_kv_self_defrag) + // - source view, destination view, copy operation + // - x2 for keys and values + //const uint32_t max_moves = max_nodes()/(6*n_layer); + // TODO: tmp fix https://github.com/ggerganov/llama.cpp/issues/6685#issuecomment-2057579516 + const uint32_t max_moves = (n_max_nodes - 2*n_layer)/(6*n_layer); + + // determine which KV cells to move where + // + // cell i moves to ids[i] + // + // if ids[i] == i || ids[i] == n_kv, then cell i is not moved + // + auto & ids = defrag_info.ids; + + ids.clear(); + ids.resize(n_kv, n_kv); + + for (uint32_t i0 = 0; i0 < n_used; ++i0) { + const auto & cell0 = cells[i0]; + + if (!cell0.is_empty()) { + ids[i0] = i0; + + continue; + } + + // found a hole - fill it with data from the end of the cache + + uint32_t nh = 1; + + // determine the size of the hole + while (i0 + nh < n_used && cells[i0 + nh].is_empty()) { + nh++; + } + + uint32_t nf = 0; + uint32_t is = n_kv - 1; + + // starting from the end, find nh non-empty cells + for (; is > i0; --is) { + const auto & cell1 = cells[is]; + + if (cell1.is_empty() || ids[is] != n_kv) { + continue; + } + + // non-empty cell which is not yet moved + nf++; + + if (nf == nh) { + break; + } + } + + // this can only happen if `n_used` is not accurate, which would be a bug + GGML_ASSERT(nf == nh && "KV defrag bug: nf != nh"); + + nf = 0; + + uint32_t i1 = is; + + // are we moving a continuous block of memory? + bool cont = false; + + // should we stop searching for the next move? + bool stop = false; + + // go back and move the nf cells to the hole + for (; i1 < n_kv; ++i1) { + auto & cell1 = cells[i1]; + + if (cell1.is_empty() || ids[i1] != n_kv) { + if (n_moves == max_moves) { + stop = true; + break; + } + + cont = false; + continue; + } + + // this cell goes to (i0 + nf) + ids[i1] = i0 + nf; + + // move the cell meta data + cells[i0 + nf] = cell1; + + // clear the old cell and move the head there + cell1 = llama_kv_cell(); + head = n_used; + + if (!cont) { + n_moves++; + cont = true; + } + + nf++; + + if (nf == nh) { + break; + } + } + + if (stop || n_moves == max_moves) { + break; + } + + //LLAMA_LOG_INFO("(tmp log) KV defrag: move [%u, %u) to [%u, %u)\n", is, i1 + 1, i0, i0 + nh); + + i0 += nh - 1; + } + + if (n_moves == 0) { + return false; + } + + LLAMA_LOG_DEBUG("(tmp log) KV defrag cell moves: %u\n", n_moves); + + LLAMA_LOG_DEBUG("expected gf nodes: %u\n", 6*n_moves*n_layer); + + return true; +} + +void llama_kv_cache_unified::state_write(llama_io_write_i & io, llama_seq_id seq_id) const { + std::vector> cell_ranges; // ranges, from inclusive, to exclusive + uint32_t cell_count = 0; + + // Count the number of cells with the specified seq_id + // Find all the ranges of cells with this seq id (or all, when -1) + uint32_t cell_range_begin = size; + for (uint32_t i = 0; i < size; ++i) { + const auto & cell = cells[i]; + if ((seq_id == -1 && !cell.is_empty()) || cell.has_seq_id(seq_id)) { + ++cell_count; + if (cell_range_begin == size) { + cell_range_begin = i; + } + } else { + if (cell_range_begin != size) { + cell_ranges.emplace_back(cell_range_begin, i); + cell_range_begin = size; + } + } + } + if (cell_range_begin != size) { + cell_ranges.emplace_back(cell_range_begin, size); + } + + // DEBUG CHECK: Sum of cell counts in ranges should equal the total cell count + uint32_t cell_count_check = 0; + for (const auto & range : cell_ranges) { + cell_count_check += range.second - range.first; + } + GGML_ASSERT(cell_count == cell_count_check); + + io.write(&cell_count, sizeof(cell_count)); + + state_write_meta(io, cell_ranges, seq_id); + state_write_data(io, cell_ranges); +} + +void llama_kv_cache_unified::state_read(llama_io_read_i & io, llama_seq_id seq_id) { + uint32_t cell_count; + io.read_to(&cell_count, sizeof(cell_count)); + + bool res = true; + res = res && state_read_meta(io, cell_count, seq_id); + res = res && state_read_data(io, cell_count); + + if (!res) { + if (seq_id == -1) { + clear(); + } else { + seq_rm(seq_id, -1, -1); + } + throw std::runtime_error("failed to restore kv cache"); + } +} + +void llama_kv_cache_unified::state_write_meta(llama_io_write_i & io, const std::vector> & cell_ranges, llama_seq_id seq_id) const { + for (const auto & range : cell_ranges) { + for (uint32_t i = range.first; i < range.second; ++i) { + const auto & cell = cells[i]; + const llama_pos pos = cell.pos; + const uint32_t n_seq_id = seq_id == -1 ? cell.seq_id.size() : 0; + + io.write(&pos, sizeof(pos)); + io.write(&n_seq_id, sizeof(n_seq_id)); + + if (n_seq_id) { + for (auto seq_id : cell.seq_id) { + io.write(&seq_id, sizeof(seq_id)); + } + } + } + } +} + +void llama_kv_cache_unified::state_write_data(llama_io_write_i & io, const std::vector> & cell_ranges) const { + const uint32_t v_trans = this->v_trans ? 1 : 0; + const uint32_t n_layer = hparams.n_layer; + + io.write(&v_trans, sizeof(v_trans)); + io.write(&n_layer, sizeof(n_layer)); + + std::vector tmp_buf; + + // Iterate and write all the keys first, each row is a cell + // Get whole range at a time + for (uint32_t il = 0; il < n_layer; ++il) { + const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa(il) + hparams.n_embd_k_s(); + + // Write key type + const int32_t k_type_i = (int32_t)k_l[il]->type; + io.write(&k_type_i, sizeof(k_type_i)); + + // Write row size of key + const uint64_t k_size_row = ggml_row_size(k_l[il]->type, n_embd_k_gqa); + io.write(&k_size_row, sizeof(k_size_row)); + + // Read each range of cells of k_size length each into tmp_buf and write out + for (const auto & range : cell_ranges) { + const size_t range_size = range.second - range.first; + const size_t buf_size = range_size * k_size_row; + io.write_tensor(k_l[il], range.first * k_size_row, buf_size); + } + } + + if (!v_trans) { + for (uint32_t il = 0; il < n_layer; ++il) { + const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il) + hparams.n_embd_v_s(); + + // Write value type + const int32_t v_type_i = (int32_t)v_l[il]->type; + io.write(&v_type_i, sizeof(v_type_i)); + + // Write row size of value + const uint64_t v_size_row = ggml_row_size(v_l[il]->type, n_embd_v_gqa); + io.write(&v_size_row, sizeof(v_size_row)); + + // Read each range of cells of v_size length each into tmp_buf and write out + for (const auto & range : cell_ranges) { + const size_t range_size = range.second - range.first; + const size_t buf_size = range_size * v_size_row; + io.write_tensor(v_l[il], range.first * v_size_row, buf_size); + } + } + } else { + // When v is transposed, we also need the element size and get the element ranges from each row + const uint32_t kv_size = size; + for (uint32_t il = 0; il < n_layer; ++il) { + const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il) + hparams.n_embd_v_s(); + + // Write value type + const int32_t v_type_i = (int32_t)v_l[il]->type; + io.write(&v_type_i, sizeof(v_type_i)); + + // Write element size + const uint32_t v_size_el = ggml_type_size(v_l[il]->type); + io.write(&v_size_el, sizeof(v_size_el)); + + // Write GQA embedding size + io.write(&n_embd_v_gqa, sizeof(n_embd_v_gqa)); + + // For each row, we get the element values of each cell + for (uint32_t j = 0; j < n_embd_v_gqa; ++j) { + // Read each range of cells of v_size_el length each into tmp_buf and write out + for (const auto & range : cell_ranges) { + const size_t range_size = range.second - range.first; + const size_t src_offset = (range.first + j * kv_size) * v_size_el; + const size_t buf_size = range_size * v_size_el; + io.write_tensor(v_l[il], src_offset, buf_size); + } + } + } + } +} + +bool llama_kv_cache_unified::state_read_meta(llama_io_read_i & io, uint32_t cell_count, llama_seq_id dest_seq_id) { + if (dest_seq_id != -1) { + // single sequence + + seq_rm(dest_seq_id, -1, -1); + + llama_sbatch sbatch; + llama_ubatch batch = sbatch.reserve_ubatch(cell_count, /* has_embd */ false); + + batch.n_tokens = cell_count; + batch.n_seq_tokens = cell_count; + batch.n_seqs = 1; + + for (uint32_t i = 0; i < cell_count; ++i) { + llama_pos pos; + uint32_t n_seq_id; + + io.read_to(&pos, sizeof(pos)); + io.read_to(&n_seq_id, sizeof(n_seq_id)); + + if (n_seq_id != 0) { + LLAMA_LOG_ERROR("%s: invalid seq_id-agnostic kv cell\n", __func__); + return false; + } + + batch.pos[i] = pos; + } + batch.n_seq_id[0] = 1; + batch.seq_id[0] = &dest_seq_id; + if (!find_slot(batch)) { + LLAMA_LOG_ERROR("%s: failed to find available cells in kv cache\n", __func__); + return false; + } + commit(); + + // DEBUG CHECK: kv.head should be our first cell, kv.head + cell_count - 1 should be our last cell (verify seq_id and pos values) + // Assume that this is one contiguous block of cells + GGML_ASSERT(head + cell_count <= size); + GGML_ASSERT(cells[head].pos == batch.pos[0]); + GGML_ASSERT(cells[head + cell_count - 1].pos == batch.pos[cell_count - 1]); + GGML_ASSERT(cells[head].has_seq_id(dest_seq_id)); + GGML_ASSERT(cells[head + cell_count - 1].has_seq_id(dest_seq_id)); + } else { + // whole KV cache restore + + if (cell_count > size) { + LLAMA_LOG_ERROR("%s: not enough cells in kv cache\n", __func__); + return false; + } + + clear(); + + for (uint32_t i = 0; i < cell_count; ++i) { + llama_kv_cell & cell = cells[i]; + + llama_pos pos; + uint32_t n_seq_id; + + io.read_to(&pos, sizeof(pos)); + io.read_to(&n_seq_id, sizeof(n_seq_id)); + + cell.pos = pos; + + for (uint32_t j = 0; j < n_seq_id; ++j) { + llama_seq_id seq_id; + io.read_to(&seq_id, sizeof(seq_id)); + + // TODO: llama_kv_cache_unified should have a notion of max sequences + //if (seq_id < 0 || (uint32_t) seq_id >= llama_n_seq_max(ctx)) { + if (seq_id < 0) { + //LLAMA_LOG_ERROR("%s: invalid seq_id, %d is out of range [0, %u)\n", __func__, seq_id, llama_n_seq_max(ctx)); + LLAMA_LOG_ERROR("%s: invalid seq_id, %d is out of range [0, inf)\n", __func__, seq_id); + return false; + } + + cell.seq_id.insert(seq_id); + + if (recurrent) { + int32_t & tail = cells[seq_id].tail; + if (tail != -1) { + LLAMA_LOG_ERROR("%s: duplicate tail for seq_id %d in cell %d and %d\n", __func__, seq_id, i, tail); + return false; + } + tail = i; + } + } + } + + head = 0; + used = cell_count; + } + + if (recurrent) { + for (uint32_t i = 0; i < cell_count; ++i) { + uint32_t cell_id = head + i; + // make sure the recurrent states will keep their restored state + cells[cell_id].src = cell_id; + } + } + + return true; +} + +bool llama_kv_cache_unified::state_read_data(llama_io_read_i & io, uint32_t cell_count) { + uint32_t v_trans; + uint32_t n_layer; + io.read_to(&v_trans, sizeof(v_trans)); + io.read_to(&n_layer, sizeof(n_layer)); + + if (n_layer != hparams.n_layer) { + LLAMA_LOG_ERROR("%s: mismatched layer count (%u instead of %u)\n", __func__, n_layer, hparams.n_layer); + return false; + } + if (cell_count > size) { + LLAMA_LOG_ERROR("%s: not enough cells in kv cache to restore state (%u > %u)\n", __func__, cell_count, size); + return false; + } + if (v_trans != (bool) v_trans) { + LLAMA_LOG_ERROR("%s: incompatible V transposition\n", __func__); + return false; + } + + // For each layer, read the keys for each cell, one row is one cell, read as one contiguous block + for (uint32_t il = 0; il < n_layer; ++il) { + const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa(il) + hparams.n_embd_k_s(); + + // Read type of key + int32_t k_type_i_ref; + io.read_to(&k_type_i_ref, sizeof(k_type_i_ref)); + const int32_t k_type_i = (int32_t) k_l[il]->type; + if (k_type_i != k_type_i_ref) { + LLAMA_LOG_ERROR("%s: mismatched key type (%d != %d, layer %d)\n", __func__, k_type_i, k_type_i_ref, il); + return false; + } + + // Read row size of key + uint64_t k_size_row_ref; + io.read_to(&k_size_row_ref, sizeof(k_size_row_ref)); + const size_t k_size_row = ggml_row_size(k_l[il]->type, n_embd_k_gqa); + if (k_size_row != k_size_row_ref) { + LLAMA_LOG_ERROR("%s: mismatched key row size (%zu != %zu, layer %d)\n", __func__, k_size_row, (size_t) k_size_row_ref, il); + return false; + } + + if (cell_count) { + // Read and set the keys for the whole cell range + ggml_backend_tensor_set(k_l[il], io.read(cell_count * k_size_row), head * k_size_row, cell_count * k_size_row); + } + } + + if (!v_trans) { + for (uint32_t il = 0; il < n_layer; ++il) { + const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il) + hparams.n_embd_v_s(); + + // Read type of value + int32_t v_type_i_ref; + io.read_to(&v_type_i_ref, sizeof(v_type_i_ref)); + const int32_t v_type_i = (int32_t)v_l[il]->type; + if (v_type_i != v_type_i_ref) { + LLAMA_LOG_ERROR("%s: mismatched value type (%d != %d, layer %d)\n", __func__, v_type_i, v_type_i_ref, il); + return false; + } + + // Read row size of value + uint64_t v_size_row_ref; + io.read_to(&v_size_row_ref, sizeof(v_size_row_ref)); + const size_t v_size_row = ggml_row_size(v_l[il]->type, n_embd_v_gqa); + if (v_size_row != v_size_row_ref) { + LLAMA_LOG_ERROR("%s: mismatched value row size (%zu != %zu, layer %d)\n", __func__, v_size_row, (size_t) v_size_row_ref, il); + return false; + } + + if (cell_count) { + // Read and set the values for the whole cell range + ggml_backend_tensor_set(v_l[il], io.read(cell_count * v_size_row), head * v_size_row, cell_count * v_size_row); + } + } + } else { + // For each layer, read the values for each cell (transposed) + for (uint32_t il = 0; il < n_layer; ++il) { + const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il) + hparams.n_embd_v_s(); + + // Read type of value + int32_t v_type_i_ref; + io.read_to(&v_type_i_ref, sizeof(v_type_i_ref)); + const int32_t v_type_i = (int32_t)v_l[il]->type; + if (v_type_i != v_type_i_ref) { + LLAMA_LOG_ERROR("%s: mismatched value type (%d != %d, layer %d)\n", __func__, v_type_i, v_type_i_ref, il); + return false; + } + + // Read element size of value + uint32_t v_size_el_ref; + io.read_to(&v_size_el_ref, sizeof(v_size_el_ref)); + const size_t v_size_el = ggml_type_size(v_l[il]->type); + if (v_size_el != v_size_el_ref) { + LLAMA_LOG_ERROR("%s: mismatched value element size (%zu != %zu, layer %d)\n", __func__, v_size_el, (size_t) v_size_el_ref, il); + return false; + } + + // Read GQA embedding size + uint32_t n_embd_v_gqa_ref; + io.read_to(&n_embd_v_gqa_ref, sizeof(n_embd_v_gqa_ref)); + if (n_embd_v_gqa != n_embd_v_gqa_ref) { + LLAMA_LOG_ERROR("%s: mismatched GQA embedding size (%u != %u, layer %d)\n", __func__, n_embd_v_gqa, n_embd_v_gqa_ref, il); + return false; + } + + if (cell_count) { + // For each row in the transposed matrix, read the values for the whole cell range + for (uint32_t j = 0; j < n_embd_v_gqa; ++j) { + const size_t dst_offset = (head + j * size) * v_size_el; + ggml_backend_tensor_set(v_l[il], io.read(cell_count * v_size_el), dst_offset, cell_count * v_size_el); + } + } + } + } + + return true; } // // kv cache view // -struct llama_kv_cache_view llama_kv_cache_view_init(const struct llama_kv_cache & kv, int32_t n_seq_max) { - struct llama_kv_cache_view result = { +llama_kv_cache_view llama_kv_cache_view_init(const llama_kv_cache & kv, int32_t n_seq_max) { + llama_kv_cache_view result = { /*.n_cells = */ 0, /*.n_seq_max = */ n_seq_max, /*.token_count = */ 0, - /*.used_cells = */ llama_get_kv_cache_used_cells(kv), + /*.used_cells = */ kv.get_used_cells(), /*.max_contiguous = */ 0, /*.max_contiguous_idx = */ -1, /*.cells = */ nullptr, @@ -642,7 +1297,7 @@ struct llama_kv_cache_view llama_kv_cache_view_init(const struct llama_kv_cache return result; } -void llama_kv_cache_view_free(struct llama_kv_cache_view * view) { +void llama_kv_cache_view_free(llama_kv_cache_view * view) { if (view->cells != nullptr) { free(view->cells); view->cells = nullptr; @@ -653,18 +1308,25 @@ void llama_kv_cache_view_free(struct llama_kv_cache_view * view) { } } -void llama_kv_cache_view_update(struct llama_kv_cache_view * view, const struct llama_kv_cache & kv) { - if (uint32_t(view->n_cells) < kv.size || view->cells == nullptr) { - view->n_cells = int32_t(kv.size); - void * p = realloc(view->cells, sizeof(struct llama_kv_cache_view_cell) * view->n_cells); +void llama_kv_cache_view_update(llama_kv_cache_view * view, const llama_kv_cache * kv) { + // TODO: rework this in the future, for now quick hack + const llama_kv_cache_unified * kvu = dynamic_cast(kv); + if (kvu == nullptr) { + LLAMA_LOG_ERROR("%s: the kv_cache_view currently works only with llama_kv_cache_unified\n", __func__); + return; + } + + if (uint32_t(view->n_cells) < kvu->size || view->cells == nullptr) { + view->n_cells = int32_t(kvu->size); + void * p = realloc(view->cells, sizeof(llama_kv_cache_view_cell) * view->n_cells); GGML_ASSERT(p != nullptr && "Failed to alloc kv_cache_view cells"); - view->cells = (struct llama_kv_cache_view_cell *)p; + view->cells = (llama_kv_cache_view_cell *)p; p = realloc(view->cells_sequences, sizeof(llama_seq_id) * view->n_seq_max * view->n_cells); GGML_ASSERT(p != nullptr && "Failed to alloc kv_cache_view cells sequences"); view->cells_sequences = (llama_seq_id *)p; } - const std::vector & kv_cells = kv.cells; + const std::vector & kv_cells = kvu->cells; llama_kv_cache_view_cell * c_curr = view->cells; llama_seq_id * cs_curr = view->cells_sequences; int32_t used_cells = 0; @@ -673,7 +1335,7 @@ void llama_kv_cache_view_update(struct llama_kv_cache_view * view, const struct uint32_t max_contig = 0; int32_t max_contig_idx = -1; - for (int32_t i = 0; i < int32_t(kv.size); i++, c_curr++, cs_curr += view->n_seq_max) { + for (int32_t i = 0; i < int32_t(kvu->size); i++, c_curr++, cs_curr += view->n_seq_max) { const size_t curr_size = kv_cells[i].seq_id.size(); token_count += curr_size; c_curr->pos = kv_cells[i].pos + kv_cells[i].delta; @@ -711,8 +1373,8 @@ void llama_kv_cache_view_update(struct llama_kv_cache_view * view, const struct view->max_contiguous_idx = max_contig_idx; view->token_count = token_count; view->used_cells = used_cells; - if (uint32_t(used_cells) != kv.used) { + if (uint32_t(used_cells) != kvu->used) { LLAMA_LOG_ERROR("%s: used cells mismatch. kv_cache says %d but we calculated %d\n", - __func__, kv.used, used_cells); + __func__, kvu->used, used_cells); } } diff --git a/src/llama-kv-cache.h b/src/llama-kv-cache.h index dca6f3998..56c74035a 100644 --- a/src/llama-kv-cache.h +++ b/src/llama-kv-cache.h @@ -1,15 +1,51 @@ #pragma once #include "llama.h" +#include "llama-io.h" +#include "llama-memory.h" #include "ggml-cpp.h" +#include #include #include +struct llama_cparams; +struct llama_hparams; +struct llama_ubatch; + +struct llama_kv_cache : public llama_memory_i { + using llama_memory_i::llama_memory_i; + + virtual void restore() = 0; // call if batch processing fails - restores the cache state + virtual void commit() = 0; // call after successful batch processing - clears any pending state + + virtual int32_t get_n_tokens() const = 0; + virtual int32_t get_used_cells() const = 0; // TODO: remove, this is too-specific to the unified cache + + virtual bool get_can_shift() const = 0; + + bool get_can_edit() const override { return get_can_shift(); } +}; + +struct llama_kv_cache_guard { + llama_kv_cache_guard(llama_kv_cache * kv) : kv(kv) {} + + ~llama_kv_cache_guard() { + kv->restore(); + } + + void commit() { + kv->commit(); + } + +private: + llama_kv_cache * kv; +}; + struct llama_kv_cell { llama_pos pos = -1; - llama_pos delta = 0; + llama_pos delta = 0; int32_t src = -1; // used by recurrent state models to copy states int32_t tail = -1; @@ -29,15 +65,112 @@ struct llama_kv_cell { }; // ring-buffer of cached KV data -struct llama_kv_cache { +// TODO: pimpl +// TODO: add notion of max sequences +class llama_kv_cache_unified : public llama_kv_cache { +public: + // can be used to query data from the model if needed + struct callbacks { + std::function get_rope_factors; + }; + + llama_kv_cache_unified( + const llama_hparams & hparams, + callbacks cbs); + + virtual ~llama_kv_cache_unified() = default; + + // TODO: become constructor + bool init( + const llama_model & model, // TODO: do not reference the model + const llama_cparams & cparams, + ggml_type type_k, + ggml_type type_v, + uint32_t kv_size, + bool offload); + + int32_t get_n_tokens() const override; + int32_t get_used_cells() const override; + + size_t total_size() const; + + // TODO: better data structures to reduce the cost of this operation + llama_pos pos_max() const; + + void clear() override; + void defrag() override; + + virtual void restore() override; + virtual void commit() override; + + bool seq_rm (llama_seq_id seq_id, llama_pos p0, llama_pos p1) override; + void seq_cp (llama_seq_id seq_id_src, llama_seq_id seq_id_dst, llama_pos p0, llama_pos p1) override; + void seq_keep(llama_seq_id seq_id) override; + void seq_add (llama_seq_id seq_id, llama_pos p0, llama_pos p1, llama_pos delta) override; + void seq_div (llama_seq_id seq_id, llama_pos p0, llama_pos p1, int d) override; + + llama_pos seq_pos_max(llama_seq_id seq_id) const override; + + bool get_can_shift() const override; + + // find an empty slot of size "n_tokens" in the cache + // updates the cache head + // Note: On success, it's important that cache.head points + // to the first cell of the slot. + bool find_slot(const llama_ubatch & batch); + + // TODO: maybe not needed + uint32_t get_padding(const llama_cparams & cparams) const; + + // find how many cells are currently in use + uint32_t cell_max() const; + + size_t size_k_bytes() const; + size_t size_v_bytes() const; + + // defrag + + struct { + std::vector ids; + } defrag_info; + + // return true if cells have been moved + bool defrag_prepare(int32_t n_max_nodes); + + // commit/restore cache + + struct slot_range { + uint32_t c0 = 0; // note: these are cell indices, not sequence positions + uint32_t c1 = 0; + }; + + // pending cell updates that are not yet committed + struct { + std::vector ranges; + } pending; + + // state write/load + + void state_write(llama_io_write_i & io, llama_seq_id seq_id = -1) const; + void state_read (llama_io_read_i & io, llama_seq_id seq_id = -1); + + // members + + const llama_hparams & hparams; + + callbacks cbs; + bool has_shift = false; bool do_defrag = false; + + // TODO: remove this and implement llama_kv_cache_recurrent instead bool recurrent = false; // with recurrent state models, a cell can hold the state for more than one past token + bool v_trans = true; // the value tensor is transposed bool can_shift = false; // Note: The value of head isn't only used to optimize searching - // for a free KV slot. llama_decode_internal also uses it, so it + // for a free KV slot. llama_decode_impl also uses it, so it // cannot be freely changed after a slot has been allocated. uint32_t head = 0; uint32_t size = 0; @@ -46,173 +179,35 @@ struct llama_kv_cache { // computed before each graph build uint32_t n = 0; + std::vector cells; + + std::vector k_l; // per layer + std::vector v_l; + +private: ggml_type type_k = GGML_TYPE_F16; ggml_type type_v = GGML_TYPE_F16; - std::vector cells; - - std::vector k_l; // per layer - std::vector v_l; - - std::vector ctxs; + std::vector ctxs; std::vector bufs; - size_t total_size() const { - size_t size = 0; - for (const auto & buf : bufs) { - size += ggml_backend_buffer_get_size(buf.get()); - } + void state_write_meta(llama_io_write_i & io, const std::vector> & cell_ranges, llama_seq_id seq_id = -1) const; + void state_write_data(llama_io_write_i & io, const std::vector> & cell_ranges) const; - return size; - } - - // TODO: better data structures to reduce the cost of this operation - llama_pos max_pos() const { - llama_pos max_pos = -1; - for (const auto & cell : cells) { - max_pos = std::max(max_pos, cell.pos); - } - - return max_pos; - } + bool state_read_meta(llama_io_read_i & io, uint32_t cell_count, llama_seq_id dest_seq_id = -1); + bool state_read_data(llama_io_read_i & io, uint32_t cell_count); }; -// a structure holds information about the slot found in llama_kv_cache_find_slot -struct llama_kv_cache_slot_info { - std::pair boundaries; // slot boundaries [begin, end) - bool found = false; // the slot was found - - explicit llama_kv_cache_slot_info(bool found_) : found{found_} {} - llama_kv_cache_slot_info(uint32_t begin, uint32_t end) : boundaries{begin, end}, found{true} {} - - operator bool() const { return found; } -}; - -// TODO: maybe not needed -uint32_t llama_kv_cache_get_padding(const struct llama_cparams & cparams); - -bool llama_kv_cache_init( - struct llama_kv_cache & cache, - const llama_model & model, - const llama_cparams & cparams, - ggml_type type_k, - ggml_type type_v, - uint32_t kv_size, - bool offload); - -// find an empty slot of size "n_tokens" in the cache -// updates the cache head -// returns a structure holding information about the slot found -// Note: On success, it's important that cache.head points -// to the first cell of the slot. -struct llama_kv_cache_slot_info llama_kv_cache_find_slot( - struct llama_kv_cache & cache, - const struct llama_ubatch & batch); - -// find how many cells are currently in use -uint32_t llama_kv_cache_cell_max(const struct llama_kv_cache & cache); - -void llama_kv_cache_clear(struct llama_kv_cache & cache); - -bool llama_kv_cache_seq_rm( - struct llama_kv_cache & cache, - llama_seq_id seq_id, - llama_pos p0, - llama_pos p1); - -void llama_kv_cache_seq_cp( - struct llama_kv_cache & cache, - llama_seq_id seq_id_src, - llama_seq_id seq_id_dst, - llama_pos p0, - llama_pos p1); - -void llama_kv_cache_seq_keep( - struct llama_kv_cache & cache, - llama_seq_id seq_id); - -void llama_kv_cache_seq_add( - struct llama_kv_cache & cache, - llama_seq_id seq_id, - llama_pos p0, - llama_pos p1, - llama_pos delta); - -void llama_kv_cache_seq_div( - struct llama_kv_cache & cache, - llama_seq_id seq_id, - llama_pos p0, - llama_pos p1, - int d); - -llama_pos llama_kv_cache_seq_pos_max( - struct llama_kv_cache & cache, - llama_seq_id seq_id); - -void llama_kv_cache_defrag(struct llama_kv_cache & cache); - -int32_t llama_get_kv_cache_token_count(const struct llama_kv_cache & kv); - -int32_t llama_get_kv_cache_used_cells(const struct llama_kv_cache & kv); - -bool llama_kv_cache_can_shift(const struct llama_kv_cache & kv); +// TODO: temporary reusing llama_kv_cache_unified -- implement recurrent cache and simplify llama_kv_cache_unified +//class llama_kv_cache_recurrent : public llama_kv_cache_unified { +//public: +// using llama_kv_cache_unified::llama_kv_cache_unified; +//}; // // kv cache view // -struct llama_kv_cache_view llama_kv_cache_view_init(const struct llama_kv_cache & kv, int32_t n_seq_max); - -void llama_kv_cache_view_update(struct llama_kv_cache_view * view, const struct llama_kv_cache & kv); - -// -// kv cache restore -// - -// saves the kv_cache state for future recovery. -// used to rollback llama_kv_cache_find_slot changes. -struct llama_kv_slot_restorer { - struct llama_kv_cache_state { - uint32_t head = 0; - uint32_t n = 0; - } old_state; - - // for non-recurrent models only - // list of slots to restore - std::vector> slot_boundaries; - - bool do_restore = false; - - explicit llama_kv_slot_restorer(const struct llama_kv_cache & cache) { - old_state.head = cache.head; - old_state.n = cache.n; - } - - // saves a slot information for future restoration - void save(const struct llama_kv_cache_slot_info & slot) { - if (slot) { - do_restore = true; - if (slot.boundaries.first != slot.boundaries.second) { - slot_boundaries.push_back(slot.boundaries); - } - } - } - - // must be explicitly called to restore the kv_cache state - // and rollback changes from all llama_kv_cache_find_slot calls - void restore(struct llama_kv_cache & cache) { - if (do_restore) { - cache.head = old_state.head; - cache.n = old_state.n; - - if (cache.recurrent) { // recurrent models like Mamba or RWKV can't have a state partially erased - llama_kv_cache_seq_rm(cache, -1, -1, -1); - } else { - for (auto & slot : slot_boundaries) { - llama_kv_cache_seq_rm(cache, -1, slot.first, slot.second); - } - } - } - } -}; +llama_kv_cache_view llama_kv_cache_view_init(const llama_kv_cache & kv, int32_t n_seq_max); +void llama_kv_cache_view_update(llama_kv_cache_view * view, const llama_kv_cache * kv); diff --git a/src/llama-memory.cpp b/src/llama-memory.cpp new file mode 100644 index 000000000..10173253e --- /dev/null +++ b/src/llama-memory.cpp @@ -0,0 +1 @@ +#include "llama-memory.h" diff --git a/src/llama-memory.h b/src/llama-memory.h new file mode 100644 index 000000000..dfa8c4e90 --- /dev/null +++ b/src/llama-memory.h @@ -0,0 +1,21 @@ +#pragma once + +#include "llama.h" + +// general concept of LLM memory +// the KV cache is a type of LLM memory, but there can be other types +class llama_memory_i { +public: + virtual void clear() = 0; + virtual void defrag() = 0; + + virtual bool seq_rm (llama_seq_id seq_id, llama_pos p0, llama_pos p1) = 0; + virtual void seq_cp (llama_seq_id seq_id_src, llama_seq_id seq_id_dst, llama_pos p0, llama_pos p1) = 0; + virtual void seq_keep(llama_seq_id seq_id) = 0; + virtual void seq_add (llama_seq_id seq_id, llama_pos p0, llama_pos p1, llama_pos delta) = 0; + virtual void seq_div (llama_seq_id seq_id, llama_pos p0, llama_pos p1, int d) = 0; + + virtual llama_pos seq_pos_max(llama_seq_id seq_id) const = 0; + + virtual bool get_can_edit() const = 0; +}; diff --git a/src/llama-mmap.cpp b/src/llama-mmap.cpp index b716630a8..9da97f1bc 100644 --- a/src/llama-mmap.cpp +++ b/src/llama-mmap.cpp @@ -8,6 +8,7 @@ #include #include #include +#include #ifdef __has_include #if __has_include() @@ -34,6 +35,10 @@ #include #endif +#if defined(__APPLE__) +#include +#endif + // TODO: consider moving to llama-impl.h if needed in more places #if defined(_WIN32) static std::string llama_format_win_err(DWORD err) { @@ -471,7 +476,11 @@ struct llama_mlock::impl { char* errmsg = std::strerror(errno); bool suggest = (errno == ENOMEM); - +#if defined(TARGET_OS_VISION) || defined(TARGET_OS_TV) || defined(_AIX) + // visionOS/tvOS dont't support RLIMIT_MEMLOCK + // Skip resource limit checks on visionOS/tvOS + suggest = false; +#else struct rlimit lock_limit; if (suggest && getrlimit(RLIMIT_MEMLOCK, &lock_limit)) { suggest = false; @@ -479,6 +488,7 @@ struct llama_mlock::impl { if (suggest && (lock_limit.rlim_max > lock_limit.rlim_cur + size)) { suggest = false; } +#endif LLAMA_LOG_WARN("warning: failed to mlock %zu-byte buffer (after previously locking %zu bytes): %s\n%s", size, this->size, errmsg, suggest ? MLOCK_SUGGESTION : ""); diff --git a/src/llama-mmap.h b/src/llama-mmap.h index 1da9ecb6b..4e5aec3f4 100644 --- a/src/llama-mmap.h +++ b/src/llama-mmap.h @@ -1,5 +1,6 @@ #pragma once +#include #include #include diff --git a/src/llama-model-loader.cpp b/src/llama-model-loader.cpp index 05d58ad90..ea73a8a7b 100644 --- a/src/llama-model-loader.cpp +++ b/src/llama-model-loader.cpp @@ -445,7 +445,8 @@ llama_model_loader::llama_model_loader( std::vector & splits, bool use_mmap, bool check_tensors, - const struct llama_model_kv_override * param_overrides_p) { + const llama_model_kv_override * param_overrides_p, + const llama_model_tensor_buft_override * param_tensor_buft_overrides_p) { int trace = 0; if (getenv("LLAMA_TRACE")) { trace = atoi(getenv("LLAMA_TRACE")); @@ -457,6 +458,8 @@ llama_model_loader::llama_model_loader( } } + tensor_buft_overrides = param_tensor_buft_overrides_p; + // Load the main GGUF struct ggml_context * ctx = NULL; struct gguf_init_params params = { @@ -600,7 +603,9 @@ llama_model_loader::llama_model_loader( if (trace > 0) { const uint16_t sid = w.idx; - LLAMA_LOG_INFO("%s: - tensor split %2d: %32s %-8s [ %s ]\n", __func__, sid, ggml_get_name(tensor), ggml_type_name(type), llama_format_tensor_shape(tensor).c_str()); + LLAMA_LOG_INFO("%s: - tensor split %2d: %32s %-8s [ %s ] %8.2f MiB\n", __func__, + sid, ggml_get_name(tensor), ggml_type_name(type), llama_format_tensor_shape(tensor).c_str(), + ggml_nbytes(tensor)/1024.0f/1024.0f); } } @@ -640,9 +645,9 @@ llama_model_loader::llama_model_loader( ftype = (llama_ftype) (ftype | LLAMA_FTYPE_GUESSED); { - const int kid = gguf_find_key(meta.get(), "general.file_type"); // TODO: use LLM_KV - if (kid >= 0) { - ftype = (llama_ftype) gguf_get_val_u32(meta.get(), kid); + uint32_t ftype_val = 0; + if (get_key(LLM_KV_GENERAL_FILE_TYPE, ftype_val, false)) { + ftype = (llama_ftype) ftype_val; } } diff --git a/src/llama-model-loader.h b/src/llama-model-loader.h index fe35404b2..0f52b011b 100644 --- a/src/llama-model-loader.h +++ b/src/llama-model-loader.h @@ -77,8 +77,9 @@ struct llama_model_loader { llama_mmaps mappings; - std::map weights_map; - std::unordered_map kv_overrides; + std::map weights_map; + std::unordered_map kv_overrides; + const llama_model_tensor_buft_override * tensor_buft_overrides; gguf_context_ptr meta; std::vector contexts; @@ -95,7 +96,8 @@ struct llama_model_loader { std::vector & splits, // optional, only need if the split does not follow naming scheme bool use_mmap, bool check_tensors, - const struct llama_model_kv_override * param_overrides_p); + const llama_model_kv_override * param_overrides_p, + const llama_model_tensor_buft_override * param_tensor_buft_overrides_p); template typename std::enable_if::value, bool>::type diff --git a/src/llama-model.cpp b/src/llama-model.cpp index 0f4b62c43..b74dd72cf 100644 --- a/src/llama-model.cpp +++ b/src/llama-model.cpp @@ -2,15 +2,22 @@ #include "llama-impl.h" #include "llama-mmap.h" +#include "llama-batch.h" +#include "llama-cparams.h" #include "llama-model-loader.h" +#include "llama-kv-cache.h" #include "ggml-cpp.h" #include #include +#include +#include #include +#include #include #include +#include #include #include @@ -26,6 +33,7 @@ const char * llm_type_name(llm_type type) { case LLM_TYPE_109M: return "109M"; case LLM_TYPE_137M: return "137M"; case LLM_TYPE_160M: return "160M"; + case LLM_TYPE_190M: return "190M"; case LLM_TYPE_220M: return "220M"; case LLM_TYPE_250M: return "250M"; case LLM_TYPE_270M: return "270M"; @@ -40,8 +48,10 @@ const char * llm_type_name(llm_type type) { case LLM_TYPE_1_4B: return "1.4B"; case LLM_TYPE_1_5B: return "1.5B"; case LLM_TYPE_1_6B: return "1.6B"; + case LLM_TYPE_1_8B: return "1.8B"; case LLM_TYPE_2B: return "2B"; case LLM_TYPE_2_8B: return "2.8B"; + case LLM_TYPE_2_9B: return "2.9B"; case LLM_TYPE_3B: return "3B"; case LLM_TYPE_4B: return "4B"; case LLM_TYPE_6B: return "6B"; @@ -79,6 +89,9 @@ const char * llm_type_name(llm_type type) { case LLM_TYPE_10B_128x3_66B: return "10B+128x3.66B"; case LLM_TYPE_57B_A14B: return "57B.A14B"; case LLM_TYPE_27B: return "27B"; + case LLM_TYPE_290B: return "290B"; + case LLM_TYPE_17B_16E: return "17Bx16E (Scout)"; + case LLM_TYPE_17B_128E: return "17Bx128E (Maverick)"; default: return "?B"; } } @@ -243,10 +256,11 @@ static ggml_backend_buffer_type_t select_weight_buft(const llama_hparams & hpara return cur_buft; } } + return nullptr; } -// CPU: ACCEL -> CPU extra -> GPU host -> CPU +// CPU: ACCEL -> GPU host -> CPU extra -> CPU static buft_list_t make_cpu_buft_list(const std::vector & devices) { buft_list_t buft_list; @@ -262,19 +276,6 @@ static buft_list_t make_cpu_buft_list(const std::vector & de } } - // add extra buffer types - auto * cpu_dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU); - auto * cpu_reg = ggml_backend_dev_backend_reg(cpu_dev); - auto ggml_backend_dev_get_extra_bufts_fn = (ggml_backend_dev_get_extra_bufts_t) - ggml_backend_reg_get_proc_address(cpu_reg, "ggml_backend_dev_get_extra_bufts"); - if (ggml_backend_dev_get_extra_bufts_fn) { - ggml_backend_buffer_type_t * extra_bufts = ggml_backend_dev_get_extra_bufts_fn(cpu_dev); - while (extra_bufts && *extra_bufts) { - buft_list.emplace_back(cpu_dev, *extra_bufts); - ++extra_bufts; - } - } - // add a host buffer type // storing the tensors in a host buffer is useful when the processing of large batches // is offloaded to a GPU device, since it reduces the time spent on data transfers @@ -289,6 +290,20 @@ static buft_list_t make_cpu_buft_list(const std::vector & de } } + // add extra buffer types, only if no GPU device is present + // ref: https://github.com/ggml-org/llama.cpp/issues/12481#issuecomment-2743136094 + auto * cpu_dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU); + auto * cpu_reg = ggml_backend_dev_backend_reg(cpu_dev); + auto ggml_backend_dev_get_extra_bufts_fn = (ggml_backend_dev_get_extra_bufts_t) + ggml_backend_reg_get_proc_address(cpu_reg, "ggml_backend_dev_get_extra_bufts"); + if (ggml_backend_dev_get_extra_bufts_fn) { + ggml_backend_buffer_type_t * extra_bufts = ggml_backend_dev_get_extra_bufts_fn(cpu_dev); + while (extra_bufts && *extra_bufts) { + buft_list.emplace_back(cpu_dev, *extra_bufts); + ++extra_bufts; + } + } + // add the CPU buffer type for (size_t i = 0; i < ggml_backend_dev_count(); ++i) { ggml_backend_dev_t dev = ggml_backend_dev_get(i); @@ -301,7 +316,7 @@ static buft_list_t make_cpu_buft_list(const std::vector & de } // GPU: split if LLAMA_SPLIT_MODE_ROW -> GPU -static buft_list_t make_gpu_buft_list(ggml_backend_dev_t dev, enum llama_split_mode split_mode, const float * tensor_split) { +static buft_list_t make_gpu_buft_list(ggml_backend_dev_t dev, llama_split_mode split_mode, const float * tensor_split) { buft_list_t buft_list; // add the device split buffer type if requested and available @@ -366,9 +381,12 @@ struct llama_model::impl { layer_dev dev_input = {}; layer_dev dev_output = {}; std::vector dev_layer; + + bool has_tensor_overrides; }; -llama_model::llama_model(const struct llama_model_params & params) : params(params), pimpl(std::make_unique()) { +llama_model::llama_model(const llama_model_params & params) : params(params), pimpl(std::make_unique()) { + pimpl->has_tensor_overrides = params.tensor_buft_overrides && params.tensor_buft_overrides[0].pattern; } llama_model::~llama_model() {} @@ -390,7 +408,7 @@ void llama_model::load_hparams(llama_model_loader & ml) { // get metadata as string for (int i = 0; i < gguf_get_n_kv(ctx); i++) { - enum gguf_type type = gguf_get_kv_type(ctx, i); + gguf_type type = gguf_get_kv_type(ctx, i); if (type == GGUF_TYPE_ARRAY) { continue; } @@ -468,6 +486,10 @@ void llama_model::load_hparams(llama_model_loader & ml) { } hparams.rope_freq_scale_train = ropescale == 0.0f ? 1.0f : 1.0f/ropescale; + // by default assume that the sliding-window layers use the same scaling type as the non-sliding-window layers + hparams.rope_freq_base_train_swa = hparams.rope_freq_base_train; + hparams.rope_freq_scale_train_swa = hparams.rope_freq_scale_train; + ml.get_key(LLM_KV_ROPE_SCALING_ATTN_FACTOR, hparams.rope_attn_factor, false); // non-transformer models do not have attention heads @@ -530,6 +552,25 @@ void llama_model::load_hparams(llama_model_loader & ml) { } } } break; + case LLM_ARCH_LLAMA4: + { + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); + ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp); + ml.get_key(LLM_KV_INTERLEAVE_MOE_LAYER_STEP, hparams.n_moe_layer_step); + hparams.n_swa_pattern = 4; // pattern: 3 chunked - 1 full + hparams.n_attn_chunk = 8192; // should this be a gguf kv? currently it's the same for Scout and Maverick + hparams.n_swa = 1; // TODO @ngxson : this is added to trigger the SWA branch (we store the chunked attn mask in the SWA tensor), will need to clean this up later + + switch (hparams.n_expert) { + case 16: type = LLM_TYPE_17B_16E; break; + case 128: type = LLM_TYPE_17B_128E; break; + default: type = LLM_TYPE_UNKNOWN; + } + + if (type == LLM_TYPE_17B_128E) { + hparams.use_kq_norm = false; + } + } break; case LLM_ARCH_DECI: { ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); @@ -746,6 +787,22 @@ void llama_model::load_hparams(llama_model_loader & ml) { default: type = LLM_TYPE_UNKNOWN; } } break; + case LLM_ARCH_QWEN3: + { + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); + switch (hparams.n_layer) { + default: type = LLM_TYPE_UNKNOWN; + } + } break; + case LLM_ARCH_QWEN3MOE: + { + ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp, false); + + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); + switch (hparams.n_layer) { + default: type = LLM_TYPE_UNKNOWN; + } + } break; case LLM_ARCH_PHI2: { ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps); @@ -773,9 +830,11 @@ void llama_model::load_hparams(llama_model_loader & ml) { hparams.n_swa = 2047; } else if (hparams.n_layer == 32 && hparams.n_head_kv(0) == 32 && hparams.n_ctx_train == 131072) { // default value for Phi-3-mini-128k-instruct + // note: this seems incorrect because the window is bigger than the train context? hparams.n_swa = 262144; } else if (hparams.n_layer == 40 && hparams.n_ctx_train == 131072) { // default value for Phi-3-medium-128k-instruct + // note: this seems incorrect because the window is equal to the train context? hparams.n_swa = 131072; } bool found_swa = ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa, false); @@ -851,11 +910,13 @@ void llama_model::load_hparams(llama_model_loader & ml) { case LLM_ARCH_GEMMA2: { hparams.n_swa = 4096; // default value of gemma 2 + hparams.n_swa_pattern = 2; + hparams.attn_soft_cap = true; + ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa, false); ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); ml.get_key(LLM_KV_ATTN_LOGIT_SOFTCAPPING, hparams.f_attn_logit_softcapping, false); ml.get_key(LLM_KV_FINAL_LOGIT_SOFTCAPPING, hparams.f_final_logit_softcapping, false); - hparams.attn_soft_cap = true; switch (hparams.n_layer) { case 26: type = LLM_TYPE_2B; break; @@ -864,6 +925,28 @@ void llama_model::load_hparams(llama_model_loader & ml) { default: type = LLM_TYPE_UNKNOWN; } } break; + case LLM_ARCH_GEMMA3: + { + hparams.n_swa_pattern = 6; + + hparams.rope_freq_base_train_swa = 10000.0f; + hparams.rope_freq_scale_train_swa = 1.0f; + + ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa); + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); + + switch (hparams.n_layer) { + case 26: type = LLM_TYPE_1B; break; + case 34: type = LLM_TYPE_4B; break; + case 48: type = LLM_TYPE_12B; break; + case 62: type = LLM_TYPE_27B; break; + default: type = LLM_TYPE_UNKNOWN; + } + + hparams.f_attention_scale = type == LLM_TYPE_27B + ? 1.0f / std::sqrt(float(hparams.n_embd / hparams.n_head(0))) + : 1.0f / std::sqrt(float(hparams.n_embd_head_k)); + } break; case LLM_ARCH_STARCODER2: { ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps); @@ -928,6 +1011,8 @@ void llama_model::load_hparams(llama_model_loader & ml) { } break; case LLM_ARCH_COHERE2: { + hparams.n_swa_pattern = 4; + ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa); ml.get_key(LLM_KV_LOGIT_SCALE, hparams.f_logit_scale); ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps); @@ -966,6 +1051,7 @@ void llama_model::load_hparams(llama_model_loader & ml) { case 16: type = LLM_TYPE_1B; break; case 32: type = LLM_TYPE_7B; break; case 40: type = LLM_TYPE_13B; break; + case 64: type = LLM_TYPE_32B; break; default: type = LLM_TYPE_UNKNOWN; } } break; @@ -1089,6 +1175,15 @@ void llama_model::load_hparams(llama_model_loader & ml) { default: type = LLM_TYPE_UNKNOWN; } } break; + case LLM_ARCH_PLM: + { + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); + ml.get_key(LLM_KV_ATTENTION_KV_LORA_RANK, hparams.n_lora_kv); + switch (hparams.n_layer) { + case 32: type = LLM_TYPE_1_8B; break; + default: type = LLM_TYPE_UNKNOWN; + } + } break; case LLM_ARCH_CHATGLM: { ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); @@ -1110,6 +1205,15 @@ void llama_model::load_hparams(llama_model_loader & ml) { default: type = LLM_TYPE_UNKNOWN; } } break; + case LLM_ARCH_GLM4: + { + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); + switch (hparams.n_layer) { + case 40: type = LLM_TYPE_9B; break; + case 61: type = LLM_TYPE_32B; break; + default: type = LLM_TYPE_UNKNOWN; + } + } break; case LLM_ARCH_BITNET: { ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); @@ -1210,6 +1314,36 @@ void llama_model::load_hparams(llama_model_loader & ml) { default: type = LLM_TYPE_UNKNOWN; } } break; + case LLM_ARCH_RWKV7: + case LLM_ARCH_ARWKV7: + { + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps, false); + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps, false); + ml.get_key(LLM_KV_WKV_HEAD_SIZE, hparams.wkv_head_size); + ml.get_key(LLM_KV_ATTENTION_DECAY_LORA_RANK, hparams.n_lora_decay); + ml.get_key(LLM_KV_ATTENTION_ICLR_LORA_RANK, hparams.n_lora_iclr); + ml.get_key(LLM_KV_ATTENTION_VALUE_RESIDUAL_MIX_LORA_RANK, hparams.n_lora_value_res_mix); + ml.get_key(LLM_KV_ATTENTION_GATE_LORA_RANK, hparams.n_lora_gate, false); + ml.get_key(LLM_KV_TOKEN_SHIFT_COUNT, hparams.token_shift_count, false); + + switch (hparams.n_layer) { + case 12: type = LLM_TYPE_190M; break; + case 24: + switch (hparams.n_embd) { + case 1024: type = LLM_TYPE_450M; break; + case 2048: type = LLM_TYPE_1_5B; break; + default: type = LLM_TYPE_UNKNOWN; + } break; + case 28: + switch (hparams.n_embd) { + case 1536: type = LLM_TYPE_1_5B; break; + case 3584: type = LLM_TYPE_7B; break; + default: type = LLM_TYPE_UNKNOWN; + } break; + case 32: type = LLM_TYPE_2_9B; break; // RWKV-7-World + default: type = LLM_TYPE_UNKNOWN; + } + } break; case LLM_ARCH_GRANITE: case LLM_ARCH_GRANITE_MOE: { @@ -1245,6 +1379,21 @@ void llama_model::load_hparams(llama_model_loader & ml) { ml.get_key(LLM_KV_ATTENTION_GROUPNORM_GROUPS, hparams.n_norm_groups); ml.get_key(LLM_KV_ATTENTION_CAUSAL, hparams.causal_attn); } break; + case LLM_ARCH_BAILINGMOE: + { + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); + ml.get_key(LLM_KV_LEADING_DENSE_BLOCK_COUNT, hparams.n_layer_dense_lead); + ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp); + ml.get_key(LLM_KV_EXPERT_SHARED_COUNT, hparams.n_expert_shared); + ml.get_key(LLM_KV_EXPERT_WEIGHTS_SCALE, hparams.expert_weights_scale); + ml.get_key(LLM_KV_EXPERT_WEIGHTS_NORM, hparams.expert_weights_norm, false); + + switch (hparams.n_layer) { + case 28: type = LLM_TYPE_16B; break; + case 88: type = LLM_TYPE_290B; break; + default: type = LLM_TYPE_UNKNOWN; + } + } break; default: throw std::runtime_error("unsupported model architecture"); } @@ -1316,13 +1465,14 @@ bool llama_model::load_tensors(llama_model_loader & ml) { const int i_gpu_start = std::max((int) hparams.n_layer - n_gpu_layers, (int) 0); const int act_gpu_layers = devices.empty() ? 0 : std::min(n_gpu_layers, (int)n_layer + 1); auto get_layer_buft_list = [&](int il) -> llama_model::impl::layer_dev { + const bool is_swa = il < (int) hparams.n_layer && hparams.is_swa(il); if (il < i_gpu_start || (il - i_gpu_start) >= act_gpu_layers) { - LLAMA_LOG_DEBUG("load_tensors: layer %3d assigned to device %s\n", il, ggml_backend_dev_name(cpu_dev)); + LLAMA_LOG_DEBUG("load_tensors: layer %3d assigned to device %s, is_swa = %d\n", il, ggml_backend_dev_name(cpu_dev), is_swa); return {cpu_dev, &pimpl->cpu_buft_list}; } const int layer_gpu = std::upper_bound(splits.begin(), splits.begin() + n_devices(), float(il - i_gpu_start)/act_gpu_layers) - splits.begin(); auto * dev = devices.at(layer_gpu); - LLAMA_LOG_DEBUG("load_tensors: layer %3d assigned to device %s\n", il, ggml_backend_dev_name(dev)); + LLAMA_LOG_DEBUG("load_tensors: layer %3d assigned to device %s, is_swa = %d\n", il, ggml_backend_dev_name(dev), is_swa); return {dev, &pimpl->gpu_buft_list.at(dev)}; }; @@ -1424,6 +1574,17 @@ bool llama_model::load_tensors(llama_model_loader & ml) { throw std::runtime_error(format("missing tensor info mapping for %s", tn.str().c_str())); } + // skip unused tensors + if (info.op == GGML_OP_NONE) { + const size_t nbytes = ggml_nbytes(t_meta); + LLAMA_LOG_WARN("model has unused tensor %s (size = %zu bytes) -- ignoring\n", tn.str().c_str(), nbytes); + + ml.size_data -= nbytes; + ml.n_created++; + + return nullptr; + } + // tensors with "bias" suffix are always used with GGML_OP_ADD ggml_op op; bool bias = tn.suffix != nullptr && strcmp(tn.suffix, "bias") == 0; @@ -1460,9 +1621,26 @@ bool llama_model::load_tensors(llama_model_loader & ml) { GGML_ABORT("invalid layer %d for tensor %s", info.layer, tn.str().c_str()); } - ggml_backend_buffer_type_t buft = select_weight_buft(hparams, t_meta, op, *buft_list); + ggml_backend_buffer_type_t buft = nullptr; + + // check overrides + if (ml.tensor_buft_overrides) { + std::string tensor_name = tn.str(); + for (const auto * overrides = ml.tensor_buft_overrides; overrides->pattern != nullptr; ++overrides) { + std::regex pattern(overrides->pattern); + if (std::regex_search(tensor_name, pattern)) { + LLAMA_LOG_DEBUG("tensor %s buffer type overriden to %s\n", tensor_name.c_str(), ggml_backend_buft_name(overrides->buft)); + buft = overrides->buft; + break; + } + } + } + if (!buft) { - throw std::runtime_error(format("failed to find a compatible buffer type for tensor %s", tn.str().c_str())); + buft = select_weight_buft(hparams, t_meta, op, *buft_list); + if (!buft) { + throw std::runtime_error(format("failed to find a compatible buffer type for tensor %s", tn.str().c_str())); + } } // avoid using a host buffer when using mmap @@ -1558,6 +1736,56 @@ bool llama_model::load_tensors(llama_model_loader & ml) { } } } break; + case LLM_ARCH_LLAMA4: + { + tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0); + + // output + output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0); + output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED); + + // if output is NULL, init from the input tok embed + if (output == NULL) { + output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED); + } + + GGML_ASSERT(hparams.n_moe_layer_step > 0 && "Llama 4 requires n_moe_layer_step > 0"); + for (int i = 0; i < n_layer; ++i) { + bool is_moe_layer = (i + 1) % hparams.n_moe_layer_step == 0; + + auto & layer = layers[i]; + + layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0); + + layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd_head_k * n_head}, 0); + layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd_k_gqa}, 0); + layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_v_gqa}, 0); + layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0); + + layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0); + + layer.rope_freqs = create_tensor(tn(LLM_TENSOR_ROPE_FREQS, "weight", i), {n_rot/2}, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0)); + + if (is_moe_layer) { + int n_ff_exp = hparams.n_ff_exp; + + layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}, 0); + layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd, n_ff_exp, n_expert}, 0); + layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), { n_ff_exp, n_embd, n_expert}, 0); + layer.ffn_up_exps = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS, "weight", i), {n_embd, n_ff_exp, n_expert}, 0); + + // Shared expert + const int64_t n_ff_shexp = n_ff_exp; + layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), { n_embd, n_ff_shexp}, 0); + layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {n_ff_shexp, n_embd }, 0); + layer.ffn_up_shexp = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP, "weight", i), { n_embd, n_ff_shexp}, 0); + } else { + layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, 0); + layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd}, 0); + layer.ffn_up = create_tensor(tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, 0); + } + } + } break; case LLM_ARCH_DECI: { tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0); @@ -1936,7 +2164,12 @@ bool llama_model::load_tensors(llama_model_loader & ml) { // output output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0); output_norm_b = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "bias"), {n_embd}, 0); - output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, 0); + output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED); + + // if output is NULL, init from the input tok embed + if (output == NULL) { + output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED); + } for (int i = 0; i < n_layer; ++i) { auto & layer = layers[i]; @@ -2121,9 +2354,9 @@ bool llama_model::load_tensors(llama_model_loader & ml) { layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0); // optional bias tensors - layer.bq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "bias", i), {n_embd}, 0); - layer.bk = create_tensor(tn(LLM_TENSOR_ATTN_K, "bias", i), {n_embd_gqa}, 0); - layer.bv = create_tensor(tn(LLM_TENSOR_ATTN_V, "bias", i), {n_embd_gqa}, 0); + layer.bq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "bias", i), {n_embd}, TENSOR_NOT_REQUIRED); + layer.bk = create_tensor(tn(LLM_TENSOR_ATTN_K, "bias", i), {n_embd_gqa}, TENSOR_NOT_REQUIRED); + layer.bv = create_tensor(tn(LLM_TENSOR_ATTN_V, "bias", i), {n_embd_gqa}, TENSOR_NOT_REQUIRED); layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0); @@ -2152,6 +2385,77 @@ bool llama_model::load_tensors(llama_model_loader & ml) { layer.ffn_up_shexp = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP, "weight", i), { n_embd, n_ff_shexp}, 0); } } break; + case LLM_ARCH_QWEN3: + { + tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0); + + // output + output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0); + output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED); + // if output is NULL, init from the input tok embed + if (output == NULL) { + output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED); + } + + for (int i = 0; i < n_layer; ++i) { + auto & layer = layers[i]; + + layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0); + + layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd_head_k * n_head}, 0); + layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd_gqa}, 0); + layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_gqa}, 0); + layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0); + + layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_head_k}, 0); + layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head_k}, 0); + + layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0); + layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, 0); + layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd}, 0); + layer.ffn_up = create_tensor(tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, 0); + } + } break; + case LLM_ARCH_QWEN3MOE: + { + tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0); + + // output + output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0); + output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, 0); + + for (int i = 0; i < n_layer; ++i) { + auto & layer = layers[i]; + + layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0); + + layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd_head_k * n_head}, 0); + layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd_gqa}, 0); + layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_gqa}, 0); + layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0); + + layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_head_k}, 0); + layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head_k}, 0); + + layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0); + + layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}, 0); + + if (n_expert == 0) { + throw std::runtime_error("n_expert must be > 0 for QWEN3MOE"); + } + if (n_expert_used == 0) { + throw std::runtime_error("n_expert_used must be > 0 for QWEN3MOE"); + } + + // MoE branch + const int64_t n_ff_exp = hparams.n_ff_exp ? hparams.n_ff_exp : n_ff / n_expert_used; + + layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), { n_embd, n_ff_exp, n_expert}, 0); + layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp, n_embd, n_expert}, 0); + layer.ffn_up_exps = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS, "weight", i), { n_embd, n_ff_exp, n_expert}, 0); + } + } break; case LLM_ARCH_PHI2: { tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0); @@ -2194,13 +2498,16 @@ bool llama_model::load_tensors(llama_model_loader & ml) { } break; case LLM_ARCH_PHI3: { - const int64_t n_embd_head = n_embd / n_head; - tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), { n_embd, n_vocab }, 0); // output output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), { n_embd }, 0); - output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), { n_embd, n_vocab }, 0); + output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED); + + // if output is NULL, init from the input tok embed + if (output == NULL) { + output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED); + } for (int i = 0; i < n_layer; ++i) { auto & layer = layers[i]; @@ -2215,8 +2522,8 @@ bool llama_model::load_tensors(llama_model_loader & ml) { layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd }, 0); layer.ffn_up = create_tensor(tn(LLM_TENSOR_FFN_UP, "weight", i), { n_embd, 2 * n_ff }, 0); - layer.rope_long = create_tensor(tn(LLM_TENSOR_ROPE_FACTORS_LONG, "weight", i), { n_embd_head/2 }, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0)); - layer.rope_short = create_tensor(tn(LLM_TENSOR_ROPE_FACTORS_SHORT, "weight", i), { n_embd_head/2 }, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0)); + layer.rope_long = create_tensor(tn(LLM_TENSOR_ROPE_FACTORS_LONG, "weight", i), { n_rot/2 }, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0)); + layer.rope_short = create_tensor(tn(LLM_TENSOR_ROPE_FACTORS_SHORT, "weight", i), { n_rot/2 }, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0)); } } break; case LLM_ARCH_PHIMOE: @@ -2237,7 +2544,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) { layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), { n_embd }, 0); layer.attn_norm_b = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "bias", i), { n_embd }, 0); - layer.wqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", i), { n_embd, n_embd + 2 * n_embd_gqa }, llama_model_loader::TENSOR_NOT_REQUIRED); + layer.wqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", i), { n_embd, n_embd + 2 * n_embd_gqa }, TENSOR_NOT_REQUIRED); if (layer.wqkv == nullptr) { layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd}, 0); layer.bq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "bias", i), {n_embd}, 0); @@ -2294,7 +2601,12 @@ bool llama_model::load_tensors(llama_model_loader & ml) { // output output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0); output_norm_b = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "bias"), {n_embd}, 0); - output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, 0); + output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED); + + // if output is NULL, init from the input tok embed + if (output == NULL) { + output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED); + } for (int i = 0; i < n_layer; ++i) { auto & layer = layers[i]; @@ -2320,7 +2632,12 @@ bool llama_model::load_tensors(llama_model_loader & ml) { } break; case LLM_ARCH_CODESHELL: { - tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0); + tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED); + + // if tok embd is NULL, init from output + if (tok_embd == NULL) { + tok_embd = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED); + } // output output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0); @@ -2443,6 +2760,40 @@ bool llama_model::load_tensors(llama_model_loader & ml) { layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0); layer.attn_post_norm = create_tensor(tn(LLM_TENSOR_ATTN_POST_NORM, "weight", i), {n_embd}, 0); + layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0); + layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, 0); + layer.ffn_up = create_tensor(tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, 0); + layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd}, 0); + layer.ffn_post_norm = create_tensor(tn(LLM_TENSOR_FFN_POST_NORM, "weight", i), {n_embd}, 0); + } + } break; + case LLM_ARCH_GEMMA3: + { + tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0); + + // output + output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0); + output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED); + + // if output is NULL, init from the input tok embed + if (output == NULL) { + output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED); + } + + for (int i = 0; i < n_layer; ++i) { + auto & layer = layers[i]; + + layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0); + + layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd_head_k * n_head}, 0); + layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd_k_gqa}, 0); + layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_v_gqa}, 0); + layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0); + + layer.attn_post_norm = create_tensor(tn(LLM_TENSOR_ATTN_POST_NORM, "weight", i), {n_embd}, 0); + layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_head_k}, 0); + layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head_k}, 0); + layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0); layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, 0); layer.ffn_up = create_tensor(tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, 0); @@ -2643,6 +2994,8 @@ bool llama_model::load_tensors(llama_model_loader & ml) { } break; case LLM_ARCH_OLMO2: { + const int64_t n_embd_head = n_embd / n_head; + tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0); // output @@ -2657,7 +3010,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) { layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_gqa}, 0); layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0); layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd}, 0); - layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd}, 0); + layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_head_kv * n_embd_head}, 0); layer.attn_post_norm = create_tensor(tn(LLM_TENSOR_ATTN_POST_NORM, "weight", i), {n_embd}, 0); layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, 0); @@ -2917,6 +3270,35 @@ bool llama_model::load_tensors(llama_model_loader & ml) { } } } break; + case LLM_ARCH_PLM: + { + const int64_t n_embd_head_qk_rope = hparams.n_rot; + const int64_t n_embd_head_qk_nope = hparams.n_embd_head_k - hparams.n_rot; + const int64_t kv_lora_rank = hparams.n_lora_kv; + + tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0); + + // output + output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0); + // output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, 0); + output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED); + + for (int i = 0; i < n_layer; ++i) { + auto & layer = layers[i]; + + layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0); + + layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd_head_k * n_head}, 0); + layer.wkv_a_mqa = create_tensor(tn(LLM_TENSOR_ATTN_KV_A_MQA, "weight", i), {n_embd, kv_lora_rank + (n_embd_head_qk_rope)}, 0); + layer.attn_kv_a_norm = create_tensor(tn(LLM_TENSOR_ATTN_KV_A_NORM, "weight", i), {kv_lora_rank}, 0); + layer.wkv_b = create_tensor(tn(LLM_TENSOR_ATTN_KV_B, "weight", i), {kv_lora_rank, n_head * (n_embd_head_qk_nope + n_embd_head_v)}, 0); + layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), { n_head * ( n_embd_head_v), n_embd}, 0); + + layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0); + layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd}, 0); + layer.ffn_up = create_tensor(tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, 0); + } + } break; case LLM_ARCH_BITNET: { tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0); @@ -3082,16 +3464,16 @@ bool llama_model::load_tensors(llama_model_loader & ml) { auto & layer = layers[i]; layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0); - layer.wqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}, llama_model_loader::TENSOR_NOT_REQUIRED); - layer.bqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "bias", i), {n_embd + 2*n_embd_gqa}, llama_model_loader::TENSOR_NOT_REQUIRED); + layer.wqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}, TENSOR_NOT_REQUIRED); + layer.bqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "bias", i), {n_embd + 2*n_embd_gqa}, TENSOR_NOT_REQUIRED); if (layer.wqkv == nullptr) { layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd_head_k * n_head}, 0); layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd_k_gqa}, 0); layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_v_gqa}, 0); - layer.bq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "bias", i), {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED); - layer.bk = create_tensor(tn(LLM_TENSOR_ATTN_K, "bias", i), {n_embd_gqa}, llama_model_loader::TENSOR_NOT_REQUIRED); - layer.bv = create_tensor(tn(LLM_TENSOR_ATTN_V, "bias", i), {n_embd_gqa}, llama_model_loader::TENSOR_NOT_REQUIRED); + layer.bq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "bias", i), {n_embd}, TENSOR_NOT_REQUIRED); + layer.bk = create_tensor(tn(LLM_TENSOR_ATTN_K, "bias", i), {n_embd_gqa}, TENSOR_NOT_REQUIRED); + layer.bv = create_tensor(tn(LLM_TENSOR_ATTN_V, "bias", i), {n_embd_gqa}, TENSOR_NOT_REQUIRED); } layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0); @@ -3103,6 +3485,45 @@ bool llama_model::load_tensors(llama_model_loader & ml) { layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, 0); } } break; + case LLM_ARCH_GLM4: + { + tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0); + + // output + output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0); + output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED); + // if output is NULL, init from the input tok embed + if (output == NULL) { + output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED); + } + + for (int i = 0; i < n_layer; ++i) { + auto & layer = layers[i]; + + layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0); + layer.wqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}, TENSOR_NOT_REQUIRED); + layer.bqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "bias", i), {n_embd + 2*n_embd_gqa}, TENSOR_NOT_REQUIRED); + + if (layer.wqkv == nullptr) { + layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd_head_k * n_head}, 0); + layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd_k_gqa}, 0); + layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_v_gqa}, 0); + layer.bq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "bias", i), {n_embd}, TENSOR_NOT_REQUIRED); + layer.bk = create_tensor(tn(LLM_TENSOR_ATTN_K, "bias", i), {n_embd_gqa}, TENSOR_NOT_REQUIRED); + layer.bv = create_tensor(tn(LLM_TENSOR_ATTN_V, "bias", i), {n_embd_gqa}, TENSOR_NOT_REQUIRED); + } + + layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0); + + layer.attn_post_norm = create_tensor(tn(LLM_TENSOR_ATTN_POST_NORM, "weight", i), {n_embd}, 0); + + layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0); + layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd}, 0); + layer.ffn_up = create_tensor(tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff * 2}, 0); + + layer.ffn_post_norm = create_tensor(tn(LLM_TENSOR_FFN_POST_NORM, "weight", i), {n_embd}, 0); + } + } break; case LLM_ARCH_NEMOTRON: { tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0); @@ -3146,7 +3567,12 @@ bool llama_model::load_tensors(llama_model_loader & ml) { // output output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0); - output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, 0); + output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED); + + // if output is NULL, init from the input tok embed + if (output == NULL) { + output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED); + } for (int i = 0; i < n_layer; ++i) { auto & layer = layers[i]; @@ -3197,12 +3623,12 @@ bool llama_model::load_tensors(llama_model_loader & ml) { layer.time_mix_w2 = create_tensor(tn(LLM_TENSOR_TIME_MIX_W2, "weight", i), {time_mix_extra_dim, n_embd, 5}, 0); layer.time_mix_lerp_x = create_tensor(tn(LLM_TENSOR_TIME_MIX_LERP_X, "weight", i), {n_embd, 1, 1}, 0); - layer.time_mix_lerp_w = create_tensor(tn(LLM_TENSOR_TIME_MIX_LERP_W, "weight", i), {n_embd, 1, 1}, llama_model_loader::TENSOR_NOT_REQUIRED); - layer.time_mix_lerp_k = create_tensor(tn(LLM_TENSOR_TIME_MIX_LERP_K, "weight", i), {n_embd, 1, 1}, llama_model_loader::TENSOR_NOT_REQUIRED); - layer.time_mix_lerp_v = create_tensor(tn(LLM_TENSOR_TIME_MIX_LERP_V, "weight", i), {n_embd, 1, 1}, llama_model_loader::TENSOR_NOT_REQUIRED); - layer.time_mix_lerp_r = create_tensor(tn(LLM_TENSOR_TIME_MIX_LERP_R, "weight", i), {n_embd, 1, 1}, llama_model_loader::TENSOR_NOT_REQUIRED); - layer.time_mix_lerp_g = create_tensor(tn(LLM_TENSOR_TIME_MIX_LERP_G, "weight", i), {n_embd, 1, 1}, llama_model_loader::TENSOR_NOT_REQUIRED); - layer.time_mix_lerp_fused = create_tensor(tn(LLM_TENSOR_TIME_MIX_LERP_FUSED, "weight", i), {n_embd, 1, 1, 5}, llama_model_loader::TENSOR_NOT_REQUIRED); + layer.time_mix_lerp_w = create_tensor(tn(LLM_TENSOR_TIME_MIX_LERP_W, "weight", i), {n_embd, 1, 1}, TENSOR_NOT_REQUIRED); + layer.time_mix_lerp_k = create_tensor(tn(LLM_TENSOR_TIME_MIX_LERP_K, "weight", i), {n_embd, 1, 1}, TENSOR_NOT_REQUIRED); + layer.time_mix_lerp_v = create_tensor(tn(LLM_TENSOR_TIME_MIX_LERP_V, "weight", i), {n_embd, 1, 1}, TENSOR_NOT_REQUIRED); + layer.time_mix_lerp_r = create_tensor(tn(LLM_TENSOR_TIME_MIX_LERP_R, "weight", i), {n_embd, 1, 1}, TENSOR_NOT_REQUIRED); + layer.time_mix_lerp_g = create_tensor(tn(LLM_TENSOR_TIME_MIX_LERP_G, "weight", i), {n_embd, 1, 1}, TENSOR_NOT_REQUIRED); + layer.time_mix_lerp_fused = create_tensor(tn(LLM_TENSOR_TIME_MIX_LERP_FUSED, "weight", i), {n_embd, 1, 1, 5}, TENSOR_NOT_REQUIRED); GGML_ASSERT(!(layer.time_mix_lerp_fused == NULL && layer.time_mix_lerp_w == NULL)); layer.time_mix_first = create_tensor(tn(LLM_TENSOR_TIME_MIX_FIRST, "weight", i), {head_size, n_embd / head_size}, 0); @@ -3232,7 +3658,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) { tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0); output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0); - output_norm_b = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "bias"), {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED); + output_norm_b = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "bias"), {n_embd}, TENSOR_NOT_REQUIRED); output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, 0); const int time_mix_extra_dim = hparams.time_mix_extra_dim; @@ -3258,7 +3684,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) { layer.time_mix_lerp_x = create_tensor(tn(LLM_TENSOR_TIME_MIX_LERP_X, "weight", i), {n_embd, 1, 1}, 0); layer.time_mix_lerp_fused = create_tensor(tn(LLM_TENSOR_TIME_MIX_LERP_FUSED, "weight", i), {n_embd, 1, 1, 5}, 0); - layer.time_mix_first = create_tensor(tn(LLM_TENSOR_TIME_MIX_FIRST, "weight", i), {head_size, n_embd / head_size}, llama_model_loader::TENSOR_NOT_REQUIRED); + layer.time_mix_first = create_tensor(tn(LLM_TENSOR_TIME_MIX_FIRST, "weight", i), {head_size, n_embd / head_size}, TENSOR_NOT_REQUIRED); layer.time_mix_decay = create_tensor(tn(LLM_TENSOR_TIME_MIX_DECAY, "weight", i), {n_embd}, 0); layer.time_mix_decay_w1 = create_tensor(tn(LLM_TENSOR_TIME_MIX_DECAY_W1, "weight", i), {n_embd, time_decay_extra_dim}, 0); layer.time_mix_decay_w2 = create_tensor(tn(LLM_TENSOR_TIME_MIX_DECAY_W2, "weight", i), {time_decay_extra_dim, attn_hidden_size}, 0); @@ -3267,9 +3693,9 @@ bool llama_model::load_tensors(llama_model_loader & ml) { layer.time_mix_receptance = create_tensor(tn(LLM_TENSOR_TIME_MIX_RECEPTANCE, "weight", i), {attn_hidden_size, n_embd}, 0); layer.time_mix_gate = create_tensor(tn(LLM_TENSOR_TIME_MIX_GATE, "weight", i), {attn_hidden_size, n_embd}, 0); // optional bias tensors - layer.time_mix_key_b = create_tensor(tn(LLM_TENSOR_TIME_MIX_KEY, "bias", i), {attn_key_value_size}, llama_model_loader::TENSOR_NOT_REQUIRED); - layer.time_mix_value_b = create_tensor(tn(LLM_TENSOR_TIME_MIX_VALUE, "bias", i), {attn_key_value_size}, llama_model_loader::TENSOR_NOT_REQUIRED); - layer.time_mix_receptance_b = create_tensor(tn(LLM_TENSOR_TIME_MIX_RECEPTANCE, "bias", i), {attn_hidden_size}, llama_model_loader::TENSOR_NOT_REQUIRED); + layer.time_mix_key_b = create_tensor(tn(LLM_TENSOR_TIME_MIX_KEY, "bias", i), {attn_key_value_size}, TENSOR_NOT_REQUIRED); + layer.time_mix_value_b = create_tensor(tn(LLM_TENSOR_TIME_MIX_VALUE, "bias", i), {attn_key_value_size}, TENSOR_NOT_REQUIRED); + layer.time_mix_receptance_b = create_tensor(tn(LLM_TENSOR_TIME_MIX_RECEPTANCE, "bias", i), {attn_hidden_size}, TENSOR_NOT_REQUIRED); layer.time_mix_output = create_tensor(tn(LLM_TENSOR_TIME_MIX_OUTPUT, "weight", i), {n_embd, attn_hidden_size}, 0); @@ -3280,6 +3706,146 @@ bool llama_model::load_tensors(llama_model_loader & ml) { layer.ffn_up = create_tensor(tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, 0); } } break; + case LLM_ARCH_RWKV7: + { + tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0); + + // Block 0, LN0 + tok_norm = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD_NORM, "weight"), {n_embd}, 0); + tok_norm_b = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD_NORM, "bias"), {n_embd}, 0); + + // output + output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0); + output_norm_b = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "bias"), {n_embd}, 0); + output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, 0); + + const int n_lora_decay = hparams.n_lora_decay; + const int n_lora_iclr = hparams.n_lora_iclr; + const int n_lora_value_res_mix = hparams.n_lora_value_res_mix; + const int n_lora_gate = hparams.n_lora_gate; + const int attn_hidden_size = n_embd; + const int ffn_size = hparams.n_ff_arr[0]; + + for (int i = 0; i < n_layer; ++i) { + auto & layer = layers[i]; + + layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0); + layer.attn_norm_b = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "bias", i), {n_embd}, 0); + + layer.attn_norm_2 = create_tensor(tn(LLM_TENSOR_ATTN_NORM_2, "weight", i), {n_embd}, 0); + layer.attn_norm_2_b = create_tensor(tn(LLM_TENSOR_ATTN_NORM_2, "bias", i), {n_embd}, 0); + + layer.time_mix_w0 = create_tensor(tn(LLM_TENSOR_TIME_MIX_W0, "weight", i), {n_embd}, 0); + layer.time_mix_w1 = create_tensor(tn(LLM_TENSOR_TIME_MIX_W1, "weight", i), {n_embd, n_lora_decay}, 0); + layer.time_mix_w2 = create_tensor(tn(LLM_TENSOR_TIME_MIX_W2, "weight", i), {n_lora_decay, n_embd}, 0); + + layer.time_mix_a0 = create_tensor(tn(LLM_TENSOR_TIME_MIX_A0, "weight", i), {n_embd}, 0); + layer.time_mix_a1 = create_tensor(tn(LLM_TENSOR_TIME_MIX_A1, "weight", i), {n_embd, n_lora_iclr}, 0); + layer.time_mix_a2 = create_tensor(tn(LLM_TENSOR_TIME_MIX_A2, "weight", i), {n_lora_iclr, n_embd}, 0); + + if (i == 0) { + // actually not used + layer.time_mix_v0 = create_tensor(tn(LLM_TENSOR_TIME_MIX_V0, "weight", i), {n_embd}, 0); + layer.time_mix_v1 = create_tensor(tn(LLM_TENSOR_TIME_MIX_V1, "weight", i), {n_embd, n_lora_iclr}, 0); + layer.time_mix_v2 = create_tensor(tn(LLM_TENSOR_TIME_MIX_V2, "weight", i), {n_lora_iclr, n_embd}, 0); + } else { + layer.time_mix_v0 = create_tensor(tn(LLM_TENSOR_TIME_MIX_V0, "weight", i), {n_embd}, 0); + layer.time_mix_v1 = create_tensor(tn(LLM_TENSOR_TIME_MIX_V1, "weight", i), {n_embd, n_lora_value_res_mix}, 0); + layer.time_mix_v2 = create_tensor(tn(LLM_TENSOR_TIME_MIX_V2, "weight", i), {n_lora_value_res_mix, n_embd}, 0); + } + + layer.time_mix_g1 = create_tensor(tn(LLM_TENSOR_TIME_MIX_G1, "weight", i), {n_embd, n_lora_gate}, 0); + layer.time_mix_g2 = create_tensor(tn(LLM_TENSOR_TIME_MIX_G2, "weight", i), {n_lora_gate, n_embd}, 0); + + layer.time_mix_lerp_fused = create_tensor(tn(LLM_TENSOR_TIME_MIX_LERP_FUSED, "weight", i), {n_embd, 1, 1, 6}, 0); + + layer.time_mix_k_k = create_tensor(tn(LLM_TENSOR_TIME_MIX_K_K, "weight", i), {attn_hidden_size}, 0); + layer.time_mix_k_a = create_tensor(tn(LLM_TENSOR_TIME_MIX_K_A, "weight", i), {attn_hidden_size}, 0); + layer.time_mix_r_k = create_tensor(tn(LLM_TENSOR_TIME_MIX_R_K, "weight", i), {attn_hidden_size}, 0); + + layer.time_mix_key = create_tensor(tn(LLM_TENSOR_TIME_MIX_KEY, "weight", i), {attn_hidden_size, n_embd}, 0); + layer.time_mix_value = create_tensor(tn(LLM_TENSOR_TIME_MIX_VALUE, "weight", i), {attn_hidden_size, n_embd}, 0); + layer.time_mix_receptance = create_tensor(tn(LLM_TENSOR_TIME_MIX_RECEPTANCE, "weight", i), {attn_hidden_size, n_embd}, 0); + + layer.time_mix_ln = create_tensor(tn(LLM_TENSOR_TIME_MIX_LN, "weight", i), {n_embd}, 0); + layer.time_mix_ln_b = create_tensor(tn(LLM_TENSOR_TIME_MIX_LN, "bias", i), {n_embd}, 0); + layer.time_mix_output = create_tensor(tn(LLM_TENSOR_TIME_MIX_OUTPUT, "weight", i), {n_embd, attn_hidden_size}, 0); + + layer.channel_mix_lerp_k = create_tensor(tn(LLM_TENSOR_CHANNEL_MIX_LERP_K, "weight", i), {n_embd, 1, 1}, 0); + + layer.channel_mix_key = create_tensor(tn(LLM_TENSOR_CHANNEL_MIX_KEY, "weight", i), {n_embd, ffn_size}, 0); + layer.channel_mix_value = create_tensor(tn(LLM_TENSOR_CHANNEL_MIX_VALUE, "weight", i), {ffn_size, n_embd}, 0); + } + + } break; + case LLM_ARCH_ARWKV7: + { + tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0); + + // output + output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0); + output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, 0); + + const int n_lora_decay = hparams.n_lora_decay; + const int n_lora_iclr = hparams.n_lora_iclr; + const int n_lora_value_res_mix = hparams.n_lora_value_res_mix; + const int n_lora_gate = hparams.n_lora_gate; + const int attn_hidden_size = n_embd; + + for (int i = 0; i < n_layer; ++i) { + auto & layer = layers[i]; + + layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0); + + layer.time_mix_w0 = create_tensor(tn(LLM_TENSOR_TIME_MIX_W0, "weight", i), {n_embd}, 0); + layer.time_mix_w1 = create_tensor(tn(LLM_TENSOR_TIME_MIX_W1, "weight", i), {n_embd, n_lora_decay}, 0); + layer.time_mix_w2 = create_tensor(tn(LLM_TENSOR_TIME_MIX_W2, "weight", i), {n_lora_decay, n_embd}, 0); + + layer.time_mix_a0 = create_tensor(tn(LLM_TENSOR_TIME_MIX_A0, "weight", i), {n_embd}, 0); + layer.time_mix_a1 = create_tensor(tn(LLM_TENSOR_TIME_MIX_A1, "weight", i), {n_embd, n_lora_iclr}, 0); + layer.time_mix_a2 = create_tensor(tn(LLM_TENSOR_TIME_MIX_A2, "weight", i), {n_lora_iclr, n_embd}, 0); + + if (i == 0) { + // actually not used + layer.time_mix_v0 = create_tensor(tn(LLM_TENSOR_TIME_MIX_V0, "weight", i), {n_embd}, 0); + layer.time_mix_v1 = create_tensor(tn(LLM_TENSOR_TIME_MIX_V1, "weight", i), {n_embd, n_lora_iclr}, 0); + layer.time_mix_v2 = create_tensor(tn(LLM_TENSOR_TIME_MIX_V2, "weight", i), {n_lora_iclr, n_embd}, 0); + } else { + layer.time_mix_v0 = create_tensor(tn(LLM_TENSOR_TIME_MIX_V0, "weight", i), {n_embd}, 0); + layer.time_mix_v1 = create_tensor(tn(LLM_TENSOR_TIME_MIX_V1, "weight", i), {n_embd, n_lora_value_res_mix}, 0); + layer.time_mix_v2 = create_tensor(tn(LLM_TENSOR_TIME_MIX_V2, "weight", i), {n_lora_value_res_mix, n_embd}, 0); + } + + layer.time_mix_g1 = create_tensor(tn(LLM_TENSOR_TIME_MIX_G1, "weight", i), {n_embd, n_lora_gate}, TENSOR_NOT_REQUIRED); + layer.time_mix_g2 = create_tensor(tn(LLM_TENSOR_TIME_MIX_G2, "weight", i), {n_lora_gate, n_embd}, TENSOR_NOT_REQUIRED); + + try { + layer.time_mix_lerp_fused = create_tensor(tn(LLM_TENSOR_TIME_MIX_LERP_FUSED, "weight", i), {n_embd, 1, 1, 6}, 0); + } catch(std::runtime_error & e) { + // ARWKV models may not have gate tensors + layer.time_mix_lerp_fused = create_tensor(tn(LLM_TENSOR_TIME_MIX_LERP_FUSED, "weight", i), {n_embd, 1, 1, 5}, 0); + } + + layer.time_mix_k_k = create_tensor(tn(LLM_TENSOR_TIME_MIX_K_K, "weight", i), {attn_hidden_size}, 0); + layer.time_mix_k_a = create_tensor(tn(LLM_TENSOR_TIME_MIX_K_A, "weight", i), {attn_hidden_size}, 0); + layer.time_mix_r_k = create_tensor(tn(LLM_TENSOR_TIME_MIX_R_K, "weight", i), {attn_hidden_size}, 0); + + layer.time_mix_key = create_tensor(tn(LLM_TENSOR_TIME_MIX_KEY, "weight", i), {attn_hidden_size, n_embd}, 0); + layer.time_mix_value = create_tensor(tn(LLM_TENSOR_TIME_MIX_VALUE, "weight", i), {attn_hidden_size, n_embd}, 0); + layer.time_mix_receptance = create_tensor(tn(LLM_TENSOR_TIME_MIX_RECEPTANCE, "weight", i), {attn_hidden_size, n_embd}, 0); + + layer.time_mix_ln = create_tensor(tn(LLM_TENSOR_TIME_MIX_LN, "weight", i), {n_embd}, TENSOR_NOT_REQUIRED); + layer.time_mix_ln_b = create_tensor(tn(LLM_TENSOR_TIME_MIX_LN, "bias", i), {n_embd}, TENSOR_NOT_REQUIRED); + layer.time_mix_output = create_tensor(tn(LLM_TENSOR_TIME_MIX_OUTPUT, "weight", i), {n_embd, attn_hidden_size}, 0); + + layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0); + + layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, 0); + layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd}, 0); + layer.ffn_up = create_tensor(tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, 0); + } + + } break; case LLM_ARCH_CHAMELEON: { tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0); @@ -3416,6 +3982,46 @@ bool llama_model::load_tensors(llama_model_loader & ml) { output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), {hparams.convnext.n_embd, n_embd}, 0); output_b = create_tensor(tn(LLM_TENSOR_OUTPUT, "bias"), {n_embd}, 0); } break; + case LLM_ARCH_BAILINGMOE: + { + const int64_t n_ff_exp = hparams.n_ff_exp; + const int64_t n_expert_shared = hparams.n_expert_shared; + + tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0); + + // output + output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0); + output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, 0); + + for (int i = 0; i < n_layer; ++i) { + auto & layer = layers[i]; + + layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0); + + layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_head * n_rot}, 0); + layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_head_kv * n_rot}, 0); + layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_head_kv * n_rot}, 0); + layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_head * n_rot, n_embd}, 0); + layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0); + + layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}, 0); + + if (n_expert == 0) { + throw std::runtime_error("n_expert must be > 0"); + } + if (n_expert_used == 0) { + throw std::runtime_error("n_expert_used must be > 0"); + } + + layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), { n_embd, n_ff_exp, n_expert}, 0); + layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp, n_embd, n_expert}, 0); + layer.ffn_up_exps = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS, "weight", i), { n_embd, n_ff_exp, n_expert}, 0); + + layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), {n_embd, n_ff_exp * n_expert_shared}, 0); + layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), { n_ff_exp * n_expert_shared, n_embd}, 0); + layer.ffn_up_shexp = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP, "weight", i), {n_embd, n_ff_exp * n_expert_shared}, 0); + } + } break; default: throw std::runtime_error("unknown architecture"); } @@ -3573,8 +4179,8 @@ size_t llama_model::size() const { return pimpl->n_bytes; } -size_t llama_model::max_nodes() const { - return std::max(8192, tensors_by_name.size()*5); +size_t llama_model::n_tensors() const { + return tensors_by_name.size(); } size_t llama_model::n_devices() const { @@ -3629,6 +4235,7 @@ void llama_model::print_info() const { LLAMA_LOG_INFO("%s: n_head_kv = %s\n", __func__, print_f([&](uint32_t il) { return hparams.n_head_kv(il); }, hparams.n_layer).c_str()); LLAMA_LOG_INFO("%s: n_rot = %u\n", __func__, hparams.n_rot); LLAMA_LOG_INFO("%s: n_swa = %u\n", __func__, hparams.n_swa); + LLAMA_LOG_INFO("%s: n_swa_pattern = %u\n", __func__, hparams.n_swa_pattern); LLAMA_LOG_INFO("%s: n_embd_head_k = %u\n", __func__, hparams.n_embd_head_k); LLAMA_LOG_INFO("%s: n_embd_head_v = %u\n", __func__, hparams.n_embd_head_v); LLAMA_LOG_INFO("%s: n_gqa = %s\n", __func__, print_f([&](uint32_t il) { return hparams.n_gqa(il); }, hparams.n_layer).c_str()); @@ -3639,6 +4246,7 @@ void llama_model::print_info() const { LLAMA_LOG_INFO("%s: f_clamp_kqv = %.1e\n", __func__, hparams.f_clamp_kqv); LLAMA_LOG_INFO("%s: f_max_alibi_bias = %.1e\n", __func__, hparams.f_max_alibi_bias); LLAMA_LOG_INFO("%s: f_logit_scale = %.1e\n", __func__, hparams.f_logit_scale); + LLAMA_LOG_INFO("%s: f_attn_scale = %.1e\n", __func__, hparams.f_attention_scale); LLAMA_LOG_INFO("%s: n_ff = %s\n", __func__, print_f([&](uint32_t il) { return hparams.n_ff(il); }, hparams.n_layer).c_str()); LLAMA_LOG_INFO("%s: n_expert = %u\n", __func__, hparams.n_expert); LLAMA_LOG_INFO("%s: n_expert_used = %u\n", __func__, hparams.n_expert_used); @@ -3686,7 +4294,7 @@ void llama_model::print_info() const { LLAMA_LOG_INFO("%s: n_expert_shared = %d\n", __func__, hparams.n_expert_shared); LLAMA_LOG_INFO("%s: expert_weights_scale = %.1f\n", __func__, hparams.expert_weights_scale); LLAMA_LOG_INFO("%s: expert_weights_norm = %d\n", __func__, hparams.expert_weights_norm); - LLAMA_LOG_INFO("%s: expert_gating_func = %s\n", __func__, llama_expert_gating_func_name((enum llama_expert_gating_func_type) hparams.expert_gating_func)); + LLAMA_LOG_INFO("%s: expert_gating_func = %s\n", __func__, llama_expert_gating_func_name((llama_expert_gating_func_type) hparams.expert_gating_func)); LLAMA_LOG_INFO("%s: rope_yarn_log_mul = %.4f\n", __func__, hparams.rope_yarn_log_mul); } @@ -3695,12 +4303,24 @@ void llama_model::print_info() const { LLAMA_LOG_INFO("%s: n_ff_shexp = %d\n", __func__, hparams.n_ff_shexp); } + if (arch == LLM_ARCH_QWEN3MOE) { + LLAMA_LOG_INFO("%s: n_ff_exp = %d\n", __func__, hparams.n_ff_exp); + } + if (arch == LLM_ARCH_MINICPM || arch == LLM_ARCH_GRANITE || arch == LLM_ARCH_GRANITE_MOE) { LLAMA_LOG_INFO("%s: f_embedding_scale = %f\n", __func__, hparams.f_embedding_scale); LLAMA_LOG_INFO("%s: f_residual_scale = %f\n", __func__, hparams.f_residual_scale); LLAMA_LOG_INFO("%s: f_attention_scale = %f\n", __func__, hparams.f_attention_scale); } + if (arch == LLM_ARCH_BAILINGMOE) { + LLAMA_LOG_INFO("%s: n_layer_dense_lead = %d\n", __func__, hparams.n_layer_dense_lead); + LLAMA_LOG_INFO("%s: n_ff_exp = %d\n", __func__, hparams.n_ff_exp); + LLAMA_LOG_INFO("%s: n_expert_shared = %d\n", __func__, hparams.n_expert_shared); + LLAMA_LOG_INFO("%s: expert_weights_scale = %.1f\n", __func__, hparams.expert_weights_scale); + LLAMA_LOG_INFO("%s: expert_weights_norm = %d\n", __func__, hparams.expert_weights_norm); + } + vocab.print_info(); } @@ -3762,9 +4382,13 @@ ggml_backend_buffer_type_t llama_model::select_buft(int il) const { }); } -const struct ggml_tensor * llama_model::get_tensor(const char * name) const { +bool llama_model::has_tensor_overrides() const { + return pimpl->has_tensor_overrides; +} + +const ggml_tensor * llama_model::get_tensor(const char * name) const { auto it = std::find_if(tensors_by_name.begin(), tensors_by_name.end(), - [name](const std::pair & it) { + [name](const std::pair & it) { return it.first == name; }); if (it == tensors_by_name.end()) { @@ -3774,13 +4398,8631 @@ const struct ggml_tensor * llama_model::get_tensor(const char * name) const { return it->second; } +struct llm_build_llama : public llm_graph_context { + llm_build_llama(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + // temperature tuning + ggml_tensor * inp_attn_scale = nullptr; + if (arch == LLM_ARCH_LLAMA4) { + inp_attn_scale = build_inp_attn_scale(); + } + + auto * inp_attn = build_attn_inp_kv_unified(); + + const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale; + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + bool use_rope = arch == LLM_ARCH_LLAMA4 + ? (il + 1) % hparams.n_no_rope_layer_step != 0 + : true; + + // norm + cur = build_norm(inpL, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self-attention + { + // rope freq factors for llama3; may return nullptr for llama2 and other models + ggml_tensor * rope_factors = static_cast(memory)->cbs.get_rope_factors(n_ctx_per_seq, il); + + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + if (model.layers[il].bq) { + Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); + cb(Qcur, "Qcur", il); + } + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + if (model.layers[il].bk) { + Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); + cb(Kcur, "Kcur", il); + } + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + if (model.layers[il].bv) { + Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); + cb(Vcur, "Vcur", il); + } + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + if (use_rope) { + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, rope_factors, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, rope_factors, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + } else if (inp_attn_scale) { + Qcur = ggml_mul(ctx0, Qcur, inp_attn_scale); + } + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + if (arch == LLM_ARCH_LLAMA4 && use_rope && hparams.use_kq_norm) { + // Llama4TextL2Norm + Qcur = ggml_rms_norm(ctx0, Qcur, hparams.f_norm_rms_eps); + Kcur = ggml_rms_norm(ctx0, Kcur, hparams.f_norm_rms_eps); + cb(Qcur, "Qcur_normed", il); + cb(Kcur, "Kcur_normed", il); + } + + cur = build_attn(inp_attn, gf, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, nullptr, kq_scale, il); + cb(cur, "attn_out", il); + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + + // For Granite architecture + if (hparams.f_residual_scale) { + cur = ggml_scale(ctx0, cur, hparams.f_residual_scale); + } + + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network (non-MoE) + if (model.layers[il].ffn_gate_inp == nullptr) { + + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(cur, + model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL, + model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL, + model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + + } else if (arch == LLM_ARCH_LLAMA4) { + // llama4 MoE + ggml_tensor * ffn_inp_normed = build_norm(ffn_inp, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + ggml_tensor * moe_out = build_moe_ffn(ffn_inp_normed, + model.layers[il].ffn_gate_inp, + model.layers[il].ffn_up_exps, + model.layers[il].ffn_gate_exps, + model.layers[il].ffn_down_exps, + nullptr, + n_expert, n_expert_used, + LLM_FFN_SILU, false, + false, 0.0, + LLAMA_EXPERT_GATING_FUNC_TYPE_SIGMOID, + il); + + // Shared experts + ggml_tensor * shexp_out = build_ffn(ffn_inp_normed, + model.layers[il].ffn_up_shexp, NULL, NULL, + model.layers[il].ffn_gate_shexp, NULL, NULL, + model.layers[il].ffn_down_shexp, NULL, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(shexp_out, "ffn_moe_shexp", il); + + cur = ggml_add(ctx0, moe_out, shexp_out); + cb(cur, "ffn_moe_out_merged", il); + + } else { + // MoE branch + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + cur = build_moe_ffn(cur, + model.layers[il].ffn_gate_inp, + model.layers[il].ffn_up_exps, + model.layers[il].ffn_gate_exps, + model.layers[il].ffn_down_exps, + nullptr, + n_expert, n_expert_used, + LLM_FFN_SILU, true, + false, 0.0, + LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, + il); + cb(cur, "ffn_moe_out", il); + } + + // For Granite architecture + if (hparams.f_residual_scale) { + cur = ggml_scale(ctx0, cur, hparams.f_residual_scale); + } + + cur = ggml_add(ctx0, cur, ffn_inp); + cb(cur, "ffn_out", il); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = build_norm(cur, + model.output_norm, NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + // For Granite architecture + if (hparams.f_logit_scale) { + cur = ggml_scale(ctx0, cur, 1.0f / hparams.f_logit_scale); + } + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +struct llm_build_deci : public llm_graph_context { + llm_build_deci(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv_unified(); + + const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale; + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + const int64_t n_head_kv = hparams.n_head_kv(il); + const int64_t n_head = hparams.n_head(il); + + if (n_head == 0) { + // attention-free layer of Llama-3_1-Nemotron-51B + cur = inpL; + } else { + // norm + cur = build_norm(inpL, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + } + + if (n_head > 0 && n_head_kv == 0) { + // "linear attention" of Llama-3_1-Nemotron-51B + cur = build_lora_mm(model.layers[il].wo, cur); + cb(cur, "wo", il); + } else if (n_head > 0) { + // self-attention + // rope freq factors for llama3; may return nullptr for llama2 and other models + ggml_tensor * rope_factors = static_cast(memory)->cbs.get_rope_factors(n_ctx_per_seq, il); + + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + if (model.layers[il].bq) { + Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); + cb(Qcur, "Qcur", il); + } + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + if (model.layers[il].bk) { + Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); + cb(Kcur, "Kcur", il); + } + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + if (model.layers[il].bv) { + Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); + cb(Vcur, "Vcur", il); + } + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, rope_factors, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, rope_factors, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, gf, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, nullptr, kq_scale, il); + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + + // For Granite architecture + if (hparams.f_residual_scale) { + cur = ggml_scale(ctx0, cur, hparams.f_residual_scale); + } + + // modified to support attention-free layer of Llama-3_1-Nemotron-51B + ggml_tensor * ffn_inp = cur; + if (n_head > 0) { + ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + } + + // feed-forward network + if (model.layers[il].ffn_gate_inp == nullptr) { + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(cur, + model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL, + model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL, + model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + } + + // For Granite architecture + if (hparams.f_residual_scale) { + cur = ggml_scale(ctx0, cur, hparams.f_residual_scale); + } + + cur = ggml_add(ctx0, cur, ffn_inp); + cb(cur, "ffn_out", il); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = build_norm(cur, + model.output_norm, NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + // For Granite architecture + if (hparams.f_logit_scale) { + cur = ggml_scale(ctx0, cur, 1.0f / hparams.f_logit_scale); + } + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +struct llm_build_baichuan : public llm_graph_context { + llm_build_baichuan(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = model.type == LLM_TYPE_7B ? build_inp_pos() : nullptr; + + auto * inp_attn = build_attn_inp_kv_unified(); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + cur = build_norm(inpL, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self-attention + { + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + switch (model.type) { + case LLM_TYPE_7B: + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + break; + case LLM_TYPE_13B: + break; + default: + GGML_ABORT("fatal error"); + } + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, gf, + model.layers[il].wo, NULL, + Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + { + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + } + + cur = ggml_add(ctx0, cur, ffn_inp); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = build_norm(cur, + model.output_norm, NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +struct llm_build_xverse : public llm_graph_context { + llm_build_xverse(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv_unified(); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + cur = build_norm(inpL, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self-attention + { + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, gf, + model.layers[il].wo, NULL, + Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + { + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + } + + cur = ggml_add(ctx0, cur, ffn_inp); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +struct llm_build_falcon : public llm_graph_context { + llm_build_falcon(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + const int64_t n_embd_gqa = hparams.n_embd_v_gqa(); + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv_unified(); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * attn_norm; + + attn_norm = build_norm(inpL, + model.layers[il].attn_norm, + model.layers[il].attn_norm_b, + LLM_NORM, il); + cb(attn_norm, "attn_norm", il); + + // self-attention + { + if (model.layers[il].attn_norm_2) { + // Falcon-40B + cur = build_norm(inpL, + model.layers[il].attn_norm_2, + model.layers[il].attn_norm_2_b, + LLM_NORM, il); + cb(cur, "attn_norm_2", il); + } else { + cur = attn_norm; + } + + cur = build_lora_mm(model.layers[il].wqkv, cur); + cb(cur, "wqkv", il); + + ggml_tensor * Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd))); + ggml_tensor * Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd))); + ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa))); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + // using mode = 2 for neox mode + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, gf, + model.layers[il].wo, NULL, + Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); + attn_norm = ggml_get_rows(ctx0, attn_norm, inp_out_ids); + } + + ggml_tensor * ffn_inp = cur; + + // feed forward + { + cur = build_ffn(attn_norm, // !! use the attn norm, not the result + model.layers[il].ffn_up, NULL, NULL, + NULL, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + LLM_FFN_GELU, LLM_FFN_SEQ, il); + cb(cur, "ffn_out", il); + } + + cur = ggml_add(ctx0, cur, ffn_inp); + cur = ggml_add(ctx0, cur, inpL); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + // norm + cur = build_norm(cur, + model.output_norm, + model.output_norm_b, + LLM_NORM, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +struct llm_build_grok : public llm_graph_context { + llm_build_grok(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // multiply by embedding_multiplier_scale of 78.38367176906169 + inpL = ggml_scale(ctx0, inpL, 78.38367176906169f); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv_unified(); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + // norm + cur = build_norm(inpL, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + + // self-attention + { + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + if (model.layers[il].bq) { + Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); + cb(Qcur, "Qcur", il); + } + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + if (model.layers[il].bk) { + Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); + cb(Kcur, "Kcur", il); + } + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + if (model.layers[il].bv) { + Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); + cb(Vcur, "Vcur", il); + } + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, gf, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, nullptr, 1.0f, il); + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + + // Grok + // if attn_out_norm is present then apply it before adding the input + if (model.layers[il].attn_out_norm) { + cur = build_norm(cur, + model.layers[il].attn_out_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_out_norm", il); + } + + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + // MoE branch + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + cur = build_moe_ffn(cur, + model.layers[il].ffn_gate_inp, + model.layers[il].ffn_up_exps, + model.layers[il].ffn_gate_exps, + model.layers[il].ffn_down_exps, + nullptr, + n_expert, n_expert_used, + LLM_FFN_GELU, true, + false, 0.0, + LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, + il); + cb(cur, "ffn_moe_out", il); + + // Grok + // if layer_out_norm is present then apply it before adding the input + // Idea: maybe ffn_out_norm is a better name + if (model.layers[il].layer_out_norm) { + cur = build_norm(cur, + model.layers[il].layer_out_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "layer_out_norm", il); + } + + cur = ggml_add(ctx0, cur, ffn_inp); + cb(cur, "ffn_out", il); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = build_norm(cur, + model.output_norm, NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + // Grok + // multiply logits by output_multiplier_scale of 0.5773502691896257 + + cur = ggml_scale(ctx0, cur, 0.5773502691896257f); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +struct llm_build_dbrx : public llm_graph_context { + llm_build_dbrx(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + const int64_t n_embd_gqa = hparams.n_embd_v_gqa(); + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv_unified(); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + // norm + cur = build_norm(inpL, + model.layers[il].attn_norm, NULL, + LLM_NORM, il); + cb(cur, "attn_norm", il); + + // self-attention + { + ggml_tensor * Qcur = nullptr; + ggml_tensor * Kcur = nullptr; + ggml_tensor * Vcur = nullptr; + + cur = build_lora_mm(model.layers[il].wqkv, cur); + cb(cur, "wqkv", il); + + cur = ggml_clamp(ctx0, cur, -hparams.f_clamp_kqv, hparams.f_clamp_kqv); + cb(cur, "wqkv_clamped", il); + + Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd))); + Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd))); + Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa))); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, gf, + model.layers[il].wo, NULL, + Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + // MoE branch + cur = build_norm(ffn_inp, + model.layers[il].attn_out_norm, NULL, + LLM_NORM, il); + cb(cur, "attn_out_norm", il); + + cur = build_moe_ffn(cur, + model.layers[il].ffn_gate_inp, + model.layers[il].ffn_up_exps, + model.layers[il].ffn_gate_exps, + model.layers[il].ffn_down_exps, + nullptr, + n_expert, n_expert_used, + LLM_FFN_SILU, true, + false, 0.0, + LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, + il); + cb(cur, "ffn_moe_out", il); + + cur = ggml_add(ctx0, cur, ffn_inp); + cb(cur, "ffn_out", il); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = build_norm(cur, + model.output_norm, NULL, + LLM_NORM, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +struct llm_build_starcoder : public llm_graph_context { + llm_build_starcoder(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + const int64_t n_embd_gqa = hparams.n_embd_v_gqa(); + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv_unified(); + + ggml_tensor * pos = ggml_get_rows(ctx0, model.pos_embd, inp_pos); + cb(pos, "pos_embd", -1); + + inpL = ggml_add(ctx0, inpL, pos); + cb(inpL, "inpL", -1); + + for (int il = 0; il < n_layer; ++il) { + cur = build_norm(inpL, + model.layers[il].attn_norm, + model.layers[il].attn_norm_b, + LLM_NORM, il); + cb(cur, "attn_norm", il); + + // self-attention + { + cur = build_lora_mm(model.layers[il].wqkv, cur); + cb(cur, "wqkv", il); + + cur = ggml_add(ctx0, cur, model.layers[il].bqkv); + cb(cur, "bqkv", il); + + ggml_tensor * Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd))); + ggml_tensor * Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd))); + ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa))); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, gf, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); + } + + // add the input + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL); + cb(ffn_inp, "ffn_inp", il); + + // FF + { + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, + model.layers[il].ffn_norm_b, + LLM_NORM, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(cur, + model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL, + NULL, NULL, NULL, + model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, + NULL, + LLM_FFN_GELU, LLM_FFN_SEQ, il); + cb(cur, "ffn_out", il); + } + + cur = ggml_add(ctx0, cur, ffn_inp); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = build_norm(inpL, + model.output_norm, + model.output_norm_b, + LLM_NORM, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +struct llm_build_refact : public llm_graph_context { + llm_build_refact(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + auto * inp_attn = build_attn_inp_kv_unified(); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + cur = build_norm(inpL, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self-attention + { + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, gf, + model.layers[il].wo, NULL, + Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + { + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + } + + cur = ggml_add(ctx0, cur, ffn_inp); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = build_norm(cur, + model.output_norm, NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +struct llm_build_bert : public llm_graph_context { + llm_build_bert(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + const int64_t n_embd_gqa = hparams.n_embd_v_gqa(); + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + + ggml_tensor * cur; + ggml_tensor * inpL; + ggml_tensor * inp_pos = nullptr; + + if (model.arch != LLM_ARCH_JINA_BERT_V2) { + inp_pos = build_inp_pos(); + } + + // construct input embeddings (token, type, position) + inpL = build_inp_embd(model.tok_embd); + + // token types are hardcoded to zero ("Sentence A") + ggml_tensor * type_row0 = ggml_view_1d(ctx0, model.type_embd, n_embd, 0); + inpL = ggml_add(ctx0, inpL, type_row0); + if (model.arch == LLM_ARCH_BERT) { + inpL = ggml_add(ctx0, ggml_get_rows(ctx0, model.pos_embd, inp_pos), inpL); + } + cb(inpL, "inp_embd", -1); + + // embed layer norm + inpL = build_norm(inpL, model.tok_norm, model.tok_norm_b, LLM_NORM, -1); + cb(inpL, "inp_norm", -1); + + auto * inp_attn = build_attn_inp_no_cache(); + + // iterate layers + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * cur = inpL; + + ggml_tensor * Qcur; + ggml_tensor * Kcur; + ggml_tensor * Vcur; + + // self-attention + if (model.arch == LLM_ARCH_BERT || model.arch == LLM_ARCH_JINA_BERT_V2) { + Qcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wq, cur), model.layers[il].bq); + + if (model.layers[il].attn_q_norm) { + Qcur = build_norm(Qcur, + model.layers[il].attn_q_norm, + model.layers[il].attn_q_norm_b, + LLM_NORM, il); + } + + Kcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wk, cur), model.layers[il].bk); + + if (model.layers[il].attn_k_norm) { + Kcur = build_norm(Kcur, + model.layers[il].attn_k_norm, + model.layers[il].attn_k_norm_b, + LLM_NORM, il); + } + + Vcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wv, cur), model.layers[il].bv); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + } else { + // compute Q and K and RoPE them + cur = build_lora_mm(model.layers[il].wqkv, cur); + cb(cur, "wqkv", il); + + Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd))); + Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd))); + Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa))); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + } + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, gf, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + cb(cur, "kqv_out", il); + + if (il == n_layer - 1 && pooling_type == LLAMA_POOLING_TYPE_NONE) { + // skip computing output for unused tokens + ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); + } + + // re-add the layer input + cur = ggml_add(ctx0, cur, inpL); + + // attention layer norm + cur = build_norm(cur, model.layers[il].attn_out_norm, model.layers[il].attn_out_norm_b, LLM_NORM, il); + + if (model.layers[il].attn_norm_2 != nullptr) { + cur = ggml_add(ctx0, cur, inpL); // re-add the layer input + cur = build_norm(cur, model.layers[il].attn_norm_2, model.layers[il].attn_norm_2_b, LLM_NORM, il); + } + + ggml_tensor * ffn_inp = cur; + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + if (model.arch == LLM_ARCH_BERT) { + cur = build_ffn(cur, + model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL, + NULL, NULL, NULL, + model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, + NULL, + LLM_FFN_GELU, LLM_FFN_SEQ, il); + } else if (model.arch == LLM_ARCH_JINA_BERT_V2) { + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, + NULL, + LLM_FFN_GELU, LLM_FFN_PAR, il); + } else { + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + } + cb(cur, "ffn_out", il); + + // attentions bypass the intermediate layer + cur = ggml_add(ctx0, cur, ffn_inp); + + // output layer norm + cur = build_norm(cur, model.layers[il].layer_out_norm, model.layers[il].layer_out_norm_b, LLM_NORM, il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cb(cur, "result_embd", -1); + res->t_embd = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +struct llm_build_bloom : public llm_graph_context { + llm_build_bloom(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + const int64_t n_embd_gqa = hparams.n_embd_v_gqa(); + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + auto * inp_attn = build_attn_inp_kv_unified(); + + inpL = build_norm(inpL, + model.tok_norm, + model.tok_norm_b, + LLM_NORM, -1); + cb(inpL, "inp_norm", -1); + + for (int il = 0; il < n_layer; ++il) { + cur = build_norm(inpL, + model.layers[il].attn_norm, + model.layers[il].attn_norm_b, + LLM_NORM, il); + cb(cur, "attn_norm", il); + + // self-attention + { + cur = build_lora_mm(model.layers[il].wqkv, cur); + cb(cur, "wqkv", il); + + cur = ggml_add(ctx0, cur, model.layers[il].bqkv); + cb(cur, "bqkv", il); + + ggml_tensor * Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd))); + ggml_tensor * Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd))); + ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa))); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, gf, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); + } + + // Add the input + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL); + cb(ffn_inp, "ffn_inp", il); + + // FF + { + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, + model.layers[il].ffn_norm_b, + LLM_NORM, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(cur, + model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL, + NULL, NULL, NULL, + model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, + NULL, + LLM_FFN_GELU, LLM_FFN_SEQ, il); + cb(cur, "ffn_out", il); + } + + cur = ggml_add(ctx0, cur, ffn_inp); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = build_norm(inpL, + model.output_norm, + model.output_norm_b, + LLM_NORM, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +struct llm_build_mpt : public llm_graph_context { + llm_build_mpt(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + const int64_t n_embd_gqa = hparams.n_embd_v_gqa(); + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + + ggml_tensor * cur; + ggml_tensor * pos; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + auto * inp_attn = build_attn_inp_kv_unified(); + + if (model.pos_embd) { + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + pos = ggml_get_rows(ctx0, model.pos_embd, inp_pos); + cb(pos, "pos_embd", -1); + + inpL = ggml_add(ctx0, inpL, pos); + cb(inpL, "inpL", -1); + } + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * attn_norm; + + attn_norm = build_norm(inpL, + model.layers[il].attn_norm, + model.layers[il].attn_norm_b, + LLM_NORM, il); + cb(attn_norm, "attn_norm", il); + + // self-attention + { + cur = attn_norm; + + cur = build_lora_mm(model.layers[il].wqkv, cur); + cb(cur, "wqkv", il); + + if (model.layers[il].bqkv){ + cur = ggml_add(ctx0, cur, model.layers[il].bqkv); + cb(cur, "bqkv", il); + } + + if (hparams.f_clamp_kqv > 0.0f) { + cur = ggml_clamp(ctx0, cur, -hparams.f_clamp_kqv, hparams.f_clamp_kqv); + cb(cur, "wqkv_clamped", il); + } + + ggml_tensor * Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd))); + ggml_tensor * Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd))); + ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa))); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + // Q/K Layernorm + if (model.layers[il].attn_q_norm) { + Qcur = build_norm(Qcur, + model.layers[il].attn_q_norm, + model.layers[il].attn_q_norm_b, + LLM_NORM, il); + cb(Qcur, "Qcur", il); + + Kcur = build_norm(Kcur, + model.layers[il].attn_k_norm, + model.layers[il].attn_k_norm_b, + LLM_NORM, il); + cb(Kcur, "Kcur", il); + } + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, gf, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); + } + + // Add the input + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL); + cb(ffn_inp, "ffn_inp", il); + + // feed forward + { + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, + model.layers[il].ffn_norm_b, + LLM_NORM, il); + cb(cur, "ffn_norm", il); + cur = build_ffn(cur, + model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL, + NULL, NULL, NULL, + model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, + model.layers[il].ffn_act, + LLM_FFN_GELU, LLM_FFN_SEQ, il); + cb(cur, "ffn_out", il); + } + + cur = ggml_add(ctx0, cur, ffn_inp); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = build_norm(cur, + model.output_norm, + model.output_norm_b, + LLM_NORM, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +struct llm_build_stablelm : public llm_graph_context { + llm_build_stablelm(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv_unified(); + + for (int il = 0; il < n_layer; ++il) { + // norm + cur = build_norm(inpL, + model.layers[il].attn_norm, + model.layers[il].attn_norm_b, + LLM_NORM, il); + cb(cur, "attn_norm", il); + + ggml_tensor * inpSA = cur; + + // self-attention + { + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + if (model.layers[il].bq) { + Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); + cb(Qcur, "Qcur", il); + } + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + if (model.layers[il].bk) { + Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); + cb(Kcur, "Kcur", il); + } + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + if (model.layers[il].bv) { + Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); + cb(Vcur, "Vcur", il); + } + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + if (model.layers[il].attn_q_norm) { + Qcur = build_norm(Qcur, + model.layers[il].attn_q_norm, + NULL, + LLM_NORM, il); + cb(Qcur, "Qcur", il); + } + + if (model.layers[il].attn_k_norm) { + Kcur = build_norm(Kcur, + model.layers[il].attn_k_norm, + NULL, + LLM_NORM, il); + cb(Kcur, "Kcur", il); + } + + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, gf, + model.layers[il].wo, NULL, + Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + { + if (model.layers[il].ffn_norm) { + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, + model.layers[il].ffn_norm_b, + LLM_NORM, il); + cb(cur, "ffn_norm", il); + } else { + // parallel residual + cur = inpSA; + } + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + } + + cur = ggml_add(ctx0, cur, ffn_inp); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = build_norm(cur, + model.output_norm, + model.output_norm_b, + LLM_NORM, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +struct llm_build_qwen : public llm_graph_context { + llm_build_qwen(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv_unified(); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + cur = build_norm(inpL, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self-attention + { + cur = build_lora_mm(model.layers[il].wqkv, cur); + cb(cur, "wqkv", il); + + cur = ggml_add(ctx0, cur, model.layers[il].bqkv); + cb(cur, "bqkv", il); + + ggml_tensor * Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd))); + ggml_tensor * Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd))); + ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 2*sizeof(float)*(n_embd))); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + // using mode = 2 for neox mode + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, gf, + model.layers[il].wo, NULL, + Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward forward + { + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + } + + cur = ggml_add(ctx0, cur, ffn_inp); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = build_norm(cur, + model.output_norm, NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +struct llm_build_qwen2 : public llm_graph_context { + llm_build_qwen2(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv_unified(); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + // norm + cur = build_norm(inpL, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self-attention + { + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); + cb(Qcur, "Qcur", il); + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); + cb(Kcur, "Kcur", il); + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); + cb(Vcur, "Vcur", il); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, gf, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + + cur = ggml_add(ctx0, cur, ffn_inp); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = build_norm(cur, + model.output_norm, NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +struct llm_build_qwen2vl : public llm_graph_context { + llm_build_qwen2vl(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv_unified(); + + int sections[4]; + std::copy(std::begin(hparams.rope_sections), std::begin(hparams.rope_sections) + 4, sections); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + // norm + cur = build_norm(inpL, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self-attention + { + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); + cb(Qcur, "Qcur", il); + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); + cb(Kcur, "Kcur", il); + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); + cb(Vcur, "Vcur", il); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + Qcur = ggml_rope_multi( + ctx0, Qcur, inp_pos, nullptr, + n_rot, sections, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + Kcur = ggml_rope_multi( + ctx0, Kcur, inp_pos, nullptr, + n_rot, sections, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, gf, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + + cur = ggml_add(ctx0, cur, ffn_inp); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = build_norm(cur, + model.output_norm, NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +struct llm_build_qwen2moe : public llm_graph_context { + llm_build_qwen2moe(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv_unified(); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + // norm + cur = build_norm(inpL, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self_attention + { + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + if (model.layers[il].bq) { + Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); + cb(Qcur, "Qcur", il); + } + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + if (model.layers[il].bk) { + Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); + cb(Kcur, "Kcur", il); + } + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + if (model.layers[il].bv) { + Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); + cb(Vcur, "Vcur", il); + } + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, gf, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // MoE branch + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + ggml_tensor * moe_out = + build_moe_ffn(cur, + model.layers[il].ffn_gate_inp, + model.layers[il].ffn_up_exps, + model.layers[il].ffn_gate_exps, + model.layers[il].ffn_down_exps, + nullptr, + n_expert, n_expert_used, + LLM_FFN_SILU, false, + false, 0.0, + LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, + il); + cb(moe_out, "ffn_moe_out", il); + + // FFN shared expert + { + ggml_tensor * cur_gate_inp = build_lora_mm(model.layers[il].ffn_gate_inp_shexp, cur); + cb(cur_gate_inp, "ffn_shexp_gate_inp", il); + + // sigmoid + ggml_tensor * cur_gate = ggml_div(ctx0, ggml_silu(ctx0, cur_gate_inp), cur_gate_inp); + cb(cur_gate, "ffn_shexp_gate", il); + + ggml_tensor * cur_ffn = build_ffn(cur, + model.layers[il].ffn_up_shexp, NULL, NULL, + model.layers[il].ffn_gate_shexp, NULL, NULL, + model.layers[il].ffn_down_shexp, NULL, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur_ffn, "ffn_shexp", il); + + ggml_tensor * ffn_shexp_out = ggml_mul(ctx0, cur_ffn, cur_gate); + cb(ffn_shexp_out, "ffn_shexp_out", il); + + moe_out = ggml_add(ctx0, moe_out, ffn_shexp_out); + cb(moe_out, "ffn_out", il); + + cur = moe_out; + } + + cur = ggml_add(ctx0, cur, ffn_inp); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = build_norm(cur, + model.output_norm, NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +struct llm_build_qwen3 : public llm_graph_context { + llm_build_qwen3(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv_unified(); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + // norm + cur = build_norm(inpL, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self-attention + { + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il); + cb(Qcur, "Qcur_normed", il); + + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il); + cb(Kcur, "Kcur_normed", il); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, gf, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + + cur = ggml_add(ctx0, cur, ffn_inp); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = build_norm(cur, + model.output_norm, NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +struct llm_build_qwen3moe : public llm_graph_context { + llm_build_qwen3moe(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv_unified(); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + // norm + cur = build_norm(inpL, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self_attention + { + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il); + cb(Qcur, "Qcur_normed", il); + + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il); + cb(Kcur, "Kcur_normed", il); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, gf, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // MoE branch + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + ggml_tensor * moe_out = + build_moe_ffn(cur, + model.layers[il].ffn_gate_inp, + model.layers[il].ffn_up_exps, + model.layers[il].ffn_gate_exps, + model.layers[il].ffn_down_exps, + nullptr, + n_expert, n_expert_used, + LLM_FFN_SILU, true, + false, 0.0, + LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, + il); + cb(moe_out, "ffn_moe_out", il); + cur = moe_out; + + cur = ggml_add(ctx0, cur, ffn_inp); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = build_norm(cur, + model.output_norm, NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +struct llm_build_phi2 : public llm_graph_context { + llm_build_phi2(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + const int64_t n_embd_gqa = hparams.n_embd_v_gqa(); + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + + ggml_tensor * cur; + ggml_tensor * attn_norm_output; + ggml_tensor * ffn_output; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv_unified(); + + for (int il = 0; il < n_layer; ++il) { + attn_norm_output = build_norm(inpL, + model.layers[il].attn_norm, + model.layers[il].attn_norm_b, + LLM_NORM, il); + cb(attn_norm_output, "attn_norm", il); + + // self-attention + { + ggml_tensor * Qcur = nullptr; + ggml_tensor * Kcur = nullptr; + ggml_tensor * Vcur = nullptr; + + if (model.layers[il].wqkv) { + cur = build_lora_mm(model.layers[il].wqkv, attn_norm_output); + cb(cur, "wqkv", il); + + cur = ggml_add(ctx0, cur, model.layers[il].bqkv); + cb(cur, "bqkv", il); + + Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd))); + Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd))); + Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa))); + } else { + Qcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wq, attn_norm_output), model.layers[il].bq); + Kcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wk, attn_norm_output), model.layers[il].bk); + Vcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wv, attn_norm_output), model.layers[il].bv); + } + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + // with phi2, we scale the Q to avoid precision issues + // ref: https://github.com/ml-explore/mlx-examples/blob/08e862336ade809bc37d1035f94b359e7d1a5152/phi2/phi2.py#L64-L66 + Qcur = ggml_scale(ctx0, Qcur, 1.0f/sqrtf(float(n_embd_head))); + + cur = build_attn(inp_attn, gf, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, nullptr, 1.0f, il); + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); + attn_norm_output = ggml_get_rows(ctx0, attn_norm_output, inp_out_ids); + } + + // FF + { + ffn_output = build_ffn(attn_norm_output, + model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL, + NULL, NULL, NULL, + model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, + NULL, + LLM_FFN_GELU, LLM_FFN_SEQ, il); + cb(ffn_output, "ffn_out", il); + } + + cur = ggml_add(ctx0, cur, ffn_output); + cur = ggml_add(ctx0, cur, inpL); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = build_norm(inpL, + model.output_norm, + model.output_norm_b, + LLM_NORM, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + cur = build_lora_mm(model.output, cur); + cb(cur, "result_output_no_bias", -1); + + cur = ggml_add(ctx0, cur, model.output_b); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +struct llm_build_phi3 : public llm_graph_context { + llm_build_phi3(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + const int64_t n_embd_gqa = hparams.n_embd_v_gqa(); + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv_unified(); + + for (int il = 0; il < n_layer; ++il) { + auto * residual = inpL; + + // self-attention + { + // rope freq factors for 128k context + ggml_tensor * rope_factors = static_cast(memory)->cbs.get_rope_factors(n_ctx_per_seq, il); + + ggml_tensor* attn_norm_output = build_norm(inpL, + model.layers[il].attn_norm, + model.layers[il].attn_norm_b, + LLM_NORM_RMS, il); + cb(attn_norm_output, "attn_norm", il); + + ggml_tensor * Qcur = nullptr; + ggml_tensor * Kcur = nullptr; + ggml_tensor * Vcur = nullptr; + + if (model.layers[il].wqkv) { + cur = build_lora_mm(model.layers[il].wqkv, attn_norm_output); + cb(cur, "wqkv", il); + + Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0 * sizeof(float) * (n_embd))); + Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1 * sizeof(float) * (n_embd))); + Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1 * sizeof(float) * (n_embd + n_embd_gqa))); + } else { + Qcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wq, attn_norm_output), model.layers[il].bq); + Kcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wk, attn_norm_output), model.layers[il].bk); + Vcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wv, attn_norm_output), model.layers[il].bv); + } + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, rope_factors, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, rope_factors, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + Qcur = ggml_scale(ctx0, Qcur, 1.0f / sqrtf(float(n_embd_head))); + cb(Qcur, "Qcur", il); + + cur = build_attn(inp_attn, gf, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, nullptr, 1.0f, il); + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + ggml_tensor* inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + residual = ggml_get_rows(ctx0, residual, inp_out_ids); + } + + cur = ggml_add(ctx0, cur, residual); + residual = cur; + + cur = build_norm(cur, + model.layers[il].ffn_norm, model.layers[il].ffn_norm_b, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + // feed-forward network + if (model.layers[il].ffn_gate_inp == nullptr) { + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + NULL, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + LLM_FFN_SWIGLU, LLM_FFN_SEQ, il); + cb(cur, "ffn_out", il); + } else { + // MoE branch + cur = build_moe_ffn(cur, + model.layers[il].ffn_gate_inp, + model.layers[il].ffn_up_exps, + model.layers[il].ffn_gate_exps, + model.layers[il].ffn_down_exps, + nullptr, + n_expert, n_expert_used, + LLM_FFN_SILU, true, + false, 0.0, + LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, + il); + cb(cur, "ffn_moe_out", il); + } + + cur = ggml_add(ctx0, residual, cur); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = build_norm(inpL, + model.output_norm, + model.output_norm_b, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + cur = build_lora_mm(model.output, cur); + + if (model.output_b != nullptr) { + cb(cur, "result_output_no_bias", -1); + cur = ggml_add(ctx0, cur, model.output_b); + } + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +struct llm_build_plamo : public llm_graph_context { + llm_build_plamo(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv_unified(); + + for (int il = 0; il < n_layer; ++il) { + + // norm + cur = build_norm(inpL, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + ggml_tensor * attention_norm = cur; + + // self-attention + { + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, nullptr, + n_embd_head, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, nullptr, + n_embd_head, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, gf, + model.layers[il].wo, NULL, + Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + } + ggml_tensor * sa_out = cur; + + cur = attention_norm; + + if (il == n_layer - 1) { + // skip computing output for unused tokens + ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + sa_out = ggml_get_rows(ctx0, sa_out, inp_out_ids); + inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); + } + + // feed-forward network + { + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + } + + cur = ggml_add(ctx0, cur, sa_out); + cur = ggml_add(ctx0, cur, inpL); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = build_norm(cur, + model.output_norm, NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +struct llm_build_gpt2 : public llm_graph_context { + llm_build_gpt2(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + const int64_t n_embd_gqa = hparams.n_embd_v_gqa(); + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + + ggml_tensor * cur; + ggml_tensor * pos; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv_unified(); + + pos = ggml_get_rows(ctx0, model.pos_embd, inp_pos); + cb(pos, "pos_embd", -1); + + inpL = ggml_add(ctx0, inpL, pos); + cb(inpL, "inpL", -1); + + for (int il = 0; il < n_layer; ++il) { + cur = build_norm(inpL, + model.layers[il].attn_norm, + model.layers[il].attn_norm_b, + LLM_NORM, il); + cb(cur, "attn_norm", il); + + // self-attention + { + cur = build_lora_mm(model.layers[il].wqkv, cur); + cb(cur, "wqkv", il); + + cur = ggml_add(ctx0, cur, model.layers[il].bqkv); + cb(cur, "bqkv", il); + + ggml_tensor * Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd))); + ggml_tensor * Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd))); + ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa))); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + cur = build_attn(inp_attn, gf, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); + } + + // add the input + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL); + cb(ffn_inp, "ffn_inp", il); + + // FF + { + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, + model.layers[il].ffn_norm_b, + LLM_NORM, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(cur, + model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL, + NULL, NULL, NULL, + model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, + NULL, + LLM_FFN_GELU, LLM_FFN_SEQ, il); + cb(cur, "ffn_out", il); + } + + cur = ggml_add(ctx0, cur, ffn_inp); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = build_norm(inpL, + model.output_norm, + model.output_norm_b, + LLM_NORM, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +struct llm_build_codeshell : public llm_graph_context { + llm_build_codeshell(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + const int64_t n_embd_gqa = hparams.n_embd_v_gqa(); + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv_unified(); + + for (int il = 0; il < n_layer; ++il) { + cur = build_norm(inpL, + model.layers[il].attn_norm, + model.layers[il].attn_norm_b, + LLM_NORM, il); + cb(cur, "attn_norm", il); + + // self-attention + { + cur = build_lora_mm(model.layers[il].wqkv, cur); + cb(cur, "wqkv", il); + + cur = ggml_add(ctx0, cur, model.layers[il].bqkv); + cb(cur, "bqkv", il); + + ggml_tensor * Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd))); + ggml_tensor * Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd))); + ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa))); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, gf, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); + } + + // add the input + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL); + cb(ffn_inp, "ffn_inp", il); + + // FF + { + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, + model.layers[il].ffn_norm_b, + LLM_NORM, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(cur, + model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL, + NULL, NULL, NULL, + model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, + NULL, + LLM_FFN_GELU, LLM_FFN_SEQ, il); + cb(cur, "ffn_out", il); + } + + cur = ggml_add(ctx0, cur, ffn_inp); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = build_norm(inpL, + model.output_norm, + model.output_norm_b, + LLM_NORM, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +struct llm_build_orion : public llm_graph_context { + llm_build_orion(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv_unified(); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + // norm + cur = build_norm(inpL, + model.layers[il].attn_norm, model.layers[il].attn_norm_b, + LLM_NORM, il); + cb(cur, "attn_norm", il); + + // self-attention + { + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + // if (model.layers[il].bq) { + // Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); + // cb(Qcur, "Qcur", il); + // } + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + // if (model.layers[il].bk) { + // Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); + // cb(Kcur, "Kcur", il); + // } + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + // if (model.layers[il].bv) { + // Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); + // cb(Vcur, "Vcur", il); + // } + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, gf, + model.layers[il].wo, NULL, + Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, model.layers[il].ffn_norm_b, + LLM_NORM, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + + cur = ggml_add(ctx0, cur, ffn_inp); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = build_norm(cur, + model.output_norm, model.output_norm_b, + LLM_NORM, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +struct llm_build_internlm2 : public llm_graph_context { + llm_build_internlm2(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv_unified(); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + // norm + cur = build_norm(inpL, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self-attention + { + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + if (model.layers[il].bq) { + Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); + cb(Qcur, "Qcur", il); + } + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + if (model.layers[il].bk) { + Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); + cb(Kcur, "Kcur", il); + } + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + if (model.layers[il].bv) { + Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); + cb(Vcur, "Vcur", il); + } + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, gf, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + + cur = ggml_add(ctx0, cur, ffn_inp); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = build_norm(cur, + model.output_norm, NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +struct llm_build_minicpm3 : public llm_graph_context { + llm_build_minicpm3(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { + //TODO: if the model varies, these parameters need to be read from the model + const int64_t n_embd_base = 256; + const float scale_embd = 12.0f; + const float scale_depth = 1.4f; + const float kq_scale = 1.0f / sqrtf(float(hparams.n_embd_head_k)); + + const uint32_t n_embd_head_qk_rope = hparams.n_rot; + const uint32_t n_embd_head_qk_nope = hparams.n_embd_head_k - hparams.n_rot; + const uint32_t kv_lora_rank = hparams.n_lora_kv; + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // scale the input embeddings + inpL = ggml_scale(ctx0, inpL, scale_embd); + cb(inpL, "inp_scaled", -1); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv_unified(); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + ggml_tensor * rope_factors = static_cast(memory)->cbs.get_rope_factors(n_ctx_per_seq, il); + + // norm + cur = build_norm(inpL, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self_attention + { + ggml_tensor * q = NULL; + // {n_embd, q_lora_rank} * {n_embd, n_tokens} -> {q_lora_rank, n_tokens} + q = ggml_mul_mat(ctx0, model.layers[il].wq_a, cur); + cb(q, "q", il); + + q = build_norm(q, + model.layers[il].attn_q_a_norm, NULL, + LLM_NORM_RMS, il); + cb(q, "q", il); + + // {q_lora_rank, n_head * hparams.n_embd_head_k} * {q_lora_rank, n_tokens} -> {n_head * hparams.n_embd_head_k, n_tokens} + q = ggml_mul_mat(ctx0, model.layers[il].wq_b, q); + cb(q, "q", il); + + // split into {n_head * n_embd_head_qk_nope, n_tokens} + ggml_tensor * q_nope = ggml_view_3d(ctx0, q, n_embd_head_qk_nope, n_head, n_tokens, + ggml_row_size(q->type, hparams.n_embd_head_k), + ggml_row_size(q->type, hparams.n_embd_head_k * n_head), + 0); + cb(q_nope, "q_nope", il); + + // and {n_head * n_embd_head_qk_rope, n_tokens} + ggml_tensor * q_pe = ggml_view_3d(ctx0, q, n_embd_head_qk_rope, n_head, n_tokens, + ggml_row_size(q->type, hparams.n_embd_head_k), + ggml_row_size(q->type, hparams.n_embd_head_k * n_head), + ggml_row_size(q->type, n_embd_head_qk_nope)); + cb(q_pe, "q_pe", il); + + // {n_embd, kv_lora_rank + n_embd_head_qk_rope} * {n_embd, n_tokens} -> {kv_lora_rank + n_embd_head_qk_rope, n_tokens} + ggml_tensor * kv_pe_compresseed = ggml_mul_mat(ctx0, model.layers[il].wkv_a_mqa, cur); + cb(kv_pe_compresseed, "kv_pe_compresseed", il); + + // split into {kv_lora_rank, n_tokens} + ggml_tensor * kv_compressed = ggml_view_2d(ctx0, kv_pe_compresseed, kv_lora_rank, n_tokens, + kv_pe_compresseed->nb[1], + 0); + cb(kv_compressed, "kv_compressed", il); + + // and {n_embd_head_qk_rope, n_tokens} + ggml_tensor * k_pe = ggml_view_3d(ctx0, kv_pe_compresseed, n_embd_head_qk_rope, 1, n_tokens, + kv_pe_compresseed->nb[1], + kv_pe_compresseed->nb[1], + ggml_row_size(kv_pe_compresseed->type, kv_lora_rank)); + cb(k_pe, "k_pe", il); + + // TODO: the CUDA backend used to not support non-cont. (RMS) norm, investigate removing ggml_cont + kv_compressed = ggml_cont(ctx0, kv_compressed); + kv_compressed = build_norm(kv_compressed, + model.layers[il].attn_kv_a_norm, NULL, + LLM_NORM_RMS, il); + cb(kv_compressed, "kv_compressed", il); + + // {kv_lora_rank, n_head * (n_embd_head_qk_nope + n_embd_head_v)} * {kv_lora_rank, n_tokens} -> {n_head * (n_embd_head_qk_nope + n_embd_head_v), n_tokens} + ggml_tensor * kv = ggml_mul_mat(ctx0, model.layers[il].wkv_b, kv_compressed); + cb(kv, "kv", il); + + // split into {n_head * n_embd_head_qk_nope, n_tokens} + ggml_tensor * k_nope = ggml_view_3d(ctx0, kv, n_embd_head_qk_nope, n_head, n_tokens, + ggml_row_size(kv->type, n_embd_head_qk_nope + hparams.n_embd_head_v), + ggml_row_size(kv->type, n_head * (n_embd_head_qk_nope + hparams.n_embd_head_v)), + 0); + cb(k_nope, "k_nope", il); + + // and {n_head * n_embd_head_v, n_tokens} + ggml_tensor * v_states = ggml_view_3d(ctx0, kv, hparams.n_embd_head_v, n_head, n_tokens, + ggml_row_size(kv->type, (n_embd_head_qk_nope + hparams.n_embd_head_v)), + ggml_row_size(kv->type, (n_embd_head_qk_nope + hparams.n_embd_head_v)*n_head), + ggml_row_size(kv->type, (n_embd_head_qk_nope))); + cb(v_states, "v_states", il); + + v_states = ggml_cont(ctx0, v_states); + cb(v_states, "v_states", il); + + v_states = ggml_view_2d(ctx0, v_states, hparams.n_embd_head_v * n_head, n_tokens, + ggml_row_size(kv->type, hparams.n_embd_head_v * n_head), + 0); + cb(v_states, "v_states", il); + + q_pe = ggml_cont(ctx0, q_pe); // TODO: the CUDA backend used to not support non-cont. RoPE, investigate removing this + q_pe = ggml_rope_ext( + ctx0, q_pe, inp_pos, rope_factors, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(q_pe, "q_pe", il); + + // shared RoPE key + k_pe = ggml_cont(ctx0, k_pe); // TODO: the CUDA backend used to not support non-cont. RoPE, investigate removing this + k_pe = ggml_rope_ext( + ctx0, k_pe, inp_pos, rope_factors, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(k_pe, "k_pe", il); + + ggml_tensor * q_states = ggml_concat(ctx0, q_nope, q_pe, 0); + cb(q_states, "q_states", il); + + ggml_tensor * k_states = ggml_concat(ctx0, k_nope, ggml_repeat(ctx0, k_pe, q_pe), 0); + cb(k_states, "k_states", il); + + cur = build_attn(inp_attn, gf, + model.layers[il].wo, NULL, + q_states, k_states, v_states, nullptr, kq_scale, il); + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + + // scale_res - scale the hidden states for residual connection + const float scale_res = scale_depth/sqrtf(float(n_layer)); + cur = ggml_scale(ctx0, cur, scale_res); + cb(cur, "hidden_scaled", il); + + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + { + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + } + + // scale the hidden states for residual connection + cur = ggml_scale(ctx0, cur, scale_res); + cb(cur, "hidden_scaled_ffn", il); + + cur = ggml_add(ctx0, cur, ffn_inp); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = build_norm(cur, + model.output_norm, NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head scaling + const float scale_lmhead = float(n_embd_base)/float(n_embd); + cur = ggml_scale(ctx0, cur, scale_lmhead); + cb(cur, "lmhead_scaling", -1); + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +struct llm_build_gemma : public llm_graph_context { + llm_build_gemma(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + inpL = ggml_scale(ctx0, inpL, sqrtf(n_embd)); + cb(inpL, "inp_scaled", -1); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv_unified(); + + for (int il = 0; il < n_layer; ++il) { + // norm + cur = build_norm(inpL, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self-attention + { + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + Qcur = ggml_scale(ctx0, Qcur, 1.0f / sqrtf(float(n_embd_head))); + cb(Qcur, "Qcur_scaled", il); + + cur = build_attn(inp_attn, gf, + model.layers[il].wo, NULL, + Qcur, Kcur, Vcur, nullptr, 1.0f, il); + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); + } + + ggml_tensor * sa_out = ggml_add(ctx0, cur, inpL); + cb(sa_out, "sa_out", il); + + cur = build_norm(sa_out, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + // feed-forward network + { + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + LLM_FFN_GELU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + } + + cur = ggml_add(ctx0, cur, sa_out); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = build_norm(cur, + model.output_norm, NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +struct llm_build_gemma2 : public llm_graph_context { + llm_build_gemma2(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_k; + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + inpL = ggml_scale(ctx0, inpL, sqrtf(n_embd)); + cb(inpL, "inp_scaled", -1); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv_unified(); + + for (int il = 0; il < n_layer; ++il) { + // norm + cur = build_norm(inpL, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self-attention + { + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + // ref: https://github.com/google/gemma_pytorch/commit/03e657582d17cb5a8617ebf333c1c16f3694670e + switch (model.type) { + case LLM_TYPE_2B: + case LLM_TYPE_9B: + case LLM_TYPE_27B: Qcur = ggml_scale(ctx0, Qcur, 1.0f / sqrtf(float(n_embd_head))); break; + default: GGML_ABORT("fatal error"); + }; + cb(Qcur, "Qcur_scaled", il); + + cur = build_attn(inp_attn, gf, + model.layers[il].wo, NULL, + Qcur, Kcur, Vcur, nullptr, 1.0f, il); + } + + cur = build_norm(cur, + model.layers[il].attn_post_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_post_norm", il); + + if (il == n_layer - 1) { + // skip computing output for unused tokens + ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); + } + + ggml_tensor * sa_out = ggml_add(ctx0, cur, inpL); + cb(sa_out, "sa_out", il); + + cur = build_norm(sa_out, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + // feed-forward network + { + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + LLM_FFN_GELU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + } + + cur = build_norm(cur, + model.layers[il].ffn_post_norm, NULL, + LLM_NORM_RMS, -1); + cb(cur, "ffn_post_norm", -1); + + cur = ggml_add(ctx0, cur, sa_out); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = build_norm(cur, + model.output_norm, NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + // final logit soft-capping + cur = ggml_scale(ctx0, cur, 1.0f / hparams.f_final_logit_softcapping); + cur = ggml_tanh(ctx0, cur); + cur = ggml_scale(ctx0, cur, hparams.f_final_logit_softcapping); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +struct llm_build_gemma3 : public llm_graph_context { + llm_build_gemma3(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_k; + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // important: do not normalize weights for raw embeddings input (i.e. encoded image emdeddings) + if (ubatch.token) { + inpL = ggml_scale(ctx0, inpL, sqrtf(n_embd)); + cb(inpL, "inp_scaled", -1); + } + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + // TODO: is causal == true correct? might need some changes + auto * inp_attn = build_attn_inp_kv_unified(); + + for (int il = 0; il < n_layer; ++il) { + const bool is_swa = hparams.is_swa(il); + + const float freq_base_l = is_swa ? hparams.rope_freq_base_train_swa : cparams.rope_freq_base; + const float freq_scale_l = is_swa ? hparams.rope_freq_scale_train_swa : cparams.rope_freq_scale; + + // norm + cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self-attention + { + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il); + cb(Qcur, "Qcur_normed", il); + + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l, + ext_factor, attn_factor, beta_fast, beta_slow); + + Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il); + cb(Kcur, "Kcur_normed", il); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l, + ext_factor, attn_factor, beta_fast, beta_slow); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, gf, + model.layers[il].wo, NULL, + Qcur, Kcur, Vcur, nullptr, hparams.f_attention_scale, il); + } + + cur = build_norm(cur, + model.layers[il].attn_post_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_post_norm", il); + + if (il == n_layer - 1) { + // skip computing output for unused tokens + ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); + } + + ggml_tensor * sa_out = ggml_add(ctx0, cur, inpL); + cb(sa_out, "sa_out", il); + + cur = build_norm(sa_out, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + // feed-forward network + { + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + LLM_FFN_GELU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + } + + cur = build_norm(cur, + model.layers[il].ffn_post_norm, NULL, + LLM_NORM_RMS, -1); + cb(cur, "ffn_post_norm", -1); + + cur = ggml_add(ctx0, cur, sa_out); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = build_norm(cur, + model.output_norm, NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +// TODO: move up next to build_starcoder +struct llm_build_starcoder2 : public llm_graph_context { + llm_build_starcoder2(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv_unified(); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + // norm + cur = build_norm(inpL, + model.layers[il].attn_norm, model.layers[il].attn_norm_b, + LLM_NORM, il); + cb(cur, "attn_norm", il); + + // self-attention + { + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + if (model.layers[il].bq) { + Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); + cb(Qcur, "Qcur", il); + } + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + if (model.layers[il].bk) { + Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); + cb(Kcur, "Kcur", il); + } + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + if (model.layers[il].bv) { + Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); + cb(Vcur, "Vcur", il); + } + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, gf, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, model.layers[il].ffn_norm_b, + LLM_NORM, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(cur, + model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL, + NULL, NULL, NULL, + model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, + NULL, + LLM_FFN_GELU, LLM_FFN_SEQ, il); + cb(cur, "ffn_out", il); + + cur = ggml_add(ctx0, cur, ffn_inp); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = build_norm(cur, + model.output_norm, model.output_norm_b, + LLM_NORM, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +struct llm_build_mamba : public llm_graph_context { + const llama_model & model; + + llm_build_mamba(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params), model(model) { + ggml_tensor * cur; + ggml_tensor * inpL; + + // {n_embd, n_tokens} + inpL = build_inp_embd(model.tok_embd); + + ggml_tensor * state_copy = build_inp_s_copy(); + ggml_tensor * state_mask = build_inp_s_mask(); + + for (int il = 0; il < n_layer; ++il) { + // norm + cur = build_norm(inpL, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + //cur = build_mamba_layer(gf, cur, state_copy, state_mask, il); + cur = build_mamba_layer(gf, cur, state_copy, state_mask, ubatch, il); + + if (il == n_layer - 1) { + // skip computing output for unused tokens + ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); + } + + // residual + cur = ggml_add(ctx0, cur, inpL); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + // final rmsnorm + cur = build_norm(inpL, + model.output_norm, NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } + + // TODO: split + ggml_tensor * build_mamba_layer( + ggml_cgraph * gf, + ggml_tensor * cur, + ggml_tensor * state_copy, + ggml_tensor * state_mask, + const llama_ubatch & ubatch, + int il) const { + const llama_kv_cache_unified * kv_self = static_cast(memory); + + const auto kv_head = kv_self->head; + + const int64_t d_conv = hparams.ssm_d_conv; + const int64_t d_inner = hparams.ssm_d_inner; + const int64_t d_state = hparams.ssm_d_state; + const int64_t dt_rank = hparams.ssm_dt_rank; + const int64_t n_seqs = ubatch.n_seqs; + // Some variants of Mamba arch (e.g. FalconMamba do apply layer norm on B and Dt layers) + const bool ssm_dt_b_c_rms = hparams.ssm_dt_b_c_rms; + // Use the same RMS norm as the final layer norm + const float norm_rms_eps = hparams.f_norm_rms_eps; + + const int64_t n_seq_tokens = ubatch.n_seq_tokens; + + GGML_ASSERT(n_seqs != 0); + GGML_ASSERT(ubatch.equal_seqs); + GGML_ASSERT(ubatch.n_tokens == n_seq_tokens * n_seqs); + + ggml_tensor * conv_states_all = kv_self->k_l[il]; + ggml_tensor * ssm_states_all = kv_self->v_l[il]; + + // (ab)using the KV cache to store the states + ggml_tensor * conv = build_copy_mask_state( + gf, conv_states_all, state_copy, state_mask, + hparams.n_embd_k_s(), n_seqs); + conv = ggml_reshape_3d(ctx0, conv, d_conv - 1, d_inner, n_seqs); + ggml_tensor * ssm = build_copy_mask_state( + gf, ssm_states_all, state_copy, state_mask, + hparams.n_embd_v_s(), n_seqs); + ssm = ggml_reshape_3d(ctx0, ssm, d_state, d_inner, n_seqs); + + // {n_embd, n_tokens} => {n_embd, n_seq_tokens, n_seqs} + cur = ggml_reshape_3d(ctx0, cur, cur->ne[0], n_seq_tokens, n_seqs); + + // {n_embd, 2*d_inner} @ {n_embd, n_seq_tokens, n_seqs} => {2*d_inner, n_seq_tokens, n_seqs} + ggml_tensor * xz = build_lora_mm(model.layers[il].ssm_in, cur); + // split the above in two + // => {d_inner, n_seq_tokens, n_seqs} + ggml_tensor * x = ggml_view_3d(ctx0, xz, d_inner, xz->ne[1], xz->ne[2], xz->nb[1], xz->nb[2], 0); + ggml_tensor * z = ggml_view_3d(ctx0, xz, d_inner, xz->ne[1], xz->ne[2], xz->nb[1], xz->nb[2], d_inner*ggml_element_size(xz)); + + // conv + { + // => {d_conv - 1 + n_seq_tokens, d_inner, n_seqs} + ggml_tensor * conv_x = ggml_concat(ctx0, conv, ggml_transpose(ctx0, x), 0); + + // copy last (d_conv - 1) columns back into the state cache + ggml_tensor * last_conv = ggml_view_3d(ctx0, conv_x, d_conv - 1, d_inner, n_seqs, conv_x->nb[1], conv_x->nb[2], n_seq_tokens*(conv_x->nb[0])); + + ggml_build_forward_expand(gf, + ggml_cpy(ctx0, last_conv, + ggml_view_1d(ctx0, conv_states_all, + (d_conv - 1)*(d_inner)*(n_seqs), + kv_head*(d_conv - 1)*(d_inner)*ggml_element_size(conv_states_all)))); + + // 1D convolution + // The equivalent is to make a self-overlapping view of conv_x + // over d_conv columns at each stride in the 3rd dimension, + // then element-wise multiply that with the conv1d weight, + // then sum the elements of each row, + // (the last two steps are a dot product over rows (also doable with mul_mat)) + // then permute away the ne[0] dimension, + // and then you're left with the resulting x tensor. + // For simultaneous sequences, all sequences need to have the same length. + x = ggml_ssm_conv(ctx0, conv_x, model.layers[il].ssm_conv1d); + + // bias + x = ggml_add(ctx0, x, model.layers[il].ssm_conv1d_b); + + x = ggml_silu(ctx0, x); + } + + // ssm + { + // {d_inner, dt_rank + 2*d_state} @ {d_inner, n_seq_tokens, n_seqs} => {dt_rank + 2*d_state, n_seq_tokens, n_seqs} + ggml_tensor * x_db = build_lora_mm(model.layers[il].ssm_x, x); + // split + ggml_tensor * dt = ggml_view_3d(ctx0, x_db, dt_rank, n_seq_tokens, n_seqs, x_db->nb[1], x_db->nb[2], 0); + ggml_tensor * B = ggml_view_3d(ctx0, x_db, d_state, n_seq_tokens, n_seqs, x_db->nb[1], x_db->nb[2], ggml_element_size(x_db)*dt_rank); + ggml_tensor * C = ggml_view_3d(ctx0, x_db, d_state, n_seq_tokens, n_seqs, x_db->nb[1], x_db->nb[2], ggml_element_size(x_db)*(dt_rank+d_state)); + + // Some Mamba variants (e.g. FalconMamba) apply RMS norm in B, C & Dt layers + if (ssm_dt_b_c_rms) { + dt = ggml_rms_norm(ctx0, dt, norm_rms_eps); + B = ggml_rms_norm(ctx0, B, norm_rms_eps); + C = ggml_rms_norm(ctx0, C, norm_rms_eps); + } + + // {dt_rank, d_inner} @ {dt_rank, n_seq_tokens, n_seqs} => {d_inner, n_seq_tokens, n_seqs} + dt = build_lora_mm(model.layers[il].ssm_dt, dt); + dt = ggml_add(ctx0, dt, model.layers[il].ssm_dt_b); + + // Custom operator to optimize the parallel associative scan + // as described in the Annex D of the Mamba paper. + // => {d_inner, n_seq_tokens, n_seqs} and {d_state, d_inner, n_seqs} + ggml_tensor * y_ssm = ggml_ssm_scan(ctx0, ssm, x, dt, model.layers[il].ssm_a, B, C); + + // store last states + ggml_build_forward_expand(gf, + ggml_cpy(ctx0, + ggml_view_1d(ctx0, y_ssm, d_state*d_inner*n_seqs, x->nb[3]), + ggml_view_1d(ctx0, ssm_states_all, d_state*d_inner*n_seqs, kv_head*d_state*d_inner*ggml_element_size(ssm_states_all)))); + + ggml_tensor * y = ggml_view_3d(ctx0, y_ssm, d_inner, n_seq_tokens, n_seqs, x->nb[1], x->nb[2], 0); + + // TODO: skip computing output earlier for unused tokens + + // {d_inner, n_seq_tokens, n_seqs} * {d_inner} => {d_inner, n_seq_tokens, n_seqs} + y = ggml_add(ctx0, y, ggml_mul(ctx0, x, model.layers[il].ssm_d)); + y = ggml_mul(ctx0, y, ggml_silu(ctx0, ggml_cont(ctx0, z))); + + // {d_inner, n_embd} @ {d_inner, n_seq_tokens, n_seqs} => {n_embd, n_seq_tokens, n_seqs} + cur = build_lora_mm(model.layers[il].ssm_out, y); + } + + // {n_embd, n_seq_tokens, n_seqs} => {n_embd, n_tokens} + cur = ggml_reshape_2d(ctx0, cur, cur->ne[0], n_seq_tokens * n_seqs); + //cb(cur, "mamba_out", il); + + return cur; + } +}; + +struct llm_build_command_r : public llm_graph_context { + llm_build_command_r(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + + const float f_logit_scale = hparams.f_logit_scale; + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv_unified(); + + for (int il = 0; il < n_layer; ++il) { + + // norm + cur = build_norm(inpL, + model.layers[il].attn_norm, NULL, + LLM_NORM, il); + cb(cur, "attn_norm", il); + ggml_tensor * ffn_inp = cur; + + // self-attention + { + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + if (model.layers[il].bq) { + Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); + cb(Qcur, "Qcur", il); + } + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + if (model.layers[il].bk) { + Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); + cb(Kcur, "Kcur", il); + } + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + if (model.layers[il].bv) { + Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); + cb(Vcur, "Vcur", il); + } + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + if (model.layers[il].attn_q_norm) { + Qcur = build_norm(Qcur, + model.layers[il].attn_q_norm, + NULL, + LLM_NORM, il); + cb(Qcur, "Qcur", il); + } + + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + if (model.layers[il].attn_k_norm) { + Kcur = build_norm(Kcur, + model.layers[il].attn_k_norm, + NULL, + LLM_NORM, il); + cb(Kcur, "Kcur", il); + } + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, gf, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); + ffn_inp = ggml_get_rows(ctx0, ffn_inp, inp_out_ids); + } + + ggml_tensor * attn_out = cur; + + // feed-forward network + { + cur = build_ffn(ffn_inp, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + } + + // add together residual + FFN + self-attention + cur = ggml_add(ctx0, cur, inpL); + cur = ggml_add(ctx0, cur, attn_out); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = build_norm(cur, + model.output_norm, NULL, + LLM_NORM, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + if (f_logit_scale) { + cur = ggml_scale(ctx0, cur, f_logit_scale); + } + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +struct llm_build_cohere2 : public llm_graph_context { + llm_build_cohere2(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + + const float f_logit_scale = hparams.f_logit_scale; + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv_unified(); + + for (int il = 0; il < n_layer; ++il) { + const bool is_swa = hparams.is_swa(il); + + // norm + cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM, il); + cb(cur, "attn_norm", il); + ggml_tensor * ffn_inp = cur; + + // self-attention + { + // rope freq factors for 128k context + ggml_tensor * rope_factors = static_cast(memory)->cbs.get_rope_factors(n_ctx_per_seq, il); + + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + if (model.layers[il].bq) { + Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); + cb(Qcur, "Qcur", il); + } + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + if (model.layers[il].bk) { + Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); + cb(Kcur, "Kcur", il); + } + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + if (model.layers[il].bv) { + Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); + cb(Vcur, "Vcur", il); + } + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + if (is_swa) { + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, rope_factors, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, rope_factors, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + } + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, gf, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); + ffn_inp = ggml_get_rows(ctx0, ffn_inp, inp_out_ids); + } + + ggml_tensor * attn_out = cur; + + // feed-forward network + { + cur = build_ffn(ffn_inp, model.layers[il].ffn_up, NULL, NULL, model.layers[il].ffn_gate, + NULL, NULL, model.layers[il].ffn_down, NULL, NULL, NULL, LLM_FFN_SILU, LLM_FFN_PAR, + il); + cb(cur, "ffn_out", il); + } + + // add together residual + FFN + self-attention + cur = ggml_add(ctx0, cur, inpL); + cur = ggml_add(ctx0, cur, attn_out); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = build_norm(cur, model.output_norm, NULL, LLM_NORM, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + if (f_logit_scale) { + cur = ggml_scale(ctx0, cur, f_logit_scale); + } + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +// ref: https://allenai.org/olmo +// based on the original build_llama() function, changes: +// * non-parametric layer norm +// * clamp qkv +// * removed bias +// * removed MoE +struct llm_build_olmo : public llm_graph_context { + llm_build_olmo(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv_unified(); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + // norm + cur = build_norm(inpL, + NULL, NULL, + LLM_NORM, il); + cb(cur, "attn_norm", il); + + // self-attention + { + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + if (hparams.f_clamp_kqv > 0.0f) { + Qcur = ggml_clamp(ctx0, Qcur, -hparams.f_clamp_kqv, hparams.f_clamp_kqv); + cb(Qcur, "Qcur", il); + } + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + if (hparams.f_clamp_kqv > 0.0f) { + Kcur = ggml_clamp(ctx0, Kcur, -hparams.f_clamp_kqv, hparams.f_clamp_kqv); + cb(Kcur, "Kcur", il); + } + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + if (hparams.f_clamp_kqv > 0.0f) { + Vcur = ggml_clamp(ctx0, Vcur, -hparams.f_clamp_kqv, hparams.f_clamp_kqv); + cb(Vcur, "Vcur", il); + } + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, gf, + model.layers[il].wo, nullptr, + Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + cur = build_norm(ffn_inp, + NULL, NULL, + LLM_NORM, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + + cur = ggml_add(ctx0, cur, ffn_inp); + cb(cur, "ffn_out", il); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = build_norm(cur, + NULL, NULL, + LLM_NORM, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +struct llm_build_olmo2 : public llm_graph_context { + llm_build_olmo2(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv_unified(); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + cur = inpL; + + // self_attention + { + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + + Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, + LLM_NORM_RMS, il); + cb(Qcur, "Qcur_normed", il); + + Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, + LLM_NORM_RMS, il); + cb(Kcur, "Kcur_normed", il); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, gf, + model.layers[il].wo, NULL, + Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + } + + cur = build_norm(cur, + model.layers[il].attn_post_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_post_norm", il); + + if (il == n_layer - 1) { + // skip computing output for unused tokens + ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + cur = build_ffn(ffn_inp, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + + cur = build_norm(cur, + model.layers[il].ffn_post_norm, NULL, + LLM_NORM_RMS, -1); + cb(cur, "ffn_post_norm", -1); + + cur = ggml_add(ctx0, cur, ffn_inp); + cb(cur, "ffn_out", il); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = build_norm(cur, + model.output_norm, NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +// based on the build_qwen2moe() function, changes: +// * removed shared experts +// * removed bias +// * added q, k norm +struct llm_build_olmoe : public llm_graph_context { + llm_build_olmoe(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv_unified(); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + // norm + cur = build_norm(inpL, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self_attention + { + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + + Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, + LLM_NORM_RMS, il); + cb(Qcur, "Qcur_normed", il); + + Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, + LLM_NORM_RMS, il); + cb(Kcur, "Kcur_normed", il); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, gf, + model.layers[il].wo, NULL, + Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // MoE branch + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + cur = build_moe_ffn(cur, + model.layers[il].ffn_gate_inp, + model.layers[il].ffn_up_exps, + model.layers[il].ffn_gate_exps, + model.layers[il].ffn_down_exps, + nullptr, + n_expert, n_expert_used, + LLM_FFN_SILU, false, + false, 0.0, + LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, + il); + cb(cur, "ffn_moe_out", il); + + cur = ggml_add(ctx0, cur, ffn_inp); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = build_norm(cur, + model.output_norm, NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +struct llm_build_openelm : public llm_graph_context { + llm_build_openelm(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + + ggml_tensor * cur; + ggml_tensor * inpL; + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv_unified(); + + for (int il = 0; il < n_layer; ++il) { + const int64_t n_head = hparams.n_head(il); + const int64_t n_head_kv = hparams.n_head_kv(il); + const int64_t n_head_qkv = 2*n_head_kv + n_head; + + cur = inpL; + ggml_tensor * residual = cur; + + // norm + cur = build_norm(inpL, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self-attention + { + cur = build_lora_mm(model.layers[il].wqkv, cur); + cb(cur, "wqkv", il); + + cur = ggml_reshape_3d(ctx0, cur, n_embd_head_k, n_head_qkv, n_tokens); + + ggml_tensor * Qcur = ggml_cont(ctx0, ggml_view_3d(ctx0, cur, n_embd_head, n_head, n_tokens, cur->nb[1], cur->nb[2], 0)); + cb(Qcur, "Qcur", il); + + ggml_tensor * Kcur = ggml_cont(ctx0, ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, cur->nb[1], cur->nb[2], cur->nb[1]*n_head)); + cb(Kcur, "Kcur", il); + + ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, cur->nb[1], cur->nb[2], cur->nb[1]*(n_head+n_head_kv))); + cb(Vcur, "Vcur", il); + + Qcur = build_norm(Qcur, + model.layers[il].attn_q_norm, NULL, + LLM_NORM_RMS, il); + cb(Qcur, "Qcur", il); + + Kcur = build_norm(Kcur, + model.layers[il].attn_k_norm, NULL, + LLM_NORM_RMS, il); + cb(Kcur, "Kcur", il); + + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, NULL, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, NULL, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Qcur, "Vcur", il); + + cur = build_attn(inp_attn, gf, + model.layers[il].wo, NULL, + Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + ggml_tensor * inp_out_ids = build_inp_out_ids(); + residual = ggml_get_rows(ctx0, residual, inp_out_ids); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + } + + ggml_tensor * ffn_inp = ggml_add(ctx0, residual, cur); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + { + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + } + + cur = ggml_add(ctx0, cur, ffn_inp); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + inpL = cur; + } + + cur = inpL; + + // norm + cur = build_norm(cur, + model.output_norm, NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +struct llm_build_gptneox : public llm_graph_context { + llm_build_gptneox(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + const int64_t n_embd_gqa = hparams.n_embd_v_gqa(); + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv_unified(); + + for (int il = 0; il < n_layer; ++il) { + cur = build_norm(inpL, + model.layers[il].attn_norm, + model.layers[il].attn_norm_b, + LLM_NORM, il); + cb(cur, "attn_norm", il); + + // self-attention + { + cur = build_lora_mm(model.layers[il].wqkv, cur); + cb(cur, "wqkv", il); + + cur = ggml_add(ctx0, cur, model.layers[il].bqkv); + cb(cur, "bqkv", il); + + ggml_tensor * Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd))); + ggml_tensor * Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd))); + ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa))); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, gf, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); + } + + // ffn + if (hparams.use_par_res) { + // attention and ffn are computed in parallel + // x = x + attn(ln1(x)) + ffn(ln2(x)) + + ggml_tensor * attn_out = cur; + + cur = build_norm(inpL, + model.layers[il].ffn_norm, + model.layers[il].ffn_norm_b, + LLM_NORM, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(cur, + model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL, + NULL, NULL, NULL, + model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, + NULL, + LLM_FFN_GELU, LLM_FFN_SEQ, il); + cb(cur, "ffn_out", il); + + cur = ggml_add(ctx0, cur, inpL); + cb(cur, "ffn_out", il); + + cur = ggml_add(ctx0, cur, attn_out); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } else { + // attention and ffn are computed sequentially + // x = x + attn(ln1(x)) + // x = x + ffn(ln2(x)) + + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL); + cb(ffn_inp, "ffn_inp", il); + + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, + model.layers[il].ffn_norm_b, + LLM_NORM, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(cur, + model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL, + NULL, NULL, NULL, + model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, + NULL, + LLM_FFN_GELU, LLM_FFN_SEQ, il); + cb(cur, "ffn_out", il); + + cur = ggml_add(ctx0, cur, ffn_inp); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + } + + cur = build_norm(inpL, + model.output_norm, + model.output_norm_b, + LLM_NORM, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +struct llm_build_arctic : public llm_graph_context { + llm_build_arctic(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv_unified(); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + // norm + cur = build_norm(inpL, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self-attention + { + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, gf, + model.layers[il].wo, NULL, + Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + + ggml_tensor * ffn_out = ggml_add(ctx0, cur, ffn_inp); + cb(ffn_out, "ffn_out", il); + + // MoE + cur = build_norm(inpSA, + model.layers[il].ffn_norm_exps, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm_exps", il); + + cur = build_moe_ffn(cur, + model.layers[il].ffn_gate_inp, + model.layers[il].ffn_up_exps, + model.layers[il].ffn_gate_exps, + model.layers[il].ffn_down_exps, + nullptr, + n_expert, n_expert_used, + LLM_FFN_SILU, true, + false, 0.0, + LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, + il); + cb(cur, "ffn_moe_out", il); + + cur = ggml_add(ctx0, cur, ffn_out); + cb(cur, "ffn_out", il); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = build_norm(cur, + model.output_norm, NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +struct llm_build_deepseek : public llm_graph_context { + llm_build_deepseek(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv_unified(); + + const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale; + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + // norm + cur = build_norm(inpL, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self-attention + { + // rope freq factors for llama3; may return nullptr for llama2 and other models + ggml_tensor * rope_factors = static_cast(memory)->cbs.get_rope_factors(n_ctx_per_seq, il); + + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + if (model.layers[il].bq) { + Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); + cb(Qcur, "Qcur", il); + } + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + if (model.layers[il].bk) { + Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); + cb(Kcur, "Kcur", il); + } + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + if (model.layers[il].bv) { + Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); + cb(Vcur, "Vcur", il); + } + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, rope_factors, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, rope_factors, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, gf, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, nullptr, kq_scale, il); + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + + + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + if ((uint32_t) il < hparams.n_layer_dense_lead) { + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + } else { + // MoE branch + ggml_tensor * moe_out = + build_moe_ffn(cur, + model.layers[il].ffn_gate_inp, + model.layers[il].ffn_up_exps, + model.layers[il].ffn_gate_exps, + model.layers[il].ffn_down_exps, + nullptr, + n_expert, n_expert_used, + LLM_FFN_SILU, false, + false, hparams.expert_weights_scale, + LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, + il); + cb(moe_out, "ffn_moe_out", il); + + // FFN shared expert + { + ggml_tensor * ffn_shexp = build_ffn(cur, + model.layers[il].ffn_up_shexp, NULL, NULL, + model.layers[il].ffn_gate_shexp, NULL, NULL, + model.layers[il].ffn_down_shexp, NULL, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(ffn_shexp, "ffn_shexp", il); + + cur = ggml_add(ctx0, moe_out, ffn_shexp); + cb(cur, "ffn_out", il); + } + } + + cur = ggml_add(ctx0, cur, ffn_inp); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = build_norm(cur, + model.output_norm, NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +struct llm_build_deepseek2 : public llm_graph_context { + llm_build_deepseek2(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { + bool is_lite = (hparams.n_layer == 27); + + // We have to pre-scale kq_scale and attn_factor to make the YaRN RoPE work correctly. + // See https://github.com/ggerganov/llama.cpp/discussions/7416 for detailed explanation. + const float mscale = attn_factor * (1.0f + hparams.rope_yarn_log_mul * logf(1.0f / freq_scale)); + const float kq_scale = 1.0f*mscale*mscale/sqrtf(float(hparams.n_embd_head_k)); + const float attn_factor_scaled = 1.0f / (1.0f + 0.1f * logf(1.0f / freq_scale)); + + const uint32_t n_embd_head_qk_rope = hparams.n_rot; + const uint32_t n_embd_head_qk_nope = hparams.n_embd_head_k - hparams.n_rot; + const uint32_t kv_lora_rank = hparams.n_lora_kv; + + ggml_tensor * cur; + ggml_tensor * inpL; + + // {n_embd, n_tokens} + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv_unified(); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + // norm + cur = build_norm(inpL, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self_attention + { + ggml_tensor * q = NULL; + if (!is_lite) { + // {n_embd, q_lora_rank} * {n_embd, n_tokens} -> {q_lora_rank, n_tokens} + q = ggml_mul_mat(ctx0, model.layers[il].wq_a, cur); + cb(q, "q", il); + + q = build_norm(q, + model.layers[il].attn_q_a_norm, NULL, + LLM_NORM_RMS, il); + cb(q, "q", il); + + // {q_lora_rank, n_head * hparams.n_embd_head_k} * {q_lora_rank, n_tokens} -> {n_head * hparams.n_embd_head_k, n_tokens} + q = ggml_mul_mat(ctx0, model.layers[il].wq_b, q); + cb(q, "q", il); + } else { + q = ggml_mul_mat(ctx0, model.layers[il].wq, cur); + cb(q, "q", il); + } + + // split into {n_head * n_embd_head_qk_nope, n_tokens} + ggml_tensor * q_nope = ggml_view_3d(ctx0, q, n_embd_head_qk_nope, n_head, n_tokens, + ggml_row_size(q->type, hparams.n_embd_head_k), + ggml_row_size(q->type, hparams.n_embd_head_k * n_head), + 0); + cb(q_nope, "q_nope", il); + + // and {n_head * n_embd_head_qk_rope, n_tokens} + ggml_tensor * q_pe = ggml_view_3d(ctx0, q, n_embd_head_qk_rope, n_head, n_tokens, + ggml_row_size(q->type, hparams.n_embd_head_k), + ggml_row_size(q->type, hparams.n_embd_head_k * n_head), + ggml_row_size(q->type, n_embd_head_qk_nope)); + cb(q_pe, "q_pe", il); + + // {n_embd, kv_lora_rank + n_embd_head_qk_rope} * {n_embd, n_tokens} -> {kv_lora_rank + n_embd_head_qk_rope, n_tokens} + ggml_tensor * kv_pe_compresseed = ggml_mul_mat(ctx0, model.layers[il].wkv_a_mqa, cur); + cb(kv_pe_compresseed, "kv_pe_compresseed", il); + + // split into {kv_lora_rank, n_tokens} + ggml_tensor * kv_compressed = ggml_view_2d(ctx0, kv_pe_compresseed, kv_lora_rank, n_tokens, + kv_pe_compresseed->nb[1], + 0); + cb(kv_compressed, "kv_compressed", il); + + // and {n_embd_head_qk_rope, n_tokens} + ggml_tensor * k_pe = ggml_view_3d(ctx0, kv_pe_compresseed, n_embd_head_qk_rope, 1, n_tokens, + kv_pe_compresseed->nb[1], + kv_pe_compresseed->nb[1], + ggml_row_size(kv_pe_compresseed->type, kv_lora_rank)); + cb(k_pe, "k_pe", il); + + // TODO: the CUDA backend used to not support non-cont. (RMS) norm, investigate removing ggml_cont + kv_compressed = ggml_cont(ctx0, kv_compressed); + kv_compressed = build_norm(kv_compressed, + model.layers[il].attn_kv_a_norm, NULL, + LLM_NORM_RMS, il); + cb(kv_compressed, "kv_compressed", il); + + // {kv_lora_rank, n_head * (n_embd_head_qk_nope + n_embd_head_v)} * {kv_lora_rank, n_tokens} -> {n_head * (n_embd_head_qk_nope + n_embd_head_v), n_tokens} + ggml_tensor * kv = ggml_mul_mat(ctx0, model.layers[il].wkv_b, kv_compressed); + cb(kv, "kv", il); + + // split into {n_head * n_embd_head_qk_nope, n_tokens} + ggml_tensor * k_nope = ggml_view_3d(ctx0, kv, n_embd_head_qk_nope, n_head, n_tokens, + ggml_row_size(kv->type, n_embd_head_qk_nope + hparams.n_embd_head_v), + ggml_row_size(kv->type, n_head * (n_embd_head_qk_nope + hparams.n_embd_head_v)), + 0); + cb(k_nope, "k_nope", il); + + // and {n_head * n_embd_head_v, n_tokens} + ggml_tensor * v_states = ggml_view_3d(ctx0, kv, hparams.n_embd_head_v, n_head, n_tokens, + ggml_row_size(kv->type, (n_embd_head_qk_nope + hparams.n_embd_head_v)), + ggml_row_size(kv->type, (n_embd_head_qk_nope + hparams.n_embd_head_v)*n_head), + ggml_row_size(kv->type, (n_embd_head_qk_nope))); + cb(v_states, "v_states", il); + + v_states = ggml_cont(ctx0, v_states); + cb(v_states, "v_states", il); + + v_states = ggml_view_2d(ctx0, v_states, hparams.n_embd_head_v * n_head, n_tokens, + ggml_row_size(kv->type, hparams.n_embd_head_v * n_head), + 0); + cb(v_states, "v_states", il); + + q_pe = ggml_cont(ctx0, q_pe); // TODO: the CUDA backend used to not support non-cont. RoPE, investigate removing this + q_pe = ggml_rope_ext( + ctx0, q_pe, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor_scaled, beta_fast, beta_slow + ); + cb(q_pe, "q_pe", il); + + // shared RoPE key + k_pe = ggml_cont(ctx0, k_pe); // TODO: the CUDA backend used to not support non-cont. RoPE, investigate removing this + k_pe = ggml_rope_ext( + ctx0, k_pe, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor_scaled, beta_fast, beta_slow + ); + cb(k_pe, "k_pe", il); + + ggml_tensor * q_states = ggml_concat(ctx0, q_nope, q_pe, 0); + cb(q_states, "q_states", il); + + ggml_tensor * k_states = ggml_concat(ctx0, k_nope, ggml_repeat(ctx0, k_pe, q_pe), 0); + cb(k_states, "k_states", il); + + cur = build_attn(inp_attn, gf, + model.layers[il].wo, NULL, + q_states, k_states, v_states, nullptr, kq_scale, il); + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + if ((uint32_t) il < hparams.n_layer_dense_lead) { + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + } else { + // MoE branch + ggml_tensor * moe_out = + build_moe_ffn(cur, + model.layers[il].ffn_gate_inp, + model.layers[il].ffn_up_exps, + model.layers[il].ffn_gate_exps, + model.layers[il].ffn_down_exps, + model.layers[il].ffn_exp_probs_b, + n_expert, n_expert_used, + LLM_FFN_SILU, hparams.expert_weights_norm, + true, hparams.expert_weights_scale, + (llama_expert_gating_func_type) hparams.expert_gating_func, + il); + cb(moe_out, "ffn_moe_out", il); + + // FFN shared expert + { + ggml_tensor * ffn_shexp = build_ffn(cur, + model.layers[il].ffn_up_shexp, NULL, NULL, + model.layers[il].ffn_gate_shexp, NULL, NULL, + model.layers[il].ffn_down_shexp, NULL, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(ffn_shexp, "ffn_shexp", il); + + cur = ggml_add(ctx0, moe_out, ffn_shexp); + cb(cur, "ffn_out", il); + } + } + + cur = ggml_add(ctx0, cur, ffn_inp); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = build_norm(cur, + model.output_norm, NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = ggml_mul_mat(ctx0, model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +struct llm_build_bitnet : public llm_graph_context { + llm_build_bitnet(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv_unified(); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + cur = build_norm(inpL, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self-attention + { + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + if (model.layers[il].wq_scale) { + Qcur = ggml_mul(ctx0, Qcur, model.layers[il].wq_scale); + } + cb(Qcur, "Qcur", il); + if (model.layers[il].bq) { + Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); + cb(Qcur, "Qcur", il); + } + + // B1.K + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + if (model.layers[il].wk_scale) { + Kcur = ggml_mul(ctx0, Kcur, model.layers[il].wk_scale); + } + cb(Kcur, "Kcur", il); + if (model.layers[il].bk) { + Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); + cb(Kcur, "Kcur", il); + } + + // B1.V + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + if (model.layers[il].wv_scale) { + Vcur = ggml_mul(ctx0, Vcur, model.layers[il].wv_scale); + } + cb(Vcur, "Vcur", il); + if (model.layers[il].bv) { + Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); + cb(Vcur, "Vcur", il); + } + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, gf, + NULL, NULL, + Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + + cur = build_norm(cur, + model.layers[il].attn_sub_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_sub_norm", il); + + cur = build_lora_mm(model.layers[il].wo, cur); + if (model.layers[il].wo_scale) { + cur = ggml_mul(ctx0, cur, model.layers[il].wo_scale); + } + if (model.layers[il].bo) { + cur = ggml_add(ctx0, cur, model.layers[il].bo); + } + cb(cur, "attn_o_out", il); + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward forward + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, model.layers[il].ffn_up_scale, + model.layers[il].ffn_gate, NULL, model.layers[il].ffn_gate_scale, + NULL, NULL, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_sub_out", il); + + cur = build_norm(cur, + model.layers[il].ffn_sub_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_sub_norm", il); + + cur = build_lora_mm(model.layers[il].ffn_down, cur); + if (model.layers[il].ffn_down_scale) { + cur = ggml_mul(ctx0, cur, model.layers[il].ffn_down_scale); + } + cb(cur, "ffn_down", il); + + cur = ggml_add(ctx0, cur, ffn_inp); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = build_norm(cur, + model.output_norm, NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + // FIXME: do not use model.tok_embd directly, duplicate as model.output + cur = build_lora_mm(model.tok_embd, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +struct llm_build_t5_enc : public llm_graph_context { + llm_build_t5_enc(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + ggml_tensor * pos_bucket_enc = build_inp_pos_bucket_enc(); + + auto * inp_attn = build_attn_inp_no_cache(); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + // norm + cur = build_norm(inpL, + model.layers[il].attn_norm_enc, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self-attention + { + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq_enc, cur); + cb(Qcur, "Qcur", il); + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk_enc, cur); + cb(Kcur, "Kcur", il); + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv_enc, cur); + cb(Vcur, "Vcur", il); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + ggml_tensor * attn_rel_b = model.layers[il].attn_rel_b_enc ? model.layers[il].attn_rel_b_enc : model.layers[0].attn_rel_b_enc; + ggml_tensor * kq_b = build_pos_bias(pos_bucket_enc, attn_rel_b); + + cur = build_attn(inp_attn, gf, + model.layers[il].wo_enc, nullptr, + Qcur, Kcur, Vcur, kq_b, 1.0f, il); + cb(cur, "kqv_out", il); + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + { + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm_enc, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + // T5 uses relu, flan-T5 uses gelu-gated + cur = build_ffn(cur, + model.layers[il].ffn_up_enc, NULL, NULL, + model.layers[il].ffn_gate_enc, NULL, NULL, + model.layers[il].ffn_down_enc, NULL, NULL, + NULL, + model.layers[il].ffn_gate_enc ? LLM_FFN_GELU : LLM_FFN_RELU, + model.layers[il].ffn_gate_enc ? LLM_FFN_PAR : LLM_FFN_SEQ, + il); + cb(cur, "ffn_out", il); + } + + cur = ggml_add(ctx0, cur, ffn_inp); + cb(cur, "ffn_out", il); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + cb(cur, "result_embd", -1); + + cur = build_norm(cur, + model.output_norm_enc, NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +struct llm_build_t5_dec : public llm_graph_context { + llm_build_t5_dec(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + //const int64_t n_embd_gqa = hparams.n_embd_v_gqa(); + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + ggml_tensor * embd_enc = build_inp_cross_embd(); + ggml_tensor * pos_bucket_dec = build_inp_pos_bucket_dec(); + + const int64_t n_outputs_enc = embd_enc->ne[1]; + + auto * inp_attn_self = build_attn_inp_kv_unified(); + auto * inp_attn_cross = build_attn_inp_cross(); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + // norm + cur = build_norm(inpL, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self-attention + { + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + ggml_tensor * attn_rel_b = model.layers[il].attn_rel_b ? model.layers[il].attn_rel_b : model.layers[0].attn_rel_b; + ggml_tensor * kq_b = build_pos_bias(pos_bucket_dec, attn_rel_b); + + cur = build_attn(inp_attn_self, gf, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, kq_b, 1.0f, il); + cb(cur, "kqv_out", il); + } + + cur = ggml_add(ctx0, cur, inpSA); + cb(cur, "cross_inp", il); + + ggml_tensor * inpCA = cur; + + // norm + cur = build_norm(cur, + model.layers[il].attn_norm_cross, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_norm_cross", il); + + // cross-attention + { + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq_cross, cur); + cb(Qcur, "Qcur", il); + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk_cross, embd_enc); + cb(Kcur, "Kcur", il); + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv_cross, embd_enc); + cb(Vcur, "Vcur", il); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_outputs_enc); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_outputs_enc); + + cur = build_attn(inp_attn_cross, gf, + model.layers[il].wo_cross, nullptr, + Qcur, Kcur, Vcur, nullptr, 1.0f, il); + cb(cur, "kqv_out", il); + + //ggml_tensor * q = ggml_permute(ctx0, Qcur, 0, 2, 1, 3); + //ggml_tensor * k = ggml_cont(ctx0, ggml_permute(ctx0, Kcur, 0, 2, 1, 3)); + + //ggml_tensor * kq = ggml_mul_mat(ctx0, k, q); + //cb(kq, "kq", il); + + //kq = ggml_soft_max_ext(ctx0, kq, KQ_mask_cross, 1.0f, hparams.f_max_alibi_bias); + //cb(kq, "kq_soft_max_ext", il); + + //ggml_tensor * v = ggml_cont(ctx0, ggml_transpose(ctx0, ggml_reshape_2d(ctx0, Vcur, n_embd_gqa, n_outputs_enc))); + //cb(v, "v", il); + + //ggml_tensor * kqv = ggml_mul_mat(ctx0, ggml_reshape_3d(ctx0, v, n_outputs_enc, n_embd_head, n_head_kv), kq); + //cb(kqv, "kqv", il); + + //ggml_tensor * kqv_merged = ggml_permute(ctx0, kqv, 0, 2, 1, 3); + //cb(kqv_merged, "kqv_merged", il); + + //cur = ggml_cont_2d(ctx0, kqv_merged, n_embd_gqa, n_tokens); + //cb(cur, "kqv_merged_cont", il); + + //ggml_build_forward_expand(gf, cur); + + //cur = build_lora_mm(model.layers[il].wo_cross, cur); + //cb(cur, "kqv_out", il); + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + inpCA = ggml_get_rows(ctx0, inpCA, inp_out_ids); + } + + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpCA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + { + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + // T5 uses relu, flan-T5 uses gelu-gated + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + model.layers[il].ffn_gate_enc ? LLM_FFN_GELU : LLM_FFN_RELU, + model.layers[il].ffn_gate_enc ? LLM_FFN_PAR : LLM_FFN_SEQ, + il); + cb(cur, "ffn_out", il); + } + + cur = ggml_add(ctx0, cur, ffn_inp); + cb(cur, "ffn_out", il); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + cb(cur, "result_embd", -1); + + cur = build_norm(cur, + model.output_norm, NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +struct llm_build_jais : public llm_graph_context { + llm_build_jais(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + const int64_t n_embd_gqa = hparams.n_embd_v_gqa(); + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + auto * inp_attn = build_attn_inp_kv_unified(); + + for (int il = 0; il < n_layer; ++il) { + cur = build_norm(inpL, + model.layers[il].attn_norm, + model.layers[il].attn_norm_b, + LLM_NORM, il); + cb(cur, "attn_norm", il); + + // self-attention + { + cur = build_lora_mm(model.layers[il].wqkv, cur); + cb(cur, "wqkv", il); + + cur = ggml_add(ctx0, cur, model.layers[il].bqkv); + cb(cur, "bqkv", il); + + ggml_tensor * Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0*cur->nb[0]*(n_embd))); + ggml_tensor * Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*cur->nb[0]*(n_embd))); + ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*cur->nb[0]*(n_embd + n_embd_gqa))); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + cur = build_attn(inp_attn, gf, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, nullptr, 1.0f/float(n_embd_head), il); + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); + } + + // add the input + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL); + cb(ffn_inp, "ffn_inp", il); + + // FF + { + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, + model.layers[il].ffn_norm_b, + LLM_NORM, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(cur, + model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL, + model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL, + model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + } + + inpL = ggml_add(ctx0, cur, ffn_inp); + cb(inpL, "l_out", il); + } + + cur = build_norm(inpL, + model.output_norm, + model.output_norm_b, + LLM_NORM, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +struct llm_build_chatglm : public llm_graph_context { + llm_build_chatglm(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + const int64_t n_embd_gqa = hparams.n_embd_v_gqa(); + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv_unified(); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + cur = build_norm(inpL, + model.layers[il].attn_norm, + NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self-attention + { + ggml_tensor * Qcur = nullptr; + ggml_tensor * Kcur = nullptr; + ggml_tensor * Vcur = nullptr; + + if (model.layers[il].wqkv == nullptr) { + Qcur = build_lora_mm(model.layers[il].wq, cur); + if (model.layers[il].bq) { + Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); + } + Kcur = build_lora_mm(model.layers[il].wk, cur); + if (model.layers[il].bk) { + Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); + } + Vcur = build_lora_mm(model.layers[il].wv, cur); + if (model.layers[il].bv) { + Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); + } + } else { + cur = build_lora_mm(model.layers[il].wqkv, cur); + cb(cur, "wqkv", il); + if (model.layers[il].bqkv) { + cur = ggml_add(ctx0, cur, model.layers[il].bqkv); + cb(cur, "bqkv", il); + } + Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd))); + Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd))); + Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa))); + } + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + //printf("freq_base: %f freq_scale: %f ext_factor: %f attn_factor: %f\n", freq_base, freq_scale, ext_factor, attn_factor); + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, gf, + model.layers[il].wo, NULL, + Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + + // Add the input + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // FF + { + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, + NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + NULL, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + LLM_FFN_SWIGLU, LLM_FFN_SEQ, il); + cb(cur, "ffn_out", il); + + } + + inpL = ggml_add(ctx0, cur, ffn_inp); + cb(inpL, "l_out", il); + } + + cur = build_norm(inpL, + model.output_norm, + NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +struct llm_build_glm4 : public llm_graph_context { + llm_build_glm4(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + const int64_t n_embd_gqa = hparams.n_embd_v_gqa(); + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv_unified(); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + // Pre-attention norm + cur = build_norm(inpL, + model.layers[il].attn_norm, + NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self-attention + { + ggml_tensor * Qcur = nullptr; + ggml_tensor * Kcur = nullptr; + ggml_tensor * Vcur = nullptr; + + if (model.layers[il].wqkv == nullptr) { + Qcur = build_lora_mm(model.layers[il].wq, cur); + if (model.layers[il].bq) { + Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); + } + Kcur = build_lora_mm(model.layers[il].wk, cur); + if (model.layers[il].bk) { + Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); + } + Vcur = build_lora_mm(model.layers[il].wv, cur); + if (model.layers[il].bv) { + Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); + } + } else { + cur = build_lora_mm(model.layers[il].wqkv, cur); + cb(cur, "wqkv", il); + if (model.layers[il].bqkv) { + cur = ggml_add(ctx0, cur, model.layers[il].bqkv); + cb(cur, "bqkv", il); + } + Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd))); + Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd))); + Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa))); + } + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, gf, + model.layers[il].wo, NULL, + Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + + // Post-attention norm (new!) + cur = build_norm(cur, + model.layers[il].attn_post_norm, + NULL, + LLM_NORM_RMS, il); + cb(cur, "post_attn_norm", il); + + // Add the input (residual connection after post-attention norm) + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // FF + { + // Pre-MLP norm + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, + NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + // MLP + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + NULL, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + LLM_FFN_SWIGLU, LLM_FFN_SEQ, il); + cb(cur, "ffn_out", il); + + // Post-MLP norm + cur = build_norm(cur, + model.layers[il].ffn_post_norm, + NULL, + LLM_NORM_RMS, il); + cb(cur, "post_mlp_norm", il); + } + + // Add residual connection after post-MLP norm + inpL = ggml_add(ctx0, cur, ffn_inp); + cb(inpL, "l_out", il); + } + + // Final norm + cur = build_norm(inpL, + model.output_norm, + NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // Output projection + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +struct llm_build_nemotron : public llm_graph_context { + llm_build_nemotron(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + //GGML_ASSERT(n_embd_head == hparams.n_rot); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv_unified(); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + // norm + cur = build_norm(inpL, + model.layers[il].attn_norm, + model.layers[il].attn_norm_b, + LLM_NORM, il); + cb(cur, "attn_norm", il); + + // self-attention + { + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + if (model.layers[il].bq) { + Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); + cb(Qcur, "Qcur", il); + } + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + if (model.layers[il].bk) { + Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); + cb(Kcur, "Kcur", il); + } + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + if (model.layers[il].bv) { + Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); + cb(Vcur, "Vcur", il); + } + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, gf, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, + model.layers[il].ffn_norm_b, + LLM_NORM, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(cur, + model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL, + NULL, NULL, NULL, + model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, + NULL, + LLM_FFN_RELU_SQR, LLM_FFN_SEQ, il); + + cur = ggml_add(ctx0, cur, ffn_inp); + cb(cur, "ffn_out", il); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = build_norm(cur, + model.output_norm, model.output_norm_b, + LLM_NORM, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +struct llm_build_exaone : public llm_graph_context { + llm_build_exaone(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv_unified(); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + // norm + cur = build_norm(inpL, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self-attention + { + // rope freq factors for llama3; may return nullptr for llama2 and other models + ggml_tensor * rope_factors = static_cast(memory)->cbs.get_rope_factors(n_ctx_per_seq, il); + + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + if (model.layers[il].bq) { + Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); + cb(Qcur, "Qcur", il); + } + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + if (model.layers[il].bk) { + Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); + cb(Kcur, "Kcur", il); + } + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + if (model.layers[il].bv) { + Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); + cb(Vcur, "Vcur", il); + } + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, rope_factors, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, rope_factors, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, gf, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + + cur = ggml_add(ctx0, cur, ffn_inp); + cb(cur, "ffn_out", il); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = build_norm(cur, + model.output_norm, NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +struct llm_build_rwkv6_base : public llm_graph_context { + const llama_model & model; + + llm_build_rwkv6_base(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params), model(model) { + } + + ggml_tensor * build_rwkv6_channel_mix( + const llama_layer * layer, + ggml_tensor * cur, + ggml_tensor * x_prev, + llm_arch arch) const { + ggml_tensor * sx = ggml_sub(ctx0, x_prev, cur); + switch (arch) { + case LLM_ARCH_RWKV6: + { + ggml_tensor * xk = ggml_add(ctx0, ggml_mul(ctx0, sx, layer->channel_mix_lerp_k), cur); + ggml_tensor * xr = ggml_add(ctx0, ggml_mul(ctx0, sx, layer->channel_mix_lerp_r), cur); + + ggml_tensor * r = ggml_sigmoid(ctx0, build_lora_mm(layer->channel_mix_receptance, xr)); + ggml_tensor * k = ggml_sqr( + ctx0, + ggml_relu( + ctx0, + build_lora_mm(layer->channel_mix_key, xk) + ) + ); + cur = ggml_mul(ctx0, r, build_lora_mm(layer->channel_mix_value, k)); + } break; + default: + GGML_ABORT("fatal error"); + } + + return cur; + } + + ggml_tensor * build_rwkv6_time_mix( + ggml_cgraph * gf, + ggml_tensor * cur, + ggml_tensor * x_prev, + ggml_tensor * state_copy, + ggml_tensor * state_mask, + const llama_ubatch & ubatch, + int il) const { + const llama_kv_cache_unified * kv_self = static_cast(memory); + + const auto n_tokens = ubatch.n_tokens; + const auto n_seqs = ubatch.n_seqs; + const auto n_seq_tokens = ubatch.n_seq_tokens; + const auto n_embd = hparams.n_embd; + const auto head_size = hparams.wkv_head_size; + const auto n_head = n_embd / head_size; + const auto n_head_kv = hparams.n_head_kv(il); + + const auto kv_head = kv_self->head; + + const auto & layer = model.layers[il]; + + bool is_qrwkv = layer.time_mix_first == nullptr; + + ggml_tensor * sx = ggml_sub(ctx0, x_prev, cur); + + sx = ggml_reshape_2d(ctx0, sx, n_embd, n_tokens); + cur = ggml_reshape_2d(ctx0, cur, n_embd, n_tokens); + + ggml_tensor * xxx = ggml_add(ctx0, ggml_mul(ctx0, sx, layer.time_mix_lerp_x), cur); + + xxx = ggml_reshape_4d( + ctx0, + ggml_tanh( + ctx0, + ggml_mul_mat(ctx0, layer.time_mix_w1, xxx) + ), + layer.time_mix_w1->ne[1] / 5, 1, 5, n_tokens + ); + + xxx = ggml_cont(ctx0, ggml_permute(ctx0, xxx, 0, 1, 3, 2)); + + xxx = ggml_mul_mat( + ctx0, + ggml_reshape_4d( + ctx0, + layer.time_mix_w2, + layer.time_mix_w2->ne[0], layer.time_mix_w2->ne[1], 1, 5 + ), + xxx + ); + + ggml_tensor *xw, *xk, *xv, *xr, *xg; + if (layer.time_mix_lerp_fused) { + // fusing these weights makes some performance improvement + sx = ggml_reshape_3d(ctx0, sx, n_embd, 1, n_tokens); + cur = ggml_reshape_3d(ctx0, cur, n_embd, 1, n_tokens); + xxx = ggml_add(ctx0, ggml_mul(ctx0, ggml_add(ctx0, xxx, layer.time_mix_lerp_fused), sx), cur); + xw = ggml_view_2d(ctx0, xxx, n_embd, n_tokens, xxx->nb[1], 0); + xk = ggml_view_2d(ctx0, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * sizeof(float)); + xv = ggml_view_2d(ctx0, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * 2 * sizeof(float)); + xr = ggml_view_2d(ctx0, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * 3 * sizeof(float)); + xg = ggml_view_2d(ctx0, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * 4 * sizeof(float)); + } else { + // for backward compatibility + xw = ggml_view_2d(ctx0, xxx, n_embd, n_tokens, xxx->nb[1], 0); + xk = ggml_view_2d(ctx0, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * sizeof(float)); + xv = ggml_view_2d(ctx0, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * 2 * sizeof(float)); + xr = ggml_view_2d(ctx0, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * 3 * sizeof(float)); + xg = ggml_view_2d(ctx0, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * 4 * sizeof(float)); + + xw = ggml_add(ctx0, ggml_mul(ctx0, ggml_add(ctx0, xw, layer.time_mix_lerp_w), sx), cur); + xk = ggml_add(ctx0, ggml_mul(ctx0, ggml_add(ctx0, xk, layer.time_mix_lerp_k), sx), cur); + xv = ggml_add(ctx0, ggml_mul(ctx0, ggml_add(ctx0, xv, layer.time_mix_lerp_v), sx), cur); + xr = ggml_add(ctx0, ggml_mul(ctx0, ggml_add(ctx0, xr, layer.time_mix_lerp_r), sx), cur); + xg = ggml_add(ctx0, ggml_mul(ctx0, ggml_add(ctx0, xg, layer.time_mix_lerp_g), sx), cur); + } + + ggml_tensor * r = build_lora_mm(layer.time_mix_receptance, xr); + ggml_tensor * k = build_lora_mm(layer.time_mix_key, xk); + ggml_tensor * v = build_lora_mm(layer.time_mix_value, xv); + if (layer.time_mix_receptance_b) { + r = ggml_add(ctx0, r, layer.time_mix_receptance_b); + } + if (layer.time_mix_key_b) { + k = ggml_add(ctx0, k, layer.time_mix_key_b); + } + if (layer.time_mix_value_b) { + v = ggml_add(ctx0, v, layer.time_mix_value_b); + } + + ggml_tensor * g = build_lora_mm(layer.time_mix_gate, xg); + if (is_qrwkv) { + g = ggml_sigmoid(ctx0, g); + } else { + g = ggml_silu(ctx0, g); + } + + if (n_head_kv != 0 && n_head_kv != n_head) { + GGML_ASSERT(n_head % n_head_kv == 0); + k = ggml_reshape_4d(ctx0, k, head_size, 1, n_head_kv, n_tokens); + v = ggml_reshape_4d(ctx0, v, head_size, 1, n_head_kv, n_tokens); + ggml_tensor * tmp = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, head_size, n_head / n_head_kv, n_head_kv, n_tokens); + k = ggml_repeat(ctx0, k, tmp); + v = ggml_repeat(ctx0, v, tmp); + } + + k = ggml_reshape_3d(ctx0, k, head_size, n_head, n_tokens); + v = ggml_reshape_3d(ctx0, v, head_size, n_head, n_tokens); + r = ggml_reshape_3d(ctx0, r, head_size, n_head, n_tokens); + + ggml_tensor * w = ggml_mul_mat( + ctx0, + layer.time_mix_decay_w2, + ggml_tanh( + ctx0, + ggml_mul_mat(ctx0, layer.time_mix_decay_w1, xw) + ) + ); + + w = ggml_add(ctx0, w, layer.time_mix_decay); + w = ggml_exp(ctx0, ggml_neg(ctx0, ggml_exp(ctx0, w))); + w = ggml_reshape_3d(ctx0, w, head_size, n_head, n_tokens); + + if (is_qrwkv) { + // k = k * (1 - w) + k = ggml_sub(ctx0, k, ggml_mul(ctx0, k, w)); + } + + ggml_tensor * wkv_state = build_copy_mask_state( + gf, kv_self->v_l[il], state_copy, state_mask, + hparams.n_embd_v_s(), n_seqs); + + ggml_tensor * wkv_output; + if (is_qrwkv) { + wkv_output = ggml_gated_linear_attn(ctx0, k, v, r, w, wkv_state, pow(head_size, -0.5f)); + } else { + wkv_output = ggml_rwkv_wkv6(ctx0, k, v, r, layer.time_mix_first, w, wkv_state); + } + cur = ggml_view_1d(ctx0, wkv_output, n_embd * n_tokens, 0); + wkv_state = ggml_view_1d(ctx0, wkv_output, n_embd * head_size * n_seqs, n_embd * n_tokens * sizeof(float)); + + ggml_build_forward_expand( + gf, + ggml_cpy( + ctx0, + wkv_state, + ggml_view_1d( + ctx0, + kv_self->v_l[il], + hparams.n_embd_v_s() * n_seqs, + hparams.n_embd_v_s() * kv_head * ggml_element_size(kv_self->v_l[il]) + ) + ) + ); + + if (!is_qrwkv) { + // group norm with head_count groups + cur = ggml_reshape_3d(ctx0, cur, n_embd / n_head, n_head, n_tokens); + cur = ggml_norm(ctx0, cur, 64e-5f); + + // Convert back to regular vectors. + cur = ggml_reshape_2d(ctx0, cur, n_embd, n_tokens); + cur = ggml_add(ctx0, ggml_mul(ctx0, cur, layer.time_mix_ln), layer.time_mix_ln_b); + } else { + cur = ggml_reshape_2d(ctx0, cur, n_embd, n_tokens); + } + + cur = ggml_mul(ctx0, cur, g); + cur = build_lora_mm(layer.time_mix_output, cur); + + return ggml_reshape_3d(ctx0, cur, n_embd, n_seq_tokens, n_seqs); + } +}; + +struct llm_build_rwkv6 : public llm_build_rwkv6_base { + llm_build_rwkv6(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_build_rwkv6_base(model, params) { + GGML_ASSERT(hparams.token_shift_count == 2); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + inpL = build_norm(inpL, model.tok_norm, model.tok_norm_b, LLM_NORM, -1); + + ggml_tensor * state_copy = build_inp_s_copy(); + ggml_tensor * state_mask = build_inp_s_mask(); + + const auto n_embd = hparams.n_embd; + const auto n_seq_tokens = ubatch.n_seq_tokens; + const auto n_seqs = ubatch.n_seqs; + + for (int il = 0; il < n_layer; ++il) { + const llama_layer * layer = &model.layers[il]; + inpL = ggml_reshape_3d(ctx0, inpL, n_embd, n_seq_tokens, n_seqs); + + ggml_tensor * token_shift = build_rwkv_token_shift_load( + gf, state_copy, state_mask, ubatch, il + ); + + ggml_tensor * att_shift = ggml_view_3d(ctx0, token_shift, n_embd, 1, n_seqs, token_shift->nb[1], token_shift->nb[2], 0); + ggml_tensor * ffn_shift = ggml_view_3d(ctx0, token_shift, n_embd, 1, n_seqs, token_shift->nb[1], token_shift->nb[2], n_embd * ggml_element_size(token_shift)); + + ggml_tensor * att_norm = build_norm(inpL, layer->attn_norm, layer->attn_norm_b, LLM_NORM, il); + cb(att_norm, "attn_norm", il); + + ggml_tensor * x_prev = ggml_concat( + ctx0, + att_shift, + ggml_view_3d(ctx0, att_norm, n_embd, n_seq_tokens - 1, n_seqs, att_norm->nb[1], att_norm->nb[2], 0), + 1 + ); + + cur = build_rwkv6_time_mix(gf, att_norm, x_prev, state_copy, state_mask, ubatch, il); + + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL); + cb(ffn_inp, "ffn_inp", il); + + ggml_tensor * ffn_norm = build_norm(ffn_inp, layer->attn_norm_2, layer->attn_norm_2_b, LLM_NORM, il); + cb(ffn_norm, "ffn_norm", il); + + x_prev = ggml_concat( + ctx0, + ffn_shift, + ggml_view_3d(ctx0, ffn_norm, n_embd, n_seq_tokens - 1, n_seqs, ffn_norm->nb[1], ffn_norm->nb[2], 0), + 1 + ); + + token_shift = ggml_concat(ctx0, + ggml_view_3d(ctx0, att_norm, n_embd, 1, n_seqs, att_norm->nb[1], att_norm->nb[2], (n_seq_tokens-1)*n_embd*ggml_element_size(att_norm)), + ggml_view_3d(ctx0, ffn_norm, n_embd, 1, n_seqs, ffn_norm->nb[1], ffn_norm->nb[2], (n_seq_tokens-1)*n_embd*ggml_element_size(ffn_norm)), + 1 + ); + ggml_build_forward_expand(gf, build_rwkv_token_shift_store(token_shift, ubatch, il)); + + if (il == n_layer - 1) { + // skip computing output for unused tokens + struct ggml_tensor * inp_out_ids = build_inp_out_ids(); + ffn_inp = ggml_get_rows(ctx0, ggml_reshape_2d(ctx0, ffn_inp, n_embd, n_tokens), inp_out_ids); + ffn_norm = ggml_get_rows(ctx0, ggml_reshape_2d(ctx0, ffn_norm, n_embd, n_tokens), inp_out_ids); + x_prev = ggml_get_rows(ctx0, ggml_reshape_2d(ctx0, x_prev, n_embd, n_tokens), inp_out_ids); + cur = ggml_get_rows(ctx0, ggml_reshape_2d(ctx0, cur, n_embd, n_tokens), inp_out_ids); + } + + cur = build_rwkv6_channel_mix(layer, ffn_norm, x_prev, LLM_ARCH_RWKV6); + cur = ggml_add(ctx0, cur, ffn_inp); + + if (hparams.rescale_every_n_layers != 0 && (il + 1) % hparams.rescale_every_n_layers == 0) { + cur = ggml_scale(ctx0, cur, 0.5F); + } + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + cur = build_norm(cur, model.output_norm, model.output_norm_b, LLM_NORM, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +// ref: https://huggingface.co/recursal/QRWKV6-32B-Instruct-Preview-v0.1/blob/main/modeling_rwkv6qwen2.py +struct llm_build_rwkv6qwen2 : public llm_build_rwkv6_base { + llm_build_rwkv6qwen2(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_build_rwkv6_base(model, params) { + GGML_ASSERT(n_embd == hparams.n_embd_k_s()); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + ggml_tensor * state_copy = build_inp_s_copy(); + ggml_tensor * state_mask = build_inp_s_mask(); + + const auto n_embd = hparams.n_embd; + const auto n_seq_tokens = ubatch.n_seq_tokens; + const auto n_seqs = ubatch.n_seqs; + + for (int il = 0; il < n_layer; ++il) { + const llama_layer * layer = &model.layers[il]; + inpL = ggml_reshape_3d(ctx0, inpL, n_embd, n_seq_tokens, n_seqs); + + ggml_tensor * token_shift = build_rwkv_token_shift_load( + gf, state_copy, state_mask, ubatch, il + ); + + ggml_tensor * att_norm = build_norm(inpL, layer->attn_norm, layer->attn_norm_b, LLM_NORM_RMS, il); + cb(att_norm, "attn_norm", il); + + ggml_tensor * x_prev = ggml_concat( + ctx0, + token_shift, + ggml_view_3d(ctx0, att_norm, n_embd, n_seq_tokens - 1, n_seqs, att_norm->nb[1], att_norm->nb[2], 0), + 1 + ); + + cur = build_rwkv6_time_mix(gf, att_norm, x_prev, state_copy, state_mask, ubatch, il); + + token_shift = ggml_view_3d(ctx0, att_norm, n_embd, 1, n_seqs, att_norm->nb[1], att_norm->nb[2], (n_seq_tokens-1)*n_embd*ggml_element_size(att_norm)); + ggml_build_forward_expand(gf, build_rwkv_token_shift_store(token_shift, ubatch, il)); + + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL); + cb(ffn_inp, "ffn_inp", il); + + if (il == n_layer - 1) { + // skip computing output for unused tokens + struct ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, ggml_reshape_2d(ctx0, cur, n_embd, n_tokens), inp_out_ids); + ffn_inp = ggml_get_rows(ctx0, ggml_reshape_2d(ctx0, ffn_inp, n_embd, n_tokens), inp_out_ids); + } + + // feed-forward network + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + + cur = ggml_add(ctx0, cur, ffn_inp); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + cur = build_norm(cur, model.output_norm, model.output_norm_b, LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +struct llm_build_rwkv7_base : public llm_graph_context { + const llama_model & model; + + llm_build_rwkv7_base(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params), model(model) { + } + + ggml_tensor * build_rwkv7_channel_mix( + const llama_layer * layer, + ggml_tensor * cur, + ggml_tensor * x_prev, + llm_arch arch) const { + ggml_tensor * sx = ggml_sub(ctx0, x_prev, cur); + switch (arch) { + case LLM_ARCH_RWKV7: + { + ggml_tensor * xk = ggml_add(ctx0, ggml_mul(ctx0, sx, layer->channel_mix_lerp_k), cur); + + ggml_tensor * k = ggml_sqr( + ctx0, + ggml_relu( + ctx0, + build_lora_mm(layer->channel_mix_key, xk) + ) + ); + + cur = build_lora_mm(layer->channel_mix_value, k); + } break; + default: + GGML_ABORT("fatal error"); + } + + return cur; + } + + ggml_tensor * build_rwkv7_time_mix( + ggml_cgraph * gf, + ggml_tensor * cur, + ggml_tensor * x_prev, + ggml_tensor * state_copy, + ggml_tensor * state_mask, + ggml_tensor *& first_layer_value, + const llama_ubatch & ubatch, + int il) const { + const llama_kv_cache_unified * kv_self = static_cast(memory); + + const auto n_tokens = ubatch.n_tokens; + const auto n_seqs = ubatch.n_seqs; + const auto n_embd = hparams.n_embd; + const auto head_size = hparams.wkv_head_size; + const auto head_count = n_embd / head_size; + const auto n_seq_tokens = ubatch.n_seq_tokens; + + const auto kv_head = kv_self->head; + + const auto & layer = model.layers[il]; + + bool has_gating = layer.time_mix_g1 && layer.time_mix_g2; + + ggml_tensor * sx = ggml_sub(ctx0, x_prev, cur); + ggml_tensor * dummy = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_embd, n_seq_tokens, n_seqs, has_gating ? 6 : 5); + sx = ggml_repeat(ctx0, sx, dummy); + + ggml_tensor * xxx = ggml_add(ctx0, ggml_mul(ctx0, sx, layer.time_mix_lerp_fused), cur); + + ggml_tensor * xr = ggml_view_2d(ctx0, xxx, n_embd, n_tokens, xxx->nb[1], 0); + ggml_tensor * xw = ggml_view_2d(ctx0, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * sizeof(float)); + ggml_tensor * xk = ggml_view_2d(ctx0, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * 2 * sizeof(float)); + ggml_tensor * xv = ggml_view_2d(ctx0, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * 3 * sizeof(float)); + ggml_tensor * xa = ggml_view_2d(ctx0, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * 4 * sizeof(float)); + ggml_tensor * xg = has_gating ? ggml_view_2d(ctx0, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * 5 * sizeof(float)) : nullptr; + + ggml_tensor * r = build_lora_mm(layer.time_mix_receptance, xr); + ggml_tensor * w = ggml_add( + ctx0, + ggml_mul_mat(ctx0, layer.time_mix_w2, ggml_tanh(ctx0, ggml_mul_mat(ctx0, layer.time_mix_w1, xw))), + layer.time_mix_w0 + ); + w = ggml_exp(ctx0, ggml_scale(ctx0, ggml_sigmoid(ctx0, w), -0.606531)); + + ggml_tensor * k = build_lora_mm(layer.time_mix_key, xk); + ggml_tensor * v = build_lora_mm(layer.time_mix_value, xv); + if (first_layer_value == nullptr) { + first_layer_value = v; + } else { + // Add the first layer value as a residual connection. + v = ggml_add(ctx0, v, + ggml_mul(ctx0, + ggml_sub(ctx0, first_layer_value, v), + ggml_sigmoid(ctx0, ggml_add(ctx0, + ggml_mul_mat(ctx0, layer.time_mix_v2, ggml_mul_mat(ctx0, layer.time_mix_v1, xv)), + layer.time_mix_v0 + ) + ) + ) + ); + } + + ggml_tensor * g = nullptr; + if (layer.time_mix_g1 && layer.time_mix_g2) { + g = ggml_mul_mat(ctx0, layer.time_mix_g2, ggml_sigmoid(ctx0, ggml_mul_mat(ctx0, layer.time_mix_g1, xg))); + } + + ggml_tensor * a = ggml_sigmoid(ctx0, + ggml_add( + ctx0, + ggml_mul_mat(ctx0, layer.time_mix_a2, ggml_mul_mat(ctx0, layer.time_mix_a1, xa)), + layer.time_mix_a0 + ) + ); + + ggml_tensor * kk = ggml_reshape_3d(ctx0, ggml_mul(ctx0, k, layer.time_mix_k_k), head_size, head_count, n_tokens); + kk = ggml_l2_norm(ctx0, kk, 1e-12); + + ggml_tensor * ka = ggml_mul(ctx0, k, layer.time_mix_k_a); + k = ggml_add(ctx0, k, ggml_sub(ctx0, ggml_mul(ctx0, a, ka), ka)); + + r = ggml_reshape_3d(ctx0, r, head_size, head_count, n_tokens); + w = ggml_reshape_3d(ctx0, w, head_size, head_count, n_tokens); + k = ggml_reshape_3d(ctx0, k, head_size, head_count, n_tokens); + v = ggml_reshape_3d(ctx0, v, head_size, head_count, n_tokens); + a = ggml_reshape_3d(ctx0, a, head_size, head_count, n_tokens); + + ggml_tensor * wkv_state = build_copy_mask_state( + gf, kv_self->v_l[il], state_copy, state_mask, + hparams.n_embd_v_s(), n_seqs); + + ggml_tensor * wkv_output = ggml_rwkv_wkv7(ctx0, r, w, k, v, ggml_neg(ctx0, kk), ggml_mul(ctx0, kk, a), wkv_state); + cur = ggml_view_1d(ctx0, wkv_output, n_embd * n_tokens, 0); + wkv_state = ggml_view_1d(ctx0, wkv_output, n_embd * head_size * n_seqs, n_embd * n_tokens * sizeof(float)); + + ggml_build_forward_expand( + gf, + ggml_cpy( + ctx0, + wkv_state, + ggml_view_1d( + ctx0, + kv_self->v_l[il], + hparams.n_embd_v_s() * n_seqs, + hparams.n_embd_v_s() * kv_head * ggml_element_size(kv_self->v_l[il]) + ) + ) + ); + + if (layer.time_mix_ln && layer.time_mix_ln_b) { + // group norm with head_count groups + cur = ggml_reshape_3d(ctx0, cur, n_embd / head_count, head_count, n_tokens); + cur = ggml_norm(ctx0, cur, 64e-5f); + + // Convert back to regular vectors. + cur = ggml_reshape_2d(ctx0, cur, n_embd, n_tokens); + cur = ggml_add(ctx0, ggml_mul(ctx0, cur, layer.time_mix_ln), layer.time_mix_ln_b); + } else { + cur = ggml_reshape_2d(ctx0, cur, n_embd, n_tokens); + } + + ggml_tensor * rk = ggml_sum_rows(ctx0, + ggml_mul(ctx0, ggml_mul(ctx0, k, r), ggml_reshape_2d(ctx0, layer.time_mix_r_k, head_size, head_count))); + cur = ggml_add(ctx0, cur, ggml_reshape_2d(ctx0, ggml_mul(ctx0, v, rk), n_embd, n_tokens)); + + if (has_gating) { + cur = ggml_mul(ctx0, cur, g); + } + cur = build_lora_mm(layer.time_mix_output, cur); + + return ggml_reshape_3d(ctx0, cur, n_embd, n_seq_tokens, n_seqs); + } +}; + +struct llm_build_rwkv7 : public llm_build_rwkv7_base { + llm_build_rwkv7(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_build_rwkv7_base(model, params) { + GGML_ASSERT(hparams.token_shift_count == 2); + + ggml_tensor * cur; + ggml_tensor * inpL; + ggml_tensor * v_first = nullptr; + + inpL = build_inp_embd(model.tok_embd); + inpL = build_norm(inpL, model.tok_norm, model.tok_norm_b, LLM_NORM, -1); + + ggml_tensor * state_copy = build_inp_s_copy(); + ggml_tensor * state_mask = build_inp_s_mask(); + + const auto n_embd = hparams.n_embd; + const auto n_seq_tokens = ubatch.n_seq_tokens; + const auto n_seqs = ubatch.n_seqs; + + for (int il = 0; il < n_layer; ++il) { + const llama_layer * layer = &model.layers[il]; + inpL = ggml_reshape_3d(ctx0, inpL, n_embd, n_seq_tokens, n_seqs); + + ggml_tensor * token_shift = build_rwkv_token_shift_load( + gf, state_copy, state_mask, ubatch, il + ); + + ggml_tensor * att_shift = ggml_view_3d(ctx0, token_shift, n_embd, 1, n_seqs, token_shift->nb[1], token_shift->nb[2], 0); + ggml_tensor * ffn_shift = ggml_view_3d(ctx0, token_shift, n_embd, 1, n_seqs, token_shift->nb[1], token_shift->nb[2], n_embd * ggml_element_size(token_shift)); + + ggml_tensor * att_norm = build_norm(inpL, layer->attn_norm, layer->attn_norm_b, LLM_NORM, il); + cb(att_norm, "attn_norm", il); + + ggml_tensor * x_prev = ggml_concat( + ctx0, + att_shift, + ggml_view_3d(ctx0, att_norm, n_embd, n_seq_tokens - 1, n_seqs, att_norm->nb[1], att_norm->nb[2], 0), + 1 + ); + + cur = build_rwkv7_time_mix(gf, att_norm, x_prev, state_copy, state_mask, v_first, ubatch, il); + + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL); + cb(ffn_inp, "ffn_inp", il); + + ggml_tensor * ffn_norm = build_norm(ffn_inp, layer->attn_norm_2, layer->attn_norm_2_b, LLM_NORM, il); + cb(ffn_norm, "ffn_norm", il); + + x_prev = ggml_concat( + ctx0, + ffn_shift, + ggml_view_3d(ctx0, ffn_norm, n_embd, n_seq_tokens - 1, n_seqs, ffn_norm->nb[1], ffn_norm->nb[2], 0), + 1 + ); + + token_shift = ggml_concat(ctx0, + ggml_view_3d(ctx0, att_norm, n_embd, 1, n_seqs, att_norm->nb[1], att_norm->nb[2], (n_seq_tokens-1)*n_embd*ggml_element_size(att_norm)), + ggml_view_3d(ctx0, ffn_norm, n_embd, 1, n_seqs, ffn_norm->nb[1], ffn_norm->nb[2], (n_seq_tokens-1)*n_embd*ggml_element_size(ffn_norm)), + 1 + ); + ggml_build_forward_expand(gf, build_rwkv_token_shift_store(token_shift, ubatch, il)); + + if (il == n_layer - 1) { + // skip computing output for unused tokens + struct ggml_tensor * inp_out_ids = build_inp_out_ids(); + ffn_inp = ggml_get_rows(ctx0, ggml_reshape_2d(ctx0, ffn_inp, n_embd, n_tokens), inp_out_ids); + ffn_norm = ggml_get_rows(ctx0, ggml_reshape_2d(ctx0, ffn_norm, n_embd, n_tokens), inp_out_ids); + x_prev = ggml_get_rows(ctx0, ggml_reshape_2d(ctx0, x_prev, n_embd, n_tokens), inp_out_ids); + } + + cur = build_rwkv7_channel_mix(layer, ffn_norm, x_prev, LLM_ARCH_RWKV7); + cur = ggml_add(ctx0, cur, ffn_inp); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + cur = build_norm(cur, model.output_norm, model.output_norm_b, LLM_NORM, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + + +struct llm_build_arwkv7 : public llm_build_rwkv7_base { + llm_build_arwkv7(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_build_rwkv7_base(model, params) { + GGML_ASSERT(n_embd == hparams.n_embd_k_s()); + + ggml_tensor * cur; + ggml_tensor * inpL; + ggml_tensor * v_first = nullptr; + + inpL = build_inp_embd(model.tok_embd); + + ggml_tensor * state_copy = build_inp_s_copy(); + ggml_tensor * state_mask = build_inp_s_mask(); + + const auto n_embd = hparams.n_embd; + const auto n_seq_tokens = ubatch.n_seq_tokens; + const auto n_seqs = ubatch.n_seqs; + + for (int il = 0; il < n_layer; ++il) { + const llama_layer * layer = &model.layers[il]; + inpL = ggml_reshape_3d(ctx0, inpL, n_embd, n_seq_tokens, n_seqs); + + ggml_tensor * token_shift = build_rwkv_token_shift_load( + gf, state_copy, state_mask, ubatch, il + ); + + ggml_tensor * att_norm = build_norm(inpL, layer->attn_norm, layer->attn_norm_b, LLM_NORM_RMS, il); + cb(att_norm, "attn_norm", il); + + ggml_tensor * x_prev = ggml_concat( + ctx0, + token_shift, + ggml_view_3d(ctx0, att_norm, n_embd, n_seq_tokens - 1, n_seqs, att_norm->nb[1], att_norm->nb[2], 0), + 1 + ); + + cur = build_rwkv7_time_mix(gf, att_norm, x_prev, state_copy, state_mask, v_first, ubatch, il); + + token_shift = ggml_view_3d(ctx0, att_norm, n_embd, 1, n_seqs, att_norm->nb[1], att_norm->nb[2], (n_seq_tokens-1)*n_embd*ggml_element_size(att_norm)); + ggml_build_forward_expand(gf, build_rwkv_token_shift_store(token_shift, ubatch, il)); + + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL); + cb(ffn_inp, "ffn_inp", il); + + if (il == n_layer - 1) { + // skip computing output for unused tokens + struct ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, ggml_reshape_2d(ctx0, cur, n_embd, n_tokens), inp_out_ids); + ffn_inp = ggml_get_rows(ctx0, ggml_reshape_2d(ctx0, ffn_inp, n_embd, n_tokens), inp_out_ids); + } + + // feed-forward network + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + + cur = ggml_add(ctx0, cur, ffn_inp); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + cur = build_norm(cur, model.output_norm, model.output_norm_b, LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +// ref: https://github.com/facebookresearch/chameleon +// based on the original build_llama() function, changes: +// * qk-norm +// * swin-norm +// * removed bias +// * removed MoE +struct llm_build_chameleon : public llm_graph_context { + llm_build_chameleon(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv_unified(); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + // norm + if (hparams.swin_norm) { + cur = inpL; + } else { + cur = build_norm(inpL, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + } + + // self-attention + { + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + + if (model.layers[il].attn_q_norm) { + Qcur = ggml_view_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens, + ggml_element_size(Qcur) * n_embd_head, + ggml_element_size(Qcur) * n_embd_head * n_head, + 0); + cb(Qcur, "Qcur", il); + + Qcur = build_norm(Qcur, + model.layers[il].attn_q_norm, + model.layers[il].attn_q_norm_b, + LLM_NORM, il); + cb(Qcur, "Qcur", il); + } + + if (model.layers[il].attn_k_norm) { + Kcur = ggml_view_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens, + ggml_element_size(Kcur) * n_embd_head, + ggml_element_size(Kcur) * n_embd_head * n_head_kv, + 0); + cb(Kcur, "Kcur", il); + + Kcur = build_norm(Kcur, + model.layers[il].attn_k_norm, + model.layers[il].attn_k_norm_b, + LLM_NORM, il); + cb(Kcur, "Kcur", il); + } + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, gf, + model.layers[il].wo, nullptr, + Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + + if (hparams.swin_norm) { + cur = build_norm(cur, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, il); + } + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + if (!hparams.swin_norm) { + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + } + + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + + if (hparams.swin_norm) { + cur = build_norm(cur, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + } + + cur = ggml_add(ctx0, cur, ffn_inp); + cb(cur, "ffn_out", il); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = build_norm(cur, + model.output_norm, NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + cb(cur, "result_output_with_img_logits", -1); + + // TODO: this suppresses the output of image tokens, which is required to enable text-only outputs. + // Needs to be removed once image outputs are supported. + int img_token_end_idx = 8196; + int img_token_start_idx = 4; + int num_img_tokens = img_token_end_idx - img_token_start_idx; + // creates 1d tensor of size num_img_tokens and values -FLT_MAX, + // which ensures that text token values are always at least larger than image token values + ggml_tensor * img_logits = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, num_img_tokens); + img_logits = ggml_clamp(ctx0, img_logits, -FLT_MAX, -FLT_MAX); + cb(img_logits, "img_logits", -1); + + cur = ggml_set_1d(ctx0, cur, img_logits, ggml_element_size(cur) * img_token_start_idx); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +struct llm_build_wavtokenizer_dec : public llm_graph_context { + llm_build_wavtokenizer_dec(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + cur = ggml_cont(ctx0, ggml_transpose(ctx0, inpL)); + + cur = ggml_conv_1d_ph(ctx0, model.conv1d, cur, 1, 1); + cur = ggml_add(ctx0, cur, model.conv1d_b); + + // posnet + for (uint32_t il = 0; il < hparams.posnet.n_layer; ++il) { + const auto & layer = model.layers[il].posnet; + + inpL = cur; + + switch (il) { + case 0: + case 1: + case 3: + case 4: + { + cur = build_norm(cur, + layer.norm1, + layer.norm1_b, + LLM_NORM_GROUP, 0); + + cur = ggml_mul(ctx0, ggml_sigmoid(ctx0, cur), cur); + + cur = ggml_conv_1d_ph(ctx0, layer.conv1, cur, 1, 1); + cur = ggml_add(ctx0, cur, layer.conv1_b); + + cur = build_norm(cur, + layer.norm2, + layer.norm2_b, + LLM_NORM_GROUP, 0); + + cur = ggml_mul(ctx0, ggml_sigmoid(ctx0, cur), cur); + + cur = ggml_conv_1d_ph(ctx0, layer.conv2, cur, 1, 1); + cur = ggml_add(ctx0, cur, layer.conv2_b); + + cur = ggml_add(ctx0, cur, inpL); + } break; + case 2: + { + cur = build_norm(cur, + layer.attn_norm, + layer.attn_norm_b, + LLM_NORM_GROUP, 0); + + ggml_tensor * q; + ggml_tensor * k; + ggml_tensor * v; + + q = ggml_conv_1d_ph(ctx0, layer.attn_q, cur, 1, 1); + k = ggml_conv_1d_ph(ctx0, layer.attn_k, cur, 1, 1); + v = ggml_conv_1d_ph(ctx0, layer.attn_v, cur, 1, 1); + + q = ggml_add(ctx0, q, layer.attn_q_b); + k = ggml_add(ctx0, k, layer.attn_k_b); + v = ggml_add(ctx0, v, layer.attn_v_b); + + q = ggml_cont(ctx0, ggml_transpose(ctx0, q)); + k = ggml_cont(ctx0, ggml_transpose(ctx0, k)); + + ggml_tensor * kq = ggml_mul_mat(ctx0, k, q); + + kq = ggml_soft_max_ext(ctx0, kq, nullptr, 1.0f/sqrtf(float(hparams.posnet.n_embd)), 0.0f); + + cur = ggml_mul_mat(ctx0, kq, v); + + cur = ggml_conv_1d_ph(ctx0, layer.attn_o, cur, 1, 1); + cur = ggml_add(ctx0, cur, layer.attn_o_b); + + cur = ggml_add(ctx0, cur, inpL); + } break; + case 5: + { + cur = build_norm(cur, + layer.norm, + layer.norm_b, + LLM_NORM_GROUP, 0); + } break; + default: GGML_ABORT("unknown posnet layer"); + }; + } + + cur = ggml_cont(ctx0, ggml_transpose(ctx0, cur)); + + cur = build_norm(cur, + model.tok_norm, + model.tok_norm_b, + LLM_NORM, -1); + + cur = ggml_cont(ctx0, ggml_transpose(ctx0, cur)); + + inpL = cur; + + // convnext + for (uint32_t il = 0; il < hparams.convnext.n_layer; ++il) { + const auto & layer = model.layers[il].convnext; + + cur = inpL; + + cur = ggml_conv_1d_dw_ph(ctx0, layer.dw, cur, 1, 1); + cur = ggml_add(ctx0, cur, layer.dw_b); + + cur = ggml_cont(ctx0, ggml_transpose(ctx0, cur)); + + cur = build_norm(cur, + layer.norm, + layer.norm_b, + LLM_NORM, -1); + + cur = build_ffn(cur, + layer.pw1, layer.pw1_b, NULL, + NULL, NULL, NULL, + layer.pw2, layer.pw2_b, NULL, + NULL, + LLM_FFN_GELU, LLM_FFN_SEQ, il); + + cur = ggml_mul(ctx0, cur, layer.gamma); + + cur = ggml_cont(ctx0, ggml_transpose(ctx0, cur)); + + inpL = ggml_add(ctx0, cur, inpL); + } + + cur = inpL; + + cur = ggml_cont(ctx0, ggml_transpose(ctx0, cur)); + + cur = build_norm(cur, + model.output_norm, + model.output_norm_b, + LLM_NORM, -1); + + // lm_head + cur = build_lora_mm(model.output, cur); + + cur = ggml_add(ctx0, cur, model.output_b); + + cb(cur, "result_embd", -1); + res->t_embd = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +struct llm_build_plm : public llm_graph_context { + llm_build_plm(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { + const float kq_scale = 1.0f/sqrtf(float(hparams.n_embd_head_k)); + + const uint32_t n_embd_head_qk_rope = hparams.n_rot; + const uint32_t n_embd_head_qk_nope = hparams.n_embd_head_k - hparams.n_rot; + const uint32_t kv_lora_rank = hparams.n_lora_kv; + + ggml_tensor * cur; + ggml_tensor * inpL; + + // {n_embd, n_tokens} + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv_unified(); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + // norm + cur = build_norm(inpL, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self_attention + { + ggml_tensor * q = NULL; + q = ggml_mul_mat(ctx0, model.layers[il].wq, cur); + cb(q, "q", il); + + // split into {n_head * n_embd_head_qk_nope, n_tokens} + ggml_tensor * q_nope = ggml_view_3d(ctx0, q, n_embd_head_qk_nope, n_head, n_tokens, + ggml_row_size(q->type, hparams.n_embd_head_k), + ggml_row_size(q->type, hparams.n_embd_head_k * n_head), + 0); + cb(q_nope, "q_nope", il); + + // and {n_head * n_embd_head_qk_rope, n_tokens} + ggml_tensor * q_pe = ggml_view_3d(ctx0, q, n_embd_head_qk_rope, n_head, n_tokens, + ggml_row_size(q->type, hparams.n_embd_head_k), + ggml_row_size(q->type, hparams.n_embd_head_k * n_head), + ggml_row_size(q->type, n_embd_head_qk_nope)); + cb(q_pe, "q_pe", il); + + // {n_embd, kv_lora_rank + n_embd_head_qk_rope} * {n_embd, n_tokens} -> {kv_lora_rank + n_embd_head_qk_rope, n_tokens} + ggml_tensor * kv_pe_compresseed = ggml_mul_mat(ctx0, model.layers[il].wkv_a_mqa, cur); + cb(kv_pe_compresseed, "kv_pe_compresseed", il); + + // split into {kv_lora_rank, n_tokens} + ggml_tensor * kv_compressed = ggml_view_2d(ctx0, kv_pe_compresseed, kv_lora_rank, n_tokens, + kv_pe_compresseed->nb[1], + 0); + cb(kv_compressed, "kv_compressed", il); + + // and {n_embd_head_qk_rope, n_tokens} + ggml_tensor * k_pe = ggml_view_3d(ctx0, kv_pe_compresseed, n_embd_head_qk_rope, 1, n_tokens, + kv_pe_compresseed->nb[1], + kv_pe_compresseed->nb[1], + ggml_row_size(kv_pe_compresseed->type, kv_lora_rank)); + cb(k_pe, "k_pe", il); + + kv_compressed = build_norm(kv_compressed, + model.layers[il].attn_kv_a_norm, NULL, + LLM_NORM_RMS, il); + cb(kv_compressed, "kv_compressed", il); + + // {kv_lora_rank, n_head * (n_embd_head_qk_nope + n_embd_head_v)} * {kv_lora_rank, n_tokens} -> {n_head * (n_embd_head_qk_nope + n_embd_head_v), n_tokens} + ggml_tensor * kv = ggml_mul_mat(ctx0, model.layers[il].wkv_b, kv_compressed); + cb(kv, "kv", il); + + // split into {n_head * n_embd_head_qk_nope, n_tokens} + ggml_tensor * k_nope = ggml_view_3d(ctx0, kv, n_embd_head_qk_nope, n_head, n_tokens, + ggml_row_size(kv->type, n_embd_head_qk_nope + hparams.n_embd_head_v), + ggml_row_size(kv->type, n_head * (n_embd_head_qk_nope + hparams.n_embd_head_v)), + 0); + cb(k_nope, "k_nope", il); + + // and {n_head * n_embd_head_v, n_tokens} + ggml_tensor * v_states = ggml_view_3d(ctx0, kv, hparams.n_embd_head_v, n_head, n_tokens, + ggml_row_size(kv->type, (n_embd_head_qk_nope + hparams.n_embd_head_v)), + ggml_row_size(kv->type, (n_embd_head_qk_nope + hparams.n_embd_head_v)*n_head), + ggml_row_size(kv->type, (n_embd_head_qk_nope))); + cb(v_states, "v_states", il); + + v_states = ggml_cont(ctx0, v_states); + cb(v_states, "v_states", il); + + v_states = ggml_view_2d(ctx0, v_states, hparams.n_embd_head_v * n_head, n_tokens, + ggml_row_size(kv->type, hparams.n_embd_head_v * n_head), + 0); + cb(v_states, "v_states", il); + + q_pe = ggml_rope_ext( + ctx0, q_pe, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(q_pe, "q_pe", il); + + // shared RoPE key + k_pe = ggml_rope_ext( + ctx0, k_pe, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(k_pe, "k_pe", il); + + ggml_tensor * q_states = ggml_concat(ctx0, q_nope, q_pe, 0); + cb(q_states, "q_states", il); + + ggml_tensor * k_states = ggml_concat(ctx0, k_nope, ggml_repeat(ctx0, k_pe, q_pe), 0); + cb(k_states, "k_states", il); + + cur = build_attn(inp_attn, gf, + model.layers[il].wo, NULL, + q_states, k_states, v_states, nullptr, kq_scale, il); + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + NULL, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + LLM_FFN_RELU_SQR, LLM_FFN_SEQ, il); + cb(cur, "ffn_out", il); + + cur = ggml_add(ctx0, cur, ffn_inp); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = build_norm(cur, + model.output_norm, NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +struct llm_build_bailingmoe : public llm_graph_context { + llm_build_bailingmoe(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv_unified(); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + // norm + cur = build_norm(inpL, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self-attention + { + // rope freq factors for llama3; may return nullptr for llama2 and other models + ggml_tensor * rope_factors = static_cast(memory)->cbs.get_rope_factors(n_ctx_per_seq, il); + + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + if (model.layers[il].bq) { + Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); + cb(Qcur, "Qcur", il); + } + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + if (model.layers[il].bk) { + Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); + cb(Kcur, "Kcur", il); + } + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + if (model.layers[il].bv) { + Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); + cb(Vcur, "Vcur", il); + } + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_rot, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_rot, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_rot, n_head_kv, n_tokens); + + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, rope_factors, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, rope_factors, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, gf, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_rot)), il); + } + + if (il == n_layer - 1) { + // skip computing output for unused tokens + ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + ggml_tensor * moe_out = + build_moe_ffn(cur, + model.layers[il].ffn_gate_inp, + model.layers[il].ffn_up_exps, + model.layers[il].ffn_gate_exps, + model.layers[il].ffn_down_exps, + nullptr, + n_expert, n_expert_used, + LLM_FFN_SILU, hparams.expert_weights_norm, + false, hparams.expert_weights_scale, + LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, + il); + cb(moe_out, "ffn_moe_out", il); + + // FFN shared expert + { + ggml_tensor * ffn_shexp = build_ffn(cur, + model.layers[il].ffn_up_shexp, NULL, NULL, + model.layers[il].ffn_gate_shexp, NULL, NULL, + model.layers[il].ffn_down_shexp, NULL, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(ffn_shexp, "ffn_shexp", il); + + cur = ggml_add(ctx0, moe_out, ffn_shexp); + cb(cur, "ffn_out", il); + } + + cur = ggml_add(ctx0, cur, ffn_inp); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = build_norm(cur, + model.output_norm, NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); + } +}; + +llama_memory_i * llama_model::create_memory() const { + llama_memory_i * res; + + switch (arch) { + case LLM_ARCH_MAMBA: + case LLM_ARCH_RWKV6: + case LLM_ARCH_RWKV6QWEN2: + case LLM_ARCH_RWKV7: + case LLM_ARCH_ARWKV7: + { + res = new llama_kv_cache_unified(hparams, { + /*.get_rope_factors =*/ nullptr + }); + } break; + default: + { + res = new llama_kv_cache_unified(hparams, { + /*.get_rope_factors =*/ [this](uint32_t n_ctx_per_seq, int il) { + // choose long/short freq factors based on the context size + if (layers[il].rope_freqs != nullptr) { + return layers[il].rope_freqs; + } + + if (n_ctx_per_seq > hparams.n_ctx_orig_yarn) { + return layers[il].rope_long; + } + + return layers[il].rope_short; + } + }); + } + } + + return res; +} + +llm_graph_result_ptr llama_model::build_graph( + const llm_graph_params & params, + ggml_cgraph * gf, + llm_graph_type type) const { + std::unique_ptr llm; + + switch (arch) { + case LLM_ARCH_LLAMA: + case LLM_ARCH_LLAMA4: + case LLM_ARCH_MINICPM: + case LLM_ARCH_GRANITE: + case LLM_ARCH_GRANITE_MOE: + { + llm = std::make_unique(*this, params, gf); + } break; + case LLM_ARCH_DECI: + { + llm = std::make_unique(*this, params, gf); + } break; + case LLM_ARCH_BAICHUAN: + { + llm = std::make_unique(*this, params, gf); + } break; + case LLM_ARCH_FALCON: + { + llm = std::make_unique(*this, params, gf); + } break; + case LLM_ARCH_GROK: + { + llm = std::make_unique(*this, params, gf); + } break; + case LLM_ARCH_STARCODER: + { + llm = std::make_unique(*this, params, gf); + } break; + case LLM_ARCH_REFACT: + { + llm = std::make_unique(*this, params, gf); + } break; + case LLM_ARCH_BERT: + case LLM_ARCH_JINA_BERT_V2: + case LLM_ARCH_NOMIC_BERT: + { + llm = std::make_unique(*this, params, gf); + } break; + case LLM_ARCH_BLOOM: + { + llm = std::make_unique(*this, params, gf); + } break; + case LLM_ARCH_MPT: + { + llm = std::make_unique(*this, params, gf); + } break; + case LLM_ARCH_STABLELM: + { + llm = std::make_unique(*this, params, gf); + } break; + case LLM_ARCH_QWEN: + { + llm = std::make_unique(*this, params, gf); + } break; + case LLM_ARCH_QWEN2: + { + llm = std::make_unique(*this, params, gf); + } break; + case LLM_ARCH_QWEN2VL: + { + llm = std::make_unique(*this, params, gf); + } break; + case LLM_ARCH_QWEN2MOE: + { + llm = std::make_unique(*this, params, gf); + } break; + case LLM_ARCH_QWEN3: + { + llm = std::make_unique(*this, params, gf); + } break; + case LLM_ARCH_QWEN3MOE: + { + llm = std::make_unique(*this, params, gf); + } break; + case LLM_ARCH_PHI2: + { + llm = std::make_unique(*this, params, gf); + } break; + case LLM_ARCH_PHI3: + case LLM_ARCH_PHIMOE: + { + llm = std::make_unique(*this, params, gf); + } break; + case LLM_ARCH_PLAMO: + { + llm = std::make_unique(*this, params, gf); + } break; + case LLM_ARCH_GPT2: + { + llm = std::make_unique(*this, params, gf); + } break; + case LLM_ARCH_CODESHELL: + { + llm = std::make_unique(*this, params, gf); + } break; + case LLM_ARCH_ORION: + { + llm = std::make_unique(*this, params, gf); + } break; + case LLM_ARCH_INTERNLM2: + { + llm = std::make_unique(*this, params, gf); + } break; + case LLM_ARCH_MINICPM3: + { + llm = std::make_unique(*this, params, gf); + } break; + case LLM_ARCH_GEMMA: + { + llm = std::make_unique(*this, params, gf); + } break; + case LLM_ARCH_GEMMA2: + { + llm = std::make_unique(*this, params, gf); + } break; + case LLM_ARCH_GEMMA3: + { + llm = std::make_unique(*this, params, gf); + } break; + case LLM_ARCH_STARCODER2: + { + llm = std::make_unique(*this, params, gf); + } break; + case LLM_ARCH_MAMBA: + { + llm = std::make_unique(*this, params, gf); + } break; + case LLM_ARCH_XVERSE: + { + llm = std::make_unique(*this, params, gf); + } break; + case LLM_ARCH_COMMAND_R: + { + llm = std::make_unique(*this, params, gf); + } break; + case LLM_ARCH_COHERE2: + { + llm = std::make_unique(*this, params, gf); + } break; + case LLM_ARCH_DBRX: + { + llm = std::make_unique(*this, params, gf); + } break; + case LLM_ARCH_OLMO: + { + llm = std::make_unique(*this, params, gf); + } break; + case LLM_ARCH_OLMO2: + { + llm = std::make_unique(*this, params, gf); + } break; + case LLM_ARCH_OLMOE: + { + llm = std::make_unique(*this, params, gf); + } break; + case LLM_ARCH_OPENELM: + { + llm = std::make_unique(*this, params, gf); + } break; + case LLM_ARCH_GPTNEOX: + { + llm = std::make_unique(*this, params, gf); + } break; + case LLM_ARCH_ARCTIC: + { + llm = std::make_unique(*this, params, gf); + } break; + case LLM_ARCH_DEEPSEEK: + { + llm = std::make_unique(*this, params, gf); + } break; + case LLM_ARCH_DEEPSEEK2: + { + llm = std::make_unique(*this, params, gf); + } break; + case LLM_ARCH_CHATGLM: + { + llm = std::make_unique(*this, params, gf); + } break; + case LLM_ARCH_GLM4: + { + llm = std::make_unique(*this, params, gf); + } break; + case LLM_ARCH_BITNET: + { + llm = std::make_unique(*this, params, gf); + } break; + case LLM_ARCH_T5: + { + switch (type) { + case LLM_GRAPH_TYPE_ENCODER: + llm = std::make_unique(*this, params, gf); + break; + case LLM_GRAPH_TYPE_DEFAULT: + case LLM_GRAPH_TYPE_DECODER: + llm = std::make_unique(*this, params, gf); + break; + default: + GGML_ABORT("invalid graph type"); + }; + } break; + case LLM_ARCH_T5ENCODER: + { + llm = std::make_unique(*this, params, gf); + } + break; + case LLM_ARCH_JAIS: + { + llm = std::make_unique(*this, params, gf); + } break; + case LLM_ARCH_NEMOTRON: + { + llm = std::make_unique(*this, params, gf); + } break; + case LLM_ARCH_EXAONE: + { + llm = std::make_unique(*this, params, gf); + } break; + case LLM_ARCH_RWKV6: + { + llm = std::make_unique(*this, params, gf); + } break; + case LLM_ARCH_RWKV6QWEN2: + { + llm = std::make_unique(*this, params, gf); + } break; + case LLM_ARCH_RWKV7: + { + llm = std::make_unique(*this, params, gf); + } break; + case LLM_ARCH_ARWKV7: + { + llm = std::make_unique(*this, params, gf); + } break; + case LLM_ARCH_CHAMELEON: + { + llm = std::make_unique(*this, params, gf); + } break; + case LLM_ARCH_WAVTOKENIZER_DEC: + { + llm = std::make_unique(*this, params, gf); + } break; + case LLM_ARCH_PLM: + { + llm = std::make_unique(*this, params, gf); + } break; + case LLM_ARCH_BAILINGMOE: + { + llm = std::make_unique(*this, params, gf); + } break; + default: + GGML_ABORT("fatal error"); + } + + // add on pooling layer + llm->build_pooling(gf, cls, cls_b, cls_out, cls_out_b); + + return std::move(llm->res); +} + // // interface implementation // -struct llama_model_params llama_model_default_params() { - struct llama_model_params result = { +llama_model_params llama_model_default_params() { + llama_model_params result = { /*.devices =*/ nullptr, + /*.tensor_buft_overrides =*/ nullptr, /*.n_gpu_layers =*/ 0, /*.split_mode =*/ LLAMA_SPLIT_MODE_LAYER, /*.main_gpu =*/ 0, @@ -3802,55 +13044,59 @@ struct llama_model_params llama_model_default_params() { return result; } -const struct llama_vocab * llama_model_get_vocab(const struct llama_model * model) { +const llama_vocab * llama_model_get_vocab(const llama_model * model) { return &model->vocab; } -void llama_free_model(struct llama_model * model) { +void llama_free_model(llama_model * model) { llama_model_free(model); } -void llama_model_free(struct llama_model * model) { +void llama_model_free(llama_model * model) { delete model; } -int32_t llama_model_n_ctx_train(const struct llama_model * model) { +int32_t llama_model_n_ctx_train(const llama_model * model) { return model->hparams.n_ctx_train; } -int32_t llama_model_n_embd(const struct llama_model * model) { +int32_t llama_model_n_embd(const llama_model * model) { return model->hparams.n_embd; } -int32_t llama_model_n_layer(const struct llama_model * model) { +int32_t llama_model_n_layer(const llama_model * model) { return model->hparams.n_layer; } -int32_t llama_model_n_head(const struct llama_model * model) { +int32_t llama_model_n_head(const llama_model * model) { return model->hparams.n_head(); } +int32_t llama_model_n_head_kv(const llama_model * model) { + return model->hparams.n_head_kv(); +} + // deprecated -int32_t llama_n_ctx_train(const struct llama_model * model) { +int32_t llama_n_ctx_train(const llama_model * model) { return llama_model_n_ctx_train(model); } // deprecated -int32_t llama_n_embd(const struct llama_model * model) { +int32_t llama_n_embd(const llama_model * model) { return llama_model_n_embd(model); } // deprecated -int32_t llama_n_layer(const struct llama_model * model) { +int32_t llama_n_layer(const llama_model * model) { return llama_model_n_layer(model); } // deprecated -int32_t llama_n_head(const struct llama_model * model) { +int32_t llama_n_head(const llama_model * model) { return llama_model_n_head(model); } -enum llama_rope_type llama_model_rope_type(const struct llama_model * model) { +llama_rope_type llama_model_rope_type(const llama_model * model) { switch (model->arch) { // these models do not use RoPE case LLM_ARCH_GPT2: @@ -3865,11 +13111,14 @@ enum llama_rope_type llama_model_rope_type(const struct llama_model * model) { case LLM_ARCH_JAIS: case LLM_ARCH_RWKV6: case LLM_ARCH_RWKV6QWEN2: + case LLM_ARCH_RWKV7: + case LLM_ARCH_ARWKV7: case LLM_ARCH_WAVTOKENIZER_DEC: return LLAMA_ROPE_TYPE_NONE; // use what we call a normal RoPE, operating on pairs of consecutive head values case LLM_ARCH_LLAMA: + case LLM_ARCH_LLAMA4: case LLM_ARCH_DECI: case LLM_ARCH_BAICHUAN: case LLM_ARCH_STARCODER: @@ -3884,10 +13133,13 @@ enum llama_rope_type llama_model_rope_type(const struct llama_model * model) { case LLM_ARCH_ARCTIC: case LLM_ARCH_DEEPSEEK: case LLM_ARCH_DEEPSEEK2: + case LLM_ARCH_PLM: case LLM_ARCH_CHATGLM: + case LLM_ARCH_GLM4: case LLM_ARCH_GRANITE: case LLM_ARCH_GRANITE_MOE: case LLM_ARCH_CHAMELEON: + case LLM_ARCH_BAILINGMOE: return LLAMA_ROPE_TYPE_NORM; // the pairs of head values are offset by n_rot/2 @@ -3901,6 +13153,8 @@ enum llama_rope_type llama_model_rope_type(const struct llama_model * model) { case LLM_ARCH_QWEN: case LLM_ARCH_QWEN2: case LLM_ARCH_QWEN2MOE: + case LLM_ARCH_QWEN3: + case LLM_ARCH_QWEN3MOE: case LLM_ARCH_OLMO2: case LLM_ARCH_OLMOE: case LLM_ARCH_PHI2: @@ -3908,6 +13162,7 @@ enum llama_rope_type llama_model_rope_type(const struct llama_model * model) { case LLM_ARCH_PHIMOE: case LLM_ARCH_GEMMA: case LLM_ARCH_GEMMA2: + case LLM_ARCH_GEMMA3: case LLM_ARCH_STARCODER2: case LLM_ARCH_OPENELM: case LLM_ARCH_GPTNEOX: @@ -3928,11 +13183,11 @@ enum llama_rope_type llama_model_rope_type(const struct llama_model * model) { return LLAMA_ROPE_TYPE_NONE; } -float llama_model_rope_freq_scale_train(const struct llama_model * model) { +float llama_model_rope_freq_scale_train(const llama_model * model) { return model->hparams.rope_freq_scale_train; } -int32_t llama_model_meta_val_str(const struct llama_model * model, const char * key, char * buf, size_t buf_size) { +int32_t llama_model_meta_val_str(const llama_model * model, const char * key, char * buf, size_t buf_size) { const auto & it = model->gguf_kv.find(key); if (it == model->gguf_kv.end()) { if (buf_size > 0) { @@ -3943,11 +13198,11 @@ int32_t llama_model_meta_val_str(const struct llama_model * model, const char * return snprintf(buf, buf_size, "%s", it->second.c_str()); } -int32_t llama_model_meta_count(const struct llama_model * model) { +int32_t llama_model_meta_count(const llama_model * model) { return (int)model->gguf_kv.size(); } -int32_t llama_model_meta_key_by_index(const struct llama_model * model, int i, char * buf, size_t buf_size) { +int32_t llama_model_meta_key_by_index(const llama_model * model, int i, char * buf, size_t buf_size) { if (i < 0 || i >= (int)model->gguf_kv.size()) { if (buf_size > 0) { buf[0] = '\0'; @@ -3959,7 +13214,7 @@ int32_t llama_model_meta_key_by_index(const struct llama_model * model, int i, c return snprintf(buf, buf_size, "%s", it->first.c_str()); } -int32_t llama_model_meta_val_str_by_index(const struct llama_model * model, int32_t i, char * buf, size_t buf_size) { +int32_t llama_model_meta_val_str_by_index(const llama_model * model, int32_t i, char * buf, size_t buf_size) { if (i < 0 || i >= (int)model->gguf_kv.size()) { if (buf_size > 0) { buf[0] = '\0'; @@ -3971,15 +13226,15 @@ int32_t llama_model_meta_val_str_by_index(const struct llama_model * model, int3 return snprintf(buf, buf_size, "%s", it->second.c_str()); } -int32_t llama_model_desc(const struct llama_model * model, char * buf, size_t buf_size) { +int32_t llama_model_desc(const llama_model * model, char * buf, size_t buf_size) { return snprintf(buf, buf_size, "%s", model->desc().c_str()); } -uint64_t llama_model_size(const struct llama_model * model) { +uint64_t llama_model_size(const llama_model * model) { return model->size(); } -const char * llama_model_chat_template(const struct llama_model * model, const char * name) { +const char * llama_model_chat_template(const llama_model * model, const char * name) { const auto key = name ? LLM_KV(model->arch, name)(LLM_KV_TOKENIZER_CHAT_TEMPLATE_N) : LLM_KV(model->arch)(LLM_KV_TOKENIZER_CHAT_TEMPLATE); const auto & it = model->gguf_kv.find(key); @@ -3990,11 +13245,11 @@ const char * llama_model_chat_template(const struct llama_model * model, const c return it->second.c_str(); } -uint64_t llama_model_n_params(const struct llama_model * model) { +uint64_t llama_model_n_params(const llama_model * model) { return model->n_elements(); } -bool llama_model_has_encoder(const struct llama_model * model) { +bool llama_model_has_encoder(const llama_model * model) { switch (model->arch) { case LLM_ARCH_T5: return true; case LLM_ARCH_T5ENCODER: return true; @@ -4002,22 +13257,28 @@ bool llama_model_has_encoder(const struct llama_model * model) { } } -bool llama_model_has_decoder(const struct llama_model * model) { +bool llama_model_has_decoder(const llama_model * model) { switch (model->arch) { case LLM_ARCH_T5ENCODER: return false; default: return true; } } -llama_token llama_model_decoder_start_token(const struct llama_model * model) { +llama_token llama_model_decoder_start_token(const llama_model * model) { return model->hparams.dec_start_token_id; } -bool llama_model_is_recurrent(const struct llama_model * model) { +bool llama_model_is_recurrent(const llama_model * model) { switch (model->arch) { - case LLM_ARCH_MAMBA: return true; - case LLM_ARCH_RWKV6: return true; - case LLM_ARCH_RWKV6QWEN2: return true; - default: return false; + case LLM_ARCH_MAMBA: return true; + case LLM_ARCH_RWKV6: return true; + case LLM_ARCH_RWKV6QWEN2: return true; + case LLM_ARCH_RWKV7: return true; + case LLM_ARCH_ARWKV7: return true; + default: return false; } } + +const std::vector> & llama_internal_get_tensor_map(const llama_model * model) { + return model->tensors_by_name; +} diff --git a/src/llama-model.h b/src/llama-model.h index a7c304447..0f18dac16 100644 --- a/src/llama-model.h +++ b/src/llama-model.h @@ -2,7 +2,9 @@ #include "llama.h" #include "llama-arch.h" +#include "llama-graph.h" #include "llama-hparams.h" +#include "llama-memory.h" #include "llama-vocab.h" #include @@ -10,6 +12,8 @@ #include #include +struct llama_cparams; +struct llama_ubatch; struct llama_model_loader; // available models @@ -25,6 +29,7 @@ enum llm_type { LLM_TYPE_109M, LLM_TYPE_137M, LLM_TYPE_160M, + LLM_TYPE_190M, LLM_TYPE_220M, LLM_TYPE_250M, LLM_TYPE_270M, @@ -39,8 +44,10 @@ enum llm_type { LLM_TYPE_1_4B, LLM_TYPE_1_5B, LLM_TYPE_1_6B, + LLM_TYPE_1_8B, LLM_TYPE_2B, LLM_TYPE_2_8B, + LLM_TYPE_2_9B, LLM_TYPE_3B, LLM_TYPE_4B, LLM_TYPE_6B, @@ -78,6 +85,9 @@ enum llm_type { LLM_TYPE_10B_128x3_66B, LLM_TYPE_57B_A14B, LLM_TYPE_27B, + LLM_TYPE_290B, + LLM_TYPE_17B_16E, // llama4 Scout + LLM_TYPE_17B_128E, // llama4 Maverick }; struct llama_layer_posnet { @@ -256,6 +266,20 @@ struct llama_layer { struct ggml_tensor * time_mix_receptance_b = nullptr; struct ggml_tensor * time_mix_gate = nullptr; + // rwkv7 + struct ggml_tensor * time_mix_w0 = nullptr; + struct ggml_tensor * time_mix_a0 = nullptr; + struct ggml_tensor * time_mix_a1 = nullptr; + struct ggml_tensor * time_mix_a2 = nullptr; + struct ggml_tensor * time_mix_v0 = nullptr; + struct ggml_tensor * time_mix_v1 = nullptr; + struct ggml_tensor * time_mix_v2 = nullptr; + struct ggml_tensor * time_mix_g1 = nullptr; + struct ggml_tensor * time_mix_g2 = nullptr; + struct ggml_tensor * time_mix_k_k = nullptr; + struct ggml_tensor * time_mix_k_a = nullptr; + struct ggml_tensor * time_mix_r_k = nullptr; + struct ggml_tensor * time_mix_ln = nullptr; struct ggml_tensor * time_mix_ln_b = nullptr; struct ggml_tensor * time_mix_output = nullptr; @@ -347,7 +371,7 @@ struct llama_model { std::string desc() const; size_t size() const; - size_t max_nodes() const; + size_t n_tensors() const; size_t n_devices() const; // total number of parameters in the model @@ -360,11 +384,26 @@ struct llama_model { ggml_backend_buffer_type_t select_buft(int il) const; + bool has_tensor_overrides() const; + const struct ggml_tensor * get_tensor(const char * name) const; + // TODO: move this to new llm_arch_model_i interface + llama_memory_i * create_memory() const; // TODO: params + + // TODO: move this to new llm_arch_model_i interface + llm_graph_result_ptr build_graph( + const llm_graph_params & params, + ggml_cgraph * gf, + llm_graph_type type) const; + private: struct impl; std::unique_ptr pimpl; }; const char * llm_type_name(llm_type type); + +// For internal test use +// TODO: remove +const std::vector> & llama_internal_get_tensor_map(const llama_model * model); diff --git a/src/llama-quant.cpp b/src/llama-quant.cpp index fb7982655..e3e10fa6c 100644 --- a/src/llama-quant.cpp +++ b/src/llama-quant.cpp @@ -527,7 +527,7 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std:: } std::vector splits = {}; - llama_model_loader ml(fname_inp, splits, use_mmap, /*check_tensors*/ true, kv_overrides); + llama_model_loader ml(fname_inp, splits, use_mmap, /*check_tensors*/ true, kv_overrides, nullptr); ml.init_mappings(false); // no prefetching llama_model model(llama_model_default_params()); @@ -756,10 +756,19 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std:: // NOTE: can't use LLM_TN here because the layer number is not known quantize &= name.find("ssm_conv1d.weight") == std::string::npos; - // do not quantize RWKV's time_mix_first tensors + // do not quantize RWKV's small yet 2D weights quantize &= name.find("time_mix_first.weight") == std::string::npos; + quantize &= name.find("time_mix_w0.weight") == std::string::npos; quantize &= name.find("time_mix_w1.weight") == std::string::npos; quantize &= name.find("time_mix_w2.weight") == std::string::npos; + quantize &= name.find("time_mix_v0.weight") == std::string::npos; + quantize &= name.find("time_mix_v1.weight") == std::string::npos; + quantize &= name.find("time_mix_v2.weight") == std::string::npos; + quantize &= name.find("time_mix_a0.weight") == std::string::npos; + quantize &= name.find("time_mix_a1.weight") == std::string::npos; + quantize &= name.find("time_mix_a2.weight") == std::string::npos; + quantize &= name.find("time_mix_g1.weight") == std::string::npos; + quantize &= name.find("time_mix_g2.weight") == std::string::npos; quantize &= name.find("time_mix_decay_w1.weight") == std::string::npos; quantize &= name.find("time_mix_decay_w2.weight") == std::string::npos; quantize &= name.find("time_mix_lerp_fused.weight") == std::string::npos; diff --git a/src/llama-sampling.cpp b/src/llama-sampling.cpp index 990b61297..d14979850 100644 --- a/src/llama-sampling.cpp +++ b/src/llama-sampling.cpp @@ -1449,7 +1449,9 @@ static struct llama_sampler * llama_sampler_init_grammar_impl( const char ** trigger_words, size_t num_trigger_words, const llama_token * trigger_tokens, - size_t num_trigger_tokens); + size_t num_trigger_tokens, + const char ** trigger_patterns, + size_t num_trigger_patterns); static void llama_sampler_grammar_reset(struct llama_sampler * smpl) { auto * ctx = (llama_sampler_grammar *) smpl->ctx; @@ -1457,12 +1459,14 @@ static void llama_sampler_grammar_reset(struct llama_sampler * smpl) { return; } - std::vector trigger_words; - for (auto & word : ctx->grammar->trigger_words) { - trigger_words.push_back(word.c_str()); + std::vector trigger_patterns_c; + trigger_patterns_c.reserve(ctx->grammar->trigger_patterns.size()); + for (auto & trigger_pattern : ctx->grammar->trigger_patterns) { + trigger_patterns_c.push_back(trigger_pattern.pattern.c_str()); } + auto * grammar_new = llama_grammar_init_impl(ctx->grammar->vocab, ctx->grammar_str.c_str(), ctx->grammar_root.c_str(), - ctx->grammar->lazy, trigger_words.data(), trigger_words.size(), + ctx->grammar->lazy, trigger_patterns_c.data(), trigger_patterns_c.size(), ctx->grammar->trigger_tokens.data(), ctx->grammar->trigger_tokens.size()); llama_grammar_free_impl(ctx->grammar); @@ -1472,7 +1476,8 @@ static void llama_sampler_grammar_reset(struct llama_sampler * smpl) { static struct llama_sampler * llama_sampler_grammar_clone(const struct llama_sampler * smpl) { const auto * ctx = (const llama_sampler_grammar *) smpl->ctx; - auto * result = llama_sampler_init_grammar_impl(ctx->vocab, nullptr, nullptr, false, nullptr, 0, nullptr, 0); + auto * result = llama_sampler_init_grammar_impl(ctx->vocab, nullptr, nullptr, false, nullptr, 0, nullptr, 0, nullptr, 0); + GGML_ASSERT(result); // copy the state { @@ -1516,16 +1521,38 @@ static struct llama_sampler * llama_sampler_init_grammar_impl( const char ** trigger_words, size_t num_trigger_words, const llama_token * trigger_tokens, - size_t num_trigger_tokens) { + size_t num_trigger_tokens, + const char ** trigger_patterns, + size_t num_trigger_patterns) { auto * ctx = new llama_sampler_grammar; if (grammar_str != nullptr && grammar_str[0] != '\0') { + // TODO: remove trigger_words support. + if (trigger_words != nullptr && num_trigger_words > 0) { + GGML_ASSERT(trigger_patterns == nullptr && num_trigger_patterns == 0); + std::string trigger_pattern("[\\s\\S]*?("); + for (size_t i = 0; i < num_trigger_words; ++i) { + static const std::regex special_chars("[.^$|()*+?\\[\\]{}\\\\]"); + if (i > 0) { + trigger_pattern += "|"; + } + trigger_pattern += std::regex_replace(trigger_words[i], special_chars, "\\$0"); + } + trigger_pattern += ")[\\s\\S]*"; + auto trigger_pattern_c = trigger_pattern.c_str(); + trigger_patterns = &trigger_pattern_c; + num_trigger_patterns = 1; + } *ctx = { /* .vocab = */ vocab, /* .grammar_str = */ grammar_str, /* .grammar_root = */ grammar_root, - /* .grammar = */ llama_grammar_init_impl(vocab, grammar_str, grammar_root, lazy, trigger_words, num_trigger_words, trigger_tokens, num_trigger_tokens), + /* .grammar = */ llama_grammar_init_impl(vocab, grammar_str, grammar_root, lazy, trigger_patterns, num_trigger_patterns, trigger_tokens, num_trigger_tokens), }; + if (!ctx->grammar) { + delete ctx; + return nullptr; + } } else { *ctx = { /* .vocab = */ vocab, @@ -1545,7 +1572,7 @@ struct llama_sampler * llama_sampler_init_grammar( const struct llama_vocab * vocab, const char * grammar_str, const char * grammar_root) { - return llama_sampler_init_grammar_impl(vocab, grammar_str, grammar_root, /* lazy= */ false, nullptr, 0, nullptr, 0); + return llama_sampler_init_grammar_impl(vocab, grammar_str, grammar_root, /* lazy= */ false, nullptr, 0, nullptr, 0, nullptr, 0); } struct llama_sampler * llama_sampler_init_grammar_lazy( @@ -1556,7 +1583,18 @@ struct llama_sampler * llama_sampler_init_grammar_lazy( size_t num_trigger_words, const llama_token * trigger_tokens, size_t num_trigger_tokens) { - return llama_sampler_init_grammar_impl(vocab, grammar_str, grammar_root, /* lazy= */ true, trigger_words, num_trigger_words, trigger_tokens, num_trigger_tokens); + return llama_sampler_init_grammar_impl(vocab, grammar_str, grammar_root, /* lazy= */ true, trigger_words, num_trigger_words, trigger_tokens, num_trigger_tokens, nullptr, 0); +} + +struct llama_sampler * llama_sampler_init_grammar_lazy_patterns( + const struct llama_vocab * vocab, + const char * grammar_str, + const char * grammar_root, + const char ** trigger_patterns, + size_t num_trigger_patterns, + const llama_token * trigger_tokens, + size_t num_trigger_tokens) { + return llama_sampler_init_grammar_impl(vocab, grammar_str, grammar_root, /* lazy= */ true, nullptr, 0, trigger_tokens, num_trigger_tokens, trigger_patterns, num_trigger_patterns); } // penalties @@ -1698,6 +1736,73 @@ struct llama_sampler * llama_sampler_init_penalties( ); } +// top-n-sigma + +struct llama_sampler_top_n_sigma { + const float n; +}; + +static const char * llama_sampler_top_n_sigma_name(const struct llama_sampler * /*smpl*/) { + return "top-n-sigma"; +} + +static void llama_sampler_top_n_sigma_apply(struct llama_sampler * smpl, llama_token_data_array * cur_p) { + const auto * ctx = (llama_sampler_top_n_sigma *) smpl->ctx; + + // find max logit and calculate mean + float max = cur_p->data[0].logit; + float logits_sum = 0; + for (size_t i = 0; i < cur_p->size; ++i) { + if (cur_p->data[i].logit > max) { + max = cur_p->data[i].logit; + } + logits_sum += cur_p->data[i].logit; + } + float mean = logits_sum/cur_p->size; + + // calculate standard deviation + float acc = 0; + for (size_t i = 0; i < cur_p->size; ++i) { + acc += pow(cur_p->data[i].logit - mean, 2); + } + float std = sqrt(acc/cur_p->size); + + //apply mask + for (size_t i = 0; i < cur_p->size; ++i) { + if (cur_p->data[i].logit < max - (ctx->n * std)) { + cur_p->data[i].logit = -INFINITY; + } + } + llama_sampler_softmax_impl(cur_p); +} + +static struct llama_sampler * llama_sampler_top_n_sigma_clone(const struct llama_sampler * smpl) { + const auto * ctx = (const llama_sampler_top_n_sigma *) smpl->ctx; + return llama_sampler_init_top_n_sigma(ctx->n); +} + +static void llama_sampler_top_n_sigma_free(struct llama_sampler * smpl) { + delete (llama_sampler_top_n_sigma *) smpl->ctx; +} + +static struct llama_sampler_i llama_sampler_top_n_sigma_i = { + /* .name = */ llama_sampler_top_n_sigma_name, + /* .accept = */ nullptr, + /* .apply = */ llama_sampler_top_n_sigma_apply, + /* .reset = */ nullptr, + /* .clone = */ llama_sampler_top_n_sigma_clone, + /* .free = */ llama_sampler_top_n_sigma_free, +}; + +struct llama_sampler * llama_sampler_init_top_n_sigma(float n) { + return llama_sampler_init( + /* .iface = */ &llama_sampler_top_n_sigma_i, + /* .ctx = */ new llama_sampler_top_n_sigma { + /* .n = */ n, + } + ); +} + // DRY struct llama_sampler_dry { diff --git a/src/llama-vocab.cpp b/src/llama-vocab.cpp index ad9ffe66a..464ff01e0 100644 --- a/src/llama-vocab.cpp +++ b/src/llama-vocab.cpp @@ -16,6 +16,7 @@ #include #include #include +#include // // helpers @@ -341,6 +342,7 @@ struct llm_tokenizer_bpe : llm_tokenizer { case LLAMA_VOCAB_PRE_TYPE_MPT: case LLAMA_VOCAB_PRE_TYPE_OLMO: case LLAMA_VOCAB_PRE_TYPE_JAIS: + case LLAMA_VOCAB_PRE_TYPE_TRILLION: regex_exprs = { "'s|'t|'re|'ve|'m|'ll|'d| ?\\p{L}+| ?\\p{N}+| ?[^\\s\\p{L}\\p{N}]+|\\s+(?!\\S)", }; @@ -392,6 +394,27 @@ struct llm_tokenizer_bpe : llm_tokenizer { "'s|'t|'re|'ve|'m|'ll|'d| ?\\p{L}+| ?\\p{N}+| ?[^\\s\\p{L}\\p{N}]+|\\s+(?!\\S)", }; break; + case LLAMA_VOCAB_PRE_TYPE_GPT4O: + regex_exprs = { + // original regex from tokenizer.json + // "[^\\r\\n\\p{L}\\p{N}]?[\\p{Lu}\\p{Lt}\\p{Lm}\\p{Lo}\\p{M}]*[\\p{Ll}\\p{Lm}\\p{Lo}\\p{M}]+(?i:'s|'t|'re|'ve|'m|'ll|'d)?|[^\\r\\n\\p{L}\\p{N}]?[\\p{Lu}\\p{Lt}\\p{Lm}\\p{Lo}\\p{M}]+[\\p{Ll}\\p{Lm}\\p{Lo}\\p{M}]*(?i:'s|'t|'re|'ve|'m|'ll|'d)?|\\p{N}{1,3}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n/]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+", + "[^\\r\\n\\p{L}\\p{N}]?((?=[\\p{L}])([^a-z]))*((?=[\\p{L}])([^A-Z]))+(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])?|[^\\r\\n\\p{L}\\p{N}]?((?=[\\p{L}])([^a-z]))+((?=[\\p{L}])([^A-Z]))*(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])?|\\p{N}{1,3}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n/]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+", + }; + break; + case LLAMA_VOCAB_PRE_TYPE_SUPERBPE: + regex_exprs = { + "\\p{N}+", + "(?=(\\d{3})+(?!\\d))", + }; + break; + case LLAMA_VOCAB_PRE_TYPE_BAILINGMOE: + regex_exprs = { + // original regex from tokenizer.json + // "'(?i:[sdmt]|ll|ve|re)|[^\\r\\n\\p{L}\\p{N}]?+\\p{L}+|\\p{N}| ?[^\\s\\p{L}\\p{N}]++[\\r\\n]*|\\s*[\\r\\n]|\\s+(?!\\S)|\\s+" + // FIXME? Changed possessive quantifiers (?+ and ++) to greedy to avoid errors and imatrix hanging (tried atomic grouping but it's not supported?) + "'(?:[sSdDmMtT]|[lL][lL]|[vV][eE]|[rR][eE])|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]|\\s+(?!\\S)|\\s+", + }; + break; default: // default regex for BPE tokenization pre-processing regex_exprs = { @@ -1549,6 +1572,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) { pre_type = LLAMA_VOCAB_PRE_TYPE_PORO; clean_spaces = false; } else if ( + tokenizer_pre == "glm4" || tokenizer_pre == "chatglm-bpe") { pre_type = LLAMA_VOCAB_PRE_TYPE_CHATGLM4; special_bos_id = LLAMA_TOKEN_NULL; @@ -1592,6 +1616,23 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) { } else if ( tokenizer_pre == "megrez") { pre_type = LLAMA_VOCAB_PRE_TYPE_QWEN2; + } else if ( + tokenizer_pre == "gpt-4o" || + tokenizer_pre == "llama4") { + pre_type = LLAMA_VOCAB_PRE_TYPE_GPT4O; + clean_spaces = false; + } else if ( + tokenizer_pre == "superbpe") { + pre_type = LLAMA_VOCAB_PRE_TYPE_SUPERBPE; + clean_spaces = false; + } else if ( + tokenizer_pre == "trillion") { + pre_type = LLAMA_VOCAB_PRE_TYPE_TRILLION; + clean_spaces = false; + } else if ( + tokenizer_pre == "bailingmoe") { + pre_type = LLAMA_VOCAB_PRE_TYPE_BAILINGMOE; + clean_spaces = false; } else { throw std::runtime_error(format("unknown pre-tokenizer type: '%s'", tokenizer_pre.c_str())); } @@ -1769,6 +1810,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) { || t.first == "" || t.first == "<|endoftext|>" || t.first == "" + || t.first == "_" || t.first == "<｜end▁of▁sentence｜>" // DeepSeek ) { special_eot_id = t.second; @@ -1801,6 +1843,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) { || t.first == "" || t.first == "<｜fim▁begin｜>" // DeepSeek || t.first == "
"
+                        || t.first == "▁
"          // CodeLlama
                         ) {
                     special_fim_pre_id = t.second;
                     if ((id_to_token[t.second].attr & LLAMA_TOKEN_ATTR_CONTROL) == 0) {
@@ -1818,6 +1861,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                         || t.first == ""
                         || t.first == "<｜fim▁hole｜>" // DeepSeek
                         || t.first == ""
+                        || t.first == "▁"         // CodeLlama
                         ) {
                     special_fim_suf_id = t.second;
                     if ((id_to_token[t.second].attr & LLAMA_TOKEN_ATTR_CONTROL) == 0) {
@@ -1835,6 +1879,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                         || t.first == ""
                         || t.first == "<｜fim▁end｜>"  // DeepSeek
                         || t.first == ""
+                        || t.first == "▁"         // CodeLlama
                         ) {
                     special_fim_mid_id = t.second;
                     if ((id_to_token[t.second].attr & LLAMA_TOKEN_ATTR_CONTROL) == 0) {
@@ -1919,6 +1964,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                     || t.first == "<|endoftext|>"
                     || t.first == "<|eom_id|>"
                     || t.first == ""
+                    || t.first == "_"
                ) {
                 special_eog_ids.insert(t.second);
                 if ((id_to_token[t.second].attr & LLAMA_TOKEN_ATTR_CONTROL) == 0) {
@@ -2177,14 +2223,12 @@ void llama_vocab::impl::tokenizer_st_partition(std::forward_list length
-                    if (match + text.length() > raw_text_base_offset + raw_text_base_length) break;
-
 #ifdef PRETOKENIZERDEBUG
                     LLAMA_LOG_WARN("FF: (%ld %ld %ld) '%s'\n", raw_text->length(), raw_text_base_offset, raw_text_base_length, raw_text->substr(raw_text_base_offset, raw_text_base_length).c_str());
 #endif
diff --git a/src/llama.cpp b/src/llama.cpp
index 607f27861..d5164720b 100644
--- a/src/llama.cpp
+++ b/src/llama.cpp
@@ -2,9370 +2,28 @@
 
 #include "llama-chat.h"
 #include "llama-mmap.h"
-#include "llama-context.h"
 #include "llama-vocab.h"
-#include "llama-sampling.h"
-#include "llama-kv-cache.h"
 #include "llama-model-loader.h"
 #include "llama-model.h"
 
 #include "ggml.h"
-#include "ggml-alloc.h"
 #include "ggml-backend.h"
-#include "ggml-cpp.h"
 
 #include 
-#include 
-#include 
-#include 
-#include 
 #include 
 #include 
 #include 
 #include 
 #include 
-#include 
 
 #if defined(_MSC_VER)
 #pragma warning(disable: 4244 4267) // possible loss of data
 #endif
 
-// Returns 0 on success, -1 on error, and -2 on cancellation via llama_progress_callback
-static int llama_model_load(const std::string & fname, std::vector & splits, llama_model & model, llama_model_params & params) {
-    // loading time will be recalculated after the first eval, so
-    // we take page faults deferred by mmap() into consideration
-    model.t_load_us = 0;
-    time_meas tm(model.t_load_us);
-
-    model.t_start_us = tm.t_start_us;
-
-    try {
-        llama_model_loader ml(fname, splits, params.use_mmap, params.check_tensors, params.kv_overrides);
-
-        ml.print_info();
-
-        model.hparams.vocab_only = params.vocab_only;
-
-        try {
-            model.load_arch(ml);
-        } catch(const std::exception & e) {
-            throw std::runtime_error("error loading model architecture: " + std::string(e.what()));
-        }
-        try {
-            model.load_hparams(ml);
-        } catch(const std::exception & e) {
-            throw std::runtime_error("error loading model hyperparameters: " + std::string(e.what()));
-        }
-        try {
-            model.load_vocab(ml);
-        } catch(const std::exception & e) {
-            throw std::runtime_error("error loading model vocabulary: " + std::string(e.what()));
-        }
-
-        model.load_stats(ml);
-        model.print_info();
-
-        if (params.vocab_only) {
-            LLAMA_LOG_INFO("%s: vocab only - skipping tensors\n", __func__);
-            return 0;
-        }
-
-        if (!model.load_tensors(ml)) {
-            return -2;
-        }
-    } catch (const std::exception & err) {
-        LLAMA_LOG_ERROR("%s: error loading model: %s\n", __func__, err.what());
-        return -1;
-    }
-
-    return 0;
-}
-
-//
-// llm_build
-//
-
-using llm_build_cb = std::function;
-
-enum llm_ffn_op_type {
-    LLM_FFN_SILU,
-    LLM_FFN_GELU,
-    LLM_FFN_RELU,
-    LLM_FFN_RELU_SQR,
-    LLM_FFN_SWIGLU,
-};
-
-enum llm_ffn_gate_type {
-    LLM_FFN_SEQ,
-    LLM_FFN_PAR, // ffn_gate is parallel to ffn_up
-};
-
-enum llm_norm_type {
-    LLM_NORM,
-    LLM_NORM_RMS,
-    LLM_NORM_GROUP,
-};
-
-static struct ggml_tensor * llm_build_inp_embd(
-        struct ggml_context * ctx,
-       struct llama_context & lctx,
-        const llama_hparams & hparams,
-         const llama_ubatch & ubatch,
-         struct ggml_tensor * tok_embd,
-         const llm_build_cb & cb) {
-    const int64_t n_embd = hparams.n_embd;
-
-    struct ggml_tensor * inpL;
-
-    if (ubatch.token) {
-        lctx.inp_tokens = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, ubatch.n_tokens);
-        cb(lctx.inp_tokens, "inp_tokens", -1);
-        ggml_set_input(lctx.inp_tokens);
-
-        inpL = ggml_get_rows(ctx, tok_embd, lctx.inp_tokens);
-
-        // apply lora for embedding tokens if needed
-        for (auto & it : lctx.lora) {
-            struct llama_adapter_lora_weight * lw = it.first->get_weight(tok_embd);
-            if (lw == nullptr) {
-                continue;
-            }
-            const float adapter_scale = it.second;
-            const float scale = lw->get_scale(it.first->alpha, adapter_scale);
-            struct ggml_tensor * inpL_delta = ggml_scale(ctx, ggml_mul_mat(
-                ctx, lw->b, // non-transposed lora_b
-                ggml_get_rows(ctx, lw->a, lctx.inp_tokens)
-            ), scale);
-            inpL = ggml_add(ctx, inpL, inpL_delta);
-        }
-    } else {
-        lctx.inp_embd = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, ubatch.n_tokens);
-        inpL = lctx.inp_embd;
-        ggml_set_input(lctx.inp_embd);
-    }
-
-    // For Granite architecture
-    if (hparams.f_embedding_scale != 0.0f) {
-        inpL = ggml_scale(ctx, inpL, hparams.f_embedding_scale);
-    }
-
-    cb(inpL, "inp_embd", -1);
-
-    return inpL;
-}
-
-static void llm_build_kv_store(
-        struct ggml_context * ctx,
-        const llama_hparams & hparams,
-        const llama_cparams & cparams,
-       const llama_kv_cache & kv,
-         struct ggml_cgraph * graph,
-         struct ggml_tensor * k_cur,
-         struct ggml_tensor * v_cur,
-                    int32_t   n_tokens,
-                    int32_t   kv_head,
-         const llm_build_cb & cb,
-                    int64_t   il) {
-    const int64_t n_ctx = cparams.n_ctx;
-
-    const int64_t n_embd_k_gqa = hparams.n_embd_k_gqa(il);
-    const int64_t n_embd_v_gqa = hparams.n_embd_v_gqa(il);
-
-    GGML_ASSERT(kv.size == n_ctx);
-
-    struct ggml_tensor * k_cache_view = ggml_view_1d(ctx, kv.k_l[il], n_tokens*n_embd_k_gqa, ggml_row_size(kv.k_l[il]->type, n_embd_k_gqa)*kv_head);
-    cb(k_cache_view, "k_cache_view", il);
-
-    // note: storing RoPE-ed version of K in the KV cache
-    ggml_build_forward_expand(graph, ggml_cpy(ctx, k_cur, k_cache_view));
-
-    assert(v_cur->ne[0] == n_embd_v_gqa && v_cur->ne[1] == n_tokens);
-
-    struct ggml_tensor * v_cache_view = nullptr;
-
-    if (cparams.flash_attn) {
-        v_cache_view = ggml_view_1d(ctx, kv.v_l[il], n_tokens*n_embd_v_gqa, ggml_row_size(kv.v_l[il]->type, n_embd_v_gqa)*kv_head);
-    } else {
-        // note: the V cache is transposed when not using flash attention
-        v_cache_view = ggml_view_2d(ctx, kv.v_l[il], n_tokens, n_embd_v_gqa,
-                (  n_ctx)*ggml_element_size(kv.v_l[il]),
-                (kv_head)*ggml_element_size(kv.v_l[il]));
-
-        v_cur = ggml_transpose(ctx, v_cur);
-    }
-    cb(v_cache_view, "v_cache_view", il);
-
-    ggml_build_forward_expand(graph, ggml_cpy(ctx, v_cur, v_cache_view));
-}
-
-// do mat_mul, while optionally apply lora
-static struct ggml_tensor * llm_build_lora_mm(
-        struct llama_context & lctx,
-         struct ggml_context * ctx0,
-          struct ggml_tensor * w,
-          struct ggml_tensor * cur) {
-    struct ggml_tensor * res = ggml_mul_mat(ctx0, w, cur);
-    for (auto & it : lctx.lora) {
-        struct llama_adapter_lora_weight * lw = it.first->get_weight(w);
-        if (lw == nullptr) {
-            continue;
-        }
-        const float adapter_scale = it.second;
-        const float scale = lw->get_scale(it.first->alpha, adapter_scale);
-        struct ggml_tensor * ab_cur = ggml_mul_mat(
-            ctx0, lw->b,
-            ggml_mul_mat(ctx0, lw->a, cur)
-        );
-        ab_cur = ggml_scale(ctx0, ab_cur, scale);
-        res = ggml_add(ctx0, res, ab_cur);
-    }
-    return res;
-}
-
-// do mat_mul_id, while optionally apply lora
-static struct ggml_tensor * llm_build_lora_mm_id(
-        struct llama_context & lctx,
-         struct ggml_context * ctx0,
-          struct ggml_tensor * w,   // struct ggml_tensor * as
-          struct ggml_tensor * cur, // struct ggml_tensor * b
-          struct ggml_tensor * ids) {
-    struct ggml_tensor * res = ggml_mul_mat_id(ctx0, w, cur, ids);
-    for (auto & it : lctx.lora) {
-        struct llama_adapter_lora_weight * lw = it.first->get_weight(w);
-        if (lw == nullptr) {
-            continue;
-        }
-        const float alpha = it.first->alpha;
-        const float rank  = (float) lw->b->ne[0];
-        const float scale = alpha ? it.second * alpha / rank : it.second;
-        struct ggml_tensor * ab_cur = ggml_mul_mat_id(
-            ctx0, lw->b,
-            ggml_mul_mat_id(ctx0, lw->a, cur, ids),
-            ids
-        );
-        ab_cur = ggml_scale(ctx0, ab_cur, scale);
-        res = ggml_add(ctx0, res, ab_cur);
-    }
-    return res;
-}
-
-static struct ggml_tensor * llm_build_norm(
-        struct ggml_context * ctx,
-         struct ggml_tensor * cur,
-        const llama_hparams & hparams,
-         struct ggml_tensor * mw,
-         struct ggml_tensor * mb,
-              llm_norm_type   type,
-         const llm_build_cb & cb,
-                        int   il) {
-    switch (type) {
-        case LLM_NORM:       cur = ggml_norm      (ctx, cur, hparams.f_norm_eps);     break;
-        case LLM_NORM_RMS:   cur = ggml_rms_norm  (ctx, cur, hparams.f_norm_rms_eps); break;
-        case LLM_NORM_GROUP:
-            {
-                cur = ggml_reshape_3d(ctx, cur, cur->ne[0], 1, cur->ne[1]);
-                cur = ggml_group_norm(ctx, cur, hparams.n_norm_groups, hparams.f_norm_group_eps);
-                cur = ggml_reshape_2d(ctx, cur, cur->ne[0],    cur->ne[2]);
-            } break;
-    }
-
-    if (mw || mb) {
-        cb(cur, "norm", il);
-    }
-
-    if (mw) {
-        cur = ggml_mul(ctx, cur, mw);
-        if (mb) {
-            cb(cur, "norm_w", il);
-        }
-    }
-
-    if (mb) {
-        cur = ggml_add(ctx, cur, mb);
-    }
-
-    return cur;
-}
-
-static struct ggml_tensor * llm_build_ffn(
-        struct ggml_context * ctx,
-       struct llama_context & lctx,
-         struct ggml_tensor * cur,
-         struct ggml_tensor * up,
-         struct ggml_tensor * up_b,
-         struct ggml_tensor * up_s,
-         struct ggml_tensor * gate,
-         struct ggml_tensor * gate_b,
-         struct ggml_tensor * gate_s,
-         struct ggml_tensor * down,
-         struct ggml_tensor * down_b,
-         struct ggml_tensor * down_s,
-         struct ggml_tensor * act_scales,
-            llm_ffn_op_type   type_op,
-          llm_ffn_gate_type   type_gate,
-         const llm_build_cb & cb,
-                        int   il) {
-    struct ggml_tensor * tmp = up ? llm_build_lora_mm(lctx, ctx, up, cur) : cur;
-    cb(tmp, "ffn_up", il);
-
-    if (up_b) {
-        tmp = ggml_add(ctx, tmp, up_b);
-        cb(tmp, "ffn_up_b", il);
-    }
-
-    if (up_s) {
-        tmp = ggml_mul(ctx, tmp, up_s);
-        cb(tmp, "ffn_up_s", il);
-    }
-
-    if (gate) {
-        switch (type_gate) {
-            case LLM_FFN_SEQ:
-                {
-                    cur = llm_build_lora_mm(lctx, ctx, gate, tmp);
-                    cb(cur, "ffn_gate", il);
-                } break;
-            case LLM_FFN_PAR:
-                {
-                    cur = llm_build_lora_mm(lctx, ctx, gate, cur);
-                    cb(cur, "ffn_gate", il);
-                } break;
-        }
-
-        if (gate_b) {
-            cur = ggml_add(ctx, cur, gate_b);
-            cb(cur, "ffn_gate_b", il);
-        }
-
-        if (gate_s) {
-            cur = ggml_mul(ctx, cur, gate_s);
-            cb(cur, "ffn_gate_s", il);
-        }
-
-    } else {
-        cur = tmp;
-    }
-
-    switch (type_op) {
-        case LLM_FFN_SILU:
-            {
-                cur = ggml_silu(ctx, cur);
-                cb(cur, "ffn_silu", il);
-            } break;
-        case LLM_FFN_GELU:
-            {
-                cur = ggml_gelu(ctx, cur);
-                cb(cur, "ffn_gelu", il);
-                if (act_scales != NULL) {
-                    cur = ggml_div(ctx, cur, act_scales);
-                    cb(cur, "ffn_act", il);
-                }
-            } break;
-        case LLM_FFN_RELU:
-            {
-                cur = ggml_relu(ctx, cur);
-                cb(cur, "ffn_relu", il);
-            } break;
-        case LLM_FFN_RELU_SQR:
-            {
-                cur = ggml_relu(ctx, cur);
-                cb(cur, "ffn_relu", il);
-
-                cur = ggml_sqr(ctx, cur);
-                cb(cur, "ffn_sqr(relu)", il);
-            } break;
-        case LLM_FFN_SWIGLU:
-            {
-                // Project to 4h. If using swiglu double the output width, see https://arxiv.org/pdf/2002.05202.pdf
-                int64_t split_point = cur->ne[0] / 2;
-                struct ggml_tensor * x0 = ggml_cont(ctx, ggml_view_2d(ctx, cur, split_point, cur->ne[1], cur->nb[1], 0));
-                struct ggml_tensor * x1 = ggml_cont(ctx, ggml_view_2d(ctx, cur, split_point, cur->ne[1], cur->nb[1], split_point * ggml_element_size(cur)));
-
-                x0 = ggml_silu(ctx, x0);
-                cb(cur, "ffn_silu", il);
-
-                cur = ggml_mul(ctx, x0, x1);
-                cb(cur, "ffn_mul", il);
-            } break;
-    }
-
-    if (type_gate == LLM_FFN_PAR) {
-        cur = ggml_mul(ctx, cur, tmp);
-        cb(cur, "ffn_gate_par", il);
-    }
-
-    if (down) {
-        cur = llm_build_lora_mm(lctx, ctx, down, cur);
-    }
-
-    if (down_b) {
-        cb(cur, "ffn_down", il);
-    }
-
-    if (down_b) {
-        cur = ggml_add(ctx, cur, down_b);
-    }
-
-    if (down_s) {
-        cur = ggml_mul(ctx, cur, down_s);
-        cb(cur, "ffn_down_s", il);
-    }
-
-    return cur;
-}
-
-static struct ggml_tensor * llm_build_moe_ffn(
-        struct ggml_context * ctx,
-       struct llama_context & lctx,
-         struct ggml_tensor * cur,
-         struct ggml_tensor * gate_inp,
-         struct ggml_tensor * up_exps,
-         struct ggml_tensor * gate_exps,
-         struct ggml_tensor * down_exps,
-         struct ggml_tensor * exp_probs_b,
-                    int64_t   n_expert,
-                    int64_t   n_expert_used,
-            llm_ffn_op_type   type_op,
-                       bool   norm_w,
-                       bool   scale_w,
-                      float   w_scale,
-llama_expert_gating_func_type gating_op,
-         const llm_build_cb & cb,
-                        int   il) {
-    int64_t n_embd = cur->ne[0];
-    int64_t n_tokens = cur->ne[1];
-
-    ggml_tensor * logits = llm_build_lora_mm(lctx, ctx, gate_inp, cur); // [n_expert, n_tokens]
-    cb(logits, "ffn_moe_logits", il);
-
-    ggml_tensor * probs = nullptr;
-    switch (gating_op) {
-        case LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX:
-            {
-                probs = ggml_soft_max(ctx, logits); // [n_expert, n_tokens]
-            } break;
-        case LLAMA_EXPERT_GATING_FUNC_TYPE_SIGMOID:
-            {
-                probs = ggml_sigmoid(ctx, logits); // [n_expert, n_tokens]
-            } break;
-        default:
-            GGML_ABORT("fatal error");
-    }
-    cb(probs, "ffn_moe_probs", il);
-
-    // add experts selection bias - introduced in DeepSeek V3
-    // leave probs unbiased as it's later used to get expert weights
-    ggml_tensor * selection_probs = probs;
-    if (exp_probs_b != nullptr) {
-        selection_probs = ggml_add(ctx, probs, exp_probs_b);
-        cb(selection_probs, "ffn_moe_probs_biased", il);
-    }
-
-    // select experts
-    ggml_tensor * selected_experts = ggml_top_k(ctx, selection_probs, n_expert_used); // [n_expert_used, n_tokens]
-    cb(selected_experts->src[0], "ffn_moe_argsort", il);
-    cb(selected_experts, "ffn_moe_topk", il);
-
-    ggml_tensor * weights = ggml_get_rows(ctx,
-            ggml_reshape_3d(ctx, probs, 1, n_expert, n_tokens), selected_experts); // [1, n_expert_used, n_tokens]
-    cb(weights, "ffn_moe_weights", il);
-
-    if (norm_w) {
-        weights = ggml_reshape_2d(ctx, weights, n_expert_used, n_tokens);
-
-        ggml_tensor * weights_sum = ggml_sum_rows(ctx, weights); // [1, n_tokens]
-        cb(weights_sum, "ffn_moe_weights_sum", il);
-
-        weights = ggml_div(ctx, weights, weights_sum); // [n_expert_used, n_tokens]
-        cb(weights, "ffn_moe_weights_norm", il);
-
-        weights = ggml_reshape_3d(ctx, weights, 1, n_expert_used, n_tokens);
-    }
-    if (scale_w) {
-        weights = ggml_scale(ctx, weights, w_scale);
-        cb(weights, "ffn_moe_weights_scaled", il);
-    }
-
-    cur = ggml_reshape_3d(ctx, cur, n_embd, 1, n_tokens);
-    ggml_tensor * up = llm_build_lora_mm_id(lctx, ctx, up_exps, cur, selected_experts); // [n_ff, n_expert_used, n_tokens]
-    cb(up, "ffn_moe_up", il);
-
-    ggml_tensor * gate = llm_build_lora_mm_id(lctx, ctx, gate_exps, cur, selected_experts); // [n_ff, n_expert_used, n_tokens]
-    cb(gate, "ffn_moe_gate", il);
-
-    switch (type_op) {
-        case LLM_FFN_SILU:
-            {
-                gate = ggml_silu(ctx, gate);
-                cb(gate, "ffn_moe_silu", il);
-            } break;
-        case LLM_FFN_GELU:
-            {
-                gate = ggml_gelu(ctx, gate);
-                cb(gate, "ffn_moe_gelu", il);
-            } break;
-        default:
-            GGML_ABORT("fatal error");
-    }
-
-    ggml_tensor * par = ggml_mul(ctx, up, gate); // [n_ff, n_expert_used, n_tokens]
-    cb(par, "ffn_moe_gate_par", il);
-
-    ggml_tensor * experts = llm_build_lora_mm_id(lctx, ctx, down_exps, par, selected_experts); // [n_embd, n_expert_used, n_tokens]
-    cb(experts, "ffn_moe_down", il);
-
-    experts = ggml_mul(ctx, experts, weights);
-
-    // aggregate experts
-    ggml_tensor * moe_out = nullptr;
-    for (int i = 0; i < n_expert_used; ++i) {
-        ggml_tensor * cur_expert = ggml_view_2d(ctx, experts, n_embd, n_tokens,
-                experts->nb[2], i*experts->nb[1]);
-
-        if (i == 0) {
-            moe_out = cur_expert;
-        } else {
-            moe_out = ggml_add(ctx, moe_out, cur_expert);
-        }
-    }
-
-    if (n_expert_used == 1) {
-        // avoid returning a non-contiguous tensor
-        moe_out = ggml_cont(ctx, moe_out);
-    }
-
-    return moe_out;
-}
-
-static struct ggml_tensor * llm_build_kqv(
-        struct ggml_context * ctx,
-       struct llama_context & lctx,
-       const llama_kv_cache & kv,
-         struct ggml_cgraph * graph,
-         struct ggml_tensor * wo,
-         struct ggml_tensor * wo_b,
-         struct ggml_tensor * q_cur,
-         struct ggml_tensor * kq_mask,
-                    int32_t   n_tokens,
-                    int32_t   n_kv,
-                    float     kq_scale,
-         const llm_build_cb & cb,
-                    int       il) {
-    const llama_model   & model   = lctx.model;
-    const llama_hparams & hparams = lctx.model.hparams;
-    const llama_cparams & cparams = lctx.cparams;
-
-    const int64_t n_ctx         = cparams.n_ctx;
-    const int64_t n_head        = hparams.n_head(il);
-    const int64_t n_head_kv     = hparams.n_head_kv(il);
-    const int64_t n_embd_head_k = hparams.n_embd_head_k;
-    const int64_t n_embd_k_gqa  = hparams.n_embd_k_gqa(il);
-    const int64_t n_embd_head_v = hparams.n_embd_head_v;
-    const int64_t n_embd_v_gqa  = hparams.n_embd_v_gqa(il);
-
-    struct ggml_tensor * q = ggml_permute(ctx, q_cur, 0, 2, 1, 3);
-    cb(q, "q", il);
-
-    struct ggml_tensor * k =
-        ggml_view_3d(ctx, kv.k_l[il],
-                n_embd_head_k, n_kv, n_head_kv,
-                ggml_row_size(kv.k_l[il]->type, n_embd_k_gqa),
-                ggml_row_size(kv.k_l[il]->type, n_embd_head_k),
-                0);
-    cb(k, "k", il);
-
-    struct ggml_tensor * cur;
-
-    if (cparams.flash_attn) {
-        GGML_UNUSED(model);
-        GGML_UNUSED(n_ctx);
-
-        // split cached v into n_head heads (not transposed)
-        struct ggml_tensor * v =
-            ggml_view_3d(ctx, kv.v_l[il],
-                    n_embd_head_v, n_kv, n_head_kv,
-                    ggml_row_size(kv.v_l[il]->type, n_embd_v_gqa),
-                    ggml_row_size(kv.v_l[il]->type, n_embd_head_v),
-                    0);
-        cb(v, "v", il);
-
-        cur = ggml_flash_attn_ext(ctx, q, k, v, kq_mask, kq_scale, hparams.f_max_alibi_bias,
-                                  hparams.attn_soft_cap ? hparams.f_attn_logit_softcapping : 0.0f);
-
-        ggml_flash_attn_ext_set_prec(cur, GGML_PREC_F32);
-
-        cur = ggml_reshape_2d(ctx, cur, n_embd_head_v*n_head, n_tokens);
-    } else {
-        struct ggml_tensor * kq = ggml_mul_mat(ctx, k, q);
-        cb(kq, "kq", il);
-
-        // note: this op tends to require high floating point range
-        //       while for some models F16 is enough, for others it is not, so we default to F32 here
-        ggml_mul_mat_set_prec(kq, GGML_PREC_F32);
-
-        if (model.arch == LLM_ARCH_GROK) {
-            // need to do the following:
-            // multiply by attn_output_multiplyer of 0.08838834764831845
-            // and then :
-            // kq = 30 * tanh(kq / 30)
-            // before the softmax below
-
-            kq = ggml_tanh(ctx, ggml_scale(ctx, kq, 0.08838834764831845f/30.0f));
-            kq = ggml_scale(ctx, kq, 30);
-        }
-
-        if (hparams.attn_soft_cap) {
-            kq = ggml_scale(ctx, kq, 1.0f / hparams.f_attn_logit_softcapping);
-            kq = ggml_tanh(ctx, kq);
-            kq = ggml_scale(ctx, kq, hparams.f_attn_logit_softcapping);
-        }
-
-        kq = ggml_soft_max_ext(ctx, kq, kq_mask, kq_scale, hparams.f_max_alibi_bias);
-        cb(kq, "kq_soft_max_ext", il);
-
-        GGML_ASSERT(kv.size == n_ctx);
-
-        // split cached v into n_head heads
-        struct ggml_tensor * v =
-            ggml_view_3d(ctx, kv.v_l[il],
-                    n_kv, n_embd_head_v, n_head_kv,
-                    ggml_element_size(kv.v_l[il])*n_ctx,
-                    ggml_element_size(kv.v_l[il])*n_ctx*n_embd_head_v,
-                    0);
-        cb(v, "v", il);
-
-        struct ggml_tensor * kqv = ggml_mul_mat(ctx, v, kq);
-        cb(kqv, "kqv", il);
-
-        struct ggml_tensor * kqv_merged = ggml_permute(ctx, kqv, 0, 2, 1, 3);
-        cb(kqv_merged, "kqv_merged", il);
-
-        cur = ggml_cont_2d(ctx, kqv_merged, n_embd_head_v*n_head, n_tokens);
-        cb(cur, "kqv_merged_cont", il);
-    }
-
-    ggml_build_forward_expand(graph, cur);
-
-    if (wo) {
-        cur = llm_build_lora_mm(lctx, ctx, wo, cur);
-    }
-
-    if (wo_b) {
-        cb(cur, "kqv_wo", il);
-    }
-
-    if (wo_b) {
-        cur = ggml_add(ctx, cur, wo_b);
-    }
-
-    return cur;
-}
-
-static struct ggml_tensor * llm_build_kv(
-        struct ggml_context * ctx,
-       struct llama_context & lctx,
-       const llama_kv_cache & kv,
-         struct ggml_cgraph * graph,
-         struct ggml_tensor * wo,
-         struct ggml_tensor * wo_b,
-         struct ggml_tensor * k_cur,
-         struct ggml_tensor * v_cur,
-         struct ggml_tensor * q_cur,
-         struct ggml_tensor * kq_mask,
-                    int32_t   n_tokens,
-                    int32_t   kv_head,
-                    int32_t   n_kv,
-                    float     kq_scale,
-         const llm_build_cb & cb,
-                    int       il) {
-    const llama_hparams & hparams = lctx.model.hparams;
-    const llama_cparams & cparams = lctx.cparams;
-
-    // these nodes are added to the graph together so that they are not reordered
-    // by doing so, the number of splits in the graph is reduced
-    ggml_build_forward_expand(graph, q_cur);
-    ggml_build_forward_expand(graph, k_cur);
-    ggml_build_forward_expand(graph, v_cur);
-
-    llm_build_kv_store(ctx, hparams, cparams, kv, graph, k_cur, v_cur, n_tokens, kv_head, cb, il);
-
-    struct ggml_tensor * cur;
-
-    cur  = llm_build_kqv(ctx, lctx, kv, graph, wo, wo_b, q_cur, kq_mask, n_tokens, n_kv, kq_scale, cb, il);
-    cb(cur, "kqv_out", il);
-
-    return cur;
-}
-
-static struct ggml_tensor * llm_build_copy_mask_state(
-        struct ggml_context * ctx,
-         struct ggml_cgraph * graph,
-         struct ggml_tensor * s,
-         struct ggml_tensor * state_copy,
-         struct ggml_tensor * state_mask,
-                    int32_t   n_state,
-                    int32_t   kv_size,
-                    int32_t   kv_head,
-                    int32_t   n_kv,
-                    int32_t   n_seqs) {
-    struct ggml_tensor * states = ggml_reshape_2d(ctx, s, n_state, kv_size);
-
-    // copy states
-    // NOTE: assuming the copy destinations are ALL contained between kv_head and kv_head + n_kv
-    // this shrinks the tensors's ne[1] to n_kv
-    states = ggml_get_rows(ctx, states, state_copy);
-
-    // clear states of sequences which are starting at the beginning of this batch
-    // FIXME: zero-out NANs?
-    states = ggml_mul(ctx, states, state_mask);
-
-    // copy states which won't be changed further (between n_seqs and n_kv)
-    ggml_build_forward_expand(graph,
-        ggml_cpy(ctx,
-            ggml_view_1d(ctx, states, n_state*(n_kv - n_seqs), n_seqs*n_state*ggml_element_size(states)),
-            ggml_view_1d(ctx, s, n_state*(n_kv - n_seqs), (kv_head + n_seqs)*n_state*ggml_element_size(s))));
-
-    // the part of the states that will be used and modified
-    return ggml_view_2d(ctx, states, n_state, n_seqs, states->nb[1], 0);
-}
-
-// TODO: split
-static struct ggml_tensor * llm_build_mamba(
-        struct ggml_context * ctx,
-       struct llama_context & lctx,
-         const llama_ubatch & ubatch,
-         struct ggml_cgraph * graph,
-         struct ggml_tensor * cur,
-         struct ggml_tensor * state_copy,
-         struct ggml_tensor * state_mask,
-                    int32_t   kv_head,
-                    int32_t   n_kv,
-         const llm_build_cb & cb,
-                    int       il) {
-    const llama_model    & model   = lctx.model;
-    const llama_hparams  & hparams = model.hparams;
-    const llama_kv_cache & kv      = lctx.kv_self;
-    const int64_t d_conv  = hparams.ssm_d_conv;
-    const int64_t d_inner = hparams.ssm_d_inner;
-    const int64_t d_state = hparams.ssm_d_state;
-    const int64_t dt_rank = hparams.ssm_dt_rank;
-    const int64_t n_seqs  = ubatch.n_seqs;
-    // Some variants of Mamba arch (e.g. FalconMamba do apply layer norm on B and Dt layers)
-    const bool ssm_dt_b_c_rms = hparams.ssm_dt_b_c_rms;
-    // Use the same RMS norm as the final layer norm
-    const float norm_rms_eps = hparams.f_norm_rms_eps;
-
-    const int64_t n_seq_tokens = ubatch.n_seq_tokens;
-
-    GGML_ASSERT(n_seqs != 0);
-    GGML_ASSERT(ubatch.equal_seqs);
-    GGML_ASSERT(ubatch.n_tokens == n_seq_tokens * n_seqs);
-
-    struct ggml_tensor * conv_states_all = kv.k_l[il];
-    struct ggml_tensor * ssm_states_all  = kv.v_l[il];
-
-    // (ab)using the KV cache to store the states
-    struct ggml_tensor * conv = llm_build_copy_mask_state(ctx,
-            graph, conv_states_all, state_copy, state_mask,
-            hparams.n_embd_k_s(), kv.size, kv_head, n_kv, n_seqs);
-    conv = ggml_reshape_3d(ctx, conv, d_conv - 1, d_inner, n_seqs);
-    struct ggml_tensor * ssm = llm_build_copy_mask_state(ctx,
-            graph, ssm_states_all, state_copy, state_mask,
-            hparams.n_embd_v_s(), kv.size, kv_head, n_kv, n_seqs);
-    ssm = ggml_reshape_3d(ctx, ssm, d_state, d_inner, n_seqs);
-
-    // {n_embd, n_tokens} => {n_embd, n_seq_tokens, n_seqs}
-    cur = ggml_reshape_3d(ctx, cur, cur->ne[0], n_seq_tokens, n_seqs);
-
-    // {n_embd, 2*d_inner} @ {n_embd, n_seq_tokens, n_seqs} => {2*d_inner, n_seq_tokens, n_seqs}
-    struct ggml_tensor * xz = llm_build_lora_mm(lctx, ctx, model.layers[il].ssm_in, cur);
-    // split the above in two
-    // => {d_inner, n_seq_tokens, n_seqs}
-    struct ggml_tensor * x = ggml_view_3d(ctx, xz, d_inner, xz->ne[1], xz->ne[2], xz->nb[1], xz->nb[2], 0);
-    struct ggml_tensor * z = ggml_view_3d(ctx, xz, d_inner, xz->ne[1], xz->ne[2], xz->nb[1], xz->nb[2], d_inner*ggml_element_size(xz));
-
-    // conv
-    {
-        // => {d_conv - 1 + n_seq_tokens, d_inner, n_seqs}
-        struct ggml_tensor * conv_x = ggml_concat(ctx, conv, ggml_transpose(ctx, x), 0);
-
-        // copy last (d_conv - 1) columns back into the state cache
-        struct ggml_tensor * last_conv = ggml_view_3d(ctx, conv_x, d_conv - 1, d_inner, n_seqs, conv_x->nb[1], conv_x->nb[2], n_seq_tokens*(conv_x->nb[0]));
-
-        ggml_build_forward_expand(graph,
-            ggml_cpy(ctx, last_conv,
-                ggml_view_1d(ctx, conv_states_all,
-                    (d_conv - 1)*(d_inner)*(n_seqs),
-                    kv_head*(d_conv - 1)*(d_inner)*ggml_element_size(conv_states_all))));
-
-        // 1D convolution
-        // The equivalent is to make a self-overlapping view of conv_x
-        // over d_conv columns at each stride in the 3rd dimension,
-        // then element-wise multiply that with the conv1d weight,
-        // then sum the elements of each row,
-        // (the last two steps are a dot product over rows (also doable with mul_mat))
-        // then permute away the ne[0] dimension,
-        // and then you're left with the resulting x tensor.
-        // For simultaneous sequences, all sequences need to have the same length.
-        x = ggml_ssm_conv(ctx, conv_x, model.layers[il].ssm_conv1d);
-
-        // bias
-        x = ggml_add(ctx, x, model.layers[il].ssm_conv1d_b);
-
-        x = ggml_silu(ctx, x);
-    }
-
-    // ssm
-    {
-        // {d_inner, dt_rank + 2*d_state} @ {d_inner, n_seq_tokens, n_seqs} => {dt_rank + 2*d_state, n_seq_tokens, n_seqs}
-        struct ggml_tensor * x_db = llm_build_lora_mm(lctx, ctx, model.layers[il].ssm_x, x);
-        // split
-        struct ggml_tensor * dt = ggml_view_3d(ctx, x_db, dt_rank, n_seq_tokens, n_seqs, x_db->nb[1], x_db->nb[2], 0);
-        struct ggml_tensor * B  = ggml_view_3d(ctx, x_db, d_state, n_seq_tokens, n_seqs, x_db->nb[1], x_db->nb[2], ggml_element_size(x_db)*dt_rank);
-        struct ggml_tensor * C  = ggml_view_3d(ctx, x_db, d_state, n_seq_tokens, n_seqs, x_db->nb[1], x_db->nb[2], ggml_element_size(x_db)*(dt_rank+d_state));
-
-        // Some Mamba variants (e.g. FalconMamba) apply RMS norm in B, C & Dt layers
-        if (ssm_dt_b_c_rms) {
-            dt = ggml_rms_norm(ctx, dt, norm_rms_eps);
-            B = ggml_rms_norm(ctx, B, norm_rms_eps);
-            C = ggml_rms_norm(ctx, C, norm_rms_eps);
-        }
-
-        // {dt_rank, d_inner} @ {dt_rank, n_seq_tokens, n_seqs} => {d_inner, n_seq_tokens, n_seqs}
-        dt = llm_build_lora_mm(lctx, ctx, model.layers[il].ssm_dt, dt);
-        dt = ggml_add(ctx, dt, model.layers[il].ssm_dt_b);
-
-        // Custom operator to optimize the parallel associative scan
-        // as described in the Annex D of the Mamba paper.
-        // => {d_inner, n_seq_tokens, n_seqs} and {d_state, d_inner, n_seqs}
-        struct ggml_tensor * y_ssm = ggml_ssm_scan(ctx, ssm, x, dt, model.layers[il].ssm_a, B, C);
-
-        // store last states
-        ggml_build_forward_expand(graph,
-            ggml_cpy(ctx,
-                ggml_view_1d(ctx, y_ssm, d_state*d_inner*n_seqs, x->nb[3]),
-                ggml_view_1d(ctx, ssm_states_all, d_state*d_inner*n_seqs, kv_head*d_state*d_inner*ggml_element_size(ssm_states_all))));
-
-        struct ggml_tensor * y = ggml_view_3d(ctx, y_ssm, d_inner, n_seq_tokens, n_seqs, x->nb[1], x->nb[2], 0);
-
-        // TODO: skip computing output earlier for unused tokens
-
-        // {d_inner, n_seq_tokens, n_seqs} * {d_inner} => {d_inner, n_seq_tokens, n_seqs}
-        y = ggml_add(ctx, y, ggml_mul(ctx, x, model.layers[il].ssm_d));
-        y = ggml_mul(ctx, y, ggml_silu(ctx, ggml_cont(ctx, z)));
-
-        // {d_inner, n_embd} @ {d_inner, n_seq_tokens, n_seqs} => {n_embd, n_seq_tokens, n_seqs}
-        cur = llm_build_lora_mm(lctx, ctx, model.layers[il].ssm_out, y);
-    }
-
-    // {n_embd, n_seq_tokens, n_seqs} => {n_embd, n_tokens}
-    cur = ggml_reshape_2d(ctx, cur, cur->ne[0], n_seq_tokens * n_seqs);
-    cb(cur, "mamba_out", il);
-
-    return cur;
-}
-
-static struct ggml_tensor * llm_build_rwkv6_time_mix(
-        struct llama_context & lctx,
-        struct ggml_context * ctx,
-        const struct llama_layer * layer,
-        struct ggml_tensor * cur,
-        struct ggml_tensor * x_prev,
-        struct ggml_tensor ** wkv_state,
-        size_t wkv_head_size,
-        size_t head_count_kv) {
-    size_t n_embd       = cur->ne[0];
-    size_t n_seq_tokens = cur->ne[1];
-    size_t n_seqs       = cur->ne[2];
-
-    size_t head_size  = wkv_head_size;
-    size_t head_count = n_embd / head_size;
-
-    size_t n_tokens = n_seqs * n_seq_tokens;
-
-    bool is_qrwkv = layer->time_mix_first == nullptr;
-
-    struct ggml_tensor * sx = ggml_sub(ctx, x_prev, cur);
-
-    sx  = ggml_reshape_2d(ctx, sx,  n_embd, n_tokens);
-    cur = ggml_reshape_2d(ctx, cur, n_embd, n_tokens);
-
-    struct ggml_tensor * xxx = ggml_add(ctx, ggml_mul(ctx, sx, layer->time_mix_lerp_x), cur);
-
-    xxx = ggml_reshape_4d(
-        ctx,
-        ggml_tanh(
-            ctx,
-            ggml_mul_mat(ctx, layer->time_mix_w1, xxx)
-        ),
-        layer->time_mix_w1->ne[1] / 5, 1, 5, n_tokens
-    );
-
-    xxx = ggml_cont(ctx, ggml_permute(ctx, xxx, 0, 1, 3, 2));
-
-    xxx = ggml_mul_mat(
-        ctx,
-        ggml_reshape_4d(
-            ctx,
-            layer->time_mix_w2,
-            layer->time_mix_w2->ne[0], layer->time_mix_w2->ne[1], 1, 5
-        ),
-        xxx
-    );
-
-    struct ggml_tensor *xw, *xk, *xv, *xr, *xg;
-    if (layer->time_mix_lerp_fused) {
-        // fusing these weights makes some performance improvement
-        sx  = ggml_reshape_3d(ctx, sx,  n_embd, 1, n_tokens);
-        cur = ggml_reshape_3d(ctx, cur, n_embd, 1, n_tokens);
-        xxx = ggml_add(ctx, ggml_mul(ctx, ggml_add(ctx, xxx, layer->time_mix_lerp_fused), sx), cur);
-        xw = ggml_view_2d(ctx, xxx, n_embd, n_tokens, xxx->nb[1], 0);
-        xk = ggml_view_2d(ctx, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * sizeof(float));
-        xv = ggml_view_2d(ctx, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * 2 * sizeof(float));
-        xr = ggml_view_2d(ctx, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * 3 * sizeof(float));
-        xg = ggml_view_2d(ctx, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * 4 * sizeof(float));
-    } else {
-        // for backward compatibility
-        xw = ggml_view_2d(ctx, xxx, n_embd, n_tokens, xxx->nb[1], 0);
-        xk = ggml_view_2d(ctx, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * sizeof(float));
-        xv = ggml_view_2d(ctx, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * 2 * sizeof(float));
-        xr = ggml_view_2d(ctx, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * 3 * sizeof(float));
-        xg = ggml_view_2d(ctx, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * 4 * sizeof(float));
-
-        xw = ggml_add(ctx, ggml_mul(ctx, ggml_add(ctx, xw, layer->time_mix_lerp_w), sx), cur);
-        xk = ggml_add(ctx, ggml_mul(ctx, ggml_add(ctx, xk, layer->time_mix_lerp_k), sx), cur);
-        xv = ggml_add(ctx, ggml_mul(ctx, ggml_add(ctx, xv, layer->time_mix_lerp_v), sx), cur);
-        xr = ggml_add(ctx, ggml_mul(ctx, ggml_add(ctx, xr, layer->time_mix_lerp_r), sx), cur);
-        xg = ggml_add(ctx, ggml_mul(ctx, ggml_add(ctx, xg, layer->time_mix_lerp_g), sx), cur);
-    }
-
-    struct ggml_tensor * r = llm_build_lora_mm(lctx, ctx, layer->time_mix_receptance, xr);
-    struct ggml_tensor * k = llm_build_lora_mm(lctx, ctx, layer->time_mix_key,        xk);
-    struct ggml_tensor * v = llm_build_lora_mm(lctx, ctx, layer->time_mix_value,      xv);
-    if (layer->time_mix_receptance_b) {
-        r = ggml_add(ctx, r, layer->time_mix_receptance_b);
-    }
-    if (layer->time_mix_key_b) {
-        k = ggml_add(ctx, k, layer->time_mix_key_b);
-    }
-    if (layer->time_mix_value_b) {
-        v = ggml_add(ctx, v, layer->time_mix_value_b);
-    }
-
-    struct ggml_tensor * g = llm_build_lora_mm(lctx, ctx, layer->time_mix_gate, xg);
-    if (is_qrwkv) {
-        g = ggml_sigmoid(ctx, g);
-    } else {
-        g = ggml_silu(ctx, g);
-    }
-
-    if (head_count_kv != head_count) {
-        GGML_ASSERT(head_count % head_count_kv == 0);
-        k = ggml_reshape_4d(ctx, k, head_size, 1, head_count_kv, n_tokens);
-        v = ggml_reshape_4d(ctx, v, head_size, 1, head_count_kv, n_tokens);
-        struct ggml_tensor * tmp = ggml_new_tensor_4d(ctx, GGML_TYPE_F32, head_size, head_count / head_count_kv, head_count_kv, n_tokens);
-        k = ggml_repeat(ctx, k, tmp);
-        v = ggml_repeat(ctx, v, tmp);
-    }
-
-    k = ggml_reshape_3d(ctx, k, head_size, head_count, n_tokens);
-    v = ggml_reshape_3d(ctx, v, head_size, head_count, n_tokens);
-    r = ggml_reshape_3d(ctx, r, head_size, head_count, n_tokens);
-
-    struct ggml_tensor * w = ggml_mul_mat(
-        ctx,
-        layer->time_mix_decay_w2,
-        ggml_tanh(
-            ctx,
-            ggml_mul_mat(ctx, layer->time_mix_decay_w1, xw)
-        )
-    );
-
-    w = ggml_add(ctx, w, layer->time_mix_decay);
-    w = ggml_exp(ctx, ggml_neg(ctx, ggml_exp(ctx, w)));
-    w = ggml_reshape_3d(ctx, w, head_size, head_count, n_tokens);
-
-    if (is_qrwkv) {
-        // k = k * (1 - w)
-        k = ggml_sub(ctx, k, ggml_mul(ctx, k, w));
-    }
-
-    struct ggml_tensor * wkv_output;
-    if (!layer->time_mix_first) {
-        wkv_output = ggml_gated_linear_attn(ctx, k, v, r, w, *wkv_state, pow(head_size, -0.5f));
-    } else {
-        wkv_output = ggml_rwkv_wkv6(ctx, k, v, r, layer->time_mix_first, w, *wkv_state);
-    }
-    cur = ggml_view_1d(ctx, wkv_output, n_embd * n_tokens, 0);
-    *wkv_state = ggml_view_1d(ctx, wkv_output, n_embd * head_size * n_seqs, n_embd * n_tokens * sizeof(float));
-
-    if (!is_qrwkv) {
-        // group norm with head_count groups
-        cur = ggml_reshape_3d(ctx, cur, n_embd / head_count, head_count, n_tokens);
-        cur = ggml_norm(ctx, cur, 64e-5f);
-
-        // Convert back to regular vectors.
-        cur = ggml_reshape_2d(ctx, cur, n_embd, n_tokens);
-        cur = ggml_add(ctx, ggml_mul(ctx, cur, layer->time_mix_ln), layer->time_mix_ln_b);
-    } else {
-        cur = ggml_reshape_2d(ctx, cur, n_embd, n_tokens);
-    }
-
-    cur = ggml_mul(ctx, cur, g);
-    cur = llm_build_lora_mm(lctx, ctx, layer->time_mix_output, cur);
-
-    return ggml_reshape_3d(ctx, cur, n_embd, n_seq_tokens, n_seqs);
-}
-
-static struct ggml_tensor * llm_build_rwkv6_channel_mix(
-        struct llama_context & lctx,
-        struct ggml_context * ctx,
-        const struct llama_layer * layer,
-        struct ggml_tensor * cur,
-        struct ggml_tensor * x_prev) {
-    struct ggml_tensor * sx = ggml_sub(ctx, x_prev, cur);
-    struct ggml_tensor * xk = ggml_add(ctx, ggml_mul(ctx, sx, layer->channel_mix_lerp_k), cur);
-    struct ggml_tensor * xr = ggml_add(ctx, ggml_mul(ctx, sx, layer->channel_mix_lerp_r), cur);
-
-    struct ggml_tensor * r = ggml_sigmoid(ctx, llm_build_lora_mm(lctx, ctx, layer->channel_mix_receptance, xr));
-    struct ggml_tensor * k = ggml_sqr(
-        ctx,
-        ggml_relu(
-            ctx,
-            llm_build_lora_mm(lctx, ctx, layer->channel_mix_key, xk)
-        )
-    );
-
-    return ggml_mul(ctx, r, llm_build_lora_mm(lctx, ctx, layer->channel_mix_value, k));
-}
-
-struct llm_build_context {
-    const llama_model    & model;
-          llama_context  & lctx;
-    const llama_hparams  & hparams;
-    const llama_cparams  & cparams;
-    const llama_ubatch   & ubatch;
-    const llama_kv_cache & kv_self;
-
-    const int64_t n_embd;
-    const int64_t n_layer;
-    const int64_t n_rot;
-    const int64_t n_ctx;       // user-specified context size (can be different from n_ctx_train)
-    const int64_t n_head;
-    const int64_t n_head_kv;
-    const int64_t n_embd_head_k;
-    const int64_t n_embd_k_gqa;
-    const int64_t n_embd_head_v;
-    const int64_t n_embd_v_gqa;
-    const int64_t n_expert;
-    const int64_t n_expert_used;
-
-    const float freq_base;
-    const float freq_scale;
-    const float ext_factor;
-    const float attn_factor;
-    const float beta_fast;
-    const float beta_slow;
-    const float norm_eps;
-    const float norm_rms_eps;
-
-    const int32_t n_tokens;
-    const int32_t n_kv;     // size of KV cache to consider (n_kv <= kv_self.size)
-    const int32_t n_outputs;
-    const int32_t n_outputs_enc;
-    const int32_t kv_head;  // index of where we store new KV data in the cache
-    const int32_t n_ctx_orig;
-
-    const bool flash_attn;
-
-    const enum llama_pooling_type pooling_type;
-    const enum llama_rope_type    rope_type;
-
-    const llm_build_cb & cb;
-
-    std::vector & buf_compute_meta;
-
-    struct ggml_context * ctx0 = nullptr;
-
-    // TODO: consider making the entire interface noexcept
-    llm_build_context(
-        llama_context  & lctx,
-    const llama_ubatch & ubatch,
-    const llm_build_cb & cb,
-                  bool   worst_case) :
-        model            (lctx.model),
-        lctx             (lctx),
-        hparams          (model.hparams),
-        cparams          (lctx.cparams),
-        ubatch           (ubatch),
-        kv_self          (lctx.kv_self),
-        n_embd           (hparams.n_embd),
-        n_layer          (hparams.n_layer),
-        n_rot            (hparams.n_rot),
-        n_ctx            (cparams.n_ctx),
-        n_head           (hparams.n_head()),
-        n_head_kv        (hparams.n_head_kv()),
-        n_embd_head_k    (hparams.n_embd_head_k),
-        n_embd_k_gqa     (hparams.n_embd_k_gqa()),
-        n_embd_head_v    (hparams.n_embd_head_v),
-        n_embd_v_gqa     (hparams.n_embd_v_gqa()),
-        n_expert         (hparams.n_expert),
-        n_expert_used    (hparams.n_expert_used),
-        freq_base        (cparams.rope_freq_base),
-        freq_scale       (cparams.rope_freq_scale),
-        ext_factor       (cparams.yarn_ext_factor),
-        attn_factor      (cparams.yarn_attn_factor),
-        beta_fast        (cparams.yarn_beta_fast),
-        beta_slow        (cparams.yarn_beta_slow),
-        norm_eps         (hparams.f_norm_eps),
-        norm_rms_eps     (hparams.f_norm_rms_eps),
-        n_tokens         (ubatch.n_tokens),
-        n_kv             (worst_case ? kv_self.size : kv_self.n),
-        n_outputs        (worst_case ? n_tokens : lctx.n_outputs),
-        n_outputs_enc    (worst_case ? n_tokens : lctx.embd_enc.size() / hparams.n_embd),
-        kv_head          (worst_case ? (kv_self.recurrent ? 0 : kv_self.size - n_tokens) : kv_self.head),
-        n_ctx_orig       (cparams.n_ctx_orig_yarn),
-        flash_attn       (cparams.flash_attn),
-        pooling_type     (cparams.pooling_type),
-        rope_type        (hparams.rope_type),
-        cb               (cb),
-        buf_compute_meta (lctx.buf_compute_meta) {
-            // all initializations should be done in init()
-        }
-
-    void init() {
-        struct ggml_init_params params = {
-            /*.mem_size   =*/ buf_compute_meta.size(),
-            /*.mem_buffer =*/ buf_compute_meta.data(),
-            /*.no_alloc   =*/ true,
-        };
-
-        ctx0 = ggml_init(params);
-
-        lctx.inp_tokens      = nullptr;
-        lctx.inp_embd        = nullptr;
-        lctx.inp_pos         = nullptr;
-        lctx.inp_out_ids     = nullptr;
-        lctx.inp_KQ_mask     = nullptr;
-        lctx.inp_KQ_mask_swa = nullptr;
-        lctx.inp_K_shift     = nullptr;
-        lctx.inp_mean        = nullptr;
-        lctx.inp_cls         = nullptr;
-        lctx.inp_s_copy      = nullptr;
-        lctx.inp_s_mask      = nullptr;
-        lctx.inp_s_seq       = nullptr;
-        lctx.inp_pos_bucket    = nullptr;
-        lctx.inp_embd_enc      = nullptr;
-        lctx.inp_KQ_mask_cross = nullptr;
-    }
-
-    void free() {
-        ggml_free(ctx0);
-        ctx0 = nullptr;
-    }
-
-    struct ggml_cgraph * build_k_shift() {
-        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false);
-
-        GGML_ASSERT(kv_self.size == n_ctx);
-
-        lctx.inp_K_shift = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_ctx);
-        cb(lctx.inp_K_shift, "K_shift", -1);
-        ggml_set_input(lctx.inp_K_shift);
-
-        for (int il = 0; il < n_layer; ++il) {
-            const int64_t n_head_kv = hparams.n_head_kv(il);
-            const int64_t n_embd_k_gqa = hparams.n_embd_k_gqa(il);
-            struct ggml_tensor * rope_factors = build_rope_factors(il);
-            struct ggml_tensor * k =
-                ggml_view_3d(ctx0, kv_self.k_l[il],
-                    n_embd_head_k, n_head_kv, n_ctx,
-                    ggml_row_size(kv_self.k_l[il]->type, n_embd_head_k),
-                    ggml_row_size(kv_self.k_l[il]->type, n_embd_k_gqa),
-                    0);
-
-            struct ggml_tensor * tmp;
-            if (ggml_is_quantized(k->type)) {
-                // dequantize to f32 -> RoPE -> quantize back
-                tmp = ggml_cast(ctx0, k, GGML_TYPE_F32);
-                cb(tmp, "K_f32", il);
-                for (auto & backend : lctx.backends) {
-                    // Figure out which backend KV cache belongs to
-                    if (ggml_backend_supports_buft(backend.get(), ggml_backend_buffer_get_type(kv_self.k_l[il]->buffer))) {
-                        ggml_backend_sched_set_tensor_backend(lctx.sched.get(), tmp, backend.get());
-                        break;
-                    }
-                }
-                tmp = ggml_rope_ext_inplace(ctx0, tmp,
-                        lctx.inp_K_shift, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow);
-                cb(tmp, "K_shifted_f32", il);
-                tmp = ggml_cpy(ctx0, tmp, k);
-            } else {
-                // we rotate only the first n_rot dimensions
-                tmp = ggml_rope_ext_inplace(ctx0, k,
-                        lctx.inp_K_shift, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow);
-            }
-            cb(tmp, "K_shifted", il);
-            ggml_build_forward_expand(gf, tmp);
-        }
-
-        return gf;
-    }
-
-    struct ggml_cgraph * build_defrag(const std::vector & ids) {
-        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false);
-
-        for (uint32_t i = 0; i < ids.size(); ++i) {
-            const uint32_t id = ids[i];
-
-            if (i == id || id == ids.size()) {
-                continue;
-            }
-
-            uint32_t nm = 1;
-
-            while (i + nm < ids.size() && ids[i + nm] == id + nm) {
-                nm++;
-            }
-
-            for (int il = 0; il < n_layer; ++il) {
-                const int64_t n_embd_k_gqa = hparams.n_embd_k_gqa(il);
-                const int64_t n_embd_v_gqa = hparams.n_embd_v_gqa(il);
-
-                ggml_tensor * view_k_src = ggml_view_2d(ctx0, kv_self.k_l[il],
-                        n_embd_k_gqa, nm,
-                        ggml_row_size(kv_self.k_l[il]->type, n_embd_k_gqa),
-                        ggml_row_size(kv_self.k_l[il]->type, n_embd_k_gqa*i));
-
-                ggml_tensor * view_k_dst = ggml_view_2d(ctx0, kv_self.k_l[il],
-                        n_embd_k_gqa, nm,
-                        ggml_row_size(kv_self.k_l[il]->type, n_embd_k_gqa),
-                        ggml_row_size(kv_self.k_l[il]->type, n_embd_k_gqa*id));
-
-                ggml_tensor * view_v_src;
-                ggml_tensor * view_v_dst;
-
-                if (flash_attn) {
-                    // NOTE: the V cache is not transposed when using flash attention
-                    view_v_src = ggml_view_2d(ctx0, kv_self.v_l[il],
-                            n_embd_v_gqa, nm,
-                            ggml_row_size(kv_self.v_l[il]->type, n_embd_v_gqa),
-                            ggml_row_size(kv_self.v_l[il]->type, n_embd_v_gqa*i));
-
-                    view_v_dst = ggml_view_2d(ctx0, kv_self.v_l[il],
-                            n_embd_v_gqa, nm,
-                            ggml_row_size(kv_self.v_l[il]->type, n_embd_v_gqa),
-                            ggml_row_size(kv_self.v_l[il]->type, n_embd_v_gqa*id));
-                } else {
-                    view_v_src = ggml_view_2d(ctx0, kv_self.v_l[il],
-                            nm, n_embd_v_gqa,
-                            ggml_row_size(kv_self.v_l[il]->type, kv_self.size),
-                            ggml_row_size(kv_self.v_l[il]->type, i));
-
-                    view_v_dst = ggml_view_2d(ctx0, kv_self.v_l[il],
-                            nm, n_embd_v_gqa,
-                            ggml_row_size(kv_self.v_l[il]->type, kv_self.size),
-                            ggml_row_size(kv_self.v_l[il]->type, id));
-                }
-
-                ggml_build_forward_expand(gf, ggml_cpy(ctx0, view_k_src, view_k_dst));
-                ggml_build_forward_expand(gf, ggml_cpy(ctx0, view_v_src, view_v_dst));
-            }
-
-            i += nm - 1;
-        }
-
-        //LLAMA_LOG_INFO("gf->n_nodes = %d\n", gf->n_nodes);
-
-        return gf;
-    }
-
-    struct ggml_tensor * build_inp_pos() {
-        lctx.inp_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens);
-        cb(lctx.inp_pos, "inp_pos", -1);
-        ggml_set_input(lctx.inp_pos);
-        return lctx.inp_pos;
-    }
-
-    struct ggml_tensor * build_rope_factors(int il) {
-        // choose long/short freq factors based on the context size
-        const auto n_ctx_pre_seq = cparams.n_ctx / cparams.n_seq_max;
-
-        if (model.layers[il].rope_freqs != nullptr) {
-            return model.layers[il].rope_freqs;
-        }
-
-        if (n_ctx_pre_seq > hparams.n_ctx_orig_yarn) {
-            return model.layers[il].rope_long;
-        }
-
-        return model.layers[il].rope_short;
-    }
-
-    struct ggml_tensor * build_inp_out_ids() {
-        lctx.inp_out_ids = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_outputs);
-        cb(lctx.inp_out_ids, "inp_out_ids", -1);
-        ggml_set_input(lctx.inp_out_ids);
-        return lctx.inp_out_ids;
-    }
-
-    struct ggml_tensor * build_inp_KQ_mask(bool causal = true) {
-        lctx.inp_KQ_mask = causal
-            ? ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_kv,     GGML_PAD(n_tokens, GGML_KQ_MASK_PAD))
-            : ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_tokens, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD));
-        cb(lctx.inp_KQ_mask, "KQ_mask", -1);
-        ggml_set_input(lctx.inp_KQ_mask);
-
-        return flash_attn ? ggml_cast(ctx0, lctx.inp_KQ_mask, GGML_TYPE_F16) : lctx.inp_KQ_mask;
-    }
-
-    struct ggml_tensor * build_inp_KQ_mask_swa(bool causal = true) {
-        GGML_ASSERT(hparams.n_swa > 0);
-
-        lctx.inp_KQ_mask_swa = causal
-            ? ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_kv,     GGML_PAD(n_tokens, GGML_KQ_MASK_PAD))
-            : ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_tokens, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD));
-        cb(lctx.inp_KQ_mask_swa, "KQ_mask_swa", -1);
-        ggml_set_input(lctx.inp_KQ_mask_swa);
-
-        return flash_attn ? ggml_cast(ctx0, lctx.inp_KQ_mask_swa, GGML_TYPE_F16) : lctx.inp_KQ_mask_swa;
-    }
-
-    struct ggml_tensor * build_inp_mean() {
-        lctx.inp_mean = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_tokens, n_tokens);
-        cb(lctx.inp_mean, "inp_mean", -1);
-        ggml_set_input(lctx.inp_mean);
-        return lctx.inp_mean;
-    }
-
-    struct ggml_tensor * build_inp_cls() {
-        lctx.inp_cls = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens);
-        cb(lctx.inp_cls, "inp_cls", -1);
-        ggml_set_input(lctx.inp_cls);
-        return lctx.inp_cls;
-    }
-
-    struct ggml_tensor * build_inp_s_copy() {
-        lctx.inp_s_copy = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_kv);
-        cb(lctx.inp_s_copy, "inp_s_copy", -1);
-        ggml_set_input(lctx.inp_s_copy);
-        return lctx.inp_s_copy;
-    }
-
-    struct ggml_tensor * build_inp_s_mask() {
-        lctx.inp_s_mask = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, 1, n_kv);
-        cb(lctx.inp_s_mask, "inp_s_mask", -1);
-        ggml_set_input(lctx.inp_s_mask);
-        return lctx.inp_s_mask;
-    }
-
-    struct ggml_cgraph * append_pooling(struct ggml_cgraph * gf) {
-        // find result_norm tensor for input
-        struct ggml_tensor * inp = nullptr;
-        for (int i = ggml_graph_n_nodes(gf) - 1; i >= 0; --i) {
-            inp = ggml_graph_node(gf, i);
-            if (strcmp(inp->name, "result_norm") == 0 || strcmp(inp->name, "result_embd") == 0) {
-                break;
-            } else {
-                inp = nullptr;
-            }
-        }
-        GGML_ASSERT(inp != nullptr && "missing result_norm/result_embd tensor");
-
-        struct ggml_tensor * cur;
-
-        switch (pooling_type) {
-            case LLAMA_POOLING_TYPE_NONE:
-                {
-                    cur = inp;
-                } break;
-            case LLAMA_POOLING_TYPE_MEAN:
-                {
-                    struct ggml_tensor * inp_mean = build_inp_mean();
-                    cur = ggml_mul_mat(ctx0, ggml_cont(ctx0, ggml_transpose(ctx0, inp)), inp_mean);
-                } break;
-            case LLAMA_POOLING_TYPE_CLS:
-            case LLAMA_POOLING_TYPE_LAST:
-                {
-                    struct ggml_tensor * inp_cls = build_inp_cls();
-                    cur = ggml_get_rows(ctx0, inp, inp_cls);
-                } break;
-            case LLAMA_POOLING_TYPE_RANK:
-                {
-                    struct ggml_tensor * inp_cls = build_inp_cls();
-                    inp = ggml_get_rows(ctx0, inp, inp_cls);
-
-                    // classification head
-                    // https://github.com/huggingface/transformers/blob/5af7d41e49bbfc8319f462eb45253dcb3863dfb7/src/transformers/models/roberta/modeling_roberta.py#L1566
-                    GGML_ASSERT(model.cls       != nullptr);
-                    GGML_ASSERT(model.cls_b     != nullptr);
-
-                    cur = ggml_add (ctx0, ggml_mul_mat(ctx0, model.cls, inp), model.cls_b);
-                    cur = ggml_tanh(ctx0, cur);
-
-                    // some models don't have `cls_out`, for example: https://huggingface.co/jinaai/jina-reranker-v1-tiny-en
-                    // https://huggingface.co/jinaai/jina-reranker-v1-tiny-en/blob/cb5347e43979c3084a890e3f99491952603ae1b7/modeling_bert.py#L884-L896
-                    if (model.cls_out) {
-                        GGML_ASSERT(model.cls_out_b != nullptr);
-
-                        cur = ggml_add (ctx0, ggml_mul_mat(ctx0, model.cls_out, cur), model.cls_out_b);
-                    }
-                } break;
-            default:
-                {
-                    GGML_ABORT("unknown pooling type");
-                }
-        }
-
-        cb(cur, "result_embd_pooled", -1);
-
-        ggml_build_forward_expand(gf, cur);
-
-        return gf;
-    }
-
-    struct ggml_tensor * llm_build_pos_bucket(bool causal) {
-        if (causal) {
-            lctx.inp_pos_bucket = ggml_new_tensor_2d(ctx0, GGML_TYPE_I32, n_kv,     n_tokens);
-        } else {
-            lctx.inp_pos_bucket = ggml_new_tensor_2d(ctx0, GGML_TYPE_I32, n_tokens, n_tokens);
-        }
-
-        ggml_set_input(lctx.inp_pos_bucket);
-        cb(lctx.inp_pos_bucket, "pos_bucket", -1);
-
-        return lctx.inp_pos_bucket;
-    }
-
-    struct ggml_tensor * llm_build_pos_bias(struct ggml_tensor * pos_bucket, struct ggml_tensor * attn_rel_b) {
-        struct ggml_tensor * pos_bucket_1d = ggml_view_1d(ctx0, pos_bucket, pos_bucket->ne[0] * pos_bucket->ne[1], 0);
-        cb(pos_bucket_1d, "pos_bucket_1d", -1);
-
-        struct ggml_tensor * pos_bias = ggml_get_rows(ctx0, attn_rel_b, pos_bucket_1d);
-        cb(pos_bias, "pos_bias", -1);
-
-        pos_bias = ggml_view_3d(ctx0, pos_bias, pos_bias->ne[0], lctx.inp_pos_bucket->ne[0], lctx.inp_pos_bucket->ne[1], ggml_element_size(pos_bias) * pos_bias->ne[0], ggml_element_size(pos_bias) * pos_bias->ne[0] * lctx.inp_pos_bucket->ne[0],  0);
-        cb(pos_bias, "pos_bias", -1);
-
-        pos_bias = ggml_permute(ctx0, pos_bias, 2, 0, 1, 3);
-        cb(pos_bias, "pos_bias", -1);
-
-        pos_bias = ggml_cont(ctx0, pos_bias);
-        cb(pos_bias, "pos_bias", -1);
-
-        return pos_bias;
-    }
-
-    struct ggml_tensor * llm_build_inp_embd_enc() {
-        const int64_t n_embd = hparams.n_embd;
-        lctx.inp_embd_enc = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, n_outputs_enc);
-        ggml_set_input(lctx.inp_embd_enc);
-        cb(lctx.inp_embd_enc, "embd_enc", -1);
-        return lctx.inp_embd_enc;
-    }
-
-    struct ggml_tensor * llm_build_inp_KQ_mask_cross() {
-        lctx.inp_KQ_mask_cross = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_outputs_enc, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD));
-        ggml_set_input(lctx.inp_KQ_mask_cross);
-        cb(lctx.inp_KQ_mask_cross, "KQ_mask_cross", -1);
-        return lctx.inp_KQ_mask_cross;
-    }
-
-    struct ggml_cgraph * build_llama() {
-        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false);
-
-        // mutable variable, needed during the last layer of the computation to skip unused tokens
-        int32_t n_tokens = this->n_tokens;
-
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-        GGML_ASSERT(n_embd_head == hparams.n_rot);
-
-        struct ggml_tensor * cur;
-        struct ggml_tensor * inpL;
-
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
-
-        // inp_pos - contains the positions
-        struct ggml_tensor * inp_pos = build_inp_pos();
-
-        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
-        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
-
-        const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
-        for (int il = 0; il < n_layer; ++il) {
-            struct ggml_tensor * inpSA = inpL;
-
-            // norm
-            cur = llm_build_norm(ctx0, inpL, hparams,
-                    model.layers[il].attn_norm, NULL,
-                    LLM_NORM_RMS, cb, il);
-            cb(cur, "attn_norm", il);
-
-            // self-attention
-            {
-                // rope freq factors for llama3; may return nullptr for llama2 and other models
-                struct ggml_tensor * rope_factors = build_rope_factors(il);
-
-                // compute Q and K and RoPE them
-                struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur);
-                cb(Qcur, "Qcur", il);
-                if (model.layers[il].bq) {
-                    Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
-                    cb(Qcur, "Qcur", il);
-                }
-
-                struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur);
-                cb(Kcur, "Kcur", il);
-                if (model.layers[il].bk) {
-                    Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
-                    cb(Kcur, "Kcur", il);
-                }
-
-                struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur);
-                cb(Vcur, "Vcur", il);
-                if (model.layers[il].bv) {
-                    Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
-                    cb(Vcur, "Vcur", il);
-                }
-
-                Qcur = ggml_rope_ext(
-                    ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, rope_factors,
-                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                    ext_factor, attn_factor, beta_fast, beta_slow
-                );
-                cb(Qcur, "Qcur", il);
-
-                Kcur = ggml_rope_ext(
-                    ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, rope_factors,
-                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                    ext_factor, attn_factor, beta_fast, beta_slow
-                );
-                cb(Kcur, "Kcur", il);
-
-                cur = llm_build_kv(ctx0, lctx, kv_self, gf,
-                        model.layers[il].wo, model.layers[il].bo,
-                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, kq_scale, cb, il);
-            }
-
-            if (il == n_layer - 1) {
-                // skip computing output for unused tokens
-                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
-                n_tokens = n_outputs;
-                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
-                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
-            }
-
-            // For Granite architecture
-            if (hparams.f_residual_scale) {
-                cur = ggml_scale(ctx0, cur, hparams.f_residual_scale);
-            }
-
-            struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
-            cb(ffn_inp, "ffn_inp", il);
-
-            // feed-forward network
-            if (model.layers[il].ffn_gate_inp == nullptr) {
-
-                cur = llm_build_norm(ctx0, ffn_inp, hparams,
-                        model.layers[il].ffn_norm, NULL,
-                        LLM_NORM_RMS, cb, il);
-                cb(cur, "ffn_norm", il);
-
-                cur = llm_build_ffn(ctx0, lctx, cur,
-                        model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
-                        model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL,
-                        model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
-                        NULL,
-                        LLM_FFN_SILU, LLM_FFN_PAR, cb, il);
-                cb(cur, "ffn_out", il);
-            } else {
-                // MoE branch
-                cur = llm_build_norm(ctx0, ffn_inp, hparams,
-                        model.layers[il].ffn_norm, NULL,
-                        LLM_NORM_RMS, cb, il);
-                cb(cur, "ffn_norm", il);
-
-                cur = llm_build_moe_ffn(ctx0, lctx, cur,
-                        model.layers[il].ffn_gate_inp,
-                        model.layers[il].ffn_up_exps,
-                        model.layers[il].ffn_gate_exps,
-                        model.layers[il].ffn_down_exps,
-                        nullptr,
-                        n_expert, n_expert_used,
-                        LLM_FFN_SILU, true,
-                        false, 0.0,
-                        LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
-                        cb, il);
-                cb(cur, "ffn_moe_out", il);
-            }
-
-            // For Granite architecture
-            if (hparams.f_residual_scale) {
-                cur = ggml_scale(ctx0, cur, hparams.f_residual_scale);
-            }
-
-            cur = ggml_add(ctx0, cur, ffn_inp);
-            cb(cur, "ffn_out", il);
-
-            cur = lctx.cvec.apply_to(ctx0, cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-
-        cur = llm_build_norm(ctx0, cur, hparams,
-                model.output_norm, NULL,
-                LLM_NORM_RMS, cb, -1);
-        cb(cur, "result_norm", -1);
-
-        // lm_head
-        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
-
-        // For Granite architecture
-        if (hparams.f_logit_scale) {
-            cur = ggml_scale(ctx0, cur, 1.0f / hparams.f_logit_scale);
-        }
-
-        cb(cur, "result_output", -1);
-
-        ggml_build_forward_expand(gf, cur);
-
-        return gf;
-    }
-
-    struct ggml_cgraph * build_deci() {
-        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false);
-
-        // mutable variable, needed during the last layer of the computation to skip unused tokens
-        int32_t n_tokens = this->n_tokens;
-
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-        GGML_ASSERT(n_embd_head == hparams.n_rot);
-
-        struct ggml_tensor * cur;
-        struct ggml_tensor * inpL;
-
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
-
-        // inp_pos - contains the positions
-        struct ggml_tensor * inp_pos = build_inp_pos();
-
-        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
-        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
-
-        const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
-        for (int il = 0; il < n_layer; ++il) {
-            struct ggml_tensor * inpSA = inpL;
-            const int64_t n_head_kv = hparams.n_head_kv(il);
-            const int64_t n_head    = hparams.n_head(il);
-
-            if (n_head == 0) {
-                // attention-free layer of Llama-3_1-Nemotron-51B
-                cur = inpL;
-            } else {
-                // norm
-                cur = llm_build_norm(ctx0, inpL, hparams,
-                        model.layers[il].attn_norm, NULL,
-                        LLM_NORM_RMS, cb, il);
-                cb(cur, "attn_norm", il);
-            }
-
-            if (n_head > 0 && n_head_kv == 0) {
-                // "linear attention" of Llama-3_1-Nemotron-51B
-                cur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wo, cur);
-                cb(cur, "wo", il);
-            } else if (n_head > 0) {
-                // self-attention
-                // rope freq factors for llama3; may return nullptr for llama2 and other models
-                struct ggml_tensor * rope_factors = build_rope_factors(il);
-
-                // compute Q and K and RoPE them
-                struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur);
-                cb(Qcur, "Qcur", il);
-                if (model.layers[il].bq) {
-                    Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
-                    cb(Qcur, "Qcur", il);
-                }
-
-                struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur);
-                cb(Kcur, "Kcur", il);
-                if (model.layers[il].bk) {
-                    Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
-                    cb(Kcur, "Kcur", il);
-                }
-
-                struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur);
-                cb(Vcur, "Vcur", il);
-                if (model.layers[il].bv) {
-                    Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
-                    cb(Vcur, "Vcur", il);
-                }
-
-                Qcur = ggml_rope_ext(
-                    ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, rope_factors,
-                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                    ext_factor, attn_factor, beta_fast, beta_slow
-                );
-                cb(Qcur, "Qcur", il);
-
-                Kcur = ggml_rope_ext(
-                    ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, rope_factors,
-                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                    ext_factor, attn_factor, beta_fast, beta_slow
-                );
-                cb(Kcur, "Kcur", il);
-
-                cur = llm_build_kv(ctx0, lctx, kv_self, gf,
-                        model.layers[il].wo, model.layers[il].bo,
-                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, kq_scale, cb, il);
-            }
-
-            if (il == n_layer - 1) {
-                // skip computing output for unused tokens
-                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
-                n_tokens = n_outputs;
-                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
-                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
-            }
-
-            // For Granite architecture
-            if (hparams.f_residual_scale) {
-                cur = ggml_scale(ctx0, cur, hparams.f_residual_scale);
-            }
-
-            // modified to support attention-free layer of Llama-3_1-Nemotron-51B
-            struct ggml_tensor * ffn_inp = cur;
-            if (n_head > 0) {
-                ffn_inp = ggml_add(ctx0, cur, inpSA);
-                cb(ffn_inp, "ffn_inp", il);
-            }
-
-            // feed-forward network
-            if (model.layers[il].ffn_gate_inp == nullptr) {
-                cur = llm_build_norm(ctx0, ffn_inp, hparams,
-                        model.layers[il].ffn_norm, NULL,
-                        LLM_NORM_RMS, cb, il);
-                cb(cur, "ffn_norm", il);
-
-                cur = llm_build_ffn(ctx0, lctx, cur,
-                        model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
-                        model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL,
-                        model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
-                        NULL,
-                        LLM_FFN_SILU, LLM_FFN_PAR, cb, il);
-                cb(cur, "ffn_out", il);
-            }
-
-            // For Granite architecture
-            if (hparams.f_residual_scale) {
-                cur = ggml_scale(ctx0, cur, hparams.f_residual_scale);
-            }
-
-            cur = ggml_add(ctx0, cur, ffn_inp);
-            cb(cur, "ffn_out", il);
-
-            cur = lctx.cvec.apply_to(ctx0, cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-
-        cur = llm_build_norm(ctx0, cur, hparams,
-                model.output_norm, NULL,
-                LLM_NORM_RMS, cb, -1);
-        cb(cur, "result_norm", -1);
-
-        // lm_head
-        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
-
-        // For Granite architecture
-        if (hparams.f_logit_scale) {
-            cur = ggml_scale(ctx0, cur, 1.0f / hparams.f_logit_scale);
-        }
-
-        cb(cur, "result_output", -1);
-
-        ggml_build_forward_expand(gf, cur);
-
-        return gf;
-    }
-
-    struct ggml_cgraph * build_baichuan() {
-        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false);
-
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-        GGML_ASSERT(n_embd_head == hparams.n_rot);
-
-        struct ggml_tensor * cur;
-        struct ggml_tensor * inpL;
-
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
-
-        // inp_pos - contains the positions
-        struct ggml_tensor * inp_pos = model.type == LLM_TYPE_7B ? build_inp_pos() : nullptr;
-
-        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
-        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
-
-        for (int il = 0; il < n_layer; ++il) {
-            struct ggml_tensor * inpSA = inpL;
-
-            cur = llm_build_norm(ctx0, inpL, hparams,
-                    model.layers[il].attn_norm, NULL,
-                    LLM_NORM_RMS, cb, il);
-            cb(cur, "attn_norm", il);
-
-            // self-attention
-            {
-                struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur);
-                cb(Qcur, "Qcur", il);
-
-                struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur);
-                cb(Kcur, "Kcur", il);
-
-                struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur);
-                cb(Vcur, "Vcur", il);
-
-                switch (model.type) {
-                    case LLM_TYPE_7B:
-                        Qcur = ggml_rope_ext(
-                            ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr,
-                            n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                            ext_factor, attn_factor, beta_fast, beta_slow
-                        );
-                        Kcur = ggml_rope_ext(
-                            ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr,
-                            n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                            ext_factor, attn_factor, beta_fast, beta_slow
-                        );
-                        break;
-                    case LLM_TYPE_13B:
-                        Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd/n_head, n_head, n_tokens);
-                        Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd/n_head, n_head, n_tokens);
-                        break;
-                    default:
-                        GGML_ABORT("fatal error");
-                }
-                cb(Qcur, "Qcur", il);
-                cb(Kcur, "Kcur", il);
-
-                cur = llm_build_kv(ctx0, lctx, kv_self, gf,
-                        model.layers[il].wo, NULL,
-                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
-            }
-
-            if (il == n_layer - 1) {
-                // skip computing output for unused tokens
-                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
-                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
-                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
-            }
-
-            struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
-            cb(ffn_inp, "ffn_inp", il);
-
-            // feed-forward network
-            {
-                cur = llm_build_norm(ctx0, ffn_inp, hparams,
-                        model.layers[il].ffn_norm, NULL,
-                        LLM_NORM_RMS, cb, il);
-                cb(cur, "ffn_norm", il);
-
-                cur = llm_build_ffn(ctx0, lctx, cur,
-                        model.layers[il].ffn_up,   NULL, NULL,
-                        model.layers[il].ffn_gate, NULL, NULL,
-                        model.layers[il].ffn_down, NULL, NULL,
-                        NULL,
-                        LLM_FFN_SILU, LLM_FFN_PAR, cb, il);
-                cb(cur, "ffn_out", il);
-            }
-
-            cur = ggml_add(ctx0, cur, ffn_inp);
-            cur = lctx.cvec.apply_to(ctx0, cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-
-        cur = llm_build_norm(ctx0, cur, hparams,
-                model.output_norm, NULL,
-                LLM_NORM_RMS, cb, -1);
-        cb(cur, "result_norm", -1);
-
-        // lm_head
-        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
-        cb(cur, "result_output", -1);
-
-        ggml_build_forward_expand(gf, cur);
-
-        return gf;
-    }
-
-    struct ggml_cgraph * build_xverse() {
-        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false);
-
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-        GGML_ASSERT(n_embd_head == hparams.n_rot);
-
-        struct ggml_tensor * cur;
-        struct ggml_tensor * inpL;
-
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
-
-        // inp_pos - contains the positions
-        struct ggml_tensor * inp_pos = build_inp_pos();
-
-        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
-        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
-
-        for (int il = 0; il < n_layer; ++il) {
-            struct ggml_tensor * inpSA = inpL;
-
-            cur = llm_build_norm(ctx0, inpL, hparams,
-                    model.layers[il].attn_norm, NULL,
-                    LLM_NORM_RMS, cb, il);
-            cb(cur, "attn_norm", il);
-
-            // self-attention
-            {
-                struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur);
-                cb(Qcur, "Qcur", il);
-
-                struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur);
-                cb(Kcur, "Kcur", il);
-
-                struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur);
-                cb(Vcur, "Vcur", il);
-
-                Qcur = ggml_rope_ext(
-                    ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr,
-                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                    ext_factor, attn_factor, beta_fast, beta_slow
-                );
-                cb(Qcur, "Qcur", il);
-
-                Kcur = ggml_rope_ext(
-                    ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr,
-                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                    ext_factor, attn_factor, beta_fast, beta_slow
-                );
-                cb(Kcur, "Kcur", il);
-                cur = llm_build_kv(ctx0, lctx, kv_self, gf,
-                        model.layers[il].wo, NULL,
-                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
-            }
-
-            if (il == n_layer - 1) {
-                // skip computing output for unused tokens
-                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
-                cur   = ggml_get_rows(ctx0,      cur, inp_out_ids);
-                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
-            }
-
-            struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
-            cb(ffn_inp, "ffn_inp", il);
-
-            // feed-forward network
-            {
-                cur = llm_build_norm(ctx0, ffn_inp, hparams,
-                        model.layers[il].ffn_norm, NULL,
-                        LLM_NORM_RMS, cb, il);
-                cb(cur, "ffn_norm", il);
-
-                cur = llm_build_ffn(ctx0, lctx, cur,
-                        model.layers[il].ffn_up,   NULL, NULL,
-                        model.layers[il].ffn_gate, NULL, NULL,
-                        model.layers[il].ffn_down, NULL, NULL,
-                        NULL,
-                        LLM_FFN_SILU, LLM_FFN_PAR, cb, il);
-                cb(cur, "ffn_out", il);
-            }
-
-            cur = ggml_add(ctx0, cur, ffn_inp);
-            cur = lctx.cvec.apply_to(ctx0, cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-
-        cur = llm_build_norm(ctx0, cur, hparams, model.output_norm, NULL, LLM_NORM_RMS, cb, -1);
-        cb(cur, "result_norm", -1);
-
-        // lm_head
-        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
-        cb(cur, "result_output", -1);
-
-        ggml_build_forward_expand(gf, cur);
-
-        return gf;
-    }
-
-    struct ggml_cgraph * build_falcon() {
-        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false);
-
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-        const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-        GGML_ASSERT(n_embd_head == hparams.n_rot);
-
-        struct ggml_tensor * cur;
-        struct ggml_tensor * inpL;
-
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
-
-        // inp_pos - contains the positions
-        struct ggml_tensor * inp_pos = build_inp_pos();
-
-        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
-        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
-
-        for (int il = 0; il < n_layer; ++il) {
-            struct ggml_tensor * attn_norm;
-
-            attn_norm = llm_build_norm(ctx0, inpL, hparams,
-                    model.layers[il].attn_norm,
-                    model.layers[il].attn_norm_b,
-                    LLM_NORM, cb, il);
-            cb(attn_norm, "attn_norm", il);
-
-            // self-attention
-            {
-                if (model.layers[il].attn_norm_2) {
-                    // Falcon-40B
-                    cur = llm_build_norm(ctx0, inpL, hparams,
-                            model.layers[il].attn_norm_2,
-                            model.layers[il].attn_norm_2_b,
-                            LLM_NORM, cb, il);
-                    cb(cur, "attn_norm_2", il);
-                } else {
-                    cur = attn_norm;
-                }
-
-                cur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wqkv, cur);
-                cb(cur, "wqkv", il);
-
-                struct ggml_tensor * Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd,     n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd)));
-                struct ggml_tensor * Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd)));
-                struct ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa)));
-
-                cb(Qcur, "Qcur", il);
-                cb(Kcur, "Kcur", il);
-                cb(Vcur, "Vcur", il);
-
-                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
-                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
-
-                // using mode = 2 for neox mode
-                Qcur = ggml_rope_ext(
-                    ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig,
-                    freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow
-                );
-                cb(Qcur, "Qcur", il);
-
-                Kcur = ggml_rope_ext(
-                    ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig,
-                    freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow
-                );
-                cb(Kcur, "Kcur", il);
-
-                cur = llm_build_kv(ctx0, lctx, kv_self, gf,
-                        model.layers[il].wo, NULL,
-                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
-            }
-
-            if (il == n_layer - 1) {
-                // skip computing output for unused tokens
-                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
-                cur       = ggml_get_rows(ctx0,       cur, inp_out_ids);
-                inpL      = ggml_get_rows(ctx0,      inpL, inp_out_ids);
-                attn_norm = ggml_get_rows(ctx0, attn_norm, inp_out_ids);
-            }
-
-            struct ggml_tensor * ffn_inp = cur;
-
-            // feed forward
-            {
-                cur = llm_build_ffn(ctx0, lctx, attn_norm, // !! use the attn norm, not the result
-                        model.layers[il].ffn_up,   NULL, NULL,
-                        NULL,                      NULL, NULL,
-                        model.layers[il].ffn_down, NULL, NULL,
-                        NULL,
-                        LLM_FFN_GELU, LLM_FFN_SEQ, cb, il);
-                cb(cur, "ffn_out", il);
-            }
-
-            cur = ggml_add(ctx0, cur, ffn_inp);
-            cur = ggml_add(ctx0, cur, inpL);
-            cur = lctx.cvec.apply_to(ctx0, cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-
-        // norm
-        cur = llm_build_norm(ctx0, cur, hparams,
-                model.output_norm,
-                model.output_norm_b,
-                LLM_NORM, cb, -1);
-        cb(cur, "result_norm", -1);
-
-        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
-        cb(cur, "result_output", -1);
-
-        ggml_build_forward_expand(gf, cur);
-
-        return gf;
-    }
-
-    struct ggml_cgraph * build_grok() {
-        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false);
-
-        // mutable variable, needed during the last layer of the computation to skip unused tokens
-        int32_t n_tokens = this->n_tokens;
-
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-        GGML_ASSERT(n_embd_head == hparams.n_rot);
-
-        struct ggml_tensor * cur;
-        struct ggml_tensor * inpL;
-
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
-
-        // multiply by embedding_multiplier_scale of 78.38367176906169
-        inpL = ggml_scale(ctx0, inpL, 78.38367176906169f);
-
-        // inp_pos - contains the positions
-        struct ggml_tensor * inp_pos = build_inp_pos();
-
-        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
-        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
-
-        for (int il = 0; il < n_layer; ++il) {
-            struct ggml_tensor * inpSA = inpL;
-
-            // norm
-            cur = llm_build_norm(ctx0, inpL, hparams,
-                    model.layers[il].attn_norm, NULL,
-                    LLM_NORM_RMS, cb, il);
-            cb(cur, "attn_norm", il);
-
-
-            // self-attention
-            {
-                // compute Q and K and RoPE them
-                struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur);
-                cb(Qcur, "Qcur", il);
-                if (model.layers[il].bq) {
-                    Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
-                    cb(Qcur, "Qcur", il);
-                }
-
-                struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur);
-                cb(Kcur, "Kcur", il);
-                if (model.layers[il].bk) {
-                    Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
-                    cb(Kcur, "Kcur", il);
-                }
-
-                struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur);
-                cb(Vcur, "Vcur", il);
-                if (model.layers[il].bv) {
-                    Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
-                    cb(Vcur, "Vcur", il);
-                }
-
-                Qcur = ggml_rope_ext(
-                    ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr,
-                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                    ext_factor, attn_factor, beta_fast, beta_slow
-                );
-                cb(Qcur, "Qcur", il);
-
-                Kcur = ggml_rope_ext(
-                    ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr,
-                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                    ext_factor, attn_factor, beta_fast, beta_slow
-                );
-                cb(Kcur, "Kcur", il);
-
-                cur = llm_build_kv(ctx0, lctx, kv_self, gf,
-                        model.layers[il].wo, model.layers[il].bo,
-                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f, cb, il);
-            }
-
-            if (il == n_layer - 1) {
-                // skip computing output for unused tokens
-                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
-                n_tokens = n_outputs;
-                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
-                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
-            }
-
-            // Grok
-            // if attn_out_norm is present then apply it before adding the input
-            if (model.layers[il].attn_out_norm) {
-                cur = llm_build_norm(ctx0, cur, hparams,
-                        model.layers[il].attn_out_norm, NULL,
-                        LLM_NORM_RMS, cb, il);
-                cb(cur, "attn_out_norm", il);
-            }
-
-            struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
-            cb(ffn_inp, "ffn_inp", il);
-
-            // feed-forward network
-            // MoE branch
-            cur = llm_build_norm(ctx0, ffn_inp, hparams,
-                    model.layers[il].ffn_norm, NULL,
-                    LLM_NORM_RMS, cb, il);
-            cb(cur, "ffn_norm", il);
-
-            cur = llm_build_moe_ffn(ctx0, lctx, cur,
-                    model.layers[il].ffn_gate_inp,
-                    model.layers[il].ffn_up_exps,
-                    model.layers[il].ffn_gate_exps,
-                    model.layers[il].ffn_down_exps,
-                    nullptr,
-                    n_expert, n_expert_used,
-                    LLM_FFN_GELU, true,
-                    false, 0.0,
-                    LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
-                    cb, il);
-            cb(cur, "ffn_moe_out", il);
-
-            // Grok
-            // if layer_out_norm is present then apply it before adding the input
-            // Idea: maybe ffn_out_norm is a better name
-            if (model.layers[il].layer_out_norm) {
-                cur = llm_build_norm(ctx0, cur, hparams,
-                        model.layers[il].layer_out_norm, NULL,
-                        LLM_NORM_RMS, cb, il);
-                cb(cur, "layer_out_norm", il);
-            }
-
-            cur = ggml_add(ctx0, cur, ffn_inp);
-            cb(cur, "ffn_out", il);
-
-            cur = lctx.cvec.apply_to(ctx0, cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-
-        cur = llm_build_norm(ctx0, cur, hparams,
-                model.output_norm, NULL,
-                LLM_NORM_RMS, cb, -1);
-        cb(cur, "result_norm", -1);
-
-        // lm_head
-        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
-
-        // Grok
-        // multiply logits by output_multiplier_scale of 0.5773502691896257
-
-        cur = ggml_scale(ctx0, cur, 0.5773502691896257f);
-
-        cb(cur, "result_output", -1);
-
-        ggml_build_forward_expand(gf, cur);
-
-        return gf;
-    }
-
-    struct ggml_cgraph * build_dbrx() {
-        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false);
-
-        // mutable variable, needed during the last layer of the computation to skip unused tokens
-        int32_t n_tokens = this->n_tokens;
-
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-        const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-        GGML_ASSERT(n_embd_head == hparams.n_rot);
-
-        struct ggml_tensor * cur;
-        struct ggml_tensor * inpL;
-
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
-
-        // inp_pos - contains the positions
-        struct ggml_tensor * inp_pos = build_inp_pos();
-
-        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
-        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
-
-        for (int il = 0; il < n_layer; ++il) {
-            struct ggml_tensor * inpSA = inpL;
-
-            // norm
-            cur = llm_build_norm(ctx0, inpL, hparams,
-                                 model.layers[il].attn_norm, NULL,
-                                 LLM_NORM, cb, il);
-            cb(cur, "attn_norm", il);
-
-            // self-attention
-            {
-                struct ggml_tensor * Qcur = nullptr;
-                struct ggml_tensor * Kcur = nullptr;
-                struct ggml_tensor * Vcur = nullptr;
-
-                cur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wqkv, cur);
-                cb(cur, "wqkv", il);
-
-                cur = ggml_clamp(ctx0, cur, -hparams.f_clamp_kqv, hparams.f_clamp_kqv);
-                cb(cur, "wqkv_clamped", il);
-
-                Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd,     n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd)));
-                Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd)));
-                Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa)));
-
-                cb(Qcur, "Qcur", il);
-                cb(Kcur, "Kcur", il);
-                cb(Vcur, "Vcur", il);
-
-                Qcur = ggml_rope_ext(
-                    ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr,
-                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                    ext_factor, attn_factor, beta_fast, beta_slow
-                );
-                cb(Qcur, "Qcur", il);
-
-                Kcur = ggml_rope_ext(
-                    ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr,
-                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                    ext_factor, attn_factor, beta_fast, beta_slow
-                );
-                cb(Kcur, "Kcur", il);
-
-                cur = llm_build_kv(ctx0, lctx, kv_self, gf,
-                        model.layers[il].wo, NULL,
-                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
-            }
-
-            if (il == n_layer - 1) {
-                // skip computing output for unused tokens
-                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
-                n_tokens = n_outputs;
-                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
-                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
-            }
-
-            struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
-            cb(ffn_inp, "ffn_inp", il);
-
-            // feed-forward network
-            // MoE branch
-            cur = llm_build_norm(ctx0, ffn_inp, hparams,
-                                 model.layers[il].attn_out_norm, NULL,
-                                 LLM_NORM, cb, il);
-            cb(cur, "attn_out_norm", il);
-
-            cur = llm_build_moe_ffn(ctx0, lctx, cur,
-                    model.layers[il].ffn_gate_inp,
-                    model.layers[il].ffn_up_exps,
-                    model.layers[il].ffn_gate_exps,
-                    model.layers[il].ffn_down_exps,
-                    nullptr,
-                    n_expert, n_expert_used,
-                    LLM_FFN_SILU, true,
-                    false, 0.0,
-                    LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
-                    cb, il);
-            cb(cur, "ffn_moe_out", il);
-
-            cur = ggml_add(ctx0, cur, ffn_inp);
-            cb(cur, "ffn_out", il);
-
-            cur = lctx.cvec.apply_to(ctx0, cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-
-        cur = llm_build_norm(ctx0, cur, hparams,
-                             model.output_norm, NULL,
-                             LLM_NORM, cb, -1);
-        cb(cur, "result_norm", -1);
-
-        // lm_head
-        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
-
-        cb(cur, "result_output", -1);
-
-        ggml_build_forward_expand(gf, cur);
-
-        return gf;
-    }
-
-    struct ggml_cgraph * build_starcoder() {
-        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false);
-
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-        const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-
-        struct ggml_tensor * cur;
-        struct ggml_tensor * inpL;
-
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
-
-        // inp_pos - contains the positions
-        struct ggml_tensor * inp_pos = build_inp_pos();
-
-        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
-        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
-
-        struct ggml_tensor * pos = ggml_get_rows(ctx0, model.pos_embd, inp_pos);
-        cb(pos, "pos_embd", -1);
-
-        inpL = ggml_add(ctx0, inpL, pos);
-        cb(inpL, "inpL", -1);
-
-        for (int il = 0; il < n_layer; ++il) {
-            cur = llm_build_norm(ctx0, inpL, hparams,
-                    model.layers[il].attn_norm,
-                    model.layers[il].attn_norm_b,
-                    LLM_NORM, cb, il);
-            cb(cur, "attn_norm", il);
-
-            // self-attention
-            {
-                cur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wqkv, cur);
-                cb(cur, "wqkv", il);
-
-                cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
-                cb(cur, "bqkv", il);
-
-                struct ggml_tensor * Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd,     n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd)));
-                struct ggml_tensor * Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd)));
-                struct ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa)));
-
-                cb(Qcur, "Qcur", il);
-                cb(Kcur, "Kcur", il);
-                cb(Vcur, "Vcur", il);
-
-                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
-
-                cur = llm_build_kv(ctx0, lctx, kv_self, gf,
-                        model.layers[il].wo, model.layers[il].bo,
-                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
-            }
-
-            if (il == n_layer - 1) {
-                // skip computing output for unused tokens
-                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
-                cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
-                inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
-            }
-
-            // add the input
-            struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL);
-            cb(ffn_inp, "ffn_inp", il);
-
-            // FF
-            {
-                cur = llm_build_norm(ctx0, ffn_inp, hparams,
-                        model.layers[il].ffn_norm,
-                        model.layers[il].ffn_norm_b,
-                        LLM_NORM, cb, il);
-                cb(cur, "ffn_norm", il);
-
-                cur = llm_build_ffn(ctx0, lctx, cur,
-                        model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
-                        NULL,                      NULL,                        NULL,
-                        model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
-                        NULL,
-                        LLM_FFN_GELU, LLM_FFN_SEQ, cb, il);
-                cb(cur, "ffn_out", il);
-            }
-
-            cur = ggml_add(ctx0, cur, ffn_inp);
-            cur = lctx.cvec.apply_to(ctx0, cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = llm_build_norm(ctx0, inpL, hparams,
-                model.output_norm,
-                model.output_norm_b,
-                LLM_NORM, cb, -1);
-        cb(cur, "result_norm", -1);
-
-        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
-        cb(cur, "result_output", -1);
-
-        ggml_build_forward_expand(gf, cur);
-
-        return gf;
-    }
-
-    struct ggml_cgraph * build_refact() {
-        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false);
-
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-
-        struct ggml_tensor * cur;
-        struct ggml_tensor * inpL;
-
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
-
-        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
-        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
-
-        for (int il = 0; il < n_layer; ++il) {
-            struct ggml_tensor * inpSA = inpL;
-
-            cur = llm_build_norm(ctx0, inpL, hparams,
-                    model.layers[il].attn_norm, NULL,
-                    LLM_NORM_RMS, cb, il);
-            cb(cur, "attn_norm", il);
-
-            // self-attention
-            {
-                struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur);
-                cb(Qcur, "Qcur", il);
-
-                struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur);
-                cb(Kcur, "Kcur", il);
-
-                struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur);
-                cb(Vcur, "Vcur", il);
-
-                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
-                cb(Kcur, "Kcur", il);
-
-                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
-                cb(Qcur, "Qcur", il);
-
-                cur = llm_build_kv(ctx0, lctx, kv_self, gf,
-                        model.layers[il].wo, NULL,
-                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
-            }
-
-            if (il == n_layer - 1) {
-                // skip computing output for unused tokens
-                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
-                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
-                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
-            }
-
-            struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
-            cb(ffn_inp, "ffn_inp", il);
-
-            // feed-forward network
-            {
-                cur = llm_build_norm(ctx0, ffn_inp, hparams,
-                        model.layers[il].ffn_norm, NULL,
-                        LLM_NORM_RMS, cb, il);
-                cb(cur, "ffn_norm", il);
-
-                cur = llm_build_ffn(ctx0, lctx, cur,
-                        model.layers[il].ffn_up,   NULL, NULL,
-                        model.layers[il].ffn_gate, NULL, NULL,
-                        model.layers[il].ffn_down, NULL, NULL,
-                        NULL,
-                        LLM_FFN_SILU, LLM_FFN_PAR, cb, il);
-                cb(cur, "ffn_out", il);
-            }
-
-            cur = ggml_add(ctx0, cur, ffn_inp);
-            cur = lctx.cvec.apply_to(ctx0, cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-
-        cur = llm_build_norm(ctx0, cur, hparams,
-                model.output_norm, NULL,
-                LLM_NORM_RMS, cb, -1);
-        cb(cur, "result_norm", -1);
-
-        // lm_head
-        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
-        cb(cur, "result_output", -1);
-
-        ggml_build_forward_expand(gf, cur);
-
-        return gf;
-    }
-
-    struct ggml_cgraph * build_bert() {
-        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false);
-
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-        const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
-
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-
-        struct ggml_tensor * cur;
-        struct ggml_tensor * inpL;
-        struct ggml_tensor * inp_pos = nullptr;
-
-        if (model.arch != LLM_ARCH_JINA_BERT_V2) {
-            inp_pos = build_inp_pos();
-        }
-
-        // construct input embeddings (token, type, position)
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
-
-        // token types are hardcoded to zero ("Sentence A")
-        struct ggml_tensor * type_row0 = ggml_view_1d(ctx0, model.type_embd, n_embd, 0);
-        inpL = ggml_add(ctx0, inpL, type_row0);
-        if (model.arch == LLM_ARCH_BERT) {
-            inpL = ggml_add(ctx0, ggml_get_rows(ctx0, model.pos_embd, inp_pos), inpL);
-        }
-        cb(inpL, "inp_embd", -1);
-
-        // embed layer norm
-        inpL = llm_build_norm(ctx0, inpL, hparams, model.tok_norm, model.tok_norm_b, LLM_NORM, cb, -1);
-        cb(inpL, "inp_norm", -1);
-
-        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
-        struct ggml_tensor * KQ_mask = build_inp_KQ_mask(false);
-
-        // iterate layers
-        for (int il = 0; il < n_layer; ++il) {
-            struct ggml_tensor * cur = inpL;
-
-            struct ggml_tensor * Qcur;
-            struct ggml_tensor * Kcur;
-            struct ggml_tensor * Vcur;
-
-            // self-attention
-            if (model.arch == LLM_ARCH_BERT || model.arch == LLM_ARCH_JINA_BERT_V2) {
-                Qcur = ggml_add(ctx0, llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur), model.layers[il].bq);
-                cb(Qcur, "Qcur", il);
-
-                if (model.layers[il].attn_q_norm) {
-                    Qcur = llm_build_norm(ctx0, Qcur, hparams,
-                            model.layers[il].attn_q_norm,
-                            model.layers[il].attn_q_norm_b,
-                            LLM_NORM, cb, il);
-                }
-
-                Kcur = ggml_add(ctx0, llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur), model.layers[il].bk);
-                cb(Kcur, "Kcur", il);
-
-                if (model.layers[il].attn_k_norm) {
-                    Kcur = llm_build_norm(ctx0, Kcur, hparams,
-                            model.layers[il].attn_k_norm,
-                            model.layers[il].attn_k_norm_b,
-                            LLM_NORM, cb, il);
-                }
-                Vcur = ggml_add(ctx0, llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur), model.layers[il].bv);
-                cb(Vcur, "Vcur", il);
-
-                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
-                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
-            } else {
-                // compute Q and K and RoPE them
-                cur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wqkv, cur);
-                cb(cur, "wqkv", il);
-
-                Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd,     n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd)));
-                Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd)));
-                Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa)));
-
-                cb(Qcur, "Qcur", il);
-                cb(Kcur, "Kcur", il);
-                cb(Vcur, "Vcur", il);
-
-                Qcur = ggml_rope_ext(
-                    ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens), inp_pos, nullptr,
-                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                    ext_factor, attn_factor, beta_fast, beta_slow
-                );
-                cb(Qcur, "Qcur", il);
-
-                Kcur = ggml_rope_ext(
-                    ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr,
-                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                    ext_factor, attn_factor, beta_fast, beta_slow
-                );
-                cb(Kcur, "Kcur", il);
-            }
-
-            struct ggml_tensor * q =                 ggml_permute(ctx0, Qcur, 0, 2, 1, 3);
-            struct ggml_tensor * k = ggml_cont(ctx0, ggml_permute(ctx0, Kcur, 0, 2, 1, 3));
-
-            struct ggml_tensor * kq = ggml_mul_mat(ctx0, k, q);
-            cb(kq, "kq", il);
-
-            kq = ggml_soft_max_ext(ctx0, kq, KQ_mask, 1.0f/sqrtf(float(n_embd_head)), hparams.f_max_alibi_bias);
-            cb(kq, "kq_soft_max_ext", il);
-
-            struct ggml_tensor * v = ggml_cont(ctx0, ggml_transpose(ctx0, ggml_reshape_2d(ctx0, Vcur, n_embd_gqa, n_tokens)));
-            cb(v, "v", il);
-
-            struct ggml_tensor * kqv = ggml_mul_mat(ctx0, ggml_reshape_3d(ctx0, v, n_tokens, n_embd_head, n_head_kv), kq);
-            cb(kqv, "kqv", il);
-
-            struct ggml_tensor * kqv_merged = ggml_permute(ctx0, kqv, 0, 2, 1, 3);
-            cb(kqv_merged, "kqv_merged", il);
-
-            cur = ggml_cont_2d(ctx0, kqv_merged, n_embd_gqa, n_tokens);
-            cb(cur, "kqv_merged_cont", il);
-
-            ggml_build_forward_expand(gf, cur);
-
-            cur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wo, cur);
-            if (model.layers[il].bo) {
-                cb(cur, "kqv_wo", il);
-            }
-
-            if (model.layers[il].bo) {
-                cur = ggml_add(ctx0, cur, model.layers[il].bo);
-            }
-            cb(cur, "kqv_out", il);
-
-            if (il == n_layer - 1 && pooling_type == LLAMA_POOLING_TYPE_NONE) {
-                // skip computing output for unused tokens
-                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
-                cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
-                inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
-            }
-
-            // re-add the layer input
-            cur = ggml_add(ctx0, cur, inpL);
-
-            // attention layer norm
-            cur = llm_build_norm(ctx0, cur, hparams, model.layers[il].attn_out_norm, model.layers[il].attn_out_norm_b, LLM_NORM, cb, il);
-
-            if (model.layers[il].attn_norm_2 != nullptr) {
-                cur = ggml_add(ctx0, cur, inpL); // re-add the layer input
-                cur = llm_build_norm(ctx0, cur, hparams, model.layers[il].attn_norm_2, model.layers[il].attn_norm_2_b, LLM_NORM, cb, il);
-            }
-
-            struct ggml_tensor * ffn_inp = cur;
-            cb(ffn_inp, "ffn_inp", il);
-
-            // feed-forward network
-            if (model.arch == LLM_ARCH_BERT) {
-                cur = llm_build_ffn(ctx0, lctx, cur,
-                        model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
-                        NULL,                      NULL,                        NULL,
-                        model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
-                        NULL,
-                        LLM_FFN_GELU, LLM_FFN_SEQ, cb, il);
-            } else if (model.arch == LLM_ARCH_JINA_BERT_V2) {
-                cur = llm_build_ffn(ctx0, lctx, cur,
-                        model.layers[il].ffn_up,   NULL,                        NULL,
-                        model.layers[il].ffn_gate, NULL,                        NULL,
-                        model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
-                        NULL,
-                        LLM_FFN_GELU, LLM_FFN_PAR, cb, il);
-            } else {
-                cur = llm_build_ffn(ctx0, lctx, cur,
-                        model.layers[il].ffn_up,   NULL, NULL,
-                        model.layers[il].ffn_gate, NULL, NULL,
-                        model.layers[il].ffn_down, NULL, NULL,
-                        NULL,
-                        LLM_FFN_SILU, LLM_FFN_PAR, cb, il);
-            }
-            cb(cur, "ffn_out", il);
-
-            // attentions bypass the intermediate layer
-            cur = ggml_add(ctx0, cur, ffn_inp);
-
-            // output layer norm
-            cur = llm_build_norm(ctx0, cur, hparams, model.layers[il].layer_out_norm, model.layers[il].layer_out_norm_b, LLM_NORM, cb, il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-
-        cb(cur, "result_embd", -1);
-
-        ggml_build_forward_expand(gf, cur);
-
-        return gf;
-    }
-
-    struct ggml_cgraph * build_bloom() {
-        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false);
-
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-        const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-
-        struct ggml_tensor * cur;
-        struct ggml_tensor * inpL;
-
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
-
-        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
-        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
-
-        inpL = llm_build_norm(ctx0, inpL, hparams,
-                model.tok_norm,
-                model.tok_norm_b,
-                LLM_NORM, cb, -1);
-        cb(inpL, "inp_norm", -1);
-
-        for (int il = 0; il < n_layer; ++il) {
-            cur = llm_build_norm(ctx0, inpL, hparams,
-                    model.layers[il].attn_norm,
-                    model.layers[il].attn_norm_b,
-                    LLM_NORM, cb, il);
-            cb(cur, "attn_norm", il);
-
-            // self-attention
-            {
-                cur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wqkv, cur);
-                cb(cur, "wqkv", il);
-
-                cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
-                cb(cur, "bqkv", il);
-
-                struct ggml_tensor * Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd,     n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd)));
-                struct ggml_tensor * Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd)));
-                struct ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa)));
-
-                cb(Qcur, "Qcur", il);
-                cb(Kcur, "Kcur", il);
-                cb(Vcur, "Vcur", il);
-
-                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
-
-                cur = llm_build_kv(ctx0, lctx, kv_self, gf,
-                        model.layers[il].wo, model.layers[il].bo,
-                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
-            }
-
-            if (il == n_layer - 1) {
-                // skip computing output for unused tokens
-                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
-                cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
-                inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
-            }
-
-            // Add the input
-            struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL);
-            cb(ffn_inp, "ffn_inp", il);
-
-            // FF
-            {
-                cur = llm_build_norm(ctx0, ffn_inp, hparams,
-                        model.layers[il].ffn_norm,
-                        model.layers[il].ffn_norm_b,
-                        LLM_NORM, cb, il);
-                cb(cur, "ffn_norm", il);
-
-                cur = llm_build_ffn(ctx0, lctx, cur,
-                        model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
-                        NULL,                      NULL,                        NULL,
-                        model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
-                        NULL,
-                        LLM_FFN_GELU, LLM_FFN_SEQ, cb, il);
-                cb(cur, "ffn_out", il);
-            }
-
-            cur = ggml_add(ctx0, cur, ffn_inp);
-            cur = lctx.cvec.apply_to(ctx0, cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = llm_build_norm(ctx0, inpL, hparams,
-                model.output_norm,
-                model.output_norm_b,
-                LLM_NORM, cb, -1);
-        cb(cur, "result_norm", -1);
-
-        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
-        cb(cur, "result_output", -1);
-
-        ggml_build_forward_expand(gf, cur);
-
-        return gf;
-    }
-
-    struct ggml_cgraph * build_mpt() {
-        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false);
-
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-        const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-
-        struct ggml_tensor * cur;
-        struct ggml_tensor * pos;
-        struct ggml_tensor * inpL;
-
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
-
-        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
-        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
-
-        if (model.pos_embd) {
-            // inp_pos - contains the positions
-            struct ggml_tensor * inp_pos = build_inp_pos();
-            pos = ggml_get_rows(ctx0, model.pos_embd, inp_pos);
-            cb(pos, "pos_embd", -1);
-
-            inpL = ggml_add(ctx0, inpL, pos);
-            cb(inpL, "inpL", -1);
-        }
-
-        for (int il = 0; il < n_layer; ++il) {
-            struct ggml_tensor * attn_norm;
-
-            attn_norm = llm_build_norm(ctx0, inpL, hparams,
-                    model.layers[il].attn_norm,
-                    model.layers[il].attn_norm_b,
-                    LLM_NORM, cb, il);
-            cb(attn_norm, "attn_norm", il);
-
-            // self-attention
-            {
-                cur = attn_norm;
-
-                cur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wqkv, cur);
-                cb(cur, "wqkv", il);
-
-                if (model.layers[il].bqkv){
-                    cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
-                    cb(cur, "bqkv", il);
-                }
-
-                if (hparams.f_clamp_kqv > 0.0f) {
-                    cur = ggml_clamp(ctx0, cur, -hparams.f_clamp_kqv, hparams.f_clamp_kqv);
-                    cb(cur, "wqkv_clamped", il);
-                }
-
-                struct ggml_tensor * Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd,     n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd)));
-                struct ggml_tensor * Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd)));
-                struct ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa)));
-
-                cb(Qcur, "Qcur", il);
-                cb(Kcur, "Kcur", il);
-                cb(Vcur, "Vcur", il);
-
-                // Q/K Layernorm
-                if (model.layers[il].attn_q_norm) {
-                    Qcur = llm_build_norm(ctx0, Qcur, hparams,
-                            model.layers[il].attn_q_norm,
-                            model.layers[il].attn_q_norm_b,
-                            LLM_NORM, cb, il);
-                    cb(Qcur, "Qcur", il);
-
-                    Kcur = llm_build_norm(ctx0, Kcur, hparams,
-                            model.layers[il].attn_k_norm,
-                            model.layers[il].attn_k_norm_b,
-                            LLM_NORM, cb, il);
-                    cb(Kcur, "Kcur", il);
-
-                    Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
-                    Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
-
-                    cur = llm_build_kv(ctx0, lctx, kv_self, gf,
-                            model.layers[il].wo, model.layers[il].bo,
-                            Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
-                } else {
-                    Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
-
-                    cur = llm_build_kv(ctx0, lctx, kv_self, gf,
-                            model.layers[il].wo, model.layers[il].bo,
-                            Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
-                }
-            }
-
-            if (il == n_layer - 1) {
-                // skip computing output for unused tokens
-                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
-                cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
-                inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
-            }
-
-            // Add the input
-            struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL);
-            cb(ffn_inp, "ffn_inp", il);
-
-            // feed forward
-            {
-                cur = llm_build_norm(ctx0, ffn_inp, hparams,
-                        model.layers[il].ffn_norm,
-                        model.layers[il].ffn_norm_b,
-                        LLM_NORM, cb, il);
-                cb(cur, "ffn_norm", il);
-                cur = llm_build_ffn(ctx0, lctx, cur,
-                        model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
-                        NULL,                      NULL,                        NULL,
-                        model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
-                        model.layers[il].ffn_act,
-                        LLM_FFN_GELU, LLM_FFN_SEQ, cb, il);
-                cb(cur, "ffn_out", il);
-            }
-
-            cur = ggml_add(ctx0, cur, ffn_inp);
-            cur = lctx.cvec.apply_to(ctx0, cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-
-        cur = llm_build_norm(ctx0, cur, hparams,
-                model.output_norm,
-                model.output_norm_b,
-                LLM_NORM, cb, -1);
-        cb(cur, "result_norm", -1);
-
-        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
-        cb(cur, "result_output", -1);
-
-        ggml_build_forward_expand(gf, cur);
-
-        return gf;
-    }
-
-    struct ggml_cgraph * build_stablelm() {
-        struct ggml_cgraph * gf = ggml_new_graph(ctx0);
-
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-
-        struct ggml_tensor * cur;
-        struct ggml_tensor * inpL;
-
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
-
-        // inp_pos - contains the positions
-        struct ggml_tensor * inp_pos = build_inp_pos();
-
-        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
-        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
-
-        for (int il = 0; il < n_layer; ++il) {
-
-
-            // norm
-            cur = llm_build_norm(ctx0, inpL, hparams,
-                    model.layers[il].attn_norm,
-                    model.layers[il].attn_norm_b,
-                    LLM_NORM, cb, il);
-            cb(cur, "attn_norm", il);
-
-            struct ggml_tensor * inpSA = cur;
-
-            // self-attention
-            {
-                // compute Q and K and RoPE them
-                struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur);
-                cb(Qcur, "Qcur", il);
-                if (model.layers[il].bq) {
-                    Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
-                    cb(Qcur, "Qcur", il);
-                }
-
-                struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur);
-                cb(Kcur, "Kcur", il);
-                if (model.layers[il].bk) {
-                    Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
-                    cb(Kcur, "Kcur", il);
-                }
-
-                struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur);
-                cb(Vcur, "Vcur", il);
-                if (model.layers[il].bv) {
-                    Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
-                    cb(Vcur, "Vcur", il);
-                }
-
-                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
-                cb(Qcur, "Qcur", il);
-                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
-                cb(Kcur, "Kcur", il);
-
-                if (model.layers[il].attn_q_norm) {
-                    Qcur = llm_build_norm(ctx0, Qcur, hparams,
-                            model.layers[il].attn_q_norm,
-                            NULL,
-                            LLM_NORM, cb, il);
-                    cb(Qcur, "Qcur", il);
-                }
-                if (model.layers[il].attn_k_norm) {
-                    Kcur = llm_build_norm(ctx0, Kcur, hparams,
-                            model.layers[il].attn_k_norm,
-                            NULL,
-                            LLM_NORM, cb, il);
-                    cb(Kcur, "Kcur", il);
-                }
-
-
-                Qcur = ggml_rope_ext(
-                    ctx0, Qcur, inp_pos, nullptr,
-                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                    ext_factor, attn_factor, beta_fast, beta_slow
-                );
-                cb(Qcur, "Qcur", il);
-
-                Kcur = ggml_rope_ext(
-                    ctx0, Kcur, inp_pos, nullptr,
-                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                    ext_factor, attn_factor, beta_fast, beta_slow
-                );
-                cb(Kcur, "Kcur", il);
-
-                cur = llm_build_kv(ctx0, lctx, kv_self, gf,
-                        model.layers[il].wo, NULL,
-                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
-            }
-
-            if (il == n_layer - 1) {
-                // skip computing output for unused tokens
-                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
-                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
-                inpL  = ggml_get_rows(ctx0,  inpL, inp_out_ids);
-                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
-            }
-
-            struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL);
-            cb(ffn_inp, "ffn_inp", il);
-
-            // feed-forward network
-            {
-                if (model.layers[il].ffn_norm) {
-                    cur = llm_build_norm(ctx0, ffn_inp, hparams,
-                            model.layers[il].ffn_norm,
-                            model.layers[il].ffn_norm_b,
-                            LLM_NORM, cb, il);
-                    cb(cur, "ffn_norm", il);
-                } else {
-                    // parallel residual
-                    cur = inpSA;
-                }
-                cur = llm_build_ffn(ctx0, lctx, cur,
-                        model.layers[il].ffn_up,   NULL, NULL,
-                        model.layers[il].ffn_gate, NULL, NULL,
-                        model.layers[il].ffn_down, NULL, NULL,
-                        NULL,
-                        LLM_FFN_SILU, LLM_FFN_PAR, cb, il);
-                cb(cur, "ffn_out", il);
-            }
-
-            cur = ggml_add(ctx0, cur, ffn_inp);
-            cur = lctx.cvec.apply_to(ctx0, cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-
-        cur = llm_build_norm(ctx0, cur, hparams,
-                model.output_norm,
-                model.output_norm_b,
-                LLM_NORM, cb, -1);
-        cb(cur, "result_norm", -1);
-
-        // lm_head
-        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
-        cb(cur, "result_output", -1);
-
-        ggml_build_forward_expand(gf, cur);
-
-        return gf;
-    }
-
-    struct ggml_cgraph * build_qwen() {
-        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false);
-
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-
-        struct ggml_tensor * cur;
-        struct ggml_tensor * inpL;
-
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
-
-        // inp_pos - contains the positions
-        struct ggml_tensor * inp_pos = build_inp_pos();
-
-        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
-        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
-
-        for (int il = 0; il < n_layer; ++il) {
-            struct ggml_tensor * inpSA = inpL;
-
-            cur = llm_build_norm(ctx0, inpL, hparams,
-                    model.layers[il].attn_norm, NULL,
-                    LLM_NORM_RMS, cb, il);
-            cb(cur, "attn_norm", il);
-
-            // self-attention
-            {
-                cur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wqkv, cur);
-                cb(cur, "wqkv", il);
-
-                cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
-                cb(cur, "bqkv", il);
-
-                struct ggml_tensor * Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd)));
-                struct ggml_tensor * Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd)));
-                struct ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 2*sizeof(float)*(n_embd)));
-
-                cb(Qcur, "Qcur", il);
-                cb(Kcur, "Kcur", il);
-                cb(Vcur, "Vcur", il);
-
-                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
-                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
-
-                // using mode = 2 for neox mode
-                Qcur = ggml_rope_ext(
-                    ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig,
-                    freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow
-                );
-                cb(Qcur, "Qcur", il);
-
-                Kcur = ggml_rope_ext(
-                    ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig,
-                    freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow
-                );
-                cb(Kcur, "Kcur", il);
-
-                cur = llm_build_kv(ctx0, lctx, kv_self, gf,
-                        model.layers[il].wo, NULL,
-                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
-            }
-
-            if (il == n_layer - 1) {
-                // skip computing output for unused tokens
-                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
-                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
-                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
-            }
-
-            struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
-            cb(ffn_inp, "ffn_inp", il);
-
-            // feed-forward forward
-            {
-                cur = llm_build_norm(ctx0, ffn_inp, hparams,
-                        model.layers[il].ffn_norm, NULL,
-                        LLM_NORM_RMS, cb, il);
-                cb(cur, "ffn_norm", il);
-
-                cur = llm_build_ffn(ctx0, lctx, cur,
-                        model.layers[il].ffn_up,   NULL, NULL,
-                        model.layers[il].ffn_gate, NULL, NULL,
-                        model.layers[il].ffn_down, NULL, NULL,
-                        NULL,
-                        LLM_FFN_SILU, LLM_FFN_PAR, cb, il);
-                cb(cur, "ffn_out", il);
-            }
-
-            cur = ggml_add(ctx0, cur, ffn_inp);
-            cur = lctx.cvec.apply_to(ctx0, cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-
-        cur = llm_build_norm(ctx0, cur, hparams,
-                model.output_norm, NULL,
-                LLM_NORM_RMS, cb, -1);
-        cb(cur, "result_norm", -1);
-
-        // lm_head
-        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
-        cb(cur, "result_output", -1);
-
-        ggml_build_forward_expand(gf, cur);
-
-        return gf;
-    }
-
-    struct ggml_cgraph * build_qwen2() {
-        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false);
-
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-        GGML_ASSERT(n_embd_head == hparams.n_rot);
-
-        struct ggml_tensor * cur;
-        struct ggml_tensor * inpL;
-
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
-
-        // inp_pos - contains the positions
-        struct ggml_tensor * inp_pos = build_inp_pos();
-
-        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
-        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
-
-        for (int il = 0; il < n_layer; ++il) {
-            struct ggml_tensor * inpSA = inpL;
-
-            // norm
-            cur = llm_build_norm(ctx0, inpL, hparams,
-                    model.layers[il].attn_norm, NULL,
-                    LLM_NORM_RMS, cb, il);
-            cb(cur, "attn_norm", il);
-
-            // self-attention
-            {
-                // compute Q and K and RoPE them
-                struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur);
-                cb(Qcur, "Qcur", il);
-                Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
-                cb(Qcur, "Qcur", il);
-
-                struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur);
-                cb(Kcur, "Kcur", il);
-                Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
-                cb(Kcur, "Kcur", il);
-
-                struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur);
-                cb(Vcur, "Vcur", il);
-                Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
-                cb(Vcur, "Vcur", il);
-
-                Qcur = ggml_rope_ext(
-                    ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens), inp_pos, nullptr,
-                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                    ext_factor, attn_factor, beta_fast, beta_slow
-                );
-                cb(Qcur, "Qcur", il);
-
-                Kcur = ggml_rope_ext(
-                    ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr,
-                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                    ext_factor, attn_factor, beta_fast, beta_slow
-                );
-                cb(Kcur, "Kcur", il);
-
-                cur = llm_build_kv(ctx0, lctx, kv_self, gf,
-                        model.layers[il].wo, model.layers[il].bo,
-                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
-            }
-
-            if (il == n_layer - 1) {
-                // skip computing output for unused tokens
-                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
-                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
-                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
-            }
-
-            struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
-            cb(ffn_inp, "ffn_inp", il);
-
-            // feed-forward network
-            cur = llm_build_norm(ctx0, ffn_inp, hparams,
-                    model.layers[il].ffn_norm, NULL,
-                    LLM_NORM_RMS, cb, il);
-            cb(cur, "ffn_norm", il);
-
-            cur = llm_build_ffn(ctx0, lctx, cur,
-                    model.layers[il].ffn_up,   NULL, NULL,
-                    model.layers[il].ffn_gate, NULL, NULL,
-                    model.layers[il].ffn_down, NULL, NULL,
-                    NULL,
-                    LLM_FFN_SILU, LLM_FFN_PAR, cb, il);
-            cb(cur, "ffn_out", il);
-
-            cur = ggml_add(ctx0, cur, ffn_inp);
-            cur = lctx.cvec.apply_to(ctx0, cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-
-        cur = llm_build_norm(ctx0, cur, hparams,
-                model.output_norm, NULL,
-                LLM_NORM_RMS, cb, -1);
-        cb(cur, "result_norm", -1);
-
-        // lm_head
-        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
-        cb(cur, "result_output", -1);
-
-        ggml_build_forward_expand(gf, cur);
-
-        return gf;
-    }
-
-    struct ggml_cgraph * build_qwen2vl() {
-        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false);
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-        GGML_ASSERT(n_embd_head == hparams.n_rot);
-
-        struct ggml_tensor * cur;
-        struct ggml_tensor * inpL;
-
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
-
-        // inp_pos - contains the positions
-        lctx.inp_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens * 4);
-        cb(lctx.inp_pos, "inp_pos", -1);
-        ggml_set_input(lctx.inp_pos);
-        struct ggml_tensor * inp_pos = lctx.inp_pos;
-
-        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
-        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
-        int sections[4];
-        std::copy(std::begin(hparams.rope_sections), std::begin(hparams.rope_sections) + 4, sections);
-
-        for (int il = 0; il < n_layer; ++il) {
-            struct ggml_tensor * inpSA = inpL;
-
-            // norm
-            cur = llm_build_norm(ctx0, inpL, hparams,
-                    model.layers[il].attn_norm, NULL,
-                    LLM_NORM_RMS, cb, il);
-            cb(cur, "attn_norm", il);
-
-            // self-attention
-            {
-                // compute Q and K and RoPE them
-                struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur);
-                cb(Qcur, "Qcur", il);
-                Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
-                cb(Qcur, "Qcur", il);
-
-                struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur);
-                cb(Kcur, "Kcur", il);
-                Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
-                cb(Kcur, "Kcur", il);
-
-                struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur);
-                cb(Vcur, "Vcur", il);
-                Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
-                cb(Vcur, "Vcur", il);
-
-                Qcur = ggml_rope_multi(
-                    ctx0,
-                    ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr,
-                    n_rot, sections, rope_type, n_ctx_orig, freq_base, freq_scale,
-                    ext_factor, attn_factor, beta_fast, beta_slow
-                );
-                cb(Qcur, "Qcur", il);
-
-                Kcur = ggml_rope_multi(
-                    ctx0,
-                    ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr,
-                    n_rot, sections, rope_type, n_ctx_orig, freq_base, freq_scale,
-                    ext_factor, attn_factor, beta_fast, beta_slow
-                );
-                cb(Kcur, "Kcur", il);
-
-                cur = llm_build_kv(ctx0, lctx, kv_self, gf,
-                        model.layers[il].wo, model.layers[il].bo,
-                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
-            }
-
-            if (il == n_layer - 1) {
-                // skip computing output for unused tokens
-                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
-                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
-                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
-            }
-
-            struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
-            cb(ffn_inp, "ffn_inp", il);
-
-            // feed-forward network
-            cur = llm_build_norm(ctx0, ffn_inp, hparams,
-                    model.layers[il].ffn_norm, NULL,
-                    LLM_NORM_RMS, cb, il);
-            cb(cur, "ffn_norm", il);
-
-            cur = llm_build_ffn(ctx0, lctx, cur,
-                    model.layers[il].ffn_up,   NULL, NULL,
-                    model.layers[il].ffn_gate, NULL, NULL,
-                    model.layers[il].ffn_down, NULL, NULL,
-                    NULL,
-                    LLM_FFN_SILU, LLM_FFN_PAR, cb, il);
-            cb(cur, "ffn_out", il);
-
-            cur = ggml_add(ctx0, cur, ffn_inp);
-            cur = lctx.cvec.apply_to(ctx0, cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-
-        cur = llm_build_norm(ctx0, cur, hparams,
-                model.output_norm, NULL,
-                LLM_NORM_RMS, cb, -1);
-        cb(cur, "result_norm", -1);
-
-        // lm_head
-        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
-        cb(cur, "result_output", -1);
-
-        ggml_build_forward_expand(gf, cur);
-
-        return gf;
-    }
-
-    struct ggml_cgraph * build_qwen2moe() {
-        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false);
-
-        // mutable variable, needed during the last layer of the computation to skip unused tokens
-        int32_t n_tokens = this->n_tokens;
-
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-        GGML_ASSERT(n_embd_head == hparams.n_rot);
-
-        struct ggml_tensor * cur;
-        struct ggml_tensor * inpL;
-
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
-
-        // inp_pos - contains the positions
-        struct ggml_tensor * inp_pos = build_inp_pos();
-
-        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
-        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
-
-        for (int il = 0; il < n_layer; ++il) {
-            struct ggml_tensor * inpSA = inpL;
-
-            // norm
-            cur = llm_build_norm(ctx0, inpL, hparams,
-                    model.layers[il].attn_norm, NULL,
-                    LLM_NORM_RMS, cb, il);
-            cb(cur, "attn_norm", il);
-
-            // self_attention
-            {
-                // compute Q and K and RoPE them
-                struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur);
-                cb(Qcur, "Qcur", il);
-                Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
-                cb(Qcur, "Qcur", il);
-
-                struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur);
-                cb(Kcur, "Kcur", il);
-                Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
-                cb(Kcur, "Kcur", il);
-
-                struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur);
-                cb(Vcur, "Vcur", il);
-                Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
-                cb(Vcur, "Vcur", il);
-
-                Qcur = ggml_rope_ext(
-                    ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr,
-                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                    ext_factor, attn_factor, beta_fast, beta_slow
-                );
-                cb(Qcur, "Qcur", il);
-
-                Kcur = ggml_rope_ext(
-                    ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr,
-                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                    ext_factor, attn_factor, beta_fast, beta_slow
-                );
-                cb(Kcur, "Kcur", il);
-
-                cur = llm_build_kv(ctx0, lctx, kv_self, gf,
-                        model.layers[il].wo, model.layers[il].bo,
-                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
-            }
-
-            if (il == n_layer - 1) {
-                // skip computing output for unused tokens
-                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
-                n_tokens = n_outputs;
-                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
-                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
-            }
-
-            struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
-            cb(ffn_inp, "ffn_inp", il);
-
-            // MoE branch
-            cur = llm_build_norm(ctx0, ffn_inp, hparams,
-                    model.layers[il].ffn_norm, NULL,
-                    LLM_NORM_RMS, cb, il);
-            cb(cur, "ffn_norm", il);
-
-            ggml_tensor * moe_out =
-                    llm_build_moe_ffn(ctx0, lctx, cur,
-                        model.layers[il].ffn_gate_inp,
-                        model.layers[il].ffn_up_exps,
-                        model.layers[il].ffn_gate_exps,
-                        model.layers[il].ffn_down_exps,
-                        nullptr,
-                        n_expert, n_expert_used,
-                        LLM_FFN_SILU, false,
-                        false, 0.0,
-                        LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
-                        cb, il);
-            cb(cur, "ffn_moe_out", il);
-
-            // FFN shared expert
-            {
-                ggml_tensor * cur_gate_inp = llm_build_lora_mm(lctx, ctx0, model.layers[il].ffn_gate_inp_shexp, cur);
-                cb(cur_gate_inp, "ffn_shexp_gate_inp", il);
-
-                // sigmoid
-                ggml_tensor * cur_gate = ggml_div(ctx0, ggml_silu(ctx0, cur_gate_inp), cur_gate_inp);
-                cb(cur_gate, "ffn_shexp_gate", il);
-
-                ggml_tensor * cur_ffn = llm_build_ffn(ctx0, lctx, cur,
-                        model.layers[il].ffn_up_shexp,   NULL, NULL,
-                        model.layers[il].ffn_gate_shexp, NULL, NULL,
-                        model.layers[il].ffn_down_shexp, NULL, NULL,
-                        NULL,
-                        LLM_FFN_SILU, LLM_FFN_PAR, cb, il);
-                cb(cur_ffn, "ffn_shexp", il);
-
-                ggml_tensor * ffn_shexp_out = ggml_mul(ctx0, cur_ffn, cur_gate);
-                cb(ffn_shexp_out, "ffn_shexp_out", il);
-
-                moe_out = ggml_add(ctx0, moe_out, ffn_shexp_out);
-                cb(moe_out, "ffn_out", il);
-
-                cur = moe_out;
-            }
-
-            cur = ggml_add(ctx0, cur, ffn_inp);
-            cur = lctx.cvec.apply_to(ctx0, cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-
-        cur = llm_build_norm(ctx0, cur, hparams,
-                model.output_norm, NULL,
-                LLM_NORM_RMS, cb, -1);
-        cb(cur, "result_norm", -1);
-
-        // lm_head
-        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
-        cb(cur, "result_output", -1);
-
-        ggml_build_forward_expand(gf, cur);
-
-        return gf;
-    }
-
-    struct ggml_cgraph * build_phi2() {
-        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false);
-
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-        const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-
-        struct ggml_tensor * cur;
-        struct ggml_tensor * attn_norm_output;
-        struct ggml_tensor * ffn_output;
-        struct ggml_tensor * inpL;
-
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
-
-        // inp_pos - contains the positions
-        struct ggml_tensor * inp_pos = build_inp_pos();
-
-        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
-        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
-
-        for (int il = 0; il < n_layer; ++il) {
-            attn_norm_output = llm_build_norm(ctx0, inpL, hparams,
-                    model.layers[il].attn_norm,
-                    model.layers[il].attn_norm_b,
-                    LLM_NORM, cb, il);
-            cb(attn_norm_output, "attn_norm", il);
-
-            // self-attention
-            {
-                struct ggml_tensor * Qcur = nullptr;
-                struct ggml_tensor * Kcur = nullptr;
-                struct ggml_tensor * Vcur = nullptr;
-
-                if (model.layers[il].wqkv) {
-                    cur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wqkv, attn_norm_output);
-                    cb(cur, "wqkv", il);
-
-                    cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
-                    cb(cur, "bqkv", il);
-
-                    Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd,     n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd)));
-                    Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd)));
-                    Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa)));
-                } else {
-                    Qcur = ggml_add(ctx0, llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, attn_norm_output), model.layers[il].bq);
-                    Kcur = ggml_add(ctx0, llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, attn_norm_output), model.layers[il].bk);
-                    Vcur = ggml_add(ctx0, llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, attn_norm_output), model.layers[il].bv);
-                }
-
-                cb(Qcur, "Qcur", il);
-                cb(Kcur, "Kcur", il);
-                cb(Vcur, "Vcur", il);
-
-                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
-                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
-
-                Qcur = ggml_rope_ext(
-                    ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig,
-                    freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow
-                );
-                cb(Qcur, "Qcur", il);
-
-                // with phi2, we scale the Q to avoid precision issues
-                // ref: https://github.com/ml-explore/mlx-examples/blob/08e862336ade809bc37d1035f94b359e7d1a5152/phi2/phi2.py#L64-L66
-                Qcur = ggml_scale(ctx0, Qcur, 1.0f/sqrtf(float(n_embd_head)));
-                cb(Qcur, "Qcur", il);
-
-                Kcur = ggml_rope_ext(
-                    ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig,
-                    freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow
-                );
-                cb(Kcur, "Kcur", il);
-
-                cur = llm_build_kv(ctx0, lctx, kv_self, gf,
-                        model.layers[il].wo, model.layers[il].bo,
-                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f, cb, il);
-            }
-
-            if (il == n_layer - 1) {
-                // skip computing output for unused tokens
-                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
-                cur              = ggml_get_rows(ctx0,              cur, inp_out_ids);
-                inpL             = ggml_get_rows(ctx0,             inpL, inp_out_ids);
-                attn_norm_output = ggml_get_rows(ctx0, attn_norm_output, inp_out_ids);
-            }
-
-            // FF
-            {
-                ffn_output = llm_build_ffn(ctx0, lctx, attn_norm_output,
-                        model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
-                        NULL,                      NULL,                        NULL,
-                        model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
-                        NULL,
-                        LLM_FFN_GELU, LLM_FFN_SEQ, cb, il);
-                cb(ffn_output, "ffn_out", il);
-            }
-
-            cur = ggml_add(ctx0, cur, ffn_output);
-            cur = ggml_add(ctx0, cur, inpL);
-            cur = lctx.cvec.apply_to(ctx0, cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = llm_build_norm(ctx0, inpL, hparams,
-                model.output_norm,
-                model.output_norm_b,
-                LLM_NORM, cb, -1);
-        cb(cur, "result_norm", -1);
-
-        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
-        cb(cur, "result_output_no_bias", -1);
-
-        cur = ggml_add(ctx0, cur, model.output_b);
-        cb(cur, "result_output", -1);
-        ggml_build_forward_expand(gf, cur);
-        return gf;
-    }
-
-    struct ggml_cgraph * build_phi3() {
-        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false);
-
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-        const int64_t n_embd_gqa = hparams.n_embd_v_gqa();
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-
-        struct ggml_tensor * cur;
-        struct ggml_tensor * inpL;
-
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
-
-        // inp_pos - contains the positions
-        struct ggml_tensor * inp_pos = build_inp_pos();
-
-        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
-        struct ggml_tensor * KQ_mask = nullptr;
-        if (hparams.n_swa == 0) {
-            // Phi-4 doesn't use sliding window attention
-            KQ_mask = build_inp_KQ_mask();
-        } else {
-            KQ_mask = build_inp_KQ_mask_swa();
-        }
-
-        for (int il = 0; il < n_layer; ++il) {
-            auto residual = inpL;
-
-            // self-attention
-            {
-                // rope freq factors for 128k context
-                struct ggml_tensor * rope_factors = build_rope_factors(il);
-
-                struct ggml_tensor* attn_norm_output = llm_build_norm(ctx0, inpL, hparams,
-                    model.layers[il].attn_norm,
-                    model.layers[il].attn_norm_b,
-                    LLM_NORM_RMS, cb, il);
-                cb(attn_norm_output, "attn_norm", il);
-
-                struct ggml_tensor * Qcur = nullptr;
-                struct ggml_tensor * Kcur = nullptr;
-                struct ggml_tensor * Vcur = nullptr;
-
-                if (model.layers[il].wqkv) {
-                    cur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wqkv, attn_norm_output);
-                    cb(cur, "wqkv", il);
-
-                    Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd,     n_tokens, cur->nb[1], 0 * sizeof(float) * (n_embd)));
-                    Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1 * sizeof(float) * (n_embd)));
-                    Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1 * sizeof(float) * (n_embd + n_embd_gqa)));
-                } else {
-                    Qcur = ggml_add(ctx0, llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, attn_norm_output), model.layers[il].bq);
-                    Kcur = ggml_add(ctx0, llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, attn_norm_output), model.layers[il].bk);
-                    Vcur = ggml_add(ctx0, llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, attn_norm_output), model.layers[il].bv);
-                }
-
-                cb(Qcur, "Qcur", il);
-                cb(Kcur, "Kcur", il);
-                cb(Vcur, "Vcur", il);
-
-                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
-                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
-
-                Qcur = ggml_rope_ext(
-                    ctx0, Qcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig,
-                    freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow
-                );
-                cb(Qcur, "Qcur", il);
-
-                Qcur = ggml_scale(ctx0, Qcur, 1.0f / sqrtf(float(n_embd_head)));
-                cb(Qcur, "Qcur", il);
-
-                Kcur = ggml_rope_ext(
-                    ctx0, Kcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig,
-                    freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow
-                );
-                cb(Kcur, "Kcur", il);
-
-                cur = llm_build_kv(ctx0, lctx, kv_self, gf,
-                        model.layers[il].wo, model.layers[il].bo,
-                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f, cb, il);
-            }
-
-            if (il == n_layer - 1) {
-                // skip computing output for unused tokens
-                struct ggml_tensor* inp_out_ids = build_inp_out_ids();
-                cur = ggml_get_rows(ctx0, cur, inp_out_ids);
-                residual = ggml_get_rows(ctx0, residual, inp_out_ids);
-            }
-
-            cur = ggml_add(ctx0, cur, residual);
-            residual = cur;
-
-            cur = llm_build_norm(ctx0, cur, hparams,
-                model.layers[il].ffn_norm, model.layers[il].ffn_norm_b,
-                LLM_NORM_RMS, cb, il);
-            cb(cur, "ffn_norm", il);
-
-            // feed-forward network
-            if (model.layers[il].ffn_gate_inp == nullptr) {
-                cur = llm_build_ffn(ctx0, lctx, cur,
-                        model.layers[il].ffn_up,   NULL, NULL,
-                        NULL,                      NULL, NULL,
-                        model.layers[il].ffn_down, NULL, NULL,
-                        NULL,
-                        LLM_FFN_SWIGLU, LLM_FFN_SEQ, cb, il);
-                cb(cur, "ffn_out", il);
-            } else {
-                // MoE branch
-                cur = llm_build_moe_ffn(ctx0, lctx, cur,
-                        model.layers[il].ffn_gate_inp,
-                        model.layers[il].ffn_up_exps,
-                        model.layers[il].ffn_gate_exps,
-                        model.layers[il].ffn_down_exps,
-                        nullptr,
-                        n_expert, n_expert_used,
-                        LLM_FFN_SILU, true,
-                        false, 0.0,
-                        LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
-                        cb, il);
-                cb(cur, "ffn_moe_out", il);
-            }
-
-            cur = ggml_add(ctx0, residual, cur);
-            cur = lctx.cvec.apply_to(ctx0, cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = llm_build_norm(ctx0, inpL, hparams,
-            model.output_norm,
-            model.output_norm_b,
-            LLM_NORM_RMS, cb, -1);
-        cb(cur, "result_norm", -1);
-
-        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
-
-        if (model.output_b != nullptr) {
-            cb(cur, "result_output_no_bias", -1);
-            cur = ggml_add(ctx0, cur, model.output_b);
-        }
-        cb(cur, "result_output", -1);
-
-        ggml_build_forward_expand(gf, cur);
-
-        return gf;
-    }
-
-
-    struct ggml_cgraph * build_plamo() {
-        struct ggml_cgraph * gf = ggml_new_graph(ctx0);
-
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-        GGML_ASSERT(n_embd_head == hparams.n_rot);
-
-        struct ggml_tensor * cur;
-        struct ggml_tensor * inpL;
-
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
-
-        // inp_pos - contains the positions
-        struct ggml_tensor * inp_pos = build_inp_pos();
-
-        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
-        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
-
-        for (int il = 0; il < n_layer; ++il) {
-
-            // norm
-            cur = llm_build_norm(ctx0, inpL, hparams,
-                    model.layers[il].attn_norm, NULL,
-                    LLM_NORM_RMS, cb, il);
-            cb(cur, "attn_norm", il);
-
-            struct ggml_tensor * attention_norm = cur;
-
-            // self-attention
-            {
-                // compute Q and K and RoPE them
-                struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur);
-                cb(Qcur, "Qcur", il);
-
-                struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur);
-                cb(Kcur, "Kcur", il);
-
-                struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur);
-                cb(Vcur, "Vcur", il);
-
-                Qcur = ggml_rope_ext(
-                        ctx0, ggml_reshape_3d(ctx0, Qcur, n_rot, n_head,    n_tokens), inp_pos, nullptr,
-                        n_embd_head, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow);
-                cb(Qcur, "Qcur", il);
-
-                Kcur = ggml_rope_ext(
-                        ctx0, ggml_reshape_3d(ctx0, Kcur, n_rot, n_head_kv, n_tokens), inp_pos, nullptr,
-                        n_embd_head, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow);
-                cb(Kcur, "Kcur", il);
-
-                cur = llm_build_kv(ctx0, lctx, kv_self, gf,
-                        model.layers[il].wo, NULL,
-                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
-            }
-            struct ggml_tensor * sa_out = cur;
-
-            cur = attention_norm;
-
-            if (il == n_layer - 1) {
-                // skip computing output for unused tokens
-                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
-                cur    = ggml_get_rows(ctx0,    cur, inp_out_ids);
-                sa_out = ggml_get_rows(ctx0, sa_out, inp_out_ids);
-                inpL   = ggml_get_rows(ctx0,   inpL, inp_out_ids);
-            }
-
-            // feed-forward network
-            {
-                cur = llm_build_ffn(ctx0, lctx, cur,
-                        model.layers[il].ffn_up,   NULL, NULL,
-                        model.layers[il].ffn_gate, NULL, NULL,
-                        model.layers[il].ffn_down, NULL, NULL,
-                        NULL,
-                        LLM_FFN_SILU, LLM_FFN_PAR, cb, il);
-                cb(cur, "ffn_out", il);
-            }
-
-            cur = ggml_add(ctx0, cur, sa_out);
-            cur = ggml_add(ctx0, cur, inpL);
-            cur = lctx.cvec.apply_to(ctx0, cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-
-        cur = llm_build_norm(ctx0, cur, hparams,
-                model.output_norm, NULL,
-                LLM_NORM_RMS, cb, -1);
-        cb(cur, "result_norm", -1);
-
-        // lm_head
-        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
-        cb(cur, "result_output", -1);
-
-        ggml_build_forward_expand(gf, cur);
-
-        return gf;
-    }
-
-    struct ggml_cgraph * build_gpt2() {
-        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false);
-
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-        const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-
-        struct ggml_tensor * cur;
-        struct ggml_tensor * pos;
-        struct ggml_tensor * inpL;
-
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
-
-        // inp_pos - contains the positions
-        struct ggml_tensor * inp_pos = build_inp_pos();
-
-        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
-        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
-
-        pos = ggml_get_rows(ctx0, model.pos_embd, inp_pos);
-        cb(pos, "pos_embd", -1);
-
-        inpL = ggml_add(ctx0, inpL, pos);
-        cb(inpL, "inpL", -1);
-
-        for (int il = 0; il < n_layer; ++il) {
-            cur = llm_build_norm(ctx0, inpL, hparams,
-                    model.layers[il].attn_norm,
-                    model.layers[il].attn_norm_b,
-                    LLM_NORM, cb, il);
-            cb(cur, "attn_norm", il);
-
-            // self-attention
-            {
-                cur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wqkv, cur);
-                cb(cur, "wqkv", il);
-
-                cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
-                cb(cur, "bqkv", il);
-
-                struct ggml_tensor * Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd,     n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd)));
-                struct ggml_tensor * Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd)));
-                struct ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa)));
-
-                cb(Qcur, "Qcur", il);
-                cb(Kcur, "Kcur", il);
-                cb(Vcur, "Vcur", il);
-
-                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
-
-                cur = llm_build_kv(ctx0, lctx, kv_self, gf,
-                        model.layers[il].wo, model.layers[il].bo,
-                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
-            }
-
-            if (il == n_layer - 1) {
-                // skip computing output for unused tokens
-                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
-                cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
-                inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
-            }
-
-            // add the input
-            struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL);
-            cb(ffn_inp, "ffn_inp", il);
-
-            // FF
-            {
-                cur = llm_build_norm(ctx0, ffn_inp, hparams,
-                        model.layers[il].ffn_norm,
-                        model.layers[il].ffn_norm_b,
-                        LLM_NORM, cb, il);
-                cb(cur, "ffn_norm", il);
-
-                cur = llm_build_ffn(ctx0, lctx, cur,
-                        model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
-                        NULL,                      NULL,                        NULL,
-                        model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
-                        NULL,
-                        LLM_FFN_GELU, LLM_FFN_SEQ, cb, il);
-                cb(cur, "ffn_out", il);
-            }
-
-            cur = ggml_add(ctx0, cur, ffn_inp);
-            cur = lctx.cvec.apply_to(ctx0, cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = llm_build_norm(ctx0, inpL, hparams,
-                model.output_norm,
-                model.output_norm_b,
-                LLM_NORM, cb, -1);
-        cb(cur, "result_norm", -1);
-
-        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
-        cb(cur, "result_output", -1);
-
-        ggml_build_forward_expand(gf, cur);
-
-        return gf;
-    }
-
-    struct ggml_cgraph * build_codeshell() {
-        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false);
-
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-        const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-        GGML_ASSERT(n_embd_head == hparams.n_rot);
-
-        struct ggml_tensor * cur;
-        struct ggml_tensor * inpL;
-
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
-
-        // inp_pos - contains the positions
-        struct ggml_tensor * inp_pos = build_inp_pos();
-
-        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
-        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
-
-        for (int il = 0; il < n_layer; ++il) {
-            cur = llm_build_norm(ctx0, inpL, hparams,
-                    model.layers[il].attn_norm,
-                    model.layers[il].attn_norm_b,
-                    LLM_NORM, cb, il);
-            cb(cur, "attn_norm", il);
-
-            // self-attention
-            {
-                cur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wqkv, cur);
-                cb(cur, "wqkv", il);
-
-                cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
-                cb(cur, "bqkv", il);
-
-                struct ggml_tensor * tmpq = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd,     n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd)));
-                struct ggml_tensor * tmpk = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd)));
-                struct ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa)));
-
-                cb(tmpq, "tmpq", il);
-                cb(tmpk, "tmpk", il);
-                cb(Vcur, "Vcur", il);
-
-                struct ggml_tensor * Qcur = ggml_rope_ext(
-                    ctx0, ggml_reshape_3d(ctx0, tmpq, n_embd_head, n_head,    n_tokens), inp_pos, nullptr,
-                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                    ext_factor, attn_factor, beta_fast, beta_slow
-                );
-                cb(Qcur, "Qcur", il);
-
-                struct ggml_tensor * Kcur = ggml_rope_ext(
-                    ctx0, ggml_reshape_3d(ctx0, tmpk, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr,
-                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                    ext_factor, attn_factor, beta_fast, beta_slow
-                );
-                cb(Kcur, "Kcur", il);
-
-                cur = llm_build_kv(ctx0, lctx, kv_self, gf,
-                        model.layers[il].wo, model.layers[il].bo,
-                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
-            }
-
-            if (il == n_layer - 1) {
-                // skip computing output for unused tokens
-                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
-                cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
-                inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
-            }
-
-            // add the input
-            struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL);
-            cb(ffn_inp, "ffn_inp", il);
-
-            // FF
-            {
-                cur = llm_build_norm(ctx0, ffn_inp, hparams,
-                        model.layers[il].ffn_norm,
-                        model.layers[il].ffn_norm_b,
-                        LLM_NORM, cb, il);
-                cb(cur, "ffn_norm", il);
-
-                cur = llm_build_ffn(ctx0, lctx, cur,
-                        model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
-                        NULL,                      NULL,                        NULL,
-                        model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
-                        NULL,
-                        LLM_FFN_GELU, LLM_FFN_SEQ, cb, il);
-                cb(cur, "ffn_out", il);
-            }
-
-            cur = ggml_add(ctx0, cur, ffn_inp);
-            cur = lctx.cvec.apply_to(ctx0, cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = llm_build_norm(ctx0, inpL, hparams,
-                model.output_norm,
-                model.output_norm_b,
-                LLM_NORM, cb, -1);
-        cb(cur, "result_norm", -1);
-
-        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
-        cb(cur, "result_output", -1);
-
-        ggml_build_forward_expand(gf, cur);
-
-        return gf;
-    }
-
-    struct ggml_cgraph * build_orion() {
-        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false);
-
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-        GGML_ASSERT(n_embd_head == hparams.n_rot);
-
-        struct ggml_tensor * cur;
-        struct ggml_tensor * inpL;
-
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
-
-        // inp_pos - contains the positions
-        struct ggml_tensor * inp_pos = build_inp_pos();
-
-        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
-        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
-
-        for (int il = 0; il < n_layer; ++il) {
-            struct ggml_tensor * inpSA = inpL;
-
-            // norm
-            cur = llm_build_norm(ctx0, inpL, hparams,
-                    model.layers[il].attn_norm, model.layers[il].attn_norm_b,
-                    LLM_NORM, cb, il);
-            cb(cur, "attn_norm", il);
-
-            // self-attention
-            {
-                // compute Q and K and RoPE them
-                struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur);
-                cb(Qcur, "Qcur", il);
-                // if (model.layers[il].bq) {
-                //     Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
-                //     cb(Qcur, "Qcur", il);
-                // }
-
-                struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur);
-                cb(Kcur, "Kcur", il);
-                // if (model.layers[il].bk) {
-                //     Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
-                //     cb(Kcur, "Kcur", il);
-                // }
-
-                struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur);
-                cb(Vcur, "Vcur", il);
-                // if (model.layers[il].bv) {
-                //     Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
-                //     cb(Vcur, "Vcur", il);
-                // }
-
-                Qcur = ggml_rope_ext(
-                    ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens), inp_pos, nullptr,
-                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                    ext_factor, attn_factor, beta_fast, beta_slow
-                );
-                cb(Qcur, "Qcur", il);
-
-                Kcur = ggml_rope_ext(
-                    ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr,
-                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                    ext_factor, attn_factor, beta_fast, beta_slow
-                );
-                cb(Kcur, "Kcur", il);
-
-                cur = llm_build_kv(ctx0, lctx, kv_self, gf,
-                        model.layers[il].wo, NULL,
-                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
-            }
-
-            if (il == n_layer - 1) {
-                // skip computing output for unused tokens
-                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
-                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
-                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
-            }
-
-            struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
-            cb(ffn_inp, "ffn_inp", il);
-
-            // feed-forward network
-            cur = llm_build_norm(ctx0, ffn_inp, hparams,
-                    model.layers[il].ffn_norm, model.layers[il].ffn_norm_b,
-                    LLM_NORM, cb, il);
-            cb(cur, "ffn_norm", il);
-
-            cur = llm_build_ffn(ctx0, lctx, cur,
-                    model.layers[il].ffn_up,   NULL, NULL,
-                    model.layers[il].ffn_gate, NULL, NULL,
-                    model.layers[il].ffn_down, NULL, NULL,
-                    NULL,
-                    LLM_FFN_SILU, LLM_FFN_PAR, cb, il);
-            cb(cur, "ffn_out", il);
-
-            cur = ggml_add(ctx0, cur, ffn_inp);
-            cur = lctx.cvec.apply_to(ctx0, cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-
-        cur = llm_build_norm(ctx0, cur, hparams,
-                model.output_norm, model.output_norm_b,
-                LLM_NORM, cb, -1);
-        cb(cur, "result_norm", -1);
-
-        // lm_head
-        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
-        cb(cur, "result_output", -1);
-
-        ggml_build_forward_expand(gf, cur);
-
-        return gf;
-    }
-
-    struct ggml_cgraph * build_internlm2() {
-        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false);
-
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-        GGML_ASSERT(n_embd_head == hparams.n_rot);
-
-        struct ggml_tensor * cur;
-        struct ggml_tensor * inpL;
-
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
-
-        // inp_pos - contains the positions
-        struct ggml_tensor * inp_pos = build_inp_pos();
-
-        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
-        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
-
-        for (int il = 0; il < n_layer; ++il) {
-            struct ggml_tensor * inpSA = inpL;
-
-            // norm
-            cur = llm_build_norm(ctx0, inpL, hparams,
-                    model.layers[il].attn_norm, NULL,
-                    LLM_NORM_RMS, cb, il);
-            cb(cur, "attn_norm", il);
-
-            // self-attention
-            {
-                // compute Q and K and RoPE them
-                struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur);
-                cb(Qcur, "Qcur", il);
-                if (model.layers[il].bq) {
-                    Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
-                    cb(Qcur, "Qcur", il);
-                }
-
-                struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur);
-                cb(Kcur, "Kcur", il);
-                if (model.layers[il].bk) {
-                    Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
-                    cb(Kcur, "Kcur", il);
-                }
-
-                struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur);
-                cb(Vcur, "Vcur", il);
-                if (model.layers[il].bv) {
-                    Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
-                    cb(Vcur, "Vcur", il);
-                }
-
-                Qcur = ggml_rope_ext(
-                    ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens), inp_pos, nullptr,
-                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                    ext_factor, attn_factor, beta_fast, beta_slow
-                );
-                cb(Qcur, "Qcur", il);
-
-                Kcur = ggml_rope_ext(
-                    ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr,
-                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                    ext_factor, attn_factor, beta_fast, beta_slow
-                );
-                cb(Kcur, "Kcur", il);
-
-                cur = llm_build_kv(ctx0, lctx, kv_self, gf,
-                        model.layers[il].wo, model.layers[il].bo,
-                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
-            }
-
-            if (il == n_layer - 1) {
-                // skip computing output for unused tokens
-                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
-                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
-                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
-            }
-
-            struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
-            cb(ffn_inp, "ffn_inp", il);
-
-            // feed-forward network
-            cur = llm_build_norm(ctx0, ffn_inp, hparams,
-                    model.layers[il].ffn_norm, NULL,
-                    LLM_NORM_RMS, cb, il);
-            cb(cur, "ffn_norm", il);
-
-            cur = llm_build_ffn(ctx0, lctx, cur,
-                    model.layers[il].ffn_up,   NULL, NULL,
-                    model.layers[il].ffn_gate, NULL, NULL,
-                    model.layers[il].ffn_down, NULL, NULL,
-                    NULL,
-                    LLM_FFN_SILU, LLM_FFN_PAR, cb, il);
-            cb(cur, "ffn_out", il);
-
-            cur = ggml_add(ctx0, cur, ffn_inp);
-            cur = lctx.cvec.apply_to(ctx0, cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-
-        cur = llm_build_norm(ctx0, cur, hparams,
-                model.output_norm, NULL,
-                LLM_NORM_RMS, cb, -1);
-        cb(cur, "result_norm", -1);
-
-        // lm_head
-        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
-        cb(cur, "result_output", -1);
-
-        ggml_build_forward_expand(gf, cur);
-
-        return gf;
-    }
-
-    struct ggml_cgraph * build_minicpm3() {
-        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false);
-
-        //TODO: if the model varies, these parameters need to be read from the model
-        const int64_t n_embd_base = 256;
-        const float scale_embd  = 12.0f;
-        const float scale_depth = 1.4f;
-        const float kq_scale = 1.0f / sqrtf(float(hparams.n_embd_head_k));
-
-        const uint32_t n_embd_head_qk_rope = hparams.n_rot;
-        const uint32_t n_embd_head_qk_nope = hparams.n_embd_head_k - hparams.n_rot;
-        const uint32_t kv_lora_rank = hparams.n_lora_kv;
-
-        struct ggml_tensor * cur;
-        struct ggml_tensor * inpL;
-
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
-
-        // scale the input embeddings
-        inpL = ggml_scale(ctx0, inpL, scale_embd);
-        cb(inpL, "inp_scaled", -1);
-
-        // inp_pos - contains the positions
-        struct ggml_tensor * inp_pos = build_inp_pos();
-
-        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
-        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
-
-        for (int il = 0; il < n_layer; ++il) {
-            struct ggml_tensor * inpSA = inpL;
-
-            struct ggml_tensor * rope_factors = build_rope_factors(il);
-            // norm
-            cur = llm_build_norm(ctx0, inpL, hparams,
-                    model.layers[il].attn_norm, NULL,
-                    LLM_NORM_RMS, cb, il);
-            cb(cur, "attn_norm", il);
-
-            // self_attention
-            {
-                struct ggml_tensor * q = NULL;
-                // {n_embd, q_lora_rank} * {n_embd, n_tokens} -> {q_lora_rank, n_tokens}
-                q = ggml_mul_mat(ctx0, model.layers[il].wq_a, cur);
-                cb(q, "q", il);
-
-                q = llm_build_norm(ctx0, q, hparams,
-                        model.layers[il].attn_q_a_norm, NULL,
-                        LLM_NORM_RMS, cb, il);
-                cb(q, "q", il);
-
-                // {q_lora_rank, n_head * hparams.n_embd_head_k} * {q_lora_rank, n_tokens} -> {n_head * hparams.n_embd_head_k, n_tokens}
-                q = ggml_mul_mat(ctx0, model.layers[il].wq_b, q);
-                cb(q, "q", il);
-
-                // split into {n_head * n_embd_head_qk_nope, n_tokens}
-                struct ggml_tensor * q_nope = ggml_view_3d(ctx0, q, n_embd_head_qk_nope, n_head, n_tokens,
-                        ggml_row_size(q->type, hparams.n_embd_head_k),
-                        ggml_row_size(q->type, hparams.n_embd_head_k * n_head),
-                        0);
-                cb(q_nope, "q_nope", il);
-
-                // and {n_head * n_embd_head_qk_rope, n_tokens}
-                struct ggml_tensor * q_pe = ggml_view_3d(ctx0, q, n_embd_head_qk_rope, n_head, n_tokens,
-                        ggml_row_size(q->type, hparams.n_embd_head_k),
-                        ggml_row_size(q->type, hparams.n_embd_head_k * n_head),
-                        ggml_row_size(q->type, n_embd_head_qk_nope));
-                cb(q_pe, "q_pe", il);
-
-                // {n_embd, kv_lora_rank + n_embd_head_qk_rope} * {n_embd, n_tokens} -> {kv_lora_rank + n_embd_head_qk_rope, n_tokens}
-                struct ggml_tensor * kv_pe_compresseed = ggml_mul_mat(ctx0, model.layers[il].wkv_a_mqa, cur);
-                cb(kv_pe_compresseed, "kv_pe_compresseed", il);
-
-                // split into {kv_lora_rank, n_tokens}
-                struct ggml_tensor * kv_compressed = ggml_view_2d(ctx0, kv_pe_compresseed, kv_lora_rank, n_tokens,
-                        kv_pe_compresseed->nb[1],
-                        0);
-                cb(kv_compressed, "kv_compressed", il);
-
-                // and {n_embd_head_qk_rope, n_tokens}
-                struct ggml_tensor * k_pe = ggml_view_3d(ctx0, kv_pe_compresseed, n_embd_head_qk_rope, 1, n_tokens,
-                        kv_pe_compresseed->nb[1],
-                        kv_pe_compresseed->nb[1],
-                        ggml_row_size(kv_pe_compresseed->type, kv_lora_rank));
-                cb(k_pe, "k_pe", il);
-
-                // TODO: the CUDA backend used to not support non-cont. (RMS) norm, investigate removing ggml_cont
-                kv_compressed = ggml_cont(ctx0, kv_compressed);
-                kv_compressed = llm_build_norm(ctx0, kv_compressed, hparams,
-                        model.layers[il].attn_kv_a_norm, NULL,
-                        LLM_NORM_RMS, cb, il);
-                cb(kv_compressed, "kv_compressed", il);
-
-                // {kv_lora_rank, n_head * (n_embd_head_qk_nope + n_embd_head_v)} * {kv_lora_rank, n_tokens} -> {n_head * (n_embd_head_qk_nope + n_embd_head_v), n_tokens}
-                struct ggml_tensor * kv = ggml_mul_mat(ctx0, model.layers[il].wkv_b, kv_compressed);
-                cb(kv, "kv", il);
-
-                // split into {n_head * n_embd_head_qk_nope, n_tokens}
-                struct ggml_tensor * k_nope = ggml_view_3d(ctx0, kv, n_embd_head_qk_nope, n_head, n_tokens,
-                        ggml_row_size(kv->type, n_embd_head_qk_nope + hparams.n_embd_head_v),
-                        ggml_row_size(kv->type, n_head * (n_embd_head_qk_nope + hparams.n_embd_head_v)),
-                        0);
-                cb(k_nope, "k_nope", il);
-
-                // and {n_head * n_embd_head_v, n_tokens}
-                struct ggml_tensor * v_states = ggml_view_3d(ctx0, kv, hparams.n_embd_head_v, n_head, n_tokens,
-                        ggml_row_size(kv->type, (n_embd_head_qk_nope + hparams.n_embd_head_v)),
-                        ggml_row_size(kv->type, (n_embd_head_qk_nope + hparams.n_embd_head_v)*n_head),
-                        ggml_row_size(kv->type, (n_embd_head_qk_nope)));
-                cb(v_states, "v_states", il);
-
-                v_states = ggml_cont(ctx0, v_states);
-                cb(v_states, "v_states", il);
-
-                v_states = ggml_view_2d(ctx0, v_states, hparams.n_embd_head_v * n_head, n_tokens,
-                    ggml_row_size(kv->type, hparams.n_embd_head_v * n_head),
-                    0);
-                cb(v_states, "v_states", il);
-
-                q_pe = ggml_cont(ctx0, q_pe); // TODO: the CUDA backend used to not support non-cont. RoPE, investigate removing this
-                q_pe = ggml_rope_ext(
-                    ctx0, q_pe, inp_pos, rope_factors,
-                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                    ext_factor, attn_factor, beta_fast, beta_slow
-                );
-                cb(q_pe, "q_pe", il);
-
-                // shared RoPE key
-                k_pe = ggml_cont(ctx0, k_pe); // TODO: the CUDA backend used to not support non-cont. RoPE, investigate removing this
-                k_pe = ggml_rope_ext(
-                    ctx0, k_pe, inp_pos, rope_factors,
-                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                    ext_factor, attn_factor, beta_fast, beta_slow
-                );
-                cb(k_pe, "k_pe", il);
-
-                struct ggml_tensor * q_states = ggml_concat(ctx0, q_nope, q_pe, 0);
-                cb(q_states, "q_states", il);
-
-                struct ggml_tensor * k_states = ggml_concat(ctx0, k_nope, ggml_repeat(ctx0, k_pe, q_pe), 0);
-                cb(k_states, "k_states", il);
-
-                cur = llm_build_kv(ctx0, lctx, kv_self, gf,
-                        model.layers[il].wo, NULL,
-                        k_states, v_states, q_states, KQ_mask, n_tokens, kv_head, n_kv, kq_scale, cb, il);
-            }
-
-            if (il == n_layer - 1) {
-                // skip computing output for unused tokens
-                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
-                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
-                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
-            }
-
-            // scale_res - scale the hidden states for residual connection
-            const float scale_res = scale_depth/sqrtf(float(n_layer));
-            cur = ggml_scale(ctx0, cur, scale_res);
-            cb(cur, "hidden_scaled", il);
-
-            struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
-            cb(ffn_inp, "ffn_inp", il);
-
-            // feed-forward network
-            {
-                cur = llm_build_norm(ctx0, ffn_inp, hparams,
-                        model.layers[il].ffn_norm, NULL,
-                        LLM_NORM_RMS, cb, il);
-                cb(cur, "ffn_norm", il);
-
-                cur = llm_build_ffn(ctx0, lctx, cur,
-                        model.layers[il].ffn_up,   NULL, NULL,
-                        model.layers[il].ffn_gate, NULL, NULL,
-                        model.layers[il].ffn_down, NULL, NULL,
-                        NULL,
-                        LLM_FFN_SILU, LLM_FFN_PAR, cb, il);
-                cb(cur, "ffn_out", il);
-            }
-
-            // scale the hidden states for residual connection
-            cur = ggml_scale(ctx0, cur, scale_res);
-            cb(cur, "hidden_scaled_ffn", il);
-
-            cur = ggml_add(ctx0, cur, ffn_inp);
-            cur = lctx.cvec.apply_to(ctx0, cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-
-        cur = llm_build_norm(ctx0, cur, hparams,
-                model.output_norm, NULL,
-                LLM_NORM_RMS, cb, -1);
-        cb(cur, "result_norm", -1);
-
-        // lm_head scaling
-        const float scale_lmhead = float(n_embd_base)/float(n_embd);
-        cur = ggml_scale(ctx0, cur, scale_lmhead);
-        cb(cur, "lmhead_scaling", -1);
-
-        // lm_head
-        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
-        cb(cur, "result_output", -1);
-
-        ggml_build_forward_expand(gf, cur);
-
-        return gf;
-    }
-
-    struct ggml_cgraph * build_gemma() {
-        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false);
-
-        const int64_t n_embd_head_k = hparams.n_embd_head_k;
-
-        struct ggml_tensor * cur;
-        struct ggml_tensor * inpL;
-
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
-
-        inpL = ggml_scale(ctx0, inpL, sqrtf(n_embd));
-        cb(inpL, "inp_scaled", -1);
-
-        // inp_pos - contains the positions
-        struct ggml_tensor * inp_pos = build_inp_pos();
-
-        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
-        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
-
-        for (int il = 0; il < n_layer; ++il) {
-            // norm
-            cur = llm_build_norm(ctx0, inpL, hparams,
-                    model.layers[il].attn_norm, NULL,
-                    LLM_NORM_RMS, cb, il);
-            cb(cur, "attn_norm", il);
-
-            // self-attention
-            {
-                // compute Q and K and RoPE them
-                struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur);
-                cb(Qcur, "Qcur", il);
-
-                struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur);
-                cb(Kcur, "Kcur", il);
-
-                struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur);
-                cb(Vcur, "Vcur", il);
-
-                Qcur = ggml_rope_ext(
-                        ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head_k, n_head,    n_tokens), inp_pos, nullptr,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow);
-                cb(Qcur, "Qcur", il);
-
-                Qcur = ggml_scale(ctx0, Qcur, 1.0f / sqrtf(float(n_embd_head_k)));
-                cb(Qcur, "Qcur_scaled", il);
-
-                Kcur = ggml_rope_ext(
-                        ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head_k, n_head_kv, n_tokens), inp_pos, nullptr,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow);
-                cb(Kcur, "Kcur", il);
-
-                cur = llm_build_kv(ctx0, lctx, kv_self, gf,
-                        model.layers[il].wo, NULL,
-                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f, cb, il);
-            }
-
-            if (il == n_layer - 1) {
-                // skip computing output for unused tokens
-                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
-                cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
-                inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
-            }
-
-            struct ggml_tensor * sa_out = ggml_add(ctx0, cur, inpL);
-            cb(sa_out, "sa_out", il);
-
-            cur = llm_build_norm(ctx0, sa_out, hparams,
-                    model.layers[il].ffn_norm, NULL,
-                    LLM_NORM_RMS, cb, il);
-            cb(cur, "ffn_norm", il);
-
-            // feed-forward network
-            {
-                cur = llm_build_ffn(ctx0, lctx, cur,
-                        model.layers[il].ffn_up,   NULL, NULL,
-                        model.layers[il].ffn_gate, NULL, NULL,
-                        model.layers[il].ffn_down, NULL, NULL,
-                        NULL,
-                        LLM_FFN_GELU, LLM_FFN_PAR, cb, il);
-                cb(cur, "ffn_out", il);
-            }
-
-            cur = ggml_add(ctx0, cur, sa_out);
-            cur = lctx.cvec.apply_to(ctx0, cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-
-        cur = llm_build_norm(ctx0, cur, hparams,
-                model.output_norm, NULL,
-                LLM_NORM_RMS, cb, -1);
-        cb(cur, "result_norm", -1);
-
-        // lm_head
-        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
-        cb(cur, "result_output", -1);
-
-        ggml_build_forward_expand(gf, cur);
-
-        return gf;
-    }
-
-    struct ggml_cgraph * build_gemma2() {
-        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false);
-
-        const int64_t n_embd_head_k = hparams.n_embd_head_k;
-
-        struct ggml_tensor * cur;
-        struct ggml_tensor * inpL;
-
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
-
-        inpL = ggml_scale(ctx0, inpL, sqrtf(n_embd));
-        cb(inpL, "inp_scaled", -1);
-
-        // inp_pos - contains the positions
-        struct ggml_tensor * inp_pos = build_inp_pos();
-
-        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
-        // gemma 2 requires different mask for layers using sliding window (SWA)
-        struct ggml_tensor * KQ_mask     = build_inp_KQ_mask(true);
-        struct ggml_tensor * KQ_mask_swa = build_inp_KQ_mask_swa(true);
-
-        for (int il = 0; il < n_layer; ++il) {
-            // (il % 2) layers use SWA
-            struct ggml_tensor * KQ_mask_l = (il % 2 == 0) ? KQ_mask_swa : KQ_mask;
-
-            // norm
-            cur = llm_build_norm(ctx0, inpL, hparams,
-                    model.layers[il].attn_norm, NULL,
-                    LLM_NORM_RMS, cb, il);
-            cb(cur, "attn_norm", il);
-
-            // self-attention
-            {
-                // compute Q and K and RoPE them
-                struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur);
-                cb(Qcur, "Qcur", il);
-
-                struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur);
-                cb(Kcur, "Kcur", il);
-
-                struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur);
-                cb(Vcur, "Vcur", il);
-
-                Qcur = ggml_rope_ext(
-                        ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head_k, n_head,    n_tokens), inp_pos, nullptr,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow);
-                cb(Qcur, "Qcur", il);
-
-                // ref: https://github.com/google/gemma_pytorch/commit/03e657582d17cb5a8617ebf333c1c16f3694670e
-                switch (model.type) {
-                    case LLM_TYPE_2B:
-                    case LLM_TYPE_9B:  Qcur = ggml_scale(ctx0, Qcur, 1.0f / sqrtf(float(n_embd_head_k)));   break;
-                    case LLM_TYPE_27B: Qcur = ggml_scale(ctx0, Qcur, 1.0f / sqrtf(float(n_embd / n_head))); break;
-                    default: GGML_ABORT("fatal error");
-                };
-                cb(Qcur, "Qcur_scaled", il);
-
-                Kcur = ggml_rope_ext(
-                        ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head_k, n_head_kv, n_tokens), inp_pos, nullptr,
-                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                        ext_factor, attn_factor, beta_fast, beta_slow);
-                cb(Kcur, "Kcur", il);
-
-                cur = llm_build_kv(ctx0, lctx, kv_self, gf,
-                        model.layers[il].wo, NULL,
-                        Kcur, Vcur, Qcur, KQ_mask_l, n_tokens, kv_head, n_kv, 1.0f, cb, il);
-            }
-
-            cur = llm_build_norm(ctx0, cur, hparams,
-                    model.layers[il].attn_post_norm, NULL,
-                    LLM_NORM_RMS, cb, il);
-            cb(cur, "attn_post_norm", il);
-
-            if (il == n_layer - 1) {
-                // skip computing output for unused tokens
-                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
-                cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
-                inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
-            }
-
-            struct ggml_tensor * sa_out = ggml_add(ctx0, cur, inpL);
-            cb(sa_out, "sa_out", il);
-
-            cur = llm_build_norm(ctx0, sa_out, hparams,
-                    model.layers[il].ffn_norm, NULL,
-                    LLM_NORM_RMS, cb, il);
-            cb(cur, "ffn_norm", il);
-
-            // feed-forward network
-            {
-                cur = llm_build_ffn(ctx0, lctx, cur,
-                        model.layers[il].ffn_up,   NULL, NULL,
-                        model.layers[il].ffn_gate, NULL, NULL,
-                        model.layers[il].ffn_down, NULL, NULL,
-                        NULL,
-                        LLM_FFN_GELU, LLM_FFN_PAR, cb, il);
-                cb(cur, "ffn_out", il);
-            }
-
-            cur = llm_build_norm(ctx0, cur, hparams,
-                model.layers[il].ffn_post_norm, NULL,
-                LLM_NORM_RMS, cb, -1);
-            cb(cur, "ffn_post_norm", -1);
-
-            cur = ggml_add(ctx0, cur, sa_out);
-            cur = lctx.cvec.apply_to(ctx0, cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-
-        cur = llm_build_norm(ctx0, cur, hparams,
-                model.output_norm, NULL,
-                LLM_NORM_RMS, cb, -1);
-        cb(cur, "result_norm", -1);
-
-        // lm_head
-        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
-
-        // final logit soft-capping
-        cur = ggml_scale(ctx0, cur, 1.0f / hparams.f_final_logit_softcapping);
-        cur = ggml_tanh(ctx0, cur);
-        cur = ggml_scale(ctx0, cur, hparams.f_final_logit_softcapping);
-
-        cb(cur, "result_output", -1);
-
-        ggml_build_forward_expand(gf, cur);
-
-        return gf;
-    }
-
-
-    struct ggml_cgraph * build_starcoder2() {
-        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false);
-
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-        GGML_ASSERT(n_embd_head == hparams.n_rot);
-
-        struct ggml_tensor * cur;
-        struct ggml_tensor * inpL;
-
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
-
-        // inp_pos - contains the positions
-        struct ggml_tensor * inp_pos = build_inp_pos();
-
-        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
-        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
-
-        for (int il = 0; il < n_layer; ++il) {
-            struct ggml_tensor * inpSA = inpL;
-
-            // norm
-            cur = llm_build_norm(ctx0, inpL, hparams,
-                    model.layers[il].attn_norm, model.layers[il].attn_norm_b,
-                    LLM_NORM, cb, il);
-            cb(cur, "attn_norm", il);
-
-            // self-attention
-            {
-                // compute Q and K and RoPE them
-                struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur);
-                cb(Qcur, "Qcur", il);
-                if (model.layers[il].bq) {
-                    Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
-                    cb(Qcur, "Qcur", il);
-                }
-
-                struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur);
-                cb(Kcur, "Kcur", il);
-                if (model.layers[il].bk) {
-                    Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
-                    cb(Kcur, "Kcur", il);
-                }
-
-                struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur);
-                cb(Vcur, "Vcur", il);
-                if (model.layers[il].bv) {
-                    Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
-                    cb(Vcur, "Vcur", il);
-                }
-
-                Qcur = ggml_rope_ext(
-                    ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr,
-                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                    ext_factor, attn_factor, beta_fast, beta_slow
-                );
-                cb(Qcur, "Qcur", il);
-
-                Kcur = ggml_rope_ext(
-                    ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr,
-                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                    ext_factor, attn_factor, beta_fast, beta_slow
-                );
-                cb(Kcur, "Kcur", il);
-
-                cur = llm_build_kv(ctx0, lctx, kv_self, gf,
-                        model.layers[il].wo, model.layers[il].bo,
-                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
-            }
-
-            if (il == n_layer - 1) {
-                // skip computing output for unused tokens
-                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
-                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
-                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
-            }
-
-            struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
-            cb(ffn_inp, "ffn_inp", il);
-
-            // feed-forward network
-
-            cur = llm_build_norm(ctx0, ffn_inp, hparams,
-                    model.layers[il].ffn_norm, model.layers[il].ffn_norm_b,
-                    LLM_NORM, cb, il);
-            cb(cur, "ffn_norm", il);
-
-            cur = llm_build_ffn(ctx0, lctx, cur,
-                        model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
-                        NULL,                      NULL,                        NULL,
-                        model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
-                        NULL,
-                        LLM_FFN_GELU, LLM_FFN_SEQ, cb, il);
-            cb(cur, "ffn_out", il);
-
-            cur = ggml_add(ctx0, cur, ffn_inp);
-            cur = lctx.cvec.apply_to(ctx0, cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-
-        cur = llm_build_norm(ctx0, cur, hparams,
-                model.output_norm, model.output_norm_b,
-                LLM_NORM, cb, -1);
-        cb(cur, "result_norm", -1);
-
-        // lm_head
-        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
-        cb(cur, "result_output", -1);
-
-        ggml_build_forward_expand(gf, cur);
-
-        return gf;
-    }
-
-    struct ggml_cgraph * build_mamba() {
-        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false);
-
-        struct ggml_tensor * cur;
-        struct ggml_tensor * inpL;
-
-        // {n_embd, n_tokens}
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
-
-        struct ggml_tensor * state_copy = build_inp_s_copy();
-        struct ggml_tensor * state_mask = build_inp_s_mask();
-
-        for (int il = 0; il < n_layer; ++il) {
-            // norm
-            cur = llm_build_norm(ctx0, inpL, hparams,
-                    model.layers[il].attn_norm, NULL,
-                    LLM_NORM_RMS, cb, il);
-            cb(cur, "attn_norm", il);
-
-            cur = llm_build_mamba(ctx0, lctx, ubatch, gf, cur,
-                    state_copy, state_mask,
-                    kv_head, n_kv, cb, il);
-
-            if (il == n_layer - 1) {
-                // skip computing output for unused tokens
-                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
-                cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
-                inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
-            }
-
-            // residual
-            cur = ggml_add(ctx0, cur, inpL);
-            cur = lctx.cvec.apply_to(ctx0, cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        // final rmsnorm
-        cur = llm_build_norm(ctx0, inpL, hparams,
-                model.output_norm, NULL,
-                LLM_NORM_RMS, cb, -1);
-        cb(cur, "result_norm", -1);
-
-        // lm_head
-        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
-        cb(cur, "result_output", -1);
-
-        ggml_build_forward_expand(gf, cur);
-
-        return gf;
-    }
-
-    struct ggml_cgraph * build_command_r() {
-
-        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false);
-
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-        const float f_logit_scale = hparams.f_logit_scale;
-
-        struct ggml_tensor * cur;
-        struct ggml_tensor * inpL;
-
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
-
-        // inp_pos - contains the positions
-        struct ggml_tensor * inp_pos = build_inp_pos();
-
-        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
-        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
-
-        for (int il = 0; il < n_layer; ++il) {
-
-            // norm
-            cur = llm_build_norm(ctx0, inpL, hparams,
-                    model.layers[il].attn_norm, NULL,
-                    LLM_NORM, cb, il);
-            cb(cur, "attn_norm", il);
-            struct ggml_tensor * ffn_inp = cur;
-
-            // self-attention
-            {
-                // compute Q and K and RoPE them
-                struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur);
-                cb(Qcur, "Qcur", il);
-                if (model.layers[il].bq) {
-                    Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
-                    cb(Qcur, "Qcur", il);
-                }
-
-                struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur);
-                cb(Kcur, "Kcur", il);
-                if (model.layers[il].bk) {
-                    Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
-                    cb(Kcur, "Kcur", il);
-                }
-
-                struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur);
-                cb(Vcur, "Vcur", il);
-                if (model.layers[il].bv) {
-                    Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
-                    cb(Vcur, "Vcur", il);
-                }
-
-                if (model.layers[il].attn_q_norm) {
-                    Qcur = ggml_view_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens,
-                                ggml_element_size(Qcur) * n_embd_head,
-                                ggml_element_size(Qcur) * n_embd_head * n_head,
-                                0);
-                    cb(Qcur, "Qcur", il);
-                    Kcur = ggml_view_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens,
-                                ggml_element_size(Kcur) * n_embd_head,
-                                ggml_element_size(Kcur) * n_embd_head * n_head_kv,
-                                0);
-                    cb(Kcur, "Kcur", il);
-
-                    Qcur = llm_build_norm(ctx0, Qcur, hparams,
-                                model.layers[il].attn_q_norm,
-                                NULL,
-                                LLM_NORM, cb, il);
-                    cb(Qcur, "Qcur", il);
-
-                    Kcur = llm_build_norm(ctx0, Kcur, hparams,
-                            model.layers[il].attn_k_norm,
-                            NULL,
-                            LLM_NORM, cb, il);
-                    cb(Kcur, "Kcur", il);
-                }
-
-                Qcur = ggml_rope_ext(
-                    ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr,
-                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                    ext_factor, attn_factor, beta_fast, beta_slow
-                );
-                cb(Qcur, "Qcur", il);
-
-                Kcur = ggml_rope_ext(
-                    ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr,
-                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                    ext_factor, attn_factor, beta_fast, beta_slow
-                );
-                cb(Kcur, "Kcur", il);
-
-                cur = llm_build_kv(ctx0, lctx, kv_self, gf,
-                        model.layers[il].wo, model.layers[il].bo,
-                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
-            }
-
-            if (il == n_layer - 1) {
-                // skip computing output for unused tokens
-                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
-                cur     = ggml_get_rows(ctx0,     cur, inp_out_ids);
-                inpL    = ggml_get_rows(ctx0,    inpL, inp_out_ids);
-                ffn_inp = ggml_get_rows(ctx0, ffn_inp, inp_out_ids);
-            }
-
-            struct ggml_tensor * attn_out = cur;
-
-            // feed-forward network
-            {
-                cur = llm_build_ffn(ctx0, lctx, ffn_inp,
-                        model.layers[il].ffn_up,   NULL, NULL,
-                        model.layers[il].ffn_gate, NULL, NULL,
-                        model.layers[il].ffn_down, NULL, NULL,
-                        NULL,
-                        LLM_FFN_SILU, LLM_FFN_PAR, cb, il);
-                cb(cur, "ffn_out", il);
-            }
-
-            // add together residual + FFN + self-attention
-            cur = ggml_add(ctx0, cur, inpL);
-            cur = ggml_add(ctx0, cur, attn_out);
-            cur = lctx.cvec.apply_to(ctx0, cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-
-        cur = llm_build_norm(ctx0, cur, hparams,
-                model.output_norm, NULL,
-                LLM_NORM, cb, -1);
-        cb(cur, "result_norm", -1);
-
-        // lm_head
-        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
-
-        if (f_logit_scale) {
-            cur = ggml_scale(ctx0, cur, f_logit_scale);
-        }
-
-        cb(cur, "result_output", -1);
-
-        ggml_build_forward_expand(gf, cur);
-
-        return gf;
-
-    }
-
-    struct ggml_cgraph * build_cohere2() {
-        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false);
-
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-        const float f_logit_scale = hparams.f_logit_scale;
-
-        struct ggml_tensor * cur;
-        struct ggml_tensor * inpL;
-
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
-
-        // inp_pos - contains the positions
-        struct ggml_tensor * inp_pos = build_inp_pos();
-
-        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
-        // cohere2 requires different mask for layers using sliding window (SWA)
-        struct ggml_tensor * KQ_mask     = build_inp_KQ_mask();
-        struct ggml_tensor * KQ_mask_swa = build_inp_KQ_mask_swa();
-
-        // sliding window switch pattern
-        const int32_t sliding_window_pattern = 4;
-
-        for (int il = 0; il < n_layer; ++il) {
-            // three layers sliding window attention (window size 4096) and ROPE
-            // fourth layer uses global attention without positional embeddings
-            const bool           is_sliding = il % sliding_window_pattern < (sliding_window_pattern - 1);
-            struct ggml_tensor * KQ_mask_l = is_sliding ? KQ_mask_swa : KQ_mask;
-
-            // norm
-            cur = llm_build_norm(ctx0, inpL, hparams, model.layers[il].attn_norm, NULL, LLM_NORM, cb, il);
-            cb(cur, "attn_norm", il);
-            struct ggml_tensor * ffn_inp = cur;
-
-            // self-attention
-            {
-                // rope freq factors for 128k context
-                struct ggml_tensor * rope_factors = build_rope_factors(il);
-
-                // compute Q and K and RoPE them
-                struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur);
-                cb(Qcur, "Qcur", il);
-                if (model.layers[il].bq) {
-                    Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
-                    cb(Qcur, "Qcur", il);
-                }
-
-                struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur);
-                cb(Kcur, "Kcur", il);
-                if (model.layers[il].bk) {
-                    Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
-                    cb(Kcur, "Kcur", il);
-                }
-
-                struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur);
-                cb(Vcur, "Vcur", il);
-                if (model.layers[il].bv) {
-                    Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
-                    cb(Vcur, "Vcur", il);
-                }
-
-                if (is_sliding) {
-                    Qcur = ggml_rope_ext(ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, rope_factors,
-                                        n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, ext_factor, attn_factor,
-                                        beta_fast, beta_slow);
-                    cb(Qcur, "Qcur", il);
-
-                    Kcur = ggml_rope_ext(ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos,
-                                        rope_factors, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, ext_factor,
-                                        attn_factor, beta_fast, beta_slow);
-                    cb(Kcur, "Kcur", il);
-                } else {
-                    // For non-sliding layers, just reshape without applying RoPE
-                    Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
-                    cb(Qcur, "Qcur", il);
-
-                    Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
-                    cb(Kcur, "Kcur", il);
-                }
-
-                cur = llm_build_kv(ctx0, lctx, kv_self, gf, model.layers[il].wo, model.layers[il].bo, Kcur, Vcur, Qcur,
-                                   KQ_mask_l, n_tokens, kv_head, n_kv, 1.0f / sqrtf(float(n_embd_head)), cb, il);
-            }
-
-            if (il == n_layer - 1) {
-                // skip computing output for unused tokens
-                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
-                cur                              = ggml_get_rows(ctx0, cur, inp_out_ids);
-                inpL                             = ggml_get_rows(ctx0, inpL, inp_out_ids);
-                ffn_inp                          = ggml_get_rows(ctx0, ffn_inp, inp_out_ids);
-            }
-
-            struct ggml_tensor * attn_out = cur;
-
-            // feed-forward network
-            {
-                cur = llm_build_ffn(ctx0, lctx, ffn_inp, model.layers[il].ffn_up, NULL, NULL, model.layers[il].ffn_gate,
-                                    NULL, NULL, model.layers[il].ffn_down, NULL, NULL, NULL, LLM_FFN_SILU, LLM_FFN_PAR,
-                                    cb, il);
-                cb(cur, "ffn_out", il);
-            }
-
-            // add together residual + FFN + self-attention
-            cur = ggml_add(ctx0, cur, inpL);
-            cur = ggml_add(ctx0, cur, attn_out);
-            cur = lctx.cvec.apply_to(ctx0, cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-
-        cur = llm_build_norm(ctx0, cur, hparams, model.output_norm, NULL, LLM_NORM, cb, -1);
-        cb(cur, "result_norm", -1);
-
-        // lm_head
-        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
-
-        if (f_logit_scale) {
-            cur = ggml_scale(ctx0, cur, f_logit_scale);
-        }
-
-        cb(cur, "result_output", -1);
-
-        ggml_build_forward_expand(gf, cur);
-
-        return gf;
-    }
-
-    // ref: https://allenai.org/olmo
-    // based on the original build_llama() function, changes:
-    //   * non-parametric layer norm
-    //   * clamp qkv
-    //   * removed bias
-    //   * removed MoE
-    struct ggml_cgraph * build_olmo() {
-        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false);
-
-        // mutable variable, needed during the last layer of the computation to skip unused tokens
-        int32_t n_tokens = this->n_tokens;
-
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-        GGML_ASSERT(n_embd_head == hparams.n_rot);
-
-        struct ggml_tensor * cur;
-        struct ggml_tensor * inpL;
-
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
-
-        // inp_pos - contains the positions
-        struct ggml_tensor * inp_pos = build_inp_pos();
-
-        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
-        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
-
-        for (int il = 0; il < n_layer; ++il) {
-            struct ggml_tensor * inpSA = inpL;
-
-            // norm
-            cur = llm_build_norm(ctx0, inpL, hparams,
-                    NULL, NULL,
-                    LLM_NORM, cb, il);
-            cb(cur, "attn_norm", il);
-
-            // self-attention
-            {
-                // compute Q and K and RoPE them
-                struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur);
-                cb(Qcur, "Qcur", il);
-                if (hparams.f_clamp_kqv > 0.0f) {
-                    Qcur = ggml_clamp(ctx0, Qcur, -hparams.f_clamp_kqv, hparams.f_clamp_kqv);
-                    cb(Qcur, "Qcur", il);
-                }
-
-                struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur);
-                cb(Kcur, "Kcur", il);
-                if (hparams.f_clamp_kqv > 0.0f) {
-                    Kcur = ggml_clamp(ctx0, Kcur, -hparams.f_clamp_kqv, hparams.f_clamp_kqv);
-                    cb(Kcur, "Kcur", il);
-                }
-
-                struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur);
-                cb(Vcur, "Vcur", il);
-                if (hparams.f_clamp_kqv > 0.0f) {
-                    Vcur = ggml_clamp(ctx0, Vcur, -hparams.f_clamp_kqv, hparams.f_clamp_kqv);
-                    cb(Vcur, "Vcur", il);
-                }
-
-                Qcur = ggml_rope_ext(
-                    ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr,
-                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                    ext_factor, attn_factor, beta_fast, beta_slow
-                );
-                cb(Qcur, "Qcur", il);
-
-                Kcur = ggml_rope_ext(
-                    ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr,
-                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                    ext_factor, attn_factor, beta_fast, beta_slow
-                );
-                cb(Kcur, "Kcur", il);
-
-                cur = llm_build_kv(ctx0, lctx, kv_self, gf,
-                        model.layers[il].wo, nullptr,
-                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
-            }
-
-            if (il == n_layer - 1) {
-                // skip computing output for unused tokens
-                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
-                n_tokens = n_outputs;
-                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
-                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
-            }
-
-            struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
-            cb(ffn_inp, "ffn_inp", il);
-
-            // feed-forward network
-            cur = llm_build_norm(ctx0, ffn_inp, hparams,
-                    NULL, NULL,
-                    LLM_NORM, cb, il);
-            cb(cur, "ffn_norm", il);
-
-            cur = llm_build_ffn(ctx0, lctx, cur,
-                    model.layers[il].ffn_up,   NULL, NULL,
-                    model.layers[il].ffn_gate, NULL, NULL,
-                    model.layers[il].ffn_down, NULL, NULL,
-                    NULL,
-                    LLM_FFN_SILU, LLM_FFN_PAR, cb, il);
-            cb(cur, "ffn_out", il);
-
-            cur = ggml_add(ctx0, cur, ffn_inp);
-            cb(cur, "ffn_out", il);
-
-            cur = lctx.cvec.apply_to(ctx0, cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-
-        cur = llm_build_norm(ctx0, cur, hparams,
-                NULL, NULL,
-                LLM_NORM, cb, -1);
-        cb(cur, "result_norm", -1);
-
-        // lm_head
-        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
-        cb(cur, "result_output", -1);
-
-        ggml_build_forward_expand(gf, cur);
-
-        return gf;
-    }
-
-    struct ggml_cgraph * build_olmo2() {
-        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false);
-
-        // mutable variable, needed during the last layer of the computation to skip unused tokens
-        int32_t n_tokens = this->n_tokens;
-
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-        GGML_ASSERT(n_embd_head == hparams.n_rot);
-
-        struct ggml_tensor * cur;
-        struct ggml_tensor * inpL;
-
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
-
-        // inp_pos - contains the positions
-        struct ggml_tensor * inp_pos = build_inp_pos();
-
-        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
-        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
-
-        for (int il = 0; il < n_layer; ++il) {
-            struct ggml_tensor * inpSA = inpL;
-
-            cur = inpL;
-
-            // self_attention
-            {
-                // compute Q and K and RoPE them
-                struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur);
-                cb(Qcur, "Qcur", il);
-
-                struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur);
-                cb(Kcur, "Kcur", il);
-
-                struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur);
-                cb(Vcur, "Vcur", il);
-
-                Qcur = llm_build_norm(ctx0, Qcur, hparams, model.layers[il].attn_q_norm, NULL,
-                        LLM_NORM_RMS, cb, il);
-                cb(Qcur, "Qcur_normed", il);
-
-                Kcur = llm_build_norm(ctx0, Kcur, hparams, model.layers[il].attn_k_norm, NULL,
-                        LLM_NORM_RMS, cb, il);
-                cb(Kcur, "Kcur_normed", il);
-
-                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
-                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
-
-                Qcur = ggml_rope_ext(
-                    ctx0, Qcur, inp_pos, nullptr,
-                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                    ext_factor, attn_factor, beta_fast, beta_slow
-                );
-                cb(Qcur, "Qcur_rope", il);
-
-                Kcur = ggml_rope_ext(
-                    ctx0, Kcur, inp_pos, nullptr,
-                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                    ext_factor, attn_factor, beta_fast, beta_slow
-                );
-                cb(Kcur, "Kcur_rope", il);
-
-                cur = llm_build_kv(ctx0, lctx, kv_self, gf,
-                        model.layers[il].wo, NULL,
-                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
-            }
-
-            cur = llm_build_norm(ctx0, cur, hparams,
-                    model.layers[il].attn_post_norm, NULL,
-                    LLM_NORM_RMS, cb, il);
-            cb(cur, "attn_post_norm", il);
-
-            if (il == n_layer - 1) {
-                // skip computing output for unused tokens
-                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
-                n_tokens = n_outputs;
-                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
-                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
-            }
-
-            struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
-            cb(ffn_inp, "ffn_inp", il);
-
-            // feed-forward network
-            cur = llm_build_ffn(ctx0, lctx, ffn_inp,
-                    model.layers[il].ffn_up,   NULL, NULL,
-                    model.layers[il].ffn_gate, NULL, NULL,
-                    model.layers[il].ffn_down, NULL, NULL,
-                    NULL,
-                    LLM_FFN_SILU, LLM_FFN_PAR, cb, il);
-            cb(cur, "ffn_out", il);
-
-            cur = llm_build_norm(ctx0, cur, hparams,
-                model.layers[il].ffn_post_norm, NULL,
-                LLM_NORM_RMS, cb, -1);
-            cb(cur, "ffn_post_norm", -1);
-
-            cur = ggml_add(ctx0, cur, ffn_inp);
-            cb(cur, "ffn_out", il);
-
-            cur = lctx.cvec.apply_to(ctx0, cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-
-        cur = llm_build_norm(ctx0, cur, hparams,
-                model.output_norm, NULL,
-                LLM_NORM_RMS, cb, -1);
-        cb(cur, "result_norm", -1);
-
-        // lm_head
-        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
-        cb(cur, "result_output", -1);
-
-        ggml_build_forward_expand(gf, cur);
-
-        return gf;
-    }
-
-    // based on the build_qwen2moe() function, changes:
-    //   * removed shared experts
-    //   * removed bias
-    //   * added q, k norm
-    struct ggml_cgraph * build_olmoe() {
-        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false);
-
-        // mutable variable, needed during the last layer of the computation to skip unused tokens
-        int32_t n_tokens = this->n_tokens;
-
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-        GGML_ASSERT(n_embd_head == hparams.n_rot);
-
-        struct ggml_tensor * cur;
-        struct ggml_tensor * inpL;
-
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
-
-        // inp_pos - contains the positions
-        struct ggml_tensor * inp_pos = build_inp_pos();
-
-        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
-        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
-
-        for (int il = 0; il < n_layer; ++il) {
-            struct ggml_tensor * inpSA = inpL;
-
-            // norm
-            cur = llm_build_norm(ctx0, inpL, hparams,
-                    model.layers[il].attn_norm, NULL,
-                    LLM_NORM_RMS, cb, il);
-            cb(cur, "attn_norm", il);
-
-            // self_attention
-            {
-                // compute Q and K and RoPE them
-                struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur);
-                cb(Qcur, "Qcur", il);
-
-                struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur);
-                cb(Kcur, "Kcur", il);
-
-                struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur);
-                cb(Vcur, "Vcur", il);
-
-                Qcur = llm_build_norm(ctx0, Qcur, hparams, model.layers[il].attn_q_norm, NULL,
-                        LLM_NORM_RMS, cb, il);
-                cb(Qcur, "Qcur_normed", il);
-
-                Kcur = llm_build_norm(ctx0, Kcur, hparams, model.layers[il].attn_k_norm, NULL,
-                        LLM_NORM_RMS, cb, il);
-                cb(Kcur, "Kcur_normed", il);
-
-                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
-                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
-
-                Qcur = ggml_rope_ext(
-                    ctx0, Qcur, inp_pos, nullptr,
-                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                    ext_factor, attn_factor, beta_fast, beta_slow
-                );
-                cb(Qcur, "Qcur_rope", il);
-
-                Kcur = ggml_rope_ext(
-                    ctx0, Kcur, inp_pos, nullptr,
-                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                    ext_factor, attn_factor, beta_fast, beta_slow
-                );
-                cb(Kcur, "Kcur_rope", il);
-
-                cur = llm_build_kv(ctx0, lctx, kv_self, gf,
-                        model.layers[il].wo, NULL,
-                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
-            }
-
-            if (il == n_layer - 1) {
-                // skip computing output for unused tokens
-                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
-                n_tokens = n_outputs;
-                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
-                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
-            }
-
-            struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
-            cb(ffn_inp, "ffn_inp", il);
-
-            // MoE branch
-            cur = llm_build_norm(ctx0, ffn_inp, hparams,
-                    model.layers[il].ffn_norm, NULL,
-                    LLM_NORM_RMS, cb, il);
-            cb(cur, "ffn_norm", il);
-
-            cur = llm_build_moe_ffn(ctx0, lctx, cur,
-                    model.layers[il].ffn_gate_inp,
-                    model.layers[il].ffn_up_exps,
-                    model.layers[il].ffn_gate_exps,
-                    model.layers[il].ffn_down_exps,
-                    nullptr,
-                    n_expert, n_expert_used,
-                    LLM_FFN_SILU, false,
-                    false, 0.0,
-                    LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
-                    cb, il);
-            cb(cur, "ffn_moe_out", il);
-
-            cur = ggml_add(ctx0, cur, ffn_inp);
-            cur = lctx.cvec.apply_to(ctx0, cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-
-        cur = llm_build_norm(ctx0, cur, hparams,
-                model.output_norm, NULL,
-                LLM_NORM_RMS, cb, -1);
-        cb(cur, "result_norm", -1);
-
-        // lm_head
-        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
-        cb(cur, "result_output", -1);
-
-        ggml_build_forward_expand(gf, cur);
-
-        return gf;
-    }
-
-    struct ggml_cgraph * build_openelm() {
-        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false);
-
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-
-        struct ggml_tensor * cur;
-        struct ggml_tensor * inpL;
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
-
-        // inp_pos - contains the positions
-        struct ggml_tensor * inp_pos = build_inp_pos();
-
-        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
-        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
-
-        for (int il = 0; il < n_layer; ++il) {
-            const int64_t n_head    = hparams.n_head(il);
-            const int64_t n_head_kv = hparams.n_head_kv(il);
-            const int64_t n_head_qkv = 2*n_head_kv + n_head;
-
-            cur = inpL;
-            struct ggml_tensor * residual = cur;
-
-            // norm
-            cur = llm_build_norm(ctx0, inpL, hparams,
-                    model.layers[il].attn_norm, NULL,
-                    LLM_NORM_RMS, cb, il);
-            cb(cur, "attn_norm", il);
-
-            // self-attention
-            {
-                cur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wqkv, cur);
-                cb(cur, "wqkv", il);
-
-                cur = ggml_reshape_3d(ctx0, cur, n_embd_head_k, n_head_qkv, n_tokens);
-
-                struct ggml_tensor * Qcur = ggml_cont(ctx0, ggml_view_3d(ctx0, cur, n_embd_head, n_head, n_tokens, cur->nb[1], cur->nb[2], 0));
-                cb(Qcur, "Qcur", il);
-
-                struct ggml_tensor * Kcur = ggml_cont(ctx0, ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, cur->nb[1], cur->nb[2], cur->nb[1]*n_head));
-                cb(Kcur, "Kcur", il);
-
-                struct ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, cur->nb[1], cur->nb[2], cur->nb[1]*(n_head+n_head_kv)));
-                cb(Vcur, "Vcur", il);
-
-                Qcur = llm_build_norm(ctx0, Qcur, hparams,
-                        model.layers[il].attn_q_norm, NULL,
-                        LLM_NORM_RMS, cb, il);
-                cb(Qcur, "Qcur", il);
-
-                Kcur = llm_build_norm(ctx0, Kcur, hparams,
-                        model.layers[il].attn_k_norm, NULL,
-                        LLM_NORM_RMS, cb, il);
-                cb(Kcur, "Kcur", il);
-
-                Qcur = ggml_rope_ext(
-                    ctx0, Qcur, inp_pos, NULL, n_rot, rope_type, n_ctx_orig,
-                    freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow
-                );
-                cb(Qcur, "Qcur", il);
-
-                Kcur = ggml_rope_ext(
-                    ctx0, Kcur, inp_pos, NULL, n_rot, rope_type, n_ctx_orig,
-                    freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow
-                );
-                cb(Kcur, "Kcur", il);
-
-                Vcur = ggml_reshape_2d(ctx0, Vcur, n_embd_head * n_head_kv, n_tokens);
-                cb(Qcur, "Vcur", il);
-
-                cur = llm_build_kv(ctx0, lctx, kv_self, gf,
-                        model.layers[il].wo, NULL,
-                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
-            }
-
-            if (il == n_layer - 1) {
-                // skip computing output for unused tokens
-                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
-                residual = ggml_get_rows(ctx0, residual, inp_out_ids);
-                cur = ggml_get_rows(ctx0, cur, inp_out_ids);
-            }
-
-            struct ggml_tensor * ffn_inp = ggml_add(ctx0, residual, cur);
-            cb(ffn_inp, "ffn_inp", il);
-
-            // feed-forward network
-            {
-                cur = llm_build_norm(ctx0, ffn_inp, hparams,
-                        model.layers[il].ffn_norm, NULL,
-                        LLM_NORM_RMS, cb, il);
-                cb(cur, "ffn_norm", il);
-
-                cur = llm_build_ffn(ctx0, lctx, cur,
-                        model.layers[il].ffn_up,   NULL, NULL,
-                        model.layers[il].ffn_gate, NULL, NULL,
-                        model.layers[il].ffn_down, NULL, NULL,
-                        NULL,
-                        LLM_FFN_SILU, LLM_FFN_PAR, cb, il);
-                cb(cur, "ffn_out", il);
-            }
-
-            cur = ggml_add(ctx0, cur, ffn_inp);
-            cur = lctx.cvec.apply_to(ctx0, cur, il);
-            cb(cur, "l_out", il);
-
-            inpL = cur;
-        }
-
-        cur = inpL;
-
-        // norm
-        cur = llm_build_norm(ctx0, cur, hparams,
-                model.output_norm, NULL,
-                LLM_NORM_RMS, cb, -1);
-        cb(cur, "result_norm", -1);
-
-        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
-        cb(cur, "result_output", -1);
-
-        ggml_build_forward_expand(gf, cur);
-
-        return gf;
-    }
-
-    struct ggml_cgraph * build_gptneox() {
-        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false);
-
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-        const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-
-        struct ggml_tensor * cur;
-        struct ggml_tensor * inpL;
-
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
-
-        // inp_pos - contains the positions
-        struct ggml_tensor * inp_pos = build_inp_pos();
-
-        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
-        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
-
-        for (int il = 0; il < n_layer; ++il) {
-            cur = llm_build_norm(ctx0, inpL, hparams,
-                    model.layers[il].attn_norm,
-                    model.layers[il].attn_norm_b,
-                    LLM_NORM, cb, il);
-            cb(cur, "attn_norm", il);
-
-            // self-attention
-            {
-                cur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wqkv, cur);
-                cb(cur, "wqkv", il);
-
-                cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
-                cb(cur, "bqkv", il);
-
-                struct ggml_tensor * Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd,     n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd)));
-                struct ggml_tensor * Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd)));
-                struct ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa)));
-
-                cb(Qcur, "Qcur", il);
-                cb(Kcur, "Kcur", il);
-                cb(Vcur, "Vcur", il);
-
-                Qcur = ggml_rope_ext(
-                    ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr,
-                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                    ext_factor, attn_factor, beta_fast, beta_slow
-                );
-                cb(Qcur, "Qcur", il);
-
-                Kcur = ggml_rope_ext(
-                    ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr,
-                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                    ext_factor, attn_factor, beta_fast, beta_slow
-                );
-                cb(Kcur, "Kcur", il);
-
-                cur = llm_build_kv(ctx0, lctx, kv_self, gf,
-                        model.layers[il].wo, model.layers[il].bo,
-                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
-            }
-
-            if (il == n_layer - 1) {
-                // skip computing output for unused tokens
-                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
-                cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
-                inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
-            }
-
-            // ffn
-            if (hparams.use_par_res) {
-                // attention and ffn are computed in parallel
-                // x = x + attn(ln1(x)) + ffn(ln2(x))
-
-                struct ggml_tensor * attn_out = cur;
-
-                cur = llm_build_norm(ctx0, inpL, hparams,
-                        model.layers[il].ffn_norm,
-                        model.layers[il].ffn_norm_b,
-                        LLM_NORM, cb, il);
-                cb(cur, "ffn_norm", il);
-
-                cur = llm_build_ffn(ctx0, lctx, cur,
-                        model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
-                        NULL,                      NULL,                        NULL,
-                        model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
-                        NULL,
-                        LLM_FFN_GELU, LLM_FFN_SEQ, cb, il);
-                cb(cur, "ffn_out", il);
-
-                cur = ggml_add(ctx0, cur, inpL);
-                cb(cur, "ffn_out", il);
-
-                cur = ggml_add(ctx0, cur, attn_out);
-                cur = lctx.cvec.apply_to(ctx0, cur, il);
-                cb(cur, "l_out", il);
-
-                // input for next layer
-                inpL = cur;
-            } else {
-                // attention and ffn are computed sequentially
-                // x = x + attn(ln1(x))
-                // x = x + ffn(ln2(x))
-
-                struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL);
-                cb(ffn_inp, "ffn_inp", il);
-
-                cur = llm_build_norm(ctx0, ffn_inp, hparams,
-                        model.layers[il].ffn_norm,
-                        model.layers[il].ffn_norm_b,
-                        LLM_NORM, cb, il);
-                cb(cur, "ffn_norm", il);
-
-                cur = llm_build_ffn(ctx0, lctx, cur,
-                        model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
-                        NULL,                      NULL,                        NULL,
-                        model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
-                        NULL,
-                        LLM_FFN_GELU, LLM_FFN_SEQ, cb, il);
-                cb(cur, "ffn_out", il);
-
-                cur = ggml_add(ctx0, cur, ffn_inp);
-                cur = lctx.cvec.apply_to(ctx0, cur, il);
-                cb(cur, "l_out", il);
-
-                // input for next layer
-                inpL = cur;
-            }
-        }
-
-        cur = llm_build_norm(ctx0, inpL, hparams,
-                model.output_norm,
-                model.output_norm_b,
-                LLM_NORM, cb, -1);
-        cb(cur, "result_norm", -1);
-
-        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
-        cb(cur, "result_output", -1);
-
-        ggml_build_forward_expand(gf, cur);
-
-        return gf;
-    }
-
-    struct ggml_cgraph * build_arctic() {
-        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false);
-
-        // mutable variable, needed during the last layer of the computation to skip unused tokens
-        int32_t n_tokens = this->n_tokens;
-
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-        GGML_ASSERT(n_embd_head == hparams.n_rot);
-
-        struct ggml_tensor * cur;
-        struct ggml_tensor * inpL;
-
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
-
-        // inp_pos - contains the positions
-        struct ggml_tensor * inp_pos = build_inp_pos();
-
-        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
-        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
-
-        for (int il = 0; il < n_layer; ++il) {
-            struct ggml_tensor * inpSA = inpL;
-
-            // norm
-            cur = llm_build_norm(ctx0, inpL, hparams,
-                    model.layers[il].attn_norm, NULL,
-                    LLM_NORM_RMS, cb, il);
-            cb(cur, "attn_norm", il);
-
-            // self-attention
-            {
-                // compute Q and K and RoPE them
-                struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur);
-                cb(Qcur, "Qcur", il);
-
-                struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur);
-                cb(Kcur, "Kcur", il);
-
-                struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur);
-                cb(Vcur, "Vcur", il);
-
-                Qcur = ggml_rope_ext(
-                    ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr,
-                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                    ext_factor, attn_factor, beta_fast, beta_slow
-                );
-                cb(Qcur, "Qcur", il);
-
-                Kcur = ggml_rope_ext(
-                    ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr,
-                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                    ext_factor, attn_factor, beta_fast, beta_slow
-                );
-                cb(Kcur, "Kcur", il);
-
-                cur = llm_build_kv(ctx0, lctx, kv_self, gf,
-                        model.layers[il].wo, NULL,
-                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
-            }
-
-            if (il == n_layer - 1) {
-                // skip computing output for unused tokens
-                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
-                n_tokens = n_outputs;
-                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
-                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
-            }
-
-            struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
-            cb(ffn_inp, "ffn_inp", il);
-
-            // feed-forward network
-            cur = llm_build_norm(ctx0, ffn_inp, hparams,
-                    model.layers[il].ffn_norm, NULL,
-                    LLM_NORM_RMS, cb, il);
-            cb(cur, "ffn_norm", il);
-
-            cur = llm_build_ffn(ctx0, lctx, cur,
-                    model.layers[il].ffn_up,   NULL, NULL,
-                    model.layers[il].ffn_gate, NULL, NULL,
-                    model.layers[il].ffn_down, NULL, NULL,
-                    NULL,
-                    LLM_FFN_SILU, LLM_FFN_PAR, cb, il);
-            cb(cur, "ffn_out", il);
-
-            struct ggml_tensor * ffn_out = ggml_add(ctx0, cur, ffn_inp);
-            cb(ffn_out, "ffn_out", il);
-
-            // MoE
-            cur = llm_build_norm(ctx0, inpSA, hparams,
-                    model.layers[il].ffn_norm_exps, NULL,
-                    LLM_NORM_RMS, cb, il);
-            cb(cur, "ffn_norm_exps", il);
-
-            cur = llm_build_moe_ffn(ctx0, lctx, cur,
-                    model.layers[il].ffn_gate_inp,
-                    model.layers[il].ffn_up_exps,
-                    model.layers[il].ffn_gate_exps,
-                    model.layers[il].ffn_down_exps,
-                    nullptr,
-                    n_expert, n_expert_used,
-                    LLM_FFN_SILU, true,
-                    false, 0.0,
-                    LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
-                    cb, il);
-            cb(cur, "ffn_moe_out", il);
-
-            cur = ggml_add(ctx0, cur, ffn_out);
-            cb(cur, "ffn_out", il);
-
-            cur = lctx.cvec.apply_to(ctx0, cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-
-        cur = llm_build_norm(ctx0, cur, hparams,
-                model.output_norm, NULL,
-                LLM_NORM_RMS, cb, -1);
-        cb(cur, "result_norm", -1);
-
-        // lm_head
-        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
-        cb(cur, "result_output", -1);
-
-        ggml_build_forward_expand(gf, cur);
-
-        return gf;
-    }
-
-    struct ggml_cgraph * build_deepseek() {
-        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false);
-
-        // mutable variable, needed during the last layer of the computation to skip unused tokens
-        int32_t n_tokens = this->n_tokens;
-
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-        GGML_ASSERT(n_embd_head == hparams.n_rot);
-
-        struct ggml_tensor * cur;
-        struct ggml_tensor * inpL;
-
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
-
-        // inp_pos - contains the positions
-        struct ggml_tensor * inp_pos = build_inp_pos();
-
-        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
-        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
-        const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
-        for (int il = 0; il < n_layer; ++il) {
-            struct ggml_tensor * inpSA = inpL;
-
-            // norm
-            cur = llm_build_norm(ctx0, inpL, hparams,
-                    model.layers[il].attn_norm, NULL,
-                    LLM_NORM_RMS, cb, il);
-            cb(cur, "attn_norm", il);
-
-            // self-attention
-            {
-                // rope freq factors for llama3; may return nullptr for llama2 and other models
-                struct ggml_tensor * rope_factors = build_rope_factors(il);
-
-                // compute Q and K and RoPE them
-                struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur);
-                cb(Qcur, "Qcur", il);
-                if (model.layers[il].bq) {
-                    Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
-                    cb(Qcur, "Qcur", il);
-                }
-
-                struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur);
-                cb(Kcur, "Kcur", il);
-                if (model.layers[il].bk) {
-                    Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
-                    cb(Kcur, "Kcur", il);
-                }
-
-                struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur);
-                cb(Vcur, "Vcur", il);
-                if (model.layers[il].bv) {
-                    Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
-                    cb(Vcur, "Vcur", il);
-                }
-
-                Qcur = ggml_rope_ext(
-                    ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, rope_factors,
-                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                    ext_factor, attn_factor, beta_fast, beta_slow
-                );
-                cb(Qcur, "Qcur", il);
-
-                Kcur = ggml_rope_ext(
-                    ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, rope_factors,
-                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                    ext_factor, attn_factor, beta_fast, beta_slow
-                );
-                cb(Kcur, "Kcur", il);
-
-                cur = llm_build_kv(ctx0, lctx, kv_self, gf,
-                        model.layers[il].wo, model.layers[il].bo,
-                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, kq_scale, cb, il);
-            }
-
-            if (il == n_layer - 1) {
-                // skip computing output for unused tokens
-                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
-                n_tokens = n_outputs;
-                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
-                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
-            }
-
-
-            struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
-            cb(ffn_inp, "ffn_inp", il);
-
-            cur = llm_build_norm(ctx0, ffn_inp, hparams,
-                    model.layers[il].ffn_norm, NULL,
-                    LLM_NORM_RMS, cb, il);
-            cb(cur, "ffn_norm", il);
-
-            if ((uint32_t) il < hparams.n_layer_dense_lead) {
-                cur = llm_build_ffn(ctx0, lctx, cur,
-                        model.layers[il].ffn_up,   NULL, NULL,
-                        model.layers[il].ffn_gate, NULL, NULL,
-                        model.layers[il].ffn_down, NULL, NULL,
-                        NULL,
-                        LLM_FFN_SILU, LLM_FFN_PAR, cb, il);
-                cb(cur, "ffn_out", il);
-            } else {
-                // MoE branch
-                ggml_tensor * moe_out =
-                        llm_build_moe_ffn(ctx0, lctx, cur,
-                            model.layers[il].ffn_gate_inp,
-                            model.layers[il].ffn_up_exps,
-                            model.layers[il].ffn_gate_exps,
-                            model.layers[il].ffn_down_exps,
-                            nullptr,
-                            n_expert, n_expert_used,
-                            LLM_FFN_SILU, false,
-                            false, hparams.expert_weights_scale,
-                            LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
-                            cb, il);
-                cb(moe_out, "ffn_moe_out", il);
-
-                // FFN shared expert
-                {
-                    ggml_tensor * ffn_shexp = llm_build_ffn(ctx0, lctx, cur,
-                            model.layers[il].ffn_up_shexp,   NULL, NULL,
-                            model.layers[il].ffn_gate_shexp, NULL, NULL,
-                            model.layers[il].ffn_down_shexp, NULL, NULL,
-                            NULL,
-                            LLM_FFN_SILU, LLM_FFN_PAR, cb, il);
-                    cb(ffn_shexp, "ffn_shexp", il);
-
-                    cur = ggml_add(ctx0, moe_out, ffn_shexp);
-                    cb(cur, "ffn_out", il);
-                }
-            }
-
-            cur = ggml_add(ctx0, cur, ffn_inp);
-            cur = lctx.cvec.apply_to(ctx0, cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-
-        cur = llm_build_norm(ctx0, cur, hparams,
-                model.output_norm, NULL,
-                LLM_NORM_RMS, cb, -1);
-        cb(cur, "result_norm", -1);
-
-        // lm_head
-        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
-
-        cb(cur, "result_output", -1);
-
-        ggml_build_forward_expand(gf, cur);
-
-        return gf;
-    }
-
-    struct ggml_cgraph * build_deepseek2() {
-        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false);
-
-        // mutable variable, needed during the last layer of the computation to skip unused tokens
-        int32_t n_tokens = this->n_tokens;
-
-        bool is_lite = (hparams.n_layer == 27);
-
-        // We have to pre-scale kq_scale and attn_factor to make the YaRN RoPE work correctly.
-        // See https://github.com/ggerganov/llama.cpp/discussions/7416 for detailed explanation.
-        const float mscale = attn_factor * (1.0f + hparams.rope_yarn_log_mul * logf(1.0f / freq_scale));
-        const float kq_scale = 1.0f*mscale*mscale/sqrtf(float(hparams.n_embd_head_k));
-        const float attn_factor_scaled = 1.0f / (1.0f + 0.1f * logf(1.0f / freq_scale));
-
-        const uint32_t n_embd_head_qk_rope = hparams.n_rot;
-        const uint32_t n_embd_head_qk_nope = hparams.n_embd_head_k - hparams.n_rot;
-        const uint32_t kv_lora_rank = hparams.n_lora_kv;
-
-        struct ggml_tensor * cur;
-        struct ggml_tensor * inpL;
-
-        // {n_embd, n_tokens}
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
-
-        // inp_pos - contains the positions
-        struct ggml_tensor * inp_pos = build_inp_pos();
-
-        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
-        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
-
-        for (int il = 0; il < n_layer; ++il) {
-            struct ggml_tensor * inpSA = inpL;
-
-            // norm
-            cur = llm_build_norm(ctx0, inpL, hparams,
-                    model.layers[il].attn_norm, NULL,
-                    LLM_NORM_RMS, cb, il);
-            cb(cur, "attn_norm", il);
-
-            // self_attention
-            {
-                struct ggml_tensor * q = NULL;
-                if (!is_lite) {
-                    // {n_embd, q_lora_rank} * {n_embd, n_tokens} -> {q_lora_rank, n_tokens}
-                    q = ggml_mul_mat(ctx0, model.layers[il].wq_a, cur);
-                    cb(q, "q", il);
-
-                    q = llm_build_norm(ctx0, q, hparams,
-                            model.layers[il].attn_q_a_norm, NULL,
-                            LLM_NORM_RMS, cb, il);
-                    cb(q, "q", il);
-
-                    // {q_lora_rank, n_head * hparams.n_embd_head_k} * {q_lora_rank, n_tokens} -> {n_head * hparams.n_embd_head_k, n_tokens}
-                    q = ggml_mul_mat(ctx0, model.layers[il].wq_b, q);
-                    cb(q, "q", il);
-                } else {
-                    q = ggml_mul_mat(ctx0, model.layers[il].wq, cur);
-                    cb(q, "q", il);
-                }
-
-                // split into {n_head * n_embd_head_qk_nope, n_tokens}
-                struct ggml_tensor * q_nope = ggml_view_3d(ctx0, q, n_embd_head_qk_nope, n_head, n_tokens,
-                        ggml_row_size(q->type, hparams.n_embd_head_k),
-                        ggml_row_size(q->type, hparams.n_embd_head_k * n_head),
-                        0);
-                cb(q_nope, "q_nope", il);
-
-                // and {n_head * n_embd_head_qk_rope, n_tokens}
-                struct ggml_tensor * q_pe = ggml_view_3d(ctx0, q, n_embd_head_qk_rope, n_head, n_tokens,
-                        ggml_row_size(q->type, hparams.n_embd_head_k),
-                        ggml_row_size(q->type, hparams.n_embd_head_k * n_head),
-                        ggml_row_size(q->type, n_embd_head_qk_nope));
-                cb(q_pe, "q_pe", il);
-
-                // {n_embd, kv_lora_rank + n_embd_head_qk_rope} * {n_embd, n_tokens} -> {kv_lora_rank + n_embd_head_qk_rope, n_tokens}
-                struct ggml_tensor * kv_pe_compresseed = ggml_mul_mat(ctx0, model.layers[il].wkv_a_mqa, cur);
-                cb(kv_pe_compresseed, "kv_pe_compresseed", il);
-
-                // split into {kv_lora_rank, n_tokens}
-                struct ggml_tensor * kv_compressed = ggml_view_2d(ctx0, kv_pe_compresseed, kv_lora_rank, n_tokens,
-                        kv_pe_compresseed->nb[1],
-                        0);
-                cb(kv_compressed, "kv_compressed", il);
-
-                // and {n_embd_head_qk_rope, n_tokens}
-                struct ggml_tensor * k_pe = ggml_view_3d(ctx0, kv_pe_compresseed, n_embd_head_qk_rope, 1, n_tokens,
-                        kv_pe_compresseed->nb[1],
-                        kv_pe_compresseed->nb[1],
-                        ggml_row_size(kv_pe_compresseed->type, kv_lora_rank));
-                cb(k_pe, "k_pe", il);
-
-                // TODO: the CUDA backend used to not support non-cont. (RMS) norm, investigate removing ggml_cont
-                kv_compressed = ggml_cont(ctx0, kv_compressed);
-                kv_compressed = llm_build_norm(ctx0, kv_compressed, hparams,
-                        model.layers[il].attn_kv_a_norm, NULL,
-                        LLM_NORM_RMS, cb, il);
-                cb(kv_compressed, "kv_compressed", il);
-
-                // {kv_lora_rank, n_head * (n_embd_head_qk_nope + n_embd_head_v)} * {kv_lora_rank, n_tokens} -> {n_head * (n_embd_head_qk_nope + n_embd_head_v), n_tokens}
-                struct ggml_tensor * kv = ggml_mul_mat(ctx0, model.layers[il].wkv_b, kv_compressed);
-                cb(kv, "kv", il);
-
-                // split into {n_head * n_embd_head_qk_nope, n_tokens}
-                struct ggml_tensor * k_nope = ggml_view_3d(ctx0, kv, n_embd_head_qk_nope, n_head, n_tokens,
-                        ggml_row_size(kv->type, n_embd_head_qk_nope + hparams.n_embd_head_v),
-                        ggml_row_size(kv->type, n_head * (n_embd_head_qk_nope + hparams.n_embd_head_v)),
-                        0);
-                cb(k_nope, "k_nope", il);
-
-                // and {n_head * n_embd_head_v, n_tokens}
-                struct ggml_tensor * v_states = ggml_view_3d(ctx0, kv, hparams.n_embd_head_v, n_head, n_tokens,
-                        ggml_row_size(kv->type, (n_embd_head_qk_nope + hparams.n_embd_head_v)),
-                        ggml_row_size(kv->type, (n_embd_head_qk_nope + hparams.n_embd_head_v)*n_head),
-                        ggml_row_size(kv->type, (n_embd_head_qk_nope)));
-                cb(v_states, "v_states", il);
-
-                v_states = ggml_cont(ctx0, v_states);
-                cb(v_states, "v_states", il);
-
-                v_states = ggml_view_2d(ctx0, v_states, hparams.n_embd_head_v * n_head, n_tokens,
-                    ggml_row_size(kv->type, hparams.n_embd_head_v * n_head),
-                    0);
-                cb(v_states, "v_states", il);
-
-                q_pe = ggml_cont(ctx0, q_pe); // TODO: the CUDA backend used to not support non-cont. RoPE, investigate removing this
-                q_pe = ggml_rope_ext(
-                    ctx0, q_pe, inp_pos, nullptr,
-                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                    ext_factor, attn_factor_scaled, beta_fast, beta_slow
-                );
-                cb(q_pe, "q_pe", il);
-
-                // shared RoPE key
-                k_pe = ggml_cont(ctx0, k_pe); // TODO: the CUDA backend used to not support non-cont. RoPE, investigate removing this
-                k_pe = ggml_rope_ext(
-                    ctx0, k_pe, inp_pos, nullptr,
-                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                    ext_factor, attn_factor_scaled, beta_fast, beta_slow
-                );
-                cb(k_pe, "k_pe", il);
-
-                struct ggml_tensor * q_states = ggml_concat(ctx0, q_nope, q_pe, 0);
-                cb(q_states, "q_states", il);
-
-                struct ggml_tensor * k_states = ggml_concat(ctx0, k_nope, ggml_repeat(ctx0, k_pe, q_pe), 0);
-                cb(k_states, "k_states", il);
-
-                cur = llm_build_kv(ctx0, lctx, kv_self, gf,
-                        model.layers[il].wo, NULL,
-                        k_states, v_states, q_states, KQ_mask, n_tokens, kv_head, n_kv, kq_scale, cb, il);
-            }
-
-            if (il == n_layer - 1) {
-                // skip computing output for unused tokens
-                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
-                n_tokens = n_outputs;
-                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
-                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
-            }
-
-            struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
-            cb(ffn_inp, "ffn_inp", il);
-
-            cur = llm_build_norm(ctx0, ffn_inp, hparams,
-                    model.layers[il].ffn_norm, NULL,
-                    LLM_NORM_RMS, cb, il);
-            cb(cur, "ffn_norm", il);
-
-            if ((uint32_t) il < hparams.n_layer_dense_lead) {
-                cur = llm_build_ffn(ctx0, lctx, cur,
-                        model.layers[il].ffn_up,   NULL, NULL,
-                        model.layers[il].ffn_gate, NULL, NULL,
-                        model.layers[il].ffn_down, NULL, NULL,
-                        NULL,
-                        LLM_FFN_SILU, LLM_FFN_PAR, cb, il);
-                cb(cur, "ffn_out", il);
-            } else {
-                // MoE branch
-                ggml_tensor * moe_out =
-                        llm_build_moe_ffn(ctx0, lctx, cur,
-                            model.layers[il].ffn_gate_inp,
-                            model.layers[il].ffn_up_exps,
-                            model.layers[il].ffn_gate_exps,
-                            model.layers[il].ffn_down_exps,
-                            model.layers[il].ffn_exp_probs_b,
-                            n_expert, n_expert_used,
-                            LLM_FFN_SILU, hparams.expert_weights_norm,
-                            true, hparams.expert_weights_scale,
-                            (enum llama_expert_gating_func_type) hparams.expert_gating_func,
-                            cb, il);
-                cb(moe_out, "ffn_moe_out", il);
-
-                // FFN shared expert
-                {
-                    ggml_tensor * ffn_shexp = llm_build_ffn(ctx0, lctx, cur,
-                            model.layers[il].ffn_up_shexp,   NULL, NULL,
-                            model.layers[il].ffn_gate_shexp, NULL, NULL,
-                            model.layers[il].ffn_down_shexp, NULL, NULL,
-                            NULL,
-                            LLM_FFN_SILU, LLM_FFN_PAR, cb, il);
-                    cb(ffn_shexp, "ffn_shexp", il);
-
-                    cur = ggml_add(ctx0, moe_out, ffn_shexp);
-                    cb(cur, "ffn_out", il);
-                }
-            }
-
-            cur = ggml_add(ctx0, cur, ffn_inp);
-            cur = lctx.cvec.apply_to(ctx0, cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-
-        cur = llm_build_norm(ctx0, cur, hparams,
-                model.output_norm, NULL,
-                LLM_NORM_RMS, cb, -1);
-        cb(cur, "result_norm", -1);
-
-        // lm_head
-        cur = ggml_mul_mat(ctx0, model.output, cur);
-        cb(cur, "result_output", -1);
-
-        ggml_build_forward_expand(gf, cur);
-
-        return gf;
-    }
-
-    struct ggml_cgraph * build_bitnet() {
-        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false);
-
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-
-        struct ggml_tensor * cur;
-        struct ggml_tensor * inpL;
-
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
-
-        // inp_pos - contains the positions
-        struct ggml_tensor * inp_pos = build_inp_pos();
-
-        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
-        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
-
-        for (int il = 0; il < n_layer; ++il) {
-            struct ggml_tensor * inpSA = inpL;
-
-            cur = llm_build_norm(ctx0, inpL, hparams,
-                    model.layers[il].attn_norm, NULL,
-                    LLM_NORM_RMS, cb, il);
-            cb(cur, "attn_norm", il);
-
-            // self-attention
-            {
-                // compute Q and K and RoPE them
-                struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur);
-                if (model.layers[il].wq_scale) {
-                    Qcur = ggml_mul(ctx0, Qcur, model.layers[il].wq_scale);
-                }
-                cb(Qcur, "Qcur", il);
-                if (model.layers[il].bq) {
-                    Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
-                    cb(Qcur, "Qcur", il);
-                }
-
-                // B1.K
-                struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur);
-                if (model.layers[il].wk_scale) {
-                    Kcur = ggml_mul(ctx0, Kcur, model.layers[il].wk_scale);
-                }
-                cb(Kcur, "Kcur", il);
-                if (model.layers[il].bk) {
-                    Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
-                    cb(Kcur, "Kcur", il);
-                }
-
-                // B1.V
-                struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur);
-                if (model.layers[il].wv_scale) {
-                    Vcur = ggml_mul(ctx0, Vcur, model.layers[il].wv_scale);
-                }
-                cb(Vcur, "Vcur", il);
-                if (model.layers[il].bv) {
-                    Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
-                    cb(Vcur, "Vcur", il);
-                }
-
-                Qcur = ggml_rope_ext(
-                    ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr,
-                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                    ext_factor, attn_factor, beta_fast, beta_slow
-                );
-                cb(Qcur, "Qcur", il);
-
-                Kcur = ggml_rope_ext(
-                    ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr,
-                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                    ext_factor, attn_factor, beta_fast, beta_slow
-                );
-                cb(Kcur, "Kcur", il);
-
-                cur = llm_build_kv(ctx0, lctx, kv_self, gf,
-                        NULL, NULL,
-                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
-
-                cur = llm_build_norm(ctx0, cur, hparams,
-                        model.layers[il].attn_sub_norm, NULL,
-                        LLM_NORM_RMS, cb, il);
-                cb(cur, "attn_sub_norm", il);
-
-                cur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wo, cur);
-                if (model.layers[il].wo_scale) {
-                    cur = ggml_mul(ctx0, cur, model.layers[il].wo_scale);
-                }
-                if (model.layers[il].bo) {
-                    cur = ggml_add(ctx0, cur, model.layers[il].bo);
-                }
-                cb(cur, "attn_o_out", il);
-            }
-
-            if (il == n_layer - 1) {
-                // skip computing output for unused tokens
-                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
-                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
-                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
-            }
-
-            struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
-            cb(ffn_inp, "ffn_inp", il);
-
-            // feed-forward forward
-            cur = llm_build_norm(ctx0, ffn_inp, hparams,
-                    model.layers[il].ffn_norm, NULL,
-                    LLM_NORM_RMS, cb, il);
-            cb(cur, "ffn_norm", il);
-
-            cur = llm_build_ffn(ctx0, lctx, cur,
-                    model.layers[il].ffn_up,   NULL, model.layers[il].ffn_up_scale,
-                    model.layers[il].ffn_gate, NULL, model.layers[il].ffn_gate_scale,
-                    NULL,                      NULL, NULL,
-                    NULL,
-                    LLM_FFN_SILU, LLM_FFN_PAR, cb, il);
-            cb(cur, "ffn_sub_out", il);
-
-            cur = llm_build_norm(ctx0, cur, hparams,
-                            model.layers[il].ffn_sub_norm, NULL,
-                            LLM_NORM_RMS, cb, il);
-            cb(cur, "ffn_sub_norm", il);
-
-            cur = llm_build_lora_mm(lctx, ctx0, model.layers[il].ffn_down, cur);
-            if (model.layers[il].ffn_down_scale) {
-                cur = ggml_mul(ctx0, cur, model.layers[il].ffn_down_scale);
-            }
-            cb(cur, "ffn_down", il);
-
-            cur = ggml_add(ctx0, cur, ffn_inp);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-
-        cur = llm_build_norm(ctx0, cur, hparams,
-                model.output_norm, NULL,
-                LLM_NORM_RMS, cb, -1);
-        cb(cur, "result_norm", -1);
-
-        // lm_head
-        // FIXME: do not use model.tok_embd directly, duplicate as model.output
-        cur = llm_build_lora_mm(lctx, ctx0, model.tok_embd, cur);
-        cb(cur, "result_output", -1);
-
-        ggml_build_forward_expand(gf, cur);
-        return gf;
-    }
-
-    struct ggml_cgraph * build_t5_enc() {
-        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false);
-
-        // mutable variable, needed during the last layer of the computation to skip unused tokens
-        int32_t n_tokens = this->n_tokens;
-
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-        const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-
-        struct ggml_tensor * cur;
-        struct ggml_tensor * inpL;
-
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
-
-        GGML_ASSERT(lctx.is_encoding);
-        struct ggml_tensor * pos_bucket_enc = llm_build_pos_bucket(false);
-
-        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
-        struct ggml_tensor * KQ_mask_enc = build_inp_KQ_mask(false);
-
-        for (int il = 0; il < n_layer; ++il) {
-            struct ggml_tensor * inpSA = inpL;
-
-            // norm
-            cur = llm_build_norm(ctx0, inpL, hparams,
-                    model.layers[il].attn_norm_enc, NULL,
-                    LLM_NORM_RMS, cb, il);
-            cb(cur, "attn_norm", il);
-
-            // self-attention
-            {
-                struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq_enc, cur);
-                cb(Qcur, "Qcur", il);
-
-                struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk_enc, cur);
-                cb(Kcur, "Kcur", il);
-
-                struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv_enc, cur);
-                cb(Vcur, "Vcur", il);
-
-                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
-                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
-
-                struct ggml_tensor * q =                 ggml_permute(ctx0, Qcur, 0, 2, 1, 3);
-                struct ggml_tensor * k = ggml_cont(ctx0, ggml_permute(ctx0, Kcur, 0, 2, 1, 3));
-
-                struct ggml_tensor * kq = ggml_mul_mat(ctx0, k, q);
-                cb(kq, "kq", il);
-
-                struct ggml_tensor * attn_rel_b = model.layers[il].attn_rel_b_enc ? model.layers[il].attn_rel_b_enc : model.layers[0].attn_rel_b_enc;
-                struct ggml_tensor * pos_bias = llm_build_pos_bias(pos_bucket_enc, attn_rel_b);
-                struct ggml_tensor * kq_b = ggml_add(ctx0, kq, pos_bias);
-                cb(kq_b, "kq_b", il);
-
-                kq = ggml_soft_max_ext(ctx0, kq_b, KQ_mask_enc, 1.0f, hparams.f_max_alibi_bias);
-                cb(kq, "kq_soft_max_ext", il);
-
-                struct ggml_tensor * v = ggml_cont(ctx0, ggml_transpose(ctx0, ggml_reshape_2d(ctx0, Vcur, n_embd_gqa, n_tokens)));
-                cb(v, "v", il);
-
-                struct ggml_tensor * kqv = ggml_mul_mat(ctx0, ggml_reshape_3d(ctx0, v, n_tokens, n_embd_head, n_head_kv), kq);
-                cb(kqv, "kqv", il);
-
-                struct ggml_tensor * kqv_merged = ggml_permute(ctx0, kqv, 0, 2, 1, 3);
-                cb(kqv_merged, "kqv_merged", il);
-
-                cur = ggml_cont_2d(ctx0, kqv_merged, n_embd_gqa, n_tokens);
-                cb(cur, "kqv_merged_cont", il);
-
-                ggml_build_forward_expand(gf, cur);
-
-                cur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wo_enc, cur);
-                cb(cur, "kqv_out", il);
-            }
-
-            if (il == n_layer - 1) {
-                // skip computing output for unused tokens
-                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
-                n_tokens = n_outputs;
-                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
-                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
-            }
-
-            struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
-            cb(ffn_inp, "ffn_inp", il);
-
-            // feed-forward network
-            {
-                cur = llm_build_norm(ctx0, ffn_inp, hparams,
-                        model.layers[il].ffn_norm_enc, NULL,
-                        LLM_NORM_RMS, cb, il);
-                cb(cur, "ffn_norm", il);
-
-                // T5 uses relu, flan-T5 uses gelu-gated
-                cur = llm_build_ffn(ctx0, lctx, cur,
-                        model.layers[il].ffn_up_enc,   NULL, NULL,
-                        model.layers[il].ffn_gate_enc, NULL, NULL,
-                        model.layers[il].ffn_down_enc, NULL, NULL,
-                        NULL,
-                        model.layers[il].ffn_gate_enc ? LLM_FFN_GELU : LLM_FFN_RELU,
-                        model.layers[il].ffn_gate_enc ? LLM_FFN_PAR  : LLM_FFN_SEQ,
-                        cb, il);
-                cb(cur, "ffn_out", il);
-            }
-
-            cur = ggml_add(ctx0, cur, ffn_inp);
-            cb(cur, "ffn_out", il);
-
-            ggml_tensor * layer_dir = lctx.cvec.tensor_for(il);
-            if (layer_dir != nullptr) {
-                cur = ggml_add(ctx0, cur, layer_dir);
-            }
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-        cb(cur, "result_embd", -1);
-
-        cur = llm_build_norm(ctx0, cur, hparams,
-                model.output_norm_enc, NULL,
-                LLM_NORM_RMS, cb, -1);
-        cb(cur, "result_norm", -1);
-
-        ggml_build_forward_expand(gf, cur);
-
-        return gf;
-    }
-
-    struct ggml_cgraph * build_t5_dec() {
-        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false);
-
-        // mutable variable, needed during the last layer of the computation to skip unused tokens
-        int32_t n_tokens = this->n_tokens;
-
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-        const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-
-        struct ggml_tensor * cur;
-        struct ggml_tensor * inpL;
-
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
-
-        GGML_ASSERT(!lctx.is_encoding);
-        GGML_ASSERT(n_outputs_enc > 0 && "call llama_encode() first");
-
-        struct ggml_tensor * embd_enc       = llm_build_inp_embd_enc();
-        struct ggml_tensor * pos_bucket_dec = llm_build_pos_bucket(true);
-
-        struct ggml_tensor * KQ_mask_dec   = build_inp_KQ_mask();
-        struct ggml_tensor * KQ_mask_cross = llm_build_inp_KQ_mask_cross();
-
-        for (int il = 0; il < n_layer; ++il) {
-            struct ggml_tensor * inpSA = inpL;
-
-            // norm
-            cur = llm_build_norm(ctx0, inpL, hparams,
-                    model.layers[il].attn_norm, NULL,
-                    LLM_NORM_RMS, cb, il);
-            cb(cur, "attn_norm", il);
-
-            // self-attention
-            {
-                struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur);
-                cb(Qcur, "Qcur", il);
-
-                struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur);
-                cb(Kcur, "Kcur", il);
-
-                struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur);
-                cb(Vcur, "Vcur", il);
-
-                llm_build_kv_store(ctx0, hparams, cparams, kv_self, gf, Kcur, Vcur, n_tokens, kv_head, cb, il);
-
-                struct ggml_tensor * k =
-                    ggml_view_3d(ctx0, kv_self.k_l[il],
-                            n_embd_head_k, n_kv, n_head_kv,
-                            ggml_row_size(kv_self.k_l[il]->type, n_embd_k_gqa),
-                            ggml_row_size(kv_self.k_l[il]->type, n_embd_head_k),
-                            0);
-                cb(k, "k", il);
-
-                struct ggml_tensor * v =
-                    ggml_view_3d(ctx0, kv_self.v_l[il],
-                            n_kv, n_embd_head_v, n_head_kv,
-                            ggml_element_size(kv_self.v_l[il])*n_ctx,
-                            ggml_element_size(kv_self.v_l[il])*n_ctx*n_embd_head_v,
-                            0);
-                cb(v, "v", il);
-
-                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
-
-                struct ggml_tensor * q = ggml_permute(ctx0, Qcur, 0, 2, 1, 3);
-
-                struct ggml_tensor * kq = ggml_mul_mat(ctx0, k, q);
-                cb(kq, "kq", il);
-
-                struct ggml_tensor * attn_rel_b = model.layers[il].attn_rel_b ? model.layers[il].attn_rel_b : model.layers[0].attn_rel_b;
-                struct ggml_tensor * pos_bias = llm_build_pos_bias(pos_bucket_dec, attn_rel_b);
-                struct ggml_tensor * kq_b = ggml_add(ctx0, kq, pos_bias);
-                cb(kq_b, "kq_b", il);
-
-                kq = ggml_soft_max_ext(ctx0, kq_b, KQ_mask_dec, 1.0f, hparams.f_max_alibi_bias);
-                cb(kq, "kq_soft_max_ext", il);
-
-                struct ggml_tensor * kqv = ggml_mul_mat(ctx0, v, kq);
-                cb(kqv, "kqv", il);
-
-                struct ggml_tensor * kqv_merged = ggml_permute(ctx0, kqv, 0, 2, 1, 3);
-                cb(kqv_merged, "kqv_merged", il);
-
-                cur = ggml_cont_2d(ctx0, kqv_merged, n_embd_gqa, n_tokens);
-                cb(cur, "kqv_merged_cont", il);
-
-                ggml_build_forward_expand(gf, cur);
-
-                cur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wo, cur);
-                cb(cur, "kqv_out", il);
-            }
-
-            cur = ggml_add(ctx0, cur, inpSA);
-            cb(cur, "cross_inp", il);
-
-            struct ggml_tensor * inpCA = cur;
-
-            // norm
-            cur = llm_build_norm(ctx0, cur, hparams,
-                    model.layers[il].attn_norm_cross, NULL,
-                    LLM_NORM_RMS, cb, il);
-            cb(cur, "attn_norm_cross", il);
-
-            // cross-attention
-            {
-                struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq_cross, cur);
-                cb(Qcur, "Qcur", il);
-
-                struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk_cross, embd_enc);
-                cb(Kcur, "Kcur", il);
-
-                struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv_cross, embd_enc);
-                cb(Vcur, "Vcur", il);
-
-                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
-                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_outputs_enc);
-
-                struct ggml_tensor * q =                 ggml_permute(ctx0, Qcur, 0, 2, 1, 3);
-                struct ggml_tensor * k = ggml_cont(ctx0, ggml_permute(ctx0, Kcur, 0, 2, 1, 3));
-
-                struct ggml_tensor * kq = ggml_mul_mat(ctx0, k, q);
-                cb(kq, "kq", il);
-
-                kq = ggml_soft_max_ext(ctx0, kq, KQ_mask_cross, 1.0f, hparams.f_max_alibi_bias);
-                cb(kq, "kq_soft_max_ext", il);
-
-                struct ggml_tensor * v = ggml_cont(ctx0, ggml_transpose(ctx0, ggml_reshape_2d(ctx0, Vcur, n_embd_gqa, n_outputs_enc)));
-                cb(v, "v", il);
-
-                struct ggml_tensor * kqv = ggml_mul_mat(ctx0, ggml_reshape_3d(ctx0, v, n_outputs_enc, n_embd_head, n_head_kv), kq);
-                cb(kqv, "kqv", il);
-
-                struct ggml_tensor * kqv_merged = ggml_permute(ctx0, kqv, 0, 2, 1, 3);
-                cb(kqv_merged, "kqv_merged", il);
-
-                cur = ggml_cont_2d(ctx0, kqv_merged, n_embd_gqa, n_tokens);
-                cb(cur, "kqv_merged_cont", il);
-
-                ggml_build_forward_expand(gf, cur);
-
-                cur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wo_cross, cur);
-                cb(cur, "kqv_out", il);
-            }
-
-            if (il == n_layer - 1) {
-                // skip computing output for unused tokens
-                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
-                n_tokens = n_outputs;
-                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
-                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
-                inpCA = ggml_get_rows(ctx0, inpCA, inp_out_ids);
-            }
-
-            struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpCA);
-            cb(ffn_inp, "ffn_inp", il);
-
-            // feed-forward network
-            {
-                cur = llm_build_norm(ctx0, ffn_inp, hparams,
-                        model.layers[il].ffn_norm, NULL,
-                        LLM_NORM_RMS, cb, il);
-                cb(cur, "ffn_norm", il);
-
-                // T5 uses relu, flan-T5 uses gelu-gated
-                cur = llm_build_ffn(ctx0, lctx, cur,
-                        model.layers[il].ffn_up,   NULL, NULL,
-                        model.layers[il].ffn_gate, NULL, NULL,
-                        model.layers[il].ffn_down, NULL, NULL,
-                        NULL,
-                        model.layers[il].ffn_gate_enc ? LLM_FFN_GELU : LLM_FFN_RELU,
-                        model.layers[il].ffn_gate_enc ? LLM_FFN_PAR : LLM_FFN_SEQ,
-                        cb, il);
-                cb(cur, "ffn_out", il);
-            }
-
-            cur = ggml_add(ctx0, cur, ffn_inp);
-            cb(cur, "ffn_out", il);
-
-            ggml_tensor * layer_dir = lctx.cvec.tensor_for(il);
-            if (layer_dir != nullptr) {
-                cur = ggml_add(ctx0, cur, layer_dir);
-            }
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-        cb(cur, "result_embd", -1);
-
-        cur = llm_build_norm(ctx0, cur, hparams,
-                model.output_norm, NULL,
-                LLM_NORM_RMS, cb, -1);
-        cb(cur, "result_norm", -1);
-
-        // lm_head
-        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
-        cb(cur, "result_output", -1);
-
-        ggml_build_forward_expand(gf, cur);
-
-        return gf;
-    }
-
-    struct ggml_cgraph * build_jais() {
-        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false);
-
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-        const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-
-        struct ggml_tensor * cur;
-        struct ggml_tensor * inpL;
-
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
-
-        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
-        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
-
-        for (int il = 0; il < n_layer; ++il) {
-            cur = llm_build_norm(ctx0, inpL, hparams,
-                    model.layers[il].attn_norm,
-                    model.layers[il].attn_norm_b,
-                    LLM_NORM, cb, il);
-            cb(cur, "attn_norm", il);
-
-            // self-attention
-            {
-                cur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wqkv, cur);
-                cb(cur, "wqkv", il);
-
-                cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
-                cb(cur, "bqkv", il);
-
-                struct ggml_tensor * Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd,     n_tokens, cur->nb[1], 0*cur->nb[0]*(n_embd)));
-                struct ggml_tensor * Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*cur->nb[0]*(n_embd)));
-                struct ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*cur->nb[0]*(n_embd + n_embd_gqa)));
-
-                cb(Qcur, "Qcur", il);
-                cb(Kcur, "Kcur", il);
-                cb(Vcur, "Vcur", il);
-
-                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
-
-                cur = llm_build_kv(ctx0, lctx, kv_self, gf,
-                        model.layers[il].wo, model.layers[il].bo,
-                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/float(n_embd_head), cb, il);
-            }
-
-            if (il == n_layer - 1) {
-                // skip computing output for unused tokens
-                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
-                cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
-                inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
-            }
-
-            // add the input
-            struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL);
-            cb(ffn_inp, "ffn_inp", il);
-
-            // FF
-            {
-                cur = llm_build_norm(ctx0, ffn_inp, hparams,
-                        model.layers[il].ffn_norm,
-                        model.layers[il].ffn_norm_b,
-                        LLM_NORM, cb, il);
-                cb(cur, "ffn_norm", il);
-
-                cur = llm_build_ffn(ctx0, lctx, cur,
-                        model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
-                        model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL,
-                        model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
-                        NULL,
-                        LLM_FFN_SILU, LLM_FFN_PAR, cb, il);
-                cb(cur, "ffn_out", il);
-            }
-
-            inpL = ggml_add(ctx0, cur, ffn_inp);
-            cb(inpL, "l_out", il);
-        }
-
-        cur = llm_build_norm(ctx0, inpL, hparams,
-                model.output_norm,
-                model.output_norm_b,
-                LLM_NORM, cb, -1);
-        cb(cur, "result_norm", -1);
-
-        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
-
-        cb(cur, "result_output", -1);
-
-        ggml_build_forward_expand(gf, cur);
-
-        return gf;
-    }
-
-    struct ggml_cgraph * build_chatglm() {
-        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false);
-
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-        const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-
-        struct ggml_tensor * cur;
-        struct ggml_tensor * inpL;
-
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
-
-        // inp_pos - contains the positions
-        struct ggml_tensor * inp_pos = build_inp_pos();
-
-        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
-        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
-
-        for (int il = 0; il < n_layer; ++il) {
-            struct ggml_tensor * inpSA = inpL;
-
-            cur = llm_build_norm(ctx0, inpL, hparams,
-                    model.layers[il].attn_norm,
-                    NULL,
-                    LLM_NORM_RMS, cb, il);
-            cb(cur, "attn_norm", il);
-
-            // self-attention
-            {
-                struct ggml_tensor * Qcur = nullptr;
-                struct ggml_tensor * Kcur = nullptr;
-                struct ggml_tensor * Vcur = nullptr;
-                if (model.layers[il].wqkv == nullptr) {
-                    Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur);
-                    if (model.layers[il].bq) {
-                        Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
-                    }
-                    Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur);
-                    if (model.layers[il].bk) {
-                        Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
-                    }
-                    Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur);
-                    if (model.layers[il].bv) {
-                        Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
-                    }
-                } else {
-                    cur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wqkv, cur);
-                    cb(cur, "wqkv", il);
-                    if (model.layers[il].bqkv) {
-                        cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
-                        cb(cur, "bqkv", il);
-                    }
-                    Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd,     n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd)));
-                    Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd)));
-                    Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa)));
-                }
-                cb(Qcur, "Qcur", il);
-                cb(Kcur, "Kcur", il);
-                cb(Vcur, "Vcur", il);
-                //printf("freq_base: %f freq_scale: %f ext_factor: %f attn_factor: %f\n", freq_base, freq_scale, ext_factor, attn_factor);
-                Qcur = ggml_rope_ext(
-                    ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr,
-                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                    ext_factor, attn_factor, beta_fast, beta_slow
-                );
-                cb(Qcur, "Qcur_rope", il);
-
-                Kcur = ggml_rope_ext(
-                    ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr,
-                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                    ext_factor, attn_factor, beta_fast, beta_slow
-                );
-                cb(Kcur, "Kcur_rope", il);
-
-                cur = llm_build_kv(ctx0, lctx, kv_self, gf,
-                        model.layers[il].wo, NULL,
-                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
-
-            }
-
-            if (il == n_layer - 1) {
-                // skip computing output for unused tokens
-                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
-                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
-                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
-            }
-
-            // Add the input
-            struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
-            cb(ffn_inp, "ffn_inp", il);
-
-            // FF
-            {
-                cur = llm_build_norm(ctx0, ffn_inp, hparams,
-                        model.layers[il].ffn_norm,
-                        NULL,
-                        LLM_NORM_RMS, cb, il);
-                cb(cur, "ffn_norm", il);
-
-                cur = llm_build_ffn(ctx0, lctx, cur,
-                        model.layers[il].ffn_up,   NULL, NULL,
-                        NULL,                      NULL, NULL,
-                        model.layers[il].ffn_down, NULL, NULL,
-                        NULL,
-                        LLM_FFN_SWIGLU, LLM_FFN_SEQ, cb, il);
-                cb(cur, "ffn_out", il);
-
-            }
-
-            inpL = ggml_add(ctx0, cur, ffn_inp);
-            cb(inpL, "l_out", il);
-        }
-
-        cur = llm_build_norm(ctx0, inpL, hparams,
-                model.output_norm,
-                NULL,
-                LLM_NORM_RMS, cb, -1);
-        cb(cur, "result_norm", -1);
-
-        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
-        cb(cur, "result_output", -1);
-
-        ggml_build_forward_expand(gf, cur);
-
-        return gf;
-    }
-
-    struct ggml_cgraph * build_nemotron() {
-        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false);
-
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-        //GGML_ASSERT(n_embd_head == hparams.n_rot);
-
-        struct ggml_tensor * cur;
-        struct ggml_tensor * inpL;
-
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
-
-        // inp_pos - contains the positions
-        struct ggml_tensor * inp_pos = build_inp_pos();
-
-        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
-        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
-
-        for (int il = 0; il < n_layer; ++il) {
-            struct ggml_tensor * inpSA = inpL;
-
-            // norm
-            cur = llm_build_norm(ctx0, inpL, hparams,
-                    model.layers[il].attn_norm,
-                    model.layers[il].attn_norm_b,
-                    LLM_NORM, cb, il);
-            cb(cur, "attn_norm", il);
-
-            // self-attention
-            {
-                // compute Q and K and RoPE them
-                struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur);
-                cb(Qcur, "Qcur", il);
-                if (model.layers[il].bq) {
-                    Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
-                    cb(Qcur, "Qcur", il);
-                }
-
-                struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur);
-                cb(Kcur, "Kcur", il);
-                if (model.layers[il].bk) {
-                    Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
-                    cb(Kcur, "Kcur", il);
-                }
-
-                struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur);
-                cb(Vcur, "Vcur", il);
-                if (model.layers[il].bv) {
-                    Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
-                    cb(Vcur, "Vcur", il);
-                }
-
-                Qcur = ggml_rope_ext(
-                    ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr,
-                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                    ext_factor, attn_factor, beta_fast, beta_slow
-                );
-                cb(Qcur, "Qcur", il);
-
-                Kcur = ggml_rope_ext(
-                    ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr,
-                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                    ext_factor, attn_factor, beta_fast, beta_slow
-                );
-                cb(Kcur, "Kcur", il);
-
-                cur = llm_build_kv(ctx0, lctx, kv_self, gf,
-                        model.layers[il].wo, model.layers[il].bo,
-                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
-            }
-
-            if (il == n_layer - 1) {
-                // skip computing output for unused tokens
-                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
-                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
-                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
-            }
-
-            struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
-            cb(ffn_inp, "ffn_inp", il);
-
-            // feed-forward network
-            cur = llm_build_norm(ctx0, ffn_inp, hparams,
-                    model.layers[il].ffn_norm,
-                    model.layers[il].ffn_norm_b,
-                    LLM_NORM, cb, il);
-            cb(cur, "ffn_norm", il);
-
-            cur = llm_build_ffn(ctx0, lctx, cur,
-                    model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
-                    NULL,                      NULL,                        NULL,
-                    model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
-                    NULL,
-                    LLM_FFN_RELU_SQR, LLM_FFN_SEQ, cb, il);
-
-            cur = ggml_add(ctx0, cur, ffn_inp);
-            cb(cur, "ffn_out", il);
-
-            cur = lctx.cvec.apply_to(ctx0, cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-
-        cur = llm_build_norm(ctx0, cur, hparams,
-                model.output_norm, model.output_norm_b,
-                LLM_NORM, cb, -1);
-        cb(cur, "result_norm", -1);
-
-        // lm_head
-        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
-        cb(cur, "result_output", -1);
-
-        ggml_build_forward_expand(gf, cur);
-
-        return gf;
-    }
-
-    struct ggml_cgraph * build_exaone() {
-        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false);
-
-        // mutable variable, needed during the last layer of the computation to skip unused tokens
-        int32_t n_tokens = this->n_tokens;
-
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-        GGML_ASSERT(n_embd_head == hparams.n_rot);
-
-        struct ggml_tensor * cur;
-        struct ggml_tensor * inpL;
-
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
-
-        // inp_pos - contains the positions
-        struct ggml_tensor * inp_pos = build_inp_pos();
-
-        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
-        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
-
-        for (int il = 0; il < n_layer; ++il) {
-            struct ggml_tensor * inpSA = inpL;
-
-            // norm
-            cur = llm_build_norm(ctx0, inpL, hparams,
-                    model.layers[il].attn_norm, NULL,
-                    LLM_NORM_RMS, cb, il);
-            cb(cur, "attn_norm", il);
-
-            // self-attention
-            {
-                // rope freq factors for llama3; may return nullptr for llama2 and other models
-                struct ggml_tensor * rope_factors = build_rope_factors(il);
-
-                // compute Q and K and RoPE them
-                struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur);
-                cb(Qcur, "Qcur", il);
-                if (model.layers[il].bq) {
-                    Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
-                    cb(Qcur, "Qcur", il);
-                }
-
-                struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur);
-                cb(Kcur, "Kcur", il);
-                if (model.layers[il].bk) {
-                    Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
-                    cb(Kcur, "Kcur", il);
-                }
-
-                struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur);
-                cb(Vcur, "Vcur", il);
-                if (model.layers[il].bv) {
-                    Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
-                    cb(Vcur, "Vcur", il);
-                }
-
-                Qcur = ggml_rope_ext(
-                    ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, rope_factors,
-                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                    ext_factor, attn_factor, beta_fast, beta_slow
-                );
-                cb(Qcur, "Qcur", il);
-
-                Kcur = ggml_rope_ext(
-                    ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, rope_factors,
-                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                    ext_factor, attn_factor, beta_fast, beta_slow
-                );
-                cb(Kcur, "Kcur", il);
-
-                cur = llm_build_kv(ctx0, lctx, kv_self, gf,
-                        model.layers[il].wo, model.layers[il].bo,
-                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
-            }
-
-            if (il == n_layer - 1) {
-                // skip computing output for unused tokens
-                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
-                n_tokens = n_outputs;
-                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
-                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
-            }
-
-            struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
-            cb(ffn_inp, "ffn_inp", il);
-
-            // feed-forward network
-            cur = llm_build_norm(ctx0, ffn_inp, hparams,
-                    model.layers[il].ffn_norm, NULL,
-                    LLM_NORM_RMS, cb, il);
-            cb(cur, "ffn_norm", il);
-
-            cur = llm_build_ffn(ctx0, lctx, cur,
-                    model.layers[il].ffn_up,   NULL, NULL,
-                    model.layers[il].ffn_gate, NULL, NULL,
-                    model.layers[il].ffn_down, NULL, NULL,
-                    NULL,
-                    LLM_FFN_SILU, LLM_FFN_PAR, cb, il);
-            cb(cur, "ffn_out", il);
-
-            cur = ggml_add(ctx0, cur, ffn_inp);
-            cb(cur, "ffn_out", il);
-
-            cur = lctx.cvec.apply_to(ctx0, cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-
-        cur = llm_build_norm(ctx0, cur, hparams,
-                model.output_norm, NULL,
-                LLM_NORM_RMS, cb, -1);
-        cb(cur, "result_norm", -1);
-
-        // lm_head
-        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
-        cb(cur, "result_output", -1);
-
-        ggml_build_forward_expand(gf, cur);
-
-        return gf;
-    }
-
-    ggml_cgraph * build_rwkv6() {
-        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false);
-
-        // Token shift state dimensions should be 2 * n_emb
-        GGML_ASSERT(n_embd == hparams.n_embd_k_s() / 2);
-
-        const int64_t n_seqs = ubatch.n_seqs;
-        const int64_t n_seq_tokens = ubatch.n_seq_tokens;
-        const int64_t n_tokens = ubatch.n_tokens;
-        GGML_ASSERT(n_seqs != 0);
-        GGML_ASSERT(ubatch.equal_seqs);
-        GGML_ASSERT(n_tokens == n_seq_tokens * n_seqs);
-
-        struct ggml_tensor * cur;
-        struct ggml_tensor * inpL;
-        struct ggml_tensor * state_copy = build_inp_s_copy();
-        struct ggml_tensor * state_mask = build_inp_s_mask();
-
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
-        inpL = llm_build_norm(ctx0, inpL, hparams, model.tok_norm, model.tok_norm_b, LLM_NORM, cb, -1);
-
-        for (int il = 0; il < n_layer; ++il) {
-            const llama_layer * layer = &model.layers[il];
-
-            // (ab)using the KV cache to store the states
-            struct ggml_tensor * token_shift = llm_build_copy_mask_state(ctx0,
-                    gf, kv_self.k_l[il], state_copy, state_mask,
-                    hparams.n_embd_k_s(), kv_self.size, kv_head, n_kv, n_seqs);
-            struct ggml_tensor * wkv_states = llm_build_copy_mask_state(ctx0,
-                    gf, kv_self.v_l[il], state_copy, state_mask,
-                    hparams.n_embd_v_s(), kv_self.size, kv_head, n_kv, n_seqs);
-
-            cur = ggml_reshape_3d(ctx0, inpL, n_embd, n_seq_tokens, n_seqs);
-            token_shift = ggml_reshape_3d(ctx0, token_shift, n_embd, 2, n_seqs);
-
-            struct ggml_tensor * att_shift = ggml_view_3d(ctx0, token_shift, n_embd, 1, n_seqs, token_shift->nb[1], token_shift->nb[2], 0);
-            struct ggml_tensor * ffn_shift = ggml_view_3d(ctx0, token_shift, n_embd, 1, n_seqs, token_shift->nb[1], token_shift->nb[2], n_embd * ggml_element_size(token_shift));
-
-            struct ggml_tensor * x_norm_att = llm_build_norm(ctx0, cur, hparams, layer->attn_norm, layer->attn_norm_b, LLM_NORM, cb, il);
-            struct ggml_tensor * x_prev = ggml_concat(
-                ctx0,
-                att_shift,
-                ggml_view_3d(ctx0, x_norm_att, n_embd, n_seq_tokens - 1, n_seqs, x_norm_att->nb[1], x_norm_att->nb[2], 0),
-                1
-            );
-
-            cur = ggml_add(ctx0, cur, llm_build_rwkv6_time_mix(lctx, ctx0, layer, x_norm_att, x_prev, &wkv_states, hparams.wkv_head_size, n_embd / hparams.wkv_head_size));
-            ggml_build_forward_expand(gf, cur);
-            ggml_build_forward_expand(
-                gf,
-                ggml_cpy(
-                    ctx0,
-                    wkv_states,
-                    ggml_view_1d(
-                        ctx0,
-                        kv_self.v_l[il],
-                        hparams.n_embd_v_s() * n_seqs,
-                        hparams.n_embd_v_s() * kv_head * ggml_element_size(kv_self.v_l[il])
-                    )
-                )
-            );
-
-            struct ggml_tensor * x_norm_ffn = llm_build_norm(ctx0, cur, hparams, layer->attn_norm_2, layer->attn_norm_2_b, LLM_NORM, cb, il);
-            x_prev = ggml_concat(
-                ctx0,
-                ffn_shift,
-                ggml_view_3d(ctx0, x_norm_ffn, n_embd, n_seq_tokens - 1, n_seqs, x_norm_ffn->nb[1], x_norm_ffn->nb[2], 0),
-                1
-            );
-            cur = ggml_add(ctx0, cur, llm_build_rwkv6_channel_mix(lctx, ctx0, layer, x_norm_ffn, x_prev));
-            ggml_build_forward_expand(gf, cur);
-
-            struct ggml_tensor * last_norm_att = ggml_view_3d(ctx0, x_norm_att, n_embd, 1, n_seqs, x_norm_att->nb[1], x_norm_att->nb[2], (n_seq_tokens-1)*n_embd*ggml_element_size(x_norm_att));
-            struct ggml_tensor * last_norm_ffn = ggml_view_3d(ctx0, x_norm_ffn, n_embd, 1, n_seqs, x_norm_ffn->nb[1], x_norm_ffn->nb[2], (n_seq_tokens-1)*n_embd*ggml_element_size(x_norm_ffn));
-
-            token_shift = ggml_concat(ctx0, last_norm_att, last_norm_ffn, 1);
-
-            ggml_build_forward_expand(
-                gf,
-                ggml_cpy(
-                    ctx0,
-                    ggml_view_1d(ctx0, token_shift, n_embd * n_seqs * 2, 0),
-                    ggml_view_1d(ctx0, kv_self.k_l[il], hparams.n_embd_k_s() * n_seqs, hparams.n_embd_k_s() * kv_head * ggml_element_size(kv_self.k_l[il]))
-                )
-            );
-
-            if (hparams.rescale_every_n_layers != 0 && (il + 1) % hparams.rescale_every_n_layers == 0) {
-                cur = ggml_scale(ctx0, cur, 0.5F);
-            }
-
-            cur = lctx.cvec.apply_to(ctx0, cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-        struct ggml_tensor * inp_out_ids = build_inp_out_ids();
-        cur = ggml_reshape_2d(ctx0, cur, n_embd, n_tokens);
-        cur = ggml_get_rows(ctx0, cur, inp_out_ids);
-
-        cur = llm_build_norm(ctx0, cur, hparams, model.output_norm, model.output_norm_b, LLM_NORM, cb, -1);
-        cb(cur, "result_norm", -1);
-
-        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
-        cb(cur, "result_output", -1);
-
-        ggml_build_forward_expand(gf, cur);
-
-        return gf;
-    }
-
-    // ref: https://huggingface.co/recursal/QRWKV6-32B-Instruct-Preview-v0.1/blob/main/modeling_rwkv6qwen2.py
-    ggml_cgraph * build_rwkv6qwen2() {
-        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false);
-
-        GGML_ASSERT(n_embd == hparams.n_embd_k_s());
-
-        const int64_t n_seqs = ubatch.n_seqs;
-        const int64_t n_seq_tokens = ubatch.n_seq_tokens;
-        const int64_t n_tokens = ubatch.n_tokens;
-        GGML_ASSERT(n_seqs != 0);
-        GGML_ASSERT(ubatch.equal_seqs);
-        GGML_ASSERT(n_tokens == n_seq_tokens * n_seqs);
-
-        struct ggml_tensor * cur;
-        struct ggml_tensor * inpL;
-        struct ggml_tensor * state_copy = build_inp_s_copy();
-        struct ggml_tensor * state_mask = build_inp_s_mask();
-
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
-
-        for (int il = 0; il < n_layer; ++il) {
-            const llama_layer * layer = &model.layers[il];
-
-            // (ab)using the KV cache to store the states
-            struct ggml_tensor * token_shift = llm_build_copy_mask_state(ctx0,
-                    gf, kv_self.k_l[il], state_copy, state_mask,
-                    hparams.n_embd_k_s(), kv_self.size, kv_head, n_kv, n_seqs);
-            struct ggml_tensor * wkv_states = llm_build_copy_mask_state(ctx0,
-                    gf, kv_self.v_l[il], state_copy, state_mask,
-                    hparams.n_embd_v_s(), kv_self.size, kv_head, n_kv, n_seqs);
-
-            cur = ggml_reshape_3d(ctx0, inpL, n_embd, n_seq_tokens, n_seqs);
-            token_shift = ggml_reshape_3d(ctx0, token_shift, n_embd, 1, n_seqs);
-
-            struct ggml_tensor * x_norm_att = llm_build_norm(ctx0, cur, hparams, layer->attn_norm, layer->attn_norm_b, LLM_NORM_RMS, cb, il);
-            struct ggml_tensor * x_prev = ggml_concat(
-                ctx0,
-                token_shift,
-                ggml_view_3d(ctx0, x_norm_att, n_embd, n_seq_tokens - 1, n_seqs, x_norm_att->nb[1], x_norm_att->nb[2], 0),
-                1
-            );
-
-            struct ggml_tensor * last_norm_att = ggml_view_3d(ctx0, x_norm_att, n_embd, 1, n_seqs, x_norm_att->nb[1], x_norm_att->nb[2], (n_seq_tokens-1)*n_embd*ggml_element_size(x_norm_att));
-            ggml_build_forward_expand(
-                gf,
-                ggml_cpy(
-                    ctx0,
-                    ggml_view_1d(ctx0, last_norm_att, n_embd * n_seqs, 0),
-                    ggml_view_1d(ctx0, kv_self.k_l[il], hparams.n_embd_k_s() * n_seqs, hparams.n_embd_k_s() * kv_head * ggml_element_size(kv_self.k_l[il]))
-                )
-            );
-
-            struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, llm_build_rwkv6_time_mix(lctx, ctx0, layer, x_norm_att, x_prev, &wkv_states, hparams.wkv_head_size, hparams.n_head_kv()));
-            ggml_build_forward_expand(gf, ffn_inp);
-            ggml_build_forward_expand(
-                gf,
-                ggml_cpy(
-                    ctx0,
-                    wkv_states,
-                    ggml_view_1d(
-                        ctx0,
-                        kv_self.v_l[il],
-                        hparams.n_embd_v_s() * n_seqs,
-                        hparams.n_embd_v_s() * kv_head * ggml_element_size(kv_self.v_l[il])
-                    )
-                )
-            );
-
-            cb(ffn_inp, "ffn_inp", il);
-
-            // feed-forward network
-            cur = llm_build_norm(ctx0, ffn_inp, hparams,
-                    model.layers[il].ffn_norm, NULL,
-                    LLM_NORM_RMS, cb, il);
-            cb(cur, "ffn_norm", il);
-
-            cur = llm_build_ffn(ctx0, lctx, cur,
-                    model.layers[il].ffn_up,   NULL, NULL,
-                    model.layers[il].ffn_gate, NULL, NULL,
-                    model.layers[il].ffn_down, NULL, NULL,
-                    NULL,
-                    LLM_FFN_SILU, LLM_FFN_PAR, cb, il);
-            cb(cur, "ffn_out", il);
-
-            cur = ggml_add(ctx0, cur, ffn_inp);
-            cur = lctx.cvec.apply_to(ctx0, cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-        struct ggml_tensor * inp_out_ids = build_inp_out_ids();
-        cur = ggml_reshape_2d(ctx0, cur, n_embd, n_tokens);
-        cur = ggml_get_rows(ctx0, cur, inp_out_ids);
-
-        cur = llm_build_norm(ctx0, cur, hparams, model.output_norm, model.output_norm_b, LLM_NORM_RMS, cb, -1);
-        cb(cur, "result_norm", -1);
-
-        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
-        cb(cur, "result_output", -1);
-
-        ggml_build_forward_expand(gf, cur);
-
-        return gf;
-    }
-
-    // ref: https://github.com/facebookresearch/chameleon
-    // based on the original build_llama() function, changes:
-    //   * qk-norm
-    //   * swin-norm
-    //   * removed bias
-    //   * removed MoE
-    struct ggml_cgraph * build_chameleon() {
-        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false);
-
-        // mutable variable, needed during the last layer of the computation to skip unused tokens
-        int32_t n_tokens = this->n_tokens;
-
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-        GGML_ASSERT(n_embd_head == hparams.n_rot);
-
-        struct ggml_tensor * cur;
-        struct ggml_tensor * inpL;
-
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
-
-        // inp_pos - contains the positions
-        struct ggml_tensor * inp_pos = build_inp_pos();
-
-        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
-        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
-
-        for (int il = 0; il < n_layer; ++il) {
-            struct ggml_tensor * inpSA = inpL;
-
-            // norm
-            if (hparams.swin_norm) {
-                cur = inpL;
-            } else {
-                cur = llm_build_norm(ctx0, inpL, hparams,
-                    model.layers[il].attn_norm, NULL,
-                    LLM_NORM_RMS, cb, il);
-                cb(cur, "attn_norm", il);
-            }
-
-            // self-attention
-            {
-                // compute Q and K and RoPE them
-                struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur);
-                cb(Qcur, "Qcur", il);
-
-                struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur);
-                cb(Kcur, "Kcur", il);
-
-                struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur);
-                cb(Vcur, "Vcur", il);
-
-                if (model.layers[il].attn_q_norm) {
-                    Qcur = ggml_view_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens,
-                                ggml_element_size(Qcur) * n_embd_head,
-                                ggml_element_size(Qcur) * n_embd_head * n_head,
-                                0);
-                    cb(Qcur, "Qcur", il);
-
-                    Qcur = llm_build_norm(ctx0, Qcur, hparams,
-                                model.layers[il].attn_q_norm,
-                                model.layers[il].attn_q_norm_b,
-                                LLM_NORM, cb, il);
-                    cb(Qcur, "Qcur", il);
-                }
-
-                if (model.layers[il].attn_k_norm) {
-                    Kcur = ggml_view_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens,
-                                ggml_element_size(Kcur) * n_embd_head,
-                                ggml_element_size(Kcur) * n_embd_head * n_head_kv,
-                                0);
-                    cb(Kcur, "Kcur", il);
-
-                    Kcur = llm_build_norm(ctx0, Kcur, hparams,
-                               model.layers[il].attn_k_norm,
-                               model.layers[il].attn_k_norm_b,
-                               LLM_NORM, cb, il);
-                    cb(Kcur, "Kcur", il);
-                }
-
-                Qcur = ggml_rope_ext(
-                    ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr,
-                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                    ext_factor, attn_factor, beta_fast, beta_slow
-                );
-                cb(Qcur, "Qcur", il);
-
-                Kcur = ggml_rope_ext(
-                    ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr,
-                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                    ext_factor, attn_factor, beta_fast, beta_slow
-                );
-                cb(Kcur, "Kcur", il);
-
-                cur = llm_build_kv(ctx0, lctx, kv_self, gf,
-                        model.layers[il].wo, nullptr,
-                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
-
-                if (hparams.swin_norm) {
-                    cur = llm_build_norm(ctx0, cur, hparams,
-                        model.layers[il].attn_norm, NULL,
-                        LLM_NORM_RMS, cb, il);
-                }
-            }
-
-            if (il == n_layer - 1) {
-                // skip computing output for unused tokens
-                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
-                n_tokens = n_outputs;
-                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
-                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
-            }
-
-            struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
-            cb(ffn_inp, "ffn_inp", il);
-
-            // feed-forward network
-            if (!hparams.swin_norm) {
-                cur = llm_build_norm(ctx0, ffn_inp, hparams,
-                        model.layers[il].ffn_norm, NULL,
-                        LLM_NORM_RMS, cb, il);
-                cb(cur, "ffn_norm", il);
-            }
-
-            cur = llm_build_ffn(ctx0, lctx, cur,
-                    model.layers[il].ffn_up,   NULL, NULL,
-                    model.layers[il].ffn_gate, NULL, NULL,
-                    model.layers[il].ffn_down, NULL, NULL,
-                    NULL,
-                    LLM_FFN_SILU, LLM_FFN_PAR, cb, il);
-            cb(cur, "ffn_out", il);
-
-            if (hparams.swin_norm) {
-                cur = llm_build_norm(ctx0, cur, hparams,
-                        model.layers[il].ffn_norm, NULL,
-                        LLM_NORM_RMS, cb, il);
-                cb(cur, "ffn_norm", il);
-            }
-
-            cur = ggml_add(ctx0, cur, ffn_inp);
-            cb(cur, "ffn_out", il);
-
-            cur = lctx.cvec.apply_to(ctx0, cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-
-        cur = llm_build_norm(ctx0, cur, hparams,
-                model.output_norm, NULL,
-                LLM_NORM_RMS, cb, -1);
-        cb(cur, "result_norm", -1);
-
-        // lm_head
-        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
-        cb(cur, "result_output_with_img_logits", -1);
-
-        // TODO: this suppresses the output of image tokens, which is required to enable text-only outputs.
-        // Needs to be removed once image outputs are supported.
-        int img_token_end_idx = 8196;
-        int img_token_start_idx = 4;
-        int num_img_tokens = img_token_end_idx - img_token_start_idx;
-        // creates 1d tensor of size num_img_tokens and values -FLT_MAX,
-        // which ensures that text token values are always at least larger than image token values
-        struct ggml_tensor * img_logits = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, num_img_tokens);
-        img_logits = ggml_clamp(ctx0, img_logits, -FLT_MAX, -FLT_MAX);
-        cb(img_logits, "img_logits", -1);
-        cur = ggml_set_1d(ctx0, cur, img_logits, ggml_element_size(cur) * img_token_start_idx);
-        cb(cur, "result_output", -1);
-
-        ggml_build_forward_expand(gf, cur);
-
-        return gf;
-    }
-
-    struct ggml_cgraph * build_wavtokenizer_dec() {
-        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, model.max_nodes(), false);
-
-        struct ggml_tensor * cur;
-        struct ggml_tensor * inpL;
-
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
-
-        cur = ggml_cont(ctx0, ggml_transpose(ctx0, inpL));
-
-        cur = ggml_conv_1d_ph(ctx0, model.conv1d, cur, 1, 1);
-        cur = ggml_add(ctx0, cur, model.conv1d_b);
-
-        // posnet
-        for (uint32_t il = 0; il < hparams.posnet.n_layer; ++il) {
-            const auto & layer = model.layers[il].posnet;
-
-            inpL = cur;
-
-            switch (il) {
-                case 0:
-                case 1:
-                case 3:
-                case 4:
-                    {
-                        cur = llm_build_norm(ctx0, cur, hparams,
-                                layer.norm1,
-                                layer.norm1_b,
-                                LLM_NORM_GROUP, cb, 0);
-
-                        cur = ggml_mul(ctx0, ggml_sigmoid(ctx0, cur), cur);
-
-                        cur = ggml_conv_1d_ph(ctx0, layer.conv1, cur, 1, 1);
-                        cur = ggml_add(ctx0, cur, layer.conv1_b);
-
-                        cur = llm_build_norm(ctx0, cur, hparams,
-                                layer.norm2,
-                                layer.norm2_b,
-                                LLM_NORM_GROUP, cb, 0);
-
-                        cur = ggml_mul(ctx0, ggml_sigmoid(ctx0, cur), cur);
-
-                        cur = ggml_conv_1d_ph(ctx0, layer.conv2, cur, 1, 1);
-                        cur = ggml_add(ctx0, cur, layer.conv2_b);
-
-                        cur = ggml_add(ctx0, cur, inpL);
-                    } break;
-                case 2:
-                    {
-                        cur = llm_build_norm(ctx0, cur, hparams,
-                                layer.attn_norm,
-                                layer.attn_norm_b,
-                                LLM_NORM_GROUP, cb, 0);
-
-                        struct ggml_tensor * q;
-                        struct ggml_tensor * k;
-                        struct ggml_tensor * v;
-
-                        q = ggml_conv_1d_ph(ctx0, layer.attn_q, cur, 1, 1);
-                        k = ggml_conv_1d_ph(ctx0, layer.attn_k, cur, 1, 1);
-                        v = ggml_conv_1d_ph(ctx0, layer.attn_v, cur, 1, 1);
-
-                        q = ggml_add(ctx0, q, layer.attn_q_b);
-                        k = ggml_add(ctx0, k, layer.attn_k_b);
-                        v = ggml_add(ctx0, v, layer.attn_v_b);
-
-                        q = ggml_cont(ctx0, ggml_transpose(ctx0, q));
-                        k = ggml_cont(ctx0, ggml_transpose(ctx0, k));
-
-                        struct ggml_tensor * kq = ggml_mul_mat(ctx0, k, q);
-
-                        kq = ggml_soft_max_ext(ctx0, kq, nullptr, 1.0f/sqrtf(float(hparams.posnet.n_embd)), 0.0f);
-
-                        cur = ggml_mul_mat(ctx0, kq, v);
-
-                        cur = ggml_conv_1d_ph(ctx0, layer.attn_o, cur, 1, 1);
-                        cur = ggml_add(ctx0, cur, layer.attn_o_b);
-
-                        cur = ggml_add(ctx0, cur, inpL);
-                    } break;
-                case 5:
-                    {
-                        cur = llm_build_norm(ctx0, cur, hparams,
-                                layer.norm,
-                                layer.norm_b,
-                                LLM_NORM_GROUP, cb, 0);
-                    } break;
-                default: GGML_ABORT("unknown posnet layer");
-            };
-        }
-
-        cur = ggml_cont(ctx0, ggml_transpose(ctx0, cur));
-
-        cur = llm_build_norm(ctx0, cur, hparams,
-                model.tok_norm,
-                model.tok_norm_b,
-                LLM_NORM, cb, -1);
-
-        cur = ggml_cont(ctx0, ggml_transpose(ctx0, cur));
-
-        inpL = cur;
-
-        // convnext
-        for (uint32_t il = 0; il < hparams.convnext.n_layer; ++il) {
-            const auto & layer = model.layers[il].convnext;
-
-            cur = inpL;
-
-            cur = ggml_conv_1d_dw_ph(ctx0, layer.dw, cur, 1, 1);
-            cur = ggml_add(ctx0, cur, layer.dw_b);
-
-            cur = ggml_cont(ctx0, ggml_transpose(ctx0, cur));
-
-            cur = llm_build_norm(ctx0, cur, hparams,
-                    layer.norm,
-                    layer.norm_b,
-                    LLM_NORM, cb, -1);
-
-            cur = llm_build_ffn(ctx0, lctx, cur,
-                    layer.pw1, layer.pw1_b, NULL,
-                    NULL,      NULL,        NULL,
-                    layer.pw2, layer.pw2_b, NULL,
-                    NULL,
-                    LLM_FFN_GELU, LLM_FFN_SEQ, cb, il);
-
-            cur = ggml_mul(ctx0, cur, layer.gamma);
-
-            cur = ggml_cont(ctx0, ggml_transpose(ctx0, cur));
-
-            inpL = ggml_add(ctx0, cur, inpL);
-        }
-
-        cur = inpL;
-
-        cur = ggml_cont(ctx0, ggml_transpose(ctx0, cur));
-
-        cur = llm_build_norm(ctx0, cur, hparams,
-                model.output_norm,
-                model.output_norm_b,
-                LLM_NORM, cb, -1);
-
-        // lm_head
-        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
-
-        cur = ggml_add(ctx0, cur, model.output_b);
-        cb(cur, "result_embd", -1);
-
-        ggml_build_forward_expand(gf, cur);
-
-        return gf;
-    }
-};
-
-static struct ggml_cgraph * llama_build_graph_defrag(llama_context & lctx, const std::vector & ids) {
-    llama_ubatch dummy = {};
-    dummy.equal_seqs = true;
-
-    llm_build_cb cb = [&](struct ggml_tensor * , const char * , int ) { };
-
-    struct llm_build_context llm(lctx, dummy, cb, false);
-
-    llm.init();
-
-    struct ggml_cgraph * result = llm.build_defrag(ids);
-
-    llm.free();
-
-    return result;
-}
-
-static struct ggml_cgraph * llama_build_graph_k_shift(llama_context & lctx) {
-    llama_ubatch dummy = {};
-    dummy.equal_seqs = true;
-
-    llm_build_cb cb = [&](struct ggml_tensor * , const char * , int ) { };
-
-    struct llm_build_context llm(lctx, dummy, cb, false);
-
-    llm.init();
-
-    struct ggml_cgraph * result = llm.build_k_shift();
-
-    llm.free();
-
-    return result;
-}
-
-static struct ggml_cgraph * llama_build_graph(
-         llama_context & lctx,
-    const llama_ubatch & ubatch,
-                  bool   worst_case) {
-    const auto & model = lctx.model;
-
-    // this callback allows us to apply custom logic to each tensor (e.g. ggml-alloc, offloading, etc.)
-    llm_build_cb cb = [&](struct ggml_tensor * cur, const char * name, int il) {
-        if (il >= 0) {
-            ggml_format_name(cur, "%s-%d", name, il);
-        } else {
-            ggml_set_name(cur, name);
-        }
-
-        if (!lctx.cparams.offload_kqv) {
-            if (strcmp(name, "kqv_merged_cont") == 0) {
-                // all nodes between the KV store and the attention output are run on the CPU
-                ggml_backend_sched_set_tensor_backend(lctx.sched.get(), cur, lctx.backend_cpu);
-            }
-        }
-
-        // norm may be automatically assigned to the backend of the previous layer, increasing data transfer between backends
-        // FIXME: fix in ggml_backend_sched
-        const bool full_offload = lctx.model.params.n_gpu_layers > (int) lctx.model.hparams.n_layer;
-        if (ubatch.n_tokens < 32 || full_offload) {
-            if (il != -1 && strcmp(name, "norm") == 0) {
-                const auto & dev_layer = lctx.model.dev_layer(il);
-                for (auto & backend : lctx.backends) {
-                    if (ggml_backend_get_device(backend.get()) == dev_layer) {
-                        if (ggml_backend_supports_op(backend.get(), cur)) {
-                            ggml_backend_sched_set_tensor_backend(lctx.sched.get(), cur, backend.get());
-                        }
-                    }
-                }
-            }
-        }
-    };
-
-    struct ggml_cgraph * result = NULL;
-
-    struct llm_build_context llm(lctx, ubatch, cb, worst_case);
-
-    llm.init();
-
-    switch (model.arch) {
-        case LLM_ARCH_LLAMA:
-        case LLM_ARCH_MINICPM:
-        case LLM_ARCH_GRANITE:
-        case LLM_ARCH_GRANITE_MOE:
-            {
-                result = llm.build_llama();
-            } break;
-        case LLM_ARCH_DECI:
-            {
-                result = llm.build_deci();
-            } break;
-        case LLM_ARCH_BAICHUAN:
-            {
-                result = llm.build_baichuan();
-            } break;
-        case LLM_ARCH_FALCON:
-            {
-                result = llm.build_falcon();
-            } break;
-        case LLM_ARCH_GROK:
-            {
-                result = llm.build_grok();
-            } break;
-        case LLM_ARCH_STARCODER:
-            {
-                result = llm.build_starcoder();
-            } break;
-        case LLM_ARCH_REFACT:
-            {
-                result = llm.build_refact();
-            } break;
-        case LLM_ARCH_BERT:
-        case LLM_ARCH_JINA_BERT_V2:
-        case LLM_ARCH_NOMIC_BERT:
-            {
-                result = llm.build_bert();
-            } break;
-        case LLM_ARCH_BLOOM:
-            {
-                result = llm.build_bloom();
-            } break;
-        case LLM_ARCH_MPT:
-            {
-                result = llm.build_mpt();
-            } break;
-         case LLM_ARCH_STABLELM:
-            {
-                result = llm.build_stablelm();
-            } break;
-        case LLM_ARCH_QWEN:
-            {
-                result = llm.build_qwen();
-            } break;
-        case LLM_ARCH_QWEN2:
-            {
-                result = llm.build_qwen2();
-            } break;
-        case LLM_ARCH_QWEN2VL:
-            {
-                lctx.n_pos_per_token = 4;
-                result = llm.build_qwen2vl();
-            } break;
-        case LLM_ARCH_QWEN2MOE:
-            {
-                result = llm.build_qwen2moe();
-            } break;
-        case LLM_ARCH_PHI2:
-            {
-                result = llm.build_phi2();
-            } break;
-        case LLM_ARCH_PHI3:
-        case LLM_ARCH_PHIMOE:
-            {
-                result = llm.build_phi3();
-            } break;
-        case LLM_ARCH_PLAMO:
-            {
-                result = llm.build_plamo();
-            } break;
-        case LLM_ARCH_GPT2:
-            {
-                result = llm.build_gpt2();
-            } break;
-        case LLM_ARCH_CODESHELL:
-            {
-                result = llm.build_codeshell();
-            } break;
-        case LLM_ARCH_ORION:
-            {
-                result = llm.build_orion();
-            } break;
-        case LLM_ARCH_INTERNLM2:
-            {
-                result = llm.build_internlm2();
-            } break;
-        case LLM_ARCH_MINICPM3:
-            {
-                result = llm.build_minicpm3();
-            } break;
-        case LLM_ARCH_GEMMA:
-            {
-                result = llm.build_gemma();
-            } break;
-        case LLM_ARCH_GEMMA2:
-            {
-                result = llm.build_gemma2();
-            } break;
-        case LLM_ARCH_STARCODER2:
-            {
-                result = llm.build_starcoder2();
-            } break;
-        case LLM_ARCH_MAMBA:
-            {
-                result = llm.build_mamba();
-            } break;
-        case LLM_ARCH_XVERSE:
-            {
-                result = llm.build_xverse();
-            } break;
-        case LLM_ARCH_COMMAND_R:
-            {
-                result = llm.build_command_r();
-            } break;
-        case LLM_ARCH_COHERE2:
-            {
-                result = llm.build_cohere2();
-            } break;
-        case LLM_ARCH_DBRX:
-            {
-                result = llm.build_dbrx();
-            } break;
-        case LLM_ARCH_OLMO:
-            {
-                result = llm.build_olmo();
-            } break;
-        case LLM_ARCH_OLMO2:
-            {
-                result = llm.build_olmo2();
-            } break;
-        case LLM_ARCH_OLMOE:
-            {
-                result = llm.build_olmoe();
-            } break;
-        case LLM_ARCH_OPENELM:
-            {
-                result = llm.build_openelm();
-            } break;
-        case LLM_ARCH_GPTNEOX:
-            {
-                result = llm.build_gptneox();
-            } break;
-        case LLM_ARCH_ARCTIC:
-            {
-                result = llm.build_arctic();
-            } break;
-        case LLM_ARCH_DEEPSEEK:
-            {
-                result = llm.build_deepseek();
-            } break;
-        case LLM_ARCH_DEEPSEEK2:
-            {
-                result = llm.build_deepseek2();
-            } break;
-        case LLM_ARCH_CHATGLM:
-            {
-                result = llm.build_chatglm();
-            } break;
-        case LLM_ARCH_BITNET:
-            {
-                result = llm.build_bitnet();
-            } break;
-        case LLM_ARCH_T5:
-            {
-                if (lctx.is_encoding) {
-                    result = llm.build_t5_enc();
-                } else {
-                    result = llm.build_t5_dec();
-                }
-            } break;
-        case LLM_ARCH_T5ENCODER:
-            {
-                result = llm.build_t5_enc();
-            } break;
-        case LLM_ARCH_JAIS:
-            {
-                result = llm.build_jais();
-            } break;
-        case LLM_ARCH_NEMOTRON:
-            {
-                result = llm.build_nemotron();
-            } break;
-        case LLM_ARCH_EXAONE:
-            {
-                result = llm.build_exaone();
-            } break;
-        case LLM_ARCH_RWKV6:
-            {
-                result = llm.build_rwkv6();
-            } break;
-        case LLM_ARCH_RWKV6QWEN2:
-            {
-                result = llm.build_rwkv6qwen2();
-            } break;
-        case LLM_ARCH_CHAMELEON:
-            {
-                result = llm.build_chameleon();
-            } break;
-        case LLM_ARCH_WAVTOKENIZER_DEC:
-            {
-                result = llm.build_wavtokenizer_dec();
-            } break;
-        default:
-            GGML_ABORT("fatal error");
-    }
-
-    // add on pooling layer
-    if (lctx.cparams.embeddings) {
-        result = llm.append_pooling(result);
-    }
-
-    llm.free();
-
-    return result;
-}
-
-// returns the result of ggml_backend_sched_graph_compute_async execution
-static enum ggml_status llama_graph_compute(
-          llama_context & lctx,
-            ggml_cgraph * gf,
-                    int   n_threads,
-        ggml_threadpool * threadpool) {
-    if (lctx.backend_cpu != nullptr) {
-        auto * reg = ggml_backend_dev_backend_reg(ggml_backend_get_device(lctx.backend_cpu));
-        auto * set_threadpool_fn = (decltype(ggml_backend_cpu_set_threadpool) *) ggml_backend_reg_get_proc_address(reg, "ggml_backend_cpu_set_threadpool");
-        set_threadpool_fn(lctx.backend_cpu, threadpool);
-    }
-
-    // set the number of threads for all the backends
-    for (const auto & set_n_threads_fn : lctx.set_n_threads_fns) {
-        set_n_threads_fn.second(set_n_threads_fn.first, n_threads);
-    }
-
-    auto status = ggml_backend_sched_graph_compute_async(lctx.sched.get(), gf);
-    if (status != GGML_STATUS_SUCCESS) {
-        LLAMA_LOG_ERROR("%s: ggml_backend_sched_graph_compute_async failed with error %d\n", __func__, status);
-    }
-
-    // fprintf(stderr, "splits: %d\n", ggml_backend_sched_get_n_splits(lctx.sched));
-
-    return status;
-}
-
-static int llama_prepare_sbatch(
-        llama_context     & lctx,
-        const llama_batch & batch,
-        uint32_t          & n_outputs) {
-    const auto & model   = lctx.model;
-    const auto & hparams = model.hparams;
-    const auto & cparams = lctx.cparams;
-
-    const uint32_t n_tokens_all = batch.n_tokens;
-    const  int64_t n_embd       = hparams.n_embd;
-
-    // this indicates we are doing pooled embedding, so we ignore batch.logits and output all tokens
-    const bool embd_pooled = cparams.embeddings && cparams.pooling_type != LLAMA_POOLING_TYPE_NONE;
-
-    GGML_ASSERT((!batch.token && batch.embd) || (batch.token && !batch.embd)); // NOLINT
-    if (batch.token) {
-        for (uint32_t i = 0; i < n_tokens_all; ++i) {
-            if (batch.token[i] < 0 || uint32_t(batch.token[i]) >= model.vocab.n_tokens()) {
-                LLAMA_LOG_ERROR("%s: invalid token[%d] = %d\n", __func__, i, batch.token[i]);
-                return -1;
-            }
-        }
-    }
-    GGML_ASSERT(n_tokens_all <= cparams.n_batch);
-    GGML_ASSERT((cparams.causal_attn || cparams.n_ubatch >= n_tokens_all) && "non-causal attention requires n_ubatch >= n_tokens");
-
-    lctx.n_queued_tokens += n_tokens_all;
-    lctx.embd_seq.clear();
-
-    // count outputs
-    if (batch.logits && !embd_pooled) {
-        for (uint32_t i = 0; i < n_tokens_all; ++i) {
-            n_outputs += batch.logits[i] != 0;
-        }
-    } else if (lctx.logits_all || embd_pooled) {
-        n_outputs = n_tokens_all;
-    } else {
-        // keep last output only
-        n_outputs = 1;
-    }
-
-    lctx.sbatch.from_batch(batch, n_embd,
-        /* simple_split */ !lctx.kv_self.recurrent,
-        /* logits_all   */ n_outputs == n_tokens_all);
-
-    // reserve output buffer
-    if (llama_output_reserve(lctx, n_outputs) < n_outputs) {
-        LLAMA_LOG_ERROR("%s: could not reserve space for batch with %u outputs\n", __func__, n_outputs);
-        return -2;
-    };
-
-    return 0;
-}
-
-static int llama_prepare_ubatch(
-        llama_context          & lctx,
-        llama_kv_slot_restorer & kv_slot_restorer,
-        llama_ubatch           & ubatch,
-        const uint32_t           n_outputs,
-        const uint32_t           n_tokens_all) {
-    GGML_ASSERT(lctx.sbatch.n_tokens > 0);
-
-    auto       & kv_self = lctx.kv_self;
-    const auto & cparams = lctx.cparams;
-    const auto & hparams = lctx.model.hparams;
-
-    // this indicates we are doing pooled embedding, so we ignore batch.logits and output all tokens
-    const bool embd_pooled = cparams.embeddings && cparams.pooling_type != LLAMA_POOLING_TYPE_NONE;
-
-    if (lctx.kv_self.recurrent) {
-        if (embd_pooled) {
-            // Pooled embeddings cannot be split across ubatches (yet)
-            ubatch = lctx.sbatch.split_seq(cparams.n_ubatch);
-        } else {
-            // recurrent model architectures are easier to implement
-            // with equal-length sequences
-            ubatch = lctx.sbatch.split_equal(cparams.n_ubatch);
-        }
-    } else {
-        ubatch = lctx.sbatch.split_simple(cparams.n_ubatch);
-    }
-
-    // count the outputs in this u_batch
-    {
-        int32_t n_outputs_new = 0;
-
-        if (n_outputs == n_tokens_all) {
-            n_outputs_new = ubatch.n_tokens;
-        } else {
-            GGML_ASSERT(ubatch.output);
-            for (uint32_t i = 0; i < ubatch.n_tokens; i++) {
-                n_outputs_new += int32_t(ubatch.output[i] != 0);
-            }
-        }
-
-        // needs to happen before the graph is built
-        lctx.n_outputs = n_outputs_new;
-    }
-
-    // non-causal masks do not use the KV cache
-    if (hparams.causal_attn) {
-        llama_kv_cache_update(&lctx);
-
-        // if we have enough unused cells before the current head ->
-        //   better to start searching from the beginning of the cache, hoping to fill it
-        if (kv_self.head > kv_self.used + 2*ubatch.n_tokens) {
-            kv_self.head = 0;
-        }
-
-        const auto slot = llama_kv_cache_find_slot(kv_self, ubatch);
-        if (!slot) {
-            return 1;
-        }
-        kv_slot_restorer.save(slot);
-
-        if (!kv_self.recurrent) {
-            // a heuristic, to avoid attending the full cache if it is not yet utilized
-            // after enough generations, the benefit from this heuristic disappears
-            // if we start defragmenting the cache, the benefit from this will be more important
-            const uint32_t pad = llama_kv_cache_get_padding(cparams);
-            kv_self.n = std::min(kv_self.size, std::max(pad, GGML_PAD(llama_kv_cache_cell_max(kv_self), pad)));
-            //kv_self.n = llama_kv_cache_cell_max(kv_self);
-        }
-    }
-
-    return 0;
-}
-
-// decode a batch of tokens by evaluating the transformer
-// in case of unsuccessful decoding (error or warning),
-// the kv_cache state will be returned to its original state
-// (for non-recurrent models) or cleaned (for recurrent models)
-//
-//   - lctx:      llama context
-//   - inp_batch: batch to evaluate
-//
-// return 0 on success
-// return positive int on warning
-// return negative int on error
-//
-static int llama_decode_impl(
-         llama_context & lctx,
-           llama_batch   inp_batch) {
-
-    lctx.is_encoding = false;
-
-    if (inp_batch.n_tokens == 0) {
-        LLAMA_LOG_ERROR("%s: n_tokens == 0\n", __func__);
-        return -1;
-    }
-
-    // temporarily allocate memory for the input batch if needed
-    llama_batch_allocr batch_allocr(inp_batch, inp_batch.pos ? -1 : lctx.kv_self.max_pos() + 1);
-    const llama_batch & batch = batch_allocr.batch;
-
-    const auto & model   = lctx.model;
-    const auto & vocab   = model.vocab;
-    const auto & hparams = model.hparams;
-    const auto & cparams = lctx.cparams;
-
-    if (lctx.t_compute_start_us == 0) {
-        lctx.t_compute_start_us = ggml_time_us();
-    }
-    auto & kv_self = lctx.kv_self;
-    llama_kv_slot_restorer kv_slot_restorer(kv_self);
-
-    const int64_t n_embd  = hparams.n_embd;
-    const int64_t n_vocab = vocab.n_tokens();
-
-    uint32_t n_outputs = 0;
-    uint32_t n_outputs_prev = 0;
-
-    {
-        const int ret = llama_prepare_sbatch(lctx, batch, n_outputs);
-        if (ret != 0) {
-            return ret;
-        }
-    }
-
-    while (lctx.sbatch.n_tokens > 0) {
-        llama_ubatch ubatch;
-        {
-            const int ret = llama_prepare_ubatch(lctx, kv_slot_restorer, ubatch, n_outputs, batch.n_tokens);
-            if (ret != 0) {
-                return ret;
-            }
-        }
-
-        const int         n_threads  = ubatch.n_tokens == 1 ? cparams.n_threads : cparams.n_threads_batch;
-        ggml_threadpool_t threadpool = ubatch.n_tokens == 1 ? lctx.threadpool   : lctx.threadpool_batch;
-
-        GGML_ASSERT(n_threads > 0);
-
-        //printf("kv_self.n = %5d, kv_self.used = %5d, kv_self.head = %5d\n", kv_self.n, kv_self.used, kv_self.head);
-
-        ggml_backend_sched_reset(lctx.sched.get());
-        ggml_backend_sched_set_eval_callback(lctx.sched.get(), lctx.cparams.cb_eval, lctx.cparams.cb_eval_user_data);
-
-        ggml_cgraph * gf = llama_build_graph(lctx, ubatch, false);
-
-        // the output is always the last tensor in the graph
-        struct ggml_tensor * res  = ggml_graph_node(gf, -1);
-        struct ggml_tensor * embd = ggml_graph_node(gf, -2);
-
-        if (lctx.n_outputs == 0) {
-            // no output
-            res  = nullptr;
-            embd = nullptr;
-        } else if (cparams.embeddings) {
-            res  = nullptr; // do not extract logits for embedding case
-            embd = nullptr;
-            for (int i = ggml_graph_n_nodes(gf) - 1; i >= 0; --i) {
-                if (strcmp(ggml_graph_node(gf, i)->name, "result_embd_pooled") == 0) {
-                    embd = ggml_graph_node(gf, i);
-                    break;
-                }
-            }
-            GGML_ASSERT(embd != nullptr && "missing embeddings tensor");
-        } else {
-            embd = nullptr; // do not extract embeddings when not needed
-            GGML_ASSERT(strcmp(res->name, "result_output") == 0 && "missing result_output tensor");
-        }
-
-        // LLAMA_LOG_INFO("graph build time: %.3f ms (%d nodes, %d leafs)\n", (ggml_time_us() - t_start_us)/1000.0, gf->n_nodes, gf->n_leafs);
-
-        ggml_backend_sched_alloc_graph(lctx.sched.get(), gf);
-
-        llama_set_inputs(lctx, ubatch);
-
-        const auto compute_status = llama_graph_compute(lctx, gf, n_threads, threadpool);
-        if (compute_status != GGML_STATUS_SUCCESS) {
-            kv_slot_restorer.restore(kv_self);
-            switch (compute_status) {
-                case GGML_STATUS_ABORTED:
-                    return 2;
-                case GGML_STATUS_ALLOC_FAILED:
-                    return -2;
-                case GGML_STATUS_FAILED:
-                default:
-                    return -3;
-            }
-        }
-
-        // update the kv ring buffer
-        {
-            kv_self.head += ubatch.n_tokens;
-
-            // Ensure kv cache head points to a valid index.
-            if (kv_self.head >= kv_self.size) {
-                kv_self.head = 0;
-            }
-        }
-
-        // plot the computation graph in dot format (for debugging purposes)
-        //if (n_past%100 == 0) {
-        //    ggml_graph_dump_dot(gf, NULL, "llama.dot");
-        //}
-
-        // extract logits
-        if (res) {
-            ggml_backend_t backend_res = ggml_backend_sched_get_tensor_backend(lctx.sched.get(), res);
-            GGML_ASSERT(backend_res != nullptr);
-            GGML_ASSERT(lctx.logits != nullptr);
-
-            float * logits_out = lctx.logits + n_outputs_prev*n_vocab;
-            const int32_t n_outputs_new = lctx.n_outputs;
-
-            if (n_outputs_new) {
-                GGML_ASSERT( n_outputs_prev + n_outputs_new <= n_outputs);
-                GGML_ASSERT((n_outputs_prev + n_outputs_new)*n_vocab <= (int64_t) lctx.logits_size);
-                ggml_backend_tensor_get_async(backend_res, res, logits_out, 0, n_outputs_new*n_vocab*sizeof(float));
-            }
-        }
-
-        // extract embeddings
-        if (embd) {
-            ggml_backend_t backend_embd = ggml_backend_sched_get_tensor_backend(lctx.sched.get(), embd);
-            GGML_ASSERT(backend_embd != nullptr);
-
-            switch (cparams.pooling_type) {
-                case LLAMA_POOLING_TYPE_NONE:
-                    {
-                        // extract token embeddings
-                        GGML_ASSERT(lctx.embd != nullptr);
-                        float * embd_out = lctx.embd + n_outputs_prev*n_embd;
-                        const int32_t n_outputs_new = lctx.n_outputs;
-
-                        if (n_outputs_new) {
-                            GGML_ASSERT( n_outputs_prev + n_outputs_new <= n_outputs);
-                            GGML_ASSERT((n_outputs_prev + n_outputs_new)*n_embd <= (int64_t) lctx.embd_size);
-                            ggml_backend_tensor_get_async(backend_embd, embd, embd_out, 0, n_outputs_new*n_embd*sizeof(float));
-                        }
-                    } break;
-                case LLAMA_POOLING_TYPE_MEAN:
-                case LLAMA_POOLING_TYPE_CLS:
-                case LLAMA_POOLING_TYPE_LAST:
-                    {
-                        // extract sequence embeddings (cleared before processing each batch)
-                        auto & embd_seq_out = lctx.embd_seq;
-
-                        for (uint32_t s = 0; s < ubatch.n_seqs; ++s) {
-                            const llama_seq_id seq_id = ubatch.seq_id[s][0];
-                            if (embd_seq_out.find(seq_id) != embd_seq_out.end()) {
-                                continue;
-                            }
-                            embd_seq_out[seq_id].resize(n_embd);
-                            ggml_backend_tensor_get_async(backend_embd, embd, embd_seq_out[seq_id].data(), (n_embd*seq_id)*sizeof(float), n_embd*sizeof(float));
-                        }
-                    } break;
-                case LLAMA_POOLING_TYPE_RANK:
-                    {
-                        // extract the rerank score - a single float per sequence
-                        auto & embd_seq_out = lctx.embd_seq;
-
-                        for (uint32_t s = 0; s < ubatch.n_seqs; ++s) {
-                            const llama_seq_id seq_id = ubatch.seq_id[s][0];
-                            if (embd_seq_out.find(seq_id) != embd_seq_out.end()) {
-                                continue;
-                            }
-                            embd_seq_out[seq_id].resize(1);
-                            ggml_backend_tensor_get_async(backend_embd, embd, embd_seq_out[seq_id].data(), (seq_id)*sizeof(float), sizeof(float));
-                        }
-                    } break;
-                case LLAMA_POOLING_TYPE_UNSPECIFIED:
-                    {
-                        GGML_ABORT("unknown pooling type");
-                    }
-            }
-        }
-        n_outputs_prev += lctx.n_outputs;
-    }
-
-    // set output mappings
-    {
-        bool sorted_output = true;
-
-        GGML_ASSERT(lctx.sbatch.out_ids.size() == n_outputs);
-
-        for (size_t i = 0; i < n_outputs; ++i) {
-            size_t out_id = lctx.sbatch.out_ids[i];
-            lctx.output_ids[out_id] = i;
-            if (out_id != i) {
-                sorted_output = false;
-            }
-        }
-
-        if (sorted_output) {
-            lctx.sbatch.out_ids.clear();
-        }
-    }
-
-    // set to total number of outputs in the batch, for use in llama_get_logits_ith
-    lctx.n_outputs = n_outputs;
-
-    // wait for the computation to finish (automatically done when obtaining the model output)
-    //llama_synchronize(&lctx);
-
-    // decide if we need to defrag the kv cache
-    if (cparams.causal_attn && cparams.defrag_thold > 0.0f) {
-        // - do not defrag small contexts (i.e. < 2048 tokens)
-        // - count the padding towards the number of used tokens
-        const float fragmentation = kv_self.n >= 2048 ? std::max(0.0f, 1.0f - float(kv_self.used + llama_kv_cache_get_padding(cparams))/float(kv_self.n)) : 0.0f;
-
-        // queue defragmentation for next llama_kv_cache_update
-        if (fragmentation > cparams.defrag_thold) {
-            LLAMA_LOG_DEBUG("%s: fragmentation: %.2f - requesting defrag\n", __func__, fragmentation);
-
-            llama_kv_cache_defrag(kv_self);
-        }
-    }
-
-    // Reset state for the next token before backend sync, to allow the CPU activities in the reset to
-    // overlap with device computation.
-    ggml_backend_sched_reset(lctx.sched.get());
-
-    return 0;
-}
-
-// encode a batch of tokens by evaluating the encoder part of the transformer
-//
-//   - lctx:      llama context
-//   - batch:     batch to evaluate
-//
-// return 0 on success
-// return positive int on warning
-// return negative int on error
-//
-static int llama_encode_impl(
-         llama_context & lctx,
-           llama_batch   inp_batch) {
-
-    lctx.is_encoding = true;
-
-    if (inp_batch.n_tokens == 0) {
-        LLAMA_LOG_ERROR("%s: n_tokens == 0\n", __func__);
-        return -1;
-    }
-
-    // temporary allocate memory for the input batch if needed
-    llama_batch_allocr batch_allocr(inp_batch, inp_batch.pos ? -1 : lctx.kv_self.max_pos() + 1);
-
-    const llama_batch & batch = batch_allocr.batch;
-    const uint32_t n_tokens = batch.n_tokens;
-
-    const auto & model   = lctx.model;
-    const auto & hparams = model.hparams;
-    const auto & cparams = lctx.cparams;
-
-    GGML_ASSERT((!batch.token && batch.embd) || (batch.token && !batch.embd)); // NOLINT
-
-    if (batch.token) {
-        for (uint32_t i = 0; i < n_tokens; ++i) {
-            if (batch.token[i] < 0 || (uint32_t) batch.token[i] >= model.vocab.n_tokens()) {
-                LLAMA_LOG_ERROR("%s: invalid token[%d] = %d\n", __func__, i, batch.token[i]);
-                return -1;
-            }
-        }
-    }
-
-    // micro-batching is not possible for non-causal encoding, so we process the batch in a single shot
-    GGML_ASSERT(cparams.n_ubatch >= n_tokens && "encoder requires n_ubatch >= n_tokens");
-
-    if (lctx.t_compute_start_us == 0) {
-        lctx.t_compute_start_us = ggml_time_us();
-    }
-
-    lctx.n_queued_tokens += n_tokens;
-
-    const int64_t n_embd = hparams.n_embd;
-
-    lctx.sbatch.from_batch(batch, n_embd, /* simple_split */ true, /* logits_all */ true);
-
-    const llama_ubatch ubatch = lctx.sbatch.split_simple(n_tokens);
-
-    // reserve output buffer
-    if (llama_output_reserve(lctx, n_tokens) < n_tokens) {
-        LLAMA_LOG_ERROR("%s: could not reserve space for batch with %u outputs\n", __func__, n_tokens);
-        return -2;
-    };
-
-    for (uint32_t i = 0; i < n_tokens; ++i) {
-        lctx.output_ids[i] = i;
-    }
-
-    lctx.inp_embd_enc = NULL;
-    lctx.n_outputs = n_tokens;
-
-    int n_threads = n_tokens == 1 ? cparams.n_threads : cparams.n_threads_batch;
-    ggml_threadpool_t threadpool = n_tokens == 1 ? lctx.threadpool : lctx.threadpool_batch;
-
-    GGML_ASSERT(n_threads > 0);
-
-    ggml_backend_sched_reset(lctx.sched.get());
-    ggml_backend_sched_set_eval_callback(lctx.sched.get(), lctx.cparams.cb_eval, lctx.cparams.cb_eval_user_data);
-
-    ggml_cgraph * gf = llama_build_graph(lctx, ubatch, false);
-
-    // the output embeddings after the final encoder normalization
-    struct ggml_tensor * embd = nullptr;
-
-    // there are two cases here
-    if (llama_model_has_decoder(&lctx.model)) {
-        // first case is an encoder-decoder T5 model where embeddings are passed to decoder
-        embd = ggml_graph_node(gf, -1);
-        GGML_ASSERT(strcmp(embd->name, "result_norm") == 0 && "missing result_output tensor");
-    } else {
-        // second case is an encoder-only T5 model
-        if (cparams.embeddings) {
-            // only output embeddings if required
-            embd = ggml_graph_node(gf, -1);
-            if (strcmp(embd->name, "result_embd_pooled") != 0) {
-                embd = ggml_graph_node(gf, -2);
-            }
-            GGML_ASSERT(strcmp(embd->name, "result_embd_pooled") == 0 && "missing embeddings tensor");
-        }
-    }
-
-    ggml_backend_sched_alloc_graph(lctx.sched.get(), gf);
-
-    llama_set_inputs(lctx, ubatch);
-
-    const auto compute_status = llama_graph_compute(lctx, gf, n_threads, threadpool);
-    switch (compute_status) {
-        case GGML_STATUS_SUCCESS:
-            break;
-        case GGML_STATUS_ABORTED:
-            return 2;
-        case GGML_STATUS_ALLOC_FAILED:
-            return -2;
-        case GGML_STATUS_FAILED:
-        default:
-            return -3;
-    }
-
-    // extract embeddings
-    if (embd) {
-        ggml_backend_t backend_embd = ggml_backend_sched_get_tensor_backend(lctx.sched.get(), embd);
-        GGML_ASSERT(backend_embd != nullptr);
-
-        if (llama_model_has_decoder(&lctx.model)) {
-            lctx.embd_enc.resize(n_tokens*n_embd);
-            float * embd_out = lctx.embd_enc.data();
-
-            ggml_backend_tensor_get_async(backend_embd, embd, embd_out, 0, n_tokens*n_embd*sizeof(float));
-            GGML_ASSERT(!ubatch.equal_seqs); // TODO: handle equal splits
-
-            // remember the sequence ids used during the encoding - needed for cross attention later
-            lctx.seq_ids_enc.resize(n_tokens);
-            for (uint32_t i = 0; i < n_tokens; i++) {
-                for (int s = 0; s < ubatch.n_seq_id[i]; s++) {
-                    llama_seq_id seq_id = ubatch.seq_id[i][s];
-                    lctx.seq_ids_enc[i].insert(seq_id);
-                }
-            }
-        } else {
-            GGML_ASSERT(lctx.embd != nullptr);
-
-            switch (cparams.pooling_type) {
-                case LLAMA_POOLING_TYPE_NONE:
-                    {
-                        // extract token embeddings
-                        GGML_ASSERT(lctx.embd != nullptr);
-                        float * embd_out = lctx.embd;
-
-                        GGML_ASSERT(n_tokens*n_embd <= (int64_t) lctx.embd_size);
-                        ggml_backend_tensor_get_async(backend_embd, embd, embd_out, 0, n_tokens*n_embd*sizeof(float));
-                    } break;
-                case LLAMA_POOLING_TYPE_MEAN:
-                case LLAMA_POOLING_TYPE_CLS:
-                case LLAMA_POOLING_TYPE_LAST:
-                    {
-                        // extract sequence embeddings
-                        auto & embd_seq_out = lctx.embd_seq;
-                        embd_seq_out.clear();
-
-                        GGML_ASSERT(!ubatch.equal_seqs); // TODO: handle equal splits
-
-                        for (uint32_t i = 0; i < n_tokens; i++) {
-                            const llama_seq_id seq_id = ubatch.seq_id[i][0];
-                            if (embd_seq_out.find(seq_id) != embd_seq_out.end()) {
-                                continue;
-                            }
-                            embd_seq_out[seq_id].resize(n_embd);
-                            ggml_backend_tensor_get_async(backend_embd, embd, embd_seq_out[seq_id].data(), (n_embd*seq_id)*sizeof(float), n_embd*sizeof(float));
-                        }
-                    } break;
-                case LLAMA_POOLING_TYPE_RANK:
-                    {
-                        // TODO: this likely should be the same logic as in llama_decoder_internal, but better to
-                        //       wait for an encoder model that requires this pooling type in order to test it
-                        //       https://github.com/ggerganov/llama.cpp/pull/9510
-                        GGML_ABORT("RANK pooling not implemented yet");
-                    }
-                case LLAMA_POOLING_TYPE_UNSPECIFIED:
-                    {
-                        GGML_ABORT("unknown pooling type");
-                    }
-            }
-        }
-    }
-
-    // Reset state for the next token before backend sync, to allow the CPU activities in the reset to
-    // overlap with device computation.
-    ggml_backend_sched_reset(lctx.sched.get());
-
-    return 0;
-}
-
-// find holes from the beginning of the KV cache and fill them by moving data from the end of the cache
-static void llama_kv_cache_defrag_impl(struct llama_context & lctx) {
-    auto & kv_self = lctx.kv_self;
-
-    const auto & hparams = lctx.model.hparams;
-
-    const uint32_t n_layer = hparams.n_layer;
-
-    const uint32_t n_kv   = llama_kv_cache_cell_max(kv_self);
-    const uint32_t n_used = kv_self.used;
-
-    assert(n_used <= n_kv);
-
-    //const int64_t t_start = ggml_time_us();
-
-    // number of cells moved
-    uint32_t n_moves = 0;
-
-    // each move requires 6*n_layer tensors (see build_defrag)
-    //   - source view, destination view, copy operation
-    //   - x2 for keys and values
-    //const uint32_t max_moves = model.max_nodes()/(6*n_layer);
-    // TODO: tmp fix https://github.com/ggerganov/llama.cpp/issues/6685#issuecomment-2057579516
-    const uint32_t max_moves = (lctx.model.max_nodes() - 2*n_layer)/(6*n_layer);
-
-    // determine which KV cells to move where
-    //
-    //  cell i moves to ids[i]
-    //
-    //  if ids[i] == i || ids[i] == n_kv, then cell i is not moved
-    //
-    std::vector ids(n_kv, n_kv);
-
-    for (uint32_t i0 = 0; i0 < n_used; ++i0) {
-        const auto & cell0 = kv_self.cells[i0];
-
-        if (!cell0.is_empty()) {
-            ids[i0] = i0;
-
-            continue;
-        }
-
-        // found a hole - fill it with data from the end of the cache
-
-        uint32_t nh = 1;
-
-        // determine the size of the hole
-        while (i0 + nh < n_used && kv_self.cells[i0 + nh].is_empty()) {
-            nh++;
-        }
-
-        uint32_t nf = 0;
-        uint32_t is = n_kv - 1;
-
-        // starting from the end, find nh non-empty cells
-        for (; is > i0; --is) {
-            const auto & cell1 = kv_self.cells[is];
-
-            if (cell1.is_empty() || ids[is] != n_kv) {
-                continue;
-            }
-
-            // non-empty cell which is not yet moved
-            nf++;
-
-            if (nf == nh) {
-                break;
-            }
-        }
-
-        // this can only happen if `n_used` is not accurate, which would be a bug
-        GGML_ASSERT(nf == nh && "KV defrag bug: nf != nh");
-
-        nf = 0;
-
-        uint32_t i1 = is;
-
-        // are we moving a continuous block of memory?
-        bool cont = false;
-
-        // should we stop searching for the next move?
-        bool stop = false;
-
-        // go back and move the nf cells to the hole
-        for (; i1 < n_kv; ++i1) {
-            auto & cell1 = kv_self.cells[i1];
-
-            if (cell1.is_empty() || ids[i1] != n_kv) {
-                if (n_moves == max_moves) {
-                    stop = true;
-                    break;
-                }
-
-                cont = false;
-                continue;
-            }
-
-            // this cell goes to (i0 + nf)
-            ids[i1] = i0 + nf;
-
-            // move the cell meta data
-            kv_self.cells[i0 + nf] = cell1;
-
-            // clear the old cell and move the head there
-            cell1 = llama_kv_cell();
-            kv_self.head = n_used;
-
-            if (!cont) {
-                n_moves++;
-                cont = true;
-            }
-
-            nf++;
-
-            if (nf == nh) {
-                break;
-            }
-        }
-
-        if (stop || n_moves == max_moves) {
-            break;
-        }
-
-        //LLAMA_LOG_INFO("(tmp log) KV defrag: move [%u, %u) to [%u, %u)\n", is, i1 + 1, i0, i0 + nh);
-
-        i0 += nh - 1;
-    }
-
-    if (n_moves == 0) {
-        return;
-    }
-
-    //LLAMA_LOG_INFO("(tmp log) KV defrag cell moves: %u\n", n_moves);
-
-    //LLAMA_LOG_INFO("expected gf nodes: %u\n", 6*n_moves*n_layer);
-
-#if 0
-    // CPU defrag
-    //
-    // TODO: optimizations are possible:
-    //       - multiple threads
-    //       - avoid copying to the host memory when already there
-    //
-    // likely not worth the effort, as we have ggml_graph based defrag
-    //
-
-    const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa();
-    const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa();
-
-    const uint32_t kv_size = kv_self.size;
-
-    std::vector buf_k;
-    std::vector buf_v;
-
-    for (uint32_t il = 0; il < n_layer; ++il) {
-        const size_t k_size_row = ggml_row_size(kv_self.k_l[il]->type, n_embd_k_gqa);
-        const size_t k_size     = ggml_row_size(kv_self.k_l[il]->type, n_embd_k_gqa*kv_size);
-
-        const size_t v_size_el = ggml_type_size(kv_self.v_l[il]->type);
-        const size_t v_size    = ggml_row_size (kv_self.v_l[il]->type, n_embd_v_gqa*kv_size);
-
-        buf_k.resize(k_size);
-        buf_v.resize(v_size);
-
-        ggml_backend_tensor_get(kv_self.k_l[il], buf_k.data(), 0, buf_k.size());
-        ggml_backend_tensor_get(kv_self.v_l[il], buf_v.data(), 0, buf_v.size());
-
-        // batch move [i, i+nm) to [id, id+nm)
-        // note: cells can move only to a lower index
-        for (uint32_t i = 0; i < n_kv; ++i) {
-            const uint32_t id = ids[i];
-
-            if (i == id || id == n_kv) {
-                continue;
-            }
-
-            uint32_t nm = 1;
-
-            while (i + nm < n_kv && ids[i + nm] == id + nm) {
-                nm++;
-            }
-
-            // move keys
-            {
-                const int64_t os =  i*k_size_row;
-                const int64_t od = id*k_size_row;
-
-                memcpy(buf_k.data() + od, buf_k.data() + os, nm*k_size_row);
-            }
-
-            // move values (note: they are transposed)
-            {
-                const int64_t os =  i;
-                const int64_t od = id;
-
-                for (uint32_t j = 0; j < n_embd_v_gqa; ++j) {
-                    memcpy(buf_v.data() + (od + j*kv_size)*v_size_el, buf_v.data() + (os + j*kv_size)*v_size_el, nm*v_size_el);
-                }
-            }
-
-            i += nm - 1;
-        }
-
-        ggml_backend_tensor_set(kv_self.k_l[il], buf_k.data(), 0, buf_k.size());
-        ggml_backend_tensor_set(kv_self.v_l[il], buf_v.data(), 0, buf_v.size());
-    }
-#else
-    // ggml_graph defrag
-
-    ggml_backend_sched_reset(lctx.sched.get());
-
-    ggml_cgraph * gf = llama_build_graph_defrag(lctx, ids);
-
-    llama_graph_compute(lctx, gf, lctx.cparams.n_threads, lctx.threadpool);
-#endif
-
-    //const int64_t t_end = ggml_time_us();
-
-    //LLAMA_LOG_INFO("(tmp log) KV defrag time: %.3f ms\n", (t_end - t_start)/1000.0);
-}
-
-static void llama_kv_cache_update_impl(struct llama_context & lctx) {
-    bool need_reserve = false;
-
-    if (lctx.kv_self.has_shift) {
-        if (!llama_kv_cache_can_shift(&lctx)) {
-            GGML_ABORT("The current context does not support K-shift");
-        }
-
-        // apply K-shift if needed
-        if (lctx.model.hparams.rope_type != LLAMA_ROPE_TYPE_NONE) {
-            ggml_backend_sched_reset(lctx.sched.get());
-
-            ggml_cgraph * gf = llama_build_graph_k_shift(lctx);
-
-            ggml_backend_sched_alloc_graph(lctx.sched.get(), gf);
-
-            llama_set_k_shift(lctx);
-
-            llama_graph_compute(lctx, gf, lctx.cparams.n_threads, lctx.threadpool);
-
-            need_reserve = true;
-        }
-
-        {
-            auto & kv_self = lctx.kv_self;
-
-            kv_self.has_shift = false;
-
-            for (uint32_t i = 0; i < kv_self.size; ++i) {
-                kv_self.cells[i].delta = 0;
-            }
-        }
-    }
-
-    // defragment the KV cache if needed
-    if (lctx.kv_self.do_defrag) {
-        llama_kv_cache_defrag_impl(lctx);
-
-        need_reserve = true;
-
-        lctx.kv_self.do_defrag = false;
-    }
-
-    // reserve a worst case graph again
-    if (need_reserve) {
-        // TODO: extract to a function
-        // build worst-case graph
-        uint32_t n_seqs = 1; // TODO: worst-case number of sequences
-        uint32_t n_tokens = std::min(lctx.cparams.n_ctx, lctx.cparams.n_ubatch);
-        llama_token token = lctx.model.vocab.token_bos(); // not actually used by llama_build_graph, but required to choose between token and embedding inputs graph
-        llama_ubatch ubatch = { true, n_tokens, n_tokens / n_seqs, n_seqs, &token, nullptr, nullptr, nullptr, nullptr, nullptr};
-        ggml_cgraph * gf = llama_build_graph(lctx, ubatch, true);
-
-        // initialize scheduler with the worst-case graph
-        ggml_backend_sched_reset(lctx.sched.get());
-        if (!ggml_backend_sched_reserve(lctx.sched.get(), gf)) {
-            LLAMA_LOG_ERROR("%s: failed to allocate compute buffers\n", __func__);
-        }
-    }
-}
-
-int32_t llama_set_adapter_lora(
-            struct llama_context * ctx,
-            struct llama_adapter_lora * adapter,
-            float scale) {
-    ctx->lora[adapter] = scale;
-    return 0;
-}
-
-int32_t llama_rm_adapter_lora(
-            struct llama_context * ctx,
-            struct llama_adapter_lora * adapter) {
-    auto pos = ctx->lora.find(adapter);
-    if (pos != ctx->lora.end()) {
-        ctx->lora.erase(pos);
-        return 0;
-    }
-
-    return -1;
-}
-
-void llama_clear_adapter_lora(struct llama_context * ctx) {
-    ctx->lora.clear();
-}
-
-int32_t llama_apply_adapter_cvec(
-        struct llama_context * ctx,
-                 const float * data,
-                      size_t   len,
-                     int32_t   n_embd,
-                     int32_t   il_start,
-                     int32_t   il_end) {
-    return ctx->cvec.apply(ctx->model, data, len, n_embd, il_start, il_end);
-}
-
 //
 // interface implementation
 //
 
-struct llama_context_params llama_context_default_params() {
-    struct llama_context_params result = {
-        /*.n_ctx                       =*/ 512,
-        /*.n_batch                     =*/ 2048,
-        /*.n_ubatch                    =*/ 512,
-        /*.n_seq_max                   =*/ 1,
-        /*.n_threads                   =*/ GGML_DEFAULT_N_THREADS, // TODO: better default
-        /*.n_threads_batch             =*/ GGML_DEFAULT_N_THREADS,
-        /*.rope_scaling_type           =*/ LLAMA_ROPE_SCALING_TYPE_UNSPECIFIED,
-        /*.pooling_type                =*/ LLAMA_POOLING_TYPE_UNSPECIFIED,
-        /*.attention_type              =*/ LLAMA_ATTENTION_TYPE_UNSPECIFIED,
-        /*.rope_freq_base              =*/ 0.0f,
-        /*.rope_freq_scale             =*/ 0.0f,
-        /*.yarn_ext_factor             =*/ -1.0f,
-        /*.yarn_attn_factor            =*/ 1.0f,
-        /*.yarn_beta_fast              =*/ 32.0f,
-        /*.yarn_beta_slow              =*/ 1.0f,
-        /*.yarn_orig_ctx               =*/ 0,
-        /*.defrag_thold                =*/ -1.0f,
-        /*.cb_eval                     =*/ nullptr,
-        /*.cb_eval_user_data           =*/ nullptr,
-        /*.type_k                      =*/ GGML_TYPE_F16,
-        /*.type_v                      =*/ GGML_TYPE_F16,
-        /*.logits_all                  =*/ false,
-        /*.embeddings                  =*/ false,
-        /*.offload_kqv                 =*/ true,
-        /*.flash_attn                  =*/ false,
-        /*.no_perf                     =*/ true,
-        /*.abort_callback              =*/ nullptr,
-        /*.abort_callback_data         =*/ nullptr,
-    };
-
-    return result;
-}
-
 struct llama_sampler_chain_params llama_sampler_chain_default_params() {
     struct llama_sampler_chain_params result = {
         /*.no_perf                     =*/ true,
@@ -9424,6 +82,57 @@ int64_t llama_time_us(void) {
     return ggml_time_us();
 }
 
+// Returns 0 on success, -1 on error, and -2 on cancellation via llama_progress_callback
+static int llama_model_load(const std::string & fname, std::vector & splits, llama_model & model, llama_model_params & params) {
+    // loading time will be recalculated after the first eval, so
+    // we take page faults deferred by mmap() into consideration
+    model.t_load_us = 0;
+    time_meas tm(model.t_load_us);
+
+    model.t_start_us = tm.t_start_us;
+
+    try {
+        llama_model_loader ml(fname, splits, params.use_mmap, params.check_tensors, params.kv_overrides, params.tensor_buft_overrides);
+
+        ml.print_info();
+
+        model.hparams.vocab_only = params.vocab_only;
+
+        try {
+            model.load_arch(ml);
+        } catch(const std::exception & e) {
+            throw std::runtime_error("error loading model architecture: " + std::string(e.what()));
+        }
+        try {
+            model.load_hparams(ml);
+        } catch(const std::exception & e) {
+            throw std::runtime_error("error loading model hyperparameters: " + std::string(e.what()));
+        }
+        try {
+            model.load_vocab(ml);
+        } catch(const std::exception & e) {
+            throw std::runtime_error("error loading model vocabulary: " + std::string(e.what()));
+        }
+
+        model.load_stats(ml);
+        model.print_info();
+
+        if (params.vocab_only) {
+            LLAMA_LOG_INFO("%s: vocab only - skipping tensors\n", __func__);
+            return 0;
+        }
+
+        if (!model.load_tensors(ml)) {
+            return -2;
+        }
+    } catch (const std::exception & err) {
+        LLAMA_LOG_ERROR("%s: error loading model: %s\n", __func__, err.what());
+        return -1;
+    }
+
+    return 0;
+}
+
 static struct llama_model * llama_model_load_from_file_impl(
         const std::string & path_model,
         std::vector & splits,
@@ -9544,460 +253,6 @@ struct llama_model * llama_model_load_from_splits(
     return llama_model_load_from_file_impl(splits.front(), splits, params);
 }
 
-struct llama_context * llama_init_from_model(
-                 struct llama_model * model,
-        struct llama_context_params   params) {
-
-    if (!model) {
-        LLAMA_LOG_ERROR("%s: model cannot be NULL\n", __func__);
-        return nullptr;
-    }
-
-    if (params.n_batch == 0 && params.n_ubatch == 0) {
-        LLAMA_LOG_ERROR("%s: n_batch and n_ubatch cannot both be zero\n", __func__);
-        return nullptr;
-    }
-
-    if (params.n_ctx == 0 && model->hparams.n_ctx_train == 0) {
-        LLAMA_LOG_ERROR("%s: n_ctx and model->hparams.n_ctx_train cannot both be zero\n", __func__);
-        return nullptr;
-    }
-
-    if (params.flash_attn && model->arch == LLM_ARCH_GROK) {
-        LLAMA_LOG_WARN("%s: flash_attn is not compatible with Grok - forcing off\n", __func__);
-        params.flash_attn = false;
-    }
-
-    if (params.flash_attn && model->hparams.n_embd_head_k != model->hparams.n_embd_head_v) {
-        LLAMA_LOG_WARN("%s: flash_attn requires n_embd_head_k == n_embd_head_v - forcing off\n", __func__);
-        params.flash_attn = false;
-    }
-
-    if (ggml_is_quantized(params.type_v) && !params.flash_attn) {
-        LLAMA_LOG_ERROR("%s: V cache quantization requires flash_attn\n", __func__);
-        return nullptr;
-    }
-
-    llama_context * ctx = new llama_context(*model);
-
-    const auto & hparams = model->hparams;
-    auto       & cparams = ctx->cparams;
-
-    cparams.n_seq_max        = std::max(1u, params.n_seq_max);
-    cparams.n_threads        = params.n_threads;
-    cparams.n_threads_batch  = params.n_threads_batch;
-    cparams.yarn_ext_factor  = params.yarn_ext_factor;
-    cparams.yarn_attn_factor = params.yarn_attn_factor;
-    cparams.yarn_beta_fast   = params.yarn_beta_fast;
-    cparams.yarn_beta_slow   = params.yarn_beta_slow;
-    cparams.defrag_thold     = params.defrag_thold;
-    cparams.embeddings       = params.embeddings;
-    cparams.offload_kqv      = params.offload_kqv;
-    cparams.flash_attn       = params.flash_attn;
-    cparams.no_perf          = params.no_perf;
-    cparams.pooling_type     = params.pooling_type;
-
-    cparams.n_ctx            = params.n_ctx           == 0    ? hparams.n_ctx_train           : params.n_ctx;
-    cparams.rope_freq_base   = params.rope_freq_base  == 0.0f ? hparams.rope_freq_base_train  : params.rope_freq_base;
-    cparams.rope_freq_scale  = params.rope_freq_scale == 0.0f ? hparams.rope_freq_scale_train : params.rope_freq_scale;
-
-    // this is necessary due to kv_self.n being padded later during inference
-    cparams.n_ctx            = GGML_PAD(cparams.n_ctx, llama_kv_cache_get_padding(cparams));
-
-    // with causal attention, the batch size is limited by the context size
-    cparams.n_batch          = hparams.causal_attn ? std::min(cparams.n_ctx, params.n_batch) : params.n_batch;
-
-    // the batch has to be at least GGML_KQ_MASK_PAD because we will be padding the KQ_mask
-    // this is required by GPU kernels in order to avoid out-of-bounds accesses (e.g. ggml_flash_attn_ext)
-    // ref: https://github.com/ggerganov/llama.cpp/pull/5021
-    if (cparams.n_batch < GGML_KQ_MASK_PAD) {
-        LLAMA_LOG_WARN("%s: n_batch is less than GGML_KQ_MASK_PAD - increasing to %d\n", __func__, GGML_KQ_MASK_PAD);
-        cparams.n_batch = GGML_KQ_MASK_PAD;
-    }
-
-    cparams.n_ubatch         = std::min(cparams.n_batch, params.n_ubatch == 0 ? params.n_batch : params.n_ubatch);
-
-    cparams.n_ctx_orig_yarn  = params.yarn_orig_ctx    != 0 ? params.yarn_orig_ctx    :
-                               hparams.n_ctx_orig_yarn != 0 ? hparams.n_ctx_orig_yarn :
-                                                              hparams.n_ctx_train;
-
-    cparams.cb_eval           = params.cb_eval;
-    cparams.cb_eval_user_data = params.cb_eval_user_data;
-
-    auto rope_scaling_type = params.rope_scaling_type;
-    if (rope_scaling_type == LLAMA_ROPE_SCALING_TYPE_UNSPECIFIED) {
-        rope_scaling_type = hparams.rope_scaling_type_train;
-    }
-
-    if (rope_scaling_type == LLAMA_ROPE_SCALING_TYPE_NONE) {
-        cparams.rope_freq_scale = 1.0f; // never scale if scaling type is none
-    }
-
-    if (cparams.yarn_ext_factor < 0.0f) { // negative indicates 'not set'
-        cparams.yarn_ext_factor = rope_scaling_type == LLAMA_ROPE_SCALING_TYPE_YARN ? 1.0f : 0.0f;
-    }
-
-    cparams.yarn_attn_factor *= hparams.rope_attn_factor;
-
-    if (cparams.pooling_type == LLAMA_POOLING_TYPE_UNSPECIFIED) {
-        if (hparams.pooling_type == LLAMA_POOLING_TYPE_UNSPECIFIED) {
-            cparams.pooling_type = LLAMA_POOLING_TYPE_NONE;
-        } else {
-            cparams.pooling_type = hparams.pooling_type;
-        }
-    }
-
-    if (params.attention_type == LLAMA_ATTENTION_TYPE_UNSPECIFIED) {
-        cparams.causal_attn = hparams.causal_attn;
-    } else {
-        cparams.causal_attn = params.attention_type == LLAMA_ATTENTION_TYPE_CAUSAL;
-    }
-
-    const uint32_t n_ctx_per_seq = cparams.n_ctx / cparams.n_seq_max;
-
-    LLAMA_LOG_INFO("%s: n_seq_max     = %u\n",   __func__, cparams.n_seq_max);
-    LLAMA_LOG_INFO("%s: n_ctx         = %u\n",   __func__, cparams.n_ctx);
-    LLAMA_LOG_INFO("%s: n_ctx_per_seq = %u\n",   __func__, n_ctx_per_seq);
-    LLAMA_LOG_INFO("%s: n_batch       = %u\n",   __func__, cparams.n_batch);
-    LLAMA_LOG_INFO("%s: n_ubatch      = %u\n",   __func__, cparams.n_ubatch);
-    LLAMA_LOG_INFO("%s: flash_attn    = %d\n",   __func__, cparams.flash_attn);
-    LLAMA_LOG_INFO("%s: freq_base     = %.1f\n", __func__, cparams.rope_freq_base);
-    LLAMA_LOG_INFO("%s: freq_scale    = %g\n",   __func__, cparams.rope_freq_scale);
-
-    if (n_ctx_per_seq < hparams.n_ctx_train) {
-        LLAMA_LOG_WARN("%s: n_ctx_per_seq (%u) < n_ctx_train (%u) -- the full capacity of the model will not be utilized\n",
-                __func__, n_ctx_per_seq, hparams.n_ctx_train);
-    }
-
-    if (n_ctx_per_seq > hparams.n_ctx_train) {
-        LLAMA_LOG_WARN("%s: n_ctx_pre_seq (%u) > n_ctx_train (%u) -- possible training context overflow\n",
-                __func__, n_ctx_per_seq, hparams.n_ctx_train);
-    }
-
-    ctx->logits_all = params.logits_all;
-
-    // build worst-case graph for encoder if a model contains encoder
-    ctx->is_encoding = llama_model_has_encoder(model);
-
-    uint32_t kv_size = cparams.n_ctx;
-    ggml_type type_k = params.type_k;
-    ggml_type type_v = params.type_v;
-
-    // Mamba only needs a constant number of KV cache cells per sequence
-    if (llama_model_is_recurrent(model)) {
-        // Mamba needs at least as many KV cells as there are sequences kept at any time
-        kv_size = std::max((uint32_t) 1, params.n_seq_max);
-        // it's probably best to keep as much precision as possible for the states
-        type_k = GGML_TYPE_F32; // required by ggml_ssm_conv for Mamba's conv_states
-        type_v = GGML_TYPE_F32; // required by ggml_ssm_scan for Mamba's ssm_states
-    }
-
-    GGML_ASSERT(hparams.n_embd_head_k % ggml_blck_size(type_k) == 0);
-    GGML_ASSERT(hparams.n_embd_head_v % ggml_blck_size(type_v) == 0);
-
-    if (!hparams.vocab_only) {
-        // GPU backends
-        for (auto * dev : model->devices) {
-            ggml_backend_t backend = ggml_backend_dev_init(dev, nullptr);
-            if (backend == nullptr) {
-                LLAMA_LOG_ERROR("%s: failed to initialize %s backend\n", __func__, ggml_backend_dev_name(dev));
-                llama_free(ctx);
-                return nullptr;
-            }
-            ctx->backends.emplace_back(backend);
-        }
-
-        // add ACCEL backends (such as BLAS)
-        for (size_t i = 0; i < ggml_backend_dev_count(); ++i) {
-            ggml_backend_dev_t dev = ggml_backend_dev_get(i);
-            if (ggml_backend_dev_type(dev) == GGML_BACKEND_DEVICE_TYPE_ACCEL) {
-                ggml_backend_t backend = ggml_backend_dev_init(dev, nullptr);
-                if (backend == nullptr) {
-                    LLAMA_LOG_ERROR("%s: failed to initialize %s backend\n", __func__, ggml_backend_dev_name(dev));
-                    llama_free(ctx);
-                    return nullptr;
-                }
-                ctx->backends.emplace_back(backend);
-            }
-        }
-
-        // add CPU backend
-        ctx->backend_cpu = ggml_backend_init_by_type(GGML_BACKEND_DEVICE_TYPE_CPU, nullptr);
-        if (ctx->backend_cpu == nullptr) {
-            LLAMA_LOG_ERROR("%s: failed to initialize CPU backend\n", __func__);
-            llama_free(ctx);
-            return nullptr;
-        }
-        ctx->backends.emplace_back(ctx->backend_cpu);
-
-        // create a list of the set_n_threads functions in the backends
-        for (auto & backend : ctx->backends) {
-            ggml_backend_dev_t dev = ggml_backend_get_device(backend.get());
-            ggml_backend_reg_t reg = dev ? ggml_backend_dev_backend_reg(dev) : nullptr;
-            if (reg) {
-                auto ggml_backend_set_n_threads_fn = (ggml_backend_set_n_threads_t) ggml_backend_reg_get_proc_address(reg, "ggml_backend_set_n_threads");
-                if (ggml_backend_set_n_threads_fn) {
-                    ctx->set_n_threads_fns.emplace_back(backend.get(), ggml_backend_set_n_threads_fn);
-                }
-            }
-        }
-
-        llama_set_abort_callback(ctx, params.abort_callback, params.abort_callback_data);
-
-        if (!llama_kv_cache_init(ctx->kv_self, ctx->model, ctx->cparams, type_k, type_v, kv_size, cparams.offload_kqv)) {
-            LLAMA_LOG_ERROR("%s: llama_kv_cache_init() failed for self-attention cache\n", __func__);
-            llama_free(ctx);
-            return nullptr;
-        }
-
-        {
-            size_t memory_size_k = 0;
-            size_t memory_size_v = 0;
-
-            for (auto & k : ctx->kv_self.k_l) {
-                memory_size_k += ggml_nbytes(k);
-            }
-
-            for (auto & v : ctx->kv_self.v_l) {
-                memory_size_v += ggml_nbytes(v);
-            }
-
-            LLAMA_LOG_INFO("%s: KV self size  = %7.2f MiB, K (%s): %7.2f MiB, V (%s): %7.2f MiB\n", __func__,
-                      (float)(memory_size_k + memory_size_v) / (1024.0f * 1024.0f),
-                ggml_type_name(type_k), (float)memory_size_k / (1024.0f * 1024.0f),
-                ggml_type_name(type_v), (float)memory_size_v / (1024.0f * 1024.0f));
-        }
-
-        // graph outputs buffer
-        {
-            // resized during inference when a batch uses more outputs
-            if (llama_output_reserve(*ctx, params.n_seq_max) < params.n_seq_max) {
-                LLAMA_LOG_ERROR("%s: failed to reserve initial output buffer\n", __func__);
-                llama_free(ctx);
-                return nullptr;
-            }
-
-            LLAMA_LOG_INFO("%s: %10s  output buffer size = %8.2f MiB\n", __func__,
-                    ggml_backend_buffer_name(ctx->buf_output.get()),
-                    ggml_backend_buffer_get_size(ctx->buf_output.get()) / 1024.0 / 1024.0);
-        }
-
-        // scheduler and compute buffers
-        {
-            // buffer types used for the compute buffer of each backend
-            std::vector backend_buft;
-            std::vector backend_ptrs;
-            for (auto & backend : ctx->backends) {
-                auto * buft = ggml_backend_get_default_buffer_type(backend.get());
-                auto backend_type = ggml_backend_dev_type(ggml_backend_get_device(backend.get()));
-                if (backend_type == GGML_BACKEND_DEVICE_TYPE_CPU && !model->devices.empty()) {
-                    // use the host buffer of the first device CPU for faster transfer of the intermediate state
-                    auto * dev = model->devices[0];
-                    auto * host_buft = ggml_backend_dev_host_buffer_type(dev);
-                    if (host_buft) {
-                        buft = host_buft;
-                    }
-                }
-                backend_buft.push_back(buft);
-                backend_ptrs.push_back(backend.get());
-            }
-
-            const size_t max_nodes = model->max_nodes();
-
-            // buffer used to store the computation graph and the tensor meta data
-            ctx->buf_compute_meta.resize(ggml_tensor_overhead()*max_nodes + ggml_graph_overhead_custom(max_nodes, false));
-
-            // TODO: move these checks to ggml_backend_sched
-            // enabling pipeline parallelism in the scheduler increases memory usage, so it is only done when necessary
-            bool pipeline_parallel =
-                model->n_devices() > 1 &&
-                model->params.n_gpu_layers > (int)model->hparams.n_layer &&
-                model->params.split_mode == LLAMA_SPLIT_MODE_LAYER &&
-                params.offload_kqv;
-
-            // pipeline parallelism requires support for async compute and events in all devices
-            if (pipeline_parallel) {
-                for (auto & backend : ctx->backends) {
-                    auto dev_type = ggml_backend_dev_type(ggml_backend_get_device(backend.get()));
-                    if (dev_type == GGML_BACKEND_DEVICE_TYPE_CPU) {
-                        // ignore CPU backend
-                        continue;
-                    }
-                    auto * dev = ggml_backend_get_device(backend.get());
-                    ggml_backend_dev_props props;
-                    ggml_backend_dev_get_props(dev, &props);
-                    if (!props.caps.async || !props.caps.events) {
-                        // device does not support async compute or events
-                        pipeline_parallel = false;
-                        break;
-                    }
-                }
-            }
-
-            ctx->sched.reset(ggml_backend_sched_new(backend_ptrs.data(), backend_buft.data(), backend_ptrs.size(), max_nodes, pipeline_parallel));
-
-            if (pipeline_parallel) {
-                LLAMA_LOG_INFO("%s: pipeline parallelism enabled (n_copies=%d)\n", __func__, ggml_backend_sched_get_n_copies(ctx->sched.get()));
-            }
-
-            // initialize scheduler with the worst-case graph
-            uint32_t n_seqs = 1; // TODO: worst-case number of sequences
-            uint32_t n_tokens = std::min(cparams.n_ctx, cparams.n_ubatch);
-            llama_token token = ctx->model.vocab.token_bos(); // not actually used by llama_build_graph, but required to choose between token and embedding inputs graph
-
-            llama_ubatch ubatch_pp = { true, n_tokens, n_tokens / n_seqs, n_seqs, &token, nullptr, nullptr, nullptr, nullptr, nullptr};
-            ggml_cgraph * gf_pp = llama_build_graph(*ctx, ubatch_pp, true);
-
-            // reserve pp graph first so that buffers are only allocated once
-            ggml_backend_sched_reserve(ctx->sched.get(), gf_pp);
-            int n_splits_pp = ggml_backend_sched_get_n_splits(ctx->sched.get());
-            int n_nodes_pp = ggml_graph_n_nodes(gf_pp);
-
-            // reserve with tg graph to get the number of splits and nodes
-            llama_ubatch ubatch_tg = { true, 1, 1, n_seqs, &token, nullptr, nullptr, nullptr, nullptr, nullptr};
-            ggml_cgraph * gf_tg = llama_build_graph(*ctx, ubatch_tg, true);
-            ggml_backend_sched_reserve(ctx->sched.get(), gf_tg);
-            int n_splits_tg = ggml_backend_sched_get_n_splits(ctx->sched.get());
-            int n_nodes_tg = ggml_graph_n_nodes(gf_tg);
-
-            // reserve again with pp graph to avoid ggml-alloc reallocations during inference
-            gf_pp = llama_build_graph(*ctx, ubatch_pp, true);
-            if (!ggml_backend_sched_reserve(ctx->sched.get(), gf_pp)) {
-                LLAMA_LOG_ERROR("%s: failed to allocate compute buffers\n", __func__);
-                llama_free(ctx);
-                return nullptr;
-            }
-
-            for (size_t i = 0; i < backend_ptrs.size(); ++i) {
-                ggml_backend_t backend = backend_ptrs[i];
-                ggml_backend_buffer_type_t buft = backend_buft[i];
-                size_t size = ggml_backend_sched_get_buffer_size(ctx->sched.get(), backend);
-                if (size > 1) {
-                    LLAMA_LOG_INFO("%s: %10s compute buffer size = %8.2f MiB\n", __func__,
-                            ggml_backend_buft_name(buft),
-                            size / 1024.0 / 1024.0);
-                }
-            }
-
-            if (n_nodes_pp == n_nodes_tg) {
-                LLAMA_LOG_INFO("%s: graph nodes  = %d\n", __func__, n_nodes_pp);
-            } else {
-                LLAMA_LOG_INFO("%s: graph nodes  = %d (with bs=%d), %d (with bs=1)\n", __func__, n_nodes_pp, n_tokens, n_nodes_tg);
-            }
-            if (n_splits_pp == n_splits_tg) {
-                LLAMA_LOG_INFO("%s: graph splits = %d\n", __func__, n_splits_pp);
-            } else {
-                LLAMA_LOG_INFO("%s: graph splits = %d (with bs=%d), %d (with bs=1)\n", __func__, n_splits_pp, n_tokens, n_splits_tg);
-            }
-        }
-    }
-
-    return ctx;
-}
-
-struct llama_context * llama_new_context_with_model(
-                 struct llama_model * model,
-        struct llama_context_params   params) {
-    return llama_init_from_model(model, params);
-}
-
-//
-// kv cache
-//
-
-// TODO: tmp bridges below until `struct llama_kv_cache` is exposed through the public API
-
-struct llama_kv_cache_view llama_kv_cache_view_init(const struct llama_context * ctx, int32_t n_seq_max) {
-    return llama_kv_cache_view_init(ctx->kv_self, n_seq_max);
-}
-
-void llama_kv_cache_view_update(const struct llama_context * ctx, struct llama_kv_cache_view * view) {
-    llama_kv_cache_view_update(view, ctx->kv_self);
-}
-
-int32_t llama_get_kv_cache_token_count(const struct llama_context * ctx) {
-    return llama_get_kv_cache_token_count(ctx->kv_self);
-}
-
-int32_t llama_get_kv_cache_used_cells(const struct llama_context * ctx) {
-    return llama_get_kv_cache_used_cells(ctx->kv_self);
-}
-
-void llama_kv_cache_clear(struct llama_context * ctx) {
-    llama_kv_cache_clear(ctx->kv_self);
-}
-
-bool llama_kv_cache_seq_rm(struct llama_context * ctx, llama_seq_id seq_id, llama_pos p0, llama_pos p1) {
-    return llama_kv_cache_seq_rm(ctx->kv_self, seq_id, p0, p1);
-}
-
-void llama_kv_cache_seq_cp(struct llama_context * ctx, llama_seq_id seq_id_src, llama_seq_id seq_id_dst, llama_pos p0, llama_pos p1) {
-    if (seq_id_src == seq_id_dst) {
-        return;
-    }
-    llama_kv_cache_seq_cp(ctx->kv_self, seq_id_src, seq_id_dst, p0, p1);
-}
-
-void llama_kv_cache_seq_keep(struct llama_context * ctx, llama_seq_id seq_id) {
-    llama_kv_cache_seq_keep(ctx->kv_self, seq_id);
-}
-
-void llama_kv_cache_seq_add(struct llama_context * ctx, llama_seq_id seq_id, llama_pos p0, llama_pos p1, llama_pos delta) {
-    if (delta == 0) {
-        return;
-    }
-
-    llama_kv_cache_seq_add(ctx->kv_self, seq_id, p0, p1, delta);
-}
-
-void llama_kv_cache_seq_div(struct llama_context * ctx, llama_seq_id seq_id, llama_pos p0, llama_pos p1, int d) {
-    if (d == 1) {
-        return;
-    }
-
-    llama_kv_cache_seq_div(ctx->kv_self, seq_id, p0, p1, d);
-}
-
-llama_pos llama_kv_cache_seq_pos_max(struct llama_context * ctx, llama_seq_id seq_id) {
-    return llama_kv_cache_seq_pos_max(ctx->kv_self, seq_id);
-}
-
-void llama_kv_cache_defrag(struct llama_context * ctx) {
-    llama_kv_cache_defrag(ctx->kv_self);
-}
-
-void llama_kv_cache_update(struct llama_context * ctx) {
-    llama_kv_cache_update_impl(*ctx);
-}
-
-bool llama_kv_cache_can_shift(struct llama_context * ctx) {
-    return llama_kv_cache_can_shift(ctx->kv_self);
-}
-
-///
-
-int32_t llama_encode(
-        struct llama_context * ctx,
-          struct llama_batch   batch) {
-    const int ret = llama_encode_impl(*ctx, batch);
-    if (ret != 0) {
-        LLAMA_LOG_ERROR("%s: failed to encode, ret = %d\n", __func__, ret);
-    }
-
-    return ret;
-}
-
-int32_t llama_decode(
-        struct llama_context * ctx,
-          struct llama_batch   batch) {
-    const int ret = llama_decode_impl(*ctx, batch);
-    if (ret != 0) {
-        LLAMA_LOG_ERROR("%s: failed to decode, ret = %d\n", __func__, ret);
-    }
-
-    return ret;
-}
-
 //
 // chat templates
 //
@@ -10065,7 +320,6 @@ const char * llama_print_system_info(void) {
     static std::string s;
     s.clear(); // Clear the string, since it's static, otherwise it will accumulate data from previous calls.
 
-
     for (size_t i = 0; i < ggml_backend_reg_count(); i++) {
         auto * reg = ggml_backend_reg_get(i);
         auto * get_features_fn = (ggml_backend_get_features_t) ggml_backend_reg_get_proc_address(reg, "ggml_backend_get_features");
@@ -10084,43 +338,3 @@ const char * llama_print_system_info(void) {
 
     return s.c_str();
 }
-
-//
-// perf
-//
-
-struct llama_perf_context_data llama_perf_context(const struct llama_context * ctx) {
-    struct llama_perf_context_data data = {};
-
-    if (ctx == nullptr) {
-        return data;
-    }
-
-    data.t_start_ms  = 1e-3 * ctx->t_start_us;
-    data.t_load_ms   = 1e-3 * ctx->t_load_us;
-    data.t_p_eval_ms = 1e-3 * ctx->t_p_eval_us;
-    data.t_eval_ms   = 1e-3 * ctx->t_eval_us;
-    data.n_p_eval    = std::max(1, ctx->n_p_eval);
-    data.n_eval      = std::max(1, ctx->n_eval);
-
-    return data;
-}
-
-void llama_perf_context_print(const struct llama_context * ctx) {
-    const auto data = llama_perf_context(ctx);
-
-    const double t_end_ms = 1e-3 * ggml_time_us();
-
-    LLAMA_LOG_INFO("%s:        load time = %10.2f ms\n", __func__, data.t_load_ms);
-    LLAMA_LOG_INFO("%s: prompt eval time = %10.2f ms / %5d tokens (%8.2f ms per token, %8.2f tokens per second)\n",
-            __func__, data.t_p_eval_ms, data.n_p_eval, data.t_p_eval_ms / data.n_p_eval, 1e3 / data.t_p_eval_ms * data.n_p_eval);
-    LLAMA_LOG_INFO("%s:        eval time = %10.2f ms / %5d runs   (%8.2f ms per token, %8.2f tokens per second)\n",
-            __func__, data.t_eval_ms, data.n_eval, data.t_eval_ms / data.n_eval, 1e3 / data.t_eval_ms * data.n_eval);
-    LLAMA_LOG_INFO("%s:       total time = %10.2f ms / %5d tokens\n", __func__, (t_end_ms - data.t_start_ms), (data.n_p_eval + data.n_eval));
-}
-
-void llama_perf_context_reset(struct llama_context * ctx) {
-    ctx->t_start_us  = ggml_time_us();
-    ctx->t_eval_us   = ctx->n_eval = 0;
-    ctx->t_p_eval_us = ctx->n_p_eval = 0;
-}
diff --git a/src/unicode.cpp b/src/unicode.cpp
index a32ae6d08..e63bb4ab0 100644
--- a/src/unicode.cpp
+++ b/src/unicode.cpp
@@ -708,7 +708,7 @@ std::vector unicode_regex_split(const std::string & text, const std
     const auto cpts = unicode_cpts_from_utf8(text);
 
     // generate a "collapsed" representation of the text, where all codepoints are replaced by a single byte
-    // ref: https://github.com/ggerganov/llama.cpp/pull/6920#issuecomment-2081479935
+    // ref: https://github.com/ggml-org/llama.cpp/pull/6920#issuecomment-2081479935
     std::string text_collapsed;
     if (need_collapse) {
         // collapse all unicode categories
diff --git a/tests/CMakeLists.txt b/tests/CMakeLists.txt
index 7a158d602..2bb210702 100644
--- a/tests/CMakeLists.txt
+++ b/tests/CMakeLists.txt
@@ -131,9 +131,13 @@ if (NOT WIN32)
 endif()
 
 llama_target_and_test(test-log.cpp)
-llama_target_and_test(test-arg-parser.cpp)
 llama_target_and_test(test-chat-template.cpp)
 
+# this fails on windows (github hosted runner) due to curl DLL not found (exit code 0xc0000135)
+if (NOT WIN32)
+    llama_target_and_test(test-arg-parser.cpp)
+endif()
+
 # llama_target_and_test(test-opt.cpp) # SLOW
 llama_target_and_test(test-gguf.cpp)
 llama_target_and_test(test-backend-ops.cpp)
diff --git a/tests/test-arg-parser.cpp b/tests/test-arg-parser.cpp
index 69604b87c..537fc63a4 100644
--- a/tests/test-arg-parser.cpp
+++ b/tests/test-arg-parser.cpp
@@ -77,7 +77,7 @@ int main(void) {
 
     argv = {"binary_name", "-m", "model_file.gguf"};
     assert(true == common_params_parse(argv.size(), list_str_to_char(argv).data(), params, LLAMA_EXAMPLE_COMMON));
-    assert(params.model == "model_file.gguf");
+    assert(params.model.path == "model_file.gguf");
 
     argv = {"binary_name", "-t", "1234"};
     assert(true == common_params_parse(argv.size(), list_str_to_char(argv).data(), params, LLAMA_EXAMPLE_COMMON));
@@ -89,7 +89,7 @@ int main(void) {
 
     argv = {"binary_name", "-m", "abc.gguf", "--predict", "6789", "--batch-size", "9090"};
     assert(true == common_params_parse(argv.size(), list_str_to_char(argv).data(), params, LLAMA_EXAMPLE_COMMON));
-    assert(params.model == "abc.gguf");
+    assert(params.model.path == "abc.gguf");
     assert(params.n_predict == 6789);
     assert(params.n_batch == 9090);
 
@@ -112,7 +112,7 @@ int main(void) {
     setenv("LLAMA_ARG_THREADS", "1010", true);
     argv = {"binary_name"};
     assert(true == common_params_parse(argv.size(), list_str_to_char(argv).data(), params, LLAMA_EXAMPLE_COMMON));
-    assert(params.model == "blah.gguf");
+    assert(params.model.path == "blah.gguf");
     assert(params.cpuparams.n_threads == 1010);
 
 
@@ -122,7 +122,7 @@ int main(void) {
     setenv("LLAMA_ARG_THREADS", "1010", true);
     argv = {"binary_name", "-m", "overwritten.gguf"};
     assert(true == common_params_parse(argv.size(), list_str_to_char(argv).data(), params, LLAMA_EXAMPLE_COMMON));
-    assert(params.model == "overwritten.gguf");
+    assert(params.model.path == "overwritten.gguf");
     assert(params.cpuparams.n_threads == 1010);
 #endif // _WIN32
 
diff --git a/tests/test-backend-ops.cpp b/tests/test-backend-ops.cpp
index 1bfd41254..3a5741c8d 100644
--- a/tests/test-backend-ops.cpp
+++ b/tests/test-backend-ops.cpp
@@ -18,20 +18,22 @@
 #include 
 #include 
 #include 
+#include 
 
 #include 
 #include 
 #include 
-#include 
-#include 
 #include 
+#include 
+#include 
+#include 
+#include 
+#include 
 #include 
 #include 
-#include 
-#include 
+#include 
 #include 
 #include 
-#include 
 #include 
 
 static void init_tensor_uniform(ggml_tensor * tensor, float min = -1.0f, float max = 1.0f) {
@@ -257,6 +259,10 @@ static std::string var_to_str(ggml_type type) {
     return ggml_type_name(type);
 }
 
+static std::string var_to_str(ggml_prec prec) {
+    return prec == GGML_PREC_F32 ? "f32" : "def";
+}
+
 static std::string var_to_str(ggml_op_pool pool) {
     switch (pool) {
         case GGML_OP_POOL_AVG:  return "avg";
@@ -265,6 +271,14 @@ static std::string var_to_str(ggml_op_pool pool) {
     }
 }
 
+static std::string var_to_str(ggml_scale_mode mode) {
+    switch (mode) {
+        case GGML_SCALE_MODE_NEAREST:  return "nearest";
+        case GGML_SCALE_MODE_BILINEAR: return "bilinear";
+        default:                      return std::to_string(mode);
+    }
+}
+
 #define VAR_TO_STR(x) (#x "=" + var_to_str(x))
 
 #define VARS_TO_STR1(a) VAR_TO_STR(a)
@@ -467,6 +481,7 @@ struct test_case {
 
         // allocate
         ggml_backend_buffer_t buf = ggml_backend_alloc_ctx_tensors(ctx, backend1);
+
         if (buf == NULL) {
             printf("failed to allocate tensors [%s] ", ggml_backend_name(backend1));
             ggml_free(ctx);
@@ -588,14 +603,13 @@ struct test_case {
             /* .mem_base = */ NULL,
             /* .no_alloc = */ true,
         };
-        ggml_context * ctx = ggml_init(params);
+        ggml_context_ptr ctx(ggml_init(params)); // smart ptr
         GGML_ASSERT(ctx);
 
-        ggml_tensor * out = build_graph(ctx);
+        ggml_tensor * out = build_graph(ctx.get());
 
         if (op_name != nullptr && op_desc(out) != op_name) {
             //printf("  %s: skipping\n", op_desc(out).c_str());
-            ggml_free(ctx);
             return true;
         }
 
@@ -605,7 +619,6 @@ struct test_case {
         // check if backends support op
         if (!ggml_backend_supports_op(backend, out)) {
             printf("not supported\n");
-            ggml_free(ctx);
             return true;
         }
 
@@ -618,22 +631,26 @@ struct test_case {
         printf("%*s", last - len, "");
 
         // allocate
-        ggml_backend_buffer_t buf = ggml_backend_alloc_ctx_tensors(ctx, backend);
+        ggml_backend_buffer_ptr buf(ggml_backend_alloc_ctx_tensors(ctx.get(), backend)); // smart ptr
+
         if (buf == NULL) {
             printf("failed to allocate tensors\n");
-            ggml_free(ctx);
             return false;
         }
 
         // randomize tensors
-        initialize_tensors(ctx);
+        initialize_tensors(ctx.get());
 
         // build graph
-        ggml_cgraph * gf = ggml_new_graph_custom(ctx, graph_nodes, false);
+        ggml_cgraph * gf = ggml_new_graph_custom(ctx.get(), graph_nodes, false);
         ggml_build_forward_expand(gf, out);
 
         // warmup run
-        ggml_backend_graph_compute(backend, gf);
+        ggml_status status = ggml_backend_graph_compute(backend, gf);
+        if (status != GGML_STATUS_SUCCESS) {
+            fprintf(stderr, "%s: ggml_backend_graph_compute failed. status=%s \n", __func__, ggml_status_to_string(status));
+            return false;
+        }
 
         // determine number of runs
         int n_runs;
@@ -684,7 +701,11 @@ struct test_case {
         int total_runs = 0;
         do {
             int64_t start_time = ggml_time_us();
-            ggml_backend_graph_compute(backend, gf);
+            ggml_status status = ggml_backend_graph_compute(backend, gf);
+            if (status != GGML_STATUS_SUCCESS) {
+                fprintf(stderr, "%s: ggml_backend_graph_compute failed. status=%s \n", __func__, ggml_status_to_string(status));
+                return false;
+            }
             int64_t end_time = ggml_time_us();
 
             total_time_us += end_time - start_time;
@@ -722,10 +743,6 @@ struct test_case {
         }
         printf("\n");
 
-        ggml_backend_buffer_free(buf);
-
-        ggml_free(ctx);
-
         return true;
     }
 
@@ -738,17 +755,16 @@ struct test_case {
             /* .mem_base = */ NULL,
             /* .no_alloc = */ true,
         };
-        ggml_context * ctx = ggml_init(params);
+        ggml_context_ptr ctx(ggml_init(params)); // smart ptr
         GGML_ASSERT(ctx);
 
-        gf = ggml_new_graph_custom(ctx, GGML_DEFAULT_GRAPH_SIZE, true);
-        gb = ggml_new_graph_custom(ctx, GGML_DEFAULT_GRAPH_SIZE, true);
+        gf = ggml_new_graph_custom(ctx.get(), GGML_DEFAULT_GRAPH_SIZE, true);
+        gb = ggml_new_graph_custom(ctx.get(), GGML_DEFAULT_GRAPH_SIZE, true);
 
-        ggml_tensor * out = build_graph(ctx);
+        ggml_tensor * out = build_graph(ctx.get());
 
         if ((op_name != nullptr && op_desc(out) != op_name) || out->op == GGML_OP_OPT_STEP_ADAMW) {
             //printf("  %s: skipping\n", op_desc(out).c_str());
-            ggml_free(ctx);
             return true;
         }
 
@@ -756,7 +772,6 @@ struct test_case {
         fflush(stdout);
 
         if (out->type != GGML_TYPE_F32) {
-            ggml_free(ctx);
             printf("not supported [%s->type != FP32]\n", out->name);
             return true;
         }
@@ -764,7 +779,7 @@ struct test_case {
         // check if the backend supports the ops
         bool supported = true;
         bool any_params = false;
-        for (ggml_tensor * t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) {
+        for (ggml_tensor * t = ggml_get_first_tensor(ctx.get()); t != NULL; t = ggml_get_next_tensor(ctx.get(), t)) {
             if (!ggml_backend_supports_op(backend, t)) {
                 printf("not supported [%s] ", ggml_backend_name(backend));
                 supported = false;
@@ -785,40 +800,38 @@ struct test_case {
         }
         if (!supported) {
             printf("\n");
-            ggml_free(ctx);
             return true;
         }
 
         int64_t ngrads = 0;
-        for (ggml_tensor * t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) {
+        for (ggml_tensor * t = ggml_get_first_tensor(ctx.get()); t != NULL; t = ggml_get_next_tensor(ctx.get(), t)) {
             if (t->flags & GGML_TENSOR_FLAG_PARAM) {
                 ngrads += ggml_nelements(t);
             }
         }
         if (ngrads > grad_nmax()) {
             printf("skipping large tensors for speed \n");
-            ggml_free(ctx);
             return true;
         }
 
 
         if (!ggml_is_scalar(out)) {
-            out = ggml_sum(ctx, out);
+            out = ggml_sum(ctx.get(), out);
             ggml_set_name(out, "sum_of_out");
         }
         ggml_set_loss(out);
 
         ggml_build_forward_expand(gf, out);
         ggml_graph_cpy(gf, gb);
-        ggml_build_backward_expand(ctx, ctx, gb, false);
+        ggml_build_backward_expand(ctx.get(), ctx.get(), gb, false);
         if (expect.size() != 1 || expect[0] != 0.0f) {
             GGML_ASSERT(ggml_graph_n_nodes(gb) > ggml_graph_n_nodes(gf));
-            for (ggml_tensor * t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) {
+            for (ggml_tensor * t = ggml_get_first_tensor(ctx.get()); t != NULL; t = ggml_get_next_tensor(ctx.get(), t)) {
                 GGML_ASSERT(!(t->flags & GGML_TENSOR_FLAG_PARAM) || ggml_graph_get_grad(gb, t)->op != GGML_OP_NONE);
             }
         }
 
-        for (ggml_tensor * t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) {
+        for (ggml_tensor * t = ggml_get_first_tensor(ctx.get()); t != NULL; t = ggml_get_next_tensor(ctx.get(), t)) {
             if (!ggml_backend_supports_op(backend, t)) {
                 printf("not supported [%s] ", ggml_backend_name(backend));
                 supported = false;
@@ -832,27 +845,32 @@ struct test_case {
         }
         if (!supported) {
             printf("\n");
-            ggml_free(ctx);
             return true;
         }
 
         // allocate
-        ggml_backend_buffer_t buf = ggml_backend_alloc_ctx_tensors(ctx, backend);
+        ggml_backend_buffer_ptr buf(ggml_backend_alloc_ctx_tensors(ctx.get(), backend)); // smart ptr
         if (buf == NULL) {
             printf("failed to allocate tensors [%s] ", ggml_backend_name(backend));
-            ggml_free(ctx);
             return false;
         }
 
-
-        initialize_tensors(ctx); // Randomizes all tensors (including gradients).
+        initialize_tensors(ctx.get()); // Randomizes all tensors (including gradients).
         ggml_graph_reset(gb);    // Sets gradients to 1 if loss, 0 otherwise.
 
-        ggml_backend_graph_compute(backend, gf);
-        ggml_backend_graph_compute(backend, gb);
+        ggml_status status = ggml_backend_graph_compute(backend, gf);
+        if (status != GGML_STATUS_SUCCESS) {
+            fprintf(stderr, "%s: ggml_backend_graph_compute failed. status=%s \n", __func__, ggml_status_to_string(status));
+            return false;
+        }
+        status = ggml_backend_graph_compute(backend, gb);
+        if (status != GGML_STATUS_SUCCESS) {
+            fprintf(stderr, "%s: ggml_backend_graph_compute failed. status=%s \n", __func__, ggml_status_to_string(status));
+            return false;
+        }
 
         bool ok = true;
-        for (struct ggml_tensor * t = ggml_get_first_tensor(ctx); t != nullptr; t = ggml_get_next_tensor(ctx, t)) {
+        for (struct ggml_tensor * t = ggml_get_first_tensor(ctx.get()); t != nullptr; t = ggml_get_next_tensor(ctx.get(), t)) {
             if (!(t->flags & GGML_TENSOR_FLAG_PARAM)) {
                 continue;
             }
@@ -897,20 +915,36 @@ struct test_case {
                 float fu, fuh, fdh, fd; // output values for xiu, xiuh, xid, xidh
 
                 ggml_backend_tensor_set(t, &xiu, i*sizeof(float), sizeof(float));
-                ggml_backend_graph_compute(backend, gf);
+                status = ggml_backend_graph_compute(backend, gf);
+                if (status != GGML_STATUS_SUCCESS) {
+                    fprintf(stderr, "%s: ggml_backend_graph_compute failed. status=%s \n", __func__, ggml_status_to_string(status));
+                    return false;
+                }
                 ggml_backend_tensor_get(out, &fu, 0, ggml_nbytes(out));
 
                 ggml_backend_tensor_set(t, &xid, i*sizeof(float), sizeof(float));
-                ggml_backend_graph_compute(backend, gf);
+                status = ggml_backend_graph_compute(backend, gf);
+                if (status != GGML_STATUS_SUCCESS) {
+                    fprintf(stderr, "%s: ggml_backend_graph_compute failed. status=%s \n", __func__, ggml_status_to_string(status));
+                    return false;
+                }
                 ggml_backend_tensor_get(out, &fd, 0, ggml_nbytes(out));
 
                 if (grad_precise()) {
                     ggml_backend_tensor_set(t, &xiuh, i*sizeof(float), sizeof(float));
-                    ggml_backend_graph_compute(backend, gf);
+                    status = ggml_backend_graph_compute(backend, gf);
+                    if (status != GGML_STATUS_SUCCESS) {
+                        fprintf(stderr, "%s: ggml_backend_graph_compute failed. status=%s \n", __func__, ggml_status_to_string(status));
+                        return false;
+                    }
                     ggml_backend_tensor_get(out, &fuh, 0, ggml_nbytes(out));
 
                     ggml_backend_tensor_set(t, &xidh, i*sizeof(float), sizeof(float));
-                    ggml_backend_graph_compute(backend, gf);
+                    status = ggml_backend_graph_compute(backend, gf);
+                    if (status != GGML_STATUS_SUCCESS) {
+                        fprintf(stderr, "%s: ggml_backend_graph_compute failed. status=%s \n", __func__, ggml_status_to_string(status));
+                        return false;
+                    }
                     ggml_backend_tensor_get(out, &fdh, 0, ggml_nbytes(out));
 
                     gn[i] = (8.0*(double)fuh + (double)fd - (8.0*(double)fdh + (double)fu)) / (6.0*(double)eps);
@@ -936,10 +970,6 @@ struct test_case {
             printf("compare failed ");
         }
 
-        ggml_backend_buffer_free(buf);
-
-        ggml_free(ctx);
-
         if (ok) {
             printf("\033[1;32mOK\033[0m\n");
             return true;
@@ -1254,7 +1284,7 @@ struct test_count_equal : public test_case {
         ggml_tensor * b = ggml_new_tensor(ctx, type, 4, ne.data());
         ggml_set_name(b, "b");
 
-        ggml_tensor * b_argmax = ggml_argmax(ctx, a);
+        ggml_tensor * b_argmax = ggml_argmax(ctx, b);
         ggml_set_name(b_argmax, "b_argmax");
 
         ggml_tensor * out = ggml_count_equal(ctx, a_argmax, b_argmax);
@@ -1441,11 +1471,13 @@ struct test_cpy : public test_case {
     const ggml_type type_src;
     const ggml_type type_dst;
     const std::array ne;
-    const std::array permute;
+    const std::array permute_src;
+    const std::array permute_dst;
     bool _src_use_permute;
+    bool _dst_use_permute;
 
     std::string vars() override {
-        return VARS_TO_STR4(type_src, type_dst, ne, permute);
+        return VARS_TO_STR5(type_src, type_dst, ne, permute_src, permute_dst);
     }
 
     double max_nmse_err() override {
@@ -1458,9 +1490,11 @@ struct test_cpy : public test_case {
 
     test_cpy(ggml_type type_src = GGML_TYPE_F32, ggml_type type_dst = GGML_TYPE_F32,
             std::array ne = {10, 10, 10, 1},
-            std::array permute = {0, 0, 0, 0})
-        : type_src(type_src), type_dst(type_dst), ne(ne), permute(permute),
-          _src_use_permute(permute[0] + permute[1] + permute[2] + permute[3] > 0) {}
+            std::array permute_src = {0, 0, 0, 0},
+            std::array permute_dst = {0, 0, 0, 0})
+        : type_src(type_src), type_dst(type_dst), ne(ne), permute_src(permute_src), permute_dst(permute_dst),
+          _src_use_permute(permute_src[0] + permute_src[1] + permute_src[2] + permute_src[3] > 0),
+          _dst_use_permute(permute_dst[0] + permute_dst[1] + permute_dst[2] + permute_dst[3] > 0) {}
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
         ggml_tensor * src = ggml_new_tensor(ctx, type_src, 4, ne.data());
@@ -1468,13 +1502,18 @@ struct test_cpy : public test_case {
         ggml_set_name(src, "src");
 
         if (_src_use_permute) {
-            src = ggml_permute(ctx, src, permute[0], permute[1], permute[2], permute[3]);
+            src = ggml_permute(ctx, src, permute_src[0], permute_src[1], permute_src[2], permute_src[3]);
             ggml_set_name(src, "src_permuted");
         }
 
-        ggml_tensor* dst = ggml_new_tensor(ctx, type_dst, 4, src->ne);
+        ggml_tensor * dst = ggml_new_tensor(ctx, type_dst, 4, src->ne);
         ggml_set_name(dst, "dst");
 
+        if (_dst_use_permute) {
+            dst = ggml_permute(ctx, dst, permute_dst[0], permute_dst[1], permute_dst[2], permute_dst[3]);
+            ggml_set_name(dst, "dst_permuted");
+        }
+
         ggml_tensor * out = ggml_cpy(ctx, src, dst);
         ggml_set_name(out, "out");
 
@@ -1511,6 +1550,7 @@ struct test_cont : public test_case {
 };
 
 // GGML_OP_ADD
+// GGML_OP_SUB
 // GGML_OP_MUL
 // GGML_OP_DIV
 struct test_bin_bcast : public test_case {
@@ -1897,6 +1937,40 @@ struct test_gla : public test_case {
     }
 };
 
+// GGML_OP_RWKV_WKV7
+struct test_rwkv_wkv7 : public test_case {
+    const ggml_type type;
+
+    const int64_t head_count;
+    const int64_t head_size;
+    const int64_t n_seq_tokens;
+    const int64_t n_seqs;
+
+    std::string vars() override {
+        return VARS_TO_STR5(type, head_count, head_size, n_seq_tokens, n_seqs);
+    }
+
+    test_rwkv_wkv7(ggml_type type = GGML_TYPE_F32,
+            int64_t head_count = 32, int64_t head_size = 64, int64_t n_seq_tokens = 32, int64_t n_seqs = 32)
+        : type(type), head_count(head_count), head_size(head_size), n_seq_tokens(n_seq_tokens), n_seqs(n_seqs) {}
+
+    ggml_tensor * build_graph(ggml_context * ctx) override {
+        const int64_t n_tokens = n_seq_tokens * n_seqs;
+        ggml_tensor * r   = ggml_new_tensor(ctx, type, 3, std::vector{ head_size, head_count, n_tokens }.data());
+        ggml_tensor * w   = ggml_new_tensor(ctx, type, 3, std::vector{ head_size, head_count, n_tokens }.data());
+        ggml_tensor * k   = ggml_new_tensor(ctx, type, 3, std::vector{ head_size, head_count, n_tokens }.data());
+        ggml_tensor * v   = ggml_new_tensor(ctx, type, 3, std::vector{ head_size, head_count, n_tokens }.data());
+        ggml_tensor * a   = ggml_new_tensor(ctx, type, 3, std::vector{ head_size, head_count, n_tokens }.data());
+        ggml_tensor * b   = ggml_new_tensor(ctx, type, 3, std::vector{ head_size, head_count, n_tokens }.data());
+        // Outputs may become NaN with long seqlen without these normalization
+        a = ggml_l2_norm(ctx, a, 1e-7F);
+        b = ggml_l2_norm(ctx, b, 1e-7F);
+        ggml_tensor * s   = ggml_new_tensor(ctx, type, 2, std::vector{ head_size * head_size * head_count, n_seqs }.data());
+        ggml_tensor * out = ggml_rwkv_wkv7(ctx, r, w, k, v, a, b, s);
+        return out;
+    }
+};
+
 // GGML_OP_MUL_MAT
 struct test_mul_mat : public test_case {
     const ggml_type type_a;
@@ -1907,9 +1981,10 @@ struct test_mul_mat : public test_case {
     const std::array bs;  // dims 3 and 4
     const std::array nr;  // repeat in dims 3 and 4
     const std::array per; // permutation of dimensions
+    const bool v; // whether a is a non-contiguous view
 
     std::string vars() override {
-        return VARS_TO_STR8(type_a, type_b, m, n, k, bs, nr, per);
+        return VARS_TO_STR9(type_a, type_b, m, n, k, bs, nr, per, v);
     }
 
     double max_nmse_err() override {
@@ -1929,8 +2004,9 @@ struct test_mul_mat : public test_case {
             int64_t m = 32, int64_t n = 32, int64_t k = 32,
             std::array bs = {10, 10},
             std::array nr = {2, 2},
-            std::array per = {0, 1, 2, 3})
-        : type_a(type_a), type_b(type_b), m(m), n(n), k(k), bs(bs), nr(nr), per(per) {}
+            std::array per = {0, 1, 2, 3},
+            bool v = false)
+        : type_a(type_a), type_b(type_b), m(m), n(n), k(k), bs(bs), nr(nr), per(per), v(v) {}
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
         // C^T = A * B^T: (k, m) * (k, n) => (m, n)
@@ -1940,6 +2016,7 @@ struct test_mul_mat : public test_case {
         const int npermuted = (per[0] != 0) + (per[1] != 1) + (per[2] != 2) + (per[3] != 3);
         if (npermuted > 0) {
             GGML_ASSERT(npermuted == 2);
+            GGML_ASSERT(!v); // not handled
             GGML_ASSERT(!ggml_is_quantized(type_a) || per[0] == 0);
             GGML_ASSERT(!ggml_is_quantized(type_b) || per[0] == 0);
 
@@ -1963,7 +2040,13 @@ struct test_mul_mat : public test_case {
             ggml_set_name(a, "a_permuted");
             ggml_set_name(b, "b_permuted");
         } else {
-            a = ggml_new_tensor_4d(ctx, type_a, k, m, bs[0],       bs[1]);
+
+            if (v) {
+                a = ggml_new_tensor_4d(ctx, type_a, k*2, m, bs[0], bs[1]);
+                a = ggml_view_4d(ctx, a, k, m, bs[0], bs[1], a->nb[1], a->nb[2], a->nb[3], 0);
+            } else {
+                a = ggml_new_tensor_4d(ctx, type_a, k, m, bs[0],       bs[1]);
+            }
             b = ggml_new_tensor_4d(ctx, type_b, k, n, bs[0]*nr[0], bs[1]*nr[1]);
             if (!ggml_is_quantized(type_a)) {
                 if (bs[1] == 1 && nr[1] == 1) {
@@ -2873,15 +2956,16 @@ struct test_upscale : public test_case {
     const std::array ne;
     const int32_t scale_factor;
     const bool transpose;
+    const ggml_scale_mode mode;
 
     std::string vars() override {
-        return VARS_TO_STR4(type, ne, scale_factor, transpose);
+        return VARS_TO_STR5(type, ne, scale_factor, mode, transpose);
     }
 
     test_upscale(ggml_type type = GGML_TYPE_F32,
             std::array ne = {512, 512, 3, 1},
-            int32_t scale_factor = 2, bool transpose = false)
-        : type(type), ne(ne), scale_factor(scale_factor), transpose(transpose) {}
+            int32_t scale_factor = 2, ggml_scale_mode mode = GGML_SCALE_MODE_NEAREST, bool transpose = false)
+        : type(type), ne(ne), scale_factor(scale_factor), transpose(transpose), mode(mode) {}
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
         ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data());
@@ -2892,7 +2976,7 @@ struct test_upscale : public test_case {
             ggml_set_name(a, "a_transposed");
         }
 
-        ggml_tensor * out = ggml_upscale(ctx, a, scale_factor);
+        ggml_tensor * out = ggml_upscale(ctx, a, scale_factor, mode);
         ggml_set_name(out, "out");
 
         return out;
@@ -2904,21 +2988,23 @@ struct test_upscale_ext : public test_case {
     const ggml_type type;
     const std::array ne;
     const std::array ne_tgt;
+    const ggml_scale_mode mode = GGML_SCALE_MODE_NEAREST;
 
     std::string vars() override {
-        return VARS_TO_STR3(type, ne, ne_tgt);
+        return VARS_TO_STR4(type, ne, ne_tgt, mode);
     }
 
     test_upscale_ext(ggml_type type = GGML_TYPE_F32,
             std::array ne     = {2, 5,  7, 11},
-            std::array ne_tgt = {5, 7, 11, 13})
-        : type(type), ne(ne), ne_tgt(ne_tgt) {}
+            std::array ne_tgt = {5, 7, 11, 13},
+            ggml_scale_mode mode = GGML_SCALE_MODE_NEAREST)
+        : type(type), ne(ne), ne_tgt(ne_tgt), mode(mode) {}
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
         ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data());
         ggml_set_name(a, "a");
 
-        ggml_tensor * out = ggml_upscale_ext(ctx, a, ne_tgt[0], ne_tgt[1],ne_tgt[2], ne_tgt[3]);
+        ggml_tensor * out = ggml_upscale_ext(ctx, a, ne_tgt[0], ne_tgt[1],ne_tgt[2], ne_tgt[3], mode);
         ggml_set_name(out, "out");
 
         return out;
@@ -2953,6 +3039,32 @@ struct test_group_norm : public test_case {
     }
 };
 
+// GGML_OP_L2_NORM
+struct test_l2_norm : public test_case {
+    const ggml_type type;
+    const std::array ne;
+    const float eps;
+
+    std::string vars() override {
+        return VARS_TO_STR2(type, ne);
+    }
+
+    test_l2_norm(ggml_type type = GGML_TYPE_F32,
+            std::array ne = {64, 64, 320, 1},
+            float eps = 1e-12f)
+        : type(type), ne(ne), eps(eps) {}
+
+    ggml_tensor * build_graph(ggml_context * ctx) override {
+        ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data());
+        ggml_set_name(a, "a");
+
+        ggml_tensor * out = ggml_l2_norm(ctx, a, eps);
+        ggml_set_name(out, "out");
+
+        return out;
+    }
+};
+
 // GGML_OP_ACC
 struct test_acc : public test_case {
     const ggml_type type;
@@ -3116,8 +3228,10 @@ struct test_leaky_relu : public test_case {
 
 // GGML_OP_FLASH_ATTN_EXT
 struct test_flash_attn_ext : public test_case {
-    const int64_t hs; // head size
+    const int64_t hsk; // K head size
+    const int64_t hsv; // V head size
     const int64_t nh; // num heads
+    const int64_t nr; // repeat in Q, tests for grouped-query attention
     const int64_t kv; // kv size
     const int64_t nb; // batch size
 
@@ -3126,11 +3240,12 @@ struct test_flash_attn_ext : public test_case {
     const float max_bias; // ALiBi
     const float logit_softcap; // Gemma 2
 
+    const ggml_prec prec;
     const ggml_type type_KV;
     std::array permute;
 
     std::string vars() override {
-        return VARS_TO_STR9(hs, nh, kv, nb, mask, max_bias, logit_softcap, type_KV, permute);
+        return VARS_TO_STR12(hsk, hsv, nh, nr, kv, nb, mask, max_bias, logit_softcap, prec, type_KV, permute);
     }
 
     double max_nmse_err() override {
@@ -3140,17 +3255,18 @@ struct test_flash_attn_ext : public test_case {
     uint64_t op_flops(ggml_tensor * t) override {
         GGML_UNUSED(t);
         // Just counting matmul costs:
-        // Q*K^T is nb x hs x kv, P*V is nb x kv x hs, per head
-        return 2 * 2 * nh * nb * hs * kv;
+        // Q*K^T is nb x hsk x kv, P*V is nb x kv x hsv, per head
+        return 2 * nh*nr * nb * (hsk + hsv) * kv;
     }
 
-    test_flash_attn_ext(int64_t hs = 128, int64_t nh = 32, int64_t kv = 96, int64_t nb = 8,
-                        bool mask = true, float max_bias = 0.0f, float logit_softcap = 0.0f, ggml_type type_KV = GGML_TYPE_F16,
-                        std::array permute = {0, 1, 2, 3})
-        : hs(hs), nh(nh), kv(kv), nb(nb), mask(mask), max_bias(max_bias), logit_softcap(logit_softcap), type_KV(type_KV), permute(permute) {}
+    test_flash_attn_ext(int64_t hsk = 128, int64_t hsv = 128, int64_t nh = 32, int64_t nr = 1, int64_t kv = 96, int64_t nb = 8,
+                        bool mask = true, float max_bias = 0.0f, float logit_softcap = 0.0f, ggml_prec prec = GGML_PREC_F32,
+                        ggml_type type_KV = GGML_TYPE_F16, std::array permute = {0, 1, 2, 3})
+        : hsk(hsk), hsv(hsv), nh(nh), nr(nr), kv(kv), nb(nb), mask(mask), max_bias(max_bias), logit_softcap(logit_softcap), prec(prec), type_KV(type_KV), permute(permute) {}
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
-        const int64_t hs_padded = GGML_PAD(hs, ggml_blck_size(type_KV));
+        const int64_t hsk_padded = GGML_PAD(hsk, ggml_blck_size(type_KV));
+        const int64_t hsv_padded = GGML_PAD(hsv, ggml_blck_size(type_KV));
 
         auto const &create_permuted = [&](ggml_type type, int64_t ne0, int64_t ne1, int64_t ne2, int64_t ne3) -> ggml_tensor * {
             int64_t ne[4] = {ne0, ne1, ne2, ne3};
@@ -3165,13 +3281,13 @@ struct test_flash_attn_ext : public test_case {
             return t;
         };
 
-        ggml_tensor * q = create_permuted(GGML_TYPE_F32, hs_padded, nb, nh, 1);
+        ggml_tensor * q = create_permuted(GGML_TYPE_F32, hsk_padded, nb, nh*nr, 1);
         ggml_set_name(q, "q");
 
-        ggml_tensor * k = create_permuted(type_KV,       hs_padded, kv, nh, 1);
+        ggml_tensor * k = create_permuted(type_KV,       hsk_padded, kv, nh,    1);
         ggml_set_name(k, "k");
 
-        ggml_tensor * v = create_permuted(type_KV,       hs_padded, kv, nh, 1);
+        ggml_tensor * v = create_permuted(type_KV,       hsv_padded, kv, nh,    1);
         ggml_set_name(v, "v");
 
         ggml_tensor * m = nullptr;
@@ -3180,7 +3296,8 @@ struct test_flash_attn_ext : public test_case {
             ggml_set_name(m, "m");
         }
 
-        ggml_tensor * out = ggml_flash_attn_ext(ctx, q, k, v, m, 1.0f/sqrtf(hs), max_bias, logit_softcap);
+        ggml_tensor * out = ggml_flash_attn_ext(ctx, q, k, v, m, 1.0f/sqrtf(hsk), max_bias, logit_softcap);
+        ggml_flash_attn_ext_set_prec(out, prec);
         ggml_set_name(out, "out");
 
         return out;
@@ -3751,10 +3868,12 @@ static std::vector> make_test_cases_eval() {
     std::default_random_engine rng(0);
 
     // unary ops
-    for (int v : {0, 1}) {
-        for (int op = 0; op < GGML_UNARY_OP_COUNT; op++) {
-            test_cases.emplace_back(new test_unary((ggml_unary_op) op, GGML_TYPE_F32, { 128, 2, 2, 2 }, v));
-            test_cases.emplace_back(new test_unary((ggml_unary_op) op, GGML_TYPE_F32, { 5, 7, 11, 13 }, v));
+    for (ggml_type type : {GGML_TYPE_F16, GGML_TYPE_F32}) {
+        for (int v : {0, 1}) {
+            for (int op = 0; op < GGML_UNARY_OP_COUNT; op++) {
+                test_cases.emplace_back(new test_unary((ggml_unary_op) op, type, { 128, 2, 2, 2 }, v));
+                test_cases.emplace_back(new test_unary((ggml_unary_op) op, type, { 5, 7, 11, 13 }, v));
+            }
         }
     }
 
@@ -3860,7 +3979,8 @@ static std::vector> make_test_cases_eval() {
     test_cases.emplace_back(new test_conv_transpose_1d({3,2,1,1}, {3,1,2,1}, 1, 0, 1));
     test_cases.emplace_back(new test_conv_transpose_1d({2,1,1,1}, {3,1,1,1}, 1, 0, 1));
 
-    test_cases.emplace_back(new test_count_equal());
+    test_cases.emplace_back(new test_count_equal(GGML_TYPE_F32, {4,  500, 1, 1}));
+    test_cases.emplace_back(new test_count_equal(GGML_TYPE_F32, {4, 5000, 1, 1}));
 
     test_cases.emplace_back(new test_argmax(GGML_TYPE_F32, {32,    1, 1, 1}));
     test_cases.emplace_back(new test_argmax(GGML_TYPE_F32, {100,  10, 1, 1}));
@@ -3885,8 +4005,6 @@ static std::vector> make_test_cases_eval() {
         test_cases.emplace_back(new test_repeat_back(GGML_TYPE_F32, {8, 6, 4, 2}, {1, 2, 1, 1}, view));
         test_cases.emplace_back(new test_repeat_back(GGML_TYPE_F32, {8, 6, 4, 2}, {1, 1, 2, 1}, view));
         test_cases.emplace_back(new test_repeat_back(GGML_TYPE_F32, {8, 6, 4, 2}, {1, 1, 1, 2}, view));
-        test_cases.emplace_back(new test_repeat_back(GGML_TYPE_I32, {8, 6, 4, 2}, {2, 1, 1, 1}, view));
-        test_cases.emplace_back(new test_repeat_back(GGML_TYPE_I16, {8, 6, 4, 2}, {1, 1, 1, 2}, view));
     }
 
     test_cases.emplace_back(new test_dup(GGML_TYPE_F32));
@@ -3908,14 +4026,25 @@ static std::vector> make_test_cases_eval() {
         test_cases.emplace_back(new test_set(GGML_TYPE_I32, GGML_TYPE_I32, {6, 5, 4, 3}, dim));
     }
 
-    for (ggml_type type_src : {GGML_TYPE_F16, GGML_TYPE_F32}) {
+    // same-type copy
+    for (ggml_type type : all_types) {
+        const auto nk = ggml_blck_size(type);
+
+        for (int k = 1; k < 4; ++k) {
+            test_cases.emplace_back(new test_cpy(type, type, {k*nk, 2, 3, 4}));
+            test_cases.emplace_back(new test_cpy(type, type, {k*nk, 2, 3, 4}, {0, 2, 1, 3}));
+            test_cases.emplace_back(new test_cpy(type, type, {k*nk, 2, 3, 4}, {0, 3, 1, 2}, {0, 2, 1, 3}));
+        }
+    }
+
+    for (ggml_type type_src : {GGML_TYPE_F16, GGML_TYPE_BF16, GGML_TYPE_F32}) {
         for (ggml_type type_dst : all_types) {
             test_cases.emplace_back(new test_cpy(type_src, type_dst, {256, 4, 4, 4}));
             test_cases.emplace_back(new test_cpy(type_src, type_dst, {256, 2, 3, 4}, {0, 2, 1, 3})); // cpy by rows
         }
     }
-    for (ggml_type type_dst : {GGML_TYPE_F32}) {
-        for (ggml_type type_src : all_types) {
+    for (ggml_type type_src : all_types) {
+        for (ggml_type type_dst : {GGML_TYPE_F32}) {
             test_cases.emplace_back(new test_cpy(type_src, type_dst, {256, 4, 4, 4}));
             test_cases.emplace_back(new test_cpy(type_src, type_dst, {256, 2, 3, 4}, {0, 2, 1, 3})); // cpy by rows
         }
@@ -3938,41 +4067,42 @@ static std::vector> make_test_cases_eval() {
     test_cases.emplace_back(new test_cont(GGML_TYPE_BF16, {2, 3, 5 ,7}));
 
     auto add_test_bin_bcast = [&](ggml_type type, std::array ne, std::array nr) {
-        for (auto op : {ggml_add, ggml_mul, ggml_div}) {
+        for (auto op : {ggml_add, ggml_sub, ggml_mul, ggml_div}) {
             test_cases.emplace_back(new test_bin_bcast(op, type, ne, nr));
         }
     };
+    for (ggml_type type : {GGML_TYPE_F16, GGML_TYPE_F32}) {
+        add_test_bin_bcast(type, {1, 1, 8, 1}, {1, 1, 1, 1});
+        add_test_bin_bcast(type, {1, 1, 1, 1}, {32, 1, 1, 1});
+        add_test_bin_bcast(type, {1, 1, 320, 320}, {1, 1, 1, 1});
+        add_test_bin_bcast(type, {10, 5, 1, 1}, {1, 1, 1, 1});
+        add_test_bin_bcast(type, {10, 5, 4, 1}, {1, 1, 1, 1});
+        add_test_bin_bcast(type, {10, 5, 4, 3}, {1, 1, 1, 1});
+        add_test_bin_bcast(type, {10, 5, 4, 3}, {2, 1, 1, 1});
+        add_test_bin_bcast(type, {10, 5, 4, 3}, {1, 2, 1, 1});
+        add_test_bin_bcast(type, {10, 5, 4, 3}, {1, 1, 2, 1});
+        add_test_bin_bcast(type, {10, 5, 4, 3}, {1, 1, 1, 2});
+        add_test_bin_bcast(type, {10, 5, 4, 3}, {1, 1, 2, 2});
+        add_test_bin_bcast(type, {10, 5, 4, 3}, {1, 2, 2, 2});
+        add_test_bin_bcast(type, {10, 5, 4, 3}, {2, 2, 2, 2});
 
-    add_test_bin_bcast(GGML_TYPE_F32, {1, 1, 8, 1}, {1, 1, 1, 1});
-    add_test_bin_bcast(GGML_TYPE_F32, {1, 1, 1, 1}, {32, 1, 1, 1});
-    add_test_bin_bcast(GGML_TYPE_F32, {1, 1, 320, 320}, {1, 1, 1, 1});
-    add_test_bin_bcast(GGML_TYPE_F32, {10, 5, 1, 1}, {1, 1, 1, 1});
-    add_test_bin_bcast(GGML_TYPE_F32, {10, 5, 4, 1}, {1, 1, 1, 1});
-    add_test_bin_bcast(GGML_TYPE_F32, {10, 5, 4, 3}, {1, 1, 1, 1});
-    add_test_bin_bcast(GGML_TYPE_F32, {10, 5, 4, 3}, {2, 1, 1, 1});
-    add_test_bin_bcast(GGML_TYPE_F32, {10, 5, 4, 3}, {1, 2, 1, 1});
-    add_test_bin_bcast(GGML_TYPE_F32, {10, 5, 4, 3}, {1, 1, 2, 1});
-    add_test_bin_bcast(GGML_TYPE_F32, {10, 5, 4, 3}, {1, 1, 1, 2});
-    add_test_bin_bcast(GGML_TYPE_F32, {10, 5, 4, 3}, {1, 1, 2, 2});
-    add_test_bin_bcast(GGML_TYPE_F32, {10, 5, 4, 3}, {1, 2, 2, 2});
-    add_test_bin_bcast(GGML_TYPE_F32, {10, 5, 4, 3}, {2, 2, 2, 2});
-
-    // stable diffusion
-    add_test_bin_bcast(GGML_TYPE_F32, {1280, 1, 1, 1}, {1, 1, 1, 1});
-    add_test_bin_bcast(GGML_TYPE_F32, {1280, 1, 1, 1}, {1, 16, 16, 1});
-    add_test_bin_bcast(GGML_TYPE_F32, {1280, 16, 16, 1}, {1, 1, 1, 1});
-    add_test_bin_bcast(GGML_TYPE_F32, {1280, 1, 1, 1}, {1, 256, 1, 1});
-    add_test_bin_bcast(GGML_TYPE_F32, {1, 1, 1280, 1}, {16, 16, 1, 1});
-    add_test_bin_bcast(GGML_TYPE_F32, {16, 16, 1280, 1}, {1, 1, 1, 1});
-    add_test_bin_bcast(GGML_TYPE_F32, {1, 1, 1920, 1}, {16, 16, 1, 1});
-    add_test_bin_bcast(GGML_TYPE_F32, {1, 1, 2560, 1}, {16, 16, 1, 1});
-    add_test_bin_bcast(GGML_TYPE_F32, {1, 1, 1280, 1}, {32, 32, 1, 1});
-    add_test_bin_bcast(GGML_TYPE_F32, {1, 1, 1920, 1}, {32, 32, 1, 1});
-    add_test_bin_bcast(GGML_TYPE_F32, {1, 1, 640, 1}, {32, 32, 1, 1});
-    add_test_bin_bcast(GGML_TYPE_F32, {5120, 1, 1, 1}, {1, 256, 1, 1});
-    add_test_bin_bcast(GGML_TYPE_F32, {640, 1, 1, 1}, {1, 1, 1, 1});
-    //add_test_bin_bcast(GGML_TYPE_F32, {3, 3, 2560, 1280}, {1, 1, 1, 1});
-    //add_test_bin_bcast(GGML_TYPE_F32, {3, 3, 2560, 1280}, {2, 1, 1, 1});
+        // stable diffusion
+        add_test_bin_bcast(type, {1280, 1, 1, 1}, {1, 1, 1, 1});
+        add_test_bin_bcast(type, {1280, 1, 1, 1}, {1, 16, 16, 1});
+        add_test_bin_bcast(type, {1280, 16, 16, 1}, {1, 1, 1, 1});
+        add_test_bin_bcast(type, {1280, 1, 1, 1}, {1, 256, 1, 1});
+        add_test_bin_bcast(type, {1, 1, 1280, 1}, {16, 16, 1, 1});
+        add_test_bin_bcast(type, {16, 16, 1280, 1}, {1, 1, 1, 1});
+        add_test_bin_bcast(type, {1, 1, 1920, 1}, {16, 16, 1, 1});
+        add_test_bin_bcast(type, {1, 1, 2560, 1}, {16, 16, 1, 1});
+        add_test_bin_bcast(type, {1, 1, 1280, 1}, {32, 32, 1, 1});
+        add_test_bin_bcast(type, {1, 1, 1920, 1}, {32, 32, 1, 1});
+        add_test_bin_bcast(type, {1, 1, 640, 1}, {32, 32, 1, 1});
+        add_test_bin_bcast(type, {5120, 1, 1, 1}, {1, 256, 1, 1});
+        add_test_bin_bcast(type, {640, 1, 1, 1}, {1, 1, 1, 1});
+        //add_test_bin_bcast(type, {3, 3, 2560, 1280}, {1, 1, 1, 1});
+        //add_test_bin_bcast(type, {3, 3, 2560, 1280}, {2, 1, 1, 1});
+    }
 
     test_cases.emplace_back(new test_add1());
     test_cases.emplace_back(new test_scale());
@@ -3984,8 +4114,11 @@ static std::vector> make_test_cases_eval() {
             test_cases.emplace_back(new test_rms_norm(GGML_TYPE_F32, {64, 5, 4, 3}, v, eps));
         }
         test_cases.emplace_back(new test_rms_norm_back(GGML_TYPE_F32, {64, 5, 4, 3}, eps));
+        test_cases.emplace_back(new test_l2_norm      (GGML_TYPE_F32, {64, 5, 4, 3}, eps));
     }
 
+    test_cases.emplace_back(new test_l2_norm(GGML_TYPE_F32, {64, 5, 4, 3}, 1e-12f));
+
     test_cases.emplace_back(new test_ssm_conv(GGML_TYPE_F32, {4, 1536, 1, 1}, {4, 1536, 1, 1}));
     test_cases.emplace_back(new test_ssm_conv(GGML_TYPE_F32, {8, 1536, 1, 1}, {4, 1536, 1, 1}));
     test_cases.emplace_back(new test_ssm_conv(GGML_TYPE_F32, {4, 1536, 4, 1}, {4, 1536, 1, 1}));
@@ -3997,6 +4130,11 @@ static std::vector> make_test_cases_eval() {
     test_cases.emplace_back(new test_rwkv_wkv6(GGML_TYPE_F32, 32, 64, 32, 4));
     test_cases.emplace_back(new test_rwkv_wkv6(GGML_TYPE_F32, 32, 64, 128, 4));
 
+    test_cases.emplace_back(new test_rwkv_wkv7(GGML_TYPE_F32, 32, 64, 1, 1));
+    test_cases.emplace_back(new test_rwkv_wkv7(GGML_TYPE_F32, 32, 64, 32, 1));
+    test_cases.emplace_back(new test_rwkv_wkv7(GGML_TYPE_F32, 32, 64, 32, 4));
+    test_cases.emplace_back(new test_rwkv_wkv7(GGML_TYPE_F32, 32, 64, 128, 4));
+
     test_cases.emplace_back(new test_gla(GGML_TYPE_F32, 32, 64, 1, 1));
     test_cases.emplace_back(new test_gla(GGML_TYPE_F32, 32, 64, 32, 1));
     test_cases.emplace_back(new test_gla(GGML_TYPE_F32, 32, 64, 32, 4));
@@ -4079,6 +4217,19 @@ static std::vector> make_test_cases_eval() {
     test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F32,  83, 2,   64, { 8,  1}, {4, 1}));
     test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F32,  64, 45, 128, { 8,  1}, {4, 1}));
     test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F32, 128, 45,  64, { 8,  1}, {4, 1}));
+    test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F32, 1056, 1, 193, {1,  1}, {4, 1}, {0, 2, 1, 3}));
+    test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F32, 1056, 1, 67,  {1,  1}, {4, 1}, {0, 2, 1, 3}));
+
+    for (auto bs : {1,2,4,8}) {
+        for (auto nr : {1,4}) {
+            for (uint32_t m = 0; m < 2; ++m) {
+                for (uint32_t k = 0; k < 2; ++k) {
+                    test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F32, 1056 + m, 1, 128 + k,  {bs,  1}, {nr, 1}, {0, 2, 1, 3}));
+                    test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F32, 128 + m,  1, 1056 + k, {bs,  1}, {nr, 1}, {0, 1, 2, 3}, true));
+                }
+            }
+        }
+    }
 
     // sycl backend will limit task global_range < MAX_INT
     // test case for f16-type-convert-to-fp32 kernel with large k under fp32 compute dtype (occurs in stable-diffusion)
@@ -4091,7 +4242,7 @@ static std::vector> make_test_cases_eval() {
             for (int n_mats : {4, 8}) {
                 for (int n_used : {1, 2, 4}) {
                     for (bool b : {false, true}) {
-                        for (int n : {1, 32}) {
+                        for (int n : {1, 32, 129}) {
                             int m = 512;
                             int k = 256;
                             test_cases.emplace_back(new test_mul_mat_id(type_a, type_b, n_mats, n_used, b, m, n, k));
@@ -4136,12 +4287,14 @@ static std::vector> make_test_cases_eval() {
         }
     }
 
-    test_cases.emplace_back(new test_sqr());
-    test_cases.emplace_back(new test_sqrt());
-    test_cases.emplace_back(new test_log());
-    test_cases.emplace_back(new test_sin());
-    test_cases.emplace_back(new test_cos());
-    test_cases.emplace_back(new test_clamp());
+    for (ggml_type type : {GGML_TYPE_F16, GGML_TYPE_F32}) {
+        test_cases.emplace_back(new test_sqr(type));
+        test_cases.emplace_back(new test_sqrt(type));
+        test_cases.emplace_back(new test_log(type));
+        test_cases.emplace_back(new test_sin(type));
+        test_cases.emplace_back(new test_cos(type));
+        test_cases.emplace_back(new test_clamp(type));
+    }
 
     test_cases.emplace_back(new test_diag_mask_inf(GGML_TYPE_F32, {10, 10, 1, 1}, 5));
     test_cases.emplace_back(new test_diag_mask_inf(GGML_TYPE_F32, {10, 10, 3, 1}, 5));
@@ -4257,12 +4410,15 @@ static std::vector> make_test_cases_eval() {
         test_cases.emplace_back(new test_argsort(GGML_TYPE_F32, {60, 10, 10, 10}, order)); // qwen
     }
 
+    for (ggml_scale_mode mode : {GGML_SCALE_MODE_NEAREST, GGML_SCALE_MODE_BILINEAR}) {
+        test_cases.emplace_back(new test_upscale(GGML_TYPE_F32, {512, 512, 3, 2}, 2, mode));
+        test_cases.emplace_back(new test_upscale(GGML_TYPE_F32, {512, 512, 3, 2}, 2, mode, true));
+        test_cases.emplace_back(new test_upscale_ext(GGML_TYPE_F32, {2, 5,  7, 11}, {5, 7, 11, 13}, mode));
+    }
+
     test_cases.emplace_back(new test_sum());
     test_cases.emplace_back(new test_sum_rows());
     test_cases.emplace_back(new test_mean());
-    test_cases.emplace_back(new test_upscale());
-    test_cases.emplace_back(new test_upscale(GGML_TYPE_F32, { 512, 512, 3, 1 }, 2, true));
-    test_cases.emplace_back(new test_upscale_ext());
     test_cases.emplace_back(new test_group_norm(GGML_TYPE_F32, {64, 64, 320, 1}));
     test_cases.emplace_back(new test_group_norm(GGML_TYPE_F32, {9, 9, 1280, 1}));
     test_cases.emplace_back(new test_acc());
@@ -4272,20 +4428,34 @@ static std::vector> make_test_cases_eval() {
     test_cases.emplace_back(new test_timestep_embedding());
     test_cases.emplace_back(new test_leaky_relu());
 
-    for (int hs : { 64, 80, 128, 256, }) {
-        for (bool mask : { true, false } ) {
-            for (float max_bias : { 0.0f, 8.0f }) {
-                if (!mask && max_bias > 0.0f) continue;
-                for (float logit_softcap : {0.0f, 10.0f}) {
-                    if (hs != 128 && logit_softcap != 0.0f) continue;
-                    for (int nh : { 32, }) {
-                        for (int kv : { 512, 1024, }) {
-                            for (int nb : { 1, 3, 32, 35, }) {
-                                for (ggml_type type_KV : {GGML_TYPE_F16, GGML_TYPE_BF16, GGML_TYPE_Q8_0, GGML_TYPE_Q4_0}) {
-                                    test_cases.emplace_back(new test_flash_attn_ext(hs, nh, kv, nb, mask, max_bias, logit_softcap, type_KV));
-                                    // run fewer test cases permuted
-                                    if (mask == true && max_bias == 0.0f && logit_softcap == 0 && kv == 512) {
-                                        test_cases.emplace_back(new test_flash_attn_ext(hs, nh, kv, nb, mask, max_bias, logit_softcap, type_KV, {0, 2, 1, 3}));
+    for (int hsk : { 64, 80, 128, 192, 256, }) {
+        for (int hsv : { 64, 80, 128, 192, 256, }) {
+            if (hsk != 192 && hsk != hsv) continue;
+            if (hsk == 192 && (hsv != 128 && hsv != 192)) continue;
+
+            for (bool mask : { true, false } ) {
+                for (float max_bias : { 0.0f, 8.0f }) {
+                    if (!mask && max_bias > 0.0f) continue;
+                    for (float logit_softcap : {0.0f, 10.0f}) {
+                        if (hsk != 128 && logit_softcap != 0.0f) continue;
+                        for (int nh : { 4, }) {
+                            for (int nr : { 1, 4, 16 }) {
+                                if (nr == 16 && hsk != 128) continue;
+                                for (int kv : { 512, 1024, }) {
+                                    if (nr != 1 && kv != 512) continue;
+                                    for (int nb : { 1, 3, 32, 35, }) {
+                                        for (ggml_prec prec : {GGML_PREC_F32, GGML_PREC_DEFAULT}) {
+                                            if (hsk != 128 && prec == GGML_PREC_DEFAULT) continue;
+                                            for (ggml_type type_KV : {GGML_TYPE_F16, GGML_TYPE_BF16, GGML_TYPE_Q8_0, GGML_TYPE_Q4_0}) {
+                                                test_cases.emplace_back(new test_flash_attn_ext(
+                                                    hsk, hsv, nh, nr, kv, nb, mask, max_bias, logit_softcap, prec, type_KV));
+                                                // run fewer test cases permuted
+                                                if (mask == true && max_bias == 0.0f && logit_softcap == 0 && kv == 512) {
+                                                    test_cases.emplace_back(new test_flash_attn_ext(
+                                                        hsk, hsv, nh, nr, kv, nb, mask, max_bias, logit_softcap, prec, type_KV, {0, 2, 1, 3}));
+                                                }
+                                            }
+                                        }
                                     }
                                 }
                             }
@@ -4337,6 +4507,9 @@ static std::vector> make_test_cases_perf() {
     test_cases.emplace_back(new test_argmax(GGML_TYPE_F32, {1024, 10, 1, 1}));
     test_cases.emplace_back(new test_argmax(GGML_TYPE_F32, {32000, 512, 1, 1}));
 
+    test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F32, 16416, 1, 128, {8,  1}, {4, 1}, {0, 2, 1, 3}));
+    test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F32, 128, 1, 16416, {8,  1}, {4, 1}, {0, 1, 2, 3}, true));
+
     for (int bs : {1, 2, 3, 4, 5, 8, 512}) {
         for (ggml_type type_a : all_types) {
             for (ggml_type type_b : {GGML_TYPE_F32}) {
@@ -4357,12 +4530,36 @@ static std::vector> make_test_cases_perf() {
         }
     }
 
+    for (int kv : { 4096, 8192, 16384, }) {
+        for (int hs : { 64, 128, }) {
+            test_cases.emplace_back(new test_flash_attn_ext(hs, hs, 8, 4, kv, 1, true, 0, 0, GGML_PREC_F32, GGML_TYPE_F16));
+        }
+    }
+
     return test_cases;
 }
 
-static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op_name) {
+static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op_name, const char * params_filter) {
+    auto filter_test_cases = [](std::vector> & test_cases, const char * params_filter) {
+        if (params_filter == nullptr) {
+            return;
+        }
+
+        std::regex params_filter_regex(params_filter);
+
+        for (auto it = test_cases.begin(); it != test_cases.end();) {
+            if (!std::regex_search((*it)->vars(), params_filter_regex)) {
+                it = test_cases.erase(it);
+                continue;
+            }
+
+            it++;
+        }
+    };
+
     if (mode == MODE_TEST) {
         auto test_cases = make_test_cases_eval();
+        filter_test_cases(test_cases, params_filter);
         ggml_backend_t backend_cpu = ggml_backend_init_by_type(GGML_BACKEND_DEVICE_TYPE_CPU, NULL);
         if (backend_cpu == NULL) {
             printf("  Failed to initialize CPU backend\n");
@@ -4384,6 +4581,7 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op
 
     if (mode == MODE_GRAD) {
         auto test_cases = make_test_cases_eval();
+        filter_test_cases(test_cases, params_filter);
         size_t n_ok = 0;
         for (auto & test : test_cases) {
             if (test->eval_grad(backend, op_name)) {
@@ -4397,6 +4595,7 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op
 
     if (mode == MODE_PERF) {
         auto test_cases = make_test_cases_perf();
+        filter_test_cases(test_cases, params_filter);
         for (auto & test : test_cases) {
             test->eval_perf(backend, op_name);
         }
@@ -4407,7 +4606,7 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op
 }
 
 static void usage(char ** argv) {
-    printf("Usage: %s [mode] [-o op] [-b backend]\n", argv[0]);
+    printf("Usage: %s [mode] [-o ] [-b ] [-p ]\n", argv[0]);
     printf("    valid modes:\n");
     printf("      - test (default, compare with CPU backend for correctness)\n");
     printf("      - grad (compare gradients from backpropagation with method of finite differences)\n");
@@ -4417,8 +4616,9 @@ static void usage(char ** argv) {
 
 int main(int argc, char ** argv) {
     test_mode mode = MODE_TEST;
-    const char * op_name_filter = NULL;
-    const char * backend_filter = NULL;
+    const char * op_name_filter = nullptr;
+    const char * backend_filter = nullptr;
+    const char * params_filter = nullptr;
 
     for (int i = 1; i < argc; i++) {
         if (strcmp(argv[i], "test") == 0) {
@@ -4441,6 +4641,13 @@ int main(int argc, char ** argv) {
                 usage(argv);
                 return 1;
             }
+        } else if (strcmp(argv[i], "-p") == 0) {
+            if (i + 1 < argc) {
+                params_filter = argv[++i];
+            } else {
+                usage(argv);
+                return 1;
+            }
         } else {
             usage(argv);
             return 1;
@@ -4487,7 +4694,7 @@ int main(int argc, char ** argv) {
         printf("  Device memory: %zu MB (%zu MB free)\n", total / 1024 / 1024, free / 1024 / 1024);
         printf("\n");
 
-        bool ok = test_backend(backend, mode, op_name_filter);
+        bool ok = test_backend(backend, mode, op_name_filter, params_filter);
 
         printf("  Backend %s: ", ggml_backend_name(backend));
         if (ok) {
diff --git a/tests/test-chat-template.cpp b/tests/test-chat-template.cpp
index e0314ae1d..be1a64006 100644
--- a/tests/test-chat-template.cpp
+++ b/tests/test-chat-template.cpp
@@ -1,13 +1,14 @@
 #include 
 #include 
 #include 
+#include 
 
 #undef NDEBUG
 #include 
 
 #include "llama.h"
 #include "common.h"
-#include "chat-template.hpp"
+#include "chat.h"
 
 static std::string normalize_newlines(const std::string & s) {
 #ifdef _WIN32
@@ -18,6 +19,15 @@ static std::string normalize_newlines(const std::string & s) {
 #endif
 }
 
+#define U8C(x) (const char*)(u8##x)
+
+static common_chat_msg simple_msg(const std::string & role, const std::string & content) {
+    common_chat_msg msg;
+    msg.role = role;
+    msg.content = content;
+    return msg;
+}
+
 int main(void) {
     std::vector conversation {
         {"system", "You are a helpful assistant"},
@@ -27,6 +37,8 @@ int main(void) {
         {"assistant", "   I am an assistant   "},
         {"user", "Another question"},
     };
+
+    // std::string wrong = /* .template_str= */ u8"[gMASK]{% for item in messages %}{% if item['tools'] is defined %}<|system|>\n你是一个名为 ChatGLM 的人工智能助手。你是基于智谱AI训练的语言模型 GLM-4 模型开发的，你的任务是针对用户的问题和要求提供适当的答复和支持。\n\n# 可用工具{% set tools = item['tools'] %}{% for tool in tools %}{% if tool['type'] == 'function' %}\n\n## {{ tool['function']['name'] }}\n\n{{ tool['function'] | tojson(indent=4) }}\n......{% endif %}{% endfor %}{% endif %}{% if item['content'] %}<|{{ item['role'] }}|>{{ item['metadata'] }}\n{{ item['content'] }}{% endif %}{% endfor %}{% if add_generation_prompt %}<|assistant|>{% endif %}";
     struct TestCase {
         std::string name;
         std::string template_str;
@@ -50,7 +62,7 @@ int main(void) {
             /* .template_str= */ "{{ bos_token }}{% for message in messages %}{% if (message['role'] == 'user') != (loop.index0 % 2 == 0) %}{{ raise_exception('Conversation roles must alternate user/assistant/user/assistant/...') }}{% endif %}{% if message['role'] == 'user' %}{{ '[INST] ' + message['content'] + ' [/INST]' }}{% elif message['role'] == 'assistant' %}{{ message['content'] + eos_token}}{% else %}{{ raise_exception('Only user and assistant roles are supported!') }}{% endif %}{% endfor %}",
             /* .expected_output= */ "[INST] You are a helpful assistant\nHello [/INST]Hi there[INST] Who are you [/INST]   I am an assistant   [INST] Another question [/INST]",
             /* .expected_output_jinja= */ "",
-            /* .bos_token= */ "",
+            /* .bos_token= */ "",
             /* .eos_token= */ "",
         },
         {
@@ -72,8 +84,8 @@ int main(void) {
         {
             /* .name= */ "mlabonne/AlphaMonarch-7B",
             /* .template_str= */ "{% for message in messages %}{{bos_token + message['role'] + '\\n' + message['content'] + eos_token + '\\n'}}{% endfor %}{% if add_generation_prompt %}{{ bos_token + 'assistant\\n' }}{% endif %}",
-            /* .expected_output= */          "system\nYou are a helpful assistant\nuser\nHello\nassistant\nHi there\nuser\nWho are you\nassistant\n   I am an assistant   \nuser\nAnother question\nassistant\n",
-            /* .expected_output_jinja= */ "system\nYou are a helpful assistant\nuser\nHello\nassistant\nHi there\nuser\nWho are you\nassistant\n   I am an assistant   \nuser\nAnother question\nassistant\n",
+            /* .expected_output= */ "system\nYou are a helpful assistant\nuser\nHello\nassistant\nHi there\nuser\nWho are you\nassistant\n   I am an assistant   \nuser\nAnother question\nassistant\n",
+            /* .expected_output_jinja= */ "",
             /* .bos_token= */ "",
             /* .eos_token= */ "",
         },
@@ -87,7 +99,7 @@ int main(void) {
             /* .name= */ "OrionStarAI/Orion-14B-Chat",
             /* .template_str= */ "{% for message in messages %}{% if loop.first %}{{ bos_token }}{% endif %}{% if message['role'] == 'user' %}{{ 'Human: ' + message['content'] + '\\n\\nAssistant: ' + eos_token }}{% elif message['role'] == 'assistant' %}{{ message['content'] + eos_token }}{% endif %}{% endfor %}",
             /* .expected_output= */       "Human: You are a helpful assistant\n\nHello\n\nAssistant: Hi thereHuman: Who are you\n\nAssistant:    I am an assistant   Human: Another question\n\nAssistant: ",
-            /* .expected_output_jinja= */ "Human: You are a helpful assistant\nHello\n\nAssistant: Hi thereHuman: Who are you\n\nAssistant:    I am an assistant   Human: Another question\n\nAssistant: ",
+            /* .expected_output_jinja= */ "Human: You are a helpful assistant\nHello\n\nAssistant: Hi thereHuman: Who are you\n\nAssistant:    I am an assistant   Human: Another question\n\nAssistant: ",
             /* .bos_token= */ "",
             /* .eos_token= */ "",
         },
@@ -169,7 +181,7 @@ int main(void) {
         },
         {
             /* .name= */ "ChatGLM4",
-            /* .template_str= */ u8"[gMASK]{% for item in messages %}{% if item['tools'] is defined %}<|system|>\n你是一个名为 ChatGLM 的人工智能助手。你是基于智谱AI训练的语言模型 GLM-4 模型开发的，你的任务是针对用户的问题和要求提供适当的答复和支持。\n\n# 可用工具{% set tools = item['tools'] %}{% for tool in tools %}{% if tool['type'] == 'function' %}\n\n## {{ tool['function']['name'] }}\n\n{{ tool['function'] | tojson(indent=4) }}\n......{% endif %}{% endfor %}{% endif %}{% if item['content'] %}<|{{ item['role'] }}|>{{ item['metadata'] }}\n{{ item['content'] }}{% endif %}{% endfor %}{% if add_generation_prompt %}<|assistant|>{% endif %}",
+            /* .template_str= */ U8C("[gMASK]{% for item in messages %}{% if item['tools'] is defined %}<|system|>\n你是一个名为 ChatGLM 的人工智能助手。你是基于智谱AI训练的语言模型 GLM-4 模型开发的，你的任务是针对用户的问题和要求提供适当的答复和支持。\n\n# 可用工具{% set tools = item['tools'] %}{% for tool in tools %}{% if tool['type'] == 'function' %}\n\n## {{ tool['function']['name'] }}\n\n{{ tool['function'] | tojson(indent=4) }}\n......{% endif %}{% endfor %}{% endif %}{% if item['content'] %}<|{{ item['role'] }}|>{{ item['metadata'] }}\n{{ item['content'] }}{% endif %}{% endfor %}{% if add_generation_prompt %}<|assistant|>{% endif %}"),
             /* .expected_output= */ "[gMASK]<|system|>\nYou are a helpful assistant<|user|>\nHello<|assistant|>\nHi there<|user|>\nWho are you<|assistant|>\n   I am an assistant   <|user|>\nAnother question<|assistant|>",
             /* .expected_output_jinja= */ "",
             /* .bos_token= */ "",
@@ -185,8 +197,8 @@ int main(void) {
         },
         {
             /* .name= */ "MiniCPM-3B-OpenHermes-2.5-v2-GGUF",
-            /* .template_str= */ u8"{% for message in messages %}{% if message['role'] == 'user' %}{{'<用户>' + message['content'].strip() + ''}}{% else %}{{message['content'].strip()}}{% endif %}{% endfor %}",
-            /* .expected_output= */ u8"You are a helpful assistant<用户>HelloHi there<用户>Who are youI am an assistant<用户>Another question",
+            /* .template_str= */ U8C("{% for message in messages %}{% if message['role'] == 'user' %}{{'<用户>' + message['content'].strip() + ''}}{% else %}{{message['content'].strip()}}{% endif %}{% endfor %}"),
+            /* .expected_output= */ U8C("You are a helpful assistant<用户>HelloHi there<用户>Who are youI am an assistant<用户>Another question"),
             /* .expected_output_jinja= */ "",
             /* .bos_token= */ "",
             /* .eos_token= */ "",
@@ -194,7 +206,7 @@ int main(void) {
         {
             /* .name= */ "DeepSeek-V2",
             /* .template_str= */ "{% if not add_generation_prompt is defined %}{% set add_generation_prompt = false %}{% endif %}{{ bos_token }}{% for message in messages %}{% if message['role'] == 'user' %}{{ 'User: ' + message['content'] + '\n\n' }}{% elif message['role'] == 'assistant' %}{{ 'Assistant: ' + message['content'] + eos_token }}{% elif message['role'] == 'system' %}{{ message['content'] + '\n\n' }}{% endif %}{% endfor %}{% if add_generation_prompt %}{{ 'Assistant:' }}{% endif %}",
-            /* .expected_output= */ u8"You are a helpful assistant\n\nUser: Hello\n\nAssistant: Hi there<｜end▁of▁sentence｜>User: Who are you\n\nAssistant:    I am an assistant   <｜end▁of▁sentence｜>User: Another question\n\nAssistant:",
+            /* .expected_output= */ U8C("You are a helpful assistant\n\nUser: Hello\n\nAssistant: Hi there<｜end▁of▁sentence｜>User: Who are you\n\nAssistant:    I am an assistant   <｜end▁of▁sentence｜>User: Another question\n\nAssistant:"),
             /* .expected_output_jinja= */ "",
             /* .bos_token= */ "",
             /* .eos_token= */ "<｜end▁of▁sentence｜>",
@@ -248,7 +260,7 @@ int main(void) {
         },
         {
             /* .name= */ "Infinigence/Megrez-3B-Instruct",
-            /* .template_str= */ u8"{% for message in messages %}{% if loop.first and messages[0]['role'] != 'system' %}{{ '<|role_start|>system<|role_end|>你是Megrez-3B-Instruct，将针对用户的问题给出详细的、积极的回答。<|turn_end|>' }}{% endif %}{{ '<|role_start|>' + message['role'] + '<|role_end|>' + message['content'] + '<|turn_end|>' }}{% endfor %}{% if add_generation_prompt %}{{ '<|role_start|>assistant<|role_end|>' }}{% endif %}",
+            /* .template_str= */ U8C("{% for message in messages %}{% if loop.first and messages[0]['role'] != 'system' %}{{ '<|role_start|>system<|role_end|>你是Megrez-3B-Instruct，将针对用户的问题给出详细的、积极的回答。<|turn_end|>' }}{% endif %}{{ '<|role_start|>' + message['role'] + '<|role_end|>' + message['content'] + '<|turn_end|>' }}{% endfor %}{% if add_generation_prompt %}{{ '<|role_start|>assistant<|role_end|>' }}{% endif %}"),
             /* .expected_output= */ "<|role_start|>system<|role_end|>You are a helpful assistant<|turn_end|><|role_start|>user<|role_end|>Hello<|turn_end|><|role_start|>assistant<|role_end|>Hi there<|turn_end|><|role_start|>user<|role_end|>Who are you<|turn_end|><|role_start|>assistant<|role_end|>   I am an assistant   <|turn_end|><|role_start|>user<|role_end|>Another question<|turn_end|><|role_start|>assistant<|role_end|>",
             /* .expected_output_jinja= */ "",
             /* .bos_token= */ "",
@@ -262,6 +274,22 @@ int main(void) {
             /* .bos_token= */ "",
             /* .eos_token= */ "",
         },
+        {
+            /* .name= */ "yandex/YandexGPT-5-Lite-8B-instruct",
+            /* .template_str= */ "{%- set names = {'assistant': ' Ассистент:', 'user': ' Пользователь:'} %}\n{%- set tools_prefix = 'Тебе доступны следующие функции:' %}\n{%- macro __render_tool(tool) %}\n    {%- set name = tool.function.name %}\n    {%- set description = tool.function.description|default('') %}\n    {%- set parameters = tool.function.parameters|tojson %}\n    {{- '\\n' }}function {{ '{' }}'name':'{{ name }}',\n    {%- if tool.function.description %}'description':'{{ description }}',{% endif %}\n'parameters':{{ parameters }}\n    {{- '}' }}\n{%- endmacro %}\n{%- macro __render_tools(tools) %}\n    {{- tools_prefix }}\n    {%- for tool in tools %}\n        {{- __render_tool(tool) }}\n    {%- endfor %}\n    {{- '\\n\\n' }}\n{%- endmacro %}\n{%- macro __render_tool_message(message) %}\n    {{- '\\n\\nРезультат вызова' }} {{ message.name }}: {{ message.content }} {{ '\\n\\n' }}\n{%- endmacro %}\n{%- if tools -%}\n    {{- __render_tools(tools) }}\n{%- endif -%}\n{%- macro __render_user_message(message) %}\n{{ names.user }} {{ message.content + '\\n\\n' }}\n{%- endmacro %}\n{%- macro __render_assistant_message(message) %}\n    {{- names.assistant }}\n    {%- set call = message['function_call'] %}\n    {%- if call %}\n        {{- '\\n[TOOL_CALL_START]' }}{{ call.name }}{{ '\\n' }}{{ call.arguments|tojson }}\n    {%- else %}\n        {{- ' ' + message.content + '\\n\\n' }}\n    {%- endif %}\n{%- endmacro %}\n{%- if not add_generation_prompt is defined %}\n{%- set add_generation_prompt = false %}\n{%- endif %}\n{%- for message in messages %}\n    {%- if message['role'] == 'user' %}\n        {{- __render_user_message(message) }}\n    {%- endif %}\n    {%- if message.role == 'assistant' and not loop.last %}\n        {{- __render_assistant_message(message) }}\n    {%- endif %}\n    {%- if message.role == 'tool' %}\n        {{- __render_tool_message(message) }}\n    {%- endif %}\n    {%- if loop.last %}\n        {{- ' Ассистент:[SEP]' }}\n    {%- endif %}\n{%- endfor %}\n",
+            /* .expected_output= */ " Пользователь: Hello\n\n Ассистент: Hi there\n\n Пользователь: Who are you\n\n Ассистент:    I am an assistant   \n\n Пользователь: Another question\n\n Ассистент:[SEP]",
+            /* .expected_output_jinja= */ " Пользователь: You are a helpful assistant\nHello\n\n Ассистент: Hi there\n\n Пользователь: Who are you\n\n Ассистент:    I am an assistant   \n\n Пользователь: Another question\n\n Ассистент:[SEP]",
+            /* .bos_token= */ "",
+            /* .eos_token= */ "",
+        },
+        {
+            /* .name= */ "inclusionAI/Ling-lite",
+            /* .template_str */ "{% for message in messages %}{% set role = message['role'] | lower %}{% if role == 'user' %}{% set role = 'HUMAN' %}{% endif %}{% set role = role | upper %}{{ '' + role + '' + message['content'] }}{% endfor %}{% if add_generation_prompt %}{{ 'ASSISTANT' }}{% endif %}",
+            /* .expected_output= */ "SYSTEMYou are a helpful assistantHUMANHelloASSISTANTHi thereHUMANWho are youASSISTANT   I am an assistant   HUMANAnother questionASSISTANT",
+            /* .expected_output_jinja= */ "",
+            /* .bos_token= */ "",
+            /* .eos_token= */ "",
+        },
     };
     std::vector formatted_chat(1024);
     int32_t res;
@@ -304,12 +332,9 @@ int main(void) {
         }
     }
 
-    json messages = json::array();
+    std::vector messages;
     for (const auto & msg : conversation) {
-        messages.push_back({
-            {"role", msg.role},
-            {"content", msg.content},
-        });
+        messages.push_back(simple_msg(msg.role, msg.content));
     }
     for (const auto & test_case : test_cases) {
         if (!test_case.supported_with_jinja) {
@@ -317,8 +342,13 @@ int main(void) {
         }
         printf("\n\n=== %s (jinja) ===\n\n", test_case.name.c_str());
         try {
-            minja::chat_template tmpl(test_case.template_str, test_case.bos_token, test_case.eos_token);
-            auto output = normalize_newlines(tmpl.apply(messages, json(), add_generation_prompt));
+            auto tmpls = common_chat_templates_init(/* model= */ nullptr, test_case.template_str.c_str(), test_case.bos_token, test_case.eos_token);
+            common_chat_templates_inputs inputs;
+            inputs.use_jinja = true;
+            inputs.messages = messages;
+            inputs.add_generation_prompt = add_generation_prompt;
+            auto output = common_chat_templates_apply(tmpls.get(), inputs).prompt;
+            output = normalize_newlines(output);
             auto expected_output = normalize_newlines(test_case.expected_output_jinja.empty() ? test_case.expected_output : test_case.expected_output_jinja);
             if (output != expected_output) {
                 printf("Expected:\n%s\n", expected_output.c_str());
@@ -336,11 +366,11 @@ int main(void) {
     // test llama_chat_format_single for system message
     printf("\n\n=== llama_chat_format_single (system message) ===\n\n");
     std::vector chat2;
-    common_chat_msg sys_msg{"system", "You are a helpful assistant", {}};
+    auto sys_msg = simple_msg("system", "You are a helpful assistant");
 
     auto fmt_sys = [&](std::string tmpl_str) {
-        minja::chat_template tmpl(tmpl_str, "", "");
-        auto output = common_chat_format_single(tmpl, chat2, sys_msg, false, /* use_jinja= */ false);
+        auto tmpls = common_chat_templates_init(/* model= */ nullptr, tmpl_str);
+        auto output = common_chat_format_single(tmpls.get(), chat2, sys_msg, false, /* use_jinja= */ false);
         printf("fmt_sys(%s) : %s\n", tmpl_str.c_str(), output.c_str());
         printf("-------------------------\n");
         return output;
@@ -360,14 +390,14 @@ int main(void) {
 
     // test llama_chat_format_single for user message
     printf("\n\n=== llama_chat_format_single (user message) ===\n\n");
-    chat2.push_back({"system", "You are a helpful assistant", {}});
-    chat2.push_back({"user", "Hello", {}});
-    chat2.push_back({"assistant", "I am assistant", {}});
-    common_chat_msg new_msg{"user", "How are you", {}};
+    chat2.push_back(simple_msg("system", "You are a helpful assistant"));
+    chat2.push_back(simple_msg("user", "Hello"));
+    chat2.push_back(simple_msg("assistant", "I am assistant"));
+    auto new_msg = simple_msg("user", "How are you");
 
-    auto fmt_single = [&](std::string tmpl_str) {
-        minja::chat_template tmpl(tmpl_str, "", "");
-        auto output = common_chat_format_single(tmpl, chat2, new_msg, true, /* use_jinja= */ false);
+    auto fmt_single = [&](const std::string & tmpl_str) {
+        auto tmpls = common_chat_templates_init(/* model= */ nullptr, tmpl_str.c_str());
+        auto output = common_chat_format_single(tmpls.get(), chat2, new_msg, true, /* use_jinja= */ false);
         printf("fmt_single(%s) : %s\n", tmpl_str.c_str(), output.c_str());
         printf("-------------------------\n");
         return output;
diff --git a/tests/test-chat.cpp b/tests/test-chat.cpp
index b78da2cdb..a0bf6affe 100644
--- a/tests/test-chat.cpp
+++ b/tests/test-chat.cpp
@@ -10,35 +10,12 @@
 #include 
 #include 
 
-#include "chat-template.hpp"
-#include "chat.hpp"
+#include "chat.h"
 #include "llama-grammar.h"
 #include "unicode.h"
 
 using json = nlohmann::ordered_json;
 
-static common_chat_msg msg_from_json(const json & message) {
-    common_chat_msg ret;
-    ret.role = "assistant";
-    if (message.contains("content") && !message.at("content").is_null()) {
-        ret.content = message.at("content");
-    }
-    if (message.contains("tool_plan")) {
-        ret.tool_plan = message.at("tool_plan");
-    }
-    auto has_tool_calls = message.contains("tool_calls");
-    if (has_tool_calls) {
-        for (const auto & tc : message.at("tool_calls")) {
-            const auto & arguments = tc.at("function").at("arguments");
-            ret.tool_calls.push_back({
-                tc.at("function").at("name").get(),
-                arguments.is_string() ? arguments.get() : arguments.dump(),
-                tc.contains("id") ? tc.at("id").get() : "",
-            });
-        }
-    }
-    return ret;
-}
 
 template  static void assert_equals(const T & expected, const T & actual) {
     if (expected != actual) {
@@ -50,7 +27,7 @@ template  static void assert_equals(const T & expected, const T & actua
 }
 
 static std::string read_file(const std::string & path) {
-    std::cerr << "# Reading: " << path << std::endl << std::flush;
+    std::cerr << "# Reading: " << path << '\n' << std::flush;
     std::ifstream fs(path, std::ios_base::binary);
     if (!fs.is_open()) {
         fs = std::ifstream("../" + path, std::ios_base::binary);
@@ -63,10 +40,14 @@ static std::string read_file(const std::string & path) {
     fs.seekg(0);
     std::string out;
     out.resize(static_cast(size));
-    fs.read(&out[0], static_cast(size));
+    fs.read(out.data(), static_cast(size));
     return out;
 }
 
+static common_chat_templates_ptr read_templates(const std::string & path) {
+    return common_chat_templates_ptr(common_chat_templates_init(/* model= */ nullptr, read_file(path)));
+}
+
 static std::unique_ptr build_grammar(const std::string & grammar_str) {
     return std::unique_ptr(
         llama_grammar_init_impl(nullptr, grammar_str.c_str(), "root", false, nullptr, 0, nullptr, 0));
@@ -87,122 +68,124 @@ static bool match_string(const std::string & input, llama_grammar * grammar) {
         }
     }
 
-    for (const auto & stack : stacks_cur) {
-        if (stack.empty()) {
-            // An empty stack means that the grammar has been completed
-            return true;
-        }
+    if (std::any_of(stacks_cur.begin(), stacks_cur.end(), [](const auto & stack) { return stack.empty(); })) {
+        // An empty stack means that the grammar has been completed
+        return true;
     }
 
     return false;
 }
 
-// Dumps `{"a": 1}` as `"{\"a\": 1}"`, unlike nlohmann::json::dump which would dump it as `"{\"a\":1}"`.
-static std::string dump(const json & j) {
-    return minja::Value(j).dump(-1, /* to_json= */ true);
-}
-
 static void assert_msg_equals(const common_chat_msg & expected, const common_chat_msg & actual) {
     assert_equals(expected.role, actual.role);
     assert_equals(expected.content, actual.content);
+    assert_equals(expected.content_parts.size(), actual.content_parts.size());
+    for (size_t i = 0; i < expected.content_parts.size(); i++) {
+        const auto & expected_part = expected.content_parts[i];
+        const auto & actual_part   = actual.content_parts[i];
+        assert_equals(expected_part.type, actual_part.type);
+        assert_equals(expected_part.text, actual_part.text);
+    }
+    assert_equals(expected.reasoning_content, actual.reasoning_content);
     assert_equals(expected.tool_calls.size(), actual.tool_calls.size());
     for (size_t i = 0; i < expected.tool_calls.size(); i++) {
         const auto & expected_tool_call = expected.tool_calls[i];
         const auto & actual_tool_call   = actual.tool_calls[i];
         assert_equals(expected_tool_call.name, actual_tool_call.name);
-        assert_equals(dump(json::parse(expected_tool_call.arguments)), dump(json::parse(actual_tool_call.arguments)));
+        assert_equals(json::parse(expected_tool_call.arguments).dump(), json::parse(actual_tool_call.arguments).dump());
         assert_equals(expected_tool_call.id, actual_tool_call.id);
     }
 }
 
-const auto special_function_tool = json::parse(R"({
-  "type": "function",
-  "function": {
-    "name": "special_function",
-    "description": "I'm special",
-    "parameters": {
-      "type": "object",
-      "properties": {
-        "arg1": {
-          "type": "integer",
-          "description": "The arg."
-        }
-      },
-      "required": ["arg1"]
-    }
-  }
-})");
-const auto python_tool           = json::parse(R"({
-  "type": "function",
-  "function": {
-    "name": "python",
-    "description": "an ipython interpreter",
-    "parameters": {
-      "type": "object",
-      "properties": {
-        "code": {
-          "type": "string",
-          "description": "Python code to execute."
-        }
-      },
-      "required": ["code"]
-    }
-  }
-})");
-const auto code_interpreter_tool = json::parse(R"({
-  "type": "function",
-  "function": {
-    "name": "code_interpreter",
-    "description": "an ipython interpreter",
-    "parameters": {
-      "type": "object",
-      "properties": {
-        "code": {
-          "type": "string",
-          "description": "Python code to execute."
-        }
-      },
-      "required": ["code"]
-    }
-  }
-})");
-const json tools                 = { special_function_tool, python_tool };
-const json llama_3_1_tools       = { special_function_tool, code_interpreter_tool };
+common_chat_tool special_function_tool {
+    /* .name = */ "special_function",
+    /* .description = */ "I'm special",
+    /* .parameters = */ R"({
+        "type": "object",
+        "properties": {
+            "arg1": {
+                "type": "integer",
+                "description": "The arg."
+            }
+        },
+        "required": ["arg1"]
+    })",
+};
+common_chat_tool python_tool {
+    /* .name = */ "python",
+    /* .description = */ "an ipython interpreter",
+    /* .parameters = */ R"({
+        "type": "object",
+        "properties": {
+            "code": {
+                "type": "string",
+                "description": "Python code to execute."
+            }
+        },
+        "required": ["code"]
+    })",
+};
+common_chat_tool code_interpreter_tool {
+    /* .name = */ "code_interpreter",
+    /* .description = */ "an ipython interpreter",
+    /* .parameters = */ R"({
+        "type": "object",
+        "properties": {
+            "code": {
+                "type": "string",
+                "description": "Python code to execute."
+            }
+        },
+        "required": ["code"]
+    })",
+};
+std::vector tools           { special_function_tool, python_tool };
+std::vector llama_3_1_tools { special_function_tool, code_interpreter_tool };
 
 struct delta_data {
     std::string        delta;
     common_chat_params params;
 };
 
-static delta_data init_delta(const common_chat_template & tmpl, const std::vector & end_tokens,
-                             const json & user_message, const json & delta_message, const json & tools,
-                             const json & tool_choice) {
-    common_chat_inputs inputs;
+static delta_data init_delta(const struct common_chat_templates * tmpls, const std::vector & end_tokens,
+                             const common_chat_msg & user_message,
+                             const common_chat_msg & delta_message,
+                             const std::vector & tools,
+                             const common_chat_tool_choice & tool_choice,
+                             bool think = false) {
+    common_chat_templates_inputs inputs;
     inputs.parallel_tool_calls = true;
-    inputs.messages            = json::array();
     inputs.messages.push_back(user_message);
     inputs.tools       = tools;
     inputs.tool_choice = tool_choice;
-    auto params_prefix = common_chat_params_init(tmpl, inputs);
+    inputs.extract_reasoning = think;
+    auto params_prefix = common_chat_templates_apply(tmpls, inputs);
 
     inputs.messages.push_back(delta_message);
     inputs.add_generation_prompt = false;
-    auto params_full             = common_chat_params_init(tmpl, inputs);
+    auto params_full             = common_chat_templates_apply(tmpls, inputs);
 
     std::string prefix = params_prefix.prompt;
     std::string full   = params_full.prompt;
 
-    // Check full starts with prefix
-    if (full.find(prefix) != 0) {
-        fprintf(stderr, "Full:\n%s\n\nPrefix:\n%s\n\n", full.c_str(), prefix.c_str());
-        throw std::runtime_error("Full message does not start with prefix");
-    }
-
     if (full == prefix) {
         throw std::runtime_error("Full message is the same as the prefix");
     }
 
-    auto delta = full.substr(prefix.size());
+    size_t common_prefix_length = 0;
+    for (size_t i = 0; i < prefix.size() && i < full.size(); ++i) {
+        if (prefix[i] != full[i]) {
+            break;
+        }
+        if (prefix[i] == '<') {
+            // DeepSeek R1's template (as of 20250209) adds a trailing  if add_generation_prompt,
+            // but it removes thinking tags for past messages.
+            // The prefix and full strings diverge at  vs. <｜tool▁calls▁begin｜>, we avoid consuming the leading <.
+            continue;
+        }
+        common_prefix_length = i + 1;
+    }
+    auto delta = full.substr(common_prefix_length);
 
     // Strip end tokens
     for (const auto & end_token : end_tokens) {
@@ -221,28 +204,29 @@ static delta_data init_delta(const common_chat_template & tmpl, const std::vecto
   gets the diff, removes any end tokens and parses the result w/ the grammar, checking that
   the parsed message is the same as the test_message
 */
-static void test_template(const common_chat_template & tmpl, const std::vector & end_tokens,
-                          const json & test_message, const json & tools = {}, const std::string & expected_delta = "",
-                          bool expect_grammar_triggered = true) {
-    common_chat_msg expected_msg = msg_from_json(test_message);
+static void test_templates(const struct common_chat_templates * tmpls, const std::vector & end_tokens,
+                          const common_chat_msg & test_message,
+                          const std::vector & tools = {},
+                          const std::string & expected_delta = "",
+                          bool expect_grammar_triggered = true,
+                          bool test_grammar_if_triggered = true,
+                          bool think = false) {
+    common_chat_msg user_message;
+    user_message.role = "user";
+    user_message.content = "Hello, world!";
 
-    auto user_message = json{
-        { "role",    "user"          },
-        { "content", "Hello, world!" }
-    };
-
-    for (const auto & tool_choice : json({ "auto", "required" })) {
-        auto data = init_delta(tmpl, end_tokens, user_message, test_message, tools, tool_choice);
+    for (const auto & tool_choice : std::vector {COMMON_CHAT_TOOL_CHOICE_AUTO, COMMON_CHAT_TOOL_CHOICE_REQUIRED}) {
+        auto data = init_delta(tmpls, end_tokens, user_message, test_message, tools, tool_choice, think);
         if (!expected_delta.empty()) {
             assert_equals(expected_delta, data.delta);
         }
 
         if (expect_grammar_triggered) {
             const auto msg = common_chat_parse(data.delta, data.params.format);
-            assert_msg_equals(expected_msg, msg);
+            assert_msg_equals(test_message, msg);
         }
 
-        if (!expected_msg.tool_calls.empty()) {
+        if (!test_message.tool_calls.empty()) {
             GGML_ASSERT(!data.params.grammar.empty());
         }
         if (!data.params.grammar.empty()) {
@@ -253,12 +237,35 @@ static void test_template(const common_chat_template & tmpl, const std::vector 0 && trigger.at_start) {
-                    fprintf(stderr, "Trigger %s not at start of message, skipping:\n\n%s\n\n", trigger.word.c_str(), constrained.c_str());
+                if (pos == std::string::npos) {
                     continue;
                 }
                 if (earliest_trigger_pos == std::string::npos || pos < earliest_trigger_pos) {
@@ -274,161 +281,364 @@ static void test_template(const common_chat_template & tmpl, const std::vectorI'm thinkingHello, world!\nWhat's up?",
+    /* .content_parts = */ {},
+    /* .tool_calls = */ {},
+    /* .reasoning_content = */ "",
+    /* .tool_name = */ "",
+    /* .tool_call_id = */ "",
+};
+const common_chat_msg message_assist_thoughts_unparsed_r7b {
+    "assistant",
+    "<|START_THINKING|>I'm thinking<|END_THINKING|>Hello, world!\nWhat's up?",
+    /* .content_parts = */ {},
+    /* .tool_calls = */ {},
+    /* .reasoning_content = */ "",
+    /* .tool_name = */ "",
+    /* .tool_call_id = */ "",
+};
+const common_chat_msg message_assist_thoughts {
+    "assistant",
+    "Hello, world!\nWhat's up?",
+    /* .content_parts = */ {},
+    /* .tool_calls = */ {},
+    /* .reasoning_content = */ "I'm thinking",
+    /* .tool_name = */ "",
+    /* .tool_call_id = */ "",
+};
+const std::vector tool_calls {
+    { "special_function", "{\"arg1\": 1}", /* .id = */ "" },
+};
+const std::vector tool_calls_idx {
+    { "special_function", "{\"arg1\": 1}", /* .id = */ "0" },
+};
+const std::vector tool_calls_id {
+    { "special_function", "{\"arg1\": 1}", /* .id = */ "123456789" },
+};
+
+const common_chat_msg message_assist_call {
+    "assistant",
+    "",
+    /* .content_parts = */ {},
+    tool_calls,
+    /* .reasoning_content = */ "",
+    /* .tool_name = */ "",
+    /* .tool_call_id = */ "",
+};
+const common_chat_msg message_assist_call_thoughts = {
+    "assistant",
+    /* .content = */ "",
+    /* .content_parts = */ {},
+    tool_calls,
+    /* .reasoning_content = */ "I'm\nthinking",
+    /* .tool_name = */ "",
+    /* .tool_call_id = */ "",
+};
+const common_chat_msg message_assist_call_thoughts_unparsed = {
+    "assistant",
+    /* .content = */ "I'm\nthinking",
+    /* .content_parts = */ {},
+    tool_calls,
+    /* .reasoning_content = */ "",
+    /* .tool_name = */ "",
+    /* .tool_call_id = */ "",
+};
+const common_chat_msg message_assist_call_id {
+    "assistant",
+    "",
+    /* .content_parts = */ {},
+    tool_calls_id,
+    /* .reasoning_content = */ "",
+    /* .tool_name = */ "",
+    /* .tool_call_id = */ "",
+};
+const common_chat_msg message_assist_call_idx {
+    "assistant",
+    "",
+    /* .content_parts = */ {},
+    tool_calls_idx,
+    /* .reasoning_content = */ "",
+    /* .tool_name = */ "",
+    /* .tool_call_id = */ "",
+};
+const common_chat_msg message_assist_call_python {
+    "assistant",
+    "",
+    /* .content_parts = */ {},
+    { { "python", "{\"code\": \"print('hey')\"}", /* .id = */ "" } },
+    /* .reasoning_content = */ "",
+    /* .tool_name = */ "",
+    /* .tool_call_id = */ "",
+};
+const common_chat_msg message_assist_call_code_interpreter {
+    "assistant",
+    "",
+    /* .content_parts = */ {},
+    { { "code_interpreter", "{\"code\": \"print('hey')\"}", /* .id = */ "" } },
+    /* .reasoning_content = */ "",
+    /* .tool_name = */ "",
+    /* .tool_call_id = */ "",
+};
+
+static void test_msgs_oaicompat_json_conversion() {
+    std::vector msgs{
+        message_user,
+        message_user_parts,
+        message_assist_call,
+        message_assist_call_thoughts,
+        message_assist_call_thoughts_unparsed,
+        message_assist_call_id,
+        message_assist_call_idx,
+        message_assist_call_python,
+        message_assist_call_code_interpreter,
+    };
+    for (const auto & msg : msgs) {
+        auto oai_json = common_chat_msgs_to_json_oaicompat({msg});
+        auto msgs2 = common_chat_msgs_parse_oaicompat(oai_json);
+        assert_equals((size_t) 1, msgs2.size());
+        auto msg2 = msgs2[0];
+        assert_msg_equals(msg, msg2);
+    }
+    assert_equals(
+        std::string(
+            "[\n"
+            "  {\n"
+            "    \"role\": \"user\",\n"
+            "    \"content\": [\n"
+            "      {\n"
+            "        \"type\": \"text\",\n"
+            "        \"text\": \"Hey\"\n"
+            "      },\n"
+            "      {\n"
+            "        \"type\": \"text\",\n"
+            "        \"text\": \"there\"\n"
+            "      }\n"
+            "    ]\n"
+            "  }\n"
+            "]"
+        ),
+        common_chat_msgs_to_json_oaicompat({message_user_parts}).dump(2));
+
+    assert_equals(
+        std::string(
+            "[\n"
+            "  {\n"
+            "    \"role\": \"assistant\",\n"
+            "    \"content\": null,\n"
+            "    \"tool_calls\": [\n"
+            "      {\n"
+            "        \"type\": \"function\",\n"
+            "        \"function\": {\n"
+            "          \"name\": \"python\",\n"
+            "          \"arguments\": \"{\\\"code\\\": \\\"print('hey')\\\"}\"\n"
+            "        }\n"
+            "      }\n"
+            "    ]\n"
+            "  }\n"
+            "]"
+        ),
+        common_chat_msgs_to_json_oaicompat({message_assist_call_python}).dump(2));
+
+    auto res = common_chat_msgs_parse_oaicompat(json::parse("[{\"role\": \"assistant\", \"tool_calls\": []}]"));
+    assert_equals(1, res.size());
+    assert_equals(res[0].role, "assistant");
+    assert_equals(true, res[0].content.empty());
+    assert_equals(true, res[0].tool_calls.empty());
+
+    try {
+        common_chat_msgs_parse_oaicompat(json::parse("[{\"role\": \"assistant\"}]"));
+        throw std::runtime_error("Expected exception");
+    } catch (const std::exception & e) {
+        if (std::string(e.what()).find("'content'") == std::string::npos) {
+            throw std::runtime_error("Expected exception about missing 'content'");
+        }
+    }
+}
+
+static void test_tools_oaicompat_json_conversion() {
+    std::vector tools{
+        special_function_tool,
+        python_tool,
+        code_interpreter_tool,
+    };
+
+    for (const auto & tool : tools) {
+        auto oai_json = common_chat_tools_to_json_oaicompat({tool});
+        auto tools2 = common_chat_tools_parse_oaicompat(oai_json);
+        assert_equals((size_t) 1, tools2.size());
+        auto tool2 = tools2[0];
+        assert_equals(tool.name, tool2.name);
+        assert_equals(tool.description, tool2.description);
+        assert_equals(json::parse(tool.parameters).dump(2), json::parse(tool2.parameters).dump(2));
+    }
+
+    assert_equals(
+        std::string(
+            "[\n"
+            "  {\n"
+            "    \"type\": \"function\",\n"
+            "    \"function\": {\n"
+            "      \"name\": \"special_function\",\n"
+            "      \"description\": \"I'm special\",\n"
+            "      \"parameters\": {\n"
+            "        \"type\": \"object\",\n"
+            "        \"properties\": {\n"
+            "          \"arg1\": {\n"
+            "            \"type\": \"integer\",\n"
+            "            \"description\": \"The arg.\"\n"
+            "          }\n"
+            "        },\n"
+            "        \"required\": [\n"
+            "          \"arg1\"\n"
+            "        ]\n"
+            "      }\n"
+            "    }\n"
+            "  }\n"
+            "]"
+        ),
+        common_chat_tools_to_json_oaicompat({special_function_tool}).dump(2));
+}
+
 static void test_template_output_parsers() {
-    json text_message {
-        { "role",    "assistant"     },
-        { "content", "Hello, world!\nWhat's up?" },
-    };
-    json tool_calls = json::array({{
-        { "type", "function" },
-        { "function", { { "name", "special_function" }, { "arguments", "{\"arg1\": 1}" } } },
-    }});
 
-    json tool_call_message {
-        { "role",       "assistant"},
-        { "content",    {}},
-        { "tool_calls", {
-            {
-                { "type", "function" },
-                { "function", {
-                    { "name", "special_function" },
-                    { "arguments", "{\"arg1\": 1}" },
-                }},
-            },
-        }},
-    };
-    json tool_call_message_with_id {
-        { "role",       "assistant"},
-        { "content",    {}},
-        { "tool_calls", {
-            {
-                { "type", "function" },
-                { "function", {
-                    { "name", "special_function" },
-                    { "arguments", "{\"arg1\": 1}" },
-                }},
-                {"id", "123456789"},
-            },
-        }},
-        { "role",       "assistant"                },
-        { "content",    {}                         },
-        { "tool_calls", tool_calls                  }
-    };
-    json tool_call_plan_message_with_idx {
-        { "role",       "assistant"},
-        { "content",    {}},
-        { "tool_plan",  "I'm not so sure"},
-        { "tool_calls", {
-            {
-                { "type", "function" },
-                { "function", {
-                    { "name", "special_function" },
-                    { "arguments", "{\"arg1\": 1}" },
-                }},
-                // Index of the tool call in the tool_calls array
-                {"id", "0"},
-            },
-        }},
-        { "role",       "assistant"                },
-        { "content",    {}                         },
-        { "tool_calls", tool_calls                  }
-    };
+    common_chat_templates_inputs inputs_no_tools;
+    inputs_no_tools.messages                = {message_user};
+    inputs_no_tools.extract_reasoning       = false;
 
-    auto python_tool_call_message = json{
-        { "role",       "assistant"                },
-        { "content",    {}                         },
-        { "tool_calls", json{ {
-                            { "type", "function" },
-                            { "function",
-                              {
-                                  { "name", "python" },
-                                  { "arguments",
-                                    {
-                                        { "code", "print('hey')" },
-                                    } },
-                              } },
-                        } } }
-    };
-    auto code_interpreter_tool_call_message = json{
-        { "role",       "assistant"                },
-        { "content",    {}                         },
-        { "tool_calls", json{ {
-                            { "type", "function" },
-                            { "function",
-                              {
-                                  { "name", "code_interpreter" },
-                                  { "arguments",
-                                    {
-                                        { "code", "print('hey')" },
-                                    } },
-                              } },
-                        } } }
-    };
+    common_chat_templates_inputs inputs_no_tools_think;
+    inputs_no_tools_think.messages          = {message_user};
+    inputs_no_tools_think.extract_reasoning = true;
 
-    common_chat_inputs inputs_no_tools;
-    inputs_no_tools.messages = {
-        { { "role", "user" }, { "content", "Hey\nThere" } }
-    };
+    common_chat_templates_inputs inputs_tools;
+    inputs_tools.messages                   = {message_user};
+    inputs_tools.tools                      = {special_function_tool};
+    inputs_tools.extract_reasoning          = false;
 
-    common_chat_inputs inputs_tools = inputs_no_tools;
-    inputs_tools.tools              = json::array();
-    inputs_tools.tools.push_back(special_function_tool);
+    common_chat_templates_inputs inputs_tools_think;
+    inputs_tools_think.messages             = {message_user};
+    inputs_tools_think.tools                = {special_function_tool};
+    inputs_tools_think.extract_reasoning    = true;
 
-    common_chat_inputs inputs_tools_builtin = inputs_no_tools;
-    inputs_tools_builtin.tools              = json::array();
-    inputs_tools_builtin.tools.push_back(python_tool);
+    common_chat_templates_inputs inputs_tools_builtin;
+    inputs_tools_builtin.messages           = {message_user};
+    inputs_tools_builtin.tools              = {python_tool};
+    inputs_tools_builtin.extract_reasoning  = false;
 
     {
         // Not supported yet
-        const common_chat_template tmpl(read_file("models/templates/CohereForAI-c4ai-command-r-plus-tool_use.jinja"), "", "");
-        assert_equals(COMMON_CHAT_FORMAT_GENERIC, common_chat_params_init(tmpl, inputs_tools).format);
+        auto tmpls = read_templates("models/templates/CohereForAI-c4ai-command-r-plus-tool_use.jinja");
+        assert_equals(COMMON_CHAT_FORMAT_CONTENT_ONLY, common_chat_templates_apply(tmpls.get(), inputs_no_tools).format);
+        assert_equals(COMMON_CHAT_FORMAT_GENERIC, common_chat_templates_apply(tmpls.get(), inputs_tools).format);
     }
     {
-        const common_chat_template tmpl(read_file("models/templates/CohereForAI-c4ai-command-r7b-12-2024-tool_use.jinja"), "", "");
+        auto tmpls = read_templates("models/templates/CohereForAI-c4ai-command-r7b-12-2024-tool_use.jinja");
         std::vector   end_tokens{ "<|END_OF_TURN_TOKEN|>" };
 
-        assert_equals(COMMON_CHAT_FORMAT_CONTENT_ONLY, common_chat_params_init(tmpl, inputs_no_tools).format);
-        assert_equals(COMMON_CHAT_FORMAT_COMMAND_R7B,    common_chat_params_init(tmpl, inputs_tools).format);
+        assert_equals(COMMON_CHAT_FORMAT_COMMAND_R7B,                   common_chat_templates_apply(tmpls.get(), inputs_no_tools).format);
+        assert_equals(COMMON_CHAT_FORMAT_COMMAND_R7B,                   common_chat_templates_apply(tmpls.get(), inputs_tools).format);
+        assert_equals(COMMON_CHAT_FORMAT_COMMAND_R7B_EXTRACT_REASONING, common_chat_templates_apply(tmpls.get(), inputs_tools_think).format);
 
-        test_template(tmpl, end_tokens, tool_call_plan_message_with_idx, tools,
-                      "<|START_THINKING|>I'm not so sure<|END_THINKING|>"
+        assert_msg_equals(message_assist,
+            common_chat_parse(
+                "Hello, world!\nWhat's up?",
+                COMMON_CHAT_FORMAT_COMMAND_R7B));
+        assert_msg_equals(message_assist,
+            common_chat_parse(
+                "Hello, world!\nWhat's up?<|END_RESPONSE|>",
+                COMMON_CHAT_FORMAT_COMMAND_R7B));
+        assert_msg_equals(message_assist,
+            common_chat_parse(
+                "<|START_RESPONSE|>Hello, world!\nWhat's up?<|END_RESPONSE|>",
+                COMMON_CHAT_FORMAT_COMMAND_R7B));
+        assert_msg_equals(message_assist_thoughts_unparsed_r7b,
+            common_chat_parse(
+                "<|START_THINKING|>I'm thinking<|END_THINKING|>"
+                "<|START_RESPONSE|>Hello, world!\nWhat's up?<|END_RESPONSE|>",
+                COMMON_CHAT_FORMAT_COMMAND_R7B));
+        assert_msg_equals(message_assist_thoughts_unparsed_r7b,
+            common_chat_parse(
+                "<|START_THINKING|>I'm thinking<|END_THINKING|>"
+                "Hello, world!\nWhat's up?<|END_RESPONSE|>",
+                COMMON_CHAT_FORMAT_COMMAND_R7B));
+
+        assert_msg_equals(message_assist_thoughts,
+            common_chat_parse(
+                "<|START_THINKING|>I'm thinking<|END_THINKING|>"
+                "<|START_RESPONSE|>Hello, world!\nWhat's up?<|END_RESPONSE|>",
+                COMMON_CHAT_FORMAT_COMMAND_R7B_EXTRACT_REASONING));
+
+        test_templates(tmpls.get(), end_tokens, message_assist_call_idx, tools,
+                      "<|START_THINKING|><|END_THINKING|>"
                       "<|START_ACTION|>[\n"
                       "    {\"tool_call_id\": \"0\", \"tool_name\": \"special_function\", \"parameters\": {\"arg1\": 1}}\n"
                       "]<|END_ACTION|>");
-        test_template(tmpl, end_tokens, text_message, tools,
+        test_templates(tmpls.get(), end_tokens, message_assist, tools,
                       "<|START_RESPONSE|>Hello, world!\n"
                       "What's up?<|END_RESPONSE|>",
                       /* expect_grammar_triggered= */ false);
     }
     {
-        const common_chat_template tmpl(read_file("models/templates/google-gemma-2-2b-it.jinja"), "", "");
+        auto tmpls = read_templates("models/templates/google-gemma-2-2b-it.jinja");
         std::vector   end_tokens{ "" };
 
-        assert_equals(COMMON_CHAT_FORMAT_CONTENT_ONLY, common_chat_params_init(tmpl, inputs_no_tools).format);
-        assert_equals(COMMON_CHAT_FORMAT_GENERIC, common_chat_params_init(tmpl, inputs_tools).format);
+        assert_equals(COMMON_CHAT_FORMAT_CONTENT_ONLY, common_chat_templates_apply(tmpls.get(), inputs_no_tools).format);
+        assert_equals(COMMON_CHAT_FORMAT_GENERIC, common_chat_templates_apply(tmpls.get(), inputs_tools).format);
         assert_equals(COMMON_CHAT_FORMAT_GENERIC,
-                      common_chat_params_init(
-                          common_chat_template(read_file("models/templates/microsoft-Phi-3.5-mini-instruct.jinja"),
-                                               "", ""),
+                      common_chat_templates_apply(
+                          read_templates("models/templates/microsoft-Phi-3.5-mini-instruct.jinja").get(),
                           inputs_tools)
                           .format);
 
         // Generic tool calls doesn't generate / parse content-only messages symmetrically.
 
-        assert_msg_equals(msg_from_json(text_message),
+        assert_msg_equals(message_assist,
                           common_chat_parse("{\n"
                                             "  \"response\": \"Hello, world!\\nWhat's up?\"\n"
                                             "}",
-                                            common_chat_params_init(tmpl, inputs_tools).format));
-        test_template(tmpl, end_tokens, tool_call_message_with_id, tools,
+                                            common_chat_templates_apply(tmpls.get(), inputs_tools).format));
+        test_templates(tmpls.get(), end_tokens, message_assist_call_id, tools,
                       "{\n"
                       "  \"tool_calls\": [\n"
                       "    {\n"
@@ -442,166 +652,331 @@ static void test_template_output_parsers() {
                       "}");
     }
     {
-        const common_chat_template tmpl(read_file("models/templates/mistralai-Mistral-Nemo-Instruct-2407.jinja"), "",
-                                        "");
+        auto tmpls = read_templates("models/templates/mistralai-Mistral-Nemo-Instruct-2407.jinja");
         std::vector   end_tokens{ "" };
 
-        assert_equals(COMMON_CHAT_FORMAT_MISTRAL_NEMO, common_chat_params_init(tmpl, inputs_tools).format);
+        assert_equals(COMMON_CHAT_FORMAT_MISTRAL_NEMO, common_chat_templates_apply(tmpls.get(), inputs_tools).format);
 
-        test_template(tmpl, end_tokens, text_message, tools, "Hello, world!\nWhat's up?", /* expect_grammar_triggered= */ false);
-        test_template(
-            tmpl, end_tokens, tool_call_message_with_id, tools,
+        test_templates(tmpls.get(), end_tokens, message_assist, tools, "Hello, world!\nWhat's up?", /* expect_grammar_triggered= */ false);
+        test_templates(
+            tmpls.get(), end_tokens, message_assist_call_id, tools,
             "[TOOL_CALLS][{\"name\": \"special_function\", \"arguments\": {\"arg1\": 1}, \"id\": \"123456789\"}]");
     }
     {
-        const common_chat_template tmpl(
-            read_file("models/templates/NousResearch-Hermes-2-Pro-Llama-3-8B-tool_use.jinja"), "", "");
+        auto tmpls = read_templates("models/templates/NousResearch-Hermes-2-Pro-Llama-3-8B-tool_use.jinja");
         std::vector end_tokens{ "<|im_end|>" };
 
-        assert_equals(COMMON_CHAT_FORMAT_HERMES_2_PRO, common_chat_params_init(tmpl, inputs_tools).format);
+        assert_equals(COMMON_CHAT_FORMAT_CONTENT_ONLY, common_chat_templates_apply(tmpls.get(), inputs_no_tools).format);
+        assert_equals(COMMON_CHAT_FORMAT_HERMES_2_PRO, common_chat_templates_apply(tmpls.get(), inputs_tools).format);
         assert_equals(
             COMMON_CHAT_FORMAT_HERMES_2_PRO,
-            common_chat_params_init(
-                common_chat_template(read_file("models/templates/NousResearch-Hermes-3-Llama-3.1-8B-tool_use.jinja"),
-                                     "", ""),
+            common_chat_templates_apply(
+                read_templates("models/templates/NousResearch-Hermes-3-Llama-3.1-8B-tool_use.jinja").get(),
                 inputs_tools)
                 .format);
         assert_equals(
             COMMON_CHAT_FORMAT_HERMES_2_PRO,
-            common_chat_params_init(
-                common_chat_template(read_file("models/templates/Qwen-Qwen2.5-7B-Instruct.jinja"), "", ""),
+            common_chat_templates_apply(
+                read_templates("models/templates/Qwen-Qwen2.5-7B-Instruct.jinja").get(),
                 inputs_tools)
                 .format);
 
-        test_template(tmpl, end_tokens, text_message, tools, "Hello, world!\nWhat's up?", /* expect_grammar_triggered= */ false);
-        test_template(tmpl, end_tokens, tool_call_message, tools,
+        // Test parsing
+        assert_msg_equals(message_assist_call, common_chat_parse(
+            "\n"
+            "{\"name\": \"special_function\", \"arguments\": {\"arg1\": 1}}\n"
+            "",
+            COMMON_CHAT_FORMAT_HERMES_2_PRO));
+        assert_msg_equals(message_assist_call, common_chat_parse(
+            "{\"arg1\": 1}",
+            COMMON_CHAT_FORMAT_HERMES_2_PRO));
+        assert_msg_equals(message_assist_call, common_chat_parse(
+            "\n"
+            "{\"arg1\": 1}\n"
+            "",
+            COMMON_CHAT_FORMAT_HERMES_2_PRO));
+        assert_msg_equals(message_assist_call, common_chat_parse(
+            "\n"
+            "  {\"name\": \"special_function\", \"arguments\": {\"arg1\": 1}}\n"
+            "",
+            COMMON_CHAT_FORMAT_HERMES_2_PRO));
+        assert_msg_equals(message_assist_call, common_chat_parse(
+            "\n"
+            "  {\"name\": \"special_function\", \"arguments\": {\"arg1\": 1}}\n"
+            "",
+            COMMON_CHAT_FORMAT_HERMES_2_PRO));
+        assert_msg_equals(message_assist_call, common_chat_parse(
+            "\n"
+            "  {\"name\": \"special_function\", \"arguments\": {\"arg1\": 1}}\n"
+            "",
+            COMMON_CHAT_FORMAT_HERMES_2_PRO));
+        assert_msg_equals(message_assist_call, common_chat_parse(
+            "```xml\n"
+            "\n"
+            "    {\"name\": \"special_function\", \"arguments\": {\"arg1\": 1}}\n"
+            "\n"
+            "```",
+            COMMON_CHAT_FORMAT_HERMES_2_PRO));
+        assert_msg_equals(message_assist_call, common_chat_parse(
+            "```xml\n"
+            "  {\"name\": \"special_function\", \"arguments\": {\"arg1\": 1}}\n"
+            "```",
+            COMMON_CHAT_FORMAT_HERMES_2_PRO));
+        assert_msg_equals(message_assist_call, common_chat_parse(
+            "```\n"
+            "  {\"name\": \"special_function\", \"arguments\": {\"arg1\": 1}}\n"
+            "```",
+            COMMON_CHAT_FORMAT_HERMES_2_PRO));
+        assert_msg_equals(message_assist_call, common_chat_parse(
+            "```\n"
+            "{\"name\": \"special_function\", \"arguments\": {\"arg1\": 1}}\n"
+            "```",
+            COMMON_CHAT_FORMAT_HERMES_2_PRO));
+        assert_msg_equals(message_assist_call, common_chat_parse(
+            "```json\n"
+            "  {\"name\": \"special_function\", \"arguments\": {\"arg1\": 1}}\n"
+            "```",
+            COMMON_CHAT_FORMAT_HERMES_2_PRO));
+        assert_msg_equals(message_assist_call, common_chat_parse(
+            "```json\n"
+            "\n"
+            "                     {\"name\": \"special_function\", \"arguments\": {\"arg1\": 1}} \n"
+            "                     \n"
+            "``` ",
+            COMMON_CHAT_FORMAT_HERMES_2_PRO));
+        assert_msg_equals(message_assist_call, common_chat_parse(
+            "\n"
+            "  {\"name\": \"special_function\", \"arguments\": {\"arg1\": 1}}\n"
+            "",
+            COMMON_CHAT_FORMAT_HERMES_2_PRO));
+        assert_msg_equals(message_assist_call, common_chat_parse(
+            "\n"
+            "  {\n"
+            "    \"name\": \"special_function\", \"arguments\": {\"arg1\": 1}\n"
+            "  }\n"
+            "",
+            COMMON_CHAT_FORMAT_HERMES_2_PRO));
+        assert_msg_equals(message_assist_call, common_chat_parse(
+            "\n"
+            "  {\"name\": \"special_function\", \"arguments\": {\"arg1\": 1}}\n"
+            "",
+            COMMON_CHAT_FORMAT_HERMES_2_PRO));
+        assert_msg_equals(message_assist_call, common_chat_parse(
+            "{\"name\": \"special_function\", \"arguments\": {\"arg1\": 1}}",
+            COMMON_CHAT_FORMAT_HERMES_2_PRO));
+        assert_msg_equals(message_assist_call, common_chat_parse(
+            "{\n  \"name\": \"special_function\", \"arguments\": {\"arg1\": 1}}",
+            COMMON_CHAT_FORMAT_HERMES_2_PRO));
+
+        assert_msg_equals(message_assist_thoughts_unparsed_think,
+            common_chat_parse("I'm thinkingHello, world!\nWhat's up?",
+            COMMON_CHAT_FORMAT_HERMES_2_PRO));
+        assert_msg_equals(message_assist_thoughts_unparsed_think,
+            common_chat_parse("I'm thinkingHello, world!\nWhat's up?",
+            COMMON_CHAT_FORMAT_HERMES_2_PRO));
+        assert_msg_equals(message_assist_thoughts,
+            common_chat_parse("I'm thinkingHello, world!\nWhat's up?",
+            COMMON_CHAT_FORMAT_HERMES_2_PRO_EXTRACT_REASONING));
+        assert_msg_equals(message_assist_thoughts,
+            common_chat_parse("I'm thinkingHello, world!\nWhat's up?",
+            COMMON_CHAT_FORMAT_HERMES_2_PRO_EXTRACT_REASONING));
+
+        test_templates(tmpls.get(), end_tokens, message_assist, tools, "Hello, world!\nWhat's up?", /* expect_grammar_triggered= */ false);
+        test_templates(tmpls.get(), end_tokens, message_assist_call, tools,
                       "\n"
                       "{\"name\": \"special_function\", \"arguments\": {\"arg1\": 1}}\n"
                       "");
-        test_template(tmpl, end_tokens, python_tool_call_message, tools,
+        test_templates(tmpls.get(), end_tokens, message_assist_call_python, tools,
                       "\n"
                       "{\"name\": \"python\", \"arguments\": {\"code\": \"print('hey')\"}}\n"
                       "");
     }
     {
-        const common_chat_template tmpl(read_file("models/templates/meta-llama-Llama-3.1-8B-Instruct.jinja"), "",
-                                        "");
+        auto tmpls = read_templates("models/templates/meta-llama-Llama-3.1-8B-Instruct.jinja");
         std::vector   end_tokens{ "<|eom_id|>", "<|eot_id|>" };
 
-        assert_equals(COMMON_CHAT_FORMAT_LLAMA_3_X, common_chat_params_init(tmpl, inputs_tools).format);
+        assert_equals(COMMON_CHAT_FORMAT_CONTENT_ONLY, common_chat_templates_apply(tmpls.get(), inputs_no_tools).format);
+        assert_equals(COMMON_CHAT_FORMAT_LLAMA_3_X, common_chat_templates_apply(tmpls.get(), inputs_tools).format);
         assert_equals(COMMON_CHAT_FORMAT_LLAMA_3_X_WITH_BUILTIN_TOOLS,
-                      common_chat_params_init(tmpl, inputs_tools_builtin).format);
+                      common_chat_templates_apply(tmpls.get(), inputs_tools_builtin).format);
         assert_equals(COMMON_CHAT_FORMAT_LLAMA_3_X_WITH_BUILTIN_TOOLS,
-                      common_chat_params_init(
-                          common_chat_template(read_file("models/templates/meta-llama-Llama-3.3-70B-Instruct.jinja"),
-                                               "", ""),
+                      common_chat_templates_apply(
+                          read_templates("models/templates/meta-llama-Llama-3.3-70B-Instruct.jinja").get(),
                           inputs_tools_builtin)
                           .format);
 
-        // test_template(tmpl, end_tokens, text_message, tools, R"(?)", /* expect_grammar_triggered= */ false);
-        test_template(tmpl, end_tokens, code_interpreter_tool_call_message, llama_3_1_tools,
+        // test_templates(tmpls.get(), end_tokens, message_assist, tools, R"(?)", /* expect_grammar_triggered= */ false);
+        test_templates(tmpls.get(), end_tokens, message_assist_call_code_interpreter, llama_3_1_tools,
                       "<|python_tag|>code_interpreter.call(code=\"print('hey')\")");
-        test_template(tmpl, end_tokens, python_tool_call_message, tools,
+        test_templates(tmpls.get(), end_tokens, message_assist_call_python, tools,
                       "<|python_tag|>python.call(code=\"print('hey')\")");
-        test_template(tmpl, end_tokens, tool_call_message, tools,
+        test_templates(tmpls.get(), end_tokens, message_assist_call, tools,
                       "{\"name\": \"special_function\", \"parameters\": {\"arg1\": 1}}");
     }
     {
-        const common_chat_template tmpl(read_file("models/templates/meta-llama-Llama-3.2-3B-Instruct.jinja"), "",
-                                        "");
+        auto tmpls = read_templates("models/templates/meta-llama-Llama-3.2-3B-Instruct.jinja");
         std::vector   end_tokens{ "<|eom_id|>", "<|eot_id|>" };
 
-        assert_equals(COMMON_CHAT_FORMAT_LLAMA_3_X, common_chat_params_init(tmpl, inputs_tools).format);
+        assert_equals(COMMON_CHAT_FORMAT_LLAMA_3_X, common_chat_templates_apply(tmpls.get(), inputs_tools).format);
+        assert_equals(COMMON_CHAT_FORMAT_CONTENT_ONLY, common_chat_templates_apply(tmpls.get(), inputs_no_tools).format);
 
-        test_template(tmpl, end_tokens, text_message, tools, "Hello, world!\nWhat's up?", /* expect_grammar_triggered= */ false);
-        test_template(tmpl, end_tokens, tool_call_message, tools,
+        test_templates(tmpls.get(), end_tokens, message_assist, tools, "Hello, world!\nWhat's up?", /* expect_grammar_triggered= */ false);
+        test_templates(tmpls.get(), end_tokens, message_assist_call, tools,
                       "{\"name\": \"special_function\", \"parameters\": {\"arg1\": 1}}");
     }
     {
-        const common_chat_template tmpl(read_file("models/templates/meetkai-functionary-medium-v3.1.jinja"), "",
-                                        "");
+        auto tmpls = read_templates("models/templates/meetkai-functionary-medium-v3.1.jinja");
         std::vector   end_tokens{ "<|eom_id|>", "<|eot_id|>" };
 
+        assert_equals(COMMON_CHAT_FORMAT_CONTENT_ONLY,
+                      common_chat_templates_apply(tmpls.get(), inputs_no_tools).format);
         assert_equals(COMMON_CHAT_FORMAT_FUNCTIONARY_V3_1_LLAMA_3_1,
-                      common_chat_params_init(tmpl, inputs_tools).format);
+                      common_chat_templates_apply(tmpls.get(), inputs_tools).format);
 
-        test_template(tmpl, end_tokens, text_message, tools, "Hello, world!\nWhat's up?", /* expect_grammar_triggered= */ false);
-        test_template(tmpl, end_tokens, tool_call_message, tools,
+        test_templates(tmpls.get(), end_tokens, message_assist, tools, "Hello, world!\nWhat's up?", /* expect_grammar_triggered= */ false);
+        test_templates(tmpls.get(), end_tokens, message_assist_call, tools,
                       "{\"arg1\": 1}");
     }
     {
-        const common_chat_template tmpl(read_file("models/templates/meetkai-functionary-medium-v3.2.jinja"), "",
-                                        "");
+        auto tmpls = read_templates("models/templates/meetkai-functionary-medium-v3.2.jinja");
         std::vector   end_tokens{ "<|eom_id|>", "<|eot_id|>" };
 
-        assert_equals(COMMON_CHAT_FORMAT_FUNCTIONARY_V3_2, common_chat_params_init(tmpl, inputs_no_tools).format);
-        assert_equals(COMMON_CHAT_FORMAT_FUNCTIONARY_V3_2, common_chat_params_init(tmpl, inputs_tools).format);
+        assert_equals(COMMON_CHAT_FORMAT_FUNCTIONARY_V3_2, common_chat_templates_apply(tmpls.get(), inputs_no_tools).format);
+        assert_equals(COMMON_CHAT_FORMAT_FUNCTIONARY_V3_2, common_chat_templates_apply(tmpls.get(), inputs_tools).format);
 
-        test_template(tmpl, end_tokens, text_message, {},
+        test_templates(tmpls.get(), end_tokens, message_assist, {},
                       "all\n"
                       "Hello, world!\n"
                       "What's up?",
                       /* expect_grammar_triggered= */ false);
-        test_template(tmpl, end_tokens, tool_call_message, tools,
+        test_templates(tmpls.get(), end_tokens, message_assist_call, tools,
                       "special_function\n"
                       "{\"arg1\": 1}");
     }
     {
-        const common_chat_template tmpl(read_file("models/templates/fireworks-ai-llama-3-firefunction-v2.jinja"), "",
-                                        "");
+        auto tmpls = read_templates("models/templates/fireworks-ai-llama-3-firefunction-v2.jinja");
         std::vector   end_tokens{ "<|eot_id|>" };
 
-        assert_equals(COMMON_CHAT_FORMAT_FIREFUNCTION_V2, common_chat_params_init(tmpl, inputs_tools).format);
+        assert_equals(COMMON_CHAT_FORMAT_CONTENT_ONLY, common_chat_templates_apply(tmpls.get(), inputs_no_tools).format);
+        assert_equals(COMMON_CHAT_FORMAT_FIREFUNCTION_V2, common_chat_templates_apply(tmpls.get(), inputs_tools).format);
 
-        test_template(tmpl, end_tokens, text_message, tools, "Hello, world!\nWhat's up?", /* expect_grammar_triggered= */ false);
-        test_template(tmpl, end_tokens, tool_call_message, tools,
+        test_templates(tmpls.get(), end_tokens, message_assist, tools, "Hello, world!\nWhat's up?", /* expect_grammar_triggered= */ false);
+        test_templates(tmpls.get(), end_tokens, message_assist_call, tools,
                       " functools[{\"name\": \"special_function\", \"arguments\": {\"arg1\": 1}}]");
     }
     {
-        const common_chat_template tmpl(read_file("models/templates/deepseek-ai-DeepSeek-R1-Distill-Llama-8B.jinja"),
-                                        "", "");
+        // Original DeepSeek R1 template. Leaves <｜tool▁calls▁begin｜> and others unclosed. Our logic fixes the prompt.
+        auto tmpls = read_templates("models/templates/deepseek-ai-DeepSeek-R1-Distill-Llama-8B.jinja");
         std::vector   end_tokens{ "<｜end▁of▁sentence｜>" };
 
-        assert_equals(COMMON_CHAT_FORMAT_DEEPSEEK_R1, common_chat_params_init(tmpl, inputs_tools).format);
+        assert_equals(COMMON_CHAT_FORMAT_DEEPSEEK_R1,                   common_chat_templates_apply(tmpls.get(), inputs_no_tools).format);
+        assert_equals(COMMON_CHAT_FORMAT_DEEPSEEK_R1,                   common_chat_templates_apply(tmpls.get(), inputs_tools).format);
+        assert_equals(COMMON_CHAT_FORMAT_DEEPSEEK_R1_EXTRACT_REASONING, common_chat_templates_apply(tmpls.get(), inputs_tools_think).format);
 
-        test_template(tmpl, end_tokens, text_message, tools, "Hello, world!\nWhat's up?", /* expect_grammar_triggered= */ false);
-        test_template(tmpl, end_tokens, tool_call_message, tools,
-                      "<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>function<｜tool▁sep｜>special_function\n"
-                      "```json\n"
-                      "{\"arg1\": 1}\n"
-                      "```<｜tool▁call▁end｜>");
+        test_templates(tmpls.get(), end_tokens, message_assist, tools, "Hello, world!\nWhat's up?", /* expect_grammar_triggered= */ false);
+        test_templates(tmpls.get(), end_tokens, message_assist_thoughts, tools, "Hello, world!\nWhat's up?", /* expect_grammar_triggered= */ false);
+        assert_msg_equals(message_assist_thoughts_unparsed_think,
+            common_chat_parse("I'm thinkingHello, world!\nWhat's up?",
+            COMMON_CHAT_FORMAT_DEEPSEEK_R1));
+        assert_msg_equals(message_assist_thoughts,
+            common_chat_parse("I'm thinkingHello, world!\nWhat's up?",
+            COMMON_CHAT_FORMAT_DEEPSEEK_R1_EXTRACT_REASONING));
+        assert_msg_equals(message_assist_thoughts,
+            // Latest template update (ast of 20250209) adds a trailing \n if add_generation_prompt is true.
+            common_chat_parse("I'm thinkingHello, world!\nWhat's up?",
+            COMMON_CHAT_FORMAT_DEEPSEEK_R1_EXTRACT_REASONING));
+        // test_templates(tmpls.get(), end_tokens, message_assist_call, tools,
+        //               "<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>function<｜tool▁sep｜>special_function\n"
+        //               "```json\n"
+        //               "{\"arg1\": 1}\n"
+        //               // Look what's not here: <｜tool▁calls▁end｜> (also missing the <｜end▁of▁sentence｜>, but that is removed lazily by the test's delta logic)
+        //               "```<｜tool▁call▁end｜>",
+        //               /* expect_grammar_triggered= */ true,
+        //               /* test_grammar_if_triggered= */ false);
+    }
+    {
+        // Replacement DeepSeek R1 template. Makes the Distill Qwen 7B/32B models happy to call tools and all.
+        auto tmpls = read_templates("models/templates/llama-cpp-deepseek-r1.jinja");
+        std::vector   end_tokens{ "<｜end▁of▁sentence｜>" };
+
+        assert_equals(COMMON_CHAT_FORMAT_DEEPSEEK_R1,                   common_chat_templates_apply(tmpls.get(), inputs_no_tools).format);
+        assert_equals(COMMON_CHAT_FORMAT_DEEPSEEK_R1,                   common_chat_templates_apply(tmpls.get(), inputs_tools).format);
+        assert_equals(COMMON_CHAT_FORMAT_DEEPSEEK_R1_EXTRACT_REASONING, common_chat_templates_apply(tmpls.get(), inputs_tools_think).format);
+
+        test_templates(tmpls.get(), end_tokens, message_assist, tools, "Hello, world!\nWhat's up?", /* expect_grammar_triggered= */ false);
+        test_templates(tmpls.get(), end_tokens, message_assist_thoughts, tools, "Hello, world!\nWhat's up?", /* expect_grammar_triggered= */ false);
+        assert_msg_equals(message_assist_thoughts_unparsed_think,
+            common_chat_parse("I'm thinkingHello, world!\nWhat's up?",
+            COMMON_CHAT_FORMAT_DEEPSEEK_R1));
+        assert_msg_equals(message_assist_thoughts,
+            common_chat_parse("I'm thinkingHello, world!\nWhat's up?",
+            COMMON_CHAT_FORMAT_DEEPSEEK_R1_EXTRACT_REASONING));
+
+        assert_msg_equals(message_assist_call_thoughts_unparsed,
+            common_chat_parse(
+                "I'm\nthinking\n\n"
+                "<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>function<｜tool▁sep｜>special_function\n"
+                "```json\n"
+                "{\"arg1\": 1}\n"
+                "```<｜tool▁call▁end｜><｜tool▁calls▁end｜>",
+                COMMON_CHAT_FORMAT_DEEPSEEK_R1));
+        assert_msg_equals(message_assist_call_thoughts,
+            common_chat_parse(
+                "I'm\nthinking\n\n"
+                "<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>function<｜tool▁sep｜>special_function\n"
+                "```json\n"
+                "{\"arg1\": 1}\n"
+                "```<｜tool▁call▁end｜><｜tool▁calls▁end｜>",
+                COMMON_CHAT_FORMAT_DEEPSEEK_R1_EXTRACT_REASONING));
+        test_templates(tmpls.get(), end_tokens, message_assist_call, tools,
+                "<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>function<｜tool▁sep｜>special_function\n"
+                "```json\n"
+                "{\"arg1\": 1}\n"
+                "```<｜tool▁call▁end｜><｜tool▁calls▁end｜>");
     }
 }
 
 int main(int argc, char ** argv) {
+    // try {
 #ifndef _WIN32
-    if (argc > 1) {
-        common_chat_inputs inputs;
-        inputs.messages = {
-            { { "role", "user" }, { "content", "Hey" } }
-        };
-        inputs.tools = json::array({ special_function_tool });
+        if (argc > 1) {
+            common_chat_templates_inputs inputs;
+            common_chat_msg msg;
+            msg.role = "user";
+            msg.content = "Hey";
+            inputs.messages = {msg};
+            inputs.tools = { special_function_tool };
 
-        std::cout << "| Template | Format |\n";
-        std::cout << "|----------|--------|\n";
+            std::cout << "| Template | Format |\n";
+            std::cout << "|----------|--------|\n";
 
-        for (int i = 1; i < argc; i++) {
-            std::string path = argv[i];
-            if (path.rfind(".jinja") != path.size() - 6) {
-                std::cerr << "Skipping non-jinja file: " << path << std::endl;
-                continue;
+            for (int i = 1; i < argc; i++) {
+                try {
+                    std::string path = argv[i];
+                    if (path.rfind(".jinja") != path.size() - 6) {
+                        std::cerr << "Skipping non-jinja file: " << path << '\n';
+                        continue;
+                    }
+                    auto tmpls = read_templates(path);
+                    auto parts  = string_split(path, "/");
+                    auto name   = parts[parts.size() - 1];
+                    auto format = common_chat_format_name(common_chat_templates_apply(tmpls.get(), inputs).format);
+                    std::cout << "| " << name << " | " << format << " |\n";
+                } catch (const std::exception & e) {
+                    std::cerr << "Failed to process " << argv[i] << ": " << e.what() << '\n';
+                }
             }
-            common_chat_template tmpl(read_file(path), "", "");
-            auto                 parts = string_split(path, "/");
-            auto                 name  = parts[parts.size() - 1];
-            std::cout << "| " << name << " | " << common_chat_format_name(common_chat_params_init(tmpl, inputs).format)
-                      << " |\n";
-        }
-    } else
+        } else
 #endif
-    {
-        test_template_output_parsers();
-        std::cout << "\n[chat] All tests passed!" << std::endl;
-    }
-    return 0;
+        {
+            test_msgs_oaicompat_json_conversion();
+            test_tools_oaicompat_json_conversion();
+            test_template_output_parsers();
+            std::cout << "\n[chat] All tests passed!" << '\n';
+        }
+        return 0;
+    // } catch (const std::exception & e) {
+    //     std::cerr << "Error: " << e.what() << '\n';
+    //     return 1;
+    // }
 }
diff --git a/tests/test-gguf.cpp b/tests/test-gguf.cpp
index 6ed696328..eaf572c66 100644
--- a/tests/test-gguf.cpp
+++ b/tests/test-gguf.cpp
@@ -697,8 +697,8 @@ static std::pair test_handcrafted_file(const unsigned int seed) {
 
 #ifdef _WIN32
         if (!file) {
-            printf("%s: failed to create tmpfile(), needs elevated privileges on Windows");
-            printf("%s: skipping tests");
+            printf("failed to create tmpfile(), needs elevated privileges on Windows");
+            printf("skipping tests");
             continue;
         }
 #else
@@ -1086,8 +1086,8 @@ static std::pair test_roundtrip(ggml_backend_dev_t dev, const unsigned
 
 #ifdef _WIN32
     if (!file) {
-        printf("%s: failed to create tmpfile(), needs elevated privileges on Windows");
-        printf("%s: skipping tests");
+        printf("failed to create tmpfile(), needs elevated privileges on Windows");
+        printf("skipping tests");
         return std::make_pair(0, 0);
     }
 #else
diff --git a/tests/test-grammar-llguidance.cpp b/tests/test-grammar-llguidance.cpp
index 8b696006b..3c19220e1 100644
--- a/tests/test-grammar-llguidance.cpp
+++ b/tests/test-grammar-llguidance.cpp
@@ -1086,6 +1086,65 @@ static void test_json_schema() {
         });
 }
 
+static void one_hot(llama_token_data_array & tok_arr, llama_token selected) {
+    auto n_vocab = tok_arr.size;
+
+    tok_arr.selected = -1;
+    tok_arr.sorted   = false;
+    for (llama_token token_id = 0; token_id < (llama_token) n_vocab; token_id++) {
+        tok_arr.data[token_id].id    = token_id;
+        tok_arr.data[token_id].logit = 0.0f;
+    }
+
+    tok_arr.data[selected].logit = 100.0f;
+}
+
+static void test_sampler_chain(void) {
+    auto sparams            = llama_sampler_chain_default_params();
+    sparams.no_perf         = false;
+    llama_sampler * sampler = llama_sampler_chain_init(sparams);
+
+    const auto grammar_data = R"(%llguidance {}
+start: /[A-Z ]*/)";
+
+    llama_sampler_chain_add(sampler, llama_sampler_init_llg(vocab, "lark", grammar_data));
+    llama_sampler_chain_add(sampler, llama_sampler_init_dist(42));
+
+    auto input  = "ALL YOUR BASE ARE BELONG TO US";
+    auto tokens = common_tokenize(vocab, input, false, false);
+
+    auto n_vocab = llama_vocab_n_tokens(vocab);
+
+    std::vector cur;
+    cur.reserve(n_vocab);
+    for (llama_token token_id = 0; token_id < (llama_token) n_vocab; token_id++) {
+        cur.emplace_back(llama_token_data{ token_id, 0.0f, 0.0f });
+    }
+    auto tok_arr = llama_token_data_array{ cur.data(), cur.size(), -1, false };
+
+    for (const auto token : tokens) {
+        one_hot(tok_arr, token);
+
+        fprintf(stderr, "applying token: %d\n", token);
+        llama_sampler_apply(sampler, &tok_arr);
+
+        auto idx = tok_arr.selected;
+        fprintf(stderr, " -> %d %f\n", cur[idx].id, cur[idx].logit);
+        assert(cur[tok_arr.selected].id == token);
+        llama_sampler_accept(sampler, token);
+    }
+
+    auto tok_eos = llama_vocab_eot(vocab);
+    if (tok_eos == LLAMA_TOKEN_NULL) {
+        tok_eos = llama_vocab_eos(vocab);
+    }
+
+    one_hot(tok_arr, tok_eos);
+
+    llama_sampler_apply(sampler, &tok_arr);
+    assert(cur[tok_arr.selected].id == tok_eos);
+}
+
 int main(int argc, const char ** argv) {
     fprintf(stdout, "Running llguidance integration tests...\n");
 
@@ -1135,6 +1194,9 @@ int main(int argc, const char ** argv) {
     test_special_chars();
     test_quantifiers();
     test_json_schema();
+
+    test_sampler_chain();
+
     fprintf(stdout, "All tests passed.\n");
     return 0;
 }
diff --git a/tests/test-json-schema-to-grammar.cpp b/tests/test-json-schema-to-grammar.cpp
index f38994c92..4d78e9142 100755
--- a/tests/test-json-schema-to-grammar.cpp
+++ b/tests/test-json-schema-to-grammar.cpp
@@ -91,7 +91,7 @@ static void test_all(const std::string & lang, std::function 0;
 }
diff --git a/tests/test-sampling.cpp b/tests/test-sampling.cpp
index 61bd67850..f1f87d454 100644
--- a/tests/test-sampling.cpp
+++ b/tests/test-sampling.cpp
@@ -181,6 +181,17 @@ static void test_dry(
     tester.check();
 }
 
+static void test_top_n_sigma(const std::vector & probs, const std::vector & probs_expected, int n) {
+    sampler_tester tester(probs, probs_expected);
+
+    DUMP(&tester.cur_p);
+    tester.apply(llama_sampler_init_top_n_sigma(n));
+    tester.apply(llama_sampler_init_dist (0));
+    DUMP(&tester.cur_p);
+
+    tester.check();
+}
+
 static void test_sampler_queue(const size_t n_vocab, const std::string & samplers_sequence, const int top_k, const float top_p, const float min_p
 ) {
     sampler_tester tester(n_vocab);
@@ -348,6 +359,10 @@ int main(void) {
     test_dry({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0, 1, 2, 0, 1}, {0.241818f, 0.241818f, 0.241818f, 0.241818f, 0.032727f}, 2.0f, 1.1f, 2, 5, {});
     test_dry({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0, 1, 2, 3, 4, 0, 1}, {0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, 1.0f, 1.1f, 4, 7, {});
 
+    test_top_n_sigma({0.1f, 0.2f, 0.3f, 0.4f}, {0.571429f, 0.428571f, 0.0f, 0.0f}, 1.00f);
+    test_top_n_sigma({0.1f, 0.2f, 0.3f, 0.4f}, {1.0f, 0.0f, 0.0f, 0.0f}, 0.00f);
+    test_top_n_sigma({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f, 0.3f, 0.2f, 0.1f}, 3.00f);
+
     test_sampler_queue(10000, "k", 10000, 1.0f, 1.0f);
     test_sampler_queue(10000, "k",     1, 1.0f, 1.0f);
     test_sampler_queue(10000, "p", 10000, 1.0f, 1.0f);