diff --git a/.devops/intel.Dockerfile b/.devops/intel.Dockerfile
index 8cad66052..9ce80a71e 100644
--- a/.devops/intel.Dockerfile
+++ b/.devops/intel.Dockerfile
@@ -49,19 +49,23 @@ COPY --from=build /app/full /app
 
 WORKDIR /app
 
-RUN apt-get update \
-    && apt-get install -y \
-    git \
-    python3 \
-    python3-pip \
-    && pip install --upgrade pip setuptools wheel \
-    && pip install -r requirements.txt \
-    && apt autoremove -y \
-    && apt clean -y \
-    && rm -rf /tmp/* /var/tmp/* \
-    && find /var/cache/apt/archives /var/lib/apt/lists -not -name lock -type f -delete \
-    && find /var/cache -type f -delete
+RUN apt-get update && \
+    apt-get install -y \
+        git \
+        python3 \
+        python3-pip \
+        python3-venv && \
+    python3 -m venv /opt/venv && \
+    . /opt/venv/bin/activate && \
+    pip install --upgrade pip setuptools wheel && \
+    pip install -r requirements.txt && \
+    apt autoremove -y && \
+    apt clean -y && \
+    rm -rf /tmp/* /var/tmp/* && \
+    find /var/cache/apt/archives /var/lib/apt/lists -not -name lock -type f -delete && \
+    find /var/cache -type f -delete
 
+ENV PATH="/opt/venv/bin:$PATH"
 
 ENTRYPOINT ["/app/tools.sh"]
 
diff --git a/.devops/tools.sh b/.devops/tools.sh
index 41a6b1e55..8a3a69340 100755
--- a/.devops/tools.sh
+++ b/.devops/tools.sh
@@ -1,4 +1,4 @@
-#!/bin/bash
+#!/usr/bin/env bash
 set -e
 
 # Read the first argument into a variable
diff --git a/.github/ISSUE_TEMPLATE/010-bug-compilation.yml b/.github/ISSUE_TEMPLATE/010-bug-compilation.yml
index b85bf5741..95a0b5cc7 100644
--- a/.github/ISSUE_TEMPLATE/010-bug-compilation.yml
+++ b/.github/ISSUE_TEMPLATE/010-bug-compilation.yml
@@ -40,7 +40,7 @@ body:
     attributes:
         label: GGML backends
         description: Which GGML backends do you know to be affected?
-        options: [AMX, BLAS, CPU, CUDA, HIP, Kompute, Metal, Musa, RPC, SYCL, Vulkan]
+        options: [AMX, BLAS, CPU, CUDA, HIP, Metal, Musa, RPC, SYCL, Vulkan, OpenCL]
         multiple: true
     validations:
       required: true
diff --git a/.github/ISSUE_TEMPLATE/011-bug-results.yml b/.github/ISSUE_TEMPLATE/011-bug-results.yml
index 1ccef0793..d1034bbb6 100644
--- a/.github/ISSUE_TEMPLATE/011-bug-results.yml
+++ b/.github/ISSUE_TEMPLATE/011-bug-results.yml
@@ -42,7 +42,7 @@ body:
     attributes:
         label: GGML backends
         description: Which GGML backends do you know to be affected?
-        options: [AMX, BLAS, CPU, CUDA, HIP, Kompute, Metal, Musa, RPC, SYCL, Vulkan]
+        options: [AMX, BLAS, CPU, CUDA, HIP, Metal, Musa, RPC, SYCL, Vulkan, OpenCL]
         multiple: true
     validations:
       required: true
diff --git a/.github/labeler.yml b/.github/labeler.yml
index 278032ef2..df6a7a40e 100644
--- a/.github/labeler.yml
+++ b/.github/labeler.yml
@@ -1,10 +1,4 @@
 # https://github.com/actions/labeler
-Kompute:
-    - changed-files:
-        - any-glob-to-any-file:
-            - ggml/include/ggml-kompute.h
-            - ggml/src/ggml-kompute/**
-            - README-kompute.md
 Apple Metal:
     - changed-files:
         - any-glob-to-any-file:
@@ -86,3 +80,15 @@ nix:
 embedding:
     - changed-files:
         - any-glob-to-any-file: examples/embedding/
+
+Ascend NPU:
+    - changed-files:
+        - any-glob-to-any-file:
+            - ggml/include/ggml-cann.h
+            - ggml/src/ggml-cann/**
+            - docs/backend/CANN.md
+OpenCL:
+    - changed-files:
+        - any-glob-to-any-file:
+            - ggml/include/ggml-opencl.h
+            - ggml/src/ggml-opencl/**
diff --git a/.github/workflows/build-cmake-pkg.yml b/.github/workflows/build-cmake-pkg.yml
new file mode 100644
index 000000000..fee2ab96b
--- /dev/null
+++ b/.github/workflows/build-cmake-pkg.yml
@@ -0,0 +1,51 @@
+name: Build relocatable cmake package
+on:
+  workflow_dispatch:
+  workflow_call:
+
+jobs:
+  linux:
+    runs-on: ubuntu-24.04
+    steps:
+      - uses: actions/checkout@v4
+        with:
+          fetch-depth: 0
+
+      - name: Install dependencies
+        run: |
+          sudo apt update
+          sudo apt install -y build-essential tcl
+
+      - name: Build
+        run: |
+          PREFIX="$(pwd)"/inst
+          cmake -S . -B build -DCMAKE_PREFIX_PATH="$PREFIX" \
+                -DLLAMA_CURL=OFF -DLLAMA_BUILD_TESTS=OFF -DLLAMA_BUILD_TOOLS=OFF \
+                -DLLAMA_BUILD_EXAMPLES=OFF -DCMAKE_BUILD_TYPE=Release
+          cmake --build build --config Release
+          cmake --install build --prefix "$PREFIX" --config Release
+
+          export LLAMA_CONFIG="$PREFIX"/lib/cmake/llama/llama-config.cmake
+          tclsh <<'EOF'
+          set build(commit)  [string trim [exec git rev-parse --short HEAD]]
+          set build(number)  [string trim [exec git rev-list  --count HEAD]]
+          set build(version) "0.0.$build(number)"
+
+          set llamaconfig [read [open "$env(LLAMA_CONFIG)" r]]
+          set checks [list "set\\(LLAMA_VERSION     \\s+$build(version)\\)" \
+                           "set\\(LLAMA_BUILD_COMMIT\\s+$build(commit)\\)" \
+                           "set\\(LLAMA_BUILD_NUMBER\\s+$build(number)\\)"]
+
+          puts -nonewline "Checking llama-config.cmake version... "
+          foreach check $checks {
+              if {![regexp -expanded -- $check $llamaconfig]} {
+                  puts "\"$check\" failed!"
+                  exit 1
+              }
+          }
+          puts "success."
+          EOF
+
+          cd examples/simple-cmake-pkg
+          cmake -S . -B build -DCMAKE_PREFIX_PATH="$PREFIX"/lib/cmake
+          cmake --build build
diff --git a/.github/workflows/build-linux-cross.yml b/.github/workflows/build-linux-cross.yml
index 92dc41f9d..7cfc82ba4 100644
--- a/.github/workflows/build-linux-cross.yml
+++ b/.github/workflows/build-linux-cross.yml
@@ -231,3 +231,116 @@ jobs:
                          -DCMAKE_FIND_ROOT_PATH_MODE_INCLUDE=BOTH
 
           cmake --build build --config Release -j $(nproc)
+
+  debian-13-loongarch64-cpu-cross:
+    runs-on: ubuntu-24.04
+    container: debian@sha256:653dfb9f86c3782e8369d5f7d29bb8faba1f4bff9025db46e807fa4c22903671
+
+    steps:
+      - uses: actions/checkout@v4
+      - name: Setup LoongArch
+        run: |
+          rm -f /etc/apt/sources.list.d/*
+          cat << EOF | tee /etc/apt/sources.list.d/debian-ports.list
+          deb http://snapshot.debian.org/archive/debian/20250515T202920Z/ trixie main
+          EOF
+          ( echo 'quiet "true";'; \
+            echo 'APT::Get::Assume-Yes "true";'; \
+            echo 'APT::Install-Recommends "false";'; \
+            echo 'Acquire::Check-Valid-Until "false";'; \
+            echo 'Acquire::Retries "5";'; \
+          ) > /etc/apt/apt.conf.d/99snapshot-repos
+
+          apt-get update
+          apt-get install -y ca-certificates debian-ports-archive-keyring cmake git zip
+          dpkg --add-architecture loong64
+
+          # Add arch-specific repositories for non-amd64 architectures
+          cat << EOF | tee /etc/apt/sources.list.d/loong64-ports.list
+          deb [arch=loong64] http://snapshot.debian.org/archive/debian-ports/20250515T194251Z/ sid main
+          EOF
+
+          apt-get update || true    ;# Prevent failure due to missing URLs.
+
+          apt-get install -y --no-install-recommends \
+                  build-essential \
+                  gcc-14-loongarch64-linux-gnu \
+                  g++-14-loongarch64-linux-gnu
+
+      - name: Build
+        run: |
+          cmake -B build -DLLAMA_CURL=OFF \
+                         -DCMAKE_BUILD_TYPE=Release \
+                         -DGGML_OPENMP=OFF \
+                         -DLLAMA_BUILD_EXAMPLES=ON \
+                         -DLLAMA_BUILD_TOOLS=ON \
+                         -DLLAMA_BUILD_TESTS=OFF \
+                         -DCMAKE_SYSTEM_NAME=Linux \
+                         -DCMAKE_SYSTEM_PROCESSOR=loongarch64 \
+                         -DCMAKE_C_COMPILER=loongarch64-linux-gnu-gcc-14 \
+                         -DCMAKE_CXX_COMPILER=loongarch64-linux-gnu-g++-14 \
+                         -DCMAKE_POSITION_INDEPENDENT_CODE=ON \
+                         -DCMAKE_FIND_ROOT_PATH=/usr/lib/loongarch64-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)
+
+  debian-13-loongarch64-vulkan-cross:
+    runs-on: ubuntu-24.04
+    container: debian@sha256:653dfb9f86c3782e8369d5f7d29bb8faba1f4bff9025db46e807fa4c22903671
+
+    steps:
+      - uses: actions/checkout@v4
+      - name: Setup LoongArch
+        run: |
+          rm -f /etc/apt/sources.list.d/*
+          cat << EOF | tee /etc/apt/sources.list.d/debian-ports.list
+          deb http://snapshot.debian.org/archive/debian/20250515T202920Z/ trixie main
+          EOF
+          ( echo 'quiet "true";'; \
+            echo 'APT::Get::Assume-Yes "true";'; \
+            echo 'APT::Install-Recommends "false";'; \
+            echo 'Acquire::Check-Valid-Until "false";'; \
+            echo 'Acquire::Retries "5";'; \
+          ) > /etc/apt/apt.conf.d/99snapshot-repos
+
+          apt-get update
+          apt-get install -y ca-certificates debian-ports-archive-keyring cmake git zip
+          dpkg --add-architecture loong64
+
+          # Add arch-specific repositories for non-amd64 architectures
+          cat << EOF | tee /etc/apt/sources.list.d/loong64-ports.list
+          deb [arch=loong64] http://snapshot.debian.org/archive/debian-ports/20250515T194251Z/ sid main
+          EOF
+
+          apt-get update || true    ;# Prevent failure due to missing URLs.
+
+          apt-get install -y --no-install-recommends \
+                  build-essential \
+                  glslc \
+                  gcc-14-loongarch64-linux-gnu \
+                  g++-14-loongarch64-linux-gnu \
+                  libvulkan-dev:loong64
+
+      - name: Build
+        run: |
+          cmake -B build -DLLAMA_CURL=OFF \
+                         -DCMAKE_BUILD_TYPE=Release \
+                         -DGGML_VULKAN=ON \
+                         -DGGML_OPENMP=OFF \
+                         -DLLAMA_BUILD_EXAMPLES=ON \
+                         -DLLAMA_BUILD_TOOLS=ON \
+                         -DLLAMA_BUILD_TESTS=OFF \
+                         -DCMAKE_SYSTEM_NAME=Linux \
+                         -DCMAKE_SYSTEM_PROCESSOR=loongarch64 \
+                         -DCMAKE_C_COMPILER=loongarch64-linux-gnu-gcc-14 \
+                         -DCMAKE_CXX_COMPILER=loongarch64-linux-gnu-g++-14 \
+                         -DCMAKE_POSITION_INDEPENDENT_CODE=ON \
+                         -DCMAKE_FIND_ROOT_PATH=/usr/lib/loongarch64-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 ee76d1799..42d63b7c5 100644
--- a/.github/workflows/build.yml
+++ b/.github/workflows/build.yml
@@ -5,10 +5,43 @@ on:
   push:
     branches:
       - master
-    paths: ['.github/workflows/build.yml', '.github/workflows/build-linux-cross.yml', '**/CMakeLists.txt', '**/.cmake', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.cuh', '**/*.swift', '**/*.m', '**/*.metal', '**/*.comp']
+    paths: [
+      '.github/workflows/build.yml',
+      '.github/workflows/build-linux-cross.yml',
+      '.github/workflows/build-cmake-pkg.yml',
+      '**/CMakeLists.txt',
+      '**/.cmake',
+      '**/*.h',
+      '**/*.hpp',
+      '**/*.c',
+      '**/*.cpp',
+      '**/*.cu',
+      '**/*.cuh',
+      '**/*.swift',
+      '**/*.m',
+      '**/*.metal',
+      '**/*.comp'
+    ]
+
   pull_request:
     types: [opened, synchronize, reopened]
-    paths: ['.github/workflows/build.yml', '.github/workflows/build-linux-cross.yml', '**/CMakeLists.txt', '**/.cmake', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.cuh', '**/*.swift', '**/*.m', '**/*.metal', '**/*.comp']
+    paths: [
+      '.github/workflows/build.yml',
+      '.github/workflows/build-linux-cross.yml',
+      '.github/workflows/build-cmake-pkg.yml',
+      '**/CMakeLists.txt',
+      '**/.cmake',
+      '**/*.h',
+      '**/*.hpp',
+      '**/*.c',
+      '**/*.cpp',
+      '**/*.cu',
+      '**/*.cuh',
+      '**/*.swift',
+      '**/*.m',
+      '**/*.metal',
+      '**/*.comp'
+    ]
 
 concurrency:
   group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
@@ -51,7 +84,8 @@ jobs:
             -DCMAKE_BUILD_RPATH="@loader_path" \
             -DLLAMA_FATAL_WARNINGS=ON \
             -DGGML_METAL_USE_BF16=ON \
-            -DGGML_METAL_EMBED_LIBRARY=ON \
+            -DGGML_METAL_EMBED_LIBRARY=OFF \
+            -DGGML_METAL_SHADER_DEBUG=ON \
             -DGGML_RPC=ON
           cmake --build build --config Release -j $(sysctl -n hw.logicalcpu)
 
@@ -306,6 +340,7 @@ jobs:
         id: cmake_test
         run: |
           cd build
+          export GGML_VK_VISIBLE_DEVICES=0
           # This is using llvmpipe and runs slower than other backends
           ctest -L main --verbose --timeout 3600
 
@@ -477,6 +512,9 @@ jobs:
   build-linux-cross:
     uses: ./.github/workflows/build-linux-cross.yml
 
+  build-cmake-pkg:
+    uses: ./.github/workflows/build-cmake-pkg.yml
+
   macOS-latest-cmake-ios:
     runs-on: macos-latest
 
@@ -627,7 +665,7 @@ jobs:
           ./build-xcframework.sh
 
   windows-msys2:
-    runs-on: windows-latest
+    runs-on: windows-2025
 
     strategy:
       fail-fast: false
@@ -677,28 +715,31 @@ jobs:
             cmake --build build --config ${{ matrix.build }} -j $(nproc)
 
   windows-latest-cmake:
-    runs-on: windows-latest
+    runs-on: windows-2025
 
     env:
       OPENBLAS_VERSION: 0.3.23
       SDE_VERSION: 9.33.0-2024-01-07
-      VULKAN_VERSION: 1.4.309.0
+      VULKAN_VERSION: 1.4.313.2
 
     strategy:
       matrix:
         include:
-          - build: 'cpu-x64'
-            defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/x64-windows-llvm.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DGGML_OPENMP=OFF'
+          - build: 'cpu-x64 (static)'
+            arch: 'x64'
+            defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/x64-windows-llvm.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DBUILD_SHARED_LIBS=OFF'
           - build: 'openblas-x64'
+            arch: 'x64'
             defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/x64-windows-llvm.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DGGML_OPENMP=OFF -DGGML_BLAS=ON -DGGML_BLAS_VENDOR=OpenBLAS -DBLAS_INCLUDE_DIRS="$env:RUNNER_TEMP/openblas/include" -DBLAS_LIBRARIES="$env:RUNNER_TEMP/openblas/lib/openblas.lib"'
           - build: 'vulkan-x64'
-            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DGGML_VULKAN=ON'
+            arch: 'x64'
+            defines: '-DCMAKE_BUILD_TYPE=Release -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DGGML_VULKAN=ON'
           - build: 'llvm-arm64'
+            arch: 'arm64'
             defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-llvm.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON'
           - build: 'llvm-arm64-opencl-adreno'
+            arch: 'arm64'
             defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-llvm.cmake -DCMAKE_PREFIX_PATH="$env:RUNNER_TEMP/opencl-arm64-release" -DGGML_OPENCL=ON -DGGML_OPENCL_USE_ADRENO_KERNELS=ON'
-         # - build: 'kompute-x64'
-         #   defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/x64-windows-llvm.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DGGML_OPENMP=OFF -DGGML_KOMPUTE=ON -DKOMPUTE_OPT_DISABLE_VULKAN_VERSION_CHECK=ON'
 
     steps:
       - name: Clone
@@ -712,12 +753,6 @@ jobs:
           variant: ccache
           evict-old-files: 1d
 
-      - name: Clone Kompute submodule
-        id: clone_kompute
-        if: ${{ matrix.build == 'kompute-x64' }}
-        run: |
-          git submodule update --init ggml/src/ggml-kompute/kompute
-
       - name: Download OpenBLAS
         id: get_openblas
         if: ${{ matrix.build == 'openblas-x64' }}
@@ -733,9 +768,9 @@ jobs:
 
       - name: Install Vulkan SDK
         id: get_vulkan
-        if: ${{ matrix.build == 'kompute-x64' || matrix.build == 'vulkan-x64' }}
+        if: ${{ matrix.build == 'vulkan-x64' }}
         run: |
-          curl.exe -o $env:RUNNER_TEMP/VulkanSDK-Installer.exe -L "https://sdk.lunarg.com/sdk/download/${env:VULKAN_VERSION}/windows/VulkanSDK-${env:VULKAN_VERSION}-Installer.exe"
+          curl.exe -o $env:RUNNER_TEMP/VulkanSDK-Installer.exe -L "https://sdk.lunarg.com/sdk/download/${env:VULKAN_VERSION}/windows/vulkansdk-windows-X64-${env:VULKAN_VERSION}.exe"
           & "$env:RUNNER_TEMP\VulkanSDK-Installer.exe" --accept-licenses --default-answer --confirm-command install
           Add-Content $env:GITHUB_ENV "VULKAN_SDK=C:\VulkanSDK\${env:VULKAN_VERSION}"
           Add-Content $env:GITHUB_PATH "C:\VulkanSDK\${env:VULKAN_VERSION}\bin"
@@ -768,6 +803,8 @@ jobs:
       - name: libCURL
         id: get_libcurl
         uses: ./.github/actions/windows-setup-curl
+        with:
+          architecture: ${{ matrix.arch == 'x64' && 'win64' || 'win64a' }}
 
       - name: Build
         id: cmake_build
@@ -777,6 +814,7 @@ jobs:
           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}
+          cp $env:CURL_PATH/bin/libcurl-*.dll build/bin/Release
 
       - name: Add libopenblas.dll
         id: add_libopenblas_dll
@@ -787,7 +825,7 @@ jobs:
 
       - name: Test
         id: cmake_test
-        if: ${{ matrix.build != 'llvm-arm64' && matrix.build != 'llvm-arm64-opencl-adreno' }}
+        if: ${{ matrix.arch == 'x64' }}
         run: |
           cd build
           ctest -L main -C Release --verbose --timeout 900
@@ -839,12 +877,12 @@ jobs:
               -DGGML_CUDA=ON
             cmake --build build
 
-  windows-2019-cmake-cuda:
-    runs-on: windows-2019
+  windows-2022-cmake-cuda:
+    runs-on: windows-2022
 
     strategy:
       matrix:
-        cuda: ['12.4', '11.7']
+        cuda: ['12.4']
 
     steps:
       - name: Clone
@@ -878,7 +916,7 @@ jobs:
         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"
+          call "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Auxiliary\Build\vcvarsall.bat" x64
           cmake -S . -B build -G "Ninja Multi-Config" ^
             -DLLAMA_BUILD_SERVER=ON ^
             -DGGML_NATIVE=OFF ^
@@ -892,7 +930,7 @@ jobs:
           cmake --build build --config Release
 
   windows-latest-cmake-sycl:
-    runs-on: windows-latest
+    runs-on: windows-2022
 
     defaults:
       run:
@@ -926,7 +964,7 @@ jobs:
 
   windows-latest-cmake-hip:
     if: ${{ github.event.inputs.create_release != 'true' }}
-    runs-on: windows-latest
+    runs-on: windows-2022
 
     steps:
       - name: Clone
diff --git a/.github/workflows/release.yml b/.github/workflows/release.yml
index 65ed24465..4ed6126f4 100644
--- a/.github/workflows/release.yml
+++ b/.github/workflows/release.yml
@@ -49,7 +49,8 @@ jobs:
         run: |
           sysctl -a
           cmake -B build \
-            -DCMAKE_BUILD_RPATH="@loader_path" \
+            -DCMAKE_INSTALL_RPATH='@loader_path' \
+            -DCMAKE_BUILD_WITH_INSTALL_RPATH=ON \
             -DLLAMA_FATAL_WARNINGS=ON \
             -DGGML_METAL_USE_BF16=ON \
             -DGGML_METAL_EMBED_LIBRARY=ON \
@@ -103,7 +104,8 @@ jobs:
           # Metal is disabled due to intermittent failures with Github runners not having a GPU:
           # https://github.com/ggml-org/llama.cpp/actions/runs/8635935781/job/23674807267#step:5:2313
           cmake -B build \
-            -DCMAKE_BUILD_RPATH="@loader_path" \
+            -DCMAKE_INSTALL_RPATH='@loader_path' \
+            -DCMAKE_BUILD_WITH_INSTALL_RPATH=ON \
             -DLLAMA_FATAL_WARNINGS=ON \
             -DGGML_METAL=OFF \
             -DGGML_RPC=ON
@@ -131,8 +133,9 @@ jobs:
         include:
           - build: 'x64'
             os: ubuntu-22.04
-          - build: 'arm64'
-            os: ubuntu-22.04-arm
+          # GGML_BACKEND_DL and GGML_CPU_ALL_VARIANTS are not currently supported on arm
+          # - build: 'arm64'
+          #   os: ubuntu-22.04-arm
 
     runs-on: ${{ matrix.os }}
 
@@ -159,6 +162,11 @@ jobs:
         id: cmake_build
         run: |
           cmake -B build \
+            -DCMAKE_INSTALL_RPATH='$ORIGIN' \
+            -DCMAKE_BUILD_WITH_INSTALL_RPATH=ON \
+            -DGGML_BACKEND_DL=ON \
+            -DGGML_NATIVE=OFF \
+            -DGGML_CPU_ALL_VARIANTS=ON \
             -DLLAMA_FATAL_WARNINGS=ON \
             ${{ env.CMAKE_ARGS }}
           cmake --build build --config Release -j $(nproc)
@@ -207,6 +215,11 @@ jobs:
         id: cmake_build
         run: |
           cmake -B build \
+            -DCMAKE_INSTALL_RPATH='$ORIGIN' \
+            -DCMAKE_BUILD_WITH_INSTALL_RPATH=ON \
+            -DGGML_BACKEND_DL=ON \
+            -DGGML_NATIVE=OFF \
+            -DGGML_CPU_ALL_VARIANTS=ON \
             -DGGML_VULKAN=ON \
             ${{ env.CMAKE_ARGS }}
           cmake --build build --config Release -j $(nproc)
@@ -228,7 +241,7 @@ jobs:
           name: llama-bin-ubuntu-vulkan-x64.zip
 
   windows-cpu:
-    runs-on: windows-latest
+    runs-on: windows-2025
 
     strategy:
       matrix:
@@ -264,7 +277,7 @@ jobs:
         env:
           CURL_PATH: ${{ steps.get_libcurl.outputs.curl_path }}
         run: |
-          call "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Auxiliary\Build\vcvarsall.bat" ${{ matrix.arch }}
+          call "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Auxiliary\Build\vcvarsall.bat" ${{ matrix.arch == 'x64' && 'x64' || 'amd64_arm64' }}
           cmake -S . -B build -G "Ninja Multi-Config" ^
             -D CMAKE_TOOLCHAIN_FILE=cmake/${{ matrix.arch }}-windows-llvm.cmake ^
             -DGGML_NATIVE=OFF ^
@@ -281,7 +294,7 @@ jobs:
           CURL_PATH: ${{ steps.get_libcurl.outputs.curl_path }}
         run: |
           Copy-Item $env:CURL_PATH\bin\libcurl-${{ matrix.arch }}.dll .\build\bin\Release\
-          Copy-Item "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Redist\MSVC\14.42.34433\debug_nonredist\${{ matrix.arch }}\Microsoft.VC143.OpenMP.LLVM\libomp140.${{ matrix.arch == 'x64' && 'x86_64' || 'aarch64' }}.dll" .\build\bin\Release\
+          Copy-Item "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Redist\MSVC\14.44.35112\debug_nonredist\${{ matrix.arch }}\Microsoft.VC143.OpenMP.LLVM\libomp140.${{ matrix.arch == 'x64' && 'x86_64' || 'aarch64' }}.dll" .\build\bin\Release\
           7z a llama-bin-win-cpu-${{ matrix.arch }}.zip .\build\bin\Release\*
 
       - name: Upload artifacts
@@ -291,11 +304,11 @@ jobs:
           name: llama-bin-win-cpu-${{ matrix.arch }}.zip
 
   windows:
-    runs-on: windows-latest
+    runs-on: windows-2025
 
     env:
       OPENBLAS_VERSION: 0.3.23
-      VULKAN_VERSION: 1.4.309.0
+      VULKAN_VERSION: 1.4.313.2
 
     strategy:
       matrix:
@@ -325,7 +338,7 @@ jobs:
         id: get_vulkan
         if: ${{ matrix.backend == 'vulkan' }}
         run: |
-          curl.exe -o $env:RUNNER_TEMP/VulkanSDK-Installer.exe -L "https://sdk.lunarg.com/sdk/download/${env:VULKAN_VERSION}/windows/VulkanSDK-${env:VULKAN_VERSION}-Installer.exe"
+          curl.exe -o $env:RUNNER_TEMP/VulkanSDK-Installer.exe -L "https://sdk.lunarg.com/sdk/download/${env:VULKAN_VERSION}/windows/vulkansdk-windows-X64-${env:VULKAN_VERSION}.exe"
           & "$env:RUNNER_TEMP\VulkanSDK-Installer.exe" --accept-licenses --default-answer --confirm-command install
           Add-Content $env:GITHUB_ENV "VULKAN_SDK=C:\VulkanSDK\${env:VULKAN_VERSION}"
           Add-Content $env:GITHUB_PATH "C:\VulkanSDK\${env:VULKAN_VERSION}\bin"
@@ -373,11 +386,11 @@ jobs:
           name: llama-bin-win-${{ matrix.backend }}-${{ matrix.arch }}.zip
 
   windows-cuda:
-    runs-on: windows-2019
+    runs-on: windows-2022
 
     strategy:
       matrix:
-        cuda: ['12.4', '11.7']
+        cuda: ['12.4']
 
     steps:
       - name: Clone
@@ -405,7 +418,7 @@ jobs:
         id: cmake_build
         shell: cmd
         run: |
-          call "C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise\VC\Auxiliary\Build\vcvars64.bat"
+          call "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Auxiliary\Build\vcvarsall.bat" x64
           cmake -S . -B build -G "Ninja Multi-Config" ^
             -DGGML_BACKEND_DL=ON ^
             -DGGML_NATIVE=OFF ^
@@ -441,7 +454,7 @@ jobs:
           name: cudart-llama-bin-win-cuda-${{ matrix.cuda }}-x64.zip
 
   windows-sycl:
-    runs-on: windows-latest
+    runs-on: windows-2022
 
     defaults:
       run:
@@ -513,7 +526,7 @@ jobs:
           name: llama-bin-win-sycl-x64.zip
 
   windows-hip:
-    runs-on: windows-latest
+    runs-on: windows-2022
 
     strategy:
       matrix:
diff --git a/.github/workflows/server.yml b/.github/workflows/server.yml
index 4baf6f6c7..f6da48857 100644
--- a/.github/workflows/server.yml
+++ b/.github/workflows/server.yml
@@ -180,7 +180,7 @@ jobs:
 
 
   server-windows:
-    runs-on: windows-2019
+    runs-on: windows-2022
 
     steps:
       - name: Clone
diff --git a/.gitmodules b/.gitmodules
index 23ce5ff05..e69de29bb 100644
--- a/.gitmodules
+++ b/.gitmodules
@@ -1,3 +0,0 @@
-[submodule "kompute"]
-	path = ggml/src/ggml-kompute/kompute
-	url = https://github.com/nomic-ai/kompute.git
diff --git a/CMakeLists.txt b/CMakeLists.txt
index ac3e90903..c79ccd09e 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -89,6 +89,14 @@ option(LLAMA_LLGUIDANCE "llama-common: include LLGuidance library for structured
 include(${CMAKE_CURRENT_SOURCE_DIR}/cmake/build-info.cmake)
 include(${CMAKE_CURRENT_SOURCE_DIR}/cmake/common.cmake)
 
+if (NOT DEFINED LLAMA_BUILD_NUMBER)
+    set(LLAMA_BUILD_NUMBER        ${BUILD_NUMBER})
+endif()
+if (NOT DEFINED LLAMA_BUILD_COMMIT)
+    set(LLAMA_BUILD_COMMIT        ${BUILD_COMMIT})
+endif()
+set(LLAMA_INSTALL_VERSION 0.0.${LLAMA_BUILD_NUMBER})
+
 # override ggml options
 set(GGML_ALL_WARNINGS   ${LLAMA_ALL_WARNINGS})
 set(GGML_FATAL_WARNINGS ${LLAMA_FATAL_WARNINGS})
@@ -112,7 +120,6 @@ endfunction()
 
 llama_option_depr(FATAL_ERROR LLAMA_CUBLAS              GGML_CUDA)
 llama_option_depr(WARNING     LLAMA_CUDA                GGML_CUDA)
-llama_option_depr(WARNING     LLAMA_KOMPUTE             GGML_KOMPUTE)
 llama_option_depr(WARNING     LLAMA_METAL               GGML_METAL)
 llama_option_depr(WARNING     LLAMA_METAL_EMBED_LIBRARY GGML_METAL_EMBED_LIBRARY)
 llama_option_depr(WARNING     LLAMA_NATIVE              GGML_NATIVE)
@@ -155,10 +162,17 @@ if (LLAMA_USE_SYSTEM_GGML)
 endif()
 
 if (NOT TARGET ggml AND NOT LLAMA_USE_SYSTEM_GGML)
+    set(GGML_BUILD_NUMBER ${LLAMA_BUILD_NUMBER})
+    set(GGML_BUILD_COMMIT ${LLAMA_BUILD_COMMIT})
     add_subdirectory(ggml)
     # ... otherwise assume ggml is added by a parent CMakeLists.txt
 endif()
 
+if (MINGW)
+    # Target Windows 8 for PrefetchVirtualMemory
+    add_compile_definitions(_WIN32_WINNT=${GGML_WIN_VER})
+endif()
+
 #
 # build the library
 #
@@ -199,10 +213,6 @@ endif()
 include(GNUInstallDirs)
 include(CMakePackageConfigHelpers)
 
-set(LLAMA_BUILD_NUMBER        ${BUILD_NUMBER})
-set(LLAMA_BUILD_COMMIT        ${BUILD_COMMIT})
-set(LLAMA_INSTALL_VERSION 0.0.${BUILD_NUMBER})
-
 set(LLAMA_INCLUDE_INSTALL_DIR ${CMAKE_INSTALL_INCLUDEDIR} CACHE PATH "Location of header  files")
 set(LLAMA_LIB_INSTALL_DIR     ${CMAKE_INSTALL_LIBDIR}     CACHE PATH "Location of library files")
 set(LLAMA_BIN_INSTALL_DIR     ${CMAKE_INSTALL_BINDIR}     CACHE PATH "Location of binary  files")
diff --git a/Makefile b/Makefile
index 958ad8f2f..ac442aec0 100644
--- a/Makefile
+++ b/Makefile
@@ -367,7 +367,7 @@ ifdef LLAMA_SERVER_SSL
 endif
 
 ifndef GGML_NO_CPU_AARCH64
-	MK_CPPFLAGS += -DGGML_USE_CPU_AARCH64
+	MK_CPPFLAGS += -DGGML_USE_CPU_REPACK
 endif
 
 # warnings
@@ -970,7 +970,7 @@ OBJ_GGML = \
 	$(DIR_GGML)/src/ggml-threading.o \
 	$(DIR_GGML)/src/ggml-cpu/ggml-cpu.o \
 	$(DIR_GGML)/src/ggml-cpu/ggml-cpu_cpp.o \
-	$(DIR_GGML)/src/ggml-cpu/ggml-cpu-aarch64.o \
+	$(DIR_GGML)/src/ggml-cpu/repack.o \
 	$(DIR_GGML)/src/ggml-cpu/ggml-cpu-hbm.o \
 	$(DIR_GGML)/src/ggml-cpu/ggml-cpu-quants.o \
 	$(DIR_GGML)/src/ggml-cpu/ggml-cpu-traits.o \
diff --git a/README.md b/README.md
index 576332bc5..90c7364df 100644
--- a/README.md
+++ b/README.md
@@ -3,9 +3,10 @@
 ![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)
+[![Release](https://img.shields.io/github/v/release/ggml-org/llama.cpp)](https://github.com/ggml-org/llama.cpp/releases)
 [![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/ggml-org/llama.cpp/discussions/3471) / [Manifesto](https://github.com/ggml-org/llama.cpp/discussions/205) / [ggml](https://github.com/ggml-org/ggml)
+[Roadmap](https://github.com/users/ggerganov/projects/7) / [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++
 
@@ -17,7 +18,6 @@ Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others)
 ## Hot topics
 
 - 🔥 Multimodal support arrived in `llama-server`: [#12898](https://github.com/ggml-org/llama.cpp/pull/12898) | [documentation](./docs/multimodal.md)
-- **GGML developer experience survey (organized and reviewed by NVIDIA):** [link](https://forms.gle/Gasw3cRgyhNEnrwK9)
 - A new binary `llama-mtmd-cli` is introduced to replace `llava-cli`, `minicpmv-cli`, `gemma3-cli` ([#13012](https://github.com/ggml-org/llama.cpp/pull/13012)) and `qwen2vl-cli` ([#13141](https://github.com/ggml-org/llama.cpp/pull/13141)), `libllava` will be deprecated
 - VS Code extension for FIM completions: https://github.com/ggml-org/llama.vscode
 - Universal [tool call support](./docs/function-calling.md) in `llama-server` https://github.com/ggml-org/llama.cpp/pull/9639
@@ -28,6 +28,30 @@ Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others)
 
 ----
 
+## Quick start
+
+Getting started with llama.cpp is straightforward. Here are several ways to install it on your machine:
+
+- Install `llama.cpp` using [brew, nix or winget](docs/install.md)
+- Run with Docker - see our [Docker documentation](docs/docker.md)
+- Download pre-built binaries from the [releases page](https://github.com/ggml-org/llama.cpp/releases)
+- Build from source by cloning this repository - check out [our build guide](docs/build.md)
+
+Once installed, you'll need a model to work with. Head to the [Obtaining and quantizing models](#obtaining-and-quantizing-models) section to learn more.
+
+Example command:
+
+```sh
+# Use a local model file
+llama-cli -m my_model.gguf
+
+# Or download and run a model directly from Hugging Face
+llama-cli -hf ggml-org/gemma-3-1b-it-GGUF
+
+# Launch OpenAI-compatible API server
+llama-server -hf ggml-org/gemma-3-1b-it-GGUF
+```
+
 ## Description
 
 The main goal of `llama.cpp` is to enable LLM inference with minimal setup and state-of-the-art performance on a wide
@@ -230,6 +254,7 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
 
 </details>
 
+
 ## Supported backends
 
 | Backend | Target devices |
@@ -246,16 +271,6 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
 | [OpenCL](docs/backend/OPENCL.md) | Adreno GPU |
 | [RPC](https://github.com/ggml-org/llama.cpp/tree/master/tools/rpc) | All |
 
-## Building the project
-
-The main product of this project is the `llama` library. Its C-style interface can be found in [include/llama.h](include/llama.h).
-The project also includes many example programs and tools using the `llama` library. The examples range from simple, minimal code snippets to sophisticated sub-projects such as an OpenAI-compatible HTTP server. Possible methods for obtaining the binaries:
-
-- 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/ggml-org/llama.cpp/releases)
-
 ## Obtaining and quantizing models
 
 The [Hugging Face](https://huggingface.co) platform hosts a [number of LLMs](https://huggingface.co/models?library=gguf&sort=trending) compatible with `llama.cpp`:
@@ -263,7 +278,11 @@ 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](https://huggingface.co/) or other model hosting sites, such as [ModelScope](https://modelscope.cn/), by using this CLI argument: `-hf <user>/<model>[: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 <user>/<model>[:quant]`. For example:
+
+```sh
+llama-cli -hf ggml-org/gemma-3-1b-it-GGUF
+```
 
 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/`.
 
diff --git a/build-xcframework.sh b/build-xcframework.sh
index a08419a80..f813984db 100755
--- a/build-xcframework.sh
+++ b/build-xcframework.sh
@@ -1,4 +1,4 @@
-#!/bin/bash
+#!/usr/bin/env bash
 #
 # Options
 IOS_MIN_OS_VERSION=16.4
diff --git a/ci/run.sh b/ci/run.sh
index b49a3a5f8..1146f86b6 100755
--- a/ci/run.sh
+++ b/ci/run.sh
@@ -1,4 +1,4 @@
-#!/bin/bash
+#!/usr/bin/env bash
 #
 # sample usage:
 #
@@ -39,14 +39,27 @@ sd=`dirname $0`
 cd $sd/../
 SRC=`pwd`
 
-CMAKE_EXTRA="-DLLAMA_FATAL_WARNINGS=ON -DLLAMA_CURL=OFF"
+CMAKE_EXTRA="-DLLAMA_FATAL_WARNINGS=ON -DLLAMA_CURL=ON"
 
 if [ ! -z ${GG_BUILD_METAL} ]; then
     CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_METAL=ON -DGGML_METAL_USE_BF16=ON"
 fi
 
 if [ ! -z ${GG_BUILD_CUDA} ]; then
-    CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_CUDA=ON -DCMAKE_CUDA_ARCHITECTURES=native"
+    CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_CUDA=ON"
+
+    if command -v nvidia-smi >/dev/null 2>&1; then
+        CUDA_ARCH=$(nvidia-smi --query-gpu=compute_cap --format=csv,noheader,nounits 2>/dev/null | head -1 | tr -d '.')
+        if [[ -n "$CUDA_ARCH" && "$CUDA_ARCH" =~ ^[0-9]+$ ]]; then
+            CMAKE_EXTRA="${CMAKE_EXTRA} -DCMAKE_CUDA_ARCHITECTURES=${CUDA_ARCH}"
+        else
+            echo "Warning: Using fallback CUDA architectures"
+            CMAKE_EXTRA="${CMAKE_EXTRA} -DCMAKE_CUDA_ARCHITECTURES=61;70;75;80;86;89"
+        fi
+    else
+        echo "Error: nvidia-smi not found, cannot build with CUDA"
+        exit 1
+    fi
 fi
 
 if [ ! -z ${GG_BUILD_SYCL} ]; then
@@ -766,7 +779,7 @@ function gg_run_rerank_tiny {
     model_f16="${path_models}/ggml-model-f16.gguf"
 
     # for this model, the SEP token is "</s>"
-    (time ./bin/llama-embedding --model ${model_f16} -p "what is panda?</s></s>hi\nwhat is panda?</s></s>it's a bear\nwhat is panda?</s></s>The giant panda (Ailuropoda melanoleuca), sometimes called a panda bear or simply panda, is a bear species endemic to China." -ngl 99 -c 0 --pooling rank --embd-normalize -1 --verbose-prompt) 2>&1 | tee -a $OUT/${ci}-rk-f16.log
+    (time ./bin/llama-embedding --model ${model_f16} -p "what is panda?\thi\nwhat is panda?\tit's a bear\nwhat is panda?\tThe giant panda (Ailuropoda melanoleuca), sometimes called a panda bear or simply panda, is a bear species endemic to China." -ngl 99 -c 0 --pooling rank --embd-normalize -1 --verbose-prompt) 2>&1 | tee -a $OUT/${ci}-rk-f16.log
 
     # sample output
     # rerank score 0:    0.029
diff --git a/common/CMakeLists.txt b/common/CMakeLists.txt
index 564af1448..f43a630c9 100644
--- a/common/CMakeLists.txt
+++ b/common/CMakeLists.txt
@@ -7,8 +7,8 @@ llama_add_compile_flags()
 # Build info header
 #
 
-if(EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/../.git")
-    set(GIT_DIR "${CMAKE_CURRENT_SOURCE_DIR}/../.git")
+if(EXISTS "${PROJECT_SOURCE_DIR}/.git")
+    set(GIT_DIR "${PROJECT_SOURCE_DIR}/.git")
 
     # Is git submodule
     if(NOT IS_DIRECTORY "${GIT_DIR}")
@@ -18,36 +18,26 @@ if(EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/../.git")
         if (SLASH_POS EQUAL 0)
             set(GIT_DIR "${REAL_GIT_DIR}")
         else()
-            set(GIT_DIR "${CMAKE_CURRENT_SOURCE_DIR}/../${REAL_GIT_DIR}")
+            set(GIT_DIR "${PROJECT_SOURCE_DIR}/${REAL_GIT_DIR}")
         endif()
     endif()
 
     if(EXISTS "${GIT_DIR}/index")
-        set(GIT_INDEX "${GIT_DIR}/index")
+        # For build-info.cpp below
+        set_property(DIRECTORY APPEND PROPERTY CMAKE_CONFIGURE_DEPENDS "${GIT_DIR}/index")
     else()
         message(WARNING "Git index not found in git repository.")
-        set(GIT_INDEX "")
     endif()
 else()
     message(WARNING "Git repository not found; to enable automatic generation of build info, make sure Git is installed and the project is a Git repository.")
-    set(GIT_INDEX "")
 endif()
 
-# Add a custom command to rebuild build-info.cpp when .git/index changes
-add_custom_command(
-    OUTPUT "${CMAKE_CURRENT_SOURCE_DIR}/build-info.cpp"
-    COMMENT "Generating build details from Git"
-    COMMAND ${CMAKE_COMMAND} -DMSVC=${MSVC} -DCMAKE_C_COMPILER_VERSION=${CMAKE_C_COMPILER_VERSION}
-            -DCMAKE_C_COMPILER_ID=${CMAKE_C_COMPILER_ID} -DCMAKE_VS_PLATFORM_NAME=${CMAKE_VS_PLATFORM_NAME}
-            -DCMAKE_C_COMPILER=${CMAKE_C_COMPILER}
-            -DCMAKE_SYSTEM_NAME=${CMAKE_SYSTEM_NAME} -DCMAKE_SYSTEM_PROCESSOR=${CMAKE_SYSTEM_PROCESSOR}
-            -P "${CMAKE_CURRENT_SOURCE_DIR}/cmake/build-info-gen-cpp.cmake"
-    WORKING_DIRECTORY "${CMAKE_CURRENT_SOURCE_DIR}/.."
-    DEPENDS "${CMAKE_CURRENT_SOURCE_DIR}/build-info.cpp.in" ${GIT_INDEX}
-    VERBATIM
-)
+set(TEMPLATE_FILE "${CMAKE_CURRENT_SOURCE_DIR}/build-info.cpp.in")
+set(OUTPUT_FILE   "${CMAKE_CURRENT_BINARY_DIR}/build-info.cpp")
+configure_file(${TEMPLATE_FILE} ${OUTPUT_FILE})
+
 set(TARGET build_info)
-add_library(${TARGET} OBJECT build-info.cpp)
+add_library(${TARGET} OBJECT ${OUTPUT_FILE})
 if (BUILD_SHARED_LIBS)
     set_target_properties(${TARGET} PROPERTIES POSITION_INDEPENDENT_CODE ON)
 endif()
diff --git a/common/arg.cpp b/common/arg.cpp
index cfa9878f9..40af7e574 100644
--- a/common/arg.cpp
+++ b/common/arg.cpp
@@ -988,10 +988,6 @@ static bool common_params_parse_ex(int argc, char ** argv, common_params_context
         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");
-    }
-
     if (!params.chat_template.empty() && !common_chat_verify_template(params.chat_template, params.use_jinja)) {
         throw std::runtime_error(string_format(
             "error: the supplied chat template is not supported: %s%s\n",
@@ -2710,6 +2706,13 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.embd_sep = value;
         }
     ).set_examples({LLAMA_EXAMPLE_EMBEDDING}));
+    add_opt(common_arg(
+        {"--cls-separator"}, "STRING",
+        "separator of classification sequences (default \\t) for example \"<#seq#>\"",
+        [](common_params & params, const std::string & value) {
+            params.cls_sep = value;
+        }
+    ).set_examples({LLAMA_EXAMPLE_EMBEDDING}));
     add_opt(common_arg(
         {"--host"}, "HOST",
         string_format("ip address to listen, or bind to an UNIX socket if the address ends with .sock (default: %s)", params.hostname.c_str()),
@@ -2747,9 +2750,10 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_EMBEDDINGS"));
     add_opt(common_arg(
         {"--reranking", "--rerank"},
-        string_format("enable reranking endpoint on server (default: %s)", params.reranking ? "enabled" : "disabled"),
+        string_format("enable reranking endpoint on server (default: %s)", "disabled"),
         [](common_params & params) {
-            params.reranking = true;
+            params.embedding = true;
+            params.pooling_type = LLAMA_POOLING_TYPE_RANK;
         }
     ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_RERANKING"));
     add_opt(common_arg(
@@ -2790,6 +2794,16 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.ssl_file_cert = value;
         }
     ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_SSL_CERT_FILE"));
+    add_opt(common_arg(
+        {"--chat-template-kwargs"}, "STRING",
+        string_format("sets additional params for the json template parser"),
+        [](common_params & params, const std::string &  value) {
+            auto parsed = json::parse(value);
+            for (const auto & item : parsed.items()) {
+                params.default_template_kwargs[item.key()] = item.value().dump();
+            }
+        }
+    ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_CHAT_TEMPLATE_KWARGS"));
     add_opt(common_arg(
         {"-to", "--timeout"}, "N",
         string_format("server read/write timeout in seconds (default: %d)", params.timeout_read),
@@ -2869,6 +2883,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         "(default: deepseek)",
         [](common_params & params, const std::string & value) {
             /**/ if (value == "deepseek") { params.reasoning_format = COMMON_REASONING_FORMAT_DEEPSEEK; }
+            else if (value == "deepseek-legacy") { params.reasoning_format = COMMON_REASONING_FORMAT_DEEPSEEK_LEGACY; }
             else if (value == "none") {     params.reasoning_format = COMMON_REASONING_FORMAT_NONE; }
             else { throw std::invalid_argument("invalid value"); }
         }
@@ -3212,6 +3227,32 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.speculative.model.path = value;
         }
     ).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_MODEL_DRAFT"));
+    add_opt(common_arg(
+        {"-ctkd", "--cache-type-k-draft"}, "TYPE",
+        string_format(
+            "KV cache data type for K for the draft model\n"
+            "allowed values: %s\n"
+            "(default: %s)",
+            get_all_kv_cache_types().c_str(),
+            ggml_type_name(params.speculative.cache_type_k)
+        ),
+        [](common_params & params, const std::string & value) {
+            params.speculative.cache_type_k = kv_cache_type_from_str(value);
+        }
+    ).set_env("LLAMA_ARG_CACHE_TYPE_K_DRAFT"));
+    add_opt(common_arg(
+        {"-ctvd", "--cache-type-v-draft"}, "TYPE",
+        string_format(
+            "KV cache data type for V for the draft model\n"
+            "allowed values: %s\n"
+            "(default: %s)",
+            get_all_kv_cache_types().c_str(),
+            ggml_type_name(params.speculative.cache_type_v)
+        ),
+        [](common_params & params, const std::string & value) {
+            params.speculative.cache_type_v = kv_cache_type_from_str(value);
+        }
+    ).set_env("LLAMA_ARG_CACHE_TYPE_V_DRAFT"));
 
     add_opt(common_arg(
         {"-mv", "--model-vocoder"}, "FNAME",
diff --git a/common/build-info.cpp.in b/common/build-info.cpp.in
index 0b945aa68..aee9d7eaf 100644
--- a/common/build-info.cpp.in
+++ b/common/build-info.cpp.in
@@ -1,4 +1,4 @@
-int LLAMA_BUILD_NUMBER = @BUILD_NUMBER@;
-char const *LLAMA_COMMIT = "@BUILD_COMMIT@";
+int LLAMA_BUILD_NUMBER = @LLAMA_BUILD_NUMBER@;
+char const *LLAMA_COMMIT = "@LLAMA_BUILD_COMMIT@";
 char const *LLAMA_COMPILER = "@BUILD_COMPILER@";
 char const *LLAMA_BUILD_TARGET = "@BUILD_TARGET@";
diff --git a/common/chat-parser.cpp b/common/chat-parser.cpp
index 65b664cb3..18a30e49a 100644
--- a/common/chat-parser.cpp
+++ b/common/chat-parser.cpp
@@ -49,6 +49,7 @@ bool common_chat_msg_parser::add_tool_call(const std::string & name, const std::
 
     // LOG_DBG("Tool call arguments:\n\traw: %s\n\tresult: %s\n", arguments.c_str(), tool_call.arguments.c_str());
     result_.tool_calls.emplace_back(tool_call);
+
     return true;
 }
 bool common_chat_msg_parser::add_tool_call(const json & tool_call) {
@@ -378,3 +379,7 @@ std::optional<common_chat_msg_parser::consume_json_result> common_chat_msg_parse
         /* .is_partial = */ found_healing_marker,
     };
 }
+
+void common_chat_msg_parser::clear_tools() {
+    result_.tool_calls.clear();
+}
diff --git a/common/chat-parser.h b/common/chat-parser.h
index 7ee355056..0e64c341a 100644
--- a/common/chat-parser.h
+++ b/common/chat-parser.h
@@ -115,4 +115,6 @@ class common_chat_msg_parser {
         const std::vector<std::vector<std::string>> & args_paths = {},
         const std::vector<std::vector<std::string>> & content_paths = {}
     );
+
+    void clear_tools();
 };
diff --git a/common/chat.cpp b/common/chat.cpp
index f1ab4c85a..114dbfccd 100644
--- a/common/chat.cpp
+++ b/common/chat.cpp
@@ -17,6 +17,8 @@
 #include <string>
 #include <vector>
 
+using json = nlohmann::ordered_json;
+
 static std::string format_time(const std::chrono::system_clock::time_point & now, const std::string & format) {
     auto time = std::chrono::system_clock::to_time_t(now);
     auto local_time = *std::localtime(&time);
@@ -82,10 +84,10 @@ json common_chat_msg::to_json_oaicompat() const
 
 std::vector<common_chat_msg_diff> common_chat_msg_diff::compute_diffs(const common_chat_msg & previous_msg, const common_chat_msg & new_msg) {
     std::vector<common_chat_msg_diff> diffs;
-    // if (previous_msg.reasoning_content != current.reasoning_content) {
-    //     auto & diff = diffs.emplace_back();
-    //     diff.reasoning_content_delta = string_diff(previous_msg.reasoning_content, current.reasoning_content);
-    // }
+    if (previous_msg.reasoning_content != new_msg.reasoning_content) {
+        auto & diff = diffs.emplace_back();
+        diff.reasoning_content_delta = string_diff(previous_msg.reasoning_content, new_msg.reasoning_content);
+    }
     if (previous_msg.content != new_msg.content) {
         auto & diff = diffs.emplace_back();
         diff.content_delta = string_diff(previous_msg.content, new_msg.content);
@@ -140,6 +142,7 @@ struct templates_params {
     bool add_generation_prompt = true;
     bool enable_thinking = true;
     std::chrono::system_clock::time_point now = std::chrono::system_clock::now();
+    json extra_context;
 };
 
 common_chat_tool_choice common_chat_tool_choice_parse_oaicompat(const std::string & tool_choice) {
@@ -385,9 +388,9 @@ json common_chat_tools_to_json_oaicompat(const std::vector<common_chat_tool> & t
 
 template <> json common_chat_msg_diff_to_json_oaicompat(const common_chat_msg_diff & diff) {
     json delta = json::object();
-    // if (!diff.reasoning_content_delta.empty()) {
-    //     delta["reasoning_content"] = msg.reasoning_content;
-    // }
+    if (!diff.reasoning_content_delta.empty()) {
+        delta["reasoning_content"] = diff.reasoning_content_delta;
+    }
     if (!diff.content_delta.empty()) {
         delta["content"] = diff.content_delta;
     }
@@ -598,6 +601,7 @@ const char * common_reasoning_format_name(common_reasoning_format format) {
     switch (format) {
         case COMMON_REASONING_FORMAT_NONE:     return "none";
         case COMMON_REASONING_FORMAT_DEEPSEEK: return "deepseek";
+        case COMMON_REASONING_FORMAT_DEEPSEEK_LEGACY: return "deepseek-legacy";
         default:
             throw std::runtime_error("Unknown reasoning format");
     }
@@ -719,16 +723,23 @@ static void foreach_function(const json & tools, const std::function<void(const
 
 static std::string apply(
     const common_chat_template & tmpl,
-    const nlohmann::ordered_json & messages,
-    const nlohmann::ordered_json & tools,
-    bool add_generation_prompt,
-    const nlohmann::ordered_json & extra_context = nlohmann::ordered_json())
+    const struct templates_params & inputs,
+    const std::optional<json> & messages_override = std::nullopt,
+    const std::optional<json> & tools_override = std::nullopt,
+    const std::optional<json> & additional_context = std::nullopt)
 {
     minja::chat_template_inputs tmpl_inputs;
-    tmpl_inputs.messages = messages;
-    tmpl_inputs.tools = tools;
-    tmpl_inputs.add_generation_prompt = add_generation_prompt;
-    tmpl_inputs.extra_context = extra_context;
+    tmpl_inputs.messages = messages_override ? *messages_override : inputs.messages;
+    if (tools_override) {
+        tmpl_inputs.tools = *tools_override;
+    } else {
+        tmpl_inputs.tools = inputs.tools.empty() ? json() : inputs.tools;
+    }
+    tmpl_inputs.add_generation_prompt = inputs.add_generation_prompt;
+    tmpl_inputs.extra_context = inputs.extra_context;
+    if (additional_context) {
+        tmpl_inputs.extra_context.merge_patch(*additional_context);
+    }
     // TODO: add flag to control date/time, if only for testing purposes.
     // tmpl_inputs.now = std::chrono::system_clock::now();
 
@@ -827,7 +838,7 @@ static common_chat_params common_chat_params_init_generic(const common_chat_temp
         inputs.messages,
         "Respond in JSON format, either with `tool_call` (a request to call tools) or with `response` reply to the user's request");
 
-    data.prompt = apply(tmpl, tweaked_messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt);
+    data.prompt = apply(tmpl, inputs, /* messages_override= */ tweaked_messages);
     data.format = COMMON_CHAT_FORMAT_GENERIC;
     return data;
 }
@@ -903,7 +914,7 @@ static common_chat_params common_chat_params_init_mistral_nemo(const common_chat
     data.preserved_tokens = {
         "[TOOL_CALLS]",
     };
-    data.prompt = apply(tmpl, inputs.messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt);
+    data.prompt = apply(tmpl, inputs);
     data.format = COMMON_CHAT_FORMAT_MISTRAL_NEMO;
     return data;
 }
@@ -933,7 +944,7 @@ static common_chat_params common_chat_params_init_command_r7b(const common_chat_
             adjusted_messages.push_back(msg);
         }
     }
-    data.prompt = apply(tmpl, adjusted_messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt, {});
+    data.prompt = apply(tmpl, inputs, /* messages_override= */ adjusted_messages);
     data.format = COMMON_CHAT_FORMAT_COMMAND_R7B;
     if (string_ends_with(data.prompt, "<|START_THINKING|>")) {
         if (!inputs.enable_thinking) {
@@ -1121,7 +1132,7 @@ static common_chat_params common_chat_params_init_llama_3_x(const common_chat_te
     } else {
         data.format = COMMON_CHAT_FORMAT_CONTENT_ONLY;
     }
-    data.prompt = apply(tmpl, inputs.messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt, {
+    data.prompt = apply(tmpl, inputs, /* messages_override =*/ std::nullopt, /* tools_override= */ std::nullopt, json {
         {"date_string", format_time(inputs.now, "%d %b %Y")},
         {"tools_in_user_message", false},
         {"builtin_tools", builtin_tools.empty() ? json() : builtin_tools},
@@ -1186,7 +1197,7 @@ static void common_chat_parse_llama_3_1(common_chat_msg_parser & builder, bool w
 
 static common_chat_params common_chat_params_init_deepseek_r1(const common_chat_template & tmpl, const struct templates_params & inputs) {
     common_chat_params data;
-    auto prompt = apply(tmpl, inputs.messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt);
+    auto prompt = apply(tmpl, inputs);
 
     // Hacks to fix the official (broken) prompt.
     // It is advisable to use --chat-template-file models/templates/llama-cpp-deepseek-r1.jinja instead,
@@ -1281,7 +1292,7 @@ static void common_chat_parse_deepseek_r1(common_chat_msg_parser & builder) {
 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, {
+    data.prompt = apply(tmpl, inputs, /* messages_override =*/ std::nullopt, /* tools_override= */ json(), json {
         {"datetime", format_time(inputs.now, "%b %d %Y %H:%M:%S GMT")},
         {"functions", json(inputs.tools.empty() ? "" : inputs.tools.dump(2))},
     });
@@ -1337,7 +1348,7 @@ static common_chat_params common_chat_params_init_functionary_v3_2(const common_
     // Using ">>>f1\n", ">>>f2\n"... as trigger words for the grammar
     // If the function is python, we also allow raw python code (if the line after `python\n` doesn't start w/ opening `{`), which the model seems to prefer for multiline code.
     common_chat_params data;
-    data.prompt = apply(tmpl, inputs.messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt);
+    data.prompt = apply(tmpl, inputs);
     data.format = COMMON_CHAT_FORMAT_FUNCTIONARY_V3_2;
     if (inputs.tools.is_array() && !inputs.tools.empty()) {
         data.grammar_lazy = inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_REQUIRED;
@@ -1464,7 +1475,7 @@ static common_chat_params common_chat_params_init_functionary_v3_1_llama_3_1(con
         data.format = COMMON_CHAT_FORMAT_CONTENT_ONLY;
     }
 
-    data.prompt = apply(tmpl, inputs.messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt);
+    data.prompt = apply(tmpl, inputs);
     // TODO: if (has_raw_python)
     return data;
 }
@@ -1497,14 +1508,15 @@ static void common_chat_parse_functionary_v3_1_llama_3_1(common_chat_msg_parser
 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;
 
-    json additional_context = {
+    json extra_context = json {
         {"enable_thinking", inputs.enable_thinking},
     };
+    extra_context.update(inputs.extra_context);
 
-    data.prompt = apply(tmpl, inputs.messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt, additional_context);
+    data.prompt = apply(tmpl, inputs, /* messages_override =*/ std::nullopt, /* tools_override= */ std::nullopt, extra_context);
     data.format = COMMON_CHAT_FORMAT_HERMES_2_PRO;
     if (string_ends_with(data.prompt, "<think>\n")) {
-        if (!inputs.enable_thinking) {
+        if (!extra_context["enable_thinking"]) {
             data.prompt += "</think>";
         } else {
             data.thinking_forced_open = true;
@@ -1690,7 +1702,7 @@ static void common_chat_parse_hermes_2_pro(common_chat_msg_parser & builder) {
 
 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.prompt = apply(tmpl, inputs);
     data.format = COMMON_CHAT_FORMAT_CONTENT_ONLY;
     data.grammar_lazy = false;
     if (!inputs.json_schema.is_null()) {
@@ -1721,6 +1733,12 @@ static common_chat_params common_chat_templates_apply_jinja(
     params.enable_thinking = inputs.enable_thinking;
     params.grammar = inputs.grammar;
     params.now = inputs.now;
+
+    params.extra_context = json::object();
+    for (auto el : inputs.chat_template_kwargs) {
+        params.extra_context[el.first] = json::parse(el.second);
+    }
+
     if (!inputs.json_schema.empty()) {
         params.json_schema = json::parse(inputs.json_schema);
     }
@@ -1837,7 +1855,7 @@ static common_chat_params common_chat_templates_apply_legacy(
     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");
+        throw std::runtime_error("this custom template is not supported, try using --jinja");
     }
 
     // if it turns out that our buffer is too small, we resize it
@@ -1920,7 +1938,9 @@ common_chat_msg common_chat_parse(const std::string & input, bool is_partial, co
     } catch (const common_chat_msg_partial_exception & ex) {
         LOG_DBG("Partial parse: %s\n", ex.what());
         if (!is_partial) {
-            throw std::runtime_error(ex.what());
+            builder.clear_tools();
+            builder.move_to(0);
+            common_chat_parse_content_only(builder);
         }
     }
     auto msg = builder.result();
diff --git a/common/chat.h b/common/chat.h
index f6b1d0ffc..ca807c145 100644
--- a/common/chat.h
+++ b/common/chat.h
@@ -7,6 +7,7 @@
 #include <chrono>
 #include <string>
 #include <vector>
+#include <map>
 
 struct common_chat_templates;
 
@@ -70,7 +71,7 @@ struct common_chat_msg {
 };
 
 struct common_chat_msg_diff {
-    // std::string reasoning_content_delta;
+    std::string reasoning_content_delta;
     std::string content_delta;
     size_t tool_call_index = std::string::npos;
     common_chat_tool_call tool_call_delta;
@@ -125,6 +126,7 @@ struct common_chat_templates_inputs {
     common_reasoning_format reasoning_format = COMMON_REASONING_FORMAT_NONE;
     bool enable_thinking = true;
     std::chrono::system_clock::time_point now = std::chrono::system_clock::now();
+    std::map<std::string, std::string> chat_template_kwargs;
 };
 
 struct common_chat_params {
diff --git a/common/cmake/build-info-gen-cpp.cmake b/common/cmake/build-info-gen-cpp.cmake
deleted file mode 100644
index fbc92b52c..000000000
--- a/common/cmake/build-info-gen-cpp.cmake
+++ /dev/null
@@ -1,24 +0,0 @@
-include(${CMAKE_CURRENT_SOURCE_DIR}/cmake/build-info.cmake)
-
-set(TEMPLATE_FILE "${CMAKE_CURRENT_SOURCE_DIR}/common/build-info.cpp.in")
-set(OUTPUT_FILE   "${CMAKE_CURRENT_SOURCE_DIR}/common/build-info.cpp")
-
-# Only write the build info if it changed
-if(EXISTS ${OUTPUT_FILE})
-    file(READ ${OUTPUT_FILE} CONTENTS)
-    string(REGEX MATCH "LLAMA_COMMIT = \"([^\"]*)\";" _ ${CONTENTS})
-    set(OLD_COMMIT ${CMAKE_MATCH_1})
-    string(REGEX MATCH "LLAMA_COMPILER = \"([^\"]*)\";" _ ${CONTENTS})
-    set(OLD_COMPILER ${CMAKE_MATCH_1})
-    string(REGEX MATCH "LLAMA_BUILD_TARGET = \"([^\"]*)\";" _ ${CONTENTS})
-    set(OLD_TARGET ${CMAKE_MATCH_1})
-    if (
-        NOT OLD_COMMIT   STREQUAL BUILD_COMMIT   OR
-        NOT OLD_COMPILER STREQUAL BUILD_COMPILER OR
-        NOT OLD_TARGET   STREQUAL BUILD_TARGET
-    )
-        configure_file(${TEMPLATE_FILE} ${OUTPUT_FILE})
-    endif()
-else()
-    configure_file(${TEMPLATE_FILE} ${OUTPUT_FILE})
-endif()
diff --git a/common/common.cpp b/common/common.cpp
index 4cc40ed8b..e4e71ad13 100644
--- a/common/common.cpp
+++ b/common/common.cpp
@@ -466,7 +466,7 @@ size_t string_find_partial_stop(const std::string_view & str, const std::string_
 
 std::string regex_escape(const std::string & s) {
     static const std::regex special_chars("[.^$|()*+?\\[\\]{}\\\\]");
-    return std::regex_replace(s, special_chars, "\\$0");
+    return std::regex_replace(s, special_chars, "\\$&");
 }
 
 std::string string_join(const std::vector<std::string> & values, const std::string & separator) {
@@ -706,11 +706,17 @@ bool fs_validate_filename(const std::string & filename) {
         // disable C++17 deprecation warning for std::codecvt_utf8
 #    pragma clang diagnostic push
 #    pragma clang diagnostic ignored "-Wdeprecated-declarations"
+#elif defined(__GNUC__)
+#    pragma GCC diagnostic push
+#    pragma GCC diagnostic ignored "-Wdeprecated-declarations"
 #endif
+
         std::wstring_convert<std::codecvt_utf8<char32_t>, char32_t> converter;
 
 #if defined(__clang__)
 #    pragma clang diagnostic pop
+#elif defined(__GNUC__)
+#    pragma GCC diagnostic pop
 #endif
 
         filename_utf32 = converter.from_bytes(filename);
@@ -767,6 +773,9 @@ bool fs_validate_filename(const std::string & filename) {
     return true;
 }
 
+#include <iostream>
+
+
 // returns true if successful, false otherwise
 bool fs_create_directory_with_parents(const std::string & path) {
 #ifdef _WIN32
@@ -784,9 +793,16 @@ bool fs_create_directory_with_parents(const std::string & path) {
     // 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);
+        pos_slash += 1;
+
+        // skip the drive letter, in some systems it can return an access denied error
+        if (subpath.length() == 2 && subpath[1] == ':') {
+            continue;
+        }
+
+        const bool success = CreateDirectoryW(subpath.c_str(), NULL);
+
         if (!success) {
             const DWORD error = GetLastError();
 
@@ -800,8 +816,6 @@ bool fs_create_directory_with_parents(const std::string & path) {
                 return false;
             }
         }
-
-        pos_slash += 1;
     }
 
     return true;
@@ -897,34 +911,6 @@ struct common_init_result common_init_from_params(common_params & params) {
 
     const llama_vocab * vocab = llama_model_get_vocab(model);
 
-    if (params.reranking) {
-        bool ok = true;
-
-        if (llama_vocab_bos(vocab) == LLAMA_TOKEN_NULL) {
-            LOG_WRN("%s: warning: vocab does not have a  BOS token, reranking will not work\n", __func__);
-            ok = false;
-        }
-
-        bool has_eos = llama_vocab_eos(vocab) != LLAMA_TOKEN_NULL;
-        bool has_sep = llama_vocab_sep(vocab) != LLAMA_TOKEN_NULL;
-
-        if (!has_eos && !has_sep) {
-            LOG_WRN("%s: warning: vocab does not have an EOS token or SEP token, reranking will not work\n", __func__);
-            ok = false;
-        } else if (!has_eos) {
-            LOG_WRN("%s: warning: vocab does not have an EOS token, using SEP token as fallback\n", __func__);
-        } else if (!has_sep) {
-            LOG_WRN("%s: warning: vocab does not have a SEP token, reranking will not work\n", __func__);
-            ok = false;
-        }
-
-        if (!ok) {
-            llama_model_free(model);
-
-            return iparams;
-        }
-    }
-
     auto cparams = common_context_params_to_llama(params);
 
     llama_context * lctx = llama_init_from_model(model, cparams);
@@ -934,7 +920,7 @@ struct common_init_result common_init_from_params(common_params & params) {
         return iparams;
     }
 
-    if (params.ctx_shift && !llama_kv_self_can_shift(lctx)) {
+    if (params.ctx_shift && !llama_memory_can_shift(llama_get_memory(lctx))) {
         LOG_WRN("%s: KV cache shifting is not supported for this context, disabling KV cache shifting\n", __func__);
         params.ctx_shift = false;
     }
@@ -966,6 +952,35 @@ struct common_init_result common_init_from_params(common_params & params) {
         }
     }
 
+    if (llama_pooling_type(lctx) == LLAMA_POOLING_TYPE_RANK) {
+        bool ok = true;
+
+        if (llama_vocab_bos(vocab) == LLAMA_TOKEN_NULL) {
+            LOG_WRN("%s: warning: vocab does not have a  BOS token, reranking will not work\n", __func__);
+            ok = false;
+        }
+
+        bool has_eos = llama_vocab_eos(vocab) != LLAMA_TOKEN_NULL;
+        bool has_sep = llama_vocab_sep(vocab) != LLAMA_TOKEN_NULL;
+
+        if (!has_eos && !has_sep) {
+            LOG_WRN("%s: warning: vocab does not have an EOS token or SEP token, reranking will not work\n", __func__);
+            ok = false;
+        } else if (!has_eos) {
+            LOG_WRN("%s: warning: vocab does not have an EOS token, using SEP token as fallback\n", __func__);
+        } else if (!has_sep) {
+            LOG_WRN("%s: warning: vocab does not have a SEP token, reranking will not work\n", __func__);
+            ok = false;
+        }
+
+        if (!ok) {
+            llama_free(lctx);
+            llama_model_free(model);
+
+            return iparams;
+        }
+    }
+
     // load and optionally apply lora adapters
     for (auto & la : params.lora_adapters) {
         llama_adapter_lora_ptr lora;
@@ -1041,7 +1056,7 @@ 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_self_clear(lctx);
+        llama_memory_clear(llama_get_memory(lctx), true);
         llama_synchronize(lctx);
         llama_perf_context_reset(lctx);
         llama_set_warmup(lctx, false);
@@ -1143,11 +1158,6 @@ struct llama_context_params common_context_params_to_llama(const common_params &
     cparams.op_offload        = !params.no_op_offload;
     cparams.swa_full          = params.swa_full;
 
-    if (params.reranking) {
-        cparams.embeddings    = true;
-        cparams.pooling_type  = LLAMA_POOLING_TYPE_RANK;
-    }
-
     cparams.type_k = params.cache_type_k;
     cparams.type_v = params.cache_type_v;
 
@@ -1280,6 +1290,9 @@ std::vector<llama_token> common_tokenize(
     int n_tokens = text.length() + 2 * add_special;
     std::vector<llama_token> result(n_tokens);
     n_tokens = llama_tokenize(vocab, text.data(), text.length(), result.data(), result.size(), add_special, parse_special);
+    if (n_tokens == std::numeric_limits<int32_t>::min()) {
+        throw std::runtime_error("Tokenization failed: input text too large, tokenization result exceeds int32_t limit");
+    }
     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);
diff --git a/common/common.h b/common/common.h
index cee1e3039..8922090e7 100644
--- a/common/common.h
+++ b/common/common.h
@@ -8,6 +8,7 @@
 #include <string>
 #include <string_view>
 #include <vector>
+#include <map>
 #include <sstream>
 
 #ifdef _WIN32
@@ -199,6 +200,9 @@ struct common_params_speculative {
     float   p_split      =  0.1f; // speculative decoding split probability
     float   p_min        = 0.75f; // minimum speculative decoding probability (greedy)
 
+    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
+
     struct cpu_params cpuparams;
     struct cpu_params cpuparams_batch;
 
@@ -215,7 +219,8 @@ struct common_params_vocoder {
 
 enum common_reasoning_format {
     COMMON_REASONING_FORMAT_NONE,
-    COMMON_REASONING_FORMAT_DEEPSEEK, // Extract thinking tag contents and return as `message.reasoning_content`
+    COMMON_REASONING_FORMAT_DEEPSEEK_LEGACY, // Extract thinking tag contents and return as `message.reasoning_content`, or leave inline in <think> tags in stream mode
+    COMMON_REASONING_FORMAT_DEEPSEEK,        // Extract thinking tag contents and return as `message.reasoning_content`, including in streaming deltas.
 };
 
 struct common_params {
@@ -354,7 +359,7 @@ struct common_params {
     int32_t embd_normalize = 2;     // normalisation for embeddings (-1=none, 0=max absolute int16, 1=taxicab, 2=euclidean, >2=p-norm)
     std::string embd_out   = "";    // empty = default, "array" = [[],[]...], "json" = openai style, "json+" = same "json" + cosine similarity matrix
     std::string embd_sep   = "\n";  // separator of embeddings
-    bool reranking         = false; // enable reranking support on server
+    std::string cls_sep    = "\t";  // separator of classification sequences
 
     // server params
     int32_t port           = 8080;         // server listens on this network port
@@ -377,6 +382,8 @@ struct common_params {
     std::string ssl_file_key  = "";                                                                         // NOLINT
     std::string ssl_file_cert = "";                                                                         // NOLINT
 
+    std::map<std::string, std::string> default_template_kwargs;
+
     // "advanced" endpoints are disabled by default for better security
     bool webui            = true;
     bool endpoint_slots   = false;
diff --git a/common/json-schema-to-grammar.cpp b/common/json-schema-to-grammar.cpp
index d38a74f95..637891f50 100644
--- a/common/json-schema-to-grammar.cpp
+++ b/common/json-schema-to-grammar.cpp
@@ -41,49 +41,6 @@ static std::string build_repetition(const std::string & item_rule, int min_items
     return result;
 }
 
-/* Minimalistic replacement for std::string_view, which is only available from C++17 onwards */
-class string_view {
-    const std::string & _str;
-    const size_t _start;
-    const size_t _end;
-public:
-    string_view(const std::string & str, size_t start = 0, size_t end  = std::string::npos) : _str(str), _start(start), _end(end == std::string::npos ? str.length() : end) {}
-
-    size_t size() const {
-        return _end - _start;
-    }
-
-    size_t length() const {
-        return size();
-    }
-
-    operator std::string() const {
-        return str();
-    }
-
-    std::string str() const {
-        return _str.substr(_start, _end - _start);
-    }
-
-    string_view substr(size_t pos, size_t len = std::string::npos) const {
-        return string_view(_str, _start + pos, len == std::string::npos ? _end : _start + pos + len);
-    }
-
-    char operator[](size_t pos) const {
-        auto index = _start + pos;
-        if (index >= _end) {
-            throw std::out_of_range("string_view index out of range");
-        }
-        return _str[_start + pos];
-    }
-
-    bool operator==(const string_view & other) const {
-        std::string this_str = *this;
-        std::string other_str = other;
-        return this_str == other_str;
-    }
-};
-
 static void _build_min_max_int(int min_value, int max_value, std::stringstream & out, int decimals_left = 16, bool top_level = true) {
     auto has_min = min_value != std::numeric_limits<int>::min();
     auto has_max = max_value != std::numeric_limits<int>::max();
@@ -112,14 +69,14 @@ static void _build_min_max_int(int min_value, int max_value, std::stringstream &
         }
         out << "}";
     };
-    std::function<void(const string_view &, const string_view &)> uniform_range =
-        [&](const string_view & from, const string_view & to) {
+    std::function<void(const std::string_view &, const std::string_view &)> uniform_range =
+        [&](const std::string_view & from, const std::string_view & to) {
             size_t i = 0;
             while (i < from.length() && i < to.length() && from[i] == to[i]) {
                 i++;
             }
             if (i > 0) {
-                out << "\"" << from.substr(0, i).str() << "\"";
+                out << "\"" << from.substr(0, i) << "\"";
             }
             if (i < from.length() && i < to.length()) {
                 if (i > 0) {
diff --git a/common/speculative.cpp b/common/speculative.cpp
index ccad70fa9..843bd1ddb 100644
--- a/common/speculative.cpp
+++ b/common/speculative.cpp
@@ -144,6 +144,8 @@ llama_tokens common_speculative_gen_draft(
     auto & smpl   = spec->smpl;
     auto & prompt = spec->prompt;
 
+    auto * mem = llama_get_memory(ctx);
+
     int reuse_i = 0;
     int reuse_n = 0;
 
@@ -173,7 +175,7 @@ llama_tokens common_speculative_gen_draft(
     result.reserve(params.n_draft);
 
     if (reuse_n == 0) {
-        llama_kv_self_clear(ctx);
+        llama_memory_clear(mem, false);
 
         prompt.clear();
     } else {
@@ -192,14 +194,14 @@ llama_tokens common_speculative_gen_draft(
         }
 
         if (reuse_i > 0) {
-            llama_kv_self_seq_rm (ctx, 0, 0, reuse_i);
-            llama_kv_self_seq_add(ctx, 0, reuse_i, -1, -reuse_i);
+            llama_memory_seq_rm (mem, 0, 0, reuse_i);
+            llama_memory_seq_add(mem, 0, reuse_i, -1, -reuse_i);
 
             prompt.erase(prompt.begin(), prompt.begin() + reuse_i);
         }
 
         if (reuse_n < (int) prompt.size()) {
-            llama_kv_self_seq_rm (ctx, 0, reuse_n, -1);
+            llama_memory_seq_rm (mem, 0, reuse_n, -1);
 
             prompt.erase(prompt.begin() + reuse_n, prompt.end());
         }
diff --git a/convert_hf_to_gguf.py b/convert_hf_to_gguf.py
index 4c566af5f..998d8723d 100755
--- a/convert_hf_to_gguf.py
+++ b/convert_hf_to_gguf.py
@@ -310,6 +310,8 @@ class ModelBase:
                             gguf.MODEL_TENSOR.POSNET_NORM2,
                             gguf.MODEL_TENSOR.V_ENC_EMBD_POS,
                             gguf.MODEL_TENSOR.A_ENC_EMBD_POS,
+                            gguf.MODEL_TENSOR.ALTUP_CORRECT_COEF,
+                            gguf.MODEL_TENSOR.ALTUP_PREDICT_COEF,
                         )
                     )
                     or not new_name.endswith(".weight")
@@ -320,7 +322,11 @@ class ModelBase:
                     self.match_model_tensor_name(new_name, key, bid)
                     for key in (
                         gguf.MODEL_TENSOR.TOKEN_EMBD,
+                        gguf.MODEL_TENSOR.PER_LAYER_TOKEN_EMBD,
                         gguf.MODEL_TENSOR.OUTPUT,
+                        gguf.MODEL_TENSOR.ALTUP_ROUTER,
+                        gguf.MODEL_TENSOR.LAUREL_L,
+                        gguf.MODEL_TENSOR.LAUREL_R,
                     )
                 ):
                     if self.ftype in (
@@ -519,7 +525,7 @@ class TextModel(ModelBase):
     def set_gguf_parameters(self):
         self.gguf_writer.add_block_count(self.block_count)
 
-        if (n_ctx := self.find_hparam(["max_position_embeddings", "n_ctx", "n_positions"], optional=True)) is not None:
+        if (n_ctx := self.find_hparam(["max_position_embeddings", "n_ctx", "n_positions", "max_length"], optional=True)) is not None:
             self.gguf_writer.add_context_length(n_ctx)
             logger.info(f"gguf: context length = {n_ctx}")
 
@@ -921,13 +927,20 @@ class TextModel(ModelBase):
         tokenizer = SentencePieceProcessor()
         tokenizer.LoadFromFile(str(tokenizer_path))
 
-        vocab_size = self.hparams.get('vocab_size', tokenizer.vocab_size())
+        vocab_size = self.find_hparam([
+            "vocab_size_per_layer_input", # gemma3n
+            "vocab_size",
+        ], optional=True) or tokenizer.vocab_size()
 
         tokens: list[bytes] = [f"[PAD{i}]".encode("utf-8") for i in range(vocab_size)]
         scores: list[float] = [-10000.0] * vocab_size
         toktypes: list[int] = [SentencePieceTokenTypes.UNUSED] * vocab_size
 
         for token_id in range(tokenizer.vocab_size()):
+            if token_id >= vocab_size:
+                logger.warning(f'ignore tokens from {token_id}: id is out of range, max={vocab_size - 1}')
+                break
+
             piece = tokenizer.IdToPiece(token_id)
             text = piece.encode("utf-8")
             score = tokenizer.GetScore(token_id)
@@ -1898,9 +1911,7 @@ class LlamaModel(TextModel):
         hparams = self.hparams
         self.gguf_writer.add_vocab_size(hparams["vocab_size"])
 
-        if "head_dim" in hparams:
-            rope_dim = hparams["head_dim"]
-        else:
+        if (rope_dim := hparams.get("head_dim")) is None:
             rope_dim = hparams["hidden_size"] // hparams["num_attention_heads"]
         self.gguf_writer.add_rope_dimension_count(rope_dim)
 
@@ -1982,7 +1993,8 @@ class LlamaModel(TextModel):
         if rope_scaling := self.find_hparam(["rope_scaling"], optional=True):
             if rope_scaling.get("rope_type", '').lower() == "llama3":
                 base = self.hparams.get("rope_theta", 10000.0)
-                dim = self.hparams.get("head_dim", self.hparams["hidden_size"] // self.hparams["num_attention_heads"])
+                if (dim := self.hparams.get("head_dim")) is None:
+                    dim = self.hparams["hidden_size"] // self.hparams["num_attention_heads"]
                 freqs = 1.0 / (base ** (torch.arange(0, dim, 2, dtype=torch.float32) / dim))
 
                 factor = rope_scaling.get("factor", 8.0)
@@ -2017,6 +2029,20 @@ class LlamaModel(TextModel):
                 raise ValueError(f"Unprocessed experts: {experts}")
 
 
+@ModelBase.register("ArceeForCausalLM")
+class ArceeModel(LlamaModel):
+    model_arch = gguf.MODEL_ARCH.ARCEE
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        self._try_set_pooling_type()
+        rope_scaling = self.hparams.get("rope_scaling") or {}
+        if rope_scaling.get("rope_type", rope_scaling.get("type")) == "yarn" and "factor" in rope_scaling:
+            self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.YARN)
+            self.gguf_writer.add_rope_scaling_factor(rope_scaling["factor"])
+            self.gguf_writer.add_rope_scaling_orig_ctx_len(rope_scaling["original_max_position_embeddings"])
+
+
 @ModelBase.register(
     "LlavaForConditionalGeneration", # pixtral
     "Mistral3ForConditionalGeneration", # mistral small 3.1
@@ -2132,7 +2158,6 @@ class Llama4Model(LlamaModel):
 
     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()
@@ -2181,7 +2206,7 @@ class Llama4VisionModel(MmprojModel):
                 name += ".weight"
             if "multi_modal_projector.linear_1" in name:
                 # despite the name with number postfix, this is a single fully connected layer
-                return [(gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.V_MMPROJ_FC], data_torch)]
+                return [(gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.V_MMPROJ_FC] + '.weight', data_torch)]
             return [(self.map_tensor_name(name), data_torch)]
         return []
 
@@ -2304,9 +2329,7 @@ class DeciModel(TextModel):
         hparams = self.hparams
         self.gguf_writer.add_vocab_size(hparams["vocab_size"])
 
-        if "head_dim" in hparams:
-            rope_dim = hparams["head_dim"]
-        else:
+        if (rope_dim := hparams.get("head_dim")) is None:
             rope_dim = hparams["hidden_size"] // hparams["num_attention_heads"]
         self.gguf_writer.add_rope_dimension_count(rope_dim)
 
@@ -2346,7 +2369,8 @@ class DeciModel(TextModel):
         if rope_scaling := self.find_hparam(["rope_scaling"], optional=True):
             if rope_scaling.get("rope_type", '').lower() == "llama3":
                 base = self.hparams.get("rope_theta", 10000.0)
-                dim = self.hparams.get("head_dim", self.hparams["hidden_size"] // self.hparams["num_attention_heads"])
+                if (dim := self.hparams.get("head_dim")) is None:
+                    dim = self.hparams["hidden_size"] // self.hparams["num_attention_heads"]
                 freqs = 1.0 / (base ** (torch.arange(0, dim, 2, dtype=torch.float32) / dim))
 
                 factor = rope_scaling.get("factor", 8.0)
@@ -2719,6 +2743,52 @@ class Qwen2Model(TextModel):
         yield from super().modify_tensors(data_torch, name, bid)
 
 
+@ModelBase.register("Ernie4_5_ForCausalLM")
+class Ernie4_5Model(TextModel):
+    model_arch = gguf.MODEL_ARCH.ERNIE4_5
+
+    def set_vocab(self):
+        self._set_vocab_sentencepiece()
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        num_heads = self.hparams["num_attention_heads"]
+        num_kv_heads = self.hparams["num_key_value_heads"]
+        head_dim = self.hparams["head_dim"]
+
+        if "ernie." in name:
+            name = name.replace("ernie.", "model.")
+        # split the qkv weights
+        # qkv_proj shape: [(num_heads + 2 * num_kv_heads) * head_dim, hidden_size]
+        if "qkv_proj" in name:
+            name_q = name.replace("qkv_proj.weight", "q_proj.weight")
+            name_k = name.replace("qkv_proj.weight", "k_proj.weight")
+            name_v = name.replace("qkv_proj.weight", "v_proj.weight")
+            total_q_dim = num_heads * head_dim
+            total_k_dim = num_kv_heads * head_dim
+            total_v_dim = num_kv_heads * head_dim
+            q_proj_weight, k_proj_weight, v_proj_weight = data_torch.split([total_q_dim, total_k_dim, total_v_dim], dim=0)
+            return [
+                (self.map_tensor_name(name_q), q_proj_weight),
+                (self.map_tensor_name(name_k), k_proj_weight),
+                (self.map_tensor_name(name_v), v_proj_weight)
+            ]
+        # split the up_gate_proj into gate and up
+        # up_gate_proj shape: [2 * intermediate_size, hidden_size]
+        if "up_gate_proj" in name:
+            name_up = name.replace("up_gate_proj.weight", "up_proj.weight")
+            name_gate = name.replace("up_gate_proj.weight", "gate_proj.weight")
+            dim_half = data_torch.shape[0] // 2
+            gate_proj_weight, up_proj_weight = data_torch.split(dim_half, dim=0)
+            return [
+                (self.map_tensor_name(name_gate), gate_proj_weight),
+                (self.map_tensor_name(name_up), up_proj_weight)
+            ]
+        return [(self.map_tensor_name(name), data_torch)]
+
+
 @ModelBase.register(
     "Qwen2VLModel",
     "Qwen2VLForConditionalGeneration",
@@ -3664,9 +3734,7 @@ class InternLM3Model(TextModel):
         hparams = self.hparams
         self.gguf_writer.add_vocab_size(hparams["vocab_size"])
 
-        if "head_dim" in hparams:
-            rope_dim = hparams["head_dim"]
-        else:
+        if (rope_dim := hparams.get("head_dim")) is None:
             rope_dim = hparams["hidden_size"] // hparams["num_attention_heads"]
         self.gguf_writer.add_rope_dimension_count(rope_dim)
 
@@ -3709,8 +3777,7 @@ class BertModel(TextModel):
         self._try_set_pooling_type()
 
         if self.cls_out_labels:
-            key_name = gguf.Keys.Classifier.OUTPUT_LABELS.format(arch = gguf.MODEL_ARCH_NAMES[self.model_arch])
-            self.gguf_writer.add_array(key_name, [v for k, v in sorted(self.cls_out_labels.items())])
+            self.gguf_writer.add_classifier_output_labels([v for k, v in sorted(self.cls_out_labels.items())])
 
     def set_vocab(self):
         tokens, toktypes, tokpre = self.get_vocab_base()
@@ -3909,9 +3976,6 @@ class BertModel(TextModel):
         special_vocab = gguf.SpecialVocab(self.dir_model, n_vocab=len(tokens))
         special_vocab.add_to_gguf(self.gguf_writer)
 
-        self.gguf_writer.add_add_bos_token(True)
-        self.gguf_writer.add_add_eos_token(True)
-
 
 @ModelBase.register("DistilBertModel", "DistilBertForMaskedLM", "DistilBertForSequenceClassification")
 class DistilBertModel(BertModel):
@@ -3953,8 +4017,6 @@ class RobertaModel(BertModel):
         bpe_tok_path = self.dir_model / "tokenizer.json"
         if bpe_tok_path.exists():
             self._set_vocab_gpt2()
-            self.gguf_writer.add_add_bos_token(True)
-            self.gguf_writer.add_add_eos_token(True)
 
             # we need this to validate the size of the token_type embeddings
             # though currently we are passing all zeros to the token_type embeddings
@@ -4060,6 +4122,34 @@ class NomicBertModel(BertModel):
         raise ValueError(f"unknown tokenizer: {toktyp}")
 
 
+@ModelBase.register("NeoBERT", "NeoBERTLMHead", "NeoBERTForSequenceClassification")
+class NeoBert(BertModel):
+    model_arch = gguf.MODEL_ARCH.NEO_BERT
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+
+        # NeoBERT uses 2/3 of the intermediate size as feed forward length
+        self.gguf_writer.add_feed_forward_length(int(2 * self.hparams["intermediate_size"] / 3))
+        self.gguf_writer.add_rope_freq_base(10000.0)  # default value for NeoBERT
+        self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.NONE)
+
+        f_rms_eps = self.hparams.get("norm_eps", 1e-6)  # default value for NeoBERT
+        self.gguf_writer.add_layer_norm_rms_eps(f_rms_eps)
+        logger.info(f"gguf: rms norm epsilon = {f_rms_eps}")
+
+        self.gguf_writer.add_pooling_type(gguf.PoolingType.CLS) # https://huggingface.co/chandar-lab/NeoBERT#how-to-use
+
+    def modify_tensors(self, data_torch, name, bid):
+        if name.startswith("decoder."):
+            return []
+
+        if name.startswith("model."):
+            name = name[6:]
+
+        return super().modify_tensors(data_torch, name, bid)
+
+
 @ModelBase.register("XLMRobertaModel", "XLMRobertaForSequenceClassification")
 class XLMRobertaModel(BertModel):
     model_arch = gguf.MODEL_ARCH.BERT
@@ -4214,6 +4304,7 @@ class Gemma2Model(TextModel):
 @ModelBase.register("Gemma3ForCausalLM", "Gemma3ForConditionalGeneration")
 class Gemma3Model(TextModel):
     model_arch = gguf.MODEL_ARCH.GEMMA3
+    norm_shift = 1.0  # Gemma3RMSNorm adds 1.0 to the norm value
 
     def set_vocab(self):
         self._set_vocab_sentencepiece()
@@ -4235,9 +4326,8 @@ class Gemma3Model(TextModel):
         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
+        # attn_logit_softcapping is 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:
@@ -4249,7 +4339,7 @@ class Gemma3Model(TextModel):
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
         del bid  # unused
 
-        if name.startswith("language_model."):
+        if "language_model." in name:
             name = name.replace("language_model.", "")
 
         elif name.startswith("multi_modal_projector.") or name.startswith("vision_tower.") \
@@ -4264,8 +4354,9 @@ class Gemma3Model(TextModel):
 
         # ref code in Gemma3RMSNorm
         # output = output * (1.0 + self.weight.float())
+        # note: this is not the case on gemma3n
         if name.endswith("norm.weight"):
-            data_torch = data_torch + 1
+            data_torch = data_torch + self.norm_shift
 
         return [(self.map_tensor_name(name), data_torch)]
 
@@ -4322,6 +4413,101 @@ class Gemma3VisionModel(MmprojModel):
         return [] # skip other tensors
 
 
+@ModelBase.register("Gemma3nForConditionalGeneration")
+class Gemma3NModel(Gemma3Model):
+    model_arch = gguf.MODEL_ARCH.GEMMA3N
+    norm_shift = 0.0 # same value with Gemma3p5RMSNorm scale_shift on python code
+
+    _altup_proj: list[Tensor] = []
+    _altup_unembd: list[Tensor] = []
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        assert self.hparams["altup_num_inputs"] == 4, "Current conversion only supports 4 altup inputs"
+        self._altup_proj = [
+            torch.Tensor(), # to be replaced
+            torch.Tensor(), # to be replaced
+            torch.Tensor(), # to be replaced
+        ]
+        self._altup_unembd = [
+            torch.Tensor(), # to be replaced
+            torch.Tensor(), # to be replaced
+            torch.Tensor(), # to be replaced
+        ]
+
+    def set_vocab(self):
+        super().set_vocab()
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        self.gguf_writer.add_altup_active_idx(self.hparams["altup_active_idx"])
+        self.gguf_writer.add_altup_num_inputs(self.hparams["altup_num_inputs"])
+        self.gguf_writer.add_embedding_length_per_layer_input(self.hparams["hidden_size_per_layer_input"])
+        self.gguf_writer.add_shared_kv_layers(self.hparams["num_kv_shared_layers"])
+
+        activation_sparsity_scale = []
+        for s in self.hparams["activation_sparsity_pattern"]:
+            normal_dist = torch.distributions.normal.Normal(0, 1)
+            std_multiplier = normal_dist.icdf(torch.tensor(s, dtype=torch.float32))
+            activation_sparsity_scale.append(std_multiplier.item())
+        self.gguf_writer.add_activation_sparsity_scale(activation_sparsity_scale)
+
+        sliding_window_pattern = []
+        for t in self.hparams["layer_types"]:
+            sliding_window_pattern.append(t == "sliding_attention")
+        self.gguf_writer.add_sliding_window_pattern(sliding_window_pattern)
+
+    def _stack_matrices(self, matrices: list[Tensor]) -> Tensor | None:
+        has_all = all(m.numel() > 0 for m in matrices)
+        if not has_all:
+            return None
+        else:
+            return torch.stack(matrices, dim=0)
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        if name.endswith("_scale"):
+            name = name + ".weight"
+
+        # TODO: implement self.prediction_coefs.weight.clamp_(...)
+
+        if "language_model." not in name:
+            return [] # skip non-language model tensors
+
+        if "altup_unembed_projections" in name:
+            data_torch = data_torch.to(device="cpu")
+            if ".0." in name:
+                self._altup_unembd[0] = data_torch
+            elif ".1." in name:
+                self._altup_unembd[1] = data_torch
+            elif ".2." in name:
+                self._altup_unembd[2] = data_torch
+            else:
+                raise ValueError(f"Unknown name: {name}")
+            out = self._stack_matrices(self._altup_unembd)
+            if out is not None:
+                return [(self.map_tensor_name("model.altup_unembed_projections.weight"), out)]
+            else:
+                return []
+
+        if "altup_projections" in name:
+            data_torch = data_torch.to(device="cpu")
+            if ".0." in name:
+                self._altup_proj[0] = data_torch
+            elif ".1." in name:
+                self._altup_proj[1] = data_torch
+            elif ".2." in name:
+                self._altup_proj[2] = data_torch
+            else:
+                raise ValueError(f"Unknown name: {name}")
+            out = self._stack_matrices(self._altup_proj)
+            if out is not None:
+                return [(self.map_tensor_name("model.altup_projections.weight"), out)]
+            else:
+                return []
+
+        return super().modify_tensors(data_torch, name, bid)
+
+
 @ModelBase.register("Starcoder2ForCausalLM")
 class StarCoder2Model(TextModel):
     model_arch = gguf.MODEL_ARCH.STARCODER2
@@ -4620,6 +4806,14 @@ class ARwkv7Model(Rwkv7Model):
 class MambaModel(TextModel):
     model_arch = gguf.MODEL_ARCH.MAMBA
 
+    def __init__(self, dir_model: Path, *args, **kwargs):
+        # Avoid using AutoConfig for hparams
+        hparams = kwargs.pop("hparams", None)
+        if hparams is None:
+            with open(dir_model / "config.json", "r", encoding="utf-8") as f:
+                hparams = json.load(f)
+        super().__init__(dir_model, *args, hparams=hparams, **kwargs)
+
     def set_vocab(self):
         vocab_size = self.hparams["vocab_size"]
         # Round vocab size to next multiple of 8
@@ -4694,6 +4888,100 @@ class MambaModel(TextModel):
         return [(new_name, data_torch)]
 
 
+@ModelBase.register("Mamba2ForCausalLM")
+class Mamba2Model(TextModel):
+    model_arch = gguf.MODEL_ARCH.MAMBA2
+
+    def __init__(self, dir_model: Path, *args, **kwargs):
+        # Avoid using AutoConfig for hparams
+        # It wrongly assumes all Mamba2 models are Mamba-Codestral-7B-v0.1
+        hparams = kwargs.pop("hparams", None)
+        if hparams is None:
+            with open(dir_model / "config.json", "r", encoding="utf-8") as f:
+                hparams = json.load(f)
+        super().__init__(dir_model, *args, hparams=hparams, **kwargs)
+
+    def set_vocab(self):
+        vocab_size = self.hparams["vocab_size"]
+        # Round vocab size to next multiple of 16
+        pad_vocab = self.hparams.get("pad_vocab_size_multiple", 16)
+        # pad using ceiling division
+        # ref: https://stackoverflow.com/a/17511341/22827863
+        vocab_size = -(vocab_size // -pad_vocab) * pad_vocab
+        self.hparams["vocab_size"] = vocab_size
+
+        if (self.dir_model / "tokenizer.model").is_file():
+            self._set_vocab_sentencepiece()
+        elif (self.dir_model / "tokenizer.model.v3").is_file():
+            # mamba-codestral
+            raise NotImplementedError(f"Please rename {self.dir_model / 'tokenizer.model.v3'} to {self.dir_model / 'tokenizer.model'}")
+        elif (self.dir_model / "tokenizer.json").is_file():
+            self._set_vocab_gpt2()
+        else:
+            # Use the GPT-NeoX tokenizer when no tokenizer files are present
+            self._set_vocab_builtin("gpt-neox", vocab_size)
+
+    def set_gguf_parameters(self):
+        d_model = self.find_hparam(["hidden_size", "d_model", "dim"])
+        d_conv  = self.find_hparam(["conv_kernel",       "d_conv"],  optional=True) or 4
+        d_inner = self.find_hparam(["intermediate_size", "d_inner"], optional=True) or 2 * d_model
+        d_state = self.find_hparam(["state_size",        "d_state"], optional=True) or 128
+        head_dim = self.find_hparam(["head_dim"],                    optional=True) or 64
+        n_group = self.find_hparam(["n_groups"],                     optional=True) or 1
+
+        rms_norm_eps = self.find_hparam(["layer_norm_epsilon", "rms_norm_eps"], optional=True) or 1e-5
+
+        # Fail early for models which don't have a block expansion factor of 2
+        # TODO: does this really matter?
+        assert d_inner == 2 * d_model
+        assert d_inner % head_dim == 0
+
+        self.gguf_writer.add_context_length(2**20)  # arbitrary value; for those who use the default
+        self.gguf_writer.add_embedding_length(d_model)
+        self.gguf_writer.add_feed_forward_length(0)  # unused, but seemingly required when loading
+        self.gguf_writer.add_head_count(0)  # unused, but seemingly required when loading
+        self.gguf_writer.add_block_count(self.block_count)
+        self.gguf_writer.add_ssm_conv_kernel(d_conv)
+        self.gguf_writer.add_ssm_inner_size(d_inner)
+        self.gguf_writer.add_ssm_state_size(d_state)
+        self.gguf_writer.add_ssm_time_step_rank(d_inner // head_dim)
+        self.gguf_writer.add_ssm_group_count(n_group)
+        self.gguf_writer.add_layer_norm_rms_eps(rms_norm_eps)
+        self.gguf_writer.add_file_type(self.ftype)
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+
+        if name.startswith("model.backbone") or name.startswith("model.lm_head"):
+            # map Mamba-Codestral-7B-v0.1 tensor names to the names used by Mamba-2
+            name = name.removeprefix("model.")
+
+        if name.endswith(".dt_bias"):
+            name = name.rpartition(".dt_bias")[0] + ".dt_proj.bias"
+
+        new_name = self.map_tensor_name(name)
+
+        if self.match_model_tensor_name(new_name, gguf.MODEL_TENSOR.SSM_CONV1D, bid):
+            data_torch = data_torch.squeeze()
+        elif any(self.match_model_tensor_name(new_name, t, bid, suffix="") for t in [
+            gguf.MODEL_TENSOR.SSM_A,
+            gguf.MODEL_TENSOR.SSM_D,
+        ]):
+            # unsqueeze A to use similar shape semantics as Mamba-1
+            # (D is also unsqueezed, but for more straightforward broadcast internally)
+            data_torch = data_torch.reshape((*data_torch.shape, 1))
+        elif self.match_model_tensor_name(new_name, gguf.MODEL_TENSOR.SSM_NORM, bid):
+            d_model = self.find_hparam(["hidden_size", "d_model", "dim"])
+            d_inner = self.find_hparam(["intermediate_size", "d_inner"], optional=True) or 2 * d_model
+            n_group = self.hparams.get("n_groups", 1)
+            data_torch = data_torch.reshape((n_group, d_inner // n_group))
+
+        if name.endswith(".A_log"):
+            logger.debug("A_log --> A ==> " + new_name)
+            data_torch = -torch.exp(data_torch)
+
+        yield (new_name, data_torch)
+
+
 @ModelBase.register("CohereForCausalLM")
 class CommandR2Model(TextModel):
     model_arch = gguf.MODEL_ARCH.COMMAND_R
@@ -4827,25 +5115,6 @@ class OlmoeModel(TextModel):
 class JinaBertV2Model(BertModel):
     model_arch = gguf.MODEL_ARCH.JINA_BERT_V2
 
-    def __init__(self, *args, **kwargs):
-        super().__init__(*args, **kwargs)
-        self.intermediate_size = self.hparams["intermediate_size"]
-
-    def get_tensors(self):
-        for name, data in super().get_tensors():
-            if 'gated_layer' in name:
-                d1 = data[:self.intermediate_size, :]
-                name1 = name.replace('gated_layers', 'gated_layers_w')
-                name1 = name1.replace('up_gated_layer', 'gated_layers_v')
-                d2 = data[self.intermediate_size:, :]
-                name2 = name.replace('gated_layers', 'gated_layers_v')
-                name2 = name2.replace('up_gated_layer', 'gated_layers_w')
-                yield name1, d1
-                yield name2, d2
-                continue
-
-            yield name, data
-
     def set_vocab(self):
         tokenizer_class = 'BertTokenizer'
         with open(self.dir_model / "tokenizer_config.json", "r", encoding="utf-8") as f:
@@ -4858,16 +5127,6 @@ class JinaBertV2Model(BertModel):
             self.gguf_writer.add_token_type_count(2)
         else:
             raise NotImplementedError(f'Tokenizer {tokenizer_class} is not supported for JinaBertModel')
-        self.gguf_writer.add_add_bos_token(True)
-        self.gguf_writer.add_add_eos_token(True)
-
-    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
-        # if name starts with "bert.", remove the prefix
-        # e.g. https://huggingface.co/jinaai/jina-reranker-v1-tiny-en
-        if name.startswith("bert."):
-            name = name[5:]
-
-        return super().modify_tensors(data_torch, name, bid)
 
 
 @ModelBase.register("OpenELMForCausalLM")
@@ -5109,9 +5368,7 @@ class DeepseekModel(TextModel):
     def set_gguf_parameters(self):
         super().set_gguf_parameters()
         hparams = self.hparams
-        if "head_dim" in hparams:
-            rope_dim = hparams["head_dim"]
-        else:
+        if (rope_dim := hparams.get("head_dim")) is None:
             rope_dim = hparams["hidden_size"] // hparams["num_attention_heads"]
 
         self.gguf_writer.add_rope_dimension_count(rope_dim)
@@ -5315,6 +5572,34 @@ class DeepseekV2Model(TextModel):
                 raise ValueError(f"Unprocessed experts: {experts}")
 
 
+@ModelBase.register("Dots1ForCausalLM")
+class Dots1Model(Qwen2MoeModel):
+    model_arch = gguf.MODEL_ARCH.DOTS1
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        self.hparams["num_experts"] = self.hparams["n_routed_experts"]
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        self.gguf_writer.add_leading_dense_block_count(self.hparams["first_k_dense_replace"])
+        self.gguf_writer.add_expert_shared_count(self.hparams["n_shared_experts"])
+        self.gguf_writer.add_expert_weights_scale(self.hparams["routed_scaling_factor"])
+        self.gguf_writer.add_expert_weights_norm(self.hparams["norm_topk_prob"])
+
+        if self.hparams["scoring_func"] == "noaux_tc":
+            self.gguf_writer.add_expert_gating_func(gguf.ExpertGatingFuncType.SIGMOID)
+        else:
+            raise ValueError(f"Unsupported scoring_func value: {self.hparams['scoring_func']}")
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None):
+        if name.endswith("e_score_correction_bias"):
+            name = name.replace("e_score_correction_bias", "e_score_correction.bias")
+        if "shared_experts" in name:
+            return [(self.map_tensor_name(name), data_torch)]
+        return super().modify_tensors(data_torch, name, bid)
+
+
 @ModelBase.register("PLMForCausalLM")
 class PLMModel(TextModel):
     model_arch = gguf.MODEL_ARCH.PLM
@@ -5443,9 +5728,6 @@ class T5Model(TextModel):
         special_vocab = gguf.SpecialVocab(self.dir_model, n_vocab=len(tokens))
         special_vocab.add_to_gguf(self.gguf_writer)
 
-        self.gguf_writer.add_add_bos_token(False)
-        self.gguf_writer.add_add_eos_token(True)
-
     def set_gguf_parameters(self):
         if (n_ctx := self.find_hparam(["n_positions"], optional=True)) is None:
             logger.warning("Couldn't find context length in config.json, assuming default value of 512")
@@ -5583,9 +5865,6 @@ class T5EncoderModel(TextModel):
         special_vocab = gguf.SpecialVocab(self.dir_model, n_vocab=len(tokens))
         special_vocab.add_to_gguf(self.gguf_writer)
 
-        self.gguf_writer.add_add_bos_token(False)
-        self.gguf_writer.add_add_eos_token(True)
-
     def set_gguf_parameters(self):
         if (n_ctx := self.find_hparam(["n_positions"], optional=True)) is None:
             logger.warning("Couldn't find context length in config.json, assuming default value of 512")
@@ -5973,7 +6252,8 @@ class ExaoneModel(TextModel):
         if rope_scaling := self.find_hparam(["rope_scaling"], optional=True):
             if rope_scaling.get("rope_type", '').lower() == "llama3":
                 base = self.hparams.get("rope_theta", 10000.0)
-                dim = self.hparams.get("head_dim", self.hparams["hidden_size"] // self.hparams["num_attention_heads"])
+                if (dim := self.hparams.get("head_dim")) is None:
+                    dim = self.hparams["hidden_size"] // self.hparams["num_attention_heads"]
                 freqs = 1.0 / (base ** (torch.arange(0, dim, 2, dtype=torch.float32) / dim))
 
                 factor = rope_scaling.get("factor", 8.0)
@@ -6085,7 +6365,8 @@ class BailingMoeModel(TextModel):
     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"]
+        if (rope_dim := hparams.get("head_dim")) is None:
+            rope_dim = hparams["hidden_size"] // hparams["num_attention_heads"]
 
         self.gguf_writer.add_rope_dimension_count(rope_dim)
         rope_scaling = self.hparams.get("rope_scaling") or {}
@@ -6117,7 +6398,8 @@ class BailingMoeModel(TextModel):
         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
+        if (head_dim := self.hparams.get("head_dim")) is None:
+            head_dim = n_embd // n_head
 
         output_name = self.format_tensor_name(gguf.MODEL_TENSOR.OUTPUT)
 
@@ -6378,8 +6660,8 @@ def parse_args() -> argparse.Namespace:
         help="model is executed on big endian machine",
     )
     parser.add_argument(
-        "model", type=Path,
-        help="directory containing model file",
+        "model", type=str,
+        help="directory containing model file or huggingface repository ID (if --remote)",
         nargs="?",
     )
     parser.add_argument(
@@ -6460,12 +6742,20 @@ def get_model_architecture(hparams: dict[str, Any], model_type: ModelType) -> st
     # maybe we should fallback to text model's arch in that case, since not many models have both
     text_config = hparams.get("text_config", {})
     vision_config = hparams.get("vision_config", {})
-    arch = hparams["architectures"][0]
+    arch = None
+    if (arches := hparams.get("architectures")) is not None and len(arches) > 0:
+        arch = arches[0]
+    elif "ssm_cfg" in hparams:
+        # For non-hf Mamba and Mamba2 models
+        arch = hparams["ssm_cfg"].get("layer", "Mamba") + "ForCausalLM"
+
     # if "architectures" is found in the sub-config, use that instead
     if model_type == ModelType.TEXT and text_config.get("architectures") is not None:
         arch = text_config["architectures"][0]
     elif model_type == ModelType.MMPROJ and vision_config.get("architectures") is not None:
         arch = vision_config["architectures"][0]
+    if arch is None:
+        raise ValueError("Failed to detect model architecture")
     return arch
 
 
@@ -6482,18 +6772,20 @@ def main() -> None:
     else:
         logging.basicConfig(level=logging.INFO)
 
-    dir_model = args.model
-
     if args.remote:
+        hf_repo_id = args.model
         from huggingface_hub import snapshot_download
         local_dir = snapshot_download(
-            repo_id=str(dir_model),
+            repo_id=hf_repo_id,
             allow_patterns=["LICENSE", "*.json", "*.md", "*.txt", "tokenizer.model"])
         dir_model = Path(local_dir)
         logger.info(f"Downloaded config and tokenizer to {local_dir}")
+    else:
+        hf_repo_id = None
+        dir_model = Path(args.model)
 
     if not dir_model.is_dir():
-        logger.error(f'Error: {args.model} is not a directory')
+        logger.error(f'Error: {dir_model} is not a directory')
         sys.exit(1)
 
     ftype_map: dict[str, gguf.LlamaFileType] = {
@@ -6513,9 +6805,9 @@ def main() -> None:
 
     if args.outfile is not None:
         fname_out = args.outfile
-    elif args.remote:
+    elif hf_repo_id:
         # if remote, use the model ID as the output file name
-        fname_out = Path("./" + str(args.model).replace("/", "-") + "-{ftype}.gguf")
+        fname_out = Path("./" + hf_repo_id.replace("/", "-") + "-{ftype}.gguf")
     else:
         fname_out = dir_model
 
@@ -6544,7 +6836,7 @@ def main() -> None:
                                      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,
-                                     remote_hf_model_id=str(args.model) if args.remote else None)
+                                     remote_hf_model_id=hf_repo_id)
 
         if args.vocab_only:
             logger.info("Exporting model vocab...")
diff --git a/docs/backend/CANN.md b/docs/backend/CANN.md
index a5ba617ca..2b001f09a 100755
--- a/docs/backend/CANN.md
+++ b/docs/backend/CANN.md
@@ -8,6 +8,7 @@
  - [DataType Supports](#datatype-supports)
  - [Docker](#docker)
  - [Linux](#linux)
+ - [Environment variable setup](#environment-variable-setup)
  - [TODO](#todo)
 
 
@@ -290,5 +291,24 @@ Authors from Peking University: Bizhao Shi (bshi@pku.edu.cn), Yuxin Yang (yxyang
 
 We would like to thank Tuo Dai, Shanni Li, and all of the project maintainers from Huawei Technologies Co., Ltd for their help during the code development and pull request.
 
+## Environment variable setup
+
+### GGML_CANN_ASYNC_MODE
+
+Enables asynchronous operator submission. Disabled by default.
+
+### GGML_CANN_MEM_POOL
+
+Specifies the memory pool management strategy:
+
+- vmm: Utilizes a virtual memory manager pool. If hardware support for VMM is unavailable, falls back to the legacy (leg) memory pool.
+
+- prio: Employs a priority queue-based memory pool management.
+- leg: Uses a fixed-size buffer pool.
+
+### GGML_CANN_DISABLE_BUF_POOL_CLEAN
+
+Controls automatic cleanup of the memory pool. This option is only effective when using the prio or leg memory pool strategies.
+
 ## TODO
 - Support more models and data types.
diff --git a/docs/backend/SYCL.md b/docs/backend/SYCL.md
index 249e73451..6e9b88935 100644
--- a/docs/backend/SYCL.md
+++ b/docs/backend/SYCL.md
@@ -757,7 +757,7 @@ 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_OPT | 0 (default) or 1 | Disable optimize features for Intel GPUs. (Recommended to 1 for intel devices older than Gen 10) |
 | 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. |
 | GGML_SYCL_DISABLE_DNN | 0 (default) or 1 | Disable running computations through oneDNN and always use oneMKL. |
 | ZES_ENABLE_SYSMAN | 0 (default) or 1 | Support to get free memory of GPU by sycl::aspect::ext_intel_free_memory.<br>Recommended to use when --split-mode = layer |
diff --git a/docs/build-s390x.md b/docs/build-s390x.md
new file mode 100644
index 000000000..4c9ebb271
--- /dev/null
+++ b/docs/build-s390x.md
@@ -0,0 +1,246 @@
+> [!IMPORTANT]
+> This build documentation is specific only to IBM Z & LinuxONE mainframes (s390x). You can find the build documentation for other architectures: [build.md](build.md).
+
+# Build llama.cpp locally (for s390x)
+
+The main product of this project is the `llama` library. Its C-style interface can be found in [include/llama.h](../include/llama.h).
+
+The project also includes many example programs and tools using the `llama` library. The examples range from simple, minimal code snippets to sophisticated sub-projects such as an OpenAI-compatible HTTP server.
+
+**To get the code:**
+
+```bash
+git clone https://github.com/ggml-org/llama.cpp
+cd llama.cpp
+```
+
+## CPU Build with BLAS
+
+Building llama.cpp with BLAS support is highly recommended as it has shown to provide performance improvements. Make sure to have OpenBLAS installed in your environment.
+
+```bash
+cmake -S . -B build             \
+    -DCMAKE_BUILD_TYPE=Release  \
+    -DGGML_BLAS=ON              \
+    -DGGML_BLAS_VENDOR=OpenBLAS
+
+cmake --build build --config Release -j $(nproc)
+```
+
+**Notes**:
+
+-   For faster repeated compilation, install [ccache](https://ccache.dev/)
+-   By default, VXE/VXE2 is enabled. To disable it (not recommended):
+
+    ```bash
+    cmake -S . -B build             \
+        -DCMAKE_BUILD_TYPE=Release  \
+        -DGGML_BLAS=ON              \
+        -DGGML_BLAS_VENDOR=OpenBLAS \
+        -DGGML_VXE=OFF
+
+    cmake --build build --config Release -j $(nproc)
+    ```
+
+-   By default, NNPA is enabled when available. To disable it (not recommended):
+
+    ```bash
+    cmake -S . -B build             \
+        -DCMAKE_BUILD_TYPE=Release  \
+        -DGGML_BLAS=ON              \
+        -DGGML_BLAS_VENDOR=OpenBLAS \
+        -DGGML_NNPA=OFF
+
+    cmake --build build --config Release -j $(nproc)
+    ```
+
+-   For debug builds:
+
+    ```bash
+    cmake -S . -B build             \
+        -DCMAKE_BUILD_TYPE=Debug    \
+        -DGGML_BLAS=ON              \
+        -DGGML_BLAS_VENDOR=OpenBLAS
+    cmake --build build --config Debug -j $(nproc)
+    ```
+
+-   For static builds, add `-DBUILD_SHARED_LIBS=OFF`:
+
+    ```bash
+    cmake -S . -B build             \
+        -DCMAKE_BUILD_TYPE=Release  \
+        -DGGML_BLAS=ON              \
+        -DGGML_BLAS_VENDOR=OpenBLAS \
+        -DBUILD_SHARED_LIBS=OFF
+
+    cmake --build build --config Release -j $(nproc)
+    ```
+
+## Getting GGUF Models
+
+All models need to be converted to Big-Endian. You can achieve this in three cases:
+
+1. **Use pre-converted models verified for use on IBM Z & LinuxONE (easiest)**
+
+    ![File Type - gguf](https://img.shields.io/badge/File_Type-gguf-fff)
+
+    You can find popular models pre-converted and verified at [s390x Ready Models](https://huggingface.co/collections/taronaeo/s390x-ready-models-672765393af438d0ccb72a08).
+
+    These models have already been converted from `safetensors` to `GGUF Big-Endian` and their respective tokenizers verified to run correctly on IBM z15 and later system.
+
+2. **Convert safetensors model to GGUF Big-Endian directly (recommended)**
+
+    ![File Type - safetensors](https://img.shields.io/badge/File_Type-safetensors-da1e28)
+
+    The model you are trying to convert must be in `safetensors` file format (for example [IBM Granite 3.3 2B](https://huggingface.co/ibm-granite/granite-3.3-2b-instruct)). Make sure you have downloaded the model repository for this case.
+
+    ```bash
+    python3 convert_hf_to_gguf.py \
+        --outfile model-name-be.f16.gguf \
+        --outtype f16 \
+        --bigendian \
+        model-directory/
+    ```
+
+    For example,
+
+    ```bash
+    python3 convert_hf_to_gguf.py \
+        --outfile granite-3.3-2b-instruct-be.f16.gguf \
+        --outtype f16 \
+        --bigendian \
+        granite-3.3-2b-instruct/
+    ```
+
+3. **Convert existing GGUF Little-Endian model to Big-Endian**
+
+    ![File Type - gguf](https://img.shields.io/badge/File_Type-gguf-fff)
+
+    The model you are trying to convert must be in `gguf` file format (for example [IBM Granite 3.3 2B](https://huggingface.co/ibm-granite/granite-3.3-2b-instruct-GGUF)). Make sure you have downloaded the model file for this case.
+
+    ```bash
+    python3 gguf-py/gguf/scripts/gguf_convert_endian.py model-name.f16.gguf BIG
+    ```
+
+    For example,
+
+    ```bash
+    python3 gguf-py/gguf/scripts/gguf_convert_endian.py granite-3.3-2b-instruct-le.f16.gguf BIG
+    mv granite-3.3-2b-instruct-le.f16.gguf granite-3.3-2b-instruct-be.f16.gguf
+    ```
+
+    **Notes:**
+
+    - The GGUF endian conversion script may not support all data types at the moment and may fail for some models/quantizations. When that happens, please try manually converting the safetensors model to GGUF Big-Endian via Step 2.
+
+## IBM Accelerators
+
+### 1. SIMD Acceleration
+
+Only available in IBM z15 or later system with the `-DGGML_VXE=ON` (turned on by default) compile flag. No hardware acceleration is possible with llama.cpp with older systems, such as IBM z14/arch12. In such systems, the APIs can still run but will use a scalar implementation.
+
+### 2. NNPA Vector Intrinsics Acceleration
+
+Only available in IBM z16 or later system with the `-DGGML_NNPA=ON` (turned on when available) compile flag. No hardware acceleration is possible with llama.cpp with older systems, such as IBM z15/arch13. In such systems, the APIs can still run but will use a scalar implementation.
+
+### 3. zDNN Accelerator
+
+_Only available in IBM z16 or later system. No direction at the moment._
+
+### 4. Spyre Accelerator
+
+_No direction at the moment._
+
+## Performance Tuning
+
+### 1. Virtualization Setup
+
+It is strongly recommended to use only LPAR (Type-1) virtualization to get the most performance.
+
+Note: Type-2 virtualization is not supported at the moment, while you can get it running, the performance will not be the best.
+
+### 2. IFL (Core) Count
+
+It is recommended to allocate a minimum of 8 shared IFLs assigned to the LPAR. Increasing the IFL count past 8 shared IFLs will only improve Prompt Processing performance but not Token Generation.
+
+Note: IFL count does not equate to vCPU count.
+
+### 3. SMT vs NOSMT (Simultaneous Multithreading)
+
+It is strongly recommended to disable SMT via the kernel boot parameters as it negatively affects performance. Please refer to your Linux distribution's guide on disabling SMT via kernel boot parameters.
+
+### 4. BLAS vs NOBLAS
+
+IBM VXE/VXE2 SIMD acceleration depends on the BLAS implementation. It is strongly recommended to use BLAS.
+
+## Frequently Asked Questions (FAQ)
+
+1. I'm getting the following error message while trying to load a model: `gguf_init_from_file_impl: failed to load model: this GGUF file version 50331648 is extremely large, is there a mismatch between the host and model endianness?`
+
+    Answer: Please ensure that the model you have downloaded/converted is GGUFv3 Big-Endian. These models are usually denoted with the `-be` suffix, i.e., `granite-3.3-2b-instruct-be.F16.gguf`.
+
+    You may refer to the [Getting GGUF Models](#getting-gguf-models) section to manually convert a `safetensors` model to `GGUF` Big Endian.
+
+2. I'm getting extremely poor performance when running inference on a model
+
+    Answer: Please refer to the [Appendix B: SIMD Support Matrix](#appendix-b-simd-support-matrix) to check if your model quantization is supported by SIMD acceleration.
+
+3. I'm building on IBM z17 and getting the following error messages: `invalid switch -march=z17`
+
+    Answer: Please ensure that your GCC compiler is of minimum GCC 15.1.0 version, and have `binutils` updated to the latest version. If this does not fix the problem, kindly open an issue.
+
+## Getting Help on IBM Z & LinuxONE
+
+1. **Bugs, Feature Requests**
+
+    Please file an issue in llama.cpp and ensure that the title contains "s390x".
+
+2. **Other Questions**
+
+    Please reach out directly to [aionz@us.ibm.com](mailto:aionz@us.ibm.com).
+
+## Appendix A: Hardware Support Matrix
+
+|         | Support | Minimum Compiler Version |
+| ------- | ------- | ------------------------ |
+| IBM z15 | ✅      |                          |
+| IBM z16 | ✅      |                          |
+| IBM z17 | ✅      | GCC 15.1.0               |
+
+-   ✅ - supported and verified to run as intended
+-   🚫 - unsupported, we are unlikely able to provide support
+
+## Appendix B: SIMD Support Matrix
+
+|            | VX/VXE/VXE2 | NNPA | zDNN | Spyre |
+| ---------- | ----------- | ---- | ---- | ----- |
+| FP32       | ✅          | ✅   | ❓   | ❓    |
+| FP16       | ✅          | ✅   | ❓   | ❓    |
+| BF16       | 🚫          | 🚫   | ❓   | ❓    |
+| Q4_0       | ✅          | ✅   | ❓   | ❓    |
+| Q4_1       | ✅          | ✅   | ❓   | ❓    |
+| Q5_0       | 🚫          | 🚫   | ❓   | ❓    |
+| Q5_1       | 🚫          | 🚫   | ❓   | ❓    |
+| Q8_0       | ✅          | ✅   | ❓   | ❓    |
+| Q2_K       | 🚫          | 🚫   | ❓   | ❓    |
+| Q3_K       | ✅          | ✅   | ❓   | ❓    |
+| Q4_K       | ✅          | ✅   | ❓   | ❓    |
+| Q5_K       | ✅          | ✅   | ❓   | ❓    |
+| Q6_K       | ✅          | ✅   | ❓   | ❓    |
+| TQ1_0      | 🚫          | 🚫   | ❓   | ❓    |
+| TQ2_0      | 🚫          | 🚫   | ❓   | ❓    |
+| IQ2_XXS    | 🚫          | 🚫   | ❓   | ❓    |
+| IQ2_XS     | 🚫          | 🚫   | ❓   | ❓    |
+| IQ2_S      | 🚫          | 🚫   | ❓   | ❓    |
+| IQ3_XXS    | 🚫          | 🚫   | ❓   | ❓    |
+| IQ3_S      | 🚫          | 🚫   | ❓   | ❓    |
+| IQ1_S      | 🚫          | 🚫   | ❓   | ❓    |
+| IQ1_M      | 🚫          | 🚫   | ❓   | ❓    |
+| IQ4_NL     | ✅          | ✅   | ❓   | ❓    |
+| IQ4_XS     | ✅          | ✅   | ❓   | ❓    |
+| FP32->FP16 | 🚫          | ✅   | ❓   | ❓    |
+| FP16->FP32 | 🚫          | ✅   | ❓   | ❓    |
+
+-   ✅ - acceleration available
+-   🚫 - acceleration unavailable, will still run using scalar implementation
+-   ❓ - acceleration unknown, please contribute if you can test it yourself
diff --git a/docs/build.md b/docs/build.md
index 32717a793..2e0b5d970 100644
--- a/docs/build.md
+++ b/docs/build.md
@@ -1,5 +1,9 @@
 # Build llama.cpp locally
 
+The main product of this project is the `llama` library. Its C-style interface can be found in [include/llama.h](../include/llama.h).
+
+The project also includes many example programs and tools using the `llama` library. The examples range from simple, minimal code snippets to sophisticated sub-projects such as an OpenAI-compatible HTTP server.
+
 **To get the Code:**
 
 ```bash
@@ -553,6 +557,10 @@ ninja
 
 To read documentation for how to build on Android, [click here](./android.md)
 
+## IBM Z & LinuxONE
+
+To read documentation for how to build on IBM Z & LinuxONE, [click here](./build-s390x.md)
+
 ## Notes about GPU-accelerated backends
 
 The GPU may still be used to accelerate some parts of the computation even when using the `-ngl 0` option. You can fully disable GPU acceleration by using `--device none`.
diff --git a/docs/docker.md b/docs/docker.md
index f8f0573c1..cbb333ee3 100644
--- a/docs/docker.md
+++ b/docs/docker.md
@@ -25,6 +25,9 @@ Additionally, there the following images, similar to the above:
 - `ghcr.io/ggml-org/llama.cpp:full-intel`: Same as `full` but compiled with SYCL support. (platforms: `linux/amd64`)
 - `ghcr.io/ggml-org/llama.cpp:light-intel`: Same as `light` but compiled with SYCL support. (platforms: `linux/amd64`)
 - `ghcr.io/ggml-org/llama.cpp:server-intel`: Same as `server` but compiled with SYCL support. (platforms: `linux/amd64`)
+- `ghcr.io/ggml-org/llama.cpp:full-vulkan`: Same as `full` but compiled with Vulkan support. (platforms: `linux/amd64`)
+- `ghcr.io/ggml-org/llama.cpp:light-vulkan`: Same as `light` but compiled with Vulkan support. (platforms: `linux/amd64`)
+- `ghcr.io/ggml-org/llama.cpp:server-vulkan`: Same as `server` but compiled with Vulkan 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).
 
diff --git a/docs/function-calling.md b/docs/function-calling.md
index fd3db9bd1..37eacaf31 100644
--- a/docs/function-calling.md
+++ b/docs/function-calling.md
@@ -11,7 +11,7 @@ Function calling is supported for all models (see https://github.com/ggml-org/ll
   - 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)
+  - Qwen 2.5 Coder
   - Mistral Nemo
   - Firefunction v2
   - Command R7B
diff --git a/docs/install.md b/docs/install.md
index 4971c1828..7200bf9b7 100644
--- a/docs/install.md
+++ b/docs/install.md
@@ -1,28 +1,42 @@
 # Install pre-built version of llama.cpp
 
-## Homebrew
+| Install via | Windows | Mac | Linux |
+|-------------|---------|-----|-------|
+| Winget      | ✅      |      |      |
+| Homebrew    |         | ✅   | ✅   |
+| MacPorts    |         | ✅   |      |
+| Nix         |         | ✅   | ✅   |
 
-On Mac and Linux, the homebrew package manager can be used via
+## Winget (Windows)
+
+```sh
+winget install llama.cpp
+```
+
+The package is automatically updated with new `llama.cpp` releases. More info: https://github.com/ggml-org/llama.cpp/issues/8188
+
+## Homebrew (Mac and Linux)
 
 ```sh
 brew install llama.cpp
 ```
+
 The formula is automatically updated with new `llama.cpp` releases. More info: https://github.com/ggml-org/llama.cpp/discussions/7668
 
-## MacPorts
+## MacPorts (Mac)
 
 ```sh
 sudo port install llama.cpp
 ```
-see also: https://ports.macports.org/port/llama.cpp/details/
 
-## Nix
+See also: https://ports.macports.org/port/llama.cpp/details/
 
-On Mac and Linux, the Nix package manager can be used via
+## Nix (Mac and Linux)
 
 ```sh
 nix profile install nixpkgs#llama-cpp
 ```
+
 For flake enabled installs.
 
 Or
@@ -34,13 +48,3 @@ nix-env --file '<nixpkgs>' --install --attr llama-cpp
 For non-flake enabled installs.
 
 This expression is automatically updated within the [nixpkgs repo](https://github.com/NixOS/nixpkgs/blob/nixos-24.05/pkgs/by-name/ll/llama-cpp/package.nix#L164).
-
-## Flox
-
-On Mac and Linux, Flox can be used to install llama.cpp within a Flox environment via
-
-```sh
-flox install llama-cpp
-```
-
-Flox follows the nixpkgs build of llama.cpp.
diff --git a/docs/multimodal.md b/docs/multimodal.md
index e849c2a0b..edbd081df 100644
--- a/docs/multimodal.md
+++ b/docs/multimodal.md
@@ -107,3 +107,7 @@ NOTE: some models may require large context window, for example: `-c 8192`
 (tool_name) -hf ggml-org/Qwen2.5-Omni-3B-GGUF
 (tool_name) -hf ggml-org/Qwen2.5-Omni-7B-GGUF
 ```
+
+## Finding more models:
+
+GGUF models on Huggingface with vision capabilities can be found here: https://huggingface.co/models?pipeline_tag=image-text-to-text&sort=trending&search=gguf
diff --git a/examples/Miku.sh b/examples/Miku.sh
index 0f6c8c878..9492bfedc 100755
--- a/examples/Miku.sh
+++ b/examples/Miku.sh
@@ -1,4 +1,4 @@
-#!/bin/bash
+#!/usr/bin/env bash
 set -e
 
 AI_NAME="${AI_NAME:-Miku}"
diff --git a/examples/batched.swift/Sources/main.swift b/examples/batched.swift/Sources/main.swift
index 514989e34..fd90bbec5 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_self_seq_cp(context, 0, Int32(i), 0, batch.n_tokens)
+    llama_memory_seq_cp(llama_get_memory(context), 0, Int32(i), 0, batch.n_tokens)
 }
 
 if n_parallel > 1 {
diff --git a/examples/chat-13B.sh b/examples/chat-13B.sh
index 1828903c3..f025a47cb 100755
--- a/examples/chat-13B.sh
+++ b/examples/chat-13B.sh
@@ -1,4 +1,4 @@
-#!/bin/bash
+#!/usr/bin/env bash
 
 set -e
 
diff --git a/examples/chat-persistent.sh b/examples/chat-persistent.sh
index 9d761ebb8..d6b6cb951 100755
--- a/examples/chat-persistent.sh
+++ b/examples/chat-persistent.sh
@@ -1,4 +1,4 @@
-#!/bin/bash
+#!/usr/bin/env bash
 
 set -euo pipefail
 
diff --git a/examples/chat-vicuna.sh b/examples/chat-vicuna.sh
index ffdd20084..c930962fd 100755
--- a/examples/chat-vicuna.sh
+++ b/examples/chat-vicuna.sh
@@ -1,4 +1,4 @@
-#!/bin/bash
+#!/usr/bin/env bash
 
 set -e
 
diff --git a/examples/chat.sh b/examples/chat.sh
index 9f85d1e26..5fec46d17 100755
--- a/examples/chat.sh
+++ b/examples/chat.sh
@@ -1,4 +1,4 @@
-#!/bin/bash
+#!/usr/bin/env bash
 
 #
 # Temporary script - will be removed in the future
diff --git a/examples/embedding/embedding.cpp b/examples/embedding/embedding.cpp
index 71f700877..0ec2999a0 100644
--- a/examples/embedding/embedding.cpp
+++ b/examples/embedding/embedding.cpp
@@ -37,7 +37,7 @@ static void batch_decode(llama_context * ctx, llama_batch & batch, float * outpu
     const enum llama_pooling_type pooling_type = llama_pooling_type(ctx);
 
     // clear previous kv_cache values (irrelevant for embeddings)
-    llama_kv_self_clear(ctx);
+    llama_memory_clear(llama_get_memory(ctx), true);
 
     // run model
     LOG_INF("%s: n_tokens = %d, n_seq = %d\n", __func__, batch.n_tokens, n_seq);
@@ -133,10 +133,36 @@ int main(int argc, char ** argv) {
     // max batch size
     const uint64_t n_batch = params.n_batch;
 
+    // get added sep and eos token, if any
+    const std::string added_sep_token = llama_vocab_get_add_sep(vocab) ? llama_vocab_get_text(vocab, llama_vocab_sep(vocab)) : "";
+    const std::string added_eos_token = llama_vocab_get_add_eos(vocab) ? llama_vocab_get_text(vocab, llama_vocab_eos(vocab)) : "";
+
     // tokenize the prompts and trim
     std::vector<std::vector<int32_t>> inputs;
     for (const auto & prompt : prompts) {
-        auto inp = common_tokenize(ctx, prompt, true, true);
+        std::vector<llama_token> inp;
+
+        // split classification pairs and insert expected separator tokens
+        if (pooling_type == LLAMA_POOLING_TYPE_RANK && prompt.find(params.cls_sep) != std::string::npos) {
+            std::vector<std::string> pairs = split_lines(prompt, params.cls_sep);
+            std::string final_prompt;
+
+            for (size_t i = 0; i < pairs.size(); i++) {
+                final_prompt += pairs[i];
+                if (i != pairs.size() - 1) {
+                    if (!added_eos_token.empty()) {
+                        final_prompt += added_eos_token;
+                    }
+                    if (!added_sep_token.empty()) {
+                        final_prompt += added_sep_token;
+                    }
+                }
+            }
+
+            inp = common_tokenize(ctx, final_prompt, true, true);
+        } else {
+            inp = common_tokenize(ctx, prompt, true, true);
+        }
         if (inp.size() > n_batch) {
             LOG_ERR("%s: number of tokens in input line (%lld) exceeds batch size (%lld), increase batch size and re-run\n",
                     __func__, (long long int) inp.size(), (long long int) n_batch);
@@ -145,11 +171,11 @@ int main(int argc, char ** argv) {
         inputs.push_back(inp);
     }
 
-    // check if the last token is SEP
+    // check if the last token is SEP/EOS
     // it should be automatically added by the tokenizer when 'tokenizer.ggml.add_eos_token' is set to 'true'
     for (auto & inp : inputs) {
-        if (inp.empty() || inp.back() != llama_vocab_sep(vocab)) {
-            LOG_WRN("%s: last token in the prompt is not SEP\n", __func__);
+        if (inp.empty() || (inp.back() != llama_vocab_sep(vocab) && inp.back() != llama_vocab_eos(vocab))) {
+            LOG_WRN("%s: last token in the prompt is not SEP or EOS\n", __func__);
             LOG_WRN("%s: 'tokenizer.ggml.add_eos_token' should be set to 'true' in the GGUF header\n", __func__);
         }
     }
@@ -236,9 +262,24 @@ int main(int argc, char ** argv) {
                 LOG("\n");
             }
         } else if (pooling_type == LLAMA_POOLING_TYPE_RANK) {
+            const uint32_t n_cls_out = llama_model_n_cls_out(model);
+            std::vector<std::string> cls_out_labels;
+
+            for (uint32_t i = 0; i < n_cls_out; i++) {
+                const char * label = llama_model_cls_label(model, i);
+                const std::string label_i(label == nullptr ? "" : label);
+                cls_out_labels.emplace_back(label_i.empty() ? std::to_string(i) : label_i);
+            }
+
             for (int j = 0; j < n_embd_count; j++) {
-                // NOTE: if you change this log - update the tests in ci/run.sh
-                LOG("rerank score %d: %8.3f\n", j, emb[j * n_embd]);
+                for (uint32_t i = 0; i < n_cls_out; i++) {
+                    // NOTE: if you change this log - update the tests in ci/run.sh
+                    if (n_cls_out == 1) {
+                        LOG("rerank score %d: %8.3f\n", j, emb[j * n_embd]);
+                    } else {
+                        LOG("rerank score %d: %8.3f [%s]\n", j, emb[j * n_embd + i], cls_out_labels[i].c_str());
+                    }
+                }
             }
         } else {
             // print the first part of the embeddings or for a single prompt, the full embedding
diff --git a/examples/eval-callback/eval-callback.cpp b/examples/eval-callback/eval-callback.cpp
index fb188f5a9..4afd80eb4 100644
--- a/examples/eval-callback/eval-callback.cpp
+++ b/examples/eval-callback/eval-callback.cpp
@@ -55,6 +55,8 @@ static void ggml_print_tensor(uint8_t * data, ggml_type type, const int64_t * ne
                         v = ggml_fp16_to_fp32(*(ggml_fp16_t *) &data[i]);
                     } else if (type == GGML_TYPE_F32) {
                         v = *(float *) &data[i];
+                    } else if (type == GGML_TYPE_I64) {
+                        v = (float) *(int64_t *) &data[i];
                     } else if (type == GGML_TYPE_I32) {
                         v = (float) *(int32_t *) &data[i];
                     } else if (type == GGML_TYPE_I16) {
@@ -134,6 +136,11 @@ static bool run(llama_context * ctx, const common_params & params) {
 
     std::vector<llama_token> tokens = common_tokenize(ctx, params.prompt, add_bos);
 
+    if (tokens.empty()) {
+        LOG_ERR("%s : there are not input tokens to process - (try to provide a prompt with '-p')\n", __func__);
+        return false;
+    }
+
     if (llama_decode(ctx, llama_batch_get_one(tokens.data(), tokens.size()))) {
         LOG_ERR("%s : failed to eval\n", __func__);
         return false;
diff --git a/examples/gritlm/gritlm.cpp b/examples/gritlm/gritlm.cpp
index 539bc4d60..bdab052c3 100644
--- a/examples/gritlm/gritlm.cpp
+++ b/examples/gritlm/gritlm.cpp
@@ -41,12 +41,11 @@ static std::vector<std::vector<float>> encode(llama_context * ctx, const std::ve
 
         // add input to batch (this increments n_tokens)
         for (int32_t j = 0; j < n_toks; j++) {
-            common_batch_add(batch, inputs[j], j, { 0 }, j >= n_inst);
+            common_batch_add(batch, inputs[j], j, { 0 }, true);
         }
 
         // clear previous kv_cache values (irrelevant for embeddings)
-        llama_kv_self_clear(ctx);
-        llama_set_embeddings(ctx, true);
+        llama_memory_clear(llama_get_memory(ctx), true);
         llama_set_causal_attn(ctx, false);
 
         // run model
@@ -102,8 +101,7 @@ static std::string generate(llama_context * ctx, llama_sampler * smpl, const std
 
     llama_token eos_token = llama_vocab_eos(vocab);
 
-    llama_kv_self_clear(ctx);
-    llama_set_embeddings(ctx, false);
+    llama_memory_clear(llama_get_memory(ctx), true);
     llama_set_causal_attn(ctx, true);
 
     llama_batch bat = llama_batch_init(llama_n_batch(ctx), 0, 1);
@@ -166,6 +164,8 @@ int main(int argc, char * argv[]) {
     llama_model_params mparams = common_model_params_to_llama(params);
     llama_context_params cparams = common_context_params_to_llama(params);
 
+    cparams.embeddings = true;
+
     llama_backend_init();
 
     llama_model * model = llama_model_load_from_file(params.model.path.c_str(), mparams);
@@ -213,6 +213,8 @@ int main(int argc, char * argv[]) {
         std::printf("Cosine similarity between \"%.50s\" and \"%.50s\" is: %.3f\n", queries[1].c_str(), documents[1].c_str(), cosine_sim_q1_d1);
     }
 
+    llama_set_embeddings(ctx, false);
+
     // ### Generation ###
     // GritLM models are not finetuned with system prompts, as you can just include system-like instructions together with your user instruction
     {
diff --git a/examples/jeopardy/jeopardy.sh b/examples/jeopardy/jeopardy.sh
index 07bcb3b8d..800df2c6a 100755
--- a/examples/jeopardy/jeopardy.sh
+++ b/examples/jeopardy/jeopardy.sh
@@ -1,4 +1,4 @@
-#!/bin/bash
+#!/usr/bin/env bash
 set -e
 
 MODEL=./models/ggml-vicuna-13b-1.1-q4_0.bin
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 9654cd53c..711ddc5d1 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_self_clear(context);
+        llama_memory_clear(llama_get_memory(context), false);
 
         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_self_clear(context);
+        llama_memory_clear(llama_get_memory(context), false);
         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_self_clear(context);
+        llama_memory_clear(llama_get_memory(context), false);
 
         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;
@@ -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_self_clear(reinterpret_cast<llama_context *>(context));
+    llama_memory_clear(llama_get_memory(reinterpret_cast<llama_context *>(context)), true);
 }
diff --git a/examples/llama.swiftui/llama.cpp.swift/LibLlama.swift b/examples/llama.swiftui/llama.cpp.swift/LibLlama.swift
index f6e31abc9..dc2bafc88 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_self_clear(context)
+            llama_memory_clear(llama_get_memory(context), false)
 
             let t_pp_start = DispatchTime.now().uptimeNanoseconds / 1000;
 
@@ -223,7 +223,7 @@ actor LlamaContext {
 
             // bench text generation
 
-            llama_kv_self_clear(context)
+            llama_memory_clear(llama_get_memory(context), false)
 
             let t_tg_start = DispatchTime.now().uptimeNanoseconds / 1000;
 
@@ -242,7 +242,7 @@ actor LlamaContext {
 
             let t_tg_end = DispatchTime.now().uptimeNanoseconds / 1000;
 
-            llama_kv_self_clear(context)
+            llama_memory_clear(llama_get_memory(context), false)
 
             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_self_clear(context)
+        llama_memory_clear(llama_get_memory(context), true)
     }
 
     private func tokenize(text: String, add_bos: Bool) -> [llama_token] {
diff --git a/examples/lookahead/lookahead.cpp b/examples/lookahead/lookahead.cpp
index 5f8620973..1e26d8221 100644
--- a/examples/lookahead/lookahead.cpp
+++ b/examples/lookahead/lookahead.cpp
@@ -60,6 +60,8 @@ int main(int argc, char ** argv) {
     llama_model * model = llama_init.model.get();
     llama_context * ctx = llama_init.context.get();
 
+    auto * mem = llama_get_memory(ctx);
+
     const llama_vocab * vocab = llama_model_get_vocab(model);
 
     // Tokenize the prompt
@@ -94,7 +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_self_seq_cp(ctx, 0, s, -1, -1);
+        llama_memory_seq_cp(mem, 0, s, -1, -1);
     }
 
     const auto t_enc_end = ggml_time_us();
@@ -427,17 +429,17 @@ 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_self_seq_rm(ctx, -1, n_past, -1);
+        llama_memory_seq_rm(mem, -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_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);
+            llama_memory_seq_keep(mem, seq_id_best);
+            llama_memory_seq_cp  (mem, seq_id_best, 0, -1, -1);
+            llama_memory_seq_rm  (mem, seq_id_best,    -1, -1);
 
             for (int s = 1; s < W + G + 1; ++s) {
-                llama_kv_self_seq_cp(ctx, 0, s, -1, -1);
+                llama_memory_seq_cp(mem, 0, s, -1, -1);
             }
         }
     }
diff --git a/examples/lookup/lookup.cpp b/examples/lookup/lookup.cpp
index 2ee502939..2bfa26b55 100644
--- a/examples/lookup/lookup.cpp
+++ b/examples/lookup/lookup.cpp
@@ -181,7 +181,7 @@ int main(int argc, char ** argv){
 
         // KV cache management
         // clean the cache of draft tokens that weren't accepted
-        llama_kv_self_seq_rm(ctx, 0, n_past, -1);
+        llama_memory_seq_rm(llama_get_memory(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/parallel/parallel.cpp b/examples/parallel/parallel.cpp
index cd85bea9a..d53e089a4 100644
--- a/examples/parallel/parallel.cpp
+++ b/examples/parallel/parallel.cpp
@@ -194,6 +194,8 @@ int main(int argc, char ** argv) {
     llama_model * model = llama_init.model.get();
     llama_context * ctx = llama_init.context.get();
 
+    auto * mem = llama_get_memory(ctx);
+
     const llama_vocab * vocab = llama_model_get_vocab(model);
 
     // load the prompts from an external file if there are any
@@ -259,7 +261,7 @@ 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_self_seq_cp(ctx, 0, i, -1, -1);
+            llama_memory_seq_cp(mem, 0, i, -1, -1);
         }
 
         LOG_INF("\n");
@@ -286,9 +288,9 @@ 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_self_seq_rm(ctx, i, -1, -1);
+                llama_memory_seq_rm(mem, i, -1, -1);
                 // but keep the system prompt
-                llama_kv_self_seq_cp(ctx, 0, i, -1, -1);
+                llama_memory_seq_cp(mem, 0, i, -1, -1);
             }
 
             LOG_INF("%s: clearing the KV cache\n", __func__);
@@ -447,8 +449,8 @@ 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_self_seq_rm(ctx,    client.id + 1, -1, -1);
-                    llama_kv_self_seq_cp(ctx, 0, client.id + 1, -1, -1);
+                    llama_memory_seq_rm(mem,    client.id + 1, -1, -1);
+                    llama_memory_seq_cp(mem, 0, client.id + 1, -1, -1);
 
                     const auto t_main_end = ggml_time_us();
 
diff --git a/examples/passkey/passkey.cpp b/examples/passkey/passkey.cpp
index 5ac881b45..8a4faa383 100644
--- a/examples/passkey/passkey.cpp
+++ b/examples/passkey/passkey.cpp
@@ -126,6 +126,8 @@ int main(int argc, char ** argv) {
 
     int n_past = 0;
 
+    auto * mem = llama_get_memory(ctx);
+
     // fill the KV cache
     for (int i = 0; i < n_ctx; i += n_batch) {
         if (i > 0 && n_grp > 1) {
@@ -133,10 +135,10 @@ int main(int argc, char ** argv) {
             const int ib = i/n_batch - 1;
             const int bd = n_batch_grp*(n_grp - 1);
 
-            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_memory_seq_add(mem, 0, n_past - n_batch,         n_past,         ib*bd);
+            llama_memory_seq_div(mem, 0, n_past - n_batch + ib*bd, n_past + ib*bd, n_grp);
 
-            n_past = llama_kv_self_seq_pos_max(ctx, 0) + 1;
+            n_past = llama_memory_seq_pos_max(mem, 0) + 1;
         }
 
         common_batch_clear(batch);
@@ -166,10 +168,10 @@ int main(int argc, char ** argv) {
 
         LOG_INF("%s: shifting KV cache with %d\n", __func__, n_discard);
 
-        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_memory_seq_rm (mem, 0, n_keep            , n_keep + n_discard);
+        llama_memory_seq_add(mem, 0, n_keep + n_discard, n_ctx,  -n_discard);
 
-        n_past = llama_kv_self_seq_pos_max(ctx, 0) + 1;
+        n_past = llama_memory_seq_pos_max(mem, 0) + 1;
 
         common_batch_clear(batch);
 
@@ -195,10 +197,10 @@ 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_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_memory_seq_rm (mem, 0, n_keep            , n_keep + n_discard);
+            llama_memory_seq_add(mem, 0, n_keep + n_discard, n_ctx,  -n_discard);
 
-            n_past = llama_kv_self_seq_pos_max(ctx, 0) + 1;
+            n_past = llama_memory_seq_pos_max(mem, 0) + 1;
         }
     }
 
diff --git a/examples/reason-act.sh b/examples/reason-act.sh
index 06d592799..3c801920d 100755
--- a/examples/reason-act.sh
+++ b/examples/reason-act.sh
@@ -1,4 +1,4 @@
-#!/bin/bash
+#!/usr/bin/env bash
 
 cd `dirname $0`
 cd ..
diff --git a/examples/retrieval/retrieval.cpp b/examples/retrieval/retrieval.cpp
index 754da1411..042e12c2b 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<int32_t> & toke
 
 static void batch_process(llama_context * ctx, llama_batch & batch, float * output, int n_seq, int n_embd) {
     // clear previous kv_cache values (irrelevant for embeddings)
-    llama_kv_self_clear(ctx);
+    llama_memory_clear(llama_get_memory(ctx), false);
 
     // run model
     LOG_INF("%s: n_tokens = %d, n_seq = %d\n", __func__, batch.n_tokens, n_seq);
diff --git a/examples/save-load-state/save-load-state.cpp b/examples/save-load-state/save-load-state.cpp
index 760ebbbf0..db79588f1 100644
--- a/examples/save-load-state/save-load-state.cpp
+++ b/examples/save-load-state/save-load-state.cpp
@@ -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_self_clear(ctx3);
+        llama_memory_clear(llama_get_memory(ctx3), true);
         fprintf(stderr, "%s : kv cache cleared\n", __func__);
 
         // restore kv into seq 1
diff --git a/examples/server-llama2-13B.sh b/examples/server-llama2-13B.sh
index 4ce79b7fa..fd5a57588 100755
--- a/examples/server-llama2-13B.sh
+++ b/examples/server-llama2-13B.sh
@@ -1,4 +1,4 @@
-#!/bin/bash
+#!/usr/bin/env bash
 
 set -e
 
diff --git a/examples/simple-chat/simple-chat.cpp b/examples/simple-chat/simple-chat.cpp
index 6608d4bea..57195df33 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_kv_self_seq_pos_max(ctx, 0) == 0;
+        const bool is_first = llama_memory_seq_pos_max(llama_get_memory(ctx), 0) == -1;
 
         // tokenize the prompt
         const int n_prompt_tokens = -llama_tokenize(vocab, prompt.c_str(), prompt.size(), NULL, 0, is_first, true);
@@ -113,15 +113,16 @@ 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_kv_self_seq_pos_max(ctx, 0);
+            int n_ctx_used = llama_memory_seq_pos_max(llama_get_memory(ctx), 0) + 1;
             if (n_ctx_used + batch.n_tokens > n_ctx) {
                 printf("\033[0m\n");
                 fprintf(stderr, "context size exceeded\n");
                 exit(0);
             }
 
-            if (llama_decode(ctx, batch)) {
-                GGML_ABORT("failed to decode\n");
+            int ret = llama_decode(ctx, batch);
+            if (ret != 0) {
+                GGML_ABORT("failed to decode, ret = %d\n", ret);
             }
 
             // sample the next token
diff --git a/examples/speculative-simple/speculative-simple.cpp b/examples/speculative-simple/speculative-simple.cpp
index 0783ed4a4..99196c9d0 100644
--- a/examples/speculative-simple/speculative-simple.cpp
+++ b/examples/speculative-simple/speculative-simple.cpp
@@ -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_self_seq_rm(ctx_tgt, 0, n_past, -1);
+            llama_memory_seq_rm(llama_get_memory(ctx_tgt), 0, n_past, -1);
         }
 
         if ((params.n_predict >= 0 && n_predict > params.n_predict) || has_eos) {
diff --git a/examples/speculative/speculative.cpp b/examples/speculative/speculative.cpp
index 561c30883..0adffdb00 100644
--- a/examples/speculative/speculative.cpp
+++ b/examples/speculative/speculative.cpp
@@ -142,6 +142,8 @@ int main(int argc, char ** argv) {
         }
     }
 
+    auto * mem_tgt = llama_get_memory(ctx_tgt);
+    auto * mem_dft = llama_get_memory(ctx_dft);
 
     // Tokenize the prompt
     std::vector<llama_token> inp;
@@ -420,14 +422,14 @@ 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_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_memory_seq_keep(mem_dft, s_keep);
+                llama_memory_seq_cp  (mem_dft, s_keep, 0, -1, -1);
+                llama_memory_seq_keep(mem_dft, 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);
+                llama_memory_seq_rm  (mem_tgt, s_keep, n_past_tgt, -1);
+                llama_memory_seq_keep(mem_tgt, s_keep);
+                llama_memory_seq_cp  (mem_tgt, s_keep, 0, -1, -1);
+                llama_memory_seq_keep(mem_tgt, 0);
             }
 
             for (int s = 0; s < n_seq_dft; ++s) {
@@ -444,7 +446,7 @@ 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_self_seq_rm(ctx_dft, 0, n_past_dft, -1);
+            llama_memory_seq_rm(mem_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 +505,8 @@ 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_self_seq_rm(ctx_dft,    n_seq_cur, -1, -1);
-                        llama_kv_self_seq_cp(ctx_dft, s, n_seq_cur, -1, -1);
+                        llama_memory_seq_rm(mem_dft,    n_seq_cur, -1, -1);
+                        llama_memory_seq_cp(mem_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 +587,9 @@ int main(int argc, char ** argv) {
 
         // evaluate the target model on the drafted tokens
         {
-            llama_kv_self_seq_keep(ctx_tgt, 0);
+            llama_memory_seq_keep(mem_tgt, 0);
             for (int s = 1; s < n_seq_dft; ++s) {
-                llama_kv_self_seq_cp(ctx_tgt, 0, s, -1, -1);
+                llama_memory_seq_cp(mem_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/build.sh b/examples/sycl/build.sh
index e72b2e261..1993520eb 100755
--- a/examples/sycl/build.sh
+++ b/examples/sycl/build.sh
@@ -1,4 +1,4 @@
-
+#!/usr/bin/env bash
 #  MIT license
 #  Copyright (C) 2024 Intel Corporation
 #  SPDX-License-Identifier: MIT
diff --git a/examples/sycl/run-llama2.sh b/examples/sycl/run-llama2.sh
index 40ce8f5b2..37195008d 100755
--- a/examples/sycl/run-llama2.sh
+++ b/examples/sycl/run-llama2.sh
@@ -1,4 +1,4 @@
-#!/bin/bash
+#!/usr/bin/env bash
 
 #  MIT license
 #  Copyright (C) 2024 Intel Corporation
diff --git a/examples/sycl/run-llama3.sh b/examples/sycl/run-llama3.sh
index 933d1b98b..8e21b017f 100755
--- a/examples/sycl/run-llama3.sh
+++ b/examples/sycl/run-llama3.sh
@@ -1,4 +1,4 @@
-#!/bin/bash
+#!/usr/bin/env bash
 
 #  MIT license
 #  Copyright (C) 2025 Intel Corporation
diff --git a/examples/ts-type-to-grammar.sh b/examples/ts-type-to-grammar.sh
index 9abba2a3d..966050407 100755
--- a/examples/ts-type-to-grammar.sh
+++ b/examples/ts-type-to-grammar.sh
@@ -1,4 +1,4 @@
-#!/bin/bash
+#!/usr/bin/env bash
 #
 # ./examples/ts-type-to-grammar.sh "{a:string,b:string,c?:string}"
 # python examples/json_schema_to_grammar.py https://json.schemastore.org/tsconfig.json
diff --git a/ggml/CMakeLists.txt b/ggml/CMakeLists.txt
index 3d01184a2..eaba9c704 100644
--- a/ggml/CMakeLists.txt
+++ b/ggml/CMakeLists.txt
@@ -105,7 +105,7 @@ 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_REPACK       "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_SSE42            "ggml: enable SSE 4.2"          ${INS_ENB})
 option(GGML_AVX              "ggml: enable AVX"              ${INS_ENB})
@@ -131,13 +131,14 @@ option(GGML_RVV              "ggml: enable rvv"              ON)
 option(GGML_RV_ZFH           "ggml: enable riscv zfh"        OFF)
 option(GGML_XTHEADVECTOR     "ggml: enable xtheadvector"     OFF)
 option(GGML_VXE              "ggml: enable vxe"              ON)
+option(GGML_NNPA             "ggml: enable nnpa"             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_POWERPC_CPUTYPE "" CACHE STRING "ggml: CPU type for PowerPC")
 
 
-if (WIN32)
+if (MINGW)
     set(GGML_WIN_VER "0x602" CACHE STRING   "ggml: Windows version")
 endif()
 
@@ -172,6 +173,7 @@ option(GGML_HIP                             "ggml: use HIP"
 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_FORCE_ROCWMMA_FATTN_GFX12   "ggml: enable rocWMMA FlashAttention on GFX12"    OFF)
 option(GGML_VULKAN                          "ggml: use Vulkan"                                OFF)
 option(GGML_VULKAN_CHECK_RESULTS            "ggml: run Vulkan op checks"                      OFF)
 option(GGML_VULKAN_DEBUG                    "ggml: enable Vulkan debug output"                OFF)
@@ -179,7 +181,6 @@ option(GGML_VULKAN_MEMORY_DEBUG             "ggml: enable Vulkan memory debug ou
 option(GGML_VULKAN_SHADER_DEBUG_INFO        "ggml: enable Vulkan shader debug info"           OFF)
 option(GGML_VULKAN_VALIDATE                 "ggml: enable Vulkan validation"                  OFF)
 option(GGML_VULKAN_RUN_TESTS                "ggml: run Vulkan tests"                          OFF)
-option(GGML_KOMPUTE                         "ggml: use Kompute"                               OFF)
 option(GGML_METAL                           "ggml: use Metal"                                 ${GGML_METAL_DEFAULT})
 option(GGML_METAL_USE_BF16                  "ggml: use bfloat if available"                   OFF)
 option(GGML_METAL_NDEBUG                    "ggml: disable Metal debugging"                   OFF)
@@ -264,7 +265,6 @@ set(GGML_PUBLIC_HEADERS
     include/ggml-cann.h
     include/ggml-cpp.h
     include/ggml-cuda.h
-    include/ggml-kompute.h
     include/ggml-opt.h
     include/ggml-metal.h
     include/ggml-rpc.h
@@ -358,6 +358,13 @@ write_basic_package_version_file(
     VERSION ${GGML_INSTALL_VERSION}
     COMPATIBILITY SameMajorVersion)
 
+target_compile_definitions(ggml-base PRIVATE
+    GGML_VERSION="${GGML_INSTALL_VERSION}"
+    GGML_COMMIT="${GGML_BUILD_COMMIT}"
+)
+message(STATUS "ggml version: ${GGML_INSTALL_VERSION}")
+message(STATUS "ggml commit:  ${GGML_BUILD_COMMIT}")
+
 install(FILES ${CMAKE_CURRENT_BINARY_DIR}/ggml-config.cmake
               ${CMAKE_CURRENT_BINARY_DIR}/ggml-version.cmake
         DESTINATION ${CMAKE_INSTALL_LIBDIR}/cmake/ggml)
@@ -367,6 +374,8 @@ if (MSVC)
         /wd4005  # Macro redefinition
         /wd4244  # Conversion from one type to another type, possible loss of data
         /wd4267  # Conversion from 'size_t' to a smaller type, possible loss of data
+        /wd4305  # Conversion from 'type1' to 'type2', possible loss of data
+        /wd4566  # Conversion from 'char' to 'wchar_t', possible loss of data
         /wd4996  # Disable POSIX deprecation warnings
         /wd4702  # Unreachable code warnings
     )
@@ -386,4 +395,46 @@ if (MSVC)
     disable_msvc_warnings(ggml-cpu-skylakex)
     disable_msvc_warnings(ggml-cpu-icelake)
     disable_msvc_warnings(ggml-cpu-alderlake)
+
+    if (GGML_BUILD_EXAMPLES)
+        disable_msvc_warnings(common-ggml)
+        disable_msvc_warnings(common)
+
+        disable_msvc_warnings(mnist-common)
+        disable_msvc_warnings(mnist-eval)
+        disable_msvc_warnings(mnist-train)
+
+        disable_msvc_warnings(gpt-2-ctx)
+        disable_msvc_warnings(gpt-2-alloc)
+        disable_msvc_warnings(gpt-2-backend)
+        disable_msvc_warnings(gpt-2-sched)
+        disable_msvc_warnings(gpt-2-quantize)
+        disable_msvc_warnings(gpt-2-batched)
+
+        disable_msvc_warnings(gpt-j)
+        disable_msvc_warnings(gpt-j-quantize)
+
+        disable_msvc_warnings(magika)
+        disable_msvc_warnings(yolov3-tiny)
+        disable_msvc_warnings(sam)
+
+        disable_msvc_warnings(simple-ctx)
+        disable_msvc_warnings(simple-backend)
+    endif()
+
+    if (GGML_BUILD_TESTS)
+        disable_msvc_warnings(test-mul-mat)
+        disable_msvc_warnings(test-arange)
+        disable_msvc_warnings(test-backend-ops)
+        disable_msvc_warnings(test-cont)
+        disable_msvc_warnings(test-conv-transpose)
+        disable_msvc_warnings(test-conv-transpose-1d)
+        disable_msvc_warnings(test-conv1d)
+        disable_msvc_warnings(test-conv2d)
+        disable_msvc_warnings(test-conv2d-dw)
+        disable_msvc_warnings(test-customop)
+        disable_msvc_warnings(test-dup)
+        disable_msvc_warnings(test-opt)
+        disable_msvc_warnings(test-pool)
+    endif ()
 endif()
diff --git a/ggml/cmake/common.cmake b/ggml/cmake/common.cmake
index bb1ec9b37..cb6638833 100644
--- a/ggml/cmake/common.cmake
+++ b/ggml/cmake/common.cmake
@@ -36,8 +36,7 @@ function(ggml_get_system_arch)
             (NOT CMAKE_OSX_ARCHITECTURES AND NOT CMAKE_GENERATOR_PLATFORM_LWR AND
             CMAKE_SYSTEM_PROCESSOR MATCHES "^(x86_64|i686|AMD64|amd64)$"))
         set(GGML_SYSTEM_ARCH "x86" PARENT_SCOPE)
-    elseif ("${CMAKE_SYSTEM_PROCESSOR} " STREQUAL "ppc64le " OR
-            "${CMAKE_SYSTEM_PROCESSOR} " STREQUAL "powerpc ")
+    elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc|power")
         set(GGML_SYSTEM_ARCH "PowerPC" PARENT_SCOPE)
     elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "loongarch64")
         set(GGML_SYSTEM_ARCH "loongarch64"  PARENT_SCOPE)
diff --git a/ggml/include/ggml-backend.h b/ggml/include/ggml-backend.h
index 778927f68..a2977ea2e 100644
--- a/ggml/include/ggml-backend.h
+++ b/ggml/include/ggml-backend.h
@@ -339,7 +339,7 @@ extern "C" {
     typedef bool (*ggml_backend_eval_callback)(int node_index, struct ggml_tensor * t1, struct ggml_tensor * t2, void * user_data);
 
     // Compare the output of two backends
-    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);
+    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, struct ggml_tensor * test_node);
 
     // Tensor initialization
     GGML_API enum ggml_status ggml_backend_tensor_alloc(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, void * addr);
diff --git a/ggml/include/ggml-cpu.h b/ggml/include/ggml-cpu.h
index de77a875e..be40b1009 100644
--- a/ggml/include/ggml-cpu.h
+++ b/ggml/include/ggml-cpu.h
@@ -101,6 +101,7 @@ extern "C" {
     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_nnpa       (void);
     GGML_BACKEND_API int ggml_cpu_has_wasm_simd  (void);
     GGML_BACKEND_API int ggml_cpu_has_llamafile  (void);
 
@@ -133,6 +134,7 @@ extern "C" {
 
     GGML_BACKEND_API ggml_backend_reg_t ggml_backend_cpu_reg(void);
 
+    GGML_BACKEND_API void ggml_cpu_fp32_to_fp32(const float *,       float *, int64_t);
     GGML_BACKEND_API void ggml_cpu_fp32_to_fp16(const float *, ggml_fp16_t *, int64_t);
     GGML_BACKEND_API void ggml_cpu_fp16_to_fp32(const ggml_fp16_t *, float *, int64_t);
     GGML_BACKEND_API void ggml_cpu_fp32_to_bf16(const float *, ggml_bf16_t *, int64_t);
diff --git a/ggml/include/ggml-kompute.h b/ggml/include/ggml-kompute.h
deleted file mode 100644
index 154aa56a7..000000000
--- a/ggml/include/ggml-kompute.h
+++ /dev/null
@@ -1,50 +0,0 @@
-#pragma once
-
-#include "ggml.h"
-#include "ggml-backend.h"
-
-#include <stdbool.h>
-#include <stddef.h>
-#include <stdint.h>
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-#define GGML_KOMPUTE_MAX_DEVICES 16
-
-struct ggml_vk_device {
-    int index;
-    int type; // same as VkPhysicalDeviceType
-    size_t heapSize;
-    const char * name;
-    const char * vendor;
-    int subgroupSize;
-    uint64_t bufferAlignment;
-    uint64_t maxAlloc;
-};
-
-struct ggml_vk_device * ggml_vk_available_devices(size_t memoryRequired, size_t * count);
-bool ggml_vk_get_device(struct ggml_vk_device * device, size_t memoryRequired, const char * name);
-bool ggml_vk_has_vulkan(void);
-bool ggml_vk_has_device(void);
-struct ggml_vk_device ggml_vk_current_device(void);
-
-//
-// backend API
-//
-
-// forward declaration
-typedef struct ggml_backend * ggml_backend_t;
-
-GGML_BACKEND_API ggml_backend_t ggml_backend_kompute_init(int device);
-
-GGML_BACKEND_API bool ggml_backend_is_kompute(ggml_backend_t backend);
-
-GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_kompute_buffer_type(int device);
-
-GGML_BACKEND_API ggml_backend_reg_t ggml_backend_kompute_reg(void);
-
-#ifdef __cplusplus
-}
-#endif
diff --git a/ggml/include/ggml.h b/ggml/include/ggml.h
index 1a57f1cd7..949eac9a5 100644
--- a/ggml/include/ggml.h
+++ b/ggml/include/ggml.h
@@ -314,6 +314,13 @@
 extern "C" {
 #endif
 
+    // Function type used in fatal error callbacks
+    typedef void (*ggml_abort_callback_t)(const char * error_message);
+
+    // Set the abort callback (passing null will restore original abort functionality: printing a message to stdout)
+    // Returns the old callback for chaining
+    GGML_API ggml_abort_callback_t ggml_set_abort_callback(ggml_abort_callback_t callback);
+
     GGML_NORETURN GGML_ATTRIBUTE_FORMAT(3, 4)
     GGML_API void ggml_abort(const char * file, int line, const char * fmt, ...);
 
@@ -470,6 +477,7 @@ extern "C" {
         GGML_OP_TRANSPOSE,
         GGML_OP_GET_ROWS,
         GGML_OP_GET_ROWS_BACK,
+        GGML_OP_SET_ROWS,
         GGML_OP_DIAG,
         GGML_OP_DIAG_MASK_INF,
         GGML_OP_DIAG_MASK_ZERO,
@@ -481,6 +489,7 @@ extern "C" {
         GGML_OP_CONV_TRANSPOSE_1D,
         GGML_OP_IM2COL,
         GGML_OP_IM2COL_BACK,
+        GGML_OP_CONV_2D,
         GGML_OP_CONV_2D_DW,
         GGML_OP_CONV_TRANSPOSE_2D,
         GGML_OP_POOL_1D,
@@ -489,6 +498,7 @@ extern "C" {
         GGML_OP_UPSCALE, // nearest interpolate
         GGML_OP_PAD,
         GGML_OP_PAD_REFLECT_1D,
+        GGML_OP_ROLL,
         GGML_OP_ARANGE,
         GGML_OP_TIMESTEP_EMBEDDING,
         GGML_OP_ARGSORT,
@@ -518,6 +528,8 @@ extern "C" {
         GGML_OP_CROSS_ENTROPY_LOSS_BACK,
         GGML_OP_OPT_STEP_ADAMW,
 
+        GGML_OP_GLU,
+
         GGML_OP_COUNT,
     };
 
@@ -541,6 +553,16 @@ extern "C" {
         GGML_UNARY_OP_COUNT,
     };
 
+    enum ggml_glu_op {
+        GGML_GLU_OP_REGLU,
+        GGML_GLU_OP_GEGLU,
+        GGML_GLU_OP_SWIGLU,
+        GGML_GLU_OP_GEGLU_ERF,
+        GGML_GLU_OP_GEGLU_QUICK,
+
+        GGML_GLU_OP_COUNT,
+    };
+
     enum ggml_object_type {
         GGML_OBJECT_TYPE_TENSOR,
         GGML_OBJECT_TYPE_GRAPH,
@@ -626,6 +648,9 @@ extern "C" {
 
     // misc
 
+    GGML_API const char * ggml_version(void);
+    GGML_API const char * ggml_commit(void);
+
     GGML_API void    ggml_time_init(void); // call this once at the beginning of the program
     GGML_API int64_t ggml_time_ms(void);
     GGML_API int64_t ggml_time_us(void);
@@ -656,6 +681,7 @@ extern "C" {
     GGML_API const char * ggml_op_symbol(enum ggml_op   op);
 
     GGML_API const char * ggml_unary_op_name(enum ggml_unary_op op);
+    GGML_API const char * ggml_glu_op_name(enum ggml_glu_op op);
     GGML_API const char * ggml_op_desc(const struct ggml_tensor * t); // unary or op name
 
     GGML_API size_t  ggml_element_size(const struct ggml_tensor * tensor);
@@ -686,6 +712,9 @@ extern "C" {
     // true for tensor that is stored in memory as CxWxHxN and has been permuted to WxHxCxN
     GGML_API bool ggml_is_contiguous_channels(const struct ggml_tensor * tensor);
 
+    // true if the elements in dimension 0 are contiguous, or there is just 1 block of elements
+    GGML_API bool ggml_is_contiguous_rows(const struct ggml_tensor * tensor);
+
     GGML_API bool ggml_are_same_shape (const struct ggml_tensor * t0, const struct ggml_tensor * t1);
     GGML_API bool ggml_are_same_stride(const struct ggml_tensor * t0, const struct ggml_tensor * t1);
 
@@ -757,6 +786,7 @@ extern "C" {
     GGML_API void ggml_unravel_index(const struct ggml_tensor * tensor, int64_t i, int64_t * i0, int64_t * i1, int64_t * i2, int64_t * i3);
 
     GGML_API enum ggml_unary_op ggml_get_unary_op(const struct ggml_tensor * tensor);
+    GGML_API enum ggml_glu_op ggml_get_glu_op(const struct ggml_tensor * tensor);
 
     GGML_API void *  ggml_get_data    (const struct ggml_tensor * tensor);
     GGML_API float * ggml_get_data_f32(const struct ggml_tensor * tensor);
@@ -1085,6 +1115,89 @@ extern "C" {
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
+    // gated linear unit ops
+    // A: n columns, r rows,
+    // result is n / 2 columns, r rows,
+    // expects gate in second half of row, unless swapped is true
+    GGML_API struct ggml_tensor * ggml_glu(
+            struct ggml_context * ctx,
+             struct ggml_tensor * a,
+             enum ggml_glu_op     op,
+             bool                 swapped);
+
+    GGML_API struct ggml_tensor * ggml_reglu(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_reglu_swapped(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_geglu(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_geglu_swapped(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_swiglu(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_swiglu_swapped(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_geglu_erf(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_geglu_erf_swapped(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_geglu_quick(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_geglu_quick_swapped(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    // A: n columns, r rows,
+    // B: n columns, r rows,
+    GGML_API struct ggml_tensor * ggml_glu_split(
+            struct ggml_context * ctx,
+             struct ggml_tensor * a,
+             struct ggml_tensor * b,
+             enum ggml_glu_op     op);
+
+    GGML_API struct ggml_tensor * ggml_reglu_split(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b);
+
+    GGML_API struct ggml_tensor * ggml_geglu_split(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b);
+
+    GGML_API struct ggml_tensor * ggml_swiglu_split(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b);
+
+    GGML_API struct ggml_tensor * ggml_geglu_erf_split(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b);
+
+    GGML_API struct ggml_tensor * ggml_geglu_quick_split(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b);
+
     // normalize along rows
     GGML_API struct ggml_tensor * ggml_norm(
             struct ggml_context * ctx,
@@ -1374,6 +1487,23 @@ extern "C" {
             struct ggml_tensor  * b,  // row indices
             struct ggml_tensor  * c); // data for ggml_get_rows, only used for its shape
 
+    // a TD  [n_embd, ne1,    ne2,    ne3]
+    // b TS  [n_embd, n_rows, ne02,   ne03] | ne02 == ne2, ne03 == ne3
+    // c I64 [n_rows, ne11,   ne12,   1]    | c[i] in [0, ne1)
+    //
+    // undefined behavior if destination rows overlap
+    //
+    // broadcast:
+    //   ne2 % ne11 == 0
+    //   ne3 % ne12 == 0
+    //
+    // return view(a)
+    GGML_API struct ggml_tensor * ggml_set_rows(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,  // destination
+            struct ggml_tensor  * b,  // source
+            struct ggml_tensor  * c); // row indices
+
     GGML_API struct ggml_tensor * ggml_diag(
         struct ggml_context     * ctx,
         struct ggml_tensor      * a);
@@ -1411,8 +1541,14 @@ extern "C" {
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
+    // a    [ne0, ne01, ne02, ne03]
+    // mask [ne0, ne11, ne12, ne13] | ne11 >= ne01, F16 or F32, optional
+    //
+    // broadcast:
+    //   ne02 % ne12 == 0
+    //   ne03 % ne13 == 0
+    //
     // fused soft_max(a*scale + mask*(ALiBi slope))
-    // mask is optional
     // max_bias = 0.0f for no ALiBi
     GGML_API struct ggml_tensor * ggml_soft_max_ext(
             struct ggml_context * ctx,
@@ -1722,6 +1858,17 @@ extern "C" {
             struct ggml_tensor  * b,
             int                   stride);
 
+    GGML_API struct ggml_tensor * ggml_conv_2d_direct(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,   // convolution kernel [KW, KH, IC, OC]
+            struct ggml_tensor  * b,   // input data [W, H, C, N]
+            int                   s0,  // stride dimension 0
+            int                   s1,  // stride dimension 1
+            int                   p0,  // padding dimension 0
+            int                   p1,  // padding dimension 1
+            int                   d0,  // dilation dimension 0
+            int                   d1); // dilation dimension 1
+
     enum ggml_op_pool {
         GGML_OP_POOL_MAX,
         GGML_OP_POOL_AVG,
@@ -1764,6 +1911,12 @@ extern "C" {
     enum ggml_scale_mode {
         GGML_SCALE_MODE_NEAREST  = 0,
         GGML_SCALE_MODE_BILINEAR = 1,
+
+        GGML_SCALE_MODE_COUNT
+    };
+
+    enum ggml_scale_flag {
+        GGML_SCALE_FLAG_ALIGN_CORNERS = (1 << 8)
     };
 
     // interpolate
@@ -1776,14 +1929,26 @@ extern "C" {
 
     // interpolate
     // interpolate scale to specified dimensions
-    GGML_API struct ggml_tensor * ggml_upscale_ext(
+    GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_upscale_ext(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
             int                   ne0,
             int                   ne1,
             int                   ne2,
             int                   ne3,
-            enum ggml_scale_mode  mode);
+            enum ggml_scale_mode  mode),
+        "use ggml_interpolate instead");
+
+    // Up- or downsamples the input to the specified size.
+    // 2D scale modes (eg. bilinear) are applied to the first two dimensions.
+    GGML_API struct ggml_tensor * ggml_interpolate(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int64_t               ne0,
+            int64_t               ne1,
+            int64_t               ne2,
+            int64_t               ne3,
+            uint32_t              mode); // ggml_scale_mode [ | ggml_scale_flag...]
 
     // pad each dimension with zeros: [x, ..., x] -> [x, ..., x, 0, ..., 0]
     GGML_API struct ggml_tensor * ggml_pad(
@@ -1801,6 +1966,17 @@ extern "C" {
             int                   p0,
             int                   p1);
 
+    // Move tensor elements by an offset given for each dimension. Elements that
+    // are shifted beyond the last position are wrapped around to the beginning.
+    GGML_API struct ggml_tensor * ggml_roll(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int                   shift0,
+            int                   shift1,
+            int                   shift2,
+            int                   shift3);
+
+
     // Ref: https://github.com/CompVis/stable-diffusion/blob/main/ldm/modules/diffusionmodules/util.py#L151
     // timesteps: [N,]
     // return: [N, dim]
@@ -1835,11 +2011,17 @@ extern "C" {
 
 #define GGML_KQ_MASK_PAD 64
 
-    // 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 !!
+    // q:    [n_embd_k, n_batch,     n_head,    ne3 ]
+    // k:    [n_embd_k, n_kv,        n_head_kv, ne3 ]
+    // v:    [n_embd_v, n_kv,        n_head_kv, ne3 ] !! not transposed !!
+    // mask: [n_kv,     n_batch_pad, ne32,      ne33] !! n_batch_pad = GGML_PAD(n_batch, GGML_KQ_MASK_PAD) !!
+    // res:  [n_embd_v, n_head,      n_batch,   ne3 ] !! permuted !!
+    //
+    // broadcast:
+    //   n_head % n_head_kv == 0
+    //   n_head % ne32      == 0
+    //   ne3    % ne33      == 0
+    //
     GGML_API struct ggml_tensor * ggml_flash_attn_ext(
             struct ggml_context * ctx,
             struct ggml_tensor  * q,
@@ -1878,7 +2060,8 @@ extern "C" {
             struct ggml_tensor  * dt,
             struct ggml_tensor  * A,
             struct ggml_tensor  * B,
-            struct ggml_tensor  * C);
+            struct ggml_tensor  * C,
+            struct ggml_tensor  * ids);
 
     // partition into non-overlapping windows with padding if needed
     // example:
diff --git a/ggml/src/CMakeLists.txt b/ggml/src/CMakeLists.txt
index 76b24bd9d..8760c2d35 100644
--- a/ggml/src/CMakeLists.txt
+++ b/ggml/src/CMakeLists.txt
@@ -125,7 +125,6 @@ if (NOT MSVC)
 endif()
 
 if (MINGW)
-    # Target Windows 8 for PrefetchVirtualMemory
     add_compile_definitions(_WIN32_WINNT=${GGML_WIN_VER})
 endif()
 
@@ -213,6 +212,7 @@ endif()
 
 add_library(ggml
             ggml-backend-reg.cpp)
+add_library(ggml::ggml ALIAS ggml)
 
 target_link_libraries(ggml PUBLIC ggml-base)
 
@@ -270,17 +270,27 @@ endfunction()
 function(ggml_add_cpu_backend_variant tag_name)
     set(GGML_CPU_TAG_NAME ${tag_name})
     # other: OPENMP LLAMAFILE CPU_HBM
-    foreach (feat NATIVE
-                  SSE42
-                  AVX AVX2 BMI2 AVX_VNNI FMA F16C
-                  AVX512 AVX512_VBMI AVX512_VNNI AVX512_BF16
-                  AMX_TILE AMX_INT8 AMX_BF16)
-        set(GGML_${feat} OFF)
-    endforeach()
+    if (GGML_SYSTEM_ARCH STREQUAL "x86")
+        foreach (feat NATIVE
+                      SSE42
+                      AVX AVX2 BMI2 AVX_VNNI FMA F16C
+                      AVX512 AVX512_VBMI AVX512_VNNI AVX512_BF16
+                      AMX_TILE AMX_INT8 AMX_BF16)
+            set(GGML_${feat} OFF)
+        endforeach()
 
-    foreach (feat ${ARGN})
-        set(GGML_${feat} ON)
-    endforeach()
+        foreach (feat ${ARGN})
+            set(GGML_${feat} ON)
+        endforeach()
+    elseif (GGML_SYSTEM_ARCH STREQUAL "ARM")
+        foreach (feat ${ARGN})
+            set(GGML_INTERNAL_${feat} ON)
+        endforeach()
+    elseif (GGML_SYSTEM_ARCH STREQUAL "PowerPC")
+        foreach (feat ${ARGN})
+            set(GGML_INTERNAL_${feat} ON)
+        endforeach()
+    endif()
 
     ggml_add_cpu_backend_variant_impl(${tag_name})
 endfunction()
@@ -290,6 +300,8 @@ ggml_add_backend(CPU)
 if (GGML_CPU_ALL_VARIANTS)
     if (NOT GGML_BACKEND_DL)
         message(FATAL_ERROR "GGML_CPU_ALL_VARIANTS requires GGML_BACKEND_DL")
+    elseif (GGML_CPU_ARM_ARCH)
+        message(FATAL_ERROR "Cannot use both GGML_CPU_ARM_ARCH and GGML_CPU_ALL_VARIANTS")
     endif()
     if (GGML_SYSTEM_ARCH STREQUAL "x86")
         ggml_add_cpu_backend_variant(x64)
@@ -303,8 +315,47 @@ if (GGML_CPU_ALL_VARIANTS)
             # MSVC doesn't support AMX
             ggml_add_cpu_backend_variant(sapphirerapids SSE42 AVX F16C AVX2 BMI2 FMA AVX512 AVX512_VBMI AVX512_VNNI AVX512_BF16 AMX_TILE AMX_INT8)
         endif()
+    elseif(GGML_SYSTEM_ARCH STREQUAL "ARM")
+        if (CMAKE_SYSTEM_NAME MATCHES "Linux")
+            # Many of these features are optional so we build versions with popular
+            # combinations and name the backends based on the version they were
+            # first released with
+            ggml_add_cpu_backend_variant(armv8.0_1)
+            ggml_add_cpu_backend_variant(armv8.2_1    DOTPROD)
+            ggml_add_cpu_backend_variant(armv8.2_2    DOTPROD FP16_VECTOR_ARITHMETIC)
+            ggml_add_cpu_backend_variant(armv8.2_3    DOTPROD FP16_VECTOR_ARITHMETIC SVE)
+            ggml_add_cpu_backend_variant(armv8.6_1    DOTPROD FP16_VECTOR_ARITHMETIC SVE MATMUL_INT8)
+            ggml_add_cpu_backend_variant(armv8.6_2    DOTPROD FP16_VECTOR_ARITHMETIC SVE MATMUL_INT8 SVE2)
+            ggml_add_cpu_backend_variant(armv9.2_1    DOTPROD FP16_VECTOR_ARITHMETIC SVE MATMUL_INT8 SME)
+            ggml_add_cpu_backend_variant(armv9.2_2    DOTPROD FP16_VECTOR_ARITHMETIC SVE MATMUL_INT8 SVE2 SME)
+        elseif (CMAKE_SYSTEM_NAME MATCHES "Android")
+            # Android-specific backends with SoC-compatible feature sets
+            ggml_add_cpu_backend_variant(android_armv8.0_1)
+            ggml_add_cpu_backend_variant(android_armv8.2_1    DOTPROD)
+            ggml_add_cpu_backend_variant(android_armv8.2_2    DOTPROD FP16_VECTOR_ARITHMETIC)
+            ggml_add_cpu_backend_variant(android_armv8.6_1    DOTPROD FP16_VECTOR_ARITHMETIC MATMUL_INT8)
+        elseif (APPLE)
+            ggml_add_cpu_backend_variant(apple_m1             DOTPROD)
+            ggml_add_cpu_backend_variant(apple_m2_m3          DOTPROD MATMUL_INT8)
+            ggml_add_cpu_backend_variant(apple_m4             DOTPROD MATMUL_INT8 NOSVE SME)
+        else()
+            message(FATAL_ERROR "Unsupported ARM target OS: ${CMAKE_SYSTEM_NAME}")
+        endif()
+    elseif (GGML_SYSTEM_ARCH STREQUAL "PowerPC")
+        if (CMAKE_SYSTEM_NAME MATCHES "Linux")
+            ggml_add_cpu_backend_variant(power0)
+            ggml_add_cpu_backend_variant(power7_1       POWER7)
+            ggml_add_cpu_backend_variant(power7_2       POWER7  VSX)
+            ggml_add_cpu_backend_variant(power8_1       POWER8)
+            ggml_add_cpu_backend_variant(power8_2       POWER8  VSX)
+            ggml_add_cpu_backend_variant(power9         POWER9  VSX)
+            ggml_add_cpu_backend_variant(power10        POWER10 VSX)
+            ggml_add_cpu_backend_variant(power11        POWER11 VSX)
+        else()
+            message(FATAL_ERROR "Unsupported PowerPC target OS: ${CMAKE_SYSTEM_NAME}")
+        endif()
     else()
-        message(FATAL_ERROR "GGML_CPU_ALL_VARIANTS not yet supported on ${GGML_SYSTEM_ARCH}")
+        message(FATAL_ERROR "GGML_CPU_ALL_VARIANTS not yet supported with ${GGML_SYSTEM_ARCH} on ${CMAKE_SYSTEM_NAME}")
     endif()
 elseif (GGML_CPU)
     ggml_add_cpu_backend_variant_impl("")
@@ -314,7 +365,6 @@ ggml_add_backend(BLAS)
 ggml_add_backend(CANN)
 ggml_add_backend(CUDA)
 ggml_add_backend(HIP)
-ggml_add_backend(Kompute)
 ggml_add_backend(METAL)
 ggml_add_backend(MUSA)
 ggml_add_backend(RPC)
diff --git a/ggml/src/ggml-backend-reg.cpp b/ggml/src/ggml-backend-reg.cpp
index 405d8e315..042ea77ac 100644
--- a/ggml/src/ggml-backend-reg.cpp
+++ b/ggml/src/ggml-backend-reg.cpp
@@ -61,14 +61,13 @@
 #include "ggml-cann.h"
 #endif
 
-#ifdef GGML_USE_KOMPUTE
-#include "ggml-kompute.h"
-#endif
-
 // disable C++17 deprecation warning for std::codecvt_utf8
 #if defined(__clang__)
 #    pragma clang diagnostic push
 #    pragma clang diagnostic ignored "-Wdeprecated-declarations"
+#elif defined(__GNUC__)
+#    pragma GCC diagnostic push
+#    pragma GCC diagnostic ignored "-Wdeprecated-declarations"
 #endif
 
 namespace fs = std::filesystem;
@@ -91,6 +90,8 @@ static std::string path_str(const fs::path & path) {
 
 #if defined(__clang__)
 #    pragma clang diagnostic pop
+#elif defined(__GNUC__)
+#    pragma GCC diagnostic pop
 #endif
 
 #ifdef _WIN32
@@ -184,9 +185,6 @@ struct ggml_backend_registry {
 #ifdef GGML_USE_RPC
         register_backend(ggml_backend_rpc_reg());
 #endif
-#ifdef GGML_USE_KOMPUTE
-        register_backend(ggml_backend_kompute_reg());
-#endif
 #ifdef GGML_USE_CPU
         register_backend(ggml_backend_cpu_reg());
 #endif
@@ -570,7 +568,6 @@ void ggml_backend_load_all_from_path(const char * dir_path) {
     ggml_backend_load_best("cann", silent, dir_path);
     ggml_backend_load_best("cuda", silent, dir_path);
     ggml_backend_load_best("hip", silent, dir_path);
-    ggml_backend_load_best("kompute", silent, dir_path);
     ggml_backend_load_best("metal", silent, dir_path);
     ggml_backend_load_best("rpc", silent, dir_path);
     ggml_backend_load_best("sycl", silent, dir_path);
diff --git a/ggml/src/ggml-backend.cpp b/ggml/src/ggml-backend.cpp
index b1050ad59..788861a36 100644
--- a/ggml/src/ggml-backend.cpp
+++ b/ggml/src/ggml-backend.cpp
@@ -817,8 +817,9 @@ static void ggml_backend_sched_print_assignments(ggml_backend_sched_t sched, str
         }
         if (sched->debug > 1) {
             ggml_backend_t tensor_backend = ggml_backend_sched_get_tensor_backend(sched, node);
-            GGML_LOG_DEBUG("node #%3d (%10.10s): %20.20s (%5.5s) [%5.5s %8.8s]:", i, ggml_op_name(node->op), node->name,
-                fmt_size(ggml_nbytes(node)), tensor_backend ? ggml_backend_name(tensor_backend) : "NULL", GET_CAUSE(node));
+            GGML_LOG_DEBUG("node #%3d (%10.10s): %20.20s (%5.5s) [%5.5s %8.8s] use=%d:", i, ggml_op_name(node->op), node->name,
+                fmt_size(ggml_nbytes(node)), tensor_backend ? ggml_backend_name(tensor_backend) : "NULL", GET_CAUSE(node),
+                graph->use_counts[ggml_hash_find(&graph->visited_hash_set, node)]);
             for (int j = 0; j < GGML_MAX_SRC; j++) {
                 struct ggml_tensor * src = node->src[j];
                 if (src == NULL) {
@@ -1826,7 +1827,7 @@ void ggml_backend_graph_copy_free(struct ggml_backend_graph_copy copy) {
     ggml_free(copy.ctx_unallocated);
 }
 
-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) {
+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, struct ggml_tensor * test_node) {
     struct ggml_backend_graph_copy copy = ggml_backend_graph_copy(backend2, graph);
     if (copy.buffer == NULL) {
         return false;
@@ -1837,28 +1838,45 @@ 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++) {
-        struct ggml_tensor * t1 = g1->nodes[i];
-        struct ggml_tensor * t2 = g2->nodes[i];
+    if (test_node != nullptr) {
+        // Compute the whole graph and only test the output for a specific tensor
+        ggml_backend_graph_compute(backend1, g1);
+        ggml_backend_graph_compute(backend2, g2);
 
-        assert(t1->op == t2->op && ggml_are_same_layout(t1, t2));
-
-        struct ggml_cgraph g1v = ggml_graph_view(g1, i, i + 1);
-        struct ggml_cgraph g2v = ggml_graph_view(g2, i, i + 1);
-
-        ggml_backend_graph_compute(backend1, &g1v);
-        ggml_backend_graph_compute(backend2, &g2v);
-
-        if (ggml_is_view_op(t1->op)) {
-            continue;
+        int test_node_idx = -1;
+        for (int i = 0; i < g1->n_nodes; i++) {
+            struct ggml_tensor * t1 = g1->nodes[i];
+            if (t1 == test_node) {
+                test_node_idx = i;
+                break;
+            }
         }
+        GGML_ASSERT(test_node_idx != -1);
 
-        // compare results, calculate rms etc
-        if (!callback(i, t1, t2, user_data)) {
-            break;
+        callback(test_node_idx, g1->nodes[test_node_idx], g2->nodes[test_node_idx], user_data);
+    } else {
+        for (int i = 0; i < g1->n_nodes; i++) {
+            struct ggml_tensor * t1 = g1->nodes[i];
+            struct ggml_tensor * t2 = g2->nodes[i];
+
+            assert(t1->op == t2->op && ggml_are_same_layout(t1, t2));
+
+            struct ggml_cgraph g1v = ggml_graph_view(g1, i, i + 1);
+            struct ggml_cgraph g2v = ggml_graph_view(g2, i, i + 1);
+
+            ggml_backend_graph_compute(backend1, &g1v);
+            ggml_backend_graph_compute(backend2, &g2v);
+
+            if (ggml_is_view_op(t1->op)) {
+                continue;
+            }
+
+            // compare results, calculate rms etc
+            if (!callback(i, t1, t2, user_data)) {
+                break;
+            }
         }
     }
-
     ggml_backend_graph_copy_free(copy);
 
     return true;
diff --git a/ggml/src/ggml-cann/aclnn_ops.cpp b/ggml/src/ggml-cann/aclnn_ops.cpp
index 437ece2d4..4d5c2c182 100755
--- a/ggml/src/ggml-cann/aclnn_ops.cpp
+++ b/ggml/src/ggml-cann/aclnn_ops.cpp
@@ -65,8 +65,9 @@
 #include <aclnnop/aclnn_eq_tensor.h>
 #include <aclnnop/aclnn_gt_scalar.h>
 #include <aclnnop/aclnn_pow.h>
-#include <aclnnop/aclnn_grouped_matmul_v2.h>
+#include <aclnnop/aclnn_grouped_matmul_v3.h>
 #include <aclnnop/aclnn_fused_infer_attention_score_v2.h>
+#include <aclnnop/aclnn_zero.h>
 #include <float.h>
 
 #include <cmath>
@@ -804,10 +805,11 @@ static aclTensor* aclnn_zero(ggml_backend_cann_context& ctx, void* buffer,
         nb[i] = nb[i - 1] * ne[i - 1];
     }
 
-    ggml_cann_async_memset(ctx, buffer, n_bytes, 0);
     aclTensor* zero =
         ggml_cann_create_tensor(buffer, type, type_size, ne, nb, dims);
+    GGML_CANN_CALL_ACLNN_OP(ctx, InplaceZero, zero);
     return zero;
+    GGML_UNUSED(n_bytes);
 }
 
 /**
@@ -2654,6 +2656,67 @@ static void ggml_cann_mul_mat_id_fp(ggml_backend_cann_context& ctx, ggml_tensor*
         memcpy(ori_src0_nb, cast_nb, sizeof(ori_src0_nb));
     }
 
+#ifdef ASCEND_310P
+    ggml_tensor src0_row = *src0;
+    ggml_tensor src1_row = *src1;
+    ggml_tensor dst_row = *dst;
+
+    if (src0->type == GGML_TYPE_F16) {
+        src0_row.type = GGML_TYPE_F32;
+    }
+
+    // src0_row [D, M, 1, 1] weight without permute
+    src0_row.ne[2] = 1;
+    src0_row.ne[3] = 1;
+    src0_row.nb[0] = ori_src0_nb[0];
+    src0_row.nb[1] = ori_src0_nb[1];
+    src0_row.nb[2] = ori_src0_nb[1];
+    src0_row.nb[3] = ori_src0_nb[1];
+
+    // src1_row [D, 1, 1, 1] -> input
+    src1_row.ne[1] = 1;
+    src1_row.ne[2] = 1;
+    src1_row.ne[3] = 1;
+    src1_row.nb[2] = nb11;
+    src1_row.nb[3] = nb11;
+
+    // dst_row [M, 1, 1, 1] -> out
+    dst_row.ne[1] = 1;
+    dst_row.ne[2] = 1;
+    dst_row.ne[3] = 1;
+    dst_row.nb[2] = nb1;
+    dst_row.nb[3] = nb1;
+
+    //create weight for one row
+    for (int64_t iid1 = 0; iid1 < ids->ne[1]; iid1++) {
+        for (int64_t id = 0; id < n_ids; id++) {
+            // expert index
+            int32_t i02 = *(int32_t *) (ids_host.data() + iid1*ids->nb[1] + id*ids->nb[0]);
+            GGML_ASSERT(i02 >= 0 && i02 < n_as);
+
+            // If B = 1 (broadcast), always use 0; otherwise, use id.
+            int64_t i11 = (ne11 == 1 ? 0 : id);
+            int64_t i12 = iid1;
+
+            int64_t i1 = id;
+            int64_t i2 = i12;
+
+            void* src0_tmp_ptr = src0_original + i02*ori_src0_nb[2];
+            void* src1_tmp_ptr = src1_original + i11*nb11 + i12*nb12;
+            void* dst_tmp_ptr  = dst_original  + i1*nb1   + i2*nb2;
+
+            src0_row.data = src0_tmp_ptr;
+            src1_row.data = src1_tmp_ptr;
+            dst_row.data = dst_tmp_ptr;
+            dst_row.src[0] = &src0_row;
+            dst_row.src[1] = &src1_row;
+
+            ggml_cann_mul_mat(ctx, &dst_row);
+        }
+    }
+    return;
+#endif
+
     std::vector<aclTensor*> src0_tensor_vec;
     std::vector<aclTensor*> src1_tensor_vec;
     std::vector<aclTensor*> dst_tensor_vec;
@@ -2701,9 +2764,9 @@ static void ggml_cann_mul_mat_id_fp(ggml_backend_cann_context& ctx, ggml_tensor*
     }
 
     size_t GROUP_SIZE = 128;
-    // GroupedMatmulV2 required tensor_list.size < 128
+    // GroupedMatmulV3 required tensor_list.size < 128
     for (size_t i = 0; i < src0_tensor_vec.size(); i += GROUP_SIZE) {
-        // split and call GroupedMatmulV2
+        // split and call GroupedMatmulV3
         size_t end = std::min(i + GROUP_SIZE, src0_tensor_vec.size());
         std::vector<aclTensor*> src0_tensor_vec_split(src0_tensor_vec.begin() + i, src0_tensor_vec.begin() + end);
         std::vector<aclTensor*> src1_tensor_vec_split(src1_tensor_vec.begin() + i, src1_tensor_vec.begin() + end);
@@ -2713,7 +2776,7 @@ static void ggml_cann_mul_mat_id_fp(ggml_backend_cann_context& ctx, ggml_tensor*
         aclTensorList* src1_tensor_list = aclCreateTensorList(src1_tensor_vec_split.data(), src1_tensor_vec_split.size());
         aclTensorList* dst_tensor_list = aclCreateTensorList(dst_tensor_vec_split.data(), dst_tensor_vec_split.size());
 
-        GGML_CANN_CALL_ACLNN_OP(ctx, GroupedMatmulV2, src1_tensor_list, src0_tensor_list,
+        GGML_CANN_CALL_ACLNN_OP(ctx, GroupedMatmulV3, src1_tensor_list, src0_tensor_list,
             nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, 0, -1, dst_tensor_list);
 
         ggml_cann_release_resources(ctx, src0_tensor_list, src1_tensor_list, dst_tensor_list);
diff --git a/ggml/src/ggml-cann/common.h b/ggml/src/ggml-cann/common.h
index 7ef80a479..8dfe3b061 100755
--- a/ggml/src/ggml-cann/common.h
+++ b/ggml/src/ggml-cann/common.h
@@ -37,6 +37,7 @@
 #include <thread>
 #include <unistd.h>
 #include <functional>
+#include <optional>
 
 #include "../include/ggml-cann.h"
 #include "../include/ggml.h"
@@ -103,6 +104,9 @@ const ggml_cann_device_info& ggml_cann_info();
 void ggml_cann_set_device(int32_t device);
 int32_t ggml_cann_get_device();
 
+std::optional<std::string> get_env(const std::string& name);
+bool parse_bool(const std::string& value);
+
 /**
  * @brief Abstract base class for memory pools used by CANN.
  */
@@ -354,7 +358,8 @@ struct ggml_backend_cann_context {
         : device(device), name("CANN" + std::to_string(device)), task_queue(1024, device) {
         ggml_cann_set_device(device);
         description = aclrtGetSocName();
-        async_mode = (getenv("GGML_CANN_ASYNC_MODE") != nullptr);
+
+        async_mode = parse_bool(get_env("GGML_CANN_ASYNC_MODE").value_or(""));
         GGML_LOG_INFO("%s: device %d async operator submission is %s\n", __func__,
             device, async_mode ? "ON" : "OFF");
     }
diff --git a/ggml/src/ggml-cann/ggml-cann.cpp b/ggml/src/ggml-cann/ggml-cann.cpp
index c0ea26002..eae575cc0 100755
--- a/ggml/src/ggml-cann/ggml-cann.cpp
+++ b/ggml/src/ggml-cann/ggml-cann.cpp
@@ -31,6 +31,8 @@
 #include <mutex>
 #include <queue>
 #include <chrono>
+#include <unordered_set>
+#include <optional>
 
 #include "ggml-impl.h"
 #include "ggml-backend-impl.h"
@@ -93,6 +95,26 @@ int32_t ggml_cann_get_device() {
     return id;
 }
 
+/**
+ * @brief Get the value of the specified environment variable (name).
+ *        if not empty, return a std::string object
+ */
+std::optional<std::string> get_env(const std::string& name) {
+    const char* val = std::getenv(name.c_str());
+    if (!val) return std::nullopt;
+    std::string res = std::string(val);
+    std::transform(res.begin(), res.end(), res.begin(), ::tolower);
+    return res;
+}
+
+/**
+ * @brief Verify whether the environment variable is a valid value.
+ */
+bool parse_bool(const std::string& value) {
+    std::unordered_set<std::string> valid_values = {"on", "1", "yes", "y", "enable", "true"};
+    return valid_values.find(value) != valid_values.end();
+}
+
 /**
  * @brief Initialize the CANN device information.
  *
@@ -214,7 +236,7 @@ struct ggml_cann_pool_buf_prio : public ggml_cann_pool {
      * @param device The device ID to associate with this buffer pool.
      */
     explicit ggml_cann_pool_buf_prio(int device) : device(device) {
-        disable_clean = getenv("GGML_CANN_DISABLE_BUF_POOL_CLEAN") != nullptr;
+        disable_clean = parse_bool(get_env("GGML_CANN_DISABLE_BUF_POOL_CLEAN").value_or(""));
     }
 
     /**
@@ -410,7 +432,7 @@ struct ggml_cann_pool_buf : public ggml_cann_pool {
      * @param device The device ID to associate with this buffer pool.
      */
     explicit ggml_cann_pool_buf(int device) : device(device) {
-        disable_clean = getenv("GGML_CANN_DISABLE_BUF_POOL_CLEAN") != nullptr;
+        disable_clean = parse_bool(get_env("GGML_CANN_DISABLE_BUF_POOL_CLEAN").value_or(""));
     }
 
     /**
@@ -731,16 +753,18 @@ struct ggml_cann_pool_vmm : public ggml_cann_pool {
  */
 std::unique_ptr<ggml_cann_pool> ggml_backend_cann_context::new_pool_for_device(
     int device) {
-    bool disable_vmm = (getenv("GGML_CANN_DISABLE_VMM_POOL") != nullptr);
-    if (!disable_vmm && ggml_cann_info().devices[device].vmm) {
-        GGML_LOG_INFO("%s: device %d use vmm pool\n", __func__, device);
-        return std::unique_ptr<ggml_cann_pool>(new ggml_cann_pool_vmm(device));
-    }
-    bool enable_buf_prio = (getenv("GGML_CANN_ENABLE_BUF_PRIO_POOL") != nullptr);
-    if (enable_buf_prio) {
+    std::string mem_pool_type = get_env("GGML_CANN_MEM_POOL").value_or("");
+
+    if (mem_pool_type == "prio") {
         GGML_LOG_INFO("%s: device %d use buffer pool with priority queue\n", __func__, device);
         return std::unique_ptr<ggml_cann_pool>(new ggml_cann_pool_buf_prio(device));
     }
+
+    if (ggml_cann_info().devices[device].vmm && mem_pool_type != "leg") {
+        GGML_LOG_INFO("%s: device %d use vmm pool\n", __func__, device);
+        return std::unique_ptr<ggml_cann_pool>(new ggml_cann_pool_vmm(device));
+    }
+
     GGML_LOG_INFO("%s: device %d use buffer pool\n", __func__, device);
     return std::unique_ptr<ggml_cann_pool>(new ggml_cann_pool_buf(device));
 }
@@ -2062,6 +2086,12 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
                     return false;
             }
         } break;
+        case GGML_OP_SET_ROWS:
+            {
+                // TODO: add support
+                // ref: https://github.com/ggml-org/llama.cpp/pull/14274
+                return false;
+            } break;
         case GGML_OP_CPY: {
             ggml_tensor *src = op->src[0];
             if ((op->type != GGML_TYPE_F32 && op->type != GGML_TYPE_F16) ||
@@ -2163,7 +2193,6 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
         case GGML_OP_SQRT:
         case GGML_OP_CLAMP:
         case GGML_OP_DIAG_MASK_INF:
-        case GGML_OP_SOFT_MAX:
         case GGML_OP_SUM_ROWS:
         case GGML_OP_ARGSORT:
         case GGML_OP_ACC:
@@ -2181,6 +2210,10 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
         case GGML_OP_PAD_REFLECT_1D:
         case GGML_OP_COUNT_EQUAL:
             return true;
+        case GGML_OP_SOFT_MAX:
+            // TODO: support broadcast
+            // ref: https://github.com/ggml-org/llama.cpp/pull/14435
+            return !op->src[1] || (op->src[1]->ne[2] == 1 && op->src[1]->ne[3] == 1);
         case GGML_OP_FLASH_ATTN_EXT:{
             // derived from [ggml-cuda.cu]
             if(op->src[1]->type != GGML_TYPE_F16 || op->src[2]->type != GGML_TYPE_F16){
@@ -2203,6 +2236,8 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
                 // DeepSeek MLA
                 return false;
             }
+            // TODO: support broadcast
+            // ref: https://github.com/ggml-org/llama.cpp/pull/14435
             if (op->src[0]->ne[3] != 1) {
                 return false;
             }
diff --git a/ggml/src/ggml-common.h b/ggml/src/ggml-common.h
index 086c822d7..fbb04426a 100644
--- a/ggml/src/ggml-common.h
+++ b/ggml/src/ggml-common.h
@@ -1074,6 +1074,10 @@ GGML_TABLE_BEGIN(uint32_t, iq3s_grid, 512)
     0x0f090307, 0x0f090501, 0x0f090b01, 0x0f0b0505, 0x0f0b0905, 0x0f0d0105, 0x0f0d0703, 0x0f0f0101,
 GGML_TABLE_END()
 
+GGML_TABLE_BEGIN(int8_t, kvalues_iq4nl, 16)
+    -127, -104, -83, -65, -49, -35, -22, -10, 1, 13, 25, 38, 53, 69, 89, 113,
+GGML_TABLE_END()
+
 #define NGRID_IQ1S 2048
 #define IQ1S_DELTA 0.125f
 #define IQ1M_DELTA 0.125f
diff --git a/ggml/src/ggml-cpu/CMakeLists.txt b/ggml/src/ggml-cpu/CMakeLists.txt
index b3237eead..66a5ad8d2 100644
--- a/ggml/src/ggml-cpu/CMakeLists.txt
+++ b/ggml/src/ggml-cpu/CMakeLists.txt
@@ -1,3 +1,17 @@
+function(ggml_add_cpu_backend_features cpu_name arch)
+    # The feature detection code is compiled as a separate target so that
+    # it can be built without the architecture flags
+    # Since multiple variants of the CPU backend may be included in the same
+    # build, using set_source_files_properties() to set the arch flags is not possible
+    set(GGML_CPU_FEATS_NAME ${cpu_name}-feats)
+    add_library(${GGML_CPU_FEATS_NAME} OBJECT ggml-cpu/arch/${arch}/cpu-feats.cpp)
+    target_include_directories(${GGML_CPU_FEATS_NAME} PRIVATE . ../include)
+    target_compile_definitions(${GGML_CPU_FEATS_NAME} PRIVATE ${ARGN})
+    target_compile_definitions(${GGML_CPU_FEATS_NAME} PRIVATE GGML_BACKEND_DL GGML_BACKEND_BUILD GGML_BACKEND_SHARED)
+    set_target_properties(${GGML_CPU_FEATS_NAME} PROPERTIES POSITION_INDEPENDENT_CODE ON)
+    target_link_libraries(${cpu_name} PRIVATE ${GGML_CPU_FEATS_NAME})
+endfunction()
+
 function(ggml_add_cpu_backend_variant_impl tag_name)
     if (tag_name)
         set(GGML_CPU_NAME ggml-cpu-${tag_name})
@@ -10,14 +24,14 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
     list (APPEND GGML_CPU_SOURCES
         ggml-cpu/ggml-cpu.c
         ggml-cpu/ggml-cpu.cpp
-        ggml-cpu/ggml-cpu-aarch64.cpp
-        ggml-cpu/ggml-cpu-aarch64.h
-        ggml-cpu/ggml-cpu-hbm.cpp
-        ggml-cpu/ggml-cpu-hbm.h
-        ggml-cpu/ggml-cpu-quants.c
-        ggml-cpu/ggml-cpu-quants.h
-        ggml-cpu/ggml-cpu-traits.cpp
-        ggml-cpu/ggml-cpu-traits.h
+        ggml-cpu/repack.cpp
+        ggml-cpu/repack.h
+        ggml-cpu/hbm.cpp
+        ggml-cpu/hbm.h
+        ggml-cpu/quants.c
+        ggml-cpu/quants.h
+        ggml-cpu/traits.cpp
+        ggml-cpu/traits.h
         ggml-cpu/amx/amx.cpp
         ggml-cpu/amx/amx.h
         ggml-cpu/amx/mmq.cpp
@@ -84,6 +98,11 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
 
     if (GGML_SYSTEM_ARCH STREQUAL "ARM")
         message(STATUS "ARM detected")
+        list(APPEND GGML_CPU_SOURCES
+            ggml-cpu/arch/arm/quants.c
+            ggml-cpu/arch/arm/repack.cpp
+            )
+
         if (MSVC AND NOT CMAKE_C_COMPILER_ID STREQUAL "Clang")
             message(FATAL_ERROR "MSVC is not supported for ARM, use clang")
         else()
@@ -138,6 +157,49 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
             else()
                 if (GGML_CPU_ARM_ARCH)
                     list(APPEND ARCH_FLAGS -march=${GGML_CPU_ARM_ARCH})
+                elseif(GGML_CPU_ALL_VARIANTS)
+                    # Begin with the lowest baseline
+                    set(ARM_MCPU "armv8-a")
+                    set(ARCH_TAGS "")
+                    set(ARCH_DEFINITIONS "")
+
+                    # When a feature is selected, bump the MCPU to the first
+                    # version that supported it
+                    if (GGML_INTERNAL_DOTPROD)
+                        set(ARM_MCPU "armv8.2-a")
+                        set(ARCH_TAGS "${ARCH_TAGS}+dotprod")
+                        list(APPEND ARCH_DEFINITIONS GGML_USE_DOTPROD)
+                    endif()
+                    if (GGML_INTERNAL_FP16_VECTOR_ARITHMETIC)
+                        set(ARM_MCPU "armv8.2-a")
+                        set(ARCH_TAGS "${ARCH_TAGS}+fp16")
+                        list(APPEND ARCH_DEFINITIONS GGML_USE_FP16_VECTOR_ARITHMETIC)
+                    endif()
+                    if (GGML_INTERNAL_SVE)
+                        set(ARM_MCPU "armv8.2-a")
+                        set(ARCH_TAGS "${ARCH_TAGS}+sve")
+                        list(APPEND ARCH_DEFINITIONS GGML_USE_SVE)
+                    endif()
+                    if (GGML_INTERNAL_MATMUL_INT8)
+                        set(ARM_MCPU "armv8.6-a")
+                        set(ARCH_TAGS "${ARCH_TAGS}+i8mm")
+                        list(APPEND ARCH_DEFINITIONS GGML_USE_MATMUL_INT8)
+                    endif()
+                    if (GGML_INTERNAL_SVE2)
+                        set(ARM_MCPU "armv8.6-a")
+                        set(ARCH_TAGS "${ARCH_TAGS}+sve2")
+                        list(APPEND ARCH_DEFINITIONS GGML_USE_SVE2)
+                    endif()
+                    if (GGML_INTERNAL_NOSVE)
+                        set(ARCH_TAGS "${ARCH_TAGS}+nosve")
+                    endif()
+                    if (GGML_INTERNAL_SME)
+                        set(ARM_MCPU "armv9.2-a")
+                        set(ARCH_TAGS "${ARCH_TAGS}+sme")
+                        list(APPEND ARCH_DEFINITIONS GGML_USE_SME)
+                    endif()
+                    list(APPEND ARCH_FLAGS "-march=${ARM_MCPU}${ARCH_TAGS}")
+                    ggml_add_cpu_backend_features(${GGML_CPU_NAME} arm ${ARCH_DEFINITIONS})
                 endif()
             endif()
 
@@ -167,6 +229,11 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
         endif()
     elseif (GGML_SYSTEM_ARCH STREQUAL "x86")
         message(STATUS "x86 detected")
+        list(APPEND GGML_CPU_SOURCES
+            ggml-cpu/arch/x86/quants.c
+            ggml-cpu/arch/x86/repack.cpp
+            )
+
         if (MSVC)
             # instruction set detection for MSVC only
             if (GGML_NATIVE)
@@ -296,21 +363,11 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
                 # the feature check relies on ARCH_DEFINITIONS, but it is not set with GGML_NATIVE
                 message(FATAL_ERROR "GGML_NATIVE is not compatible with GGML_BACKEND_DL, consider using GGML_CPU_ALL_VARIANTS")
             endif()
-
-            # The feature detection code is compiled as a separate target so that
-            # it can be built without the architecture flags
-            # Since multiple variants of the CPU backend may be included in the same
-            # build, using set_source_files_properties() to set the arch flags is not possible
-            set(GGML_CPU_FEATS_NAME ${GGML_CPU_NAME}-feats)
-            add_library(${GGML_CPU_FEATS_NAME} OBJECT ggml-cpu/cpu-feats-x86.cpp)
-            target_include_directories(${GGML_CPU_FEATS_NAME} PRIVATE . .. ../include)
-            target_compile_definitions(${GGML_CPU_FEATS_NAME} PRIVATE ${ARCH_DEFINITIONS})
-            target_compile_definitions(${GGML_CPU_FEATS_NAME} PRIVATE GGML_BACKEND_DL GGML_BACKEND_BUILD GGML_BACKEND_SHARED)
-            set_target_properties(${GGML_CPU_FEATS_NAME} PROPERTIES POSITION_INDEPENDENT_CODE ON)
-            target_link_libraries(${GGML_CPU_NAME} PRIVATE ${GGML_CPU_FEATS_NAME})
+            ggml_add_cpu_backend_features(${GGML_CPU_NAME} x86 ${ARCH_DEFINITIONS})
         endif()
     elseif (GGML_SYSTEM_ARCH STREQUAL "PowerPC")
         message(STATUS "PowerPC detected")
+        list(APPEND GGML_CPU_SOURCES ggml-cpu/arch/powerpc/quants.c)
         if (GGML_NATIVE)
             if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64")
                 file(READ "/proc/cpuinfo" POWER10_M)
@@ -318,7 +375,8 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
                 execute_process(COMMAND bash -c "prtconf |grep 'Implementation' | head -n 1" OUTPUT_VARIABLE POWER10_M)
             endif()
 
-            string(REGEX MATCHALL "POWER *([0-9]+)" MATCHED_STRING "${POWER10_M}")
+            string(TOUPPER "${POWER10_M}" POWER10_M_UPPER)
+            string(REGEX MATCHALL "POWER *([0-9]+)" MATCHED_STRING "${POWER10_M_UPPER}")
             string(REGEX REPLACE "POWER *([0-9]+)" "\\1" EXTRACTED_NUMBER "${MATCHED_STRING}")
 
             if (EXTRACTED_NUMBER GREATER_EQUAL 10)
@@ -330,6 +388,27 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
             else()
                 list(APPEND ARCH_FLAGS -mcpu=native -mtune=native -mpowerpc64)
             endif()
+        elseif(GGML_CPU_ALL_VARIANTS)
+            # Begin with the lowest baseline
+            set(ARCH_DEFINITIONS "")
+
+            # When a feature is selected, bump the MCPU to the first
+            # version that supported it
+            foreach(PVER RANGE 7 11)
+                if(DEFINED GGML_INTERNAL_POWER${PVER})
+                    set(POWERPC_MCPU "power${PVER}")
+                    list(APPEND ARCH_DEFINITIONS GGML_USE_POWER${PVER})
+                endif()
+            endforeach()
+            if (GGML_INTERNAL_VSX)
+                list(APPEND ARCH_DEFINITIONS GGML_USE_VSX)
+                list(APPEND ARCH_FLAGS -mvsx)
+            endif()
+
+            if (DEFINED POWERPC_MCPU)
+                list(APPEND ARCH_FLAGS -mcpu=${POWERPC_MCPU})
+            endif()
+            ggml_add_cpu_backend_features(${GGML_CPU_NAME} powerpc ${ARCH_DEFINITIONS})
         else()
             if (GGML_CPU_POWERPC_CPUTYPE)
                 list(APPEND ARCH_FLAGS -mcpu=${GGML_CPU_POWERPC_CPUTYPE})
@@ -337,6 +416,8 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
         endif()
     elseif (GGML_SYSTEM_ARCH STREQUAL "loongarch64")
         message(STATUS "loongarch64 detected")
+        list(APPEND GGML_CPU_SOURCES ggml-cpu/arch/loongarch/quants.c)
+
         list(APPEND ARCH_FLAGS -march=loongarch64)
         if (GGML_LASX)
             list(APPEND ARCH_FLAGS -mlasx)
@@ -346,6 +427,10 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
         endif()
     elseif (GGML_SYSTEM_ARCH STREQUAL "riscv64")
         message(STATUS "riscv64 detected")
+        list(APPEND GGML_CPU_SOURCES
+            ggml-cpu/arch/riscv/quants.c
+            ggml-cpu/arch/riscv/repack.cpp
+            )
         if (GGML_RVV)
             if (GGML_XTHEADVECTOR)
                 list(APPEND ARCH_FLAGS -march=rv64gc_xtheadvector -mabi=lp64d)
@@ -357,11 +442,13 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
         endif()
     elseif (GGML_SYSTEM_ARCH STREQUAL "s390x")
         message(STATUS "s390x detected")
+        list(APPEND GGML_CPU_SOURCES ggml-cpu/arch/s390/quants.c)
         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 (${S390X_M} MATCHES "8561|8562")
+            set(GGML_NNPA OFF)
             message(STATUS "z15 target")
             list(APPEND ARCH_FLAGS -march=z15)
         elseif (${S390X_M} MATCHES "3931")
@@ -378,14 +465,25 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
         endif()
 
         if (GGML_VXE)
+            message(STATUS "VX/VXE/VXE2 enabled")
             list(APPEND ARCH_FLAGS -mvx -mzvector)
+            list(APPEND ARCH_DEFINITIONS GGML_VXE)
         endif()
+
+        if (GGML_NNPA)
+            message(STATUS "NNPA enabled")
+            list(APPEND ARCH_DEFINITIONS GGML_NNPA)
+        endif()
+    elseif (CMAKE_SYSTEM_PROCESSOR MATCHES "wasm")
+        message(STATUS "Wasm detected")
+        list (APPEND GGML_CPU_SOURCES ggml-cpu/arch/wasm/quants.c)
     else()
-        message(STATUS "Unknown architecture")
+        message(WARNING "Unknown CPU architecture. Falling back to generic implementations.")
+        list(APPEND ARCH_FLAGS -DGGML_CPU_GENERIC)
     endif()
 
-    if (GGML_CPU_AARCH64)
-        target_compile_definitions(${GGML_CPU_NAME} PRIVATE GGML_USE_CPU_AARCH64)
+    if (GGML_CPU_REPACK)
+        target_compile_definitions(${GGML_CPU_NAME} PRIVATE GGML_USE_CPU_REPACK)
     endif()
 
     if (GGML_CPU_KLEIDIAI)
@@ -396,9 +494,9 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
 
         # Fetch KleidiAI sources:
         include(FetchContent)
-        set(KLEIDIAI_COMMIT_TAG "v1.6.0")
+        set(KLEIDIAI_COMMIT_TAG "v1.9.0")
         set(KLEIDIAI_DOWNLOAD_URL "https://github.com/ARM-software/kleidiai/archive/refs/tags/${KLEIDIAI_COMMIT_TAG}.tar.gz")
-        set(KLEIDIAI_ARCHIVE_MD5  "75b4ad68f25ab673dcc01065e5a0b05f")
+        set(KLEIDIAI_ARCHIVE_MD5  "2a8e1bb55d201557553545536489a017")
 
         if (POLICY CMP0135)
             cmake_policy(SET CMP0135 NEW)
@@ -491,4 +589,9 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
     if (EMSCRIPTEN)
         set_target_properties(${GGML_CPU_NAME} PROPERTIES COMPILE_FLAGS "-msimd128")
     endif()
+
+    if (CMAKE_CXX_COMPILER_ID STREQUAL "IntelLLVM")
+        # The compiler automatically enables "-ffast-math" which can cause NaNs in tests due to "-fassociative-math"
+        target_compile_options(${GGML_CPU_NAME} PRIVATE "-fno-associative-math")
+    endif()
 endfunction()
diff --git a/ggml/src/ggml-cpu/amx/amx.cpp b/ggml/src/ggml-cpu/amx/amx.cpp
index 0f067137d..258857b00 100644
--- a/ggml/src/ggml-cpu/amx/amx.cpp
+++ b/ggml/src/ggml-cpu/amx/amx.cpp
@@ -5,7 +5,7 @@
 #include "ggml-backend.h"
 #include "ggml-impl.h"
 #include "ggml-cpu.h"
-#include "ggml-cpu-traits.h"
+#include "traits.h"
 
 #if defined(__gnu_linux__)
 #include <sys/syscall.h>
diff --git a/ggml/src/ggml-cpu/amx/mmq.cpp b/ggml/src/ggml-cpu/amx/mmq.cpp
index 0ea91596b..47c61b881 100644
--- a/ggml/src/ggml-cpu/amx/mmq.cpp
+++ b/ggml/src/ggml-cpu/amx/mmq.cpp
@@ -8,7 +8,8 @@
 #include "mmq.h"
 #include "ggml-impl.h"
 #include "ggml-cpu-impl.h"
-#include "ggml-cpu-quants.h"
+#include "simd-mappings.h"
+#include "quants.h"
 #include "ggml-quants.h"
 #include <algorithm>
 #include <type_traits>
@@ -453,7 +454,7 @@ void quantize_row_q8_K_vnni(const float * RESTRICT x, void * RESTRICT vy, int64_
 
         // Quantize these floats
         const float iscale = 127.f / amax;
-        y[i].d = GGML_FP32_TO_FP16(1 / iscale);
+        y[i].d = GGML_CPU_FP32_TO_FP16(1 / iscale);
         const float id = ( amax != 0.0f ) ? iscale : 0.f;
         const __m512 vscale = _mm512_set1_ps(id);
 
@@ -1090,7 +1091,7 @@ struct acc_C<block_q8_0, block_q4_0, is_acc> {
         const __m512 vd0 = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)((const char *)packed_B + offset)));
 
         for (int m = 0; m < nr; ++m) {
-            const __m512 vd1 = _mm512_set1_ps(GGML_FP16_TO_FP32(A[m * lda].d));
+            const __m512 vd1 = _mm512_set1_ps(GGML_CPU_FP16_TO_FP32(A[m * lda].d));
             const __m512 vtile = _mm512_cvtepi32_ps(_mm512_loadu_si512(tile + m * TILE_N));
 
             __m512 vsum;
@@ -1113,8 +1114,8 @@ struct acc_C<block_q8_1, block_q4_1, is_acc> {
         const __m512 vm0 = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)((const char *)packed_B + offset + TILE_N * sizeof(ggml_half))));
 
         for (int m = 0; m < nr; ++m) {
-            const __m512 vd1 = _mm512_set1_ps(GGML_FP16_TO_FP32(A[m * lda].d));
-            const __m512 vs1 = _mm512_set1_ps(GGML_FP16_TO_FP32(A[m * lda].s));
+            const __m512 vd1 = _mm512_set1_ps(GGML_CPU_FP16_TO_FP32(A[m * lda].d));
+            const __m512 vs1 = _mm512_set1_ps(GGML_CPU_FP16_TO_FP32(A[m * lda].s));
             const __m512 vtile = _mm512_cvtepi32_ps(_mm512_loadu_si512(tile + m * TILE_N));
 
             __m512 vsum;
@@ -1137,7 +1138,7 @@ struct acc_C<block_q8_0, block_q8_0, is_acc> {
         const __m512 vd0 = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)((const char *)packed_B + offset)));
 
         for (int m = 0; m < nr; ++m) {
-            const __m512 vd1 = _mm512_set1_ps(GGML_FP16_TO_FP32(A[m * lda].d));
+            const __m512 vd1 = _mm512_set1_ps(GGML_CPU_FP16_TO_FP32(A[m * lda].d));
             const __m512 vtile = _mm512_cvtepi32_ps(_mm512_loadu_si512(tile + m * TILE_N));
 
             __m512 vsum;
@@ -1437,7 +1438,7 @@ struct tinygemm_kernel_vnni<block_q8_0, block_q4_0, float, BLOCK_M, BLOCK_N, BLO
                     va[k] = _mm512_set1_epi32(a_ptr[k]);
                     vcomp = _mm512_dpbusd_epi32(vcomp, off, va[k]);
                 }
-                vd1 = _mm512_set1_ps(GGML_FP16_TO_FP32(A[0 * KB + i].d));
+                vd1 = _mm512_set1_ps(GGML_CPU_FP16_TO_FP32(A[0 * KB + i].d));
             }
 
             // load b
@@ -1498,8 +1499,8 @@ struct tinygemm_kernel_vnni<block_q8_1, block_q4_1, float, 1, BLOCK_N, BLOCK_K>
                 for (int k = 0; k < 8; ++k) {
                     va[k] = _mm512_set1_epi32(a_ptr[k]);
                 }
-                vd1 = _mm512_set1_ps(GGML_FP16_TO_FP32(A[0 * KB + i].d));
-                vs1 = _mm512_set1_ps(GGML_FP16_TO_FP32(A[0 * KB + i].s));
+                vd1 = _mm512_set1_ps(GGML_CPU_FP16_TO_FP32(A[0 * KB + i].d));
+                vs1 = _mm512_set1_ps(GGML_CPU_FP16_TO_FP32(A[0 * KB + i].s));
             }
 
             // load b
@@ -1571,7 +1572,7 @@ struct tinygemm_kernel_vnni<block_q8_0, block_q8_0, float, BLOCK_M, BLOCK_N, BLO
                     va[k] = _mm512_set1_epi32(a_ptr[k]);
                     va[k] = _mm512_add_epi8(va[k], off);
                 }
-                vd1 = _mm512_set1_ps(GGML_FP16_TO_FP32(A[0 * KB + i].d));
+                vd1 = _mm512_set1_ps(GGML_CPU_FP16_TO_FP32(A[0 * KB + i].d));
             }
 
             // load b
diff --git a/ggml/src/ggml-cpu/arch-fallback.h b/ggml/src/ggml-cpu/arch-fallback.h
new file mode 100644
index 000000000..10e534251
--- /dev/null
+++ b/ggml/src/ggml-cpu/arch-fallback.h
@@ -0,0 +1,184 @@
+#pragma once
+
+// Rename `_generic` functions if no native implementation is available.
+// This effectively selects the generic implementation.
+
+#if defined(GGML_CPU_GENERIC)
+// quants.c
+#define quantize_row_q8_0_generic quantize_row_q8_0
+#define quantize_row_q8_1_generic quantize_row_q8_1
+#define quantize_row_q8_K_generic quantize_row_q8_K
+#define ggml_vec_dot_q4_0_q8_0_generic ggml_vec_dot_q4_0_q8_0
+#define ggml_vec_dot_q4_1_q8_1_generic ggml_vec_dot_q4_1_q8_1
+#define ggml_vec_dot_q5_0_q8_0_generic ggml_vec_dot_q5_0_q8_0
+#define ggml_vec_dot_q5_1_q8_1_generic ggml_vec_dot_q5_1_q8_1
+#define ggml_vec_dot_q8_0_q8_0_generic ggml_vec_dot_q8_0_q8_0
+#define ggml_vec_dot_tq1_0_q8_K_generic ggml_vec_dot_tq1_0_q8_K
+#define ggml_vec_dot_tq2_0_q8_K_generic ggml_vec_dot_tq2_0_q8_K
+#define ggml_vec_dot_q2_K_q8_K_generic ggml_vec_dot_q2_K_q8_K
+#define ggml_vec_dot_q3_K_q8_K_generic ggml_vec_dot_q3_K_q8_K
+#define ggml_vec_dot_q4_K_q8_K_generic ggml_vec_dot_q4_K_q8_K
+#define ggml_vec_dot_q5_K_q8_K_generic ggml_vec_dot_q5_K_q8_K
+#define ggml_vec_dot_q6_K_q8_K_generic ggml_vec_dot_q6_K_q8_K
+#define ggml_vec_dot_iq2_xxs_q8_K_generic ggml_vec_dot_iq2_xxs_q8_K
+#define ggml_vec_dot_iq2_xs_q8_K_generic ggml_vec_dot_iq2_xs_q8_K
+#define ggml_vec_dot_iq2_s_q8_K_generic ggml_vec_dot_iq2_s_q8_K
+#define ggml_vec_dot_iq3_xxs_q8_K_generic ggml_vec_dot_iq3_xxs_q8_K
+#define ggml_vec_dot_iq3_s_q8_K_generic ggml_vec_dot_iq3_s_q8_K
+#define ggml_vec_dot_iq1_s_q8_K_generic ggml_vec_dot_iq1_s_q8_K
+#define ggml_vec_dot_iq1_m_q8_K_generic ggml_vec_dot_iq1_m_q8_K
+#define ggml_vec_dot_iq4_nl_q8_0_generic ggml_vec_dot_iq4_nl_q8_0
+#define ggml_vec_dot_iq4_xs_q8_K_generic ggml_vec_dot_iq4_xs_q8_K
+// repack.cpp
+#define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
+#define ggml_quantize_mat_q8_0_4x8_generic ggml_quantize_mat_q8_0_4x8
+#define ggml_quantize_mat_q8_K_4x8_generic ggml_quantize_mat_q8_K_4x8
+#define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0
+#define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
+#define ggml_gemv_q4_0_8x8_q8_0_generic ggml_gemv_q4_0_8x8_q8_0
+#define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
+#define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
+#define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
+#define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
+#define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0
+#define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
+#define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
+#elif defined(__aarch64__) || defined(__arm__) || defined(_M_ARM) || defined(_M_ARM64)
+// repack.cpp
+#define ggml_quantize_mat_q8_K_4x8_generic ggml_quantize_mat_q8_K_4x8
+#define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
+#define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
+#elif defined(__x86_64__) || defined(__i386__) || defined(_M_IX86) || defined(_M_X64)
+// repack.cpp
+#define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
+#define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0
+#define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
+#define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
+#define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
+#define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
+#define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
+#elif defined(__POWERPC__) || defined(__powerpc__)
+// ref: https://github.com/ggml-org/llama.cpp/pull/14146#issuecomment-2972561679
+// quants.c
+#define quantize_row_q8_K_generic quantize_row_q8_K
+#define ggml_vec_dot_tq1_0_q8_K_generic ggml_vec_dot_tq1_0_q8_K
+#define ggml_vec_dot_tq2_0_q8_K_generic ggml_vec_dot_tq2_0_q8_K
+#define ggml_vec_dot_iq1_m_q8_K_generic ggml_vec_dot_iq1_m_q8_K
+// repack.cpp
+#define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
+#define ggml_quantize_mat_q8_0_4x8_generic ggml_quantize_mat_q8_0_4x8
+#define ggml_quantize_mat_q8_K_4x8_generic ggml_quantize_mat_q8_K_4x8
+#define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0
+#define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
+#define ggml_gemv_q4_0_8x8_q8_0_generic ggml_gemv_q4_0_8x8_q8_0
+#define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
+#define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
+#define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
+#define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
+#define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0
+#define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
+#define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
+#elif defined(__loongarch64)
+// quants.c
+#define quantize_row_q8_K_generic quantize_row_q8_K
+#define ggml_vec_dot_tq1_0_q8_K_generic ggml_vec_dot_tq1_0_q8_K
+#define ggml_vec_dot_tq2_0_q8_K_generic ggml_vec_dot_tq2_0_q8_K
+#define ggml_vec_dot_iq1_m_q8_K_generic ggml_vec_dot_iq1_m_q8_K
+// repack.cpp
+#define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
+#define ggml_quantize_mat_q8_0_4x8_generic ggml_quantize_mat_q8_0_4x8
+#define ggml_quantize_mat_q8_K_4x8_generic ggml_quantize_mat_q8_K_4x8
+#define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0
+#define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
+#define ggml_gemv_q4_0_8x8_q8_0_generic ggml_gemv_q4_0_8x8_q8_0
+#define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
+#define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
+#define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
+#define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
+#define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0
+#define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
+#define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
+#elif defined(__riscv)
+// quants.c
+#define quantize_row_q8_K_generic quantize_row_q8_K
+#define ggml_vec_dot_tq1_0_q8_K_generic ggml_vec_dot_tq1_0_q8_K
+#define ggml_vec_dot_tq2_0_q8_K_generic ggml_vec_dot_tq2_0_q8_K
+#define ggml_vec_dot_iq2_xxs_q8_K_generic ggml_vec_dot_iq2_xxs_q8_K
+#define ggml_vec_dot_iq2_xs_q8_K_generic ggml_vec_dot_iq2_xs_q8_K
+#define ggml_vec_dot_iq2_s_q8_K_generic ggml_vec_dot_iq2_s_q8_K
+#define ggml_vec_dot_iq3_xxs_q8_K_generic ggml_vec_dot_iq3_xxs_q8_K
+#define ggml_vec_dot_iq3_s_q8_K_generic ggml_vec_dot_iq3_s_q8_K
+#define ggml_vec_dot_iq1_s_q8_K_generic ggml_vec_dot_iq1_s_q8_K
+#define ggml_vec_dot_iq1_m_q8_K_generic ggml_vec_dot_iq1_m_q8_K
+#define ggml_vec_dot_iq4_nl_q8_0_generic ggml_vec_dot_iq4_nl_q8_0
+#define ggml_vec_dot_iq4_xs_q8_K_generic ggml_vec_dot_iq4_xs_q8_K
+// repack.cpp
+#define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
+#define ggml_quantize_mat_q8_0_4x8_generic ggml_quantize_mat_q8_0_4x8
+#define ggml_quantize_mat_q8_K_4x8_generic ggml_quantize_mat_q8_K_4x8
+#define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0
+#define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
+#define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
+#define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
+#define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
+#define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
+#define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
+#define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
+#elif defined(__s390x__)
+// quants.c
+#define quantize_row_q8_K_generic quantize_row_q8_K
+#define ggml_vec_dot_q5_0_q8_0_generic ggml_vec_dot_q5_0_q8_0
+#define ggml_vec_dot_q5_1_q8_1_generic ggml_vec_dot_q5_1_q8_1
+#define ggml_vec_dot_tq1_0_q8_K_generic ggml_vec_dot_tq1_0_q8_K
+#define ggml_vec_dot_tq2_0_q8_K_generic ggml_vec_dot_tq2_0_q8_K
+#define ggml_vec_dot_q2_K_q8_K_generic ggml_vec_dot_q2_K_q8_K
+#define ggml_vec_dot_iq2_xxs_q8_K_generic ggml_vec_dot_iq2_xxs_q8_K
+#define ggml_vec_dot_iq2_xs_q8_K_generic ggml_vec_dot_iq2_xs_q8_K
+#define ggml_vec_dot_iq2_s_q8_K_generic ggml_vec_dot_iq2_s_q8_K
+#define ggml_vec_dot_iq3_xxs_q8_K_generic ggml_vec_dot_iq3_xxs_q8_K
+#define ggml_vec_dot_iq3_s_q8_K_generic ggml_vec_dot_iq3_s_q8_K
+#define ggml_vec_dot_iq1_s_q8_K_generic ggml_vec_dot_iq1_s_q8_K
+#define ggml_vec_dot_iq1_m_q8_K_generic ggml_vec_dot_iq1_m_q8_K
+// repack.cpp
+#define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
+#define ggml_quantize_mat_q8_0_4x8_generic ggml_quantize_mat_q8_0_4x8
+#define ggml_quantize_mat_q8_K_4x8_generic ggml_quantize_mat_q8_K_4x8
+#define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0
+#define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
+#define ggml_gemv_q4_0_8x8_q8_0_generic ggml_gemv_q4_0_8x8_q8_0
+#define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
+#define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
+#define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
+#define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
+#define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0
+#define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
+#define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
+#elif defined(__wasm__)
+// quants.c
+#define ggml_vec_dot_q4_1_q8_1_generic ggml_vec_dot_q4_1_q8_1
+#define ggml_vec_dot_tq1_0_q8_K_generic ggml_vec_dot_tq1_0_q8_K
+#define ggml_vec_dot_tq2_0_q8_K_generic ggml_vec_dot_tq2_0_q8_K
+#define ggml_vec_dot_iq2_xxs_q8_K_generic ggml_vec_dot_iq2_xxs_q8_K
+#define ggml_vec_dot_iq2_xs_q8_K_generic ggml_vec_dot_iq2_xs_q8_K
+#define ggml_vec_dot_iq2_s_q8_K_generic ggml_vec_dot_iq2_s_q8_K
+#define ggml_vec_dot_iq3_xxs_q8_K_generic ggml_vec_dot_iq3_xxs_q8_K
+#define ggml_vec_dot_iq3_s_q8_K_generic ggml_vec_dot_iq3_s_q8_K
+#define ggml_vec_dot_iq1_s_q8_K_generic ggml_vec_dot_iq1_s_q8_K
+#define ggml_vec_dot_iq1_m_q8_K_generic ggml_vec_dot_iq1_m_q8_K
+#define ggml_vec_dot_iq4_nl_q8_0_generic ggml_vec_dot_iq4_nl_q8_0
+#define ggml_vec_dot_iq4_xs_q8_K_generic ggml_vec_dot_iq4_xs_q8_K
+// repack.cpp
+#define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
+#define ggml_quantize_mat_q8_0_4x8_generic ggml_quantize_mat_q8_0_4x8
+#define ggml_quantize_mat_q8_K_4x8_generic ggml_quantize_mat_q8_K_4x8
+#define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0
+#define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
+#define ggml_gemv_q4_0_8x8_q8_0_generic ggml_gemv_q4_0_8x8_q8_0
+#define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
+#define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
+#define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
+#define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
+#define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0
+#define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
+#define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
+#endif
diff --git a/ggml/src/ggml-cpu/arch/arm/cpu-feats.cpp b/ggml/src/ggml-cpu/arch/arm/cpu-feats.cpp
new file mode 100644
index 000000000..67369147c
--- /dev/null
+++ b/ggml/src/ggml-cpu/arch/arm/cpu-feats.cpp
@@ -0,0 +1,94 @@
+#include "ggml-backend-impl.h"
+
+#if defined(__aarch64__)
+
+#if defined(__linux__)
+#include <sys/auxv.h>
+#elif defined(__APPLE__)
+#include <sys/sysctl.h>
+#endif
+
+#if !defined(HWCAP2_I8MM)
+#define HWCAP2_I8MM (1 << 13)
+#endif
+
+#if !defined(HWCAP2_SME)
+#define HWCAP2_SME (1 << 23)
+#endif
+
+struct aarch64_features {
+    // has_neon not needed, aarch64 has NEON guaranteed
+    bool has_dotprod     = false;
+    bool has_fp16_va     = false;
+    bool has_sve         = false;
+    bool has_sve2        = false;
+    bool has_i8mm        = false;
+    bool has_sme         = false;
+
+    aarch64_features() {
+#if defined(__linux__)
+        uint32_t hwcap = getauxval(AT_HWCAP);
+        uint32_t hwcap2 = getauxval(AT_HWCAP2);
+
+        has_dotprod = !!(hwcap & HWCAP_ASIMDDP);
+        has_fp16_va = !!(hwcap & HWCAP_FPHP);
+        has_sve     = !!(hwcap & HWCAP_SVE);
+        has_sve2    = !!(hwcap2 & HWCAP2_SVE2);
+        has_i8mm    = !!(hwcap2 & HWCAP2_I8MM);
+        has_sme     = !!(hwcap2 & HWCAP2_SME);
+#elif defined(__APPLE__)
+        int oldp = 0;
+        size_t size = sizeof(oldp);
+
+        if (sysctlbyname("hw.optional.arm.FEAT_DotProd", &oldp, &size, NULL, 0) == 0) {
+            has_dotprod = static_cast<bool>(oldp);
+        }
+
+        if (sysctlbyname("hw.optional.arm.FEAT_I8MM", &oldp, &size, NULL, 0) == 0) {
+            has_i8mm = static_cast<bool>(oldp);
+        }
+
+        if (sysctlbyname("hw.optional.arm.FEAT_SME", &oldp, &size, NULL, 0) == 0) {
+            has_sme = static_cast<bool>(oldp);
+        }
+
+        // Apple apparently does not implement SVE yet
+#endif
+    }
+};
+
+static int ggml_backend_cpu_aarch64_score() {
+    int score = 1;
+    aarch64_features af;
+
+#ifdef GGML_USE_DOTPROD
+    if (!af.has_dotprod) { return 0; }
+    score += 1<<1;
+#endif
+#ifdef GGML_USE_FP16_VECTOR_ARITHMETIC
+    if (!af.has_fp16_va) { return 0; }
+    score += 1<<2;
+#endif
+#ifdef GGML_USE_SVE
+    if (!af.has_sve) { return 0; }
+    score += 1<<3;
+#endif
+#ifdef GGML_USE_MATMUL_INT8
+    if (!af.has_i8mm) { return 0; }
+    score += 1<<4;
+#endif
+#ifdef GGML_USE_SVE2
+    if (!af.has_sve2) { return 0; }
+    score += 1<<5;
+#endif
+#ifdef GGML_USE_SME
+    if (!af.has_sme) { return 0; }
+    score += 1<<6;
+#endif
+
+    return score;
+}
+
+GGML_BACKEND_DL_SCORE_IMPL(ggml_backend_cpu_aarch64_score)
+
+# endif // defined(__aarch64__)
diff --git a/ggml/src/ggml-cpu/arch/arm/quants.c b/ggml/src/ggml-cpu/arch/arm/quants.c
new file mode 100644
index 000000000..3e2d3d03d
--- /dev/null
+++ b/ggml/src/ggml-cpu/arch/arm/quants.c
@@ -0,0 +1,4114 @@
+#define GGML_COMMON_IMPL_C
+#include "ggml-common.h"
+#include "ggml-quants.h"
+#include "ggml-impl.h"
+#include "ggml-cpu.h"
+#include "simd-mappings.h"
+
+#include "../../quants.h"
+#include "../../ggml-cpu-impl.h"
+
+#include <math.h>
+#include <string.h>
+#include <assert.h>
+#include <float.h>
+#include <stdlib.h> // for qsort
+#include <stdio.h>  // for GGML_ASSERT
+
+#define GROUP_MAX_EPS 1e-15f
+#define GROUP_MAX_EPS_IQ3_XXS 1e-8f
+#define GROUP_MAX_EPS_IQ2_S 1e-8f
+#define GROUP_MAX_EPS_IQ1_M 1e-7f
+#define GROUP_MAX_EPS_IQ1_S 1e-12f
+
+#define UNUSED GGML_UNUSED
+
+#if defined(__ARM_NEON)
+#define B1(c,s,n)  0x ## n ## c ,  0x ## n ## s
+#define B2(c,s,n) B1(c,s,n ## c), B1(c,s,n ## s)
+#define B3(c,s,n) B2(c,s,n ## c), B2(c,s,n ## s)
+#define B4(c,s,n) B3(c,s,n ## c), B3(c,s,n ## s)
+#define B5(c,s,n) B4(c,s,n ## c), B4(c,s,n ## s)
+#define B6(c,s,n) B5(c,s,n ## c), B5(c,s,n ## s)
+#define B7(c,s,n) B6(c,s,n ## c), B6(c,s,n ## s)
+#define B8(c,s  ) B7(c,s,     c), B7(c,s,     s)
+
+// precomputed tables for expanding 8bits to 8 bytes:
+static const uint64_t table_b2b_0[1 << 8] = { B8(00, 10) }; // ( b) << 4
+static const uint64_t table_b2b_1[1 << 8] = { B8(10, 00) }; // (!b) << 4
+#endif
+
+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 * GGML_RESTRICT y = vy;
+
+#if defined(__ARM_NEON)
+    for (int i = 0; i < nb; i++) {
+        float32x4_t srcv [8];
+        float32x4_t asrcv[8];
+        float32x4_t amaxv[8];
+
+        for (int j = 0; j < 8; j++) srcv[j]  = vld1q_f32(x + i*32 + 4*j);
+        for (int j = 0; j < 8; j++) asrcv[j] = vabsq_f32(srcv[j]);
+
+        for (int j = 0; j < 4; j++) amaxv[2*j] = vmaxq_f32(asrcv[2*j], asrcv[2*j+1]);
+        for (int j = 0; j < 2; j++) amaxv[4*j] = vmaxq_f32(amaxv[4*j], amaxv[4*j+2]);
+        for (int j = 0; j < 1; j++) amaxv[8*j] = vmaxq_f32(amaxv[8*j], amaxv[8*j+4]);
+
+        const float amax = vmaxvq_f32(amaxv[0]);
+
+        const float d = amax / ((1 << 7) - 1);
+        const float id = d ? 1.0f/d : 0.0f;
+
+        y[i].d = GGML_CPU_FP32_TO_FP16(d);
+
+        for (int j = 0; j < 8; j++) {
+            const float32x4_t v  = vmulq_n_f32(srcv[j], id);
+            const int32x4_t   vi = vcvtnq_s32_f32(v);
+
+            y[i].qs[4*j + 0] = vgetq_lane_s32(vi, 0);
+            y[i].qs[4*j + 1] = vgetq_lane_s32(vi, 1);
+            y[i].qs[4*j + 2] = vgetq_lane_s32(vi, 2);
+            y[i].qs[4*j + 3] = vgetq_lane_s32(vi, 3);
+        }
+    }
+#else
+    GGML_UNUSED(nb);
+    // scalar
+    quantize_row_q8_0_ref(x, y, k);
+#endif
+}
+
+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 * GGML_RESTRICT y = vy;
+#if defined(__ARM_NEON)
+    for (int i = 0; i < nb; i++) {
+        float32x4_t srcv [8];
+        float32x4_t asrcv[8];
+        float32x4_t amaxv[8];
+
+        for (int j = 0; j < 8; j++) srcv[j]  = vld1q_f32(x + i*32 + 4*j);
+        for (int j = 0; j < 8; j++) asrcv[j] = vabsq_f32(srcv[j]);
+
+        for (int j = 0; j < 4; j++) amaxv[2*j] = vmaxq_f32(asrcv[2*j], asrcv[2*j+1]);
+        for (int j = 0; j < 2; j++) amaxv[4*j] = vmaxq_f32(amaxv[4*j], amaxv[4*j+2]);
+        for (int j = 0; j < 1; j++) amaxv[8*j] = vmaxq_f32(amaxv[8*j], amaxv[8*j+4]);
+
+        const float amax = vmaxvq_f32(amaxv[0]);
+
+        const float d = amax / ((1 << 7) - 1);
+        const float id = d ? 1.0f/d : 0.0f;
+
+        y[i].d = GGML_CPU_FP32_TO_FP16(d);
+
+        int32x4_t accv = vdupq_n_s32(0);
+
+        for (int j = 0; j < 8; j++) {
+            const float32x4_t v  = vmulq_n_f32(srcv[j], id);
+            const int32x4_t   vi = vcvtnq_s32_f32(v);
+
+            y[i].qs[4*j + 0] = vgetq_lane_s32(vi, 0);
+            y[i].qs[4*j + 1] = vgetq_lane_s32(vi, 1);
+            y[i].qs[4*j + 2] = vgetq_lane_s32(vi, 2);
+            y[i].qs[4*j + 3] = vgetq_lane_s32(vi, 3);
+
+            accv = vaddq_s32(accv, vi);
+        }
+
+        y[i].s = GGML_CPU_FP32_TO_FP16(d * vaddvq_s32(accv));
+    }
+#else
+    GGML_UNUSED(nb);
+    // scalar
+    quantize_row_q8_1_ref(x, y, k);
+#endif
+}
+
+// placeholder implementation for Apple targets
+void quantize_row_q8_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k) {
+    quantize_row_q8_K_ref(x, y, k);
+}
+
+//===================================== Dot products =================================
+
+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;
+
+    assert(n % qk == 0);
+#if defined(__ARM_FEATURE_MATMUL_INT8)
+    assert((nrc == 2) || (nrc == 1));
+#else
+    assert(nrc == 1);
+#endif
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    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 * 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 * 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);
+
+            const uint8x16_t v0_0 = vld1q_u8(b_x0->qs);
+            const uint8x16_t v0_1 = vld1q_u8(b_x1->qs);
+
+            // 4-bit -> 8-bit
+            const int8x16_t v0_0l = vreinterpretq_s8_u8(vandq_u8  (v0_0, m4b));
+            const int8x16_t v0_0h = vreinterpretq_s8_u8(vshrq_n_u8(v0_0, 4));
+            const int8x16_t v0_1l = vreinterpretq_s8_u8(vandq_u8  (v0_1, m4b));
+            const int8x16_t v0_1h = vreinterpretq_s8_u8(vshrq_n_u8(v0_1, 4));
+
+            // sub 8
+            const int8x16_t x0_l = vsubq_s8(v0_0l, s8b);
+            const int8x16_t x0_h = vsubq_s8(v0_0h, s8b);
+            const int8x16_t x1_l = vsubq_s8(v0_1l, s8b);
+            const int8x16_t x1_h = vsubq_s8(v0_1h, s8b);
+
+            // load y
+            const int8x16_t y0_l = vld1q_s8(b_y0->qs);
+            const int8x16_t y0_h = vld1q_s8(b_y0->qs + 16);
+            const int8x16_t y1_l = vld1q_s8(b_y1->qs);
+            const int8x16_t y1_h = vld1q_s8(b_y1->qs + 16);
+
+            float32_t _scale[4] = {
+                GGML_CPU_FP16_TO_FP32(b_x0->d)*GGML_CPU_FP16_TO_FP32(b_y0->d),
+                GGML_CPU_FP16_TO_FP32(b_x0->d)*GGML_CPU_FP16_TO_FP32(b_y1->d),
+                GGML_CPU_FP16_TO_FP32(b_x1->d)*GGML_CPU_FP16_TO_FP32(b_y0->d),
+                GGML_CPU_FP16_TO_FP32(b_x1->d)*GGML_CPU_FP16_TO_FP32(b_y1->d)
+            };
+            float32x4_t scale = vld1q_f32(_scale);
+
+            int8x16_t l0 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(x0_l), vreinterpretq_s64_s8(x1_l)));
+            int8x16_t l1 = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(x0_l), vreinterpretq_s64_s8(x1_l)));
+
+            int8x16_t l2 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(x0_h), vreinterpretq_s64_s8(x1_h)));
+            int8x16_t l3 = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(x0_h), vreinterpretq_s64_s8(x1_h)));
+
+            int8x16_t r0 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(y0_l), vreinterpretq_s64_s8(y1_l)));
+            int8x16_t r1 = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(y0_l), vreinterpretq_s64_s8(y1_l)));
+
+            int8x16_t r2 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(y0_h), vreinterpretq_s64_s8(y1_h)));
+            int8x16_t r3 = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(y0_h), vreinterpretq_s64_s8(y1_h)));
+
+            sumv0 = vmlaq_f32(sumv0,(vcvtq_f32_s32(vmmlaq_s32((vmmlaq_s32((vmmlaq_s32((vmmlaq_s32(vdupq_n_s32(0), l0, r0)),
+                                                l1, r1)), l2, r2)), l3, r3))), scale);
+        }
+
+        float32x4_t sumv1 = vextq_f32 (sumv0, sumv0, 2);
+        float32x4_t sumv2 = vzip1q_f32(sumv0, sumv1);
+
+        vst1_f32(s,      vget_low_f32 (sumv2));
+        vst1_f32(s + bs, vget_high_f32(sumv2));
+
+        return;
+    }
+#endif
+
+    int ib = 0;
+    float sumf = 0;
+
+#if defined(__ARM_FEATURE_SVE)
+    svfloat32_t sumv0 = svdup_n_f32(0.0f);
+    svfloat32_t sumv1 = svdup_n_f32(0.0f);
+
+    const int vector_length = ggml_cpu_get_sve_cnt()*8;
+
+    // VLA Implementation using switch case
+    switch (vector_length) {
+        case 128:
+            {
+                // predicate for activating higher lanes for 4 float32 elements
+                const svbool_t ph4 = svptrue_pat_b32(SV_VL4);
+
+                for (; ib + 1 < nb; ib += 2) {
+                    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);
+                    const svuint8_t qx1r = svld1rq_u8(svptrue_b8(), x1->qs);
+
+                    // 4-bit -> 8-bit
+                    const svint8_t qx0l = svreinterpret_s8_u8(svand_n_u8_m(svptrue_b8(), qx0r, 0x0F));
+                    const svint8_t qx0h = svreinterpret_s8_u8(svlsr_n_u8_m(svptrue_b8(), qx0r, 0x04));
+                    const svint8_t qx1l = svreinterpret_s8_u8(svand_n_u8_m(svptrue_b8(), qx1r, 0x0F));
+                    const svint8_t qx1h = svreinterpret_s8_u8(svlsr_n_u8_m(svptrue_b8(), qx1r, 0x04));
+
+                    // sub 8
+                    const svint8_t qx0ls = svsub_n_s8_x(svptrue_b8(), qx0h, 8);
+                    const svint8_t qx0hs = svsub_n_s8_x(svptrue_b8(), qx0l, 8);
+                    const svint8_t qx1ls = svsub_n_s8_x(svptrue_b8(), qx1h, 8);
+                    const svint8_t qx1hs = svsub_n_s8_x(svptrue_b8(), qx1l, 8);
+
+                    // load y
+                    const svint8_t qy0h = svld1_s8(svptrue_b8(), y0->qs);
+                    const svint8_t qy0l = svld1_s8(svptrue_b8(), y0->qs + 16);
+                    const svint8_t qy1h = svld1_s8(svptrue_b8(), y1->qs);
+                    const svint8_t qy1l = svld1_s8(svptrue_b8(), y1->qs + 16);
+
+                    // dot product
+                    sumv0 = svmla_n_f32_x(ph4, sumv0, svcvt_f32_s32_x(ph4, svadd_x(ph4,
+                                    svdot_s32(svdup_n_s32(0), qx0ls, qy0l),
+                                    svdot_s32(svdup_n_s32(0), qx0hs, qy0h))), GGML_CPU_FP16_TO_FP32(x0->d)*GGML_CPU_FP16_TO_FP32(y0->d));
+                    sumv1 = svmla_n_f32_x(ph4, sumv1, svcvt_f32_s32_x(ph4, svadd_x(ph4,
+                                    svdot_s32(svdup_n_s32(0), qx1ls, qy1l),
+                                    svdot_s32(svdup_n_s32(0), qx1hs, qy1h))), GGML_CPU_FP16_TO_FP32(x1->d)*GGML_CPU_FP16_TO_FP32(y1->d));
+                }
+
+                sumf = svaddv_f32(svptrue_b32(), svadd_f32_x(svptrue_b32(), sumv0, sumv1));
+            } break;
+        case 256:
+            {
+                // predicate for activating higher lanes for 16 int8 elements
+                const svbool_t ph16 = svptrue_pat_b8(SV_VL16);
+                // predicate for activating lower lanes for  16 int8 elements
+                const svbool_t pl16 = svnot_b_z(svptrue_b8(), ph16);
+
+                for (; ib + 1 < nb; ib += 2) {
+                    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);
+                    const svuint8_t qx1r = svld1rq_u8(svptrue_b8(), x1->qs);
+
+                    // 4-bit -> 8-bit
+                    const svint8_t qx0 = svreinterpret_s8_u8(svlsr_n_u8_m(pl16, svand_n_u8_m(ph16, qx0r, 0x0F), 0x04));
+                    const svint8_t qx1 = svreinterpret_s8_u8(svlsr_n_u8_m(pl16, svand_n_u8_m(ph16, qx1r, 0x0F), 0x04));
+
+                    // sub 8
+                    const svint8_t qx0s = svsub_n_s8_x(svptrue_b8(), qx0, 8);
+                    const svint8_t qx1s = svsub_n_s8_x(svptrue_b8(), qx1, 8);
+
+                    // load y
+                    const svint8_t qy0 = svld1_s8(svptrue_b8(), y0->qs);
+                    const svint8_t qy1 = svld1_s8(svptrue_b8(), y1->qs);
+
+                    // dot product
+                    sumv0 = svmla_n_f32_x(svptrue_b32(), sumv0, svcvt_f32_s32_x(svptrue_b32(),
+                                svdot_s32(svdup_n_s32(0), qx0s, qy0)), GGML_CPU_FP16_TO_FP32(x0->d)*GGML_CPU_FP16_TO_FP32(y0->d));
+                    sumv1 = svmla_n_f32_x(svptrue_b32(), sumv1, svcvt_f32_s32_x(svptrue_b32(),
+                                svdot_s32(svdup_n_s32(0), qx1s, qy1)), GGML_CPU_FP16_TO_FP32(x1->d)*GGML_CPU_FP16_TO_FP32(y1->d));
+                }
+
+                sumf = svaddv_f32(svptrue_b32(), svadd_f32_x(svptrue_b32(), sumv0, sumv1));
+            } break;
+        case 512:
+            {
+                // predicate for activating higher lanes for 32 int8 elements
+                const svbool_t ph32 = svptrue_pat_b8(SV_VL32);
+
+                // predicate for activating higher lanes for 16 int8 elements
+                const svbool_t ph16 = svptrue_pat_b8(SV_VL16);
+                // predicate for activating lower lanes for 16 int8 elements from first 32 int8 activated lanes
+                const svbool_t pl16 = svnot_b_z(ph32, ph16);
+
+                for (; ib + 1 < nb; ib += 2) {
+                    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);
+                    const svuint8_t qx1r = svld1rq_u8(ph32, x1->qs);
+
+                    // 4-bit -> 8-bit
+                    const svint8_t qx0 = svreinterpret_s8_u8(svlsr_n_u8_m(pl16, svand_n_u8_m(ph16, qx0r, 0x0F), 0x04));
+                    const svint8_t qx1 = svreinterpret_s8_u8(svlsr_n_u8_m(pl16, svand_n_u8_m(ph16, qx1r, 0x0F), 0x04));
+
+                    // sub 8
+                    const svint8_t qx0s = svsub_n_s8_x(ph32, qx0, 8);
+                    const svint8_t qx1s = svsub_n_s8_x(ph32, qx1, 8);
+
+                    // load y
+                    const svint8_t qy0 = svld1_s8(ph32, y0->qs);
+                    const svint8_t qy1 = svld1_s8(ph32, y1->qs);
+
+                    // dot product
+                    sumv0 = svmla_n_f32_x(ph32, sumv0, svcvt_f32_s32_x(ph32,
+                                svdot_s32(svdup_n_s32(0), qx0s, qy0)), GGML_CPU_FP16_TO_FP32(x0->d)*GGML_CPU_FP16_TO_FP32(y0->d));
+                    sumv1 = svmla_n_f32_x(ph32, sumv1, svcvt_f32_s32_x(ph32,
+                                svdot_s32(svdup_n_s32(0), qx1s, qy1)), GGML_CPU_FP16_TO_FP32(x1->d)*GGML_CPU_FP16_TO_FP32(y1->d));
+                }
+
+                sumf = svaddv_f32(ph32, svadd_f32_x(ph32, sumv0, sumv1));
+            } break;
+        default:
+            assert(false && "Unsupported vector length");
+            break;
+    }
+
+#elif defined(__ARM_NEON)
+    float32x4_t sumv0 = vdupq_n_f32(0.0f);
+    float32x4_t sumv1 = vdupq_n_f32(0.0f);
+
+    for (; ib + 1 < nb; ib += 2) {
+        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);
+
+        const uint8x16_t v0_0 = vld1q_u8(x0->qs);
+        const uint8x16_t v0_1 = vld1q_u8(x1->qs);
+
+        // 4-bit -> 8-bit
+        const int8x16_t v0_0l = vreinterpretq_s8_u8(vandq_u8  (v0_0, m4b));
+        const int8x16_t v0_0h = vreinterpretq_s8_u8(vshrq_n_u8(v0_0, 4));
+        const int8x16_t v0_1l = vreinterpretq_s8_u8(vandq_u8  (v0_1, m4b));
+        const int8x16_t v0_1h = vreinterpretq_s8_u8(vshrq_n_u8(v0_1, 4));
+
+        // sub 8
+        const int8x16_t v0_0ls = vsubq_s8(v0_0l, s8b);
+        const int8x16_t v0_0hs = vsubq_s8(v0_0h, s8b);
+        const int8x16_t v0_1ls = vsubq_s8(v0_1l, s8b);
+        const int8x16_t v0_1hs = vsubq_s8(v0_1h, s8b);
+
+        // load y
+        const int8x16_t v1_0l = vld1q_s8(y0->qs);
+        const int8x16_t v1_0h = vld1q_s8(y0->qs + 16);
+        const int8x16_t v1_1l = vld1q_s8(y1->qs);
+        const int8x16_t v1_1h = vld1q_s8(y1->qs + 16);
+
+        // dot product into int32x4_t
+        const int32x4_t p_0 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), v0_0ls, v1_0l), v0_0hs, v1_0h);
+        const int32x4_t p_1 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), v0_1ls, v1_1l), v0_1hs, v1_1h);
+
+        sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(p_0), GGML_CPU_FP16_TO_FP32(x0->d)*GGML_CPU_FP16_TO_FP32(y0->d));
+        sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(p_1), GGML_CPU_FP16_TO_FP32(x1->d)*GGML_CPU_FP16_TO_FP32(y1->d));
+    }
+
+    sumf = vaddvq_f32(sumv0) + vaddvq_f32(sumv1);
+#endif
+    for (; ib < nb; ++ib) {
+        int sumi0 = 0;
+        int sumi1 = 0;
+
+        for (int j = 0; j < qk/2; ++j) {
+            const int v0 = (x[ib].qs[j] & 0x0F) - 8;
+            const int v1 = (x[ib].qs[j] >>   4) - 8;
+
+            sumi0 += (v0 * y[ib].qs[j]);
+            sumi1 += (v1 * y[ib].qs[j + qk/2]);
+        }
+
+        int sumi = sumi0 + sumi1;
+        sumf += sumi*GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d);
+    }
+
+    *s = sumf;
+}
+
+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;
+
+    assert(n % qk == 0);
+#if defined(__ARM_FEATURE_MATMUL_INT8)
+    assert((nrc == 2) || (nrc == 1));
+#else
+    assert(nrc == 1);
+#endif
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    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 * 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 * 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_CPU_FP16_TO_FP32(b_x0->m) * GGML_CPU_FP16_TO_FP32(b_y0->s),
+                GGML_CPU_FP16_TO_FP32(b_x1->m) * GGML_CPU_FP16_TO_FP32(b_y0->s),
+                GGML_CPU_FP16_TO_FP32(b_x0->m) * GGML_CPU_FP16_TO_FP32(b_y1->s),
+                GGML_CPU_FP16_TO_FP32(b_x1->m) * GGML_CPU_FP16_TO_FP32(b_y1->s)
+            };
+            summs0 = vaddq_f32(summs0, vld1q_f32(summs_t));
+
+            const uint8x16_t m4b = vdupq_n_u8(0x0F);
+
+            const uint8x16_t v0_0 = vld1q_u8(b_x0->qs);
+            const uint8x16_t v0_1 = vld1q_u8(b_x1->qs);
+
+            // 4-bit -> 8-bit
+            const int8x16_t x0_l = vreinterpretq_s8_u8(vandq_u8  (v0_0, m4b));
+            const int8x16_t x0_h = vreinterpretq_s8_u8(vshrq_n_u8(v0_0, 4));
+            const int8x16_t x1_l = vreinterpretq_s8_u8(vandq_u8  (v0_1, m4b));
+            const int8x16_t x1_h = vreinterpretq_s8_u8(vshrq_n_u8(v0_1, 4));
+
+            // load y
+            const int8x16_t y0_l = vld1q_s8(b_y0->qs);
+            const int8x16_t y0_h = vld1q_s8(b_y0->qs + 16);
+            const int8x16_t y1_l = vld1q_s8(b_y1->qs);
+            const int8x16_t y1_h = vld1q_s8(b_y1->qs + 16);
+
+            // mmla into int32x4_t
+            float32_t _scale[4] = {
+                GGML_CPU_FP16_TO_FP32(b_x0->d)*GGML_CPU_FP16_TO_FP32(b_y0->d),
+                GGML_CPU_FP16_TO_FP32(b_x0->d)*GGML_CPU_FP16_TO_FP32(b_y1->d),
+                GGML_CPU_FP16_TO_FP32(b_x1->d)*GGML_CPU_FP16_TO_FP32(b_y0->d),
+                GGML_CPU_FP16_TO_FP32(b_x1->d)*GGML_CPU_FP16_TO_FP32(b_y1->d)
+            };
+            float32x4_t scale = vld1q_f32(_scale);
+
+            int8x16_t l0 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(x0_l), vreinterpretq_s64_s8(x1_l)));
+            int8x16_t l1 = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(x0_l), vreinterpretq_s64_s8(x1_l)));
+
+            int8x16_t l2 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(x0_h), vreinterpretq_s64_s8(x1_h)));
+            int8x16_t l3 = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(x0_h), vreinterpretq_s64_s8(x1_h)));
+
+            int8x16_t r0 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(y0_l), vreinterpretq_s64_s8(y1_l)));
+            int8x16_t r1 = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(y0_l), vreinterpretq_s64_s8(y1_l)));
+
+            int8x16_t r2 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(y0_h), vreinterpretq_s64_s8(y1_h)));
+            int8x16_t r3 = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(y0_h), vreinterpretq_s64_s8(y1_h)));
+            sumv0 = vmlaq_f32(sumv0,(vcvtq_f32_s32(vmmlaq_s32((vmmlaq_s32((vmmlaq_s32((vmmlaq_s32(vdupq_n_s32(0), l0, r0)),
+                                                l1, r1)), l2, r2)), l3, r3))), scale);
+        }
+
+        float32x4_t sumv1 = vextq_f32 (sumv0, sumv0, 2);
+        float32x4_t sumv2 = vzip1q_f32(sumv0, sumv1);
+
+        sumv2 = vaddq_f32(sumv2, summs0);
+
+        vst1_f32(s,      vget_low_f32 (sumv2));
+        vst1_f32(s + bs, vget_high_f32(sumv2));
+
+        return;
+    }
+#endif
+
+    int ib = 0;
+    float sumf = 0;
+
+#if defined(__ARM_NEON)
+    float32x4_t sumv0 = vdupq_n_f32(0.0f);
+    float32x4_t sumv1 = vdupq_n_f32(0.0f);
+
+    float summs = 0;
+
+    for (; ib + 1 < nb; ib += 2) {
+        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_CPU_FP16_TO_FP32(x0->m) * GGML_CPU_FP16_TO_FP32(y0->s) + GGML_CPU_FP16_TO_FP32(x1->m) * GGML_CPU_FP16_TO_FP32(y1->s);
+
+        const uint8x16_t m4b = vdupq_n_u8(0x0F);
+
+        const uint8x16_t v0_0 = vld1q_u8(x0->qs);
+        const uint8x16_t v0_1 = vld1q_u8(x1->qs);
+
+        // 4-bit -> 8-bit
+        const int8x16_t v0_0l = vreinterpretq_s8_u8(vandq_u8  (v0_0, m4b));
+        const int8x16_t v0_0h = vreinterpretq_s8_u8(vshrq_n_u8(v0_0, 4));
+        const int8x16_t v0_1l = vreinterpretq_s8_u8(vandq_u8  (v0_1, m4b));
+        const int8x16_t v0_1h = vreinterpretq_s8_u8(vshrq_n_u8(v0_1, 4));
+
+        // load y
+        const int8x16_t v1_0l = vld1q_s8(y0->qs);
+        const int8x16_t v1_0h = vld1q_s8(y0->qs + 16);
+        const int8x16_t v1_1l = vld1q_s8(y1->qs);
+        const int8x16_t v1_1h = vld1q_s8(y1->qs + 16);
+
+        // dot product into int32x4_t
+        const int32x4_t p_0 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), v0_0l, v1_0l), v0_0h, v1_0h);
+        const int32x4_t p_1 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), v0_1l, v1_1l), v0_1h, v1_1h);
+
+        sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(p_0), GGML_CPU_FP16_TO_FP32(x0->d)*GGML_CPU_FP16_TO_FP32(y0->d));
+        sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(p_1), GGML_CPU_FP16_TO_FP32(x1->d)*GGML_CPU_FP16_TO_FP32(y1->d));
+    }
+
+    sumf = vaddvq_f32(sumv0) + vaddvq_f32(sumv1) + summs;
+
+#endif
+    for (; ib < nb; ++ib) {
+        int sumi0 = 0;
+        int sumi1 = 0;
+
+        for (int j = 0; j < qk/2; ++j) {
+            const int v0 = (x[ib].qs[j] & 0x0F);
+            const int v1 = (x[ib].qs[j] >>   4);
+
+            sumi0 += (v0 * y[ib].qs[j]);
+            sumi1 += (v1 * y[ib].qs[j + qk/2]);
+        }
+
+        int sumi = sumi0 + sumi1;
+        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
+    }
+
+    *s = sumf;
+}
+
+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;
+
+    int ib = 0;
+    float sumf = 0;
+
+    assert(n % qk == 0);
+    assert(qk == QK5_0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    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);
+    float32x4_t sumv1 = vdupq_n_f32(0.0f);
+
+    uint32_t qh0;
+    uint32_t qh1;
+
+    uint64_t tmp0[4];
+    uint64_t tmp1[4];
+
+    for (; ib + 1 < nb; ib += 2) {
+        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);
+
+        // extract the 5th bit via lookup table ((!b) << 4)
+        memcpy(&qh0, x0->qh, sizeof(qh0));
+        memcpy(&qh1, x1->qh, sizeof(qh1));
+
+        tmp0[0] = table_b2b_1[(qh0 >>  0) & 0xFF];
+        tmp0[1] = table_b2b_1[(qh0 >>  8) & 0xFF];
+        tmp0[2] = table_b2b_1[(qh0 >> 16) & 0xFF];
+        tmp0[3] = table_b2b_1[(qh0 >> 24)       ];
+
+        tmp1[0] = table_b2b_1[(qh1 >>  0) & 0xFF];
+        tmp1[1] = table_b2b_1[(qh1 >>  8) & 0xFF];
+        tmp1[2] = table_b2b_1[(qh1 >> 16) & 0xFF];
+        tmp1[3] = table_b2b_1[(qh1 >> 24)       ];
+
+        const int8x16_t qhl0 = vld1q_s8((const int8_t *)(tmp0 + 0));
+        const int8x16_t qhh0 = vld1q_s8((const int8_t *)(tmp0 + 2));
+        const int8x16_t qhl1 = vld1q_s8((const int8_t *)(tmp1 + 0));
+        const int8x16_t qhh1 = vld1q_s8((const int8_t *)(tmp1 + 2));
+
+        const uint8x16_t v0_0 = vld1q_u8(x0->qs);
+        const uint8x16_t v0_1 = vld1q_u8(x1->qs);
+
+        // 4-bit -> 8-bit
+        int8x16_t v0_0l = vreinterpretq_s8_u8(vandq_u8  (v0_0, m4b));
+        int8x16_t v0_0h = vreinterpretq_s8_u8(vshrq_n_u8(v0_0, 4));
+        int8x16_t v0_1l = vreinterpretq_s8_u8(vandq_u8  (v0_1, m4b));
+        int8x16_t v0_1h = vreinterpretq_s8_u8(vshrq_n_u8(v0_1, 4));
+
+        // add high bit and sub 16 (equivalent to sub 0x10 when bit is zero)
+        const int8x16_t v0_0lf = vsubq_s8(v0_0l, qhl0);
+        const int8x16_t v0_0hf = vsubq_s8(v0_0h, qhh0);
+        const int8x16_t v0_1lf = vsubq_s8(v0_1l, qhl1);
+        const int8x16_t v0_1hf = vsubq_s8(v0_1h, qhh1);
+
+        // load y
+        const int8x16_t v1_0l = vld1q_s8(y0->qs);
+        const int8x16_t v1_0h = vld1q_s8(y0->qs + 16);
+        const int8x16_t v1_1l = vld1q_s8(y1->qs);
+        const int8x16_t v1_1h = vld1q_s8(y1->qs + 16);
+
+        sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(
+                        ggml_vdotq_s32(vdupq_n_s32(0), v0_0lf, v1_0l),
+                        ggml_vdotq_s32(vdupq_n_s32(0), v0_0hf, v1_0h))), GGML_CPU_FP16_TO_FP32(x0->d)*GGML_CPU_FP16_TO_FP32(y0->d));
+        sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(
+                        ggml_vdotq_s32(vdupq_n_s32(0), v0_1lf, v1_1l),
+                        ggml_vdotq_s32(vdupq_n_s32(0), v0_1hf, v1_1h))), GGML_CPU_FP16_TO_FP32(x1->d)*GGML_CPU_FP16_TO_FP32(y1->d));
+    }
+
+    sumf = vaddvq_f32(sumv0) + vaddvq_f32(sumv1);
+
+#endif
+    for (; ib < nb; ++ib) {
+        uint32_t qh;
+        memcpy(&qh, x[ib].qh, sizeof(qh));
+
+        int sumi0 = 0;
+        int sumi1 = 0;
+
+        for (int j = 0; j < qk/2; ++j) {
+            const uint8_t xh_0 = ((qh & (1u << (j + 0 ))) >> (j + 0 )) << 4;
+            const uint8_t xh_1 = ((qh & (1u << (j + 16))) >> (j + 12));
+
+            const int32_t x0 = (int8_t)(((x[ib].qs[j] & 0x0F) | xh_0) - 16);
+            const int32_t x1 = (int8_t)(((x[ib].qs[j] >>   4) | xh_1) - 16);
+
+            sumi0 += (x0 * y[ib].qs[j]);
+            sumi1 += (x1 * y[ib].qs[j + qk/2]);
+        }
+
+        int sumi = sumi0 + sumi1;
+        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d)) * sumi;
+    }
+
+    *s = sumf;
+}
+
+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;
+
+    int ib = 0;
+    float sumf = 0;
+
+    assert(n % qk == 0);
+    assert(qk == QK5_1);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    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);
+    float32x4_t sumv1 = vdupq_n_f32(0.0f);
+
+    float summs0 = 0.0f;
+    float summs1 = 0.0f;
+
+    uint32_t qh0;
+    uint32_t qh1;
+
+    uint64_t tmp0[4];
+    uint64_t tmp1[4];
+
+    for (; ib + 1 < nb; ib += 2) {
+        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);
+
+        summs0 += GGML_CPU_FP16_TO_FP32(x0->m) * GGML_CPU_FP16_TO_FP32(y0->s);
+        summs1 += GGML_CPU_FP16_TO_FP32(x1->m) * GGML_CPU_FP16_TO_FP32(y1->s);
+
+        // extract the 5th bit via lookup table ((b) << 4)
+        memcpy(&qh0, x0->qh, sizeof(qh0));
+        memcpy(&qh1, x1->qh, sizeof(qh1));
+
+        tmp0[0] = table_b2b_0[(qh0 >>  0) & 0xFF];
+        tmp0[1] = table_b2b_0[(qh0 >>  8) & 0xFF];
+        tmp0[2] = table_b2b_0[(qh0 >> 16) & 0xFF];
+        tmp0[3] = table_b2b_0[(qh0 >> 24)       ];
+
+        tmp1[0] = table_b2b_0[(qh1 >>  0) & 0xFF];
+        tmp1[1] = table_b2b_0[(qh1 >>  8) & 0xFF];
+        tmp1[2] = table_b2b_0[(qh1 >> 16) & 0xFF];
+        tmp1[3] = table_b2b_0[(qh1 >> 24)       ];
+
+        const int8x16_t qhl0 = vld1q_s8((const int8_t *)(tmp0 + 0));
+        const int8x16_t qhh0 = vld1q_s8((const int8_t *)(tmp0 + 2));
+        const int8x16_t qhl1 = vld1q_s8((const int8_t *)(tmp1 + 0));
+        const int8x16_t qhh1 = vld1q_s8((const int8_t *)(tmp1 + 2));
+
+        const uint8x16_t v0_0 = vld1q_u8(x0->qs);
+        const uint8x16_t v0_1 = vld1q_u8(x1->qs);
+
+        // 4-bit -> 8-bit
+        const int8x16_t v0_0l = vreinterpretq_s8_u8(vandq_u8  (v0_0, m4b));
+        const int8x16_t v0_0h = vreinterpretq_s8_u8(vshrq_n_u8(v0_0, 4));
+        const int8x16_t v0_1l = vreinterpretq_s8_u8(vandq_u8  (v0_1, m4b));
+        const int8x16_t v0_1h = vreinterpretq_s8_u8(vshrq_n_u8(v0_1, 4));
+
+        // add high bit
+        const int8x16_t v0_0lf = vorrq_s8(v0_0l, qhl0);
+        const int8x16_t v0_0hf = vorrq_s8(v0_0h, qhh0);
+        const int8x16_t v0_1lf = vorrq_s8(v0_1l, qhl1);
+        const int8x16_t v0_1hf = vorrq_s8(v0_1h, qhh1);
+
+        // load y
+        const int8x16_t v1_0l = vld1q_s8(y0->qs);
+        const int8x16_t v1_0h = vld1q_s8(y0->qs + 16);
+        const int8x16_t v1_1l = vld1q_s8(y1->qs);
+        const int8x16_t v1_1h = vld1q_s8(y1->qs + 16);
+
+        sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(
+                        ggml_vdotq_s32(vdupq_n_s32(0), v0_0lf, v1_0l),
+                        ggml_vdotq_s32(vdupq_n_s32(0), v0_0hf, v1_0h))), GGML_CPU_FP16_TO_FP32(x0->d)*GGML_CPU_FP16_TO_FP32(y0->d));
+        sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(
+                        ggml_vdotq_s32(vdupq_n_s32(0), v0_1lf, v1_1l),
+                        ggml_vdotq_s32(vdupq_n_s32(0), v0_1hf, v1_1h))), GGML_CPU_FP16_TO_FP32(x1->d)*GGML_CPU_FP16_TO_FP32(y1->d));
+    }
+
+    sumf = vaddvq_f32(sumv0) + vaddvq_f32(sumv1) + summs0 + summs1;
+
+#endif
+    for (; ib < nb; ++ib) {
+        uint32_t qh;
+        memcpy(&qh, x[ib].qh, sizeof(qh));
+
+        int sumi0 = 0;
+        int sumi1 = 0;
+
+        for (int j = 0; j < qk/2; ++j) {
+            const uint8_t xh_0 = ((qh >> (j +  0)) << 4) & 0x10;
+            const uint8_t xh_1 = ((qh >> (j + 12))     ) & 0x10;
+
+            const int32_t x0 = (x[ib].qs[j] & 0xF) | xh_0;
+            const int32_t x1 = (x[ib].qs[j] >>  4) | xh_1;
+
+            sumi0 += (x0 * y[ib].qs[j]);
+            sumi1 += (x1 * y[ib].qs[j + qk/2]);
+        }
+
+        int sumi = sumi0 + sumi1;
+        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
+    }
+
+    *s = sumf;
+}
+
+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;
+
+    assert(n % qk == 0);
+#if defined(__ARM_FEATURE_MATMUL_INT8)
+    assert((nrc == 2) || (nrc == 1));
+#else
+    assert(nrc == 1);
+#endif
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    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 * 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 * GGML_RESTRICT b_x0 = &vx0[i];
+            const block_q8_0 * GGML_RESTRICT b_y0 = &vy0[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);
+            const int8x16_t x1_l = vld1q_s8(b_x1->qs);
+            const int8x16_t x1_h = vld1q_s8(b_x1->qs + 16);
+
+            // load y
+            const int8x16_t y0_l = vld1q_s8(b_y0->qs);
+            const int8x16_t y0_h = vld1q_s8(b_y0->qs + 16);
+            const int8x16_t y1_l = vld1q_s8(b_y1->qs);
+            const int8x16_t y1_h = vld1q_s8(b_y1->qs + 16);
+
+            float32_t _scale[4] = {
+                GGML_CPU_FP16_TO_FP32(b_x0->d)*GGML_CPU_FP16_TO_FP32(b_y0->d),
+                GGML_CPU_FP16_TO_FP32(b_x0->d)*GGML_CPU_FP16_TO_FP32(b_y1->d),
+                GGML_CPU_FP16_TO_FP32(b_x1->d)*GGML_CPU_FP16_TO_FP32(b_y0->d),
+                GGML_CPU_FP16_TO_FP32(b_x1->d)*GGML_CPU_FP16_TO_FP32(b_y1->d)
+            };
+            float32x4_t scale = vld1q_f32(_scale);
+
+            int8x16_t l0 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(x0_l), vreinterpretq_s64_s8(x1_l)));
+            int8x16_t l1 = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(x0_l), vreinterpretq_s64_s8(x1_l)));
+
+            int8x16_t l2 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(x0_h), vreinterpretq_s64_s8(x1_h)));
+            int8x16_t l3 = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(x0_h), vreinterpretq_s64_s8(x1_h)));
+
+            int8x16_t r0 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(y0_l), vreinterpretq_s64_s8(y1_l)));
+            int8x16_t r1 = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(y0_l), vreinterpretq_s64_s8(y1_l)));
+
+            int8x16_t r2 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(y0_h), vreinterpretq_s64_s8(y1_h)));
+            int8x16_t r3 = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(y0_h), vreinterpretq_s64_s8(y1_h)));
+
+            sumv0 = vmlaq_f32(sumv0,(vcvtq_f32_s32(vmmlaq_s32((vmmlaq_s32((vmmlaq_s32((vmmlaq_s32(vdupq_n_s32(0), l0, r0)),
+                                                l1, r1)), l2, r2)), l3, r3))), scale);
+        }
+
+        float32x4_t sumv1 = vextq_f32 (sumv0, sumv0, 2);
+        float32x4_t sumv2 = vzip1q_f32(sumv0, sumv1);
+
+        vst1_f32(s,      vget_low_f32 (sumv2));
+        vst1_f32(s + bs, vget_high_f32(sumv2));
+
+        return;
+    }
+#endif
+
+    int ib = 0;
+    float sumf = 0;
+
+#if defined(__ARM_FEATURE_SVE)
+    svfloat32_t sumv0 = svdup_n_f32(0.0f);
+    svfloat32_t sumv1 = svdup_n_f32(0.0f);
+
+    const int vector_length = ggml_cpu_get_sve_cnt()*8;
+
+    //VLA Implemenation for SVE
+    switch (vector_length) {
+        case 128:
+            {
+                // predicate for activating lanes for 16 Int8 elements
+                const svbool_t ph16 = svptrue_pat_b8 (SV_VL16);
+                const svbool_t pl16 = svptrue_pat_b32(SV_VL4);
+
+                for (; ib + 1 < nb; ib += 2) {
+                    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);
+                    const svint8_t qx0_1 = svld1_s8(ph16, x0->qs+16);
+                    const svint8_t qx1_0 = svld1_s8(ph16, x1->qs);
+                    const svint8_t qx1_1 = svld1_s8(ph16, x1->qs+16);
+
+                    // load y
+                    const svint8_t qy0_0 = svld1_s8(ph16, y0->qs);
+                    const svint8_t qy0_1 = svld1_s8(ph16, y0->qs+16);
+                    const svint8_t qy1_0 = svld1_s8(ph16, y1->qs);
+                    const svint8_t qy1_1 = svld1_s8(ph16, y1->qs+16);
+
+                    sumv0 = svmla_n_f32_x(pl16, sumv0, svcvt_f32_s32_x(pl16, svadd_x(pl16,
+                                    svdot_s32(svdup_n_s32(0), qx0_0, qy0_0),
+                                    svdot_s32(svdup_n_s32(0), qx0_1, qy0_1))), GGML_CPU_FP16_TO_FP32(x0->d)*GGML_CPU_FP16_TO_FP32(y0->d));
+                    sumv1 = svmla_n_f32_x(pl16, sumv1, svcvt_f32_s32_x(pl16, svadd_x(pl16,
+                                    svdot_s32(svdup_n_s32(0), qx1_0, qy1_0),
+                                    svdot_s32(svdup_n_s32(0), qx1_1, qy1_1))), GGML_CPU_FP16_TO_FP32(x1->d)*GGML_CPU_FP16_TO_FP32(y1->d));
+                }
+
+                sumf = svaddv_f32(pl16, svadd_f32_x(pl16, sumv0, sumv1));
+            } break;
+        case 256:
+            {
+                //printf("sve256");
+                for (; ib + 1 < nb; ib += 2) {
+                    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);
+                    const svint8_t qx1 = svld1_s8(svptrue_b8(), x1->qs);
+
+                    // load y
+                    const svint8_t qy0 = svld1_s8(svptrue_b8(), y0->qs);
+                    const svint8_t qy1 = svld1_s8(svptrue_b8(), y1->qs);
+
+                    sumv0 = svmla_n_f32_x(svptrue_b32(), sumv0, svcvt_f32_s32_x(svptrue_b32(),
+                                svdot_s32(svdup_n_s32(0), qx0, qy0)), GGML_CPU_FP16_TO_FP32(x0->d)*GGML_CPU_FP16_TO_FP32(y0->d));
+                    sumv1 = svmla_n_f32_x(svptrue_b32(), sumv1, svcvt_f32_s32_x(svptrue_b32(),
+                                svdot_s32(svdup_n_s32(0), qx1, qy1)), GGML_CPU_FP16_TO_FP32(x1->d)*GGML_CPU_FP16_TO_FP32(y1->d));
+                }
+
+                sumf = svaddv_f32(svptrue_b32(), svadd_f32_x(svptrue_b32(), sumv0, sumv1));
+            } break;
+        case 512:
+            {
+                // predicate for activating high 256 bit
+                const svbool_t ph32 = svptrue_pat_b8(SV_VL32);
+                // predicate for activating low 256 bit
+                const svbool_t pl32 = svnot_b_z(svptrue_b8(), ph32);
+
+                // predicate for activating high lanes for 8 float32 elements
+                const svbool_t ph8 = svptrue_pat_b32(SV_VL8);
+                // predicate for activating low lanes for 8 float32 elements
+                const svbool_t pl8 = svnot_b_z(svptrue_b32(), ph8);
+
+                svfloat32_t sumv00 = svdup_n_f32(0.0f);
+
+                for (; ib + 1 < nb; ib += 2) {
+                    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
+                    // load x
+                    const svint8_t qx_32 = svld1_s8(ph32, x0->qs);
+                          svint8_t qx_64 = svld1_s8(pl32, x0->qs + 2);
+
+                    qx_64 = svadd_s8_x(svptrue_b8(), qx_32, qx_64);
+
+                    // load y
+                    const svint8_t qy_32 = svld1_s8(ph32, y0->qs);
+                          svint8_t qy_64 = svld1_s8(pl32, y0->qs + 2);
+
+                    qy_64 = svadd_s8_x(svptrue_b8(), qy_32, qy_64);
+
+                    // scale creation
+                    const float32_t deq1 = GGML_CPU_FP16_TO_FP32(x0->d)*GGML_CPU_FP16_TO_FP32(y0->d);
+                    const float32_t deq2 = GGML_CPU_FP16_TO_FP32(x1->d)*GGML_CPU_FP16_TO_FP32(y1->d);
+
+                    // duplicate deq1 in first half of vector and deq2 in second half of vector
+                    const svfloat32_t temp = svdup_f32_m(svdup_f32_z(ph8, deq1), pl8, deq2);
+
+                    const svfloat32_t sumvt = svcvt_f32_s32_x(svptrue_b32(), svdot_s32(svdup_n_s32(0), qx_64, qy_64));
+
+                    sumv00 = svmla_f32_m(svptrue_b32(), sumv00, sumvt, temp);
+                }
+
+                sumf = svaddv_f32(svptrue_b32(), sumv00);
+                break;
+            }
+        default:
+            assert(false && "Unsupported vector length");
+            break;
+    }
+#elif defined(__ARM_NEON)
+    float32x4_t sumv0 = vdupq_n_f32(0.0f);
+    float32x4_t sumv1 = vdupq_n_f32(0.0f);
+
+    for (; ib + 1 < nb; ib += 2) {
+        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);
+        const int8x16_t x1_0 = vld1q_s8(x1->qs);
+        const int8x16_t x1_1 = vld1q_s8(x1->qs + 16);
+
+        // load y
+        const int8x16_t y0_0 = vld1q_s8(y0->qs);
+        const int8x16_t y0_1 = vld1q_s8(y0->qs + 16);
+        const int8x16_t y1_0 = vld1q_s8(y1->qs);
+        const int8x16_t y1_1 = vld1q_s8(y1->qs + 16);
+
+        sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(
+                        ggml_vdotq_s32(vdupq_n_s32(0), x0_0, y0_0),
+                        ggml_vdotq_s32(vdupq_n_s32(0), x0_1, y0_1))), GGML_CPU_FP16_TO_FP32(x0->d)*GGML_CPU_FP16_TO_FP32(y0->d));
+
+        sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(
+                        ggml_vdotq_s32(vdupq_n_s32(0), x1_0, y1_0),
+                        ggml_vdotq_s32(vdupq_n_s32(0), x1_1, y1_1))), GGML_CPU_FP16_TO_FP32(x1->d)*GGML_CPU_FP16_TO_FP32(y1->d));
+    }
+
+    sumf = vaddvq_f32(sumv0) + vaddvq_f32(sumv1);
+#endif
+    for (; ib < nb; ++ib) {
+        int sumi = 0;
+
+        for (int j = 0; j < qk; j++) {
+            sumi += x[ib].qs[j]*y[ib].qs[j];
+        }
+
+        sumf += sumi*(GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d));
+    }
+
+    *s = sumf;
+}
+
+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 * GGML_RESTRICT x = vx;
+    const block_q8_K  * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+#if defined(__ARM_NEON)
+    float sumf = 0.0f;
+
+    uint8_t k_shift[16] = {1, 1, 1, 1, 3, 3, 3, 3, 9, 9, 9, 9, 27, 27, 27, 27};
+
+    const uint8x16_t shift = vld1q_u8(k_shift);
+
+    for (int i = 0; i < nb; ++i) {
+#if defined(__ARM_FEATURE_DOTPROD)
+        int32x4_t sumi0 = vdupq_n_s32(0);
+        int32x4_t sumi1 = vdupq_n_s32(0);
+#else
+        int16x8_t sumi0 = vdupq_n_s16(0);
+        int16x8_t sumi1 = vdupq_n_s16(0);
+#endif
+
+        // first 32 bytes of 5 elements
+        {
+            uint8x16_t qx0 = vld1q_u8(x[i].qs + 0);
+            uint8x16_t qx1 = vld1q_u8(x[i].qs + 16);
+            uint8x16_t qx2 = vmulq_u8(qx0, vdupq_n_u8(3));
+            uint8x16_t qx3 = vmulq_u8(qx1, vdupq_n_u8(3));
+            uint8x16_t qx4 = vmulq_u8(qx0, vdupq_n_u8(9));
+            uint8x16_t qx5 = vmulq_u8(qx1, vdupq_n_u8(9));
+            uint8x16_t qx6 = vmulq_u8(qx0, vdupq_n_u8(27));
+            uint8x16_t qx7 = vmulq_u8(qx1, vdupq_n_u8(27));
+            uint8x16_t qx8 = vmulq_u8(qx0, vdupq_n_u8(81));
+            uint8x16_t qx9 = vmulq_u8(qx1, vdupq_n_u8(81));
+
+            // multiply by 3 and keep the 2 bits above 8 bits
+            int8x16_t sqx0 = vreinterpretq_s8_u8(vshrq_n_u8(vhaddq_u8(qx0, vshrq_n_u8(qx0, 1)), 6));
+            int8x16_t sqx1 = vreinterpretq_s8_u8(vshrq_n_u8(vhaddq_u8(qx1, vshrq_n_u8(qx1, 1)), 6));
+            int8x16_t sqx2 = vreinterpretq_s8_u8(vshrq_n_u8(vhaddq_u8(qx2, vshrq_n_u8(qx2, 1)), 6));
+            int8x16_t sqx3 = vreinterpretq_s8_u8(vshrq_n_u8(vhaddq_u8(qx3, vshrq_n_u8(qx3, 1)), 6));
+            int8x16_t sqx4 = vreinterpretq_s8_u8(vshrq_n_u8(vhaddq_u8(qx4, vshrq_n_u8(qx4, 1)), 6));
+            int8x16_t sqx5 = vreinterpretq_s8_u8(vshrq_n_u8(vhaddq_u8(qx5, vshrq_n_u8(qx5, 1)), 6));
+            int8x16_t sqx6 = vreinterpretq_s8_u8(vshrq_n_u8(vhaddq_u8(qx6, vshrq_n_u8(qx6, 1)), 6));
+            int8x16_t sqx7 = vreinterpretq_s8_u8(vshrq_n_u8(vhaddq_u8(qx7, vshrq_n_u8(qx7, 1)), 6));
+            int8x16_t sqx8 = vreinterpretq_s8_u8(vshrq_n_u8(vhaddq_u8(qx8, vshrq_n_u8(qx8, 1)), 6));
+            int8x16_t sqx9 = vreinterpretq_s8_u8(vshrq_n_u8(vhaddq_u8(qx9, vshrq_n_u8(qx9, 1)), 6));
+
+            const int8x16_t qy0 = vld1q_s8(y[i].qs +   0);
+            const int8x16_t qy1 = vld1q_s8(y[i].qs +  16);
+            const int8x16_t qy2 = vld1q_s8(y[i].qs +  32);
+            const int8x16_t qy3 = vld1q_s8(y[i].qs +  48);
+            const int8x16_t qy4 = vld1q_s8(y[i].qs +  64);
+            const int8x16_t qy5 = vld1q_s8(y[i].qs +  80);
+            const int8x16_t qy6 = vld1q_s8(y[i].qs +  96);
+            const int8x16_t qy7 = vld1q_s8(y[i].qs + 112);
+            const int8x16_t qy8 = vld1q_s8(y[i].qs + 128);
+            const int8x16_t qy9 = vld1q_s8(y[i].qs + 144);
+
+#if defined(__ARM_FEATURE_DOTPROD)
+            sumi0 = vdotq_s32(sumi0, sqx0, qy0);
+            sumi1 = vdotq_s32(sumi1, sqx1, qy1);
+            sumi0 = vdotq_s32(sumi0, sqx2, qy2);
+            sumi1 = vdotq_s32(sumi1, sqx3, qy3);
+            sumi0 = vdotq_s32(sumi0, sqx4, qy4);
+            sumi1 = vdotq_s32(sumi1, sqx5, qy5);
+            sumi0 = vdotq_s32(sumi0, sqx6, qy6);
+            sumi1 = vdotq_s32(sumi1, sqx7, qy7);
+            sumi0 = vdotq_s32(sumi0, sqx8, qy8);
+            sumi1 = vdotq_s32(sumi1, sqx9, qy9);
+#else
+            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx0), vget_low_s8(qy0));
+            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx0), vget_high_s8(qy0));
+            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx1), vget_low_s8(qy1));
+            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx1), vget_high_s8(qy1));
+            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx2), vget_low_s8(qy2));
+            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx2), vget_high_s8(qy2));
+            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx3), vget_low_s8(qy3));
+            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx3), vget_high_s8(qy3));
+            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx4), vget_low_s8(qy4));
+            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx4), vget_high_s8(qy4));
+            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx5), vget_low_s8(qy5));
+            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx5), vget_high_s8(qy5));
+            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx6), vget_low_s8(qy6));
+            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx6), vget_high_s8(qy6));
+            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx7), vget_low_s8(qy7));
+            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx7), vget_high_s8(qy7));
+            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx8), vget_low_s8(qy8));
+            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx8), vget_high_s8(qy8));
+            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx9), vget_low_s8(qy9));
+            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx9), vget_high_s8(qy9));
+#endif
+        }
+
+        // last 16 bytes of 5-element, along with the 4 bytes of 4 elements
+        {
+            uint8x16_t qx0 = vld1q_u8(x[i].qs + 32);
+            uint8x16_t qx1 = vmulq_u8(qx0, vdupq_n_u8(3));
+            uint8x16_t qx2 = vmulq_u8(qx0, vdupq_n_u8(9));
+            uint8x16_t qx3 = vmulq_u8(qx0, vdupq_n_u8(27));
+            uint8x16_t qx4 = vmulq_u8(qx0, vdupq_n_u8(81));
+            uint32_t qh;
+            memcpy(&qh, x[i].qh, sizeof(qh)); // potentially unaligned
+            uint8x16_t qx5 = vreinterpretq_u8_u32(vdupq_n_u32(qh));
+            qx5 = vmulq_u8(qx5, shift);
+
+            // multiply by 3 and keep the 2 bits above 8 bits
+            int8x16_t sqx0 = vreinterpretq_s8_u8(vshrq_n_u8(vhaddq_u8(qx0, vshrq_n_u8(qx0, 1)), 6));
+            int8x16_t sqx1 = vreinterpretq_s8_u8(vshrq_n_u8(vhaddq_u8(qx1, vshrq_n_u8(qx1, 1)), 6));
+            int8x16_t sqx2 = vreinterpretq_s8_u8(vshrq_n_u8(vhaddq_u8(qx2, vshrq_n_u8(qx2, 1)), 6));
+            int8x16_t sqx3 = vreinterpretq_s8_u8(vshrq_n_u8(vhaddq_u8(qx3, vshrq_n_u8(qx3, 1)), 6));
+            int8x16_t sqx4 = vreinterpretq_s8_u8(vshrq_n_u8(vhaddq_u8(qx4, vshrq_n_u8(qx4, 1)), 6));
+            int8x16_t sqx5 = vreinterpretq_s8_u8(vshrq_n_u8(vhaddq_u8(qx5, vshrq_n_u8(qx5, 1)), 6));
+
+            const int8x16_t qy0 = vld1q_s8(y[i].qs + 160);
+            const int8x16_t qy1 = vld1q_s8(y[i].qs + 176);
+            const int8x16_t qy2 = vld1q_s8(y[i].qs + 192);
+            const int8x16_t qy3 = vld1q_s8(y[i].qs + 208);
+            const int8x16_t qy4 = vld1q_s8(y[i].qs + 224);
+            const int8x16_t qy5 = vld1q_s8(y[i].qs + 240);
+
+#if defined(__ARM_FEATURE_DOTPROD)
+            sumi0 = vdotq_s32(sumi0, sqx0, qy0);
+            sumi1 = vdotq_s32(sumi1, sqx1, qy1);
+            sumi0 = vdotq_s32(sumi0, sqx2, qy2);
+            sumi1 = vdotq_s32(sumi1, sqx3, qy3);
+            sumi0 = vdotq_s32(sumi0, sqx4, qy4);
+            sumi1 = vdotq_s32(sumi1, sqx5, qy5);
+#else
+            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx0), vget_low_s8(qy0));
+            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx0), vget_high_s8(qy0));
+            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx1), vget_low_s8(qy1));
+            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx1), vget_high_s8(qy1));
+            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx2), vget_low_s8(qy2));
+            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx2), vget_high_s8(qy2));
+            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx3), vget_low_s8(qy3));
+            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx3), vget_high_s8(qy3));
+            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx4), vget_low_s8(qy4));
+            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx4), vget_high_s8(qy4));
+            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx5), vget_low_s8(qy5));
+            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx5), vget_high_s8(qy5));
+#endif
+        }
+
+        const int16x8_t ysum0 = vld1q_s16(y[i].bsums);
+        const int16x8_t ysum1 = vld1q_s16(y[i].bsums + 8);
+
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
+
+#if defined(__ARM_FEATURE_DOTPROD)
+        sumi0 = vaddq_s32(sumi0, sumi1);
+        sumi0 = vsubq_s32(sumi0, vpaddlq_s16(vaddq_s16(ysum0, ysum1)));
+
+        sumf += d * (float) vaddvq_s32(sumi0);
+#else
+        sumi0 = vaddq_s16(sumi0, sumi1);
+        sumi0 = vsubq_s16(sumi0, vaddq_s16(ysum0, ysum1));
+
+        sumf += d * (float) vaddlvq_s16(sumi0);
+#endif
+    }
+
+    *s = sumf;
+
+#else
+    const uint8_t pow3[6] = {1, 3, 9, 27, 81, 243};
+
+    float sumf = 0.0f;
+
+    for (int i = 0; i < nb; ++i) {
+        int sum = 0;
+
+        for (size_t j = 0; j < sizeof(x->qs) - sizeof(x->qs) % 32; j += 32) {
+            for (size_t l = 0; l < 5; ++l) {
+                for (size_t m = 0; m < 32; ++m) {
+                    uint8_t q = x[i].qs[j + m] * pow3[l];
+                    uint16_t xi = ((uint16_t) q * 3) >> 8;
+                    sum += (xi - 1) * y[i].qs[j*5 + l*32 + m];
+                }
+            }
+        }
+        for (size_t j = sizeof(x->qs) - sizeof(x->qs) % 32; j < sizeof(x->qs); j += 16) {
+            for (size_t l = 0; l < 5; ++l) {
+                for (size_t m = 0; m < 16; ++m) {
+                    uint8_t q = x[i].qs[j + m] * pow3[l];
+                    uint16_t xi = ((uint16_t) q * 3) >> 8;
+                    sum += (xi - 1) * y[i].qs[j*5 + l*16 + m];
+                }
+            }
+        }
+
+        for (size_t l = 0; l < 4; ++l) {
+            for (size_t j = 0; j < sizeof(x->qh); ++j) {
+                uint8_t q = x[i].qh[j] * pow3[l];
+                uint16_t xi = ((uint16_t) q * 3) >> 8;
+                sum += (xi - 1) * y[i].qs[sizeof(x->qs)*5 + l*sizeof(x->qh) + j];
+            }
+        }
+
+        sumf += (float) sum * (GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d);
+    }
+
+    *s = sumf;
+#endif
+}
+
+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 * GGML_RESTRICT x = vx;
+    const block_q8_K  * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+#if defined(__ARM_NEON)
+    float sumf = 0.0f;
+
+    const uint8x16_t m3 = vdupq_n_u8(3);
+
+    for (int i = 0; i < nb; ++i) {
+#if defined(__ARM_FEATURE_DOTPROD)
+        int32x4_t sumi0 = vdupq_n_s32(0);
+        int32x4_t sumi1 = vdupq_n_s32(0);
+#else
+        int16x8_t sumi0 = vdupq_n_s16(0);
+        int16x8_t sumi1 = vdupq_n_s16(0);
+#endif
+
+        for (size_t j = 0; j < sizeof(x->qs); j += 32) {
+            uint8x16_t qx0 = vld1q_u8(x[i].qs + j);
+            uint8x16_t qx1 = vld1q_u8(x[i].qs + j + 16);
+            uint8x16_t qx2 = vshrq_n_u8(qx0, 2);
+            uint8x16_t qx3 = vshrq_n_u8(qx1, 2);
+            uint8x16_t qx4 = vshrq_n_u8(qx0, 4);
+            uint8x16_t qx5 = vshrq_n_u8(qx1, 4);
+            uint8x16_t qx6 = vshrq_n_u8(qx0, 6);
+            uint8x16_t qx7 = vshrq_n_u8(qx1, 6);
+
+            int8x16_t sqx0 = vreinterpretq_s8_u8(vandq_u8(qx0, m3));
+            int8x16_t sqx1 = vreinterpretq_s8_u8(vandq_u8(qx1, m3));
+            int8x16_t sqx2 = vreinterpretq_s8_u8(vandq_u8(qx2, m3));
+            int8x16_t sqx3 = vreinterpretq_s8_u8(vandq_u8(qx3, m3));
+            int8x16_t sqx4 = vreinterpretq_s8_u8(vandq_u8(qx4, m3));
+            int8x16_t sqx5 = vreinterpretq_s8_u8(vandq_u8(qx5, m3));
+            int8x16_t sqx6 = vreinterpretq_s8_u8(vandq_u8(qx6, m3));
+            int8x16_t sqx7 = vreinterpretq_s8_u8(vandq_u8(qx7, m3));
+
+            const int8x16_t qy0 = vld1q_s8(y[i].qs + j*4 +   0);
+            const int8x16_t qy1 = vld1q_s8(y[i].qs + j*4 +  16);
+            const int8x16_t qy2 = vld1q_s8(y[i].qs + j*4 +  32);
+            const int8x16_t qy3 = vld1q_s8(y[i].qs + j*4 +  48);
+            const int8x16_t qy4 = vld1q_s8(y[i].qs + j*4 +  64);
+            const int8x16_t qy5 = vld1q_s8(y[i].qs + j*4 +  80);
+            const int8x16_t qy6 = vld1q_s8(y[i].qs + j*4 +  96);
+            const int8x16_t qy7 = vld1q_s8(y[i].qs + j*4 + 112);
+
+#if defined(__ARM_FEATURE_DOTPROD)
+            sumi0 = vdotq_s32(sumi0, sqx0, qy0);
+            sumi1 = vdotq_s32(sumi1, sqx1, qy1);
+            sumi0 = vdotq_s32(sumi0, sqx2, qy2);
+            sumi1 = vdotq_s32(sumi1, sqx3, qy3);
+            sumi0 = vdotq_s32(sumi0, sqx4, qy4);
+            sumi1 = vdotq_s32(sumi1, sqx5, qy5);
+            sumi0 = vdotq_s32(sumi0, sqx6, qy6);
+            sumi1 = vdotq_s32(sumi1, sqx7, qy7);
+#else
+            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx0), vget_low_s8(qy0));
+            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx0), vget_high_s8(qy0));
+            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx1), vget_low_s8(qy1));
+            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx1), vget_high_s8(qy1));
+            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx2), vget_low_s8(qy2));
+            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx2), vget_high_s8(qy2));
+            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx3), vget_low_s8(qy3));
+            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx3), vget_high_s8(qy3));
+            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx4), vget_low_s8(qy4));
+            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx4), vget_high_s8(qy4));
+            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx5), vget_low_s8(qy5));
+            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx5), vget_high_s8(qy5));
+            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx6), vget_low_s8(qy6));
+            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx6), vget_high_s8(qy6));
+            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx7), vget_low_s8(qy7));
+            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx7), vget_high_s8(qy7));
+#endif
+        }
+
+        const int16x8_t ysum0 = vld1q_s16(y[i].bsums);
+        const int16x8_t ysum1 = vld1q_s16(y[i].bsums + 8);
+
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
+
+#if defined(__ARM_FEATURE_DOTPROD)
+        sumi0 = vaddq_s32(sumi0, sumi1);
+        sumi0 = vsubq_s32(sumi0, vpaddlq_s16(vaddq_s16(ysum0, ysum1)));
+
+        sumf += d * (float) vaddvq_s32(sumi0);
+#else
+        sumi0 = vaddq_s16(sumi0, sumi1);
+        sumi0 = vsubq_s16(sumi0, vaddq_s16(ysum0, ysum1));
+
+        sumf += d * (float) vaddlvq_s16(sumi0);
+#endif
+    }
+
+    *s = sumf;
+
+#else
+    float sumf = 0.0f;
+
+    for (int i = 0; i < nb; ++i) {
+        int32_t sumi = 0;
+
+        for (size_t j = 0; j < sizeof(x->qs); j += 32) {
+            for (size_t l = 0; l < 4; ++l) {
+                for (size_t k = 0; k < 32; ++k) {
+                    sumi += y[i].qs[j*4 + l*32 + k] * (((x[i].qs[j + k] >> (l*2)) & 3) - 1);
+                }
+            }
+        }
+
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+
+        sumf += (float) sumi * d;
+    }
+
+    *s = sumf;
+#endif
+}
+
+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 * GGML_RESTRICT x = vx;
+    const block_q8_K * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+#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_CPU_FP16_TO_FP32(x[i].d);
+                svfloat32_t d_broad = svdup_n_f32((float32_t)d);
+                const float dmin = -y[i].d * GGML_CPU_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_CPU_FP16_TO_FP32(x[i].d);
+                svfloat32_t d_broad = svdup_n_f32((float32_t)d);
+                const float dmin = -y[i].d * GGML_CPU_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);
+
+    const int32x4_t vzero = vdupq_n_s32(0);
+
+    ggml_int8x16x2_t q2bytes;
+    uint8_t aux[16];
+
+    float sum = 0;
+
+    for (int i = 0; i < nb; ++i) {
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+        const float dmin = -y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
+
+        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);
+        vst1q_u8(aux, scales);
+
+        const uint8x16_t mins = vshrq_n_u8(mins_and_scales, 4);
+        const ggml_int16x8x2_t q8sums = ggml_vld1q_s16_x2(y[i].bsums);
+        const ggml_int16x8x2_t mins16 = {{vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(mins))), vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(mins)))}};
+        const int32x4_t s0 = vaddq_s32(vmull_s16(vget_low_s16 (mins16.val[0]), vget_low_s16 (q8sums.val[0])),
+                                       vmull_s16(vget_high_s16(mins16.val[0]), vget_high_s16(q8sums.val[0])));
+        const int32x4_t s1 = vaddq_s32(vmull_s16(vget_low_s16 (mins16.val[1]), vget_low_s16 (q8sums.val[1])),
+                                       vmull_s16(vget_high_s16(mins16.val[1]), vget_high_s16(q8sums.val[1])));
+        sum += dmin * vaddvq_s32(vaddq_s32(s0, s1));
+
+        int isum = 0;
+        int is = 0;
+
+// We use this macro instead of a function call because for some reason
+// the code runs 2-3% slower, even if the function is declared inline
+#define MULTIPLY_ACCUM_WITH_SCALE(index)\
+        isum += vaddvq_s32(ggml_vdotq_s32(vzero, q2bytes.val[0], q8bytes.val[0])) * aux[is+(index)];\
+        isum += vaddvq_s32(ggml_vdotq_s32(vzero, q2bytes.val[1], q8bytes.val[1])) * aux[is+1+(index)];
+
+#define SHIFT_MULTIPLY_ACCUM_WITH_SCALE(shift, index)\
+        q8bytes = ggml_vld1q_s8_x2(q8); q8 += 32;\
+        q2bytes.val[0] = vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q2bits.val[0], (shift)), m3));\
+        q2bytes.val[1] = vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q2bits.val[1], (shift)), m3));\
+        MULTIPLY_ACCUM_WITH_SCALE((index));
+
+        for (int j = 0; j < QK_K/128; ++j) {
+            const ggml_uint8x16x2_t q2bits = ggml_vld1q_u8_x2(q2); q2 += 32;
+
+            ggml_int8x16x2_t q8bytes = ggml_vld1q_s8_x2(q8); q8 += 32;
+            q2bytes.val[0] = vreinterpretq_s8_u8(vandq_u8(q2bits.val[0], m3));
+            q2bytes.val[1] = vreinterpretq_s8_u8(vandq_u8(q2bits.val[1], m3));
+
+            MULTIPLY_ACCUM_WITH_SCALE(0);
+
+            SHIFT_MULTIPLY_ACCUM_WITH_SCALE(2, 2);
+            SHIFT_MULTIPLY_ACCUM_WITH_SCALE(4, 4);
+            SHIFT_MULTIPLY_ACCUM_WITH_SCALE(6, 6);
+
+            is += 8;
+        }
+
+        sum += d * isum;
+    }
+
+    *s = sum;
+
+#else
+
+    float sumf = 0;
+
+    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;
+
+        int summs = 0;
+        for (int j = 0; j < 16; ++j) {
+            summs += y[i].bsums[j] * (sc[j] >> 4);
+        }
+
+        const float dall = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+        const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
+
+        int isum = 0;
+        int is = 0;
+        int d;
+        for (int k = 0; k < QK_K/128; ++k) {
+            int shift = 0;
+            for (int j = 0; j < 4; ++j) {
+                d = sc[is++] & 0xF;
+                int isuml = 0;
+                for (int l =  0; l < 16; ++l) isuml += q8[l] * ((q2[l] >> shift) & 3);
+                isum += d * isuml;
+                d = sc[is++] & 0xF;
+                isuml = 0;
+                for (int l = 16; l < 32; ++l) isuml += q8[l] * ((q2[l] >> shift) & 3);
+                isum += d * isuml;
+                shift += 2;
+                q8 += 32;
+            }
+            q2 += 32;
+        }
+        sumf += dall * isum - dmin * summs;
+    }
+    *s = sumf;
+#endif
+}
+
+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);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const uint32_t kmask1 = 0x03030303;
+    const uint32_t kmask2 = 0x0f0f0f0f;
+
+    const block_q3_K * GGML_RESTRICT x = vx;
+    const block_q8_K * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+#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_CPU_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];
+
+    const uint8x16_t m3b = vdupq_n_u8(0x3);
+    const int32x4_t  vzero = vdupq_n_s32(0);
+
+    const uint8x16_t m0 = vdupq_n_u8(1);
+    const uint8x16_t m1 = vshlq_n_u8(m0, 1);
+    const uint8x16_t m2 = vshlq_n_u8(m0, 2);
+    const uint8x16_t m3 = vshlq_n_u8(m0, 3);
+    const int8_t m32 = 32;
+
+    ggml_int8x16x4_t q3bytes;
+
+    float sum = 0;
+
+    for (int i = 0; i < nb; ++i) {
+
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+
+        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);
+
+        ggml_uint8x16x4_t q3h;
+
+        int32_t isum = 0;
+
+        // 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;
+
+        for (int j = 0; j < QK_K/128; ++j) {
+
+            const ggml_uint8x16x2_t q3bits = ggml_vld1q_u8_x2(q3); q3 += 32;
+            const ggml_int8x16x4_t q8bytes_1 = ggml_vld1q_s8_x4(q8); q8 += 64;
+            const ggml_int8x16x4_t q8bytes_2 = ggml_vld1q_s8_x4(q8); q8 += 64;
+
+            q3h.val[0] = vshlq_n_u8(vbicq_u8(m0, qhbits.val[0]), 2);
+            q3h.val[1] = vshlq_n_u8(vbicq_u8(m0, qhbits.val[1]), 2);
+            q3h.val[2] = vshlq_n_u8(vbicq_u8(m1, qhbits.val[0]), 1);
+            q3h.val[3] = vshlq_n_u8(vbicq_u8(m1, qhbits.val[1]), 1);
+
+            q3bytes.val[0] = vsubq_s8(vreinterpretq_s8_u8(vandq_u8(q3bits.val[0], m3b)), vreinterpretq_s8_u8(q3h.val[0]));
+            q3bytes.val[1] = vsubq_s8(vreinterpretq_s8_u8(vandq_u8(q3bits.val[1], m3b)), vreinterpretq_s8_u8(q3h.val[1]));
+            q3bytes.val[2] = vsubq_s8(vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q3bits.val[0], 2), m3b)), vreinterpretq_s8_u8(q3h.val[2]));
+            q3bytes.val[3] = vsubq_s8(vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q3bits.val[1], 2), m3b)), vreinterpretq_s8_u8(q3h.val[3]));
+
+            isum += vaddvq_s32(ggml_vdotq_s32(vzero, q3bytes.val[0], q8bytes_1.val[0])) * scale[0];
+            isum += vaddvq_s32(ggml_vdotq_s32(vzero, q3bytes.val[1], q8bytes_1.val[1])) * scale[1];
+            isum += vaddvq_s32(ggml_vdotq_s32(vzero, q3bytes.val[2], q8bytes_1.val[2])) * scale[2];
+            isum += vaddvq_s32(ggml_vdotq_s32(vzero, q3bytes.val[3], q8bytes_1.val[3])) * scale[3];
+
+            scale += 4;
+
+            q3h.val[0] = vbicq_u8(m2, qhbits.val[0]);
+            q3h.val[1] = vbicq_u8(m2, qhbits.val[1]);
+            q3h.val[2] = vshrq_n_u8(vbicq_u8(m3, qhbits.val[0]), 1);
+            q3h.val[3] = vshrq_n_u8(vbicq_u8(m3, qhbits.val[1]), 1);
+
+            q3bytes.val[0] = vsubq_s8(vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q3bits.val[0], 4), m3b)), vreinterpretq_s8_u8(q3h.val[0]));
+            q3bytes.val[1] = vsubq_s8(vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q3bits.val[1], 4), m3b)), vreinterpretq_s8_u8(q3h.val[1]));
+            q3bytes.val[2] = vsubq_s8(vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q3bits.val[0], 6), m3b)), vreinterpretq_s8_u8(q3h.val[2]));
+            q3bytes.val[3] = vsubq_s8(vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q3bits.val[1], 6), m3b)), vreinterpretq_s8_u8(q3h.val[3]));
+
+            isum += vaddvq_s32(ggml_vdotq_s32(vzero, q3bytes.val[0], q8bytes_2.val[0])) * scale[0];
+            isum += vaddvq_s32(ggml_vdotq_s32(vzero, q3bytes.val[1], q8bytes_2.val[1])) * scale[1];
+            isum += vaddvq_s32(ggml_vdotq_s32(vzero, q3bytes.val[2], q8bytes_2.val[2])) * scale[2];
+            isum += vaddvq_s32(ggml_vdotq_s32(vzero, q3bytes.val[3], q8bytes_2.val[3])) * scale[3];
+
+            scale += 4;
+
+            if (j == 0) {
+                qhbits.val[0] = vshrq_n_u8(qhbits.val[0], 4);
+                qhbits.val[1] = vshrq_n_u8(qhbits.val[1], 4);
+            }
+
+        }
+        sum += d * isum;
+
+    }
+
+    *s = sum;
+
+#else
+    // scalar version
+    // This function is written like this so the compiler can manage to vectorize most of it
+    // Using -Ofast, GCC and clang manage to produce code that is within a factor of 2 or so from the
+    // manually vectorized version above. Every other version I tried would run at least 4 times slower.
+    // The ideal situation would be if we could just write the code once, and the compiler would
+    // automatically produce the best possible set of machine instructions, instead of us having to manually
+    // write vectorized versions for AVX, ARM_NEON, etc.
+
+    int8_t  aux8[QK_K];
+    int16_t aux16[8];
+    float   sums [8];
+    int32_t aux32[8];
+    memset(sums, 0, 8*sizeof(float));
+
+    uint32_t auxs[4];
+    const int8_t * scales = (const int8_t*)auxs;
+
+    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;
+        memset(aux32, 0, 8*sizeof(int32_t));
+        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;
+            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
+            a += 32; m <<= 1;
+            for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 2) & 3;
+            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
+            a += 32; m <<= 1;
+            for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 4) & 3;
+            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
+            a += 32; m <<= 1;
+            for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 6) & 3;
+            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
+            a += 32; m <<= 1;
+            q3 += 32;
+        }
+        a = aux8;
+
+        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);
+        for (int j = 0; j < QK_K/16; ++j) {
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
+            q8 += 8; a += 8;
+        }
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
+        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
+    }
+    for (int l = 0; l < 8; ++l) sumf += sums[l];
+    *s = sumf;
+
+#endif
+
+}
+
+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);
+#ifdef __ARM_FEATURE_MATMUL_INT8
+    assert((nrc == 2) || (nrc == 1));
+#else
+    assert(nrc == 1);
+#endif
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q4_K * GGML_RESTRICT x = vx;
+    const block_q8_K * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+    static const uint32_t kmask1 = 0x3f3f3f3f;
+    static const uint32_t kmask2 = 0x0f0f0f0f;
+    static const uint32_t kmask3 = 0x03030303;
+
+    uint32_t utmp[4];
+
+#if defined(__ARM_FEATURE_MATMUL_INT8)
+    if (nrc == 2) {
+        const block_q4_K * GGML_RESTRICT x0 = x;
+        const block_q4_K * GGML_RESTRICT x1 = (const block_q4_K *) ((const uint8_t *)vx + bx);
+        const block_q8_K * GGML_RESTRICT y0 = y;
+        const block_q8_K * GGML_RESTRICT y1 = (const block_q8_K *) ((const uint8_t *)vy + by);
+
+        const uint8x16_t m4b = vdupq_n_u8(0x0f);
+
+        float32x4_t vfsum = vdupq_n_f32(0.0f);
+
+        for (int i = 0; i < nb; ++i, ++x0, ++x1, ++y0, ++y1) {
+            const uint8_t * GGML_RESTRICT qx0 = x0->qs;
+            const uint8_t * GGML_RESTRICT qx1 = x1->qs;
+            const  int8_t * GGML_RESTRICT qy0 = y0->qs;
+            const  int8_t * GGML_RESTRICT qy1 = y1->qs;
+
+            // decode scales and mins
+            int8_t x0_scales[8], x1_scales[8];
+            int16x8_t x0_mins, x1_mins;
+            {
+                uint32_t scales_mins[3];
+                memcpy(scales_mins, x0->scales, 12);
+                const uint32_t mins_0_3 = scales_mins[1] & kmask1;
+                const uint32_t mins_4_7 = ((scales_mins[2] >> 4) & kmask2) | (((scales_mins[1] >> 6) & kmask3) << 4);
+                const uint32x2_t mins = {mins_0_3, mins_4_7};
+                x0_mins = vreinterpretq_s16_u16(vmovl_u8(vreinterpret_u8_u32(mins)));
+                uint32_t scales[2];
+                scales[0] = scales_mins[0] & kmask1; // scales 0~3
+                scales[1] = (scales_mins[2] & kmask2) | (((scales_mins[0] >> 6) & kmask3) << 4); // scales 4~7
+                memcpy(x0_scales, scales, 8);
+            }
+            {
+                uint32_t scales_mins[3];
+                memcpy(scales_mins, x1->scales, 12);
+                const uint32_t mins_0_3 = scales_mins[1] & kmask1;
+                const uint32_t mins_4_7 = ((scales_mins[2] >> 4) & kmask2) | (((scales_mins[1] >> 6) & kmask3) << 4);
+                const uint32x2_t mins = {mins_0_3, mins_4_7};
+                x1_mins = vreinterpretq_s16_u16(vmovl_u8(vreinterpret_u8_u32(mins)));
+                uint32_t scales[2];
+                scales[0] = scales_mins[0] & kmask1; // scales 0~3
+                scales[1] = (scales_mins[2] & kmask2) | (((scales_mins[0] >> 6) & kmask3) << 4); // scales 4~7
+                memcpy(x1_scales, scales, 8);
+            }
+
+            int32x4_t visum = {0};
+
+            // process 64 data points per iteration, totally 256 data points
+            for (int j = 0; j < QK_K / 64; ++j, qx0 += 32, qx1 += 32, qy0 += 64, qy1 += 64) {
+                const int8x16x4_t vy0 = vld1q_s8_x4(qy0);
+                const int8x16x4_t vy1 = vld1q_s8_x4(qy1);
+
+                int8x16_t vx0[4], vx1[4];
+                {
+                    const uint8x16x2_t vv = vld1q_u8_x2(qx0);
+                    vx0[0] = vreinterpretq_s8_u8(vandq_u8(vv.val[0], m4b));
+                    vx0[1] = vreinterpretq_s8_u8(vandq_u8(vv.val[1], m4b));
+                    vx0[2] = vreinterpretq_s8_u8(vshrq_n_u8(vv.val[0], 4));
+                    vx0[3] = vreinterpretq_s8_u8(vshrq_n_u8(vv.val[1], 4));
+                }
+                {
+                    const uint8x16x2_t vv = vld1q_u8_x2(qx1);
+                    vx1[0] = vreinterpretq_s8_u8(vandq_u8(vv.val[0], m4b));
+                    vx1[1] = vreinterpretq_s8_u8(vandq_u8(vv.val[1], m4b));
+                    vx1[2] = vreinterpretq_s8_u8(vshrq_n_u8(vv.val[0], 4));
+                    vx1[3] = vreinterpretq_s8_u8(vshrq_n_u8(vv.val[1], 4));
+                }
+
+                // process 32 data points (share same block scale) per iteration
+                for (int k = 0; k < 2; ++k) {
+                    const int blk = j * 2 + k;
+                    const int32x4_t block_scale = {
+                        x0_scales[blk],
+                        x0_scales[blk],
+                        x1_scales[blk],
+                        x1_scales[blk],
+                    };
+
+                    int32x4_t vr = {0};
+                    for (int l = 0; l < 2; ++l) {
+                        const int idx = k * 2 + l;
+                        const int64x2_t vx0_s64 = vreinterpretq_s64_s8(vx0[idx]);
+                        const int64x2_t vx1_s64 = vreinterpretq_s64_s8(vx1[idx]);
+                        const int64x2_t vy0_s64 = vreinterpretq_s64_s8(vy0.val[idx]);
+                        const int64x2_t vy1_s64 = vreinterpretq_s64_s8(vy1.val[idx]);
+                        const int8x16_t vx_l = vreinterpretq_s8_s64(vzip1q_s64(vx0_s64, vx1_s64));
+                        const int8x16_t vx_h = vreinterpretq_s8_s64(vzip2q_s64(vx0_s64, vx1_s64));
+                        const int8x16_t vy_l = vreinterpretq_s8_s64(vzip1q_s64(vy0_s64, vy1_s64));
+                        const int8x16_t vy_h = vreinterpretq_s8_s64(vzip2q_s64(vy0_s64, vy1_s64));
+                        vr = vmmlaq_s32(vr, vx_l, vy_l);
+                        vr = vmmlaq_s32(vr, vx_h, vy_h);
+                    }
+                    // apply block scale, will NOT overflow
+                    // block_scale * sum_256(int4*int8) <= 2^(8+8+4+8) = 28 bits
+                    visum = vmlaq_s32(visum, vr, block_scale);
+                }
+            }
+
+            // adjust bias, apply superblock scale
+            {
+                int32_t bias[4];
+                // no obvious uplift from sve sdot-16, just use neon mul add
+                const int16x8_t y0_sums = vpaddq_s16(vld1q_s16(y0->bsums), vld1q_s16(y0->bsums+8));
+                const int16x8_t y1_sums = vpaddq_s16(vld1q_s16(y1->bsums), vld1q_s16(y1->bsums+8));
+                bias[0] = vaddvq_s32(vaddq_s32(vmull_s16(vget_low_s16(y0_sums), vget_low_s16(x0_mins)),
+                                               vmull_s16(vget_high_s16(y0_sums), vget_high_s16(x0_mins))));
+                bias[1] = vaddvq_s32(vaddq_s32(vmull_s16(vget_low_s16(y1_sums), vget_low_s16(x0_mins)),
+                                               vmull_s16(vget_high_s16(y1_sums), vget_high_s16(x0_mins))));
+                bias[2] = vaddvq_s32(vaddq_s32(vmull_s16(vget_low_s16(y0_sums), vget_low_s16(x1_mins)),
+                                               vmull_s16(vget_high_s16(y0_sums), vget_high_s16(x1_mins))));
+                bias[3] = vaddvq_s32(vaddq_s32(vmull_s16(vget_low_s16(y1_sums), vget_low_s16(x1_mins)),
+                                               vmull_s16(vget_high_s16(y1_sums), vget_high_s16(x1_mins))));
+                const float32x4_t dmins = {
+                    GGML_CPU_FP16_TO_FP32(x0->dmin) * y0->d,
+                    GGML_CPU_FP16_TO_FP32(x0->dmin) * y1->d,
+                    GGML_CPU_FP16_TO_FP32(x1->dmin) * y0->d,
+                    GGML_CPU_FP16_TO_FP32(x1->dmin) * y1->d,
+                };
+                vfsum = vmlsq_f32(vfsum, vcvtq_f32_s32(vld1q_s32(bias)), dmins);
+
+                const float32x4_t superblock_scale = {
+                    GGML_CPU_FP16_TO_FP32(x0->d) * y0->d,
+                    GGML_CPU_FP16_TO_FP32(x0->d) * y1->d,
+                    GGML_CPU_FP16_TO_FP32(x1->d) * y0->d,
+                    GGML_CPU_FP16_TO_FP32(x1->d) * y1->d,
+                };
+                vfsum = vmlaq_f32(vfsum, vcvtq_f32_s32(visum), superblock_scale);
+            }
+        }
+
+        // vfsum = ABCD -> ACBD
+        // AC -> s, BD -> (s+bs)
+        vfsum = vzip1q_f32(vfsum, vextq_f32(vfsum, vfsum, 2));
+        vst1_f32(s,      vget_low_f32 (vfsum));
+        vst1_f32(s + bs, vget_high_f32(vfsum));
+
+        return;
+    }
+#endif
+
+#ifdef __ARM_FEATURE_SVE
+    float sumf = 0;
+    for (int i = 0; i < nb; ++i) {
+
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+        const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
+
+        const int16x8_t q8sums = vpaddq_s16(vld1q_s16(y[i].bsums), vld1q_s16(y[i].bsums + 8));
+
+        memcpy(utmp, x[i].scales, K_SCALE_SIZE);
+
+        uint32x2_t mins8 = { 0 };
+        mins8 = vset_lane_u32(utmp[1] & kmask1, mins8, 0);
+        mins8 = vset_lane_u32(((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4), mins8, 1);
+
+        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
+        utmp[0] &= kmask1;
+
+        const int16x8_t mins = vreinterpretq_s16_u16(vmovl_u8(vreinterpret_u8_u32(mins8)));
+        const int32x4_t prod = vaddq_s32(vmull_s16(vget_low_s16 (q8sums), vget_low_s16 (mins)),
+                                         vmull_s16(vget_high_s16(q8sums), vget_high_s16(mins)));
+        sumf -= dmin * vaddvq_s32(prod);
+
+        const uint8_t * scales = (const uint8_t *)utmp;
+
+        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);
+        const svint32_t mzero = svdup_n_s32(0);
+        svint32_t sumi1 = svdup_n_s32(0);
+        svint32_t sumi1_1 = svdup_n_s32(0);
+        svint32_t sumi1_2 = svdup_n_s32(0);
+        svint32_t sumi2 = svdup_n_s32(0);
+        svint32_t sumi2_1 = svdup_n_s32(0);
+        svint32_t sumi2_2 = svdup_n_s32(0);
+        switch (vector_length) {
+            case 128:
+                {
+                    for (int j = 0; j < QK_K/64; ++j) {
+                        svint8_t q4bytes = svreinterpret_s8_u8(svand_u8_x(svptrue_b8(), svld1_u8(svptrue_b8(), q4), m4b));
+                        svint8_t q8bytes = svld1_s8(svptrue_b8(), q8); q8 += 16;
+                        sumi1_1 = svmla_n_s32_x(svptrue_b32(), sumi1_1, svdot_s32(mzero, q4bytes, q8bytes), scales[2*j+0]);
+                        q4bytes = svreinterpret_s8_u8(svand_u8_x(svptrue_b8(), svld1_u8(svptrue_b8(), q4+16), m4b));
+                        q8bytes = svld1_s8(svptrue_b8(), q8); q8 += 16;
+                        sumi1_2 = svmla_n_s32_x(svptrue_b32(), sumi1_2, svdot_s32(mzero, q4bytes, q8bytes), scales[2*j+0]);
+
+                        q4bytes = svreinterpret_s8_u8(svlsr_n_u8_x(svptrue_b8(), svld1_u8(svptrue_b8(), q4), 4));
+                        q8bytes = svld1_s8(svptrue_b8(), q8); q8 += 16;
+                        sumi2_1 = svmla_n_s32_x(svptrue_b32(), sumi2_1, svdot_s32(mzero, q4bytes, q8bytes), scales[2*j+1]);
+                        q4bytes = svreinterpret_s8_u8(svlsr_n_u8_x(svptrue_b8(), svld1_u8(svptrue_b8(), q4+16), 4));
+                        q8bytes = svld1_s8(svptrue_b8(), q8); q8 += 16;
+                        sumi2_2 = svmla_n_s32_x(svptrue_b32(), sumi2_2, svdot_s32(mzero, q4bytes, q8bytes), scales[2*j+1]);
+                        q4 += 32;
+                    }
+                    sumi1 = svadd_s32_x(svptrue_b32(), sumi1_1, sumi1_2);
+                    sumi2 = svadd_s32_x(svptrue_b32(), sumi2_1, sumi2_2);
+                    sumf += d * (svaddv_s32(svptrue_b32(), svadd_s32_x(svptrue_b32(), sumi1, sumi2)));
+                } break;
+            case 256:
+            case 512:
+                {
+                    for (int j = 0; j < QK_K/64; ++j) {
+                        const svuint8_t q4bits  = svld1_u8(svptrue_pat_b8(SV_VL32), q4); q4 += 32;
+                        svint8_t q4bytes = svreinterpret_s8_u8(svand_u8_x(svptrue_pat_b8(SV_VL32), q4bits, m4b));
+                        svint8_t q8bytes = svld1_s8(svptrue_pat_b8(SV_VL32), q8); q8 += 32;
+                        sumi1 = svmla_n_s32_x(svptrue_pat_b32(SV_VL8), sumi1, svdot_s32(mzero, q4bytes, q8bytes), scales[2*j+0]);
+
+                        q4bytes = svreinterpret_s8_u8(svlsr_n_u8_x(svptrue_pat_b8(SV_VL32), q4bits, 4));
+                        q8bytes = svld1_s8(svptrue_pat_b8(SV_VL32), q8); q8 += 32;
+                        sumi2 = svmla_n_s32_x(svptrue_pat_b32(SV_VL8), sumi2, svdot_s32(mzero, q4bytes, q8bytes), scales[2*j+1]);
+                    }
+                    sumf += d * (svaddv_s32(svptrue_pat_b32(SV_VL8), svadd_s32_x(svptrue_pat_b32(SV_VL8), sumi1, sumi2)));
+                } break;
+            default:
+                assert(false && "Unsupported vector length");
+                break;
+        }
+    }
+    *s = sumf;
+#elif defined __ARM_NEON
+    const uint8x16_t m4b = vdupq_n_u8(0xf);
+    const int32x4_t mzero = vdupq_n_s32(0);
+
+    ggml_int8x16x2_t q4bytes;
+    ggml_int8x16x2_t q8bytes;
+
+    float sumf = 0;
+
+    for (int i = 0; i < nb; ++i) {
+
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+        const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
+
+        const int16x8_t q8sums = vpaddq_s16(vld1q_s16(y[i].bsums), vld1q_s16(y[i].bsums + 8));
+
+        memcpy(utmp, x[i].scales, 12);
+
+        uint32x2_t mins8 = { 0 };
+        mins8 = vset_lane_u32(utmp[1] & kmask1, mins8, 0);
+        mins8 = vset_lane_u32(((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4), mins8, 1);
+
+        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
+        utmp[0] &= kmask1;
+
+        const int16x8_t mins = vreinterpretq_s16_u16(vmovl_u8(vreinterpret_u8_u32(mins8)));
+        const int32x4_t prod = vaddq_s32(vmull_s16(vget_low_s16 (q8sums), vget_low_s16 (mins)),
+                                         vmull_s16(vget_high_s16(q8sums), vget_high_s16(mins)));
+        sumf -= dmin * vaddvq_s32(prod);
+
+        const uint8_t * scales = (const uint8_t *)utmp;
+
+        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;
+
+        for (int j = 0; j < QK_K/64; ++j) {
+            const ggml_uint8x16x2_t q4bits = ggml_vld1q_u8_x2(q4); q4 += 32;
+
+            q8bytes = ggml_vld1q_s8_x2(q8); q8 += 32;
+            q4bytes.val[0] = vreinterpretq_s8_u8(vandq_u8  (q4bits.val[0], m4b));
+            q4bytes.val[1] = vreinterpretq_s8_u8(vandq_u8  (q4bits.val[1], m4b));
+
+            const int32x4_t p1 = ggml_vdotq_s32(ggml_vdotq_s32(mzero, q4bytes.val[0], q8bytes.val[0]), q4bytes.val[1], q8bytes.val[1]);
+            sumi1 += vaddvq_s32(p1) * scales[2*j+0];
+
+            q8bytes = ggml_vld1q_s8_x2(q8); q8 += 32;
+            q4bytes.val[0] = vreinterpretq_s8_u8(vshrq_n_u8(q4bits.val[0], 4));
+            q4bytes.val[1] = vreinterpretq_s8_u8(vshrq_n_u8(q4bits.val[1], 4));
+
+            const int32x4_t p2 = ggml_vdotq_s32(ggml_vdotq_s32(mzero, q4bytes.val[0], q8bytes.val[0]), q4bytes.val[1], q8bytes.val[1]);
+
+            sumi2 += vaddvq_s32(p2) * scales[2*j+1];
+        }
+
+        sumf += d * (sumi1 + sumi2);
+
+    }
+
+    *s = sumf;
+
+#else
+
+    const uint8_t * scales = (const uint8_t*)&utmp[0];
+    const uint8_t * mins   = (const uint8_t*)&utmp[2];
+
+    int8_t  aux8[QK_K];
+    int16_t aux16[8];
+    float   sums [8];
+    int32_t aux32[8];
+    memset(sums, 0, 8*sizeof(float));
+
+    float sumf = 0;
+    for (int i = 0; i < nb; ++i) {
+        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 * 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;
+            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l]  >> 4);
+            a += 32; q4 += 32;
+        }
+        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;
+
+        int sumi = 0;
+        for (int j = 0; j < QK_K/16; ++j) sumi += y[i].bsums[j] * mins[j/2];
+        a = aux8;
+        int is = 0;
+        for (int j = 0; j < QK_K/32; ++j) {
+            int32_t scale = scales[is++];
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+        }
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
+        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
+        const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
+        sumf -= dmin * sumi;
+    }
+    for (int l = 0; l < 8; ++l) sumf += sums[l];
+    *s = sumf;
+#endif
+}
+
+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);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q5_K * GGML_RESTRICT x = vx;
+    const block_q8_K * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+    static const uint32_t kmask1 = 0x3f3f3f3f;
+    static const uint32_t kmask2 = 0x0f0f0f0f;
+    static const uint32_t kmask3 = 0x03030303;
+
+    uint32_t utmp[4];
+
+
+#ifdef __ARM_NEON
+    const uint8x16_t m4b = vdupq_n_u8(0xf);
+    const uint8x16_t mone = vdupq_n_u8(1);
+    const uint8x16_t mtwo = vdupq_n_u8(2);
+    const int32x4_t mzero = vdupq_n_s32(0);
+
+    ggml_int8x16x4_t q5bytes;
+
+    float sumf = 0;
+
+    for (int i = 0; i < nb; ++i) {
+
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+        const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
+
+        const int16x8_t q8sums = vpaddq_s16(vld1q_s16(y[i].bsums), vld1q_s16(y[i].bsums + 8));
+
+        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 uint8x8_t mins8 = vld1_u8((const uint8_t*)utmp + 8);
+        const int16x8_t mins = vreinterpretq_s16_u16(vmovl_u8(mins8));
+        const int32x4_t prod = vaddq_s32(vmull_s16(vget_low_s16 (q8sums), vget_low_s16 (mins)),
+                                         vmull_s16(vget_high_s16(q8sums), vget_high_s16(mins)));
+        int32_t sumi_mins = vaddvq_s32(prod);
+
+        const uint8_t * scales = (const uint8_t *)utmp;
+
+        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);
+
+        ggml_uint8x16x4_t q5h;
+
+        int32_t sumi = 0;
+
+        for (int j = 0; j < QK_K/64; ++j) {
+
+            const ggml_uint8x16x2_t q5bits = ggml_vld1q_u8_x2(q5); q5 += 32;
+            const ggml_int8x16x4_t q8bytes = ggml_vld1q_s8_x4(q8); q8 += 64;
+
+            q5h.val[0] = vshlq_n_u8(vandq_u8(mone, qhbits.val[0]), 4);
+            q5h.val[1] = vshlq_n_u8(vandq_u8(mone, qhbits.val[1]), 4);
+            q5h.val[2] = vshlq_n_u8(vandq_u8(mtwo, qhbits.val[0]), 3);
+            q5h.val[3] = vshlq_n_u8(vandq_u8(mtwo, qhbits.val[1]), 3);
+            qhbits.val[0] = vshrq_n_u8(qhbits.val[0], 2);
+            qhbits.val[1] = vshrq_n_u8(qhbits.val[1], 2);
+
+            q5bytes.val[0] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(q5bits.val[0], m4b), q5h.val[0]));
+            q5bytes.val[1] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(q5bits.val[1], m4b), q5h.val[1]));
+            q5bytes.val[2] = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q5bits.val[0], 4), q5h.val[2]));
+            q5bytes.val[3] = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q5bits.val[1], 4), q5h.val[3]));
+
+            sumi += vaddvq_s32(ggml_vdotq_s32(ggml_vdotq_s32(mzero, q5bytes.val[0], q8bytes.val[0]), q5bytes.val[1], q8bytes.val[1])) * *scales++;
+            sumi += vaddvq_s32(ggml_vdotq_s32(ggml_vdotq_s32(mzero, q5bytes.val[2], q8bytes.val[2]), q5bytes.val[3], q8bytes.val[3])) * *scales++;
+        }
+
+        sumf += d * sumi - dmin * sumi_mins;
+    }
+
+    *s = sumf;
+
+#else
+
+    const uint8_t * scales = (const uint8_t*)&utmp[0];
+    const uint8_t * mins   = (const uint8_t*)&utmp[2];
+
+    int8_t  aux8[QK_K];
+    int16_t aux16[8];
+    float   sums [8];
+    int32_t aux32[8];
+    memset(sums, 0, 8*sizeof(float));
+
+    float sumf = 0;
+    for (int i = 0; i < nb; ++i) {
+        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 * 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);
+            for (int l = 0; l < 32; ++l) a[l] += (hm[l] & m ? 16 : 0);
+            a += 32; m <<= 1;
+            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l]  >> 4);
+            for (int l = 0; l < 32; ++l) a[l] += (hm[l] & m ? 16 : 0);
+            a += 32; m <<= 1;
+            q4 += 32;
+        }
+        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;
+
+        int sumi = 0;
+        for (int j = 0; j < QK_K/16; ++j) sumi += y[i].bsums[j] * mins[j/2];
+        a = aux8;
+        int is = 0;
+        for (int j = 0; j < QK_K/32; ++j) {
+            int32_t scale = scales[is++];
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+        }
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
+        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
+        const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
+        sumf -= dmin * sumi;
+    }
+    for (int l = 0; l < 8; ++l) sumf += sums[l];
+    *s = sumf;
+#endif
+}
+
+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);
+#ifdef __ARM_FEATURE_MATMUL_INT8
+    assert((nrc == 2) || (nrc == 1));
+#else
+    assert(nrc == 1);
+#endif
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q6_K * GGML_RESTRICT x = vx;
+    const block_q8_K * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+#if defined(__ARM_FEATURE_MATMUL_INT8)
+    if (nrc == 2) {
+        const block_q6_K * GGML_RESTRICT x0 = x;
+        const block_q6_K * GGML_RESTRICT x1 = (const block_q6_K *) ((const uint8_t *)vx + bx);
+        const block_q8_K * GGML_RESTRICT y0 = y;
+        const block_q8_K * GGML_RESTRICT y1 = (const block_q8_K *) ((const uint8_t *)vy + by);
+
+        float32x4_t vfsum = vdupq_n_f32(0.0f);
+
+        for (int i = 0; i < nb; ++i, ++x0, ++x1, ++y0, ++y1) {
+            const uint8_t * GGML_RESTRICT ql0 = x0->ql;
+            const uint8_t * GGML_RESTRICT ql1 = x1->ql;
+            const uint8_t * GGML_RESTRICT qh0 = x0->qh;
+            const uint8_t * GGML_RESTRICT qh1 = x1->qh;
+            const  int8_t * GGML_RESTRICT qy0 = y0->qs;
+            const  int8_t * GGML_RESTRICT qy1 = y1->qs;
+
+            const uint8x16_t mone = vdupq_n_u8(0x30);
+            const uint8x16_t  m4b = vdupq_n_u8(0x0f);
+
+            int32x4_t visum = vdupq_n_s32(0);
+
+            // process 8 blocks per iteration, totally 16 blocks
+            for (int j = 0; j < 2; ++j, qh0 += 32, ql0 += 64, qh1 += 32, ql1 += 64) {
+                int8x16_t vx0[8], vx1[8];
+
+                // de-quantize vx0[8]
+                {
+                    const uint8x16x2_t qh_bits = vld1q_u8_x2(qh0);
+                    const uint8x16x4_t ql_bits = vld1q_u8_x4(ql0);
+
+                    uint8x16_t q6h_0 = vandq_u8(mone, vshlq_n_u8(qh_bits.val[0], 4));
+                    uint8x16_t q6h_1 = vandq_u8(mone, vshlq_n_u8(qh_bits.val[1], 4));
+                    uint8x16_t q6h_2 = vandq_u8(mone, vshlq_n_u8(qh_bits.val[0], 2));
+                    uint8x16_t q6h_3 = vandq_u8(mone, vshlq_n_u8(qh_bits.val[1], 2));
+
+                    vx0[0] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(ql_bits.val[0], m4b), q6h_0));
+                    vx0[1] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(ql_bits.val[1], m4b), q6h_1));
+                    vx0[2] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(ql_bits.val[2], m4b), q6h_2));
+                    vx0[3] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(ql_bits.val[3], m4b), q6h_3));
+
+                    q6h_0 = vandq_u8(mone, qh_bits.val[0]);
+                    q6h_1 = vandq_u8(mone, qh_bits.val[1]);
+                    q6h_2 = vandq_u8(mone, vshrq_n_u8(qh_bits.val[0], 2));
+                    q6h_3 = vandq_u8(mone, vshrq_n_u8(qh_bits.val[1], 2));
+
+                    vx0[4] = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(ql_bits.val[0], 4), q6h_0));
+                    vx0[5] = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(ql_bits.val[1], 4), q6h_1));
+                    vx0[6] = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(ql_bits.val[2], 4), q6h_2));
+                    vx0[7] = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(ql_bits.val[3], 4), q6h_3));
+                }
+
+                // de-quantize vx1[8]
+                {
+                    const uint8x16x2_t qh_bits = vld1q_u8_x2(qh1);
+                    const uint8x16x4_t ql_bits = vld1q_u8_x4(ql1);
+
+                    uint8x16_t q6h_0 = vandq_u8(mone, vshlq_n_u8(qh_bits.val[0], 4));
+                    uint8x16_t q6h_1 = vandq_u8(mone, vshlq_n_u8(qh_bits.val[1], 4));
+                    uint8x16_t q6h_2 = vandq_u8(mone, vshlq_n_u8(qh_bits.val[0], 2));
+                    uint8x16_t q6h_3 = vandq_u8(mone, vshlq_n_u8(qh_bits.val[1], 2));
+
+                    vx1[0] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(ql_bits.val[0], m4b), q6h_0));
+                    vx1[1] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(ql_bits.val[1], m4b), q6h_1));
+                    vx1[2] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(ql_bits.val[2], m4b), q6h_2));
+                    vx1[3] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(ql_bits.val[3], m4b), q6h_3));
+
+                    q6h_0 = vandq_u8(mone, qh_bits.val[0]);
+                    q6h_1 = vandq_u8(mone, qh_bits.val[1]);
+                    q6h_2 = vandq_u8(mone, vshrq_n_u8(qh_bits.val[0], 2));
+                    q6h_3 = vandq_u8(mone, vshrq_n_u8(qh_bits.val[1], 2));
+
+                    vx1[4] = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(ql_bits.val[0], 4), q6h_0));
+                    vx1[5] = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(ql_bits.val[1], 4), q6h_1));
+                    vx1[6] = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(ql_bits.val[2], 4), q6h_2));
+                    vx1[7] = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(ql_bits.val[3], 4), q6h_3));
+                }
+
+                // process 16 elements (one block with same scale) per iteration
+                // - vx = concat(ql, qh) - 32
+                // - r1,r2,r3,r4 = smmla(vx, vy)
+                for (int k = 0; k < 8; ++k) {
+                    const int blk = j * 8 + k;
+
+                    const int8x16_t vy0 = vld1q_s8(qy0);
+                    const int8x16_t vy1 = vld1q_s8(qy1);
+                    qy0 += 16;
+                    qy1 += 16;
+
+                    const int32x4_t block_scale = {
+                        x0->scales[blk],
+                        x0->scales[blk],
+                        x1->scales[blk],
+                        x1->scales[blk],
+                    };
+
+                    // calculate four results at once with outer product
+                    const int8x16_t vx_l = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(vx0[k]), vreinterpretq_s64_s8(vx1[k])));
+                    const int8x16_t vx_h = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(vx0[k]), vreinterpretq_s64_s8(vx1[k])));
+                    const int8x16_t vy_l = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(vy0), vreinterpretq_s64_s8(vy1)));
+                    const int8x16_t vy_h = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(vy0), vreinterpretq_s64_s8(vy1)));
+                    int32x4_t vr = vdupq_n_s32(0);
+                    vr = vmmlaq_s32(vr, vx_l, vy_l);
+                    vr = vmmlaq_s32(vr, vx_h, vy_h);
+
+                    // apply block scale, will NOT overflow
+                    // block_scale * sum_256(int6*int8) <= 2^(8+8+6+8) = 30 bits
+                    visum = vmlaq_s32(visum, vr, block_scale);
+                }
+            }
+
+            // adjust bias, apply superblock scale
+            {
+                int32_t bias[4];
+#ifdef __ARM_FEATURE_SVE
+                const svbool_t pg16_8 = svptrue_pat_b16(SV_VL8);
+                const svbool_t pg8_8 = svptrue_pat_b8(SV_VL8);
+                const svint16_t y0_q8sums_0 = svld1_s16(pg16_8, y0->bsums);
+                const svint16_t y0_q8sums_1 = svld1_s16(pg16_8, y0->bsums + 8);
+                const svint16_t y1_q8sums_0 = svld1_s16(pg16_8, y1->bsums);
+                const svint16_t y1_q8sums_1 = svld1_s16(pg16_8, y1->bsums + 8);
+                const svint16_t x0_q6scales_0 = svunpklo_s16(svld1_s8(pg8_8, x0->scales));
+                const svint16_t x0_q6scales_1 = svunpklo_s16(svld1_s8(pg8_8, x0->scales + 8));
+                const svint16_t x1_q6scales_0 = svunpklo_s16(svld1_s8(pg8_8, x1->scales));
+                const svint16_t x1_q6scales_1 = svunpklo_s16(svld1_s8(pg8_8, x1->scales + 8));
+                const svint64_t zero = svdup_n_s64(0);
+                bias[0] = svaddv_s64(svptrue_b64(), svadd_s64_x(svptrue_b64(), svdot_s64(zero, y0_q8sums_0, x0_q6scales_0),
+                                                                               svdot_s64(zero, y0_q8sums_1, x0_q6scales_1)));
+                bias[1] = svaddv_s64(svptrue_b64(), svadd_s64_x(svptrue_b64(), svdot_s64(zero, y1_q8sums_0, x0_q6scales_0),
+                                                                               svdot_s64(zero, y1_q8sums_1, x0_q6scales_1)));
+                bias[2] = svaddv_s64(svptrue_b64(), svadd_s64_x(svptrue_b64(), svdot_s64(zero, y0_q8sums_0, x1_q6scales_0),
+                                                                               svdot_s64(zero, y0_q8sums_1, x1_q6scales_1)));
+                bias[3] = svaddv_s64(svptrue_b64(), svadd_s64_x(svptrue_b64(), svdot_s64(zero, y1_q8sums_0, x1_q6scales_0),
+                                                                               svdot_s64(zero, y1_q8sums_1, x1_q6scales_1)));
+#else
+                // NEON doesn't support int16 dot product, fallback to separated mul and add
+                const int16x8x2_t q8sums0 = vld1q_s16_x2(y0->bsums);
+                const int16x8x2_t q8sums1 = vld1q_s16_x2(y1->bsums);
+
+                int8x16_t scales_s8 = vld1q_s8(x0->scales);
+                const int16x8x2_t q6scales0 = {{vmovl_s8(vget_low_s8(scales_s8)), vmovl_s8(vget_high_s8(scales_s8))}};
+                scales_s8 = vld1q_s8(x1->scales);
+                const int16x8x2_t q6scales1 = {{vmovl_s8(vget_low_s8(scales_s8)), vmovl_s8(vget_high_s8(scales_s8))}};
+
+                int32x4_t prod;
+                prod = vaddq_s32(vaddq_s32(vmull_s16(vget_low_s16 (q8sums0.val[0]), vget_low_s16 (q6scales0.val[0])),
+                                           vmull_s16(vget_high_s16(q8sums0.val[0]), vget_high_s16(q6scales0.val[0]))),
+                                 vaddq_s32(vmull_s16(vget_low_s16 (q8sums0.val[1]), vget_low_s16 (q6scales0.val[1])),
+                                           vmull_s16(vget_high_s16(q8sums0.val[1]), vget_high_s16(q6scales0.val[1]))));
+                bias[0] = vaddvq_s32(prod);
+                prod = vaddq_s32(vaddq_s32(vmull_s16(vget_low_s16 (q8sums1.val[0]), vget_low_s16 (q6scales0.val[0])),
+                                           vmull_s16(vget_high_s16(q8sums1.val[0]), vget_high_s16(q6scales0.val[0]))),
+                                 vaddq_s32(vmull_s16(vget_low_s16 (q8sums1.val[1]), vget_low_s16 (q6scales0.val[1])),
+                                           vmull_s16(vget_high_s16(q8sums1.val[1]), vget_high_s16(q6scales0.val[1]))));
+                bias[1] = vaddvq_s32(prod);
+                prod = vaddq_s32(vaddq_s32(vmull_s16(vget_low_s16 (q8sums0.val[0]), vget_low_s16 (q6scales1.val[0])),
+                                           vmull_s16(vget_high_s16(q8sums0.val[0]), vget_high_s16(q6scales1.val[0]))),
+                                 vaddq_s32(vmull_s16(vget_low_s16 (q8sums0.val[1]), vget_low_s16 (q6scales1.val[1])),
+                                           vmull_s16(vget_high_s16(q8sums0.val[1]), vget_high_s16(q6scales1.val[1]))));
+                bias[2] = vaddvq_s32(prod);
+                prod = vaddq_s32(vaddq_s32(vmull_s16(vget_low_s16 (q8sums1.val[0]), vget_low_s16 (q6scales1.val[0])),
+                                           vmull_s16(vget_high_s16(q8sums1.val[0]), vget_high_s16(q6scales1.val[0]))),
+                                 vaddq_s32(vmull_s16(vget_low_s16 (q8sums1.val[1]), vget_low_s16 (q6scales1.val[1])),
+                                           vmull_s16(vget_high_s16(q8sums1.val[1]), vget_high_s16(q6scales1.val[1]))));
+                bias[3] = vaddvq_s32(prod);
+
+#endif
+                const int32x4_t vibias = vmulq_n_s32(vld1q_s32(bias), 32);
+
+                const float32x4_t superblock_scale = {
+                    GGML_CPU_FP16_TO_FP32(x0->d) * y0->d,
+                    GGML_CPU_FP16_TO_FP32(x0->d) * y1->d,
+                    GGML_CPU_FP16_TO_FP32(x1->d) * y0->d,
+                    GGML_CPU_FP16_TO_FP32(x1->d) * y1->d,
+                };
+
+                visum = vsubq_s32(visum, vibias);
+                vfsum = vmlaq_f32(vfsum, vcvtq_f32_s32(visum), superblock_scale);
+            }
+        }
+
+        // vfsum = ABCD -> ACBD
+        // AC -> s, BD -> (s+bs)
+        vfsum = vzip1q_f32(vfsum, vextq_f32(vfsum, vfsum, 2));
+        vst1_f32(s,      vget_low_f32 (vfsum));
+        vst1_f32(s + bs, vget_high_f32(vfsum));
+
+        return;
+    }
+#endif
+
+#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_CPU_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);
+    const int32x4_t  vzero = vdupq_n_s32(0);
+    //const int8x16_t  m32s = vdupq_n_s8(32);
+
+    const uint8x16_t mone = vdupq_n_u8(3);
+
+    ggml_int8x16x4_t q6bytes;
+    ggml_uint8x16x4_t q6h;
+
+    for (int i = 0; i < nb; ++i) {
+
+        const float d_all = GGML_CPU_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 ggml_int16x8x2_t q8sums = ggml_vld1q_s16_x2(y[i].bsums);
+        const int8x16_t scales = vld1q_s8(scale);
+        const ggml_int16x8x2_t q6scales = {{vmovl_s8(vget_low_s8(scales)), vmovl_s8(vget_high_s8(scales))}};
+
+        const int32x4_t prod = vaddq_s32(vaddq_s32(vmull_s16(vget_low_s16 (q8sums.val[0]), vget_low_s16 (q6scales.val[0])),
+                                                   vmull_s16(vget_high_s16(q8sums.val[0]), vget_high_s16(q6scales.val[0]))),
+                                         vaddq_s32(vmull_s16(vget_low_s16 (q8sums.val[1]), vget_low_s16 (q6scales.val[1])),
+                                                   vmull_s16(vget_high_s16(q8sums.val[1]), vget_high_s16(q6scales.val[1]))));
+        int32_t isum_mins = vaddvq_s32(prod);
+
+        int32_t isum = 0;
+
+        for (int j = 0; j < QK_K/128; ++j) {
+
+            ggml_uint8x16x2_t qhbits = ggml_vld1q_u8_x2(qh); qh += 32;
+            ggml_uint8x16x4_t q6bits = ggml_vld1q_u8_x4(q6); q6 += 64;
+            ggml_int8x16x4_t q8bytes = ggml_vld1q_s8_x4(q8); q8 += 64;
+
+            q6h.val[0] = vshlq_n_u8(vandq_u8(mone, qhbits.val[0]), 4);
+            q6h.val[1] = vshlq_n_u8(vandq_u8(mone, qhbits.val[1]), 4);
+            uint8x16_t shifted = vshrq_n_u8(qhbits.val[0], 2);
+            q6h.val[2] = vshlq_n_u8(vandq_u8(mone, shifted), 4);
+            shifted = vshrq_n_u8(qhbits.val[1], 2);
+            q6h.val[3] = vshlq_n_u8(vandq_u8(mone, shifted), 4);
+
+            //q6bytes.val[0] = vsubq_s8(vreinterpretq_s8_u8(vorrq_u8(vandq_u8(q6bits.val[0], m4b), q6h.val[0])), m32s);
+            //q6bytes.val[1] = vsubq_s8(vreinterpretq_s8_u8(vorrq_u8(vandq_u8(q6bits.val[1], m4b), q6h.val[1])), m32s);
+            //q6bytes.val[2] = vsubq_s8(vreinterpretq_s8_u8(vorrq_u8(vandq_u8(q6bits.val[2], m4b), q6h.val[2])), m32s);
+            //q6bytes.val[3] = vsubq_s8(vreinterpretq_s8_u8(vorrq_u8(vandq_u8(q6bits.val[3], m4b), q6h.val[3])), m32s);
+            q6bytes.val[0] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(q6bits.val[0], m4b), q6h.val[0]));
+            q6bytes.val[1] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(q6bits.val[1], m4b), q6h.val[1]));
+            q6bytes.val[2] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(q6bits.val[2], m4b), q6h.val[2]));
+            q6bytes.val[3] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(q6bits.val[3], m4b), q6h.val[3]));
+
+            isum += vaddvq_s32(ggml_vdotq_s32(vzero, q6bytes.val[0], q8bytes.val[0])) * scale[0] +
+                    vaddvq_s32(ggml_vdotq_s32(vzero, q6bytes.val[1], q8bytes.val[1])) * scale[1] +
+                    vaddvq_s32(ggml_vdotq_s32(vzero, q6bytes.val[2], q8bytes.val[2])) * scale[2] +
+                    vaddvq_s32(ggml_vdotq_s32(vzero, q6bytes.val[3], q8bytes.val[3])) * scale[3];
+
+            scale += 4;
+
+            q8bytes = ggml_vld1q_s8_x4(q8); q8 += 64;
+
+            shifted = vshrq_n_u8(qhbits.val[0], 4);
+            q6h.val[0] = vshlq_n_u8(vandq_u8(mone, shifted), 4);
+            shifted = vshrq_n_u8(qhbits.val[1], 4);
+            q6h.val[1] = vshlq_n_u8(vandq_u8(mone, shifted), 4);
+            shifted = vshrq_n_u8(qhbits.val[0], 6);
+            q6h.val[2] = vshlq_n_u8(vandq_u8(mone, shifted), 4);
+            shifted = vshrq_n_u8(qhbits.val[1], 6);
+            q6h.val[3] = vshlq_n_u8(vandq_u8(mone, shifted), 4);
+
+            //q6bytes.val[0] = vsubq_s8(vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6bits.val[0], 4), q6h.val[0])), m32s);
+            //q6bytes.val[1] = vsubq_s8(vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6bits.val[1], 4), q6h.val[1])), m32s);
+            //q6bytes.val[2] = vsubq_s8(vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6bits.val[2], 4), q6h.val[2])), m32s);
+            //q6bytes.val[3] = vsubq_s8(vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6bits.val[3], 4), q6h.val[3])), m32s);
+            q6bytes.val[0] = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6bits.val[0], 4), q6h.val[0]));
+            q6bytes.val[1] = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6bits.val[1], 4), q6h.val[1]));
+            q6bytes.val[2] = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6bits.val[2], 4), q6h.val[2]));
+            q6bytes.val[3] = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6bits.val[3], 4), q6h.val[3]));
+
+            isum += vaddvq_s32(ggml_vdotq_s32(vzero, q6bytes.val[0], q8bytes.val[0])) * scale[0] +
+                    vaddvq_s32(ggml_vdotq_s32(vzero, q6bytes.val[1], q8bytes.val[1])) * scale[1] +
+                    vaddvq_s32(ggml_vdotq_s32(vzero, q6bytes.val[2], q8bytes.val[2])) * scale[2] +
+                    vaddvq_s32(ggml_vdotq_s32(vzero, q6bytes.val[3], q8bytes.val[3])) * scale[3];
+            scale += 4;
+        }
+        //sum += isum * d_all * y[i].d;
+        sum += d_all * y[i].d * (isum - 32 * isum_mins);
+
+    }
+    *s = sum;
+#else
+
+    int8_t  aux8[QK_K];
+    int16_t aux16[8];
+    float   sums [8];
+    int32_t aux32[8];
+    memset(sums, 0, 8*sizeof(float));
+
+    float sumf = 0;
+    for (int i = 0; i < nb; ++i) {
+        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 * 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;
+                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;
+        }
+        a = aux8;
+        int is = 0;
+        for (int j = 0; j < QK_K/16; ++j) {
+            int scale = x[i].scales[is++];
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+        }
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
+        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
+    }
+    for (int l = 0; l < 8; ++l) sumf += sums[l];
+    *s = sumf;
+#endif
+}
+
+#if defined (__ARM_NEON)
+static const int8_t keven_signs_q2xs[1024] = {
+     1,  1,  1,  1,  1,  1,  1,  1, -1,  1,  1,  1,  1,  1,  1, -1,  1, -1,  1,  1,  1,  1,  1, -1, -1, -1,  1,  1,  1,  1,  1,  1,
+     1,  1, -1,  1,  1,  1,  1, -1, -1,  1, -1,  1,  1,  1,  1,  1,  1, -1, -1,  1,  1,  1,  1,  1, -1, -1, -1,  1,  1,  1,  1, -1,
+     1,  1,  1, -1,  1,  1,  1, -1, -1,  1,  1, -1,  1,  1,  1,  1,  1, -1,  1, -1,  1,  1,  1,  1, -1, -1,  1, -1,  1,  1,  1, -1,
+     1,  1, -1, -1,  1,  1,  1,  1, -1,  1, -1, -1,  1,  1,  1, -1,  1, -1, -1, -1,  1,  1,  1, -1, -1, -1, -1, -1,  1,  1,  1,  1,
+     1,  1,  1,  1, -1,  1,  1, -1, -1,  1,  1,  1, -1,  1,  1,  1,  1, -1,  1,  1, -1,  1,  1,  1, -1, -1,  1,  1, -1,  1,  1, -1,
+     1,  1, -1,  1, -1,  1,  1,  1, -1,  1, -1,  1, -1,  1,  1, -1,  1, -1, -1,  1, -1,  1,  1, -1, -1, -1, -1,  1, -1,  1,  1,  1,
+     1,  1,  1, -1, -1,  1,  1,  1, -1,  1,  1, -1, -1,  1,  1, -1,  1, -1,  1, -1, -1,  1,  1, -1, -1, -1,  1, -1, -1,  1,  1,  1,
+     1,  1, -1, -1, -1,  1,  1, -1, -1,  1, -1, -1, -1,  1,  1,  1,  1, -1, -1, -1, -1,  1,  1,  1, -1, -1, -1, -1, -1,  1,  1, -1,
+     1,  1,  1,  1,  1, -1,  1, -1, -1,  1,  1,  1,  1, -1,  1,  1,  1, -1,  1,  1,  1, -1,  1,  1, -1, -1,  1,  1,  1, -1,  1, -1,
+     1,  1, -1,  1,  1, -1,  1,  1, -1,  1, -1,  1,  1, -1,  1, -1,  1, -1, -1,  1,  1, -1,  1, -1, -1, -1, -1,  1,  1, -1,  1,  1,
+     1,  1,  1, -1,  1, -1,  1,  1, -1,  1,  1, -1,  1, -1,  1, -1,  1, -1,  1, -1,  1, -1,  1, -1, -1, -1,  1, -1,  1, -1,  1,  1,
+     1,  1, -1, -1,  1, -1,  1, -1, -1,  1, -1, -1,  1, -1,  1,  1,  1, -1, -1, -1,  1, -1,  1,  1, -1, -1, -1, -1,  1, -1,  1, -1,
+     1,  1,  1,  1, -1, -1,  1,  1, -1,  1,  1,  1, -1, -1,  1, -1,  1, -1,  1,  1, -1, -1,  1, -1, -1, -1,  1,  1, -1, -1,  1,  1,
+     1,  1, -1,  1, -1, -1,  1, -1, -1,  1, -1,  1, -1, -1,  1,  1,  1, -1, -1,  1, -1, -1,  1,  1, -1, -1, -1,  1, -1, -1,  1, -1,
+     1,  1,  1, -1, -1, -1,  1, -1, -1,  1,  1, -1, -1, -1,  1,  1,  1, -1,  1, -1, -1, -1,  1,  1, -1, -1,  1, -1, -1, -1,  1, -1,
+     1,  1, -1, -1, -1, -1,  1,  1, -1,  1, -1, -1, -1, -1,  1, -1,  1, -1, -1, -1, -1, -1,  1, -1, -1, -1, -1, -1, -1, -1,  1,  1,
+     1,  1,  1,  1,  1,  1, -1, -1, -1,  1,  1,  1,  1,  1, -1,  1,  1, -1,  1,  1,  1,  1, -1,  1, -1, -1,  1,  1,  1,  1, -1, -1,
+     1,  1, -1,  1,  1,  1, -1,  1, -1,  1, -1,  1,  1,  1, -1, -1,  1, -1, -1,  1,  1,  1, -1, -1, -1, -1, -1,  1,  1,  1, -1,  1,
+     1,  1,  1, -1,  1,  1, -1,  1, -1,  1,  1, -1,  1,  1, -1, -1,  1, -1,  1, -1,  1,  1, -1, -1, -1, -1,  1, -1,  1,  1, -1,  1,
+     1,  1, -1, -1,  1,  1, -1, -1, -1,  1, -1, -1,  1,  1, -1,  1,  1, -1, -1, -1,  1,  1, -1,  1, -1, -1, -1, -1,  1,  1, -1, -1,
+     1,  1,  1,  1, -1,  1, -1,  1, -1,  1,  1,  1, -1,  1, -1, -1,  1, -1,  1,  1, -1,  1, -1, -1, -1, -1,  1,  1, -1,  1, -1,  1,
+     1,  1, -1,  1, -1,  1, -1, -1, -1,  1, -1,  1, -1,  1, -1,  1,  1, -1, -1,  1, -1,  1, -1,  1, -1, -1, -1,  1, -1,  1, -1, -1,
+     1,  1,  1, -1, -1,  1, -1, -1, -1,  1,  1, -1, -1,  1, -1,  1,  1, -1,  1, -1, -1,  1, -1,  1, -1, -1,  1, -1, -1,  1, -1, -1,
+     1,  1, -1, -1, -1,  1, -1,  1, -1,  1, -1, -1, -1,  1, -1, -1,  1, -1, -1, -1, -1,  1, -1, -1, -1, -1, -1, -1, -1,  1, -1,  1,
+     1,  1,  1,  1,  1, -1, -1,  1, -1,  1,  1,  1,  1, -1, -1, -1,  1, -1,  1,  1,  1, -1, -1, -1, -1, -1,  1,  1,  1, -1, -1,  1,
+     1,  1, -1,  1,  1, -1, -1, -1, -1,  1, -1,  1,  1, -1, -1,  1,  1, -1, -1,  1,  1, -1, -1,  1, -1, -1, -1,  1,  1, -1, -1, -1,
+     1,  1,  1, -1,  1, -1, -1, -1, -1,  1,  1, -1,  1, -1, -1,  1,  1, -1,  1, -1,  1, -1, -1,  1, -1, -1,  1, -1,  1, -1, -1, -1,
+     1,  1, -1, -1,  1, -1, -1,  1, -1,  1, -1, -1,  1, -1, -1, -1,  1, -1, -1, -1,  1, -1, -1, -1, -1, -1, -1, -1,  1, -1, -1,  1,
+     1,  1,  1,  1, -1, -1, -1, -1, -1,  1,  1,  1, -1, -1, -1,  1,  1, -1,  1,  1, -1, -1, -1,  1, -1, -1,  1,  1, -1, -1, -1, -1,
+     1,  1, -1,  1, -1, -1, -1,  1, -1,  1, -1,  1, -1, -1, -1, -1,  1, -1, -1,  1, -1, -1, -1, -1, -1, -1, -1,  1, -1, -1, -1,  1,
+     1,  1,  1, -1, -1, -1, -1,  1, -1,  1,  1, -1, -1, -1, -1, -1,  1, -1,  1, -1, -1, -1, -1, -1, -1, -1,  1, -1, -1, -1, -1,  1,
+     1,  1, -1, -1, -1, -1, -1, -1, -1,  1, -1, -1, -1, -1, -1,  1,  1, -1, -1, -1, -1, -1, -1,  1, -1, -1, -1, -1, -1, -1, -1, -1,
+};
+#endif
+
+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);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_iq2_xxs * GGML_RESTRICT x = vx;
+    const block_q8_K    * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+#if defined(__ARM_NEON)
+
+    const uint64_t * signs64 = (const uint64_t *)keven_signs_q2xs;
+
+    uint32_t aux32[4];
+    const uint8_t * aux8 = (const uint8_t *)aux32;
+
+    ggml_int8x16x4_t q2u;
+    ggml_int8x16x4_t q2s;
+    ggml_int8x16x4_t q8b;
+
+    float sumf = 0;
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_CPU_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; ib32 += 2) {
+            q8b = ggml_vld1q_s8_x4(q8); q8 += 64;
+            memcpy(aux32, q2, 4*sizeof(uint32_t)); q2 += 8;
+            q2u.val[0] = vcombine_s8(vld1_s8((const void *)(iq2xxs_grid + aux8[ 0])), vld1_s8((const void *)(iq2xxs_grid + aux8[ 1])));
+            q2u.val[1] = vcombine_s8(vld1_s8((const void *)(iq2xxs_grid + aux8[ 2])), vld1_s8((const void *)(iq2xxs_grid + aux8[ 3])));
+            q2u.val[2] = vcombine_s8(vld1_s8((const void *)(iq2xxs_grid + aux8[ 8])), vld1_s8((const void *)(iq2xxs_grid + aux8[ 9])));
+            q2u.val[3] = vcombine_s8(vld1_s8((const void *)(iq2xxs_grid + aux8[10])), vld1_s8((const void *)(iq2xxs_grid + aux8[11])));
+            q2s.val[0] = vcombine_s8(vld1_s8((const void *)(signs64 + ((aux32[1] >>  0) & 127))), vld1_s8((const void *)(signs64 + ((aux32[1] >>  7) & 127))));
+            q2s.val[1] = vcombine_s8(vld1_s8((const void *)(signs64 + ((aux32[1] >> 14) & 127))), vld1_s8((const void *)(signs64 + ((aux32[1] >> 21) & 127))));
+            q2s.val[2] = vcombine_s8(vld1_s8((const void *)(signs64 + ((aux32[3] >>  0) & 127))), vld1_s8((const void *)(signs64 + ((aux32[3] >>  7) & 127))));
+            q2s.val[3] = vcombine_s8(vld1_s8((const void *)(signs64 + ((aux32[3] >> 14) & 127))), vld1_s8((const void *)(signs64 + ((aux32[3] >> 21) & 127))));
+            q2u.val[0] = vmulq_s8(q2u.val[0], q2s.val[0]);
+            q2u.val[1] = vmulq_s8(q2u.val[1], q2s.val[1]);
+            q2u.val[2] = vmulq_s8(q2u.val[2], q2s.val[2]);
+            q2u.val[3] = vmulq_s8(q2u.val[3], q2s.val[3]);
+            const int32x4_t p1 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), q2u.val[0], q8b.val[0]), q2u.val[1], q8b.val[1]);
+            const int32x4_t p2 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), q2u.val[2], q8b.val[2]), q2u.val[3], q8b.val[3]);
+            sumf1 += vaddvq_s32(p1) * (0.5f + (aux32[1] >> 28));
+            sumf2 += vaddvq_s32(p2) * (0.5f + (aux32[3] >> 28));
+        }
+        sumf += d*(sumf1 + sumf2);
+    }
+    *s = 0.25f * sumf;
+
+#else
+
+    uint32_t aux32[2];
+    const uint8_t * aux8 = (const uint8_t *)aux32;
+
+    float sumf = 0.f;
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_CPU_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;
+        int32_t bsum = 0;
+        for (int ib32 = 0; ib32 < QK_K/32; ++ib32) {
+            memcpy(aux32, q2, 2*sizeof(uint32_t));
+            q2 += 4;
+            const uint32_t ls = 2*(aux32[1] >> 28) + 1;
+            int32_t sumi = 0;
+            for (int l = 0; l < 4; ++l) {
+                const uint8_t * grid = (const uint8_t *)(iq2xxs_grid + aux8[l]);
+                const uint8_t  signs = ksigns_iq2xs[(aux32[1] >> 7*l) & 127];
+                for (int j = 0; j < 8; ++j) {
+                    sumi += grid[j] * q8[j] * (signs & kmask_iq2xs[j] ? -1 : 1);
+                }
+                q8 += 8;
+            }
+            bsum += sumi * ls;
+        }
+        sumf += d * bsum;
+    }
+    *s = 0.125f * sumf;
+#endif
+}
+
+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);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_iq2_xs * GGML_RESTRICT x = vx;
+    const block_q8_K   * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+#if defined(__ARM_NEON)
+
+    const uint64_t * signs64 = (const uint64_t *)keven_signs_q2xs;
+
+    ggml_int8x16x4_t q2u;
+    ggml_int8x16x4_t q2s;
+    ggml_int8x16x4_t q8b;
+
+    int32x4x4_t scales32;
+
+    float sumf = 0;
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_CPU_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;
+        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);
+        uint8x16_t scales = vcombine_u8(vzip1_u8(scales_l, scales_h), vzip2_u8(scales_l, scales_h));
+        scales = vaddq_u8(vshlq_n_u8(scales, 1), vdupq_n_u8(1));
+        const uint16x8_t scales1 = vmovl_u8(vget_low_u8(scales));
+        const uint16x8_t scales2 = vmovl_u8(vget_high_u8(scales));
+        scales32.val[0] = vreinterpretq_s32_u32(vmovl_u16(vget_low_u16(scales1)));
+        scales32.val[1] = vreinterpretq_s32_u32(vmovl_u16(vget_high_u16(scales1)));
+        scales32.val[2] = vreinterpretq_s32_u32(vmovl_u16(vget_low_u16(scales2)));
+        scales32.val[3] = vreinterpretq_s32_u32(vmovl_u16(vget_high_u16(scales2)));
+        int32x4_t sumi = vdupq_n_s32(0);
+        for (int ib64 = 0; ib64 < QK_K/64; ++ib64) {
+            q8b = ggml_vld1q_s8_x4(q8); q8 += 64;
+            q2u.val[0] = vcombine_s8(vld1_s8((const void *)(iq2xs_grid + (q2[0] & 511))), vld1_s8((const void *)(iq2xs_grid + (q2[1] & 511))));
+            q2u.val[1] = vcombine_s8(vld1_s8((const void *)(iq2xs_grid + (q2[2] & 511))), vld1_s8((const void *)(iq2xs_grid + (q2[3] & 511))));
+            q2u.val[2] = vcombine_s8(vld1_s8((const void *)(iq2xs_grid + (q2[4] & 511))), vld1_s8((const void *)(iq2xs_grid + (q2[5] & 511))));
+            q2u.val[3] = vcombine_s8(vld1_s8((const void *)(iq2xs_grid + (q2[6] & 511))), vld1_s8((const void *)(iq2xs_grid + (q2[7] & 511))));
+            q2s.val[0] = vcombine_s8(vld1_s8((const void *)(signs64 + (q2[0] >> 9))), vld1_s8((const void *)(signs64 + (q2[1] >> 9))));
+            q2s.val[1] = vcombine_s8(vld1_s8((const void *)(signs64 + (q2[2] >> 9))), vld1_s8((const void *)(signs64 + (q2[3] >> 9))));
+            q2s.val[2] = vcombine_s8(vld1_s8((const void *)(signs64 + (q2[4] >> 9))), vld1_s8((const void *)(signs64 + (q2[5] >> 9))));
+            q2s.val[3] = vcombine_s8(vld1_s8((const void *)(signs64 + (q2[6] >> 9))), vld1_s8((const void *)(signs64 + (q2[7] >> 9))));
+            q2u.val[0] = vmulq_s8(q2u.val[0], q2s.val[0]);
+            q2u.val[1] = vmulq_s8(q2u.val[1], q2s.val[1]);
+            q2u.val[2] = vmulq_s8(q2u.val[2], q2s.val[2]);
+            q2u.val[3] = vmulq_s8(q2u.val[3], q2s.val[3]);
+            const int32x4_t p1 = ggml_vdotq_s32(vdupq_n_s32(0), q2u.val[0], q8b.val[0]);
+            const int32x4_t p2 = ggml_vdotq_s32(vdupq_n_s32(0), q2u.val[1], q8b.val[1]);
+            const int32x4_t p3 = ggml_vdotq_s32(vdupq_n_s32(0), q2u.val[2], q8b.val[2]);
+            const int32x4_t p4 = ggml_vdotq_s32(vdupq_n_s32(0), q2u.val[3], q8b.val[3]);
+            const int32x4_t p = vpaddq_s32(vpaddq_s32(p1, p2), vpaddq_s32(p3, p4));
+            sumi = vmlaq_s32(sumi, p, scales32.val[ib64]);
+            q2 += 8;
+        }
+        sumf += d*vaddvq_s32(sumi);
+    }
+    *s = 0.125f * sumf;
+
+#else
+
+    float sumf = 0.f;
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
+        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;
+            const uint16_t ls2 = 2*(sc[ib32] >>  4) + 1;
+            int32_t sumi = 0;
+            for (int l = 0; l < 2; ++l) {
+                const uint8_t * grid = (const uint8_t *)(iq2xs_grid + (q2[l] & 511));
+                const uint8_t  signs = ksigns_iq2xs[q2[l] >> 9];
+                for (int j = 0; j < 8; ++j) {
+                    sumi += grid[j] * q8[j] * (signs & kmask_iq2xs[j] ? -1 : 1);
+                }
+                q8 += 8;
+            }
+            bsum += sumi * ls1;
+            sumi = 0;
+            for (int l = 2; l < 4; ++l) {
+                const uint8_t * grid = (const uint8_t *)(iq2xs_grid + (q2[l] & 511));
+                const uint8_t  signs = ksigns_iq2xs[q2[l] >> 9];
+                for (int j = 0; j < 8; ++j) {
+                    sumi += grid[j] * q8[j] * (signs & kmask_iq2xs[j] ? -1 : 1);
+                }
+                q8 += 8;
+            }
+            bsum += sumi * ls2;
+            q2 += 4;
+        }
+        sumf += d * bsum;
+    }
+    *s = 0.125f * sumf;
+#endif
+}
+
+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);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_iq2_s * GGML_RESTRICT x = vx;
+    const block_q8_K  * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+#if defined(__ARM_NEON)
+
+   static const uint8_t k_mask1[32] = {0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01,
+                                       0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03
+   };
+
+    static const uint8_t k_mask2[16] = {0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,};
+
+    const ggml_uint8x16x2_t mask1 = ggml_vld1q_u8_x2(k_mask1);
+    const uint8x16_t        mask2 = vld1q_u8(k_mask2);
+    const uint8x16_t m1 = vdupq_n_u8(1);
+    const int32x4_t vzero = vdupq_n_s32(0);
+
+    uint8x16x2_t vs;
+    ggml_int8x16x4_t q2s;
+    ggml_int8x16x4_t q8b;
+
+    float sumf = 0;
+    for (int i = 0; i < nb; ++i) {
+
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
+
+        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) {
+            q8b = ggml_vld1q_s8_x4(q8); q8 += 64;
+            q2s.val[0] = vcombine_s8(vld1_s8((const int8_t *)(iq2s_grid + (qs[0] | ((qh[ib32+0] << 8) & 0x300)))),
+                                     vld1_s8((const int8_t *)(iq2s_grid + (qs[1] | ((qh[ib32+0] << 6) & 0x300)))));
+            q2s.val[1] = vcombine_s8(vld1_s8((const int8_t *)(iq2s_grid + (qs[2] | ((qh[ib32+0] << 4) & 0x300)))),
+                                     vld1_s8((const int8_t *)(iq2s_grid + (qs[3] | ((qh[ib32+0] << 2) & 0x300)))));
+            q2s.val[2] = vcombine_s8(vld1_s8((const int8_t *)(iq2s_grid + (qs[4] | ((qh[ib32+1] << 8) & 0x300)))),
+                                     vld1_s8((const int8_t *)(iq2s_grid + (qs[5] | ((qh[ib32+1] << 6) & 0x300)))));
+            q2s.val[3] = vcombine_s8(vld1_s8((const int8_t *)(iq2s_grid + (qs[6] | ((qh[ib32+1] << 4) & 0x300)))),
+                                     vld1_s8((const int8_t *)(iq2s_grid + (qs[7] | ((qh[ib32+1] << 2) & 0x300)))));
+            qs += 8;
+
+            vs.val[0] = vreinterpretq_u8_u32(vdupq_n_u32(signs[0] | ((uint32_t) signs[1] << 16)));
+            vs.val[1] = vandq_u8(ggml_vqtbl1q_u8(vs.val[0], mask1.val[1]), mask2);
+            vs.val[0] = vandq_u8(ggml_vqtbl1q_u8(vs.val[0], mask1.val[0]), mask2);
+            vs.val[0] = vceqq_u8(vs.val[0], mask2);
+            vs.val[1] = vceqq_u8(vs.val[1], mask2);
+
+            q2s.val[0] = vmulq_s8(vreinterpretq_s8_u8(vorrq_u8(vs.val[0], m1)), q2s.val[0]);
+            q2s.val[1] = vmulq_s8(vreinterpretq_s8_u8(vorrq_u8(vs.val[1], m1)), q2s.val[1]);
+
+            vs.val[0] = vreinterpretq_u8_u32(vdupq_n_u32(signs[2] | ((uint32_t) signs[3] << 16)));
+            vs.val[1] = vandq_u8(ggml_vqtbl1q_u8(vs.val[0], mask1.val[1]), mask2);
+            vs.val[0] = vandq_u8(ggml_vqtbl1q_u8(vs.val[0], mask1.val[0]), mask2);
+            vs.val[0] = vceqq_u8(vs.val[0], mask2);
+            vs.val[1] = vceqq_u8(vs.val[1], mask2);
+
+            signs += 4;
+
+            q2s.val[2] = vmulq_s8(vreinterpretq_s8_u8(vorrq_u8(vs.val[0], m1)), q2s.val[2]);
+            q2s.val[3] = vmulq_s8(vreinterpretq_s8_u8(vorrq_u8(vs.val[1], m1)), q2s.val[3]);
+
+            const int32x4_t p1 = ggml_vdotq_s32(vzero, q2s.val[0], q8b.val[0]);
+            const int32x4_t p2 = ggml_vdotq_s32(vzero, q2s.val[1], q8b.val[1]);
+            const int32x4_t p3 = ggml_vdotq_s32(vzero, q2s.val[2], q8b.val[2]);
+            const int32x4_t p4 = ggml_vdotq_s32(vzero, q2s.val[3], q8b.val[3]);
+
+            sumi1 += vaddvq_s32(p1) * (1 + 2*(x[i].scales[ib32+0] & 0xf));
+            sumi2 += vaddvq_s32(p2) * (1 + 2*(x[i].scales[ib32+0] >>  4));
+            sumi1 += vaddvq_s32(p3) * (1 + 2*(x[i].scales[ib32+1] & 0xf));
+            sumi2 += vaddvq_s32(p4) * (1 + 2*(x[i].scales[ib32+1] >>  4));
+        }
+        sumf += d*(sumi1 + sumi2);
+    }
+
+    *s = 0.125f * sumf;
+
+#else
+
+    float sumf = 0;
+    for (int i = 0; i < nb; i++) {
+
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
+        const int8_t  * q8 = y[i].qs;
+        const uint8_t * qs = x[i].qs;
+        const uint8_t * qh = x[i].qh;
+        const uint8_t * signs = qs + QK_K/8;
+
+        int bsum = 0;
+        for (int ib32 = 0; ib32 < QK_K/32; ++ib32) {
+            int ls1 = 1 + 2*(x[i].scales[ib32] & 0xf);
+            int ls2 = 1 + 2*(x[i].scales[ib32] >>  4);
+            int sumi1 = 0, sumi2 = 0;
+            for (int l = 0; l < 2; ++l) {
+                const uint8_t * grid = (const uint8_t *)(iq2s_grid + (qs[l] | (qh[ib32] << (8-2*l) & 0x300)));
+                for (int j = 0; j < 8; ++j) {
+                    sumi1 += q8[j] * grid[j] * (signs[l] & kmask_iq2xs[j] ? -1 : 1);
+                }
+                q8 += 8;
+            }
+            for (int l = 2; l < 4; ++l) {
+                const uint8_t * grid = (const uint8_t *)(iq2s_grid + (qs[l] | (qh[ib32] << (8-2*l) & 0x300)));
+                for (int j = 0; j < 8; ++j) {
+                    sumi2 += q8[j] * grid[j] * (signs[l] & kmask_iq2xs[j] ? -1 : 1);
+                }
+                q8 += 8;
+            }
+            bsum += ls1 * sumi1 + ls2 * sumi2;
+            qs += 4;
+            signs += 4;
+        }
+
+        sumf += d * bsum;
+    }
+
+    *s = 0.125f * sumf;
+
+#endif
+
+}
+
+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);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_iq3_xxs * GGML_RESTRICT x = vx;
+    const block_q8_K    * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+#if defined(__ARM_NEON)
+
+    const uint64_t * signs64 = (const uint64_t *)keven_signs_q2xs;
+
+    uint32_t aux32[2];
+
+    ggml_int8x16x4_t q3s;
+    ggml_int8x16x4_t q8b;
+
+    float sumf = 0;
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
+        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;
+            memcpy(aux32, gas, 2*sizeof(uint32_t)); gas += 2*sizeof(uint32_t);
+            const uint32x4_t aux32x4_0 = ggml_vld1q_u32(iq3xxs_grid[q3[ 0]], iq3xxs_grid[q3[ 1]], iq3xxs_grid[q3[ 2]], iq3xxs_grid[q3[ 3]]);
+            const uint32x4_t aux32x4_1 = ggml_vld1q_u32(iq3xxs_grid[q3[ 4]], iq3xxs_grid[q3[ 5]], iq3xxs_grid[q3[ 6]], iq3xxs_grid[q3[ 7]]);
+            const uint32x4_t aux32x4_2 = ggml_vld1q_u32(iq3xxs_grid[q3[ 8]], iq3xxs_grid[q3[ 9]], iq3xxs_grid[q3[10]], iq3xxs_grid[q3[11]]);
+            const uint32x4_t aux32x4_3 = ggml_vld1q_u32(iq3xxs_grid[q3[12]], iq3xxs_grid[q3[13]], iq3xxs_grid[q3[14]], iq3xxs_grid[q3[15]]);
+            q3 += 16;
+            q3s.val[0] = vcombine_s8(vld1_s8((const void *)(signs64 + ((aux32[0] >>  0) & 127))), vld1_s8((const void *)(signs64 + ((aux32[0] >>  7) & 127))));
+            q3s.val[1] = vcombine_s8(vld1_s8((const void *)(signs64 + ((aux32[0] >> 14) & 127))), vld1_s8((const void *)(signs64 + ((aux32[0] >> 21) & 127))));
+            q3s.val[2] = vcombine_s8(vld1_s8((const void *)(signs64 + ((aux32[1] >>  0) & 127))), vld1_s8((const void *)(signs64 + ((aux32[1] >>  7) & 127))));
+            q3s.val[3] = vcombine_s8(vld1_s8((const void *)(signs64 + ((aux32[1] >> 14) & 127))), vld1_s8((const void *)(signs64 + ((aux32[1] >> 21) & 127))));
+            q3s.val[0] = vmulq_s8(q3s.val[0], vreinterpretq_s8_u32(aux32x4_0));
+            q3s.val[1] = vmulq_s8(q3s.val[1], vreinterpretq_s8_u32(aux32x4_1));
+            q3s.val[2] = vmulq_s8(q3s.val[2], vreinterpretq_s8_u32(aux32x4_2));
+            q3s.val[3] = vmulq_s8(q3s.val[3], vreinterpretq_s8_u32(aux32x4_3));
+            const int32x4_t p1 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), q3s.val[0], q8b.val[0]), q3s.val[1], q8b.val[1]);
+            const int32x4_t p2 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), q3s.val[2], q8b.val[2]), q3s.val[3], q8b.val[3]);
+            sumf1 += vaddvq_s32(p1) * (0.5f + (aux32[0] >> 28));
+            sumf2 += vaddvq_s32(p2) * (0.5f + (aux32[1] >> 28));
+        }
+        sumf += d*(sumf1 + sumf2);
+    }
+    *s = 0.5f * sumf;
+
+#else
+
+    uint32_t aux32;
+
+    float sumf = 0.f;
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
+        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);
+            const uint32_t ls = 2*(aux32 >> 28) + 1;
+            int32_t sumi = 0;
+            for (int l = 0; l < 4; ++l) {
+                const uint8_t * grid1 = (const uint8_t *)(iq3xxs_grid + q3[2*l+0]);
+                const uint8_t * grid2 = (const uint8_t *)(iq3xxs_grid + q3[2*l+1]);
+                const uint8_t  signs = ksigns_iq2xs[(aux32 >> 7*l) & 127];
+                for (int j = 0; j < 4; ++j) {
+                    sumi += grid1[j] * q8[j+0] * (signs & kmask_iq2xs[j+0] ? -1 : 1);
+                    sumi += grid2[j] * q8[j+4] * (signs & kmask_iq2xs[j+4] ? -1 : 1);
+                }
+                q8 += 8;
+            }
+            q3 += 8;
+            bsum += sumi * ls;
+        }
+        sumf += d * bsum;
+    }
+    *s = 0.25f * sumf;
+#endif
+}
+
+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);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_iq3_s * GGML_RESTRICT x = vx;
+    const block_q8_K  * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+#if defined(__ARM_NEON)
+
+    typedef union {
+        uint16x8_t vec_index;
+        uint16_t   index[8];
+    } vec_index_t;
+
+   static const uint8_t k_mask1[32] = {0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01,
+                                       0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03
+   };
+
+    static const uint8_t k_mask2[16] = {0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,};
+
+    static const int16_t k_shift[8] = {8, 7, 6, 5, 4, 3, 2, 1};
+
+    const ggml_uint8x16x2_t mask1 = ggml_vld1q_u8_x2(k_mask1);
+    const uint8x16_t        mask2 = vld1q_u8(k_mask2);
+
+    const int16x8_t  hshift = vld1q_s16(k_shift);
+    const uint16x8_t m256   = vdupq_n_u16(256);
+    const uint8x16_t m1     = vdupq_n_u8(1);
+
+    uint8x16x2_t vs;
+    ggml_int8x16x4_t q3s;
+    ggml_int8x16x4_t q8b;
+    vec_index_t idx;
+
+    uint32_t scales32[2];
+    const uint8_t * scales8 = (const uint8_t *)scales32;
+
+    float sumf = 0;
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
+        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;
+        scales32[0] = ((scales32[0] & 0x0f0f0f0f) << 1) | 0x01010101;
+
+        int sumi1 = 0, sumi2 = 0;
+        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
+            q8b = ggml_vld1q_s8_x4(q8); q8 += 64;
+
+            const uint8x16_t idx_l = vld1q_u8(qs); qs += 16;
+            idx.vec_index = vorrq_u16(vmovl_u8(vget_low_u8 (idx_l)), vandq_u16(vshlq_u16(vdupq_n_u16(qh[ib32+0]), hshift), m256));
+            const uint32x4_t aux32x4_0 = ggml_vld1q_u32(iq3s_grid[idx.index[0]], iq3s_grid[idx.index[1]],
+                                                        iq3s_grid[idx.index[2]], iq3s_grid[idx.index[3]]);
+            const uint32x4_t aux32x4_1 = ggml_vld1q_u32(iq3s_grid[idx.index[4]], iq3s_grid[idx.index[5]],
+                                                        iq3s_grid[idx.index[6]], iq3s_grid[idx.index[7]]);
+            idx.vec_index = vorrq_u16(vmovl_u8(vget_high_u8(idx_l)), vandq_u16(vshlq_u16(vdupq_n_u16(qh[ib32+1]), hshift), m256));
+            const uint32x4_t aux32x4_2 = ggml_vld1q_u32(iq3s_grid[idx.index[0]], iq3s_grid[idx.index[1]],
+                                                        iq3s_grid[idx.index[2]], iq3s_grid[idx.index[3]]);
+            const uint32x4_t aux32x4_3 = ggml_vld1q_u32(iq3s_grid[idx.index[4]], iq3s_grid[idx.index[5]],
+                                                        iq3s_grid[idx.index[6]], iq3s_grid[idx.index[7]]);
+
+
+            vs.val[0] = vreinterpretq_u8_u32(vdupq_n_u32(signs[0] | ((uint32_t) signs[1] << 16)));
+            vs.val[1] = vandq_u8(ggml_vqtbl1q_u8(vs.val[0], mask1.val[1]), mask2);
+            vs.val[0] = vandq_u8(ggml_vqtbl1q_u8(vs.val[0], mask1.val[0]), mask2);
+            vs.val[0] = vorrq_u8(vceqq_u8(vs.val[0], mask2), m1);
+            vs.val[1] = vorrq_u8(vceqq_u8(vs.val[1], mask2), m1);
+
+            q3s.val[0] = vmulq_s8(vreinterpretq_s8_u8(vs.val[0]), vreinterpretq_s8_u32(aux32x4_0));
+            q3s.val[1] = vmulq_s8(vreinterpretq_s8_u8(vs.val[1]), vreinterpretq_s8_u32(aux32x4_1));
+
+            vs.val[0] = vreinterpretq_u8_u32(vdupq_n_u32(signs[2] | ((uint32_t) signs[3] << 16)));
+            vs.val[1] = vandq_u8(ggml_vqtbl1q_u8(vs.val[0], mask1.val[1]), mask2);
+            vs.val[0] = vandq_u8(ggml_vqtbl1q_u8(vs.val[0], mask1.val[0]), mask2);
+            vs.val[0] = vorrq_u8(vceqq_u8(vs.val[0], mask2), m1);
+            vs.val[1] = vorrq_u8(vceqq_u8(vs.val[1], mask2), m1);
+
+            signs += 4;
+
+            q3s.val[2] = vmulq_s8(vreinterpretq_s8_u8(vs.val[0]), vreinterpretq_s8_u32(aux32x4_2));
+            q3s.val[3] = vmulq_s8(vreinterpretq_s8_u8(vs.val[1]), vreinterpretq_s8_u32(aux32x4_3));
+
+            const int32x4_t p1 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), q3s.val[0], q8b.val[0]), q3s.val[1], q8b.val[1]);
+            const int32x4_t p2 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), q3s.val[2], q8b.val[2]), q3s.val[3], q8b.val[3]);
+
+            sumi1 += vaddvq_s32(p1) * scales8[ib32/2+0];
+            sumi2 += vaddvq_s32(p2) * scales8[ib32/2+4];
+        }
+        sumf += d*(sumi1 + sumi2);
+    }
+    *s = sumf;
+
+#else
+
+    float sumf = 0.f;
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
+        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;
+            const uint32_t ls2 = 2*(x[i].scales[ib32/2] >>  4) + 1;
+            int32_t sumi = 0;
+            for (int l = 0; l < 4; ++l) {
+                const uint8_t * grid1 = (const uint8_t *)(iq3s_grid + (qs[2*l+0] | ((qh[ib32+0] << (8-2*l)) & 256)));
+                const uint8_t * grid2 = (const uint8_t *)(iq3s_grid + (qs[2*l+1] | ((qh[ib32+0] << (7-2*l)) & 256)));
+                for (int j = 0; j < 4; ++j) {
+                    sumi += grid1[j] * q8[j+0] * (signs[l] & kmask_iq2xs[j+0] ? -1 : 1);
+                    sumi += grid2[j] * q8[j+4] * (signs[l] & kmask_iq2xs[j+4] ? -1 : 1);
+                }
+                q8 += 8;
+            }
+            qs += 8;
+            signs += 4;
+            bsum += sumi * ls1;
+            sumi = 0;
+            for (int l = 0; l < 4; ++l) {
+                const uint8_t * grid1 = (const uint8_t *)(iq3s_grid + (qs[2*l+0] | ((qh[ib32+1] << (8-2*l)) & 256)));
+                const uint8_t * grid2 = (const uint8_t *)(iq3s_grid + (qs[2*l+1] | ((qh[ib32+1] << (7-2*l)) & 256)));
+                for (int j = 0; j < 4; ++j) {
+                    sumi += grid1[j] * q8[j+0] * (signs[l] & kmask_iq2xs[j+0] ? -1 : 1);
+                    sumi += grid2[j] * q8[j+4] * (signs[l] & kmask_iq2xs[j+4] ? -1 : 1);
+                }
+                q8 += 8;
+            }
+            qs += 8;
+            signs += 4;
+            bsum += sumi * ls2;
+        }
+        sumf += d * bsum;
+    }
+    *s = sumf;
+#endif
+}
+
+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);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_iq1_s * GGML_RESTRICT x = vx;
+    const block_q8_K  * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+#if defined __ARM_NEON
+
+    ggml_int8x16x4_t q1b;
+    ggml_int8x16x4_t q8b;
+
+    float sumf = 0;
+    for (int i = 0; i < nb; ++i) {
+
+        const int8_t   * q8 = y[i].qs;
+        const uint8_t  * qs = x[i].qs;
+        const uint16_t * qh = x[i].qh;
+
+        int sumi1 = 0, sumi2 = 0, sumi3 = 0;
+
+        for (int ib = 0; ib < QK_K/32; ib += 2) {
+
+            q1b.val[0] = vcombine_s8(vld1_s8((const int8_t *)(iq1s_grid + (qs[0] | ((qh[ib+0] << 8) & 0x700)))),
+                                     vld1_s8((const int8_t *)(iq1s_grid + (qs[1] | ((qh[ib+0] << 5) & 0x700)))));
+            q1b.val[1] = vcombine_s8(vld1_s8((const int8_t *)(iq1s_grid + (qs[2] | ((qh[ib+0] << 2) & 0x700)))),
+                                     vld1_s8((const int8_t *)(iq1s_grid + (qs[3] | ((qh[ib+0] >> 1) & 0x700)))));
+            q1b.val[2] = vcombine_s8(vld1_s8((const int8_t *)(iq1s_grid + (qs[4] | ((qh[ib+1] << 8) & 0x700)))),
+                                     vld1_s8((const int8_t *)(iq1s_grid + (qs[5] | ((qh[ib+1] << 5) & 0x700)))));
+            q1b.val[3] = vcombine_s8(vld1_s8((const int8_t *)(iq1s_grid + (qs[6] | ((qh[ib+1] << 2) & 0x700)))),
+                                     vld1_s8((const int8_t *)(iq1s_grid + (qs[7] | ((qh[ib+1] >> 1) & 0x700)))));
+            qs += 8;
+
+            q8b = ggml_vld1q_s8_x4(q8); q8 += 64;
+
+            const int32x4_t p1 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), q1b.val[0], q8b.val[0]), q1b.val[1], q8b.val[1]);
+            const int32x4_t p2 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), q1b.val[2], q8b.val[2]), q1b.val[3], q8b.val[3]);
+
+            const int ls1 = 2*((qh[ib+0] >> 12) & 7) + 1;
+            const int ls2 = 2*((qh[ib+1] >> 12) & 7) + 1;
+            sumi1 += vaddvq_s32(p1) * ls1;
+            sumi2 += vaddvq_s32(p2) * ls2;
+            sumi3 += (y[i].bsums[2*ib+0] + y[i].bsums[2*ib+1]) * ls1 * (qh[ib+0] & 0x8000 ? -1 : 1)
+                   + (y[i].bsums[2*ib+2] + y[i].bsums[2*ib+3]) * ls2 * (qh[ib+1] & 0x8000 ? -1 : 1);
+
+        }
+
+        sumf += y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d) * (sumi1 + sumi2 + IQ1S_DELTA * sumi3);
+    }
+
+    *s = sumf;
+
+#else
+
+    float sumf = 0;
+    for (int i = 0; i < nb; i++) {
+
+        const int8_t   * q8 = y[i].qs;
+        const uint8_t  * qs = x[i].qs;
+        const uint16_t * qh = x[i].qh;
+
+        int sumi = 0, sumi1 = 0;
+        for (int ib = 0; ib < QK_K/32; ++ib) {
+            const int ls = 2*((qh[ib] >> 12) & 7) + 1;
+            const int delta = qh[ib] & 0x8000 ? -1 : 1;
+            int lsum = 0;
+            for (int l = 0; l < 4; ++l) {
+                const int8_t * grid = (const int8_t *)(iq1s_grid + (qs[l] | (((qh[ib] >> 3*l) & 7) << 8)));
+                for (int j = 0; j < 8; ++j) {
+                    lsum += q8[j] * grid[j];
+                }
+                q8 += 8;
+            }
+            sumi  += ls * lsum;
+            sumi1 += ls * delta * (y[i].bsums[2*ib+0] + y[i].bsums[2*ib+1]);
+            qs += 4;
+        }
+
+        sumf += GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d * (sumi + IQ1S_DELTA * sumi1);
+    }
+
+    *s = sumf;
+
+#endif
+}
+
+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);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_iq1_m * GGML_RESTRICT x = vx;
+    const block_q8_K  * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+    iq1m_scale_t scale;
+
+#if defined __ARM_NEON
+    const int32x4_t mask  = vdupq_n_s32(0x7);
+    const int32x4_t mone  = vdupq_n_s32(1);
+    const int32x4_t mzero = vdupq_n_s32(0);
+
+    ggml_int8x16x4_t deltas;
+    deltas.val[0] = vcombine_s8(vdup_n_s8(+1), vdup_n_s8(+1));
+    deltas.val[1] = vcombine_s8(vdup_n_s8(-1), vdup_n_s8(+1));
+    deltas.val[2] = vcombine_s8(vdup_n_s8(+1), vdup_n_s8(-1));
+    deltas.val[3] = vcombine_s8(vdup_n_s8(-1), vdup_n_s8(-1));
+
+    ggml_int8x16x4_t q1b;
+    ggml_int8x16x4_t q8b;
+
+    uint32_t aux32;
+    const uint8_t * aux8 = (const uint8_t *)&aux32;
+
+    float sumf = 0;
+    for (int i = 0; i < nb; ++i) {
+
+        const int8_t   * q8 = y[i].qs;
+        const uint8_t  * qs = x[i].qs;
+        const uint8_t  * qh = x[i].qh;
+        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);
+
+        int32x4_t sumi1 = mzero;
+        int32x4_t sumi2 = mzero;
+
+        for (int ib = 0; ib < QK_K/32; ib += 2) {
+
+            q1b.val[0] = vcombine_s8(vld1_s8((const int8_t *)(iq1s_grid + (qs[0] | ((qh[0] << 8) & 0x700)))),
+                                     vld1_s8((const int8_t *)(iq1s_grid + (qs[1] | ((qh[0] << 4) & 0x700)))));
+            q1b.val[1] = vcombine_s8(vld1_s8((const int8_t *)(iq1s_grid + (qs[2] | ((qh[1] << 8) & 0x700)))),
+                                     vld1_s8((const int8_t *)(iq1s_grid + (qs[3] | ((qh[1] << 4) & 0x700)))));
+            q1b.val[2] = vcombine_s8(vld1_s8((const int8_t *)(iq1s_grid + (qs[4] | ((qh[2] << 8) & 0x700)))),
+                                     vld1_s8((const int8_t *)(iq1s_grid + (qs[5] | ((qh[2] << 4) & 0x700)))));
+            q1b.val[3] = vcombine_s8(vld1_s8((const int8_t *)(iq1s_grid + (qs[6] | ((qh[3] << 8) & 0x700)))),
+                                     vld1_s8((const int8_t *)(iq1s_grid + (qs[7] | ((qh[3] << 4) & 0x700)))));
+
+            q8b = ggml_vld1q_s8_x4(q8); q8 += 64;
+
+            const int32x4_t p1 = vpaddq_s32(ggml_vdotq_s32(mzero, q1b.val[0], q8b.val[0]), ggml_vdotq_s32(mzero, q1b.val[1], q8b.val[1]));
+            const int32x4_t p2 = vpaddq_s32(ggml_vdotq_s32(mzero, q1b.val[2], q8b.val[2]), ggml_vdotq_s32(mzero, q1b.val[3], q8b.val[3]));
+            const int32x4_t p12 = vpaddq_s32(p1, p2);
+
+            const uint32_t * qh32 = (const uint32_t *)qh; // we are 4-byte aligned, so we can do that
+            aux32 = ((qh32[0] >> 3) & 0x01010101) | ((qh32[0] >> 6) & 0x02020202);
+
+            const int32x4_t p3 = vpaddq_s32(ggml_vdotq_s32(mzero, deltas.val[aux8[0]], q8b.val[0]), ggml_vdotq_s32(mzero, deltas.val[aux8[1]], q8b.val[1]));
+            const int32x4_t p4 = vpaddq_s32(ggml_vdotq_s32(mzero, deltas.val[aux8[2]], q8b.val[2]), ggml_vdotq_s32(mzero, deltas.val[aux8[3]], q8b.val[3]));
+            const int32x4_t p34 = vpaddq_s32(p3, p4);
+
+            int32x4_t scales_4 = ggml_vld1q_u32(sc[ib/2] >> 0, sc[ib/2] >> 3, sc[ib/2] >> 6, sc[ib/2] >> 9);
+
+            scales_4 = vaddq_s32(vshlq_n_s32(vandq_s32(scales_4, mask), 1), mone);
+
+            sumi1 = vmlaq_s32(sumi1, scales_4, p12);
+            sumi2 = vmlaq_s32(sumi2, scales_4, p34);
+
+            qs += 8; qh += 4;
+
+        }
+
+        sumf += y[i].d * GGML_CPU_FP16_TO_FP32(scale.f16) * (vaddvq_s32(sumi1) + IQ1M_DELTA * vaddvq_s32(sumi2));
+    }
+
+    *s = sumf;
+
+#else
+
+    int sum1[2], sum2[2], delta[4];
+
+    float sumf = 0;
+    for (int i = 0; i < nb; i++) {
+
+        const int8_t   * q8 = y[i].qs;
+        const uint8_t  * qs = x[i].qs;
+        const uint8_t  * qh = x[i].qh;
+        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);
+
+        int sumi1 = 0, sumi2 = 0;
+        for (int ib = 0; ib < QK_K/32; ++ib) {
+            delta[0] = qh[0] & 0x08 ? -1 : 1;
+            delta[1] = qh[0] & 0x80 ? -1 : 1;
+            delta[2] = qh[1] & 0x08 ? -1 : 1;
+            delta[3] = qh[1] & 0x80 ? -1 : 1;
+            sum1[0] = sum1[1] = sum2[0] = sum2[1] = 0;
+            for (int l = 0; l < 4; ++l) {
+                const int8_t * grid = (const int8_t *)(iq1s_grid + (qs[l] | (((uint16_t)qh[l/2] << (8 - 4*(l%2))) & 0x700)));
+                int lsum1 = 0, lsum2 = 0;
+                for (int j = 0; j < 8; ++j) {
+                    lsum1 += q8[j] * grid[j];
+                    lsum2 += q8[j];
+                }
+                q8 += 8;
+                sum1[l/2] += lsum1;
+                sum2[l/2] += lsum2*delta[l];
+            }
+
+            const int ls1 = 2*((sc[ib/2] >> (6*(ib%2)+0)) & 0x7) + 1;
+            const int ls2 = 2*((sc[ib/2] >> (6*(ib%2)+3)) & 0x7) + 1;
+
+            sumi1 += sum1[0] * ls1 + sum1[1] * ls2;
+            sumi2 += sum2[0] * ls1 + sum2[1] * ls2;
+            qs += 4;
+            qh += 2;
+        }
+
+        sumf += GGML_CPU_FP16_TO_FP32(scale.f16) * y[i].d * (sumi1 + IQ1M_DELTA * sumi2);
+    }
+
+    *s = sumf;
+
+#endif
+}
+
+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);
+    UNUSED(by);
+    UNUSED(bs);
+    assert(n % QK4_NL == 0);
+    static_assert(QK4_NL == QK8_0, "QK4_NL and QK8_0 must be the same");
+
+    const block_iq4_nl * GGML_RESTRICT x = vx;
+    const block_q8_0   * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK4_NL;
+
+    int ib = 0;
+    float sumf = 0;
+
+#if defined __ARM_NEON
+    const int8x16_t values = vld1q_s8(kvalues_iq4nl);
+    const uint8x16_t m4b = vdupq_n_u8(0x0f);
+    uint8x16x2_t q4bits;
+    int8x16x4_t q4b;
+    int8x16x4_t q8b;
+    int32x4_t prod_1, prod_2;
+
+    for (; ib + 1 < nb; ib += 2) {
+
+        q4bits.val[0] = vld1q_u8(x[ib + 0].qs);
+        q4bits.val[1] = vld1q_u8(x[ib + 1].qs);
+        q8b.val[0]    = vld1q_s8(y[ib + 0].qs);
+        q8b.val[1]    = vld1q_s8(y[ib + 0].qs + 16);
+        q8b.val[2]    = vld1q_s8(y[ib + 1].qs);
+        q8b.val[3]    = vld1q_s8(y[ib + 1].qs + 16);
+
+        q4b.val[0] = ggml_vqtbl1q_s8(values, vandq_u8  (q4bits.val[0], m4b));
+        q4b.val[1] = ggml_vqtbl1q_s8(values, vshrq_n_u8(q4bits.val[0], 4));
+        q4b.val[2] = ggml_vqtbl1q_s8(values, vandq_u8  (q4bits.val[1], m4b));
+        q4b.val[3] = ggml_vqtbl1q_s8(values, vshrq_n_u8(q4bits.val[1], 4));
+
+        prod_1 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), q4b.val[0], q8b.val[0]), q4b.val[1], q8b.val[1]);
+        prod_2 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), q4b.val[2], q8b.val[2]), q4b.val[3], q8b.val[3]);
+
+        sumf +=
+            GGML_CPU_FP16_TO_FP32(x[ib+0].d) * GGML_CPU_FP16_TO_FP32(y[ib + 0].d) * vaddvq_s32(prod_1) +
+            GGML_CPU_FP16_TO_FP32(x[ib+1].d) * GGML_CPU_FP16_TO_FP32(y[ib + 1].d) * vaddvq_s32(prod_2);
+    }
+
+#endif
+    for (; ib < nb; ++ib) {
+        const float d = GGML_CPU_FP16_TO_FP32(y[ib].d)*GGML_CPU_FP16_TO_FP32(x[ib].d);
+        int sumi1 = 0, sumi2 = 0;
+        for (int j = 0; j < QK4_NL/2; ++j) {
+            sumi1 += y[ib].qs[j+       0] * kvalues_iq4nl[x[ib].qs[j] & 0xf];
+            sumi2 += y[ib].qs[j+QK4_NL/2] * kvalues_iq4nl[x[ib].qs[j] >>  4];
+        }
+        sumf += d * (sumi1 + sumi2);
+    }
+    *s = sumf;
+}
+
+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);
+    UNUSED(by);
+    UNUSED(bs);
+    assert(n % QK_K == 0);
+
+    const block_iq4_xs * GGML_RESTRICT x = vx;
+    const block_q8_K   * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+#if defined __ARM_NEON
+    const int8x16_t values = vld1q_s8(kvalues_iq4nl);
+    const uint8x16_t m4b = vdupq_n_u8(0x0f);
+    ggml_uint8x16x2_t q4bits;
+    ggml_int8x16x4_t q4b;
+    ggml_int8x16x4_t q8b;
+    int32x4_t prod_1, prod_2;
+
+    float sumf = 0;
+
+    for (int ibl = 0; ibl < nb; ++ibl) {
+
+        const int8_t  * q8 = y[ibl].qs;
+        const uint8_t * q4 = x[ibl].qs;
+        uint16_t h = x[ibl].scales_h;
+
+        int sumi1 = 0, sumi2 = 0;
+        for (int ib = 0; ib < QK_K/64; ++ib) {
+
+            q4bits = ggml_vld1q_u8_x2(q4); q4 += 32;
+            q8b    = ggml_vld1q_s8_x4(q8); q8 += 64;
+
+            q4b.val[0] = ggml_vqtbl1q_s8(values, vandq_u8  (q4bits.val[0], m4b));
+            q4b.val[1] = ggml_vqtbl1q_s8(values, vshrq_n_u8(q4bits.val[0], 4));
+            q4b.val[2] = ggml_vqtbl1q_s8(values, vandq_u8  (q4bits.val[1], m4b));
+            q4b.val[3] = ggml_vqtbl1q_s8(values, vshrq_n_u8(q4bits.val[1], 4));
+
+            prod_1 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), q4b.val[0], q8b.val[0]), q4b.val[1], q8b.val[1]);
+            prod_2 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), q4b.val[2], q8b.val[2]), q4b.val[3], q8b.val[3]);
+
+            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 += vaddvq_s32(prod_1) * ls1;
+            sumi2 += vaddvq_s32(prod_2) * ls2;
+
+        }
+
+        sumf += GGML_CPU_FP16_TO_FP32(x[ibl].d) * y[ibl].d * (sumi1 + sumi2);
+    }
+
+    *s = sumf;
+
+#else
+    float sumf = 0;
+    for (int ibl = 0; ibl < nb; ++ibl) {
+        const float d4d8 = GGML_CPU_FP16_TO_FP32(x[ibl].d) * y[ibl].d;
+        uint16_t h = x[ibl].scales_h;
+        const uint8_t * qs = x[ibl].qs;
+        const int8_t  * q8 = y[ibl].qs;
+        for (int ib = 0; ib < QK_K/32; ib += 2) {
+            const uint8_t ls1 = (x[ibl].scales_l[ib/2] & 0xf) | ((h << 4) & 0x30);
+            const uint8_t ls2 = (x[ibl].scales_l[ib/2] >>  4) | ((h << 2) & 0x30);
+            h >>= 4;
+            const float d1 = d4d8*(ls1 - 32);
+            const float d2 = d4d8*(ls2 - 32);
+            int sumi1 = 0, sumi2 = 0;
+            for (int j = 0; j < 16; ++j) {
+                sumi1 += q8[j+ 0] * kvalues_iq4nl[qs[j] & 0xf];
+                sumi2 += q8[j+16] * kvalues_iq4nl[qs[j] >>  4];
+            }
+            sumf += d1 * (sumi1 + sumi2);
+            qs += 16;
+            q8 += 32;
+            sumi1 = sumi2 = 0;
+            for (int j = 0; j < 16; ++j) {
+                sumi1 += q8[j+ 0] * kvalues_iq4nl[qs[j] & 0xf];
+                sumi2 += q8[j+16] * kvalues_iq4nl[qs[j] >>  4];
+            }
+            sumf += d2 * (sumi1 + sumi2);
+            qs += 16;
+            q8 += 32;
+        }
+    }
+    *s = sumf;
+#endif
+}
+
diff --git a/ggml/src/ggml-cpu/arch/arm/repack.cpp b/ggml/src/ggml-cpu/arch/arm/repack.cpp
new file mode 100644
index 000000000..2f8bc9e25
--- /dev/null
+++ b/ggml/src/ggml-cpu/arch/arm/repack.cpp
@@ -0,0 +1,2163 @@
+#define GGML_COMMON_IMPL_CPP
+#define GGML_COMMON_DECL_CPP
+#include "ggml-common.h"
+#include "ggml-backend-impl.h"
+
+#include "ggml-impl.h"
+#include "ggml-cpu.h"
+#include "ggml-cpu-impl.h"
+#include "simd-mappings.h"
+#include "traits.h"
+
+#include <cmath>
+#include <cstring>
+#include <cassert>
+#include <cstdlib> // for qsort
+#include <cstdio>  // for GGML_ASSERT
+
+#define GGML_CPU_CLANG_WORKAROUND
+#include "../../repack.h"
+
+#if defined(__GNUC__)
+#pragma GCC diagnostic ignored "-Woverlength-strings"
+#endif
+
+#define UNUSED GGML_UNUSED
+
+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;
+
+    block_q8_0x4 * GGML_RESTRICT y = (block_q8_0x4 *) vy;
+
+#if defined(__ARM_NEON)
+    float32x4_t srcv[4][8];
+    float id[4];
+
+    for (int i = 0; i < nb; i++) {
+        float32x4_t asrcv[8];
+        float32x4_t amaxv[8];
+
+        for (int row_iter = 0; row_iter < 4; row_iter++) {
+            for (int j = 0; j < 8; j++) srcv[row_iter][j] = vld1q_f32(x + row_iter * k + i * 32 + 4 * j);
+            for (int j = 0; j < 8; j++) asrcv[j] = vabsq_f32(srcv[row_iter][j]);
+
+            for (int j = 0; j < 4; j++) amaxv[2 * j] = vmaxq_f32(asrcv[2 * j], asrcv[2 * j + 1]);
+            for (int j = 0; j < 2; j++) amaxv[4 * j] = vmaxq_f32(amaxv[4 * j], amaxv[4 * j + 2]);
+            for (int j = 0; j < 1; j++) amaxv[8 * j] = vmaxq_f32(amaxv[8 * j], amaxv[8 * j + 4]);
+
+            const float amax = vmaxvq_f32(amaxv[0]);
+
+            const float d = amax / ((1 << 7) - 1);
+            id[row_iter] = d ? 1.0f / d : 0.0f;
+
+            y[i].d[row_iter] = GGML_CPU_FP32_TO_FP16(d);
+        }
+
+        for (int j = 0; j < 8; j++) {
+            float32x4_t v = vmulq_n_f32(srcv[0][j], id[0]);
+            int32x4_t vi = vcvtnq_s32_f32(v);
+            y[i].qs[16 * j + 0] = vgetq_lane_s32(vi, 0);
+            y[i].qs[16 * j + 1] = vgetq_lane_s32(vi, 1);
+            y[i].qs[16 * j + 2] = vgetq_lane_s32(vi, 2);
+            y[i].qs[16 * j + 3] = vgetq_lane_s32(vi, 3);
+
+            v = vmulq_n_f32(srcv[1][j], id[1]);
+            vi = vcvtnq_s32_f32(v);
+            y[i].qs[16 * j + 4] = vgetq_lane_s32(vi, 0);
+            y[i].qs[16 * j + 5] = vgetq_lane_s32(vi, 1);
+            y[i].qs[16 * j + 6] = vgetq_lane_s32(vi, 2);
+            y[i].qs[16 * j + 7] = vgetq_lane_s32(vi, 3);
+
+            v = vmulq_n_f32(srcv[2][j], id[2]);
+            vi = vcvtnq_s32_f32(v);
+            y[i].qs[16 * j + 8] = vgetq_lane_s32(vi, 0);
+            y[i].qs[16 * j + 9] = vgetq_lane_s32(vi, 1);
+            y[i].qs[16 * j + 10] = vgetq_lane_s32(vi, 2);
+            y[i].qs[16 * j + 11] = vgetq_lane_s32(vi, 3);
+
+            v = vmulq_n_f32(srcv[3][j], id[3]);
+            vi = vcvtnq_s32_f32(v);
+            y[i].qs[16 * j + 12] = vgetq_lane_s32(vi, 0);
+            y[i].qs[16 * j + 13] = vgetq_lane_s32(vi, 1);
+            y[i].qs[16 * j + 14] = vgetq_lane_s32(vi, 2);
+            y[i].qs[16 * j + 15] = vgetq_lane_s32(vi, 3);
+        }
+    }
+#else
+    // scalar
+    const int blck_size_interleave = 4;
+    float srcv[4][QK8_0];
+    float id[4];
+
+    for (int i = 0; i < nb; i++) {
+        for (int row_iter = 0; row_iter < 4; row_iter++) {
+            float amax = 0.0f; // absolute max
+
+            for (int j = 0; j < QK8_0; j++) {
+                srcv[row_iter][j] = x[row_iter * k + i * QK8_0 + j];
+                amax = MAX(amax, fabsf(srcv[row_iter][j]));
+            }
+
+            const float d = amax / ((1 << 7) - 1);
+            id[row_iter] = d ? 1.0f / d : 0.0f;
+
+            y[i].d[row_iter] = GGML_CPU_FP32_TO_FP16(d);
+        }
+
+        for (int j = 0; j < QK8_0 * 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);
+
+            float x0 = srcv[src_id][src_offset] * id[src_id];
+            y[i].qs[j] = roundf(x0);
+        }
+    }
+#endif
+}
+
+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;
+
+    block_q8_0x4 * GGML_RESTRICT y = (block_q8_0x4 *) vy;
+
+#if defined(__ARM_NEON)
+    float32x4_t srcv[4][8];
+    float id[4];
+
+    for (int i = 0; i < nb; i++) {
+        float32x4_t asrcv[8];
+        float32x4_t amaxv[8];
+
+        for (int row_iter = 0; row_iter < 4; row_iter++) {
+            for (int j = 0; j < 8; j++) srcv[row_iter][j] = vld1q_f32(x + row_iter * k + i * 32 + 4 * j);
+            for (int j = 0; j < 8; j++) asrcv[j] = vabsq_f32(srcv[row_iter][j]);
+
+            for (int j = 0; j < 4; j++) amaxv[2 * j] = vmaxq_f32(asrcv[2 * j], asrcv[2 * j + 1]);
+            for (int j = 0; j < 2; j++) amaxv[4 * j] = vmaxq_f32(amaxv[4 * j], amaxv[4 * j + 2]);
+            for (int j = 0; j < 1; j++) amaxv[8 * j] = vmaxq_f32(amaxv[8 * j], amaxv[8 * j + 4]);
+
+            const float amax = vmaxvq_f32(amaxv[0]);
+
+            const float d = amax / ((1 << 7) - 1);
+            id[row_iter] = d ? 1.0f / d : 0.0f;
+
+            y[i].d[row_iter] = GGML_CPU_FP32_TO_FP16(d);
+        }
+
+        for (int j = 0; j < 4; j++) {
+            float32x4_t v = vmulq_n_f32(srcv[0][2 * j], id[0]);
+            int32x4_t vi = vcvtnq_s32_f32(v);
+            y[i].qs[32 * j + 0] = vgetq_lane_s32(vi, 0);
+            y[i].qs[32 * j + 1] = vgetq_lane_s32(vi, 1);
+            y[i].qs[32 * j + 2] = vgetq_lane_s32(vi, 2);
+            y[i].qs[32 * j + 3] = vgetq_lane_s32(vi, 3);
+            v = vmulq_n_f32(srcv[0][2 * j + 1], id[0]);
+            vi = vcvtnq_s32_f32(v);
+            y[i].qs[32 * j + 4] = vgetq_lane_s32(vi, 0);
+            y[i].qs[32 * j + 5] = vgetq_lane_s32(vi, 1);
+            y[i].qs[32 * j + 6] = vgetq_lane_s32(vi, 2);
+            y[i].qs[32 * j + 7] = vgetq_lane_s32(vi, 3);
+
+            v = vmulq_n_f32(srcv[1][2 * j], id[1]);
+            vi = vcvtnq_s32_f32(v);
+            y[i].qs[32 * j + 8] = vgetq_lane_s32(vi, 0);
+            y[i].qs[32 * j + 9] = vgetq_lane_s32(vi, 1);
+            y[i].qs[32 * j + 10] = vgetq_lane_s32(vi, 2);
+            y[i].qs[32 * j + 11] = vgetq_lane_s32(vi, 3);
+            v = vmulq_n_f32(srcv[1][2 * j + 1], id[1]);
+            vi = vcvtnq_s32_f32(v);
+            y[i].qs[32 * j + 12] = vgetq_lane_s32(vi, 0);
+            y[i].qs[32 * j + 13] = vgetq_lane_s32(vi, 1);
+            y[i].qs[32 * j + 14] = vgetq_lane_s32(vi, 2);
+            y[i].qs[32 * j + 15] = vgetq_lane_s32(vi, 3);
+
+            v = vmulq_n_f32(srcv[2][2 * j], id[2]);
+            vi = vcvtnq_s32_f32(v);
+            y[i].qs[32 * j + 16] = vgetq_lane_s32(vi, 0);
+            y[i].qs[32 * j + 17] = vgetq_lane_s32(vi, 1);
+            y[i].qs[32 * j + 18] = vgetq_lane_s32(vi, 2);
+            y[i].qs[32 * j + 19] = vgetq_lane_s32(vi, 3);
+            v = vmulq_n_f32(srcv[2][2 * j + 1], id[2]);
+            vi = vcvtnq_s32_f32(v);
+            y[i].qs[32 * j + 20] = vgetq_lane_s32(vi, 0);
+            y[i].qs[32 * j + 21] = vgetq_lane_s32(vi, 1);
+            y[i].qs[32 * j + 22] = vgetq_lane_s32(vi, 2);
+            y[i].qs[32 * j + 23] = vgetq_lane_s32(vi, 3);
+
+            v = vmulq_n_f32(srcv[3][2 * j], id[3]);
+            vi = vcvtnq_s32_f32(v);
+            y[i].qs[32 * j + 24] = vgetq_lane_s32(vi, 0);
+            y[i].qs[32 * j + 25] = vgetq_lane_s32(vi, 1);
+            y[i].qs[32 * j + 26] = vgetq_lane_s32(vi, 2);
+            y[i].qs[32 * j + 27] = vgetq_lane_s32(vi, 3);
+            v = vmulq_n_f32(srcv[3][2 * j + 1], id[3]);
+            vi = vcvtnq_s32_f32(v);
+            y[i].qs[32 * j + 28] = vgetq_lane_s32(vi, 0);
+            y[i].qs[32 * j + 29] = vgetq_lane_s32(vi, 1);
+            y[i].qs[32 * j + 30] = vgetq_lane_s32(vi, 2);
+            y[i].qs[32 * j + 31] = vgetq_lane_s32(vi, 3);
+        }
+    }
+
+#else
+    // scalar
+    const int blck_size_interleave = 8;
+    float srcv[4][QK8_0];
+    float id[4];
+
+    for (int i = 0; i < nb; i++) {
+        for (int row_iter = 0; row_iter < 4; row_iter++) {
+            float amax = 0.0f; // absolute max
+
+            for (int j = 0; j < QK8_0; j++) {
+                srcv[row_iter][j] = x[row_iter * k + i * QK8_0 + j];
+                amax = MAX(amax, fabsf(srcv[row_iter][j]));
+            }
+
+            const float d = amax / ((1 << 7) - 1);
+            id[row_iter] = d ? 1.0f / d : 0.0f;
+
+            y[i].d[row_iter] = GGML_CPU_FP32_TO_FP16(d);
+        }
+
+        for (int j = 0; j < QK8_0 * 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);
+
+            float x0 = srcv[src_id][src_offset] * id[src_id];
+            y[i].qs[j] = roundf(x0);
+        }
+    }
+#endif
+}
+
+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) {
+    const int qk = QK8_0;
+    const int nb = n / qk;
+    const int ncols_interleaved = 4;
+    const int blocklen = 4;
+
+    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(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD)
+    const block_q4_0x4 * b_ptr = (const block_q4_0x4 *) vx;
+
+    for (int c = 0; c < nc; c += ncols_interleaved) {
+        const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
+        float32x4_t acc = vdupq_n_f32(0);
+        for (int b = 0; b < nb; b++) {
+            int8x16_t b0 = vld1q_s8((const int8_t *) b_ptr->qs);
+            int8x16_t b1 = vld1q_s8((const int8_t *) b_ptr->qs + 16);
+            int8x16_t b2 = vld1q_s8((const int8_t *) b_ptr->qs + 32);
+            int8x16_t b3 = vld1q_s8((const int8_t *) b_ptr->qs + 48);
+            float16x4_t bd = vld1_f16((const __fp16 *) b_ptr->d);
+
+            int8x16_t a0 = vld1q_s8(a_ptr->qs);
+            int8x16_t a1 = vld1q_s8(a_ptr->qs + qk/2);
+            float16x4_t ad = vld1_dup_f16((const __fp16 *) &a_ptr->d);
+
+            int32x4_t ret = vdupq_n_s32(0);
+
+            ret = vdotq_laneq_s32(ret, b0 << 4, a0, 0);
+            ret = vdotq_laneq_s32(ret, b1 << 4, a0, 1);
+            ret = vdotq_laneq_s32(ret, b2 << 4, a0, 2);
+            ret = vdotq_laneq_s32(ret, b3 << 4, a0, 3);
+
+            ret = vdotq_laneq_s32(ret, b0 & 0xf0U, a1, 0);
+            ret = vdotq_laneq_s32(ret, b1 & 0xf0U, a1, 1);
+            ret = vdotq_laneq_s32(ret, b2 & 0xf0U, a1, 2);
+            ret = vdotq_laneq_s32(ret, b3 & 0xf0U, a1, 3);
+
+            acc = vfmaq_f32(acc, vcvtq_n_f32_s32(ret, 4),
+                            vmulq_f32(vcvt_f32_f16(ad), vcvt_f32_f16(bd)));
+            a_ptr++;
+            b_ptr++;
+        }
+        vst1q_f32(s, acc);
+        s += ncols_interleaved;
+    }
+    return;
+#endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD)
+    float sumf[4];
+    int sumi;
+
+    const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
+    for (int x = 0; x < nc / ncols_interleaved; x++) {
+        const block_q4_0x4 * b_ptr = (const block_q4_0x4 *) vx + (x * nb);
+
+        for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0;
+        for (int l = 0; l < nb; l++) {
+            for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                for (int j = 0; j < ncols_interleaved; j++) {
+                    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] << 4);
+                        const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
+                        sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4;
+                    }
+                    sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
+                }
+            }
+        }
+        for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j];
+    }
+}
+
+void ggml_gemv_q4_0_4x8_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;
+    const int ncols_interleaved = 4;
+    const int blocklen = 8;
+
+    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(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD)
+    const block_q4_0x4 * b_ptr = (const block_q4_0x4 *) vx;
+
+    for (int c = 0; c < nc; c += ncols_interleaved) {
+        const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
+        float32x4_t acc = vdupq_n_f32(0);
+        for (int b = 0; b < nb; b++) {
+            int8x16_t b0 = vld1q_s8((const int8_t *) b_ptr->qs);
+            int8x16_t b1 = vld1q_s8((const int8_t *) b_ptr->qs + 16);
+            int8x16_t b2 = vld1q_s8((const int8_t *) b_ptr->qs + 32);
+            int8x16_t b3 = vld1q_s8((const int8_t *) b_ptr->qs + 48);
+            float16x4_t bd = vld1_f16((const __fp16 *) b_ptr->d);
+
+            int8x16_t a0 = (int8x16_t) vld1q_dup_s64((const int64_t *) a_ptr->qs);
+            int8x16_t a1 = (int8x16_t) vld1q_dup_s64((const int64_t *) a_ptr->qs + 1);
+            int8x16_t a2 = (int8x16_t) vld1q_dup_s64((const int64_t *) a_ptr->qs + 2);
+            int8x16_t a3 = (int8x16_t) vld1q_dup_s64((const int64_t *) a_ptr->qs + 3);
+            float16x4_t ad = vld1_dup_f16((const __fp16 *) &a_ptr->d);
+
+            int32x4_t ret0 = vdupq_n_s32(0);
+            int32x4_t ret1 = vdupq_n_s32(0);
+
+            ret0 = vdotq_s32(ret0, b0 << 4, a0);
+            ret1 = vdotq_s32(ret1, b1 << 4, a0);
+            ret0 = vdotq_s32(ret0, b2 << 4, a1);
+            ret1 = vdotq_s32(ret1, b3 << 4, a1);
+
+            ret0 = vdotq_s32(ret0, b0 & 0xf0U, a2);
+            ret1 = vdotq_s32(ret1, b1 & 0xf0U, a2);
+            ret0 = vdotq_s32(ret0, b2 & 0xf0U, a3);
+            ret1 = vdotq_s32(ret1, b3 & 0xf0U, a3);
+
+            int32x4_t ret = vpaddq_s32(ret0, ret1);
+
+            acc = vfmaq_f32(acc, vcvtq_n_f32_s32(ret, 4),
+                    vmulq_f32(vcvt_f32_f16(ad), vcvt_f32_f16(bd)));
+            a_ptr++;
+            b_ptr++;
+        }
+        vst1q_f32(s, acc);
+        s += ncols_interleaved;
+    }
+    return;
+#endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD)
+    float sumf[4];
+    int sumi;
+
+    const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
+    for (int x = 0; x < nc / ncols_interleaved; x++) {
+        const block_q4_0x4 * b_ptr = (const block_q4_0x4 *) vx + (x * nb);
+
+        for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0;
+        for (int l = 0; l < nb; l++) {
+            for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                for (int j = 0; j < ncols_interleaved; j++) {
+                    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] << 4);
+                        const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
+                        sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4;
+                    }
+                    sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
+                }
+            }
+        }
+        for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j];
+    }
+}
+
+void ggml_gemv_q4_0_8x8_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;
+    const int ncols_interleaved = 8;
+    const int blocklen = 8;
+
+    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(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__)
+#if defined(__ARM_FEATURE_SVE)
+    if (ggml_cpu_get_sve_cnt() == QK8_0) {
+        const void * b_ptr = vx;
+        const void * a_ptr = vy;
+        float * res_ptr = s;
+
+        __asm__ __volatile__(
+            "ptrue p0.b\n"
+            "add %x[b_ptr], %x[b_ptr], #0x10\n"
+            "1:"  // Column loop
+            "add x22, %x[a_ptr], #0x2\n"
+            "mov z31.b, #0x0\n"
+            "mov x21, %x[nb]\n"
+            "2:"  // Block loop
+            "ld1b { z30.b }, p0/Z, [%x[b_ptr]]\n"
+            "ld1b { z29.b }, p0/Z, [%x[b_ptr], #1, MUL VL]\n"
+            "mov z28.s, #0x0\n"
+            "mov z27.s, #0x0\n"
+            "ld1rd { z26.d }, p0/Z, [x22]\n"
+            "ld1b { z25.b }, p0/Z, [%x[b_ptr], #2, MUL VL]\n"
+            "sub x20, x22, #0x2\n"
+            "sub x21, x21, #0x1\n"
+            "ld1b { z24.b }, p0/Z, [%x[b_ptr], #3, MUL VL]\n"
+            "ld1rd { z23.d }, p0/Z, [x22, #8]\n"
+            "lsl z22.b, z30.b, #0x4\n"
+            "lsl z16.b, z29.b, #0x4\n"
+            "and z30.b, z30.b, #0xf0\n"
+            "and z29.b, z29.b, #0xf0\n"
+            "ld1rd { z21.d }, p0/Z, [x22, #16]\n"
+            "ld1rd { z20.d }, p0/Z, [x22, #24]\n"
+            "lsl z19.b, z25.b, #0x4\n"
+            "and z25.b, z25.b, #0xf0\n"
+            "ld1rh { z17.h }, p0/Z, [x20]\n"
+            "ld1h { z18.s }, p0/Z, [%x[b_ptr], #-1, MUL VL]\n"
+            "sdot z28.s, z22.b, z26.b\n"
+            "sdot z27.s, z16.b, z26.b\n"
+            "lsl z16.b, z24.b, #0x4\n"
+            "add x22, x22, #0x22\n"
+            "and z24.b, z24.b, #0xf0\n"
+            "add %x[b_ptr], %x[b_ptr], #0x90\n"
+            "fcvt z17.s, p0/m, z17.h\n"
+            "fcvt z18.s, p0/m, z18.h\n"
+            "sdot z28.s, z19.b, z23.b\n"
+            "sdot z27.s, z16.b, z23.b\n"
+            "fmul z18.s, z18.s, z17.s\n"
+            "sdot z28.s, z30.b, z21.b\n"
+            "sdot z27.s, z29.b, z21.b\n"
+            "sdot z28.s, z25.b, z20.b\n"
+            "sdot z27.s, z24.b, z20.b\n"
+            "uzp1 z17.s, z28.s, z27.s\n"
+            "uzp2 z16.s, z28.s, z27.s\n"
+            "add z17.s, z17.s, z16.s\n"
+            "asr z17.s, z17.s, #0x4\n"
+            "scvtf z17.s, p0/m, z17.s\n"
+            "fmla z31.s, p0/M, z17.s, z18.s\n"
+            "cbnz x21, 2b\n"
+            "sub %x[nc], %x[nc], #0x8\n"
+            "st1w { z31.s }, p0, [%x[res_ptr]]\n"
+            "add %x[res_ptr], %x[res_ptr], #0x20\n"
+            "cbnz %x[nc], 1b\n"
+            : [b_ptr] "+&r" (b_ptr), [res_ptr] "+&r" (res_ptr), [nc] "+&r" (nc)
+            : [a_ptr] "r" (a_ptr), [nb] "r" (nb)
+            : "memory", "p0", "x20", "x21", "x22", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
+        );
+        return;
+    }
+#endif // #if defined(__ARM_FEATURE_SVE)
+
+#endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__)
+    {
+        float sumf[8];
+        int sumi;
+
+        const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
+        for (int x = 0; x < nc / ncols_interleaved; x++) {
+            const block_q4_0x8 * b_ptr = (const block_q4_0x8 *) vx + (x * nb);
+
+            for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0;
+            for (int l = 0; l < nb; l++) {
+                for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                    for (int j = 0; j < ncols_interleaved; j++) {
+                        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] << 4);
+                            const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
+                            sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4;
+                        }
+                        sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
+                    }
+                }
+            }
+            for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j];
+        }
+    }
+}
+
+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;
+    const int ncols_interleaved = 4;
+    const int blocklen = 4;
+
+    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(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD)
+    const int8x16_t kvalues = vld1q_s8(kvalues_iq4nl);
+    const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
+    float * res_ptr = s;
+
+    for (int x = 0; x < nc / ncols_interleaved; x++) {
+        const block_iq4_nlx4 * b_ptr = (const block_iq4_nlx4 *) vx + (x * nb);
+
+        float32x4_t sumf = vdupq_n_f32(0);
+        for (int l = 0; l < nb; l++) {
+            uint8x16_t b_0 = vld1q_u8(b_ptr[l].qs + 0);
+            uint8x16_t b_1 = vld1q_u8(b_ptr[l].qs + 16);
+            uint8x16_t b_2 = vld1q_u8(b_ptr[l].qs + 32);
+            uint8x16_t b_3 = vld1q_u8(b_ptr[l].qs + 48);
+
+            int8x16_t b_0_hi = vqtbl1q_s8(kvalues, b_0 >> 4);
+            int8x16_t b_0_lo = vqtbl1q_s8(kvalues, b_0 & 0x0F);
+            int8x16_t b_1_hi = vqtbl1q_s8(kvalues, b_1 >> 4);
+            int8x16_t b_1_lo = vqtbl1q_s8(kvalues, b_1 & 0x0F);
+            int8x16_t b_2_hi = vqtbl1q_s8(kvalues, b_2 >> 4);
+            int8x16_t b_2_lo = vqtbl1q_s8(kvalues, b_2 & 0x0F);
+            int8x16_t b_3_hi = vqtbl1q_s8(kvalues, b_3 >> 4);
+            int8x16_t b_3_lo = vqtbl1q_s8(kvalues, b_3 & 0x0F);
+
+            int8x16_t a_0 = vld1q_s8(a_ptr[l].qs + 0);
+            int8x16_t a_1 = vld1q_s8(a_ptr[l].qs + 16);
+
+            int32x4_t sumi = vdupq_n_s32(0);
+            sumi = vdotq_laneq_s32(sumi, b_0_lo, a_0, 0);
+            sumi = vdotq_laneq_s32(sumi, b_0_hi, a_1, 0);
+            sumi = vdotq_laneq_s32(sumi, b_1_lo, a_0, 1);
+            sumi = vdotq_laneq_s32(sumi, b_1_hi, a_1, 1);
+            sumi = vdotq_laneq_s32(sumi, b_2_lo, a_0, 2);
+            sumi = vdotq_laneq_s32(sumi, b_2_hi, a_1, 2);
+            sumi = vdotq_laneq_s32(sumi, b_3_lo, a_0, 3);
+            sumi = vdotq_laneq_s32(sumi, b_3_hi, a_1, 3);
+
+            float32x4_t a_d = vcvt_f32_f16(vld1_dup_f16((const float16_t *)&a_ptr[l].d));
+            float32x4_t b_d = vcvt_f32_f16(vld1_f16((const float16_t *)b_ptr[l].d));
+            float32x4_t d = a_d * b_d;
+
+            sumf = vmlaq_f32(sumf, d, vcvtq_f32_s32(sumi));
+        }
+
+        vst1q_f32(res_ptr + x * 4, sumf);
+    }
+    return;
+#endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON)
+    {
+        float sumf[4];
+        int sumi;
+
+        const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
+        for (int x = 0; x < nc / ncols_interleaved; x++) {
+            const block_iq4_nlx4 * b_ptr = (const block_iq4_nlx4 *) vx + (x * nb);
+
+            for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0;
+            for (int l = 0; l < nb; l++) {
+                for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                    for (int j = 0; j < ncols_interleaved; j++) {
+                        sumi = 0;
+                        for (int i = 0; i < blocklen; ++i) {
+                            const int v0 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0x0F];
+                            const int v1 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4];
+                            sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2]));
+                        }
+                        sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
+                    }
+                }
+            }
+            for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j];
+        }
+    }
+}
+
+void ggml_gemm_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) {
+    const int qk = QK8_0;
+    const int nb = n / qk;
+    const int ncols_interleaved = 4;
+    const int blocklen = 4;
+
+    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(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD)
+    const void * b_ptr = vx;
+    const void * a_ptr = vy;
+    float * res_ptr = s;
+    size_t res_stride = bs * sizeof(float);
+
+    __asm__ __volatile__(
+        "mov x10, %x[nr]\n"
+        "mov x9, #0x88\n"
+        "cmp x10, #0x10\n"
+        "mul x9, %x[nb], x9\n"
+        "blt 4f\n"
+        "1:"  // Row loop
+        "add x28, %x[b_ptr], #0x8\n"
+        "mov x27, %x[nc]\n"
+        "add x26, %x[res_ptr], %x[res_stride], LSL #4\n"
+        "2:"  // Column loop
+        "add x25, %x[a_ptr], #0x8\n"
+        "movi v15.16b, #0x0\n"
+        "movi v19.16b, #0x0\n"
+        "mov x24, %x[nb]\n"
+        "add x23, x25, x9\n"
+        "movi v18.16b, #0x0\n"
+        "movi v14.16b, #0x0\n"
+        "add x22, x23, x9\n"
+        "movi v11.16b, #0x0\n"
+        "movi v13.16b, #0x0\n"
+        "add x21, x22, x9\n"
+        "movi v23.16b, #0x0\n"
+        "movi v16.16b, #0x0\n"
+        "movi v25.16b, #0x0\n"
+        "movi v7.16b, #0x0\n"
+        "movi v0.16b, #0x0\n"
+        "movi v4.16b, #0x0\n"
+        "movi v5.16b, #0x0\n"
+        "movi v21.16b, #0x0\n"
+        "movi v8.16b, #0x0\n"
+        "movi v1.16b, #0x0\n"
+        "3:"  // Block loop
+        "ldr q3, [x28, #0x0]\n"
+        "ldr q31, [x25, #0x0]\n"
+        "movi v28.16b, #0x4\n"
+        "movi v10.4s, #0x0\n"
+        "ldr q22, [x28, #0x10]\n"
+        "ldr q6, [x25, #0x10]\n"
+        "movi v29.4s, #0x0\n"
+        "movi v9.4s, #0x0\n"
+        "ldr q27, [x28, #0x20]\n"
+        "ldr q30, [x28, #0x30]\n"
+        "movi v20.4s, #0x0\n"
+        "movi v24.16b, #0xf0\n"
+        "ldr d2, [x25, #-0x8]\n"
+        "ldr d26, [x23, #-0x8]\n"
+        "sshl v12.16b, v3.16b, v28.16b\n"
+        "sub x20, x28, #0x8\n"
+        "ldr d17, [x20, #0x0]\n"
+        "and v3.16b, v3.16b, v24.16b\n"
+        "subs x24, x24, #0x1\n"
+        "add x28, x28, #0x48\n"
+        ".inst 0x4f9fe18a  // sdot v10.4s, v12.16b, v31.4b[0]\n"
+        ".inst 0x4fbfe19d  // sdot v29.4s, v12.16b, v31.4b[1]\n"
+        ".inst 0x4f9fe989  // sdot v9.4s, v12.16b, v31.4b[2]\n"
+        ".inst 0x4fbfe994  // sdot v20.4s, v12.16b, v31.4b[3]\n"
+        "sshl v31.16b, v22.16b, v28.16b\n"
+        "and v22.16b, v22.16b, v24.16b\n"
+        "fcvtl v17.4s, v17.4h\n"
+        "fcvtl v2.4s, v2.4h\n"
+        "fcvtl v26.4s, v26.4h\n"
+        ".inst 0x4f86e3ea  // sdot v10.4s, v31.16b, v6.4b[0]\n"
+        ".inst 0x4fa6e3fd  // sdot v29.4s, v31.16b, v6.4b[1]\n"
+        ".inst 0x4f86ebe9  // sdot v9.4s, v31.16b, v6.4b[2]\n"
+        ".inst 0x4fa6ebf4  // sdot v20.4s, v31.16b, v6.4b[3]\n"
+        "sshl v6.16b, v27.16b, v28.16b\n"
+        "sshl v28.16b, v30.16b, v28.16b\n"
+        "and v27.16b, v27.16b, v24.16b\n"
+        "and v30.16b, v30.16b, v24.16b\n"
+        "ldr q24, [x25, #0x20]\n"
+        ".inst 0x4f98e0ca  // sdot v10.4s, v6.16b, v24.4b[0]\n"
+        ".inst 0x4fb8e0dd  // sdot v29.4s, v6.16b, v24.4b[1]\n"
+        ".inst 0x4f98e8c9  // sdot v9.4s, v6.16b, v24.4b[2]\n"
+        ".inst 0x4fb8e8d4  // sdot v20.4s, v6.16b, v24.4b[3]\n"
+        "ldr q24, [x25, #0x30]\n"
+        ".inst 0x4f98e38a  // sdot v10.4s, v28.16b, v24.4b[0]\n"
+        ".inst 0x4fb8e39d  // sdot v29.4s, v28.16b, v24.4b[1]\n"
+        ".inst 0x4f98eb89  // sdot v9.4s, v28.16b, v24.4b[2]\n"
+        ".inst 0x4fb8eb94  // sdot v20.4s, v28.16b, v24.4b[3]\n"
+        "ldr q24, [x25, #0x40]\n"
+        ".inst 0x4f98e06a  // sdot v10.4s, v3.16b, v24.4b[0]\n"
+        ".inst 0x4fb8e07d  // sdot v29.4s, v3.16b, v24.4b[1]\n"
+        ".inst 0x4f98e869  // sdot v9.4s, v3.16b, v24.4b[2]\n"
+        ".inst 0x4fb8e874  // sdot v20.4s, v3.16b, v24.4b[3]\n"
+        "ldr q24, [x25, #0x50]\n"
+        ".inst 0x4f98e2ca  // sdot v10.4s, v22.16b, v24.4b[0]\n"
+        ".inst 0x4fb8e2dd  // sdot v29.4s, v22.16b, v24.4b[1]\n"
+        ".inst 0x4f98eac9  // sdot v9.4s, v22.16b, v24.4b[2]\n"
+        ".inst 0x4fb8ead4  // sdot v20.4s, v22.16b, v24.4b[3]\n"
+        "ldr q24, [x25, #0x60]\n"
+        ".inst 0x4f98e36a  // sdot v10.4s, v27.16b, v24.4b[0]\n"
+        ".inst 0x4fb8e37d  // sdot v29.4s, v27.16b, v24.4b[1]\n"
+        ".inst 0x4f98eb69  // sdot v9.4s, v27.16b, v24.4b[2]\n"
+        ".inst 0x4fb8eb74  // sdot v20.4s, v27.16b, v24.4b[3]\n"
+        "ldr q24, [x25, #0x70]\n"
+        "add x25, x25, #0x88\n"
+        ".inst 0x4f98e3ca  // sdot v10.4s, v30.16b, v24.4b[0]\n"
+        ".inst 0x4fb8e3dd  // sdot v29.4s, v30.16b, v24.4b[1]\n"
+        ".inst 0x4f98ebc9  // sdot v9.4s, v30.16b, v24.4b[2]\n"
+        ".inst 0x4fb8ebd4  // sdot v20.4s, v30.16b, v24.4b[3]\n"
+        "fmul v24.4s, v17.4s, v2.s[0]\n"
+        "scvtf v10.4s, v10.4s, #0x4\n"
+        "scvtf v29.4s, v29.4s, #0x4\n"
+        "scvtf v9.4s, v9.4s, #0x4\n"
+        "scvtf v20.4s, v20.4s, #0x4\n"
+        "fmla v15.4s, v10.4s, v24.4s\n"
+        "ldr q24, [x23, #0x0]\n"
+        "fmul v10.4s, v17.4s, v2.s[1]\n"
+        "fmla v19.4s, v29.4s, v10.4s\n"
+        "ldr q10, [x23, #0x10]\n"
+        "fmul v29.4s, v17.4s, v2.s[2]\n"
+        "fmul v2.4s, v17.4s, v2.s[3]\n"
+        "fmla v18.4s, v9.4s, v29.4s\n"
+        "movi v9.4s, #0x0\n"
+        "movi v29.4s, #0x0\n"
+        ".inst 0x4f98e189  // sdot v9.4s, v12.16b, v24.4b[0]\n"
+        ".inst 0x4fb8e19d  // sdot v29.4s, v12.16b, v24.4b[1]\n"
+        "fmla v14.4s, v20.4s, v2.4s\n"
+        "movi v20.4s, #0x0\n"
+        "movi v2.4s, #0x0\n"
+        ".inst 0x4f98e994  // sdot v20.4s, v12.16b, v24.4b[2]\n"
+        ".inst 0x4fb8e982  // sdot v2.4s, v12.16b, v24.4b[3]\n"
+        "ldr q24, [x23, #0x20]\n"
+        ".inst 0x4f8ae3e9  // sdot v9.4s, v31.16b, v10.4b[0]\n"
+        ".inst 0x4faae3fd  // sdot v29.4s, v31.16b, v10.4b[1]\n"
+        ".inst 0x4f8aebf4  // sdot v20.4s, v31.16b, v10.4b[2]\n"
+        ".inst 0x4faaebe2  // sdot v2.4s, v31.16b, v10.4b[3]\n"
+        "ldr q10, [x23, #0x30]\n"
+        ".inst 0x4f98e0c9  // sdot v9.4s, v6.16b, v24.4b[0]\n"
+        ".inst 0x4fb8e0dd  // sdot v29.4s, v6.16b, v24.4b[1]\n"
+        ".inst 0x4f98e8d4  // sdot v20.4s, v6.16b, v24.4b[2]\n"
+        ".inst 0x4fb8e8c2  // sdot v2.4s, v6.16b, v24.4b[3]\n"
+        "ldr q24, [x23, #0x40]\n"
+        ".inst 0x4f8ae389  // sdot v9.4s, v28.16b, v10.4b[0]\n"
+        ".inst 0x4faae39d  // sdot v29.4s, v28.16b, v10.4b[1]\n"
+        ".inst 0x4f8aeb94  // sdot v20.4s, v28.16b, v10.4b[2]\n"
+        ".inst 0x4faaeb82  // sdot v2.4s, v28.16b, v10.4b[3]\n"
+        "ldr q10, [x23, #0x50]\n"
+        ".inst 0x4f98e069  // sdot v9.4s, v3.16b, v24.4b[0]\n"
+        ".inst 0x4fb8e07d  // sdot v29.4s, v3.16b, v24.4b[1]\n"
+        ".inst 0x4f98e874  // sdot v20.4s, v3.16b, v24.4b[2]\n"
+        ".inst 0x4fb8e862  // sdot v2.4s, v3.16b, v24.4b[3]\n"
+        "ldr q24, [x23, #0x60]\n"
+        ".inst 0x4f8ae2c9  // sdot v9.4s, v22.16b, v10.4b[0]\n"
+        ".inst 0x4faae2dd  // sdot v29.4s, v22.16b, v10.4b[1]\n"
+        ".inst 0x4f8aead4  // sdot v20.4s, v22.16b, v10.4b[2]\n"
+        ".inst 0x4faaeac2  // sdot v2.4s, v22.16b, v10.4b[3]\n"
+        "ldr q10, [x23, #0x70]\n"
+        "add x23, x23, #0x88\n"
+        ".inst 0x4f98e369  // sdot v9.4s, v27.16b, v24.4b[0]\n"
+        ".inst 0x4fb8e37d  // sdot v29.4s, v27.16b, v24.4b[1]\n"
+        ".inst 0x4f98eb74  // sdot v20.4s, v27.16b, v24.4b[2]\n"
+        ".inst 0x4fb8eb62  // sdot v2.4s, v27.16b, v24.4b[3]\n"
+        "ldr q24, [x22, #0x0]\n"
+        ".inst 0x4f8ae3c9  // sdot v9.4s, v30.16b, v10.4b[0]\n"
+        ".inst 0x4faae3dd  // sdot v29.4s, v30.16b, v10.4b[1]\n"
+        ".inst 0x4f8aebd4  // sdot v20.4s, v30.16b, v10.4b[2]\n"
+        ".inst 0x4faaebc2  // sdot v2.4s, v30.16b, v10.4b[3]\n"
+        "fmul v10.4s, v17.4s, v26.s[0]\n"
+        "scvtf v9.4s, v9.4s, #0x4\n"
+        "scvtf v29.4s, v29.4s, #0x4\n"
+        "scvtf v20.4s, v20.4s, #0x4\n"
+        "scvtf v2.4s, v2.4s, #0x4\n"
+        "fmla v11.4s, v9.4s, v10.4s\n"
+        "ldr q9, [x22, #0x10]\n"
+        "fmul v10.4s, v17.4s, v26.s[1]\n"
+        "fmla v13.4s, v29.4s, v10.4s\n"
+        "ldr d29, [x22, #-0x8]\n"
+        "fmul v10.4s, v17.4s, v26.s[2]\n"
+        "fmul v26.4s, v17.4s, v26.s[3]\n"
+        "fcvtl v29.4s, v29.4h\n"
+        "fmla v23.4s, v20.4s, v10.4s\n"
+        "movi v20.4s, #0x0\n"
+        "movi v10.4s, #0x0\n"
+        "fmla v16.4s, v2.4s, v26.4s\n"
+        "movi v26.4s, #0x0\n"
+        "movi v2.4s, #0x0\n"
+        ".inst 0x4f98e194  // sdot v20.4s, v12.16b, v24.4b[0]\n"
+        ".inst 0x4fb8e18a  // sdot v10.4s, v12.16b, v24.4b[1]\n"
+        ".inst 0x4f98e99a  // sdot v26.4s, v12.16b, v24.4b[2]\n"
+        ".inst 0x4fb8e982  // sdot v2.4s, v12.16b, v24.4b[3]\n"
+        "ldr q24, [x22, #0x20]\n"
+        ".inst 0x4f89e3f4  // sdot v20.4s, v31.16b, v9.4b[0]\n"
+        ".inst 0x4fa9e3ea  // sdot v10.4s, v31.16b, v9.4b[1]\n"
+        ".inst 0x4f89ebfa  // sdot v26.4s, v31.16b, v9.4b[2]\n"
+        ".inst 0x4fa9ebe2  // sdot v2.4s, v31.16b, v9.4b[3]\n"
+        "ldr q9, [x22, #0x30]\n"
+        ".inst 0x4f98e0d4  // sdot v20.4s, v6.16b, v24.4b[0]\n"
+        ".inst 0x4fb8e0ca  // sdot v10.4s, v6.16b, v24.4b[1]\n"
+        ".inst 0x4f98e8da  // sdot v26.4s, v6.16b, v24.4b[2]\n"
+        ".inst 0x4fb8e8c2  // sdot v2.4s, v6.16b, v24.4b[3]\n"
+        "ldr q24, [x22, #0x40]\n"
+        ".inst 0x4f89e394  // sdot v20.4s, v28.16b, v9.4b[0]\n"
+        ".inst 0x4fa9e38a  // sdot v10.4s, v28.16b, v9.4b[1]\n"
+        ".inst 0x4f89eb9a  // sdot v26.4s, v28.16b, v9.4b[2]\n"
+        ".inst 0x4fa9eb82  // sdot v2.4s, v28.16b, v9.4b[3]\n"
+        "ldr q9, [x22, #0x50]\n"
+        ".inst 0x4f98e074  // sdot v20.4s, v3.16b, v24.4b[0]\n"
+        ".inst 0x4fb8e06a  // sdot v10.4s, v3.16b, v24.4b[1]\n"
+        ".inst 0x4f98e87a  // sdot v26.4s, v3.16b, v24.4b[2]\n"
+        ".inst 0x4fb8e862  // sdot v2.4s, v3.16b, v24.4b[3]\n"
+        "ldr q24, [x22, #0x60]\n"
+        ".inst 0x4f89e2d4  // sdot v20.4s, v22.16b, v9.4b[0]\n"
+        ".inst 0x4fa9e2ca  // sdot v10.4s, v22.16b, v9.4b[1]\n"
+        ".inst 0x4f89eada  // sdot v26.4s, v22.16b, v9.4b[2]\n"
+        ".inst 0x4fa9eac2  // sdot v2.4s, v22.16b, v9.4b[3]\n"
+        "ldr q9, [x22, #0x70]\n"
+        "add x22, x22, #0x88\n"
+        ".inst 0x4f98e374  // sdot v20.4s, v27.16b, v24.4b[0]\n"
+        ".inst 0x4fb8e36a  // sdot v10.4s, v27.16b, v24.4b[1]\n"
+        ".inst 0x4f98eb7a  // sdot v26.4s, v27.16b, v24.4b[2]\n"
+        ".inst 0x4fb8eb62  // sdot v2.4s, v27.16b, v24.4b[3]\n"
+        "ldr q24, [x21, #0x0]\n"
+        ".inst 0x4f89e3d4  // sdot v20.4s, v30.16b, v9.4b[0]\n"
+        ".inst 0x4fa9e3ca  // sdot v10.4s, v30.16b, v9.4b[1]\n"
+        ".inst 0x4f89ebda  // sdot v26.4s, v30.16b, v9.4b[2]\n"
+        ".inst 0x4fa9ebc2  // sdot v2.4s, v30.16b, v9.4b[3]\n"
+        "fmul v9.4s, v17.4s, v29.s[0]\n"
+        "scvtf v20.4s, v20.4s, #0x4\n"
+        "scvtf v10.4s, v10.4s, #0x4\n"
+        "scvtf v26.4s, v26.4s, #0x4\n"
+        "scvtf v2.4s, v2.4s, #0x4\n"
+        "fmla v25.4s, v20.4s, v9.4s\n"
+        "ldr q9, [x21, #0x10]\n"
+        "fmul v20.4s, v17.4s, v29.s[1]\n"
+        "fmla v7.4s, v10.4s, v20.4s\n"
+        "ldr d20, [x21, #-0x8]\n"
+        "fmul v10.4s, v17.4s, v29.s[2]\n"
+        "fmul v29.4s, v17.4s, v29.s[3]\n"
+        "fcvtl v20.4s, v20.4h\n"
+        "fmla v0.4s, v26.4s, v10.4s\n"
+        "movi v26.4s, #0x0\n"
+        "movi v10.4s, #0x0\n"
+        "fmla v4.4s, v2.4s, v29.4s\n"
+        "movi v2.4s, #0x0\n"
+        "movi v29.4s, #0x0\n"
+        ".inst 0x4f98e19a  // sdot v26.4s, v12.16b, v24.4b[0]\n"
+        ".inst 0x4fb8e18a  // sdot v10.4s, v12.16b, v24.4b[1]\n"
+        ".inst 0x4f98e982  // sdot v2.4s, v12.16b, v24.4b[2]\n"
+        ".inst 0x4fb8e99d  // sdot v29.4s, v12.16b, v24.4b[3]\n"
+        "ldr q12, [x21, #0x20]\n"
+        "fmul v24.4s, v17.4s, v20.s[0]\n"
+        ".inst 0x4f89e3fa  // sdot v26.4s, v31.16b, v9.4b[0]\n"
+        ".inst 0x4fa9e3ea  // sdot v10.4s, v31.16b, v9.4b[1]\n"
+        ".inst 0x4f89ebe2  // sdot v2.4s, v31.16b, v9.4b[2]\n"
+        ".inst 0x4fa9ebfd  // sdot v29.4s, v31.16b, v9.4b[3]\n"
+        "ldr q9, [x21, #0x30]\n"
+        "fmul v31.4s, v17.4s, v20.s[1]\n"
+        ".inst 0x4f8ce0da  // sdot v26.4s, v6.16b, v12.4b[0]\n"
+        ".inst 0x4face0ca  // sdot v10.4s, v6.16b, v12.4b[1]\n"
+        ".inst 0x4f8ce8c2  // sdot v2.4s, v6.16b, v12.4b[2]\n"
+        ".inst 0x4face8dd  // sdot v29.4s, v6.16b, v12.4b[3]\n"
+        "ldr q12, [x21, #0x40]\n"
+        "fmul v6.4s, v17.4s, v20.s[2]\n"
+        "fmul v20.4s, v17.4s, v20.s[3]\n"
+        ".inst 0x4f89e39a  // sdot v26.4s, v28.16b, v9.4b[0]\n"
+        ".inst 0x4fa9e38a  // sdot v10.4s, v28.16b, v9.4b[1]\n"
+        ".inst 0x4f89eb82  // sdot v2.4s, v28.16b, v9.4b[2]\n"
+        ".inst 0x4fa9eb9d  // sdot v29.4s, v28.16b, v9.4b[3]\n"
+        "ldr q9, [x21, #0x50]\n"
+        ".inst 0x4f8ce07a  // sdot v26.4s, v3.16b, v12.4b[0]\n"
+        ".inst 0x4face06a  // sdot v10.4s, v3.16b, v12.4b[1]\n"
+        ".inst 0x4f8ce862  // sdot v2.4s, v3.16b, v12.4b[2]\n"
+        ".inst 0x4face87d  // sdot v29.4s, v3.16b, v12.4b[3]\n"
+        "ldr q12, [x21, #0x60]\n"
+        ".inst 0x4f89e2da  // sdot v26.4s, v22.16b, v9.4b[0]\n"
+        ".inst 0x4fa9e2ca  // sdot v10.4s, v22.16b, v9.4b[1]\n"
+        ".inst 0x4f89eac2  // sdot v2.4s, v22.16b, v9.4b[2]\n"
+        ".inst 0x4fa9eadd  // sdot v29.4s, v22.16b, v9.4b[3]\n"
+        "ldr q17, [x21, #0x70]\n"
+        "add x21, x21, #0x88\n"
+        ".inst 0x4f8ce37a  // sdot v26.4s, v27.16b, v12.4b[0]\n"
+        ".inst 0x4face36a  // sdot v10.4s, v27.16b, v12.4b[1]\n"
+        ".inst 0x4f8ceb62  // sdot v2.4s, v27.16b, v12.4b[2]\n"
+        ".inst 0x4faceb7d  // sdot v29.4s, v27.16b, v12.4b[3]\n"
+        ".inst 0x4f91e3da  // sdot v26.4s, v30.16b, v17.4b[0]\n"
+        ".inst 0x4fb1e3ca  // sdot v10.4s, v30.16b, v17.4b[1]\n"
+        ".inst 0x4f91ebc2  // sdot v2.4s, v30.16b, v17.4b[2]\n"
+        ".inst 0x4fb1ebdd  // sdot v29.4s, v30.16b, v17.4b[3]\n"
+        "scvtf v26.4s, v26.4s, #0x4\n"
+        "scvtf v10.4s, v10.4s, #0x4\n"
+        "fmla v5.4s, v26.4s, v24.4s\n"
+        "scvtf v2.4s, v2.4s, #0x4\n"
+        "scvtf v29.4s, v29.4s, #0x4\n"
+        "fmla v21.4s, v10.4s, v31.4s\n"
+        "fmla v8.4s, v2.4s, v6.4s\n"
+        "fmla v1.4s, v29.4s, v20.4s\n"
+        "bgt 3b\n"
+        "mov x20, %x[res_ptr]\n"
+        "subs x27, x27, #0x4\n"
+        "add %x[res_ptr], %x[res_ptr], #0x10\n"
+        "str q15, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "str q19, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "str q18, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "str q14, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "str q11, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "str q13, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "str q23, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "str q16, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "str q25, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "str q7, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "str q0, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "str q4, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "str q5, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "str q21, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "str q8, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "str q1, [x20, #0x0]\n"
+        "bne 2b\n"
+        "mov x20, #0x4\n"
+        "sub x10, x10, #0x10\n"
+        "cmp x10, #0x10\n"
+        "mov %x[res_ptr], x26\n"
+        "madd %x[a_ptr], x20, x9, %x[a_ptr]\n"
+        "bge 1b\n"
+        "4:"  // Row loop skip
+        "cbz x10, 9f\n"
+        "5:"  // Row tail: Row loop
+        "add x24, %x[b_ptr], #0x8\n"
+        "mov x23, %x[nc]\n"
+        "add x22, %x[res_ptr], %x[res_stride], LSL #2\n"
+        "6:"  // Row tail: Column loop
+        "movi v15.16b, #0x0\n"
+        "movi v19.16b, #0x0\n"
+        "add x25, %x[a_ptr], #0x8\n"
+        "mov x21, %x[nb]\n"
+        "movi v18.16b, #0x0\n"
+        "movi v14.16b, #0x0\n"
+        "7:"  // Row tail: Block loop
+        "ldr q7, [x24, #0x0]\n"
+        "ldr q5, [x25, #0x0]\n"
+        "movi v9.16b, #0x4\n"
+        "movi v4.4s, #0x0\n"
+        "ldr q3, [x24, #0x10]\n"
+        "ldr q2, [x25, #0x10]\n"
+        "movi v1.4s, #0x0\n"
+        "movi v0.4s, #0x0\n"
+        "ldr q13, [x24, #0x20]\n"
+        "ldr q31, [x25, #0x20]\n"
+        "movi v30.4s, #0x0\n"
+        "movi v29.16b, #0xf0\n"
+        "ldr q28, [x24, #0x30]\n"
+        "ldr q27, [x25, #0x30]\n"
+        "sshl v20.16b, v7.16b, v9.16b\n"
+        "sub x20, x24, #0x8\n"
+        "ldr q26, [x25, #0x40]\n"
+        "ldr q25, [x25, #0x50]\n"
+        "sshl v17.16b, v3.16b, v9.16b\n"
+        "and v7.16b, v7.16b, v29.16b\n"
+        "ldr q24, [x25, #0x60]\n"
+        "ldr q16, [x25, #0x70]\n"
+        "sshl v22.16b, v13.16b, v9.16b\n"
+        "and v3.16b, v3.16b, v29.16b\n"
+        "ldr d21, [x20, #0x0]\n"
+        "ldr d12, [x25, #-0x8]\n"
+        ".inst 0x4f85e284  // sdot v4.4s, v20.16b, v5.4b[0]\n"
+        ".inst 0x4fa5e281  // sdot v1.4s, v20.16b, v5.4b[1]\n"
+        ".inst 0x4f85ea80  // sdot v0.4s, v20.16b, v5.4b[2]\n"
+        ".inst 0x4fa5ea9e  // sdot v30.4s, v20.16b, v5.4b[3]\n"
+        "sshl v9.16b, v28.16b, v9.16b\n"
+        "subs x21, x21, #0x1\n"
+        "and v13.16b, v13.16b, v29.16b\n"
+        "and v28.16b, v28.16b, v29.16b\n"
+        "add x25, x25, #0x88\n"
+        "add x24, x24, #0x48\n"
+        "fcvtl v21.4s, v21.4h\n"
+        "fcvtl v12.4s, v12.4h\n"
+        ".inst 0x4f82e224  // sdot v4.4s, v17.16b, v2.4b[0]\n"
+        ".inst 0x4fa2e221  // sdot v1.4s, v17.16b, v2.4b[1]\n"
+        ".inst 0x4f82ea20  // sdot v0.4s, v17.16b, v2.4b[2]\n"
+        ".inst 0x4fa2ea3e  // sdot v30.4s, v17.16b, v2.4b[3]\n"
+        "fmul v11.4s, v21.4s, v12.s[0]\n"
+        "fmul v23.4s, v21.4s, v12.s[1]\n"
+        "fmul v17.4s, v21.4s, v12.s[2]\n"
+        ".inst 0x4f9fe2c4  // sdot v4.4s, v22.16b, v31.4b[0]\n"
+        "fmul v6.4s, v21.4s, v12.s[3]\n"
+        ".inst 0x4fbfe2c1  // sdot v1.4s, v22.16b, v31.4b[1]\n"
+        ".inst 0x4f9feac0  // sdot v0.4s, v22.16b, v31.4b[2]\n"
+        ".inst 0x4fbfeade  // sdot v30.4s, v22.16b, v31.4b[3]\n"
+        ".inst 0x4f9be124  // sdot v4.4s, v9.16b, v27.4b[0]\n"
+        ".inst 0x4fbbe121  // sdot v1.4s, v9.16b, v27.4b[1]\n"
+        ".inst 0x4f9be920  // sdot v0.4s, v9.16b, v27.4b[2]\n"
+        ".inst 0x4fbbe93e  // sdot v30.4s, v9.16b, v27.4b[3]\n"
+        ".inst 0x4f9ae0e4  // sdot v4.4s, v7.16b, v26.4b[0]\n"
+        ".inst 0x4fbae0e1  // sdot v1.4s, v7.16b, v26.4b[1]\n"
+        ".inst 0x4f9ae8e0  // sdot v0.4s, v7.16b, v26.4b[2]\n"
+        ".inst 0x4fbae8fe  // sdot v30.4s, v7.16b, v26.4b[3]\n"
+        ".inst 0x4f99e064  // sdot v4.4s, v3.16b, v25.4b[0]\n"
+        ".inst 0x4fb9e061  // sdot v1.4s, v3.16b, v25.4b[1]\n"
+        ".inst 0x4f99e860  // sdot v0.4s, v3.16b, v25.4b[2]\n"
+        ".inst 0x4fb9e87e  // sdot v30.4s, v3.16b, v25.4b[3]\n"
+        ".inst 0x4f98e1a4  // sdot v4.4s, v13.16b, v24.4b[0]\n"
+        ".inst 0x4fb8e1a1  // sdot v1.4s, v13.16b, v24.4b[1]\n"
+        ".inst 0x4f98e9a0  // sdot v0.4s, v13.16b, v24.4b[2]\n"
+        ".inst 0x4fb8e9be  // sdot v30.4s, v13.16b, v24.4b[3]\n"
+        ".inst 0x4f90e384  // sdot v4.4s, v28.16b, v16.4b[0]\n"
+        ".inst 0x4fb0e381  // sdot v1.4s, v28.16b, v16.4b[1]\n"
+        ".inst 0x4f90eb80  // sdot v0.4s, v28.16b, v16.4b[2]\n"
+        ".inst 0x4fb0eb9e  // sdot v30.4s, v28.16b, v16.4b[3]\n"
+        "scvtf v4.4s, v4.4s, #0x4\n"
+        "scvtf v1.4s, v1.4s, #0x4\n"
+        "scvtf v0.4s, v0.4s, #0x4\n"
+        "fmla v15.4s, v4.4s, v11.4s\n"
+        "scvtf v30.4s, v30.4s, #0x4\n"
+        "fmla v19.4s, v1.4s, v23.4s\n"
+        "fmla v18.4s, v0.4s, v17.4s\n"
+        "fmla v14.4s, v30.4s, v6.4s\n"
+        "bgt 7b\n"
+        "mov x20, %x[res_ptr]\n"
+        "cmp x10, #0x1\n"
+        "str q15, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "ble 8f\n"
+        "cmp x10, #0x2\n"
+        "str q19, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "ble 8f\n"
+        "cmp x10, #0x3\n"
+        "str q18, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "ble 8f\n"
+        "str q14, [x20, #0x0]\n"
+        "8:"  // Row tail: Accumulator store skip
+        "subs x23, x23, #0x4\n"
+        "add %x[res_ptr], %x[res_ptr], #0x10\n"
+        "bne 6b\n"
+        "subs x10, x10, #0x4\n"
+        "add %x[a_ptr], %x[a_ptr], x9\n"
+        "mov %x[res_ptr], x22\n"
+        "bgt 5b\n"
+        "9:"  // Row tail: Row loop skip
+        : [a_ptr] "+&r" (a_ptr), [res_ptr] "+&r" (res_ptr)
+        : [b_ptr] "r" (b_ptr), [nr] "r" (nr), [nb] "r" (nb), [res_stride] "r" (res_stride), [nc] "r" (nc)
+        : "cc", "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", "x9", "x10", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
+    );
+    return;
+#endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON)
+    {
+        float sumf[4][4];
+        int sumi;
+
+        for (int y = 0; y < nr / 4; y++) {
+            const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
+            for (int x = 0; x < nc / ncols_interleaved; x++) {
+                const block_q4_0x4 * b_ptr = (const block_q4_0x4 *) vx + (x * nb);
+                for (int m = 0; m < 4; m++) {
+                    for (int j = 0; j < ncols_interleaved; j++) sumf[m][j] = 0.0;
+                }
+                for (int l = 0; l < nb; l++) {
+                    for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                        for (int m = 0; m < 4; m++) {
+                            for (int j = 0; j < ncols_interleaved; j++) {
+                                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] << 4);
+                                    const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
+                                    sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
+                                            (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4;
+                                }
+                                sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(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];
+                }
+            }
+        }
+    }
+}
+
+void ggml_gemm_q4_0_4x8_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;
+    const int ncols_interleaved = 4;
+    const int blocklen = 8;
+
+    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(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_MATMUL_INT8)
+    const void * b_ptr = vx;
+    const void * a_ptr = vy;
+    float * res_ptr = s;
+    size_t res_stride = bs * sizeof(float);
+
+    __asm__ __volatile__(
+        "mov x10, %x[nr]\n"
+        "mov x9, #0x88\n"
+        "cmp x10, #0x10\n"
+        "mul x9, %x[nb], x9\n"
+        "blt 4f\n"
+        "1:"  // Row loop
+        "add x28, %x[b_ptr], #0x8\n"
+        "mov x27, %x[nc]\n"
+        "add x26, %x[res_ptr], %x[res_stride], LSL #4\n"
+        "2:"  // Column loop
+        "add x25, %x[a_ptr], #0x8\n"
+        "movi v2.16b, #0x0\n"
+        "movi v10.16b, #0x0\n"
+        "mov x24, %x[nb]\n"
+        "add x23, x25, x9\n"
+        "movi v12.16b, #0x0\n"
+        "movi v28.16b, #0x0\n"
+        "add x22, x23, x9\n"
+        "movi v11.16b, #0x0\n"
+        "movi v13.16b, #0x0\n"
+        "add x21, x22, x9\n"
+        "movi v22.16b, #0x0\n"
+        "movi v23.16b, #0x0\n"
+        "movi v25.16b, #0x0\n"
+        "movi v5.16b, #0x0\n"
+        "movi v7.16b, #0x0\n"
+        "movi v4.16b, #0x0\n"
+        "movi v6.16b, #0x0\n"
+        "movi v30.16b, #0x0\n"
+        "movi v24.16b, #0x0\n"
+        "movi v14.16b, #0x0\n"
+        "3:"  // Block loop
+        "ldr q21, [x28, #0x0]\n"
+        "ldr q16, [x28, #0x10]\n"
+        "movi v1.16b, #0x4\n"
+        "movi v19.4s, #0x0\n"
+        "ldr q27, [x25, #0x0]\n"
+        "ldr q15, [x25, #0x10]\n"
+        "movi v26.4s, #0x0\n"
+        "movi v18.4s, #0x0\n"
+        "ldr q29, [x28, #0x20]\n"
+        "ldr q3, [x28, #0x30]\n"
+        "movi v17.4s, #0x0\n"
+        "movi v0.16b, #0xf0\n"
+        "ldr d20, [x25, #-0x8]\n"
+        "ldr d9, [x23, #-0x8]\n"
+        "sshl v8.16b, v21.16b, v1.16b\n"
+        "sshl v31.16b, v16.16b, v1.16b\n"
+        "and v21.16b, v21.16b, v0.16b\n"
+        "and v16.16b, v16.16b, v0.16b\n"
+        "sub x20, x28, #0x8\n"
+        "subs x24, x24, #0x1\n"
+        "add x28, x28, #0x48\n"
+        ".inst 0x4e88a773  // smmla v19.4s, v27.16b, v8.16b\n"
+        ".inst 0x4e9fa77a  // smmla v26.4s, v27.16b, v31.16b\n"
+        "ldr q27, [x25, #0x20]\n"
+        ".inst 0x4e88a5f2  // smmla v18.4s, v15.16b, v8.16b\n"
+        ".inst 0x4e9fa5f1  // smmla v17.4s, v15.16b, v31.16b\n"
+        "sshl v15.16b, v29.16b, v1.16b\n"
+        "sshl v1.16b, v3.16b, v1.16b\n"
+        "and v29.16b, v29.16b, v0.16b\n"
+        "and v3.16b, v3.16b, v0.16b\n"
+        "ldr q0, [x25, #0x30]\n"
+        "fcvtl v20.4s, v20.4h\n"
+        ".inst 0x4e8fa773  // smmla v19.4s, v27.16b, v15.16b\n"
+        "fcvtl v9.4s, v9.4h\n"
+        ".inst 0x4e81a77a  // smmla v26.4s, v27.16b, v1.16b\n"
+        "ldr q27, [x25, #0x40]\n"
+        ".inst 0x4e8fa412  // smmla v18.4s, v0.16b, v15.16b\n"
+        ".inst 0x4e81a411  // smmla v17.4s, v0.16b, v1.16b\n"
+        "ldr q0, [x25, #0x50]\n"
+        ".inst 0x4e95a773  // smmla v19.4s, v27.16b, v21.16b\n"
+        ".inst 0x4e90a77a  // smmla v26.4s, v27.16b, v16.16b\n"
+        "ldr q27, [x25, #0x60]\n"
+        ".inst 0x4e95a412  // smmla v18.4s, v0.16b, v21.16b\n"
+        ".inst 0x4e90a411  // smmla v17.4s, v0.16b, v16.16b\n"
+        "ldr q0, [x25, #0x70]\n"
+        "add x25, x25, #0x88\n"
+        ".inst 0x4e9da773  // smmla v19.4s, v27.16b, v29.16b\n"
+        ".inst 0x4e83a77a  // smmla v26.4s, v27.16b, v3.16b\n"
+        "ldr d27, [x20, #0x0]\n"
+        ".inst 0x4e9da412  // smmla v18.4s, v0.16b, v29.16b\n"
+        ".inst 0x4e83a411  // smmla v17.4s, v0.16b, v3.16b\n"
+        "fcvtl v27.4s, v27.4h\n"
+        "uzp1 v0.2d, v19.2d, v26.2d\n"
+        "uzp2 v26.2d, v19.2d, v26.2d\n"
+        "fmul v19.4s, v27.4s, v20.s[0]\n"
+        "scvtf v0.4s, v0.4s, #0x4\n"
+        "scvtf v26.4s, v26.4s, #0x4\n"
+        "fmla v2.4s, v0.4s, v19.4s\n"
+        "ldr q19, [x23, #0x0]\n"
+        "uzp1 v0.2d, v18.2d, v17.2d\n"
+        "uzp2 v18.2d, v18.2d, v17.2d\n"
+        "fmul v17.4s, v27.4s, v20.s[1]\n"
+        "scvtf v0.4s, v0.4s, #0x4\n"
+        "scvtf v18.4s, v18.4s, #0x4\n"
+        "fmla v10.4s, v26.4s, v17.4s\n"
+        "ldr q17, [x23, #0x10]\n"
+        "fmul v26.4s, v27.4s, v20.s[2]\n"
+        "fmul v20.4s, v27.4s, v20.s[3]\n"
+        "fmla v12.4s, v0.4s, v26.4s\n"
+        "ldr d0, [x22, #-0x8]\n"
+        "ldr d26, [x21, #-0x8]\n"
+        "fcvtl v0.4s, v0.4h\n"
+        "fmla v28.4s, v18.4s, v20.4s\n"
+        "movi v20.4s, #0x0\n"
+        "movi v18.4s, #0x0\n"
+        ".inst 0x4e88a674  // smmla v20.4s, v19.16b, v8.16b\n"
+        ".inst 0x4e9fa672  // smmla v18.4s, v19.16b, v31.16b\n"
+        "ldr q19, [x23, #0x20]\n"
+        "fcvtl v26.4s, v26.4h\n"
+        ".inst 0x4e8fa674  // smmla v20.4s, v19.16b, v15.16b\n"
+        ".inst 0x4e81a672  // smmla v18.4s, v19.16b, v1.16b\n"
+        "ldr q19, [x23, #0x40]\n"
+        ".inst 0x4e95a674  // smmla v20.4s, v19.16b, v21.16b\n"
+        ".inst 0x4e90a672  // smmla v18.4s, v19.16b, v16.16b\n"
+        "ldr q19, [x23, #0x60]\n"
+        ".inst 0x4e9da674  // smmla v20.4s, v19.16b, v29.16b\n"
+        ".inst 0x4e83a672  // smmla v18.4s, v19.16b, v3.16b\n"
+        "uzp1 v19.2d, v20.2d, v18.2d\n"
+        "scvtf v19.4s, v19.4s, #0x4\n"
+        "uzp2 v20.2d, v20.2d, v18.2d\n"
+        "fmul v18.4s, v27.4s, v9.s[0]\n"
+        "scvtf v20.4s, v20.4s, #0x4\n"
+        "fmla v11.4s, v19.4s, v18.4s\n"
+        "ldr q18, [x22, #0x0]\n"
+        "fmul v19.4s, v27.4s, v9.s[1]\n"
+        "fmla v13.4s, v20.4s, v19.4s\n"
+        "movi v19.4s, #0x0\n"
+        "movi v20.4s, #0x0\n"
+        ".inst 0x4e88a633  // smmla v19.4s, v17.16b, v8.16b\n"
+        ".inst 0x4e9fa634  // smmla v20.4s, v17.16b, v31.16b\n"
+        "ldr q17, [x23, #0x30]\n"
+        ".inst 0x4e8fa633  // smmla v19.4s, v17.16b, v15.16b\n"
+        ".inst 0x4e81a634  // smmla v20.4s, v17.16b, v1.16b\n"
+        "ldr q17, [x23, #0x50]\n"
+        ".inst 0x4e95a633  // smmla v19.4s, v17.16b, v21.16b\n"
+        ".inst 0x4e90a634  // smmla v20.4s, v17.16b, v16.16b\n"
+        "ldr q17, [x23, #0x70]\n"
+        "add x23, x23, #0x88\n"
+        ".inst 0x4e9da633  // smmla v19.4s, v17.16b, v29.16b\n"
+        ".inst 0x4e83a634  // smmla v20.4s, v17.16b, v3.16b\n"
+        "uzp1 v17.2d, v19.2d, v20.2d\n"
+        "scvtf v17.4s, v17.4s, #0x4\n"
+        "uzp2 v20.2d, v19.2d, v20.2d\n"
+        "fmul v19.4s, v27.4s, v9.s[2]\n"
+        "fmul v9.4s, v27.4s, v9.s[3]\n"
+        "scvtf v20.4s, v20.4s, #0x4\n"
+        "fmla v22.4s, v17.4s, v19.4s\n"
+        "ldr q17, [x22, #0x10]\n"
+        "movi v19.4s, #0x0\n"
+        ".inst 0x4e88a653  // smmla v19.4s, v18.16b, v8.16b\n"
+        "fmla v23.4s, v20.4s, v9.4s\n"
+        "movi v20.4s, #0x0\n"
+        "movi v9.4s, #0x0\n"
+        ".inst 0x4e9fa654  // smmla v20.4s, v18.16b, v31.16b\n"
+        "ldr q18, [x22, #0x20]\n"
+        ".inst 0x4e88a629  // smmla v9.4s, v17.16b, v8.16b\n"
+        ".inst 0x4e8fa653  // smmla v19.4s, v18.16b, v15.16b\n"
+        ".inst 0x4e81a654  // smmla v20.4s, v18.16b, v1.16b\n"
+        "ldr q18, [x22, #0x40]\n"
+        ".inst 0x4e95a653  // smmla v19.4s, v18.16b, v21.16b\n"
+        ".inst 0x4e90a654  // smmla v20.4s, v18.16b, v16.16b\n"
+        "ldr q18, [x22, #0x60]\n"
+        ".inst 0x4e9da653  // smmla v19.4s, v18.16b, v29.16b\n"
+        ".inst 0x4e83a654  // smmla v20.4s, v18.16b, v3.16b\n"
+        "movi v18.4s, #0x0\n"
+        ".inst 0x4e9fa632  // smmla v18.4s, v17.16b, v31.16b\n"
+        "ldr q17, [x22, #0x30]\n"
+        ".inst 0x4e8fa629  // smmla v9.4s, v17.16b, v15.16b\n"
+        ".inst 0x4e81a632  // smmla v18.4s, v17.16b, v1.16b\n"
+        "ldr q17, [x22, #0x50]\n"
+        ".inst 0x4e95a629  // smmla v9.4s, v17.16b, v21.16b\n"
+        ".inst 0x4e90a632  // smmla v18.4s, v17.16b, v16.16b\n"
+        "ldr q17, [x22, #0x70]\n"
+        "add x22, x22, #0x88\n"
+        ".inst 0x4e9da629  // smmla v9.4s, v17.16b, v29.16b\n"
+        ".inst 0x4e83a632  // smmla v18.4s, v17.16b, v3.16b\n"
+        "uzp1 v17.2d, v19.2d, v20.2d\n"
+        "uzp2 v20.2d, v19.2d, v20.2d\n"
+        "fmul v19.4s, v27.4s, v0.s[0]\n"
+        "scvtf v17.4s, v17.4s, #0x4\n"
+        "scvtf v20.4s, v20.4s, #0x4\n"
+        "fmla v25.4s, v17.4s, v19.4s\n"
+        "ldr q19, [x21, #0x0]\n"
+        "fmul v17.4s, v27.4s, v0.s[1]\n"
+        "fmla v5.4s, v20.4s, v17.4s\n"
+        "ldr q17, [x21, #0x10]\n"
+        "uzp1 v20.2d, v9.2d, v18.2d\n"
+        "uzp2 v9.2d, v9.2d, v18.2d\n"
+        "fmul v18.4s, v27.4s, v0.s[2]\n"
+        "fmul v0.4s, v27.4s, v0.s[3]\n"
+        "scvtf v20.4s, v20.4s, #0x4\n"
+        "scvtf v9.4s, v9.4s, #0x4\n"
+        "fmla v7.4s, v20.4s, v18.4s\n"
+        "movi v20.4s, #0x0\n"
+        "movi v18.4s, #0x0\n"
+        ".inst 0x4e88a674  // smmla v20.4s, v19.16b, v8.16b\n"
+        ".inst 0x4e9fa672  // smmla v18.4s, v19.16b, v31.16b\n"
+        "ldr q19, [x21, #0x20]\n"
+        "fmla v4.4s, v9.4s, v0.4s\n"
+        "movi v9.4s, #0x0\n"
+        "movi v0.4s, #0x0\n"
+        ".inst 0x4e88a629  // smmla v9.4s, v17.16b, v8.16b\n"
+        "fmul v8.4s, v27.4s, v26.s[0]\n"
+        ".inst 0x4e9fa620  // smmla v0.4s, v17.16b, v31.16b\n"
+        "ldr q17, [x21, #0x30]\n"
+        ".inst 0x4e8fa674  // smmla v20.4s, v19.16b, v15.16b\n"
+        "fmul v31.4s, v27.4s, v26.s[1]\n"
+        ".inst 0x4e81a672  // smmla v18.4s, v19.16b, v1.16b\n"
+        "ldr q19, [x21, #0x40]\n"
+        ".inst 0x4e8fa629  // smmla v9.4s, v17.16b, v15.16b\n"
+        "fmul v15.4s, v27.4s, v26.s[2]\n"
+        "fmul v27.4s, v27.4s, v26.s[3]\n"
+        ".inst 0x4e81a620  // smmla v0.4s, v17.16b, v1.16b\n"
+        "ldr q1, [x21, #0x50]\n"
+        ".inst 0x4e95a674  // smmla v20.4s, v19.16b, v21.16b\n"
+        ".inst 0x4e90a672  // smmla v18.4s, v19.16b, v16.16b\n"
+        "ldr q26, [x21, #0x60]\n"
+        ".inst 0x4e95a429  // smmla v9.4s, v1.16b, v21.16b\n"
+        ".inst 0x4e90a420  // smmla v0.4s, v1.16b, v16.16b\n"
+        "ldr q21, [x21, #0x70]\n"
+        "add x21, x21, #0x88\n"
+        ".inst 0x4e9da754  // smmla v20.4s, v26.16b, v29.16b\n"
+        ".inst 0x4e83a752  // smmla v18.4s, v26.16b, v3.16b\n"
+        ".inst 0x4e9da6a9  // smmla v9.4s, v21.16b, v29.16b\n"
+        ".inst 0x4e83a6a0  // smmla v0.4s, v21.16b, v3.16b\n"
+        "uzp1 v29.2d, v20.2d, v18.2d\n"
+        "uzp2 v21.2d, v20.2d, v18.2d\n"
+        "scvtf v29.4s, v29.4s, #0x4\n"
+        "uzp1 v18.2d, v9.2d, v0.2d\n"
+        "uzp2 v16.2d, v9.2d, v0.2d\n"
+        "scvtf v21.4s, v21.4s, #0x4\n"
+        "fmla v6.4s, v29.4s, v8.4s\n"
+        "scvtf v18.4s, v18.4s, #0x4\n"
+        "scvtf v16.4s, v16.4s, #0x4\n"
+        "fmla v30.4s, v21.4s, v31.4s\n"
+        "fmla v24.4s, v18.4s, v15.4s\n"
+        "fmla v14.4s, v16.4s, v27.4s\n"
+        "bgt 3b\n"
+        "mov x20, %x[res_ptr]\n"
+        "subs x27, x27, #0x4\n"
+        "add %x[res_ptr], %x[res_ptr], #0x10\n"
+        "str q2, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "str q10, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "str q12, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "str q28, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "str q11, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "str q13, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "str q22, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "str q23, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "str q25, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "str q5, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "str q7, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "str q4, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "str q6, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "str q30, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "str q24, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "str q14, [x20, #0x0]\n"
+        "bne 2b\n"
+        "mov x20, #0x4\n"
+        "sub x10, x10, #0x10\n"
+        "cmp x10, #0x10\n"
+        "mov %x[res_ptr], x26\n"
+        "madd %x[a_ptr], x20, x9, %x[a_ptr]\n"
+        "bge 1b\n"
+        "4:"  // Row loop skip
+        "cbz x10, 9f\n"
+        "5:"  // Row tail: Row loop
+        "add x24, %x[b_ptr], #0x8\n"
+        "mov x23, %x[nc]\n"
+        "add x22, %x[res_ptr], %x[res_stride], LSL #2\n"
+        "6:"  // Row tail: Column loop
+        "movi v2.16b, #0x0\n"
+        "movi v10.16b, #0x0\n"
+        "add x25, %x[a_ptr], #0x8\n"
+        "mov x21, %x[nb]\n"
+        "movi v12.16b, #0x0\n"
+        "movi v28.16b, #0x0\n"
+        "7:"  // Row tail: Block loop
+        "ldr q6, [x24, #0x0]\n"
+        "ldr q5, [x24, #0x10]\n"
+        "movi v17.16b, #0x4\n"
+        "movi v8.4s, #0x0\n"
+        "ldr q4, [x25, #0x0]\n"
+        "ldr q13, [x25, #0x10]\n"
+        "movi v27.4s, #0x0\n"
+        "movi v0.4s, #0x0\n"
+        "ldr q31, [x24, #0x20]\n"
+        "ldr q14, [x24, #0x30]\n"
+        "movi v29.4s, #0x0\n"
+        "movi v22.16b, #0xf0\n"
+        "ldr q11, [x25, #0x20]\n"
+        "ldr q23, [x25, #0x30]\n"
+        "sshl v21.16b, v6.16b, v17.16b\n"
+        "sshl v16.16b, v5.16b, v17.16b\n"
+        "ldr q20, [x25, #0x40]\n"
+        "ldr q26, [x25, #0x50]\n"
+        "and v6.16b, v6.16b, v22.16b\n"
+        "and v5.16b, v5.16b, v22.16b\n"
+        "ldr q25, [x25, #0x60]\n"
+        "ldr q3, [x25, #0x70]\n"
+        "sshl v19.16b, v31.16b, v17.16b\n"
+        "sshl v18.16b, v14.16b, v17.16b\n"
+        "ldr d17, [x25, #-0x8]\n"
+        ".inst 0x4e95a488  // smmla v8.4s, v4.16b, v21.16b\n"
+        ".inst 0x4e90a49b  // smmla v27.4s, v4.16b, v16.16b\n"
+        "and v31.16b, v31.16b, v22.16b\n"
+        ".inst 0x4e95a5a0  // smmla v0.4s, v13.16b, v21.16b\n"
+        ".inst 0x4e90a5bd  // smmla v29.4s, v13.16b, v16.16b\n"
+        "and v14.16b, v14.16b, v22.16b\n"
+        "sub x20, x24, #0x8\n"
+        "ldr d16, [x20, #0x0]\n"
+        "subs x21, x21, #0x1\n"
+        "add x25, x25, #0x88\n"
+        "fcvtl v17.4s, v17.4h\n"
+        "add x24, x24, #0x48\n"
+        ".inst 0x4e93a568  // smmla v8.4s, v11.16b, v19.16b\n"
+        ".inst 0x4e92a57b  // smmla v27.4s, v11.16b, v18.16b\n"
+        ".inst 0x4e93a6e0  // smmla v0.4s, v23.16b, v19.16b\n"
+        ".inst 0x4e92a6fd  // smmla v29.4s, v23.16b, v18.16b\n"
+        "fcvtl v16.4s, v16.4h\n"
+        ".inst 0x4e86a688  // smmla v8.4s, v20.16b, v6.16b\n"
+        ".inst 0x4e85a69b  // smmla v27.4s, v20.16b, v5.16b\n"
+        "fmul v23.4s, v16.4s, v17.s[0]\n"
+        "fmul v21.4s, v16.4s, v17.s[1]\n"
+        "fmul v1.4s, v16.4s, v17.s[2]\n"
+        "fmul v20.4s, v16.4s, v17.s[3]\n"
+        ".inst 0x4e86a740  // smmla v0.4s, v26.16b, v6.16b\n"
+        ".inst 0x4e85a75d  // smmla v29.4s, v26.16b, v5.16b\n"
+        ".inst 0x4e9fa728  // smmla v8.4s, v25.16b, v31.16b\n"
+        ".inst 0x4e8ea73b  // smmla v27.4s, v25.16b, v14.16b\n"
+        ".inst 0x4e9fa460  // smmla v0.4s, v3.16b, v31.16b\n"
+        ".inst 0x4e8ea47d  // smmla v29.4s, v3.16b, v14.16b\n"
+        "uzp1 v19.2d, v8.2d, v27.2d\n"
+        "uzp2 v18.2d, v8.2d, v27.2d\n"
+        "scvtf v19.4s, v19.4s, #0x4\n"
+        "uzp1 v17.2d, v0.2d, v29.2d\n"
+        "uzp2 v16.2d, v0.2d, v29.2d\n"
+        "scvtf v18.4s, v18.4s, #0x4\n"
+        "fmla v2.4s, v19.4s, v23.4s\n"
+        "scvtf v17.4s, v17.4s, #0x4\n"
+        "scvtf v16.4s, v16.4s, #0x4\n"
+        "fmla v10.4s, v18.4s, v21.4s\n"
+        "fmla v12.4s, v17.4s, v1.4s\n"
+        "fmla v28.4s, v16.4s, v20.4s\n"
+        "bgt 7b\n"
+        "mov x20, %x[res_ptr]\n"
+        "cmp x10, #0x1\n"
+        "str q2, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "ble 8f\n"
+        "cmp x10, #0x2\n"
+        "str q10, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "ble 8f\n"
+        "cmp x10, #0x3\n"
+        "str q12, [x20, #0x0]\n"
+        "add x20, x20, %x[res_stride]\n"
+        "ble 8f\n"
+        "str q28, [x20, #0x0]\n"
+        "8:"  // Row tail: Accumulator store skip
+        "subs x23, x23, #0x4\n"
+        "add %x[res_ptr], %x[res_ptr], #0x10\n"
+        "bne 6b\n"
+        "subs x10, x10, #0x4\n"
+        "add %x[a_ptr], %x[a_ptr], x9\n"
+        "mov %x[res_ptr], x22\n"
+        "bgt 5b\n"
+        "9:"  // Row tail: Row loop skip
+        : [a_ptr] "+&r" (a_ptr), [res_ptr] "+&r" (res_ptr)
+        : [b_ptr] "r" (b_ptr), [nr] "r" (nr), [nb] "r" (nb), [res_stride] "r" (res_stride), [nc] "r" (nc)
+        : "cc", "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", "x9", "x10", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
+    );
+    return;
+#endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_MATMUL_INT8)
+    float sumf[4][4];
+    int sumi;
+
+    for (int y = 0; y < nr / 4; y++) {
+        const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
+        for (int x = 0; x < nc / ncols_interleaved; x++) {
+            const block_q4_0x4 * b_ptr = (const block_q4_0x4 *) vx + (x * nb);
+            for (int m = 0; m < 4; m++) {
+                for (int j = 0; j < ncols_interleaved; j++) sumf[m][j] = 0.0;
+            }
+            for (int l = 0; l < nb; l++) {
+                for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                    for (int m = 0; m < 4; m++) {
+                        for (int j = 0; j < ncols_interleaved; j++) {
+                            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] << 4);
+                                const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
+                                sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
+                                        (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4;
+                            }
+                            sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(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];
+            }
+        }
+    }
+}
+
+void ggml_gemm_q4_0_8x8_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;
+    const int ncols_interleaved = 8;
+    const int blocklen = 8;
+
+    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(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__)
+#if defined(__ARM_FEATURE_SVE) && defined(__ARM_FEATURE_MATMUL_INT8)
+    if (ggml_cpu_get_sve_cnt() == QK8_0) {
+        const void * b_ptr = vx;
+        const void * a_ptr = vy;
+        float * res_ptr = s;
+        size_t res_stride = bs * sizeof(float);
+
+        __asm__ __volatile__(
+            "mov x20, #0x4\n"
+            "mov x13, %x[nr]\n"
+            "mov z28.s, #-0x4\n"
+            "mov x12, #0x88\n"
+            "ptrue p1.b\n"
+            "whilelt p0.s, XZR, x20\n"
+            "cmp x13, #0x10\n"
+            "mul x12, %x[nb], x12\n"
+            "blt 4f\n"
+            "1:"  // Row loop
+            "add x11, %x[b_ptr], #0x10\n"
+            "mov x10, %x[nc]\n"
+            "add x9, %x[res_ptr], %x[res_stride], LSL #4\n"
+            "2:"  // Column loop
+            "add x28, %x[a_ptr], #0x8\n"
+            "mov z24.b, #0x0\n"
+            "mov z15.b, #0x0\n"
+            "mov x27, %x[nb]\n"
+            "add x26, x28, x12\n"
+            "mov z12.b, #0x0\n"
+            "mov z0.b, #0x0\n"
+            "add x25, x26, x12\n"
+            "mov z13.b, #0x0\n"
+            "mov z1.b, #0x0\n"
+            "add x24, x25, x12\n"
+            "mov z20.b, #0x0\n"
+            "mov z25.b, #0x0\n"
+            "mov z11.b, #0x0\n"
+            "mov z16.b, #0x0\n"
+            "mov z19.b, #0x0\n"
+            "mov z26.b, #0x0\n"
+            "mov z8.b, #0x0\n"
+            "mov z29.b, #0x0\n"
+            "mov z27.b, #0x0\n"
+            "mov z10.b, #0x0\n"
+            "3:"  // Block loop
+            "ld1b { z30.b }, p1/Z, [x11]\n"
+            "ld1b { z21.b }, p1/Z, [x11, #1, MUL VL]\n"
+            "mov z18.s, #0x0\n"
+            "mov z7.s, #0x0\n"
+            "ld1rqb { z3.b }, p1/Z, [x28]\n"
+            "ld1rqb { z5.b }, p1/Z, [x28, #16]\n"
+            "mov z9.s, #0x0\n"
+            "mov z22.s, #0x0\n"
+            "ld1b { z4.b }, p1/Z, [x11, #2, MUL VL]\n"
+            "ld1b { z17.b }, p1/Z, [x11, #3, MUL VL]\n"
+            "sub x20, x11, #0x10\n"
+            "sub x23, x28, #0x8\n"
+            "lsl z31.b, z30.b, #0x4\n"
+            "lsl z6.b, z21.b, #0x4\n"
+            "ld1h { z23.s }, p1/Z, [x20]\n"
+            "sub x22, x26, #0x8\n"
+            "and z30.b, z30.b, #0xf0\n"
+            "and z21.b, z21.b, #0xf0\n"
+            "sub x21, x25, #0x8\n"
+            "sub x20, x24, #0x8\n"
+            "lsl z14.b, z4.b, #0x4\n"
+            "lsl z2.b, z17.b, #0x4\n"
+            "subs x27, x27, #0x1\n"
+            "add x11, x11, #0x90\n"
+            ".inst 0x451f9872  // smmla z18.s, z3.b, z31.b\n"
+            ".inst 0x45069867  // smmla z7.s, z3.b, z6.b\n"
+            "ld1rqb { z3.b }, p1/Z, [x28, #32]\n"
+            "and z4.b, z4.b, #0xf0\n"
+            ".inst 0x451f98a9  // smmla z9.s, z5.b, z31.b\n"
+            ".inst 0x450698b6  // smmla z22.s, z5.b, z6.b\n"
+            "ld1rqb { z5.b }, p1/Z, [x28, #48]\n"
+            "and z17.b, z17.b, #0xf0\n"
+            "fcvt z23.s, p1/m, z23.h\n"
+            ".inst 0x450e9872  // smmla z18.s, z3.b, z14.b\n"
+            ".inst 0x45029867  // smmla z7.s, z3.b, z2.b\n"
+            "ld1rqb { z3.b }, p1/Z, [x28, #64]\n"
+            ".inst 0x450e98a9  // smmla z9.s, z5.b, z14.b\n"
+            ".inst 0x450298b6  // smmla z22.s, z5.b, z2.b\n"
+            "ld1rqb { z5.b }, p1/Z, [x28, #80]\n"
+            "fscale z23.s, p1/m, z23.s, z28.s\n"
+            ".inst 0x451e9872  // smmla z18.s, z3.b, z30.b\n"
+            ".inst 0x45159867  // smmla z7.s, z3.b, z21.b\n"
+            "ld1rqb { z3.b }, p1/Z, [x28, #96]\n"
+            ".inst 0x451e98a9  // smmla z9.s, z5.b, z30.b\n"
+            ".inst 0x451598b6  // smmla z22.s, z5.b, z21.b\n"
+            "ld1rqb { z5.b }, p1/Z, [x28, #112]\n"
+            "add x28, x28, #0x88\n"
+            ".inst 0x45049872  // smmla z18.s, z3.b, z4.b\n"
+            ".inst 0x45119867  // smmla z7.s, z3.b, z17.b\n"
+            "ld1h { z3.s }, p0/Z, [x23]\n"
+            ".inst 0x450498a9  // smmla z9.s, z5.b, z4.b\n"
+            ".inst 0x451198b6  // smmla z22.s, z5.b, z17.b\n"
+            "fcvt z3.s, p1/m, z3.h\n"
+            "uzp1 z5.d, z18.d, z7.d\n"
+            "uzp2 z18.d, z18.d, z7.d\n"
+            "mov z3.q, z3.q[0]\n"
+            "uzp1 z7.d, z9.d, z22.d\n"
+            "uzp2 z22.d, z9.d, z22.d\n"
+            "fmul z9.s, z23.s, z3.s[0]\n"
+            "scvtf z5.s, p1/m, z5.s\n"
+            "scvtf z18.s, p1/m, z18.s\n"
+            "scvtf z7.s, p1/m, z7.s\n"
+            "scvtf z22.s, p1/m, z22.s\n"
+            "fmla z24.s, p1/M, z5.s, z9.s\n"
+            "ld1rqb { z5.b }, p1/Z, [x26]\n"
+            "fmul z9.s, z23.s, z3.s[1]\n"
+            "fmla z15.s, p1/M, z18.s, z9.s\n"
+            "ld1rqb { z18.b }, p1/Z, [x26, #16]\n"
+            "fmul z9.s, z23.s, z3.s[2]\n"
+            "fmul z3.s, z23.s, z3.s[3]\n"
+            "fmla z12.s, p1/M, z7.s, z9.s\n"
+            "mov z9.s, #0x0\n"
+            "ld1h { z7.s }, p0/Z, [x22]\n"
+            ".inst 0x451f98a9  // smmla z9.s, z5.b, z31.b\n"
+            "fmla z0.s, p1/M, z22.s, z3.s\n"
+            "mov z22.s, #0x0\n"
+            "ld1h { z3.s }, p0/Z, [x21]\n"
+            ".inst 0x450698b6  // smmla z22.s, z5.b, z6.b\n"
+            "ld1rqb { z5.b }, p1/Z, [x26, #32]\n"
+            "fcvt z7.s, p1/m, z7.h\n"
+            "fcvt z3.s, p1/m, z3.h\n"
+            ".inst 0x450e98a9  // smmla z9.s, z5.b, z14.b\n"
+            ".inst 0x450298b6  // smmla z22.s, z5.b, z2.b\n"
+            "ld1rqb { z5.b }, p1/Z, [x26, #64]\n"
+            "mov z7.q, z7.q[0]\n"
+            "mov z3.q, z3.q[0]\n"
+            ".inst 0x451e98a9  // smmla z9.s, z5.b, z30.b\n"
+            ".inst 0x451598b6  // smmla z22.s, z5.b, z21.b\n"
+            "ld1rqb { z5.b }, p1/Z, [x26, #96]\n"
+            ".inst 0x450498a9  // smmla z9.s, z5.b, z4.b\n"
+            ".inst 0x451198b6  // smmla z22.s, z5.b, z17.b\n"
+            "uzp1 z5.d, z9.d, z22.d\n"
+            "scvtf z5.s, p1/m, z5.s\n"
+            "uzp2 z22.d, z9.d, z22.d\n"
+            "fmul z9.s, z23.s, z7.s[0]\n"
+            "scvtf z22.s, p1/m, z22.s\n"
+            "fmla z13.s, p1/M, z5.s, z9.s\n"
+            "ld1rqb { z9.b }, p1/Z, [x25]\n"
+            "fmul z5.s, z23.s, z7.s[1]\n"
+            "fmla z1.s, p1/M, z22.s, z5.s\n"
+            "mov z5.s, #0x0\n"
+            "mov z22.s, #0x0\n"
+            ".inst 0x451f9a45  // smmla z5.s, z18.b, z31.b\n"
+            ".inst 0x45069a56  // smmla z22.s, z18.b, z6.b\n"
+            "ld1rqb { z18.b }, p1/Z, [x26, #48]\n"
+            ".inst 0x450e9a45  // smmla z5.s, z18.b, z14.b\n"
+            ".inst 0x45029a56  // smmla z22.s, z18.b, z2.b\n"
+            "ld1rqb { z18.b }, p1/Z, [x26, #80]\n"
+            ".inst 0x451e9a45  // smmla z5.s, z18.b, z30.b\n"
+            ".inst 0x45159a56  // smmla z22.s, z18.b, z21.b\n"
+            "ld1rqb { z18.b }, p1/Z, [x26, #112]\n"
+            "add x26, x26, #0x88\n"
+            ".inst 0x45049a45  // smmla z5.s, z18.b, z4.b\n"
+            ".inst 0x45119a56  // smmla z22.s, z18.b, z17.b\n"
+            "uzp1 z18.d, z5.d, z22.d\n"
+            "scvtf z18.s, p1/m, z18.s\n"
+            "uzp2 z22.d, z5.d, z22.d\n"
+            "fmul z5.s, z23.s, z7.s[2]\n"
+            "fmul z7.s, z23.s, z7.s[3]\n"
+            "scvtf z22.s, p1/m, z22.s\n"
+            "fmla z20.s, p1/M, z18.s, z5.s\n"
+            "ld1rqb { z18.b }, p1/Z, [x25, #16]\n"
+            "ld1h { z5.s }, p0/Z, [x20]\n"
+            "fcvt z5.s, p1/m, z5.h\n"
+            "fmla z25.s, p1/M, z22.s, z7.s\n"
+            "mov z22.s, #0x0\n"
+            "mov z7.s, #0x0\n"
+            ".inst 0x451f9936  // smmla z22.s, z9.b, z31.b\n"
+            ".inst 0x45069927  // smmla z7.s, z9.b, z6.b\n"
+            "ld1rqb { z9.b }, p1/Z, [x25, #32]\n"
+            "mov z5.q, z5.q[0]\n"
+            ".inst 0x450e9936  // smmla z22.s, z9.b, z14.b\n"
+            ".inst 0x45029927  // smmla z7.s, z9.b, z2.b\n"
+            "ld1rqb { z9.b }, p1/Z, [x25, #64]\n"
+            ".inst 0x451e9936  // smmla z22.s, z9.b, z30.b\n"
+            ".inst 0x45159927  // smmla z7.s, z9.b, z21.b\n"
+            "ld1rqb { z9.b }, p1/Z, [x25, #96]\n"
+            ".inst 0x45049936  // smmla z22.s, z9.b, z4.b\n"
+            ".inst 0x45119927  // smmla z7.s, z9.b, z17.b\n"
+            "uzp1 z9.d, z22.d, z7.d\n"
+            "scvtf z9.s, p1/m, z9.s\n"
+            "uzp2 z22.d, z22.d, z7.d\n"
+            "fmul z7.s, z23.s, z3.s[0]\n"
+            "scvtf z22.s, p1/m, z22.s\n"
+            "fmla z11.s, p1/M, z9.s, z7.s\n"
+            "ld1rqb { z9.b }, p1/Z, [x24]\n"
+            "fmul z7.s, z23.s, z3.s[1]\n"
+            "fmla z16.s, p1/M, z22.s, z7.s\n"
+            "mov z22.s, #0x0\n"
+            "mov z7.s, #0x0\n"
+            ".inst 0x451f9a56  // smmla z22.s, z18.b, z31.b\n"
+            ".inst 0x45069a47  // smmla z7.s, z18.b, z6.b\n"
+            "ld1rqb { z18.b }, p1/Z, [x25, #48]\n"
+            ".inst 0x450e9a56  // smmla z22.s, z18.b, z14.b\n"
+            ".inst 0x45029a47  // smmla z7.s, z18.b, z2.b\n"
+            "ld1rqb { z18.b }, p1/Z, [x25, #80]\n"
+            ".inst 0x451e9a56  // smmla z22.s, z18.b, z30.b\n"
+            ".inst 0x45159a47  // smmla z7.s, z18.b, z21.b\n"
+            "ld1rqb { z18.b }, p1/Z, [x25, #112]\n"
+            "add x25, x25, #0x88\n"
+            ".inst 0x45049a56  // smmla z22.s, z18.b, z4.b\n"
+            ".inst 0x45119a47  // smmla z7.s, z18.b, z17.b\n"
+            "uzp1 z18.d, z22.d, z7.d\n"
+            "scvtf z18.s, p1/m, z18.s\n"
+            "uzp2 z7.d, z22.d, z7.d\n"
+            "fmul z22.s, z23.s, z3.s[2]\n"
+            "fmul z3.s, z23.s, z3.s[3]\n"
+            "scvtf z7.s, p1/m, z7.s\n"
+            "fmla z19.s, p1/M, z18.s, z22.s\n"
+            "ld1rqb { z18.b }, p1/Z, [x24, #16]\n"
+            "fmul z22.s, z23.s, z5.s[0]\n"
+            "fmla z26.s, p1/M, z7.s, z3.s\n"
+            "mov z3.s, #0x0\n"
+            "mov z7.s, #0x0\n"
+            ".inst 0x451f9923  // smmla z3.s, z9.b, z31.b\n"
+            ".inst 0x45069927  // smmla z7.s, z9.b, z6.b\n"
+            "ld1rqb { z9.b }, p1/Z, [x24, #32]\n"
+            ".inst 0x450e9923  // smmla z3.s, z9.b, z14.b\n"
+            ".inst 0x45029927  // smmla z7.s, z9.b, z2.b\n"
+            "mov z9.s, #0x0\n"
+            ".inst 0x451f9a49  // smmla z9.s, z18.b, z31.b\n"
+            "mov z31.s, #0x0\n"
+            ".inst 0x45069a5f  // smmla z31.s, z18.b, z6.b\n"
+            "ld1rqb { z6.b }, p1/Z, [x24, #48]\n"
+            "ld1rqb { z18.b }, p1/Z, [x24, #64]\n"
+            ".inst 0x450e98c9  // smmla z9.s, z6.b, z14.b\n"
+            "fmul z14.s, z23.s, z5.s[1]\n"
+            ".inst 0x450298df  // smmla z31.s, z6.b, z2.b\n"
+            "ld1rqb { z6.b }, p1/Z, [x24, #80]\n"
+            "fmul z2.s, z23.s, z5.s[2]\n"
+            "fmul z23.s, z23.s, z5.s[3]\n"
+            ".inst 0x451e9a43  // smmla z3.s, z18.b, z30.b\n"
+            ".inst 0x45159a47  // smmla z7.s, z18.b, z21.b\n"
+            "ld1rqb { z5.b }, p1/Z, [x24, #96]\n"
+            ".inst 0x451e98c9  // smmla z9.s, z6.b, z30.b\n"
+            ".inst 0x451598df  // smmla z31.s, z6.b, z21.b\n"
+            "ld1rqb { z18.b }, p1/Z, [x24, #112]\n"
+            "add x24, x24, #0x88\n"
+            ".inst 0x450498a3  // smmla z3.s, z5.b, z4.b\n"
+            ".inst 0x451198a7  // smmla z7.s, z5.b, z17.b\n"
+            ".inst 0x45049a49  // smmla z9.s, z18.b, z4.b\n"
+            ".inst 0x45119a5f  // smmla z31.s, z18.b, z17.b\n"
+            "uzp1 z18.d, z3.d, z7.d\n"
+            "uzp2 z5.d, z3.d, z7.d\n"
+            "scvtf z18.s, p1/m, z18.s\n"
+            "uzp1 z6.d, z9.d, z31.d\n"
+            "uzp2 z9.d, z9.d, z31.d\n"
+            "scvtf z5.s, p1/m, z5.s\n"
+            "fmla z8.s, p1/M, z18.s, z22.s\n"
+            "scvtf z6.s, p1/m, z6.s\n"
+            "scvtf z9.s, p1/m, z9.s\n"
+            "fmla z29.s, p1/M, z5.s, z14.s\n"
+            "fmla z27.s, p1/M, z6.s, z2.s\n"
+            "fmla z10.s, p1/M, z9.s, z23.s\n"
+            "bgt 3b\n"
+            "mov x20, %x[res_ptr]\n"
+            "subs x10, x10, #0x8\n"
+            "add %x[res_ptr], %x[res_ptr], #0x20\n"
+            "st1w { z24.s }, p1, [x20]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "st1w { z15.s }, p1, [x20]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "st1w { z12.s }, p1, [x20]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "st1w { z0.s }, p1, [x20]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "st1w { z13.s }, p1, [x20]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "st1w { z1.s }, p1, [x20]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "st1w { z20.s }, p1, [x20]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "st1w { z25.s }, p1, [x20]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "st1w { z11.s }, p1, [x20]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "st1w { z16.s }, p1, [x20]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "st1w { z19.s }, p1, [x20]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "st1w { z26.s }, p1, [x20]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "st1w { z8.s }, p1, [x20]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "st1w { z29.s }, p1, [x20]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "st1w { z27.s }, p1, [x20]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "st1w { z10.s }, p1, [x20]\n"
+            "bne 2b\n"
+            "mov x20, #0x4\n"
+            "sub x13, x13, #0x10\n"
+            "cmp x13, #0x10\n"
+            "mov %x[res_ptr], x9\n"
+            "madd %x[a_ptr], x20, x12, %x[a_ptr]\n"
+            "bge 1b\n"
+            "4:"  // Row loop skip
+            "cbz x13, 9f\n"
+            "5:"  // Row tail: Row loop
+            "add x25, %x[b_ptr], #0x10\n"
+            "mov x24, %x[nc]\n"
+            "add x23, %x[res_ptr], %x[res_stride], LSL #2\n"
+            "6:"  // Row tail: Column loop
+            "mov z24.b, #0x0\n"
+            "mov z15.b, #0x0\n"
+            "add x28, %x[a_ptr], #0x8\n"
+            "mov x22, %x[nb]\n"
+            "mov z12.b, #0x0\n"
+            "mov z0.b, #0x0\n"
+            "7:"  // Row tail: Block loop
+            "ld1b { z3.b }, p1/Z, [x25]\n"
+            "ld1b { z6.b }, p1/Z, [x25, #1, MUL VL]\n"
+            "mov z2.s, #0x0\n"
+            "mov z25.s, #0x0\n"
+            "ld1rqb { z26.b }, p1/Z, [x28]\n"
+            "ld1rqb { z21.b }, p1/Z, [x28, #16]\n"
+            "mov z27.s, #0x0\n"
+            "mov z19.s, #0x0\n"
+            "ld1b { z29.b }, p1/Z, [x25, #2, MUL VL]\n"
+            "ld1b { z16.b }, p1/Z, [x25, #3, MUL VL]\n"
+            "sub x21, x25, #0x10\n"
+            "sub x20, x28, #0x8\n"
+            "lsl z20.b, z3.b, #0x4\n"
+            "lsl z4.b, z6.b, #0x4\n"
+            "ld1rqb { z10.b }, p1/Z, [x28, #32]\n"
+            "ld1rqb { z23.b }, p1/Z, [x28, #48]\n"
+            "and z3.b, z3.b, #0xf0\n"
+            "and z6.b, z6.b, #0xf0\n"
+            "ld1rqb { z11.b }, p1/Z, [x28, #64]\n"
+            "ld1rqb { z7.b }, p1/Z, [x28, #80]\n"
+            "lsl z8.b, z29.b, #0x4\n"
+            "lsl z14.b, z16.b, #0x4\n"
+            "ld1rqb { z18.b }, p1/Z, [x28, #96]\n"
+            "ld1rqb { z30.b }, p1/Z, [x28, #112]\n"
+            ".inst 0x45149b42  // smmla z2.s, z26.b, z20.b\n"
+            ".inst 0x45049b59  // smmla z25.s, z26.b, z4.b\n"
+            "and z29.b, z29.b, #0xf0\n"
+            "ld1h { z17.s }, p1/Z, [x21]\n"
+            ".inst 0x45149abb  // smmla z27.s, z21.b, z20.b\n"
+            ".inst 0x45049ab3  // smmla z19.s, z21.b, z4.b\n"
+            "and z16.b, z16.b, #0xf0\n"
+            "ld1h { z4.s }, p0/Z, [x20]\n"
+            "subs x22, x22, #0x1\n"
+            "add x28, x28, #0x88\n"
+            "fcvt z17.s, p1/m, z17.h\n"
+            "add x25, x25, #0x90\n"
+            ".inst 0x45089942  // smmla z2.s, z10.b, z8.b\n"
+            ".inst 0x450e9959  // smmla z25.s, z10.b, z14.b\n"
+            "fcvt z4.s, p1/m, z4.h\n"
+            ".inst 0x45089afb  // smmla z27.s, z23.b, z8.b\n"
+            ".inst 0x450e9af3  // smmla z19.s, z23.b, z14.b\n"
+            "fscale z17.s, p1/m, z17.s, z28.s\n"
+            "mov z4.q, z4.q[0]\n"
+            ".inst 0x45039962  // smmla z2.s, z11.b, z3.b\n"
+            ".inst 0x45069979  // smmla z25.s, z11.b, z6.b\n"
+            "fmul z23.s, z17.s, z4.s[0]\n"
+            "fmul z9.s, z17.s, z4.s[1]\n"
+            "fmul z21.s, z17.s, z4.s[2]\n"
+            "fmul z4.s, z17.s, z4.s[3]\n"
+            ".inst 0x450398fb  // smmla z27.s, z7.b, z3.b\n"
+            ".inst 0x450698f3  // smmla z19.s, z7.b, z6.b\n"
+            ".inst 0x451d9a42  // smmla z2.s, z18.b, z29.b\n"
+            ".inst 0x45109a59  // smmla z25.s, z18.b, z16.b\n"
+            ".inst 0x451d9bdb  // smmla z27.s, z30.b, z29.b\n"
+            ".inst 0x45109bd3  // smmla z19.s, z30.b, z16.b\n"
+            "uzp1 z31.d, z2.d, z25.d\n"
+            "uzp2 z13.d, z2.d, z25.d\n"
+            "scvtf z31.s, p1/m, z31.s\n"
+            "uzp1 z17.d, z27.d, z19.d\n"
+            "uzp2 z18.d, z27.d, z19.d\n"
+            "scvtf z13.s, p1/m, z13.s\n"
+            "fmla z24.s, p1/M, z31.s, z23.s\n"
+            "scvtf z17.s, p1/m, z17.s\n"
+            "scvtf z18.s, p1/m, z18.s\n"
+            "fmla z15.s, p1/M, z13.s, z9.s\n"
+            "fmla z12.s, p1/M, z17.s, z21.s\n"
+            "fmla z0.s, p1/M, z18.s, z4.s\n"
+            "bgt 7b\n"
+            "mov x20, %x[res_ptr]\n"
+            "cmp x13, #0x1\n"
+            "st1w { z24.s }, p1, [x20]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "ble 8f\n"
+            "cmp x13, #0x2\n"
+            "st1w { z15.s }, p1, [x20]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "ble 8f\n"
+            "cmp x13, #0x3\n"
+            "st1w { z12.s }, p1, [x20]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "ble 8f\n"
+            "st1w { z0.s }, p1, [x20]\n"
+            "8:"  // Row tail: Accumulator store skip
+            "subs x24, x24, #0x8\n"
+            "add %x[res_ptr], %x[res_ptr], #0x20\n"
+            "bne 6b\n"
+            "subs x13, x13, #0x4\n"
+            "add %x[a_ptr], %x[a_ptr], x12\n"
+            "mov %x[res_ptr], x23\n"
+            "bgt 5b\n"
+            "9:"  // Row tail: Row loop skip
+            : [a_ptr] "+&r" (a_ptr), [res_ptr] "+&r" (res_ptr)
+            : [b_ptr] "r" (b_ptr), [nr] "r" (nr), [nb] "r" (nb), [res_stride] "r" (res_stride), [nc] "r" (nc)
+            : "cc", "memory", "p0", "p1", "x9", "x10", "x11", "x12", "x13", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
+        );
+        return;
+    }
+#endif // #if defined(__ARM_FEATURE_SVE) && defined(__ARM_FEATURE_MATMUL_INT8)
+
+#endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__)
+    float sumf[4][8];
+    int sumi;
+
+    for (int y = 0; y < nr / 4; y++) {
+        const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
+        for (int x = 0; x < nc / ncols_interleaved; x++) {
+            const block_q4_0x8 * b_ptr = (const block_q4_0x8 *) vx + (x * nb);
+            for (int m = 0; m < 4; m++) {
+                for (int j = 0; j < ncols_interleaved; j++) sumf[m][j] = 0.0;
+            }
+            for (int l = 0; l < nb; l++) {
+                for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                    for (int m = 0; m < 4; m++) {
+                        for (int j = 0; j < ncols_interleaved; j++) {
+                            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] << 4);
+                                const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
+                                sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
+                                         (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4;
+                            }
+                            sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(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];
+            }
+        }
+    }
+}
+
+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;
+    const int ncols_interleaved = 4;
+    const int blocklen = 4;
+
+    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(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD)
+    const int8x16_t kvalues = vld1q_s8(kvalues_iq4nl);
+
+    for (int y = 0; y < nr / 4; y++) {
+        const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
+        for (int x = 0; x < nc / ncols_interleaved; x++) {
+            const block_iq4_nlx4 * b_ptr = (const block_iq4_nlx4 *) vx + (x * nb);
+
+            float32x4_t sumf[4];
+            for (int m = 0; m < 4; m++) {
+                sumf[m] = vdupq_n_f32(0);
+            }
+
+            for (int l = 0; l < nb; l++) {
+                float32x4_t a_d = vcvt_f32_f16(vld1_f16((const float16_t *)a_ptr[l].d));
+                float32x4_t b_d = vcvt_f32_f16(vld1_f16((const float16_t *)b_ptr[l].d));
+
+                int32x4_t sumi_0 = vdupq_n_s32(0);
+                int32x4_t sumi_1 = vdupq_n_s32(0);
+                int32x4_t sumi_2 = vdupq_n_s32(0);
+                int32x4_t sumi_3 = vdupq_n_s32(0);
+
+                for (int k = 0; k < 4; k++) {
+                    int8x16_t a_0 = vld1q_s8(a_ptr[l].qs + 16 * k + 0);
+                    int8x16_t a_1 = vld1q_s8(a_ptr[l].qs + 16 * k + 64);
+
+                    uint8x16_t b = vld1q_u8(b_ptr[l].qs + 16 * k);
+                    int8x16_t b_hi = vqtbl1q_s8(kvalues, b >> 4);
+                    int8x16_t b_lo = vqtbl1q_s8(kvalues, b & 0xF);
+
+                    sumi_0 = vdotq_laneq_s32(sumi_0, b_lo, a_0, 0);
+                    sumi_1 = vdotq_laneq_s32(sumi_1, b_lo, a_0, 1);
+                    sumi_2 = vdotq_laneq_s32(sumi_2, b_lo, a_0, 2);
+                    sumi_3 = vdotq_laneq_s32(sumi_3, b_lo, a_0, 3);
+                    sumi_0 = vdotq_laneq_s32(sumi_0, b_hi, a_1, 0);
+                    sumi_1 = vdotq_laneq_s32(sumi_1, b_hi, a_1, 1);
+                    sumi_2 = vdotq_laneq_s32(sumi_2, b_hi, a_1, 2);
+                    sumi_3 = vdotq_laneq_s32(sumi_3, b_hi, a_1, 3);
+                }
+
+                sumf[0] = vmlaq_f32(sumf[0], vmulq_laneq_f32(b_d, a_d, 0), vcvtq_f32_s32(sumi_0));
+                sumf[1] = vmlaq_f32(sumf[1], vmulq_laneq_f32(b_d, a_d, 1), vcvtq_f32_s32(sumi_1));
+                sumf[2] = vmlaq_f32(sumf[2], vmulq_laneq_f32(b_d, a_d, 2), vcvtq_f32_s32(sumi_2));
+                sumf[3] = vmlaq_f32(sumf[3], vmulq_laneq_f32(b_d, a_d, 3), vcvtq_f32_s32(sumi_3));
+            }
+
+            for (int m = 0; m < 4; m++) {
+                vst1q_f32(s + (y * 4 + m) * bs + x * 4, sumf[m]);
+            }
+        }
+    }
+    return;
+#endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON)
+    {
+        float sumf[4][4];
+        int sumi;
+
+        for (int y = 0; y < nr / 4; y++) {
+            const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
+            for (int x = 0; x < nc / ncols_interleaved; x++) {
+                const block_iq4_nlx4 * b_ptr = (const block_iq4_nlx4 *) vx + (x * nb);
+                for (int m = 0; m < 4; m++) {
+                    for (int j = 0; j < ncols_interleaved; j++) sumf[m][j] = 0.0;
+                }
+                for (int l = 0; l < nb; l++) {
+                    for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                        for (int m = 0; m < 4; m++) {
+                            for (int j = 0; j < ncols_interleaved; j++) {
+                                sumi = 0;
+                                for (int i = 0; i < blocklen; ++i) {
+                                    const int v0 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0x0F];
+                                    const int v1 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4];
+                                    sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
+                                            (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4]));
+                                }
+                                sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(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];
+                }
+            }
+        }
+    }
+}
diff --git a/ggml/src/ggml-cpu/arch/loongarch/quants.c b/ggml/src/ggml-cpu/arch/loongarch/quants.c
new file mode 100644
index 000000000..9e33fb322
--- /dev/null
+++ b/ggml/src/ggml-cpu/arch/loongarch/quants.c
@@ -0,0 +1,2639 @@
+#define GGML_COMMON_IMPL_C
+#include "ggml-common.h"
+#include "ggml-quants.h"
+#include "ggml-impl.h"
+#include "ggml-cpu.h"
+#include "simd-mappings.h"
+
+#include "../../quants.h"
+#include "../../ggml-cpu-impl.h"
+
+#include <math.h>
+#include <string.h>
+#include <assert.h>
+#include <float.h>
+#include <stdlib.h> // for qsort
+#include <stdio.h>  // for GGML_ASSERT
+
+#define GROUP_MAX_EPS 1e-15f
+#define GROUP_MAX_EPS_IQ3_XXS 1e-8f
+#define GROUP_MAX_EPS_IQ2_S 1e-8f
+#define GROUP_MAX_EPS_IQ1_M 1e-7f
+#define GROUP_MAX_EPS_IQ1_S 1e-12f
+
+#define UNUSED GGML_UNUSED
+
+#if defined(__loongarch_sx)
+
+static __m128i lsx_packs_w(__m128i a, __m128i b) {
+    __m128i tmp, tmp1;
+    tmp = __lsx_vsat_w(a, 15);
+    tmp1 = __lsx_vsat_w(b, 15);
+    return __lsx_vpickev_h(tmp1, tmp);
+}
+
+static __m128i lsx_packs_h(__m128i a, __m128i b) {
+    __m128i tmp, tmp1;
+    tmp = __lsx_vsat_h(a, 7);
+    tmp1 = __lsx_vsat_h(b, 7);
+    return __lsx_vpickev_b(tmp1, tmp);
+}
+
+static __m128i lsx_packus_h(__m128i a, __m128i b) {
+    __m128i tmp, tmp1;
+    tmp = __lsx_vsat_hu(a, 7);
+    tmp1 = __lsx_vsat_hu(b, 7);
+    return __lsx_vpickev_b(tmp1, tmp);
+}
+
+static __m128i lsx_maddubs_h(__m128i a, __m128i b) {
+    __m128i tmp1, tmp2;
+    tmp1 = __lsx_vmulwev_h_b(a, b);
+    tmp2 = __lsx_vmulwod_h_b(a, b);
+    return __lsx_vsadd_h(tmp1, tmp2);
+}
+
+static __m128i lsx_madd_h(__m128i a, __m128i b) {
+    __m128i tmp1, tmp2;
+    tmp1 = __lsx_vmulwev_w_h(a, b);
+    tmp2 = __lsx_vmulwod_w_h(a, b);
+    return __lsx_vadd_w(tmp1, tmp2);
+}
+
+static __m128i lsx_set_w(int32_t a, int32_t b, int32_t c, int32_t d) {
+    v4i32 __ret = {d, c, b, a};
+    return (__m128i)__ret;
+}
+
+static __m128i lsx_shuffle_b(__m128i a, __m128i b) {
+    __m128i mask_f, zero, tmp0, tmp2, mask;
+    int f = 0x8f;
+    mask_f = __lsx_vreplgr2vr_b(f);
+    zero = __lsx_vldi(0);
+    tmp0 = __lsx_vand_v(b, mask_f); // get mask with low 4 bit and sign bits
+    tmp0 = __lsx_vori_b(tmp0, 0x10); // make each mask or  with 0x10 prepare for positive
+    mask = __lsx_vsle_b(zero, tmp0); // if mask >= 0, set mask
+    tmp2 = __lsx_vand_v(tmp0, mask); // maskout the in2 < ones
+    return __lsx_vshuf_b(a, zero, tmp2);
+}
+
+static __m128i lsx_hadd_h(__m128i a, __m128i b) {
+    __m128i tmp1 = __lsx_vpickev_h(b, a);
+    __m128i tmp2 = __lsx_vpickod_h(b, a);
+    return __lsx_vadd_h(tmp1, tmp2);
+}
+
+static __m128i lsx_hadd_w(__m128i a, __m128i b) {
+    __m128i tmp1 = __lsx_vpickev_w(b, a);
+    __m128i tmp2 = __lsx_vpickod_w(b, a);
+    return __lsx_vadd_w(tmp1, tmp2);
+}
+
+static __m128 lsx_hadd_s(__m128 a, __m128 b) {
+    __m128 tmp1 = (__m128)__lsx_vpickev_w((__m128i)b, (__m128i)a);
+    __m128 tmp2 = (__m128)__lsx_vpickod_w((__m128i)b, (__m128i)a);
+
+    return __lsx_vfadd_s(tmp1, tmp2);
+}
+
+static inline float hsum_float_4x4(const __m128 a, const __m128 b, const __m128 c, const __m128 d) {
+    __m128 res_0 =lsx_hadd_s(a, b);
+    __m128 res_1 =lsx_hadd_s(c, d);
+    __m128 res =lsx_hadd_s(res_0, res_1);
+    res =lsx_hadd_s(res, res);
+    res =lsx_hadd_s(res, res);
+
+    return ((v4f32)res)[0];
+}
+#endif
+
+#if defined(__loongarch_asx)
+
+#ifdef __clang__
+#define VREGS_PREFIX "$vr"
+#define XREGS_PREFIX "$xr"
+#else // GCC
+#define VREGS_PREFIX "$f"
+#define XREGS_PREFIX "$f"
+#endif
+#define __ALL_REGS "0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,21,22,23,24,25,26,27,28,29,30,31"
+// Convert __m128i to __m256i
+static inline __m256i ____m256i(__m128i in) {
+    __m256i out = __lasx_xvldi(0);
+    __asm__ volatile (
+        ".irp i," __ALL_REGS                "\n\t"
+        " .ifc %[out], " XREGS_PREFIX"\\i    \n\t"
+        "  .irp j," __ALL_REGS              "\n\t"
+        "   .ifc %[in], " VREGS_PREFIX "\\j  \n\t"
+        "    xvpermi.q $xr\\i, $xr\\j, 0x20  \n\t"
+        "   .endif                           \n\t"
+        "  .endr                             \n\t"
+        " .endif                             \n\t"
+        ".endr                               \n\t"
+        : [out] "+f" (out) : [in] "f" (in)
+    );
+    return out;
+}
+// Convert two __m128i to __m256i
+static inline __m256i lasx_set_q(__m128i inhi, __m128i inlo) {
+    __m256i out;
+    __asm__ volatile (
+        ".irp i," __ALL_REGS                "\n\t"
+        " .ifc %[hi], " VREGS_PREFIX "\\i    \n\t"
+        "  .irp j," __ALL_REGS              "\n\t"
+        "   .ifc %[lo], " VREGS_PREFIX "\\j  \n\t"
+        "    xvpermi.q $xr\\i, $xr\\j, 0x20  \n\t"
+        "   .endif                           \n\t"
+        "  .endr                             \n\t"
+        " .endif                             \n\t"
+        ".endr                               \n\t"
+        ".ifnc %[out], %[hi]                 \n\t"
+        ".irp i," __ALL_REGS                "\n\t"
+        " .ifc %[out], " XREGS_PREFIX "\\i   \n\t"
+        "  .irp j," __ALL_REGS              "\n\t"
+        "   .ifc %[hi], " VREGS_PREFIX "\\j  \n\t"
+        "    xvori.b $xr\\i, $xr\\j, 0       \n\t"
+        "   .endif                           \n\t"
+        "  .endr                             \n\t"
+        " .endif                             \n\t"
+        ".endr                               \n\t"
+        ".endif                              \n\t"
+        : [out] "=f" (out), [hi] "+f" (inhi)
+        : [lo] "f" (inlo)
+    );
+    return out;
+}
+// Convert __m256i low part to __m128i
+static inline __m128i lasx_extracti128_lo(__m256i in) {
+    __m128i out;
+    __asm__ volatile (
+        ".ifnc %[out], %[in]                 \n\t"
+        ".irp i," __ALL_REGS                "\n\t"
+        " .ifc %[out], " VREGS_PREFIX "\\i   \n\t"
+        "  .irp j," __ALL_REGS              "\n\t"
+        "   .ifc %[in], " XREGS_PREFIX "\\j  \n\t"
+        "    vori.b $vr\\i, $vr\\j, 0        \n\t"
+        "   .endif                           \n\t"
+        "  .endr                             \n\t"
+        " .endif                             \n\t"
+        ".endr                               \n\t"
+        ".endif                              \n\t"
+        : [out] "=f" (out) : [in] "f" (in)
+    );
+    return out;
+}
+// Convert __m256i high part to __m128i
+static inline __m128i lasx_extracti128_hi(__m256i in) {
+    __m128i out;
+    __asm__ volatile (
+        ".irp i," __ALL_REGS                "\n\t"
+        " .ifc %[out], " VREGS_PREFIX "\\i   \n\t"
+        "  .irp j," __ALL_REGS              "\n\t"
+        "   .ifc %[in], " XREGS_PREFIX "\\j  \n\t"
+        "    xvpermi.q $xr\\i, $xr\\j, 0x11  \n\t"
+        "   .endif                           \n\t"
+        "  .endr                             \n\t"
+        " .endif                             \n\t"
+        ".endr                               \n\t"
+        : [out] "=f" (out) : [in] "f" (in)
+    );
+    return out;
+}
+
+static __m256i lasx_set_w(int e7, int e6, int e5, int e4, int e3, int e2, int e1, int e0) {
+    v8i32 __ret = {e0, e1, e2, e3, e4, e5, e6, e7};
+    return (__m256i)__ret;
+}
+
+static __m256i lasx_set_d(int64_t a, int64_t b, int64_t c, int64_t d) {
+    v4i64 __ret = {d, c, b, a};
+    return (__m256i)__ret;
+}
+
+static __m256i lasx_insertf128( __m128i x, __m128i y) {
+    return lasx_set_q(x, y);
+}
+
+static __m256i lasx_shuffle_b(__m256i a, __m256i b) {
+    __m256i mask_f, zero, tmp0, tmp2, mask;
+    int f = 0x8f;
+    mask_f = __lasx_xvreplgr2vr_b(f);
+    zero = __lasx_xvldi(0);
+    tmp0 = __lasx_xvand_v(b, mask_f); // get mask with low 4 bit and sign bits
+    tmp0 = __lasx_xvori_b(tmp0, 0x10); // make each mask or  with 0x10 prepare for positive
+    mask = __lasx_xvsle_b(zero, tmp0); // if mask >= 0, set mask
+    tmp2 = __lasx_xvand_v(tmp0, mask); // maskout the in2 < ones
+    return __lasx_xvshuf_b(a, zero, tmp2);
+}
+
+static __m256i lasx_extu8_16(__m128i a) {
+    return __lasx_vext2xv_hu_bu(____m256i(a));
+}
+
+static __m256i lasx_ext8_16(__m128i a) {
+    return __lasx_vext2xv_h_b(____m256i(a));
+}
+
+static __m256i lasx_ext16_32(__m128i a) {
+    return __lasx_vext2xv_w_h(____m256i(a));
+}
+
+static __m128i lasx_extracti128( __m256i a, int pos) {
+    __m128i ret;
+    if( pos == 0)
+    {
+       ret = lasx_extracti128_lo(a);
+    } else {
+       ret = lasx_extracti128_hi(a);
+    }
+    return ret;
+}
+
+static __m128 lasx_extractf128( __m256 a, int pos) {
+    __m128 ret;
+    if( pos == 0)
+    {
+       ret = (__m128)lasx_extracti128_lo((__m256i)a);
+    } else {
+       ret = (__m128)lasx_extracti128_hi((__m256i)a);
+    }
+    return ret;
+}
+
+static __m256i lasx_maddubs_h(__m256i a, __m256i b) {
+    __m256i tmp1, tmp2;
+    tmp1 = __lasx_xvmulwev_h_b(a, b);
+    tmp2 = __lasx_xvmulwod_h_b(a, b);
+    return __lasx_xvsadd_h(tmp1, tmp2);
+}
+
+static __m256i lasx_madd_h(__m256i a, __m256i b) {
+    __m256i tmp1, tmp2;
+    tmp1 = __lasx_xvmulwev_w_h(a, b);
+    tmp2 = __lasx_xvmulwod_w_h(a, b);
+    return __lasx_xvadd_w(tmp1, tmp2);
+}
+
+static __m256i lasx_packs_w(__m256i a, __m256i b) {
+    __m256i tmp, tmp1;
+    tmp = __lasx_xvsat_w(a, 15);
+    tmp1 = __lasx_xvsat_w(b, 15);
+    return __lasx_xvpickev_h(tmp1, tmp);
+}
+
+static __m256i lasx_packs_h(__m256i a, __m256i b) {
+    __m256i tmp, tmp1;
+    tmp = __lasx_xvsat_h(a, 7);
+    tmp1 = __lasx_xvsat_h(b, 7);
+    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
+    const __m128i ax = __lsx_vsigncov_b(x, x);
+    // Sign the values of the y vectors
+    const __m128i sy = __lsx_vsigncov_b(x, y);
+    // Perform multiplication and create 16-bit values
+    const __m128i dot = lsx_maddubs_h(ax, sy);
+    const __m128i ones = __lsx_vreplgr2vr_h(1);
+    return lsx_madd_h(ones, dot);
+}
+
+// horizontally add 8 floats
+static inline float hsum_float_8(const __m256 x) {
+    __m128 res = lasx_extractf128(x, 1);
+    res = __lsx_vfadd_s(res, lasx_extractf128(x, 0));
+    res = __lsx_vfadd_s(res, (__m128)__lsx_vpickod_d((__m128i)res, (__m128i)res));
+    res = __lsx_vfadd_s(res, (__m128)__lsx_vinsgr2vr_w(__lsx_vldi(0), __lsx_vpickve2gr_w(res, 1), 0));
+    return ((v4f32)res)[0];
+}
+
+// horizontally add 8 int32_t
+static inline int hsum_i32_8(const __m256i a) {
+
+    __m256i tmp1 = __lasx_xvpermi_q(a, a, 0x11);
+    __m256i tmp2 = __lasx_xvpermi_q(a, a, 0x00);
+
+    __m128i  tmp1_128 = lasx_extracti128_lo(tmp1);
+    __m128i  tmp2_128 = lasx_extracti128_lo(tmp2);
+
+    __m128i sum128 = __lsx_vadd_w(tmp1_128, tmp2_128);
+
+    __m128i ev = __lsx_vpickev_w(sum128, sum128);
+    __m128i od = __lsx_vpickod_w(sum128, sum128);
+    __m128i sum64 = __lsx_vadd_w(ev, od);
+
+    int sum64_1, sum64_2;
+    sum64_1 = __lsx_vpickve2gr_w(sum64, 0);
+    sum64_2 = __lsx_vpickve2gr_w(sum64, 1);
+
+    return  sum64_1 + sum64_2;
+}
+
+// horizontally add 4 int32_t
+static inline int hsum_i32_4(const __m128i a) {
+    __m128i ev = __lsx_vpickev_w(a, a);
+    __m128i od = __lsx_vpickod_w(a, a);
+    __m128i sum64 = __lsx_vadd_w(ev, od);
+
+    int sum64_1, sum64_2;
+    sum64_1 = __lsx_vpickve2gr_w(sum64, 0);
+    sum64_2 = __lsx_vpickve2gr_w(sum64, 1);
+
+    return  sum64_1 + sum64_2;
+}
+
+// spread 32 bits to 32 bytes { 0x00, 0xFF }
+static inline __m256i bytes_from_bits_32(const uint8_t * x) {
+
+    uint32_t x32;
+    memcpy(&x32, x, sizeof(uint32_t));
+    const __m256i shuf_mask = lasx_set_d(
+            0x0303030303030303, 0x0202020202020202,
+            0x0101010101010101, 0x0000000000000000);
+
+    __m256i bytes = lasx_shuffle_b(__lasx_xvreplgr2vr_w(x32), shuf_mask);
+    const __m256i bit_mask = __lasx_xvreplgr2vr_d(0x7fbfdfeff7fbfdfe);
+    bytes = __lasx_xvor_v(bytes, bit_mask);
+    return __lasx_xvseq_b(bytes, __lasx_xvreplgr2vr_d(-1));
+}
+
+// Unpack 32 4-bit fields into 32 bytes
+// The output vector contains 32 bytes, each one in [ 0 .. 15 ] interval
+static inline __m256i bytes_from_nibbles_32(const uint8_t * rsi) {
+    const __m128i lo = __lsx_vld((const __m128i *)rsi, 0);
+    __m128i hi = __lsx_vsrli_h(lo, 4);
+    return __lasx_xvandi_b(lasx_insertf128(hi, lo), 0xf);
+}
+
+// add int16_t pairwise and return as float vector
+static inline __m256 sum_i16_pairs_float(const __m256i x) {
+    __m256i v = __lasx_xvpackod_h(x, x);
+    __m256i summed_pairs = __lasx_xvaddwev_w_h(x, v);
+    return __lasx_xvffint_s_w(summed_pairs);
+}
+
+static inline __m256 mul_sum_us8_pairs_float(const __m256i ax, const __m256i sy) {
+    // Perform multiplication and create 16-bit values
+    const __m256i dot = lasx_maddubs_h(ax, sy);
+    return sum_i16_pairs_float(dot);
+}
+
+// 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) {
+    const __m256i dot = lasx_madd_h_b(x, y);
+    return sum_i16_pairs_float(dot);
+}
+
+static inline __m128i packNibbles( __m256i bytes ) {
+    // Move bits within 16-bit lanes from 0000_abcd_0000_efgh into 0000_0000_abcd_efgh
+    const __m256i lowByte = __lasx_xvreplgr2vr_h(0xFF);
+     __m256i high = __lasx_xvandn_v(lowByte, bytes);
+    __m256i low = __lasx_xvand_v(lowByte, bytes);
+    high = __lasx_xvsrli_h(high, 4);
+    bytes = __lasx_xvor_v(low, high);
+    // Compress uint16_t lanes into bytes
+    __m128i *r0 = (__m128i *)&bytes;
+    __m256i tmp_h128 = __lasx_xvpermi_q(bytes, bytes, 0x11);
+    __m128i *r1 = (__m128i *)&tmp_h128;
+
+    __m128i zero = __lsx_vldi(0);
+    __m128i tmp, tmp2, tmp3;
+
+    tmp = __lsx_vmax_h(zero, *r0);
+    tmp2 = __lsx_vsat_hu(tmp, 7);
+
+    tmp = __lsx_vmax_h(zero, *r1);
+    tmp3 = __lsx_vsat_hu(tmp, 7);
+    return  __lsx_vpickev_b(tmp3, tmp2);
+}
+#endif  //__loongarch_asx
+
+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 * GGML_RESTRICT y = vy;
+
+#if defined(__loongarch_asx)
+    for (int i = 0; i < nb; i++) {
+        __m256 v0 = (__m256)__lasx_xvld( x , 0);
+        __m256 v1 = (__m256)__lasx_xvld( x , 32);
+        __m256 v2 = (__m256)__lasx_xvld( x , 64);
+        __m256 v3 = (__m256)__lasx_xvld( x , 96);
+        x += 32;
+
+        // Compute max(abs(e)) for the block
+        const __m256 sign_bit = __lasx_xvreplfr2vr_s( -0.0f );
+        __m256 max_abs = (__m256)__lasx_xvandn_v( (__m256i)sign_bit, (__m256i)v0 );
+        max_abs = __lasx_xvfmax_s( max_abs, (__m256)__lasx_xvandn_v( (__m256i)sign_bit, (__m256i)v1 ) );
+        max_abs = __lasx_xvfmax_s( max_abs, (__m256)__lasx_xvandn_v( (__m256i)sign_bit, (__m256i)v2 ) );
+        max_abs = __lasx_xvfmax_s( max_abs, (__m256)__lasx_xvandn_v( (__m256i)sign_bit, (__m256i)v3 ) );
+
+        __m128 max4 = __lsx_vfmax_s( lasx_extractf128( max_abs, 1 ), lasx_extractf128( max_abs , 0) );
+        max4 = __lsx_vfmax_s( max4, (__m128)__lsx_vpickod_d((__m128i) max4, (__m128i)max4 ) );
+        __m128 tmp = max4;
+        max4 = __lsx_vfmax_s( max4, (__m128)__lsx_vinsgr2vr_w(tmp, __lsx_vpickve2gr_w( max4, 1 ), 0 ));
+        const float max_scalar = ((v4f32)max4)[0];
+
+        // Quantize these floats
+        const float d = max_scalar / 127.f;
+        y[i].d = GGML_CPU_FP32_TO_FP16(d);
+        const float id = ( max_scalar != 0.0f ) ? 127.f / max_scalar : 0.0f;
+        const __m256 mul = (__m256)__lasx_xvreplfr2vr_s( id );
+
+        // Apply the multiplier
+        v0 = __lasx_xvfmul_s( v0, mul );
+        v1 = __lasx_xvfmul_s( v1, mul );
+        v2 = __lasx_xvfmul_s( v2, mul );
+        v3 = __lasx_xvfmul_s( v3, mul );
+
+        // Round to nearest integer
+        __m256i i0 = __lasx_xvftintrne_w_s( v0 );
+        __m256i i1 = __lasx_xvftintrne_w_s( v1 );
+        __m256i i2 = __lasx_xvftintrne_w_s( v2 );
+        __m256i i3 = __lasx_xvftintrne_w_s( v3 );
+
+        __m128i ni0 = lasx_extracti128( i0, 0 );
+        __m128i ni1 = lasx_extracti128( i0, 1);
+        __m128i ni2 = lasx_extracti128( i1, 0);
+        __m128i ni3 = lasx_extracti128( i1, 1);
+        __m128i ni4 = lasx_extracti128( i2, 0);
+        __m128i ni5 = lasx_extracti128( i2, 1);
+        __m128i ni6 = lasx_extracti128( i3, 0);
+        __m128i ni7 = lasx_extracti128( i3, 1);
+
+        // Convert int32 to int16
+        ni0 = lsx_packs_w( ni0, ni1 );
+        ni2 = lsx_packs_w( ni2, ni3 );
+        ni4 = lsx_packs_w( ni4, ni5 );
+        ni6 = lsx_packs_w( ni6, ni7 );
+        // Convert int16 to int8
+        ni0 = lsx_packs_h( ni0, ni2 );
+        ni4 = lsx_packs_h( ni4, ni6 );
+
+        __lsx_vst(ni0, (__m128i *)(y[i].qs +  0), 0);
+        __lsx_vst(ni4, (__m128i *)(y[i].qs + 16), 0);
+
+    }
+#else
+    GGML_UNUSED(nb);
+    // scalar
+    quantize_row_q8_0_ref(x, y, k);
+#endif
+}
+
+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 * GGML_RESTRICT y = vy;
+
+#if defined(__loongarch_asx)
+    for (int i = 0; i < nb; i++) {
+        __m256 v0 = (__m256)__lasx_xvld( x , 0 );
+        __m256 v1 = (__m256)__lasx_xvld( x , 32 );
+        __m256 v2 = (__m256)__lasx_xvld( x , 64 );
+        __m256 v3 = (__m256)__lasx_xvld( x , 96 );
+        x += 32;
+
+        // Compute max(abs(e)) for the block
+        const __m256 sign_bit = __lasx_xvreplfr2vr_s( -0.0f );
+        __m256 max_abs = (__m256)__lasx_xvandn_v( (__m256i)sign_bit, (__m256i)v0 );
+        max_abs = __lasx_xvfmax_s( max_abs, (__m256)__lasx_xvandn_v( (__m256i)sign_bit, (__m256i)v1 ) );
+        max_abs = __lasx_xvfmax_s( max_abs, (__m256)__lasx_xvandn_v( (__m256i)sign_bit, (__m256i)v2 ) );
+        max_abs = __lasx_xvfmax_s( max_abs, (__m256)__lasx_xvandn_v( (__m256i)sign_bit, (__m256i)v3 ) );
+
+        __m128 max4 = __lsx_vfmax_s( lasx_extractf128( max_abs, 1 ), lasx_extractf128( max_abs, 0) );
+        max4 = __lsx_vfmax_s( max4, (__m128)__lsx_vpickod_d((__m128i) max4, (__m128i)max4 ) );
+        __m128 tmp = max4;
+        max4 = __lsx_vfmax_s( max4, (__m128)__lsx_vextrins_w((__m128i)tmp, (__m128i)max4, 0x10 ));
+        const float max_scalar = ((v4f32)max4)[0];
+
+        // Quantize these floats
+        const float d = max_scalar / 127.f;
+        y[i].d = GGML_CPU_FP32_TO_FP16(d);
+        const float id = ( max_scalar != 0.0f ) ? 127.f / max_scalar : 0.0f;
+        const __m256 mul = __lasx_xvreplfr2vr_s( id );
+
+        // Apply the multiplier
+        v0 = __lasx_xvfmul_s( v0, mul );
+        v1 = __lasx_xvfmul_s( v1, mul );
+        v2 = __lasx_xvfmul_s( v2, mul );
+        v3 = __lasx_xvfmul_s( v3, mul );
+
+        // Round to nearest integer
+        __m256i i0 = __lasx_xvftintrne_w_s( v0 );
+        __m256i i1 = __lasx_xvftintrne_w_s( v1 );
+        __m256i i2 = __lasx_xvftintrne_w_s( v2 );
+        __m256i i3 = __lasx_xvftintrne_w_s( v3 );
+
+        __m128i ni0 = lasx_extracti128(i0, 0);
+        __m128i ni1 = lasx_extracti128( i0, 1);
+        __m128i ni2 = lasx_extracti128( i1, 0);
+        __m128i ni3 = lasx_extracti128( i1, 1);
+        __m128i ni4 = lasx_extracti128( i2, 0 );
+        __m128i ni5 = lasx_extracti128( i2, 1);
+        __m128i ni6 = lasx_extracti128( i3, 0);
+        __m128i ni7 = lasx_extracti128( i3, 1);
+
+        // Compute the sum of the quants and set y[i].s
+        const __m128i s0 = __lsx_vadd_w(__lsx_vadd_w(ni0, ni1), __lsx_vadd_w(ni2, ni3));
+        const __m128i s1 = __lsx_vadd_w(__lsx_vadd_w(ni4, ni5), __lsx_vadd_w(ni6, ni7));
+        y[i].s = GGML_CPU_FP32_TO_FP16(d * hsum_i32_4(__lsx_vadd_w(s0, s1)));
+
+        // Convert int32 to int16
+        ni0 = lsx_packs_w( ni0, ni1 );
+        ni2 = lsx_packs_w( ni2, ni3 );
+        ni4 = lsx_packs_w( ni4, ni5 );
+        ni6 = lsx_packs_w( ni6, ni7 );
+        // Convert int16 to int8
+        ni0 = lsx_packs_h( ni0, ni2 );
+        ni4 = lsx_packs_h( ni4, ni6 );
+
+        __lsx_vst(ni0, (__m128i *)(y[i].qs +  0), 0);
+        __lsx_vst(ni4, (__m128i *)(y[i].qs + 16), 0);
+    }
+#else
+    GGML_UNUSED(nb);
+    // scalar
+    quantize_row_q8_1_ref(x, y, k);
+#endif
+}
+
+
+//===================================== Dot products =================================
+
+//
+// Helper functions
+//
+
+#if defined(__loongarch_asx)
+// shuffles to pick the required scales in dot products
+static inline __m256i get_scale_shuffle_q3k(int i) {
+    static const uint8_t k_shuffle[128] = {
+         0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1,     2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3,
+         4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5,     6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7,
+         8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9,    10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,
+        12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,    14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,
+    };
+    return __lasx_xvld((const __m256i*)k_shuffle + i, 0);
+}
+static inline __m256i get_scale_shuffle_k4(int i) {
+    static const uint8_t k_shuffle[256] = {
+         0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1,
+         2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3,
+         4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5,
+         6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7,
+         8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9,
+        10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,
+        12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,
+        14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15
+    };
+    return __lasx_xvld((const __m256i*)k_shuffle + i, 0);
+}
+static inline __m128i get_scale_shuffle(int i) {
+    static const uint8_t k_shuffle[128] = {
+         0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1,
+         2, 2, 2, 2, 2, 2, 2, 2, 3, 3, 3, 3, 3, 3, 3, 3,
+         4, 4, 4, 4, 4, 4, 4, 4, 5, 5, 5, 5, 5, 5, 5, 5,
+         6, 6, 6, 6, 6, 6, 6, 6, 7, 7, 7, 7, 7, 7, 7, 7,
+         8, 8, 8, 8, 8, 8, 8, 8, 9, 9, 9, 9, 9, 9, 9, 9,
+        10,10,10,10,10,10,10,10, 11,11,11,11,11,11,11,11,
+        12,12,12,12,12,12,12,12, 13,13,13,13,13,13,13,13,
+        14,14,14,14,14,14,14,14, 15,15,15,15,15,15,15,15
+    };
+    return __lsx_vld((const __m128i*)k_shuffle + i, 0);
+}
+#endif
+
+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;
+
+    assert(n % qk == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q4_0 * GGML_RESTRICT x = vx;
+    const block_q8_0 * GGML_RESTRICT y = vy;
+
+    int ib = 0;
+    float sumf = 0;
+
+#if defined(__loongarch_asx)
+    // Initialize accumulator with zeros
+    __m256 acc = (__m256)__lasx_xvldi(0);
+
+    // Main loop
+    for (; ib < nb; ++ib) {
+        /* Compute combined scale for the block */
+        const __m256 d = __lasx_xvreplfr2vr_s( GGML_CPU_FP16_TO_FP32(x[ib].d) * GGML_CPU_FP16_TO_FP32(y[ib].d) );
+
+        __m256i qx = bytes_from_nibbles_32(x[ib].qs);
+
+        // Now we have a vector with bytes in [ 0 .. 15 ] interval. Offset them into [ -8 .. +7 ] interval.
+        const __m256i off = __lasx_xvreplgr2vr_b( 8 );
+        qx = __lasx_xvsub_b( qx, off );
+
+        __m256i qy = __lasx_xvld((const __m256i *)y[ib].qs, 0);
+
+        const __m256 q = mul_sum_i8_pairs_float(qx, qy);
+
+        /* Multiply q with scale and accumulate */
+        acc = __lasx_xvfmadd_s( d, q, acc );
+    }
+
+    sumf = hsum_float_8(acc);
+
+#elif defined(__loongarch_sx)
+    // set constants
+    const __m128i low_mask = __lsx_vreplgr2vr_b(0xF);
+    const __m128i off = __lsx_vreplgr2vr_b(8);
+
+    // Initialize accumulator with zeros
+    __m128 acc_0 = (__m128)__lsx_vldi(0);
+    __m128 acc_1 = (__m128)__lsx_vldi(0);
+    __m128 acc_2 = (__m128)__lsx_vldi(0);
+    __m128 acc_3 = (__m128)__lsx_vldi(0);
+
+    for (; ib + 1 < nb; ib += 2) {
+
+        // Compute combined scale for the block 0 and 1
+        const __m128 d_0_1 = (__m128)__lsx_vreplgr2vr_w( GGML_CPU_FP16_TO_FP32(x[ib].d) * GGML_CPU_FP16_TO_FP32(y[ib].d) );
+
+        const __m128i tmp_0_1 = __lsx_vld((const __m128i *)x[ib].qs, 0);
+
+        __m128i bx_0 = __lsx_vand_v(low_mask, tmp_0_1);
+        __m128i by_0 = __lsx_vld((const __m128i *)y[ib].qs, 0);
+        bx_0 = __lsx_vsub_b(bx_0, off);
+        const __m128i i32_0 = mul_sum_i8_pairs(bx_0, by_0);
+
+        __m128i bx_1 = __lsx_vand_v(low_mask, __lsx_vsrli_d(tmp_0_1, 4));
+        __m128i by_1 = __lsx_vld((const __m128i *)(y[ib].qs + 16), 0);
+        bx_1 = __lsx_vsub_b(bx_1, off);
+        const __m128i i32_1 = mul_sum_i8_pairs(bx_1, by_1);
+
+        //_mm_prefetch(&x[ib] + 2 * sizeof(block_q4_0), _MM_HINT_T0);
+        //_mm_prefetch(&y[ib] + 2 * sizeof(block_q8_0), _MM_HINT_T0);
+
+        // Compute combined scale for the block 2 and 3
+        const __m128 d_2_3 = (__m128)__lsx_vreplgr2vr_w( GGML_CPU_FP16_TO_FP32(x[ib + 1].d) * GGML_CPU_FP16_TO_FP32(y[ib + 1].d) );
+
+        const __m128i tmp_2_3 = __lsx_vld((const __m128i *)x[ib + 1].qs, 0);
+
+        __m128i bx_2 = __lsx_vand_v(low_mask, tmp_2_3);
+        __m128i by_2 = __lsx_vld((const __m128i *)y[ib + 1].qs, 0);
+        bx_2 = __lsx_vsub_b(bx_2, off);
+        const __m128i i32_2 = mul_sum_i8_pairs(bx_2, by_2);
+
+        __m128i bx_3 = __lsx_vand_v(low_mask, __lsx_vsrli_d(tmp_2_3, 4));
+        __m128i by_3 = __lsx_vld((const __m128i *)(y[ib + 1].qs + 16), 0);
+        bx_3 = __lsx_vsub_b(bx_3, off);
+        const __m128i i32_3 = mul_sum_i8_pairs(bx_3, by_3);
+
+        // Convert int32_t to float
+        __m128 p0 = __lsx_vffint_s_w(i32_0);
+        __m128 p1 = __lsx_vffint_s_w(i32_1);
+        __m128 p2 = __lsx_vffint_s_w(i32_2);
+        __m128 p3 = __lsx_vffint_s_w(i32_3);
+
+        // Apply the scale
+        __m128 p0_d = __lsx_vfmul_s( d_0_1, p0 );
+        __m128 p1_d = __lsx_vfmul_s( d_0_1, p1 );
+        __m128 p2_d = __lsx_vfmul_s( d_2_3, p2 );
+        __m128 p3_d = __lsx_vfmul_s( d_2_3, p3 );
+
+        // Acummulate
+        acc_0 = __lsx_vfadd_s(p0_d, acc_0);
+        acc_1 = __lsx_vfadd_s(p1_d, acc_1);
+        acc_2 = __lsx_vfadd_s(p2_d, acc_2);
+        acc_3 = __lsx_vfadd_s(p3_d, acc_3);
+    }
+
+    sumf = hsum_float_4x4(acc_0, acc_1, acc_2, acc_3);
+
+#endif
+    for (; ib < nb; ++ib) {
+        int sumi0 = 0;
+        int sumi1 = 0;
+
+        for (int j = 0; j < qk/2; ++j) {
+            const int v0 = (x[ib].qs[j] & 0x0F) - 8;
+            const int v1 = (x[ib].qs[j] >>   4) - 8;
+
+            sumi0 += (v0 * y[ib].qs[j]);
+            sumi1 += (v1 * y[ib].qs[j + qk/2]);
+        }
+
+        int sumi = sumi0 + sumi1;
+        sumf += sumi*GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d);
+    }
+
+    *s = sumf;
+}
+
+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;
+
+    assert(n % qk == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q4_1 * GGML_RESTRICT x = vx;
+    const block_q8_1 * GGML_RESTRICT y = vy;
+
+    int ib = 0;
+    float sumf = 0;
+
+#if defined(__loongarch_asx)
+    // Initialize accumulator with zeros
+    __m256 acc = (__m256)__lasx_xvldi(0);
+
+    float summs = 0;
+
+    // Main loop
+    for (; ib < nb; ++ib) {
+        const float d0 = GGML_CPU_FP16_TO_FP32(x[ib].d);
+        const float d1 = GGML_CPU_FP16_TO_FP32(y[ib].d);
+
+        summs += GGML_CPU_FP16_TO_FP32(x[ib].m) * GGML_CPU_FP16_TO_FP32(y[ib].s);
+
+        const __m256 d0v = __lasx_xvreplfr2vr_s( d0 );
+        const __m256 d1v = __lasx_xvreplfr2vr_s( d1 );
+
+        // Compute combined scales
+        const __m256 d0d1 = __lasx_xvfmul_s( d0v, d1v );
+
+        // Load 16 bytes, and unpack 4 bit fields into bytes, making 32 bytes
+        const __m256i qx = bytes_from_nibbles_32(x[ib].qs);
+        const __m256i qy = __lasx_xvld( (const __m256i *)y[ib].qs, 0);
+
+        const __m256 xy = mul_sum_us8_pairs_float(qx, qy);
+
+        // Accumulate d0*d1*x*y
+        acc = __lasx_xvfmadd_s( d0d1, xy, acc );
+    }
+
+    sumf = hsum_float_8(acc) + summs;
+
+#endif
+    for (; ib < nb; ++ib) {
+        int sumi0 = 0;
+        int sumi1 = 0;
+
+        for (int j = 0; j < qk/2; ++j) {
+            const int v0 = (x[ib].qs[j] & 0x0F);
+            const int v1 = (x[ib].qs[j] >>   4);
+
+            sumi0 += (v0 * y[ib].qs[j]);
+            sumi1 += (v1 * y[ib].qs[j + qk/2]);
+        }
+
+        int sumi = sumi0 + sumi1;
+        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
+    }
+
+    *s = sumf;
+}
+
+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;
+
+    int ib = 0;
+    float sumf = 0;
+
+    assert(n % qk == 0);
+    assert(qk == QK5_0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q5_0 * GGML_RESTRICT x = vx;
+    const block_q8_0 * GGML_RESTRICT y = vy;
+
+#if defined(__loongarch_asx)
+    // Initialize accumulator with zeros
+    __m256 acc = (__m256)__lasx_xvldi(0);
+
+    // Main loop
+    for (; ib < nb; ++ib) {
+        /* Compute combined scale for the block */
+        const __m256 d = __lasx_xvreplfr2vr_s(GGML_CPU_FP16_TO_FP32(x[ib].d) * GGML_CPU_FP16_TO_FP32(y[ib].d)); //FIXME
+
+        __m256i qx = bytes_from_nibbles_32(x[ib].qs);
+        __m256i bxhi = bytes_from_bits_32(x[ib].qh);
+        bxhi = __lasx_xvandn_v(bxhi, __lasx_xvreplgr2vr_b((char)0xF0));
+        qx = __lasx_xvor_v(qx, bxhi);
+
+        __m256i qy = __lasx_xvld((const __m256i *)y[ib].qs, 0);
+
+        const __m256 q = mul_sum_i8_pairs_float(qx, qy);
+
+        /* Multiply q with scale and accumulate */
+        acc = __lasx_xvfmadd_s(d, q, acc);
+    }
+
+    sumf = hsum_float_8(acc);
+
+#endif
+    for (; ib < nb; ++ib) {
+        uint32_t qh;
+        memcpy(&qh, x[ib].qh, sizeof(qh));
+
+        int sumi0 = 0;
+        int sumi1 = 0;
+
+        for (int j = 0; j < qk/2; ++j) {
+            const uint8_t xh_0 = ((qh & (1u << (j + 0 ))) >> (j + 0 )) << 4;
+            const uint8_t xh_1 = ((qh & (1u << (j + 16))) >> (j + 12));
+
+            const int32_t x0 = (int8_t)(((x[ib].qs[j] & 0x0F) | xh_0) - 16);
+            const int32_t x1 = (int8_t)(((x[ib].qs[j] >>   4) | xh_1) - 16);
+
+            sumi0 += (x0 * y[ib].qs[j]);
+            sumi1 += (x1 * y[ib].qs[j + qk/2]);
+        }
+
+        int sumi = sumi0 + sumi1;
+        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d)) * sumi;
+    }
+
+    *s = sumf;
+}
+
+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;
+
+    int ib = 0;
+    float sumf = 0;
+
+    assert(n % qk == 0);
+    assert(qk == QK5_1);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q5_1 * GGML_RESTRICT x = vx;
+    const block_q8_1 * GGML_RESTRICT y = vy;
+
+#if defined(__loongarch_asx)
+    // Initialize accumulator with zeros
+    __m256 acc = (__m256)__lasx_xvldi(0);
+
+    float summs = 0.0f;
+
+    // Main loop
+    for (; ib < nb; ++ib) {
+        const __m256 dx = __lasx_xvreplfr2vr_s(GGML_CPU_FP16_TO_FP32(x[ib].d));
+
+        summs += GGML_CPU_FP16_TO_FP32(x[ib].m) * GGML_CPU_FP16_TO_FP32(y[ib].s);
+
+        __m256i qx = bytes_from_nibbles_32(x[ib].qs);
+        __m256i bxhi = bytes_from_bits_32(x[ib].qh);
+        bxhi = __lasx_xvand_v(bxhi, __lasx_xvreplgr2vr_b(0x10));
+        qx = __lasx_xvor_v(qx, bxhi);
+
+        const __m256 dy = __lasx_xvreplfr2vr_s(GGML_CPU_FP16_TO_FP32(y[ib].d));
+        const __m256i qy = __lasx_xvld((const __m256i *)y[ib].qs, 0);
+
+        const __m256 q = mul_sum_us8_pairs_float(qx, qy);
+
+        acc = __lasx_xvfmadd_s(q, __lasx_xvfmul_s(dx, dy), acc);
+    }
+
+    sumf = hsum_float_8(acc) + summs;
+
+#endif
+    for (; ib < nb; ++ib) {
+        uint32_t qh;
+        memcpy(&qh, x[ib].qh, sizeof(qh));
+
+        int sumi0 = 0;
+        int sumi1 = 0;
+
+        for (int j = 0; j < qk/2; ++j) {
+            const uint8_t xh_0 = ((qh >> (j +  0)) << 4) & 0x10;
+            const uint8_t xh_1 = ((qh >> (j + 12))     ) & 0x10;
+
+            const int32_t x0 = (x[ib].qs[j] & 0xF) | xh_0;
+            const int32_t x1 = (x[ib].qs[j] >>  4) | xh_1;
+
+            sumi0 += (x0 * y[ib].qs[j]);
+            sumi1 += (x1 * y[ib].qs[j + qk/2]);
+        }
+
+        int sumi = sumi0 + sumi1;
+        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
+    }
+
+    *s = sumf;
+}
+
+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;
+
+    assert(n % qk == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q8_0 * GGML_RESTRICT x = vx;
+    const block_q8_0 * GGML_RESTRICT y = vy;
+
+    int ib = 0;
+    float sumf = 0;
+
+#if defined(__loongarch_asx)
+    // Initialize accumulator with zeros
+    __m256 acc = (__m256)__lasx_xvldi(0);
+
+    // Main loop
+    for (; ib < nb; ++ib) {
+        // Compute combined scale for the block
+        const __m256 d = __lasx_xvreplfr2vr_s(GGML_CPU_FP16_TO_FP32(x[ib].d) * GGML_CPU_FP16_TO_FP32(y[ib].d));
+        __m256i qx = __lasx_xvld((const __m256i *)x[ib].qs, 0);
+        __m256i qy = __lasx_xvld((const __m256i *)y[ib].qs, 0);
+
+        const __m256 q = mul_sum_i8_pairs_float(qx, qy);
+
+        // Multiply q with scale and accumulate
+        acc = __lasx_xvfmadd_s( d, q, acc );
+    }
+
+    sumf = hsum_float_8(acc);
+
+#endif
+    for (; ib < nb; ++ib) {
+        int sumi = 0;
+
+        for (int j = 0; j < qk; j++) {
+            sumi += x[ib].qs[j]*y[ib].qs[j];
+        }
+
+        sumf += sumi*(GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d));
+    }
+
+    *s = sumf;
+}
+
+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 * GGML_RESTRICT x = vx;
+    const block_q8_K * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+#if defined __loongarch_asx
+
+    __m256 acc = (__m256)__lasx_xvldi(0);
+
+    for (int i = 0; i < nb; ++i) {
+
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+        const float dmin = -y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
+
+        const uint8_t * GGML_RESTRICT q2 = x[i].qs;
+        const int8_t  * GGML_RESTRICT q8 = y[i].qs;
+
+        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 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);
+
+        for (int j = 0; j < QK_K/128; ++j) {
+
+            const __m256i q2bits = __lasx_xvld((const __m256i*)q2, 0); q2 += 32;
+
+            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;
+
+            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_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_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);
+
+            sumi = __lasx_xvadd_w(sumi, __lasx_xvadd_w(p0, p2));
+        }
+
+        acc = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(d), __lasx_xvffint_s_w(sumi), acc);
+
+    }
+
+    *s = hsum_float_8(acc);
+
+#else
+
+    float sumf = 0;
+
+    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;
+
+        int summs = 0;
+        for (int j = 0; j < 16; ++j) {
+            summs += y[i].bsums[j] * (sc[j] >> 4);
+        }
+
+        const float dall = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+        const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
+
+        int isum = 0;
+        int is = 0;
+        int d;
+        for (int k = 0; k < QK_K/128; ++k) {
+            int shift = 0;
+            for (int j = 0; j < 4; ++j) {
+                d = sc[is++] & 0xF;
+                int isuml = 0;
+                for (int l =  0; l < 16; ++l) isuml += q8[l] * ((q2[l] >> shift) & 3);
+                isum += d * isuml;
+                d = sc[is++] & 0xF;
+                isuml = 0;
+                for (int l = 16; l < 32; ++l) isuml += q8[l] * ((q2[l] >> shift) & 3);
+                isum += d * isuml;
+                shift += 2;
+                q8 += 32;
+            }
+            q2 += 32;
+        }
+        sumf += dall * isum - dmin * summs;
+    }
+    *s = sumf;
+#endif
+}
+
+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);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const uint32_t kmask1 = 0x03030303;
+    const uint32_t kmask2 = 0x0f0f0f0f;
+
+    const block_q3_K * GGML_RESTRICT x = vx;
+    const block_q8_K * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+#if defined __loongarch_asx
+
+    const __m128i m32 = __lsx_vreplgr2vr_b(32);
+
+    __m256 acc = (__m256)__lasx_xvldi(0);
+
+    uint32_t aux[3];
+
+    for (int i = 0; i < nb; ++i) {
+
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+        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(
+                ((aux[1] >> 4) & kmask2) | (((aux[2] >> 6) & kmask1) << 4),
+                ((aux[0] >> 4) & kmask2) | (((aux[2] >> 4) & kmask1) << 4),
+                (aux[1] & kmask2) | (((aux[2] >> 2) & kmask1) << 4),
+                (aux[0] & kmask2) | (((aux[2] >> 0) & kmask1) << 4));
+        scales128 = __lsx_vsub_b(scales128, m32);
+
+        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);
+
+        // integer accumulator
+        __m256i sumi = __lasx_xvldi(0);
+
+        for (int j = 0; j < QK_K/128; ++j) {
+            // load low 2 bits
+            const __m256i q3bits = __lasx_xvld((const __m256i*)q3, 0); q3 += 32;
+
+            // prepare low and high bits
+            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;
+            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 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_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);
+            p16_2 = __lasx_xvadd_w(p16_2, p16_3);
+            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);
+    }
+
+    *s = hsum_float_8(acc);
+
+#else
+    // scalar version
+    // This function is written like this so the compiler can manage to vectorize most of it
+    // Using -Ofast, GCC and clang manage to produce code that is within a factor of 2 or so from the
+    // manually vectorized version above. Every other version I tried would run at least 4 times slower.
+    // The ideal situation would be if we could just write the code once, and the compiler would
+    // automatically produce the best possible set of machine instructions, instead of us having to manually
+    // write vectorized versions for AVX, ARM_NEON, etc.
+
+    int8_t  aux8[QK_K];
+    int16_t aux16[8];
+    float   sums [8];
+    int32_t aux32[8];
+    memset(sums, 0, 8*sizeof(float));
+
+    uint32_t auxs[4];
+    const int8_t * scales = (const int8_t*)auxs;
+
+    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;
+        memset(aux32, 0, 8*sizeof(int32_t));
+        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;
+            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
+            a += 32; m <<= 1;
+            for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 2) & 3;
+            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
+            a += 32; m <<= 1;
+            for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 4) & 3;
+            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
+            a += 32; m <<= 1;
+            for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 6) & 3;
+            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
+            a += 32; m <<= 1;
+            q3 += 32;
+        }
+        a = aux8;
+
+        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);
+        for (int j = 0; j < QK_K/16; ++j) {
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
+            q8 += 8; a += 8;
+        }
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
+        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
+    }
+    for (int l = 0; l < 8; ++l) sumf += sums[l];
+    *s = sumf;
+
+#endif
+
+}
+
+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);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q4_K * GGML_RESTRICT x = vx;
+    const block_q8_K * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+    static const uint32_t kmask1 = 0x3f3f3f3f;
+    static const uint32_t kmask2 = 0x0f0f0f0f;
+    static const uint32_t kmask3 = 0x03030303;
+
+    uint32_t utmp[4];
+
+#if defined __loongarch_asx
+
+    __m256 acc = (__m256)__lasx_xvldi(0);
+    __m128 acc_m = (__m128)__lsx_vldi(0);
+
+   for (int i = 0; i < nb; ++i) {
+
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+        const float dmin = -y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
+
+        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 uint8_t * GGML_RESTRICT q4 = x[i].qs;
+        const int8_t  * GGML_RESTRICT q8 = y[i].qs;
+
+        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(mins128, q8s);
+        acc_m = __lsx_vfmadd_s(__lsx_vreplfr2vr_s(dmin), __lsx_vffint_s_w(prod), acc_m);
+
+        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_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_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_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_madd_h_b(q4h, q8h);
+            p16h = lasx_madd_h(scale_h, p16h);
+            const __m256i sumj = __lasx_xvadd_w(p16l, p16h);
+
+            sumi = __lasx_xvadd_w(sumi, sumj);
+        }
+
+        __m256 vd = __lasx_xvreplfr2vr_s(d);
+        acc = __lasx_xvfmadd_s(vd, __lasx_xvffint_s_w(sumi), acc);
+
+    }
+
+    acc_m = __lsx_vfadd_s(acc_m, (__m128)__lsx_vpermi_w((__m128i)acc_m, (__m128i)acc_m, 0xee));
+    __m128i tmp1 = __lsx_vinsgr2vr_w(__lsx_vldi(0), __lsx_vpickve2gr_w((__m128i)acc_m, 1), 0);
+    acc_m = __lsx_vfadd_s(acc_m, (__m128)tmp1);
+
+
+    *s = hsum_float_8(acc) + ((v4f32)acc_m)[0];
+
+#else
+
+    const uint8_t * scales = (const uint8_t*)&utmp[0];
+    const uint8_t * mins   = (const uint8_t*)&utmp[2];
+
+    int8_t  aux8[QK_K];
+    int16_t aux16[8];
+    float   sums [8];
+    int32_t aux32[8];
+    memset(sums, 0, 8*sizeof(float));
+
+    float sumf = 0;
+    for (int i = 0; i < nb; ++i) {
+        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 * 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;
+            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l]  >> 4);
+            a += 32; q4 += 32;
+        }
+        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;
+
+        int sumi = 0;
+        for (int j = 0; j < QK_K/16; ++j) sumi += y[i].bsums[j] * mins[j/2];
+        a = aux8;
+        int is = 0;
+        for (int j = 0; j < QK_K/32; ++j) {
+            int32_t scale = scales[is++];
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+        }
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
+        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
+        const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
+        sumf -= dmin * sumi;
+    }
+    for (int l = 0; l < 8; ++l) sumf += sums[l];
+    *s = sumf;
+#endif
+}
+
+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);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q5_K * GGML_RESTRICT x = vx;
+    const block_q8_K * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+    static const uint32_t kmask1 = 0x3f3f3f3f;
+    static const uint32_t kmask2 = 0x0f0f0f0f;
+    static const uint32_t kmask3 = 0x03030303;
+
+    uint32_t utmp[4];
+
+#if defined __loongarch_asx
+
+    __m256 acc = (__m256)__lasx_xvldi(0);
+    __m128 acc_m = (__m128)__lsx_vldi(0);
+
+    for (int i = 0; i < nb; ++i) {
+
+        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_CPU_FP16_TO_FP32(x[i].d);
+        const float dmin = -y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
+
+        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 __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(mins128, q8s);
+        acc_m = __lsx_vfmadd_s(__lsx_vreplfr2vr_s(dmin), __lsx_vffint_s_w(prod), acc_m);
+
+        const __m256i scales = lasx_insertf128(scales128, scales128);
+
+        const __m256i hbits = __lasx_xvld((const __m256i*)x[i].qh, 0);
+
+        __m256i sumi = __lasx_xvldi(0);
+
+        for (int j = 0; j < QK_K/64; ++j) {
+
+            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;
+
+            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_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);
+
+            sumi = __lasx_xvadd_w(sumi, __lasx_xvadd_w(p16_0, p16_1));
+
+        }
+
+        __m256 vd = __lasx_xvreplfr2vr_s(d);
+        acc = __lasx_xvfmadd_s(vd, __lasx_xvffint_s_w(sumi), acc);
+
+    }
+
+    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];
+
+#else
+
+    const uint8_t * scales = (const uint8_t*)&utmp[0];
+    const uint8_t * mins   = (const uint8_t*)&utmp[2];
+
+    int8_t  aux8[QK_K];
+    int16_t aux16[8];
+    float   sums [8];
+    int32_t aux32[8];
+    memset(sums, 0, 8*sizeof(float));
+
+    float sumf = 0;
+    for (int i = 0; i < nb; ++i) {
+        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 * 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);
+            for (int l = 0; l < 32; ++l) a[l] += (hm[l] & m ? 16 : 0);
+            a += 32; m <<= 1;
+            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l]  >> 4);
+            for (int l = 0; l < 32; ++l) a[l] += (hm[l] & m ? 16 : 0);
+            a += 32; m <<= 1;
+            q4 += 32;
+        }
+        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;
+
+        int sumi = 0;
+        for (int j = 0; j < QK_K/16; ++j) sumi += y[i].bsums[j] * mins[j/2];
+        a = aux8;
+        int is = 0;
+        for (int j = 0; j < QK_K/32; ++j) {
+            int32_t scale = scales[is++];
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+        }
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
+        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
+        const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
+        sumf -= dmin * sumi;
+    }
+    for (int l = 0; l < 8; ++l) sumf += sums[l];
+    *s = sumf;
+#endif
+}
+
+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);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q6_K * GGML_RESTRICT x = vx;
+    const block_q8_K * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+#if defined __loongarch_asx
+
+    const __m256i m32s = __lasx_xvreplgr2vr_b(32);
+
+    __m256 acc = (__m256)__lasx_xvldi(0);
+
+    for (int i = 0; i < nb; ++i) {
+
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+
+        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 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);
+
+        for (int j = 0; j < QK_K/128; ++j) {
+
+            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_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_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 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);
+
+            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));
+        }
+
+        acc = __lasx_xvfmadd_s((__m256)__lasx_xvreplfr2vr_s(d), __lasx_xvffint_s_w(sumi), acc);
+    }
+
+    *s = hsum_float_8(acc);
+
+#else
+
+    int8_t  aux8[QK_K];
+    int16_t aux16[8];
+    float   sums [8];
+    int32_t aux32[8];
+    memset(sums, 0, 8*sizeof(float));
+
+    float sumf = 0;
+    for (int i = 0; i < nb; ++i) {
+        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 * 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;
+                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;
+        }
+        a = aux8;
+        int is = 0;
+        for (int j = 0; j < QK_K/16; ++j) {
+            int scale = x[i].scales[is++];
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+        }
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
+        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
+    }
+    for (int l = 0; l < 8; ++l) sumf += sums[l];
+    *s = sumf;
+#endif
+}
+
+#if defined(__loongarch_asx)
+static const int8_t keven_signs_q2xs[1024] = {
+     1,  1,  1,  1,  1,  1,  1,  1, -1,  1,  1,  1,  1,  1,  1, -1,  1, -1,  1,  1,  1,  1,  1, -1, -1, -1,  1,  1,  1,  1,  1,  1,
+     1,  1, -1,  1,  1,  1,  1, -1, -1,  1, -1,  1,  1,  1,  1,  1,  1, -1, -1,  1,  1,  1,  1,  1, -1, -1, -1,  1,  1,  1,  1, -1,
+     1,  1,  1, -1,  1,  1,  1, -1, -1,  1,  1, -1,  1,  1,  1,  1,  1, -1,  1, -1,  1,  1,  1,  1, -1, -1,  1, -1,  1,  1,  1, -1,
+     1,  1, -1, -1,  1,  1,  1,  1, -1,  1, -1, -1,  1,  1,  1, -1,  1, -1, -1, -1,  1,  1,  1, -1, -1, -1, -1, -1,  1,  1,  1,  1,
+     1,  1,  1,  1, -1,  1,  1, -1, -1,  1,  1,  1, -1,  1,  1,  1,  1, -1,  1,  1, -1,  1,  1,  1, -1, -1,  1,  1, -1,  1,  1, -1,
+     1,  1, -1,  1, -1,  1,  1,  1, -1,  1, -1,  1, -1,  1,  1, -1,  1, -1, -1,  1, -1,  1,  1, -1, -1, -1, -1,  1, -1,  1,  1,  1,
+     1,  1,  1, -1, -1,  1,  1,  1, -1,  1,  1, -1, -1,  1,  1, -1,  1, -1,  1, -1, -1,  1,  1, -1, -1, -1,  1, -1, -1,  1,  1,  1,
+     1,  1, -1, -1, -1,  1,  1, -1, -1,  1, -1, -1, -1,  1,  1,  1,  1, -1, -1, -1, -1,  1,  1,  1, -1, -1, -1, -1, -1,  1,  1, -1,
+     1,  1,  1,  1,  1, -1,  1, -1, -1,  1,  1,  1,  1, -1,  1,  1,  1, -1,  1,  1,  1, -1,  1,  1, -1, -1,  1,  1,  1, -1,  1, -1,
+     1,  1, -1,  1,  1, -1,  1,  1, -1,  1, -1,  1,  1, -1,  1, -1,  1, -1, -1,  1,  1, -1,  1, -1, -1, -1, -1,  1,  1, -1,  1,  1,
+     1,  1,  1, -1,  1, -1,  1,  1, -1,  1,  1, -1,  1, -1,  1, -1,  1, -1,  1, -1,  1, -1,  1, -1, -1, -1,  1, -1,  1, -1,  1,  1,
+     1,  1, -1, -1,  1, -1,  1, -1, -1,  1, -1, -1,  1, -1,  1,  1,  1, -1, -1, -1,  1, -1,  1,  1, -1, -1, -1, -1,  1, -1,  1, -1,
+     1,  1,  1,  1, -1, -1,  1,  1, -1,  1,  1,  1, -1, -1,  1, -1,  1, -1,  1,  1, -1, -1,  1, -1, -1, -1,  1,  1, -1, -1,  1,  1,
+     1,  1, -1,  1, -1, -1,  1, -1, -1,  1, -1,  1, -1, -1,  1,  1,  1, -1, -1,  1, -1, -1,  1,  1, -1, -1, -1,  1, -1, -1,  1, -1,
+     1,  1,  1, -1, -1, -1,  1, -1, -1,  1,  1, -1, -1, -1,  1,  1,  1, -1,  1, -1, -1, -1,  1,  1, -1, -1,  1, -1, -1, -1,  1, -1,
+     1,  1, -1, -1, -1, -1,  1,  1, -1,  1, -1, -1, -1, -1,  1, -1,  1, -1, -1, -1, -1, -1,  1, -1, -1, -1, -1, -1, -1, -1,  1,  1,
+     1,  1,  1,  1,  1,  1, -1, -1, -1,  1,  1,  1,  1,  1, -1,  1,  1, -1,  1,  1,  1,  1, -1,  1, -1, -1,  1,  1,  1,  1, -1, -1,
+     1,  1, -1,  1,  1,  1, -1,  1, -1,  1, -1,  1,  1,  1, -1, -1,  1, -1, -1,  1,  1,  1, -1, -1, -1, -1, -1,  1,  1,  1, -1,  1,
+     1,  1,  1, -1,  1,  1, -1,  1, -1,  1,  1, -1,  1,  1, -1, -1,  1, -1,  1, -1,  1,  1, -1, -1, -1, -1,  1, -1,  1,  1, -1,  1,
+     1,  1, -1, -1,  1,  1, -1, -1, -1,  1, -1, -1,  1,  1, -1,  1,  1, -1, -1, -1,  1,  1, -1,  1, -1, -1, -1, -1,  1,  1, -1, -1,
+     1,  1,  1,  1, -1,  1, -1,  1, -1,  1,  1,  1, -1,  1, -1, -1,  1, -1,  1,  1, -1,  1, -1, -1, -1, -1,  1,  1, -1,  1, -1,  1,
+     1,  1, -1,  1, -1,  1, -1, -1, -1,  1, -1,  1, -1,  1, -1,  1,  1, -1, -1,  1, -1,  1, -1,  1, -1, -1, -1,  1, -1,  1, -1, -1,
+     1,  1,  1, -1, -1,  1, -1, -1, -1,  1,  1, -1, -1,  1, -1,  1,  1, -1,  1, -1, -1,  1, -1,  1, -1, -1,  1, -1, -1,  1, -1, -1,
+     1,  1, -1, -1, -1,  1, -1,  1, -1,  1, -1, -1, -1,  1, -1, -1,  1, -1, -1, -1, -1,  1, -1, -1, -1, -1, -1, -1, -1,  1, -1,  1,
+     1,  1,  1,  1,  1, -1, -1,  1, -1,  1,  1,  1,  1, -1, -1, -1,  1, -1,  1,  1,  1, -1, -1, -1, -1, -1,  1,  1,  1, -1, -1,  1,
+     1,  1, -1,  1,  1, -1, -1, -1, -1,  1, -1,  1,  1, -1, -1,  1,  1, -1, -1,  1,  1, -1, -1,  1, -1, -1, -1,  1,  1, -1, -1, -1,
+     1,  1,  1, -1,  1, -1, -1, -1, -1,  1,  1, -1,  1, -1, -1,  1,  1, -1,  1, -1,  1, -1, -1,  1, -1, -1,  1, -1,  1, -1, -1, -1,
+     1,  1, -1, -1,  1, -1, -1,  1, -1,  1, -1, -1,  1, -1, -1, -1,  1, -1, -1, -1,  1, -1, -1, -1, -1, -1, -1, -1,  1, -1, -1,  1,
+     1,  1,  1,  1, -1, -1, -1, -1, -1,  1,  1,  1, -1, -1, -1,  1,  1, -1,  1,  1, -1, -1, -1,  1, -1, -1,  1,  1, -1, -1, -1, -1,
+     1,  1, -1,  1, -1, -1, -1,  1, -1,  1, -1,  1, -1, -1, -1, -1,  1, -1, -1,  1, -1, -1, -1, -1, -1, -1, -1,  1, -1, -1, -1,  1,
+     1,  1,  1, -1, -1, -1, -1,  1, -1,  1,  1, -1, -1, -1, -1, -1,  1, -1,  1, -1, -1, -1, -1, -1, -1, -1,  1, -1, -1, -1, -1,  1,
+     1,  1, -1, -1, -1, -1, -1, -1, -1,  1, -1, -1, -1, -1, -1,  1,  1, -1, -1, -1, -1, -1, -1,  1, -1, -1, -1, -1, -1, -1, -1, -1,
+};
+#endif
+
+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);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_iq2_xxs * GGML_RESTRICT x = vx;
+    const block_q8_K    * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+#if defined(__loongarch_asx)
+
+    const uint64_t * signs64 = (const uint64_t *)keven_signs_q2xs;
+
+    uint32_t aux32[4];
+    const uint8_t * aux8 = (const uint8_t *)aux32;
+
+    __m256 accumf = (__m256)__lasx_xvldi(0);
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_CPU_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;
+        __m256i sumi1 = __lasx_xvldi(0);
+        __m256i sumi2 = __lasx_xvldi(0);
+        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
+            const __m256i q8_1 = __lasx_xvld((const __m256i *)q8, 0); q8 += 32;
+            const __m256i q8_2 = __lasx_xvld((const __m256i *)q8, 0); q8 += 32;
+            memcpy(aux32, q2, 4*sizeof(uint32_t)); q2 += 8;
+
+            const __m256i q2_1 = lasx_set_d(iq2xxs_grid[aux8[ 3]], iq2xxs_grid[aux8[ 2]], iq2xxs_grid[aux8[1]], iq2xxs_grid[aux8[0]]);
+            const __m256i q2_2 = lasx_set_d(iq2xxs_grid[aux8[11]], iq2xxs_grid[aux8[10]], iq2xxs_grid[aux8[9]], iq2xxs_grid[aux8[8]]);
+            const __m256i s2_1 = lasx_set_d(signs64[(aux32[1] >> 21) & 127], signs64[(aux32[1] >> 14) & 127],
+                                                   signs64[(aux32[1] >>  7) & 127], signs64[(aux32[1] >>  0) & 127]);
+            const __m256i s2_2 = lasx_set_d(signs64[(aux32[3] >> 21) & 127], signs64[(aux32[3] >> 14) & 127],
+                                                   signs64[(aux32[3] >>  7) & 127], signs64[(aux32[3] >>  0) & 127]);
+            const __m256i q8s_1 = __lasx_xvsigncov_b(s2_1, q8_1);
+            const __m256i q8s_2 = __lasx_xvsigncov_b(s2_2, q8_2);
+            const __m256i dot1  = lasx_maddubs_h(q2_1, q8s_1);
+            const __m256i dot2  = lasx_maddubs_h(q2_2, q8s_2);
+            const uint16_t ls1 = aux32[1] >> 28;
+            const uint16_t ls2 = aux32[3] >> 28;
+            const __m256i p1 = lasx_madd_h(dot1, __lasx_xvreplgr2vr_h(2*ls1+1));
+            const __m256i p2 = lasx_madd_h(dot2, __lasx_xvreplgr2vr_h(2*ls2+1));
+            sumi1 = __lasx_xvadd_w(sumi1, p1);
+            sumi2 = __lasx_xvadd_w(sumi2, p2);
+        }
+
+        accumf = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(d), __lasx_xvffint_s_w(__lasx_xvadd_w(sumi1, sumi2)), accumf);
+    }
+
+    *s = 0.125f * hsum_float_8(accumf);
+
+#else
+
+    uint32_t aux32[2];
+    const uint8_t * aux8 = (const uint8_t *)aux32;
+
+    float sumf = 0.f;
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_CPU_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;
+        int32_t bsum = 0;
+        for (int ib32 = 0; ib32 < QK_K/32; ++ib32) {
+            memcpy(aux32, q2, 2*sizeof(uint32_t));
+            q2 += 4;
+            const uint32_t ls = 2*(aux32[1] >> 28) + 1;
+            int32_t sumi = 0;
+            for (int l = 0; l < 4; ++l) {
+                const uint8_t * grid = (const uint8_t *)(iq2xxs_grid + aux8[l]);
+                const uint8_t  signs = ksigns_iq2xs[(aux32[1] >> 7*l) & 127];
+                for (int j = 0; j < 8; ++j) {
+                    sumi += grid[j] * q8[j] * (signs & kmask_iq2xs[j] ? -1 : 1);
+                }
+                q8 += 8;
+            }
+            bsum += sumi * ls;
+        }
+        sumf += d * bsum;
+    }
+    *s = 0.125f * sumf;
+#endif
+}
+
+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);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_iq2_xs * GGML_RESTRICT x = vx;
+    const block_q8_K   * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+#if defined(__loongarch_asx)
+
+    const __m256i mone = __lasx_xvreplgr2vr_b(1);
+    static const char block_sign_shuffle_mask_1[32] = {
+        0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02,
+        0x04, 0x04, 0x04, 0x04, 0x04, 0x04, 0x04, 0x04, 0x06, 0x06, 0x06, 0x06, 0x06, 0x06, 0x06, 0x06,
+    };
+    static const char block_sign_shuffle_mask_2[32] = {
+        0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x0a, 0x0a, 0x0a, 0x0a, 0x0a, 0x0a, 0x0a, 0x0a,
+        0x0c, 0x0c, 0x0c, 0x0c, 0x0c, 0x0c, 0x0c, 0x0c, 0x0e, 0x0e, 0x0e, 0x0e, 0x0e, 0x0e, 0x0e, 0x0e,
+    };
+    static const uint8_t bit_selector_mask_bytes[32] = {
+        0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
+        0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
+    };
+
+    const __m256i bit_selector_mask = __lasx_xvld((const __m256i*)bit_selector_mask_bytes, 0);
+    const __m256i block_sign_shuffle_1 = __lasx_xvld((const __m256i*)block_sign_shuffle_mask_1, 0);
+    const __m256i block_sign_shuffle_2 = __lasx_xvld((const __m256i*)block_sign_shuffle_mask_2, 0);
+
+    static const uint8_t k_bit_helper[32] = {
+        0x00, 0x80, 0x80, 0x00, 0x80, 0x00, 0x00, 0x80, 0x80, 0x00, 0x00, 0x80, 0x00, 0x80, 0x80, 0x00,
+        0x00, 0x80, 0x80, 0x00, 0x80, 0x00, 0x00, 0x80, 0x80, 0x00, 0x00, 0x80, 0x00, 0x80, 0x80, 0x00,
+    };
+    const __m256i bit_helper = __lasx_xvld((const __m256i*)k_bit_helper, 0);
+    const __m256i m511 = __lasx_xvreplgr2vr_h(511);
+    const __m128i m4 = __lsx_vreplgr2vr_b(0xf);
+    const __m128i m1 = __lsx_vreplgr2vr_b(1);
+
+    uint64_t aux64;
+
+    // somewhat hacky, but gives a significant boost in performance
+    __m256i aux_gindex;
+    const uint16_t * gindex = (const uint16_t *)&aux_gindex;
+
+    __m256 accumf = (__m256)__lasx_xvldi(0);
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_CPU_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;
+
+        memcpy(&aux64, x[i].scales, 8);
+        __m128i stmp = __lsx_vreplgr2vr_d(aux64);
+        stmp = __lsx_vilvl_b( __lsx_vand_v(__lsx_vsrli_h(stmp, 4), m4), __lsx_vand_v(stmp, m4));
+        const __m128i scales = __lsx_vadd_b(__lsx_vslli_h(stmp, 1), m1);
+
+        __m256i sumi1 = __lasx_xvldi(0);
+        __m256i sumi2 = __lasx_xvldi(0);
+        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 4) {
+
+            const __m256i q2_data = __lasx_xvld((const __m256i*)q2, 0);  q2 += 16;
+            aux_gindex = __lasx_xvand_v(q2_data, m511);
+
+            const __m256i partial_sign_bits = __lasx_xvsrli_h(q2_data, 9);
+            const __m256i partial_sign_bits_upper = __lasx_xvsrli_h(q2_data, 13);
+            const __m256i partial_sign_bits_for_counting = __lasx_xvxor_v(partial_sign_bits, partial_sign_bits_upper);
+
+            const __m256i odd_bits = lasx_shuffle_b(bit_helper, partial_sign_bits_for_counting);
+            const __m256i full_sign_bits = __lasx_xvor_v(partial_sign_bits, odd_bits);
+
+            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;
+            const __m256i q8_4 = __lasx_xvld((const __m256i *)q8, 0); q8 += 32;
+
+            const __m256i q2_1 = lasx_set_d(iq2xs_grid[gindex[ 3]], iq2xs_grid[gindex[ 2]],
+                                                   iq2xs_grid[gindex[ 1]], iq2xs_grid[gindex[ 0]]);
+            const __m256i q2_2 = lasx_set_d(iq2xs_grid[gindex[ 7]], iq2xs_grid[gindex[ 6]],
+                                                   iq2xs_grid[gindex[ 5]], iq2xs_grid[gindex[ 4]]);
+            const __m256i q2_3 = lasx_set_d(iq2xs_grid[gindex[11]], iq2xs_grid[gindex[10]],
+                                                   iq2xs_grid[gindex[ 9]], iq2xs_grid[gindex[ 8]]);
+            const __m256i q2_4 = lasx_set_d(iq2xs_grid[gindex[15]], iq2xs_grid[gindex[14]],
+                                                   iq2xs_grid[gindex[13]], iq2xs_grid[gindex[12]]);
+
+            const __m128i full_signs_l = lasx_extracti128(full_sign_bits, 0);
+            const __m128i full_signs_h = lasx_extracti128(full_sign_bits, 1);
+            const __m256i full_signs_1 = lasx_insertf128(full_signs_l, full_signs_l);
+            const __m256i full_signs_2 = lasx_insertf128(full_signs_h, full_signs_h);
+
+            __m256i signs;
+            signs = lasx_shuffle_b(full_signs_1, block_sign_shuffle_1);
+            signs = __lasx_xvseq_b(__lasx_xvand_v(signs, bit_selector_mask), bit_selector_mask);
+            const __m256i q8s_1 = __lasx_xvsigncov_b(__lasx_xvor_v(signs, mone), q8_1);
+
+            signs = lasx_shuffle_b(full_signs_1, block_sign_shuffle_2);
+            signs = __lasx_xvseq_b(__lasx_xvand_v(signs, bit_selector_mask), bit_selector_mask);
+            const __m256i q8s_2 = __lasx_xvsigncov_b(__lasx_xvor_v(signs, mone), q8_2);
+
+            signs = lasx_shuffle_b(full_signs_2, block_sign_shuffle_1);
+            signs = __lasx_xvseq_b(__lasx_xvand_v(signs, bit_selector_mask), bit_selector_mask);
+            const __m256i q8s_3 = __lasx_xvsigncov_b(__lasx_xvor_v(signs, mone), q8_3);
+
+            signs = lasx_shuffle_b(full_signs_2, block_sign_shuffle_2);
+            signs = __lasx_xvseq_b(__lasx_xvand_v(signs, bit_selector_mask), bit_selector_mask);
+            const __m256i q8s_4 = __lasx_xvsigncov_b(__lasx_xvor_v(signs, mone), q8_4);
+
+            const __m256i dot1  = lasx_maddubs_h(q2_1, q8s_1);
+            const __m256i dot2  = lasx_maddubs_h(q2_2, q8s_2);
+            const __m256i dot3  = lasx_maddubs_h(q2_3, q8s_3);
+            const __m256i dot4  = lasx_maddubs_h(q2_4, q8s_4);
+
+            const __m256i sc1 = lasx_ext8_16(lsx_shuffle_b(scales, get_scale_shuffle(ib32+0)));
+            const __m256i sc2 = lasx_ext8_16(lsx_shuffle_b(scales, get_scale_shuffle(ib32+1)));
+            const __m256i sc3 = lasx_ext8_16(lsx_shuffle_b(scales, get_scale_shuffle(ib32+2)));
+            const __m256i sc4 = lasx_ext8_16(lsx_shuffle_b(scales, get_scale_shuffle(ib32+3)));
+
+            sumi1 = __lasx_xvadd_w(sumi1, lasx_madd_h(dot1, sc1));
+            sumi2 = __lasx_xvadd_w(sumi2, lasx_madd_h(dot2, sc2));
+            sumi1 = __lasx_xvadd_w(sumi1, lasx_madd_h(dot3, sc3));
+            sumi2 = __lasx_xvadd_w(sumi2, lasx_madd_h(dot4, sc4));
+        }
+
+        accumf = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(d), __lasx_xvffint_s_w(__lasx_xvadd_w(sumi1, sumi2)), accumf);
+
+    }
+
+    *s = 0.125f * hsum_float_8(accumf);
+
+#else
+
+    float sumf = 0.f;
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
+        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;
+            const uint16_t ls2 = 2*(sc[ib32] >>  4) + 1;
+            int32_t sumi = 0;
+            for (int l = 0; l < 2; ++l) {
+                const uint8_t * grid = (const uint8_t *)(iq2xs_grid + (q2[l] & 511));
+                const uint8_t  signs = ksigns_iq2xs[q2[l] >> 9];
+                for (int j = 0; j < 8; ++j) {
+                    sumi += grid[j] * q8[j] * (signs & kmask_iq2xs[j] ? -1 : 1);
+                }
+                q8 += 8;
+            }
+            bsum += sumi * ls1;
+            sumi = 0;
+            for (int l = 2; l < 4; ++l) {
+                const uint8_t * grid = (const uint8_t *)(iq2xs_grid + (q2[l] & 511));
+                const uint8_t  signs = ksigns_iq2xs[q2[l] >> 9];
+                for (int j = 0; j < 8; ++j) {
+                    sumi += grid[j] * q8[j] * (signs & kmask_iq2xs[j] ? -1 : 1);
+                }
+                q8 += 8;
+            }
+            bsum += sumi * ls2;
+            q2 += 4;
+        }
+        sumf += d * bsum;
+    }
+    *s = 0.125f * sumf;
+#endif
+}
+
+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);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_iq2_s * GGML_RESTRICT x = vx;
+    const block_q8_K  * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+#if defined(__loongarch_asx)
+
+   static const uint8_t k_mask1[32] = {0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01,
+                                       0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03
+   };
+
+    static const uint8_t k_mask2[32] = {0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
+                                        0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
+    };
+
+
+    const __m128i m4 = __lsx_vreplgr2vr_b(0xf);
+    const __m128i m1 = __lsx_vreplgr2vr_b(1);
+
+    const __m256i mask1 = __lasx_xvld((const __m256i*)k_mask1, 0);
+    const __m256i mask2 = __lasx_xvld((const __m256i*)k_mask2, 0);
+    uint64_t aux64;
+
+    __m256 accumf = (__m256)__lasx_xvldi(0);
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
+        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);
+        tmp1 = __lsx_vinsgr2vr_d(tmp1, aux64, 0);
+        tmp1 = __lsx_vinsgr2vr_d(tmp1, aux64 >> 4, 1);
+        const __m128i scales8 = __lsx_vadd_b(__lsx_vslli_h(__lsx_vand_v(tmp1, m4), 1), m1);
+        const __m256i scales16 = lasx_ext8_16(scales8); // 0 2 4 6 8 10 12 14 1 3 5 7 9 11 13 15
+
+        __m256i sumi1 = __lasx_xvldi(0);
+        __m256i sumi2 = __lasx_xvldi(0);
+        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
+            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 q2_1 = lasx_set_d(iq2s_grid[qs[3] | ((qh[ib32+0] << 2) & 0x300)],
+                                                   iq2s_grid[qs[2] | ((qh[ib32+0] << 4) & 0x300)],
+                                                   iq2s_grid[qs[1] | ((qh[ib32+0] << 6) & 0x300)],
+                                                   iq2s_grid[qs[0] | ((qh[ib32+0] << 8) & 0x300)]);
+            const __m256i q2_2 = lasx_set_d(iq2s_grid[qs[7] | ((qh[ib32+1] << 2) & 0x300)],
+                                                   iq2s_grid[qs[6] | ((qh[ib32+1] << 4) & 0x300)],
+                                                   iq2s_grid[qs[5] | ((qh[ib32+1] << 6) & 0x300)],
+                                                   iq2s_grid[qs[4] | ((qh[ib32+1] << 8) & 0x300)]);
+            qs += 8;
+
+            __m256i aux256 = __lasx_xvreplgr2vr_w(signs[0] | ((uint32_t) signs[1] << 16));
+            aux256 = __lasx_xvand_v(lasx_shuffle_b(aux256,mask1), mask2);
+            const __m256i s2_1 = __lasx_xvseq_b(aux256, mask2);
+            const __m256i q8s_1 = __lasx_xvsub_b(__lasx_xvxor_v(s2_1, q8_1), s2_1);
+
+            aux256 = __lasx_xvreplgr2vr_w(signs[2] | ((uint32_t) signs[3] << 16));
+            aux256 = __lasx_xvand_v(lasx_shuffle_b(aux256,mask1), mask2);
+            const __m256i s2_2 = __lasx_xvseq_b(aux256, mask2);
+            const __m256i q8s_2 = __lasx_xvsub_b(__lasx_xvxor_v(s2_2, q8_2), s2_2);
+
+            signs += 4;
+
+            const __m256i dot1  = lasx_maddubs_h(q2_1, q8s_1); // blocks 2*ib32+0, 2*ib32+1
+            const __m256i dot2  = lasx_maddubs_h(q2_2, q8s_2); // blocks 2*ib32+2, 2*ib32+3
+
+            const __m256i p1 = lasx_madd_h(dot1, lasx_shuffle_b(scales16, get_scale_shuffle_k4(ib32+0)));
+            const __m256i p2 = lasx_madd_h(dot2, lasx_shuffle_b(scales16, get_scale_shuffle_k4(ib32+1)));
+            sumi1 = __lasx_xvadd_w(sumi1, p1);
+            sumi2 = __lasx_xvadd_w(sumi2, p2);
+        }
+
+        accumf = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(d), __lasx_xvffint_s_w(__lasx_xvadd_w(sumi1, sumi2)), accumf);
+    }
+
+    *s = 0.125f * hsum_float_8(accumf);
+
+#else
+
+    float sumf = 0;
+    for (int i = 0; i < nb; i++) {
+
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
+        const int8_t  * q8 = y[i].qs;
+        const uint8_t * qs = x[i].qs;
+        const uint8_t * qh = x[i].qh;
+        const uint8_t * signs = qs + QK_K/8;
+
+        int bsum = 0;
+        for (int ib32 = 0; ib32 < QK_K/32; ++ib32) {
+            int ls1 = 1 + 2*(x[i].scales[ib32] & 0xf);
+            int ls2 = 1 + 2*(x[i].scales[ib32] >>  4);
+            int sumi1 = 0, sumi2 = 0;
+            for (int l = 0; l < 2; ++l) {
+                const uint8_t * grid = (const uint8_t *)(iq2s_grid + (qs[l] | (qh[ib32] << (8-2*l) & 0x300)));
+                for (int j = 0; j < 8; ++j) {
+                    sumi1 += q8[j] * grid[j] * (signs[l] & kmask_iq2xs[j] ? -1 : 1);
+                }
+                q8 += 8;
+            }
+            for (int l = 2; l < 4; ++l) {
+                const uint8_t * grid = (const uint8_t *)(iq2s_grid + (qs[l] | (qh[ib32] << (8-2*l) & 0x300)));
+                for (int j = 0; j < 8; ++j) {
+                    sumi2 += q8[j] * grid[j] * (signs[l] & kmask_iq2xs[j] ? -1 : 1);
+                }
+                q8 += 8;
+            }
+            bsum += ls1 * sumi1 + ls2 * sumi2;
+            qs += 4;
+            signs += 4;
+        }
+
+        sumf += d * bsum;
+    }
+
+    *s = 0.125f * sumf;
+
+#endif
+
+}
+
+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);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_iq3_xxs * GGML_RESTRICT x = vx;
+    const block_q8_K    * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+#if defined(__loongarch_asx)
+
+    const uint64_t * signs64 = (const uint64_t *)keven_signs_q2xs;
+
+    uint32_t aux32[2];
+
+    __m256 accumf = (__m256)__lasx_xvldi(0);
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
+        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) {
+            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 q2_1 = lasx_set_w(iq3xxs_grid[q3[7]], iq3xxs_grid[q3[6]], iq3xxs_grid[q3[5]], iq3xxs_grid[q3[4]],
+                                                iq3xxs_grid[q3[3]], iq3xxs_grid[q3[2]], iq3xxs_grid[q3[1]], iq3xxs_grid[q3[0]]);
+            q3 += 8;
+            const __m256i q2_2 = lasx_set_w(iq3xxs_grid[q3[7]], iq3xxs_grid[q3[6]], iq3xxs_grid[q3[5]], iq3xxs_grid[q3[4]],
+                                                iq3xxs_grid[q3[3]], iq3xxs_grid[q3[2]], iq3xxs_grid[q3[1]], iq3xxs_grid[q3[0]]);
+            q3 += 8;
+            memcpy(aux32, gas, 8); gas += 8;
+
+            const __m256i s2_1 = lasx_set_d(signs64[(aux32[0] >> 21) & 127], signs64[(aux32[0] >> 14) & 127],
+                                                   signs64[(aux32[0] >>  7) & 127], signs64[(aux32[0] >>  0) & 127]);
+            const __m256i s2_2 = lasx_set_d(signs64[(aux32[1] >> 21) & 127], signs64[(aux32[1] >> 14) & 127],
+                                                   signs64[(aux32[1] >>  7) & 127], signs64[(aux32[1] >>  0) & 127]);
+            const __m256i q8s_1 = __lasx_xvsigncov_b(s2_1, q8_1);
+            const __m256i q8s_2 = __lasx_xvsigncov_b(s2_2, q8_2);
+            const __m256i dot1  = lasx_maddubs_h(q2_1, q8s_1);
+            const __m256i dot2  = lasx_maddubs_h(q2_2, q8s_2);
+            const uint16_t ls1 = aux32[0] >> 28;
+            const uint16_t ls2 = aux32[1] >> 28;
+
+            const __m256i p1 = lasx_madd_h(dot1, __lasx_xvreplgr2vr_h(2*ls1+1));
+            const __m256i p2 = lasx_madd_h(dot2, __lasx_xvreplgr2vr_h(2*ls2+1));
+            sumi1 = __lasx_xvadd_w(sumi1, p1);
+            sumi2 = __lasx_xvadd_w(sumi2, p2);
+        }
+
+        accumf = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(d), __lasx_xvffint_s_w(__lasx_xvadd_w(sumi1, sumi2)), accumf);
+    }
+
+    *s = 0.25f * hsum_float_8(accumf);
+
+#else
+
+    uint32_t aux32;
+
+    float sumf = 0.f;
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
+        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);
+            const uint32_t ls = 2*(aux32 >> 28) + 1;
+            int32_t sumi = 0;
+            for (int l = 0; l < 4; ++l) {
+                const uint8_t * grid1 = (const uint8_t *)(iq3xxs_grid + q3[2*l+0]);
+                const uint8_t * grid2 = (const uint8_t *)(iq3xxs_grid + q3[2*l+1]);
+                const uint8_t  signs = ksigns_iq2xs[(aux32 >> 7*l) & 127];
+                for (int j = 0; j < 4; ++j) {
+                    sumi += grid1[j] * q8[j+0] * (signs & kmask_iq2xs[j+0] ? -1 : 1);
+                    sumi += grid2[j] * q8[j+4] * (signs & kmask_iq2xs[j+4] ? -1 : 1);
+                }
+                q8 += 8;
+            }
+            q3 += 8;
+            bsum += sumi * ls;
+        }
+        sumf += d * bsum;
+    }
+    *s = 0.25f * sumf;
+#endif
+}
+
+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);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_iq3_s * GGML_RESTRICT x = vx;
+    const block_q8_K  * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+#if defined(__loongarch_asx)
+
+   static const uint8_t k_mask1[32] = {0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01,
+                                       0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03
+   };
+
+    static const uint8_t k_mask2[32] = {0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
+                                        0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
+    };
+
+    const __m256i mask1 = __lasx_xvld((const __m256i*)k_mask1, 0);
+    const __m256i mask2 = __lasx_xvld((const __m256i*)k_mask2, 0);
+
+    __m256i idx_shift = lasx_set_w(1, 2, 3, 4, 5, 6, 7, 8);
+    const __m256i idx_mask  = __lasx_xvreplgr2vr_w(256);
+
+    typedef union {
+        __m256i  vec[2];
+        uint32_t index[16];
+    } index_t;
+
+    index_t idx;
+
+    __m256 accumf = (__m256)__lasx_xvldi(0);
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
+        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) {
+            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 idx_l = lasx_extu8_16(__lsx_vld(qs, 0)); qs += 16;
+            idx.vec[0] = __lasx_xvreplgr2vr_w(qh[ib32+0]);
+            idx.vec[1] = __lasx_xvreplgr2vr_w(qh[ib32+1]);
+            idx.vec[0] = __lasx_xvand_v(__lasx_xvsll_w(idx.vec[0], idx_shift), idx_mask);
+            idx.vec[1] = __lasx_xvand_v(__lasx_xvsll_w(idx.vec[1], idx_shift), idx_mask);
+            idx.vec[0] = __lasx_xvor_v(idx.vec[0], lasx_ext16_32(lasx_extracti128(idx_l, 0)));
+            idx.vec[1] = __lasx_xvor_v(idx.vec[1], lasx_ext16_32(lasx_extracti128(idx_l, 1)));
+
+            // At leat on my CPU (Ryzen 7950X), using _mm256_i32gather_epi32 is slower than _mm256_set_epi32. Strange.
+            //const __m256i q2_1 = _mm256_i32gather_epi32((const int *)iq3s_grid, idx.vec[0], 4);
+            //const __m256i q2_2 = _mm256_i32gather_epi32((const int *)iq3s_grid, idx.vec[1], 4);
+            const __m256i q2_1 = lasx_set_w(
+                    iq3s_grid[idx.index[7]], iq3s_grid[idx.index[6]], iq3s_grid[idx.index[5]], iq3s_grid[idx.index[4]],
+                    iq3s_grid[idx.index[3]], iq3s_grid[idx.index[2]], iq3s_grid[idx.index[1]], iq3s_grid[idx.index[0]]
+            );
+            const __m256i q2_2 = lasx_set_w(
+                    iq3s_grid[idx.index[15]], iq3s_grid[idx.index[14]], iq3s_grid[idx.index[13]], iq3s_grid[idx.index[12]],
+                    iq3s_grid[idx.index[11]], iq3s_grid[idx.index[10]], iq3s_grid[idx.index[ 9]], iq3s_grid[idx.index[ 8]]
+            );
+
+            __m256i aux256 = __lasx_xvreplgr2vr_w(signs[0] | (signs[1] << 16));
+            aux256 = __lasx_xvand_v(lasx_shuffle_b(aux256,mask1), mask2);
+            const __m256i s2_1 = __lasx_xvseq_b(aux256, mask2);
+            const __m256i q8s_1 = __lasx_xvsub_b(__lasx_xvxor_v(s2_1, q8_1), s2_1);
+
+            aux256 = __lasx_xvreplgr2vr_w(signs[2] | (signs[3] << 16));
+            aux256 = __lasx_xvand_v(lasx_shuffle_b(aux256,mask1), mask2);
+            const __m256i s2_2 = __lasx_xvseq_b(aux256, mask2);
+            const __m256i q8s_2 = __lasx_xvsub_b(__lasx_xvxor_v(s2_2, q8_2), s2_2);
+
+            signs += 4;
+
+            const __m256i dot1 = lasx_maddubs_h(q2_1, q8s_1);
+            const __m256i dot2  = lasx_maddubs_h(q2_2, q8s_2);
+            const uint16_t ls1 = x[i].scales[ib32/2] & 0xf;
+            const uint16_t ls2 = x[i].scales[ib32/2] >>  4;
+            const __m256i p1 = lasx_madd_h(dot1, __lasx_xvreplgr2vr_h(2*ls1+1));
+            const __m256i p2 = lasx_madd_h(dot2, __lasx_xvreplgr2vr_h(2*ls2+1));
+            sumi1 = __lasx_xvadd_w(sumi1, p1);
+            sumi2 = __lasx_xvadd_w(sumi2, p2);
+        }
+
+        accumf = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(d), __lasx_xvffint_s_w(__lasx_xvadd_w(sumi1, sumi2)), accumf);
+    }
+
+    *s = hsum_float_8(accumf);
+
+#else
+
+    float sumf = 0.f;
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
+        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;
+            const uint32_t ls2 = 2*(x[i].scales[ib32/2] >>  4) + 1;
+            int32_t sumi = 0;
+            for (int l = 0; l < 4; ++l) {
+                const uint8_t * grid1 = (const uint8_t *)(iq3s_grid + (qs[2*l+0] | ((qh[ib32+0] << (8-2*l)) & 256)));
+                const uint8_t * grid2 = (const uint8_t *)(iq3s_grid + (qs[2*l+1] | ((qh[ib32+0] << (7-2*l)) & 256)));
+                for (int j = 0; j < 4; ++j) {
+                    sumi += grid1[j] * q8[j+0] * (signs[l] & kmask_iq2xs[j+0] ? -1 : 1);
+                    sumi += grid2[j] * q8[j+4] * (signs[l] & kmask_iq2xs[j+4] ? -1 : 1);
+                }
+                q8 += 8;
+            }
+            qs += 8;
+            signs += 4;
+            bsum += sumi * ls1;
+            sumi = 0;
+            for (int l = 0; l < 4; ++l) {
+                const uint8_t * grid1 = (const uint8_t *)(iq3s_grid + (qs[2*l+0] | ((qh[ib32+1] << (8-2*l)) & 256)));
+                const uint8_t * grid2 = (const uint8_t *)(iq3s_grid + (qs[2*l+1] | ((qh[ib32+1] << (7-2*l)) & 256)));
+                for (int j = 0; j < 4; ++j) {
+                    sumi += grid1[j] * q8[j+0] * (signs[l] & kmask_iq2xs[j+0] ? -1 : 1);
+                    sumi += grid2[j] * q8[j+4] * (signs[l] & kmask_iq2xs[j+4] ? -1 : 1);
+                }
+                q8 += 8;
+            }
+            qs += 8;
+            signs += 4;
+            bsum += sumi * ls2;
+        }
+        sumf += d * bsum;
+    }
+    *s = sumf;
+#endif
+}
+
+#if defined(__loongarch_asx)
+static inline __m256i mul_add_epi8(const __m256i x, const __m256i y) {
+    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 * 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);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_iq1_s * GGML_RESTRICT x = vx;
+    const block_q8_K  * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+#if defined(__loongarch_asx)
+
+    __m256 accum = (__m256)__lasx_xvldi(0);
+    float accum1 = 0;
+    for (int i = 0; i < nb; ++i) {
+
+        const int8_t   * q8 = y[i].qs;
+        const uint8_t  * qs = x[i].qs;
+        const uint16_t * qh = x[i].qh;
+
+        __m256i sumi = __lasx_xvldi(0);
+        int sumi1 = 0;
+        for (int ib = 0; ib < QK_K/32; ib += 2) {
+            __m256i q1b_1 = __lasx_xvinsgr2vr_d(q1b_1, iq1s_grid[qs[0] | ((qh[ib+0] << 8) & 0x700)], 0);
+            q1b_1 = __lasx_xvinsgr2vr_d(q1b_1, iq1s_grid[qs[1] | ((qh[ib+0] << 5) & 0x700)], 1);
+            q1b_1 = __lasx_xvinsgr2vr_d(q1b_1, iq1s_grid[qs[2] | ((qh[ib+0] << 2) & 0x700)], 2);
+            q1b_1 = __lasx_xvinsgr2vr_d(q1b_1, iq1s_grid[qs[3] | ((qh[ib+0] >> 1) & 0x700)], 3);
+
+            __m256i q1b_2 = __lasx_xvinsgr2vr_d(q1b_2, iq1s_grid[qs[4] | ((qh[ib+1] << 8) & 0x700)], 0);
+            q1b_2 = __lasx_xvinsgr2vr_d(q1b_2, iq1s_grid[qs[5] | ((qh[ib+1] << 5) & 0x700)], 1);
+            q1b_2 = __lasx_xvinsgr2vr_d(q1b_2, iq1s_grid[qs[6] | ((qh[ib+1] << 2) & 0x700)], 2);
+            q1b_2 = __lasx_xvinsgr2vr_d(q1b_2, iq1s_grid[qs[7] | ((qh[ib+1] >> 1) & 0x700)], 3);
+
+            qs += 8;
+            const __m256i q8b_1 = __lasx_xvld((const __m256i*)q8, 0); q8 += 32;
+            const __m256i q8b_2 = __lasx_xvld((const __m256i*)q8, 0); q8 += 32;
+
+            const __m256i dot1 = mul_add_epi8(q1b_1, q8b_1);
+            const __m256i dot2 = mul_add_epi8(q1b_2, q8b_2);
+            const int16_t ls1 = 2*((qh[ib+0] >> 12) & 7) + 1;
+            const int16_t ls2 = 2*((qh[ib+1] >> 12) & 7) + 1;
+
+            __m256i tmp1, tmp5, tmp6;
+            tmp1 = __lasx_xvreplgr2vr_h(ls1);
+            tmp5 = __lasx_xvmulwev_w_h(dot1, tmp1);
+            tmp6 = __lasx_xvmulwod_w_h(dot1, tmp1);
+            const __m256i p1 = __lasx_xvadd_w(tmp5, tmp6);
+
+            tmp1 = __lasx_xvreplgr2vr_h(ls2);
+            tmp5 = __lasx_xvmulwev_w_h(dot2, tmp1);
+            tmp6 = __lasx_xvmulwod_w_h(dot2, tmp1);
+            const __m256i p2 = __lasx_xvadd_w(tmp5, tmp6);
+
+            sumi = __lasx_xvadd_w(sumi, __lasx_xvadd_w(p1, p2));
+            sumi1 += (y[i].bsums[2*ib+0] + y[i].bsums[2*ib+1]) * (qh[ib+0] & 0x8000 ? -1 : 1) * ls1
+                   + (y[i].bsums[2*ib+2] + y[i].bsums[2*ib+3]) * (qh[ib+1] & 0x8000 ? -1 : 1) * ls2;
+        }
+
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+        accum = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(d), __lasx_xvffint_s_w(sumi), accum);
+        accum1 += d * sumi1;
+    }
+
+    *s = hsum_float_8(accum) + IQ1S_DELTA * accum1;
+
+#else
+
+    float sumf = 0;
+    for (int i = 0; i < nb; i++) {
+
+        const int8_t   * q8 = y[i].qs;
+        const uint8_t  * qs = x[i].qs;
+        const uint16_t * qh = x[i].qh;
+
+        int sumi = 0, sumi1 = 0;
+        for (int ib = 0; ib < QK_K/32; ++ib) {
+            const int ls = 2*((qh[ib] >> 12) & 7) + 1;
+            const int delta = qh[ib] & 0x8000 ? -1 : 1;
+            int lsum = 0;
+            for (int l = 0; l < 4; ++l) {
+                const int8_t * grid = (const int8_t *)(iq1s_grid + (qs[l] | (((qh[ib] >> 3*l) & 7) << 8)));
+                for (int j = 0; j < 8; ++j) {
+                    lsum += q8[j] * grid[j];
+                }
+                q8 += 8;
+            }
+            sumi  += ls * lsum;
+            sumi1 += ls * delta * (y[i].bsums[2*ib+0] + y[i].bsums[2*ib+1]);
+            qs += 4;
+        }
+
+        sumf += GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d * (sumi + IQ1S_DELTA * sumi1);
+    }
+
+    *s = sumf;
+
+#endif
+}
+
+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);
+    UNUSED(by);
+    UNUSED(bs);
+    assert(n % QK4_NL == 0);
+    static_assert(QK4_NL == QK8_0, "QK4_NL and QK8_0 must be the same");
+
+    const block_iq4_nl * GGML_RESTRICT x = vx;
+    const block_q8_0   * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK4_NL;
+
+    int ib = 0;
+    float sumf = 0;
+
+#if defined (__loongarch_asx)
+
+    const __m128i values128 = __lsx_vld((const __m128i*)kvalues_iq4nl, 0);
+    const __m128i m4b  = __lsx_vreplgr2vr_b(0x0f);
+    const __m256i mone = __lasx_xvreplgr2vr_h(1);
+
+    __m256 accum1 = (__m256)__lasx_xvldi(0);
+    __m256 accum2 = (__m256)__lasx_xvldi(0);
+    for (; ib + 1 < nb; ib += 2) {
+        const __m128i q4bits_1 = __lsx_vld((const __m128i*)x[ib + 0].qs, 0);
+        const __m128i q4bits_2 = __lsx_vld((const __m128i*)x[ib + 1].qs, 0);
+        const __m256i q8b_1 = __lasx_xvld((const __m256i *)y[ib + 0].qs, 0);
+        const __m256i q8b_2 = __lasx_xvld((const __m256i *)y[ib + 1].qs, 0);
+        const __m256i q4b_1 = lasx_insertf128(lsx_shuffle_b(values128, __lsx_vand_v(__lsx_vsrli_h(q4bits_1, 4), m4b)),
+                                              lsx_shuffle_b(values128, __lsx_vand_v(q4bits_1, m4b)));
+        const __m256i q4b_2 = lasx_insertf128(lsx_shuffle_b(values128, __lsx_vand_v(__lsx_vsrli_h(q4bits_2, 4), m4b)),
+                                              lsx_shuffle_b(values128, __lsx_vand_v(q4bits_2, m4b)));
+        const __m256i p16_1 = mul_add_epi8(q4b_1, q8b_1);
+        const __m256i p16_2 = mul_add_epi8(q4b_2, q8b_2);
+        const __m256i p_1 = lasx_madd_h(p16_1, mone);
+        const __m256i p_2 = lasx_madd_h(p16_2, mone);
+        accum1 = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(GGML_CPU_FP16_TO_FP32(y[ib + 0].d)*GGML_CPU_FP16_TO_FP32(x[ib + 0].d)),
+                __lasx_xvffint_s_w(p_1), accum1);
+        accum2 = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(GGML_CPU_FP16_TO_FP32(y[ib + 1].d)*GGML_CPU_FP16_TO_FP32(x[ib + 1].d)),
+                __lasx_xvffint_s_w(p_2), accum2);
+    }
+
+    sumf = hsum_float_8(__lasx_xvfadd_s(accum1, accum2));
+
+#endif
+    for (; ib < nb; ++ib) {
+        const float d = GGML_CPU_FP16_TO_FP32(y[ib].d)*GGML_CPU_FP16_TO_FP32(x[ib].d);
+        int sumi1 = 0, sumi2 = 0;
+        for (int j = 0; j < QK4_NL/2; ++j) {
+            sumi1 += y[ib].qs[j+       0] * kvalues_iq4nl[x[ib].qs[j] & 0xf];
+            sumi2 += y[ib].qs[j+QK4_NL/2] * kvalues_iq4nl[x[ib].qs[j] >>  4];
+        }
+        sumf += d * (sumi1 + sumi2);
+    }
+    *s = sumf;
+}
+
+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);
+    UNUSED(by);
+    UNUSED(bs);
+    assert(n % QK_K == 0);
+
+    const block_iq4_xs * GGML_RESTRICT x = vx;
+    const block_q8_K   * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+#if defined(__loongarch_asx)
+
+    const __m128i values128 = __lsx_vld((const __m128i*)kvalues_iq4nl, 0);
+
+    __m256 accum = (__m256)__lasx_xvldi(0);
+
+    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);
+        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 __m256i q8b_1 = __lasx_xvld((const __m256i *)q8, 0); q8 += 32;
+            const __m256i q8b_2 = __lasx_xvld((const __m256i *)q8, 0); q8 += 32;
+            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;
+            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);
+        }
+        accum = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(GGML_CPU_FP16_TO_FP32(x[ibl].d)*y[ibl].d),
+                __lasx_xvffint_s_w(__lasx_xvadd_w(sumi1, sumi2)), accum);
+    }
+
+    *s = hsum_float_8(accum);
+
+#else
+    float sumf = 0;
+    for (int ibl = 0; ibl < nb; ++ibl) {
+        const float d4d8 = GGML_CPU_FP16_TO_FP32(x[ibl].d) * y[ibl].d;
+        uint16_t h = x[ibl].scales_h;
+        const uint8_t * qs = x[ibl].qs;
+        const int8_t  * q8 = y[ibl].qs;
+        for (int ib = 0; ib < QK_K/32; ib += 2) {
+            const uint8_t ls1 = (x[ibl].scales_l[ib/2] & 0xf) | ((h << 4) & 0x30);
+            const uint8_t ls2 = (x[ibl].scales_l[ib/2] >>  4) | ((h << 2) & 0x30);
+            h >>= 4;
+            const float d1 = d4d8*(ls1 - 32);
+            const float d2 = d4d8*(ls2 - 32);
+            int sumi1 = 0, sumi2 = 0;
+            for (int j = 0; j < 16; ++j) {
+                sumi1 += q8[j+ 0] * kvalues_iq4nl[qs[j] & 0xf];
+                sumi2 += q8[j+16] * kvalues_iq4nl[qs[j] >>  4];
+            }
+            sumf += d1 * (sumi1 + sumi2);
+            qs += 16;
+            q8 += 32;
+            sumi1 = sumi2 = 0;
+            for (int j = 0; j < 16; ++j) {
+                sumi1 += q8[j+ 0] * kvalues_iq4nl[qs[j] & 0xf];
+                sumi2 += q8[j+16] * kvalues_iq4nl[qs[j] >>  4];
+            }
+            sumf += d2 * (sumi1 + sumi2);
+            qs += 16;
+            q8 += 32;
+        }
+    }
+    *s = sumf;
+#endif
+}
+
diff --git a/ggml/src/ggml-cpu/arch/powerpc/cpu-feats.cpp b/ggml/src/ggml-cpu/arch/powerpc/cpu-feats.cpp
new file mode 100644
index 000000000..fedd64302
--- /dev/null
+++ b/ggml/src/ggml-cpu/arch/powerpc/cpu-feats.cpp
@@ -0,0 +1,82 @@
+# include "ggml-backend-impl.h"
+
+#if defined(__powerpc64__) || defined(__ppc64__) || defined(__PPC64__)
+
+#if defined(__linux__)
+#include <sys/auxv.h>
+#endif
+
+#include <string>
+
+struct powerpc_features {
+    std::string platform = "";
+    int power_version    = -1;
+
+    bool has_vsx         = false;
+
+    powerpc_features() {
+#if defined(__linux__)
+        unsigned long auxval = getauxval(AT_PLATFORM);
+        if (auxval) {
+            platform = std::string(reinterpret_cast<const char*>(auxval));
+            // TBD: Do systems exist that return this in uppercase?
+            if (platform.substr(0, 5) == "power") {
+                // Extractt a numeric suffix, if one exists
+                int vpos = -1;
+                for (int i = platform.length() - 1; i >= 0; i--) {
+                    if (std::isdigit(platform[i])) {
+                        vpos = i;
+                    } else {
+                        break;
+                    }
+                }
+                if (vpos > -1) {
+                    power_version = std::stoi(platform.substr(vpos));
+                }
+            }
+        }
+#endif
+        if (power_version >= 9) {
+            has_vsx = true;
+        }
+    }
+};
+
+static int ggml_backend_cpu_powerpc_score() {
+    int score = 1;
+    powerpc_features pf;
+
+// Platform scores
+#if defined(GGML_USE_POWER7)
+    if (pf.power_version < 7) { return 0; }
+    score += 1<<1;
+#endif
+#if defined(GGML_USE_POWER8)
+    if (pf.power_version < 8) { return 0; }
+    score += 1<<2;
+#endif
+#if defined(GGML_USE_POWER9)
+    if (pf.power_version < 9) { return 0; }
+    score += 1<<3;
+#endif
+#if defined(GGML_USE_POWER10)
+    if (pf.power_version < 10) { return 0; }
+    score += 1<<4;
+#endif
+#if defined(GGML_USE_POWER11)
+    if (pf.power_version < 11) { return 0; }
+    score += 1<<5;
+#endif
+
+// Feature scores
+#if defined(GGML_USE_VSX)
+    if (!pf.has_vsx) { return 0; }
+    score += 1<<6;
+#endif
+
+    return score;
+}
+
+GGML_BACKEND_DL_SCORE_IMPL(ggml_backend_cpu_powerpc_score)
+
+#endif // defined(__powerpc64__) || defined(__ppc64__) || defined(__PPC64__)
diff --git a/ggml/src/ggml-cpu/arch/powerpc/quants.c b/ggml/src/ggml-cpu/arch/powerpc/quants.c
new file mode 100644
index 000000000..053d5cbdc
--- /dev/null
+++ b/ggml/src/ggml-cpu/arch/powerpc/quants.c
@@ -0,0 +1,2732 @@
+#define GGML_COMMON_IMPL_C
+#include "ggml-common.h"
+#include "ggml-quants.h"
+#include "ggml-impl.h"
+#include "ggml-cpu.h"
+#include "simd-mappings.h"
+
+#include "../../quants.h"
+#include "../../ggml-cpu-impl.h"
+
+#include <math.h>
+#include <string.h>
+#include <assert.h>
+#include <float.h>
+#include <stdlib.h> // for qsort
+#include <stdio.h>  // for GGML_ASSERT
+
+#define GROUP_MAX_EPS 1e-15f
+#define GROUP_MAX_EPS_IQ3_XXS 1e-8f
+#define GROUP_MAX_EPS_IQ2_S 1e-8f
+#define GROUP_MAX_EPS_IQ1_M 1e-7f
+#define GROUP_MAX_EPS_IQ1_S 1e-12f
+
+#define UNUSED GGML_UNUSED
+
+#if defined(__POWER9_VECTOR__)
+#define B1(c,s,n)  0x ## n ## c ,  0x ## n ## s
+#define B2(c,s,n) B1(c,s,n ## c), B1(c,s,n ## s)
+#define B3(c,s,n) B2(c,s,n ## c), B2(c,s,n ## s)
+#define B4(c,s,n) B3(c,s,n ## c), B3(c,s,n ## s)
+#define B5(c,s,n) B4(c,s,n ## c), B4(c,s,n ## s)
+#define B6(c,s,n) B5(c,s,n ## c), B5(c,s,n ## s)
+#define B7(c,s,n) B6(c,s,n ## c), B6(c,s,n ## s)
+#define B8(c,s  ) B7(c,s,     c), B7(c,s,     s)
+
+// precomputed tables for expanding 8bits to 8 bytes:
+static const uint64_t table_b2b_0[1 << 8] = { B8(00, 10) }; // ( b) << 4
+static const uint64_t table_b2b_1[1 << 8] = { B8(10, 00) }; // (!b) << 4
+#endif
+
+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 * GGML_RESTRICT y = vy;
+
+#if defined(__POWER9_VECTOR__)
+    for (int i = 0; i < nb; i++) {
+        vector float srcv [8];
+        vector float asrcv[8];
+        vector float amaxv[8];
+        vector signed int vi[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;
+        const vector float vid = vec_splats(id);
+
+        y[i].d = GGML_CPU_FP32_TO_FP16(d);
+
+        for (int j = 0; j < 8; j++) {
+            const vector float v  = vec_round(vec_mul(srcv[j], vid));
+            vi[j] = vec_cts(v, 0);
+        }
+        vec_xst(vec_pack(vec_pack(vi[0], vi[1]), vec_pack(vi[2], vi[3])),  0, &y[i].qs[0]);
+        vec_xst(vec_pack(vec_pack(vi[4], vi[5]), vec_pack(vi[6], vi[7])), 16, &y[i].qs[0]);
+    }
+#else
+    GGML_UNUSED(nb);
+    // scalar
+    quantize_row_q8_0_ref(x, y, k);
+#endif
+}
+
+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 * GGML_RESTRICT y = vy;
+
+#if defined(__POWER9_VECTOR__)
+    for (int i = 0; i < nb; i++) {
+        vector float srcv [8];
+        vector float asrcv[8];
+        vector float amaxv[8];
+        vector signed int vi[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;
+        const vector float vid = vec_splats(id);
+
+        y[i].d = GGML_CPU_FP32_TO_FP16(d);
+
+        vector int accv = vec_splats(0);
+
+        for (int j = 0; j < 8; j++) {
+            const vector float v  = vec_round(vec_mul(srcv[j], vid));
+            vi[j] = vec_cts(v, 0);
+
+            accv = vec_add(accv, vi[j]);
+        }
+        vec_xst(vec_pack(vec_pack(vi[0], vi[1]), vec_pack(vi[2], vi[3])),  0, &y[i].qs[0]);
+        vec_xst(vec_pack(vec_pack(vi[4], vi[5]), vec_pack(vi[6], vi[7])), 16, &y[i].qs[0]);
+
+        accv = vec_add(accv, vec_sld(accv, accv, 4));
+        accv = vec_add(accv, vec_sld(accv, accv, 8));
+        y[i].s = GGML_CPU_FP32_TO_FP16(d * vec_extract(accv, 0));
+    }
+
+#else
+    GGML_UNUSED(nb);
+    // scalar
+    quantize_row_q8_1_ref(x, y, k);
+#endif
+}
+
+
+//===================================== Dot products =================================
+
+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;
+
+    assert(n % qk == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q4_0 * GGML_RESTRICT x = vx;
+    const block_q8_0 * GGML_RESTRICT y = vy;
+
+    int ib = 0;
+    float sumf = 0;
+
+#if defined(__POWER9_VECTOR__)
+    const vector signed char lowMask = vec_splats((signed char)0xF);
+    const vector signed int v0 = vec_splats((int32_t)0);
+    const vector unsigned char v4 = vec_splats((unsigned char)0x4);
+    const vector signed char v8 = vec_splats((signed char)0x8);
+
+    vector float vsumf0 = 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);
+
+        vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[ib].d));
+        vector float vyd = vec_splats(GGML_CPU_FP16_TO_FP32(y[ib].d));
+        vector float vd = vec_mul(vxd, vyd);
+
+        vector signed char qxs = (vector signed char)vec_xl( 0, x[ib].qs);
+        vector signed char q8y0 = vec_xl( 0, y[ib].qs);
+        vector signed char q8y1 = vec_xl(16, y[ib].qs);
+
+        vector signed char q4x0 = vec_and(qxs, lowMask);
+        vector signed char q4x1 = vec_sr(qxs, v4);
+
+        q4x0 = vec_sub(q4x0, v8);
+        q4x1 = vec_sub(q4x1, v8);
+
+        vector signed short qv0 = vec_add(vec_mule(q4x0, q8y0), vec_mulo(q4x0, q8y0));
+        vector signed short qv1 = vec_add(vec_mule(q4x1, q8y1), vec_mulo(q4x1, q8y1));
+
+        vector signed int vsumi0 = v0;
+
+        vsumi0 = vec_sum4s(qv0, vsumi0);
+        vsumi0 = vec_sum4s(qv1, vsumi0);
+
+        vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
+    }
+
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4));
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8));
+
+    sumf = vec_extract(vsumf0, 0);
+
+#endif
+    for (; ib < nb; ++ib) {
+        int sumi0 = 0;
+        int sumi1 = 0;
+
+        for (int j = 0; j < qk/2; ++j) {
+            const int v0 = (x[ib].qs[j] & 0x0F) - 8;
+            const int v1 = (x[ib].qs[j] >>   4) - 8;
+
+            sumi0 += (v0 * y[ib].qs[j]);
+            sumi1 += (v1 * y[ib].qs[j + qk/2]);
+        }
+
+        int sumi = sumi0 + sumi1;
+        sumf += sumi*GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d);
+    }
+
+    *s = sumf;
+}
+
+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;
+
+    assert(n % qk == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q4_1 * GGML_RESTRICT x = vx;
+    const block_q8_1 * GGML_RESTRICT y = vy;
+
+    int ib = 0;
+    float sumf = 0;
+
+#if defined(__POWER9_VECTOR__)
+    const vector signed char lowMask = vec_splats((signed char)0xF);
+    const vector signed int v0 = vec_splats((int32_t)0);
+    const vector unsigned char v4 = vec_splats((unsigned char)0x4);
+
+    vector float vsumf0 = vec_splats(0.0f);
+
+#pragma GCC unroll 4
+    for (; ib < nb; ++ib) {
+        __builtin_prefetch(x[ib].qs, 0, 1);
+        __builtin_prefetch(y[ib].qs, 0, 1);
+
+        vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[ib].d));
+        vector float vyd = vec_splats(GGML_CPU_FP16_TO_FP32(y[ib].d));
+        vector float vd = vec_mul(vxd, vyd);
+
+        vector float vxmin = vec_splats(GGML_CPU_FP16_TO_FP32(x[ib].m));
+        vector float vys = {GGML_CPU_FP16_TO_FP32(y[ib].s), 0.0f, 0.0f, 0.0f};
+        vsumf0 = vec_madd(vxmin, vys, vsumf0);
+
+        vector signed char qxs = (vector signed char)vec_xl( 0, x[ib].qs);
+        vector signed char q8y0 = vec_xl( 0, y[ib].qs);
+        vector signed char q8y1 = vec_xl(16, y[ib].qs);
+
+        vector unsigned char q4x0 = (vector unsigned char)vec_and(qxs, lowMask);
+        vector unsigned char q4x1 = (vector unsigned char)vec_sr(qxs, v4);
+
+        vector signed int vsumi0 = v0;
+
+        vsumi0 = vec_msum(q8y0, q4x0, vsumi0);
+        vsumi0 = vec_msum(q8y1, q4x1, vsumi0);
+
+        vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
+    }
+
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4));
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8));
+
+    sumf = vec_extract(vsumf0, 0);
+
+#endif
+    for (; ib < nb; ++ib) {
+        int sumi0 = 0;
+        int sumi1 = 0;
+
+        for (int j = 0; j < qk/2; ++j) {
+            const int v0 = (x[ib].qs[j] & 0x0F);
+            const int v1 = (x[ib].qs[j] >>   4);
+
+            sumi0 += (v0 * y[ib].qs[j]);
+            sumi1 += (v1 * y[ib].qs[j + qk/2]);
+        }
+
+        int sumi = sumi0 + sumi1;
+        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
+    }
+
+    *s = sumf;
+}
+
+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;
+
+    int ib = 0;
+    float sumf = 0;
+
+    assert(n % qk == 0);
+    assert(qk == QK5_0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q5_0 * GGML_RESTRICT x = vx;
+    const block_q8_0 * GGML_RESTRICT y = vy;
+
+#if defined(__POWER9_VECTOR__)
+    const vector signed char lowMask = vec_splats((signed char)0xF);
+    const vector unsigned char v4 = vec_splats((unsigned char)4);
+
+    vector float vsumf0 = vec_splats(0.0f);
+
+#pragma GCC unroll 4
+    for (; ib < nb; ++ib) {
+        __builtin_prefetch(x[ib].qs, 0, 1);
+        __builtin_prefetch(y[ib].qs, 0, 1);
+
+        vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[ib].d));
+        vector float vyd = vec_splats(GGML_CPU_FP16_TO_FP32(y[ib].d));
+        vector float vd = vec_mul(vxd, vyd);
+
+        vector signed long long aux64x2_0 = {(uint64_t)(table_b2b_1[x[ib].qh[0]]), (uint64_t)(table_b2b_1[x[ib].qh[1]])};
+        vector signed long long aux64x2_1 = {(uint64_t)(table_b2b_1[x[ib].qh[2]]), (uint64_t)(table_b2b_1[x[ib].qh[3]])};
+
+        vector signed char qh0 = (vector signed char)aux64x2_0;
+        vector signed char qh1 = (vector signed char)aux64x2_1;
+
+        vector signed char qxs = (vector signed char)vec_xl( 0, x[ib].qs);
+
+        vector signed char q5x0 = vec_sub(vec_and (qxs, lowMask), qh0);
+        vector signed char q5x1 = vec_sub(vec_sr(qxs, v4), qh1);
+
+        vector signed char q8y0 = vec_xl(  0, y[ib].qs);
+        vector signed char q8y1 = vec_xl( 16, y[ib].qs);
+
+        vector signed short qv0 = vec_add(vec_mule(q5x0, q8y0), vec_mulo(q5x0, q8y0));
+        vector signed short qv1 = vec_add(vec_mule(q5x1, q8y1), vec_mulo(q5x1, q8y1));
+
+        qv0 = vec_add(qv0, qv1);
+
+        vector signed int vsumi0 = vec_add(vec_unpackh(qv0), vec_unpackl(qv0));
+
+        vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
+    }
+
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4));
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8));
+
+    sumf = vec_extract(vsumf0, 0);
+
+#endif
+    for (; ib < nb; ++ib) {
+        uint32_t qh;
+        memcpy(&qh, x[ib].qh, sizeof(qh));
+
+        int sumi0 = 0;
+        int sumi1 = 0;
+
+        for (int j = 0; j < qk/2; ++j) {
+            const uint8_t xh_0 = ((qh & (1u << (j + 0 ))) >> (j + 0 )) << 4;
+            const uint8_t xh_1 = ((qh & (1u << (j + 16))) >> (j + 12));
+
+            const int32_t x0 = (int8_t)(((x[ib].qs[j] & 0x0F) | xh_0) - 16);
+            const int32_t x1 = (int8_t)(((x[ib].qs[j] >>   4) | xh_1) - 16);
+
+            sumi0 += (x0 * y[ib].qs[j]);
+            sumi1 += (x1 * y[ib].qs[j + qk/2]);
+        }
+
+        int sumi = sumi0 + sumi1;
+        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d)) * sumi;
+    }
+
+    *s = sumf;
+}
+
+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;
+
+    int ib = 0;
+    float sumf = 0;
+
+    assert(n % qk == 0);
+    assert(qk == QK5_1);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q5_1 * GGML_RESTRICT x = vx;
+    const block_q8_1 * GGML_RESTRICT y = vy;
+
+#if defined(__POWER9_VECTOR__)
+    const vector signed char lowMask = vec_splats((signed char)0xF);
+    const vector signed int v0 = vec_splats((int32_t)0);
+    const vector unsigned char v4 = vec_splats((unsigned char)0x4);
+
+    vector float vsumf0 = vec_splats(0.0f);
+
+#pragma GCC unroll 4
+    for (; ib < nb; ++ib) {
+        __builtin_prefetch(x[ib].qs, 0, 1);
+        __builtin_prefetch(y[ib].qs, 0, 1);
+
+        vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[ib].d));
+        vector float vyd = vec_splats(GGML_CPU_FP16_TO_FP32(y[ib].d));
+        vector float vd = vec_mul(vxd, vyd);
+
+        vector float vxmin = vec_splats(GGML_CPU_FP16_TO_FP32(x[ib].m));
+        vector float vys = {GGML_CPU_FP16_TO_FP32(y[ib].s), 0.f, 0.f, 0.f};
+        vsumf0 = vec_madd(vxmin, vys, vsumf0);
+
+        vector unsigned long long aux64x2_0 = {(uint64_t)(table_b2b_0[x[ib].qh[0]]), (uint64_t)(table_b2b_0[x[ib].qh[1]])};
+        vector unsigned long long aux64x2_1 = {(uint64_t)(table_b2b_0[x[ib].qh[2]]), (uint64_t)(table_b2b_0[x[ib].qh[3]])};
+
+        vector signed char qh0 = (vector signed char)aux64x2_0;
+        vector signed char qh1 = (vector signed char)aux64x2_1;
+
+        vector signed char qxs = (vector signed char)vec_xl( 0, x[ib].qs);
+
+        vector unsigned char q5x0 = (vector unsigned char)vec_or(vec_and(qxs, lowMask), qh0);
+        vector unsigned char q5x1 = (vector unsigned char)vec_or(vec_sr(qxs, v4), qh1);
+
+        vector signed char q8y0 = vec_xl(  0, y[ib].qs);
+        vector signed char q8y1 = vec_xl( 16, y[ib].qs);
+
+        vector signed int vsumi0 = v0;
+
+        vsumi0 = vec_msum(q8y0, q5x0, vsumi0);
+        vsumi0 = vec_msum(q8y1, q5x1, vsumi0);
+
+        vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
+    }
+
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4));
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8));
+
+    sumf = vec_extract(vsumf0, 0);
+
+#endif
+    for (; ib < nb; ++ib) {
+        uint32_t qh;
+        memcpy(&qh, x[ib].qh, sizeof(qh));
+
+        int sumi0 = 0;
+        int sumi1 = 0;
+
+        for (int j = 0; j < qk/2; ++j) {
+            const uint8_t xh_0 = ((qh >> (j +  0)) << 4) & 0x10;
+            const uint8_t xh_1 = ((qh >> (j + 12))     ) & 0x10;
+
+            const int32_t x0 = (x[ib].qs[j] & 0xF) | xh_0;
+            const int32_t x1 = (x[ib].qs[j] >>  4) | xh_1;
+
+            sumi0 += (x0 * y[ib].qs[j]);
+            sumi1 += (x1 * y[ib].qs[j + qk/2]);
+        }
+
+        int sumi = sumi0 + sumi1;
+        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
+    }
+
+    *s = sumf;
+}
+
+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;
+
+    assert(n % qk == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q8_0 * GGML_RESTRICT x = vx;
+    const block_q8_0 * GGML_RESTRICT y = vy;
+
+    int ib = 0;
+    float sumf = 0;
+
+#if defined(__POWER9_VECTOR__)
+    const vector signed int v0 = vec_splats((int32_t)0);
+    vector float vsumf0 = 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);
+
+        vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[ib].d));
+        vector float vyd = vec_splats(GGML_CPU_FP16_TO_FP32(y[ib].d));
+        vector float vd = vec_mul(vxd, vyd);
+
+        vector signed char q8x0 = vec_xl( 0, x[ib].qs);
+        vector signed char q8x1 = vec_xl(16, x[ib].qs);
+        vector signed char q8y0 = vec_xl( 0, y[ib].qs);
+        vector signed char q8y1 = vec_xl(16, y[ib].qs);
+
+        vector signed short qv0 = vec_mule(q8x0, q8y0);
+        vector signed short qv1 = vec_mulo(q8x0, q8y0);
+        vector signed short qv2 = vec_mule(q8x1, q8y1);
+        vector signed short qv3 = vec_mulo(q8x1, q8y1);
+
+        vector signed int vsumi0 = v0;
+        vector signed int vsumi1 = v0;
+
+        vsumi0 = vec_sum4s(qv0, vsumi0);
+        vsumi1 = vec_sum4s(qv1, vsumi1);
+        vsumi0 = vec_sum4s(qv2, vsumi0);
+        vsumi1 = vec_sum4s(qv3, vsumi1);
+
+        vsumi0 = vec_add(vsumi0, vsumi1);
+
+        vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
+    }
+
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4));
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8));
+
+    sumf = vec_extract(vsumf0, 0);
+
+#endif
+    for (; ib < nb; ++ib) {
+        int sumi = 0;
+
+        for (int j = 0; j < qk; j++) {
+            sumi += x[ib].qs[j]*y[ib].qs[j];
+        }
+
+        sumf += sumi*(GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d));
+    }
+
+    *s = sumf;
+}
+
+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 * GGML_RESTRICT x = vx;
+    const block_q8_K * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+#if defined(__POWER9_VECTOR__)
+    const vector signed char lowMask = vec_splats((signed char)0x3);
+    const vector signed char lowScaleMask = vec_splats((signed char)0xF);
+    const vector int v0 = vec_splats((int32_t)0);
+    const vector unsigned char v2 = vec_splats((unsigned char)0x2);
+    const vector unsigned char v6 = vec_splats((unsigned char)0x6);
+    const vector unsigned char v4 = vec_splats((unsigned char)0x4);
+
+    vector float vsumf0 = vec_splats(0.0f);
+    vector float vsumf1 = vec_splats(0.0f);
+    vector float vsumf2 = vec_splats(0.0f);
+    vector float vsumf3 = vec_splats(0.0f);
+
+    for (int i = 0; i < nb; ++i) {
+        vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[i].d));
+        vector float vyd = vec_splats(y[i].d);
+        vector float vd = vec_mul(vxd, vyd);
+
+        vector float vxmin = vec_splats(GGML_CPU_FP16_TO_FP32(x[i].dmin));
+        vector float vdmin = vec_mul(vxmin, vyd);
+
+        vector signed short q8ysums0 = vec_xl( 0, y[i].bsums);
+        vector signed short q8ysums1 = vec_xl(16, y[i].bsums);
+
+        vector signed char q2xmins = (vector signed char)vec_xl( 0, x[i].scales);
+        vector signed char vscales = vec_and(q2xmins, lowScaleMask);
+
+        q2xmins = vec_sr(q2xmins, v4);
+        vector signed short q2xmins0 = vec_unpackh(q2xmins);
+        vector signed short q2xmins1 = vec_unpackl(q2xmins);
+
+        vector signed int prod0 = vec_mule(q2xmins0, q8ysums0);
+        vector signed int prod1 = vec_mulo(q2xmins0, q8ysums0);
+        vector signed int prod2 = vec_mule(q2xmins1, q8ysums1);
+        vector signed int prod3 = vec_mulo(q2xmins1, q8ysums1);
+
+        vsumf0 = vec_nmsub(vec_ctf(prod0, 0), vdmin, vsumf0);
+        vsumf1 = vec_nmsub(vec_ctf(prod1, 0), vdmin, vsumf1);
+        vsumf2 = vec_nmsub(vec_ctf(prod2, 0), vdmin, vsumf2);
+        vsumf3 = vec_nmsub(vec_ctf(prod3, 0), vdmin, vsumf3);
+
+        vector signed int vsumi0 = v0;
+        vector signed int vsumi1 = v0;
+        vector signed int vsumi2 = v0;
+        vector signed int vsumi3 = v0;
+        vector signed int vsumi4 = v0;
+        vector signed int vsumi5 = v0;
+        vector signed int vsumi6 = v0;
+        vector signed int vsumi7 = v0;
+
+        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);
+            __builtin_prefetch(q8, 0, 1);
+
+            vector signed char qxs0 = (vector signed char)vec_xl( 0, q2);
+            vector signed char qxs1 = (vector signed char)vec_xl(16, q2);
+            q2 += 32;
+
+            vector unsigned char q2x00 = (vector unsigned char)vec_and(qxs0, lowMask);
+            vector unsigned char q2x01 = (vector unsigned char)vec_and(vec_sr(qxs0, v2), lowMask);
+            vector unsigned char q2x02 = (vector unsigned char)vec_and(vec_sr(qxs0, v4), lowMask);
+            vector unsigned char q2x03 = (vector unsigned char)vec_and(vec_sr(qxs0, v6), lowMask);
+            vector unsigned char q2x10 = (vector unsigned char)vec_and(qxs1, lowMask);
+            vector unsigned char q2x11 = (vector unsigned char)vec_and(vec_sr(qxs1, v2), lowMask);
+            vector unsigned char q2x12 = (vector unsigned char)vec_and(vec_sr(qxs1, v4), lowMask);
+            vector unsigned char q2x13 = (vector unsigned char)vec_and(vec_sr(qxs1, v6), lowMask);
+
+            vector signed char q8y00 = vec_xl(  0, q8);
+            vector signed char q8y10 = vec_xl( 16, q8);
+            vector signed char q8y01 = vec_xl( 32, q8);
+            vector signed char q8y11 = vec_xl( 48, q8);
+            vector signed char q8y02 = vec_xl( 64, q8);
+            vector signed char q8y12 = vec_xl( 80, q8);
+            vector signed char q8y03 = vec_xl( 96, q8);
+            vector signed char q8y13 = vec_xl(112, q8);
+            q8 += 128;
+
+            vector signed int qv0 = vec_msum(q8y00, q2x00, v0);
+            vector signed int qv1 = vec_msum(q8y01, q2x01, v0);
+            vector signed int qv2 = vec_msum(q8y02, q2x02, v0);
+            vector signed int qv3 = vec_msum(q8y03, q2x03, v0);
+            vector signed int qv4 = vec_msum(q8y10, q2x10, v0);
+            vector signed int qv5 = vec_msum(q8y11, q2x11, v0);
+            vector signed int qv6 = vec_msum(q8y12, q2x12, v0);
+            vector signed int qv7 = vec_msum(q8y13, q2x13, v0);
+
+            vector signed short vscales_07 = vec_unpackh(vscales);
+            vector signed int vscales_03 = vec_unpackh(vscales_07);
+            vector signed int vscales_47 = vec_unpackl(vscales_07);
+            vector signed int vs0 = vec_splat(vscales_03, 0);
+            vector signed int vs1 = vec_splat(vscales_03, 1);
+            vector signed int vs2 = vec_splat(vscales_03, 2);
+            vector signed int vs3 = vec_splat(vscales_03, 3);
+            vector signed int vs4 = vec_splat(vscales_47, 0);
+            vector signed int vs5 = vec_splat(vscales_47, 1);
+            vector signed int vs6 = vec_splat(vscales_47, 2);
+            vector signed int vs7 = vec_splat(vscales_47, 3);
+            vscales = vec_sld(vscales, vscales, 8);
+
+            vsumi0 = vec_add(vec_mul(qv0, vs0), vsumi0);
+            vsumi1 = vec_add(vec_mul(qv1, vs2), vsumi1);
+            vsumi2 = vec_add(vec_mul(qv2, vs4), vsumi2);
+            vsumi3 = vec_add(vec_mul(qv3, vs6), vsumi3);
+            vsumi4 = vec_add(vec_mul(qv4, vs1), vsumi4);
+            vsumi5 = vec_add(vec_mul(qv5, vs3), vsumi5);
+            vsumi6 = vec_add(vec_mul(qv6, vs5), vsumi6);
+            vsumi7 = vec_add(vec_mul(qv7, vs7), vsumi7);
+        }
+
+        vsumi0 = vec_add(vsumi0, vsumi4);
+        vsumi1 = vec_add(vsumi1, vsumi5);
+        vsumi2 = vec_add(vsumi2, vsumi6);
+        vsumi3 = vec_add(vsumi3, vsumi7);
+
+        vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
+        vsumf1 = vec_madd(vec_ctf(vsumi1, 0), vd, vsumf1);
+        vsumf2 = vec_madd(vec_ctf(vsumi2, 0), vd, vsumf2);
+        vsumf3 = vec_madd(vec_ctf(vsumi3, 0), vd, vsumf3);
+    }
+
+    vsumf0 = vec_add(vsumf0, vsumf2);
+    vsumf1 = vec_add(vsumf1, vsumf3);
+
+    vsumf0 = vec_add(vsumf0, vsumf1);
+
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4));
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8));
+
+    *s = vec_extract(vsumf0, 0);
+
+#else
+
+    float sumf = 0;
+
+    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;
+
+        int summs = 0;
+        for (int j = 0; j < 16; ++j) {
+            summs += y[i].bsums[j] * (sc[j] >> 4);
+        }
+
+        const float dall = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+        const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
+
+        int isum = 0;
+        int is = 0;
+        int d;
+        for (int k = 0; k < QK_K/128; ++k) {
+            int shift = 0;
+            for (int j = 0; j < 4; ++j) {
+                d = sc[is++] & 0xF;
+                int isuml = 0;
+                for (int l =  0; l < 16; ++l) isuml += q8[l] * ((q2[l] >> shift) & 3);
+                isum += d * isuml;
+                d = sc[is++] & 0xF;
+                isuml = 0;
+                for (int l = 16; l < 32; ++l) isuml += q8[l] * ((q2[l] >> shift) & 3);
+                isum += d * isuml;
+                shift += 2;
+                q8 += 32;
+            }
+            q2 += 32;
+        }
+        sumf += dall * isum - dmin * summs;
+    }
+    *s = sumf;
+#endif
+}
+
+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);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const uint32_t kmask1 = 0x03030303;
+    const uint32_t kmask2 = 0x0f0f0f0f;
+
+    const block_q3_K * GGML_RESTRICT x = vx;
+    const block_q8_K * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+#if defined(__POWER9_VECTOR__)
+    const vector signed char lowMask = vec_splats((signed char)0x3);
+    const vector signed char lowMask1 = vec_splats((int8_t)0xf);
+    const vector signed char lowMask2 = vec_splats((int8_t)0x30);
+    const vector int v0 = vec_splats((int32_t)0);
+    const vector signed char v1 = vec_splats((signed char)0x1);
+    const vector unsigned char v2 = vec_splats((unsigned char)0x2);
+    const vector unsigned char v3 = vec_splats((unsigned char)0x3);
+    const vector unsigned char v4 = vec_splats((unsigned char)0x4);
+    const vector unsigned char v6 = vec_splats((unsigned char)0x6);
+    const vector signed char off = vec_splats((signed char)0x20);
+
+    vector float vsumf0 = vec_splats(0.0f);
+    vector float vsumf1 = vec_splats(0.0f);
+    vector float vsumf2 = vec_splats(0.0f);
+    vector float vsumf3 = vec_splats(0.0f);
+
+    for (int i = 0; i < nb; ++i) {
+        vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[i].d));
+        vector float vyd = vec_splats(y[i].d);
+        vector float vd = vec_mul(vxd, vyd);
+
+        UNUSED(kmask1);
+        UNUSED(kmask2);
+
+        vector signed char u0 = (vector signed char)vec_xl_len(x[i].scales, 8);
+        vector signed char u1 = vec_and(u0, lowMask1);
+        vector signed char u2 = (vector signed char)vec_xl_len(x[i].scales + 8, 4);
+        vector signed char u3 = (vector signed char)vec_mergeh((vector signed int)u2, (vector signed int)vec_sr(u2, v2));
+        vector signed char u30 = vec_sl(vec_and(u3, lowMask), v4);
+        vector signed char u31 = vec_and(u3, lowMask2);
+
+        u1 = vec_or(u1, u30);
+        u2 = vec_or(vec_sr(u0, v4), u31);
+
+        vector signed char vscales = (vector signed char)vec_mergeh((vector signed long long)u1, (vector signed long long)u2);
+        vector signed char qxhs0 = (vector signed char)vec_xl( 0, x[i].hmask);
+        vector signed char qxhs1 = (vector signed char)vec_xl(16, x[i].hmask);
+
+        vscales = vec_sub(vscales, off);
+
+        vector signed int vsumi0 = v0;
+        vector signed int vsumi1 = v0;
+        vector signed int vsumi2 = v0;
+        vector signed int vsumi3 = v0;
+        vector signed int vsumi4 = v0;
+        vector signed int vsumi5 = v0;
+        vector signed int vsumi6 = v0;
+        vector signed int vsumi7 = v0;
+
+        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);
+            __builtin_prefetch(q8, 0, 1);
+
+            vector signed char qxs0 = (vector signed char)vec_xl( 0, q3);
+            vector signed char qxs1 = (vector signed char)vec_xl(16, q3);
+            q3 += 32;
+
+            //the low 2 bits
+            vector signed char qxs00 = vec_and(qxs0, lowMask);
+            vector signed char qxs01 = vec_and(vec_sr(qxs0, v2), lowMask);
+            vector signed char qxs02 = vec_and(vec_sr(qxs0, v4), lowMask);
+            vector signed char qxs03 = vec_and(vec_sr(qxs0, v6), lowMask);
+            vector signed char qxs10 = vec_and(qxs1, lowMask);
+            vector signed char qxs11 = vec_and(vec_sr(qxs1, v2), lowMask);
+            vector signed char qxs12 = vec_and(vec_sr(qxs1, v4), lowMask);
+            vector signed char qxs13 = vec_and(vec_sr(qxs1, v6), lowMask);
+
+            //the 3rd bit
+            vector signed char qxh00 = vec_sl(vec_andc(v1, qxhs0), v2);
+            vector signed char qxh01 = vec_sl(vec_andc(v1, vec_sr(qxhs0, (vector unsigned char)v1)), v2);
+            vector signed char qxh02 = vec_sl(vec_andc(v1, vec_sr(qxhs0, v2)), v2);
+            vector signed char qxh03 = vec_sl(vec_andc(v1, vec_sr(qxhs0, v3)), v2);
+            vector signed char qxh10 = vec_sl(vec_andc(v1, qxhs1), v2);
+            vector signed char qxh11 = vec_sl(vec_andc(v1, vec_sr(qxhs1, (vector unsigned char)v1)), v2);
+            vector signed char qxh12 = vec_sl(vec_andc(v1, vec_sr(qxhs1, v2)), v2);
+            vector signed char qxh13 = vec_sl(vec_andc(v1, vec_sr(qxhs1, v3)), v2);
+            qxhs0 = vec_sr(qxhs0, v4);
+            qxhs1 = vec_sr(qxhs1, v4);
+
+            vector signed char q3x00 = vec_sub(qxs00, qxh00);
+            vector signed char q3x01 = vec_sub(qxs01, qxh01);
+            vector signed char q3x02 = vec_sub(qxs02, qxh02);
+            vector signed char q3x03 = vec_sub(qxs03, qxh03);
+            vector signed char q3x10 = vec_sub(qxs10, qxh10);
+            vector signed char q3x11 = vec_sub(qxs11, qxh11);
+            vector signed char q3x12 = vec_sub(qxs12, qxh12);
+            vector signed char q3x13 = vec_sub(qxs13, qxh13);
+
+            vector signed char q8y00 = vec_xl(  0, q8);
+            vector signed char q8y10 = vec_xl( 16, q8);
+            vector signed char q8y01 = vec_xl( 32, q8);
+            vector signed char q8y11 = vec_xl( 48, q8);
+            vector signed char q8y02 = vec_xl( 64, q8);
+            vector signed char q8y12 = vec_xl( 80, q8);
+            vector signed char q8y03 = vec_xl( 96, q8);
+            vector signed char q8y13 = vec_xl(112, q8);
+            q8 += 128;
+
+            vector signed short vscales_h = vec_unpackh(vscales);
+            vector signed short vs0 = vec_splat(vscales_h, 0);
+            vector signed short vs1 = vec_splat(vscales_h, 1);
+            vector signed short vs2 = vec_splat(vscales_h, 2);
+            vector signed short vs3 = vec_splat(vscales_h, 3);
+            vector signed short vs4 = vec_splat(vscales_h, 4);
+            vector signed short vs5 = vec_splat(vscales_h, 5);
+            vector signed short vs6 = vec_splat(vscales_h, 6);
+            vector signed short vs7 = vec_splat(vscales_h, 7);
+            vscales = vec_sld(vscales, vscales, 8);
+
+            vector signed short qv00 = vec_add(vec_mule(q3x00, q8y00), vec_mulo(q3x00, q8y00));
+            vector signed short qv01 = vec_add(vec_mule(q3x01, q8y01), vec_mulo(q3x01, q8y01));
+            vector signed short qv02 = vec_add(vec_mule(q3x02, q8y02), vec_mulo(q3x02, q8y02));
+            vector signed short qv03 = vec_add(vec_mule(q3x03, q8y03), vec_mulo(q3x03, q8y03));
+            vector signed short qv10 = vec_add(vec_mule(q3x10, q8y10), vec_mulo(q3x10, q8y10));
+            vector signed short qv11 = vec_add(vec_mule(q3x11, q8y11), vec_mulo(q3x11, q8y11));
+            vector signed short qv12 = vec_add(vec_mule(q3x12, q8y12), vec_mulo(q3x12, q8y12));
+            vector signed short qv13 = vec_add(vec_mule(q3x13, q8y13), vec_mulo(q3x13, q8y13));
+
+            vsumi0 = vec_msum(qv00, vs0, vsumi0);
+            vsumi1 = vec_msum(qv01, vs2, vsumi1);
+            vsumi2 = vec_msum(qv02, vs4, vsumi2);
+            vsumi3 = vec_msum(qv03, vs6, vsumi3);
+            vsumi4 = vec_msum(qv10, vs1, vsumi4);
+            vsumi5 = vec_msum(qv11, vs3, vsumi5);
+            vsumi6 = vec_msum(qv12, vs5, vsumi6);
+            vsumi7 = vec_msum(qv13, vs7, vsumi7);
+        }
+
+        vsumi0 = vec_add(vsumi0, vsumi4);
+        vsumi1 = vec_add(vsumi1, vsumi5);
+        vsumi2 = vec_add(vsumi2, vsumi6);
+        vsumi3 = vec_add(vsumi3, vsumi7);
+
+        vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
+        vsumf1 = vec_madd(vec_ctf(vsumi1, 0), vd, vsumf1);
+        vsumf2 = vec_madd(vec_ctf(vsumi2, 0), vd, vsumf2);
+        vsumf3 = vec_madd(vec_ctf(vsumi3, 0), vd, vsumf3);
+    }
+
+    vsumf0 = vec_add(vsumf0, vsumf2);
+    vsumf1 = vec_add(vsumf1, vsumf3);
+
+    vsumf0 = vec_add(vsumf0, vsumf1);
+
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4));
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8));
+
+    *s = vec_extract(vsumf0, 0);
+
+#else
+    // scalar version
+    // This function is written like this so the compiler can manage to vectorize most of it
+    // Using -Ofast, GCC and clang manage to produce code that is within a factor of 2 or so from the
+    // manually vectorized version above. Every other version I tried would run at least 4 times slower.
+    // The ideal situation would be if we could just write the code once, and the compiler would
+    // automatically produce the best possible set of machine instructions, instead of us having to manually
+    // write vectorized versions for AVX, ARM_NEON, etc.
+
+    int8_t  aux8[QK_K];
+    int16_t aux16[8];
+    float   sums [8];
+    int32_t aux32[8];
+    memset(sums, 0, 8*sizeof(float));
+
+    uint32_t auxs[4];
+    const int8_t * scales = (const int8_t*)auxs;
+
+    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;
+        memset(aux32, 0, 8*sizeof(int32_t));
+        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;
+            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
+            a += 32; m <<= 1;
+            for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 2) & 3;
+            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
+            a += 32; m <<= 1;
+            for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 4) & 3;
+            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
+            a += 32; m <<= 1;
+            for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 6) & 3;
+            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
+            a += 32; m <<= 1;
+            q3 += 32;
+        }
+        a = aux8;
+
+        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);
+        for (int j = 0; j < QK_K/16; ++j) {
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
+            q8 += 8; a += 8;
+        }
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
+        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
+    }
+    for (int l = 0; l < 8; ++l) sumf += sums[l];
+    *s = sumf;
+
+#endif
+
+}
+
+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);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q4_K * GGML_RESTRICT x = vx;
+    const block_q8_K * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+    static const uint32_t kmask1 = 0x3f3f3f3f;
+    static const uint32_t kmask2 = 0x0f0f0f0f;
+    static const uint32_t kmask3 = 0x03030303;
+
+    uint32_t utmp[4];
+
+#if defined(__POWER9_VECTOR__)
+    const vector signed char lowMask = vec_splats((signed char)0xF);
+    const vector signed char lowMask1 = vec_splats((int8_t)0x3f);
+    const vector signed char lowMask2 = vec_splats((int8_t)0x30);
+    const vector int v0 = vec_splats((int32_t)0);
+    const vector unsigned char v2 = vec_splats((uint8_t)2);
+    const vector unsigned char v4 = vec_splats((unsigned char)0x4);
+
+    vector float vsumf0 = vec_splats(0.0f);
+    vector float vsumf1 = vec_splats(0.0f);
+    vector float vsumf2 = vec_splats(0.0f);
+    vector float vsumf3 = vec_splats(0.0f);
+
+    for (int i = 0; i < nb; ++i) {
+        vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[i].d));
+        vector float vyd = vec_splats(y[i].d);
+        vector float vd = vec_mul(vxd, vyd);
+
+        vector float vxmin = vec_splats(GGML_CPU_FP16_TO_FP32(x[i].dmin));
+        vector float vdmin = vec_mul(vxmin, vyd);
+
+        vector signed short q8ysums0 = vec_xl( 0, y[i].bsums);
+        vector signed short q8ysums1 = vec_xl(16, y[i].bsums);
+
+        UNUSED(kmask1);
+        UNUSED(kmask2);
+        UNUSED(kmask3);
+        UNUSED(utmp);
+
+        vector signed char u0 = (vector signed char)vec_xl_len(x[i].scales, 8);
+        vector signed char u1 = vec_and(vec_sr(u0, v2), lowMask2);
+        vector signed char u2 = (vector signed char)vec_xl_len(x[i].scales + 8, 4);
+        vector signed char u3 = vec_sr(u2, v4);
+
+        vector signed char u30 = u1;
+        vector signed char u31 = (vector signed char)vec_mergeh((vector signed int)vec_and(u2, lowMask), (vector signed int)u3);
+
+        u1 = vec_and(u0, lowMask1);
+        u2 = vec_or(u30, u31);
+
+        vector signed char utmps = (vector signed char)vec_mergeh((vector signed int)u1, (vector signed int)u2);
+
+        vector signed short vscales = vec_unpackh(utmps);
+        vector signed short q4xmins = vec_unpackl(utmps);
+        vector signed short q4xmins0 = vec_mergeh(q4xmins, q4xmins);
+        vector signed short q4xmins1 = vec_mergel(q4xmins, q4xmins);
+
+        vector signed int prod0 = vec_mule(q4xmins0, q8ysums0);
+        vector signed int prod1 = vec_mule(q4xmins1, q8ysums1);
+        vector signed int prod2 = vec_mulo(q4xmins0, q8ysums0);
+        vector signed int prod3 = vec_mulo(q4xmins1, q8ysums1);
+
+        vsumf0 = vec_nmsub(vec_ctf(prod0, 0), vdmin, vsumf0);
+        vsumf1 = vec_nmsub(vec_ctf(prod1, 0), vdmin, vsumf1);
+        vsumf2 = vec_nmsub(vec_ctf(prod2, 0), vdmin, vsumf2);
+        vsumf3 = vec_nmsub(vec_ctf(prod3, 0), vdmin, vsumf3);
+
+        vector signed int vsumi0 = v0;
+        vector signed int vsumi1 = v0;
+        vector signed int vsumi2 = v0;
+        vector signed int vsumi3 = v0;
+
+        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);
+            __builtin_prefetch(q8, 0, 1);
+
+            vector signed char qxs0 = (vector signed char)vec_xl( 0, q4);
+            vector signed char qxs1 = (vector signed char)vec_xl(16, q4);
+            vector signed char qxs2 = (vector signed char)vec_xl(32, q4);
+            vector signed char qxs3 = (vector signed char)vec_xl(48, q4);
+            q4 += 64;
+
+            vector unsigned char q4x00 = (vector unsigned char)vec_and(qxs0, lowMask);
+            vector unsigned char q4x01 = (vector unsigned char)vec_sr(qxs0, v4);
+            vector unsigned char q4x10 = (vector unsigned char)vec_and(qxs1, lowMask);
+            vector unsigned char q4x11 = (vector unsigned char)vec_sr(qxs1, v4);
+            vector unsigned char q4x20 = (vector unsigned char)vec_and(qxs2, lowMask);
+            vector unsigned char q4x21 = (vector unsigned char)vec_sr(qxs2, v4);
+            vector unsigned char q4x30 = (vector unsigned char)vec_and(qxs3, lowMask);
+            vector unsigned char q4x31 = (vector unsigned char)vec_sr(qxs3, v4);
+
+            vector signed char q8y00 = vec_xl(  0, q8);
+            vector signed char q8y10 = vec_xl( 16, q8);
+            vector signed char q8y01 = vec_xl( 32, q8);
+            vector signed char q8y11 = vec_xl( 48, q8);
+            vector signed char q8y20 = vec_xl( 64, q8);
+            vector signed char q8y30 = vec_xl( 80, q8);
+            vector signed char q8y21 = vec_xl( 96, q8);
+            vector signed char q8y31 = vec_xl(112, q8);
+            q8 += 128;
+
+            vector signed int qv00 = vec_msum(q8y00, q4x00, v0);
+            vector signed int qv01 = vec_msum(q8y01, q4x01, v0);
+            vector signed int qv10 = vec_msum(q8y10, q4x10, v0);
+            vector signed int qv11 = vec_msum(q8y11, q4x11, v0);
+            vector signed int qv20 = vec_msum(q8y20, q4x20, v0);
+            vector signed int qv21 = vec_msum(q8y21, q4x21, v0);
+            vector signed int qv30 = vec_msum(q8y30, q4x30, v0);
+            vector signed int qv31 = vec_msum(q8y31, q4x31, v0);
+
+            vector signed int vscales_h = vec_unpackh(vscales);
+            vector signed int vs0 = vec_splat(vscales_h, 0);
+            vector signed int vs1 = vec_splat(vscales_h, 1);
+            vector signed int vs2 = vec_splat(vscales_h, 2);
+            vector signed int vs3 = vec_splat(vscales_h, 3);
+            vscales = vec_sld(vscales, vscales, 8);
+
+            vsumi0 = vec_add(vec_mul(qv00, vs0), vsumi0);
+            vsumi1 = vec_add(vec_mul(qv01, vs1), vsumi1);
+            vsumi2 = vec_add(vec_mul(qv20, vs2), vsumi2);
+            vsumi3 = vec_add(vec_mul(qv21, vs3), vsumi3);
+
+            vsumi0 = vec_add(vec_mul(qv10, vs0), vsumi0);
+            vsumi1 = vec_add(vec_mul(qv11, vs1), vsumi1);
+            vsumi2 = vec_add(vec_mul(qv30, vs2), vsumi2);
+            vsumi3 = vec_add(vec_mul(qv31, vs3), vsumi3);
+        }
+
+        vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
+        vsumf1 = vec_madd(vec_ctf(vsumi1, 0), vd, vsumf1);
+        vsumf2 = vec_madd(vec_ctf(vsumi2, 0), vd, vsumf2);
+        vsumf3 = vec_madd(vec_ctf(vsumi3, 0), vd, vsumf3);
+    }
+
+    vsumf0 = vec_add(vsumf0, vsumf2);
+    vsumf1 = vec_add(vsumf1, vsumf3);
+
+    vsumf0 = vec_add(vsumf0, vsumf1);
+
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4));
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8));
+
+    *s = vec_extract(vsumf0, 0);
+
+#else
+
+    const uint8_t * scales = (const uint8_t*)&utmp[0];
+    const uint8_t * mins   = (const uint8_t*)&utmp[2];
+
+    int8_t  aux8[QK_K];
+    int16_t aux16[8];
+    float   sums [8];
+    int32_t aux32[8];
+    memset(sums, 0, 8*sizeof(float));
+
+    float sumf = 0;
+    for (int i = 0; i < nb; ++i) {
+        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 * 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;
+            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l]  >> 4);
+            a += 32; q4 += 32;
+        }
+        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;
+
+        int sumi = 0;
+        for (int j = 0; j < QK_K/16; ++j) sumi += y[i].bsums[j] * mins[j/2];
+        a = aux8;
+        int is = 0;
+        for (int j = 0; j < QK_K/32; ++j) {
+            int32_t scale = scales[is++];
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+        }
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
+        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
+        const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
+        sumf -= dmin * sumi;
+    }
+    for (int l = 0; l < 8; ++l) sumf += sums[l];
+    *s = sumf;
+#endif
+}
+
+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);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q5_K * GGML_RESTRICT x = vx;
+    const block_q8_K * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+    static const uint32_t kmask1 = 0x3f3f3f3f;
+    static const uint32_t kmask2 = 0x0f0f0f0f;
+    static const uint32_t kmask3 = 0x03030303;
+
+    uint32_t utmp[4];
+
+#if defined(__POWER9_VECTOR__)
+    const vector signed char lowMask = vec_splats((signed char)0xF);
+    const vector signed char lowMask1 = vec_splats((int8_t)0x3f);
+    const vector signed char lowMask2 = vec_splats((int8_t)0x30);
+    const vector int v0 = vec_splats((int32_t)0);
+    const vector unsigned char v1 = vec_splats((unsigned char)0x1);
+    const vector unsigned char v2 = vec_splats((unsigned char)0x2);
+    const vector unsigned char v3 = vec_splats((unsigned char)0x3);
+    const vector unsigned char v4 = vec_splats((unsigned char)0x4);
+
+    vector float vsumf0 = vec_splats(0.0f);
+    vector float vsumf1 = vec_splats(0.0f);
+    vector float vsumf2 = vec_splats(0.0f);
+    vector float vsumf3 = vec_splats(0.0f);
+
+    for (int i = 0; i < nb; ++i) {
+        vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[i].d));
+        vector float vyd = vec_splats(y[i].d);
+        vector float vd = vec_mul(vxd, vyd);
+
+        vector float vxmin = vec_splats(GGML_CPU_FP16_TO_FP32(x[i].dmin));
+        vector float vdmin = vec_mul(vxmin, vyd);
+
+        UNUSED(kmask1);
+        UNUSED(kmask2);
+        UNUSED(kmask3);
+        UNUSED(utmp);
+
+        vector signed char u0 = (vector signed char)vec_xl_len(x[i].scales, 8);
+        vector signed char u1 = vec_and(vec_sr(u0, v2), lowMask2);
+        vector signed char u2 = (vector signed char)vec_xl_len(x[i].scales + 8, 4);
+        vector signed char u3 = vec_sr(u2, v4);
+
+        vector signed char u30 = u1;
+        vector signed char u31 = (vector signed char)vec_mergeh((vector signed int)vec_and(u2, lowMask), (vector signed int)u3);
+
+        u1 = vec_and(u0, lowMask1);
+        u2 = vec_or(u30, u31);
+
+        vector signed char utmps = (vector signed char)vec_mergeh((vector signed int)u1, (vector signed int)u2);
+
+        vector signed short q8ysums0 = vec_xl( 0, y[i].bsums);
+        vector signed short q8ysums1 = vec_xl(16, y[i].bsums);
+
+        vector signed short vscales = vec_unpackh(utmps);
+
+        vector signed short q5xmins = vec_unpackl(utmps);
+        vector signed short q5xmins0 = vec_mergeh(q5xmins, q5xmins);
+        vector signed short q5xmins1 = vec_mergel(q5xmins, q5xmins);
+
+        vector signed int prod0 = vec_mule(q5xmins0, q8ysums0);
+        vector signed int prod1 = vec_mule(q5xmins1, q8ysums1);
+        vector signed int prod2 = vec_mulo(q5xmins0, q8ysums0);
+        vector signed int prod3 = vec_mulo(q5xmins1, q8ysums1);
+
+        vsumf0 = vec_nmsub(vec_ctf(prod0, 0), vdmin, vsumf0);
+        vsumf1 = vec_nmsub(vec_ctf(prod1, 0), vdmin, vsumf1);
+        vsumf2 = vec_nmsub(vec_ctf(prod2, 0), vdmin, vsumf2);
+        vsumf3 = vec_nmsub(vec_ctf(prod3, 0), vdmin, vsumf3);
+
+        vector signed char qxhs0 = (vector signed char)vec_xl( 0, x[i].qh);
+        vector signed char qxhs1 = (vector signed char)vec_xl(16, x[i].qh);
+
+        vector signed int vsumi0 = v0;
+        vector signed int vsumi1 = v0;
+        vector signed int vsumi2 = v0;
+        vector signed int vsumi3 = v0;
+
+        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);
+            __builtin_prefetch(q8, 0, 1);
+
+            vector signed char qxs0 = (vector signed char)vec_xl( 0, q5);
+            vector signed char qxs1 = (vector signed char)vec_xl(16, q5);
+            q5 += 32;
+
+            vector signed char qxs00 = vec_and(qxs0, lowMask);
+            vector signed char qxs01 = vec_sr(qxs0, v4);
+            vector signed char qxs10 = vec_and(qxs1, lowMask);
+            vector signed char qxs11 = vec_sr(qxs1, v4);
+
+            vector signed char q5h00 = vec_sl(vec_and((vector signed char)v1, qxhs0), v4);
+            vector signed char q5h01 = vec_sl(vec_and((vector signed char)v2, qxhs0), v3);
+            vector signed char q5h10 = vec_sl(vec_and((vector signed char)v1, qxhs1), v4);
+            vector signed char q5h11 = vec_sl(vec_and((vector signed char)v2, qxhs1), v3);
+            qxhs0 = vec_sr(qxhs0, v2);
+            qxhs1 = vec_sr(qxhs1, v2);
+
+            vector unsigned char q5x00 = (vector unsigned char)vec_or(q5h00, qxs00);
+            vector unsigned char q5x01 = (vector unsigned char)vec_or(q5h01, qxs01);
+            vector unsigned char q5x10 = (vector unsigned char)vec_or(q5h10, qxs10);
+            vector unsigned char q5x11 = (vector unsigned char)vec_or(q5h11, qxs11);
+
+            vector signed char q8y00 = vec_xl( 0, q8);
+            vector signed char q8y10 = vec_xl(16, q8);
+            vector signed char q8y01 = vec_xl(32, q8);
+            vector signed char q8y11 = vec_xl(48, q8);
+            q8 += 64;
+
+            vector signed int qv00 = vec_msum(q8y00, q5x00, v0);
+            vector signed int qv01 = vec_msum(q8y01, q5x01, v0);
+            vector signed int qv10 = vec_msum(q8y10, q5x10, v0);
+            vector signed int qv11 = vec_msum(q8y11, q5x11, v0);
+
+            vector signed int vscales_h = vec_unpackh(vscales);
+            vector signed int vs0 = vec_splat(vscales_h, 0);
+            vector signed int vs1 = vec_splat(vscales_h, 1);
+            vscales = vec_sld(vscales, vscales, 12);
+
+            vsumi0 = vec_add(vec_mul(qv00, vs0), vsumi0);
+            vsumi1 = vec_add(vec_mul(qv10, vs0), vsumi1);
+            vsumi2 = vec_add(vec_mul(qv01, vs1), vsumi2);
+            vsumi3 = vec_add(vec_mul(qv11, vs1), vsumi3);
+        }
+
+        vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
+        vsumf1 = vec_madd(vec_ctf(vsumi1, 0), vd, vsumf1);
+        vsumf2 = vec_madd(vec_ctf(vsumi2, 0), vd, vsumf2);
+        vsumf3 = vec_madd(vec_ctf(vsumi3, 0), vd, vsumf3);
+    }
+
+    vsumf0 = vec_add(vsumf0, vsumf2);
+    vsumf1 = vec_add(vsumf1, vsumf3);
+
+    vsumf0 = vec_add(vsumf0, vsumf1);
+
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4));
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8));
+
+    *s = vec_extract(vsumf0, 0);
+
+#else
+
+    const uint8_t * scales = (const uint8_t*)&utmp[0];
+    const uint8_t * mins   = (const uint8_t*)&utmp[2];
+
+    int8_t  aux8[QK_K];
+    int16_t aux16[8];
+    float   sums [8];
+    int32_t aux32[8];
+    memset(sums, 0, 8*sizeof(float));
+
+    float sumf = 0;
+    for (int i = 0; i < nb; ++i) {
+        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 * 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);
+            for (int l = 0; l < 32; ++l) a[l] += (hm[l] & m ? 16 : 0);
+            a += 32; m <<= 1;
+            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l]  >> 4);
+            for (int l = 0; l < 32; ++l) a[l] += (hm[l] & m ? 16 : 0);
+            a += 32; m <<= 1;
+            q4 += 32;
+        }
+        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;
+
+        int sumi = 0;
+        for (int j = 0; j < QK_K/16; ++j) sumi += y[i].bsums[j] * mins[j/2];
+        a = aux8;
+        int is = 0;
+        for (int j = 0; j < QK_K/32; ++j) {
+            int32_t scale = scales[is++];
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+        }
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
+        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
+        const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
+        sumf -= dmin * sumi;
+    }
+    for (int l = 0; l < 8; ++l) sumf += sums[l];
+    *s = sumf;
+#endif
+}
+
+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);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q6_K * GGML_RESTRICT x = vx;
+    const block_q8_K * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+#if defined(__POWER9_VECTOR__)
+    const vector signed char lowMask = vec_splats((signed char)0xF);
+    const vector int v0 = vec_splats((int32_t)0);
+    const vector unsigned char v2 = vec_splats((unsigned char)0x2);
+    const vector unsigned char v3 = vec_splats((unsigned char)0x3);
+    const vector unsigned char v4 = vec_splats((unsigned char)0x4);
+    const vector unsigned char v6 = vec_splats((unsigned char)0x6);
+    const vector signed char off = vec_splats((signed char)0x20);
+
+    vector float vsumf0 = vec_splats(0.0f);
+    vector float vsumf1 = vec_splats(0.0f);
+    vector float vsumf2 = vec_splats(0.0f);
+    vector float vsumf3 = vec_splats(0.0f);
+
+    for (int i = 0; i < nb; ++i) {
+        vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[i].d));
+        vector float vyd = vec_splats(y[i].d);
+        vector float vd = vec_mul(vxd, vyd);
+
+        vector signed int vsumi0 = v0;
+        vector signed int vsumi1 = v0;
+        vector signed int vsumi2 = v0;
+        vector signed int vsumi3 = v0;
+        vector signed int vsumi4 = v0;
+        vector signed int vsumi5 = v0;
+        vector signed int vsumi6 = v0;
+        vector signed int vsumi7 = v0;
+
+        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);
+            __builtin_prefetch(qh, 0, 0);
+            __builtin_prefetch(q8, 0, 0);
+
+            vector signed char qxs0 = (vector signed char)vec_xl( 0, q6);
+            vector signed char qxs1 = (vector signed char)vec_xl(16, q6);
+            vector signed char qxs2 = (vector signed char)vec_xl(32, q6);
+            vector signed char qxs3 = (vector signed char)vec_xl(48, q6);
+            q6 += 64;
+
+            vector signed char qxs00 = vec_and(qxs0, lowMask);
+            vector signed char qxs01 = vec_sr(qxs0, v4);
+            vector signed char qxs10 = vec_and(qxs1, lowMask);
+            vector signed char qxs11 = vec_sr(qxs1, v4);
+            vector signed char qxs20 = vec_and(qxs2, lowMask);
+            vector signed char qxs21 = vec_sr(qxs2, v4);
+            vector signed char qxs30 = vec_and(qxs3, lowMask);
+            vector signed char qxs31 = vec_sr(qxs3, v4);
+
+            vector signed char qxhs0 = (vector signed char)vec_xl( 0, qh);
+            vector signed char qxhs1 = (vector signed char)vec_xl(16, qh);
+            qh += 32;
+
+            vector signed char qxh00 = vec_sl(vec_and((vector signed char)v3, qxhs0), v4);
+            vector signed char qxh01 = vec_sl(vec_and((vector signed char)v3, vec_sr(qxhs0, v4)), v4);
+            vector signed char qxh10 = vec_sl(vec_and((vector signed char)v3, qxhs1), v4);
+            vector signed char qxh11 = vec_sl(vec_and((vector signed char)v3, vec_sr(qxhs1, v4)), v4);
+            vector signed char qxh20 = vec_sl(vec_and((vector signed char)v3, vec_sr(qxhs0, v2)), v4);
+            vector signed char qxh21 = vec_sl(vec_and((vector signed char)v3, vec_sr(qxhs0, v6)), v4);
+            vector signed char qxh30 = vec_sl(vec_and((vector signed char)v3, vec_sr(qxhs1, v2)), v4);
+            vector signed char qxh31 = vec_sl(vec_and((vector signed char)v3, vec_sr(qxhs1, v6)), v4);
+
+            vector signed char q6x00 = vec_sub(vec_or(qxh00, qxs00), off);
+            vector signed char q6x01 = vec_sub(vec_or(qxh01, qxs01), off);
+            vector signed char q6x10 = vec_sub(vec_or(qxh10, qxs10), off);
+            vector signed char q6x11 = vec_sub(vec_or(qxh11, qxs11), off);
+            vector signed char q6x20 = vec_sub(vec_or(qxh20, qxs20), off);
+            vector signed char q6x21 = vec_sub(vec_or(qxh21, qxs21), off);
+            vector signed char q6x30 = vec_sub(vec_or(qxh30, qxs30), off);
+            vector signed char q6x31 = vec_sub(vec_or(qxh31, qxs31), off);
+
+            vector signed char q8y00 = vec_xl(  0, q8);
+            vector signed char q8y10 = vec_xl( 16, q8);
+            vector signed char q8y20 = vec_xl( 32, q8);
+            vector signed char q8y30 = vec_xl( 48, q8);
+            vector signed char q8y01 = vec_xl( 64, q8);
+            vector signed char q8y11 = vec_xl( 80, q8);
+            vector signed char q8y21 = vec_xl( 96, q8);
+            vector signed char q8y31 = vec_xl(112, q8);
+            q8 += 128;
+
+            vector signed short qv00 = vec_add(vec_mule(q6x00, q8y00), vec_mulo(q6x00, q8y00));
+            vector signed short qv10 = vec_add(vec_mule(q6x10, q8y10), vec_mulo(q6x10, q8y10));
+            vector signed short qv20 = vec_add(vec_mule(q6x20, q8y20), vec_mulo(q6x20, q8y20));
+            vector signed short qv30 = vec_add(vec_mule(q6x30, q8y30), vec_mulo(q6x30, q8y30));
+            vector signed short qv01 = vec_add(vec_mule(q6x01, q8y01), vec_mulo(q6x01, q8y01));
+            vector signed short qv11 = vec_add(vec_mule(q6x11, q8y11), vec_mulo(q6x11, q8y11));
+            vector signed short qv21 = vec_add(vec_mule(q6x21, q8y21), vec_mulo(q6x21, q8y21));
+            vector signed short qv31 = vec_add(vec_mule(q6x31, q8y31), vec_mulo(q6x31, q8y31));
+
+            vector signed short vscales = vec_unpackh(vec_xl_len(qs, 8));
+            qs += 8;
+
+            vector signed short vs0 = vec_splat(vscales, 0);
+            vector signed short vs1 = vec_splat(vscales, 1);
+            vector signed short vs2 = vec_splat(vscales, 2);
+            vector signed short vs3 = vec_splat(vscales, 3);
+            vector signed short vs4 = vec_splat(vscales, 4);
+            vector signed short vs5 = vec_splat(vscales, 5);
+            vector signed short vs6 = vec_splat(vscales, 6);
+            vector signed short vs7 = vec_splat(vscales, 7);
+
+            vsumi0 = vec_msum(qv00, vs0, vsumi0);
+            vsumi1 = vec_msum(qv01, vs4, vsumi1);
+            vsumi2 = vec_msum(qv10, vs1, vsumi2);
+            vsumi3 = vec_msum(qv11, vs5, vsumi3);
+            vsumi4 = vec_msum(qv20, vs2, vsumi4);
+            vsumi5 = vec_msum(qv21, vs6, vsumi5);
+            vsumi6 = vec_msum(qv30, vs3, vsumi6);
+            vsumi7 = vec_msum(qv31, vs7, vsumi7);
+        }
+
+        vsumi0 = vec_add(vsumi0, vsumi4);
+        vsumi1 = vec_add(vsumi1, vsumi5);
+        vsumi2 = vec_add(vsumi2, vsumi6);
+        vsumi3 = vec_add(vsumi3, vsumi7);
+
+        vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
+        vsumf1 = vec_madd(vec_ctf(vsumi1, 0), vd, vsumf1);
+        vsumf2 = vec_madd(vec_ctf(vsumi2, 0), vd, vsumf2);
+        vsumf3 = vec_madd(vec_ctf(vsumi3, 0), vd, vsumf3);
+    }
+
+    vsumf0 = vec_add(vsumf0, vsumf2);
+    vsumf1 = vec_add(vsumf1, vsumf3);
+
+    vsumf0 = vec_add(vsumf0, vsumf1);
+
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4));
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8));
+
+    *s = vec_extract(vsumf0, 0);
+
+#else
+
+    int8_t  aux8[QK_K];
+    int16_t aux16[8];
+    float   sums [8];
+    int32_t aux32[8];
+    memset(sums, 0, 8*sizeof(float));
+
+    float sumf = 0;
+    for (int i = 0; i < nb; ++i) {
+        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 * 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;
+                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;
+        }
+        a = aux8;
+        int is = 0;
+        for (int j = 0; j < QK_K/16; ++j) {
+            int scale = x[i].scales[is++];
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+        }
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
+        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
+    }
+    for (int l = 0; l < 8; ++l) sumf += sums[l];
+    *s = sumf;
+#endif
+}
+
+#if defined (__POWER9_VECTOR__)
+static const int8_t keven_signs_q2xs[1024] = {
+     1,  1,  1,  1,  1,  1,  1,  1, -1,  1,  1,  1,  1,  1,  1, -1,  1, -1,  1,  1,  1,  1,  1, -1, -1, -1,  1,  1,  1,  1,  1,  1,
+     1,  1, -1,  1,  1,  1,  1, -1, -1,  1, -1,  1,  1,  1,  1,  1,  1, -1, -1,  1,  1,  1,  1,  1, -1, -1, -1,  1,  1,  1,  1, -1,
+     1,  1,  1, -1,  1,  1,  1, -1, -1,  1,  1, -1,  1,  1,  1,  1,  1, -1,  1, -1,  1,  1,  1,  1, -1, -1,  1, -1,  1,  1,  1, -1,
+     1,  1, -1, -1,  1,  1,  1,  1, -1,  1, -1, -1,  1,  1,  1, -1,  1, -1, -1, -1,  1,  1,  1, -1, -1, -1, -1, -1,  1,  1,  1,  1,
+     1,  1,  1,  1, -1,  1,  1, -1, -1,  1,  1,  1, -1,  1,  1,  1,  1, -1,  1,  1, -1,  1,  1,  1, -1, -1,  1,  1, -1,  1,  1, -1,
+     1,  1, -1,  1, -1,  1,  1,  1, -1,  1, -1,  1, -1,  1,  1, -1,  1, -1, -1,  1, -1,  1,  1, -1, -1, -1, -1,  1, -1,  1,  1,  1,
+     1,  1,  1, -1, -1,  1,  1,  1, -1,  1,  1, -1, -1,  1,  1, -1,  1, -1,  1, -1, -1,  1,  1, -1, -1, -1,  1, -1, -1,  1,  1,  1,
+     1,  1, -1, -1, -1,  1,  1, -1, -1,  1, -1, -1, -1,  1,  1,  1,  1, -1, -1, -1, -1,  1,  1,  1, -1, -1, -1, -1, -1,  1,  1, -1,
+     1,  1,  1,  1,  1, -1,  1, -1, -1,  1,  1,  1,  1, -1,  1,  1,  1, -1,  1,  1,  1, -1,  1,  1, -1, -1,  1,  1,  1, -1,  1, -1,
+     1,  1, -1,  1,  1, -1,  1,  1, -1,  1, -1,  1,  1, -1,  1, -1,  1, -1, -1,  1,  1, -1,  1, -1, -1, -1, -1,  1,  1, -1,  1,  1,
+     1,  1,  1, -1,  1, -1,  1,  1, -1,  1,  1, -1,  1, -1,  1, -1,  1, -1,  1, -1,  1, -1,  1, -1, -1, -1,  1, -1,  1, -1,  1,  1,
+     1,  1, -1, -1,  1, -1,  1, -1, -1,  1, -1, -1,  1, -1,  1,  1,  1, -1, -1, -1,  1, -1,  1,  1, -1, -1, -1, -1,  1, -1,  1, -1,
+     1,  1,  1,  1, -1, -1,  1,  1, -1,  1,  1,  1, -1, -1,  1, -1,  1, -1,  1,  1, -1, -1,  1, -1, -1, -1,  1,  1, -1, -1,  1,  1,
+     1,  1, -1,  1, -1, -1,  1, -1, -1,  1, -1,  1, -1, -1,  1,  1,  1, -1, -1,  1, -1, -1,  1,  1, -1, -1, -1,  1, -1, -1,  1, -1,
+     1,  1,  1, -1, -1, -1,  1, -1, -1,  1,  1, -1, -1, -1,  1,  1,  1, -1,  1, -1, -1, -1,  1,  1, -1, -1,  1, -1, -1, -1,  1, -1,
+     1,  1, -1, -1, -1, -1,  1,  1, -1,  1, -1, -1, -1, -1,  1, -1,  1, -1, -1, -1, -1, -1,  1, -1, -1, -1, -1, -1, -1, -1,  1,  1,
+     1,  1,  1,  1,  1,  1, -1, -1, -1,  1,  1,  1,  1,  1, -1,  1,  1, -1,  1,  1,  1,  1, -1,  1, -1, -1,  1,  1,  1,  1, -1, -1,
+     1,  1, -1,  1,  1,  1, -1,  1, -1,  1, -1,  1,  1,  1, -1, -1,  1, -1, -1,  1,  1,  1, -1, -1, -1, -1, -1,  1,  1,  1, -1,  1,
+     1,  1,  1, -1,  1,  1, -1,  1, -1,  1,  1, -1,  1,  1, -1, -1,  1, -1,  1, -1,  1,  1, -1, -1, -1, -1,  1, -1,  1,  1, -1,  1,
+     1,  1, -1, -1,  1,  1, -1, -1, -1,  1, -1, -1,  1,  1, -1,  1,  1, -1, -1, -1,  1,  1, -1,  1, -1, -1, -1, -1,  1,  1, -1, -1,
+     1,  1,  1,  1, -1,  1, -1,  1, -1,  1,  1,  1, -1,  1, -1, -1,  1, -1,  1,  1, -1,  1, -1, -1, -1, -1,  1,  1, -1,  1, -1,  1,
+     1,  1, -1,  1, -1,  1, -1, -1, -1,  1, -1,  1, -1,  1, -1,  1,  1, -1, -1,  1, -1,  1, -1,  1, -1, -1, -1,  1, -1,  1, -1, -1,
+     1,  1,  1, -1, -1,  1, -1, -1, -1,  1,  1, -1, -1,  1, -1,  1,  1, -1,  1, -1, -1,  1, -1,  1, -1, -1,  1, -1, -1,  1, -1, -1,
+     1,  1, -1, -1, -1,  1, -1,  1, -1,  1, -1, -1, -1,  1, -1, -1,  1, -1, -1, -1, -1,  1, -1, -1, -1, -1, -1, -1, -1,  1, -1,  1,
+     1,  1,  1,  1,  1, -1, -1,  1, -1,  1,  1,  1,  1, -1, -1, -1,  1, -1,  1,  1,  1, -1, -1, -1, -1, -1,  1,  1,  1, -1, -1,  1,
+     1,  1, -1,  1,  1, -1, -1, -1, -1,  1, -1,  1,  1, -1, -1,  1,  1, -1, -1,  1,  1, -1, -1,  1, -1, -1, -1,  1,  1, -1, -1, -1,
+     1,  1,  1, -1,  1, -1, -1, -1, -1,  1,  1, -1,  1, -1, -1,  1,  1, -1,  1, -1,  1, -1, -1,  1, -1, -1,  1, -1,  1, -1, -1, -1,
+     1,  1, -1, -1,  1, -1, -1,  1, -1,  1, -1, -1,  1, -1, -1, -1,  1, -1, -1, -1,  1, -1, -1, -1, -1, -1, -1, -1,  1, -1, -1,  1,
+     1,  1,  1,  1, -1, -1, -1, -1, -1,  1,  1,  1, -1, -1, -1,  1,  1, -1,  1,  1, -1, -1, -1,  1, -1, -1,  1,  1, -1, -1, -1, -1,
+     1,  1, -1,  1, -1, -1, -1,  1, -1,  1, -1,  1, -1, -1, -1, -1,  1, -1, -1,  1, -1, -1, -1, -1, -1, -1, -1,  1, -1, -1, -1,  1,
+     1,  1,  1, -1, -1, -1, -1,  1, -1,  1,  1, -1, -1, -1, -1, -1,  1, -1,  1, -1, -1, -1, -1, -1, -1, -1,  1, -1, -1, -1, -1,  1,
+     1,  1, -1, -1, -1, -1, -1, -1, -1,  1, -1, -1, -1, -1, -1,  1,  1, -1, -1, -1, -1, -1, -1,  1, -1, -1, -1, -1, -1, -1, -1, -1,
+};
+#endif
+
+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);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_iq2_xxs * GGML_RESTRICT x = vx;
+    const block_q8_K    * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+#if defined(__POWER9_VECTOR__)
+    const vector int v0 = vec_splats((int32_t)0);
+    vector float vsumf0 = vec_splats(0.0f);
+    vector float vsumf1 = vec_splats(0.0f);
+    vector float vsumf2 = vec_splats(0.0f);
+    vector float vsumf3 = vec_splats(0.0f);
+
+    const uint64_t * signs64 = (const uint64_t *)keven_signs_q2xs;
+
+    for (int i = 0; i < nb; ++i) {
+        vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[i].d));
+        vector float vyd = vec_splats(y[i].d);
+        vector float vd = vec_mul(vxd, vyd);
+
+        vector signed int vsumi0 = v0;
+        vector signed int vsumi1 = v0;
+        vector signed int vsumi2 = v0;
+        vector signed int vsumi3 = v0;
+
+        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);
+            __builtin_prefetch(q8, 0, 1);
+
+            uint32_t aux32[4];
+            const uint8_t * aux8 = (const uint8_t *)aux32;
+
+            memcpy(aux32, q2, 4*sizeof(uint32_t));
+            q2 += 8;
+
+            vector signed long long aux64x2_0 = {*(const int64_t *)(iq2xxs_grid + aux8[ 0]), *(const int64_t *)(iq2xxs_grid + aux8[ 1])};
+            vector signed long long aux64x2_1 = {*(const int64_t *)(iq2xxs_grid + aux8[ 2]), *(const int64_t *)(iq2xxs_grid + aux8[ 3])};
+            vector signed long long aux64x2_2 = {*(const int64_t *)(iq2xxs_grid + aux8[ 8]), *(const int64_t *)(iq2xxs_grid + aux8[ 9])};
+            vector signed long long aux64x2_3 = {*(const int64_t *)(iq2xxs_grid + aux8[10]), *(const int64_t *)(iq2xxs_grid + aux8[11])};
+
+            vector signed long long vsigns0 = {*(const int64_t *)(signs64 + ((aux32[1] >>  0) & 127)), *(const int64_t *)(signs64 + ((aux32[1] >>  7) & 127))};
+            vector signed long long vsigns1 = {*(const int64_t *)(signs64 + ((aux32[1] >> 14) & 127)), *(const int64_t *)(signs64 + ((aux32[1] >> 21) & 127))};
+            vector signed long long vsigns2 = {*(const int64_t *)(signs64 + ((aux32[3] >>  0) & 127)), *(const int64_t *)(signs64 + ((aux32[3] >>  7) & 127))};
+            vector signed long long vsigns3 = {*(const int64_t *)(signs64 + ((aux32[3] >> 14) & 127)), *(const int64_t *)(signs64 + ((aux32[3] >> 21) & 127))};
+
+            vector signed char q2x0 = (vector signed char)vec_mul((vector signed char)vsigns0, (vector signed char)aux64x2_0);
+            vector signed char q2x1 = (vector signed char)vec_mul((vector signed char)vsigns1, (vector signed char)aux64x2_1);
+            vector signed char q2x2 = (vector signed char)vec_mul((vector signed char)vsigns2, (vector signed char)aux64x2_2);
+            vector signed char q2x3 = (vector signed char)vec_mul((vector signed char)vsigns3, (vector signed char)aux64x2_3);
+
+            vector signed char q8y0 = vec_xl( 0, q8);
+            vector signed char q8y1 = vec_xl(16, q8);
+            vector signed char q8y2 = vec_xl(32, q8);
+            vector signed char q8y3 = vec_xl(48, q8);
+            q8 += 64;
+
+            vector signed short qv0 = vec_add(vec_mule(q2x0, q8y0), vec_mulo(q2x0, q8y0));
+            vector signed short qv1 = vec_add(vec_mule(q2x1, q8y1), vec_mulo(q2x1, q8y1));
+            vector signed short qv2 = vec_add(vec_mule(q2x2, q8y2), vec_mulo(q2x2, q8y2));
+            vector signed short qv3 = vec_add(vec_mule(q2x3, q8y3), vec_mulo(q2x3, q8y3));
+
+            const uint16_t ls0 = aux32[1] >> 28;
+            const uint16_t ls1 = aux32[3] >> 28;
+
+            vector signed short vscales01 = vec_splats((int16_t)(2*ls0+1));
+            vector signed short vscales23 = vec_splats((int16_t)(2*ls1+1));
+
+            vsumi0 = vec_msum(qv0, vscales01, vsumi0);
+            vsumi1 = vec_msum(qv1, vscales01, vsumi1);
+            vsumi2 = vec_msum(qv2, vscales23, vsumi2);
+            vsumi3 = vec_msum(qv3, vscales23, vsumi3);
+        }
+
+        vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
+        vsumf1 = vec_madd(vec_ctf(vsumi1, 0), vd, vsumf1);
+        vsumf2 = vec_madd(vec_ctf(vsumi2, 0), vd, vsumf2);
+        vsumf3 = vec_madd(vec_ctf(vsumi3, 0), vd, vsumf3);
+    }
+
+    vsumf0 = vec_add(vsumf0, vsumf2);
+    vsumf1 = vec_add(vsumf1, vsumf3);
+
+    vsumf0 = vec_add(vsumf0, vsumf1);
+
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4));
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8));
+
+    *s = 0.125f * vec_extract(vsumf0, 0);
+
+#else
+
+    uint32_t aux32[2];
+    const uint8_t * aux8 = (const uint8_t *)aux32;
+
+    float sumf = 0.f;
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_CPU_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;
+        int32_t bsum = 0;
+        for (int ib32 = 0; ib32 < QK_K/32; ++ib32) {
+            memcpy(aux32, q2, 2*sizeof(uint32_t));
+            q2 += 4;
+            const uint32_t ls = 2*(aux32[1] >> 28) + 1;
+            int32_t sumi = 0;
+            for (int l = 0; l < 4; ++l) {
+                const uint8_t * grid = (const uint8_t *)(iq2xxs_grid + aux8[l]);
+                const uint8_t  signs = ksigns_iq2xs[(aux32[1] >> 7*l) & 127];
+                for (int j = 0; j < 8; ++j) {
+                    sumi += grid[j] * q8[j] * (signs & kmask_iq2xs[j] ? -1 : 1);
+                }
+                q8 += 8;
+            }
+            bsum += sumi * ls;
+        }
+        sumf += d * bsum;
+    }
+    *s = 0.125f * sumf;
+#endif
+}
+
+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);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_iq2_xs * GGML_RESTRICT x = vx;
+    const block_q8_K   * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+#if defined(__POWER9_VECTOR__)
+    const vector int v0 = vec_splats((int32_t)0);
+    vector float vsumf0 = vec_splats(0.0f);
+    vector float vsumf1 = vec_splats(0.0f);
+    vector float vsumf2 = vec_splats(0.0f);
+    vector float vsumf3 = vec_splats(0.0f);
+
+    const uint64_t * signs64 = (const uint64_t *)keven_signs_q2xs;
+
+    for (int i = 0; i < nb; ++i) {
+        vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[i].d));
+        vector float vyd = vec_splats(y[i].d);
+        vector float vd = vec_mul(vxd, vyd);
+
+        vector signed int vsumi0 = v0;
+        vector signed int vsumi1 = v0;
+        vector signed int vsumi2 = v0;
+        vector signed int vsumi3 = v0;
+
+        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);
+            __builtin_prefetch(q8, 0, 1);
+
+            vector signed long long aux64x2_0 = {*(const int64_t *)(iq2xs_grid + (q2[0] & 511)), *(const int64_t *)(iq2xs_grid + (q2[1] & 511))};
+            vector signed long long aux64x2_1 = {*(const int64_t *)(iq2xs_grid + (q2[2] & 511)), *(const int64_t *)(iq2xs_grid + (q2[3] & 511))};
+            vector signed long long aux64x2_2 = {*(const int64_t *)(iq2xs_grid + (q2[4] & 511)), *(const int64_t *)(iq2xs_grid + (q2[5] & 511))};
+            vector signed long long aux64x2_3 = {*(const int64_t *)(iq2xs_grid + (q2[6] & 511)), *(const int64_t *)(iq2xs_grid + (q2[7] & 511))};
+
+            vector signed long long vsigns0 = {*(const int64_t *)(signs64 + ((q2[0] >> 9))), *(const int64_t *)(signs64 + ((q2[1] >> 9)))};
+            vector signed long long vsigns1 = {*(const int64_t *)(signs64 + ((q2[2] >> 9))), *(const int64_t *)(signs64 + ((q2[3] >> 9)))};
+            vector signed long long vsigns2 = {*(const int64_t *)(signs64 + ((q2[4] >> 9))), *(const int64_t *)(signs64 + ((q2[5] >> 9)))};
+            vector signed long long vsigns3 = {*(const int64_t *)(signs64 + ((q2[6] >> 9))), *(const int64_t *)(signs64 + ((q2[7] >> 9)))};
+            q2 += 8;
+
+            vector signed char q2x0 = (vector signed char)vec_mul((vector signed char)vsigns0, (vector signed char)aux64x2_0);
+            vector signed char q2x1 = (vector signed char)vec_mul((vector signed char)vsigns1, (vector signed char)aux64x2_1);
+            vector signed char q2x2 = (vector signed char)vec_mul((vector signed char)vsigns2, (vector signed char)aux64x2_2);
+            vector signed char q2x3 = (vector signed char)vec_mul((vector signed char)vsigns3, (vector signed char)aux64x2_3);
+
+            vector signed char q8y0 = vec_xl( 0, q8);
+            vector signed char q8y1 = vec_xl(16, q8);
+            vector signed char q8y2 = vec_xl(32, q8);
+            vector signed char q8y3 = vec_xl(48, q8);
+            q8 += 64;
+
+            vector signed short qv0 = vec_add(vec_mule(q2x0, q8y0), vec_mulo(q2x0, q8y0));
+            vector signed short qv1 = vec_add(vec_mule(q2x1, q8y1), vec_mulo(q2x1, q8y1));
+            vector signed short qv2 = vec_add(vec_mule(q2x2, q8y2), vec_mulo(q2x2, q8y2));
+            vector signed short qv3 = vec_add(vec_mule(q2x3, q8y3), vec_mulo(q2x3, q8y3));
+
+            const uint16_t ls0 = (uint16_t)(sc[0] & 0xf);
+            const uint16_t ls1 = (uint16_t)(sc[0] >>  4);
+            const uint16_t ls2 = (uint16_t)(sc[1] & 0xf);
+            const uint16_t ls3 = (uint16_t)(sc[1] >>  4);
+            sc += 2;
+
+            vector signed short vscales0 = vec_splats((int16_t)(2*ls0+1));
+            vector signed short vscales1 = vec_splats((int16_t)(2*ls1+1));
+            vector signed short vscales2 = vec_splats((int16_t)(2*ls2+1));
+            vector signed short vscales3 = vec_splats((int16_t)(2*ls3+1));
+
+            vsumi0 = vec_msum(qv0, vscales0, vsumi0);
+            vsumi1 = vec_msum(qv1, vscales1, vsumi1);
+            vsumi2 = vec_msum(qv2, vscales2, vsumi2);
+            vsumi3 = vec_msum(qv3, vscales3, vsumi3);
+        }
+
+        vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
+        vsumf1 = vec_madd(vec_ctf(vsumi1, 0), vd, vsumf1);
+        vsumf2 = vec_madd(vec_ctf(vsumi2, 0), vd, vsumf2);
+        vsumf3 = vec_madd(vec_ctf(vsumi3, 0), vd, vsumf3);
+    }
+
+    vsumf0 = vec_add(vsumf0, vsumf2);
+    vsumf1 = vec_add(vsumf1, vsumf3);
+
+    vsumf0 = vec_add(vsumf0, vsumf1);
+
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4));
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8));
+
+    *s = 0.125f * vec_extract(vsumf0, 0);
+
+#else
+
+    float sumf = 0.f;
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
+        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;
+            const uint16_t ls2 = 2*(sc[ib32] >>  4) + 1;
+            int32_t sumi = 0;
+            for (int l = 0; l < 2; ++l) {
+                const uint8_t * grid = (const uint8_t *)(iq2xs_grid + (q2[l] & 511));
+                const uint8_t  signs = ksigns_iq2xs[q2[l] >> 9];
+                for (int j = 0; j < 8; ++j) {
+                    sumi += grid[j] * q8[j] * (signs & kmask_iq2xs[j] ? -1 : 1);
+                }
+                q8 += 8;
+            }
+            bsum += sumi * ls1;
+            sumi = 0;
+            for (int l = 2; l < 4; ++l) {
+                const uint8_t * grid = (const uint8_t *)(iq2xs_grid + (q2[l] & 511));
+                const uint8_t  signs = ksigns_iq2xs[q2[l] >> 9];
+                for (int j = 0; j < 8; ++j) {
+                    sumi += grid[j] * q8[j] * (signs & kmask_iq2xs[j] ? -1 : 1);
+                }
+                q8 += 8;
+            }
+            bsum += sumi * ls2;
+            q2 += 4;
+        }
+        sumf += d * bsum;
+    }
+    *s = 0.125f * sumf;
+#endif
+}
+
+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);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_iq2_s * GGML_RESTRICT x = vx;
+    const block_q8_K  * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+#if defined(__POWER9_VECTOR__)
+    static const uint8_t k_mask1[32] = {0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01,
+                                        0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03
+    };
+
+    static const uint8_t k_mask2[16] = {0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,};
+
+    const vector int v0 = vec_splats((int32_t)0);
+
+    vector float vsumf0 = vec_splats(0.0f);
+    vector float vsumf1 = vec_splats(0.0f);
+    vector float vsumf2 = vec_splats(0.0f);
+    vector float vsumf3 = vec_splats(0.0f);
+
+    const vector unsigned char mask0 = vec_xl( 0, k_mask1);
+    const vector unsigned char mask1 = vec_xl(16, k_mask1);
+    const vector signed char mask2 = (vector signed char)vec_xl( 0, k_mask2);
+
+    for (int i = 0; i < nb; ++i) {
+        vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[i].d));
+        vector float vyd = vec_splats(y[i].d);
+        vector float vd = vec_mul(vxd, vyd);
+
+        vector signed int vsumi0 = v0;
+        vector signed int vsumi1 = v0;
+        vector signed int vsumi2 = v0;
+        vector signed int vsumi3 = v0;
+
+        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);
+            __builtin_prefetch(q8, 0, 1);
+
+            vector signed long long aux64x2_0 = {*(const int64_t *)(iq2s_grid + (q2[0] | ((qh[0] << 8) & 0x300))), *(const int64_t *)(iq2s_grid + (q2[1] | ((qh[0] << 6) & 0x300)))};
+            vector signed long long aux64x2_1 = {*(const int64_t *)(iq2s_grid + (q2[2] | ((qh[0] << 4) & 0x300))), *(const int64_t *)(iq2s_grid + (q2[3] | ((qh[0] << 2) & 0x300)))};
+            vector signed long long aux64x2_2 = {*(const int64_t *)(iq2s_grid + (q2[4] | ((qh[1] << 8) & 0x300))), *(const int64_t *)(iq2s_grid + (q2[5] | ((qh[1] << 6) & 0x300)))};
+            vector signed long long aux64x2_3 = {*(const int64_t *)(iq2s_grid + (q2[6] | ((qh[1] << 4) & 0x300))), *(const int64_t *)(iq2s_grid + (q2[7] | ((qh[1] << 2) & 0x300)))};
+            q2 += 8;
+            qh += 2;
+
+            vector signed char vsigns01 = (vector signed char)vec_splats(*(const uint32_t *)&signs[0]);
+            vector signed char vsigns23 = (vector signed char)vec_splats(*(const uint32_t *)&signs[2]);
+            signs += 4;
+
+            vector signed char vsigns0 = vec_perm(vsigns01, vsigns01, mask0);
+            vector signed char vsigns1 = vec_perm(vsigns01, vsigns01, mask1);
+            vector signed char vsigns2 = vec_perm(vsigns23, vsigns23, mask0);
+            vector signed char vsigns3 = vec_perm(vsigns23, vsigns23, mask1);
+
+            vsigns0 = (vector signed char)vec_cmpeq(vec_and(vsigns0, mask2), mask2);
+            vsigns1 = (vector signed char)vec_cmpeq(vec_and(vsigns1, mask2), mask2);
+            vsigns2 = (vector signed char)vec_cmpeq(vec_and(vsigns2, mask2), mask2);
+            vsigns3 = (vector signed char)vec_cmpeq(vec_and(vsigns3, mask2), mask2);
+
+            vector signed char q2x0 = vec_sub(vec_xor(vsigns0, (vector signed char)aux64x2_0), vsigns0);
+            vector signed char q2x1 = vec_sub(vec_xor(vsigns1, (vector signed char)aux64x2_1), vsigns1);
+            vector signed char q2x2 = vec_sub(vec_xor(vsigns2, (vector signed char)aux64x2_2), vsigns2);
+            vector signed char q2x3 = vec_sub(vec_xor(vsigns3, (vector signed char)aux64x2_3), vsigns3);
+
+            vector signed char q8y0 = vec_xl( 0, q8);
+            vector signed char q8y1 = vec_xl(16, q8);
+            vector signed char q8y2 = vec_xl(32, q8);
+            vector signed char q8y3 = vec_xl(48, q8);
+            q8 += 64;
+
+            vector signed short qv0 = vec_add(vec_mule(q2x0, q8y0), vec_mulo(q2x0, q8y0));
+            vector signed short qv1 = vec_add(vec_mule(q2x1, q8y1), vec_mulo(q2x1, q8y1));
+            vector signed short qv2 = vec_add(vec_mule(q2x2, q8y2), vec_mulo(q2x2, q8y2));
+            vector signed short qv3 = vec_add(vec_mule(q2x3, q8y3), vec_mulo(q2x3, q8y3));
+
+            const uint16_t ls0 = (uint16_t)(sc[0] & 0xf);
+            const uint16_t ls1 = (uint16_t)(sc[0] >>  4);
+            const uint16_t ls2 = (uint16_t)(sc[1] & 0xf);
+            const uint16_t ls3 = (uint16_t)(sc[1] >>  4);
+            sc += 2;
+
+            vector signed short vscales0 = vec_splats((int16_t)(2*ls0+1));
+            vector signed short vscales1 = vec_splats((int16_t)(2*ls1+1));
+            vector signed short vscales2 = vec_splats((int16_t)(2*ls2+1));
+            vector signed short vscales3 = vec_splats((int16_t)(2*ls3+1));
+
+            vsumi0 = vec_msum(qv0, vscales0, vsumi0);
+            vsumi1 = vec_msum(qv1, vscales1, vsumi1);
+            vsumi2 = vec_msum(qv2, vscales2, vsumi2);
+            vsumi3 = vec_msum(qv3, vscales3, vsumi3);
+        }
+
+        vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
+        vsumf1 = vec_madd(vec_ctf(vsumi1, 0), vd, vsumf1);
+        vsumf2 = vec_madd(vec_ctf(vsumi2, 0), vd, vsumf2);
+        vsumf3 = vec_madd(vec_ctf(vsumi3, 0), vd, vsumf3);
+    }
+
+    vsumf0 = vec_add(vsumf0, vsumf2);
+    vsumf1 = vec_add(vsumf1, vsumf3);
+
+    vsumf0 = vec_add(vsumf0, vsumf1);
+
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4));
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8));
+
+    *s = 0.125f * vec_extract(vsumf0, 0);
+
+#else
+
+    float sumf = 0;
+    for (int i = 0; i < nb; i++) {
+
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
+        const int8_t  * q8 = y[i].qs;
+        const uint8_t * qs = x[i].qs;
+        const uint8_t * qh = x[i].qh;
+        const uint8_t * signs = qs + QK_K/8;
+
+        int bsum = 0;
+        for (int ib32 = 0; ib32 < QK_K/32; ++ib32) {
+            int ls1 = 1 + 2*(x[i].scales[ib32] & 0xf);
+            int ls2 = 1 + 2*(x[i].scales[ib32] >>  4);
+            int sumi1 = 0, sumi2 = 0;
+            for (int l = 0; l < 2; ++l) {
+                const uint8_t * grid = (const uint8_t *)(iq2s_grid + (qs[l] | (qh[ib32] << (8-2*l) & 0x300)));
+                for (int j = 0; j < 8; ++j) {
+                    sumi1 += q8[j] * grid[j] * (signs[l] & kmask_iq2xs[j] ? -1 : 1);
+                }
+                q8 += 8;
+            }
+            for (int l = 2; l < 4; ++l) {
+                const uint8_t * grid = (const uint8_t *)(iq2s_grid + (qs[l] | (qh[ib32] << (8-2*l) & 0x300)));
+                for (int j = 0; j < 8; ++j) {
+                    sumi2 += q8[j] * grid[j] * (signs[l] & kmask_iq2xs[j] ? -1 : 1);
+                }
+                q8 += 8;
+            }
+            bsum += ls1 * sumi1 + ls2 * sumi2;
+            qs += 4;
+            signs += 4;
+        }
+
+        sumf += d * bsum;
+    }
+
+    *s = 0.125f * sumf;
+
+#endif
+
+}
+
+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);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_iq3_xxs * GGML_RESTRICT x = vx;
+    const block_q8_K    * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+#if defined(__POWER9_VECTOR__)
+    const uint64_t * signs64 = (const uint64_t *)keven_signs_q2xs;
+
+    const vector int v0 = vec_splats((int32_t)0);
+
+    vector float vsumf0 = vec_splats(0.0f);
+    vector float vsumf1 = vec_splats(0.0f);
+    vector float vsumf2 = vec_splats(0.0f);
+    vector float vsumf3 = vec_splats(0.0f);
+
+    for (int i = 0; i < nb; ++i) {
+        vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[i].d));
+        vector float vyd = vec_splats(y[i].d);
+        vector float vd = vec_mul(vxd, vyd);
+
+        vector signed int vsumi0 = v0;
+        vector signed int vsumi1 = v0;
+        vector signed int vsumi2 = v0;
+        vector signed int vsumi3 = v0;
+
+        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) {
+            __builtin_prefetch(q3, 0, 1);
+            __builtin_prefetch(q8, 0, 1);
+
+            vector unsigned int aux32x4_0 = {iq3xxs_grid[q3[ 0]], iq3xxs_grid[q3[ 1]], iq3xxs_grid[q3[ 2]], iq3xxs_grid[q3[ 3]]};
+            vector unsigned int aux32x4_1 = {iq3xxs_grid[q3[ 4]], iq3xxs_grid[q3[ 5]], iq3xxs_grid[q3[ 6]], iq3xxs_grid[q3[ 7]]};
+            vector unsigned int aux32x4_2 = {iq3xxs_grid[q3[ 8]], iq3xxs_grid[q3[ 9]], iq3xxs_grid[q3[10]], iq3xxs_grid[q3[11]]};
+            vector unsigned int aux32x4_3 = {iq3xxs_grid[q3[12]], iq3xxs_grid[q3[13]], iq3xxs_grid[q3[14]], iq3xxs_grid[q3[15]]};
+            q3 += 16;
+
+            vector unsigned long long aux64x2_0 = {(uint64_t)(signs64[(signs[0] >>  0) & 127]), (uint64_t)(signs64[(signs[0] >>  7) & 127])};
+            vector unsigned long long aux64x2_1 = {(uint64_t)(signs64[(signs[0] >> 14) & 127]), (uint64_t)(signs64[(signs[0] >> 21) & 127])};
+            vector unsigned long long aux64x2_2 = {(uint64_t)(signs64[(signs[1] >>  0) & 127]), (uint64_t)(signs64[(signs[1] >>  7) & 127])};
+            vector unsigned long long aux64x2_3 = {(uint64_t)(signs64[(signs[1] >> 14) & 127]), (uint64_t)(signs64[(signs[1] >> 21) & 127])};
+
+            vector signed char q3x0 = vec_mul((vector signed char)aux64x2_0, (vector signed char)aux32x4_0);
+            vector signed char q3x1 = vec_mul((vector signed char)aux64x2_1, (vector signed char)aux32x4_1);
+            vector signed char q3x2 = vec_mul((vector signed char)aux64x2_2, (vector signed char)aux32x4_2);
+            vector signed char q3x3 = vec_mul((vector signed char)aux64x2_3, (vector signed char)aux32x4_3);
+
+            vector signed char q8y0 = vec_xl( 0, q8);
+            vector signed char q8y1 = vec_xl(16, q8);
+            vector signed char q8y2 = vec_xl(32, q8);
+            vector signed char q8y3 = vec_xl(48, q8);
+            q8 += 64;
+
+            vector signed short qv0 = vec_add(vec_mule(q3x0, q8y0), vec_mulo(q3x0, q8y0));
+            vector signed short qv1 = vec_add(vec_mule(q3x1, q8y1), vec_mulo(q3x1, q8y1));
+            vector signed short qv2 = vec_add(vec_mule(q3x2, q8y2), vec_mulo(q3x2, q8y2));
+            vector signed short qv3 = vec_add(vec_mule(q3x3, q8y3), vec_mulo(q3x3, q8y3));
+
+            const uint16_t ls0 = (uint16_t)(signs[0] >> 28);
+            const uint16_t ls1 = (uint16_t)(signs[1] >> 28);
+            signs += 2;
+
+            vector signed short vscales01 = (vector signed short)vec_splats((uint16_t)(2*ls0+1));
+            vector signed short vscales23 = (vector signed short)vec_splats((uint16_t)(2*ls1+1));
+
+            vsumi0 = vec_msum(qv0, vscales01, vsumi0);
+            vsumi1 = vec_msum(qv1, vscales01, vsumi1);
+            vsumi2 = vec_msum(qv2, vscales23, vsumi2);
+            vsumi3 = vec_msum(qv3, vscales23, vsumi3);
+        }
+
+        vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
+        vsumf1 = vec_madd(vec_ctf(vsumi1, 0), vd, vsumf1);
+        vsumf2 = vec_madd(vec_ctf(vsumi2, 0), vd, vsumf2);
+        vsumf3 = vec_madd(vec_ctf(vsumi3, 0), vd, vsumf3);
+    }
+
+    vsumf0 = vec_add(vsumf0, vsumf2);
+    vsumf1 = vec_add(vsumf1, vsumf3);
+
+    vsumf0 = vec_add(vsumf0, vsumf1);
+
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4));
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8));
+
+    *s = 0.25f * vec_extract(vsumf0, 0);
+
+#else
+
+    uint32_t aux32;
+
+    float sumf = 0.f;
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
+        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);
+            const uint32_t ls = 2*(aux32 >> 28) + 1;
+            int32_t sumi = 0;
+            for (int l = 0; l < 4; ++l) {
+                const uint8_t * grid1 = (const uint8_t *)(iq3xxs_grid + q3[2*l+0]);
+                const uint8_t * grid2 = (const uint8_t *)(iq3xxs_grid + q3[2*l+1]);
+                const uint8_t  signs = ksigns_iq2xs[(aux32 >> 7*l) & 127];
+                for (int j = 0; j < 4; ++j) {
+                    sumi += grid1[j] * q8[j+0] * (signs & kmask_iq2xs[j+0] ? -1 : 1);
+                    sumi += grid2[j] * q8[j+4] * (signs & kmask_iq2xs[j+4] ? -1 : 1);
+                }
+                q8 += 8;
+            }
+            q3 += 8;
+            bsum += sumi * ls;
+        }
+        sumf += d * bsum;
+    }
+    *s = 0.25f * sumf;
+#endif
+}
+
+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);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_iq3_s * GGML_RESTRICT x = vx;
+    const block_q8_K  * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+#if defined(__POWER9_VECTOR__)
+    static const uint8_t k_mask1[32] = {0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01,
+                                        0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03
+    };
+
+    static const uint8_t k_mask2[16] = {0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,};
+
+    const vector int v0 = vec_splats((int32_t)0);
+
+    vector float vsumf0 = vec_splats(0.0f);
+    vector float vsumf1 = vec_splats(0.0f);
+    vector float vsumf2 = vec_splats(0.0f);
+    vector float vsumf3 = vec_splats(0.0f);
+
+    const vector unsigned char mask0 = vec_xl( 0, k_mask1);
+    const vector unsigned char mask1 = vec_xl(16, k_mask1);
+    const vector signed char mask2 = (vector signed char)vec_xl( 0, k_mask2);
+
+    for (int i = 0; i < nb; ++i) {
+        vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[i].d));
+        vector float vyd = vec_splats(y[i].d);
+        vector float vd = vec_mul(vxd, vyd);
+
+        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;
+        vector signed int vsumi2 = v0;
+        vector signed int vsumi3 = v0;
+
+        for (int j = 0; j < QK_K/32; j += 2) {
+            __builtin_prefetch(q3, 0, 1);
+            __builtin_prefetch(q8, 0, 1);
+
+            vector unsigned int aux32x4_0 = {iq3s_grid[q3[ 0] | ((qh[0] << 8) & 256)], iq3s_grid[q3[ 1] | ((qh[0] << 7) & 256)],
+                                             iq3s_grid[q3[ 2] | ((qh[0] << 6) & 256)], iq3s_grid[q3[ 3] | ((qh[0] << 5) & 256)]};
+            vector unsigned int aux32x4_1 = {iq3s_grid[q3[ 4] | ((qh[0] << 4) & 256)], iq3s_grid[q3[ 5] | ((qh[0] << 3) & 256)],
+                                             iq3s_grid[q3[ 6] | ((qh[0] << 2) & 256)], iq3s_grid[q3[ 7] | ((qh[0] << 1) & 256)]};
+            vector unsigned int aux32x4_2 = {iq3s_grid[q3[ 8] | ((qh[1] << 8) & 256)], iq3s_grid[q3[ 9] | ((qh[1] << 7) & 256)],
+                                             iq3s_grid[q3[10] | ((qh[1] << 6) & 256)], iq3s_grid[q3[11] | ((qh[1] << 5) & 256)]};
+            vector unsigned int aux32x4_3 = {iq3s_grid[q3[12] | ((qh[1] << 4) & 256)], iq3s_grid[q3[13] | ((qh[1] << 3) & 256)],
+                                             iq3s_grid[q3[14] | ((qh[1] << 2) & 256)], iq3s_grid[q3[15] | ((qh[1] << 1) & 256)]};
+            q3 += 16;
+            qh += 2;
+
+            vector signed char vsigns01 = (vector signed char)vec_splats(*(const uint32_t *)&signs[0]);
+            vector signed char vsigns02 = (vector signed char)vec_splats(*(const uint32_t *)&signs[2]);
+            signs += 4;
+
+            vector signed char vsigns0 = vec_perm(vsigns01, vsigns01, mask0);
+            vector signed char vsigns1 = vec_perm(vsigns01, vsigns01, mask1);
+            vector signed char vsigns2 = vec_perm(vsigns02, vsigns02, mask0);
+            vector signed char vsigns3 = vec_perm(vsigns02, vsigns02, mask1);
+
+            vsigns0 = (vector signed char)vec_cmpeq(vec_and(vsigns0, mask2), mask2);
+            vsigns1 = (vector signed char)vec_cmpeq(vec_and(vsigns1, mask2), mask2);
+            vsigns2 = (vector signed char)vec_cmpeq(vec_and(vsigns2, mask2), mask2);
+            vsigns3 = (vector signed char)vec_cmpeq(vec_and(vsigns3, mask2), mask2);
+
+            vector signed char q3x0 = vec_sub(vec_xor(vsigns0, (vector signed char)aux32x4_0), vsigns0);
+            vector signed char q3x1 = vec_sub(vec_xor(vsigns1, (vector signed char)aux32x4_1), vsigns1);
+            vector signed char q3x2 = vec_sub(vec_xor(vsigns2, (vector signed char)aux32x4_2), vsigns2);
+            vector signed char q3x3 = vec_sub(vec_xor(vsigns3, (vector signed char)aux32x4_3), vsigns3);
+
+            vector signed char q8y0 = vec_xl( 0, q8);
+            vector signed char q8y1 = vec_xl(16, q8);
+            vector signed char q8y2 = vec_xl(32, q8);
+            vector signed char q8y3 = vec_xl(48, q8);
+            q8 += 64;
+
+            vector signed short qv0 = vec_add(vec_mule(q3x0, q8y0), vec_mulo(q3x0, q8y0));
+            vector signed short qv1 = vec_add(vec_mule(q3x1, q8y1), vec_mulo(q3x1, q8y1));
+            vector signed short qv2 = vec_add(vec_mule(q3x2, q8y2), vec_mulo(q3x2, q8y2));
+            vector signed short qv3 = vec_add(vec_mule(q3x3, q8y3), vec_mulo(q3x3, q8y3));
+
+            const uint16_t ls0 = (uint16_t)(sc[0] & 0xf);
+            const uint16_t ls1 = (uint16_t)(sc[0] >>  4);
+            sc ++;
+
+            vector signed short vscales01 = (vector signed short)vec_splats((uint16_t)(2*ls0+1));
+            vector signed short vscales23 = (vector signed short)vec_splats((uint16_t)(2*ls1+1));
+
+            vsumi0 = vec_msum(qv0, vscales01, vsumi0);
+            vsumi1 = vec_msum(qv1, vscales01, vsumi1);
+            vsumi2 = vec_msum(qv2, vscales23, vsumi2);
+            vsumi3 = vec_msum(qv3, vscales23, vsumi3);
+        }
+
+        vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
+        vsumf1 = vec_madd(vec_ctf(vsumi1, 0), vd, vsumf1);
+        vsumf2 = vec_madd(vec_ctf(vsumi2, 0), vd, vsumf2);
+        vsumf3 = vec_madd(vec_ctf(vsumi3, 0), vd, vsumf3);
+    }
+
+    vsumf0 = vec_add(vsumf0, vsumf2);
+    vsumf1 = vec_add(vsumf1, vsumf3);
+
+    vsumf0 = vec_add(vsumf0, vsumf1);
+
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4));
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8));
+
+    *s = vec_extract(vsumf0, 0);
+
+#else
+
+    float sumf = 0.f;
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
+        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;
+            const uint32_t ls2 = 2*(x[i].scales[ib32/2] >>  4) + 1;
+            int32_t sumi = 0;
+            for (int l = 0; l < 4; ++l) {
+                const uint8_t * grid1 = (const uint8_t *)(iq3s_grid + (qs[2*l+0] | ((qh[ib32+0] << (8-2*l)) & 256)));
+                const uint8_t * grid2 = (const uint8_t *)(iq3s_grid + (qs[2*l+1] | ((qh[ib32+0] << (7-2*l)) & 256)));
+                for (int j = 0; j < 4; ++j) {
+                    sumi += grid1[j] * q8[j+0] * (signs[l] & kmask_iq2xs[j+0] ? -1 : 1);
+                    sumi += grid2[j] * q8[j+4] * (signs[l] & kmask_iq2xs[j+4] ? -1 : 1);
+                }
+                q8 += 8;
+            }
+            qs += 8;
+            signs += 4;
+            bsum += sumi * ls1;
+            sumi = 0;
+            for (int l = 0; l < 4; ++l) {
+                const uint8_t * grid1 = (const uint8_t *)(iq3s_grid + (qs[2*l+0] | ((qh[ib32+1] << (8-2*l)) & 256)));
+                const uint8_t * grid2 = (const uint8_t *)(iq3s_grid + (qs[2*l+1] | ((qh[ib32+1] << (7-2*l)) & 256)));
+                for (int j = 0; j < 4; ++j) {
+                    sumi += grid1[j] * q8[j+0] * (signs[l] & kmask_iq2xs[j+0] ? -1 : 1);
+                    sumi += grid2[j] * q8[j+4] * (signs[l] & kmask_iq2xs[j+4] ? -1 : 1);
+                }
+                q8 += 8;
+            }
+            qs += 8;
+            signs += 4;
+            bsum += sumi * ls2;
+        }
+        sumf += d * bsum;
+    }
+    *s = sumf;
+#endif
+}
+
+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);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_iq1_s * GGML_RESTRICT x = vx;
+    const block_q8_K  * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+#if defined(__POWER9_VECTOR__)
+    const vector unsigned char v0 = vec_splats((unsigned char)0x0);
+    const vector unsigned short vsign = vec_splats((unsigned short)0x8000);
+
+    vector float vsumf0 = vec_splats(0.0f);
+    vector float vsumf1 = vec_splats(0.0f);
+    vector float vsumf2 = vec_splats(0.0f);
+    vector float vsumf3 = vec_splats(0.0f);
+
+    for (int i = 0; i < nb; ++i) {
+        vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[i].d));
+        vector float vyd = vec_splats(y[i].d);
+        vector float vd = vec_mul(vxd, vyd);
+
+        vector signed int vsumi0 = vec_splats((int32_t)0);
+        vector signed int vsumi1 = vec_splats((int32_t)0);
+        vector signed int vsumi2 = vec_splats((int32_t)0);
+        vector signed int vsumi3 = vec_splats((int32_t)0);
+        vector signed int vsumi8 = vec_splats((int32_t)0);
+
+        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);
+            __builtin_prefetch(qh, 0, 1);
+            __builtin_prefetch(q8, 0, 1);
+
+            vector signed long long aux64x2_0 = {*(const int64_t *)(iq1s_grid + (q1[0] | ((qh[0] << 8) & 0x700))), *(const int64_t *)(iq1s_grid + (q1[1] | ((qh[0] << 5) & 0x700)))};
+            vector signed long long aux64x2_1 = {*(const int64_t *)(iq1s_grid + (q1[2] | ((qh[0] << 2) & 0x700))), *(const int64_t *)(iq1s_grid + (q1[3] | ((qh[0] >> 1) & 0x700)))};
+            vector signed long long aux64x2_2 = {*(const int64_t *)(iq1s_grid + (q1[4] | ((qh[1] << 8) & 0x700))), *(const int64_t *)(iq1s_grid + (q1[5] | ((qh[1] << 5) & 0x700)))};
+            vector signed long long aux64x2_3 = {*(const int64_t *)(iq1s_grid + (q1[6] | ((qh[1] << 2) & 0x700))), *(const int64_t *)(iq1s_grid + (q1[7] | ((qh[1] >> 1) & 0x700)))};
+            q1 += 8;
+
+            vector signed char q1x0 = (vector signed char)aux64x2_0;
+            vector signed char q1x1 = (vector signed char)aux64x2_1;
+            vector signed char q1x2 = (vector signed char)aux64x2_2;
+            vector signed char q1x3 = (vector signed char)aux64x2_3;
+
+            vector signed char q8y0 = vec_xl( 0, q8);
+            vector signed char q8y1 = vec_xl(16, q8);
+            vector signed char q8y2 = vec_xl(32, q8);
+            vector signed char q8y3 = vec_xl(48, q8);
+            q8 += 64;
+
+            vector signed short qv0 = vec_add(vec_mule(q1x0, q8y0), vec_mulo(q1x0, q8y0));
+            vector signed short qv1 = vec_add(vec_mule(q1x1, q8y1), vec_mulo(q1x1, q8y1));
+            vector signed short qv2 = vec_add(vec_mule(q1x2, q8y2), vec_mulo(q1x2, q8y2));
+            vector signed short qv3 = vec_add(vec_mule(q1x3, q8y3), vec_mulo(q1x3, q8y3));
+
+            const uint16_t ls0 = (uint16_t)((qh[0] >> 12) & 7);
+            const uint16_t ls1 = (uint16_t)((qh[1] >> 12) & 7);
+
+            vector signed short vscales01 = (vector signed short)vec_splats((uint16_t)(2*ls0+1));
+            vector signed short vscales23 = (vector signed short)vec_splats((uint16_t)(2*ls1+1));
+            vector signed short vscales = vec_sld(vscales23, vscales01, 8);
+
+            vsumi0 = vec_msum(qv0, vscales01, vsumi0);
+            vsumi1 = vec_msum(qv1, vscales01, vsumi1);
+            vsumi2 = vec_msum(qv2, vscales23, vsumi2);
+            vsumi3 = vec_msum(qv3, vscales23, vsumi3);
+
+            vector signed short q8ysums = vec_xl_len(qs, 8);
+            qs += 4;
+            q8ysums = vec_mergeh(q8ysums, (vector signed short)v0);
+
+            vector signed short qxh = (vector signed short)vec_sld(vec_splats(qh[1]), vec_splats(qh[0]), 8);
+            qh += 2;
+            vector __bool short vsel = vec_cmpge(qxh, (vector signed short)v0);
+
+            vector signed short q8ysum = vec_sel((vector signed short)vec_xor((vector unsigned short)q8ysums, vsign), q8ysums, vsel);
+
+            vsumi8 = vec_add(vec_mule(q8ysum, vscales), vsumi8);
+        }
+
+        vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
+        vsumf1 = vec_madd(vec_ctf(vsumi1, 0), vd, vsumf1);
+        vsumf2 = vec_madd(vec_ctf(vsumi2, 0), vd, vsumf2);
+        vsumf3 = vec_madd(vec_ctf(vsumi3, 0), vd, vsumf3);
+
+        vsumf0 = vec_madd(vec_ctf(vsumi8, 0), vec_mul(vd, vec_splats(IQ1S_DELTA)), vsumf0);
+    }
+
+    vsumf0 = vec_add(vsumf0, vsumf2);
+    vsumf1 = vec_add(vsumf1, vsumf3);
+
+    vsumf0 = vec_add(vsumf0, vsumf1);
+
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4));
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8));
+
+    *s = vec_extract(vsumf0, 0);
+
+#else
+
+    float sumf = 0;
+    for (int i = 0; i < nb; i++) {
+
+        const int8_t   * q8 = y[i].qs;
+        const uint8_t  * qs = x[i].qs;
+        const uint16_t * qh = x[i].qh;
+
+        int sumi = 0, sumi1 = 0;
+        for (int ib = 0; ib < QK_K/32; ++ib) {
+            const int ls = 2*((qh[ib] >> 12) & 7) + 1;
+            const int delta = qh[ib] & 0x8000 ? -1 : 1;
+            int lsum = 0;
+            for (int l = 0; l < 4; ++l) {
+                const int8_t * grid = (const int8_t *)(iq1s_grid + (qs[l] | (((qh[ib] >> 3*l) & 7) << 8)));
+                for (int j = 0; j < 8; ++j) {
+                    lsum += q8[j] * grid[j];
+                }
+                q8 += 8;
+            }
+            sumi  += ls * lsum;
+            sumi1 += ls * delta * (y[i].bsums[2*ib+0] + y[i].bsums[2*ib+1]);
+            qs += 4;
+        }
+
+        sumf += GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d * (sumi + IQ1S_DELTA * sumi1);
+    }
+
+    *s = sumf;
+
+#endif
+}
+
+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);
+    UNUSED(by);
+    UNUSED(bs);
+    assert(n % QK4_NL == 0);
+    static_assert(QK4_NL == QK8_0, "QK4_NL and QK8_0 must be the same");
+
+    const block_iq4_nl * GGML_RESTRICT x = vx;
+    const block_q8_0   * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK4_NL;
+
+    int ib = 0;
+    float sumf = 0;
+
+#if defined(__POWER9_VECTOR__)
+    const vector signed char lowMask = vec_splats((signed char)0xF);
+    const vector signed int v0 = vec_splats((int32_t)0);
+    const vector unsigned char v4 = vec_splats((unsigned char)0x4);
+
+    vector float vsumf0 = vec_splats(0.0f);
+    vector float vsumf1 = vec_splats(0.0f);
+
+    const vector signed char values = vec_xl( 0, kvalues_iq4nl);
+
+#pragma GCC unroll 4
+    for (; ib < nb; ++ib) {
+        __builtin_prefetch(x[ib].qs, 0, 1);
+        __builtin_prefetch(y[ib].qs, 0, 1);
+
+
+        vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[ib].d));
+        vector float vyd = vec_splats(GGML_CPU_FP16_TO_FP32(y[ib].d));
+        vector float vd = vec_mul(vxd, vyd);
+
+        vector signed char qxs = (vector signed char)vec_xl( 0, x[ib].qs);
+        vector signed char q4x0 = vec_and(qxs, lowMask);
+        vector signed char q4x1 = vec_sr(qxs, v4);
+
+        q4x0 = vec_perm(values, values, (vector unsigned char)q4x0);
+        q4x1 = vec_perm(values, values, (vector unsigned char)q4x1);
+
+        vector signed char q8y0 = vec_xl( 0, y[ib].qs);
+        vector signed char q8y1 = vec_xl(16, y[ib].qs);
+
+        vector signed short qv0 = vec_add(vec_mule(q4x0, q8y0), vec_mulo(q4x0, q8y0));
+        vector signed short qv1 = vec_add(vec_mule(q4x1, q8y1), vec_mulo(q4x1, q8y1));
+
+        vector signed int vsumi0 = v0;
+        vector signed int vsumi1 = v0;
+
+        vsumi0 = vec_sum4s(qv0, vsumi0);
+        vsumi1 = vec_sum4s(qv1, vsumi1);
+
+        vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
+        vsumf1 = vec_madd(vec_ctf(vsumi1, 0), vd, vsumf1);
+    }
+
+    vsumf0 = vec_add(vsumf0, vsumf1);
+
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4));
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8));
+
+    sumf = vec_extract(vsumf0, 0);
+
+#endif
+    for (; ib < nb; ++ib) {
+        const float d = GGML_CPU_FP16_TO_FP32(y[ib].d)*GGML_CPU_FP16_TO_FP32(x[ib].d);
+        int sumi1 = 0, sumi2 = 0;
+        for (int j = 0; j < QK4_NL/2; ++j) {
+            sumi1 += y[ib].qs[j+       0] * kvalues_iq4nl[x[ib].qs[j] & 0xf];
+            sumi2 += y[ib].qs[j+QK4_NL/2] * kvalues_iq4nl[x[ib].qs[j] >>  4];
+        }
+        sumf += d * (sumi1 + sumi2);
+    }
+    *s = sumf;
+}
+
+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);
+    UNUSED(by);
+    UNUSED(bs);
+    assert(n % QK_K == 0);
+
+    const block_iq4_xs * GGML_RESTRICT x = vx;
+    const block_q8_K   * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+#if defined(__POWER9_VECTOR__)
+    const vector signed char lowMask = vec_splats((signed char)0xF);
+    const vector int v0 = vec_splats((int32_t)0);
+    const vector unsigned char v4 = vec_splats((unsigned char)0x4);
+
+    vector float vsumf0 = vec_splats(0.0f);
+    vector float vsumf1 = vec_splats(0.0f);
+    vector float vsumf2 = vec_splats(0.0f);
+    vector float vsumf3 = vec_splats(0.0f);
+
+    const vector signed char values = vec_xl( 0, kvalues_iq4nl);
+
+    for (int ibl = 0; ibl < nb; ++ibl) {
+
+        vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[ibl].d));
+        vector float vyd = vec_splats(y[ibl].d);
+        vector float vd = vec_mul(vxd, vyd);
+
+        vector signed int vsumi0 = v0;
+        vector signed int vsumi1 = v0;
+        vector signed int vsumi2 = v0;
+        vector signed int vsumi3 = v0;
+
+        uint16_t h = x[ibl].scales_h;
+
+        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);
+            __builtin_prefetch(q8, 0, 1);
+
+            vector signed char qxs0 = (vector signed char)vec_xl( 0, q4);
+            vector signed char qxs1 = (vector signed char)vec_xl(16, q4);
+            q4 += 32;
+
+            vector signed char q4x00 = (vector signed char)vec_and(qxs0, lowMask);
+            vector signed char q4x01 = (vector signed char)vec_sr(qxs0, v4);
+            vector signed char q4x10 = (vector signed char)vec_and(qxs1, lowMask);
+            vector signed char q4x11 = (vector signed char)vec_sr(qxs1, v4);
+
+            q4x00 = vec_perm(values, values, (vector unsigned char)q4x00);
+            q4x01 = vec_perm(values, values, (vector unsigned char)q4x01);
+            q4x10 = vec_perm(values, values, (vector unsigned char)q4x10);
+            q4x11 = vec_perm(values, values, (vector unsigned char)q4x11);
+
+            vector signed char q8y0 = vec_xl( 0, q8);
+            vector signed char q8y1 = vec_xl(16, q8);
+            vector signed char q8y2 = vec_xl(32, q8);
+            vector signed char q8y3 = vec_xl(48, q8);
+            q8 += 64;
+
+            vector signed short qv0 = vec_add(vec_mule(q4x00, q8y0), vec_mulo(q4x00, q8y0));
+            vector signed short qv1 = vec_add(vec_mule(q4x01, q8y1), vec_mulo(q4x01, q8y1));
+            vector signed short qv2 = vec_add(vec_mule(q4x10, q8y2), vec_mulo(q4x10, q8y2));
+            vector signed short qv3 = vec_add(vec_mule(q4x11, q8y3), vec_mulo(q4x11, q8y3));
+
+            const uint16_t ls0 = (uint16_t)(((sc[0] & 0xf) | ((h << 4) & 0x30)) - 32);
+            const uint16_t ls1 = (uint16_t)(((sc[0] >>  4) | ((h << 2) & 0x30)) - 32);
+            h >>= 4;
+            sc ++;
+
+            vector signed short vscales01 = vec_splats((int16_t)ls0);
+            vector signed short vscales23 = vec_splats((int16_t)ls1);
+
+            vsumi0 = vec_msum(qv0, vscales01, vsumi0);
+            vsumi1 = vec_msum(qv1, vscales01, vsumi1);
+            vsumi2 = vec_msum(qv2, vscales23, vsumi2);
+            vsumi3 = vec_msum(qv3, vscales23, vsumi3);
+        }
+
+        vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
+        vsumf1 = vec_madd(vec_ctf(vsumi1, 0), vd, vsumf1);
+        vsumf2 = vec_madd(vec_ctf(vsumi2, 0), vd, vsumf2);
+        vsumf3 = vec_madd(vec_ctf(vsumi3, 0), vd, vsumf3);
+    }
+
+    vsumf0 = vec_add(vsumf0, vsumf2);
+    vsumf1 = vec_add(vsumf1, vsumf3);
+
+    vsumf0 = vec_add(vsumf0, vsumf1);
+
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4));
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8));
+
+    *s = vec_extract(vsumf0, 0);
+
+#else
+    float sumf = 0;
+    for (int ibl = 0; ibl < nb; ++ibl) {
+        const float d4d8 = GGML_CPU_FP16_TO_FP32(x[ibl].d) * y[ibl].d;
+        uint16_t h = x[ibl].scales_h;
+        const uint8_t * qs = x[ibl].qs;
+        const int8_t  * q8 = y[ibl].qs;
+        for (int ib = 0; ib < QK_K/32; ib += 2) {
+            const uint8_t ls1 = (x[ibl].scales_l[ib/2] & 0xf) | ((h << 4) & 0x30);
+            const uint8_t ls2 = (x[ibl].scales_l[ib/2] >>  4) | ((h << 2) & 0x30);
+            h >>= 4;
+            const float d1 = d4d8*(ls1 - 32);
+            const float d2 = d4d8*(ls2 - 32);
+            int sumi1 = 0, sumi2 = 0;
+            for (int j = 0; j < 16; ++j) {
+                sumi1 += q8[j+ 0] * kvalues_iq4nl[qs[j] & 0xf];
+                sumi2 += q8[j+16] * kvalues_iq4nl[qs[j] >>  4];
+            }
+            sumf += d1 * (sumi1 + sumi2);
+            qs += 16;
+            q8 += 32;
+            sumi1 = sumi2 = 0;
+            for (int j = 0; j < 16; ++j) {
+                sumi1 += q8[j+ 0] * kvalues_iq4nl[qs[j] & 0xf];
+                sumi2 += q8[j+16] * kvalues_iq4nl[qs[j] >>  4];
+            }
+            sumf += d2 * (sumi1 + sumi2);
+            qs += 16;
+            q8 += 32;
+        }
+    }
+    *s = sumf;
+#endif
+}
+
diff --git a/ggml/src/ggml-cpu/arch/riscv/quants.c b/ggml/src/ggml-cpu/arch/riscv/quants.c
new file mode 100644
index 000000000..8b64d8adc
--- /dev/null
+++ b/ggml/src/ggml-cpu/arch/riscv/quants.c
@@ -0,0 +1,2069 @@
+#define GGML_COMMON_IMPL_C
+#include "ggml-common.h"
+#include "ggml-quants.h"
+#include "ggml-impl.h"
+#include "ggml-cpu.h"
+#include "simd-mappings.h"
+
+#include "../../quants.h"
+#include "../../ggml-cpu-impl.h"
+
+#include <math.h>
+#include <string.h>
+#include <assert.h>
+#include <float.h>
+#include <stdlib.h> // for qsort
+#include <stdio.h>  // for GGML_ASSERT
+
+#define GROUP_MAX_EPS 1e-15f
+#define GROUP_MAX_EPS_IQ3_XXS 1e-8f
+#define GROUP_MAX_EPS_IQ2_S 1e-8f
+#define GROUP_MAX_EPS_IQ1_M 1e-7f
+#define GROUP_MAX_EPS_IQ1_S 1e-12f
+
+#define UNUSED GGML_UNUSED
+
+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 * GGML_RESTRICT y = vy;
+
+#if defined(__riscv_v)
+
+    size_t vl = QK8_0;
+
+    for (int i = 0; i < nb; i++) {
+        // load elements
+        vfloat32m8_t v_x   = __riscv_vle32_v_f32m8(x+i*QK8_0, 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_f32m8_f32m1(vfabs, tmp, vl);
+        float amax = __riscv_vfmv_f_s_f32m1_f32(vmax);
+
+        const float d = amax / ((1 << 7) - 1);
+        const float id = d ? 1.0f/d : 0.0f;
+
+        y[i].d = GGML_CPU_FP32_TO_FP16(d);
+
+        vfloat32m8_t x0 = __riscv_vfmul_vf_f32m8(v_x, id, vl);
+
+        // convert to integer
+        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_i8m2(y[i].qs , vs, vl);
+    }
+#else
+    GGML_UNUSED(nb);
+    // scalar
+    quantize_row_q8_0_ref(x, y, k);
+#endif
+}
+
+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 * GGML_RESTRICT y = vy;
+
+#if defined(__riscv_v)
+
+    size_t vl = QK8_1;
+
+    for (int i = 0; i < nb; i++) {
+        // load elements
+        vfloat32m8_t v_x   = __riscv_vle32_v_f32m8(x+i*QK8_1, 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_f32m8_f32m1(vfabs, tmp, vl);
+        float amax = __riscv_vfmv_f_s_f32m1_f32(vmax);
+
+        const float d  = amax / ((1 << 7) - 1);
+        const float id = d ? 1.0f/d : 0.0f;
+
+        y[i].d = GGML_CPU_FP32_TO_FP16(d);
+
+        vfloat32m8_t x0 = __riscv_vfmul_vf_f32m8(v_x, id, vl);
+
+        // convert to integer
+        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_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_i8m2_i16m1(vs, tmp2, vl);
+
+        // set y[i].s
+        int sum = __riscv_vmv_x_s_i16m1_i16(vwrs);
+        y[i].s = GGML_CPU_FP32_TO_FP16(sum*d);
+    }
+
+#else
+    GGML_UNUSED(nb);
+    // scalar
+    quantize_row_q8_1_ref(x, y, k);
+#endif
+}
+
+//===================================== Dot products =================================
+
+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;
+
+    assert(n % qk == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q4_0 * GGML_RESTRICT x = vx;
+    const block_q8_0 * GGML_RESTRICT y = vy;
+
+    int ib = 0;
+    float sumf = 0;
+
+#if defined(__riscv_v)
+    size_t vl = qk / 2;
+
+    for (; ib < nb; ++ib) {
+        // load elements
+        vuint8m1_t tx = __riscv_vle8_v_u8m1(x[ib].qs, 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
+        vuint8m1_t x_a = __riscv_vand_vx_u8m1(tx, 0x0F, vl);
+        vuint8m1_t x_l = __riscv_vsrl_vx_u8m1(tx, 0x04, vl);
+
+        vint8m1_t x_ai = __riscv_vreinterpret_v_u8m1_i8m1(x_a);
+        vint8m1_t x_li = __riscv_vreinterpret_v_u8m1_i8m1(x_l);
+
+        // subtract offset
+        vint8m1_t v0 = __riscv_vsub_vx_i8m1(x_ai, 8, vl);
+        vint8m1_t v1 = __riscv_vsub_vx_i8m1(x_li, 8, 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 vs2 = __riscv_vwredsum_vs_i16m2_i32m1(vec_mul2, vec_zero, vl);
+
+        int sumi = __riscv_vmv_x_s_i32m1_i32(vs2);
+
+        sumf += sumi*GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d);
+    }
+
+#endif
+    for (; ib < nb; ++ib) {
+        int sumi0 = 0;
+        int sumi1 = 0;
+
+        for (int j = 0; j < qk/2; ++j) {
+            const int v0 = (x[ib].qs[j] & 0x0F) - 8;
+            const int v1 = (x[ib].qs[j] >>   4) - 8;
+
+            sumi0 += (v0 * y[ib].qs[j]);
+            sumi1 += (v1 * y[ib].qs[j + qk/2]);
+        }
+
+        int sumi = sumi0 + sumi1;
+        sumf += sumi*GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d);
+    }
+
+    *s = sumf;
+}
+
+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;
+
+    assert(n % qk == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q4_1 * GGML_RESTRICT x = vx;
+    const block_q8_1 * GGML_RESTRICT y = vy;
+
+    int ib = 0;
+    float sumf = 0;
+
+#if defined(__riscv_v)
+    size_t vl = qk / 2;
+
+    for (; ib < nb; ++ib) {
+        // load elements
+        vuint8m1_t tx = __riscv_vle8_v_u8m1(x[ib].qs, 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
+        vuint8m1_t x_a = __riscv_vand_vx_u8m1(tx, 0x0F, vl);
+        vuint8m1_t x_l = __riscv_vsrl_vx_u8m1(tx, 0x04, vl);
+
+        vint8m1_t v0 = __riscv_vreinterpret_v_u8m1_i8m1(x_a);
+        vint8m1_t v1 = __riscv_vreinterpret_v_u8m1_i8m1(x_l);
+
+        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 vs2 = __riscv_vwredsum_vs_i16m2_i32m1(vec_mul2, vec_zero, vl);
+
+        int sumi = __riscv_vmv_x_s_i32m1_i32(vs2);
+
+        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
+    }
+
+#endif
+    for (; ib < nb; ++ib) {
+        int sumi0 = 0;
+        int sumi1 = 0;
+
+        for (int j = 0; j < qk/2; ++j) {
+            const int v0 = (x[ib].qs[j] & 0x0F);
+            const int v1 = (x[ib].qs[j] >>   4);
+
+            sumi0 += (v0 * y[ib].qs[j]);
+            sumi1 += (v1 * y[ib].qs[j + qk/2]);
+        }
+
+        int sumi = sumi0 + sumi1;
+        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
+    }
+
+    *s = sumf;
+}
+
+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;
+
+    int ib = 0;
+    float sumf = 0;
+
+    assert(n % qk == 0);
+    assert(qk == QK5_0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q5_0 * GGML_RESTRICT x = vx;
+    const block_q8_0 * GGML_RESTRICT y = vy;
+
+#if defined(__riscv_v)
+    size_t vl;
+    size_t vlenb = __riscv_vlenb();
+
+    for (; ib < nb; ++ib) {
+        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);
+        }
+
+        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);
+
+        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d) * GGML_CPU_FP16_TO_FP32(y[ib].d)) * sumi;
+    }
+
+#endif
+    for (; ib < nb; ++ib) {
+        uint32_t qh;
+        memcpy(&qh, x[ib].qh, sizeof(qh));
+
+        int sumi0 = 0;
+        int sumi1 = 0;
+
+        for (int j = 0; j < qk/2; ++j) {
+            const uint8_t xh_0 = ((qh & (1u << (j + 0 ))) >> (j + 0 )) << 4;
+            const uint8_t xh_1 = ((qh & (1u << (j + 16))) >> (j + 12));
+
+            const int32_t x0 = (int8_t)(((x[ib].qs[j] & 0x0F) | xh_0) - 16);
+            const int32_t x1 = (int8_t)(((x[ib].qs[j] >>   4) | xh_1) - 16);
+
+            sumi0 += (x0 * y[ib].qs[j]);
+            sumi1 += (x1 * y[ib].qs[j + qk/2]);
+        }
+
+        int sumi = sumi0 + sumi1;
+        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d)) * sumi;
+    }
+
+    *s = sumf;
+}
+
+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;
+
+    int ib = 0;
+    float sumf = 0;
+
+    assert(n % qk == 0);
+    assert(qk == QK5_1);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q5_1 * GGML_RESTRICT x = vx;
+    const block_q8_1 * GGML_RESTRICT y = vy;
+
+#if defined(__riscv_v)
+    size_t vl;
+    size_t vlenb = __riscv_vlenb();
+
+    for (; ib < nb; ++ib) {
+        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);
+        }
+
+        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_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
+    }
+
+#endif
+    for (; ib < nb; ++ib) {
+        uint32_t qh;
+        memcpy(&qh, x[ib].qh, sizeof(qh));
+
+        int sumi0 = 0;
+        int sumi1 = 0;
+
+        for (int j = 0; j < qk/2; ++j) {
+            const uint8_t xh_0 = ((qh >> (j +  0)) << 4) & 0x10;
+            const uint8_t xh_1 = ((qh >> (j + 12))     ) & 0x10;
+
+            const int32_t x0 = (x[ib].qs[j] & 0xF) | xh_0;
+            const int32_t x1 = (x[ib].qs[j] >>  4) | xh_1;
+
+            sumi0 += (x0 * y[ib].qs[j]);
+            sumi1 += (x1 * y[ib].qs[j + qk/2]);
+        }
+
+        int sumi = sumi0 + sumi1;
+        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
+    }
+
+    *s = sumf;
+}
+
+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;
+
+    assert(n % qk == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q8_0 * GGML_RESTRICT x = vx;
+    const block_q8_0 * GGML_RESTRICT y = vy;
+
+    int ib = 0;
+    float sumf = 0;
+
+#if defined(__riscv_v)
+    size_t vl = qk;
+
+    for (; ib < nb; ++ib) {
+        // load elements
+        vint8m2_t bx_0 = __riscv_vle8_v_i8m2(x[ib].qs, vl);
+        vint8m2_t by_0 = __riscv_vle8_v_i8m2(y[ib].qs, 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_i16m4_i32m1(vw_mul, v_zero, vl);
+
+        int sumi = __riscv_vmv_x_s_i32m1_i32(v_sum);
+
+        sumf += sumi*(GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d));
+    }
+
+#endif
+    for (; ib < nb; ++ib) {
+        int sumi = 0;
+
+        for (int j = 0; j < qk; j++) {
+            sumi += x[ib].qs[j]*y[ib].qs[j];
+        }
+
+        sumf += sumi*(GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d));
+    }
+
+    *s = sumf;
+}
+
+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 * GGML_RESTRICT x = vx;
+    const block_q8_K * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+#if defined __riscv_xtheadvector
+
+    float sumf = 0;
+    uint8_t atmp[16];
+
+    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_CPU_FP16_TO_FP32(x[i].d);
+        const float dmin = -y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
+        uint8_t *patmp = atmp;
+        int vsums;
+        int tmp;
+        __asm__ __volatile__(
+            "th.vsetvli zero, %[vl16], e8, m1\n\t"
+            "th.vmv.v.x v8, zero\n\t"
+            "th.vlb.v v1, (%[sc])\n\t"
+            "th.vand.vi v0, v1, 0xF\n\t"
+            "th.vsrl.vi v1, v1, 4\n\t"
+            "th.vsb.v v0, (%[scale])\n\t"
+            "th.vwaddu.vx v16, v1, zero\n\t"
+            "th.vsetvli zero, %[vl16], e16, m2\n\t"
+            "th.vlh.v v2, (%[bsums])\n\t"
+            "th.vwmul.vv v4, v16, v2\n\t"
+            "th.vsetvli zero, %[vl16], e32, m4\n\t"
+            "th.vredsum.vs v8, v4, v8\n\t"
+            "th.vmv.x.s %[vsums], v8"
+            : [tmp] "=&r" (tmp), [vsums] "=&r" (vsums)
+            : [sc] "r" (sc), [scale] "r" (atmp), [bsums] "r" (y[i].bsums)
+            , [vl16] "r" (16)
+            : "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__(
+                "th.vsetvli zero, %[vl32], e8, m2\n\t"
+                "th.vlb.v v0, (%[q2])\n\t"
+                "th.vsrl.vi v2, v0, 2\n\t"
+                "th.vsrl.vi v4, v0, 4\n\t"
+                "th.vsrl.vi v6, v0, 6\n\t"
+                "th.vand.vi v0, v0, 0x3\n\t"
+                "th.vand.vi v2, v2, 0x3\n\t"
+                "th.vand.vi v4, v4, 0x3\n\t"
+                "th.vsetvli zero, %[vl128], e8, m8\n\t"
+                "th.vlb.v v8, (%[q8])\n\t"
+                "th.vsetvli zero, %[vl64], e8, m4\n\t"
+                "th.vwmul.vv v16, v0, v8\n\t"
+                "th.vwmul.vv v24, v4, v12\n\t"
+                "th.vsetvli zero, %[vl16], e16, m2\n\t"
+                "th.vmv.v.x v0, zero\n\t"
+                "th.vwredsum.vs v10, v16, v0\n\t"
+                "th.vwredsum.vs v9, v18, v0\n\t"
+                "th.vwredsum.vs v8, v20, v0\n\t"
+                "th.vwredsum.vs v7, v22, v0\n\t"
+                "th.vwredsum.vs v11, v24, v0\n\t"
+                "th.vwredsum.vs v12, v26, v0\n\t"
+                "th.vwredsum.vs v13, v28, v0\n\t"
+                "th.vwredsum.vs v14, v30, v0\n\t"
+                "li %[tmp], 4\n\t"
+                "th.vsetvli zero, %[tmp], e32, m1\n\t"
+                "th.vslideup.vi v10, v9, 1\n\t"
+                "th.vslideup.vi v8, v7, 1\n\t"
+                "th.vslideup.vi v11, v12, 1\n\t"
+                "th.vslideup.vi v13, v14, 1\n\t"
+                "th.vslideup.vi v10, v8, 2\n\t"
+                "th.vslideup.vi v11, v13, 2\n\t"
+                "li %[tmp], 8\n\t"
+                "th.vsetvli zero, %[tmp], e32, m2\n\t"
+                "th.vlbu.v v12, (%[scale])\n\t"
+                "th.vmul.vv v10, v10, v12\n\t"
+                "th.vredsum.vs v0, v10, v0\n\t"
+                "th.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)
+                , [vl16] "r" (16), [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;
+    }
+
+    *s = sumf;
+
+#elif defined __riscv_v
+
+    float sumf = 0;
+    uint8_t atmp[16];
+
+    const int vector_length = __riscv_vlenb() * 8;
+    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 };
+
+    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_CPU_FP16_TO_FP32(x[i].d);
+            const float dmin = -y[i].d * GGML_CPU_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_CPU_FP16_TO_FP32(x[i].d);
+            const float dmin = -y[i].d * GGML_CPU_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;
+
+#else
+
+    float sumf = 0;
+
+    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;
+
+        int summs = 0;
+        for (int j = 0; j < 16; ++j) {
+            summs += y[i].bsums[j] * (sc[j] >> 4);
+        }
+
+        const float dall = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+        const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
+
+        int isum = 0;
+        int is = 0;
+        int d;
+        for (int k = 0; k < QK_K/128; ++k) {
+            int shift = 0;
+            for (int j = 0; j < 4; ++j) {
+                d = sc[is++] & 0xF;
+                int isuml = 0;
+                for (int l =  0; l < 16; ++l) isuml += q8[l] * ((q2[l] >> shift) & 3);
+                isum += d * isuml;
+                d = sc[is++] & 0xF;
+                isuml = 0;
+                for (int l = 16; l < 32; ++l) isuml += q8[l] * ((q2[l] >> shift) & 3);
+                isum += d * isuml;
+                shift += 2;
+                q8 += 32;
+            }
+            q2 += 32;
+        }
+        sumf += dall * isum - dmin * summs;
+    }
+    *s = sumf;
+#endif
+}
+
+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);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const uint32_t kmask1 = 0x03030303;
+    const uint32_t kmask2 = 0x0f0f0f0f;
+
+    const block_q3_K * GGML_RESTRICT x = vx;
+    const block_q8_K * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+#if defined __riscv_xtheadvector
+
+    uint32_t utmp[4];
+    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;
+
+        int8_t * scale = (int8_t *)utmp;
+        int tmp;
+        __asm__ __volatile__(
+            "li %[tmp], 12\n\t"
+            "th.vsetvli zero, %[tmp], e8, m1\n\t"
+            "th.vlb.v v0, (%[s6b])\n\t"
+            "th.vmv.v.v v2, v0\n\t"
+            "li %[tmp], 2\n\t"
+            "th.vsetvli zero, %[tmp], e64, m1\n\t"
+            "th.vmv.v.x v9, %[sh]\n\t"\
+            "th.vslidedown.vi v1, v0, 1\n\t"
+            "th.vslide1up.vx v8, v9, zero\n\t" // {0, 0, 4, 4}
+            "th.vslideup.vi v0, v2, 1\n\t" // {aux[0], aux[1], aux[0], aux[1]}
+            "li %[tmp], 4\n\t"
+            "th.vsetvli zero, %[tmp], e32, m1\n\t"
+            "th.vid.v v9\n\t"
+            "th.vmv.x.s %[tmp], v1\n\t"
+            "th.vsll.vi v9, v9, 1\n\t" // {0, 2, 4, 6}
+            "th.vmv.v.x v1, %[tmp]\n\t" // {aux[2], aux[2], aux[2], aux[2]}
+            "th.vsrl.vv v4, v1, v9\n\t"
+            "th.vsrl.vv v2, v0, v8\n\t"
+            "th.vand.vx v5, v4, %[kmask1]\n\t"
+            "th.vand.vx v3, v2, %[kmask2]\n\t"
+            "th.vsll.vi v6, v5, 4\n\t"
+            "th.vor.vv v7, v6, v3\n\t"
+            "li %[tmp], 16\n\t"
+            "th.vsetvli zero, %[tmp], e8, m1\n\t"
+            "th.vsub.vx v0, v7, %[c]\n\t"
+            "th.vsb.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"
+        );
+
+        uint8_t m = 1;
+        int isum = 0;
+        for (int j = 0; j < QK_K; j += 128) {
+            __asm__ __volatile__(
+                // fixme: use v0p7 mask layout directly
+                "th.vsetvli zero, %[vl32], e8, m2\n\t"
+                "th.vlb.v v8, (%[q3])\n\t"
+                "th.vsrl.vi v10, v8, 2\n\t"
+                "th.vsrl.vi v12, v8, 4\n\t"
+                "th.vsrl.vi v14, v8, 6\n\t"
+                "th.vand.vi v8, v8, 3\n\t"
+                "th.vand.vi v10, v10, 3\n\t"
+                "th.vand.vi v12, v12, 3\n\t"
+                "th.vlb.v v2, (%[qh])\n\t"
+                "th.vand.vx v4, v2, %[m]\n\t"
+                "slli %[m], %[m], 1\n\t"
+                "th.vmseq.vx v0, v4, zero\n\t"
+                "th.vadd.vi v8, v8, -4, v0.t\n\t"
+                "th.vand.vx v4, v2, %[m]\n\t"
+                "slli %[m], %[m], 1\n\t"
+                "th.vmseq.vx v0, v4, zero\n\t"
+                "th.vadd.vi v10, v10, -4, v0.t\n\t"
+                "th.vand.vx v4, v2, %[m]\n\t"
+                "slli %[m], %[m], 1\n\t"
+                "th.vmseq.vx v0, v4, zero\n\t"
+                "th.vadd.vi v12, v12, -4, v0.t\n\t"
+                "th.vand.vx v4, v2, %[m]\n\t"
+                "slli %[m], %[m], 1\n\t"
+                "th.vmseq.vx v0, v4, zero\n\t"
+                "th.vadd.vi v14, v14, -4, v0.t\n\t"
+                "th.vsetvli zero, %[vl128], e8, m8\n\t"
+                "th.vlb.v v0, (%[q8])\n\t"
+                "th.vsetvli zero, %[vl64], e8, m4\n\t"
+                "th.vwmul.vv v16, v0, v8\n\t"
+                "th.vwmul.vv v24, v4, v12\n\t"
+                "li %[tmp], 16\n\t"
+                "th.vsetvli zero, %[tmp], e16, m2\n\t"
+                "th.vmv.v.x v0, zero\n\t"
+                "th.vwredsum.vs v10, v16, v0\n\t"
+                "th.vwredsum.vs v9, v18, v0\n\t"
+                "th.vwredsum.vs v8, v20, v0\n\t"
+                "th.vwredsum.vs v7, v22, v0\n\t"
+                "th.vwredsum.vs v11, v24, v0\n\t"
+                "th.vwredsum.vs v12, v26, v0\n\t"
+                "th.vwredsum.vs v13, v28, v0\n\t"
+                "th.vwredsum.vs v14, v30, v0\n\t"
+                "li %[tmp], 4\n\t"
+                "th.vsetvli zero, %[tmp], e32, m1\n\t"
+                "th.vslideup.vi v10, v9, 1\n\t"
+                "th.vslideup.vi v8, v7, 1\n\t"
+                "th.vslideup.vi v11, v12, 1\n\t"
+                "th.vslideup.vi v13, v14, 1\n\t"
+                "th.vslideup.vi v10, v8, 2\n\t"
+                "th.vslideup.vi v11, v13, 2\n\t"
+                "li %[tmp], 8\n\t"
+                "th.vsetvli zero, %[tmp], e32, m2\n\t"
+                "th.vlb.v v12, (%[scale])\n\t"
+                "th.vmul.vv v10, v10, v12\n\t"
+                "th.vredsum.vs v0, v10, v0\n\t"
+                "th.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_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
+        sumf += d * isum;
+    }
+
+    *s = sumf;
+
+#elif defined __riscv_v
+
+    uint32_t utmp[4];
+    float sumf = 0;
+    uint32_t aux[3];
+    const int vector_length = __riscv_vlenb() * 8;
+
+    switch (vector_length) {
+    case 256:
+        for (int i = 0; i < nb; ++i) {
+
+            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;
+
+            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;
+
+            vint32m1_t vzero = __riscv_vmv_v_x_i32m1(0, 1);
+            vuint8m1_t vqh = __riscv_vle8_v_u8m1(qh, vl);
+
+            int sum_t = 0;
+
+            for (int j = 0; j < QK_K; j += 128) {
+
+                vl = 32;
+
+                // 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));
+
+                // 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_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;
+
+                // 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;
+
+                // 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);
+
+                sum_t +=  __riscv_vmv_x_s_i32m1_i32(isum3);
+
+                q3 += 32;    q8 += 128;   scale += 8;
+
+            }
+
+            const float d = GGML_CPU_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;
+
+            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"
+            );
+
+            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_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
+            sumf += d * isum;
+        }
+        break;
+    default:
+        assert(false && "Unsupported vector length");
+        break;
+    }
+
+    *s = sumf;
+
+#else
+    // scalar version
+    // This function is written like this so the compiler can manage to vectorize most of it
+    // Using -Ofast, GCC and clang manage to produce code that is within a factor of 2 or so from the
+    // manually vectorized version above. Every other version I tried would run at least 4 times slower.
+    // The ideal situation would be if we could just write the code once, and the compiler would
+    // automatically produce the best possible set of machine instructions, instead of us having to manually
+    // write vectorized versions for AVX, ARM_NEON, etc.
+
+    int8_t  aux8[QK_K];
+    int16_t aux16[8];
+    float   sums [8];
+    int32_t aux32[8];
+    memset(sums, 0, 8*sizeof(float));
+
+    uint32_t auxs[4];
+    const int8_t * scales = (const int8_t*)auxs;
+
+    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;
+        memset(aux32, 0, 8*sizeof(int32_t));
+        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;
+            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
+            a += 32; m <<= 1;
+            for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 2) & 3;
+            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
+            a += 32; m <<= 1;
+            for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 4) & 3;
+            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
+            a += 32; m <<= 1;
+            for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 6) & 3;
+            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
+            a += 32; m <<= 1;
+            q3 += 32;
+        }
+        a = aux8;
+
+        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);
+        for (int j = 0; j < QK_K/16; ++j) {
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
+            q8 += 8; a += 8;
+        }
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
+        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
+    }
+    for (int l = 0; l < 8; ++l) sumf += sums[l];
+    *s = sumf;
+
+#endif
+
+}
+
+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);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q4_K * GGML_RESTRICT x = vx;
+    const block_q8_K * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+    static const uint32_t kmask1 = 0x3f3f3f3f;
+    static const uint32_t kmask2 = 0x0f0f0f0f;
+    static const uint32_t kmask3 = 0x03030303;
+
+    uint32_t utmp[4];
+
+#if defined __riscv_xtheadvector
+
+    const uint8_t * scales = (const uint8_t*)&utmp[0];
+    const uint8_t * mins   = (const uint8_t*)&utmp[2];
+
+    float sumf = 0;
+
+    for (int i = 0; i < nb; ++i) {
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+        const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
+
+        int tmp, tmp2, sumi;
+        __asm__ __volatile__(
+            "li %[t1], 12\n\t"
+            "th.vsetvli zero, %[t1], e8, m1\n\t"
+            "th.vlb.v v1, (%[s6b])\n\t" // {aux[0], aux[1], aux[2]}
+            "li %[t1], 4\n\t"
+            "th.vsetvli zero, %[t1], e32, m1\n\t"
+            "th.vslidedown.vi v2, v1, 2\n\t"
+            "th.vmv.v.v v3, v2\n\t"
+            "th.vslideup.vi v2, v3, 1\n\t" // {aux[2], aux[2]}
+            "li %[t1], 2\n\t"
+            "th.vsetvli zero, %[t1], e32, m1\n\t"
+            "th.vmv.v.i v4, 4\n\t"
+            "th.vand.vx v8, v1, %[kmask1]\n\t"
+            "th.vslide1up.vx v5, v4, zero\n\t" // {0, 4}
+            "th.vsrl.vi v6, v1, 6\n\t"
+            "th.vsrl.vv v7, v2, v5\n\t"
+            "th.vand.vx v0, v6, %[kmask3]\n\t"
+            "th.vand.vx v2, v7, %[kmask2]\n\t"
+            "th.vsll.vi v6, v0, 4\n\t"
+            "li %[t2], 8\n\t"
+            "addi %[t1], %[utmp], 4\n\t"
+            "th.vor.vv v1, v6, v2\n\t"
+            "th.vssw.v v8, (%[utmp]), %[t2]\n\t"
+            "th.vssw.v v1, (%[t1]), %[t2]\n\t"
+            "th.vsetvli zero, zero, e32, m2\n\t" // vl == 8
+            "th.vlw.v v2, (%[bsums])\n\t"
+            "th.vsetvli zero, %[t2], e16, m1\n\t"
+            "th.vnsrl.vi v0, v2, 0\n\t"
+            "th.vnsrl.vi v1, v2, 16\n\t"
+            "th.vadd.vv v2, v0, v1\n\t"
+            "th.vlbu.v v4, (%[mins])\n\t"
+            "th.vwmul.vv v6, v4, v2\n\t"
+            "th.vmv.v.x v0, zero\n\t"
+            "th.vsetvli zero, %[t2], e32, m2\n\t"
+            "th.vredsum.vs v0, v6, v0\n\t"
+            "th.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__(
+                "th.vsetvli zero, %[vl128], e8, m8\n\t"
+                "th.vlb.v v8, (%[q8])\n\t"
+                "th.vsetvli zero, %[vl64], e8, m4\n\t"
+                "th.vlb.v v0, (%[q4])\n\t"
+                "th.vsrl.vi v4, v0, 4\n\t"
+                "th.vand.vi v0, v0, 0xF\n\t"
+                "th.vsetvli zero, %[vl32], e8, m2\n\t"
+                "th.vwmul.vv v28, v6, v14\n\t"
+                "th.vwmul.vv v20, v4, v10\n\t"
+                "th.vwmul.vv v24, v2, v12\n\t"
+                "th.vwmul.vv v16, v0, v8\n\t"
+                "li %[tmp], 4\n\t"
+                "th.vsetvli zero, %[tmp], e32, m1\n\t"
+                "th.vlbu.v v1, (%[scale])\n\t"
+                "th.vmv.v.x v0, zero\n\t"
+                "th.vsetvli zero, %[vl32], e16, m4\n\t"
+                "th.vwredsum.vs v6, v24, v0\n\t"
+                "th.vwredsum.vs v7, v28, v0\n\t"
+                "th.vwredsum.vs v4, v16, v0\n\t"
+                "th.vwredsum.vs v5, v20, v0\n\t"
+                "th.vsetvli zero, %[tmp], e32, m1\n\t"
+                "th.vslideup.vi v6, v7, 1\n\t"
+                "th.vslideup.vi v4, v5, 1\n\t"
+                "th.vslideup.vi v4, v6, 2\n\t"
+                "th.vmul.vv v8, v4, v1\n\t"
+                "th.vredsum.vs v0, v8, v0\n\t"
+                "th.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;
+
+    }
+
+    *s = sumf;
+
+#elif defined __riscv_v
+
+    const uint8_t * scales = (const uint8_t*)&utmp[0];
+    const uint8_t * mins   = (const uint8_t*)&utmp[2];
+
+    float sumf = 0;
+    const int vector_length = __riscv_vlenb() * 8;
+
+    switch (vector_length) {
+    case 256:
+        for (int i = 0; i < nb; ++i) {
+
+            size_t vl = 8;
+
+            const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+            const float dmin = y[i].d * GGML_CPU_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);
+
+            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);
+
+            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 * GGML_RESTRICT q4 = x[i].qs;
+            const int8_t  * GGML_RESTRICT q8 = y[i].qs;
+
+            vl = 32;
+
+            int32_t sum_1 = 0;
+            int32_t sum_2 = 0;
+
+            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);
+
+                // 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];
+
+                // 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];
+
+                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_CPU_FP16_TO_FP32(x[i].d);
+            const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
+
+            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;
+
+#else
+
+    const uint8_t * scales = (const uint8_t*)&utmp[0];
+    const uint8_t * mins   = (const uint8_t*)&utmp[2];
+
+    int8_t  aux8[QK_K];
+    int16_t aux16[8];
+    float   sums [8];
+    int32_t aux32[8];
+    memset(sums, 0, 8*sizeof(float));
+
+    float sumf = 0;
+    for (int i = 0; i < nb; ++i) {
+        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 * 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;
+            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l]  >> 4);
+            a += 32; q4 += 32;
+        }
+        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;
+
+        int sumi = 0;
+        for (int j = 0; j < QK_K/16; ++j) sumi += y[i].bsums[j] * mins[j/2];
+        a = aux8;
+        int is = 0;
+        for (int j = 0; j < QK_K/32; ++j) {
+            int32_t scale = scales[is++];
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+        }
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
+        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
+        const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
+        sumf -= dmin * sumi;
+    }
+    for (int l = 0; l < 8; ++l) sumf += sums[l];
+    *s = sumf;
+#endif
+}
+
+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);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q5_K * GGML_RESTRICT x = vx;
+    const block_q8_K * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+    static const uint32_t kmask1 = 0x3f3f3f3f;
+    static const uint32_t kmask2 = 0x0f0f0f0f;
+    static const uint32_t kmask3 = 0x03030303;
+
+    uint32_t utmp[4];
+
+#if defined __riscv_v
+
+    const uint8_t * scales = (const uint8_t*)&utmp[0];
+    const uint8_t * mins   = (const uint8_t*)&utmp[2];
+
+    float sumf = 0;
+    float sums = 0.0;
+
+    size_t vl;
+
+    for (int i = 0; i < nb; ++i) {
+
+        vl = 8;
+
+        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_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
+
+        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);
+        const uint32_t uaux = utmp[1] & kmask1;
+        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
+        utmp[2] = uaux;
+        utmp[0] &= kmask1;
+
+        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_i32m2_i32m1(prod, __riscv_vmv_v_x_i32m1(0, 1), vl);
+        sumf -= dmin * __riscv_vmv_x_s_i32m1_i32(sumi);
+
+        vl = 32;
+        int32_t aux32 = 0;
+        int is = 0;
+
+        uint8_t m = 1;
+        vint32m1_t vzero = __riscv_vmv_v_x_i32m1(0, 1);
+        vuint8m2_t vqh = __riscv_vle8_v_u8m2(hm, vl);
+
+        for (int j = 0; j < QK_K/64; ++j) {
+            // load Q5 and Q8
+            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
+            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;
+
+            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;
+
+            vint16m4_t v0 = __riscv_vwmul_vv_i16m4(q5_m1, q8_y1, vl);
+            vint16m4_t v1 = __riscv_vwmul_vv_i16m4(q5_m2, q8_y2, 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_i32m8_i32m1(vs1, vzero, vl);
+            vint32m1_t vacc2 = __riscv_vredsum_vs_i32m8_i32m1(vs2, vacc1, vl);
+
+            aux32 += __riscv_vmv_x_s_i32m1_i32(vacc2);
+            q5 += 32;    q8 += 64;
+
+        }
+
+        sums += aux32 * d;
+
+    }
+
+    *s = sumf+sums;
+
+#else
+
+    const uint8_t * scales = (const uint8_t*)&utmp[0];
+    const uint8_t * mins   = (const uint8_t*)&utmp[2];
+
+    int8_t  aux8[QK_K];
+    int16_t aux16[8];
+    float   sums [8];
+    int32_t aux32[8];
+    memset(sums, 0, 8*sizeof(float));
+
+    float sumf = 0;
+    for (int i = 0; i < nb; ++i) {
+        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 * 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);
+            for (int l = 0; l < 32; ++l) a[l] += (hm[l] & m ? 16 : 0);
+            a += 32; m <<= 1;
+            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l]  >> 4);
+            for (int l = 0; l < 32; ++l) a[l] += (hm[l] & m ? 16 : 0);
+            a += 32; m <<= 1;
+            q4 += 32;
+        }
+        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;
+
+        int sumi = 0;
+        for (int j = 0; j < QK_K/16; ++j) sumi += y[i].bsums[j] * mins[j/2];
+        a = aux8;
+        int is = 0;
+        for (int j = 0; j < QK_K/32; ++j) {
+            int32_t scale = scales[is++];
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+        }
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
+        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
+        const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
+        sumf -= dmin * sumi;
+    }
+    for (int l = 0; l < 8; ++l) sumf += sums[l];
+    *s = sumf;
+#endif
+}
+
+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);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q6_K * GGML_RESTRICT x = vx;
+    const block_q8_K * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+#if defined __riscv_xtheadvector
+
+    float sumf = 0;
+
+    for (int i = 0; i < nb; ++i) {
+
+        const float d = GGML_CPU_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__(
+                "th.vsetvli zero, %[vl32], e8, m2\n\t" // vl == 32
+                "th.vlb.v v4, (%[qh])\n\t"
+                "th.vsll.vi v0, v4, 4\n\t"
+                "th.vsll.vi v2, v4, 2\n\t"
+                "th.vsrl.vi v6, v4, 2\n\t"
+                "th.vsetvli zero, %[vl64], e8, m4\n\t" // vl == 64
+                "th.vlb.v v8, (%[q6])\n\t"
+                "th.vsrl.vi v12, v8, 4\n\t"
+                "th.vand.vi v8, v8, 0xF\n\t"
+                "th.vsetvli zero, %[vl128], e8, m8\n\t" // vl == 128
+                "th.vand.vx v0, v0, %[mask]\n\t"
+                "th.vor.vv v8, v8, v0\n\t"
+                "th.vlb.v v0, (%[q8])\n\t"
+                "th.vsub.vx v8, v8, %[vl32]\n\t"
+                "th.vsetvli zero, %[vl64], e8, m4\n\t" // vl == 64
+                "th.vwmul.vv v16, v0, v8\n\t"
+                "th.vwmul.vv v24, v4, v12\n\t"
+                "li %[t0], 16\n\t"
+                "th.vsetvli zero, %[t0], e16, m2\n\t" // vl == 16
+                "th.vmv.v.x v0, zero\n\t"
+                "th.vwredsum.vs v10, v16, v0\n\t"
+                "th.vwredsum.vs v9, v18, v0\n\t"
+                "th.vwredsum.vs v8, v20, v0\n\t"
+                "th.vwredsum.vs v7, v22, v0\n\t"
+                "th.vwredsum.vs v11, v24, v0\n\t"
+                "th.vwredsum.vs v12, v26, v0\n\t"
+                "th.vwredsum.vs v13, v28, v0\n\t"
+                "th.vwredsum.vs v14, v30, v0\n\t"
+                "li %[t0], 4\n\t"
+                "th.vsetvli zero, %[t0], e32, m1\n\t" // vl == 4
+                "th.vslideup.vi v10, v9, 1\n\t"
+                "th.vslideup.vi v8, v7, 1\n\t"
+                "th.vslideup.vi v11, v12, 1\n\t"
+                "th.vslideup.vi v13, v14, 1\n\t"
+                "th.vslideup.vi v10, v8, 2\n\t"
+                "th.vslideup.vi v11, v13, 2\n\t"
+                "li %[t0], 8\n\t"
+                "th.vsetvli zero, %[t0], e32, m2\n\t" // vl == 8
+                "th.vlb.v v4, (%[scale])\n\t"
+                "th.vmul.vv v2, v4, v10\n\t"
+                "th.vredsum.vs v0, v2, v0\n\t"
+                "th.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;
+
+    }
+
+    *s = sumf;
+
+#elif defined __riscv_v
+
+    float sumf = 0;
+    const int vector_length = __riscv_vlenb() * 8;
+
+    switch (vector_length) {
+    case 256:
+        for (int i = 0; i < nb; ++i) {
+
+            const float d = GGML_CPU_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_CPU_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;
+
+#else
+
+    int8_t  aux8[QK_K];
+    int16_t aux16[8];
+    float   sums [8];
+    int32_t aux32[8];
+    memset(sums, 0, 8*sizeof(float));
+
+    float sumf = 0;
+    for (int i = 0; i < nb; ++i) {
+        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 * 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;
+                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;
+        }
+        a = aux8;
+        int is = 0;
+        for (int j = 0; j < QK_K/16; ++j) {
+            int scale = x[i].scales[is++];
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+        }
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
+        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
+    }
+    for (int l = 0; l < 8; ++l) sumf += sums[l];
+    *s = sumf;
+#endif
+}
+
diff --git a/ggml/src/ggml-cpu/arch/riscv/repack.cpp b/ggml/src/ggml-cpu/arch/riscv/repack.cpp
new file mode 100644
index 000000000..45c91a694
--- /dev/null
+++ b/ggml/src/ggml-cpu/arch/riscv/repack.cpp
@@ -0,0 +1,397 @@
+#define GGML_COMMON_IMPL_CPP
+#define GGML_COMMON_DECL_CPP
+#include "ggml-common.h"
+#include "ggml-backend-impl.h"
+
+#include "ggml-impl.h"
+#include "ggml-cpu.h"
+#include "ggml-cpu-impl.h"
+#include "simd-mappings.h"
+#include "traits.h"
+
+#include <cmath>
+#include <cstring>
+#include <cassert>
+#include <cstdlib> // for qsort
+#include <cstdio>  // for GGML_ASSERT
+
+#define GGML_CPU_CLANG_WORKAROUND
+#include "../../repack.h"
+
+#if defined(__GNUC__)
+#pragma GCC diagnostic ignored "-Woverlength-strings"
+#endif
+
+#define UNUSED GGML_UNUSED
+
+void ggml_gemv_q4_0_8x8_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;
+    const int ncols_interleaved = 8;
+    const int blocklen = 8;
+
+    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 __riscv_v
+    if (__riscv_vlenb() >= QK4_0) {
+        const size_t vl = QK4_0;
+
+        const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
+        for (int x = 0; x < nc / ncols_interleaved; x++) {
+            const block_q4_0x8 * b_ptr = (const block_q4_0x8 *) vx + (x * nb);
+
+            vfloat32m1_t sumf = __riscv_vfmv_v_f_f32m1(0.0, vl / 4);
+            for (int l = 0; l < nb; l++) {
+                const int64_t a0 = *(const int64_t *)&a_ptr[l].qs[0];
+                const int64_t a1 = *(const int64_t *)&a_ptr[l].qs[8];
+                const int64_t a2 = *(const int64_t *)&a_ptr[l].qs[16];
+                const int64_t a3 = *(const int64_t *)&a_ptr[l].qs[24];
+                __asm__ __volatile__("" ::: "memory"); // prevent gcc from emitting fused vlse64, violating alignment constraints
+                const vint8m2_t lhs_0_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(a0, vl / 4));
+                const vint8m2_t lhs_1_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(a1, vl / 4));
+                const vint8m2_t lhs_2_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(a2, vl / 4));
+                const vint8m2_t lhs_3_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(a3, vl / 4));
+
+                const vint8m4_t rhs_raw_vec = __riscv_vle8_v_i8m4((const int8_t *)b_ptr[l].qs, vl * 4);
+                const vint8m4_t rhs_vec_lo = __riscv_vsra_vx_i8m4(__riscv_vsll_vx_i8m4(rhs_raw_vec, 4, vl * 4), 4, vl * 4);
+                const vint8m4_t rhs_vec_hi = __riscv_vsra_vx_i8m4(rhs_raw_vec, 4, vl * 4);
+                const vint8m2_t rhs_vec_lo_0 = __riscv_vget_v_i8m4_i8m2(rhs_vec_lo, 0);
+                const vint8m2_t rhs_vec_lo_1 = __riscv_vget_v_i8m4_i8m2(rhs_vec_lo, 1);
+                const vint8m2_t rhs_vec_hi_0 = __riscv_vget_v_i8m4_i8m2(rhs_vec_hi, 0);
+                const vint8m2_t rhs_vec_hi_1 = __riscv_vget_v_i8m4_i8m2(rhs_vec_hi, 1);
+
+                const vint16m4_t sumi_lo_0 = __riscv_vwmul_vv_i16m4(rhs_vec_lo_0, lhs_0_8, vl * 2);
+                const vint16m4_t sumi_lo_1 = __riscv_vwmacc_vv_i16m4(sumi_lo_0, rhs_vec_lo_1, lhs_1_8, vl * 2);
+                const vint16m4_t sumi_hi_0 = __riscv_vwmacc_vv_i16m4(sumi_lo_1, rhs_vec_hi_0, lhs_2_8, vl * 2);
+                const vint16m4_t sumi_hi_m = __riscv_vwmacc_vv_i16m4(sumi_hi_0, rhs_vec_hi_1, lhs_3_8, vl * 2);
+
+                const vuint32m4_t sumi_i32 = __riscv_vreinterpret_v_i32m4_u32m4(__riscv_vreinterpret_v_i16m4_i32m4(sumi_hi_m));
+                const vuint16m2_t sumi_h2_0 = __riscv_vnsrl_wx_u16m2(sumi_i32, 0, vl);
+                const vuint16m2_t sumi_h2_1 = __riscv_vnsrl_wx_u16m2(sumi_i32, 16, vl);
+                const vuint16m2_t sumi_h2 = __riscv_vadd_vv_u16m2(sumi_h2_0, sumi_h2_1, vl);
+                const vuint32m2_t sumi_h2_i32 = __riscv_vreinterpret_v_u16m2_u32m2(sumi_h2);
+                const vuint16m1_t sumi_h4_0 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 0, vl / 2);
+                const vuint16m1_t sumi_h4_1 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 16, vl / 2);
+                const vuint16m1_t sumi_h4 = __riscv_vadd_vv_u16m1(sumi_h4_0, sumi_h4_1, vl / 2);
+                const vuint32m1_t sumi_h4_i32 = __riscv_vreinterpret_v_u16m1_u32m1(sumi_h4);
+                const vint16mf2_t sumi_h8_0 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 0, vl / 4));
+                const vint16mf2_t sumi_h8_1 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 16, vl / 4));
+                const vint32m1_t sumi_h8 = __riscv_vwadd_vv_i32m1(sumi_h8_0, sumi_h8_1, vl / 4);
+                const vfloat32m1_t facc = __riscv_vfcvt_f_x_v_f32m1(sumi_h8, vl / 4);
+
+                // vector version needs Zvfhmin extension
+                const float a_scale = GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
+                const float b_scales[8] = {
+                    GGML_CPU_FP16_TO_FP32(b_ptr[l].d[0]),
+                    GGML_CPU_FP16_TO_FP32(b_ptr[l].d[1]),
+                    GGML_CPU_FP16_TO_FP32(b_ptr[l].d[2]),
+                    GGML_CPU_FP16_TO_FP32(b_ptr[l].d[3]),
+                    GGML_CPU_FP16_TO_FP32(b_ptr[l].d[4]),
+                    GGML_CPU_FP16_TO_FP32(b_ptr[l].d[5]),
+                    GGML_CPU_FP16_TO_FP32(b_ptr[l].d[6]),
+                    GGML_CPU_FP16_TO_FP32(b_ptr[l].d[7])
+                };
+                const vfloat32m1_t b_scales_vec = __riscv_vle32_v_f32m1(b_scales, vl / 4);
+                const vfloat32m1_t tmp1 = __riscv_vfmul_vf_f32m1(facc, a_scale, vl / 4);
+                sumf = __riscv_vfmacc_vv_f32m1(sumf, tmp1, b_scales_vec, vl / 4);
+            }
+            __riscv_vse32_v_f32m1(s + x * ncols_interleaved, sumf, vl / 4);
+        }
+        return;
+    }
+
+#endif
+    {
+        float sumf[8];
+        int sumi;
+
+        const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
+        for (int x = 0; x < nc / ncols_interleaved; x++) {
+            const block_q4_0x8 * b_ptr = (const block_q4_0x8 *) vx + (x * nb);
+
+            for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0;
+            for (int l = 0; l < nb; l++) {
+                for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                    for (int j = 0; j < ncols_interleaved; j++) {
+                        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] << 4);
+                            const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
+                            sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4;
+                        }
+                        sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
+                    }
+                }
+            }
+            for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j];
+        }
+    }
+}
+
+void ggml_gemm_q4_0_8x8_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;
+    const int ncols_interleaved = 8;
+    const int blocklen = 8;
+
+    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 __riscv_v
+    if (__riscv_vlenb() >= QK4_0) {
+        const size_t vl = QK4_0;
+
+        for (int y = 0; y < nr / 4; y++) {
+            const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
+            for (int x = 0; x < nc / ncols_interleaved; x++) {
+                const block_q4_0x8 * b_ptr = (const block_q4_0x8 *) vx + (x * nb);
+                vfloat32m1_t sumf0 = __riscv_vfmv_v_f_f32m1(0.0, vl / 4);
+                vfloat32m1_t sumf1 = __riscv_vfmv_v_f_f32m1(0.0, vl / 4);
+                vfloat32m1_t sumf2 = __riscv_vfmv_v_f_f32m1(0.0, vl / 4);
+                vfloat32m1_t sumf3 = __riscv_vfmv_v_f_f32m1(0.0, vl / 4);
+                for (int l = 0; l < nb; l++) {
+                    const vint8m4_t rhs_raw_vec = __riscv_vle8_v_i8m4((const int8_t *)b_ptr[l].qs, vl * 4);
+                    const vint8m4_t rhs_vec_lo = __riscv_vsra_vx_i8m4(__riscv_vsll_vx_i8m4(rhs_raw_vec, 4, vl * 4), 4, vl * 4);
+                    const vint8m4_t rhs_vec_hi = __riscv_vsra_vx_i8m4(rhs_raw_vec, 4, vl * 4);
+                    const vint8m2_t rhs_vec_lo_0 = __riscv_vget_v_i8m4_i8m2(rhs_vec_lo, 0);
+                    const vint8m2_t rhs_vec_lo_1 = __riscv_vget_v_i8m4_i8m2(rhs_vec_lo, 1);
+                    const vint8m2_t rhs_vec_hi_0 = __riscv_vget_v_i8m4_i8m2(rhs_vec_hi, 0);
+                    const vint8m2_t rhs_vec_hi_1 = __riscv_vget_v_i8m4_i8m2(rhs_vec_hi, 1);
+
+                    // vector version needs Zvfhmin extension
+                    const float a_scales[4] = {
+                        GGML_CPU_FP16_TO_FP32(a_ptr[l].d[0]),
+                        GGML_CPU_FP16_TO_FP32(a_ptr[l].d[1]),
+                        GGML_CPU_FP16_TO_FP32(a_ptr[l].d[2]),
+                        GGML_CPU_FP16_TO_FP32(a_ptr[l].d[3])
+                    };
+                    const float b_scales[8] = {
+                        GGML_CPU_FP16_TO_FP32(b_ptr[l].d[0]),
+                        GGML_CPU_FP16_TO_FP32(b_ptr[l].d[1]),
+                        GGML_CPU_FP16_TO_FP32(b_ptr[l].d[2]),
+                        GGML_CPU_FP16_TO_FP32(b_ptr[l].d[3]),
+                        GGML_CPU_FP16_TO_FP32(b_ptr[l].d[4]),
+                        GGML_CPU_FP16_TO_FP32(b_ptr[l].d[5]),
+                        GGML_CPU_FP16_TO_FP32(b_ptr[l].d[6]),
+                        GGML_CPU_FP16_TO_FP32(b_ptr[l].d[7])
+                    };
+                    const vfloat32m1_t b_scales_vec = __riscv_vle32_v_f32m1(b_scales, vl / 4);
+
+                    const int64_t A0 = *(const int64_t *)&a_ptr[l].qs[0];
+                    const int64_t A4 = *(const int64_t *)&a_ptr[l].qs[32];
+                    const int64_t A8 = *(const int64_t *)&a_ptr[l].qs[64];
+                    const int64_t Ac = *(const int64_t *)&a_ptr[l].qs[96];
+                    __asm__ __volatile__("" ::: "memory"); // prevent gcc from emitting fused vlse64, violating alignment
+                    vint16m4_t sumi_l0;
+                    {
+                        const vint8m2_t lhs_0_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A0, vl / 4));
+                        const vint8m2_t lhs_1_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A4, vl / 4));
+                        const vint8m2_t lhs_2_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A8, vl / 4));
+                        const vint8m2_t lhs_3_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(Ac, vl / 4));
+                        const vint16m4_t sumi_lo_0 = __riscv_vwmul_vv_i16m4(rhs_vec_lo_0, lhs_0_8, vl * 2);
+                        const vint16m4_t sumi_lo_1 = __riscv_vwmacc_vv_i16m4(sumi_lo_0, rhs_vec_lo_1, lhs_1_8, vl * 2);
+                        const vint16m4_t sumi_hi_0 = __riscv_vwmacc_vv_i16m4(sumi_lo_1, rhs_vec_hi_0, lhs_2_8, vl * 2);
+                        const vint16m4_t sumi_hi_m = __riscv_vwmacc_vv_i16m4(sumi_hi_0, rhs_vec_hi_1, lhs_3_8, vl * 2);
+
+                        sumi_l0 = sumi_hi_m;
+                    }
+
+                    {
+                        const vuint32m4_t sumi_i32 = __riscv_vreinterpret_v_i32m4_u32m4(__riscv_vreinterpret_v_i16m4_i32m4(sumi_l0));
+                        const vuint16m2_t sumi_h2_0 = __riscv_vnsrl_wx_u16m2(sumi_i32, 0, vl);
+                        const vuint16m2_t sumi_h2_1 = __riscv_vnsrl_wx_u16m2(sumi_i32, 16, vl);
+                        const vuint16m2_t sumi_h2 = __riscv_vadd_vv_u16m2(sumi_h2_0, sumi_h2_1, vl);
+                        const vuint32m2_t sumi_h2_i32 = __riscv_vreinterpret_v_u16m2_u32m2(sumi_h2);
+                        const vuint16m1_t sumi_h4_0 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 0, vl / 2);
+                        const vuint16m1_t sumi_h4_1 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 16, vl / 2);
+                        const vuint16m1_t sumi_h4 = __riscv_vadd_vv_u16m1(sumi_h4_0, sumi_h4_1, vl / 2);
+                        const vuint32m1_t sumi_h4_i32 = __riscv_vreinterpret_v_u16m1_u32m1(sumi_h4);
+                        const vint16mf2_t sumi_h8_0 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 0, vl / 4));
+                        const vint16mf2_t sumi_h8_1 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 16, vl / 4));
+                        const vint32m1_t sumi_h8 = __riscv_vwadd_vv_i32m1(sumi_h8_0, sumi_h8_1, vl / 4);
+                        const vfloat32m1_t facc = __riscv_vfcvt_f_x_v_f32m1(sumi_h8, vl / 4);
+
+                        const vfloat32m1_t tmp1 = __riscv_vfmul_vf_f32m1(facc, a_scales[0], vl / 4);
+                        sumf0 = __riscv_vfmacc_vv_f32m1(sumf0, tmp1, b_scales_vec, vl / 4);
+                    }
+
+                    const int64_t A1 = *(const int64_t *)&a_ptr[l].qs[8];
+                    const int64_t A5 = *(const int64_t *)&a_ptr[l].qs[40];
+                    const int64_t A9 = *(const int64_t *)&a_ptr[l].qs[72];
+                    const int64_t Ad = *(const int64_t *)&a_ptr[l].qs[104];
+                    __asm__ __volatile__("" ::: "memory"); // prevent gcc from emitting fused vlse64, violating alignment
+                    vint16m4_t sumi_l1;
+                    {
+                        const vint8m2_t lhs_0_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A1, vl / 4));
+                        const vint8m2_t lhs_1_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A5, vl / 4));
+                        const vint8m2_t lhs_2_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A9, vl / 4));
+                        const vint8m2_t lhs_3_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(Ad, vl / 4));
+                        const vint16m4_t sumi_lo_0 = __riscv_vwmul_vv_i16m4(rhs_vec_lo_0, lhs_0_8, vl * 2);
+                        const vint16m4_t sumi_lo_1 = __riscv_vwmacc_vv_i16m4(sumi_lo_0, rhs_vec_lo_1, lhs_1_8, vl * 2);
+                        const vint16m4_t sumi_hi_0 = __riscv_vwmacc_vv_i16m4(sumi_lo_1, rhs_vec_hi_0, lhs_2_8, vl * 2);
+                        const vint16m4_t sumi_hi_m = __riscv_vwmacc_vv_i16m4(sumi_hi_0, rhs_vec_hi_1, lhs_3_8, vl * 2);
+
+                        sumi_l1 = sumi_hi_m;
+                    }
+
+                    {
+                        const vuint32m4_t sumi_i32 = __riscv_vreinterpret_v_i32m4_u32m4(__riscv_vreinterpret_v_i16m4_i32m4(sumi_l1));
+                        const vuint16m2_t sumi_h2_0 = __riscv_vnsrl_wx_u16m2(sumi_i32, 0, vl);
+                        const vuint16m2_t sumi_h2_1 = __riscv_vnsrl_wx_u16m2(sumi_i32, 16, vl);
+                        const vuint16m2_t sumi_h2 = __riscv_vadd_vv_u16m2(sumi_h2_0, sumi_h2_1, vl);
+                        const vuint32m2_t sumi_h2_i32 = __riscv_vreinterpret_v_u16m2_u32m2(sumi_h2);
+                        const vuint16m1_t sumi_h4_0 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 0, vl / 2);
+                        const vuint16m1_t sumi_h4_1 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 16, vl / 2);
+                        const vuint16m1_t sumi_h4 = __riscv_vadd_vv_u16m1(sumi_h4_0, sumi_h4_1, vl / 2);
+                        const vuint32m1_t sumi_h4_i32 = __riscv_vreinterpret_v_u16m1_u32m1(sumi_h4);
+                        const vint16mf2_t sumi_h8_0 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 0, vl / 4));
+                        const vint16mf2_t sumi_h8_1 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 16, vl / 4));
+                        const vint32m1_t sumi_h8 = __riscv_vwadd_vv_i32m1(sumi_h8_0, sumi_h8_1, vl / 4);
+                        const vfloat32m1_t facc = __riscv_vfcvt_f_x_v_f32m1(sumi_h8, vl / 4);
+
+                        const vfloat32m1_t tmp1 = __riscv_vfmul_vf_f32m1(facc, a_scales[1], vl / 4);
+                        sumf1 = __riscv_vfmacc_vv_f32m1(sumf1, tmp1, b_scales_vec, vl / 4);
+                    }
+
+                    const int64_t A2 = *(const int64_t *)&a_ptr[l].qs[16];
+                    const int64_t A6 = *(const int64_t *)&a_ptr[l].qs[48];
+                    const int64_t Aa = *(const int64_t *)&a_ptr[l].qs[80];
+                    const int64_t Ae = *(const int64_t *)&a_ptr[l].qs[112];
+                    __asm__ __volatile__("" ::: "memory"); // prevent gcc from emitting fused vlse64, violating alignment
+                    vint16m4_t sumi_l2;
+                    {
+                        const vint8m2_t lhs_0_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A2, vl / 4));
+                        const vint8m2_t lhs_1_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A6, vl / 4));
+                        const vint8m2_t lhs_2_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(Aa, vl / 4));
+                        const vint8m2_t lhs_3_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(Ae, vl / 4));
+                        const vint16m4_t sumi_lo_0 = __riscv_vwmul_vv_i16m4(rhs_vec_lo_0, lhs_0_8, vl * 2);
+                        const vint16m4_t sumi_lo_1 = __riscv_vwmacc_vv_i16m4(sumi_lo_0, rhs_vec_lo_1, lhs_1_8, vl * 2);
+                        const vint16m4_t sumi_hi_0 = __riscv_vwmacc_vv_i16m4(sumi_lo_1, rhs_vec_hi_0, lhs_2_8, vl * 2);
+                        const vint16m4_t sumi_hi_m = __riscv_vwmacc_vv_i16m4(sumi_hi_0, rhs_vec_hi_1, lhs_3_8, vl * 2);
+
+                        sumi_l2 = sumi_hi_m;
+                    }
+
+                    {
+                        const vuint32m4_t sumi_i32 = __riscv_vreinterpret_v_i32m4_u32m4(__riscv_vreinterpret_v_i16m4_i32m4(sumi_l2));
+                        const vuint16m2_t sumi_h2_0 = __riscv_vnsrl_wx_u16m2(sumi_i32, 0, vl);
+                        const vuint16m2_t sumi_h2_1 = __riscv_vnsrl_wx_u16m2(sumi_i32, 16, vl);
+                        const vuint16m2_t sumi_h2 = __riscv_vadd_vv_u16m2(sumi_h2_0, sumi_h2_1, vl);
+                        const vuint32m2_t sumi_h2_i32 = __riscv_vreinterpret_v_u16m2_u32m2(sumi_h2);
+                        const vuint16m1_t sumi_h4_0 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 0, vl / 2);
+                        const vuint16m1_t sumi_h4_1 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 16, vl / 2);
+                        const vuint16m1_t sumi_h4 = __riscv_vadd_vv_u16m1(sumi_h4_0, sumi_h4_1, vl / 2);
+                        const vuint32m1_t sumi_h4_i32 = __riscv_vreinterpret_v_u16m1_u32m1(sumi_h4);
+                        const vint16mf2_t sumi_h8_0 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 0, vl / 4));
+                        const vint16mf2_t sumi_h8_1 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 16, vl / 4));
+                        const vint32m1_t sumi_h8 = __riscv_vwadd_vv_i32m1(sumi_h8_0, sumi_h8_1, vl / 4);
+                        const vfloat32m1_t facc = __riscv_vfcvt_f_x_v_f32m1(sumi_h8, vl / 4);
+
+                        const vfloat32m1_t tmp1 = __riscv_vfmul_vf_f32m1(facc, a_scales[2], vl / 4);
+                        sumf2 = __riscv_vfmacc_vv_f32m1(sumf2, tmp1, b_scales_vec, vl / 4);
+                    }
+
+                    const int64_t A3 = *(const int64_t *)&a_ptr[l].qs[24];
+                    const int64_t A7 = *(const int64_t *)&a_ptr[l].qs[56];
+                    const int64_t Ab = *(const int64_t *)&a_ptr[l].qs[88];
+                    const int64_t Af = *(const int64_t *)&a_ptr[l].qs[120];
+                    __asm__ __volatile__("" ::: "memory"); // prevent gcc from emitting fused vlse64, violating alignment
+                    vint16m4_t sumi_l3;
+                    {
+                        const vint8m2_t lhs_0_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A3, vl / 4));
+                        const vint8m2_t lhs_1_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A7, vl / 4));
+                        const vint8m2_t lhs_2_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(Ab, vl / 4));
+                        const vint8m2_t lhs_3_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(Af, vl / 4));
+                        const vint16m4_t sumi_lo_0 = __riscv_vwmul_vv_i16m4(rhs_vec_lo_0, lhs_0_8, vl * 2);
+                        const vint16m4_t sumi_lo_1 = __riscv_vwmacc_vv_i16m4(sumi_lo_0, rhs_vec_lo_1, lhs_1_8, vl * 2);
+                        const vint16m4_t sumi_hi_0 = __riscv_vwmacc_vv_i16m4(sumi_lo_1, rhs_vec_hi_0, lhs_2_8, vl * 2);
+                        const vint16m4_t sumi_hi_m = __riscv_vwmacc_vv_i16m4(sumi_hi_0, rhs_vec_hi_1, lhs_3_8, vl * 2);
+
+                        sumi_l3 = sumi_hi_m;
+                    }
+
+                    {
+                        const vuint32m4_t sumi_i32 = __riscv_vreinterpret_v_i32m4_u32m4(__riscv_vreinterpret_v_i16m4_i32m4(sumi_l3));
+                        const vuint16m2_t sumi_h2_0 = __riscv_vnsrl_wx_u16m2(sumi_i32, 0, vl);
+                        const vuint16m2_t sumi_h2_1 = __riscv_vnsrl_wx_u16m2(sumi_i32, 16, vl);
+                        const vuint16m2_t sumi_h2 = __riscv_vadd_vv_u16m2(sumi_h2_0, sumi_h2_1, vl);
+                        const vuint32m2_t sumi_h2_i32 = __riscv_vreinterpret_v_u16m2_u32m2(sumi_h2);
+                        const vuint16m1_t sumi_h4_0 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 0, vl / 2);
+                        const vuint16m1_t sumi_h4_1 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 16, vl / 2);
+                        const vuint16m1_t sumi_h4 = __riscv_vadd_vv_u16m1(sumi_h4_0, sumi_h4_1, vl / 2);
+                        const vuint32m1_t sumi_h4_i32 = __riscv_vreinterpret_v_u16m1_u32m1(sumi_h4);
+                        const vint16mf2_t sumi_h8_0 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 0, vl / 4));
+                        const vint16mf2_t sumi_h8_1 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 16, vl / 4));
+                        const vint32m1_t sumi_h8 = __riscv_vwadd_vv_i32m1(sumi_h8_0, sumi_h8_1, vl / 4);
+                        const vfloat32m1_t facc = __riscv_vfcvt_f_x_v_f32m1(sumi_h8, vl / 4);
+
+                        const vfloat32m1_t tmp1 = __riscv_vfmul_vf_f32m1(facc, a_scales[3], vl / 4);
+                        sumf3 = __riscv_vfmacc_vv_f32m1(sumf3, tmp1, b_scales_vec, vl / 4);
+                    }
+                }
+                __riscv_vse32_v_f32m1(&s[(y * 4 + 0) * bs + x * ncols_interleaved], sumf0, vl / 4);
+                __riscv_vse32_v_f32m1(&s[(y * 4 + 1) * bs + x * ncols_interleaved], sumf1, vl / 4);
+                __riscv_vse32_v_f32m1(&s[(y * 4 + 2) * bs + x * ncols_interleaved], sumf2, vl / 4);
+                __riscv_vse32_v_f32m1(&s[(y * 4 + 3) * bs + x * ncols_interleaved], sumf3, vl / 4);
+            }
+        }
+
+        return;
+    }
+
+#endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__)
+    float sumf[4][8];
+    int sumi;
+
+    for (int y = 0; y < nr / 4; y++) {
+        const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
+        for (int x = 0; x < nc / ncols_interleaved; x++) {
+            const block_q4_0x8 * b_ptr = (const block_q4_0x8 *) vx + (x * nb);
+            for (int m = 0; m < 4; m++) {
+                for (int j = 0; j < ncols_interleaved; j++) sumf[m][j] = 0.0;
+            }
+            for (int l = 0; l < nb; l++) {
+                for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                    for (int m = 0; m < 4; m++) {
+                        for (int j = 0; j < ncols_interleaved; j++) {
+                            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] << 4);
+                                const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
+                                sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
+                                         (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4;
+                            }
+                            sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(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];
+            }
+        }
+    }
+}
diff --git a/ggml/src/ggml-cpu/arch/s390/quants.c b/ggml/src/ggml-cpu/arch/s390/quants.c
new file mode 100644
index 000000000..a840219a4
--- /dev/null
+++ b/ggml/src/ggml-cpu/arch/s390/quants.c
@@ -0,0 +1,1300 @@
+#define GGML_COMMON_IMPL_C
+#include "ggml-common.h"
+#include "ggml-quants.h"
+#include "ggml-impl.h"
+#include "ggml-cpu.h"
+#include "simd-mappings.h"
+
+#include "../../quants.h"
+#include "../../ggml-cpu-impl.h"
+
+#include <math.h>
+#include <string.h>
+#include <assert.h>
+#include <float.h>
+#include <stdlib.h> // for qsort
+#include <stdio.h>  // for GGML_ASSERT
+
+#define GROUP_MAX_EPS 1e-15f
+#define GROUP_MAX_EPS_IQ3_XXS 1e-8f
+#define GROUP_MAX_EPS_IQ2_S 1e-8f
+#define GROUP_MAX_EPS_IQ1_M 1e-7f
+#define GROUP_MAX_EPS_IQ1_S 1e-12f
+
+#define UNUSED GGML_UNUSED
+
+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 * GGML_RESTRICT y = vy;
+
+#if 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_CPU_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
+    quantize_row_q8_0_ref(x, y, k);
+#endif
+}
+
+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 * GGML_RESTRICT y = vy;
+
+#if 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_CPU_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_CPU_FP32_TO_FP16(d * (acc[0] + acc[1] + acc[2] + acc[3]));
+    }
+#else
+    GGML_UNUSED(nb);
+    // scalar
+    quantize_row_q8_1_ref(x, y, k);
+#endif
+}
+
+
+//===================================== Dot products =================================
+
+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;
+
+    assert(n % qk == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q4_0 * GGML_RESTRICT x = vx;
+    const block_q8_0 * GGML_RESTRICT y = vy;
+
+    int ib = 0;
+    float sumf = 0;
+
+#if 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_CPU_FP16_TO_FP32(x[ib].d) * GGML_CPU_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;
+        int sumi1 = 0;
+
+        for (int j = 0; j < qk/2; ++j) {
+            const int v0 = (x[ib].qs[j] & 0x0F) - 8;
+            const int v1 = (x[ib].qs[j] >>   4) - 8;
+
+            sumi0 += (v0 * y[ib].qs[j]);
+            sumi1 += (v1 * y[ib].qs[j + qk/2]);
+        }
+
+        int sumi = sumi0 + sumi1;
+        sumf += sumi*GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d);
+    }
+
+    *s = sumf;
+}
+
+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;
+
+    assert(n % qk == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q4_1 * GGML_RESTRICT x = vx;
+    const block_q8_1 * GGML_RESTRICT y = vy;
+
+    int ib = 0;
+    float sumf = 0;
+
+#if 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_CPU_FP16_TO_FP32(x[ib].m) * GGML_CPU_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_CPU_FP16_TO_FP32(x[ib].d) * GGML_CPU_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;
+        int sumi1 = 0;
+
+        for (int j = 0; j < qk/2; ++j) {
+            const int v0 = (x[ib].qs[j] & 0x0F);
+            const int v1 = (x[ib].qs[j] >>   4);
+
+            sumi0 += (v0 * y[ib].qs[j]);
+            sumi1 += (v1 * y[ib].qs[j + qk/2]);
+        }
+
+        int sumi = sumi0 + sumi1;
+        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
+    }
+
+    *s = sumf;
+}
+
+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;
+
+    assert(n % qk == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q8_0 * GGML_RESTRICT x = vx;
+    const block_q8_0 * GGML_RESTRICT y = vy;
+
+    int ib = 0;
+    float sumf = 0;
+
+#if 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_CPU_FP16_TO_FP32(x[ib].d) * GGML_CPU_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;
+
+        for (int j = 0; j < qk; j++) {
+            sumi += x[ib].qs[j]*y[ib].qs[j];
+        }
+
+        sumf += sumi*(GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d));
+    }
+
+    *s = sumf;
+}
+
+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);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const uint32_t kmask1 = 0x03030303;
+    const uint32_t kmask2 = 0x0f0f0f0f;
+
+    const block_q3_K * GGML_RESTRICT x = vx;
+    const block_q8_K * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+#if defined(__VXE__) || defined(__VXE2__)
+    uint32_t aux[3];
+    uint32_t utmp[4];
+
+    const int32x4_t v_z = vec_splat_s32(0);
+    const uint8x16_t v_3m = vec_splat_u8(0x03);
+
+    const uint8x16_t v_0c = vec_splat_u8(1);
+    const uint8x16_t v_1c = vec_sl(v_0c, 1);
+    const uint8x16_t v_2c = vec_sl(v_0c, 2);
+    const uint8x16_t v_3c = vec_sl(v_0c, 3);
+
+    uint8x16_t q3h[4];
+    uint8x16_t q3b[2];
+    int8x16_t q3bytes[4];
+    int8x16_t q8bytes[4];
+    uint8x16_t qhbits[2];
+
+    float sum = 0;
+
+    for (int i = 0; i < nb; ++i) {
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+
+        const uint8_t * restrict x0l = x[i].qs;
+        const uint8_t * restrict x0h = x[i].hmask;
+        const int8_t  * restrict y0  = y[i].qs;
+
+        qhbits[0] = vec_xl(0 , x0h);
+        qhbits[1] = vec_xl(16, x0h);
+
+        int32_t isum = 0;
+
+        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;
+
+        for (int j = 0; j < QK_K/128; ++j) {
+            int32x4_t isum0, isum1, isum2, isum3;
+
+            q3b[0] = vec_xl(0 , x0l);
+            q3b[1] = vec_xl(16, x0l);
+            x0l += 32;
+
+            q8bytes[0] = vec_xl(0  , y0);
+            q8bytes[1] = vec_xl(16 , y0);
+            q8bytes[2] = vec_xl(32 , y0);
+            q8bytes[3] = vec_xl(48 , y0);
+            q8bytes[4] = vec_xl(64 , y0);
+            q8bytes[5] = vec_xl(80 , y0);
+            q8bytes[6] = vec_xl(96 , y0);
+            q8bytes[7] = vec_xl(112, y0);
+            y0 += 128;
+
+            q3h[0] = vec_sl(vec_andc(v_0c, qhbits[0]), 2);
+            q3h[1] = vec_sl(vec_andc(v_0c, qhbits[1]), 2);
+            q3h[2] = vec_sl(vec_andc(v_1c, qhbits[0]), 1);
+            q3h[3] = vec_sl(vec_andc(v_1c, qhbits[1]), 1);
+
+            q3bytes[0] = vec_sub((int8x16_t)vec_and(q3b[0], v_3m), (int8x16_t)q3h[0]);
+            q3bytes[1] = vec_sub((int8x16_t)vec_and(q3b[1], v_3m), (int8x16_t)q3h[1]);
+            q3bytes[2] = vec_sub((int8x16_t)vec_and(vec_sr(q3b[0], 2), v_3m), (int8x16_t)q3h[2]);
+            q3bytes[3] = vec_sub((int8x16_t)vec_and(vec_sr(q3b[1], 2), v_3m), (int8x16_t)q3h[3]);
+
+            isum0 = ggml_vec_dot(v_z, q3bytes[0], q8bytes[0]);
+            isum1 = ggml_vec_dot(v_z, q3bytes[1], q8bytes[1]);
+            isum2 = ggml_vec_dot(v_z, q3bytes[2], q8bytes[2]);
+            isum3 = ggml_vec_dot(v_z, q3bytes[3], q8bytes[3]);
+
+            isum += (isum0[0] + isum0[1] + isum0[2] + isum0[3]) * scale[0];
+            isum += (isum1[0] + isum1[1] + isum1[2] + isum1[3]) * scale[1];
+            isum += (isum2[0] + isum2[1] + isum2[2] + isum2[3]) * scale[2];
+            isum += (isum3[0] + isum3[1] + isum3[2] + isum3[3]) * scale[3];
+
+            scale += 4;
+
+            q3h[0] = vec_andc(v_2c, qhbits[0]);
+            q3h[1] = vec_andc(v_2c, qhbits[1]);
+            q3h[2] = vec_sr(vec_andc(v_3c, qhbits[0]), 1);
+            q3h[3] = vec_sr(vec_andc(v_3c, qhbits[1]), 1);
+
+            q3bytes[0] = vec_sub((int8x16_t)vec_and(vec_sr(q3b[0], 4), v_3m), (int8x16_t)q3h[0]);
+            q3bytes[1] = vec_sub((int8x16_t)vec_and(vec_sr(q3b[1], 4), v_3m), (int8x16_t)q3h[1]);
+            q3bytes[2] = vec_sub((int8x16_t)vec_and(vec_sr(q3b[0], 6), v_3m), (int8x16_t)q3h[2]);
+            q3bytes[3] = vec_sub((int8x16_t)vec_and(vec_sr(q3b[1], 6), v_3m), (int8x16_t)q3h[3]);
+
+            isum0 = ggml_vec_dot(v_z, q3bytes[0], q8bytes[4]);
+            isum1 = ggml_vec_dot(v_z, q3bytes[1], q8bytes[5]);
+            isum2 = ggml_vec_dot(v_z, q3bytes[2], q8bytes[6]);
+            isum3 = ggml_vec_dot(v_z, q3bytes[3], q8bytes[7]);
+
+            isum += (isum0[0] + isum0[1] + isum0[2] + isum0[3]) * scale[0];
+            isum += (isum1[0] + isum1[1] + isum1[2] + isum1[3]) * scale[1];
+            isum += (isum2[0] + isum2[1] + isum2[2] + isum2[3]) * scale[2];
+            isum += (isum3[0] + isum3[1] + isum3[2] + isum3[3]) * scale[3];
+
+            scale += 4;
+
+            if (j == 0) {
+                qhbits[0] = vec_sr(qhbits[0], 4);
+                qhbits[1] = vec_sr(qhbits[1], 4);
+            }
+        }
+
+        sum += d * isum;
+    }
+
+    *s = sum;
+
+#else
+    // scalar version
+    // This function is written like this so the compiler can manage to vectorize most of it
+    // Using -Ofast, GCC and clang manage to produce code that is within a factor of 2 or so from the
+    // manually vectorized version above. Every other version I tried would run at least 4 times slower.
+    // The ideal situation would be if we could just write the code once, and the compiler would
+    // automatically produce the best possible set of machine instructions, instead of us having to manually
+    // write vectorized versions for AVX, ARM_NEON, etc.
+
+    int8_t  aux8[QK_K];
+    int16_t aux16[8];
+    float   sums [8];
+    int32_t aux32[8];
+    memset(sums, 0, 8*sizeof(float));
+
+    uint32_t auxs[4];
+    const int8_t * scales = (const int8_t*)auxs;
+
+    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;
+        memset(aux32, 0, 8*sizeof(int32_t));
+        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;
+            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
+            a += 32; m <<= 1;
+            for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 2) & 3;
+            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
+            a += 32; m <<= 1;
+            for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 4) & 3;
+            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
+            a += 32; m <<= 1;
+            for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 6) & 3;
+            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
+            a += 32; m <<= 1;
+            q3 += 32;
+        }
+        a = aux8;
+
+        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);
+        for (int j = 0; j < QK_K/16; ++j) {
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
+            q8 += 8; a += 8;
+        }
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
+        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
+    }
+    for (int l = 0; l < 8; ++l) sumf += sums[l];
+    *s = sumf;
+
+#endif
+
+}
+
+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);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q4_K * GGML_RESTRICT x = vx;
+    const block_q8_K * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+    static const uint32_t kmask1 = 0x3f3f3f3f;
+    static const uint32_t kmask2 = 0x0f0f0f0f;
+    static const uint32_t kmask3 = 0x03030303;
+
+    uint32_t utmp[4];
+
+#if 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_CPU_FP16_TO_FP32(x[i].d);
+        const float dmin = y[i].d * GGML_CPU_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];
+    const uint8_t * mins   = (const uint8_t*)&utmp[2];
+
+    int8_t  aux8[QK_K];
+    int16_t aux16[8];
+    float   sums [8];
+    int32_t aux32[8];
+    memset(sums, 0, 8*sizeof(float));
+
+    float sumf = 0;
+    for (int i = 0; i < nb; ++i) {
+        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 * 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;
+            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l]  >> 4);
+            a += 32; q4 += 32;
+        }
+        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;
+
+        int sumi = 0;
+        for (int j = 0; j < QK_K/16; ++j) sumi += y[i].bsums[j] * mins[j/2];
+        a = aux8;
+        int is = 0;
+        for (int j = 0; j < QK_K/32; ++j) {
+            int32_t scale = scales[is++];
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+        }
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
+        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
+        const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
+        sumf -= dmin * sumi;
+    }
+    for (int l = 0; l < 8; ++l) sumf += sums[l];
+    *s = sumf;
+#endif
+}
+
+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);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q5_K * GGML_RESTRICT x = vx;
+    const block_q8_K * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+    static const uint32_t kmask1 = 0x3f3f3f3f;
+    static const uint32_t kmask2 = 0x0f0f0f0f;
+    static const uint32_t kmask3 = 0x03030303;
+
+    uint32_t utmp[4];
+
+#if 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_CPU_FP16_TO_FP32(x[i].d);
+        const float dmin = y[i].d * GGML_CPU_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];
+    const uint8_t * mins   = (const uint8_t*)&utmp[2];
+
+    int8_t  aux8[QK_K];
+    int16_t aux16[8];
+    float   sums [8];
+    int32_t aux32[8];
+    memset(sums, 0, 8*sizeof(float));
+
+    float sumf = 0;
+    for (int i = 0; i < nb; ++i) {
+        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 * 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);
+            for (int l = 0; l < 32; ++l) a[l] += (hm[l] & m ? 16 : 0);
+            a += 32; m <<= 1;
+            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l]  >> 4);
+            for (int l = 0; l < 32; ++l) a[l] += (hm[l] & m ? 16 : 0);
+            a += 32; m <<= 1;
+            q4 += 32;
+        }
+        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;
+
+        int sumi = 0;
+        for (int j = 0; j < QK_K/16; ++j) sumi += y[i].bsums[j] * mins[j/2];
+        a = aux8;
+        int is = 0;
+        for (int j = 0; j < QK_K/32; ++j) {
+            int32_t scale = scales[is++];
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+        }
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
+        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
+        const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
+        sumf -= dmin * sumi;
+    }
+    for (int l = 0; l < 8; ++l) sumf += sums[l];
+    *s = sumf;
+#endif
+}
+
+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);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q6_K * GGML_RESTRICT x = vx;
+    const block_q8_K * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+#if 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_CPU_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];
+    int16_t aux16[8];
+    float   sums [8];
+    int32_t aux32[8];
+    memset(sums, 0, 8*sizeof(float));
+
+    float sumf = 0;
+    for (int i = 0; i < nb; ++i) {
+        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 * 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;
+                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;
+        }
+        a = aux8;
+        int is = 0;
+        for (int j = 0; j < QK_K/16; ++j) {
+            int scale = x[i].scales[is++];
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+        }
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
+        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
+    }
+    for (int l = 0; l < 8; ++l) sumf += sums[l];
+    *s = sumf;
+#endif
+}
+
+// #if defined(__VXE__) || defined(__VXE2__)
+// static const int8_t keven_signs_q2xs[1024] = {
+//      1,  1,  1,  1,  1,  1,  1,  1, -1,  1,  1,  1,  1,  1,  1, -1,  1, -1,  1,  1,  1,  1,  1, -1, -1, -1,  1,  1,  1,  1,  1,  1,
+//      1,  1, -1,  1,  1,  1,  1, -1, -1,  1, -1,  1,  1,  1,  1,  1,  1, -1, -1,  1,  1,  1,  1,  1, -1, -1, -1,  1,  1,  1,  1, -1,
+//      1,  1,  1, -1,  1,  1,  1, -1, -1,  1,  1, -1,  1,  1,  1,  1,  1, -1,  1, -1,  1,  1,  1,  1, -1, -1,  1, -1,  1,  1,  1, -1,
+//      1,  1, -1, -1,  1,  1,  1,  1, -1,  1, -1, -1,  1,  1,  1, -1,  1, -1, -1, -1,  1,  1,  1, -1, -1, -1, -1, -1,  1,  1,  1,  1,
+//      1,  1,  1,  1, -1,  1,  1, -1, -1,  1,  1,  1, -1,  1,  1,  1,  1, -1,  1,  1, -1,  1,  1,  1, -1, -1,  1,  1, -1,  1,  1, -1,
+//      1,  1, -1,  1, -1,  1,  1,  1, -1,  1, -1,  1, -1,  1,  1, -1,  1, -1, -1,  1, -1,  1,  1, -1, -1, -1, -1,  1, -1,  1,  1,  1,
+//      1,  1,  1, -1, -1,  1,  1,  1, -1,  1,  1, -1, -1,  1,  1, -1,  1, -1,  1, -1, -1,  1,  1, -1, -1, -1,  1, -1, -1,  1,  1,  1,
+//      1,  1, -1, -1, -1,  1,  1, -1, -1,  1, -1, -1, -1,  1,  1,  1,  1, -1, -1, -1, -1,  1,  1,  1, -1, -1, -1, -1, -1,  1,  1, -1,
+//      1,  1,  1,  1,  1, -1,  1, -1, -1,  1,  1,  1,  1, -1,  1,  1,  1, -1,  1,  1,  1, -1,  1,  1, -1, -1,  1,  1,  1, -1,  1, -1,
+//      1,  1, -1,  1,  1, -1,  1,  1, -1,  1, -1,  1,  1, -1,  1, -1,  1, -1, -1,  1,  1, -1,  1, -1, -1, -1, -1,  1,  1, -1,  1,  1,
+//      1,  1,  1, -1,  1, -1,  1,  1, -1,  1,  1, -1,  1, -1,  1, -1,  1, -1,  1, -1,  1, -1,  1, -1, -1, -1,  1, -1,  1, -1,  1,  1,
+//      1,  1, -1, -1,  1, -1,  1, -1, -1,  1, -1, -1,  1, -1,  1,  1,  1, -1, -1, -1,  1, -1,  1,  1, -1, -1, -1, -1,  1, -1,  1, -1,
+//      1,  1,  1,  1, -1, -1,  1,  1, -1,  1,  1,  1, -1, -1,  1, -1,  1, -1,  1,  1, -1, -1,  1, -1, -1, -1,  1,  1, -1, -1,  1,  1,
+//      1,  1, -1,  1, -1, -1,  1, -1, -1,  1, -1,  1, -1, -1,  1,  1,  1, -1, -1,  1, -1, -1,  1,  1, -1, -1, -1,  1, -1, -1,  1, -1,
+//      1,  1,  1, -1, -1, -1,  1, -1, -1,  1,  1, -1, -1, -1,  1,  1,  1, -1,  1, -1, -1, -1,  1,  1, -1, -1,  1, -1, -1, -1,  1, -1,
+//      1,  1, -1, -1, -1, -1,  1,  1, -1,  1, -1, -1, -1, -1,  1, -1,  1, -1, -1, -1, -1, -1,  1, -1, -1, -1, -1, -1, -1, -1,  1,  1,
+//      1,  1,  1,  1,  1,  1, -1, -1, -1,  1,  1,  1,  1,  1, -1,  1,  1, -1,  1,  1,  1,  1, -1,  1, -1, -1,  1,  1,  1,  1, -1, -1,
+//      1,  1, -1,  1,  1,  1, -1,  1, -1,  1, -1,  1,  1,  1, -1, -1,  1, -1, -1,  1,  1,  1, -1, -1, -1, -1, -1,  1,  1,  1, -1,  1,
+//      1,  1,  1, -1,  1,  1, -1,  1, -1,  1,  1, -1,  1,  1, -1, -1,  1, -1,  1, -1,  1,  1, -1, -1, -1, -1,  1, -1,  1,  1, -1,  1,
+//      1,  1, -1, -1,  1,  1, -1, -1, -1,  1, -1, -1,  1,  1, -1,  1,  1, -1, -1, -1,  1,  1, -1,  1, -1, -1, -1, -1,  1,  1, -1, -1,
+//      1,  1,  1,  1, -1,  1, -1,  1, -1,  1,  1,  1, -1,  1, -1, -1,  1, -1,  1,  1, -1,  1, -1, -1, -1, -1,  1,  1, -1,  1, -1,  1,
+//      1,  1, -1,  1, -1,  1, -1, -1, -1,  1, -1,  1, -1,  1, -1,  1,  1, -1, -1,  1, -1,  1, -1,  1, -1, -1, -1,  1, -1,  1, -1, -1,
+//      1,  1,  1, -1, -1,  1, -1, -1, -1,  1,  1, -1, -1,  1, -1,  1,  1, -1,  1, -1, -1,  1, -1,  1, -1, -1,  1, -1, -1,  1, -1, -1,
+//      1,  1, -1, -1, -1,  1, -1,  1, -1,  1, -1, -1, -1,  1, -1, -1,  1, -1, -1, -1, -1,  1, -1, -1, -1, -1, -1, -1, -1,  1, -1,  1,
+//      1,  1,  1,  1,  1, -1, -1,  1, -1,  1,  1,  1,  1, -1, -1, -1,  1, -1,  1,  1,  1, -1, -1, -1, -1, -1,  1,  1,  1, -1, -1,  1,
+//      1,  1, -1,  1,  1, -1, -1, -1, -1,  1, -1,  1,  1, -1, -1,  1,  1, -1, -1,  1,  1, -1, -1,  1, -1, -1, -1,  1,  1, -1, -1, -1,
+//      1,  1,  1, -1,  1, -1, -1, -1, -1,  1,  1, -1,  1, -1, -1,  1,  1, -1,  1, -1,  1, -1, -1,  1, -1, -1,  1, -1,  1, -1, -1, -1,
+//      1,  1, -1, -1,  1, -1, -1,  1, -1,  1, -1, -1,  1, -1, -1, -1,  1, -1, -1, -1,  1, -1, -1, -1, -1, -1, -1, -1,  1, -1, -1,  1,
+//      1,  1,  1,  1, -1, -1, -1, -1, -1,  1,  1,  1, -1, -1, -1,  1,  1, -1,  1,  1, -1, -1, -1,  1, -1, -1,  1,  1, -1, -1, -1, -1,
+//      1,  1, -1,  1, -1, -1, -1,  1, -1,  1, -1,  1, -1, -1, -1, -1,  1, -1, -1,  1, -1, -1, -1, -1, -1, -1, -1,  1, -1, -1, -1,  1,
+//      1,  1,  1, -1, -1, -1, -1,  1, -1,  1,  1, -1, -1, -1, -1, -1,  1, -1,  1, -1, -1, -1, -1, -1, -1, -1,  1, -1, -1, -1, -1,  1,
+//      1,  1, -1, -1, -1, -1, -1, -1, -1,  1, -1, -1, -1, -1, -1,  1,  1, -1, -1, -1, -1, -1, -1,  1, -1, -1, -1, -1, -1, -1, -1, -1,
+// };
+// #endif
+
+// 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);
+//     UNUSED(bx);
+//     UNUSED(by);
+//     UNUSED(bs);
+
+//     const block_iq2_xxs * GGML_RESTRICT x = vx;
+//     const block_q8_K    * GGML_RESTRICT y = vy;
+
+//     const int nb = n / QK_K;
+
+// #if 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_CPU_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];
+//     const uint8_t * aux8 = (const uint8_t *)aux32;
+
+//     float sumf = 0.f;
+//     for (int i = 0; i < nb; ++i) {
+//         const float d = GGML_CPU_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;
+//         int32_t bsum = 0;
+//         for (int ib32 = 0; ib32 < QK_K/32; ++ib32) {
+//             memcpy(aux32, q2, 2*sizeof(uint32_t));
+//             q2 += 4;
+//             const uint32_t ls = 2*(aux32[1] >> 28) + 1;
+//             int32_t sumi = 0;
+//             for (int l = 0; l < 4; ++l) {
+//                 const uint8_t * grid = (const uint8_t *)(iq2xxs_grid + aux8[l]);
+//                 const uint8_t  signs = ksigns_iq2xs[(aux32[1] >> 7*l) & 127];
+//                 for (int j = 0; j < 8; ++j) {
+//                     sumi += grid[j] * q8[j] * (signs & kmask_iq2xs[j] ? -1 : 1);
+//                 }
+//                 q8 += 8;
+//             }
+//             bsum += sumi * ls;
+//         }
+//         sumf += d * bsum;
+//     }
+//     *s = 0.125f * sumf;
+// #endif
+// }
+
+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);
+    UNUSED(by);
+    UNUSED(bs);
+    assert(n % QK4_NL == 0);
+    static_assert(QK4_NL == QK8_0, "QK4_NL and QK8_0 must be the same");
+
+    const block_iq4_nl * GGML_RESTRICT x = vx;
+    const block_q8_0   * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK4_NL;
+
+    int ib = 0;
+    float sumf = 0;
+
+#if 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_CPU_FP16_TO_FP32(x0->d) * GGML_CPU_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_CPU_FP16_TO_FP32(y[ib].d)*GGML_CPU_FP16_TO_FP32(x[ib].d);
+        int sumi1 = 0, sumi2 = 0;
+        for (int j = 0; j < QK4_NL/2; ++j) {
+            sumi1 += y[ib].qs[j+       0] * kvalues_iq4nl[x[ib].qs[j] & 0xf];
+            sumi2 += y[ib].qs[j+QK4_NL/2] * kvalues_iq4nl[x[ib].qs[j] >>  4];
+        }
+        sumf += d * (sumi1 + sumi2);
+    }
+    *s = sumf;
+}
+
+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);
+    UNUSED(by);
+    UNUSED(bs);
+    assert(n % QK_K == 0);
+
+    const block_iq4_xs * GGML_RESTRICT x = vx;
+    const block_q8_K   * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+#if 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_CPU_FP16_TO_FP32(x[ibl].d) * y[ibl].d * (sumi1 + sumi2);
+    }
+
+    *s = sumf;
+
+#else
+    float sumf = 0;
+    for (int ibl = 0; ibl < nb; ++ibl) {
+        const float d4d8 = GGML_CPU_FP16_TO_FP32(x[ibl].d) * y[ibl].d;
+        uint16_t h = x[ibl].scales_h;
+        const uint8_t * qs = x[ibl].qs;
+        const int8_t  * q8 = y[ibl].qs;
+        for (int ib = 0; ib < QK_K/32; ib += 2) {
+            const uint8_t ls1 = (x[ibl].scales_l[ib/2] & 0xf) | ((h << 4) & 0x30);
+            const uint8_t ls2 = (x[ibl].scales_l[ib/2] >>  4) | ((h << 2) & 0x30);
+            h >>= 4;
+            const float d1 = d4d8*(ls1 - 32);
+            const float d2 = d4d8*(ls2 - 32);
+            int sumi1 = 0, sumi2 = 0;
+            for (int j = 0; j < 16; ++j) {
+                sumi1 += q8[j+ 0] * kvalues_iq4nl[qs[j] & 0xf];
+                sumi2 += q8[j+16] * kvalues_iq4nl[qs[j] >>  4];
+            }
+            sumf += d1 * (sumi1 + sumi2);
+            qs += 16;
+            q8 += 32;
+            sumi1 = sumi2 = 0;
+            for (int j = 0; j < 16; ++j) {
+                sumi1 += q8[j+ 0] * kvalues_iq4nl[qs[j] & 0xf];
+                sumi2 += q8[j+16] * kvalues_iq4nl[qs[j] >>  4];
+            }
+            sumf += d2 * (sumi1 + sumi2);
+            qs += 16;
+            q8 += 32;
+        }
+    }
+    *s = sumf;
+#endif
+}
+
diff --git a/ggml/src/ggml-cpu/arch/wasm/quants.c b/ggml/src/ggml-cpu/arch/wasm/quants.c
new file mode 100644
index 000000000..b0904d8a3
--- /dev/null
+++ b/ggml/src/ggml-cpu/arch/wasm/quants.c
@@ -0,0 +1,1481 @@
+#define GGML_COMMON_IMPL_C
+#include "ggml-common.h"
+#include "ggml-quants.h"
+#include "ggml-impl.h"
+#include "ggml-cpu.h"
+#include "simd-mappings.h"
+
+#include "../../quants.h"
+#include "../../ggml-cpu-impl.h"
+
+#include <math.h>
+#include <string.h>
+#include <assert.h>
+#include <float.h>
+#include <stdlib.h> // for qsort
+#include <stdio.h>  // for GGML_ASSERT
+
+#define GROUP_MAX_EPS 1e-15f
+#define GROUP_MAX_EPS_IQ3_XXS 1e-8f
+#define GROUP_MAX_EPS_IQ2_S 1e-8f
+#define GROUP_MAX_EPS_IQ1_M 1e-7f
+#define GROUP_MAX_EPS_IQ1_S 1e-12f
+
+#define UNUSED GGML_UNUSED
+
+#if defined(__wasm_simd128__)
+#define B1(c,s,n)  0x ## n ## c ,  0x ## n ## s
+#define B2(c,s,n) B1(c,s,n ## c), B1(c,s,n ## s)
+#define B3(c,s,n) B2(c,s,n ## c), B2(c,s,n ## s)
+#define B4(c,s,n) B3(c,s,n ## c), B3(c,s,n ## s)
+#define B5(c,s,n) B4(c,s,n ## c), B4(c,s,n ## s)
+#define B6(c,s,n) B5(c,s,n ## c), B5(c,s,n ## s)
+#define B7(c,s,n) B6(c,s,n ## c), B6(c,s,n ## s)
+#define B8(c,s  ) B7(c,s,     c), B7(c,s,     s)
+
+// precomputed tables for expanding 8bits to 8 bytes:
+static const uint64_t table_b2b_0[1 << 8] = { B8(00, 10) }; // ( b) << 4
+static const uint64_t table_b2b_1[1 << 8] = { B8(10, 00) }; // (!b) << 4
+#endif
+
+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 * GGML_RESTRICT y = vy;
+
+#if defined __wasm_simd128__
+    for (int i = 0; i < nb; i++) {
+        v128_t srcv [8];
+        v128_t asrcv[8];
+        v128_t amaxv[8];
+
+        for (int j = 0; j < 8; j++) srcv[j]  = wasm_v128_load(x + i*32 + 4*j);
+        for (int j = 0; j < 8; j++) asrcv[j] = wasm_f32x4_abs(srcv[j]);
+
+        for (int j = 0; j < 4; j++) amaxv[2*j] = wasm_f32x4_max(asrcv[2*j], asrcv[2*j+1]);
+        for (int j = 0; j < 2; j++) amaxv[4*j] = wasm_f32x4_max(amaxv[4*j], amaxv[4*j+2]);
+        for (int j = 0; j < 1; j++) amaxv[8*j] = wasm_f32x4_max(amaxv[8*j], amaxv[8*j+4]);
+
+        const float amax = MAX(MAX(wasm_f32x4_extract_lane(amaxv[0], 0),
+                                   wasm_f32x4_extract_lane(amaxv[0], 1)),
+                               MAX(wasm_f32x4_extract_lane(amaxv[0], 2),
+                                   wasm_f32x4_extract_lane(amaxv[0], 3)));
+
+        const float d = amax / ((1 << 7) - 1);
+        const float id = d ? 1.0f/d : 0.0f;
+
+        y[i].d = GGML_CPU_FP32_TO_FP16(d);
+
+        for (int j = 0; j < 8; j++) {
+            const v128_t v  = wasm_f32x4_mul(srcv[j], wasm_f32x4_splat(id));
+            const v128_t vi = wasm_i32x4_trunc_sat_f32x4(v);
+
+            y[i].qs[4*j + 0] = wasm_i32x4_extract_lane(vi, 0);
+            y[i].qs[4*j + 1] = wasm_i32x4_extract_lane(vi, 1);
+            y[i].qs[4*j + 2] = wasm_i32x4_extract_lane(vi, 2);
+            y[i].qs[4*j + 3] = wasm_i32x4_extract_lane(vi, 3);
+        }
+    }
+#else
+    GGML_UNUSED(nb);
+    // scalar
+    quantize_row_q8_0_ref(x, y, k);
+#endif
+}
+
+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 * GGML_RESTRICT y = vy;
+#if defined __wasm_simd128__
+    for (int i = 0; i < nb; i++) {
+        v128_t srcv [8];
+        v128_t asrcv[8];
+        v128_t amaxv[8];
+
+        for (int j = 0; j < 8; j++) srcv[j]  = wasm_v128_load(x + i*32 + 4*j);
+        for (int j = 0; j < 8; j++) asrcv[j] = wasm_f32x4_abs(srcv[j]);
+
+        for (int j = 0; j < 4; j++) amaxv[2*j] = wasm_f32x4_max(asrcv[2*j], asrcv[2*j+1]);
+        for (int j = 0; j < 2; j++) amaxv[4*j] = wasm_f32x4_max(amaxv[4*j], amaxv[4*j+2]);
+        for (int j = 0; j < 1; j++) amaxv[8*j] = wasm_f32x4_max(amaxv[8*j], amaxv[8*j+4]);
+
+        const float amax = MAX(MAX(wasm_f32x4_extract_lane(amaxv[0], 0),
+                                   wasm_f32x4_extract_lane(amaxv[0], 1)),
+                               MAX(wasm_f32x4_extract_lane(amaxv[0], 2),
+                                   wasm_f32x4_extract_lane(amaxv[0], 3)));
+
+        const float d = amax / ((1 << 7) - 1);
+        const float id = d ? 1.0f/d : 0.0f;
+
+        y[i].d = GGML_CPU_FP32_TO_FP16(d);
+
+        v128_t accv = wasm_i32x4_splat(0);
+
+        for (int j = 0; j < 8; j++) {
+            const v128_t v  = wasm_f32x4_mul(srcv[j], wasm_f32x4_splat(id));
+            const v128_t vi = wasm_i32x4_trunc_sat_f32x4(v);
+
+            y[i].qs[4*j + 0] = wasm_i32x4_extract_lane(vi, 0);
+            y[i].qs[4*j + 1] = wasm_i32x4_extract_lane(vi, 1);
+            y[i].qs[4*j + 2] = wasm_i32x4_extract_lane(vi, 2);
+            y[i].qs[4*j + 3] = wasm_i32x4_extract_lane(vi, 3);
+
+            accv = wasm_i32x4_add(accv, vi);
+        }
+
+        y[i].s = GGML_CPU_FP32_TO_FP16(
+                d * (wasm_i32x4_extract_lane(accv, 0) +
+                     wasm_i32x4_extract_lane(accv, 1) +
+                     wasm_i32x4_extract_lane(accv, 2) +
+                     wasm_i32x4_extract_lane(accv, 3)));
+    }
+#else
+    GGML_UNUSED(nb);
+    // scalar
+    quantize_row_q8_1_ref(x, y, k);
+#endif
+}
+
+//===================================== Q8_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 =================================
+
+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;
+
+    assert(n % qk == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q4_0 * GGML_RESTRICT x = vx;
+    const block_q8_0 * GGML_RESTRICT y = vy;
+
+    int ib = 0;
+    float sumf = 0;
+
+#if 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_CPU_FP16_TO_FP32(x0->d) * GGML_CPU_FP16_TO_FP32(y0->d);
+        float scale1 = GGML_CPU_FP16_TO_FP32(x1->d) * GGML_CPU_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);
+
+#endif
+    for (; ib < nb; ++ib) {
+        int sumi0 = 0;
+        int sumi1 = 0;
+
+        for (int j = 0; j < qk/2; ++j) {
+            const int v0 = (x[ib].qs[j] & 0x0F) - 8;
+            const int v1 = (x[ib].qs[j] >>   4) - 8;
+
+            sumi0 += (v0 * y[ib].qs[j]);
+            sumi1 += (v1 * y[ib].qs[j + qk/2]);
+        }
+
+        int sumi = sumi0 + sumi1;
+        sumf += sumi*GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d);
+    }
+
+    *s = sumf;
+}
+
+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;
+
+    int ib = 0;
+    float sumf = 0;
+
+    assert(n % qk == 0);
+    assert(qk == QK5_0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q5_0 * GGML_RESTRICT x = vx;
+    const block_q8_0 * GGML_RESTRICT y = vy;
+
+#if defined __wasm_simd128__
+    v128_t sumv = wasm_f32x4_splat(0.0f);
+
+    uint32_t qh_;
+    uint64_t tmp[4];
+
+    // TODO: check if unrolling this is better
+    for (; ib < nb; ++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_));
+
+        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);
+
+        const v128_t v0 = wasm_v128_load(x0->qs);
+
+        // 4-bit -> 8-bit
+        const v128_t v0l = wasm_v128_and (v0, m4b);
+        const v128_t v0h = wasm_u8x16_shr(v0, 4);
+
+        // add high bit and sub 16 (equivalent to sub 0x10 when bit is zero)
+        const v128_t v0lf = wasm_i8x16_sub(v0l, qhl);
+        const v128_t v0hf = wasm_i8x16_sub(v0h, qhh);
+
+        // load y
+        const v128_t v1l = wasm_v128_load(y0->qs);
+        const v128_t v1h = wasm_v128_load(y0->qs + 16);
+
+        // int8x16 -> int16x8
+        const v128_t v0lfl = wasm_i16x8_extend_low_i8x16 (v0lf);
+        const v128_t v0lfh = wasm_i16x8_extend_high_i8x16(v0lf);
+        const v128_t v0hfl = wasm_i16x8_extend_low_i8x16 (v0hf);
+        const v128_t v0hfh = wasm_i16x8_extend_high_i8x16(v0hf);
+
+        const v128_t v1ll = wasm_i16x8_extend_low_i8x16 (v1l);
+        const v128_t v1lh = wasm_i16x8_extend_high_i8x16(v1l);
+        const v128_t v1hl = wasm_i16x8_extend_low_i8x16 (v1h);
+        const v128_t v1hh = wasm_i16x8_extend_high_i8x16(v1h);
+
+        // dot product
+        sumv = wasm_f32x4_add(sumv, wasm_f32x4_mul(wasm_f32x4_convert_i32x4(
+                        wasm_i32x4_add(
+                            wasm_i32x4_add(wasm_i32x4_dot_i16x8(v0lfl, v1ll),
+                                           wasm_i32x4_dot_i16x8(v0lfh, v1lh)),
+                            wasm_i32x4_add(wasm_i32x4_dot_i16x8(v0hfl, v1hl),
+                                           wasm_i32x4_dot_i16x8(v0hfh, v1hh)))),
+                    wasm_f32x4_splat(GGML_CPU_FP16_TO_FP32(x0->d) * GGML_CPU_FP16_TO_FP32(y0->d))));
+    }
+
+    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);
+
+#endif
+    for (; ib < nb; ++ib) {
+        uint32_t qh;
+        memcpy(&qh, x[ib].qh, sizeof(qh));
+
+        int sumi0 = 0;
+        int sumi1 = 0;
+
+        for (int j = 0; j < qk/2; ++j) {
+            const uint8_t xh_0 = ((qh & (1u << (j + 0 ))) >> (j + 0 )) << 4;
+            const uint8_t xh_1 = ((qh & (1u << (j + 16))) >> (j + 12));
+
+            const int32_t x0 = (int8_t)(((x[ib].qs[j] & 0x0F) | xh_0) - 16);
+            const int32_t x1 = (int8_t)(((x[ib].qs[j] >>   4) | xh_1) - 16);
+
+            sumi0 += (x0 * y[ib].qs[j]);
+            sumi1 += (x1 * y[ib].qs[j + qk/2]);
+        }
+
+        int sumi = sumi0 + sumi1;
+        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d)) * sumi;
+    }
+
+    *s = sumf;
+}
+
+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;
+
+    int ib = 0;
+    float sumf = 0;
+
+    assert(n % qk == 0);
+    assert(qk == QK5_1);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q5_1 * GGML_RESTRICT x = vx;
+    const block_q8_1 * GGML_RESTRICT y = vy;
+
+#if defined __wasm_simd128__
+    v128_t sumv = wasm_f32x4_splat(0.0f);
+
+    float summs = 0.0f;
+
+    uint32_t qh_;
+    uint64_t tmp[4];
+
+    // TODO: check if unrolling this is better
+    for (; ib < nb; ++ib) {
+        const block_q5_1 * GGML_RESTRICT x0 = &x[ib];
+        const block_q8_1 * GGML_RESTRICT y0 = &y[ib];
+
+        summs += GGML_CPU_FP16_TO_FP32(x0->m) * GGML_CPU_FP16_TO_FP32(y0->s);
+
+        const v128_t m4b = wasm_i8x16_splat(0x0F);
+
+        // extract the 5th bit
+        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)       ];
+
+        const v128_t qhl = wasm_v128_load(tmp + 0);
+        const v128_t qhh = wasm_v128_load(tmp + 2);
+
+        const v128_t v0 = wasm_v128_load(x0->qs);
+
+        // 4-bit -> 8-bit
+        const v128_t v0l = wasm_v128_and (v0, m4b);
+        const v128_t v0h = wasm_u8x16_shr(v0, 4);
+
+        // add high bit
+        const v128_t v0lf = wasm_v128_or(v0l, qhl);
+        const v128_t v0hf = wasm_v128_or(v0h, qhh);
+
+        // load y
+        const v128_t v1l = wasm_v128_load(y0->qs);
+        const v128_t v1h = wasm_v128_load(y0->qs + 16);
+
+        // int8x16 -> int16x8
+        const v128_t v0lfl = wasm_i16x8_extend_low_i8x16 (v0lf);
+        const v128_t v0lfh = wasm_i16x8_extend_high_i8x16(v0lf);
+        const v128_t v0hfl = wasm_i16x8_extend_low_i8x16 (v0hf);
+        const v128_t v0hfh = wasm_i16x8_extend_high_i8x16(v0hf);
+
+        const v128_t v1ll = wasm_i16x8_extend_low_i8x16 (v1l);
+        const v128_t v1lh = wasm_i16x8_extend_high_i8x16(v1l);
+        const v128_t v1hl = wasm_i16x8_extend_low_i8x16 (v1h);
+        const v128_t v1hh = wasm_i16x8_extend_high_i8x16(v1h);
+
+        // dot product
+        sumv = wasm_f32x4_add(sumv,
+                wasm_f32x4_mul(wasm_f32x4_convert_i32x4(wasm_i32x4_add(
+                            wasm_i32x4_add(wasm_i32x4_dot_i16x8(v0lfl, v1ll),
+                                           wasm_i32x4_dot_i16x8(v0lfh, v1lh)),
+                            wasm_i32x4_add(wasm_i32x4_dot_i16x8(v0hfl, v1hl),
+                                           wasm_i32x4_dot_i16x8(v0hfh, v1hh)))),
+                    wasm_f32x4_splat(GGML_CPU_FP16_TO_FP32(x0->d) * GGML_CPU_FP16_TO_FP32(y0->d))));
+    }
+
+    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) + summs;
+
+#endif
+    for (; ib < nb; ++ib) {
+        uint32_t qh;
+        memcpy(&qh, x[ib].qh, sizeof(qh));
+
+        int sumi0 = 0;
+        int sumi1 = 0;
+
+        for (int j = 0; j < qk/2; ++j) {
+            const uint8_t xh_0 = ((qh >> (j +  0)) << 4) & 0x10;
+            const uint8_t xh_1 = ((qh >> (j + 12))     ) & 0x10;
+
+            const int32_t x0 = (x[ib].qs[j] & 0xF) | xh_0;
+            const int32_t x1 = (x[ib].qs[j] >>  4) | xh_1;
+
+            sumi0 += (x0 * y[ib].qs[j]);
+            sumi1 += (x1 * y[ib].qs[j + qk/2]);
+        }
+
+        int sumi = sumi0 + sumi1;
+        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
+    }
+
+    *s = sumf;
+}
+
+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;
+
+    assert(n % qk == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q8_0 * GGML_RESTRICT x = vx;
+    const block_q8_0 * GGML_RESTRICT y = vy;
+
+    int ib = 0;
+    float sumf = 0;
+
+#if 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_CPU_FP16_TO_FP32(x0->d) * GGML_CPU_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);
+
+#endif
+    for (; ib < nb; ++ib) {
+        int sumi = 0;
+
+        for (int j = 0; j < qk; j++) {
+            sumi += x[ib].qs[j]*y[ib].qs[j];
+        }
+
+        sumf += sumi*(GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d));
+    }
+
+    *s = sumf;
+}
+
+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 * GGML_RESTRICT x = vx;
+    const block_q8_K * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+#if defined __wasm_simd128__
+    float sumf = 0;
+
+    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;
+
+        // 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);
+
+            v128_t sc_low = wasm_u16x8_extend_low_u8x16(sc_upper);
+            v128_t sc_high = wasm_u16x8_extend_high_u8x16(sc_upper);
+
+            v128_t bsums1 = wasm_v128_load(&y[i].bsums[0]);
+            v128_t bsums2 = wasm_v128_load(&y[i].bsums[8]);
+
+            summs_vec = wasm_i32x4_add(
+                wasm_i32x4_add(wasm_i32x4_dot_i16x8(sc_low, bsums1),
+                               wasm_i32x4_dot_i16x8(sc_high, bsums2)),
+                summs_vec
+            );
+
+            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);
+
+        // Vectorized isum calculation
+        int32_t isum = 0;
+        const uint8_t * sc_ptr = sc;
+        const int k_iters = QK_K/128;
+
+        for (int k = 0; k < k_iters; ++k) {
+            v128_t isum_vec = wasm_i32x4_splat(0);
+            int shift = 0;
+
+            for (int j = 0; j < 4; ++j) {
+                const int d0 = (sc_ptr[0] & 0xF);
+                const int d1 = (sc_ptr[1] & 0xF);
+                sc_ptr += 2;
+
+                // 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));
+
+                // 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));
+
+                // 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)
+                );
+
+                // 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))
+                );
+
+                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);
+        }
+
+        const float dall = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
+        sumf += dall * isum - dmin * summs;
+    }
+
+    *s = sumf;
+
+#else
+
+    float sumf = 0;
+
+    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;
+
+        int summs = 0;
+        for (int j = 0; j < 16; ++j) {
+            summs += y[i].bsums[j] * (sc[j] >> 4);
+        }
+
+        const float dall = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+        const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
+
+        int isum = 0;
+        int is = 0;
+        int d;
+        for (int k = 0; k < QK_K/128; ++k) {
+            int shift = 0;
+            for (int j = 0; j < 4; ++j) {
+                d = sc[is++] & 0xF;
+                int isuml = 0;
+                for (int l =  0; l < 16; ++l) isuml += q8[l] * ((q2[l] >> shift) & 3);
+                isum += d * isuml;
+                d = sc[is++] & 0xF;
+                isuml = 0;
+                for (int l = 16; l < 32; ++l) isuml += q8[l] * ((q2[l] >> shift) & 3);
+                isum += d * isuml;
+                shift += 2;
+                q8 += 32;
+            }
+            q2 += 32;
+        }
+        sumf += dall * isum - dmin * summs;
+    }
+    *s = sumf;
+#endif
+}
+
+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);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const uint32_t kmask1 = 0x03030303;
+    const uint32_t kmask2 = 0x0f0f0f0f;
+
+    const block_q3_K * GGML_RESTRICT x = vx;
+    const block_q8_K * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+#if 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_CPU_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;
+
+#else
+    // scalar version
+    // This function is written like this so the compiler can manage to vectorize most of it
+    // Using -Ofast, GCC and clang manage to produce code that is within a factor of 2 or so from the
+    // manually vectorized version above. Every other version I tried would run at least 4 times slower.
+    // The ideal situation would be if we could just write the code once, and the compiler would
+    // automatically produce the best possible set of machine instructions, instead of us having to manually
+    // write vectorized versions for AVX, ARM_NEON, etc.
+
+    int8_t  aux8[QK_K];
+    int16_t aux16[8];
+    float   sums [8];
+    int32_t aux32[8];
+    memset(sums, 0, 8*sizeof(float));
+
+    uint32_t auxs[4];
+    const int8_t * scales = (const int8_t*)auxs;
+
+    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;
+        memset(aux32, 0, 8*sizeof(int32_t));
+        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;
+            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
+            a += 32; m <<= 1;
+            for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 2) & 3;
+            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
+            a += 32; m <<= 1;
+            for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 4) & 3;
+            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
+            a += 32; m <<= 1;
+            for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 6) & 3;
+            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
+            a += 32; m <<= 1;
+            q3 += 32;
+        }
+        a = aux8;
+
+        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);
+        for (int j = 0; j < QK_K/16; ++j) {
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
+            q8 += 8; a += 8;
+        }
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
+        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
+    }
+    for (int l = 0; l < 8; ++l) sumf += sums[l];
+    *s = sumf;
+
+#endif
+
+}
+
+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);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q4_K * GGML_RESTRICT x = vx;
+    const block_q8_K * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+    static const uint32_t kmask1 = 0x3f3f3f3f;
+    static const uint32_t kmask2 = 0x0f0f0f0f;
+    static const uint32_t kmask3 = 0x03030303;
+
+    uint32_t utmp[4];
+
+#if 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_CPU_FP16_TO_FP32(x[i].d);
+        const float dmin = y[i].d * GGML_CPU_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;
+
+#else
+
+    const uint8_t * scales = (const uint8_t*)&utmp[0];
+    const uint8_t * mins   = (const uint8_t*)&utmp[2];
+
+    int8_t  aux8[QK_K];
+    int16_t aux16[8];
+    float   sums [8];
+    int32_t aux32[8];
+    memset(sums, 0, 8*sizeof(float));
+
+    float sumf = 0;
+    for (int i = 0; i < nb; ++i) {
+        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 * 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;
+            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l]  >> 4);
+            a += 32; q4 += 32;
+        }
+        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;
+
+        int sumi = 0;
+        for (int j = 0; j < QK_K/16; ++j) sumi += y[i].bsums[j] * mins[j/2];
+        a = aux8;
+        int is = 0;
+        for (int j = 0; j < QK_K/32; ++j) {
+            int32_t scale = scales[is++];
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+        }
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
+        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
+        const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
+        sumf -= dmin * sumi;
+    }
+    for (int l = 0; l < 8; ++l) sumf += sums[l];
+    *s = sumf;
+#endif
+}
+
+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);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q5_K * GGML_RESTRICT x = vx;
+    const block_q8_K * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+    static const uint32_t kmask1 = 0x3f3f3f3f;
+    static const uint32_t kmask2 = 0x0f0f0f0f;
+    static const uint32_t kmask3 = 0x03030303;
+
+    uint32_t utmp[4];
+
+#if 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_CPU_FP16_TO_FP32(x[i].d);
+        const float dmin = y[i].d * GGML_CPU_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;
+
+#else
+
+    const uint8_t * scales = (const uint8_t*)&utmp[0];
+    const uint8_t * mins   = (const uint8_t*)&utmp[2];
+
+    int8_t  aux8[QK_K];
+    int16_t aux16[8];
+    float   sums [8];
+    int32_t aux32[8];
+    memset(sums, 0, 8*sizeof(float));
+
+    float sumf = 0;
+    for (int i = 0; i < nb; ++i) {
+        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 * 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);
+            for (int l = 0; l < 32; ++l) a[l] += (hm[l] & m ? 16 : 0);
+            a += 32; m <<= 1;
+            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l]  >> 4);
+            for (int l = 0; l < 32; ++l) a[l] += (hm[l] & m ? 16 : 0);
+            a += 32; m <<= 1;
+            q4 += 32;
+        }
+        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;
+
+        int sumi = 0;
+        for (int j = 0; j < QK_K/16; ++j) sumi += y[i].bsums[j] * mins[j/2];
+        a = aux8;
+        int is = 0;
+        for (int j = 0; j < QK_K/32; ++j) {
+            int32_t scale = scales[is++];
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+        }
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
+        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
+        const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
+        sumf -= dmin * sumi;
+    }
+    for (int l = 0; l < 8; ++l) sumf += sums[l];
+    *s = sumf;
+#endif
+}
+
+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);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q6_K * GGML_RESTRICT x = vx;
+    const block_q8_K * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+#if 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};
+
+    for (int i = 0; i < nb; ++i) {
+        // 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;
+        }
+
+        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_CPU_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;
+
+#else
+
+    int8_t  aux8[QK_K];
+    int16_t aux16[8];
+    float   sums [8];
+    int32_t aux32[8];
+    memset(sums, 0, 8*sizeof(float));
+
+    float sumf = 0;
+    for (int i = 0; i < nb; ++i) {
+        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 * 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;
+                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;
+        }
+        a = aux8;
+        int is = 0;
+        for (int j = 0; j < QK_K/16; ++j) {
+            int scale = x[i].scales[is++];
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+        }
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
+        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
+    }
+    for (int l = 0; l < 8; ++l) sumf += sums[l];
+    *s = sumf;
+#endif
+}
+
diff --git a/ggml/src/ggml-cpu/cpu-feats-x86.cpp b/ggml/src/ggml-cpu/arch/x86/cpu-feats.cpp
similarity index 100%
rename from ggml/src/ggml-cpu/cpu-feats-x86.cpp
rename to ggml/src/ggml-cpu/arch/x86/cpu-feats.cpp
diff --git a/ggml/src/ggml-cpu/arch/x86/quants.c b/ggml/src/ggml-cpu/arch/x86/quants.c
new file mode 100644
index 000000000..e7527c00a
--- /dev/null
+++ b/ggml/src/ggml-cpu/arch/x86/quants.c
@@ -0,0 +1,4311 @@
+#define GGML_COMMON_IMPL_C
+#include "ggml-common.h"
+#include "ggml-quants.h"
+#include "ggml-impl.h"
+#include "ggml-cpu.h"
+#include "simd-mappings.h"
+
+#include "../../quants.h"
+#include "../../ggml-cpu-impl.h"
+
+#include <math.h>
+#include <string.h>
+#include <assert.h>
+#include <stdlib.h> // for qsort
+#include <stdio.h>  // for GGML_ASSERT
+
+#define GROUP_MAX_EPS 1e-15f
+#define GROUP_MAX_EPS_IQ3_XXS 1e-8f
+#define GROUP_MAX_EPS_IQ2_S 1e-8f
+#define GROUP_MAX_EPS_IQ1_M 1e-7f
+#define GROUP_MAX_EPS_IQ1_S 1e-12f
+
+#define UNUSED GGML_UNUSED
+
+// some compilers don't provide _mm256_set_m128i, e.g. gcc 7
+#define MM256_SET_M128I(a, b) _mm256_insertf128_si256(_mm256_castsi128_si256(b), (a), 1)
+
+#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__) || defined(__SSSE3__)
+// 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
+    const __m128i ax = _mm_sign_epi8(x, x);
+    // Sign the values of the y vectors
+    const __m128i sy = _mm_sign_epi8(y, x);
+    // Perform multiplication and create 16-bit values
+    const __m128i dot = _mm_maddubs_epi16(ax, sy);
+    const __m128i ones = _mm_set1_epi16(1);
+    return _mm_madd_epi16(ones, dot);
+}
+
+#if __AVX__ || __AVX2__ || __AVX512F__
+// horizontally add 8 floats
+static inline float hsum_float_8(const __m256 x) {
+    __m128 res = _mm256_extractf128_ps(x, 1);
+    res = _mm_add_ps(res, _mm256_castps256_ps128(x));
+    res = _mm_add_ps(res, _mm_movehl_ps(res, res));
+    res = _mm_add_ss(res, _mm_movehdup_ps(res));
+    return _mm_cvtss_f32(res);
+}
+
+// horizontally add 8 int32_t
+static inline int hsum_i32_8(const __m256i a) {
+    const __m128i sum128 = _mm_add_epi32(_mm256_castsi256_si128(a), _mm256_extractf128_si256(a, 1));
+    const __m128i hi64 = _mm_unpackhi_epi64(sum128, sum128);
+    const __m128i sum64 = _mm_add_epi32(hi64, sum128);
+    const __m128i hi32  = _mm_shuffle_epi32(sum64, _MM_SHUFFLE(2, 3, 0, 1));
+    return _mm_cvtsi128_si32(_mm_add_epi32(sum64, hi32));
+}
+
+// horizontally add 4 int32_t
+static inline int hsum_i32_4(const __m128i a) {
+    const __m128i hi64 = _mm_unpackhi_epi64(a, a);
+    const __m128i sum64 = _mm_add_epi32(hi64, a);
+    const __m128i hi32  = _mm_shuffle_epi32(sum64, _MM_SHUFFLE(2, 3, 0, 1));
+    return _mm_cvtsi128_si32(_mm_add_epi32(sum64, hi32));
+}
+
+#if defined(__AVX2__) || defined(__AVX512F__)
+// spread 32 bits to 32 bytes { 0x00, 0xFF }
+static inline __m256i bytes_from_bits_32(const uint8_t * x) {
+    uint32_t x32;
+    memcpy(&x32, x, sizeof(uint32_t));
+    const __m256i shuf_mask = _mm256_set_epi64x(
+            0x0303030303030303, 0x0202020202020202,
+            0x0101010101010101, 0x0000000000000000);
+    __m256i bytes = _mm256_shuffle_epi8(_mm256_set1_epi32(x32), shuf_mask);
+    const __m256i bit_mask = _mm256_set1_epi64x(0x7fbfdfeff7fbfdfe);
+    bytes = _mm256_or_si256(bytes, bit_mask);
+    return _mm256_cmpeq_epi8(bytes, _mm256_set1_epi64x(-1));
+}
+
+// Unpack 32 4-bit fields into 32 bytes
+// The output vector contains 32 bytes, each one in [ 0 .. 15 ] interval
+static inline __m256i bytes_from_nibbles_32(const uint8_t * rsi)
+{
+    const __m128i tmp = _mm_loadu_si128((const __m128i *)rsi);
+    const __m256i bytes = MM256_SET_M128I(_mm_srli_epi16(tmp, 4), tmp);
+    const __m256i lowMask = _mm256_set1_epi8( 0xF );
+    return _mm256_and_si256(lowMask, bytes);
+}
+
+// add int16_t pairwise and return as float vector
+static inline __m256 sum_i16_pairs_float(const __m256i x) {
+    const __m256i ones = _mm256_set1_epi16(1);
+    const __m256i summed_pairs = _mm256_madd_epi16(ones, x);
+    return _mm256_cvtepi32_ps(summed_pairs);
+}
+
+static inline __m256 mul_sum_us8_pairs_float(const __m256i ax, const __m256i sy) {
+#if defined(__AVX512VNNI__) && defined(__AVX512VL__)
+    const __m256i zero = _mm256_setzero_si256();
+    const __m256i summed_pairs = _mm256_dpbusd_epi32(zero, ax, sy);
+    return _mm256_cvtepi32_ps(summed_pairs);
+#elif defined(__AVXVNNI__)
+    const __m256i zero = _mm256_setzero_si256();
+    const __m256i summed_pairs = _mm256_dpbusd_avx_epi32(zero, ax, sy);
+    return _mm256_cvtepi32_ps(summed_pairs);
+#else
+    // Perform multiplication and create 16-bit values
+    const __m256i dot = _mm256_maddubs_epi16(ax, sy);
+    return sum_i16_pairs_float(dot);
+#endif
+}
+
+// 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) {
+#if __AVXVNNIINT8__
+    const __m256i zero = _mm256_setzero_si256();
+    const __m256i summed_pairs = _mm256_dpbssd_epi32(zero, x, y);
+    return _mm256_cvtepi32_ps(summed_pairs);
+#else
+    // Get absolute values of x vectors
+    const __m256i ax = _mm256_sign_epi8(x, x);
+    // Sign the values of the y vectors
+    const __m256i sy = _mm256_sign_epi8(y, x);
+    return mul_sum_us8_pairs_float(ax, sy);
+#endif
+}
+
+static inline __m128i packNibbles( __m256i bytes )
+{
+    // Move bits within 16-bit lanes from 0000_abcd_0000_efgh into 0000_0000_abcd_efgh
+#if __AVX512F__
+    const __m256i bytes_srli_4 = _mm256_srli_epi16(bytes, 4);   // 0000_0000_abcd_0000
+    bytes = _mm256_or_si256(bytes, bytes_srli_4);               // 0000_abcd_abcd_efgh
+    return _mm256_cvtepi16_epi8(bytes);                         // abcd_efgh
+#else
+    const __m256i lowByte = _mm256_set1_epi16( 0xFF );
+    __m256i high = _mm256_andnot_si256( lowByte, bytes );
+    __m256i low = _mm256_and_si256( lowByte, bytes );
+    high = _mm256_srli_epi16( high, 4 );
+    bytes = _mm256_or_si256( low, high );
+
+    // Compress uint16_t lanes into bytes
+    __m128i r0 = _mm256_castsi256_si128( bytes );
+    __m128i r1 = _mm256_extracti128_si256( bytes, 1 );
+    return _mm_packus_epi16( r0, r1 );
+#endif
+}
+#elif defined(__AVX__)
+static inline __m128i packNibbles( __m128i bytes1, __m128i bytes2 )
+{
+    // Move bits within 16-bit lanes from 0000_abcd_0000_efgh into 0000_0000_abcd_efgh
+    const __m128i lowByte = _mm_set1_epi16( 0xFF );
+    __m128i high = _mm_andnot_si128( lowByte, bytes1 );
+    __m128i low = _mm_and_si128( lowByte, bytes1 );
+    high = _mm_srli_epi16( high, 4 );
+    bytes1 = _mm_or_si128( low, high );
+    high = _mm_andnot_si128( lowByte, bytes2 );
+    low = _mm_and_si128( lowByte, bytes2 );
+    high = _mm_srli_epi16( high, 4 );
+    bytes2 = _mm_or_si128( low, high );
+
+    return _mm_packus_epi16( bytes1, bytes2);
+}
+
+static inline __m128i mul_add_epi8_sse(const __m128i x, const __m128i y) {
+    const __m128i ax = _mm_sign_epi8(x, x);
+    const __m128i sy = _mm_sign_epi8(y, x);
+    return _mm_maddubs_epi16(ax, sy);
+}
+
+// spread 32 bits to 32 bytes { 0x00, 0xFF }
+static inline __m256i bytes_from_bits_32(const uint8_t * x) {
+    uint32_t x32;
+    memcpy(&x32, x, sizeof(uint32_t));
+    const __m128i shuf_maskl = _mm_set_epi64x(0x0101010101010101, 0x0000000000000000);
+    const __m128i shuf_maskh = _mm_set_epi64x(0x0303030303030303, 0x0202020202020202);
+    __m128i bytesl = _mm_shuffle_epi8(_mm_set1_epi32(x32), shuf_maskl);
+    __m128i bytesh = _mm_shuffle_epi8(_mm_set1_epi32(x32), shuf_maskh);
+    const __m128i bit_mask = _mm_set1_epi64x(0x7fbfdfeff7fbfdfe);
+    bytesl = _mm_or_si128(bytesl, bit_mask);
+    bytesh = _mm_or_si128(bytesh, bit_mask);
+    bytesl = _mm_cmpeq_epi8(bytesl, _mm_set1_epi64x(-1));
+    bytesh = _mm_cmpeq_epi8(bytesh, _mm_set1_epi64x(-1));
+    return MM256_SET_M128I(bytesh, bytesl);
+}
+
+// Unpack 32 4-bit fields into 32 bytes
+// The output vector contains 32 bytes, each one in [ 0 .. 15 ] interval
+static inline __m256i bytes_from_nibbles_32(const uint8_t * rsi)
+{
+    // Load 16 bytes from memory
+    __m128i tmpl = _mm_loadu_si128((const __m128i *)rsi);
+    __m128i tmph = _mm_srli_epi16(tmpl, 4);
+    const __m128i lowMask = _mm_set1_epi8(0xF);
+    tmpl = _mm_and_si128(lowMask, tmpl);
+    tmph = _mm_and_si128(lowMask, tmph);
+    return MM256_SET_M128I(tmph, tmpl);
+}
+
+// add int16_t pairwise and return as float vector
+static inline __m256 sum_i16_pairs_float(const __m128i xh, const __m128i xl) {
+    const __m128i ones = _mm_set1_epi16(1);
+    const __m128i summed_pairsl = _mm_madd_epi16(ones, xl);
+    const __m128i summed_pairsh = _mm_madd_epi16(ones, xh);
+    const __m256i summed_pairs = MM256_SET_M128I(summed_pairsh, summed_pairsl);
+    return _mm256_cvtepi32_ps(summed_pairs);
+}
+
+static inline __m256 mul_sum_us8_pairs_float(const __m256i ax, const __m256i sy) {
+    const __m128i axl = _mm256_castsi256_si128(ax);
+    const __m128i axh = _mm256_extractf128_si256(ax, 1);
+    const __m128i syl = _mm256_castsi256_si128(sy);
+    const __m128i syh = _mm256_extractf128_si256(sy, 1);
+    // Perform multiplication and create 16-bit values
+    const __m128i dotl = _mm_maddubs_epi16(axl, syl);
+    const __m128i doth = _mm_maddubs_epi16(axh, syh);
+    return sum_i16_pairs_float(doth, dotl);
+}
+
+// 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) {
+    const __m128i xl = _mm256_castsi256_si128(x);
+    const __m128i xh = _mm256_extractf128_si256(x, 1);
+    const __m128i yl = _mm256_castsi256_si128(y);
+    const __m128i yh = _mm256_extractf128_si256(y, 1);
+    // Get absolute values of x vectors
+    const __m128i axl = _mm_sign_epi8(xl, xl);
+    const __m128i axh = _mm_sign_epi8(xh, xh);
+    // Sign the values of the y vectors
+    const __m128i syl = _mm_sign_epi8(yl, xl);
+    const __m128i syh = _mm_sign_epi8(yh, xh);
+    // Perform multiplication and create 16-bit values
+    const __m128i dotl = _mm_maddubs_epi16(axl, syl);
+    const __m128i doth = _mm_maddubs_epi16(axh, syh);
+    return sum_i16_pairs_float(doth, dotl);
+}
+
+// larger version of mul_sum_i8_pairs_float where x and y are each represented by four 128-bit vectors
+static inline __m256 mul_sum_i8_quad_float(const __m128i x_1_0, const __m128i x_1_1, const __m128i x_2_0, const __m128i x_2_1,
+                                           const __m128i y_1_0, const __m128i y_1_1, const __m128i y_2_0, const __m128i y_2_1) {
+    const __m128i mone = _mm_set1_epi16(1);
+
+    const __m128i p16_1_0 = mul_add_epi8_sse(x_1_0, y_1_0);
+    const __m128i p16_1_1 = mul_add_epi8_sse(x_1_1, y_1_1);
+    const __m128i p16_2_0 = mul_add_epi8_sse(x_2_0, y_2_0);
+    const __m128i p16_2_1 = mul_add_epi8_sse(x_2_1, y_2_1);
+    const __m128i p_1_0 = _mm_madd_epi16(p16_1_0, mone);
+    const __m128i p_1_1 = _mm_madd_epi16(p16_1_1, mone);
+    const __m128i p_2_0 = _mm_madd_epi16(p16_2_0, mone);
+    const __m128i p_2_1 = _mm_madd_epi16(p16_2_1, mone);
+    const __m128i p_1 = _mm_add_epi32(p_1_0, p_1_1);
+    const __m128i p_2 = _mm_add_epi32(p_2_0, p_2_1);
+    return _mm256_cvtepi32_ps(MM256_SET_M128I(p_2, p_1));
+}
+
+// quad fp16 delta calculation
+static inline __m256 quad_fp16_delta_float(const float x0, const float y0, const float x1, const float y1) {
+    // GGML_CPU_FP16_TO_FP32 is faster than Intel F16C
+    return _mm256_set_m128(_mm_set1_ps(GGML_CPU_FP16_TO_FP32(x1) * GGML_CPU_FP16_TO_FP32(y1)),
+                           _mm_set1_ps(GGML_CPU_FP16_TO_FP32(x0) * GGML_CPU_FP16_TO_FP32(y0)));
+}
+#endif
+#elif defined(__SSSE3__)
+// horizontally add 4x4 floats
+static inline float hsum_float_4x4(const __m128 a, const __m128 b, const __m128 c, const __m128 d) {
+    __m128 res_0 =_mm_hadd_ps(a, b);
+    __m128 res_1 =_mm_hadd_ps(c, d);
+    __m128 res =_mm_hadd_ps(res_0, res_1);
+    res =_mm_hadd_ps(res, res);
+    res =_mm_hadd_ps(res, res);
+
+    return _mm_cvtss_f32(res);
+}
+#endif // __AVX__ || __AVX2__ || __AVX512F__
+#endif // defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__) || defined(__SSSE3__)
+
+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 * GGML_RESTRICT y = vy;
+
+#if defined(__AVX2__) || defined(__AVX__)
+    for (int i = 0; i < nb; i++) {
+        // Load elements into 4 AVX vectors
+        __m256 v0 = _mm256_loadu_ps( x );
+        __m256 v1 = _mm256_loadu_ps( x + 8 );
+        __m256 v2 = _mm256_loadu_ps( x + 16 );
+        __m256 v3 = _mm256_loadu_ps( x + 24 );
+        x += 32;
+
+        // Compute max(abs(e)) for the block
+        const __m256 signBit = _mm256_set1_ps( -0.0f );
+        __m256 maxAbs = _mm256_andnot_ps( signBit, v0 );
+        maxAbs = _mm256_max_ps( maxAbs, _mm256_andnot_ps( signBit, v1 ) );
+        maxAbs = _mm256_max_ps( maxAbs, _mm256_andnot_ps( signBit, v2 ) );
+        maxAbs = _mm256_max_ps( maxAbs, _mm256_andnot_ps( signBit, 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 );
+
+        // Quantize these floats
+        const float d = maxScalar / 127.f;
+        y[i].d = GGML_CPU_FP32_TO_FP16(d);
+        const float id = ( maxScalar != 0.0f ) ? 127.f / maxScalar : 0.0f;
+        const __m256 mul = _mm256_set1_ps( id );
+
+        // Apply the multiplier
+        v0 = _mm256_mul_ps( v0, mul );
+        v1 = _mm256_mul_ps( v1, mul );
+        v2 = _mm256_mul_ps( v2, mul );
+        v3 = _mm256_mul_ps( v3, mul );
+
+        // 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 );
+
+#if defined(__AVX2__)
+        // Convert int32 to int16
+        i0 = _mm256_packs_epi32( i0, i1 );	// 0, 1, 2, 3,  8, 9, 10, 11,  4, 5, 6, 7, 12, 13, 14, 15
+        i2 = _mm256_packs_epi32( i2, i3 );	// 16, 17, 18, 19,  24, 25, 26, 27,  20, 21, 22, 23, 28, 29, 30, 31
+                                            // Convert int16 to int8
+        i0 = _mm256_packs_epi16( i0, i2 );	// 0, 1, 2, 3,  8, 9, 10, 11,  16, 17, 18, 19,  24, 25, 26, 27,  4, 5, 6, 7, 12, 13, 14, 15, 20, 21, 22, 23, 28, 29, 30, 31
+
+        // We got our precious signed bytes, but the order is now wrong
+        // These AVX2 pack instructions process 16-byte pieces independently
+        // The following instruction is fixing the order
+        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, i0);
+#else
+        // Since we don't have in AVX some necessary functions,
+        // we split the registers in half and call AVX2 analogs from SSE
+        __m128i ni0 = _mm256_castsi256_si128( i0 );
+        __m128i ni1 = _mm256_extractf128_si256( i0, 1);
+        __m128i ni2 = _mm256_castsi256_si128( i1 );
+        __m128i ni3 = _mm256_extractf128_si256( i1, 1);
+        __m128i ni4 = _mm256_castsi256_si128( i2 );
+        __m128i ni5 = _mm256_extractf128_si256( i2, 1);
+        __m128i ni6 = _mm256_castsi256_si128( i3 );
+        __m128i ni7 = _mm256_extractf128_si256( i3, 1);
+
+        // Convert int32 to int16
+        ni0 = _mm_packs_epi32( ni0, ni1 );
+        ni2 = _mm_packs_epi32( ni2, ni3 );
+        ni4 = _mm_packs_epi32( ni4, ni5 );
+        ni6 = _mm_packs_epi32( ni6, ni7 );
+        // Convert int16 to int8
+        ni0 = _mm_packs_epi16( ni0, ni2 );
+        ni4 = _mm_packs_epi16( ni4, ni6 );
+
+        _mm_storeu_si128((__m128i *)(y[i].qs +  0), ni0);
+        _mm_storeu_si128((__m128i *)(y[i].qs + 16), ni4);
+#endif
+    }
+#else
+    GGML_UNUSED(nb);
+    // scalar
+    quantize_row_q8_0_ref(x, y, k);
+#endif
+}
+
+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 * GGML_RESTRICT y = vy;
+#if defined(__AVX2__) || defined(__AVX__)
+    for (int i = 0; i < nb; i++) {
+        // Load elements into 4 AVX vectors
+        __m256 v0 = _mm256_loadu_ps( x );
+        __m256 v1 = _mm256_loadu_ps( x + 8 );
+        __m256 v2 = _mm256_loadu_ps( x + 16 );
+        __m256 v3 = _mm256_loadu_ps( x + 24 );
+        x += 32;
+
+        // Compute max(abs(e)) for the block
+        const __m256 signBit = _mm256_set1_ps( -0.0f );
+        __m256 maxAbs = _mm256_andnot_ps( signBit, v0 );
+        maxAbs = _mm256_max_ps( maxAbs, _mm256_andnot_ps( signBit, v1 ) );
+        maxAbs = _mm256_max_ps( maxAbs, _mm256_andnot_ps( signBit, v2 ) );
+        maxAbs = _mm256_max_ps( maxAbs, _mm256_andnot_ps( signBit, 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 max_scalar = _mm_cvtss_f32( max4 );
+
+        // Quantize these floats
+        const float d = max_scalar / 127.f;
+        y[i].d = GGML_CPU_FP32_TO_FP16(d);
+        const float id = ( max_scalar != 0.0f ) ? 127.f / max_scalar : 0.0f;
+        const __m256 mul = _mm256_set1_ps( id );
+
+        // Apply the multiplier
+        v0 = _mm256_mul_ps( v0, mul );
+        v1 = _mm256_mul_ps( v1, mul );
+        v2 = _mm256_mul_ps( v2, mul );
+        v3 = _mm256_mul_ps( v3, mul );
+
+        // 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 );
+
+#if defined(__AVX2__)
+        // Compute the sum of the quants and set y[i].s
+        y[i].s = GGML_CPU_FP32_TO_FP16(d * hsum_i32_8(_mm256_add_epi32(_mm256_add_epi32(i0, i1), _mm256_add_epi32(i2, i3))));
+
+        // Convert int32 to int16
+        i0 = _mm256_packs_epi32( i0, i1 );	// 0, 1, 2, 3,  8, 9, 10, 11,  4, 5, 6, 7, 12, 13, 14, 15
+        i2 = _mm256_packs_epi32( i2, i3 );	// 16, 17, 18, 19,  24, 25, 26, 27,  20, 21, 22, 23, 28, 29, 30, 31
+                                            // Convert int16 to int8
+        i0 = _mm256_packs_epi16( i0, i2 );	// 0, 1, 2, 3,  8, 9, 10, 11,  16, 17, 18, 19,  24, 25, 26, 27,  4, 5, 6, 7, 12, 13, 14, 15, 20, 21, 22, 23, 28, 29, 30, 31
+
+        // We got our precious signed bytes, but the order is now wrong
+        // These AVX2 pack instructions process 16-byte pieces independently
+        // The following instruction is fixing the order
+        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, i0);
+#else
+        // Since we don't have in AVX some necessary functions,
+        // we split the registers in half and call AVX2 analogs from SSE
+        __m128i ni0 = _mm256_castsi256_si128( i0 );
+        __m128i ni1 = _mm256_extractf128_si256( i0, 1);
+        __m128i ni2 = _mm256_castsi256_si128( i1 );
+        __m128i ni3 = _mm256_extractf128_si256( i1, 1);
+        __m128i ni4 = _mm256_castsi256_si128( i2 );
+        __m128i ni5 = _mm256_extractf128_si256( i2, 1);
+        __m128i ni6 = _mm256_castsi256_si128( i3 );
+        __m128i ni7 = _mm256_extractf128_si256( i3, 1);
+
+        // Compute the sum of the quants and set y[i].s
+        const __m128i s0 = _mm_add_epi32(_mm_add_epi32(ni0, ni1), _mm_add_epi32(ni2, ni3));
+        const __m128i s1 = _mm_add_epi32(_mm_add_epi32(ni4, ni5), _mm_add_epi32(ni6, ni7));
+        y[i].s = GGML_CPU_FP32_TO_FP16(d * hsum_i32_4(_mm_add_epi32(s0, s1)));
+
+        // Convert int32 to int16
+        ni0 = _mm_packs_epi32( ni0, ni1 );
+        ni2 = _mm_packs_epi32( ni2, ni3 );
+        ni4 = _mm_packs_epi32( ni4, ni5 );
+        ni6 = _mm_packs_epi32( ni6, ni7 );
+        // Convert int16 to int8
+        ni0 = _mm_packs_epi16( ni0, ni2 );
+        ni4 = _mm_packs_epi16( ni4, ni6 );
+
+        _mm_storeu_si128((__m128i *)(y[i].qs +  0), ni0);
+        _mm_storeu_si128((__m128i *)(y[i].qs + 16), ni4);
+#endif
+    }
+#else
+    GGML_UNUSED(nb);
+    // scalar
+    quantize_row_q8_1_ref(x, y, k);
+#endif
+}
+
+// placeholder implementation for Apple targets
+void quantize_row_q8_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k) {
+    quantize_row_q8_K_ref(x, y, k);
+}
+
+//===================================== Dot products =================================
+
+//
+// Helper functions
+//
+
+#if __AVX__ || __AVX2__ || __AVX512F__
+
+// shuffles to pick the required scales in dot products
+static inline __m256i get_scale_shuffle_q3k(int i) {
+    static const uint8_t k_shuffle[128] = {
+         0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1,     2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3,
+         4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5,     6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7,
+         8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9,    10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,
+        12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,    14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,
+    };
+    return _mm256_loadu_si256((const __m256i*)k_shuffle + i);
+}
+static inline __m256i get_scale_shuffle_k4(int i) {
+    static const uint8_t k_shuffle[256] = {
+         0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1,
+         2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3,
+         4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5,
+         6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7,
+         8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9,
+        10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,
+        12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,
+        14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15
+    };
+    return _mm256_loadu_si256((const __m256i*)k_shuffle + i);
+}
+static inline __m128i get_scale_shuffle(int i) {
+    static const uint8_t k_shuffle[128] = {
+         0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1,
+         2, 2, 2, 2, 2, 2, 2, 2, 3, 3, 3, 3, 3, 3, 3, 3,
+         4, 4, 4, 4, 4, 4, 4, 4, 5, 5, 5, 5, 5, 5, 5, 5,
+         6, 6, 6, 6, 6, 6, 6, 6, 7, 7, 7, 7, 7, 7, 7, 7,
+         8, 8, 8, 8, 8, 8, 8, 8, 9, 9, 9, 9, 9, 9, 9, 9,
+        10,10,10,10,10,10,10,10, 11,11,11,11,11,11,11,11,
+        12,12,12,12,12,12,12,12, 13,13,13,13,13,13,13,13,
+        14,14,14,14,14,14,14,14, 15,15,15,15,15,15,15,15
+    };
+    return _mm_loadu_si128((const __m128i*)k_shuffle + i);
+}
+#endif
+
+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;
+
+    assert(n % qk == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q4_0 * GGML_RESTRICT x = vx;
+    const block_q8_0 * GGML_RESTRICT y = vy;
+
+    int ib = 0;
+    float sumf = 0;
+
+#if defined(__AVX2__)
+    // Initialize accumulator with zeros
+    __m256 acc = _mm256_setzero_ps();
+
+    // Main loop
+    for (; ib < nb; ++ib) {
+        /* Compute combined scale for the block */
+        const __m256 d = _mm256_set1_ps( GGML_CPU_FP16_TO_FP32(x[ib].d) * GGML_CPU_FP16_TO_FP32(y[ib].d) );
+
+        __m256i qx = bytes_from_nibbles_32(x[ib].qs);
+
+        // Now we have a vector with bytes in [ 0 .. 15 ] interval. Offset them into [ -8 .. +7 ] interval.
+        const __m256i off = _mm256_set1_epi8( 8 );
+        qx = _mm256_sub_epi8( qx, off );
+
+        __m256i qy = _mm256_loadu_si256((const __m256i *)y[ib].qs);
+
+        const __m256 q = mul_sum_i8_pairs_float(qx, qy);
+
+        /* Multiply q with scale and accumulate */
+        acc = _mm256_fmadd_ps( d, q, acc );
+    }
+
+    sumf = hsum_float_8(acc);
+#elif defined(__AVX__)
+    __m256 accum = _mm256_setzero_ps();
+    for (; ib + 1 < nb; ib += 2) {
+        const __m128i q4bits_1 = _mm_loadu_si128((const __m128i *)x[ib + 0].qs);
+        const __m128i q4bits_2 = _mm_loadu_si128((const __m128i *)x[ib + 1].qs);
+        const __m128i q8b_1_0 = _mm_loadu_si128((const __m128i *)y[ib + 0].qs);
+        const __m128i q8b_1_1 = _mm_loadu_si128((const __m128i *)y[ib + 0].qs + 1);
+        const __m128i q8b_2_0 = _mm_loadu_si128((const __m128i *)y[ib + 1].qs);
+        const __m128i q8b_2_1 = _mm_loadu_si128((const __m128i *)y[ib + 1].qs + 1);
+
+        const __m128i q4b_1_0 = _mm_sub_epi8(_mm_and_si128(_mm_set1_epi8(15), q4bits_1), _mm_set1_epi8(8));
+        const __m128i q4b_1_1 = _mm_sub_epi8(_mm_and_si128(_mm_set1_epi8(15), _mm_srli_epi16(q4bits_1, 4)), _mm_set1_epi8(8));
+        const __m128i q4b_2_0 = _mm_sub_epi8(_mm_and_si128(_mm_set1_epi8(15), q4bits_2), _mm_set1_epi8(8));
+        const __m128i q4b_2_1 = _mm_sub_epi8(_mm_and_si128(_mm_set1_epi8(15), _mm_srli_epi16(q4bits_2, 4)), _mm_set1_epi8(8));
+
+        const __m128i p16_1_0 = mul_add_epi8_sse(q4b_1_0, q8b_1_0);
+        const __m128i p16_1_1 = mul_add_epi8_sse(q4b_1_1, q8b_1_1);
+        const __m128i p16_2_0 = mul_add_epi8_sse(q4b_2_0, q8b_2_0);
+        const __m128i p16_2_1 = mul_add_epi8_sse(q4b_2_1, q8b_2_1);
+        const __m128i p_1 = _mm_add_epi16(p16_1_0, p16_1_1);
+        const __m128i p_2 = _mm_add_epi16(p16_2_0, p16_2_1);
+        const __m256 p =  sum_i16_pairs_float(p_2, p_1);
+
+        const __m256 deltas = quad_fp16_delta_float(x[ib].d, y[ib].d, x[ib + 1].d, y[ib + 1].d);
+        accum = _mm256_add_ps(_mm256_mul_ps(deltas, p), accum);
+    }
+
+    sumf = hsum_float_8(accum);
+#elif defined(__SSSE3__)
+    // set constants
+    const __m128i lowMask = _mm_set1_epi8(0xF);
+    const __m128i off = _mm_set1_epi8(8);
+
+    // Initialize accumulator with zeros
+    __m128 acc_0 = _mm_setzero_ps();
+    __m128 acc_1 = _mm_setzero_ps();
+    __m128 acc_2 = _mm_setzero_ps();
+    __m128 acc_3 = _mm_setzero_ps();
+
+    for (; ib + 1 < nb; ib += 2) {
+        _mm_prefetch(&x[ib] + sizeof(block_q4_0), _MM_HINT_T0);
+        _mm_prefetch(&y[ib] + sizeof(block_q8_0), _MM_HINT_T0);
+
+        // Compute combined scale for the block 0 and 1
+        const __m128 d_0_1 = _mm_set1_ps( GGML_CPU_FP16_TO_FP32(x[ib].d) * GGML_CPU_FP16_TO_FP32(y[ib].d) );
+
+        const __m128i tmp_0_1 = _mm_loadu_si128((const __m128i *)x[ib].qs);
+
+        __m128i bx_0 = _mm_and_si128(lowMask, tmp_0_1);
+        __m128i by_0 = _mm_loadu_si128((const __m128i *)y[ib].qs);
+        bx_0 = _mm_sub_epi8(bx_0, off);
+        const __m128i i32_0 = mul_sum_i8_pairs(bx_0, by_0);
+
+        __m128i bx_1 = _mm_and_si128(lowMask, _mm_srli_epi64(tmp_0_1, 4));
+        __m128i by_1 = _mm_loadu_si128((const __m128i *)(y[ib].qs + 16));
+        bx_1 = _mm_sub_epi8(bx_1, off);
+        const __m128i i32_1 = mul_sum_i8_pairs(bx_1, by_1);
+
+        _mm_prefetch(&x[ib] + 2 * sizeof(block_q4_0), _MM_HINT_T0);
+        _mm_prefetch(&y[ib] + 2 * sizeof(block_q8_0), _MM_HINT_T0);
+
+        // Compute combined scale for the block 2 and 3
+        const __m128 d_2_3 = _mm_set1_ps( GGML_CPU_FP16_TO_FP32(x[ib + 1].d) * GGML_CPU_FP16_TO_FP32(y[ib + 1].d) );
+
+        const __m128i tmp_2_3 = _mm_loadu_si128((const __m128i *)x[ib + 1].qs);
+
+        __m128i bx_2 = _mm_and_si128(lowMask, tmp_2_3);
+        __m128i by_2 = _mm_loadu_si128((const __m128i *)y[ib + 1].qs);
+        bx_2 = _mm_sub_epi8(bx_2, off);
+        const __m128i i32_2 = mul_sum_i8_pairs(bx_2, by_2);
+
+        __m128i bx_3 = _mm_and_si128(lowMask, _mm_srli_epi64(tmp_2_3, 4));
+        __m128i by_3 = _mm_loadu_si128((const __m128i *)(y[ib + 1].qs + 16));
+        bx_3 = _mm_sub_epi8(bx_3, off);
+        const __m128i i32_3 = mul_sum_i8_pairs(bx_3, by_3);
+
+        // Convert int32_t to float
+        __m128 p0 = _mm_cvtepi32_ps(i32_0);
+        __m128 p1 = _mm_cvtepi32_ps(i32_1);
+        __m128 p2 = _mm_cvtepi32_ps(i32_2);
+        __m128 p3 = _mm_cvtepi32_ps(i32_3);
+
+        // Apply the scale
+        __m128 p0_d = _mm_mul_ps( d_0_1, p0 );
+        __m128 p1_d = _mm_mul_ps( d_0_1, p1 );
+        __m128 p2_d = _mm_mul_ps( d_2_3, p2 );
+        __m128 p3_d = _mm_mul_ps( d_2_3, p3 );
+
+        // Acummulate
+        acc_0 = _mm_add_ps(p0_d, acc_0);
+        acc_1 = _mm_add_ps(p1_d, acc_1);
+        acc_2 = _mm_add_ps(p2_d, acc_2);
+        acc_3 = _mm_add_ps(p3_d, acc_3);
+    }
+
+    sumf = hsum_float_4x4(acc_0, acc_1, acc_2, acc_3);
+
+#endif
+    for (; ib < nb; ++ib) {
+        int sumi0 = 0;
+        int sumi1 = 0;
+
+        for (int j = 0; j < qk/2; ++j) {
+            const int v0 = (x[ib].qs[j] & 0x0F) - 8;
+            const int v1 = (x[ib].qs[j] >>   4) - 8;
+
+            sumi0 += (v0 * y[ib].qs[j]);
+            sumi1 += (v1 * y[ib].qs[j + qk/2]);
+        }
+
+        int sumi = sumi0 + sumi1;
+        sumf += sumi*GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d);
+    }
+
+    *s = sumf;
+}
+
+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;
+
+    assert(n % qk == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q4_1 * GGML_RESTRICT x = vx;
+    const block_q8_1 * GGML_RESTRICT y = vy;
+
+    int ib = 0;
+    float sumf = 0;
+
+#if defined(__AVX2__) || defined(__AVX__)
+    // Initialize accumulator with zeros
+    __m256 acc = _mm256_setzero_ps();
+
+    float summs = 0;
+
+    // Main loop
+    for (; ib < nb; ++ib) {
+        const float d0 = GGML_CPU_FP16_TO_FP32(x[ib].d);
+        const float d1 = GGML_CPU_FP16_TO_FP32(y[ib].d);
+
+        summs += GGML_CPU_FP16_TO_FP32(x[ib].m) * GGML_CPU_FP16_TO_FP32(y[ib].s);
+
+        const __m256 d0v = _mm256_set1_ps( d0 );
+        const __m256 d1v = _mm256_set1_ps( d1 );
+
+        // Compute combined scales
+        const __m256 d0d1 = _mm256_mul_ps( d0v, d1v );
+
+        // Load 16 bytes, and unpack 4 bit fields into bytes, making 32 bytes
+        const __m256i qx = bytes_from_nibbles_32(x[ib].qs);
+        const __m256i qy = _mm256_loadu_si256( (const __m256i *)y[ib].qs );
+
+        const __m256 xy = mul_sum_us8_pairs_float(qx, qy);
+
+        // Accumulate d0*d1*x*y
+#if defined(__AVX2__)
+        acc = _mm256_fmadd_ps( d0d1, xy, acc );
+#else
+        acc = _mm256_add_ps( _mm256_mul_ps( d0d1, xy ), acc );
+#endif
+    }
+
+    sumf = hsum_float_8(acc) + summs;
+
+#endif
+    for (; ib < nb; ++ib) {
+        int sumi0 = 0;
+        int sumi1 = 0;
+
+        for (int j = 0; j < qk/2; ++j) {
+            const int v0 = (x[ib].qs[j] & 0x0F);
+            const int v1 = (x[ib].qs[j] >>   4);
+
+            sumi0 += (v0 * y[ib].qs[j]);
+            sumi1 += (v1 * y[ib].qs[j + qk/2]);
+        }
+
+        int sumi = sumi0 + sumi1;
+        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
+    }
+
+    *s = sumf;
+}
+
+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;
+
+    int ib = 0;
+    float sumf = 0;
+
+    assert(n % qk == 0);
+    assert(qk == QK5_0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q5_0 * GGML_RESTRICT x = vx;
+    const block_q8_0 * GGML_RESTRICT y = vy;
+
+#if defined(__AVX2__)
+    // Initialize accumulator with zeros
+    __m256 acc = _mm256_setzero_ps();
+
+    // Main loop
+    for (; ib < nb; ++ib) {
+        /* Compute combined scale for the block */
+        const __m256 d = _mm256_set1_ps(GGML_CPU_FP16_TO_FP32(x[ib].d) * GGML_CPU_FP16_TO_FP32(y[ib].d));
+
+        __m256i qx = bytes_from_nibbles_32(x[ib].qs);
+        __m256i bxhi = bytes_from_bits_32(x[ib].qh);
+        bxhi = _mm256_andnot_si256(bxhi, _mm256_set1_epi8((char)0xF0));
+        qx = _mm256_or_si256(qx, bxhi);
+
+        __m256i qy = _mm256_loadu_si256((const __m256i *)y[ib].qs);
+
+        const __m256 q = mul_sum_i8_pairs_float(qx, qy);
+
+        /* Multiply q with scale and accumulate */
+        acc = _mm256_fmadd_ps(d, q, acc);
+    }
+
+    sumf = hsum_float_8(acc);
+#elif defined(__AVX__)
+    // Initialize accumulator with zeros
+    __m256 acc = _mm256_setzero_ps();
+    __m128i mask = _mm_set1_epi8((char)0xF0);
+
+    // Main loop
+    for (; ib < nb; ++ib) {
+        /* Compute combined scale for the block */
+        const __m256 d = _mm256_set1_ps(GGML_CPU_FP16_TO_FP32(x[ib].d) * GGML_CPU_FP16_TO_FP32(y[ib].d));
+
+        __m256i bx_0 = bytes_from_nibbles_32(x[ib].qs);
+        const __m256i bxhi = bytes_from_bits_32(x[ib].qh);
+        __m128i bxhil = _mm256_castsi256_si128(bxhi);
+        __m128i bxhih = _mm256_extractf128_si256(bxhi, 1);
+        bxhil = _mm_andnot_si128(bxhil, mask);
+        bxhih = _mm_andnot_si128(bxhih, mask);
+        __m128i bxl = _mm256_castsi256_si128(bx_0);
+        __m128i bxh = _mm256_extractf128_si256(bx_0, 1);
+        bxl = _mm_or_si128(bxl, bxhil);
+        bxh = _mm_or_si128(bxh, bxhih);
+        bx_0 = MM256_SET_M128I(bxh, bxl);
+
+        const __m256i by_0 = _mm256_loadu_si256((const __m256i *)y[ib].qs);
+
+        const __m256 q = mul_sum_i8_pairs_float(bx_0, by_0);
+
+        /* Multiply q with scale and accumulate */
+        acc = _mm256_add_ps(_mm256_mul_ps(d, q), acc);
+    }
+
+    sumf = hsum_float_8(acc);
+
+#endif
+    for (; ib < nb; ++ib) {
+        uint32_t qh;
+        memcpy(&qh, x[ib].qh, sizeof(qh));
+
+        int sumi0 = 0;
+        int sumi1 = 0;
+
+        for (int j = 0; j < qk/2; ++j) {
+            const uint8_t xh_0 = ((qh & (1u << (j + 0 ))) >> (j + 0 )) << 4;
+            const uint8_t xh_1 = ((qh & (1u << (j + 16))) >> (j + 12));
+
+            const int32_t x0 = (int8_t)(((x[ib].qs[j] & 0x0F) | xh_0) - 16);
+            const int32_t x1 = (int8_t)(((x[ib].qs[j] >>   4) | xh_1) - 16);
+
+            sumi0 += (x0 * y[ib].qs[j]);
+            sumi1 += (x1 * y[ib].qs[j + qk/2]);
+        }
+
+        int sumi = sumi0 + sumi1;
+        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d)) * sumi;
+    }
+
+    *s = sumf;
+}
+
+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;
+
+    int ib = 0;
+    float sumf = 0;
+
+    assert(n % qk == 0);
+    assert(qk == QK5_1);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q5_1 * GGML_RESTRICT x = vx;
+    const block_q8_1 * GGML_RESTRICT y = vy;
+
+#if defined(__AVX2__)
+    // Initialize accumulator with zeros
+    __m256 acc = _mm256_setzero_ps();
+
+    float summs = 0.0f;
+
+    // Main loop
+    for (; ib < nb; ++ib) {
+        const __m256 dx = _mm256_set1_ps(GGML_CPU_FP16_TO_FP32(x[ib].d));
+
+        summs += GGML_CPU_FP16_TO_FP32(x[ib].m) * GGML_CPU_FP16_TO_FP32(y[ib].s);
+
+        __m256i qx = bytes_from_nibbles_32(x[ib].qs);
+        __m256i bxhi = bytes_from_bits_32(x[ib].qh);
+        bxhi = _mm256_and_si256(bxhi, _mm256_set1_epi8(0x10));
+        qx = _mm256_or_si256(qx, bxhi);
+
+        const __m256 dy = _mm256_set1_ps(GGML_CPU_FP16_TO_FP32(y[ib].d));
+        const __m256i qy = _mm256_loadu_si256((const __m256i *)y[ib].qs);
+
+        const __m256 q = mul_sum_us8_pairs_float(qx, qy);
+
+        acc = _mm256_fmadd_ps(q, _mm256_mul_ps(dx, dy), acc);
+    }
+
+    sumf = hsum_float_8(acc) + summs;
+#elif defined(__AVX__)
+    // Initialize accumulator with zeros
+    __m256 acc = _mm256_setzero_ps();
+    __m128i mask = _mm_set1_epi8(0x10);
+
+    float summs = 0.0f;
+
+    // Main loop
+    for (; ib < nb; ++ib) {
+        const __m256 dx = _mm256_set1_ps(GGML_CPU_FP16_TO_FP32(x[ib].d));
+
+        summs += GGML_CPU_FP16_TO_FP32(x[ib].m) * GGML_CPU_FP16_TO_FP32(y[ib].s);
+
+        __m256i bx_0 = bytes_from_nibbles_32(x[ib].qs);
+        const __m256i bxhi = bytes_from_bits_32(x[ib].qh);
+        __m128i bxhil = _mm256_castsi256_si128(bxhi);
+        __m128i bxhih = _mm256_extractf128_si256(bxhi, 1);
+        bxhil = _mm_and_si128(bxhil, mask);
+        bxhih = _mm_and_si128(bxhih, mask);
+        __m128i bxl = _mm256_castsi256_si128(bx_0);
+        __m128i bxh = _mm256_extractf128_si256(bx_0, 1);
+        bxl = _mm_or_si128(bxl, bxhil);
+        bxh = _mm_or_si128(bxh, bxhih);
+        bx_0 = MM256_SET_M128I(bxh, bxl);
+
+        const __m256 dy = _mm256_set1_ps(GGML_CPU_FP16_TO_FP32(y[ib].d));
+        const __m256i by_0 = _mm256_loadu_si256((const __m256i *)y[ib].qs);
+
+        const __m256 q = mul_sum_us8_pairs_float(bx_0, by_0);
+
+        acc = _mm256_add_ps(_mm256_mul_ps(q, _mm256_mul_ps(dx, dy)), acc);
+    }
+
+    sumf = hsum_float_8(acc) + summs;
+
+#endif
+    for (; ib < nb; ++ib) {
+        uint32_t qh;
+        memcpy(&qh, x[ib].qh, sizeof(qh));
+
+        int sumi0 = 0;
+        int sumi1 = 0;
+
+        for (int j = 0; j < qk/2; ++j) {
+            const uint8_t xh_0 = ((qh >> (j +  0)) << 4) & 0x10;
+            const uint8_t xh_1 = ((qh >> (j + 12))     ) & 0x10;
+
+            const int32_t x0 = (x[ib].qs[j] & 0xF) | xh_0;
+            const int32_t x1 = (x[ib].qs[j] >>  4) | xh_1;
+
+            sumi0 += (x0 * y[ib].qs[j]);
+            sumi1 += (x1 * y[ib].qs[j + qk/2]);
+        }
+
+        int sumi = sumi0 + sumi1;
+        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
+    }
+
+    *s = sumf;
+}
+
+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;
+
+    assert(n % qk == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q8_0 * GGML_RESTRICT x = vx;
+    const block_q8_0 * GGML_RESTRICT y = vy;
+
+    int ib = 0;
+    float sumf = 0;
+
+#if defined(__AVX2__)
+    // Initialize accumulator with zeros
+    __m256 acc = _mm256_setzero_ps();
+
+    // Main loop
+    for (; ib < nb; ++ib) {
+        // Compute combined scale for the block
+        const __m256 d = _mm256_set1_ps(GGML_CPU_FP16_TO_FP32(x[ib].d) * GGML_CPU_FP16_TO_FP32(y[ib].d));
+        __m256i qx = _mm256_loadu_si256((const __m256i *)x[ib].qs);
+        __m256i qy = _mm256_loadu_si256((const __m256i *)y[ib].qs);
+
+        const __m256 q = mul_sum_i8_pairs_float(qx, qy);
+
+        // Multiply q with scale and accumulate
+        acc = _mm256_fmadd_ps( d, q, acc );
+    }
+
+    sumf = hsum_float_8(acc);
+#elif defined(__AVX__)
+    __m256 accum = _mm256_setzero_ps();
+
+    for (; ib + 1 < nb; ib += 2) {
+        const __m128i qx_1_0 = _mm_loadu_si128((const __m128i *)x[ib].qs);
+        const __m128i qx_1_1 = _mm_loadu_si128((const __m128i *)x[ib].qs + 1);
+        const __m128i qx_2_0 = _mm_loadu_si128((const __m128i *)x[ib + 1].qs);
+        const __m128i qx_2_1 = _mm_loadu_si128((const __m128i *)x[ib + 1].qs + 1);
+        const __m128i qy_1_0 = _mm_loadu_si128((const __m128i *)y[ib].qs);
+        const __m128i qy_1_1 = _mm_loadu_si128((const __m128i *)y[ib].qs + 1);
+        const __m128i qy_2_0 = _mm_loadu_si128((const __m128i *)y[ib + 1].qs);
+        const __m128i qy_2_1 = _mm_loadu_si128((const __m128i *)y[ib + 1].qs + 1);
+
+        const __m256 p = mul_sum_i8_quad_float(qx_1_0, qx_1_1, qx_2_0, qx_2_1, qy_1_0, qy_1_1, qy_2_0, qy_2_1);
+        const __m256 deltas = quad_fp16_delta_float(x[ib].d, y[ib].d, x[ib + 1].d, y[ib + 1].d);
+        accum = _mm256_add_ps(_mm256_mul_ps(deltas, p), accum);
+    }
+
+    sumf = hsum_float_8(accum);
+
+#endif
+    for (; ib < nb; ++ib) {
+        int sumi = 0;
+
+        for (int j = 0; j < qk; j++) {
+            sumi += x[ib].qs[j]*y[ib].qs[j];
+        }
+
+        sumf += sumi*(GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d));
+    }
+
+    *s = sumf;
+}
+
+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 * GGML_RESTRICT x = vx;
+    const block_q8_K  * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+#if defined(__AVX2__)
+    __m256 sumf = _mm256_setzero_ps();
+
+    for (int i = 0; i < nb; ++i) {
+        // 16-bit sums
+        __m256i sumi0 = _mm256_setzero_si256();
+        __m256i sumi1 = _mm256_setzero_si256();
+        __m256i sumi2 = _mm256_setzero_si256();
+
+        // first 32 bytes of 5 elements
+        {
+            __m256i qx0 = _mm256_loadu_si256((const __m256i *) (x[i].qs));
+            // 8-bit multiplies with shifts, masks and adds
+            __m256i qx1 = _mm256_add_epi8(qx0, _mm256_add_epi8(qx0, qx0)); // 1 * 3
+            __m256i qx2 = _mm256_add_epi8(_mm256_and_si256(_mm256_slli_epi16(qx0, 3), _mm256_set1_epi8(-8)), qx0); // 1 * 9
+            __m256i qx3 = _mm256_add_epi8(_mm256_and_si256(_mm256_slli_epi16(qx1, 3), _mm256_set1_epi8(-8)), qx1); // 3 * 9
+            __m256i qx4 = _mm256_add_epi8(_mm256_and_si256(_mm256_slli_epi16(qx2, 3), _mm256_set1_epi8(-8)), qx2); // 9 * 9
+
+            // TODO: can _mm256_mulhi_epu16 be faster even if 16-bits?
+
+            // Cancel the +1 from avg so that it behaves like a halving add
+            qx0 = _mm256_subs_epu8(qx0, _mm256_set1_epi8(1));
+            qx1 = _mm256_subs_epu8(qx1, _mm256_set1_epi8(1));
+            qx2 = _mm256_subs_epu8(qx2, _mm256_set1_epi8(1));
+            qx3 = _mm256_subs_epu8(qx3, _mm256_set1_epi8(1));
+            qx4 = _mm256_subs_epu8(qx4, _mm256_set1_epi8(1));
+            // Multiply by 3 and get the top 2 bits
+            qx0 = _mm256_avg_epu8(qx0, _mm256_avg_epu8(qx0, _mm256_setzero_si256()));
+            qx1 = _mm256_avg_epu8(qx1, _mm256_avg_epu8(qx1, _mm256_setzero_si256()));
+            qx2 = _mm256_avg_epu8(qx2, _mm256_avg_epu8(qx2, _mm256_setzero_si256()));
+            qx3 = _mm256_avg_epu8(qx3, _mm256_avg_epu8(qx3, _mm256_setzero_si256()));
+            qx4 = _mm256_avg_epu8(qx4, _mm256_avg_epu8(qx4, _mm256_setzero_si256()));
+            qx0 = _mm256_and_si256(_mm256_srli_epi16(qx0, 6), _mm256_set1_epi8(3));
+            qx1 = _mm256_and_si256(_mm256_srli_epi16(qx1, 6), _mm256_set1_epi8(3));
+            qx2 = _mm256_and_si256(_mm256_srli_epi16(qx2, 6), _mm256_set1_epi8(3));
+            qx3 = _mm256_and_si256(_mm256_srli_epi16(qx3, 6), _mm256_set1_epi8(3));
+            qx4 = _mm256_and_si256(_mm256_srli_epi16(qx4, 6), _mm256_set1_epi8(3));
+
+            const __m256i qy0 = _mm256_loadu_si256((const __m256i *) (y[i].qs +   0));
+            const __m256i qy1 = _mm256_loadu_si256((const __m256i *) (y[i].qs +  32));
+            const __m256i qy2 = _mm256_loadu_si256((const __m256i *) (y[i].qs +  64));
+            const __m256i qy3 = _mm256_loadu_si256((const __m256i *) (y[i].qs +  96));
+            const __m256i qy4 = _mm256_loadu_si256((const __m256i *) (y[i].qs + 128));
+
+            qx0 = _mm256_maddubs_epi16(qx0, qy0);
+            qx1 = _mm256_maddubs_epi16(qx1, qy1);
+            qx2 = _mm256_maddubs_epi16(qx2, qy2);
+            qx3 = _mm256_maddubs_epi16(qx3, qy3);
+            qx4 = _mm256_maddubs_epi16(qx4, qy4);
+
+            sumi0 = _mm256_add_epi16(sumi0, _mm256_add_epi16(qx0, qx1));
+            sumi1 = _mm256_add_epi16(sumi1, _mm256_add_epi16(qx2, qx3));
+            sumi2 = _mm256_add_epi16(sumi2, qx4);
+        }
+
+        // last 16 bytes of 5-element, along with the 4 bytes of 4 elements
+        {
+            __m128i qx0 = _mm_loadu_si128((const __m128i *) (x[i].qs + 32));
+            uint32_t qh;
+            memcpy(&qh, x[i].qh, sizeof(qh)); // potentially unaligned
+            __m256i qx5_l = _mm256_cvtepu8_epi16(_mm_set1_epi32(qh));
+            __m128i qx1 = _mm_add_epi8(qx0, _mm_add_epi8(qx0, qx0)); // 1 * 3
+            __m128i qx2 = _mm_add_epi8(_mm_and_si128(_mm_slli_epi16(qx0, 3), _mm_set1_epi8(-8)), qx0); // 1 * 9
+            __m128i qx3 = _mm_add_epi8(_mm_and_si128(_mm_slli_epi16(qx1, 3), _mm_set1_epi8(-8)), qx1); // 3 * 9
+            __m128i qx4 = _mm_add_epi8(_mm_and_si128(_mm_slli_epi16(qx2, 3), _mm_set1_epi8(-8)), qx2); // 9 * 9
+            __m256i qx01 = MM256_SET_M128I(qx1, qx0);
+            __m256i qx23 = MM256_SET_M128I(qx3, qx2);
+
+            // avx2 does not have 8-bit multiplies, so 16-bit it is.
+            qx5_l = _mm256_mullo_epi16(qx5_l, _mm256_set_epi16(27, 27, 27, 27, 9, 9, 9, 9, 3, 3, 3, 3, 1, 1, 1, 1));
+            qx5_l = _mm256_and_si256(qx5_l, _mm256_set1_epi16(0xFF));
+            __m128i qx5 = _mm_packus_epi16(_mm256_castsi256_si128(qx5_l), _mm256_extracti128_si256(qx5_l, 1));
+
+            __m256i qx45 = MM256_SET_M128I(qx5, qx4);
+
+            // Cancel the +1 from avg so that it behaves like a halving add
+            qx01 = _mm256_subs_epu8(qx01, _mm256_set1_epi8(1));
+            qx23 = _mm256_subs_epu8(qx23, _mm256_set1_epi8(1));
+            qx45 = _mm256_subs_epu8(qx45, _mm256_set1_epi8(1));
+            // Multiply by 3 and get the top 2 bits
+            qx01 = _mm256_avg_epu8(qx01, _mm256_avg_epu8(qx01, _mm256_setzero_si256()));
+            qx23 = _mm256_avg_epu8(qx23, _mm256_avg_epu8(qx23, _mm256_setzero_si256()));
+            qx45 = _mm256_avg_epu8(qx45, _mm256_avg_epu8(qx45, _mm256_setzero_si256()));
+            qx01 = _mm256_and_si256(_mm256_srli_epi16(qx01, 6), _mm256_set1_epi8(3));
+            qx23 = _mm256_and_si256(_mm256_srli_epi16(qx23, 6), _mm256_set1_epi8(3));
+            qx45 = _mm256_and_si256(_mm256_srli_epi16(qx45, 6), _mm256_set1_epi8(3));
+
+            const __m256i qy01 = _mm256_loadu_si256((const __m256i *) (y[i].qs + 160));
+            const __m256i qy23 = _mm256_loadu_si256((const __m256i *) (y[i].qs + 192));
+            const __m256i qy45 = _mm256_loadu_si256((const __m256i *) (y[i].qs + 224));
+
+            qx01 = _mm256_maddubs_epi16(qx01, qy01);
+            qx23 = _mm256_maddubs_epi16(qx23, qy23);
+            qx45 = _mm256_maddubs_epi16(qx45, qy45);
+
+            sumi0 = _mm256_add_epi16(sumi0, qx01);
+            sumi1 = _mm256_add_epi16(sumi1, qx23);
+            sumi2 = _mm256_add_epi16(sumi2, qx45);
+        }
+
+        const __m256i ysum = _mm256_loadu_si256((const __m256i *) y[i].bsums);
+        const __m256 d = _mm256_set1_ps(y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d));
+
+        sumi0 = _mm256_sub_epi16(sumi0, ysum);
+        sumi0 = _mm256_add_epi16(sumi0, _mm256_add_epi16(sumi1, sumi2));
+        sumi0 = _mm256_madd_epi16(sumi0, _mm256_set1_epi16(1));
+
+        sumf = _mm256_add_ps(_mm256_mul_ps(_mm256_cvtepi32_ps(sumi0), d), sumf);
+    }
+
+    *s = hsum_float_8(sumf);
+
+#else
+    const uint8_t pow3[6] = {1, 3, 9, 27, 81, 243};
+
+    float sumf = 0.0f;
+
+    for (int i = 0; i < nb; ++i) {
+        int sum = 0;
+
+        for (size_t j = 0; j < sizeof(x->qs) - sizeof(x->qs) % 32; j += 32) {
+            for (size_t l = 0; l < 5; ++l) {
+                for (size_t m = 0; m < 32; ++m) {
+                    uint8_t q = x[i].qs[j + m] * pow3[l];
+                    uint16_t xi = ((uint16_t) q * 3) >> 8;
+                    sum += (xi - 1) * y[i].qs[j*5 + l*32 + m];
+                }
+            }
+        }
+        for (size_t j = sizeof(x->qs) - sizeof(x->qs) % 32; j < sizeof(x->qs); j += 16) {
+            for (size_t l = 0; l < 5; ++l) {
+                for (size_t m = 0; m < 16; ++m) {
+                    uint8_t q = x[i].qs[j + m] * pow3[l];
+                    uint16_t xi = ((uint16_t) q * 3) >> 8;
+                    sum += (xi - 1) * y[i].qs[j*5 + l*16 + m];
+                }
+            }
+        }
+
+        for (size_t l = 0; l < 4; ++l) {
+            for (size_t j = 0; j < sizeof(x->qh); ++j) {
+                uint8_t q = x[i].qh[j] * pow3[l];
+                uint16_t xi = ((uint16_t) q * 3) >> 8;
+                sum += (xi - 1) * y[i].qs[sizeof(x->qs)*5 + l*sizeof(x->qh) + j];
+            }
+        }
+
+        sumf += (float) sum * (GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d);
+    }
+
+    *s = sumf;
+#endif
+}
+
+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 * GGML_RESTRICT x = vx;
+    const block_q8_K  * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+#if defined(__AVX2__)
+    __m256 sumf = _mm256_setzero_ps();
+
+    for (int i = 0; i < nb; ++i) {
+        // 16-bit sums, because 256*127 still fits
+        __m256i sumi0 = _mm256_setzero_si256();
+        __m256i sumi1 = _mm256_setzero_si256();
+
+        for (size_t j = 0; j < sizeof(x->qs); j += 32) {
+            __m256i qx0 = _mm256_loadu_si256((const __m256i *) (x[i].qs + j));
+            __m256i qx1 = _mm256_srli_epi16(qx0, 2);
+            __m256i qx2 = _mm256_srli_epi16(qx0, 4);
+            __m256i qx3 = _mm256_srli_epi16(qx0, 6);
+
+            // 0, 1, 2 (should not be 3)
+            qx0 = _mm256_and_si256(qx0, _mm256_set1_epi8(3));
+            qx1 = _mm256_and_si256(qx1, _mm256_set1_epi8(3));
+            qx2 = _mm256_and_si256(qx2, _mm256_set1_epi8(3));
+            qx3 = _mm256_and_si256(qx3, _mm256_set1_epi8(3));
+
+            const __m256i qy0 = _mm256_loadu_si256((const __m256i *) (y[i].qs + j*4 +  0));
+            const __m256i qy1 = _mm256_loadu_si256((const __m256i *) (y[i].qs + j*4 + 32));
+            const __m256i qy2 = _mm256_loadu_si256((const __m256i *) (y[i].qs + j*4 + 64));
+            const __m256i qy3 = _mm256_loadu_si256((const __m256i *) (y[i].qs + j*4 + 96));
+
+            qx0 = _mm256_maddubs_epi16(qx0, qy0);
+            qx1 = _mm256_maddubs_epi16(qx1, qy1);
+            qx2 = _mm256_maddubs_epi16(qx2, qy2);
+            qx3 = _mm256_maddubs_epi16(qx3, qy3);
+
+            sumi0 = _mm256_add_epi16(sumi0, _mm256_add_epi16(qx0, qx1));
+            sumi1 = _mm256_add_epi16(sumi1, _mm256_add_epi16(qx2, qx3));
+        }
+
+        const __m256i ysum = _mm256_loadu_si256((const __m256i *) y[i].bsums);
+        const __m256 d = _mm256_set1_ps(y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d));
+
+        sumi0 = _mm256_add_epi16(sumi0, sumi1);
+        sumi0 = _mm256_sub_epi16(sumi0, ysum);
+        sumi0 = _mm256_madd_epi16(sumi0, _mm256_set1_epi16(1));
+
+        sumf = _mm256_add_ps(_mm256_mul_ps(_mm256_cvtepi32_ps(sumi0), d), sumf);
+    }
+
+    *s = hsum_float_8(sumf);
+
+#else
+    float sumf = 0.0f;
+
+    for (int i = 0; i < nb; ++i) {
+        int32_t sumi = 0;
+
+        for (size_t j = 0; j < sizeof(x->qs); j += 32) {
+            for (size_t l = 0; l < 4; ++l) {
+                for (size_t k = 0; k < 32; ++k) {
+                    sumi += y[i].qs[j*4 + l*32 + k] * (((x[i].qs[j + k] >> (l*2)) & 3) - 1);
+                }
+            }
+        }
+
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+
+        sumf += (float) sumi * d;
+    }
+
+    *s = sumf;
+#endif
+}
+
+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 * GGML_RESTRICT x = vx;
+    const block_q8_K * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+#if defined __AVX2__
+
+    const __m256i m3 = _mm256_set1_epi8(3);
+    const __m128i m4 = _mm_set1_epi8(0xF);
+
+    __m256 acc = _mm256_setzero_ps();
+
+    for (int i = 0; i < nb; ++i) {
+
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+        const float dmin = -y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
+
+        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);
+        const __m128i mins8 = _mm_and_si128(_mm_srli_epi16(mins_and_scales, 4), m4);
+        const __m256i mins = _mm256_cvtepi8_epi16(mins8);
+        const __m256i prod = _mm256_madd_epi16(mins, _mm256_loadu_si256((const __m256i*)y[i].bsums));
+
+        acc = _mm256_fmadd_ps(_mm256_broadcast_ss(&dmin), _mm256_cvtepi32_ps(prod), acc);
+
+        const __m256i all_scales = _mm256_cvtepi8_epi16(scales8);
+        const __m128i l_scales = _mm256_extracti128_si256(all_scales, 0);
+        const __m128i h_scales = _mm256_extracti128_si256(all_scales, 1);
+        const __m256i scales[2] = {MM256_SET_M128I(l_scales, l_scales), MM256_SET_M128I(h_scales, h_scales)};
+
+        __m256i sumi = _mm256_setzero_si256();
+
+        for (int j = 0; j < QK_K/128; ++j) {
+
+            const __m256i q2bits = _mm256_loadu_si256((const __m256i*)q2); q2 += 32;
+
+            const __m256i q8_0 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
+            const __m256i q8_1 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
+            const __m256i q8_2 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
+            const __m256i q8_3 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
+
+            const __m256i q2_0 = _mm256_and_si256(q2bits, m3);
+            const __m256i q2_1 = _mm256_and_si256(_mm256_srli_epi16(q2bits, 2), m3);
+            const __m256i q2_2 = _mm256_and_si256(_mm256_srli_epi16(q2bits, 4), m3);
+            const __m256i q2_3 = _mm256_and_si256(_mm256_srli_epi16(q2bits, 6), m3);
+
+            __m256i p0 = _mm256_maddubs_epi16(q2_0, q8_0);
+            __m256i p1 = _mm256_maddubs_epi16(q2_1, q8_1);
+            __m256i p2 = _mm256_maddubs_epi16(q2_2, q8_2);
+            __m256i p3 = _mm256_maddubs_epi16(q2_3, q8_3);
+
+            p0 = _mm256_madd_epi16(_mm256_shuffle_epi8(scales[j], get_scale_shuffle_q3k(0)), p0);
+            p1 = _mm256_madd_epi16(_mm256_shuffle_epi8(scales[j], get_scale_shuffle_q3k(1)), p1);
+            p2 = _mm256_madd_epi16(_mm256_shuffle_epi8(scales[j], get_scale_shuffle_q3k(2)), p2);
+            p3 = _mm256_madd_epi16(_mm256_shuffle_epi8(scales[j], get_scale_shuffle_q3k(3)), p3);
+
+            p0 = _mm256_add_epi32(p0, p1);
+            p2 = _mm256_add_epi32(p2, p3);
+
+            sumi = _mm256_add_epi32(sumi, _mm256_add_epi32(p0, p2));
+        }
+
+        acc = _mm256_fmadd_ps(_mm256_broadcast_ss(&d), _mm256_cvtepi32_ps(sumi), acc);
+
+    }
+
+    *s = hsum_float_8(acc);
+
+#elif defined __AVX__
+
+    const __m128i m3 = _mm_set1_epi8(0x3);
+    const __m128i m4 = _mm_set1_epi8(0xF);
+    const __m128i m2 = _mm_set1_epi8(0x2);
+
+    __m256 acc = _mm256_setzero_ps();
+
+    for (int i = 0; i < nb; ++i) {
+
+        const float dall = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+        const float dmin = -y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
+
+        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);
+        const __m128i scales16 = _mm_and_si128(mins_and_scales, m4);
+        const __m128i mins16 = _mm_and_si128(_mm_srli_epi16(mins_and_scales, 4), m4);
+        const __m128i mins_0 = _mm_cvtepi8_epi16(mins16);
+        const __m128i mins_1 = _mm_cvtepi8_epi16(_mm_unpackhi_epi64(mins16, mins16));
+
+        // summs = y[i].bsums * (x[i].scales >> 4) in 16bits*8*2 to 32bits*4*2
+        const __m128i summs_0 = _mm_madd_epi16(mins_0, _mm_loadu_si128((const __m128i*)&y[i].bsums[0]));
+        const __m128i summs_1 = _mm_madd_epi16(mins_1, _mm_loadu_si128((const __m128i*)&y[i].bsums[8]));
+
+        // sumf += -dmin * summs in 32bits*8
+        acc = _mm256_add_ps(_mm256_mul_ps(_mm256_broadcast_ss(&dmin), _mm256_cvtepi32_ps(MM256_SET_M128I(summs_1, summs_0))), acc);
+
+        const __m128i scales_0 = _mm_cvtepi8_epi16(scales16);
+        const __m128i scales_1 = _mm_cvtepi8_epi16(_mm_unpackhi_epi64(scales16, scales16));
+        const __m128i scales[2] = { scales_0, scales_1 };
+
+        __m128i sumi_0 = _mm_setzero_si128();
+        __m128i sumi_1 = _mm_setzero_si128();
+
+        for (int j = 0; j < QK_K/128; ++j) {
+
+            // load Q8 quants int8*16*8 from block_q8_K.qs[QK_K]
+            const __m128i q8_0 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            const __m128i q8_1 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            const __m128i q8_2 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            const __m128i q8_3 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            const __m128i q8_4 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            const __m128i q8_5 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            const __m128i q8_6 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            const __m128i q8_7 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+
+            // load 2bits*16*8 from block_q2_K.qs[QK_K/4]
+            __m128i q2bits = _mm_loadu_si128((const __m128i*)q2); q2 += 16;
+            const __m128i q2_0 = _mm_and_si128(q2bits, m3);
+            const __m128i q2_2 = _mm_and_si128(_mm_srli_epi16(q2bits, 2), m3);
+            const __m128i q2_4 = _mm_and_si128(_mm_srli_epi16(q2bits, 4), m3);
+            const __m128i q2_6 = _mm_and_si128(_mm_srli_epi16(q2bits, 6), m3);
+            q2bits = _mm_loadu_si128((const __m128i*)q2); q2 += 16;
+            const __m128i q2_1 = _mm_and_si128(q2bits, m3);
+            const __m128i q2_3 = _mm_and_si128(_mm_srli_epi16(q2bits, 2), m3);
+            const __m128i q2_5 = _mm_and_si128(_mm_srli_epi16(q2bits, 4), m3);
+            const __m128i q2_7 = _mm_and_si128(_mm_srli_epi16(q2bits, 6), m3);
+
+            // isuml = q8[l] * ((q2[l] >> shift) & 3) in 8bits*16*8 to 16bits*8*8
+            __m128i p0 = _mm_maddubs_epi16(q2_0, q8_0);
+            __m128i p1 = _mm_maddubs_epi16(q2_1, q8_1);
+            __m128i p2 = _mm_maddubs_epi16(q2_2, q8_2);
+            __m128i p3 = _mm_maddubs_epi16(q2_3, q8_3);
+            __m128i p4 = _mm_maddubs_epi16(q2_4, q8_4);
+            __m128i p5 = _mm_maddubs_epi16(q2_5, q8_5);
+            __m128i p6 = _mm_maddubs_epi16(q2_6, q8_6);
+            __m128i p7 = _mm_maddubs_epi16(q2_7, q8_7);
+
+            // isum += (x[i].scales[is++] & 0xF) * isuml in 16bits*8*8 to 32bits*4*8
+            __m128i shuffle = _mm_set1_epi16(0x0100);
+            p0 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p0);
+            shuffle = _mm_add_epi16(shuffle, m2);
+            p1 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p1);
+            shuffle = _mm_add_epi16(shuffle, m2);
+            p2 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p2);
+            shuffle = _mm_add_epi16(shuffle, m2);
+            p3 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p3);
+            shuffle = _mm_add_epi16(shuffle, m2);
+            p4 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p4);
+            shuffle = _mm_add_epi16(shuffle, m2);
+            p5 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p5);
+            shuffle = _mm_add_epi16(shuffle, m2);
+            p6 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p6);
+            shuffle = _mm_add_epi16(shuffle, m2);
+            p7 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p7);
+
+            p0 = _mm_add_epi32(p0, p1);
+            p2 = _mm_add_epi32(p2, p3);
+            p4 = _mm_add_epi32(p4, p5);
+            p6 = _mm_add_epi32(p6, p7);
+
+            // isum in 32bits*4*2
+            sumi_0 = _mm_add_epi32(sumi_0, _mm_add_epi32(p0, p2));
+            sumi_1 = _mm_add_epi32(sumi_1, _mm_add_epi32(p4, p6));
+        }
+
+        // sumf += dall * isum - dmin * summs in 32bits
+        __m256i sumi = MM256_SET_M128I(sumi_1, sumi_0);
+        acc = _mm256_add_ps(_mm256_mul_ps(_mm256_broadcast_ss(&dall), _mm256_cvtepi32_ps(sumi)), acc);
+    }
+
+    *s = hsum_float_8(acc);
+
+#else
+
+    float sumf = 0;
+
+    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;
+
+        int summs = 0;
+        for (int j = 0; j < 16; ++j) {
+            summs += y[i].bsums[j] * (sc[j] >> 4);
+        }
+
+        const float dall = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+        const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
+
+        int isum = 0;
+        int is = 0;
+        int d;
+        for (int k = 0; k < QK_K/128; ++k) {
+            int shift = 0;
+            for (int j = 0; j < 4; ++j) {
+                d = sc[is++] & 0xF;
+                int isuml = 0;
+                for (int l =  0; l < 16; ++l) isuml += q8[l] * ((q2[l] >> shift) & 3);
+                isum += d * isuml;
+                d = sc[is++] & 0xF;
+                isuml = 0;
+                for (int l = 16; l < 32; ++l) isuml += q8[l] * ((q2[l] >> shift) & 3);
+                isum += d * isuml;
+                shift += 2;
+                q8 += 32;
+            }
+            q2 += 32;
+        }
+        sumf += dall * isum - dmin * summs;
+    }
+    *s = sumf;
+#endif
+}
+
+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);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const uint32_t kmask1 = 0x03030303;
+    const uint32_t kmask2 = 0x0f0f0f0f;
+
+    const block_q3_K * GGML_RESTRICT x = vx;
+    const block_q8_K * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+#if defined __AVX2__
+
+    const __m256i m3 = _mm256_set1_epi8(3);
+    const __m256i mone = _mm256_set1_epi8(1);
+    const __m128i m32 = _mm_set1_epi8(32);
+
+    __m256 acc = _mm256_setzero_ps();
+
+    uint32_t aux[3];
+
+    for (int i = 0; i < nb; ++i) {
+
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+
+        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 = _mm_set_epi32(
+                ((aux[1] >> 4) & kmask2) | (((aux[2] >> 6) & kmask1) << 4),
+                ((aux[0] >> 4) & kmask2) | (((aux[2] >> 4) & kmask1) << 4),
+                (aux[1] & kmask2) | (((aux[2] >> 2) & kmask1) << 4),
+                (aux[0] & kmask2) | (((aux[2] >> 0) & kmask1) << 4));
+        scales128 = _mm_sub_epi8(scales128, m32);
+        const __m256i all_scales = _mm256_cvtepi8_epi16(scales128);
+        const __m128i l_scales = _mm256_extracti128_si256(all_scales, 0);
+        const __m128i h_scales = _mm256_extracti128_si256(all_scales, 1);
+        const __m256i scales[2] = {MM256_SET_M128I(l_scales, l_scales), MM256_SET_M128I(h_scales, h_scales)};
+
+        // high bit
+        const __m256i hbits = _mm256_loadu_si256((const __m256i*)x[i].hmask);
+
+        // integer accumulator
+        __m256i sumi = _mm256_setzero_si256();
+
+        int bit = 0;
+        int is  = 0;
+
+        for (int j = 0; j < QK_K/128; ++j) {
+            // load low 2 bits
+            const __m256i q3bits = _mm256_loadu_si256((const __m256i*)q3); q3 += 32;
+
+            // prepare low and high bits
+            const __m256i q3l_0 = _mm256_and_si256(q3bits, m3);
+            const __m256i q3h_0 = _mm256_slli_epi16(_mm256_srli_epi16(_mm256_andnot_si256(hbits, _mm256_slli_epi16(mone, bit)), bit), 2);
+            ++bit;
+
+            const __m256i q3l_1 = _mm256_and_si256(_mm256_srli_epi16(q3bits, 2), m3);
+            const __m256i q3h_1 = _mm256_slli_epi16(_mm256_srli_epi16(_mm256_andnot_si256(hbits, _mm256_slli_epi16(mone, bit)), bit), 2);
+            ++bit;
+
+            const __m256i q3l_2 = _mm256_and_si256(_mm256_srli_epi16(q3bits, 4), m3);
+            const __m256i q3h_2 = _mm256_slli_epi16(_mm256_srli_epi16(_mm256_andnot_si256(hbits, _mm256_slli_epi16(mone, bit)), bit), 2);
+            ++bit;
+
+            const __m256i q3l_3 = _mm256_and_si256(_mm256_srli_epi16(q3bits, 6), m3);
+            const __m256i q3h_3 = _mm256_slli_epi16(_mm256_srli_epi16(_mm256_andnot_si256(hbits, _mm256_slli_epi16(mone, bit)), bit), 2);
+            ++bit;
+
+            // load Q8 quants
+            const __m256i q8_0 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
+            const __m256i q8_1 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
+            const __m256i q8_2 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
+            const __m256i q8_3 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
+
+            // Dot product: we multiply the 2 low bits and 1 high bit part separately, so we can use _mm256_maddubs_epi16,
+            // 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 = _mm256_maddubs_epi16(q3h_0, q8_0);
+            __m256i q8s_1 = _mm256_maddubs_epi16(q3h_1, q8_1);
+            __m256i q8s_2 = _mm256_maddubs_epi16(q3h_2, q8_2);
+            __m256i q8s_3 = _mm256_maddubs_epi16(q3h_3, q8_3);
+
+            __m256i p16_0 = _mm256_maddubs_epi16(q3l_0, q8_0);
+            __m256i p16_1 = _mm256_maddubs_epi16(q3l_1, q8_1);
+            __m256i p16_2 = _mm256_maddubs_epi16(q3l_2, q8_2);
+            __m256i p16_3 = _mm256_maddubs_epi16(q3l_3, q8_3);
+
+            p16_0 = _mm256_sub_epi16(p16_0, q8s_0);
+            p16_1 = _mm256_sub_epi16(p16_1, q8s_1);
+            p16_2 = _mm256_sub_epi16(p16_2, q8s_2);
+            p16_3 = _mm256_sub_epi16(p16_3, q8s_3);
+
+            // multiply with scales
+            p16_0 = _mm256_madd_epi16(_mm256_shuffle_epi8(scales[j], get_scale_shuffle_q3k(is + 0)), p16_0);
+            p16_1 = _mm256_madd_epi16(_mm256_shuffle_epi8(scales[j], get_scale_shuffle_q3k(is + 1)), p16_1);
+            p16_2 = _mm256_madd_epi16(_mm256_shuffle_epi8(scales[j], get_scale_shuffle_q3k(is + 2)), p16_2);
+            p16_3 = _mm256_madd_epi16(_mm256_shuffle_epi8(scales[j], get_scale_shuffle_q3k(is + 3)), p16_3);
+
+            // accumulate
+            p16_0 = _mm256_add_epi32(p16_0, p16_1);
+            p16_2 = _mm256_add_epi32(p16_2, p16_3);
+            sumi  = _mm256_add_epi32(sumi, _mm256_add_epi32(p16_0, p16_2));
+
+        }
+
+        // multiply with block scale and accumulate
+        acc = _mm256_fmadd_ps(_mm256_broadcast_ss(&d), _mm256_cvtepi32_ps(sumi), acc);
+
+    }
+
+    *s = hsum_float_8(acc);
+
+#elif defined __AVX__
+
+    const __m128i m3 = _mm_set1_epi8(3);
+    const __m128i mone = _mm_set1_epi8(1);
+    const __m128i m32 = _mm_set1_epi8(32);
+    const __m128i m2 = _mm_set1_epi8(2);
+
+    __m256 acc = _mm256_setzero_ps();
+
+    const uint32_t *aux;
+
+    for (int i = 0; i < nb; ++i) {
+
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+
+        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;
+        __m128i scales128 = _mm_set_epi32(
+                ((aux[1] >> 4) & kmask2) | (((aux[2] >> 6) & kmask1) << 4),
+                ((aux[0] >> 4) & kmask2) | (((aux[2] >> 4) & kmask1) << 4),
+                (aux[1] & kmask2) | (((aux[2] >> 2) & kmask1) << 4),
+                (aux[0] & kmask2) | (((aux[2] >> 0) & kmask1) << 4));
+        scales128 = _mm_sub_epi8(scales128, m32);
+        const __m128i scales_0 = _mm_cvtepi8_epi16(scales128);
+        const __m128i scales_1 = _mm_cvtepi8_epi16(_mm_unpackhi_epi64(scales128, scales128));
+        const __m128i scales[2] = { scales_0, scales_1 };
+
+        // high bit *128*2 from block_q3_K.hmask[QK_K/8]
+        const __m128i hbits_0 = _mm_loadu_si128((const __m128i*)&x[i].hmask[0]);
+        const __m128i hbits_1 = _mm_loadu_si128((const __m128i*)&x[i].hmask[16]);
+
+        // integer accumulator
+        __m128i sumi_0 = _mm_setzero_si128();
+        __m128i sumi_1 = _mm_setzero_si128();
+
+        for (int j = 0; j < QK_K/128; ++j) {
+            // load low 2 bits *64*2 from block_q3_K.qs[QK_K/4]
+            const __m128i q3bits_0 = _mm_loadu_si128((const __m128i*)q3); q3 += 16;
+            const __m128i q3bits_1 = _mm_loadu_si128((const __m128i*)q3); q3 += 16;
+
+            // prepare low and high bits
+            const int bit = j << 2;
+
+            const __m128i q3l_0 = _mm_and_si128(q3bits_0, m3);
+            const __m128i q3l_1 = _mm_and_si128(q3bits_1, m3);
+            const __m128i q3h_0 = _mm_slli_epi16(_mm_srli_epi16(_mm_andnot_si128(hbits_0, _mm_slli_epi16(mone, bit)), bit), 2);
+            const __m128i q3h_1 = _mm_slli_epi16(_mm_srli_epi16(_mm_andnot_si128(hbits_1, _mm_slli_epi16(mone, bit)), bit), 2);
+
+            const __m128i q3l_2 = _mm_and_si128(_mm_srli_epi16(q3bits_0, 2), m3);
+            const __m128i q3l_3 = _mm_and_si128(_mm_srli_epi16(q3bits_1, 2), m3);
+            const __m128i q3h_2 = _mm_slli_epi16(_mm_srli_epi16(_mm_andnot_si128(hbits_0, _mm_slli_epi16(mone, bit+1)), bit+1), 2);
+            const __m128i q3h_3 = _mm_slli_epi16(_mm_srli_epi16(_mm_andnot_si128(hbits_1, _mm_slli_epi16(mone, bit+1)), bit+1), 2);
+
+            const __m128i q3l_4 = _mm_and_si128(_mm_srli_epi16(q3bits_0, 4), m3);
+            const __m128i q3l_5 = _mm_and_si128(_mm_srli_epi16(q3bits_1, 4), m3);
+            const __m128i q3h_4 = _mm_slli_epi16(_mm_srli_epi16(_mm_andnot_si128(hbits_0, _mm_slli_epi16(mone, bit+2)), bit+2), 2);
+            const __m128i q3h_5 = _mm_slli_epi16(_mm_srli_epi16(_mm_andnot_si128(hbits_1, _mm_slli_epi16(mone, bit+2)), bit+2), 2);
+
+            const __m128i q3l_6 = _mm_and_si128(_mm_srli_epi16(q3bits_0, 6), m3);
+            const __m128i q3l_7 = _mm_and_si128(_mm_srli_epi16(q3bits_1, 6), m3);
+            const __m128i q3h_6 = _mm_slli_epi16(_mm_srli_epi16(_mm_andnot_si128(hbits_0, _mm_slli_epi16(mone, bit+3)), bit+3), 2);
+            const __m128i q3h_7 = _mm_slli_epi16(_mm_srli_epi16(_mm_andnot_si128(hbits_1, _mm_slli_epi16(mone, bit+3)), bit+3), 2);
+
+            // load Q8 quants from block_q8_K.qs[QK_K]
+            const __m128i q8_0 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            const __m128i q8_1 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            const __m128i q8_2 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            const __m128i q8_3 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            const __m128i q8_4 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            const __m128i q8_5 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            const __m128i q8_6 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            const __m128i q8_7 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+
+            // Dot product: we multiply the 2 low bits and 1 high bit part separately, so we can use _mm256_maddubs_epi16,
+            // 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)
+            __m128i q8s_0 = _mm_maddubs_epi16(q3h_0, q8_0);
+            __m128i q8s_1 = _mm_maddubs_epi16(q3h_1, q8_1);
+            __m128i q8s_2 = _mm_maddubs_epi16(q3h_2, q8_2);
+            __m128i q8s_3 = _mm_maddubs_epi16(q3h_3, q8_3);
+            __m128i q8s_4 = _mm_maddubs_epi16(q3h_4, q8_4);
+            __m128i q8s_5 = _mm_maddubs_epi16(q3h_5, q8_5);
+            __m128i q8s_6 = _mm_maddubs_epi16(q3h_6, q8_6);
+            __m128i q8s_7 = _mm_maddubs_epi16(q3h_7, q8_7);
+
+            __m128i p16_0 = _mm_maddubs_epi16(q3l_0, q8_0);
+            __m128i p16_1 = _mm_maddubs_epi16(q3l_1, q8_1);
+            __m128i p16_2 = _mm_maddubs_epi16(q3l_2, q8_2);
+            __m128i p16_3 = _mm_maddubs_epi16(q3l_3, q8_3);
+            __m128i p16_4 = _mm_maddubs_epi16(q3l_4, q8_4);
+            __m128i p16_5 = _mm_maddubs_epi16(q3l_5, q8_5);
+            __m128i p16_6 = _mm_maddubs_epi16(q3l_6, q8_6);
+            __m128i p16_7 = _mm_maddubs_epi16(q3l_7, q8_7);
+
+            p16_0 = _mm_sub_epi16(p16_0, q8s_0);
+            p16_1 = _mm_sub_epi16(p16_1, q8s_1);
+            p16_2 = _mm_sub_epi16(p16_2, q8s_2);
+            p16_3 = _mm_sub_epi16(p16_3, q8s_3);
+            p16_4 = _mm_sub_epi16(p16_4, q8s_4);
+            p16_5 = _mm_sub_epi16(p16_5, q8s_5);
+            p16_6 = _mm_sub_epi16(p16_6, q8s_6);
+            p16_7 = _mm_sub_epi16(p16_7, q8s_7);
+
+            // multiply with scales
+            __m128i shuffle = _mm_set1_epi16(0x0100);
+            p16_0 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p16_0);
+            shuffle = _mm_add_epi16(shuffle, m2);
+            p16_1 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p16_1);
+            shuffle = _mm_add_epi16(shuffle, m2);
+            p16_2 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p16_2);
+            shuffle = _mm_add_epi16(shuffle, m2);
+            p16_3 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p16_3);
+            shuffle = _mm_add_epi16(shuffle, m2);
+            p16_4 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p16_4);
+            shuffle = _mm_add_epi16(shuffle, m2);
+            p16_5 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p16_5);
+            shuffle = _mm_add_epi16(shuffle, m2);
+            p16_6 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p16_6);
+            shuffle = _mm_add_epi16(shuffle, m2);
+            p16_7 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p16_7);
+
+            // accumulate
+            p16_0 = _mm_add_epi32(p16_0, p16_1);
+            p16_2 = _mm_add_epi32(p16_2, p16_3);
+            p16_4 = _mm_add_epi32(p16_4, p16_5);
+            p16_6 = _mm_add_epi32(p16_6, p16_7);
+            sumi_0 = _mm_add_epi32(sumi_0, _mm_add_epi32(p16_0, p16_2));
+            sumi_1 = _mm_add_epi32(sumi_1, _mm_add_epi32(p16_4, p16_6));
+
+        }
+
+        // multiply with block scale and accumulate
+        __m256i sumi = MM256_SET_M128I(sumi_1, sumi_0);
+        acc = _mm256_add_ps(_mm256_mul_ps(_mm256_broadcast_ss(&d), _mm256_cvtepi32_ps(sumi)), acc);
+
+    }
+
+    *s = hsum_float_8(acc);
+
+#else
+    // scalar version
+    // This function is written like this so the compiler can manage to vectorize most of it
+    // Using -Ofast, GCC and clang manage to produce code that is within a factor of 2 or so from the
+    // manually vectorized version above. Every other version I tried would run at least 4 times slower.
+    // The ideal situation would be if we could just write the code once, and the compiler would
+    // automatically produce the best possible set of machine instructions, instead of us having to manually
+    // write vectorized versions for AVX, ARM_NEON, etc.
+
+    int8_t  aux8[QK_K];
+    int16_t aux16[8];
+    float   sums [8];
+    int32_t aux32[8];
+    memset(sums, 0, 8*sizeof(float));
+
+    uint32_t auxs[4];
+    const int8_t * scales = (const int8_t*)auxs;
+
+    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;
+        memset(aux32, 0, 8*sizeof(int32_t));
+        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;
+            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
+            a += 32; m <<= 1;
+            for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 2) & 3;
+            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
+            a += 32; m <<= 1;
+            for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 4) & 3;
+            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
+            a += 32; m <<= 1;
+            for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 6) & 3;
+            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
+            a += 32; m <<= 1;
+            q3 += 32;
+        }
+        a = aux8;
+
+        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);
+        for (int j = 0; j < QK_K/16; ++j) {
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
+            q8 += 8; a += 8;
+        }
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
+        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
+    }
+    for (int l = 0; l < 8; ++l) sumf += sums[l];
+    *s = sumf;
+
+#endif
+
+}
+
+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);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q4_K * GGML_RESTRICT x = vx;
+    const block_q8_K * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+    static const uint32_t kmask1 = 0x3f3f3f3f;
+    static const uint32_t kmask2 = 0x0f0f0f0f;
+    static const uint32_t kmask3 = 0x03030303;
+
+    uint32_t utmp[4];
+
+#if defined __AVX2__
+
+    const __m256i m4 = _mm256_set1_epi8(0xF);
+
+    __m256 acc = _mm256_setzero_ps();
+    __m128 acc_m = _mm_setzero_ps();
+
+   for (int i = 0; i < nb; ++i) {
+
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+        const float dmin = -y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
+
+        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 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]));
+
+        const __m256i q8sums = _mm256_loadu_si256((const __m256i*)y[i].bsums);
+        const __m128i q8s = _mm_hadd_epi16(_mm256_extracti128_si256(q8sums, 0), _mm256_extracti128_si256(q8sums, 1));
+        const __m128i prod = _mm_madd_epi16(_mm256_extracti128_si256(mins_and_scales, 1), q8s);
+        acc_m = _mm_fmadd_ps(_mm_set1_ps(dmin), _mm_cvtepi32_ps(prod), acc_m);
+
+        const __m128i sc128  = _mm256_extracti128_si256(mins_and_scales, 0);
+        const __m256i scales = MM256_SET_M128I(sc128, sc128);
+
+        __m256i sumi = _mm256_setzero_si256();
+
+        for (int j = 0; j < QK_K/64; ++j) {
+
+            const __m256i scale_l = _mm256_shuffle_epi8(scales, get_scale_shuffle_k4(2*j+0));
+            const __m256i scale_h = _mm256_shuffle_epi8(scales, get_scale_shuffle_k4(2*j+1));
+
+            const __m256i q4bits = _mm256_loadu_si256((const __m256i*)q4); q4 += 32;
+            const __m256i q4l = _mm256_and_si256(q4bits, m4);
+            const __m256i q4h = _mm256_and_si256(_mm256_srli_epi16(q4bits, 4), m4);
+
+            const __m256i q8l = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
+            __m256i p16l = _mm256_maddubs_epi16(q4l, q8l);
+            p16l = _mm256_madd_epi16(scale_l, p16l);
+
+            const __m256i q8h = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
+            __m256i p16h = _mm256_maddubs_epi16(q4h, q8h);
+            p16h = _mm256_madd_epi16(scale_h, p16h);
+            const __m256i sumj = _mm256_add_epi32(p16l, p16h);
+
+            sumi = _mm256_add_epi32(sumi, sumj);
+        }
+
+        __m256 vd = _mm256_set1_ps(d);
+        acc = _mm256_fmadd_ps(vd, _mm256_cvtepi32_ps(sumi), acc);
+
+    }
+
+    acc_m = _mm_add_ps(acc_m, _mm_movehl_ps(acc_m, acc_m));
+    acc_m = _mm_add_ss(acc_m, _mm_movehdup_ps(acc_m));
+
+    *s = hsum_float_8(acc) + _mm_cvtss_f32(acc_m);
+
+#elif defined __AVX__
+
+    const __m128i m4 = _mm_set1_epi8(0xF);
+    const __m128i m2 = _mm_set1_epi8(0x2);
+
+    __m256 acc = _mm256_setzero_ps();
+    __m128 acc_m = _mm_setzero_ps();
+
+   for (int i = 0; i < nb; ++i) {
+
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+        const float dmin = -y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
+
+        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);
+        const uint32_t uaux = utmp[1] & kmask1;
+        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
+        utmp[2] = uaux;
+        utmp[0] &= kmask1;
+
+        const __m128i utmps = _mm_set_epi32(utmp[3], utmp[2], utmp[1], utmp[0]);
+        const __m128i scales = _mm_cvtepu8_epi16(utmps);
+        const __m128i mins = _mm_cvtepu8_epi16(_mm_unpackhi_epi64(utmps, utmps));
+
+        const __m128i q8sums_0 = _mm_loadu_si128((const __m128i*)&y[i].bsums[0]);
+        const __m128i q8sums_1 = _mm_loadu_si128((const __m128i*)&y[i].bsums[8]);
+        const __m128i q8s = _mm_hadd_epi16(q8sums_0, q8sums_1);
+        const __m128i prod = _mm_madd_epi16(mins, q8s);
+        acc_m = _mm_add_ps(_mm_mul_ps(_mm_set1_ps(dmin), _mm_cvtepi32_ps(prod)), acc_m);
+
+        __m128i sumi_0 = _mm_setzero_si128();
+        __m128i sumi_1 = _mm_setzero_si128();
+
+        __m128i shuffle = _mm_set1_epi16(0x0100);
+        for (int j = 0; j < QK_K/64; ++j) {
+
+            const __m128i scale_l = _mm_shuffle_epi8(scales, shuffle);
+            shuffle = _mm_add_epi16(shuffle, m2);
+            const __m128i scale_h = _mm_shuffle_epi8(scales, shuffle);
+            shuffle = _mm_add_epi16(shuffle, m2);
+
+            __m128i q4bits = _mm_loadu_si128((const __m128i*)q4); q4 += 16;
+            const __m128i q4l_0 = _mm_and_si128(q4bits, m4);
+            const __m128i q4h_0 = _mm_and_si128(_mm_srli_epi16(q4bits, 4), m4);
+            q4bits = _mm_loadu_si128((const __m128i*)q4); q4 += 16;
+            const __m128i q4l_1 = _mm_and_si128(q4bits, m4);
+            const __m128i q4h_1 = _mm_and_si128(_mm_srli_epi16(q4bits, 4), m4);
+
+            const __m128i q8l_0 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            __m128i p16l = _mm_maddubs_epi16(q4l_0, q8l_0);
+            p16l = _mm_madd_epi16(scale_l, p16l);
+            sumi_0 = _mm_add_epi32(sumi_0, p16l);
+            const __m128i q8l_1 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            p16l = _mm_maddubs_epi16(q4l_1, q8l_1);
+            p16l = _mm_madd_epi16(scale_l, p16l);
+            sumi_1 = _mm_add_epi32(sumi_1, p16l);
+
+            const __m128i q8h_0 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            __m128i p16h = _mm_maddubs_epi16(q4h_0, q8h_0);
+            p16h = _mm_madd_epi16(scale_h, p16h);
+            sumi_0 = _mm_add_epi32(sumi_0, p16h);
+            const __m128i q8h_1 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            p16h = _mm_maddubs_epi16(q4h_1, q8h_1);
+            p16h = _mm_madd_epi16(scale_h, p16h);
+            sumi_1 = _mm_add_epi32(sumi_1, p16h);
+
+        }
+
+        __m256 vd = _mm256_set1_ps(d);
+        __m256i sumi = MM256_SET_M128I(sumi_1, sumi_0);
+        acc = _mm256_add_ps(_mm256_mul_ps(vd, _mm256_cvtepi32_ps(sumi)), acc);
+
+    }
+
+    acc_m = _mm_add_ps(acc_m, _mm_movehl_ps(acc_m, acc_m));
+    acc_m = _mm_add_ss(acc_m, _mm_movehdup_ps(acc_m));
+
+    *s = hsum_float_8(acc) + _mm_cvtss_f32(acc_m);
+
+#else
+
+    const uint8_t * scales = (const uint8_t*)&utmp[0];
+    const uint8_t * mins   = (const uint8_t*)&utmp[2];
+
+    int8_t  aux8[QK_K];
+    int16_t aux16[8];
+    float   sums [8];
+    int32_t aux32[8];
+    memset(sums, 0, 8*sizeof(float));
+
+    float sumf = 0;
+    for (int i = 0; i < nb; ++i) {
+        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 * 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;
+            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l]  >> 4);
+            a += 32; q4 += 32;
+        }
+        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;
+
+        int sumi = 0;
+        for (int j = 0; j < QK_K/16; ++j) sumi += y[i].bsums[j] * mins[j/2];
+        a = aux8;
+        int is = 0;
+        for (int j = 0; j < QK_K/32; ++j) {
+            int32_t scale = scales[is++];
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+        }
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
+        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
+        const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
+        sumf -= dmin * sumi;
+    }
+    for (int l = 0; l < 8; ++l) sumf += sums[l];
+    *s = sumf;
+#endif
+}
+
+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);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q5_K * GGML_RESTRICT x = vx;
+    const block_q8_K * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+    static const uint32_t kmask1 = 0x3f3f3f3f;
+    static const uint32_t kmask2 = 0x0f0f0f0f;
+    static const uint32_t kmask3 = 0x03030303;
+
+    uint32_t utmp[4];
+
+#if defined __AVX2__
+
+    const __m256i m4 = _mm256_set1_epi8(0xF);
+    const __m128i mzero = _mm_setzero_si128();
+    const __m256i mone  = _mm256_set1_epi8(1);
+
+    __m256 acc = _mm256_setzero_ps();
+
+    float summs = 0.f;
+
+    for (int i = 0; i < nb; ++i) {
+        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_CPU_FP16_TO_FP32(x[i].d);
+        const float dmin = -y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
+
+        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 __m256i mins_and_scales = _mm256_cvtepu8_epi16(_mm_set_epi32(utmp[3], utmp[2], utmp[1], utmp[0]));
+
+        const __m256i q8sums = _mm256_loadu_si256((const __m256i*)y[i].bsums);
+        const __m128i q8s = _mm_hadd_epi16(_mm256_extracti128_si256(q8sums, 0), _mm256_extracti128_si256(q8sums, 1));
+        const __m128i prod = _mm_madd_epi16(_mm256_extracti128_si256(mins_and_scales, 1), q8s);
+        const __m128i hsum = _mm_hadd_epi32(_mm_hadd_epi32(prod, mzero), mzero);
+        summs += dmin * _mm_extract_epi32(hsum, 0);
+
+        const __m128i sc128  = _mm256_extracti128_si256(mins_and_scales, 0);
+        const __m256i scales = MM256_SET_M128I(sc128, sc128);
+
+        const __m256i hbits = _mm256_loadu_si256((const __m256i*)x[i].qh);
+        __m256i hmask = mone;
+
+        __m256i sumi = _mm256_setzero_si256();
+
+        int bit = 0;
+
+        for (int j = 0; j < QK_K/64; ++j) {
+
+            const __m256i scale_0 = _mm256_shuffle_epi8(scales, get_scale_shuffle_k4(2*j+0));
+            const __m256i scale_1 = _mm256_shuffle_epi8(scales, get_scale_shuffle_k4(2*j+1));
+
+            const __m256i q5bits = _mm256_loadu_si256((const __m256i*)q5); q5 += 32;
+
+            const __m256i q5l_0 = _mm256_and_si256(q5bits, m4);
+            const __m256i q5h_0 = _mm256_slli_epi16(_mm256_srli_epi16(_mm256_and_si256(hbits, hmask), bit++), 4);
+            const __m256i q5_0  = _mm256_add_epi8(q5l_0, q5h_0);
+            hmask = _mm256_slli_epi16(hmask, 1);
+
+            const __m256i q5l_1 = _mm256_and_si256(_mm256_srli_epi16(q5bits, 4), m4);
+            const __m256i q5h_1 = _mm256_slli_epi16(_mm256_srli_epi16(_mm256_and_si256(hbits, hmask), bit++), 4);
+            const __m256i q5_1  = _mm256_add_epi8(q5l_1, q5h_1);
+            hmask = _mm256_slli_epi16(hmask, 1);
+
+            const __m256i q8_0 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
+            const __m256i q8_1 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
+
+            __m256i p16_0 = _mm256_maddubs_epi16(q5_0, q8_0);
+            __m256i p16_1 = _mm256_maddubs_epi16(q5_1, q8_1);
+
+            p16_0 = _mm256_madd_epi16(scale_0, p16_0);
+            p16_1 = _mm256_madd_epi16(scale_1, p16_1);
+
+            sumi = _mm256_add_epi32(sumi, _mm256_add_epi32(p16_0, p16_1));
+
+        }
+
+        __m256 vd = _mm256_set1_ps(d);
+        acc = _mm256_fmadd_ps(vd, _mm256_cvtepi32_ps(sumi), acc);
+
+    }
+
+    *s = hsum_float_8(acc) + summs;
+
+#elif defined __AVX__
+
+    const __m128i m4 = _mm_set1_epi8(0xF);
+    const __m128i mzero = _mm_setzero_si128();
+    const __m128i mone  = _mm_set1_epi8(1);
+    const __m128i m2 = _mm_set1_epi8(2);
+
+    __m256 acc = _mm256_setzero_ps();
+
+    float summs = 0.f;
+
+    for (int i = 0; i < nb; ++i) {
+
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+        const float dmin = -y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
+
+        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);
+        const uint32_t uaux = utmp[1] & kmask1;
+        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
+        utmp[2] = uaux;
+        utmp[0] &= kmask1;
+
+        const __m128i utmps = _mm_set_epi32(utmp[3], utmp[2], utmp[1], utmp[0]);
+        const __m128i scales = _mm_cvtepu8_epi16(utmps);
+        const __m128i mins = _mm_cvtepu8_epi16(_mm_unpackhi_epi64(utmps, utmps));
+
+        const __m128i q8sums_0 = _mm_loadu_si128((const __m128i*)&y[i].bsums[0]);
+        const __m128i q8sums_1 = _mm_loadu_si128((const __m128i*)&y[i].bsums[8]);
+        const __m128i q8s = _mm_hadd_epi16(q8sums_0, q8sums_1);
+        const __m128i prod = _mm_madd_epi16(mins, q8s);
+        const __m128i hsum = _mm_hadd_epi32(_mm_hadd_epi32(prod, mzero), mzero);
+        summs += dmin * _mm_extract_epi32(hsum, 0);
+
+        const __m128i hbits_0 = _mm_loadu_si128((const __m128i*)&x[i].qh[0]);
+        const __m128i hbits_1 = _mm_loadu_si128((const __m128i*)&x[i].qh[16]);
+        __m128i hmask = mone;
+
+        __m128i sumi_0 = _mm_setzero_si128();
+        __m128i sumi_1 = _mm_setzero_si128();
+
+        int bit = 0;
+
+        __m128i shuffle = _mm_set1_epi16(0x0100);
+        for (int j = 0; j < QK_K/64; ++j) {
+
+            const __m128i scale_0 = _mm_shuffle_epi8(scales, shuffle);
+            shuffle = _mm_add_epi16(shuffle, m2);
+            const __m128i scale_1 = _mm_shuffle_epi8(scales, shuffle);
+            shuffle = _mm_add_epi16(shuffle, m2);
+
+            const __m128i q5bits_0 = _mm_loadu_si128((const __m128i*)q5); q5 += 16;
+            const __m128i q5bits_1 = _mm_loadu_si128((const __m128i*)q5); q5 += 16;
+
+            __m128i q5l_0 = _mm_and_si128(q5bits_0, m4);
+            __m128i q5l_1 = _mm_and_si128(q5bits_1, m4);
+            __m128i q5h_0 = _mm_slli_epi16(_mm_srli_epi16(_mm_and_si128(hbits_0, hmask), bit), 4);
+            __m128i q5h_1 = _mm_slli_epi16(_mm_srli_epi16(_mm_and_si128(hbits_1, hmask), bit++), 4);
+            __m128i q5_0  = _mm_add_epi8(q5l_0, q5h_0);
+            __m128i q5_1  = _mm_add_epi8(q5l_1, q5h_1);
+            hmask = _mm_slli_epi16(hmask, 1);
+
+            __m128i q8_0 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            __m128i q8_1 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            __m128i p16_0 = _mm_maddubs_epi16(q5_0, q8_0);
+            __m128i p16_1 = _mm_maddubs_epi16(q5_1, q8_1);
+            p16_0 = _mm_madd_epi16(scale_0, p16_0);
+            p16_1 = _mm_madd_epi16(scale_0, p16_1);
+
+            q5l_0 = _mm_and_si128(_mm_srli_epi16(q5bits_0, 4), m4);
+            q5l_1 = _mm_and_si128(_mm_srli_epi16(q5bits_1, 4), m4);
+            q5h_0 = _mm_slli_epi16(_mm_srli_epi16(_mm_and_si128(hbits_0, hmask), bit), 4);
+            q5h_1 = _mm_slli_epi16(_mm_srli_epi16(_mm_and_si128(hbits_1, hmask), bit++), 4);
+            q5_0  = _mm_add_epi8(q5l_0, q5h_0);
+            q5_1  = _mm_add_epi8(q5l_1, q5h_1);
+            hmask = _mm_slli_epi16(hmask, 1);
+
+            q8_0 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            q8_1 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            __m128i p16_2 = _mm_maddubs_epi16(q5_0, q8_0);
+            __m128i p16_3 = _mm_maddubs_epi16(q5_1, q8_1);
+            p16_2 = _mm_madd_epi16(scale_1, p16_2);
+            p16_3 = _mm_madd_epi16(scale_1, p16_3);
+
+            sumi_0 = _mm_add_epi32(sumi_0, _mm_add_epi32(p16_0, p16_2));
+            sumi_1 = _mm_add_epi32(sumi_1, _mm_add_epi32(p16_1, p16_3));
+
+        }
+
+        __m256 vd = _mm256_set1_ps(d);
+        __m256i sumi = MM256_SET_M128I(sumi_1, sumi_0);
+        acc = _mm256_add_ps(_mm256_mul_ps(vd, _mm256_cvtepi32_ps(sumi)), acc);
+
+    }
+
+    *s = hsum_float_8(acc) + summs;
+
+#else
+
+    const uint8_t * scales = (const uint8_t*)&utmp[0];
+    const uint8_t * mins   = (const uint8_t*)&utmp[2];
+
+    int8_t  aux8[QK_K];
+    int16_t aux16[8];
+    float   sums [8];
+    int32_t aux32[8];
+    memset(sums, 0, 8*sizeof(float));
+
+    float sumf = 0;
+    for (int i = 0; i < nb; ++i) {
+        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 * 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);
+            for (int l = 0; l < 32; ++l) a[l] += (hm[l] & m ? 16 : 0);
+            a += 32; m <<= 1;
+            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l]  >> 4);
+            for (int l = 0; l < 32; ++l) a[l] += (hm[l] & m ? 16 : 0);
+            a += 32; m <<= 1;
+            q4 += 32;
+        }
+        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;
+
+        int sumi = 0;
+        for (int j = 0; j < QK_K/16; ++j) sumi += y[i].bsums[j] * mins[j/2];
+        a = aux8;
+        int is = 0;
+        for (int j = 0; j < QK_K/32; ++j) {
+            int32_t scale = scales[is++];
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+        }
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
+        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
+        const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
+        sumf -= dmin * sumi;
+    }
+    for (int l = 0; l < 8; ++l) sumf += sums[l];
+    *s = sumf;
+#endif
+}
+
+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);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q6_K * GGML_RESTRICT x = vx;
+    const block_q8_K * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+#if defined __AVX2__
+
+    const __m256i m4 = _mm256_set1_epi8(0xF);
+    const __m256i m2 = _mm256_set1_epi8(3);
+    const __m256i m32s = _mm256_set1_epi8(32);
+
+    __m256 acc = _mm256_setzero_ps();
+
+    for (int i = 0; i < nb; ++i) {
+
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+
+        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);
+
+        __m256i sumi = _mm256_setzero_si256();
+
+        int is = 0;
+
+        for (int j = 0; j < QK_K/128; ++j) {
+
+            const __m128i scale_0 = _mm_shuffle_epi8(scales, get_scale_shuffle(is + 0));
+            const __m128i scale_1 = _mm_shuffle_epi8(scales, get_scale_shuffle(is + 1));
+            const __m128i scale_2 = _mm_shuffle_epi8(scales, get_scale_shuffle(is + 2));
+            const __m128i scale_3 = _mm_shuffle_epi8(scales, get_scale_shuffle(is + 3));
+            is += 4;
+
+            const __m256i q4bits1 = _mm256_loadu_si256((const __m256i*)q4); q4 += 32;
+            const __m256i q4bits2 = _mm256_loadu_si256((const __m256i*)q4); q4 += 32;
+            const __m256i q4bitsH = _mm256_loadu_si256((const __m256i*)qh); qh += 32;
+
+            const __m256i q4h_0 = _mm256_slli_epi16(_mm256_and_si256(q4bitsH, m2), 4);
+            const __m256i q4h_1 = _mm256_slli_epi16(_mm256_and_si256(_mm256_srli_epi16(q4bitsH, 2), m2), 4);
+            const __m256i q4h_2 = _mm256_slli_epi16(_mm256_and_si256(_mm256_srli_epi16(q4bitsH, 4), m2), 4);
+            const __m256i q4h_3 = _mm256_slli_epi16(_mm256_and_si256(_mm256_srli_epi16(q4bitsH, 6), m2), 4);
+
+            const __m256i q4_0 = _mm256_or_si256(_mm256_and_si256(q4bits1, m4), q4h_0);
+            const __m256i q4_1 = _mm256_or_si256(_mm256_and_si256(q4bits2, m4), q4h_1);
+            const __m256i q4_2 = _mm256_or_si256(_mm256_and_si256(_mm256_srli_epi16(q4bits1, 4), m4), q4h_2);
+            const __m256i q4_3 = _mm256_or_si256(_mm256_and_si256(_mm256_srli_epi16(q4bits2, 4), m4), q4h_3);
+
+            const __m256i q8_0 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
+            const __m256i q8_1 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
+            const __m256i q8_2 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
+            const __m256i q8_3 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
+
+            __m256i q8s_0 = _mm256_maddubs_epi16(m32s, q8_0);
+            __m256i q8s_1 = _mm256_maddubs_epi16(m32s, q8_1);
+            __m256i q8s_2 = _mm256_maddubs_epi16(m32s, q8_2);
+            __m256i q8s_3 = _mm256_maddubs_epi16(m32s, q8_3);
+
+            __m256i p16_0 = _mm256_maddubs_epi16(q4_0, q8_0);
+            __m256i p16_1 = _mm256_maddubs_epi16(q4_1, q8_1);
+            __m256i p16_2 = _mm256_maddubs_epi16(q4_2, q8_2);
+            __m256i p16_3 = _mm256_maddubs_epi16(q4_3, q8_3);
+
+            p16_0 = _mm256_sub_epi16(p16_0, q8s_0);
+            p16_1 = _mm256_sub_epi16(p16_1, q8s_1);
+            p16_2 = _mm256_sub_epi16(p16_2, q8s_2);
+            p16_3 = _mm256_sub_epi16(p16_3, q8s_3);
+
+            p16_0 = _mm256_madd_epi16(_mm256_cvtepi8_epi16(scale_0), p16_0);
+            p16_1 = _mm256_madd_epi16(_mm256_cvtepi8_epi16(scale_1), p16_1);
+            p16_2 = _mm256_madd_epi16(_mm256_cvtepi8_epi16(scale_2), p16_2);
+            p16_3 = _mm256_madd_epi16(_mm256_cvtepi8_epi16(scale_3), p16_3);
+
+            sumi = _mm256_add_epi32(sumi, _mm256_add_epi32(p16_0, p16_1));
+            sumi = _mm256_add_epi32(sumi, _mm256_add_epi32(p16_2, p16_3));
+
+        }
+
+        acc = _mm256_fmadd_ps(_mm256_broadcast_ss(&d), _mm256_cvtepi32_ps(sumi), acc);
+    }
+
+    *s = hsum_float_8(acc);
+
+#elif defined __AVX__
+
+    const __m128i m3 = _mm_set1_epi8(3);
+    const __m128i m15 = _mm_set1_epi8(15);
+
+    __m256 acc = _mm256_setzero_ps();
+
+    for (int i = 0; i < nb; ++i) {
+
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+
+        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);
+        const __m128i q8sums_1 = _mm_loadu_si128((const __m128i*)y[i].bsums + 1);
+        const __m128i scales = _mm_loadu_si128((const __m128i*)x[i].scales);
+        const __m128i scales_16_0 = _mm_cvtepi8_epi16(scales);
+        const __m128i scales_16_1 = _mm_cvtepi8_epi16(_mm_bsrli_si128(scales, 8));
+        const __m128i q8sclsub_0 = _mm_slli_epi32(_mm_madd_epi16(q8sums_0, scales_16_0), 5);
+        const __m128i q8sclsub_1 = _mm_slli_epi32(_mm_madd_epi16(q8sums_1, scales_16_1), 5);
+
+        __m128i sumi_0 = _mm_setzero_si128();
+        __m128i sumi_1 = _mm_setzero_si128();
+
+        int is = 0;
+
+        for (int j = 0; j < QK_K/128; ++j) {
+
+            const __m128i q4bitsH_0 = _mm_loadu_si128((const __m128i*)qh); qh += 16;
+            const __m128i q4bitsH_1 = _mm_loadu_si128((const __m128i*)qh); qh += 16;
+
+            const __m128i q4h_0 = _mm_slli_epi16(_mm_and_si128(q4bitsH_0, m3), 4);
+            const __m128i q4h_1 = _mm_slli_epi16(_mm_and_si128(q4bitsH_1, m3), 4);
+            const __m128i q4h_2 = _mm_slli_epi16(_mm_and_si128(q4bitsH_0, _mm_set1_epi8(12)), 2);
+            const __m128i q4h_3 = _mm_slli_epi16(_mm_and_si128(q4bitsH_1, _mm_set1_epi8(12)), 2);
+            const __m128i q4h_4 = _mm_and_si128(q4bitsH_0, _mm_set1_epi8(48));
+            const __m128i q4h_5 = _mm_and_si128(q4bitsH_1, _mm_set1_epi8(48));
+            const __m128i q4h_6 = _mm_srli_epi16(_mm_and_si128(q4bitsH_0, _mm_set1_epi8(-64)), 2);
+            const __m128i q4h_7 = _mm_srli_epi16(_mm_and_si128(q4bitsH_1, _mm_set1_epi8(-64)), 2);
+
+            const __m128i q4bits1_0 = _mm_loadu_si128((const __m128i*)q4); q4 += 16;
+            const __m128i q4bits1_1 = _mm_loadu_si128((const __m128i*)q4); q4 += 16;
+            const __m128i q4bits2_0 = _mm_loadu_si128((const __m128i*)q4); q4 += 16;
+            const __m128i q4bits2_1 = _mm_loadu_si128((const __m128i*)q4); q4 += 16;
+
+            const __m128i q4_0 = _mm_or_si128(_mm_and_si128(q4bits1_0, m15), q4h_0);
+            const __m128i q4_1 = _mm_or_si128(_mm_and_si128(q4bits1_1, m15), q4h_1);
+            const __m128i q4_2 = _mm_or_si128(_mm_and_si128(q4bits2_0, m15), q4h_2);
+            const __m128i q4_3 = _mm_or_si128(_mm_and_si128(q4bits2_1, m15), q4h_3);
+            const __m128i q4_4 = _mm_or_si128(_mm_and_si128(_mm_srli_epi16(q4bits1_0, 4), m15), q4h_4);
+            const __m128i q4_5 = _mm_or_si128(_mm_and_si128(_mm_srli_epi16(q4bits1_1, 4), m15), q4h_5);
+            const __m128i q4_6 = _mm_or_si128(_mm_and_si128(_mm_srli_epi16(q4bits2_0, 4), m15), q4h_6);
+            const __m128i q4_7 = _mm_or_si128(_mm_and_si128(_mm_srli_epi16(q4bits2_1, 4), m15), q4h_7);
+
+            const __m128i q8_0 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            const __m128i q8_1 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            const __m128i q8_2 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            const __m128i q8_3 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            const __m128i q8_4 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            const __m128i q8_5 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            const __m128i q8_6 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            const __m128i q8_7 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+
+            __m128i p16_0 = _mm_maddubs_epi16(q4_0, q8_0);
+            __m128i p16_1 = _mm_maddubs_epi16(q4_1, q8_1);
+            __m128i p16_2 = _mm_maddubs_epi16(q4_2, q8_2);
+            __m128i p16_3 = _mm_maddubs_epi16(q4_3, q8_3);
+            __m128i p16_4 = _mm_maddubs_epi16(q4_4, q8_4);
+            __m128i p16_5 = _mm_maddubs_epi16(q4_5, q8_5);
+            __m128i p16_6 = _mm_maddubs_epi16(q4_6, q8_6);
+            __m128i p16_7 = _mm_maddubs_epi16(q4_7, q8_7);
+
+            const __m128i scale_0 = _mm_shuffle_epi8(scales, get_scale_shuffle(is + 0));
+            const __m128i scale_1 = _mm_shuffle_epi8(scales, get_scale_shuffle(is + 1));
+            const __m128i scale_2 = _mm_shuffle_epi8(scales, get_scale_shuffle(is + 2));
+            const __m128i scale_3 = _mm_shuffle_epi8(scales, get_scale_shuffle(is + 3));
+            is += 4;
+
+            p16_0 = _mm_madd_epi16(_mm_cvtepi8_epi16(scale_0), p16_0);
+            p16_1 = _mm_madd_epi16(_mm_cvtepi8_epi16(_mm_bsrli_si128(scale_0, 8)), p16_1);
+            p16_2 = _mm_madd_epi16(_mm_cvtepi8_epi16(scale_1), p16_2);
+            p16_3 = _mm_madd_epi16(_mm_cvtepi8_epi16(_mm_bsrli_si128(scale_1, 8)), p16_3);
+            p16_4 = _mm_madd_epi16(_mm_cvtepi8_epi16(scale_2), p16_4);
+            p16_5 = _mm_madd_epi16(_mm_cvtepi8_epi16(_mm_bsrli_si128(scale_2, 8)), p16_5);
+            p16_6 = _mm_madd_epi16(_mm_cvtepi8_epi16(scale_3), p16_6);
+            p16_7 = _mm_madd_epi16(_mm_cvtepi8_epi16(_mm_bsrli_si128(scale_3, 8)), p16_7);
+
+            sumi_0 = _mm_add_epi32(sumi_0, _mm_add_epi32(p16_0, p16_2));
+            sumi_1 = _mm_add_epi32(sumi_1, _mm_add_epi32(p16_1, p16_3));
+            sumi_0 = _mm_add_epi32(sumi_0, _mm_add_epi32(p16_4, p16_6));
+            sumi_1 = _mm_add_epi32(sumi_1, _mm_add_epi32(p16_5, p16_7));
+
+        }
+
+        sumi_0 = _mm_sub_epi32(sumi_0, q8sclsub_0);
+        sumi_1 = _mm_sub_epi32(sumi_1, q8sclsub_1);
+        const __m256i sumi = MM256_SET_M128I(sumi_1, sumi_0);
+        acc = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(sumi)), acc);
+    }
+
+    *s = hsum_float_8(acc);
+
+#else
+
+    int8_t  aux8[QK_K];
+    int16_t aux16[8];
+    float   sums [8];
+    int32_t aux32[8];
+    memset(sums, 0, 8*sizeof(float));
+
+    float sumf = 0;
+    for (int i = 0; i < nb; ++i) {
+        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 * 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;
+                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;
+        }
+        a = aux8;
+        int is = 0;
+        for (int j = 0; j < QK_K/16; ++j) {
+            int scale = x[i].scales[is++];
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+        }
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
+        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
+    }
+    for (int l = 0; l < 8; ++l) sumf += sums[l];
+    *s = sumf;
+#endif
+}
+
+#if defined (__AVX__) || defined (__AVX2__)
+static const int8_t keven_signs_q2xs[1024] = {
+     1,  1,  1,  1,  1,  1,  1,  1, -1,  1,  1,  1,  1,  1,  1, -1,  1, -1,  1,  1,  1,  1,  1, -1, -1, -1,  1,  1,  1,  1,  1,  1,
+     1,  1, -1,  1,  1,  1,  1, -1, -1,  1, -1,  1,  1,  1,  1,  1,  1, -1, -1,  1,  1,  1,  1,  1, -1, -1, -1,  1,  1,  1,  1, -1,
+     1,  1,  1, -1,  1,  1,  1, -1, -1,  1,  1, -1,  1,  1,  1,  1,  1, -1,  1, -1,  1,  1,  1,  1, -1, -1,  1, -1,  1,  1,  1, -1,
+     1,  1, -1, -1,  1,  1,  1,  1, -1,  1, -1, -1,  1,  1,  1, -1,  1, -1, -1, -1,  1,  1,  1, -1, -1, -1, -1, -1,  1,  1,  1,  1,
+     1,  1,  1,  1, -1,  1,  1, -1, -1,  1,  1,  1, -1,  1,  1,  1,  1, -1,  1,  1, -1,  1,  1,  1, -1, -1,  1,  1, -1,  1,  1, -1,
+     1,  1, -1,  1, -1,  1,  1,  1, -1,  1, -1,  1, -1,  1,  1, -1,  1, -1, -1,  1, -1,  1,  1, -1, -1, -1, -1,  1, -1,  1,  1,  1,
+     1,  1,  1, -1, -1,  1,  1,  1, -1,  1,  1, -1, -1,  1,  1, -1,  1, -1,  1, -1, -1,  1,  1, -1, -1, -1,  1, -1, -1,  1,  1,  1,
+     1,  1, -1, -1, -1,  1,  1, -1, -1,  1, -1, -1, -1,  1,  1,  1,  1, -1, -1, -1, -1,  1,  1,  1, -1, -1, -1, -1, -1,  1,  1, -1,
+     1,  1,  1,  1,  1, -1,  1, -1, -1,  1,  1,  1,  1, -1,  1,  1,  1, -1,  1,  1,  1, -1,  1,  1, -1, -1,  1,  1,  1, -1,  1, -1,
+     1,  1, -1,  1,  1, -1,  1,  1, -1,  1, -1,  1,  1, -1,  1, -1,  1, -1, -1,  1,  1, -1,  1, -1, -1, -1, -1,  1,  1, -1,  1,  1,
+     1,  1,  1, -1,  1, -1,  1,  1, -1,  1,  1, -1,  1, -1,  1, -1,  1, -1,  1, -1,  1, -1,  1, -1, -1, -1,  1, -1,  1, -1,  1,  1,
+     1,  1, -1, -1,  1, -1,  1, -1, -1,  1, -1, -1,  1, -1,  1,  1,  1, -1, -1, -1,  1, -1,  1,  1, -1, -1, -1, -1,  1, -1,  1, -1,
+     1,  1,  1,  1, -1, -1,  1,  1, -1,  1,  1,  1, -1, -1,  1, -1,  1, -1,  1,  1, -1, -1,  1, -1, -1, -1,  1,  1, -1, -1,  1,  1,
+     1,  1, -1,  1, -1, -1,  1, -1, -1,  1, -1,  1, -1, -1,  1,  1,  1, -1, -1,  1, -1, -1,  1,  1, -1, -1, -1,  1, -1, -1,  1, -1,
+     1,  1,  1, -1, -1, -1,  1, -1, -1,  1,  1, -1, -1, -1,  1,  1,  1, -1,  1, -1, -1, -1,  1,  1, -1, -1,  1, -1, -1, -1,  1, -1,
+     1,  1, -1, -1, -1, -1,  1,  1, -1,  1, -1, -1, -1, -1,  1, -1,  1, -1, -1, -1, -1, -1,  1, -1, -1, -1, -1, -1, -1, -1,  1,  1,
+     1,  1,  1,  1,  1,  1, -1, -1, -1,  1,  1,  1,  1,  1, -1,  1,  1, -1,  1,  1,  1,  1, -1,  1, -1, -1,  1,  1,  1,  1, -1, -1,
+     1,  1, -1,  1,  1,  1, -1,  1, -1,  1, -1,  1,  1,  1, -1, -1,  1, -1, -1,  1,  1,  1, -1, -1, -1, -1, -1,  1,  1,  1, -1,  1,
+     1,  1,  1, -1,  1,  1, -1,  1, -1,  1,  1, -1,  1,  1, -1, -1,  1, -1,  1, -1,  1,  1, -1, -1, -1, -1,  1, -1,  1,  1, -1,  1,
+     1,  1, -1, -1,  1,  1, -1, -1, -1,  1, -1, -1,  1,  1, -1,  1,  1, -1, -1, -1,  1,  1, -1,  1, -1, -1, -1, -1,  1,  1, -1, -1,
+     1,  1,  1,  1, -1,  1, -1,  1, -1,  1,  1,  1, -1,  1, -1, -1,  1, -1,  1,  1, -1,  1, -1, -1, -1, -1,  1,  1, -1,  1, -1,  1,
+     1,  1, -1,  1, -1,  1, -1, -1, -1,  1, -1,  1, -1,  1, -1,  1,  1, -1, -1,  1, -1,  1, -1,  1, -1, -1, -1,  1, -1,  1, -1, -1,
+     1,  1,  1, -1, -1,  1, -1, -1, -1,  1,  1, -1, -1,  1, -1,  1,  1, -1,  1, -1, -1,  1, -1,  1, -1, -1,  1, -1, -1,  1, -1, -1,
+     1,  1, -1, -1, -1,  1, -1,  1, -1,  1, -1, -1, -1,  1, -1, -1,  1, -1, -1, -1, -1,  1, -1, -1, -1, -1, -1, -1, -1,  1, -1,  1,
+     1,  1,  1,  1,  1, -1, -1,  1, -1,  1,  1,  1,  1, -1, -1, -1,  1, -1,  1,  1,  1, -1, -1, -1, -1, -1,  1,  1,  1, -1, -1,  1,
+     1,  1, -1,  1,  1, -1, -1, -1, -1,  1, -1,  1,  1, -1, -1,  1,  1, -1, -1,  1,  1, -1, -1,  1, -1, -1, -1,  1,  1, -1, -1, -1,
+     1,  1,  1, -1,  1, -1, -1, -1, -1,  1,  1, -1,  1, -1, -1,  1,  1, -1,  1, -1,  1, -1, -1,  1, -1, -1,  1, -1,  1, -1, -1, -1,
+     1,  1, -1, -1,  1, -1, -1,  1, -1,  1, -1, -1,  1, -1, -1, -1,  1, -1, -1, -1,  1, -1, -1, -1, -1, -1, -1, -1,  1, -1, -1,  1,
+     1,  1,  1,  1, -1, -1, -1, -1, -1,  1,  1,  1, -1, -1, -1,  1,  1, -1,  1,  1, -1, -1, -1,  1, -1, -1,  1,  1, -1, -1, -1, -1,
+     1,  1, -1,  1, -1, -1, -1,  1, -1,  1, -1,  1, -1, -1, -1, -1,  1, -1, -1,  1, -1, -1, -1, -1, -1, -1, -1,  1, -1, -1, -1,  1,
+     1,  1,  1, -1, -1, -1, -1,  1, -1,  1,  1, -1, -1, -1, -1, -1,  1, -1,  1, -1, -1, -1, -1, -1, -1, -1,  1, -1, -1, -1, -1,  1,
+     1,  1, -1, -1, -1, -1, -1, -1, -1,  1, -1, -1, -1, -1, -1,  1,  1, -1, -1, -1, -1, -1, -1,  1, -1, -1, -1, -1, -1, -1, -1, -1,
+};
+#endif
+
+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);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_iq2_xxs * GGML_RESTRICT x = vx;
+    const block_q8_K    * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+#if defined(__AVX2__)
+
+    const uint64_t * signs64 = (const uint64_t *)keven_signs_q2xs;
+
+    uint32_t aux32[4];
+    const uint8_t * aux8 = (const uint8_t *)aux32;
+
+    __m256 accumf = _mm256_setzero_ps();
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_CPU_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;
+        __m256i sumi1 = _mm256_setzero_si256();
+        __m256i sumi2 = _mm256_setzero_si256();
+        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
+            const __m256i q8_1 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32;
+            const __m256i q8_2 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32;
+            memcpy(aux32, q2, 4*sizeof(uint32_t)); q2 += 8;
+            const __m256i q2_1 = _mm256_set_epi64x(iq2xxs_grid[aux8[ 3]], iq2xxs_grid[aux8[ 2]], iq2xxs_grid[aux8[1]], iq2xxs_grid[aux8[0]]);
+            const __m256i q2_2 = _mm256_set_epi64x(iq2xxs_grid[aux8[11]], iq2xxs_grid[aux8[10]], iq2xxs_grid[aux8[9]], iq2xxs_grid[aux8[8]]);
+            const __m256i s2_1 = _mm256_set_epi64x(signs64[(aux32[1] >> 21) & 127], signs64[(aux32[1] >> 14) & 127],
+                                                   signs64[(aux32[1] >>  7) & 127], signs64[(aux32[1] >>  0) & 127]);
+            const __m256i s2_2 = _mm256_set_epi64x(signs64[(aux32[3] >> 21) & 127], signs64[(aux32[3] >> 14) & 127],
+                                                   signs64[(aux32[3] >>  7) & 127], signs64[(aux32[3] >>  0) & 127]);
+            const __m256i q8s_1 = _mm256_sign_epi8(q8_1, s2_1);
+            const __m256i q8s_2 = _mm256_sign_epi8(q8_2, s2_2);
+            const __m256i dot1  = _mm256_maddubs_epi16(q2_1, q8s_1);
+            const __m256i dot2  = _mm256_maddubs_epi16(q2_2, q8s_2);
+            const uint16_t ls1 = aux32[1] >> 28;
+            const uint16_t ls2 = aux32[3] >> 28;
+            const __m256i p1 = _mm256_madd_epi16(dot1, _mm256_set1_epi16(2*ls1+1));
+            const __m256i p2 = _mm256_madd_epi16(dot2, _mm256_set1_epi16(2*ls2+1));
+            sumi1 = _mm256_add_epi32(sumi1, p1);
+            sumi2 = _mm256_add_epi32(sumi2, p2);
+        }
+
+        accumf = _mm256_fmadd_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(_mm256_add_epi32(sumi1, sumi2)), accumf);
+
+    }
+
+    *s = 0.125f * hsum_float_8(accumf);
+
+#elif defined(__AVX__)
+    const uint64_t * signs64 = (const uint64_t *)keven_signs_q2xs;
+
+    uint32_t aux32[4];
+    const uint8_t * aux8 = (const uint8_t *)aux32;
+
+    __m256 accumf = _mm256_setzero_ps();
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_CPU_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;
+        __m128i sumi1_0 = _mm_setzero_si128();
+        __m128i sumi1_1 = _mm_setzero_si128();
+        __m128i sumi2_0 = _mm_setzero_si128();
+        __m128i sumi2_1 = _mm_setzero_si128();
+        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
+            const __m128i q8_1_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8_1_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8_2_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8_2_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            memcpy(aux32, q2, 4*sizeof(uint32_t)); q2 += 8;
+            const __m128i q2_1_0 = _mm_set_epi64x(iq2xxs_grid[aux8[1]], iq2xxs_grid[aux8[0]]);
+            const __m128i q2_1_1 = _mm_set_epi64x(iq2xxs_grid[aux8[3]], iq2xxs_grid[aux8[2]]);
+            const __m128i q2_2_0 = _mm_set_epi64x(iq2xxs_grid[aux8[9]], iq2xxs_grid[aux8[8]]);
+            const __m128i q2_2_1 = _mm_set_epi64x(iq2xxs_grid[aux8[11]], iq2xxs_grid[aux8[10]]);
+            const __m128i s2_1_0 = _mm_set_epi64x(signs64[(aux32[1] >>  7) & 127], signs64[(aux32[1] >>  0) & 127]);
+            const __m128i s2_1_1 = _mm_set_epi64x(signs64[(aux32[1] >> 21) & 127], signs64[(aux32[1] >> 14) & 127]);
+            const __m128i s2_2_0 = _mm_set_epi64x(signs64[(aux32[3] >>  7) & 127], signs64[(aux32[3] >>  0) & 127]);
+            const __m128i s2_2_1 = _mm_set_epi64x(signs64[(aux32[3] >> 21) & 127], signs64[(aux32[3] >> 14) & 127]);
+            const __m128i q8s_1_0 = _mm_sign_epi8(q8_1_0, s2_1_0);
+            const __m128i q8s_1_1 = _mm_sign_epi8(q8_1_1, s2_1_1);
+            const __m128i q8s_2_0 = _mm_sign_epi8(q8_2_0, s2_2_0);
+            const __m128i q8s_2_1 = _mm_sign_epi8(q8_2_1, s2_2_1);
+            const __m128i dot1_0  = _mm_maddubs_epi16(q2_1_0, q8s_1_0);
+            const __m128i dot1_1  = _mm_maddubs_epi16(q2_1_1, q8s_1_1);
+            const __m128i dot2_0  = _mm_maddubs_epi16(q2_2_0, q8s_2_0);
+            const __m128i dot2_1  = _mm_maddubs_epi16(q2_2_1, q8s_2_1);
+            const uint16_t ls1 = aux32[1] >> 28;
+            const uint16_t ls2 = aux32[3] >> 28;
+            const __m128i p1_0 = _mm_madd_epi16(dot1_0, _mm_set1_epi16(2*ls1+1));
+            const __m128i p1_1 = _mm_madd_epi16(dot1_1, _mm_set1_epi16(2*ls1+1));
+            const __m128i p2_0 = _mm_madd_epi16(dot2_0, _mm_set1_epi16(2*ls2+1));
+            const __m128i p2_1 = _mm_madd_epi16(dot2_1, _mm_set1_epi16(2*ls2+1));
+            sumi1_0 = _mm_add_epi32(sumi1_0, p1_0);
+            sumi1_1 = _mm_add_epi32(sumi1_1, p1_1);
+            sumi2_0 = _mm_add_epi32(sumi2_0, p2_0);
+            sumi2_1 = _mm_add_epi32(sumi2_1, p2_1);
+        }
+
+        accumf = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(MM256_SET_M128I(_mm_add_epi32(sumi1_1, sumi2_1), _mm_add_epi32(sumi1_0, sumi2_0)))), accumf);
+
+    }
+
+    *s = 0.125f * hsum_float_8(accumf);
+
+#else
+
+    uint32_t aux32[2];
+    const uint8_t * aux8 = (const uint8_t *)aux32;
+
+    float sumf = 0.f;
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_CPU_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;
+        int32_t bsum = 0;
+        for (int ib32 = 0; ib32 < QK_K/32; ++ib32) {
+            memcpy(aux32, q2, 2*sizeof(uint32_t));
+            q2 += 4;
+            const uint32_t ls = 2*(aux32[1] >> 28) + 1;
+            int32_t sumi = 0;
+            for (int l = 0; l < 4; ++l) {
+                const uint8_t * grid = (const uint8_t *)(iq2xxs_grid + aux8[l]);
+                const uint8_t  signs = ksigns_iq2xs[(aux32[1] >> 7*l) & 127];
+                for (int j = 0; j < 8; ++j) {
+                    sumi += grid[j] * q8[j] * (signs & kmask_iq2xs[j] ? -1 : 1);
+                }
+                q8 += 8;
+            }
+            bsum += sumi * ls;
+        }
+        sumf += d * bsum;
+    }
+    *s = 0.125f * sumf;
+#endif
+}
+
+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);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_iq2_xs * GGML_RESTRICT x = vx;
+    const block_q8_K   * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+#if defined(__AVX2__)
+
+    const __m256i mone = _mm256_set1_epi8(1);
+    static const char block_sign_shuffle_mask_1[32] = {
+        0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02,
+        0x04, 0x04, 0x04, 0x04, 0x04, 0x04, 0x04, 0x04, 0x06, 0x06, 0x06, 0x06, 0x06, 0x06, 0x06, 0x06,
+    };
+    static const char block_sign_shuffle_mask_2[32] = {
+        0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x0a, 0x0a, 0x0a, 0x0a, 0x0a, 0x0a, 0x0a, 0x0a,
+        0x0c, 0x0c, 0x0c, 0x0c, 0x0c, 0x0c, 0x0c, 0x0c, 0x0e, 0x0e, 0x0e, 0x0e, 0x0e, 0x0e, 0x0e, 0x0e,
+    };
+    static const uint8_t bit_selector_mask_bytes[32] = {
+        0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
+        0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
+    };
+
+    const __m256i bit_selector_mask = _mm256_loadu_si256((const __m256i*)bit_selector_mask_bytes);
+    const __m256i block_sign_shuffle_1 = _mm256_loadu_si256((const __m256i*)block_sign_shuffle_mask_1);
+    const __m256i block_sign_shuffle_2 = _mm256_loadu_si256((const __m256i*)block_sign_shuffle_mask_2);
+
+    static const uint8_t k_bit_helper[32] = {
+        0x00, 0x80, 0x80, 0x00, 0x80, 0x00, 0x00, 0x80, 0x80, 0x00, 0x00, 0x80, 0x00, 0x80, 0x80, 0x00,
+        0x00, 0x80, 0x80, 0x00, 0x80, 0x00, 0x00, 0x80, 0x80, 0x00, 0x00, 0x80, 0x00, 0x80, 0x80, 0x00,
+    };
+    const __m256i bit_helper = _mm256_loadu_si256((const __m256i*)k_bit_helper);
+    const __m256i m511 = _mm256_set1_epi16(511);
+    const __m128i m4 = _mm_set1_epi8(0xf);
+    const __m128i m1 = _mm_set1_epi8(1);
+
+    uint64_t aux64;
+
+    // somewhat hacky, but gives a significant boost in performance
+    __m256i aux_gindex;
+    const uint16_t * gindex = (const uint16_t *)&aux_gindex;
+
+    __m256 accumf = _mm256_setzero_ps();
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_CPU_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;
+
+        memcpy(&aux64, x[i].scales, 8);
+        __m128i stmp = _mm_set1_epi64x(aux64);
+        stmp = _mm_unpacklo_epi8(_mm_and_si128(stmp, m4), _mm_and_si128(_mm_srli_epi16(stmp, 4), m4));
+        const __m128i scales = _mm_add_epi8(_mm_slli_epi16(stmp, 1), m1);
+
+        __m256i sumi1 = _mm256_setzero_si256();
+        __m256i sumi2 = _mm256_setzero_si256();
+        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 4) {
+
+            const __m256i q2_data = _mm256_loadu_si256((const __m256i*)q2);  q2 += 16;
+            aux_gindex = _mm256_and_si256(q2_data, m511);
+
+            const __m256i partial_sign_bits = _mm256_srli_epi16(q2_data, 9);
+            const __m256i partial_sign_bits_upper = _mm256_srli_epi16(q2_data, 13);
+            const __m256i partial_sign_bits_for_counting = _mm256_xor_si256(partial_sign_bits, partial_sign_bits_upper);
+
+            const __m256i odd_bits = _mm256_shuffle_epi8(bit_helper, partial_sign_bits_for_counting);
+            const __m256i full_sign_bits = _mm256_or_si256(partial_sign_bits, odd_bits);
+
+            const __m256i q8_1 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32;
+            const __m256i q8_2 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32;
+            const __m256i q8_3 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32;
+            const __m256i q8_4 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32;
+
+            const __m256i q2_1 = _mm256_set_epi64x(iq2xs_grid[gindex[ 3]], iq2xs_grid[gindex[ 2]],
+                                                   iq2xs_grid[gindex[ 1]], iq2xs_grid[gindex[ 0]]);
+            const __m256i q2_2 = _mm256_set_epi64x(iq2xs_grid[gindex[ 7]], iq2xs_grid[gindex[ 6]],
+                                                   iq2xs_grid[gindex[ 5]], iq2xs_grid[gindex[ 4]]);
+            const __m256i q2_3 = _mm256_set_epi64x(iq2xs_grid[gindex[11]], iq2xs_grid[gindex[10]],
+                                                   iq2xs_grid[gindex[ 9]], iq2xs_grid[gindex[ 8]]);
+            const __m256i q2_4 = _mm256_set_epi64x(iq2xs_grid[gindex[15]], iq2xs_grid[gindex[14]],
+                                                   iq2xs_grid[gindex[13]], iq2xs_grid[gindex[12]]);
+
+            const __m128i full_signs_l = _mm256_castsi256_si128(full_sign_bits);
+            const __m128i full_signs_h = _mm256_extractf128_si256(full_sign_bits, 1);
+            const __m256i full_signs_1 = MM256_SET_M128I(full_signs_l, full_signs_l);
+            const __m256i full_signs_2 = MM256_SET_M128I(full_signs_h, full_signs_h);
+
+            __m256i signs;
+            signs = _mm256_shuffle_epi8(full_signs_1, block_sign_shuffle_1);
+            signs = _mm256_cmpeq_epi8(_mm256_and_si256(signs, bit_selector_mask), bit_selector_mask);
+            const __m256i q8s_1 = _mm256_sign_epi8(q8_1, _mm256_or_si256(signs, mone));
+
+            signs = _mm256_shuffle_epi8(full_signs_1, block_sign_shuffle_2);
+            signs = _mm256_cmpeq_epi8(_mm256_and_si256(signs, bit_selector_mask), bit_selector_mask);
+            const __m256i q8s_2 = _mm256_sign_epi8(q8_2, _mm256_or_si256(signs, mone));
+
+            signs = _mm256_shuffle_epi8(full_signs_2, block_sign_shuffle_1);
+            signs = _mm256_cmpeq_epi8(_mm256_and_si256(signs, bit_selector_mask), bit_selector_mask);
+            const __m256i q8s_3 = _mm256_sign_epi8(q8_3, _mm256_or_si256(signs, mone));
+
+            signs = _mm256_shuffle_epi8(full_signs_2, block_sign_shuffle_2);
+            signs = _mm256_cmpeq_epi8(_mm256_and_si256(signs, bit_selector_mask), bit_selector_mask);
+            const __m256i q8s_4 = _mm256_sign_epi8(q8_4, _mm256_or_si256(signs, mone));
+
+            const __m256i dot1  = _mm256_maddubs_epi16(q2_1, q8s_1);
+            const __m256i dot2  = _mm256_maddubs_epi16(q2_2, q8s_2);
+            const __m256i dot3  = _mm256_maddubs_epi16(q2_3, q8s_3);
+            const __m256i dot4  = _mm256_maddubs_epi16(q2_4, q8s_4);
+
+            const __m256i sc1 = _mm256_cvtepi8_epi16(_mm_shuffle_epi8(scales, get_scale_shuffle(ib32+0)));
+            const __m256i sc2 = _mm256_cvtepi8_epi16(_mm_shuffle_epi8(scales, get_scale_shuffle(ib32+1)));
+            const __m256i sc3 = _mm256_cvtepi8_epi16(_mm_shuffle_epi8(scales, get_scale_shuffle(ib32+2)));
+            const __m256i sc4 = _mm256_cvtepi8_epi16(_mm_shuffle_epi8(scales, get_scale_shuffle(ib32+3)));
+
+            sumi1 = _mm256_add_epi32(sumi1, _mm256_madd_epi16(dot1, sc1));
+            sumi2 = _mm256_add_epi32(sumi2, _mm256_madd_epi16(dot2, sc2));
+            sumi1 = _mm256_add_epi32(sumi1, _mm256_madd_epi16(dot3, sc3));
+            sumi2 = _mm256_add_epi32(sumi2, _mm256_madd_epi16(dot4, sc4));
+        }
+
+        accumf = _mm256_fmadd_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(_mm256_add_epi32(sumi1, sumi2)), accumf);
+
+    }
+
+    *s = 0.125f * hsum_float_8(accumf);
+
+#elif defined(__AVX__)
+    const __m128i mone = _mm_set1_epi8(1);
+    static const char block_sign_shuffle_mask_1[32] = {
+        0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02,
+        0x04, 0x04, 0x04, 0x04, 0x04, 0x04, 0x04, 0x04, 0x06, 0x06, 0x06, 0x06, 0x06, 0x06, 0x06, 0x06,
+    };
+    static const char block_sign_shuffle_mask_2[32] = {
+        0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x0a, 0x0a, 0x0a, 0x0a, 0x0a, 0x0a, 0x0a, 0x0a,
+        0x0c, 0x0c, 0x0c, 0x0c, 0x0c, 0x0c, 0x0c, 0x0c, 0x0e, 0x0e, 0x0e, 0x0e, 0x0e, 0x0e, 0x0e, 0x0e,
+    };
+    static const uint8_t bit_selector_mask_bytes[32] = {
+        0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
+        0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
+    };
+
+    const __m128i bit_selector_mask_0 = _mm_loadu_si128((const __m128i*)bit_selector_mask_bytes);
+    const __m128i bit_selector_mask_1 = _mm_loadu_si128((const __m128i*)bit_selector_mask_bytes + 1);
+    const __m128i block_sign_shuffle_1_0 = _mm_loadu_si128((const __m128i*)block_sign_shuffle_mask_1);
+    const __m128i block_sign_shuffle_1_1 = _mm_loadu_si128((const __m128i*)block_sign_shuffle_mask_1 + 1);
+    const __m128i block_sign_shuffle_2_0 = _mm_loadu_si128((const __m128i*)block_sign_shuffle_mask_2);
+    const __m128i block_sign_shuffle_2_1 = _mm_loadu_si128((const __m128i*)block_sign_shuffle_mask_2 + 1);
+
+    static const uint8_t k_bit_helper[32] = {
+        0x00, 0x80, 0x80, 0x00, 0x80, 0x00, 0x00, 0x80, 0x80, 0x00, 0x00, 0x80, 0x00, 0x80, 0x80, 0x00,
+        0x00, 0x80, 0x80, 0x00, 0x80, 0x00, 0x00, 0x80, 0x80, 0x00, 0x00, 0x80, 0x00, 0x80, 0x80, 0x00,
+    };
+    const __m128i bit_helper_0 = _mm_loadu_si128((const __m128i*)k_bit_helper);
+    const __m128i bit_helper_1 = _mm_loadu_si128((const __m128i*)k_bit_helper + 1);
+    const __m128i m511 = _mm_set1_epi16(511);
+    const __m128i m4 = _mm_set1_epi8(0xf);
+    const __m128i m1 = _mm_set1_epi8(1);
+
+    uint64_t aux64;
+
+    // somewhat hacky, but gives a significant boost in performance
+    __m256i aux_gindex;
+    const uint16_t * gindex = (const uint16_t *)&aux_gindex;
+
+    __m256 accumf = _mm256_setzero_ps();
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_CPU_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;
+
+        memcpy(&aux64, x[i].scales, 8);
+        __m128i stmp = _mm_set1_epi64x(aux64);
+        stmp = _mm_unpacklo_epi8(_mm_and_si128(stmp, m4), _mm_and_si128(_mm_srli_epi16(stmp, 4), m4));
+        const __m128i scales = _mm_add_epi8(_mm_slli_epi16(stmp, 1), m1);
+
+        __m128i sumi1_0 = _mm_setzero_si128();
+        __m128i sumi1_1 = _mm_setzero_si128();
+        __m128i sumi2_0 = _mm_setzero_si128();
+        __m128i sumi2_1 = _mm_setzero_si128();
+        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 4) {
+
+            const __m128i q2_data_0 = _mm_loadu_si128((const __m128i*)q2);
+            const __m128i q2_data_1 = _mm_loadu_si128((const __m128i*)q2 + 1);  q2 += 16;
+            aux_gindex = MM256_SET_M128I(_mm_and_si128(q2_data_1, m511), _mm_and_si128(q2_data_0, m511));
+
+            const __m128i partial_sign_bits_0 = _mm_srli_epi16(q2_data_0, 9);
+            const __m128i partial_sign_bits_1 = _mm_srli_epi16(q2_data_1, 9);
+            const __m128i partial_sign_bits_upper_0 = _mm_srli_epi16(q2_data_0, 13);
+            const __m128i partial_sign_bits_upper_1 = _mm_srli_epi16(q2_data_1, 13);
+            const __m128i partial_sign_bits_for_counting_0 = _mm_xor_si128(partial_sign_bits_0, partial_sign_bits_upper_0);
+            const __m128i partial_sign_bits_for_counting_1 = _mm_xor_si128(partial_sign_bits_1, partial_sign_bits_upper_1);
+
+            const __m128i odd_bits_0 = _mm_shuffle_epi8(bit_helper_0, partial_sign_bits_for_counting_0);
+            const __m128i odd_bits_1 = _mm_shuffle_epi8(bit_helper_1, partial_sign_bits_for_counting_1);
+            const __m128i full_sign_bits_0 = _mm_or_si128(partial_sign_bits_0, odd_bits_0);
+            const __m128i full_sign_bits_1 = _mm_or_si128(partial_sign_bits_1, odd_bits_1);
+
+            const __m128i q8_1_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8_1_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8_2_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8_2_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8_3_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8_3_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8_4_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8_4_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+
+            const __m128i q2_1_0 = _mm_set_epi64x(iq2xs_grid[gindex[1]], iq2xs_grid[gindex[0]]);
+            const __m128i q2_1_1 = _mm_set_epi64x(iq2xs_grid[gindex[3]], iq2xs_grid[gindex[2]]);
+            const __m128i q2_2_0 = _mm_set_epi64x(iq2xs_grid[gindex[5]], iq2xs_grid[gindex[4]]);
+            const __m128i q2_2_1 = _mm_set_epi64x(iq2xs_grid[gindex[7]], iq2xs_grid[gindex[6]]);
+            const __m128i q2_3_0 = _mm_set_epi64x(iq2xs_grid[gindex[9]], iq2xs_grid[gindex[8]]);
+            const __m128i q2_3_1 = _mm_set_epi64x(iq2xs_grid[gindex[11]], iq2xs_grid[gindex[10]]);
+            const __m128i q2_4_0 = _mm_set_epi64x(iq2xs_grid[gindex[13]], iq2xs_grid[gindex[12]]);
+            const __m128i q2_4_1 = _mm_set_epi64x(iq2xs_grid[gindex[15]], iq2xs_grid[gindex[14]]);
+
+            // AVX2 full_signs_1 is full_sign_bits_0 here
+            // AVX2 full_signs_2 is full_sign_bits_1 here
+            __m128i signs_0, signs_1;
+            signs_0 = _mm_shuffle_epi8(full_sign_bits_0, block_sign_shuffle_1_0);
+            signs_1 = _mm_shuffle_epi8(full_sign_bits_0, block_sign_shuffle_1_1);
+            signs_0 = _mm_cmpeq_epi8(_mm_and_si128(signs_0, bit_selector_mask_0), bit_selector_mask_0);
+            signs_1 = _mm_cmpeq_epi8(_mm_and_si128(signs_1, bit_selector_mask_1), bit_selector_mask_1);
+            const __m128i q8s_1_0 = _mm_sign_epi8(q8_1_0, _mm_or_si128(signs_0, mone));
+            const __m128i q8s_1_1 = _mm_sign_epi8(q8_1_1, _mm_or_si128(signs_1, mone));
+
+            signs_0 = _mm_shuffle_epi8(full_sign_bits_0, block_sign_shuffle_2_0);
+            signs_1 = _mm_shuffle_epi8(full_sign_bits_0, block_sign_shuffle_2_1);
+            signs_0 = _mm_cmpeq_epi8(_mm_and_si128(signs_0, bit_selector_mask_0), bit_selector_mask_0);
+            signs_1 = _mm_cmpeq_epi8(_mm_and_si128(signs_1, bit_selector_mask_1), bit_selector_mask_1);
+            const __m128i q8s_2_0 = _mm_sign_epi8(q8_2_0, _mm_or_si128(signs_0, mone));
+            const __m128i q8s_2_1 = _mm_sign_epi8(q8_2_1, _mm_or_si128(signs_1, mone));
+
+            signs_0 = _mm_shuffle_epi8(full_sign_bits_1, block_sign_shuffle_1_0);
+            signs_1 = _mm_shuffle_epi8(full_sign_bits_1, block_sign_shuffle_1_1);
+            signs_0 = _mm_cmpeq_epi8(_mm_and_si128(signs_0, bit_selector_mask_0), bit_selector_mask_0);
+            signs_1 = _mm_cmpeq_epi8(_mm_and_si128(signs_1, bit_selector_mask_1), bit_selector_mask_1);
+            const __m128i q8s_3_0 = _mm_sign_epi8(q8_3_0, _mm_or_si128(signs_0, mone));
+            const __m128i q8s_3_1 = _mm_sign_epi8(q8_3_1, _mm_or_si128(signs_1, mone));
+
+            signs_0 = _mm_shuffle_epi8(full_sign_bits_1, block_sign_shuffle_2_0);
+            signs_1 = _mm_shuffle_epi8(full_sign_bits_1, block_sign_shuffle_2_1);
+            signs_0 = _mm_cmpeq_epi8(_mm_and_si128(signs_0, bit_selector_mask_0), bit_selector_mask_0);
+            signs_1 = _mm_cmpeq_epi8(_mm_and_si128(signs_1, bit_selector_mask_1), bit_selector_mask_1);
+            const __m128i q8s_4_0 = _mm_sign_epi8(q8_4_0, _mm_or_si128(signs_0, mone));
+            const __m128i q8s_4_1 = _mm_sign_epi8(q8_4_1, _mm_or_si128(signs_1, mone));
+
+            const __m128i dot1_0  = _mm_maddubs_epi16(q2_1_0, q8s_1_0);
+            const __m128i dot1_1  = _mm_maddubs_epi16(q2_1_1, q8s_1_1);
+            const __m128i dot2_0  = _mm_maddubs_epi16(q2_2_0, q8s_2_0);
+            const __m128i dot2_1  = _mm_maddubs_epi16(q2_2_1, q8s_2_1);
+            const __m128i dot3_0  = _mm_maddubs_epi16(q2_3_0, q8s_3_0);
+            const __m128i dot3_1  = _mm_maddubs_epi16(q2_3_1, q8s_3_1);
+            const __m128i dot4_0  = _mm_maddubs_epi16(q2_4_0, q8s_4_0);
+            const __m128i dot4_1  = _mm_maddubs_epi16(q2_4_1, q8s_4_1);
+
+            __m128i sc_tmp = _mm_shuffle_epi8(scales, get_scale_shuffle(ib32+0));
+            const __m128i sc1_0 = _mm_cvtepi8_epi16(sc_tmp);
+            const __m128i sc1_1 = _mm_cvtepi8_epi16(_mm_srli_si128(sc_tmp, 8));
+            sc_tmp = _mm_shuffle_epi8(scales, get_scale_shuffle(ib32+1));
+            const __m128i sc2_0 = _mm_cvtepi8_epi16(sc_tmp);
+            const __m128i sc2_1 = _mm_cvtepi8_epi16(_mm_srli_si128(sc_tmp, 8));
+            sc_tmp = _mm_shuffle_epi8(scales, get_scale_shuffle(ib32+2));
+            const __m128i sc3_0 = _mm_cvtepi8_epi16(sc_tmp);
+            const __m128i sc3_1 = _mm_cvtepi8_epi16(_mm_srli_si128(sc_tmp, 8));
+            sc_tmp = _mm_shuffle_epi8(scales, get_scale_shuffle(ib32+3));
+            const __m128i sc4_0 = _mm_cvtepi8_epi16(sc_tmp);
+            const __m128i sc4_1 = _mm_cvtepi8_epi16(_mm_srli_si128(sc_tmp, 8));
+
+            sumi1_0 = _mm_add_epi32(sumi1_0, _mm_madd_epi16(dot1_0, sc1_0));
+            sumi1_1 = _mm_add_epi32(sumi1_1, _mm_madd_epi16(dot1_1, sc1_1));
+            sumi2_0 = _mm_add_epi32(sumi2_0, _mm_madd_epi16(dot2_0, sc2_0));
+            sumi2_1 = _mm_add_epi32(sumi2_1, _mm_madd_epi16(dot2_1, sc2_1));
+            sumi1_0 = _mm_add_epi32(sumi1_0, _mm_madd_epi16(dot3_0, sc3_0));
+            sumi1_1 = _mm_add_epi32(sumi1_1, _mm_madd_epi16(dot3_1, sc3_1));
+            sumi2_0 = _mm_add_epi32(sumi2_0, _mm_madd_epi16(dot4_0, sc4_0));
+            sumi2_1 = _mm_add_epi32(sumi2_1, _mm_madd_epi16(dot4_1, sc4_1));
+        }
+
+        accumf = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(MM256_SET_M128I(_mm_add_epi32(sumi1_1, sumi2_1), _mm_add_epi32(sumi1_0, sumi2_0)))), accumf);
+
+    }
+
+    *s = 0.125f * hsum_float_8(accumf);
+
+#else
+
+    float sumf = 0.f;
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
+        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;
+            const uint16_t ls2 = 2*(sc[ib32] >>  4) + 1;
+            int32_t sumi = 0;
+            for (int l = 0; l < 2; ++l) {
+                const uint8_t * grid = (const uint8_t *)(iq2xs_grid + (q2[l] & 511));
+                const uint8_t  signs = ksigns_iq2xs[q2[l] >> 9];
+                for (int j = 0; j < 8; ++j) {
+                    sumi += grid[j] * q8[j] * (signs & kmask_iq2xs[j] ? -1 : 1);
+                }
+                q8 += 8;
+            }
+            bsum += sumi * ls1;
+            sumi = 0;
+            for (int l = 2; l < 4; ++l) {
+                const uint8_t * grid = (const uint8_t *)(iq2xs_grid + (q2[l] & 511));
+                const uint8_t  signs = ksigns_iq2xs[q2[l] >> 9];
+                for (int j = 0; j < 8; ++j) {
+                    sumi += grid[j] * q8[j] * (signs & kmask_iq2xs[j] ? -1 : 1);
+                }
+                q8 += 8;
+            }
+            bsum += sumi * ls2;
+            q2 += 4;
+        }
+        sumf += d * bsum;
+    }
+    *s = 0.125f * sumf;
+#endif
+}
+
+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);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_iq2_s * GGML_RESTRICT x = vx;
+    const block_q8_K  * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+#if defined(__AVX2__)
+
+   static const uint8_t k_mask1[32] = {0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01,
+                                       0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03
+   };
+
+    static const uint8_t k_mask2[32] = {0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
+                                        0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
+    };
+
+    const __m128i m4 = _mm_set1_epi8(0xf);
+    const __m128i m1 = _mm_set1_epi8(1);
+
+    const __m256i mask1 = _mm256_loadu_si256((const __m256i*)k_mask1);
+    const __m256i mask2 = _mm256_loadu_si256((const __m256i*)k_mask2);
+
+    uint64_t aux64;
+
+    __m256 accumf = _mm256_setzero_ps();
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
+        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);
+        const __m256i scales16 = _mm256_cvtepi8_epi16(scales8); // 0 2 4 6 8 10 12 14 1 3 5 7 9 11 13 15
+
+        __m256i sumi1 = _mm256_setzero_si256();
+        __m256i sumi2 = _mm256_setzero_si256();
+        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
+            const __m256i q8_1 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32;
+            const __m256i q8_2 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32;
+            const __m256i q2_1 = _mm256_set_epi64x(iq2s_grid[qs[3] | ((qh[ib32+0] << 2) & 0x300)],
+                                                   iq2s_grid[qs[2] | ((qh[ib32+0] << 4) & 0x300)],
+                                                   iq2s_grid[qs[1] | ((qh[ib32+0] << 6) & 0x300)],
+                                                   iq2s_grid[qs[0] | ((qh[ib32+0] << 8) & 0x300)]);
+            const __m256i q2_2 = _mm256_set_epi64x(iq2s_grid[qs[7] | ((qh[ib32+1] << 2) & 0x300)],
+                                                   iq2s_grid[qs[6] | ((qh[ib32+1] << 4) & 0x300)],
+                                                   iq2s_grid[qs[5] | ((qh[ib32+1] << 6) & 0x300)],
+                                                   iq2s_grid[qs[4] | ((qh[ib32+1] << 8) & 0x300)]);
+            qs += 8;
+
+            __m256i aux256 = _mm256_set1_epi32(signs[0] | ((uint32_t) signs[1] << 16));
+            aux256 = _mm256_and_si256(_mm256_shuffle_epi8(aux256,mask1), mask2);
+            const __m256i s2_1 = _mm256_cmpeq_epi8(aux256, mask2);
+            const __m256i q8s_1 = _mm256_sub_epi8(_mm256_xor_si256(s2_1, q8_1), s2_1);
+
+            aux256 = _mm256_set1_epi32(signs[2] | ((uint32_t) signs[3] << 16));
+            aux256 = _mm256_and_si256(_mm256_shuffle_epi8(aux256,mask1), mask2);
+            const __m256i s2_2 = _mm256_cmpeq_epi8(aux256, mask2);
+            const __m256i q8s_2 = _mm256_sub_epi8(_mm256_xor_si256(s2_2, q8_2), s2_2);
+
+            signs += 4;
+
+            const __m256i dot1  = _mm256_maddubs_epi16(q2_1, q8s_1); // blocks 2*ib32+0, 2*ib32+1
+            const __m256i dot2  = _mm256_maddubs_epi16(q2_2, q8s_2); // blocks 2*ib32+2, 2*ib32+3
+
+            const __m256i p1 = _mm256_madd_epi16(dot1, _mm256_shuffle_epi8(scales16, get_scale_shuffle_k4(ib32+0)));
+            const __m256i p2 = _mm256_madd_epi16(dot2, _mm256_shuffle_epi8(scales16, get_scale_shuffle_k4(ib32+1)));
+            sumi1 = _mm256_add_epi32(sumi1, p1);
+            sumi2 = _mm256_add_epi32(sumi2, p2);
+        }
+
+        accumf = _mm256_fmadd_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(_mm256_add_epi32(sumi1, sumi2)), accumf);
+
+    }
+
+    *s = 0.125f * hsum_float_8(accumf);
+
+#elif defined(__AVX__)
+   static const uint8_t k_mask1[32] = {0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01,
+                                       0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03
+   };
+
+    static const uint8_t k_mask2[32] = {0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
+                                        0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
+    };
+
+    const __m128i m4 = _mm_set1_epi8(0xf);
+    const __m128i m1 = _mm_set1_epi8(1);
+
+    const __m128i mask1_0 = _mm_loadu_si128((const __m128i*)k_mask1);
+    const __m128i mask1_1 = _mm_loadu_si128((const __m128i*)k_mask1 + 1);
+    const __m128i mask2_0 = _mm_loadu_si128((const __m128i*)k_mask2);
+    const __m128i mask2_1 = _mm_loadu_si128((const __m128i*)k_mask2 + 1);
+
+    uint64_t aux64;
+
+    __m256 accumf = _mm256_setzero_ps();
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
+        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);
+        const __m128i scales16_0 = _mm_cvtepi8_epi16(scales8);
+        const __m128i scales16_1 = _mm_cvtepi8_epi16(_mm_srli_si128(scales8, 8));
+
+        __m128i sumi1_0 = _mm_setzero_si128();
+        __m128i sumi1_1 = _mm_setzero_si128();
+        __m128i sumi2_0 = _mm_setzero_si128();
+        __m128i sumi2_1 = _mm_setzero_si128();
+        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
+            const __m128i q8_1_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8_1_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8_2_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8_2_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q2_1_0 = _mm_set_epi64x(iq2s_grid[qs[1] | ((qh[ib32+0] << 6) & 0x300)],
+                                                  iq2s_grid[qs[0] | ((qh[ib32+0] << 8) & 0x300)]);
+            const __m128i q2_1_1 = _mm_set_epi64x(iq2s_grid[qs[3] | ((qh[ib32+0] << 2) & 0x300)],
+                                                  iq2s_grid[qs[2] | ((qh[ib32+0] << 4) & 0x300)]);
+            const __m128i q2_2_0 = _mm_set_epi64x(iq2s_grid[qs[5] | ((qh[ib32+1] << 6) & 0x300)],
+                                                  iq2s_grid[qs[4] | ((qh[ib32+1] << 8) & 0x300)]);
+            const __m128i q2_2_1 = _mm_set_epi64x(iq2s_grid[qs[7] | ((qh[ib32+1] << 2) & 0x300)],
+                                                  iq2s_grid[qs[6] | ((qh[ib32+1] << 4) & 0x300)]);
+            qs += 8;
+
+            __m128i aux128_0 = _mm_set1_epi32(signs[0] | ((uint32_t) signs[1] << 16));
+            __m128i aux128_1 = aux128_0;
+            aux128_0 = _mm_and_si128(_mm_shuffle_epi8(aux128_0,mask1_0), mask2_0);
+            aux128_1 = _mm_and_si128(_mm_shuffle_epi8(aux128_1,mask1_1), mask2_1);
+            const __m128i s2_1_0 = _mm_cmpeq_epi8(aux128_0, mask2_0);
+            const __m128i s2_1_1 = _mm_cmpeq_epi8(aux128_1, mask2_1);
+            const __m128i q8s_1_0 = _mm_sub_epi8(_mm_xor_si128(s2_1_0, q8_1_0), s2_1_0);
+            const __m128i q8s_1_1 = _mm_sub_epi8(_mm_xor_si128(s2_1_1, q8_1_1), s2_1_1);
+
+            aux128_0 = _mm_set1_epi32(signs[2] | ((uint32_t) signs[3] << 16));
+            aux128_1 = aux128_0;
+            aux128_0 = _mm_and_si128(_mm_shuffle_epi8(aux128_0,mask1_0), mask2_0);
+            aux128_1 = _mm_and_si128(_mm_shuffle_epi8(aux128_1,mask1_1), mask2_1);
+            const __m128i s2_2_0 = _mm_cmpeq_epi8(aux128_0, mask2_0);
+            const __m128i s2_2_1 = _mm_cmpeq_epi8(aux128_1, mask2_1);
+            const __m128i q8s_2_0 = _mm_sub_epi8(_mm_xor_si128(s2_2_0, q8_2_0), s2_2_0);
+            const __m128i q8s_2_1 = _mm_sub_epi8(_mm_xor_si128(s2_2_1, q8_2_1), s2_2_1);
+
+            signs += 4;
+
+            const __m128i dot1_0  = _mm_maddubs_epi16(q2_1_0, q8s_1_0);
+            const __m128i dot1_1  = _mm_maddubs_epi16(q2_1_1, q8s_1_1);
+            const __m128i dot2_0  = _mm_maddubs_epi16(q2_2_0, q8s_2_0);
+            const __m128i dot2_1  = _mm_maddubs_epi16(q2_2_1, q8s_2_1);
+
+            const __m128i p1_0 = _mm_madd_epi16(dot1_0, _mm_shuffle_epi8(scales16_0, _mm256_extractf128_si256(get_scale_shuffle_k4(ib32+0), 0)));
+            const __m128i p1_1 = _mm_madd_epi16(dot1_1, _mm_shuffle_epi8(scales16_1, _mm256_extractf128_si256(get_scale_shuffle_k4(ib32+0), 1)));
+            const __m128i p2_0 = _mm_madd_epi16(dot2_0, _mm_shuffle_epi8(scales16_0, _mm256_extractf128_si256(get_scale_shuffle_k4(ib32+1), 0)));
+            const __m128i p2_1 = _mm_madd_epi16(dot2_1, _mm_shuffle_epi8(scales16_1, _mm256_extractf128_si256(get_scale_shuffle_k4(ib32+1), 1)));
+            sumi1_0 = _mm_add_epi32(sumi1_0, p1_0);
+            sumi1_1 = _mm_add_epi32(sumi1_1, p1_1);
+            sumi2_0 = _mm_add_epi32(sumi2_0, p2_0);
+            sumi2_1 = _mm_add_epi32(sumi2_1, p2_1);
+        }
+
+        accumf = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(MM256_SET_M128I(_mm_add_epi32(sumi1_1, sumi2_1), _mm_add_epi32(sumi1_0, sumi2_0)))), accumf);
+
+    }
+
+    *s = 0.125f * hsum_float_8(accumf);
+
+#else
+
+    float sumf = 0;
+    for (int i = 0; i < nb; i++) {
+
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
+        const int8_t  * q8 = y[i].qs;
+        const uint8_t * qs = x[i].qs;
+        const uint8_t * qh = x[i].qh;
+        const uint8_t * signs = qs + QK_K/8;
+
+        int bsum = 0;
+        for (int ib32 = 0; ib32 < QK_K/32; ++ib32) {
+            int ls1 = 1 + 2*(x[i].scales[ib32] & 0xf);
+            int ls2 = 1 + 2*(x[i].scales[ib32] >>  4);
+            int sumi1 = 0, sumi2 = 0;
+            for (int l = 0; l < 2; ++l) {
+                const uint8_t * grid = (const uint8_t *)(iq2s_grid + (qs[l] | (qh[ib32] << (8-2*l) & 0x300)));
+                for (int j = 0; j < 8; ++j) {
+                    sumi1 += q8[j] * grid[j] * (signs[l] & kmask_iq2xs[j] ? -1 : 1);
+                }
+                q8 += 8;
+            }
+            for (int l = 2; l < 4; ++l) {
+                const uint8_t * grid = (const uint8_t *)(iq2s_grid + (qs[l] | (qh[ib32] << (8-2*l) & 0x300)));
+                for (int j = 0; j < 8; ++j) {
+                    sumi2 += q8[j] * grid[j] * (signs[l] & kmask_iq2xs[j] ? -1 : 1);
+                }
+                q8 += 8;
+            }
+            bsum += ls1 * sumi1 + ls2 * sumi2;
+            qs += 4;
+            signs += 4;
+        }
+
+        sumf += d * bsum;
+    }
+
+    *s = 0.125f * sumf;
+
+#endif
+
+}
+
+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);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_iq3_xxs * GGML_RESTRICT x = vx;
+    const block_q8_K    * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+#if defined(__AVX2__)
+
+    const uint64_t * signs64 = (const uint64_t *)keven_signs_q2xs;
+
+    uint32_t aux32[2];
+
+    __m256 accumf = _mm256_setzero_ps();
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
+        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) {
+            const __m256i q8_1 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32;
+            const __m256i q8_2 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32;
+            const __m256i q2_1 = _mm256_set_epi32(iq3xxs_grid[q3[7]], iq3xxs_grid[q3[6]], iq3xxs_grid[q3[5]], iq3xxs_grid[q3[4]],
+                                                  iq3xxs_grid[q3[3]], iq3xxs_grid[q3[2]], iq3xxs_grid[q3[1]], iq3xxs_grid[q3[0]]);
+            q3 += 8;
+            const __m256i q2_2 = _mm256_set_epi32(iq3xxs_grid[q3[7]], iq3xxs_grid[q3[6]], iq3xxs_grid[q3[5]], iq3xxs_grid[q3[4]],
+                                                  iq3xxs_grid[q3[3]], iq3xxs_grid[q3[2]], iq3xxs_grid[q3[1]], iq3xxs_grid[q3[0]]);
+            q3 += 8;
+            memcpy(aux32, gas, 8); gas += 8;
+            const __m256i s2_1 = _mm256_set_epi64x(signs64[(aux32[0] >> 21) & 127], signs64[(aux32[0] >> 14) & 127],
+                                                   signs64[(aux32[0] >>  7) & 127], signs64[(aux32[0] >>  0) & 127]);
+            const __m256i s2_2 = _mm256_set_epi64x(signs64[(aux32[1] >> 21) & 127], signs64[(aux32[1] >> 14) & 127],
+                                                   signs64[(aux32[1] >>  7) & 127], signs64[(aux32[1] >>  0) & 127]);
+            const __m256i q8s_1 = _mm256_sign_epi8(q8_1, s2_1);
+            const __m256i q8s_2 = _mm256_sign_epi8(q8_2, s2_2);
+            const __m256i dot1  = _mm256_maddubs_epi16(q2_1, q8s_1);
+            const __m256i dot2  = _mm256_maddubs_epi16(q2_2, q8s_2);
+            const uint16_t ls1 = aux32[0] >> 28;
+            const uint16_t ls2 = aux32[1] >> 28;
+            const __m256i p1 = _mm256_madd_epi16(dot1, _mm256_set1_epi16(2*ls1+1));
+            const __m256i p2 = _mm256_madd_epi16(dot2, _mm256_set1_epi16(2*ls2+1));
+            sumi1 = _mm256_add_epi32(sumi1, p1);
+            sumi2 = _mm256_add_epi32(sumi2, p2);
+        }
+
+        accumf = _mm256_fmadd_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(_mm256_add_epi32(sumi1, sumi2)), accumf);
+
+    }
+
+    *s = 0.25f * hsum_float_8(accumf);
+
+#elif defined(__AVX__)
+    const uint64_t * signs64 = (const uint64_t *)keven_signs_q2xs;
+
+    uint32_t aux32[2];
+
+    __m256 accumf = _mm256_setzero_ps();
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
+        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();
+        __m128i sumi2_1 = _mm_setzero_si128();
+        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
+            const __m128i q8_1_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8_1_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8_2_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8_2_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q2_1_0 = _mm_set_epi32(iq3xxs_grid[q3[3]], iq3xxs_grid[q3[2]], iq3xxs_grid[q3[1]], iq3xxs_grid[q3[0]]);
+            const __m128i q2_1_1 = _mm_set_epi32(iq3xxs_grid[q3[7]], iq3xxs_grid[q3[6]], iq3xxs_grid[q3[5]], iq3xxs_grid[q3[4]]);
+            q3 += 8;
+            const __m128i q2_2_0 = _mm_set_epi32(iq3xxs_grid[q3[3]], iq3xxs_grid[q3[2]], iq3xxs_grid[q3[1]], iq3xxs_grid[q3[0]]);
+            const __m128i q2_2_1 = _mm_set_epi32(iq3xxs_grid[q3[7]], iq3xxs_grid[q3[6]], iq3xxs_grid[q3[5]], iq3xxs_grid[q3[4]]);
+            q3 += 8;
+            memcpy(aux32, gas, 8); gas += 8;
+            const __m128i s2_1_0 = _mm_set_epi64x(signs64[(aux32[0] >>  7) & 127], signs64[(aux32[0] >>  0) & 127]);
+            const __m128i s2_1_1 = _mm_set_epi64x(signs64[(aux32[0] >> 21) & 127], signs64[(aux32[0] >> 14) & 127]);
+            const __m128i s2_2_0 = _mm_set_epi64x(signs64[(aux32[1] >>  7) & 127], signs64[(aux32[1] >>  0) & 127]);
+            const __m128i s2_2_1 = _mm_set_epi64x(signs64[(aux32[1] >> 21) & 127], signs64[(aux32[1] >> 14) & 127]);
+            const __m128i q8s_1_0 = _mm_sign_epi8(q8_1_0, s2_1_0);
+            const __m128i q8s_1_1 = _mm_sign_epi8(q8_1_1, s2_1_1);
+            const __m128i q8s_2_0 = _mm_sign_epi8(q8_2_0, s2_2_0);
+            const __m128i q8s_2_1 = _mm_sign_epi8(q8_2_1, s2_2_1);
+            const __m128i dot1_0  = _mm_maddubs_epi16(q2_1_0, q8s_1_0);
+            const __m128i dot1_1  = _mm_maddubs_epi16(q2_1_1, q8s_1_1);
+            const __m128i dot2_0  = _mm_maddubs_epi16(q2_2_0, q8s_2_0);
+            const __m128i dot2_1  = _mm_maddubs_epi16(q2_2_1, q8s_2_1);
+            const uint16_t ls1 = aux32[0] >> 28;
+            const uint16_t ls2 = aux32[1] >> 28;
+            const __m128i p1_0 = _mm_madd_epi16(dot1_0, _mm_set1_epi16(2*ls1+1));
+            const __m128i p1_1 = _mm_madd_epi16(dot1_1, _mm_set1_epi16(2*ls1+1));
+            const __m128i p2_0 = _mm_madd_epi16(dot2_0, _mm_set1_epi16(2*ls2+1));
+            const __m128i p2_1 = _mm_madd_epi16(dot2_1, _mm_set1_epi16(2*ls2+1));
+            sumi1_0 = _mm_add_epi32(sumi1_0, p1_0);
+            sumi1_1 = _mm_add_epi32(sumi1_1, p1_1);
+            sumi2_0 = _mm_add_epi32(sumi2_0, p2_0);
+            sumi2_1 = _mm_add_epi32(sumi2_1, p2_1);
+        }
+
+        accumf = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(MM256_SET_M128I(_mm_add_epi32(sumi1_1, sumi2_1), _mm_add_epi32(sumi1_0, sumi2_0)))), accumf);
+
+    }
+
+    *s = 0.25f * hsum_float_8(accumf);
+
+#else
+
+    uint32_t aux32;
+
+    float sumf = 0.f;
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
+        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);
+            const uint32_t ls = 2*(aux32 >> 28) + 1;
+            int32_t sumi = 0;
+            for (int l = 0; l < 4; ++l) {
+                const uint8_t * grid1 = (const uint8_t *)(iq3xxs_grid + q3[2*l+0]);
+                const uint8_t * grid2 = (const uint8_t *)(iq3xxs_grid + q3[2*l+1]);
+                const uint8_t  signs = ksigns_iq2xs[(aux32 >> 7*l) & 127];
+                for (int j = 0; j < 4; ++j) {
+                    sumi += grid1[j] * q8[j+0] * (signs & kmask_iq2xs[j+0] ? -1 : 1);
+                    sumi += grid2[j] * q8[j+4] * (signs & kmask_iq2xs[j+4] ? -1 : 1);
+                }
+                q8 += 8;
+            }
+            q3 += 8;
+            bsum += sumi * ls;
+        }
+        sumf += d * bsum;
+    }
+    *s = 0.25f * sumf;
+#endif
+}
+
+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);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_iq3_s * GGML_RESTRICT x = vx;
+    const block_q8_K  * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+#if defined(__AVX2__)
+
+   static const uint8_t k_mask1[32] = {0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01,
+                                       0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03
+   };
+
+    static const uint8_t k_mask2[32] = {0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
+                                        0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
+    };
+
+    const __m256i mask1 = _mm256_loadu_si256((const __m256i*)k_mask1);
+    const __m256i mask2 = _mm256_loadu_si256((const __m256i*)k_mask2);
+
+    const __m256i idx_shift = _mm256_set_epi32(1, 2, 3, 4, 5, 6, 7, 8);
+    const __m256i idx_mask  = _mm256_set1_epi32(256);
+
+    typedef union {
+        __m256i  vec[2];
+        uint32_t index[16];
+    } index_t;
+
+    index_t idx;
+
+    __m256 accumf = _mm256_setzero_ps();
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
+        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) {
+            const __m256i q8_1 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32;
+            const __m256i q8_2 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32;
+            const __m256i idx_l = _mm256_cvtepu8_epi16(_mm_loadu_si128((const __m128i *)qs)); qs += 16;
+            idx.vec[0] = _mm256_set1_epi32(qh[ib32+0]);
+            idx.vec[1] = _mm256_set1_epi32(qh[ib32+1]);
+            idx.vec[0] = _mm256_and_si256(_mm256_sllv_epi32(idx.vec[0], idx_shift), idx_mask);
+            idx.vec[1] = _mm256_and_si256(_mm256_sllv_epi32(idx.vec[1], idx_shift), idx_mask);
+            idx.vec[0] = _mm256_or_si256(idx.vec[0], _mm256_cvtepi16_epi32(_mm256_castsi256_si128(idx_l)));
+            idx.vec[1] = _mm256_or_si256(idx.vec[1], _mm256_cvtepi16_epi32(_mm256_extractf128_si256(idx_l, 1)));
+
+            // At leat on my CPU (Ryzen 7950X), using _mm256_i32gather_epi32 is slower than _mm256_set_epi32. Strange.
+            //const __m256i q2_1 = _mm256_i32gather_epi32((const int *)iq3s_grid, idx.vec[0], 4);
+            //const __m256i q2_2 = _mm256_i32gather_epi32((const int *)iq3s_grid, idx.vec[1], 4);
+            const __m256i q2_1 = _mm256_set_epi32(
+                    iq3s_grid[idx.index[7]], iq3s_grid[idx.index[6]], iq3s_grid[idx.index[5]], iq3s_grid[idx.index[4]],
+                    iq3s_grid[idx.index[3]], iq3s_grid[idx.index[2]], iq3s_grid[idx.index[1]], iq3s_grid[idx.index[0]]
+            );
+            const __m256i q2_2 = _mm256_set_epi32(
+                    iq3s_grid[idx.index[15]], iq3s_grid[idx.index[14]], iq3s_grid[idx.index[13]], iq3s_grid[idx.index[12]],
+                    iq3s_grid[idx.index[11]], iq3s_grid[idx.index[10]], iq3s_grid[idx.index[ 9]], iq3s_grid[idx.index[ 8]]
+            );
+
+            __m256i aux256 = _mm256_set1_epi32(signs[0] | (signs[1] << 16));
+            aux256 = _mm256_and_si256(_mm256_shuffle_epi8(aux256,mask1), mask2);
+            const __m256i s2_1 = _mm256_cmpeq_epi8(aux256, mask2);
+            const __m256i q8s_1 = _mm256_sub_epi8(_mm256_xor_si256(s2_1, q8_1), s2_1);
+
+            aux256 = _mm256_set1_epi32(signs[2] | (signs[3] << 16));
+            aux256 = _mm256_and_si256(_mm256_shuffle_epi8(aux256,mask1), mask2);
+            const __m256i s2_2 = _mm256_cmpeq_epi8(aux256, mask2);
+            const __m256i q8s_2 = _mm256_sub_epi8(_mm256_xor_si256(s2_2, q8_2), s2_2);
+
+            signs += 4;
+
+            const __m256i dot1  = _mm256_maddubs_epi16(q2_1, q8s_1);
+            const __m256i dot2  = _mm256_maddubs_epi16(q2_2, q8s_2);
+            const uint16_t ls1 = x[i].scales[ib32/2] & 0xf;
+            const uint16_t ls2 = x[i].scales[ib32/2] >>  4;
+            const __m256i p1 = _mm256_madd_epi16(dot1, _mm256_set1_epi16(2*ls1+1));
+            const __m256i p2 = _mm256_madd_epi16(dot2, _mm256_set1_epi16(2*ls2+1));
+            sumi1 = _mm256_add_epi32(sumi1, p1);
+            sumi2 = _mm256_add_epi32(sumi2, p2);
+        }
+
+        accumf = _mm256_fmadd_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(_mm256_add_epi32(sumi1, sumi2)), accumf);
+
+    }
+
+    *s = hsum_float_8(accumf);
+
+#elif defined(__AVX__)
+   static const uint8_t k_mask1[32] = {0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01,
+                                       0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03
+   };
+
+    static const uint8_t k_mask2[32] = {0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
+                                        0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
+    };
+
+    const __m128i mask1_0 = _mm_loadu_si128((const __m128i*)k_mask1);
+    const __m128i mask1_1 = _mm_loadu_si128((const __m128i*)k_mask1 + 1);
+    const __m128i mask2_0 = _mm_loadu_si128((const __m128i*)k_mask2);
+    const __m128i mask2_1 = _mm_loadu_si128((const __m128i*)k_mask2 + 1);
+
+    const __m128i idx_mul_0 = _mm_set_epi32(32, 64, 128, 256);
+    const __m128i idx_mul_1 = _mm_set_epi32(2, 4, 8, 16);
+    const __m128i idx_mask  = _mm_set1_epi32(256);
+
+    typedef union {
+        __m128i  vec[4];
+        uint32_t index[16];
+    } index_t;
+
+    index_t idx;
+
+    __m256 accumf = _mm256_setzero_ps();
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
+        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();
+        __m128i sumi2_1 = _mm_setzero_si128();
+        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
+            const __m128i q8_1_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8_1_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8_2_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8_2_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i qs_tmp = _mm_loadu_si128((const __m128i *)qs);
+            const __m128i idx_l_0 = _mm_cvtepu8_epi16(qs_tmp);
+            const __m128i idx_l_1 = _mm_cvtepu8_epi16(_mm_srli_si128(qs_tmp, 8)); qs += 16;
+            idx.vec[0] = _mm_set1_epi32(qh[ib32+0]);
+            idx.vec[1] = idx.vec[0];
+            idx.vec[2] = _mm_set1_epi32(qh[ib32+1]);
+            idx.vec[3] = idx.vec[2];
+
+            idx.vec[0] = _mm_and_si128(_mm_mullo_epi32(idx.vec[0], idx_mul_0), idx_mask);
+            idx.vec[1] = _mm_and_si128(_mm_mullo_epi32(idx.vec[1], idx_mul_1), idx_mask);
+            idx.vec[2] = _mm_and_si128(_mm_mullo_epi32(idx.vec[2], idx_mul_0), idx_mask);
+            idx.vec[3] = _mm_and_si128(_mm_mullo_epi32(idx.vec[3], idx_mul_1), idx_mask);
+
+            idx.vec[0] = _mm_or_si128(idx.vec[0], _mm_cvtepi16_epi32(idx_l_0));
+            idx.vec[1] = _mm_or_si128(idx.vec[1], _mm_cvtepi16_epi32(_mm_srli_si128(idx_l_0, 8)));
+            idx.vec[2] = _mm_or_si128(idx.vec[2], _mm_cvtepi16_epi32(idx_l_1));
+            idx.vec[3] = _mm_or_si128(idx.vec[3], _mm_cvtepi16_epi32(_mm_srli_si128(idx_l_1, 8)));
+
+            const __m128i q2_1_0 = _mm_set_epi32(iq3s_grid[idx.index[3]], iq3s_grid[idx.index[2]], iq3s_grid[idx.index[1]], iq3s_grid[idx.index[0]]);
+            const __m128i q2_1_1 = _mm_set_epi32(iq3s_grid[idx.index[7]], iq3s_grid[idx.index[6]], iq3s_grid[idx.index[5]], iq3s_grid[idx.index[4]]);
+            const __m128i q2_2_0 = _mm_set_epi32(iq3s_grid[idx.index[11]], iq3s_grid[idx.index[10]], iq3s_grid[idx.index[9]], iq3s_grid[idx.index[8]]);
+            const __m128i q2_2_1 = _mm_set_epi32(iq3s_grid[idx.index[15]], iq3s_grid[idx.index[14]], iq3s_grid[idx.index[13]], iq3s_grid[idx.index[12]]);
+
+            __m128i aux128_0 = _mm_set1_epi32(signs[0] | (signs[1] << 16));
+            __m128i aux128_1 = aux128_0;
+            aux128_0 = _mm_and_si128(_mm_shuffle_epi8(aux128_0,mask1_0), mask2_0);
+            aux128_1 = _mm_and_si128(_mm_shuffle_epi8(aux128_1,mask1_1), mask2_1);
+            const __m128i s2_1_0 = _mm_cmpeq_epi8(aux128_0, mask2_0);
+            const __m128i s2_1_1 = _mm_cmpeq_epi8(aux128_1, mask2_1);
+            const __m128i q8s_1_0 = _mm_sub_epi8(_mm_xor_si128(s2_1_0, q8_1_0), s2_1_0);
+            const __m128i q8s_1_1 = _mm_sub_epi8(_mm_xor_si128(s2_1_1, q8_1_1), s2_1_1);
+
+            aux128_0 = _mm_set1_epi32(signs[2] | (signs[3] << 16));
+            aux128_1 = aux128_0;
+            aux128_0 = _mm_and_si128(_mm_shuffle_epi8(aux128_0,mask1_0), mask2_0);
+            aux128_1 = _mm_and_si128(_mm_shuffle_epi8(aux128_1,mask1_1), mask2_1);
+            const __m128i s2_2_0 = _mm_cmpeq_epi8(aux128_0, mask2_0);
+            const __m128i s2_2_1 = _mm_cmpeq_epi8(aux128_1, mask2_1);
+            const __m128i q8s_2_0 = _mm_sub_epi8(_mm_xor_si128(s2_2_0, q8_2_0), s2_2_0);
+            const __m128i q8s_2_1 = _mm_sub_epi8(_mm_xor_si128(s2_2_1, q8_2_1), s2_2_1);
+
+            signs += 4;
+
+            const __m128i dot1_0  = _mm_maddubs_epi16(q2_1_0, q8s_1_0);
+            const __m128i dot1_1  = _mm_maddubs_epi16(q2_1_1, q8s_1_1);
+            const __m128i dot2_0  = _mm_maddubs_epi16(q2_2_0, q8s_2_0);
+            const __m128i dot2_1  = _mm_maddubs_epi16(q2_2_1, q8s_2_1);
+            const uint16_t ls1 = x[i].scales[ib32/2] & 0xf;
+            const uint16_t ls2 = x[i].scales[ib32/2] >>  4;
+            const __m128i p1_0 = _mm_madd_epi16(dot1_0, _mm_set1_epi16(2*ls1+1));
+            const __m128i p1_1 = _mm_madd_epi16(dot1_1, _mm_set1_epi16(2*ls1+1));
+            const __m128i p2_0 = _mm_madd_epi16(dot2_0, _mm_set1_epi16(2*ls2+1));
+            const __m128i p2_1 = _mm_madd_epi16(dot2_1, _mm_set1_epi16(2*ls2+1));
+            sumi1_0 = _mm_add_epi32(sumi1_0, p1_0);
+            sumi1_1 = _mm_add_epi32(sumi1_1, p1_1);
+            sumi2_0 = _mm_add_epi32(sumi2_0, p2_0);
+            sumi2_1 = _mm_add_epi32(sumi2_1, p2_1);
+        }
+
+        accumf = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(MM256_SET_M128I(_mm_add_epi32(sumi1_1, sumi2_1), _mm_add_epi32(sumi1_0, sumi2_0)))), accumf);
+
+    }
+
+    *s = hsum_float_8(accumf);
+
+#else
+
+    float sumf = 0.f;
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
+        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;
+            const uint32_t ls2 = 2*(x[i].scales[ib32/2] >>  4) + 1;
+            int32_t sumi = 0;
+            for (int l = 0; l < 4; ++l) {
+                const uint8_t * grid1 = (const uint8_t *)(iq3s_grid + (qs[2*l+0] | ((qh[ib32+0] << (8-2*l)) & 256)));
+                const uint8_t * grid2 = (const uint8_t *)(iq3s_grid + (qs[2*l+1] | ((qh[ib32+0] << (7-2*l)) & 256)));
+                for (int j = 0; j < 4; ++j) {
+                    sumi += grid1[j] * q8[j+0] * (signs[l] & kmask_iq2xs[j+0] ? -1 : 1);
+                    sumi += grid2[j] * q8[j+4] * (signs[l] & kmask_iq2xs[j+4] ? -1 : 1);
+                }
+                q8 += 8;
+            }
+            qs += 8;
+            signs += 4;
+            bsum += sumi * ls1;
+            sumi = 0;
+            for (int l = 0; l < 4; ++l) {
+                const uint8_t * grid1 = (const uint8_t *)(iq3s_grid + (qs[2*l+0] | ((qh[ib32+1] << (8-2*l)) & 256)));
+                const uint8_t * grid2 = (const uint8_t *)(iq3s_grid + (qs[2*l+1] | ((qh[ib32+1] << (7-2*l)) & 256)));
+                for (int j = 0; j < 4; ++j) {
+                    sumi += grid1[j] * q8[j+0] * (signs[l] & kmask_iq2xs[j+0] ? -1 : 1);
+                    sumi += grid2[j] * q8[j+4] * (signs[l] & kmask_iq2xs[j+4] ? -1 : 1);
+                }
+                q8 += 8;
+            }
+            qs += 8;
+            signs += 4;
+            bsum += sumi * ls2;
+        }
+        sumf += d * bsum;
+    }
+    *s = sumf;
+#endif
+}
+
+#if defined(__AVX2__)
+static inline __m256i mul_add_epi8(const __m256i x, const __m256i y) {
+    const __m256i ax = _mm256_sign_epi8(x, x);
+    const __m256i sy = _mm256_sign_epi8(y, x);
+    return _mm256_maddubs_epi16(ax, sy);
+}
+#endif
+
+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);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_iq1_s * GGML_RESTRICT x = vx;
+    const block_q8_K  * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+#if defined __AVX2__
+
+    __m256 accum = _mm256_setzero_ps();
+    float accum1 = 0;
+    for (int i = 0; i < nb; ++i) {
+
+        const int8_t   * q8 = y[i].qs;
+        const uint8_t  * qs = x[i].qs;
+        const uint16_t * qh = x[i].qh;
+
+        __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;
+
+            const __m256i dot1 = mul_add_epi8(q1b_1, q8b_1);
+            const __m256i dot2 = mul_add_epi8(q1b_2, q8b_2);
+            const int16_t ls1 = 2*((qh[ib+0] >> 12) & 7) + 1;
+            const int16_t ls2 = 2*((qh[ib+1] >> 12) & 7) + 1;
+            const __m256i p1 = _mm256_madd_epi16(dot1, _mm256_set1_epi16(ls1));
+            const __m256i p2 = _mm256_madd_epi16(dot2, _mm256_set1_epi16(ls2));
+
+            sumi = _mm256_add_epi32(sumi, _mm256_add_epi32(p1, p2));
+            sumi1 += (y[i].bsums[2*ib+0] + y[i].bsums[2*ib+1]) * (qh[ib+0] & 0x8000 ? -1 : 1) * ls1
+                   + (y[i].bsums[2*ib+2] + y[i].bsums[2*ib+3]) * (qh[ib+1] & 0x8000 ? -1 : 1) * ls2;
+        }
+
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+        accum = _mm256_fmadd_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(sumi), accum);
+        accum1 += d * sumi1;
+
+    }
+
+    *s = hsum_float_8(accum) + IQ1S_DELTA * accum1;
+
+#elif defined __AVX__
+    __m256 accum = _mm256_setzero_ps();
+    float accum1 = 0;
+    for (int i = 0; i < nb; ++i) {
+
+        const int8_t   * q8 = y[i].qs;
+        const uint8_t  * qs = x[i].qs;
+        const uint16_t * qh = x[i].qh;
+
+        __m128i sumi1_0 = _mm_setzero_si128();
+        __m128i sumi1_1 = _mm_setzero_si128();
+        int sumi1 = 0;
+        for (int ib = 0; ib < QK_K/32; ib += 2) {
+            const __m128i q1b_1_0 = _mm_set_epi64x(iq1s_grid[qs[1] | ((qh[ib+0] << 5) & 0x700)], iq1s_grid[qs[0] | ((qh[ib+0] << 8) & 0x700)]);
+            const __m128i q1b_1_1 = _mm_set_epi64x(iq1s_grid[qs[3] | ((qh[ib+0] >> 1) & 0x700)], iq1s_grid[qs[2] | ((qh[ib+0] << 2) & 0x700)]);
+            const __m128i q1b_2_0 = _mm_set_epi64x(iq1s_grid[qs[5] | ((qh[ib+1] << 5) & 0x700)], iq1s_grid[qs[4] | ((qh[ib+1] << 8) & 0x700)]);
+            const __m128i q1b_2_1 = _mm_set_epi64x(iq1s_grid[qs[7] | ((qh[ib+1] >> 1) & 0x700)], iq1s_grid[qs[6] | ((qh[ib+1] << 2) & 0x700)]);
+            qs += 8;
+            const __m128i q8b_1_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8b_1_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8b_2_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8b_2_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+
+            const __m128i dot1_0 = mul_add_epi8_sse(q1b_1_0, q8b_1_0);
+            const __m128i dot1_1 = mul_add_epi8_sse(q1b_1_1, q8b_1_1);
+            const __m128i dot2_0 = mul_add_epi8_sse(q1b_2_0, q8b_2_0);
+            const __m128i dot2_1 = mul_add_epi8_sse(q1b_2_1, q8b_2_1);
+            const int16_t ls1 = 2*((qh[ib+0] >> 12) & 7) + 1;
+            const int16_t ls2 = 2*((qh[ib+1] >> 12) & 7) + 1;
+            const __m128i p1_0 = _mm_madd_epi16(dot1_0, _mm_set1_epi16(ls1));
+            const __m128i p1_1 = _mm_madd_epi16(dot1_1, _mm_set1_epi16(ls1));
+            const __m128i p2_0 = _mm_madd_epi16(dot2_0, _mm_set1_epi16(ls2));
+            const __m128i p2_1 = _mm_madd_epi16(dot2_1, _mm_set1_epi16(ls2));
+
+            sumi1_0 = _mm_add_epi32(sumi1_0, _mm_add_epi32(p1_0, p2_0));
+            sumi1_1 = _mm_add_epi32(sumi1_1, _mm_add_epi32(p1_1, p2_1));
+            sumi1 += (y[i].bsums[2*ib+0] + y[i].bsums[2*ib+1]) * (qh[ib+0] & 0x8000 ? -1 : 1) * ls1
+                   + (y[i].bsums[2*ib+2] + y[i].bsums[2*ib+3]) * (qh[ib+1] & 0x8000 ? -1 : 1) * ls2;
+        }
+
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+        accum = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(MM256_SET_M128I(sumi1_1, sumi1_0))), accum);
+        accum1 += d * sumi1;
+
+    }
+
+    *s = hsum_float_8(accum) + IQ1S_DELTA * accum1;
+
+#else
+
+    float sumf = 0;
+    for (int i = 0; i < nb; i++) {
+
+        const int8_t   * q8 = y[i].qs;
+        const uint8_t  * qs = x[i].qs;
+        const uint16_t * qh = x[i].qh;
+
+        int sumi = 0, sumi1 = 0;
+        for (int ib = 0; ib < QK_K/32; ++ib) {
+            const int ls = 2*((qh[ib] >> 12) & 7) + 1;
+            const int delta = qh[ib] & 0x8000 ? -1 : 1;
+            int lsum = 0;
+            for (int l = 0; l < 4; ++l) {
+                const int8_t * grid = (const int8_t *)(iq1s_grid + (qs[l] | (((qh[ib] >> 3*l) & 7) << 8)));
+                for (int j = 0; j < 8; ++j) {
+                    lsum += q8[j] * grid[j];
+                }
+                q8 += 8;
+            }
+            sumi  += ls * lsum;
+            sumi1 += ls * delta * (y[i].bsums[2*ib+0] + y[i].bsums[2*ib+1]);
+            qs += 4;
+        }
+
+        sumf += GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d * (sumi + IQ1S_DELTA * sumi1);
+    }
+
+    *s = sumf;
+
+#endif
+}
+
+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);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_iq1_m * GGML_RESTRICT x = vx;
+    const block_q8_K  * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+    iq1m_scale_t scale;
+
+#if defined __AVX2__
+
+    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();
+    for (int i = 0; i < nb; ++i) {
+
+        const int8_t   * q8 = y[i].qs;
+        const uint8_t  * qs = x[i].qs;
+        const uint8_t  * qh = x[i].qh;
+        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)]
+            );
+            const __m256i q1b_2 = _mm256_set_epi64x(
+                    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 delta1 = _mm256_set_epi64x(qh[1] & 0x80 ? 0xffffffffffffffff : 0x0101010101010101,
+                                                     qh[1] & 0x08 ? 0xffffffffffffffff : 0x0101010101010101,
+                                                     qh[0] & 0x80 ? 0xffffffffffffffff : 0x0101010101010101,
+                                                     qh[0] & 0x08 ? 0xffffffffffffffff : 0x0101010101010101);
+            const __m256i delta2 = _mm256_set_epi64x(qh[3] & 0x80 ? 0xffffffffffffffff : 0x0101010101010101,
+                                                     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 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_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);
+
+            const __m256i p1 = _mm256_madd_epi16(dot1, scale1);
+            const __m256i p2 = _mm256_madd_epi16(dot2, scale2);
+            const __m256i p3 = _mm256_madd_epi16(dot3, scale1);
+            const __m256i p4 = _mm256_madd_epi16(dot4, scale2);
+
+            sumi1 = _mm256_add_epi32(sumi1, _mm256_add_epi32(p1, p2));
+            sumi2 = _mm256_add_epi32(sumi2, _mm256_add_epi32(p3, p4));
+
+            qs += 8; qh += 4;
+        }
+
+        const __m256 d = _mm256_set1_ps(y[i].d * GGML_CPU_FP16_TO_FP32(scale.f16));
+
+        accum1 = _mm256_fmadd_ps(d, _mm256_cvtepi32_ps(sumi1), accum1);
+        accum2 = _mm256_fmadd_ps(d, _mm256_cvtepi32_ps(sumi2), accum2);
+    }
+
+    *s = hsum_float_8(accum1) + IQ1M_DELTA * hsum_float_8(accum2);
+
+#elif defined __AVX__
+    const __m128i mask = _mm_set1_epi16(0x7);
+    const __m128i mone = _mm_set1_epi16(1);
+
+    __m256 accum1 = _mm256_setzero_ps();
+    __m256 accum2 = _mm256_setzero_ps();
+    for (int i = 0; i < nb; ++i) {
+
+        const int8_t   * q8 = y[i].qs;
+        const uint8_t  * qs = x[i].qs;
+        const uint8_t  * qh = x[i].qh;
+        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);
+
+        __m128i sumi1_0 = _mm_setzero_si128();
+        __m128i sumi1_1 = _mm_setzero_si128();
+        __m128i sumi2_0 = _mm_setzero_si128();
+        __m128i sumi2_1 = _mm_setzero_si128();
+        for (int ib = 0; ib < QK_K/32; ib += 2) {
+            const __m128i q1b_1_0 = _mm_set_epi64x(
+                    iq1s_grid[qs[1] | (((uint16_t)qh[0] << 4) & 0x700)], iq1s_grid[qs[0] | (((uint16_t)qh[0] << 8) & 0x700)]);
+            const __m128i q1b_1_1 = _mm_set_epi64x(
+                    iq1s_grid[qs[3] | (((uint16_t)qh[1] << 4) & 0x700)], iq1s_grid[qs[2] | (((uint16_t)qh[1] << 8) & 0x700)]);
+            const __m128i q1b_2_0 = _mm_set_epi64x(
+                    iq1s_grid[qs[5] | (((uint16_t)qh[2] << 4) & 0x700)], iq1s_grid[qs[4] | (((uint16_t)qh[2] << 8) & 0x700)]);
+            const __m128i q1b_2_1 = _mm_set_epi64x(
+                    iq1s_grid[qs[7] | (((uint16_t)qh[3] << 4) & 0x700)], iq1s_grid[qs[6] | (((uint16_t)qh[3] << 8) & 0x700)]);
+            const __m128i q8b_1_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8b_1_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8b_2_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8b_2_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+
+            const __m128i dot1_0 = mul_add_epi8_sse(q1b_1_0, q8b_1_0);
+            const __m128i dot1_1 = mul_add_epi8_sse(q1b_1_1, q8b_1_1);
+            const __m128i dot2_0 = mul_add_epi8_sse(q1b_2_0, q8b_2_0);
+            const __m128i dot2_1 = mul_add_epi8_sse(q1b_2_1, q8b_2_1);
+
+            const __m128i delta1_0 = _mm_set_epi64x(qh[0] & 0x80 ? 0xffffffffffffffff : 0x0101010101010101,
+                                                     qh[0] & 0x08 ? 0xffffffffffffffff : 0x0101010101010101);
+            const __m128i delta1_1 = _mm_set_epi64x(qh[1] & 0x80 ? 0xffffffffffffffff : 0x0101010101010101,
+                                                     qh[1] & 0x08 ? 0xffffffffffffffff : 0x0101010101010101);
+            const __m128i delta2_0 = _mm_set_epi64x(qh[2] & 0x80 ? 0xffffffffffffffff : 0x0101010101010101,
+                                                     qh[2] & 0x08 ? 0xffffffffffffffff : 0x0101010101010101);
+            const __m128i delta2_1 = _mm_set_epi64x(qh[3] & 0x80 ? 0xffffffffffffffff : 0x0101010101010101,
+                                                     qh[3] & 0x08 ? 0xffffffffffffffff : 0x0101010101010101);
+
+            const __m128i dot3_0 = mul_add_epi8_sse(delta1_0, q8b_1_0);
+            const __m128i dot3_1 = mul_add_epi8_sse(delta1_1, q8b_1_1);
+            const __m128i dot4_0 = mul_add_epi8_sse(delta2_0, q8b_2_0);
+            const __m128i dot4_1 = mul_add_epi8_sse(delta2_1, q8b_2_1);
+
+            __m128i scale1_0 = _mm_set1_epi16(sc[ib/2] >> 0);
+            __m128i scale1_1 = _mm_set1_epi16(sc[ib/2] >> 3);
+            __m128i scale2_0 = _mm_set1_epi16(sc[ib/2] >> 6);
+            __m128i scale2_1 = _mm_set1_epi16(sc[ib/2] >> 9);
+
+            scale1_0 = _mm_add_epi16(_mm_slli_epi16(_mm_and_si128(scale1_0, mask), 1), mone);
+            scale1_1 = _mm_add_epi16(_mm_slli_epi16(_mm_and_si128(scale1_1, mask), 1), mone);
+            scale2_0 = _mm_add_epi16(_mm_slli_epi16(_mm_and_si128(scale2_0, mask), 1), mone);
+            scale2_1 = _mm_add_epi16(_mm_slli_epi16(_mm_and_si128(scale2_1, mask), 1), mone);
+            const __m128i p1_0 = _mm_madd_epi16(dot1_0, scale1_0);
+            const __m128i p1_1 = _mm_madd_epi16(dot1_1, scale1_1);
+            const __m128i p2_0 = _mm_madd_epi16(dot2_0, scale2_0);
+            const __m128i p2_1 = _mm_madd_epi16(dot2_1, scale2_1);
+            const __m128i p3_0 = _mm_madd_epi16(dot3_0, scale1_0);
+            const __m128i p3_1 = _mm_madd_epi16(dot3_1, scale1_1);
+            const __m128i p4_0 = _mm_madd_epi16(dot4_0, scale2_0);
+            const __m128i p4_1 = _mm_madd_epi16(dot4_1, scale2_1);
+
+            sumi1_0 = _mm_add_epi32(sumi1_0, _mm_add_epi32(p1_0, p2_0));
+            sumi1_1 = _mm_add_epi32(sumi1_1, _mm_add_epi32(p1_1, p2_1));
+            sumi2_0 = _mm_add_epi32(sumi2_0, _mm_add_epi32(p3_0, p4_0));
+            sumi2_1 = _mm_add_epi32(sumi2_1, _mm_add_epi32(p3_1, p4_1));
+
+            qs += 8; qh += 4;
+        }
+
+        const __m256 d = _mm256_set1_ps(y[i].d * GGML_CPU_FP16_TO_FP32(scale.f16));
+
+        accum1 = _mm256_add_ps(_mm256_mul_ps(d, _mm256_cvtepi32_ps(MM256_SET_M128I(sumi1_1, sumi1_0))), accum1);
+        accum2 = _mm256_add_ps(_mm256_mul_ps(d, _mm256_cvtepi32_ps(MM256_SET_M128I(sumi2_1, sumi2_0))), accum2);
+    }
+
+    *s = hsum_float_8(accum1) + IQ1M_DELTA * hsum_float_8(accum2);
+
+#else
+
+    int sum1[2], sum2[2], delta[4];
+
+    float sumf = 0;
+    for (int i = 0; i < nb; i++) {
+
+        const int8_t   * q8 = y[i].qs;
+        const uint8_t  * qs = x[i].qs;
+        const uint8_t  * qh = x[i].qh;
+        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);
+
+        int sumi1 = 0, sumi2 = 0;
+        for (int ib = 0; ib < QK_K/32; ++ib) {
+            delta[0] = qh[0] & 0x08 ? -1 : 1;
+            delta[1] = qh[0] & 0x80 ? -1 : 1;
+            delta[2] = qh[1] & 0x08 ? -1 : 1;
+            delta[3] = qh[1] & 0x80 ? -1 : 1;
+            sum1[0] = sum1[1] = sum2[0] = sum2[1] = 0;
+            for (int l = 0; l < 4; ++l) {
+                const int8_t * grid = (const int8_t *)(iq1s_grid + (qs[l] | (((uint16_t)qh[l/2] << (8 - 4*(l%2))) & 0x700)));
+                int lsum1 = 0, lsum2 = 0;
+                for (int j = 0; j < 8; ++j) {
+                    lsum1 += q8[j] * grid[j];
+                    lsum2 += q8[j];
+                }
+                q8 += 8;
+                sum1[l/2] += lsum1;
+                sum2[l/2] += lsum2*delta[l];
+            }
+
+            const int ls1 = 2*((sc[ib/2] >> (6*(ib%2)+0)) & 0x7) + 1;
+            const int ls2 = 2*((sc[ib/2] >> (6*(ib%2)+3)) & 0x7) + 1;
+
+            sumi1 += sum1[0] * ls1 + sum1[1] * ls2;
+            sumi2 += sum2[0] * ls1 + sum2[1] * ls2;
+            qs += 4;
+            qh += 2;
+        }
+
+        sumf += GGML_CPU_FP16_TO_FP32(scale.f16) * y[i].d * (sumi1 + IQ1M_DELTA * sumi2);
+    }
+
+    *s = sumf;
+
+#endif
+}
+
+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);
+    UNUSED(by);
+    UNUSED(bs);
+    assert(n % QK4_NL == 0);
+    static_assert(QK4_NL == QK8_0, "QK4_NL and QK8_0 must be the same");
+
+    const block_iq4_nl * GGML_RESTRICT x = vx;
+    const block_q8_0   * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK4_NL;
+
+    int ib = 0;
+    float sumf = 0;
+
+#if defined __AVX2__
+
+    const __m128i values128 = _mm_loadu_si128((const __m128i*)kvalues_iq4nl);
+    const __m128i m4b  = _mm_set1_epi8(0x0f);
+    const __m256i mone = _mm256_set1_epi16(1);
+
+    __m256 accum1 = _mm256_setzero_ps();
+    __m256 accum2 = _mm256_setzero_ps();
+    for (; ib + 1 < nb; ib += 2) {
+        const __m128i q4bits_1 = _mm_loadu_si128((const __m128i*)x[ib + 0].qs);
+        const __m128i q4bits_2 = _mm_loadu_si128((const __m128i*)x[ib + 1].qs);
+        const __m256i q8b_1 = _mm256_loadu_si256((const __m256i *)y[ib + 0].qs);
+        const __m256i q8b_2 = _mm256_loadu_si256((const __m256i *)y[ib + 1].qs);
+        const __m256i q4b_1 = MM256_SET_M128I(_mm_shuffle_epi8(values128, _mm_and_si128(_mm_srli_epi16(q4bits_1, 4), m4b)),
+                                              _mm_shuffle_epi8(values128, _mm_and_si128(q4bits_1, m4b)));
+        const __m256i q4b_2 = MM256_SET_M128I(_mm_shuffle_epi8(values128, _mm_and_si128(_mm_srli_epi16(q4bits_2, 4), m4b)),
+                                              _mm_shuffle_epi8(values128, _mm_and_si128(q4bits_2, m4b)));
+        const __m256i p16_1 = mul_add_epi8(q4b_1, q8b_1);
+        const __m256i p16_2 = mul_add_epi8(q4b_2, q8b_2);
+        const __m256i p_1 = _mm256_madd_epi16(p16_1, mone);
+        const __m256i p_2 = _mm256_madd_epi16(p16_2, mone);
+        accum1 = _mm256_fmadd_ps(_mm256_set1_ps(GGML_CPU_FP16_TO_FP32(y[ib + 0].d)*GGML_CPU_FP16_TO_FP32(x[ib + 0].d)),
+                _mm256_cvtepi32_ps(p_1), accum1);
+        accum2 = _mm256_fmadd_ps(_mm256_set1_ps(GGML_CPU_FP16_TO_FP32(y[ib + 1].d)*GGML_CPU_FP16_TO_FP32(x[ib + 1].d)),
+                _mm256_cvtepi32_ps(p_2), accum2);
+    }
+
+    sumf = hsum_float_8(_mm256_add_ps(accum1, accum2));
+
+#elif defined __AVX__
+    const __m128i values128 = _mm_loadu_si128((const __m128i*)kvalues_iq4nl);
+    const __m128i m4b  = _mm_set1_epi8(0x0f);
+
+    __m256 accum = _mm256_setzero_ps();
+    for (; ib + 1 < nb; ib += 2) {
+        const __m128i q4bits_1 = _mm_loadu_si128((const __m128i *)x[ib + 0].qs);
+        const __m128i q4bits_2 = _mm_loadu_si128((const __m128i *)x[ib + 1].qs);
+        const __m128i q8b_1_0 = _mm_loadu_si128((const __m128i *)y[ib + 0].qs);
+        const __m128i q8b_1_1 = _mm_loadu_si128((const __m128i *)y[ib + 0].qs + 1);
+        const __m128i q8b_2_0 = _mm_loadu_si128((const __m128i *)y[ib + 1].qs);
+        const __m128i q8b_2_1 = _mm_loadu_si128((const __m128i *)y[ib + 1].qs + 1);
+
+        const __m128i q4b_1_0 = _mm_shuffle_epi8(values128, _mm_and_si128(q4bits_1, m4b));
+        const __m128i q4b_1_1 = _mm_shuffle_epi8(values128, _mm_and_si128(_mm_srli_epi16(q4bits_1, 4), m4b));
+        const __m128i q4b_2_0 = _mm_shuffle_epi8(values128, _mm_and_si128(q4bits_2, m4b));
+        const __m128i q4b_2_1 = _mm_shuffle_epi8(values128, _mm_and_si128(_mm_srli_epi16(q4bits_2, 4), m4b));
+
+        const __m256 p = mul_sum_i8_quad_float(q4b_1_0, q4b_1_1, q4b_2_0, q4b_2_1, q8b_1_0, q8b_1_1, q8b_2_0, q8b_2_1);
+        const __m256 deltas = quad_fp16_delta_float(x[ib].d, y[ib].d, x[ib + 1].d, y[ib + 1].d);
+        accum = _mm256_add_ps(_mm256_mul_ps(deltas, p), accum);
+    }
+
+    sumf = hsum_float_8(accum);
+
+#endif
+    for (; ib < nb; ++ib) {
+        const float d = GGML_CPU_FP16_TO_FP32(y[ib].d)*GGML_CPU_FP16_TO_FP32(x[ib].d);
+        int sumi1 = 0, sumi2 = 0;
+        for (int j = 0; j < QK4_NL/2; ++j) {
+            sumi1 += y[ib].qs[j+       0] * kvalues_iq4nl[x[ib].qs[j] & 0xf];
+            sumi2 += y[ib].qs[j+QK4_NL/2] * kvalues_iq4nl[x[ib].qs[j] >>  4];
+        }
+        sumf += d * (sumi1 + sumi2);
+    }
+    *s = sumf;
+}
+
+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);
+    UNUSED(by);
+    UNUSED(bs);
+    assert(n % QK_K == 0);
+
+    const block_iq4_xs * GGML_RESTRICT x = vx;
+    const block_q8_K   * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+#if defined __AVX2__
+
+    const __m128i values128 = _mm_loadu_si128((const __m128i*)kvalues_iq4nl);
+    const __m128i m4b  = _mm_set1_epi8(0x0f);
+
+    __m256 accum = _mm256_setzero_ps();
+    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 = _mm256_setzero_si256();
+        __m256i sumi2 = _mm256_setzero_si256();
+        for (int ib = 0; ib < QK_K/32; ib += 2) {
+            const __m128i q4bits_1 = _mm_loadu_si128((const __m128i*)qs);  qs += 16;
+            const __m128i q4bits_2 = _mm_loadu_si128((const __m128i*)qs);  qs += 16;
+            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 q4b_1 = MM256_SET_M128I(_mm_shuffle_epi8(values128, _mm_and_si128(_mm_srli_epi16(q4bits_1, 4), m4b)),
+                                                  _mm_shuffle_epi8(values128, _mm_and_si128(q4bits_1, m4b)));
+            const __m256i q4b_2 = MM256_SET_M128I(_mm_shuffle_epi8(values128, _mm_and_si128(_mm_srli_epi16(q4bits_2, 4), m4b)),
+                                                  _mm_shuffle_epi8(values128, _mm_and_si128(q4bits_2, m4b)));
+            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;
+            const __m256i p_1 = _mm256_madd_epi16(p16_1, _mm256_set1_epi16(ls1));
+            const __m256i p_2 = _mm256_madd_epi16(p16_2, _mm256_set1_epi16(ls2));
+            sumi1 = _mm256_add_epi32(p_1, sumi1);
+            sumi2 = _mm256_add_epi32(p_2, sumi2);
+        }
+        accum = _mm256_fmadd_ps(_mm256_set1_ps(GGML_CPU_FP16_TO_FP32(x[ibl].d)*y[ibl].d),
+                _mm256_cvtepi32_ps(_mm256_add_epi32(sumi1, sumi2)), accum);
+    }
+
+    *s = hsum_float_8(accum);
+
+#elif defined __AVX__
+    const __m128i values128 = _mm_loadu_si128((const __m128i*)kvalues_iq4nl);
+    const __m128i m4b  = _mm_set1_epi8(0x0f);
+
+    __m256 accum = _mm256_setzero_ps();
+    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;
+        __m128i sumi1_0 = _mm_setzero_si128();
+        __m128i sumi1_1 = _mm_setzero_si128();
+        __m128i sumi2_0 = _mm_setzero_si128();
+        __m128i sumi2_1 = _mm_setzero_si128();
+        for (int ib = 0; ib < QK_K/32; ib += 2) {
+            const __m128i q4bits_1 = _mm_loadu_si128((const __m128i *)qs); qs += 16;
+            const __m128i q4bits_2 = _mm_loadu_si128((const __m128i *)qs); qs += 16;
+            const __m128i q8b_1_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8b_1_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8b_2_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8b_2_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q4b_1_0 = _mm_shuffle_epi8(values128, _mm_and_si128(q4bits_1, m4b));
+            const __m128i q4b_1_1 = _mm_shuffle_epi8(values128, _mm_and_si128(_mm_srli_epi16(q4bits_1, 4), m4b));
+            const __m128i q4b_2_0 = _mm_shuffle_epi8(values128, _mm_and_si128(q4bits_2, m4b));
+            const __m128i q4b_2_1 = _mm_shuffle_epi8(values128, _mm_and_si128(_mm_srli_epi16(q4bits_2, 4), m4b));
+            const __m128i p16_1_0 = mul_add_epi8_sse(q4b_1_0, q8b_1_0);
+            const __m128i p16_1_1 = mul_add_epi8_sse(q4b_1_1, q8b_1_1);
+            const __m128i p16_2_0 = mul_add_epi8_sse(q4b_2_0, q8b_2_0);
+            const __m128i p16_2_1 = mul_add_epi8_sse(q4b_2_1, q8b_2_1);
+            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;
+            const __m128i p_1_0 = _mm_madd_epi16(p16_1_0, _mm_set1_epi16(ls1));
+            const __m128i p_1_1 = _mm_madd_epi16(p16_1_1, _mm_set1_epi16(ls1));
+            const __m128i p_2_0 = _mm_madd_epi16(p16_2_0, _mm_set1_epi16(ls2));
+            const __m128i p_2_1 = _mm_madd_epi16(p16_2_1, _mm_set1_epi16(ls2));
+            sumi1_0 = _mm_add_epi32(p_1_0, sumi1_0);
+            sumi1_1 = _mm_add_epi32(p_1_1, sumi1_1);
+            sumi2_0 = _mm_add_epi32(p_2_0, sumi2_0);
+            sumi2_1 = _mm_add_epi32(p_2_1, sumi2_1);
+        }
+        __m128i sumi12_0 = _mm_add_epi32(sumi1_0, sumi2_0);
+        __m128i sumi12_1 = _mm_add_epi32(sumi1_1, sumi2_1);
+        accum = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(GGML_CPU_FP16_TO_FP32(x[ibl].d)*y[ibl].d),
+                _mm256_cvtepi32_ps(MM256_SET_M128I(sumi12_1, sumi12_0))), accum);
+    }
+
+    *s = hsum_float_8(accum);
+
+#else
+    float sumf = 0;
+    for (int ibl = 0; ibl < nb; ++ibl) {
+        const float d4d8 = GGML_CPU_FP16_TO_FP32(x[ibl].d) * y[ibl].d;
+        uint16_t h = x[ibl].scales_h;
+        const uint8_t * qs = x[ibl].qs;
+        const int8_t  * q8 = y[ibl].qs;
+        for (int ib = 0; ib < QK_K/32; ib += 2) {
+            const uint8_t ls1 = (x[ibl].scales_l[ib/2] & 0xf) | ((h << 4) & 0x30);
+            const uint8_t ls2 = (x[ibl].scales_l[ib/2] >>  4) | ((h << 2) & 0x30);
+            h >>= 4;
+            const float d1 = d4d8*(ls1 - 32);
+            const float d2 = d4d8*(ls2 - 32);
+            int sumi1 = 0, sumi2 = 0;
+            for (int j = 0; j < 16; ++j) {
+                sumi1 += q8[j+ 0] * kvalues_iq4nl[qs[j] & 0xf];
+                sumi2 += q8[j+16] * kvalues_iq4nl[qs[j] >>  4];
+            }
+            sumf += d1 * (sumi1 + sumi2);
+            qs += 16;
+            q8 += 32;
+            sumi1 = sumi2 = 0;
+            for (int j = 0; j < 16; ++j) {
+                sumi1 += q8[j+ 0] * kvalues_iq4nl[qs[j] & 0xf];
+                sumi2 += q8[j+16] * kvalues_iq4nl[qs[j] >>  4];
+            }
+            sumf += d2 * (sumi1 + sumi2);
+            qs += 16;
+            q8 += 32;
+        }
+    }
+    *s = sumf;
+#endif
+}
+
diff --git a/ggml/src/ggml-cpu/ggml-cpu-aarch64.cpp b/ggml/src/ggml-cpu/arch/x86/repack.cpp
similarity index 67%
rename from ggml/src/ggml-cpu/ggml-cpu-aarch64.cpp
rename to ggml/src/ggml-cpu/arch/x86/repack.cpp
index 0a3ff867c..c00c1e541 100644
--- a/ggml/src/ggml-cpu/ggml-cpu-aarch64.cpp
+++ b/ggml/src/ggml-cpu/arch/x86/repack.cpp
@@ -3,72 +3,20 @@
 #include "ggml-common.h"
 #include "ggml-backend-impl.h"
 
-#include "ggml-quants.h"
 #include "ggml-impl.h"
 #include "ggml-cpu.h"
 #include "ggml-cpu-impl.h"
-#include "ggml-cpu-traits.h"
+#include "simd-mappings.h"
+#include "traits.h"
 
 #include <cmath>
 #include <cstring>
 #include <cassert>
-#include <cfloat>
 #include <cstdlib> // for qsort
 #include <cstdio>  // for GGML_ASSERT
 
-#include "ggml-cpu-aarch64.h"
-
-// TODO: move to include file?
-template <int K> constexpr int QK_0() {
-    if constexpr (K == 4) {
-        return QK4_0;
-    }
-    if constexpr (K == 8) {
-        return QK8_0;
-    }
-    return -1;
-}
-
-template <int K, int N> struct block {
-    ggml_half d[N];                         // deltas for N qK_0 blocks
-    int8_t    qs[(QK_0<K>() * N * K) / 8];  // quants for N qK_0 blocks
-};
-
-// control size
-static_assert(sizeof(block<4, 4>) == 4 * sizeof(ggml_half) + QK8_0 * 2, "wrong block<4,4> size/padding");
-static_assert(sizeof(block<4, 8>) == 8 * sizeof(ggml_half) + QK8_0 * 4, "wrong block<4,8> size/padding");
-static_assert(sizeof(block<8, 4>) == 4 * sizeof(ggml_half) + QK8_0 * 4, "wrong block<8,4> size/padding");
-static_assert(sizeof(block<8, 8>) == 8 * sizeof(ggml_half) + QK8_0 * 8, "wrong block<8,8> size/padding");
-
-using block_q4_0x4 = block<4, 4>;
-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
-};
-
-static_assert(sizeof(block_iq4_nlx4) == 4 * sizeof(ggml_half) + QK4_NL * 2, "wrong iq4_nlx4 block size/padding");
+#define GGML_CPU_CLANG_WORKAROUND
+#include "../../repack.h"
 
 #if defined(__GNUC__)
 #pragma GCC diagnostic ignored "-Woverlength-strings"
@@ -76,27 +24,6 @@ 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
-// returns an interleaved block_q4_0x4
-// in the interleaved block_q4_0x4, place deltas for 4 block_q4_0 blocks
-// first, then interleave quants from 4 block_q4_0s in blocks of blck_size_interleave
-//
-// - in                  : an array of block_q4_0 pointers
-// - blck_size_interleave : the block_q4_0 quants bytes are interleaved in blocks of
-//                         blck_size_interleave bytes
-// - xor_mask            : the mask to convert the nibbles in block_q4_0 quants bytes
-//                         from bias offset form to pure sign form (this saves subtract
-//                         operations durin unpacking)
-//
 #if defined(__AVX__)
 #if defined(__F16C__)
 #if defined(__AVX512F__)
@@ -113,11 +40,11 @@ static inline __m512 __avx512_f32cx8x2_load(ggml_fp16_t *x, ggml_fp16_t *y) {
     float tmp[16];
 
     for (int i = 0; i < 8; i++) {
-        tmp[i] = GGML_FP16_TO_FP32(x[i]);
+        tmp[i] = GGML_CPU_FP16_TO_FP32(x[i]);
     }
 
     for (int i = 0; i < 8; i++) {
-        tmp[i + 8] = GGML_FP16_TO_FP32(y[i]);
+        tmp[i + 8] = GGML_CPU_FP16_TO_FP32(y[i]);
     }
 
     return _mm512_loadu_ps(tmp);
@@ -128,10 +55,10 @@ static inline __m512 __avx512_repeat_f32cx16_load(__m128i x) {
     _mm_storeu_si128((__m128i*)tmphalf, x);
 
     for (int i = 0; i < 4; i++) {
-        tmp[i] = GGML_FP16_TO_FP32(tmphalf[i]);
-        tmp[i + 4] = GGML_FP16_TO_FP32(tmphalf[i]);
-        tmp[i + 8] = GGML_FP16_TO_FP32(tmphalf[i]);
-        tmp[i + 12] = GGML_FP16_TO_FP32(tmphalf[i]);
+        tmp[i] = GGML_CPU_FP16_TO_FP32(tmphalf[i]);
+        tmp[i + 4] = GGML_CPU_FP16_TO_FP32(tmphalf[i]);
+        tmp[i + 8] = GGML_CPU_FP16_TO_FP32(tmphalf[i]);
+        tmp[i + 12] = GGML_CPU_FP16_TO_FP32(tmphalf[i]);
     }
 
     return _mm512_loadu_ps(tmp);
@@ -141,7 +68,7 @@ static inline __m256 __avx_f32cx8_load(ggml_fp16_t *x) {
     float tmp[8];
 
     for (int i = 0; i < 8; i++) {
-        tmp[i] = GGML_FP16_TO_FP32(x[i]);
+        tmp[i] = GGML_CPU_FP16_TO_FP32(x[i]);
     }
 
     return _mm256_loadu_ps(tmp);
@@ -150,8 +77,8 @@ static inline __m256 __avx_repeat_f32cx8_load(ggml_fp16_t *x) {
     float tmp[8];
 
     for (int i = 0; i < 4; i++) {
-        tmp[i] = GGML_FP16_TO_FP32(x[i]);
-        tmp[i + 4] = GGML_FP16_TO_FP32(x[i]);
+        tmp[i] = GGML_CPU_FP16_TO_FP32(x[i]);
+        tmp[i + 4] = GGML_CPU_FP16_TO_FP32(x[i]);
     }
 
     return _mm256_loadu_ps(tmp);
@@ -162,7 +89,7 @@ static inline __m256 __avx_rearranged_f32cx8_load(ggml_fp16_t *x, __m128i arrang
 
     _mm_storeu_si128((__m128i*)tmphalf, _mm_shuffle_epi8(_mm_loadu_si128((const __m128i *) x), arrangeMask));
     for (int i = 0; i < 8; i++) {
-        tmp[i] = GGML_FP16_TO_FP32(tmphalf[i]);
+        tmp[i] = GGML_CPU_FP16_TO_FP32(tmphalf[i]);
     }
 
     return _mm256_loadu_ps(tmp);
@@ -178,6 +105,12 @@ static inline __m256 __avx_rearranged_f32cx8_load(ggml_fp16_t *x, __m128i arrang
 #endif
 #endif
 
+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;
+}
 
 #if defined(__AVX2__) || defined(__AVX512F__)
 #if defined(__AVX512F__)
@@ -242,188 +175,14 @@ static inline __m256i mul_sum_i8_pairs_acc_int32x8(const __m256i acc, const __m2
 }
 #endif
 
-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 ggml_quantize_mat_q8_0_4x4(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) {
+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;
 
     block_q8_0x4 * GGML_RESTRICT y = (block_q8_0x4 *) vy;
 
-#if defined(__ARM_NEON)
-    float32x4_t srcv[4][8];
-    float id[4];
-
-    for (int i = 0; i < nb; i++) {
-        float32x4_t asrcv[8];
-        float32x4_t amaxv[8];
-
-        for (int row_iter = 0; row_iter < 4; row_iter++) {
-            for (int j = 0; j < 8; j++) srcv[row_iter][j] = vld1q_f32(x + row_iter * k + i * 32 + 4 * j);
-            for (int j = 0; j < 8; j++) asrcv[j] = vabsq_f32(srcv[row_iter][j]);
-
-            for (int j = 0; j < 4; j++) amaxv[2 * j] = vmaxq_f32(asrcv[2 * j], asrcv[2 * j + 1]);
-            for (int j = 0; j < 2; j++) amaxv[4 * j] = vmaxq_f32(amaxv[4 * j], amaxv[4 * j + 2]);
-            for (int j = 0; j < 1; j++) amaxv[8 * j] = vmaxq_f32(amaxv[8 * j], amaxv[8 * j + 4]);
-
-            const float amax = vmaxvq_f32(amaxv[0]);
-
-            const float d = amax / ((1 << 7) - 1);
-            id[row_iter] = d ? 1.0f / d : 0.0f;
-
-            y[i].d[row_iter] = GGML_FP32_TO_FP16(d);
-        }
-
-        for (int j = 0; j < 8; j++) {
-            float32x4_t v = vmulq_n_f32(srcv[0][j], id[0]);
-            int32x4_t vi = vcvtnq_s32_f32(v);
-            y[i].qs[16 * j + 0] = vgetq_lane_s32(vi, 0);
-            y[i].qs[16 * j + 1] = vgetq_lane_s32(vi, 1);
-            y[i].qs[16 * j + 2] = vgetq_lane_s32(vi, 2);
-            y[i].qs[16 * j + 3] = vgetq_lane_s32(vi, 3);
-
-            v = vmulq_n_f32(srcv[1][j], id[1]);
-            vi = vcvtnq_s32_f32(v);
-            y[i].qs[16 * j + 4] = vgetq_lane_s32(vi, 0);
-            y[i].qs[16 * j + 5] = vgetq_lane_s32(vi, 1);
-            y[i].qs[16 * j + 6] = vgetq_lane_s32(vi, 2);
-            y[i].qs[16 * j + 7] = vgetq_lane_s32(vi, 3);
-
-            v = vmulq_n_f32(srcv[2][j], id[2]);
-            vi = vcvtnq_s32_f32(v);
-            y[i].qs[16 * j + 8] = vgetq_lane_s32(vi, 0);
-            y[i].qs[16 * j + 9] = vgetq_lane_s32(vi, 1);
-            y[i].qs[16 * j + 10] = vgetq_lane_s32(vi, 2);
-            y[i].qs[16 * j + 11] = vgetq_lane_s32(vi, 3);
-
-            v = vmulq_n_f32(srcv[3][j], id[3]);
-            vi = vcvtnq_s32_f32(v);
-            y[i].qs[16 * j + 12] = vgetq_lane_s32(vi, 0);
-            y[i].qs[16 * j + 13] = vgetq_lane_s32(vi, 1);
-            y[i].qs[16 * j + 14] = vgetq_lane_s32(vi, 2);
-            y[i].qs[16 * j + 15] = vgetq_lane_s32(vi, 3);
-        }
-    }
-#else
-    // scalar
-    const int blck_size_interleave = 4;
-    float srcv[4][QK8_0];
-    float id[4];
-
-    for (int i = 0; i < nb; i++) {
-        for (int row_iter = 0; row_iter < 4; row_iter++) {
-            float amax = 0.0f; // absolute max
-
-            for (int j = 0; j < QK8_0; j++) {
-                srcv[row_iter][j] = x[row_iter * k + i * QK8_0 + j];
-                amax = MAX(amax, fabsf(srcv[row_iter][j]));
-            }
-
-            const float d = amax / ((1 << 7) - 1);
-            id[row_iter] = d ? 1.0f / d : 0.0f;
-
-            y[i].d[row_iter] = GGML_FP32_TO_FP16(d);
-        }
-
-        for (int j = 0; j < QK8_0 * 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);
-
-            float x0 = srcv[src_id][src_offset] * id[src_id];
-            y[i].qs[j] = roundf(x0);
-        }
-    }
-#endif
-}
-
-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;
-
-    block_q8_0x4 * GGML_RESTRICT y = (block_q8_0x4 *) vy;
-
-#if defined(__ARM_NEON)
-    float32x4_t srcv[4][8];
-    float id[4];
-
-    for (int i = 0; i < nb; i++) {
-        float32x4_t asrcv[8];
-        float32x4_t amaxv[8];
-
-        for (int row_iter = 0; row_iter < 4; row_iter++) {
-            for (int j = 0; j < 8; j++) srcv[row_iter][j] = vld1q_f32(x + row_iter * k + i * 32 + 4 * j);
-            for (int j = 0; j < 8; j++) asrcv[j] = vabsq_f32(srcv[row_iter][j]);
-
-            for (int j = 0; j < 4; j++) amaxv[2 * j] = vmaxq_f32(asrcv[2 * j], asrcv[2 * j + 1]);
-            for (int j = 0; j < 2; j++) amaxv[4 * j] = vmaxq_f32(amaxv[4 * j], amaxv[4 * j + 2]);
-            for (int j = 0; j < 1; j++) amaxv[8 * j] = vmaxq_f32(amaxv[8 * j], amaxv[8 * j + 4]);
-
-            const float amax = vmaxvq_f32(amaxv[0]);
-
-            const float d = amax / ((1 << 7) - 1);
-            id[row_iter] = d ? 1.0f / d : 0.0f;
-
-            y[i].d[row_iter] = GGML_FP32_TO_FP16(d);
-        }
-
-        for (int j = 0; j < 4; j++) {
-            float32x4_t v = vmulq_n_f32(srcv[0][2 * j], id[0]);
-            int32x4_t vi = vcvtnq_s32_f32(v);
-            y[i].qs[32 * j + 0] = vgetq_lane_s32(vi, 0);
-            y[i].qs[32 * j + 1] = vgetq_lane_s32(vi, 1);
-            y[i].qs[32 * j + 2] = vgetq_lane_s32(vi, 2);
-            y[i].qs[32 * j + 3] = vgetq_lane_s32(vi, 3);
-            v = vmulq_n_f32(srcv[0][2 * j + 1], id[0]);
-            vi = vcvtnq_s32_f32(v);
-            y[i].qs[32 * j + 4] = vgetq_lane_s32(vi, 0);
-            y[i].qs[32 * j + 5] = vgetq_lane_s32(vi, 1);
-            y[i].qs[32 * j + 6] = vgetq_lane_s32(vi, 2);
-            y[i].qs[32 * j + 7] = vgetq_lane_s32(vi, 3);
-
-            v = vmulq_n_f32(srcv[1][2 * j], id[1]);
-            vi = vcvtnq_s32_f32(v);
-            y[i].qs[32 * j + 8] = vgetq_lane_s32(vi, 0);
-            y[i].qs[32 * j + 9] = vgetq_lane_s32(vi, 1);
-            y[i].qs[32 * j + 10] = vgetq_lane_s32(vi, 2);
-            y[i].qs[32 * j + 11] = vgetq_lane_s32(vi, 3);
-            v = vmulq_n_f32(srcv[1][2 * j + 1], id[1]);
-            vi = vcvtnq_s32_f32(v);
-            y[i].qs[32 * j + 12] = vgetq_lane_s32(vi, 0);
-            y[i].qs[32 * j + 13] = vgetq_lane_s32(vi, 1);
-            y[i].qs[32 * j + 14] = vgetq_lane_s32(vi, 2);
-            y[i].qs[32 * j + 15] = vgetq_lane_s32(vi, 3);
-
-            v = vmulq_n_f32(srcv[2][2 * j], id[2]);
-            vi = vcvtnq_s32_f32(v);
-            y[i].qs[32 * j + 16] = vgetq_lane_s32(vi, 0);
-            y[i].qs[32 * j + 17] = vgetq_lane_s32(vi, 1);
-            y[i].qs[32 * j + 18] = vgetq_lane_s32(vi, 2);
-            y[i].qs[32 * j + 19] = vgetq_lane_s32(vi, 3);
-            v = vmulq_n_f32(srcv[2][2 * j + 1], id[2]);
-            vi = vcvtnq_s32_f32(v);
-            y[i].qs[32 * j + 20] = vgetq_lane_s32(vi, 0);
-            y[i].qs[32 * j + 21] = vgetq_lane_s32(vi, 1);
-            y[i].qs[32 * j + 22] = vgetq_lane_s32(vi, 2);
-            y[i].qs[32 * j + 23] = vgetq_lane_s32(vi, 3);
-
-            v = vmulq_n_f32(srcv[3][2 * j], id[3]);
-            vi = vcvtnq_s32_f32(v);
-            y[i].qs[32 * j + 24] = vgetq_lane_s32(vi, 0);
-            y[i].qs[32 * j + 25] = vgetq_lane_s32(vi, 1);
-            y[i].qs[32 * j + 26] = vgetq_lane_s32(vi, 2);
-            y[i].qs[32 * j + 27] = vgetq_lane_s32(vi, 3);
-            v = vmulq_n_f32(srcv[3][2 * j + 1], id[3]);
-            vi = vcvtnq_s32_f32(v);
-            y[i].qs[32 * j + 28] = vgetq_lane_s32(vi, 0);
-            y[i].qs[32 * j + 29] = vgetq_lane_s32(vi, 1);
-            y[i].qs[32 * j + 30] = vgetq_lane_s32(vi, 2);
-            y[i].qs[32 * j + 31] = vgetq_lane_s32(vi, 3);
-        }
-    }
-#elif defined(__AVX2__) || defined(__AVX__)
+#if defined(__AVX2__) || defined(__AVX__)
     float id[4];
     __m256 srcv[4][4];
     __m256 idvec[4];
@@ -453,7 +212,7 @@ static void ggml_quantize_mat_q8_0_4x8(const float * GGML_RESTRICT x, void * GGM
             id[row_iter] = ( maxScalar != 0.0f ) ? 127.f / maxScalar : 0.0f; //d ? 1.0f / d : 0.0f;
 
             // Store the scale for the individual block
-            y[i].d[row_iter] = GGML_FP32_TO_FP16(d);
+            y[i].d[row_iter] = GGML_CPU_FP32_TO_FP16(d);
 
             // Store the values in blocks of eight values - Aim is to use these later for block interleaving
             srcv[row_iter][0] = v0;
@@ -520,6 +279,7 @@ static void ggml_quantize_mat_q8_0_4x8(const float * GGML_RESTRICT x, void * GGM
 #endif
         }
     }
+
 #else
     // scalar
     const int blck_size_interleave = 8;
@@ -538,7 +298,7 @@ static void ggml_quantize_mat_q8_0_4x8(const float * GGML_RESTRICT x, void * GGM
             const float d = amax / ((1 << 7) - 1);
             id[row_iter] = d ? 1.0f / d : 0.0f;
 
-            y[i].d[row_iter] = GGML_FP32_TO_FP16(d);
+            y[i].d[row_iter] = GGML_CPU_FP32_TO_FP16(d);
         }
 
         for (int j = 0; j < QK8_0 * 4; j++) {
@@ -553,7 +313,7 @@ static void ggml_quantize_mat_q8_0_4x8(const float * GGML_RESTRICT x, void * GGM
 #endif
 }
 
-static void ggml_quantize_mat_q8_K_4x8(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) {
+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;
@@ -817,203 +577,7 @@ static void ggml_quantize_mat_q8_K_4x8(const float * GGML_RESTRICT x, void * GGM
 #endif
 }
 
-template <int64_t INTER_SIZE, ggml_type PARAM_TYPE>
-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);
-    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) {
-    const int qk = QK8_0;
-    const int nb = n / qk;
-    const int ncols_interleaved = 4;
-    const int blocklen = 4;
-
-    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(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD)
-    if (ggml_cpu_has_neon() && ggml_cpu_has_dotprod()) {
-        const block_q4_0x4 * b_ptr = (const block_q4_0x4 *) vx;
-
-        for (int c = 0; c < nc; c += ncols_interleaved) {
-            const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
-            float32x4_t acc = vdupq_n_f32(0);
-            for (int b = 0; b < nb; b++) {
-                int8x16_t b0 = vld1q_s8((const int8_t *) b_ptr->qs);
-                int8x16_t b1 = vld1q_s8((const int8_t *) b_ptr->qs + 16);
-                int8x16_t b2 = vld1q_s8((const int8_t *) b_ptr->qs + 32);
-                int8x16_t b3 = vld1q_s8((const int8_t *) b_ptr->qs + 48);
-                float16x4_t bd = vld1_f16((const __fp16 *) b_ptr->d);
-
-                int8x16_t a0 = vld1q_s8(a_ptr->qs);
-                int8x16_t a1 = vld1q_s8(a_ptr->qs + qk/2);
-                float16x4_t ad = vld1_dup_f16((const __fp16 *) &a_ptr->d);
-
-                int32x4_t ret = vdupq_n_s32(0);
-
-                ret = vdotq_laneq_s32(ret, b0 << 4, a0, 0);
-                ret = vdotq_laneq_s32(ret, b1 << 4, a0, 1);
-                ret = vdotq_laneq_s32(ret, b2 << 4, a0, 2);
-                ret = vdotq_laneq_s32(ret, b3 << 4, a0, 3);
-
-                ret = vdotq_laneq_s32(ret, b0 & 0xf0U, a1, 0);
-                ret = vdotq_laneq_s32(ret, b1 & 0xf0U, a1, 1);
-                ret = vdotq_laneq_s32(ret, b2 & 0xf0U, a1, 2);
-                ret = vdotq_laneq_s32(ret, b3 & 0xf0U, a1, 3);
-
-                acc = vfmaq_f32(acc, vcvtq_n_f32_s32(ret, 4),
-                                vmulq_f32(vcvt_f32_f16(ad), vcvt_f32_f16(bd)));
-                a_ptr++;
-                b_ptr++;
-            }
-            vst1q_f32(s, acc);
-            s += ncols_interleaved;
-        }
-        return;
-    }
-#endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD)
-    float sumf[4];
-    int sumi;
-
-    const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
-    for (int x = 0; x < nc / ncols_interleaved; x++) {
-        const block_q4_0x4 * b_ptr = (const block_q4_0x4 *) vx + (x * nb);
-
-        for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0;
-        for (int l = 0; l < nb; l++) {
-            for (int k = 0; k < (qk / (2 * blocklen)); k++) {
-                for (int j = 0; j < ncols_interleaved; j++) {
-                    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] << 4);
-                        const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
-                        sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4;
-                    }
-                    sumf[j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d);
-                }
-            }
-        }
-        for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j];
-    }
-}
-
-static void ggml_gemv_q4_0_4x8_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;
-    const int ncols_interleaved = 4;
-    const int blocklen = 8;
-
-    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(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD)
-    if (ggml_cpu_has_neon() && ggml_cpu_has_dotprod()) {
-        const block_q4_0x4 * b_ptr = (const block_q4_0x4 *) vx;
-
-        for (int c = 0; c < nc; c += ncols_interleaved) {
-            const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
-            float32x4_t acc = vdupq_n_f32(0);
-            for (int b = 0; b < nb; b++) {
-                int8x16_t b0 = vld1q_s8((const int8_t *) b_ptr->qs);
-                int8x16_t b1 = vld1q_s8((const int8_t *) b_ptr->qs + 16);
-                int8x16_t b2 = vld1q_s8((const int8_t *) b_ptr->qs + 32);
-                int8x16_t b3 = vld1q_s8((const int8_t *) b_ptr->qs + 48);
-                float16x4_t bd = vld1_f16((const __fp16 *) b_ptr->d);
-
-                int8x16_t a0 = (int8x16_t) vld1q_dup_s64((const int64_t *) a_ptr->qs);
-                int8x16_t a1 = (int8x16_t) vld1q_dup_s64((const int64_t *) a_ptr->qs + 1);
-                int8x16_t a2 = (int8x16_t) vld1q_dup_s64((const int64_t *) a_ptr->qs + 2);
-                int8x16_t a3 = (int8x16_t) vld1q_dup_s64((const int64_t *) a_ptr->qs + 3);
-                float16x4_t ad = vld1_dup_f16((const __fp16 *) &a_ptr->d);
-
-                int32x4_t ret0 = vdupq_n_s32(0);
-                int32x4_t ret1 = vdupq_n_s32(0);
-
-                ret0 = vdotq_s32(ret0, b0 << 4, a0);
-                ret1 = vdotq_s32(ret1, b1 << 4, a0);
-                ret0 = vdotq_s32(ret0, b2 << 4, a1);
-                ret1 = vdotq_s32(ret1, b3 << 4, a1);
-
-                ret0 = vdotq_s32(ret0, b0 & 0xf0U, a2);
-                ret1 = vdotq_s32(ret1, b1 & 0xf0U, a2);
-                ret0 = vdotq_s32(ret0, b2 & 0xf0U, a3);
-                ret1 = vdotq_s32(ret1, b3 & 0xf0U, a3);
-
-                int32x4_t ret = vpaddq_s32(ret0, ret1);
-
-                acc = vfmaq_f32(acc, vcvtq_n_f32_s32(ret, 4),
-                        vmulq_f32(vcvt_f32_f16(ad), vcvt_f32_f16(bd)));
-                a_ptr++;
-                b_ptr++;
-            }
-            vst1q_f32(s, acc);
-            s += ncols_interleaved;
-        }
-        return;
-    }
-#endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD)
-    float sumf[4];
-    int sumi;
-
-    const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
-    for (int x = 0; x < nc / ncols_interleaved; x++) {
-        const block_q4_0x4 * b_ptr = (const block_q4_0x4 *) vx + (x * nb);
-
-        for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0;
-        for (int l = 0; l < nb; l++) {
-            for (int k = 0; k < (qk / (2 * blocklen)); k++) {
-                for (int j = 0; j < ncols_interleaved; j++) {
-                    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] << 4);
-                        const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
-                        sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4;
-                    }
-                    sumf[j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d);
-                }
-            }
-        }
-        for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j];
-    }
-}
-
-static void ggml_gemv_q4_0_8x8_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) {
+void ggml_gemv_q4_0_8x8_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;
     const int ncols_interleaved = 8;
@@ -1032,75 +596,7 @@ static void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, c
     UNUSED(ncols_interleaved);
     UNUSED(blocklen);
 
-#if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__)
-#if defined(__ARM_FEATURE_SVE)
-    if (ggml_cpu_has_sve() && ggml_cpu_get_sve_cnt() == QK8_0) {
-        const void * b_ptr = vx;
-        const void * a_ptr = vy;
-        float * res_ptr = s;
-
-        __asm__ __volatile__(
-            "ptrue p0.b\n"
-            "add %x[b_ptr], %x[b_ptr], #0x10\n"
-            "1:"  // Column loop
-            "add x22, %x[a_ptr], #0x2\n"
-            "mov z31.b, #0x0\n"
-            "mov x21, %x[nb]\n"
-            "2:"  // Block loop
-            "ld1b { z30.b }, p0/Z, [%x[b_ptr]]\n"
-            "ld1b { z29.b }, p0/Z, [%x[b_ptr], #1, MUL VL]\n"
-            "mov z28.s, #0x0\n"
-            "mov z27.s, #0x0\n"
-            "ld1rd { z26.d }, p0/Z, [x22]\n"
-            "ld1b { z25.b }, p0/Z, [%x[b_ptr], #2, MUL VL]\n"
-            "sub x20, x22, #0x2\n"
-            "sub x21, x21, #0x1\n"
-            "ld1b { z24.b }, p0/Z, [%x[b_ptr], #3, MUL VL]\n"
-            "ld1rd { z23.d }, p0/Z, [x22, #8]\n"
-            "lsl z22.b, z30.b, #0x4\n"
-            "lsl z16.b, z29.b, #0x4\n"
-            "and z30.b, z30.b, #0xf0\n"
-            "and z29.b, z29.b, #0xf0\n"
-            "ld1rd { z21.d }, p0/Z, [x22, #16]\n"
-            "ld1rd { z20.d }, p0/Z, [x22, #24]\n"
-            "lsl z19.b, z25.b, #0x4\n"
-            "and z25.b, z25.b, #0xf0\n"
-            "ld1rh { z17.h }, p0/Z, [x20]\n"
-            "ld1h { z18.s }, p0/Z, [%x[b_ptr], #-1, MUL VL]\n"
-            "sdot z28.s, z22.b, z26.b\n"
-            "sdot z27.s, z16.b, z26.b\n"
-            "lsl z16.b, z24.b, #0x4\n"
-            "add x22, x22, #0x22\n"
-            "and z24.b, z24.b, #0xf0\n"
-            "add %x[b_ptr], %x[b_ptr], #0x90\n"
-            "fcvt z17.s, p0/m, z17.h\n"
-            "fcvt z18.s, p0/m, z18.h\n"
-            "sdot z28.s, z19.b, z23.b\n"
-            "sdot z27.s, z16.b, z23.b\n"
-            "fmul z18.s, z18.s, z17.s\n"
-            "sdot z28.s, z30.b, z21.b\n"
-            "sdot z27.s, z29.b, z21.b\n"
-            "sdot z28.s, z25.b, z20.b\n"
-            "sdot z27.s, z24.b, z20.b\n"
-            "uzp1 z17.s, z28.s, z27.s\n"
-            "uzp2 z16.s, z28.s, z27.s\n"
-            "add z17.s, z17.s, z16.s\n"
-            "asr z17.s, z17.s, #0x4\n"
-            "scvtf z17.s, p0/m, z17.s\n"
-            "fmla z31.s, p0/M, z17.s, z18.s\n"
-            "cbnz x21, 2b\n"
-            "sub %x[nc], %x[nc], #0x8\n"
-            "st1w { z31.s }, p0, [%x[res_ptr]]\n"
-            "add %x[res_ptr], %x[res_ptr], #0x20\n"
-            "cbnz %x[nc], 1b\n"
-            : [b_ptr] "+&r" (b_ptr), [res_ptr] "+&r" (res_ptr), [nc] "+&r" (nc)
-            : [a_ptr] "r" (a_ptr), [nb] "r" (nb)
-            : "memory", "p0", "x20", "x21", "x22", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
-        );
-        return;
-    }
-#endif // #if defined(__ARM_FEATURE_SVE)
-#elif defined(__AVX2__)
+#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);
@@ -1152,7 +648,7 @@ static void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, c
                 const __m256 col_scale_f32 = GGML_F32Cx8_REARRANGE_LOAD(b_ptr[b].d, changemask);
 
                 // Load and convert to FP32 scale from block_q8_0
-                const __m256 row_scale_f32 = _mm256_set1_ps(GGML_FP16_TO_FP32(a_ptr[b].d));
+                const __m256 row_scale_f32 = _mm256_set1_ps(GGML_CPU_FP16_TO_FP32(a_ptr[b].d));
 
                 // Load the block values in block_q8_0 in batches of 16 bytes and replicate the same across 256 bit vector
                 __m256i lhs_vec_0 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)a_ptr[b].qs));
@@ -1191,74 +687,8 @@ static void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, c
         }
     }
     return;
-#elif defined __riscv_v
-    if (__riscv_vlenb() >= QK4_0) {
-        const size_t vl = QK4_0;
 
-        const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
-        for (int x = 0; x < nc / ncols_interleaved; x++) {
-            const block_q4_0x8 * b_ptr = (const block_q4_0x8 *) vx + (x * nb);
-
-            vfloat32m1_t sumf = __riscv_vfmv_v_f_f32m1(0.0, vl / 4);
-            for (int l = 0; l < nb; l++) {
-                const int64_t a0 = *(const int64_t *)&a_ptr[l].qs[0];
-                const int64_t a1 = *(const int64_t *)&a_ptr[l].qs[8];
-                const int64_t a2 = *(const int64_t *)&a_ptr[l].qs[16];
-                const int64_t a3 = *(const int64_t *)&a_ptr[l].qs[24];
-                __asm__ __volatile__("" ::: "memory"); // prevent gcc from emitting fused vlse64, violating alignment
-                const vint8m2_t lhs_0_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(a0, vl / 4));
-                const vint8m2_t lhs_1_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(a1, vl / 4));
-                const vint8m2_t lhs_2_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(a2, vl / 4));
-                const vint8m2_t lhs_3_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(a3, vl / 4));
-
-                const vint8m4_t rhs_raw_vec = __riscv_vle8_v_i8m4((const int8_t *)b_ptr[l].qs, vl * 4);
-                const vint8m4_t rhs_vec_lo = __riscv_vsra_vx_i8m4(__riscv_vsll_vx_i8m4(rhs_raw_vec, 4, vl * 4), 4, vl * 4);
-                const vint8m4_t rhs_vec_hi = __riscv_vsra_vx_i8m4(rhs_raw_vec, 4, vl * 4);
-                const vint8m2_t rhs_vec_lo_0 = __riscv_vget_v_i8m4_i8m2(rhs_vec_lo, 0);
-                const vint8m2_t rhs_vec_lo_1 = __riscv_vget_v_i8m4_i8m2(rhs_vec_lo, 1);
-                const vint8m2_t rhs_vec_hi_0 = __riscv_vget_v_i8m4_i8m2(rhs_vec_hi, 0);
-                const vint8m2_t rhs_vec_hi_1 = __riscv_vget_v_i8m4_i8m2(rhs_vec_hi, 1);
-
-                const vint16m4_t sumi_lo_0 = __riscv_vwmul_vv_i16m4(rhs_vec_lo_0, lhs_0_8, vl * 2);
-                const vint16m4_t sumi_lo_1 = __riscv_vwmacc_vv_i16m4(sumi_lo_0, rhs_vec_lo_1, lhs_1_8, vl * 2);
-                const vint16m4_t sumi_hi_0 = __riscv_vwmacc_vv_i16m4(sumi_lo_1, rhs_vec_hi_0, lhs_2_8, vl * 2);
-                const vint16m4_t sumi_hi_m = __riscv_vwmacc_vv_i16m4(sumi_hi_0, rhs_vec_hi_1, lhs_3_8, vl * 2);
-
-                const vuint32m4_t sumi_i32 = __riscv_vreinterpret_v_i32m4_u32m4(__riscv_vreinterpret_v_i16m4_i32m4(sumi_hi_m));
-                const vuint16m2_t sumi_h2_0 = __riscv_vnsrl_wx_u16m2(sumi_i32, 0, vl);
-                const vuint16m2_t sumi_h2_1 = __riscv_vnsrl_wx_u16m2(sumi_i32, 16, vl);
-                const vuint16m2_t sumi_h2 = __riscv_vadd_vv_u16m2(sumi_h2_0, sumi_h2_1, vl);
-                const vuint32m2_t sumi_h2_i32 = __riscv_vreinterpret_v_u16m2_u32m2(sumi_h2);
-                const vuint16m1_t sumi_h4_0 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 0, vl / 2);
-                const vuint16m1_t sumi_h4_1 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 16, vl / 2);
-                const vuint16m1_t sumi_h4 = __riscv_vadd_vv_u16m1(sumi_h4_0, sumi_h4_1, vl / 2);
-                const vuint32m1_t sumi_h4_i32 = __riscv_vreinterpret_v_u16m1_u32m1(sumi_h4);
-                const vint16mf2_t sumi_h8_0 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 0, vl / 4));
-                const vint16mf2_t sumi_h8_1 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 16, vl / 4));
-                const vint32m1_t sumi_h8 = __riscv_vwadd_vv_i32m1(sumi_h8_0, sumi_h8_1, vl / 4);
-                const vfloat32m1_t facc = __riscv_vfcvt_f_x_v_f32m1(sumi_h8, vl / 4);
-
-                // vector version needs Zvfhmin extension
-                const float a_scale = GGML_FP16_TO_FP32(a_ptr[l].d);
-                const float b_scales[8] = {
-                    GGML_FP16_TO_FP32(b_ptr[l].d[0]),
-                    GGML_FP16_TO_FP32(b_ptr[l].d[1]),
-                    GGML_FP16_TO_FP32(b_ptr[l].d[2]),
-                    GGML_FP16_TO_FP32(b_ptr[l].d[3]),
-                    GGML_FP16_TO_FP32(b_ptr[l].d[4]),
-                    GGML_FP16_TO_FP32(b_ptr[l].d[5]),
-                    GGML_FP16_TO_FP32(b_ptr[l].d[6]),
-                    GGML_FP16_TO_FP32(b_ptr[l].d[7])
-                };
-                const vfloat32m1_t b_scales_vec = __riscv_vle32_v_f32m1(b_scales, vl / 4);
-                const vfloat32m1_t tmp1 = __riscv_vfmul_vf_f32m1(facc, a_scale, vl / 4);
-                sumf = __riscv_vfmacc_vv_f32m1(sumf, tmp1, b_scales_vec, vl / 4);
-            }
-            __riscv_vse32_v_f32m1(s + x * ncols_interleaved, sumf, vl / 4);
-        }
-        return;
-    }
-#endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__)
+#endif
     {
         float sumf[8];
         int sumi;
@@ -1277,7 +707,7 @@ static void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, c
                             const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
                             sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4;
                         }
-                        sumf[j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d);
+                        sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
                     }
                 }
             }
@@ -1286,7 +716,7 @@ 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) {
+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;
@@ -1543,13 +973,13 @@ static void ggml_gemv_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, c
                         sumi2 = sumi2 * scales_1[j];
                         sumi += sumi1 + sumi2;
                     }
-                    sumf[j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d;
+                    sumf[j] += sumi * GGML_CPU_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;
+                    sum_minf[j] += mins[j] * (a_ptr[l].bsums[sb * 2] + a_ptr[l].bsums[sb * 2 + 1]) * GGML_CPU_FP16_TO_FP32(b_ptr[l].dmin[j]) * a_ptr[l].d;
                 }
             }
         }
@@ -1560,1074 +990,7 @@ static void ggml_gemv_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, c
 #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;
-    const int ncols_interleaved = 4;
-    const int blocklen = 4;
-
-    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(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD)
-    if (ggml_cpu_has_neon() && ggml_cpu_has_dotprod()) {
-        const int8x16_t kvalues = vld1q_s8(kvalues_iq4nl);
-        const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
-        float * res_ptr = s;
-
-        for (int x = 0; x < nc / ncols_interleaved; x++) {
-            const block_iq4_nlx4 * b_ptr = (const block_iq4_nlx4 *) vx + (x * nb);
-
-            float32x4_t sumf = vdupq_n_f32(0);
-            for (int l = 0; l < nb; l++) {
-                uint8x16_t b_0 = vld1q_u8(b_ptr[l].qs + 0);
-                uint8x16_t b_1 = vld1q_u8(b_ptr[l].qs + 16);
-                uint8x16_t b_2 = vld1q_u8(b_ptr[l].qs + 32);
-                uint8x16_t b_3 = vld1q_u8(b_ptr[l].qs + 48);
-
-                int8x16_t b_0_hi = vqtbl1q_s8(kvalues, b_0 >> 4);
-                int8x16_t b_0_lo = vqtbl1q_s8(kvalues, b_0 & 0x0F);
-                int8x16_t b_1_hi = vqtbl1q_s8(kvalues, b_1 >> 4);
-                int8x16_t b_1_lo = vqtbl1q_s8(kvalues, b_1 & 0x0F);
-                int8x16_t b_2_hi = vqtbl1q_s8(kvalues, b_2 >> 4);
-                int8x16_t b_2_lo = vqtbl1q_s8(kvalues, b_2 & 0x0F);
-                int8x16_t b_3_hi = vqtbl1q_s8(kvalues, b_3 >> 4);
-                int8x16_t b_3_lo = vqtbl1q_s8(kvalues, b_3 & 0x0F);
-
-                int8x16_t a_0 = vld1q_s8(a_ptr[l].qs + 0);
-                int8x16_t a_1 = vld1q_s8(a_ptr[l].qs + 16);
-
-                int32x4_t sumi = vdupq_n_s32(0);
-                sumi = vdotq_laneq_s32(sumi, b_0_lo, a_0, 0);
-                sumi = vdotq_laneq_s32(sumi, b_0_hi, a_1, 0);
-                sumi = vdotq_laneq_s32(sumi, b_1_lo, a_0, 1);
-                sumi = vdotq_laneq_s32(sumi, b_1_hi, a_1, 1);
-                sumi = vdotq_laneq_s32(sumi, b_2_lo, a_0, 2);
-                sumi = vdotq_laneq_s32(sumi, b_2_hi, a_1, 2);
-                sumi = vdotq_laneq_s32(sumi, b_3_lo, a_0, 3);
-                sumi = vdotq_laneq_s32(sumi, b_3_hi, a_1, 3);
-
-                float32x4_t a_d = vcvt_f32_f16(vld1_dup_f16((const float16_t *)&a_ptr[l].d));
-                float32x4_t b_d = vcvt_f32_f16(vld1_f16((const float16_t *)b_ptr[l].d));
-                float32x4_t d = a_d * b_d;
-
-                sumf = vmlaq_f32(sumf, d, vcvtq_f32_s32(sumi));
-            }
-
-            vst1q_f32(res_ptr + x * 4, sumf);
-        }
-        return;
-    }
-#endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON)
-    {
-        float sumf[4];
-        int sumi;
-
-        const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
-        for (int x = 0; x < nc / ncols_interleaved; x++) {
-            const block_iq4_nlx4 * b_ptr = (const block_iq4_nlx4 *) vx + (x * nb);
-
-            for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0;
-            for (int l = 0; l < nb; l++) {
-                for (int k = 0; k < (qk / (2 * blocklen)); k++) {
-                    for (int j = 0; j < ncols_interleaved; j++) {
-                        sumi = 0;
-                        for (int i = 0; i < blocklen; ++i) {
-                            const int v0 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0x0F];
-                            const int v1 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4];
-                            sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2]));
-                        }
-                        sumf[j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d);
-                    }
-                }
-            }
-            for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j];
-        }
-    }
-}
-
-static void ggml_gemm_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) {
-    const int qk = QK8_0;
-    const int nb = n / qk;
-    const int ncols_interleaved = 4;
-    const int blocklen = 4;
-
-    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(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON)
-    if (ggml_cpu_has_neon() && ggml_cpu_has_dotprod()) {
-        const void * b_ptr = vx;
-        const void * a_ptr = vy;
-        float * res_ptr = s;
-        size_t res_stride = bs * sizeof(float);
-
-        __asm__ __volatile__(
-            "mov x10, %x[nr]\n"
-            "mov x9, #0x88\n"
-            "cmp x10, #0x10\n"
-            "mul x9, %x[nb], x9\n"
-            "blt 4f\n"
-            "1:"  // Row loop
-            "add x28, %x[b_ptr], #0x8\n"
-            "mov x27, %x[nc]\n"
-            "add x26, %x[res_ptr], %x[res_stride], LSL #4\n"
-            "2:"  // Column loop
-            "add x25, %x[a_ptr], #0x8\n"
-            "movi v15.16b, #0x0\n"
-            "movi v19.16b, #0x0\n"
-            "mov x24, %x[nb]\n"
-            "add x23, x25, x9\n"
-            "movi v18.16b, #0x0\n"
-            "movi v14.16b, #0x0\n"
-            "add x22, x23, x9\n"
-            "movi v11.16b, #0x0\n"
-            "movi v13.16b, #0x0\n"
-            "add x21, x22, x9\n"
-            "movi v23.16b, #0x0\n"
-            "movi v16.16b, #0x0\n"
-            "movi v25.16b, #0x0\n"
-            "movi v7.16b, #0x0\n"
-            "movi v0.16b, #0x0\n"
-            "movi v4.16b, #0x0\n"
-            "movi v5.16b, #0x0\n"
-            "movi v21.16b, #0x0\n"
-            "movi v8.16b, #0x0\n"
-            "movi v1.16b, #0x0\n"
-            "3:"  // Block loop
-            "ldr q3, [x28, #0x0]\n"
-            "ldr q31, [x25, #0x0]\n"
-            "movi v28.16b, #0x4\n"
-            "movi v10.4s, #0x0\n"
-            "ldr q22, [x28, #0x10]\n"
-            "ldr q6, [x25, #0x10]\n"
-            "movi v29.4s, #0x0\n"
-            "movi v9.4s, #0x0\n"
-            "ldr q27, [x28, #0x20]\n"
-            "ldr q30, [x28, #0x30]\n"
-            "movi v20.4s, #0x0\n"
-            "movi v24.16b, #0xf0\n"
-            "ldr d2, [x25, #-0x8]\n"
-            "ldr d26, [x23, #-0x8]\n"
-            "sshl v12.16b, v3.16b, v28.16b\n"
-            "sub x20, x28, #0x8\n"
-            "ldr d17, [x20, #0x0]\n"
-            "and v3.16b, v3.16b, v24.16b\n"
-            "subs x24, x24, #0x1\n"
-            "add x28, x28, #0x48\n"
-            ".inst 0x4f9fe18a  // sdot v10.4s, v12.16b, v31.4b[0]\n"
-            ".inst 0x4fbfe19d  // sdot v29.4s, v12.16b, v31.4b[1]\n"
-            ".inst 0x4f9fe989  // sdot v9.4s, v12.16b, v31.4b[2]\n"
-            ".inst 0x4fbfe994  // sdot v20.4s, v12.16b, v31.4b[3]\n"
-            "sshl v31.16b, v22.16b, v28.16b\n"
-            "and v22.16b, v22.16b, v24.16b\n"
-            "fcvtl v17.4s, v17.4h\n"
-            "fcvtl v2.4s, v2.4h\n"
-            "fcvtl v26.4s, v26.4h\n"
-            ".inst 0x4f86e3ea  // sdot v10.4s, v31.16b, v6.4b[0]\n"
-            ".inst 0x4fa6e3fd  // sdot v29.4s, v31.16b, v6.4b[1]\n"
-            ".inst 0x4f86ebe9  // sdot v9.4s, v31.16b, v6.4b[2]\n"
-            ".inst 0x4fa6ebf4  // sdot v20.4s, v31.16b, v6.4b[3]\n"
-            "sshl v6.16b, v27.16b, v28.16b\n"
-            "sshl v28.16b, v30.16b, v28.16b\n"
-            "and v27.16b, v27.16b, v24.16b\n"
-            "and v30.16b, v30.16b, v24.16b\n"
-            "ldr q24, [x25, #0x20]\n"
-            ".inst 0x4f98e0ca  // sdot v10.4s, v6.16b, v24.4b[0]\n"
-            ".inst 0x4fb8e0dd  // sdot v29.4s, v6.16b, v24.4b[1]\n"
-            ".inst 0x4f98e8c9  // sdot v9.4s, v6.16b, v24.4b[2]\n"
-            ".inst 0x4fb8e8d4  // sdot v20.4s, v6.16b, v24.4b[3]\n"
-            "ldr q24, [x25, #0x30]\n"
-            ".inst 0x4f98e38a  // sdot v10.4s, v28.16b, v24.4b[0]\n"
-            ".inst 0x4fb8e39d  // sdot v29.4s, v28.16b, v24.4b[1]\n"
-            ".inst 0x4f98eb89  // sdot v9.4s, v28.16b, v24.4b[2]\n"
-            ".inst 0x4fb8eb94  // sdot v20.4s, v28.16b, v24.4b[3]\n"
-            "ldr q24, [x25, #0x40]\n"
-            ".inst 0x4f98e06a  // sdot v10.4s, v3.16b, v24.4b[0]\n"
-            ".inst 0x4fb8e07d  // sdot v29.4s, v3.16b, v24.4b[1]\n"
-            ".inst 0x4f98e869  // sdot v9.4s, v3.16b, v24.4b[2]\n"
-            ".inst 0x4fb8e874  // sdot v20.4s, v3.16b, v24.4b[3]\n"
-            "ldr q24, [x25, #0x50]\n"
-            ".inst 0x4f98e2ca  // sdot v10.4s, v22.16b, v24.4b[0]\n"
-            ".inst 0x4fb8e2dd  // sdot v29.4s, v22.16b, v24.4b[1]\n"
-            ".inst 0x4f98eac9  // sdot v9.4s, v22.16b, v24.4b[2]\n"
-            ".inst 0x4fb8ead4  // sdot v20.4s, v22.16b, v24.4b[3]\n"
-            "ldr q24, [x25, #0x60]\n"
-            ".inst 0x4f98e36a  // sdot v10.4s, v27.16b, v24.4b[0]\n"
-            ".inst 0x4fb8e37d  // sdot v29.4s, v27.16b, v24.4b[1]\n"
-            ".inst 0x4f98eb69  // sdot v9.4s, v27.16b, v24.4b[2]\n"
-            ".inst 0x4fb8eb74  // sdot v20.4s, v27.16b, v24.4b[3]\n"
-            "ldr q24, [x25, #0x70]\n"
-            "add x25, x25, #0x88\n"
-            ".inst 0x4f98e3ca  // sdot v10.4s, v30.16b, v24.4b[0]\n"
-            ".inst 0x4fb8e3dd  // sdot v29.4s, v30.16b, v24.4b[1]\n"
-            ".inst 0x4f98ebc9  // sdot v9.4s, v30.16b, v24.4b[2]\n"
-            ".inst 0x4fb8ebd4  // sdot v20.4s, v30.16b, v24.4b[3]\n"
-            "fmul v24.4s, v17.4s, v2.s[0]\n"
-            "scvtf v10.4s, v10.4s, #0x4\n"
-            "scvtf v29.4s, v29.4s, #0x4\n"
-            "scvtf v9.4s, v9.4s, #0x4\n"
-            "scvtf v20.4s, v20.4s, #0x4\n"
-            "fmla v15.4s, v10.4s, v24.4s\n"
-            "ldr q24, [x23, #0x0]\n"
-            "fmul v10.4s, v17.4s, v2.s[1]\n"
-            "fmla v19.4s, v29.4s, v10.4s\n"
-            "ldr q10, [x23, #0x10]\n"
-            "fmul v29.4s, v17.4s, v2.s[2]\n"
-            "fmul v2.4s, v17.4s, v2.s[3]\n"
-            "fmla v18.4s, v9.4s, v29.4s\n"
-            "movi v9.4s, #0x0\n"
-            "movi v29.4s, #0x0\n"
-            ".inst 0x4f98e189  // sdot v9.4s, v12.16b, v24.4b[0]\n"
-            ".inst 0x4fb8e19d  // sdot v29.4s, v12.16b, v24.4b[1]\n"
-            "fmla v14.4s, v20.4s, v2.4s\n"
-            "movi v20.4s, #0x0\n"
-            "movi v2.4s, #0x0\n"
-            ".inst 0x4f98e994  // sdot v20.4s, v12.16b, v24.4b[2]\n"
-            ".inst 0x4fb8e982  // sdot v2.4s, v12.16b, v24.4b[3]\n"
-            "ldr q24, [x23, #0x20]\n"
-            ".inst 0x4f8ae3e9  // sdot v9.4s, v31.16b, v10.4b[0]\n"
-            ".inst 0x4faae3fd  // sdot v29.4s, v31.16b, v10.4b[1]\n"
-            ".inst 0x4f8aebf4  // sdot v20.4s, v31.16b, v10.4b[2]\n"
-            ".inst 0x4faaebe2  // sdot v2.4s, v31.16b, v10.4b[3]\n"
-            "ldr q10, [x23, #0x30]\n"
-            ".inst 0x4f98e0c9  // sdot v9.4s, v6.16b, v24.4b[0]\n"
-            ".inst 0x4fb8e0dd  // sdot v29.4s, v6.16b, v24.4b[1]\n"
-            ".inst 0x4f98e8d4  // sdot v20.4s, v6.16b, v24.4b[2]\n"
-            ".inst 0x4fb8e8c2  // sdot v2.4s, v6.16b, v24.4b[3]\n"
-            "ldr q24, [x23, #0x40]\n"
-            ".inst 0x4f8ae389  // sdot v9.4s, v28.16b, v10.4b[0]\n"
-            ".inst 0x4faae39d  // sdot v29.4s, v28.16b, v10.4b[1]\n"
-            ".inst 0x4f8aeb94  // sdot v20.4s, v28.16b, v10.4b[2]\n"
-            ".inst 0x4faaeb82  // sdot v2.4s, v28.16b, v10.4b[3]\n"
-            "ldr q10, [x23, #0x50]\n"
-            ".inst 0x4f98e069  // sdot v9.4s, v3.16b, v24.4b[0]\n"
-            ".inst 0x4fb8e07d  // sdot v29.4s, v3.16b, v24.4b[1]\n"
-            ".inst 0x4f98e874  // sdot v20.4s, v3.16b, v24.4b[2]\n"
-            ".inst 0x4fb8e862  // sdot v2.4s, v3.16b, v24.4b[3]\n"
-            "ldr q24, [x23, #0x60]\n"
-            ".inst 0x4f8ae2c9  // sdot v9.4s, v22.16b, v10.4b[0]\n"
-            ".inst 0x4faae2dd  // sdot v29.4s, v22.16b, v10.4b[1]\n"
-            ".inst 0x4f8aead4  // sdot v20.4s, v22.16b, v10.4b[2]\n"
-            ".inst 0x4faaeac2  // sdot v2.4s, v22.16b, v10.4b[3]\n"
-            "ldr q10, [x23, #0x70]\n"
-            "add x23, x23, #0x88\n"
-            ".inst 0x4f98e369  // sdot v9.4s, v27.16b, v24.4b[0]\n"
-            ".inst 0x4fb8e37d  // sdot v29.4s, v27.16b, v24.4b[1]\n"
-            ".inst 0x4f98eb74  // sdot v20.4s, v27.16b, v24.4b[2]\n"
-            ".inst 0x4fb8eb62  // sdot v2.4s, v27.16b, v24.4b[3]\n"
-            "ldr q24, [x22, #0x0]\n"
-            ".inst 0x4f8ae3c9  // sdot v9.4s, v30.16b, v10.4b[0]\n"
-            ".inst 0x4faae3dd  // sdot v29.4s, v30.16b, v10.4b[1]\n"
-            ".inst 0x4f8aebd4  // sdot v20.4s, v30.16b, v10.4b[2]\n"
-            ".inst 0x4faaebc2  // sdot v2.4s, v30.16b, v10.4b[3]\n"
-            "fmul v10.4s, v17.4s, v26.s[0]\n"
-            "scvtf v9.4s, v9.4s, #0x4\n"
-            "scvtf v29.4s, v29.4s, #0x4\n"
-            "scvtf v20.4s, v20.4s, #0x4\n"
-            "scvtf v2.4s, v2.4s, #0x4\n"
-            "fmla v11.4s, v9.4s, v10.4s\n"
-            "ldr q9, [x22, #0x10]\n"
-            "fmul v10.4s, v17.4s, v26.s[1]\n"
-            "fmla v13.4s, v29.4s, v10.4s\n"
-            "ldr d29, [x22, #-0x8]\n"
-            "fmul v10.4s, v17.4s, v26.s[2]\n"
-            "fmul v26.4s, v17.4s, v26.s[3]\n"
-            "fcvtl v29.4s, v29.4h\n"
-            "fmla v23.4s, v20.4s, v10.4s\n"
-            "movi v20.4s, #0x0\n"
-            "movi v10.4s, #0x0\n"
-            "fmla v16.4s, v2.4s, v26.4s\n"
-            "movi v26.4s, #0x0\n"
-            "movi v2.4s, #0x0\n"
-            ".inst 0x4f98e194  // sdot v20.4s, v12.16b, v24.4b[0]\n"
-            ".inst 0x4fb8e18a  // sdot v10.4s, v12.16b, v24.4b[1]\n"
-            ".inst 0x4f98e99a  // sdot v26.4s, v12.16b, v24.4b[2]\n"
-            ".inst 0x4fb8e982  // sdot v2.4s, v12.16b, v24.4b[3]\n"
-            "ldr q24, [x22, #0x20]\n"
-            ".inst 0x4f89e3f4  // sdot v20.4s, v31.16b, v9.4b[0]\n"
-            ".inst 0x4fa9e3ea  // sdot v10.4s, v31.16b, v9.4b[1]\n"
-            ".inst 0x4f89ebfa  // sdot v26.4s, v31.16b, v9.4b[2]\n"
-            ".inst 0x4fa9ebe2  // sdot v2.4s, v31.16b, v9.4b[3]\n"
-            "ldr q9, [x22, #0x30]\n"
-            ".inst 0x4f98e0d4  // sdot v20.4s, v6.16b, v24.4b[0]\n"
-            ".inst 0x4fb8e0ca  // sdot v10.4s, v6.16b, v24.4b[1]\n"
-            ".inst 0x4f98e8da  // sdot v26.4s, v6.16b, v24.4b[2]\n"
-            ".inst 0x4fb8e8c2  // sdot v2.4s, v6.16b, v24.4b[3]\n"
-            "ldr q24, [x22, #0x40]\n"
-            ".inst 0x4f89e394  // sdot v20.4s, v28.16b, v9.4b[0]\n"
-            ".inst 0x4fa9e38a  // sdot v10.4s, v28.16b, v9.4b[1]\n"
-            ".inst 0x4f89eb9a  // sdot v26.4s, v28.16b, v9.4b[2]\n"
-            ".inst 0x4fa9eb82  // sdot v2.4s, v28.16b, v9.4b[3]\n"
-            "ldr q9, [x22, #0x50]\n"
-            ".inst 0x4f98e074  // sdot v20.4s, v3.16b, v24.4b[0]\n"
-            ".inst 0x4fb8e06a  // sdot v10.4s, v3.16b, v24.4b[1]\n"
-            ".inst 0x4f98e87a  // sdot v26.4s, v3.16b, v24.4b[2]\n"
-            ".inst 0x4fb8e862  // sdot v2.4s, v3.16b, v24.4b[3]\n"
-            "ldr q24, [x22, #0x60]\n"
-            ".inst 0x4f89e2d4  // sdot v20.4s, v22.16b, v9.4b[0]\n"
-            ".inst 0x4fa9e2ca  // sdot v10.4s, v22.16b, v9.4b[1]\n"
-            ".inst 0x4f89eada  // sdot v26.4s, v22.16b, v9.4b[2]\n"
-            ".inst 0x4fa9eac2  // sdot v2.4s, v22.16b, v9.4b[3]\n"
-            "ldr q9, [x22, #0x70]\n"
-            "add x22, x22, #0x88\n"
-            ".inst 0x4f98e374  // sdot v20.4s, v27.16b, v24.4b[0]\n"
-            ".inst 0x4fb8e36a  // sdot v10.4s, v27.16b, v24.4b[1]\n"
-            ".inst 0x4f98eb7a  // sdot v26.4s, v27.16b, v24.4b[2]\n"
-            ".inst 0x4fb8eb62  // sdot v2.4s, v27.16b, v24.4b[3]\n"
-            "ldr q24, [x21, #0x0]\n"
-            ".inst 0x4f89e3d4  // sdot v20.4s, v30.16b, v9.4b[0]\n"
-            ".inst 0x4fa9e3ca  // sdot v10.4s, v30.16b, v9.4b[1]\n"
-            ".inst 0x4f89ebda  // sdot v26.4s, v30.16b, v9.4b[2]\n"
-            ".inst 0x4fa9ebc2  // sdot v2.4s, v30.16b, v9.4b[3]\n"
-            "fmul v9.4s, v17.4s, v29.s[0]\n"
-            "scvtf v20.4s, v20.4s, #0x4\n"
-            "scvtf v10.4s, v10.4s, #0x4\n"
-            "scvtf v26.4s, v26.4s, #0x4\n"
-            "scvtf v2.4s, v2.4s, #0x4\n"
-            "fmla v25.4s, v20.4s, v9.4s\n"
-            "ldr q9, [x21, #0x10]\n"
-            "fmul v20.4s, v17.4s, v29.s[1]\n"
-            "fmla v7.4s, v10.4s, v20.4s\n"
-            "ldr d20, [x21, #-0x8]\n"
-            "fmul v10.4s, v17.4s, v29.s[2]\n"
-            "fmul v29.4s, v17.4s, v29.s[3]\n"
-            "fcvtl v20.4s, v20.4h\n"
-            "fmla v0.4s, v26.4s, v10.4s\n"
-            "movi v26.4s, #0x0\n"
-            "movi v10.4s, #0x0\n"
-            "fmla v4.4s, v2.4s, v29.4s\n"
-            "movi v2.4s, #0x0\n"
-            "movi v29.4s, #0x0\n"
-            ".inst 0x4f98e19a  // sdot v26.4s, v12.16b, v24.4b[0]\n"
-            ".inst 0x4fb8e18a  // sdot v10.4s, v12.16b, v24.4b[1]\n"
-            ".inst 0x4f98e982  // sdot v2.4s, v12.16b, v24.4b[2]\n"
-            ".inst 0x4fb8e99d  // sdot v29.4s, v12.16b, v24.4b[3]\n"
-            "ldr q12, [x21, #0x20]\n"
-            "fmul v24.4s, v17.4s, v20.s[0]\n"
-            ".inst 0x4f89e3fa  // sdot v26.4s, v31.16b, v9.4b[0]\n"
-            ".inst 0x4fa9e3ea  // sdot v10.4s, v31.16b, v9.4b[1]\n"
-            ".inst 0x4f89ebe2  // sdot v2.4s, v31.16b, v9.4b[2]\n"
-            ".inst 0x4fa9ebfd  // sdot v29.4s, v31.16b, v9.4b[3]\n"
-            "ldr q9, [x21, #0x30]\n"
-            "fmul v31.4s, v17.4s, v20.s[1]\n"
-            ".inst 0x4f8ce0da  // sdot v26.4s, v6.16b, v12.4b[0]\n"
-            ".inst 0x4face0ca  // sdot v10.4s, v6.16b, v12.4b[1]\n"
-            ".inst 0x4f8ce8c2  // sdot v2.4s, v6.16b, v12.4b[2]\n"
-            ".inst 0x4face8dd  // sdot v29.4s, v6.16b, v12.4b[3]\n"
-            "ldr q12, [x21, #0x40]\n"
-            "fmul v6.4s, v17.4s, v20.s[2]\n"
-            "fmul v20.4s, v17.4s, v20.s[3]\n"
-            ".inst 0x4f89e39a  // sdot v26.4s, v28.16b, v9.4b[0]\n"
-            ".inst 0x4fa9e38a  // sdot v10.4s, v28.16b, v9.4b[1]\n"
-            ".inst 0x4f89eb82  // sdot v2.4s, v28.16b, v9.4b[2]\n"
-            ".inst 0x4fa9eb9d  // sdot v29.4s, v28.16b, v9.4b[3]\n"
-            "ldr q9, [x21, #0x50]\n"
-            ".inst 0x4f8ce07a  // sdot v26.4s, v3.16b, v12.4b[0]\n"
-            ".inst 0x4face06a  // sdot v10.4s, v3.16b, v12.4b[1]\n"
-            ".inst 0x4f8ce862  // sdot v2.4s, v3.16b, v12.4b[2]\n"
-            ".inst 0x4face87d  // sdot v29.4s, v3.16b, v12.4b[3]\n"
-            "ldr q12, [x21, #0x60]\n"
-            ".inst 0x4f89e2da  // sdot v26.4s, v22.16b, v9.4b[0]\n"
-            ".inst 0x4fa9e2ca  // sdot v10.4s, v22.16b, v9.4b[1]\n"
-            ".inst 0x4f89eac2  // sdot v2.4s, v22.16b, v9.4b[2]\n"
-            ".inst 0x4fa9eadd  // sdot v29.4s, v22.16b, v9.4b[3]\n"
-            "ldr q17, [x21, #0x70]\n"
-            "add x21, x21, #0x88\n"
-            ".inst 0x4f8ce37a  // sdot v26.4s, v27.16b, v12.4b[0]\n"
-            ".inst 0x4face36a  // sdot v10.4s, v27.16b, v12.4b[1]\n"
-            ".inst 0x4f8ceb62  // sdot v2.4s, v27.16b, v12.4b[2]\n"
-            ".inst 0x4faceb7d  // sdot v29.4s, v27.16b, v12.4b[3]\n"
-            ".inst 0x4f91e3da  // sdot v26.4s, v30.16b, v17.4b[0]\n"
-            ".inst 0x4fb1e3ca  // sdot v10.4s, v30.16b, v17.4b[1]\n"
-            ".inst 0x4f91ebc2  // sdot v2.4s, v30.16b, v17.4b[2]\n"
-            ".inst 0x4fb1ebdd  // sdot v29.4s, v30.16b, v17.4b[3]\n"
-            "scvtf v26.4s, v26.4s, #0x4\n"
-            "scvtf v10.4s, v10.4s, #0x4\n"
-            "fmla v5.4s, v26.4s, v24.4s\n"
-            "scvtf v2.4s, v2.4s, #0x4\n"
-            "scvtf v29.4s, v29.4s, #0x4\n"
-            "fmla v21.4s, v10.4s, v31.4s\n"
-            "fmla v8.4s, v2.4s, v6.4s\n"
-            "fmla v1.4s, v29.4s, v20.4s\n"
-            "bgt 3b\n"
-            "mov x20, %x[res_ptr]\n"
-            "subs x27, x27, #0x4\n"
-            "add %x[res_ptr], %x[res_ptr], #0x10\n"
-            "str q15, [x20, #0x0]\n"
-            "add x20, x20, %x[res_stride]\n"
-            "str q19, [x20, #0x0]\n"
-            "add x20, x20, %x[res_stride]\n"
-            "str q18, [x20, #0x0]\n"
-            "add x20, x20, %x[res_stride]\n"
-            "str q14, [x20, #0x0]\n"
-            "add x20, x20, %x[res_stride]\n"
-            "str q11, [x20, #0x0]\n"
-            "add x20, x20, %x[res_stride]\n"
-            "str q13, [x20, #0x0]\n"
-            "add x20, x20, %x[res_stride]\n"
-            "str q23, [x20, #0x0]\n"
-            "add x20, x20, %x[res_stride]\n"
-            "str q16, [x20, #0x0]\n"
-            "add x20, x20, %x[res_stride]\n"
-            "str q25, [x20, #0x0]\n"
-            "add x20, x20, %x[res_stride]\n"
-            "str q7, [x20, #0x0]\n"
-            "add x20, x20, %x[res_stride]\n"
-            "str q0, [x20, #0x0]\n"
-            "add x20, x20, %x[res_stride]\n"
-            "str q4, [x20, #0x0]\n"
-            "add x20, x20, %x[res_stride]\n"
-            "str q5, [x20, #0x0]\n"
-            "add x20, x20, %x[res_stride]\n"
-            "str q21, [x20, #0x0]\n"
-            "add x20, x20, %x[res_stride]\n"
-            "str q8, [x20, #0x0]\n"
-            "add x20, x20, %x[res_stride]\n"
-            "str q1, [x20, #0x0]\n"
-            "bne 2b\n"
-            "mov x20, #0x4\n"
-            "sub x10, x10, #0x10\n"
-            "cmp x10, #0x10\n"
-            "mov %x[res_ptr], x26\n"
-            "madd %x[a_ptr], x20, x9, %x[a_ptr]\n"
-            "bge 1b\n"
-            "4:"  // Row loop skip
-            "cbz x10, 9f\n"
-            "5:"  // Row tail: Row loop
-            "add x24, %x[b_ptr], #0x8\n"
-            "mov x23, %x[nc]\n"
-            "add x22, %x[res_ptr], %x[res_stride], LSL #2\n"
-            "6:"  // Row tail: Column loop
-            "movi v15.16b, #0x0\n"
-            "movi v19.16b, #0x0\n"
-            "add x25, %x[a_ptr], #0x8\n"
-            "mov x21, %x[nb]\n"
-            "movi v18.16b, #0x0\n"
-            "movi v14.16b, #0x0\n"
-            "7:"  // Row tail: Block loop
-            "ldr q7, [x24, #0x0]\n"
-            "ldr q5, [x25, #0x0]\n"
-            "movi v9.16b, #0x4\n"
-            "movi v4.4s, #0x0\n"
-            "ldr q3, [x24, #0x10]\n"
-            "ldr q2, [x25, #0x10]\n"
-            "movi v1.4s, #0x0\n"
-            "movi v0.4s, #0x0\n"
-            "ldr q13, [x24, #0x20]\n"
-            "ldr q31, [x25, #0x20]\n"
-            "movi v30.4s, #0x0\n"
-            "movi v29.16b, #0xf0\n"
-            "ldr q28, [x24, #0x30]\n"
-            "ldr q27, [x25, #0x30]\n"
-            "sshl v20.16b, v7.16b, v9.16b\n"
-            "sub x20, x24, #0x8\n"
-            "ldr q26, [x25, #0x40]\n"
-            "ldr q25, [x25, #0x50]\n"
-            "sshl v17.16b, v3.16b, v9.16b\n"
-            "and v7.16b, v7.16b, v29.16b\n"
-            "ldr q24, [x25, #0x60]\n"
-            "ldr q16, [x25, #0x70]\n"
-            "sshl v22.16b, v13.16b, v9.16b\n"
-            "and v3.16b, v3.16b, v29.16b\n"
-            "ldr d21, [x20, #0x0]\n"
-            "ldr d12, [x25, #-0x8]\n"
-            ".inst 0x4f85e284  // sdot v4.4s, v20.16b, v5.4b[0]\n"
-            ".inst 0x4fa5e281  // sdot v1.4s, v20.16b, v5.4b[1]\n"
-            ".inst 0x4f85ea80  // sdot v0.4s, v20.16b, v5.4b[2]\n"
-            ".inst 0x4fa5ea9e  // sdot v30.4s, v20.16b, v5.4b[3]\n"
-            "sshl v9.16b, v28.16b, v9.16b\n"
-            "subs x21, x21, #0x1\n"
-            "and v13.16b, v13.16b, v29.16b\n"
-            "and v28.16b, v28.16b, v29.16b\n"
-            "add x25, x25, #0x88\n"
-            "add x24, x24, #0x48\n"
-            "fcvtl v21.4s, v21.4h\n"
-            "fcvtl v12.4s, v12.4h\n"
-            ".inst 0x4f82e224  // sdot v4.4s, v17.16b, v2.4b[0]\n"
-            ".inst 0x4fa2e221  // sdot v1.4s, v17.16b, v2.4b[1]\n"
-            ".inst 0x4f82ea20  // sdot v0.4s, v17.16b, v2.4b[2]\n"
-            ".inst 0x4fa2ea3e  // sdot v30.4s, v17.16b, v2.4b[3]\n"
-            "fmul v11.4s, v21.4s, v12.s[0]\n"
-            "fmul v23.4s, v21.4s, v12.s[1]\n"
-            "fmul v17.4s, v21.4s, v12.s[2]\n"
-            ".inst 0x4f9fe2c4  // sdot v4.4s, v22.16b, v31.4b[0]\n"
-            "fmul v6.4s, v21.4s, v12.s[3]\n"
-            ".inst 0x4fbfe2c1  // sdot v1.4s, v22.16b, v31.4b[1]\n"
-            ".inst 0x4f9feac0  // sdot v0.4s, v22.16b, v31.4b[2]\n"
-            ".inst 0x4fbfeade  // sdot v30.4s, v22.16b, v31.4b[3]\n"
-            ".inst 0x4f9be124  // sdot v4.4s, v9.16b, v27.4b[0]\n"
-            ".inst 0x4fbbe121  // sdot v1.4s, v9.16b, v27.4b[1]\n"
-            ".inst 0x4f9be920  // sdot v0.4s, v9.16b, v27.4b[2]\n"
-            ".inst 0x4fbbe93e  // sdot v30.4s, v9.16b, v27.4b[3]\n"
-            ".inst 0x4f9ae0e4  // sdot v4.4s, v7.16b, v26.4b[0]\n"
-            ".inst 0x4fbae0e1  // sdot v1.4s, v7.16b, v26.4b[1]\n"
-            ".inst 0x4f9ae8e0  // sdot v0.4s, v7.16b, v26.4b[2]\n"
-            ".inst 0x4fbae8fe  // sdot v30.4s, v7.16b, v26.4b[3]\n"
-            ".inst 0x4f99e064  // sdot v4.4s, v3.16b, v25.4b[0]\n"
-            ".inst 0x4fb9e061  // sdot v1.4s, v3.16b, v25.4b[1]\n"
-            ".inst 0x4f99e860  // sdot v0.4s, v3.16b, v25.4b[2]\n"
-            ".inst 0x4fb9e87e  // sdot v30.4s, v3.16b, v25.4b[3]\n"
-            ".inst 0x4f98e1a4  // sdot v4.4s, v13.16b, v24.4b[0]\n"
-            ".inst 0x4fb8e1a1  // sdot v1.4s, v13.16b, v24.4b[1]\n"
-            ".inst 0x4f98e9a0  // sdot v0.4s, v13.16b, v24.4b[2]\n"
-            ".inst 0x4fb8e9be  // sdot v30.4s, v13.16b, v24.4b[3]\n"
-            ".inst 0x4f90e384  // sdot v4.4s, v28.16b, v16.4b[0]\n"
-            ".inst 0x4fb0e381  // sdot v1.4s, v28.16b, v16.4b[1]\n"
-            ".inst 0x4f90eb80  // sdot v0.4s, v28.16b, v16.4b[2]\n"
-            ".inst 0x4fb0eb9e  // sdot v30.4s, v28.16b, v16.4b[3]\n"
-            "scvtf v4.4s, v4.4s, #0x4\n"
-            "scvtf v1.4s, v1.4s, #0x4\n"
-            "scvtf v0.4s, v0.4s, #0x4\n"
-            "fmla v15.4s, v4.4s, v11.4s\n"
-            "scvtf v30.4s, v30.4s, #0x4\n"
-            "fmla v19.4s, v1.4s, v23.4s\n"
-            "fmla v18.4s, v0.4s, v17.4s\n"
-            "fmla v14.4s, v30.4s, v6.4s\n"
-            "bgt 7b\n"
-            "mov x20, %x[res_ptr]\n"
-            "cmp x10, #0x1\n"
-            "str q15, [x20, #0x0]\n"
-            "add x20, x20, %x[res_stride]\n"
-            "ble 8f\n"
-            "cmp x10, #0x2\n"
-            "str q19, [x20, #0x0]\n"
-            "add x20, x20, %x[res_stride]\n"
-            "ble 8f\n"
-            "cmp x10, #0x3\n"
-            "str q18, [x20, #0x0]\n"
-            "add x20, x20, %x[res_stride]\n"
-            "ble 8f\n"
-            "str q14, [x20, #0x0]\n"
-            "8:"  // Row tail: Accumulator store skip
-            "subs x23, x23, #0x4\n"
-            "add %x[res_ptr], %x[res_ptr], #0x10\n"
-            "bne 6b\n"
-            "subs x10, x10, #0x4\n"
-            "add %x[a_ptr], %x[a_ptr], x9\n"
-            "mov %x[res_ptr], x22\n"
-            "bgt 5b\n"
-            "9:"  // Row tail: Row loop skip
-            : [a_ptr] "+&r" (a_ptr), [res_ptr] "+&r" (res_ptr)
-            : [b_ptr] "r" (b_ptr), [nr] "r" (nr), [nb] "r" (nb), [res_stride] "r" (res_stride), [nc] "r" (nc)
-            : "cc", "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", "x9", "x10", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
-        );
-        return;
-    }
-#endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON)
-    {
-        float sumf[4][4];
-        int sumi;
-
-        for (int y = 0; y < nr / 4; y++) {
-            const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
-            for (int x = 0; x < nc / ncols_interleaved; x++) {
-                const block_q4_0x4 * b_ptr = (const block_q4_0x4 *) vx + (x * nb);
-                for (int m = 0; m < 4; m++) {
-                    for (int j = 0; j < ncols_interleaved; j++) sumf[m][j] = 0.0;
-                }
-                for (int l = 0; l < nb; l++) {
-                    for (int k = 0; k < (qk / (2 * blocklen)); k++) {
-                        for (int m = 0; m < 4; m++) {
-                            for (int j = 0; j < ncols_interleaved; j++) {
-                                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] << 4);
-                                    const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
-                                    sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
-                                            (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4;
-                                }
-                                sumf[m][j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(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];
-                }
-            }
-        }
-    }
-}
-
-static void ggml_gemm_q4_0_4x8_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;
-    const int ncols_interleaved = 4;
-    const int blocklen = 8;
-
-    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(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_MATMUL_INT8)
-    if (ggml_cpu_has_neon() && ggml_cpu_has_matmul_int8()) {
-        const void * b_ptr = vx;
-        const void * a_ptr = vy;
-        float * res_ptr = s;
-        size_t res_stride = bs * sizeof(float);
-
-        __asm__ __volatile__(
-            "mov x10, %x[nr]\n"
-            "mov x9, #0x88\n"
-            "cmp x10, #0x10\n"
-            "mul x9, %x[nb], x9\n"
-            "blt 4f\n"
-            "1:"  // Row loop
-            "add x28, %x[b_ptr], #0x8\n"
-            "mov x27, %x[nc]\n"
-            "add x26, %x[res_ptr], %x[res_stride], LSL #4\n"
-            "2:"  // Column loop
-            "add x25, %x[a_ptr], #0x8\n"
-            "movi v2.16b, #0x0\n"
-            "movi v10.16b, #0x0\n"
-            "mov x24, %x[nb]\n"
-            "add x23, x25, x9\n"
-            "movi v12.16b, #0x0\n"
-            "movi v28.16b, #0x0\n"
-            "add x22, x23, x9\n"
-            "movi v11.16b, #0x0\n"
-            "movi v13.16b, #0x0\n"
-            "add x21, x22, x9\n"
-            "movi v22.16b, #0x0\n"
-            "movi v23.16b, #0x0\n"
-            "movi v25.16b, #0x0\n"
-            "movi v5.16b, #0x0\n"
-            "movi v7.16b, #0x0\n"
-            "movi v4.16b, #0x0\n"
-            "movi v6.16b, #0x0\n"
-            "movi v30.16b, #0x0\n"
-            "movi v24.16b, #0x0\n"
-            "movi v14.16b, #0x0\n"
-            "3:"  // Block loop
-            "ldr q21, [x28, #0x0]\n"
-            "ldr q16, [x28, #0x10]\n"
-            "movi v1.16b, #0x4\n"
-            "movi v19.4s, #0x0\n"
-            "ldr q27, [x25, #0x0]\n"
-            "ldr q15, [x25, #0x10]\n"
-            "movi v26.4s, #0x0\n"
-            "movi v18.4s, #0x0\n"
-            "ldr q29, [x28, #0x20]\n"
-            "ldr q3, [x28, #0x30]\n"
-            "movi v17.4s, #0x0\n"
-            "movi v0.16b, #0xf0\n"
-            "ldr d20, [x25, #-0x8]\n"
-            "ldr d9, [x23, #-0x8]\n"
-            "sshl v8.16b, v21.16b, v1.16b\n"
-            "sshl v31.16b, v16.16b, v1.16b\n"
-            "and v21.16b, v21.16b, v0.16b\n"
-            "and v16.16b, v16.16b, v0.16b\n"
-            "sub x20, x28, #0x8\n"
-            "subs x24, x24, #0x1\n"
-            "add x28, x28, #0x48\n"
-            ".inst 0x4e88a773  // smmla v19.4s, v27.16b, v8.16b\n"
-            ".inst 0x4e9fa77a  // smmla v26.4s, v27.16b, v31.16b\n"
-            "ldr q27, [x25, #0x20]\n"
-            ".inst 0x4e88a5f2  // smmla v18.4s, v15.16b, v8.16b\n"
-            ".inst 0x4e9fa5f1  // smmla v17.4s, v15.16b, v31.16b\n"
-            "sshl v15.16b, v29.16b, v1.16b\n"
-            "sshl v1.16b, v3.16b, v1.16b\n"
-            "and v29.16b, v29.16b, v0.16b\n"
-            "and v3.16b, v3.16b, v0.16b\n"
-            "ldr q0, [x25, #0x30]\n"
-            "fcvtl v20.4s, v20.4h\n"
-            ".inst 0x4e8fa773  // smmla v19.4s, v27.16b, v15.16b\n"
-            "fcvtl v9.4s, v9.4h\n"
-            ".inst 0x4e81a77a  // smmla v26.4s, v27.16b, v1.16b\n"
-            "ldr q27, [x25, #0x40]\n"
-            ".inst 0x4e8fa412  // smmla v18.4s, v0.16b, v15.16b\n"
-            ".inst 0x4e81a411  // smmla v17.4s, v0.16b, v1.16b\n"
-            "ldr q0, [x25, #0x50]\n"
-            ".inst 0x4e95a773  // smmla v19.4s, v27.16b, v21.16b\n"
-            ".inst 0x4e90a77a  // smmla v26.4s, v27.16b, v16.16b\n"
-            "ldr q27, [x25, #0x60]\n"
-            ".inst 0x4e95a412  // smmla v18.4s, v0.16b, v21.16b\n"
-            ".inst 0x4e90a411  // smmla v17.4s, v0.16b, v16.16b\n"
-            "ldr q0, [x25, #0x70]\n"
-            "add x25, x25, #0x88\n"
-            ".inst 0x4e9da773  // smmla v19.4s, v27.16b, v29.16b\n"
-            ".inst 0x4e83a77a  // smmla v26.4s, v27.16b, v3.16b\n"
-            "ldr d27, [x20, #0x0]\n"
-            ".inst 0x4e9da412  // smmla v18.4s, v0.16b, v29.16b\n"
-            ".inst 0x4e83a411  // smmla v17.4s, v0.16b, v3.16b\n"
-            "fcvtl v27.4s, v27.4h\n"
-            "uzp1 v0.2d, v19.2d, v26.2d\n"
-            "uzp2 v26.2d, v19.2d, v26.2d\n"
-            "fmul v19.4s, v27.4s, v20.s[0]\n"
-            "scvtf v0.4s, v0.4s, #0x4\n"
-            "scvtf v26.4s, v26.4s, #0x4\n"
-            "fmla v2.4s, v0.4s, v19.4s\n"
-            "ldr q19, [x23, #0x0]\n"
-            "uzp1 v0.2d, v18.2d, v17.2d\n"
-            "uzp2 v18.2d, v18.2d, v17.2d\n"
-            "fmul v17.4s, v27.4s, v20.s[1]\n"
-            "scvtf v0.4s, v0.4s, #0x4\n"
-            "scvtf v18.4s, v18.4s, #0x4\n"
-            "fmla v10.4s, v26.4s, v17.4s\n"
-            "ldr q17, [x23, #0x10]\n"
-            "fmul v26.4s, v27.4s, v20.s[2]\n"
-            "fmul v20.4s, v27.4s, v20.s[3]\n"
-            "fmla v12.4s, v0.4s, v26.4s\n"
-            "ldr d0, [x22, #-0x8]\n"
-            "ldr d26, [x21, #-0x8]\n"
-            "fcvtl v0.4s, v0.4h\n"
-            "fmla v28.4s, v18.4s, v20.4s\n"
-            "movi v20.4s, #0x0\n"
-            "movi v18.4s, #0x0\n"
-            ".inst 0x4e88a674  // smmla v20.4s, v19.16b, v8.16b\n"
-            ".inst 0x4e9fa672  // smmla v18.4s, v19.16b, v31.16b\n"
-            "ldr q19, [x23, #0x20]\n"
-            "fcvtl v26.4s, v26.4h\n"
-            ".inst 0x4e8fa674  // smmla v20.4s, v19.16b, v15.16b\n"
-            ".inst 0x4e81a672  // smmla v18.4s, v19.16b, v1.16b\n"
-            "ldr q19, [x23, #0x40]\n"
-            ".inst 0x4e95a674  // smmla v20.4s, v19.16b, v21.16b\n"
-            ".inst 0x4e90a672  // smmla v18.4s, v19.16b, v16.16b\n"
-            "ldr q19, [x23, #0x60]\n"
-            ".inst 0x4e9da674  // smmla v20.4s, v19.16b, v29.16b\n"
-            ".inst 0x4e83a672  // smmla v18.4s, v19.16b, v3.16b\n"
-            "uzp1 v19.2d, v20.2d, v18.2d\n"
-            "scvtf v19.4s, v19.4s, #0x4\n"
-            "uzp2 v20.2d, v20.2d, v18.2d\n"
-            "fmul v18.4s, v27.4s, v9.s[0]\n"
-            "scvtf v20.4s, v20.4s, #0x4\n"
-            "fmla v11.4s, v19.4s, v18.4s\n"
-            "ldr q18, [x22, #0x0]\n"
-            "fmul v19.4s, v27.4s, v9.s[1]\n"
-            "fmla v13.4s, v20.4s, v19.4s\n"
-            "movi v19.4s, #0x0\n"
-            "movi v20.4s, #0x0\n"
-            ".inst 0x4e88a633  // smmla v19.4s, v17.16b, v8.16b\n"
-            ".inst 0x4e9fa634  // smmla v20.4s, v17.16b, v31.16b\n"
-            "ldr q17, [x23, #0x30]\n"
-            ".inst 0x4e8fa633  // smmla v19.4s, v17.16b, v15.16b\n"
-            ".inst 0x4e81a634  // smmla v20.4s, v17.16b, v1.16b\n"
-            "ldr q17, [x23, #0x50]\n"
-            ".inst 0x4e95a633  // smmla v19.4s, v17.16b, v21.16b\n"
-            ".inst 0x4e90a634  // smmla v20.4s, v17.16b, v16.16b\n"
-            "ldr q17, [x23, #0x70]\n"
-            "add x23, x23, #0x88\n"
-            ".inst 0x4e9da633  // smmla v19.4s, v17.16b, v29.16b\n"
-            ".inst 0x4e83a634  // smmla v20.4s, v17.16b, v3.16b\n"
-            "uzp1 v17.2d, v19.2d, v20.2d\n"
-            "scvtf v17.4s, v17.4s, #0x4\n"
-            "uzp2 v20.2d, v19.2d, v20.2d\n"
-            "fmul v19.4s, v27.4s, v9.s[2]\n"
-            "fmul v9.4s, v27.4s, v9.s[3]\n"
-            "scvtf v20.4s, v20.4s, #0x4\n"
-            "fmla v22.4s, v17.4s, v19.4s\n"
-            "ldr q17, [x22, #0x10]\n"
-            "movi v19.4s, #0x0\n"
-            ".inst 0x4e88a653  // smmla v19.4s, v18.16b, v8.16b\n"
-            "fmla v23.4s, v20.4s, v9.4s\n"
-            "movi v20.4s, #0x0\n"
-            "movi v9.4s, #0x0\n"
-            ".inst 0x4e9fa654  // smmla v20.4s, v18.16b, v31.16b\n"
-            "ldr q18, [x22, #0x20]\n"
-            ".inst 0x4e88a629  // smmla v9.4s, v17.16b, v8.16b\n"
-            ".inst 0x4e8fa653  // smmla v19.4s, v18.16b, v15.16b\n"
-            ".inst 0x4e81a654  // smmla v20.4s, v18.16b, v1.16b\n"
-            "ldr q18, [x22, #0x40]\n"
-            ".inst 0x4e95a653  // smmla v19.4s, v18.16b, v21.16b\n"
-            ".inst 0x4e90a654  // smmla v20.4s, v18.16b, v16.16b\n"
-            "ldr q18, [x22, #0x60]\n"
-            ".inst 0x4e9da653  // smmla v19.4s, v18.16b, v29.16b\n"
-            ".inst 0x4e83a654  // smmla v20.4s, v18.16b, v3.16b\n"
-            "movi v18.4s, #0x0\n"
-            ".inst 0x4e9fa632  // smmla v18.4s, v17.16b, v31.16b\n"
-            "ldr q17, [x22, #0x30]\n"
-            ".inst 0x4e8fa629  // smmla v9.4s, v17.16b, v15.16b\n"
-            ".inst 0x4e81a632  // smmla v18.4s, v17.16b, v1.16b\n"
-            "ldr q17, [x22, #0x50]\n"
-            ".inst 0x4e95a629  // smmla v9.4s, v17.16b, v21.16b\n"
-            ".inst 0x4e90a632  // smmla v18.4s, v17.16b, v16.16b\n"
-            "ldr q17, [x22, #0x70]\n"
-            "add x22, x22, #0x88\n"
-            ".inst 0x4e9da629  // smmla v9.4s, v17.16b, v29.16b\n"
-            ".inst 0x4e83a632  // smmla v18.4s, v17.16b, v3.16b\n"
-            "uzp1 v17.2d, v19.2d, v20.2d\n"
-            "uzp2 v20.2d, v19.2d, v20.2d\n"
-            "fmul v19.4s, v27.4s, v0.s[0]\n"
-            "scvtf v17.4s, v17.4s, #0x4\n"
-            "scvtf v20.4s, v20.4s, #0x4\n"
-            "fmla v25.4s, v17.4s, v19.4s\n"
-            "ldr q19, [x21, #0x0]\n"
-            "fmul v17.4s, v27.4s, v0.s[1]\n"
-            "fmla v5.4s, v20.4s, v17.4s\n"
-            "ldr q17, [x21, #0x10]\n"
-            "uzp1 v20.2d, v9.2d, v18.2d\n"
-            "uzp2 v9.2d, v9.2d, v18.2d\n"
-            "fmul v18.4s, v27.4s, v0.s[2]\n"
-            "fmul v0.4s, v27.4s, v0.s[3]\n"
-            "scvtf v20.4s, v20.4s, #0x4\n"
-            "scvtf v9.4s, v9.4s, #0x4\n"
-            "fmla v7.4s, v20.4s, v18.4s\n"
-            "movi v20.4s, #0x0\n"
-            "movi v18.4s, #0x0\n"
-            ".inst 0x4e88a674  // smmla v20.4s, v19.16b, v8.16b\n"
-            ".inst 0x4e9fa672  // smmla v18.4s, v19.16b, v31.16b\n"
-            "ldr q19, [x21, #0x20]\n"
-            "fmla v4.4s, v9.4s, v0.4s\n"
-            "movi v9.4s, #0x0\n"
-            "movi v0.4s, #0x0\n"
-            ".inst 0x4e88a629  // smmla v9.4s, v17.16b, v8.16b\n"
-            "fmul v8.4s, v27.4s, v26.s[0]\n"
-            ".inst 0x4e9fa620  // smmla v0.4s, v17.16b, v31.16b\n"
-            "ldr q17, [x21, #0x30]\n"
-            ".inst 0x4e8fa674  // smmla v20.4s, v19.16b, v15.16b\n"
-            "fmul v31.4s, v27.4s, v26.s[1]\n"
-            ".inst 0x4e81a672  // smmla v18.4s, v19.16b, v1.16b\n"
-            "ldr q19, [x21, #0x40]\n"
-            ".inst 0x4e8fa629  // smmla v9.4s, v17.16b, v15.16b\n"
-            "fmul v15.4s, v27.4s, v26.s[2]\n"
-            "fmul v27.4s, v27.4s, v26.s[3]\n"
-            ".inst 0x4e81a620  // smmla v0.4s, v17.16b, v1.16b\n"
-            "ldr q1, [x21, #0x50]\n"
-            ".inst 0x4e95a674  // smmla v20.4s, v19.16b, v21.16b\n"
-            ".inst 0x4e90a672  // smmla v18.4s, v19.16b, v16.16b\n"
-            "ldr q26, [x21, #0x60]\n"
-            ".inst 0x4e95a429  // smmla v9.4s, v1.16b, v21.16b\n"
-            ".inst 0x4e90a420  // smmla v0.4s, v1.16b, v16.16b\n"
-            "ldr q21, [x21, #0x70]\n"
-            "add x21, x21, #0x88\n"
-            ".inst 0x4e9da754  // smmla v20.4s, v26.16b, v29.16b\n"
-            ".inst 0x4e83a752  // smmla v18.4s, v26.16b, v3.16b\n"
-            ".inst 0x4e9da6a9  // smmla v9.4s, v21.16b, v29.16b\n"
-            ".inst 0x4e83a6a0  // smmla v0.4s, v21.16b, v3.16b\n"
-            "uzp1 v29.2d, v20.2d, v18.2d\n"
-            "uzp2 v21.2d, v20.2d, v18.2d\n"
-            "scvtf v29.4s, v29.4s, #0x4\n"
-            "uzp1 v18.2d, v9.2d, v0.2d\n"
-            "uzp2 v16.2d, v9.2d, v0.2d\n"
-            "scvtf v21.4s, v21.4s, #0x4\n"
-            "fmla v6.4s, v29.4s, v8.4s\n"
-            "scvtf v18.4s, v18.4s, #0x4\n"
-            "scvtf v16.4s, v16.4s, #0x4\n"
-            "fmla v30.4s, v21.4s, v31.4s\n"
-            "fmla v24.4s, v18.4s, v15.4s\n"
-            "fmla v14.4s, v16.4s, v27.4s\n"
-            "bgt 3b\n"
-            "mov x20, %x[res_ptr]\n"
-            "subs x27, x27, #0x4\n"
-            "add %x[res_ptr], %x[res_ptr], #0x10\n"
-            "str q2, [x20, #0x0]\n"
-            "add x20, x20, %x[res_stride]\n"
-            "str q10, [x20, #0x0]\n"
-            "add x20, x20, %x[res_stride]\n"
-            "str q12, [x20, #0x0]\n"
-            "add x20, x20, %x[res_stride]\n"
-            "str q28, [x20, #0x0]\n"
-            "add x20, x20, %x[res_stride]\n"
-            "str q11, [x20, #0x0]\n"
-            "add x20, x20, %x[res_stride]\n"
-            "str q13, [x20, #0x0]\n"
-            "add x20, x20, %x[res_stride]\n"
-            "str q22, [x20, #0x0]\n"
-            "add x20, x20, %x[res_stride]\n"
-            "str q23, [x20, #0x0]\n"
-            "add x20, x20, %x[res_stride]\n"
-            "str q25, [x20, #0x0]\n"
-            "add x20, x20, %x[res_stride]\n"
-            "str q5, [x20, #0x0]\n"
-            "add x20, x20, %x[res_stride]\n"
-            "str q7, [x20, #0x0]\n"
-            "add x20, x20, %x[res_stride]\n"
-            "str q4, [x20, #0x0]\n"
-            "add x20, x20, %x[res_stride]\n"
-            "str q6, [x20, #0x0]\n"
-            "add x20, x20, %x[res_stride]\n"
-            "str q30, [x20, #0x0]\n"
-            "add x20, x20, %x[res_stride]\n"
-            "str q24, [x20, #0x0]\n"
-            "add x20, x20, %x[res_stride]\n"
-            "str q14, [x20, #0x0]\n"
-            "bne 2b\n"
-            "mov x20, #0x4\n"
-            "sub x10, x10, #0x10\n"
-            "cmp x10, #0x10\n"
-            "mov %x[res_ptr], x26\n"
-            "madd %x[a_ptr], x20, x9, %x[a_ptr]\n"
-            "bge 1b\n"
-            "4:"  // Row loop skip
-            "cbz x10, 9f\n"
-            "5:"  // Row tail: Row loop
-            "add x24, %x[b_ptr], #0x8\n"
-            "mov x23, %x[nc]\n"
-            "add x22, %x[res_ptr], %x[res_stride], LSL #2\n"
-            "6:"  // Row tail: Column loop
-            "movi v2.16b, #0x0\n"
-            "movi v10.16b, #0x0\n"
-            "add x25, %x[a_ptr], #0x8\n"
-            "mov x21, %x[nb]\n"
-            "movi v12.16b, #0x0\n"
-            "movi v28.16b, #0x0\n"
-            "7:"  // Row tail: Block loop
-            "ldr q6, [x24, #0x0]\n"
-            "ldr q5, [x24, #0x10]\n"
-            "movi v17.16b, #0x4\n"
-            "movi v8.4s, #0x0\n"
-            "ldr q4, [x25, #0x0]\n"
-            "ldr q13, [x25, #0x10]\n"
-            "movi v27.4s, #0x0\n"
-            "movi v0.4s, #0x0\n"
-            "ldr q31, [x24, #0x20]\n"
-            "ldr q14, [x24, #0x30]\n"
-            "movi v29.4s, #0x0\n"
-            "movi v22.16b, #0xf0\n"
-            "ldr q11, [x25, #0x20]\n"
-            "ldr q23, [x25, #0x30]\n"
-            "sshl v21.16b, v6.16b, v17.16b\n"
-            "sshl v16.16b, v5.16b, v17.16b\n"
-            "ldr q20, [x25, #0x40]\n"
-            "ldr q26, [x25, #0x50]\n"
-            "and v6.16b, v6.16b, v22.16b\n"
-            "and v5.16b, v5.16b, v22.16b\n"
-            "ldr q25, [x25, #0x60]\n"
-            "ldr q3, [x25, #0x70]\n"
-            "sshl v19.16b, v31.16b, v17.16b\n"
-            "sshl v18.16b, v14.16b, v17.16b\n"
-            "ldr d17, [x25, #-0x8]\n"
-            ".inst 0x4e95a488  // smmla v8.4s, v4.16b, v21.16b\n"
-            ".inst 0x4e90a49b  // smmla v27.4s, v4.16b, v16.16b\n"
-            "and v31.16b, v31.16b, v22.16b\n"
-            ".inst 0x4e95a5a0  // smmla v0.4s, v13.16b, v21.16b\n"
-            ".inst 0x4e90a5bd  // smmla v29.4s, v13.16b, v16.16b\n"
-            "and v14.16b, v14.16b, v22.16b\n"
-            "sub x20, x24, #0x8\n"
-            "ldr d16, [x20, #0x0]\n"
-            "subs x21, x21, #0x1\n"
-            "add x25, x25, #0x88\n"
-            "fcvtl v17.4s, v17.4h\n"
-            "add x24, x24, #0x48\n"
-            ".inst 0x4e93a568  // smmla v8.4s, v11.16b, v19.16b\n"
-            ".inst 0x4e92a57b  // smmla v27.4s, v11.16b, v18.16b\n"
-            ".inst 0x4e93a6e0  // smmla v0.4s, v23.16b, v19.16b\n"
-            ".inst 0x4e92a6fd  // smmla v29.4s, v23.16b, v18.16b\n"
-            "fcvtl v16.4s, v16.4h\n"
-            ".inst 0x4e86a688  // smmla v8.4s, v20.16b, v6.16b\n"
-            ".inst 0x4e85a69b  // smmla v27.4s, v20.16b, v5.16b\n"
-            "fmul v23.4s, v16.4s, v17.s[0]\n"
-            "fmul v21.4s, v16.4s, v17.s[1]\n"
-            "fmul v1.4s, v16.4s, v17.s[2]\n"
-            "fmul v20.4s, v16.4s, v17.s[3]\n"
-            ".inst 0x4e86a740  // smmla v0.4s, v26.16b, v6.16b\n"
-            ".inst 0x4e85a75d  // smmla v29.4s, v26.16b, v5.16b\n"
-            ".inst 0x4e9fa728  // smmla v8.4s, v25.16b, v31.16b\n"
-            ".inst 0x4e8ea73b  // smmla v27.4s, v25.16b, v14.16b\n"
-            ".inst 0x4e9fa460  // smmla v0.4s, v3.16b, v31.16b\n"
-            ".inst 0x4e8ea47d  // smmla v29.4s, v3.16b, v14.16b\n"
-            "uzp1 v19.2d, v8.2d, v27.2d\n"
-            "uzp2 v18.2d, v8.2d, v27.2d\n"
-            "scvtf v19.4s, v19.4s, #0x4\n"
-            "uzp1 v17.2d, v0.2d, v29.2d\n"
-            "uzp2 v16.2d, v0.2d, v29.2d\n"
-            "scvtf v18.4s, v18.4s, #0x4\n"
-            "fmla v2.4s, v19.4s, v23.4s\n"
-            "scvtf v17.4s, v17.4s, #0x4\n"
-            "scvtf v16.4s, v16.4s, #0x4\n"
-            "fmla v10.4s, v18.4s, v21.4s\n"
-            "fmla v12.4s, v17.4s, v1.4s\n"
-            "fmla v28.4s, v16.4s, v20.4s\n"
-            "bgt 7b\n"
-            "mov x20, %x[res_ptr]\n"
-            "cmp x10, #0x1\n"
-            "str q2, [x20, #0x0]\n"
-            "add x20, x20, %x[res_stride]\n"
-            "ble 8f\n"
-            "cmp x10, #0x2\n"
-            "str q10, [x20, #0x0]\n"
-            "add x20, x20, %x[res_stride]\n"
-            "ble 8f\n"
-            "cmp x10, #0x3\n"
-            "str q12, [x20, #0x0]\n"
-            "add x20, x20, %x[res_stride]\n"
-            "ble 8f\n"
-            "str q28, [x20, #0x0]\n"
-            "8:"  // Row tail: Accumulator store skip
-            "subs x23, x23, #0x4\n"
-            "add %x[res_ptr], %x[res_ptr], #0x10\n"
-            "bne 6b\n"
-            "subs x10, x10, #0x4\n"
-            "add %x[a_ptr], %x[a_ptr], x9\n"
-            "mov %x[res_ptr], x22\n"
-            "bgt 5b\n"
-            "9:"  // Row tail: Row loop skip
-            : [a_ptr] "+&r" (a_ptr), [res_ptr] "+&r" (res_ptr)
-            : [b_ptr] "r" (b_ptr), [nr] "r" (nr), [nb] "r" (nb), [res_stride] "r" (res_stride), [nc] "r" (nc)
-            : "cc", "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", "x9", "x10", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
-        );
-        return;
-    }
-#endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_MATMUL_INT8)
-    float sumf[4][4];
-    int sumi;
-
-    for (int y = 0; y < nr / 4; y++) {
-        const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
-        for (int x = 0; x < nc / ncols_interleaved; x++) {
-            const block_q4_0x4 * b_ptr = (const block_q4_0x4 *) vx + (x * nb);
-            for (int m = 0; m < 4; m++) {
-                for (int j = 0; j < ncols_interleaved; j++) sumf[m][j] = 0.0;
-            }
-            for (int l = 0; l < nb; l++) {
-                for (int k = 0; k < (qk / (2 * blocklen)); k++) {
-                    for (int m = 0; m < 4; m++) {
-                        for (int j = 0; j < ncols_interleaved; j++) {
-                            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] << 4);
-                                const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
-                                sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
-                                        (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4;
-                            }
-                            sumf[m][j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(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];
-            }
-        }
-    }
-}
-
-static void ggml_gemm_q4_0_8x8_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) {
+void ggml_gemm_q4_0_8x8_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;
     const int ncols_interleaved = 8;
@@ -2647,420 +1010,7 @@ static void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, c
     UNUSED(ncols_interleaved);
     UNUSED(blocklen);
 
-#if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__)
-#if defined(__ARM_FEATURE_SVE) && defined(__ARM_FEATURE_MATMUL_INT8)
-    if (ggml_cpu_has_sve() && ggml_cpu_has_matmul_int8() && ggml_cpu_get_sve_cnt() == QK8_0) {
-        const void * b_ptr = vx;
-        const void * a_ptr = vy;
-        float * res_ptr = s;
-        size_t res_stride = bs * sizeof(float);
-
-        __asm__ __volatile__(
-            "mov x20, #0x4\n"
-            "mov x13, %x[nr]\n"
-            "mov z28.s, #-0x4\n"
-            "mov x12, #0x88\n"
-            "ptrue p1.b\n"
-            "whilelt p0.s, XZR, x20\n"
-            "cmp x13, #0x10\n"
-            "mul x12, %x[nb], x12\n"
-            "blt 4f\n"
-            "1:"  // Row loop
-            "add x11, %x[b_ptr], #0x10\n"
-            "mov x10, %x[nc]\n"
-            "add x9, %x[res_ptr], %x[res_stride], LSL #4\n"
-            "2:"  // Column loop
-            "add x28, %x[a_ptr], #0x8\n"
-            "mov z24.b, #0x0\n"
-            "mov z15.b, #0x0\n"
-            "mov x27, %x[nb]\n"
-            "add x26, x28, x12\n"
-            "mov z12.b, #0x0\n"
-            "mov z0.b, #0x0\n"
-            "add x25, x26, x12\n"
-            "mov z13.b, #0x0\n"
-            "mov z1.b, #0x0\n"
-            "add x24, x25, x12\n"
-            "mov z20.b, #0x0\n"
-            "mov z25.b, #0x0\n"
-            "mov z11.b, #0x0\n"
-            "mov z16.b, #0x0\n"
-            "mov z19.b, #0x0\n"
-            "mov z26.b, #0x0\n"
-            "mov z8.b, #0x0\n"
-            "mov z29.b, #0x0\n"
-            "mov z27.b, #0x0\n"
-            "mov z10.b, #0x0\n"
-            "3:"  // Block loop
-            "ld1b { z30.b }, p1/Z, [x11]\n"
-            "ld1b { z21.b }, p1/Z, [x11, #1, MUL VL]\n"
-            "mov z18.s, #0x0\n"
-            "mov z7.s, #0x0\n"
-            "ld1rqb { z3.b }, p1/Z, [x28]\n"
-            "ld1rqb { z5.b }, p1/Z, [x28, #16]\n"
-            "mov z9.s, #0x0\n"
-            "mov z22.s, #0x0\n"
-            "ld1b { z4.b }, p1/Z, [x11, #2, MUL VL]\n"
-            "ld1b { z17.b }, p1/Z, [x11, #3, MUL VL]\n"
-            "sub x20, x11, #0x10\n"
-            "sub x23, x28, #0x8\n"
-            "lsl z31.b, z30.b, #0x4\n"
-            "lsl z6.b, z21.b, #0x4\n"
-            "ld1h { z23.s }, p1/Z, [x20]\n"
-            "sub x22, x26, #0x8\n"
-            "and z30.b, z30.b, #0xf0\n"
-            "and z21.b, z21.b, #0xf0\n"
-            "sub x21, x25, #0x8\n"
-            "sub x20, x24, #0x8\n"
-            "lsl z14.b, z4.b, #0x4\n"
-            "lsl z2.b, z17.b, #0x4\n"
-            "subs x27, x27, #0x1\n"
-            "add x11, x11, #0x90\n"
-            ".inst 0x451f9872  // smmla z18.s, z3.b, z31.b\n"
-            ".inst 0x45069867  // smmla z7.s, z3.b, z6.b\n"
-            "ld1rqb { z3.b }, p1/Z, [x28, #32]\n"
-            "and z4.b, z4.b, #0xf0\n"
-            ".inst 0x451f98a9  // smmla z9.s, z5.b, z31.b\n"
-            ".inst 0x450698b6  // smmla z22.s, z5.b, z6.b\n"
-            "ld1rqb { z5.b }, p1/Z, [x28, #48]\n"
-            "and z17.b, z17.b, #0xf0\n"
-            "fcvt z23.s, p1/m, z23.h\n"
-            ".inst 0x450e9872  // smmla z18.s, z3.b, z14.b\n"
-            ".inst 0x45029867  // smmla z7.s, z3.b, z2.b\n"
-            "ld1rqb { z3.b }, p1/Z, [x28, #64]\n"
-            ".inst 0x450e98a9  // smmla z9.s, z5.b, z14.b\n"
-            ".inst 0x450298b6  // smmla z22.s, z5.b, z2.b\n"
-            "ld1rqb { z5.b }, p1/Z, [x28, #80]\n"
-            "fscale z23.s, p1/m, z23.s, z28.s\n"
-            ".inst 0x451e9872  // smmla z18.s, z3.b, z30.b\n"
-            ".inst 0x45159867  // smmla z7.s, z3.b, z21.b\n"
-            "ld1rqb { z3.b }, p1/Z, [x28, #96]\n"
-            ".inst 0x451e98a9  // smmla z9.s, z5.b, z30.b\n"
-            ".inst 0x451598b6  // smmla z22.s, z5.b, z21.b\n"
-            "ld1rqb { z5.b }, p1/Z, [x28, #112]\n"
-            "add x28, x28, #0x88\n"
-            ".inst 0x45049872  // smmla z18.s, z3.b, z4.b\n"
-            ".inst 0x45119867  // smmla z7.s, z3.b, z17.b\n"
-            "ld1h { z3.s }, p0/Z, [x23]\n"
-            ".inst 0x450498a9  // smmla z9.s, z5.b, z4.b\n"
-            ".inst 0x451198b6  // smmla z22.s, z5.b, z17.b\n"
-            "fcvt z3.s, p1/m, z3.h\n"
-            "uzp1 z5.d, z18.d, z7.d\n"
-            "uzp2 z18.d, z18.d, z7.d\n"
-            "mov z3.q, z3.q[0]\n"
-            "uzp1 z7.d, z9.d, z22.d\n"
-            "uzp2 z22.d, z9.d, z22.d\n"
-            "fmul z9.s, z23.s, z3.s[0]\n"
-            "scvtf z5.s, p1/m, z5.s\n"
-            "scvtf z18.s, p1/m, z18.s\n"
-            "scvtf z7.s, p1/m, z7.s\n"
-            "scvtf z22.s, p1/m, z22.s\n"
-            "fmla z24.s, p1/M, z5.s, z9.s\n"
-            "ld1rqb { z5.b }, p1/Z, [x26]\n"
-            "fmul z9.s, z23.s, z3.s[1]\n"
-            "fmla z15.s, p1/M, z18.s, z9.s\n"
-            "ld1rqb { z18.b }, p1/Z, [x26, #16]\n"
-            "fmul z9.s, z23.s, z3.s[2]\n"
-            "fmul z3.s, z23.s, z3.s[3]\n"
-            "fmla z12.s, p1/M, z7.s, z9.s\n"
-            "mov z9.s, #0x0\n"
-            "ld1h { z7.s }, p0/Z, [x22]\n"
-            ".inst 0x451f98a9  // smmla z9.s, z5.b, z31.b\n"
-            "fmla z0.s, p1/M, z22.s, z3.s\n"
-            "mov z22.s, #0x0\n"
-            "ld1h { z3.s }, p0/Z, [x21]\n"
-            ".inst 0x450698b6  // smmla z22.s, z5.b, z6.b\n"
-            "ld1rqb { z5.b }, p1/Z, [x26, #32]\n"
-            "fcvt z7.s, p1/m, z7.h\n"
-            "fcvt z3.s, p1/m, z3.h\n"
-            ".inst 0x450e98a9  // smmla z9.s, z5.b, z14.b\n"
-            ".inst 0x450298b6  // smmla z22.s, z5.b, z2.b\n"
-            "ld1rqb { z5.b }, p1/Z, [x26, #64]\n"
-            "mov z7.q, z7.q[0]\n"
-            "mov z3.q, z3.q[0]\n"
-            ".inst 0x451e98a9  // smmla z9.s, z5.b, z30.b\n"
-            ".inst 0x451598b6  // smmla z22.s, z5.b, z21.b\n"
-            "ld1rqb { z5.b }, p1/Z, [x26, #96]\n"
-            ".inst 0x450498a9  // smmla z9.s, z5.b, z4.b\n"
-            ".inst 0x451198b6  // smmla z22.s, z5.b, z17.b\n"
-            "uzp1 z5.d, z9.d, z22.d\n"
-            "scvtf z5.s, p1/m, z5.s\n"
-            "uzp2 z22.d, z9.d, z22.d\n"
-            "fmul z9.s, z23.s, z7.s[0]\n"
-            "scvtf z22.s, p1/m, z22.s\n"
-            "fmla z13.s, p1/M, z5.s, z9.s\n"
-            "ld1rqb { z9.b }, p1/Z, [x25]\n"
-            "fmul z5.s, z23.s, z7.s[1]\n"
-            "fmla z1.s, p1/M, z22.s, z5.s\n"
-            "mov z5.s, #0x0\n"
-            "mov z22.s, #0x0\n"
-            ".inst 0x451f9a45  // smmla z5.s, z18.b, z31.b\n"
-            ".inst 0x45069a56  // smmla z22.s, z18.b, z6.b\n"
-            "ld1rqb { z18.b }, p1/Z, [x26, #48]\n"
-            ".inst 0x450e9a45  // smmla z5.s, z18.b, z14.b\n"
-            ".inst 0x45029a56  // smmla z22.s, z18.b, z2.b\n"
-            "ld1rqb { z18.b }, p1/Z, [x26, #80]\n"
-            ".inst 0x451e9a45  // smmla z5.s, z18.b, z30.b\n"
-            ".inst 0x45159a56  // smmla z22.s, z18.b, z21.b\n"
-            "ld1rqb { z18.b }, p1/Z, [x26, #112]\n"
-            "add x26, x26, #0x88\n"
-            ".inst 0x45049a45  // smmla z5.s, z18.b, z4.b\n"
-            ".inst 0x45119a56  // smmla z22.s, z18.b, z17.b\n"
-            "uzp1 z18.d, z5.d, z22.d\n"
-            "scvtf z18.s, p1/m, z18.s\n"
-            "uzp2 z22.d, z5.d, z22.d\n"
-            "fmul z5.s, z23.s, z7.s[2]\n"
-            "fmul z7.s, z23.s, z7.s[3]\n"
-            "scvtf z22.s, p1/m, z22.s\n"
-            "fmla z20.s, p1/M, z18.s, z5.s\n"
-            "ld1rqb { z18.b }, p1/Z, [x25, #16]\n"
-            "ld1h { z5.s }, p0/Z, [x20]\n"
-            "fcvt z5.s, p1/m, z5.h\n"
-            "fmla z25.s, p1/M, z22.s, z7.s\n"
-            "mov z22.s, #0x0\n"
-            "mov z7.s, #0x0\n"
-            ".inst 0x451f9936  // smmla z22.s, z9.b, z31.b\n"
-            ".inst 0x45069927  // smmla z7.s, z9.b, z6.b\n"
-            "ld1rqb { z9.b }, p1/Z, [x25, #32]\n"
-            "mov z5.q, z5.q[0]\n"
-            ".inst 0x450e9936  // smmla z22.s, z9.b, z14.b\n"
-            ".inst 0x45029927  // smmla z7.s, z9.b, z2.b\n"
-            "ld1rqb { z9.b }, p1/Z, [x25, #64]\n"
-            ".inst 0x451e9936  // smmla z22.s, z9.b, z30.b\n"
-            ".inst 0x45159927  // smmla z7.s, z9.b, z21.b\n"
-            "ld1rqb { z9.b }, p1/Z, [x25, #96]\n"
-            ".inst 0x45049936  // smmla z22.s, z9.b, z4.b\n"
-            ".inst 0x45119927  // smmla z7.s, z9.b, z17.b\n"
-            "uzp1 z9.d, z22.d, z7.d\n"
-            "scvtf z9.s, p1/m, z9.s\n"
-            "uzp2 z22.d, z22.d, z7.d\n"
-            "fmul z7.s, z23.s, z3.s[0]\n"
-            "scvtf z22.s, p1/m, z22.s\n"
-            "fmla z11.s, p1/M, z9.s, z7.s\n"
-            "ld1rqb { z9.b }, p1/Z, [x24]\n"
-            "fmul z7.s, z23.s, z3.s[1]\n"
-            "fmla z16.s, p1/M, z22.s, z7.s\n"
-            "mov z22.s, #0x0\n"
-            "mov z7.s, #0x0\n"
-            ".inst 0x451f9a56  // smmla z22.s, z18.b, z31.b\n"
-            ".inst 0x45069a47  // smmla z7.s, z18.b, z6.b\n"
-            "ld1rqb { z18.b }, p1/Z, [x25, #48]\n"
-            ".inst 0x450e9a56  // smmla z22.s, z18.b, z14.b\n"
-            ".inst 0x45029a47  // smmla z7.s, z18.b, z2.b\n"
-            "ld1rqb { z18.b }, p1/Z, [x25, #80]\n"
-            ".inst 0x451e9a56  // smmla z22.s, z18.b, z30.b\n"
-            ".inst 0x45159a47  // smmla z7.s, z18.b, z21.b\n"
-            "ld1rqb { z18.b }, p1/Z, [x25, #112]\n"
-            "add x25, x25, #0x88\n"
-            ".inst 0x45049a56  // smmla z22.s, z18.b, z4.b\n"
-            ".inst 0x45119a47  // smmla z7.s, z18.b, z17.b\n"
-            "uzp1 z18.d, z22.d, z7.d\n"
-            "scvtf z18.s, p1/m, z18.s\n"
-            "uzp2 z7.d, z22.d, z7.d\n"
-            "fmul z22.s, z23.s, z3.s[2]\n"
-            "fmul z3.s, z23.s, z3.s[3]\n"
-            "scvtf z7.s, p1/m, z7.s\n"
-            "fmla z19.s, p1/M, z18.s, z22.s\n"
-            "ld1rqb { z18.b }, p1/Z, [x24, #16]\n"
-            "fmul z22.s, z23.s, z5.s[0]\n"
-            "fmla z26.s, p1/M, z7.s, z3.s\n"
-            "mov z3.s, #0x0\n"
-            "mov z7.s, #0x0\n"
-            ".inst 0x451f9923  // smmla z3.s, z9.b, z31.b\n"
-            ".inst 0x45069927  // smmla z7.s, z9.b, z6.b\n"
-            "ld1rqb { z9.b }, p1/Z, [x24, #32]\n"
-            ".inst 0x450e9923  // smmla z3.s, z9.b, z14.b\n"
-            ".inst 0x45029927  // smmla z7.s, z9.b, z2.b\n"
-            "mov z9.s, #0x0\n"
-            ".inst 0x451f9a49  // smmla z9.s, z18.b, z31.b\n"
-            "mov z31.s, #0x0\n"
-            ".inst 0x45069a5f  // smmla z31.s, z18.b, z6.b\n"
-            "ld1rqb { z6.b }, p1/Z, [x24, #48]\n"
-            "ld1rqb { z18.b }, p1/Z, [x24, #64]\n"
-            ".inst 0x450e98c9  // smmla z9.s, z6.b, z14.b\n"
-            "fmul z14.s, z23.s, z5.s[1]\n"
-            ".inst 0x450298df  // smmla z31.s, z6.b, z2.b\n"
-            "ld1rqb { z6.b }, p1/Z, [x24, #80]\n"
-            "fmul z2.s, z23.s, z5.s[2]\n"
-            "fmul z23.s, z23.s, z5.s[3]\n"
-            ".inst 0x451e9a43  // smmla z3.s, z18.b, z30.b\n"
-            ".inst 0x45159a47  // smmla z7.s, z18.b, z21.b\n"
-            "ld1rqb { z5.b }, p1/Z, [x24, #96]\n"
-            ".inst 0x451e98c9  // smmla z9.s, z6.b, z30.b\n"
-            ".inst 0x451598df  // smmla z31.s, z6.b, z21.b\n"
-            "ld1rqb { z18.b }, p1/Z, [x24, #112]\n"
-            "add x24, x24, #0x88\n"
-            ".inst 0x450498a3  // smmla z3.s, z5.b, z4.b\n"
-            ".inst 0x451198a7  // smmla z7.s, z5.b, z17.b\n"
-            ".inst 0x45049a49  // smmla z9.s, z18.b, z4.b\n"
-            ".inst 0x45119a5f  // smmla z31.s, z18.b, z17.b\n"
-            "uzp1 z18.d, z3.d, z7.d\n"
-            "uzp2 z5.d, z3.d, z7.d\n"
-            "scvtf z18.s, p1/m, z18.s\n"
-            "uzp1 z6.d, z9.d, z31.d\n"
-            "uzp2 z9.d, z9.d, z31.d\n"
-            "scvtf z5.s, p1/m, z5.s\n"
-            "fmla z8.s, p1/M, z18.s, z22.s\n"
-            "scvtf z6.s, p1/m, z6.s\n"
-            "scvtf z9.s, p1/m, z9.s\n"
-            "fmla z29.s, p1/M, z5.s, z14.s\n"
-            "fmla z27.s, p1/M, z6.s, z2.s\n"
-            "fmla z10.s, p1/M, z9.s, z23.s\n"
-            "bgt 3b\n"
-            "mov x20, %x[res_ptr]\n"
-            "subs x10, x10, #0x8\n"
-            "add %x[res_ptr], %x[res_ptr], #0x20\n"
-            "st1w { z24.s }, p1, [x20]\n"
-            "add x20, x20, %x[res_stride]\n"
-            "st1w { z15.s }, p1, [x20]\n"
-            "add x20, x20, %x[res_stride]\n"
-            "st1w { z12.s }, p1, [x20]\n"
-            "add x20, x20, %x[res_stride]\n"
-            "st1w { z0.s }, p1, [x20]\n"
-            "add x20, x20, %x[res_stride]\n"
-            "st1w { z13.s }, p1, [x20]\n"
-            "add x20, x20, %x[res_stride]\n"
-            "st1w { z1.s }, p1, [x20]\n"
-            "add x20, x20, %x[res_stride]\n"
-            "st1w { z20.s }, p1, [x20]\n"
-            "add x20, x20, %x[res_stride]\n"
-            "st1w { z25.s }, p1, [x20]\n"
-            "add x20, x20, %x[res_stride]\n"
-            "st1w { z11.s }, p1, [x20]\n"
-            "add x20, x20, %x[res_stride]\n"
-            "st1w { z16.s }, p1, [x20]\n"
-            "add x20, x20, %x[res_stride]\n"
-            "st1w { z19.s }, p1, [x20]\n"
-            "add x20, x20, %x[res_stride]\n"
-            "st1w { z26.s }, p1, [x20]\n"
-            "add x20, x20, %x[res_stride]\n"
-            "st1w { z8.s }, p1, [x20]\n"
-            "add x20, x20, %x[res_stride]\n"
-            "st1w { z29.s }, p1, [x20]\n"
-            "add x20, x20, %x[res_stride]\n"
-            "st1w { z27.s }, p1, [x20]\n"
-            "add x20, x20, %x[res_stride]\n"
-            "st1w { z10.s }, p1, [x20]\n"
-            "bne 2b\n"
-            "mov x20, #0x4\n"
-            "sub x13, x13, #0x10\n"
-            "cmp x13, #0x10\n"
-            "mov %x[res_ptr], x9\n"
-            "madd %x[a_ptr], x20, x12, %x[a_ptr]\n"
-            "bge 1b\n"
-            "4:"  // Row loop skip
-            "cbz x13, 9f\n"
-            "5:"  // Row tail: Row loop
-            "add x25, %x[b_ptr], #0x10\n"
-            "mov x24, %x[nc]\n"
-            "add x23, %x[res_ptr], %x[res_stride], LSL #2\n"
-            "6:"  // Row tail: Column loop
-            "mov z24.b, #0x0\n"
-            "mov z15.b, #0x0\n"
-            "add x28, %x[a_ptr], #0x8\n"
-            "mov x22, %x[nb]\n"
-            "mov z12.b, #0x0\n"
-            "mov z0.b, #0x0\n"
-            "7:"  // Row tail: Block loop
-            "ld1b { z3.b }, p1/Z, [x25]\n"
-            "ld1b { z6.b }, p1/Z, [x25, #1, MUL VL]\n"
-            "mov z2.s, #0x0\n"
-            "mov z25.s, #0x0\n"
-            "ld1rqb { z26.b }, p1/Z, [x28]\n"
-            "ld1rqb { z21.b }, p1/Z, [x28, #16]\n"
-            "mov z27.s, #0x0\n"
-            "mov z19.s, #0x0\n"
-            "ld1b { z29.b }, p1/Z, [x25, #2, MUL VL]\n"
-            "ld1b { z16.b }, p1/Z, [x25, #3, MUL VL]\n"
-            "sub x21, x25, #0x10\n"
-            "sub x20, x28, #0x8\n"
-            "lsl z20.b, z3.b, #0x4\n"
-            "lsl z4.b, z6.b, #0x4\n"
-            "ld1rqb { z10.b }, p1/Z, [x28, #32]\n"
-            "ld1rqb { z23.b }, p1/Z, [x28, #48]\n"
-            "and z3.b, z3.b, #0xf0\n"
-            "and z6.b, z6.b, #0xf0\n"
-            "ld1rqb { z11.b }, p1/Z, [x28, #64]\n"
-            "ld1rqb { z7.b }, p1/Z, [x28, #80]\n"
-            "lsl z8.b, z29.b, #0x4\n"
-            "lsl z14.b, z16.b, #0x4\n"
-            "ld1rqb { z18.b }, p1/Z, [x28, #96]\n"
-            "ld1rqb { z30.b }, p1/Z, [x28, #112]\n"
-            ".inst 0x45149b42  // smmla z2.s, z26.b, z20.b\n"
-            ".inst 0x45049b59  // smmla z25.s, z26.b, z4.b\n"
-            "and z29.b, z29.b, #0xf0\n"
-            "ld1h { z17.s }, p1/Z, [x21]\n"
-            ".inst 0x45149abb  // smmla z27.s, z21.b, z20.b\n"
-            ".inst 0x45049ab3  // smmla z19.s, z21.b, z4.b\n"
-            "and z16.b, z16.b, #0xf0\n"
-            "ld1h { z4.s }, p0/Z, [x20]\n"
-            "subs x22, x22, #0x1\n"
-            "add x28, x28, #0x88\n"
-            "fcvt z17.s, p1/m, z17.h\n"
-            "add x25, x25, #0x90\n"
-            ".inst 0x45089942  // smmla z2.s, z10.b, z8.b\n"
-            ".inst 0x450e9959  // smmla z25.s, z10.b, z14.b\n"
-            "fcvt z4.s, p1/m, z4.h\n"
-            ".inst 0x45089afb  // smmla z27.s, z23.b, z8.b\n"
-            ".inst 0x450e9af3  // smmla z19.s, z23.b, z14.b\n"
-            "fscale z17.s, p1/m, z17.s, z28.s\n"
-            "mov z4.q, z4.q[0]\n"
-            ".inst 0x45039962  // smmla z2.s, z11.b, z3.b\n"
-            ".inst 0x45069979  // smmla z25.s, z11.b, z6.b\n"
-            "fmul z23.s, z17.s, z4.s[0]\n"
-            "fmul z9.s, z17.s, z4.s[1]\n"
-            "fmul z21.s, z17.s, z4.s[2]\n"
-            "fmul z4.s, z17.s, z4.s[3]\n"
-            ".inst 0x450398fb  // smmla z27.s, z7.b, z3.b\n"
-            ".inst 0x450698f3  // smmla z19.s, z7.b, z6.b\n"
-            ".inst 0x451d9a42  // smmla z2.s, z18.b, z29.b\n"
-            ".inst 0x45109a59  // smmla z25.s, z18.b, z16.b\n"
-            ".inst 0x451d9bdb  // smmla z27.s, z30.b, z29.b\n"
-            ".inst 0x45109bd3  // smmla z19.s, z30.b, z16.b\n"
-            "uzp1 z31.d, z2.d, z25.d\n"
-            "uzp2 z13.d, z2.d, z25.d\n"
-            "scvtf z31.s, p1/m, z31.s\n"
-            "uzp1 z17.d, z27.d, z19.d\n"
-            "uzp2 z18.d, z27.d, z19.d\n"
-            "scvtf z13.s, p1/m, z13.s\n"
-            "fmla z24.s, p1/M, z31.s, z23.s\n"
-            "scvtf z17.s, p1/m, z17.s\n"
-            "scvtf z18.s, p1/m, z18.s\n"
-            "fmla z15.s, p1/M, z13.s, z9.s\n"
-            "fmla z12.s, p1/M, z17.s, z21.s\n"
-            "fmla z0.s, p1/M, z18.s, z4.s\n"
-            "bgt 7b\n"
-            "mov x20, %x[res_ptr]\n"
-            "cmp x13, #0x1\n"
-            "st1w { z24.s }, p1, [x20]\n"
-            "add x20, x20, %x[res_stride]\n"
-            "ble 8f\n"
-            "cmp x13, #0x2\n"
-            "st1w { z15.s }, p1, [x20]\n"
-            "add x20, x20, %x[res_stride]\n"
-            "ble 8f\n"
-            "cmp x13, #0x3\n"
-            "st1w { z12.s }, p1, [x20]\n"
-            "add x20, x20, %x[res_stride]\n"
-            "ble 8f\n"
-            "st1w { z0.s }, p1, [x20]\n"
-            "8:"  // Row tail: Accumulator store skip
-            "subs x24, x24, #0x8\n"
-            "add %x[res_ptr], %x[res_ptr], #0x20\n"
-            "bne 6b\n"
-            "subs x13, x13, #0x4\n"
-            "add %x[a_ptr], %x[a_ptr], x12\n"
-            "mov %x[res_ptr], x23\n"
-            "bgt 5b\n"
-            "9:"  // Row tail: Row loop skip
-            : [a_ptr] "+&r" (a_ptr), [res_ptr] "+&r" (res_ptr)
-            : [b_ptr] "r" (b_ptr), [nr] "r" (nr), [nb] "r" (nb), [res_stride] "r" (res_stride), [nc] "r" (nc)
-            : "cc", "memory", "p0", "p1", "x9", "x10", "x11", "x12", "x13", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
-        );
-        return;
-    }
-#endif // #if defined(__ARM_FEATURE_SVE) && defined(__ARM_FEATURE_MATMUL_INT8)
-#elif defined(__AVX2__) || defined(__AVX512F__)
+#if defined(__AVX2__) || defined(__AVX512F__)
     {
         const block_q4_0x8 * b_ptr_start = (const block_q4_0x8 *)vx;
         const block_q8_0x4 * a_ptr_start = (const block_q8_0x4 *)vy;
@@ -3783,207 +1733,7 @@ static void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, c
         }
         return;
     }
-#elif defined __riscv_v
-    if (__riscv_vlenb() >= QK4_0) {
-        const size_t vl = QK4_0;
 
-        for (int y = 0; y < nr / 4; y++) {
-            const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
-            for (int x = 0; x < nc / ncols_interleaved; x++) {
-                const block_q4_0x8 * b_ptr = (const block_q4_0x8 *) vx + (x * nb);
-                vfloat32m1_t sumf0 = __riscv_vfmv_v_f_f32m1(0.0, vl / 4);
-                vfloat32m1_t sumf1 = __riscv_vfmv_v_f_f32m1(0.0, vl / 4);
-                vfloat32m1_t sumf2 = __riscv_vfmv_v_f_f32m1(0.0, vl / 4);
-                vfloat32m1_t sumf3 = __riscv_vfmv_v_f_f32m1(0.0, vl / 4);
-                for (int l = 0; l < nb; l++) {
-                    const vint8m4_t rhs_raw_vec = __riscv_vle8_v_i8m4((const int8_t *)b_ptr[l].qs, vl * 4);
-                    const vint8m4_t rhs_vec_lo = __riscv_vsra_vx_i8m4(__riscv_vsll_vx_i8m4(rhs_raw_vec, 4, vl * 4), 4, vl * 4);
-                    const vint8m4_t rhs_vec_hi = __riscv_vsra_vx_i8m4(rhs_raw_vec, 4, vl * 4);
-                    const vint8m2_t rhs_vec_lo_0 = __riscv_vget_v_i8m4_i8m2(rhs_vec_lo, 0);
-                    const vint8m2_t rhs_vec_lo_1 = __riscv_vget_v_i8m4_i8m2(rhs_vec_lo, 1);
-                    const vint8m2_t rhs_vec_hi_0 = __riscv_vget_v_i8m4_i8m2(rhs_vec_hi, 0);
-                    const vint8m2_t rhs_vec_hi_1 = __riscv_vget_v_i8m4_i8m2(rhs_vec_hi, 1);
-
-                    // vector version needs Zvfhmin extension
-                    const float a_scales[4] = {
-                        GGML_FP16_TO_FP32(a_ptr[l].d[0]),
-                        GGML_FP16_TO_FP32(a_ptr[l].d[1]),
-                        GGML_FP16_TO_FP32(a_ptr[l].d[2]),
-                        GGML_FP16_TO_FP32(a_ptr[l].d[3])
-                    };
-                    const float b_scales[8] = {
-                        GGML_FP16_TO_FP32(b_ptr[l].d[0]),
-                        GGML_FP16_TO_FP32(b_ptr[l].d[1]),
-                        GGML_FP16_TO_FP32(b_ptr[l].d[2]),
-                        GGML_FP16_TO_FP32(b_ptr[l].d[3]),
-                        GGML_FP16_TO_FP32(b_ptr[l].d[4]),
-                        GGML_FP16_TO_FP32(b_ptr[l].d[5]),
-                        GGML_FP16_TO_FP32(b_ptr[l].d[6]),
-                        GGML_FP16_TO_FP32(b_ptr[l].d[7])
-                    };
-                    const vfloat32m1_t b_scales_vec = __riscv_vle32_v_f32m1(b_scales, vl / 4);
-
-                    const int64_t A0 = *(const int64_t *)&a_ptr[l].qs[0];
-                    const int64_t A4 = *(const int64_t *)&a_ptr[l].qs[32];
-                    const int64_t A8 = *(const int64_t *)&a_ptr[l].qs[64];
-                    const int64_t Ac = *(const int64_t *)&a_ptr[l].qs[96];
-                    __asm__ __volatile__("" ::: "memory"); // prevent gcc from emitting fused vlse64, violating alignment
-                    vint16m4_t sumi_l0;
-                    {
-                        const vint8m2_t lhs_0_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A0, vl / 4));
-                        const vint8m2_t lhs_1_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A4, vl / 4));
-                        const vint8m2_t lhs_2_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A8, vl / 4));
-                        const vint8m2_t lhs_3_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(Ac, vl / 4));
-                        const vint16m4_t sumi_lo_0 = __riscv_vwmul_vv_i16m4(rhs_vec_lo_0, lhs_0_8, vl * 2);
-                        const vint16m4_t sumi_lo_1 = __riscv_vwmacc_vv_i16m4(sumi_lo_0, rhs_vec_lo_1, lhs_1_8, vl * 2);
-                        const vint16m4_t sumi_hi_0 = __riscv_vwmacc_vv_i16m4(sumi_lo_1, rhs_vec_hi_0, lhs_2_8, vl * 2);
-                        const vint16m4_t sumi_hi_m = __riscv_vwmacc_vv_i16m4(sumi_hi_0, rhs_vec_hi_1, lhs_3_8, vl * 2);
-
-                        sumi_l0 = sumi_hi_m;
-                    }
-
-                    {
-                        const vuint32m4_t sumi_i32 = __riscv_vreinterpret_v_i32m4_u32m4(__riscv_vreinterpret_v_i16m4_i32m4(sumi_l0));
-                        const vuint16m2_t sumi_h2_0 = __riscv_vnsrl_wx_u16m2(sumi_i32, 0, vl);
-                        const vuint16m2_t sumi_h2_1 = __riscv_vnsrl_wx_u16m2(sumi_i32, 16, vl);
-                        const vuint16m2_t sumi_h2 = __riscv_vadd_vv_u16m2(sumi_h2_0, sumi_h2_1, vl);
-                        const vuint32m2_t sumi_h2_i32 = __riscv_vreinterpret_v_u16m2_u32m2(sumi_h2);
-                        const vuint16m1_t sumi_h4_0 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 0, vl / 2);
-                        const vuint16m1_t sumi_h4_1 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 16, vl / 2);
-                        const vuint16m1_t sumi_h4 = __riscv_vadd_vv_u16m1(sumi_h4_0, sumi_h4_1, vl / 2);
-                        const vuint32m1_t sumi_h4_i32 = __riscv_vreinterpret_v_u16m1_u32m1(sumi_h4);
-                        const vint16mf2_t sumi_h8_0 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 0, vl / 4));
-                        const vint16mf2_t sumi_h8_1 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 16, vl / 4));
-                        const vint32m1_t sumi_h8 = __riscv_vwadd_vv_i32m1(sumi_h8_0, sumi_h8_1, vl / 4);
-                        const vfloat32m1_t facc = __riscv_vfcvt_f_x_v_f32m1(sumi_h8, vl / 4);
-
-                        const vfloat32m1_t tmp1 = __riscv_vfmul_vf_f32m1(facc, a_scales[0], vl / 4);
-                        sumf0 = __riscv_vfmacc_vv_f32m1(sumf0, tmp1, b_scales_vec, vl / 4);
-                    }
-
-                    const int64_t A1 = *(const int64_t *)&a_ptr[l].qs[8];
-                    const int64_t A5 = *(const int64_t *)&a_ptr[l].qs[40];
-                    const int64_t A9 = *(const int64_t *)&a_ptr[l].qs[72];
-                    const int64_t Ad = *(const int64_t *)&a_ptr[l].qs[104];
-                    __asm__ __volatile__("" ::: "memory"); // prevent gcc from emitting fused vlse64, violating alignment
-                    vint16m4_t sumi_l1;
-                    {
-                        const vint8m2_t lhs_0_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A1, vl / 4));
-                        const vint8m2_t lhs_1_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A5, vl / 4));
-                        const vint8m2_t lhs_2_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A9, vl / 4));
-                        const vint8m2_t lhs_3_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(Ad, vl / 4));
-                        const vint16m4_t sumi_lo_0 = __riscv_vwmul_vv_i16m4(rhs_vec_lo_0, lhs_0_8, vl * 2);
-                        const vint16m4_t sumi_lo_1 = __riscv_vwmacc_vv_i16m4(sumi_lo_0, rhs_vec_lo_1, lhs_1_8, vl * 2);
-                        const vint16m4_t sumi_hi_0 = __riscv_vwmacc_vv_i16m4(sumi_lo_1, rhs_vec_hi_0, lhs_2_8, vl * 2);
-                        const vint16m4_t sumi_hi_m = __riscv_vwmacc_vv_i16m4(sumi_hi_0, rhs_vec_hi_1, lhs_3_8, vl * 2);
-
-                        sumi_l1 = sumi_hi_m;
-                    }
-
-                    {
-                        const vuint32m4_t sumi_i32 = __riscv_vreinterpret_v_i32m4_u32m4(__riscv_vreinterpret_v_i16m4_i32m4(sumi_l1));
-                        const vuint16m2_t sumi_h2_0 = __riscv_vnsrl_wx_u16m2(sumi_i32, 0, vl);
-                        const vuint16m2_t sumi_h2_1 = __riscv_vnsrl_wx_u16m2(sumi_i32, 16, vl);
-                        const vuint16m2_t sumi_h2 = __riscv_vadd_vv_u16m2(sumi_h2_0, sumi_h2_1, vl);
-                        const vuint32m2_t sumi_h2_i32 = __riscv_vreinterpret_v_u16m2_u32m2(sumi_h2);
-                        const vuint16m1_t sumi_h4_0 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 0, vl / 2);
-                        const vuint16m1_t sumi_h4_1 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 16, vl / 2);
-                        const vuint16m1_t sumi_h4 = __riscv_vadd_vv_u16m1(sumi_h4_0, sumi_h4_1, vl / 2);
-                        const vuint32m1_t sumi_h4_i32 = __riscv_vreinterpret_v_u16m1_u32m1(sumi_h4);
-                        const vint16mf2_t sumi_h8_0 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 0, vl / 4));
-                        const vint16mf2_t sumi_h8_1 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 16, vl / 4));
-                        const vint32m1_t sumi_h8 = __riscv_vwadd_vv_i32m1(sumi_h8_0, sumi_h8_1, vl / 4);
-                        const vfloat32m1_t facc = __riscv_vfcvt_f_x_v_f32m1(sumi_h8, vl / 4);
-
-                        const vfloat32m1_t tmp1 = __riscv_vfmul_vf_f32m1(facc, a_scales[1], vl / 4);
-                        sumf1 = __riscv_vfmacc_vv_f32m1(sumf1, tmp1, b_scales_vec, vl / 4);
-                    }
-
-                    const int64_t A2 = *(const int64_t *)&a_ptr[l].qs[16];
-                    const int64_t A6 = *(const int64_t *)&a_ptr[l].qs[48];
-                    const int64_t Aa = *(const int64_t *)&a_ptr[l].qs[80];
-                    const int64_t Ae = *(const int64_t *)&a_ptr[l].qs[112];
-                    __asm__ __volatile__("" ::: "memory"); // prevent gcc from emitting fused vlse64, violating alignment
-                    vint16m4_t sumi_l2;
-                    {
-                        const vint8m2_t lhs_0_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A2, vl / 4));
-                        const vint8m2_t lhs_1_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A6, vl / 4));
-                        const vint8m2_t lhs_2_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(Aa, vl / 4));
-                        const vint8m2_t lhs_3_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(Ae, vl / 4));
-                        const vint16m4_t sumi_lo_0 = __riscv_vwmul_vv_i16m4(rhs_vec_lo_0, lhs_0_8, vl * 2);
-                        const vint16m4_t sumi_lo_1 = __riscv_vwmacc_vv_i16m4(sumi_lo_0, rhs_vec_lo_1, lhs_1_8, vl * 2);
-                        const vint16m4_t sumi_hi_0 = __riscv_vwmacc_vv_i16m4(sumi_lo_1, rhs_vec_hi_0, lhs_2_8, vl * 2);
-                        const vint16m4_t sumi_hi_m = __riscv_vwmacc_vv_i16m4(sumi_hi_0, rhs_vec_hi_1, lhs_3_8, vl * 2);
-
-                        sumi_l2 = sumi_hi_m;
-                    }
-
-                    {
-                        const vuint32m4_t sumi_i32 = __riscv_vreinterpret_v_i32m4_u32m4(__riscv_vreinterpret_v_i16m4_i32m4(sumi_l2));
-                        const vuint16m2_t sumi_h2_0 = __riscv_vnsrl_wx_u16m2(sumi_i32, 0, vl);
-                        const vuint16m2_t sumi_h2_1 = __riscv_vnsrl_wx_u16m2(sumi_i32, 16, vl);
-                        const vuint16m2_t sumi_h2 = __riscv_vadd_vv_u16m2(sumi_h2_0, sumi_h2_1, vl);
-                        const vuint32m2_t sumi_h2_i32 = __riscv_vreinterpret_v_u16m2_u32m2(sumi_h2);
-                        const vuint16m1_t sumi_h4_0 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 0, vl / 2);
-                        const vuint16m1_t sumi_h4_1 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 16, vl / 2);
-                        const vuint16m1_t sumi_h4 = __riscv_vadd_vv_u16m1(sumi_h4_0, sumi_h4_1, vl / 2);
-                        const vuint32m1_t sumi_h4_i32 = __riscv_vreinterpret_v_u16m1_u32m1(sumi_h4);
-                        const vint16mf2_t sumi_h8_0 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 0, vl / 4));
-                        const vint16mf2_t sumi_h8_1 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 16, vl / 4));
-                        const vint32m1_t sumi_h8 = __riscv_vwadd_vv_i32m1(sumi_h8_0, sumi_h8_1, vl / 4);
-                        const vfloat32m1_t facc = __riscv_vfcvt_f_x_v_f32m1(sumi_h8, vl / 4);
-
-                        const vfloat32m1_t tmp1 = __riscv_vfmul_vf_f32m1(facc, a_scales[2], vl / 4);
-                        sumf2 = __riscv_vfmacc_vv_f32m1(sumf2, tmp1, b_scales_vec, vl / 4);
-                    }
-
-                    const int64_t A3 = *(const int64_t *)&a_ptr[l].qs[24];
-                    const int64_t A7 = *(const int64_t *)&a_ptr[l].qs[56];
-                    const int64_t Ab = *(const int64_t *)&a_ptr[l].qs[88];
-                    const int64_t Af = *(const int64_t *)&a_ptr[l].qs[120];
-                    __asm__ __volatile__("" ::: "memory"); // prevent gcc from emitting fused vlse64, violating alignment
-                    vint16m4_t sumi_l3;
-                    {
-                        const vint8m2_t lhs_0_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A3, vl / 4));
-                        const vint8m2_t lhs_1_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A7, vl / 4));
-                        const vint8m2_t lhs_2_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(Ab, vl / 4));
-                        const vint8m2_t lhs_3_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(Af, vl / 4));
-                        const vint16m4_t sumi_lo_0 = __riscv_vwmul_vv_i16m4(rhs_vec_lo_0, lhs_0_8, vl * 2);
-                        const vint16m4_t sumi_lo_1 = __riscv_vwmacc_vv_i16m4(sumi_lo_0, rhs_vec_lo_1, lhs_1_8, vl * 2);
-                        const vint16m4_t sumi_hi_0 = __riscv_vwmacc_vv_i16m4(sumi_lo_1, rhs_vec_hi_0, lhs_2_8, vl * 2);
-                        const vint16m4_t sumi_hi_m = __riscv_vwmacc_vv_i16m4(sumi_hi_0, rhs_vec_hi_1, lhs_3_8, vl * 2);
-
-                        sumi_l3 = sumi_hi_m;
-                    }
-
-                    {
-                        const vuint32m4_t sumi_i32 = __riscv_vreinterpret_v_i32m4_u32m4(__riscv_vreinterpret_v_i16m4_i32m4(sumi_l3));
-                        const vuint16m2_t sumi_h2_0 = __riscv_vnsrl_wx_u16m2(sumi_i32, 0, vl);
-                        const vuint16m2_t sumi_h2_1 = __riscv_vnsrl_wx_u16m2(sumi_i32, 16, vl);
-                        const vuint16m2_t sumi_h2 = __riscv_vadd_vv_u16m2(sumi_h2_0, sumi_h2_1, vl);
-                        const vuint32m2_t sumi_h2_i32 = __riscv_vreinterpret_v_u16m2_u32m2(sumi_h2);
-                        const vuint16m1_t sumi_h4_0 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 0, vl / 2);
-                        const vuint16m1_t sumi_h4_1 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 16, vl / 2);
-                        const vuint16m1_t sumi_h4 = __riscv_vadd_vv_u16m1(sumi_h4_0, sumi_h4_1, vl / 2);
-                        const vuint32m1_t sumi_h4_i32 = __riscv_vreinterpret_v_u16m1_u32m1(sumi_h4);
-                        const vint16mf2_t sumi_h8_0 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 0, vl / 4));
-                        const vint16mf2_t sumi_h8_1 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 16, vl / 4));
-                        const vint32m1_t sumi_h8 = __riscv_vwadd_vv_i32m1(sumi_h8_0, sumi_h8_1, vl / 4);
-                        const vfloat32m1_t facc = __riscv_vfcvt_f_x_v_f32m1(sumi_h8, vl / 4);
-
-                        const vfloat32m1_t tmp1 = __riscv_vfmul_vf_f32m1(facc, a_scales[3], vl / 4);
-                        sumf3 = __riscv_vfmacc_vv_f32m1(sumf3, tmp1, b_scales_vec, vl / 4);
-                    }
-                }
-                __riscv_vse32_v_f32m1(&s[(y * 4 + 0) * bs + x * ncols_interleaved], sumf0, vl / 4);
-                __riscv_vse32_v_f32m1(&s[(y * 4 + 1) * bs + x * ncols_interleaved], sumf1, vl / 4);
-                __riscv_vse32_v_f32m1(&s[(y * 4 + 2) * bs + x * ncols_interleaved], sumf2, vl / 4);
-                __riscv_vse32_v_f32m1(&s[(y * 4 + 3) * bs + x * ncols_interleaved], sumf3, vl / 4);
-            }
-        }
-
-        return;
-    }
 #endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__)
     float sumf[4][8];
     int sumi;
@@ -4006,7 +1756,7 @@ static void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, c
                                 sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
                                          (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4;
                             }
-                            sumf[m][j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d[m]);
+                            sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m]);
                         }
                     }
                 }
@@ -4019,7 +1769,7 @@ 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) {
+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;
@@ -5510,7 +3260,7 @@ static void ggml_gemm_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, c
                                 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];
+                            sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d[m];
                         }
                     }
                 }
@@ -5519,7 +3269,7 @@ static void ggml_gemm_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, c
                     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];
+                            sum_minf[m][j] += mins[j] * (bsums[0] + bsums[1]) * GGML_CPU_FP16_TO_FP32(b_ptr[l].dmin[j]) * a_ptr[l].d[m];
                         }
                     }
                 }
@@ -5533,899 +3283,3 @@ static void ggml_gemm_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, c
     }
 #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;
-    const int ncols_interleaved = 4;
-    const int blocklen = 4;
-
-    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(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD)
-    if (ggml_cpu_has_neon() && ggml_cpu_has_dotprod()) {
-        const int8x16_t kvalues = vld1q_s8(kvalues_iq4nl);
-
-        for (int y = 0; y < nr / 4; y++) {
-            const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
-            for (int x = 0; x < nc / ncols_interleaved; x++) {
-                const block_iq4_nlx4 * b_ptr = (const block_iq4_nlx4 *) vx + (x * nb);
-
-                float32x4_t sumf[4];
-                for (int m = 0; m < 4; m++) {
-                    sumf[m] = vdupq_n_f32(0);
-                }
-
-                for (int l = 0; l < nb; l++) {
-                    float32x4_t a_d = vcvt_f32_f16(vld1_f16((const float16_t *)a_ptr[l].d));
-                    float32x4_t b_d = vcvt_f32_f16(vld1_f16((const float16_t *)b_ptr[l].d));
-
-                    int32x4_t sumi_0 = vdupq_n_s32(0);
-                    int32x4_t sumi_1 = vdupq_n_s32(0);
-                    int32x4_t sumi_2 = vdupq_n_s32(0);
-                    int32x4_t sumi_3 = vdupq_n_s32(0);
-
-                    for (int k = 0; k < 4; k++) {
-                        int8x16_t a_0 = vld1q_s8(a_ptr[l].qs + 16 * k + 0);
-                        int8x16_t a_1 = vld1q_s8(a_ptr[l].qs + 16 * k + 64);
-
-                        uint8x16_t b = vld1q_u8(b_ptr[l].qs + 16 * k);
-                        int8x16_t b_hi = vqtbl1q_s8(kvalues, b >> 4);
-                        int8x16_t b_lo = vqtbl1q_s8(kvalues, b & 0xF);
-
-                        sumi_0 = vdotq_laneq_s32(sumi_0, b_lo, a_0, 0);
-                        sumi_1 = vdotq_laneq_s32(sumi_1, b_lo, a_0, 1);
-                        sumi_2 = vdotq_laneq_s32(sumi_2, b_lo, a_0, 2);
-                        sumi_3 = vdotq_laneq_s32(sumi_3, b_lo, a_0, 3);
-                        sumi_0 = vdotq_laneq_s32(sumi_0, b_hi, a_1, 0);
-                        sumi_1 = vdotq_laneq_s32(sumi_1, b_hi, a_1, 1);
-                        sumi_2 = vdotq_laneq_s32(sumi_2, b_hi, a_1, 2);
-                        sumi_3 = vdotq_laneq_s32(sumi_3, b_hi, a_1, 3);
-                    }
-
-                    sumf[0] = vmlaq_f32(sumf[0], vmulq_laneq_f32(b_d, a_d, 0), vcvtq_f32_s32(sumi_0));
-                    sumf[1] = vmlaq_f32(sumf[1], vmulq_laneq_f32(b_d, a_d, 1), vcvtq_f32_s32(sumi_1));
-                    sumf[2] = vmlaq_f32(sumf[2], vmulq_laneq_f32(b_d, a_d, 2), vcvtq_f32_s32(sumi_2));
-                    sumf[3] = vmlaq_f32(sumf[3], vmulq_laneq_f32(b_d, a_d, 3), vcvtq_f32_s32(sumi_3));
-                }
-
-                for (int m = 0; m < 4; m++) {
-                    vst1q_f32(s + (y * 4 + m) * bs + x * 4, sumf[m]);
-                }
-            }
-        }
-        return;
-    }
-#endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON)
-    {
-        float sumf[4][4];
-        int sumi;
-
-        for (int y = 0; y < nr / 4; y++) {
-            const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
-            for (int x = 0; x < nc / ncols_interleaved; x++) {
-                const block_iq4_nlx4 * b_ptr = (const block_iq4_nlx4 *) vx + (x * nb);
-                for (int m = 0; m < 4; m++) {
-                    for (int j = 0; j < ncols_interleaved; j++) sumf[m][j] = 0.0;
-                }
-                for (int l = 0; l < nb; l++) {
-                    for (int k = 0; k < (qk / (2 * blocklen)); k++) {
-                        for (int m = 0; m < 4; m++) {
-                            for (int j = 0; j < ncols_interleaved; j++) {
-                                sumi = 0;
-                                for (int i = 0; i < blocklen; ++i) {
-                                    const int v0 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0x0F];
-                                    const int v1 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4];
-                                    sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
-                                            (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4]));
-                                }
-                                sumf[m][j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(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];
-                }
-            }
-        }
-    }
-}
-
-static block_q4_0x4 make_block_q4_0x4(block_q4_0 * in, unsigned int blck_size_interleave) {
-    block_q4_0x4 out;
-
-    for (int i = 0; i < 4; i++) {
-        out.d[i] = in[i].d;
-    }
-
-    const int end = QK4_0 * 2 / blck_size_interleave;
-
-    if (blck_size_interleave == 8) {
-        const uint64_t xor_mask = 0x8888888888888888ULL;
-        for (int i = 0; i < end; ++i) {
-            int src_id = i % 4;
-            int src_offset = (i / 4) * blck_size_interleave;
-            int dst_offset = i * blck_size_interleave;
-
-            uint64_t elems;
-            // Using memcpy to avoid unaligned memory accesses
-            memcpy(&elems, &in[src_id].qs[src_offset], sizeof(uint64_t));
-            elems ^= xor_mask;
-            memcpy(&out.qs[dst_offset], &elems, sizeof(uint64_t));
-        }
-    } else if (blck_size_interleave == 4) {
-        const uint32_t xor_mask = 0x88888888;
-        for (int i = 0; i < end; ++i) {
-            int src_id = i % 4;
-            int src_offset = (i / 4) * blck_size_interleave;
-            int dst_offset = i * blck_size_interleave;
-
-            uint32_t elems;
-            memcpy(&elems, &in[src_id].qs[src_offset], sizeof(uint32_t));
-            elems ^= xor_mask;
-            memcpy(&out.qs[dst_offset], &elems, sizeof(uint32_t));
-        }
-    } else {
-        GGML_ASSERT(false);
-    }
-
-    return out;
-}
-
-// interleave 8 block_q4_0s in blocks of blck_size_interleave
-// returns an interleaved block_q4_0x8
-// in the interleaved block_q4_0x8, place deltas for 8 block_q4_0 blocks
-// first, then interleave quants from 8 block_q4_0s in blocks of blck_size_interleave
-static block_q4_0x8 make_block_q4_0x8(block_q4_0 * in, unsigned int blck_size_interleave) {
-    block_q4_0x8 out;
-
-    for (int i = 0; i < 8; i++) {
-        out.d[i] = in[i].d;
-    }
-
-    const int end = QK4_0 * 4 / blck_size_interleave;
-    const uint64_t xor_mask = 0x8888888888888888ULL;
-
-    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));
-        elems ^= xor_mask;
-        memcpy(&out.qs[dst_offset], &elems, sizeof(uint64_t));
-    }
-
-    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);
-    constexpr int nrows_interleaved = 4;
-
-    block_q4_0x4 * dst = (block_q4_0x4 *)t->data;
-    const block_q4_0 * src = (const block_q4_0 *)data;
-    block_q4_0 dst_tmp[4];
-    int nrow = ggml_nrows(t);
-    int nblocks = t->ne[0] / QK4_0;
-
-    GGML_ASSERT(data_size == nrow * nblocks * sizeof(block_q4_0));
-
-    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_0x4(dst_tmp, interleave_block);
-        }
-        src += nrows_interleaved * nblocks;
-    }
-    return 0;
-
-    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);
-    GGML_ASSERT(interleave_block == 8);
-    constexpr int nrows_interleaved = 8;
-
-    block_q4_0x8 * dst = (block_q4_0x8*)t->data;
-    const block_q4_0 * src = (const block_q4_0*) data;
-    block_q4_0 dst_tmp[8];
-    int nrow = ggml_nrows(t);
-    int nblocks = t->ne[0] / QK4_0;
-
-    GGML_ASSERT(data_size == nrow * nblocks * sizeof(block_q4_0));
-
-    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_0x8(dst_tmp, interleave_block);
-        }
-        src += nrows_interleaved * nblocks;
-    }
-    return 0;
-
-    GGML_UNUSED(data_size);
-}
-
-static block_iq4_nlx4 make_block_iq4_nlx4(block_iq4_nl * in, unsigned int blck_size_interleave) {
-    block_iq4_nlx4 out;
-
-    for (int i = 0; i < 4; i++) {
-        out.d[i] = in[i].d;
-    }
-
-    const int end = QK4_NL * 2 / blck_size_interleave;
-
-    // TODO: this branch seems wrong
-    //if (blck_size_interleave == 8) {
-    //    for (int i = 0; i < end; ++i) {
-    //        int src_id = i % 4;
-    //        int src_offset = (i / 4) * blck_size_interleave;
-    //        int dst_offset = i * blck_size_interleave;
-
-    //        // Using memcpy to avoid unaligned memory accesses
-    //        memcpy(&out.qs[dst_offset], &in[src_id].qs[src_offset], sizeof(uint64_t));
-    //    }
-    //} else
-    if (blck_size_interleave == 4) {
-        for (int i = 0; i < end; ++i) {
-            int src_id = i % 4;
-            int src_offset = (i / 4) * blck_size_interleave;
-            int dst_offset = i * blck_size_interleave;
-
-            memcpy(&out.qs[dst_offset], &in[src_id].qs[src_offset], sizeof(uint32_t));
-        }
-    } else {
-        GGML_ASSERT(false);
-    }
-
-    return out;
-}
-
-static int repack_iq4_nl_to_iq4_nl_4_bl(struct ggml_tensor * t, int interleave_block, const void * GGML_RESTRICT data, size_t data_size) {
-    GGML_ASSERT(t->type == GGML_TYPE_IQ4_NL);
-    //GGML_ASSERT(interleave_block == 4 || interleave_block == 8);
-    GGML_ASSERT(interleave_block == 4);
-
-    block_iq4_nlx4 * dst = (block_iq4_nlx4 *)t->data;
-    const block_iq4_nl * src = (const block_iq4_nl *)data;
-    block_iq4_nl dst_tmp[4];
-    int nrow = ggml_nrows(t);
-    int nrows_interleaved = 4;
-    int nblocks = t->ne[0] / QK4_0;
-
-    GGML_ASSERT(data_size == nrow * nblocks * sizeof(block_iq4_nl));
-
-    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_iq4_nlx4(dst_tmp, interleave_block);
-        }
-        src += nrows_interleaved * nblocks;
-    }
-    return 0;
-
-    GGML_UNUSED(data_size);
-}
-
-namespace ggml::cpu::aarch64 {
-// repack
-template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS>
-int repack(struct ggml_tensor *, const void *, size_t);
-
-// TODO: generalise.
-template <> int repack<block_q4_0, 4, 4>(struct ggml_tensor * t, const void * data, size_t data_size) {
-    return repack_q4_0_to_q4_0_4_bl(t, 4, data, data_size);
-}
-
-template <> int repack<block_q4_0, 8, 4>(struct ggml_tensor * t, const void * data, size_t data_size) {
-    return repack_q4_0_to_q4_0_4_bl(t, 8, data, data_size);
-}
-
-template <> int repack<block_q4_0, 8, 8>(struct ggml_tensor * t, const void * data, size_t data_size) {
-    return repack_q4_0_to_q4_0_8_bl(t, 8, data, data_size);
-}
-
-template <> int repack<block_q4_K, 8, 8>(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<block_iq4_nl, 4, 4>(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);
-}
-
-// TODO: needs to be revisited
-//template <> int repack<block_iq4_nl, 8, 4>(struct ggml_tensor * t, const void * data, size_t data_size) {
-//    return repack_iq4_nl_to_iq4_nl_4_bl(t, 8, data, data_size);
-//}
-
-// gemv
-template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS, ggml_type PARAM_TYPE>
-void gemv(int, float *, size_t, const void *, const void *, int, int);
-
-template <> void gemv<block_q4_0, 4, 4, GGML_TYPE_Q8_0>(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<block_q4_0, 8, 4, GGML_TYPE_Q8_0>(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<block_q4_0, 8, 8, GGML_TYPE_Q8_0>(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<block_q4_K, 8, 8, GGML_TYPE_Q8_K>(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<block_iq4_nl, 4, 4, GGML_TYPE_Q8_0>(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 <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS, ggml_type PARAM_TYPE>
-void gemm(int, float *, size_t, const void *, const void *, int, int);
-
-template <> void gemm<block_q4_0, 4, 4, GGML_TYPE_Q8_0>(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<block_q4_0, 8, 4, GGML_TYPE_Q8_0>(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<block_q4_0, 8, 8, GGML_TYPE_Q8_0>(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<block_q4_K, 8, 8, GGML_TYPE_Q8_K>(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<block_iq4_nl, 4, 4, GGML_TYPE_Q8_0>(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);
-}
-
-class tensor_traits_base : public ggml::cpu::tensor_traits {
-  public:
-    virtual int repack(struct ggml_tensor * t, const void * data, size_t data_size) = 0;
-};
-
-template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS, ggml_type PARAM_TYPE> 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(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;
-        }
-        return false;
-    }
-
-    void forward_mul_mat(ggml_compute_params * params, ggml_tensor * op) {
-        const ggml_tensor * src0 = op->src[0];
-        const ggml_tensor * src1 = op->src[1];
-        ggml_tensor *       dst  = op;
-
-        GGML_TENSOR_BINARY_OP_LOCALS
-
-        const int ith = params->ith;
-        const int nth = params->nth;
-
-        GGML_ASSERT(ne0 == ne01);
-        GGML_ASSERT(ne1 == ne11);
-        GGML_ASSERT(ne2 == ne12);
-        GGML_ASSERT(ne3 == ne13);
-
-        // dst cannot be transposed or permuted
-        GGML_ASSERT(nb0 == sizeof(float));
-        GGML_ASSERT(nb0 <= nb1);
-        GGML_ASSERT(nb1 <= nb2);
-        GGML_ASSERT(nb2 <= nb3);
-
-        GGML_ASSERT(src1->type == GGML_TYPE_F32);
-
-        GGML_ASSERT(ggml_n_dims(op->src[0]) == 2);
-        // GGML_ASSERT(ggml_n_dims(op->src[1]) == 2);
-
-        char *       wdata = static_cast<char *>(params->wdata);
-        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(PARAM_TYPE)->from_float;
-
-        int64_t i11_processed = 0;
-        for (int64_t i11 = ith * 4; i11 < ne11 - ne11 % 4; i11 += nth * 4) {
-            ggml_quantize_mat_t<INTER_SIZE, PARAM_TYPE>((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);
-        }
-
-        ggml_barrier(params->threadpool);
-
-        const void * src1_wdata      = params->wdata;
-        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;
-        if (src0_start >= src0_end) {
-            return;
-        }
-
-        // If there are more than three rows in src1, use gemm; otherwise, use gemv.
-        if (ne11 > 3) {
-            gemm<BLOC_TYPE, INTER_SIZE, NB_COLS, PARAM_TYPE>(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<BLOC_TYPE, INTER_SIZE, NB_COLS, PARAM_TYPE>(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);
-        }
-    }
-
-    void forward_mul_mat_id(ggml_compute_params * params, ggml_tensor * op) {
-        const ggml_tensor * src0 = op->src[0];
-        const ggml_tensor * src1 = op->src[1];
-        const ggml_tensor * ids  = op->src[2];
-        ggml_tensor *       dst  = op;
-
-        GGML_TENSOR_BINARY_OP_LOCALS
-
-        const int ith = params->ith;
-        const int nth = params->nth;
-
-        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));
-        GGML_ASSERT(nb10 == ggml_type_size(src1->type));
-
-        // dst cannot be transposed or permuted
-        GGML_ASSERT(nb0 == sizeof(float));
-        GGML_ASSERT(nb0 <= nb1);
-        GGML_ASSERT(nb1 <= nb2);
-        GGML_ASSERT(nb2 <= nb3);
-
-        GGML_ASSERT(ne03 == 1);
-        GGML_ASSERT(ne13 == 1);
-        GGML_ASSERT(ne3  == 1);
-
-        GGML_ASSERT(src1->type == GGML_TYPE_F32);
-
-        // row groups
-        const int n_ids = ids->ne[0]; // n_expert_used
-        const int n_as  = ne02;       // n_expert
-
-        const size_t nbw1 = ggml_row_size(PARAM_TYPE, ne10);
-        const size_t nbw2 = nbw1*ne11;
-        const size_t nbw3 = nbw2*ne12;
-
-        struct mmid_row_mapping {
-            int32_t i1;
-            int32_t i2;
-        };
-
-        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));
-        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 => 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),
-                           (void *)               (wdata + i12 * nbw2 + i11 * nbw1),
-                           ne10);
-            }
-        }
-
-#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));
-
-            // group rows by src0 matrix
-            for (int32_t iid1 = 0; iid1 < ids->ne[1]; ++iid1) {
-                for (int32_t id = 0; id < n_ids; ++id) {
-                    const int32_t i02 =
-                        *(const int32_t *) ((const char *) ids->data + iid1 * ids->nb[1] + id * ids->nb[0]);
-
-                    GGML_ASSERT(i02 >= 0 && i02 < n_as);
-
-                    MMID_MATRIX_ROW(i02, matrix_row_counts[i02]) = { id, iid1 };
-                    matrix_row_counts[i02] += 1;
-                }
-            }
-        }
-
-        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];
-
-            if (cne1 == 0) {
-                continue;
-            }
-
-            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;
-            }
-
-            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 int64_t i1 = id;  // selected expert index
-                const int64_t i2 = i12; // row
-
-                const auto * src1_col = (const char *) wdata + (i11 * nbw1 + i12 * nbw2);
-
-                gemv<BLOC_TYPE, INTER_SIZE, NB_COLS, PARAM_TYPE>(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);
-            }
-        }
-#undef MMID_MATRIX_ROW
-    }
-
-    int repack(struct ggml_tensor * t, const void * data, size_t data_size) override {
-        GGML_LOG_DEBUG("%s: repack tensor %s with %s_%dx%d\n", __func__, t->name, ggml_type_name(t->type),
-                       (int) NB_COLS, (int) INTER_SIZE);
-        return ggml::cpu::aarch64::repack<BLOC_TYPE, INTER_SIZE, NB_COLS>(t, data, data_size);
-    }
-};
-
-// instance for Q4
-static const tensor_traits<block_q4_0, 4, 4, GGML_TYPE_Q8_0> q4_0_4x4_q8_0;
-static const tensor_traits<block_q4_0, 8, 4, GGML_TYPE_Q8_0> q4_0_4x8_q8_0;
-static const tensor_traits<block_q4_0, 8, 8, GGML_TYPE_Q8_0> q4_0_8x8_q8_0;
-static const tensor_traits<block_q4_K, 8, 8, GGML_TYPE_Q8_K> q4_K_8x8_q8_K;
-
-// instance for IQ4
-static const tensor_traits<block_iq4_nl, 4, 4, GGML_TYPE_Q8_0> iq4_nl_4x4_q8_0;
-
-}  // namespace ggml::cpu::aarch64
-
-static const ggml::cpu::tensor_traits * ggml_aarch64_get_optimal_repack_type(const struct ggml_tensor * cur) {
-    if (cur->type == GGML_TYPE_Q4_0) {
-        if (ggml_cpu_has_avx2() || (ggml_cpu_has_sve() && ggml_cpu_has_matmul_int8() && ggml_cpu_get_sve_cnt() == QK8_0)) {
-            if (cur->ne[1] % 8 == 0) {
-                return &ggml::cpu::aarch64::q4_0_8x8_q8_0;
-            }
-        }
-        if (ggml_cpu_has_neon() && ggml_cpu_has_matmul_int8()) {
-            if (cur->ne[1] % 4 == 0) {
-                return &ggml::cpu::aarch64::q4_0_4x8_q8_0;
-            }
-        }
-        if (ggml_cpu_has_neon() && ggml_cpu_has_dotprod()) {
-            if (cur->ne[1] % 4 == 0) {
-                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) {
-                return &ggml::cpu::aarch64::iq4_nl_4x4_q8_0;
-            }
-        }
-    }
-
-    return nullptr;
-}
-
-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::cpu::tensor_traits *>(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,
-                                                       const void * data, size_t offset, size_t size) {
-    GGML_ASSERT(offset == 0);
-    GGML_ASSERT(size == ggml_nbytes(tensor));
-
-    auto tensor_traits = (ggml::cpu::aarch64::tensor_traits_base *) tensor->extra;
-    auto OK            = tensor_traits->repack(tensor, data, size);
-
-    GGML_ASSERT(OK == 0);
-    GGML_UNUSED(buffer);
-}
-
-static const char * ggml_backend_cpu_aarch64_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
-    return "CPU_AARCH64";
-
-    GGML_UNUSED(buft);
-}
-
-static ggml_backend_buffer_t ggml_backend_cpu_aarch64_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_aarch64_buffer_init_tensor;
-    buffer->iface.set_tensor  = ggml_backend_cpu_aarch64_buffer_set_tensor;
-    buffer->iface.get_tensor  = nullptr;
-    buffer->iface.cpy_tensor  = nullptr;
-    return buffer;
-}
-
-static size_t ggml_backend_cpu_aarch64_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
-    return TENSOR_ALIGNMENT;
-
-    GGML_UNUSED(buft);
-}
-
-namespace ggml::cpu::aarch64 {
-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]->buffer &&
-                (ggml_n_dims(op->src[0]) == 2) &&
-                op->src[0]->buffer->buft == ggml_backend_cpu_aarch64_buffer_type() &&
-                ggml_aarch64_get_optimal_repack_type(op->src[0])
-                ) {
-            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) {
-                return true;
-            }
-            //if (op->src[1]->type == GGML_TYPE_Q8_0) {
-            //    return true;
-            //}
-            // may be possible if Q8_0 packed...
-        } else if (op->op == GGML_OP_MUL_MAT_ID
-                && op->src[0]->buffer
-                && (ggml_n_dims(op->src[0]) == 3)
-                && op->src[0]->buffer->buft == ggml_backend_cpu_aarch64_buffer_type()
-                && ggml_aarch64_get_optimal_repack_type(op->src[0])
-                ) {
-            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) {
-                return true;
-            }
-            //if (op->src[1]->type == GGML_TYPE_Q8_0) {
-            //    return true;
-            //}
-        }
-        return false;
-    }
-
-    ggml::cpu::tensor_traits * get_tensor_traits(const struct ggml_tensor * op) override {
-        if (op->op == GGML_OP_MUL_MAT || op->op == GGML_OP_MUL_MAT_ID) {
-            if (op->src[0]->buffer && op->src[0]->buffer->buft == ggml_backend_cpu_aarch64_buffer_type()) {
-                return (ggml::cpu::tensor_traits *) op->src[0]->extra;
-            }
-        }
-        return nullptr;
-    }
-};
-}  // namespace ggml::cpu::aarch64
-
-ggml_backend_buffer_type_t ggml_backend_cpu_aarch64_buffer_type(void) {
-    static struct ggml_backend_buffer_type ggml_backend_cpu_buffer_type_aarch64 = {
-        /* .iface    = */ {
-                           /* .get_name         = */ ggml_backend_cpu_aarch64_buffer_type_get_name,
-                           /* .alloc_buffer     = */ ggml_backend_cpu_aarch64_buffer_type_alloc_buffer,
-                           /* .get_alignment    = */ ggml_backend_cpu_aarch64_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 = */ new ggml::cpu::aarch64::extra_buffer_type(),
-    };
-
-    return &ggml_backend_cpu_buffer_type_aarch64;
-}
diff --git a/ggml/src/ggml-cpu/common.h b/ggml/src/ggml-cpu/common.h
index 3df01c1ed..353563dc3 100644
--- a/ggml/src/ggml-cpu/common.h
+++ b/ggml/src/ggml-cpu/common.h
@@ -1,9 +1,10 @@
 #pragma once
 
 #include "ggml.h"
-#include "ggml-cpu-traits.h"
+#include "traits.h"
 #include "ggml-cpu-impl.h"
 #include "ggml-impl.h"
+#include "simd-mappings.h"
 
 #ifdef __cplusplus
 
@@ -12,11 +13,11 @@
 // 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);
+    return GGML_CPU_FP32_TO_FP16(x);
 }
 
 static inline float f16_to_f32(ggml_fp16_t x) {
-    return GGML_FP16_TO_FP32(x);
+    return GGML_CPU_FP16_TO_FP32(x);
 }
 
 static inline ggml_bf16_t f32_to_bf16(float x) {
diff --git a/ggml/src/ggml-cpu/ggml-cpu-aarch64.h b/ggml/src/ggml-cpu/ggml-cpu-aarch64.h
deleted file mode 100644
index 6e84c826b..000000000
--- a/ggml/src/ggml-cpu/ggml-cpu-aarch64.h
+++ /dev/null
@@ -1,8 +0,0 @@
-#pragma once
-
-#include "ggml-cpu-traits.h"
-#include "ggml.h"
-
-// GGML internal header
-
-ggml_backend_buffer_type_t ggml_backend_cpu_aarch64_buffer_type(void);
diff --git a/ggml/src/ggml-cpu/ggml-cpu-impl.h b/ggml/src/ggml-cpu/ggml-cpu-impl.h
index b3f1b5ca7..d839cf5c5 100644
--- a/ggml/src/ggml-cpu/ggml-cpu-impl.h
+++ b/ggml/src/ggml-cpu/ggml-cpu-impl.h
@@ -62,11 +62,17 @@ struct ggml_compute_params {
 #if defined(__s390x__) && defined(__VEC__)
 #ifndef __VXE__
 #define __VXE__
-#endif
+#endif  // __VXE__
 #ifndef __VXE2__
 #define __VXE2__
-#endif
-#endif
+#endif  // __VXE2__
+#endif  // __s390x__ && __VEC__
+
+#if defined(__s390x__) && defined(GGML_NNPA)
+#ifndef __NNPA__
+#define __NNPA__
+#endif  // __NNPA__
+#endif  // __s390x__ && GGML_NNPA
 
 #if defined(__ARM_FEATURE_SVE)
 #include <sys/prctl.h>
@@ -371,7 +377,7 @@ inline static int32x4_t ggml_vdotq_s32(int32x4_t acc, int8x16_t a, int8x16_t b)
 #define vec_xor(a, b) ((a) ^ (b)) // Vector XOR
 #endif
 
-typedef signed char char8x16_t __attribute__((vector_size(16)));
+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)));
@@ -382,10 +388,10 @@ 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 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 signed   long long long64x2_t  __attribute__((vector_size(16)));
 typedef unsigned long long ulong64x2_t __attribute__((vector_size(16)));
 
 typedef struct ggml_uint8x16x2_t {
@@ -503,6 +509,9 @@ static __m256 __lasx_xvreplfr2vr_s(const float val) {
 // TODO: move to ggml-threading
 void ggml_barrier(struct ggml_threadpool * tp);
 
+void ggml_threadpool_chunk_set(struct ggml_threadpool * tp, int value);
+int  ggml_threadpool_chunk_add(struct ggml_threadpool * tp, int value);
+
 #ifdef __cplusplus
 }
 #endif
diff --git a/ggml/src/ggml-cpu/ggml-cpu-quants.c b/ggml/src/ggml-cpu/ggml-cpu-quants.c
deleted file mode 100644
index 40bded476..000000000
--- a/ggml/src/ggml-cpu/ggml-cpu-quants.c
+++ /dev/null
@@ -1,13891 +0,0 @@
-#define GGML_COMMON_IMPL_C
-#include "ggml-common.h"
-
-#include "ggml-quants.h"
-#include "ggml-cpu-quants.h"
-#include "ggml-impl.h"
-#include "ggml-cpu-impl.h"
-#include "ggml-cpu.h"
-
-#include <math.h>
-#include <string.h>
-#include <assert.h>
-#include <float.h>
-#include <stdlib.h> // for qsort
-#include <stdio.h>  // for GGML_ASSERT
-
-#define GROUP_MAX_EPS 1e-15f
-#define GROUP_MAX_EPS_IQ3_XXS 1e-8f
-#define GROUP_MAX_EPS_IQ2_S 1e-8f
-#define GROUP_MAX_EPS_IQ1_M 1e-7f
-#define GROUP_MAX_EPS_IQ1_S 1e-12f
-
-#define UNUSED GGML_UNUSED
-
-// some compilers don't provide _mm256_set_m128i, e.g. gcc 7
-#define MM256_SET_M128I(a, b) _mm256_insertf128_si256(_mm256_castsi128_si256(b), (a), 1)
-
-#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__) || defined(__SSSE3__)
-// 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
-    const __m128i ax = _mm_sign_epi8(x, x);
-    // Sign the values of the y vectors
-    const __m128i sy = _mm_sign_epi8(y, x);
-    // Perform multiplication and create 16-bit values
-    const __m128i dot = _mm_maddubs_epi16(ax, sy);
-    const __m128i ones = _mm_set1_epi16(1);
-    return _mm_madd_epi16(ones, dot);
-}
-
-#if __AVX__ || __AVX2__ || __AVX512F__
-// horizontally add 8 floats
-static inline float hsum_float_8(const __m256 x) {
-    __m128 res = _mm256_extractf128_ps(x, 1);
-    res = _mm_add_ps(res, _mm256_castps256_ps128(x));
-    res = _mm_add_ps(res, _mm_movehl_ps(res, res));
-    res = _mm_add_ss(res, _mm_movehdup_ps(res));
-    return _mm_cvtss_f32(res);
-}
-
-// horizontally add 8 int32_t
-static inline int hsum_i32_8(const __m256i a) {
-    const __m128i sum128 = _mm_add_epi32(_mm256_castsi256_si128(a), _mm256_extractf128_si256(a, 1));
-    const __m128i hi64 = _mm_unpackhi_epi64(sum128, sum128);
-    const __m128i sum64 = _mm_add_epi32(hi64, sum128);
-    const __m128i hi32  = _mm_shuffle_epi32(sum64, _MM_SHUFFLE(2, 3, 0, 1));
-    return _mm_cvtsi128_si32(_mm_add_epi32(sum64, hi32));
-}
-
-// horizontally add 4 int32_t
-static inline int hsum_i32_4(const __m128i a) {
-    const __m128i hi64 = _mm_unpackhi_epi64(a, a);
-    const __m128i sum64 = _mm_add_epi32(hi64, a);
-    const __m128i hi32  = _mm_shuffle_epi32(sum64, _MM_SHUFFLE(2, 3, 0, 1));
-    return _mm_cvtsi128_si32(_mm_add_epi32(sum64, hi32));
-}
-
-#if defined(__AVX2__) || defined(__AVX512F__)
-// spread 32 bits to 32 bytes { 0x00, 0xFF }
-static inline __m256i bytes_from_bits_32(const uint8_t * x) {
-    uint32_t x32;
-    memcpy(&x32, x, sizeof(uint32_t));
-    const __m256i shuf_mask = _mm256_set_epi64x(
-            0x0303030303030303, 0x0202020202020202,
-            0x0101010101010101, 0x0000000000000000);
-    __m256i bytes = _mm256_shuffle_epi8(_mm256_set1_epi32(x32), shuf_mask);
-    const __m256i bit_mask = _mm256_set1_epi64x(0x7fbfdfeff7fbfdfe);
-    bytes = _mm256_or_si256(bytes, bit_mask);
-    return _mm256_cmpeq_epi8(bytes, _mm256_set1_epi64x(-1));
-}
-
-// Unpack 32 4-bit fields into 32 bytes
-// The output vector contains 32 bytes, each one in [ 0 .. 15 ] interval
-static inline __m256i bytes_from_nibbles_32(const uint8_t * rsi)
-{
-    const __m128i tmp = _mm_loadu_si128((const __m128i *)rsi);
-    const __m256i bytes = MM256_SET_M128I(_mm_srli_epi16(tmp, 4), tmp);
-    const __m256i lowMask = _mm256_set1_epi8( 0xF );
-    return _mm256_and_si256(lowMask, bytes);
-}
-
-// add int16_t pairwise and return as float vector
-static inline __m256 sum_i16_pairs_float(const __m256i x) {
-    const __m256i ones = _mm256_set1_epi16(1);
-    const __m256i summed_pairs = _mm256_madd_epi16(ones, x);
-    return _mm256_cvtepi32_ps(summed_pairs);
-}
-
-static inline __m256 mul_sum_us8_pairs_float(const __m256i ax, const __m256i sy) {
-#if defined(__AVX512VNNI__) && defined(__AVX512VL__)
-    const __m256i zero = _mm256_setzero_si256();
-    const __m256i summed_pairs = _mm256_dpbusd_epi32(zero, ax, sy);
-    return _mm256_cvtepi32_ps(summed_pairs);
-#elif defined(__AVXVNNI__)
-    const __m256i zero = _mm256_setzero_si256();
-    const __m256i summed_pairs = _mm256_dpbusd_avx_epi32(zero, ax, sy);
-    return _mm256_cvtepi32_ps(summed_pairs);
-#else
-    // Perform multiplication and create 16-bit values
-    const __m256i dot = _mm256_maddubs_epi16(ax, sy);
-    return sum_i16_pairs_float(dot);
-#endif
-}
-
-// 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) {
-#if __AVXVNNIINT8__
-    const __m256i zero = _mm256_setzero_si256();
-    const __m256i summed_pairs = _mm256_dpbssd_epi32(zero, x, y);
-    return _mm256_cvtepi32_ps(summed_pairs);
-#else
-    // Get absolute values of x vectors
-    const __m256i ax = _mm256_sign_epi8(x, x);
-    // Sign the values of the y vectors
-    const __m256i sy = _mm256_sign_epi8(y, x);
-    return mul_sum_us8_pairs_float(ax, sy);
-#endif
-}
-
-static inline __m128i packNibbles( __m256i bytes )
-{
-    // Move bits within 16-bit lanes from 0000_abcd_0000_efgh into 0000_0000_abcd_efgh
-#if __AVX512F__
-    const __m256i bytes_srli_4 = _mm256_srli_epi16(bytes, 4);   // 0000_0000_abcd_0000
-    bytes = _mm256_or_si256(bytes, bytes_srli_4);               // 0000_abcd_abcd_efgh
-    return _mm256_cvtepi16_epi8(bytes);                         // abcd_efgh
-#else
-    const __m256i lowByte = _mm256_set1_epi16( 0xFF );
-    __m256i high = _mm256_andnot_si256( lowByte, bytes );
-    __m256i low = _mm256_and_si256( lowByte, bytes );
-    high = _mm256_srli_epi16( high, 4 );
-    bytes = _mm256_or_si256( low, high );
-
-    // Compress uint16_t lanes into bytes
-    __m128i r0 = _mm256_castsi256_si128( bytes );
-    __m128i r1 = _mm256_extracti128_si256( bytes, 1 );
-    return _mm_packus_epi16( r0, r1 );
-#endif
-}
-#elif defined(__AVX__)
-static inline __m128i packNibbles( __m128i bytes1, __m128i bytes2 )
-{
-    // Move bits within 16-bit lanes from 0000_abcd_0000_efgh into 0000_0000_abcd_efgh
-    const __m128i lowByte = _mm_set1_epi16( 0xFF );
-    __m128i high = _mm_andnot_si128( lowByte, bytes1 );
-    __m128i low = _mm_and_si128( lowByte, bytes1 );
-    high = _mm_srli_epi16( high, 4 );
-    bytes1 = _mm_or_si128( low, high );
-    high = _mm_andnot_si128( lowByte, bytes2 );
-    low = _mm_and_si128( lowByte, bytes2 );
-    high = _mm_srli_epi16( high, 4 );
-    bytes2 = _mm_or_si128( low, high );
-
-    return _mm_packus_epi16( bytes1, bytes2);
-}
-
-static inline __m128i mul_add_epi8_sse(const __m128i x, const __m128i y) {
-    const __m128i ax = _mm_sign_epi8(x, x);
-    const __m128i sy = _mm_sign_epi8(y, x);
-    return _mm_maddubs_epi16(ax, sy);
-}
-
-// spread 32 bits to 32 bytes { 0x00, 0xFF }
-static inline __m256i bytes_from_bits_32(const uint8_t * x) {
-    uint32_t x32;
-    memcpy(&x32, x, sizeof(uint32_t));
-    const __m128i shuf_maskl = _mm_set_epi64x(0x0101010101010101, 0x0000000000000000);
-    const __m128i shuf_maskh = _mm_set_epi64x(0x0303030303030303, 0x0202020202020202);
-    __m128i bytesl = _mm_shuffle_epi8(_mm_set1_epi32(x32), shuf_maskl);
-    __m128i bytesh = _mm_shuffle_epi8(_mm_set1_epi32(x32), shuf_maskh);
-    const __m128i bit_mask = _mm_set1_epi64x(0x7fbfdfeff7fbfdfe);
-    bytesl = _mm_or_si128(bytesl, bit_mask);
-    bytesh = _mm_or_si128(bytesh, bit_mask);
-    bytesl = _mm_cmpeq_epi8(bytesl, _mm_set1_epi64x(-1));
-    bytesh = _mm_cmpeq_epi8(bytesh, _mm_set1_epi64x(-1));
-    return MM256_SET_M128I(bytesh, bytesl);
-}
-
-// Unpack 32 4-bit fields into 32 bytes
-// The output vector contains 32 bytes, each one in [ 0 .. 15 ] interval
-static inline __m256i bytes_from_nibbles_32(const uint8_t * rsi)
-{
-    // Load 16 bytes from memory
-    __m128i tmpl = _mm_loadu_si128((const __m128i *)rsi);
-    __m128i tmph = _mm_srli_epi16(tmpl, 4);
-    const __m128i lowMask = _mm_set1_epi8(0xF);
-    tmpl = _mm_and_si128(lowMask, tmpl);
-    tmph = _mm_and_si128(lowMask, tmph);
-    return MM256_SET_M128I(tmph, tmpl);
-}
-
-// add int16_t pairwise and return as float vector
-static inline __m256 sum_i16_pairs_float(const __m128i xh, const __m128i xl) {
-    const __m128i ones = _mm_set1_epi16(1);
-    const __m128i summed_pairsl = _mm_madd_epi16(ones, xl);
-    const __m128i summed_pairsh = _mm_madd_epi16(ones, xh);
-    const __m256i summed_pairs = MM256_SET_M128I(summed_pairsh, summed_pairsl);
-    return _mm256_cvtepi32_ps(summed_pairs);
-}
-
-static inline __m256 mul_sum_us8_pairs_float(const __m256i ax, const __m256i sy) {
-    const __m128i axl = _mm256_castsi256_si128(ax);
-    const __m128i axh = _mm256_extractf128_si256(ax, 1);
-    const __m128i syl = _mm256_castsi256_si128(sy);
-    const __m128i syh = _mm256_extractf128_si256(sy, 1);
-    // Perform multiplication and create 16-bit values
-    const __m128i dotl = _mm_maddubs_epi16(axl, syl);
-    const __m128i doth = _mm_maddubs_epi16(axh, syh);
-    return sum_i16_pairs_float(doth, dotl);
-}
-
-// 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) {
-    const __m128i xl = _mm256_castsi256_si128(x);
-    const __m128i xh = _mm256_extractf128_si256(x, 1);
-    const __m128i yl = _mm256_castsi256_si128(y);
-    const __m128i yh = _mm256_extractf128_si256(y, 1);
-    // Get absolute values of x vectors
-    const __m128i axl = _mm_sign_epi8(xl, xl);
-    const __m128i axh = _mm_sign_epi8(xh, xh);
-    // Sign the values of the y vectors
-    const __m128i syl = _mm_sign_epi8(yl, xl);
-    const __m128i syh = _mm_sign_epi8(yh, xh);
-    // Perform multiplication and create 16-bit values
-    const __m128i dotl = _mm_maddubs_epi16(axl, syl);
-    const __m128i doth = _mm_maddubs_epi16(axh, syh);
-    return sum_i16_pairs_float(doth, dotl);
-}
-
-// larger version of mul_sum_i8_pairs_float where x and y are each represented by four 128-bit vectors
-static inline __m256 mul_sum_i8_quad_float(const __m128i x_1_0, const __m128i x_1_1, const __m128i x_2_0, const __m128i x_2_1,
-                                           const __m128i y_1_0, const __m128i y_1_1, const __m128i y_2_0, const __m128i y_2_1) {
-    const __m128i mone = _mm_set1_epi16(1);
-
-    const __m128i p16_1_0 = mul_add_epi8_sse(x_1_0, y_1_0);
-    const __m128i p16_1_1 = mul_add_epi8_sse(x_1_1, y_1_1);
-    const __m128i p16_2_0 = mul_add_epi8_sse(x_2_0, y_2_0);
-    const __m128i p16_2_1 = mul_add_epi8_sse(x_2_1, y_2_1);
-    const __m128i p_1_0 = _mm_madd_epi16(p16_1_0, mone);
-    const __m128i p_1_1 = _mm_madd_epi16(p16_1_1, mone);
-    const __m128i p_2_0 = _mm_madd_epi16(p16_2_0, mone);
-    const __m128i p_2_1 = _mm_madd_epi16(p16_2_1, mone);
-    const __m128i p_1 = _mm_add_epi32(p_1_0, p_1_1);
-    const __m128i p_2 = _mm_add_epi32(p_2_0, p_2_1);
-    return _mm256_cvtepi32_ps(MM256_SET_M128I(p_2, p_1));
-}
-
-// quad fp16 delta calculation
-static inline __m256 quad_fp16_delta_float(const float x0, const float y0, const float x1, const float y1) {
-    // GGML_FP16_TO_FP32 is faster than Intel F16C
-    return _mm256_set_m128(_mm_set1_ps(GGML_FP16_TO_FP32(x1) * GGML_FP16_TO_FP32(y1)),
-                           _mm_set1_ps(GGML_FP16_TO_FP32(x0) * GGML_FP16_TO_FP32(y0)));
-}
-#endif
-#elif defined(__SSSE3__)
-// horizontally add 4x4 floats
-static inline float hsum_float_4x4(const __m128 a, const __m128 b, const __m128 c, const __m128 d) {
-    __m128 res_0 =_mm_hadd_ps(a, b);
-    __m128 res_1 =_mm_hadd_ps(c, d);
-    __m128 res =_mm_hadd_ps(res_0, res_1);
-    res =_mm_hadd_ps(res, res);
-    res =_mm_hadd_ps(res, res);
-
-    return _mm_cvtss_f32(res);
-}
-#endif // __AVX__ || __AVX2__ || __AVX512F__
-#endif // defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__) || defined(__SSSE3__)
-
-#if defined(__ARM_NEON) || defined(__wasm_simd128__) || defined(__POWER9_VECTOR__)
-#define B1(c,s,n)  0x ## n ## c ,  0x ## n ## s
-#define B2(c,s,n) B1(c,s,n ## c), B1(c,s,n ## s)
-#define B3(c,s,n) B2(c,s,n ## c), B2(c,s,n ## s)
-#define B4(c,s,n) B3(c,s,n ## c), B3(c,s,n ## s)
-#define B5(c,s,n) B4(c,s,n ## c), B4(c,s,n ## s)
-#define B6(c,s,n) B5(c,s,n ## c), B5(c,s,n ## s)
-#define B7(c,s,n) B6(c,s,n ## c), B6(c,s,n ## s)
-#define B8(c,s  ) B7(c,s,     c), B7(c,s,     s)
-
-// precomputed tables for expanding 8bits to 8 bytes:
-static const uint64_t table_b2b_0[1 << 8] = { B8(00, 10) }; // ( b) << 4
-static const uint64_t table_b2b_1[1 << 8] = { B8(10, 00) }; // (!b) << 4
-#endif
-
-#if defined(__loongarch_sx)
-
-static __m128i lsx_packs_w(__m128i a, __m128i b) {
-    __m128i tmp, tmp1;
-    tmp = __lsx_vsat_w(a, 15);
-    tmp1 = __lsx_vsat_w(b, 15);
-    return __lsx_vpickev_h(tmp1, tmp);
-}
-
-static __m128i lsx_packs_h(__m128i a, __m128i b) {
-    __m128i tmp, tmp1;
-    tmp = __lsx_vsat_h(a, 7);
-    tmp1 = __lsx_vsat_h(b, 7);
-    return __lsx_vpickev_b(tmp1, tmp);
-}
-
-static __m128i lsx_packus_h(__m128i a, __m128i b) {
-    __m128i tmp, tmp1;
-    tmp = __lsx_vsat_hu(a, 7);
-    tmp1 = __lsx_vsat_hu(b, 7);
-    return __lsx_vpickev_b(tmp1, tmp);
-}
-
-static __m128i lsx_maddubs_h(__m128i a, __m128i b) {
-    __m128i tmp1, tmp2;
-    tmp1 = __lsx_vmulwev_h_b(a, b);
-    tmp2 = __lsx_vmulwod_h_b(a, b);
-    return __lsx_vsadd_h(tmp1, tmp2);
-}
-
-static __m128i lsx_madd_h(__m128i a, __m128i b) {
-    __m128i tmp1, tmp2;
-    tmp1 = __lsx_vmulwev_w_h(a, b);
-    tmp2 = __lsx_vmulwod_w_h(a, b);
-    return __lsx_vadd_w(tmp1, tmp2);
-}
-
-static __m128i lsx_set_w(int32_t a, int32_t b, int32_t c, int32_t d) {
-    v4i32 __ret = {d, c, b, a};
-    return (__m128i)__ret;
-}
-
-static __m128i lsx_shuffle_b(__m128i a, __m128i b) {
-    __m128i mask_f, zero, tmp0, tmp2, mask;
-    int f = 0x8f;
-    mask_f = __lsx_vreplgr2vr_b(f);
-    zero = __lsx_vldi(0);
-    tmp0 = __lsx_vand_v(b, mask_f); // get mask with low 4 bit and sign bits
-    tmp0 = __lsx_vori_b(tmp0, 0x10); // make each mask or  with 0x10 prepare for positive
-    mask = __lsx_vsle_b(zero, tmp0); // if mask >= 0, set mask
-    tmp2 = __lsx_vand_v(tmp0, mask); // maskout the in2 < ones
-    return __lsx_vshuf_b(a, zero, tmp2);
-}
-
-static __m128i lsx_hadd_h(__m128i a, __m128i b) {
-    __m128i tmp1 = __lsx_vpickev_h(b, a);
-    __m128i tmp2 = __lsx_vpickod_h(b, a);
-    return __lsx_vadd_h(tmp1, tmp2);
-}
-
-static __m128i lsx_hadd_w(__m128i a, __m128i b) {
-    __m128i tmp1 = __lsx_vpickev_w(b, a);
-    __m128i tmp2 = __lsx_vpickod_w(b, a);
-    return __lsx_vadd_w(tmp1, tmp2);
-}
-
-static __m128 lsx_hadd_s(__m128 a, __m128 b) {
-    __m128 tmp1 = (__m128)__lsx_vpickev_w((__m128i)b, (__m128i)a);
-    __m128 tmp2 = (__m128)__lsx_vpickod_w((__m128i)b, (__m128i)a);
-
-    return __lsx_vfadd_s(tmp1, tmp2);
-}
-
-static inline float hsum_float_4x4(const __m128 a, const __m128 b, const __m128 c, const __m128 d) {
-    __m128 res_0 =lsx_hadd_s(a, b);
-    __m128 res_1 =lsx_hadd_s(c, d);
-    __m128 res =lsx_hadd_s(res_0, res_1);
-    res =lsx_hadd_s(res, res);
-    res =lsx_hadd_s(res, res);
-
-    return ((v4f32)res)[0];
-}
-#endif
-
-#if defined(__loongarch_asx)
-
-#ifdef __clang__
-#define VREGS_PREFIX "$vr"
-#define XREGS_PREFIX "$xr"
-#else // GCC
-#define VREGS_PREFIX "$f"
-#define XREGS_PREFIX "$f"
-#endif
-#define __ALL_REGS "0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,21,22,23,24,25,26,27,28,29,30,31"
-// Convert __m128i to __m256i
-static inline __m256i ____m256i(__m128i in) {
-    __m256i out = __lasx_xvldi(0);
-    __asm__ volatile (
-        ".irp i," __ALL_REGS                "\n\t"
-        " .ifc %[out], " XREGS_PREFIX"\\i    \n\t"
-        "  .irp j," __ALL_REGS              "\n\t"
-        "   .ifc %[in], " VREGS_PREFIX "\\j  \n\t"
-        "    xvpermi.q $xr\\i, $xr\\j, 0x20  \n\t"
-        "   .endif                           \n\t"
-        "  .endr                             \n\t"
-        " .endif                             \n\t"
-        ".endr                               \n\t"
-        : [out] "+f" (out) : [in] "f" (in)
-    );
-    return out;
-}
-// Convert two __m128i to __m256i
-static inline __m256i lasx_set_q(__m128i inhi, __m128i inlo) {
-    __m256i out;
-    __asm__ volatile (
-        ".irp i," __ALL_REGS                "\n\t"
-        " .ifc %[hi], " VREGS_PREFIX "\\i    \n\t"
-        "  .irp j," __ALL_REGS              "\n\t"
-        "   .ifc %[lo], " VREGS_PREFIX "\\j  \n\t"
-        "    xvpermi.q $xr\\i, $xr\\j, 0x20  \n\t"
-        "   .endif                           \n\t"
-        "  .endr                             \n\t"
-        " .endif                             \n\t"
-        ".endr                               \n\t"
-        ".ifnc %[out], %[hi]                 \n\t"
-        ".irp i," __ALL_REGS                "\n\t"
-        " .ifc %[out], " XREGS_PREFIX "\\i   \n\t"
-        "  .irp j," __ALL_REGS              "\n\t"
-        "   .ifc %[hi], " VREGS_PREFIX "\\j  \n\t"
-        "    xvori.b $xr\\i, $xr\\j, 0       \n\t"
-        "   .endif                           \n\t"
-        "  .endr                             \n\t"
-        " .endif                             \n\t"
-        ".endr                               \n\t"
-        ".endif                              \n\t"
-        : [out] "=f" (out), [hi] "+f" (inhi)
-        : [lo] "f" (inlo)
-    );
-    return out;
-}
-// Convert __m256i low part to __m128i
-static inline __m128i lasx_extracti128_lo(__m256i in) {
-    __m128i out;
-    __asm__ volatile (
-        ".ifnc %[out], %[in]                 \n\t"
-        ".irp i," __ALL_REGS                "\n\t"
-        " .ifc %[out], " VREGS_PREFIX "\\i   \n\t"
-        "  .irp j," __ALL_REGS              "\n\t"
-        "   .ifc %[in], " XREGS_PREFIX "\\j  \n\t"
-        "    vori.b $vr\\i, $vr\\j, 0        \n\t"
-        "   .endif                           \n\t"
-        "  .endr                             \n\t"
-        " .endif                             \n\t"
-        ".endr                               \n\t"
-        ".endif                              \n\t"
-        : [out] "=f" (out) : [in] "f" (in)
-    );
-    return out;
-}
-// Convert __m256i high part to __m128i
-static inline __m128i lasx_extracti128_hi(__m256i in) {
-    __m128i out;
-    __asm__ volatile (
-        ".irp i," __ALL_REGS                "\n\t"
-        " .ifc %[out], " VREGS_PREFIX "\\i   \n\t"
-        "  .irp j," __ALL_REGS              "\n\t"
-        "   .ifc %[in], " XREGS_PREFIX "\\j  \n\t"
-        "    xvpermi.q $xr\\i, $xr\\j, 0x11  \n\t"
-        "   .endif                           \n\t"
-        "  .endr                             \n\t"
-        " .endif                             \n\t"
-        ".endr                               \n\t"
-        : [out] "=f" (out) : [in] "f" (in)
-    );
-    return out;
-}
-
-static __m256i lasx_set_w(int e7, int e6, int e5, int e4, int e3, int e2, int e1, int e0) {
-    v8i32 __ret = {e0, e1, e2, e3, e4, e5, e6, e7};
-    return (__m256i)__ret;
-}
-
-static __m256i lasx_set_d(int64_t a, int64_t b, int64_t c, int64_t d) {
-    v4i64 __ret = {d, c, b, a};
-    return (__m256i)__ret;
-}
-
-static __m256i lasx_insertf128( __m128i x, __m128i y) {
-    return lasx_set_q(x, y);
-}
-
-static __m256i lasx_shuffle_b(__m256i a, __m256i b) {
-    __m256i mask_f, zero, tmp0, tmp2, mask;
-    int f = 0x8f;
-    mask_f = __lasx_xvreplgr2vr_b(f);
-    zero = __lasx_xvldi(0);
-    tmp0 = __lasx_xvand_v(b, mask_f); // get mask with low 4 bit and sign bits
-    tmp0 = __lasx_xvori_b(tmp0, 0x10); // make each mask or  with 0x10 prepare for positive
-    mask = __lasx_xvsle_b(zero, tmp0); // if mask >= 0, set mask
-    tmp2 = __lasx_xvand_v(tmp0, mask); // maskout the in2 < ones
-    return __lasx_xvshuf_b(a, zero, tmp2);
-}
-
-static __m256i lasx_extu8_16(__m128i a) {
-    return __lasx_vext2xv_hu_bu(____m256i(a));
-}
-
-static __m256i lasx_ext8_16(__m128i a) {
-    return __lasx_vext2xv_h_b(____m256i(a));
-}
-
-static __m256i lasx_ext16_32(__m128i a) {
-    return __lasx_vext2xv_w_h(____m256i(a));
-}
-
-static __m128i lasx_extracti128( __m256i a, int pos) {
-    __m128i ret;
-    if( pos == 0)
-    {
-       ret = lasx_extracti128_lo(a);
-    } else {
-       ret = lasx_extracti128_hi(a);
-    }
-    return ret;
-}
-
-static __m128 lasx_extractf128( __m256 a, int pos) {
-    __m128 ret;
-    if( pos == 0)
-    {
-       ret = (__m128)lasx_extracti128_lo((__m256i)a);
-    } else {
-       ret = (__m128)lasx_extracti128_hi((__m256i)a);
-    }
-    return ret;
-}
-
-static __m256i lasx_maddubs_h(__m256i a, __m256i b) {
-    __m256i tmp1, tmp2;
-    tmp1 = __lasx_xvmulwev_h_b(a, b);
-    tmp2 = __lasx_xvmulwod_h_b(a, b);
-    return __lasx_xvsadd_h(tmp1, tmp2);
-}
-
-static __m256i lasx_madd_h(__m256i a, __m256i b) {
-    __m256i tmp1, tmp2;
-    tmp1 = __lasx_xvmulwev_w_h(a, b);
-    tmp2 = __lasx_xvmulwod_w_h(a, b);
-    return __lasx_xvadd_w(tmp1, tmp2);
-}
-
-static __m256i lasx_packs_w(__m256i a, __m256i b) {
-    __m256i tmp, tmp1;
-    tmp = __lasx_xvsat_w(a, 15);
-    tmp1 = __lasx_xvsat_w(b, 15);
-    return __lasx_xvpickev_h(tmp1, tmp);
-}
-
-static __m256i lasx_packs_h(__m256i a, __m256i b) {
-    __m256i tmp, tmp1;
-    tmp = __lasx_xvsat_h(a, 7);
-    tmp1 = __lasx_xvsat_h(b, 7);
-    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
-    const __m128i ax = __lsx_vsigncov_b(x, x);
-    // Sign the values of the y vectors
-    const __m128i sy = __lsx_vsigncov_b(x, y);
-    // Perform multiplication and create 16-bit values
-    const __m128i dot = lsx_maddubs_h(ax, sy);
-    const __m128i ones = __lsx_vreplgr2vr_h(1);
-    return lsx_madd_h(ones, dot);
-}
-
-// horizontally add 8 floats
-static inline float hsum_float_8(const __m256 x) {
-    __m128 res = lasx_extractf128(x, 1);
-    res = __lsx_vfadd_s(res, lasx_extractf128(x, 0));
-    res = __lsx_vfadd_s(res, (__m128)__lsx_vpickod_d((__m128i)res, (__m128i)res));
-    res = __lsx_vfadd_s(res, (__m128)__lsx_vinsgr2vr_w(__lsx_vldi(0), __lsx_vpickve2gr_w(res, 1), 0));
-    return ((v4f32)res)[0];
-}
-
-// horizontally add 8 int32_t
-static inline int hsum_i32_8(const __m256i a) {
-
-    __m256i tmp1 = __lasx_xvpermi_q(a, a, 0x11);
-    __m256i tmp2 = __lasx_xvpermi_q(a, a, 0x00);
-
-    __m128i  tmp1_128 = lasx_extracti128_lo(tmp1);
-    __m128i  tmp2_128 = lasx_extracti128_lo(tmp2);
-
-    __m128i sum128 = __lsx_vadd_w(tmp1_128, tmp2_128);
-
-    __m128i ev = __lsx_vpickev_w(sum128, sum128);
-    __m128i od = __lsx_vpickod_w(sum128, sum128);
-    __m128i sum64 = __lsx_vadd_w(ev, od);
-
-    int sum64_1, sum64_2;
-    sum64_1 = __lsx_vpickve2gr_w(sum64, 0);
-    sum64_2 = __lsx_vpickve2gr_w(sum64, 1);
-
-    return  sum64_1 + sum64_2;
-}
-
-// horizontally add 4 int32_t
-static inline int hsum_i32_4(const __m128i a) {
-    __m128i ev = __lsx_vpickev_w(a, a);
-    __m128i od = __lsx_vpickod_w(a, a);
-    __m128i sum64 = __lsx_vadd_w(ev, od);
-
-    int sum64_1, sum64_2;
-    sum64_1 = __lsx_vpickve2gr_w(sum64, 0);
-    sum64_2 = __lsx_vpickve2gr_w(sum64, 1);
-
-    return  sum64_1 + sum64_2;
-}
-
-// spread 32 bits to 32 bytes { 0x00, 0xFF }
-static inline __m256i bytes_from_bits_32(const uint8_t * x) {
-
-    uint32_t x32;
-    memcpy(&x32, x, sizeof(uint32_t));
-    const __m256i shuf_mask = lasx_set_d(
-            0x0303030303030303, 0x0202020202020202,
-            0x0101010101010101, 0x0000000000000000);
-
-    __m256i bytes = lasx_shuffle_b(__lasx_xvreplgr2vr_w(x32), shuf_mask);
-    const __m256i bit_mask = __lasx_xvreplgr2vr_d(0x7fbfdfeff7fbfdfe);
-    bytes = __lasx_xvor_v(bytes, bit_mask);
-    return __lasx_xvseq_b(bytes, __lasx_xvreplgr2vr_d(-1));
-}
-
-// Unpack 32 4-bit fields into 32 bytes
-// The output vector contains 32 bytes, each one in [ 0 .. 15 ] interval
-static inline __m256i bytes_from_nibbles_32(const uint8_t * rsi) {
-    const __m128i lo = __lsx_vld((const __m128i *)rsi, 0);
-    __m128i hi = __lsx_vsrli_h(lo, 4);
-    return __lasx_xvandi_b(lasx_insertf128(hi, lo), 0xf);
-}
-
-// add int16_t pairwise and return as float vector
-static inline __m256 sum_i16_pairs_float(const __m256i x) {
-    __m256i v = __lasx_xvpackod_h(x, x);
-    __m256i summed_pairs = __lasx_xvaddwev_w_h(x, v);
-    return __lasx_xvffint_s_w(summed_pairs);
-}
-
-static inline __m256 mul_sum_us8_pairs_float(const __m256i ax, const __m256i sy) {
-    // Perform multiplication and create 16-bit values
-    const __m256i dot = lasx_maddubs_h(ax, sy);
-    return sum_i16_pairs_float(dot);
-}
-
-// 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) {
-    const __m256i dot = lasx_madd_h_b(x, y);
-    return sum_i16_pairs_float(dot);
-}
-
-static inline __m128i packNibbles( __m256i bytes ) {
-    // Move bits within 16-bit lanes from 0000_abcd_0000_efgh into 0000_0000_abcd_efgh
-    const __m256i lowByte = __lasx_xvreplgr2vr_h(0xFF);
-     __m256i high = __lasx_xvandn_v(lowByte, bytes);
-    __m256i low = __lasx_xvand_v(lowByte, bytes);
-    high = __lasx_xvsrli_h(high, 4);
-    bytes = __lasx_xvor_v(low, high);
-    // Compress uint16_t lanes into bytes
-    __m128i *r0 = (__m128i *)&bytes;
-    __m256i tmp_h128 = __lasx_xvpermi_q(bytes, bytes, 0x11);
-    __m128i *r1 = (__m128i *)&tmp_h128;
-
-    __m128i zero = __lsx_vldi(0);
-    __m128i tmp, tmp2, tmp3;
-
-    tmp = __lsx_vmax_h(zero, *r0);
-    tmp2 = __lsx_vsat_hu(tmp, 7);
-
-    tmp = __lsx_vmax_h(zero, *r1);
-    tmp3 = __lsx_vsat_hu(tmp, 7);
-    return  __lsx_vpickev_b(tmp3, tmp2);
-}
-#endif  //__loongarch_asx
-
-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 * 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 * 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 * 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 * 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 * GGML_RESTRICT y = vy;
-
-#if defined(__ARM_NEON)
-    for (int i = 0; i < nb; i++) {
-        float32x4_t srcv [8];
-        float32x4_t asrcv[8];
-        float32x4_t amaxv[8];
-
-        for (int j = 0; j < 8; j++) srcv[j]  = vld1q_f32(x + i*32 + 4*j);
-        for (int j = 0; j < 8; j++) asrcv[j] = vabsq_f32(srcv[j]);
-
-        for (int j = 0; j < 4; j++) amaxv[2*j] = vmaxq_f32(asrcv[2*j], asrcv[2*j+1]);
-        for (int j = 0; j < 2; j++) amaxv[4*j] = vmaxq_f32(amaxv[4*j], amaxv[4*j+2]);
-        for (int j = 0; j < 1; j++) amaxv[8*j] = vmaxq_f32(amaxv[8*j], amaxv[8*j+4]);
-
-        const float amax = vmaxvq_f32(amaxv[0]);
-
-        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 float32x4_t v  = vmulq_n_f32(srcv[j], id);
-            const int32x4_t   vi = vcvtnq_s32_f32(v);
-
-            y[i].qs[4*j + 0] = vgetq_lane_s32(vi, 0);
-            y[i].qs[4*j + 1] = vgetq_lane_s32(vi, 1);
-            y[i].qs[4*j + 2] = vgetq_lane_s32(vi, 2);
-            y[i].qs[4*j + 3] = vgetq_lane_s32(vi, 3);
-        }
-    }
-#elif defined __wasm_simd128__
-    for (int i = 0; i < nb; i++) {
-        v128_t srcv [8];
-        v128_t asrcv[8];
-        v128_t amaxv[8];
-
-        for (int j = 0; j < 8; j++) srcv[j]  = wasm_v128_load(x + i*32 + 4*j);
-        for (int j = 0; j < 8; j++) asrcv[j] = wasm_f32x4_abs(srcv[j]);
-
-        for (int j = 0; j < 4; j++) amaxv[2*j] = wasm_f32x4_max(asrcv[2*j], asrcv[2*j+1]);
-        for (int j = 0; j < 2; j++) amaxv[4*j] = wasm_f32x4_max(amaxv[4*j], amaxv[4*j+2]);
-        for (int j = 0; j < 1; j++) amaxv[8*j] = wasm_f32x4_max(amaxv[8*j], amaxv[8*j+4]);
-
-        const float amax = MAX(MAX(wasm_f32x4_extract_lane(amaxv[0], 0),
-                                   wasm_f32x4_extract_lane(amaxv[0], 1)),
-                               MAX(wasm_f32x4_extract_lane(amaxv[0], 2),
-                                   wasm_f32x4_extract_lane(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 v128_t v  = wasm_f32x4_mul(srcv[j], wasm_f32x4_splat(id));
-            const v128_t vi = wasm_i32x4_trunc_sat_f32x4(v);
-
-            y[i].qs[4*j + 0] = wasm_i32x4_extract_lane(vi, 0);
-            y[i].qs[4*j + 1] = wasm_i32x4_extract_lane(vi, 1);
-            y[i].qs[4*j + 2] = wasm_i32x4_extract_lane(vi, 2);
-            y[i].qs[4*j + 3] = wasm_i32x4_extract_lane(vi, 3);
-        }
-    }
-#elif defined(__AVX2__) || defined(__AVX__)
-    for (int i = 0; i < nb; i++) {
-        // Load elements into 4 AVX vectors
-        __m256 v0 = _mm256_loadu_ps( x );
-        __m256 v1 = _mm256_loadu_ps( x + 8 );
-        __m256 v2 = _mm256_loadu_ps( x + 16 );
-        __m256 v3 = _mm256_loadu_ps( x + 24 );
-        x += 32;
-
-        // Compute max(abs(e)) for the block
-        const __m256 signBit = _mm256_set1_ps( -0.0f );
-        __m256 maxAbs = _mm256_andnot_ps( signBit, v0 );
-        maxAbs = _mm256_max_ps( maxAbs, _mm256_andnot_ps( signBit, v1 ) );
-        maxAbs = _mm256_max_ps( maxAbs, _mm256_andnot_ps( signBit, v2 ) );
-        maxAbs = _mm256_max_ps( maxAbs, _mm256_andnot_ps( signBit, 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 );
-
-        // Quantize these floats
-        const float d = maxScalar / 127.f;
-        y[i].d = GGML_FP32_TO_FP16(d);
-        const float id = ( maxScalar != 0.0f ) ? 127.f / maxScalar : 0.0f;
-        const __m256 mul = _mm256_set1_ps( id );
-
-        // Apply the multiplier
-        v0 = _mm256_mul_ps( v0, mul );
-        v1 = _mm256_mul_ps( v1, mul );
-        v2 = _mm256_mul_ps( v2, mul );
-        v3 = _mm256_mul_ps( v3, mul );
-
-        // 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 );
-
-#if defined(__AVX2__)
-        // Convert int32 to int16
-        i0 = _mm256_packs_epi32( i0, i1 );	// 0, 1, 2, 3,  8, 9, 10, 11,  4, 5, 6, 7, 12, 13, 14, 15
-        i2 = _mm256_packs_epi32( i2, i3 );	// 16, 17, 18, 19,  24, 25, 26, 27,  20, 21, 22, 23, 28, 29, 30, 31
-                                            // Convert int16 to int8
-        i0 = _mm256_packs_epi16( i0, i2 );	// 0, 1, 2, 3,  8, 9, 10, 11,  16, 17, 18, 19,  24, 25, 26, 27,  4, 5, 6, 7, 12, 13, 14, 15, 20, 21, 22, 23, 28, 29, 30, 31
-
-        // We got our precious signed bytes, but the order is now wrong
-        // These AVX2 pack instructions process 16-byte pieces independently
-        // The following instruction is fixing the order
-        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, i0);
-#else
-        // Since we don't have in AVX some necessary functions,
-        // we split the registers in half and call AVX2 analogs from SSE
-        __m128i ni0 = _mm256_castsi256_si128( i0 );
-        __m128i ni1 = _mm256_extractf128_si256( i0, 1);
-        __m128i ni2 = _mm256_castsi256_si128( i1 );
-        __m128i ni3 = _mm256_extractf128_si256( i1, 1);
-        __m128i ni4 = _mm256_castsi256_si128( i2 );
-        __m128i ni5 = _mm256_extractf128_si256( i2, 1);
-        __m128i ni6 = _mm256_castsi256_si128( i3 );
-        __m128i ni7 = _mm256_extractf128_si256( i3, 1);
-
-        // Convert int32 to int16
-        ni0 = _mm_packs_epi32( ni0, ni1 );
-        ni2 = _mm_packs_epi32( ni2, ni3 );
-        ni4 = _mm_packs_epi32( ni4, ni5 );
-        ni6 = _mm_packs_epi32( ni6, ni7 );
-        // Convert int16 to int8
-        ni0 = _mm_packs_epi16( ni0, ni2 );
-        ni4 = _mm_packs_epi16( ni4, ni6 );
-
-        _mm_storeu_si128((__m128i *)(y[i].qs +  0), ni0);
-        _mm_storeu_si128((__m128i *)(y[i].qs + 16), ni4);
-#endif
-    }
-#elif defined(__riscv_v)
-
-    size_t vl = QK8_0;
-
-    for (int i = 0; i < nb; i++) {
-        // load elements
-        vfloat32m8_t v_x   = __riscv_vle32_v_f32m8(x+i*QK8_0, 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_f32m8_f32m1(vfabs, tmp, vl);
-        float amax = __riscv_vfmv_f_s_f32m1_f32(vmax);
-
-        const float d = amax / ((1 << 7) - 1);
-        const float id = d ? 1.0f/d : 0.0f;
-
-        y[i].d = GGML_FP32_TO_FP16(d);
-
-        vfloat32m8_t x0 = __riscv_vfmul_vf_f32m8(v_x, id, vl);
-
-        // convert to integer
-        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_i8m2(y[i].qs , vs, vl);
-    }
-
-#elif defined(__POWER9_VECTOR__)
-    for (int i = 0; i < nb; i++) {
-        vector float srcv [8];
-        vector float asrcv[8];
-        vector float amaxv[8];
-        vector signed int vi[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;
-        const vector float vid = vec_splats(id);
-
-        y[i].d = GGML_FP32_TO_FP16(d);
-
-        for (int j = 0; j < 8; j++) {
-            const vector float v  = vec_round(vec_mul(srcv[j], vid));
-            vi[j] = vec_cts(v, 0);
-        }
-        vec_xst(vec_pack(vec_pack(vi[0], vi[1]), vec_pack(vi[2], vi[3])),  0, &y[i].qs[0]);
-        vec_xst(vec_pack(vec_pack(vi[4], vi[5]), vec_pack(vi[6], vi[7])), 16, &y[i].qs[0]);
-    }
-
-#elif defined(__loongarch_asx)
-    for (int i = 0; i < nb; i++) {
-        __m256 v0 = (__m256)__lasx_xvld( x , 0);
-        __m256 v1 = (__m256)__lasx_xvld( x , 32);
-        __m256 v2 = (__m256)__lasx_xvld( x , 64);
-        __m256 v3 = (__m256)__lasx_xvld( x , 96);
-        x += 32;
-
-        // Compute max(abs(e)) for the block
-        const __m256 sign_bit = __lasx_xvreplfr2vr_s( -0.0f );
-        __m256 max_abs = (__m256)__lasx_xvandn_v( (__m256i)sign_bit, (__m256i)v0 );
-        max_abs = __lasx_xvfmax_s( max_abs, (__m256)__lasx_xvandn_v( (__m256i)sign_bit, (__m256i)v1 ) );
-        max_abs = __lasx_xvfmax_s( max_abs, (__m256)__lasx_xvandn_v( (__m256i)sign_bit, (__m256i)v2 ) );
-        max_abs = __lasx_xvfmax_s( max_abs, (__m256)__lasx_xvandn_v( (__m256i)sign_bit, (__m256i)v3 ) );
-
-        __m128 max4 = __lsx_vfmax_s( lasx_extractf128( max_abs, 1 ), lasx_extractf128( max_abs , 0) );
-        max4 = __lsx_vfmax_s( max4, (__m128)__lsx_vpickod_d((__m128i) max4, (__m128i)max4 ) );
-        __m128 tmp = max4;
-        max4 = __lsx_vfmax_s( max4, (__m128)__lsx_vinsgr2vr_w(tmp, __lsx_vpickve2gr_w( max4, 1 ), 0 ));
-        const float max_scalar = ((v4f32)max4)[0];
-
-        // Quantize these floats
-        const float d = max_scalar / 127.f;
-        y[i].d = GGML_FP32_TO_FP16(d);
-        const float id = ( max_scalar != 0.0f ) ? 127.f / max_scalar : 0.0f;
-        const __m256 mul = (__m256)__lasx_xvreplfr2vr_s( id );
-
-        // Apply the multiplier
-        v0 = __lasx_xvfmul_s( v0, mul );
-        v1 = __lasx_xvfmul_s( v1, mul );
-        v2 = __lasx_xvfmul_s( v2, mul );
-        v3 = __lasx_xvfmul_s( v3, mul );
-
-        // Round to nearest integer
-        __m256i i0 = __lasx_xvftintrne_w_s( v0 );
-        __m256i i1 = __lasx_xvftintrne_w_s( v1 );
-        __m256i i2 = __lasx_xvftintrne_w_s( v2 );
-        __m256i i3 = __lasx_xvftintrne_w_s( v3 );
-
-        __m128i ni0 = lasx_extracti128( i0, 0 );
-        __m128i ni1 = lasx_extracti128( i0, 1);
-        __m128i ni2 = lasx_extracti128( i1, 0);
-        __m128i ni3 = lasx_extracti128( i1, 1);
-        __m128i ni4 = lasx_extracti128( i2, 0);
-        __m128i ni5 = lasx_extracti128( i2, 1);
-        __m128i ni6 = lasx_extracti128( i3, 0);
-        __m128i ni7 = lasx_extracti128( i3, 1);
-
-        // Convert int32 to int16
-        ni0 = lsx_packs_w( ni0, ni1 );
-        ni2 = lsx_packs_w( ni2, ni3 );
-        ni4 = lsx_packs_w( ni4, ni5 );
-        ni6 = lsx_packs_w( ni6, ni7 );
-        // Convert int16 to int8
-        ni0 = lsx_packs_h( ni0, ni2 );
-        ni4 = lsx_packs_h( ni4, ni6 );
-
-        __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);
-
-        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
-    quantize_row_q8_0_ref(x, y, k);
-#endif
-}
-
-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 * GGML_RESTRICT y = vy;
-
-#if defined(__ARM_NEON)
-    for (int i = 0; i < nb; i++) {
-        float32x4_t srcv [8];
-        float32x4_t asrcv[8];
-        float32x4_t amaxv[8];
-
-        for (int j = 0; j < 8; j++) srcv[j]  = vld1q_f32(x + i*32 + 4*j);
-        for (int j = 0; j < 8; j++) asrcv[j] = vabsq_f32(srcv[j]);
-
-        for (int j = 0; j < 4; j++) amaxv[2*j] = vmaxq_f32(asrcv[2*j], asrcv[2*j+1]);
-        for (int j = 0; j < 2; j++) amaxv[4*j] = vmaxq_f32(amaxv[4*j], amaxv[4*j+2]);
-        for (int j = 0; j < 1; j++) amaxv[8*j] = vmaxq_f32(amaxv[8*j], amaxv[8*j+4]);
-
-        const float amax = vmaxvq_f32(amaxv[0]);
-
-        const float d = amax / ((1 << 7) - 1);
-        const float id = d ? 1.0f/d : 0.0f;
-
-        y[i].d = GGML_FP32_TO_FP16(d);
-
-        int32x4_t accv = vdupq_n_s32(0);
-
-        for (int j = 0; j < 8; j++) {
-            const float32x4_t v  = vmulq_n_f32(srcv[j], id);
-            const int32x4_t   vi = vcvtnq_s32_f32(v);
-
-            y[i].qs[4*j + 0] = vgetq_lane_s32(vi, 0);
-            y[i].qs[4*j + 1] = vgetq_lane_s32(vi, 1);
-            y[i].qs[4*j + 2] = vgetq_lane_s32(vi, 2);
-            y[i].qs[4*j + 3] = vgetq_lane_s32(vi, 3);
-
-            accv = vaddq_s32(accv, vi);
-        }
-
-        y[i].s = GGML_FP32_TO_FP16(d * vaddvq_s32(accv));
-    }
-#elif defined __wasm_simd128__
-    for (int i = 0; i < nb; i++) {
-        v128_t srcv [8];
-        v128_t asrcv[8];
-        v128_t amaxv[8];
-
-        for (int j = 0; j < 8; j++) srcv[j]  = wasm_v128_load(x + i*32 + 4*j);
-        for (int j = 0; j < 8; j++) asrcv[j] = wasm_f32x4_abs(srcv[j]);
-
-        for (int j = 0; j < 4; j++) amaxv[2*j] = wasm_f32x4_max(asrcv[2*j], asrcv[2*j+1]);
-        for (int j = 0; j < 2; j++) amaxv[4*j] = wasm_f32x4_max(amaxv[4*j], amaxv[4*j+2]);
-        for (int j = 0; j < 1; j++) amaxv[8*j] = wasm_f32x4_max(amaxv[8*j], amaxv[8*j+4]);
-
-        const float amax = MAX(MAX(wasm_f32x4_extract_lane(amaxv[0], 0),
-                                   wasm_f32x4_extract_lane(amaxv[0], 1)),
-                               MAX(wasm_f32x4_extract_lane(amaxv[0], 2),
-                                   wasm_f32x4_extract_lane(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);
-
-        v128_t accv = wasm_i32x4_splat(0);
-
-        for (int j = 0; j < 8; j++) {
-            const v128_t v  = wasm_f32x4_mul(srcv[j], wasm_f32x4_splat(id));
-            const v128_t vi = wasm_i32x4_trunc_sat_f32x4(v);
-
-            y[i].qs[4*j + 0] = wasm_i32x4_extract_lane(vi, 0);
-            y[i].qs[4*j + 1] = wasm_i32x4_extract_lane(vi, 1);
-            y[i].qs[4*j + 2] = wasm_i32x4_extract_lane(vi, 2);
-            y[i].qs[4*j + 3] = wasm_i32x4_extract_lane(vi, 3);
-
-            accv = wasm_i32x4_add(accv, vi);
-        }
-
-        y[i].s = GGML_FP32_TO_FP16(
-                d * (wasm_i32x4_extract_lane(accv, 0) +
-                     wasm_i32x4_extract_lane(accv, 1) +
-                     wasm_i32x4_extract_lane(accv, 2) +
-                     wasm_i32x4_extract_lane(accv, 3)));
-    }
-#elif defined(__AVX2__) || defined(__AVX__)
-    for (int i = 0; i < nb; i++) {
-        // Load elements into 4 AVX vectors
-        __m256 v0 = _mm256_loadu_ps( x );
-        __m256 v1 = _mm256_loadu_ps( x + 8 );
-        __m256 v2 = _mm256_loadu_ps( x + 16 );
-        __m256 v3 = _mm256_loadu_ps( x + 24 );
-        x += 32;
-
-        // Compute max(abs(e)) for the block
-        const __m256 signBit = _mm256_set1_ps( -0.0f );
-        __m256 maxAbs = _mm256_andnot_ps( signBit, v0 );
-        maxAbs = _mm256_max_ps( maxAbs, _mm256_andnot_ps( signBit, v1 ) );
-        maxAbs = _mm256_max_ps( maxAbs, _mm256_andnot_ps( signBit, v2 ) );
-        maxAbs = _mm256_max_ps( maxAbs, _mm256_andnot_ps( signBit, 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 max_scalar = _mm_cvtss_f32( max4 );
-
-        // Quantize these floats
-        const float d = max_scalar / 127.f;
-        y[i].d = GGML_FP32_TO_FP16(d);
-        const float id = ( max_scalar != 0.0f ) ? 127.f / max_scalar : 0.0f;
-        const __m256 mul = _mm256_set1_ps( id );
-
-        // Apply the multiplier
-        v0 = _mm256_mul_ps( v0, mul );
-        v1 = _mm256_mul_ps( v1, mul );
-        v2 = _mm256_mul_ps( v2, mul );
-        v3 = _mm256_mul_ps( v3, mul );
-
-        // 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 );
-
-#if defined(__AVX2__)
-        // Compute the sum of the quants and set y[i].s
-        y[i].s = GGML_FP32_TO_FP16(d * hsum_i32_8(_mm256_add_epi32(_mm256_add_epi32(i0, i1), _mm256_add_epi32(i2, i3))));
-
-        // Convert int32 to int16
-        i0 = _mm256_packs_epi32( i0, i1 );	// 0, 1, 2, 3,  8, 9, 10, 11,  4, 5, 6, 7, 12, 13, 14, 15
-        i2 = _mm256_packs_epi32( i2, i3 );	// 16, 17, 18, 19,  24, 25, 26, 27,  20, 21, 22, 23, 28, 29, 30, 31
-                                            // Convert int16 to int8
-        i0 = _mm256_packs_epi16( i0, i2 );	// 0, 1, 2, 3,  8, 9, 10, 11,  16, 17, 18, 19,  24, 25, 26, 27,  4, 5, 6, 7, 12, 13, 14, 15, 20, 21, 22, 23, 28, 29, 30, 31
-
-        // We got our precious signed bytes, but the order is now wrong
-        // These AVX2 pack instructions process 16-byte pieces independently
-        // The following instruction is fixing the order
-        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, i0);
-#else
-        // Since we don't have in AVX some necessary functions,
-        // we split the registers in half and call AVX2 analogs from SSE
-        __m128i ni0 = _mm256_castsi256_si128( i0 );
-        __m128i ni1 = _mm256_extractf128_si256( i0, 1);
-        __m128i ni2 = _mm256_castsi256_si128( i1 );
-        __m128i ni3 = _mm256_extractf128_si256( i1, 1);
-        __m128i ni4 = _mm256_castsi256_si128( i2 );
-        __m128i ni5 = _mm256_extractf128_si256( i2, 1);
-        __m128i ni6 = _mm256_castsi256_si128( i3 );
-        __m128i ni7 = _mm256_extractf128_si256( i3, 1);
-
-        // Compute the sum of the quants and set y[i].s
-        const __m128i s0 = _mm_add_epi32(_mm_add_epi32(ni0, ni1), _mm_add_epi32(ni2, ni3));
-        const __m128i s1 = _mm_add_epi32(_mm_add_epi32(ni4, ni5), _mm_add_epi32(ni6, ni7));
-        y[i].s = GGML_FP32_TO_FP16(d * hsum_i32_4(_mm_add_epi32(s0, s1)));
-
-        // Convert int32 to int16
-        ni0 = _mm_packs_epi32( ni0, ni1 );
-        ni2 = _mm_packs_epi32( ni2, ni3 );
-        ni4 = _mm_packs_epi32( ni4, ni5 );
-        ni6 = _mm_packs_epi32( ni6, ni7 );
-        // Convert int16 to int8
-        ni0 = _mm_packs_epi16( ni0, ni2 );
-        ni4 = _mm_packs_epi16( ni4, ni6 );
-
-        _mm_storeu_si128((__m128i *)(y[i].qs +  0), ni0);
-        _mm_storeu_si128((__m128i *)(y[i].qs + 16), ni4);
-#endif
-    }
-#elif defined(__riscv_v)
-
-    size_t vl = QK8_1;
-
-    for (int i = 0; i < nb; i++) {
-        // load elements
-        vfloat32m8_t v_x   = __riscv_vle32_v_f32m8(x+i*QK8_1, 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_f32m8_f32m1(vfabs, tmp, vl);
-        float amax = __riscv_vfmv_f_s_f32m1_f32(vmax);
-
-        const float d  = amax / ((1 << 7) - 1);
-        const float id = d ? 1.0f/d : 0.0f;
-
-        y[i].d = GGML_FP32_TO_FP16(d);
-
-        vfloat32m8_t x0 = __riscv_vfmul_vf_f32m8(v_x, id, vl);
-
-        // convert to integer
-        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_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_i8m2_i16m1(vs, tmp2, vl);
-
-        // set y[i].s
-        int sum = __riscv_vmv_x_s_i16m1_i16(vwrs);
-        y[i].s = GGML_FP32_TO_FP16(sum*d);
-    }
-
-#elif defined(__POWER9_VECTOR__)
-    for (int i = 0; i < nb; i++) {
-        vector float srcv [8];
-        vector float asrcv[8];
-        vector float amaxv[8];
-        vector signed int vi[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;
-        const vector float vid = vec_splats(id);
-
-        y[i].d = GGML_FP32_TO_FP16(d);
-
-        vector int accv = vec_splats(0);
-
-        for (int j = 0; j < 8; j++) {
-            const vector float v  = vec_round(vec_mul(srcv[j], vid));
-            vi[j] = vec_cts(v, 0);
-
-            accv = vec_add(accv, vi[j]);
-        }
-        vec_xst(vec_pack(vec_pack(vi[0], vi[1]), vec_pack(vi[2], vi[3])),  0, &y[i].qs[0]);
-        vec_xst(vec_pack(vec_pack(vi[4], vi[5]), vec_pack(vi[6], vi[7])), 16, &y[i].qs[0]);
-
-        accv = vec_add(accv, vec_sld(accv, accv, 4));
-        accv = vec_add(accv, vec_sld(accv, accv, 8));
-        y[i].s = GGML_FP32_TO_FP16(d * vec_extract(accv, 0));
-    }
-
-#elif defined(__loongarch_asx)
-    for (int i = 0; i < nb; i++) {
-        __m256 v0 = (__m256)__lasx_xvld( x , 0 );
-        __m256 v1 = (__m256)__lasx_xvld( x , 32 );
-        __m256 v2 = (__m256)__lasx_xvld( x , 64 );
-        __m256 v3 = (__m256)__lasx_xvld( x , 96 );
-        x += 32;
-
-        // Compute max(abs(e)) for the block
-        const __m256 sign_bit = __lasx_xvreplfr2vr_s( -0.0f );
-        __m256 max_abs = (__m256)__lasx_xvandn_v( (__m256i)sign_bit, (__m256i)v0 );
-        max_abs = __lasx_xvfmax_s( max_abs, (__m256)__lasx_xvandn_v( (__m256i)sign_bit, (__m256i)v1 ) );
-        max_abs = __lasx_xvfmax_s( max_abs, (__m256)__lasx_xvandn_v( (__m256i)sign_bit, (__m256i)v2 ) );
-        max_abs = __lasx_xvfmax_s( max_abs, (__m256)__lasx_xvandn_v( (__m256i)sign_bit, (__m256i)v3 ) );
-
-        __m128 max4 = __lsx_vfmax_s( lasx_extractf128( max_abs, 1 ), lasx_extractf128( max_abs, 0) );
-        max4 = __lsx_vfmax_s( max4, (__m128)__lsx_vpickod_d((__m128i) max4, (__m128i)max4 ) );
-        __m128 tmp = max4;
-        max4 = __lsx_vfmax_s( max4, (__m128)__lsx_vextrins_w((__m128i)tmp, (__m128i)max4, 0x10 ));
-        const float max_scalar = ((v4f32)max4)[0];
-
-        // Quantize these floats
-        const float d = max_scalar / 127.f;
-        y[i].d = GGML_FP32_TO_FP16(d);
-        const float id = ( max_scalar != 0.0f ) ? 127.f / max_scalar : 0.0f;
-        const __m256 mul = __lasx_xvreplfr2vr_s( id );
-
-        // Apply the multiplier
-        v0 = __lasx_xvfmul_s( v0, mul );
-        v1 = __lasx_xvfmul_s( v1, mul );
-        v2 = __lasx_xvfmul_s( v2, mul );
-        v3 = __lasx_xvfmul_s( v3, mul );
-
-        // Round to nearest integer
-        __m256i i0 = __lasx_xvftintrne_w_s( v0 );
-        __m256i i1 = __lasx_xvftintrne_w_s( v1 );
-        __m256i i2 = __lasx_xvftintrne_w_s( v2 );
-        __m256i i3 = __lasx_xvftintrne_w_s( v3 );
-
-        __m128i ni0 = lasx_extracti128(i0, 0);
-        __m128i ni1 = lasx_extracti128( i0, 1);
-        __m128i ni2 = lasx_extracti128( i1, 0);
-        __m128i ni3 = lasx_extracti128( i1, 1);
-        __m128i ni4 = lasx_extracti128( i2, 0 );
-        __m128i ni5 = lasx_extracti128( i2, 1);
-        __m128i ni6 = lasx_extracti128( i3, 0);
-        __m128i ni7 = lasx_extracti128( i3, 1);
-
-        // Compute the sum of the quants and set y[i].s
-        const __m128i s0 = __lsx_vadd_w(__lsx_vadd_w(ni0, ni1), __lsx_vadd_w(ni2, ni3));
-        const __m128i s1 = __lsx_vadd_w(__lsx_vadd_w(ni4, ni5), __lsx_vadd_w(ni6, ni7));
-        y[i].s = GGML_FP32_TO_FP16(d * hsum_i32_4(__lsx_vadd_w(s0, s1)));
-
-        // Convert int32 to int16
-        ni0 = lsx_packs_w( ni0, ni1 );
-        ni2 = lsx_packs_w( ni2, ni3 );
-        ni4 = lsx_packs_w( ni4, ni5 );
-        ni6 = lsx_packs_w( ni6, ni7 );
-        // Convert int16 to int8
-        ni0 = lsx_packs_h( ni0, ni2 );
-        ni4 = lsx_packs_h( ni4, ni6 );
-
-        __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
-    quantize_row_q8_1_ref(x, y, k);
-#endif
-}
-
-//
-// 2-6 bit quantization in super-blocks
-//
-
-//
-// ===================== Helper functions
-//
-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;
-}
-
-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) {
-        float ax = fabsf(x[i]);
-        if (ax > amax) { amax = ax; max = x[i]; }
-    }
-    if (amax < GROUP_MAX_EPS) { // all zero
-        for (int i = 0; i < n; ++i) {
-            L[i] = 0;
-        }
-        return 0.f;
-    }
-    float iscale = -nmax / max;
-    if (rmse_type == 0) {
-        for (int i = 0; i < n; ++i) {
-            int l = nearest_int(iscale * x[i]);
-            L[i] = nmax + MAX(-nmax, MIN(nmax-1, l));
-        }
-        return 1/iscale;
-    }
-    bool return_early = false;
-    if (rmse_type < 0) {
-        rmse_type = -rmse_type;
-        return_early = true;
-    }
-    float sumlx = 0;
-    float suml2 = 0;
-#ifdef HAVE_BUGGY_APPLE_LINKER
-    // use 'volatile' to prevent unroll and work around a bug in Apple ld64 1015.7
-    for (volatile int i = 0; i < n; ++i) {
-#else
-    for (int i = 0; i < n; ++i) {
-#endif
-        int l = nearest_int(iscale * x[i]);
-        l = MAX(-nmax, MIN(nmax-1, l));
-        L[i] = l + nmax;
-        float w = qw ? qw[i] : rmse_type == 1 ? x[i] * x[i] : rmse_type == 2 ? 1 : rmse_type == 3 ? fabsf(x[i]) : sqrtf(fabsf(x[i]));
-        sumlx += w*x[i]*l;
-        suml2 += w*l*l;
-    }
-    float scale = suml2 ? sumlx/suml2 : 0.0f;
-    if (return_early) return suml2 > 0 ? 0.5f*(scale + 1/iscale) : 1/iscale;
-    float best = scale * sumlx;
-    for (int is = -9; is <= 9; ++is) {
-        if (is == 0) {
-            continue;
-        }
-        iscale = -(nmax + 0.1f*is) / max;
-        sumlx = suml2 = 0;
-        for (int i = 0; i < n; ++i) {
-            int l = nearest_int(iscale * x[i]);
-            l = MAX(-nmax, MIN(nmax-1, l));
-            float w = qw ? qw[i] : rmse_type == 1 ? x[i] * x[i] : rmse_type == 2 ? 1 : rmse_type == 3 ? fabsf(x[i]) : sqrtf(fabsf(x[i]));
-            sumlx += w*x[i]*l;
-            suml2 += w*l*l;
-        }
-        if (suml2 > 0 && sumlx*sumlx > best*suml2) {
-            for (int i = 0; i < n; ++i) {
-                int l = nearest_int(iscale * x[i]);
-                L[i] = nmax + MAX(-nmax, MIN(nmax-1, l));
-            }
-            scale = sumlx/suml2; best = scale*sumlx;
-        }
-    }
-    return scale;
-}
-
-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) {
-        float ax = fabsf(x[i]);
-        if (ax > amax) { amax = ax; max = x[i]; }
-    }
-    if (amax < GROUP_MAX_EPS) { // all zero
-        for (int i = 0; i < n; ++i) { L[i] = 0; }
-        return 0.f;
-    }
-    float iscale = -nmax / max;
-    if (do_rmse) {
-        float sumlx = 0;
-        float suml2 = 0;
-        for (int i = 0; i < n; ++i) {
-            int l = nearest_int(iscale * x[i]);
-            l = MAX(-nmax, MIN(nmax-1, l));
-            L[i] = l;
-            float w = x[i]*x[i];
-            sumlx += w*x[i]*l;
-            suml2 += w*l*l;
-        }
-        for (int itry = 0; itry < 5; ++itry) {
-            int n_changed = 0;
-            for (int i = 0; i < n; ++i) {
-                float w = x[i]*x[i];
-                float slx = sumlx - w*x[i]*L[i];
-                if (slx > 0) {
-                    float sl2 = suml2 - w*L[i]*L[i];
-                    int new_l = nearest_int(x[i] * sl2 / slx);
-                    new_l = MAX(-nmax, MIN(nmax-1, new_l));
-                    if (new_l != L[i]) {
-                        slx += w*x[i]*new_l;
-                        sl2 += w*new_l*new_l;
-                        if (sl2 > 0 && slx*slx*suml2 > sumlx*sumlx*sl2) {
-                            L[i] = new_l; sumlx = slx; suml2 = sl2;
-                            ++n_changed;
-                        }
-                    }
-                }
-            }
-            if (!n_changed) {
-                break;
-            }
-        }
-        for (int i = 0; i < n; ++i) {
-            L[i] += nmax;
-        }
-        return sumlx / suml2;
-    }
-    for (int i = 0; i < n; ++i) {
-        int l = nearest_int(iscale * x[i]);
-        l = MAX(-nmax, MIN(nmax-1, l));
-        L[i] = l + nmax;
-    }
-    return 1/iscale;
-}
-
-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];
-    for (int i = 1; i < n; ++i) {
-        if (x[i] < min) min = x[i];
-        if (x[i] > max) max = x[i];
-    }
-    if (max == min) {
-        for (int i = 0; i < n; ++i) L[i] = 0;
-        *the_min = 0;
-        return 0.f;
-    }
-    if (min > 0) min = 0;
-    float iscale = nmax/(max - min);
-    float scale = 1/iscale;
-    for (int itry = 0; itry < ntry; ++itry) {
-        float sumlx = 0; int suml2 = 0;
-        bool did_change = false;
-        for (int i = 0; i < n; ++i) {
-            int l = nearest_int(iscale*(x[i] - min));
-            l = MAX(0, MIN(nmax, l));
-            if (l != L[i]) {
-                L[i] = l;
-                did_change = true;
-            }
-            sumlx += (x[i] - min)*l;
-            suml2 += l*l;
-        }
-        scale = sumlx/suml2;
-        float sum = 0;
-        for (int i = 0; i < n; ++i) {
-            sum += x[i] - scale*L[i];
-        }
-        min = alpha*min + (1 - alpha)*sum/n;
-        if (min > 0) min = 0;
-        iscale = 1/scale;
-        if (!did_change) break;
-    }
-    *the_min = -min;
-    return scale;
-}
-
-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];
-    float sum_w = weights[0];
-    float sum_x = sum_w * x[0];
-#ifdef HAVE_BUGGY_APPLE_LINKER
-    // use 'volatile' to prevent unroll and work around a bug in Apple ld64 1015.7
-    for (volatile int i = 1; i < n; ++i) {
-#else
-    for (int i = 1; i < n; ++i) {
-#endif
-        if (x[i] < min) min = x[i];
-        if (x[i] > max) max = x[i];
-        float w = weights[i];
-        sum_w += w;
-        sum_x += w * x[i];
-    }
-    if (min > 0) min = 0;
-    if (max == min) {
-        for (int i = 0; i < n; ++i) L[i] = 0;
-        *the_min = -min;
-        return 0.f;
-    }
-    float iscale = nmax/(max - min);
-    float scale = 1/iscale;
-    float best_mad = 0;
-    for (int i = 0; i < n; ++i) {
-        int l = nearest_int(iscale*(x[i] - min));
-        L[i] = MAX(0, MIN(nmax, l));
-        float diff = scale * L[i] + min - x[i];
-        diff = use_mad ? fabsf(diff) : diff * diff;
-        float w = weights[i];
-        best_mad += w * diff;
-    }
-    if (nstep < 1) {
-        *the_min = -min;
-        return scale;
-    }
-    for (int is = 0; is <= nstep; ++is) {
-        iscale = (rmin + rdelta*is + nmax)/(max - min);
-        float sum_l = 0, sum_l2 = 0, sum_xl = 0;
-        for (int i = 0; i < n; ++i) {
-            int l = nearest_int(iscale*(x[i] - min));
-            l = MAX(0, MIN(nmax, l));
-            Laux[i] = l;
-            float w = weights[i];
-            sum_l += w*l;
-            sum_l2 += w*l*l;
-            sum_xl += w*l*x[i];
-        }
-        float D = sum_w * sum_l2 - sum_l * sum_l;
-        if (D > 0) {
-            float this_scale = (sum_w * sum_xl - sum_x * sum_l)/D;
-            float this_min   = (sum_l2 * sum_x - sum_l * sum_xl)/D;
-            if (this_min > 0) {
-                this_min = 0;
-                this_scale = sum_xl / sum_l2;
-            }
-            float mad = 0;
-            for (int i = 0; i < n; ++i) {
-                float diff = this_scale * Laux[i] + this_min - x[i];
-                diff = use_mad ? fabsf(diff) : diff * diff;
-                float w = weights[i];
-                mad += w * diff;
-            }
-            if (mad < best_mad) {
-                for (int i = 0; i < n; ++i) {
-                    L[i] = Laux[i];
-                }
-                best_mad = mad;
-                scale = this_scale;
-                min = this_min;
-            }
-        }
-    }
-    *the_min = -min;
-    return scale;
-}
-
-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 {
-        *d = (q[j+4] & 0xF) | ((q[j-4] >> 6) << 4);
-        *m = (q[j+4] >>  4) | ((q[j-0] >> 6) << 4);
-    }
-}
-
-//========================- 2-bit (de)-quantization
-
-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 * 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 * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) {
-    assert(k % QK_K == 0);
-    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 * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) {
-    assert(k % QK_K == 0);
-    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 * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) {
-    assert(k % QK_K == 0);
-    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 * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) {
-    assert(k % QK_K == 0);
-    block_tq1_0 * GGML_RESTRICT y = vy;
-    quantize_row_tq1_0_ref(x, y, 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 * GGML_RESTRICT y = vy;
-    quantize_row_tq2_0_ref(x, y, k);
-}
-
-static const int8_t kvalues_iq4nl[16] = {-127, -104, -83, -65, -49, -35, -22, -10, 1, 13, 25, 38, 53, 69, 89, 113};
-
-//===================================== Q8_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 =================================
-
-//
-// Helper functions
-//
-#if __AVX__ || __AVX2__ || __AVX512F__
-
-// shuffles to pick the required scales in dot products
-static inline __m256i get_scale_shuffle_q3k(int i) {
-    static const uint8_t k_shuffle[128] = {
-         0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1,     2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3,
-         4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5,     6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7,
-         8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9,    10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,
-        12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,    14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,
-    };
-    return _mm256_loadu_si256((const __m256i*)k_shuffle + i);
-}
-static inline __m256i get_scale_shuffle_k4(int i) {
-    static const uint8_t k_shuffle[256] = {
-         0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1,
-         2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3,
-         4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5,
-         6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7,
-         8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9,
-        10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,
-        12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,
-        14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15
-    };
-    return _mm256_loadu_si256((const __m256i*)k_shuffle + i);
-}
-static inline __m128i get_scale_shuffle(int i) {
-    static const uint8_t k_shuffle[128] = {
-         0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1,
-         2, 2, 2, 2, 2, 2, 2, 2, 3, 3, 3, 3, 3, 3, 3, 3,
-         4, 4, 4, 4, 4, 4, 4, 4, 5, 5, 5, 5, 5, 5, 5, 5,
-         6, 6, 6, 6, 6, 6, 6, 6, 7, 7, 7, 7, 7, 7, 7, 7,
-         8, 8, 8, 8, 8, 8, 8, 8, 9, 9, 9, 9, 9, 9, 9, 9,
-        10,10,10,10,10,10,10,10, 11,11,11,11,11,11,11,11,
-        12,12,12,12,12,12,12,12, 13,13,13,13,13,13,13,13,
-        14,14,14,14,14,14,14,14, 15,15,15,15,15,15,15,15
-    };
-    return _mm_loadu_si128((const __m128i*)k_shuffle + i);
-}
-#elif defined(__loongarch_asx)
-// shuffles to pick the required scales in dot products
-static inline __m256i get_scale_shuffle_q3k(int i) {
-    static const uint8_t k_shuffle[128] = {
-         0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1,     2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3,
-         4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5,     6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7,
-         8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9,    10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,
-        12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,    14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,
-    };
-    return __lasx_xvld((const __m256i*)k_shuffle + i, 0);
-}
-static inline __m256i get_scale_shuffle_k4(int i) {
-    static const uint8_t k_shuffle[256] = {
-         0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1,
-         2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3,
-         4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5,
-         6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7,
-         8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9,
-        10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,
-        12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,
-        14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15
-    };
-    return __lasx_xvld((const __m256i*)k_shuffle + i, 0);
-}
-static inline __m128i get_scale_shuffle(int i) {
-    static const uint8_t k_shuffle[128] = {
-         0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1,
-         2, 2, 2, 2, 2, 2, 2, 2, 3, 3, 3, 3, 3, 3, 3, 3,
-         4, 4, 4, 4, 4, 4, 4, 4, 5, 5, 5, 5, 5, 5, 5, 5,
-         6, 6, 6, 6, 6, 6, 6, 6, 7, 7, 7, 7, 7, 7, 7, 7,
-         8, 8, 8, 8, 8, 8, 8, 8, 9, 9, 9, 9, 9, 9, 9, 9,
-        10,10,10,10,10,10,10,10, 11,11,11,11,11,11,11,11,
-        12,12,12,12,12,12,12,12, 13,13,13,13,13,13,13,13,
-        14,14,14,14,14,14,14,14, 15,15,15,15,15,15,15,15
-    };
-    return __lsx_vld((const __m128i*)k_shuffle + i, 0);
-}
-#endif
-
-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;
-
-    assert(n % qk == 0);
-#if defined(__ARM_FEATURE_MATMUL_INT8)
-    assert((nrc == 2) || (nrc == 1));
-#else
-    assert(nrc == 1);
-#endif
-    UNUSED(nrc);
-    UNUSED(bx);
-    UNUSED(by);
-    UNUSED(bs);
-
-    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 * 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 * 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);
-
-            const uint8x16_t v0_0 = vld1q_u8(b_x0->qs);
-            const uint8x16_t v0_1 = vld1q_u8(b_x1->qs);
-
-            // 4-bit -> 8-bit
-            const int8x16_t v0_0l = vreinterpretq_s8_u8(vandq_u8  (v0_0, m4b));
-            const int8x16_t v0_0h = vreinterpretq_s8_u8(vshrq_n_u8(v0_0, 4));
-            const int8x16_t v0_1l = vreinterpretq_s8_u8(vandq_u8  (v0_1, m4b));
-            const int8x16_t v0_1h = vreinterpretq_s8_u8(vshrq_n_u8(v0_1, 4));
-
-            // sub 8
-            const int8x16_t x0_l = vsubq_s8(v0_0l, s8b);
-            const int8x16_t x0_h = vsubq_s8(v0_0h, s8b);
-            const int8x16_t x1_l = vsubq_s8(v0_1l, s8b);
-            const int8x16_t x1_h = vsubq_s8(v0_1h, s8b);
-
-            // load y
-            const int8x16_t y0_l = vld1q_s8(b_y0->qs);
-            const int8x16_t y0_h = vld1q_s8(b_y0->qs + 16);
-            const int8x16_t y1_l = vld1q_s8(b_y1->qs);
-            const int8x16_t y1_h = vld1q_s8(b_y1->qs + 16);
-
-            float32_t _scale[4] = {
-                GGML_FP16_TO_FP32(b_x0->d)*GGML_FP16_TO_FP32(b_y0->d),
-                GGML_FP16_TO_FP32(b_x0->d)*GGML_FP16_TO_FP32(b_y1->d),
-                GGML_FP16_TO_FP32(b_x1->d)*GGML_FP16_TO_FP32(b_y0->d),
-                GGML_FP16_TO_FP32(b_x1->d)*GGML_FP16_TO_FP32(b_y1->d)
-            };
-            float32x4_t scale = vld1q_f32(_scale);
-
-            int8x16_t l0 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(x0_l), vreinterpretq_s64_s8(x1_l)));
-            int8x16_t l1 = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(x0_l), vreinterpretq_s64_s8(x1_l)));
-
-            int8x16_t l2 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(x0_h), vreinterpretq_s64_s8(x1_h)));
-            int8x16_t l3 = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(x0_h), vreinterpretq_s64_s8(x1_h)));
-
-            int8x16_t r0 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(y0_l), vreinterpretq_s64_s8(y1_l)));
-            int8x16_t r1 = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(y0_l), vreinterpretq_s64_s8(y1_l)));
-
-            int8x16_t r2 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(y0_h), vreinterpretq_s64_s8(y1_h)));
-            int8x16_t r3 = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(y0_h), vreinterpretq_s64_s8(y1_h)));
-
-            sumv0 = vmlaq_f32(sumv0,(vcvtq_f32_s32(vmmlaq_s32((vmmlaq_s32((vmmlaq_s32((vmmlaq_s32(vdupq_n_s32(0), l0, r0)),
-                                                l1, r1)), l2, r2)), l3, r3))), scale);
-        }
-
-        float32x4_t sumv1 = vextq_f32 (sumv0, sumv0, 2);
-        float32x4_t sumv2 = vzip1q_f32(sumv0, sumv1);
-
-        vst1_f32(s,      vget_low_f32 (sumv2));
-        vst1_f32(s + bs, vget_high_f32(sumv2));
-
-        return;
-    }
-#endif
-
-    int ib = 0;
-    float sumf = 0;
-
-#if defined(__ARM_FEATURE_SVE)
-    svfloat32_t sumv0 = svdup_n_f32(0.0f);
-    svfloat32_t sumv1 = svdup_n_f32(0.0f);
-
-    const int vector_length = ggml_cpu_get_sve_cnt()*8;
-
-    // VLA Implementation using switch case
-    switch (vector_length) {
-        case 128:
-            {
-                // predicate for activating higher lanes for 4 float32 elements
-                const svbool_t ph4 = svptrue_pat_b32(SV_VL4);
-
-                for (; ib + 1 < nb; ib += 2) {
-                    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);
-                    const svuint8_t qx1r = svld1rq_u8(svptrue_b8(), x1->qs);
-
-                    // 4-bit -> 8-bit
-                    const svint8_t qx0l = svreinterpret_s8_u8(svand_n_u8_m(svptrue_b8(), qx0r, 0x0F));
-                    const svint8_t qx0h = svreinterpret_s8_u8(svlsr_n_u8_m(svptrue_b8(), qx0r, 0x04));
-                    const svint8_t qx1l = svreinterpret_s8_u8(svand_n_u8_m(svptrue_b8(), qx1r, 0x0F));
-                    const svint8_t qx1h = svreinterpret_s8_u8(svlsr_n_u8_m(svptrue_b8(), qx1r, 0x04));
-
-                    // sub 8
-                    const svint8_t qx0ls = svsub_n_s8_x(svptrue_b8(), qx0h, 8);
-                    const svint8_t qx0hs = svsub_n_s8_x(svptrue_b8(), qx0l, 8);
-                    const svint8_t qx1ls = svsub_n_s8_x(svptrue_b8(), qx1h, 8);
-                    const svint8_t qx1hs = svsub_n_s8_x(svptrue_b8(), qx1l, 8);
-
-                    // load y
-                    const svint8_t qy0h = svld1_s8(svptrue_b8(), y0->qs);
-                    const svint8_t qy0l = svld1_s8(svptrue_b8(), y0->qs + 16);
-                    const svint8_t qy1h = svld1_s8(svptrue_b8(), y1->qs);
-                    const svint8_t qy1l = svld1_s8(svptrue_b8(), y1->qs + 16);
-
-                    // dot product
-                    sumv0 = svmla_n_f32_x(ph4, sumv0, svcvt_f32_s32_x(ph4, svadd_x(ph4,
-                                    svdot_s32(svdup_n_s32(0), qx0ls, qy0l),
-                                    svdot_s32(svdup_n_s32(0), qx0hs, qy0h))), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
-                    sumv1 = svmla_n_f32_x(ph4, sumv1, svcvt_f32_s32_x(ph4, svadd_x(ph4,
-                                    svdot_s32(svdup_n_s32(0), qx1ls, qy1l),
-                                    svdot_s32(svdup_n_s32(0), qx1hs, qy1h))), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
-                }
-
-                sumf = svaddv_f32(svptrue_b32(), svadd_f32_x(svptrue_b32(), sumv0, sumv1));
-            } break;
-        case 256:
-            {
-                // predicate for activating higher lanes for 16 int8 elements
-                const svbool_t ph16 = svptrue_pat_b8(SV_VL16);
-                // predicate for activating lower lanes for  16 int8 elements
-                const svbool_t pl16 = svnot_b_z(svptrue_b8(), ph16);
-
-                for (; ib + 1 < nb; ib += 2) {
-                    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);
-                    const svuint8_t qx1r = svld1rq_u8(svptrue_b8(), x1->qs);
-
-                    // 4-bit -> 8-bit
-                    const svint8_t qx0 = svreinterpret_s8_u8(svlsr_n_u8_m(pl16, svand_n_u8_m(ph16, qx0r, 0x0F), 0x04));
-                    const svint8_t qx1 = svreinterpret_s8_u8(svlsr_n_u8_m(pl16, svand_n_u8_m(ph16, qx1r, 0x0F), 0x04));
-
-                    // sub 8
-                    const svint8_t qx0s = svsub_n_s8_x(svptrue_b8(), qx0, 8);
-                    const svint8_t qx1s = svsub_n_s8_x(svptrue_b8(), qx1, 8);
-
-                    // load y
-                    const svint8_t qy0 = svld1_s8(svptrue_b8(), y0->qs);
-                    const svint8_t qy1 = svld1_s8(svptrue_b8(), y1->qs);
-
-                    // dot product
-                    sumv0 = svmla_n_f32_x(svptrue_b32(), sumv0, svcvt_f32_s32_x(svptrue_b32(),
-                                svdot_s32(svdup_n_s32(0), qx0s, qy0)), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
-                    sumv1 = svmla_n_f32_x(svptrue_b32(), sumv1, svcvt_f32_s32_x(svptrue_b32(),
-                                svdot_s32(svdup_n_s32(0), qx1s, qy1)), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
-                }
-
-                sumf = svaddv_f32(svptrue_b32(), svadd_f32_x(svptrue_b32(), sumv0, sumv1));
-            } break;
-        case 512:
-            {
-                // predicate for activating higher lanes for 32 int8 elements
-                const svbool_t ph32 = svptrue_pat_b8(SV_VL32);
-
-                // predicate for activating higher lanes for 16 int8 elements
-                const svbool_t ph16 = svptrue_pat_b8(SV_VL16);
-                // predicate for activating lower lanes for 16 int8 elements from first 32 int8 activated lanes
-                const svbool_t pl16 = svnot_b_z(ph32, ph16);
-
-                for (; ib + 1 < nb; ib += 2) {
-                    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);
-                    const svuint8_t qx1r = svld1rq_u8(ph32, x1->qs);
-
-                    // 4-bit -> 8-bit
-                    const svint8_t qx0 = svreinterpret_s8_u8(svlsr_n_u8_m(pl16, svand_n_u8_m(ph16, qx0r, 0x0F), 0x04));
-                    const svint8_t qx1 = svreinterpret_s8_u8(svlsr_n_u8_m(pl16, svand_n_u8_m(ph16, qx1r, 0x0F), 0x04));
-
-                    // sub 8
-                    const svint8_t qx0s = svsub_n_s8_x(ph32, qx0, 8);
-                    const svint8_t qx1s = svsub_n_s8_x(ph32, qx1, 8);
-
-                    // load y
-                    const svint8_t qy0 = svld1_s8(ph32, y0->qs);
-                    const svint8_t qy1 = svld1_s8(ph32, y1->qs);
-
-                    // dot product
-                    sumv0 = svmla_n_f32_x(ph32, sumv0, svcvt_f32_s32_x(ph32,
-                                svdot_s32(svdup_n_s32(0), qx0s, qy0)), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
-                    sumv1 = svmla_n_f32_x(ph32, sumv1, svcvt_f32_s32_x(ph32,
-                                svdot_s32(svdup_n_s32(0), qx1s, qy1)), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
-                }
-
-                sumf = svaddv_f32(ph32, svadd_f32_x(ph32, sumv0, sumv1));
-            } break;
-        default:
-            assert(false && "Unsupported vector length");
-            break;
-    }
-
-#elif defined(__ARM_NEON)
-    float32x4_t sumv0 = vdupq_n_f32(0.0f);
-    float32x4_t sumv1 = vdupq_n_f32(0.0f);
-
-    for (; ib + 1 < nb; ib += 2) {
-        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);
-
-        const uint8x16_t v0_0 = vld1q_u8(x0->qs);
-        const uint8x16_t v0_1 = vld1q_u8(x1->qs);
-
-        // 4-bit -> 8-bit
-        const int8x16_t v0_0l = vreinterpretq_s8_u8(vandq_u8  (v0_0, m4b));
-        const int8x16_t v0_0h = vreinterpretq_s8_u8(vshrq_n_u8(v0_0, 4));
-        const int8x16_t v0_1l = vreinterpretq_s8_u8(vandq_u8  (v0_1, m4b));
-        const int8x16_t v0_1h = vreinterpretq_s8_u8(vshrq_n_u8(v0_1, 4));
-
-        // sub 8
-        const int8x16_t v0_0ls = vsubq_s8(v0_0l, s8b);
-        const int8x16_t v0_0hs = vsubq_s8(v0_0h, s8b);
-        const int8x16_t v0_1ls = vsubq_s8(v0_1l, s8b);
-        const int8x16_t v0_1hs = vsubq_s8(v0_1h, s8b);
-
-        // load y
-        const int8x16_t v1_0l = vld1q_s8(y0->qs);
-        const int8x16_t v1_0h = vld1q_s8(y0->qs + 16);
-        const int8x16_t v1_1l = vld1q_s8(y1->qs);
-        const int8x16_t v1_1h = vld1q_s8(y1->qs + 16);
-
-        // dot product into int32x4_t
-        const int32x4_t p_0 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), v0_0ls, v1_0l), v0_0hs, v1_0h);
-        const int32x4_t p_1 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), v0_1ls, v1_1l), v0_1hs, v1_1h);
-
-        sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(p_0), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
-        sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(p_1), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
-    }
-
-    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();
-
-    // Main loop
-    for (; ib < nb; ++ib) {
-        /* Compute combined scale for the block */
-        const __m256 d = _mm256_set1_ps( GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d) );
-
-        __m256i qx = bytes_from_nibbles_32(x[ib].qs);
-
-        // Now we have a vector with bytes in [ 0 .. 15 ] interval. Offset them into [ -8 .. +7 ] interval.
-        const __m256i off = _mm256_set1_epi8( 8 );
-        qx = _mm256_sub_epi8( qx, off );
-
-        __m256i qy = _mm256_loadu_si256((const __m256i *)y[ib].qs);
-
-        const __m256 q = mul_sum_i8_pairs_float(qx, qy);
-
-        /* Multiply q with scale and accumulate */
-        acc = _mm256_fmadd_ps( d, q, acc );
-    }
-
-    sumf = hsum_float_8(acc);
-#elif defined(__AVX__)
-    __m256 accum = _mm256_setzero_ps();
-    for (; ib + 1 < nb; ib += 2) {
-        const __m128i q4bits_1 = _mm_loadu_si128((const __m128i *)x[ib + 0].qs);
-        const __m128i q4bits_2 = _mm_loadu_si128((const __m128i *)x[ib + 1].qs);
-        const __m128i q8b_1_0 = _mm_loadu_si128((const __m128i *)y[ib + 0].qs);
-        const __m128i q8b_1_1 = _mm_loadu_si128((const __m128i *)y[ib + 0].qs + 1);
-        const __m128i q8b_2_0 = _mm_loadu_si128((const __m128i *)y[ib + 1].qs);
-        const __m128i q8b_2_1 = _mm_loadu_si128((const __m128i *)y[ib + 1].qs + 1);
-
-        const __m128i q4b_1_0 = _mm_sub_epi8(_mm_and_si128(_mm_set1_epi8(15), q4bits_1), _mm_set1_epi8(8));
-        const __m128i q4b_1_1 = _mm_sub_epi8(_mm_and_si128(_mm_set1_epi8(15), _mm_srli_epi16(q4bits_1, 4)), _mm_set1_epi8(8));
-        const __m128i q4b_2_0 = _mm_sub_epi8(_mm_and_si128(_mm_set1_epi8(15), q4bits_2), _mm_set1_epi8(8));
-        const __m128i q4b_2_1 = _mm_sub_epi8(_mm_and_si128(_mm_set1_epi8(15), _mm_srli_epi16(q4bits_2, 4)), _mm_set1_epi8(8));
-
-        const __m128i p16_1_0 = mul_add_epi8_sse(q4b_1_0, q8b_1_0);
-        const __m128i p16_1_1 = mul_add_epi8_sse(q4b_1_1, q8b_1_1);
-        const __m128i p16_2_0 = mul_add_epi8_sse(q4b_2_0, q8b_2_0);
-        const __m128i p16_2_1 = mul_add_epi8_sse(q4b_2_1, q8b_2_1);
-        const __m128i p_1 = _mm_add_epi16(p16_1_0, p16_1_1);
-        const __m128i p_2 = _mm_add_epi16(p16_2_0, p16_2_1);
-        const __m256 p =  sum_i16_pairs_float(p_2, p_1);
-
-        const __m256 deltas = quad_fp16_delta_float(x[ib].d, y[ib].d, x[ib + 1].d, y[ib + 1].d);
-        accum = _mm256_add_ps(_mm256_mul_ps(deltas, p), accum);
-    }
-
-    sumf = hsum_float_8(accum);
-#elif defined(__SSSE3__)
-    // set constants
-    const __m128i lowMask = _mm_set1_epi8(0xF);
-    const __m128i off = _mm_set1_epi8(8);
-
-    // Initialize accumulator with zeros
-    __m128 acc_0 = _mm_setzero_ps();
-    __m128 acc_1 = _mm_setzero_ps();
-    __m128 acc_2 = _mm_setzero_ps();
-    __m128 acc_3 = _mm_setzero_ps();
-
-    for (; ib + 1 < nb; ib += 2) {
-        _mm_prefetch(&x[ib] + sizeof(block_q4_0), _MM_HINT_T0);
-        _mm_prefetch(&y[ib] + sizeof(block_q8_0), _MM_HINT_T0);
-
-        // Compute combined scale for the block 0 and 1
-        const __m128 d_0_1 = _mm_set1_ps( GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d) );
-
-        const __m128i tmp_0_1 = _mm_loadu_si128((const __m128i *)x[ib].qs);
-
-        __m128i bx_0 = _mm_and_si128(lowMask, tmp_0_1);
-        __m128i by_0 = _mm_loadu_si128((const __m128i *)y[ib].qs);
-        bx_0 = _mm_sub_epi8(bx_0, off);
-        const __m128i i32_0 = mul_sum_i8_pairs(bx_0, by_0);
-
-        __m128i bx_1 = _mm_and_si128(lowMask, _mm_srli_epi64(tmp_0_1, 4));
-        __m128i by_1 = _mm_loadu_si128((const __m128i *)(y[ib].qs + 16));
-        bx_1 = _mm_sub_epi8(bx_1, off);
-        const __m128i i32_1 = mul_sum_i8_pairs(bx_1, by_1);
-
-        _mm_prefetch(&x[ib] + 2 * sizeof(block_q4_0), _MM_HINT_T0);
-        _mm_prefetch(&y[ib] + 2 * sizeof(block_q8_0), _MM_HINT_T0);
-
-        // Compute combined scale for the block 2 and 3
-        const __m128 d_2_3 = _mm_set1_ps( GGML_FP16_TO_FP32(x[ib + 1].d) * GGML_FP16_TO_FP32(y[ib + 1].d) );
-
-        const __m128i tmp_2_3 = _mm_loadu_si128((const __m128i *)x[ib + 1].qs);
-
-        __m128i bx_2 = _mm_and_si128(lowMask, tmp_2_3);
-        __m128i by_2 = _mm_loadu_si128((const __m128i *)y[ib + 1].qs);
-        bx_2 = _mm_sub_epi8(bx_2, off);
-        const __m128i i32_2 = mul_sum_i8_pairs(bx_2, by_2);
-
-        __m128i bx_3 = _mm_and_si128(lowMask, _mm_srli_epi64(tmp_2_3, 4));
-        __m128i by_3 = _mm_loadu_si128((const __m128i *)(y[ib + 1].qs + 16));
-        bx_3 = _mm_sub_epi8(bx_3, off);
-        const __m128i i32_3 = mul_sum_i8_pairs(bx_3, by_3);
-
-        // Convert int32_t to float
-        __m128 p0 = _mm_cvtepi32_ps(i32_0);
-        __m128 p1 = _mm_cvtepi32_ps(i32_1);
-        __m128 p2 = _mm_cvtepi32_ps(i32_2);
-        __m128 p3 = _mm_cvtepi32_ps(i32_3);
-
-        // Apply the scale
-        __m128 p0_d = _mm_mul_ps( d_0_1, p0 );
-        __m128 p1_d = _mm_mul_ps( d_0_1, p1 );
-        __m128 p2_d = _mm_mul_ps( d_2_3, p2 );
-        __m128 p3_d = _mm_mul_ps( d_2_3, p3 );
-
-        // Acummulate
-        acc_0 = _mm_add_ps(p0_d, acc_0);
-        acc_1 = _mm_add_ps(p1_d, acc_1);
-        acc_2 = _mm_add_ps(p2_d, acc_2);
-        acc_3 = _mm_add_ps(p3_d, acc_3);
-    }
-
-    sumf = hsum_float_4x4(acc_0, acc_1, acc_2, acc_3);
-#elif defined(__riscv_v)
-    size_t vl = qk / 2;
-
-    for (; ib < nb; ++ib) {
-        // load elements
-        vuint8m1_t tx = __riscv_vle8_v_u8m1(x[ib].qs, 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
-        vuint8m1_t x_a = __riscv_vand_vx_u8m1(tx, 0x0F, vl);
-        vuint8m1_t x_l = __riscv_vsrl_vx_u8m1(tx, 0x04, vl);
-
-        vint8m1_t x_ai = __riscv_vreinterpret_v_u8m1_i8m1(x_a);
-        vint8m1_t x_li = __riscv_vreinterpret_v_u8m1_i8m1(x_l);
-
-        // subtract offset
-        vint8m1_t v0 = __riscv_vsub_vx_i8m1(x_ai, 8, vl);
-        vint8m1_t v1 = __riscv_vsub_vx_i8m1(x_li, 8, 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 vs2 = __riscv_vwredsum_vs_i16m2_i32m1(vec_mul2, vec_zero, vl);
-
-        int sumi = __riscv_vmv_x_s_i32m1_i32(vs2);
-
-        sumf += sumi*GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d);
-    }
-
-#elif defined(__POWER9_VECTOR__)
-    const vector signed char lowMask = vec_splats((signed char)0xF);
-    const vector signed int v0 = vec_splats((int32_t)0);
-    const vector unsigned char v4 = vec_splats((unsigned char)0x4);
-    const vector signed char v8 = vec_splats((signed char)0x8);
-
-    vector float vsumf0 = 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);
-
-        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[ib].d));
-        vector float vyd = vec_splats(GGML_FP16_TO_FP32(y[ib].d));
-        vector float vd = vec_mul(vxd, vyd);
-
-        vector signed char qxs = (vector signed char)vec_xl( 0, x[ib].qs);
-        vector signed char q8y0 = vec_xl( 0, y[ib].qs);
-        vector signed char q8y1 = vec_xl(16, y[ib].qs);
-
-        vector signed char q4x0 = vec_and(qxs, lowMask);
-        vector signed char q4x1 = vec_sr(qxs, v4);
-
-        q4x0 = vec_sub(q4x0, v8);
-        q4x1 = vec_sub(q4x1, v8);
-
-        vector signed short qv0 = vec_add(vec_mule(q4x0, q8y0), vec_mulo(q4x0, q8y0));
-        vector signed short qv1 = vec_add(vec_mule(q4x1, q8y1), vec_mulo(q4x1, q8y1));
-
-        vector signed int vsumi0 = v0;
-
-        vsumi0 = vec_sum4s(qv0, vsumi0);
-        vsumi0 = vec_sum4s(qv1, vsumi0);
-
-        vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
-    }
-
-    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4));
-    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8));
-
-    sumf = vec_extract(vsumf0, 0);
-
-#elif defined(__loongarch_asx)
-    // Initialize accumulator with zeros
-    __m256 acc = (__m256)__lasx_xvldi(0);
-
-    // Main loop
-    for (; ib < nb; ++ib) {
-        /* Compute combined scale for the block */
-        const __m256 d = __lasx_xvreplfr2vr_s( GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d) );
-
-        __m256i qx = bytes_from_nibbles_32(x[ib].qs);
-
-        // Now we have a vector with bytes in [ 0 .. 15 ] interval. Offset them into [ -8 .. +7 ] interval.
-        const __m256i off = __lasx_xvreplgr2vr_b( 8 );
-        qx = __lasx_xvsub_b( qx, off );
-
-        __m256i qy = __lasx_xvld((const __m256i *)y[ib].qs, 0);
-
-        const __m256 q = mul_sum_i8_pairs_float(qx, qy);
-
-        /* Multiply q with scale and accumulate */
-        acc = __lasx_xvfmadd_s( d, q, acc );
-    }
-
-    sumf = hsum_float_8(acc);
-
-#elif defined(__loongarch_sx)
-    // set constants
-    const __m128i low_mask = __lsx_vreplgr2vr_b(0xF);
-    const __m128i off = __lsx_vreplgr2vr_b(8);
-
-    // Initialize accumulator with zeros
-    __m128 acc_0 = (__m128)__lsx_vldi(0);
-    __m128 acc_1 = (__m128)__lsx_vldi(0);
-    __m128 acc_2 = (__m128)__lsx_vldi(0);
-    __m128 acc_3 = (__m128)__lsx_vldi(0);
-
-    for (; ib + 1 < nb; ib += 2) {
-
-        // Compute combined scale for the block 0 and 1
-        const __m128 d_0_1 = (__m128)__lsx_vreplgr2vr_w( GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d) );
-
-        const __m128i tmp_0_1 = __lsx_vld((const __m128i *)x[ib].qs, 0);
-
-        __m128i bx_0 = __lsx_vand_v(low_mask, tmp_0_1);
-        __m128i by_0 = __lsx_vld((const __m128i *)y[ib].qs, 0);
-        bx_0 = __lsx_vsub_b(bx_0, off);
-        const __m128i i32_0 = mul_sum_i8_pairs(bx_0, by_0);
-
-        __m128i bx_1 = __lsx_vand_v(low_mask, __lsx_vsrli_d(tmp_0_1, 4));
-        __m128i by_1 = __lsx_vld((const __m128i *)(y[ib].qs + 16), 0);
-        bx_1 = __lsx_vsub_b(bx_1, off);
-        const __m128i i32_1 = mul_sum_i8_pairs(bx_1, by_1);
-
-        //_mm_prefetch(&x[ib] + 2 * sizeof(block_q4_0), _MM_HINT_T0);
-        //_mm_prefetch(&y[ib] + 2 * sizeof(block_q8_0), _MM_HINT_T0);
-
-        // Compute combined scale for the block 2 and 3
-        const __m128 d_2_3 = (__m128)__lsx_vreplgr2vr_w( GGML_FP16_TO_FP32(x[ib + 1].d) * GGML_FP16_TO_FP32(y[ib + 1].d) );
-
-        const __m128i tmp_2_3 = __lsx_vld((const __m128i *)x[ib + 1].qs, 0);
-
-        __m128i bx_2 = __lsx_vand_v(low_mask, tmp_2_3);
-        __m128i by_2 = __lsx_vld((const __m128i *)y[ib + 1].qs, 0);
-        bx_2 = __lsx_vsub_b(bx_2, off);
-        const __m128i i32_2 = mul_sum_i8_pairs(bx_2, by_2);
-
-        __m128i bx_3 = __lsx_vand_v(low_mask, __lsx_vsrli_d(tmp_2_3, 4));
-        __m128i by_3 = __lsx_vld((const __m128i *)(y[ib + 1].qs + 16), 0);
-        bx_3 = __lsx_vsub_b(bx_3, off);
-        const __m128i i32_3 = mul_sum_i8_pairs(bx_3, by_3);
-
-        // Convert int32_t to float
-        __m128 p0 = __lsx_vffint_s_w(i32_0);
-        __m128 p1 = __lsx_vffint_s_w(i32_1);
-        __m128 p2 = __lsx_vffint_s_w(i32_2);
-        __m128 p3 = __lsx_vffint_s_w(i32_3);
-
-        // Apply the scale
-        __m128 p0_d = __lsx_vfmul_s( d_0_1, p0 );
-        __m128 p1_d = __lsx_vfmul_s( d_0_1, p1 );
-        __m128 p2_d = __lsx_vfmul_s( d_2_3, p2 );
-        __m128 p3_d = __lsx_vfmul_s( d_2_3, p3 );
-
-        // Acummulate
-        acc_0 = __lsx_vfadd_s(p0_d, acc_0);
-        acc_1 = __lsx_vfadd_s(p1_d, acc_1);
-        acc_2 = __lsx_vfadd_s(p2_d, acc_2);
-        acc_3 = __lsx_vfadd_s(p3_d, acc_3);
-    }
-
-    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;
-        int sumi1 = 0;
-
-        for (int j = 0; j < qk/2; ++j) {
-            const int v0 = (x[ib].qs[j] & 0x0F) - 8;
-            const int v1 = (x[ib].qs[j] >>   4) - 8;
-
-            sumi0 += (v0 * y[ib].qs[j]);
-            sumi1 += (v1 * y[ib].qs[j + qk/2]);
-        }
-
-        int sumi = sumi0 + sumi1;
-        sumf += sumi*GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d);
-    }
-
-    *s = sumf;
-}
-
-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;
-
-    assert(n % qk == 0);
-#if defined(__ARM_FEATURE_MATMUL_INT8)
-    assert((nrc == 2) || (nrc == 1));
-#else
-    assert(nrc == 1);
-#endif
-    UNUSED(nrc);
-    UNUSED(bx);
-    UNUSED(by);
-    UNUSED(bs);
-
-    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 * 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 * 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),
-                GGML_FP16_TO_FP32(b_x1->m) * GGML_FP16_TO_FP32(b_y0->s),
-                GGML_FP16_TO_FP32(b_x0->m) * GGML_FP16_TO_FP32(b_y1->s),
-                GGML_FP16_TO_FP32(b_x1->m) * GGML_FP16_TO_FP32(b_y1->s)
-            };
-            summs0 = vaddq_f32(summs0, vld1q_f32(summs_t));
-
-            const uint8x16_t m4b = vdupq_n_u8(0x0F);
-
-            const uint8x16_t v0_0 = vld1q_u8(b_x0->qs);
-            const uint8x16_t v0_1 = vld1q_u8(b_x1->qs);
-
-            // 4-bit -> 8-bit
-            const int8x16_t x0_l = vreinterpretq_s8_u8(vandq_u8  (v0_0, m4b));
-            const int8x16_t x0_h = vreinterpretq_s8_u8(vshrq_n_u8(v0_0, 4));
-            const int8x16_t x1_l = vreinterpretq_s8_u8(vandq_u8  (v0_1, m4b));
-            const int8x16_t x1_h = vreinterpretq_s8_u8(vshrq_n_u8(v0_1, 4));
-
-            // load y
-            const int8x16_t y0_l = vld1q_s8(b_y0->qs);
-            const int8x16_t y0_h = vld1q_s8(b_y0->qs + 16);
-            const int8x16_t y1_l = vld1q_s8(b_y1->qs);
-            const int8x16_t y1_h = vld1q_s8(b_y1->qs + 16);
-
-            // mmla into int32x4_t
-            float32_t _scale[4] = {
-                GGML_FP16_TO_FP32(b_x0->d)*GGML_FP16_TO_FP32(b_y0->d),
-                GGML_FP16_TO_FP32(b_x0->d)*GGML_FP16_TO_FP32(b_y1->d),
-                GGML_FP16_TO_FP32(b_x1->d)*GGML_FP16_TO_FP32(b_y0->d),
-                GGML_FP16_TO_FP32(b_x1->d)*GGML_FP16_TO_FP32(b_y1->d)
-            };
-            float32x4_t scale = vld1q_f32(_scale);
-
-            int8x16_t l0 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(x0_l), vreinterpretq_s64_s8(x1_l)));
-            int8x16_t l1 = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(x0_l), vreinterpretq_s64_s8(x1_l)));
-
-            int8x16_t l2 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(x0_h), vreinterpretq_s64_s8(x1_h)));
-            int8x16_t l3 = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(x0_h), vreinterpretq_s64_s8(x1_h)));
-
-            int8x16_t r0 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(y0_l), vreinterpretq_s64_s8(y1_l)));
-            int8x16_t r1 = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(y0_l), vreinterpretq_s64_s8(y1_l)));
-
-            int8x16_t r2 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(y0_h), vreinterpretq_s64_s8(y1_h)));
-            int8x16_t r3 = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(y0_h), vreinterpretq_s64_s8(y1_h)));
-            sumv0 = vmlaq_f32(sumv0,(vcvtq_f32_s32(vmmlaq_s32((vmmlaq_s32((vmmlaq_s32((vmmlaq_s32(vdupq_n_s32(0), l0, r0)),
-                                                l1, r1)), l2, r2)), l3, r3))), scale);
-        }
-
-        float32x4_t sumv1 = vextq_f32 (sumv0, sumv0, 2);
-        float32x4_t sumv2 = vzip1q_f32(sumv0, sumv1);
-
-        sumv2 = vaddq_f32(sumv2, summs0);
-
-        vst1_f32(s,      vget_low_f32 (sumv2));
-        vst1_f32(s + bs, vget_high_f32(sumv2));
-
-        return;
-    }
-#endif
-
-    int ib = 0;
-    float sumf = 0;
-
-    // TODO: add WASM SIMD
-#if defined(__ARM_NEON)
-    float32x4_t sumv0 = vdupq_n_f32(0.0f);
-    float32x4_t sumv1 = vdupq_n_f32(0.0f);
-
-    float summs = 0;
-
-    for (; ib + 1 < nb; ib += 2) {
-        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);
-
-        const uint8x16_t m4b = vdupq_n_u8(0x0F);
-
-        const uint8x16_t v0_0 = vld1q_u8(x0->qs);
-        const uint8x16_t v0_1 = vld1q_u8(x1->qs);
-
-        // 4-bit -> 8-bit
-        const int8x16_t v0_0l = vreinterpretq_s8_u8(vandq_u8  (v0_0, m4b));
-        const int8x16_t v0_0h = vreinterpretq_s8_u8(vshrq_n_u8(v0_0, 4));
-        const int8x16_t v0_1l = vreinterpretq_s8_u8(vandq_u8  (v0_1, m4b));
-        const int8x16_t v0_1h = vreinterpretq_s8_u8(vshrq_n_u8(v0_1, 4));
-
-        // load y
-        const int8x16_t v1_0l = vld1q_s8(y0->qs);
-        const int8x16_t v1_0h = vld1q_s8(y0->qs + 16);
-        const int8x16_t v1_1l = vld1q_s8(y1->qs);
-        const int8x16_t v1_1h = vld1q_s8(y1->qs + 16);
-
-        // dot product into int32x4_t
-        const int32x4_t p_0 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), v0_0l, v1_0l), v0_0h, v1_0h);
-        const int32x4_t p_1 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), v0_1l, v1_1l), v0_1h, v1_1h);
-
-        sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(p_0), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
-        sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(p_1), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
-    }
-
-    sumf = vaddvq_f32(sumv0) + vaddvq_f32(sumv1) + summs;
-#elif defined(__AVX2__) || defined(__AVX__)
-    // Initialize accumulator with zeros
-    __m256 acc = _mm256_setzero_ps();
-
-    float summs = 0;
-
-    // Main loop
-    for (; ib < nb; ++ib) {
-        const float d0 = GGML_FP16_TO_FP32(x[ib].d);
-        const float d1 = GGML_FP16_TO_FP32(y[ib].d);
-
-        summs += GGML_FP16_TO_FP32(x[ib].m) * GGML_FP16_TO_FP32(y[ib].s);
-
-        const __m256 d0v = _mm256_set1_ps( d0 );
-        const __m256 d1v = _mm256_set1_ps( d1 );
-
-        // Compute combined scales
-        const __m256 d0d1 = _mm256_mul_ps( d0v, d1v );
-
-        // Load 16 bytes, and unpack 4 bit fields into bytes, making 32 bytes
-        const __m256i qx = bytes_from_nibbles_32(x[ib].qs);
-        const __m256i qy = _mm256_loadu_si256( (const __m256i *)y[ib].qs );
-
-        const __m256 xy = mul_sum_us8_pairs_float(qx, qy);
-
-        // Accumulate d0*d1*x*y
-#if defined(__AVX2__)
-        acc = _mm256_fmadd_ps( d0d1, xy, acc );
-#else
-        acc = _mm256_add_ps( _mm256_mul_ps( d0d1, xy ), acc );
-#endif
-    }
-
-    sumf = hsum_float_8(acc) + summs;
-#elif defined(__riscv_v)
-    size_t vl = qk / 2;
-
-    for (; ib < nb; ++ib) {
-        // load elements
-        vuint8m1_t tx = __riscv_vle8_v_u8m1(x[ib].qs, 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
-        vuint8m1_t x_a = __riscv_vand_vx_u8m1(tx, 0x0F, vl);
-        vuint8m1_t x_l = __riscv_vsrl_vx_u8m1(tx, 0x04, vl);
-
-        vint8m1_t v0 = __riscv_vreinterpret_v_u8m1_i8m1(x_a);
-        vint8m1_t v1 = __riscv_vreinterpret_v_u8m1_i8m1(x_l);
-
-        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 vs2 = __riscv_vwredsum_vs_i16m2_i32m1(vec_mul2, vec_zero, 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 + GGML_FP16_TO_FP32(x[ib].m)*GGML_FP16_TO_FP32(y[ib].s);
-    }
-
-#elif defined(__POWER9_VECTOR__)
-    const vector signed char lowMask = vec_splats((signed char)0xF);
-    const vector signed int v0 = vec_splats((int32_t)0);
-    const vector unsigned char v4 = vec_splats((unsigned char)0x4);
-
-    vector float vsumf0 = vec_splats(0.0f);
-
-#pragma GCC unroll 4
-    for (; ib < nb; ++ib) {
-        __builtin_prefetch(x[ib].qs, 0, 1);
-        __builtin_prefetch(y[ib].qs, 0, 1);
-
-        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[ib].d));
-        vector float vyd = vec_splats(GGML_FP16_TO_FP32(y[ib].d));
-        vector float vd = vec_mul(vxd, vyd);
-
-        vector float vxmin = vec_splats(GGML_FP16_TO_FP32(x[ib].m));
-        vector float vys = {GGML_FP16_TO_FP32(y[ib].s), 0.0f, 0.0f, 0.0f};
-        vsumf0 = vec_madd(vxmin, vys, vsumf0);
-
-        vector signed char qxs = (vector signed char)vec_xl( 0, x[ib].qs);
-        vector signed char q8y0 = vec_xl( 0, y[ib].qs);
-        vector signed char q8y1 = vec_xl(16, y[ib].qs);
-
-        vector unsigned char q4x0 = (vector unsigned char)vec_and(qxs, lowMask);
-        vector unsigned char q4x1 = (vector unsigned char)vec_sr(qxs, v4);
-
-        vector signed int vsumi0 = v0;
-
-        vsumi0 = vec_msum(q8y0, q4x0, vsumi0);
-        vsumi0 = vec_msum(q8y1, q4x1, vsumi0);
-
-        vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
-    }
-
-    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4));
-    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8));
-
-    sumf = vec_extract(vsumf0, 0);
-
-#elif defined(__loongarch_asx)
-    // Initialize accumulator with zeros
-    __m256 acc = (__m256)__lasx_xvldi(0);
-
-    float summs = 0;
-
-    // Main loop
-    for (; ib < nb; ++ib) {
-        const float d0 = GGML_FP16_TO_FP32(x[ib].d);
-        const float d1 = GGML_FP16_TO_FP32(y[ib].d);
-
-        summs += GGML_FP16_TO_FP32(x[ib].m) * GGML_FP16_TO_FP32(y[ib].s);
-
-        const __m256 d0v = __lasx_xvreplfr2vr_s( d0 );
-        const __m256 d1v = __lasx_xvreplfr2vr_s( d1 );
-
-        // Compute combined scales
-        const __m256 d0d1 = __lasx_xvfmul_s( d0v, d1v );
-
-        // Load 16 bytes, and unpack 4 bit fields into bytes, making 32 bytes
-        const __m256i qx = bytes_from_nibbles_32(x[ib].qs);
-        const __m256i qy = __lasx_xvld( (const __m256i *)y[ib].qs, 0);
-
-        const __m256 xy = mul_sum_us8_pairs_float(qx, qy);
-
-        // Accumulate d0*d1*x*y
-        acc = __lasx_xvfmadd_s( d0d1, xy, acc );
-    }
-
-    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;
-        int sumi1 = 0;
-
-        for (int j = 0; j < qk/2; ++j) {
-            const int v0 = (x[ib].qs[j] & 0x0F);
-            const int v1 = (x[ib].qs[j] >>   4);
-
-            sumi0 += (v0 * y[ib].qs[j]);
-            sumi1 += (v1 * y[ib].qs[j + qk/2]);
-        }
-
-        int sumi = sumi0 + sumi1;
-        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);
-    }
-
-    *s = sumf;
-}
-
-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;
-
-    int ib = 0;
-    float sumf = 0;
-
-    assert(n % qk == 0);
-    assert(qk == QK5_0);
-    assert(nrc == 1);
-    UNUSED(nrc);
-    UNUSED(bx);
-    UNUSED(by);
-    UNUSED(bs);
-
-    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);
-    float32x4_t sumv1 = vdupq_n_f32(0.0f);
-
-    uint32_t qh0;
-    uint32_t qh1;
-
-    uint64_t tmp0[4];
-    uint64_t tmp1[4];
-
-    for (; ib + 1 < nb; ib += 2) {
-        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);
-
-        // extract the 5th bit via lookup table ((!b) << 4)
-        memcpy(&qh0, x0->qh, sizeof(qh0));
-        memcpy(&qh1, x1->qh, sizeof(qh1));
-
-        tmp0[0] = table_b2b_1[(qh0 >>  0) & 0xFF];
-        tmp0[1] = table_b2b_1[(qh0 >>  8) & 0xFF];
-        tmp0[2] = table_b2b_1[(qh0 >> 16) & 0xFF];
-        tmp0[3] = table_b2b_1[(qh0 >> 24)       ];
-
-        tmp1[0] = table_b2b_1[(qh1 >>  0) & 0xFF];
-        tmp1[1] = table_b2b_1[(qh1 >>  8) & 0xFF];
-        tmp1[2] = table_b2b_1[(qh1 >> 16) & 0xFF];
-        tmp1[3] = table_b2b_1[(qh1 >> 24)       ];
-
-        const int8x16_t qhl0 = vld1q_s8((const int8_t *)(tmp0 + 0));
-        const int8x16_t qhh0 = vld1q_s8((const int8_t *)(tmp0 + 2));
-        const int8x16_t qhl1 = vld1q_s8((const int8_t *)(tmp1 + 0));
-        const int8x16_t qhh1 = vld1q_s8((const int8_t *)(tmp1 + 2));
-
-        const uint8x16_t v0_0 = vld1q_u8(x0->qs);
-        const uint8x16_t v0_1 = vld1q_u8(x1->qs);
-
-        // 4-bit -> 8-bit
-        int8x16_t v0_0l = vreinterpretq_s8_u8(vandq_u8  (v0_0, m4b));
-        int8x16_t v0_0h = vreinterpretq_s8_u8(vshrq_n_u8(v0_0, 4));
-        int8x16_t v0_1l = vreinterpretq_s8_u8(vandq_u8  (v0_1, m4b));
-        int8x16_t v0_1h = vreinterpretq_s8_u8(vshrq_n_u8(v0_1, 4));
-
-        // add high bit and sub 16 (equivalent to sub 0x10 when bit is zero)
-        const int8x16_t v0_0lf = vsubq_s8(v0_0l, qhl0);
-        const int8x16_t v0_0hf = vsubq_s8(v0_0h, qhh0);
-        const int8x16_t v0_1lf = vsubq_s8(v0_1l, qhl1);
-        const int8x16_t v0_1hf = vsubq_s8(v0_1h, qhh1);
-
-        // load y
-        const int8x16_t v1_0l = vld1q_s8(y0->qs);
-        const int8x16_t v1_0h = vld1q_s8(y0->qs + 16);
-        const int8x16_t v1_1l = vld1q_s8(y1->qs);
-        const int8x16_t v1_1h = vld1q_s8(y1->qs + 16);
-
-        sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(
-                        ggml_vdotq_s32(vdupq_n_s32(0), v0_0lf, v1_0l),
-                        ggml_vdotq_s32(vdupq_n_s32(0), v0_0hf, v1_0h))), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
-        sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(
-                        ggml_vdotq_s32(vdupq_n_s32(0), v0_1lf, v1_1l),
-                        ggml_vdotq_s32(vdupq_n_s32(0), v0_1hf, v1_1h))), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
-    }
-
-    sumf = vaddvq_f32(sumv0) + vaddvq_f32(sumv1);
-#elif defined __wasm_simd128__
-    v128_t sumv = wasm_f32x4_splat(0.0f);
-
-    uint32_t qh_;
-    uint64_t tmp[4];
-
-    // TODO: check if unrolling this is better
-    for (; ib < nb; ++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_));
-
-        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);
-
-        const v128_t v0 = wasm_v128_load(x0->qs);
-
-        // 4-bit -> 8-bit
-        const v128_t v0l = wasm_v128_and (v0, m4b);
-        const v128_t v0h = wasm_u8x16_shr(v0, 4);
-
-        // add high bit and sub 16 (equivalent to sub 0x10 when bit is zero)
-        const v128_t v0lf = wasm_i8x16_sub(v0l, qhl);
-        const v128_t v0hf = wasm_i8x16_sub(v0h, qhh);
-
-        // load y
-        const v128_t v1l = wasm_v128_load(y0->qs);
-        const v128_t v1h = wasm_v128_load(y0->qs + 16);
-
-        // int8x16 -> int16x8
-        const v128_t v0lfl = wasm_i16x8_extend_low_i8x16 (v0lf);
-        const v128_t v0lfh = wasm_i16x8_extend_high_i8x16(v0lf);
-        const v128_t v0hfl = wasm_i16x8_extend_low_i8x16 (v0hf);
-        const v128_t v0hfh = wasm_i16x8_extend_high_i8x16(v0hf);
-
-        const v128_t v1ll = wasm_i16x8_extend_low_i8x16 (v1l);
-        const v128_t v1lh = wasm_i16x8_extend_high_i8x16(v1l);
-        const v128_t v1hl = wasm_i16x8_extend_low_i8x16 (v1h);
-        const v128_t v1hh = wasm_i16x8_extend_high_i8x16(v1h);
-
-        // dot product
-        sumv = wasm_f32x4_add(sumv, wasm_f32x4_mul(wasm_f32x4_convert_i32x4(
-                        wasm_i32x4_add(
-                            wasm_i32x4_add(wasm_i32x4_dot_i16x8(v0lfl, v1ll),
-                                           wasm_i32x4_dot_i16x8(v0lfh, v1lh)),
-                            wasm_i32x4_add(wasm_i32x4_dot_i16x8(v0hfl, v1hl),
-                                           wasm_i32x4_dot_i16x8(v0hfh, v1hh)))),
-                    wasm_f32x4_splat(GGML_FP16_TO_FP32(x0->d) * GGML_FP16_TO_FP32(y0->d))));
-    }
-
-    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();
-
-    // Main loop
-    for (; ib < nb; ++ib) {
-        /* Compute combined scale for the block */
-        const __m256 d = _mm256_set1_ps(GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d));
-
-        __m256i qx = bytes_from_nibbles_32(x[ib].qs);
-        __m256i bxhi = bytes_from_bits_32(x[ib].qh);
-        bxhi = _mm256_andnot_si256(bxhi, _mm256_set1_epi8((char)0xF0));
-        qx = _mm256_or_si256(qx, bxhi);
-
-        __m256i qy = _mm256_loadu_si256((const __m256i *)y[ib].qs);
-
-        const __m256 q = mul_sum_i8_pairs_float(qx, qy);
-
-        /* Multiply q with scale and accumulate */
-        acc = _mm256_fmadd_ps(d, q, acc);
-    }
-
-    sumf = hsum_float_8(acc);
-#elif defined(__AVX__)
-    // Initialize accumulator with zeros
-    __m256 acc = _mm256_setzero_ps();
-    __m128i mask = _mm_set1_epi8((char)0xF0);
-
-    // Main loop
-    for (; ib < nb; ++ib) {
-        /* Compute combined scale for the block */
-        const __m256 d = _mm256_set1_ps(GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d));
-
-        __m256i bx_0 = bytes_from_nibbles_32(x[ib].qs);
-        const __m256i bxhi = bytes_from_bits_32(x[ib].qh);
-        __m128i bxhil = _mm256_castsi256_si128(bxhi);
-        __m128i bxhih = _mm256_extractf128_si256(bxhi, 1);
-        bxhil = _mm_andnot_si128(bxhil, mask);
-        bxhih = _mm_andnot_si128(bxhih, mask);
-        __m128i bxl = _mm256_castsi256_si128(bx_0);
-        __m128i bxh = _mm256_extractf128_si256(bx_0, 1);
-        bxl = _mm_or_si128(bxl, bxhil);
-        bxh = _mm_or_si128(bxh, bxhih);
-        bx_0 = MM256_SET_M128I(bxh, bxl);
-
-        const __m256i by_0 = _mm256_loadu_si256((const __m256i *)y[ib].qs);
-
-        const __m256 q = mul_sum_i8_pairs_float(bx_0, by_0);
-
-        /* Multiply q with scale and accumulate */
-        acc = _mm256_add_ps(_mm256_mul_ps(d, q), acc);
-    }
-
-    sumf = hsum_float_8(acc);
-#elif defined(__riscv_v)
-    size_t vl;
-    size_t vlenb = __riscv_vlenb();
-
-    for (; ib < nb; ++ib) {
-        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);
-        }
-
-        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);
-
-        sumf += (GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d)) * sumi;
-    }
-
-#elif defined(__POWER9_VECTOR__)
-    const vector signed char lowMask = vec_splats((signed char)0xF);
-    const vector unsigned char v4 = vec_splats((unsigned char)4);
-
-    vector float vsumf0 = vec_splats(0.0f);
-
-#pragma GCC unroll 4
-    for (; ib < nb; ++ib) {
-        __builtin_prefetch(x[ib].qs, 0, 1);
-        __builtin_prefetch(y[ib].qs, 0, 1);
-
-        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[ib].d));
-        vector float vyd = vec_splats(GGML_FP16_TO_FP32(y[ib].d));
-        vector float vd = vec_mul(vxd, vyd);
-
-        vector signed long long aux64x2_0 = {(uint64_t)(table_b2b_1[x[ib].qh[0]]), (uint64_t)(table_b2b_1[x[ib].qh[1]])};
-        vector signed long long aux64x2_1 = {(uint64_t)(table_b2b_1[x[ib].qh[2]]), (uint64_t)(table_b2b_1[x[ib].qh[3]])};
-
-        vector signed char qh0 = (vector signed char)aux64x2_0;
-        vector signed char qh1 = (vector signed char)aux64x2_1;
-
-        vector signed char qxs = (vector signed char)vec_xl( 0, x[ib].qs);
-
-        vector signed char q5x0 = vec_sub(vec_and (qxs, lowMask), qh0);
-        vector signed char q5x1 = vec_sub(vec_sr(qxs, v4), qh1);
-
-        vector signed char q8y0 = vec_xl(  0, y[ib].qs);
-        vector signed char q8y1 = vec_xl( 16, y[ib].qs);
-
-        vector signed short qv0 = vec_add(vec_mule(q5x0, q8y0), vec_mulo(q5x0, q8y0));
-        vector signed short qv1 = vec_add(vec_mule(q5x1, q8y1), vec_mulo(q5x1, q8y1));
-
-        qv0 = vec_add(qv0, qv1);
-
-        vector signed int vsumi0 = vec_add(vec_unpackh(qv0), vec_unpackl(qv0));
-
-        vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
-    }
-
-    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4));
-    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8));
-
-    sumf = vec_extract(vsumf0, 0);
-
-#elif defined(__loongarch_asx)
-    // Initialize accumulator with zeros
-    __m256 acc = (__m256)__lasx_xvldi(0);
-
-    // Main loop
-    for (; ib < nb; ++ib) {
-        /* Compute combined scale for the block */
-        const __m256 d = __lasx_xvreplfr2vr_s(GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d)); //FIXME
-
-        __m256i qx = bytes_from_nibbles_32(x[ib].qs);
-        __m256i bxhi = bytes_from_bits_32(x[ib].qh);
-        bxhi = __lasx_xvandn_v(bxhi, __lasx_xvreplgr2vr_b((char)0xF0));
-        qx = __lasx_xvor_v(qx, bxhi);
-
-        __m256i qy = __lasx_xvld((const __m256i *)y[ib].qs, 0);
-
-        const __m256 q = mul_sum_i8_pairs_float(qx, qy);
-
-        /* Multiply q with scale and accumulate */
-        acc = __lasx_xvfmadd_s(d, q, acc);
-    }
-
-    sumf = hsum_float_8(acc);
-#endif
-    for (; ib < nb; ++ib) {
-        uint32_t qh;
-        memcpy(&qh, x[ib].qh, sizeof(qh));
-
-        int sumi0 = 0;
-        int sumi1 = 0;
-
-        for (int j = 0; j < qk/2; ++j) {
-            const uint8_t xh_0 = ((qh & (1u << (j + 0 ))) >> (j + 0 )) << 4;
-            const uint8_t xh_1 = ((qh & (1u << (j + 16))) >> (j + 12));
-
-            const int32_t x0 = (int8_t)(((x[ib].qs[j] & 0x0F) | xh_0) - 16);
-            const int32_t x1 = (int8_t)(((x[ib].qs[j] >>   4) | xh_1) - 16);
-
-            sumi0 += (x0 * y[ib].qs[j]);
-            sumi1 += (x1 * y[ib].qs[j + qk/2]);
-        }
-
-        int sumi = sumi0 + sumi1;
-        sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d)) * sumi;
-    }
-
-    *s = sumf;
-}
-
-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;
-
-    int ib = 0;
-    float sumf = 0;
-
-    assert(n % qk == 0);
-    assert(qk == QK5_1);
-    assert(nrc == 1);
-    UNUSED(nrc);
-    UNUSED(bx);
-    UNUSED(by);
-    UNUSED(bs);
-
-    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);
-    float32x4_t sumv1 = vdupq_n_f32(0.0f);
-
-    float summs0 = 0.0f;
-    float summs1 = 0.0f;
-
-    uint32_t qh0;
-    uint32_t qh1;
-
-    uint64_t tmp0[4];
-    uint64_t tmp1[4];
-
-    for (; ib + 1 < nb; ib += 2) {
-        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);
-
-        summs0 += GGML_FP16_TO_FP32(x0->m) * GGML_FP16_TO_FP32(y0->s);
-        summs1 += GGML_FP16_TO_FP32(x1->m) * GGML_FP16_TO_FP32(y1->s);
-
-        // extract the 5th bit via lookup table ((b) << 4)
-        memcpy(&qh0, x0->qh, sizeof(qh0));
-        memcpy(&qh1, x1->qh, sizeof(qh1));
-
-        tmp0[0] = table_b2b_0[(qh0 >>  0) & 0xFF];
-        tmp0[1] = table_b2b_0[(qh0 >>  8) & 0xFF];
-        tmp0[2] = table_b2b_0[(qh0 >> 16) & 0xFF];
-        tmp0[3] = table_b2b_0[(qh0 >> 24)       ];
-
-        tmp1[0] = table_b2b_0[(qh1 >>  0) & 0xFF];
-        tmp1[1] = table_b2b_0[(qh1 >>  8) & 0xFF];
-        tmp1[2] = table_b2b_0[(qh1 >> 16) & 0xFF];
-        tmp1[3] = table_b2b_0[(qh1 >> 24)       ];
-
-        const int8x16_t qhl0 = vld1q_s8((const int8_t *)(tmp0 + 0));
-        const int8x16_t qhh0 = vld1q_s8((const int8_t *)(tmp0 + 2));
-        const int8x16_t qhl1 = vld1q_s8((const int8_t *)(tmp1 + 0));
-        const int8x16_t qhh1 = vld1q_s8((const int8_t *)(tmp1 + 2));
-
-        const uint8x16_t v0_0 = vld1q_u8(x0->qs);
-        const uint8x16_t v0_1 = vld1q_u8(x1->qs);
-
-        // 4-bit -> 8-bit
-        const int8x16_t v0_0l = vreinterpretq_s8_u8(vandq_u8  (v0_0, m4b));
-        const int8x16_t v0_0h = vreinterpretq_s8_u8(vshrq_n_u8(v0_0, 4));
-        const int8x16_t v0_1l = vreinterpretq_s8_u8(vandq_u8  (v0_1, m4b));
-        const int8x16_t v0_1h = vreinterpretq_s8_u8(vshrq_n_u8(v0_1, 4));
-
-        // add high bit
-        const int8x16_t v0_0lf = vorrq_s8(v0_0l, qhl0);
-        const int8x16_t v0_0hf = vorrq_s8(v0_0h, qhh0);
-        const int8x16_t v0_1lf = vorrq_s8(v0_1l, qhl1);
-        const int8x16_t v0_1hf = vorrq_s8(v0_1h, qhh1);
-
-        // load y
-        const int8x16_t v1_0l = vld1q_s8(y0->qs);
-        const int8x16_t v1_0h = vld1q_s8(y0->qs + 16);
-        const int8x16_t v1_1l = vld1q_s8(y1->qs);
-        const int8x16_t v1_1h = vld1q_s8(y1->qs + 16);
-
-        sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(
-                        ggml_vdotq_s32(vdupq_n_s32(0), v0_0lf, v1_0l),
-                        ggml_vdotq_s32(vdupq_n_s32(0), v0_0hf, v1_0h))), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
-        sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(
-                        ggml_vdotq_s32(vdupq_n_s32(0), v0_1lf, v1_1l),
-                        ggml_vdotq_s32(vdupq_n_s32(0), v0_1hf, v1_1h))), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
-    }
-
-    sumf = vaddvq_f32(sumv0) + vaddvq_f32(sumv1) + summs0 + summs1;
-#elif defined __wasm_simd128__
-    v128_t sumv = wasm_f32x4_splat(0.0f);
-
-    float summs = 0.0f;
-
-    uint32_t qh_;
-    uint64_t tmp[4];
-
-    // TODO: check if unrolling this is better
-    for (; ib < nb; ++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_));
-
-        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);
-
-        const v128_t v0 = wasm_v128_load(x0->qs);
-
-        // 4-bit -> 8-bit
-        const v128_t v0l = wasm_v128_and (v0, m4b);
-        const v128_t v0h = wasm_u8x16_shr(v0, 4);
-
-        // add high bit
-        const v128_t v0lf = wasm_v128_or(v0l, qhl);
-        const v128_t v0hf = wasm_v128_or(v0h, qhh);
-
-        // load y
-        const v128_t v1l = wasm_v128_load(y0->qs);
-        const v128_t v1h = wasm_v128_load(y0->qs + 16);
-
-        // int8x16 -> int16x8
-        const v128_t v0lfl = wasm_i16x8_extend_low_i8x16 (v0lf);
-        const v128_t v0lfh = wasm_i16x8_extend_high_i8x16(v0lf);
-        const v128_t v0hfl = wasm_i16x8_extend_low_i8x16 (v0hf);
-        const v128_t v0hfh = wasm_i16x8_extend_high_i8x16(v0hf);
-
-        const v128_t v1ll = wasm_i16x8_extend_low_i8x16 (v1l);
-        const v128_t v1lh = wasm_i16x8_extend_high_i8x16(v1l);
-        const v128_t v1hl = wasm_i16x8_extend_low_i8x16 (v1h);
-        const v128_t v1hh = wasm_i16x8_extend_high_i8x16(v1h);
-
-        // dot product
-        sumv = wasm_f32x4_add(sumv,
-                wasm_f32x4_mul(wasm_f32x4_convert_i32x4(wasm_i32x4_add(
-                            wasm_i32x4_add(wasm_i32x4_dot_i16x8(v0lfl, v1ll),
-                                           wasm_i32x4_dot_i16x8(v0lfh, v1lh)),
-                            wasm_i32x4_add(wasm_i32x4_dot_i16x8(v0hfl, v1hl),
-                                           wasm_i32x4_dot_i16x8(v0hfh, v1hh)))),
-                    wasm_f32x4_splat(GGML_FP16_TO_FP32(x0->d) * GGML_FP16_TO_FP32(y0->d))));
-    }
-
-    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) + summs;
-#elif defined(__AVX2__)
-    // Initialize accumulator with zeros
-    __m256 acc = _mm256_setzero_ps();
-
-    float summs = 0.0f;
-
-    // Main loop
-    for (; ib < nb; ++ib) {
-        const __m256 dx = _mm256_set1_ps(GGML_FP16_TO_FP32(x[ib].d));
-
-        summs += GGML_FP16_TO_FP32(x[ib].m) * GGML_FP16_TO_FP32(y[ib].s);
-
-        __m256i qx = bytes_from_nibbles_32(x[ib].qs);
-        __m256i bxhi = bytes_from_bits_32(x[ib].qh);
-        bxhi = _mm256_and_si256(bxhi, _mm256_set1_epi8(0x10));
-        qx = _mm256_or_si256(qx, bxhi);
-
-        const __m256 dy = _mm256_set1_ps(GGML_FP16_TO_FP32(y[ib].d));
-        const __m256i qy = _mm256_loadu_si256((const __m256i *)y[ib].qs);
-
-        const __m256 q = mul_sum_us8_pairs_float(qx, qy);
-
-        acc = _mm256_fmadd_ps(q, _mm256_mul_ps(dx, dy), acc);
-    }
-
-    sumf = hsum_float_8(acc) + summs;
-#elif defined(__AVX__)
-    // Initialize accumulator with zeros
-    __m256 acc = _mm256_setzero_ps();
-    __m128i mask = _mm_set1_epi8(0x10);
-
-    float summs = 0.0f;
-
-    // Main loop
-    for (; ib < nb; ++ib) {
-        const __m256 dx = _mm256_set1_ps(GGML_FP16_TO_FP32(x[ib].d));
-
-        summs += GGML_FP16_TO_FP32(x[ib].m) * GGML_FP16_TO_FP32(y[ib].s);
-
-        __m256i bx_0 = bytes_from_nibbles_32(x[ib].qs);
-        const __m256i bxhi = bytes_from_bits_32(x[ib].qh);
-        __m128i bxhil = _mm256_castsi256_si128(bxhi);
-        __m128i bxhih = _mm256_extractf128_si256(bxhi, 1);
-        bxhil = _mm_and_si128(bxhil, mask);
-        bxhih = _mm_and_si128(bxhih, mask);
-        __m128i bxl = _mm256_castsi256_si128(bx_0);
-        __m128i bxh = _mm256_extractf128_si256(bx_0, 1);
-        bxl = _mm_or_si128(bxl, bxhil);
-        bxh = _mm_or_si128(bxh, bxhih);
-        bx_0 = MM256_SET_M128I(bxh, bxl);
-
-        const __m256 dy = _mm256_set1_ps(GGML_FP16_TO_FP32(y[ib].d));
-        const __m256i by_0 = _mm256_loadu_si256((const __m256i *)y[ib].qs);
-
-        const __m256 q = mul_sum_us8_pairs_float(bx_0, by_0);
-
-        acc = _mm256_add_ps(_mm256_mul_ps(q, _mm256_mul_ps(dx, dy)), acc);
-    }
-
-    sumf = hsum_float_8(acc) + summs;
-#elif defined(__riscv_v)
-    size_t vl;
-    size_t vlenb = __riscv_vlenb();
-
-    for (; ib < nb; ++ib) {
-        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);
-        }
-
-        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);
-    }
-
-#elif defined(__POWER9_VECTOR__)
-    const vector signed char lowMask = vec_splats((signed char)0xF);
-    const vector signed int v0 = vec_splats((int32_t)0);
-    const vector unsigned char v4 = vec_splats((unsigned char)0x4);
-
-    vector float vsumf0 = vec_splats(0.0f);
-
-#pragma GCC unroll 4
-    for (; ib < nb; ++ib) {
-        __builtin_prefetch(x[ib].qs, 0, 1);
-        __builtin_prefetch(y[ib].qs, 0, 1);
-
-        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[ib].d));
-        vector float vyd = vec_splats(GGML_FP16_TO_FP32(y[ib].d));
-        vector float vd = vec_mul(vxd, vyd);
-
-        vector float vxmin = vec_splats(GGML_FP16_TO_FP32(x[ib].m));
-        vector float vys = {GGML_FP16_TO_FP32(y[ib].s), 0.f, 0.f, 0.f};
-        vsumf0 = vec_madd(vxmin, vys, vsumf0);
-
-        vector unsigned long long aux64x2_0 = {(uint64_t)(table_b2b_0[x[ib].qh[0]]), (uint64_t)(table_b2b_0[x[ib].qh[1]])};
-        vector unsigned long long aux64x2_1 = {(uint64_t)(table_b2b_0[x[ib].qh[2]]), (uint64_t)(table_b2b_0[x[ib].qh[3]])};
-
-        vector signed char qh0 = (vector signed char)aux64x2_0;
-        vector signed char qh1 = (vector signed char)aux64x2_1;
-
-        vector signed char qxs = (vector signed char)vec_xl( 0, x[ib].qs);
-
-        vector unsigned char q5x0 = (vector unsigned char)vec_or(vec_and(qxs, lowMask), qh0);
-        vector unsigned char q5x1 = (vector unsigned char)vec_or(vec_sr(qxs, v4), qh1);
-
-        vector signed char q8y0 = vec_xl(  0, y[ib].qs);
-        vector signed char q8y1 = vec_xl( 16, y[ib].qs);
-
-        vector signed int vsumi0 = v0;
-
-        vsumi0 = vec_msum(q8y0, q5x0, vsumi0);
-        vsumi0 = vec_msum(q8y1, q5x1, vsumi0);
-
-        vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
-    }
-
-    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4));
-    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8));
-
-    sumf = vec_extract(vsumf0, 0);
-
-#elif defined(__loongarch_asx)
-    // Initialize accumulator with zeros
-    __m256 acc = (__m256)__lasx_xvldi(0);
-
-    float summs = 0.0f;
-
-    // Main loop
-    for (; ib < nb; ++ib) {
-        const __m256 dx = __lasx_xvreplfr2vr_s(GGML_FP16_TO_FP32(x[ib].d));
-
-        summs += GGML_FP16_TO_FP32(x[ib].m) * GGML_FP16_TO_FP32(y[ib].s);
-
-        __m256i qx = bytes_from_nibbles_32(x[ib].qs);
-        __m256i bxhi = bytes_from_bits_32(x[ib].qh);
-        bxhi = __lasx_xvand_v(bxhi, __lasx_xvreplgr2vr_b(0x10));
-        qx = __lasx_xvor_v(qx, bxhi);
-
-        const __m256 dy = __lasx_xvreplfr2vr_s(GGML_FP16_TO_FP32(y[ib].d));
-        const __m256i qy = __lasx_xvld((const __m256i *)y[ib].qs, 0);
-
-        const __m256 q = mul_sum_us8_pairs_float(qx, qy);
-
-        acc = __lasx_xvfmadd_s(q, __lasx_xvfmul_s(dx, dy), acc);
-    }
-
-    sumf = hsum_float_8(acc) + summs;
-#endif
-    for (; ib < nb; ++ib) {
-        uint32_t qh;
-        memcpy(&qh, x[ib].qh, sizeof(qh));
-
-        int sumi0 = 0;
-        int sumi1 = 0;
-
-        for (int j = 0; j < qk/2; ++j) {
-            const uint8_t xh_0 = ((qh >> (j +  0)) << 4) & 0x10;
-            const uint8_t xh_1 = ((qh >> (j + 12))     ) & 0x10;
-
-            const int32_t x0 = (x[ib].qs[j] & 0xF) | xh_0;
-            const int32_t x1 = (x[ib].qs[j] >>  4) | xh_1;
-
-            sumi0 += (x0 * y[ib].qs[j]);
-            sumi1 += (x1 * y[ib].qs[j + qk/2]);
-        }
-
-        int sumi = sumi0 + sumi1;
-        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);
-    }
-
-    *s = sumf;
-}
-
-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;
-
-    assert(n % qk == 0);
-#if defined(__ARM_FEATURE_MATMUL_INT8)
-    assert((nrc == 2) || (nrc == 1));
-#else
-    assert(nrc == 1);
-#endif
-    UNUSED(nrc);
-    UNUSED(bx);
-    UNUSED(by);
-    UNUSED(bs);
-
-    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 * 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 * GGML_RESTRICT b_x0 = &vx0[i];
-            const block_q8_0 * GGML_RESTRICT b_y0 = &vy0[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);
-            const int8x16_t x1_l = vld1q_s8(b_x1->qs);
-            const int8x16_t x1_h = vld1q_s8(b_x1->qs + 16);
-
-            // load y
-            const int8x16_t y0_l = vld1q_s8(b_y0->qs);
-            const int8x16_t y0_h = vld1q_s8(b_y0->qs + 16);
-            const int8x16_t y1_l = vld1q_s8(b_y1->qs);
-            const int8x16_t y1_h = vld1q_s8(b_y1->qs + 16);
-
-            float32_t _scale[4] = {
-                GGML_FP16_TO_FP32(b_x0->d)*GGML_FP16_TO_FP32(b_y0->d),
-                GGML_FP16_TO_FP32(b_x0->d)*GGML_FP16_TO_FP32(b_y1->d),
-                GGML_FP16_TO_FP32(b_x1->d)*GGML_FP16_TO_FP32(b_y0->d),
-                GGML_FP16_TO_FP32(b_x1->d)*GGML_FP16_TO_FP32(b_y1->d)
-            };
-            float32x4_t scale = vld1q_f32(_scale);
-
-            int8x16_t l0 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(x0_l), vreinterpretq_s64_s8(x1_l)));
-            int8x16_t l1 = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(x0_l), vreinterpretq_s64_s8(x1_l)));
-
-            int8x16_t l2 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(x0_h), vreinterpretq_s64_s8(x1_h)));
-            int8x16_t l3 = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(x0_h), vreinterpretq_s64_s8(x1_h)));
-
-            int8x16_t r0 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(y0_l), vreinterpretq_s64_s8(y1_l)));
-            int8x16_t r1 = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(y0_l), vreinterpretq_s64_s8(y1_l)));
-
-            int8x16_t r2 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(y0_h), vreinterpretq_s64_s8(y1_h)));
-            int8x16_t r3 = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(y0_h), vreinterpretq_s64_s8(y1_h)));
-
-            sumv0 = vmlaq_f32(sumv0,(vcvtq_f32_s32(vmmlaq_s32((vmmlaq_s32((vmmlaq_s32((vmmlaq_s32(vdupq_n_s32(0), l0, r0)),
-                                                l1, r1)), l2, r2)), l3, r3))), scale);
-        }
-
-        float32x4_t sumv1 = vextq_f32 (sumv0, sumv0, 2);
-        float32x4_t sumv2 = vzip1q_f32(sumv0, sumv1);
-
-        vst1_f32(s,      vget_low_f32 (sumv2));
-        vst1_f32(s + bs, vget_high_f32(sumv2));
-
-        return;
-    }
-#endif
-
-    int ib = 0;
-    float sumf = 0;
-
-#if defined(__ARM_FEATURE_SVE)
-    svfloat32_t sumv0 = svdup_n_f32(0.0f);
-    svfloat32_t sumv1 = svdup_n_f32(0.0f);
-
-    const int vector_length = ggml_cpu_get_sve_cnt()*8;
-
-    //VLA Implemenation for SVE
-    switch (vector_length) {
-        case 128:
-            {
-                // predicate for activating lanes for 16 Int8 elements
-                const svbool_t ph16 = svptrue_pat_b8 (SV_VL16);
-                const svbool_t pl16 = svptrue_pat_b32(SV_VL4);
-
-                for (; ib + 1 < nb; ib += 2) {
-                    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);
-                    const svint8_t qx0_1 = svld1_s8(ph16, x0->qs+16);
-                    const svint8_t qx1_0 = svld1_s8(ph16, x1->qs);
-                    const svint8_t qx1_1 = svld1_s8(ph16, x1->qs+16);
-
-                    // load y
-                    const svint8_t qy0_0 = svld1_s8(ph16, y0->qs);
-                    const svint8_t qy0_1 = svld1_s8(ph16, y0->qs+16);
-                    const svint8_t qy1_0 = svld1_s8(ph16, y1->qs);
-                    const svint8_t qy1_1 = svld1_s8(ph16, y1->qs+16);
-
-                    sumv0 = svmla_n_f32_x(pl16, sumv0, svcvt_f32_s32_x(pl16, svadd_x(pl16,
-                                    svdot_s32(svdup_n_s32(0), qx0_0, qy0_0),
-                                    svdot_s32(svdup_n_s32(0), qx0_1, qy0_1))), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
-                    sumv1 = svmla_n_f32_x(pl16, sumv1, svcvt_f32_s32_x(pl16, svadd_x(pl16,
-                                    svdot_s32(svdup_n_s32(0), qx1_0, qy1_0),
-                                    svdot_s32(svdup_n_s32(0), qx1_1, qy1_1))), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
-                }
-
-                sumf = svaddv_f32(pl16, svadd_f32_x(pl16, sumv0, sumv1));
-            } break;
-        case 256:
-            {
-                //printf("sve256");
-                for (; ib + 1 < nb; ib += 2) {
-                    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);
-                    const svint8_t qx1 = svld1_s8(svptrue_b8(), x1->qs);
-
-                    // load y
-                    const svint8_t qy0 = svld1_s8(svptrue_b8(), y0->qs);
-                    const svint8_t qy1 = svld1_s8(svptrue_b8(), y1->qs);
-
-                    sumv0 = svmla_n_f32_x(svptrue_b32(), sumv0, svcvt_f32_s32_x(svptrue_b32(),
-                                svdot_s32(svdup_n_s32(0), qx0, qy0)), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
-                    sumv1 = svmla_n_f32_x(svptrue_b32(), sumv1, svcvt_f32_s32_x(svptrue_b32(),
-                                svdot_s32(svdup_n_s32(0), qx1, qy1)), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
-                }
-
-                sumf = svaddv_f32(svptrue_b32(), svadd_f32_x(svptrue_b32(), sumv0, sumv1));
-            } break;
-        case 512:
-            {
-                // predicate for activating high 256 bit
-                const svbool_t ph32 = svptrue_pat_b8(SV_VL32);
-                // predicate for activating low 256 bit
-                const svbool_t pl32 = svnot_b_z(svptrue_b8(), ph32);
-
-                // predicate for activating high lanes for 8 float32 elements
-                const svbool_t ph8 = svptrue_pat_b32(SV_VL8);
-                // predicate for activating low lanes for 8 float32 elements
-                const svbool_t pl8 = svnot_b_z(svptrue_b32(), ph8);
-
-                svfloat32_t sumv00 = svdup_n_f32(0.0f);
-
-                for (; ib + 1 < nb; ib += 2) {
-                    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
-                    // load x
-                    const svint8_t qx_32 = svld1_s8(ph32, x0->qs);
-                          svint8_t qx_64 = svld1_s8(pl32, x0->qs + 2);
-
-                    qx_64 = svadd_s8_x(svptrue_b8(), qx_32, qx_64);
-
-                    // load y
-                    const svint8_t qy_32 = svld1_s8(ph32, y0->qs);
-                          svint8_t qy_64 = svld1_s8(pl32, y0->qs + 2);
-
-                    qy_64 = svadd_s8_x(svptrue_b8(), qy_32, qy_64);
-
-                    // scale creation
-                    const float32_t deq1 = GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d);
-                    const float32_t deq2 = GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d);
-
-                    // duplicate deq1 in first half of vector and deq2 in second half of vector
-                    const svfloat32_t temp = svdup_f32_m(svdup_f32_z(ph8, deq1), pl8, deq2);
-
-                    const svfloat32_t sumvt = svcvt_f32_s32_x(svptrue_b32(), svdot_s32(svdup_n_s32(0), qx_64, qy_64));
-
-                    sumv00 = svmla_f32_m(svptrue_b32(), sumv00, sumvt, temp);
-                }
-
-                sumf = svaddv_f32(svptrue_b32(), sumv00);
-                break;
-            }
-        default:
-            assert(false && "Unsupported vector length");
-            break;
-    }
-#elif defined(__ARM_NEON)
-    float32x4_t sumv0 = vdupq_n_f32(0.0f);
-    float32x4_t sumv1 = vdupq_n_f32(0.0f);
-
-    for (; ib + 1 < nb; ib += 2) {
-        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);
-        const int8x16_t x1_0 = vld1q_s8(x1->qs);
-        const int8x16_t x1_1 = vld1q_s8(x1->qs + 16);
-
-        // load y
-        const int8x16_t y0_0 = vld1q_s8(y0->qs);
-        const int8x16_t y0_1 = vld1q_s8(y0->qs + 16);
-        const int8x16_t y1_0 = vld1q_s8(y1->qs);
-        const int8x16_t y1_1 = vld1q_s8(y1->qs + 16);
-
-        sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(
-                        ggml_vdotq_s32(vdupq_n_s32(0), x0_0, y0_0),
-                        ggml_vdotq_s32(vdupq_n_s32(0), x0_1, y0_1))), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
-
-        sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(
-                        ggml_vdotq_s32(vdupq_n_s32(0), x1_0, y1_0),
-                        ggml_vdotq_s32(vdupq_n_s32(0), x1_1, y1_1))), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
-    }
-
-    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();
-
-    // Main loop
-    for (; ib < nb; ++ib) {
-        // Compute combined scale for the block
-        const __m256 d = _mm256_set1_ps(GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d));
-        __m256i qx = _mm256_loadu_si256((const __m256i *)x[ib].qs);
-        __m256i qy = _mm256_loadu_si256((const __m256i *)y[ib].qs);
-
-        const __m256 q = mul_sum_i8_pairs_float(qx, qy);
-
-        // Multiply q with scale and accumulate
-        acc = _mm256_fmadd_ps( d, q, acc );
-    }
-
-    sumf = hsum_float_8(acc);
-#elif defined(__AVX__)
-    __m256 accum = _mm256_setzero_ps();
-
-    for (; ib + 1 < nb; ib += 2) {
-        const __m128i qx_1_0 = _mm_loadu_si128((const __m128i *)x[ib].qs);
-        const __m128i qx_1_1 = _mm_loadu_si128((const __m128i *)x[ib].qs + 1);
-        const __m128i qx_2_0 = _mm_loadu_si128((const __m128i *)x[ib + 1].qs);
-        const __m128i qx_2_1 = _mm_loadu_si128((const __m128i *)x[ib + 1].qs + 1);
-        const __m128i qy_1_0 = _mm_loadu_si128((const __m128i *)y[ib].qs);
-        const __m128i qy_1_1 = _mm_loadu_si128((const __m128i *)y[ib].qs + 1);
-        const __m128i qy_2_0 = _mm_loadu_si128((const __m128i *)y[ib + 1].qs);
-        const __m128i qy_2_1 = _mm_loadu_si128((const __m128i *)y[ib + 1].qs + 1);
-
-        const __m256 p = mul_sum_i8_quad_float(qx_1_0, qx_1_1, qx_2_0, qx_2_1, qy_1_0, qy_1_1, qy_2_0, qy_2_1);
-        const __m256 deltas = quad_fp16_delta_float(x[ib].d, y[ib].d, x[ib + 1].d, y[ib + 1].d);
-        accum = _mm256_add_ps(_mm256_mul_ps(deltas, p), accum);
-    }
-
-    sumf = hsum_float_8(accum);
-#elif defined(__riscv_v)
-    size_t vl = qk;
-
-    for (; ib < nb; ++ib) {
-        // load elements
-        vint8m2_t bx_0 = __riscv_vle8_v_i8m2(x[ib].qs, vl);
-        vint8m2_t by_0 = __riscv_vle8_v_i8m2(y[ib].qs, 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_i16m4_i32m1(vw_mul, v_zero, vl);
-
-        int sumi = __riscv_vmv_x_s_i32m1_i32(v_sum);
-
-        sumf += sumi*(GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d));
-    }
-#elif defined(__POWER9_VECTOR__)
-    const vector signed int v0 = vec_splats((int32_t)0);
-    vector float vsumf0 = 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);
-
-        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[ib].d));
-        vector float vyd = vec_splats(GGML_FP16_TO_FP32(y[ib].d));
-        vector float vd = vec_mul(vxd, vyd);
-
-        vector signed char q8x0 = vec_xl( 0, x[ib].qs);
-        vector signed char q8x1 = vec_xl(16, x[ib].qs);
-        vector signed char q8y0 = vec_xl( 0, y[ib].qs);
-        vector signed char q8y1 = vec_xl(16, y[ib].qs);
-
-        vector signed short qv0 = vec_mule(q8x0, q8y0);
-        vector signed short qv1 = vec_mulo(q8x0, q8y0);
-        vector signed short qv2 = vec_mule(q8x1, q8y1);
-        vector signed short qv3 = vec_mulo(q8x1, q8y1);
-
-        vector signed int vsumi0 = v0;
-        vector signed int vsumi1 = v0;
-
-        vsumi0 = vec_sum4s(qv0, vsumi0);
-        vsumi1 = vec_sum4s(qv1, vsumi1);
-        vsumi0 = vec_sum4s(qv2, vsumi0);
-        vsumi1 = vec_sum4s(qv3, vsumi1);
-
-        vsumi0 = vec_add(vsumi0, vsumi1);
-
-        vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
-    }
-
-    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4));
-    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8));
-
-    sumf = vec_extract(vsumf0, 0);
-
-#elif defined(__loongarch_asx)
-    // Initialize accumulator with zeros
-    __m256 acc = (__m256)__lasx_xvldi(0);
-
-    // Main loop
-    for (; ib < nb; ++ib) {
-        // Compute combined scale for the block
-        const __m256 d = __lasx_xvreplfr2vr_s(GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d));
-        __m256i qx = __lasx_xvld((const __m256i *)x[ib].qs, 0);
-        __m256i qy = __lasx_xvld((const __m256i *)y[ib].qs, 0);
-
-        const __m256 q = mul_sum_i8_pairs_float(qx, qy);
-
-        // Multiply q with scale and accumulate
-        acc = __lasx_xvfmadd_s( d, q, acc );
-    }
-
-    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;
-
-        for (int j = 0; j < qk; j++) {
-            sumi += x[ib].qs[j]*y[ib].qs[j];
-        }
-
-        sumf += sumi*(GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d));
-    }
-
-    *s = sumf;
-}
-
-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 * GGML_RESTRICT x = vx;
-    const block_q8_K  * GGML_RESTRICT y = vy;
-
-    const int nb = n / QK_K;
-
-#if defined(__ARM_NEON)
-    float sumf = 0.0f;
-
-    uint8_t k_shift[16] = {1, 1, 1, 1, 3, 3, 3, 3, 9, 9, 9, 9, 27, 27, 27, 27};
-
-    const uint8x16_t shift = vld1q_u8(k_shift);
-
-    for (int i = 0; i < nb; ++i) {
-#if defined(__ARM_FEATURE_DOTPROD)
-        int32x4_t sumi0 = vdupq_n_s32(0);
-        int32x4_t sumi1 = vdupq_n_s32(0);
-#else
-        int16x8_t sumi0 = vdupq_n_s16(0);
-        int16x8_t sumi1 = vdupq_n_s16(0);
-#endif
-
-        // first 32 bytes of 5 elements
-        {
-            uint8x16_t qx0 = vld1q_u8(x[i].qs + 0);
-            uint8x16_t qx1 = vld1q_u8(x[i].qs + 16);
-            uint8x16_t qx2 = vmulq_u8(qx0, vdupq_n_u8(3));
-            uint8x16_t qx3 = vmulq_u8(qx1, vdupq_n_u8(3));
-            uint8x16_t qx4 = vmulq_u8(qx0, vdupq_n_u8(9));
-            uint8x16_t qx5 = vmulq_u8(qx1, vdupq_n_u8(9));
-            uint8x16_t qx6 = vmulq_u8(qx0, vdupq_n_u8(27));
-            uint8x16_t qx7 = vmulq_u8(qx1, vdupq_n_u8(27));
-            uint8x16_t qx8 = vmulq_u8(qx0, vdupq_n_u8(81));
-            uint8x16_t qx9 = vmulq_u8(qx1, vdupq_n_u8(81));
-
-            // multiply by 3 and keep the 2 bits above 8 bits
-            int8x16_t sqx0 = vreinterpretq_s8_u8(vshrq_n_u8(vhaddq_u8(qx0, vshrq_n_u8(qx0, 1)), 6));
-            int8x16_t sqx1 = vreinterpretq_s8_u8(vshrq_n_u8(vhaddq_u8(qx1, vshrq_n_u8(qx1, 1)), 6));
-            int8x16_t sqx2 = vreinterpretq_s8_u8(vshrq_n_u8(vhaddq_u8(qx2, vshrq_n_u8(qx2, 1)), 6));
-            int8x16_t sqx3 = vreinterpretq_s8_u8(vshrq_n_u8(vhaddq_u8(qx3, vshrq_n_u8(qx3, 1)), 6));
-            int8x16_t sqx4 = vreinterpretq_s8_u8(vshrq_n_u8(vhaddq_u8(qx4, vshrq_n_u8(qx4, 1)), 6));
-            int8x16_t sqx5 = vreinterpretq_s8_u8(vshrq_n_u8(vhaddq_u8(qx5, vshrq_n_u8(qx5, 1)), 6));
-            int8x16_t sqx6 = vreinterpretq_s8_u8(vshrq_n_u8(vhaddq_u8(qx6, vshrq_n_u8(qx6, 1)), 6));
-            int8x16_t sqx7 = vreinterpretq_s8_u8(vshrq_n_u8(vhaddq_u8(qx7, vshrq_n_u8(qx7, 1)), 6));
-            int8x16_t sqx8 = vreinterpretq_s8_u8(vshrq_n_u8(vhaddq_u8(qx8, vshrq_n_u8(qx8, 1)), 6));
-            int8x16_t sqx9 = vreinterpretq_s8_u8(vshrq_n_u8(vhaddq_u8(qx9, vshrq_n_u8(qx9, 1)), 6));
-
-            const int8x16_t qy0 = vld1q_s8(y[i].qs +   0);
-            const int8x16_t qy1 = vld1q_s8(y[i].qs +  16);
-            const int8x16_t qy2 = vld1q_s8(y[i].qs +  32);
-            const int8x16_t qy3 = vld1q_s8(y[i].qs +  48);
-            const int8x16_t qy4 = vld1q_s8(y[i].qs +  64);
-            const int8x16_t qy5 = vld1q_s8(y[i].qs +  80);
-            const int8x16_t qy6 = vld1q_s8(y[i].qs +  96);
-            const int8x16_t qy7 = vld1q_s8(y[i].qs + 112);
-            const int8x16_t qy8 = vld1q_s8(y[i].qs + 128);
-            const int8x16_t qy9 = vld1q_s8(y[i].qs + 144);
-
-#if defined(__ARM_FEATURE_DOTPROD)
-            sumi0 = vdotq_s32(sumi0, sqx0, qy0);
-            sumi1 = vdotq_s32(sumi1, sqx1, qy1);
-            sumi0 = vdotq_s32(sumi0, sqx2, qy2);
-            sumi1 = vdotq_s32(sumi1, sqx3, qy3);
-            sumi0 = vdotq_s32(sumi0, sqx4, qy4);
-            sumi1 = vdotq_s32(sumi1, sqx5, qy5);
-            sumi0 = vdotq_s32(sumi0, sqx6, qy6);
-            sumi1 = vdotq_s32(sumi1, sqx7, qy7);
-            sumi0 = vdotq_s32(sumi0, sqx8, qy8);
-            sumi1 = vdotq_s32(sumi1, sqx9, qy9);
-#else
-            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx0), vget_low_s8(qy0));
-            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx0), vget_high_s8(qy0));
-            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx1), vget_low_s8(qy1));
-            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx1), vget_high_s8(qy1));
-            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx2), vget_low_s8(qy2));
-            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx2), vget_high_s8(qy2));
-            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx3), vget_low_s8(qy3));
-            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx3), vget_high_s8(qy3));
-            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx4), vget_low_s8(qy4));
-            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx4), vget_high_s8(qy4));
-            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx5), vget_low_s8(qy5));
-            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx5), vget_high_s8(qy5));
-            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx6), vget_low_s8(qy6));
-            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx6), vget_high_s8(qy6));
-            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx7), vget_low_s8(qy7));
-            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx7), vget_high_s8(qy7));
-            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx8), vget_low_s8(qy8));
-            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx8), vget_high_s8(qy8));
-            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx9), vget_low_s8(qy9));
-            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx9), vget_high_s8(qy9));
-#endif
-        }
-
-        // last 16 bytes of 5-element, along with the 4 bytes of 4 elements
-        {
-            uint8x16_t qx0 = vld1q_u8(x[i].qs + 32);
-            uint8x16_t qx1 = vmulq_u8(qx0, vdupq_n_u8(3));
-            uint8x16_t qx2 = vmulq_u8(qx0, vdupq_n_u8(9));
-            uint8x16_t qx3 = vmulq_u8(qx0, vdupq_n_u8(27));
-            uint8x16_t qx4 = vmulq_u8(qx0, vdupq_n_u8(81));
-            uint32_t qh;
-            memcpy(&qh, x[i].qh, sizeof(qh)); // potentially unaligned
-            uint8x16_t qx5 = vreinterpretq_u8_u32(vdupq_n_u32(qh));
-            qx5 = vmulq_u8(qx5, shift);
-
-            // multiply by 3 and keep the 2 bits above 8 bits
-            int8x16_t sqx0 = vreinterpretq_s8_u8(vshrq_n_u8(vhaddq_u8(qx0, vshrq_n_u8(qx0, 1)), 6));
-            int8x16_t sqx1 = vreinterpretq_s8_u8(vshrq_n_u8(vhaddq_u8(qx1, vshrq_n_u8(qx1, 1)), 6));
-            int8x16_t sqx2 = vreinterpretq_s8_u8(vshrq_n_u8(vhaddq_u8(qx2, vshrq_n_u8(qx2, 1)), 6));
-            int8x16_t sqx3 = vreinterpretq_s8_u8(vshrq_n_u8(vhaddq_u8(qx3, vshrq_n_u8(qx3, 1)), 6));
-            int8x16_t sqx4 = vreinterpretq_s8_u8(vshrq_n_u8(vhaddq_u8(qx4, vshrq_n_u8(qx4, 1)), 6));
-            int8x16_t sqx5 = vreinterpretq_s8_u8(vshrq_n_u8(vhaddq_u8(qx5, vshrq_n_u8(qx5, 1)), 6));
-
-            const int8x16_t qy0 = vld1q_s8(y[i].qs + 160);
-            const int8x16_t qy1 = vld1q_s8(y[i].qs + 176);
-            const int8x16_t qy2 = vld1q_s8(y[i].qs + 192);
-            const int8x16_t qy3 = vld1q_s8(y[i].qs + 208);
-            const int8x16_t qy4 = vld1q_s8(y[i].qs + 224);
-            const int8x16_t qy5 = vld1q_s8(y[i].qs + 240);
-
-#if defined(__ARM_FEATURE_DOTPROD)
-            sumi0 = vdotq_s32(sumi0, sqx0, qy0);
-            sumi1 = vdotq_s32(sumi1, sqx1, qy1);
-            sumi0 = vdotq_s32(sumi0, sqx2, qy2);
-            sumi1 = vdotq_s32(sumi1, sqx3, qy3);
-            sumi0 = vdotq_s32(sumi0, sqx4, qy4);
-            sumi1 = vdotq_s32(sumi1, sqx5, qy5);
-#else
-            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx0), vget_low_s8(qy0));
-            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx0), vget_high_s8(qy0));
-            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx1), vget_low_s8(qy1));
-            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx1), vget_high_s8(qy1));
-            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx2), vget_low_s8(qy2));
-            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx2), vget_high_s8(qy2));
-            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx3), vget_low_s8(qy3));
-            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx3), vget_high_s8(qy3));
-            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx4), vget_low_s8(qy4));
-            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx4), vget_high_s8(qy4));
-            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx5), vget_low_s8(qy5));
-            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx5), vget_high_s8(qy5));
-#endif
-        }
-
-        const int16x8_t ysum0 = vld1q_s16(y[i].bsums);
-        const int16x8_t ysum1 = vld1q_s16(y[i].bsums + 8);
-
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
-
-#if defined(__ARM_FEATURE_DOTPROD)
-        sumi0 = vaddq_s32(sumi0, sumi1);
-        sumi0 = vsubq_s32(sumi0, vpaddlq_s16(vaddq_s16(ysum0, ysum1)));
-
-        sumf += d * (float) vaddvq_s32(sumi0);
-#else
-        sumi0 = vaddq_s16(sumi0, sumi1);
-        sumi0 = vsubq_s16(sumi0, vaddq_s16(ysum0, ysum1));
-
-        sumf += d * (float) vaddlvq_s16(sumi0);
-#endif
-    }
-
-    *s = sumf;
-
-#elif defined(__AVX2__)
-    __m256 sumf = _mm256_setzero_ps();
-
-    for (int i = 0; i < nb; ++i) {
-        // 16-bit sums
-        __m256i sumi0 = _mm256_setzero_si256();
-        __m256i sumi1 = _mm256_setzero_si256();
-        __m256i sumi2 = _mm256_setzero_si256();
-
-        // first 32 bytes of 5 elements
-        {
-            __m256i qx0 = _mm256_loadu_si256((const __m256i *) (x[i].qs));
-            // 8-bit multiplies with shifts, masks and adds
-            __m256i qx1 = _mm256_add_epi8(qx0, _mm256_add_epi8(qx0, qx0)); // 1 * 3
-            __m256i qx2 = _mm256_add_epi8(_mm256_and_si256(_mm256_slli_epi16(qx0, 3), _mm256_set1_epi8(-8)), qx0); // 1 * 9
-            __m256i qx3 = _mm256_add_epi8(_mm256_and_si256(_mm256_slli_epi16(qx1, 3), _mm256_set1_epi8(-8)), qx1); // 3 * 9
-            __m256i qx4 = _mm256_add_epi8(_mm256_and_si256(_mm256_slli_epi16(qx2, 3), _mm256_set1_epi8(-8)), qx2); // 9 * 9
-
-            // TODO: can _mm256_mulhi_epu16 be faster even if 16-bits?
-
-            // Cancel the +1 from avg so that it behaves like a halving add
-            qx0 = _mm256_subs_epu8(qx0, _mm256_set1_epi8(1));
-            qx1 = _mm256_subs_epu8(qx1, _mm256_set1_epi8(1));
-            qx2 = _mm256_subs_epu8(qx2, _mm256_set1_epi8(1));
-            qx3 = _mm256_subs_epu8(qx3, _mm256_set1_epi8(1));
-            qx4 = _mm256_subs_epu8(qx4, _mm256_set1_epi8(1));
-            // Multiply by 3 and get the top 2 bits
-            qx0 = _mm256_avg_epu8(qx0, _mm256_avg_epu8(qx0, _mm256_setzero_si256()));
-            qx1 = _mm256_avg_epu8(qx1, _mm256_avg_epu8(qx1, _mm256_setzero_si256()));
-            qx2 = _mm256_avg_epu8(qx2, _mm256_avg_epu8(qx2, _mm256_setzero_si256()));
-            qx3 = _mm256_avg_epu8(qx3, _mm256_avg_epu8(qx3, _mm256_setzero_si256()));
-            qx4 = _mm256_avg_epu8(qx4, _mm256_avg_epu8(qx4, _mm256_setzero_si256()));
-            qx0 = _mm256_and_si256(_mm256_srli_epi16(qx0, 6), _mm256_set1_epi8(3));
-            qx1 = _mm256_and_si256(_mm256_srli_epi16(qx1, 6), _mm256_set1_epi8(3));
-            qx2 = _mm256_and_si256(_mm256_srli_epi16(qx2, 6), _mm256_set1_epi8(3));
-            qx3 = _mm256_and_si256(_mm256_srli_epi16(qx3, 6), _mm256_set1_epi8(3));
-            qx4 = _mm256_and_si256(_mm256_srli_epi16(qx4, 6), _mm256_set1_epi8(3));
-
-            const __m256i qy0 = _mm256_loadu_si256((const __m256i *) (y[i].qs +   0));
-            const __m256i qy1 = _mm256_loadu_si256((const __m256i *) (y[i].qs +  32));
-            const __m256i qy2 = _mm256_loadu_si256((const __m256i *) (y[i].qs +  64));
-            const __m256i qy3 = _mm256_loadu_si256((const __m256i *) (y[i].qs +  96));
-            const __m256i qy4 = _mm256_loadu_si256((const __m256i *) (y[i].qs + 128));
-
-            qx0 = _mm256_maddubs_epi16(qx0, qy0);
-            qx1 = _mm256_maddubs_epi16(qx1, qy1);
-            qx2 = _mm256_maddubs_epi16(qx2, qy2);
-            qx3 = _mm256_maddubs_epi16(qx3, qy3);
-            qx4 = _mm256_maddubs_epi16(qx4, qy4);
-
-            sumi0 = _mm256_add_epi16(sumi0, _mm256_add_epi16(qx0, qx1));
-            sumi1 = _mm256_add_epi16(sumi1, _mm256_add_epi16(qx2, qx3));
-            sumi2 = _mm256_add_epi16(sumi2, qx4);
-        }
-
-        // last 16 bytes of 5-element, along with the 4 bytes of 4 elements
-        {
-            __m128i qx0 = _mm_loadu_si128((const __m128i *) (x[i].qs + 32));
-            uint32_t qh;
-            memcpy(&qh, x[i].qh, sizeof(qh)); // potentially unaligned
-            __m256i qx5_l = _mm256_cvtepu8_epi16(_mm_set1_epi32(qh));
-            __m128i qx1 = _mm_add_epi8(qx0, _mm_add_epi8(qx0, qx0)); // 1 * 3
-            __m128i qx2 = _mm_add_epi8(_mm_and_si128(_mm_slli_epi16(qx0, 3), _mm_set1_epi8(-8)), qx0); // 1 * 9
-            __m128i qx3 = _mm_add_epi8(_mm_and_si128(_mm_slli_epi16(qx1, 3), _mm_set1_epi8(-8)), qx1); // 3 * 9
-            __m128i qx4 = _mm_add_epi8(_mm_and_si128(_mm_slli_epi16(qx2, 3), _mm_set1_epi8(-8)), qx2); // 9 * 9
-            __m256i qx01 = MM256_SET_M128I(qx1, qx0);
-            __m256i qx23 = MM256_SET_M128I(qx3, qx2);
-
-            // avx2 does not have 8-bit multiplies, so 16-bit it is.
-            qx5_l = _mm256_mullo_epi16(qx5_l, _mm256_set_epi16(27, 27, 27, 27, 9, 9, 9, 9, 3, 3, 3, 3, 1, 1, 1, 1));
-            qx5_l = _mm256_and_si256(qx5_l, _mm256_set1_epi16(0xFF));
-            __m128i qx5 = _mm_packus_epi16(_mm256_castsi256_si128(qx5_l), _mm256_extracti128_si256(qx5_l, 1));
-
-            __m256i qx45 = MM256_SET_M128I(qx5, qx4);
-
-            // Cancel the +1 from avg so that it behaves like a halving add
-            qx01 = _mm256_subs_epu8(qx01, _mm256_set1_epi8(1));
-            qx23 = _mm256_subs_epu8(qx23, _mm256_set1_epi8(1));
-            qx45 = _mm256_subs_epu8(qx45, _mm256_set1_epi8(1));
-            // Multiply by 3 and get the top 2 bits
-            qx01 = _mm256_avg_epu8(qx01, _mm256_avg_epu8(qx01, _mm256_setzero_si256()));
-            qx23 = _mm256_avg_epu8(qx23, _mm256_avg_epu8(qx23, _mm256_setzero_si256()));
-            qx45 = _mm256_avg_epu8(qx45, _mm256_avg_epu8(qx45, _mm256_setzero_si256()));
-            qx01 = _mm256_and_si256(_mm256_srli_epi16(qx01, 6), _mm256_set1_epi8(3));
-            qx23 = _mm256_and_si256(_mm256_srli_epi16(qx23, 6), _mm256_set1_epi8(3));
-            qx45 = _mm256_and_si256(_mm256_srli_epi16(qx45, 6), _mm256_set1_epi8(3));
-
-            const __m256i qy01 = _mm256_loadu_si256((const __m256i *) (y[i].qs + 160));
-            const __m256i qy23 = _mm256_loadu_si256((const __m256i *) (y[i].qs + 192));
-            const __m256i qy45 = _mm256_loadu_si256((const __m256i *) (y[i].qs + 224));
-
-            qx01 = _mm256_maddubs_epi16(qx01, qy01);
-            qx23 = _mm256_maddubs_epi16(qx23, qy23);
-            qx45 = _mm256_maddubs_epi16(qx45, qy45);
-
-            sumi0 = _mm256_add_epi16(sumi0, qx01);
-            sumi1 = _mm256_add_epi16(sumi1, qx23);
-            sumi2 = _mm256_add_epi16(sumi2, qx45);
-        }
-
-        const __m256i ysum = _mm256_loadu_si256((const __m256i *) y[i].bsums);
-        const __m256 d = _mm256_set1_ps(y[i].d * GGML_FP16_TO_FP32(x[i].d));
-
-        sumi0 = _mm256_sub_epi16(sumi0, ysum);
-        sumi0 = _mm256_add_epi16(sumi0, _mm256_add_epi16(sumi1, sumi2));
-        sumi0 = _mm256_madd_epi16(sumi0, _mm256_set1_epi16(1));
-
-        sumf = _mm256_add_ps(_mm256_mul_ps(_mm256_cvtepi32_ps(sumi0), d), sumf);
-    }
-
-    *s = hsum_float_8(sumf);
-
-#else
-    const uint8_t pow3[6] = {1, 3, 9, 27, 81, 243};
-
-    float sumf = 0.0f;
-
-    for (int i = 0; i < nb; ++i) {
-        int sum = 0;
-
-        for (size_t j = 0; j < sizeof(x->qs) - sizeof(x->qs) % 32; j += 32) {
-            for (size_t l = 0; l < 5; ++l) {
-                for (size_t m = 0; m < 32; ++m) {
-                    uint8_t q = x[i].qs[j + m] * pow3[l];
-                    uint16_t xi = ((uint16_t) q * 3) >> 8;
-                    sum += (xi - 1) * y[i].qs[j*5 + l*32 + m];
-                }
-            }
-        }
-        for (size_t j = sizeof(x->qs) - sizeof(x->qs) % 32; j < sizeof(x->qs); j += 16) {
-            for (size_t l = 0; l < 5; ++l) {
-                for (size_t m = 0; m < 16; ++m) {
-                    uint8_t q = x[i].qs[j + m] * pow3[l];
-                    uint16_t xi = ((uint16_t) q * 3) >> 8;
-                    sum += (xi - 1) * y[i].qs[j*5 + l*16 + m];
-                }
-            }
-        }
-
-        for (size_t l = 0; l < 4; ++l) {
-            for (size_t j = 0; j < sizeof(x->qh); ++j) {
-                uint8_t q = x[i].qh[j] * pow3[l];
-                uint16_t xi = ((uint16_t) q * 3) >> 8;
-                sum += (xi - 1) * y[i].qs[sizeof(x->qs)*5 + l*sizeof(x->qh) + j];
-            }
-        }
-
-        sumf += (float) sum * (GGML_FP16_TO_FP32(x[i].d) * y[i].d);
-    }
-
-    *s = sumf;
-#endif
-}
-
-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 * GGML_RESTRICT x = vx;
-    const block_q8_K  * GGML_RESTRICT y = vy;
-
-    const int nb = n / QK_K;
-
-#if defined(__ARM_NEON)
-    float sumf = 0.0f;
-
-    const uint8x16_t m3 = vdupq_n_u8(3);
-
-    for (int i = 0; i < nb; ++i) {
-#if defined(__ARM_FEATURE_DOTPROD)
-        int32x4_t sumi0 = vdupq_n_s32(0);
-        int32x4_t sumi1 = vdupq_n_s32(0);
-#else
-        int16x8_t sumi0 = vdupq_n_s16(0);
-        int16x8_t sumi1 = vdupq_n_s16(0);
-#endif
-
-        for (size_t j = 0; j < sizeof(x->qs); j += 32) {
-            uint8x16_t qx0 = vld1q_u8(x[i].qs + j);
-            uint8x16_t qx1 = vld1q_u8(x[i].qs + j + 16);
-            uint8x16_t qx2 = vshrq_n_u8(qx0, 2);
-            uint8x16_t qx3 = vshrq_n_u8(qx1, 2);
-            uint8x16_t qx4 = vshrq_n_u8(qx0, 4);
-            uint8x16_t qx5 = vshrq_n_u8(qx1, 4);
-            uint8x16_t qx6 = vshrq_n_u8(qx0, 6);
-            uint8x16_t qx7 = vshrq_n_u8(qx1, 6);
-
-            int8x16_t sqx0 = vreinterpretq_s8_u8(vandq_u8(qx0, m3));
-            int8x16_t sqx1 = vreinterpretq_s8_u8(vandq_u8(qx1, m3));
-            int8x16_t sqx2 = vreinterpretq_s8_u8(vandq_u8(qx2, m3));
-            int8x16_t sqx3 = vreinterpretq_s8_u8(vandq_u8(qx3, m3));
-            int8x16_t sqx4 = vreinterpretq_s8_u8(vandq_u8(qx4, m3));
-            int8x16_t sqx5 = vreinterpretq_s8_u8(vandq_u8(qx5, m3));
-            int8x16_t sqx6 = vreinterpretq_s8_u8(vandq_u8(qx6, m3));
-            int8x16_t sqx7 = vreinterpretq_s8_u8(vandq_u8(qx7, m3));
-
-            const int8x16_t qy0 = vld1q_s8(y[i].qs + j*4 +   0);
-            const int8x16_t qy1 = vld1q_s8(y[i].qs + j*4 +  16);
-            const int8x16_t qy2 = vld1q_s8(y[i].qs + j*4 +  32);
-            const int8x16_t qy3 = vld1q_s8(y[i].qs + j*4 +  48);
-            const int8x16_t qy4 = vld1q_s8(y[i].qs + j*4 +  64);
-            const int8x16_t qy5 = vld1q_s8(y[i].qs + j*4 +  80);
-            const int8x16_t qy6 = vld1q_s8(y[i].qs + j*4 +  96);
-            const int8x16_t qy7 = vld1q_s8(y[i].qs + j*4 + 112);
-
-#if defined(__ARM_FEATURE_DOTPROD)
-            sumi0 = vdotq_s32(sumi0, sqx0, qy0);
-            sumi1 = vdotq_s32(sumi1, sqx1, qy1);
-            sumi0 = vdotq_s32(sumi0, sqx2, qy2);
-            sumi1 = vdotq_s32(sumi1, sqx3, qy3);
-            sumi0 = vdotq_s32(sumi0, sqx4, qy4);
-            sumi1 = vdotq_s32(sumi1, sqx5, qy5);
-            sumi0 = vdotq_s32(sumi0, sqx6, qy6);
-            sumi1 = vdotq_s32(sumi1, sqx7, qy7);
-#else
-            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx0), vget_low_s8(qy0));
-            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx0), vget_high_s8(qy0));
-            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx1), vget_low_s8(qy1));
-            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx1), vget_high_s8(qy1));
-            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx2), vget_low_s8(qy2));
-            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx2), vget_high_s8(qy2));
-            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx3), vget_low_s8(qy3));
-            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx3), vget_high_s8(qy3));
-            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx4), vget_low_s8(qy4));
-            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx4), vget_high_s8(qy4));
-            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx5), vget_low_s8(qy5));
-            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx5), vget_high_s8(qy5));
-            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx6), vget_low_s8(qy6));
-            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx6), vget_high_s8(qy6));
-            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx7), vget_low_s8(qy7));
-            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx7), vget_high_s8(qy7));
-#endif
-        }
-
-        const int16x8_t ysum0 = vld1q_s16(y[i].bsums);
-        const int16x8_t ysum1 = vld1q_s16(y[i].bsums + 8);
-
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
-
-#if defined(__ARM_FEATURE_DOTPROD)
-        sumi0 = vaddq_s32(sumi0, sumi1);
-        sumi0 = vsubq_s32(sumi0, vpaddlq_s16(vaddq_s16(ysum0, ysum1)));
-
-        sumf += d * (float) vaddvq_s32(sumi0);
-#else
-        sumi0 = vaddq_s16(sumi0, sumi1);
-        sumi0 = vsubq_s16(sumi0, vaddq_s16(ysum0, ysum1));
-
-        sumf += d * (float) vaddlvq_s16(sumi0);
-#endif
-    }
-
-    *s = sumf;
-
-#elif defined(__AVX2__)
-    __m256 sumf = _mm256_setzero_ps();
-
-    for (int i = 0; i < nb; ++i) {
-        // 16-bit sums, because 256*127 still fits
-        __m256i sumi0 = _mm256_setzero_si256();
-        __m256i sumi1 = _mm256_setzero_si256();
-
-        for (size_t j = 0; j < sizeof(x->qs); j += 32) {
-            __m256i qx0 = _mm256_loadu_si256((const __m256i *) (x[i].qs + j));
-            __m256i qx1 = _mm256_srli_epi16(qx0, 2);
-            __m256i qx2 = _mm256_srli_epi16(qx0, 4);
-            __m256i qx3 = _mm256_srli_epi16(qx0, 6);
-
-            // 0, 1, 2 (should not be 3)
-            qx0 = _mm256_and_si256(qx0, _mm256_set1_epi8(3));
-            qx1 = _mm256_and_si256(qx1, _mm256_set1_epi8(3));
-            qx2 = _mm256_and_si256(qx2, _mm256_set1_epi8(3));
-            qx3 = _mm256_and_si256(qx3, _mm256_set1_epi8(3));
-
-            const __m256i qy0 = _mm256_loadu_si256((const __m256i *) (y[i].qs + j*4 +  0));
-            const __m256i qy1 = _mm256_loadu_si256((const __m256i *) (y[i].qs + j*4 + 32));
-            const __m256i qy2 = _mm256_loadu_si256((const __m256i *) (y[i].qs + j*4 + 64));
-            const __m256i qy3 = _mm256_loadu_si256((const __m256i *) (y[i].qs + j*4 + 96));
-
-            qx0 = _mm256_maddubs_epi16(qx0, qy0);
-            qx1 = _mm256_maddubs_epi16(qx1, qy1);
-            qx2 = _mm256_maddubs_epi16(qx2, qy2);
-            qx3 = _mm256_maddubs_epi16(qx3, qy3);
-
-            sumi0 = _mm256_add_epi16(sumi0, _mm256_add_epi16(qx0, qx1));
-            sumi1 = _mm256_add_epi16(sumi1, _mm256_add_epi16(qx2, qx3));
-        }
-
-        const __m256i ysum = _mm256_loadu_si256((const __m256i *) y[i].bsums);
-        const __m256 d = _mm256_set1_ps(y[i].d * GGML_FP16_TO_FP32(x[i].d));
-
-        sumi0 = _mm256_add_epi16(sumi0, sumi1);
-        sumi0 = _mm256_sub_epi16(sumi0, ysum);
-        sumi0 = _mm256_madd_epi16(sumi0, _mm256_set1_epi16(1));
-
-        sumf = _mm256_add_ps(_mm256_mul_ps(_mm256_cvtepi32_ps(sumi0), d), sumf);
-    }
-
-    *s = hsum_float_8(sumf);
-
-#else
-    float sumf = 0.0f;
-
-    for (int i = 0; i < nb; ++i) {
-        int32_t sumi = 0;
-
-        for (size_t j = 0; j < sizeof(x->qs); j += 32) {
-            for (size_t l = 0; l < 4; ++l) {
-                for (size_t k = 0; k < 32; ++k) {
-                    sumi += y[i].qs[j*4 + l*32 + k] * (((x[i].qs[j + k] >> (l*2)) & 3) - 1);
-                }
-            }
-        }
-
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-
-        sumf += (float) sumi * d;
-    }
-
-    *s = sumf;
-#endif
-}
-
-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 * GGML_RESTRICT x = vx;
-    const block_q8_K * GGML_RESTRICT y = vy;
-
-    const int nb = n / QK_K;
-
-#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);
-
-    const int32x4_t vzero = vdupq_n_s32(0);
-
-    ggml_int8x16x2_t q2bytes;
-    uint8_t aux[16];
-
-    float sum = 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 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);
-        vst1q_u8(aux, scales);
-
-        const uint8x16_t mins = vshrq_n_u8(mins_and_scales, 4);
-        const ggml_int16x8x2_t q8sums = ggml_vld1q_s16_x2(y[i].bsums);
-        const ggml_int16x8x2_t mins16 = {{vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(mins))), vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(mins)))}};
-        const int32x4_t s0 = vaddq_s32(vmull_s16(vget_low_s16 (mins16.val[0]), vget_low_s16 (q8sums.val[0])),
-                                       vmull_s16(vget_high_s16(mins16.val[0]), vget_high_s16(q8sums.val[0])));
-        const int32x4_t s1 = vaddq_s32(vmull_s16(vget_low_s16 (mins16.val[1]), vget_low_s16 (q8sums.val[1])),
-                                       vmull_s16(vget_high_s16(mins16.val[1]), vget_high_s16(q8sums.val[1])));
-        sum += dmin * vaddvq_s32(vaddq_s32(s0, s1));
-
-        int isum = 0;
-        int is = 0;
-
-// We use this macro instead of a function call because for some reason
-// the code runs 2-3% slower, even if the function is declared inline
-#define MULTIPLY_ACCUM_WITH_SCALE(index)\
-        isum += vaddvq_s32(ggml_vdotq_s32(vzero, q2bytes.val[0], q8bytes.val[0])) * aux[is+(index)];\
-        isum += vaddvq_s32(ggml_vdotq_s32(vzero, q2bytes.val[1], q8bytes.val[1])) * aux[is+1+(index)];
-
-#define SHIFT_MULTIPLY_ACCUM_WITH_SCALE(shift, index)\
-        q8bytes = ggml_vld1q_s8_x2(q8); q8 += 32;\
-        q2bytes.val[0] = vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q2bits.val[0], (shift)), m3));\
-        q2bytes.val[1] = vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q2bits.val[1], (shift)), m3));\
-        MULTIPLY_ACCUM_WITH_SCALE((index));
-
-        for (int j = 0; j < QK_K/128; ++j) {
-            const ggml_uint8x16x2_t q2bits = ggml_vld1q_u8_x2(q2); q2 += 32;
-
-            ggml_int8x16x2_t q8bytes = ggml_vld1q_s8_x2(q8); q8 += 32;
-            q2bytes.val[0] = vreinterpretq_s8_u8(vandq_u8(q2bits.val[0], m3));
-            q2bytes.val[1] = vreinterpretq_s8_u8(vandq_u8(q2bits.val[1], m3));
-
-            MULTIPLY_ACCUM_WITH_SCALE(0);
-
-            SHIFT_MULTIPLY_ACCUM_WITH_SCALE(2, 2);
-            SHIFT_MULTIPLY_ACCUM_WITH_SCALE(4, 4);
-            SHIFT_MULTIPLY_ACCUM_WITH_SCALE(6, 6);
-
-            is += 8;
-        }
-
-        sum += d * isum;
-    }
-
-    *s = sum;
-
-#elif defined __AVX2__
-
-    const __m256i m3 = _mm256_set1_epi8(3);
-    const __m128i m4 = _mm_set1_epi8(0xF);
-
-    __m256 acc = _mm256_setzero_ps();
-
-    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 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);
-        const __m128i mins8 = _mm_and_si128(_mm_srli_epi16(mins_and_scales, 4), m4);
-        const __m256i mins = _mm256_cvtepi8_epi16(mins8);
-        const __m256i prod = _mm256_madd_epi16(mins, _mm256_loadu_si256((const __m256i*)y[i].bsums));
-
-        acc = _mm256_fmadd_ps(_mm256_broadcast_ss(&dmin), _mm256_cvtepi32_ps(prod), acc);
-
-        const __m256i all_scales = _mm256_cvtepi8_epi16(scales8);
-        const __m128i l_scales = _mm256_extracti128_si256(all_scales, 0);
-        const __m128i h_scales = _mm256_extracti128_si256(all_scales, 1);
-        const __m256i scales[2] = {MM256_SET_M128I(l_scales, l_scales), MM256_SET_M128I(h_scales, h_scales)};
-
-        __m256i sumi = _mm256_setzero_si256();
-
-        for (int j = 0; j < QK_K/128; ++j) {
-
-            const __m256i q2bits = _mm256_loadu_si256((const __m256i*)q2); q2 += 32;
-
-            const __m256i q8_0 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
-            const __m256i q8_1 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
-            const __m256i q8_2 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
-            const __m256i q8_3 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
-
-            const __m256i q2_0 = _mm256_and_si256(q2bits, m3);
-            const __m256i q2_1 = _mm256_and_si256(_mm256_srli_epi16(q2bits, 2), m3);
-            const __m256i q2_2 = _mm256_and_si256(_mm256_srli_epi16(q2bits, 4), m3);
-            const __m256i q2_3 = _mm256_and_si256(_mm256_srli_epi16(q2bits, 6), m3);
-
-            __m256i p0 = _mm256_maddubs_epi16(q2_0, q8_0);
-            __m256i p1 = _mm256_maddubs_epi16(q2_1, q8_1);
-            __m256i p2 = _mm256_maddubs_epi16(q2_2, q8_2);
-            __m256i p3 = _mm256_maddubs_epi16(q2_3, q8_3);
-
-            p0 = _mm256_madd_epi16(_mm256_shuffle_epi8(scales[j], get_scale_shuffle_q3k(0)), p0);
-            p1 = _mm256_madd_epi16(_mm256_shuffle_epi8(scales[j], get_scale_shuffle_q3k(1)), p1);
-            p2 = _mm256_madd_epi16(_mm256_shuffle_epi8(scales[j], get_scale_shuffle_q3k(2)), p2);
-            p3 = _mm256_madd_epi16(_mm256_shuffle_epi8(scales[j], get_scale_shuffle_q3k(3)), p3);
-
-            p0 = _mm256_add_epi32(p0, p1);
-            p2 = _mm256_add_epi32(p2, p3);
-
-            sumi = _mm256_add_epi32(sumi, _mm256_add_epi32(p0, p2));
-        }
-
-        acc = _mm256_fmadd_ps(_mm256_broadcast_ss(&d), _mm256_cvtepi32_ps(sumi), acc);
-
-    }
-
-    *s = hsum_float_8(acc);
-
-#elif defined __AVX__
-
-    const __m128i m3 = _mm_set1_epi8(0x3);
-    const __m128i m4 = _mm_set1_epi8(0xF);
-    const __m128i m2 = _mm_set1_epi8(0x2);
-
-    __m256 acc = _mm256_setzero_ps();
-
-    for (int i = 0; i < nb; ++i) {
-
-        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 * 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);
-        const __m128i scales16 = _mm_and_si128(mins_and_scales, m4);
-        const __m128i mins16 = _mm_and_si128(_mm_srli_epi16(mins_and_scales, 4), m4);
-        const __m128i mins_0 = _mm_cvtepi8_epi16(mins16);
-        const __m128i mins_1 = _mm_cvtepi8_epi16(_mm_unpackhi_epi64(mins16, mins16));
-
-        // summs = y[i].bsums * (x[i].scales >> 4) in 16bits*8*2 to 32bits*4*2
-        const __m128i summs_0 = _mm_madd_epi16(mins_0, _mm_loadu_si128((const __m128i*)&y[i].bsums[0]));
-        const __m128i summs_1 = _mm_madd_epi16(mins_1, _mm_loadu_si128((const __m128i*)&y[i].bsums[8]));
-
-        // sumf += -dmin * summs in 32bits*8
-        acc = _mm256_add_ps(_mm256_mul_ps(_mm256_broadcast_ss(&dmin), _mm256_cvtepi32_ps(MM256_SET_M128I(summs_1, summs_0))), acc);
-
-        const __m128i scales_0 = _mm_cvtepi8_epi16(scales16);
-        const __m128i scales_1 = _mm_cvtepi8_epi16(_mm_unpackhi_epi64(scales16, scales16));
-        const __m128i scales[2] = { scales_0, scales_1 };
-
-        __m128i sumi_0 = _mm_setzero_si128();
-        __m128i sumi_1 = _mm_setzero_si128();
-
-        for (int j = 0; j < QK_K/128; ++j) {
-
-            // load Q8 quants int8*16*8 from block_q8_K.qs[QK_K]
-            const __m128i q8_0 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
-            const __m128i q8_1 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
-            const __m128i q8_2 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
-            const __m128i q8_3 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
-            const __m128i q8_4 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
-            const __m128i q8_5 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
-            const __m128i q8_6 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
-            const __m128i q8_7 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
-
-            // load 2bits*16*8 from block_q2_K.qs[QK_K/4]
-            __m128i q2bits = _mm_loadu_si128((const __m128i*)q2); q2 += 16;
-            const __m128i q2_0 = _mm_and_si128(q2bits, m3);
-            const __m128i q2_2 = _mm_and_si128(_mm_srli_epi16(q2bits, 2), m3);
-            const __m128i q2_4 = _mm_and_si128(_mm_srli_epi16(q2bits, 4), m3);
-            const __m128i q2_6 = _mm_and_si128(_mm_srli_epi16(q2bits, 6), m3);
-            q2bits = _mm_loadu_si128((const __m128i*)q2); q2 += 16;
-            const __m128i q2_1 = _mm_and_si128(q2bits, m3);
-            const __m128i q2_3 = _mm_and_si128(_mm_srli_epi16(q2bits, 2), m3);
-            const __m128i q2_5 = _mm_and_si128(_mm_srli_epi16(q2bits, 4), m3);
-            const __m128i q2_7 = _mm_and_si128(_mm_srli_epi16(q2bits, 6), m3);
-
-            // isuml = q8[l] * ((q2[l] >> shift) & 3) in 8bits*16*8 to 16bits*8*8
-            __m128i p0 = _mm_maddubs_epi16(q2_0, q8_0);
-            __m128i p1 = _mm_maddubs_epi16(q2_1, q8_1);
-            __m128i p2 = _mm_maddubs_epi16(q2_2, q8_2);
-            __m128i p3 = _mm_maddubs_epi16(q2_3, q8_3);
-            __m128i p4 = _mm_maddubs_epi16(q2_4, q8_4);
-            __m128i p5 = _mm_maddubs_epi16(q2_5, q8_5);
-            __m128i p6 = _mm_maddubs_epi16(q2_6, q8_6);
-            __m128i p7 = _mm_maddubs_epi16(q2_7, q8_7);
-
-            // isum += (x[i].scales[is++] & 0xF) * isuml in 16bits*8*8 to 32bits*4*8
-            __m128i shuffle = _mm_set1_epi16(0x0100);
-            p0 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p0);
-            shuffle = _mm_add_epi16(shuffle, m2);
-            p1 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p1);
-            shuffle = _mm_add_epi16(shuffle, m2);
-            p2 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p2);
-            shuffle = _mm_add_epi16(shuffle, m2);
-            p3 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p3);
-            shuffle = _mm_add_epi16(shuffle, m2);
-            p4 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p4);
-            shuffle = _mm_add_epi16(shuffle, m2);
-            p5 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p5);
-            shuffle = _mm_add_epi16(shuffle, m2);
-            p6 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p6);
-            shuffle = _mm_add_epi16(shuffle, m2);
-            p7 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p7);
-
-            p0 = _mm_add_epi32(p0, p1);
-            p2 = _mm_add_epi32(p2, p3);
-            p4 = _mm_add_epi32(p4, p5);
-            p6 = _mm_add_epi32(p6, p7);
-
-            // isum in 32bits*4*2
-            sumi_0 = _mm_add_epi32(sumi_0, _mm_add_epi32(p0, p2));
-            sumi_1 = _mm_add_epi32(sumi_1, _mm_add_epi32(p4, p6));
-        }
-
-        // sumf += dall * isum - dmin * summs in 32bits
-        __m256i sumi = MM256_SET_M128I(sumi_1, sumi_0);
-        acc = _mm256_add_ps(_mm256_mul_ps(_mm256_broadcast_ss(&dall), _mm256_cvtepi32_ps(sumi)), acc);
-    }
-
-    *s = hsum_float_8(acc);
-
-#elif defined __wasm_simd128__
-    float sumf = 0;
-
-    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;
-
-        // 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);
-
-            v128_t sc_low = wasm_u16x8_extend_low_u8x16(sc_upper);
-            v128_t sc_high = wasm_u16x8_extend_high_u8x16(sc_upper);
-
-            v128_t bsums1 = wasm_v128_load(&y[i].bsums[0]);
-            v128_t bsums2 = wasm_v128_load(&y[i].bsums[8]);
-
-            summs_vec = wasm_i32x4_add(
-                wasm_i32x4_add(wasm_i32x4_dot_i16x8(sc_low, bsums1),
-                               wasm_i32x4_dot_i16x8(sc_high, bsums2)),
-                summs_vec
-            );
-
-            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);
-
-        // Vectorized isum calculation
-        int32_t isum = 0;
-        const uint8_t * sc_ptr = sc;
-        const int k_iters = QK_K/128;
-
-        for (int k = 0; k < k_iters; ++k) {
-            v128_t isum_vec = wasm_i32x4_splat(0);
-            int shift = 0;
-
-            for (int j = 0; j < 4; ++j) {
-                const int d0 = (sc_ptr[0] & 0xF);
-                const int d1 = (sc_ptr[1] & 0xF);
-                sc_ptr += 2;
-
-                // 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));
-
-                // 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));
-
-                // 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)
-                );
-
-                // 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))
-                );
-
-                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);
-        }
-
-        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_xtheadvector
-
-    float sumf = 0;
-    uint8_t atmp[16];
-
-    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__(
-            "th.vsetvli zero, %[vl16], e8, m1\n\t"
-            "th.vmv.v.x v8, zero\n\t"
-            "th.vlb.v v1, (%[sc])\n\t"
-            "th.vand.vi v0, v1, 0xF\n\t"
-            "th.vsrl.vi v1, v1, 4\n\t"
-            "th.vsb.v v0, (%[scale])\n\t"
-            "th.vwaddu.vx v16, v1, zero\n\t"
-            "th.vsetvli zero, %[vl16], e16, m2\n\t"
-            "th.vlh.v v2, (%[bsums])\n\t"
-            "th.vwmul.vv v4, v16, v2\n\t"
-            "th.vsetvli zero, %[vl16], e32, m4\n\t"
-            "th.vredsum.vs v8, v4, v8\n\t"
-            "th.vmv.x.s %[vsums], v8"
-            : [tmp] "=&r" (tmp), [vsums] "=&r" (vsums)
-            : [sc] "r" (sc), [scale] "r" (atmp), [bsums] "r" (y[i].bsums)
-            , [vl16] "r" (16)
-            : "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__(
-                "th.vsetvli zero, %[vl32], e8, m2\n\t"
-                "th.vlb.v v0, (%[q2])\n\t"
-                "th.vsrl.vi v2, v0, 2\n\t"
-                "th.vsrl.vi v4, v0, 4\n\t"
-                "th.vsrl.vi v6, v0, 6\n\t"
-                "th.vand.vi v0, v0, 0x3\n\t"
-                "th.vand.vi v2, v2, 0x3\n\t"
-                "th.vand.vi v4, v4, 0x3\n\t"
-                "th.vsetvli zero, %[vl128], e8, m8\n\t"
-                "th.vlb.v v8, (%[q8])\n\t"
-                "th.vsetvli zero, %[vl64], e8, m4\n\t"
-                "th.vwmul.vv v16, v0, v8\n\t"
-                "th.vwmul.vv v24, v4, v12\n\t"
-                "th.vsetvli zero, %[vl16], e16, m2\n\t"
-                "th.vmv.v.x v0, zero\n\t"
-                "th.vwredsum.vs v10, v16, v0\n\t"
-                "th.vwredsum.vs v9, v18, v0\n\t"
-                "th.vwredsum.vs v8, v20, v0\n\t"
-                "th.vwredsum.vs v7, v22, v0\n\t"
-                "th.vwredsum.vs v11, v24, v0\n\t"
-                "th.vwredsum.vs v12, v26, v0\n\t"
-                "th.vwredsum.vs v13, v28, v0\n\t"
-                "th.vwredsum.vs v14, v30, v0\n\t"
-                "li %[tmp], 4\n\t"
-                "th.vsetvli zero, %[tmp], e32, m1\n\t"
-                "th.vslideup.vi v10, v9, 1\n\t"
-                "th.vslideup.vi v8, v7, 1\n\t"
-                "th.vslideup.vi v11, v12, 1\n\t"
-                "th.vslideup.vi v13, v14, 1\n\t"
-                "th.vslideup.vi v10, v8, 2\n\t"
-                "th.vslideup.vi v11, v13, 2\n\t"
-                "li %[tmp], 8\n\t"
-                "th.vsetvli zero, %[tmp], e32, m2\n\t"
-                "th.vlbu.v v12, (%[scale])\n\t"
-                "th.vmul.vv v10, v10, v12\n\t"
-                "th.vredsum.vs v0, v10, v0\n\t"
-                "th.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)
-                , [vl16] "r" (16), [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;
-    }
-
-    *s = sumf;
-
-#elif defined __riscv_v
-
-    float sumf = 0;
-    uint8_t atmp[16];
-
-    const int vector_length = __riscv_vlenb() * 8;
-    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 };
-
-    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;
-
-#elif defined(__POWER9_VECTOR__)
-    const vector signed char lowMask = vec_splats((signed char)0x3);
-    const vector signed char lowScaleMask = vec_splats((signed char)0xF);
-    const vector int v0 = vec_splats((int32_t)0);
-    const vector unsigned char v2 = vec_splats((unsigned char)0x2);
-    const vector unsigned char v6 = vec_splats((unsigned char)0x6);
-    const vector unsigned char v4 = vec_splats((unsigned char)0x4);
-
-    vector float vsumf0 = vec_splats(0.0f);
-    vector float vsumf1 = vec_splats(0.0f);
-    vector float vsumf2 = vec_splats(0.0f);
-    vector float vsumf3 = vec_splats(0.0f);
-
-    for (int i = 0; i < nb; ++i) {
-        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d));
-        vector float vyd = vec_splats(y[i].d);
-        vector float vd = vec_mul(vxd, vyd);
-
-        vector float vxmin = vec_splats(GGML_FP16_TO_FP32(x[i].dmin));
-        vector float vdmin = vec_mul(vxmin, vyd);
-
-        vector signed short q8ysums0 = vec_xl( 0, y[i].bsums);
-        vector signed short q8ysums1 = vec_xl(16, y[i].bsums);
-
-        vector signed char q2xmins = (vector signed char)vec_xl( 0, x[i].scales);
-        vector signed char vscales = vec_and(q2xmins, lowScaleMask);
-
-        q2xmins = vec_sr(q2xmins, v4);
-        vector signed short q2xmins0 = vec_unpackh(q2xmins);
-        vector signed short q2xmins1 = vec_unpackl(q2xmins);
-
-        vector signed int prod0 = vec_mule(q2xmins0, q8ysums0);
-        vector signed int prod1 = vec_mulo(q2xmins0, q8ysums0);
-        vector signed int prod2 = vec_mule(q2xmins1, q8ysums1);
-        vector signed int prod3 = vec_mulo(q2xmins1, q8ysums1);
-
-        vsumf0 = vec_nmsub(vec_ctf(prod0, 0), vdmin, vsumf0);
-        vsumf1 = vec_nmsub(vec_ctf(prod1, 0), vdmin, vsumf1);
-        vsumf2 = vec_nmsub(vec_ctf(prod2, 0), vdmin, vsumf2);
-        vsumf3 = vec_nmsub(vec_ctf(prod3, 0), vdmin, vsumf3);
-
-        vector signed int vsumi0 = v0;
-        vector signed int vsumi1 = v0;
-        vector signed int vsumi2 = v0;
-        vector signed int vsumi3 = v0;
-        vector signed int vsumi4 = v0;
-        vector signed int vsumi5 = v0;
-        vector signed int vsumi6 = v0;
-        vector signed int vsumi7 = v0;
-
-        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);
-            __builtin_prefetch(q8, 0, 1);
-
-            vector signed char qxs0 = (vector signed char)vec_xl( 0, q2);
-            vector signed char qxs1 = (vector signed char)vec_xl(16, q2);
-            q2 += 32;
-
-            vector unsigned char q2x00 = (vector unsigned char)vec_and(qxs0, lowMask);
-            vector unsigned char q2x01 = (vector unsigned char)vec_and(vec_sr(qxs0, v2), lowMask);
-            vector unsigned char q2x02 = (vector unsigned char)vec_and(vec_sr(qxs0, v4), lowMask);
-            vector unsigned char q2x03 = (vector unsigned char)vec_and(vec_sr(qxs0, v6), lowMask);
-            vector unsigned char q2x10 = (vector unsigned char)vec_and(qxs1, lowMask);
-            vector unsigned char q2x11 = (vector unsigned char)vec_and(vec_sr(qxs1, v2), lowMask);
-            vector unsigned char q2x12 = (vector unsigned char)vec_and(vec_sr(qxs1, v4), lowMask);
-            vector unsigned char q2x13 = (vector unsigned char)vec_and(vec_sr(qxs1, v6), lowMask);
-
-            vector signed char q8y00 = vec_xl(  0, q8);
-            vector signed char q8y10 = vec_xl( 16, q8);
-            vector signed char q8y01 = vec_xl( 32, q8);
-            vector signed char q8y11 = vec_xl( 48, q8);
-            vector signed char q8y02 = vec_xl( 64, q8);
-            vector signed char q8y12 = vec_xl( 80, q8);
-            vector signed char q8y03 = vec_xl( 96, q8);
-            vector signed char q8y13 = vec_xl(112, q8);
-            q8 += 128;
-
-            vector signed int qv0 = vec_msum(q8y00, q2x00, v0);
-            vector signed int qv1 = vec_msum(q8y01, q2x01, v0);
-            vector signed int qv2 = vec_msum(q8y02, q2x02, v0);
-            vector signed int qv3 = vec_msum(q8y03, q2x03, v0);
-            vector signed int qv4 = vec_msum(q8y10, q2x10, v0);
-            vector signed int qv5 = vec_msum(q8y11, q2x11, v0);
-            vector signed int qv6 = vec_msum(q8y12, q2x12, v0);
-            vector signed int qv7 = vec_msum(q8y13, q2x13, v0);
-
-            vector signed short vscales_07 = vec_unpackh(vscales);
-            vector signed int vscales_03 = vec_unpackh(vscales_07);
-            vector signed int vscales_47 = vec_unpackl(vscales_07);
-            vector signed int vs0 = vec_splat(vscales_03, 0);
-            vector signed int vs1 = vec_splat(vscales_03, 1);
-            vector signed int vs2 = vec_splat(vscales_03, 2);
-            vector signed int vs3 = vec_splat(vscales_03, 3);
-            vector signed int vs4 = vec_splat(vscales_47, 0);
-            vector signed int vs5 = vec_splat(vscales_47, 1);
-            vector signed int vs6 = vec_splat(vscales_47, 2);
-            vector signed int vs7 = vec_splat(vscales_47, 3);
-            vscales = vec_sld(vscales, vscales, 8);
-
-            vsumi0 = vec_add(vec_mul(qv0, vs0), vsumi0);
-            vsumi1 = vec_add(vec_mul(qv1, vs2), vsumi1);
-            vsumi2 = vec_add(vec_mul(qv2, vs4), vsumi2);
-            vsumi3 = vec_add(vec_mul(qv3, vs6), vsumi3);
-            vsumi4 = vec_add(vec_mul(qv4, vs1), vsumi4);
-            vsumi5 = vec_add(vec_mul(qv5, vs3), vsumi5);
-            vsumi6 = vec_add(vec_mul(qv6, vs5), vsumi6);
-            vsumi7 = vec_add(vec_mul(qv7, vs7), vsumi7);
-        }
-
-        vsumi0 = vec_add(vsumi0, vsumi4);
-        vsumi1 = vec_add(vsumi1, vsumi5);
-        vsumi2 = vec_add(vsumi2, vsumi6);
-        vsumi3 = vec_add(vsumi3, vsumi7);
-
-        vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
-        vsumf1 = vec_madd(vec_ctf(vsumi1, 0), vd, vsumf1);
-        vsumf2 = vec_madd(vec_ctf(vsumi2, 0), vd, vsumf2);
-        vsumf3 = vec_madd(vec_ctf(vsumi3, 0), vd, vsumf3);
-    }
-
-    vsumf0 = vec_add(vsumf0, vsumf2);
-    vsumf1 = vec_add(vsumf1, vsumf3);
-
-    vsumf0 = vec_add(vsumf0, vsumf1);
-
-    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4));
-    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8));
-
-    *s = vec_extract(vsumf0, 0);
-
-#elif defined __loongarch_asx
-
-    __m256 acc = (__m256)__lasx_xvldi(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 uint8_t * GGML_RESTRICT q2 = x[i].qs;
-        const int8_t  * GGML_RESTRICT q8 = y[i].qs;
-
-        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 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);
-
-        for (int j = 0; j < QK_K/128; ++j) {
-
-            const __m256i q2bits = __lasx_xvld((const __m256i*)q2, 0); q2 += 32;
-
-            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;
-
-            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_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_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);
-
-            sumi = __lasx_xvadd_w(sumi, __lasx_xvadd_w(p0, p2));
-        }
-
-        acc = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(d), __lasx_xvffint_s_w(sumi), acc);
-
-    }
-
-    *s = hsum_float_8(acc);
-
-#else
-
-    float sumf = 0;
-
-    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;
-
-        int summs = 0;
-        for (int j = 0; j < 16; ++j) {
-            summs += y[i].bsums[j] * (sc[j] >> 4);
-        }
-
-        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);
-
-        int isum = 0;
-        int is = 0;
-        int d;
-        for (int k = 0; k < QK_K/128; ++k) {
-            int shift = 0;
-            for (int j = 0; j < 4; ++j) {
-                d = sc[is++] & 0xF;
-                int isuml = 0;
-                for (int l =  0; l < 16; ++l) isuml += q8[l] * ((q2[l] >> shift) & 3);
-                isum += d * isuml;
-                d = sc[is++] & 0xF;
-                isuml = 0;
-                for (int l = 16; l < 32; ++l) isuml += q8[l] * ((q2[l] >> shift) & 3);
-                isum += d * isuml;
-                shift += 2;
-                q8 += 32;
-            }
-            q2 += 32;
-        }
-        sumf += dall * isum - dmin * summs;
-    }
-    *s = sumf;
-#endif
-}
-
-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);
-    UNUSED(bx);
-    UNUSED(by);
-    UNUSED(bs);
-
-    const uint32_t kmask1 = 0x03030303;
-    const uint32_t kmask2 = 0x0f0f0f0f;
-
-    const block_q3_K * GGML_RESTRICT x = vx;
-    const block_q8_K * GGML_RESTRICT y = vy;
-
-    const int nb = n / QK_K;
-
-#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];
-
-    const uint8x16_t m3b = vdupq_n_u8(0x3);
-    const int32x4_t  vzero = vdupq_n_s32(0);
-
-    const uint8x16_t m0 = vdupq_n_u8(1);
-    const uint8x16_t m1 = vshlq_n_u8(m0, 1);
-    const uint8x16_t m2 = vshlq_n_u8(m0, 2);
-    const uint8x16_t m3 = vshlq_n_u8(m0, 3);
-    const int8_t m32 = 32;
-
-    ggml_int8x16x4_t q3bytes;
-
-    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 = 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);
-
-        ggml_uint8x16x4_t q3h;
-
-        int32_t isum = 0;
-
-        // 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;
-
-        for (int j = 0; j < QK_K/128; ++j) {
-
-            const ggml_uint8x16x2_t q3bits = ggml_vld1q_u8_x2(q3); q3 += 32;
-            const ggml_int8x16x4_t q8bytes_1 = ggml_vld1q_s8_x4(q8); q8 += 64;
-            const ggml_int8x16x4_t q8bytes_2 = ggml_vld1q_s8_x4(q8); q8 += 64;
-
-            q3h.val[0] = vshlq_n_u8(vbicq_u8(m0, qhbits.val[0]), 2);
-            q3h.val[1] = vshlq_n_u8(vbicq_u8(m0, qhbits.val[1]), 2);
-            q3h.val[2] = vshlq_n_u8(vbicq_u8(m1, qhbits.val[0]), 1);
-            q3h.val[3] = vshlq_n_u8(vbicq_u8(m1, qhbits.val[1]), 1);
-
-            q3bytes.val[0] = vsubq_s8(vreinterpretq_s8_u8(vandq_u8(q3bits.val[0], m3b)), vreinterpretq_s8_u8(q3h.val[0]));
-            q3bytes.val[1] = vsubq_s8(vreinterpretq_s8_u8(vandq_u8(q3bits.val[1], m3b)), vreinterpretq_s8_u8(q3h.val[1]));
-            q3bytes.val[2] = vsubq_s8(vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q3bits.val[0], 2), m3b)), vreinterpretq_s8_u8(q3h.val[2]));
-            q3bytes.val[3] = vsubq_s8(vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q3bits.val[1], 2), m3b)), vreinterpretq_s8_u8(q3h.val[3]));
-
-            isum += vaddvq_s32(ggml_vdotq_s32(vzero, q3bytes.val[0], q8bytes_1.val[0])) * scale[0];
-            isum += vaddvq_s32(ggml_vdotq_s32(vzero, q3bytes.val[1], q8bytes_1.val[1])) * scale[1];
-            isum += vaddvq_s32(ggml_vdotq_s32(vzero, q3bytes.val[2], q8bytes_1.val[2])) * scale[2];
-            isum += vaddvq_s32(ggml_vdotq_s32(vzero, q3bytes.val[3], q8bytes_1.val[3])) * scale[3];
-
-            scale += 4;
-
-            q3h.val[0] = vbicq_u8(m2, qhbits.val[0]);
-            q3h.val[1] = vbicq_u8(m2, qhbits.val[1]);
-            q3h.val[2] = vshrq_n_u8(vbicq_u8(m3, qhbits.val[0]), 1);
-            q3h.val[3] = vshrq_n_u8(vbicq_u8(m3, qhbits.val[1]), 1);
-
-            q3bytes.val[0] = vsubq_s8(vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q3bits.val[0], 4), m3b)), vreinterpretq_s8_u8(q3h.val[0]));
-            q3bytes.val[1] = vsubq_s8(vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q3bits.val[1], 4), m3b)), vreinterpretq_s8_u8(q3h.val[1]));
-            q3bytes.val[2] = vsubq_s8(vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q3bits.val[0], 6), m3b)), vreinterpretq_s8_u8(q3h.val[2]));
-            q3bytes.val[3] = vsubq_s8(vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q3bits.val[1], 6), m3b)), vreinterpretq_s8_u8(q3h.val[3]));
-
-            isum += vaddvq_s32(ggml_vdotq_s32(vzero, q3bytes.val[0], q8bytes_2.val[0])) * scale[0];
-            isum += vaddvq_s32(ggml_vdotq_s32(vzero, q3bytes.val[1], q8bytes_2.val[1])) * scale[1];
-            isum += vaddvq_s32(ggml_vdotq_s32(vzero, q3bytes.val[2], q8bytes_2.val[2])) * scale[2];
-            isum += vaddvq_s32(ggml_vdotq_s32(vzero, q3bytes.val[3], q8bytes_2.val[3])) * scale[3];
-
-            scale += 4;
-
-            if (j == 0) {
-                qhbits.val[0] = vshrq_n_u8(qhbits.val[0], 4);
-                qhbits.val[1] = vshrq_n_u8(qhbits.val[1], 4);
-            }
-
-        }
-        sum += d * isum;
-
-    }
-
-    *s = sum;
-
-#elif defined __AVX2__
-
-    const __m256i m3 = _mm256_set1_epi8(3);
-    const __m256i mone = _mm256_set1_epi8(1);
-    const __m128i m32 = _mm_set1_epi8(32);
-
-    __m256 acc = _mm256_setzero_ps();
-
-    uint32_t aux[3];
-
-    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 = x[i].qs;
-        const int8_t  * GGML_RESTRICT q8 = y[i].qs;
-
-        // Set up scales
-        memcpy(aux, x[i].scales, 12);
-        __m128i scales128 = _mm_set_epi32(
-                ((aux[1] >> 4) & kmask2) | (((aux[2] >> 6) & kmask1) << 4),
-                ((aux[0] >> 4) & kmask2) | (((aux[2] >> 4) & kmask1) << 4),
-                (aux[1] & kmask2) | (((aux[2] >> 2) & kmask1) << 4),
-                (aux[0] & kmask2) | (((aux[2] >> 0) & kmask1) << 4));
-        scales128 = _mm_sub_epi8(scales128, m32);
-        const __m256i all_scales = _mm256_cvtepi8_epi16(scales128);
-        const __m128i l_scales = _mm256_extracti128_si256(all_scales, 0);
-        const __m128i h_scales = _mm256_extracti128_si256(all_scales, 1);
-        const __m256i scales[2] = {MM256_SET_M128I(l_scales, l_scales), MM256_SET_M128I(h_scales, h_scales)};
-
-        // high bit
-        const __m256i hbits = _mm256_loadu_si256((const __m256i*)x[i].hmask);
-
-        // integer accumulator
-        __m256i sumi = _mm256_setzero_si256();
-
-        int bit = 0;
-        int is  = 0;
-
-        for (int j = 0; j < QK_K/128; ++j) {
-            // load low 2 bits
-            const __m256i q3bits = _mm256_loadu_si256((const __m256i*)q3); q3 += 32;
-
-            // prepare low and high bits
-            const __m256i q3l_0 = _mm256_and_si256(q3bits, m3);
-            const __m256i q3h_0 = _mm256_slli_epi16(_mm256_srli_epi16(_mm256_andnot_si256(hbits, _mm256_slli_epi16(mone, bit)), bit), 2);
-            ++bit;
-
-            const __m256i q3l_1 = _mm256_and_si256(_mm256_srli_epi16(q3bits, 2), m3);
-            const __m256i q3h_1 = _mm256_slli_epi16(_mm256_srli_epi16(_mm256_andnot_si256(hbits, _mm256_slli_epi16(mone, bit)), bit), 2);
-            ++bit;
-
-            const __m256i q3l_2 = _mm256_and_si256(_mm256_srli_epi16(q3bits, 4), m3);
-            const __m256i q3h_2 = _mm256_slli_epi16(_mm256_srli_epi16(_mm256_andnot_si256(hbits, _mm256_slli_epi16(mone, bit)), bit), 2);
-            ++bit;
-
-            const __m256i q3l_3 = _mm256_and_si256(_mm256_srli_epi16(q3bits, 6), m3);
-            const __m256i q3h_3 = _mm256_slli_epi16(_mm256_srli_epi16(_mm256_andnot_si256(hbits, _mm256_slli_epi16(mone, bit)), bit), 2);
-            ++bit;
-
-            // load Q8 quants
-            const __m256i q8_0 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
-            const __m256i q8_1 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
-            const __m256i q8_2 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
-            const __m256i q8_3 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
-
-            // Dot product: we multiply the 2 low bits and 1 high bit part separately, so we can use _mm256_maddubs_epi16,
-            // 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 = _mm256_maddubs_epi16(q3h_0, q8_0);
-            __m256i q8s_1 = _mm256_maddubs_epi16(q3h_1, q8_1);
-            __m256i q8s_2 = _mm256_maddubs_epi16(q3h_2, q8_2);
-            __m256i q8s_3 = _mm256_maddubs_epi16(q3h_3, q8_3);
-
-            __m256i p16_0 = _mm256_maddubs_epi16(q3l_0, q8_0);
-            __m256i p16_1 = _mm256_maddubs_epi16(q3l_1, q8_1);
-            __m256i p16_2 = _mm256_maddubs_epi16(q3l_2, q8_2);
-            __m256i p16_3 = _mm256_maddubs_epi16(q3l_3, q8_3);
-
-            p16_0 = _mm256_sub_epi16(p16_0, q8s_0);
-            p16_1 = _mm256_sub_epi16(p16_1, q8s_1);
-            p16_2 = _mm256_sub_epi16(p16_2, q8s_2);
-            p16_3 = _mm256_sub_epi16(p16_3, q8s_3);
-
-            // multiply with scales
-            p16_0 = _mm256_madd_epi16(_mm256_shuffle_epi8(scales[j], get_scale_shuffle_q3k(is + 0)), p16_0);
-            p16_1 = _mm256_madd_epi16(_mm256_shuffle_epi8(scales[j], get_scale_shuffle_q3k(is + 1)), p16_1);
-            p16_2 = _mm256_madd_epi16(_mm256_shuffle_epi8(scales[j], get_scale_shuffle_q3k(is + 2)), p16_2);
-            p16_3 = _mm256_madd_epi16(_mm256_shuffle_epi8(scales[j], get_scale_shuffle_q3k(is + 3)), p16_3);
-
-            // accumulate
-            p16_0 = _mm256_add_epi32(p16_0, p16_1);
-            p16_2 = _mm256_add_epi32(p16_2, p16_3);
-            sumi  = _mm256_add_epi32(sumi, _mm256_add_epi32(p16_0, p16_2));
-
-        }
-
-        // multiply with block scale and accumulate
-        acc = _mm256_fmadd_ps(_mm256_broadcast_ss(&d), _mm256_cvtepi32_ps(sumi), acc);
-
-    }
-
-    *s = hsum_float_8(acc);
-
-#elif defined __AVX__
-
-    const __m128i m3 = _mm_set1_epi8(3);
-    const __m128i mone = _mm_set1_epi8(1);
-    const __m128i m32 = _mm_set1_epi8(32);
-    const __m128i m2 = _mm_set1_epi8(2);
-
-    __m256 acc = _mm256_setzero_ps();
-
-    const uint32_t *aux;
-
-    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 = x[i].qs;
-        const int8_t  * GGML_RESTRICT q8 = y[i].qs;
-
-        // Set up scales
-        aux = (const uint32_t *)x[i].scales;
-        __m128i scales128 = _mm_set_epi32(
-                ((aux[1] >> 4) & kmask2) | (((aux[2] >> 6) & kmask1) << 4),
-                ((aux[0] >> 4) & kmask2) | (((aux[2] >> 4) & kmask1) << 4),
-                (aux[1] & kmask2) | (((aux[2] >> 2) & kmask1) << 4),
-                (aux[0] & kmask2) | (((aux[2] >> 0) & kmask1) << 4));
-        scales128 = _mm_sub_epi8(scales128, m32);
-        const __m128i scales_0 = _mm_cvtepi8_epi16(scales128);
-        const __m128i scales_1 = _mm_cvtepi8_epi16(_mm_unpackhi_epi64(scales128, scales128));
-        const __m128i scales[2] = { scales_0, scales_1 };
-
-        // high bit *128*2 from block_q3_K.hmask[QK_K/8]
-        const __m128i hbits_0 = _mm_loadu_si128((const __m128i*)&x[i].hmask[0]);
-        const __m128i hbits_1 = _mm_loadu_si128((const __m128i*)&x[i].hmask[16]);
-
-        // integer accumulator
-        __m128i sumi_0 = _mm_setzero_si128();
-        __m128i sumi_1 = _mm_setzero_si128();
-
-        for (int j = 0; j < QK_K/128; ++j) {
-            // load low 2 bits *64*2 from block_q3_K.qs[QK_K/4]
-            const __m128i q3bits_0 = _mm_loadu_si128((const __m128i*)q3); q3 += 16;
-            const __m128i q3bits_1 = _mm_loadu_si128((const __m128i*)q3); q3 += 16;
-
-            // prepare low and high bits
-            const int bit = j << 2;
-
-            const __m128i q3l_0 = _mm_and_si128(q3bits_0, m3);
-            const __m128i q3l_1 = _mm_and_si128(q3bits_1, m3);
-            const __m128i q3h_0 = _mm_slli_epi16(_mm_srli_epi16(_mm_andnot_si128(hbits_0, _mm_slli_epi16(mone, bit)), bit), 2);
-            const __m128i q3h_1 = _mm_slli_epi16(_mm_srli_epi16(_mm_andnot_si128(hbits_1, _mm_slli_epi16(mone, bit)), bit), 2);
-
-            const __m128i q3l_2 = _mm_and_si128(_mm_srli_epi16(q3bits_0, 2), m3);
-            const __m128i q3l_3 = _mm_and_si128(_mm_srli_epi16(q3bits_1, 2), m3);
-            const __m128i q3h_2 = _mm_slli_epi16(_mm_srli_epi16(_mm_andnot_si128(hbits_0, _mm_slli_epi16(mone, bit+1)), bit+1), 2);
-            const __m128i q3h_3 = _mm_slli_epi16(_mm_srli_epi16(_mm_andnot_si128(hbits_1, _mm_slli_epi16(mone, bit+1)), bit+1), 2);
-
-            const __m128i q3l_4 = _mm_and_si128(_mm_srli_epi16(q3bits_0, 4), m3);
-            const __m128i q3l_5 = _mm_and_si128(_mm_srli_epi16(q3bits_1, 4), m3);
-            const __m128i q3h_4 = _mm_slli_epi16(_mm_srli_epi16(_mm_andnot_si128(hbits_0, _mm_slli_epi16(mone, bit+2)), bit+2), 2);
-            const __m128i q3h_5 = _mm_slli_epi16(_mm_srli_epi16(_mm_andnot_si128(hbits_1, _mm_slli_epi16(mone, bit+2)), bit+2), 2);
-
-            const __m128i q3l_6 = _mm_and_si128(_mm_srli_epi16(q3bits_0, 6), m3);
-            const __m128i q3l_7 = _mm_and_si128(_mm_srli_epi16(q3bits_1, 6), m3);
-            const __m128i q3h_6 = _mm_slli_epi16(_mm_srli_epi16(_mm_andnot_si128(hbits_0, _mm_slli_epi16(mone, bit+3)), bit+3), 2);
-            const __m128i q3h_7 = _mm_slli_epi16(_mm_srli_epi16(_mm_andnot_si128(hbits_1, _mm_slli_epi16(mone, bit+3)), bit+3), 2);
-
-            // load Q8 quants from block_q8_K.qs[QK_K]
-            const __m128i q8_0 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
-            const __m128i q8_1 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
-            const __m128i q8_2 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
-            const __m128i q8_3 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
-            const __m128i q8_4 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
-            const __m128i q8_5 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
-            const __m128i q8_6 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
-            const __m128i q8_7 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
-
-            // Dot product: we multiply the 2 low bits and 1 high bit part separately, so we can use _mm256_maddubs_epi16,
-            // 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)
-            __m128i q8s_0 = _mm_maddubs_epi16(q3h_0, q8_0);
-            __m128i q8s_1 = _mm_maddubs_epi16(q3h_1, q8_1);
-            __m128i q8s_2 = _mm_maddubs_epi16(q3h_2, q8_2);
-            __m128i q8s_3 = _mm_maddubs_epi16(q3h_3, q8_3);
-            __m128i q8s_4 = _mm_maddubs_epi16(q3h_4, q8_4);
-            __m128i q8s_5 = _mm_maddubs_epi16(q3h_5, q8_5);
-            __m128i q8s_6 = _mm_maddubs_epi16(q3h_6, q8_6);
-            __m128i q8s_7 = _mm_maddubs_epi16(q3h_7, q8_7);
-
-            __m128i p16_0 = _mm_maddubs_epi16(q3l_0, q8_0);
-            __m128i p16_1 = _mm_maddubs_epi16(q3l_1, q8_1);
-            __m128i p16_2 = _mm_maddubs_epi16(q3l_2, q8_2);
-            __m128i p16_3 = _mm_maddubs_epi16(q3l_3, q8_3);
-            __m128i p16_4 = _mm_maddubs_epi16(q3l_4, q8_4);
-            __m128i p16_5 = _mm_maddubs_epi16(q3l_5, q8_5);
-            __m128i p16_6 = _mm_maddubs_epi16(q3l_6, q8_6);
-            __m128i p16_7 = _mm_maddubs_epi16(q3l_7, q8_7);
-
-            p16_0 = _mm_sub_epi16(p16_0, q8s_0);
-            p16_1 = _mm_sub_epi16(p16_1, q8s_1);
-            p16_2 = _mm_sub_epi16(p16_2, q8s_2);
-            p16_3 = _mm_sub_epi16(p16_3, q8s_3);
-            p16_4 = _mm_sub_epi16(p16_4, q8s_4);
-            p16_5 = _mm_sub_epi16(p16_5, q8s_5);
-            p16_6 = _mm_sub_epi16(p16_6, q8s_6);
-            p16_7 = _mm_sub_epi16(p16_7, q8s_7);
-
-            // multiply with scales
-            __m128i shuffle = _mm_set1_epi16(0x0100);
-            p16_0 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p16_0);
-            shuffle = _mm_add_epi16(shuffle, m2);
-            p16_1 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p16_1);
-            shuffle = _mm_add_epi16(shuffle, m2);
-            p16_2 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p16_2);
-            shuffle = _mm_add_epi16(shuffle, m2);
-            p16_3 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p16_3);
-            shuffle = _mm_add_epi16(shuffle, m2);
-            p16_4 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p16_4);
-            shuffle = _mm_add_epi16(shuffle, m2);
-            p16_5 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p16_5);
-            shuffle = _mm_add_epi16(shuffle, m2);
-            p16_6 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p16_6);
-            shuffle = _mm_add_epi16(shuffle, m2);
-            p16_7 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p16_7);
-
-            // accumulate
-            p16_0 = _mm_add_epi32(p16_0, p16_1);
-            p16_2 = _mm_add_epi32(p16_2, p16_3);
-            p16_4 = _mm_add_epi32(p16_4, p16_5);
-            p16_6 = _mm_add_epi32(p16_6, p16_7);
-            sumi_0 = _mm_add_epi32(sumi_0, _mm_add_epi32(p16_0, p16_2));
-            sumi_1 = _mm_add_epi32(sumi_1, _mm_add_epi32(p16_4, p16_6));
-
-        }
-
-        // multiply with block scale and accumulate
-        __m256i sumi = MM256_SET_M128I(sumi_1, sumi_0);
-        acc = _mm256_add_ps(_mm256_mul_ps(_mm256_broadcast_ss(&d), _mm256_cvtepi32_ps(sumi)), acc);
-
-    }
-
-    *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_xtheadvector
-
-    uint32_t utmp[4];
-    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;
-
-        int8_t * scale = (int8_t *)utmp;
-        int tmp;
-        __asm__ __volatile__(
-            "li %[tmp], 12\n\t"
-            "th.vsetvli zero, %[tmp], e8, m1\n\t"
-            "th.vlb.v v0, (%[s6b])\n\t"
-            "th.vmv.v.v v2, v0\n\t"
-            "li %[tmp], 2\n\t"
-            "th.vsetvli zero, %[tmp], e64, m1\n\t"
-            "th.vmv.v.x v9, %[sh]\n\t"\
-            "th.vslidedown.vi v1, v0, 1\n\t"
-            "th.vslide1up.vx v8, v9, zero\n\t" // {0, 0, 4, 4}
-            "th.vslideup.vi v0, v2, 1\n\t" // {aux[0], aux[1], aux[0], aux[1]}
-            "li %[tmp], 4\n\t"
-            "th.vsetvli zero, %[tmp], e32, m1\n\t"
-            "th.vid.v v9\n\t"
-            "th.vmv.x.s %[tmp], v1\n\t"
-            "th.vsll.vi v9, v9, 1\n\t" // {0, 2, 4, 6}
-            "th.vmv.v.x v1, %[tmp]\n\t" // {aux[2], aux[2], aux[2], aux[2]}
-            "th.vsrl.vv v4, v1, v9\n\t"
-            "th.vsrl.vv v2, v0, v8\n\t"
-            "th.vand.vx v5, v4, %[kmask1]\n\t"
-            "th.vand.vx v3, v2, %[kmask2]\n\t"
-            "th.vsll.vi v6, v5, 4\n\t"
-            "th.vor.vv v7, v6, v3\n\t"
-            "li %[tmp], 16\n\t"
-            "th.vsetvli zero, %[tmp], e8, m1\n\t"
-            "th.vsub.vx v0, v7, %[c]\n\t"
-            "th.vsb.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"
-        );
-
-        uint8_t m = 1;
-        int isum = 0;
-        for (int j = 0; j < QK_K; j += 128) {
-            __asm__ __volatile__(
-                // fixme: use v0p7 mask layout directly
-                "th.vsetvli zero, %[vl32], e8, m2\n\t"
-                "th.vlb.v v8, (%[q3])\n\t"
-                "th.vsrl.vi v10, v8, 2\n\t"
-                "th.vsrl.vi v12, v8, 4\n\t"
-                "th.vsrl.vi v14, v8, 6\n\t"
-                "th.vand.vi v8, v8, 3\n\t"
-                "th.vand.vi v10, v10, 3\n\t"
-                "th.vand.vi v12, v12, 3\n\t"
-                "th.vlb.v v2, (%[qh])\n\t"
-                "th.vand.vx v4, v2, %[m]\n\t"
-                "slli %[m], %[m], 1\n\t"
-                "th.vmseq.vx v0, v4, zero\n\t"
-                "th.vadd.vi v8, v8, -4, v0.t\n\t"
-                "th.vand.vx v4, v2, %[m]\n\t"
-                "slli %[m], %[m], 1\n\t"
-                "th.vmseq.vx v0, v4, zero\n\t"
-                "th.vadd.vi v10, v10, -4, v0.t\n\t"
-                "th.vand.vx v4, v2, %[m]\n\t"
-                "slli %[m], %[m], 1\n\t"
-                "th.vmseq.vx v0, v4, zero\n\t"
-                "th.vadd.vi v12, v12, -4, v0.t\n\t"
-                "th.vand.vx v4, v2, %[m]\n\t"
-                "slli %[m], %[m], 1\n\t"
-                "th.vmseq.vx v0, v4, zero\n\t"
-                "th.vadd.vi v14, v14, -4, v0.t\n\t"
-                "th.vsetvli zero, %[vl128], e8, m8\n\t"
-                "th.vlb.v v0, (%[q8])\n\t"
-                "th.vsetvli zero, %[vl64], e8, m4\n\t"
-                "th.vwmul.vv v16, v0, v8\n\t"
-                "th.vwmul.vv v24, v4, v12\n\t"
-                "li %[tmp], 16\n\t"
-                "th.vsetvli zero, %[tmp], e16, m2\n\t"
-                "th.vmv.v.x v0, zero\n\t"
-                "th.vwredsum.vs v10, v16, v0\n\t"
-                "th.vwredsum.vs v9, v18, v0\n\t"
-                "th.vwredsum.vs v8, v20, v0\n\t"
-                "th.vwredsum.vs v7, v22, v0\n\t"
-                "th.vwredsum.vs v11, v24, v0\n\t"
-                "th.vwredsum.vs v12, v26, v0\n\t"
-                "th.vwredsum.vs v13, v28, v0\n\t"
-                "th.vwredsum.vs v14, v30, v0\n\t"
-                "li %[tmp], 4\n\t"
-                "th.vsetvli zero, %[tmp], e32, m1\n\t"
-                "th.vslideup.vi v10, v9, 1\n\t"
-                "th.vslideup.vi v8, v7, 1\n\t"
-                "th.vslideup.vi v11, v12, 1\n\t"
-                "th.vslideup.vi v13, v14, 1\n\t"
-                "th.vslideup.vi v10, v8, 2\n\t"
-                "th.vslideup.vi v11, v13, 2\n\t"
-                "li %[tmp], 8\n\t"
-                "th.vsetvli zero, %[tmp], e32, m2\n\t"
-                "th.vlb.v v12, (%[scale])\n\t"
-                "th.vmul.vv v10, v10, v12\n\t"
-                "th.vredsum.vs v0, v10, v0\n\t"
-                "th.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;
-    }
-
-    *s = sumf;
-
-#elif defined __riscv_v
-
-    uint32_t utmp[4];
-    float sumf = 0;
-    uint32_t aux[3];
-    const int vector_length = __riscv_vlenb() * 8;
-
-    switch (vector_length) {
-    case 256:
-        for (int i = 0; i < nb; ++i) {
-
-            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;
-
-            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;
-
-            vint32m1_t vzero = __riscv_vmv_v_x_i32m1(0, 1);
-            vuint8m1_t vqh = __riscv_vle8_v_u8m1(qh, vl);
-
-            int sum_t = 0;
-
-            for (int j = 0; j < QK_K; j += 128) {
-
-                vl = 32;
-
-                // 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));
-
-                // 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_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;
-
-                // 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;
-
-                // 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);
-
-                sum_t +=  __riscv_vmv_x_s_i32m1_i32(isum3);
-
-                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;
-
-            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"
-            );
-
-            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;
-
-#elif defined(__POWER9_VECTOR__)
-    const vector signed char lowMask = vec_splats((signed char)0x3);
-    const vector signed char lowMask1 = vec_splats((int8_t)0xf);
-    const vector signed char lowMask2 = vec_splats((int8_t)0x30);
-    const vector int v0 = vec_splats((int32_t)0);
-    const vector signed char v1 = vec_splats((signed char)0x1);
-    const vector unsigned char v2 = vec_splats((unsigned char)0x2);
-    const vector unsigned char v3 = vec_splats((unsigned char)0x3);
-    const vector unsigned char v4 = vec_splats((unsigned char)0x4);
-    const vector unsigned char v6 = vec_splats((unsigned char)0x6);
-    const vector signed char off = vec_splats((signed char)0x20);
-
-    vector float vsumf0 = vec_splats(0.0f);
-    vector float vsumf1 = vec_splats(0.0f);
-    vector float vsumf2 = vec_splats(0.0f);
-    vector float vsumf3 = vec_splats(0.0f);
-
-    for (int i = 0; i < nb; ++i) {
-        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d));
-        vector float vyd = vec_splats(y[i].d);
-        vector float vd = vec_mul(vxd, vyd);
-
-        UNUSED(kmask1);
-        UNUSED(kmask2);
-
-        vector signed char u0 = (vector signed char)vec_xl_len(x[i].scales, 8);
-        vector signed char u1 = vec_and(u0, lowMask1);
-        vector signed char u2 = (vector signed char)vec_xl_len(x[i].scales + 8, 4);
-        vector signed char u3 = (vector signed char)vec_mergeh((vector signed int)u2, (vector signed int)vec_sr(u2, v2));
-        vector signed char u30 = vec_sl(vec_and(u3, lowMask), v4);
-        vector signed char u31 = vec_and(u3, lowMask2);
-
-        u1 = vec_or(u1, u30);
-        u2 = vec_or(vec_sr(u0, v4), u31);
-
-        vector signed char vscales = (vector signed char)vec_mergeh((vector signed long long)u1, (vector signed long long)u2);
-        vector signed char qxhs0 = (vector signed char)vec_xl( 0, x[i].hmask);
-        vector signed char qxhs1 = (vector signed char)vec_xl(16, x[i].hmask);
-
-        vscales = vec_sub(vscales, off);
-
-        vector signed int vsumi0 = v0;
-        vector signed int vsumi1 = v0;
-        vector signed int vsumi2 = v0;
-        vector signed int vsumi3 = v0;
-        vector signed int vsumi4 = v0;
-        vector signed int vsumi5 = v0;
-        vector signed int vsumi6 = v0;
-        vector signed int vsumi7 = v0;
-
-        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);
-            __builtin_prefetch(q8, 0, 1);
-
-            vector signed char qxs0 = (vector signed char)vec_xl( 0, q3);
-            vector signed char qxs1 = (vector signed char)vec_xl(16, q3);
-            q3 += 32;
-
-            //the low 2 bits
-            vector signed char qxs00 = vec_and(qxs0, lowMask);
-            vector signed char qxs01 = vec_and(vec_sr(qxs0, v2), lowMask);
-            vector signed char qxs02 = vec_and(vec_sr(qxs0, v4), lowMask);
-            vector signed char qxs03 = vec_and(vec_sr(qxs0, v6), lowMask);
-            vector signed char qxs10 = vec_and(qxs1, lowMask);
-            vector signed char qxs11 = vec_and(vec_sr(qxs1, v2), lowMask);
-            vector signed char qxs12 = vec_and(vec_sr(qxs1, v4), lowMask);
-            vector signed char qxs13 = vec_and(vec_sr(qxs1, v6), lowMask);
-
-            //the 3rd bit
-            vector signed char qxh00 = vec_sl(vec_andc(v1, qxhs0), v2);
-            vector signed char qxh01 = vec_sl(vec_andc(v1, vec_sr(qxhs0, (vector unsigned char)v1)), v2);
-            vector signed char qxh02 = vec_sl(vec_andc(v1, vec_sr(qxhs0, v2)), v2);
-            vector signed char qxh03 = vec_sl(vec_andc(v1, vec_sr(qxhs0, v3)), v2);
-            vector signed char qxh10 = vec_sl(vec_andc(v1, qxhs1), v2);
-            vector signed char qxh11 = vec_sl(vec_andc(v1, vec_sr(qxhs1, (vector unsigned char)v1)), v2);
-            vector signed char qxh12 = vec_sl(vec_andc(v1, vec_sr(qxhs1, v2)), v2);
-            vector signed char qxh13 = vec_sl(vec_andc(v1, vec_sr(qxhs1, v3)), v2);
-            qxhs0 = vec_sr(qxhs0, v4);
-            qxhs1 = vec_sr(qxhs1, v4);
-
-            vector signed char q3x00 = vec_sub(qxs00, qxh00);
-            vector signed char q3x01 = vec_sub(qxs01, qxh01);
-            vector signed char q3x02 = vec_sub(qxs02, qxh02);
-            vector signed char q3x03 = vec_sub(qxs03, qxh03);
-            vector signed char q3x10 = vec_sub(qxs10, qxh10);
-            vector signed char q3x11 = vec_sub(qxs11, qxh11);
-            vector signed char q3x12 = vec_sub(qxs12, qxh12);
-            vector signed char q3x13 = vec_sub(qxs13, qxh13);
-
-            vector signed char q8y00 = vec_xl(  0, q8);
-            vector signed char q8y10 = vec_xl( 16, q8);
-            vector signed char q8y01 = vec_xl( 32, q8);
-            vector signed char q8y11 = vec_xl( 48, q8);
-            vector signed char q8y02 = vec_xl( 64, q8);
-            vector signed char q8y12 = vec_xl( 80, q8);
-            vector signed char q8y03 = vec_xl( 96, q8);
-            vector signed char q8y13 = vec_xl(112, q8);
-            q8 += 128;
-
-            vector signed short vscales_h = vec_unpackh(vscales);
-            vector signed short vs0 = vec_splat(vscales_h, 0);
-            vector signed short vs1 = vec_splat(vscales_h, 1);
-            vector signed short vs2 = vec_splat(vscales_h, 2);
-            vector signed short vs3 = vec_splat(vscales_h, 3);
-            vector signed short vs4 = vec_splat(vscales_h, 4);
-            vector signed short vs5 = vec_splat(vscales_h, 5);
-            vector signed short vs6 = vec_splat(vscales_h, 6);
-            vector signed short vs7 = vec_splat(vscales_h, 7);
-            vscales = vec_sld(vscales, vscales, 8);
-
-            vector signed short qv00 = vec_add(vec_mule(q3x00, q8y00), vec_mulo(q3x00, q8y00));
-            vector signed short qv01 = vec_add(vec_mule(q3x01, q8y01), vec_mulo(q3x01, q8y01));
-            vector signed short qv02 = vec_add(vec_mule(q3x02, q8y02), vec_mulo(q3x02, q8y02));
-            vector signed short qv03 = vec_add(vec_mule(q3x03, q8y03), vec_mulo(q3x03, q8y03));
-            vector signed short qv10 = vec_add(vec_mule(q3x10, q8y10), vec_mulo(q3x10, q8y10));
-            vector signed short qv11 = vec_add(vec_mule(q3x11, q8y11), vec_mulo(q3x11, q8y11));
-            vector signed short qv12 = vec_add(vec_mule(q3x12, q8y12), vec_mulo(q3x12, q8y12));
-            vector signed short qv13 = vec_add(vec_mule(q3x13, q8y13), vec_mulo(q3x13, q8y13));
-
-            vsumi0 = vec_msum(qv00, vs0, vsumi0);
-            vsumi1 = vec_msum(qv01, vs2, vsumi1);
-            vsumi2 = vec_msum(qv02, vs4, vsumi2);
-            vsumi3 = vec_msum(qv03, vs6, vsumi3);
-            vsumi4 = vec_msum(qv10, vs1, vsumi4);
-            vsumi5 = vec_msum(qv11, vs3, vsumi5);
-            vsumi6 = vec_msum(qv12, vs5, vsumi6);
-            vsumi7 = vec_msum(qv13, vs7, vsumi7);
-        }
-
-        vsumi0 = vec_add(vsumi0, vsumi4);
-        vsumi1 = vec_add(vsumi1, vsumi5);
-        vsumi2 = vec_add(vsumi2, vsumi6);
-        vsumi3 = vec_add(vsumi3, vsumi7);
-
-        vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
-        vsumf1 = vec_madd(vec_ctf(vsumi1, 0), vd, vsumf1);
-        vsumf2 = vec_madd(vec_ctf(vsumi2, 0), vd, vsumf2);
-        vsumf3 = vec_madd(vec_ctf(vsumi3, 0), vd, vsumf3);
-    }
-
-    vsumf0 = vec_add(vsumf0, vsumf2);
-    vsumf1 = vec_add(vsumf1, vsumf3);
-
-    vsumf0 = vec_add(vsumf0, vsumf1);
-
-    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4));
-    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8));
-
-    *s = vec_extract(vsumf0, 0);
-
-#elif defined __loongarch_asx
-
-    const __m128i m32 = __lsx_vreplgr2vr_b(32);
-
-    __m256 acc = (__m256)__lasx_xvldi(0);
-
-    uint32_t aux[3];
-
-    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 = 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(
-                ((aux[1] >> 4) & kmask2) | (((aux[2] >> 6) & kmask1) << 4),
-                ((aux[0] >> 4) & kmask2) | (((aux[2] >> 4) & kmask1) << 4),
-                (aux[1] & kmask2) | (((aux[2] >> 2) & kmask1) << 4),
-                (aux[0] & kmask2) | (((aux[2] >> 0) & kmask1) << 4));
-        scales128 = __lsx_vsub_b(scales128, m32);
-
-        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);
-
-        // integer accumulator
-        __m256i sumi = __lasx_xvldi(0);
-
-        for (int j = 0; j < QK_K/128; ++j) {
-            // load low 2 bits
-            const __m256i q3bits = __lasx_xvld((const __m256i*)q3, 0); q3 += 32;
-
-            // prepare low and high bits
-            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;
-            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 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_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);
-            p16_2 = __lasx_xvadd_w(p16_2, p16_3);
-            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);
-    }
-
-    *s = hsum_float_8(acc);
-#elif defined(__VXE__) || defined(__VXE2__)
-    uint32_t aux[3];
-    uint32_t utmp[4];
-
-    const int32x4_t v_z = vec_splat_s32(0);
-    const uint8x16_t v_3m = vec_splat_u8(0x03);
-
-    const uint8x16_t v_0c = vec_splat_u8(1);
-    const uint8x16_t v_1c = vec_sl(v_0c, 1);
-    const uint8x16_t v_2c = vec_sl(v_0c, 2);
-    const uint8x16_t v_3c = vec_sl(v_0c, 3);
-
-    uint8x16_t q3h[4];
-    uint8x16_t q3b[2];
-    int8x16_t q3bytes[4];
-    int8x16_t q8bytes[4];
-    uint8x16_t qhbits[2];
-
-    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 * restrict x0l = x[i].qs;
-        const uint8_t * restrict x0h = x[i].hmask;
-        const int8_t  * restrict y0  = y[i].qs;
-
-        qhbits[0] = vec_xl(0 , x0h);
-        qhbits[1] = vec_xl(16, x0h);
-
-        int32_t isum = 0;
-
-        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;
-
-        for (int j = 0; j < QK_K/128; ++j) {
-            int32x4_t isum0, isum1, isum2, isum3;
-
-            q3b[0] = vec_xl(0 , x0l);
-            q3b[1] = vec_xl(16, x0l);
-            x0l += 32;
-
-            q8bytes[0] = vec_xl(0  , y0);
-            q8bytes[1] = vec_xl(16 , y0);
-            q8bytes[2] = vec_xl(32 , y0);
-            q8bytes[3] = vec_xl(48 , y0);
-            q8bytes[4] = vec_xl(64 , y0);
-            q8bytes[5] = vec_xl(80 , y0);
-            q8bytes[6] = vec_xl(96 , y0);
-            q8bytes[7] = vec_xl(112, y0);
-            y0 += 128;
-
-            q3h[0] = vec_sl(vec_andc(v_0c, qhbits[0]), 2);
-            q3h[1] = vec_sl(vec_andc(v_0c, qhbits[1]), 2);
-            q3h[2] = vec_sl(vec_andc(v_1c, qhbits[0]), 1);
-            q3h[3] = vec_sl(vec_andc(v_1c, qhbits[1]), 1);
-
-            q3bytes[0] = vec_sub((int8x16_t)vec_and(q3b[0], v_3m), (int8x16_t)q3h[0]);
-            q3bytes[1] = vec_sub((int8x16_t)vec_and(q3b[1], v_3m), (int8x16_t)q3h[1]);
-            q3bytes[2] = vec_sub((int8x16_t)vec_and(vec_sr(q3b[0], 2), v_3m), (int8x16_t)q3h[2]);
-            q3bytes[3] = vec_sub((int8x16_t)vec_and(vec_sr(q3b[1], 2), v_3m), (int8x16_t)q3h[3]);
-
-            isum0 = ggml_vec_dot(v_z, q3bytes[0], q8bytes[0]);
-            isum1 = ggml_vec_dot(v_z, q3bytes[1], q8bytes[1]);
-            isum2 = ggml_vec_dot(v_z, q3bytes[2], q8bytes[2]);
-            isum3 = ggml_vec_dot(v_z, q3bytes[3], q8bytes[3]);
-
-            isum += (isum0[0] + isum0[1] + isum0[2] + isum0[3]) * scale[0];
-            isum += (isum1[0] + isum1[1] + isum1[2] + isum1[3]) * scale[1];
-            isum += (isum2[0] + isum2[1] + isum2[2] + isum2[3]) * scale[2];
-            isum += (isum3[0] + isum3[1] + isum3[2] + isum3[3]) * scale[3];
-
-            scale += 4;
-
-            q3h[0] = vec_andc(v_2c, qhbits[0]);
-            q3h[1] = vec_andc(v_2c, qhbits[1]);
-            q3h[2] = vec_sr(vec_andc(v_3c, qhbits[0]), 1);
-            q3h[3] = vec_sr(vec_andc(v_3c, qhbits[1]), 1);
-
-            q3bytes[0] = vec_sub((int8x16_t)vec_and(vec_sr(q3b[0], 4), v_3m), (int8x16_t)q3h[0]);
-            q3bytes[1] = vec_sub((int8x16_t)vec_and(vec_sr(q3b[1], 4), v_3m), (int8x16_t)q3h[1]);
-            q3bytes[2] = vec_sub((int8x16_t)vec_and(vec_sr(q3b[0], 6), v_3m), (int8x16_t)q3h[2]);
-            q3bytes[3] = vec_sub((int8x16_t)vec_and(vec_sr(q3b[1], 6), v_3m), (int8x16_t)q3h[3]);
-
-            isum0 = ggml_vec_dot(v_z, q3bytes[0], q8bytes[4]);
-            isum1 = ggml_vec_dot(v_z, q3bytes[1], q8bytes[5]);
-            isum2 = ggml_vec_dot(v_z, q3bytes[2], q8bytes[6]);
-            isum3 = ggml_vec_dot(v_z, q3bytes[3], q8bytes[7]);
-
-            isum += (isum0[0] + isum0[1] + isum0[2] + isum0[3]) * scale[0];
-            isum += (isum1[0] + isum1[1] + isum1[2] + isum1[3]) * scale[1];
-            isum += (isum2[0] + isum2[1] + isum2[2] + isum2[3]) * scale[2];
-            isum += (isum3[0] + isum3[1] + isum3[2] + isum3[3]) * scale[3];
-
-            scale += 4;
-
-            if (j == 0) {
-                qhbits[0] = vec_sr(qhbits[0], 4);
-                qhbits[1] = vec_sr(qhbits[1], 4);
-            }
-        }
-
-        sum += d * isum;
-    }
-
-    *s = sum;
-#else
-    // scalar version
-    // This function is written like this so the compiler can manage to vectorize most of it
-    // Using -Ofast, GCC and clang manage to produce code that is within a factor of 2 or so from the
-    // manually vectorized version above. Every other version I tried would run at least 4 times slower.
-    // The ideal situation would be if we could just write the code once, and the compiler would
-    // automatically produce the best possible set of machine instructions, instead of us having to manually
-    // write vectorized versions for AVX, ARM_NEON, etc.
-
-    int8_t  aux8[QK_K];
-    int16_t aux16[8];
-    float   sums [8];
-    int32_t aux32[8];
-    memset(sums, 0, 8*sizeof(float));
-
-    uint32_t auxs[4];
-    const int8_t * scales = (const int8_t*)auxs;
-
-    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;
-        memset(aux32, 0, 8*sizeof(int32_t));
-        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;
-            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
-            a += 32; m <<= 1;
-            for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 2) & 3;
-            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
-            a += 32; m <<= 1;
-            for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 4) & 3;
-            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
-            a += 32; m <<= 1;
-            for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 6) & 3;
-            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
-            a += 32; m <<= 1;
-            q3 += 32;
-        }
-        a = aux8;
-
-        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);
-        for (int j = 0; j < QK_K/16; ++j) {
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
-            q8 += 8; a += 8;
-        }
-        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];
-    }
-    for (int l = 0; l < 8; ++l) sumf += sums[l];
-    *s = sumf;
-
-#endif
-
-}
-
-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);
-#ifdef __ARM_FEATURE_MATMUL_INT8
-    assert((nrc == 2) || (nrc == 1));
-#else
-    assert(nrc == 1);
-#endif
-    UNUSED(nrc);
-    UNUSED(bx);
-    UNUSED(by);
-    UNUSED(bs);
-
-    const block_q4_K * GGML_RESTRICT x = vx;
-    const block_q8_K * GGML_RESTRICT y = vy;
-
-    const int nb = n / QK_K;
-
-    static const uint32_t kmask1 = 0x3f3f3f3f;
-    static const uint32_t kmask2 = 0x0f0f0f0f;
-    static const uint32_t kmask3 = 0x03030303;
-
-    uint32_t utmp[4];
-
-#if defined(__ARM_FEATURE_MATMUL_INT8)
-    if (nrc == 2) {
-        const block_q4_K * GGML_RESTRICT x0 = x;
-        const block_q4_K * GGML_RESTRICT x1 = (const block_q4_K *) ((const uint8_t *)vx + bx);
-        const block_q8_K * GGML_RESTRICT y0 = y;
-        const block_q8_K * GGML_RESTRICT y1 = (const block_q8_K *) ((const uint8_t *)vy + by);
-
-        const uint8x16_t m4b = vdupq_n_u8(0x0f);
-
-        float32x4_t vfsum = vdupq_n_f32(0.0f);
-
-        for (int i = 0; i < nb; ++i, ++x0, ++x1, ++y0, ++y1) {
-            const uint8_t * GGML_RESTRICT qx0 = x0->qs;
-            const uint8_t * GGML_RESTRICT qx1 = x1->qs;
-            const  int8_t * GGML_RESTRICT qy0 = y0->qs;
-            const  int8_t * GGML_RESTRICT qy1 = y1->qs;
-
-            // decode scales and mins
-            int8_t x0_scales[8], x1_scales[8];
-            int16x8_t x0_mins, x1_mins;
-            {
-                uint32_t scales_mins[3];
-                memcpy(scales_mins, x0->scales, 12);
-                const uint32_t mins_0_3 = scales_mins[1] & kmask1;
-                const uint32_t mins_4_7 = ((scales_mins[2] >> 4) & kmask2) | (((scales_mins[1] >> 6) & kmask3) << 4);
-                const uint32x2_t mins = {mins_0_3, mins_4_7};
-                x0_mins = vreinterpretq_s16_u16(vmovl_u8(vreinterpret_u8_u32(mins)));
-                uint32_t scales[2];
-                scales[0] = scales_mins[0] & kmask1; // scales 0~3
-                scales[1] = (scales_mins[2] & kmask2) | (((scales_mins[0] >> 6) & kmask3) << 4); // scales 4~7
-                memcpy(x0_scales, scales, 8);
-            }
-            {
-                uint32_t scales_mins[3];
-                memcpy(scales_mins, x1->scales, 12);
-                const uint32_t mins_0_3 = scales_mins[1] & kmask1;
-                const uint32_t mins_4_7 = ((scales_mins[2] >> 4) & kmask2) | (((scales_mins[1] >> 6) & kmask3) << 4);
-                const uint32x2_t mins = {mins_0_3, mins_4_7};
-                x1_mins = vreinterpretq_s16_u16(vmovl_u8(vreinterpret_u8_u32(mins)));
-                uint32_t scales[2];
-                scales[0] = scales_mins[0] & kmask1; // scales 0~3
-                scales[1] = (scales_mins[2] & kmask2) | (((scales_mins[0] >> 6) & kmask3) << 4); // scales 4~7
-                memcpy(x1_scales, scales, 8);
-            }
-
-            int32x4_t visum = {0};
-
-            // process 64 data points per iteration, totally 256 data points
-            for (int j = 0; j < QK_K / 64; ++j, qx0 += 32, qx1 += 32, qy0 += 64, qy1 += 64) {
-                const int8x16x4_t vy0 = vld1q_s8_x4(qy0);
-                const int8x16x4_t vy1 = vld1q_s8_x4(qy1);
-
-                int8x16_t vx0[4], vx1[4];
-                {
-                    const uint8x16x2_t vv = vld1q_u8_x2(qx0);
-                    vx0[0] = vreinterpretq_s8_u8(vandq_u8(vv.val[0], m4b));
-                    vx0[1] = vreinterpretq_s8_u8(vandq_u8(vv.val[1], m4b));
-                    vx0[2] = vreinterpretq_s8_u8(vshrq_n_u8(vv.val[0], 4));
-                    vx0[3] = vreinterpretq_s8_u8(vshrq_n_u8(vv.val[1], 4));
-                }
-                {
-                    const uint8x16x2_t vv = vld1q_u8_x2(qx1);
-                    vx1[0] = vreinterpretq_s8_u8(vandq_u8(vv.val[0], m4b));
-                    vx1[1] = vreinterpretq_s8_u8(vandq_u8(vv.val[1], m4b));
-                    vx1[2] = vreinterpretq_s8_u8(vshrq_n_u8(vv.val[0], 4));
-                    vx1[3] = vreinterpretq_s8_u8(vshrq_n_u8(vv.val[1], 4));
-                }
-
-                // process 32 data points (share same block scale) per iteration
-                for (int k = 0; k < 2; ++k) {
-                    const int blk = j * 2 + k;
-                    const int32x4_t block_scale = {
-                        x0_scales[blk],
-                        x0_scales[blk],
-                        x1_scales[blk],
-                        x1_scales[blk],
-                    };
-
-                    int32x4_t vr = {0};
-                    for (int l = 0; l < 2; ++l) {
-                        const int idx = k * 2 + l;
-                        const int64x2_t vx0_s64 = vreinterpretq_s64_s8(vx0[idx]);
-                        const int64x2_t vx1_s64 = vreinterpretq_s64_s8(vx1[idx]);
-                        const int64x2_t vy0_s64 = vreinterpretq_s64_s8(vy0.val[idx]);
-                        const int64x2_t vy1_s64 = vreinterpretq_s64_s8(vy1.val[idx]);
-                        const int8x16_t vx_l = vreinterpretq_s8_s64(vzip1q_s64(vx0_s64, vx1_s64));
-                        const int8x16_t vx_h = vreinterpretq_s8_s64(vzip2q_s64(vx0_s64, vx1_s64));
-                        const int8x16_t vy_l = vreinterpretq_s8_s64(vzip1q_s64(vy0_s64, vy1_s64));
-                        const int8x16_t vy_h = vreinterpretq_s8_s64(vzip2q_s64(vy0_s64, vy1_s64));
-                        vr = vmmlaq_s32(vr, vx_l, vy_l);
-                        vr = vmmlaq_s32(vr, vx_h, vy_h);
-                    }
-                    // apply block scale, will NOT overflow
-                    // block_scale * sum_256(int4*int8) <= 2^(8+8+4+8) = 28 bits
-                    visum = vmlaq_s32(visum, vr, block_scale);
-                }
-            }
-
-            // adjust bias, apply superblock scale
-            {
-                int32_t bias[4];
-                // no obvious uplift from sve sdot-16, just use neon mul add
-                const int16x8_t y0_sums = vpaddq_s16(vld1q_s16(y0->bsums), vld1q_s16(y0->bsums+8));
-                const int16x8_t y1_sums = vpaddq_s16(vld1q_s16(y1->bsums), vld1q_s16(y1->bsums+8));
-                bias[0] = vaddvq_s32(vaddq_s32(vmull_s16(vget_low_s16(y0_sums), vget_low_s16(x0_mins)),
-                                               vmull_s16(vget_high_s16(y0_sums), vget_high_s16(x0_mins))));
-                bias[1] = vaddvq_s32(vaddq_s32(vmull_s16(vget_low_s16(y1_sums), vget_low_s16(x0_mins)),
-                                               vmull_s16(vget_high_s16(y1_sums), vget_high_s16(x0_mins))));
-                bias[2] = vaddvq_s32(vaddq_s32(vmull_s16(vget_low_s16(y0_sums), vget_low_s16(x1_mins)),
-                                               vmull_s16(vget_high_s16(y0_sums), vget_high_s16(x1_mins))));
-                bias[3] = vaddvq_s32(vaddq_s32(vmull_s16(vget_low_s16(y1_sums), vget_low_s16(x1_mins)),
-                                               vmull_s16(vget_high_s16(y1_sums), vget_high_s16(x1_mins))));
-                const float32x4_t dmins = {
-                    GGML_FP16_TO_FP32(x0->dmin) * y0->d,
-                    GGML_FP16_TO_FP32(x0->dmin) * y1->d,
-                    GGML_FP16_TO_FP32(x1->dmin) * y0->d,
-                    GGML_FP16_TO_FP32(x1->dmin) * y1->d,
-                };
-                vfsum = vmlsq_f32(vfsum, vcvtq_f32_s32(vld1q_s32(bias)), dmins);
-
-                const float32x4_t superblock_scale = {
-                    GGML_FP16_TO_FP32(x0->d) * y0->d,
-                    GGML_FP16_TO_FP32(x0->d) * y1->d,
-                    GGML_FP16_TO_FP32(x1->d) * y0->d,
-                    GGML_FP16_TO_FP32(x1->d) * y1->d,
-                };
-                vfsum = vmlaq_f32(vfsum, vcvtq_f32_s32(visum), superblock_scale);
-            }
-        }
-
-        // vfsum = ABCD -> ACBD
-        // AC -> s, BD -> (s+bs)
-        vfsum = vzip1q_f32(vfsum, vextq_f32(vfsum, vfsum, 2));
-        vst1_f32(s,      vget_low_f32 (vfsum));
-        vst1_f32(s + bs, vget_high_f32(vfsum));
-
-        return;
-    }
-#endif
-
-#ifdef __ARM_FEATURE_SVE
-    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 q8sums = vpaddq_s16(vld1q_s16(y[i].bsums), vld1q_s16(y[i].bsums + 8));
-
-        memcpy(utmp, x[i].scales, K_SCALE_SIZE);
-
-        uint32x2_t mins8 = { 0 };
-        mins8 = vset_lane_u32(utmp[1] & kmask1, mins8, 0);
-        mins8 = vset_lane_u32(((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4), mins8, 1);
-
-        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
-        utmp[0] &= kmask1;
-
-        const int16x8_t mins = vreinterpretq_s16_u16(vmovl_u8(vreinterpret_u8_u32(mins8)));
-        const int32x4_t prod = vaddq_s32(vmull_s16(vget_low_s16 (q8sums), vget_low_s16 (mins)),
-                                         vmull_s16(vget_high_s16(q8sums), vget_high_s16(mins)));
-        sumf -= dmin * vaddvq_s32(prod);
-
-        const uint8_t * scales = (const uint8_t *)utmp;
-
-        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);
-        const svint32_t mzero = svdup_n_s32(0);
-        svint32_t sumi1 = svdup_n_s32(0);
-        svint32_t sumi1_1 = svdup_n_s32(0);
-        svint32_t sumi1_2 = svdup_n_s32(0);
-        svint32_t sumi2 = svdup_n_s32(0);
-        svint32_t sumi2_1 = svdup_n_s32(0);
-        svint32_t sumi2_2 = svdup_n_s32(0);
-        switch (vector_length) {
-            case 128:
-                {
-                    for (int j = 0; j < QK_K/64; ++j) {
-                        svint8_t q4bytes = svreinterpret_s8_u8(svand_u8_x(svptrue_b8(), svld1_u8(svptrue_b8(), q4), m4b));
-                        svint8_t q8bytes = svld1_s8(svptrue_b8(), q8); q8 += 16;
-                        sumi1_1 = svmla_n_s32_x(svptrue_b32(), sumi1_1, svdot_s32(mzero, q4bytes, q8bytes), scales[2*j+0]);
-                        q4bytes = svreinterpret_s8_u8(svand_u8_x(svptrue_b8(), svld1_u8(svptrue_b8(), q4+16), m4b));
-                        q8bytes = svld1_s8(svptrue_b8(), q8); q8 += 16;
-                        sumi1_2 = svmla_n_s32_x(svptrue_b32(), sumi1_2, svdot_s32(mzero, q4bytes, q8bytes), scales[2*j+0]);
-
-                        q4bytes = svreinterpret_s8_u8(svlsr_n_u8_x(svptrue_b8(), svld1_u8(svptrue_b8(), q4), 4));
-                        q8bytes = svld1_s8(svptrue_b8(), q8); q8 += 16;
-                        sumi2_1 = svmla_n_s32_x(svptrue_b32(), sumi2_1, svdot_s32(mzero, q4bytes, q8bytes), scales[2*j+1]);
-                        q4bytes = svreinterpret_s8_u8(svlsr_n_u8_x(svptrue_b8(), svld1_u8(svptrue_b8(), q4+16), 4));
-                        q8bytes = svld1_s8(svptrue_b8(), q8); q8 += 16;
-                        sumi2_2 = svmla_n_s32_x(svptrue_b32(), sumi2_2, svdot_s32(mzero, q4bytes, q8bytes), scales[2*j+1]);
-                        q4 += 32;
-                    }
-                    sumi1 = svadd_s32_x(svptrue_b32(), sumi1_1, sumi1_2);
-                    sumi2 = svadd_s32_x(svptrue_b32(), sumi2_1, sumi2_2);
-                    sumf += d * (svaddv_s32(svptrue_b32(), svadd_s32_x(svptrue_b32(), sumi1, sumi2)));
-                } break;
-            case 256:
-            case 512:
-                {
-                    for (int j = 0; j < QK_K/64; ++j) {
-                        const svuint8_t q4bits  = svld1_u8(svptrue_pat_b8(SV_VL32), q4); q4 += 32;
-                        svint8_t q4bytes = svreinterpret_s8_u8(svand_u8_x(svptrue_pat_b8(SV_VL32), q4bits, m4b));
-                        svint8_t q8bytes = svld1_s8(svptrue_pat_b8(SV_VL32), q8); q8 += 32;
-                        sumi1 = svmla_n_s32_x(svptrue_pat_b32(SV_VL8), sumi1, svdot_s32(mzero, q4bytes, q8bytes), scales[2*j+0]);
-
-                        q4bytes = svreinterpret_s8_u8(svlsr_n_u8_x(svptrue_pat_b8(SV_VL32), q4bits, 4));
-                        q8bytes = svld1_s8(svptrue_pat_b8(SV_VL32), q8); q8 += 32;
-                        sumi2 = svmla_n_s32_x(svptrue_pat_b32(SV_VL8), sumi2, svdot_s32(mzero, q4bytes, q8bytes), scales[2*j+1]);
-                    }
-                    sumf += d * (svaddv_s32(svptrue_pat_b32(SV_VL8), svadd_s32_x(svptrue_pat_b32(SV_VL8), sumi1, sumi2)));
-                } break;
-            default:
-                assert(false && "Unsupported vector length");
-                break;
-        }
-    }
-    *s = sumf;
-#elif defined __ARM_NEON
-    const uint8x16_t m4b = vdupq_n_u8(0xf);
-    const int32x4_t mzero = vdupq_n_s32(0);
-
-    ggml_int8x16x2_t q4bytes;
-    ggml_int8x16x2_t q8bytes;
-
-    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 q8sums = vpaddq_s16(vld1q_s16(y[i].bsums), vld1q_s16(y[i].bsums + 8));
-
-        memcpy(utmp, x[i].scales, 12);
-
-        uint32x2_t mins8 = { 0 };
-        mins8 = vset_lane_u32(utmp[1] & kmask1, mins8, 0);
-        mins8 = vset_lane_u32(((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4), mins8, 1);
-
-        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
-        utmp[0] &= kmask1;
-
-        const int16x8_t mins = vreinterpretq_s16_u16(vmovl_u8(vreinterpret_u8_u32(mins8)));
-        const int32x4_t prod = vaddq_s32(vmull_s16(vget_low_s16 (q8sums), vget_low_s16 (mins)),
-                                         vmull_s16(vget_high_s16(q8sums), vget_high_s16(mins)));
-        sumf -= dmin * vaddvq_s32(prod);
-
-        const uint8_t * scales = (const uint8_t *)utmp;
-
-        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;
-
-        for (int j = 0; j < QK_K/64; ++j) {
-            const ggml_uint8x16x2_t q4bits = ggml_vld1q_u8_x2(q4); q4 += 32;
-
-            q8bytes = ggml_vld1q_s8_x2(q8); q8 += 32;
-            q4bytes.val[0] = vreinterpretq_s8_u8(vandq_u8  (q4bits.val[0], m4b));
-            q4bytes.val[1] = vreinterpretq_s8_u8(vandq_u8  (q4bits.val[1], m4b));
-
-            const int32x4_t p1 = ggml_vdotq_s32(ggml_vdotq_s32(mzero, q4bytes.val[0], q8bytes.val[0]), q4bytes.val[1], q8bytes.val[1]);
-            sumi1 += vaddvq_s32(p1) * scales[2*j+0];
-
-            q8bytes = ggml_vld1q_s8_x2(q8); q8 += 32;
-            q4bytes.val[0] = vreinterpretq_s8_u8(vshrq_n_u8(q4bits.val[0], 4));
-            q4bytes.val[1] = vreinterpretq_s8_u8(vshrq_n_u8(q4bits.val[1], 4));
-
-            const int32x4_t p2 = ggml_vdotq_s32(ggml_vdotq_s32(mzero, q4bytes.val[0], q8bytes.val[0]), q4bytes.val[1], q8bytes.val[1]);
-
-            sumi2 += vaddvq_s32(p2) * scales[2*j+1];
-        }
-
-        sumf += d * (sumi1 + sumi2);
-
-    }
-
-    *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);
-
-    __m256 acc = _mm256_setzero_ps();
-    __m128 acc_m = _mm_setzero_ps();
-
-   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);
-
-        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 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]));
-
-        const __m256i q8sums = _mm256_loadu_si256((const __m256i*)y[i].bsums);
-        const __m128i q8s = _mm_hadd_epi16(_mm256_extracti128_si256(q8sums, 0), _mm256_extracti128_si256(q8sums, 1));
-        const __m128i prod = _mm_madd_epi16(_mm256_extracti128_si256(mins_and_scales, 1), q8s);
-        acc_m = _mm_fmadd_ps(_mm_set1_ps(dmin), _mm_cvtepi32_ps(prod), acc_m);
-
-        const __m128i sc128  = _mm256_extracti128_si256(mins_and_scales, 0);
-        const __m256i scales = MM256_SET_M128I(sc128, sc128);
-
-        __m256i sumi = _mm256_setzero_si256();
-
-        for (int j = 0; j < QK_K/64; ++j) {
-
-            const __m256i scale_l = _mm256_shuffle_epi8(scales, get_scale_shuffle_k4(2*j+0));
-            const __m256i scale_h = _mm256_shuffle_epi8(scales, get_scale_shuffle_k4(2*j+1));
-
-            const __m256i q4bits = _mm256_loadu_si256((const __m256i*)q4); q4 += 32;
-            const __m256i q4l = _mm256_and_si256(q4bits, m4);
-            const __m256i q4h = _mm256_and_si256(_mm256_srli_epi16(q4bits, 4), m4);
-
-            const __m256i q8l = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
-            __m256i p16l = _mm256_maddubs_epi16(q4l, q8l);
-            p16l = _mm256_madd_epi16(scale_l, p16l);
-
-            const __m256i q8h = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
-            __m256i p16h = _mm256_maddubs_epi16(q4h, q8h);
-            p16h = _mm256_madd_epi16(scale_h, p16h);
-            const __m256i sumj = _mm256_add_epi32(p16l, p16h);
-
-            sumi = _mm256_add_epi32(sumi, sumj);
-        }
-
-        __m256 vd = _mm256_set1_ps(d);
-        acc = _mm256_fmadd_ps(vd, _mm256_cvtepi32_ps(sumi), acc);
-
-    }
-
-    acc_m = _mm_add_ps(acc_m, _mm_movehl_ps(acc_m, acc_m));
-    acc_m = _mm_add_ss(acc_m, _mm_movehdup_ps(acc_m));
-
-    *s = hsum_float_8(acc) + _mm_cvtss_f32(acc_m);
-
-#elif defined __AVX__
-
-    const __m128i m4 = _mm_set1_epi8(0xF);
-    const __m128i m2 = _mm_set1_epi8(0x2);
-
-    __m256 acc = _mm256_setzero_ps();
-    __m128 acc_m = _mm_setzero_ps();
-
-   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 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);
-        const uint32_t uaux = utmp[1] & kmask1;
-        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
-        utmp[2] = uaux;
-        utmp[0] &= kmask1;
-
-        const __m128i utmps = _mm_set_epi32(utmp[3], utmp[2], utmp[1], utmp[0]);
-        const __m128i scales = _mm_cvtepu8_epi16(utmps);
-        const __m128i mins = _mm_cvtepu8_epi16(_mm_unpackhi_epi64(utmps, utmps));
-
-        const __m128i q8sums_0 = _mm_loadu_si128((const __m128i*)&y[i].bsums[0]);
-        const __m128i q8sums_1 = _mm_loadu_si128((const __m128i*)&y[i].bsums[8]);
-        const __m128i q8s = _mm_hadd_epi16(q8sums_0, q8sums_1);
-        const __m128i prod = _mm_madd_epi16(mins, q8s);
-        acc_m = _mm_add_ps(_mm_mul_ps(_mm_set1_ps(dmin), _mm_cvtepi32_ps(prod)), acc_m);
-
-        __m128i sumi_0 = _mm_setzero_si128();
-        __m128i sumi_1 = _mm_setzero_si128();
-
-        __m128i shuffle = _mm_set1_epi16(0x0100);
-        for (int j = 0; j < QK_K/64; ++j) {
-
-            const __m128i scale_l = _mm_shuffle_epi8(scales, shuffle);
-            shuffle = _mm_add_epi16(shuffle, m2);
-            const __m128i scale_h = _mm_shuffle_epi8(scales, shuffle);
-            shuffle = _mm_add_epi16(shuffle, m2);
-
-            __m128i q4bits = _mm_loadu_si128((const __m128i*)q4); q4 += 16;
-            const __m128i q4l_0 = _mm_and_si128(q4bits, m4);
-            const __m128i q4h_0 = _mm_and_si128(_mm_srli_epi16(q4bits, 4), m4);
-            q4bits = _mm_loadu_si128((const __m128i*)q4); q4 += 16;
-            const __m128i q4l_1 = _mm_and_si128(q4bits, m4);
-            const __m128i q4h_1 = _mm_and_si128(_mm_srli_epi16(q4bits, 4), m4);
-
-            const __m128i q8l_0 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
-            __m128i p16l = _mm_maddubs_epi16(q4l_0, q8l_0);
-            p16l = _mm_madd_epi16(scale_l, p16l);
-            sumi_0 = _mm_add_epi32(sumi_0, p16l);
-            const __m128i q8l_1 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
-            p16l = _mm_maddubs_epi16(q4l_1, q8l_1);
-            p16l = _mm_madd_epi16(scale_l, p16l);
-            sumi_1 = _mm_add_epi32(sumi_1, p16l);
-
-            const __m128i q8h_0 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
-            __m128i p16h = _mm_maddubs_epi16(q4h_0, q8h_0);
-            p16h = _mm_madd_epi16(scale_h, p16h);
-            sumi_0 = _mm_add_epi32(sumi_0, p16h);
-            const __m128i q8h_1 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
-            p16h = _mm_maddubs_epi16(q4h_1, q8h_1);
-            p16h = _mm_madd_epi16(scale_h, p16h);
-            sumi_1 = _mm_add_epi32(sumi_1, p16h);
-
-        }
-
-        __m256 vd = _mm256_set1_ps(d);
-        __m256i sumi = MM256_SET_M128I(sumi_1, sumi_0);
-        acc = _mm256_add_ps(_mm256_mul_ps(vd, _mm256_cvtepi32_ps(sumi)), acc);
-
-    }
-
-    acc_m = _mm_add_ps(acc_m, _mm_movehl_ps(acc_m, acc_m));
-    acc_m = _mm_add_ss(acc_m, _mm_movehdup_ps(acc_m));
-
-    *s = hsum_float_8(acc) + _mm_cvtss_f32(acc_m);
-
-#elif defined __riscv_xtheadvector
-
-    const uint8_t * scales = (const uint8_t*)&utmp[0];
-    const uint8_t * mins   = (const uint8_t*)&utmp[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);
-
-        int tmp, tmp2, sumi;
-        __asm__ __volatile__(
-            "li %[t1], 12\n\t"
-            "th.vsetvli zero, %[t1], e8, m1\n\t"
-            "th.vlb.v v1, (%[s6b])\n\t" // {aux[0], aux[1], aux[2]}
-            "li %[t1], 4\n\t"
-            "th.vsetvli zero, %[t1], e32, m1\n\t"
-            "th.vslidedown.vi v2, v1, 2\n\t"
-            "th.vmv.v.v v3, v2\n\t"
-            "th.vslideup.vi v2, v3, 1\n\t" // {aux[2], aux[2]}
-            "li %[t1], 2\n\t"
-            "th.vsetvli zero, %[t1], e32, m1\n\t"
-            "th.vmv.v.i v4, 4\n\t"
-            "th.vand.vx v8, v1, %[kmask1]\n\t"
-            "th.vslide1up.vx v5, v4, zero\n\t" // {0, 4}
-            "th.vsrl.vi v6, v1, 6\n\t"
-            "th.vsrl.vv v7, v2, v5\n\t"
-            "th.vand.vx v0, v6, %[kmask3]\n\t"
-            "th.vand.vx v2, v7, %[kmask2]\n\t"
-            "th.vsll.vi v6, v0, 4\n\t"
-            "li %[t2], 8\n\t"
-            "addi %[t1], %[utmp], 4\n\t"
-            "th.vor.vv v1, v6, v2\n\t"
-            "th.vssw.v v8, (%[utmp]), %[t2]\n\t"
-            "th.vssw.v v1, (%[t1]), %[t2]\n\t"
-            "th.vsetvli zero, zero, e32, m2\n\t" // vl == 8
-            "th.vlw.v v2, (%[bsums])\n\t"
-            "th.vsetvli zero, %[t2], e16, m1\n\t"
-            "th.vnsrl.vi v0, v2, 0\n\t"
-            "th.vnsrl.vi v1, v2, 16\n\t"
-            "th.vadd.vv v2, v0, v1\n\t"
-            "th.vlbu.v v4, (%[mins])\n\t"
-            "th.vwmul.vv v6, v4, v2\n\t"
-            "th.vmv.v.x v0, zero\n\t"
-            "th.vsetvli zero, %[t2], e32, m2\n\t"
-            "th.vredsum.vs v0, v6, v0\n\t"
-            "th.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__(
-                "th.vsetvli zero, %[vl128], e8, m8\n\t"
-                "th.vlb.v v8, (%[q8])\n\t"
-                "th.vsetvli zero, %[vl64], e8, m4\n\t"
-                "th.vlb.v v0, (%[q4])\n\t"
-                "th.vsrl.vi v4, v0, 4\n\t"
-                "th.vand.vi v0, v0, 0xF\n\t"
-                "th.vsetvli zero, %[vl32], e8, m2\n\t"
-                "th.vwmul.vv v28, v6, v14\n\t"
-                "th.vwmul.vv v20, v4, v10\n\t"
-                "th.vwmul.vv v24, v2, v12\n\t"
-                "th.vwmul.vv v16, v0, v8\n\t"
-                "li %[tmp], 4\n\t"
-                "th.vsetvli zero, %[tmp], e32, m1\n\t"
-                "th.vlbu.v v1, (%[scale])\n\t"
-                "th.vmv.v.x v0, zero\n\t"
-                "th.vsetvli zero, %[vl32], e16, m4\n\t"
-                "th.vwredsum.vs v6, v24, v0\n\t"
-                "th.vwredsum.vs v7, v28, v0\n\t"
-                "th.vwredsum.vs v4, v16, v0\n\t"
-                "th.vwredsum.vs v5, v20, v0\n\t"
-                "th.vsetvli zero, %[tmp], e32, m1\n\t"
-                "th.vslideup.vi v6, v7, 1\n\t"
-                "th.vslideup.vi v4, v5, 1\n\t"
-                "th.vslideup.vi v4, v6, 2\n\t"
-                "th.vmul.vv v8, v4, v1\n\t"
-                "th.vredsum.vs v0, v8, v0\n\t"
-                "th.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;
-
-    }
-
-    *s = sumf;
-
-#elif defined __riscv_v
-
-    const uint8_t * scales = (const uint8_t*)&utmp[0];
-    const uint8_t * mins   = (const uint8_t*)&utmp[2];
-
-    float sumf = 0;
-    const int vector_length = __riscv_vlenb() * 8;
-
-    switch (vector_length) {
-    case 256:
-        for (int i = 0; i < nb; ++i) {
-
-            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);
-
-            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;
-
-            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);
-
-            const uint8_t * GGML_RESTRICT q4 = x[i].qs;
-            const int8_t  * GGML_RESTRICT q8 = y[i].qs;
-
-            vl = 32;
-
-            int32_t sum_1 = 0;
-            int32_t sum_2 = 0;
-
-            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);
-
-                // 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];
-
-                // 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];
-
-                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);
-
-            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;
-
-#elif defined(__POWER9_VECTOR__)
-    const vector signed char lowMask = vec_splats((signed char)0xF);
-    const vector signed char lowMask1 = vec_splats((int8_t)0x3f);
-    const vector signed char lowMask2 = vec_splats((int8_t)0x30);
-    const vector int v0 = vec_splats((int32_t)0);
-    const vector unsigned char v2 = vec_splats((uint8_t)2);
-    const vector unsigned char v4 = vec_splats((unsigned char)0x4);
-
-    vector float vsumf0 = vec_splats(0.0f);
-    vector float vsumf1 = vec_splats(0.0f);
-    vector float vsumf2 = vec_splats(0.0f);
-    vector float vsumf3 = vec_splats(0.0f);
-
-    for (int i = 0; i < nb; ++i) {
-        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d));
-        vector float vyd = vec_splats(y[i].d);
-        vector float vd = vec_mul(vxd, vyd);
-
-        vector float vxmin = vec_splats(GGML_FP16_TO_FP32(x[i].dmin));
-        vector float vdmin = vec_mul(vxmin, vyd);
-
-        vector signed short q8ysums0 = vec_xl( 0, y[i].bsums);
-        vector signed short q8ysums1 = vec_xl(16, y[i].bsums);
-
-        UNUSED(kmask1);
-        UNUSED(kmask2);
-        UNUSED(kmask3);
-        UNUSED(utmp);
-
-        vector signed char u0 = (vector signed char)vec_xl_len(x[i].scales, 8);
-        vector signed char u1 = vec_and(vec_sr(u0, v2), lowMask2);
-        vector signed char u2 = (vector signed char)vec_xl_len(x[i].scales + 8, 4);
-        vector signed char u3 = vec_sr(u2, v4);
-
-        vector signed char u30 = u1;
-        vector signed char u31 = (vector signed char)vec_mergeh((vector signed int)vec_and(u2, lowMask), (vector signed int)u3);
-
-        u1 = vec_and(u0, lowMask1);
-        u2 = vec_or(u30, u31);
-
-        vector signed char utmps = (vector signed char)vec_mergeh((vector signed int)u1, (vector signed int)u2);
-
-        vector signed short vscales = vec_unpackh(utmps);
-        vector signed short q4xmins = vec_unpackl(utmps);
-        vector signed short q4xmins0 = vec_mergeh(q4xmins, q4xmins);
-        vector signed short q4xmins1 = vec_mergel(q4xmins, q4xmins);
-
-        vector signed int prod0 = vec_mule(q4xmins0, q8ysums0);
-        vector signed int prod1 = vec_mule(q4xmins1, q8ysums1);
-        vector signed int prod2 = vec_mulo(q4xmins0, q8ysums0);
-        vector signed int prod3 = vec_mulo(q4xmins1, q8ysums1);
-
-        vsumf0 = vec_nmsub(vec_ctf(prod0, 0), vdmin, vsumf0);
-        vsumf1 = vec_nmsub(vec_ctf(prod1, 0), vdmin, vsumf1);
-        vsumf2 = vec_nmsub(vec_ctf(prod2, 0), vdmin, vsumf2);
-        vsumf3 = vec_nmsub(vec_ctf(prod3, 0), vdmin, vsumf3);
-
-        vector signed int vsumi0 = v0;
-        vector signed int vsumi1 = v0;
-        vector signed int vsumi2 = v0;
-        vector signed int vsumi3 = v0;
-
-        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);
-            __builtin_prefetch(q8, 0, 1);
-
-            vector signed char qxs0 = (vector signed char)vec_xl( 0, q4);
-            vector signed char qxs1 = (vector signed char)vec_xl(16, q4);
-            vector signed char qxs2 = (vector signed char)vec_xl(32, q4);
-            vector signed char qxs3 = (vector signed char)vec_xl(48, q4);
-            q4 += 64;
-
-            vector unsigned char q4x00 = (vector unsigned char)vec_and(qxs0, lowMask);
-            vector unsigned char q4x01 = (vector unsigned char)vec_sr(qxs0, v4);
-            vector unsigned char q4x10 = (vector unsigned char)vec_and(qxs1, lowMask);
-            vector unsigned char q4x11 = (vector unsigned char)vec_sr(qxs1, v4);
-            vector unsigned char q4x20 = (vector unsigned char)vec_and(qxs2, lowMask);
-            vector unsigned char q4x21 = (vector unsigned char)vec_sr(qxs2, v4);
-            vector unsigned char q4x30 = (vector unsigned char)vec_and(qxs3, lowMask);
-            vector unsigned char q4x31 = (vector unsigned char)vec_sr(qxs3, v4);
-
-            vector signed char q8y00 = vec_xl(  0, q8);
-            vector signed char q8y10 = vec_xl( 16, q8);
-            vector signed char q8y01 = vec_xl( 32, q8);
-            vector signed char q8y11 = vec_xl( 48, q8);
-            vector signed char q8y20 = vec_xl( 64, q8);
-            vector signed char q8y30 = vec_xl( 80, q8);
-            vector signed char q8y21 = vec_xl( 96, q8);
-            vector signed char q8y31 = vec_xl(112, q8);
-            q8 += 128;
-
-            vector signed int qv00 = vec_msum(q8y00, q4x00, v0);
-            vector signed int qv01 = vec_msum(q8y01, q4x01, v0);
-            vector signed int qv10 = vec_msum(q8y10, q4x10, v0);
-            vector signed int qv11 = vec_msum(q8y11, q4x11, v0);
-            vector signed int qv20 = vec_msum(q8y20, q4x20, v0);
-            vector signed int qv21 = vec_msum(q8y21, q4x21, v0);
-            vector signed int qv30 = vec_msum(q8y30, q4x30, v0);
-            vector signed int qv31 = vec_msum(q8y31, q4x31, v0);
-
-            vector signed int vscales_h = vec_unpackh(vscales);
-            vector signed int vs0 = vec_splat(vscales_h, 0);
-            vector signed int vs1 = vec_splat(vscales_h, 1);
-            vector signed int vs2 = vec_splat(vscales_h, 2);
-            vector signed int vs3 = vec_splat(vscales_h, 3);
-            vscales = vec_sld(vscales, vscales, 8);
-
-            vsumi0 = vec_add(vec_mul(qv00, vs0), vsumi0);
-            vsumi1 = vec_add(vec_mul(qv01, vs1), vsumi1);
-            vsumi2 = vec_add(vec_mul(qv20, vs2), vsumi2);
-            vsumi3 = vec_add(vec_mul(qv21, vs3), vsumi3);
-
-            vsumi0 = vec_add(vec_mul(qv10, vs0), vsumi0);
-            vsumi1 = vec_add(vec_mul(qv11, vs1), vsumi1);
-            vsumi2 = vec_add(vec_mul(qv30, vs2), vsumi2);
-            vsumi3 = vec_add(vec_mul(qv31, vs3), vsumi3);
-        }
-
-        vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
-        vsumf1 = vec_madd(vec_ctf(vsumi1, 0), vd, vsumf1);
-        vsumf2 = vec_madd(vec_ctf(vsumi2, 0), vd, vsumf2);
-        vsumf3 = vec_madd(vec_ctf(vsumi3, 0), vd, vsumf3);
-    }
-
-    vsumf0 = vec_add(vsumf0, vsumf2);
-    vsumf1 = vec_add(vsumf1, vsumf3);
-
-    vsumf0 = vec_add(vsumf0, vsumf1);
-
-    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4));
-    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8));
-
-    *s = vec_extract(vsumf0, 0);
-
-#elif defined __loongarch_asx
-
-    __m256 acc = (__m256)__lasx_xvldi(0);
-    __m128 acc_m = (__m128)__lsx_vldi(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);
-
-        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 uint8_t * GGML_RESTRICT q4 = x[i].qs;
-        const int8_t  * GGML_RESTRICT q8 = y[i].qs;
-
-        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(mins128, q8s);
-        acc_m = __lsx_vfmadd_s(__lsx_vreplfr2vr_s(dmin), __lsx_vffint_s_w(prod), acc_m);
-
-        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_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_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_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_madd_h_b(q4h, q8h);
-            p16h = lasx_madd_h(scale_h, p16h);
-            const __m256i sumj = __lasx_xvadd_w(p16l, p16h);
-
-            sumi = __lasx_xvadd_w(sumi, sumj);
-        }
-
-        __m256 vd = __lasx_xvreplfr2vr_s(d);
-        acc = __lasx_xvfmadd_s(vd, __lasx_xvffint_s_w(sumi), acc);
-
-    }
-
-    acc_m = __lsx_vfadd_s(acc_m, (__m128)__lsx_vpermi_w((__m128i)acc_m, (__m128i)acc_m, 0xee));
-    __m128i tmp1 = __lsx_vinsgr2vr_w(__lsx_vldi(0), __lsx_vpickve2gr_w((__m128i)acc_m, 1), 0);
-    acc_m = __lsx_vfadd_s(acc_m, (__m128)tmp1);
-
-
-    *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];
-    const uint8_t * mins   = (const uint8_t*)&utmp[2];
-
-    int8_t  aux8[QK_K];
-    int16_t aux16[8];
-    float   sums [8];
-    int32_t aux32[8];
-    memset(sums, 0, 8*sizeof(float));
-
-    float sumf = 0;
-    for (int i = 0; i < nb; ++i) {
-        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 * 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;
-            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l]  >> 4);
-            a += 32; q4 += 32;
-        }
-        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;
-
-        int sumi = 0;
-        for (int j = 0; j < QK_K/16; ++j) sumi += y[i].bsums[j] * mins[j/2];
-        a = aux8;
-        int is = 0;
-        for (int j = 0; j < QK_K/32; ++j) {
-            int32_t scale = scales[is++];
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-        }
-        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];
-        const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d;
-        sumf -= dmin * sumi;
-    }
-    for (int l = 0; l < 8; ++l) sumf += sums[l];
-    *s = sumf;
-#endif
-}
-
-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);
-    UNUSED(bx);
-    UNUSED(by);
-    UNUSED(bs);
-
-    const block_q5_K * GGML_RESTRICT x = vx;
-    const block_q8_K * GGML_RESTRICT y = vy;
-
-    const int nb = n / QK_K;
-
-    static const uint32_t kmask1 = 0x3f3f3f3f;
-    static const uint32_t kmask2 = 0x0f0f0f0f;
-    static const uint32_t kmask3 = 0x03030303;
-
-    uint32_t utmp[4];
-
-#ifdef __ARM_NEON
-    const uint8x16_t m4b = vdupq_n_u8(0xf);
-    const uint8x16_t mone = vdupq_n_u8(1);
-    const uint8x16_t mtwo = vdupq_n_u8(2);
-    const int32x4_t mzero = vdupq_n_s32(0);
-
-    ggml_int8x16x4_t q5bytes;
-
-    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 q8sums = vpaddq_s16(vld1q_s16(y[i].bsums), vld1q_s16(y[i].bsums + 8));
-
-        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 uint8x8_t mins8 = vld1_u8((const uint8_t*)utmp + 8);
-        const int16x8_t mins = vreinterpretq_s16_u16(vmovl_u8(mins8));
-        const int32x4_t prod = vaddq_s32(vmull_s16(vget_low_s16 (q8sums), vget_low_s16 (mins)),
-                                         vmull_s16(vget_high_s16(q8sums), vget_high_s16(mins)));
-        int32_t sumi_mins = vaddvq_s32(prod);
-
-        const uint8_t * scales = (const uint8_t *)utmp;
-
-        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);
-
-        ggml_uint8x16x4_t q5h;
-
-        int32_t sumi = 0;
-
-        for (int j = 0; j < QK_K/64; ++j) {
-
-            const ggml_uint8x16x2_t q5bits = ggml_vld1q_u8_x2(q5); q5 += 32;
-            const ggml_int8x16x4_t q8bytes = ggml_vld1q_s8_x4(q8); q8 += 64;
-
-            q5h.val[0] = vshlq_n_u8(vandq_u8(mone, qhbits.val[0]), 4);
-            q5h.val[1] = vshlq_n_u8(vandq_u8(mone, qhbits.val[1]), 4);
-            q5h.val[2] = vshlq_n_u8(vandq_u8(mtwo, qhbits.val[0]), 3);
-            q5h.val[3] = vshlq_n_u8(vandq_u8(mtwo, qhbits.val[1]), 3);
-            qhbits.val[0] = vshrq_n_u8(qhbits.val[0], 2);
-            qhbits.val[1] = vshrq_n_u8(qhbits.val[1], 2);
-
-            q5bytes.val[0] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(q5bits.val[0], m4b), q5h.val[0]));
-            q5bytes.val[1] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(q5bits.val[1], m4b), q5h.val[1]));
-            q5bytes.val[2] = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q5bits.val[0], 4), q5h.val[2]));
-            q5bytes.val[3] = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q5bits.val[1], 4), q5h.val[3]));
-
-            sumi += vaddvq_s32(ggml_vdotq_s32(ggml_vdotq_s32(mzero, q5bytes.val[0], q8bytes.val[0]), q5bytes.val[1], q8bytes.val[1])) * *scales++;
-            sumi += vaddvq_s32(ggml_vdotq_s32(ggml_vdotq_s32(mzero, q5bytes.val[2], q8bytes.val[2]), q5bytes.val[3], q8bytes.val[3])) * *scales++;
-        }
-
-        sumf += d * sumi - dmin * sumi_mins;
-    }
-
-    *s = sumf;
-
-#elif defined __AVX2__
-
-    const __m256i m4 = _mm256_set1_epi8(0xF);
-    const __m128i mzero = _mm_setzero_si128();
-    const __m256i mone  = _mm256_set1_epi8(1);
-
-    __m256 acc = _mm256_setzero_ps();
-
-    float summs = 0.f;
-
-    for (int i = 0; i < nb; ++i) {
-        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);
-
-        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 __m256i mins_and_scales = _mm256_cvtepu8_epi16(_mm_set_epi32(utmp[3], utmp[2], utmp[1], utmp[0]));
-
-        const __m256i q8sums = _mm256_loadu_si256((const __m256i*)y[i].bsums);
-        const __m128i q8s = _mm_hadd_epi16(_mm256_extracti128_si256(q8sums, 0), _mm256_extracti128_si256(q8sums, 1));
-        const __m128i prod = _mm_madd_epi16(_mm256_extracti128_si256(mins_and_scales, 1), q8s);
-        const __m128i hsum = _mm_hadd_epi32(_mm_hadd_epi32(prod, mzero), mzero);
-        summs += dmin * _mm_extract_epi32(hsum, 0);
-
-        const __m128i sc128  = _mm256_extracti128_si256(mins_and_scales, 0);
-        const __m256i scales = MM256_SET_M128I(sc128, sc128);
-
-        const __m256i hbits = _mm256_loadu_si256((const __m256i*)x[i].qh);
-        __m256i hmask = mone;
-
-        __m256i sumi = _mm256_setzero_si256();
-
-        int bit = 0;
-
-        for (int j = 0; j < QK_K/64; ++j) {
-
-            const __m256i scale_0 = _mm256_shuffle_epi8(scales, get_scale_shuffle_k4(2*j+0));
-            const __m256i scale_1 = _mm256_shuffle_epi8(scales, get_scale_shuffle_k4(2*j+1));
-
-            const __m256i q5bits = _mm256_loadu_si256((const __m256i*)q5); q5 += 32;
-
-            const __m256i q5l_0 = _mm256_and_si256(q5bits, m4);
-            const __m256i q5h_0 = _mm256_slli_epi16(_mm256_srli_epi16(_mm256_and_si256(hbits, hmask), bit++), 4);
-            const __m256i q5_0  = _mm256_add_epi8(q5l_0, q5h_0);
-            hmask = _mm256_slli_epi16(hmask, 1);
-
-            const __m256i q5l_1 = _mm256_and_si256(_mm256_srli_epi16(q5bits, 4), m4);
-            const __m256i q5h_1 = _mm256_slli_epi16(_mm256_srli_epi16(_mm256_and_si256(hbits, hmask), bit++), 4);
-            const __m256i q5_1  = _mm256_add_epi8(q5l_1, q5h_1);
-            hmask = _mm256_slli_epi16(hmask, 1);
-
-            const __m256i q8_0 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
-            const __m256i q8_1 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
-
-            __m256i p16_0 = _mm256_maddubs_epi16(q5_0, q8_0);
-            __m256i p16_1 = _mm256_maddubs_epi16(q5_1, q8_1);
-
-            p16_0 = _mm256_madd_epi16(scale_0, p16_0);
-            p16_1 = _mm256_madd_epi16(scale_1, p16_1);
-
-            sumi = _mm256_add_epi32(sumi, _mm256_add_epi32(p16_0, p16_1));
-
-        }
-
-        __m256 vd = _mm256_set1_ps(d);
-        acc = _mm256_fmadd_ps(vd, _mm256_cvtepi32_ps(sumi), acc);
-
-    }
-
-    *s = hsum_float_8(acc) + summs;
-
-#elif defined __AVX__
-
-    const __m128i m4 = _mm_set1_epi8(0xF);
-    const __m128i mzero = _mm_setzero_si128();
-    const __m128i mone  = _mm_set1_epi8(1);
-    const __m128i m2 = _mm_set1_epi8(2);
-
-    __m256 acc = _mm256_setzero_ps();
-
-    float summs = 0.f;
-
-    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 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);
-        const uint32_t uaux = utmp[1] & kmask1;
-        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
-        utmp[2] = uaux;
-        utmp[0] &= kmask1;
-
-        const __m128i utmps = _mm_set_epi32(utmp[3], utmp[2], utmp[1], utmp[0]);
-        const __m128i scales = _mm_cvtepu8_epi16(utmps);
-        const __m128i mins = _mm_cvtepu8_epi16(_mm_unpackhi_epi64(utmps, utmps));
-
-        const __m128i q8sums_0 = _mm_loadu_si128((const __m128i*)&y[i].bsums[0]);
-        const __m128i q8sums_1 = _mm_loadu_si128((const __m128i*)&y[i].bsums[8]);
-        const __m128i q8s = _mm_hadd_epi16(q8sums_0, q8sums_1);
-        const __m128i prod = _mm_madd_epi16(mins, q8s);
-        const __m128i hsum = _mm_hadd_epi32(_mm_hadd_epi32(prod, mzero), mzero);
-        summs += dmin * _mm_extract_epi32(hsum, 0);
-
-        const __m128i hbits_0 = _mm_loadu_si128((const __m128i*)&x[i].qh[0]);
-        const __m128i hbits_1 = _mm_loadu_si128((const __m128i*)&x[i].qh[16]);
-        __m128i hmask = mone;
-
-        __m128i sumi_0 = _mm_setzero_si128();
-        __m128i sumi_1 = _mm_setzero_si128();
-
-        int bit = 0;
-
-        __m128i shuffle = _mm_set1_epi16(0x0100);
-        for (int j = 0; j < QK_K/64; ++j) {
-
-            const __m128i scale_0 = _mm_shuffle_epi8(scales, shuffle);
-            shuffle = _mm_add_epi16(shuffle, m2);
-            const __m128i scale_1 = _mm_shuffle_epi8(scales, shuffle);
-            shuffle = _mm_add_epi16(shuffle, m2);
-
-            const __m128i q5bits_0 = _mm_loadu_si128((const __m128i*)q5); q5 += 16;
-            const __m128i q5bits_1 = _mm_loadu_si128((const __m128i*)q5); q5 += 16;
-
-            __m128i q5l_0 = _mm_and_si128(q5bits_0, m4);
-            __m128i q5l_1 = _mm_and_si128(q5bits_1, m4);
-            __m128i q5h_0 = _mm_slli_epi16(_mm_srli_epi16(_mm_and_si128(hbits_0, hmask), bit), 4);
-            __m128i q5h_1 = _mm_slli_epi16(_mm_srli_epi16(_mm_and_si128(hbits_1, hmask), bit++), 4);
-            __m128i q5_0  = _mm_add_epi8(q5l_0, q5h_0);
-            __m128i q5_1  = _mm_add_epi8(q5l_1, q5h_1);
-            hmask = _mm_slli_epi16(hmask, 1);
-
-            __m128i q8_0 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
-            __m128i q8_1 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
-            __m128i p16_0 = _mm_maddubs_epi16(q5_0, q8_0);
-            __m128i p16_1 = _mm_maddubs_epi16(q5_1, q8_1);
-            p16_0 = _mm_madd_epi16(scale_0, p16_0);
-            p16_1 = _mm_madd_epi16(scale_0, p16_1);
-
-            q5l_0 = _mm_and_si128(_mm_srli_epi16(q5bits_0, 4), m4);
-            q5l_1 = _mm_and_si128(_mm_srli_epi16(q5bits_1, 4), m4);
-            q5h_0 = _mm_slli_epi16(_mm_srli_epi16(_mm_and_si128(hbits_0, hmask), bit), 4);
-            q5h_1 = _mm_slli_epi16(_mm_srli_epi16(_mm_and_si128(hbits_1, hmask), bit++), 4);
-            q5_0  = _mm_add_epi8(q5l_0, q5h_0);
-            q5_1  = _mm_add_epi8(q5l_1, q5h_1);
-            hmask = _mm_slli_epi16(hmask, 1);
-
-            q8_0 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
-            q8_1 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
-            __m128i p16_2 = _mm_maddubs_epi16(q5_0, q8_0);
-            __m128i p16_3 = _mm_maddubs_epi16(q5_1, q8_1);
-            p16_2 = _mm_madd_epi16(scale_1, p16_2);
-            p16_3 = _mm_madd_epi16(scale_1, p16_3);
-
-            sumi_0 = _mm_add_epi32(sumi_0, _mm_add_epi32(p16_0, p16_2));
-            sumi_1 = _mm_add_epi32(sumi_1, _mm_add_epi32(p16_1, p16_3));
-
-        }
-
-        __m256 vd = _mm256_set1_ps(d);
-        __m256i sumi = MM256_SET_M128I(sumi_1, sumi_0);
-        acc = _mm256_add_ps(_mm256_mul_ps(vd, _mm256_cvtepi32_ps(sumi)), acc);
-
-    }
-
-    *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
-
-    const uint8_t * scales = (const uint8_t*)&utmp[0];
-    const uint8_t * mins   = (const uint8_t*)&utmp[2];
-
-    float sumf = 0;
-    float sums = 0.0;
-
-    size_t vl;
-
-    for (int i = 0; i < nb; ++i) {
-
-        vl = 8;
-
-        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;
-
-        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);
-        const uint32_t uaux = utmp[1] & kmask1;
-        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
-        utmp[2] = uaux;
-        utmp[0] &= kmask1;
-
-        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_i32m2_i32m1(prod, __riscv_vmv_v_x_i32m1(0, 1), vl);
-        sumf -= dmin * __riscv_vmv_x_s_i32m1_i32(sumi);
-
-        vl = 32;
-        int32_t aux32 = 0;
-        int is = 0;
-
-        uint8_t m = 1;
-        vint32m1_t vzero = __riscv_vmv_v_x_i32m1(0, 1);
-        vuint8m2_t vqh = __riscv_vle8_v_u8m2(hm, vl);
-
-        for (int j = 0; j < QK_K/64; ++j) {
-            // load Q5 and Q8
-            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
-            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;
-
-            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;
-
-            vint16m4_t v0 = __riscv_vwmul_vv_i16m4(q5_m1, q8_y1, vl);
-            vint16m4_t v1 = __riscv_vwmul_vv_i16m4(q5_m2, q8_y2, 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_i32m8_i32m1(vs1, vzero, vl);
-            vint32m1_t vacc2 = __riscv_vredsum_vs_i32m8_i32m1(vs2, vacc1, vl);
-
-            aux32 += __riscv_vmv_x_s_i32m1_i32(vacc2);
-            q5 += 32;    q8 += 64;
-
-        }
-
-        sums += aux32 * d;
-
-    }
-
-    *s = sumf+sums;
-
-#elif defined(__POWER9_VECTOR__)
-    const vector signed char lowMask = vec_splats((signed char)0xF);
-    const vector signed char lowMask1 = vec_splats((int8_t)0x3f);
-    const vector signed char lowMask2 = vec_splats((int8_t)0x30);
-    const vector int v0 = vec_splats((int32_t)0);
-    const vector unsigned char v1 = vec_splats((unsigned char)0x1);
-    const vector unsigned char v2 = vec_splats((unsigned char)0x2);
-    const vector unsigned char v3 = vec_splats((unsigned char)0x3);
-    const vector unsigned char v4 = vec_splats((unsigned char)0x4);
-
-    vector float vsumf0 = vec_splats(0.0f);
-    vector float vsumf1 = vec_splats(0.0f);
-    vector float vsumf2 = vec_splats(0.0f);
-    vector float vsumf3 = vec_splats(0.0f);
-
-    for (int i = 0; i < nb; ++i) {
-        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d));
-        vector float vyd = vec_splats(y[i].d);
-        vector float vd = vec_mul(vxd, vyd);
-
-        vector float vxmin = vec_splats(GGML_FP16_TO_FP32(x[i].dmin));
-        vector float vdmin = vec_mul(vxmin, vyd);
-
-        UNUSED(kmask1);
-        UNUSED(kmask2);
-        UNUSED(kmask3);
-        UNUSED(utmp);
-
-        vector signed char u0 = (vector signed char)vec_xl_len(x[i].scales, 8);
-        vector signed char u1 = vec_and(vec_sr(u0, v2), lowMask2);
-        vector signed char u2 = (vector signed char)vec_xl_len(x[i].scales + 8, 4);
-        vector signed char u3 = vec_sr(u2, v4);
-
-        vector signed char u30 = u1;
-        vector signed char u31 = (vector signed char)vec_mergeh((vector signed int)vec_and(u2, lowMask), (vector signed int)u3);
-
-        u1 = vec_and(u0, lowMask1);
-        u2 = vec_or(u30, u31);
-
-        vector signed char utmps = (vector signed char)vec_mergeh((vector signed int)u1, (vector signed int)u2);
-
-        vector signed short q8ysums0 = vec_xl( 0, y[i].bsums);
-        vector signed short q8ysums1 = vec_xl(16, y[i].bsums);
-
-        vector signed short vscales = vec_unpackh(utmps);
-
-        vector signed short q5xmins = vec_unpackl(utmps);
-        vector signed short q5xmins0 = vec_mergeh(q5xmins, q5xmins);
-        vector signed short q5xmins1 = vec_mergel(q5xmins, q5xmins);
-
-        vector signed int prod0 = vec_mule(q5xmins0, q8ysums0);
-        vector signed int prod1 = vec_mule(q5xmins1, q8ysums1);
-        vector signed int prod2 = vec_mulo(q5xmins0, q8ysums0);
-        vector signed int prod3 = vec_mulo(q5xmins1, q8ysums1);
-
-        vsumf0 = vec_nmsub(vec_ctf(prod0, 0), vdmin, vsumf0);
-        vsumf1 = vec_nmsub(vec_ctf(prod1, 0), vdmin, vsumf1);
-        vsumf2 = vec_nmsub(vec_ctf(prod2, 0), vdmin, vsumf2);
-        vsumf3 = vec_nmsub(vec_ctf(prod3, 0), vdmin, vsumf3);
-
-        vector signed char qxhs0 = (vector signed char)vec_xl( 0, x[i].qh);
-        vector signed char qxhs1 = (vector signed char)vec_xl(16, x[i].qh);
-
-        vector signed int vsumi0 = v0;
-        vector signed int vsumi1 = v0;
-        vector signed int vsumi2 = v0;
-        vector signed int vsumi3 = v0;
-
-        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);
-            __builtin_prefetch(q8, 0, 1);
-
-            vector signed char qxs0 = (vector signed char)vec_xl( 0, q5);
-            vector signed char qxs1 = (vector signed char)vec_xl(16, q5);
-            q5 += 32;
-
-            vector signed char qxs00 = vec_and(qxs0, lowMask);
-            vector signed char qxs01 = vec_sr(qxs0, v4);
-            vector signed char qxs10 = vec_and(qxs1, lowMask);
-            vector signed char qxs11 = vec_sr(qxs1, v4);
-
-            vector signed char q5h00 = vec_sl(vec_and((vector signed char)v1, qxhs0), v4);
-            vector signed char q5h01 = vec_sl(vec_and((vector signed char)v2, qxhs0), v3);
-            vector signed char q5h10 = vec_sl(vec_and((vector signed char)v1, qxhs1), v4);
-            vector signed char q5h11 = vec_sl(vec_and((vector signed char)v2, qxhs1), v3);
-            qxhs0 = vec_sr(qxhs0, v2);
-            qxhs1 = vec_sr(qxhs1, v2);
-
-            vector unsigned char q5x00 = (vector unsigned char)vec_or(q5h00, qxs00);
-            vector unsigned char q5x01 = (vector unsigned char)vec_or(q5h01, qxs01);
-            vector unsigned char q5x10 = (vector unsigned char)vec_or(q5h10, qxs10);
-            vector unsigned char q5x11 = (vector unsigned char)vec_or(q5h11, qxs11);
-
-            vector signed char q8y00 = vec_xl( 0, q8);
-            vector signed char q8y10 = vec_xl(16, q8);
-            vector signed char q8y01 = vec_xl(32, q8);
-            vector signed char q8y11 = vec_xl(48, q8);
-            q8 += 64;
-
-            vector signed int qv00 = vec_msum(q8y00, q5x00, v0);
-            vector signed int qv01 = vec_msum(q8y01, q5x01, v0);
-            vector signed int qv10 = vec_msum(q8y10, q5x10, v0);
-            vector signed int qv11 = vec_msum(q8y11, q5x11, v0);
-
-            vector signed int vscales_h = vec_unpackh(vscales);
-            vector signed int vs0 = vec_splat(vscales_h, 0);
-            vector signed int vs1 = vec_splat(vscales_h, 1);
-            vscales = vec_sld(vscales, vscales, 12);
-
-            vsumi0 = vec_add(vec_mul(qv00, vs0), vsumi0);
-            vsumi1 = vec_add(vec_mul(qv10, vs0), vsumi1);
-            vsumi2 = vec_add(vec_mul(qv01, vs1), vsumi2);
-            vsumi3 = vec_add(vec_mul(qv11, vs1), vsumi3);
-        }
-
-        vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
-        vsumf1 = vec_madd(vec_ctf(vsumi1, 0), vd, vsumf1);
-        vsumf2 = vec_madd(vec_ctf(vsumi2, 0), vd, vsumf2);
-        vsumf3 = vec_madd(vec_ctf(vsumi3, 0), vd, vsumf3);
-    }
-
-    vsumf0 = vec_add(vsumf0, vsumf2);
-    vsumf1 = vec_add(vsumf1, vsumf3);
-
-    vsumf0 = vec_add(vsumf0, vsumf1);
-
-    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4));
-    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8));
-
-    *s = vec_extract(vsumf0, 0);
-
-#elif defined __loongarch_asx
-
-    __m256 acc = (__m256)__lasx_xvldi(0);
-    __m128 acc_m = (__m128)__lsx_vldi(0);
-
-    for (int i = 0; i < nb; ++i) {
-
-        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);
-
-        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 __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(mins128, q8s);
-        acc_m = __lsx_vfmadd_s(__lsx_vreplfr2vr_s(dmin), __lsx_vffint_s_w(prod), acc_m);
-
-        const __m256i scales = lasx_insertf128(scales128, scales128);
-
-        const __m256i hbits = __lasx_xvld((const __m256i*)x[i].qh, 0);
-
-        __m256i sumi = __lasx_xvldi(0);
-
-        for (int j = 0; j < QK_K/64; ++j) {
-
-            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;
-
-            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_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);
-
-            sumi = __lasx_xvadd_w(sumi, __lasx_xvadd_w(p16_0, p16_1));
-
-        }
-
-        __m256 vd = __lasx_xvreplfr2vr_s(d);
-        acc = __lasx_xvfmadd_s(vd, __lasx_xvffint_s_w(sumi), acc);
-
-    }
-
-    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];
-    const uint8_t * mins   = (const uint8_t*)&utmp[2];
-
-    int8_t  aux8[QK_K];
-    int16_t aux16[8];
-    float   sums [8];
-    int32_t aux32[8];
-    memset(sums, 0, 8*sizeof(float));
-
-    float sumf = 0;
-    for (int i = 0; i < nb; ++i) {
-        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 * 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);
-            for (int l = 0; l < 32; ++l) a[l] += (hm[l] & m ? 16 : 0);
-            a += 32; m <<= 1;
-            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l]  >> 4);
-            for (int l = 0; l < 32; ++l) a[l] += (hm[l] & m ? 16 : 0);
-            a += 32; m <<= 1;
-            q4 += 32;
-        }
-        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;
-
-        int sumi = 0;
-        for (int j = 0; j < QK_K/16; ++j) sumi += y[i].bsums[j] * mins[j/2];
-        a = aux8;
-        int is = 0;
-        for (int j = 0; j < QK_K/32; ++j) {
-            int32_t scale = scales[is++];
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-        }
-        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];
-        const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d;
-        sumf -= dmin * sumi;
-    }
-    for (int l = 0; l < 8; ++l) sumf += sums[l];
-    *s = sumf;
-#endif
-}
-
-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);
-#ifdef __ARM_FEATURE_MATMUL_INT8
-    assert((nrc == 2) || (nrc == 1));
-#else
-    assert(nrc == 1);
-#endif
-    UNUSED(nrc);
-    UNUSED(bx);
-    UNUSED(by);
-    UNUSED(bs);
-
-    const block_q6_K * GGML_RESTRICT x = vx;
-    const block_q8_K * GGML_RESTRICT y = vy;
-
-    const int nb = n / QK_K;
-
-#if defined(__ARM_FEATURE_MATMUL_INT8)
-    if (nrc == 2) {
-        const block_q6_K * GGML_RESTRICT x0 = x;
-        const block_q6_K * GGML_RESTRICT x1 = (const block_q6_K *) ((const uint8_t *)vx + bx);
-        const block_q8_K * GGML_RESTRICT y0 = y;
-        const block_q8_K * GGML_RESTRICT y1 = (const block_q8_K *) ((const uint8_t *)vy + by);
-
-        float32x4_t vfsum = vdupq_n_f32(0.0f);
-
-        for (int i = 0; i < nb; ++i, ++x0, ++x1, ++y0, ++y1) {
-            const uint8_t * GGML_RESTRICT ql0 = x0->ql;
-            const uint8_t * GGML_RESTRICT ql1 = x1->ql;
-            const uint8_t * GGML_RESTRICT qh0 = x0->qh;
-            const uint8_t * GGML_RESTRICT qh1 = x1->qh;
-            const  int8_t * GGML_RESTRICT qy0 = y0->qs;
-            const  int8_t * GGML_RESTRICT qy1 = y1->qs;
-
-            const uint8x16_t mone = vdupq_n_u8(0x30);
-            const uint8x16_t  m4b = vdupq_n_u8(0x0f);
-
-            int32x4_t visum = vdupq_n_s32(0);
-
-            // process 8 blocks per iteration, totally 16 blocks
-            for (int j = 0; j < 2; ++j, qh0 += 32, ql0 += 64, qh1 += 32, ql1 += 64) {
-                int8x16_t vx0[8], vx1[8];
-
-                // de-quantize vx0[8]
-                {
-                    const uint8x16x2_t qh_bits = vld1q_u8_x2(qh0);
-                    const uint8x16x4_t ql_bits = vld1q_u8_x4(ql0);
-
-                    uint8x16_t q6h_0 = vandq_u8(mone, vshlq_n_u8(qh_bits.val[0], 4));
-                    uint8x16_t q6h_1 = vandq_u8(mone, vshlq_n_u8(qh_bits.val[1], 4));
-                    uint8x16_t q6h_2 = vandq_u8(mone, vshlq_n_u8(qh_bits.val[0], 2));
-                    uint8x16_t q6h_3 = vandq_u8(mone, vshlq_n_u8(qh_bits.val[1], 2));
-
-                    vx0[0] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(ql_bits.val[0], m4b), q6h_0));
-                    vx0[1] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(ql_bits.val[1], m4b), q6h_1));
-                    vx0[2] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(ql_bits.val[2], m4b), q6h_2));
-                    vx0[3] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(ql_bits.val[3], m4b), q6h_3));
-
-                    q6h_0 = vandq_u8(mone, qh_bits.val[0]);
-                    q6h_1 = vandq_u8(mone, qh_bits.val[1]);
-                    q6h_2 = vandq_u8(mone, vshrq_n_u8(qh_bits.val[0], 2));
-                    q6h_3 = vandq_u8(mone, vshrq_n_u8(qh_bits.val[1], 2));
-
-                    vx0[4] = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(ql_bits.val[0], 4), q6h_0));
-                    vx0[5] = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(ql_bits.val[1], 4), q6h_1));
-                    vx0[6] = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(ql_bits.val[2], 4), q6h_2));
-                    vx0[7] = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(ql_bits.val[3], 4), q6h_3));
-                }
-
-                // de-quantize vx1[8]
-                {
-                    const uint8x16x2_t qh_bits = vld1q_u8_x2(qh1);
-                    const uint8x16x4_t ql_bits = vld1q_u8_x4(ql1);
-
-                    uint8x16_t q6h_0 = vandq_u8(mone, vshlq_n_u8(qh_bits.val[0], 4));
-                    uint8x16_t q6h_1 = vandq_u8(mone, vshlq_n_u8(qh_bits.val[1], 4));
-                    uint8x16_t q6h_2 = vandq_u8(mone, vshlq_n_u8(qh_bits.val[0], 2));
-                    uint8x16_t q6h_3 = vandq_u8(mone, vshlq_n_u8(qh_bits.val[1], 2));
-
-                    vx1[0] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(ql_bits.val[0], m4b), q6h_0));
-                    vx1[1] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(ql_bits.val[1], m4b), q6h_1));
-                    vx1[2] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(ql_bits.val[2], m4b), q6h_2));
-                    vx1[3] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(ql_bits.val[3], m4b), q6h_3));
-
-                    q6h_0 = vandq_u8(mone, qh_bits.val[0]);
-                    q6h_1 = vandq_u8(mone, qh_bits.val[1]);
-                    q6h_2 = vandq_u8(mone, vshrq_n_u8(qh_bits.val[0], 2));
-                    q6h_3 = vandq_u8(mone, vshrq_n_u8(qh_bits.val[1], 2));
-
-                    vx1[4] = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(ql_bits.val[0], 4), q6h_0));
-                    vx1[5] = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(ql_bits.val[1], 4), q6h_1));
-                    vx1[6] = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(ql_bits.val[2], 4), q6h_2));
-                    vx1[7] = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(ql_bits.val[3], 4), q6h_3));
-                }
-
-                // process 16 elements (one block with same scale) per iteration
-                // - vx = concat(ql, qh) - 32
-                // - r1,r2,r3,r4 = smmla(vx, vy)
-                for (int k = 0; k < 8; ++k) {
-                    const int blk = j * 8 + k;
-
-                    const int8x16_t vy0 = vld1q_s8(qy0);
-                    const int8x16_t vy1 = vld1q_s8(qy1);
-                    qy0 += 16;
-                    qy1 += 16;
-
-                    const int32x4_t block_scale = {
-                        x0->scales[blk],
-                        x0->scales[blk],
-                        x1->scales[blk],
-                        x1->scales[blk],
-                    };
-
-                    // calculate four results at once with outer product
-                    const int8x16_t vx_l = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(vx0[k]), vreinterpretq_s64_s8(vx1[k])));
-                    const int8x16_t vx_h = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(vx0[k]), vreinterpretq_s64_s8(vx1[k])));
-                    const int8x16_t vy_l = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(vy0), vreinterpretq_s64_s8(vy1)));
-                    const int8x16_t vy_h = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(vy0), vreinterpretq_s64_s8(vy1)));
-                    int32x4_t vr = vdupq_n_s32(0);
-                    vr = vmmlaq_s32(vr, vx_l, vy_l);
-                    vr = vmmlaq_s32(vr, vx_h, vy_h);
-
-                    // apply block scale, will NOT overflow
-                    // block_scale * sum_256(int6*int8) <= 2^(8+8+6+8) = 30 bits
-                    visum = vmlaq_s32(visum, vr, block_scale);
-                }
-            }
-
-            // adjust bias, apply superblock scale
-            {
-                int32_t bias[4];
-#ifdef __ARM_FEATURE_SVE
-                const svbool_t pg16_8 = svptrue_pat_b16(SV_VL8);
-                const svbool_t pg8_8 = svptrue_pat_b8(SV_VL8);
-                const svint16_t y0_q8sums_0 = svld1_s16(pg16_8, y0->bsums);
-                const svint16_t y0_q8sums_1 = svld1_s16(pg16_8, y0->bsums + 8);
-                const svint16_t y1_q8sums_0 = svld1_s16(pg16_8, y1->bsums);
-                const svint16_t y1_q8sums_1 = svld1_s16(pg16_8, y1->bsums + 8);
-                const svint16_t x0_q6scales_0 = svunpklo_s16(svld1_s8(pg8_8, x0->scales));
-                const svint16_t x0_q6scales_1 = svunpklo_s16(svld1_s8(pg8_8, x0->scales + 8));
-                const svint16_t x1_q6scales_0 = svunpklo_s16(svld1_s8(pg8_8, x1->scales));
-                const svint16_t x1_q6scales_1 = svunpklo_s16(svld1_s8(pg8_8, x1->scales + 8));
-                const svint64_t zero = svdup_n_s64(0);
-                bias[0] = svaddv_s64(svptrue_b64(), svadd_s64_x(svptrue_b64(), svdot_s64(zero, y0_q8sums_0, x0_q6scales_0),
-                                                                               svdot_s64(zero, y0_q8sums_1, x0_q6scales_1)));
-                bias[1] = svaddv_s64(svptrue_b64(), svadd_s64_x(svptrue_b64(), svdot_s64(zero, y1_q8sums_0, x0_q6scales_0),
-                                                                               svdot_s64(zero, y1_q8sums_1, x0_q6scales_1)));
-                bias[2] = svaddv_s64(svptrue_b64(), svadd_s64_x(svptrue_b64(), svdot_s64(zero, y0_q8sums_0, x1_q6scales_0),
-                                                                               svdot_s64(zero, y0_q8sums_1, x1_q6scales_1)));
-                bias[3] = svaddv_s64(svptrue_b64(), svadd_s64_x(svptrue_b64(), svdot_s64(zero, y1_q8sums_0, x1_q6scales_0),
-                                                                               svdot_s64(zero, y1_q8sums_1, x1_q6scales_1)));
-#else
-                // NEON doesn't support int16 dot product, fallback to separated mul and add
-                const int16x8x2_t q8sums0 = vld1q_s16_x2(y0->bsums);
-                const int16x8x2_t q8sums1 = vld1q_s16_x2(y1->bsums);
-
-                int8x16_t scales_s8 = vld1q_s8(x0->scales);
-                const int16x8x2_t q6scales0 = {{vmovl_s8(vget_low_s8(scales_s8)), vmovl_s8(vget_high_s8(scales_s8))}};
-                scales_s8 = vld1q_s8(x1->scales);
-                const int16x8x2_t q6scales1 = {{vmovl_s8(vget_low_s8(scales_s8)), vmovl_s8(vget_high_s8(scales_s8))}};
-
-                int32x4_t prod;
-                prod = vaddq_s32(vaddq_s32(vmull_s16(vget_low_s16 (q8sums0.val[0]), vget_low_s16 (q6scales0.val[0])),
-                                           vmull_s16(vget_high_s16(q8sums0.val[0]), vget_high_s16(q6scales0.val[0]))),
-                                 vaddq_s32(vmull_s16(vget_low_s16 (q8sums0.val[1]), vget_low_s16 (q6scales0.val[1])),
-                                           vmull_s16(vget_high_s16(q8sums0.val[1]), vget_high_s16(q6scales0.val[1]))));
-                bias[0] = vaddvq_s32(prod);
-                prod = vaddq_s32(vaddq_s32(vmull_s16(vget_low_s16 (q8sums1.val[0]), vget_low_s16 (q6scales0.val[0])),
-                                           vmull_s16(vget_high_s16(q8sums1.val[0]), vget_high_s16(q6scales0.val[0]))),
-                                 vaddq_s32(vmull_s16(vget_low_s16 (q8sums1.val[1]), vget_low_s16 (q6scales0.val[1])),
-                                           vmull_s16(vget_high_s16(q8sums1.val[1]), vget_high_s16(q6scales0.val[1]))));
-                bias[1] = vaddvq_s32(prod);
-                prod = vaddq_s32(vaddq_s32(vmull_s16(vget_low_s16 (q8sums0.val[0]), vget_low_s16 (q6scales1.val[0])),
-                                           vmull_s16(vget_high_s16(q8sums0.val[0]), vget_high_s16(q6scales1.val[0]))),
-                                 vaddq_s32(vmull_s16(vget_low_s16 (q8sums0.val[1]), vget_low_s16 (q6scales1.val[1])),
-                                           vmull_s16(vget_high_s16(q8sums0.val[1]), vget_high_s16(q6scales1.val[1]))));
-                bias[2] = vaddvq_s32(prod);
-                prod = vaddq_s32(vaddq_s32(vmull_s16(vget_low_s16 (q8sums1.val[0]), vget_low_s16 (q6scales1.val[0])),
-                                           vmull_s16(vget_high_s16(q8sums1.val[0]), vget_high_s16(q6scales1.val[0]))),
-                                 vaddq_s32(vmull_s16(vget_low_s16 (q8sums1.val[1]), vget_low_s16 (q6scales1.val[1])),
-                                           vmull_s16(vget_high_s16(q8sums1.val[1]), vget_high_s16(q6scales1.val[1]))));
-                bias[3] = vaddvq_s32(prod);
-
-#endif
-                const int32x4_t vibias = vmulq_n_s32(vld1q_s32(bias), 32);
-
-                const float32x4_t superblock_scale = {
-                    GGML_FP16_TO_FP32(x0->d) * y0->d,
-                    GGML_FP16_TO_FP32(x0->d) * y1->d,
-                    GGML_FP16_TO_FP32(x1->d) * y0->d,
-                    GGML_FP16_TO_FP32(x1->d) * y1->d,
-                };
-
-                visum = vsubq_s32(visum, vibias);
-                vfsum = vmlaq_f32(vfsum, vcvtq_f32_s32(visum), superblock_scale);
-            }
-        }
-
-        // vfsum = ABCD -> ACBD
-        // AC -> s, BD -> (s+bs)
-        vfsum = vzip1q_f32(vfsum, vextq_f32(vfsum, vfsum, 2));
-        vst1_f32(s,      vget_low_f32 (vfsum));
-        vst1_f32(s + bs, vget_high_f32(vfsum));
-
-        return;
-    }
-#endif
-
-#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);
-    const int32x4_t  vzero = vdupq_n_s32(0);
-    //const int8x16_t  m32s = vdupq_n_s8(32);
-
-    const uint8x16_t mone = vdupq_n_u8(3);
-
-    ggml_int8x16x4_t q6bytes;
-    ggml_uint8x16x4_t q6h;
-
-    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 ggml_int16x8x2_t q8sums = ggml_vld1q_s16_x2(y[i].bsums);
-        const int8x16_t scales = vld1q_s8(scale);
-        const ggml_int16x8x2_t q6scales = {{vmovl_s8(vget_low_s8(scales)), vmovl_s8(vget_high_s8(scales))}};
-
-        const int32x4_t prod = vaddq_s32(vaddq_s32(vmull_s16(vget_low_s16 (q8sums.val[0]), vget_low_s16 (q6scales.val[0])),
-                                                   vmull_s16(vget_high_s16(q8sums.val[0]), vget_high_s16(q6scales.val[0]))),
-                                         vaddq_s32(vmull_s16(vget_low_s16 (q8sums.val[1]), vget_low_s16 (q6scales.val[1])),
-                                                   vmull_s16(vget_high_s16(q8sums.val[1]), vget_high_s16(q6scales.val[1]))));
-        int32_t isum_mins = vaddvq_s32(prod);
-
-        int32_t isum = 0;
-
-        for (int j = 0; j < QK_K/128; ++j) {
-
-            ggml_uint8x16x2_t qhbits = ggml_vld1q_u8_x2(qh); qh += 32;
-            ggml_uint8x16x4_t q6bits = ggml_vld1q_u8_x4(q6); q6 += 64;
-            ggml_int8x16x4_t q8bytes = ggml_vld1q_s8_x4(q8); q8 += 64;
-
-            q6h.val[0] = vshlq_n_u8(vandq_u8(mone, qhbits.val[0]), 4);
-            q6h.val[1] = vshlq_n_u8(vandq_u8(mone, qhbits.val[1]), 4);
-            uint8x16_t shifted = vshrq_n_u8(qhbits.val[0], 2);
-            q6h.val[2] = vshlq_n_u8(vandq_u8(mone, shifted), 4);
-            shifted = vshrq_n_u8(qhbits.val[1], 2);
-            q6h.val[3] = vshlq_n_u8(vandq_u8(mone, shifted), 4);
-
-            //q6bytes.val[0] = vsubq_s8(vreinterpretq_s8_u8(vorrq_u8(vandq_u8(q6bits.val[0], m4b), q6h.val[0])), m32s);
-            //q6bytes.val[1] = vsubq_s8(vreinterpretq_s8_u8(vorrq_u8(vandq_u8(q6bits.val[1], m4b), q6h.val[1])), m32s);
-            //q6bytes.val[2] = vsubq_s8(vreinterpretq_s8_u8(vorrq_u8(vandq_u8(q6bits.val[2], m4b), q6h.val[2])), m32s);
-            //q6bytes.val[3] = vsubq_s8(vreinterpretq_s8_u8(vorrq_u8(vandq_u8(q6bits.val[3], m4b), q6h.val[3])), m32s);
-            q6bytes.val[0] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(q6bits.val[0], m4b), q6h.val[0]));
-            q6bytes.val[1] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(q6bits.val[1], m4b), q6h.val[1]));
-            q6bytes.val[2] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(q6bits.val[2], m4b), q6h.val[2]));
-            q6bytes.val[3] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(q6bits.val[3], m4b), q6h.val[3]));
-
-            isum += vaddvq_s32(ggml_vdotq_s32(vzero, q6bytes.val[0], q8bytes.val[0])) * scale[0] +
-                    vaddvq_s32(ggml_vdotq_s32(vzero, q6bytes.val[1], q8bytes.val[1])) * scale[1] +
-                    vaddvq_s32(ggml_vdotq_s32(vzero, q6bytes.val[2], q8bytes.val[2])) * scale[2] +
-                    vaddvq_s32(ggml_vdotq_s32(vzero, q6bytes.val[3], q8bytes.val[3])) * scale[3];
-
-            scale += 4;
-
-            q8bytes = ggml_vld1q_s8_x4(q8); q8 += 64;
-
-            shifted = vshrq_n_u8(qhbits.val[0], 4);
-            q6h.val[0] = vshlq_n_u8(vandq_u8(mone, shifted), 4);
-            shifted = vshrq_n_u8(qhbits.val[1], 4);
-            q6h.val[1] = vshlq_n_u8(vandq_u8(mone, shifted), 4);
-            shifted = vshrq_n_u8(qhbits.val[0], 6);
-            q6h.val[2] = vshlq_n_u8(vandq_u8(mone, shifted), 4);
-            shifted = vshrq_n_u8(qhbits.val[1], 6);
-            q6h.val[3] = vshlq_n_u8(vandq_u8(mone, shifted), 4);
-
-            //q6bytes.val[0] = vsubq_s8(vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6bits.val[0], 4), q6h.val[0])), m32s);
-            //q6bytes.val[1] = vsubq_s8(vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6bits.val[1], 4), q6h.val[1])), m32s);
-            //q6bytes.val[2] = vsubq_s8(vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6bits.val[2], 4), q6h.val[2])), m32s);
-            //q6bytes.val[3] = vsubq_s8(vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6bits.val[3], 4), q6h.val[3])), m32s);
-            q6bytes.val[0] = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6bits.val[0], 4), q6h.val[0]));
-            q6bytes.val[1] = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6bits.val[1], 4), q6h.val[1]));
-            q6bytes.val[2] = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6bits.val[2], 4), q6h.val[2]));
-            q6bytes.val[3] = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6bits.val[3], 4), q6h.val[3]));
-
-            isum += vaddvq_s32(ggml_vdotq_s32(vzero, q6bytes.val[0], q8bytes.val[0])) * scale[0] +
-                    vaddvq_s32(ggml_vdotq_s32(vzero, q6bytes.val[1], q8bytes.val[1])) * scale[1] +
-                    vaddvq_s32(ggml_vdotq_s32(vzero, q6bytes.val[2], q8bytes.val[2])) * scale[2] +
-                    vaddvq_s32(ggml_vdotq_s32(vzero, q6bytes.val[3], q8bytes.val[3])) * scale[3];
-            scale += 4;
-        }
-        //sum += isum * d_all * y[i].d;
-        sum += d_all * y[i].d * (isum - 32 * isum_mins);
-
-    }
-    *s = sum;
-
-#elif defined __AVX2__
-
-    const __m256i m4 = _mm256_set1_epi8(0xF);
-    const __m256i m2 = _mm256_set1_epi8(3);
-    const __m256i m32s = _mm256_set1_epi8(32);
-
-    __m256 acc = _mm256_setzero_ps();
-
-    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 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);
-
-        __m256i sumi = _mm256_setzero_si256();
-
-        int is = 0;
-
-        for (int j = 0; j < QK_K/128; ++j) {
-
-            const __m128i scale_0 = _mm_shuffle_epi8(scales, get_scale_shuffle(is + 0));
-            const __m128i scale_1 = _mm_shuffle_epi8(scales, get_scale_shuffle(is + 1));
-            const __m128i scale_2 = _mm_shuffle_epi8(scales, get_scale_shuffle(is + 2));
-            const __m128i scale_3 = _mm_shuffle_epi8(scales, get_scale_shuffle(is + 3));
-            is += 4;
-
-            const __m256i q4bits1 = _mm256_loadu_si256((const __m256i*)q4); q4 += 32;
-            const __m256i q4bits2 = _mm256_loadu_si256((const __m256i*)q4); q4 += 32;
-            const __m256i q4bitsH = _mm256_loadu_si256((const __m256i*)qh); qh += 32;
-
-            const __m256i q4h_0 = _mm256_slli_epi16(_mm256_and_si256(q4bitsH, m2), 4);
-            const __m256i q4h_1 = _mm256_slli_epi16(_mm256_and_si256(_mm256_srli_epi16(q4bitsH, 2), m2), 4);
-            const __m256i q4h_2 = _mm256_slli_epi16(_mm256_and_si256(_mm256_srli_epi16(q4bitsH, 4), m2), 4);
-            const __m256i q4h_3 = _mm256_slli_epi16(_mm256_and_si256(_mm256_srli_epi16(q4bitsH, 6), m2), 4);
-
-            const __m256i q4_0 = _mm256_or_si256(_mm256_and_si256(q4bits1, m4), q4h_0);
-            const __m256i q4_1 = _mm256_or_si256(_mm256_and_si256(q4bits2, m4), q4h_1);
-            const __m256i q4_2 = _mm256_or_si256(_mm256_and_si256(_mm256_srli_epi16(q4bits1, 4), m4), q4h_2);
-            const __m256i q4_3 = _mm256_or_si256(_mm256_and_si256(_mm256_srli_epi16(q4bits2, 4), m4), q4h_3);
-
-            const __m256i q8_0 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
-            const __m256i q8_1 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
-            const __m256i q8_2 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
-            const __m256i q8_3 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
-
-            __m256i q8s_0 = _mm256_maddubs_epi16(m32s, q8_0);
-            __m256i q8s_1 = _mm256_maddubs_epi16(m32s, q8_1);
-            __m256i q8s_2 = _mm256_maddubs_epi16(m32s, q8_2);
-            __m256i q8s_3 = _mm256_maddubs_epi16(m32s, q8_3);
-
-            __m256i p16_0 = _mm256_maddubs_epi16(q4_0, q8_0);
-            __m256i p16_1 = _mm256_maddubs_epi16(q4_1, q8_1);
-            __m256i p16_2 = _mm256_maddubs_epi16(q4_2, q8_2);
-            __m256i p16_3 = _mm256_maddubs_epi16(q4_3, q8_3);
-
-            p16_0 = _mm256_sub_epi16(p16_0, q8s_0);
-            p16_1 = _mm256_sub_epi16(p16_1, q8s_1);
-            p16_2 = _mm256_sub_epi16(p16_2, q8s_2);
-            p16_3 = _mm256_sub_epi16(p16_3, q8s_3);
-
-            p16_0 = _mm256_madd_epi16(_mm256_cvtepi8_epi16(scale_0), p16_0);
-            p16_1 = _mm256_madd_epi16(_mm256_cvtepi8_epi16(scale_1), p16_1);
-            p16_2 = _mm256_madd_epi16(_mm256_cvtepi8_epi16(scale_2), p16_2);
-            p16_3 = _mm256_madd_epi16(_mm256_cvtepi8_epi16(scale_3), p16_3);
-
-            sumi = _mm256_add_epi32(sumi, _mm256_add_epi32(p16_0, p16_1));
-            sumi = _mm256_add_epi32(sumi, _mm256_add_epi32(p16_2, p16_3));
-
-        }
-
-        acc = _mm256_fmadd_ps(_mm256_broadcast_ss(&d), _mm256_cvtepi32_ps(sumi), acc);
-    }
-
-    *s = hsum_float_8(acc);
-
-#elif defined __AVX__
-
-    const __m128i m3 = _mm_set1_epi8(3);
-    const __m128i m15 = _mm_set1_epi8(15);
-
-    __m256 acc = _mm256_setzero_ps();
-
-    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 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);
-        const __m128i q8sums_1 = _mm_loadu_si128((const __m128i*)y[i].bsums + 1);
-        const __m128i scales = _mm_loadu_si128((const __m128i*)x[i].scales);
-        const __m128i scales_16_0 = _mm_cvtepi8_epi16(scales);
-        const __m128i scales_16_1 = _mm_cvtepi8_epi16(_mm_bsrli_si128(scales, 8));
-        const __m128i q8sclsub_0 = _mm_slli_epi32(_mm_madd_epi16(q8sums_0, scales_16_0), 5);
-        const __m128i q8sclsub_1 = _mm_slli_epi32(_mm_madd_epi16(q8sums_1, scales_16_1), 5);
-
-        __m128i sumi_0 = _mm_setzero_si128();
-        __m128i sumi_1 = _mm_setzero_si128();
-
-        int is = 0;
-
-        for (int j = 0; j < QK_K/128; ++j) {
-
-            const __m128i q4bitsH_0 = _mm_loadu_si128((const __m128i*)qh); qh += 16;
-            const __m128i q4bitsH_1 = _mm_loadu_si128((const __m128i*)qh); qh += 16;
-
-            const __m128i q4h_0 = _mm_slli_epi16(_mm_and_si128(q4bitsH_0, m3), 4);
-            const __m128i q4h_1 = _mm_slli_epi16(_mm_and_si128(q4bitsH_1, m3), 4);
-            const __m128i q4h_2 = _mm_slli_epi16(_mm_and_si128(q4bitsH_0, _mm_set1_epi8(12)), 2);
-            const __m128i q4h_3 = _mm_slli_epi16(_mm_and_si128(q4bitsH_1, _mm_set1_epi8(12)), 2);
-            const __m128i q4h_4 = _mm_and_si128(q4bitsH_0, _mm_set1_epi8(48));
-            const __m128i q4h_5 = _mm_and_si128(q4bitsH_1, _mm_set1_epi8(48));
-            const __m128i q4h_6 = _mm_srli_epi16(_mm_and_si128(q4bitsH_0, _mm_set1_epi8(-64)), 2);
-            const __m128i q4h_7 = _mm_srli_epi16(_mm_and_si128(q4bitsH_1, _mm_set1_epi8(-64)), 2);
-
-            const __m128i q4bits1_0 = _mm_loadu_si128((const __m128i*)q4); q4 += 16;
-            const __m128i q4bits1_1 = _mm_loadu_si128((const __m128i*)q4); q4 += 16;
-            const __m128i q4bits2_0 = _mm_loadu_si128((const __m128i*)q4); q4 += 16;
-            const __m128i q4bits2_1 = _mm_loadu_si128((const __m128i*)q4); q4 += 16;
-
-            const __m128i q4_0 = _mm_or_si128(_mm_and_si128(q4bits1_0, m15), q4h_0);
-            const __m128i q4_1 = _mm_or_si128(_mm_and_si128(q4bits1_1, m15), q4h_1);
-            const __m128i q4_2 = _mm_or_si128(_mm_and_si128(q4bits2_0, m15), q4h_2);
-            const __m128i q4_3 = _mm_or_si128(_mm_and_si128(q4bits2_1, m15), q4h_3);
-            const __m128i q4_4 = _mm_or_si128(_mm_and_si128(_mm_srli_epi16(q4bits1_0, 4), m15), q4h_4);
-            const __m128i q4_5 = _mm_or_si128(_mm_and_si128(_mm_srli_epi16(q4bits1_1, 4), m15), q4h_5);
-            const __m128i q4_6 = _mm_or_si128(_mm_and_si128(_mm_srli_epi16(q4bits2_0, 4), m15), q4h_6);
-            const __m128i q4_7 = _mm_or_si128(_mm_and_si128(_mm_srli_epi16(q4bits2_1, 4), m15), q4h_7);
-
-            const __m128i q8_0 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
-            const __m128i q8_1 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
-            const __m128i q8_2 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
-            const __m128i q8_3 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
-            const __m128i q8_4 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
-            const __m128i q8_5 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
-            const __m128i q8_6 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
-            const __m128i q8_7 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
-
-            __m128i p16_0 = _mm_maddubs_epi16(q4_0, q8_0);
-            __m128i p16_1 = _mm_maddubs_epi16(q4_1, q8_1);
-            __m128i p16_2 = _mm_maddubs_epi16(q4_2, q8_2);
-            __m128i p16_3 = _mm_maddubs_epi16(q4_3, q8_3);
-            __m128i p16_4 = _mm_maddubs_epi16(q4_4, q8_4);
-            __m128i p16_5 = _mm_maddubs_epi16(q4_5, q8_5);
-            __m128i p16_6 = _mm_maddubs_epi16(q4_6, q8_6);
-            __m128i p16_7 = _mm_maddubs_epi16(q4_7, q8_7);
-
-            const __m128i scale_0 = _mm_shuffle_epi8(scales, get_scale_shuffle(is + 0));
-            const __m128i scale_1 = _mm_shuffle_epi8(scales, get_scale_shuffle(is + 1));
-            const __m128i scale_2 = _mm_shuffle_epi8(scales, get_scale_shuffle(is + 2));
-            const __m128i scale_3 = _mm_shuffle_epi8(scales, get_scale_shuffle(is + 3));
-            is += 4;
-
-            p16_0 = _mm_madd_epi16(_mm_cvtepi8_epi16(scale_0), p16_0);
-            p16_1 = _mm_madd_epi16(_mm_cvtepi8_epi16(_mm_bsrli_si128(scale_0, 8)), p16_1);
-            p16_2 = _mm_madd_epi16(_mm_cvtepi8_epi16(scale_1), p16_2);
-            p16_3 = _mm_madd_epi16(_mm_cvtepi8_epi16(_mm_bsrli_si128(scale_1, 8)), p16_3);
-            p16_4 = _mm_madd_epi16(_mm_cvtepi8_epi16(scale_2), p16_4);
-            p16_5 = _mm_madd_epi16(_mm_cvtepi8_epi16(_mm_bsrli_si128(scale_2, 8)), p16_5);
-            p16_6 = _mm_madd_epi16(_mm_cvtepi8_epi16(scale_3), p16_6);
-            p16_7 = _mm_madd_epi16(_mm_cvtepi8_epi16(_mm_bsrli_si128(scale_3, 8)), p16_7);
-
-            sumi_0 = _mm_add_epi32(sumi_0, _mm_add_epi32(p16_0, p16_2));
-            sumi_1 = _mm_add_epi32(sumi_1, _mm_add_epi32(p16_1, p16_3));
-            sumi_0 = _mm_add_epi32(sumi_0, _mm_add_epi32(p16_4, p16_6));
-            sumi_1 = _mm_add_epi32(sumi_1, _mm_add_epi32(p16_5, p16_7));
-
-        }
-
-        sumi_0 = _mm_sub_epi32(sumi_0, q8sclsub_0);
-        sumi_1 = _mm_sub_epi32(sumi_1, q8sclsub_1);
-        const __m256i sumi = MM256_SET_M128I(sumi_1, sumi_0);
-        acc = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(sumi)), acc);
-    }
-
-    *s = hsum_float_8(acc);
-
-#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};
-
-    for (int i = 0; i < nb; ++i) {
-        // 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;
-        }
-
-        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_xtheadvector
-
-    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;
-
-        int sum_t = 0;
-        int t0;
-
-        for (int j = 0; j < QK_K/128; ++j) {
-            __asm__ __volatile__(
-                "th.vsetvli zero, %[vl32], e8, m2\n\t" // vl == 32
-                "th.vlb.v v4, (%[qh])\n\t"
-                "th.vsll.vi v0, v4, 4\n\t"
-                "th.vsll.vi v2, v4, 2\n\t"
-                "th.vsrl.vi v6, v4, 2\n\t"
-                "th.vsetvli zero, %[vl64], e8, m4\n\t" // vl == 64
-                "th.vlb.v v8, (%[q6])\n\t"
-                "th.vsrl.vi v12, v8, 4\n\t"
-                "th.vand.vi v8, v8, 0xF\n\t"
-                "th.vsetvli zero, %[vl128], e8, m8\n\t" // vl == 128
-                "th.vand.vx v0, v0, %[mask]\n\t"
-                "th.vor.vv v8, v8, v0\n\t"
-                "th.vlb.v v0, (%[q8])\n\t"
-                "th.vsub.vx v8, v8, %[vl32]\n\t"
-                "th.vsetvli zero, %[vl64], e8, m4\n\t" // vl == 64
-                "th.vwmul.vv v16, v0, v8\n\t"
-                "th.vwmul.vv v24, v4, v12\n\t"
-                "li %[t0], 16\n\t"
-                "th.vsetvli zero, %[t0], e16, m2\n\t" // vl == 16
-                "th.vmv.v.x v0, zero\n\t"
-                "th.vwredsum.vs v10, v16, v0\n\t"
-                "th.vwredsum.vs v9, v18, v0\n\t"
-                "th.vwredsum.vs v8, v20, v0\n\t"
-                "th.vwredsum.vs v7, v22, v0\n\t"
-                "th.vwredsum.vs v11, v24, v0\n\t"
-                "th.vwredsum.vs v12, v26, v0\n\t"
-                "th.vwredsum.vs v13, v28, v0\n\t"
-                "th.vwredsum.vs v14, v30, v0\n\t"
-                "li %[t0], 4\n\t"
-                "th.vsetvli zero, %[t0], e32, m1\n\t" // vl == 4
-                "th.vslideup.vi v10, v9, 1\n\t"
-                "th.vslideup.vi v8, v7, 1\n\t"
-                "th.vslideup.vi v11, v12, 1\n\t"
-                "th.vslideup.vi v13, v14, 1\n\t"
-                "th.vslideup.vi v10, v8, 2\n\t"
-                "th.vslideup.vi v11, v13, 2\n\t"
-                "li %[t0], 8\n\t"
-                "th.vsetvli zero, %[t0], e32, m2\n\t" // vl == 8
-                "th.vlb.v v4, (%[scale])\n\t"
-                "th.vmul.vv v2, v4, v10\n\t"
-                "th.vredsum.vs v0, v2, v0\n\t"
-                "th.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;
-
-    }
-
-    *s = sumf;
-
-#elif defined __riscv_v
-
-    float sumf = 0;
-    const int vector_length = __riscv_vlenb() * 8;
-
-    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;
-
-#elif defined(__POWER9_VECTOR__)
-    const vector signed char lowMask = vec_splats((signed char)0xF);
-    const vector int v0 = vec_splats((int32_t)0);
-    const vector unsigned char v2 = vec_splats((unsigned char)0x2);
-    const vector unsigned char v3 = vec_splats((unsigned char)0x3);
-    const vector unsigned char v4 = vec_splats((unsigned char)0x4);
-    const vector unsigned char v6 = vec_splats((unsigned char)0x6);
-    const vector signed char off = vec_splats((signed char)0x20);
-
-    vector float vsumf0 = vec_splats(0.0f);
-    vector float vsumf1 = vec_splats(0.0f);
-    vector float vsumf2 = vec_splats(0.0f);
-    vector float vsumf3 = vec_splats(0.0f);
-
-    for (int i = 0; i < nb; ++i) {
-        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d));
-        vector float vyd = vec_splats(y[i].d);
-        vector float vd = vec_mul(vxd, vyd);
-
-        vector signed int vsumi0 = v0;
-        vector signed int vsumi1 = v0;
-        vector signed int vsumi2 = v0;
-        vector signed int vsumi3 = v0;
-        vector signed int vsumi4 = v0;
-        vector signed int vsumi5 = v0;
-        vector signed int vsumi6 = v0;
-        vector signed int vsumi7 = v0;
-
-        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);
-            __builtin_prefetch(qh, 0, 0);
-            __builtin_prefetch(q8, 0, 0);
-
-            vector signed char qxs0 = (vector signed char)vec_xl( 0, q6);
-            vector signed char qxs1 = (vector signed char)vec_xl(16, q6);
-            vector signed char qxs2 = (vector signed char)vec_xl(32, q6);
-            vector signed char qxs3 = (vector signed char)vec_xl(48, q6);
-            q6 += 64;
-
-            vector signed char qxs00 = vec_and(qxs0, lowMask);
-            vector signed char qxs01 = vec_sr(qxs0, v4);
-            vector signed char qxs10 = vec_and(qxs1, lowMask);
-            vector signed char qxs11 = vec_sr(qxs1, v4);
-            vector signed char qxs20 = vec_and(qxs2, lowMask);
-            vector signed char qxs21 = vec_sr(qxs2, v4);
-            vector signed char qxs30 = vec_and(qxs3, lowMask);
-            vector signed char qxs31 = vec_sr(qxs3, v4);
-
-            vector signed char qxhs0 = (vector signed char)vec_xl( 0, qh);
-            vector signed char qxhs1 = (vector signed char)vec_xl(16, qh);
-            qh += 32;
-
-            vector signed char qxh00 = vec_sl(vec_and((vector signed char)v3, qxhs0), v4);
-            vector signed char qxh01 = vec_sl(vec_and((vector signed char)v3, vec_sr(qxhs0, v4)), v4);
-            vector signed char qxh10 = vec_sl(vec_and((vector signed char)v3, qxhs1), v4);
-            vector signed char qxh11 = vec_sl(vec_and((vector signed char)v3, vec_sr(qxhs1, v4)), v4);
-            vector signed char qxh20 = vec_sl(vec_and((vector signed char)v3, vec_sr(qxhs0, v2)), v4);
-            vector signed char qxh21 = vec_sl(vec_and((vector signed char)v3, vec_sr(qxhs0, v6)), v4);
-            vector signed char qxh30 = vec_sl(vec_and((vector signed char)v3, vec_sr(qxhs1, v2)), v4);
-            vector signed char qxh31 = vec_sl(vec_and((vector signed char)v3, vec_sr(qxhs1, v6)), v4);
-
-            vector signed char q6x00 = vec_sub(vec_or(qxh00, qxs00), off);
-            vector signed char q6x01 = vec_sub(vec_or(qxh01, qxs01), off);
-            vector signed char q6x10 = vec_sub(vec_or(qxh10, qxs10), off);
-            vector signed char q6x11 = vec_sub(vec_or(qxh11, qxs11), off);
-            vector signed char q6x20 = vec_sub(vec_or(qxh20, qxs20), off);
-            vector signed char q6x21 = vec_sub(vec_or(qxh21, qxs21), off);
-            vector signed char q6x30 = vec_sub(vec_or(qxh30, qxs30), off);
-            vector signed char q6x31 = vec_sub(vec_or(qxh31, qxs31), off);
-
-            vector signed char q8y00 = vec_xl(  0, q8);
-            vector signed char q8y10 = vec_xl( 16, q8);
-            vector signed char q8y20 = vec_xl( 32, q8);
-            vector signed char q8y30 = vec_xl( 48, q8);
-            vector signed char q8y01 = vec_xl( 64, q8);
-            vector signed char q8y11 = vec_xl( 80, q8);
-            vector signed char q8y21 = vec_xl( 96, q8);
-            vector signed char q8y31 = vec_xl(112, q8);
-            q8 += 128;
-
-            vector signed short qv00 = vec_add(vec_mule(q6x00, q8y00), vec_mulo(q6x00, q8y00));
-            vector signed short qv10 = vec_add(vec_mule(q6x10, q8y10), vec_mulo(q6x10, q8y10));
-            vector signed short qv20 = vec_add(vec_mule(q6x20, q8y20), vec_mulo(q6x20, q8y20));
-            vector signed short qv30 = vec_add(vec_mule(q6x30, q8y30), vec_mulo(q6x30, q8y30));
-            vector signed short qv01 = vec_add(vec_mule(q6x01, q8y01), vec_mulo(q6x01, q8y01));
-            vector signed short qv11 = vec_add(vec_mule(q6x11, q8y11), vec_mulo(q6x11, q8y11));
-            vector signed short qv21 = vec_add(vec_mule(q6x21, q8y21), vec_mulo(q6x21, q8y21));
-            vector signed short qv31 = vec_add(vec_mule(q6x31, q8y31), vec_mulo(q6x31, q8y31));
-
-            vector signed short vscales = vec_unpackh(vec_xl_len(qs, 8));
-            qs += 8;
-
-            vector signed short vs0 = vec_splat(vscales, 0);
-            vector signed short vs1 = vec_splat(vscales, 1);
-            vector signed short vs2 = vec_splat(vscales, 2);
-            vector signed short vs3 = vec_splat(vscales, 3);
-            vector signed short vs4 = vec_splat(vscales, 4);
-            vector signed short vs5 = vec_splat(vscales, 5);
-            vector signed short vs6 = vec_splat(vscales, 6);
-            vector signed short vs7 = vec_splat(vscales, 7);
-
-            vsumi0 = vec_msum(qv00, vs0, vsumi0);
-            vsumi1 = vec_msum(qv01, vs4, vsumi1);
-            vsumi2 = vec_msum(qv10, vs1, vsumi2);
-            vsumi3 = vec_msum(qv11, vs5, vsumi3);
-            vsumi4 = vec_msum(qv20, vs2, vsumi4);
-            vsumi5 = vec_msum(qv21, vs6, vsumi5);
-            vsumi6 = vec_msum(qv30, vs3, vsumi6);
-            vsumi7 = vec_msum(qv31, vs7, vsumi7);
-        }
-
-        vsumi0 = vec_add(vsumi0, vsumi4);
-        vsumi1 = vec_add(vsumi1, vsumi5);
-        vsumi2 = vec_add(vsumi2, vsumi6);
-        vsumi3 = vec_add(vsumi3, vsumi7);
-
-        vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
-        vsumf1 = vec_madd(vec_ctf(vsumi1, 0), vd, vsumf1);
-        vsumf2 = vec_madd(vec_ctf(vsumi2, 0), vd, vsumf2);
-        vsumf3 = vec_madd(vec_ctf(vsumi3, 0), vd, vsumf3);
-    }
-
-    vsumf0 = vec_add(vsumf0, vsumf2);
-    vsumf1 = vec_add(vsumf1, vsumf3);
-
-    vsumf0 = vec_add(vsumf0, vsumf1);
-
-    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4));
-    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8));
-
-    *s = vec_extract(vsumf0, 0);
-
-#elif defined __loongarch_asx
-
-    const __m256i m32s = __lasx_xvreplgr2vr_b(32);
-
-    __m256 acc = (__m256)__lasx_xvldi(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 q4 = x[i].ql;
-        const uint8_t * GGML_RESTRICT qh = x[i].qh;
-        const int8_t  * GGML_RESTRICT q8 = y[i].qs;
-
-        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);
-
-        for (int j = 0; j < QK_K/128; ++j) {
-
-            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_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_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 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);
-
-            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));
-        }
-
-        acc = __lasx_xvfmadd_s((__m256)__lasx_xvreplfr2vr_s(d), __lasx_xvffint_s_w(sumi), acc);
-    }
-
-    *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];
-    int16_t aux16[8];
-    float   sums [8];
-    int32_t aux32[8];
-    memset(sums, 0, 8*sizeof(float));
-
-    float sumf = 0;
-    for (int i = 0; i < nb; ++i) {
-        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 * 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;
-                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;
-        }
-        a = aux8;
-        int is = 0;
-        for (int j = 0; j < QK_K/16; ++j) {
-            int scale = x[i].scales[is++];
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-        }
-        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];
-    }
-    for (int l = 0; l < 8; ++l) sumf += sums[l];
-    *s = sumf;
-#endif
-}
-
-#if defined (__AVX__) || defined (__AVX2__) || defined (__ARM_NEON) || defined (__POWER9_VECTOR__) || defined(__loongarch_asx)
-static const int8_t keven_signs_q2xs[1024] = {
-     1,  1,  1,  1,  1,  1,  1,  1, -1,  1,  1,  1,  1,  1,  1, -1,  1, -1,  1,  1,  1,  1,  1, -1, -1, -1,  1,  1,  1,  1,  1,  1,
-     1,  1, -1,  1,  1,  1,  1, -1, -1,  1, -1,  1,  1,  1,  1,  1,  1, -1, -1,  1,  1,  1,  1,  1, -1, -1, -1,  1,  1,  1,  1, -1,
-     1,  1,  1, -1,  1,  1,  1, -1, -1,  1,  1, -1,  1,  1,  1,  1,  1, -1,  1, -1,  1,  1,  1,  1, -1, -1,  1, -1,  1,  1,  1, -1,
-     1,  1, -1, -1,  1,  1,  1,  1, -1,  1, -1, -1,  1,  1,  1, -1,  1, -1, -1, -1,  1,  1,  1, -1, -1, -1, -1, -1,  1,  1,  1,  1,
-     1,  1,  1,  1, -1,  1,  1, -1, -1,  1,  1,  1, -1,  1,  1,  1,  1, -1,  1,  1, -1,  1,  1,  1, -1, -1,  1,  1, -1,  1,  1, -1,
-     1,  1, -1,  1, -1,  1,  1,  1, -1,  1, -1,  1, -1,  1,  1, -1,  1, -1, -1,  1, -1,  1,  1, -1, -1, -1, -1,  1, -1,  1,  1,  1,
-     1,  1,  1, -1, -1,  1,  1,  1, -1,  1,  1, -1, -1,  1,  1, -1,  1, -1,  1, -1, -1,  1,  1, -1, -1, -1,  1, -1, -1,  1,  1,  1,
-     1,  1, -1, -1, -1,  1,  1, -1, -1,  1, -1, -1, -1,  1,  1,  1,  1, -1, -1, -1, -1,  1,  1,  1, -1, -1, -1, -1, -1,  1,  1, -1,
-     1,  1,  1,  1,  1, -1,  1, -1, -1,  1,  1,  1,  1, -1,  1,  1,  1, -1,  1,  1,  1, -1,  1,  1, -1, -1,  1,  1,  1, -1,  1, -1,
-     1,  1, -1,  1,  1, -1,  1,  1, -1,  1, -1,  1,  1, -1,  1, -1,  1, -1, -1,  1,  1, -1,  1, -1, -1, -1, -1,  1,  1, -1,  1,  1,
-     1,  1,  1, -1,  1, -1,  1,  1, -1,  1,  1, -1,  1, -1,  1, -1,  1, -1,  1, -1,  1, -1,  1, -1, -1, -1,  1, -1,  1, -1,  1,  1,
-     1,  1, -1, -1,  1, -1,  1, -1, -1,  1, -1, -1,  1, -1,  1,  1,  1, -1, -1, -1,  1, -1,  1,  1, -1, -1, -1, -1,  1, -1,  1, -1,
-     1,  1,  1,  1, -1, -1,  1,  1, -1,  1,  1,  1, -1, -1,  1, -1,  1, -1,  1,  1, -1, -1,  1, -1, -1, -1,  1,  1, -1, -1,  1,  1,
-     1,  1, -1,  1, -1, -1,  1, -1, -1,  1, -1,  1, -1, -1,  1,  1,  1, -1, -1,  1, -1, -1,  1,  1, -1, -1, -1,  1, -1, -1,  1, -1,
-     1,  1,  1, -1, -1, -1,  1, -1, -1,  1,  1, -1, -1, -1,  1,  1,  1, -1,  1, -1, -1, -1,  1,  1, -1, -1,  1, -1, -1, -1,  1, -1,
-     1,  1, -1, -1, -1, -1,  1,  1, -1,  1, -1, -1, -1, -1,  1, -1,  1, -1, -1, -1, -1, -1,  1, -1, -1, -1, -1, -1, -1, -1,  1,  1,
-     1,  1,  1,  1,  1,  1, -1, -1, -1,  1,  1,  1,  1,  1, -1,  1,  1, -1,  1,  1,  1,  1, -1,  1, -1, -1,  1,  1,  1,  1, -1, -1,
-     1,  1, -1,  1,  1,  1, -1,  1, -1,  1, -1,  1,  1,  1, -1, -1,  1, -1, -1,  1,  1,  1, -1, -1, -1, -1, -1,  1,  1,  1, -1,  1,
-     1,  1,  1, -1,  1,  1, -1,  1, -1,  1,  1, -1,  1,  1, -1, -1,  1, -1,  1, -1,  1,  1, -1, -1, -1, -1,  1, -1,  1,  1, -1,  1,
-     1,  1, -1, -1,  1,  1, -1, -1, -1,  1, -1, -1,  1,  1, -1,  1,  1, -1, -1, -1,  1,  1, -1,  1, -1, -1, -1, -1,  1,  1, -1, -1,
-     1,  1,  1,  1, -1,  1, -1,  1, -1,  1,  1,  1, -1,  1, -1, -1,  1, -1,  1,  1, -1,  1, -1, -1, -1, -1,  1,  1, -1,  1, -1,  1,
-     1,  1, -1,  1, -1,  1, -1, -1, -1,  1, -1,  1, -1,  1, -1,  1,  1, -1, -1,  1, -1,  1, -1,  1, -1, -1, -1,  1, -1,  1, -1, -1,
-     1,  1,  1, -1, -1,  1, -1, -1, -1,  1,  1, -1, -1,  1, -1,  1,  1, -1,  1, -1, -1,  1, -1,  1, -1, -1,  1, -1, -1,  1, -1, -1,
-     1,  1, -1, -1, -1,  1, -1,  1, -1,  1, -1, -1, -1,  1, -1, -1,  1, -1, -1, -1, -1,  1, -1, -1, -1, -1, -1, -1, -1,  1, -1,  1,
-     1,  1,  1,  1,  1, -1, -1,  1, -1,  1,  1,  1,  1, -1, -1, -1,  1, -1,  1,  1,  1, -1, -1, -1, -1, -1,  1,  1,  1, -1, -1,  1,
-     1,  1, -1,  1,  1, -1, -1, -1, -1,  1, -1,  1,  1, -1, -1,  1,  1, -1, -1,  1,  1, -1, -1,  1, -1, -1, -1,  1,  1, -1, -1, -1,
-     1,  1,  1, -1,  1, -1, -1, -1, -1,  1,  1, -1,  1, -1, -1,  1,  1, -1,  1, -1,  1, -1, -1,  1, -1, -1,  1, -1,  1, -1, -1, -1,
-     1,  1, -1, -1,  1, -1, -1,  1, -1,  1, -1, -1,  1, -1, -1, -1,  1, -1, -1, -1,  1, -1, -1, -1, -1, -1, -1, -1,  1, -1, -1,  1,
-     1,  1,  1,  1, -1, -1, -1, -1, -1,  1,  1,  1, -1, -1, -1,  1,  1, -1,  1,  1, -1, -1, -1,  1, -1, -1,  1,  1, -1, -1, -1, -1,
-     1,  1, -1,  1, -1, -1, -1,  1, -1,  1, -1,  1, -1, -1, -1, -1,  1, -1, -1,  1, -1, -1, -1, -1, -1, -1, -1,  1, -1, -1, -1,  1,
-     1,  1,  1, -1, -1, -1, -1,  1, -1,  1,  1, -1, -1, -1, -1, -1,  1, -1,  1, -1, -1, -1, -1, -1, -1, -1,  1, -1, -1, -1, -1,  1,
-     1,  1, -1, -1, -1, -1, -1, -1, -1,  1, -1, -1, -1, -1, -1,  1,  1, -1, -1, -1, -1, -1, -1,  1, -1, -1, -1, -1, -1, -1, -1, -1,
-};
-#endif
-
-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);
-    UNUSED(bx);
-    UNUSED(by);
-    UNUSED(bs);
-
-    const block_iq2_xxs * GGML_RESTRICT x = vx;
-    const block_q8_K    * GGML_RESTRICT y = vy;
-
-    const int nb = n / QK_K;
-
-#if defined(__ARM_NEON)
-
-    const uint64_t * signs64 = (const uint64_t *)keven_signs_q2xs;
-
-    uint32_t aux32[4];
-    const uint8_t * aux8 = (const uint8_t *)aux32;
-
-    ggml_int8x16x4_t q2u;
-    ggml_int8x16x4_t q2s;
-    ggml_int8x16x4_t q8b;
-
-    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; ib32 += 2) {
-            q8b = ggml_vld1q_s8_x4(q8); q8 += 64;
-            memcpy(aux32, q2, 4*sizeof(uint32_t)); q2 += 8;
-            q2u.val[0] = vcombine_s8(vld1_s8((const void *)(iq2xxs_grid + aux8[ 0])), vld1_s8((const void *)(iq2xxs_grid + aux8[ 1])));
-            q2u.val[1] = vcombine_s8(vld1_s8((const void *)(iq2xxs_grid + aux8[ 2])), vld1_s8((const void *)(iq2xxs_grid + aux8[ 3])));
-            q2u.val[2] = vcombine_s8(vld1_s8((const void *)(iq2xxs_grid + aux8[ 8])), vld1_s8((const void *)(iq2xxs_grid + aux8[ 9])));
-            q2u.val[3] = vcombine_s8(vld1_s8((const void *)(iq2xxs_grid + aux8[10])), vld1_s8((const void *)(iq2xxs_grid + aux8[11])));
-            q2s.val[0] = vcombine_s8(vld1_s8((const void *)(signs64 + ((aux32[1] >>  0) & 127))), vld1_s8((const void *)(signs64 + ((aux32[1] >>  7) & 127))));
-            q2s.val[1] = vcombine_s8(vld1_s8((const void *)(signs64 + ((aux32[1] >> 14) & 127))), vld1_s8((const void *)(signs64 + ((aux32[1] >> 21) & 127))));
-            q2s.val[2] = vcombine_s8(vld1_s8((const void *)(signs64 + ((aux32[3] >>  0) & 127))), vld1_s8((const void *)(signs64 + ((aux32[3] >>  7) & 127))));
-            q2s.val[3] = vcombine_s8(vld1_s8((const void *)(signs64 + ((aux32[3] >> 14) & 127))), vld1_s8((const void *)(signs64 + ((aux32[3] >> 21) & 127))));
-            q2u.val[0] = vmulq_s8(q2u.val[0], q2s.val[0]);
-            q2u.val[1] = vmulq_s8(q2u.val[1], q2s.val[1]);
-            q2u.val[2] = vmulq_s8(q2u.val[2], q2s.val[2]);
-            q2u.val[3] = vmulq_s8(q2u.val[3], q2s.val[3]);
-            const int32x4_t p1 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), q2u.val[0], q8b.val[0]), q2u.val[1], q8b.val[1]);
-            const int32x4_t p2 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), q2u.val[2], q8b.val[2]), q2u.val[3], q8b.val[3]);
-            sumf1 += vaddvq_s32(p1) * (0.5f + (aux32[1] >> 28));
-            sumf2 += vaddvq_s32(p2) * (0.5f + (aux32[3] >> 28));
-        }
-        sumf += d*(sumf1 + sumf2);
-    }
-    *s = 0.25f * sumf;
-
-#elif defined(__AVX2__)
-
-    const uint64_t * signs64 = (const uint64_t *)keven_signs_q2xs;
-
-    uint32_t aux32[4];
-    const uint8_t * aux8 = (const uint8_t *)aux32;
-
-    __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 * 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) {
-            const __m256i q8_1 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32;
-            const __m256i q8_2 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32;
-            memcpy(aux32, q2, 4*sizeof(uint32_t)); q2 += 8;
-            const __m256i q2_1 = _mm256_set_epi64x(iq2xxs_grid[aux8[ 3]], iq2xxs_grid[aux8[ 2]], iq2xxs_grid[aux8[1]], iq2xxs_grid[aux8[0]]);
-            const __m256i q2_2 = _mm256_set_epi64x(iq2xxs_grid[aux8[11]], iq2xxs_grid[aux8[10]], iq2xxs_grid[aux8[9]], iq2xxs_grid[aux8[8]]);
-            const __m256i s2_1 = _mm256_set_epi64x(signs64[(aux32[1] >> 21) & 127], signs64[(aux32[1] >> 14) & 127],
-                                                   signs64[(aux32[1] >>  7) & 127], signs64[(aux32[1] >>  0) & 127]);
-            const __m256i s2_2 = _mm256_set_epi64x(signs64[(aux32[3] >> 21) & 127], signs64[(aux32[3] >> 14) & 127],
-                                                   signs64[(aux32[3] >>  7) & 127], signs64[(aux32[3] >>  0) & 127]);
-            const __m256i q8s_1 = _mm256_sign_epi8(q8_1, s2_1);
-            const __m256i q8s_2 = _mm256_sign_epi8(q8_2, s2_2);
-            const __m256i dot1  = _mm256_maddubs_epi16(q2_1, q8s_1);
-            const __m256i dot2  = _mm256_maddubs_epi16(q2_2, q8s_2);
-            const uint16_t ls1 = aux32[1] >> 28;
-            const uint16_t ls2 = aux32[3] >> 28;
-            const __m256i p1 = _mm256_madd_epi16(dot1, _mm256_set1_epi16(2*ls1+1));
-            const __m256i p2 = _mm256_madd_epi16(dot2, _mm256_set1_epi16(2*ls2+1));
-            sumi1 = _mm256_add_epi32(sumi1, p1);
-            sumi2 = _mm256_add_epi32(sumi2, p2);
-        }
-
-        accumf = _mm256_fmadd_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(_mm256_add_epi32(sumi1, sumi2)), accumf);
-
-    }
-
-    *s = 0.125f * hsum_float_8(accumf);
-
-#elif defined(__AVX__)
-    const uint64_t * signs64 = (const uint64_t *)keven_signs_q2xs;
-
-    uint32_t aux32[4];
-    const uint8_t * aux8 = (const uint8_t *)aux32;
-
-    __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 * 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();
-        __m128i sumi2_1 = _mm_setzero_si128();
-        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
-            const __m128i q8_1_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
-            const __m128i q8_1_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
-            const __m128i q8_2_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
-            const __m128i q8_2_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
-            memcpy(aux32, q2, 4*sizeof(uint32_t)); q2 += 8;
-            const __m128i q2_1_0 = _mm_set_epi64x(iq2xxs_grid[aux8[1]], iq2xxs_grid[aux8[0]]);
-            const __m128i q2_1_1 = _mm_set_epi64x(iq2xxs_grid[aux8[3]], iq2xxs_grid[aux8[2]]);
-            const __m128i q2_2_0 = _mm_set_epi64x(iq2xxs_grid[aux8[9]], iq2xxs_grid[aux8[8]]);
-            const __m128i q2_2_1 = _mm_set_epi64x(iq2xxs_grid[aux8[11]], iq2xxs_grid[aux8[10]]);
-            const __m128i s2_1_0 = _mm_set_epi64x(signs64[(aux32[1] >>  7) & 127], signs64[(aux32[1] >>  0) & 127]);
-            const __m128i s2_1_1 = _mm_set_epi64x(signs64[(aux32[1] >> 21) & 127], signs64[(aux32[1] >> 14) & 127]);
-            const __m128i s2_2_0 = _mm_set_epi64x(signs64[(aux32[3] >>  7) & 127], signs64[(aux32[3] >>  0) & 127]);
-            const __m128i s2_2_1 = _mm_set_epi64x(signs64[(aux32[3] >> 21) & 127], signs64[(aux32[3] >> 14) & 127]);
-            const __m128i q8s_1_0 = _mm_sign_epi8(q8_1_0, s2_1_0);
-            const __m128i q8s_1_1 = _mm_sign_epi8(q8_1_1, s2_1_1);
-            const __m128i q8s_2_0 = _mm_sign_epi8(q8_2_0, s2_2_0);
-            const __m128i q8s_2_1 = _mm_sign_epi8(q8_2_1, s2_2_1);
-            const __m128i dot1_0  = _mm_maddubs_epi16(q2_1_0, q8s_1_0);
-            const __m128i dot1_1  = _mm_maddubs_epi16(q2_1_1, q8s_1_1);
-            const __m128i dot2_0  = _mm_maddubs_epi16(q2_2_0, q8s_2_0);
-            const __m128i dot2_1  = _mm_maddubs_epi16(q2_2_1, q8s_2_1);
-            const uint16_t ls1 = aux32[1] >> 28;
-            const uint16_t ls2 = aux32[3] >> 28;
-            const __m128i p1_0 = _mm_madd_epi16(dot1_0, _mm_set1_epi16(2*ls1+1));
-            const __m128i p1_1 = _mm_madd_epi16(dot1_1, _mm_set1_epi16(2*ls1+1));
-            const __m128i p2_0 = _mm_madd_epi16(dot2_0, _mm_set1_epi16(2*ls2+1));
-            const __m128i p2_1 = _mm_madd_epi16(dot2_1, _mm_set1_epi16(2*ls2+1));
-            sumi1_0 = _mm_add_epi32(sumi1_0, p1_0);
-            sumi1_1 = _mm_add_epi32(sumi1_1, p1_1);
-            sumi2_0 = _mm_add_epi32(sumi2_0, p2_0);
-            sumi2_1 = _mm_add_epi32(sumi2_1, p2_1);
-        }
-
-        accumf = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(MM256_SET_M128I(_mm_add_epi32(sumi1_1, sumi2_1), _mm_add_epi32(sumi1_0, sumi2_0)))), accumf);
-
-    }
-
-    *s = 0.125f * hsum_float_8(accumf);
-
-#elif defined(__POWER9_VECTOR__)
-    const vector int v0 = vec_splats((int32_t)0);
-    vector float vsumf0 = vec_splats(0.0f);
-    vector float vsumf1 = vec_splats(0.0f);
-    vector float vsumf2 = vec_splats(0.0f);
-    vector float vsumf3 = vec_splats(0.0f);
-
-    const uint64_t * signs64 = (const uint64_t *)keven_signs_q2xs;
-
-    for (int i = 0; i < nb; ++i) {
-        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d));
-        vector float vyd = vec_splats(y[i].d);
-        vector float vd = vec_mul(vxd, vyd);
-
-        vector signed int vsumi0 = v0;
-        vector signed int vsumi1 = v0;
-        vector signed int vsumi2 = v0;
-        vector signed int vsumi3 = v0;
-
-        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);
-            __builtin_prefetch(q8, 0, 1);
-
-            uint32_t aux32[4];
-            const uint8_t * aux8 = (const uint8_t *)aux32;
-
-            memcpy(aux32, q2, 4*sizeof(uint32_t));
-            q2 += 8;
-
-            vector signed long long aux64x2_0 = {*(const int64_t *)(iq2xxs_grid + aux8[ 0]), *(const int64_t *)(iq2xxs_grid + aux8[ 1])};
-            vector signed long long aux64x2_1 = {*(const int64_t *)(iq2xxs_grid + aux8[ 2]), *(const int64_t *)(iq2xxs_grid + aux8[ 3])};
-            vector signed long long aux64x2_2 = {*(const int64_t *)(iq2xxs_grid + aux8[ 8]), *(const int64_t *)(iq2xxs_grid + aux8[ 9])};
-            vector signed long long aux64x2_3 = {*(const int64_t *)(iq2xxs_grid + aux8[10]), *(const int64_t *)(iq2xxs_grid + aux8[11])};
-
-            vector signed long long vsigns0 = {*(const int64_t *)(signs64 + ((aux32[1] >>  0) & 127)), *(const int64_t *)(signs64 + ((aux32[1] >>  7) & 127))};
-            vector signed long long vsigns1 = {*(const int64_t *)(signs64 + ((aux32[1] >> 14) & 127)), *(const int64_t *)(signs64 + ((aux32[1] >> 21) & 127))};
-            vector signed long long vsigns2 = {*(const int64_t *)(signs64 + ((aux32[3] >>  0) & 127)), *(const int64_t *)(signs64 + ((aux32[3] >>  7) & 127))};
-            vector signed long long vsigns3 = {*(const int64_t *)(signs64 + ((aux32[3] >> 14) & 127)), *(const int64_t *)(signs64 + ((aux32[3] >> 21) & 127))};
-
-            vector signed char q2x0 = (vector signed char)vec_mul((vector signed char)vsigns0, (vector signed char)aux64x2_0);
-            vector signed char q2x1 = (vector signed char)vec_mul((vector signed char)vsigns1, (vector signed char)aux64x2_1);
-            vector signed char q2x2 = (vector signed char)vec_mul((vector signed char)vsigns2, (vector signed char)aux64x2_2);
-            vector signed char q2x3 = (vector signed char)vec_mul((vector signed char)vsigns3, (vector signed char)aux64x2_3);
-
-            vector signed char q8y0 = vec_xl( 0, q8);
-            vector signed char q8y1 = vec_xl(16, q8);
-            vector signed char q8y2 = vec_xl(32, q8);
-            vector signed char q8y3 = vec_xl(48, q8);
-            q8 += 64;
-
-            vector signed short qv0 = vec_add(vec_mule(q2x0, q8y0), vec_mulo(q2x0, q8y0));
-            vector signed short qv1 = vec_add(vec_mule(q2x1, q8y1), vec_mulo(q2x1, q8y1));
-            vector signed short qv2 = vec_add(vec_mule(q2x2, q8y2), vec_mulo(q2x2, q8y2));
-            vector signed short qv3 = vec_add(vec_mule(q2x3, q8y3), vec_mulo(q2x3, q8y3));
-
-            const uint16_t ls0 = aux32[1] >> 28;
-            const uint16_t ls1 = aux32[3] >> 28;
-
-            vector signed short vscales01 = vec_splats((int16_t)(2*ls0+1));
-            vector signed short vscales23 = vec_splats((int16_t)(2*ls1+1));
-
-            vsumi0 = vec_msum(qv0, vscales01, vsumi0);
-            vsumi1 = vec_msum(qv1, vscales01, vsumi1);
-            vsumi2 = vec_msum(qv2, vscales23, vsumi2);
-            vsumi3 = vec_msum(qv3, vscales23, vsumi3);
-        }
-
-        vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
-        vsumf1 = vec_madd(vec_ctf(vsumi1, 0), vd, vsumf1);
-        vsumf2 = vec_madd(vec_ctf(vsumi2, 0), vd, vsumf2);
-        vsumf3 = vec_madd(vec_ctf(vsumi3, 0), vd, vsumf3);
-    }
-
-    vsumf0 = vec_add(vsumf0, vsumf2);
-    vsumf1 = vec_add(vsumf1, vsumf3);
-
-    vsumf0 = vec_add(vsumf0, vsumf1);
-
-    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4));
-    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8));
-
-    *s = 0.125f * vec_extract(vsumf0, 0);
-
-#elif defined(__loongarch_asx)
-
-    const uint64_t * signs64 = (const uint64_t *)keven_signs_q2xs;
-
-    uint32_t aux32[4];
-    const uint8_t * aux8 = (const uint8_t *)aux32;
-
-    __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 * 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) {
-            const __m256i q8_1 = __lasx_xvld((const __m256i *)q8, 0); q8 += 32;
-            const __m256i q8_2 = __lasx_xvld((const __m256i *)q8, 0); q8 += 32;
-            memcpy(aux32, q2, 4*sizeof(uint32_t)); q2 += 8;
-
-            const __m256i q2_1 = lasx_set_d(iq2xxs_grid[aux8[ 3]], iq2xxs_grid[aux8[ 2]], iq2xxs_grid[aux8[1]], iq2xxs_grid[aux8[0]]);
-            const __m256i q2_2 = lasx_set_d(iq2xxs_grid[aux8[11]], iq2xxs_grid[aux8[10]], iq2xxs_grid[aux8[9]], iq2xxs_grid[aux8[8]]);
-            const __m256i s2_1 = lasx_set_d(signs64[(aux32[1] >> 21) & 127], signs64[(aux32[1] >> 14) & 127],
-                                                   signs64[(aux32[1] >>  7) & 127], signs64[(aux32[1] >>  0) & 127]);
-            const __m256i s2_2 = lasx_set_d(signs64[(aux32[3] >> 21) & 127], signs64[(aux32[3] >> 14) & 127],
-                                                   signs64[(aux32[3] >>  7) & 127], signs64[(aux32[3] >>  0) & 127]);
-            const __m256i q8s_1 = __lasx_xvsigncov_b(s2_1, q8_1);
-            const __m256i q8s_2 = __lasx_xvsigncov_b(s2_2, q8_2);
-            const __m256i dot1  = lasx_maddubs_h(q2_1, q8s_1);
-            const __m256i dot2  = lasx_maddubs_h(q2_2, q8s_2);
-            const uint16_t ls1 = aux32[1] >> 28;
-            const uint16_t ls2 = aux32[3] >> 28;
-            const __m256i p1 = lasx_madd_h(dot1, __lasx_xvreplgr2vr_h(2*ls1+1));
-            const __m256i p2 = lasx_madd_h(dot2, __lasx_xvreplgr2vr_h(2*ls2+1));
-            sumi1 = __lasx_xvadd_w(sumi1, p1);
-            sumi2 = __lasx_xvadd_w(sumi2, p2);
-        }
-
-        accumf = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(d), __lasx_xvffint_s_w(__lasx_xvadd_w(sumi1, sumi2)), accumf);
-    }
-
-    *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];
-    const uint8_t * aux8 = (const uint8_t *)aux32;
-
-    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 * 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));
-            q2 += 4;
-            const uint32_t ls = 2*(aux32[1] >> 28) + 1;
-            int32_t sumi = 0;
-            for (int l = 0; l < 4; ++l) {
-                const uint8_t * grid = (const uint8_t *)(iq2xxs_grid + aux8[l]);
-                const uint8_t  signs = ksigns_iq2xs[(aux32[1] >> 7*l) & 127];
-                for (int j = 0; j < 8; ++j) {
-                    sumi += grid[j] * q8[j] * (signs & kmask_iq2xs[j] ? -1 : 1);
-                }
-                q8 += 8;
-            }
-            bsum += sumi * ls;
-        }
-        sumf += d * bsum;
-    }
-    *s = 0.125f * sumf;
-#endif
-}
-
-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);
-    UNUSED(bx);
-    UNUSED(by);
-    UNUSED(bs);
-
-    const block_iq2_xs * GGML_RESTRICT x = vx;
-    const block_q8_K   * GGML_RESTRICT y = vy;
-
-    const int nb = n / QK_K;
-
-#if defined(__ARM_NEON)
-
-    const uint64_t * signs64 = (const uint64_t *)keven_signs_q2xs;
-
-    ggml_int8x16x4_t q2u;
-    ggml_int8x16x4_t q2s;
-    ggml_int8x16x4_t q8b;
-
-    int32x4x4_t scales32;
-
-    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;
-        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);
-        uint8x16_t scales = vcombine_u8(vzip1_u8(scales_l, scales_h), vzip2_u8(scales_l, scales_h));
-        scales = vaddq_u8(vshlq_n_u8(scales, 1), vdupq_n_u8(1));
-        const uint16x8_t scales1 = vmovl_u8(vget_low_u8(scales));
-        const uint16x8_t scales2 = vmovl_u8(vget_high_u8(scales));
-        scales32.val[0] = vreinterpretq_s32_u32(vmovl_u16(vget_low_u16(scales1)));
-        scales32.val[1] = vreinterpretq_s32_u32(vmovl_u16(vget_high_u16(scales1)));
-        scales32.val[2] = vreinterpretq_s32_u32(vmovl_u16(vget_low_u16(scales2)));
-        scales32.val[3] = vreinterpretq_s32_u32(vmovl_u16(vget_high_u16(scales2)));
-        int32x4_t sumi = vdupq_n_s32(0);
-        for (int ib64 = 0; ib64 < QK_K/64; ++ib64) {
-            q8b = ggml_vld1q_s8_x4(q8); q8 += 64;
-            q2u.val[0] = vcombine_s8(vld1_s8((const void *)(iq2xs_grid + (q2[0] & 511))), vld1_s8((const void *)(iq2xs_grid + (q2[1] & 511))));
-            q2u.val[1] = vcombine_s8(vld1_s8((const void *)(iq2xs_grid + (q2[2] & 511))), vld1_s8((const void *)(iq2xs_grid + (q2[3] & 511))));
-            q2u.val[2] = vcombine_s8(vld1_s8((const void *)(iq2xs_grid + (q2[4] & 511))), vld1_s8((const void *)(iq2xs_grid + (q2[5] & 511))));
-            q2u.val[3] = vcombine_s8(vld1_s8((const void *)(iq2xs_grid + (q2[6] & 511))), vld1_s8((const void *)(iq2xs_grid + (q2[7] & 511))));
-            q2s.val[0] = vcombine_s8(vld1_s8((const void *)(signs64 + (q2[0] >> 9))), vld1_s8((const void *)(signs64 + (q2[1] >> 9))));
-            q2s.val[1] = vcombine_s8(vld1_s8((const void *)(signs64 + (q2[2] >> 9))), vld1_s8((const void *)(signs64 + (q2[3] >> 9))));
-            q2s.val[2] = vcombine_s8(vld1_s8((const void *)(signs64 + (q2[4] >> 9))), vld1_s8((const void *)(signs64 + (q2[5] >> 9))));
-            q2s.val[3] = vcombine_s8(vld1_s8((const void *)(signs64 + (q2[6] >> 9))), vld1_s8((const void *)(signs64 + (q2[7] >> 9))));
-            q2u.val[0] = vmulq_s8(q2u.val[0], q2s.val[0]);
-            q2u.val[1] = vmulq_s8(q2u.val[1], q2s.val[1]);
-            q2u.val[2] = vmulq_s8(q2u.val[2], q2s.val[2]);
-            q2u.val[3] = vmulq_s8(q2u.val[3], q2s.val[3]);
-            const int32x4_t p1 = ggml_vdotq_s32(vdupq_n_s32(0), q2u.val[0], q8b.val[0]);
-            const int32x4_t p2 = ggml_vdotq_s32(vdupq_n_s32(0), q2u.val[1], q8b.val[1]);
-            const int32x4_t p3 = ggml_vdotq_s32(vdupq_n_s32(0), q2u.val[2], q8b.val[2]);
-            const int32x4_t p4 = ggml_vdotq_s32(vdupq_n_s32(0), q2u.val[3], q8b.val[3]);
-            const int32x4_t p = vpaddq_s32(vpaddq_s32(p1, p2), vpaddq_s32(p3, p4));
-            sumi = vmlaq_s32(sumi, p, scales32.val[ib64]);
-            q2 += 8;
-        }
-        sumf += d*vaddvq_s32(sumi);
-    }
-    *s = 0.125f * sumf;
-
-#elif defined(__AVX2__)
-
-    const __m256i mone = _mm256_set1_epi8(1);
-    static const char block_sign_shuffle_mask_1[32] = {
-        0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02,
-        0x04, 0x04, 0x04, 0x04, 0x04, 0x04, 0x04, 0x04, 0x06, 0x06, 0x06, 0x06, 0x06, 0x06, 0x06, 0x06,
-    };
-    static const char block_sign_shuffle_mask_2[32] = {
-        0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x0a, 0x0a, 0x0a, 0x0a, 0x0a, 0x0a, 0x0a, 0x0a,
-        0x0c, 0x0c, 0x0c, 0x0c, 0x0c, 0x0c, 0x0c, 0x0c, 0x0e, 0x0e, 0x0e, 0x0e, 0x0e, 0x0e, 0x0e, 0x0e,
-    };
-    static const uint8_t bit_selector_mask_bytes[32] = {
-        0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
-        0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
-    };
-
-    const __m256i bit_selector_mask = _mm256_loadu_si256((const __m256i*)bit_selector_mask_bytes);
-    const __m256i block_sign_shuffle_1 = _mm256_loadu_si256((const __m256i*)block_sign_shuffle_mask_1);
-    const __m256i block_sign_shuffle_2 = _mm256_loadu_si256((const __m256i*)block_sign_shuffle_mask_2);
-
-    static const uint8_t k_bit_helper[32] = {
-        0x00, 0x80, 0x80, 0x00, 0x80, 0x00, 0x00, 0x80, 0x80, 0x00, 0x00, 0x80, 0x00, 0x80, 0x80, 0x00,
-        0x00, 0x80, 0x80, 0x00, 0x80, 0x00, 0x00, 0x80, 0x80, 0x00, 0x00, 0x80, 0x00, 0x80, 0x80, 0x00,
-    };
-    const __m256i bit_helper = _mm256_loadu_si256((const __m256i*)k_bit_helper);
-    const __m256i m511 = _mm256_set1_epi16(511);
-    const __m128i m4 = _mm_set1_epi8(0xf);
-    const __m128i m1 = _mm_set1_epi8(1);
-
-    uint64_t aux64;
-
-    // somewhat hacky, but gives a significant boost in performance
-    __m256i aux_gindex;
-    const uint16_t * gindex = (const uint16_t *)&aux_gindex;
-
-    __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 * 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);
-        stmp = _mm_unpacklo_epi8(_mm_and_si128(stmp, m4), _mm_and_si128(_mm_srli_epi16(stmp, 4), m4));
-        const __m128i scales = _mm_add_epi8(_mm_slli_epi16(stmp, 1), m1);
-
-        __m256i sumi1 = _mm256_setzero_si256();
-        __m256i sumi2 = _mm256_setzero_si256();
-        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 4) {
-
-            const __m256i q2_data = _mm256_loadu_si256((const __m256i*)q2);  q2 += 16;
-            aux_gindex = _mm256_and_si256(q2_data, m511);
-
-            const __m256i partial_sign_bits = _mm256_srli_epi16(q2_data, 9);
-            const __m256i partial_sign_bits_upper = _mm256_srli_epi16(q2_data, 13);
-            const __m256i partial_sign_bits_for_counting = _mm256_xor_si256(partial_sign_bits, partial_sign_bits_upper);
-
-            const __m256i odd_bits = _mm256_shuffle_epi8(bit_helper, partial_sign_bits_for_counting);
-            const __m256i full_sign_bits = _mm256_or_si256(partial_sign_bits, odd_bits);
-
-            const __m256i q8_1 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32;
-            const __m256i q8_2 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32;
-            const __m256i q8_3 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32;
-            const __m256i q8_4 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32;
-
-            const __m256i q2_1 = _mm256_set_epi64x(iq2xs_grid[gindex[ 3]], iq2xs_grid[gindex[ 2]],
-                                                   iq2xs_grid[gindex[ 1]], iq2xs_grid[gindex[ 0]]);
-            const __m256i q2_2 = _mm256_set_epi64x(iq2xs_grid[gindex[ 7]], iq2xs_grid[gindex[ 6]],
-                                                   iq2xs_grid[gindex[ 5]], iq2xs_grid[gindex[ 4]]);
-            const __m256i q2_3 = _mm256_set_epi64x(iq2xs_grid[gindex[11]], iq2xs_grid[gindex[10]],
-                                                   iq2xs_grid[gindex[ 9]], iq2xs_grid[gindex[ 8]]);
-            const __m256i q2_4 = _mm256_set_epi64x(iq2xs_grid[gindex[15]], iq2xs_grid[gindex[14]],
-                                                   iq2xs_grid[gindex[13]], iq2xs_grid[gindex[12]]);
-
-            const __m128i full_signs_l = _mm256_castsi256_si128(full_sign_bits);
-            const __m128i full_signs_h = _mm256_extractf128_si256(full_sign_bits, 1);
-            const __m256i full_signs_1 = MM256_SET_M128I(full_signs_l, full_signs_l);
-            const __m256i full_signs_2 = MM256_SET_M128I(full_signs_h, full_signs_h);
-
-            __m256i signs;
-            signs = _mm256_shuffle_epi8(full_signs_1, block_sign_shuffle_1);
-            signs = _mm256_cmpeq_epi8(_mm256_and_si256(signs, bit_selector_mask), bit_selector_mask);
-            const __m256i q8s_1 = _mm256_sign_epi8(q8_1, _mm256_or_si256(signs, mone));
-
-            signs = _mm256_shuffle_epi8(full_signs_1, block_sign_shuffle_2);
-            signs = _mm256_cmpeq_epi8(_mm256_and_si256(signs, bit_selector_mask), bit_selector_mask);
-            const __m256i q8s_2 = _mm256_sign_epi8(q8_2, _mm256_or_si256(signs, mone));
-
-            signs = _mm256_shuffle_epi8(full_signs_2, block_sign_shuffle_1);
-            signs = _mm256_cmpeq_epi8(_mm256_and_si256(signs, bit_selector_mask), bit_selector_mask);
-            const __m256i q8s_3 = _mm256_sign_epi8(q8_3, _mm256_or_si256(signs, mone));
-
-            signs = _mm256_shuffle_epi8(full_signs_2, block_sign_shuffle_2);
-            signs = _mm256_cmpeq_epi8(_mm256_and_si256(signs, bit_selector_mask), bit_selector_mask);
-            const __m256i q8s_4 = _mm256_sign_epi8(q8_4, _mm256_or_si256(signs, mone));
-
-            const __m256i dot1  = _mm256_maddubs_epi16(q2_1, q8s_1);
-            const __m256i dot2  = _mm256_maddubs_epi16(q2_2, q8s_2);
-            const __m256i dot3  = _mm256_maddubs_epi16(q2_3, q8s_3);
-            const __m256i dot4  = _mm256_maddubs_epi16(q2_4, q8s_4);
-
-            const __m256i sc1 = _mm256_cvtepi8_epi16(_mm_shuffle_epi8(scales, get_scale_shuffle(ib32+0)));
-            const __m256i sc2 = _mm256_cvtepi8_epi16(_mm_shuffle_epi8(scales, get_scale_shuffle(ib32+1)));
-            const __m256i sc3 = _mm256_cvtepi8_epi16(_mm_shuffle_epi8(scales, get_scale_shuffle(ib32+2)));
-            const __m256i sc4 = _mm256_cvtepi8_epi16(_mm_shuffle_epi8(scales, get_scale_shuffle(ib32+3)));
-
-            sumi1 = _mm256_add_epi32(sumi1, _mm256_madd_epi16(dot1, sc1));
-            sumi2 = _mm256_add_epi32(sumi2, _mm256_madd_epi16(dot2, sc2));
-            sumi1 = _mm256_add_epi32(sumi1, _mm256_madd_epi16(dot3, sc3));
-            sumi2 = _mm256_add_epi32(sumi2, _mm256_madd_epi16(dot4, sc4));
-        }
-
-        accumf = _mm256_fmadd_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(_mm256_add_epi32(sumi1, sumi2)), accumf);
-
-    }
-
-    *s = 0.125f * hsum_float_8(accumf);
-
-#elif defined(__AVX__)
-    const __m128i mone = _mm_set1_epi8(1);
-    static const char block_sign_shuffle_mask_1[32] = {
-        0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02,
-        0x04, 0x04, 0x04, 0x04, 0x04, 0x04, 0x04, 0x04, 0x06, 0x06, 0x06, 0x06, 0x06, 0x06, 0x06, 0x06,
-    };
-    static const char block_sign_shuffle_mask_2[32] = {
-        0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x0a, 0x0a, 0x0a, 0x0a, 0x0a, 0x0a, 0x0a, 0x0a,
-        0x0c, 0x0c, 0x0c, 0x0c, 0x0c, 0x0c, 0x0c, 0x0c, 0x0e, 0x0e, 0x0e, 0x0e, 0x0e, 0x0e, 0x0e, 0x0e,
-    };
-    static const uint8_t bit_selector_mask_bytes[32] = {
-        0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
-        0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
-    };
-
-    const __m128i bit_selector_mask_0 = _mm_loadu_si128((const __m128i*)bit_selector_mask_bytes);
-    const __m128i bit_selector_mask_1 = _mm_loadu_si128((const __m128i*)bit_selector_mask_bytes + 1);
-    const __m128i block_sign_shuffle_1_0 = _mm_loadu_si128((const __m128i*)block_sign_shuffle_mask_1);
-    const __m128i block_sign_shuffle_1_1 = _mm_loadu_si128((const __m128i*)block_sign_shuffle_mask_1 + 1);
-    const __m128i block_sign_shuffle_2_0 = _mm_loadu_si128((const __m128i*)block_sign_shuffle_mask_2);
-    const __m128i block_sign_shuffle_2_1 = _mm_loadu_si128((const __m128i*)block_sign_shuffle_mask_2 + 1);
-
-    static const uint8_t k_bit_helper[32] = {
-        0x00, 0x80, 0x80, 0x00, 0x80, 0x00, 0x00, 0x80, 0x80, 0x00, 0x00, 0x80, 0x00, 0x80, 0x80, 0x00,
-        0x00, 0x80, 0x80, 0x00, 0x80, 0x00, 0x00, 0x80, 0x80, 0x00, 0x00, 0x80, 0x00, 0x80, 0x80, 0x00,
-    };
-    const __m128i bit_helper_0 = _mm_loadu_si128((const __m128i*)k_bit_helper);
-    const __m128i bit_helper_1 = _mm_loadu_si128((const __m128i*)k_bit_helper + 1);
-    const __m128i m511 = _mm_set1_epi16(511);
-    const __m128i m4 = _mm_set1_epi8(0xf);
-    const __m128i m1 = _mm_set1_epi8(1);
-
-    uint64_t aux64;
-
-    // somewhat hacky, but gives a significant boost in performance
-    __m256i aux_gindex;
-    const uint16_t * gindex = (const uint16_t *)&aux_gindex;
-
-    __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 * 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);
-        stmp = _mm_unpacklo_epi8(_mm_and_si128(stmp, m4), _mm_and_si128(_mm_srli_epi16(stmp, 4), m4));
-        const __m128i scales = _mm_add_epi8(_mm_slli_epi16(stmp, 1), m1);
-
-        __m128i sumi1_0 = _mm_setzero_si128();
-        __m128i sumi1_1 = _mm_setzero_si128();
-        __m128i sumi2_0 = _mm_setzero_si128();
-        __m128i sumi2_1 = _mm_setzero_si128();
-        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 4) {
-
-            const __m128i q2_data_0 = _mm_loadu_si128((const __m128i*)q2);
-            const __m128i q2_data_1 = _mm_loadu_si128((const __m128i*)q2 + 1);  q2 += 16;
-            aux_gindex = MM256_SET_M128I(_mm_and_si128(q2_data_1, m511), _mm_and_si128(q2_data_0, m511));
-
-            const __m128i partial_sign_bits_0 = _mm_srli_epi16(q2_data_0, 9);
-            const __m128i partial_sign_bits_1 = _mm_srli_epi16(q2_data_1, 9);
-            const __m128i partial_sign_bits_upper_0 = _mm_srli_epi16(q2_data_0, 13);
-            const __m128i partial_sign_bits_upper_1 = _mm_srli_epi16(q2_data_1, 13);
-            const __m128i partial_sign_bits_for_counting_0 = _mm_xor_si128(partial_sign_bits_0, partial_sign_bits_upper_0);
-            const __m128i partial_sign_bits_for_counting_1 = _mm_xor_si128(partial_sign_bits_1, partial_sign_bits_upper_1);
-
-            const __m128i odd_bits_0 = _mm_shuffle_epi8(bit_helper_0, partial_sign_bits_for_counting_0);
-            const __m128i odd_bits_1 = _mm_shuffle_epi8(bit_helper_1, partial_sign_bits_for_counting_1);
-            const __m128i full_sign_bits_0 = _mm_or_si128(partial_sign_bits_0, odd_bits_0);
-            const __m128i full_sign_bits_1 = _mm_or_si128(partial_sign_bits_1, odd_bits_1);
-
-            const __m128i q8_1_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
-            const __m128i q8_1_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
-            const __m128i q8_2_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
-            const __m128i q8_2_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
-            const __m128i q8_3_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
-            const __m128i q8_3_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
-            const __m128i q8_4_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
-            const __m128i q8_4_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
-
-            const __m128i q2_1_0 = _mm_set_epi64x(iq2xs_grid[gindex[1]], iq2xs_grid[gindex[0]]);
-            const __m128i q2_1_1 = _mm_set_epi64x(iq2xs_grid[gindex[3]], iq2xs_grid[gindex[2]]);
-            const __m128i q2_2_0 = _mm_set_epi64x(iq2xs_grid[gindex[5]], iq2xs_grid[gindex[4]]);
-            const __m128i q2_2_1 = _mm_set_epi64x(iq2xs_grid[gindex[7]], iq2xs_grid[gindex[6]]);
-            const __m128i q2_3_0 = _mm_set_epi64x(iq2xs_grid[gindex[9]], iq2xs_grid[gindex[8]]);
-            const __m128i q2_3_1 = _mm_set_epi64x(iq2xs_grid[gindex[11]], iq2xs_grid[gindex[10]]);
-            const __m128i q2_4_0 = _mm_set_epi64x(iq2xs_grid[gindex[13]], iq2xs_grid[gindex[12]]);
-            const __m128i q2_4_1 = _mm_set_epi64x(iq2xs_grid[gindex[15]], iq2xs_grid[gindex[14]]);
-
-            // AVX2 full_signs_1 is full_sign_bits_0 here
-            // AVX2 full_signs_2 is full_sign_bits_1 here
-            __m128i signs_0, signs_1;
-            signs_0 = _mm_shuffle_epi8(full_sign_bits_0, block_sign_shuffle_1_0);
-            signs_1 = _mm_shuffle_epi8(full_sign_bits_0, block_sign_shuffle_1_1);
-            signs_0 = _mm_cmpeq_epi8(_mm_and_si128(signs_0, bit_selector_mask_0), bit_selector_mask_0);
-            signs_1 = _mm_cmpeq_epi8(_mm_and_si128(signs_1, bit_selector_mask_1), bit_selector_mask_1);
-            const __m128i q8s_1_0 = _mm_sign_epi8(q8_1_0, _mm_or_si128(signs_0, mone));
-            const __m128i q8s_1_1 = _mm_sign_epi8(q8_1_1, _mm_or_si128(signs_1, mone));
-
-            signs_0 = _mm_shuffle_epi8(full_sign_bits_0, block_sign_shuffle_2_0);
-            signs_1 = _mm_shuffle_epi8(full_sign_bits_0, block_sign_shuffle_2_1);
-            signs_0 = _mm_cmpeq_epi8(_mm_and_si128(signs_0, bit_selector_mask_0), bit_selector_mask_0);
-            signs_1 = _mm_cmpeq_epi8(_mm_and_si128(signs_1, bit_selector_mask_1), bit_selector_mask_1);
-            const __m128i q8s_2_0 = _mm_sign_epi8(q8_2_0, _mm_or_si128(signs_0, mone));
-            const __m128i q8s_2_1 = _mm_sign_epi8(q8_2_1, _mm_or_si128(signs_1, mone));
-
-            signs_0 = _mm_shuffle_epi8(full_sign_bits_1, block_sign_shuffle_1_0);
-            signs_1 = _mm_shuffle_epi8(full_sign_bits_1, block_sign_shuffle_1_1);
-            signs_0 = _mm_cmpeq_epi8(_mm_and_si128(signs_0, bit_selector_mask_0), bit_selector_mask_0);
-            signs_1 = _mm_cmpeq_epi8(_mm_and_si128(signs_1, bit_selector_mask_1), bit_selector_mask_1);
-            const __m128i q8s_3_0 = _mm_sign_epi8(q8_3_0, _mm_or_si128(signs_0, mone));
-            const __m128i q8s_3_1 = _mm_sign_epi8(q8_3_1, _mm_or_si128(signs_1, mone));
-
-            signs_0 = _mm_shuffle_epi8(full_sign_bits_1, block_sign_shuffle_2_0);
-            signs_1 = _mm_shuffle_epi8(full_sign_bits_1, block_sign_shuffle_2_1);
-            signs_0 = _mm_cmpeq_epi8(_mm_and_si128(signs_0, bit_selector_mask_0), bit_selector_mask_0);
-            signs_1 = _mm_cmpeq_epi8(_mm_and_si128(signs_1, bit_selector_mask_1), bit_selector_mask_1);
-            const __m128i q8s_4_0 = _mm_sign_epi8(q8_4_0, _mm_or_si128(signs_0, mone));
-            const __m128i q8s_4_1 = _mm_sign_epi8(q8_4_1, _mm_or_si128(signs_1, mone));
-
-            const __m128i dot1_0  = _mm_maddubs_epi16(q2_1_0, q8s_1_0);
-            const __m128i dot1_1  = _mm_maddubs_epi16(q2_1_1, q8s_1_1);
-            const __m128i dot2_0  = _mm_maddubs_epi16(q2_2_0, q8s_2_0);
-            const __m128i dot2_1  = _mm_maddubs_epi16(q2_2_1, q8s_2_1);
-            const __m128i dot3_0  = _mm_maddubs_epi16(q2_3_0, q8s_3_0);
-            const __m128i dot3_1  = _mm_maddubs_epi16(q2_3_1, q8s_3_1);
-            const __m128i dot4_0  = _mm_maddubs_epi16(q2_4_0, q8s_4_0);
-            const __m128i dot4_1  = _mm_maddubs_epi16(q2_4_1, q8s_4_1);
-
-            __m128i sc_tmp = _mm_shuffle_epi8(scales, get_scale_shuffle(ib32+0));
-            const __m128i sc1_0 = _mm_cvtepi8_epi16(sc_tmp);
-            const __m128i sc1_1 = _mm_cvtepi8_epi16(_mm_srli_si128(sc_tmp, 8));
-            sc_tmp = _mm_shuffle_epi8(scales, get_scale_shuffle(ib32+1));
-            const __m128i sc2_0 = _mm_cvtepi8_epi16(sc_tmp);
-            const __m128i sc2_1 = _mm_cvtepi8_epi16(_mm_srli_si128(sc_tmp, 8));
-            sc_tmp = _mm_shuffle_epi8(scales, get_scale_shuffle(ib32+2));
-            const __m128i sc3_0 = _mm_cvtepi8_epi16(sc_tmp);
-            const __m128i sc3_1 = _mm_cvtepi8_epi16(_mm_srli_si128(sc_tmp, 8));
-            sc_tmp = _mm_shuffle_epi8(scales, get_scale_shuffle(ib32+3));
-            const __m128i sc4_0 = _mm_cvtepi8_epi16(sc_tmp);
-            const __m128i sc4_1 = _mm_cvtepi8_epi16(_mm_srli_si128(sc_tmp, 8));
-
-            sumi1_0 = _mm_add_epi32(sumi1_0, _mm_madd_epi16(dot1_0, sc1_0));
-            sumi1_1 = _mm_add_epi32(sumi1_1, _mm_madd_epi16(dot1_1, sc1_1));
-            sumi2_0 = _mm_add_epi32(sumi2_0, _mm_madd_epi16(dot2_0, sc2_0));
-            sumi2_1 = _mm_add_epi32(sumi2_1, _mm_madd_epi16(dot2_1, sc2_1));
-            sumi1_0 = _mm_add_epi32(sumi1_0, _mm_madd_epi16(dot3_0, sc3_0));
-            sumi1_1 = _mm_add_epi32(sumi1_1, _mm_madd_epi16(dot3_1, sc3_1));
-            sumi2_0 = _mm_add_epi32(sumi2_0, _mm_madd_epi16(dot4_0, sc4_0));
-            sumi2_1 = _mm_add_epi32(sumi2_1, _mm_madd_epi16(dot4_1, sc4_1));
-        }
-
-        accumf = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(MM256_SET_M128I(_mm_add_epi32(sumi1_1, sumi2_1), _mm_add_epi32(sumi1_0, sumi2_0)))), accumf);
-
-    }
-
-    *s = 0.125f * hsum_float_8(accumf);
-
-#elif defined(__loongarch_asx)
-
-    const __m256i mone = __lasx_xvreplgr2vr_b(1);
-    static const char block_sign_shuffle_mask_1[32] = {
-        0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02,
-        0x04, 0x04, 0x04, 0x04, 0x04, 0x04, 0x04, 0x04, 0x06, 0x06, 0x06, 0x06, 0x06, 0x06, 0x06, 0x06,
-    };
-    static const char block_sign_shuffle_mask_2[32] = {
-        0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x0a, 0x0a, 0x0a, 0x0a, 0x0a, 0x0a, 0x0a, 0x0a,
-        0x0c, 0x0c, 0x0c, 0x0c, 0x0c, 0x0c, 0x0c, 0x0c, 0x0e, 0x0e, 0x0e, 0x0e, 0x0e, 0x0e, 0x0e, 0x0e,
-    };
-    static const uint8_t bit_selector_mask_bytes[32] = {
-        0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
-        0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
-    };
-
-    const __m256i bit_selector_mask = __lasx_xvld((const __m256i*)bit_selector_mask_bytes, 0);
-    const __m256i block_sign_shuffle_1 = __lasx_xvld((const __m256i*)block_sign_shuffle_mask_1, 0);
-    const __m256i block_sign_shuffle_2 = __lasx_xvld((const __m256i*)block_sign_shuffle_mask_2, 0);
-
-    static const uint8_t k_bit_helper[32] = {
-        0x00, 0x80, 0x80, 0x00, 0x80, 0x00, 0x00, 0x80, 0x80, 0x00, 0x00, 0x80, 0x00, 0x80, 0x80, 0x00,
-        0x00, 0x80, 0x80, 0x00, 0x80, 0x00, 0x00, 0x80, 0x80, 0x00, 0x00, 0x80, 0x00, 0x80, 0x80, 0x00,
-    };
-    const __m256i bit_helper = __lasx_xvld((const __m256i*)k_bit_helper, 0);
-    const __m256i m511 = __lasx_xvreplgr2vr_h(511);
-    const __m128i m4 = __lsx_vreplgr2vr_b(0xf);
-    const __m128i m1 = __lsx_vreplgr2vr_b(1);
-
-    uint64_t aux64;
-
-    // somewhat hacky, but gives a significant boost in performance
-    __m256i aux_gindex;
-    const uint16_t * gindex = (const uint16_t *)&aux_gindex;
-
-    __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 * 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);
-        stmp = __lsx_vilvl_b( __lsx_vand_v(__lsx_vsrli_h(stmp, 4), m4), __lsx_vand_v(stmp, m4));
-        const __m128i scales = __lsx_vadd_b(__lsx_vslli_h(stmp, 1), m1);
-
-        __m256i sumi1 = __lasx_xvldi(0);
-        __m256i sumi2 = __lasx_xvldi(0);
-        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 4) {
-
-            const __m256i q2_data = __lasx_xvld((const __m256i*)q2, 0);  q2 += 16;
-            aux_gindex = __lasx_xvand_v(q2_data, m511);
-
-            const __m256i partial_sign_bits = __lasx_xvsrli_h(q2_data, 9);
-            const __m256i partial_sign_bits_upper = __lasx_xvsrli_h(q2_data, 13);
-            const __m256i partial_sign_bits_for_counting = __lasx_xvxor_v(partial_sign_bits, partial_sign_bits_upper);
-
-            const __m256i odd_bits = lasx_shuffle_b(bit_helper, partial_sign_bits_for_counting);
-            const __m256i full_sign_bits = __lasx_xvor_v(partial_sign_bits, odd_bits);
-
-            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;
-            const __m256i q8_4 = __lasx_xvld((const __m256i *)q8, 0); q8 += 32;
-
-            const __m256i q2_1 = lasx_set_d(iq2xs_grid[gindex[ 3]], iq2xs_grid[gindex[ 2]],
-                                                   iq2xs_grid[gindex[ 1]], iq2xs_grid[gindex[ 0]]);
-            const __m256i q2_2 = lasx_set_d(iq2xs_grid[gindex[ 7]], iq2xs_grid[gindex[ 6]],
-                                                   iq2xs_grid[gindex[ 5]], iq2xs_grid[gindex[ 4]]);
-            const __m256i q2_3 = lasx_set_d(iq2xs_grid[gindex[11]], iq2xs_grid[gindex[10]],
-                                                   iq2xs_grid[gindex[ 9]], iq2xs_grid[gindex[ 8]]);
-            const __m256i q2_4 = lasx_set_d(iq2xs_grid[gindex[15]], iq2xs_grid[gindex[14]],
-                                                   iq2xs_grid[gindex[13]], iq2xs_grid[gindex[12]]);
-
-            const __m128i full_signs_l = lasx_extracti128(full_sign_bits, 0);
-            const __m128i full_signs_h = lasx_extracti128(full_sign_bits, 1);
-            const __m256i full_signs_1 = lasx_insertf128(full_signs_l, full_signs_l);
-            const __m256i full_signs_2 = lasx_insertf128(full_signs_h, full_signs_h);
-
-            __m256i signs;
-            signs = lasx_shuffle_b(full_signs_1, block_sign_shuffle_1);
-            signs = __lasx_xvseq_b(__lasx_xvand_v(signs, bit_selector_mask), bit_selector_mask);
-            const __m256i q8s_1 = __lasx_xvsigncov_b(__lasx_xvor_v(signs, mone), q8_1);
-
-            signs = lasx_shuffle_b(full_signs_1, block_sign_shuffle_2);
-            signs = __lasx_xvseq_b(__lasx_xvand_v(signs, bit_selector_mask), bit_selector_mask);
-            const __m256i q8s_2 = __lasx_xvsigncov_b(__lasx_xvor_v(signs, mone), q8_2);
-
-            signs = lasx_shuffle_b(full_signs_2, block_sign_shuffle_1);
-            signs = __lasx_xvseq_b(__lasx_xvand_v(signs, bit_selector_mask), bit_selector_mask);
-            const __m256i q8s_3 = __lasx_xvsigncov_b(__lasx_xvor_v(signs, mone), q8_3);
-
-            signs = lasx_shuffle_b(full_signs_2, block_sign_shuffle_2);
-            signs = __lasx_xvseq_b(__lasx_xvand_v(signs, bit_selector_mask), bit_selector_mask);
-            const __m256i q8s_4 = __lasx_xvsigncov_b(__lasx_xvor_v(signs, mone), q8_4);
-
-            const __m256i dot1  = lasx_maddubs_h(q2_1, q8s_1);
-            const __m256i dot2  = lasx_maddubs_h(q2_2, q8s_2);
-            const __m256i dot3  = lasx_maddubs_h(q2_3, q8s_3);
-            const __m256i dot4  = lasx_maddubs_h(q2_4, q8s_4);
-
-            const __m256i sc1 = lasx_ext8_16(lsx_shuffle_b(scales, get_scale_shuffle(ib32+0)));
-            const __m256i sc2 = lasx_ext8_16(lsx_shuffle_b(scales, get_scale_shuffle(ib32+1)));
-            const __m256i sc3 = lasx_ext8_16(lsx_shuffle_b(scales, get_scale_shuffle(ib32+2)));
-            const __m256i sc4 = lasx_ext8_16(lsx_shuffle_b(scales, get_scale_shuffle(ib32+3)));
-
-            sumi1 = __lasx_xvadd_w(sumi1, lasx_madd_h(dot1, sc1));
-            sumi2 = __lasx_xvadd_w(sumi2, lasx_madd_h(dot2, sc2));
-            sumi1 = __lasx_xvadd_w(sumi1, lasx_madd_h(dot3, sc3));
-            sumi2 = __lasx_xvadd_w(sumi2, lasx_madd_h(dot4, sc4));
-        }
-
-        accumf = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(d), __lasx_xvffint_s_w(__lasx_xvadd_w(sumi1, sumi2)), accumf);
-
-    }
-
-    *s = 0.125f * hsum_float_8(accumf);
-#elif defined(__POWER9_VECTOR__)
-    const vector int v0 = vec_splats((int32_t)0);
-    vector float vsumf0 = vec_splats(0.0f);
-    vector float vsumf1 = vec_splats(0.0f);
-    vector float vsumf2 = vec_splats(0.0f);
-    vector float vsumf3 = vec_splats(0.0f);
-
-    const uint64_t * signs64 = (const uint64_t *)keven_signs_q2xs;
-
-    for (int i = 0; i < nb; ++i) {
-        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d));
-        vector float vyd = vec_splats(y[i].d);
-        vector float vd = vec_mul(vxd, vyd);
-
-        vector signed int vsumi0 = v0;
-        vector signed int vsumi1 = v0;
-        vector signed int vsumi2 = v0;
-        vector signed int vsumi3 = v0;
-
-        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);
-            __builtin_prefetch(q8, 0, 1);
-
-            vector signed long long aux64x2_0 = {*(const int64_t *)(iq2xs_grid + (q2[0] & 511)), *(const int64_t *)(iq2xs_grid + (q2[1] & 511))};
-            vector signed long long aux64x2_1 = {*(const int64_t *)(iq2xs_grid + (q2[2] & 511)), *(const int64_t *)(iq2xs_grid + (q2[3] & 511))};
-            vector signed long long aux64x2_2 = {*(const int64_t *)(iq2xs_grid + (q2[4] & 511)), *(const int64_t *)(iq2xs_grid + (q2[5] & 511))};
-            vector signed long long aux64x2_3 = {*(const int64_t *)(iq2xs_grid + (q2[6] & 511)), *(const int64_t *)(iq2xs_grid + (q2[7] & 511))};
-
-            vector signed long long vsigns0 = {*(const int64_t *)(signs64 + ((q2[0] >> 9))), *(const int64_t *)(signs64 + ((q2[1] >> 9)))};
-            vector signed long long vsigns1 = {*(const int64_t *)(signs64 + ((q2[2] >> 9))), *(const int64_t *)(signs64 + ((q2[3] >> 9)))};
-            vector signed long long vsigns2 = {*(const int64_t *)(signs64 + ((q2[4] >> 9))), *(const int64_t *)(signs64 + ((q2[5] >> 9)))};
-            vector signed long long vsigns3 = {*(const int64_t *)(signs64 + ((q2[6] >> 9))), *(const int64_t *)(signs64 + ((q2[7] >> 9)))};
-            q2 += 8;
-
-            vector signed char q2x0 = (vector signed char)vec_mul((vector signed char)vsigns0, (vector signed char)aux64x2_0);
-            vector signed char q2x1 = (vector signed char)vec_mul((vector signed char)vsigns1, (vector signed char)aux64x2_1);
-            vector signed char q2x2 = (vector signed char)vec_mul((vector signed char)vsigns2, (vector signed char)aux64x2_2);
-            vector signed char q2x3 = (vector signed char)vec_mul((vector signed char)vsigns3, (vector signed char)aux64x2_3);
-
-            vector signed char q8y0 = vec_xl( 0, q8);
-            vector signed char q8y1 = vec_xl(16, q8);
-            vector signed char q8y2 = vec_xl(32, q8);
-            vector signed char q8y3 = vec_xl(48, q8);
-            q8 += 64;
-
-            vector signed short qv0 = vec_add(vec_mule(q2x0, q8y0), vec_mulo(q2x0, q8y0));
-            vector signed short qv1 = vec_add(vec_mule(q2x1, q8y1), vec_mulo(q2x1, q8y1));
-            vector signed short qv2 = vec_add(vec_mule(q2x2, q8y2), vec_mulo(q2x2, q8y2));
-            vector signed short qv3 = vec_add(vec_mule(q2x3, q8y3), vec_mulo(q2x3, q8y3));
-
-            const uint16_t ls0 = (uint16_t)(sc[0] & 0xf);
-            const uint16_t ls1 = (uint16_t)(sc[0] >>  4);
-            const uint16_t ls2 = (uint16_t)(sc[1] & 0xf);
-            const uint16_t ls3 = (uint16_t)(sc[1] >>  4);
-            sc += 2;
-
-            vector signed short vscales0 = vec_splats((int16_t)(2*ls0+1));
-            vector signed short vscales1 = vec_splats((int16_t)(2*ls1+1));
-            vector signed short vscales2 = vec_splats((int16_t)(2*ls2+1));
-            vector signed short vscales3 = vec_splats((int16_t)(2*ls3+1));
-
-            vsumi0 = vec_msum(qv0, vscales0, vsumi0);
-            vsumi1 = vec_msum(qv1, vscales1, vsumi1);
-            vsumi2 = vec_msum(qv2, vscales2, vsumi2);
-            vsumi3 = vec_msum(qv3, vscales3, vsumi3);
-        }
-
-        vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
-        vsumf1 = vec_madd(vec_ctf(vsumi1, 0), vd, vsumf1);
-        vsumf2 = vec_madd(vec_ctf(vsumi2, 0), vd, vsumf2);
-        vsumf3 = vec_madd(vec_ctf(vsumi3, 0), vd, vsumf3);
-    }
-
-    vsumf0 = vec_add(vsumf0, vsumf2);
-    vsumf1 = vec_add(vsumf1, vsumf3);
-
-    vsumf0 = vec_add(vsumf0, vsumf1);
-
-    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4));
-    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8));
-
-    *s = 0.125f * vec_extract(vsumf0, 0);
-#else
-
-    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 * 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;
-            const uint16_t ls2 = 2*(sc[ib32] >>  4) + 1;
-            int32_t sumi = 0;
-            for (int l = 0; l < 2; ++l) {
-                const uint8_t * grid = (const uint8_t *)(iq2xs_grid + (q2[l] & 511));
-                const uint8_t  signs = ksigns_iq2xs[q2[l] >> 9];
-                for (int j = 0; j < 8; ++j) {
-                    sumi += grid[j] * q8[j] * (signs & kmask_iq2xs[j] ? -1 : 1);
-                }
-                q8 += 8;
-            }
-            bsum += sumi * ls1;
-            sumi = 0;
-            for (int l = 2; l < 4; ++l) {
-                const uint8_t * grid = (const uint8_t *)(iq2xs_grid + (q2[l] & 511));
-                const uint8_t  signs = ksigns_iq2xs[q2[l] >> 9];
-                for (int j = 0; j < 8; ++j) {
-                    sumi += grid[j] * q8[j] * (signs & kmask_iq2xs[j] ? -1 : 1);
-                }
-                q8 += 8;
-            }
-            bsum += sumi * ls2;
-            q2 += 4;
-        }
-        sumf += d * bsum;
-    }
-    *s = 0.125f * sumf;
-#endif
-}
-
-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);
-    UNUSED(bx);
-    UNUSED(by);
-    UNUSED(bs);
-
-    const block_iq2_s * GGML_RESTRICT x = vx;
-    const block_q8_K  * GGML_RESTRICT y = vy;
-
-    const int nb = n / QK_K;
-
-#if defined(__ARM_NEON)
-
-   static const uint8_t k_mask1[32] = {0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01,
-                                       0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03
-   };
-
-    static const uint8_t k_mask2[16] = {0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,};
-
-    const ggml_uint8x16x2_t mask1 = ggml_vld1q_u8_x2(k_mask1);
-    const uint8x16_t        mask2 = vld1q_u8(k_mask2);
-    const uint8x16_t m1 = vdupq_n_u8(1);
-    const int32x4_t vzero = vdupq_n_s32(0);
-
-    uint8x16x2_t vs;
-    ggml_int8x16x4_t q2s;
-    ggml_int8x16x4_t q8b;
-
-    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 * 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) {
-            q8b = ggml_vld1q_s8_x4(q8); q8 += 64;
-            q2s.val[0] = vcombine_s8(vld1_s8((const int8_t *)(iq2s_grid + (qs[0] | ((qh[ib32+0] << 8) & 0x300)))),
-                                     vld1_s8((const int8_t *)(iq2s_grid + (qs[1] | ((qh[ib32+0] << 6) & 0x300)))));
-            q2s.val[1] = vcombine_s8(vld1_s8((const int8_t *)(iq2s_grid + (qs[2] | ((qh[ib32+0] << 4) & 0x300)))),
-                                     vld1_s8((const int8_t *)(iq2s_grid + (qs[3] | ((qh[ib32+0] << 2) & 0x300)))));
-            q2s.val[2] = vcombine_s8(vld1_s8((const int8_t *)(iq2s_grid + (qs[4] | ((qh[ib32+1] << 8) & 0x300)))),
-                                     vld1_s8((const int8_t *)(iq2s_grid + (qs[5] | ((qh[ib32+1] << 6) & 0x300)))));
-            q2s.val[3] = vcombine_s8(vld1_s8((const int8_t *)(iq2s_grid + (qs[6] | ((qh[ib32+1] << 4) & 0x300)))),
-                                     vld1_s8((const int8_t *)(iq2s_grid + (qs[7] | ((qh[ib32+1] << 2) & 0x300)))));
-            qs += 8;
-
-            vs.val[0] = vreinterpretq_u8_u32(vdupq_n_u32(signs[0] | ((uint32_t) signs[1] << 16)));
-            vs.val[1] = vandq_u8(ggml_vqtbl1q_u8(vs.val[0], mask1.val[1]), mask2);
-            vs.val[0] = vandq_u8(ggml_vqtbl1q_u8(vs.val[0], mask1.val[0]), mask2);
-            vs.val[0] = vceqq_u8(vs.val[0], mask2);
-            vs.val[1] = vceqq_u8(vs.val[1], mask2);
-
-            q2s.val[0] = vmulq_s8(vreinterpretq_s8_u8(vorrq_u8(vs.val[0], m1)), q2s.val[0]);
-            q2s.val[1] = vmulq_s8(vreinterpretq_s8_u8(vorrq_u8(vs.val[1], m1)), q2s.val[1]);
-
-            vs.val[0] = vreinterpretq_u8_u32(vdupq_n_u32(signs[2] | ((uint32_t) signs[3] << 16)));
-            vs.val[1] = vandq_u8(ggml_vqtbl1q_u8(vs.val[0], mask1.val[1]), mask2);
-            vs.val[0] = vandq_u8(ggml_vqtbl1q_u8(vs.val[0], mask1.val[0]), mask2);
-            vs.val[0] = vceqq_u8(vs.val[0], mask2);
-            vs.val[1] = vceqq_u8(vs.val[1], mask2);
-
-            signs += 4;
-
-            q2s.val[2] = vmulq_s8(vreinterpretq_s8_u8(vorrq_u8(vs.val[0], m1)), q2s.val[2]);
-            q2s.val[3] = vmulq_s8(vreinterpretq_s8_u8(vorrq_u8(vs.val[1], m1)), q2s.val[3]);
-
-            const int32x4_t p1 = ggml_vdotq_s32(vzero, q2s.val[0], q8b.val[0]);
-            const int32x4_t p2 = ggml_vdotq_s32(vzero, q2s.val[1], q8b.val[1]);
-            const int32x4_t p3 = ggml_vdotq_s32(vzero, q2s.val[2], q8b.val[2]);
-            const int32x4_t p4 = ggml_vdotq_s32(vzero, q2s.val[3], q8b.val[3]);
-
-            sumi1 += vaddvq_s32(p1) * (1 + 2*(x[i].scales[ib32+0] & 0xf));
-            sumi2 += vaddvq_s32(p2) * (1 + 2*(x[i].scales[ib32+0] >>  4));
-            sumi1 += vaddvq_s32(p3) * (1 + 2*(x[i].scales[ib32+1] & 0xf));
-            sumi2 += vaddvq_s32(p4) * (1 + 2*(x[i].scales[ib32+1] >>  4));
-        }
-        sumf += d*(sumi1 + sumi2);
-    }
-
-    *s = 0.125f * sumf;
-
-#elif defined(__AVX2__)
-
-   static const uint8_t k_mask1[32] = {0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01,
-                                       0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03
-   };
-
-    static const uint8_t k_mask2[32] = {0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
-                                        0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
-    };
-
-    const __m128i m4 = _mm_set1_epi8(0xf);
-    const __m128i m1 = _mm_set1_epi8(1);
-
-    const __m256i mask1 = _mm256_loadu_si256((const __m256i*)k_mask1);
-    const __m256i mask2 = _mm256_loadu_si256((const __m256i*)k_mask2);
-
-    uint64_t aux64;
-
-    __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 * 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);
-        const __m256i scales16 = _mm256_cvtepi8_epi16(scales8); // 0 2 4 6 8 10 12 14 1 3 5 7 9 11 13 15
-
-        __m256i sumi1 = _mm256_setzero_si256();
-        __m256i sumi2 = _mm256_setzero_si256();
-        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
-            const __m256i q8_1 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32;
-            const __m256i q8_2 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32;
-            const __m256i q2_1 = _mm256_set_epi64x(iq2s_grid[qs[3] | ((qh[ib32+0] << 2) & 0x300)],
-                                                   iq2s_grid[qs[2] | ((qh[ib32+0] << 4) & 0x300)],
-                                                   iq2s_grid[qs[1] | ((qh[ib32+0] << 6) & 0x300)],
-                                                   iq2s_grid[qs[0] | ((qh[ib32+0] << 8) & 0x300)]);
-            const __m256i q2_2 = _mm256_set_epi64x(iq2s_grid[qs[7] | ((qh[ib32+1] << 2) & 0x300)],
-                                                   iq2s_grid[qs[6] | ((qh[ib32+1] << 4) & 0x300)],
-                                                   iq2s_grid[qs[5] | ((qh[ib32+1] << 6) & 0x300)],
-                                                   iq2s_grid[qs[4] | ((qh[ib32+1] << 8) & 0x300)]);
-            qs += 8;
-
-            __m256i aux256 = _mm256_set1_epi32(signs[0] | ((uint32_t) signs[1] << 16));
-            aux256 = _mm256_and_si256(_mm256_shuffle_epi8(aux256,mask1), mask2);
-            const __m256i s2_1 = _mm256_cmpeq_epi8(aux256, mask2);
-            const __m256i q8s_1 = _mm256_sub_epi8(_mm256_xor_si256(s2_1, q8_1), s2_1);
-
-            aux256 = _mm256_set1_epi32(signs[2] | ((uint32_t) signs[3] << 16));
-            aux256 = _mm256_and_si256(_mm256_shuffle_epi8(aux256,mask1), mask2);
-            const __m256i s2_2 = _mm256_cmpeq_epi8(aux256, mask2);
-            const __m256i q8s_2 = _mm256_sub_epi8(_mm256_xor_si256(s2_2, q8_2), s2_2);
-
-            signs += 4;
-
-            const __m256i dot1  = _mm256_maddubs_epi16(q2_1, q8s_1); // blocks 2*ib32+0, 2*ib32+1
-            const __m256i dot2  = _mm256_maddubs_epi16(q2_2, q8s_2); // blocks 2*ib32+2, 2*ib32+3
-
-            const __m256i p1 = _mm256_madd_epi16(dot1, _mm256_shuffle_epi8(scales16, get_scale_shuffle_k4(ib32+0)));
-            const __m256i p2 = _mm256_madd_epi16(dot2, _mm256_shuffle_epi8(scales16, get_scale_shuffle_k4(ib32+1)));
-            sumi1 = _mm256_add_epi32(sumi1, p1);
-            sumi2 = _mm256_add_epi32(sumi2, p2);
-        }
-
-        accumf = _mm256_fmadd_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(_mm256_add_epi32(sumi1, sumi2)), accumf);
-
-    }
-
-    *s = 0.125f * hsum_float_8(accumf);
-
-#elif defined(__AVX__)
-   static const uint8_t k_mask1[32] = {0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01,
-                                       0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03
-   };
-
-    static const uint8_t k_mask2[32] = {0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
-                                        0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
-    };
-
-    const __m128i m4 = _mm_set1_epi8(0xf);
-    const __m128i m1 = _mm_set1_epi8(1);
-
-    const __m128i mask1_0 = _mm_loadu_si128((const __m128i*)k_mask1);
-    const __m128i mask1_1 = _mm_loadu_si128((const __m128i*)k_mask1 + 1);
-    const __m128i mask2_0 = _mm_loadu_si128((const __m128i*)k_mask2);
-    const __m128i mask2_1 = _mm_loadu_si128((const __m128i*)k_mask2 + 1);
-
-    uint64_t aux64;
-
-    __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 * 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);
-        const __m128i scales16_0 = _mm_cvtepi8_epi16(scales8);
-        const __m128i scales16_1 = _mm_cvtepi8_epi16(_mm_srli_si128(scales8, 8));
-
-        __m128i sumi1_0 = _mm_setzero_si128();
-        __m128i sumi1_1 = _mm_setzero_si128();
-        __m128i sumi2_0 = _mm_setzero_si128();
-        __m128i sumi2_1 = _mm_setzero_si128();
-        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
-            const __m128i q8_1_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
-            const __m128i q8_1_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
-            const __m128i q8_2_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
-            const __m128i q8_2_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
-            const __m128i q2_1_0 = _mm_set_epi64x(iq2s_grid[qs[1] | ((qh[ib32+0] << 6) & 0x300)],
-                                                  iq2s_grid[qs[0] | ((qh[ib32+0] << 8) & 0x300)]);
-            const __m128i q2_1_1 = _mm_set_epi64x(iq2s_grid[qs[3] | ((qh[ib32+0] << 2) & 0x300)],
-                                                  iq2s_grid[qs[2] | ((qh[ib32+0] << 4) & 0x300)]);
-            const __m128i q2_2_0 = _mm_set_epi64x(iq2s_grid[qs[5] | ((qh[ib32+1] << 6) & 0x300)],
-                                                  iq2s_grid[qs[4] | ((qh[ib32+1] << 8) & 0x300)]);
-            const __m128i q2_2_1 = _mm_set_epi64x(iq2s_grid[qs[7] | ((qh[ib32+1] << 2) & 0x300)],
-                                                  iq2s_grid[qs[6] | ((qh[ib32+1] << 4) & 0x300)]);
-            qs += 8;
-
-            __m128i aux128_0 = _mm_set1_epi32(signs[0] | ((uint32_t) signs[1] << 16));
-            __m128i aux128_1 = aux128_0;
-            aux128_0 = _mm_and_si128(_mm_shuffle_epi8(aux128_0,mask1_0), mask2_0);
-            aux128_1 = _mm_and_si128(_mm_shuffle_epi8(aux128_1,mask1_1), mask2_1);
-            const __m128i s2_1_0 = _mm_cmpeq_epi8(aux128_0, mask2_0);
-            const __m128i s2_1_1 = _mm_cmpeq_epi8(aux128_1, mask2_1);
-            const __m128i q8s_1_0 = _mm_sub_epi8(_mm_xor_si128(s2_1_0, q8_1_0), s2_1_0);
-            const __m128i q8s_1_1 = _mm_sub_epi8(_mm_xor_si128(s2_1_1, q8_1_1), s2_1_1);
-
-            aux128_0 = _mm_set1_epi32(signs[2] | ((uint32_t) signs[3] << 16));
-            aux128_1 = aux128_0;
-            aux128_0 = _mm_and_si128(_mm_shuffle_epi8(aux128_0,mask1_0), mask2_0);
-            aux128_1 = _mm_and_si128(_mm_shuffle_epi8(aux128_1,mask1_1), mask2_1);
-            const __m128i s2_2_0 = _mm_cmpeq_epi8(aux128_0, mask2_0);
-            const __m128i s2_2_1 = _mm_cmpeq_epi8(aux128_1, mask2_1);
-            const __m128i q8s_2_0 = _mm_sub_epi8(_mm_xor_si128(s2_2_0, q8_2_0), s2_2_0);
-            const __m128i q8s_2_1 = _mm_sub_epi8(_mm_xor_si128(s2_2_1, q8_2_1), s2_2_1);
-
-            signs += 4;
-
-            const __m128i dot1_0  = _mm_maddubs_epi16(q2_1_0, q8s_1_0);
-            const __m128i dot1_1  = _mm_maddubs_epi16(q2_1_1, q8s_1_1);
-            const __m128i dot2_0  = _mm_maddubs_epi16(q2_2_0, q8s_2_0);
-            const __m128i dot2_1  = _mm_maddubs_epi16(q2_2_1, q8s_2_1);
-
-            const __m128i p1_0 = _mm_madd_epi16(dot1_0, _mm_shuffle_epi8(scales16_0, _mm256_extractf128_si256(get_scale_shuffle_k4(ib32+0), 0)));
-            const __m128i p1_1 = _mm_madd_epi16(dot1_1, _mm_shuffle_epi8(scales16_1, _mm256_extractf128_si256(get_scale_shuffle_k4(ib32+0), 1)));
-            const __m128i p2_0 = _mm_madd_epi16(dot2_0, _mm_shuffle_epi8(scales16_0, _mm256_extractf128_si256(get_scale_shuffle_k4(ib32+1), 0)));
-            const __m128i p2_1 = _mm_madd_epi16(dot2_1, _mm_shuffle_epi8(scales16_1, _mm256_extractf128_si256(get_scale_shuffle_k4(ib32+1), 1)));
-            sumi1_0 = _mm_add_epi32(sumi1_0, p1_0);
-            sumi1_1 = _mm_add_epi32(sumi1_1, p1_1);
-            sumi2_0 = _mm_add_epi32(sumi2_0, p2_0);
-            sumi2_1 = _mm_add_epi32(sumi2_1, p2_1);
-        }
-
-        accumf = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(MM256_SET_M128I(_mm_add_epi32(sumi1_1, sumi2_1), _mm_add_epi32(sumi1_0, sumi2_0)))), accumf);
-
-    }
-
-    *s = 0.125f * hsum_float_8(accumf);
-
-#elif defined(__POWER9_VECTOR__)
-    static const uint8_t k_mask1[32] = {0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01,
-                                        0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03
-    };
-
-    static const uint8_t k_mask2[16] = {0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,};
-
-    const vector int v0 = vec_splats((int32_t)0);
-
-    vector float vsumf0 = vec_splats(0.0f);
-    vector float vsumf1 = vec_splats(0.0f);
-    vector float vsumf2 = vec_splats(0.0f);
-    vector float vsumf3 = vec_splats(0.0f);
-
-    const vector unsigned char mask0 = vec_xl( 0, k_mask1);
-    const vector unsigned char mask1 = vec_xl(16, k_mask1);
-    const vector signed char mask2 = (vector signed char)vec_xl( 0, k_mask2);
-
-    for (int i = 0; i < nb; ++i) {
-        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d));
-        vector float vyd = vec_splats(y[i].d);
-        vector float vd = vec_mul(vxd, vyd);
-
-        vector signed int vsumi0 = v0;
-        vector signed int vsumi1 = v0;
-        vector signed int vsumi2 = v0;
-        vector signed int vsumi3 = v0;
-
-        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);
-            __builtin_prefetch(q8, 0, 1);
-
-            vector signed long long aux64x2_0 = {*(const int64_t *)(iq2s_grid + (q2[0] | ((qh[0] << 8) & 0x300))), *(const int64_t *)(iq2s_grid + (q2[1] | ((qh[0] << 6) & 0x300)))};
-            vector signed long long aux64x2_1 = {*(const int64_t *)(iq2s_grid + (q2[2] | ((qh[0] << 4) & 0x300))), *(const int64_t *)(iq2s_grid + (q2[3] | ((qh[0] << 2) & 0x300)))};
-            vector signed long long aux64x2_2 = {*(const int64_t *)(iq2s_grid + (q2[4] | ((qh[1] << 8) & 0x300))), *(const int64_t *)(iq2s_grid + (q2[5] | ((qh[1] << 6) & 0x300)))};
-            vector signed long long aux64x2_3 = {*(const int64_t *)(iq2s_grid + (q2[6] | ((qh[1] << 4) & 0x300))), *(const int64_t *)(iq2s_grid + (q2[7] | ((qh[1] << 2) & 0x300)))};
-            q2 += 8;
-            qh += 2;
-
-            vector signed char vsigns01 = (vector signed char)vec_splats(*(const uint32_t *)&signs[0]);
-            vector signed char vsigns23 = (vector signed char)vec_splats(*(const uint32_t *)&signs[2]);
-            signs += 4;
-
-            vector signed char vsigns0 = vec_perm(vsigns01, vsigns01, mask0);
-            vector signed char vsigns1 = vec_perm(vsigns01, vsigns01, mask1);
-            vector signed char vsigns2 = vec_perm(vsigns23, vsigns23, mask0);
-            vector signed char vsigns3 = vec_perm(vsigns23, vsigns23, mask1);
-
-            vsigns0 = (vector signed char)vec_cmpeq(vec_and(vsigns0, mask2), mask2);
-            vsigns1 = (vector signed char)vec_cmpeq(vec_and(vsigns1, mask2), mask2);
-            vsigns2 = (vector signed char)vec_cmpeq(vec_and(vsigns2, mask2), mask2);
-            vsigns3 = (vector signed char)vec_cmpeq(vec_and(vsigns3, mask2), mask2);
-
-            vector signed char q2x0 = vec_sub(vec_xor(vsigns0, (vector signed char)aux64x2_0), vsigns0);
-            vector signed char q2x1 = vec_sub(vec_xor(vsigns1, (vector signed char)aux64x2_1), vsigns1);
-            vector signed char q2x2 = vec_sub(vec_xor(vsigns2, (vector signed char)aux64x2_2), vsigns2);
-            vector signed char q2x3 = vec_sub(vec_xor(vsigns3, (vector signed char)aux64x2_3), vsigns3);
-
-            vector signed char q8y0 = vec_xl( 0, q8);
-            vector signed char q8y1 = vec_xl(16, q8);
-            vector signed char q8y2 = vec_xl(32, q8);
-            vector signed char q8y3 = vec_xl(48, q8);
-            q8 += 64;
-
-            vector signed short qv0 = vec_add(vec_mule(q2x0, q8y0), vec_mulo(q2x0, q8y0));
-            vector signed short qv1 = vec_add(vec_mule(q2x1, q8y1), vec_mulo(q2x1, q8y1));
-            vector signed short qv2 = vec_add(vec_mule(q2x2, q8y2), vec_mulo(q2x2, q8y2));
-            vector signed short qv3 = vec_add(vec_mule(q2x3, q8y3), vec_mulo(q2x3, q8y3));
-
-            const uint16_t ls0 = (uint16_t)(sc[0] & 0xf);
-            const uint16_t ls1 = (uint16_t)(sc[0] >>  4);
-            const uint16_t ls2 = (uint16_t)(sc[1] & 0xf);
-            const uint16_t ls3 = (uint16_t)(sc[1] >>  4);
-            sc += 2;
-
-            vector signed short vscales0 = vec_splats((int16_t)(2*ls0+1));
-            vector signed short vscales1 = vec_splats((int16_t)(2*ls1+1));
-            vector signed short vscales2 = vec_splats((int16_t)(2*ls2+1));
-            vector signed short vscales3 = vec_splats((int16_t)(2*ls3+1));
-
-            vsumi0 = vec_msum(qv0, vscales0, vsumi0);
-            vsumi1 = vec_msum(qv1, vscales1, vsumi1);
-            vsumi2 = vec_msum(qv2, vscales2, vsumi2);
-            vsumi3 = vec_msum(qv3, vscales3, vsumi3);
-        }
-
-        vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
-        vsumf1 = vec_madd(vec_ctf(vsumi1, 0), vd, vsumf1);
-        vsumf2 = vec_madd(vec_ctf(vsumi2, 0), vd, vsumf2);
-        vsumf3 = vec_madd(vec_ctf(vsumi3, 0), vd, vsumf3);
-    }
-
-    vsumf0 = vec_add(vsumf0, vsumf2);
-    vsumf1 = vec_add(vsumf1, vsumf3);
-
-    vsumf0 = vec_add(vsumf0, vsumf1);
-
-    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4));
-    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8));
-
-    *s = 0.125f * vec_extract(vsumf0, 0);
-
-#elif defined(__loongarch_asx)
-
-   static const uint8_t k_mask1[32] = {0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01,
-                                       0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03
-   };
-
-    static const uint8_t k_mask2[32] = {0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
-                                        0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
-    };
-
-
-    const __m128i m4 = __lsx_vreplgr2vr_b(0xf);
-    const __m128i m1 = __lsx_vreplgr2vr_b(1);
-
-    const __m256i mask1 = __lasx_xvld((const __m256i*)k_mask1, 0);
-    const __m256i mask2 = __lasx_xvld((const __m256i*)k_mask2, 0);
-    uint64_t aux64;
-
-    __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 * 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);
-        tmp1 = __lsx_vinsgr2vr_d(tmp1, aux64, 0);
-        tmp1 = __lsx_vinsgr2vr_d(tmp1, aux64 >> 4, 1);
-        const __m128i scales8 = __lsx_vadd_b(__lsx_vslli_h(__lsx_vand_v(tmp1, m4), 1), m1);
-        const __m256i scales16 = lasx_ext8_16(scales8); // 0 2 4 6 8 10 12 14 1 3 5 7 9 11 13 15
-
-        __m256i sumi1 = __lasx_xvldi(0);
-        __m256i sumi2 = __lasx_xvldi(0);
-        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
-            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 q2_1 = lasx_set_d(iq2s_grid[qs[3] | ((qh[ib32+0] << 2) & 0x300)],
-                                                   iq2s_grid[qs[2] | ((qh[ib32+0] << 4) & 0x300)],
-                                                   iq2s_grid[qs[1] | ((qh[ib32+0] << 6) & 0x300)],
-                                                   iq2s_grid[qs[0] | ((qh[ib32+0] << 8) & 0x300)]);
-            const __m256i q2_2 = lasx_set_d(iq2s_grid[qs[7] | ((qh[ib32+1] << 2) & 0x300)],
-                                                   iq2s_grid[qs[6] | ((qh[ib32+1] << 4) & 0x300)],
-                                                   iq2s_grid[qs[5] | ((qh[ib32+1] << 6) & 0x300)],
-                                                   iq2s_grid[qs[4] | ((qh[ib32+1] << 8) & 0x300)]);
-            qs += 8;
-
-            __m256i aux256 = __lasx_xvreplgr2vr_w(signs[0] | ((uint32_t) signs[1] << 16));
-            aux256 = __lasx_xvand_v(lasx_shuffle_b(aux256,mask1), mask2);
-            const __m256i s2_1 = __lasx_xvseq_b(aux256, mask2);
-            const __m256i q8s_1 = __lasx_xvsub_b(__lasx_xvxor_v(s2_1, q8_1), s2_1);
-
-            aux256 = __lasx_xvreplgr2vr_w(signs[2] | ((uint32_t) signs[3] << 16));
-            aux256 = __lasx_xvand_v(lasx_shuffle_b(aux256,mask1), mask2);
-            const __m256i s2_2 = __lasx_xvseq_b(aux256, mask2);
-            const __m256i q8s_2 = __lasx_xvsub_b(__lasx_xvxor_v(s2_2, q8_2), s2_2);
-
-            signs += 4;
-
-            const __m256i dot1  = lasx_maddubs_h(q2_1, q8s_1); // blocks 2*ib32+0, 2*ib32+1
-            const __m256i dot2  = lasx_maddubs_h(q2_2, q8s_2); // blocks 2*ib32+2, 2*ib32+3
-
-            const __m256i p1 = lasx_madd_h(dot1, lasx_shuffle_b(scales16, get_scale_shuffle_k4(ib32+0)));
-            const __m256i p2 = lasx_madd_h(dot2, lasx_shuffle_b(scales16, get_scale_shuffle_k4(ib32+1)));
-            sumi1 = __lasx_xvadd_w(sumi1, p1);
-            sumi2 = __lasx_xvadd_w(sumi2, p2);
-        }
-
-        accumf = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(d), __lasx_xvffint_s_w(__lasx_xvadd_w(sumi1, sumi2)), accumf);
-    }
-
-    *s = 0.125f * hsum_float_8(accumf);
-
-#else
-
-    float sumf = 0;
-    for (int i = 0; i < nb; i++) {
-
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
-        const int8_t  * q8 = y[i].qs;
-        const uint8_t * qs = x[i].qs;
-        const uint8_t * qh = x[i].qh;
-        const uint8_t * signs = qs + QK_K/8;
-
-        int bsum = 0;
-        for (int ib32 = 0; ib32 < QK_K/32; ++ib32) {
-            int ls1 = 1 + 2*(x[i].scales[ib32] & 0xf);
-            int ls2 = 1 + 2*(x[i].scales[ib32] >>  4);
-            int sumi1 = 0, sumi2 = 0;
-            for (int l = 0; l < 2; ++l) {
-                const uint8_t * grid = (const uint8_t *)(iq2s_grid + (qs[l] | (qh[ib32] << (8-2*l) & 0x300)));
-                for (int j = 0; j < 8; ++j) {
-                    sumi1 += q8[j] * grid[j] * (signs[l] & kmask_iq2xs[j] ? -1 : 1);
-                }
-                q8 += 8;
-            }
-            for (int l = 2; l < 4; ++l) {
-                const uint8_t * grid = (const uint8_t *)(iq2s_grid + (qs[l] | (qh[ib32] << (8-2*l) & 0x300)));
-                for (int j = 0; j < 8; ++j) {
-                    sumi2 += q8[j] * grid[j] * (signs[l] & kmask_iq2xs[j] ? -1 : 1);
-                }
-                q8 += 8;
-            }
-            bsum += ls1 * sumi1 + ls2 * sumi2;
-            qs += 4;
-            signs += 4;
-        }
-
-        sumf += d * bsum;
-    }
-
-    *s = 0.125f * sumf;
-
-#endif
-
-}
-
-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);
-    UNUSED(bx);
-    UNUSED(by);
-    UNUSED(bs);
-
-    const block_iq3_xxs * GGML_RESTRICT x = vx;
-    const block_q8_K    * GGML_RESTRICT y = vy;
-
-    const int nb = n / QK_K;
-
-#if defined(__ARM_NEON)
-
-    const uint64_t * signs64 = (const uint64_t *)keven_signs_q2xs;
-
-    uint32_t aux32[2];
-
-    ggml_int8x16x4_t q3s;
-    ggml_int8x16x4_t q8b;
-
-    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 * 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;
-            memcpy(aux32, gas, 2*sizeof(uint32_t)); gas += 2*sizeof(uint32_t);
-            const uint32x4_t aux32x4_0 = ggml_vld1q_u32(iq3xxs_grid[q3[ 0]], iq3xxs_grid[q3[ 1]], iq3xxs_grid[q3[ 2]], iq3xxs_grid[q3[ 3]]);
-            const uint32x4_t aux32x4_1 = ggml_vld1q_u32(iq3xxs_grid[q3[ 4]], iq3xxs_grid[q3[ 5]], iq3xxs_grid[q3[ 6]], iq3xxs_grid[q3[ 7]]);
-            const uint32x4_t aux32x4_2 = ggml_vld1q_u32(iq3xxs_grid[q3[ 8]], iq3xxs_grid[q3[ 9]], iq3xxs_grid[q3[10]], iq3xxs_grid[q3[11]]);
-            const uint32x4_t aux32x4_3 = ggml_vld1q_u32(iq3xxs_grid[q3[12]], iq3xxs_grid[q3[13]], iq3xxs_grid[q3[14]], iq3xxs_grid[q3[15]]);
-            q3 += 16;
-            q3s.val[0] = vcombine_s8(vld1_s8((const void *)(signs64 + ((aux32[0] >>  0) & 127))), vld1_s8((const void *)(signs64 + ((aux32[0] >>  7) & 127))));
-            q3s.val[1] = vcombine_s8(vld1_s8((const void *)(signs64 + ((aux32[0] >> 14) & 127))), vld1_s8((const void *)(signs64 + ((aux32[0] >> 21) & 127))));
-            q3s.val[2] = vcombine_s8(vld1_s8((const void *)(signs64 + ((aux32[1] >>  0) & 127))), vld1_s8((const void *)(signs64 + ((aux32[1] >>  7) & 127))));
-            q3s.val[3] = vcombine_s8(vld1_s8((const void *)(signs64 + ((aux32[1] >> 14) & 127))), vld1_s8((const void *)(signs64 + ((aux32[1] >> 21) & 127))));
-            q3s.val[0] = vmulq_s8(q3s.val[0], vreinterpretq_s8_u32(aux32x4_0));
-            q3s.val[1] = vmulq_s8(q3s.val[1], vreinterpretq_s8_u32(aux32x4_1));
-            q3s.val[2] = vmulq_s8(q3s.val[2], vreinterpretq_s8_u32(aux32x4_2));
-            q3s.val[3] = vmulq_s8(q3s.val[3], vreinterpretq_s8_u32(aux32x4_3));
-            const int32x4_t p1 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), q3s.val[0], q8b.val[0]), q3s.val[1], q8b.val[1]);
-            const int32x4_t p2 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), q3s.val[2], q8b.val[2]), q3s.val[3], q8b.val[3]);
-            sumf1 += vaddvq_s32(p1) * (0.5f + (aux32[0] >> 28));
-            sumf2 += vaddvq_s32(p2) * (0.5f + (aux32[1] >> 28));
-        }
-        sumf += d*(sumf1 + sumf2);
-    }
-    *s = 0.5f * sumf;
-
-#elif defined(__AVX2__)
-
-    const uint64_t * signs64 = (const uint64_t *)keven_signs_q2xs;
-
-    uint32_t aux32[2];
-
-    __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 * 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) {
-            const __m256i q8_1 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32;
-            const __m256i q8_2 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32;
-            const __m256i q2_1 = _mm256_set_epi32(iq3xxs_grid[q3[7]], iq3xxs_grid[q3[6]], iq3xxs_grid[q3[5]], iq3xxs_grid[q3[4]],
-                                                  iq3xxs_grid[q3[3]], iq3xxs_grid[q3[2]], iq3xxs_grid[q3[1]], iq3xxs_grid[q3[0]]);
-            q3 += 8;
-            const __m256i q2_2 = _mm256_set_epi32(iq3xxs_grid[q3[7]], iq3xxs_grid[q3[6]], iq3xxs_grid[q3[5]], iq3xxs_grid[q3[4]],
-                                                  iq3xxs_grid[q3[3]], iq3xxs_grid[q3[2]], iq3xxs_grid[q3[1]], iq3xxs_grid[q3[0]]);
-            q3 += 8;
-            memcpy(aux32, gas, 8); gas += 8;
-            const __m256i s2_1 = _mm256_set_epi64x(signs64[(aux32[0] >> 21) & 127], signs64[(aux32[0] >> 14) & 127],
-                                                   signs64[(aux32[0] >>  7) & 127], signs64[(aux32[0] >>  0) & 127]);
-            const __m256i s2_2 = _mm256_set_epi64x(signs64[(aux32[1] >> 21) & 127], signs64[(aux32[1] >> 14) & 127],
-                                                   signs64[(aux32[1] >>  7) & 127], signs64[(aux32[1] >>  0) & 127]);
-            const __m256i q8s_1 = _mm256_sign_epi8(q8_1, s2_1);
-            const __m256i q8s_2 = _mm256_sign_epi8(q8_2, s2_2);
-            const __m256i dot1  = _mm256_maddubs_epi16(q2_1, q8s_1);
-            const __m256i dot2  = _mm256_maddubs_epi16(q2_2, q8s_2);
-            const uint16_t ls1 = aux32[0] >> 28;
-            const uint16_t ls2 = aux32[1] >> 28;
-            const __m256i p1 = _mm256_madd_epi16(dot1, _mm256_set1_epi16(2*ls1+1));
-            const __m256i p2 = _mm256_madd_epi16(dot2, _mm256_set1_epi16(2*ls2+1));
-            sumi1 = _mm256_add_epi32(sumi1, p1);
-            sumi2 = _mm256_add_epi32(sumi2, p2);
-        }
-
-        accumf = _mm256_fmadd_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(_mm256_add_epi32(sumi1, sumi2)), accumf);
-
-    }
-
-    *s = 0.25f * hsum_float_8(accumf);
-
-#elif defined(__AVX__)
-    const uint64_t * signs64 = (const uint64_t *)keven_signs_q2xs;
-
-    uint32_t aux32[2];
-
-    __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 * 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();
-        __m128i sumi2_1 = _mm_setzero_si128();
-        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
-            const __m128i q8_1_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
-            const __m128i q8_1_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
-            const __m128i q8_2_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
-            const __m128i q8_2_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
-            const __m128i q2_1_0 = _mm_set_epi32(iq3xxs_grid[q3[3]], iq3xxs_grid[q3[2]], iq3xxs_grid[q3[1]], iq3xxs_grid[q3[0]]);
-            const __m128i q2_1_1 = _mm_set_epi32(iq3xxs_grid[q3[7]], iq3xxs_grid[q3[6]], iq3xxs_grid[q3[5]], iq3xxs_grid[q3[4]]);
-            q3 += 8;
-            const __m128i q2_2_0 = _mm_set_epi32(iq3xxs_grid[q3[3]], iq3xxs_grid[q3[2]], iq3xxs_grid[q3[1]], iq3xxs_grid[q3[0]]);
-            const __m128i q2_2_1 = _mm_set_epi32(iq3xxs_grid[q3[7]], iq3xxs_grid[q3[6]], iq3xxs_grid[q3[5]], iq3xxs_grid[q3[4]]);
-            q3 += 8;
-            memcpy(aux32, gas, 8); gas += 8;
-            const __m128i s2_1_0 = _mm_set_epi64x(signs64[(aux32[0] >>  7) & 127], signs64[(aux32[0] >>  0) & 127]);
-            const __m128i s2_1_1 = _mm_set_epi64x(signs64[(aux32[0] >> 21) & 127], signs64[(aux32[0] >> 14) & 127]);
-            const __m128i s2_2_0 = _mm_set_epi64x(signs64[(aux32[1] >>  7) & 127], signs64[(aux32[1] >>  0) & 127]);
-            const __m128i s2_2_1 = _mm_set_epi64x(signs64[(aux32[1] >> 21) & 127], signs64[(aux32[1] >> 14) & 127]);
-            const __m128i q8s_1_0 = _mm_sign_epi8(q8_1_0, s2_1_0);
-            const __m128i q8s_1_1 = _mm_sign_epi8(q8_1_1, s2_1_1);
-            const __m128i q8s_2_0 = _mm_sign_epi8(q8_2_0, s2_2_0);
-            const __m128i q8s_2_1 = _mm_sign_epi8(q8_2_1, s2_2_1);
-            const __m128i dot1_0  = _mm_maddubs_epi16(q2_1_0, q8s_1_0);
-            const __m128i dot1_1  = _mm_maddubs_epi16(q2_1_1, q8s_1_1);
-            const __m128i dot2_0  = _mm_maddubs_epi16(q2_2_0, q8s_2_0);
-            const __m128i dot2_1  = _mm_maddubs_epi16(q2_2_1, q8s_2_1);
-            const uint16_t ls1 = aux32[0] >> 28;
-            const uint16_t ls2 = aux32[1] >> 28;
-            const __m128i p1_0 = _mm_madd_epi16(dot1_0, _mm_set1_epi16(2*ls1+1));
-            const __m128i p1_1 = _mm_madd_epi16(dot1_1, _mm_set1_epi16(2*ls1+1));
-            const __m128i p2_0 = _mm_madd_epi16(dot2_0, _mm_set1_epi16(2*ls2+1));
-            const __m128i p2_1 = _mm_madd_epi16(dot2_1, _mm_set1_epi16(2*ls2+1));
-            sumi1_0 = _mm_add_epi32(sumi1_0, p1_0);
-            sumi1_1 = _mm_add_epi32(sumi1_1, p1_1);
-            sumi2_0 = _mm_add_epi32(sumi2_0, p2_0);
-            sumi2_1 = _mm_add_epi32(sumi2_1, p2_1);
-        }
-
-        accumf = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(MM256_SET_M128I(_mm_add_epi32(sumi1_1, sumi2_1), _mm_add_epi32(sumi1_0, sumi2_0)))), accumf);
-
-    }
-
-    *s = 0.25f * hsum_float_8(accumf);
-
-#elif defined(__POWER9_VECTOR__)
-    const uint64_t * signs64 = (const uint64_t *)keven_signs_q2xs;
-
-    const vector int v0 = vec_splats((int32_t)0);
-
-    vector float vsumf0 = vec_splats(0.0f);
-    vector float vsumf1 = vec_splats(0.0f);
-    vector float vsumf2 = vec_splats(0.0f);
-    vector float vsumf3 = vec_splats(0.0f);
-
-    for (int i = 0; i < nb; ++i) {
-        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d));
-        vector float vyd = vec_splats(y[i].d);
-        vector float vd = vec_mul(vxd, vyd);
-
-        vector signed int vsumi0 = v0;
-        vector signed int vsumi1 = v0;
-        vector signed int vsumi2 = v0;
-        vector signed int vsumi3 = v0;
-
-        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) {
-            __builtin_prefetch(q3, 0, 1);
-            __builtin_prefetch(q8, 0, 1);
-
-            vector unsigned int aux32x4_0 = {iq3xxs_grid[q3[ 0]], iq3xxs_grid[q3[ 1]], iq3xxs_grid[q3[ 2]], iq3xxs_grid[q3[ 3]]};
-            vector unsigned int aux32x4_1 = {iq3xxs_grid[q3[ 4]], iq3xxs_grid[q3[ 5]], iq3xxs_grid[q3[ 6]], iq3xxs_grid[q3[ 7]]};
-            vector unsigned int aux32x4_2 = {iq3xxs_grid[q3[ 8]], iq3xxs_grid[q3[ 9]], iq3xxs_grid[q3[10]], iq3xxs_grid[q3[11]]};
-            vector unsigned int aux32x4_3 = {iq3xxs_grid[q3[12]], iq3xxs_grid[q3[13]], iq3xxs_grid[q3[14]], iq3xxs_grid[q3[15]]};
-            q3 += 16;
-
-            vector unsigned long long aux64x2_0 = {(uint64_t)(signs64[(signs[0] >>  0) & 127]), (uint64_t)(signs64[(signs[0] >>  7) & 127])};
-            vector unsigned long long aux64x2_1 = {(uint64_t)(signs64[(signs[0] >> 14) & 127]), (uint64_t)(signs64[(signs[0] >> 21) & 127])};
-            vector unsigned long long aux64x2_2 = {(uint64_t)(signs64[(signs[1] >>  0) & 127]), (uint64_t)(signs64[(signs[1] >>  7) & 127])};
-            vector unsigned long long aux64x2_3 = {(uint64_t)(signs64[(signs[1] >> 14) & 127]), (uint64_t)(signs64[(signs[1] >> 21) & 127])};
-
-            vector signed char q3x0 = vec_mul((vector signed char)aux64x2_0, (vector signed char)aux32x4_0);
-            vector signed char q3x1 = vec_mul((vector signed char)aux64x2_1, (vector signed char)aux32x4_1);
-            vector signed char q3x2 = vec_mul((vector signed char)aux64x2_2, (vector signed char)aux32x4_2);
-            vector signed char q3x3 = vec_mul((vector signed char)aux64x2_3, (vector signed char)aux32x4_3);
-
-            vector signed char q8y0 = vec_xl( 0, q8);
-            vector signed char q8y1 = vec_xl(16, q8);
-            vector signed char q8y2 = vec_xl(32, q8);
-            vector signed char q8y3 = vec_xl(48, q8);
-            q8 += 64;
-
-            vector signed short qv0 = vec_add(vec_mule(q3x0, q8y0), vec_mulo(q3x0, q8y0));
-            vector signed short qv1 = vec_add(vec_mule(q3x1, q8y1), vec_mulo(q3x1, q8y1));
-            vector signed short qv2 = vec_add(vec_mule(q3x2, q8y2), vec_mulo(q3x2, q8y2));
-            vector signed short qv3 = vec_add(vec_mule(q3x3, q8y3), vec_mulo(q3x3, q8y3));
-
-            const uint16_t ls0 = (uint16_t)(signs[0] >> 28);
-            const uint16_t ls1 = (uint16_t)(signs[1] >> 28);
-            signs += 2;
-
-            vector signed short vscales01 = (vector signed short)vec_splats((uint16_t)(2*ls0+1));
-            vector signed short vscales23 = (vector signed short)vec_splats((uint16_t)(2*ls1+1));
-
-            vsumi0 = vec_msum(qv0, vscales01, vsumi0);
-            vsumi1 = vec_msum(qv1, vscales01, vsumi1);
-            vsumi2 = vec_msum(qv2, vscales23, vsumi2);
-            vsumi3 = vec_msum(qv3, vscales23, vsumi3);
-        }
-
-        vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
-        vsumf1 = vec_madd(vec_ctf(vsumi1, 0), vd, vsumf1);
-        vsumf2 = vec_madd(vec_ctf(vsumi2, 0), vd, vsumf2);
-        vsumf3 = vec_madd(vec_ctf(vsumi3, 0), vd, vsumf3);
-    }
-
-    vsumf0 = vec_add(vsumf0, vsumf2);
-    vsumf1 = vec_add(vsumf1, vsumf3);
-
-    vsumf0 = vec_add(vsumf0, vsumf1);
-
-    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4));
-    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8));
-
-    *s = 0.25f * vec_extract(vsumf0, 0);
-
-#elif defined(__loongarch_asx)
-
-    const uint64_t * signs64 = (const uint64_t *)keven_signs_q2xs;
-
-    uint32_t aux32[2];
-
-    __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 * 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) {
-            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 q2_1 = lasx_set_w(iq3xxs_grid[q3[7]], iq3xxs_grid[q3[6]], iq3xxs_grid[q3[5]], iq3xxs_grid[q3[4]],
-                                                iq3xxs_grid[q3[3]], iq3xxs_grid[q3[2]], iq3xxs_grid[q3[1]], iq3xxs_grid[q3[0]]);
-            q3 += 8;
-            const __m256i q2_2 = lasx_set_w(iq3xxs_grid[q3[7]], iq3xxs_grid[q3[6]], iq3xxs_grid[q3[5]], iq3xxs_grid[q3[4]],
-                                                iq3xxs_grid[q3[3]], iq3xxs_grid[q3[2]], iq3xxs_grid[q3[1]], iq3xxs_grid[q3[0]]);
-            q3 += 8;
-            memcpy(aux32, gas, 8); gas += 8;
-
-            const __m256i s2_1 = lasx_set_d(signs64[(aux32[0] >> 21) & 127], signs64[(aux32[0] >> 14) & 127],
-                                                   signs64[(aux32[0] >>  7) & 127], signs64[(aux32[0] >>  0) & 127]);
-            const __m256i s2_2 = lasx_set_d(signs64[(aux32[1] >> 21) & 127], signs64[(aux32[1] >> 14) & 127],
-                                                   signs64[(aux32[1] >>  7) & 127], signs64[(aux32[1] >>  0) & 127]);
-            const __m256i q8s_1 = __lasx_xvsigncov_b(s2_1, q8_1);
-            const __m256i q8s_2 = __lasx_xvsigncov_b(s2_2, q8_2);
-            const __m256i dot1  = lasx_maddubs_h(q2_1, q8s_1);
-            const __m256i dot2  = lasx_maddubs_h(q2_2, q8s_2);
-            const uint16_t ls1 = aux32[0] >> 28;
-            const uint16_t ls2 = aux32[1] >> 28;
-
-            const __m256i p1 = lasx_madd_h(dot1, __lasx_xvreplgr2vr_h(2*ls1+1));
-            const __m256i p2 = lasx_madd_h(dot2, __lasx_xvreplgr2vr_h(2*ls2+1));
-            sumi1 = __lasx_xvadd_w(sumi1, p1);
-            sumi2 = __lasx_xvadd_w(sumi2, p2);
-        }
-
-        accumf = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(d), __lasx_xvffint_s_w(__lasx_xvadd_w(sumi1, sumi2)), accumf);
-    }
-
-    *s = 0.25f * hsum_float_8(accumf);
-
-#else
-
-    uint32_t aux32;
-
-    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 * 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);
-            const uint32_t ls = 2*(aux32 >> 28) + 1;
-            int32_t sumi = 0;
-            for (int l = 0; l < 4; ++l) {
-                const uint8_t * grid1 = (const uint8_t *)(iq3xxs_grid + q3[2*l+0]);
-                const uint8_t * grid2 = (const uint8_t *)(iq3xxs_grid + q3[2*l+1]);
-                const uint8_t  signs = ksigns_iq2xs[(aux32 >> 7*l) & 127];
-                for (int j = 0; j < 4; ++j) {
-                    sumi += grid1[j] * q8[j+0] * (signs & kmask_iq2xs[j+0] ? -1 : 1);
-                    sumi += grid2[j] * q8[j+4] * (signs & kmask_iq2xs[j+4] ? -1 : 1);
-                }
-                q8 += 8;
-            }
-            q3 += 8;
-            bsum += sumi * ls;
-        }
-        sumf += d * bsum;
-    }
-    *s = 0.25f * sumf;
-#endif
-}
-
-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);
-    UNUSED(bx);
-    UNUSED(by);
-    UNUSED(bs);
-
-    const block_iq3_s * GGML_RESTRICT x = vx;
-    const block_q8_K  * GGML_RESTRICT y = vy;
-
-    const int nb = n / QK_K;
-
-#if defined(__ARM_NEON)
-
-    typedef union {
-        uint16x8_t vec_index;
-        uint16_t   index[8];
-    } vec_index_t;
-
-   static const uint8_t k_mask1[32] = {0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01,
-                                       0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03
-   };
-
-    static const uint8_t k_mask2[16] = {0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,};
-
-    static const int16_t k_shift[8] = {8, 7, 6, 5, 4, 3, 2, 1};
-
-    const ggml_uint8x16x2_t mask1 = ggml_vld1q_u8_x2(k_mask1);
-    const uint8x16_t        mask2 = vld1q_u8(k_mask2);
-
-    const int16x8_t  hshift = vld1q_s16(k_shift);
-    const uint16x8_t m256   = vdupq_n_u16(256);
-    const uint8x16_t m1     = vdupq_n_u8(1);
-
-    uint8x16x2_t vs;
-    ggml_int8x16x4_t q3s;
-    ggml_int8x16x4_t q8b;
-    vec_index_t idx;
-
-    uint32_t scales32[2];
-    const uint8_t * scales8 = (const uint8_t *)scales32;
-
-    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 * 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;
-        scales32[0] = ((scales32[0] & 0x0f0f0f0f) << 1) | 0x01010101;
-
-        int sumi1 = 0, sumi2 = 0;
-        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
-            q8b = ggml_vld1q_s8_x4(q8); q8 += 64;
-
-            const uint8x16_t idx_l = vld1q_u8(qs); qs += 16;
-            idx.vec_index = vorrq_u16(vmovl_u8(vget_low_u8 (idx_l)), vandq_u16(vshlq_u16(vdupq_n_u16(qh[ib32+0]), hshift), m256));
-            const uint32x4_t aux32x4_0 = ggml_vld1q_u32(iq3s_grid[idx.index[0]], iq3s_grid[idx.index[1]],
-                                                        iq3s_grid[idx.index[2]], iq3s_grid[idx.index[3]]);
-            const uint32x4_t aux32x4_1 = ggml_vld1q_u32(iq3s_grid[idx.index[4]], iq3s_grid[idx.index[5]],
-                                                        iq3s_grid[idx.index[6]], iq3s_grid[idx.index[7]]);
-            idx.vec_index = vorrq_u16(vmovl_u8(vget_high_u8(idx_l)), vandq_u16(vshlq_u16(vdupq_n_u16(qh[ib32+1]), hshift), m256));
-            const uint32x4_t aux32x4_2 = ggml_vld1q_u32(iq3s_grid[idx.index[0]], iq3s_grid[idx.index[1]],
-                                                        iq3s_grid[idx.index[2]], iq3s_grid[idx.index[3]]);
-            const uint32x4_t aux32x4_3 = ggml_vld1q_u32(iq3s_grid[idx.index[4]], iq3s_grid[idx.index[5]],
-                                                        iq3s_grid[idx.index[6]], iq3s_grid[idx.index[7]]);
-
-
-            vs.val[0] = vreinterpretq_u8_u32(vdupq_n_u32(signs[0] | ((uint32_t) signs[1] << 16)));
-            vs.val[1] = vandq_u8(ggml_vqtbl1q_u8(vs.val[0], mask1.val[1]), mask2);
-            vs.val[0] = vandq_u8(ggml_vqtbl1q_u8(vs.val[0], mask1.val[0]), mask2);
-            vs.val[0] = vorrq_u8(vceqq_u8(vs.val[0], mask2), m1);
-            vs.val[1] = vorrq_u8(vceqq_u8(vs.val[1], mask2), m1);
-
-            q3s.val[0] = vmulq_s8(vreinterpretq_s8_u8(vs.val[0]), vreinterpretq_s8_u32(aux32x4_0));
-            q3s.val[1] = vmulq_s8(vreinterpretq_s8_u8(vs.val[1]), vreinterpretq_s8_u32(aux32x4_1));
-
-            vs.val[0] = vreinterpretq_u8_u32(vdupq_n_u32(signs[2] | ((uint32_t) signs[3] << 16)));
-            vs.val[1] = vandq_u8(ggml_vqtbl1q_u8(vs.val[0], mask1.val[1]), mask2);
-            vs.val[0] = vandq_u8(ggml_vqtbl1q_u8(vs.val[0], mask1.val[0]), mask2);
-            vs.val[0] = vorrq_u8(vceqq_u8(vs.val[0], mask2), m1);
-            vs.val[1] = vorrq_u8(vceqq_u8(vs.val[1], mask2), m1);
-
-            signs += 4;
-
-            q3s.val[2] = vmulq_s8(vreinterpretq_s8_u8(vs.val[0]), vreinterpretq_s8_u32(aux32x4_2));
-            q3s.val[3] = vmulq_s8(vreinterpretq_s8_u8(vs.val[1]), vreinterpretq_s8_u32(aux32x4_3));
-
-            const int32x4_t p1 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), q3s.val[0], q8b.val[0]), q3s.val[1], q8b.val[1]);
-            const int32x4_t p2 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), q3s.val[2], q8b.val[2]), q3s.val[3], q8b.val[3]);
-
-            sumi1 += vaddvq_s32(p1) * scales8[ib32/2+0];
-            sumi2 += vaddvq_s32(p2) * scales8[ib32/2+4];
-        }
-        sumf += d*(sumi1 + sumi2);
-    }
-    *s = sumf;
-
-#elif defined(__AVX2__)
-
-   static const uint8_t k_mask1[32] = {0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01,
-                                       0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03
-   };
-
-    static const uint8_t k_mask2[32] = {0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
-                                        0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
-    };
-
-    const __m256i mask1 = _mm256_loadu_si256((const __m256i*)k_mask1);
-    const __m256i mask2 = _mm256_loadu_si256((const __m256i*)k_mask2);
-
-    const __m256i idx_shift = _mm256_set_epi32(1, 2, 3, 4, 5, 6, 7, 8);
-    const __m256i idx_mask  = _mm256_set1_epi32(256);
-
-    typedef union {
-        __m256i  vec[2];
-        uint32_t index[16];
-    } index_t;
-
-    index_t idx;
-
-    __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 * 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) {
-            const __m256i q8_1 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32;
-            const __m256i q8_2 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32;
-            const __m256i idx_l = _mm256_cvtepu8_epi16(_mm_loadu_si128((const __m128i *)qs)); qs += 16;
-            idx.vec[0] = _mm256_set1_epi32(qh[ib32+0]);
-            idx.vec[1] = _mm256_set1_epi32(qh[ib32+1]);
-            idx.vec[0] = _mm256_and_si256(_mm256_sllv_epi32(idx.vec[0], idx_shift), idx_mask);
-            idx.vec[1] = _mm256_and_si256(_mm256_sllv_epi32(idx.vec[1], idx_shift), idx_mask);
-            idx.vec[0] = _mm256_or_si256(idx.vec[0], _mm256_cvtepi16_epi32(_mm256_castsi256_si128(idx_l)));
-            idx.vec[1] = _mm256_or_si256(idx.vec[1], _mm256_cvtepi16_epi32(_mm256_extractf128_si256(idx_l, 1)));
-
-            // At leat on my CPU (Ryzen 7950X), using _mm256_i32gather_epi32 is slower than _mm256_set_epi32. Strange.
-            //const __m256i q2_1 = _mm256_i32gather_epi32((const int *)iq3s_grid, idx.vec[0], 4);
-            //const __m256i q2_2 = _mm256_i32gather_epi32((const int *)iq3s_grid, idx.vec[1], 4);
-            const __m256i q2_1 = _mm256_set_epi32(
-                    iq3s_grid[idx.index[7]], iq3s_grid[idx.index[6]], iq3s_grid[idx.index[5]], iq3s_grid[idx.index[4]],
-                    iq3s_grid[idx.index[3]], iq3s_grid[idx.index[2]], iq3s_grid[idx.index[1]], iq3s_grid[idx.index[0]]
-            );
-            const __m256i q2_2 = _mm256_set_epi32(
-                    iq3s_grid[idx.index[15]], iq3s_grid[idx.index[14]], iq3s_grid[idx.index[13]], iq3s_grid[idx.index[12]],
-                    iq3s_grid[idx.index[11]], iq3s_grid[idx.index[10]], iq3s_grid[idx.index[ 9]], iq3s_grid[idx.index[ 8]]
-            );
-
-            __m256i aux256 = _mm256_set1_epi32(signs[0] | (signs[1] << 16));
-            aux256 = _mm256_and_si256(_mm256_shuffle_epi8(aux256,mask1), mask2);
-            const __m256i s2_1 = _mm256_cmpeq_epi8(aux256, mask2);
-            const __m256i q8s_1 = _mm256_sub_epi8(_mm256_xor_si256(s2_1, q8_1), s2_1);
-
-            aux256 = _mm256_set1_epi32(signs[2] | (signs[3] << 16));
-            aux256 = _mm256_and_si256(_mm256_shuffle_epi8(aux256,mask1), mask2);
-            const __m256i s2_2 = _mm256_cmpeq_epi8(aux256, mask2);
-            const __m256i q8s_2 = _mm256_sub_epi8(_mm256_xor_si256(s2_2, q8_2), s2_2);
-
-            signs += 4;
-
-            const __m256i dot1  = _mm256_maddubs_epi16(q2_1, q8s_1);
-            const __m256i dot2  = _mm256_maddubs_epi16(q2_2, q8s_2);
-            const uint16_t ls1 = x[i].scales[ib32/2] & 0xf;
-            const uint16_t ls2 = x[i].scales[ib32/2] >>  4;
-            const __m256i p1 = _mm256_madd_epi16(dot1, _mm256_set1_epi16(2*ls1+1));
-            const __m256i p2 = _mm256_madd_epi16(dot2, _mm256_set1_epi16(2*ls2+1));
-            sumi1 = _mm256_add_epi32(sumi1, p1);
-            sumi2 = _mm256_add_epi32(sumi2, p2);
-        }
-
-        accumf = _mm256_fmadd_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(_mm256_add_epi32(sumi1, sumi2)), accumf);
-
-    }
-
-    *s = hsum_float_8(accumf);
-
-#elif defined(__AVX__)
-   static const uint8_t k_mask1[32] = {0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01,
-                                       0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03
-   };
-
-    static const uint8_t k_mask2[32] = {0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
-                                        0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
-    };
-
-    const __m128i mask1_0 = _mm_loadu_si128((const __m128i*)k_mask1);
-    const __m128i mask1_1 = _mm_loadu_si128((const __m128i*)k_mask1 + 1);
-    const __m128i mask2_0 = _mm_loadu_si128((const __m128i*)k_mask2);
-    const __m128i mask2_1 = _mm_loadu_si128((const __m128i*)k_mask2 + 1);
-
-    const __m128i idx_mul_0 = _mm_set_epi32(32, 64, 128, 256);
-    const __m128i idx_mul_1 = _mm_set_epi32(2, 4, 8, 16);
-    const __m128i idx_mask  = _mm_set1_epi32(256);
-
-    typedef union {
-        __m128i  vec[4];
-        uint32_t index[16];
-    } index_t;
-
-    index_t idx;
-
-    __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 * 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();
-        __m128i sumi2_1 = _mm_setzero_si128();
-        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
-            const __m128i q8_1_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
-            const __m128i q8_1_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
-            const __m128i q8_2_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
-            const __m128i q8_2_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
-            const __m128i qs_tmp = _mm_loadu_si128((const __m128i *)qs);
-            const __m128i idx_l_0 = _mm_cvtepu8_epi16(qs_tmp);
-            const __m128i idx_l_1 = _mm_cvtepu8_epi16(_mm_srli_si128(qs_tmp, 8)); qs += 16;
-            idx.vec[0] = _mm_set1_epi32(qh[ib32+0]);
-            idx.vec[1] = idx.vec[0];
-            idx.vec[2] = _mm_set1_epi32(qh[ib32+1]);
-            idx.vec[3] = idx.vec[2];
-
-            idx.vec[0] = _mm_and_si128(_mm_mullo_epi32(idx.vec[0], idx_mul_0), idx_mask);
-            idx.vec[1] = _mm_and_si128(_mm_mullo_epi32(idx.vec[1], idx_mul_1), idx_mask);
-            idx.vec[2] = _mm_and_si128(_mm_mullo_epi32(idx.vec[2], idx_mul_0), idx_mask);
-            idx.vec[3] = _mm_and_si128(_mm_mullo_epi32(idx.vec[3], idx_mul_1), idx_mask);
-
-            idx.vec[0] = _mm_or_si128(idx.vec[0], _mm_cvtepi16_epi32(idx_l_0));
-            idx.vec[1] = _mm_or_si128(idx.vec[1], _mm_cvtepi16_epi32(_mm_srli_si128(idx_l_0, 8)));
-            idx.vec[2] = _mm_or_si128(idx.vec[2], _mm_cvtepi16_epi32(idx_l_1));
-            idx.vec[3] = _mm_or_si128(idx.vec[3], _mm_cvtepi16_epi32(_mm_srli_si128(idx_l_1, 8)));
-
-            const __m128i q2_1_0 = _mm_set_epi32(iq3s_grid[idx.index[3]], iq3s_grid[idx.index[2]], iq3s_grid[idx.index[1]], iq3s_grid[idx.index[0]]);
-            const __m128i q2_1_1 = _mm_set_epi32(iq3s_grid[idx.index[7]], iq3s_grid[idx.index[6]], iq3s_grid[idx.index[5]], iq3s_grid[idx.index[4]]);
-            const __m128i q2_2_0 = _mm_set_epi32(iq3s_grid[idx.index[11]], iq3s_grid[idx.index[10]], iq3s_grid[idx.index[9]], iq3s_grid[idx.index[8]]);
-            const __m128i q2_2_1 = _mm_set_epi32(iq3s_grid[idx.index[15]], iq3s_grid[idx.index[14]], iq3s_grid[idx.index[13]], iq3s_grid[idx.index[12]]);
-
-            __m128i aux128_0 = _mm_set1_epi32(signs[0] | (signs[1] << 16));
-            __m128i aux128_1 = aux128_0;
-            aux128_0 = _mm_and_si128(_mm_shuffle_epi8(aux128_0,mask1_0), mask2_0);
-            aux128_1 = _mm_and_si128(_mm_shuffle_epi8(aux128_1,mask1_1), mask2_1);
-            const __m128i s2_1_0 = _mm_cmpeq_epi8(aux128_0, mask2_0);
-            const __m128i s2_1_1 = _mm_cmpeq_epi8(aux128_1, mask2_1);
-            const __m128i q8s_1_0 = _mm_sub_epi8(_mm_xor_si128(s2_1_0, q8_1_0), s2_1_0);
-            const __m128i q8s_1_1 = _mm_sub_epi8(_mm_xor_si128(s2_1_1, q8_1_1), s2_1_1);
-
-            aux128_0 = _mm_set1_epi32(signs[2] | (signs[3] << 16));
-            aux128_1 = aux128_0;
-            aux128_0 = _mm_and_si128(_mm_shuffle_epi8(aux128_0,mask1_0), mask2_0);
-            aux128_1 = _mm_and_si128(_mm_shuffle_epi8(aux128_1,mask1_1), mask2_1);
-            const __m128i s2_2_0 = _mm_cmpeq_epi8(aux128_0, mask2_0);
-            const __m128i s2_2_1 = _mm_cmpeq_epi8(aux128_1, mask2_1);
-            const __m128i q8s_2_0 = _mm_sub_epi8(_mm_xor_si128(s2_2_0, q8_2_0), s2_2_0);
-            const __m128i q8s_2_1 = _mm_sub_epi8(_mm_xor_si128(s2_2_1, q8_2_1), s2_2_1);
-
-            signs += 4;
-
-            const __m128i dot1_0  = _mm_maddubs_epi16(q2_1_0, q8s_1_0);
-            const __m128i dot1_1  = _mm_maddubs_epi16(q2_1_1, q8s_1_1);
-            const __m128i dot2_0  = _mm_maddubs_epi16(q2_2_0, q8s_2_0);
-            const __m128i dot2_1  = _mm_maddubs_epi16(q2_2_1, q8s_2_1);
-            const uint16_t ls1 = x[i].scales[ib32/2] & 0xf;
-            const uint16_t ls2 = x[i].scales[ib32/2] >>  4;
-            const __m128i p1_0 = _mm_madd_epi16(dot1_0, _mm_set1_epi16(2*ls1+1));
-            const __m128i p1_1 = _mm_madd_epi16(dot1_1, _mm_set1_epi16(2*ls1+1));
-            const __m128i p2_0 = _mm_madd_epi16(dot2_0, _mm_set1_epi16(2*ls2+1));
-            const __m128i p2_1 = _mm_madd_epi16(dot2_1, _mm_set1_epi16(2*ls2+1));
-            sumi1_0 = _mm_add_epi32(sumi1_0, p1_0);
-            sumi1_1 = _mm_add_epi32(sumi1_1, p1_1);
-            sumi2_0 = _mm_add_epi32(sumi2_0, p2_0);
-            sumi2_1 = _mm_add_epi32(sumi2_1, p2_1);
-        }
-
-        accumf = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(MM256_SET_M128I(_mm_add_epi32(sumi1_1, sumi2_1), _mm_add_epi32(sumi1_0, sumi2_0)))), accumf);
-
-    }
-
-    *s = hsum_float_8(accumf);
-
-#elif defined(__POWER9_VECTOR__)
-    static const uint8_t k_mask1[32] = {0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01,
-                                        0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03
-    };
-
-    static const uint8_t k_mask2[16] = {0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,};
-
-    const vector int v0 = vec_splats((int32_t)0);
-
-    vector float vsumf0 = vec_splats(0.0f);
-    vector float vsumf1 = vec_splats(0.0f);
-    vector float vsumf2 = vec_splats(0.0f);
-    vector float vsumf3 = vec_splats(0.0f);
-
-    const vector unsigned char mask0 = vec_xl( 0, k_mask1);
-    const vector unsigned char mask1 = vec_xl(16, k_mask1);
-    const vector signed char mask2 = (vector signed char)vec_xl( 0, k_mask2);
-
-    for (int i = 0; i < nb; ++i) {
-        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d));
-        vector float vyd = vec_splats(y[i].d);
-        vector float vd = vec_mul(vxd, vyd);
-
-        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;
-        vector signed int vsumi2 = v0;
-        vector signed int vsumi3 = v0;
-
-        for (int j = 0; j < QK_K/32; j += 2) {
-            __builtin_prefetch(q3, 0, 1);
-            __builtin_prefetch(q8, 0, 1);
-
-            vector unsigned int aux32x4_0 = {iq3s_grid[q3[ 0] | ((qh[0] << 8) & 256)], iq3s_grid[q3[ 1] | ((qh[0] << 7) & 256)],
-                                             iq3s_grid[q3[ 2] | ((qh[0] << 6) & 256)], iq3s_grid[q3[ 3] | ((qh[0] << 5) & 256)]};
-            vector unsigned int aux32x4_1 = {iq3s_grid[q3[ 4] | ((qh[0] << 4) & 256)], iq3s_grid[q3[ 5] | ((qh[0] << 3) & 256)],
-                                             iq3s_grid[q3[ 6] | ((qh[0] << 2) & 256)], iq3s_grid[q3[ 7] | ((qh[0] << 1) & 256)]};
-            vector unsigned int aux32x4_2 = {iq3s_grid[q3[ 8] | ((qh[1] << 8) & 256)], iq3s_grid[q3[ 9] | ((qh[1] << 7) & 256)],
-                                             iq3s_grid[q3[10] | ((qh[1] << 6) & 256)], iq3s_grid[q3[11] | ((qh[1] << 5) & 256)]};
-            vector unsigned int aux32x4_3 = {iq3s_grid[q3[12] | ((qh[1] << 4) & 256)], iq3s_grid[q3[13] | ((qh[1] << 3) & 256)],
-                                             iq3s_grid[q3[14] | ((qh[1] << 2) & 256)], iq3s_grid[q3[15] | ((qh[1] << 1) & 256)]};
-            q3 += 16;
-            qh += 2;
-
-            vector signed char vsigns01 = (vector signed char)vec_splats(*(const uint32_t *)&signs[0]);
-            vector signed char vsigns02 = (vector signed char)vec_splats(*(const uint32_t *)&signs[2]);
-            signs += 4;
-
-            vector signed char vsigns0 = vec_perm(vsigns01, vsigns01, mask0);
-            vector signed char vsigns1 = vec_perm(vsigns01, vsigns01, mask1);
-            vector signed char vsigns2 = vec_perm(vsigns02, vsigns02, mask0);
-            vector signed char vsigns3 = vec_perm(vsigns02, vsigns02, mask1);
-
-            vsigns0 = (vector signed char)vec_cmpeq(vec_and(vsigns0, mask2), mask2);
-            vsigns1 = (vector signed char)vec_cmpeq(vec_and(vsigns1, mask2), mask2);
-            vsigns2 = (vector signed char)vec_cmpeq(vec_and(vsigns2, mask2), mask2);
-            vsigns3 = (vector signed char)vec_cmpeq(vec_and(vsigns3, mask2), mask2);
-
-            vector signed char q3x0 = vec_sub(vec_xor(vsigns0, (vector signed char)aux32x4_0), vsigns0);
-            vector signed char q3x1 = vec_sub(vec_xor(vsigns1, (vector signed char)aux32x4_1), vsigns1);
-            vector signed char q3x2 = vec_sub(vec_xor(vsigns2, (vector signed char)aux32x4_2), vsigns2);
-            vector signed char q3x3 = vec_sub(vec_xor(vsigns3, (vector signed char)aux32x4_3), vsigns3);
-
-            vector signed char q8y0 = vec_xl( 0, q8);
-            vector signed char q8y1 = vec_xl(16, q8);
-            vector signed char q8y2 = vec_xl(32, q8);
-            vector signed char q8y3 = vec_xl(48, q8);
-            q8 += 64;
-
-            vector signed short qv0 = vec_add(vec_mule(q3x0, q8y0), vec_mulo(q3x0, q8y0));
-            vector signed short qv1 = vec_add(vec_mule(q3x1, q8y1), vec_mulo(q3x1, q8y1));
-            vector signed short qv2 = vec_add(vec_mule(q3x2, q8y2), vec_mulo(q3x2, q8y2));
-            vector signed short qv3 = vec_add(vec_mule(q3x3, q8y3), vec_mulo(q3x3, q8y3));
-
-            const uint16_t ls0 = (uint16_t)(sc[0] & 0xf);
-            const uint16_t ls1 = (uint16_t)(sc[0] >>  4);
-            sc ++;
-
-            vector signed short vscales01 = (vector signed short)vec_splats((uint16_t)(2*ls0+1));
-            vector signed short vscales23 = (vector signed short)vec_splats((uint16_t)(2*ls1+1));
-
-            vsumi0 = vec_msum(qv0, vscales01, vsumi0);
-            vsumi1 = vec_msum(qv1, vscales01, vsumi1);
-            vsumi2 = vec_msum(qv2, vscales23, vsumi2);
-            vsumi3 = vec_msum(qv3, vscales23, vsumi3);
-        }
-
-        vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
-        vsumf1 = vec_madd(vec_ctf(vsumi1, 0), vd, vsumf1);
-        vsumf2 = vec_madd(vec_ctf(vsumi2, 0), vd, vsumf2);
-        vsumf3 = vec_madd(vec_ctf(vsumi3, 0), vd, vsumf3);
-    }
-
-    vsumf0 = vec_add(vsumf0, vsumf2);
-    vsumf1 = vec_add(vsumf1, vsumf3);
-
-    vsumf0 = vec_add(vsumf0, vsumf1);
-
-    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4));
-    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8));
-
-    *s = vec_extract(vsumf0, 0);
-
-#elif defined(__loongarch_asx)
-
-   static const uint8_t k_mask1[32] = {0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01,
-                                       0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03
-   };
-
-    static const uint8_t k_mask2[32] = {0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
-                                        0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
-    };
-
-    const __m256i mask1 = __lasx_xvld((const __m256i*)k_mask1, 0);
-    const __m256i mask2 = __lasx_xvld((const __m256i*)k_mask2, 0);
-
-    __m256i idx_shift = lasx_set_w(1, 2, 3, 4, 5, 6, 7, 8);
-    const __m256i idx_mask  = __lasx_xvreplgr2vr_w(256);
-
-    typedef union {
-        __m256i  vec[2];
-        uint32_t index[16];
-    } index_t;
-
-    index_t idx;
-
-    __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 * 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) {
-            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 idx_l = lasx_extu8_16(__lsx_vld(qs, 0)); qs += 16;
-            idx.vec[0] = __lasx_xvreplgr2vr_w(qh[ib32+0]);
-            idx.vec[1] = __lasx_xvreplgr2vr_w(qh[ib32+1]);
-            idx.vec[0] = __lasx_xvand_v(__lasx_xvsll_w(idx.vec[0], idx_shift), idx_mask);
-            idx.vec[1] = __lasx_xvand_v(__lasx_xvsll_w(idx.vec[1], idx_shift), idx_mask);
-            idx.vec[0] = __lasx_xvor_v(idx.vec[0], lasx_ext16_32(lasx_extracti128(idx_l, 0)));
-            idx.vec[1] = __lasx_xvor_v(idx.vec[1], lasx_ext16_32(lasx_extracti128(idx_l, 1)));
-
-            // At leat on my CPU (Ryzen 7950X), using _mm256_i32gather_epi32 is slower than _mm256_set_epi32. Strange.
-            //const __m256i q2_1 = _mm256_i32gather_epi32((const int *)iq3s_grid, idx.vec[0], 4);
-            //const __m256i q2_2 = _mm256_i32gather_epi32((const int *)iq3s_grid, idx.vec[1], 4);
-            const __m256i q2_1 = lasx_set_w(
-                    iq3s_grid[idx.index[7]], iq3s_grid[idx.index[6]], iq3s_grid[idx.index[5]], iq3s_grid[idx.index[4]],
-                    iq3s_grid[idx.index[3]], iq3s_grid[idx.index[2]], iq3s_grid[idx.index[1]], iq3s_grid[idx.index[0]]
-            );
-            const __m256i q2_2 = lasx_set_w(
-                    iq3s_grid[idx.index[15]], iq3s_grid[idx.index[14]], iq3s_grid[idx.index[13]], iq3s_grid[idx.index[12]],
-                    iq3s_grid[idx.index[11]], iq3s_grid[idx.index[10]], iq3s_grid[idx.index[ 9]], iq3s_grid[idx.index[ 8]]
-            );
-
-            __m256i aux256 = __lasx_xvreplgr2vr_w(signs[0] | (signs[1] << 16));
-            aux256 = __lasx_xvand_v(lasx_shuffle_b(aux256,mask1), mask2);
-            const __m256i s2_1 = __lasx_xvseq_b(aux256, mask2);
-            const __m256i q8s_1 = __lasx_xvsub_b(__lasx_xvxor_v(s2_1, q8_1), s2_1);
-
-            aux256 = __lasx_xvreplgr2vr_w(signs[2] | (signs[3] << 16));
-            aux256 = __lasx_xvand_v(lasx_shuffle_b(aux256,mask1), mask2);
-            const __m256i s2_2 = __lasx_xvseq_b(aux256, mask2);
-            const __m256i q8s_2 = __lasx_xvsub_b(__lasx_xvxor_v(s2_2, q8_2), s2_2);
-
-            signs += 4;
-
-            const __m256i dot1 = lasx_maddubs_h(q2_1, q8s_1);
-            const __m256i dot2  = lasx_maddubs_h(q2_2, q8s_2);
-            const uint16_t ls1 = x[i].scales[ib32/2] & 0xf;
-            const uint16_t ls2 = x[i].scales[ib32/2] >>  4;
-            const __m256i p1 = lasx_madd_h(dot1, __lasx_xvreplgr2vr_h(2*ls1+1));
-            const __m256i p2 = lasx_madd_h(dot2, __lasx_xvreplgr2vr_h(2*ls2+1));
-            sumi1 = __lasx_xvadd_w(sumi1, p1);
-            sumi2 = __lasx_xvadd_w(sumi2, p2);
-        }
-
-        accumf = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(d), __lasx_xvffint_s_w(__lasx_xvadd_w(sumi1, sumi2)), accumf);
-    }
-
-    *s = hsum_float_8(accumf);
-
-#else
-
-    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 * 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;
-            const uint32_t ls2 = 2*(x[i].scales[ib32/2] >>  4) + 1;
-            int32_t sumi = 0;
-            for (int l = 0; l < 4; ++l) {
-                const uint8_t * grid1 = (const uint8_t *)(iq3s_grid + (qs[2*l+0] | ((qh[ib32+0] << (8-2*l)) & 256)));
-                const uint8_t * grid2 = (const uint8_t *)(iq3s_grid + (qs[2*l+1] | ((qh[ib32+0] << (7-2*l)) & 256)));
-                for (int j = 0; j < 4; ++j) {
-                    sumi += grid1[j] * q8[j+0] * (signs[l] & kmask_iq2xs[j+0] ? -1 : 1);
-                    sumi += grid2[j] * q8[j+4] * (signs[l] & kmask_iq2xs[j+4] ? -1 : 1);
-                }
-                q8 += 8;
-            }
-            qs += 8;
-            signs += 4;
-            bsum += sumi * ls1;
-            sumi = 0;
-            for (int l = 0; l < 4; ++l) {
-                const uint8_t * grid1 = (const uint8_t *)(iq3s_grid + (qs[2*l+0] | ((qh[ib32+1] << (8-2*l)) & 256)));
-                const uint8_t * grid2 = (const uint8_t *)(iq3s_grid + (qs[2*l+1] | ((qh[ib32+1] << (7-2*l)) & 256)));
-                for (int j = 0; j < 4; ++j) {
-                    sumi += grid1[j] * q8[j+0] * (signs[l] & kmask_iq2xs[j+0] ? -1 : 1);
-                    sumi += grid2[j] * q8[j+4] * (signs[l] & kmask_iq2xs[j+4] ? -1 : 1);
-                }
-                q8 += 8;
-            }
-            qs += 8;
-            signs += 4;
-            bsum += sumi * ls2;
-        }
-        sumf += d * bsum;
-    }
-    *s = sumf;
-#endif
-}
-
-#if defined(__AVX2__)
-static inline __m256i mul_add_epi8(const __m256i x, const __m256i y) {
-    const __m256i ax = _mm256_sign_epi8(x, x);
-    const __m256i sy = _mm256_sign_epi8(y, x);
-    return _mm256_maddubs_epi16(ax, sy);
-}
-#elif defined(__loongarch_asx)
-static inline __m256i mul_add_epi8(const __m256i x, const __m256i y) {
-    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 * 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);
-    UNUSED(bx);
-    UNUSED(by);
-    UNUSED(bs);
-
-    const block_iq1_s * GGML_RESTRICT x = vx;
-    const block_q8_K  * GGML_RESTRICT y = vy;
-
-    const int nb = n / QK_K;
-
-#if defined __ARM_NEON
-
-    ggml_int8x16x4_t q1b;
-    ggml_int8x16x4_t q8b;
-
-    float sumf = 0;
-    for (int i = 0; i < nb; ++i) {
-
-        const int8_t   * q8 = y[i].qs;
-        const uint8_t  * qs = x[i].qs;
-        const uint16_t * qh = x[i].qh;
-
-        int sumi1 = 0, sumi2 = 0, sumi3 = 0;
-
-        for (int ib = 0; ib < QK_K/32; ib += 2) {
-
-            q1b.val[0] = vcombine_s8(vld1_s8((const int8_t *)(iq1s_grid + (qs[0] | ((qh[ib+0] << 8) & 0x700)))),
-                                     vld1_s8((const int8_t *)(iq1s_grid + (qs[1] | ((qh[ib+0] << 5) & 0x700)))));
-            q1b.val[1] = vcombine_s8(vld1_s8((const int8_t *)(iq1s_grid + (qs[2] | ((qh[ib+0] << 2) & 0x700)))),
-                                     vld1_s8((const int8_t *)(iq1s_grid + (qs[3] | ((qh[ib+0] >> 1) & 0x700)))));
-            q1b.val[2] = vcombine_s8(vld1_s8((const int8_t *)(iq1s_grid + (qs[4] | ((qh[ib+1] << 8) & 0x700)))),
-                                     vld1_s8((const int8_t *)(iq1s_grid + (qs[5] | ((qh[ib+1] << 5) & 0x700)))));
-            q1b.val[3] = vcombine_s8(vld1_s8((const int8_t *)(iq1s_grid + (qs[6] | ((qh[ib+1] << 2) & 0x700)))),
-                                     vld1_s8((const int8_t *)(iq1s_grid + (qs[7] | ((qh[ib+1] >> 1) & 0x700)))));
-            qs += 8;
-
-            q8b = ggml_vld1q_s8_x4(q8); q8 += 64;
-
-            const int32x4_t p1 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), q1b.val[0], q8b.val[0]), q1b.val[1], q8b.val[1]);
-            const int32x4_t p2 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), q1b.val[2], q8b.val[2]), q1b.val[3], q8b.val[3]);
-
-            const int ls1 = 2*((qh[ib+0] >> 12) & 7) + 1;
-            const int ls2 = 2*((qh[ib+1] >> 12) & 7) + 1;
-            sumi1 += vaddvq_s32(p1) * ls1;
-            sumi2 += vaddvq_s32(p2) * ls2;
-            sumi3 += (y[i].bsums[2*ib+0] + y[i].bsums[2*ib+1]) * ls1 * (qh[ib+0] & 0x8000 ? -1 : 1)
-                   + (y[i].bsums[2*ib+2] + y[i].bsums[2*ib+3]) * ls2 * (qh[ib+1] & 0x8000 ? -1 : 1);
-
-        }
-
-        sumf += y[i].d * GGML_FP16_TO_FP32(x[i].d) * (sumi1 + sumi2 + IQ1S_DELTA * sumi3);
-    }
-
-    *s = sumf;
-
-#elif defined __AVX2__
-
-    __m256 accum = _mm256_setzero_ps();
-    float accum1 = 0;
-    for (int i = 0; i < nb; ++i) {
-
-        const int8_t   * q8 = y[i].qs;
-        const uint8_t  * qs = x[i].qs;
-        const uint16_t * qh = x[i].qh;
-
-        __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;
-
-            const __m256i dot1 = mul_add_epi8(q1b_1, q8b_1);
-            const __m256i dot2 = mul_add_epi8(q1b_2, q8b_2);
-            const int16_t ls1 = 2*((qh[ib+0] >> 12) & 7) + 1;
-            const int16_t ls2 = 2*((qh[ib+1] >> 12) & 7) + 1;
-            const __m256i p1 = _mm256_madd_epi16(dot1, _mm256_set1_epi16(ls1));
-            const __m256i p2 = _mm256_madd_epi16(dot2, _mm256_set1_epi16(ls2));
-
-            sumi = _mm256_add_epi32(sumi, _mm256_add_epi32(p1, p2));
-            sumi1 += (y[i].bsums[2*ib+0] + y[i].bsums[2*ib+1]) * (qh[ib+0] & 0x8000 ? -1 : 1) * ls1
-                   + (y[i].bsums[2*ib+2] + y[i].bsums[2*ib+3]) * (qh[ib+1] & 0x8000 ? -1 : 1) * ls2;
-        }
-
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        accum = _mm256_fmadd_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(sumi), accum);
-        accum1 += d * sumi1;
-
-    }
-
-    *s = hsum_float_8(accum) + IQ1S_DELTA * accum1;
-
-#elif defined __AVX__
-    __m256 accum = _mm256_setzero_ps();
-    float accum1 = 0;
-    for (int i = 0; i < nb; ++i) {
-
-        const int8_t   * q8 = y[i].qs;
-        const uint8_t  * qs = x[i].qs;
-        const uint16_t * qh = x[i].qh;
-
-        __m128i sumi1_0 = _mm_setzero_si128();
-        __m128i sumi1_1 = _mm_setzero_si128();
-        int sumi1 = 0;
-        for (int ib = 0; ib < QK_K/32; ib += 2) {
-            const __m128i q1b_1_0 = _mm_set_epi64x(iq1s_grid[qs[1] | ((qh[ib+0] << 5) & 0x700)], iq1s_grid[qs[0] | ((qh[ib+0] << 8) & 0x700)]);
-            const __m128i q1b_1_1 = _mm_set_epi64x(iq1s_grid[qs[3] | ((qh[ib+0] >> 1) & 0x700)], iq1s_grid[qs[2] | ((qh[ib+0] << 2) & 0x700)]);
-            const __m128i q1b_2_0 = _mm_set_epi64x(iq1s_grid[qs[5] | ((qh[ib+1] << 5) & 0x700)], iq1s_grid[qs[4] | ((qh[ib+1] << 8) & 0x700)]);
-            const __m128i q1b_2_1 = _mm_set_epi64x(iq1s_grid[qs[7] | ((qh[ib+1] >> 1) & 0x700)], iq1s_grid[qs[6] | ((qh[ib+1] << 2) & 0x700)]);
-            qs += 8;
-            const __m128i q8b_1_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
-            const __m128i q8b_1_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
-            const __m128i q8b_2_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
-            const __m128i q8b_2_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
-
-            const __m128i dot1_0 = mul_add_epi8_sse(q1b_1_0, q8b_1_0);
-            const __m128i dot1_1 = mul_add_epi8_sse(q1b_1_1, q8b_1_1);
-            const __m128i dot2_0 = mul_add_epi8_sse(q1b_2_0, q8b_2_0);
-            const __m128i dot2_1 = mul_add_epi8_sse(q1b_2_1, q8b_2_1);
-            const int16_t ls1 = 2*((qh[ib+0] >> 12) & 7) + 1;
-            const int16_t ls2 = 2*((qh[ib+1] >> 12) & 7) + 1;
-            const __m128i p1_0 = _mm_madd_epi16(dot1_0, _mm_set1_epi16(ls1));
-            const __m128i p1_1 = _mm_madd_epi16(dot1_1, _mm_set1_epi16(ls1));
-            const __m128i p2_0 = _mm_madd_epi16(dot2_0, _mm_set1_epi16(ls2));
-            const __m128i p2_1 = _mm_madd_epi16(dot2_1, _mm_set1_epi16(ls2));
-
-            sumi1_0 = _mm_add_epi32(sumi1_0, _mm_add_epi32(p1_0, p2_0));
-            sumi1_1 = _mm_add_epi32(sumi1_1, _mm_add_epi32(p1_1, p2_1));
-            sumi1 += (y[i].bsums[2*ib+0] + y[i].bsums[2*ib+1]) * (qh[ib+0] & 0x8000 ? -1 : 1) * ls1
-                   + (y[i].bsums[2*ib+2] + y[i].bsums[2*ib+3]) * (qh[ib+1] & 0x8000 ? -1 : 1) * ls2;
-        }
-
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        accum = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(MM256_SET_M128I(sumi1_1, sumi1_0))), accum);
-        accum1 += d * sumi1;
-
-    }
-
-    *s = hsum_float_8(accum) + IQ1S_DELTA * accum1;
-
-#elif defined(__POWER9_VECTOR__)
-    const vector unsigned char v0 = vec_splats((unsigned char)0x0);
-    const vector unsigned short vsign = vec_splats((unsigned short)0x8000);
-
-    vector float vsumf0 = vec_splats(0.0f);
-    vector float vsumf1 = vec_splats(0.0f);
-    vector float vsumf2 = vec_splats(0.0f);
-    vector float vsumf3 = vec_splats(0.0f);
-
-    for (int i = 0; i < nb; ++i) {
-        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d));
-        vector float vyd = vec_splats(y[i].d);
-        vector float vd = vec_mul(vxd, vyd);
-
-        vector signed int vsumi0 = vec_splats((int32_t)0);
-        vector signed int vsumi1 = vec_splats((int32_t)0);
-        vector signed int vsumi2 = vec_splats((int32_t)0);
-        vector signed int vsumi3 = vec_splats((int32_t)0);
-        vector signed int vsumi8 = vec_splats((int32_t)0);
-
-        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);
-            __builtin_prefetch(qh, 0, 1);
-            __builtin_prefetch(q8, 0, 1);
-
-            vector signed long long aux64x2_0 = {*(const int64_t *)(iq1s_grid + (q1[0] | ((qh[0] << 8) & 0x700))), *(const int64_t *)(iq1s_grid + (q1[1] | ((qh[0] << 5) & 0x700)))};
-            vector signed long long aux64x2_1 = {*(const int64_t *)(iq1s_grid + (q1[2] | ((qh[0] << 2) & 0x700))), *(const int64_t *)(iq1s_grid + (q1[3] | ((qh[0] >> 1) & 0x700)))};
-            vector signed long long aux64x2_2 = {*(const int64_t *)(iq1s_grid + (q1[4] | ((qh[1] << 8) & 0x700))), *(const int64_t *)(iq1s_grid + (q1[5] | ((qh[1] << 5) & 0x700)))};
-            vector signed long long aux64x2_3 = {*(const int64_t *)(iq1s_grid + (q1[6] | ((qh[1] << 2) & 0x700))), *(const int64_t *)(iq1s_grid + (q1[7] | ((qh[1] >> 1) & 0x700)))};
-            q1 += 8;
-
-            vector signed char q1x0 = (vector signed char)aux64x2_0;
-            vector signed char q1x1 = (vector signed char)aux64x2_1;
-            vector signed char q1x2 = (vector signed char)aux64x2_2;
-            vector signed char q1x3 = (vector signed char)aux64x2_3;
-
-            vector signed char q8y0 = vec_xl( 0, q8);
-            vector signed char q8y1 = vec_xl(16, q8);
-            vector signed char q8y2 = vec_xl(32, q8);
-            vector signed char q8y3 = vec_xl(48, q8);
-            q8 += 64;
-
-            vector signed short qv0 = vec_add(vec_mule(q1x0, q8y0), vec_mulo(q1x0, q8y0));
-            vector signed short qv1 = vec_add(vec_mule(q1x1, q8y1), vec_mulo(q1x1, q8y1));
-            vector signed short qv2 = vec_add(vec_mule(q1x2, q8y2), vec_mulo(q1x2, q8y2));
-            vector signed short qv3 = vec_add(vec_mule(q1x3, q8y3), vec_mulo(q1x3, q8y3));
-
-            const uint16_t ls0 = (uint16_t)((qh[0] >> 12) & 7);
-            const uint16_t ls1 = (uint16_t)((qh[1] >> 12) & 7);
-
-            vector signed short vscales01 = (vector signed short)vec_splats((uint16_t)(2*ls0+1));
-            vector signed short vscales23 = (vector signed short)vec_splats((uint16_t)(2*ls1+1));
-            vector signed short vscales = vec_sld(vscales23, vscales01, 8);
-
-            vsumi0 = vec_msum(qv0, vscales01, vsumi0);
-            vsumi1 = vec_msum(qv1, vscales01, vsumi1);
-            vsumi2 = vec_msum(qv2, vscales23, vsumi2);
-            vsumi3 = vec_msum(qv3, vscales23, vsumi3);
-
-            vector signed short q8ysums = vec_xl_len(qs, 8);
-            qs += 4;
-            q8ysums = vec_mergeh(q8ysums, (vector signed short)v0);
-
-            vector signed short qxh = (vector signed short)vec_sld(vec_splats(qh[1]), vec_splats(qh[0]), 8);
-            qh += 2;
-            vector __bool short vsel = vec_cmpge(qxh, (vector signed short)v0);
-
-            vector signed short q8ysum = vec_sel((vector signed short)vec_xor((vector unsigned short)q8ysums, vsign), q8ysums, vsel);
-
-            vsumi8 = vec_add(vec_mule(q8ysum, vscales), vsumi8);
-        }
-
-        vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
-        vsumf1 = vec_madd(vec_ctf(vsumi1, 0), vd, vsumf1);
-        vsumf2 = vec_madd(vec_ctf(vsumi2, 0), vd, vsumf2);
-        vsumf3 = vec_madd(vec_ctf(vsumi3, 0), vd, vsumf3);
-
-        vsumf0 = vec_madd(vec_ctf(vsumi8, 0), vec_mul(vd, vec_splats(IQ1S_DELTA)), vsumf0);
-    }
-
-    vsumf0 = vec_add(vsumf0, vsumf2);
-    vsumf1 = vec_add(vsumf1, vsumf3);
-
-    vsumf0 = vec_add(vsumf0, vsumf1);
-
-    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4));
-    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8));
-
-    *s = vec_extract(vsumf0, 0);
-
-#elif defined(__loongarch_asx)
-
-    __m256 accum = (__m256)__lasx_xvldi(0);
-    float accum1 = 0;
-    for (int i = 0; i < nb; ++i) {
-
-        const int8_t   * q8 = y[i].qs;
-        const uint8_t  * qs = x[i].qs;
-        const uint16_t * qh = x[i].qh;
-
-        __m256i sumi = __lasx_xvldi(0);
-        int sumi1 = 0;
-        for (int ib = 0; ib < QK_K/32; ib += 2) {
-            __m256i q1b_1 = __lasx_xvinsgr2vr_d(q1b_1, iq1s_grid[qs[0] | ((qh[ib+0] << 8) & 0x700)], 0);
-            q1b_1 = __lasx_xvinsgr2vr_d(q1b_1, iq1s_grid[qs[1] | ((qh[ib+0] << 5) & 0x700)], 1);
-            q1b_1 = __lasx_xvinsgr2vr_d(q1b_1, iq1s_grid[qs[2] | ((qh[ib+0] << 2) & 0x700)], 2);
-            q1b_1 = __lasx_xvinsgr2vr_d(q1b_1, iq1s_grid[qs[3] | ((qh[ib+0] >> 1) & 0x700)], 3);
-
-            __m256i q1b_2 = __lasx_xvinsgr2vr_d(q1b_2, iq1s_grid[qs[4] | ((qh[ib+1] << 8) & 0x700)], 0);
-            q1b_2 = __lasx_xvinsgr2vr_d(q1b_2, iq1s_grid[qs[5] | ((qh[ib+1] << 5) & 0x700)], 1);
-            q1b_2 = __lasx_xvinsgr2vr_d(q1b_2, iq1s_grid[qs[6] | ((qh[ib+1] << 2) & 0x700)], 2);
-            q1b_2 = __lasx_xvinsgr2vr_d(q1b_2, iq1s_grid[qs[7] | ((qh[ib+1] >> 1) & 0x700)], 3);
-
-            qs += 8;
-            const __m256i q8b_1 = __lasx_xvld((const __m256i*)q8, 0); q8 += 32;
-            const __m256i q8b_2 = __lasx_xvld((const __m256i*)q8, 0); q8 += 32;
-
-            const __m256i dot1 = mul_add_epi8(q1b_1, q8b_1);
-            const __m256i dot2 = mul_add_epi8(q1b_2, q8b_2);
-            const int16_t ls1 = 2*((qh[ib+0] >> 12) & 7) + 1;
-            const int16_t ls2 = 2*((qh[ib+1] >> 12) & 7) + 1;
-
-            __m256i tmp1, tmp5, tmp6;
-            tmp1 = __lasx_xvreplgr2vr_h(ls1);
-            tmp5 = __lasx_xvmulwev_w_h(dot1, tmp1);
-            tmp6 = __lasx_xvmulwod_w_h(dot1, tmp1);
-            const __m256i p1 = __lasx_xvadd_w(tmp5, tmp6);
-
-            tmp1 = __lasx_xvreplgr2vr_h(ls2);
-            tmp5 = __lasx_xvmulwev_w_h(dot2, tmp1);
-            tmp6 = __lasx_xvmulwod_w_h(dot2, tmp1);
-            const __m256i p2 = __lasx_xvadd_w(tmp5, tmp6);
-
-            sumi = __lasx_xvadd_w(sumi, __lasx_xvadd_w(p1, p2));
-            sumi1 += (y[i].bsums[2*ib+0] + y[i].bsums[2*ib+1]) * (qh[ib+0] & 0x8000 ? -1 : 1) * ls1
-                   + (y[i].bsums[2*ib+2] + y[i].bsums[2*ib+3]) * (qh[ib+1] & 0x8000 ? -1 : 1) * ls2;
-        }
-
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        accum = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(d), __lasx_xvffint_s_w(sumi), accum);
-        accum1 += d * sumi1;
-    }
-
-    *s = hsum_float_8(accum) + IQ1S_DELTA * accum1;
-
-#else
-
-    float sumf = 0;
-    for (int i = 0; i < nb; i++) {
-
-        const int8_t   * q8 = y[i].qs;
-        const uint8_t  * qs = x[i].qs;
-        const uint16_t * qh = x[i].qh;
-
-        int sumi = 0, sumi1 = 0;
-        for (int ib = 0; ib < QK_K/32; ++ib) {
-            const int ls = 2*((qh[ib] >> 12) & 7) + 1;
-            const int delta = qh[ib] & 0x8000 ? -1 : 1;
-            int lsum = 0;
-            for (int l = 0; l < 4; ++l) {
-                const int8_t * grid = (const int8_t *)(iq1s_grid + (qs[l] | (((qh[ib] >> 3*l) & 7) << 8)));
-                for (int j = 0; j < 8; ++j) {
-                    lsum += q8[j] * grid[j];
-                }
-                q8 += 8;
-            }
-            sumi  += ls * lsum;
-            sumi1 += ls * delta * (y[i].bsums[2*ib+0] + y[i].bsums[2*ib+1]);
-            qs += 4;
-        }
-
-        sumf += GGML_FP16_TO_FP32(x[i].d) * y[i].d * (sumi + IQ1S_DELTA * sumi1);
-    }
-
-    *s = sumf;
-
-#endif
-}
-
-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);
-    UNUSED(bx);
-    UNUSED(by);
-    UNUSED(bs);
-
-    const block_iq1_m * GGML_RESTRICT x = vx;
-    const block_q8_K  * GGML_RESTRICT y = vy;
-
-    const int nb = n / QK_K;
-
-    iq1m_scale_t scale;
-
-#if defined __ARM_NEON
-    const int32x4_t mask  = vdupq_n_s32(0x7);
-    const int32x4_t mone  = vdupq_n_s32(1);
-    const int32x4_t mzero = vdupq_n_s32(0);
-
-    ggml_int8x16x4_t deltas;
-    deltas.val[0] = vcombine_s8(vdup_n_s8(+1), vdup_n_s8(+1));
-    deltas.val[1] = vcombine_s8(vdup_n_s8(-1), vdup_n_s8(+1));
-    deltas.val[2] = vcombine_s8(vdup_n_s8(+1), vdup_n_s8(-1));
-    deltas.val[3] = vcombine_s8(vdup_n_s8(-1), vdup_n_s8(-1));
-
-    ggml_int8x16x4_t q1b;
-    ggml_int8x16x4_t q8b;
-
-    uint32_t aux32;
-    const uint8_t * aux8 = (const uint8_t *)&aux32;
-
-    float sumf = 0;
-    for (int i = 0; i < nb; ++i) {
-
-        const int8_t   * q8 = y[i].qs;
-        const uint8_t  * qs = x[i].qs;
-        const uint8_t  * qh = x[i].qh;
-        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);
-
-        int32x4_t sumi1 = mzero;
-        int32x4_t sumi2 = mzero;
-
-        for (int ib = 0; ib < QK_K/32; ib += 2) {
-
-            q1b.val[0] = vcombine_s8(vld1_s8((const int8_t *)(iq1s_grid + (qs[0] | ((qh[0] << 8) & 0x700)))),
-                                     vld1_s8((const int8_t *)(iq1s_grid + (qs[1] | ((qh[0] << 4) & 0x700)))));
-            q1b.val[1] = vcombine_s8(vld1_s8((const int8_t *)(iq1s_grid + (qs[2] | ((qh[1] << 8) & 0x700)))),
-                                     vld1_s8((const int8_t *)(iq1s_grid + (qs[3] | ((qh[1] << 4) & 0x700)))));
-            q1b.val[2] = vcombine_s8(vld1_s8((const int8_t *)(iq1s_grid + (qs[4] | ((qh[2] << 8) & 0x700)))),
-                                     vld1_s8((const int8_t *)(iq1s_grid + (qs[5] | ((qh[2] << 4) & 0x700)))));
-            q1b.val[3] = vcombine_s8(vld1_s8((const int8_t *)(iq1s_grid + (qs[6] | ((qh[3] << 8) & 0x700)))),
-                                     vld1_s8((const int8_t *)(iq1s_grid + (qs[7] | ((qh[3] << 4) & 0x700)))));
-
-            q8b = ggml_vld1q_s8_x4(q8); q8 += 64;
-
-            const int32x4_t p1 = vpaddq_s32(ggml_vdotq_s32(mzero, q1b.val[0], q8b.val[0]), ggml_vdotq_s32(mzero, q1b.val[1], q8b.val[1]));
-            const int32x4_t p2 = vpaddq_s32(ggml_vdotq_s32(mzero, q1b.val[2], q8b.val[2]), ggml_vdotq_s32(mzero, q1b.val[3], q8b.val[3]));
-            const int32x4_t p12 = vpaddq_s32(p1, p2);
-
-            const uint32_t * qh32 = (const uint32_t *)qh; // we are 4-byte aligned, so we can do that
-            aux32 = ((qh32[0] >> 3) & 0x01010101) | ((qh32[0] >> 6) & 0x02020202);
-
-            const int32x4_t p3 = vpaddq_s32(ggml_vdotq_s32(mzero, deltas.val[aux8[0]], q8b.val[0]), ggml_vdotq_s32(mzero, deltas.val[aux8[1]], q8b.val[1]));
-            const int32x4_t p4 = vpaddq_s32(ggml_vdotq_s32(mzero, deltas.val[aux8[2]], q8b.val[2]), ggml_vdotq_s32(mzero, deltas.val[aux8[3]], q8b.val[3]));
-            const int32x4_t p34 = vpaddq_s32(p3, p4);
-
-            int32x4_t scales_4 = ggml_vld1q_u32(sc[ib/2] >> 0, sc[ib/2] >> 3, sc[ib/2] >> 6, sc[ib/2] >> 9);
-
-            scales_4 = vaddq_s32(vshlq_n_s32(vandq_s32(scales_4, mask), 1), mone);
-
-            sumi1 = vmlaq_s32(sumi1, scales_4, p12);
-            sumi2 = vmlaq_s32(sumi2, scales_4, p34);
-
-            qs += 8; qh += 4;
-
-        }
-
-        sumf += y[i].d * GGML_FP16_TO_FP32(scale.f16) * (vaddvq_s32(sumi1) + IQ1M_DELTA * vaddvq_s32(sumi2));
-    }
-
-    *s = sumf;
-
-#elif defined __AVX2__
-
-    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();
-    for (int i = 0; i < nb; ++i) {
-
-        const int8_t   * q8 = y[i].qs;
-        const uint8_t  * qs = x[i].qs;
-        const uint8_t  * qh = x[i].qh;
-        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)]
-            );
-            const __m256i q1b_2 = _mm256_set_epi64x(
-                    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 delta1 = _mm256_set_epi64x(qh[1] & 0x80 ? 0xffffffffffffffff : 0x0101010101010101,
-                                                     qh[1] & 0x08 ? 0xffffffffffffffff : 0x0101010101010101,
-                                                     qh[0] & 0x80 ? 0xffffffffffffffff : 0x0101010101010101,
-                                                     qh[0] & 0x08 ? 0xffffffffffffffff : 0x0101010101010101);
-            const __m256i delta2 = _mm256_set_epi64x(qh[3] & 0x80 ? 0xffffffffffffffff : 0x0101010101010101,
-                                                     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 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_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);
-
-            const __m256i p1 = _mm256_madd_epi16(dot1, scale1);
-            const __m256i p2 = _mm256_madd_epi16(dot2, scale2);
-            const __m256i p3 = _mm256_madd_epi16(dot3, scale1);
-            const __m256i p4 = _mm256_madd_epi16(dot4, scale2);
-
-            sumi1 = _mm256_add_epi32(sumi1, _mm256_add_epi32(p1, p2));
-            sumi2 = _mm256_add_epi32(sumi2, _mm256_add_epi32(p3, p4));
-
-            qs += 8; qh += 4;
-        }
-
-        const __m256 d = _mm256_set1_ps(y[i].d * GGML_FP16_TO_FP32(scale.f16));
-
-        accum1 = _mm256_fmadd_ps(d, _mm256_cvtepi32_ps(sumi1), accum1);
-        accum2 = _mm256_fmadd_ps(d, _mm256_cvtepi32_ps(sumi2), accum2);
-    }
-
-    *s = hsum_float_8(accum1) + IQ1M_DELTA * hsum_float_8(accum2);
-
-#elif defined __AVX__
-    const __m128i mask = _mm_set1_epi16(0x7);
-    const __m128i mone = _mm_set1_epi16(1);
-
-    __m256 accum1 = _mm256_setzero_ps();
-    __m256 accum2 = _mm256_setzero_ps();
-    for (int i = 0; i < nb; ++i) {
-
-        const int8_t   * q8 = y[i].qs;
-        const uint8_t  * qs = x[i].qs;
-        const uint8_t  * qh = x[i].qh;
-        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);
-
-        __m128i sumi1_0 = _mm_setzero_si128();
-        __m128i sumi1_1 = _mm_setzero_si128();
-        __m128i sumi2_0 = _mm_setzero_si128();
-        __m128i sumi2_1 = _mm_setzero_si128();
-        for (int ib = 0; ib < QK_K/32; ib += 2) {
-            const __m128i q1b_1_0 = _mm_set_epi64x(
-                    iq1s_grid[qs[1] | (((uint16_t)qh[0] << 4) & 0x700)], iq1s_grid[qs[0] | (((uint16_t)qh[0] << 8) & 0x700)]);
-            const __m128i q1b_1_1 = _mm_set_epi64x(
-                    iq1s_grid[qs[3] | (((uint16_t)qh[1] << 4) & 0x700)], iq1s_grid[qs[2] | (((uint16_t)qh[1] << 8) & 0x700)]);
-            const __m128i q1b_2_0 = _mm_set_epi64x(
-                    iq1s_grid[qs[5] | (((uint16_t)qh[2] << 4) & 0x700)], iq1s_grid[qs[4] | (((uint16_t)qh[2] << 8) & 0x700)]);
-            const __m128i q1b_2_1 = _mm_set_epi64x(
-                    iq1s_grid[qs[7] | (((uint16_t)qh[3] << 4) & 0x700)], iq1s_grid[qs[6] | (((uint16_t)qh[3] << 8) & 0x700)]);
-            const __m128i q8b_1_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
-            const __m128i q8b_1_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
-            const __m128i q8b_2_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
-            const __m128i q8b_2_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
-
-            const __m128i dot1_0 = mul_add_epi8_sse(q1b_1_0, q8b_1_0);
-            const __m128i dot1_1 = mul_add_epi8_sse(q1b_1_1, q8b_1_1);
-            const __m128i dot2_0 = mul_add_epi8_sse(q1b_2_0, q8b_2_0);
-            const __m128i dot2_1 = mul_add_epi8_sse(q1b_2_1, q8b_2_1);
-
-            const __m128i delta1_0 = _mm_set_epi64x(qh[0] & 0x80 ? 0xffffffffffffffff : 0x0101010101010101,
-                                                     qh[0] & 0x08 ? 0xffffffffffffffff : 0x0101010101010101);
-            const __m128i delta1_1 = _mm_set_epi64x(qh[1] & 0x80 ? 0xffffffffffffffff : 0x0101010101010101,
-                                                     qh[1] & 0x08 ? 0xffffffffffffffff : 0x0101010101010101);
-            const __m128i delta2_0 = _mm_set_epi64x(qh[2] & 0x80 ? 0xffffffffffffffff : 0x0101010101010101,
-                                                     qh[2] & 0x08 ? 0xffffffffffffffff : 0x0101010101010101);
-            const __m128i delta2_1 = _mm_set_epi64x(qh[3] & 0x80 ? 0xffffffffffffffff : 0x0101010101010101,
-                                                     qh[3] & 0x08 ? 0xffffffffffffffff : 0x0101010101010101);
-
-            const __m128i dot3_0 = mul_add_epi8_sse(delta1_0, q8b_1_0);
-            const __m128i dot3_1 = mul_add_epi8_sse(delta1_1, q8b_1_1);
-            const __m128i dot4_0 = mul_add_epi8_sse(delta2_0, q8b_2_0);
-            const __m128i dot4_1 = mul_add_epi8_sse(delta2_1, q8b_2_1);
-
-            __m128i scale1_0 = _mm_set1_epi16(sc[ib/2] >> 0);
-            __m128i scale1_1 = _mm_set1_epi16(sc[ib/2] >> 3);
-            __m128i scale2_0 = _mm_set1_epi16(sc[ib/2] >> 6);
-            __m128i scale2_1 = _mm_set1_epi16(sc[ib/2] >> 9);
-
-            scale1_0 = _mm_add_epi16(_mm_slli_epi16(_mm_and_si128(scale1_0, mask), 1), mone);
-            scale1_1 = _mm_add_epi16(_mm_slli_epi16(_mm_and_si128(scale1_1, mask), 1), mone);
-            scale2_0 = _mm_add_epi16(_mm_slli_epi16(_mm_and_si128(scale2_0, mask), 1), mone);
-            scale2_1 = _mm_add_epi16(_mm_slli_epi16(_mm_and_si128(scale2_1, mask), 1), mone);
-            const __m128i p1_0 = _mm_madd_epi16(dot1_0, scale1_0);
-            const __m128i p1_1 = _mm_madd_epi16(dot1_1, scale1_1);
-            const __m128i p2_0 = _mm_madd_epi16(dot2_0, scale2_0);
-            const __m128i p2_1 = _mm_madd_epi16(dot2_1, scale2_1);
-            const __m128i p3_0 = _mm_madd_epi16(dot3_0, scale1_0);
-            const __m128i p3_1 = _mm_madd_epi16(dot3_1, scale1_1);
-            const __m128i p4_0 = _mm_madd_epi16(dot4_0, scale2_0);
-            const __m128i p4_1 = _mm_madd_epi16(dot4_1, scale2_1);
-
-            sumi1_0 = _mm_add_epi32(sumi1_0, _mm_add_epi32(p1_0, p2_0));
-            sumi1_1 = _mm_add_epi32(sumi1_1, _mm_add_epi32(p1_1, p2_1));
-            sumi2_0 = _mm_add_epi32(sumi2_0, _mm_add_epi32(p3_0, p4_0));
-            sumi2_1 = _mm_add_epi32(sumi2_1, _mm_add_epi32(p3_1, p4_1));
-
-            qs += 8; qh += 4;
-        }
-
-        const __m256 d = _mm256_set1_ps(y[i].d * GGML_FP16_TO_FP32(scale.f16));
-
-        accum1 = _mm256_add_ps(_mm256_mul_ps(d, _mm256_cvtepi32_ps(MM256_SET_M128I(sumi1_1, sumi1_0))), accum1);
-        accum2 = _mm256_add_ps(_mm256_mul_ps(d, _mm256_cvtepi32_ps(MM256_SET_M128I(sumi2_1, sumi2_0))), accum2);
-    }
-
-    *s = hsum_float_8(accum1) + IQ1M_DELTA * hsum_float_8(accum2);
-
-#else
-
-    int sum1[2], sum2[2], delta[4];
-
-    float sumf = 0;
-    for (int i = 0; i < nb; i++) {
-
-        const int8_t   * q8 = y[i].qs;
-        const uint8_t  * qs = x[i].qs;
-        const uint8_t  * qh = x[i].qh;
-        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);
-
-        int sumi1 = 0, sumi2 = 0;
-        for (int ib = 0; ib < QK_K/32; ++ib) {
-            delta[0] = qh[0] & 0x08 ? -1 : 1;
-            delta[1] = qh[0] & 0x80 ? -1 : 1;
-            delta[2] = qh[1] & 0x08 ? -1 : 1;
-            delta[3] = qh[1] & 0x80 ? -1 : 1;
-            sum1[0] = sum1[1] = sum2[0] = sum2[1] = 0;
-            for (int l = 0; l < 4; ++l) {
-                const int8_t * grid = (const int8_t *)(iq1s_grid + (qs[l] | (((uint16_t)qh[l/2] << (8 - 4*(l%2))) & 0x700)));
-                int lsum1 = 0, lsum2 = 0;
-                for (int j = 0; j < 8; ++j) {
-                    lsum1 += q8[j] * grid[j];
-                    lsum2 += q8[j];
-                }
-                q8 += 8;
-                sum1[l/2] += lsum1;
-                sum2[l/2] += lsum2*delta[l];
-            }
-
-            const int ls1 = 2*((sc[ib/2] >> (6*(ib%2)+0)) & 0x7) + 1;
-            const int ls2 = 2*((sc[ib/2] >> (6*(ib%2)+3)) & 0x7) + 1;
-
-            sumi1 += sum1[0] * ls1 + sum1[1] * ls2;
-            sumi2 += sum2[0] * ls1 + sum2[1] * ls2;
-            qs += 4;
-            qh += 2;
-        }
-
-        sumf += GGML_FP16_TO_FP32(scale.f16) * y[i].d * (sumi1 + IQ1M_DELTA * sumi2);
-    }
-
-    *s = sumf;
-
-#endif
-}
-
-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);
-    UNUSED(by);
-    UNUSED(bs);
-    assert(n % QK4_NL == 0);
-    static_assert(QK4_NL == QK8_0, "QK4_NL and QK8_0 must be the same");
-
-    const block_iq4_nl * GGML_RESTRICT x = vx;
-    const block_q8_0   * GGML_RESTRICT y = vy;
-
-    const int nb = n / QK4_NL;
-
-    int ib = 0;
-    float sumf = 0;
-
-#if defined __ARM_NEON
-    const int8x16_t values = vld1q_s8(kvalues_iq4nl);
-    const uint8x16_t m4b = vdupq_n_u8(0x0f);
-    uint8x16x2_t q4bits;
-    int8x16x4_t q4b;
-    int8x16x4_t q8b;
-    int32x4_t prod_1, prod_2;
-
-    for (; ib + 1 < nb; ib += 2) {
-
-        q4bits.val[0] = vld1q_u8(x[ib + 0].qs);
-        q4bits.val[1] = vld1q_u8(x[ib + 1].qs);
-        q8b.val[0]    = vld1q_s8(y[ib + 0].qs);
-        q8b.val[1]    = vld1q_s8(y[ib + 0].qs + 16);
-        q8b.val[2]    = vld1q_s8(y[ib + 1].qs);
-        q8b.val[3]    = vld1q_s8(y[ib + 1].qs + 16);
-
-        q4b.val[0] = ggml_vqtbl1q_s8(values, vandq_u8  (q4bits.val[0], m4b));
-        q4b.val[1] = ggml_vqtbl1q_s8(values, vshrq_n_u8(q4bits.val[0], 4));
-        q4b.val[2] = ggml_vqtbl1q_s8(values, vandq_u8  (q4bits.val[1], m4b));
-        q4b.val[3] = ggml_vqtbl1q_s8(values, vshrq_n_u8(q4bits.val[1], 4));
-
-        prod_1 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), q4b.val[0], q8b.val[0]), q4b.val[1], q8b.val[1]);
-        prod_2 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), q4b.val[2], q8b.val[2]), q4b.val[3], q8b.val[3]);
-
-        sumf +=
-            GGML_FP16_TO_FP32(x[ib+0].d) * GGML_FP16_TO_FP32(y[ib + 0].d) * vaddvq_s32(prod_1) +
-            GGML_FP16_TO_FP32(x[ib+1].d) * GGML_FP16_TO_FP32(y[ib + 1].d) * vaddvq_s32(prod_2);
-    }
-
-#elif defined __AVX2__
-
-    const __m128i values128 = _mm_loadu_si128((const __m128i*)kvalues_iq4nl);
-    const __m128i m4b  = _mm_set1_epi8(0x0f);
-    const __m256i mone = _mm256_set1_epi16(1);
-
-    __m256 accum1 = _mm256_setzero_ps();
-    __m256 accum2 = _mm256_setzero_ps();
-    for (; ib + 1 < nb; ib += 2) {
-        const __m128i q4bits_1 = _mm_loadu_si128((const __m128i*)x[ib + 0].qs);
-        const __m128i q4bits_2 = _mm_loadu_si128((const __m128i*)x[ib + 1].qs);
-        const __m256i q8b_1 = _mm256_loadu_si256((const __m256i *)y[ib + 0].qs);
-        const __m256i q8b_2 = _mm256_loadu_si256((const __m256i *)y[ib + 1].qs);
-        const __m256i q4b_1 = MM256_SET_M128I(_mm_shuffle_epi8(values128, _mm_and_si128(_mm_srli_epi16(q4bits_1, 4), m4b)),
-                                              _mm_shuffle_epi8(values128, _mm_and_si128(q4bits_1, m4b)));
-        const __m256i q4b_2 = MM256_SET_M128I(_mm_shuffle_epi8(values128, _mm_and_si128(_mm_srli_epi16(q4bits_2, 4), m4b)),
-                                              _mm_shuffle_epi8(values128, _mm_and_si128(q4bits_2, m4b)));
-        const __m256i p16_1 = mul_add_epi8(q4b_1, q8b_1);
-        const __m256i p16_2 = mul_add_epi8(q4b_2, q8b_2);
-        const __m256i p_1 = _mm256_madd_epi16(p16_1, mone);
-        const __m256i p_2 = _mm256_madd_epi16(p16_2, mone);
-        accum1 = _mm256_fmadd_ps(_mm256_set1_ps(GGML_FP16_TO_FP32(y[ib + 0].d)*GGML_FP16_TO_FP32(x[ib + 0].d)),
-                _mm256_cvtepi32_ps(p_1), accum1);
-        accum2 = _mm256_fmadd_ps(_mm256_set1_ps(GGML_FP16_TO_FP32(y[ib + 1].d)*GGML_FP16_TO_FP32(x[ib + 1].d)),
-                _mm256_cvtepi32_ps(p_2), accum2);
-    }
-
-    sumf = hsum_float_8(_mm256_add_ps(accum1, accum2));
-
-#elif defined __AVX__
-    const __m128i values128 = _mm_loadu_si128((const __m128i*)kvalues_iq4nl);
-    const __m128i m4b  = _mm_set1_epi8(0x0f);
-
-    __m256 accum = _mm256_setzero_ps();
-    for (; ib + 1 < nb; ib += 2) {
-        const __m128i q4bits_1 = _mm_loadu_si128((const __m128i *)x[ib + 0].qs);
-        const __m128i q4bits_2 = _mm_loadu_si128((const __m128i *)x[ib + 1].qs);
-        const __m128i q8b_1_0 = _mm_loadu_si128((const __m128i *)y[ib + 0].qs);
-        const __m128i q8b_1_1 = _mm_loadu_si128((const __m128i *)y[ib + 0].qs + 1);
-        const __m128i q8b_2_0 = _mm_loadu_si128((const __m128i *)y[ib + 1].qs);
-        const __m128i q8b_2_1 = _mm_loadu_si128((const __m128i *)y[ib + 1].qs + 1);
-
-        const __m128i q4b_1_0 = _mm_shuffle_epi8(values128, _mm_and_si128(q4bits_1, m4b));
-        const __m128i q4b_1_1 = _mm_shuffle_epi8(values128, _mm_and_si128(_mm_srli_epi16(q4bits_1, 4), m4b));
-        const __m128i q4b_2_0 = _mm_shuffle_epi8(values128, _mm_and_si128(q4bits_2, m4b));
-        const __m128i q4b_2_1 = _mm_shuffle_epi8(values128, _mm_and_si128(_mm_srli_epi16(q4bits_2, 4), m4b));
-
-        const __m256 p = mul_sum_i8_quad_float(q4b_1_0, q4b_1_1, q4b_2_0, q4b_2_1, q8b_1_0, q8b_1_1, q8b_2_0, q8b_2_1);
-        const __m256 deltas = quad_fp16_delta_float(x[ib].d, y[ib].d, x[ib + 1].d, y[ib + 1].d);
-        accum = _mm256_add_ps(_mm256_mul_ps(deltas, p), accum);
-    }
-
-    sumf = hsum_float_8(accum);
-
-#elif defined(__POWER9_VECTOR__)
-    const vector signed char lowMask = vec_splats((signed char)0xF);
-    const vector signed int v0 = vec_splats((int32_t)0);
-    const vector unsigned char v4 = vec_splats((unsigned char)0x4);
-
-    vector float vsumf0 = vec_splats(0.0f);
-    vector float vsumf1 = vec_splats(0.0f);
-
-    const vector signed char values = vec_xl( 0, kvalues_iq4nl);
-
-#pragma GCC unroll 4
-    for (; ib < nb; ++ib) {
-        __builtin_prefetch(x[ib].qs, 0, 1);
-        __builtin_prefetch(y[ib].qs, 0, 1);
-
-
-        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[ib].d));
-        vector float vyd = vec_splats(GGML_FP16_TO_FP32(y[ib].d));
-        vector float vd = vec_mul(vxd, vyd);
-
-        vector signed char qxs = (vector signed char)vec_xl( 0, x[ib].qs);
-        vector signed char q4x0 = vec_and(qxs, lowMask);
-        vector signed char q4x1 = vec_sr(qxs, v4);
-
-        q4x0 = vec_perm(values, values, (vector unsigned char)q4x0);
-        q4x1 = vec_perm(values, values, (vector unsigned char)q4x1);
-
-        vector signed char q8y0 = vec_xl( 0, y[ib].qs);
-        vector signed char q8y1 = vec_xl(16, y[ib].qs);
-
-        vector signed short qv0 = vec_add(vec_mule(q4x0, q8y0), vec_mulo(q4x0, q8y0));
-        vector signed short qv1 = vec_add(vec_mule(q4x1, q8y1), vec_mulo(q4x1, q8y1));
-
-        vector signed int vsumi0 = v0;
-        vector signed int vsumi1 = v0;
-
-        vsumi0 = vec_sum4s(qv0, vsumi0);
-        vsumi1 = vec_sum4s(qv1, vsumi1);
-
-        vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
-        vsumf1 = vec_madd(vec_ctf(vsumi1, 0), vd, vsumf1);
-    }
-
-    vsumf0 = vec_add(vsumf0, vsumf1);
-
-    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4));
-    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8));
-
-    sumf = vec_extract(vsumf0, 0);
-
-#elif defined (__loongarch_asx)
-
-    const __m128i values128 = __lsx_vld((const __m128i*)kvalues_iq4nl, 0);
-    const __m128i m4b  = __lsx_vreplgr2vr_b(0x0f);
-    const __m256i mone = __lasx_xvreplgr2vr_h(1);
-
-    __m256 accum1 = (__m256)__lasx_xvldi(0);
-    __m256 accum2 = (__m256)__lasx_xvldi(0);
-    for (; ib + 1 < nb; ib += 2) {
-        const __m128i q4bits_1 = __lsx_vld((const __m128i*)x[ib + 0].qs, 0);
-        const __m128i q4bits_2 = __lsx_vld((const __m128i*)x[ib + 1].qs, 0);
-        const __m256i q8b_1 = __lasx_xvld((const __m256i *)y[ib + 0].qs, 0);
-        const __m256i q8b_2 = __lasx_xvld((const __m256i *)y[ib + 1].qs, 0);
-        const __m256i q4b_1 = lasx_insertf128(lsx_shuffle_b(values128, __lsx_vand_v(__lsx_vsrli_h(q4bits_1, 4), m4b)),
-                                              lsx_shuffle_b(values128, __lsx_vand_v(q4bits_1, m4b)));
-        const __m256i q4b_2 = lasx_insertf128(lsx_shuffle_b(values128, __lsx_vand_v(__lsx_vsrli_h(q4bits_2, 4), m4b)),
-                                              lsx_shuffle_b(values128, __lsx_vand_v(q4bits_2, m4b)));
-        const __m256i p16_1 = mul_add_epi8(q4b_1, q8b_1);
-        const __m256i p16_2 = mul_add_epi8(q4b_2, q8b_2);
-        const __m256i p_1 = lasx_madd_h(p16_1, mone);
-        const __m256i p_2 = lasx_madd_h(p16_2, mone);
-        accum1 = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(GGML_FP16_TO_FP32(y[ib + 0].d)*GGML_FP16_TO_FP32(x[ib + 0].d)),
-                __lasx_xvffint_s_w(p_1), accum1);
-        accum2 = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(GGML_FP16_TO_FP32(y[ib + 1].d)*GGML_FP16_TO_FP32(x[ib + 1].d)),
-                __lasx_xvffint_s_w(p_2), accum2);
-    }
-
-    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);
-        int sumi1 = 0, sumi2 = 0;
-        for (int j = 0; j < QK4_NL/2; ++j) {
-            sumi1 += y[ib].qs[j+       0] * kvalues_iq4nl[x[ib].qs[j] & 0xf];
-            sumi2 += y[ib].qs[j+QK4_NL/2] * kvalues_iq4nl[x[ib].qs[j] >>  4];
-        }
-        sumf += d * (sumi1 + sumi2);
-    }
-    *s = sumf;
-}
-
-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);
-    UNUSED(by);
-    UNUSED(bs);
-    assert(n % QK_K == 0);
-
-    const block_iq4_xs * GGML_RESTRICT x = vx;
-    const block_q8_K   * GGML_RESTRICT y = vy;
-
-    const int nb = n / QK_K;
-
-#if defined __ARM_NEON
-    const int8x16_t values = vld1q_s8(kvalues_iq4nl);
-    const uint8x16_t m4b = vdupq_n_u8(0x0f);
-    ggml_uint8x16x2_t q4bits;
-    ggml_int8x16x4_t q4b;
-    ggml_int8x16x4_t q8b;
-    int32x4_t prod_1, prod_2;
-
-    float sumf = 0;
-
-    for (int ibl = 0; ibl < nb; ++ibl) {
-
-        const int8_t  * q8 = y[ibl].qs;
-        const uint8_t * q4 = x[ibl].qs;
-        uint16_t h = x[ibl].scales_h;
-
-        int sumi1 = 0, sumi2 = 0;
-        for (int ib = 0; ib < QK_K/64; ++ib) {
-
-            q4bits = ggml_vld1q_u8_x2(q4); q4 += 32;
-            q8b    = ggml_vld1q_s8_x4(q8); q8 += 64;
-
-            q4b.val[0] = ggml_vqtbl1q_s8(values, vandq_u8  (q4bits.val[0], m4b));
-            q4b.val[1] = ggml_vqtbl1q_s8(values, vshrq_n_u8(q4bits.val[0], 4));
-            q4b.val[2] = ggml_vqtbl1q_s8(values, vandq_u8  (q4bits.val[1], m4b));
-            q4b.val[3] = ggml_vqtbl1q_s8(values, vshrq_n_u8(q4bits.val[1], 4));
-
-            prod_1 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), q4b.val[0], q8b.val[0]), q4b.val[1], q8b.val[1]);
-            prod_2 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), q4b.val[2], q8b.val[2]), q4b.val[3], q8b.val[3]);
-
-            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 += vaddvq_s32(prod_1) * ls1;
-            sumi2 += vaddvq_s32(prod_2) * ls2;
-
-        }
-
-        sumf += GGML_FP16_TO_FP32(x[ibl].d) * y[ibl].d * (sumi1 + sumi2);
-    }
-
-    *s = sumf;
-
-#elif defined __AVX2__
-
-    const __m128i values128 = _mm_loadu_si128((const __m128i*)kvalues_iq4nl);
-    const __m128i m4b  = _mm_set1_epi8(0x0f);
-
-    __m256 accum = _mm256_setzero_ps();
-    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 = _mm256_setzero_si256();
-        __m256i sumi2 = _mm256_setzero_si256();
-        for (int ib = 0; ib < QK_K/32; ib += 2) {
-            const __m128i q4bits_1 = _mm_loadu_si128((const __m128i*)qs);  qs += 16;
-            const __m128i q4bits_2 = _mm_loadu_si128((const __m128i*)qs);  qs += 16;
-            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 q4b_1 = MM256_SET_M128I(_mm_shuffle_epi8(values128, _mm_and_si128(_mm_srli_epi16(q4bits_1, 4), m4b)),
-                                                  _mm_shuffle_epi8(values128, _mm_and_si128(q4bits_1, m4b)));
-            const __m256i q4b_2 = MM256_SET_M128I(_mm_shuffle_epi8(values128, _mm_and_si128(_mm_srli_epi16(q4bits_2, 4), m4b)),
-                                                  _mm_shuffle_epi8(values128, _mm_and_si128(q4bits_2, m4b)));
-            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;
-            const __m256i p_1 = _mm256_madd_epi16(p16_1, _mm256_set1_epi16(ls1));
-            const __m256i p_2 = _mm256_madd_epi16(p16_2, _mm256_set1_epi16(ls2));
-            sumi1 = _mm256_add_epi32(p_1, sumi1);
-            sumi2 = _mm256_add_epi32(p_2, sumi2);
-        }
-        accum = _mm256_fmadd_ps(_mm256_set1_ps(GGML_FP16_TO_FP32(x[ibl].d)*y[ibl].d),
-                _mm256_cvtepi32_ps(_mm256_add_epi32(sumi1, sumi2)), accum);
-    }
-
-    *s = hsum_float_8(accum);
-
-#elif defined __AVX__
-    const __m128i values128 = _mm_loadu_si128((const __m128i*)kvalues_iq4nl);
-    const __m128i m4b  = _mm_set1_epi8(0x0f);
-
-    __m256 accum = _mm256_setzero_ps();
-    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;
-        __m128i sumi1_0 = _mm_setzero_si128();
-        __m128i sumi1_1 = _mm_setzero_si128();
-        __m128i sumi2_0 = _mm_setzero_si128();
-        __m128i sumi2_1 = _mm_setzero_si128();
-        for (int ib = 0; ib < QK_K/32; ib += 2) {
-            const __m128i q4bits_1 = _mm_loadu_si128((const __m128i *)qs); qs += 16;
-            const __m128i q4bits_2 = _mm_loadu_si128((const __m128i *)qs); qs += 16;
-            const __m128i q8b_1_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
-            const __m128i q8b_1_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
-            const __m128i q8b_2_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
-            const __m128i q8b_2_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
-            const __m128i q4b_1_0 = _mm_shuffle_epi8(values128, _mm_and_si128(q4bits_1, m4b));
-            const __m128i q4b_1_1 = _mm_shuffle_epi8(values128, _mm_and_si128(_mm_srli_epi16(q4bits_1, 4), m4b));
-            const __m128i q4b_2_0 = _mm_shuffle_epi8(values128, _mm_and_si128(q4bits_2, m4b));
-            const __m128i q4b_2_1 = _mm_shuffle_epi8(values128, _mm_and_si128(_mm_srli_epi16(q4bits_2, 4), m4b));
-            const __m128i p16_1_0 = mul_add_epi8_sse(q4b_1_0, q8b_1_0);
-            const __m128i p16_1_1 = mul_add_epi8_sse(q4b_1_1, q8b_1_1);
-            const __m128i p16_2_0 = mul_add_epi8_sse(q4b_2_0, q8b_2_0);
-            const __m128i p16_2_1 = mul_add_epi8_sse(q4b_2_1, q8b_2_1);
-            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;
-            const __m128i p_1_0 = _mm_madd_epi16(p16_1_0, _mm_set1_epi16(ls1));
-            const __m128i p_1_1 = _mm_madd_epi16(p16_1_1, _mm_set1_epi16(ls1));
-            const __m128i p_2_0 = _mm_madd_epi16(p16_2_0, _mm_set1_epi16(ls2));
-            const __m128i p_2_1 = _mm_madd_epi16(p16_2_1, _mm_set1_epi16(ls2));
-            sumi1_0 = _mm_add_epi32(p_1_0, sumi1_0);
-            sumi1_1 = _mm_add_epi32(p_1_1, sumi1_1);
-            sumi2_0 = _mm_add_epi32(p_2_0, sumi2_0);
-            sumi2_1 = _mm_add_epi32(p_2_1, sumi2_1);
-        }
-        __m128i sumi12_0 = _mm_add_epi32(sumi1_0, sumi2_0);
-        __m128i sumi12_1 = _mm_add_epi32(sumi1_1, sumi2_1);
-        accum = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(GGML_FP16_TO_FP32(x[ibl].d)*y[ibl].d),
-                _mm256_cvtepi32_ps(MM256_SET_M128I(sumi12_1, sumi12_0))), accum);
-    }
-
-    *s = hsum_float_8(accum);
-
-#elif defined(__POWER9_VECTOR__)
-    const vector signed char lowMask = vec_splats((signed char)0xF);
-    const vector int v0 = vec_splats((int32_t)0);
-    const vector unsigned char v4 = vec_splats((unsigned char)0x4);
-
-    vector float vsumf0 = vec_splats(0.0f);
-    vector float vsumf1 = vec_splats(0.0f);
-    vector float vsumf2 = vec_splats(0.0f);
-    vector float vsumf3 = vec_splats(0.0f);
-
-    const vector signed char values = vec_xl( 0, kvalues_iq4nl);
-
-    for (int ibl = 0; ibl < nb; ++ibl) {
-
-        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[ibl].d));
-        vector float vyd = vec_splats(y[ibl].d);
-        vector float vd = vec_mul(vxd, vyd);
-
-        vector signed int vsumi0 = v0;
-        vector signed int vsumi1 = v0;
-        vector signed int vsumi2 = v0;
-        vector signed int vsumi3 = v0;
-
-        uint16_t h = x[ibl].scales_h;
-
-        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);
-            __builtin_prefetch(q8, 0, 1);
-
-            vector signed char qxs0 = (vector signed char)vec_xl( 0, q4);
-            vector signed char qxs1 = (vector signed char)vec_xl(16, q4);
-            q4 += 32;
-
-            vector signed char q4x00 = (vector signed char)vec_and(qxs0, lowMask);
-            vector signed char q4x01 = (vector signed char)vec_sr(qxs0, v4);
-            vector signed char q4x10 = (vector signed char)vec_and(qxs1, lowMask);
-            vector signed char q4x11 = (vector signed char)vec_sr(qxs1, v4);
-
-            q4x00 = vec_perm(values, values, (vector unsigned char)q4x00);
-            q4x01 = vec_perm(values, values, (vector unsigned char)q4x01);
-            q4x10 = vec_perm(values, values, (vector unsigned char)q4x10);
-            q4x11 = vec_perm(values, values, (vector unsigned char)q4x11);
-
-            vector signed char q8y0 = vec_xl( 0, q8);
-            vector signed char q8y1 = vec_xl(16, q8);
-            vector signed char q8y2 = vec_xl(32, q8);
-            vector signed char q8y3 = vec_xl(48, q8);
-            q8 += 64;
-
-            vector signed short qv0 = vec_add(vec_mule(q4x00, q8y0), vec_mulo(q4x00, q8y0));
-            vector signed short qv1 = vec_add(vec_mule(q4x01, q8y1), vec_mulo(q4x01, q8y1));
-            vector signed short qv2 = vec_add(vec_mule(q4x10, q8y2), vec_mulo(q4x10, q8y2));
-            vector signed short qv3 = vec_add(vec_mule(q4x11, q8y3), vec_mulo(q4x11, q8y3));
-
-            const uint16_t ls0 = (uint16_t)(((sc[0] & 0xf) | ((h << 4) & 0x30)) - 32);
-            const uint16_t ls1 = (uint16_t)(((sc[0] >>  4) | ((h << 2) & 0x30)) - 32);
-            h >>= 4;
-            sc ++;
-
-            vector signed short vscales01 = vec_splats((int16_t)ls0);
-            vector signed short vscales23 = vec_splats((int16_t)ls1);
-
-            vsumi0 = vec_msum(qv0, vscales01, vsumi0);
-            vsumi1 = vec_msum(qv1, vscales01, vsumi1);
-            vsumi2 = vec_msum(qv2, vscales23, vsumi2);
-            vsumi3 = vec_msum(qv3, vscales23, vsumi3);
-        }
-
-        vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
-        vsumf1 = vec_madd(vec_ctf(vsumi1, 0), vd, vsumf1);
-        vsumf2 = vec_madd(vec_ctf(vsumi2, 0), vd, vsumf2);
-        vsumf3 = vec_madd(vec_ctf(vsumi3, 0), vd, vsumf3);
-    }
-
-    vsumf0 = vec_add(vsumf0, vsumf2);
-    vsumf1 = vec_add(vsumf1, vsumf3);
-
-    vsumf0 = vec_add(vsumf0, vsumf1);
-
-    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4));
-    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8));
-
-    *s = vec_extract(vsumf0, 0);
-
-#elif defined(__loongarch_asx)
-
-    const __m128i values128 = __lsx_vld((const __m128i*)kvalues_iq4nl, 0);
-
-    __m256 accum = (__m256)__lasx_xvldi(0);
-
-    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);
-        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 __m256i q8b_1 = __lasx_xvld((const __m256i *)q8, 0); q8 += 32;
-            const __m256i q8b_2 = __lasx_xvld((const __m256i *)q8, 0); q8 += 32;
-            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;
-            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);
-        }
-        accum = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(GGML_FP16_TO_FP32(x[ibl].d)*y[ibl].d),
-                __lasx_xvffint_s_w(__lasx_xvadd_w(sumi1, sumi2)), accum);
-    }
-
-    *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;
-    for (int ibl = 0; ibl < nb; ++ibl) {
-        const float d4d8 = GGML_FP16_TO_FP32(x[ibl].d) * y[ibl].d;
-        uint16_t h = x[ibl].scales_h;
-        const uint8_t * qs = x[ibl].qs;
-        const int8_t  * q8 = y[ibl].qs;
-        for (int ib = 0; ib < QK_K/32; ib += 2) {
-            const uint8_t ls1 = (x[ibl].scales_l[ib/2] & 0xf) | ((h << 4) & 0x30);
-            const uint8_t ls2 = (x[ibl].scales_l[ib/2] >>  4) | ((h << 2) & 0x30);
-            h >>= 4;
-            const float d1 = d4d8*(ls1 - 32);
-            const float d2 = d4d8*(ls2 - 32);
-            int sumi1 = 0, sumi2 = 0;
-            for (int j = 0; j < 16; ++j) {
-                sumi1 += q8[j+ 0] * kvalues_iq4nl[qs[j] & 0xf];
-                sumi2 += q8[j+16] * kvalues_iq4nl[qs[j] >>  4];
-            }
-            sumf += d1 * (sumi1 + sumi2);
-            qs += 16;
-            q8 += 32;
-            sumi1 = sumi2 = 0;
-            for (int j = 0; j < 16; ++j) {
-                sumi1 += q8[j+ 0] * kvalues_iq4nl[qs[j] & 0xf];
-                sumi2 += q8[j+16] * kvalues_iq4nl[qs[j] >>  4];
-            }
-            sumf += d2 * (sumi1 + sumi2);
-            qs += 16;
-            q8 += 32;
-        }
-    }
-    *s = sumf;
-#endif
-}
-
-// ============================ 4-bit non-linear quants
-
-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 * 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 c7426df2b..c5271b775 100644
--- a/ggml/src/ggml-cpu/ggml-cpu.c
+++ b/ggml/src/ggml-cpu/ggml-cpu.c
@@ -3,11 +3,11 @@
 
 #include "ggml-backend-impl.h"
 #include "ggml-backend.h"
-#include "ggml-cpu-traits.h"
+#include "traits.h"
 #include "ggml-cpu-impl.h"
 #include "ggml-cpu.h"
 #include "ggml-impl.h"
-#include "ggml-cpu-quants.h"
+#include "quants.h"
 #include "ggml-threading.h"
 #include "unary-ops.h"
 #include "binary-ops.h"
@@ -72,15 +72,13 @@
 #define UNUSED GGML_UNUSED
 #define SWAP(x, y, T) do { T SWAP = x; (x) = y; (y) = SWAP; } while (0)
 
+// precomputed f32 table for f16 (256 KB) (simd-mappings.h)
+float ggml_table_f32_f16[1 << 16];
+
 #if defined(__ARM_ARCH)
 struct ggml_arm_arch_features_type {
-    int has_neon;
-    int has_dotprod;
-    int has_i8mm;
-    int has_sve;
     int sve_cnt;
-    int has_sme;
-} ggml_arm_arch_features = {-1, -1, -1, -1, 0, -1};
+} ggml_arm_arch_features = { 0 };
 #endif
 
 
@@ -197,6 +195,7 @@ typedef pthread_t ggml_thread_t;
 
 static const struct ggml_type_traits_cpu type_traits_cpu[GGML_TYPE_COUNT] = {
     [GGML_TYPE_F32] = {
+        .from_float               = (ggml_from_float_t) ggml_cpu_fp32_to_fp32,
         .vec_dot                  = (ggml_vec_dot_t) ggml_vec_dot_f32,
         .vec_dot_type             = GGML_TYPE_F32,
         .nrows                    = 1,
@@ -559,6 +558,14 @@ void ggml_barrier(struct ggml_threadpool * tp) {
 #endif
 }
 
+void ggml_threadpool_chunk_set(struct ggml_threadpool * tp, int value) {
+    atomic_store_explicit(&tp->current_chunk, value, memory_order_relaxed);
+}
+
+int ggml_threadpool_chunk_add(struct ggml_threadpool * tp, int value) {
+    return atomic_fetch_add_explicit(&tp->current_chunk, value, memory_order_relaxed);
+}
+
 #if defined(__gnu_linux__)
 static cpu_set_t ggml_get_numa_affinity(void) {
     cpu_set_t cpuset;
@@ -670,87 +677,15 @@ bool ggml_is_numa(void) {
 
 #if defined(__linux__) && defined(__aarch64__)
 #include <sys/auxv.h>
-#elif defined(__APPLE__)
-#include <sys/sysctl.h>
-#endif
-
-#if !defined(HWCAP2_I8MM)
-#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_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_sme     = !!(hwcap2 & HWCAP2_SME);
-
-#if defined(__ARM_FEATURE_SVE)
+#if defined(__linux__) && defined(__aarch64__) && defined(__ARM_FEATURE_SVE)
     ggml_arm_arch_features.sve_cnt = PR_SVE_VL_LEN_MASK & prctl(PR_SVE_GET_VL);
 #endif
-#elif defined(__APPLE__)
-    int oldp = 0;
-    size_t size = sizeof(oldp);
-    if (sysctlbyname("hw.optional.AdvSIMD", &oldp, &size, NULL, 0) != 0) {
-        oldp = 0;
-    }
-    ggml_arm_arch_features.has_neon = oldp;
-
-    if (sysctlbyname("hw.optional.arm.FEAT_DotProd", &oldp, &size, NULL, 0) != 0) {
-        oldp = 0;
-    }
-    ggml_arm_arch_features.has_dotprod = oldp;
-
-    if (sysctlbyname("hw.optional.arm.FEAT_I8MM", &oldp, &size, NULL, 0) != 0) {
-        oldp = 0;
-    }
-    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
-// Run-time CPU feature detection not implemented for this platform, fallback to compile time
-#if defined(__ARM_NEON)
-    ggml_arm_arch_features.has_neon = 1;
-#else
-    ggml_arm_arch_features.has_neon = 0;
-#endif
-
-#if defined(__ARM_FEATURE_MATMUL_INT8)
-    ggml_arm_arch_features.has_i8mm = 1;
-#else
-    ggml_arm_arch_features.has_i8mm = 0;
-#endif
-
-#if defined(__ARM_FEATURE_SVE)
-    ggml_arm_arch_features.has_sve = 1;
-    ggml_arm_arch_features.sve_cnt = 16;
-#else
-    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
+
+#endif // __ARM_ARCH
 
 struct ggml_tensor * ggml_new_i32(struct ggml_context * ctx, int32_t value) {
     GGML_ASSERT(!ggml_get_no_alloc(ctx));
@@ -805,7 +740,7 @@ struct ggml_tensor * ggml_set_i32 (struct ggml_tensor * tensor, int32_t value) {
             {
                 assert(tensor->nb[0] == sizeof(ggml_fp16_t));
                 for (int i = 0; i < n; i++) {
-                    ggml_vec_set_f16(nc, (ggml_fp16_t *)(data + i*n1), GGML_FP32_TO_FP16(value));
+                    ggml_vec_set_f16(nc, (ggml_fp16_t *)(data + i*n1), GGML_CPU_FP32_TO_FP16(value));
                 }
             } break;
         case GGML_TYPE_BF16:
@@ -864,7 +799,7 @@ struct ggml_tensor * ggml_set_f32(struct ggml_tensor * tensor, float value) {
             {
                 assert(tensor->nb[0] == sizeof(ggml_fp16_t));
                 for (int i = 0; i < n; i++) {
-                    ggml_vec_set_f16(nc, (ggml_fp16_t *)(data + i*n1), GGML_FP32_TO_FP16(value));
+                    ggml_vec_set_f16(nc, (ggml_fp16_t *)(data + i*n1), GGML_CPU_FP32_TO_FP16(value));
                 }
             } break;
         case GGML_TYPE_BF16:
@@ -915,7 +850,7 @@ int32_t ggml_get_i32_1d(const struct ggml_tensor * tensor, int i) {
         case GGML_TYPE_F16:
             {
                 GGML_ASSERT(tensor->nb[0] == sizeof(ggml_fp16_t));
-                return GGML_FP16_TO_FP32(((ggml_fp16_t *)(tensor->data))[i]);
+                return GGML_CPU_FP16_TO_FP32(((ggml_fp16_t *)(tensor->data))[i]);
             }
         case GGML_TYPE_BF16:
             {
@@ -960,7 +895,7 @@ void ggml_set_i32_1d(const struct ggml_tensor * tensor, int i, int32_t value) {
         case GGML_TYPE_F16:
             {
                 GGML_ASSERT(tensor->nb[0] == sizeof(ggml_fp16_t));
-                ((ggml_fp16_t *)(tensor->data))[i] = GGML_FP32_TO_FP16(value);
+                ((ggml_fp16_t *)(tensor->data))[i] = GGML_CPU_FP32_TO_FP16(value);
             } break;
         case GGML_TYPE_BF16:
             {
@@ -989,7 +924,7 @@ int32_t ggml_get_i32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i
         case GGML_TYPE_I32:
             return ((int32_t *) data)[0];
         case GGML_TYPE_F16:
-            return GGML_FP16_TO_FP32(((ggml_fp16_t *) data)[0]);
+            return GGML_CPU_FP16_TO_FP32(((ggml_fp16_t *) data)[0]);
         case GGML_TYPE_BF16:
             return GGML_BF16_TO_FP32(((ggml_bf16_t *) data)[0]);
         case GGML_TYPE_F32:
@@ -1016,7 +951,7 @@ void ggml_set_i32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2,
             } break;
         case GGML_TYPE_F16:
             {
-                ((ggml_fp16_t *)(data))[0] = GGML_FP32_TO_FP16(value);
+                ((ggml_fp16_t *)(data))[0] = GGML_CPU_FP32_TO_FP16(value);
             } break;
         case GGML_TYPE_BF16:
             {
@@ -1054,7 +989,7 @@ float ggml_get_f32_1d(const struct ggml_tensor * tensor, int i) {
             }
         case GGML_TYPE_F16:
             {
-                return GGML_FP16_TO_FP32(((ggml_fp16_t *)(tensor->data))[i]);
+                return GGML_CPU_FP16_TO_FP32(((ggml_fp16_t *)(tensor->data))[i]);
             }
         case GGML_TYPE_BF16:
             {
@@ -1093,7 +1028,7 @@ void ggml_set_f32_1d(const struct ggml_tensor * tensor, int i, float value) {
             } break;
         case GGML_TYPE_F16:
             {
-                ((ggml_fp16_t *)(tensor->data))[i] = GGML_FP32_TO_FP16(value);
+                ((ggml_fp16_t *)(tensor->data))[i] = GGML_CPU_FP32_TO_FP16(value);
             } break;
         case GGML_TYPE_BF16:
             {
@@ -1120,7 +1055,7 @@ float ggml_get_f32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2,
         case GGML_TYPE_I32:
             return ((int32_t *) data)[0];
         case GGML_TYPE_F16:
-            return GGML_FP16_TO_FP32(((ggml_fp16_t *) data)[0]);
+            return GGML_CPU_FP16_TO_FP32(((ggml_fp16_t *) data)[0]);
         case GGML_TYPE_BF16:
             return GGML_BF16_TO_FP32(((ggml_bf16_t *) data)[0]);
         case GGML_TYPE_F32:
@@ -1147,7 +1082,7 @@ void ggml_set_f32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2,
             } break;
         case GGML_TYPE_F16:
             {
-                ((ggml_fp16_t *)(data))[0] = GGML_FP32_TO_FP16(value);
+                ((ggml_fp16_t *)(data))[0] = GGML_CPU_FP32_TO_FP16(value);
             } break;
         case GGML_TYPE_BF16:
             {
@@ -1258,7 +1193,7 @@ static void ggml_compute_forward_mul_mat_one_chunk(
     }
 }
 
-static void ggml_compute_forward_mul_mat(
+void ggml_compute_forward_mul_mat(
         const struct ggml_compute_params * params,
               struct ggml_tensor * dst) {
 
@@ -1883,6 +1818,10 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
             {
                 ggml_compute_forward_get_rows_back(params, tensor);
             } break;
+        case GGML_OP_SET_ROWS:
+            {
+                ggml_compute_forward_set_rows(params, tensor);
+            } break;
         case GGML_OP_DIAG:
             {
                 ggml_compute_forward_diag(params, tensor);
@@ -1927,6 +1866,10 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
             {
                 ggml_compute_forward_im2col_back_f32(params, tensor);
             } break;
+        case GGML_OP_CONV_2D:
+            {
+                ggml_compute_forward_conv_2d(params, tensor);
+            } break;
         case GGML_OP_CONV_2D_DW:
             {
                 ggml_compute_forward_conv_2d_dw(params, tensor);
@@ -1959,6 +1902,10 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
             {
                 ggml_compute_forward_pad_reflect_1d(params, tensor);
             } break;
+        case GGML_OP_ROLL:
+            {
+                ggml_compute_forward_roll(params, tensor);
+            } break;
         case GGML_OP_ARANGE:
             {
                 ggml_compute_forward_arange(params, tensor);
@@ -2006,6 +1953,10 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
             {
                 ggml_compute_forward_unary(params, tensor);
             } break;
+        case GGML_OP_GLU:
+            {
+                ggml_compute_forward_glu(params, tensor);
+            } break;
         case GGML_OP_GET_REL_POS:
             {
                 ggml_compute_forward_get_rel_pos(params, tensor);
@@ -2216,6 +2167,20 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
                     GGML_ABORT("fatal error");
             }
             break;
+        case GGML_OP_GLU:
+            switch (ggml_get_glu_op(node)) {
+                case GGML_GLU_OP_REGLU:
+                case GGML_GLU_OP_GEGLU:
+                case GGML_GLU_OP_SWIGLU:
+                case GGML_GLU_OP_GEGLU_ERF:
+                case GGML_GLU_OP_GEGLU_QUICK:
+                    {
+                        n_tasks = n_threads;
+                    } break;
+                default:
+                    GGML_ABORT("fatal error");
+            }
+            break;
         case GGML_OP_SILU_BACK:
         case GGML_OP_MUL:
         case GGML_OP_DIV:
@@ -2232,6 +2197,7 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
                 n_tasks = n_threads;
             } break;
         case GGML_OP_GET_ROWS:
+        case GGML_OP_SET_ROWS:
             {
                 // FIXME: get_rows can use additional threads, but the cost of launching additional threads
                 // decreases performance with GPU offloading
@@ -2268,6 +2234,7 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
             } break;
         case GGML_OP_IM2COL:
         case GGML_OP_IM2COL_BACK:
+        case GGML_OP_CONV_2D:
         case GGML_OP_CONV_2D_DW:
         case GGML_OP_CONV_TRANSPOSE_1D:
         case GGML_OP_CONV_TRANSPOSE_2D:
@@ -2283,6 +2250,7 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
         case GGML_OP_UPSCALE:
         case GGML_OP_PAD:
         case GGML_OP_PAD_REFLECT_1D:
+        case GGML_OP_ROLL:
         case GGML_OP_ARANGE:
         case GGML_OP_TIMESTEP_EMBEDDING:
         case GGML_OP_ARGSORT:
@@ -2785,6 +2753,10 @@ struct ggml_cplan ggml_graph_plan(
                             GGML_ABORT("fatal error");
                         }
                     } break;
+                case GGML_OP_CONV_2D:
+                    {
+                        cur = GGML_IM2COL_WORK_SIZE;
+                    } break;
                 case GGML_OP_CONV_TRANSPOSE_2D:
                     {
                         const int64_t ne00 = node->src[0]->ne[0]; // W
@@ -3185,6 +3157,10 @@ enum ggml_status ggml_graph_compute_with_ctx(struct ggml_context * ctx, struct g
     return ggml_graph_compute(cgraph, &cplan);
 }
 
+void ggml_cpu_fp32_to_fp32(const float * x, float * y, int64_t n) {
+    memcpy(y, x, n * sizeof(float));
+}
+
 void ggml_cpu_fp32_to_fp16(const float * x, ggml_fp16_t * y, int64_t n) {
     int64_t i = 0;
 #if defined(__F16C__)
@@ -3205,9 +3181,24 @@ void ggml_cpu_fp32_to_fp16(const float * x, ggml_fp16_t * y, int64_t n) {
         __m128i y_vec = _mm_cvtps_ph(x_vec, _MM_FROUND_TO_NEAREST_INT);
         _mm_storel_epi64((__m128i *)(y + i), y_vec);
     }
+#elif defined(__NNPA__)
+    for (; i + 7 < n; i += 8) {
+        float32x4_t v_xh = vec_xl(0, (const float *)(x + i + 0));
+        float32x4_t v_xl = vec_xl(0, (const float *)(x + i + 4));
+        uint16x8_t v_yd = vec_round_from_fp32(v_xh, v_xl, 0);
+        uint16x8_t v_y = vec_convert_to_fp16(v_yd, 0);
+        vec_xst(v_y, 0, (ggml_fp16_t *)(y + i));
+    }
+    for (; i + 3 < n; i += 4) {
+        float32x4_t v_x = vec_xl(0, (const float *)(x + i));
+        float32x4_t v_zero = vec_splats(0.0f);
+        uint16x8_t v_yd = vec_round_from_fp32(v_x, v_zero, 0);
+        uint16x8_t v_y = vec_convert_to_fp16(v_yd, 0);
+        vec_xst(v_y, 0, (ggml_fp16_t *)(y + i));
+    }
 #endif
     for (; i < n; ++i) {
-        y[i] = GGML_FP32_TO_FP16(x[i]);
+        y[i] = GGML_CPU_FP32_TO_FP16(x[i]);
     }
 }
 
@@ -3231,9 +3222,25 @@ void ggml_cpu_fp16_to_fp32(const ggml_fp16_t * x, float * y, int64_t n) {
         __m128 y_vec = _mm_cvtph_ps(x_vec);
         _mm_storeu_ps(y + i, y_vec);
     }
+#elif defined(__NNPA__)
+    for (; i + 7 < n; i += 8) {
+        uint16x8_t v_x = vec_xl(0, (const ggml_fp16_t *)(x + i));
+        uint16x8_t v_yd = vec_convert_from_fp16(v_x, 0);
+        float32x4_t v_yh = vec_extend_to_fp32_hi(v_yd, 0);
+        float32x4_t v_yl = vec_extend_to_fp32_lo(v_yd, 0);
+        vec_xst(v_yh, 0, (float *)(y + i + 0));
+        vec_xst(v_yl, 0, (float *)(y + i + 4));
+    }
+    for (; i + 3 < n; i += 4) {
+        uint16x8_t v_x = vec_xl(0, (const ggml_fp16_t *)(x + i));
+        uint16x8_t v_yd = vec_convert_from_fp16(v_x, 0);
+        float32x4_t v_yh = vec_extend_to_fp32_hi(v_yd, 0);
+        vec_xst(v_yh, 0, (float *)(y + i));
+    }
 #endif
+
     for (; i < n; ++i) {
-        y[i] = GGML_FP16_TO_FP32(x[i]);
+        y[i] = GGML_CPU_FP16_TO_FP32(x[i]);
     }
 }
 
@@ -3433,9 +3440,17 @@ int ggml_cpu_has_vxe(void) {
 #endif
 }
 
+int ggml_cpu_has_nnpa(void) {
+#if defined(GGML_NNPA)
+    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;
+    return 1;
 #else
     return 0;
 #endif
@@ -3443,7 +3458,7 @@ int ggml_cpu_has_neon(void) {
 
 int ggml_cpu_has_dotprod(void) {
 #if defined(__ARM_ARCH) && defined(__ARM_FEATURE_DOTPROD)
-    return ggml_arm_arch_features.has_dotprod;
+    return 1;
 #else
     return 0;
 #endif
@@ -3451,7 +3466,7 @@ int ggml_cpu_has_dotprod(void) {
 
 int ggml_cpu_has_sve(void) {
 #if defined(__ARM_ARCH) && defined(__ARM_FEATURE_SVE)
-    return ggml_arm_arch_features.has_sve;
+    return 1;
 #else
     return 0;
 #endif
@@ -3459,7 +3474,7 @@ int ggml_cpu_has_sve(void) {
 
 int ggml_cpu_has_matmul_int8(void) {
 #if defined(__ARM_ARCH) && defined(__ARM_FEATURE_MATMUL_INT8)
-    return ggml_arm_arch_features.has_i8mm;
+    return 1;
 #else
     return 0;
 #endif
@@ -3475,14 +3490,14 @@ int ggml_cpu_get_sve_cnt(void) {
 
 int ggml_cpu_has_sme(void) {
 #if defined(__ARM_ARCH) && defined(__ARM_FEATURE_SME)
-    return ggml_arm_arch_features.has_sme;
+    return 1;
 #else
     return 0;
 #endif
 }
 
 void ggml_cpu_init(void) {
-    // needed to initialize f16 tables
+    // needed to initialize ggml_time
     {
         struct ggml_init_params params = { 0, NULL, false };
         struct ggml_context * ctx = ggml_init(params);
@@ -3503,9 +3518,10 @@ void ggml_cpu_init(void) {
                     uint16_t u16;
                     ggml_fp16_t fp16;
                 } u = {i};
-                float f = GGML_FP16_TO_FP32(u.fp16);
-                ggml_table_gelu_f16[i] = GGML_FP32_TO_FP16(ggml_gelu_f32(f));
-                ggml_table_gelu_quick_f16[i] = GGML_FP32_TO_FP16(ggml_gelu_quick_f32(f));
+                float f = GGML_COMPUTE_FP16_TO_FP32(u.fp16);
+                ggml_table_f32_f16[i] = f;
+                ggml_table_gelu_f16[i] = GGML_CPU_FP32_TO_FP16(ggml_gelu_f32(f));
+                ggml_table_gelu_quick_f16[i] = GGML_CPU_FP32_TO_FP16(ggml_gelu_quick_f32(f));
             }
 
             const uint64_t t_end = ggml_time_us(); UNUSED(t_end);
diff --git a/ggml/src/ggml-cpu/ggml-cpu.cpp b/ggml/src/ggml-cpu/ggml-cpu.cpp
index e013e8b41..c9daa4c39 100644
--- a/ggml/src/ggml-cpu/ggml-cpu.cpp
+++ b/ggml/src/ggml-cpu/ggml-cpu.cpp
@@ -1,8 +1,8 @@
 #include "ggml-backend.h"
 #include "ggml-backend-impl.h"
 #include "ggml-cpu.h"
-#include "ggml-cpu-aarch64.h"
-#include "ggml-cpu-traits.h"
+#include "repack.h"
+#include "traits.h"
 #include "ggml-impl.h"
 #include "amx/amx.h"
 
@@ -11,7 +11,7 @@
 #include <vector>
 
 #ifdef GGML_USE_CPU_HBM
-#    include "ggml-cpu-hbm.h"
+#    include "hbm.h"
 #endif
 
 #ifdef GGML_USE_CPU_KLEIDIAI
@@ -51,9 +51,9 @@ std::vector<ggml_backend_buffer_type_t>& ggml_backend_cpu_get_extra_buffers_type
         }
 #endif
 
-#ifdef GGML_USE_CPU_AARCH64
-        if (ggml_backend_cpu_aarch64_buffer_type()) {
-            bufts.push_back(ggml_backend_cpu_aarch64_buffer_type());
+#ifdef GGML_USE_CPU_REPACK
+        if (ggml_backend_cpu_repack_buffer_type()) {
+            bufts.push_back(ggml_backend_cpu_repack_buffer_type());
         }
 #endif
 
@@ -416,6 +416,7 @@ static bool ggml_backend_cpu_device_supports_op(ggml_backend_dev_t dev, const st
 
     switch (op->op) {
         case GGML_OP_CPY:
+        case GGML_OP_SET_ROWS:
             return
                 op->type != GGML_TYPE_IQ3_XXS &&
                 op->type != GGML_TYPE_IQ3_S   &&
@@ -578,6 +579,9 @@ static ggml_backend_feature * ggml_backend_cpu_get_features(ggml_backend_reg_t r
         if (ggml_cpu_has_vxe()) {
             features.push_back({ "VXE", "1" });
         }
+        if (ggml_cpu_has_nnpa()) {
+            features.push_back({ "NNPA", "1" });
+        }
         if (ggml_cpu_has_wasm_simd()) {
             features.push_back({ "WASM_SIMD", "1" });
         }
@@ -596,8 +600,8 @@ static ggml_backend_feature * ggml_backend_cpu_get_features(ggml_backend_reg_t r
     #ifdef GGML_USE_CPU_KLEIDIAI
         features.push_back({ "KLEIDIAI", "1" });
     #endif
-    #ifdef GGML_USE_CPU_AARCH64
-        features.push_back({ "AARCH64_REPACK", "1" });
+    #ifdef GGML_USE_CPU_REPACK
+        features.push_back({ "REPACK", "1" });
     #endif
 
         features.push_back({ nullptr, nullptr });
diff --git a/ggml/src/ggml-cpu/ggml-cpu-hbm.cpp b/ggml/src/ggml-cpu/hbm.cpp
similarity index 98%
rename from ggml/src/ggml-cpu/ggml-cpu-hbm.cpp
rename to ggml/src/ggml-cpu/hbm.cpp
index fa8dea2af..a4073c15e 100644
--- a/ggml/src/ggml-cpu/ggml-cpu-hbm.cpp
+++ b/ggml/src/ggml-cpu/hbm.cpp
@@ -5,7 +5,7 @@
 #include "ggml-cpu.h"
 #include "ggml-impl.h"
 
-#include "ggml-cpu-hbm.h"
+#include "hbm.h"
 
 // buffer type HBM
 
diff --git a/ggml/src/ggml-cpu/ggml-cpu-hbm.h b/ggml/src/ggml-cpu/hbm.h
similarity index 100%
rename from ggml/src/ggml-cpu/ggml-cpu-hbm.h
rename to ggml/src/ggml-cpu/hbm.h
diff --git a/ggml/src/ggml-cpu/kleidiai/kleidiai.cpp b/ggml/src/ggml-cpu/kleidiai/kleidiai.cpp
index 15f0cd154..fafe45e6c 100644
--- a/ggml/src/ggml-cpu/kleidiai/kleidiai.cpp
+++ b/ggml/src/ggml-cpu/kleidiai/kleidiai.cpp
@@ -26,7 +26,7 @@
 #include "ggml-impl.h"
 #include "ggml-backend-impl.h"
 #include "ggml-threading.h"
-#include "ggml-cpu-traits.h"
+#include "traits.h"
 
 #include "kernels.h"
 
diff --git a/ggml/src/ggml-cpu/llamafile/sgemm.cpp b/ggml/src/ggml-cpu/llamafile/sgemm.cpp
index 1d46158f9..ed61869a5 100644
--- a/ggml/src/ggml-cpu/llamafile/sgemm.cpp
+++ b/ggml/src/ggml-cpu/llamafile/sgemm.cpp
@@ -52,8 +52,8 @@
 #include "ggml-impl.h"
 #include "ggml-cpu-impl.h"
 #include "ggml-quants.h"
+#include "simd-mappings.h"
 
-#include <atomic>
 #include <array>
 #include <type_traits>
 
@@ -63,7 +63,7 @@
 #define NOINLINE __attribute__((__noinline__))
 #endif
 
-#if defined(__ARM_NEON) || defined(__AVX512F__)
+#if defined(__ARM_NEON) || defined(__AVX512F__) || defined(__VXE__) || defined(__VXE2__)
 #define VECTOR_REGISTERS 32
 #else
 #define VECTOR_REGISTERS 16
@@ -74,7 +74,7 @@
 namespace {
 
 inline float unhalf(ggml_fp16_t d) {
-    return GGML_FP16_TO_FP32(d);
+    return GGML_CPU_FP16_TO_FP32(d);
 }
 
 ////////////////////////////////////////////////////////////////////////////////////////////////////
@@ -110,6 +110,12 @@ inline float16x8_t sub(float16x8_t x, float16x8_t y) { return vsubq_f16(x, y); }
 inline float16x8_t mul(float16x8_t x, float16x8_t y) { return vmulq_f16(x, y); }
 #endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
 
+#if defined(__VXE__) || defined(__VXE2__)
+inline float32x4_t add(float32x4_t x, float32x4_t y) { return vec_add(x, y); }
+inline float32x4_t sub(float32x4_t x, float32x4_t y) { return vec_sub(x, y); }
+inline float32x4_t mul(float32x4_t x, float32x4_t y) { return vec_mul(x, y); }
+#endif
+
 #if defined(__MMA__)
 typedef vector unsigned char vec_t;
 typedef __vector_quad acc_t;
@@ -163,6 +169,13 @@ inline float16x8_t madd(float16x8_t a, float16x8_t b, float16x8_t c) {
 #endif
 #endif
 
+#if defined(__VXE__) || defined(__VXE2__)
+template <>
+inline float32x4_t madd(float32x4_t a, float32x4_t b, float32x4_t c) {
+    return vec_madd(a, b, c);
+}
+#endif
+
 ////////////////////////////////////////////////////////////////////////////////////////////////////
 // VECTORIZED HORIZONTAL SUM
 
@@ -179,6 +192,13 @@ inline float hsum(float16x8_t x) {
 }
 #endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
 
+#if defined(__VXE__) || defined(__VXE2__)
+inline float hsum(float32x4_t x) {
+    float32x4_t tmp = x + vec_reve(x);
+    return tmp[0] + tmp[1];
+}
+#endif
+
 #if defined(__SSE__) || defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__)
 inline float hsum(__m128 x) {
 #if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__)
@@ -228,6 +248,21 @@ template <> inline float32x4_t load(const ggml_fp16_t *p) {
 #endif // _MSC_VER
 #endif // __ARM_NEON
 
+#if defined(__VXE__) || defined(__VXE2__)
+template <> inline float32x4_t load(const ggml_fp16_t * p) {
+    float tmp[4];
+
+    for (int i = 0; i < 4; i++) {
+        tmp[i] = GGML_CPU_FP16_TO_FP32(p[i]);
+    }
+
+    return vec_xl(0, (const float *)(tmp));
+}
+template <> inline float32x4_t load(const float * p) {
+    return vec_xl(0, p);
+}
+#endif
+
 #if defined(__SSE__) || defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__)
 template <> inline __m128 load(const float *p) {
     return _mm_loadu_ps(p);
@@ -394,8 +429,6 @@ class tinyBLAS {
 
     template <int RM, int RN, int BM>
     NOINLINE void gemm(int64_t m, int64_t n, int64_t BN) {
-        static std::atomic<int64_t> current_chunk;
-
         GGML_ASSERT(m % (RM * BM) == 0);
         const int64_t ytiles = m / (RM * BM);
         const int64_t xtiles = (n + RN -1) / RN;
@@ -410,7 +443,7 @@ class tinyBLAS {
         if (params->ith == 0) {
             GGML_ASSERT( jj_BN * SIZE_BN + (NB_BN - jj_BN) * (SIZE_BN - 1) == xtiles);
             // Every thread starts at ith, so the first unprocessed chunk is nth.  This save a bit of coordination right at the start.
-            std::atomic_store_explicit(&current_chunk, (int64_t)params->nth, std::memory_order_relaxed);
+            ggml_threadpool_chunk_set(params->threadpool, params->nth);
         }
 
         ggml_barrier(params->threadpool);
@@ -439,8 +472,7 @@ class tinyBLAS {
                 GGML_ASSERT(jj == jj2);
             }
 
-            // next step.
-            job = std::atomic_fetch_add_explicit(&current_chunk, (int64_t)1, std::memory_order_relaxed);
+            job = ggml_threadpool_chunk_add(params->threadpool, 1);
         }
 
         ggml_barrier(params->threadpool);
@@ -3323,6 +3355,14 @@ bool llamafile_sgemm(const struct ggml_compute_params * params, int64_t m, int64
             (const float *)B, ldb,
             (float *)C, ldc};
         return tb.matmul(m, n);
+#elif defined(__VXE__) || defined(__VXE2__)
+        if (n < 4)
+            return false;
+        tinyBLAS<4, float32x4_t, float32x4_t, float, float, float> tb{ params,
+            k, (const float *)A, lda,
+            (const float *)B, ldb,
+            (float *)C, ldc};
+        return tb.matmul(m, n);
 #elif defined(__MMA__)
         if (k % 8)
             return false;
@@ -3414,6 +3454,16 @@ bool llamafile_sgemm(const struct ggml_compute_params * params, int64_t m, int64
                 (float *)C, ldc};
             return tb.matmul(m, n);
         }
+#elif defined(__VXE__) || defined(__VXE2__)
+        if (n < 4)
+            return false;
+        if (Btype == GGML_TYPE_F16) {
+            tinyBLAS<4, float32x4_t, float32x4_t, ggml_fp16_t, ggml_fp16_t, float> tb{ params,
+                k, (const ggml_fp16_t *)A, lda,
+                (const ggml_fp16_t *)B, ldb,
+                (float *)C, ldc};
+            return tb.matmul(m, n);
+        }
 #endif
         return false;
     }
diff --git a/ggml/src/ggml-cpu/llamafile/sgemm.h b/ggml/src/ggml-cpu/llamafile/sgemm.h
index 3d2909515..729e8853d 100644
--- a/ggml/src/ggml-cpu/llamafile/sgemm.h
+++ b/ggml/src/ggml-cpu/llamafile/sgemm.h
@@ -1,6 +1,11 @@
 #pragma once
 #include <stdint.h>
 #include <stdbool.h>
+
+#if defined(__VXE__) || defined(__VXE2__)
+#include <vecintrin.h>
+#endif
+
 #ifdef __cplusplus
 extern "C" {
 #endif
diff --git a/ggml/src/ggml-cpu/ops.cpp b/ggml/src/ggml-cpu/ops.cpp
index d8de7531b..aaeee614a 100644
--- a/ggml/src/ggml-cpu/ops.cpp
+++ b/ggml/src/ggml-cpu/ops.cpp
@@ -3,6 +3,7 @@
 #include "ggml-cpu.h"
 #include "ggml-impl.h"
 #include "binary-ops.h"
+#include "ggml.h"
 #include "unary-ops.h"
 #include "vec.h"
 
@@ -108,7 +109,7 @@ static void ggml_compute_forward_dup_f16(
                         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]);
+                                dst_ptr[id] = GGML_CPU_FP16_TO_FP32(src0_ptr[i00]);
                                 id++;
                             }
                         }
@@ -130,7 +131,7 @@ static void ggml_compute_forward_dup_f16(
                             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]);
+                                src0_f32[i00] = GGML_CPU_FP16_TO_FP32(src0_ptr[i00]);
                             }
 
                             quantize_row_q(src0_f32, dst_ptr + id, ne00);
@@ -156,7 +157,7 @@ static void ggml_compute_forward_dup_f16(
                             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);
+                                dst_ptr[id] = GGML_CPU_FP16_TO_FP32(*src0_ptr);
                                 id++;
                             }
                         }
@@ -267,7 +268,7 @@ static void ggml_compute_forward_dup_f16(
                         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);
+                        *(float *) dst_ptr = GGML_CPU_FP16_TO_FP32(*(const ggml_fp16_t *) src0_ptr);
 
                         if (++i10 == ne0) {
                             i10 = 0;
@@ -372,7 +373,7 @@ static void ggml_compute_forward_dup_bf16(
                         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]));
+                                dst_ptr[id] = GGML_CPU_FP32_TO_FP16(GGML_BF16_TO_FP32(src0_ptr[i00]));
                                 id++;
                             }
                         }
@@ -473,7 +474,7 @@ static void ggml_compute_forward_dup_bf16(
                             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));
+                                dst_ptr[id] = GGML_CPU_FP32_TO_FP16(GGML_BF16_TO_FP32(*src0_ptr));
                                 id++;
                             }
                         }
@@ -566,7 +567,7 @@ static void ggml_compute_forward_dup_bf16(
                         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));
+                        *(ggml_fp16_t *) dst_ptr = GGML_CPU_FP32_TO_FP16(GGML_BF16_TO_FP32(*(const ggml_bf16_t *) src0_ptr));
 
                         if (++i10 == ne0) {
                             i10 = 0;
@@ -696,24 +697,8 @@ static void ggml_compute_forward_dup_f32(
     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;
+            if (ggml_get_type_traits_cpu(dst->type)->from_float) {
+                ggml_from_float_t const from_float = ggml_get_type_traits_cpu(dst->type)->from_float;
 
                 size_t id = 0;
                 size_t rs = nb0 * (ne00 / ggml_blck_size(dst->type));
@@ -724,7 +709,7 @@ static void ggml_compute_forward_dup_f32(
                         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);
+                            from_float(src0_ptr, dst_ptr + id, ne00);
                             id += rs;
                         }
                         id += rs * (ne01 - ir1);
@@ -765,7 +750,7 @@ static void ggml_compute_forward_dup_f32(
                             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);
+                                dst_ptr[id] = GGML_CPU_FP32_TO_FP16(*src0_ptr);
                                 id++;
                             }
                         }
@@ -878,7 +863,7 @@ static void ggml_compute_forward_dup_f32(
                         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);
+                        *(ggml_fp16_t *) dst_ptr = GGML_CPU_FP32_TO_FP16(*(const float *) src0_ptr);
 
                         if (++i10 == ne0) {
                             i10 = 0;
@@ -1419,7 +1404,7 @@ static void ggml_compute_forward_add1_f16_f32(
         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);
+            dst_ptr[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(src0_ptr[i]) + v);
         }
     }
 }
@@ -1435,7 +1420,7 @@ static void ggml_compute_forward_add1_f16_f16(
     GGML_ASSERT(ggml_is_scalar(src1));
 
     // scalar to add
-    const float v = GGML_FP16_TO_FP32(*(ggml_fp16_t *) src1->data);
+    const float v = GGML_CPU_FP16_TO_FP32(*(ggml_fp16_t *) src1->data);
 
     const int ith = params->ith;
     const int nth = params->nth;
@@ -1467,7 +1452,7 @@ static void ggml_compute_forward_add1_f16_f16(
         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);
+            dst_ptr[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(src0_ptr[i]) + v);
         }
     }
 }
@@ -1889,7 +1874,7 @@ static void ggml_compute_forward_sum_f16(
             }
         }
     }
-    ((ggml_fp16_t *) dst->data)[0] = GGML_FP32_TO_FP16(sum);
+    ((ggml_fp16_t *) dst->data)[0] = GGML_CPU_FP32_TO_FP16(sum);
 }
 
 static void ggml_compute_forward_sum_bf16(
@@ -2300,6 +2285,12 @@ void ggml_compute_forward_repeat(
             {
                 ggml_compute_forward_repeat_f32(params, dst);
             } break;
+        // TODO: templateify the implemenation and support for I64
+        //       ref https://github.com/ggml-org/llama.cpp/pull/14274#discussion_r2169492225
+        //case GGML_TYPE_I64:
+        //    {
+        //        ggml_compute_forward_repeat_i64(params, dst);
+        //    } break;
         default:
             {
                 GGML_ABORT("fatal error");
@@ -2660,7 +2651,7 @@ static void ggml_compute_forward_gelu_f16(
 #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);
+            const float v = GGML_CPU_FP16_TO_FP32(x);
             GGML_UNUSED(v);
             assert(!isnan(v));
             assert(!isinf(v));
@@ -2763,7 +2754,7 @@ static void ggml_compute_forward_gelu_erf_f16(
 #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);
+            const float v = GGML_CPU_FP16_TO_FP32(x);
             GGML_UNUSED(v);
             assert(!isnan(v));
             assert(!isinf(v));
@@ -2866,7 +2857,7 @@ static void ggml_compute_forward_gelu_quick_f16(
 #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);
+            const float v = GGML_CPU_FP16_TO_FP32(x);
             GGML_UNUSED(v);
             assert(!isnan(v));
             assert(!isinf(v));
@@ -2969,7 +2960,7 @@ static void ggml_compute_forward_silu_f16(
 #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);
+            const float v = GGML_CPU_FP16_TO_FP32(x);
             GGML_UNUSED(v);
             assert(!isnan(v));
             assert(!isinf(v));
@@ -3163,7 +3154,7 @@ static void ggml_compute_forward_silu_back_f16(
     #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);
+            const float v = GGML_CPU_FP16_TO_FP32(x);
             GGML_UNUSED(v);
             assert(!isnan(v));
             assert(!isinf(v));
@@ -3194,6 +3185,721 @@ void ggml_compute_forward_silu_back(
     }
 }
 
+// ggml_compute_forward_reglu
+
+static void ggml_compute_forward_reglu_f32(
+        const ggml_compute_params * params,
+        ggml_tensor * dst) {
+
+    const ggml_tensor * src0 = dst->src[0];
+    const ggml_tensor * src1 = dst->src[1];
+    char * src0_d = (char *) src0->data;
+    char * src1_d = (char *) (src1 ? src1->data : src0->data);
+    const size_t src0_o = src0->nb[1];
+    const size_t src1_o = src1 ? src1->nb[1] : src0->nb[1];
+
+    GGML_ASSERT(ggml_is_contiguous_1(src0));
+    GGML_ASSERT(ggml_is_contiguous_1(dst));
+
+    if (src1) {
+        GGML_ASSERT(ggml_is_contiguous_1(src1));
+        GGML_ASSERT(src0->type == src1->type);
+    }
+
+    const int ith = params->ith;
+    const int nth = params->nth;
+
+    const int nc = src1 ? src0->ne[0] : src0->ne[0] / 2;
+    const int nr = ggml_nrows(src0);
+
+    GGML_ASSERT(dst->ne[0] == nc);
+    GGML_ASSERT(ggml_nrows(dst) == nr);
+
+    const int32_t swapped = ggml_get_op_params_i32(dst, 1);
+
+    // 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 * src0_p = (float *) (src0_d + i1*src0_o);
+        float * src1_p = (float *) (src1_d + i1*src1_o);
+
+        if (!src1) {
+            src0_p += swapped ? nc : 0;
+            src1_p += swapped ? 0 : nc;
+        }
+
+        ggml_vec_reglu_f32(nc, (float *) ((char *) dst->data + i1*(dst->nb[1])), src0_p, src1_p);
+
+#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_reglu_f16(
+    const ggml_compute_params * params,
+    ggml_tensor * dst) {
+
+    const ggml_tensor * src0 = dst->src[0];
+    const ggml_tensor * src1 = dst->src[1];
+    char * src0_d = (char *) src0->data;
+    char * src1_d = (char *) (src1 ? src1->data : src0->data);
+    const size_t src0_o = src0->nb[1];
+    const size_t src1_o = src1 ? src1->nb[1] : src0->nb[1];
+
+    GGML_ASSERT(ggml_is_contiguous_1(src0));
+    GGML_ASSERT(ggml_is_contiguous_1(dst));
+
+    if (src1) {
+        GGML_ASSERT(ggml_is_contiguous_1(src1));
+        GGML_ASSERT(src0->type == src1->type);
+    }
+
+    const int ith = params->ith;
+    const int nth = params->nth;
+
+    const int nc = src1 ? src0->ne[0] : src0->ne[0] / 2;
+    const int nr = ggml_nrows(src0);
+
+    GGML_ASSERT(dst->ne[0] == nc);
+    GGML_ASSERT(ggml_nrows(dst) == nr);
+
+    const int32_t swapped = ggml_get_op_params_i32(dst, 1);
+
+    // 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_fp16_t * src0_p = (ggml_fp16_t *) (src0_d + i1*src0_o);
+        ggml_fp16_t * src1_p = (ggml_fp16_t *) (src1_d + i1*src1_o);
+
+        if (!src1) {
+            src0_p += swapped ? nc : 0;
+            src1_p += swapped ? 0 : nc;
+        }
+
+        ggml_vec_reglu_f16(nc, (ggml_fp16_t *) ((char *) dst->data + i1*(dst->nb[1])), src0_p, src1_p);
+
+#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_reglu(
+        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_reglu_f32(params, dst);
+            } break;
+        case GGML_TYPE_F16:
+            {
+                ggml_compute_forward_reglu_f16(params, dst);
+            } break;
+        default:
+            {
+                GGML_ABORT("fatal error");
+            }
+    }
+}
+
+// ggml_compute_forward_geglu
+
+static void ggml_compute_forward_geglu_f32(
+        const ggml_compute_params * params,
+        ggml_tensor * dst) {
+
+    const ggml_tensor * src0 = dst->src[0];
+    const ggml_tensor * src1 = dst->src[1];
+    char * src0_d = (char *) src0->data;
+    char * src1_d = (char *) (src1 ? src1->data : src0->data);
+    const size_t src0_o = src0->nb[1];
+    const size_t src1_o = src1 ? src1->nb[1] : src0->nb[1];
+
+    GGML_ASSERT(ggml_is_contiguous_1(src0));
+    GGML_ASSERT(ggml_is_contiguous_1(dst));
+
+    if (src1) {
+        GGML_ASSERT(ggml_is_contiguous_1(src1));
+        GGML_ASSERT(src0->type == src1->type);
+    }
+
+    const int ith = params->ith;
+    const int nth = params->nth;
+
+    const int nc = src1 ? src0->ne[0] : src0->ne[0] / 2;
+    const int nr = ggml_nrows(src0);
+
+    GGML_ASSERT(dst->ne[0] == nc);
+    GGML_ASSERT(ggml_nrows(dst) == nr);
+
+    const int32_t swapped = ggml_get_op_params_i32(dst, 1);
+
+    // 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 * src0_p = (float *) (src0_d + i1*src0_o);
+        float * src1_p = (float *) (src1_d + i1*src1_o);
+
+        if (!src1) {
+            src0_p += swapped ? nc : 0;
+            src1_p += swapped ? 0 : nc;
+        }
+
+        ggml_vec_geglu_f32(nc, (float *) ((char *) dst->data + i1*(dst->nb[1])), src0_p, src1_p);
+
+#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_geglu_f16(
+    const ggml_compute_params * params,
+    ggml_tensor * dst) {
+
+    const ggml_tensor * src0 = dst->src[0];
+    const ggml_tensor * src1 = dst->src[1];
+    char * src0_d = (char *) src0->data;
+    char * src1_d = (char *) (src1 ? src1->data : src0->data);
+    const size_t src0_o = src0->nb[1];
+    const size_t src1_o = src1 ? src1->nb[1] : src0->nb[1];
+
+    GGML_ASSERT(ggml_is_contiguous_1(src0));
+    GGML_ASSERT(ggml_is_contiguous_1(dst));
+
+    if (src1) {
+        GGML_ASSERT(ggml_is_contiguous_1(src1));
+        GGML_ASSERT(src0->type == src1->type);
+    }
+
+    const int ith = params->ith;
+    const int nth = params->nth;
+
+    const int nc = src1 ? src0->ne[0] : src0->ne[0] / 2;
+    const int nr = ggml_nrows(src0);
+
+    GGML_ASSERT(dst->ne[0] == nc);
+    GGML_ASSERT(ggml_nrows(dst) == nr);
+
+    const int32_t swapped = ggml_get_op_params_i32(dst, 1);
+
+    // 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_fp16_t * src0_p = (ggml_fp16_t *) (src0_d + i1*src0_o);
+        ggml_fp16_t * src1_p = (ggml_fp16_t *) (src1_d + i1*src1_o);
+
+        if (!src1) {
+            src0_p += swapped ? nc : 0;
+            src1_p += swapped ? 0 : nc;
+        }
+
+        ggml_vec_geglu_f16(nc, (ggml_fp16_t *) ((char *) dst->data + i1*(dst->nb[1])), src0_p, src1_p);
+
+#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_geglu(
+        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_geglu_f32(params, dst);
+            } break;
+        case GGML_TYPE_F16:
+            {
+                ggml_compute_forward_geglu_f16(params, dst);
+            } break;
+        default:
+            {
+                GGML_ABORT("fatal error");
+            }
+    }
+}
+
+// ggml_compute_forward_swiglu
+
+static void ggml_compute_forward_swiglu_f32(
+        const ggml_compute_params * params,
+        ggml_tensor * dst) {
+
+    const ggml_tensor * src0 = dst->src[0];
+    const ggml_tensor * src1 = dst->src[1];
+    char * src0_d = (char *) src0->data;
+    char * src1_d = (char *) (src1 ? src1->data : src0->data);
+    const size_t src0_o = src0->nb[1];
+    const size_t src1_o = src1 ? src1->nb[1] : src0->nb[1];
+
+    GGML_ASSERT(ggml_is_contiguous_1(src0));
+    GGML_ASSERT(ggml_is_contiguous_1(dst));
+
+    if (src1) {
+        GGML_ASSERT(ggml_is_contiguous_1(src1));
+        GGML_ASSERT(src0->type == src1->type);
+    }
+
+    const int ith = params->ith;
+    const int nth = params->nth;
+
+    const int nc = src1 ? src0->ne[0] : src0->ne[0] / 2;
+    const int nr = ggml_nrows(src0);
+
+    GGML_ASSERT(dst->ne[0] == nc);
+    GGML_ASSERT(ggml_nrows(dst) == nr);
+
+    const int32_t swapped = ggml_get_op_params_i32(dst, 1);
+
+    // 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 * src0_p = (float *) (src0_d + i1*src0_o);
+        float * src1_p = (float *) (src1_d + i1*src1_o);
+
+        if (!src1) {
+            src0_p += swapped ? nc : 0;
+            src1_p += swapped ? 0 : nc;
+        }
+
+        ggml_vec_swiglu_f32(nc, (float *) ((char *) dst->data + i1*(dst->nb[1])), src0_p, src1_p);
+
+#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_swiglu_f16(
+    const ggml_compute_params * params,
+    ggml_tensor * dst) {
+
+    const ggml_tensor * src0 = dst->src[0];
+    const ggml_tensor * src1 = dst->src[1];
+    char * src0_d = (char *) src0->data;
+    char * src1_d = (char *) (src1 ? src1->data : src0->data);
+    const size_t src0_o = src0->nb[1];
+    const size_t src1_o = src1 ? src1->nb[1] : src0->nb[1];
+
+    GGML_ASSERT(ggml_is_contiguous_1(src0));
+    GGML_ASSERT(ggml_is_contiguous_1(dst));
+
+    if (src1) {
+        GGML_ASSERT(ggml_is_contiguous_1(src1));
+        GGML_ASSERT(src0->type == src1->type);
+    }
+
+    const int ith = params->ith;
+    const int nth = params->nth;
+
+    const int nc = src1 ? src0->ne[0] : src0->ne[0] / 2;
+    const int nr = ggml_nrows(src0);
+
+    GGML_ASSERT(dst->ne[0] == nc);
+    GGML_ASSERT(ggml_nrows(dst) == nr);
+
+    const int32_t swapped = ggml_get_op_params_i32(dst, 1);
+
+    // 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_fp16_t * src0_p = (ggml_fp16_t *) (src0_d + i1*src0_o);
+        ggml_fp16_t * src1_p = (ggml_fp16_t *) (src1_d + i1*src1_o);
+
+        if (!src1) {
+            src0_p += swapped ? nc : 0;
+            src1_p += swapped ? 0 : nc;
+        }
+
+        ggml_vec_swiglu_f16(nc, (ggml_fp16_t *) ((char *) dst->data + i1*(dst->nb[1])), src0_p, src1_p);
+
+#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_swiglu(
+        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_swiglu_f32(params, dst);
+            } break;
+        case GGML_TYPE_F16:
+            {
+                ggml_compute_forward_swiglu_f16(params, dst);
+            } break;
+        default:
+            {
+                GGML_ABORT("fatal error");
+            }
+    }
+}
+
+// ggml_compute_forward_geglu_erf
+
+static void ggml_compute_forward_geglu_erf_f32(
+        const ggml_compute_params * params,
+        ggml_tensor * dst) {
+
+    const ggml_tensor * src0 = dst->src[0];
+    const ggml_tensor * src1 = dst->src[1];
+    char * src0_d = (char *) src0->data;
+    char * src1_d = (char *) (src1 ? src1->data : src0->data);
+    const size_t src0_o = src0->nb[1];
+    const size_t src1_o = src1 ? src1->nb[1] : src0->nb[1];
+
+    GGML_ASSERT(ggml_is_contiguous_1(src0));
+    GGML_ASSERT(ggml_is_contiguous_1(dst));
+
+    if (src1) {
+        GGML_ASSERT(ggml_is_contiguous_1(src1));
+        GGML_ASSERT(src0->type == src1->type);
+    }
+
+    const int ith = params->ith;
+    const int nth = params->nth;
+
+    const int nc = src1 ? src0->ne[0] : src0->ne[0] / 2;
+    const int nr = ggml_nrows(src0);
+
+    GGML_ASSERT(dst->ne[0] == nc);
+    GGML_ASSERT(ggml_nrows(dst) == nr);
+
+    const int32_t swapped = ggml_get_op_params_i32(dst, 1);
+
+    // 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 * src0_p = (float *) (src0_d + i1*src0_o);
+        float * src1_p = (float *) (src1_d + i1*src1_o);
+
+        if (!src1) {
+            src0_p += swapped ? nc : 0;
+            src1_p += swapped ? 0 : nc;
+        }
+
+        ggml_vec_geglu_erf_f32(nc, (float *) ((char *) dst->data + i1*(dst->nb[1])), src0_p, src1_p);
+
+#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_geglu_erf_f16(
+    const ggml_compute_params * params,
+    ggml_tensor * dst) {
+
+    const ggml_tensor * src0 = dst->src[0];
+    const ggml_tensor * src1 = dst->src[1];
+    char * src0_d = (char *) src0->data;
+    char * src1_d = (char *) (src1 ? src1->data : src0->data);
+    const size_t src0_o = src0->nb[1];
+    const size_t src1_o = src1 ? src1->nb[1] : src0->nb[1];
+
+    GGML_ASSERT(ggml_is_contiguous_1(src0));
+    GGML_ASSERT(ggml_is_contiguous_1(dst));
+
+    if (src1) {
+        GGML_ASSERT(ggml_is_contiguous_1(src1));
+        GGML_ASSERT(src0->type == src1->type);
+    }
+
+    const int ith = params->ith;
+    const int nth = params->nth;
+
+    const int nc = src1 ? src0->ne[0] : src0->ne[0] / 2;
+    const int nr = ggml_nrows(src0);
+
+    GGML_ASSERT(dst->ne[0] == nc);
+    GGML_ASSERT(ggml_nrows(dst) == nr);
+
+    const int32_t swapped = ggml_get_op_params_i32(dst, 1);
+
+    // 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_fp16_t * src0_p = (ggml_fp16_t *) (src0_d + i1*src0_o);
+        ggml_fp16_t * src1_p = (ggml_fp16_t *) (src1_d + i1*src1_o);
+
+        if (!src1) {
+            src0_p += swapped ? nc : 0;
+            src1_p += swapped ? 0 : nc;
+        }
+
+        ggml_vec_geglu_erf_f16(nc, (ggml_fp16_t *) ((char *) dst->data + i1*(dst->nb[1])), src0_p, src1_p);
+
+#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_geglu_erf(
+        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_geglu_erf_f32(params, dst);
+            } break;
+        case GGML_TYPE_F16:
+            {
+                ggml_compute_forward_geglu_erf_f16(params, dst);
+            } break;
+        default:
+            {
+                GGML_ABORT("fatal error");
+            }
+    }
+}
+
+// ggml_compute_forward_geglu_quick
+
+static void ggml_compute_forward_geglu_quick_f32(
+        const ggml_compute_params * params,
+        ggml_tensor * dst) {
+
+    const ggml_tensor * src0 = dst->src[0];
+    const ggml_tensor * src1 = dst->src[1];
+    char * src0_d = (char *) src0->data;
+    char * src1_d = (char *) (src1 ? src1->data : src0->data);
+    const size_t src0_o = src0->nb[1];
+    const size_t src1_o = src1 ? src1->nb[1] : src0->nb[1];
+
+    GGML_ASSERT(ggml_is_contiguous_1(src0));
+    GGML_ASSERT(ggml_is_contiguous_1(dst));
+
+    if (src1) {
+        GGML_ASSERT(ggml_is_contiguous_1(src1));
+        GGML_ASSERT(src0->type == src1->type);
+    }
+
+    const int ith = params->ith;
+    const int nth = params->nth;
+
+    const int nc = src1 ? src0->ne[0] : src0->ne[0] / 2;
+    const int nr = ggml_nrows(src0);
+
+    GGML_ASSERT(dst->ne[0] == nc);
+    GGML_ASSERT(ggml_nrows(dst) == nr);
+
+    const int32_t swapped = ggml_get_op_params_i32(dst, 1);
+
+    // 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 * src0_p = (float *) (src0_d + i1*src0_o);
+        float * src1_p = (float *) (src1_d + i1*src1_o);
+
+        if (!src1) {
+            src0_p += swapped ? nc : 0;
+            src1_p += swapped ? 0 : nc;
+        }
+
+        ggml_vec_geglu_quick_f32(nc, (float *) ((char *) dst->data + i1*(dst->nb[1])), src0_p, src1_p);
+
+#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_geglu_quick_f16(
+    const ggml_compute_params * params,
+    ggml_tensor * dst) {
+
+    const ggml_tensor * src0 = dst->src[0];
+    const ggml_tensor * src1 = dst->src[1];
+    char * src0_d = (char *) src0->data;
+    char * src1_d = (char *) (src1 ? src1->data : src0->data);
+    const size_t src0_o = src0->nb[1];
+    const size_t src1_o = src1 ? src1->nb[1] : src0->nb[1];
+
+    GGML_ASSERT(ggml_is_contiguous_1(src0));
+    GGML_ASSERT(ggml_is_contiguous_1(dst));
+
+    if (src1) {
+        GGML_ASSERT(ggml_is_contiguous_1(src1));
+        GGML_ASSERT(src0->type == src1->type);
+    }
+
+    const int ith = params->ith;
+    const int nth = params->nth;
+
+    const int nc = src1 ? src0->ne[0] : src0->ne[0] / 2;
+    const int nr = ggml_nrows(src0);
+
+    GGML_ASSERT(dst->ne[0] == nc);
+    GGML_ASSERT(ggml_nrows(dst) == nr);
+
+    const int32_t swapped = ggml_get_op_params_i32(dst, 1);
+
+    // 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_fp16_t * src0_p = (ggml_fp16_t *) (src0_d + i1*src0_o);
+        ggml_fp16_t * src1_p = (ggml_fp16_t *) (src1_d + i1*src1_o);
+
+        if (!src1) {
+            src0_p += swapped ? nc : 0;
+            src1_p += swapped ? 0 : nc;
+        }
+
+        ggml_vec_geglu_quick_f16(nc, (ggml_fp16_t *) ((char *) dst->data + i1*(dst->nb[1])), src0_p, src1_p);
+
+#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_geglu_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_geglu_quick_f32(params, dst);
+            } break;
+        case GGML_TYPE_F16:
+            {
+                ggml_compute_forward_geglu_quick_f16(params, dst);
+            } break;
+        default:
+            {
+                GGML_ABORT("fatal error");
+            }
+    }
+}
+
 // ggml_compute_forward_norm
 
 static void ggml_compute_forward_norm_f32(
@@ -4470,6 +5176,74 @@ void ggml_compute_forward_get_rows(
     //}
 }
 
+static void ggml_compute_forward_set_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 = ne01;
+
+    assert(ne0  == nc);
+    assert(ne2  == ne02);
+    assert(ne3  == ne03);
+    assert(src0->type == GGML_TYPE_F32);
+    assert(ne02 % ne11 == 0);
+    assert(ne03 % ne12 == 0);
+
+    const int ith = params->ith;
+    const int nth = params->nth;
+
+    // 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 = std::min(ir0 + dr, nr);
+
+    ggml_from_float_t const from_float = ggml_get_type_traits_cpu(dst->type)->from_float;
+
+    for (int64_t i03 = 0; i03 < ne03; ++i03) {
+        for (int64_t i02 = 0; i02 < ne02; ++i02) {
+            for (int64_t i = ir0; i < ir1; ++i) {
+                const int64_t i12 = i03%ne12;
+                const int64_t i11 = i02%ne11;
+                const int64_t i10 = i;
+
+                const int64_t i1 = *(int64_t *) ((char *) src1->data + i10*nb10 + i11*nb11 + i12*nb12);
+
+                GGML_ASSERT(i1 >= 0 && i1 < ne1);
+
+                from_float(
+                        (const float *) ((char *) src0->data +  i*nb01 + i02*nb02 + i03*nb03),
+                                        ((char *)  dst->data + i1*nb1  + i02*nb2  + i03*nb3), nc);
+            }
+        }
+    }
+}
+
+void ggml_compute_forward_set_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_set_rows_f32(params, dst);
+            } break;
+        default:
+            {
+                GGML_ABORT("src0->type = %d (%s) not supported", src0->type, ggml_type_name(src0->type));
+            }
+    }
+}
+
 // ggml_compute_forward_get_rows_back
 
 static void ggml_compute_forward_get_rows_back_f32_f16(
@@ -4500,7 +5274,7 @@ static void ggml_compute_forward_get_rows_back_f32_f16(
 
         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);
+            ((float *) ((char *) dst->data + r*dst->nb[1]))[j] += GGML_CPU_FP16_TO_FP32(v);
         }
     }
 }
@@ -4744,14 +5518,17 @@ static void ggml_compute_forward_soft_max_f32(
     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;
+    const int64_t nb11 = src1 ? src1->nb[1] : 1;
+    const int64_t nb12 = src1 ? src1->nb[2] : 1;
+    const int64_t nb13 = src1 ? src1->nb[3] : 1;
+
+    const int64_t ne12 = src1 ? src1->ne[2] : 1;
+    const int64_t ne13 = src1 ? src1->ne[3] : 1;
 
     // TODO: is this supposed to be ceil instead of floor?
     //       https://huggingface.co/mosaicml/mpt-7b/blob/main/attention.py#L370
@@ -4761,68 +5538,66 @@ static void ggml_compute_forward_soft_max_f32(
     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;
+    float * wp = (float *) params->wdata + (ne00 + 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;
+    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 int64_t i11 = i01;
+                const int64_t i12 = i02%ne12;
+                const int64_t i13 = i03%ne13;
 
-        float * sp = (float *)((char *) src0->data + i1*src0->nb[1]);
-        float * dp = (float *)((char *)  dst->data +  i1*dst->nb[1]);
+                // ALiBi
+                const uint32_t h = i02; // 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;
 
-        // 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;
+                float * sp = (float *)((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03);
+                float * dp = (float *)((char *)  dst->data + i01*nb1  + i02*nb2  + i03*nb3);
 
-        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]);
+                // broadcast the mask across rows
+                ggml_fp16_t * mp_f16 = src1 ? (ggml_fp16_t *)((char *) src1->data + i11*nb11 + i12*nb12 + i13*nb13) : NULL;
+                float       * mp_f32 = src1 ? (float       *)((char *) src1->data + i11*nb11 + i12*nb12 + i13*nb13) : NULL;
+
+                ggml_vec_cpy_f32  (ne00, wp, sp);
+                ggml_vec_scale_f32(ne00, wp, scale);
+                if (mp_f32) {
+                    if (use_f16) {
+                        for (int i = 0; i < ne00; ++i) {
+                            wp[i] += slope*GGML_CPU_FP16_TO_FP32(mp_f16[i]);
+                        }
+                    } else {
+                        for (int i = 0; i < ne00; ++i) {
+                            wp[i] += slope*mp_f32[i];
+                        }
+                    }
                 }
-            } else {
-                for (int i = 0; i < nc; ++i) {
-                    wp[i] += slope*mp_f32[i];
+
+#ifndef NDEBUG
+                for (int i = 0; i < ne00; ++i) {
+                    //printf("p[%d] = %f\n", i, p[i]);
+                    assert(!isnan(wp[i]));
                 }
+#endif
+
+                float max = -INFINITY;
+                ggml_vec_max_f32(ne00, &max, wp);
+
+                ggml_float sum = ggml_vec_soft_max_f32(ne00, dp, wp, max);
+                assert(sum > 0.0);
+
+                sum = 1.0/sum;
+                ggml_vec_scale_f32(ne00, dp, sum);
+
+#ifndef NDEBUG
+                for (int i = 0; i < ne00; ++i) {
+                    assert(!isnan(dp[i]));
+                    assert(!isinf(dp[i]));
+                }
+#endif
             }
         }
-
-#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
     }
 }
 
@@ -5018,8 +5793,8 @@ static void ggml_compute_forward_clamp_f16(
         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));
+            float v = GGML_CPU_FP16_TO_FP32(src0_ptr[i]);
+            dst_ptr[i] = GGML_CPU_FP32_TO_FP16(MAX(MIN(v, max), min));
         }
     }
 }
@@ -5476,11 +6251,11 @@ static void ggml_compute_forward_rope_f16(
                             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]);
+                            const float x0 = GGML_CPU_FP16_TO_FP32(src[0]);
+                            const float x1 = GGML_CPU_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);
+                            dst_data[0]      = GGML_CPU_FP32_TO_FP16(x0*cos_theta - x1*sin_theta);
+                            dst_data[n_dims] = GGML_CPU_FP32_TO_FP16(x0*sin_theta + x1*cos_theta);
                         }
                     } else {
                         for (int64_t i0 = 0; i0 < n_dims; i0 += 2) {
@@ -5492,11 +6267,11 @@ static void ggml_compute_forward_rope_f16(
                             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]);
+                            const float x0 = GGML_CPU_FP16_TO_FP32(src[0]);
+                            const float x1 = GGML_CPU_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);
+                            dst_data[0]        = GGML_CPU_FP32_TO_FP16(x0*cos_theta - x1*sin_theta);
+                            dst_data[n_dims/2] = GGML_CPU_FP32_TO_FP16(x0*sin_theta + x1*cos_theta);
                         }
                     }
                 } else {
@@ -5507,11 +6282,11 @@ static void ggml_compute_forward_rope_f16(
                         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]);
+                        const float x0 = GGML_CPU_FP16_TO_FP32(src[0]);
+                        const float x1 = GGML_CPU_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);
+                        dst_data[0] = GGML_CPU_FP32_TO_FP16(x0*cos_theta - x1*sin_theta);
+                        dst_data[1] = GGML_CPU_FP32_TO_FP16(x0*sin_theta + x1*cos_theta);
                     }
                 }
 
@@ -5525,11 +6300,11 @@ static void ggml_compute_forward_rope_f16(
                         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]);
+                        const float x0 = GGML_CPU_FP16_TO_FP32(src[0]);
+                        const float x1 = GGML_CPU_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);
+                        dst_data[0]      = GGML_CPU_FP32_TO_FP16(x0*cos_theta - x1*sin_theta);
+                        dst_data[n_dims] = GGML_CPU_FP32_TO_FP16(x0*sin_theta + x1*cos_theta);
                     }
                 } else {
                     for (int64_t i0 = n_dims; i0 < ne0; i0 += 2) {
@@ -5640,7 +6415,7 @@ static void ggml_compute_forward_conv_transpose_1d_f16_f32(
             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]);
+                    dst_data[i10*ne11 + i11] = GGML_CPU_FP32_TO_FP16(src[i10]);
                 }
             }
         }
@@ -5933,7 +6708,7 @@ static void ggml_compute_forward_im2col_f16(
                                 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]);
+                                    dst_data[iic*(KH*KW) + ikh*KW + ikw] = GGML_CPU_FP32_TO_FP16(src_data[iih*IW + iiw]);
                                 }
                             }
                         }
@@ -6058,6 +6833,186 @@ void ggml_compute_forward_im2col_back_f32(
     }
 }
 
+static void ggml_call_mul_mat(ggml_type type, const ggml_compute_params * params, int64_t m, int64_t n, int64_t k,
+                              void * a, void * b, float * c) {
+    const ggml_type_traits * traits = ggml_get_type_traits(type);
+    struct ggml_tensor src1 = {};
+    src1.type  = type;
+    src1.ne[0] = k;
+    src1.ne[1] = m;
+    src1.ne[2] = 1;
+    src1.ne[3] = 1;
+    src1.nb[0] = traits->type_size;
+    src1.nb[1] = k * traits->type_size;
+    src1.nb[2] = src1.nb[1];
+    src1.nb[3] = src1.nb[2];
+    src1.data  = a;
+
+    struct ggml_tensor src0 = {};
+    src0.type  = type;
+    src0.ne[0] = k;
+    src0.ne[1] = n;
+    src0.ne[2] = 1;
+    src0.ne[3] = 1;
+    src0.nb[0] = traits->type_size;
+    src0.nb[1] = k * traits->type_size;
+    src0.nb[2] = src0.nb[1];
+    src0.nb[3] = src0.nb[2];
+    src0.data  = b;
+
+    struct ggml_tensor dst = {};
+    dst.ne[0] = n;
+    dst.ne[1] = m;
+    dst.ne[2] = 1;
+    dst.ne[3] = 1;
+    dst.nb[0] = sizeof(float);
+    dst.nb[1] = n * sizeof(float);
+    dst.nb[2] = dst.nb[1];
+    dst.nb[3] = dst.nb[2];
+    dst.data  = c;
+    dst.src[0] = &src0;
+    dst.src[1] = &src1;
+
+    ggml_compute_forward_mul_mat(params, &dst);
+}
+
+// ggml_compute_forward_conv_2d
+
+static void ggml_compute_forward_conv_2d_impl(const ggml_compute_params * params,
+                                              const ggml_tensor *         kernel,  // [KW, KH, IC, OC]
+                                              const ggml_tensor *         src,     // [W, H, C, N]
+                                              ggml_tensor *               dst,     // [OW, OH, OC, N]
+                                              ggml_type                   kernel_type) {
+
+    GGML_ASSERT(ggml_is_contiguous(kernel));
+    GGML_ASSERT(kernel_type == GGML_TYPE_F16 || kernel_type == GGML_TYPE_F32);
+    GGML_ASSERT(kernel->type == kernel_type);
+
+    const ggml_type_traits * traits = ggml_get_type_traits(kernel_type);
+
+    const int32_t stride_x   = dst->op_params[0];
+    const int32_t stride_y   = dst->op_params[1];
+    const int32_t pad_x      = dst->op_params[2];
+    const int32_t pad_y      = dst->op_params[3];
+    const int32_t dilation_x = dst->op_params[4];
+    const int32_t dilation_y = dst->op_params[5];
+
+    const int64_t c_in  = src->ne[2];
+    const int64_t c_out = kernel->ne[3];
+    GGML_ASSERT(c_in == kernel->ne[2]);
+
+    const int64_t src_w = src->ne[0];
+    const int64_t src_h = src->ne[1];
+    const int64_t knl_w = kernel->ne[0];
+    const int64_t knl_h = kernel->ne[1];
+    const int64_t dst_w = dst->ne[0];
+    const int64_t dst_h = dst->ne[1];
+
+    const float * src_data = (float *) src->data;
+    void  * knl_data       = kernel->data;
+    float * dst_data       = (float *) dst->data;
+
+    const int64_t knl_n           = knl_w * knl_h * c_in;
+    const int64_t patch_total     = dst->ne[3] * dst_w * dst_h;
+
+    const int64_t space_per_patch   = knl_n * traits->type_size + c_out * sizeof(float);
+    const int64_t batch_size        = params->wsize / space_per_patch;
+    const int64_t patches_per_batch = batch_size > 8 ? (batch_size / 8) * 8 : batch_size;
+    const int64_t batch_n           = (patch_total + patches_per_batch - 1) / patches_per_batch;
+
+    GGML_ASSERT(patches_per_batch > 0 && batch_size >= 1);
+
+    void * tmp = params->wdata;
+
+    for (int64_t batch_i = 0; batch_i < batch_n; ++batch_i) {
+
+        const int64_t patch_start_batch = batch_i * patches_per_batch;
+        const int64_t patch_end_batch   = std::min(patch_start_batch + patches_per_batch,
+                                              patch_total);
+        const int64_t patch_n           = patch_end_batch - patch_start_batch;
+
+        const int64_t patch_per_thread  = (patch_n + params->nth - 1) / params->nth;
+        const int64_t patch_start       = patch_start_batch + params->ith * patch_per_thread;
+        const int64_t patch_end         = std::min(patch_start + patch_per_thread, patch_end_batch);
+
+        //im2col for a patch
+        for (int64_t p = patch_start; p < patch_end; ++p) {
+            const int64_t  batch_n     =  p / (dst_w * dst_h);
+            const int64_t  src_x       = (p / dst_w) % dst_h;
+            const int64_t  src_y       =  p % dst_w;
+
+            const float * src_base = (const float *)((const char *)src_data + batch_n * src->nb[3]);
+            char *        dst_row  = (char *) tmp + (p % patches_per_batch) * knl_n * traits->type_size;
+
+            for (int64_t ic = 0; ic < c_in; ++ic) {
+                for (int64_t ky = 0; ky < knl_h; ++ky) {
+                    for (int64_t kx = 0; kx < knl_w; ++kx) {
+                        const int64_t sy = src_x * stride_y + ky * dilation_y - pad_y;
+                        const int64_t sx = src_y * stride_x + kx * dilation_x - pad_x;
+
+                        int64_t dst_idx = ic * (knl_h * knl_w) + ky * knl_w + kx;
+
+                        float src_val;
+                        if (sy < 0 || sy >= src_h || sx < 0 || sx >= src_w) {
+                            src_val = 0.0f;
+                        } else {
+                            const float * src_ptr = (const float *)((const char *)src_base + sx * src->nb[0] + sy * src->nb[1] + ic * src->nb[2]);
+                            src_val               = *src_ptr;
+                        }
+
+                        char * element_ptr = dst_row + dst_idx * traits->type_size;
+                        if (kernel_type == GGML_TYPE_F32) {
+                            *(float *) element_ptr = src_val;
+                        } else if (kernel_type == GGML_TYPE_F16) {
+                            *(ggml_fp16_t *) element_ptr = GGML_CPU_FP32_TO_FP16(src_val);
+                        }
+                    }
+                }
+            }
+        }   // patches handled by this thread
+
+        ggml_barrier(params->threadpool);
+
+        float * gemm_output = (float *) ((char *) tmp + patches_per_batch * knl_n * traits->type_size);
+
+        GGML_ASSERT(gemm_output + patch_n * c_out <= (float*)tmp + params->wsize);
+
+        // GEMM: patches[patch_n, knl_n] × kernel[knl_n, c_out] = output[patch_n, c_out]
+        ggml_call_mul_mat(kernel_type, params, patch_n, c_out, knl_n, tmp, knl_data, gemm_output);
+
+        ggml_barrier(params->threadpool);
+
+
+        //permute back [OC, N, OH, OW] to [N, OC, OH, OW]
+        const int64_t permute_per_thread = (patch_n + params->nth - 1) / params->nth;
+        const int64_t permute_start = params->ith * permute_per_thread;
+        const int64_t permute_end = std::min(permute_start + permute_per_thread, patch_n);
+
+        for (int64_t i = permute_start; i < permute_end; ++i) {
+            const int64_t p       = patch_start_batch + i;
+            const int64_t batch_n = p / (dst_w * dst_h);
+            const int64_t dst_y   = (p / dst_w) % dst_h;
+            const int64_t dst_x   = p % dst_w;
+
+            for (int64_t oc = 0; oc < c_out; ++oc) {
+                const float value = gemm_output[i * c_out + oc];
+                float * dst_ptr = (float *)((char *)dst_data + dst_x * dst->nb[0] + dst_y * dst->nb[1] + oc * dst->nb[2] + batch_n * dst->nb[3]);
+                *dst_ptr = value;
+            }
+        }
+    }
+}
+
+void ggml_compute_forward_conv_2d(
+        const ggml_compute_params * params,
+        ggml_tensor * dst) {
+
+    const ggml_tensor * src0 = dst->src[0];
+    const ggml_tensor * src1 = dst->src[1];
+
+    ggml_compute_forward_conv_2d_impl(params, src0, src1, dst, src0->type);
+}
+
 // ggml_compute_forward_conv_transpose_2d
 
 void ggml_compute_forward_conv_transpose_2d(
@@ -6109,7 +7064,7 @@ void ggml_compute_forward_conv_transpose_2d(
                     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]);
+                        dst_data[i10*ne12 + i12] = GGML_CPU_FP32_TO_FP16(src[i10]);
                     }
                 }
             }
@@ -6358,7 +7313,7 @@ static void ggml_compute_forward_pool_1d_sk_p0(
                 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]);
+                const float srow_j = (src->type == GGML_TYPE_F32) ? ((const float*)srow)[j] : GGML_CPU_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;
@@ -6450,7 +7405,7 @@ void ggml_compute_forward_pool_2d(
                     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]);
+                        const float srow_j = (src->type == GGML_TYPE_F32) ? ((const float*)srow)[j] : GGML_CPU_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;
@@ -6538,7 +7493,7 @@ void ggml_compute_forward_pool_2d_back(
                             }
 
                             const float val = dst->type == GGML_TYPE_F32 ?
-                                ((const float *) drowf)[j] : GGML_FP16_TO_FP32(((const ggml_fp16_t *) drowf)[j]);
+                                ((const float *) drowf)[j] : GGML_CPU_FP16_TO_FP32(((const ggml_fp16_t *) drowf)[j]);
                             if (val <= maxval) {
                                 continue;
                             }
@@ -6558,7 +7513,7 @@ void ggml_compute_forward_pool_2d_back(
                     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]));
+                        ((ggml_fp16_t *) drow)[j] = GGML_CPU_FP32_TO_FP16(grad0 + GGML_CPU_FP16_TO_FP32(((const ggml_fp16_t *) drow)[j]));
                     }
                 } else if (op == GGML_OP_POOL_AVG) {
                     const float grad = grad0 / ka;
@@ -6577,7 +7532,7 @@ void ggml_compute_forward_pool_2d_back(
                             if (dst->type == GGML_TYPE_F32) {
                                 ((float *) drow)[j] += grad;
                             } else {
-                                ((ggml_fp16_t *) drow)[j] += GGML_FP32_TO_FP16(grad);
+                                ((ggml_fp16_t *) drow)[j] += GGML_CPU_FP32_TO_FP16(grad);
                             }
                         }
                     }
@@ -6608,12 +7563,13 @@ static void ggml_compute_forward_upscale_f32(
 
     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];
+    float sf0 = (float)ne0/src0->ne[0];
+    float sf1 = (float)ne1/src0->ne[1];
+    float sf2 = (float)ne2/src0->ne[2];
+    float sf3 = (float)ne3/src0->ne[3];
 
-    const ggml_scale_mode mode = (ggml_scale_mode) ggml_get_op_params_i32(dst, 0);
+    const int32_t mode_flags = ggml_get_op_params_i32(dst, 0);
+    const ggml_scale_mode mode = (ggml_scale_mode) (mode_flags & 0xFF);
 
     if (mode == GGML_SCALE_MODE_NEAREST) {
         for (int64_t i3 = 0; i3 < ne3; i3++) {
@@ -6634,8 +7590,12 @@ static void ggml_compute_forward_upscale_f32(
             }
         }
     } 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;
+        float pixel_offset = 0.5f;
+        if (mode_flags & GGML_SCALE_FLAG_ALIGN_CORNERS) {
+            pixel_offset = 0.0f;
+            sf0 = (float)(ne0 - 1) / (src0->ne[0] - 1);
+            sf1 = (float)(ne1 - 1) / (src0->ne[1] - 1);
+        }
 
         for (int64_t i3 = 0; i3 < ne3; i3++) {
             const int64_t i03 = i3 / sf3;
@@ -6793,6 +7753,73 @@ void ggml_compute_forward_pad_reflect_1d(
     }
 }
 
+// ggml_compute_forward_roll
+
+static int64_t ggml_wrap_index(int64_t i, int64_t ne) {
+    if (i < 0) {
+        return i + ne;
+    } else if (i >= ne) {
+        return i - ne;
+    }
+    return i;
+}
+
+static void ggml_compute_forward_roll_f32(
+        const ggml_compute_params * params,
+        ggml_tensor * dst) {
+
+    const ggml_tensor * src0 = dst->src[0];
+    const float * src_data = (const float *) src0->data;
+    float * dst_data = (float *) dst->data;
+
+    GGML_TENSOR_UNARY_OP_LOCALS
+
+    const int s0 = ggml_get_op_params_i32(dst, 0);
+    const int s1 = ggml_get_op_params_i32(dst, 1);
+    const int s2 = ggml_get_op_params_i32(dst, 2);
+    const int s3 = ggml_get_op_params_i32(dst, 3);
+
+    const int64_t total = ne1 * ne2 * ne3;
+    const int64_t per_thread = (total + params->nth) / params->nth;
+    const int64_t start = params->ith * per_thread;
+    const int64_t end   = std::min(start + per_thread, total);
+
+    for (int64_t i = start; i < end; ++i) {
+        const int64_t i1 = i % ne1;
+        const int64_t i2 = (i / ne1) % ne2;
+        const int64_t i3 = i / (ne2 * ne1);
+        float * dst_row = dst_data + (i3*nb3 + i2*nb2 + i1*nb1) / sizeof(float);
+
+        const int64_t i01 = ggml_wrap_index(i1 - s1, ne01);
+        const int64_t i02 = ggml_wrap_index(i2 - s2, ne02);
+        const int64_t i03 = ggml_wrap_index(i3 - s3, ne03);
+        const float * src_row = src_data + (i03*nb03 + i02*nb02 + i01*nb01) / sizeof(float);
+
+        const int64_t s = ggml_wrap_index(-s0, ne00);
+        const int64_t n = ne00 - s;
+        ggml_vec_cpy_f32(n, dst_row,     src_row + s);
+        ggml_vec_cpy_f32(s, dst_row + n, src_row);
+    }
+}
+
+void ggml_compute_forward_roll(
+        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_roll_f32(params, dst);
+            } break;
+        default:
+            {
+                GGML_ABORT("fatal error");
+            }
+    }
+}
+
 // ggml_compute_forward_arange
 
 static void ggml_compute_forward_arange_f32(
@@ -7026,7 +8053,7 @@ static void ggml_compute_forward_flash_attn_ext_f16(
     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_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;
@@ -7058,7 +8085,7 @@ static void ggml_compute_forward_flash_attn_ext_f16(
             memset(VKQ32, 0, DV*sizeof(float));
         }
 
-        const ggml_fp16_t * mp = mask ? (ggml_fp16_t *)((char *) mask->data + iq1*mask->nb[1]) : NULL;
+        const ggml_fp16_t * mp = mask ? (ggml_fp16_t *)((char *) mask->data + iq1*mask->nb[1] + (iq2%mask->ne[2])*mask->nb[2] + (iq3%mask->ne[3])*mask->nb[3]) : NULL;
 
         // k indices
         const int ik3 = iq3 / rk3;
@@ -7075,7 +8102,7 @@ static void ggml_compute_forward_flash_attn_ext_f16(
         // 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;
+            const float mv = mp ? slope*GGML_CPU_FP16_TO_FP32(mp[ic]) : 0.0f;
             if (mv == -INFINITY) {
                 continue;
             }
@@ -7143,7 +8170,7 @@ static void ggml_compute_forward_flash_attn_ext_f16(
 
         if (v->type == GGML_TYPE_F16) {
             for (int64_t d = 0; d < DV; ++d) {
-                VKQ32[d] = GGML_FP16_TO_FP32(VKQ16[d]);
+                VKQ32[d] = GGML_CPU_FP16_TO_FP32(VKQ16[d]);
             }
         }
 
@@ -7596,120 +8623,210 @@ void ggml_compute_forward_ssm_conv(
 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 ggml_tensor * src0 = dst->src[0]; // s  {d_state, dim, n_head, n_seqs+}
+    const ggml_tensor * src1 = dst->src[1]; // x  {dim, n_head, n_seq_tokens, n_seqs}
+    const ggml_tensor * src2 = dst->src[2]; // dt {n_head, n_seq_tokens, n_seqs}
+    const ggml_tensor * src3 = dst->src[3]; // A  {d_state, n_head} or {1, n_head}
+    const ggml_tensor * src4 = dst->src[4]; // B  {d_state, n_group, n_seq_tokens, n_seqs}
+    const ggml_tensor * src5 = dst->src[5]; // C  {d_state, n_group, n_seq_tokens, n_seqs}
+    const ggml_tensor * src6 = dst->src[6]; // ids {n_seqs}
 
     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
+    const int64_t nc = src0->ne[0]; // d_state
+    const int64_t nr = src0->ne[1]; // dim
+    const int64_t nh = src1->ne[1]; // n_head
+    const int64_t ng = src4->ne[1];
+    const int64_t nt = src1->ne[2]; // number of tokens per sequence
+    const int64_t ns = src1->ne[3]; // number of sequences in the batch
 
-    GGML_ASSERT(ggml_nelements(src1) + ggml_nelements(src0) == ggml_nelements(dst));
+    // can't use ggml_nbytes because src1 is not necessarily contiguous
+    const int64_t s_off = ggml_nelements(src1) * ggml_element_size(src1);
+
+    GGML_ASSERT(ggml_nelements(src1) + nc*nr*nh*ns == 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));
+    GGML_ASSERT(src6->nb[0] == sizeof(int32_t));
+    // allows optimizing the modulo since n_group should be a power of 2
+    GGML_ASSERT((ng & -ng) == ng);
 
-    // rows per thread
-    const int dr = (nr + nth - 1)/nth;
+    // heads per thread
+    const int dh = (nh + 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;
+    // head range for this thread
+    const int ih0 = dh*ith;
+    const int ih1 = MIN(ih0 + dh, nh);
 
-    #ifdef __ARM_FEATURE_SVE
-        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}
+    const int32_t * ids = (const int32_t *) src6->data;
 
-                // use the output as the source for the next token-wise iterations
-                if (i2 > 0) { s0 = s; }
+    for (int i3 = 0; i3 < ns; ++i3) {
+        const float * s0 = (const float *) ((const char *) src0->data + ids[i3]*(src0->nb[3])); // {d_state, dim, nh, ns}
+              float * s  = (      float *) ((      char *) dst->data  + i3*(src0->nb[3]) + s_off); // {d_state, dim, nh, ns}
 
-                // d_inner
-                for (int i1 = 0; i1 < ir; ++i1) {
-                    float dt_soft_plus = dt[i1] <= 20.0f ? log1pf(expf(dt[i1])) : dt[i1];
-                    float x_dt = x[i1] * dt_soft_plus;
-                    svfloat32_t vx_dt = GGML_F32_VEC_SET1(x_dt);
-                    svfloat32_t vdt_soft_plus = GGML_F32_VEC_SET1(dt_soft_plus);
-                    svfloat32_t r1_vector = GGML_F32_VEC_ZERO;
+        for (int i2 = 0; i2 < nt; ++i2) {
+            const float * x  = (const float *) ((const char *) src1->data + i2*(src1->nb[2]) + i3*(src1->nb[3])); // {dim, nh, nt, ns}
+            const float * dt = (const float *) ((const char *) src2->data + i2*(src2->nb[1]) + i3*(src2->nb[2])); // {nh, nt, ns}
+            const float * A  = (const float *) ((const char *) src3->data); // {d_state, nh} or {1, nh}
+            const float * B  = (const float *) ((const char *) src4->data + i2*(src4->nb[2]) + i3*(src4->nb[3])); // {d_state, ng, nt, ns}
+            const float * C  = (const float *) ((const char *) src5->data + i2*(src5->nb[2]) + i3*(src5->nb[3])); // {d_state, ng, nt, ns}
+                  float * y  = (      float *) ((      char *) dst->data + i2*(nh*nr*sizeof(float)) + i3*(nt*nh*nr*sizeof(float))); // {dim, nh, nt, ns}
 
-                    for (int64_t k = 0; k < nc; k += svcntw()) {
-                        svfloat32_t vA = GGML_F32_VEC_LOAD(&A[i1*nc + k]);
-                        svfloat32_t vB = GGML_F32_VEC_LOAD(&B[k]);
-                        svfloat32_t vC = GGML_F32_VEC_LOAD(&C[k]);
-                        svfloat32_t vs0 = GGML_F32_VEC_LOAD(&s0[i1*nc + k]);
+            if (src3->ne[0] == 1) {
+                // Mamba-2 has a scalar decay factor per head; dA can be outside the state-wise loop
 
-                        svfloat32_t t1 = GGML_F32_VEC_MUL(vdt_soft_plus, vA);
-                        t1 = exp_ps_sve(svptrue_b32(), t1);
-                        svfloat32_t t2 = GGML_F32_VEC_MUL(vx_dt, vB);
+                // n_head
+                for (int h = ih0; h < ih1; ++h) {
+                    // ref: https://github.com/state-spaces/mamba/blob/62db608da60f6fc790b8ed9f4b3225e95ca15fde/mamba_ssm/ops/triton/softplus.py#L16
+                    const float dt_soft_plus = dt[h] <= 20.0f ? log1pf(expf(dt[h])) : dt[h];
+                    const float dA = expf(dt_soft_plus * A[h]);
 
-                        vs0 = GGML_F32_VEC_FMA(vs0, t1, t2);
-                        r1_vector = GGML_F32_VEC_ADD(GGML_F32_VEC_MUL(vs0, vC), r1_vector);
+                    // dim
+                    for (int i1 = 0; i1 < nr; ++i1) {
+                        const int ii = i1 + h*nr;
+                        const float x_dt = x[ii] * dt_soft_plus;
+                        float sumf = 0.0f;
+#if defined(GGML_SIMD)
+    #if defined(__ARM_FEATURE_SVE)
+                        const int ggml_f32_epr = svcntw();
+                        const int ggml_f32_step = 1 * ggml_f32_epr;
 
-                        GGML_F32_VEC_STORE(&s[i1*nc + k], vs0);
-                    }
-                    y[i1] = GGML_F32xt_REDUCE_ONE(r1_vector);
-                }
-            }
-        }
+                        const int np = (nc & ~(ggml_f32_step - 1));
+
+                        GGML_F32_VEC sum = GGML_F32_VEC_ZERO;
+
+                        GGML_F32_VEC adA = GGML_F32_VEC_SET1(dA);
+                        GGML_F32_VEC axdt = GGML_F32_VEC_SET1(x_dt);
+
+                        for (int i = 0; i < np; i += ggml_f32_step) {
+                            // TODO: maybe unroll more?
+                            for (int j = 0; j < 1; j++) {
+                                GGML_F32_VEC t0 = GGML_F32_VEC_LOAD(s0 + i + j*ggml_f32_epr + ii*nc);
+                                GGML_F32_VEC t1 = GGML_F32_VEC_LOAD(B + i + j*ggml_f32_epr + (h & (ng - 1))*nc);
+                                GGML_F32_VEC t2 = GGML_F32_VEC_LOAD(C + i + j*ggml_f32_epr + (h & (ng - 1))*nc);
+
+                                t0 = GGML_F32_VEC_MUL(t0, adA);
+                                t1 = GGML_F32_VEC_MUL(t1, axdt);
+
+                                t0 = GGML_F32_VEC_ADD(t0, t1);
+
+                                sum = GGML_F32_VEC_FMA(sum, t0, t2);
+
+                                GGML_F32_VEC_STORE(s + i + j*ggml_f32_epr + ii*nc, t0);
+                            }
+                        }
+
+                        sumf = GGML_F32xt_REDUCE_ONE(sum);
     #else
-        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}
+                        const int np = (nc & ~(GGML_F32_STEP - 1));
 
-                // use the output as the source for the next token-wise iterations
-                if (i2 > 0) { s0 = s; }
+                        GGML_F32_VEC sum[GGML_F32_ARR] = { GGML_F32_VEC_ZERO };
 
-                // 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;
+                        GGML_F32_VEC adA = GGML_F32_VEC_SET1(dA);
+                        GGML_F32_VEC axdt = GGML_F32_VEC_SET1(x_dt);
+
+                        GGML_F32_VEC ax[GGML_F32_ARR];
+                        GGML_F32_VEC ay[GGML_F32_ARR];
+                        GGML_F32_VEC az[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(s0 + i + j*GGML_F32_EPR + ii*nc);
+                                ay[j] = GGML_F32_VEC_LOAD(B + i + j*GGML_F32_EPR + (h & (ng - 1))*nc);
+                                az[j] = GGML_F32_VEC_LOAD(C + i + j*GGML_F32_EPR + (h & (ng - 1))*nc);
+
+                                ax[j] = GGML_F32_VEC_MUL(ax[j], adA);
+                                ay[j] = GGML_F32_VEC_MUL(ay[j], axdt);
+
+                                ax[j] = GGML_F32_VEC_ADD(ax[j], ay[j]);
+
+                                sum[j] = GGML_F32_VEC_FMA(sum[j], ax[j], az[j]);
+
+                                GGML_F32_VEC_STORE(s + i + j*GGML_F32_EPR + ii*nc, ax[j]);
+                            }
+                        }
+
+                        // reduce sum0..sum3 to sum0
+                        GGML_F32_VEC_REDUCE(sumf, sum);
+    #endif
+#else
+                        const int np = 0;
+#endif
+                        // d_state
+                        for (int i0 = np; i0 < nc; ++i0) {
+                            const int i = i0 + ii*nc;
+                            const int ig = i0 + (h & (ng - 1))*nc;
+                            // state = prev_state * dA + dB * x
+                            const float state = (s0[i] * dA) + (B[ig] * x_dt);
+                            // y = rowwise_dotprod(state, C)
+                            sumf += state * C[ig];
+                            s[i] = state;
+                        }
+                        y[ii] = sumf;
+                    }
+                }
+            } else {
+                // Mamba-1 has an element-wise decay factor for the states
+
+                // n_head
+                for (int h = ih0; h < ih1; ++h) {
+                    // ref: https://github.com/state-spaces/mamba/blob/62db608da60f6fc790b8ed9f4b3225e95ca15fde/mamba_ssm/ops/triton/softplus.py#L16
+                    const float dt_soft_plus = dt[h] <= 20.0f ? log1pf(expf(dt[h])) : dt[h];
+
+                    // dim
+                    for (int i1 = 0; i1 < nr; ++i1) {
+                        const int ii = i1 + h*nr;
+                        const float x_dt = x[ii] * dt_soft_plus;
+#if defined(__ARM_FEATURE_SVE)
+                        svfloat32_t vx_dt = GGML_F32_VEC_SET1(x_dt);
+                        svfloat32_t vdt_soft_plus = GGML_F32_VEC_SET1(dt_soft_plus);
+                        svfloat32_t r1_vector = GGML_F32_VEC_ZERO;
+
+                        // d_state
+                        // TODO: what happens when (d_state % svcntw()) != 0?
+                        for (int64_t k = 0; k < nc; k += svcntw()) {
+                            svfloat32_t vA = GGML_F32_VEC_LOAD(&A[h*nc + k]);
+                            svfloat32_t vB = GGML_F32_VEC_LOAD(&B[k + (h & (ng - 1))*nc]);
+                            svfloat32_t vC = GGML_F32_VEC_LOAD(&C[k + (h & (ng - 1))*nc]);
+                            svfloat32_t vs0 = GGML_F32_VEC_LOAD(&s0[ii*nc + k]);
+
+                            svfloat32_t t1 = GGML_F32_VEC_MUL(vdt_soft_plus, vA);
+                            t1 = exp_ps_sve(svptrue_b32(), t1);
+                            svfloat32_t t2 = GGML_F32_VEC_MUL(vx_dt, vB);
+
+                            vs0 = GGML_F32_VEC_FMA(t2, vs0, t1);
+                            r1_vector = GGML_F32_VEC_ADD(GGML_F32_VEC_MUL(vs0, vC), r1_vector);
+
+                            GGML_F32_VEC_STORE(&s[ii*nc + k], vs0);
+                        }
+                        y[ii] = GGML_F32xt_REDUCE_ONE(r1_vector);
+#else
+                        float sumf = 0.0f;
+                        // NOTE: can't really use GGML_SIMD here because d_state is usually 16
+                        //       and also because expf is used within the loop.
+                        // d_state
+                        for (int i0 = 0; i0 < nc; ++i0) {
+                            const int i = i0 + ii*nc;
+                            const int ig = i0 + (h & (ng - 1))*nc;
+                            // state = prev_state * dA + dB * x
+                            const float state = (s0[i] * expf(dt_soft_plus * A[i0 + h*nc])) + (B[ig] * x_dt);
+                            // y = rowwise_dotprod(state, C)
+                            sumf += state * C[ig];
+                            s[i] = state;
+                        }
+                        y[ii] = sumf;
+#endif
                     }
-                    y[i1] = sumf;
                 }
             }
+            // use the output as the source when it's not the first token-wise iteration
+            s0 = s;
         }
-    #endif
+    }
 }
 
 void ggml_compute_forward_ssm_scan(
@@ -7927,6 +9044,42 @@ void ggml_compute_forward_unary(
     }
 }
 
+//ggml_compute_forward_glu
+
+void ggml_compute_forward_glu(
+        const ggml_compute_params * params,
+        ggml_tensor * dst) {
+
+    const ggml_glu_op op = ggml_get_glu_op(dst);
+
+    switch (op) {
+        case GGML_GLU_OP_REGLU:
+            {
+                ggml_compute_forward_reglu(params, dst);
+            } break;
+        case GGML_GLU_OP_GEGLU:
+            {
+                ggml_compute_forward_geglu(params, dst);
+            } break;
+        case GGML_GLU_OP_SWIGLU:
+            {
+                ggml_compute_forward_swiglu(params, dst);
+            } break;
+        case GGML_GLU_OP_GEGLU_ERF:
+            {
+                ggml_compute_forward_geglu_erf(params, dst);
+            } break;
+        case GGML_GLU_OP_GEGLU_QUICK:
+            {
+                ggml_compute_forward_geglu_quick(params, dst);
+            } break;
+        default:
+            {
+                GGML_ABORT("fatal error");
+            }
+    }
+}
+
 // ggml_compute_forward_get_rel_pos
 
 static void ggml_compute_forward_get_rel_pos_f16(
@@ -8132,8 +9285,8 @@ static void ggml_compute_forward_rwkv_wkv6_f32(
         #define WKV_VECTOR_SIZE 4
     #endif
 
-    int wkv_vector_size;
     #ifdef WKV_VECTOR_SIZE
+        int wkv_vector_size;
         #if defined(__ARM_FEATURE_SVE)
             wkv_vector_size = svcntw();
         #else
@@ -8348,8 +9501,8 @@ static void ggml_compute_forward_gla_f32(
         #define GLA_VECTOR_SIZE 4
     #endif
 
-    int gla_vector_size;
     #ifdef GLA_VECTOR_SIZE
+        int gla_vector_size;
         #if defined(__ARM_FEATURE_SVE)
             gla_vector_size = svcntw();
         #else
diff --git a/ggml/src/ggml-cpu/ops.h b/ggml/src/ggml-cpu/ops.h
index dc081b9e6..3a32ec20d 100644
--- a/ggml/src/ggml-cpu/ops.h
+++ b/ggml/src/ggml-cpu/ops.h
@@ -20,6 +20,9 @@
 
 static const size_t CACHE_LINE_SIZE_F32 = CACHE_LINE_SIZE/sizeof(float);
 
+// Work buffer size for im2col operations in CONV2D
+#define GGML_IM2COL_WORK_SIZE (16 * 1024 * 1024)
+
 #ifdef __cplusplus
 extern "C" {
 #endif
@@ -53,6 +56,7 @@ void ggml_compute_forward_permute(const struct ggml_compute_params * params, str
 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_set_rows(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);
@@ -64,6 +68,7 @@ void ggml_compute_forward_clamp(const struct ggml_compute_params * params, struc
 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_2d(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_conv_2d_dw(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);
@@ -72,6 +77,7 @@ void ggml_compute_forward_pool_2d_back(const struct ggml_compute_params * params
 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_roll(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);
@@ -92,6 +98,7 @@ void ggml_compute_forward_ssm_scan(const struct ggml_compute_params * params, st
 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_glu(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);
@@ -104,6 +111,7 @@ void ggml_compute_forward_custom(const struct ggml_compute_params * params, stru
 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);
+void ggml_compute_forward_mul_mat(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 
 #ifdef __cplusplus
 }
diff --git a/ggml/src/ggml-cpu/quants.c b/ggml/src/ggml-cpu/quants.c
new file mode 100644
index 000000000..ee35ab42f
--- /dev/null
+++ b/ggml/src/ggml-cpu/quants.c
@@ -0,0 +1,1158 @@
+#define GGML_COMMON_IMPL_C
+#include "ggml-common.h"
+
+#include "ggml-cpu-impl.h"
+#include "simd-mappings.h"
+#include "ggml-quants.h"
+#include "quants.h"
+
+#include "arch-fallback.h"
+
+#include <string.h>
+#include <assert.h>
+#include <float.h>
+#include <stdlib.h> // for qsort
+#include <stdio.h>  // for GGML_ASSERT
+
+#define GROUP_MAX_EPS 1e-15f
+#define GROUP_MAX_EPS_IQ3_XXS 1e-8f
+#define GROUP_MAX_EPS_IQ2_S 1e-8f
+#define GROUP_MAX_EPS_IQ1_M 1e-7f
+#define GROUP_MAX_EPS_IQ1_S 1e-12f
+
+#define UNUSED GGML_UNUSED
+
+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 * 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 * 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 * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k) {
+    quantize_row_q5_1_ref(x, y, k);
+}
+
+void quantize_row_q8_0_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k) {
+    quantize_row_q8_0_ref(x, y, k);
+}
+
+void quantize_row_q8_1_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k) {
+    quantize_row_q8_1_ref(x, y, k);
+}
+
+//
+// 2-6 bit quantization in super-blocks
+//
+
+//========================- 2-bit (de)-quantization
+
+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 * 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 * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) {
+    assert(k % QK_K == 0);
+    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 * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) {
+    assert(k % QK_K == 0);
+    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 * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) {
+    assert(k % QK_K == 0);
+    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 * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) {
+    assert(k % QK_K == 0);
+    block_tq1_0 * GGML_RESTRICT y = vy;
+    quantize_row_tq1_0_ref(x, y, 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 * GGML_RESTRICT y = vy;
+    quantize_row_tq2_0_ref(x, y, k);
+}
+
+//===================================== Q8_K ==============================================
+
+void quantize_row_q8_K_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k) {
+    quantize_row_q8_K_ref(x, y, k);
+}
+
+//===================================== Dot products =================================
+
+void ggml_vec_dot_q4_0_q8_0_generic(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;
+
+    assert(n % qk == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q4_0 * GGML_RESTRICT x = vx;
+    const block_q8_0 * GGML_RESTRICT y = vy;
+
+    int ib = 0;
+    float sumf = 0;
+
+    for (; ib < nb; ++ib) {
+        int sumi0 = 0;
+        int sumi1 = 0;
+
+        for (int j = 0; j < qk/2; ++j) {
+            const int v0 = (x[ib].qs[j] & 0x0F) - 8;
+            const int v1 = (x[ib].qs[j] >>   4) - 8;
+
+            sumi0 += (v0 * y[ib].qs[j]);
+            sumi1 += (v1 * y[ib].qs[j + qk/2]);
+        }
+
+        int sumi = sumi0 + sumi1;
+        sumf += sumi*GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d);
+    }
+
+    *s = sumf;
+}
+
+// TODO: add WASM SIMD
+void ggml_vec_dot_q4_1_q8_1_generic(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;
+
+    assert(n % qk == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q4_1 * GGML_RESTRICT x = vx;
+    const block_q8_1 * GGML_RESTRICT y = vy;
+
+    int ib = 0;
+    float sumf = 0;
+
+    for (; ib < nb; ++ib) {
+        int sumi0 = 0;
+        int sumi1 = 0;
+
+        for (int j = 0; j < qk/2; ++j) {
+            const int v0 = (x[ib].qs[j] & 0x0F);
+            const int v1 = (x[ib].qs[j] >>   4);
+
+            sumi0 += (v0 * y[ib].qs[j]);
+            sumi1 += (v1 * y[ib].qs[j + qk/2]);
+        }
+
+        int sumi = sumi0 + sumi1;
+        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
+    }
+
+    *s = sumf;
+}
+
+void ggml_vec_dot_q5_0_q8_0_generic(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;
+
+    int ib = 0;
+    float sumf = 0;
+
+    assert(n % qk == 0);
+    assert(qk == QK5_0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q5_0 * GGML_RESTRICT x = vx;
+    const block_q8_0 * GGML_RESTRICT y = vy;
+
+    for (; ib < nb; ++ib) {
+        uint32_t qh;
+        memcpy(&qh, x[ib].qh, sizeof(qh));
+
+        int sumi0 = 0;
+        int sumi1 = 0;
+
+        for (int j = 0; j < qk/2; ++j) {
+            const uint8_t xh_0 = ((qh & (1u << (j + 0 ))) >> (j + 0 )) << 4;
+            const uint8_t xh_1 = ((qh & (1u << (j + 16))) >> (j + 12));
+
+            const int32_t x0 = (int8_t)(((x[ib].qs[j] & 0x0F) | xh_0) - 16);
+            const int32_t x1 = (int8_t)(((x[ib].qs[j] >>   4) | xh_1) - 16);
+
+            sumi0 += (x0 * y[ib].qs[j]);
+            sumi1 += (x1 * y[ib].qs[j + qk/2]);
+        }
+
+        int sumi = sumi0 + sumi1;
+        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d)) * sumi;
+    }
+
+    *s = sumf;
+}
+
+void ggml_vec_dot_q5_1_q8_1_generic(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;
+
+    int ib = 0;
+    float sumf = 0;
+
+    assert(n % qk == 0);
+    assert(qk == QK5_1);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q5_1 * GGML_RESTRICT x = vx;
+    const block_q8_1 * GGML_RESTRICT y = vy;
+
+    for (; ib < nb; ++ib) {
+        uint32_t qh;
+        memcpy(&qh, x[ib].qh, sizeof(qh));
+
+        int sumi0 = 0;
+        int sumi1 = 0;
+
+        for (int j = 0; j < qk/2; ++j) {
+            const uint8_t xh_0 = ((qh >> (j +  0)) << 4) & 0x10;
+            const uint8_t xh_1 = ((qh >> (j + 12))     ) & 0x10;
+
+            const int32_t x0 = (x[ib].qs[j] & 0xF) | xh_0;
+            const int32_t x1 = (x[ib].qs[j] >>  4) | xh_1;
+
+            sumi0 += (x0 * y[ib].qs[j]);
+            sumi1 += (x1 * y[ib].qs[j + qk/2]);
+        }
+
+        int sumi = sumi0 + sumi1;
+        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
+    }
+
+    *s = sumf;
+}
+
+void ggml_vec_dot_q8_0_q8_0_generic(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;
+
+    assert(n % qk == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q8_0 * GGML_RESTRICT x = vx;
+    const block_q8_0 * GGML_RESTRICT y = vy;
+
+    int ib = 0;
+    float sumf = 0;
+
+    for (; ib < nb; ++ib) {
+        int sumi = 0;
+
+        for (int j = 0; j < qk; j++) {
+            sumi += x[ib].qs[j]*y[ib].qs[j];
+        }
+
+        sumf += sumi*(GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d));
+    }
+
+    *s = sumf;
+}
+
+void ggml_vec_dot_tq1_0_q8_K_generic(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 * GGML_RESTRICT x = vx;
+    const block_q8_K  * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+    const uint8_t pow3[6] = {1, 3, 9, 27, 81, 243};
+
+    float sumf = 0.0f;
+
+    for (int i = 0; i < nb; ++i) {
+        int sum = 0;
+
+        for (size_t j = 0; j < sizeof(x->qs) - sizeof(x->qs) % 32; j += 32) {
+            for (size_t l = 0; l < 5; ++l) {
+                for (size_t m = 0; m < 32; ++m) {
+                    uint8_t q = x[i].qs[j + m] * pow3[l];
+                    uint16_t xi = ((uint16_t) q * 3) >> 8;
+                    sum += (xi - 1) * y[i].qs[j*5 + l*32 + m];
+                }
+            }
+        }
+        for (size_t j = sizeof(x->qs) - sizeof(x->qs) % 32; j < sizeof(x->qs); j += 16) {
+            for (size_t l = 0; l < 5; ++l) {
+                for (size_t m = 0; m < 16; ++m) {
+                    uint8_t q = x[i].qs[j + m] * pow3[l];
+                    uint16_t xi = ((uint16_t) q * 3) >> 8;
+                    sum += (xi - 1) * y[i].qs[j*5 + l*16 + m];
+                }
+            }
+        }
+
+        for (size_t l = 0; l < 4; ++l) {
+            for (size_t j = 0; j < sizeof(x->qh); ++j) {
+                uint8_t q = x[i].qh[j] * pow3[l];
+                uint16_t xi = ((uint16_t) q * 3) >> 8;
+                sum += (xi - 1) * y[i].qs[sizeof(x->qs)*5 + l*sizeof(x->qh) + j];
+            }
+        }
+
+        sumf += (float) sum * (GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d);
+    }
+
+    *s = sumf;
+}
+
+void ggml_vec_dot_tq2_0_q8_K_generic(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 * GGML_RESTRICT x = vx;
+    const block_q8_K  * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+    float sumf = 0.0f;
+
+    for (int i = 0; i < nb; ++i) {
+        int32_t sumi = 0;
+
+        for (size_t j = 0; j < sizeof(x->qs); j += 32) {
+            for (size_t l = 0; l < 4; ++l) {
+                for (size_t k = 0; k < 32; ++k) {
+                    sumi += y[i].qs[j*4 + l*32 + k] * (((x[i].qs[j + k] >> (l*2)) & 3) - 1);
+                }
+            }
+        }
+
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+
+        sumf += (float) sumi * d;
+    }
+
+    *s = sumf;
+}
+
+void ggml_vec_dot_q2_K_q8_K_generic(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 * GGML_RESTRICT x = vx;
+    const block_q8_K * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+    float sumf = 0;
+
+    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;
+
+        int summs = 0;
+        for (int j = 0; j < 16; ++j) {
+            summs += y[i].bsums[j] * (sc[j] >> 4);
+        }
+
+        const float dall = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+        const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
+
+        int isum = 0;
+        int is = 0;
+        int d;
+        for (int k = 0; k < QK_K/128; ++k) {
+            int shift = 0;
+            for (int j = 0; j < 4; ++j) {
+                d = sc[is++] & 0xF;
+                int isuml = 0;
+                for (int l =  0; l < 16; ++l) isuml += q8[l] * ((q2[l] >> shift) & 3);
+                isum += d * isuml;
+                d = sc[is++] & 0xF;
+                isuml = 0;
+                for (int l = 16; l < 32; ++l) isuml += q8[l] * ((q2[l] >> shift) & 3);
+                isum += d * isuml;
+                shift += 2;
+                q8 += 32;
+            }
+            q2 += 32;
+        }
+        sumf += dall * isum - dmin * summs;
+    }
+    *s = sumf;
+}
+
+void ggml_vec_dot_q3_K_q8_K_generic(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);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const uint32_t kmask1 = 0x03030303;
+    const uint32_t kmask2 = 0x0f0f0f0f;
+
+    const block_q3_K * GGML_RESTRICT x = vx;
+    const block_q8_K * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+    // scalar version
+    // This function is written like this so the compiler can manage to vectorize most of it
+    // Using -Ofast, GCC and clang manage to produce code that is within a factor of 2 or so from the
+    // manually vectorized version above. Every other version I tried would run at least 4 times slower.
+    // The ideal situation would be if we could just write the code once, and the compiler would
+    // automatically produce the best possible set of machine instructions, instead of us having to manually
+    // write vectorized versions for AVX, ARM_NEON, etc.
+
+    int8_t  aux8[QK_K];
+    int16_t aux16[8];
+    float   sums [8];
+    int32_t aux32[8];
+    memset(sums, 0, 8*sizeof(float));
+
+    uint32_t auxs[4];
+    const int8_t * scales = (const int8_t*)auxs;
+
+    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;
+        memset(aux32, 0, 8*sizeof(int32_t));
+        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;
+            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
+            a += 32; m <<= 1;
+            for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 2) & 3;
+            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
+            a += 32; m <<= 1;
+            for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 4) & 3;
+            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
+            a += 32; m <<= 1;
+            for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 6) & 3;
+            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
+            a += 32; m <<= 1;
+            q3 += 32;
+        }
+        a = aux8;
+
+        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);
+        for (int j = 0; j < QK_K/16; ++j) {
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
+            q8 += 8; a += 8;
+        }
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
+        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
+    }
+    for (int l = 0; l < 8; ++l) sumf += sums[l];
+    *s = sumf;
+}
+
+void ggml_vec_dot_q4_K_q8_K_generic(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);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q4_K * GGML_RESTRICT x = vx;
+    const block_q8_K * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+    static const uint32_t kmask1 = 0x3f3f3f3f;
+    static const uint32_t kmask2 = 0x0f0f0f0f;
+    static const uint32_t kmask3 = 0x03030303;
+
+    uint32_t utmp[4];
+
+    const uint8_t * scales = (const uint8_t*)&utmp[0];
+    const uint8_t * mins   = (const uint8_t*)&utmp[2];
+
+    int8_t  aux8[QK_K];
+    int16_t aux16[8];
+    float   sums [8];
+    int32_t aux32[8];
+    memset(sums, 0, 8*sizeof(float));
+
+    float sumf = 0;
+    for (int i = 0; i < nb; ++i) {
+        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 * 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;
+            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l]  >> 4);
+            a += 32; q4 += 32;
+        }
+        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;
+
+        int sumi = 0;
+        for (int j = 0; j < QK_K/16; ++j) sumi += y[i].bsums[j] * mins[j/2];
+        a = aux8;
+        int is = 0;
+        for (int j = 0; j < QK_K/32; ++j) {
+            int32_t scale = scales[is++];
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+        }
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
+        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
+        const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
+        sumf -= dmin * sumi;
+    }
+    for (int l = 0; l < 8; ++l) sumf += sums[l];
+    *s = sumf;
+}
+
+void ggml_vec_dot_q5_K_q8_K_generic(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);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q5_K * GGML_RESTRICT x = vx;
+    const block_q8_K * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+    static const uint32_t kmask1 = 0x3f3f3f3f;
+    static const uint32_t kmask2 = 0x0f0f0f0f;
+    static const uint32_t kmask3 = 0x03030303;
+
+    uint32_t utmp[4];
+
+    const uint8_t * scales = (const uint8_t*)&utmp[0];
+    const uint8_t * mins   = (const uint8_t*)&utmp[2];
+
+    int8_t  aux8[QK_K];
+    int16_t aux16[8];
+    float   sums [8];
+    int32_t aux32[8];
+    memset(sums, 0, 8*sizeof(float));
+
+    float sumf = 0;
+    for (int i = 0; i < nb; ++i) {
+        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 * 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);
+            for (int l = 0; l < 32; ++l) a[l] += (hm[l] & m ? 16 : 0);
+            a += 32; m <<= 1;
+            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l]  >> 4);
+            for (int l = 0; l < 32; ++l) a[l] += (hm[l] & m ? 16 : 0);
+            a += 32; m <<= 1;
+            q4 += 32;
+        }
+        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;
+
+        int sumi = 0;
+        for (int j = 0; j < QK_K/16; ++j) sumi += y[i].bsums[j] * mins[j/2];
+        a = aux8;
+        int is = 0;
+        for (int j = 0; j < QK_K/32; ++j) {
+            int32_t scale = scales[is++];
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+        }
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
+        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
+        const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
+        sumf -= dmin * sumi;
+    }
+    for (int l = 0; l < 8; ++l) sumf += sums[l];
+    *s = sumf;
+}
+
+void ggml_vec_dot_q6_K_q8_K_generic(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);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q6_K * GGML_RESTRICT x = vx;
+    const block_q8_K * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+    int8_t  aux8[QK_K];
+    int16_t aux16[8];
+    float   sums [8];
+    int32_t aux32[8];
+    memset(sums, 0, 8*sizeof(float));
+
+    float sumf = 0;
+    for (int i = 0; i < nb; ++i) {
+        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 * 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;
+                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;
+        }
+        a = aux8;
+        int is = 0;
+        for (int j = 0; j < QK_K/16; ++j) {
+            int scale = x[i].scales[is++];
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+        }
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
+        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
+    }
+    for (int l = 0; l < 8; ++l) sumf += sums[l];
+    *s = sumf;
+}
+
+void ggml_vec_dot_iq2_xxs_q8_K_generic(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);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_iq2_xxs * GGML_RESTRICT x = vx;
+    const block_q8_K    * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+    uint32_t aux32[2];
+    const uint8_t * aux8 = (const uint8_t *)aux32;
+
+    float sumf = 0.f;
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_CPU_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;
+        int32_t bsum = 0;
+        for (int ib32 = 0; ib32 < QK_K/32; ++ib32) {
+            memcpy(aux32, q2, 2*sizeof(uint32_t));
+            q2 += 4;
+            const uint32_t ls = 2*(aux32[1] >> 28) + 1;
+            int32_t sumi = 0;
+            for (int l = 0; l < 4; ++l) {
+                const uint8_t * grid = (const uint8_t *)(iq2xxs_grid + aux8[l]);
+                const uint8_t  signs = ksigns_iq2xs[(aux32[1] >> 7*l) & 127];
+                for (int j = 0; j < 8; ++j) {
+                    sumi += grid[j] * q8[j] * (signs & kmask_iq2xs[j] ? -1 : 1);
+                }
+                q8 += 8;
+            }
+            bsum += sumi * ls;
+        }
+        sumf += d * bsum;
+    }
+    *s = 0.125f * sumf;
+}
+
+void ggml_vec_dot_iq2_xs_q8_K_generic(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);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_iq2_xs * GGML_RESTRICT x = vx;
+    const block_q8_K   * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+    float sumf = 0.f;
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
+        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;
+            const uint16_t ls2 = 2*(sc[ib32] >>  4) + 1;
+            int32_t sumi = 0;
+            for (int l = 0; l < 2; ++l) {
+                const uint8_t * grid = (const uint8_t *)(iq2xs_grid + (q2[l] & 511));
+                const uint8_t  signs = ksigns_iq2xs[q2[l] >> 9];
+                for (int j = 0; j < 8; ++j) {
+                    sumi += grid[j] * q8[j] * (signs & kmask_iq2xs[j] ? -1 : 1);
+                }
+                q8 += 8;
+            }
+            bsum += sumi * ls1;
+            sumi = 0;
+            for (int l = 2; l < 4; ++l) {
+                const uint8_t * grid = (const uint8_t *)(iq2xs_grid + (q2[l] & 511));
+                const uint8_t  signs = ksigns_iq2xs[q2[l] >> 9];
+                for (int j = 0; j < 8; ++j) {
+                    sumi += grid[j] * q8[j] * (signs & kmask_iq2xs[j] ? -1 : 1);
+                }
+                q8 += 8;
+            }
+            bsum += sumi * ls2;
+            q2 += 4;
+        }
+        sumf += d * bsum;
+    }
+    *s = 0.125f * sumf;
+}
+
+void ggml_vec_dot_iq2_s_q8_K_generic(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);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_iq2_s * GGML_RESTRICT x = vx;
+    const block_q8_K  * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+    float sumf = 0;
+    for (int i = 0; i < nb; i++) {
+
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
+        const int8_t  * q8 = y[i].qs;
+        const uint8_t * qs = x[i].qs;
+        const uint8_t * qh = x[i].qh;
+        const uint8_t * signs = qs + QK_K/8;
+
+        int bsum = 0;
+        for (int ib32 = 0; ib32 < QK_K/32; ++ib32) {
+            int ls1 = 1 + 2*(x[i].scales[ib32] & 0xf);
+            int ls2 = 1 + 2*(x[i].scales[ib32] >>  4);
+            int sumi1 = 0, sumi2 = 0;
+            for (int l = 0; l < 2; ++l) {
+                const uint8_t * grid = (const uint8_t *)(iq2s_grid + (qs[l] | (qh[ib32] << (8-2*l) & 0x300)));
+                for (int j = 0; j < 8; ++j) {
+                    sumi1 += q8[j] * grid[j] * (signs[l] & kmask_iq2xs[j] ? -1 : 1);
+                }
+                q8 += 8;
+            }
+            for (int l = 2; l < 4; ++l) {
+                const uint8_t * grid = (const uint8_t *)(iq2s_grid + (qs[l] | (qh[ib32] << (8-2*l) & 0x300)));
+                for (int j = 0; j < 8; ++j) {
+                    sumi2 += q8[j] * grid[j] * (signs[l] & kmask_iq2xs[j] ? -1 : 1);
+                }
+                q8 += 8;
+            }
+            bsum += ls1 * sumi1 + ls2 * sumi2;
+            qs += 4;
+            signs += 4;
+        }
+
+        sumf += d * bsum;
+    }
+
+    *s = 0.125f * sumf;
+}
+
+void ggml_vec_dot_iq3_xxs_q8_K_generic(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);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_iq3_xxs * GGML_RESTRICT x = vx;
+    const block_q8_K    * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+    uint32_t aux32;
+
+    float sumf = 0.f;
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
+        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);
+            const uint32_t ls = 2*(aux32 >> 28) + 1;
+            int32_t sumi = 0;
+            for (int l = 0; l < 4; ++l) {
+                const uint8_t * grid1 = (const uint8_t *)(iq3xxs_grid + q3[2*l+0]);
+                const uint8_t * grid2 = (const uint8_t *)(iq3xxs_grid + q3[2*l+1]);
+                const uint8_t  signs = ksigns_iq2xs[(aux32 >> 7*l) & 127];
+                for (int j = 0; j < 4; ++j) {
+                    sumi += grid1[j] * q8[j+0] * (signs & kmask_iq2xs[j+0] ? -1 : 1);
+                    sumi += grid2[j] * q8[j+4] * (signs & kmask_iq2xs[j+4] ? -1 : 1);
+                }
+                q8 += 8;
+            }
+            q3 += 8;
+            bsum += sumi * ls;
+        }
+        sumf += d * bsum;
+    }
+    *s = 0.25f * sumf;
+}
+
+void ggml_vec_dot_iq3_s_q8_K_generic(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);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_iq3_s * GGML_RESTRICT x = vx;
+    const block_q8_K  * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+    float sumf = 0.f;
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
+        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;
+            const uint32_t ls2 = 2*(x[i].scales[ib32/2] >>  4) + 1;
+            int32_t sumi = 0;
+            for (int l = 0; l < 4; ++l) {
+                const uint8_t * grid1 = (const uint8_t *)(iq3s_grid + (qs[2*l+0] | ((qh[ib32+0] << (8-2*l)) & 256)));
+                const uint8_t * grid2 = (const uint8_t *)(iq3s_grid + (qs[2*l+1] | ((qh[ib32+0] << (7-2*l)) & 256)));
+                for (int j = 0; j < 4; ++j) {
+                    sumi += grid1[j] * q8[j+0] * (signs[l] & kmask_iq2xs[j+0] ? -1 : 1);
+                    sumi += grid2[j] * q8[j+4] * (signs[l] & kmask_iq2xs[j+4] ? -1 : 1);
+                }
+                q8 += 8;
+            }
+            qs += 8;
+            signs += 4;
+            bsum += sumi * ls1;
+            sumi = 0;
+            for (int l = 0; l < 4; ++l) {
+                const uint8_t * grid1 = (const uint8_t *)(iq3s_grid + (qs[2*l+0] | ((qh[ib32+1] << (8-2*l)) & 256)));
+                const uint8_t * grid2 = (const uint8_t *)(iq3s_grid + (qs[2*l+1] | ((qh[ib32+1] << (7-2*l)) & 256)));
+                for (int j = 0; j < 4; ++j) {
+                    sumi += grid1[j] * q8[j+0] * (signs[l] & kmask_iq2xs[j+0] ? -1 : 1);
+                    sumi += grid2[j] * q8[j+4] * (signs[l] & kmask_iq2xs[j+4] ? -1 : 1);
+                }
+                q8 += 8;
+            }
+            qs += 8;
+            signs += 4;
+            bsum += sumi * ls2;
+        }
+        sumf += d * bsum;
+    }
+    *s = sumf;
+}
+
+void ggml_vec_dot_iq1_s_q8_K_generic(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);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_iq1_s * GGML_RESTRICT x = vx;
+    const block_q8_K  * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+    float sumf = 0;
+    for (int i = 0; i < nb; i++) {
+
+        const int8_t   * q8 = y[i].qs;
+        const uint8_t  * qs = x[i].qs;
+        const uint16_t * qh = x[i].qh;
+
+        int sumi = 0, sumi1 = 0;
+        for (int ib = 0; ib < QK_K/32; ++ib) {
+            const int ls = 2*((qh[ib] >> 12) & 7) + 1;
+            const int delta = qh[ib] & 0x8000 ? -1 : 1;
+            int lsum = 0;
+            for (int l = 0; l < 4; ++l) {
+                const int8_t * grid = (const int8_t *)(iq1s_grid + (qs[l] | (((qh[ib] >> 3*l) & 7) << 8)));
+                for (int j = 0; j < 8; ++j) {
+                    lsum += q8[j] * grid[j];
+                }
+                q8 += 8;
+            }
+            sumi  += ls * lsum;
+            sumi1 += ls * delta * (y[i].bsums[2*ib+0] + y[i].bsums[2*ib+1]);
+            qs += 4;
+        }
+
+        sumf += GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d * (sumi + IQ1S_DELTA * sumi1);
+    }
+
+    *s = sumf;
+}
+
+void ggml_vec_dot_iq1_m_q8_K_generic(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);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_iq1_m * GGML_RESTRICT x = vx;
+    const block_q8_K  * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+    iq1m_scale_t scale;
+
+    int sum1[2], sum2[2], delta[4];
+
+    float sumf = 0;
+    for (int i = 0; i < nb; i++) {
+
+        const int8_t   * q8 = y[i].qs;
+        const uint8_t  * qs = x[i].qs;
+        const uint8_t  * qh = x[i].qh;
+        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);
+
+        int sumi1 = 0, sumi2 = 0;
+        for (int ib = 0; ib < QK_K/32; ++ib) {
+            delta[0] = qh[0] & 0x08 ? -1 : 1;
+            delta[1] = qh[0] & 0x80 ? -1 : 1;
+            delta[2] = qh[1] & 0x08 ? -1 : 1;
+            delta[3] = qh[1] & 0x80 ? -1 : 1;
+            sum1[0] = sum1[1] = sum2[0] = sum2[1] = 0;
+            for (int l = 0; l < 4; ++l) {
+                const int8_t * grid = (const int8_t *)(iq1s_grid + (qs[l] | (((uint16_t)qh[l/2] << (8 - 4*(l%2))) & 0x700)));
+                int lsum1 = 0, lsum2 = 0;
+                for (int j = 0; j < 8; ++j) {
+                    lsum1 += q8[j] * grid[j];
+                    lsum2 += q8[j];
+                }
+                q8 += 8;
+                sum1[l/2] += lsum1;
+                sum2[l/2] += lsum2*delta[l];
+            }
+
+            const int ls1 = 2*((sc[ib/2] >> (6*(ib%2)+0)) & 0x7) + 1;
+            const int ls2 = 2*((sc[ib/2] >> (6*(ib%2)+3)) & 0x7) + 1;
+
+            sumi1 += sum1[0] * ls1 + sum1[1] * ls2;
+            sumi2 += sum2[0] * ls1 + sum2[1] * ls2;
+            qs += 4;
+            qh += 2;
+        }
+
+        sumf += GGML_CPU_FP16_TO_FP32(scale.f16) * y[i].d * (sumi1 + IQ1M_DELTA * sumi2);
+    }
+
+    *s = sumf;
+}
+
+void ggml_vec_dot_iq4_nl_q8_0_generic(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);
+    assert(n % QK4_NL == 0);
+    static_assert(QK4_NL == QK8_0, "QK4_NL and QK8_0 must be the same");
+
+    const block_iq4_nl * GGML_RESTRICT x = vx;
+    const block_q8_0   * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK4_NL;
+
+    int ib = 0;
+    float sumf = 0;
+
+    for (; ib < nb; ++ib) {
+        const float d = GGML_CPU_FP16_TO_FP32(y[ib].d)*GGML_CPU_FP16_TO_FP32(x[ib].d);
+        int sumi1 = 0, sumi2 = 0;
+        for (int j = 0; j < QK4_NL/2; ++j) {
+            sumi1 += y[ib].qs[j+       0] * kvalues_iq4nl[x[ib].qs[j] & 0xf];
+            sumi2 += y[ib].qs[j+QK4_NL/2] * kvalues_iq4nl[x[ib].qs[j] >>  4];
+        }
+        sumf += d * (sumi1 + sumi2);
+    }
+    *s = sumf;
+}
+
+void ggml_vec_dot_iq4_xs_q8_K_generic(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);
+    assert(n % QK_K == 0);
+
+    const block_iq4_xs * GGML_RESTRICT x = vx;
+    const block_q8_K   * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+    float sumf = 0;
+    for (int ibl = 0; ibl < nb; ++ibl) {
+        const float d4d8 = GGML_CPU_FP16_TO_FP32(x[ibl].d) * y[ibl].d;
+        uint16_t h = x[ibl].scales_h;
+        const uint8_t * qs = x[ibl].qs;
+        const int8_t  * q8 = y[ibl].qs;
+        for (int ib = 0; ib < QK_K/32; ib += 2) {
+            const uint8_t ls1 = (x[ibl].scales_l[ib/2] & 0xf) | ((h << 4) & 0x30);
+            const uint8_t ls2 = (x[ibl].scales_l[ib/2] >>  4) | ((h << 2) & 0x30);
+            h >>= 4;
+            const float d1 = d4d8*(ls1 - 32);
+            const float d2 = d4d8*(ls2 - 32);
+            int sumi1 = 0, sumi2 = 0;
+            for (int j = 0; j < 16; ++j) {
+                sumi1 += q8[j+ 0] * kvalues_iq4nl[qs[j] & 0xf];
+                sumi2 += q8[j+16] * kvalues_iq4nl[qs[j] >>  4];
+            }
+            sumf += d1 * (sumi1 + sumi2);
+            qs += 16;
+            q8 += 32;
+            sumi1 = sumi2 = 0;
+            for (int j = 0; j < 16; ++j) {
+                sumi1 += q8[j+ 0] * kvalues_iq4nl[qs[j] & 0xf];
+                sumi2 += q8[j+16] * kvalues_iq4nl[qs[j] >>  4];
+            }
+            sumf += d2 * (sumi1 + sumi2);
+            qs += 16;
+            q8 += 32;
+        }
+    }
+    *s = sumf;
+}
+
+// ============================ 4-bit non-linear quants
+
+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 * 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-quants.h b/ggml/src/ggml-cpu/quants.h
similarity index 56%
rename from ggml/src/ggml-cpu/ggml-cpu-quants.h
rename to ggml/src/ggml-cpu/quants.h
index e33d9d473..dc4342c87 100644
--- a/ggml/src/ggml-cpu/ggml-cpu-quants.h
+++ b/ggml/src/ggml-cpu/quants.h
@@ -58,6 +58,32 @@ void ggml_vec_dot_iq4_nl_q8_0 (int n, float * GGML_RESTRICT s, size_t bs, const
 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);
 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);
 
+// Generic implementation
+void quantize_row_q8_0_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
+void quantize_row_q8_1_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
+void quantize_row_q8_K_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void ggml_vec_dot_q4_0_q8_0_generic(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);
+void ggml_vec_dot_q4_1_q8_1_generic(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);
+void ggml_vec_dot_q5_0_q8_0_generic(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);
+void ggml_vec_dot_q5_1_q8_1_generic(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);
+void ggml_vec_dot_q8_0_q8_0_generic(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);
+void ggml_vec_dot_tq1_0_q8_K_generic(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);
+void ggml_vec_dot_tq2_0_q8_K_generic(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);
+void ggml_vec_dot_q2_K_q8_K_generic(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);
+void ggml_vec_dot_q3_K_q8_K_generic(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);
+void ggml_vec_dot_q4_K_q8_K_generic(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);
+void ggml_vec_dot_q5_K_q8_K_generic(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);
+void ggml_vec_dot_q6_K_q8_K_generic(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);
+void ggml_vec_dot_iq2_xxs_q8_K_generic(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);
+void ggml_vec_dot_iq2_xs_q8_K_generic(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);
+void ggml_vec_dot_iq2_s_q8_K_generic(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);
+void ggml_vec_dot_iq3_xxs_q8_K_generic(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);
+void ggml_vec_dot_iq3_s_q8_K_generic(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);
+void ggml_vec_dot_iq1_s_q8_K_generic(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);
+void ggml_vec_dot_iq1_m_q8_K_generic(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);
+void ggml_vec_dot_iq4_nl_q8_0_generic(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);
+void ggml_vec_dot_iq4_xs_q8_K_generic(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);
+
 #ifdef __cplusplus
 }
 #endif
diff --git a/ggml/src/ggml-cpu/repack.cpp b/ggml/src/ggml-cpu/repack.cpp
new file mode 100644
index 000000000..72ee93a5a
--- /dev/null
+++ b/ggml/src/ggml-cpu/repack.cpp
@@ -0,0 +1,1571 @@
+#define GGML_COMMON_IMPL_CPP
+#define GGML_COMMON_DECL_CPP
+#include "ggml-common.h"
+#include "ggml-backend-impl.h"
+
+#include "ggml-impl.h"
+#include "ggml-cpu.h"
+#include "ggml-cpu-impl.h"
+#include "simd-mappings.h"
+#include "traits.h"
+
+#include "arch-fallback.h"
+
+#include <cmath>
+#include <cstring>
+#include <cassert>
+#include <cstdlib> // for qsort
+#include <cstdio>  // for GGML_ASSERT
+
+#include "repack.h"
+
+#if defined(__GNUC__)
+#pragma GCC diagnostic ignored "-Woverlength-strings"
+#endif
+
+#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
+// returns an interleaved block_q4_0x4
+// in the interleaved block_q4_0x4, place deltas for 4 block_q4_0 blocks
+// first, then interleave quants from 4 block_q4_0s in blocks of blck_size_interleave
+//
+// - in                  : an array of block_q4_0 pointers
+// - blck_size_interleave : the block_q4_0 quants bytes are interleaved in blocks of
+//                         blck_size_interleave bytes
+// - xor_mask            : the mask to convert the nibbles in block_q4_0 quants bytes
+//                         from bias offset form to pure sign form (this saves subtract
+//                         operations durin unpacking)
+//
+
+extern "C" {
+
+void ggml_quantize_mat_q8_0_4x4_generic(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_0x4 * GGML_RESTRICT y = (block_q8_0x4 *) vy;
+
+    // scalar
+    const int blck_size_interleave = 4;
+    float srcv[4][QK8_0];
+    float id[4];
+
+    for (int i = 0; i < nb; i++) {
+        for (int row_iter = 0; row_iter < 4; row_iter++) {
+            float amax = 0.0f; // absolute max
+
+            for (int j = 0; j < QK8_0; j++) {
+                srcv[row_iter][j] = x[row_iter * k + i * QK8_0 + j];
+                amax = MAX(amax, fabsf(srcv[row_iter][j]));
+            }
+
+            const float d = amax / ((1 << 7) - 1);
+            id[row_iter] = d ? 1.0f / d : 0.0f;
+
+            y[i].d[row_iter] = GGML_CPU_FP32_TO_FP16(d);
+        }
+
+        for (int j = 0; j < QK8_0 * 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);
+
+            float x0 = srcv[src_id][src_offset] * id[src_id];
+            y[i].qs[j] = roundf(x0);
+        }
+    }
+}
+
+void ggml_quantize_mat_q8_0_4x8_generic(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_0x4 * GGML_RESTRICT y = (block_q8_0x4 *) vy;
+
+    // scalar
+    const int blck_size_interleave = 8;
+    float srcv[4][QK8_0];
+    float id[4];
+
+    for (int i = 0; i < nb; i++) {
+        for (int row_iter = 0; row_iter < 4; row_iter++) {
+            float amax = 0.0f; // absolute max
+
+            for (int j = 0; j < QK8_0; j++) {
+                srcv[row_iter][j] = x[row_iter * k + i * QK8_0 + j];
+                amax = MAX(amax, fabsf(srcv[row_iter][j]));
+            }
+
+            const float d = amax / ((1 << 7) - 1);
+            id[row_iter] = d ? 1.0f / d : 0.0f;
+
+            y[i].d[row_iter] = GGML_CPU_FP32_TO_FP16(d);
+        }
+
+        for (int j = 0; j < QK8_0 * 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);
+
+            float x0 = srcv[src_id][src_offset] * id[src_id];
+            y[i].qs[j] = roundf(x0);
+        }
+    }
+}
+
+void ggml_quantize_mat_q8_K_4x8_generic(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;
+
+    // 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];
+        }
+    }
+}
+
+} // extern "C"
+
+template <int64_t INTER_SIZE, ggml_type PARAM_TYPE>
+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);
+    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);
+}
+
+extern "C" {
+
+void ggml_gemv_q4_0_4x4_q8_0_generic(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;
+    const int ncols_interleaved = 4;
+    const int blocklen = 4;
+
+    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);
+
+    float sumf[4];
+    int sumi;
+
+    const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
+    for (int x = 0; x < nc / ncols_interleaved; x++) {
+        const block_q4_0x4 * b_ptr = (const block_q4_0x4 *) vx + (x * nb);
+
+        for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0;
+        for (int l = 0; l < nb; l++) {
+            for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                for (int j = 0; j < ncols_interleaved; j++) {
+                    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] << 4);
+                        const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
+                        sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4;
+                    }
+                    sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
+                }
+            }
+        }
+        for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j];
+    }
+}
+
+void ggml_gemv_q4_0_4x8_q8_0_generic(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;
+    const int ncols_interleaved = 4;
+    const int blocklen = 8;
+
+    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);
+
+    float sumf[4];
+    int sumi;
+
+    const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
+    for (int x = 0; x < nc / ncols_interleaved; x++) {
+        const block_q4_0x4 * b_ptr = (const block_q4_0x4 *) vx + (x * nb);
+
+        for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0;
+        for (int l = 0; l < nb; l++) {
+            for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                for (int j = 0; j < ncols_interleaved; j++) {
+                    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] << 4);
+                        const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
+                        sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4;
+                    }
+                    sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
+                }
+            }
+        }
+        for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j];
+    }
+}
+
+void ggml_gemv_q4_0_8x8_q8_0_generic(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;
+    const int ncols_interleaved = 8;
+    const int blocklen = 8;
+
+    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);
+
+    {
+        float sumf[8];
+        int sumi;
+
+        const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
+        for (int x = 0; x < nc / ncols_interleaved; x++) {
+            const block_q4_0x8 * b_ptr = (const block_q4_0x8 *) vx + (x * nb);
+
+            for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0;
+            for (int l = 0; l < nb; l++) {
+                for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                    for (int j = 0; j < ncols_interleaved; j++) {
+                        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] << 4);
+                            const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
+                            sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4;
+                        }
+                        sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
+                    }
+                }
+            }
+            for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j];
+        }
+    }
+}
+
+void ggml_gemv_q4_K_8x8_q8_K_generic(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);
+
+    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_CPU_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_CPU_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];
+        }
+    }
+}
+
+void ggml_gemv_iq4_nl_4x4_q8_0_generic(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;
+    const int ncols_interleaved = 4;
+    const int blocklen = 4;
+
+    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);
+
+    {
+        float sumf[4];
+        int sumi;
+
+        const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
+        for (int x = 0; x < nc / ncols_interleaved; x++) {
+            const block_iq4_nlx4 * b_ptr = (const block_iq4_nlx4 *) vx + (x * nb);
+
+            for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0;
+            for (int l = 0; l < nb; l++) {
+                for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                    for (int j = 0; j < ncols_interleaved; j++) {
+                        sumi = 0;
+                        for (int i = 0; i < blocklen; ++i) {
+                            const int v0 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0x0F];
+                            const int v1 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4];
+                            sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2]));
+                        }
+                        sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
+                    }
+                }
+            }
+            for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j];
+        }
+    }
+}
+
+void ggml_gemm_q4_0_4x4_q8_0_generic(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;
+    const int ncols_interleaved = 4;
+    const int blocklen = 4;
+
+    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);
+
+    {
+        float sumf[4][4];
+        int sumi;
+
+        for (int y = 0; y < nr / 4; y++) {
+            const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
+            for (int x = 0; x < nc / ncols_interleaved; x++) {
+                const block_q4_0x4 * b_ptr = (const block_q4_0x4 *) vx + (x * nb);
+                for (int m = 0; m < 4; m++) {
+                    for (int j = 0; j < ncols_interleaved; j++) sumf[m][j] = 0.0;
+                }
+                for (int l = 0; l < nb; l++) {
+                    for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                        for (int m = 0; m < 4; m++) {
+                            for (int j = 0; j < ncols_interleaved; j++) {
+                                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] << 4);
+                                    const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
+                                    sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
+                                            (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4;
+                                }
+                                sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(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];
+                }
+            }
+        }
+    }
+}
+
+void ggml_gemm_q4_0_4x8_q8_0_generic(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;
+    const int ncols_interleaved = 4;
+    const int blocklen = 8;
+
+    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);
+
+    float sumf[4][4];
+    int sumi;
+
+    for (int y = 0; y < nr / 4; y++) {
+        const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
+        for (int x = 0; x < nc / ncols_interleaved; x++) {
+            const block_q4_0x4 * b_ptr = (const block_q4_0x4 *) vx + (x * nb);
+            for (int m = 0; m < 4; m++) {
+                for (int j = 0; j < ncols_interleaved; j++) sumf[m][j] = 0.0;
+            }
+            for (int l = 0; l < nb; l++) {
+                for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                    for (int m = 0; m < 4; m++) {
+                        for (int j = 0; j < ncols_interleaved; j++) {
+                            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] << 4);
+                                const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
+                                sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
+                                        (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4;
+                            }
+                            sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(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];
+            }
+        }
+    }
+}
+
+void ggml_gemm_q4_0_8x8_q8_0_generic(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;
+    const int ncols_interleaved = 8;
+    const int blocklen = 8;
+
+    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);
+
+    float sumf[4][8];
+    int sumi;
+
+    for (int y = 0; y < nr / 4; y++) {
+        const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
+        for (int x = 0; x < nc / ncols_interleaved; x++) {
+            const block_q4_0x8 * b_ptr = (const block_q4_0x8 *) vx + (x * nb);
+            for (int m = 0; m < 4; m++) {
+                for (int j = 0; j < ncols_interleaved; j++) sumf[m][j] = 0.0;
+            }
+            for (int l = 0; l < nb; l++) {
+                for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                    for (int m = 0; m < 4; m++) {
+                        for (int j = 0; j < ncols_interleaved; j++) {
+                            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] << 4);
+                                const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
+                                sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
+                                         (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4;
+                            }
+                            sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(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];
+            }
+        }
+    }
+}
+
+void ggml_gemm_q4_K_8x8_q8_K_generic(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);
+
+    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_CPU_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_CPU_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];
+                }
+            }
+        }
+    }
+}
+
+void ggml_gemm_iq4_nl_4x4_q8_0_generic(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;
+    const int ncols_interleaved = 4;
+    const int blocklen = 4;
+
+    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);
+
+    {
+        float sumf[4][4];
+        int sumi;
+
+        for (int y = 0; y < nr / 4; y++) {
+            const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
+            for (int x = 0; x < nc / ncols_interleaved; x++) {
+                const block_iq4_nlx4 * b_ptr = (const block_iq4_nlx4 *) vx + (x * nb);
+                for (int m = 0; m < 4; m++) {
+                    for (int j = 0; j < ncols_interleaved; j++) sumf[m][j] = 0.0;
+                }
+                for (int l = 0; l < nb; l++) {
+                    for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                        for (int m = 0; m < 4; m++) {
+                            for (int j = 0; j < ncols_interleaved; j++) {
+                                sumi = 0;
+                                for (int i = 0; i < blocklen; ++i) {
+                                    const int v0 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0x0F];
+                                    const int v1 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4];
+                                    sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
+                                            (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4]));
+                                }
+                                sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(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];
+                }
+            }
+        }
+    }
+}
+
+} // extern "C"
+
+static block_q4_0x4 make_block_q4_0x4(block_q4_0 * in, unsigned int blck_size_interleave) {
+    block_q4_0x4 out;
+
+    for (int i = 0; i < 4; i++) {
+        out.d[i] = in[i].d;
+    }
+
+    const int end = QK4_0 * 2 / blck_size_interleave;
+
+    if (blck_size_interleave == 8) {
+        const uint64_t xor_mask = 0x8888888888888888ULL;
+        for (int i = 0; i < end; ++i) {
+            int src_id = i % 4;
+            int src_offset = (i / 4) * blck_size_interleave;
+            int dst_offset = i * blck_size_interleave;
+
+            uint64_t elems;
+            // Using memcpy to avoid unaligned memory accesses
+            memcpy(&elems, &in[src_id].qs[src_offset], sizeof(uint64_t));
+            elems ^= xor_mask;
+            memcpy(&out.qs[dst_offset], &elems, sizeof(uint64_t));
+        }
+    } else if (blck_size_interleave == 4) {
+        const uint32_t xor_mask = 0x88888888;
+        for (int i = 0; i < end; ++i) {
+            int src_id = i % 4;
+            int src_offset = (i / 4) * blck_size_interleave;
+            int dst_offset = i * blck_size_interleave;
+
+            uint32_t elems;
+            memcpy(&elems, &in[src_id].qs[src_offset], sizeof(uint32_t));
+            elems ^= xor_mask;
+            memcpy(&out.qs[dst_offset], &elems, sizeof(uint32_t));
+        }
+    } else {
+        GGML_ASSERT(false);
+    }
+
+    return out;
+}
+
+// interleave 8 block_q4_0s in blocks of blck_size_interleave
+// returns an interleaved block_q4_0x8
+// in the interleaved block_q4_0x8, place deltas for 8 block_q4_0 blocks
+// first, then interleave quants from 8 block_q4_0s in blocks of blck_size_interleave
+static block_q4_0x8 make_block_q4_0x8(block_q4_0 * in, unsigned int blck_size_interleave) {
+    block_q4_0x8 out;
+
+    for (int i = 0; i < 8; i++) {
+        out.d[i] = in[i].d;
+    }
+
+    const int end = QK4_0 * 4 / blck_size_interleave;
+    const uint64_t xor_mask = 0x8888888888888888ULL;
+
+    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));
+        elems ^= xor_mask;
+        memcpy(&out.qs[dst_offset], &elems, sizeof(uint64_t));
+    }
+
+    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);
+    constexpr int nrows_interleaved = 4;
+
+    block_q4_0x4 * dst = (block_q4_0x4 *)t->data;
+    const block_q4_0 * src = (const block_q4_0 *)data;
+    block_q4_0 dst_tmp[4];
+    int nrow = ggml_nrows(t);
+    int nblocks = t->ne[0] / QK4_0;
+
+    GGML_ASSERT(data_size == nrow * nblocks * sizeof(block_q4_0));
+
+    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_0x4(dst_tmp, interleave_block);
+        }
+        src += nrows_interleaved * nblocks;
+    }
+    return 0;
+
+    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);
+    GGML_ASSERT(interleave_block == 8);
+    constexpr int nrows_interleaved = 8;
+
+    block_q4_0x8 * dst = (block_q4_0x8*)t->data;
+    const block_q4_0 * src = (const block_q4_0*) data;
+    block_q4_0 dst_tmp[8];
+    int nrow = ggml_nrows(t);
+    int nblocks = t->ne[0] / QK4_0;
+
+    GGML_ASSERT(data_size == nrow * nblocks * sizeof(block_q4_0));
+
+    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_0x8(dst_tmp, interleave_block);
+        }
+        src += nrows_interleaved * nblocks;
+    }
+    return 0;
+
+    GGML_UNUSED(data_size);
+}
+
+static block_iq4_nlx4 make_block_iq4_nlx4(block_iq4_nl * in, unsigned int blck_size_interleave) {
+    block_iq4_nlx4 out;
+
+    for (int i = 0; i < 4; i++) {
+        out.d[i] = in[i].d;
+    }
+
+    const int end = QK4_NL * 2 / blck_size_interleave;
+
+    // TODO: this branch seems wrong
+    //if (blck_size_interleave == 8) {
+    //    for (int i = 0; i < end; ++i) {
+    //        int src_id = i % 4;
+    //        int src_offset = (i / 4) * blck_size_interleave;
+    //        int dst_offset = i * blck_size_interleave;
+
+    //        // Using memcpy to avoid unaligned memory accesses
+    //        memcpy(&out.qs[dst_offset], &in[src_id].qs[src_offset], sizeof(uint64_t));
+    //    }
+    //} else
+    if (blck_size_interleave == 4) {
+        for (int i = 0; i < end; ++i) {
+            int src_id = i % 4;
+            int src_offset = (i / 4) * blck_size_interleave;
+            int dst_offset = i * blck_size_interleave;
+
+            memcpy(&out.qs[dst_offset], &in[src_id].qs[src_offset], sizeof(uint32_t));
+        }
+    } else {
+        GGML_ASSERT(false);
+    }
+
+    return out;
+}
+
+static int repack_iq4_nl_to_iq4_nl_4_bl(struct ggml_tensor * t, int interleave_block, const void * GGML_RESTRICT data, size_t data_size) {
+    GGML_ASSERT(t->type == GGML_TYPE_IQ4_NL);
+    //GGML_ASSERT(interleave_block == 4 || interleave_block == 8);
+    GGML_ASSERT(interleave_block == 4);
+
+    block_iq4_nlx4 * dst = (block_iq4_nlx4 *)t->data;
+    const block_iq4_nl * src = (const block_iq4_nl *)data;
+    block_iq4_nl dst_tmp[4];
+    int nrow = ggml_nrows(t);
+    int nrows_interleaved = 4;
+    int nblocks = t->ne[0] / QK4_0;
+
+    GGML_ASSERT(data_size == nrow * nblocks * sizeof(block_iq4_nl));
+
+    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_iq4_nlx4(dst_tmp, interleave_block);
+        }
+        src += nrows_interleaved * nblocks;
+    }
+    return 0;
+
+    GGML_UNUSED(data_size);
+}
+
+namespace ggml::cpu::repack {
+// repack
+template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS>
+int repack(struct ggml_tensor *, const void *, size_t);
+
+// TODO: generalise.
+template <> int repack<block_q4_0, 4, 4>(struct ggml_tensor * t, const void * data, size_t data_size) {
+    return repack_q4_0_to_q4_0_4_bl(t, 4, data, data_size);
+}
+
+template <> int repack<block_q4_0, 8, 4>(struct ggml_tensor * t, const void * data, size_t data_size) {
+    return repack_q4_0_to_q4_0_4_bl(t, 8, data, data_size);
+}
+
+template <> int repack<block_q4_0, 8, 8>(struct ggml_tensor * t, const void * data, size_t data_size) {
+    return repack_q4_0_to_q4_0_8_bl(t, 8, data, data_size);
+}
+
+template <> int repack<block_q4_K, 8, 8>(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<block_iq4_nl, 4, 4>(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);
+}
+
+// TODO: needs to be revisited
+//template <> int repack<block_iq4_nl, 8, 4>(struct ggml_tensor * t, const void * data, size_t data_size) {
+//    return repack_iq4_nl_to_iq4_nl_4_bl(t, 8, data, data_size);
+//}
+
+// gemv
+template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS, ggml_type PARAM_TYPE>
+void gemv(int, float *, size_t, const void *, const void *, int, int);
+
+template <> void gemv<block_q4_0, 4, 4, GGML_TYPE_Q8_0>(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<block_q4_0, 8, 4, GGML_TYPE_Q8_0>(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<block_q4_0, 8, 8, GGML_TYPE_Q8_0>(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<block_q4_K, 8, 8, GGML_TYPE_Q8_K>(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<block_iq4_nl, 4, 4, GGML_TYPE_Q8_0>(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 <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS, ggml_type PARAM_TYPE>
+void gemm(int, float *, size_t, const void *, const void *, int, int);
+
+template <> void gemm<block_q4_0, 4, 4, GGML_TYPE_Q8_0>(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<block_q4_0, 8, 4, GGML_TYPE_Q8_0>(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<block_q4_0, 8, 8, GGML_TYPE_Q8_0>(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<block_q4_K, 8, 8, GGML_TYPE_Q8_K>(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<block_iq4_nl, 4, 4, GGML_TYPE_Q8_0>(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);
+}
+
+class tensor_traits_base : public ggml::cpu::tensor_traits {
+  public:
+    virtual int repack(struct ggml_tensor * t, const void * data, size_t data_size) = 0;
+};
+
+template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS, ggml_type PARAM_TYPE> 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(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.
+
+                    const int64_t ne02 = op->src[0]->ne[2]; // n_as, n_expert
+                    const int64_t ne12 = op->src[1]->ne[2]; // n_tokens
+
+                    const size_t sizeof_mmid_row_mapping = sizeof(int64_t);
+
+                    size += sizeof_mmid_row_mapping*ne02*(ne12 + 1);
+
+                    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;
+        }
+        return false;
+    }
+
+    void forward_mul_mat(ggml_compute_params * params, ggml_tensor * op) {
+        const ggml_tensor * src0 = op->src[0];
+        const ggml_tensor * src1 = op->src[1];
+        ggml_tensor *       dst  = op;
+
+        GGML_TENSOR_BINARY_OP_LOCALS
+
+        const int ith = params->ith;
+        const int nth = params->nth;
+
+        GGML_ASSERT(ne0 == ne01);
+        GGML_ASSERT(ne1 == ne11);
+        GGML_ASSERT(ne2 == ne12);
+        GGML_ASSERT(ne3 == ne13);
+
+        // dst cannot be transposed or permuted
+        GGML_ASSERT(nb0 == sizeof(float));
+        GGML_ASSERT(nb0 <= nb1);
+        GGML_ASSERT(nb1 <= nb2);
+        GGML_ASSERT(nb2 <= nb3);
+
+        GGML_ASSERT(src1->type == GGML_TYPE_F32);
+
+        GGML_ASSERT(ggml_n_dims(op->src[0]) == 2);
+        // GGML_ASSERT(ggml_n_dims(op->src[1]) == 2);
+
+        char *       wdata = static_cast<char *>(params->wdata);
+        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(PARAM_TYPE)->from_float;
+
+        int64_t i11_processed = 0;
+        for (int64_t i11 = ith * 4; i11 < ne11 - ne11 % 4; i11 += nth * 4) {
+            ggml_quantize_mat_t<INTER_SIZE, PARAM_TYPE>((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);
+        }
+
+        ggml_barrier(params->threadpool);
+
+        const void * src1_wdata      = params->wdata;
+        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;
+        if (src0_start >= src0_end) {
+            return;
+        }
+
+        // If there are more than three rows in src1, use gemm; otherwise, use gemv.
+        if (ne11 > 3) {
+            gemm<BLOC_TYPE, INTER_SIZE, NB_COLS, PARAM_TYPE>(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<BLOC_TYPE, INTER_SIZE, NB_COLS, PARAM_TYPE>(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);
+        }
+    }
+
+    void forward_mul_mat_id(ggml_compute_params * params, ggml_tensor * op) {
+        const ggml_tensor * src0 = op->src[0];
+        const ggml_tensor * src1 = op->src[1];
+        const ggml_tensor * ids  = op->src[2];
+        ggml_tensor *       dst  = op;
+
+        GGML_TENSOR_BINARY_OP_LOCALS
+
+        const int ith = params->ith;
+        const int nth = params->nth;
+
+        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));
+        GGML_ASSERT(nb10 == ggml_type_size(src1->type));
+
+        // dst cannot be transposed or permuted
+        GGML_ASSERT(nb0 == sizeof(float));
+        GGML_ASSERT(nb0 <= nb1);
+        GGML_ASSERT(nb1 <= nb2);
+        GGML_ASSERT(nb2 <= nb3);
+
+        GGML_ASSERT(ne03 == 1);
+        GGML_ASSERT(ne13 == 1);
+        GGML_ASSERT(ne3  == 1);
+
+        GGML_ASSERT(src1->type == GGML_TYPE_F32);
+
+        // row groups
+        const int n_ids = ids->ne[0]; // n_expert_used
+        const int n_as  = ne02;       // n_expert
+
+        const size_t nbw1 = ggml_row_size(PARAM_TYPE, ne10);
+        const size_t nbw2 = nbw1*ne11;
+        const size_t nbw3 = nbw2*ne12;
+
+        struct mmid_row_mapping {
+            int32_t i1;
+            int32_t i2;
+        };
+
+        GGML_ASSERT(params->wsize >=
+                (GGML_PAD(nbw3, sizeof(int64_t)) +
+                 n_as*(ne12 + 1)*sizeof(mmid_row_mapping))
+                );
+
+        auto * wdata          = (char *)params->wdata;
+        auto * wdata_src1_end = (char *)wdata + GGML_PAD(nbw3, sizeof(int64_t));
+
+        // total of [n_as][ne12 + 1] elemets of type mmid_row_mapping (2*int32_t = 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 => 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),
+                           (void *)               (wdata + i12 * nbw2 + i11 * nbw1),
+                           ne10);
+            }
+        }
+
+#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));
+
+            // group rows by src0 matrix
+            for (int32_t iid1 = 0; iid1 < ids->ne[1]; ++iid1) {
+                for (int32_t id = 0; id < n_ids; ++id) {
+                    const int32_t i02 =
+                        *(const int32_t *) ((const char *) ids->data + iid1 * ids->nb[1] + id * ids->nb[0]);
+
+                    GGML_ASSERT(i02 >= 0 && i02 < n_as);
+
+                    MMID_MATRIX_ROW(i02, matrix_row_counts[i02]) = { id, iid1 };
+                    matrix_row_counts[i02] += 1;
+                }
+            }
+        }
+
+        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];
+
+            if (cne1 == 0) {
+                continue;
+            }
+
+            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;
+            }
+
+            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 int64_t i1 = id;  // selected expert index
+                const int64_t i2 = i12; // row
+
+                const auto * src1_col = (const char *) wdata + (i11 * nbw1 + i12 * nbw2);
+
+                gemv<BLOC_TYPE, INTER_SIZE, NB_COLS, PARAM_TYPE>(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);
+            }
+        }
+#undef MMID_MATRIX_ROW
+    }
+
+    int repack(struct ggml_tensor * t, const void * data, size_t data_size) override {
+        GGML_LOG_DEBUG("%s: repack tensor %s with %s_%dx%d\n", __func__, t->name, ggml_type_name(t->type),
+                       (int) NB_COLS, (int) INTER_SIZE);
+        return ggml::cpu::repack::repack<BLOC_TYPE, INTER_SIZE, NB_COLS>(t, data, data_size);
+    }
+};
+
+}  // namespace ggml::cpu::repack
+
+static const ggml::cpu::tensor_traits * ggml_repack_get_optimal_repack_type(const struct ggml_tensor * cur) {
+
+    // instance for Q4
+    static const ggml::cpu::repack::tensor_traits<block_q4_0, 4, 4, GGML_TYPE_Q8_0> q4_0_4x4_q8_0;
+    static const ggml::cpu::repack::tensor_traits<block_q4_0, 8, 4, GGML_TYPE_Q8_0> q4_0_4x8_q8_0;
+    static const ggml::cpu::repack::tensor_traits<block_q4_0, 8, 8, GGML_TYPE_Q8_0> q4_0_8x8_q8_0;
+    static const ggml::cpu::repack::tensor_traits<block_q4_K, 8, 8, GGML_TYPE_Q8_K> q4_K_8x8_q8_K;
+
+    // instance for IQ4
+    static const ggml::cpu::repack::tensor_traits<block_iq4_nl, 4, 4, GGML_TYPE_Q8_0> iq4_nl_4x4_q8_0;
+
+    if (cur->type == GGML_TYPE_Q4_0) {
+        if (ggml_cpu_has_avx2() || (ggml_cpu_has_sve() && ggml_cpu_has_matmul_int8() && ggml_cpu_get_sve_cnt() == QK8_0)) {
+            if (cur->ne[1] % 8 == 0) {
+                return &q4_0_8x8_q8_0;
+            }
+        }
+        if (ggml_cpu_has_neon() && ggml_cpu_has_matmul_int8()) {
+            if (cur->ne[1] % 4 == 0) {
+                return &q4_0_4x8_q8_0;
+            }
+        }
+        if (ggml_cpu_has_neon() && ggml_cpu_has_dotprod()) {
+            if (cur->ne[1] % 4 == 0) {
+                return &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 &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) {
+                return &iq4_nl_4x4_q8_0;
+            }
+        }
+    }
+
+    return nullptr;
+}
+
+static enum ggml_status ggml_backend_cpu_repack_buffer_init_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) {
+    tensor->extra = (void *) const_cast<ggml::cpu::tensor_traits *>(ggml_repack_get_optimal_repack_type(tensor));
+
+    GGML_UNUSED(buffer);
+    return GGML_STATUS_SUCCESS;
+}
+
+static void ggml_backend_cpu_repack_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::repack::tensor_traits_base *) tensor->extra;
+    auto OK            = tensor_traits->repack(tensor, data, size);
+
+    GGML_ASSERT(OK == 0);
+    GGML_UNUSED(buffer);
+}
+
+static const char * ggml_backend_cpu_repack_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
+    return "CPU_REPACK";
+
+    GGML_UNUSED(buft);
+}
+
+static ggml_backend_buffer_t ggml_backend_cpu_repack_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_repack_buffer_init_tensor;
+    buffer->iface.set_tensor  = ggml_backend_cpu_repack_buffer_set_tensor;
+    buffer->iface.get_tensor  = nullptr;
+    buffer->iface.cpy_tensor  = nullptr;
+    return buffer;
+}
+
+static size_t ggml_backend_cpu_repack_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
+    return TENSOR_ALIGNMENT;
+
+    GGML_UNUSED(buft);
+}
+
+namespace ggml::cpu::repack {
+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]->buffer &&
+                (ggml_n_dims(op->src[0]) == 2) &&
+                op->src[0]->buffer->buft == ggml_backend_cpu_repack_buffer_type() &&
+                ggml_repack_get_optimal_repack_type(op->src[0])
+                ) {
+            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) {
+                return true;
+            }
+            //if (op->src[1]->type == GGML_TYPE_Q8_0) {
+            //    return true;
+            //}
+            // may be possible if Q8_0 packed...
+        } else if (op->op == GGML_OP_MUL_MAT_ID
+                && op->src[0]->buffer
+                && (ggml_n_dims(op->src[0]) == 3)
+                && op->src[0]->buffer->buft == ggml_backend_cpu_repack_buffer_type()
+                && ggml_repack_get_optimal_repack_type(op->src[0])
+                ) {
+            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) {
+                return true;
+            }
+            //if (op->src[1]->type == GGML_TYPE_Q8_0) {
+            //    return true;
+            //}
+        }
+        return false;
+    }
+
+    ggml::cpu::tensor_traits * get_tensor_traits(const struct ggml_tensor * op) override {
+        if (op->op == GGML_OP_MUL_MAT || op->op == GGML_OP_MUL_MAT_ID) {
+            if (op->src[0]->buffer && op->src[0]->buffer->buft == ggml_backend_cpu_repack_buffer_type()) {
+                return (ggml::cpu::tensor_traits *) op->src[0]->extra;
+            }
+        }
+        return nullptr;
+    }
+};
+}  // namespace ggml::cpu::repack
+
+ggml_backend_buffer_type_t ggml_backend_cpu_repack_buffer_type(void) {
+    static struct ggml_backend_buffer_type ggml_backend_cpu_buffer_type_repack = {
+        /* .iface    = */ {
+                           /* .get_name         = */ ggml_backend_cpu_repack_buffer_type_get_name,
+                           /* .alloc_buffer     = */ ggml_backend_cpu_repack_buffer_type_alloc_buffer,
+                           /* .get_alignment    = */ ggml_backend_cpu_repack_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 = */ new ggml::cpu::repack::extra_buffer_type(),
+    };
+
+    return &ggml_backend_cpu_buffer_type_repack;
+}
diff --git a/ggml/src/ggml-cpu/repack.h b/ggml/src/ggml-cpu/repack.h
new file mode 100644
index 000000000..4421e5f8e
--- /dev/null
+++ b/ggml/src/ggml-cpu/repack.h
@@ -0,0 +1,98 @@
+#pragma once
+
+#define GGML_COMMON_DECL_CPP
+#include "ggml-common.h"
+
+#include "traits.h"
+#include "ggml.h"
+
+// GGML internal header
+
+ggml_backend_buffer_type_t ggml_backend_cpu_repack_buffer_type(void);
+
+template <int K> constexpr int QK_0() {
+    if constexpr (K == 4) {
+        return QK4_0;
+    }
+    if constexpr (K == 8) {
+        return QK8_0;
+    }
+    return -1;
+}
+
+template <int K, int N> struct block {
+    ggml_half d[N];                         // deltas for N qK_0 blocks
+    int8_t    qs[(QK_0<K>() * N * K) / 8];  // quants for N qK_0 blocks
+};
+
+// control size
+static_assert(sizeof(block<4, 4>) == 4 * sizeof(ggml_half) + QK8_0 * 2, "wrong block<4,4> size/padding");
+static_assert(sizeof(block<4, 8>) == 8 * sizeof(ggml_half) + QK8_0 * 4, "wrong block<4,8> size/padding");
+static_assert(sizeof(block<8, 4>) == 4 * sizeof(ggml_half) + QK8_0 * 4, "wrong block<8,4> size/padding");
+static_assert(sizeof(block<8, 8>) == 8 * sizeof(ggml_half) + QK8_0 * 8, "wrong block<8,8> size/padding");
+
+using block_q4_0x4 = block<4, 4>;
+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
+};
+
+static_assert(sizeof(block_iq4_nlx4) == 4 * sizeof(ggml_half) + QK4_NL * 2, "wrong iq4_nlx4 block size/padding");
+
+#if defined(__cplusplus)
+extern "C" {
+#endif
+
+void ggml_quantize_mat_q8_0_4x4(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
+void ggml_quantize_mat_q8_0_4x8(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
+void ggml_quantize_mat_q8_K_4x8(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
+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);
+void ggml_gemv_q4_0_4x8_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);
+void ggml_gemv_q4_0_8x8_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);
+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);
+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);
+void ggml_gemm_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);
+void ggml_gemm_q4_0_4x8_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);
+void ggml_gemm_q4_0_8x8_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);
+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);
+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);
+
+// Native implementations
+void ggml_quantize_mat_q8_0_4x4_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
+void ggml_quantize_mat_q8_0_4x8_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
+void ggml_quantize_mat_q8_K_4x8_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
+void ggml_gemv_q4_0_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_q4_0_4x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_q4_0_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_q4_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q4_0_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q4_0_4x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q4_0_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q4_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+
+#if defined(__cplusplus)
+} // extern "C"
+#endif
diff --git a/ggml/src/ggml-cpu/simd-mappings.h b/ggml/src/ggml-cpu/simd-mappings.h
index 2e3669c01..b4ad68c9f 100644
--- a/ggml/src/ggml-cpu/simd-mappings.h
+++ b/ggml/src/ggml-cpu/simd-mappings.h
@@ -2,10 +2,167 @@
 
 #include "ggml-cpu-impl.h"
 
+#ifdef __ARM_FEATURE_SVE
+#include <arm_sve.h>
+#endif // __ARM_FEATURE_SVE
+
+#if defined(__ARM_NEON) && !defined(__CUDACC__) && !defined(__MUSACC__)
+// if YCM cannot find <arm_neon.h>, 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 <arm_neon.h>
+#endif
+
+#if defined(__F16C__)
+#include <immintrin.h>
+#endif
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
 //
 // simd mappings
 //
 
+// FP16 to FP32 conversion
+
+// 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_CPU_COMPUTE_FP16_TO_FP32(x) neon_compute_fp16_to_fp32(x)
+    #define GGML_CPU_COMPUTE_FP32_TO_FP16(x) neon_compute_fp32_to_fp16(x)
+
+    #define GGML_CPU_FP16_TO_FP32(x) GGML_CPU_COMPUTE_FP16_TO_FP32(x)
+
+    static inline float neon_compute_fp16_to_fp32(ggml_fp16_t h) {
+        __fp16 tmp;
+        memcpy(&tmp, &h, sizeof(ggml_fp16_t));
+        return (float)tmp;
+    }
+
+    static inline ggml_fp16_t neon_compute_fp32_to_fp16(float f) {
+        ggml_fp16_t res;
+        __fp16 tmp = f;
+        memcpy(&res, &tmp, sizeof(ggml_fp16_t));
+        return res;
+    }
+#elif defined(__F16C__)
+    #ifdef _MSC_VER
+        #define GGML_CPU_COMPUTE_FP16_TO_FP32(x) _mm_cvtss_f32(_mm_cvtph_ps(_mm_cvtsi32_si128(x)))
+        #define GGML_CPU_COMPUTE_FP32_TO_FP16(x) _mm_extract_epi16(_mm_cvtps_ph(_mm_set_ss(x), 0), 0)
+    #else
+        #define GGML_CPU_COMPUTE_FP16_TO_FP32(x) _cvtsh_ss(x)
+        #define GGML_CPU_COMPUTE_FP32_TO_FP16(x) _cvtss_sh(x, 0)
+    #endif
+#elif defined(__POWER9_VECTOR__)
+    #define GGML_CPU_COMPUTE_FP16_TO_FP32(x) power_compute_fp16_to_fp32(x)
+    #define GGML_CPU_COMPUTE_FP32_TO_FP16(x) power_compute_fp32_to_fp16(x)
+    /* the inline asm below is about 12% faster than the lookup method */
+    #define GGML_CPU_FP16_TO_FP32(x) GGML_CPU_COMPUTE_FP16_TO_FP32(x)
+    #define GGML_CPU_FP32_TO_FP16(x) GGML_CPU_COMPUTE_FP32_TO_FP16(x)
+
+    static inline float power_compute_fp16_to_fp32(ggml_fp16_t h) {
+        float f;
+        double d;
+        __asm__(
+            "mtfprd %0,%2\n"
+            "xscvhpdp %0,%0\n"
+            "frsp %1,%0\n" :
+            /* temp */ "=d"(d),
+            /* out */  "=f"(f):
+            /* in */   "r"(h));
+        return f;
+    }
+
+    static inline ggml_fp16_t power_compute_fp32_to_fp16(float f) {
+        double d;
+        ggml_fp16_t r;
+        __asm__( /* xscvdphp can work on double or single precision */
+            "xscvdphp %0,%2\n"
+            "mffprd %1,%0\n" :
+            /* temp */ "=d"(d),
+            /* out */  "=r"(r):
+            /* in */   "f"(f));
+        return r;
+    }
+#elif defined(__riscv) && defined(__riscv_zfhmin)
+    static inline float riscv_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 riscv_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_CPU_COMPUTE_FP16_TO_FP32(x) riscv_compute_fp16_to_fp32(x)
+    #define GGML_CPU_COMPUTE_FP32_TO_FP16(x) riscv_compute_fp32_to_fp16(x)
+    #define GGML_CPU_FP16_TO_FP32(x) GGML_CPU_COMPUTE_FP16_TO_FP32(x)
+    #define GGML_CPU_FP32_TO_FP16(x) GGML_CPU_COMPUTE_FP32_TO_FP16(x)
+#elif defined(__NNPA__)
+    #define GGML_CPU_COMPUTE_FP16_TO_FP32(x) nnpa_compute_fp16_to_fp32(x)
+    #define GGML_CPU_COMPUTE_FP32_TO_FP16(x) nnpa_compute_fp32_to_fp16(x)
+
+    #define GGML_CPU_FP16_TO_FP32(x) GGML_CPU_COMPUTE_FP16_TO_FP32(x)
+    #define GGML_CPU_FP32_TO_FP16(x) GGML_CPU_COMPUTE_FP32_TO_FP16(x)
+
+    static inline float nnpa_compute_fp16_to_fp32(ggml_fp16_t h) {
+        uint16x8_t v_h = vec_splats(h);
+        uint16x8_t v_hd = vec_convert_from_fp16(v_h, 0);
+        return vec_extend_to_fp32_hi(v_hd, 0)[0];
+    }
+
+    static inline ggml_fp16_t nnpa_compute_fp32_to_fp16(float f) {
+        float32x4_t v_f = vec_splats(f);
+        float32x4_t v_zero = vec_splats(0.0f);
+        uint16x8_t v_hd = vec_round_from_fp32(v_f, v_zero, 0);
+        uint16x8_t v_h = vec_convert_to_fp16(v_hd, 0);
+        return vec_extract(v_h, 0);
+    }
+#endif
+
+// precomputed f32 table for f16 (256 KB)
+// defined in ggml-cpu.c, initialized in ggml_cpu_init()
+extern float ggml_table_f32_f16[1 << 16];
+
+// On ARM NEON, it's quicker to directly convert x -> x instead of calling into ggml_lookup_fp16_to_fp32,
+// so we define GGML_CPU_FP16_TO_FP32 and GGML_CPU_FP32_TO_FP16 elsewhere for NEON.
+// This is also true for POWER9.
+#if !defined(GGML_CPU_FP16_TO_FP32)
+inline static float ggml_lookup_fp16_to_fp32(ggml_fp16_t f) {
+    uint16_t s;
+    memcpy(&s, &f, sizeof(uint16_t));
+    return ggml_table_f32_f16[s];
+}
+
+#define GGML_CPU_FP16_TO_FP32(x) ggml_lookup_fp16_to_fp32(x)
+#endif
+
+#if !defined(GGML_CPU_FP32_TO_FP16)
+#define GGML_CPU_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x)
+#endif
+
+
 // 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
@@ -32,7 +189,7 @@
 #define GGML_F32xt_LOAD(...)              GGML_F32xt_LOAD_IMPL(DEFAULT_PG, __VA_ARGS__)
 #define GGML_F32xt_STORE_IMPL(pg,a,b)     svst1_f32(pg, a, b)
 #define GGML_F32xt_STORE(...)             GGML_F32xt_STORE_IMPL(DEFAULT_PG, __VA_ARGS__)
-#define GGML_F32xt_FMA_IMPL(pg, a, b, c)  svmad_f32_m(pg, a, b, c)
+#define GGML_F32xt_FMA_IMPL(pg, a, b, c)  svmad_f32_m(pg, b, c, a)
 #define GGML_F32xt_FMA(...)               GGML_F32xt_FMA_IMPL(DEFAULT_PG, __VA_ARGS__)
 #define GGML_F32xt_ADD_IMPL(pg, a, b)     svadd_f32_m(pg, a, b)
 #define GGML_F32xt_ADD(...)               GGML_F32xt_ADD_IMPL(DEFAULT_PG, __VA_ARGS__)
@@ -415,7 +572,7 @@ 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]);
+        tmp[i] = GGML_CPU_FP16_TO_FP32(x[i]);
     }
 
     return _mm256_loadu_ps(tmp);
@@ -426,7 +583,7 @@ static inline void __avx_f32cx8_store(ggml_fp16_t *x, __m256 y) {
     _mm256_storeu_ps(arr, y);
 
     for (int i = 0; i < 8; i++)
-        x[i] = GGML_FP32_TO_FP16(arr[i]);
+        x[i] = GGML_CPU_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)
@@ -574,10 +731,10 @@ static inline unsigned char ggml_endian_byte(int i) {
 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]);
+    tmp[0] = GGML_CPU_FP16_TO_FP32(p[0]);
+    tmp[1] = GGML_CPU_FP16_TO_FP32(p[1]);
+    tmp[2] = GGML_CPU_FP16_TO_FP32(p[2]);
+    tmp[3] = GGML_CPU_FP16_TO_FP32(p[3]);
 
     return wasm_v128_load(tmp);
 }
@@ -587,10 +744,10 @@ inline static void __wasm_f16x4_store(ggml_fp16_t * p, v128_t x) {
 
     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]);
+    p[0] = GGML_CPU_FP32_TO_FP16(tmp[0]);
+    p[1] = GGML_CPU_FP32_TO_FP16(tmp[1]);
+    p[2] = GGML_CPU_FP32_TO_FP16(tmp[2]);
+    p[3] = GGML_CPU_FP32_TO_FP16(tmp[3]);
 }
 
 #define GGML_F16x4             v128_t
@@ -690,10 +847,10 @@ inline static void __wasm_f16x4_store(ggml_fp16_t * p, v128_t x) {
 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]);
+    tmp[0] = GGML_CPU_FP16_TO_FP32(x[0]);
+    tmp[1] = GGML_CPU_FP16_TO_FP32(x[1]);
+    tmp[2] = GGML_CPU_FP16_TO_FP32(x[2]);
+    tmp[3] = GGML_CPU_FP16_TO_FP32(x[3]);
 
     return _mm_loadu_ps(tmp);
 }
@@ -703,10 +860,10 @@ static inline void __sse_f16x4_store(ggml_fp16_t * x, __m128 y) {
 
     _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]);
+    x[0] = GGML_CPU_FP32_TO_FP16(arr[0]);
+    x[1] = GGML_CPU_FP32_TO_FP16(arr[1]);
+    x[2] = GGML_CPU_FP32_TO_FP16(arr[2]);
+    x[3] = GGML_CPU_FP32_TO_FP16(arr[3]);
 }
 
 #define GGML_F32Cx4             __m128
@@ -828,7 +985,7 @@ static inline void __lasx_f32cx8_store(ggml_fp16_t * x, __m256 y) {
 #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_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
@@ -874,10 +1031,10 @@ static inline void __lasx_f32cx8_store(ggml_fp16_t * x, __m256 y) {
 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]);
+    tmp[0] = GGML_CPU_FP16_TO_FP32(x[0]);
+    tmp[1] = GGML_CPU_FP16_TO_FP32(x[1]);
+    tmp[2] = GGML_CPU_FP16_TO_FP32(x[2]);
+    tmp[3] = GGML_CPU_FP16_TO_FP32(x[3]);
 
     return __lsx_vld(tmp, 0);
 }
@@ -887,10 +1044,10 @@ static inline void __lsx_f16x4_store(ggml_fp16_t * x, __m128 y) {
 
     __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]);
+    x[0] = GGML_CPU_FP32_TO_FP16(arr[0]);
+    x[1] = GGML_CPU_FP32_TO_FP16(arr[1]);
+    x[2] = GGML_CPU_FP32_TO_FP16(arr[2]);
+    x[3] = GGML_CPU_FP32_TO_FP16(arr[3]);
 }
 
 #define GGML_F32Cx4             __m128
@@ -922,7 +1079,7 @@ static inline void __lsx_f16x4_store(ggml_fp16_t * x, __m128 y) {
 #define GGML_F32_STEP 32
 #define GGML_F32_EPR  4
 
-#define GGML_F32x4              __vector float
+#define GGML_F32x4              float32x4_t
 #define GGML_F32x4_ZERO         vec_splats(0.0f)
 #define GGML_F32x4_SET1         vec_splats
 #define GGML_F32x4_LOAD(p)      vec_xl(0, p)
@@ -944,10 +1101,8 @@ static inline void __lsx_f16x4_store(ggml_fp16_t * x, __m128 y) {
     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);                     \
+    float32x4_t tmp = x[0] + vec_reve(x[0]);        \
+    res = tmp[0] + tmp[1];                          \
 }
 
 #define GGML_F32_VEC        GGML_F32x4
@@ -964,28 +1119,45 @@ static inline void __lsx_f16x4_store(ggml_fp16_t * x, __m128 y) {
 #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) {
+static inline float32x4_t __lzs_f16cx4_load(const ggml_fp16_t * x) {
+#if defined(__NNPA__)
+    uint16x8_t v_x = vec_xl(0, (const ggml_fp16_t *)x);
+    uint16x8_t v_xd = vec_convert_from_fp16(v_x, 0);
+    return vec_extend_to_fp32_hi(v_xd, 0);
+#else
     float tmp[4];
 
     for (int i = 0; i < 4; i++) {
-        tmp[i] = GGML_FP16_TO_FP32(x[i]);
+        tmp[i] = GGML_CPU_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));
+#endif
 }
 
-static inline void __lzs_f16cx4_store(ggml_fp16_t * x, __vector float y) {
+static inline void __lzs_f16cx4_store(ggml_fp16_t * x, float32x4_t v_y) {
+#if defined(__NNPA__)
+    float32x4_t v_zero = vec_splats(0.0f);
+    uint16x8_t v_xd = vec_round_from_fp32(v_y, v_zero, 0);
+    uint16x8_t v_x = vec_convert_to_fp16(v_xd, 0);
+
+    x[0] = vec_extract(v_x, 0);
+    x[1] = vec_extract(v_x, 1);
+    x[2] = vec_extract(v_x, 2);
+    x[3] = vec_extract(v_x, 3);
+#else
     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));
+    vec_xst(v_y, 0, (float *)(arr));
 
     for (int i = 0; i < 4; i++) {
-        x[i] = GGML_FP32_TO_FP16(arr[i]);
+        x[i] = GGML_CPU_FP32_TO_FP16(arr[i]);
     }
+#endif
 }
 
 #define GGML_F16_VEC                GGML_F32x4
@@ -1006,3 +1178,7 @@ static inline void __lzs_f16cx4_store(ggml_fp16_t * x, __vector float y) {
 #define GGML_F32_ARR (GGML_F32_STEP/GGML_F32_EPR)
 #define GGML_F16_ARR (GGML_F16_STEP/GGML_F16_EPR)
 #endif
+
+#ifdef __cplusplus
+}
+#endif
diff --git a/ggml/src/ggml-cpu/ggml-cpu-traits.cpp b/ggml/src/ggml-cpu/traits.cpp
similarity index 97%
rename from ggml/src/ggml-cpu/ggml-cpu-traits.cpp
rename to ggml/src/ggml-cpu/traits.cpp
index 62a0712da..139fa5964 100644
--- a/ggml/src/ggml-cpu/ggml-cpu-traits.cpp
+++ b/ggml/src/ggml-cpu/traits.cpp
@@ -1,4 +1,4 @@
-#include "ggml-cpu-traits.h"
+#include "traits.h"
 
 #include "ggml-backend-impl.h"
 #include "ggml-backend.h"
diff --git a/ggml/src/ggml-cpu/ggml-cpu-traits.h b/ggml/src/ggml-cpu/traits.h
similarity index 100%
rename from ggml/src/ggml-cpu/ggml-cpu-traits.h
rename to ggml/src/ggml-cpu/traits.h
diff --git a/ggml/src/ggml-cpu/vec.cpp b/ggml/src/ggml-cpu/vec.cpp
index f7614568e..a8156011e 100644
--- a/ggml/src/ggml-cpu/vec.cpp
+++ b/ggml/src/ggml-cpu/vec.cpp
@@ -37,35 +37,35 @@ void ggml_vec_dot_f32(int n, float * GGML_RESTRICT s, size_t bs, const float * G
         for (int i = 0; i < np; i += ggml_f32_step) {
             ax1 = GGML_F32_VEC_LOAD(x + i);
             ay1 = GGML_F32_VEC_LOAD(y + i);
-            sum1 = GGML_F32_VEC_FMA(ax1, ay1, sum1);
+            sum1 = GGML_F32_VEC_FMA(sum1, ax1, ay1);
 
             ax2 = GGML_F32_VEC_LOAD(x + i + 1*ggml_f32_epr);
             ay2 = GGML_F32_VEC_LOAD(y + i + 1*ggml_f32_epr);
-            sum2 = GGML_F32_VEC_FMA(ax2, ay2, sum2);
+            sum2 = GGML_F32_VEC_FMA(sum2, ax2, ay2);
 
             ax3 = GGML_F32_VEC_LOAD(x + i + 2*ggml_f32_epr);
             ay3 = GGML_F32_VEC_LOAD(y + i + 2*ggml_f32_epr);
-            sum3 = GGML_F32_VEC_FMA(ax3, ay3, sum3);
+            sum3 = GGML_F32_VEC_FMA(sum3, ax3, ay3);
 
             ax4 = GGML_F32_VEC_LOAD(x + i + 3*ggml_f32_epr);
             ay4 = GGML_F32_VEC_LOAD(y + i + 3*ggml_f32_epr);
-            sum4 = GGML_F32_VEC_FMA(ax4, ay4, sum4);
+            sum4 = GGML_F32_VEC_FMA(sum4, ax4, ay4);
 
             ax5 = GGML_F32_VEC_LOAD(x + i + 4*ggml_f32_epr);
             ay5 = GGML_F32_VEC_LOAD(y + i + 4*ggml_f32_epr);
-            sum5 = GGML_F32_VEC_FMA(ax5, ay5, sum5);
+            sum5 = GGML_F32_VEC_FMA(sum5, ax5, ay5);
 
             ax6 = GGML_F32_VEC_LOAD(x + i + 5*ggml_f32_epr);
             ay6 = GGML_F32_VEC_LOAD(y + i + 5*ggml_f32_epr);
-            sum6 = GGML_F32_VEC_FMA(ax6, ay6, sum6);
+            sum6 = GGML_F32_VEC_FMA(sum6, ax6, ay6);
 
             ax7 = GGML_F32_VEC_LOAD(x + i + 6*ggml_f32_epr);
             ay7 = GGML_F32_VEC_LOAD(y + i + 6*ggml_f32_epr);
-            sum7 = GGML_F32_VEC_FMA(ax7, ay7, sum7);
+            sum7 = GGML_F32_VEC_FMA(sum7, ax7, ay7);
 
             ax8 = GGML_F32_VEC_LOAD(x + i + 7*ggml_f32_epr);
             ay8 = GGML_F32_VEC_LOAD(y + i + 7*ggml_f32_epr);
-            sum8 = GGML_F32_VEC_FMA(ax8, ay8, sum8);
+            sum8 = GGML_F32_VEC_FMA(sum8, ax8, ay8);
         }
         // leftovers
         // Since 8 unrolls are done in above loop, leftovers lie in range [0, ggml_f32_step] which is handled in below loop
@@ -73,7 +73,7 @@ void ggml_vec_dot_f32(int n, float * GGML_RESTRICT s, size_t bs, const float * G
         for (int i = np; i < np2; i += ggml_f32_epr) {
             ax1 = GGML_F32_VEC_LOAD(x + i);
             ay1 = GGML_F32_VEC_LOAD(y + i);
-            sum1 = GGML_F32_VEC_FMA(ax1, ay1, sum1);
+            sum1 = GGML_F32_VEC_FMA(sum1, ax1, ay1);
         }
         // maximum number of leftover elements will be less that ggml_f32_epr. Apply predicated svmad on available elements only
         if (np2 < n) {
@@ -219,11 +219,11 @@ void ggml_vec_dot_f16(int n, float * GGML_RESTRICT s, size_t bs, ggml_fp16_t * G
 
     // leftovers
     for (int i = np; i < n; ++i) {
-        sumf += (ggml_float)(GGML_FP16_TO_FP32(x[i])*GGML_FP16_TO_FP32(y[i]));
+        sumf += (ggml_float)(GGML_CPU_FP16_TO_FP32(x[i])*GGML_CPU_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]));
+        sumf += (ggml_float)(GGML_CPU_FP16_TO_FP32(x[i])*GGML_CPU_FP16_TO_FP32(y[i]));
     }
 #endif
 
@@ -254,6 +254,30 @@ void ggml_vec_silu_f32(const int n, float * y, const float * x) {
     }
 }
 
+void ggml_vec_swiglu_f32(const int n, float * y, const float * x, const float * g) {
+    int i = 0;
+#if defined(__AVX512F__) && defined(__AVX512DQ__)
+    for (; i + 15 < n; i += 16) {
+        _mm512_storeu_ps(y + i, _mm512_mul_ps(ggml_v_silu(_mm512_loadu_ps(x + i)), _mm512_loadu_ps(g + i)));
+    }
+#elif defined(__AVX2__) && defined(__FMA__)
+    for (; i + 7 < n; i += 8) {
+        _mm256_storeu_ps(y + i, _mm256_mul_ps(ggml_v_silu(_mm256_loadu_ps(x + i)), _mm256_loadu_ps(g + i)));
+    }
+#elif defined(__SSE2__)
+    for (; i + 3 < n; i += 4) {
+        _mm_storeu_ps(y + i, _mm_mul_ps(ggml_v_silu(_mm_loadu_ps(x + i)), _mm_loadu_ps(g + i)));
+    }
+#elif defined(__ARM_NEON) && defined(__aarch64__)
+    for (; i + 3 < n; i += 4) {
+        vst1q_f32(y + i, vmulq_f32(ggml_v_silu(vld1q_f32(x + i)), vld1q_f32(g + i)));
+    }
+#endif
+    for (; i < n; ++i) {
+        y[i] = ggml_silu_f32(x[i]) * g[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;
diff --git a/ggml/src/ggml-cpu/vec.h b/ggml/src/ggml-cpu/vec.h
index 09dbade21..1f5857a23 100644
--- a/ggml/src/ggml-cpu/vec.h
+++ b/ggml/src/ggml-cpu/vec.h
@@ -58,7 +58,7 @@ inline static void ggml_vec_set_bf16(const int n, ggml_bf16_t * x, const ggml_bf
 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]));
+        z[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(x[i]) + GGML_CPU_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;    }
@@ -67,7 +67,7 @@ inline static void ggml_vec_acc1_f32(const int n, float * y, const float   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]));
+        z[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(x[i]) - GGML_CPU_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;           }
@@ -75,20 +75,20 @@ inline static void ggml_vec_cpy_f32 (const int n, float * y, const float * x)
 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]));
+        y[i] = GGML_CPU_FP32_TO_FP16(-GGML_CPU_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]));
+        z[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(x[i]) * GGML_CPU_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]));
+        z[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(x[i]) / GGML_CPU_FP16_TO_FP32(y[i]));
     }
 }
 
@@ -131,13 +131,13 @@ inline static void ggml_vec_dot_f16_unroll(const int n, const int xs, float * GG
     // 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]));
+            sumf[j] += (ggml_float)(GGML_CPU_FP16_TO_FP32(x[j][i])*GGML_CPU_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]));
+            sumf[j] += (ggml_float)(GGML_CPU_FP16_TO_FP32(x[j][i])*GGML_CPU_FP16_TO_FP32(y[i]));
         }
     }
 #endif
@@ -163,49 +163,49 @@ inline static void ggml_vec_mad_f32(const int n, float * GGML_RESTRICT y, const
 
             ax1 = GGML_F32_VEC_LOAD(x + i);
             ay1 = GGML_F32_VEC_LOAD(y + i);
-            ay1 = GGML_F32_VEC_FMA(ax1, vx, ay1);
+            ay1 = GGML_F32_VEC_FMA(ay1, ax1, vx);
 
             GGML_F32_VEC_STORE(y + i, ay1);
 
             ax2 = GGML_F32_VEC_LOAD(x + i + 1*ggml_f32_epr);
             ay2 = GGML_F32_VEC_LOAD(y + i + 1*ggml_f32_epr);
-            ay2 = GGML_F32_VEC_FMA(ax2, vx, ay2);
+            ay2 = GGML_F32_VEC_FMA(ay2, ax2, vx);
 
             GGML_F32_VEC_STORE(y + i + 1*ggml_f32_epr, ay2);
 
             ax3 = GGML_F32_VEC_LOAD(x + i + 2*ggml_f32_epr);
             ay3 = GGML_F32_VEC_LOAD(y + i + 2*ggml_f32_epr);
-            ay3 = GGML_F32_VEC_FMA(ax3, vx, ay3);
+            ay3 = GGML_F32_VEC_FMA(ay3, ax3, vx);
 
             GGML_F32_VEC_STORE(y + i + 2*ggml_f32_epr, ay3);
 
             ax4 = GGML_F32_VEC_LOAD(x + i + 3*ggml_f32_epr);
             ay4 = GGML_F32_VEC_LOAD(y + i + 3*ggml_f32_epr);
-            ay4 = GGML_F32_VEC_FMA(ax4, vx, ay4);
+            ay4 = GGML_F32_VEC_FMA(ay4, ax4, vx);
 
             GGML_F32_VEC_STORE(y + i + 3*ggml_f32_epr, ay4);
 
             ax5 = GGML_F32_VEC_LOAD(x + i + 4*ggml_f32_epr);
             ay5 = GGML_F32_VEC_LOAD(y + i + 4*ggml_f32_epr);
-            ay5 = GGML_F32_VEC_FMA(ax5, vx, ay5);
+            ay5 = GGML_F32_VEC_FMA(ay5, ax5, vx);
 
             GGML_F32_VEC_STORE(y + i + 4*ggml_f32_epr, ay5);
 
             ax6 = GGML_F32_VEC_LOAD(x + i + 5*ggml_f32_epr);
             ay6 = GGML_F32_VEC_LOAD(y + i + 5*ggml_f32_epr);
-            ay6 = GGML_F32_VEC_FMA(ax6, vx, ay6);
+            ay6 = GGML_F32_VEC_FMA(ay6, ax6, vx);
 
             GGML_F32_VEC_STORE(y + i + 5*ggml_f32_epr, ay6);
 
             ax7 = GGML_F32_VEC_LOAD(x + i + 6*ggml_f32_epr);
             ay7 = GGML_F32_VEC_LOAD(y + i + 6*ggml_f32_epr);
-            ay7 = GGML_F32_VEC_FMA(ax7, vx, ay7);
+            ay7 = GGML_F32_VEC_FMA(ay7, ax7, vx);
 
             GGML_F32_VEC_STORE(y + i + 6*ggml_f32_epr, ay7);
 
             ax8 = GGML_F32_VEC_LOAD(x + i + 7*ggml_f32_epr);
             ay8 = GGML_F32_VEC_LOAD(y + i + 7*ggml_f32_epr);
-            ay8 = GGML_F32_VEC_FMA(ax8, vx, ay8);
+            ay8 = GGML_F32_VEC_FMA(ay8, ax8, vx);
 
             GGML_F32_VEC_STORE(y + i + 7*ggml_f32_epr, ay8);
         }
@@ -215,7 +215,7 @@ inline static void ggml_vec_mad_f32(const int n, float * GGML_RESTRICT y, const
         for (int i = np; i < np2; i += ggml_f32_epr) {
             ax1 = GGML_F32_VEC_LOAD(x + i);
             ay1 = GGML_F32_VEC_LOAD(y + i);
-            ay1 = GGML_F32_VEC_FMA(ax1, vx, ay1);
+            ay1 = GGML_F32_VEC_FMA(ay1, ax1, vx);
 
             GGML_F32_VEC_STORE(y + i, ay1);
         }
@@ -280,12 +280,12 @@ inline static void ggml_vec_mad_f16(const int n, ggml_fp16_t * GGML_RESTRICT y,
 
     // 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);
+        y[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(y[i]) + GGML_CPU_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);
+        y[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(y[i]) + GGML_CPU_FP16_TO_FP32(x[i])*v);
     }
 #endif
 }
@@ -430,12 +430,12 @@ inline static void ggml_vec_scale_f16(const int n, ggml_fp16_t * y, const float
 
     // leftovers
     for (int i = np; i < n; ++i) {
-        y[i] = GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(y[i])*v);
+        y[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_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);
+        y[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(y[i])*v);
     }
 #endif
 }
@@ -444,103 +444,103 @@ inline static void ggml_vec_norm_f32 (const int n, float * s, const float * x) {
 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);
+        float v = GGML_CPU_FP16_TO_FP32(x[i]);
+        y[i] = GGML_CPU_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])));
+        y[i] = GGML_CPU_FP32_TO_FP16(sqrtf(GGML_CPU_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])));
+        y[i] = GGML_CPU_FP32_TO_FP16(logf(GGML_CPU_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])));
+        y[i] = GGML_CPU_FP32_TO_FP16(sinf(GGML_CPU_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])));
+        y[i] = GGML_CPU_FP32_TO_FP16(cosf(GGML_CPU_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])));
+        y[i] = GGML_CPU_FP32_TO_FP16(fabsf(GGML_CPU_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));
+        float v = GGML_CPU_FP16_TO_FP32(x[i]);
+        y[i] = GGML_CPU_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);
+        y[i] = GGML_CPU_FP32_TO_FP16((GGML_CPU_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])));
+        y[i] = GGML_CPU_FP32_TO_FP16(tanhf(GGML_CPU_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])));
+        y[i] = GGML_CPU_FP32_TO_FP16(expm1f(GGML_CPU_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);
+        float v = GGML_CPU_FP16_TO_FP32(x[i]);
+        y[i] = GGML_CPU_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));
+        float v = GGML_CPU_FP16_TO_FP32(x[i]);
+        y[i] = GGML_CPU_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]))));
+        y[i] = GGML_CPU_FP32_TO_FP16(1.f / (1.f + expf(-GGML_CPU_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)));
+        float v = GGML_CPU_FP16_TO_FP32(x[i]);
+        y[i] = GGML_CPU_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)));
+        y[i] = GGML_CPU_FP32_TO_FP16(fminf(1.0f, fmaxf(0.0f, (GGML_CPU_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])));
+        y[i] = GGML_CPU_FP32_TO_FP16(expf(GGML_CPU_FP16_TO_FP32(x[i])));
     }
 }
 
@@ -562,9 +562,9 @@ inline static void ggml_vec_gelu_f16(const int n, ggml_fp16_t * y, const ggml_fp
 
 inline static void ggml_vec_gelu_erf_f16(const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
     for (int i = 0; i < n; ++i) {
-        float xi = GGML_FP16_TO_FP32(x[i]);
+        float xi = GGML_CPU_FP16_TO_FP32(x[i]);
         float res = 0.5f*xi*(1.0f + erff(xi*SQRT_2_INV));
-        y[i] = GGML_FP32_TO_FP16(res);
+        y[i] = GGML_CPU_FP32_TO_FP16(res);
     }
 }
 
@@ -577,9 +577,9 @@ inline static void ggml_vec_gelu_f32(const int n, float * y, const float * x) {
         } else if (x[i] >= 10.0f) {
             y[i] = x[i];
         } else {
-            ggml_fp16_t fp16 = GGML_FP32_TO_FP16(x[i]);
+            ggml_fp16_t fp16 = GGML_CPU_FP32_TO_FP16(x[i]);
             memcpy(&t, &fp16, sizeof(uint16_t));
-            y[i] = GGML_FP16_TO_FP32(ggml_table_gelu_f16[t]);
+            y[i] = GGML_CPU_FP16_TO_FP32(ggml_table_gelu_f16[t]);
         }
     }
 }
@@ -613,9 +613,9 @@ inline static float ggml_gelu_quick_f32(float x) {
 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]);
+        ggml_fp16_t fp16 = GGML_CPU_FP32_TO_FP16(x[i]);
         memcpy(&t, &fp16, sizeof(uint16_t));
-        y[i] = GGML_FP16_TO_FP32(ggml_table_gelu_quick_f16[t]);
+        y[i] = GGML_CPU_FP16_TO_FP32(ggml_table_gelu_quick_f16[t]);
     }
 }
 #else
@@ -628,8 +628,8 @@ inline static void ggml_vec_gelu_quick_f32(const int n, float * y, const float *
 
 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))));
+        float v = GGML_CPU_FP16_TO_FP32(x[i]);
+        y[i] = GGML_CPU_FP32_TO_FP16(v*(1.0f/(1.0f+expf(GELU_QUICK_COEF*v))));
     }
 }
 
@@ -638,8 +638,8 @@ 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)));
+    float v = GGML_CPU_FP16_TO_FP32(x);
+    return GGML_CPU_FP32_TO_FP16(v/(1.0f + expf(-v)));
 }
 
 #if __FINITE_MATH_ONLY__
@@ -888,9 +888,9 @@ inline static float ggml_silu_backward_f32(float x, float dy) {
 }
 
 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 v = GGML_CPU_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)));
+    return GGML_CPU_FP32_TO_FP16(GGML_CPU_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) {
@@ -905,6 +905,100 @@ inline static void ggml_vec_silu_backward_f16(const int n, ggml_fp16_t * dx, con
     }
 }
 
+inline static void ggml_vec_reglu_f32 (const int n, float * y, const float * x, const float * g) {
+    for (int i = 0; i < n; ++i) {
+        y[i] = (x[i] > 0.f) ? x[i] * g[i] : 0.f;
+    }
+}
+
+inline static void ggml_vec_reglu_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x, const ggml_fp16_t * g) {
+    for (int i = 0; i < n; ++i) {
+        float v = GGML_CPU_FP16_TO_FP32(x[i]);
+        y[i] = GGML_CPU_FP32_TO_FP16((v > 0.f) ? v * GGML_CPU_FP16_TO_FP32(g[i]) : 0.f);
+    }
+}
+
+#ifdef GGML_GELU_FP16
+inline static void ggml_vec_geglu_f32(const int n, float * y, const float * x, const float * g) {
+    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] * g[i];
+        } else {
+            ggml_fp16_t fp16 = GGML_CPU_FP32_TO_FP16(x[i]);
+            memcpy(&t, &fp16, sizeof(uint16_t));
+            y[i] = GGML_CPU_FP16_TO_FP32(ggml_table_gelu_f16[t]) * g[i];
+        }
+    }
+}
+#else
+inline static void ggml_vec_geglu_f32(const int n, float * y, const float * x, const float * g) {
+    for (int i = 0; i < n; ++i) {
+        y[i] = ggml_gelu_f32(x[i]) * g[i];
+    }
+}
+#endif
+
+inline static void ggml_vec_geglu_f16(const int n, ggml_fp16_t * y, const ggml_fp16_t * x, const ggml_fp16_t * g) {
+    const uint16_t * i16 = (const uint16_t *) x;
+    for (int i = 0; i < n; ++i) {
+        float v = GGML_CPU_FP16_TO_FP32(g[i]);
+        y[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(ggml_table_gelu_f16[i16[i]]) * v);
+    }
+}
+
+void ggml_vec_swiglu_f32(const int n, float * y, const float * x, const float * g);
+
+inline static void ggml_vec_swiglu_f16(const int n, ggml_fp16_t * y, const ggml_fp16_t * x, const ggml_fp16_t * g) {
+    for (int i = 0; i < n; ++i) {
+        float v = GGML_CPU_FP16_TO_FP32(x[i]);
+        float w = GGML_CPU_FP16_TO_FP32(g[i]);
+        y[i] = GGML_CPU_FP32_TO_FP16((v/(1.0f + expf(-v))) * w);
+    }
+}
+
+inline static void ggml_vec_geglu_erf_f32(const int n, float * y, const float * x, const float * g) {
+    for (int i = 0; i < n; ++i) {
+        float xi = x[i];
+        y[i] = 0.5f * xi * (1.0f + erff(xi*SQRT_2_INV)) * g[i];
+    }
+}
+
+inline static void ggml_vec_geglu_erf_f16(const int n, ggml_fp16_t * y, const ggml_fp16_t * x, const ggml_fp16_t * g) {
+    for (int i = 0; i < n; ++i) {
+        float xi = GGML_CPU_FP16_TO_FP32(x[i]);
+        float gi = GGML_CPU_FP16_TO_FP32(g[i]);
+        y[i] = GGML_CPU_FP32_TO_FP16(0.5f * xi * (1.0f + erff(xi*SQRT_2_INV)) * gi);
+    }
+}
+
+#ifdef GGML_GELU_QUICK_FP16
+inline static void ggml_vec_geglu_quick_f32(const int n, float * y, const float * x, const float * g) {
+    uint16_t t;
+    for (int i = 0; i < n; ++i) {
+        ggml_fp16_t fp16 = GGML_CPU_FP32_TO_FP16(x[i]);
+        memcpy(&t, &fp16, sizeof(uint16_t));
+        y[i] = GGML_CPU_FP16_TO_FP32(ggml_table_gelu_quick_f16[t]) * g[i];
+    }
+}
+#else
+inline static void ggml_vec_geglu_quick_f32(const int n, float * y, const float * x, const float * g) {
+    for (int i = 0; i < n; ++i) {
+        y[i] = ggml_gelu_quick_f32(x[i]) * g[i];
+    }
+}
+#endif
+
+inline static void ggml_vec_geglu_quick_f16(const int n, ggml_fp16_t * y, const ggml_fp16_t * x, const ggml_fp16_t * g) {
+    const uint16_t * i16 = (const uint16_t *) x;
+    for (int i = 0; i < n; ++i) {
+        float v = GGML_CPU_FP16_TO_FP32(g[i]);
+        y[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(ggml_table_gelu_quick_f16[i16[i]]) * v);
+    }
+}
+
 inline static void ggml_vec_sum_f32(const int n, float * s, const float * x) {
 #ifndef GGML_USE_ACCELERATE
     ggml_float sum = 0.0;
@@ -928,7 +1022,7 @@ inline static void ggml_vec_sum_f32_ggf(const int n, ggml_float * s, const float
 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]);
+        sum += GGML_CPU_FP16_TO_FP32(x[i]);
     }
     *s = sum;
 }
diff --git a/ggml/src/ggml-cuda/common.cuh b/ggml/src/ggml-cuda/common.cuh
index e1ce1d4cd..954f74d40 100644
--- a/ggml/src/ggml-cuda/common.cuh
+++ b/ggml/src/ggml-cuda/common.cuh
@@ -19,10 +19,10 @@
 #endif
 #include "ggml-common.h"
 
-#include <cstdio>
 #include <array>
 #include <cassert>
 #include <cfloat>
+#include <cstdio>
 #include <string>
 #include <vector>
 
@@ -76,11 +76,9 @@
 #define GGML_CUDA_CC_IS_CDNA(cc)  (cc >= GGML_CUDA_CC_CDNA && cc < GGML_CUDA_CC_RDNA1)
 
 // Moore Threads
-#define GGML_CUDA_MUSA_ARCH_IS_QY1 (__MUSA_ARCH__ <= 210)
-
-#define GGML_CUDA_CC_QY1  (GGML_CUDA_CC_OFFSET_MTHREADS + 0x210) // MTT S80, MTT S3000
-#define GGML_CUDA_CC_QY2  (GGML_CUDA_CC_OFFSET_MTHREADS + 0x220) // MTT S4000
-#define GGML_CUDA_CC_NG   (GGML_CUDA_CC_OFFSET_MTHREADS + 0x310) // TBD
+#define GGML_CUDA_CC_QY1 (GGML_CUDA_CC_OFFSET_MTHREADS + 0x210) // MTT S80, MTT S3000
+#define GGML_CUDA_CC_QY2 (GGML_CUDA_CC_OFFSET_MTHREADS + 0x220) // MTT S4000
+#define GGML_CUDA_CC_NG  (GGML_CUDA_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)
@@ -177,6 +175,20 @@ 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 !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && !defined(GGML_USE_MUSA)
+#define CUDA_SET_SHARED_MEMORY_LIMIT(kernel, nbytes) \
+    do { \
+        static bool shared_memory_limit_raised[GGML_CUDA_MAX_DEVICES] = {false}; \
+        const int id = ggml_cuda_get_device(); \
+        if (!shared_memory_limit_raised[id]) { \
+            CUDA_CHECK(cudaFuncSetAttribute(kernel, cudaFuncAttributeMaxDynamicSharedMemorySize, nbytes)); \
+            shared_memory_limit_raised[id] = true; \
+        } \
+    } while (0)
+#else
+#define CUDA_SET_SHARED_MEMORY_LIMIT(kernel, nbytes) do {} while (0)
+#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && !defined(GGML_USE_MUSA)
+
 #if CUDART_VERSION >= 11010 || defined(GGML_USE_MUSA)
 #define GGML_CUDA_ASSUME(x) __builtin_assume(x)
 #else
@@ -203,13 +215,13 @@ typedef float2 dfloat2;
 #define FAST_FP16_AVAILABLE
 #endif // defined(FP16_AVAILABLE) && __CUDA_ARCH__ != 610
 
-#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA
+#if (!defined(GGML_USE_HIP) && __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA) || defined(GGML_USE_MUSA)
 #define FP16_MMA_AVAILABLE
-#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA
+#endif // (!defined(GGML_USE_HIP) && __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA) || defined(GGML_USE_MUSA)
 
-#if defined(GGML_HIP_ROCWMMA_FATTN) && (defined(CDNA) || defined(RDNA3) || defined(RDNA4))
+#if defined(GGML_HIP_ROCWMMA_FATTN) && (defined(CDNA) || defined(RDNA3) || (defined(GGML_HIP_ROCWMMA_FATTN_GFX12) && defined(RDNA4)))
 #define FP16_MMA_AVAILABLE
-#endif // defined(GGML_HIP_ROCWMMA_FATTN) && (defined(CDNA) || defined(RDNA3) || defined(RDNA4))
+#endif // defined(GGML_HIP_ROCWMMA_FATTN) && (defined(CDNA) || defined(RDNA3) || (defined(GGML_HIP_ROCWMMA_FATTN_GFX12) && defined(RDNA4)))
 
 #if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= GGML_CUDA_CC_TURING
 #define NEW_MMA_AVAILABLE
@@ -219,9 +231,9 @@ typedef float2 dfloat2;
 #define CP_ASYNC_AVAILABLE
 #endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE
 
-#if !defined(GGML_CUDA_NO_FA) && !(defined(GGML_USE_MUSA) && GGML_CUDA_MUSA_ARCH_IS_QY1)
+#if !defined(GGML_CUDA_NO_FA) && !(defined(GGML_USE_MUSA) && __MUSA_ARCH__ < 220)
 #define FLASH_ATTN_AVAILABLE
-#endif // !defined(GGML_CUDA_NO_FA) && !(defined(GGML_USE_MUSA) && GGML_CUDA_MUSA_ARCH_IS_QY1)
+#endif // !defined(GGML_CUDA_NO_FA) && !(defined(GGML_USE_MUSA) && __MUSA_ARCH__ < 220)
 
 static bool fp16_available(const int cc) {
     return ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_PASCAL;
@@ -233,7 +245,8 @@ static bool fast_fp16_available(const int 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);
+    return (GGML_CUDA_CC_IS_NVIDIA(cc) && cc >= GGML_CUDA_CC_PASCAL && cc != 610) || GGML_CUDA_CC_IS_AMD(cc) ||
+        (GGML_CUDA_CC_IS_MTHREADS(cc) && cc >= GGML_CUDA_CC_QY2);
 }
 
 // Any FP16 tensor core instructions are available for ggml code.
@@ -241,15 +254,35 @@ 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);
+    if ((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_MTHREADS(cc)) {
+        return true;
+    } else if (GGML_CUDA_CC_IS_RDNA4(cc)) {
+#if defined(GGML_HIP_ROCWMMA_FATTN) && defined(GGML_HIP_ROCWMMA_FATTN_GFX12)
+        return true;
+#else
+        return false;
+#endif // defined(GGML_HIP_ROCWMMA_FATTN) && defined(GGML_HIP_ROCWMMA_FATTN_GFX12)
+    } else {
+        return false;
+    }
 #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);
+        GGML_CUDA_CC_IS_CDNA(cc) || GGML_CUDA_CC_IS_RDNA3(cc) || GGML_CUDA_CC_IS_RDNA4(cc) ||
+        (GGML_CUDA_CC_IS_MTHREADS(cc) && cc >= GGML_CUDA_CC_QY2);
+}
+
+static bool bf16_mma_hardware_available(const int cc) {
+    return (GGML_CUDA_CC_IS_NVIDIA(cc) && cc >= GGML_CUDA_CC_AMPERE) || GGML_CUDA_CC_IS_CDNA(cc) || cc >= GGML_CUDA_CC_RDNA3;
+}
+
+static bool fp32_mma_hardware_available(const int cc) {
+    return GGML_CUDA_CC_IS_CDNA(cc);
 }
 
 // Volta technically had FP16 tensor cores but they work very differently compared to Turing and later.
@@ -262,11 +295,11 @@ static bool cp_async_available(const int cc) {
 }
 
 static constexpr __device__ int ggml_cuda_get_physical_warp_size() {
-#if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
-    return __AMDGCN_WAVEFRONT_SIZE;
+#if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__) && (defined(__GFX9__) || defined(__GFX8__))
+    return 64;
 #else
     return 32;
-#endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
+#endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__) && (defined(__GFX9__) || defined(__GFX8__))
 }
 
 [[noreturn]]
@@ -362,6 +395,26 @@ static __device__ __forceinline__ half2 warp_reduce_sum(half2 a) {
 #endif // FP16_AVAILABLE
 }
 
+// Row reduction kernel template - compute sum (norm=false) or mean (norm=true)
+template<bool norm>
+static __global__ void reduce_rows_f32(const float * x, float * dst, const int ncols) {
+    const int row = blockIdx.x;
+    const int col = threadIdx.x;
+
+    float sum = 0.0f;
+    for (int i = col; i < ncols; i += blockDim.x) {
+        sum += x[row * ncols + i];
+    }
+
+    sum = warp_reduce_sum(sum);
+
+    if (col != 0) {
+        return;
+    }
+
+    dst[row] = norm ? sum / ncols : sum;
+}
+
 template<int width = WARP_SIZE>
 static __device__ __forceinline__ float warp_reduce_max(float x) {
 #pragma unroll
@@ -466,9 +519,6 @@ static __device__ __forceinline__ int ggml_cuda_dp4a(const int a, const int b, i
 #endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
 }
 
-// TODO: move to ggml-common.h
-static constexpr __device__ int8_t kvalues_iq4nl[16] = {-127, -104, -83, -65, -49, -35, -22, -10, 1, 13, 25, 38, 53, 69, 89, 113};
-
 typedef void (*dequantize_kernel_t)(const void * vx, const int64_t ib, const int iqs, dfloat2 & v);
 
 static __device__ __forceinline__ float get_alibi_slope(
@@ -770,21 +820,7 @@ struct ggml_backend_cuda_context {
         name(GGML_CUDA_NAME + std::to_string(device)) {
     }
 
-    ~ggml_backend_cuda_context() {
-        if (copy_event != nullptr) {
-            CUDA_CHECK(cudaEventDestroy(copy_event));
-        }
-        for (int i = 0; i < GGML_CUDA_MAX_DEVICES; ++i) {
-            for (int j = 0; j < GGML_CUDA_MAX_STREAMS; ++j) {
-                if (streams[i][j] != nullptr) {
-                    CUDA_CHECK(cudaStreamDestroy(streams[i][j]));
-                }
-            }
-            if (cublas_handles[i] != nullptr) {
-                CUBLAS_CHECK(cublasDestroy(cublas_handles[i]));
-            }
-        }
-    }
+    ~ggml_backend_cuda_context();
 
     cudaStream_t stream(int device, int stream) {
         if (streams[device][stream] == nullptr) {
diff --git a/ggml/src/ggml-cuda/conv2d-dw.cu b/ggml/src/ggml-cuda/conv2d-dw.cu
new file mode 100644
index 000000000..7583233b1
--- /dev/null
+++ b/ggml/src/ggml-cuda/conv2d-dw.cu
@@ -0,0 +1,161 @@
+#include "conv2d-dw.cuh"
+
+struct conv_params {
+    int in_w, in_h;
+    int out_w, out_h;
+    int kernel_w, kernel_h;
+    int stride_x, stride_y;
+    int padding_x, padding_y;
+    int dilation_x, dilation_y;
+    int channels, batches;
+};
+
+struct kernel_bounds {
+    int y_min, y_max;
+    int x_min, x_max;
+};
+
+__device__ __forceinline__ kernel_bounds calculate_kernel_bounds(int out_x, int out_y, const conv_params & params) {
+    kernel_bounds bounds;
+    bounds.y_min = max(0, (params.padding_y - out_y * params.stride_y + params.dilation_y - 1) / params.dilation_y);
+    bounds.y_max =
+        min(params.kernel_h,
+            (params.in_h + params.padding_y - out_y * params.stride_y + params.dilation_y - 1) / params.dilation_y);
+    bounds.x_min = max(0, (params.padding_x - out_x * params.stride_x + params.dilation_x - 1) / params.dilation_x);
+    bounds.x_max =
+        min(params.kernel_w,
+            (params.in_w + params.padding_x - out_x * params.stride_x + params.dilation_x - 1) / params.dilation_x);
+    return bounds;
+}
+
+__device__ __forceinline__ int calculate_input_coord(int out_coord, int kern_coord, int stride, int dilation, int padding) {
+    return out_coord * stride + kern_coord * dilation - padding;
+}
+
+struct whcn_layout {
+    __device__ static int input_index(int n, int c, int y, int x, const conv_params & params) {
+        return n * (params.channels * params.in_w * params.in_h) + c * params.in_w * params.in_h + y * params.in_w + x;
+    }
+
+    __device__ static int kernel_index(int c, int ky, int kx, const conv_params & params) {
+        return c * params.kernel_h * params.kernel_w + ky * params.kernel_w + kx;
+    }
+
+    __device__ static int output_index(int n, int c, int y, int x, const conv_params & params) {
+        return n * (params.channels * params.out_w * params.out_h) + c * params.out_w * params.out_h +
+               y * params.out_w + x;
+    }
+
+    __device__ static void unpack_indices(int global_idx, const conv_params & params, int & n, int & c, int & out_y,
+                                          int & out_x) {
+        out_x = global_idx % params.out_w;
+        out_y = (global_idx / params.out_w) % params.out_h;
+        c     = (global_idx / (params.out_w * params.out_h)) % params.channels;
+        n     = global_idx / (params.out_w * params.out_h * params.channels);
+    }
+};
+
+struct cwhn_layout {
+    __device__ static int input_index(int n, int c, int y, int x, const conv_params & params) {
+        return n * (params.channels * params.in_w * params.in_h) + (y * params.in_w + x) * params.channels + c;
+    }
+
+    __device__ static int kernel_index(int c, int ky, int kx, const conv_params & params) {
+        return (ky * params.kernel_w + kx) * params.channels + c;
+    }
+
+    __device__ static int output_index(int n, int c, int y, int x, const conv_params & params) {
+        return n * (params.channels * params.out_w * params.out_h) + y * (params.out_w * params.channels) +
+               x * params.channels + c;
+    }
+
+    __device__ static void unpack_indices(int global_idx, const conv_params & params, int & n, int & c, int & out_y,
+                                          int & out_x) {
+        c     = global_idx % params.channels;
+        out_x = (global_idx / params.channels) % params.out_w;
+        out_y = (global_idx / (params.channels * params.out_w)) % params.out_h;
+        n     = global_idx / (params.channels * params.out_w * params.out_h);
+    }
+};
+
+template <typename T, typename Layout>
+__global__ void conv2d_dw_kernel(const T * __restrict__ input, const T * __restrict__ kernel, T * __restrict__ output,
+                                 const int in_w, const int in_h, const int out_w, const int out_h,
+                                 const int kernel_w, const int kernel_h, const int stride_x, const int stride_y,
+                                 const int padding_x, const int padding_y, const int dilation_x, const int dilation_y,
+                                 const int channels, const int batches) {
+    const int global_idx     = blockIdx.x * blockDim.x + threadIdx.x;
+    const int total_elements = batches * channels * out_h * out_w;
+
+    if (global_idx >= total_elements) {
+        return;
+    }
+
+    conv_params params = { in_w,     in_h,      out_w,     out_h,      kernel_w,   kernel_h, stride_x,
+                           stride_y, padding_x, padding_y, dilation_x, dilation_y, channels, batches };
+
+    int batch_idx, channel_idx, out_y_idx, out_x_idx;
+    Layout::unpack_indices(global_idx, params, batch_idx, channel_idx, out_y_idx, out_x_idx);
+
+    T accumulator = 0;
+    kernel_bounds bounds = calculate_kernel_bounds(out_x_idx, out_y_idx, params);
+
+    for (int kern_y = bounds.y_min; kern_y < bounds.y_max; ++kern_y) {
+        int in_y_idx = calculate_input_coord(out_y_idx, kern_y, params.stride_y, params.dilation_y, params.padding_y);
+
+        for (int kern_x = bounds.x_min; kern_x < bounds.x_max; ++kern_x) {
+            int in_x_idx = calculate_input_coord(out_x_idx, kern_x, params.stride_x, params.dilation_x, params.padding_x);
+
+            const T input_val  = input[Layout::input_index(batch_idx, channel_idx, in_y_idx, in_x_idx, params)];
+            const T kernel_val = kernel[Layout::kernel_index(channel_idx, kern_y, kern_x, params)];
+
+            accumulator += input_val * kernel_val;
+        }
+    }
+
+    output[Layout::output_index(batch_idx, channel_idx, out_y_idx, out_x_idx, params)] = accumulator;
+}
+
+void ggml_cuda_op_conv2d_dw(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * kernel = dst->src[0];
+    const ggml_tensor * input  = dst->src[1];
+
+    GGML_ASSERT(kernel->type == GGML_TYPE_F32 && input->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32);
+    const float * w_d = (const float *) kernel->data;
+    const float * x_d = (const float *) input->data;
+    float *       y_d = (float *) dst->data;
+
+    const int32_t * p          = (const int32_t *) dst->op_params;
+    const int       stride_x   = p[0];
+    const int       stride_y   = p[1];
+    const int       padding_x  = p[2];
+    const int       padding_y  = p[3];
+    const int       dilation_x = p[4];
+    const int       dilation_y = p[5];
+
+    const int in_w     = input->ne[0];
+    const int in_h     = input->ne[1];
+    const int kernel_w = kernel->ne[0];
+    const int kernel_h = kernel->ne[1];
+    const int out_w    = dst->ne[0];
+    const int out_h    = dst->ne[1];
+    const int channels = dst->ne[2];
+    const int batches  = dst->ne[3];
+
+    cudaStream_t st = ctx.stream();
+
+    const int total  = batches * channels * out_h * out_w;
+    const int blocks = (total + CUDA_CONV2D_DW_BLOCK_SIZE - 1) / CUDA_CONV2D_DW_BLOCK_SIZE;
+
+    if (ggml_is_contiguous(input)) {
+        conv2d_dw_kernel<float, whcn_layout><<<blocks, CUDA_CONV2D_DW_BLOCK_SIZE, 0, st>>>(
+            x_d, w_d, y_d, in_w, in_h, out_w, out_h, kernel_w, kernel_h, stride_x, stride_y, padding_x, padding_y,
+            dilation_x, dilation_y, channels, batches);
+    } else if (ggml_is_contiguous_channels(input)) {
+        conv2d_dw_kernel<float, cwhn_layout><<<blocks, CUDA_CONV2D_DW_BLOCK_SIZE, 0, st>>>(
+            x_d, w_d, y_d, in_w, in_h, out_w, out_h, kernel_w, kernel_h, stride_x, stride_y, padding_x, padding_y,
+            dilation_x, dilation_y, channels, batches);
+    } else {
+        GGML_ABORT("Unsupported memory layout for conv_2d_dw");
+    }
+}
diff --git a/ggml/src/ggml-cuda/conv2d-dw.cuh b/ggml/src/ggml-cuda/conv2d-dw.cuh
new file mode 100644
index 000000000..b5d5a69d3
--- /dev/null
+++ b/ggml/src/ggml-cuda/conv2d-dw.cuh
@@ -0,0 +1,5 @@
+#pragma once
+#include "common.cuh"
+
+#define CUDA_CONV2D_DW_BLOCK_SIZE 256
+void ggml_cuda_op_conv2d_dw(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
diff --git a/ggml/src/ggml-cuda/conv2d-transpose.cu b/ggml/src/ggml-cuda/conv2d-transpose.cu
new file mode 100644
index 000000000..03224e404
--- /dev/null
+++ b/ggml/src/ggml-cuda/conv2d-transpose.cu
@@ -0,0 +1,91 @@
+#include <algorithm>
+
+#include "conv2d-transpose.cuh"
+#include "ggml.h"
+
+__global__ void conv2d_transpose_kernel(const float * __restrict__ input, const half * __restrict__ kernel,
+                                        float * __restrict__ output, const int in_w, const int in_h, const int out_w,
+                                        const int out_h, const int kernel_w, const int kernel_h, const int stride,
+                                        const int c_in, const int c_out, const int batches) {
+    const int global_idx = blockIdx.x * blockDim.x + threadIdx.x;
+
+    const int total_elements = out_w * out_h * c_out * batches;
+
+    if (global_idx >= total_elements) {
+        return;
+    }
+
+    const int out_x_idx = global_idx % out_w;
+    const int out_y_idx = (global_idx / out_w) % out_h;
+    const int c_idx     = (global_idx / (out_w * out_h)) % c_out;
+    const int n_idx     = global_idx / (out_w * out_h * c_out);
+
+    float accumulator = 0;
+    // For each output idx, find the inputs that contribute to it by checking stride alignment and bounds
+
+    for (int c_in_idx = 0; c_in_idx < c_in; c_in_idx++) {
+        for (int kh = 0; kh < kernel_h; ++kh) {
+            int in_y = out_y_idx - kh;
+            if (in_y < 0 || in_y % stride) continue;
+            in_y /= stride;
+            if (in_y >= in_h) continue;
+
+            for (int kw = 0; kw < kernel_w; ++kw) {
+                int in_x = out_x_idx - kw;
+                if (in_x < 0 || in_x % stride) continue;
+                in_x /= stride;
+                if (in_x >= in_w) continue;
+
+                const int input_idx = (in_w * in_h * c_in) * n_idx + (in_w * in_h) * c_in_idx + (in_w) *in_y + in_x;
+                const int kernel_idx =
+                    (kernel_h * kernel_w * c_out) * c_in_idx + (kernel_h * kernel_w) * c_idx + (kernel_w) *kh + kw;
+
+                float input_val = input[input_idx];
+                half  kern_val  = kernel[kernel_idx];
+
+                accumulator += input_val * (float) kern_val;
+            }
+        }
+    }
+
+    output[(out_w * out_h * c_out) * n_idx + (out_w * out_h) * c_idx + (out_w) *out_y_idx + out_x_idx] = accumulator;
+}
+
+//input is (W, H, C_in, N), Kernel is (W, H, C_out, C_in)
+void ggml_cuda_conv_2d_transpose_p0(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * kernel = dst->src[0];
+    const ggml_tensor * input  = dst->src[1];
+
+    GGML_ASSERT(kernel->type == GGML_TYPE_F16 && input->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32);
+
+    const float * input_data  = (const float *) input->data;
+    float *       output_data = (float *) dst->data;
+    const half * kernel_data = (const half *) kernel->data;
+
+    const int input_w      = input->ne[0];
+    const int input_h      = input->ne[1];
+    const int output_w     = dst->ne[0];
+    const int output_h     = dst->ne[1];
+    const int channels_in  = input->ne[2];
+    const int channels_out = kernel->ne[2];
+    const int kernel_w     = kernel->ne[0];
+    const int kernel_h     = kernel->ne[1];
+    const int stride       = dst->op_params[0];
+    const int batches      = input->ne[3];
+
+    GGML_ASSERT(channels_in == kernel->ne[3]);
+    GGML_ASSERT(stride > 0);
+
+    cudaStream_t st = ctx.stream();
+
+    GGML_ASSERT(ggml_is_contiguous(input));
+    GGML_ASSERT(ggml_is_contiguous(kernel));
+    GGML_ASSERT(ggml_is_contiguous(dst));
+
+    const int total  = (output_w * output_h * channels_out * batches);
+    const int blocks = (total + CUDA_CONV2D_TRANSPOSE_BLOCK_SIZE - 1) / CUDA_CONV2D_TRANSPOSE_BLOCK_SIZE;
+
+    conv2d_transpose_kernel<<<blocks, CUDA_CONV2D_TRANSPOSE_BLOCK_SIZE, 0, st>>>(
+        input_data, kernel_data, output_data, input_w, input_h, output_w, output_h, kernel_w, kernel_h, stride,
+        channels_in, channels_out, batches);
+}
diff --git a/ggml/src/ggml-cuda/conv2d-transpose.cuh b/ggml/src/ggml-cuda/conv2d-transpose.cuh
new file mode 100644
index 000000000..c9430b248
--- /dev/null
+++ b/ggml/src/ggml-cuda/conv2d-transpose.cuh
@@ -0,0 +1,4 @@
+#include "common.cuh"
+
+#define CUDA_CONV2D_TRANSPOSE_BLOCK_SIZE 256
+void ggml_cuda_conv_2d_transpose_p0(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
diff --git a/ggml/src/ggml-cuda/convert.cu b/ggml/src/ggml-cuda/convert.cu
index c6dec4276..eeaa14bf5 100644
--- a/ggml/src/ggml-cuda/convert.cu
+++ b/ggml/src/ggml-cuda/convert.cu
@@ -728,3 +728,25 @@ to_fp16_nc_cuda_t ggml_get_to_fp16_nc_cuda(ggml_type type) {
             return nullptr;
     }
 }
+
+to_bf16_nc_cuda_t ggml_get_to_bf16_nc_cuda(ggml_type type) {
+    switch (type) {
+        case GGML_TYPE_F32:
+            return convert_unary_cuda<float, nv_bfloat16>;
+        case GGML_TYPE_F16:
+            return convert_unary_cuda<half, nv_bfloat16>;
+        default:
+            return nullptr;
+    }
+}
+
+to_fp32_nc_cuda_t ggml_get_to_fp32_nc_cuda(ggml_type type) {
+    switch (type) {
+        case GGML_TYPE_F16:
+            return convert_unary_cuda<half, float>;
+        case GGML_TYPE_BF16:
+            return convert_unary_cuda<nv_bfloat16, float>;
+        default:
+            return nullptr;
+    }
+}
diff --git a/ggml/src/ggml-cuda/convert.cuh b/ggml/src/ggml-cuda/convert.cuh
index b65b98e08..f04214be1 100644
--- a/ggml/src/ggml-cuda/convert.cuh
+++ b/ggml/src/ggml-cuda/convert.cuh
@@ -22,5 +22,10 @@ using to_t_nc_cuda_t = void (*)(const void * x, T * y,
     int64_t ne00, int64_t ne01, int64_t ne02, int64_t ne03,
     int64_t s01, int64_t s02, int64_t s03, cudaStream_t stream);
 
+typedef to_t_nc_cuda_t<float> to_fp32_nc_cuda_t;
 typedef to_t_nc_cuda_t<half> to_fp16_nc_cuda_t;
+typedef to_t_nc_cuda_t<nv_bfloat16> to_bf16_nc_cuda_t;
+
+to_fp32_nc_cuda_t ggml_get_to_fp32_nc_cuda(ggml_type type);
 to_fp16_nc_cuda_t ggml_get_to_fp16_nc_cuda(ggml_type type);
+to_bf16_nc_cuda_t ggml_get_to_bf16_nc_cuda(ggml_type type);
diff --git a/ggml/src/ggml-cuda/cross-entropy-loss.cu b/ggml/src/ggml-cuda/cross-entropy-loss.cu
index 0ce4afbb2..0c8b08197 100644
--- a/ggml/src/ggml-cuda/cross-entropy-loss.cu
+++ b/ggml/src/ggml-cuda/cross-entropy-loss.cu
@@ -123,13 +123,7 @@ void ggml_cuda_cross_entropy_loss(ggml_backend_cuda_context & ctx, ggml_tensor *
     ggml_cuda_pool_alloc<float> dst_tmp(pool, blocks_num.x);
 
     if (nbytes_shared <= smpbo) {
-#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_f32<true>, cudaFuncAttributeMaxDynamicSharedMemorySize, smpbo));
-            shared_memory_limit_raised[id] = true;
-        }
-#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && !defined(GGML_USE_MUSA)
+        CUDA_SET_SHARED_MEMORY_LIMIT((cross_entropy_loss_f32<true>), smpbo);
         cross_entropy_loss_f32<true><<<blocks_num, blocks_dim, nbytes_shared, stream>>>(src0_d, src1_d, dst_tmp.ptr, ne00, nrows);
     } else {
         cross_entropy_loss_f32<false><<<blocks_num, blocks_dim, 0, stream>>>(src0_d, src1_d, dst_tmp.ptr, ne00, nrows);
@@ -175,13 +169,7 @@ 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__)) && !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<true>, cudaFuncAttributeMaxDynamicSharedMemorySize, smpbo));
-            shared_memory_limit_raised[id] = true;
-        }
-#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && !defined(GGML_USE_MUSA)
+        CUDA_SET_SHARED_MEMORY_LIMIT((cross_entropy_loss_back_f32<true>), smpbo);
         cross_entropy_loss_back_f32<true><<<blocks_num, blocks_dim, nbytes_shared, stream>>>(grad_d, src0f_d, src1f_d, dst_d, ne00);
     } else {
         cross_entropy_loss_back_f32<false><<<blocks_num, blocks_dim, 0, stream>>>(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 cfab2b5eb..075f14a49 100644
--- a/ggml/src/ggml-cuda/fattn-common.cuh
+++ b/ggml/src/ggml-cuda/fattn-common.cuh
@@ -32,7 +32,9 @@ typedef void (* fattn_kernel_t)(
         const int ne12,
         const int ne13,
         const int ne31,
+        const int ne32,
         const int nb31,
+        const int nb32,
         const int nb01,
         const int nb02,
         const int nb03,
@@ -851,7 +853,8 @@ void launch_fattn(
         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],
-        mask ? mask->ne[1] : 0, mask ?  mask->nb[1] : 0,
+        mask ? mask->ne[1] : 0, mask ? mask->ne[2] : 0,
+        mask ? mask->nb[1] : 0, mask ? mask->nb[2] : 0,
         Q->nb[1], Q->nb[2], Q->nb[3],
         nb11, nb12, nb13,
         nb21, nb22, nb23,
diff --git a/ggml/src/ggml-cuda/fattn-mma-f16.cuh b/ggml/src/ggml-cuda/fattn-mma-f16.cuh
index 925f39e89..709589854 100644
--- a/ggml/src/ggml-cuda/fattn-mma-f16.cuh
+++ b/ggml/src/ggml-cuda/fattn-mma-f16.cuh
@@ -652,9 +652,12 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
         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]);
+            const float KQ_max_diff = KQ_max[col] - KQ_max_new[col];
+            KQ_max_scale[col] = expf(KQ_max_diff);
             KQ_max[col] = KQ_max_new[col];
 
+            *((uint32_t *) &KQ_max_scale[col]) *= KQ_max_diff >= SOFTMAX_FTZ_THRESHOLD;
+
             // 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];
         }
@@ -1220,7 +1223,9 @@ static __global__ void flash_attn_ext_f16(
         const int ne12,
         const int ne13,
         const int ne31,
+        const int ne32,
         const int nb31,
+        const int nb32,
         const int nb01,
         const int nb02,
         const int nb03,
@@ -1285,7 +1290,8 @@ static __global__ void flash_attn_ext_f16(
 
         const float2 * Q_f2    = (const float2 *) (Q + nb02* channel*ncols2);
         const half2  * K_h2    = (const half2  *) (K + nb12*(channel*ncols2 / gqa_ratio));
-        const half2  * mask_h2 = ncols2 > 1 || mask ? (const half2  *) mask + (nb31/sizeof(half2))*jt*ncols1 : nullptr;
+        const half2  * mask_h2 = ncols2 == 1 && !mask ? nullptr :
+            (const half2  *) (mask + nb32*(channel % ne32) + nb31*jt*ncols1);
         float2       * dstk    = ((float2 *) dst) + channel*(ncols2 * DV/2);
 
         const half2 * V_h2 = mla ? K_h2 + (DKQ/2 - DV/2) : (const half2 *) (V + nb22*(channel*ncols2 / gqa_ratio));
@@ -1324,7 +1330,8 @@ static __global__ void flash_attn_ext_f16(
 
     const float2 * Q_f2    = (const float2 *) (Q + nb02* channel*ncols2);
     const half2  * K_h2    = (const half2  *) (K + nb12*(channel*ncols2 / gqa_ratio));
-    const half2  * mask_h2 = ncols2 > 1 || mask ? (const half2  *) mask + (nb31/sizeof(half2))*jt*ncols1 : nullptr;
+    const half2  * mask_h2 = ncols2 == 1 && !mask ? nullptr :
+        (const half2  *) (mask + nb32*(channel % ne32) + nb31*jt*ncols1);
     float2       * dstk    = ((float2 *) dst) + channel*(ncols2 * DV/2);
 
     const half2 * V_h2 = mla ? K_h2 + (DKQ/2 - DV/2) : (const half2 *) (V + nb22*(channel*ncols2 / gqa_ratio));
@@ -1345,8 +1352,8 @@ static __global__ void flash_attn_ext_f16(
     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(ne11); GGML_UNUSED(ne12); GGML_UNUSED(ne13); GGML_UNUSED(ne31); GGML_UNUSED(ne32);
+    GGML_UNUSED(nb31); GGML_UNUSED(nb32); 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);
diff --git a/ggml/src/ggml-cuda/fattn-tile-f16.cu b/ggml/src/ggml-cuda/fattn-tile-f16.cu
index 9283560d5..0c967f178 100644
--- a/ggml/src/ggml-cuda/fattn-tile-f16.cu
+++ b/ggml/src/ggml-cuda/fattn-tile-f16.cu
@@ -6,7 +6,7 @@
 
 template<int D, int ncols, int nwarps, bool use_logit_softcap> // D == head size
 #if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
-__launch_bounds__(nwarps*WARP_SIZE, 1)
+__launch_bounds__(nwarps*WARP_SIZE, 2)
 #endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
 static __global__ void flash_attn_tile_ext_f16(
         const char * __restrict__ Q,
@@ -30,7 +30,9 @@ static __global__ void flash_attn_tile_ext_f16(
         const int ne12,
         const int ne13,
         const int ne31,
+        const int ne32,
         const int nb31,
+        const int nb32,
         const int nb01,
         const int nb02,
         const int nb03,
@@ -64,7 +66,7 @@ static __global__ void flash_attn_tile_ext_f16(
     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 half   * maskh = (const half   *) (mask + nb32*(blockIdx.z % ne32)      + nb31*ic0);
 
     const int stride_KV2 = nb11 / sizeof(half2);
 
@@ -288,8 +290,8 @@ static __global__ void flash_attn_tile_ext_f16(
     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(ne12); GGML_UNUSED(ne13); GGML_UNUSED(ne31); GGML_UNUSED(ne32);
+    GGML_UNUSED(nb31); GGML_UNUSED(nb32); 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);
diff --git a/ggml/src/ggml-cuda/fattn-tile-f32.cu b/ggml/src/ggml-cuda/fattn-tile-f32.cu
index 32673adb5..124d5d3e8 100644
--- a/ggml/src/ggml-cuda/fattn-tile-f32.cu
+++ b/ggml/src/ggml-cuda/fattn-tile-f32.cu
@@ -6,7 +6,7 @@
 
 template<int D, int ncols, int nwarps, bool use_logit_softcap> // D == head size
 #if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
-__launch_bounds__(nwarps*WARP_SIZE, 1)
+__launch_bounds__(nwarps*WARP_SIZE, 2)
 #endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
 static __global__ void flash_attn_tile_ext_f32(
         const char * __restrict__ Q,
@@ -30,7 +30,9 @@ static __global__ void flash_attn_tile_ext_f32(
         const int ne12,
         const int ne13,
         const int ne31,
+        const int ne32,
         const int nb31,
+        const int nb32,
         const int nb01,
         const int nb02,
         const int nb03,
@@ -58,8 +60,8 @@ static __global__ void flash_attn_tile_ext_f32(
         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(ne12); GGML_UNUSED(ne13); GGML_UNUSED(ne31); GGML_UNUSED(ne32);
+        GGML_UNUSED(nb31); GGML_UNUSED(nb32); 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);
@@ -76,7 +78,7 @@ static __global__ void flash_attn_tile_ext_f32(
     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 half   * maskh = (const half   *) (mask + nb32*(blockIdx.z % ne32)      + nb31*ic0);
 
     const int stride_KV2 = nb11 / sizeof(half2);
 
diff --git a/ggml/src/ggml-cuda/fattn-vec-f16.cuh b/ggml/src/ggml-cuda/fattn-vec-f16.cuh
index 35e649cb3..e78fb1819 100644
--- a/ggml/src/ggml-cuda/fattn-vec-f16.cuh
+++ b/ggml/src/ggml-cuda/fattn-vec-f16.cuh
@@ -27,7 +27,9 @@ static __global__ void flash_attn_vec_ext_f16(
         const int ne12,
         const int ne13,
         const int ne31,
+        const int ne32,
         const int nb31,
+        const int nb32,
         const int nb01,
         const int nb02,
         const int nb03,
@@ -68,7 +70,7 @@ static __global__ void flash_attn_vec_ext_f16(
     K += nb12*(blockIdx.z / gqa_ratio);
     V += nb22*(blockIdx.z / gqa_ratio);
 
-    const half * maskh = (const half   *)  mask + ne11*ic0;
+    const half * maskh = (const half *) (mask + nb32*(blockIdx.z % ne32) + nb31*ic0);
 
     const float slopef = get_alibi_slope(max_bias, blockIdx.z, n_head_log2, m0, m1);
     const half  slopeh = __float2half(slopef);
@@ -342,8 +344,8 @@ static __global__ void flash_attn_vec_ext_f16(
     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(ne12); GGML_UNUSED(ne13); GGML_UNUSED(ne31); GGML_UNUSED(ne32);
+    GGML_UNUSED(nb31); GGML_UNUSED(nb32); 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);
diff --git a/ggml/src/ggml-cuda/fattn-vec-f32.cuh b/ggml/src/ggml-cuda/fattn-vec-f32.cuh
index 953967917..c22baf417 100644
--- a/ggml/src/ggml-cuda/fattn-vec-f32.cuh
+++ b/ggml/src/ggml-cuda/fattn-vec-f32.cuh
@@ -27,7 +27,9 @@ static __global__ void flash_attn_vec_ext_f32(
         const int ne12,
         const int ne13,
         const int ne31,
+        const int ne32,
         const int nb31,
+        const int nb32,
         const int nb01,
         const int nb02,
         const int nb03,
@@ -51,8 +53,8 @@ static __global__ void flash_attn_vec_ext_f32(
         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(ne12); GGML_UNUSED(ne13); GGML_UNUSED(ne31); GGML_UNUSED(ne32);
+        GGML_UNUSED(nb31); GGML_UNUSED(nb32); 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);
@@ -79,7 +81,8 @@ static __global__ void flash_attn_vec_ext_f32(
     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 half * maskh = (const half *) (mask + nb32*(blockIdx.z % ne32) + nb31*ic0);
 
     const float slope = get_alibi_slope(max_bias, blockIdx.z, n_head_log2, m0, m1);
 
diff --git a/ggml/src/ggml-cuda/fattn-wmma-f16.cu b/ggml/src/ggml-cuda/fattn-wmma-f16.cu
index c5668adb1..c95ca7b1f 100644
--- a/ggml/src/ggml-cuda/fattn-wmma-f16.cu
+++ b/ggml/src/ggml-cuda/fattn-wmma-f16.cu
@@ -9,7 +9,11 @@
 #ifdef FP16_MMA_AVAILABLE
 #if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
 #include <mma.h>
+#ifdef GGML_USE_MUSA
+namespace wmma = mtmusa::wmma;
+#else // GGML_USE_MUSA
 namespace wmma = nvcuda::wmma;
+#endif // GGML_USE_MUSA
 #elif defined(GGML_HIP_ROCWMMA_FATTN) && defined(FP16_MMA_AVAILABLE)
 #undef HIP_ENABLE_WARP_SYNC_BUILTINS // conflicts with rocWMMA headers
 #include <rocwmma/rocwmma.hpp>
@@ -42,7 +46,9 @@ static __global__ void flash_attn_ext_f16(
         const int ne12,
         const int ne13,
         const int ne31,
+        const int ne32,
         const int nb31,
+        const int nb32,
         const int nb01,
         const int nb02,
         const int nb03,
@@ -90,11 +96,11 @@ 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.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 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 + nb32*(blockIdx.z % ne32)      + nb31*ic0);
+    const half2 * mask2 = (const half2 *)  maskh;
 
     const int stride_Q  = nb01 / sizeof(float);
     const int stride_KV = nb11 / sizeof(half);
@@ -436,7 +442,7 @@ static __global__ void flash_attn_ext_f16(
     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(ne31); GGML_UNUSED(ne32); GGML_UNUSED(nb31); GGML_UNUSED(nb32); 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);
diff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cu
index 2a6f7f108..af5ad1ed5 100644
--- a/ggml/src/ggml-cuda/ggml-cuda.cu
+++ b/ggml/src/ggml-cuda/ggml-cuda.cu
@@ -11,6 +11,8 @@
 #include "ggml-cuda/clamp.cuh"
 #include "ggml-cuda/concat.cuh"
 #include "ggml-cuda/conv-transpose-1d.cuh"
+#include "ggml-cuda/conv2d-dw.cuh"
+#include "ggml-cuda/conv2d-transpose.cuh"
 #include "ggml-cuda/convert.cuh"
 #include "ggml-cuda/count-equal.cuh"
 #include "ggml-cuda/cpy.cuh"
@@ -35,6 +37,7 @@
 #include "ggml-cuda/ssm-scan.cuh"
 #include "ggml-cuda/sum.cuh"
 #include "ggml-cuda/sumrows.cuh"
+#include "ggml-cuda/mean.cuh"
 #include "ggml-cuda/tsembd.cuh"
 #include "ggml-cuda/unary.cuh"
 #include "ggml-cuda/upscale.cuh"
@@ -47,6 +50,7 @@
 #include <atomic>
 #include <charconv>
 #include <cinttypes>
+#include <condition_variable>
 #include <cstddef>
 #include <cstdint>
 #include <float.h>
@@ -54,9 +58,8 @@
 #include <map>
 #include <memory>
 #include <mutex>
-#include <stdint.h>
-#include <stdio.h>
 #include <stdarg.h>
+#include <stdio.h>
 #include <stdlib.h>
 #include <string>
 #include <vector>
@@ -97,8 +100,7 @@ int ggml_cuda_get_device() {
 static cudaError_t ggml_cuda_device_malloc(void ** ptr, size_t size, int device) {
     ggml_cuda_set_device(device);
     cudaError_t err;
-    if (getenv("GGML_CUDA_ENABLE_UNIFIED_MEMORY") != nullptr)
-    {
+    if (getenv("GGML_CUDA_ENABLE_UNIFIED_MEMORY") != nullptr) {
         err = cudaMallocManaged(ptr, size);
 #if defined(GGML_USE_HIP)
         if (err == hipSuccess) {
@@ -116,9 +118,7 @@ static cudaError_t ggml_cuda_device_malloc(void ** ptr, size_t size, int device)
             err = cudaMalloc(ptr, size);
         }
 #endif // defined(GGML_USE_HIP)
-    }
-    else
-    {
+    } else {
         err = cudaMalloc(ptr, size);
     }
     return err;
@@ -514,6 +514,33 @@ std::unique_ptr<ggml_cuda_pool> ggml_backend_cuda_context::new_pool_for_device(i
     return std::unique_ptr<ggml_cuda_pool>(new ggml_cuda_pool_leg(device));
 }
 
+// destroying a cuBLAS handle while a graph is being captured in a different thread can result in a CUDA error
+// this lock is used to ensure that no cuBLAS handle is destroyed while a graph is being captured
+
+static std::mutex ggml_cuda_lock;
+static std::condition_variable ggml_cuda_lock_cv;
+static std::atomic<int> ggml_cuda_lock_counter;
+
+ggml_backend_cuda_context::~ggml_backend_cuda_context() {
+    std::unique_lock<std::mutex> lock(ggml_cuda_lock);
+    ggml_cuda_lock_cv.wait(lock, []{ return ggml_cuda_lock_counter.load(std::memory_order_relaxed) == 0; });
+
+    if (copy_event != nullptr) {
+        CUDA_CHECK(cudaEventDestroy(copy_event));
+    }
+    for (int i = 0; i < GGML_CUDA_MAX_DEVICES; ++i) {
+        for (int j = 0; j < GGML_CUDA_MAX_STREAMS; ++j) {
+            if (streams[i][j] != nullptr) {
+                CUDA_CHECK(cudaStreamDestroy(streams[i][j]));
+            }
+        }
+        if (cublas_handles[i] != nullptr) {
+            CUBLAS_CHECK(cublasDestroy(cublas_handles[i]));
+        }
+    }
+}
+
+
 // cuda buffer
 
 struct ggml_backend_cuda_buffer_context {
@@ -615,9 +642,8 @@ static void ggml_backend_cuda_buffer_clear(ggml_backend_buffer_t buffer, uint8_t
     ggml_backend_cuda_buffer_context * ctx = (ggml_backend_cuda_buffer_context *)buffer->context;
 
     ggml_cuda_set_device(ctx->device);
-    CUDA_CHECK(cudaDeviceSynchronize());
-    CUDA_CHECK(cudaMemset(ctx->dev_ptr, value, buffer->size));
-    CUDA_CHECK(cudaDeviceSynchronize());
+    CUDA_CHECK(cudaMemsetAsync(ctx->dev_ptr, value, buffer->size, cudaStreamPerThread));
+    CUDA_CHECK(cudaStreamSynchronize(cudaStreamPerThread));
 }
 
 static const ggml_backend_buffer_i ggml_backend_cuda_buffer_interface = {
@@ -1144,7 +1170,6 @@ typedef void (*ggml_cuda_op_mul_mat_t)(
 static cudaError_t ggml_cuda_cpy_tensor_2d(
     void * dst, const struct ggml_tensor * src, int64_t i3, int64_t i2, int64_t i1_low, int64_t i1_high, cudaStream_t stream) {
 
-    GGML_ASSERT(ggml_backend_buffer_is_cuda(src->buffer));
     const char * src_ptr = (const char *) src->data;
     char       * dst_ptr = (char       *) dst;
 
@@ -1202,9 +1227,12 @@ static void ggml_cuda_op_mul_mat_cublas(
 
     const int cc = ggml_cuda_info().devices[id].cc;
 
+    const bool supports_bf16 = GGML_CUDA_CC_IS_NVIDIA(cc) || GGML_CUDA_CC_IS_AMD(cc) ||
+        (GGML_CUDA_CC_IS_MTHREADS(cc) && cc >= GGML_CUDA_CC_QY2);
+
     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 (src0->type == GGML_TYPE_BF16 && ggml_is_contiguous(src0) && row_diff == src0->ne[1]) {
+    if (supports_bf16 && src0->type == GGML_TYPE_BF16 && ggml_is_contiguous(src0) && row_diff == src0->ne[1]) {
         ggml_cuda_pool_alloc<nv_bfloat16> 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);
@@ -1232,7 +1260,7 @@ static void ggml_cuda_op_mul_mat_cublas(
 
         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) {
+    } else if (fast_fp16_hardware_available(cc) && use_fp16) {
         // convert src0 and src1 to fp16, multiply as fp16, convert dst to fp32
         ggml_cuda_pool_alloc<half> src0_as_f16(ctx.pool(id));
         if (src0->type != GGML_TYPE_F16) {
@@ -1427,8 +1455,6 @@ static void ggml_cuda_op_mul_mat(
     const int64_t nb2 = dst->nb[2];
     const int64_t nb3 = dst->nb[3];
 
-    GGML_ASSERT(ggml_backend_buffer_is_cuda(dst->buffer));
-    GGML_ASSERT(ggml_backend_buffer_is_cuda(src1->buffer));
     ggml_backend_cuda_buffer_context * src1_ctx = (ggml_backend_cuda_buffer_context *) src1->buffer->context;
     ggml_backend_cuda_buffer_context * dst_ctx  = (ggml_backend_cuda_buffer_context *) dst->buffer->context;
 
@@ -1723,7 +1749,7 @@ static void ggml_cuda_op_mul_mat(
 }
 
 static __global__ void k_compute_batched_ptrs(
-        const half * src0_as_f16, const half * src1_as_f16, char * dst,
+        const void * src0_as_f16, const void * src1_as_f16, char * dst,
         const void ** ptrs_src, void ** ptrs_dst,
         int64_t ne12, int64_t ne13,
         int64_t ne23,
@@ -1746,83 +1772,131 @@ static __global__ void k_compute_batched_ptrs(
     ptrs_dst[0*ne23 + i12 + i13*ne12] = (      char *)         dst + i12*nbd2 + i13*nbd3;
 }
 
-static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+// Type traits for mapping ggml types to CUDA/cuBLAS types
+template<ggml_type T>
+struct batched_mul_mat_traits;
+
+template<>
+struct batched_mul_mat_traits<GGML_TYPE_F32> {
+    using cuda_type = float;
+    static inline const cublasComputeType_t compute_type = CUBLAS_COMPUTE_32F;
+    static inline const cudaDataType_t data_type = CUDA_R_32F;
+    static inline const ggml_type ggml_type_val = GGML_TYPE_F32;
+    static inline const float alpha = 1.0f;
+    static inline const float beta = 0.0f;
+    static inline const void* get_alpha() { static const float val = alpha; return &val; }
+    static inline const void* get_beta() { static const float val = beta; return &val; }
+    static inline auto get_nc_converter(ggml_type src_type) { return ggml_get_to_fp32_nc_cuda(src_type); }
+};
+
+template<>
+struct batched_mul_mat_traits<GGML_TYPE_BF16> {
+    using cuda_type = nv_bfloat16;
+    static inline const cublasComputeType_t compute_type = CUBLAS_COMPUTE_32F;
+    static inline const cudaDataType_t data_type = CUDA_R_16BF;
+    static inline const ggml_type ggml_type_val = GGML_TYPE_BF16;
+    static inline const float alpha = 1.0f;
+    static inline const float beta = 0.0f;
+    static inline const void* get_alpha() { static const float val = alpha; return &val; }
+    static inline const void* get_beta() { static const float val = beta; return &val; }
+    static inline auto get_nc_converter(ggml_type src_type) { return ggml_get_to_bf16_nc_cuda(src_type); }
+};
+
+template<>
+struct batched_mul_mat_traits<GGML_TYPE_F16> {
+    using cuda_type = half;
+    static inline const cublasComputeType_t compute_type = CUBLAS_COMPUTE_16F;
+    static inline const cudaDataType_t data_type = CUDA_R_16F;
+    static inline const ggml_type ggml_type_val = GGML_TYPE_F16;
+    static inline const half alpha = 1.0;
+    static inline const half beta = 0.0;
+    static inline const void* get_alpha() { static const half val = alpha; return &val; }
+    static inline const void* get_beta() { static const half val = beta; return &val; }
+    static inline auto get_nc_converter(ggml_type src_type) { return ggml_get_to_fp16_nc_cuda(src_type); }
+};
+
+template<ggml_type src0_type>
+static void ggml_cuda_mul_mat_batched_cublas_impl(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    using traits = batched_mul_mat_traits<src0_type>;
+    using cuda_t = typename traits::cuda_type;
+
     GGML_ASSERT(!ggml_is_transposed(src0));
     GGML_ASSERT(!ggml_is_transposed(src1));
-
-    GGML_ASSERT(ggml_backend_buffer_is_cuda(src0->buffer));
-    GGML_ASSERT(src0->type == GGML_TYPE_F16);
+    GGML_ASSERT(!ggml_backend_buft_is_cuda_split(src0->buffer->buft));
+    GGML_ASSERT(src0->type == src0_type);
+    GGML_ASSERT(ggml_is_contiguous(dst));
 
     // Byte offsets and tensor dimensions are currently used in an inconsistent way for dst.
     // As long as dst is contiguous this does not matter though.
-    GGML_ASSERT(ggml_is_contiguous(dst));
 
     GGML_TENSOR_BINARY_OP_LOCALS
 
     const int64_t ne_dst = ggml_nelements(dst);
-
     cudaStream_t main_stream = ctx.stream();
-
     CUBLAS_CHECK(cublasSetStream(ctx.cublas_handle(), main_stream));
 
-    const half * src0_f16 = (const half *) src0->data;
     float * dst_ddf = (float *) dst->data;
-
-    const half * src1_f16 = (const half *) src1->data;
     const size_t ts_src1 = ggml_type_size(src1->type);
     GGML_ASSERT(nb10 == ts_src1);
     int64_t s11 = nb11 / ts_src1;
     int64_t s12 = nb12 / ts_src1;
     int64_t s13 = nb13 / ts_src1;
-    ggml_cuda_pool_alloc<half> src1_f16_alloc(ctx.pool());
 
-    // convert src1 to fp16
-    if (src1->type != GGML_TYPE_F16) {
-        const to_fp16_nc_cuda_t to_fp16_cuda = ggml_get_to_fp16_nc_cuda(src1->type);
+    const cuda_t * src0_ptr = nullptr;
+    const cuda_t * src1_ptr = nullptr;
+
+    ggml_cuda_pool_alloc<cuda_t> src0_alloc(ctx.pool());
+    ggml_cuda_pool_alloc<cuda_t> src1_alloc(ctx.pool());
+
+    // Handle src0
+    src0_ptr = (const cuda_t *) src0->data;
+
+    // Handle src1 - convert if necessary
+    if (src1->type == src0_type) {
+        src1_ptr = (const cuda_t *) src1->data;
+    } else {
+        // Convert src1 to target type using traits conversion functions
         const int64_t ne_src1 = ggml_nelements(src1);
-        src1_f16_alloc.alloc(ne_src1);
-        GGML_ASSERT(to_fp16_cuda != nullptr);
+        src1_alloc.alloc(ne_src1);
 
-        to_fp16_cuda(src1_f16, src1_f16_alloc.get(), ne10, ne11, ne12, ne13, s11, s12, s13, main_stream);
-
-        src1_f16 = src1_f16_alloc.get();
+        const auto convert_func = traits::get_nc_converter(src1->type);
+        GGML_ASSERT(convert_func != nullptr);
+        convert_func(src1->data, src1_alloc.get(), ne10, ne11, ne12, ne13, s11, s12, s13, main_stream);
+        src1_ptr = src1_alloc.get();
         s11 = ne10;
         s12 = ne11*s11;
         s13 = ne12*s12;
     }
 
-    ggml_cuda_pool_alloc<half> dst_f16(ctx.pool());
+    // Setup destination buffer
+    ggml_cuda_pool_alloc<cuda_t> dst_temp(ctx.pool());
     char * dst_t;
-
-    cublasComputeType_t cu_compute_type = CUBLAS_COMPUTE_16F;
-    cudaDataType_t      cu_data_type    = CUDA_R_16F;
-
-    // dst strides
     size_t nbd2 = dst->nb[2];
     size_t nbd3 = dst->nb[3];
 
-    const half  alpha_f16 = 1.0f;
-    const half  beta_f16  = 0.0f;
-
+    cublasComputeType_t cu_compute_type = traits::compute_type;
+    cudaDataType_t cu_data_type = traits::data_type;
+    cudaDataType_t cu_data_type_a = traits::data_type;
+    cudaDataType_t cu_data_type_b = traits::data_type;
+    const void * alpha = traits::get_alpha();
+    const void * beta = traits::get_beta();
     const float alpha_f32 = 1.0f;
-    const float beta_f32  = 0.0f;
-
-    const void * alpha = &alpha_f16;
-    const void * beta  = &beta_f16;
+    const float beta_f32 = 0.0f;
 
     if (dst->op_params[0] == GGML_PREC_DEFAULT) {
-        dst_t = (char *) dst_f16.alloc(ne_dst);
-
-        nbd2 /= sizeof(float) / sizeof(half);
-        nbd3 /= sizeof(float) / sizeof(half);
+        if constexpr (src0_type == GGML_TYPE_F32) {
+            dst_t = (char *) dst_ddf;  // Direct F32 output
+        } else {
+            dst_t = (char *) dst_temp.alloc(ne_dst);
+            nbd2 /= sizeof(float) / sizeof(cuda_t);
+            nbd3 /= sizeof(float) / sizeof(cuda_t);
+        }
     } else {
         dst_t = (char *) dst_ddf;
-
         cu_compute_type = CUBLAS_COMPUTE_32F;
-        cu_data_type    = CUDA_R_32F;
-
+        cu_data_type = CUDA_R_32F;
         alpha = &alpha_f32;
-        beta  = &beta_f32;
+        beta = &beta_f32;
     }
 
     int id = ggml_cuda_get_device();
@@ -1830,7 +1904,7 @@ static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, co
     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;
+        beta = &beta_f32;
     }
 
     GGML_ASSERT(ne12 % ne02 == 0);
@@ -1840,35 +1914,15 @@ static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, co
     const int64_t r2 = ne12/ne02;
     const int64_t r3 = ne13/ne03;
 
-#if 0
-    // use cublasGemmEx
-    {
-        for (int i13 = 0; i13 < ne13; ++i13) {
-            for (int i12 = 0; i12 < ne12; ++i12) {
-                int i03 = i13 / r3;
-                int i02 = i12 / r2;
-
-                CUBLAS_CHECK(
-                cublasGemmEx(ctx.cublas_handle(), CUBLAS_OP_T, CUBLAS_OP_N,
-                    ne01, ne11, ne10,
-                    alpha, (const char *) src0_f16 + i03*nb03 + i02*nb02, CUDA_R_16F,   nb01/sizeof(half),
-                                          src1_f16 + i13*s13  + i12*s12,  CUDA_R_16F,   s11,
-                    beta,  (      char *)    dst_t + i13*nbd3 + i12*nbd2, cu_data_type, ne0,
-                    cu_compute_type,
-                    CUBLAS_GEMM_DEFAULT_TENSOR_OP));
-            }
-        }
-    }
-#else
     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
         CUBLAS_CHECK(
         cublasGemmStridedBatchedEx(ctx.cublas_handle(), CUBLAS_OP_T, CUBLAS_OP_N,
                 ne01, ne11, ne10,
-                alpha, src0_f16, CUDA_R_16F,   nb01/nb00, nb02/nb00, // strideA
-                       src1_f16, CUDA_R_16F,   s11,       s12,       // strideB
-                beta,     dst_t, cu_data_type, ne0,       ne1*ne0,   // strideC
+                alpha, src0_ptr, cu_data_type_a, nb01/nb00, nb02/nb00, // strideA
+                       src1_ptr, cu_data_type_b, s11,       s12,       // strideB
+                beta,     dst_t, cu_data_type,   ne0,       ne1*ne0,   // strideC
                 ne12*ne13,
                 cu_compute_type,
                 CUBLAS_GEMM_DEFAULT_TENSOR_OP));
@@ -1879,34 +1933,55 @@ static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, co
         ggml_cuda_pool_alloc<const void *> ptrs_src(ctx.pool(), 2*ne23);
         ggml_cuda_pool_alloc<      void *> ptrs_dst(ctx.pool(), 1*ne23);
 
+        size_t src1_stride_size = sizeof(cuda_t);
+
         dim3 block_dims(ne13, ne12);
         k_compute_batched_ptrs<<<1, block_dims, 0, main_stream>>>(
-                src0_f16, src1_f16, dst_t,
+                src0_ptr, src1_ptr, dst_t,
                 ptrs_src.get(), ptrs_dst.get(),
                 ne12, ne13,
                 ne23,
                 nb02, nb03,
-                src1->type == GGML_TYPE_F16 ? nb12 : s12*sizeof(half),
-                src1->type == GGML_TYPE_F16 ? nb13 : s13*sizeof(half),
+                (src1->type == src0_type) ? nb12 : s12*src1_stride_size,
+                (src1->type == src0_type) ? nb13 : s13*src1_stride_size,
                 nbd2, nbd3,
                 r2, r3);
+
         CUDA_CHECK(cudaGetLastError());
 
         CUBLAS_CHECK(
         cublasGemmBatchedEx(ctx.cublas_handle(), CUBLAS_OP_T, CUBLAS_OP_N,
                 ne01, ne11, ne10,
-                alpha, (const void **) (ptrs_src.get() + 0*ne23), CUDA_R_16F,   nb01/nb00,
-                       (const void **) (ptrs_src.get() + 1*ne23), CUDA_R_16F,   s11,
-                beta,  (      void **) (ptrs_dst.get() + 0*ne23), cu_data_type, ne0,
+                alpha, (const void **) (ptrs_src.get() + 0*ne23), cu_data_type_a, nb01/nb00,
+                       (const void **) (ptrs_src.get() + 1*ne23), cu_data_type_b, s11,
+                beta,  (      void **) (ptrs_dst.get() + 0*ne23), cu_data_type,   ne0,
                 ne23,
                 cu_compute_type,
                 CUBLAS_GEMM_DEFAULT_TENSOR_OP));
     }
-#endif
 
-    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);
+    // Convert output back to F32 if needed
+    if (dst->op_params[0] == GGML_PREC_DEFAULT && cu_data_type != CUDA_R_32F) {
+        const to_fp32_cuda_t to_fp32_cuda = ggml_get_to_fp32_cuda(traits::ggml_type_val);
+        to_fp32_cuda(dst_temp.get(), dst_ddf, ne_dst, main_stream);
+    }
+}
+
+static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(src0->type == GGML_TYPE_F16 || src0->type == GGML_TYPE_BF16 || src0->type == GGML_TYPE_F32);
+
+    switch (src0->type) {
+        case GGML_TYPE_F32:
+            ggml_cuda_mul_mat_batched_cublas_impl<GGML_TYPE_F32>(ctx, src0, src1, dst);
+            break;
+        case GGML_TYPE_BF16:
+            ggml_cuda_mul_mat_batched_cublas_impl<GGML_TYPE_BF16>(ctx, src0, src1, dst);
+            break;
+        case GGML_TYPE_F16:
+            ggml_cuda_mul_mat_batched_cublas_impl<GGML_TYPE_F16>(ctx, src0, src1, dst);
+            break;
+        default:
+            GGML_ABORT("Unsupported type");
     }
 }
 
@@ -1920,16 +1995,14 @@ static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor
         && ggml_nbytes(src0) != ggml_backend_buffer_get_alloc_size(src0->buffer, src0) && src0->view_src;
 
     bool use_mul_mat_vec   = (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || src0->type == GGML_TYPE_BF16)
-        && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32
-        && src0->ne[0] % 2 == 0 && src1->ne[1] == 1;
+        && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32;
     bool use_mul_mat_vec_q = ggml_is_quantized(src0->type) && !bad_padding_clear
         && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32
         && src1->ne[1] <= MMVQ_MAX_BATCH_SIZE;
     bool use_mul_mat_q     = ggml_is_quantized(src0->type) && !bad_padding_clear
         && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32;
 
-    bool any_gpus_with_slow_fp16   = false;
-    bool any_gpus_without_fp16_mma = false;
+    bool any_gpus_with_slow_fp16 = false;
 
     if (split) {
         ggml_backend_cuda_split_buffer_type_context * buft_ctx = (ggml_backend_cuda_split_buffer_type_context *) src0->buffer->buft->context;
@@ -1940,16 +2013,16 @@ static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor
                 continue;
             }
 
-            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_hardware_available(cc);
-            any_gpus_without_fp16_mma = any_gpus_without_fp16_mma || !fp16_mma_hardware_available(cc);
+            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]);
+            use_mul_mat_vec         = use_mul_mat_vec           && ggml_cuda_should_use_mmv(src0->type, cc, src0->ne, src1->ne[1]);
+            any_gpus_with_slow_fp16 = any_gpus_with_slow_fp16   || !fast_fp16_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_hardware_available(cc);
-        any_gpus_without_fp16_mma = any_gpus_without_fp16_mma || !fp16_mma_hardware_available(cc);
+        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]);
+        use_mul_mat_vec         = use_mul_mat_vec           && ggml_cuda_should_use_mmv(src0->type, cc, src0->ne, src1->ne[1]);
+        any_gpus_with_slow_fp16 = any_gpus_with_slow_fp16   || !fast_fp16_hardware_available(cc);
     }
 
     // debug helpers
@@ -1960,7 +2033,13 @@ static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor
     //printf("src0 is contiguous %d, transposed %d, type = %s, name = %s\n", ggml_is_contiguous(src0), ggml_is_transposed(src0), ggml_type_name(src0->type), src0->name);
     //printf("src1 is contiguous %d, transposed %d, type = %s, name = %s\n", ggml_is_contiguous(src1), ggml_is_transposed(src1), ggml_type_name(src1->type), src1->name);
 
-    if (!split && use_mul_mat_vec && (src0->ne[1] <= MMV_MAX_ROWS || any_gpus_without_fp16_mma)) {
+    //TODO update for generic tensor parallelism
+    const int cc                     = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
+    bool use_batched_cublas_f16  = src0->type == GGML_TYPE_F16 && (src1->type == GGML_TYPE_F16 || !any_gpus_with_slow_fp16);
+    bool use_batched_cublas_bf16 = src0->type == GGML_TYPE_BF16 && bf16_mma_hardware_available(cc);
+    bool use_batched_cublas_f32  = src0->type == GGML_TYPE_F32;
+
+    if (!split && use_mul_mat_vec) {
         // the custom F16 vector kernel can be used over batched cuBLAS GEMM
         // but this is only faster for GPUs without tensor cores or with a thin src0 matrix (particularly KQV in attention)
         ggml_cuda_mul_mat_vec(ctx, src0, src1, nullptr, dst);
@@ -1968,8 +2047,8 @@ static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor
         ggml_cuda_mul_mat_vec_q(ctx, src0, src1, nullptr, dst);
     } else if (!split && use_mul_mat_q) {
         ggml_cuda_mul_mat_q(ctx, src0, src1, nullptr, dst);
-    } else if (!split && src0->type == GGML_TYPE_F16 && (src1->type == GGML_TYPE_F16 || !any_gpus_with_slow_fp16) &&
-            !ggml_is_transposed(src0) && !ggml_is_transposed(src1) && src1->ne[2]*src1->ne[3] > 1) {
+    } else if (!split && (use_batched_cublas_f16 || use_batched_cublas_bf16 || use_batched_cublas_f32)
+        && !ggml_is_transposed(src0) && !ggml_is_transposed(src1) && src1->ne[2]*src1->ne[3] > 1) {
         // general KQ + KQV multi-batch without FlashAttention
         ggml_cuda_mul_mat_batched_cublas(ctx, src0, src1, dst);
     } else if (use_mul_mat_vec) {
@@ -2224,6 +2303,27 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
                     return false;
             }
             break;
+        case GGML_OP_GLU:
+            switch (ggml_get_glu_op(dst)) {
+                case GGML_GLU_OP_REGLU:
+                    ggml_cuda_op_reglu(ctx, dst);
+                    break;
+                case GGML_GLU_OP_GEGLU:
+                    ggml_cuda_op_geglu(ctx, dst);
+                    break;
+                case GGML_GLU_OP_SWIGLU:
+                    ggml_cuda_op_swiglu(ctx, dst);
+                    break;
+                case GGML_GLU_OP_GEGLU_ERF:
+                    ggml_cuda_op_geglu_erf(ctx, dst);
+                    break;
+                case GGML_GLU_OP_GEGLU_QUICK:
+                    ggml_cuda_op_geglu_quick(ctx, dst);
+                    break;
+                default:
+                    return false;
+            }
+            break;
         case GGML_OP_NORM:
             ggml_cuda_op_norm(ctx, dst);
             break;
@@ -2314,6 +2414,12 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
         case GGML_OP_IM2COL:
             ggml_cuda_op_im2col(ctx, dst);
             break;
+        case GGML_OP_CONV_2D_DW:
+            ggml_cuda_op_conv2d_dw(ctx, dst);
+            break;
+        case GGML_OP_CONV_TRANSPOSE_2D:
+            ggml_cuda_conv_2d_transpose_p0(ctx, dst);
+            break;
         case GGML_OP_CONV_TRANSPOSE_1D:
             ggml_cuda_op_conv_transpose_1d(ctx,dst);
             break;
@@ -2326,6 +2432,9 @@ 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_MEAN:
+            ggml_cuda_op_mean(ctx, dst);
+            break;
         case GGML_OP_SSM_CONV:
             ggml_cuda_op_ssm_conv(ctx, dst);
             break;
@@ -2668,7 +2777,9 @@ static void evaluate_and_capture_cuda_graph(ggml_backend_cuda_context * cuda_ctx
                                ggml_backend_buft_is_cuda_split(node->src[j]->buffer->buft) || (integrated && ggml_backend_buft_is_cuda_host(node->src[j]->buffer->buft)));
                     }
                 }
-#endif
+#else
+                GGML_UNUSED(integrated);
+#endif // NDEBUG
 
                 bool ok = ggml_cuda_compute_forward(*cuda_ctx, node);
                 if (!ok) {
@@ -2687,6 +2798,11 @@ static void evaluate_and_capture_cuda_graph(ggml_backend_cuda_context * cuda_ctx
 
             CUDA_CHECK(cudaStreamEndCapture(cuda_ctx->stream(), &cuda_ctx->cuda_graph->graph));
             graph_evaluated_or_captured = true; // CUDA graph has been captured
+
+            std::lock_guard<std::mutex> lock(ggml_cuda_lock);
+            if (ggml_cuda_lock_counter.fetch_sub(1, std::memory_order_relaxed) == 1) {
+                ggml_cuda_lock_cv.notify_all();
+            }
         } else {
             graph_evaluated_or_captured = true; // ggml graph has been directly evaluated
         }
@@ -2762,7 +2878,13 @@ static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t backend,
         }
     }
 
-    if (use_cuda_graph && cuda_graph_update_required) { // Start CUDA graph capture
+    if (use_cuda_graph && cuda_graph_update_required) {
+        // Start CUDA graph capture
+        {
+            std::lock_guard<std::mutex> lock(ggml_cuda_lock);
+            ggml_cuda_lock_counter.fetch_add(1, std::memory_order_relaxed);
+        }
+
         CUDA_CHECK(cudaStreamBeginCapture(cuda_ctx->stream(), cudaStreamCaptureModeRelaxed));
     }
 
@@ -2995,6 +3117,18 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
                     return false;
             }
             break;
+        case GGML_OP_GLU:
+            switch (ggml_get_glu_op(op)) {
+                case GGML_GLU_OP_REGLU:
+                case GGML_GLU_OP_GEGLU:
+                case GGML_GLU_OP_SWIGLU:
+                case GGML_GLU_OP_GEGLU_ERF:
+                case GGML_GLU_OP_GEGLU_QUICK:
+                    return ggml_is_contiguous_1(op->src[0]);
+                default:
+                    return false;
+            }
+            break;
         case GGML_OP_MUL_MAT:
         case GGML_OP_MUL_MAT_ID:
             {
@@ -3018,9 +3152,16 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
                     return false;
                 }
 #ifdef GGML_USE_MUSA
-                if (b->type == GGML_TYPE_F16 && b->ne[2]*b->ne[3] > 1 &&
-                    !ggml_is_transposed(a) && !ggml_is_transposed(b)) {
-                    return false;
+                const int cc = ggml_cuda_info().devices[dev_ctx->device].cc;
+                if (b->ne[2]*b->ne[3] > 1 && !ggml_is_transposed(a) && !ggml_is_transposed(b)) {
+                    if (GGML_CUDA_CC_IS_QY1(cc) && op->op == GGML_OP_MUL_MAT &&
+                            a->type == GGML_TYPE_F16 && b->type == GGML_TYPE_F16) {
+                        return false;
+                    }
+                    if (GGML_CUDA_CC_IS_QY2(cc) && op->op == GGML_OP_MUL_MAT_ID &&
+                            a->type == GGML_TYPE_Q2_K && b->type == GGML_TYPE_F32) {
+                        return false;
+                    }
                 }
 #endif // GGML_USE_MUSA
                 switch (a->type) {
@@ -3047,11 +3188,6 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
                     case GGML_TYPE_IQ4_NL:
                     case GGML_TYPE_IQ4_XS:
                     case GGML_TYPE_BF16:
-#ifdef GGML_USE_MUSA
-                        if (a->type == GGML_TYPE_Q3_K) {
-                            return false;
-                        }
-#endif // GGML_USE_MUSA
                         return true;
                     default:
                         return false;
@@ -3193,12 +3329,26 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
         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_SSM_SCAN: {
+            if (op->src[3]->ne[0] == 1) {
+                // Mamba2
+                // (kernel only supports d_state == 128 && d_head % 16 == 0)
+                return op->src[0]->ne[0] == 128 && op->src[0]->ne[1] % 16 == 0;
+            } else {
+                // Mamba
+                // (kernel only supports d_state == 16, d_head == 1, n_head % 128 == 0, n_group == 1)
+                return op->src[0]->ne[0] == 16 && op->src[0]->ne[1] == 1 && op->src[0]->ne[2] % 128 == 0 && op->src[4]->ne[1] == 1;
+            }
+        }
+        case GGML_OP_SSM_CONV: {
+            // assumes d_inner % threads == 0
+            return op->src[0]->ne[1] % 128 == 0;
+        }
         case GGML_OP_CONT:
             return op->src[0]->type != GGML_TYPE_BF16;
         case GGML_OP_DIAG_MASK_INF:
+            return true;
         case GGML_OP_SOFT_MAX:
             return true;
         case GGML_OP_SOFT_MAX_BACK: {
@@ -3211,9 +3361,12 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
             return op->src[0]->nb[0] == ggml_type_size(op->src[0]->type) && ggml_is_contiguous_2(op->src[0]);
         }
         case GGML_OP_IM2COL:
+        case GGML_OP_CONV_2D_DW:
+        case GGML_OP_CONV_TRANSPOSE_2D:
         case GGML_OP_POOL_2D:
         case GGML_OP_SUM:
         case GGML_OP_SUM_ROWS:
+        case GGML_OP_MEAN:
         case GGML_OP_ARGSORT:
         case GGML_OP_ACC:
             return true;
@@ -3244,6 +3397,9 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
             if (op->src[0]->ne[0] == 192) {
                 return false;
             }
+            // TODO: support broadcast
+            // note: this was initially implemented in https://github.com/ggml-org/llama.cpp/pull/14500, but
+            //       the interface of ggml_flash_attn_ext() changed in https://github.com/ggml-org/llama.cpp/pull/14505
             if (op->src[0]->ne[3] != 1) {
                 return false;
             }
diff --git a/ggml/src/ggml-cuda/mean.cu b/ggml/src/ggml-cuda/mean.cu
new file mode 100644
index 000000000..4b238a399
--- /dev/null
+++ b/ggml/src/ggml-cuda/mean.cu
@@ -0,0 +1,19 @@
+#include "mean.cuh"
+
+void ggml_cuda_op_mean(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_ASSERT(ggml_is_contiguous(src0));
+
+    const int64_t ncols = src0->ne[0];
+    const int64_t nrows = ggml_nrows(src0);
+
+    const dim3 block_dims(WARP_SIZE, 1, 1);
+    const dim3 block_nums(nrows, 1, 1);
+    reduce_rows_f32</*norm*/ true><<<block_nums, block_dims, 0, stream>>>(src0_d, dst_d, ncols);
+}
diff --git a/ggml/src/ggml-cuda/mean.cuh b/ggml/src/ggml-cuda/mean.cuh
new file mode 100644
index 000000000..2b9b10433
--- /dev/null
+++ b/ggml/src/ggml-cuda/mean.cuh
@@ -0,0 +1,3 @@
+#include "common.cuh"
+
+void ggml_cuda_op_mean(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
diff --git a/ggml/src/ggml-cuda/mmq.cuh b/ggml/src/ggml-cuda/mmq.cuh
index 80baf459c..9696a3204 100644
--- a/ggml/src/ggml-cuda/mmq.cuh
+++ b/ggml/src/ggml-cuda/mmq.cuh
@@ -3016,14 +3016,8 @@ static void launch_mul_mat_q(ggml_backend_cuda_context & ctx, const mmq_args & a
 
     const int nbytes_shared = mmq_get_nbytes_shared<type>(mmq_x, mmq_y, cc);
 
-#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(mul_mat_q<type, mmq_x, MMQ_NWARPS, false>, cudaFuncAttributeMaxDynamicSharedMemorySize, nbytes_shared));
-        CUDA_CHECK(cudaFuncSetAttribute(mul_mat_q<type, mmq_x, MMQ_NWARPS, true>,  cudaFuncAttributeMaxDynamicSharedMemorySize, nbytes_shared));
-        shared_memory_limit_raised[id] = true;
-    }
-#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && !defined(GGML_USE_MUSA)
+    CUDA_SET_SHARED_MEMORY_LIMIT((mul_mat_q<type, mmq_x, MMQ_NWARPS, false>), nbytes_shared);
+    CUDA_SET_SHARED_MEMORY_LIMIT((mul_mat_q<type, mmq_x, MMQ_NWARPS, true>),  nbytes_shared);
 
     const int nty  = (args.nrows_x   + mmq_y - 1) / mmq_y;
     const int ntx  = (args.ncols_dst + mmq_x - 1) / mmq_x;
diff --git a/ggml/src/ggml-cuda/mmv.cu b/ggml/src/ggml-cuda/mmv.cu
index d8c385e23..e14c93516 100644
--- a/ggml/src/ggml-cuda/mmv.cu
+++ b/ggml/src/ggml-cuda/mmv.cu
@@ -2,25 +2,26 @@
 #include "common.cuh"
 #include "mmv.cuh"
 
-template <typename T, typename type_acc, int block_size>
+template <typename T, typename type_acc, int ncols_dst, int block_size>
 static __global__ void mul_mat_vec(
         const T * __restrict__ x, const float * __restrict__ y, const int32_t * __restrict__ ids, float * __restrict__ dst,
-        const int64_t ncols2, const int64_t nchannels_y, const int64_t stride_row,
-        const int64_t channel_ratio, const int64_t stride_channel_x, const int64_t stride_channel_y, const int64_t stride_channel_dst,
-        const int64_t sample_ratio, const int64_t stride_sample_x, const int64_t stride_sample_y, const int64_t stride_sample_dst) {
-    const int64_t row         = blockIdx.x;
-    const int64_t channel_dst = blockIdx.y;
-    const int64_t channel_x   = ids ? ids[channel_dst]          : channel_dst / channel_ratio;
-    const int64_t channel_y   = ids ? channel_dst % nchannels_y : channel_dst;
-    const int64_t sample_dst  = blockIdx.z;
-    const int64_t sample_x    = sample_dst / sample_ratio;
-    const int64_t sample_y    = sample_dst;
-    const int     tid         = threadIdx.x;
+        const int ncols2, const int nchannels_y, const int stride_row, const int stride_col_y2, const int stride_col_dst,
+        const int channel_ratio, const int stride_channel_x, const int stride_channel_y, const int stride_channel_dst,
+        const int sample_ratio, const int stride_sample_x, const int stride_sample_y, const int stride_sample_dst) {
+    const int row         = blockIdx.x;
+    const int channel_dst = blockIdx.y;
+    const int channel_x   = ids ? ids[channel_dst]          : channel_dst / channel_ratio;
+    const int channel_y   = ids ? channel_dst % nchannels_y : channel_dst;
+    const int sample_dst  = blockIdx.z;
+    const int sample_x    = sample_dst / sample_ratio;
+    const int sample_y    = sample_dst;
+    const int tid         = threadIdx.x;
+
     constexpr int warp_size   = ggml_cuda_get_physical_warp_size();
 
-    x   += sample_x  *stride_sample_x   + channel_x  *stride_channel_x   + row*stride_row;
-    y   += sample_y  *stride_sample_y   + channel_y  *stride_channel_y;
-    dst += sample_dst*stride_sample_dst + channel_dst*stride_channel_dst;
+    x   += int64_t(sample_x)  *stride_sample_x   + channel_x  *stride_channel_x   + row*stride_row;
+    y   += int64_t(sample_y)  *stride_sample_y   + channel_y  *stride_channel_y;
+    dst += int64_t(sample_dst)*stride_sample_dst + channel_dst*stride_channel_dst;
 
     const float2 * y2 = (const float2 *) y;
 
@@ -34,81 +35,108 @@ static __global__ void mul_mat_vec(
         __syncthreads();
     }
 
-    float sumf = 0.0f;
+    float sumf[ncols_dst] = {0.0f};
 
     if constexpr (std::is_same<T, float>::value) {
         const float2 * x2 = (const float2 *) x;
 
-        for (int64_t col2 = tid; col2 < ncols2; col2 += block_size) {
+        for (int col2 = tid; col2 < ncols2; col2 += block_size) {
             const float2 tmpx = x2[col2];
-            const float2 tmpy = y2[col2];
-            sumf += tmpx.x*tmpy.x;
-            sumf += tmpx.y*tmpy.y;
+
+#pragma unroll
+            for (int j = 0; j < ncols_dst; ++j) {
+                const float2 tmpy = y2[j*stride_col_y2 + col2];
+                sumf[j] += tmpx.x*tmpy.x;
+                sumf[j] += tmpx.y*tmpy.y;
+            }
         }
     } else if constexpr (std::is_same<T, half>::value) {
         const half2 * x2 = (const half2 *) x;
 
         if (std::is_same<type_acc, float>::value) {
-            for (int64_t col2 = tid; col2 < ncols2; col2 += block_size) {
+            for (int col2 = tid; col2 < ncols2; col2 += block_size) {
                 const float2 tmpx = __half22float2(x2[col2]);
-                const float2 tmpy = y2[col2];
-                sumf += tmpx.x * tmpy.x;
-                sumf += tmpx.y * tmpy.y;
+
+#pragma unroll
+                for (int j = 0; j < ncols_dst; ++j) {
+                    const float2 tmpy = y2[j*stride_col_y2 + col2];
+                    sumf[j] += tmpx.x * tmpy.x;
+                    sumf[j] += tmpx.y * tmpy.y;
+                }
             }
         } else {
 #ifdef FP16_AVAILABLE
-            half2 sumh2 = make_half2(0.0f, 0.0f);
+            half2 sumh2[ncols_dst] = {{0.0f, 0.0f}};
 
-            for (int64_t col2 = tid; col2 < ncols2; col2 += block_size) {
-                const float2 tmp = y2[col2];
-                sumh2 += x2[col2] * make_half2(tmp.x, tmp.y);
+            for (int col2 = tid; col2 < ncols2; col2 += block_size) {
+                const half2 tmpx = x2[col2];
+
+#pragma unroll
+                for (int j = 0; j < ncols_dst; ++j) {
+                    const float2 tmpy = y2[j*stride_col_y2 + col2];
+                    sumh2[j] += tmpx * make_half2(tmpy.x, tmpy.y);
+                }
             }
 
-            sumf = __low2float(sumh2) + __high2float(sumh2);
+#pragma unroll
+            for (int j = 0; j < ncols_dst; ++j) {
+                sumf[j] = __low2float(sumh2[j]) + __high2float(sumh2[j]);
+            }
 #else
             NO_DEVICE_CODE;
 #endif // FP16_AVAILABLE
         }
     } else if constexpr (std::is_same<T, nv_bfloat16>::value) {
         const int * x2 = (const int *) x;
-        for (int64_t col2 = tid; col2 < ncols2; col2 += block_size) {
-            const int    tmpx = x2[col2];
-            const float2 tmpy = y2[col2];
-            sumf += float(reinterpret_cast<const nv_bfloat16 *>(&tmpx)[0]) * tmpy.x;
-            sumf += float(reinterpret_cast<const nv_bfloat16 *>(&tmpx)[1]) * tmpy.y;
+        for (int col2 = tid; col2 < ncols2; col2 += block_size) {
+            const int tmpx = x2[col2];
+#pragma unroll
+            for (int j = 0; j < ncols_dst; ++j) {
+                const float2 tmpy = y2[j*stride_col_y2 + col2];
+                sumf[j] += float(reinterpret_cast<const nv_bfloat16 *>(&tmpx)[0]) * tmpy.x;
+                sumf[j] += float(reinterpret_cast<const nv_bfloat16 *>(&tmpx)[1]) * tmpy.y;
+            }
         }
     } else {
         static_assert(std::is_same<T, void>::value, "unsupported type");
     }
 
-    sumf = warp_reduce_sum<warp_size>(sumf);
+#pragma unroll
+    for (int j = 0; j < ncols_dst; ++j) {
+        sumf[j] = warp_reduce_sum<warp_size>(sumf[j]);
 
-    if (block_size > warp_size) {
-        buf_iw[tid/warp_size] = sumf;
-        __syncthreads();
-        if (tid >= warp_size) {
-            return;
+        if (block_size > warp_size) {
+            buf_iw[tid/warp_size] = sumf[j];
+            __syncthreads();
+            if (tid < warp_size) {
+                sumf[j] = buf_iw[tid];
+                sumf[j] = warp_reduce_sum<warp_size>(sumf[j]);
+            }
+            if (j < ncols_dst) {
+                __syncthreads();
+            }
         }
-        sumf = buf_iw[tid];
-        sumf = warp_reduce_sum<warp_size>(sumf);
     }
 
-    if (tid != 0) {
+    if (tid >= ncols_dst) {
         return;
     }
 
-    dst[row] = sumf;
+    dst[tid*stride_col_dst + row] = sumf[tid];
 }
 
-template <typename T, typename type_acc>
+template <typename T, typename type_acc, int ncols_dst>
 static void launch_mul_mat_vec_cuda(
         const T * x, const float * y, const int32_t * ids, float * dst,
-        const int64_t ncols, const int64_t nrows, const int64_t stride_row, const int64_t nchannels_x, const int64_t nchannels_y, const int64_t nchannels_dst,
+        const int64_t ncols, const int64_t nrows,
+        const int64_t stride_row, const int64_t stride_col_y, const int64_t stride_col_dst,
+        const int64_t nchannels_x, const int64_t nchannels_y, const int64_t nchannels_dst,
         const int64_t stride_channel_x, const int64_t stride_channel_y, const int64_t stride_channel_dst, const int64_t nsamples_x,
         const int64_t nsamples_dst, const int64_t stride_sample_x, const int64_t stride_sample_y, const int64_t stride_sample_dst,
         cudaStream_t stream) {
-    GGML_ASSERT(ncols      % 2 == 0);
-    GGML_ASSERT(stride_row % 2 == 0);
+    GGML_ASSERT(ncols        % 2 == 0);
+    GGML_ASSERT(stride_row   % 2 == 0);
+    GGML_ASSERT(stride_col_y % 2 == 0);
     GGML_ASSERT(ids || nchannels_dst % nchannels_x == 0);
     GGML_ASSERT(       nsamples_dst  % nsamples_x  == 0);
     const int64_t channel_ratio = nchannels_dst / nchannels_x;
@@ -138,44 +166,52 @@ static void launch_mul_mat_vec_cuda(
     const dim3 block_dims(block_size_best, 1, 1);
     switch (block_size_best) {
         case   32: {
-            mul_mat_vec<T, type_acc,  32><<<block_nums, block_dims, smem, stream>>>
-                (x, y, ids, dst, ncols/2, nchannels_y, stride_row, channel_ratio, stride_channel_x, stride_channel_y,
-                 stride_channel_dst, sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
+            mul_mat_vec<T, type_acc, ncols_dst,  32><<<block_nums, block_dims, smem, stream>>>
+                (x, y, ids, dst, ncols/2, nchannels_y, stride_row, stride_col_y/2, stride_col_dst,
+                 channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
         } break;
         case   64: {
-            mul_mat_vec<T, type_acc,  64><<<block_nums, block_dims, smem, stream>>>
-                (x, y, ids, dst, ncols/2, nchannels_y, stride_row, channel_ratio, stride_channel_x, stride_channel_y,
-                 stride_channel_dst, sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
+            mul_mat_vec<T, type_acc, ncols_dst,  64><<<block_nums, block_dims, smem, stream>>>
+                (x, y, ids, dst, ncols/2, nchannels_y, stride_row, stride_col_y/2, stride_col_dst,
+                 channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
         } break;
         case   96: {
-            mul_mat_vec<T, type_acc,  96><<<block_nums, block_dims, smem, stream>>>
-                (x, y, ids, dst, ncols/2, nchannels_y, stride_row, channel_ratio, stride_channel_x, stride_channel_y,
-                 stride_channel_dst, sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
+            mul_mat_vec<T, type_acc, ncols_dst,  96><<<block_nums, block_dims, smem, stream>>>
+                (x, y, ids, dst, ncols/2, nchannels_y, stride_row, stride_col_y/2, stride_col_dst,
+                 channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
         } break;
         case  128: {
-            mul_mat_vec<T, type_acc, 128><<<block_nums, block_dims, smem, stream>>>
-                (x, y, ids, dst, ncols/2, nchannels_y, stride_row, channel_ratio, stride_channel_x, stride_channel_y,
-                 stride_channel_dst, sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
+            mul_mat_vec<T, type_acc, ncols_dst, 128><<<block_nums, block_dims, smem, stream>>>
+                (x, y, ids, dst, ncols/2, nchannels_y, stride_row, stride_col_y/2, stride_col_dst,
+                 channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
         } break;
         case  160: {
-            mul_mat_vec<T, type_acc, 160><<<block_nums, block_dims, smem, stream>>>
-                (x, y, ids, dst, ncols/2, nchannels_y, stride_row, channel_ratio, stride_channel_x, stride_channel_y,
-                 stride_channel_dst, sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
+            mul_mat_vec<T, type_acc, ncols_dst, 160><<<block_nums, block_dims, smem, stream>>>
+                (x, y, ids, dst, ncols/2, nchannels_y, stride_row, stride_col_y/2, stride_col_dst,
+                 channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
         } break;
         case  192: {
-            mul_mat_vec<T, type_acc, 192><<<block_nums, block_dims, smem, stream>>>
-                (x, y, ids, dst, ncols/2, nchannels_y, stride_row, channel_ratio, stride_channel_x, stride_channel_y,
-                 stride_channel_dst, sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
+            mul_mat_vec<T, type_acc, ncols_dst, 192><<<block_nums, block_dims, smem, stream>>>
+                (x, y, ids, dst, ncols/2, nchannels_y, stride_row, stride_col_y/2, stride_col_dst,
+                 channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
         } break;
         case  224: {
-            mul_mat_vec<T, type_acc, 224><<<block_nums, block_dims, smem, stream>>>
-                (x, y, ids, dst, ncols/2, nchannels_y, stride_row, channel_ratio, stride_channel_x, stride_channel_y,
-                 stride_channel_dst, sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
+            mul_mat_vec<T, type_acc, ncols_dst, 224><<<block_nums, block_dims, smem, stream>>>
+                (x, y, ids, dst, ncols/2, nchannels_y, stride_row, stride_col_y/2, stride_col_dst,
+                 channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
         } break;
         case  256: {
-            mul_mat_vec<T, type_acc, 256><<<block_nums, block_dims, smem, stream>>>
-                (x, y, ids, dst, ncols/2, nchannels_y, stride_row, channel_ratio, stride_channel_x, stride_channel_y,
-                 stride_channel_dst, sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
+            mul_mat_vec<T, type_acc, ncols_dst, 256><<<block_nums, block_dims, smem, stream>>>
+                (x, y, ids, dst, ncols/2, nchannels_y, stride_row, stride_col_y/2, stride_col_dst,
+                 channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
         } break;
         default: {
             GGML_ABORT("fatal error");
@@ -183,23 +219,91 @@ static void launch_mul_mat_vec_cuda(
     }
 }
 
+template <typename T, typename type_acc>
+static void mul_mat_vec_cuda_switch_ncols_dst(
+        const T * x, const float * y, const int32_t * ids, float * dst,
+        const int64_t ncols, const int64_t nrows, const int64_t ncols_dst,
+        const int64_t stride_row, const int64_t stride_col_y, const int64_t stride_col_dst,
+        const int64_t nchannels_x, const int64_t nchannels_y, const int64_t nchannels_dst,
+        const int64_t stride_channel_x, const int64_t stride_channel_y, const int64_t stride_channel_dst, const int64_t nsamples_x,
+        const int64_t nsamples_dst, const int64_t stride_sample_x, const int64_t stride_sample_y, const int64_t stride_sample_dst,
+        cudaStream_t stream) {
+    switch (ncols_dst) {
+        case 1:
+            launch_mul_mat_vec_cuda<T, type_acc, 1>
+                (x, y, ids, dst, ncols, nrows, stride_row, stride_col_y, stride_col_dst,
+                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y,
+                 stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+            break;
+        case 2:
+            launch_mul_mat_vec_cuda<T, type_acc, 2>
+                (x, y, ids, dst, ncols, nrows, stride_row, stride_col_y, stride_col_dst,
+                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y,
+                 stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+            break;
+        case 3:
+            launch_mul_mat_vec_cuda<T, type_acc, 3>
+                (x, y, ids, dst, ncols, nrows, stride_row, stride_col_y, stride_col_dst,
+                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y,
+                 stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+            break;
+        case 4:
+            launch_mul_mat_vec_cuda<T, type_acc, 4>
+                (x, y, ids, dst, ncols, nrows, stride_row, stride_col_y, stride_col_dst,
+                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y,
+                 stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+            break;
+        case 5:
+            launch_mul_mat_vec_cuda<T, type_acc, 5>
+                (x, y, ids, dst, ncols, nrows, stride_row, stride_col_y, stride_col_dst,
+                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y,
+                 stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+            break;
+        case 6:
+            launch_mul_mat_vec_cuda<T, type_acc, 6>
+                (x, y, ids, dst, ncols, nrows, stride_row, stride_col_y, stride_col_dst,
+                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y,
+                 stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+            break;
+        case 7:
+            launch_mul_mat_vec_cuda<T, type_acc, 7>
+                (x, y, ids, dst, ncols, nrows, stride_row, stride_col_y, stride_col_dst,
+                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y,
+                 stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+            break;
+        case 8:
+            launch_mul_mat_vec_cuda<T, type_acc, 8>
+                (x, y, ids, dst, ncols, nrows, stride_row, stride_col_y, stride_col_dst,
+                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y,
+                 stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+            break;
+        default:
+            GGML_ABORT("fatal error");
+            break;
+    }
+}
+
 template<typename T>
 static void mul_mat_vec_cuda(
         const T * x, const float * y, const int32_t * ids, float * dst,
-        const int64_t ncols, const int64_t nrows, const int64_t stride_row, const int64_t nchannels_x, const int64_t nchannels_y, const int64_t nchannels_dst,
+        const int64_t ncols, const int64_t nrows, const int64_t ncols_dst,
+        const int64_t stride_row, const int64_t stride_col_y, const int stride_col_dst,
+        const int64_t nchannels_x, const int64_t nchannels_y, const int64_t nchannels_dst,
         const int64_t stride_channel_x, const int64_t stride_channel_y, const int64_t stride_channel_dst, const int64_t nsamples_x,
         const int64_t nsamples_dst, const int64_t stride_sample_x, const int64_t stride_sample_y, const int64_t stride_sample_dst,
         enum ggml_prec prec, cudaStream_t stream) {
     if constexpr(std::is_same<T, half>::value) {
         if (prec == GGML_PREC_DEFAULT) {
-            launch_mul_mat_vec_cuda<T, half>
-                (x, y, ids, dst, ncols, nrows, stride_row, nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y,
+            mul_mat_vec_cuda_switch_ncols_dst<T, half>
+                (x, y, ids, dst, ncols, nrows, ncols_dst, stride_row, stride_col_y, stride_col_dst,
+                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y,
                  stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
             return;
         }
     }
-    launch_mul_mat_vec_cuda<T, float>
-        (x, y, ids, dst, ncols, nrows, stride_row, nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y,
+    mul_mat_vec_cuda_switch_ncols_dst<T, float>
+        (x, y, ids, dst, ncols, nrows, ncols_dst, stride_row, stride_col_y, stride_col_dst,
+         nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y,
          stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
 }
 
@@ -246,24 +350,24 @@ void ggml_cuda_mul_mat_vec(ggml_backend_cuda_context & ctx, const ggml_tensor *
     const int64_t stride_channel_dst = ids ? s1   : s2;
     const int64_t stride_channel_y   = ids ? s11  : s12;
 
-    GGML_ASSERT(ncols_dst == 1);
+    GGML_ASSERT(!ids || ncols_dst == 1);
 
     switch (src0->type) {
         case GGML_TYPE_F32: {
             const float * src0_d = (const float *) src0->data;
-            mul_mat_vec_cuda(src0_d, src1_d, ids_d, dst_d, ne00, ne01, s01,
+            mul_mat_vec_cuda(src0_d, src1_d, ids_d, dst_d, ne00, ne01, ncols_dst, s01, s11, s1,
                 ne02, nchannels_y, nchannels_dst, s02, stride_channel_y, stride_channel_dst,
                 ne03,              ne3,           s03, s13,              s3,                 prec, ctx.stream());
         } break;
         case GGML_TYPE_F16: {
             const half * src0_d = (const half *) src0->data;
-            mul_mat_vec_cuda(src0_d, src1_d, ids_d, dst_d, ne00, ne01, s01,
+            mul_mat_vec_cuda(src0_d, src1_d, ids_d, dst_d, ne00, ne01, ncols_dst, s01, s11, s1,
                 ne02, nchannels_y, nchannels_dst, s02, stride_channel_y, stride_channel_dst,
                 ne03,              ne3,           s03, s13,              s3,                 prec, ctx.stream());
         } break;
         case GGML_TYPE_BF16: {
             const nv_bfloat16 * src0_d = (const nv_bfloat16 *) src0->data;
-            mul_mat_vec_cuda(src0_d, src1_d, ids_d, dst_d, ne00, ne01, s01,
+            mul_mat_vec_cuda(src0_d, src1_d, ids_d, dst_d, ne00, ne01, ncols_dst, s01, s11, s1,
                 ne02, nchannels_y, nchannels_dst, s02, stride_channel_y, stride_channel_dst,
                 ne03,              ne3,           s03, s13,              s3,                 prec, ctx.stream());
         } break;
@@ -282,16 +386,19 @@ void ggml_cuda_op_mul_mat_vec(
     GGML_ASSERT(dst->type  == GGML_TYPE_F32);
 
     const int64_t ne00 = src0->ne[0];
+    const int64_t ne10 = src1->ne[0];
+    const int64_t ne0  =  dst->ne[0];
     const int64_t row_diff = row_high - row_low;
 
-    GGML_ASSERT(src1_ncols == 1);
-
-    const int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
+    const int id = ggml_cuda_get_device();
+    const int cc = ggml_cuda_info().devices[id].cc;
     const enum ggml_prec prec = fast_fp16_available(cc) ? ggml_prec(dst->op_params[0]) : GGML_PREC_F32;
 
 
     // ggml_cuda_op provides single, contiguous matrices
     const int64_t stride_row         = ne00;
+    const int64_t stride_col_y       = ne10;
+    const int64_t stride_col_dst     = id == ctx.device ? ne0 : row_diff; // main device has larger memory buffer
     const int64_t nchannels_x        = 1;
     const int64_t nchannels_y        = 1;
     const int64_t nchannels_dst      = 1;
@@ -307,19 +414,19 @@ void ggml_cuda_op_mul_mat_vec(
     switch (src0->type) {
         case GGML_TYPE_F32: {
             const float * src0_d = (const float *) src0_dd_i;
-            mul_mat_vec_cuda(src0_d, src1_ddf_i, nullptr, dst_dd_i, ne00, row_diff, stride_row,
+            mul_mat_vec_cuda(src0_d, src1_ddf_i, nullptr, dst_dd_i, ne00, row_diff, src1_ncols, stride_row, stride_col_y, stride_col_dst,
                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, prec, stream);
         } break;
         case GGML_TYPE_F16: {
             const half * src0_d = (const half *) src0_dd_i;
-            mul_mat_vec_cuda(src0_d, src1_ddf_i, nullptr, dst_dd_i, ne00, row_diff, stride_row,
+            mul_mat_vec_cuda(src0_d, src1_ddf_i, nullptr, dst_dd_i, ne00, row_diff, src1_ncols, stride_row, stride_col_y, stride_col_dst,
                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, prec, stream);
         } break;
         case GGML_TYPE_BF16: {
             const nv_bfloat16 * src0_d = (const nv_bfloat16 *) src0_dd_i;
-            mul_mat_vec_cuda(src0_d, src1_ddf_i, nullptr, dst_dd_i, ne00, row_diff, stride_row,
+            mul_mat_vec_cuda(src0_d, src1_ddf_i, nullptr, dst_dd_i, ne00, row_diff, src1_ncols, stride_row, stride_col_y, stride_col_dst,
                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, prec, stream);
         } break;
@@ -334,3 +441,66 @@ void ggml_cuda_op_mul_mat_vec(
     GGML_UNUSED(src1_ncols);
     GGML_UNUSED(src1_padded_row_size);
 }
+
+bool ggml_cuda_should_use_mmv(enum ggml_type type, int cc, const int64_t * src0_ne, int64_t ne11) {
+    if (src0_ne[0] % 2 != 0) {
+        return false;
+    }
+    switch (type) {
+        case GGML_TYPE_F32:
+            if (GGML_CUDA_CC_IS_NVIDIA(cc)) {
+                if (cc >= GGML_CUDA_CC_ADA_LOVELACE) {
+                    return ne11 <= 8;
+                }
+                if (cc >= GGML_CUDA_CC_TURING) {
+                    return ne11 <= 4;
+                }
+                return ne11 <= 3;
+            } else if (GGML_CUDA_CC_IS_AMD(cc)) {
+                if (fp32_mma_hardware_available(cc)) {
+                    return ne11 <= 3;
+                }
+                return ne11 <= 8;
+            }
+            return ne11 <= 8;
+        case GGML_TYPE_F16:
+            if (GGML_CUDA_CC_IS_NVIDIA(cc)) {
+                const bool src0_small = (src0_ne[1] <= 512 || src0_ne[2]*src0_ne[3] == 1);
+                if (cc >= GGML_CUDA_CC_ADA_LOVELACE) {
+                    return src0_small && ne11 <= 4;
+                }
+                if (fp16_mma_hardware_available(cc)) {
+                    return src0_small && ne11 <= 3;
+                }
+                return ne11 <= 8;
+            } else if (GGML_CUDA_CC_IS_AMD(cc)) {
+                if (fp16_mma_hardware_available(cc)) {
+                    if (GGML_CUDA_CC_IS_RDNA3(cc) || GGML_CUDA_CC_IS_RDNA4(cc)) {
+                        return ne11 <= 5;
+                    }
+                    return ne11 <= 2;
+                }
+                return ne11 <= 8;
+            }
+            return ne11 <= 8;
+        case GGML_TYPE_BF16:
+            if (GGML_CUDA_CC_IS_NVIDIA(cc)) {
+                const bool src0_small = (src0_ne[1] <= 512 || src0_ne[2]*src0_ne[3] == 1);
+                if (cc >= GGML_CUDA_CC_ADA_LOVELACE) {
+                    return src0_small && ne11 <= 4;
+                }
+                if (bf16_mma_hardware_available(cc)) {
+                    return src0_small && ne11 <= 3;
+                }
+                return ne11 <= 8;
+            } else if (GGML_CUDA_CC_IS_AMD(cc)) {
+                if (bf16_mma_hardware_available(cc)) {
+                    return ne11 <= 3;
+                }
+                return ne11 <= 8;
+            }
+            return ne11 <= 8;
+        default:
+            return false;
+    }
+}
diff --git a/ggml/src/ggml-cuda/mmv.cuh b/ggml/src/ggml-cuda/mmv.cuh
index 756e7e1cc..1330bcb6a 100644
--- a/ggml/src/ggml-cuda/mmv.cuh
+++ b/ggml/src/ggml-cuda/mmv.cuh
@@ -1,8 +1,5 @@
 #include "common.cuh"
 
-// maximum number of src0 rows with which to use mul_mat_vec over cuBLAS if FP16 tensor cores are available
-#define MMV_MAX_ROWS 512
-
 void ggml_cuda_mul_mat_vec(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * ids, ggml_tensor * dst);
 
 void ggml_cuda_op_mul_mat_vec(
@@ -10,3 +7,5 @@ void ggml_cuda_op_mul_mat_vec(
     const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i,
     const char * src1_ddq_i, float * dst_dd_i, const int64_t row_low, const int64_t row_high, const int64_t src1_ncols,
     const int64_t src1_padded_row_size, cudaStream_t stream);
+
+bool ggml_cuda_should_use_mmv(enum ggml_type type, int cc, const int64_t * src0_ne, int64_t ne11);
diff --git a/ggml/src/ggml-cuda/softmax.cu b/ggml/src/ggml-cuda/softmax.cu
index aac6e0999..14543e978 100644
--- a/ggml/src/ggml-cuda/softmax.cu
+++ b/ggml/src/ggml-cuda/softmax.cu
@@ -2,6 +2,7 @@
 #include "ggml.h"
 #include "softmax.cuh"
 #include <cstdint>
+#include <utility>
 
 template <typename T>
 static __device__ __forceinline__ float t2f32(T val) {
@@ -13,6 +14,29 @@ __device__ float __forceinline__ t2f32<half>(half val) {
     return __half2float(val);
 }
 
+struct soft_max_params {
+
+    int64_t nheads;
+    uint32_t n_head_log2;
+    int64_t ncols;
+    int64_t nrows_x;
+    int64_t nrows_y;
+    int64_t ne00;
+    int64_t ne01;
+    int64_t ne02;
+    int64_t ne03;
+    int64_t nb11;
+    int64_t nb12;
+    int64_t nb13;
+
+    int64_t ne12;
+    int64_t ne13;
+    float scale;
+    float max_bias;
+    float m0;
+    float m1;
+};
+
 // When ncols_template == 0 the bounds for the loops in this function are not known and can't be unrolled.
 // As we want to keep pragma unroll for all other cases we supress the clang transformation warning here.
 #ifdef __clang__
@@ -21,16 +45,24 @@ __device__ float __forceinline__ t2f32<half>(half val) {
 #endif // __clang__
 template <bool use_shared, int ncols_template, int block_size_template, typename T>
 static __global__ void soft_max_f32(
-        const float * x, const T * mask, float * dst, const int ncols_par, const int nrows_y,
-        const float scale, const float max_bias, const float m0, const float m1, uint32_t n_head_log2) {
-    const int ncols = ncols_template == 0 ? ncols_par : ncols_template;
+        const float * x, const T * mask, float * dst, const soft_max_params p) {
+    const int ncols = ncols_template == 0 ? p.ncols : ncols_template;
 
     const int tid  = threadIdx.x;
-    const int rowx = blockIdx.x;
-    const int rowy = rowx % nrows_y; // broadcast the mask in the row dimension
+
+    const int64_t i03 = blockIdx.z;
+    const int64_t i02 = blockIdx.y;
+    const int64_t i01 = blockIdx.x;
+
+    //TODO: noncontigous inputs/outputs
+    const int rowx = blockIdx.x + blockIdx.y * gridDim.x + blockIdx.z * gridDim.x * gridDim.y;
+
+    const int64_t i11 = i01;
+    const int64_t i12 = i02 % p.ne12;
+    const int64_t i13 = i03 % p.ne13;
 
     x    += int64_t(rowx)*ncols;
-    mask += int64_t(rowy)*ncols * (mask != nullptr);
+    mask += (i11*p.nb11 + i12*p.nb12 + i13*p.nb13) / sizeof(T) * (mask != nullptr);
     dst  += int64_t(rowx)*ncols;
 
     const int block_size = block_size_template == 0 ? blockDim.x : block_size_template;
@@ -38,7 +70,7 @@ static __global__ void soft_max_f32(
     const int warp_id = threadIdx.x / WARP_SIZE;
     const int lane_id = threadIdx.x % WARP_SIZE;
 
-    const float slope = get_alibi_slope(max_bias, rowx/nrows_y, n_head_log2, m0, m1);
+    const float slope = get_alibi_slope(p.max_bias, i02, p.n_head_log2, p.m0, p.m1);
 
     extern __shared__ float data_soft_max_f32[];
     float * buf_iw = data_soft_max_f32; // shared memory buffer for inter-warp communication
@@ -55,7 +87,7 @@ static __global__ void soft_max_f32(
             break;
         }
 
-        const float val = x[col]*scale + (mask ? slope*t2f32(mask[col]) : 0.0f);
+        const float val = x[col]*p.scale + (mask ? slope*t2f32(mask[col]) : 0.0f);
 
         vals[col] = val;
         max_val = max(max_val, val);
@@ -150,64 +182,58 @@ static __global__ void soft_max_back_f32(
     }
 }
 
+template<int... Ns, typename T>
+static void launch_soft_max_kernels(const float * x, const T * mask, float * dst,
+                             const soft_max_params & p, cudaStream_t stream, dim3 block_dims, dim3 block_nums, size_t nbytes_shared)
+{
+    const int id       = ggml_cuda_get_device();
+    const size_t smpbo = ggml_cuda_info().devices[id].smpbo;
+
+    auto launch_kernel = [=](auto I) -> bool {
+        constexpr int ncols = decltype(I)::value;
+        constexpr int block = (ncols > 1024 ? 1024 : ncols);
+
+        if (p.ncols == ncols) {
+            CUDA_SET_SHARED_MEMORY_LIMIT((soft_max_f32<true, ncols, block, T>), smpbo);
+            soft_max_f32<true, ncols, block><<<block_nums, block_dims, nbytes_shared, stream>>>
+                (x, mask, dst, p);
+            return true;
+        }
+        return false;
+    };
+
+    // unary fold over launch_kernel
+    if ((launch_kernel(std::integral_constant<int, Ns>{}) || ...)) {
+        return;
+    }
+
+    //default case
+    CUDA_SET_SHARED_MEMORY_LIMIT((soft_max_f32<true, 0, 0, T>), smpbo);
+    soft_max_f32<true, 0, 0><<<block_nums, block_dims, nbytes_shared, stream>>>(x, mask, dst, p);
+}
+
+
 template<typename T>
-static void soft_max_f32_cuda(const float * x, const T * mask, float * dst, const int ncols_x, const int nrows_x, const int nrows_y, const float scale, const float max_bias, cudaStream_t stream) {
+static void soft_max_f32_cuda(const float * x, const T * mask, float * dst, const soft_max_params & params, cudaStream_t stream) {
     int nth = WARP_SIZE;
+    const int64_t ncols_x = params.ncols;
+
     while (nth < ncols_x && nth < CUDA_SOFT_MAX_BLOCK_SIZE) nth *= 2;
     const dim3 block_dims(nth,     1, 1);
-    const dim3 block_nums(nrows_x, 1, 1);
+    const dim3 block_nums(params.ne01, params.ne02, params.ne03);
     const size_t nbytes_shared = (GGML_PAD(ncols_x, WARP_SIZE) + WARP_SIZE)*sizeof(float);
     static_assert(CUDA_SOFT_MAX_BLOCK_SIZE == 1024, "These values need to be adjusted.");
 
-    const uint32_t n_head      = nrows_x/nrows_y;
-    const uint32_t n_head_log2 = 1u << (uint32_t) floorf(log2f((float) 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 id       = ggml_cuda_get_device();
+    const size_t smpbo = ggml_cuda_info().devices[id].smpbo;
 
-    // FIXME: this limit could be raised by ~2-4x on Ampere or newer
-    if (nbytes_shared < ggml_cuda_info().devices[ggml_cuda_get_device()].smpb) {
-        switch (ncols_x) {
-            case 32:
-                soft_max_f32<true,   32,   32><<<block_nums, block_dims, nbytes_shared, stream>>>
-                    (x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
-                break;
-            case 64:
-                soft_max_f32<true,   64,   64><<<block_nums, block_dims, nbytes_shared, stream>>>
-                    (x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
-                break;
-            case 128:
-                soft_max_f32<true,  128,  128><<<block_nums, block_dims, nbytes_shared, stream>>>
-                    (x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
-                break;
-            case 256:
-                soft_max_f32<true,  256,  256><<<block_nums, block_dims, nbytes_shared, stream>>>
-                    (x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
-                break;
-            case 512:
-                soft_max_f32<true,  512,  512><<<block_nums, block_dims, nbytes_shared, stream>>>
-                    (x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
-                break;
-            case 1024:
-                soft_max_f32<true, 1024, 1024><<<block_nums, block_dims, nbytes_shared, stream>>>
-                    (x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
-                break;
-            case 2048:
-                soft_max_f32<true, 2048, 1024><<<block_nums, block_dims, nbytes_shared, stream>>>
-                    (x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
-                break;
-            case 4096:
-                soft_max_f32<true, 4096, 1024><<<block_nums, block_dims, nbytes_shared, stream>>>
-                    (x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
-                break;
-            default:
-                soft_max_f32<true,    0,    0><<<block_nums, block_dims, nbytes_shared, stream>>>
-                    (x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
-                break;
-        }
+
+    if (nbytes_shared <= smpbo) {
+        launch_soft_max_kernels<32, 64, 128, 256, 512, 1024, 2048, 4096>(x, mask, dst, params, stream, block_dims, block_nums, nbytes_shared);
     } else {
         const size_t nbytes_shared_low = WARP_SIZE*sizeof(float);
-        soft_max_f32<false, 0, 0><<<block_nums, block_dims, nbytes_shared_low, stream>>>(x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
+        soft_max_f32<false, 0, 0><<<block_nums, block_dims, nbytes_shared_low, stream>>>(x, mask, dst, params);
     }
 }
 
@@ -235,10 +261,11 @@ void ggml_cuda_op_soft_max(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
 
     GGML_ASSERT(!src1 || src1->type == GGML_TYPE_F16 || src1->type == GGML_TYPE_F32); // src1 contains mask and it is optional
 
-    const int64_t ne00    = src0->ne[0];
     const int64_t nrows_x = ggml_nrows(src0);
     const int64_t nrows_y = src0->ne[1];
 
+    const int64_t ne00 = src0->ne[0];
+
     float scale    = 1.0f;
     float max_bias = 0.0f;
 
@@ -247,10 +274,44 @@ void ggml_cuda_op_soft_max(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
 
     const bool use_f16 = (src1 && src1->type == GGML_TYPE_F16);
 
+    const int64_t nb11 = src1 ? src1->nb[1] : 1;
+    const int64_t nb12 = src1 ? src1->nb[2] : 1;
+    const int64_t nb13 = src1 ? src1->nb[3] : 1;
+
+    const int64_t ne12 = src1 ? src1->ne[2] : 1;
+    const int64_t ne13 = src1 ? src1->ne[3] : 1;
+
+    const uint32_t n_head      = src0->ne[2];
+    const uint32_t n_head_log2 = 1u << (uint32_t) floorf(log2f((float) 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);
+
+
+    soft_max_params params = {};
+    params.nheads = src0->ne[2];
+    params.n_head_log2 = n_head_log2;
+    params.ncols = ne00;
+    params.nrows_x = nrows_x;
+    params.nrows_y = nrows_y;
+    params.ne00 = src0->ne[0];
+    params.ne01 = src0->ne[1];
+    params.ne02 = src0->ne[2];
+    params.ne03 = src0->ne[3];
+    params.nb11 = nb11;
+    params.nb12 = nb12;
+    params.nb13 = nb13;
+    params.ne12 = ne12;
+    params.ne13 = ne13;
+    params.scale = scale;
+    params.max_bias = max_bias;
+    params.m0 = m0;
+    params.m1 = m1;
+
     if (use_f16) {
-        soft_max_f32_cuda(src0_d, (const half  *) src1_d, dst_d, ne00, nrows_x, nrows_y, scale, max_bias, stream);
+        soft_max_f32_cuda(src0_d, (const half  *) src1_d, dst_d, params, stream);
     } else {
-        soft_max_f32_cuda(src0_d, (const float *) src1_d, dst_d, ne00, nrows_x, nrows_y, scale, max_bias, stream);
+        soft_max_f32_cuda(src0_d, (const float *) src1_d, dst_d, params, stream);
     }
 }
 
diff --git a/ggml/src/ggml-cuda/ssm-scan.cu b/ggml/src/ggml-cuda/ssm-scan.cu
index 37ee208c0..dc3b1a9a8 100644
--- a/ggml/src/ggml-cuda/ssm-scan.cu
+++ b/ggml/src/ggml-cuda/ssm-scan.cu
@@ -4,14 +4,15 @@ template <size_t splitD, size_t N>
 __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 int32_t * __restrict__ src6, float * __restrict__ dst,
+                 const int src0_nb2, const int src0_nb3, const int src1_nb2, const int src1_nb3,
+                 const int src2_nb1, const int src2_nb2, const int src3_nb1,
+                 const int src4_nb2, const int src4_nb3, const int src5_nb2, const int src5_nb3,
+                 const int64_t s_off, const int64_t d_inner, const int64_t L) {
+
+    constexpr int warp_size = ggml_cuda_get_physical_warp_size();
+    const int bidx = blockIdx.x;  // split along B (sequences)
+    const int bidy = blockIdx.y;  // split along D (d_inner)
     const int tid  = threadIdx.x;
     const int wid  = tid / 32;
     const int wtid = tid % 32;
@@ -22,38 +23,38 @@ __global__ void __launch_bounds__(splitD, 2)
     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 * s0_block = (const float *) ((const char *) src0 + src6[bidx] * src0_nb3 + bidy * splitD * src0_nb2);
+    const float * x_block  = (const float *) ((const char *) src1 + (bidx * src1_nb3) + 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 float * B_block  = (const float *) ((const char *) src4 + (bidx * src4_nb3));
+    const float * C_block  = (const float *) ((const char *) src5 + (bidx * src5_nb3));
+    float *       y_block  = (float *) ((char *) dst + (bidx * d_inner * L * sizeof(float)) + bidy * splitD * sizeof(float));
+    float *       s_block  = (float *) ((char *) dst + s_off + bidx * src0_nb3 + bidy * splitD * src0_nb2);
 
-    const int stride_s0 = src0_nb1 / sizeof(float);
-    const int stride_x  = src1_nb1 / sizeof(float);
+    const int stride_s0 = src0_nb2 / sizeof(float);
+    const int stride_x  = src1_nb2 / 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_B  = src4_nb2 / sizeof(float);
+    const int stride_C  = src5_nb2 / sizeof(float);
     const int stride_s  = stride_s0;
-    const int stride_y  = stride_x;
+    const int stride_y  = d_inner;
 
     // 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];
+            float value = A_block[(wid * warp_size + 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;
+            smem_A[(wid * warp_size + 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;
+            float value = s0_block[(wid * warp_size + i) * stride_s0 + wtid];
+            smem_s0[(wid * warp_size + i) * stride_ss0 + wtid + ((wtid / 16) > 0 ? 1 : 0)] = value;
         }
     }
 
@@ -82,24 +83,156 @@ __global__ void __launch_bounds__(splitD, 2)
     }
 }
 
+// assumes as many threads as d_state
+template <int splitH, int d_state>
+__global__ void __launch_bounds__(d_state, 1)
+    ssm_scan_f32_group(
+        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 int32_t * __restrict__ src6, float * __restrict__ dst,
+        const int src0_nb2, const int src0_nb3, const int src1_nb2, const int src1_nb3,
+        const int src2_nb1, const int src2_nb2, const int src3_nb1,
+        const int src4_nb2, const int src4_nb3, const int src5_nb2, const int src5_nb3,
+        const int64_t s_off, const int64_t n_head, const int64_t d_head, const int64_t n_group, const int64_t n_tok) {
+
+    const int head_idx = (blockIdx.x * splitH) / d_head;
+    const int head_off = ((blockIdx.x * splitH) % d_head) * sizeof(float);
+    const int seq_idx = blockIdx.y;
+
+    const int group_off = (head_idx & (n_group - 1)) * d_state * sizeof(float);
+
+    const float * s0_block = (const float *) ((const char *) src0 + src6[seq_idx] * src0_nb3 + head_idx * src0_nb2 + head_off * d_state);
+    const float * x_block  = (const float *) ((const char *) src1 + (seq_idx * src1_nb3) + blockIdx.x * splitH * sizeof(float));
+    const float * dt_block = (const float *) ((const char *) src2 + (seq_idx * src2_nb2) + head_idx * sizeof(float));
+    const float * A_block  = (const float *) ((const char *) src3 + head_idx * src3_nb1);
+    const float * B_block  = (const float *) ((const char *) src4 + (seq_idx * src4_nb3) + (group_off));
+    const float * C_block  = (const float *) ((const char *) src5 + (seq_idx * src5_nb3) + (group_off));
+    float *       y_block  = dst + (seq_idx * n_tok * n_head * d_head) + blockIdx.x * splitH;
+    float *       s_block  = (float *) ((char *) dst + s_off + seq_idx * src0_nb3 + head_idx * src0_nb2 + head_off * d_state);
+
+    // strides across n_seq_tokens
+    const int stride_x  = src1_nb2 / sizeof(float);
+    const int stride_dt = src2_nb1 / sizeof(float);
+    const int stride_B  = src4_nb2 / sizeof(float);
+    const int stride_C  = src5_nb2 / sizeof(float);
+    const int stride_y  = n_head * d_head;
+
+    float state[splitH];
+    // for the parallel accumulation
+    __shared__ float stateC[splitH * d_state];
+
+#pragma unroll
+    for (int j = 0; j < splitH; j++) {
+        state[j] = s0_block[j * d_state + threadIdx.x];
+    }
+
+    for (int64_t i = 0; i < n_tok; i++) {
+        // TODO: only calculate dA and dt_soft_plus once per head instead of every splitH head elements
+        // TODO: only calculate B and C once per head group
+        // NOTE: dt_soft_plus, dA and x_dt have the same value across threads here.
+        float dt_soft_plus = dt_block[i * stride_dt];
+        if (dt_soft_plus <= 20.0f) {
+            dt_soft_plus = log1pf(expf(dt_soft_plus));
+        }
+        const float dA = expf(dt_soft_plus * A_block[0]);
+        const float B = B_block[i * stride_B + threadIdx.x];
+        const float C = C_block[i * stride_C + threadIdx.x];
+
+        // across d_head
+#pragma unroll
+        for (int j = 0; j < splitH; j++) {
+            const float x_dt = x_block[i * stride_x + j] * dt_soft_plus;
+
+            state[j] = (state[j] * dA) + (B * x_dt);
+
+            stateC[j * d_state + threadIdx.x] = state[j] * C;
+        }
+
+        __syncthreads();
+
+        // parallel accumulation for stateC
+        // TODO: simplify
+        {
+            static_assert((d_state & -d_state) == d_state, "the state size has to be a power of 2");
+            static_assert((splitH & -splitH) == splitH, "splitH has to be a power of 2");
+
+            // reduce until w matches the warp size
+            // TODO: does this work even when the physical warp size is 64?
+#pragma unroll
+            for (int w = d_state; w > WARP_SIZE; w >>= 1) {
+                // (assuming there are d_state threads)
+#pragma unroll
+                for (int j = 0; j < ((w >> 1) * splitH + d_state - 1) / d_state; j++) {
+                    // TODO: check for bank conflicts
+                    const int k = (threadIdx.x % (w >> 1)) + (d_state * (threadIdx.x / (w >> 1))) + j * d_state * (d_state / (w >> 1));
+                    stateC[k] += stateC[k + (w >> 1)];
+
+                }
+                __syncthreads();
+            }
+
+            static_assert(splitH >= d_state / WARP_SIZE);
+
+#pragma unroll
+            for (int j = 0; j < splitH / (d_state / WARP_SIZE); j++) {
+                float y = stateC[(threadIdx.x % WARP_SIZE) + d_state * (threadIdx.x / WARP_SIZE) + j * d_state * (d_state / WARP_SIZE)];
+                y = warp_reduce_sum(y);
+
+                // store the above accumulations
+                if (threadIdx.x % WARP_SIZE == 0) {
+                    const int k = threadIdx.x / WARP_SIZE + j * (d_state / WARP_SIZE);
+                    y_block[i * stride_y + k] = y;
+                }
+            }
+        }
+    }
+
+    // write back the state
+#pragma unroll
+    for (int j = 0; j < splitH; j++) {
+        s_block[j * d_state + threadIdx.x] = state[j];
+    }
+}
+
 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,
+                              const float * src4, const float * src5, const int32_t * src6, float * dst,
+                              const int src0_nb2, const int src0_nb3, const int src1_nb2, const int src1_nb3, const int src2_nb1,
+                              const int src2_nb2, const int src3_nb1, const int src4_nb2, const int src4_nb3, const int src5_nb2,
+                              const int src5_nb3, const int64_t s_off, const int64_t d_state, const int64_t head_dim,
+                              const int64_t n_head, const int64_t n_group, const int64_t n_tok, const int64_t n_seq,
                               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><<<blocks, threads, smem_size, stream>>>(
-            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);
+    // NOTE: if you change conditions here, be sure to update the corresponding supports_op condition!
+    if (src3_nb1 == sizeof(float)) {
+        // Mamba-2
+        if (d_state == 128) {
+            GGML_ASSERT(d_state % threads == 0);
+            // NOTE: can be any power of two between 4 and 64
+            const int splitH = 16;
+            GGML_ASSERT(head_dim % splitH == 0);
+            const dim3 blocks((n_head * head_dim + (splitH - 1)) / splitH, n_seq, 1);
+            ssm_scan_f32_group<16, 128><<<blocks, threads, 0, stream>>>(
+                    src0, src1, src2, src3, src4, src5, src6, dst,
+                    src0_nb2, src0_nb3, src1_nb2, src1_nb3, src2_nb1, src2_nb2, src3_nb1,
+                    src4_nb2, src4_nb3, src5_nb2, src5_nb3, s_off, n_head, head_dim, n_group, n_tok);
+        } else {
+            GGML_ABORT("doesn't support d_state!=128.");
+        }
     } else {
-        GGML_ABORT("doesn't support N!=16.");
+        // Mamba-1
+        GGML_ASSERT(n_head % threads == 0);
+        GGML_ASSERT(head_dim == 1);
+        GGML_ASSERT(n_group == 1);
+        const dim3 blocks(n_seq, (n_head + threads - 1) / threads, 1);
+        const int  smem_size = (threads * (d_state + 1) * 2) * sizeof(float);
+        if (d_state == 16) {
+            ssm_scan_f32<128, 16><<<blocks, threads, smem_size, stream>>>(
+                src0, src1, src2, src3, src4, src5, src6, dst,
+                src0_nb2, src0_nb3, src1_nb2, src1_nb3, src2_nb1, src2_nb2,
+                src3_nb1, src4_nb2, src4_nb3, src5_nb2, src5_nb3, s_off, n_head, n_tok);
+        } else {
+            GGML_ABORT("doesn't support d_state!=16.");
+        }
     }
 }
 
@@ -110,30 +243,25 @@ void ggml_cuda_op_ssm_scan(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     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 struct ggml_tensor * src6 = dst->src[6];  // ids
 
     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
+    const int64_t nr  = src0->ne[1];  // head_dim or 1
+    const int64_t nh  = src1->ne[1];  // n_head
+    const int64_t ng  = src4->ne[1];  // n_group
+    const int64_t n_t = src1->ne[2];  // number of tokens per sequence
+    const int64_t n_s = src1->ne[3];  // number of sequences in the batch
 
-    GGML_ASSERT(ggml_nelements(src1) + ggml_nelements(src0) == ggml_nelements(dst));
+    const int64_t s_off = ggml_nelements(src1) * sizeof(float);
+
+    GGML_ASSERT(ggml_nelements(src1) + nc*nr*nh*n_s == 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));
+    GGML_ASSERT(src6->nb[0] == sizeof(int32_t));
 
     const float * src0_d = (const float *) src0->data;
     const float * src1_d = (const float *) src1->data;
@@ -141,13 +269,16 @@ void ggml_cuda_op_ssm_scan(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     const float * src3_d = (const float *) src3->data;
     const float * src4_d = (const float *) src4->data;
     const float * src5_d = (const float *) src5->data;
+    const int32_t * src6_d = (const int32_t *) src6->data;
     float *       dst_d  = (float *) dst->data;
     cudaStream_t  stream = ctx.stream();
 
     GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT(src6->type == GGML_TYPE_I32);
     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);
+    ssm_scan_f32_cuda(src0_d, src1_d, src2_d, src3_d, src4_d, src5_d, src6_d, dst_d,
+                      src0->nb[2], src0->nb[3], src1->nb[2], src1->nb[3], src2->nb[1], src2->nb[2],
+                      src3->nb[1], src4->nb[2], src4->nb[3], src5->nb[2], src5->nb[3],
+                      s_off, nc, nr, nh, ng, n_t, n_s, stream);
 }
diff --git a/ggml/src/ggml-cuda/sumrows.cu b/ggml/src/ggml-cuda/sumrows.cu
index 38dbf1b5e..2eee08fa0 100644
--- a/ggml/src/ggml-cuda/sumrows.cu
+++ b/ggml/src/ggml-cuda/sumrows.cu
@@ -1,25 +1,9 @@
 #include "sumrows.cuh"
 
-static __global__ void k_sum_rows_f32(const float * x, float * dst, const int ncols) {
-    const int row = blockIdx.x;
-    const int col = threadIdx.x;
-
-    float sum = 0.0f;
-    for (int i = col; i < ncols; i += blockDim.x) {
-        sum += x[row * ncols + i];
-    }
-
-    sum = warp_reduce_sum(sum);
-
-    if (col == 0) {
-        dst[row] = sum;
-    }
-}
-
 void sum_rows_f32_cuda(const float * x, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
     const dim3 block_dims(WARP_SIZE, 1, 1);
     const dim3 block_nums(nrows, 1, 1);
-    k_sum_rows_f32<<<block_nums, block_dims, 0, stream>>>(x, dst, ncols);
+    reduce_rows_f32</*norm*/false><<<block_nums, block_dims, 0, stream>>>(x, dst, ncols);
 }
 
 void ggml_cuda_op_sum_rows(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
@@ -35,5 +19,8 @@ void ggml_cuda_op_sum_rows(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     const int64_t ncols = src0->ne[0];
     const int64_t nrows = ggml_nrows(src0);
 
-    sum_rows_f32_cuda(src0_d, dst_d, ncols, nrows, stream);
+    const dim3 block_dims(WARP_SIZE, 1, 1);
+    const dim3 block_nums(nrows, 1, 1);
+
+    reduce_rows_f32</*norm=*/false><<<block_nums, block_dims, 0, stream>>>(src0_d, dst_d, ncols);
 }
diff --git a/ggml/src/ggml-cuda/sumrows.cuh b/ggml/src/ggml-cuda/sumrows.cuh
index 191db1c13..3431c599b 100644
--- a/ggml/src/ggml-cuda/sumrows.cuh
+++ b/ggml/src/ggml-cuda/sumrows.cuh
@@ -1,5 +1,4 @@
 #include "common.cuh"
 
 void sum_rows_f32_cuda(const float * x, float * dst, const int ncols, const int nrows, cudaStream_t stream);
-
 void ggml_cuda_op_sum_rows(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
diff --git a/ggml/src/ggml-cuda/unary.cu b/ggml/src/ggml-cuda/unary.cu
index 2c0375fbe..f9c7b83c4 100644
--- a/ggml/src/ggml-cuda/unary.cu
+++ b/ggml/src/ggml-cuda/unary.cu
@@ -196,6 +196,103 @@ void ggml_cuda_op_log(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     ggml_cuda_op_unary<op_log>(ctx, dst);
 }
 
+/* gated ops */
+
+template <float (*op)(float), typename T>
+static __global__ void unary_gated_op_kernel(const T * x, const T * g, T * dst, const int64_t k, const int64_t n, const int64_t o0, const int64_t o1) {
+    const int64_t i = int64_t(blockDim.x)*blockIdx.x + threadIdx.x;
+
+    if (i >= k) {
+        return;
+    }
+
+    // perform base op and multiply with gate (either offset in same tensor or a separate one)
+    const int64_t j0 = (i / n) * o0 + (i % n);
+    const int64_t j1 = o0 == o1 ? j0 : (i / n) * o1 + (i % n);
+
+    dst[i] = (T)(op((float)x[j0]) * (float)g[j1]);
+}
+
+template <float (*op)(float), typename T>
+static void unary_gated_cuda(const T * x, const T * g, T * dst, const int64_t k, const int64_t n, const int64_t o0, const int64_t o1, cudaStream_t stream) {
+    const int64_t num_blocks = (k + CUDA_GLU_BLOCK_SIZE - 1) / CUDA_GLU_BLOCK_SIZE;
+    unary_gated_op_kernel<op><<<num_blocks, CUDA_GLU_BLOCK_SIZE, 0, stream>>>(x, g, dst, k, n, o0, o1);
+}
+
+template <float (*op)(float)>
+void ggml_cuda_op_unary_gated(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const ggml_tensor * src1 = dst->src[1];
+    void * src0_d = src0->data;
+    void * src1_d = src1 ? src1->data : src0->data;
+    const int64_t src0_o = src0->nb[1];
+    const int64_t src1_o = src1 ? src1->nb[1] : src0->nb[1];
+    void * dst_d = dst->data;
+    const int64_t nc = src1 ? src0->ne[0] : src0->ne[0] / 2;
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(ggml_is_contiguous_1(src0));
+    GGML_ASSERT(src0->nb[0] == ggml_element_size(src0));
+    GGML_ASSERT(ggml_is_contiguous(dst));
+
+    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);
+    GGML_ASSERT(dst->ne[0] == nc);
+    GGML_ASSERT(ggml_nrows(dst) == ggml_nrows(src0));
+
+    if (src1) {
+        GGML_ASSERT(ggml_is_contiguous_1(src1));
+        GGML_ASSERT(src1->nb[0] == ggml_element_size(src1));
+        GGML_ASSERT(src1->ne[0] == nc);
+        GGML_ASSERT(src0->type == src1->type);
+    }
+
+    const int32_t swapped = ((const int32_t *) dst->op_params)[1];
+
+    if (src0->type == GGML_TYPE_F16) {
+        half * src0_p = (half *) src0_d;
+        half * src1_p = (half *) src1_d;
+
+        if (!src1) {
+            src0_p += swapped ? nc : 0;
+            src1_p += swapped ? 0 : nc;
+        }
+
+        unary_gated_cuda<op>(src0_p, src1_p, (half *)dst_d, ggml_nelements(dst), nc, src0_o / sizeof(half), src1_o / sizeof(half), stream);
+    } else {
+        float * src0_p = (float *) src0_d;
+        float * src1_p = (float *) src1_d;
+
+        if (!src1) {
+            src0_p += swapped ? nc : 0;
+            src1_p += swapped ? 0 : nc;
+        }
+
+        unary_gated_cuda<op>(src0_p, src1_p, (float *)dst_d, ggml_nelements(dst), nc, src0_o / sizeof(float), src1_o / sizeof(float), stream);
+    }
+}
+
+void ggml_cuda_op_reglu(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    ggml_cuda_op_unary_gated<op_relu>(ctx, dst);
+}
+
+void ggml_cuda_op_geglu(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    ggml_cuda_op_unary_gated<op_gelu>(ctx, dst);
+}
+
+void ggml_cuda_op_swiglu(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    ggml_cuda_op_unary_gated<op_silu>(ctx, dst);
+}
+
+void ggml_cuda_op_geglu_erf(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    ggml_cuda_op_unary_gated<op_gelu_erf>(ctx, dst);
+}
+
+void ggml_cuda_op_geglu_quick(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    ggml_cuda_op_unary_gated<op_gelu_quick>(ctx, dst);
+}
+
 /* silu_back */
 
 static __device__ __forceinline__ float op_silu_back(float grad, float x) {
diff --git a/ggml/src/ggml-cuda/unary.cuh b/ggml/src/ggml-cuda/unary.cuh
index 6686fc17e..289d690e5 100644
--- a/ggml/src/ggml-cuda/unary.cuh
+++ b/ggml/src/ggml-cuda/unary.cuh
@@ -15,6 +15,7 @@
 #define CUDA_SQRT_BLOCK_SIZE 256
 #define CUDA_SIN_BLOCK_SIZE 256
 #define CUDA_COS_BLOCK_SIZE 256
+#define CUDA_GLU_BLOCK_SIZE 256
 
 void ggml_cuda_op_abs(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
 
@@ -57,3 +58,13 @@ 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);
+
+void ggml_cuda_op_reglu(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
+void ggml_cuda_op_geglu(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
+void ggml_cuda_op_swiglu(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
+void ggml_cuda_op_geglu_erf(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
+void ggml_cuda_op_geglu_quick(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
diff --git a/ggml/src/ggml-hip/CMakeLists.txt b/ggml/src/ggml-hip/CMakeLists.txt
index 1fe8fe3b8..e29df9856 100644
--- a/ggml/src/ggml-hip/CMakeLists.txt
+++ b/ggml/src/ggml-hip/CMakeLists.txt
@@ -113,6 +113,10 @@ if (GGML_HIP_ROCWMMA_FATTN)
     add_compile_definitions(GGML_HIP_ROCWMMA_FATTN)
 endif()
 
+if (GGML_HIP_FORCE_ROCWMMA_FATTN_GFX12 OR ${hip_VERSION} VERSION_GREATER_EQUAL 7.0)
+    add_compile_definitions(GGML_HIP_ROCWMMA_FATTN_GFX12)
+endif()
+
 if (NOT GGML_CUDA_FA)
     add_compile_definitions(GGML_CUDA_NO_FA)
 endif()
diff --git a/ggml/src/ggml-impl.h b/ggml/src/ggml-impl.h
index 6dc5ce0d9..4972558c9 100644
--- a/ggml/src/ggml-impl.h
+++ b/ggml/src/ggml-impl.h
@@ -301,6 +301,7 @@ struct ggml_cgraph {
     struct ggml_tensor ** grads;     // the outputs of these tensors are the gradients of the nodes
     struct ggml_tensor ** grad_accs; // accumulators for node gradients
     struct ggml_tensor ** leafs;     // tensors with constant data
+    int32_t             * use_counts;// number of uses of each tensor, indexed by hash table slot
 
     struct ggml_hash_set visited_hash_set;
 
@@ -317,203 +318,81 @@ struct ggml_cgraph ggml_graph_view(struct ggml_cgraph * cgraph, int i0, int i1);
 GGML_API void * ggml_aligned_malloc(size_t size);
 GGML_API void ggml_aligned_free(void * ptr, size_t size);
 
-// FP16 to FP32 conversion
+// FP16 <-> FP32
+// ref: https://github.com/Maratyszcza/FP16
 
-// 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) {
-        __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;
-        __fp16 tmp = f;
-        memcpy(&res, &tmp, sizeof(ggml_fp16_t));
-        return res;
-    }
-
-#elif defined(__F16C__)
-
-    #ifdef _MSC_VER
-        #define GGML_COMPUTE_FP16_TO_FP32(x) _mm_cvtss_f32(_mm_cvtph_ps(_mm_cvtsi32_si128(x)))
-        #define GGML_COMPUTE_FP32_TO_FP16(x) _mm_extract_epi16(_mm_cvtps_ph(_mm_set_ss(x), 0), 0)
-    #else
-        #define GGML_COMPUTE_FP16_TO_FP32(x) _cvtsh_ss(x)
-        #define GGML_COMPUTE_FP32_TO_FP16(x) _cvtss_sh(x, 0)
-    #endif
-
-#elif defined(__POWER9_VECTOR__)
-
-    #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)
-    /* the inline asm below is about 12% faster than the lookup method */
-    #define GGML_FP16_TO_FP32(x) GGML_COMPUTE_FP16_TO_FP32(x)
-    #define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x)
-
-    static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
-        float f;
-        double d;
-        __asm__(
-            "mtfprd %0,%2\n"
-            "xscvhpdp %0,%0\n"
-            "frsp %1,%0\n" :
-            /* temp */ "=d"(d),
-            /* out */  "=f"(f):
-            /* in */   "r"(h));
-        return f;
-    }
-
-    static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
-        double d;
-        ggml_fp16_t r;
-        __asm__( /* xscvdphp can work on double or single precision */
-            "xscvdphp %0,%2\n"
-            "mffprd %1,%0\n" :
-            /* temp */ "=d"(d),
-            /* out */  "=r"(r):
-            /* in */   "f"(f));
-        return r;
-    }
-
-#elif defined(__riscv) && defined(__riscv_zfhmin)
-
-    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
-    // ref: https://github.com/Maratyszcza/FP16
-
-    static inline float fp32_from_bits(uint32_t w) {
-        union {
-            uint32_t as_bits;
-            float as_value;
-        } fp32;
-        fp32.as_bits = w;
-        return fp32.as_value;
-    }
-
-    static inline uint32_t fp32_to_bits(float f) {
-        union {
-            float as_value;
-            uint32_t as_bits;
-        } fp32;
-        fp32.as_value = f;
-        return fp32.as_bits;
-    }
-
-    static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
-        const uint32_t w = (uint32_t) h << 16;
-        const uint32_t sign = w & UINT32_C(0x80000000);
-        const uint32_t two_w = w + w;
-
-        const uint32_t exp_offset = UINT32_C(0xE0) << 23;
-    #if (defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) || defined(__GNUC__) && !defined(__STRICT_ANSI__)) && (!defined(__cplusplus) || __cplusplus >= 201703L)
-        const float exp_scale = 0x1.0p-112f;
-    #else
-        const float exp_scale = fp32_from_bits(UINT32_C(0x7800000));
-    #endif
-        const float normalized_value = fp32_from_bits((two_w >> 4) + exp_offset) * exp_scale;
-
-        const uint32_t magic_mask = UINT32_C(126) << 23;
-        const float magic_bias = 0.5f;
-        const float denormalized_value = fp32_from_bits((two_w >> 17) | magic_mask) - magic_bias;
-
-        const uint32_t denormalized_cutoff = UINT32_C(1) << 27;
-        const uint32_t result = sign |
-            (two_w < denormalized_cutoff ? fp32_to_bits(denormalized_value) : fp32_to_bits(normalized_value));
-        return fp32_from_bits(result);
-    }
-
-    static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
-    #if (defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) || defined(__GNUC__) && !defined(__STRICT_ANSI__)) && (!defined(__cplusplus) || __cplusplus >= 201703L)
-        const float scale_to_inf = 0x1.0p+112f;
-        const float scale_to_zero = 0x1.0p-110f;
-    #else
-        const float scale_to_inf = fp32_from_bits(UINT32_C(0x77800000));
-        const float scale_to_zero = fp32_from_bits(UINT32_C(0x08800000));
-    #endif
-        float base = (fabsf(f) * scale_to_inf) * scale_to_zero;
-
-        const uint32_t w = fp32_to_bits(f);
-        const uint32_t shl1_w = w + w;
-        const uint32_t sign = w & UINT32_C(0x80000000);
-        uint32_t bias = shl1_w & UINT32_C(0xFF000000);
-        if (bias < UINT32_C(0x71000000)) {
-            bias = UINT32_C(0x71000000);
-        }
-
-        base = fp32_from_bits((bias >> 1) + UINT32_C(0x07800000)) + base;
-        const uint32_t bits = fp32_to_bits(base);
-        const uint32_t exp_bits = (bits >> 13) & UINT32_C(0x00007C00);
-        const uint32_t mantissa_bits = bits & UINT32_C(0x00000FFF);
-        const uint32_t nonsign = exp_bits + mantissa_bits;
-        return (sign >> 16) | (shl1_w > UINT32_C(0xFF000000) ? UINT16_C(0x7E00) : nonsign);
-    }
-
-    #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(__CUDACC__) && __CUDACC_VER_MAJOR__ <= 11) && !defined(__MUSACC__)
-
-// precomputed f32 table for f16 (256 KB)
-// defined in ggml.c, initialized in ggml_init()
-GGML_API float ggml_table_f32_f16[1 << 16];
-
-// On ARM NEON, it's quicker to directly convert x -> x instead of calling into ggml_lookup_fp16_to_fp32,
-// so we define GGML_FP16_TO_FP32 and GGML_FP32_TO_FP16 elsewhere for NEON.
-// This is also true for POWER9.
-#if !defined(GGML_FP16_TO_FP32)
-inline static float ggml_lookup_fp16_to_fp32(ggml_fp16_t f) {
-    uint16_t s;
-    memcpy(&s, &f, sizeof(uint16_t));
-    return ggml_table_f32_f16[s];
+static inline float fp32_from_bits(uint32_t w) {
+    union {
+        uint32_t as_bits;
+        float as_value;
+    } fp32;
+    fp32.as_bits = w;
+    return fp32.as_value;
 }
 
-#define GGML_FP16_TO_FP32(x) ggml_lookup_fp16_to_fp32(x)
-#endif
+static inline uint32_t fp32_to_bits(float f) {
+    union {
+        float as_value;
+        uint32_t as_bits;
+    } fp32;
+    fp32.as_value = f;
+    return fp32.as_bits;
+}
 
-#if !defined(GGML_FP32_TO_FP16)
-#define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x)
+static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
+    const uint32_t w = (uint32_t) h << 16;
+    const uint32_t sign = w & UINT32_C(0x80000000);
+    const uint32_t two_w = w + w;
+
+    const uint32_t exp_offset = UINT32_C(0xE0) << 23;
+#if (defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) || defined(__GNUC__) && !defined(__STRICT_ANSI__)) && (!defined(__cplusplus) || __cplusplus >= 201703L)
+    const float exp_scale = 0x1.0p-112f;
+#else
+    const float exp_scale = fp32_from_bits(UINT32_C(0x7800000));
 #endif
+    const float normalized_value = fp32_from_bits((two_w >> 4) + exp_offset) * exp_scale;
+
+    const uint32_t magic_mask = UINT32_C(126) << 23;
+    const float magic_bias = 0.5f;
+    const float denormalized_value = fp32_from_bits((two_w >> 17) | magic_mask) - magic_bias;
+
+    const uint32_t denormalized_cutoff = UINT32_C(1) << 27;
+    const uint32_t result = sign |
+        (two_w < denormalized_cutoff ? fp32_to_bits(denormalized_value) : fp32_to_bits(normalized_value));
+    return fp32_from_bits(result);
+}
+
+static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
+#if (defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) || defined(__GNUC__) && !defined(__STRICT_ANSI__)) && (!defined(__cplusplus) || __cplusplus >= 201703L)
+    const float scale_to_inf = 0x1.0p+112f;
+    const float scale_to_zero = 0x1.0p-110f;
+#else
+    const float scale_to_inf = fp32_from_bits(UINT32_C(0x77800000));
+    const float scale_to_zero = fp32_from_bits(UINT32_C(0x08800000));
+#endif
+    float base = (fabsf(f) * scale_to_inf) * scale_to_zero;
+
+    const uint32_t w = fp32_to_bits(f);
+    const uint32_t shl1_w = w + w;
+    const uint32_t sign = w & UINT32_C(0x80000000);
+    uint32_t bias = shl1_w & UINT32_C(0xFF000000);
+    if (bias < UINT32_C(0x71000000)) {
+        bias = UINT32_C(0x71000000);
+    }
+
+    base = fp32_from_bits((bias >> 1) + UINT32_C(0x07800000)) + base;
+    const uint32_t bits = fp32_to_bits(base);
+    const uint32_t exp_bits = (bits >> 13) & UINT32_C(0x00007C00);
+    const uint32_t mantissa_bits = bits & UINT32_C(0x00000FFF);
+    const uint32_t nonsign = exp_bits + mantissa_bits;
+    return (sign >> 16) | (shl1_w > UINT32_C(0xFF000000) ? UINT16_C(0x7E00) : nonsign);
+}
+
+#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)
 
 /**
  * Converts brain16 to float32.
@@ -589,13 +468,76 @@ static inline ggml_bf16_t ggml_compute_fp32_to_bf16(float s) {
 #define GGML_FP32_TO_BF16(x) ggml_compute_fp32_to_bf16(x)
 #define GGML_BF16_TO_FP32(x) ggml_compute_bf16_to_fp32(x)
 
+// return true if the node's results are only used by N other nodes
+// and can be fused into their calculations.
+static inline bool ggml_node_has_n_uses(const struct ggml_cgraph * cgraph, int node_idx, int32_t n_uses) {
+    const struct ggml_tensor * node = cgraph->nodes[node_idx];
+
+    // check the use count against how many we're replacing
+    size_t hash_pos = ggml_hash_find(&cgraph->visited_hash_set, node);
+    if (!ggml_bitset_get(cgraph->visited_hash_set.used, hash_pos) || cgraph->use_counts[hash_pos] != n_uses) {
+        return false;
+    }
+
+    // if node is a view, some other node might be using the intermediate result
+    // via the view source.
+    if (node->view_src) {
+        return false;
+    }
+
+    // If the user requested output for the node, can't fuse
+    if (node->flags & GGML_TENSOR_FLAG_OUTPUT) {
+        return false;
+    }
+
+    return true;
+}
+
+// Returns true if nodes [i, i+ops.size()) are the sequence of ggml_ops in ops[]
+// and are fusable. Nodes are considered fusable according to this function if:
+// - all nodes except the last have only one use and are not views/outputs (see ggml_node_has_N_uses).
+// - all nodes except the last are a src of the following node.
+// - all nodes are the same shape.
+// TODO: Consider allowing GGML_OP_NONE nodes in between
+static inline bool ggml_can_fuse(const struct ggml_cgraph * cgraph, int node_idx, const enum ggml_op * ops, int num_ops) {
+    if (node_idx + num_ops > cgraph->n_nodes) {
+        return false;
+    }
+
+    for (int i = 0; i < num_ops; ++i) {
+        struct ggml_tensor * node = cgraph->nodes[node_idx + i];
+        if (node->op != ops[i]) {
+            return false;
+        }
+        if (i < num_ops - 1 && !ggml_node_has_n_uses(cgraph, node_idx + i, 1)) {
+            return false;
+        }
+        if (i > 0) {
+            struct ggml_tensor * prev = cgraph->nodes[node_idx + i - 1];
+            if (node->src[0] != prev && node->src[1] != prev) {
+                return false;
+            }
+            if (!ggml_are_same_shape(node, prev)) {
+                return false;
+            }
+        }
+    }
+    return true;
+}
+
 #ifdef __cplusplus
 }
 #endif
 
 #ifdef __cplusplus
+#include <initializer_list>
 #include <vector>
 
+// nicer C++ syntax for ggml_can_fuse
+inline bool ggml_can_fuse(const struct ggml_cgraph * cgraph, int node_idx, std::initializer_list<enum ggml_op> ops) {
+    return ggml_can_fuse(cgraph, node_idx, ops.begin(), (int)ops.size());
+}
+
 // expose GGUF internals for test code
 GGML_API size_t gguf_type_size(enum gguf_type type);
 GGML_API struct gguf_context * gguf_init_from_file_impl(FILE * file, struct gguf_init_params params);
diff --git a/ggml/src/ggml-kompute/CMakeLists.txt b/ggml/src/ggml-kompute/CMakeLists.txt
deleted file mode 100644
index c9109d5e8..000000000
--- a/ggml/src/ggml-kompute/CMakeLists.txt
+++ /dev/null
@@ -1,166 +0,0 @@
-
-find_package(Vulkan COMPONENTS glslc REQUIRED)
-find_program(glslc_executable NAMES glslc HINTS Vulkan::glslc)
-
-if (NOT glslc_executable)
-    message(FATAL_ERROR "glslc not found")
-endif()
-
-ggml_add_backend_library(ggml-kompute
-                         ggml-kompute.cpp
-                         ../../include/ggml-kompute.h
-                        )
-
-target_link_libraries(ggml-kompute PRIVATE ggml-base kompute)
-target_include_directories(ggml-kompute PRIVATE ${CMAKE_CURRENT_BINARY_DIR})
-
-add_compile_definitions(VULKAN_HPP_DISPATCH_LOADER_DYNAMIC=1)
-
-function(compile_shader)
-    set(options)
-    set(oneValueArgs)
-    set(multiValueArgs SOURCES)
-    cmake_parse_arguments(compile_shader "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
-    foreach(source ${compile_shader_SOURCES})
-        get_filename_component(filename ${source} NAME)
-        set(spv_file ${filename}.spv)
-        add_custom_command(
-            OUTPUT ${spv_file}
-            DEPENDS ${CMAKE_CURRENT_SOURCE_DIR}/${source}
-            ${CMAKE_CURRENT_SOURCE_DIR}/kompute-shaders/common.comp
-            ${CMAKE_CURRENT_SOURCE_DIR}/kompute-shaders/op_getrows.comp
-            ${CMAKE_CURRENT_SOURCE_DIR}/kompute-shaders/op_mul_mv_q_n_pre.comp
-            ${CMAKE_CURRENT_SOURCE_DIR}/kompute-shaders/op_mul_mv_q_n.comp
-            COMMAND ${glslc_executable} --target-env=vulkan1.2 -o ${spv_file} ${CMAKE_CURRENT_SOURCE_DIR}/${source}
-            COMMENT "Compiling ${source} to ${spv_file}"
-            )
-
-        get_filename_component(RAW_FILE_NAME ${spv_file} NAME)
-        set(FILE_NAME "shader${RAW_FILE_NAME}")
-        string(REPLACE ".comp.spv" ".h" HEADER_FILE ${FILE_NAME})
-        string(TOUPPER ${HEADER_FILE} HEADER_FILE_DEFINE)
-        string(REPLACE "." "_" HEADER_FILE_DEFINE "${HEADER_FILE_DEFINE}")
-        set(OUTPUT_HEADER_FILE "${HEADER_FILE}")
-        message(STATUS "${HEADER_FILE} generating ${HEADER_FILE_DEFINE}")
-        if(CMAKE_GENERATOR MATCHES "Visual Studio")
-            add_custom_command(
-                OUTPUT ${OUTPUT_HEADER_FILE}
-                COMMAND ${CMAKE_COMMAND} -E echo "/*THIS FILE HAS BEEN AUTOMATICALLY GENERATED - DO NOT EDIT*/" > ${OUTPUT_HEADER_FILE}
-                COMMAND ${CMAKE_COMMAND} -E echo \"\#ifndef ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
-                COMMAND ${CMAKE_COMMAND} -E echo \"\#define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
-                COMMAND ${CMAKE_COMMAND} -E echo "namespace kp {" >> ${OUTPUT_HEADER_FILE}
-                COMMAND ${CMAKE_COMMAND} -E echo "namespace shader_data {" >> ${OUTPUT_HEADER_FILE}
-                COMMAND ${CMAKE_BINARY_DIR}/bin/$<CONFIG>/xxd -i ${RAW_FILE_NAME} >> ${OUTPUT_HEADER_FILE}
-                COMMAND ${CMAKE_COMMAND} -E echo "}}" >> ${OUTPUT_HEADER_FILE}
-                COMMAND ${CMAKE_COMMAND} -E echo \"\#endif // define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
-                DEPENDS ${spv_file} xxd
-                COMMENT "Converting to hpp: ${FILE_NAME} ${CMAKE_BINARY_DIR}/bin/$<CONFIG>/xxd"
-                )
-        else()
-            add_custom_command(
-                OUTPUT ${OUTPUT_HEADER_FILE}
-                COMMAND ${CMAKE_COMMAND} -E echo "/*THIS FILE HAS BEEN AUTOMATICALLY GENERATED - DO NOT EDIT*/" > ${OUTPUT_HEADER_FILE}
-                COMMAND ${CMAKE_COMMAND} -E echo \"\#ifndef ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
-                COMMAND ${CMAKE_COMMAND} -E echo \"\#define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
-                COMMAND ${CMAKE_COMMAND} -E echo "namespace kp {" >> ${OUTPUT_HEADER_FILE}
-                COMMAND ${CMAKE_COMMAND} -E echo "namespace shader_data {" >> ${OUTPUT_HEADER_FILE}
-                COMMAND ${CMAKE_BINARY_DIR}/bin/xxd -i ${RAW_FILE_NAME} >> ${OUTPUT_HEADER_FILE}
-                COMMAND ${CMAKE_COMMAND} -E echo "}}" >> ${OUTPUT_HEADER_FILE}
-                COMMAND ${CMAKE_COMMAND} -E echo \"\#endif // define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
-                DEPENDS ${spv_file} xxd
-                COMMENT "Converting to hpp: ${FILE_NAME} ${CMAKE_BINARY_DIR}/bin/xxd"
-                )
-        endif()
-    endforeach()
-endfunction()
-
-if (EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/kompute/CMakeLists.txt")
-    message(STATUS "Kompute found")
-    set(KOMPUTE_OPT_LOG_LEVEL Error CACHE STRING "Kompute log level")
-    add_subdirectory(kompute)
-
-    # Compile our shaders
-    compile_shader(SOURCES
-        kompute-shaders/op_scale.comp
-        kompute-shaders/op_scale_8.comp
-        kompute-shaders/op_add.comp
-        kompute-shaders/op_addrow.comp
-        kompute-shaders/op_mul.comp
-        kompute-shaders/op_silu.comp
-        kompute-shaders/op_relu.comp
-        kompute-shaders/op_gelu.comp
-        kompute-shaders/op_softmax.comp
-        kompute-shaders/op_norm.comp
-        kompute-shaders/op_rmsnorm.comp
-        kompute-shaders/op_diagmask.comp
-        kompute-shaders/op_mul_mat_mat_f32.comp
-        kompute-shaders/op_mul_mat_f16.comp
-        kompute-shaders/op_mul_mat_q8_0.comp
-        kompute-shaders/op_mul_mat_q4_0.comp
-        kompute-shaders/op_mul_mat_q4_1.comp
-        kompute-shaders/op_mul_mat_q4_k.comp
-        kompute-shaders/op_mul_mat_q6_k.comp
-        kompute-shaders/op_getrows_f32.comp
-        kompute-shaders/op_getrows_f16.comp
-        kompute-shaders/op_getrows_q4_0.comp
-        kompute-shaders/op_getrows_q4_1.comp
-        kompute-shaders/op_getrows_q6_k.comp
-        kompute-shaders/op_rope_norm_f16.comp
-        kompute-shaders/op_rope_norm_f32.comp
-        kompute-shaders/op_rope_neox_f16.comp
-        kompute-shaders/op_rope_neox_f32.comp
-        kompute-shaders/op_cpy_f16_f16.comp
-        kompute-shaders/op_cpy_f16_f32.comp
-        kompute-shaders/op_cpy_f32_f16.comp
-        kompute-shaders/op_cpy_f32_f32.comp
-    )
-
-    # Create a custom target for our generated shaders
-    add_custom_target(generated_shaders DEPENDS
-        shaderop_scale.h
-        shaderop_scale_8.h
-        shaderop_add.h
-        shaderop_addrow.h
-        shaderop_mul.h
-        shaderop_silu.h
-        shaderop_relu.h
-        shaderop_gelu.h
-        shaderop_softmax.h
-        shaderop_norm.h
-        shaderop_rmsnorm.h
-        shaderop_diagmask.h
-        shaderop_mul_mat_mat_f32.h
-        shaderop_mul_mat_f16.h
-        shaderop_mul_mat_q8_0.h
-        shaderop_mul_mat_q4_0.h
-        shaderop_mul_mat_q4_1.h
-        shaderop_mul_mat_q4_k.h
-        shaderop_mul_mat_q6_k.h
-        shaderop_getrows_f32.h
-        shaderop_getrows_f16.h
-        shaderop_getrows_q4_0.h
-        shaderop_getrows_q4_1.h
-        shaderop_getrows_q6_k.h
-        shaderop_rope_norm_f16.h
-        shaderop_rope_norm_f32.h
-        shaderop_rope_neox_f16.h
-        shaderop_rope_neox_f32.h
-        shaderop_cpy_f16_f16.h
-        shaderop_cpy_f16_f32.h
-        shaderop_cpy_f32_f16.h
-        shaderop_cpy_f32_f32.h
-    )
-
-    # Create a custom command that depends on the generated_shaders
-    add_custom_command(
-        OUTPUT ${CMAKE_CURRENT_BINARY_DIR}/ggml-kompute.stamp
-        COMMAND ${CMAKE_COMMAND} -E touch ${CMAKE_CURRENT_BINARY_DIR}/ggml-kompute.stamp
-        DEPENDS generated_shaders
-        COMMENT "Ensuring shaders are generated before compiling ggml-kompute.cpp"
-    )
-
-    # Add the stamp to the main sources to ensure dependency tracking
-    target_sources(ggml-kompute PRIVATE ${CMAKE_CURRENT_BINARY_DIR}/ggml-kompute.stamp)
-else()
-    message(WARNING "Kompute not found")
-endif()
diff --git a/ggml/src/ggml-kompute/ggml-kompute.cpp b/ggml/src/ggml-kompute/ggml-kompute.cpp
deleted file mode 100644
index 505792271..000000000
--- a/ggml/src/ggml-kompute/ggml-kompute.cpp
+++ /dev/null
@@ -1,2251 +0,0 @@
-#include "ggml-impl.h"
-#include "ggml-backend.h"
-#include "ggml-backend-impl.h"
-#include "ggml-kompute.h"
-
-// These are generated at build time by cmake custom command
-#include "shaderop_scale.h"
-#include "shaderop_scale_8.h"
-#include "shaderop_add.h"
-#include "shaderop_addrow.h"
-#include "shaderop_mul.h"
-#include "shaderop_silu.h"
-#include "shaderop_relu.h"
-#include "shaderop_gelu.h"
-#include "shaderop_softmax.h"
-#include "shaderop_norm.h"
-#include "shaderop_rmsnorm.h"
-#include "shaderop_diagmask.h"
-#include "shaderop_mul_mat_f16.h"
-#include "shaderop_mul_mat_q8_0.h"
-#include "shaderop_mul_mat_q4_0.h"
-#include "shaderop_mul_mat_q4_1.h"
-#include "shaderop_mul_mat_q4_k.h"
-#include "shaderop_mul_mat_q6_k.h"
-#include "shaderop_mul_mat_mat_f32.h"
-#include "shaderop_getrows_f32.h"
-#include "shaderop_getrows_f16.h"
-#include "shaderop_getrows_q4_0.h"
-#include "shaderop_getrows_q4_1.h"
-#include "shaderop_getrows_q6_k.h"
-#include "shaderop_rope_norm_f16.h"
-#include "shaderop_rope_norm_f32.h"
-#include "shaderop_rope_neox_f16.h"
-#include "shaderop_rope_neox_f32.h"
-#include "shaderop_cpy_f16_f16.h"
-#include "shaderop_cpy_f16_f32.h"
-#include "shaderop_cpy_f32_f16.h"
-#include "shaderop_cpy_f32_f32.h"
-
-#include <algorithm>
-#include <array>
-#include <cassert>
-#include <cstdint>
-#include <cstdio>
-#include <cstring>
-#include <iostream>
-#include <memory>
-#include <mutex>
-#include <stdexcept>
-#include <string>
-#include <unordered_map>
-#include <utility>
-#include <vector>
-
-#include <kompute/Kompute.hpp>
-#include <vulkan/vulkan.hpp>
-
-#ifdef __linux__
-#include <cstdlib> // for setenv
-#endif
-
-#define QK4_0 32
-#define QR4_0 2
-#define QK4_1 32
-#define QK_NL 16
-
-typedef ggml_fp16_t half;
-
-static std::string ggml_kompute_format_name(int device) {
-    return "Kompute" + std::to_string(device);
-}
-
-struct ggml_kompute_context {
-    int device;
-    std::string name;
-    std::shared_ptr<vk::DescriptorPool> pool;
-
-    ggml_kompute_context(int device)
-        : device(device), name(ggml_kompute_format_name(device)) {}
-};
-
-// FIXME: It would be good to consolidate the kompute manager and the kompute context into one object
-// and consolidate the init functions and simplify object lifetime management. As it currently stands,
-// we *have* to have the kompute manager no matter what for device discovery, but the kompute context
-// is only created when a device is set and vulkan is explicitly turned on.
-static ggml_kompute_context *s_kompute_context = nullptr;
-
-class kompute_manager {
-    kp::Manager *s_mgr = nullptr;
-
-public:
-    kp::Manager *operator()() {
-        if (s_mgr && !s_mgr->hasInstance()) {
-            destroy();
-        }
-        if (!s_mgr) {
-            s_mgr = new kp::Manager;
-        }
-        return s_mgr;
-    }
-
-    void destroy() {
-        delete s_mgr;
-        s_mgr = nullptr;
-    }
-};
-
-static kompute_manager komputeManager;
-
-struct ggml_vk_memory {
-    void *data = nullptr;
-    size_t size = 0;
-    vk::DeviceMemory *primaryMemory = nullptr;
-    vk::Buffer *primaryBuffer = nullptr;
-    vk::DeviceMemory *stagingMemory = nullptr;
-    vk::Buffer *stagingBuffer = nullptr;
-};
-
-#ifdef __linux__
-__attribute__((constructor))
-static void enable_sam() {
-    setenv("RADV_PERFTEST", "sam", false);
-}
-#endif
-
-static bool ggml_vk_checkPhysicalDeviceFeatures(vk::PhysicalDevice physical_device) {
-    vk::PhysicalDeviceFeatures availableFeatures;
-    physical_device.getFeatures(&availableFeatures);
-
-    if (!availableFeatures.shaderInt16)
-        return false;
-
-    vk::PhysicalDeviceVulkan11Features availableFeatures11;
-    vk::PhysicalDeviceVulkan12Features availableFeatures12;
-
-    availableFeatures11.pNext = &availableFeatures12;
-    availableFeatures12.pNext = nullptr;
-
-    vk::PhysicalDeviceFeatures2 features2;
-    features2.pNext = &availableFeatures11;
-
-    physical_device.getFeatures2(&features2);
-
-    if (!availableFeatures11.uniformAndStorageBuffer16BitAccess ||
-        !availableFeatures11.storageBuffer16BitAccess) {
-        return false;
-    }
-
-    if (!availableFeatures12.storageBuffer8BitAccess ||
-        !availableFeatures12.uniformAndStorageBuffer8BitAccess ||
-        !availableFeatures12.shaderFloat16 ||
-        !availableFeatures12.shaderInt8) {
-        return false;
-    }
-
-    return true;
-}
-
-static const char * ggml_vk_getVendorName(uint32_t vendorID) {
-    switch (vendorID) {
-        case 0x10DE:
-            return "nvidia";
-        case 0x1002:
-            return "amd";
-        case 0x8086:
-            return "intel";
-        default:
-            return "unknown";
-    }
-}
-
-static std::vector<ggml_vk_device> ggml_vk_available_devices_internal(size_t memoryRequired) {
-    std::vector<ggml_vk_device> results;
-    if (!komputeManager()->hasVulkan() || !komputeManager()->hasInstance())
-        return results;
-
-    std::vector<vk::PhysicalDevice> physical_devices;
-    try {
-        physical_devices = komputeManager()->listDevices();
-    } catch (vk::SystemError & err) {
-        std::cerr << __func__ << ": ignoring Vulkan exception: " << err.what() << "\n";
-        return results;
-    }
-
-    uint32_t deviceCount = physical_devices.size();
-    if (deviceCount == 0)
-        return results;
-
-    std::unordered_map<std::string, size_t> count_by_name;
-
-    for (uint32_t i = 0; i < deviceCount; i++) {
-        const auto & physical_device = physical_devices[i];
-
-        VkPhysicalDeviceProperties dev_props = physical_device.getProperties();
-        VkPhysicalDeviceMemoryProperties memoryProperties = physical_device.getMemoryProperties();
-        const uint32_t major = VK_VERSION_MAJOR(dev_props.apiVersion);
-        const uint32_t minor = VK_VERSION_MINOR(dev_props.apiVersion);
-        if (major < 1 || minor < 2)
-            continue;
-
-        if (!ggml_vk_checkPhysicalDeviceFeatures(physical_device))
-            continue;
-
-        size_t heapSize = 0;
-        for (uint32_t j = 0; j < memoryProperties.memoryHeapCount; ++j) {
-            VkMemoryHeap heap = memoryProperties.memoryHeaps[j];
-            if (heap.flags & VK_MEMORY_HEAP_DEVICE_LOCAL_BIT) {
-                heapSize = heap.size;
-                break;
-            }
-        }
-
-        if (heapSize < memoryRequired)
-            continue;
-
-        auto ext_props = physical_device.enumerateDeviceExtensionProperties();
-        bool has_maintenance4 = false;
-
-        // Check if maintenance4 is supported
-        for (const auto & properties : ext_props) {
-            if (strcmp("VK_KHR_maintenance4", properties.extensionName) == 0) {
-                has_maintenance4 = true;
-            }
-        }
-
-        vk::PhysicalDeviceSubgroupProperties subgroup_props;
-        vk::PhysicalDeviceProperties2 dev_props2;
-        vk::PhysicalDeviceMaintenance3Properties dev_props3;
-        vk::PhysicalDeviceMaintenance4Properties dev_props4;
-        dev_props2.pNext = &dev_props3;
-        dev_props3.pNext = &subgroup_props;
-        if (has_maintenance4) {
-            subgroup_props.pNext = &dev_props4;
-        }
-        physical_device.getProperties2(&dev_props2);
-
-        if (subgroup_props.subgroupSize < 32)
-            continue;
-
-        ggml_vk_device d;
-        d.index = i;
-        d.type = dev_props.deviceType;
-        d.heapSize = heapSize;
-        d.vendor = strdup(ggml_vk_getVendorName(dev_props.vendorID));
-        d.subgroupSize = subgroup_props.subgroupSize;
-        d.bufferAlignment = dev_props.limits.minStorageBufferOffsetAlignment;
-
-        if (has_maintenance4) {
-            d.maxAlloc = std::min(dev_props3.maxMemoryAllocationSize, dev_props4.maxBufferSize);
-        } else {
-            d.maxAlloc = dev_props3.maxMemoryAllocationSize;
-        }
-
-        std::string name(dev_props.deviceName);
-        size_t n_idx = ++count_by_name[name];
-        if (n_idx > 1) {
-            name += " (" + std::to_string(n_idx) + ")";
-        }
-        d.name = strdup(name.c_str());
-
-        results.push_back(d);
-    }
-
-    std::stable_sort(results.begin(), results.end(),
-        [](const ggml_vk_device& lhs, const ggml_vk_device& rhs) -> bool {
-            if (lhs.type != rhs.type) {
-                if (lhs.type == VK_PHYSICAL_DEVICE_TYPE_DISCRETE_GPU) return true;
-                if (rhs.type == VK_PHYSICAL_DEVICE_TYPE_DISCRETE_GPU) return false;
-
-                if (lhs.type == VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU) return true;
-                if (rhs.type == VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU) return false;
-            }
-            return lhs.heapSize < rhs.heapSize;
-        }
-    );
-
-    return results;
-}
-
-static std::vector<ggml_vk_device>& ggml_vk_available_devices() {
-    static std::vector<ggml_vk_device> devices = ggml_vk_available_devices_internal(0);
-    return devices;
-}
-
-static void ggml_vk_filterByVendor(std::vector<ggml_vk_device>& devices, const std::string& targetVendor) {
-    devices.erase(
-        std::remove_if(devices.begin(), devices.end(),
-            [&targetVendor](const ggml_vk_device& device) {
-                return device.vendor != targetVendor;
-            }),
-        devices.end()
-    );
-}
-
-static void ggml_vk_filterByName(std::vector<ggml_vk_device>& devices, const std::string& targetName) {
-    devices.erase(
-        std::remove_if(devices.begin(), devices.end(),
-            [&targetName](const ggml_vk_device& device) {
-                return device.name != targetName;
-            }),
-        devices.end()
-    );
-}
-
-static bool ggml_vk_get_device(ggml_vk_device * device, size_t memoryRequired, const std::string & name) {
-    if (name.empty())
-        return false;
-
-    auto devices = ggml_vk_available_devices_internal(memoryRequired);
-    if (name == "amd" || name == "nvidia" || name == "intel") {
-        ggml_vk_filterByVendor(devices, name);
-    } else if (name != "gpu") {
-        ggml_vk_filterByName(devices, name);
-    }
-
-    if (devices.empty())
-        return false;
-
-    *device = devices.front();
-    return true;
-}
-
-bool ggml_vk_get_device(ggml_vk_device * device, size_t memoryRequired, const char * name) {
-    return ggml_vk_get_device(device, memoryRequired, std::string(name));
-}
-
-bool ggml_vk_has_vulkan() {
-    return komputeManager()->hasVulkan();
-}
-
-bool ggml_vk_has_device() {
-    return komputeManager()->hasDevice();
-}
-
-ggml_vk_device ggml_vk_current_device() {
-    if (!komputeManager()->hasDevice())
-        return ggml_vk_device();
-
-    auto devices = ggml_vk_available_devices();
-    ggml_vk_filterByName(devices, komputeManager()->physicalDevice()->getProperties().deviceName.data());
-    GGML_ASSERT(!devices.empty());
-    return devices.front();
-}
-
-static
-void ggml_vk_allocate_descriptor_pool(struct ggml_kompute_context * ctx, size_t size) {
-    std::vector<vk::DescriptorPoolSize> descriptorPoolSizes = {
-        vk::DescriptorPoolSize(
-          vk::DescriptorType::eStorageBuffer,
-          4 * size // Descriptor count is number of possible tensors to pass into an algorithm
-          )
-    };
-
-    vk::DescriptorPoolCreateInfo descriptorPoolInfo(
-      vk::DescriptorPoolCreateFlags(),
-      size, // Max sets
-      static_cast<uint32_t>(descriptorPoolSizes.size()),
-      descriptorPoolSizes.data());
-
-    ctx->pool = std::make_shared<vk::DescriptorPool>();
-    vk::Result r = komputeManager()->device()->createDescriptorPool(
-      &descriptorPoolInfo, nullptr, ctx->pool.get());
-    if (r != vk::Result::eSuccess)
-        std::cerr << "Error allocating descriptor pool" << vk::to_string(r);
-}
-
-static
-void ggml_vk_free_descriptor_pool(struct ggml_kompute_context * ctx) {
-    if (ctx->pool) {
-        komputeManager()->device()->destroy(
-          *ctx->pool,
-          (vk::Optional<const vk::AllocationCallbacks>)nullptr);
-        ctx->pool = nullptr;
-    }
-}
-
-static
-vk::Buffer *ggml_vk_allocate_buffer(size_t size) {
-    vk::BufferCreateInfo bufferCreateInfo;
-    bufferCreateInfo.size = size;
-    bufferCreateInfo.usage = vk::BufferUsageFlagBits::eStorageBuffer |
-                             vk::BufferUsageFlagBits::eTransferSrc |
-                             vk::BufferUsageFlagBits::eTransferDst;
-    bufferCreateInfo.sharingMode = vk::SharingMode::eExclusive;
-
-    vk::Buffer *vkBuffer = new vk::Buffer;
-    vk::Result r = komputeManager()->device()->createBuffer(&bufferCreateInfo, nullptr, vkBuffer);
-    if (r != vk::Result::eSuccess)
-        std::cerr << "Error allocating buffer " << vk::to_string(r) << std::endl;
-    return vkBuffer;
-}
-
-static
-vk::DeviceMemory *ggml_vk_allocate(size_t size, vk::MemoryPropertyFlags flags, vk::MemoryRequirements requirements, bool *isHostVisible) {
-
-    uint32_t memoryTypeIndex = -1;
-    bool memoryTypeIndexFound = false;
-    vk::PhysicalDeviceMemoryProperties memoryProperties = komputeManager()->physicalDevice()->getMemoryProperties();
-    for (uint32_t i = 0; i < memoryProperties.memoryTypeCount; i++) {
-        const vk::MemoryType &memoryType = memoryProperties.memoryTypes[i];
-        const vk::MemoryHeap &memoryHeap = memoryProperties.memoryHeaps[memoryType.heapIndex];
-        if (memoryHeap.size < size) {
-            continue;
-        }
-
-        if (requirements.memoryTypeBits & (1 << i)) {
-            if (((memoryProperties.memoryTypes[i]).propertyFlags &
-                 flags) == flags) {
-                memoryTypeIndex = i;
-                memoryTypeIndexFound = true;
-                if (isHostVisible && (memoryProperties.memoryTypes[i].propertyFlags & vk::MemoryPropertyFlagBits::eHostVisible)) {
-                    *isHostVisible = true;
-                }
-                break;
-            }
-        }
-    }
-    if (!memoryTypeIndexFound) {
-        throw std::runtime_error(
-          "Memory type index for buffer creation not found");
-    }
-
-    vk::MemoryAllocateInfo allocInfo;
-    allocInfo.allocationSize = size;
-    allocInfo.memoryTypeIndex = memoryTypeIndex;
-    vk::DeviceMemory *vkDeviceMemory =  new vk::DeviceMemory;
-    vk::Result r = komputeManager()->device()->allocateMemory(&allocInfo, nullptr, vkDeviceMemory);
-    if (r != vk::Result::eSuccess) {
-        std::cerr << "Error allocating memory " << vk::to_string(r) << std::endl;
-        throw std::runtime_error("Error allocating vulkan memory.");
-    }
-    return vkDeviceMemory;
-}
-
-static size_t ggml_vk_aligned_offset(ggml_backend_buffer_t buffer, size_t offset) {
-    size_t minStorageBufferOffsetAlignment = ggml_backend_buffer_get_alignment(buffer);
-
-    // If offset is already aligned, return it directly
-    if (offset % minStorageBufferOffsetAlignment == 0) {
-        return offset;
-    }
-
-    // Otherwise, return the largest multiple of minStorageBufferOffsetAlignment less than offset
-    return (offset / minStorageBufferOffsetAlignment) * minStorageBufferOffsetAlignment;
-}
-
-static ggml_vk_memory ggml_vk_allocate(size_t size) {
-    ggml_vk_memory memory;
-    bool isHostVisible = false;
-    {
-        memory.primaryBuffer = ggml_vk_allocate_buffer(size);
-        vk::MemoryRequirements memoryRequirements = komputeManager()->device()->getBufferMemoryRequirements(*memory.primaryBuffer);
-        vk::MemoryPropertyFlags memoryPropertyFlags = vk::MemoryPropertyFlagBits::eDeviceLocal;
-        memory.primaryMemory = ggml_vk_allocate(size, memoryPropertyFlags, memoryRequirements, &isHostVisible);
-        komputeManager()->device()->bindBufferMemory(*memory.primaryBuffer, *memory.primaryMemory, 0);
-        if (isHostVisible) {
-            vk::Result r = komputeManager()->device()->mapMemory(*memory.primaryMemory, 0, size, vk::MemoryMapFlags(), &memory.data);
-            if (r != vk::Result::eSuccess)
-                std::cerr << "Error mapping memory" << vk::to_string(r);
-        }
-    }
-
-    if (!isHostVisible) {
-        memory.stagingBuffer = ggml_vk_allocate_buffer(size);
-        vk::MemoryRequirements memoryRequirements = komputeManager()->device()->getBufferMemoryRequirements(*memory.stagingBuffer);
-        vk::MemoryPropertyFlags memoryPropertyFlags = vk::MemoryPropertyFlagBits::eHostVisible |
-                                                      vk::MemoryPropertyFlagBits::eHostCoherent |
-                                                      vk::MemoryPropertyFlagBits::eHostCached;
-        memory.stagingMemory = ggml_vk_allocate(size, memoryPropertyFlags, memoryRequirements, &isHostVisible);
-        komputeManager()->device()->bindBufferMemory(*memory.stagingBuffer, *memory.stagingMemory, 0);
-        vk::Result r = komputeManager()->device()->mapMemory(*memory.stagingMemory, 0, size, vk::MemoryMapFlags(), &memory.data);
-        if (r != vk::Result::eSuccess)
-            std::cerr << "Error mapping memory" << vk::to_string(r);
-    }
-
-    memory.size = size;
-    return memory;
-}
-
-static void ggml_vk_free_memory(ggml_vk_memory &memory)
-{
-    komputeManager()->device()->destroy(
-      *memory.primaryBuffer,
-      (vk::Optional<const vk::AllocationCallbacks>)nullptr);
-    if (memory.stagingBuffer) {
-        komputeManager()->device()->destroy(
-          *memory.stagingBuffer,
-          (vk::Optional<const vk::AllocationCallbacks>)nullptr);
-    }
-    komputeManager()->device()->freeMemory(
-      *memory.primaryMemory,
-      (vk::Optional<const vk::AllocationCallbacks>)nullptr);
-    if (memory.stagingMemory) {
-        komputeManager()->device()->freeMemory(
-          *memory.stagingMemory,
-          (vk::Optional<const vk::AllocationCallbacks>)nullptr);
-    }
-}
-
-static const char * ggml_backend_kompute_buffer_type_get_name(ggml_backend_buffer_type_t buft);
-
-static
-ggml_vk_memory * ggml_vk_find_tensor(const struct ggml_tensor * t, uint64_t & offset) {
-    ggml_backend_buffer_t buffer = t->view_src ? t->view_src->buffer : t->buffer;
-
-    // compatibility with ggml-backend
-    GGML_ASSERT(buffer && buffer->buft->iface.get_name == ggml_backend_kompute_buffer_type_get_name);
-
-    ggml_vk_memory * buf_ctx = static_cast<ggml_vk_memory *>(buffer->context);
-
-    const intptr_t ioffs = intptr_t(t->data) - intptr_t(buf_ctx->data);
-
-    GGML_ASSERT(ioffs >= 0 && ioffs + int64_t(ggml_nbytes(t)) <= int64_t(buffer->size));
-
-    offset = uint64_t(ioffs);
-    return buf_ctx;
-}
-
-static
-const std::shared_ptr<kp::Tensor> ggml_vk_get_tensor(const struct ggml_tensor * t, uint32_t * alignedOffset = nullptr) {
-    uint64_t originalOffset = 0;
-    auto * res = ggml_vk_find_tensor(t, originalOffset);
-    if (!res) {
-        static std::shared_ptr<kp::Tensor> nullTensor = nullptr;
-        return nullTensor;
-    }
-
-    // Create a tensor whose memory will be composed of our buffers at the correct offset
-    const size_t nelements = ggml_nelements(t);
-    size_t nbytes = ggml_nbytes(t);
-
-    size_t vulkanOffset = ggml_vk_aligned_offset(t->buffer, originalOffset);
-    if (alignedOffset) {
-        *alignedOffset = originalOffset - vulkanOffset;
-        nbytes += *alignedOffset;
-    }
-
-    return komputeManager()->tensor(
-        t->data,
-        nelements,
-        nbytes, kp::Tensor::TensorDataTypes::eFloat,
-        res->primaryMemory, res->primaryBuffer,
-        res->stagingMemory, res->stagingBuffer,
-        vulkanOffset);
-}
-
-static std::vector<uint32_t> getSpirvShader(const unsigned char* rawData, size_t size) {
-    if (size % sizeof(uint32_t) != 0) {
-        throw std::runtime_error("Invalid size: must be divisible by sizeof(uint32_t)");
-    }
-
-    const uint32_t* data_ptr = reinterpret_cast<const uint32_t*>(rawData);
-    size_t count = size / sizeof(uint32_t);
-    return std::vector<uint32_t>(data_ptr, data_ptr + count);
-}
-
-inline static
-uint32_t safe_divide(uint32_t a, uint32_t b) {
-    if (b <= 1) {
-        return a;
-    }
-    if ((a % b) != 0) {
-        fprintf(stderr, "((%u %% %u) == %u) != 0\n", a, b, a % b);
-        GGML_ABORT("safe_divide result would've had remainder");
-    }
-    return a / b;
-}
-
-static void ggml_vk_add(
-    kp::Sequence& seq,
-    const std::shared_ptr<kp::Tensor>& inA,
-    const std::shared_ptr<kp::Tensor>& inB,
-    const std::shared_ptr<kp::Tensor>& out,
-    uint32_t inAOff, uint32_t inBOff, uint32_t outOff,
-    int32_t ne00, int32_t ne01, int32_t ne02, int32_t ne03,
-    int32_t nb00, int32_t nb01, int32_t nb02, int32_t nb03,
-    int32_t ne10, int32_t ne11, int32_t ne12, int32_t ne13,
-    int32_t nb10, int32_t nb11, int32_t nb12, int32_t nb13,
-    int32_t ne0,
-    int32_t nb0,  int32_t nb1,  int32_t nb2,  int32_t nb3
-) {
-    const static auto spirv = getSpirvShader(kp::shader_data::op_add_comp_spv,
-        kp::shader_data::op_add_comp_spv_len);
-
-    struct PushConstants {
-        uint32_t inAOff, inBOff, outOff;
-        int32_t ne00;
-        int32_t nb00, nb01, nb02, nb03;
-        int32_t ne10, ne11, ne12, ne13;
-        int32_t nb10, nb11, nb12, nb13;
-        int32_t ne0;
-        int32_t nb0, nb1, nb2, nb3;
-    } const pushConsts {
-        safe_divide(inAOff, 4), safe_divide(inBOff, 4), safe_divide(outOff, 4),
-        ne00,
-        nb00, nb01, nb02, nb03,
-        ne10, ne11, ne12, ne13,
-        nb10, nb11, nb12, nb13,
-        ne0,
-        nb0, nb1, nb2, nb3
-    };
-
-    std::shared_ptr<kp::Algorithm> s_algo = nullptr;
-    if (!komputeManager()->hasAlgorithm(__func__)) {
-        s_algo = komputeManager()->algorithm<float, PushConstants>(__func__, s_kompute_context->pool.get(), {inA, inB, out}, spirv, {unsigned(ne01), unsigned(ne02), unsigned(ne03)}, {}, {pushConsts});
-    } else {
-        s_algo = komputeManager()->getAlgorithm(__func__);
-        s_algo->setTensors({inA, inB, out});
-        s_algo->setWorkgroup({unsigned(ne01), unsigned(ne02), unsigned(ne03)});
-        s_algo->setPushConstants<PushConstants>({pushConsts});
-        s_algo->updateDescriptors(s_kompute_context->pool.get());
-    }
-    seq.record<kp::OpAlgoDispatch>(s_algo);
-}
-
-static void ggml_vk_addrow(kp::Sequence& seq,
-                 const std::shared_ptr<kp::Tensor>& inA,
-                 const std::shared_ptr<kp::Tensor>& inB,
-                 const std::shared_ptr<kp::Tensor>& out,
-                 uint32_t inAOff, uint32_t inBOff, uint32_t outOff,
-                 uint32_t size, uint32_t row = 0) {
-
-    const static auto spirv = getSpirvShader(kp::shader_data::op_addrow_comp_spv,
-        kp::shader_data::op_addrow_comp_spv_len);
-
-    struct PushConstants {
-        uint32_t inAOff, inBOff, outOff;
-        uint32_t row;
-    } const pushConsts {
-        safe_divide(inAOff, 4), safe_divide(inBOff, 4), safe_divide(outOff, 4),
-        row
-    };
-
-    std::shared_ptr<kp::Algorithm> s_algo = nullptr;
-    if (!komputeManager()->hasAlgorithm(__func__))
-        s_algo = komputeManager()->algorithm<float, PushConstants>(__func__, s_kompute_context->pool.get(), {inA, inB, out}, spirv, {size}, {}, {pushConsts});
-    else {
-        s_algo = komputeManager()->getAlgorithm(__func__);
-        s_algo->setTensors({inA, inB, out});
-        s_algo->setWorkgroup({size});
-        s_algo->setPushConstants<PushConstants>({pushConsts});
-        s_algo->updateDescriptors(s_kompute_context->pool.get());
-    }
-    seq.record<kp::OpAlgoDispatch>(s_algo);
-}
-
-static void ggml_vk_mul(
-    kp::Sequence& seq,
-    const std::shared_ptr<kp::Tensor>& inA,
-    const std::shared_ptr<kp::Tensor>& inB,
-    const std::shared_ptr<kp::Tensor>& out,
-    uint32_t inAOff, uint32_t inBOff, uint32_t outOff,
-    int32_t ne00, int32_t ne01, int32_t ne02, int32_t ne03,
-    int32_t nb00, int32_t nb01, int32_t nb02, int32_t nb03,
-    int32_t ne10, int32_t ne11, int32_t ne12, int32_t ne13,
-    int32_t nb10, int32_t nb11, int32_t nb12, int32_t nb13,
-    int32_t ne0,
-    int32_t nb0,  int32_t nb1,  int32_t nb2,  int32_t nb3
-) {
-    const static auto spirv = getSpirvShader(kp::shader_data::op_mul_comp_spv,
-        kp::shader_data::op_mul_comp_spv_len);
-
-    struct PushConstants {
-        uint32_t inAOff, inBOff, outOff;
-        int32_t ne00;
-        int32_t nb00, nb01, nb02, nb03;
-        int32_t ne10, ne11, ne12, ne13;
-        int32_t nb10, nb11, nb12, nb13;
-        int32_t ne0;
-        int32_t nb0, nb1, nb2, nb3;
-    } const pushConsts {
-        safe_divide(inAOff, 4), safe_divide(inBOff, 4), safe_divide(outOff, 4),
-        ne00,
-        nb00, nb01, nb02, nb03,
-        ne10, ne11, ne12, ne13,
-        nb10, nb11, nb12, nb13,
-        ne0,
-        nb0, nb1, nb2, nb3
-    };
-
-    std::shared_ptr<kp::Algorithm> s_algo = nullptr;
-    if (!komputeManager()->hasAlgorithm(__func__)) {
-        s_algo = komputeManager()->algorithm<float, PushConstants>(__func__, s_kompute_context->pool.get(), {inA, inB, out}, spirv, {unsigned(ne01), unsigned(ne02), unsigned(ne03)}, {}, {pushConsts});
-    } else {
-        s_algo = komputeManager()->getAlgorithm(__func__);
-        s_algo->setTensors({inA, inB, out});
-        s_algo->setWorkgroup({unsigned(ne01), unsigned(ne02), unsigned(ne03)});
-        s_algo->setPushConstants<PushConstants>({pushConsts});
-        s_algo->updateDescriptors(s_kompute_context->pool.get());
-    }
-    seq.record<kp::OpAlgoDispatch>(s_algo);
-}
-
-static void ggml_vk_scale(kp::Sequence& seq,
-                   const std::shared_ptr<kp::Tensor>& in,
-                   const std::shared_ptr<kp::Tensor>& out,
-                   uint32_t inOff, uint32_t outOff,
-                   uint32_t size, float scale) {
-    const static auto spirv_1 = getSpirvShader(
-        kp::shader_data::op_scale_comp_spv, kp::shader_data::op_scale_comp_spv_len
-    );
-    const static auto spirv_8 = getSpirvShader(
-        kp::shader_data::op_scale_8_comp_spv, kp::shader_data::op_scale_8_comp_spv_len
-    );
-
-    struct PushConstants {
-        uint32_t inOff, outOff;
-        float scale;
-    } const pushConsts {
-        safe_divide(inOff, 4), safe_divide(outOff, 4),
-        scale
-    };
-
-    const auto * spirv = &spirv_1;
-    std::string name(__func__);
-    if (size % 8 == 0) {
-        size /= 8;
-        name += "_8";
-        spirv = &spirv_8;
-    }
-
-    std::shared_ptr<kp::Algorithm> s_algo = nullptr;
-    if (!komputeManager()->hasAlgorithm(name)) {
-        s_algo = komputeManager()->algorithm<float, PushConstants>(name, s_kompute_context->pool.get(), {in, out}, *spirv, {size}, {}, {pushConsts});
-    } else {
-        s_algo = komputeManager()->getAlgorithm(name);
-        s_algo->setTensors({in, out});
-        s_algo->setWorkgroup({size});
-        s_algo->setPushConstants<PushConstants>({pushConsts});
-        s_algo->updateDescriptors(s_kompute_context->pool.get());
-    }
-    seq.record<kp::OpAlgoDispatch>(s_algo);
-}
-
-static void ggml_vk_xxlu(
-    const std::vector<uint32_t>& spirv, const char * suffix, kp::Sequence& seq,
-    const std::shared_ptr<kp::Tensor>& in,
-    const std::shared_ptr<kp::Tensor>& out,
-    uint32_t inOff, uint32_t outOff,
-    uint32_t size
-) {
-    struct PushConstants {
-        uint32_t inOff, outOff;
-    } const pushConsts {
-        safe_divide(inOff, 4), safe_divide(outOff, 4),
-    };
-
-    auto name = std::string(__func__) + "_" + suffix;
-    std::shared_ptr<kp::Algorithm> s_algo = nullptr;
-    if (!komputeManager()->hasAlgorithm(name)) {
-        s_algo = komputeManager()->algorithm<float, PushConstants>(name, s_kompute_context->pool.get(), {in, out}, spirv, {size}, {}, {pushConsts});
-    } else {
-        s_algo = komputeManager()->getAlgorithm(name);
-        s_algo->setTensors({in, out});
-        s_algo->setWorkgroup({size});
-        s_algo->setPushConstants<PushConstants>({pushConsts});
-        s_algo->updateDescriptors(s_kompute_context->pool.get());
-    }
-    seq.record<kp::OpAlgoDispatch>(s_algo);
-}
-
-template <typename... Args>
-static void ggml_vk_silu(Args&&... args) {
-    const static auto spirv = getSpirvShader(kp::shader_data::op_silu_comp_spv,
-        kp::shader_data::op_silu_comp_spv_len);
-
-    ggml_vk_xxlu(spirv, "silu", std::forward<Args>(args)...);
-}
-
-template <typename... Args>
-static void ggml_vk_relu(Args&&... args) {
-    const static auto spirv = getSpirvShader(kp::shader_data::op_relu_comp_spv,
-        kp::shader_data::op_relu_comp_spv_len);
-
-    ggml_vk_xxlu(spirv, "relu", std::forward<Args>(args)...);
-}
-
-template <typename... Args>
-static void ggml_vk_gelu(Args&&... args) {
-    const static auto spirv = getSpirvShader(kp::shader_data::op_gelu_comp_spv,
-        kp::shader_data::op_gelu_comp_spv_len);
-
-    ggml_vk_xxlu(spirv, "gelu", std::forward<Args>(args)...);
-}
-
-static void ggml_vk_soft_max(
-    kp::Sequence& seq,
-    const std::shared_ptr<kp::Tensor>& inA,
-    const std::shared_ptr<kp::Tensor>& inB,
-    const std::shared_ptr<kp::Tensor>& out,
-    uint32_t inAOff, uint32_t inBOff, uint32_t outOff,
-    int32_t ne00, int32_t ne01, int32_t ne02, uint32_t ne03,
-    float scale, float max_bias, float m0, float m1,
-    uint32_t n_head_log2
-) {
-    const static auto spirv = getSpirvShader(kp::shader_data::op_softmax_comp_spv,
-        kp::shader_data::op_softmax_comp_spv_len);
-
-    struct PushConstants {
-        uint32_t inAOff, inBOff, outOff;
-        int32_t ne00, ne01, ne02;
-        float scale, max_bias, m0, m1;
-        uint32_t n_head_log2;
-        int32_t mask;
-    } pushConsts {
-        safe_divide(inAOff, 4), safe_divide(inBOff, 4), safe_divide(outOff, 4),
-        ne00, ne01, ne02,
-        scale, max_bias, m0, m1,
-        n_head_log2,
-        bool(inB)
-    };
-
-    auto & inB_ = inB ? inB : inA;
-
-    std::shared_ptr<kp::Algorithm> s_algo = nullptr;
-    if (!komputeManager()->hasAlgorithm(__func__)) {
-        // FIXME: The softmax kernel needs to be fixed to use the subgroupsize which can vary by device
-        const uint32_t local_x = 32;
-        s_algo = komputeManager()->algorithm<uint32_t, PushConstants>(__func__, s_kompute_context->pool.get(), {inA, inB_, out}, spirv, {unsigned(ne01), unsigned(ne02), unsigned(ne03)}, {local_x}, {pushConsts});
-    } else {
-        s_algo = komputeManager()->getAlgorithm(__func__);
-        s_algo->setTensors({inA, inB_, out});
-        s_algo->setWorkgroup({unsigned(ne01), unsigned(ne02), unsigned(ne03)});
-        s_algo->setPushConstants<PushConstants>({pushConsts});
-        s_algo->updateDescriptors(s_kompute_context->pool.get());
-    }
-    seq.record<kp::OpAlgoDispatch>(s_algo);
-}
-
-static void ggml_vk_norm_(
-    const std::vector<uint32_t>& spirv, const char * suffix, kp::Sequence& seq,
-    const std::shared_ptr<kp::Tensor>& in,
-    const std::shared_ptr<kp::Tensor>& out,
-    uint32_t inOff, uint32_t outOff,
-    int32_t ne00, int32_t nb01,
-    int32_t nrows, float epsilon
-) {
-    GGML_ASSERT(nb01%sizeof(float) == 0);
-    GGML_ASSERT(ne00%sizeof(float) == 0);
-
-    struct PushConstants {
-        uint32_t inOff, outOff;
-        uint32_t ne00, nb01;
-        float eps;
-    } pushConsts {
-        safe_divide(inOff, 4), safe_divide(outOff, 4),
-        (uint32_t)ne00, (uint32_t)nb01, epsilon
-    };
-
-    auto name = std::string(__func__) + "_" + suffix;
-    std::shared_ptr<kp::Algorithm> s_algo = nullptr;
-    if (!komputeManager()->hasAlgorithm(name)) {
-        s_algo = komputeManager()->algorithm<float, PushConstants>(name, s_kompute_context->pool.get(), {in, out}, spirv, {(uint32_t)nrows}, {}, {pushConsts});
-    } else {
-        s_algo = komputeManager()->getAlgorithm(name);
-        s_algo->setTensors({in, out});
-        s_algo->setWorkgroup({(uint32_t)nrows});
-        s_algo->setPushConstants<PushConstants>({pushConsts});
-        s_algo->updateDescriptors(s_kompute_context->pool.get());
-    }
-    seq.record<kp::OpAlgoDispatch>(s_algo);
-}
-
-template <typename... Args>
-static void ggml_vk_norm(Args&&... args) {
-    const static auto spirv = getSpirvShader(kp::shader_data::op_norm_comp_spv,
-        kp::shader_data::op_norm_comp_spv_len);
-
-    ggml_vk_norm_(spirv, "norm", std::forward<Args>(args)...);
-}
-
-template <typename... Args>
-static void ggml_vk_rms_norm(Args&&... args) {
-    const static auto spirv = getSpirvShader(kp::shader_data::op_rmsnorm_comp_spv,
-        kp::shader_data::op_rmsnorm_comp_spv_len);
-
-    ggml_vk_norm_(spirv, "rms", std::forward<Args>(args)...);
-}
-
-static void ggml_vk_diag_mask_inf(kp::Sequence& seq,
-                           const std::shared_ptr<kp::Tensor>& in,
-                           const std::shared_ptr<kp::Tensor>& out,
-                           uint32_t inOff, uint32_t outOff,
-                           uint32_t n_past,
-                           int32_t ne00, int32_t ne01, int32_t ne02) {
-    const static auto spirv = getSpirvShader(kp::shader_data::op_diagmask_comp_spv,
-        kp::shader_data::op_diagmask_comp_spv_len);
-
-    struct PushConstants {
-        uint32_t inOff, outOff;
-        uint32_t n_past;
-        int32_t ne00, ne01;
-    } pushConsts {
-        safe_divide(inOff, 4), safe_divide(outOff, 4),
-        n_past,
-        ne00, ne01
-    };
-
-    std::shared_ptr<kp::Algorithm> s_algo = nullptr;
-    if (!komputeManager()->hasAlgorithm(__func__))
-        s_algo = komputeManager()->algorithm<float, PushConstants>(__func__, s_kompute_context->pool.get(), {in, out}, spirv, {unsigned(ne00), unsigned(ne01), unsigned(ne02)}, {}, {pushConsts});
-    else {
-        s_algo = komputeManager()->getAlgorithm(__func__);
-        s_algo->setTensors({in, out});
-        s_algo->setWorkgroup({unsigned(ne00), unsigned(ne01), unsigned(ne02)});
-        s_algo->setPushConstants<PushConstants>({pushConsts});
-        s_algo->updateDescriptors(s_kompute_context->pool.get());
-    }
-    seq.record<kp::OpAlgoDispatch>(s_algo);
-}
-
-static void ggml_vk_mul_mat_f16(
-    kp::Sequence& seq,
-    const std::shared_ptr<kp::Tensor>& inA,
-    const std::shared_ptr<kp::Tensor>& inB,
-    const std::shared_ptr<kp::Tensor>& out,
-    uint32_t inAOff, uint32_t inBOff, uint32_t outOff,
-    int32_t ne00, int32_t ne01, int32_t ne02,
-    uint32_t nb00, uint32_t nb01, uint32_t nb02, uint32_t nb03,
-    int32_t ne10, int32_t ne11, int32_t ne12, int32_t ne13,
-    uint32_t nb10, uint32_t nb11, uint32_t nb12, uint32_t nb13,
-    int32_t ne0, int32_t ne1,
-    uint32_t r2, uint32_t r3
-) {
-    const static auto spirv = getSpirvShader(kp::shader_data::op_mul_mat_f16_comp_spv,
-        kp::shader_data::op_mul_mat_f16_comp_spv_len);
-
-    struct PushConstants {
-        uint32_t inAOff, inBOff, outOff;
-        int32_t ne00, ne01, ne02;
-        uint32_t nb00, nb01, nb02, nb03;
-        int32_t ne10, ne11, ne12;
-        uint32_t nb10, nb11, nb12, nb13;
-        int32_t ne0, ne1;
-        uint32_t r2, r3;
-    } pushConsts {
-        safe_divide(inAOff, 2), safe_divide(inBOff, 4), safe_divide(outOff, 4),
-        ne00, ne01, ne02,
-        nb00, nb01, nb02, nb03,
-        ne10, ne11, ne12,
-        nb10, nb11, nb12, nb13,
-        ne0, ne1,
-        r2, r3
-    };
-
-    const unsigned ny = unsigned((ne11 + 4 - 1)/4);
-
-    std::shared_ptr<kp::Algorithm> s_algo = nullptr;
-    if (!komputeManager()->hasAlgorithm(__func__)) {
-        const uint32_t local_x = ggml_vk_current_device().subgroupSize * 2;
-        s_algo = komputeManager()->algorithm<uint32_t, PushConstants>(__func__, s_kompute_context->pool.get(), {inA, inB, out}, spirv, {unsigned(ne01), ny, unsigned(ne12*ne13)}, {local_x}, {pushConsts});
-    } else {
-        s_algo = komputeManager()->getAlgorithm(__func__);
-        s_algo->setTensors({inA, inB, out});
-        s_algo->setWorkgroup({unsigned(ne01), ny, unsigned(ne12*ne13)});
-        s_algo->setPushConstants<PushConstants>({pushConsts});
-        s_algo->updateDescriptors(s_kompute_context->pool.get());
-    }
-    seq.record<kp::OpAlgoDispatch>(s_algo);
-}
-
-static void ggml_vk_mul_mat_mat_f32(kp::Sequence& seq,
-                         const std::shared_ptr<kp::Tensor>& inA,
-                         const std::shared_ptr<kp::Tensor>& inB,
-                         const std::shared_ptr<kp::Tensor>& out,
-                         uint32_t inAOff, uint32_t inBOff, uint32_t outOff,
-                         int32_t ne00, int32_t ne01, int32_t ne02,
-                         uint32_t nb01, uint32_t nb02,
-                         int32_t ne11, int32_t ne12,
-                         uint32_t nb11, uint32_t nb12,
-                         uint32_t nb1, uint32_t nb2) {
-    const static auto spirv = getSpirvShader(kp::shader_data::op_mul_mat_mat_f32_comp_spv,
-        kp::shader_data::op_mul_mat_mat_f32_comp_spv_len);
-
-    struct PushConstants {
-        uint32_t inAOff, inBOff, outOff;
-        int32_t ne00, ne01, ne02, ne11, ne12;
-        uint32_t nb01, nb02;
-        uint32_t nb11, nb12;
-        uint32_t nb1, nb2;
-    } pushConsts {
-        safe_divide(inAOff, 4), safe_divide(inBOff, 4), safe_divide(outOff, 4),
-        ne00, ne01, ne02, ne11, ne12,
-        nb01, nb02, nb11, nb12,
-        nb1, nb2
-    };
-
-    const uint32_t local_x = ggml_vk_current_device().subgroupSize;
-    std::shared_ptr<kp::Algorithm> s_algo = nullptr;
-    if (!komputeManager()->hasAlgorithm(__func__)) {
-        s_algo = komputeManager()->algorithm<uint32_t, PushConstants>(__func__, s_kompute_context->pool.get(),
-        {inA, inB, out}, spirv,
-        {unsigned(ne01),
-         unsigned(ne11),
-         unsigned(std::max(ne12, ne02))
-         },
-        {local_x},
-        {pushConsts});
-    } else {
-        s_algo = komputeManager()->getAlgorithm(__func__);
-        s_algo->setTensors({inA, inB, out});
-        s_algo->setWorkgroup({unsigned(ne01),
-                              unsigned(ne11),
-                              unsigned(std::max(ne12, ne02)),
-                              });
-        s_algo->setPushConstants<PushConstants>({pushConsts});
-        s_algo->updateDescriptors(s_kompute_context->pool.get());
-    }
-    seq.record<kp::OpAlgoDispatch>(s_algo);
-}
-
-static void ggml_vk_mul_mat_impl(
-    const std::vector<uint32_t>& spirv, const char * suffix, uint32_t block_size, kp::Sequence& seq,
-    const std::shared_ptr<kp::Tensor>& inA,
-    const std::shared_ptr<kp::Tensor>& inB,
-    const std::shared_ptr<kp::Tensor>& out,
-    uint32_t inAOff, uint32_t inBOff, uint32_t outOff,
-    int32_t ne00, int32_t ne01, int32_t ne02,
-    int32_t ne10, int32_t ne11, int32_t ne12, int32_t ne13,
-    int32_t ne0, int32_t ne1,
-    uint32_t nb01, uint32_t nb02, uint32_t nb03,
-    uint32_t nb11, uint32_t nb12, uint32_t nb13,
-    uint32_t r2, uint32_t r3
-) {
-    struct PushConstants {
-        uint32_t inAOff, inBOff, outOff;
-        int32_t ne00, ne01, ne02;
-        int32_t ne10, ne12;
-        int32_t ne0, ne1;
-        uint32_t nb01, nb02, nb03;
-        uint32_t nb11, nb12, nb13;
-        uint32_t r2, r3;
-    } pushConsts {
-        safe_divide(inAOff, block_size), safe_divide(inBOff, 4), safe_divide(outOff, 4),
-        ne00, ne01, ne02,
-        ne10, ne12,
-        ne0, ne1,
-        nb01, nb02, nb03,
-        nb11, nb12, nb13,
-        r2, r3
-    };
-
-    auto name = std::string(__func__) + "_" + suffix;
-    std::shared_ptr<kp::Algorithm> s_algo = nullptr;
-    if (!komputeManager()->hasAlgorithm(name)) {
-        const uint32_t local_x = (ggml_vk_current_device().subgroupSize * 2) / 8;
-        s_algo = komputeManager()->algorithm<uint32_t, PushConstants>(name, s_kompute_context->pool.get(), {inA, inB, out}, spirv, {unsigned((ne01 + 7)/8), unsigned(ne11), unsigned(ne12*ne13)}, {local_x}, {pushConsts});
-    } else {
-        s_algo = komputeManager()->getAlgorithm(name);
-        s_algo->setTensors({inA, inB, out});
-        s_algo->setWorkgroup({unsigned((ne01 + 7)/8), unsigned(ne11), unsigned(ne12*ne13)});
-        s_algo->setPushConstants<PushConstants>({pushConsts});
-        s_algo->updateDescriptors(s_kompute_context->pool.get());
-    }
-    seq.record<kp::OpAlgoDispatch>(s_algo);
-}
-
-template <typename... Args>
-static void ggml_vk_mul_mat_q4_0(Args&&... args) {
-    const static auto spirv = getSpirvShader(kp::shader_data::op_mul_mat_q4_0_comp_spv,
-        kp::shader_data::op_mul_mat_q4_0_comp_spv_len);
-
-    ggml_vk_mul_mat_impl(spirv, "q4_0", 1/*We access blocks unaligned*/, std::forward<Args>(args)...);
-}
-
-template <typename... Args>
-static void ggml_vk_mul_mat_q4_1(Args&&... args) {
-    const static auto spirv = getSpirvShader(kp::shader_data::op_mul_mat_q4_1_comp_spv,
-        kp::shader_data::op_mul_mat_q4_1_comp_spv_len);
-
-    ggml_vk_mul_mat_impl(spirv, "q4_1", 1/*We access blocks unaligned*/, std::forward<Args>(args)...);
-}
-
-template <typename... Args>
-static void ggml_vk_mul_mat_q8_0(Args&&... args) {
-    const static auto spirv = getSpirvShader(kp::shader_data::op_mul_mat_q8_0_comp_spv,
-        kp::shader_data::op_mul_mat_q8_0_comp_spv_len);
-
-    ggml_vk_mul_mat_impl(spirv, "q8_0", 1/*We access blocks unaligned*/, std::forward<Args>(args)...);
-}
-
-static void ggml_vk_mul_mat_q4_k(
-    kp::Sequence& seq,
-    const std::shared_ptr<kp::Tensor>& inA,
-    const std::shared_ptr<kp::Tensor>& inB,
-    const std::shared_ptr<kp::Tensor>& out,
-    uint32_t inAOff, uint32_t inBOff, uint32_t outOff,
-    int32_t ne00, int32_t ne01, int32_t ne02,
-    int32_t ne10, int32_t ne11, int32_t ne12, int32_t ne13,
-    int32_t ne0, int32_t ne1,
-    uint32_t nb01, uint32_t nb02, uint32_t nb03,
-    uint32_t nb11, uint32_t nb12, uint32_t nb13,
-    uint32_t r2, uint32_t r3
-) {
-    const static auto spirv = getSpirvShader(kp::shader_data::op_mul_mat_q4_k_comp_spv,
-        kp::shader_data::op_mul_mat_q4_k_comp_spv_len);
-
-    struct PushConstants {
-        uint32_t inAOff, inBOff, outOff;
-        int32_t ne00, ne10, ne0, ne1, ne01, ne02, ne12;
-        uint32_t nb01, nb02, nb03, nb11, nb12, nb13;
-        uint32_t r2, r3;
-    } pushConsts {
-        inAOff, safe_divide(inBOff, 4), safe_divide(outOff, 4),
-        ne00, ne10, ne0, ne1, ne01, ne02, ne12,
-        nb01, nb02, nb03, nb11, nb12, nb13,
-        r2, r3
-    };
-
-    std::shared_ptr<kp::Algorithm> s_algo = nullptr;
-    if (!komputeManager()->hasAlgorithm(__func__)) {
-        s_algo = komputeManager()->algorithm<uint32_t, PushConstants>(__func__, s_kompute_context->pool.get(), {inA, inB, out}, spirv, {unsigned((ne01 + 3)/4), unsigned(ne11), unsigned(ne12) * unsigned(ne13)}, {}, {pushConsts});
-    } else {
-        s_algo = komputeManager()->getAlgorithm(__func__);
-        s_algo->setTensors({inA, inB, out});
-        s_algo->setWorkgroup({unsigned((ne01 + 3)/4), unsigned(ne11), unsigned(ne12) * unsigned(ne13)});
-        s_algo->setPushConstants<PushConstants>({pushConsts});
-        s_algo->updateDescriptors(s_kompute_context->pool.get());
-    }
-    seq.record<kp::OpAlgoDispatch>(s_algo);
-}
-
-static void ggml_vk_mul_mat_q6_k(
-    kp::Sequence& seq,
-    const std::shared_ptr<kp::Tensor>& inA,
-    const std::shared_ptr<kp::Tensor>& inB,
-    const std::shared_ptr<kp::Tensor>& out,
-    uint32_t inAOff, uint32_t inBOff, uint32_t outOff,
-    int32_t ne00, int32_t ne01, int32_t ne02,
-    int32_t ne10, int32_t ne11, int32_t ne12, int32_t ne13,
-    int32_t ne0, int32_t ne1,
-    uint32_t nb01, uint32_t nb02, uint32_t nb03,
-    uint32_t nb11, uint32_t nb12, uint32_t nb13,
-    uint32_t r2, uint32_t r3
-) {
-    const static auto spirv = getSpirvShader(kp::shader_data::op_mul_mat_q6_k_comp_spv,
-        kp::shader_data::op_mul_mat_q6_k_comp_spv_len);
-
-    struct PushConstants {
-        uint32_t inAOff, inBOff, outOff;
-        int32_t ne00, ne10, ne0, ne1, ne01, ne02, ne12;
-        uint32_t nb01, nb02, nb03, nb11, nb12, nb13;
-        uint32_t r2, r3;
-    } pushConsts {
-        inAOff, safe_divide(inBOff, 4), safe_divide(outOff, 4),
-        ne00, ne10, ne0, ne1, ne01, ne02, ne12,
-        nb01, nb02, nb03, nb11, nb12, nb13,
-        r2, r3
-    };
-
-    std::shared_ptr<kp::Algorithm> s_algo = nullptr;
-    if (!komputeManager()->hasAlgorithm(__func__)) {
-        const uint32_t local_x = 2;
-        const uint32_t local_y = ggml_vk_current_device().subgroupSize;
-        s_algo = komputeManager()->algorithm<uint32_t, PushConstants>(__func__, s_kompute_context->pool.get(), {inA, inB, out}, spirv, {unsigned((ne01 + 1)/2), unsigned(ne11), unsigned(ne12)*unsigned(ne13)}, {local_x, local_y}, {pushConsts});
-    } else {
-        s_algo = komputeManager()->getAlgorithm(__func__);
-        s_algo->setTensors({inA, inB, out});
-        s_algo->setWorkgroup({unsigned((ne01 + 1)/2), unsigned(ne11), unsigned(ne12)*unsigned(ne13)});
-        s_algo->setPushConstants<PushConstants>({pushConsts});
-        s_algo->updateDescriptors(s_kompute_context->pool.get());
-    }
-    seq.record<kp::OpAlgoDispatch>(s_algo);
-}
-
-static void ggml_vk_get_rows(
-    const std::vector<uint32_t>& spirv,
-    const char * suffix,
-    unsigned element_size, unsigned qk,
-    kp::Sequence& seq,
-    const std::shared_ptr<kp::Tensor>& inA,
-    const std::shared_ptr<kp::Tensor>& inB,
-    const std::shared_ptr<kp::Tensor>& out,
-    uint32_t inAOff, uint32_t inBOff, uint32_t outOff,
-    int32_t ne00, int32_t nb01, int32_t nb1,
-    uint32_t size
-) {
-    GGML_ASSERT(nb01%element_size == 0);
-    GGML_ASSERT(nb1%sizeof(float) == 0);
-    if (qk) GGML_ASSERT(ne00%qk == 0);
-
-    struct PushConstants {
-        uint32_t inAOff, inBOff, outOff;
-        int32_t ne00, nb01, nb1;
-    } pushConsts {
-        safe_divide(inAOff, element_size), safe_divide(inBOff, 4), safe_divide(outOff, 4),
-        ne00, nb01, nb1
-    };
-
-    auto name = std::string(__func__) + "_" + suffix;
-    std::shared_ptr<kp::Algorithm> s_algo = nullptr;
-    if (!komputeManager()->hasAlgorithm(name)) {
-        s_algo = komputeManager()->algorithm<float, PushConstants>(name, s_kompute_context->pool.get(), {inA, inB, out}, spirv, {size}, {}, {pushConsts});
-    } else {
-        s_algo = komputeManager()->getAlgorithm(name);
-        s_algo->setTensors({inA, inB, out});
-        s_algo->setWorkgroup({size});
-        s_algo->setPushConstants<PushConstants>({pushConsts});
-        s_algo->updateDescriptors(s_kompute_context->pool.get());
-    }
-    seq.record<kp::OpAlgoDispatch>(s_algo);
-}
-
-template <typename... Args>
-static void ggml_vk_get_rows_f32(Args&&... args) {
-    const static auto spirv = getSpirvShader(kp::shader_data::op_getrows_f32_comp_spv,
-        kp::shader_data::op_getrows_f32_comp_spv_len);
-
-    ggml_vk_get_rows(spirv, "f32", sizeof(float), 0, std::forward<Args>(args)...);
-}
-
-template <typename... Args>
-static void ggml_vk_get_rows_f16(Args&&... args) {
-    const static auto spirv = getSpirvShader(kp::shader_data::op_getrows_f16_comp_spv,
-        kp::shader_data::op_getrows_f16_comp_spv_len);
-
-    ggml_vk_get_rows(spirv, "f16", sizeof(half), 0, std::forward<Args>(args)...);
-}
-
-template <typename... Args>
-static void ggml_vk_get_rows_q4_0(Args&&... args) {
-    const static auto spirv = getSpirvShader(kp::shader_data::op_getrows_q4_0_comp_spv,
-        kp::shader_data::op_getrows_q4_0_comp_spv_len);
-
-    ggml_vk_get_rows(spirv, "q4_0", 1/*We access blocks unaligned*/, QK4_0, std::forward<Args>(args)...);
-}
-
-template <typename... Args>
-static void ggml_vk_get_rows_q4_1(Args&&... args) {
-    const static auto spirv = getSpirvShader(kp::shader_data::op_getrows_q4_1_comp_spv,
-        kp::shader_data::op_getrows_q4_1_comp_spv_len);
-
-    ggml_vk_get_rows(spirv, "q4_1", 1/*We access blocks unaligned*/, QK4_1, std::forward<Args>(args)...);
-}
-
-template <typename... Args>
-static void ggml_vk_get_rows_q6_k(Args&&... args) {
-    const static auto spirv = getSpirvShader(kp::shader_data::op_getrows_q6_k_comp_spv,
-        kp::shader_data::op_getrows_q6_k_comp_spv_len);
-    ggml_vk_get_rows(spirv, "q6_k", 1/*We access blocks unaligned*/, QK_NL, std::forward<Args>(args)...);
-}
-
-static void ggml_vk_rope(
-    kp::Sequence& seq,
-    const std::shared_ptr<kp::Tensor>& inA,
-    const std::shared_ptr<kp::Tensor>& inB,
-    const std::shared_ptr<kp::Tensor>& inC,
-    const std::shared_ptr<kp::Tensor>& out,
-    uint32_t inAOff, uint32_t inBOff, uint32_t inCOff, uint32_t outOff,
-    ggml_type src0t, int32_t n_dims, int32_t mode, int32_t n_ctx_orig,
-    float freq_base, float freq_scale, bool has_freq_factors, float ext_factor, float attn_factor, float beta_fast, float beta_slow,
-    int32_t ne01, int32_t ne02, int32_t ne03,
-    uint32_t nb00, uint32_t nb01, uint32_t nb02, uint32_t nb03,
-    int32_t ne0,
-    uint32_t nb0, uint32_t nb1, uint32_t nb2, uint32_t nb3
-) {
-    GGML_ASSERT(src0t == GGML_TYPE_F16 || src0t == GGML_TYPE_F32);
-
-    static const auto spirv_norm_f16 = getSpirvShader(
-        kp::shader_data::op_rope_norm_f16_comp_spv, kp::shader_data::op_rope_norm_f16_comp_spv_len
-    );
-    static const auto spirv_norm_f32 = getSpirvShader(
-        kp::shader_data::op_rope_norm_f32_comp_spv, kp::shader_data::op_rope_norm_f32_comp_spv_len
-    );
-    static const auto spirv_neox_f16 = getSpirvShader(
-        kp::shader_data::op_rope_neox_f16_comp_spv, kp::shader_data::op_rope_neox_f16_comp_spv_len
-    );
-    static const auto spirv_neox_f32 = getSpirvShader(
-        kp::shader_data::op_rope_neox_f32_comp_spv, kp::shader_data::op_rope_neox_f32_comp_spv_len
-    );
-
-    int type_size = src0t == GGML_TYPE_F16 ? 2 : 4;
-
-    GGML_ASSERT(nb03 % type_size == 0);
-    GGML_ASSERT(nb02 % type_size == 0);
-    GGML_ASSERT(nb01 % type_size == 0);
-    GGML_ASSERT(nb00 % type_size == 0);
-    GGML_ASSERT(nb3  % type_size == 0);
-    GGML_ASSERT(nb2  % type_size == 0);
-    GGML_ASSERT(nb1  % type_size == 0);
-    GGML_ASSERT(nb0  % type_size == 0);
-
-    struct PushConstants {
-        uint32_t inAOff, inBOff, inCOff, outOff;
-        int32_t n_dims, mode, n_ctx_orig;
-        float freq_base, freq_scale;
-        bool has_freq_factors;
-        float ext_factor, attn_factor, beta_fast, beta_slow;
-        uint32_t nb00, nb01, nb02, nb03;
-        int32_t ne0;
-        uint32_t nb0, nb1, nb2, nb3;
-    } pushConsts {
-        safe_divide(inAOff, type_size), safe_divide(inBOff, 4), safe_divide(inCOff, type_size), safe_divide(outOff, type_size),
-        n_dims, mode, n_ctx_orig,
-        freq_base, freq_scale,
-        has_freq_factors,
-        ext_factor, attn_factor, beta_fast, beta_slow,
-        nb00, nb01, nb02, nb03,
-        ne0,
-        nb0, nb1, nb2, nb3
-    };
-
-    auto & inC_ = inC ? inC : inA;
-    const bool is_neox = mode & GGML_ROPE_TYPE_NEOX;
-    const bool is_f16 = src0t == GGML_TYPE_F16;
-
-    auto name = std::string(__func__) + (is_neox ? "_neox" : "_norm") + (src0t == GGML_TYPE_F16 ? "_f16" : "_f32");
-    std::shared_ptr<kp::Algorithm> s_algo = nullptr;
-    if (!komputeManager()->hasAlgorithm(name)) {
-        auto & spirv = is_neox ? is_f16 ? spirv_neox_f16 : spirv_neox_f32 : is_f16 ? spirv_norm_f16 : spirv_norm_f32;
-        s_algo = komputeManager()->algorithm<float, PushConstants>(
-            name, s_kompute_context->pool.get(), {inA, inB, inC_, out}, spirv,
-            {unsigned(ne01), unsigned(ne02), unsigned(ne03)}, {}, {pushConsts}
-        );
-    } else {
-        s_algo = komputeManager()->getAlgorithm(name);
-        s_algo->setTensors({inA, inB, inC_, out});
-        s_algo->setWorkgroup({unsigned(ne01), unsigned(ne02), unsigned(ne03)});
-        s_algo->setPushConstants<PushConstants>({pushConsts});
-        s_algo->updateDescriptors(s_kompute_context->pool.get());
-    }
-    seq.record<kp::OpAlgoDispatch>(s_algo);
-}
-
-static void ggml_vk_cpy(
-    const std::vector<uint32_t>& spirv,
-    uint32_t in_element_size, uint32_t out_element_size,
-    kp::Sequence& seq,
-    const std::shared_ptr<kp::Tensor>& in,
-    const std::shared_ptr<kp::Tensor>& out,
-    uint32_t inOff, uint32_t outOff,
-    int32_t ne00, int32_t ne01, int32_t ne02, int32_t ne03,
-    uint32_t nb00, uint32_t nb01, uint32_t nb02, uint32_t nb03,
-    int32_t ne0, int32_t ne1, int32_t ne2,
-    uint32_t nb0, uint32_t nb1, uint32_t nb2, uint32_t nb3
-) {
-    struct PushConstants {
-        uint32_t inOff, outOff;
-        int32_t ne00, ne01, ne02;
-        uint32_t nb00, nb01, nb02, nb03;
-        int32_t ne0, ne1, ne2;
-        uint32_t nb0, nb1, nb2, nb3;
-    } pushConsts {
-        safe_divide(inOff, in_element_size), safe_divide(outOff, out_element_size),
-        ne00, ne01, ne02,
-        nb00, nb01, nb02, nb03,
-        ne0, ne1, ne2,
-        nb0, nb1, nb2, nb3
-    };
-
-    std::string name = std::string(__func__)
-                       + "_i_" + std::to_string(in_element_size)
-                       + "_o_" + std::to_string(out_element_size);
-    std::shared_ptr<kp::Algorithm> s_algo = nullptr;
-    if (!komputeManager()->hasAlgorithm(name))
-        s_algo = komputeManager()->algorithm<float, PushConstants>(name, s_kompute_context->pool.get(), {in, out}, spirv, {unsigned(ne01), unsigned(ne02), unsigned(ne03)}, {}, {pushConsts});
-    else {
-        s_algo = komputeManager()->getAlgorithm(name);
-        s_algo->setTensors({in, out});
-        s_algo->setWorkgroup({unsigned(ne01), unsigned(ne02), unsigned(ne03)});
-        s_algo->setPushConstants<PushConstants>({pushConsts});
-        s_algo->updateDescriptors(s_kompute_context->pool.get());
-    }
-    seq.record<kp::OpAlgoDispatch>(s_algo);
-}
-
-template <typename... Args>
-static void ggml_vk_cpy_f32_f16(Args&&... args) {
-    const static auto spirv = getSpirvShader(kp::shader_data::op_cpy_f32_f16_comp_spv,
-        kp::shader_data::op_cpy_f32_f16_comp_spv_len);
-    ggml_vk_cpy(spirv, 4, 2, std::forward<Args>(args)...);
-}
-
-template <typename... Args>
-static void ggml_vk_cpy_f32_f32(Args&&... args) {
-    const static auto spirv = getSpirvShader(kp::shader_data::op_cpy_f32_f32_comp_spv,
-        kp::shader_data::op_cpy_f32_f32_comp_spv_len);
-    ggml_vk_cpy(spirv, 4, 4, std::forward<Args>(args)...);
-}
-
-template <typename... Args>
-static void ggml_vk_cpy_f16_f16(Args&&... args) {
-    const static auto spirv = getSpirvShader(kp::shader_data::op_cpy_f16_f16_comp_spv,
-        kp::shader_data::op_cpy_f16_f16_comp_spv_len);
-    ggml_vk_cpy(spirv, 2, 2, std::forward<Args>(args)...);
-}
-
-template <typename... Args>
-static void ggml_vk_cpy_f16_f32(Args&&... args) {
-    const static auto spirv = getSpirvShader(kp::shader_data::op_cpy_f16_f32_comp_spv,
-        kp::shader_data::op_cpy_f16_f32_comp_spv_len);
-    ggml_vk_cpy(spirv, 2, 4, std::forward<Args>(args)...);
-}
-
-static bool ggml_backend_kompute_device_supports_op(ggml_backend_dev_t dev, const struct ggml_tensor * op) {
-    int64_t n = ggml_nelements(op);
-    switch (op->op) {
-        case GGML_OP_UNARY:
-            if (n % 4 != 0) return false;
-            switch (ggml_get_unary_op(op)) {
-                case GGML_UNARY_OP_GELU:
-                    if (n % 8 != 0) return false;
-                    // fall through
-                case GGML_UNARY_OP_RELU:
-                case GGML_UNARY_OP_SILU:
-                    return ggml_is_contiguous(op->src[0]);
-                default:
-                    ;
-            }
-            break;
-        case GGML_OP_NONE:
-        case GGML_OP_RESHAPE:
-        case GGML_OP_VIEW:
-        case GGML_OP_TRANSPOSE:
-        case GGML_OP_PERMUTE:
-        case GGML_OP_ADD:
-        case GGML_OP_MUL:
-        case GGML_OP_SCALE:
-        case GGML_OP_SOFT_MAX:
-        case GGML_OP_RMS_NORM:
-        case GGML_OP_NORM:
-            return true;
-        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 true;
-            }
-        case GGML_OP_DUP:
-        case GGML_OP_CPY:
-        case GGML_OP_CONT:
-            switch (op->src[0]->type) {
-                case GGML_TYPE_F32:
-                case GGML_TYPE_F16:
-                    break;
-                default:
-                    return false;
-            }
-            switch (op->type) {
-                case GGML_TYPE_F32:
-                case GGML_TYPE_F16:
-                    break;
-                default:
-                    return false;
-            }
-            return true;
-        case GGML_OP_DIAG_MASK_INF:
-            return op->ne[3] == 1;
-        case GGML_OP_GET_ROWS:
-            switch (op->src[0]->type) {
-                case GGML_TYPE_F32:
-                case GGML_TYPE_F16:
-                case GGML_TYPE_Q4_0:
-                case GGML_TYPE_Q4_1:
-                case GGML_TYPE_Q6_K:
-                    return op->ne[2] == 1 && op->ne[3] == 1;
-                default:
-                    ;
-            }
-            return false;
-        case GGML_OP_MUL_MAT:
-            if (op->src[1]->type != GGML_TYPE_F32 || ggml_is_transposed(op->src[0]) || ggml_is_transposed(op->src[1]))
-                return false;
-
-            switch (op->src[0]->type) {
-                case GGML_TYPE_F32:
-                    return op->ne[3] == 1;
-                case GGML_TYPE_Q6_K:
-                case GGML_TYPE_F16:
-                case GGML_TYPE_Q8_0:
-                case GGML_TYPE_Q4_0:
-                case GGML_TYPE_Q4_1:
-                case GGML_TYPE_Q4_K:
-                    return true;
-                default:
-                    ;
-            }
-        default:
-            ;
-    }
-    return false;
-
-    GGML_UNUSED(dev);
-}
-
-static void ggml_vk_graph_compute(struct ggml_kompute_context * ctx, struct ggml_cgraph * gf) {
-    const int n_seq = 8;
-
-    // FIXME: Figure out if we can somehow optimize the size of the pool... right now we're setting
-    // it to the size of the graph, but I think it can be made smaller?
-    ggml_vk_allocate_descriptor_pool(ctx, gf->n_nodes);
-
-    std::vector<std::shared_ptr<kp::Sequence>> sequences(n_seq);
-
-    for (auto& sequence : sequences) {
-        sequence = komputeManager()->sequence();
-    }
-    for (int seq_idx = 0; seq_idx < n_seq; ++seq_idx) {
-        const int n_nodes_per_seq = (gf->n_nodes + n_seq - 1) / n_seq;
-
-        auto& seq = *sequences[seq_idx];
-
-        const int node_start = (seq_idx + 0) * n_nodes_per_seq;
-        const int node_end   = std::min((seq_idx == n_seq - 1) ? gf->n_nodes : (seq_idx + 1) * n_nodes_per_seq, gf->n_nodes);
-
-        bool any_commands_recorded = false;
-
-        for (int i = node_start; i < node_end; ++i) {
-            struct ggml_tensor * src0 = gf->nodes[i]->src[0];
-            struct ggml_tensor * src1 = gf->nodes[i]->src[1];
-            struct ggml_tensor * src2 = gf->nodes[i]->src[2]; GGML_UNUSED(src2);
-            struct ggml_tensor * dst = gf->nodes[i];
-            GGML_ASSERT(dst->data != nullptr);
-
-            if (ggml_is_empty(dst)) {
-                continue;
-            }
-
-            switch (dst->op) {
-                case GGML_OP_NONE:
-                case GGML_OP_RESHAPE:
-                case GGML_OP_VIEW:
-                case GGML_OP_TRANSPOSE:
-                case GGML_OP_PERMUTE:
-                    continue; // noop -> next node
-                default:
-                    break;
-            }
-
-            any_commands_recorded = true;
-
-            const int32_t ne00 = src0 ? src0->ne[0] : 0;
-            const int32_t ne01 = src0 ? src0->ne[1] : 0;
-            const int32_t ne02 = src0 ? src0->ne[2] : 0;
-            const int32_t ne03 = src0 ? src0->ne[3] : 0;
-
-            const uint32_t nb00 = src0 ? src0->nb[0] : 0;
-            const uint32_t nb01 = src0 ? src0->nb[1] : 0;
-            const uint32_t nb02 = src0 ? src0->nb[2] : 0;
-            const uint32_t nb03 = src0 ? src0->nb[3] : 0;
-
-            const int32_t ne10 = src1 ? src1->ne[0] : 0;
-            const int32_t ne11 = src1 ? src1->ne[1] : 0;
-            const int32_t ne12 = src1 ? src1->ne[2] : 0;
-            const int32_t ne13 = src1 ? src1->ne[3] : 0;
-
-            const uint32_t nb10 = src1 ? src1->nb[0] : 0;
-            const uint32_t nb11 = src1 ? src1->nb[1] : 0;
-            const uint32_t nb12 = src1 ? src1->nb[2] : 0;
-            const uint32_t nb13 = src1 ? src1->nb[3] : 0;
-
-            const int32_t ne0 = dst ? dst->ne[0] : 0;
-            const int32_t ne1 = dst ? dst->ne[1] : 0;
-            const int32_t ne2 = dst ? dst->ne[2] : 0;
-//            const int32_t ne3 = dst ? dst->ne[3] : 0;
-
-            const uint32_t nb0 = dst ? dst->nb[0] : 0;
-            const uint32_t nb1 = dst ? dst->nb[1] : 0;
-            const uint32_t nb2 = dst ? dst->nb[2] : 0;
-            const uint32_t nb3 = dst ? dst->nb[3] : 0;
-
-            const enum ggml_type src0t = src0 ? src0->type : GGML_TYPE_COUNT;
-            const enum ggml_type src1t = src1 ? src1->type : GGML_TYPE_COUNT;
-            const enum ggml_type dstt = dst ? dst->type : GGML_TYPE_COUNT;
-
-            const static std::shared_ptr<kp::Tensor> nullTensor = nullptr;
-            uint32_t off_src0 = 0;
-            uint32_t off_src1 = 0;
-            uint32_t off_src2 = 0;
-            uint32_t off_dst  = 0;
-            const std::shared_ptr<kp::Tensor>& id_src0 = src0 ? ggml_vk_get_tensor(src0, &off_src0) : nullTensor;
-            const std::shared_ptr<kp::Tensor>& id_src1 = src1 ? ggml_vk_get_tensor(src1, &off_src1) : nullTensor;
-            const std::shared_ptr<kp::Tensor>& id_src2 = src2 ? ggml_vk_get_tensor(src2, &off_src2) : nullTensor;
-            const std::shared_ptr<kp::Tensor>& id_dst  = dst  ? ggml_vk_get_tensor(dst,  &off_dst)  : nullTensor;
-
-            switch (dst->op) {
-                case GGML_OP_ADD:
-                    {
-                        if (ggml_nelements(src1) == ne10 && ggml_is_contiguous(src1) && ne00 % 4 == 0 && ne10 % 4 == 0) {
-                            // src1 is a row
-                            ggml_vk_addrow(seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, ggml_nelements(dst)/4, ne00);
-                        } else {
-                            ggml_vk_add(
-                                seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst,
-                                ne00, ne01, ne02, ne03,
-                                nb00, nb01, nb02, nb03,
-                                ne10, ne11, ne12, ne13,
-                                nb10, nb11, nb12, nb13,
-                                ne0,
-                                nb0, nb1, nb2, nb3
-                            );
-                        }
-                    } break;
-                case GGML_OP_MUL:
-                    {
-                        ggml_vk_mul(
-                            seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst,
-                            ne00, ne01, ne02, ne03,
-                            nb00, nb01, nb02, nb03,
-                            ne10, ne11, ne12, ne13,
-                            nb10, nb11, nb12, nb13,
-                            ne0,
-                            nb0, nb1, nb2, nb3
-                        );
-                    } break;
-                case GGML_OP_SCALE:
-                    {
-                        float scale; memcpy(&scale, dst->op_params, sizeof(float));
-
-                        ggml_vk_scale(seq, id_src0, id_dst, off_src0, off_dst, ggml_nelements(dst), scale);
-                    } break;
-                case GGML_OP_UNARY:
-                    {
-                        int64_t n = ggml_nelements(dst);
-                        GGML_ASSERT(n % 4 == 0);
-                        switch (ggml_get_unary_op(gf->nodes[i])) {
-                            case GGML_UNARY_OP_SILU:
-                                {
-                                    ggml_vk_silu(seq, id_src0, id_dst, off_src0, off_dst, n/4);
-                                } break;
-                            case GGML_UNARY_OP_RELU:
-                                {
-                                    ggml_vk_relu(seq, id_src0, id_dst, off_src0, off_dst, n/4);
-                                } break;
-                            case GGML_UNARY_OP_GELU:
-                                {
-                                    GGML_ASSERT(n % 8 == 0);
-                                    ggml_vk_gelu(seq, id_src0, id_dst, off_src0, off_dst, n/8);
-                                } break;
-                            default:
-                                {
-                                    fprintf(stderr, "%s: node %3d, op = %8s not implemented\n", __func__, i, ggml_op_name(dst->op));
-                                    GGML_ABORT("fatal error");
-                                }
-                        }
-                    } break;
-                case GGML_OP_SOFT_MAX:
-                    {
-                        float scale;
-                        float max_bias;
-
-                        memcpy(&scale,    (float *)dst->op_params + 0, sizeof(float));
-                        memcpy(&max_bias, (float *)dst->op_params + 1, sizeof(float));
-
-#pragma message("TODO: add ggml_vk_soft_max() F16 src1 support")
-#pragma message("ref:  https://github.com/ggerganov/llama.cpp/pull/5021")
-                        GGML_ASSERT(!src1 || src1t == GGML_TYPE_F32);
-
-                        const int64_t nrows_x = ggml_nrows(src0);
-                        const int64_t nrows_y = src0->ne[1];
-
-                        const uint32_t n_head      = nrows_x/nrows_y;
-                        const uint32_t n_head_log2 = 1u << (uint32_t) floorf(log2f((float) 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_vk_soft_max(seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, ne00, ne01, ne02, ne03, scale, max_bias, m0, m1, n_head_log2);
-                    } break;
-                case GGML_OP_DIAG_MASK_INF:
-                    {
-                        const int n_past = ((int32_t *)(dst->op_params))[0];
-                        ggml_vk_diag_mask_inf(seq, id_src0, id_dst, off_src0, off_dst, n_past, ne00, ne01, ne02);
-                    } break;
-                case GGML_OP_NORM:
-                    {
-                        float eps;
-                        memcpy(&eps, dst->op_params, sizeof(float));
-                        ggml_vk_norm(seq, id_src0, id_dst, off_src0, off_dst, ne00, nb01, ggml_nrows(src0), eps);
-                    } break;
-                case GGML_OP_RMS_NORM:
-                    {
-                        GGML_ASSERT(ne00 % 4 == 0);
-
-                        float eps;
-                        memcpy(&eps, dst->op_params, sizeof(float));
-                        ggml_vk_rms_norm(seq, id_src0, id_dst, off_src0, off_dst, ne00, nb01, ggml_nrows(src0), eps);
-                    } break;
-                case GGML_OP_MUL_MAT:
-                    {
-                        GGML_ASSERT(ne00 == ne10);
-
-                        GGML_ASSERT(ne12 % ne02 == 0);
-                        GGML_ASSERT(ne13 % ne03 == 0);
-
-                        const uint32_t r2 = ne12/ne02;
-                        const uint32_t r3 = ne13/ne03;
-
-                        if (src1t != GGML_TYPE_F32) {
-                            fprintf(stderr, "%s: %s: Unsupported src1 type: %u/%u\n", __func__, ggml_op_name(dst->op), src0t, src1t);
-                            goto not_implemented;
-                        }
-
-                        if (ggml_is_transposed(src0) ||
-                            ggml_is_transposed(src1)) {
-                            fprintf(stderr, "%s: %s: matmul on tranposed tensor not supported: %u/%u\n", __func__, ggml_op_name(dst->op), src0t, src1t);
-                            goto not_implemented;
-                        }
-
-                        switch (src0t) {
-                            case GGML_TYPE_F32:
-                                ggml_vk_mul_mat_mat_f32(
-                                    seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst,
-                                    ne00, ne01, ne02, nb01, nb02, ne11, ne12, nb11, nb12, nb1, nb2
-                                );
-                                break;
-                            case GGML_TYPE_F16:
-                                ggml_vk_mul_mat_f16(
-                                    seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst,
-                                    ne00, ne01, ne02, nb00, nb01, nb02, nb03,
-                                    ne10, ne11, ne12, ne13, nb10, nb11, nb12, nb13,
-                                    ne0, ne1, r2, r3
-                                );
-                                break;
-                            case GGML_TYPE_Q8_0:
-                                ggml_vk_mul_mat_q8_0(
-                                    seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst,
-                                    ne00, ne01, ne02, ne10, ne11, ne12, ne13, ne0, ne1,
-                                    nb01, nb02, nb03, nb11, nb12, nb13, r2, r3
-                                );
-                                break;
-                            case GGML_TYPE_Q4_0:
-                                ggml_vk_mul_mat_q4_0(
-                                    seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst,
-                                    ne00, ne01, ne02, ne10, ne11, ne12, ne13, ne0, ne1,
-                                    nb01, nb02, nb03, nb11, nb12, nb13, r2, r3
-                                );
-                                break;
-                            case GGML_TYPE_Q4_1:
-                                ggml_vk_mul_mat_q4_1(
-                                    seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst,
-                                    ne00, ne01, ne02, ne10, ne11, ne12, ne13, ne0, ne1,
-                                    nb01, nb02, nb03, nb11, nb12, nb13, r2, r3
-                                );
-                                break;
-                            case GGML_TYPE_Q4_K:
-                                ggml_vk_mul_mat_q4_k(
-                                    seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst,
-                                    ne00, ne01, ne02, ne10, ne11, ne12, ne13, ne0, ne1,
-                                    nb01, nb02, nb03, nb11, nb12, nb13, r2, r3
-                                );
-                                break;
-                            case GGML_TYPE_Q6_K:
-                                ggml_vk_mul_mat_q6_k(
-                                    seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst,
-                                    ne00, ne01, ne02, ne10, ne11, ne12, ne13, ne0, ne1,
-                                    nb01, nb02, nb03, nb11, nb12, nb13, r2, r3
-                                );
-                                break;
-                            default: {
-                                fprintf(stderr, "%s: %s: Unsupported quantization: %u/%u\n", __func__, ggml_op_name(dst->op), src0t, src1t);
-                                goto not_implemented;
-                            }
-                        }
-
-                    } break;
-                case GGML_OP_GET_ROWS:
-                    {
-                        if (src0t == GGML_TYPE_F32) {
-                            ggml_vk_get_rows_f32(seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, ne00, nb01, nb1, ggml_nelements(src1));
-                        } else if (src0t == GGML_TYPE_F16) {
-                            ggml_vk_get_rows_f16(seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, ne00, nb01, nb1, ggml_nelements(src1));
-                        } else if (src0t == GGML_TYPE_Q4_0) {
-                            ggml_vk_get_rows_q4_0(seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, ne00, nb01, nb1, ggml_nelements(src1));
-                        } else if (src0t == GGML_TYPE_Q4_1) {
-                            ggml_vk_get_rows_q4_1(seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, ne00, nb01, nb1, ggml_nelements(src1));
-                        } else if (src0t == GGML_TYPE_Q6_K) {
-                            ggml_vk_get_rows_q6_k(seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, ne00, nb01, nb1, ggml_nelements(src1));
-                        } else {
-                            fprintf(stderr, "%s: %s: Unsupported quantization: %u\n", __func__, ggml_op_name(dst->op), src0t);
-                            goto not_implemented;
-                        }
-                    } break;
-                case GGML_OP_ROPE:
-                    {
-                        GGML_ASSERT(ne10 == ne02);
-                        GGML_ASSERT(src0t == dstt);
-                        // 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];
-                        // skip 3, n_ctx used in GLM RoPE, unimplemented in Vulkan
-                        const int n_ctx_orig = ((int32_t *) dst->op_params)[4];
-
-                        const bool has_freq_factors = dst->src[2] != nullptr;
-
-                        float freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow;
-                        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));
-                        ggml_vk_rope(
-                            seq, id_src0, id_src1, id_src2, id_dst, off_src0, off_src1, off_src2, off_dst, src0t, n_dims, mode, n_ctx_orig,
-                            freq_base, freq_scale, has_freq_factors, ext_factor, attn_factor, beta_fast, beta_slow,
-                            ne01, ne02, ne03, nb00, nb01, nb02, nb03, ne0, nb0, nb1, nb2, nb3
-                        );
-                    } break;
-                case GGML_OP_DUP:
-                case GGML_OP_CPY:
-                case GGML_OP_CONT:
-                    {
-                        switch (src0t) {
-                            case GGML_TYPE_F32:
-                                {
-                                    switch (dstt) {
-                                        case GGML_TYPE_F16: ggml_vk_cpy_f32_f16(seq, id_src0, id_dst, off_src0, off_dst, ne00, ne01, ne02, ne03, nb00, nb01, nb02, nb03, ne0, ne1, ne2, nb0, nb1, nb2, nb3); break;
-                                        case GGML_TYPE_F32: ggml_vk_cpy_f32_f32(seq, id_src0, id_dst, off_src0, off_dst, ne00, ne01, ne02, ne03, nb00, nb01, nb02, nb03, ne0, ne1, ne2, nb0, nb1, nb2, nb3); break;
-                                        default: goto not_implemented;
-                                    }
-                                } break;
-                            case GGML_TYPE_F16:
-                                {
-                                    switch (dstt) {
-                                        case GGML_TYPE_F16: ggml_vk_cpy_f16_f16(seq, id_src0, id_dst, off_src0, off_dst, ne00, ne01, ne02, ne03, nb00, nb01, nb02, nb03, ne0, ne1, ne2, nb0, nb1, nb2, nb3); break;
-                                        case GGML_TYPE_F32: ggml_vk_cpy_f16_f32(seq, id_src0, id_dst, off_src0, off_dst, ne00, ne01, ne02, ne03, nb00, nb01, nb02, nb03, ne0, ne1, ne2, nb0, nb1, nb2, nb3); break;
-                                    default: goto not_implemented;
-                                } break;
-                            default: goto not_implemented;
-                            }
-                        }
-                    } break;
-                default: goto not_implemented;
-            }
-            continue;
-            not_implemented: {}
-            fprintf(stderr, "%s: node %3d, op = %8s not implemented\n", __func__, i, ggml_op_name(dst->op));
-            //GGML_ABORT("fatal error");
-        }
-
-        // Evaluate sequence
-        if (any_commands_recorded) {
-            seq.evalAsync();
-        }
-    }
-
-    // Wait for all sequences to finish
-    for (auto& sequence : sequences) {
-        if (sequence->isRunning())
-            sequence->evalAwait();
-    }
-
-    ggml_vk_free_descriptor_pool(ctx);
-}
-
-template<>
-kp::Tensor::TensorDataTypes
-kp::TensorT<half>::dataType()
-{
-    return TensorDataTypes::eFloat;
-}
-
-template<>
-kp::Tensor::TensorDataTypes
-kp::TensorT<uint8_t>::dataType()
-{
-    return TensorDataTypes::eUnsignedInt;
-}
-
-////////////////////////////////////////////////////////////////////////////////
-
-// backend interface
-
-struct ggml_backend_kompute_buffer_type_context {
-    int         device;
-    int         device_ref = 0;
-    uint64_t    buffer_alignment;
-    uint64_t    max_alloc;
-    std::string name;
-
-    ggml_backend_kompute_buffer_type_context(int device, uint64_t buffer_alignment, uint64_t max_alloc)
-        : device(device), buffer_alignment(buffer_alignment), max_alloc(max_alloc), name(ggml_kompute_format_name(device)) {}
-};
-
-static void ggml_backend_kompute_device_ref(ggml_backend_buffer_type_t buft) {
-    auto * ctx = static_cast<ggml_backend_kompute_buffer_type_context *>(buft->context);
-
-    if (!ctx->device_ref) {
-        komputeManager()->initializeDevice(
-            ctx->device, {}, {
-                "VK_KHR_shader_float16_int8", "VK_KHR_8bit_storage",
-                "VK_KHR_16bit_storage", "VK_KHR_shader_non_semantic_info"
-            }
-        );
-    }
-
-    assert(ggml_vk_has_device());
-    ctx->device_ref++;
-}
-
-static void ggml_backend_kompute_device_unref(ggml_backend_buffer_type_t buft) {
-    auto * ctx = static_cast<ggml_backend_kompute_buffer_type_context *>(buft->context);
-
-    assert(ctx->device_ref > 0);
-
-    ctx->device_ref--;
-
-    if (!ctx->device_ref) {
-        komputeManager.destroy();
-    }
-}
-
-static void ggml_backend_kompute_buffer_free_buffer(ggml_backend_buffer_t buffer) {
-    auto * memory = (ggml_vk_memory *)buffer->context;
-    if (ggml_vk_has_device()) {
-        ggml_vk_free_memory(*memory);
-    }
-    delete memory;
-}
-
-static void * ggml_backend_kompute_buffer_get_base(ggml_backend_buffer_t buffer) {
-    return ((ggml_vk_memory *)buffer->context)->data;
-}
-
-static void ggml_backend_kompute_buffer_set_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
-    GGML_UNUSED(buffer);
-
-    const auto res = ggml_vk_get_tensor(tensor);
-    GGML_ASSERT(res);
-
-    memcpy((char *)tensor->data + offset, data, size);
-
-    komputeManager()->sequence()->eval<kp::OpTensorSyncDevice>({res});
-}
-
-static void ggml_backend_kompute_buffer_get_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * tensor, void * data, size_t offset, size_t size) {
-    GGML_UNUSED(buffer);
-
-    const auto res = ggml_vk_get_tensor(tensor);
-    GGML_ASSERT(res);
-
-    komputeManager()->sequence()->eval<kp::OpTensorSyncLocal>({res});
-
-    memcpy(data, (const char *)tensor->data + offset, size);
-}
-
-static void ggml_backend_kompute_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
-    auto * memory = (ggml_vk_memory *)buffer->context;
-    memset(memory->data, value, buffer->size);
-
-    if (memory->stagingBuffer)
-        komputeManager()->sequence()->eval<kp::OpBufferSyncDevice>(memory->primaryBuffer, memory->stagingBuffer, memory->size);
-}
-
-static ggml_backend_buffer_i ggml_backend_kompute_buffer_i = {
-    /* .free_buffer     = */ ggml_backend_kompute_buffer_free_buffer,
-    /* .get_base        = */ ggml_backend_kompute_buffer_get_base,
-    /* .init_tensor     = */ NULL,
-    /* .memset_tensor   = */ NULL,
-    /* .set_tensor      = */ ggml_backend_kompute_buffer_set_tensor,
-    /* .get_tensor      = */ ggml_backend_kompute_buffer_get_tensor,
-    /* .cpy_tensor      = */ NULL,
-    /* .clear           = */ ggml_backend_kompute_buffer_clear,
-    /* .reset           = */ NULL,
-};
-
-// default buffer type
-
-static const char * ggml_backend_kompute_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
-    auto * ctx = static_cast<ggml_backend_kompute_buffer_type_context *>(buft->context);
-    return ctx->name.c_str();
-}
-
-static ggml_backend_buffer_t ggml_backend_kompute_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
-    ggml_backend_kompute_device_ref(buft);
-    auto * ctx = new ggml_vk_memory(ggml_vk_allocate(size));
-    return ggml_backend_buffer_init(buft, ggml_backend_kompute_buffer_i, ctx, size);
-}
-
-static size_t ggml_backend_kompute_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
-    auto * ctx = static_cast<ggml_backend_kompute_buffer_type_context *>(buft->context);
-    return ctx->buffer_alignment;
-}
-
-static size_t ggml_backend_vk_buffer_type_get_max_size(ggml_backend_buffer_type_t buft) {
-    auto * ctx = static_cast<ggml_backend_kompute_buffer_type_context *>(buft->context);
-    return ctx->max_alloc;
-}
-
-static ggml_backend_buffer_type_i ggml_backend_kompute_buffer_type_interface = {
-    /* .get_name         = */ ggml_backend_kompute_buffer_type_get_name,
-    /* .alloc_buffer     = */ ggml_backend_kompute_buffer_type_alloc_buffer,
-    /* .get_alignment    = */ ggml_backend_kompute_buffer_type_get_alignment,
-    /* .get_max_size     = */ ggml_backend_vk_buffer_type_get_max_size,
-    /* .get_alloc_size   = */ NULL, // defaults to ggml_nbytes
-    /* .is_host          = */ NULL,
-};
-
-ggml_backend_buffer_type_t ggml_backend_kompute_buffer_type(int device) {
-    static std::mutex mutex;
-    std::lock_guard<std::mutex> lock(mutex);
-
-    auto devices = ggml_vk_available_devices();
-    int32_t device_count = (int32_t) devices.size();
-    GGML_ASSERT(device < device_count);
-    GGML_ASSERT(devices.size() <= GGML_KOMPUTE_MAX_DEVICES);
-
-    static ggml_backend_buffer_type
-        ggml_backend_kompute_buffer_types[GGML_KOMPUTE_MAX_DEVICES];
-
-    static bool ggml_backend_kompute_buffer_type_initialized = false;
-
-    if (!ggml_backend_kompute_buffer_type_initialized) {
-        for (int32_t i = 0; i < device_count; i++) {
-            ggml_backend_kompute_buffer_types[i] = {
-                /* .iface    = */ ggml_backend_kompute_buffer_type_interface,
-                /* .device   = */ ggml_backend_reg_dev_get(ggml_backend_kompute_reg(), i),
-                /* .context  = */ new ggml_backend_kompute_buffer_type_context{ i, devices[i].bufferAlignment, devices[i].maxAlloc },
-            };
-        }
-        ggml_backend_kompute_buffer_type_initialized = true;
-    }
-
-    return &ggml_backend_kompute_buffer_types[device];
-}
-
-// backend
-
-static const char * ggml_backend_kompute_name(ggml_backend_t backend) {
-    auto * ctx = static_cast<ggml_kompute_context *>(backend->context);
-    return ctx->name.c_str();
-}
-
-static void ggml_backend_kompute_free(ggml_backend_t backend) {
-    auto * ctx = static_cast<ggml_kompute_context *>(backend->context);
-
-    assert(ctx == s_kompute_context);
-    s_kompute_context = nullptr;
-    if (ctx != nullptr) {
-        delete ctx;
-    }
-
-    delete backend;
-}
-
-static ggml_status ggml_backend_kompute_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
-    auto * ctx = static_cast<ggml_kompute_context *>(backend->context);
-    ggml_vk_graph_compute(ctx, cgraph);
-    return GGML_STATUS_SUCCESS;
-}
-
-static struct ggml_backend_i kompute_backend_i = {
-    /* .get_name                = */ ggml_backend_kompute_name,
-    /* .free                    = */ ggml_backend_kompute_free,
-    /* .set_tensor_async        = */ NULL,
-    /* .get_tensor_async        = */ NULL,
-    /* .cpy_tensor_async        = */ NULL,
-    /* .synchronize             = */ NULL,
-    /* .graph_plan_create       = */ NULL,
-    /* .graph_plan_free         = */ NULL,
-    /* .graph_plan_update       = */ NULL,
-    /* .graph_plan_compute      = */ NULL,
-    /* .graph_compute           = */ ggml_backend_kompute_graph_compute,
-    /* .event_record            = */ NULL,
-    /* .event_wait              = */ NULL,
-};
-
-static ggml_guid_t ggml_backend_kompute_guid() {
-    static ggml_guid guid = { 0x7b, 0x57, 0xdc, 0xaf, 0xde, 0x12, 0x1d, 0x49, 0xfb, 0x35, 0xfa, 0x9b, 0x18, 0x31, 0x1d, 0xca };
-    return &guid;
-}
-
-ggml_backend_t ggml_backend_kompute_init(int device) {
-    GGML_ASSERT(s_kompute_context == nullptr);
-    s_kompute_context = new ggml_kompute_context(device);
-
-    ggml_backend_t kompute_backend = new ggml_backend {
-        /* .guid      = */ ggml_backend_kompute_guid(),
-        /* .interface = */ kompute_backend_i,
-        /* .device    = */ ggml_backend_reg_dev_get(ggml_backend_kompute_reg(), device),
-        /* .context   = */ s_kompute_context,
-    };
-
-    return kompute_backend;
-}
-
-bool ggml_backend_is_kompute(ggml_backend_t backend) {
-    return backend != NULL && ggml_guid_matches(backend->guid, ggml_backend_kompute_guid());
-}
-
-static size_t ggml_backend_kompute_get_device_count() {
-    auto devices = ggml_vk_available_devices();
-    return devices.size();
-}
-
-static void ggml_backend_kompute_get_device_description(int device, char * description, size_t description_size) {
-    auto devices = ggml_vk_available_devices();
-    GGML_ASSERT((size_t) device < devices.size());
-    snprintf(description, description_size, "%s", devices[device].name);
-}
-
-static void ggml_backend_kompute_get_device_memory(int device, size_t * free, size_t * total) {
-    auto devices = ggml_vk_available_devices();
-    GGML_ASSERT((size_t) device < devices.size());
-    *total = devices[device].heapSize;
-    *free = devices[device].heapSize;
-}
-
-//////////////////////////
-
-struct ggml_backend_kompute_device_context {
-    int device;
-    std::string name;
-    std::string description;
-};
-
-static const char * ggml_backend_kompute_device_get_name(ggml_backend_dev_t dev) {
-    ggml_backend_kompute_device_context * ctx = (ggml_backend_kompute_device_context *)dev->context;
-    return ctx->name.c_str();
-}
-
-static const char * ggml_backend_kompute_device_get_description(ggml_backend_dev_t dev) {
-    ggml_backend_kompute_device_context * ctx = (ggml_backend_kompute_device_context *)dev->context;
-    return ctx->description.c_str();
-}
-
-static void ggml_backend_kompute_device_get_memory(ggml_backend_dev_t dev, size_t * free, size_t * total) {
-    ggml_backend_kompute_device_context * ctx = (ggml_backend_kompute_device_context *)dev->context;
-    ggml_backend_kompute_get_device_memory(ctx->device, free, total);
-}
-
-static ggml_backend_buffer_type_t ggml_backend_kompute_device_get_buffer_type(ggml_backend_dev_t dev) {
-    ggml_backend_kompute_device_context * ctx = (ggml_backend_kompute_device_context *)dev->context;
-    return ggml_backend_kompute_buffer_type(ctx->device);
-}
-
-static bool ggml_backend_kompute_device_supports_buft(ggml_backend_dev_t dev, ggml_backend_buffer_type_t buft) {
-    if (buft->iface.get_name != ggml_backend_kompute_buffer_type_get_name) {
-        return false;
-    }
-
-    ggml_backend_kompute_device_context * ctx = (ggml_backend_kompute_device_context *)dev->context;
-    ggml_backend_kompute_buffer_type_context * buft_ctx = (ggml_backend_kompute_buffer_type_context *)buft->context;
-
-    return buft_ctx->device == ctx->device;
-}
-
-static enum ggml_backend_dev_type ggml_backend_kompute_device_get_type(ggml_backend_dev_t dev) {
-    GGML_UNUSED(dev);
-    return GGML_BACKEND_DEVICE_TYPE_GPU;
-}
-
-static void ggml_backend_kompute_device_get_props(ggml_backend_dev_t dev, struct ggml_backend_dev_props * props) {
-    props->name        = ggml_backend_kompute_device_get_name(dev);
-    props->description = ggml_backend_kompute_device_get_description(dev);
-    props->type        = ggml_backend_kompute_device_get_type(dev);
-    ggml_backend_kompute_device_get_memory(dev, &props->memory_free, &props->memory_total);
-    props->caps = {
-        /* async                  = */ false,
-        /* host_buffer            = */ false,
-        /* .buffer_from_host_ptr  = */ false,
-        /* events                 = */ false,
-    };
-}
-
-static ggml_backend_t ggml_backend_kompute_device_init(ggml_backend_dev_t dev, const char * params) {
-    GGML_UNUSED(params);
-    ggml_backend_kompute_device_context * ctx = (ggml_backend_kompute_device_context *)dev->context;
-    return ggml_backend_kompute_init(ctx->device);
-}
-
-static bool ggml_backend_kompute_device_offload_op(ggml_backend_dev_t dev, const ggml_tensor * op) {
-    const int min_batch_size = 32;
-
-    return (op->ne[1] >= min_batch_size && op->op != GGML_OP_GET_ROWS) ||
-           (op->ne[2] >= min_batch_size && op->op == GGML_OP_MUL_MAT_ID);
-
-    GGML_UNUSED(dev);
-}
-
-static const struct ggml_backend_device_i ggml_backend_kompute_device_i = {
-    /* .get_name             = */ ggml_backend_kompute_device_get_name,
-    /* .get_description      = */ ggml_backend_kompute_device_get_description,
-    /* .get_memory           = */ ggml_backend_kompute_device_get_memory,
-    /* .get_type             = */ ggml_backend_kompute_device_get_type,
-    /* .get_props            = */ ggml_backend_kompute_device_get_props,
-    /* .init_backend         = */ ggml_backend_kompute_device_init,
-    /* .get_buffer_type      = */ ggml_backend_kompute_device_get_buffer_type,
-    /* .get_host_buffer_type = */ NULL,
-    /* .buffer_from_host_ptr = */ NULL,
-    /* .supports_op          = */ ggml_backend_kompute_device_supports_op,
-    /* .supports_buft        = */ ggml_backend_kompute_device_supports_buft,
-    /* .offload_op           = */ ggml_backend_kompute_device_offload_op,
-    /* .event_new            = */ NULL,
-    /* .event_free           = */ NULL,
-    /* .event_synchronize    = */ NULL,
-};
-
-static const char * ggml_backend_kompute_reg_get_name(ggml_backend_reg_t reg) {
-    GGML_UNUSED(reg);
-    return "Kompute";
-}
-
-static size_t ggml_backend_kompute_reg_get_device_count(ggml_backend_reg_t reg) {
-    GGML_UNUSED(reg);
-    return ggml_backend_kompute_get_device_count();
-}
-
-static ggml_backend_dev_t ggml_backend_kompute_reg_get_device(ggml_backend_reg_t reg, size_t device) {
-    static std::vector<ggml_backend_dev_t> devices;
-
-    static bool initialized = false;
-
-    {
-        static std::mutex mutex;
-        std::lock_guard<std::mutex> lock(mutex);
-        if (!initialized) {
-            for (size_t i = 0; i < ggml_backend_kompute_get_device_count(); i++) {
-                ggml_backend_kompute_device_context * ctx = new ggml_backend_kompute_device_context;
-                char desc[256];
-                ggml_backend_kompute_get_device_description(i, desc, sizeof(desc));
-                ctx->device = i;
-                ctx->name = "Kompute" + std::to_string(i);
-                ctx->description = desc;
-                devices.push_back(new ggml_backend_device {
-                    /* .iface   = */ ggml_backend_kompute_device_i,
-                    /* .reg     = */ reg,
-                    /* .context = */ ctx,
-                });
-            }
-            initialized = true;
-        }
-    }
-
-    GGML_ASSERT(device < devices.size());
-    return devices[device];
-}
-
-static const struct ggml_backend_reg_i ggml_backend_kompute_reg_i = {
-    /* .get_name         = */ ggml_backend_kompute_reg_get_name,
-    /* .get_device_count = */ ggml_backend_kompute_reg_get_device_count,
-    /* .get_device       = */ ggml_backend_kompute_reg_get_device,
-    /* .get_proc_address = */ NULL,
-};
-
-ggml_backend_reg_t ggml_backend_kompute_reg() {
-    static ggml_backend_reg reg = {
-        /* .api_version = */ GGML_BACKEND_API_VERSION,
-        /* .iface       = */ ggml_backend_kompute_reg_i,
-        /* .context     = */ nullptr,
-    };
-
-    return &reg;
-}
-
-GGML_BACKEND_DL_IMPL(ggml_backend_kompute_reg)
diff --git a/ggml/src/ggml-kompute/kompute b/ggml/src/ggml-kompute/kompute
deleted file mode 160000
index 4565194ed..000000000
--- a/ggml/src/ggml-kompute/kompute
+++ /dev/null
@@ -1 +0,0 @@
-Subproject commit 4565194ed7c32d1d2efa32ceab4d3c6cae006306
diff --git a/ggml/src/ggml-kompute/kompute-shaders/common.comp b/ggml/src/ggml-kompute/kompute-shaders/common.comp
deleted file mode 100644
index dbe4cf804..000000000
--- a/ggml/src/ggml-kompute/kompute-shaders/common.comp
+++ /dev/null
@@ -1,112 +0,0 @@
-#extension GL_EXT_shader_16bit_storage: require
-#extension GL_EXT_shader_8bit_storage: require
-#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
-#extension GL_EXT_shader_explicit_arithmetic_types_int8: require
-#extension GL_EXT_shader_explicit_arithmetic_types_int16: require
-#extension GL_EXT_shader_explicit_arithmetic_types_int64: require
-#extension GL_EXT_control_flow_attributes: enable
-#extension GL_KHR_shader_subgroup_arithmetic : require
-#extension GL_EXT_debug_printf : enable
-
-#define QK4_0 32
-#define QK4_1 32
-
-#define GELU_COEF_A 0.044715
-#define SQRT_2_OVER_PI 0.79788456080286535587989211986876
-#define TWOPI_F 6.283185307179586f
-
-#define QK_K 256
-#define K_SCALE_SIZE 12
-
-#define u8BufToU16(buf, idx) (((uint16_t(buf[idx + 1]) << 8)) | buf[idx])
-#define u8BufToFloat16(buf, idx) uint16BitsToHalf u8BufToU16(buf, idx)
-#define u8BufToU32(buf, idx) (((uint32_t u8BufToU16(buf, idx + 2) << 8 | buf[idx + 1]) << 8) | buf[idx])
-#define u8BufToFloat(buf, idx) uintBitsToFloat u8BufToU32(buf, idx)
-
-#define sizeof_block_q4_0 0x12
-struct block_q4_0 {
-    float16_t d;
-    uint8_t qs[QK4_0 / 2];
-};
-mat4 dequantize_q4_0(const block_q4_0 xb, uint il) {
-    const float d1 = il != 0 ? (xb.d / 16.f) : xb.d;
-    const float d2 = d1 / 256.f;
-    const float md = -8.f * xb.d;
-    const uint16_t mask0 = il != 0 ? uint16_t(0x00F0) : uint16_t(0x000F);
-    const uint16_t mask1 = mask0 << 8;
-
-    mat4 reg;
-    for (int i=0;i<8;i++) {
-        uint16_t b = (uint16_t(xb.qs[2 * i + 1]) << 8) | uint16_t(xb.qs[2 * i]);
-        reg[i/2][2*(i%2)+0] = d1 * (b & mask0) + md;
-        reg[i/2][2*(i%2)+1] = d2 * (b & mask1) + md;
-    }
-    return reg;
-}
-
-#define sizeof_block_q4_1 0x14
-struct block_q4_1 {
-    float16_t d;
-    float16_t m;
-    uint8_t qs[QK4_1 / 2];
-};
-mat4 dequantize_q4_1(const block_q4_1 xb, uint il) {
-    const float d1 = il != 0 ? (xb.d / 16.f) : xb.d;
-    const float d2 = d1 / 256.f;
-    const float  m = xb.m;
-    const uint16_t mask0 = il != 0 ? uint16_t(0x00F0) : uint16_t(0x000F);
-    const uint16_t mask1 = mask0 << 8;
-
-    mat4 reg;
-    for (int i=0;i<8;i++) {
-        uint16_t b = (uint16_t(xb.qs[2 * i + 1]) << 8) | uint16_t(xb.qs[2 * i]);
-        reg[i/2][2*(i%2)+0] = ((b & mask0) * d1) + m;
-        reg[i/2][2*(i%2)+1] = ((b & mask1) * d2) + m;
-    }
-    return reg;
-}
-
-#define sizeof_block_q4_k 144
-struct block_q4_k {
-    float16_t d;
-    float16_t dmin;
-    uint8_t scales[K_SCALE_SIZE];
-    uint8_t qs[QK_K/2];
-};
-
-#define sizeof_block_q6_k 210
-struct block_q6_k {
-    uint8_t ql[QK_K/2];      // quants, lower 4 bits
-    uint8_t qh[QK_K/4];      // quants, upper 2 bits
-    int8_t  scales[QK_K/16]; // scales, quantized with 8 bits
-    float16_t d;             // super-block scale
-};
-mat4 dequantize_q6_k(const block_q6_k xb, uint il) {
-    const float16_t d_all = xb.d;
-
-    const uint qlIndex = 64*(il/8) + 32*((il/2)&1) + 16*(il&1);
-    const uint qhIndex = 32*(il/8) + 16*(il&1);
-    float16_t sc = xb.scales[(il%2) + 2 * ((il/2))];
-    il = (il/2) & 3;
-
-    const uint16_t  kmask1 = il>1 ? uint16_t(il>2 ? 192 : 48) : uint16_t(il>0 ? 12 : 3);
-    const uint16_t  kmask2 = il>1 ? uint8_t(0xF0)             : uint8_t(0x0F);
-    const float16_t coef   = il>1 ? float16_t(1.f/16.f)       : float16_t(1.f);
-    const float16_t ml = float16_t(d_all * sc * 32.f);
-    const float16_t dl = float16_t(d_all * sc * coef);
-    mat4 reg;
-    for (int i = 0; i < 16; ++i) {
-        const float16_t q = (il&1) != 0 ? ((xb.ql[qlIndex + i] & kmask2) | ((xb.qh[qhIndex + i] & kmask1) << 2))
-                                        : ((xb.ql[qlIndex + i] & kmask2) | ((xb.qh[qhIndex + i] & kmask1) << 4));
-        reg[i/4][i%4] = dl * q - ml;
-    }
-    return reg;
-}
-
-
-#define QK8_0 32
-// struct block_q8_0 {
-//     float16_t d;         // delta
-//     int8_t    qs[QK8_0]; // quants
-// };
-#define sizeof_block_q8_0 34
diff --git a/ggml/src/ggml-kompute/kompute-shaders/op_add.comp b/ggml/src/ggml-kompute/kompute-shaders/op_add.comp
deleted file mode 100644
index b7b76a79d..000000000
--- a/ggml/src/ggml-kompute/kompute-shaders/op_add.comp
+++ /dev/null
@@ -1,58 +0,0 @@
-#version 450
-
-#include "common.comp"
-
-layout(local_size_x = 1024) in;
-
-layout(binding = 0) buffer restrict readonly tensorInA { float inA[]; };
-layout(binding = 1) buffer restrict readonly tensorInB { float inB[]; };
-layout(binding = 2) buffer restrict writeonly tensorOut { float out_[]; };
-
-layout(push_constant) uniform PushConstants {
-    uint inAOff;
-    uint inBOff;
-    uint outOff;
-    int ne00;
-    int nb00;
-    int nb01;
-    int nb02;
-    int nb03;
-    int ne10;
-    int ne11;
-    int ne12;
-    int ne13;
-    int nb10;
-    int nb11;
-    int nb12;
-    int nb13;
-    int ne0;
-    int nb0;
-    int nb1;
-    int nb2;
-    int nb3;
-  //int offs; // TODO: needed for GGML_OP_ACC, see metal code
-} pcs;
-
-// general-purpose kernel for addition of two tensors
-// pros: works for non-contiguous tensors, supports broadcast across dims 1, 2 and 3
-// cons: not very efficient
-void main() {
-    const uint i03 = gl_WorkGroupID.z;
-    const uint i02 = gl_WorkGroupID.y;
-    const uint i01 = gl_WorkGroupID.x;
-
-    const uint i13 = i03 % pcs.ne13;
-    const uint i12 = i02 % pcs.ne12;
-    const uint i11 = i01 % pcs.ne11;
-
-    int offs = 0; // TMP (see above)
-
-    uint src0_off = uint((i03*pcs.nb03 + i02*pcs.nb02 + i01*pcs.nb01 + offs) / 4);
-    uint src1_off = uint((i13*pcs.nb13 + i12*pcs.nb12 + i11*pcs.nb11       ) / 4);
-    uint dst_off  = uint((i03*pcs.nb3  + i02*pcs.nb2  + i01*pcs.nb1  + offs) / 4);
-
-    for (uint i0 = gl_LocalInvocationID.x; i0 < pcs.ne0; i0 += gl_WorkGroupSize.x) {
-        const uint i10 = i0 % pcs.ne10;
-        out_[pcs.outOff + dst_off + i0] = inA[pcs.inAOff + src0_off + i0] + inB[pcs.inBOff + src1_off + i10];
-    }
-}
diff --git a/ggml/src/ggml-kompute/kompute-shaders/op_addrow.comp b/ggml/src/ggml-kompute/kompute-shaders/op_addrow.comp
deleted file mode 100644
index 2376a6b8f..000000000
--- a/ggml/src/ggml-kompute/kompute-shaders/op_addrow.comp
+++ /dev/null
@@ -1,25 +0,0 @@
-#version 450
-
-#include "common.comp"
-
-layout(local_size_x = 1) in;
-
-layout(binding = 0) buffer restrict readonly tensorInA { float inA[]; };
-layout(binding = 1) buffer restrict readonly tensorInB { float inB[]; };
-layout(binding = 2) buffer restrict writeonly tensorOut { float out_[]; };
-
-layout(push_constant) uniform PushConstants {
-    uint inAOff;
-    uint inBOff;
-    uint outOff;
-    uint row;
-} pcs;
-
-void main() {
-    const uint baseIndex = gl_WorkGroupID.x * 4;
-
-    for (uint x = 0; x < 4; x++) {
-        const uint i = baseIndex + x;
-        out_[i + pcs.outOff] = inA[i + pcs.inAOff] + inB[(i % pcs.row) + pcs.inBOff];
-    }
-}
diff --git a/ggml/src/ggml-kompute/kompute-shaders/op_cpy_f16_f16.comp b/ggml/src/ggml-kompute/kompute-shaders/op_cpy_f16_f16.comp
deleted file mode 100644
index d57247d2d..000000000
--- a/ggml/src/ggml-kompute/kompute-shaders/op_cpy_f16_f16.comp
+++ /dev/null
@@ -1,52 +0,0 @@
-#version 450
-
-#include "common.comp"
-
-#define IN_TYPE float16_t
-#define IN_TYPE_SIZE 2
-#define OUT_TYPE float16_t
-#define OUT_TYPE_SIZE 2
-
-layout(local_size_x = 1024) in;
-
-layout (binding = 0) readonly buffer tensorIn { IN_TYPE in_[]; };
-layout (binding = 1) writeonly buffer tensorOut { OUT_TYPE out_[]; };
-
-layout (push_constant) uniform parameter {
-    uint inOff;
-    uint outOff;
-    int ne00;
-    int ne01;
-    int ne02;
-    uint nb00;
-    uint nb01;
-    uint nb02;
-    uint nb03;
-    int ne0;
-    int ne1;
-    int ne2;
-    uint nb0;
-    uint nb1;
-    uint nb2;
-    uint nb3;
-} pcs;
-
-void main() {
-    const uint i03 = gl_WorkGroupID.z;
-    const uint i02 = gl_WorkGroupID.y;
-    const uint i01 = gl_WorkGroupID.x;
-
-    const int n = int(i03)*pcs.ne02*pcs.ne01*pcs.ne00 + int(i02)*pcs.ne01*pcs.ne00 + int(i01)*pcs.ne00;
-
-    const int i3 = n / (pcs.ne2*pcs.ne1*pcs.ne0);
-    const int i2 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0) / (pcs.ne1*pcs.ne0);
-    const int i1 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0 - i2*pcs.ne1*pcs.ne0) / pcs.ne0;
-    const int i0 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0 - i2*pcs.ne1*pcs.ne0 - i1*pcs.ne0);
-
-    const uint dst_data = (i3*pcs.nb3 + i2*pcs.nb2 + i1*pcs.nb1 + i0*pcs.nb0) / OUT_TYPE_SIZE + pcs.outOff; // Based from out_
-
-    for (uint i00 = gl_LocalInvocationID.x; i00 < pcs.ne00; i00 += gl_WorkGroupSize.x) {
-        const uint src = uint((i03*pcs.nb03 + i02*pcs.nb02 + i01*pcs.nb01 + i00*pcs.nb00) / IN_TYPE_SIZE) + pcs.inOff; // Based from in_
-        out_[dst_data+i00] = OUT_TYPE(in_[src]);
-    }
-}
diff --git a/ggml/src/ggml-kompute/kompute-shaders/op_cpy_f16_f32.comp b/ggml/src/ggml-kompute/kompute-shaders/op_cpy_f16_f32.comp
deleted file mode 100644
index b568bcd7b..000000000
--- a/ggml/src/ggml-kompute/kompute-shaders/op_cpy_f16_f32.comp
+++ /dev/null
@@ -1,52 +0,0 @@
-#version 450
-
-#include "common.comp"
-
-#define IN_TYPE float16_t
-#define IN_TYPE_SIZE 2
-#define OUT_TYPE float
-#define OUT_TYPE_SIZE 4
-
-layout(local_size_x = 1024) in;
-
-layout (binding = 0) readonly buffer tensorIn { IN_TYPE in_[]; };
-layout (binding = 1) writeonly buffer tensorOut { OUT_TYPE out_[]; };
-
-layout (push_constant) uniform parameter {
-    uint inOff;
-    uint outOff;
-    int ne00;
-    int ne01;
-    int ne02;
-    uint nb00;
-    uint nb01;
-    uint nb02;
-    uint nb03;
-    int ne0;
-    int ne1;
-    int ne2;
-    uint nb0;
-    uint nb1;
-    uint nb2;
-    uint nb3;
-} pcs;
-
-void main() {
-    const uint i03 = gl_WorkGroupID.z;
-    const uint i02 = gl_WorkGroupID.y;
-    const uint i01 = gl_WorkGroupID.x;
-
-    const int n = int(i03)*pcs.ne02*pcs.ne01*pcs.ne00 + int(i02)*pcs.ne01*pcs.ne00 + int(i01)*pcs.ne00;
-
-    const int i3 = n / (pcs.ne2*pcs.ne1*pcs.ne0);
-    const int i2 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0) / (pcs.ne1*pcs.ne0);
-    const int i1 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0 - i2*pcs.ne1*pcs.ne0) / pcs.ne0;
-    const int i0 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0 - i2*pcs.ne1*pcs.ne0 - i1*pcs.ne0);
-
-    const uint dst_data = (i3*pcs.nb3 + i2*pcs.nb2 + i1*pcs.nb1 + i0*pcs.nb0) / OUT_TYPE_SIZE + pcs.outOff; // Based from out_
-
-    for (uint i00 = gl_LocalInvocationID.x; i00 < pcs.ne00; i00 += gl_WorkGroupSize.x) {
-        const uint src = uint((i03*pcs.nb03 + i02*pcs.nb02 + i01*pcs.nb01 + i00*pcs.nb00) / IN_TYPE_SIZE) + pcs.inOff; // Based from in_
-        out_[dst_data+i00] = OUT_TYPE(in_[src]);
-    }
-}
diff --git a/ggml/src/ggml-kompute/kompute-shaders/op_cpy_f32_f16.comp b/ggml/src/ggml-kompute/kompute-shaders/op_cpy_f32_f16.comp
deleted file mode 100644
index 99b228343..000000000
--- a/ggml/src/ggml-kompute/kompute-shaders/op_cpy_f32_f16.comp
+++ /dev/null
@@ -1,52 +0,0 @@
-#version 450
-
-#include "common.comp"
-
-#define IN_TYPE float
-#define IN_TYPE_SIZE 4
-#define OUT_TYPE float16_t
-#define OUT_TYPE_SIZE 2
-
-layout(local_size_x = 1024) in;
-
-layout (binding = 0) readonly buffer tensorIn { IN_TYPE in_[]; };
-layout (binding = 1) writeonly buffer tensorOut { OUT_TYPE out_[]; };
-
-layout (push_constant) uniform parameter {
-    uint inOff;
-    uint outOff;
-    int ne00;
-    int ne01;
-    int ne02;
-    uint nb00;
-    uint nb01;
-    uint nb02;
-    uint nb03;
-    int ne0;
-    int ne1;
-    int ne2;
-    uint nb0;
-    uint nb1;
-    uint nb2;
-    uint nb3;
-} pcs;
-
-void main() {
-    const uint i03 = gl_WorkGroupID.z;
-    const uint i02 = gl_WorkGroupID.y;
-    const uint i01 = gl_WorkGroupID.x;
-
-    const int n = int(i03)*pcs.ne02*pcs.ne01*pcs.ne00 + int(i02)*pcs.ne01*pcs.ne00 + int(i01)*pcs.ne00;
-
-    const int i3 = n / (pcs.ne2*pcs.ne1*pcs.ne0);
-    const int i2 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0) / (pcs.ne1*pcs.ne0);
-    const int i1 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0 - i2*pcs.ne1*pcs.ne0) / pcs.ne0;
-    const int i0 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0 - i2*pcs.ne1*pcs.ne0 - i1*pcs.ne0);
-
-    const uint dst_data = (i3*pcs.nb3 + i2*pcs.nb2 + i1*pcs.nb1 + i0*pcs.nb0) / OUT_TYPE_SIZE + pcs.outOff; // Based from out_
-
-    for (uint i00 = gl_LocalInvocationID.x; i00 < pcs.ne00; i00 += gl_WorkGroupSize.x) {
-        const uint src = uint((i03*pcs.nb03 + i02*pcs.nb02 + i01*pcs.nb01 + i00*pcs.nb00) / IN_TYPE_SIZE) + pcs.inOff; // Based from in_
-        out_[dst_data+i00] = OUT_TYPE(in_[src]);
-    }
-}
diff --git a/ggml/src/ggml-kompute/kompute-shaders/op_cpy_f32_f32.comp b/ggml/src/ggml-kompute/kompute-shaders/op_cpy_f32_f32.comp
deleted file mode 100644
index 2fc998492..000000000
--- a/ggml/src/ggml-kompute/kompute-shaders/op_cpy_f32_f32.comp
+++ /dev/null
@@ -1,52 +0,0 @@
-#version 450
-
-#include "common.comp"
-
-#define IN_TYPE float
-#define IN_TYPE_SIZE 4
-#define OUT_TYPE float
-#define OUT_TYPE_SIZE 4
-
-layout(local_size_x = 1024) in;
-
-layout (binding = 0) readonly buffer tensorIn { IN_TYPE in_[]; };
-layout (binding = 1) writeonly buffer tensorOut { OUT_TYPE out_[]; };
-
-layout (push_constant) uniform parameter {
-    uint inOff;
-    uint outOff;
-    int ne00;
-    int ne01;
-    int ne02;
-    uint nb00;
-    uint nb01;
-    uint nb02;
-    uint nb03;
-    int ne0;
-    int ne1;
-    int ne2;
-    uint nb0;
-    uint nb1;
-    uint nb2;
-    uint nb3;
-} pcs;
-
-void main() {
-    const uint i03 = gl_WorkGroupID.z;
-    const uint i02 = gl_WorkGroupID.y;
-    const uint i01 = gl_WorkGroupID.x;
-
-    const int n = int(i03)*pcs.ne02*pcs.ne01*pcs.ne00 + int(i02)*pcs.ne01*pcs.ne00 + int(i01)*pcs.ne00;
-
-    const int i3 = n / (pcs.ne2*pcs.ne1*pcs.ne0);
-    const int i2 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0) / (pcs.ne1*pcs.ne0);
-    const int i1 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0 - i2*pcs.ne1*pcs.ne0) / pcs.ne0;
-    const int i0 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0 - i2*pcs.ne1*pcs.ne0 - i1*pcs.ne0);
-
-    const uint dst_data = (i3*pcs.nb3 + i2*pcs.nb2 + i1*pcs.nb1 + i0*pcs.nb0) / OUT_TYPE_SIZE + pcs.outOff; // Based from out_
-
-    for (uint i00 = gl_LocalInvocationID.x; i00 < pcs.ne00; i00 += gl_WorkGroupSize.x) {
-        const uint src = uint((i03*pcs.nb03 + i02*pcs.nb02 + i01*pcs.nb01 + i00*pcs.nb00) / IN_TYPE_SIZE) + pcs.inOff; // Based from in_
-        out_[dst_data+i00] = OUT_TYPE(in_[src]);
-    }
-}
diff --git a/ggml/src/ggml-kompute/kompute-shaders/op_diagmask.comp b/ggml/src/ggml-kompute/kompute-shaders/op_diagmask.comp
deleted file mode 100644
index 291c3fc18..000000000
--- a/ggml/src/ggml-kompute/kompute-shaders/op_diagmask.comp
+++ /dev/null
@@ -1,30 +0,0 @@
-#version 450
-
-#include "common.comp"
-
-layout(local_size_x = 1) in;
-
-layout(binding = 0) buffer restrict readonly tensorIn { float in_[]; };
-layout(binding = 1) buffer restrict writeonly tensorOut { float out_[]; };
-
-layout(push_constant) uniform PushConstants {
-    uint inOff;
-    uint outOff;
-    uint n_past;
-    int ne00;
-    int ne01;
-} pcs;
-
-void main() {
-    const uint i02 = gl_WorkGroupID.z;
-    const uint i01 = gl_WorkGroupID.y;
-    const uint i00 = gl_WorkGroupID.x;
-
-    const uint index = i02*pcs.ne01*pcs.ne00 + i01*pcs.ne00 + i00;
-
-    if (i00 > pcs.n_past + i01) {
-        out_[index + pcs.outOff] = uintBitsToFloat(0xFF800000);
-    } else {
-        out_[index + pcs.outOff] = in_[index + pcs.inOff];
-    }
-}
diff --git a/ggml/src/ggml-kompute/kompute-shaders/op_gelu.comp b/ggml/src/ggml-kompute/kompute-shaders/op_gelu.comp
deleted file mode 100644
index 9d8c53710..000000000
--- a/ggml/src/ggml-kompute/kompute-shaders/op_gelu.comp
+++ /dev/null
@@ -1,22 +0,0 @@
-#version 450
-
-#include "common.comp"
-
-layout(local_size_x = 1) in;
-
-layout(binding = 0) buffer restrict readonly tensorIn { float in_[]; };
-layout(binding = 1) buffer restrict writeonly tensorOut { float out_[]; };
-layout(push_constant) uniform PushConstants {
-    uint inOff;
-    uint outOff;
-} pcs;
-
-void main() {
-    const uint baseIndex = gl_WorkGroupID.x * 8;
-
-    for (uint x = 0; x < 8; x++) {
-        const uint i = baseIndex + x;
-        const float y = in_[i + pcs.inOff];
-        out_[i + pcs.outOff] = 0.5*y*(1.0 + tanh(clamp(SQRT_2_OVER_PI*y*(1.0 + GELU_COEF_A*y*y), -15.0, 15.0)));
-    }
-}
diff --git a/ggml/src/ggml-kompute/kompute-shaders/op_getrows.comp b/ggml/src/ggml-kompute/kompute-shaders/op_getrows.comp
deleted file mode 100644
index 1a5581b23..000000000
--- a/ggml/src/ggml-kompute/kompute-shaders/op_getrows.comp
+++ /dev/null
@@ -1,17 +0,0 @@
-void main() {
-    const uint i = gl_WorkGroupID.x;
-    const int r = inB[i + pcs.inBOff];
-
-    int z = 0;
-    for (uint ind = gl_LocalInvocationID.x; ind < pcs.ne00/16; ind += gl_WorkGroupSize.x) {
-        const uint inIndex = (r * pcs.nb01 + pcs.inAOff) + ind/NL * SIZE_OF_BLOCK;
-        const mat4 result = dequantize_block(inIndex, ind%NL);
-        for (uint j = 0; j < 4; ++j) {
-            for (uint k = 0; k < 4; ++k) {
-                const uint outIndex = i * pcs.nb1/BYTES_FOR_TYPE + pcs.outOff + z;
-                out_[outIndex] = result[j][k];
-                ++z;
-            }
-        }
-    }
-}
diff --git a/ggml/src/ggml-kompute/kompute-shaders/op_getrows_f16.comp b/ggml/src/ggml-kompute/kompute-shaders/op_getrows_f16.comp
deleted file mode 100644
index 48c936108..000000000
--- a/ggml/src/ggml-kompute/kompute-shaders/op_getrows_f16.comp
+++ /dev/null
@@ -1,31 +0,0 @@
-#version 450
-
-#include "common.comp"
-
-layout(local_size_x = 1) in;
-
-layout (binding = 0) readonly buffer tensorInA { float16_t inA[]; };
-layout (binding = 1) readonly buffer tensorInB { int inB[]; };
-layout (binding = 2) writeonly buffer tensorOut { float out_[]; };
-
-layout (push_constant) uniform parameter {
-    uint inAOff;
-    uint inBOff;
-    uint outOff;
-    int ne00;
-    int nb01;
-    int nb1;
-} pcs;
-
-void dequantize_row_f16(uint x /*Based from inA unaligned*/, uint y /*Based from out_*/, int k) {
-    for (int j = 0; j < k; j++) {
-        out_[y + j] = inA[x + j];
-    }
-}
-
-void main() {
-    const uint i = gl_WorkGroupID.x;
-    const int r = inB[i + pcs.inBOff];
-
-    dequantize_row_f16(r*pcs.nb01/2/*bytes for float16*/ + pcs.inAOff, i*pcs.nb1/4 + pcs.outOff, pcs.ne00);
-}
diff --git a/ggml/src/ggml-kompute/kompute-shaders/op_getrows_f32.comp b/ggml/src/ggml-kompute/kompute-shaders/op_getrows_f32.comp
deleted file mode 100644
index 9d7acdaf8..000000000
--- a/ggml/src/ggml-kompute/kompute-shaders/op_getrows_f32.comp
+++ /dev/null
@@ -1,31 +0,0 @@
-#version 450
-
-#include "common.comp"
-
-layout(local_size_x = 1) in;
-
-layout (binding = 0) readonly buffer tensorInA { float inA[]; };
-layout (binding = 1) readonly buffer tensorInB { int inB[]; };
-layout (binding = 2) writeonly buffer tensorOut { float out_[]; };
-
-layout (push_constant) uniform parameter {
-    uint inAOff;
-    uint inBOff;
-    uint outOff;
-    int ne00;
-    int nb01;
-    int nb1;
-} pcs;
-
-void dequantize_row_f32(uint x /*Based from inA unaligned*/, uint y /*Based from out_*/, int k) {
-    for (int j = 0; j < k; j++) {
-        out_[y + j] = inA[x + j];
-    }
-}
-
-void main() {
-    const uint i = gl_WorkGroupID.x;
-    const int r = inB[i + pcs.inBOff];
-
-    dequantize_row_f32(r*pcs.nb01/4 + pcs.inAOff, i*pcs.nb1/4 + pcs.outOff, pcs.ne00);
-}
diff --git a/ggml/src/ggml-kompute/kompute-shaders/op_getrows_q4_0.comp b/ggml/src/ggml-kompute/kompute-shaders/op_getrows_q4_0.comp
deleted file mode 100644
index 32b2e891e..000000000
--- a/ggml/src/ggml-kompute/kompute-shaders/op_getrows_q4_0.comp
+++ /dev/null
@@ -1,38 +0,0 @@
-#version 450
-
-#include "common.comp"
-
-#define NL 2
-#define BYTES_FOR_TYPE 4 /*bytes for float*/
-#define SIZE_OF_BLOCK sizeof_block_q4_0
-
-layout(local_size_x = 1) in;
-
-layout (binding = 0) readonly buffer tensorInA { uint8_t inA[]; };
-layout (binding = 1) readonly buffer tensorInB { int inB[]; };
-layout (binding = 2) writeonly buffer tensorOut { float out_[]; };
-
-layout (push_constant) uniform parameter {
-    uint inAOff;
-    uint inBOff;
-    uint outOff;
-    int ne00;
-    int nb01;
-    int nb1;
-} pcs;
-
-block_q4_0 get_unaligned_block_q4_0(uint index) {
-    block_q4_0 fres;
-    fres.d = u8BufToFloat16(inA, index);
-    [[unroll]] for (uint it = 0; it != QK4_0 / 2; it++) {
-        fres.qs[it] = inA[index+2+it];
-    }
-    return fres;
-}
-
-mat4 dequantize_block(uint index, uint il) {
-    const block_q4_0 block = get_unaligned_block_q4_0(index);
-    return dequantize_q4_0(block, il);
-}
-
-#include "op_getrows.comp"
diff --git a/ggml/src/ggml-kompute/kompute-shaders/op_getrows_q4_1.comp b/ggml/src/ggml-kompute/kompute-shaders/op_getrows_q4_1.comp
deleted file mode 100644
index 87f2fbe17..000000000
--- a/ggml/src/ggml-kompute/kompute-shaders/op_getrows_q4_1.comp
+++ /dev/null
@@ -1,39 +0,0 @@
-#version 450
-
-#include "common.comp"
-
-#define NL 2
-#define BYTES_FOR_TYPE 4 /*bytes for float*/
-#define SIZE_OF_BLOCK sizeof_block_q4_1
-
-layout(local_size_x = 1) in;
-
-layout (binding = 0) readonly buffer tensorInA { uint8_t inA[]; };
-layout (binding = 1) readonly buffer tensorInB { int inB[]; };
-layout (binding = 2) writeonly buffer tensorOut { float out_[]; };
-
-layout (push_constant) uniform parameter {
-    uint inAOff;
-    uint inBOff;
-    uint outOff;
-    int ne00;
-    int nb01;
-    int nb1;
-} pcs;
-
-block_q4_1 get_unaligned_block_q4_1(uint index) {
-    block_q4_1 fres;
-    fres.d = u8BufToFloat16(inA, index);
-    fres.m = u8BufToFloat16(inA, index+2);
-    [[unroll]] for (uint it = 0; it != QK4_1 / 2; it++) {
-        fres.qs[it] = inA[index+4+it];
-    }
-    return fres;
-}
-
-mat4 dequantize_block(uint index, uint il) {
-    const block_q4_1 block = get_unaligned_block_q4_1(index);
-    return dequantize_q4_1(block, il);
-}
-
-#include "op_getrows.comp"
diff --git a/ggml/src/ggml-kompute/kompute-shaders/op_getrows_q6_k.comp b/ggml/src/ggml-kompute/kompute-shaders/op_getrows_q6_k.comp
deleted file mode 100644
index 9ce3545d1..000000000
--- a/ggml/src/ggml-kompute/kompute-shaders/op_getrows_q6_k.comp
+++ /dev/null
@@ -1,44 +0,0 @@
-#version 450
-
-#include "common.comp"
-
-#define NL 16
-#define BYTES_FOR_TYPE 4 /*bytes for float*/
-#define SIZE_OF_BLOCK sizeof_block_q6_k
-
-layout(local_size_x = 1) in;
-
-layout (binding = 0) readonly buffer tensorInA { uint8_t inA[]; };
-layout (binding = 1) readonly buffer tensorInB { int inB[]; };
-layout (binding = 2) writeonly buffer tensorOut { float out_[]; };
-
-layout (push_constant) uniform parameter {
-    uint inAOff;
-    uint inBOff;
-    uint outOff;
-    int ne00;
-    int nb01;
-    int nb1;
-} pcs;
-
-block_q6_k get_unaligned_block_q6_k(uint index) {
-    block_q6_k fres;
-    [[unroll]] for (uint it = 0; it != QK_K / 2; it++) {
-        fres.ql[it] = inA[index + it];
-    }
-    [[unroll]] for (uint it = 0; it != QK_K / 4; it++) {
-        fres.qh[it] = inA[index + QK_K/2 + it];
-    }
-    [[unroll]] for (uint it = 0; it != QK_K / 16; it++) {
-        fres.scales[it] = int8_t(inA[index + QK_K/2 + QK_K/4 + it]);
-    }
-    fres.d = u8BufToFloat16(inA, index + QK_K/2 + QK_K/4 + QK_K/16);
-    return fres;
-}
-
-mat4 dequantize_block(uint index, uint il) {
-    const block_q6_k block = get_unaligned_block_q6_k(index);
-    return dequantize_q6_k(block, il);
-}
-
-#include "op_getrows.comp"
diff --git a/ggml/src/ggml-kompute/kompute-shaders/op_mul.comp b/ggml/src/ggml-kompute/kompute-shaders/op_mul.comp
deleted file mode 100644
index c92647c4d..000000000
--- a/ggml/src/ggml-kompute/kompute-shaders/op_mul.comp
+++ /dev/null
@@ -1,52 +0,0 @@
-#version 450
-
-#include "common.comp"
-
-layout(local_size_x = 1024) in;
-
-layout(binding = 0) buffer restrict readonly tensorInA { float inA[]; };
-layout(binding = 1) buffer restrict readonly tensorInB { float inB[]; };
-layout(binding = 2) buffer restrict writeonly tensorOut { float out_[]; };
-
-layout(push_constant) uniform PushConstants {
-    uint inAOff;
-    uint inBOff;
-    uint outOff;
-    int ne00;
-    int nb00;
-    int nb01;
-    int nb02;
-    int nb03;
-    int ne10;
-    int ne11;
-    int ne12;
-    int ne13;
-    int nb10;
-    int nb11;
-    int nb12;
-    int nb13;
-    int ne0;
-    int nb0;
-    int nb1;
-    int nb2;
-    int nb3;
-} pcs;
-
-void main() {
-    const uint i03 = gl_WorkGroupID.z;
-    const uint i02 = gl_WorkGroupID.y;
-    const uint i01 = gl_WorkGroupID.x;
-
-    const uint i13 = i03 % pcs.ne13;
-    const uint i12 = i02 % pcs.ne12;
-    const uint i11 = i01 % pcs.ne11;
-
-    uint src0_off = uint((i03*pcs.nb03 + i02*pcs.nb02 + i01*pcs.nb01) / 4);
-    uint src1_off = uint((i13*pcs.nb13 + i12*pcs.nb12 + i11*pcs.nb11) / 4);
-    uint dst_off  = uint((i03*pcs.nb3  + i02*pcs.nb2  + i01*pcs.nb1)  / 4);
-
-    for (uint i0 = gl_LocalInvocationID.x; i0 < pcs.ne0; i0 += gl_WorkGroupSize.x) {
-        const uint i10 = i0 % pcs.ne10;
-        out_[pcs.outOff + dst_off + i0] = inA[pcs.inAOff + src0_off + i0] * inB[pcs.inBOff + src1_off + i10];
-    }
-}
diff --git a/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_f16.comp b/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_f16.comp
deleted file mode 100644
index 0ab1b2fc2..000000000
--- a/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_f16.comp
+++ /dev/null
@@ -1,69 +0,0 @@
-#version 450
-
-#include "common.comp"
-
-#extension GL_KHR_shader_subgroup_arithmetic : require
-
-layout(local_size_x_id = 0) in;
-
-layout (binding = 0) readonly buffer tensorInA { float16_t inA[]; };
-layout (binding = 1) readonly buffer tensorInB { float inB[]; };
-layout (binding = 2) writeonly buffer tensorOut { float out_[]; };
-
-layout (push_constant) uniform parameter {
-    uint inAOff;
-    uint inBOff;
-    uint outOff;
-    int ne00;
-    int ne01;
-    int ne02;
-    uint nb00;
-    uint nb01;
-    uint nb02;
-    uint nb03;
-    int ne10;
-    int ne11;
-    int ne12;
-    uint nb10;
-    uint nb11;
-    uint nb12;
-    uint nb13;
-    int ne0;
-    int ne1;
-    uint r2;
-    uint r3;
-} pcs;
-
-#define N_F16_F32 4
-
-void main() {
-    const uint r0 = gl_WorkGroupID.x;
-    const uint rb = gl_WorkGroupID.y*N_F16_F32;
-    const uint im = gl_WorkGroupID.z;
-
-    const uint i12 = im%pcs.ne12;
-    const uint i13 = im/pcs.ne12;
-
-    const uint offset0 = r0*pcs.nb01 + (i12/pcs.r2)*pcs.nb02 + (i13/pcs.r3)*pcs.nb03;
-
-    const uint x = offset0 / 2 + pcs.inAOff; // Based from inA
-
-    for (uint row = 0; row < N_F16_F32; ++row) {
-        uint r1 = rb + row;
-        if (r1 >= pcs.ne11) {
-            break;
-        }
-
-        const uint y = (r1*pcs.nb11 + i12*pcs.nb12 + i13*pcs.nb13) / 4 + pcs.inBOff;
-
-        float sumf = 0;
-        for (uint i = gl_SubgroupInvocationID.x; i < pcs.ne00; i += gl_SubgroupSize) {
-            sumf += float(inA[x+i]) * float(inB[y+i]);
-        }
-
-        const float all_sum = subgroupAdd(sumf);
-        if (subgroupElect()) {
-            out_[im*pcs.ne1*pcs.ne0 + r1*pcs.ne0 + r0 + pcs.outOff] = all_sum;
-        }
-    }
-}
diff --git a/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_mat_f32.comp b/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_mat_f32.comp
deleted file mode 100644
index d1ca4ad6c..000000000
--- a/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_mat_f32.comp
+++ /dev/null
@@ -1,51 +0,0 @@
-#version 450
-
-#include "common.comp"
-
-#extension GL_KHR_shader_subgroup_arithmetic : require
-#extension GL_EXT_debug_printf : enable
-
-// device subgroup size
-layout (local_size_x_id = 0) in;
-
-layout(binding = 0) readonly buffer tensorInA { float inA[]; };
-layout(binding = 1) readonly buffer tensorInB { float inB[]; };
-layout(binding = 2) writeonly buffer tensorOut { float out_[]; };
-
-layout(push_constant) uniform parameter {
-  uint inAOff;
-  uint inBOff;
-  uint outOff;
-  int ne00;
-  int ne01;
-  int ne02;
-  int ne11;
-  int ne12;
-  uint nb01;
-  uint nb02;
-  uint nb11;
-  uint nb12;
-  uint nb1;
-  uint nb2;
-}
-pcs;
-
-
-void main() {
-  uvec3 gid = gl_WorkGroupID;
-
-  uint bc_ab = pcs.ne12 > pcs.ne02 ? gid.z / (pcs.ne12 / pcs.ne02) : gid.z;
-  uint bc_ba = pcs.ne02 > pcs.ne12 ? gid.z / (pcs.ne02 / pcs.ne12) : gid.z;
-
-  const uint x = (gid.x*pcs.nb01 + bc_ab*pcs.nb02) / 4 + pcs.inAOff; // Based from inA
-  const uint y = (gid.y*pcs.nb11 + bc_ba*pcs.nb12) / 4 + pcs.inBOff; // based from inB
-  float sum = 0.0f;
-  for (uint i = gl_SubgroupInvocationID.x; i < pcs.ne00; i += gl_SubgroupSize) {
-      sum += float(inA[x+i]) * float(inB[y+i]);
-  }
-
-  const float all_sum = subgroupAdd(sum);
-  if (subgroupElect()) {
-    out_[gid.z*(pcs.nb2/4) + gid.y*(pcs.nb1/4) + gid.x + pcs.outOff] = all_sum;
-  }
-}
diff --git a/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q4_0.comp b/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q4_0.comp
deleted file mode 100644
index b0cea8bbe..000000000
--- a/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q4_0.comp
+++ /dev/null
@@ -1,33 +0,0 @@
-#version 450
-
-#include "common.comp"
-
-#define BLOCKS_IN_QUANT QK4_0
-#define SIZE_OF_BLOCK sizeof_block_q4_0
-#define N_ROWS 4
-
-#include "op_mul_mv_q_n_pre.comp"
-
-// The q4_0 version of this function
-float block_q_n_dot_y(uint block_index, uint yb, uint il) {
-    vec2 acc = vec2(0.0, 0.0);
-    const uint index = (block_index) * SIZE_OF_BLOCK + pcs.inAOff;
-    float d = float(u8BufToFloat16(inA, index));
-    float sumy = 0.0f;
-    for (int i = 0; i < BLOCKS_IN_QUANT/4; i+=2) {
-        const uint16_t b = u8BufToU16(inA, index + 2 + il + i);
-
-        const float yl0 = inB[yb + i];
-        const float yl1 = inB[yb + i + 1];
-        const float yl8 = inB[yb + i + BLOCKS_IN_QUANT/2];
-        const float yl9 = inB[yb + i + BLOCKS_IN_QUANT/2 + 1];
-
-        sumy += yl0 + yl1 + yl8 + yl9;
-
-        acc[0] += yl0 * (b & 0x000F) + yl1 / 256.f * (b & 0x0F00);
-        acc[1] += yl8 / 16.f * (b & 0x00F0) + yl9 / 4096.f * (b & 0xF000);
-    }
-    return d * (sumy * -8.f + acc[0] + acc[1]);
-}
-
-#include "op_mul_mv_q_n.comp"
diff --git a/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q4_1.comp b/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q4_1.comp
deleted file mode 100644
index 8582c61a3..000000000
--- a/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q4_1.comp
+++ /dev/null
@@ -1,35 +0,0 @@
-#version 450
-
-#include "common.comp"
-
-#define BLOCKS_IN_QUANT QK4_1
-#define SIZE_OF_BLOCK sizeof_block_q4_1
-#define N_ROWS 4
-
-#include "op_mul_mv_q_n_pre.comp"
-
-// The q4_1 version of this function
-float block_q_n_dot_y(uint block_index, uint yb, uint il) {
-    vec2 acc = vec2(0.0, 0.0);
-    const uint index = (block_index) * SIZE_OF_BLOCK + pcs.inAOff;
-    float d = float(u8BufToFloat16(inA, index));
-    float m = float(u8BufToFloat16(inA, index+2));
-
-    float sumy = 0.0f;
-    for (int i = 0; i < BLOCKS_IN_QUANT/4; i+=2) {
-        const uint16_t b = u8BufToU16(inA, index + 4 + il + i);
-
-        const float yl0 = inB[yb + i];
-        const float yl1 = inB[yb + i + 1];
-        const float yl8 = inB[yb + i + BLOCKS_IN_QUANT/2];
-        const float yl9 = inB[yb + i + BLOCKS_IN_QUANT/2 + 1];
-
-        sumy += yl0 + yl1 + yl8 + yl9;
-
-        acc[0] += yl0 * (b & 0x000F) + yl1 / 256.f * (b & 0x0F00);
-        acc[1] += yl8 / 16.f * (b & 0x00F0) + yl9 / 4096.f * (b & 0xF000);
-    }
-    return d * (acc[0] + acc[1]) + sumy * m;
-}
-
-#include "op_mul_mv_q_n.comp"
diff --git a/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q4_k.comp b/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q4_k.comp
deleted file mode 100644
index a5752a3a0..000000000
--- a/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q4_k.comp
+++ /dev/null
@@ -1,140 +0,0 @@
-#version 450
-
-#include "common.comp"
-
-#define N_DST 4
-#define SIZE_OF_BLOCK sizeof_block_q4_k
-
-layout(local_size_x = 4) in;
-layout(local_size_y = 8) in;
-layout(local_size_z = 1) in;
-
-layout (binding = 0) readonly buffer tensorInA { block_q4_k inA[]; };
-layout (binding = 1) readonly buffer tensorInB { float inB[]; };
-layout (binding = 2) writeonly buffer tensorOut { float out_[]; };
-
-layout (push_constant) uniform parameter {
-    uint inAOff;
-    uint inBOff;
-    uint outOff;
-    int ne00;
-    int ne10;
-    int ne0;
-    int ne1;
-    int ne01;
-    int ne02;
-    int ne12;
-    uint nb01;
-    uint nb02;
-    uint nb03;
-    uint nb11;
-    uint nb12;
-    uint nb13;
-    uint r2;
-    uint r3;
-} pcs;
-
-void main() {
-    const uint16_t kmask1 = uint16_t(0x3f3f);
-    const uint16_t kmask2 = uint16_t(0x0f0f);
-    const uint16_t kmask3 = uint16_t(0xc0c0);
-
-    const uint ix = gl_SubgroupInvocationID/8;  // 0...3
-    const uint it = gl_SubgroupInvocationID%8;  // 0...7
-    const uint iq = it/4;     // 0 or 1
-    const uint ir = it%4;     // 0...3
-
-    const uint nb = pcs.ne00/QK_K;
-
-    const uint r0 = gl_WorkGroupID.x;
-    const uint r1 = gl_WorkGroupID.y;
-    const uint im = gl_WorkGroupID.z;
-
-    const uint first_row = r0 * N_DST;
-    const uint ib_row = first_row * nb;
-
-    const uint i12 = im%pcs.ne12;
-    const uint i13 = im/pcs.ne12;
-
-    const uint offset0 = first_row*(pcs.nb01/SIZE_OF_BLOCK) + (i12/pcs.r2)*(pcs.nb02/SIZE_OF_BLOCK) + (i13/pcs.r3)*(pcs.nb03/SIZE_OF_BLOCK);
-    const uint offset1 =        r1*pcs.nb11 + (i12       )*pcs.nb12 + (i13       )*pcs.nb13;
-
-    const uint xblk = offset0 + pcs.inAOff;
-    const uint y = (offset1 / 4) + pcs.inBOff;
-
-    float yl[16];
-    float yh[16];
-    float sumf[N_DST] = {0.f, 0.f, 0.f, 0.f};
-    float all_sum = 0.f;
-
-    uint y4 = y + ix * QK_K + 64 * iq + 8 * ir;
-
-    for (uint ib = ix; ib < nb; ib += 4) {
-        const uint blk_idx = ib + xblk;
-
-        float sumy[4] = {0.f, 0.f, 0.f, 0.f};
-        for (int i = 0; i < 8; ++i) {
-            yl[i+0] = inB[y4+i+  0]; sumy[0] += yl[i+0];
-            yl[i+8] = inB[y4+i+ 32]; sumy[1] += yl[i+8];
-            yh[i+0] = inB[y4+i+128]; sumy[2] += yh[i+0];
-            yh[i+8] = inB[y4+i+160]; sumy[3] += yh[i+8];
-        }
-
-        for (int row = 0; row < N_DST; row++) {
-            uint row_idx = row * (pcs.nb01 / SIZE_OF_BLOCK);
-
-            uint16_t sc_0 = u8BufToU16(inA[blk_idx + row_idx].scales, iq * 2 + 0);
-            uint16_t sc_1 = u8BufToU16(inA[blk_idx + row_idx].scales, iq * 2 + 2);
-            uint16_t sc_2 = u8BufToU16(inA[blk_idx + row_idx].scales, iq * 2 + 4);
-            uint16_t sc_3 = u8BufToU16(inA[blk_idx + row_idx].scales, iq * 2 + 6);
-            uint16_t sc_4 = u8BufToU16(inA[blk_idx + row_idx].scales, iq * 2 + 8);
-
-            uint16_t sc16[4];
-            sc16[0] = sc_0 & kmask1;
-            sc16[1] = sc_2 & kmask1;
-            sc16[2] = ((sc_4 >> 0) & kmask2) | ((sc_0 & kmask3) >> 2);
-            sc16[3] = ((sc_4 >> 4) & kmask2) | ((sc_2 & kmask3) >> 2);
-
-            float acc1[4] = {0.f, 0.f, 0.f, 0.f};
-            float acc2[4] = {0.f, 0.f, 0.f, 0.f};
-            for (int i = 0; i < 8; i += 2) {
-                uint16_t q1 = u8BufToU16(inA[blk_idx + row_idx].qs, 32 * iq + 8 * ir + i);
-                uint16_t q2 = u8BufToU16(inA[blk_idx + row_idx].qs, 64 + 32 * iq + 8 * ir + i);
-                acc1[0] += yl[i+0] * (q1 & 0x000F);
-                acc1[1] += yl[i+1] * (q1 & 0x0F00);
-                acc1[2] += yl[i+8] * (q1 & 0x00F0);
-                acc1[3] += yl[i+9] * (q1 & 0xF000);
-                acc2[0] += yh[i+0] * (q2 & 0x000F);
-                acc2[1] += yh[i+1] * (q2 & 0x0F00);
-                acc2[2] += yh[i+8] * (q2 & 0x00F0);
-                acc2[3] += yh[i+9] * (q2 & 0xF000);
-            }
-
-            uint8_t sc8_0 = uint8_t(sc16[0] & 0xFF);
-            uint8_t sc8_1 = uint8_t(sc16[0] >> 8 );
-            uint8_t sc8_2 = uint8_t(sc16[1] & 0xFF);
-            uint8_t sc8_3 = uint8_t(sc16[1] >> 8 );
-            uint8_t sc8_4 = uint8_t(sc16[2] & 0xFF);
-            uint8_t sc8_5 = uint8_t(sc16[2] >> 8 );
-            uint8_t sc8_6 = uint8_t(sc16[3] & 0xFF);
-            uint8_t sc8_7 = uint8_t(sc16[3] >> 8 );
-
-            float dall = float(inA[blk_idx + row_idx].d);
-            float dmin = float(inA[blk_idx + row_idx].dmin);
-            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 +
-                               (acc2[2] + 1.f/256.f * acc2[3]) * sc8_5 * 1.f/16.f) -
-                dmin * (sumy[0] * sc8_2 + sumy[1] * sc8_3 + sumy[2] * sc8_6 + sumy[3] * sc8_7);
-        }
-
-        y4 += 4 * QK_K;
-    }
-
-    for (int row = 0; row < N_DST; ++row) {
-        all_sum = subgroupAdd(sumf[row]);
-        if (subgroupElect()) {
-            out_[r1*pcs.ne0 + im*pcs.ne0*pcs.ne1 + first_row + row + pcs.outOff] = all_sum;
-        }
-    }
-}
diff --git a/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q6_k.comp b/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q6_k.comp
deleted file mode 100644
index d331d1a70..000000000
--- a/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q6_k.comp
+++ /dev/null
@@ -1,106 +0,0 @@
-#version 450
-
-#include "common.comp"
-
-#define SIZE_OF_BLOCK sizeof_block_q6_k
-
-layout(local_size_x_id = 0) in;
-layout(local_size_y_id = 1) in;
-layout(local_size_z = 1) in;
-
-layout (binding = 0) readonly buffer tensorInA { uint8_t inA[]; };
-layout (binding = 1) readonly buffer tensorInB { float inB[]; };
-layout (binding = 2) writeonly buffer tensorOut { float out_[]; };
-
-layout (push_constant) uniform parameter {
-    uint inAOff;
-    uint inBOff;
-    uint outOff;
-    int ne00;
-    int ne10;
-    int ne0;
-    int ne1;
-    int ne01;
-    int ne02;
-    int ne12;
-    uint nb01;
-    uint nb02;
-    uint nb03;
-    uint nb11;
-    uint nb12;
-    uint nb13;
-    uint r2;
-    uint r3;
-} pcs;
-
-void main() {
-    const uint8_t kmask1 = uint8_t(0x03);
-    const uint8_t kmask2 = uint8_t(0x0C);
-    const uint8_t kmask3 = uint8_t(0x30);
-    const uint8_t kmask4 = uint8_t(0xC0);
-
-    const uint nb = pcs.ne00/QK_K;
-
-    const uint r0 = gl_WorkGroupID.x;
-    const uint r1 = gl_WorkGroupID.y;
-    const uint im = gl_WorkGroupID.z;
-
-    const uint row = (r0 * gl_NumSubgroups + gl_SubgroupID);
-
-    const uint i12 = im%pcs.ne12;
-    const uint i13 = im/pcs.ne12;
-
-    const uint x = row*(pcs.nb01/SIZE_OF_BLOCK) + (i12/pcs.r2)*(pcs.nb02/SIZE_OF_BLOCK) + (i13/pcs.r3)*(pcs.nb03/SIZE_OF_BLOCK);
-    const uint yy = (r1*pcs.nb11 + i12*pcs.nb12 + i13*pcs.nb13) / 4 + pcs.inBOff;
-
-    float sumf = 0;
-
-    // bits of invocation ID for gl_SubgroupSize=32:
-    //  x   x   x   x   x
-    //  4   3   2   1   0
-    // (     tid     ) ix
-    //  ip (   il    )
-
-    const uint block_stride = gl_SubgroupSize / 16;         // number of blocks each subgroup processes
-    const uint tid  = gl_SubgroupInvocationID/block_stride; // first block_stride groups have tid=0
-    const uint ix   = gl_SubgroupInvocationID%block_stride; // first block is 0..block_stride-1
-    const uint ip   = tid/8;        // first or second half of block (0 or 1)
-    const uint il   = tid%8;        // each half has 8 parts, one per scale
-    const uint n    = 4;            // 4 scales at a time (and 4 sums)
-    const uint l0   = n*il;         // offset into half-block, 0..28
-    const uint is   = 8*ip + l0/16; // 0, 1, 8, 9
-
-    const uint y_offset = 128*ip + l0;
-    const uint q_offset_l = 64*ip + l0;
-    const uint q_offset_h = 32*ip + l0;
-
-    for (uint i = ix; i < nb; i += block_stride) {
-
-        const uint baseIndex = (x + i) * SIZE_OF_BLOCK + pcs.inAOff;
-
-        const uint qlIndex = q_offset_l;
-        const uint q2Index = qlIndex + QK_K/8;
-        const uint qhIndex = q_offset_h;
-        const uint y = yy + i * QK_K + y_offset;
-
-        float sums[4] = {0.0f, 0.0f, 0.0f, 0.0f};
-        for (uint l = 0; l < n; ++l) {
-            const uint8_t currentQ1 = inA[baseIndex + qlIndex + l];
-            const uint8_t currentQ2 = inA[baseIndex + q2Index + l];
-            const uint8_t currentQh = inA[baseIndex + QK_K/2 + qhIndex + l];
-
-            sums[0] += inB[y+l+ 0] * (int8_t((currentQ1 & 0xF) | ((currentQh & kmask1) << 4)) - 32);
-            sums[1] += inB[y+l+32] * (int8_t((currentQ2 & 0xF) | ((currentQh & kmask2) << 2)) - 32);
-            sums[2] += inB[y+l+64] * (int8_t((currentQ1  >> 4) | ((currentQh & kmask3) << 0)) - 32);
-            sums[3] += inB[y+l+96] * (int8_t((currentQ2  >> 4) | ((currentQh & kmask4) >> 2)) - 32);
-        }
-
-        float d = u8BufToFloat16(inA, baseIndex + QK_K/2 + QK_K/4 + QK_K/16);
-        sumf += d * (sums[0] * int8_t(inA[baseIndex + QK_K/2 + QK_K/4 + is]) + sums[1] * int8_t(inA[baseIndex + QK_K/2 + QK_K/4 + 2 + is]) + sums[2] * int8_t(inA[baseIndex + QK_K/2 + QK_K/4 + 4 + is]) + sums[3] * int8_t(inA[baseIndex + QK_K/2 + QK_K/4 + 6 + is]));
-    }
-
-    const float tot = subgroupAdd(sumf);
-    if (subgroupElect()) {
-        out_[r1*pcs.ne0 + im*pcs.ne0*pcs.ne1 + row + pcs.outOff] = tot;
-    }
-}
diff --git a/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q8_0.comp b/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q8_0.comp
deleted file mode 100644
index 34d015e90..000000000
--- a/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q8_0.comp
+++ /dev/null
@@ -1,73 +0,0 @@
-#version 450
-
-#include "common.comp"
-
-#include "op_mul_mv_q_n_pre.comp"
-
-#define SIZE_OF_D 2
-
-#define N_DST 4 // each SIMD group works on 4 rows
-#define N_SIMDGROUP 2 // number of SIMD groups in a thread group
-#define N_SIMDWIDTH 32 // assuming SIMD group size is 32
-
-#define NB_Q8_0 8
-
-void main() {
-    // NB: hack to make compatible with AMD GPUs that have a subgroup size of 64
-    if (gl_SubgroupInvocationID > 31)
-        return;
-
-    const int nr  = N_DST;
-    const int nsg = N_SIMDGROUP;
-    const int nw  = N_SIMDWIDTH;
-
-    const int nb = pcs.ne00/QK8_0;
-    const uint r0 = gl_WorkGroupID.x;
-    const uint r1 = gl_WorkGroupID.y;
-    const uint im = gl_WorkGroupID.z;
-
-    const uint first_row = (r0 * nsg + gl_SubgroupID) * nr;
-
-    const uint i12 = im%pcs.ne12;
-    const uint i13 = im/pcs.ne12;
-
-    const uint offset0 = first_row * nb + (i12/pcs.r2)*(nb*pcs.ne01) + (i13/pcs.r3)*(nb*pcs.ne01*pcs.ne02);
-
-    const uint x = offset0*sizeof_block_q8_0 + pcs.inAOff; // Based from inA
-    const uint y = r1*pcs.ne10 + im*pcs.ne00*pcs.ne1 + pcs.inBOff; // based from inB
-
-    float yl[NB_Q8_0];
-    float sumf[N_DST]={0.f, 0.f, 0.f, 0.f};
-
-    const uint ix = gl_SubgroupInvocationID.x/4;
-    const uint il = gl_SubgroupInvocationID.x%4;
-
-    uint yb = y + ix * QK8_0 + NB_Q8_0*il;
-
-    // each thread in a SIMD group deals with NB_Q8_0 quants at a time
-    for (uint ib = ix; ib < nb; ib += nw/4) {
-        for (int i = 0; i < NB_Q8_0; ++i) {
-            yl[i] = inB[yb + i];
-        }
-
-        for (int row = 0; row < nr; row++) {
-            const uint block_offset = (ib+row*nb) * sizeof_block_q8_0;
-            float sumq = 0.f;
-            for (int iq = 0; iq < NB_Q8_0; ++iq) {
-                const int8_t qs_iq = int8_t(inA[x + block_offset + SIZE_OF_D + NB_Q8_0*il + iq]);
-                sumq += qs_iq * yl[iq];
-            }
-            const float16_t d = u8BufToFloat16(inA, x + block_offset);
-            sumf[row] += sumq*d;
-        }
-
-        yb += NB_Q8_0 * nw;
-    }
-
-    for (int row = 0; row < nr; ++row) {
-        const float tot = subgroupAdd(sumf[row]);
-        if (subgroupElect() && first_row + row < pcs.ne01) {
-            out_[r1*pcs.ne0 + im*pcs.ne0*pcs.ne1 + first_row + row] = tot;
-        }
-    }
-}
diff --git a/ggml/src/ggml-kompute/kompute-shaders/op_mul_mv_q_n.comp b/ggml/src/ggml-kompute/kompute-shaders/op_mul_mv_q_n.comp
deleted file mode 100644
index a6517cc1f..000000000
--- a/ggml/src/ggml-kompute/kompute-shaders/op_mul_mv_q_n.comp
+++ /dev/null
@@ -1,52 +0,0 @@
-void main() {
-    // NB: hack to make compatible with AMD GPUs that have a subgroup size of 64
-    if (gl_SubgroupInvocationID > 31)
-        return;
-
-    const uint nb = uint(pcs.ne00/BLOCKS_IN_QUANT);
-
-    const uint r0 = gl_WorkGroupID.x;
-    const uint r1 = gl_WorkGroupID.y;
-    const uint im = gl_WorkGroupID.z;
-
-    const uint first_row = (r0 * gl_NumSubgroups + gl_SubgroupID) * N_ROWS;
-
-    const uint i12 = im%pcs.ne12;
-    const uint i13 = im/pcs.ne12;
-
-    // pointers to src0 rows
-    uint ax[N_ROWS];
-    for (int row = 0; row < N_ROWS; ++row) {
-        const uint offset0 = (first_row + row)*(pcs.nb01/SIZE_OF_BLOCK) + (i12/pcs.r2)*(pcs.nb02/SIZE_OF_BLOCK) + (i13/pcs.r3)*(pcs.nb03/SIZE_OF_BLOCK);
-
-        ax[row] = offset0 + pcs.inAOff;
-    }
-
-    const uint y = (r1*pcs.nb11 + i12*pcs.nb12 + i13*pcs.nb13) / 4 + pcs.inBOff;
-
-    float sumf[N_ROWS] = {0.0f, 0.0f, 0.0f, 0.0f};
-
-    const uint ix = gl_SubgroupInvocationID/2;
-    const uint il = (BLOCKS_IN_QUANT/4)*(gl_SubgroupInvocationID%2);
-
-    uint yb = y + ix * BLOCKS_IN_QUANT + il;
-
-    //debugPrintfEXT("gl_NumSubgroups=%d, gl_SubgroupID=%d, gl_SubgroupInvocationID=%d, glSubgroupSize=%d, gl_WorkGroupSize.x=%d, gl_WorkGroupSize.y=%d, gl_WorkGroupSize.z=%d\n",
-    //    gl_NumSubgroups, gl_SubgroupID, gl_SubgroupInvocationID, gl_SubgroupSize,
-    //    gl_WorkGroupSize.x, gl_WorkGroupSize.y, gl_WorkGroupSize.z);
-
-    for (uint ib = ix; ib < nb; ib += 16) {
-        for (int row = 0; row < N_ROWS; row++) {
-            sumf[row] += block_q_n_dot_y(ax[row] + ib, yb, il);
-        }
-
-        yb += BLOCKS_IN_QUANT * 16;
-    }
-
-    for (int row = 0; row < N_ROWS; ++row) {
-        const float tot = subgroupAdd(sumf[row]);
-        if (first_row + row < pcs.ne01 && subgroupElect()) {
-            out_[r1*pcs.ne0 + im*pcs.ne0*pcs.ne1 + first_row + row + pcs.outOff] = tot;
-        }
-    }
-}
diff --git a/ggml/src/ggml-kompute/kompute-shaders/op_mul_mv_q_n_pre.comp b/ggml/src/ggml-kompute/kompute-shaders/op_mul_mv_q_n_pre.comp
deleted file mode 100644
index a9a2f2218..000000000
--- a/ggml/src/ggml-kompute/kompute-shaders/op_mul_mv_q_n_pre.comp
+++ /dev/null
@@ -1,28 +0,0 @@
-layout(local_size_x_id = 0) in;
-layout(local_size_y = 8) in;
-layout(local_size_z = 1) in;
-
-layout (binding = 0) readonly buffer tensorInA { uint8_t inA[]; };
-layout (binding = 1) readonly buffer tensorInB { float inB[]; };
-layout (binding = 2) writeonly buffer tensorOut { float out_[]; };
-
-layout (push_constant) uniform parameter {
-    uint inAOff;
-    uint inBOff;
-    uint outOff;
-    int  ne00;
-    int  ne01;
-    int  ne02;
-    int  ne10;
-    int  ne12;
-    int  ne0;
-    int  ne1;
-    uint nb01;
-    uint nb02;
-    uint nb03;
-    uint nb11;
-    uint nb12;
-    uint nb13;
-    uint r2;
-    uint r3;
-} pcs;
diff --git a/ggml/src/ggml-kompute/kompute-shaders/op_norm.comp b/ggml/src/ggml-kompute/kompute-shaders/op_norm.comp
deleted file mode 100644
index ad0c3c01b..000000000
--- a/ggml/src/ggml-kompute/kompute-shaders/op_norm.comp
+++ /dev/null
@@ -1,84 +0,0 @@
-#version 450
-
-#include "common.comp"
-
-layout(local_size_x = 256) in;
-
-layout(binding = 0) buffer restrict readonly tensorIn { float in_[]; };
-layout(binding = 1) buffer restrict tensorOut { float out_[]; };
-
-layout(push_constant) uniform PushConstants {
-    uint inOff;
-    uint outOff;
-    uint ne00;
-    uint nb01;
-    float eps;
-} pcs;
-
-shared float sum[gl_WorkGroupSize.x];
-
-void main() {
-    const uint x = (gl_WorkGroupID.x*pcs.nb01/4) + pcs.inOff; // Based from in_
-    // MEAN
-    // parallel sum
-    sum[gl_LocalInvocationID.x] = 0.0;
-    for (uint i00 = gl_LocalInvocationID.x; i00 < pcs.ne00; i00 += gl_WorkGroupSize.x) {
-        sum[gl_LocalInvocationID.x] += in_[x+i00];
-    }
-
-    // reduce
-    barrier();
-    memoryBarrierShared();
-    [[unroll]] for (uint i = gl_WorkGroupSize.x/2; i > 0; i /= 2) {
-        if (gl_LocalInvocationID.x < i) {
-            sum[gl_LocalInvocationID.x] += sum[gl_LocalInvocationID.x + i];
-        }
-        barrier();
-        memoryBarrierShared();
-    }
-
-    // broadcast
-    if (gl_LocalInvocationID.x == 0) {
-        sum[0] /= float(pcs.ne00);
-    }
-    barrier();
-    memoryBarrierShared();
-    const float mean = sum[0];
-
-    // recenter
-    const uint y = (gl_WorkGroupID.x*pcs.ne00) + pcs.outOff; // Based from out_
-    for (uint i00 = gl_LocalInvocationID.x; i00 < pcs.ne00; i00 += gl_WorkGroupSize.x) {
-        out_[y+i00] = in_[x+i00] - mean;
-    }
-
-    // VARIANCE
-    // parallel sum
-    sum[gl_LocalInvocationID.x] = 0.0;
-    for (uint i00 = gl_LocalInvocationID.x; i00 < pcs.ne00; i00 += gl_WorkGroupSize.x) {
-        sum[gl_LocalInvocationID.x] += out_[y+i00] * out_[y+i00];
-    }
-
-    // reduce
-    barrier();
-    memoryBarrierShared();
-    [[unroll]] for (uint i = gl_WorkGroupSize.x/2; i > 0; i /= 2) {
-        if (gl_LocalInvocationID.x < i) {
-            sum[gl_LocalInvocationID.x] += sum[gl_LocalInvocationID.x + i];
-        }
-        barrier();
-        memoryBarrierShared();
-    }
-
-    // broadcast
-    if (gl_LocalInvocationID.x == 0) {
-        sum[0] /= float(pcs.ne00);
-    }
-    barrier();
-    memoryBarrierShared();
-    const float variance = sum[0];
-
-    const float scale = 1.0f/sqrt(variance + pcs.eps);
-    for (uint i00 = gl_LocalInvocationID.x; i00 < pcs.ne00; i00 += gl_WorkGroupSize.x) {
-        out_[y+i00] *= scale;
-    }
-}
diff --git a/ggml/src/ggml-kompute/kompute-shaders/op_relu.comp b/ggml/src/ggml-kompute/kompute-shaders/op_relu.comp
deleted file mode 100644
index 52a601fe6..000000000
--- a/ggml/src/ggml-kompute/kompute-shaders/op_relu.comp
+++ /dev/null
@@ -1,21 +0,0 @@
-#version 450
-
-#include "common.comp"
-
-layout(local_size_x = 1) in;
-
-layout(binding = 0) buffer restrict readonly tensorIn { float in_[]; };
-layout(binding = 1) buffer restrict writeonly tensorOut { float out_[]; };
-layout(push_constant) uniform PushConstants {
-    uint inOff;
-    uint outOff;
-} pcs;
-
-void main() {
-    const uint baseIndex = gl_WorkGroupID.x * 4;
-
-    for (uint x = 0; x < 4; x++) {
-        const uint i = baseIndex + x;
-        out_[i + pcs.outOff] = max(0.0, in_[i + pcs.inOff]);
-    }
-}
diff --git a/ggml/src/ggml-kompute/kompute-shaders/op_rmsnorm.comp b/ggml/src/ggml-kompute/kompute-shaders/op_rmsnorm.comp
deleted file mode 100644
index da658c160..000000000
--- a/ggml/src/ggml-kompute/kompute-shaders/op_rmsnorm.comp
+++ /dev/null
@@ -1,53 +0,0 @@
-#version 450
-
-#include "common.comp"
-
-layout(local_size_x = 512) in;
-
-layout(binding = 0) buffer restrict readonly tensorIn { float in_[]; };
-layout(binding = 1) buffer restrict tensorOut { float out_[]; };
-
-layout(push_constant) uniform PushConstants {
-    uint inOff;
-    uint outOff;
-    uint ne00;
-    uint nb01;
-    float eps;
-} pcs;
-
-shared float sum[gl_WorkGroupSize.x];
-
-void main() {
-    const uint x = (gl_WorkGroupID.x*pcs.nb01/4) + pcs.inOff; // Based from in_
-
-    // parallel sum
-    sum[gl_LocalInvocationID.x] = 0.0;
-    for (uint i00 = gl_LocalInvocationID.x; i00 < pcs.ne00; i00 += gl_WorkGroupSize.x) {
-        sum[gl_LocalInvocationID.x] += in_[x+i00] * in_[x+i00];
-    }
-
-    // reduce
-    barrier();
-    memoryBarrierShared();
-    [[unroll]] for (uint i = gl_WorkGroupSize.x/2; i > 0; i /= 2) {
-        if (gl_LocalInvocationID.x < i) {
-            sum[gl_LocalInvocationID.x] += sum[gl_LocalInvocationID.x + i];
-        }
-        barrier();
-        memoryBarrierShared();
-    }
-
-    // broadcast
-    if (gl_LocalInvocationID.x == 0) {
-        sum[0] /= float(pcs.ne00);
-    }
-    barrier();
-    memoryBarrierShared();
-
-    const float scale = 1.0f/sqrt(sum[0] + pcs.eps);
-
-    const uint y = (gl_WorkGroupID.x*pcs.ne00) + pcs.outOff; // Based from out_
-    for (uint i00 = gl_LocalInvocationID.x; i00 < pcs.ne00; i00 += gl_WorkGroupSize.x) {
-        out_[y+i00] = in_[x+i00] * scale;
-    }
-}
diff --git a/ggml/src/ggml-kompute/kompute-shaders/op_rope_neox_f16.comp b/ggml/src/ggml-kompute/kompute-shaders/op_rope_neox_f16.comp
deleted file mode 100644
index 63659cbfe..000000000
--- a/ggml/src/ggml-kompute/kompute-shaders/op_rope_neox_f16.comp
+++ /dev/null
@@ -1,52 +0,0 @@
-#version 450
-
-#include "rope_common.comp"
-
-layout(binding = 0) buffer restrict readonly  tensorInA { float16_t inA[]; };
-layout(binding = 1) buffer restrict readonly  tensorInB { int       inB[]; };
-layout(binding = 2) buffer restrict readonly  tensorInC { float     inC[]; };
-layout(binding = 3) buffer restrict writeonly tensorOut { float16_t out_[]; };
-
-void main() {
-    const uint i3 = gl_WorkGroupID.z;
-    const uint i2 = gl_WorkGroupID.y;
-    const uint i1 = gl_WorkGroupID.x;
-
-    float corr_dims[2];
-    rope_yarn_corr_dims(pcs.n_dims, pcs.n_ctx_orig, pcs.freq_base, pcs.beta_fast, pcs.beta_slow, corr_dims);
-
-    const float theta_scale = pow(pcs.freq_base, -2.0/pcs.n_dims);
-
-    float theta_base = float(inB[pcs.inBOff + i2]);
-    float inv_ndims = -1.f/pcs.n_dims;
-
-    float cos_theta;
-    float sin_theta;
-
-    for (uint i0 = 2*gl_LocalInvocationIndex; i0 < pcs.ne0; i0 += 2*gl_WorkGroupSize.x) {
-        if (i0 < pcs.n_dims) {
-            uint ic = i0/2;
-
-            float theta = theta_base * pow(pcs.freq_base, inv_ndims*i0);
-
-            const float freq_factor = pcs.has_freq_factors ? inC[pcs.inCOff + ic] : 1.0f;
-
-            rope_yarn(theta/freq_factor, pcs.freq_scale, corr_dims, i0, pcs.ext_factor, pcs.attn_factor, cos_theta, sin_theta);
-
-            const uint src      = uint((i3*pcs.nb03 + i2*pcs.nb02 + i1*pcs.nb01 + ic*pcs.nb00) / 2) + pcs.inAOff; // Based from in
-            const uint dst_data = uint((i3*pcs.nb3  + i2*pcs.nb2  + i1*pcs.nb1  + ic*pcs.nb0)  / 2) + pcs.outOff; // Based from out_
-
-            const float x0 = float(inA[src]);
-            const float x1 = float(inA[src+pcs.n_dims/2]);
-
-            out_[dst_data]              = float16_t(x0*cos_theta - x1*sin_theta);
-            out_[dst_data+pcs.n_dims/2] = float16_t(x0*sin_theta + x1*cos_theta);
-        } else {
-            const uint src      = uint((i3*pcs.nb03 + i2*pcs.nb02 + i1*pcs.nb01 + i0*pcs.nb00) / 2) + pcs.inAOff; // Based from in
-            const uint dst_data = uint((i3*pcs.nb3  + i2*pcs.nb2  + i1*pcs.nb1  + i0*pcs.nb0)  / 2) + pcs.outOff; // Based from out_
-
-            out_[dst_data]   = inA[src];
-            out_[dst_data+1] = inA[src+1];
-        }
-    }
-}
diff --git a/ggml/src/ggml-kompute/kompute-shaders/op_rope_neox_f32.comp b/ggml/src/ggml-kompute/kompute-shaders/op_rope_neox_f32.comp
deleted file mode 100644
index 4df56204d..000000000
--- a/ggml/src/ggml-kompute/kompute-shaders/op_rope_neox_f32.comp
+++ /dev/null
@@ -1,52 +0,0 @@
-#version 450
-
-#include "rope_common.comp"
-
-layout(binding = 0) buffer restrict readonly  tensorInA { float inA[]; };
-layout(binding = 1) buffer restrict readonly  tensorInB { int       inB[]; };
-layout(binding = 2) buffer restrict readonly  tensorInC { float inC[]; };
-layout(binding = 3) buffer restrict writeonly tensorOut { float out_[]; };
-
-void main() {
-    const uint i3 = gl_WorkGroupID.z;
-    const uint i2 = gl_WorkGroupID.y;
-    const uint i1 = gl_WorkGroupID.x;
-
-    float corr_dims[2];
-    rope_yarn_corr_dims(pcs.n_dims, pcs.n_ctx_orig, pcs.freq_base, pcs.beta_fast, pcs.beta_slow, corr_dims);
-
-    const float theta_scale = pow(pcs.freq_base, -2.0/pcs.n_dims);
-
-    float theta_base = float(inB[pcs.inBOff + i2]);
-    float inv_ndims = -1.f/pcs.n_dims;
-
-    float cos_theta;
-    float sin_theta;
-
-    for (uint i0 = 2*gl_LocalInvocationIndex; i0 < pcs.ne0; i0 += 2*gl_WorkGroupSize.x) {
-        if (i0 < pcs.n_dims) {
-            uint ic = i0/2;
-
-            float theta = theta_base * pow(pcs.freq_base, inv_ndims*i0);
-
-            const float freq_factor = pcs.has_freq_factors ? inC[pcs.inCOff + ic] : 1.0f;
-
-            rope_yarn(theta/freq_factor, pcs.freq_scale, corr_dims, i0, pcs.ext_factor, pcs.attn_factor, cos_theta, sin_theta);
-
-            const uint src      = uint((i3*pcs.nb03 + i2*pcs.nb02 + i1*pcs.nb01 + ic*pcs.nb00) / 4) + pcs.inAOff; // Based from in
-            const uint dst_data = uint((i3*pcs.nb3  + i2*pcs.nb2  + i1*pcs.nb1  + ic*pcs.nb0)  / 4) + pcs.outOff; // Based from out_
-
-            const float x0 = inA[src];
-            const float x1 = inA[src+pcs.n_dims/2];
-
-            out_[dst_data]              = x0*cos_theta - x1*sin_theta;
-            out_[dst_data+pcs.n_dims/2] = x0*sin_theta + x1*cos_theta;
-        } else {
-            const uint src      = uint((i3*pcs.nb03 + i2*pcs.nb02 + i1*pcs.nb01 + i0*pcs.nb00) / 4) + pcs.inAOff; // Based from in
-            const uint dst_data = uint((i3*pcs.nb3  + i2*pcs.nb2  + i1*pcs.nb1  + i0*pcs.nb0)  / 4) + pcs.outOff; // Based from out_
-
-            out_[dst_data]   = inA[src];
-            out_[dst_data+1] = inA[src+1];
-        }
-    }
-}
diff --git a/ggml/src/ggml-kompute/kompute-shaders/op_rope_norm_f16.comp b/ggml/src/ggml-kompute/kompute-shaders/op_rope_norm_f16.comp
deleted file mode 100644
index a3c0eda8b..000000000
--- a/ggml/src/ggml-kompute/kompute-shaders/op_rope_norm_f16.comp
+++ /dev/null
@@ -1,52 +0,0 @@
-#version 450
-
-#include "rope_common.comp"
-
-layout(binding = 0) buffer restrict readonly  tensorInA { float16_t inA[]; };
-layout(binding = 1) buffer restrict readonly  tensorInB { int       inB[]; };
-layout(binding = 2) buffer restrict readonly  tensorInC { float     inC[]; };
-layout(binding = 3) buffer restrict writeonly tensorOut { float16_t out_[]; };
-
-void main() {
-    const uint i3 = gl_WorkGroupID.z;
-    const uint i2 = gl_WorkGroupID.y;
-    const uint i1 = gl_WorkGroupID.x;
-
-    float corr_dims[2];
-    rope_yarn_corr_dims(pcs.n_dims, pcs.n_ctx_orig, pcs.freq_base, pcs.beta_fast, pcs.beta_slow, corr_dims);
-
-    const float theta_scale = pow(pcs.freq_base, -2.0/pcs.n_dims);
-
-    float theta_base = float(inB[pcs.inBOff + i2]);
-    float inv_ndims = -1.f/pcs.n_dims;
-
-    float cos_theta;
-    float sin_theta;
-
-    for (uint i0 = 2*gl_LocalInvocationIndex; i0 < pcs.ne0; i0 += 2*gl_WorkGroupSize.x) {
-        if (i0 < pcs.n_dims) {
-            uint ic = i0/2;
-
-            float theta = theta_base * pow(pcs.freq_base, inv_ndims*i0);
-
-            const float freq_factor = pcs.has_freq_factors ? inC[pcs.inCOff + ic] : 1.0f;
-
-            rope_yarn(theta/freq_factor, pcs.freq_scale, corr_dims, i0, pcs.ext_factor, pcs.attn_factor, cos_theta, sin_theta);
-
-            const uint src      = uint((i3*pcs.nb03 + i2*pcs.nb02 + i1*pcs.nb01 + i0*pcs.nb00) / 2) + pcs.inAOff; // Based from in
-            const uint dst_data = uint((i3*pcs.nb3  + i2*pcs.nb2  + i1*pcs.nb1  + i0*pcs.nb0)  / 2) + pcs.outOff; // Based from out_
-
-            const float x0 = float(inA[src]);
-            const float x1 = float(inA[src+1]);
-
-            out_[dst_data]   = float16_t(x0*cos_theta - x1*sin_theta);
-            out_[dst_data+1] = float16_t(x0*sin_theta + x1*cos_theta);
-        } else {
-            const uint src      = uint((i3*pcs.nb03 + i2*pcs.nb02 + i1*pcs.nb01 + i0*pcs.nb00) / 2) + pcs.inAOff; // Based from in
-            const uint dst_data = uint((i3*pcs.nb3  + i2*pcs.nb2  + i1*pcs.nb1  + i0*pcs.nb0)  / 2) + pcs.outOff; // Based from out_
-
-            out_[dst_data]   = inA[src];
-            out_[dst_data+1] = inA[src+1];
-        }
-    }
-}
diff --git a/ggml/src/ggml-kompute/kompute-shaders/op_rope_norm_f32.comp b/ggml/src/ggml-kompute/kompute-shaders/op_rope_norm_f32.comp
deleted file mode 100644
index b7963ae72..000000000
--- a/ggml/src/ggml-kompute/kompute-shaders/op_rope_norm_f32.comp
+++ /dev/null
@@ -1,52 +0,0 @@
-#version 450
-
-#include "rope_common.comp"
-
-layout(binding = 0) buffer restrict readonly  tensorInA { float inA[]; };
-layout(binding = 1) buffer restrict readonly  tensorInB { int   inB[]; };
-layout(binding = 2) buffer restrict readonly  tensorInC { float inC[]; };
-layout(binding = 3) buffer restrict writeonly tensorOut { float out_[]; };
-
-void main() {
-    const uint i3 = gl_WorkGroupID.z;
-    const uint i2 = gl_WorkGroupID.y;
-    const uint i1 = gl_WorkGroupID.x;
-
-    float corr_dims[2];
-    rope_yarn_corr_dims(pcs.n_dims, pcs.n_ctx_orig, pcs.freq_base, pcs.beta_fast, pcs.beta_slow, corr_dims);
-
-    const float theta_scale = pow(pcs.freq_base, -2.0/pcs.n_dims);
-
-    float theta_base = float(inB[pcs.inBOff + i2]);
-    float inv_ndims = -1.f/pcs.n_dims;
-
-    float cos_theta;
-    float sin_theta;
-
-    for (uint i0 = 2*gl_LocalInvocationIndex; i0 < pcs.ne0; i0 += 2*gl_WorkGroupSize.x) {
-        if (i0 < pcs.n_dims) {
-            uint ic = i0/2;
-
-            float theta = theta_base * pow(pcs.freq_base, inv_ndims*i0);
-
-            const float freq_factor = pcs.has_freq_factors ? inC[pcs.inCOff + ic] : 1.0f;
-
-            rope_yarn(theta/freq_factor, pcs.freq_scale, corr_dims, i0, pcs.ext_factor, pcs.attn_factor, cos_theta, sin_theta);
-
-            const uint src      = uint((i3*pcs.nb03 + i2*pcs.nb02 + i1*pcs.nb01 + i0*pcs.nb00) / 4) + pcs.inAOff; // Based from in
-            const uint dst_data = uint((i3*pcs.nb3  + i2*pcs.nb2  + i1*pcs.nb1  + i0*pcs.nb0)  / 4) + pcs.outOff; // Based from out_
-
-            const float x0 = inA[src];
-            const float x1 = inA[src+1];
-
-            out_[dst_data]   = x0*cos_theta - x1*sin_theta;
-            out_[dst_data+1] = x0*sin_theta + x1*cos_theta;
-        } else {
-            const uint src      = uint((i3*pcs.nb03 + i2*pcs.nb02 + i1*pcs.nb01 + i0*pcs.nb00) / 4) + pcs.inAOff; // Based from in
-            const uint dst_data = uint((i3*pcs.nb3  + i2*pcs.nb2  + i1*pcs.nb1  + i0*pcs.nb0)  / 4) + pcs.outOff; // Based from out_
-
-            out_[dst_data]   = inA[src];
-            out_[dst_data+1] = inA[src+1];
-        }
-    }
-}
diff --git a/ggml/src/ggml-kompute/kompute-shaders/op_scale.comp b/ggml/src/ggml-kompute/kompute-shaders/op_scale.comp
deleted file mode 100644
index bdae26738..000000000
--- a/ggml/src/ggml-kompute/kompute-shaders/op_scale.comp
+++ /dev/null
@@ -1,19 +0,0 @@
-#version 450
-
-#include "common.comp"
-
-layout(local_size_x = 1) in;
-
-layout(binding = 0) buffer restrict readonly tensorIn { float in_[]; };
-layout(binding = 1) buffer restrict writeonly tensorOut { float out_[]; };
-
-layout(push_constant) uniform PushConstants {
-    uint inOff;
-    uint outOff;
-    float scale;
-} pcs;
-
-void main() {
-    const uint i = gl_WorkGroupID.x;
-    out_[i + pcs.outOff] = in_[i + pcs.inOff] * pcs.scale;
-}
diff --git a/ggml/src/ggml-kompute/kompute-shaders/op_scale_8.comp b/ggml/src/ggml-kompute/kompute-shaders/op_scale_8.comp
deleted file mode 100644
index ada69754b..000000000
--- a/ggml/src/ggml-kompute/kompute-shaders/op_scale_8.comp
+++ /dev/null
@@ -1,23 +0,0 @@
-#version 450
-
-#include "common.comp"
-
-layout(local_size_x = 1) in;
-
-layout(binding = 0) buffer restrict readonly tensorIn { float in_[]; };
-layout(binding = 1) buffer restrict writeonly tensorOut { float out_[]; };
-
-layout(push_constant) uniform PushConstants {
-    uint inOff;
-    uint outOff;
-    float scale;
-} pcs;
-
-void main() {
-    const uint baseIndex = gl_WorkGroupID.x * 8;
-
-    for (uint x = 0; x < 8; x++) {
-        const uint i = baseIndex + x;
-        out_[i + pcs.outOff] = in_[i + pcs.inOff] * pcs.scale;
-    }
-}
diff --git a/ggml/src/ggml-kompute/kompute-shaders/op_silu.comp b/ggml/src/ggml-kompute/kompute-shaders/op_silu.comp
deleted file mode 100644
index 0fb8e4b74..000000000
--- a/ggml/src/ggml-kompute/kompute-shaders/op_silu.comp
+++ /dev/null
@@ -1,22 +0,0 @@
-#version 450
-
-#include "common.comp"
-
-layout(local_size_x = 1) in;
-
-layout(binding = 0) buffer restrict readonly tensorIn { float in_[]; };
-layout(binding = 1) buffer restrict writeonly tensorOut { float out_[]; };
-layout(push_constant) uniform PushConstants {
-    uint inOff;
-    uint outOff;
-} pcs;
-
-void main() {
-    const uint baseIndex = gl_WorkGroupID.x * 4;
-
-    for (uint x = 0; x < 4; x++) {
-        const uint i = baseIndex + x;
-        const float y = in_[i + pcs.inOff];
-        out_[i + pcs.outOff] = y / (1.0 + exp(-y));
-    }
-}
diff --git a/ggml/src/ggml-kompute/kompute-shaders/op_softmax.comp b/ggml/src/ggml-kompute/kompute-shaders/op_softmax.comp
deleted file mode 100644
index 4165295bf..000000000
--- a/ggml/src/ggml-kompute/kompute-shaders/op_softmax.comp
+++ /dev/null
@@ -1,72 +0,0 @@
-// TODO: implement multi-simd softmax (llama.cpp commit e16b9fa4)
-
-#version 450
-
-#include "common.comp"
-
-layout(local_size_x_id = 0) in;
-
-layout(binding = 0) buffer restrict readonly tensorInA { float inA[]; };
-layout(binding = 1) buffer restrict readonly tensorInB { float inB[]; };
-layout(binding = 2) buffer restrict writeonly tensorOut { float out_[]; };
-
-layout(push_constant) uniform PushConstants {
-    uint inAOff;
-    uint inBOff;
-    uint outOff;
-    int ne00;
-    int ne01;
-    int ne02;
-    float scale;
-    float max_bias;
-    float m0;
-    float m1;
-    uint n_head_log2;
-    int mask;
-} pcs;
-
-void main() {
-    if (gl_SubgroupInvocationID > 31)
-        return;
-
-    const uint i03 = gl_WorkGroupID.z;
-    const uint i02 = gl_WorkGroupID.y;
-    const uint i01 = gl_WorkGroupID.x;
-
-    const uint extra_off = i03*pcs.ne02*pcs.ne01*pcs.ne00 + i02*pcs.ne01*pcs.ne00 + i01*pcs.ne00;
-    const uint psrc0 = extra_off + pcs.inAOff; // Based from inA
-    const uint pmask = i01*pcs.ne00 + pcs.inBOff; // Based from inB
-    const uint pdst = extra_off + pcs.outOff; // Based from out_
-
-    float slope = 1.0f;
-
-    // ALiBi
-    if (pcs.max_bias > 0.0f) {
-        int64_t h = i02;
-
-        float base = h < pcs.n_head_log2 ? pcs.m0 : pcs.m1;
-        int64_t exp = h < pcs.n_head_log2 ? h + 1 : 2*(h - pcs.n_head_log2) + 1;
-
-        slope = pow(base, float(exp));
-    }
-
-    // parallel max
-    float localMax = uintBitsToFloat(0xFF800000);
-    for (uint i00 = gl_SubgroupInvocationID.x; i00 < pcs.ne00; i00 += 32) {
-        localMax = max(localMax, inA[psrc0 + i00]*pcs.scale + (pcs.mask!=0 ? slope*inB[pmask + i00] : 0.0f));
-    }
-    float max_ = subgroupMax(localMax);
-
-    // parallel sum
-    float localSum = 0.0f;
-    for (uint i00 = gl_SubgroupInvocationID.x; i00 < pcs.ne00; i00 += 32) {
-        const float exp_psrc0 = exp(inA[psrc0 + i00]*pcs.scale + (pcs.mask!=0 ? slope*inB[pmask + i00] : 0.0f) - max_);
-        localSum += exp_psrc0;
-        out_[pdst + i00] = exp_psrc0;
-    }
-
-    const float sum = subgroupAdd(localSum);
-    for (uint i00 = gl_SubgroupInvocationID.x; i00 < pcs.ne00; i00 += 32) {
-        out_[pdst + i00] /= sum;
-    }
-}
diff --git a/ggml/src/ggml-kompute/kompute-shaders/rope_common.comp b/ggml/src/ggml-kompute/kompute-shaders/rope_common.comp
deleted file mode 100644
index 0fca640dc..000000000
--- a/ggml/src/ggml-kompute/kompute-shaders/rope_common.comp
+++ /dev/null
@@ -1,71 +0,0 @@
-#include "common.comp"
-
-#define GGML_ROPE_TYPE_NEOX 2
-
-// TODO: use a local size of 32 or more (Metal uses 1024)
-layout(local_size_x = 1) in;
-
-layout (push_constant) uniform parameter {
-    uint inAOff;
-    uint inBOff;
-    uint inCOff;
-    uint outOff;
-    int n_dims;
-    int mode;
-    int n_ctx_orig;
-    float freq_base;
-    float freq_scale;
-    bool has_freq_factors;
-    float ext_factor;
-    float attn_factor;
-    float beta_fast;
-    float beta_slow;
-    uint nb00;
-    uint nb01;
-    uint nb02;
-    uint nb03;
-    int ne0;
-    uint nb0;
-    uint nb1;
-    uint nb2;
-    uint nb3;
-} pcs;
-
-float rope_yarn_ramp(const float low, const float high, const float i0) {
-    const float y = (i0 / 2 - low) / max(0.001f, high - low);
-    return 1.0f - min(1.0f, max(0.0f, y));
-}
-
-// YaRN algorithm based on LlamaYaRNScaledRotaryEmbedding.py from https://github.com/jquesnelle/yarn
-// MIT licensed. Copyright (c) 2023 Jeffrey Quesnelle and Bowen Peng.
-void rope_yarn(
-    float theta_extrap, float freq_scale, float corr_dims[2], float i0, float ext_factor, float mscale,
-    out float cos_theta, out 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 * log(1.0f / freq_scale);
-    }
-    cos_theta = cos(theta) * mscale;
-    sin_theta = sin(theta) * mscale;
-}
-
-// Apparently solving `n_rot = 2pi * x * base^((2 * max_pos_emb) / n_dims)` for x, we get
-// `corr_fac(n_rot) = n_dims * log(max_pos_emb / (n_rot * 2pi)) / (2 * log(base))`
-float rope_yarn_corr_factor(int n_dims, int n_ctx_orig, float n_rot, float base) {
-    return n_dims * log(n_ctx_orig / (n_rot * TWOPI_F)) / (2 * log(base));
-}
-
-void rope_yarn_corr_dims(
-    int n_dims, int n_ctx_orig, float freq_base, float beta_fast, float beta_slow, out float dims[2]
-) {
-    // start and end correction dims
-    dims[0] = max(0.0f,         floor(rope_yarn_corr_factor(n_dims, n_ctx_orig, beta_fast, freq_base)));
-    dims[1] = min(n_dims - 1.0f, ceil(rope_yarn_corr_factor(n_dims, n_ctx_orig, beta_slow, freq_base)));
-}
diff --git a/ggml/src/ggml-metal/CMakeLists.txt b/ggml/src/ggml-metal/CMakeLists.txt
index e22232780..0ca8a3c55 100644
--- a/ggml/src/ggml-metal/CMakeLists.txt
+++ b/ggml/src/ggml-metal/CMakeLists.txt
@@ -44,21 +44,22 @@ if (GGML_METAL_EMBED_LIBRARY)
     set(METALLIB_SOURCE_EMBED_TMP "${CMAKE_BINARY_DIR}/autogenerated/ggml-metal-embed.metal.tmp")
 
     add_custom_command(
-        OUTPUT ${METALLIB_EMBED_ASM}
+        OUTPUT "${METALLIB_EMBED_ASM}"
         COMMAND echo "Embedding Metal library"
-        COMMAND sed -e '/__embed_ggml-common.h__/r         ${METALLIB_COMMON}' -e '/__embed_ggml-common.h__/d'         < ${METALLIB_SOURCE}           > ${METALLIB_SOURCE_EMBED_TMP}
-        COMMAND sed -e '/\#include \"ggml-metal-impl.h\"/r ${METALLIB_IMPL}'   -e '/\#include \"ggml-metal-impl.h\"/d' < ${METALLIB_SOURCE_EMBED_TMP} > ${METALLIB_SOURCE_EMBED}
-        COMMAND echo ".section __DATA,__ggml_metallib"          >  ${METALLIB_EMBED_ASM}
-        COMMAND echo ".globl _ggml_metallib_start"              >> ${METALLIB_EMBED_ASM}
-        COMMAND echo "_ggml_metallib_start:"                    >> ${METALLIB_EMBED_ASM}
-        COMMAND echo ".incbin \\\"${METALLIB_SOURCE_EMBED}\\\"" >> ${METALLIB_EMBED_ASM}
-        COMMAND echo ".globl _ggml_metallib_end"                >> ${METALLIB_EMBED_ASM}
-        COMMAND echo "_ggml_metallib_end:"                      >> ${METALLIB_EMBED_ASM}
+        COMMAND sed -e "/__embed_ggml-common.h__/r ${METALLIB_COMMON}"       -e "/__embed_ggml-common.h__/d"         < "${METALLIB_SOURCE}"           > "${METALLIB_SOURCE_EMBED_TMP}"
+        COMMAND sed -e "/\#include \"ggml-metal-impl.h\"/r ${METALLIB_IMPL}" -e "/\#include \"ggml-metal-impl.h\"/d" < "${METALLIB_SOURCE_EMBED_TMP}" > "${METALLIB_SOURCE_EMBED}"
+        COMMAND echo ".section __DATA,__ggml_metallib"          >  "${METALLIB_EMBED_ASM}"
+        COMMAND echo ".globl _ggml_metallib_start"              >> "${METALLIB_EMBED_ASM}"
+        COMMAND echo "_ggml_metallib_start:"                    >> "${METALLIB_EMBED_ASM}"
+        COMMAND echo .incbin "\"${METALLIB_SOURCE_EMBED}\""     >> "${METALLIB_EMBED_ASM}"
+        COMMAND echo ".globl _ggml_metallib_end"                >> "${METALLIB_EMBED_ASM}"
+        COMMAND echo "_ggml_metallib_end:"                      >> "${METALLIB_EMBED_ASM}"
         DEPENDS ../ggml-common.h ggml-metal.metal ggml-metal-impl.h
         COMMENT "Generate assembly for embedded Metal library"
+        VERBATIM
     )
 
-    target_sources(ggml-metal PRIVATE ${METALLIB_EMBED_ASM})
+    target_sources(ggml-metal PRIVATE "${METALLIB_EMBED_ASM}")
 else()
     if (GGML_METAL_SHADER_DEBUG)
         # custom command to do the following:
@@ -70,7 +71,9 @@ else()
         # note: adding -fno-inline fixes the tests when using MTL_SHADER_VALIDATION=1
         # note: unfortunately, we have to call it default.metallib instead of ggml.metallib
         #       ref: https://github.com/ggerganov/whisper.cpp/issues/1720
-        set(XC_FLAGS -fno-fast-math -fno-inline -g)
+        # note: adding -g causes segmentation fault during compile
+        #set(XC_FLAGS -fno-fast-math -fno-inline -g)
+        set(XC_FLAGS -fno-fast-math -fno-inline)
     else()
         set(XC_FLAGS -O3)
     endif()
@@ -89,7 +92,7 @@ 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 - |
-            xcrun -sdk macosx metallib - -o ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
+                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 ${METALLIB_COMMON}
diff --git a/ggml/src/ggml-metal/ggml-metal-impl.h b/ggml/src/ggml-metal/ggml-metal-impl.h
index 17eab976f..752d55c21 100644
--- a/ggml/src/ggml-metal/ggml-metal-impl.h
+++ b/ggml/src/ggml-metal/ggml-metal-impl.h
@@ -229,7 +229,11 @@ typedef struct {
     uint64_t nb21;
     uint64_t nb22;
     uint64_t nb23;
+    int32_t  ne32;
+    int32_t  ne33;
     uint64_t nb31;
+    uint64_t nb32;
+    uint64_t nb33;
     int32_t  ne1;
     int32_t  ne2;
     float    scale;
@@ -422,6 +426,17 @@ typedef struct {
     int32_t  KHW; // KH * KW, pre-computed on CPU to save GPU resources
 } ggml_metal_kargs_im2col;
 
+typedef struct{
+    int32_t  ne00;
+    uint64_t nb01;
+    int32_t  ne10;
+    uint64_t nb11;
+    int32_t  ne0;
+    uint64_t nb1;
+    int32_t  i00;
+    int32_t  i10;
+} ggml_metal_kargs_glu;
+
 typedef struct {
     int64_t  ne00;
     int64_t  ne01;
@@ -450,9 +465,21 @@ typedef struct {
 } ggml_metal_kargs_sum_rows;
 
 typedef struct {
-    int64_t  ne00;
-    int64_t  ne01;
-    int64_t  ne02;
+    int32_t  ne00;
+    int32_t  ne01;
+    int32_t  ne02;
+    uint64_t nb01;
+    uint64_t nb02;
+    uint64_t nb03;
+    int32_t  ne11;
+    int32_t  ne12;
+    int32_t  ne13;
+    uint64_t nb11;
+    uint64_t nb12;
+    uint64_t nb13;
+    uint64_t nb1;
+    uint64_t nb2;
+    uint64_t nb3;
     float    scale;
     float    max_bias;
     float    m0;
@@ -488,26 +515,25 @@ typedef struct {
 typedef struct {
     int64_t  d_state;
     int64_t  d_inner;
+    int64_t  n_head;
+    int64_t  n_group;
     int64_t  n_seq_tokens;
     int64_t  n_seqs;
-    uint64_t nb00;
     uint64_t nb01;
     uint64_t nb02;
-    uint64_t nb10;
+    uint64_t nb03;
     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 nb43;
     uint64_t nb51;
     uint64_t nb52;
+    uint64_t nb53;
 } ggml_metal_kargs_ssm_scan;
 
 typedef struct {
@@ -521,6 +547,22 @@ typedef struct {
     uint64_t nb2;
 } ggml_metal_kargs_get_rows;
 
+typedef struct {
+    int32_t  nk0;
+    int32_t  ne01;
+    uint64_t nb01;
+    uint64_t nb02;
+    uint64_t nb03;
+    int32_t  ne11;
+    int32_t  ne12;
+    uint64_t nb10;
+    uint64_t nb11;
+    uint64_t nb12;
+    uint64_t nb1;
+    uint64_t nb2;
+    uint64_t nb3;
+} ggml_metal_kargs_set_rows;
+
 typedef struct {
     int64_t  ne00;
     int64_t  ne01;
diff --git a/ggml/src/ggml-metal/ggml-metal.m b/ggml/src/ggml-metal/ggml-metal.m
index f78e7eee5..40fc315e8 100644
--- a/ggml/src/ggml-metal/ggml-metal.m
+++ b/ggml/src/ggml-metal/ggml-metal.m
@@ -48,22 +48,28 @@ static struct ggml_backend_metal_device_context {
     int            mtl_device_ref_count;
     id<MTLLibrary> mtl_library;
 
+    NSLock * mtl_lock;
+
     bool has_simdgroup_reduction;
     bool has_simdgroup_mm;
     bool has_residency_sets;
     bool has_bfloat;
     bool use_bfloat;
 
+    size_t max_size;
+
     char name[128];
 } g_ggml_ctx_dev_main = {
     /*.mtl_device              =*/ nil,
     /*.mtl_device_ref_count    =*/ 0,
     /*.mtl_library             =*/ nil,
+    /*.mtl_lock                =*/ nil,
     /*.has_simdgroup_reduction =*/ false,
     /*.has_simdgroup_mm        =*/ false,
     /*.has_residency_sets      =*/ false,
     /*.has_bfloat              =*/ false,
     /*.use_bfloat              =*/ false,
+    /*.max_size                =*/ 0,
     /*.name                    =*/ "",
 };
 
@@ -71,6 +77,10 @@ static struct ggml_backend_metal_device_context {
 static id<MTLDevice> ggml_backend_metal_device_acq(struct ggml_backend_metal_device_context * ctx) {
     assert(ctx != NULL);
 
+    if (ctx->mtl_lock == nil) {
+        ctx->mtl_lock = [[NSLock alloc] init];
+    }
+
     if (ctx->mtl_device == nil) {
         ctx->mtl_device = MTLCreateSystemDefaultDevice();
     }
@@ -94,6 +104,8 @@ static id<MTLDevice> ggml_backend_metal_device_acq(struct ggml_backend_metal_dev
         ctx->use_bfloat = false;
 #endif
 
+        ctx->max_size = ctx->mtl_device.maxBufferLength;
+
         strncpy(ctx->name, [[ctx->mtl_device name] UTF8String], sizeof(ctx->name) - 1);
     }
 
@@ -110,6 +122,11 @@ 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_lock) {
+            [ctx->mtl_lock release];
+            ctx->mtl_lock = nil;
+        }
+
         if (ctx->mtl_library) {
             [ctx->mtl_library release];
             ctx->mtl_library = nil;
@@ -185,20 +202,33 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ4_NL,
     GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ4_XS,
     GGML_METAL_KERNEL_TYPE_GET_ROWS_I32,
+    GGML_METAL_KERNEL_TYPE_SET_ROWS_F32,
+    GGML_METAL_KERNEL_TYPE_SET_ROWS_F16,
+    GGML_METAL_KERNEL_TYPE_SET_ROWS_BF16,
+    GGML_METAL_KERNEL_TYPE_SET_ROWS_Q8_0,
+    GGML_METAL_KERNEL_TYPE_SET_ROWS_Q4_0,
+    GGML_METAL_KERNEL_TYPE_SET_ROWS_Q4_1,
+    GGML_METAL_KERNEL_TYPE_SET_ROWS_Q5_0,
+    GGML_METAL_KERNEL_TYPE_SET_ROWS_Q5_1,
+    GGML_METAL_KERNEL_TYPE_SET_ROWS_IQ4_NL,
     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_SSM_SCAN_F32_GROUP,
     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_F32_F32_C4,
     GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32,
+    GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_C4,
     GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_1ROW,
     GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_L4,
     GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F16,
     GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_F32,
+    GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_F32_C4,
     GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_F32_1ROW,
     GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_F32_L4,
     GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_BF16,
@@ -497,7 +527,13 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_SIN,
     GGML_METAL_KERNEL_TYPE_COS,
     GGML_METAL_KERNEL_TYPE_NEG,
+    GGML_METAL_KERNEL_TYPE_REGLU,
+    GGML_METAL_KERNEL_TYPE_GEGLU,
+    GGML_METAL_KERNEL_TYPE_SWIGLU,
+    GGML_METAL_KERNEL_TYPE_GEGLU_ERF,
+    GGML_METAL_KERNEL_TYPE_GEGLU_QUICK,
     GGML_METAL_KERNEL_TYPE_SUM_ROWS,
+    GGML_METAL_KERNEL_TYPE_MEAN,
     GGML_METAL_KERNEL_TYPE_POOL_2D_AVG_F32,
     GGML_METAL_KERNEL_TYPE_POOL_2D_MAX_F32,
     GGML_METAL_KERNEL_TYPE_ARGMAX,
@@ -976,7 +1012,7 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
     struct ggml_backend_metal_context * ctx = calloc(1, sizeof(struct ggml_backend_metal_context));
     struct ggml_backend_metal_device_context * ctx_dev = dev->context;
 
-    id<MTLDevice> device = ggml_backend_metal_device_acq(ctx_dev);
+    id<MTLDevice> device = ctx_dev->mtl_device;
 
     GGML_LOG_INFO("%s: picking default device: %s\n", __func__, [[device name] UTF8String]);
 
@@ -990,9 +1026,16 @@ 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);
 
     // load library
-    if (ctx_dev->mtl_library == nil) {
-        ctx_dev->mtl_library = ggml_metal_load_library(device, ctx_dev->use_bfloat);
+    {
+        [ctx_dev->mtl_lock lock];
+
+        if (ctx_dev->mtl_library == nil) {
+            ctx_dev->mtl_library = ggml_metal_load_library(device, ctx_dev->use_bfloat);
+        }
+
+        [ctx_dev->mtl_lock unlock];
     }
+
     id<MTLLibrary> metal_library = ctx_dev->mtl_library;
     if (metal_library == nil) {
         GGML_LOG_ERROR("%s: error: metal library is nil\n", __func__);
@@ -1141,20 +1184,33 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
         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_SET_ROWS_F32,                    set_rows_f32,                    true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SET_ROWS_F16,                    set_rows_f16,                    true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SET_ROWS_BF16,                   set_rows_bf16,                   use_bfloat);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SET_ROWS_Q8_0,                   set_rows_q8_0,                   true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SET_ROWS_Q4_0,                   set_rows_q4_0,                   true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SET_ROWS_Q4_1,                   set_rows_q4_1,                   true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SET_ROWS_Q5_0,                   set_rows_q5_0,                   true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SET_ROWS_Q5_1,                   set_rows_q5_1,                   true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SET_ROWS_IQ4_NL,                 set_rows_iq4_nl,                 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_SSM_SCAN_F32_GROUP,              ssm_scan_f32_group,              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_F32_F32_C4,               mul_mv_f32_f32_c4,               true);
         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_C4,              mul_mv_bf16_f32_c4,              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_C4,               mul_mv_f16_f32_c4,               true);
         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);
@@ -1453,7 +1509,13 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
         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_NEG,                             neg,                             true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_REGLU,                           reglu,                           true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GEGLU,                           geglu,                           true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SWIGLU,                          swiglu,                          true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GEGLU_ERF,                       geglu_erf,                       true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GEGLU_QUICK,                     geglu_quick,                     true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SUM_ROWS,                        sum_rows,                        true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MEAN,                            mean,                            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);
@@ -1603,6 +1665,10 @@ static bool ggml_metal_supports_op(const struct ggml_backend_metal_device_contex
     const bool use_bfloat              = ctx_dev->use_bfloat;
 
     if (!use_bfloat) {
+        if (op->type == GGML_TYPE_BF16) {
+            return false;
+        }
+
         for (size_t i = 0, n = 3; i < n; ++i) {
             if (op->src[i] != NULL && op->src[i]->type == GGML_TYPE_BF16) {
                 return false;
@@ -1626,6 +1692,17 @@ static bool ggml_metal_supports_op(const struct ggml_backend_metal_device_contex
                 default:
                     return false;
             }
+        case GGML_OP_GLU:
+            switch (ggml_get_glu_op(op)) {
+                case GGML_GLU_OP_REGLU:
+                case GGML_GLU_OP_GEGLU:
+                case GGML_GLU_OP_SWIGLU:
+                case GGML_GLU_OP_GEGLU_ERF:
+                case GGML_GLU_OP_GEGLU_QUICK:
+                    return ggml_is_contiguous_1(op->src[0]) && op->src[0]->type == GGML_TYPE_F32;
+               default:
+                    return false;
+            }
         case GGML_OP_NONE:
         case GGML_OP_RESHAPE:
         case GGML_OP_VIEW:
@@ -1653,9 +1730,10 @@ static bool ggml_metal_supports_op(const struct ggml_backend_metal_device_contex
         case GGML_OP_LOG:
             return false; // TODO: implement
         case GGML_OP_SUM_ROWS:
+        case GGML_OP_MEAN:
         case GGML_OP_SOFT_MAX:
         case GGML_OP_GROUP_NORM:
-            return has_simdgroup_reduction && ggml_is_contiguous(op->src[0]);
+            return has_simdgroup_reduction && ggml_is_contiguous_rows(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]));
@@ -1771,6 +1849,27 @@ static bool ggml_metal_supports_op(const struct ggml_backend_metal_device_contex
             {
                 return op->ne[3] == 1;
             }
+        case GGML_OP_SET_ROWS:
+            {
+                if (op->src[0]->type != GGML_TYPE_F32) {
+                    return false;
+                }
+
+                switch (op->type) {
+                    case GGML_TYPE_F32:
+                    case GGML_TYPE_F16:
+                    case GGML_TYPE_BF16:
+                    case GGML_TYPE_Q8_0:
+                    case GGML_TYPE_Q4_0:
+                    case GGML_TYPE_Q4_1:
+                    case GGML_TYPE_Q5_0:
+                    case GGML_TYPE_Q5_1:
+                    case GGML_TYPE_IQ4_NL:
+                        return true;
+                    default:
+                        return false;
+                };
+            }
         default:
             return false;
     }
@@ -2343,6 +2442,68 @@ static bool ggml_metal_encode_node(
                     GGML_ABORT("fatal error");
                 }
             } break;
+        case GGML_OP_GLU:
+            {
+                GGML_ASSERT(ggml_is_contiguous_1(src0));
+
+                if (src1) {
+                    GGML_ASSERT(ggml_are_same_shape(src0, src1));
+                }
+
+                id<MTLComputePipelineState> pipeline = nil;
+
+                switch (ggml_get_glu_op(node)) {
+                    case GGML_GLU_OP_REGLU:
+                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_REGLU].pipeline;
+                        break;
+                    case GGML_GLU_OP_GEGLU:
+                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GEGLU].pipeline;
+                        break;
+                    case GGML_GLU_OP_SWIGLU:
+                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SWIGLU].pipeline;
+                        break;
+                    case GGML_GLU_OP_GEGLU_ERF:
+                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GEGLU_ERF].pipeline;
+                        break;
+                    case GGML_GLU_OP_GEGLU_QUICK:
+                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GEGLU_QUICK].pipeline;
+                        break;
+                    default:
+                        GGML_ABORT("fatal error");
+                }
+
+                const int32_t swp = ((const int32_t *) dst->op_params)[1];
+
+                const int32_t i00 = swp ? ne0 : 0;
+                const int32_t i10 = swp ? 0 : ne0;
+
+                ggml_metal_kargs_glu args = {
+                    /*.ne00 =*/ ne00,
+                    /*.nb01 =*/ nb01,
+                    /*.ne10 =*/ src1 ? ne10 : ne00,
+                    /*.nb11 =*/ src1 ? nb11 : nb01,
+                    /*.ne0  =*/ ne0,
+                    /*.nb1  =*/ nb1,
+                    /*.i00  =*/ src1 ? 0 : i00,
+                    /*.i10  =*/ src1 ? 0 : i10,
+                };
+
+                [encoder setComputePipelineState:pipeline];
+                [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                if (src1) {
+                    [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
+                } else {
+                    [encoder setBuffer:id_src0 offset:offs_src0 atIndex:1];
+                }
+                [encoder setBuffer:id_dst  offset:offs_dst  atIndex:2];
+                [encoder setBytes:&args length:sizeof(args) atIndex:3];
+
+                const int64_t nrows = ggml_nrows(src0);
+
+                const int32_t nth = MIN((int) pipeline.maxTotalThreadsPerThreadgroup, ne00/2);
+
+                [encoder dispatchThreadgroups:MTLSizeMake(nrows, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
+            } break;
         case GGML_OP_SQR:
             {
                 GGML_ASSERT(ggml_is_contiguous(src0));
@@ -2400,11 +2561,31 @@ static bool ggml_metal_encode_node(
                 [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
             } break;
         case GGML_OP_SUM_ROWS:
+        case GGML_OP_MEAN:
             {
                 GGML_ASSERT(src0->nb[0] == ggml_type_size(src0->type));
 
-                id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SUM_ROWS].pipeline;
+                id<MTLComputePipelineState> pipeline = nil;
 
+                switch (dst->op) {
+                    case GGML_OP_SUM_ROWS:
+                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SUM_ROWS].pipeline;
+                        break;
+                    case GGML_OP_MEAN:
+                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MEAN].pipeline;
+                        break;
+                    default:
+                        GGML_ABORT("fatal error");
+                }
+
+                int nth = 32; // SIMD width
+
+                while (nth < ne00 && nth < (int) pipeline.maxTotalThreadsPerThreadgroup) {
+                    nth *= 2;
+                }
+
+                nth = MIN(nth, (int) pipeline.maxTotalThreadsPerThreadgroup);
+                nth = MIN(nth, ne00);
 
                 ggml_metal_kargs_sum_rows args = {
                    /*.ne00 =*/ ne00,
@@ -2434,11 +2615,12 @@ static bool ggml_metal_encode_node(
                 };
 
                 [encoder setComputePipelineState:pipeline];
-                [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 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];
 
-                [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
             } break;
         case GGML_OP_SOFT_MAX:
             {
@@ -2476,10 +2658,7 @@ static bool ggml_metal_encode_node(
                 memcpy(&scale,    ((const int32_t *) dst->op_params) + 0, sizeof(scale));
                 memcpy(&max_bias, ((const int32_t *) dst->op_params) + 1, sizeof(max_bias));
 
-                const int64_t nrows_x = ggml_nrows(src0);
-                const int64_t nrows_y = src0->ne[1];
-
-                const uint32_t n_head      = nrows_x/nrows_y;
+                const uint32_t n_head      = src0->ne[2];
                 const uint32_t n_head_log2 = 1u << (uint32_t) floorf(log2f((float) n_head));
 
                 const float m0 = powf(2.0f, -(max_bias       ) / n_head_log2);
@@ -2539,6 +2718,18 @@ static bool ggml_metal_encode_node(
                     /*.ne00        =*/ ne00,
                     /*.ne01        =*/ ne01,
                     /*.ne02        =*/ ne02,
+                    /*.nb01        =*/ nb01,
+                    /*.nb02        =*/ nb02,
+                    /*.nb03        =*/ nb03,
+                    /*.ne11        =*/ ne11,
+                    /*.ne12        =*/ ne12,
+                    /*.ne13        =*/ ne13,
+                    /*.nb11        =*/ nb11,
+                    /*.nb12        =*/ nb12,
+                    /*.nb13        =*/ nb13,
+                    /*.nb1         =*/ nb1,
+                    /*.nb2         =*/ nb2,
+                    /*.nb3         =*/ nb3,
                     /*.scale       =*/ scale,
                     /*.max_bias    =*/ max_bias,
                     /*.m0          =*/ m0,
@@ -2558,7 +2749,7 @@ static bool ggml_metal_encode_node(
 
                 [encoder setThreadgroupMemoryLength:32*sizeof(float) atIndex:0];
 
-                [encoder dispatchThreadgroups:MTLSizeMake(ne01*ne02*ne03, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
+                [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
             } break;
         case GGML_OP_DIAG_MASK_INF:
             {
@@ -2632,71 +2823,91 @@ static bool ggml_metal_encode_node(
                 struct ggml_tensor * src3 = node->src[3];
                 struct ggml_tensor * src4 = node->src[4];
                 struct ggml_tensor * src5 = node->src[5];
+                struct ggml_tensor * src6 = node->src[6];
 
                 GGML_ASSERT(src3);
                 GGML_ASSERT(src4);
                 GGML_ASSERT(src5);
+                GGML_ASSERT(src6);
 
                 size_t offs_src3 = 0;
                 size_t offs_src4 = 0;
                 size_t offs_src5 = 0;
+                size_t offs_src6 = 0;
 
                 id<MTLBuffer> id_src3 = src3 ? ggml_metal_get_buffer(src3, &offs_src3) : nil;
                 id<MTLBuffer> id_src4 = src4 ? ggml_metal_get_buffer(src4, &offs_src4) : nil;
                 id<MTLBuffer> id_src5 = src5 ? ggml_metal_get_buffer(src5, &offs_src5) : nil;
+                id<MTLBuffer> id_src6 = src6 ? ggml_metal_get_buffer(src6, &offs_src6) : nil;
 
-                const int64_t  ne30 = src3->ne[0]; GGML_UNUSED(ne30);
+                const int64_t  ne30 = src3->ne[0];
                 const int64_t  ne31 = src3->ne[1]; GGML_UNUSED(ne31);
 
-                const uint64_t nb30 = src3->nb[0];
+                const uint64_t nb30 = src3->nb[0]; GGML_UNUSED(nb30);
                 const uint64_t nb31 = src3->nb[1];
 
                 const int64_t  ne40 = src4->ne[0]; GGML_UNUSED(ne40);
-                const int64_t  ne41 = src4->ne[1]; GGML_UNUSED(ne41);
+                const int64_t  ne41 = src4->ne[1];
                 const int64_t  ne42 = src4->ne[2]; GGML_UNUSED(ne42);
+                const int64_t  ne43 = src4->ne[3]; GGML_UNUSED(ne43);
 
-                const uint64_t nb40 = src4->nb[0];
+                const uint64_t nb40 = src4->nb[0]; GGML_UNUSED(nb40);
                 const uint64_t nb41 = src4->nb[1];
                 const uint64_t nb42 = src4->nb[2];
+                const uint64_t nb43 = src4->nb[3];
 
                 const int64_t  ne50 = src5->ne[0]; GGML_UNUSED(ne50);
                 const int64_t  ne51 = src5->ne[1]; GGML_UNUSED(ne51);
                 const int64_t  ne52 = src5->ne[2]; GGML_UNUSED(ne52);
+                const int64_t  ne53 = src5->ne[3]; GGML_UNUSED(ne53);
 
-                const uint64_t nb50 = src5->nb[0];
+                const uint64_t nb50 = src5->nb[0]; GGML_UNUSED(nb50);
                 const uint64_t nb51 = src5->nb[1];
                 const uint64_t nb52 = src5->nb[2];
+                const uint64_t nb53 = src5->nb[3];
+
+                const int64_t  ne60 = src6->ne[0]; GGML_UNUSED(ne60);
+
+                const uint64_t nb60 = src6->nb[0]; GGML_UNUSED(nb60);
 
                 const int64_t d_state      = ne00;
                 const int64_t d_inner      = ne01;
-                const int64_t n_seq_tokens = ne11;
-                const int64_t n_seqs       = ne02;
+                const int64_t n_head       = ne02;
+                const int64_t n_group      = ne41;
+                const int64_t n_seq_tokens = ne12;
+                const int64_t n_seqs       = ne13;
 
-                id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SSM_SCAN_F32].pipeline;
+                id<MTLComputePipelineState> pipeline = nil;
+
+                if (ne30 == 1) {
+                    // Mamba-2
+                    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SSM_SCAN_F32_GROUP].pipeline;
+                } else {
+                    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SSM_SCAN_F32].pipeline;
+                }
 
                 ggml_metal_kargs_ssm_scan args = {
-                    /*.d_state =*/ d_state,
-                    /*.d_inner =*/ d_inner,
+                    /*.d_state      =*/ d_state,
+                    /*.d_inner      =*/ d_inner,
+                    /*.n_head       =*/ n_head,
+                    /*.n_group      =*/ n_group,
                     /*.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,
+                    /*.n_seqs       =*/ n_seqs,
+                    /*.nb01         =*/ nb01,
+                    /*.nb02         =*/ nb02,
+                    /*.nb03         =*/ nb03,
+                    /*.nb11         =*/ nb11,
+                    /*.nb12         =*/ nb12,
+                    /*.nb13         =*/ nb13,
+                    /*.nb21         =*/ nb21,
+                    /*.nb22         =*/ nb22,
+                    /*.nb31         =*/ nb31,
+                    /*.nb41         =*/ nb41,
+                    /*.nb42         =*/ nb42,
+                    /*.nb43         =*/ nb43,
+                    /*.nb51         =*/ nb51,
+                    /*.nb52         =*/ nb52,
+                    /*.nb53         =*/ nb53,
                 };
 
                 [encoder setComputePipelineState:pipeline];
@@ -2706,10 +2917,17 @@ static bool ggml_metal_encode_node(
                 [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 setBuffer:id_src6 offset:offs_src6 atIndex:6];
+                [encoder setBuffer:id_dst  offset:offs_dst  atIndex:7];
+                [encoder setBytes:&args    length:sizeof(args) atIndex:8];
 
-                [encoder dispatchThreadgroups:MTLSizeMake(d_inner, n_seqs, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                if (ne30 == 1) {
+                    // Mamba-2
+                    [encoder dispatchThreadgroups:MTLSizeMake(d_inner, n_head, n_seqs) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                } else {
+                    GGML_ASSERT(d_inner == 1);
+                    [encoder dispatchThreadgroups:MTLSizeMake(n_head, n_seqs, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                }
             } break;
         case GGML_OP_RWKV_WKV6:
             {
@@ -3063,14 +3281,23 @@ static bool ggml_metal_encode_node(
                                 nsg = 1;
                                 nr0 = 1;
                                 nr1 = 4;
-                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_F32_F32].pipeline;
+                                if (ne00 == 4) {
+                                    nr0 = 32;
+                                    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_F32_F32_C4].pipeline;
+                                } else {
+                                    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_F32_F32].pipeline;
+                                }
                             } break;
                         case GGML_TYPE_F16:
                             {
                                 nsg = 1;
                                 nr0 = 1;
                                 if (src1t == GGML_TYPE_F32) {
-                                    if (ne11 * ne12 < 4) {
+                                    if (ne00 == 4) {
+                                        nr0 = 32;
+                                        nr1 = 4;
+                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_C4].pipeline;
+                                    } else 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;
@@ -3089,7 +3316,11 @@ static bool ggml_metal_encode_node(
                                 nsg = 1;
                                 nr0 = 1;
                                 if (src1t == GGML_TYPE_F32) {
-                                    if (ne11 * ne12 < 4) {
+                                    if (ne00 == 4) {
+                                        nr0 = 32;
+                                        nr1 = 4;
+                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_F32_C4].pipeline;
+                                    } else 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;
@@ -3710,13 +3941,74 @@ static bool ggml_metal_encode_node(
                 };
 
                 [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:&args length:sizeof(args) atIndex:3];
+                [encoder setBytes:&args    length:sizeof(args) atIndex:0];
+                [encoder setBuffer:id_src0 offset:offs_src0    atIndex:1];
+                [encoder setBuffer:id_src1 offset:offs_src1    atIndex:2];
+                [encoder setBuffer:id_dst  offset:offs_dst     atIndex:3];
 
                 [encoder dispatchThreadgroups:MTLSizeMake(ne10, ne11, 1) threadsPerThreadgroup:MTLSizeMake(32, 1, 1)];
             } break;
+        case GGML_OP_SET_ROWS:
+            {
+                id<MTLComputePipelineState> pipeline = nil;
+
+                switch (dst->type) {
+                    case GGML_TYPE_F32:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SET_ROWS_F32   ].pipeline; break;
+                    case GGML_TYPE_F16:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SET_ROWS_F16   ].pipeline; break;
+                    case GGML_TYPE_BF16:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SET_ROWS_BF16  ].pipeline; break;
+                    case GGML_TYPE_Q8_0:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SET_ROWS_Q8_0  ].pipeline; break;
+                    case GGML_TYPE_Q4_0:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SET_ROWS_Q4_0  ].pipeline; break;
+                    case GGML_TYPE_Q4_1:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SET_ROWS_Q4_1  ].pipeline; break;
+                    case GGML_TYPE_Q5_0:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SET_ROWS_Q5_0  ].pipeline; break;
+                    case GGML_TYPE_Q5_1:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SET_ROWS_Q5_1  ].pipeline; break;
+                    case GGML_TYPE_IQ4_NL: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SET_ROWS_IQ4_NL].pipeline; break;
+                    default: GGML_ABORT("not implemented");
+                }
+
+                const int32_t nk0 = ne0/ggml_blck_size(dst->type);
+
+                int nth = 32; // SIMD width
+
+                while (nth < nk0 && nth < (int) pipeline.maxTotalThreadsPerThreadgroup) {
+                    nth *= 2;
+                }
+
+                int nrptg = 1;
+                if (nth > nk0) {
+                    nrptg = (nth + nk0 - 1)/nk0;
+                    nth   = nk0;
+
+                    if (nrptg*nth > (int) pipeline.maxTotalThreadsPerThreadgroup) {
+                        nrptg--;
+                    }
+                }
+
+                nth = MIN(nth, nk0);
+
+                ggml_metal_kargs_set_rows args = {
+                    /*.nk0  =*/ nk0,
+                    /*.ne01 =*/ ne01,
+                    /*.nb01 =*/ nb01,
+                    /*.nb02 =*/ nb02,
+                    /*.nb03 =*/ nb03,
+                    /*.ne11 =*/ ne11,
+                    /*.ne12 =*/ ne12,
+                    /*.nb10 =*/ nb10,
+                    /*.nb11 =*/ nb11,
+                    /*.nb12 =*/ nb12,
+                    /*.nb1  =*/ nb1,
+                    /*.nb2  =*/ nb2,
+                    /*.nb3  =*/ nb3,
+                };
+
+                [encoder setComputePipelineState:pipeline];
+                [encoder setBytes:&args    length:sizeof(args) atIndex:0];
+                [encoder setBuffer:id_src0 offset:offs_src0    atIndex:1];
+                [encoder setBuffer:id_src1 offset:offs_src1    atIndex:2];
+                [encoder setBuffer:id_dst  offset:offs_dst     atIndex:3];
+
+                [encoder dispatchThreadgroups:MTLSizeMake((ne01 + nrptg - 1)/nrptg, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, nrptg, 1)];
+            } break;
         case GGML_OP_RMS_NORM:
             {
                 GGML_ASSERT(ne00 % 4 == 0);
@@ -3733,6 +4025,7 @@ static bool ggml_metal_encode_node(
                     nth *= 2;
                 }
 
+                nth = MIN(nth, (int) pipeline.maxTotalThreadsPerThreadgroup);
                 nth = MIN(nth, ne00/4);
 
                 ggml_metal_kargs_rms_norm args = {
@@ -3769,6 +4062,7 @@ static bool ggml_metal_encode_node(
                     nth *= 2;
                 }
 
+                nth = MIN(nth, (int) pipeline.maxTotalThreadsPerThreadgroup);
                 nth = MIN(nth, ne00/4);
 
                 ggml_metal_kargs_l2_norm args = {
@@ -3841,6 +4135,7 @@ static bool ggml_metal_encode_node(
                     nth *= 2;
                 }
 
+                nth = MIN(nth, (int) pipeline.maxTotalThreadsPerThreadgroup);
                 nth = MIN(nth, ne00/4);
 
                 ggml_metal_kargs_norm args = {
@@ -4734,7 +5029,11 @@ static bool ggml_metal_encode_node(
                     /*.nb21          =*/ nb21,
                     /*.nb22          =*/ nb22,
                     /*.nb23          =*/ nb23,
+                    /*.ne32          =*/ ne32,
+                    /*.ne33          =*/ ne33,
                     /*.nb31          =*/ nb31,
+                    /*.nb32          =*/ nb32,
+                    /*.nb33          =*/ nb33,
                     /*.ne1           =*/ ne1,
                     /*.ne2           =*/ ne2,
                     /*.scale         =*/ scale,
@@ -4766,6 +5065,8 @@ static bool ggml_metal_encode_node(
                     GGML_ASSERT(nqptg  % 8  == 0);
                     GGML_ASSERT(ncpsg  % 32 == 0);
 
+                    const int is_q = ggml_is_quantized(src1->type) ? 1 : 0;
+
                     // 2*(2*ncpsg + nqptg)*(nsg)
                     // ncpsg soft_max values + ncpsg mask values + a diagonal scaling matrix (in float)
                     //
@@ -4773,7 +5074,7 @@ static bool ggml_metal_encode_node(
                     // the shared memory needed for the simdgroups to load the KV cache
                     // each thread loads (dequantizes) 16 head elements, there are 32 threads in th SG
                     //
-#define FATTN_SMEM(nsg) (GGML_PAD((nqptg*(ne00 + 2*(2*ncpsg + nqptg)*(nsg)) + 16*32*(nsg))*(sizeof(float)/2), 16))
+#define FATTN_SMEM(nsg) (GGML_PAD((nqptg*(2*ne00 + 2*(2*ncpsg + nqptg)*(nsg)) + is_q*(16*32*(nsg)))*(sizeof(float)/2), 16))
 
                     int64_t nsgmax = 2;
 
@@ -4810,9 +5111,9 @@ static bool ggml_metal_encode_node(
                     // and store the soft_max values and the mask
                     //
                     // ne00*(nsg)
-                    // each simdgroup has a full f16 head vector in shared mem to accumulate results
+                    // each simdgroup has a full f32 head vector in shared mem to accumulate results
                     //
-#define FATTN_SMEM(nsg) (GGML_PAD((nqptg*(GGML_PAD(ne00, 128) + 4*ncpsg*(nsg)) + ne20*(nsg))*(sizeof(float)/2), 16))
+#define FATTN_SMEM(nsg) (GGML_PAD((nqptg*(GGML_PAD(ne00, 128) + 4*ncpsg*(nsg)) + 2*ne20*(nsg))*(sizeof(float)/2), 16))
 
                     int64_t nsgmax = 2;
                     while (true) {
@@ -4925,8 +5226,39 @@ static bool ggml_metal_encode_node(
                     default: GGML_ABORT("not implemented");
                 }
 
+                GGML_ASSERT(ne00 % ggml_blck_size(src0->type) == 0);
+
+                // TODO: support
+                //const int32_t nk00 = ne00/ggml_blck_size(dst->type);
+                const int32_t nk00 = ne00;
+
+                int nth = 32; // SIMD width
+
+                while (nth < nk00 && nth < (int) pipeline.maxTotalThreadsPerThreadgroup) {
+                    nth *= 2;
+                }
+
+                nth = MIN(nth, (int) pipeline.maxTotalThreadsPerThreadgroup);
+
+                // when rows are small, we can batch them together in a single threadgroup
+                int nrptg = 1;
+
+                // TODO: relax this constraint in the future
+                if (ggml_blck_size(src0->type) == 1 && ggml_blck_size(dst->type) == 1) {
+                    if (nth > nk00) {
+                        nrptg = (nth + nk00 - 1)/nk00;
+                        nth   = nk00;
+
+                        if (nrptg*nth > (int) pipeline.maxTotalThreadsPerThreadgroup) {
+                            nrptg--;
+                        }
+                    }
+                }
+
+                nth = MIN(nth, nk00);
+
                 ggml_metal_kargs_cpy args = {
-                    /*.ne00 =*/ ne00,
+                    /*.ne00 =*/ nk00,
                     /*.ne01 =*/ ne01,
                     /*.ne02 =*/ ne02,
                     /*.ne03 =*/ ne03,
@@ -4949,11 +5281,7 @@ static bool 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)];
-
+                [encoder dispatchThreadgroups:MTLSizeMake((ne01 + nrptg - 1)/nrptg, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, nrptg, 1)];
             } break;
         case GGML_OP_SET:
             {
@@ -5259,7 +5587,6 @@ static void ggml_backend_metal_buffer_free_buffer(ggml_backend_buffer_t buffer)
     }
 
     ggml_backend_metal_buffer_rset_free(ctx);
-    ggml_backend_metal_device_rel(buffer->buft->device->context);
 
     if (ctx->owned) {
 #if TARGET_OS_OSX
@@ -5368,7 +5695,10 @@ static ggml_backend_buffer_t ggml_backend_metal_buffer_type_alloc_buffer(ggml_ba
     }
 
     struct ggml_backend_metal_device_context * ctx_dev = (struct ggml_backend_metal_device_context *)buft->device->context;
-    id<MTLDevice> device = ggml_backend_metal_device_acq(ctx_dev);
+
+    GGML_ASSERT(ctx_dev->mtl_device != nil);
+
+    id<MTLDevice> device = ctx_dev->mtl_device;
 
     ctx->all_data = ggml_metal_host_malloc(size_aligned);
     ctx->all_size = size_aligned;
@@ -5391,14 +5721,12 @@ static ggml_backend_buffer_t ggml_backend_metal_buffer_type_alloc_buffer(ggml_ba
     if (size_aligned > 0 && (ctx->all_data == NULL || ctx->buffers[0].metal == nil)) {
         GGML_LOG_ERROR("%s: error: failed to allocate buffer, size = %8.2f MiB\n", __func__, size_aligned / 1024.0 / 1024.0);
         free(ctx);
-        ggml_backend_metal_device_rel(ctx_dev);
         return NULL;
     }
 
     if (!ggml_backend_metal_buffer_rset_init(ctx, ctx_dev, device)) {
         GGML_LOG_ERROR("%s: error: failed to initialize residency set\n", __func__);
         free(ctx);
-        ggml_backend_metal_device_rel(ctx_dev);
         return NULL;
     }
 
@@ -5409,17 +5737,14 @@ static ggml_backend_buffer_t ggml_backend_metal_buffer_type_alloc_buffer(ggml_ba
 
 static size_t ggml_backend_metal_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
     return 32;
+
     GGML_UNUSED(buft);
 }
 
 static size_t ggml_backend_metal_buffer_type_get_max_size(ggml_backend_buffer_type_t buft) {
-    id<MTLDevice> device = ggml_backend_metal_device_acq(buft->device->context);
-    const size_t max_size = device.maxBufferLength;
-    ggml_backend_metal_device_rel(buft->device->context);
+    const size_t max_size = ((struct ggml_backend_metal_device_context *)buft->device->context)->max_size;
 
     return max_size;
-
-    GGML_UNUSED(buft);
 }
 
 static bool ggml_backend_metal_buffer_type_is_host(ggml_backend_buffer_type_t buft) {
@@ -5492,7 +5817,10 @@ ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data, size_t siz
     }
 
     struct ggml_backend_metal_device_context * ctx_dev = &g_ggml_ctx_dev_main;
-    id<MTLDevice> device = ggml_backend_metal_device_acq(ctx_dev);
+
+    GGML_ASSERT(ctx_dev->mtl_device != nil);
+
+    id<MTLDevice> device = ctx_dev->mtl_device;
 
     // the buffer fits into the max buffer size allowed by the device
     if (size_aligned <= device.maxBufferLength) {
@@ -5548,7 +5876,6 @@ ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data, size_t siz
     if (!ggml_backend_metal_buffer_rset_init(ctx, ctx_dev, device)) {
         GGML_LOG_ERROR("%s: error: failed to initialize residency set\n", __func__);
         free(ctx);
-        ggml_backend_metal_device_rel(ctx_dev);
         return NULL;
     }
 
@@ -5564,10 +5891,8 @@ static const char * ggml_backend_metal_name(ggml_backend_t backend) {
 }
 
 static void ggml_backend_metal_free(ggml_backend_t backend) {
-    struct ggml_backend_metal_context        * ctx     = backend->context;
-    struct ggml_backend_metal_device_context * ctx_dev = backend->device->context;
+    struct ggml_backend_metal_context * ctx = backend->context;
 
-    ggml_backend_metal_device_rel(ctx_dev);
     ggml_metal_free(ctx);
 
     free(backend);
@@ -5707,6 +6032,8 @@ bool ggml_backend_metal_supports_family(ggml_backend_t backend, int family) {
 
     struct ggml_backend_metal_device_context * ctx_dev = backend->device->context;
 
+    GGML_ASSERT(ctx_dev->mtl_device != nil);
+
     return [ctx_dev->mtl_device supportsFamily:(MTLGPUFamilyApple1 + family - 1)];
 }
 
@@ -5726,10 +6053,7 @@ static const char * ggml_backend_metal_device_get_name(ggml_backend_dev_t dev) {
 }
 
 static const char * ggml_backend_metal_device_get_description(ggml_backend_dev_t dev) {
-    // acq/rel just to populate ctx->name in case it hasn't been done yet
     struct ggml_backend_metal_device_context * ctx_dev = (struct ggml_backend_metal_device_context *)dev->context;
-    ggml_backend_metal_device_acq(ctx_dev);
-    ggml_backend_metal_device_rel(ctx_dev);
 
     return ctx_dev->name;
 }
@@ -5737,12 +6061,10 @@ static const char * ggml_backend_metal_device_get_description(ggml_backend_dev_t
 static void ggml_backend_metal_device_get_memory(ggml_backend_dev_t dev, size_t * free, size_t * total) {
     if (@available(macOS 10.12, iOS 16.0, *)) {
         struct ggml_backend_metal_device_context * ctx_dev = (struct ggml_backend_metal_device_context *)dev->context;
-        id<MTLDevice> device = ggml_backend_metal_device_acq(ctx_dev);
+        id<MTLDevice> device = ctx_dev->mtl_device;
 
         *total = device.recommendedMaxWorkingSetSize;
         *free  = *total - device.currentAllocatedSize;
-
-        ggml_backend_metal_device_rel(ctx_dev);
     } else {
         *free = 1;
         *total = 1;
@@ -5820,7 +6142,10 @@ static ggml_backend_buffer_t ggml_backend_metal_device_buffer_from_ptr(ggml_back
     }
 
     struct ggml_backend_metal_device_context * ctx_dev = (struct ggml_backend_metal_device_context *)dev->context;
-    id<MTLDevice> device = ggml_backend_metal_device_acq(ctx_dev);
+
+    GGML_ASSERT(ctx_dev->mtl_device != nil);
+
+    id<MTLDevice> device = ctx_dev->mtl_device;
 
     // the buffer fits into the max buffer size allowed by the device
     if (size_aligned <= device.maxBufferLength) {
@@ -5876,7 +6201,6 @@ static ggml_backend_buffer_t ggml_backend_metal_device_buffer_from_ptr(ggml_back
     if (!ggml_backend_metal_buffer_rset_init(ctx, ctx_dev, device)) {
         GGML_LOG_ERROR("%s: error: failed to initialize residency set\n", __func__);
         free(ctx);
-        ggml_backend_metal_device_rel(ctx_dev);
         return NULL;
     }
 
@@ -5890,8 +6214,9 @@ static bool ggml_backend_metal_device_supports_op(ggml_backend_dev_t dev, const
 }
 
 static bool ggml_backend_metal_device_supports_buft(ggml_backend_dev_t dev, ggml_backend_buffer_type_t buft) {
-    return buft->iface.get_name == ggml_backend_metal_buffer_type_get_name ||
-            buft->iface.get_name == ggml_backend_metal_buffer_from_ptr_type_get_name;
+    return
+        buft->iface.get_name == ggml_backend_metal_buffer_type_get_name ||
+        buft->iface.get_name == ggml_backend_metal_buffer_from_ptr_type_get_name;
 
     GGML_UNUSED(dev);
 }
@@ -5976,8 +6301,19 @@ static struct ggml_backend_reg_i ggml_backend_metal_reg_i = {
     /* .get_proc_address = */ ggml_backend_metal_get_proc_address,
 };
 
+// called upon program exit
+static void ggml_metal_cleanup(void) {
+    ggml_backend_metal_device_rel(&g_ggml_ctx_dev_main);
+}
+
+// TODO: make thread-safe
 ggml_backend_reg_t ggml_backend_metal_reg(void) {
-    // TODO: make this thread-safe somehow?
+    ggml_backend_metal_device_acq(&g_ggml_ctx_dev_main);
+
+    // register cleanup callback
+    // TODO: not ideal, but not sure if there is a better way to do this in Objective-C
+    atexit(ggml_metal_cleanup);
+
     {
         g_ggml_backend_metal_reg = (struct ggml_backend_reg) {
             /* .api_version = */ GGML_BACKEND_API_VERSION,
diff --git a/ggml/src/ggml-metal/ggml-metal.metal b/ggml/src/ggml-metal/ggml-metal.metal
index 59899550e..22240bab4 100644
--- a/ggml/src/ggml-metal/ggml-metal.metal
+++ b/ggml/src/ggml-metal/ggml-metal.metal
@@ -35,6 +35,17 @@ constexpr constant static float kvalues_iq4nl_f[16] = {
     -127.f, -104.f, -83.f, -65.f, -49.f, -35.f, -22.f, -10.f, 1.f, 13.f, 25.f, 38.f, 53.f, 69.f, 89.f, 113.f
 };
 
+static inline int best_index_int8(int n, constant float * 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;
+}
+
 // NOTE: this is not dequantizing - we are simply fitting the template
 template <typename type4x4>
 void dequantize_f32(device const float4x4 * src, short il, thread type4x4 & reg) {
@@ -97,6 +108,178 @@ void dequantize_q4_0_t4(device const block_q4_0 * xb, short il, thread type4 & r
     }
 }
 
+void quantize_q4_0(device const float * src, device block_q4_0 & dst) {
+#pragma METAL fp math_mode(safe)
+    float amax = 0.0f; // absolute max
+    float max  = 0.0f;
+
+    for (int j = 0; j < QK4_0; j++) {
+        const float v = src[j];
+        if (amax < fabs(v)) {
+            amax = fabs(v);
+            max  = v;
+        }
+    }
+
+    const float d = max / -8;
+    const float id = d ? 1.0f/d : 0.0f;
+
+    dst.d = d;
+
+    for (int j = 0; j < QK4_0/2; ++j) {
+        const float x0 = src[0       + j]*id;
+        const float x1 = src[QK4_0/2 + j]*id;
+
+        const uint8_t xi0 = MIN(15, (int8_t)(x0 + 8.5f));
+        const uint8_t xi1 = MIN(15, (int8_t)(x1 + 8.5f));
+
+        dst.qs[j]  = xi0;
+        dst.qs[j] |= xi1 << 4;
+    }
+}
+
+void quantize_q4_1(device const float * src, device block_q4_1 & dst) {
+#pragma METAL fp math_mode(safe)
+    float min = FLT_MAX;
+    float max = -FLT_MAX;
+
+    for (int j = 0; j < QK4_1; j++) {
+        const float v = src[j];
+        if (min > v) min = v;
+        if (max < v) max = v;
+    }
+
+    const float d = (max - min) / ((1 << 4) - 1);
+    const float id = d ? 1.0f/d : 0.0f;
+
+    dst.d = d;
+    dst.m = min;
+
+    for (int j = 0; j < QK4_1/2; ++j) {
+        const float x0 = (src[0       + j] - min)*id;
+        const float x1 = (src[QK4_1/2 + j] - min)*id;
+
+        const uint8_t xi0 = MIN(15, (int8_t)(x0 + 0.5f));
+        const uint8_t xi1 = MIN(15, (int8_t)(x1 + 0.5f));
+
+        dst.qs[j]  = xi0;
+        dst.qs[j] |= xi1 << 4;
+    }
+}
+
+void quantize_q5_0(device const float * src, device block_q5_0 & dst) {
+#pragma METAL fp math_mode(safe)
+    float amax = 0.0f; // absolute max
+    float max  = 0.0f;
+
+    for (int j = 0; j < QK5_0; j++) {
+        const float v = src[j];
+        if (amax < fabs(v)) {
+            amax = fabs(v);
+            max  = v;
+        }
+    }
+
+    const float d = max / -16;
+    const float id = d ? 1.0f/d : 0.0f;
+
+    dst.d = d;
+
+    uint32_t qh = 0;
+    for (int j = 0; j < QK5_0/2; ++j) {
+        const float x0 = src[0       + j]*id;
+        const float x1 = src[QK5_0/2 + j]*id;
+
+        const uint8_t xi0 = MIN(31, (int8_t)(x0 + 16.5f));
+        const uint8_t xi1 = MIN(31, (int8_t)(x1 + 16.5f));
+
+        dst.qs[j] = (xi0 & 0xf) | ((xi1 & 0xf) << 4);
+        qh |= ((xi0 & 0x10u) >> 4) << (j + 0);
+        qh |= ((xi1 & 0x10u) >> 4) << (j + QK5_0/2);
+    }
+
+    thread const uint8_t * qh8 = (thread const uint8_t *)&qh;
+
+    for (int j = 0; j < 4; ++j) {
+        dst.qh[j] = qh8[j];
+    }
+}
+
+void quantize_q5_1(device const float * src, device block_q5_1 & dst) {
+#pragma METAL fp math_mode(safe)
+    float max = src[0];
+    float min = src[0];
+
+    for (int j = 1; j < QK5_1; j++) {
+        const float v = src[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;
+
+    dst.d = d;
+    dst.m = min;
+
+    uint32_t qh = 0;
+    for (int j = 0; j < QK5_1/2; ++j) {
+        const float x0 = (src[0       + j] - min)*id;
+        const float x1 = (src[QK5_1/2 + j] - min)*id;
+
+        const uint8_t xi0 = (uint8_t)(x0 + 0.5f);
+        const uint8_t xi1 = (uint8_t)(x1 + 0.5f);
+
+        dst.qs[j] = (xi0 & 0xf) | ((xi1 & 0xf) << 4);
+        qh |= ((xi0 & 0x10u) >> 4) << (j + 0);
+        qh |= ((xi1 & 0x10u) >> 4) << (j + QK5_1/2);
+    }
+
+    thread const uint8_t * qh8 = (thread const uint8_t *)&qh;
+
+    for (int j = 0; j < 4; ++j) {
+        dst.qh[j] = qh8[j];
+    }
+}
+
+void quantize_iq4_nl(device const float * src, device block_iq4_nl & dst) {
+#pragma METAL fp math_mode(safe)
+    float amax = 0.0f; // absolute max
+    float max  = 0.0f;
+
+    for (int j = 0; j < QK4_NL; j++) {
+        const float v = src[j];
+        if (amax < fabs(v)) {
+            amax = fabs(v);
+            max  = v;
+        }
+    }
+
+    const float d = max / kvalues_iq4nl_f[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 = src[0        + j]*id;
+        const float x1 = src[QK4_NL/2 + j]*id;
+
+        const uint8_t xi0 = best_index_int8(16, kvalues_iq4nl_f, x0);
+        const uint8_t xi1 = best_index_int8(16, kvalues_iq4nl_f, x1);
+
+        dst.qs[j] = xi0 | (xi1 << 4);
+
+        const float v0 = kvalues_iq4nl_f[xi0];
+        const float v1 = kvalues_iq4nl_f[xi1];
+        const float w0 = src[0        + j]*src[0        + j];
+        const float w1 = src[QK4_NL/2 + j]*src[QK4_NL/2 + j];
+        sumqx += w0*v0*src[j] + w1*v1*src[QK4_NL/2 + j];
+        sumq2 += w0*v0*v0 + w1*v1*v1;
+
+    }
+
+    dst.d = sumq2 > 0 ? sumqx/sumq2 : d;
+}
+
 template <typename type4x4>
 void dequantize_q4_1(device const block_q4_1 * xb, short il, thread type4x4 & reg) {
     device const uint16_t * qs = ((device const uint16_t *)xb + 2);
@@ -279,6 +462,27 @@ void dequantize_q8_0_t4(device const block_q8_0 *xb, short il, thread type4 & re
     }
 }
 
+void quantize_q8_0(device const float * src, device block_q8_0 & dst) {
+#pragma METAL fp math_mode(safe)
+    float amax = 0.0f; // absolute max
+
+    for (int j = 0; j < QK8_0; j++) {
+        const float v = src[j];
+        amax = MAX(amax, fabs(v));
+    }
+
+    const float d = amax / ((1 << 7) - 1);
+    const float id = d ? 1.0f/d : 0.0f;
+
+    dst.d = d;
+
+    for (int j = 0; j < QK8_0; ++j) {
+        const float x0 = src[j]*id;
+
+        dst.qs[j] = round(x0);
+    }
+}
+
 template <typename type4x4>
 void dequantize_q2_K(device const block_q2_K *xb, short il, thread type4x4 & reg) {
     const float d = xb->d;
@@ -993,31 +1197,169 @@ kernel void kernel_neg(
     dst[tpig] = -src0[tpig];
 }
 
+kernel void kernel_reglu(
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        constant ggml_metal_kargs_glu & args,
+        uint tgpig[[threadgroup_position_in_grid]],
+        uint tpitg[[thread_position_in_threadgroup]],
+        uint   ntg[[threads_per_threadgroup]]) {
+    device const float * src0_row = (device const float *) ((device const char *) src0 + tgpig*args.nb01) + args.i00;
+    device const float * src1_row = (device const float *) ((device const char *) src1 + tgpig*args.nb11) + args.i10;
+    device       float * dst_row  = (device       float *) ((device       char *) dst  + tgpig*args.nb1);
+
+    for (int i0 = tpitg; i0 < args.ne0; i0 += ntg) {
+        const float x0 = src0_row[i0];
+        const float x1 = src1_row[i0];
+
+        dst_row[i0] = x0*x1*(x0 > 0.0f);
+    }
+}
+
+kernel void kernel_geglu(
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        constant ggml_metal_kargs_glu & args,
+        uint tgpig[[threadgroup_position_in_grid]],
+        uint tpitg[[thread_position_in_threadgroup]],
+        uint   ntg[[threads_per_threadgroup]]) {
+    device const float * src0_row = (device const float *) ((device const char *) src0 + tgpig*args.nb01) + args.i00;
+    device const float * src1_row = (device const float *) ((device const char *) src1 + tgpig*args.nb11) + args.i10;
+    device       float * dst_row  = (device       float *) ((device       char *) dst  + tgpig*args.nb1);
+
+    for (int i0 = tpitg; i0 < args.ne0; i0 += ntg) {
+        const float x0 = src0_row[i0];
+        const float x1 = src1_row[i0];
+
+        const float gelu = 0.5f*x0*(1.0f + precise::tanh(SQRT_2_OVER_PI*x0*(1.0f + GELU_COEF_A*x0*x0)));
+
+        dst_row[i0] = gelu*x1;
+    }
+}
+
+kernel void kernel_swiglu(
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        constant ggml_metal_kargs_glu & args,
+        uint tgpig[[threadgroup_position_in_grid]],
+        uint tpitg[[thread_position_in_threadgroup]],
+        uint   ntg[[threads_per_threadgroup]]) {
+    device const float * src0_row = (device const float *) ((device const char *) src0 + tgpig*args.nb01) + args.i00;
+    device const float * src1_row = (device const float *) ((device const char *) src1 + tgpig*args.nb11) + args.i10;
+    device       float * dst_row  = (device       float *) ((device       char *) dst  + tgpig*args.nb1);
+
+    for (int i0 = tpitg; i0 < args.ne0; i0 += ntg) {
+        const float x0 = src0_row[i0];
+        const float x1 = src1_row[i0];
+
+        const float silu = x0 / (1.0f + exp(-x0));
+
+        dst_row[i0] = silu*x1;
+    }
+}
+
+kernel void kernel_geglu_erf(
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        constant ggml_metal_kargs_glu & args,
+        uint tgpig[[threadgroup_position_in_grid]],
+        uint tpitg[[thread_position_in_threadgroup]],
+        uint   ntg[[threads_per_threadgroup]]) {
+    device const float * src0_row = (device const float *) ((device const char *) src0 + tgpig*args.nb01) + args.i00;
+    device const float * src1_row = (device const float *) ((device const char *) src1 + tgpig*args.nb11) + args.i10;
+    device       float * dst_row  = (device       float *) ((device       char *) dst  + tgpig*args.nb1);
+
+    for (int i0 = tpitg; i0 < args.ne0; i0 += ntg) {
+        const float x0 = src0_row[i0];
+        const float x1 = src1_row[i0];
+
+        const float gelu_erf = 0.5f*x0*(1.0f+erf_approx<float>(x0*SQRT_2_INV));
+
+        dst_row[i0] = gelu_erf*x1;
+    }
+}
+
+kernel void kernel_geglu_quick(
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        constant ggml_metal_kargs_glu & args,
+        uint tgpig[[threadgroup_position_in_grid]],
+        uint tpitg[[thread_position_in_threadgroup]],
+        uint   ntg[[threads_per_threadgroup]]) {
+    device const float * src0_row = (device const float *) ((device const char *) src0 + tgpig*args.nb01) + args.i00;
+    device const float * src1_row = (device const float *) ((device const char *) src1 + tgpig*args.nb11) + args.i10;
+    device       float * dst_row  = (device       float *) ((device       char *) dst  + tgpig*args.nb1);
+
+    for (int i0 = tpitg; i0 < args.ne0; i0 += ntg) {
+        const float x0 = src0_row[i0];
+        const float x1 = src1_row[i0];
+
+        const float gelu_quick = x0*(1.0f/(1.0f+exp(GELU_QUICK_COEF*x0)));
+
+        dst_row[i0] = gelu_quick*x1;
+    }
+}
+
+template <bool norm>
 kernel void kernel_sum_rows(
+        constant ggml_metal_kargs_sum_rows & args,
         device const float * src0,
         device       float * dst,
-        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;
+        threadgroup  float * shmem_f32 [[threadgroup(0)]],
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort  sgitg[[simdgroup_index_in_threadgroup]],
+        ushort  tiisg[[thread_index_in_simdgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    int64_t i3 = tgpig.z;
+    int64_t i2 = tgpig.y;
+    int64_t i1 = tgpig.x;
 
     if (i3 >= args.ne03 || i2 >= args.ne02 || i1 >= args.ne01) {
         return;
     }
 
+    if (sgitg == 0) {
+        shmem_f32[tiisg] = 0.0f;
+    }
+
     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;
+    float sumf = 0;
 
-    for (int64_t i0 = 0; i0 < args.ne00; i0++) {
-        row_sum += src_row[i0];
+    for (int64_t i0 = tpitg.x; i0 < args.ne00; i0 += ntg.x) {
+        sumf += src_row[i0];
     }
 
-    dst_row[0] = row_sum;
+    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);
+
+    if (tpitg.x == 0) {
+        dst_row[0] = norm ? sumf / args.ne00 : sumf;
+    }
 }
 
+typedef decltype(kernel_sum_rows<false>) kernel_sum_rows_t;
+
+template [[host_name("kernel_sum_rows")]] kernel kernel_sum_rows_t kernel_sum_rows<false>;
+template [[host_name("kernel_mean")]]     kernel kernel_sum_rows_t kernel_sum_rows<true>;
+
 template<typename T>
 kernel void kernel_soft_max(
         device const  char * src0,
@@ -1025,24 +1367,28 @@ kernel void kernel_soft_max(
         device        char * dst,
         constant ggml_metal_kargs_soft_max & args,
         threadgroup  float * buf [[threadgroup(0)]],
-        uint  tgpig[[threadgroup_position_in_grid]],
-        uint  tpitg[[thread_position_in_threadgroup]],
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3 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) / (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);
+        uint3  tptg[[threads_per_threadgroup]]) {
+    const int32_t i03 = tgpig.z;
+    const int32_t i02 = tgpig.y;
+    const int32_t i01 = tgpig.x;
 
-    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);
+    const int32_t i13 = i03%args.ne13;
+    const int32_t i12 = i02%args.ne12;
+    const int32_t i11 = i01;
+
+    device const float * psrc0 =                (device const float *) (src0 + i01*args.nb01 + i02*args.nb02 + i03*args.nb03);
+    device const     T * pmask = src1 != src0 ? (device const T *    ) (src1 + i11*args.nb11 + i12*args.nb12 + i13*args.nb13) : nullptr;
+    device       float * pdst  =                (device       float *) (dst  + i01*args.nb1  + i02*args.nb2  + i03*args.nb3);
 
     float slope = 1.0f;
 
     // ALiBi
     if (args.max_bias > 0.0f) {
-        const int64_t h = i02;
+        const int32_t h = i02;
 
         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;
@@ -1053,13 +1399,13 @@ kernel void kernel_soft_max(
     // parallel max
     float lmax = -INFINITY;
 
-    for (int i00 = tpitg; i00 < args.ne00; i00 += ntg) {
+    for (int i00 = tpitg.x; i00 < args.ne00; i00 += tptg.x) {
         lmax = MAX(lmax, psrc0[i00]*args.scale + (pmask ? slope*pmask[i00] : 0.0f));
     }
 
     // find the max value in the block
     float max_val = simd_max(lmax);
-    if (ntg > N_SIMDWIDTH) {
+    if (tptg.x > N_SIMDWIDTH) {
         if (sgitg == 0) {
             buf[tiisg] = -INFINITY;
         }
@@ -1078,7 +1424,7 @@ kernel void kernel_soft_max(
 
     // parallel sum
     float lsum = 0.0f;
-    for (int i00 = tpitg; i00 < args.ne00; i00 += ntg) {
+    for (int i00 = tpitg.x; i00 < args.ne00; i00 += tptg.x) {
         const float exp_psrc0 = exp((psrc0[i00]*args.scale + (pmask ? slope*pmask[i00] : 0.0f)) - max_val);
         lsum += exp_psrc0;
         pdst[i00] = exp_psrc0;
@@ -1090,7 +1436,7 @@ kernel void kernel_soft_max(
 
     float sum = simd_sum(lsum);
 
-    if (ntg > N_SIMDWIDTH) {
+    if (tptg.x > N_SIMDWIDTH) {
         if (sgitg == 0) {
             buf[tiisg] = 0.0f;
         }
@@ -1109,7 +1455,7 @@ kernel void kernel_soft_max(
 
     const float inv_sum = 1.0f/sum;
 
-    for (int i00 = tpitg; i00 < args.ne00; i00 += ntg) {
+    for (int i00 = tpitg.x; i00 < args.ne00; i00 += tptg.x) {
         pdst[i00] *= inv_sum;
     }
 }
@@ -1121,23 +1467,27 @@ kernel void kernel_soft_max_4(
         device        char * dst,
         constant ggml_metal_kargs_soft_max & args,
         threadgroup  float * buf [[threadgroup(0)]],
-        uint  tgpig[[threadgroup_position_in_grid]],
-        uint  tpitg[[thread_position_in_threadgroup]],
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3 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) / (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);
+        uint3  tptg[[threads_per_threadgroup]]) {
+    const int32_t i03 = tgpig.z;
+    const int32_t i02 = tgpig.y;
+    const int32_t i01 = tgpig.x;
 
-    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;
+    const int32_t i13 = i03%args.ne13;
+    const int32_t i12 = i02%args.ne12;
+    const int32_t i11 = i01;
+
+    device const float4 * psrc4 =                (device const float4 *) (src0 + i01*args.nb01 + i02*args.nb02 + i03*args.nb03);
+    device const      T * pmask = src1 != src0 ? (device const T *     ) (src1 + i11*args.nb11 + i12*args.nb12 + i13*args.nb13) : nullptr;
+    device       float4 * pdst4 =                (device       float4 *) (dst  + i01*args.nb1  + i02*args.nb2  + i03*args.nb3);
 
     float slope = 1.0f;
 
     if (args.max_bias > 0.0f) {
-        const int64_t h = i02;
+        const int32_t h = i02;
 
         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;
@@ -1148,14 +1498,14 @@ kernel void kernel_soft_max_4(
     // parallel max
     float4 lmax4 = -INFINITY;
 
-    for (int i00 = tpitg; i00 < args.ne00/4; i00 += ntg) {
+    for (int i00 = tpitg.x; i00 < args.ne00/4; i00 += tptg.x) {
         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]));
 
     float max_val = simd_max(lmax);
-    if (ntg > N_SIMDWIDTH) {
+    if (tptg.x > N_SIMDWIDTH) {
         if (sgitg == 0) {
             buf[tiisg] = -INFINITY;
         }
@@ -1174,7 +1524,7 @@ kernel void kernel_soft_max_4(
 
     // parallel sum
     float4 lsum4 = 0.0f;
-    for (int i00 = tpitg; i00 < args.ne00/4; i00 += ntg) {
+    for (int i00 = tpitg.x; i00 < args.ne00/4; i00 += tptg.x) {
         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;
@@ -1188,7 +1538,7 @@ kernel void kernel_soft_max_4(
 
     float sum = simd_sum(lsum);
 
-    if (ntg > N_SIMDWIDTH) {
+    if (tptg.x > N_SIMDWIDTH) {
         if (sgitg == 0) {
             buf[tiisg] = 0.0f;
         }
@@ -1207,7 +1557,7 @@ kernel void kernel_soft_max_4(
 
     const float inv_sum = 1.0f/sum;
 
-    for (int i00 = tpitg; i00 < args.ne00/4; i00 += ntg) {
+    for (int i00 = tpitg.x; i00 < args.ne00/4; i00 += tptg.x) {
         pdst4[i00] *= inv_sum;
     }
 }
@@ -1293,7 +1643,7 @@ kernel void kernel_ssm_conv_f32(
     x[0] = sumf;
 }
 
-// ref: ggml.c:ggml_compute_forward_ssm_scan_f32
+// ref: ggml.c:ggml_compute_forward_ssm_scan_f32, Mamba-1 part
 kernel void kernel_ssm_scan_f32(
         device const void * src0,
         device const void * src1,
@@ -1301,46 +1651,119 @@ kernel void kernel_ssm_scan_f32(
         device const void * src3,
         device const void * src4,
         device const void * src5,
+        device const void * src6,
         device      float * dst,
         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 i1 = 0;
+    const int64_t ir = tgpig.x; // current head
+    const int64_t i3 = tgpig.y; // current seq
+
+    const uint64_t nb00 = sizeof(float);
+    const uint64_t nb10 = sizeof(float);
+    const uint64_t nb20 = sizeof(float);
 
     const int64_t nc  = args.d_state;
-    // const int64_t nr  = args.d_inner;
+    const int64_t nr  = args.d_inner;
+    const int64_t nh  = args.n_head;
+    const int64_t ng  = args.n_group;
     const int64_t n_t = args.n_seq_tokens;
-    // const int64_t n_s = args.n_seqs;
+
+    const int64_t s_off = nr * nh * n_t * args.n_seqs * sizeof(float);
+
+    device const int32_t * ids = (device const int32_t *) src6;
+
+    device const float * s0 = (device const float *) ((device const char *) src0 + ir*args.nb02 + ids[i3]*args.nb03);
+    device       float * s  = (device       float *) ((device       char *) dst  + ir*args.nb02 +      i3*args.nb03 + s_off);
 
     for (int64_t i2 = 0; i2 < n_t; ++i2) {
-        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);
+        device const float * x  = (device const float *) ((device const char *) src1 + i1*nb10 + ir*args.nb11 + i2*args.nb12 + i3*args.nb13); // {dim, nh, nt, ns}
+        device const float * dt = (device const float *) ((device const char *) src2 + ir*nb20 + i2*args.nb21 + i3*args.nb22); // {nh, nt, ns}
+        device const float * A  = (device const float *) ((device const char *) src3 + ir*args.nb31); // {d_state, nh}
+        device const float * B  = (device const float *) ((device const char *) src4 + (ir & (ng - 1))*args.nb41 + i2*args.nb42 + i3*args.nb43); // {d_state, ng, nt, ns}
+        device const float * C  = (device const float *) ((device const char *) src5 + (ir & (ng - 1))*args.nb51 + i2*args.nb52 + i3*args.nb53); // {d_state, ng, nt, ns}
+        device       float * y  = (device       float *) ((device       char *) dst  + (i1 + ir*(nr) + i2*(nh*nr) + i3*(n_t*nh*nr))*nb00); // {dim, nh, nt, ns}
 
-        if (i2 > 0) {
-            s0 = s;
-        }
-
-        // i1 == 0
-        float dt_soft_plus = dt[0] <= 20.0f ? log(1.0f + exp(dt[0])) : dt[0];
-        float x_dt = x[0] * dt_soft_plus;
+        const float dt_soft_plus = dt[0] <= 20.0f ? log(1.0f + exp(dt[0])) : dt[0];
+        const float x_dt = x[0] * dt_soft_plus;
         float sumf = 0.0f;
 
         for (int64_t i0 = 0; i0 < nc; ++i0) {
-            int64_t i = i0;
-            float state = (s0[i] * exp(dt_soft_plus * A[i])) + (B[i0] * x_dt);
+            const int64_t i = i0 + i1*nc;
+            const float state = (s0[i] * exp(dt_soft_plus * A[i0])) + (B[i0] * x_dt);
             sumf += state * C[i0];
             s[i] = state;
         }
 
         y[0] = sumf;
+
+        // recurse
+        s0 = s;
+    }
+}
+
+// ref: ggml.c:ggml_compute_forward_ssm_scan_f32, Mamba-2 part
+// TODO: optimize (e.g. by parallelizing over d_state)
+kernel void kernel_ssm_scan_f32_group(
+        device const void * src0,
+        device const void * src1,
+        device const void * src2,
+        device const void * src3,
+        device const void * src4,
+        device const void * src5,
+        device const void * src6,
+        device      float * dst,
+        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 i1 = tgpig.x;
+    const int64_t ir = tgpig.y; // current head
+    const int64_t i3 = tgpig.z; // current seq
+
+    const uint64_t nb00 = sizeof(float);
+    const uint64_t nb10 = sizeof(float);
+    const uint64_t nb20 = sizeof(float);
+
+    const int64_t nc  = args.d_state;
+    const int64_t nr  = args.d_inner;
+    const int64_t nh  = args.n_head;
+    const int64_t ng  = args.n_group;
+    const int64_t n_t = args.n_seq_tokens;
+
+    const int64_t s_off = nr * nh * n_t * args.n_seqs * sizeof(float);
+
+    device const int32_t * ids = (device const int32_t *) src6;
+
+    device const float * s0 = (device const float *) ((device const char *) src0 + ir*args.nb02 + ids[i3]*args.nb03);
+    device       float * s  = (device       float *) ((device       char *) dst  + ir*args.nb02 +      i3*args.nb03 + s_off);
+
+    for (int64_t i2 = 0; i2 < n_t; ++i2) {
+        device const float * x  = (device const float *) ((device const char *) src1 + i1*nb10 + ir*args.nb11 + i2*args.nb12 + i3*args.nb13); // {dim, nh, nt, ns}
+        device const float * dt = (device const float *) ((device const char *) src2 + ir*nb20 + i2*args.nb21 + i3*args.nb22); // {nh, nt, ns}
+        device const float * A  = (device const float *) ((device const char *) src3 + ir*args.nb31); // {1, nh}
+        device const float * B  = (device const float *) ((device const char *) src4 + (ir & (ng - 1))*args.nb41 + i2*args.nb42 + i3*args.nb43); // {d_state, ng, nt, ns}
+        device const float * C  = (device const float *) ((device const char *) src5 + (ir & (ng - 1))*args.nb51 + i2*args.nb52 + i3*args.nb53); // {d_state, ng, nt, ns}
+        device       float * y  = (device       float *) ((device       char *) dst  + (i1 + ir*(nr) + i2*(nh*nr) + i3*(n_t*nh*nr))*nb00); // {dim, nh, nt, ns}
+
+        const float dt_soft_plus = dt[0] <= 20.0f ? log(1.0f + exp(dt[0])) : dt[0];
+        const float x_dt = x[0] * dt_soft_plus;
+        const float dA = exp(dt_soft_plus * A[0]);
+        float sumf = 0.0f;
+
+        for (int64_t i0 = 0; i0 < nc; ++i0) {
+            const int64_t i = i0 + i1*nc;
+            const float state = (s0[i] * dA) + (B[i0] * x_dt);
+            sumf += state * C[i0];
+            s[i] = state;
+        }
+
+        y[0] = sumf;
+
+        // recurse
+        s0 = s;
     }
 }
 
@@ -2502,6 +2925,70 @@ template [[host_name("kernel_mul_mv_bf16_f32")]]  kernel mul_mv_t kernel_mul_mv<
 template [[host_name("kernel_mul_mv_bf16_bf16")]] kernel mul_mv_t kernel_mul_mv<bfloat, bfloat4, bfloat, bfloat4>;
 #endif
 
+template<typename T04, typename T14, typename args_t>
+void kernel_mul_mv_c4_impl(
+        args_t args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        uint3  tgpig,
+        ushort tiisg) {
+    const int r0 = tgpig.x*32 + tiisg;
+    const int rb = tgpig.y*N_MV_T_T;
+    const int im = tgpig.z;
+
+    if (r0 >= args.ne01) {
+        return;
+    }
+
+    const uint i12 = im%args.ne12;
+    const uint i13 = im/args.ne12;
+
+    const uint64_t offset0 = r0*args.nb01 + (i12/args.r2)*args.nb02 + (i13/args.r3)*args.nb03;
+
+    device const T04 * x = (device const T04 *) (src0 + offset0);
+
+    device float * dst_f32 = (device float *) dst + (uint64_t)im*args.ne0*args.ne1;
+
+    for (int row = 0; row < N_MV_T_T; ++row) {
+        int r1 = rb + row;
+        if (r1 >= args.ne11) {
+            break;
+        }
+
+        const uint64_t offset1 = r1*args.nb11 + (i12   )*args.nb12 + (i13   )*args.nb13;
+
+        device const T14 * y = (device const T14 *) (src1 + offset1);
+
+        dst_f32[(uint64_t)r1*args.ne0 + r0] = dot((float4) x[0], (float4) y[0]);
+    }
+}
+
+template<typename T04, typename T14>
+kernel void kernel_mul_mv_c4(
+        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]]) {
+    kernel_mul_mv_c4_impl<T04, T14, constant ggml_metal_kargs_mul_mv &>(
+        args,
+        src0,
+        src1,
+        dst,
+        tgpig,
+        tiisg);
+}
+
+typedef decltype(kernel_mul_mv_c4<half4, half4>) mul_mv_c4_t;
+
+template [[host_name("kernel_mul_mv_f32_f32_c4")]]  kernel mul_mv_c4_t kernel_mul_mv_c4<float4,  float4>;
+template [[host_name("kernel_mul_mv_f16_f32_c4")]]  kernel mul_mv_c4_t kernel_mul_mv_c4<half4,   float4>;
+#if defined(GGML_METAL_USE_BF16)
+template [[host_name("kernel_mul_mv_bf16_f32_c4")]] kernel mul_mv_c4_t kernel_mul_mv_c4<bfloat4, float4>;
+#endif
+
 template<typename T, typename T4>
 kernel void kernel_mul_mv_1row(
         constant ggml_metal_kargs_mul_mv & args,
@@ -3328,14 +3815,12 @@ kernel void kernel_flash_attn_ext(
     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  = DK + 2*TS; // shared memory size per query in (half)
+    const short TS = nsg*SH;      // shared memory size per query in (s_t == float)
+    const short T  = 2*DK + 2*TS; // shared memory size per query in (half)
 
-    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 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 + 2*sgitg*SH + 2*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
@@ -3354,7 +3839,7 @@ kernel void kernel_flash_attn_ext(
             if (iq1 + j < args.ne01) {
                 sq4[j*DK4 + i] = (q4_t) q4[i];
             } else {
-                sq4[j*DK4 + i] = (q4_t) 0.0f;
+                sq4[j*DK4 + i] = 0;
             }
         }
     }
@@ -3419,7 +3904,7 @@ kernel void kernel_flash_attn_ext(
                 // 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);
+                    device const half * pm = (device const half *) ((device const char *) mask + (iq1 + j)*args.nb31 + (iq2%args.ne32)*args.nb32 + (iq3%args.ne33)*args.nb33);
 
                     const float m = pm[ic + tiisg];
 
@@ -3548,20 +4033,20 @@ kernel void kernel_flash_attn_ext(
 
             // O = diag(ms)*O
             {
-                s8x8_t mm;
-                simdgroup_load(mm, ss + 2*C, TS, 0, false);
+                s8x8_t ms;
+                simdgroup_load(ms, ss + 2*C, TS, 0, false);
 
                 #pragma unroll(DV8)
                 for (short i = 0; i < DV8; ++i) {
-                    simdgroup_multiply(lo[i], mm, lo[i]);
+                    simdgroup_multiply(lo[i], ms, lo[i]);
                 }
             }
 
             // O = O + (Q*K^T)*V
             {
                 for (short cc = 0; cc < C/8; ++cc) {
-                    s8x8_t ms;
-                    simdgroup_load(ms, ss + 8*cc, TS, 0, false);
+                    s8x8_t vs;
+                    simdgroup_load(vs, ss + 8*cc, TS, 0, false);
 
                     if (is_same<vd4x4_t, v4x4_t>::value) {
                         // we can read directly from global memory
@@ -3572,7 +4057,7 @@ kernel void kernel_flash_attn_ext(
                             v8x8_t mv;
                             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]);
+                            simdgroup_multiply_accumulate(lo[i], vs, mv, lo[i]);
                         }
                     } else {
                         for (short ii = 0; ii < DV16; ii += 4) {
@@ -3593,10 +4078,10 @@ kernel void kernel_flash_attn_ext(
                                     v8x8_t mv;
 
                                     simdgroup_load(mv, sv + 16*k + 0*8, 4*16, 0, false);
-                                    simdgroup_multiply_accumulate(lo[2*(ii + k) + 0], ms, mv, lo[2*(ii + k) + 0]);
+                                    simdgroup_multiply_accumulate(lo[2*(ii + k) + 0], vs, mv, lo[2*(ii + k) + 0]);
 
                                     simdgroup_load(mv, sv + 16*k + 1*8, 4*16, 0, false);
-                                    simdgroup_multiply_accumulate(lo[2*(ii + k) + 1], ms, mv, lo[2*(ii + k) + 1]);
+                                    simdgroup_multiply_accumulate(lo[2*(ii + k) + 1], vs, mv, lo[2*(ii + k) + 1]);
                                 }
                             } else {
                                 if (ii + tx < DV16) {
@@ -3611,10 +4096,10 @@ kernel void kernel_flash_attn_ext(
                                     v8x8_t mv;
 
                                     simdgroup_load(mv, sv + 16*k + 0*8, 4*16, 0, false);
-                                    simdgroup_multiply_accumulate(lo[2*(ii + k) + 0], ms, mv, lo[2*(ii + k) + 0]);
+                                    simdgroup_multiply_accumulate(lo[2*(ii + k) + 0], vs, mv, lo[2*(ii + k) + 0]);
 
                                     simdgroup_load(mv, sv + 16*k + 1*8, 4*16, 0, false);
-                                    simdgroup_multiply_accumulate(lo[2*(ii + k) + 1], ms, mv, lo[2*(ii + k) + 1]);
+                                    simdgroup_multiply_accumulate(lo[2*(ii + k) + 1], vs, mv, lo[2*(ii + k) + 1]);
                                 }
                             }
                         }
@@ -3624,93 +4109,89 @@ kernel void kernel_flash_attn_ext(
         }
 
         // these are needed for reducing the results from the simdgroups (reuse the ss buffer)
-        for (short j = 0; j < Q; ++j) {
-            if (tiisg == 0) {
-                ss[j*TS + 0] = S[j];
-                ss[j*TS + 1] = M[j];
-            }
+        for (short j = tiisg; j < Q; j += NW) {
+            ss[j*TS + 0] = S[j];
+            ss[j*TS + 1] = M[j];
         }
     }
 
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    threadgroup float  * so  = (threadgroup float  *) (shmem_f16 + 0*DK); // reuse query data for accumulation
+    threadgroup float4 * so4 = (threadgroup float4 *) (shmem_f16 + 0*DK);
+
+    // store result to shared memory in F32
+    if (sgitg == 0) {
+        for (short i = 0; i < DV8; ++i) {
+            //simdgroup_store(lo[i], so + i*8, DV, 0, false);
+            simdgroup_float8x8 t(1.0f);
+            simdgroup_multiply(t, lo[i], t);
+            simdgroup_store(t, so + i*8, DV, 0, false);
+        }
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
     // reduce the warps sequentially
     for (ushort sg = 1; sg < nsg; ++sg) {
-        float S = { 0.0f };
-        float M = { -__FLT_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 < DV8; ++i) {
-                simdgroup_store(lo[i], so + i*8, DV, 0, false);
+            for (short j = tiisg; j < Q; j += NW) {
+                const float S0 = ss[j*TS - 1*SH + 0];
+                const float S1 = ss[j*TS        + 0];
+
+                const float M0 = ss[j*TS - 1*SH + 1];
+                const float M1 = ss[j*TS        + 1];
+
+                const float M = max(M0, M1);
+
+                float ms0 = exp(M0 - M);
+                float ms1 = exp(M1 - M);
+
+                const float S = S0*ms0 + S1*ms1;
+
+                ss[j*TS + 0] = S;
+                ss[j*TS + 1] = M;
+
+                ss[j*TS + 2*C + j - 1*SH] = ms0;
+                ss[j*TS + 2*C + j       ] = ms1;
             }
-        }
 
-        threadgroup_barrier(mem_flags::mem_threadgroup);
-
-        // the first simdgroup accumulates the results from the other simdgroups
-        if (sgitg == 0) {
-            for (short j = 0; j < Q; ++j) {
-                const float S0 = ss[j*TS +         0];
-                const float S1 = ss[j*TS + sg*SH + 0];
-
-                const float M0 = ss[j*TS +         1];
-                const float M1 = ss[j*TS + sg*SH + 1];
-
-                M = max(M0, M1);
-
-                const float ms0 = exp(M0 - M);
-                const float ms1 = exp(M1 - M);
-
-                S = S0*ms0 + S1*ms1;
-
-                if (tiisg == 0) {
-                    ss[j*TS + 0] = S;
-                    ss[j*TS + 1] = M;
-
-                    ss[j*TS + 2*C + j        ] = ms0;
-                    ss[j*TS + 2*C + j + sg*SH] = ms1;
-                }
-            }
+            //simdgroup_barrier(mem_flags::mem_threadgroup);
 
             // O_0 = diag(ms0)*O_0 + diag(ms1)*O_1
             {
                 s8x8_t ms0;
                 s8x8_t ms1;
 
-                simdgroup_load(ms0, ss + 2*C,         TS, 0, false);
-                simdgroup_load(ms1, ss + 2*C + sg*SH, TS, 0, false);
+                simdgroup_load(ms0, ss + 2*C - 1*SH, TS, 0, false);
+                simdgroup_load(ms1, ss + 2*C,        TS, 0, false);
 
                 #pragma unroll(DV8)
                 for (short i = 0; i < DV8; ++i) {
-                    o8x8_t t;
+                    simdgroup_float8x8 t;
 
                     simdgroup_load    (t, so + i*8, DV, 0, false);
-                    simdgroup_multiply(t, ms1, t);
+                    simdgroup_multiply(t, ms0, t);
 
-                    simdgroup_multiply_accumulate(lo[i], ms0, lo[i], t);
+                    simdgroup_multiply_accumulate(t, ms1, lo[i], t);
+                    simdgroup_store(t, so + i*8, DV, 0, false);
                 }
             }
         }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
     }
 
-    // store result to shared memory (reuse sq)
-    if (sgitg == 0) {
-        for (short i = 0; i < DV8; ++i) {
-            simdgroup_store(lo[i], so + i*8, DV, 0, false);
-        }
-    }
-
-    device float4 * dst4 = (device float4 *) dst;
+    threadgroup s_t * sf = (threadgroup s_t *) (shmem_f16 + 2*(nsg-1)*SH + 2*Q*DK);
 
     // final rescale with 1/S and store to global memory
-    if (sgitg == 0) {
-        for (short j = 0; j < Q && iq1 + j < args.ne01; ++j) {
-            const float S = ss[j*TS + 0];
+    for (short j = sgitg; j < Q && iq1 + j < args.ne01; j += nsg) {
+        const float S = 1.0f/sf[j*TS + 0];
 
-            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;
-            }
+        device float4 * dst4 = (device float4 *) dst + ((uint64_t)iq3*args.ne2*args.ne1 + iq2 + (uint64_t)(iq1 + j)*args.ne1)*DV4;
+
+        for (short i = tiisg; i < DV4; i += NW) {
+            dst4[i] = (float4) so4[j*DV4 + i]*S;
         }
     }
 }
@@ -3719,12 +4200,22 @@ kernel void kernel_flash_attn_ext(
 //       template to be able to explore different combinations
 //
 #define FA_TYPES \
-    half,  half4,   simdgroup_half8x8,  \
-    half,  half4x4, simdgroup_half8x8,  \
-    half,  half4x4, simdgroup_half8x8,  \
-    float,          simdgroup_float8x8, \
-    float,          simdgroup_float8x8, \
-    half,  half4,   simdgroup_half8x8
+    float,  float4,    simdgroup_float8x8, \
+    half,   half4x4,   simdgroup_half8x8,  \
+    half,   half4x4,   simdgroup_half8x8,  \
+    float,             simdgroup_float8x8, \
+    float,             simdgroup_float8x8, \
+    half,   half4,     simdgroup_half8x8
+    //float,  float4,    simdgroup_float8x8
+
+#define FA_TYPES_BF \
+    bfloat, bfloat4,   simdgroup_bfloat8x8, \
+    bfloat, bfloat4x4, simdgroup_bfloat8x8, \
+    bfloat, bfloat4x4, simdgroup_bfloat8x8, \
+    float,             simdgroup_float8x8,  \
+    float,             simdgroup_float8x8,  \
+    half,   half4,     simdgroup_half8x8
+    //float,  float4,    simdgroup_float8x8
 
 typedef decltype(kernel_flash_attn_ext<FA_TYPES, half4x4, 1, dequantize_f16, half4x4, 1, dequantize_f16, 64, 64>) flash_attn_ext_t;
 
@@ -3739,15 +4230,15 @@ template [[host_name("kernel_flash_attn_ext_f16_h256")]]         kernel flash_at
 template [[host_name("kernel_flash_attn_ext_f16_hk576_hv512")]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, half4x4,    1, dequantize_f16,  half4x4,    1, dequantize_f16,  576, 512>;
 
 #if defined(GGML_METAL_USE_BF16)
-template [[host_name("kernel_flash_attn_ext_bf16_h64" )]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 64,  64>;
-template [[host_name("kernel_flash_attn_ext_bf16_h80" )]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 80,  80>;
-template [[host_name("kernel_flash_attn_ext_bf16_h96" )]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 96,  96>;
-template [[host_name("kernel_flash_attn_ext_bf16_h112")]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 112, 112>;
-template [[host_name("kernel_flash_attn_ext_bf16_h128")]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 128, 128>;
-template [[host_name("kernel_flash_attn_ext_bf16_h192")]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 192, 192>;
-template [[host_name("kernel_flash_attn_ext_bf16_hk192_hv128")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 192, 128>;
-template [[host_name("kernel_flash_attn_ext_bf16_h256")]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 256, 256>;
-template [[host_name("kernel_flash_attn_ext_bf16_hk576_hv512")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 576, 512>;
+template [[host_name("kernel_flash_attn_ext_bf16_h64" )]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_BF, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 64,  64>;
+template [[host_name("kernel_flash_attn_ext_bf16_h80" )]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_BF, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 80,  80>;
+template [[host_name("kernel_flash_attn_ext_bf16_h96" )]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_BF, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 96,  96>;
+template [[host_name("kernel_flash_attn_ext_bf16_h112")]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_BF, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 112, 112>;
+template [[host_name("kernel_flash_attn_ext_bf16_h128")]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_BF, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 128, 128>;
+template [[host_name("kernel_flash_attn_ext_bf16_h192")]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_BF, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 192, 192>;
+template [[host_name("kernel_flash_attn_ext_bf16_hk192_hv128")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_BF, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 192, 128>;
+template [[host_name("kernel_flash_attn_ext_bf16_h256")]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_BF, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 256, 256>;
+template [[host_name("kernel_flash_attn_ext_bf16_hk576_hv512")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_BF, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 576, 512>;
 #endif
 
 template [[host_name("kernel_flash_attn_ext_q4_0_h64" )]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q4_0, 2, dequantize_q4_0, block_q4_0, 2, dequantize_q4_0, 64,  64>;
@@ -3801,6 +4292,7 @@ template [[host_name("kernel_flash_attn_ext_q8_0_h256")]]        kernel flash_at
 template [[host_name("kernel_flash_attn_ext_q8_0_hk576_hv512")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q8_0, 2, dequantize_q8_0, block_q8_0, 2, dequantize_q8_0, 576, 512>;
 
 #undef FA_TYPES
+#undef FA_TYPES_BF
 
 template<
     typename q4_t,  // query types in shared memory
@@ -3847,12 +4339,12 @@ kernel void kernel_flash_attn_ext_vec(
 
     const short T = DK + nsg*SH; // shared memory size per query in (half)
 
-  //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
+  //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 + 2*sgitg*DV       + Q*T);  // scratch buffer for the results
 
     // store the result for all queries in local memory (the O matrix from the paper)
     o4_t lo[DV4/NL];
@@ -3898,7 +4390,7 @@ kernel void kernel_flash_attn_ext_vec(
         const bool has_mask = mask != q;
 
         // pointer to the mask
-        device const half * pm = (device const half *) (mask + iq1*args.nb31);
+        device const half * pm = (device const half *) (mask + iq1*args.nb31 + (iq2%args.ne32)*args.nb32 + (iq3%args.ne33)*args.nb33);
 
         float slope = 1.0f;
 
@@ -4157,7 +4649,7 @@ kernel void kernel_flash_attn_ext_vec(
            half4,  \
     float,         \
     float, float4, \
-           half4
+           float4
 
 typedef decltype(kernel_flash_attn_ext_vec<FA_TYPES, half4, 1, dequantize_f16_t4, half4, 1, dequantize_f16_t4, 128, 128, 4>) flash_attn_ext_vec_t;
 
@@ -4271,11 +4763,16 @@ kernel void kernel_cpy(
         device  const char * src0,
         device        char * dst,
         uint3   tgpig[[threadgroup_position_in_grid]],
+        uint    tiitg[[thread_index_in_threadgroup]],
         ushort3 tpitg[[thread_position_in_threadgroup]],
-        ushort3   ntg[[threads_per_threadgroup]]) {
+        ushort3  tptg[[threads_per_threadgroup]]) {
     const int i03 = tgpig[2];
     const int i02 = tgpig[1];
-    const int i01 = tgpig[0];
+    const int i01 = tgpig[0]*tptg.y + tiitg/tptg.x;
+
+    if (i01 >= args.ne01) {
+        return;
+    }
 
     const int64_t n = i03*args.ne02*args.ne01*args.ne00 + i02*args.ne01*args.ne00 + i01*args.ne00;
 
@@ -4286,7 +4783,7 @@ kernel void kernel_cpy(
 
     device T1 * dst_data = (device T1 *) (dst + i3*args.nb3 + i2*args.nb2 + i1*args.nb1 + i0*args.nb0);
 
-    for (int64_t i00 = tpitg.x; i00 < args.ne00; i00 += ntg.x) {
+    for (int64_t i00 = tiitg%tptg.x; i00 < args.ne00; i00 += tptg.x) {
         device const T0 * src = (device T0 *)(src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01 + i00*args.nb00);
         dst_data[i00] = (T1) src[0];
     }
@@ -4306,6 +4803,7 @@ template [[host_name("kernel_cpy_bf16_f32")]]  kernel kernel_cpy_t kernel_cpy<bf
 template [[host_name("kernel_cpy_bf16_bf16")]] kernel kernel_cpy_t kernel_cpy<bfloat, bfloat>;
 #endif
 
+// TODO: templetify these kernels
 kernel void kernel_cpy_f32_q8_0(
         constant ggml_metal_kargs_cpy & args,
         device const char * src0,
@@ -4329,23 +4827,7 @@ kernel void kernel_cpy_f32_q8_0(
     for (int64_t i00 = tpitg.x*QK8_0; i00 < args.ne00; i00 += ntg.x*QK8_0) {
         device const float * src = (device float *)(src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01 + i00*args.nb00);
 
-        float amax = 0.0f; // absolute max
-
-        for (int j = 0; j < QK8_0; j++) {
-            const float v = src[j];
-            amax = MAX(amax, fabs(v));
-        }
-
-        const float d = amax / ((1 << 7) - 1);
-        const float id = d ? 1.0f/d : 0.0f;
-
-        dst_data[i00/QK8_0].d = d;
-
-        for (int j = 0; j < QK8_0; ++j) {
-            const float x0 = src[j]*id;
-
-            dst_data[i00/QK8_0].qs[j] = round(x0);
-        }
+        quantize_q8_0(src, dst_data[i00/QK8_0]);
     }
 }
 
@@ -4372,32 +4854,7 @@ kernel void kernel_cpy_f32_q4_0(
     for (int64_t i00 = tpitg.x*QK4_0; i00 < args.ne00; i00 += ntg.x*QK4_0) {
         device const float * src = (device float *)(src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01 + i00*args.nb00);
 
-        float amax = 0.0f; // absolute max
-        float max  = 0.0f;
-
-        for (int j = 0; j < QK4_0; j++) {
-            const float v = src[j];
-            if (amax < fabs(v)) {
-                amax = fabs(v);
-                max  = v;
-            }
-        }
-
-        const float d = max / -8;
-        const float id = d ? 1.0f/d : 0.0f;
-
-        dst_data[i00/QK4_0].d = d;
-
-        for (int j = 0; j < QK4_0/2; ++j) {
-            const float x0 = src[0       + j]*id;
-            const float x1 = src[QK4_0/2 + j]*id;
-
-            const uint8_t xi0 = MIN(15, (int8_t)(x0 + 8.5f));
-            const uint8_t xi1 = MIN(15, (int8_t)(x1 + 8.5f));
-
-            dst_data[i00/QK4_0].qs[j]  = xi0;
-            dst_data[i00/QK4_0].qs[j] |= xi1 << 4;
-        }
+        quantize_q4_0(src, dst_data[i00/QK4_0]);
     }
 }
 
@@ -4424,31 +4881,7 @@ kernel void kernel_cpy_f32_q4_1(
     for (int64_t i00 = tpitg.x*QK4_1; i00 < args.ne00; i00 += ntg.x*QK4_1) {
         device const float * src = (device float *)(src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01 + i00*args.nb00);
 
-        float min = FLT_MAX;
-        float max = -FLT_MAX;
-
-        for (int j = 0; j < QK4_1; j++) {
-            const float v = src[j];
-            if (min > v) min = v;
-            if (max < v) max = v;
-        }
-
-        const float d = (max - min) / ((1 << 4) - 1);
-        const float id = d ? 1.0f/d : 0.0f;
-
-        dst_data[i00/QK4_1].d = d;
-        dst_data[i00/QK4_1].m = min;
-
-        for (int j = 0; j < QK4_1/2; ++j) {
-            const float x0 = (src[0       + j] - min)*id;
-            const float x1 = (src[QK4_1/2 + j] - min)*id;
-
-            const uint8_t xi0 = MIN(15, (int8_t)(x0 + 0.5f));
-            const uint8_t xi1 = MIN(15, (int8_t)(x1 + 0.5f));
-
-            dst_data[i00/QK4_1].qs[j]  = xi0;
-            dst_data[i00/QK4_1].qs[j] |= xi1 << 4;
-        }
+        quantize_q4_1(src, dst_data[i00/QK4_1]);
     }
 }
 
@@ -4475,38 +4908,7 @@ kernel void kernel_cpy_f32_q5_0(
     for (int64_t i00 = tpitg.x*QK5_0; i00 < args.ne00; i00 += ntg.x*QK5_0) {
         device const float * src = (device float *)(src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01 + i00*args.nb00);
 
-        float amax = 0.0f; // absolute max
-        float max  = 0.0f;
-
-        for (int j = 0; j < QK5_0; j++) {
-            const float v = src[j];
-            if (amax < fabs(v)) {
-                amax = fabs(v);
-                max  = v;
-            }
-        }
-
-        const float d = max / -16;
-        const float id = d ? 1.0f/d : 0.0f;
-
-        dst_data[i00/QK5_0].d = d;
-
-        uint32_t qh = 0;
-        for (int j = 0; j < QK5_0/2; ++j) {
-            const float x0 = src[0       + j]*id;
-            const float x1 = src[QK5_0/2 + j]*id;
-
-            const uint8_t xi0 = MIN(31, (int8_t)(x0 + 16.5f));
-            const uint8_t xi1 = MIN(31, (int8_t)(x1 + 16.5f));
-
-            dst_data[i00/QK5_0].qs[j] = (xi0 & 0xf) | ((xi1 & 0xf) << 4);
-            qh |= ((xi0 & 0x10u) >> 4) << (j + 0);
-            qh |= ((xi1 & 0x10u) >> 4) << (j + QK5_0/2);
-        }
-        thread const uint8_t * qh8 = (thread const uint8_t *)&qh;
-        for (int j = 0; j < 4; ++j) {
-            dst_data[i00/QK5_0].qh[j] = qh8[j];
-        }
+        quantize_q5_0(src, dst_data[i00/QK5_0]);
     }
 }
 
@@ -4533,51 +4935,10 @@ kernel void kernel_cpy_f32_q5_1(
     for (int64_t i00 = tpitg.x*QK5_1; i00 < args.ne00; i00 += ntg.x*QK5_1) {
         device const float * src = (device float *)(src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01 + i00*args.nb00);
 
-        float max = src[0];
-        float min = src[0];
-
-        for (int j = 1; j < QK5_1; j++) {
-            const float v = src[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;
-
-        dst_data[i00/QK5_1].d = d;
-        dst_data[i00/QK5_1].m = min;
-
-        uint32_t qh = 0;
-        for (int j = 0; j < QK5_1/2; ++j) {
-            const float x0 = (src[0       + j] - min)*id;
-            const float x1 = (src[QK5_1/2 + j] - min)*id;
-
-            const uint8_t xi0 = (uint8_t)(x0 + 0.5f);
-            const uint8_t xi1 = (uint8_t)(x1 + 0.5f);
-
-            dst_data[i00/QK5_1].qs[j] = (xi0 & 0xf) | ((xi1 & 0xf) << 4);
-            qh |= ((xi0 & 0x10u) >> 4) << (j + 0);
-            qh |= ((xi1 & 0x10u) >> 4) << (j + QK5_1/2);
-        }
-        thread const uint8_t * qh8 = (thread const uint8_t *)&qh;
-        for (int j = 0; j < 4; ++j) {
-            dst_data[i00/QK5_1].qh[j] = qh8[j];
-        }
+        quantize_q5_1(src, dst_data[i00/QK5_1]);
     }
 }
 
-static inline int best_index_int8(int n, constant float * 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;
-}
-
 kernel void kernel_cpy_f32_iq4_nl(
         constant ggml_metal_kargs_cpy & args,
         device const char * src0,
@@ -4601,40 +4962,7 @@ kernel void kernel_cpy_f32_iq4_nl(
     for (int64_t i00 = tpitg.x*QK4_NL; i00 < args.ne00; i00 += ntg.x*QK4_NL) {
         device const float * src = (device float *)(src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01 + i00*args.nb00);
 
-        float amax = 0.0f; // absolute max
-        float max  = 0.0f;
-
-        for (int j = 0; j < QK4_NL; j++) {
-            const float v = src[j];
-            if (amax < fabs(v)) {
-                amax = fabs(v);
-                max  = v;
-            }
-        }
-
-        const float d = max / kvalues_iq4nl_f[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 = src[0        + j]*id;
-            const float x1 = src[QK4_NL/2 + j]*id;
-
-            const uint8_t xi0 = best_index_int8(16, kvalues_iq4nl_f, x0);
-            const uint8_t xi1 = best_index_int8(16, kvalues_iq4nl_f, x1);
-
-            dst_data[i00/QK4_NL].qs[j] = xi0 | (xi1 << 4);
-
-            const float v0 = kvalues_iq4nl_f[xi0];
-            const float v1 = kvalues_iq4nl_f[xi1];
-            const float w0 = src[0        + j]*src[0        + j];
-            const float w1 = src[QK4_NL/2 + j]*src[QK4_NL/2 + j];
-            sumqx += w0*v0*src[j] + w1*v1*src[QK4_NL/2 + j];
-            sumq2 += w0*v0*v0 + w1*v1*v1;
-
-        }
-
-        dst_data[i00/QK4_NL].d = sumq2 > 0 ? sumqx/sumq2 : d;
+        quantize_iq4_nl(src, dst_data[i00/QK4_NL]);
     }
 }
 
@@ -6315,10 +6643,10 @@ kernel void kernel_mul_mv_iq4_xs_f32(
 
 template<typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread float4x4 &)>
 kernel void kernel_get_rows_q(
+        constant ggml_metal_kargs_get_rows & args,
         device const  void * src0,
         device const  void * src1,
         device       float * dst,
-        constant ggml_metal_kargs_get_rows & args,
         uint3                tgpig[[threadgroup_position_in_grid]],
         uint                 tiitg[[thread_index_in_threadgroup]],
         uint3                tptg [[threads_per_threadgroup]]) {
@@ -6338,10 +6666,10 @@ kernel void kernel_get_rows_q(
 
 template<typename T>
 kernel void kernel_get_rows_f(
+        constant ggml_metal_kargs_get_rows & args,
         device const  void * src0,
         device const  void * src1,
         device       float * dst,
-        constant ggml_metal_kargs_get_rows & args,
         uint3                tgpig[[threadgroup_position_in_grid]],
         uint                 tiitg[[thread_index_in_threadgroup]],
         uint3                tptg [[threads_per_threadgroup]]) {
@@ -6359,10 +6687,10 @@ kernel void kernel_get_rows_f(
 }
 
 kernel void kernel_get_rows_i32(
+        constant ggml_metal_kargs_get_rows & args,
         device const  void * src0,
         device const  void * src1,
         device     int32_t * dst,
-        constant ggml_metal_kargs_get_rows & args,
         uint3                tgpig[[threadgroup_position_in_grid]],
         uint                 tiitg[[thread_index_in_threadgroup]],
         uint3                tptg [[threads_per_threadgroup]]) {
@@ -6379,6 +6707,67 @@ kernel void kernel_get_rows_i32(
     }
 }
 
+template<typename block_q, void (*quantize_func)(device const float *, device block_q &)>
+kernel void kernel_set_rows_q32(
+        constant ggml_metal_kargs_set_rows & args,
+        device const  void * src0,
+        device const  void * src1,
+        device       float * dst,
+        uint3                tgpig[[threadgroup_position_in_grid]],
+        uint                 tiitg[[thread_index_in_threadgroup]],
+        uint3                tptg [[threads_per_threadgroup]]) {
+    const int32_t i03 = tgpig.z;
+    const int32_t i02 = tgpig.y;
+
+    const int32_t i12 = i03%args.ne12;
+    const int32_t i11 = i02%args.ne11;
+
+    const int32_t i01 = tgpig.x*tptg.y + tiitg/tptg.x;
+    if (i01 >= args.ne01) {
+        return;
+    }
+
+    const int32_t i10 = i01;
+    const int64_t i1 = ((const device int64_t *) ((const device char *) src1 + i10*args.nb10 + i11*args.nb11 + i12*args.nb12))[0];
+
+          device block_q * dst_row = (      device block_q *) ((      device char *) dst  +  i1*args.nb1  + i02*args.nb2  + i03*args.nb3);
+    const device float   * src_row = (const device float   *) ((const device char *) src0 + i01*args.nb01 + i02*args.nb02 + i03*args.nb03);
+
+    for (int ind = tiitg%tptg.x; ind < args.nk0; ind += tptg.x) {
+        quantize_func(src_row + 32*ind, dst_row[ind]);
+    }
+}
+
+template<typename T>
+kernel void kernel_set_rows_f(
+        constant ggml_metal_kargs_set_rows & args,
+        device const  void * src0,
+        device const  void * src1,
+        device       float * dst,
+        uint3                tgpig[[threadgroup_position_in_grid]],
+        uint                 tiitg[[thread_index_in_threadgroup]],
+        uint3                tptg [[threads_per_threadgroup]]) {
+    const int32_t i03 = tgpig.z;
+    const int32_t i02 = tgpig.y;
+
+    const int32_t i12 = i03%args.ne12;
+    const int32_t i11 = i02%args.ne11;
+
+    const int32_t i01 = tgpig.x*tptg.y + tiitg/tptg.x;
+    if (i01 >= args.ne01) {
+        return;
+    }
+
+    const int32_t i10 = i01;
+    const int64_t i1 = ((const device int64_t *) ((const device char *) src1 + i10*args.nb10 + i11*args.nb11 + i12*args.nb12))[0];
+
+    device T     * dst_row = (      device T     *) ((      device char *) dst  +  i1*args.nb1  + i02*args.nb2  + i03*args.nb3);
+    const device float * src_row = (const device float *) ((const device char *) src0 + i01*args.nb01 + i02*args.nb02 + i03*args.nb03);
+
+    for (int ind = tiitg%tptg.x; ind < args.nk0; ind += tptg.x) {
+        dst_row[ind] = (T) src_row[ind];
+    }
+}
 
 #define BLOCK_SIZE_M 64 // 8 simdgroup matrices from matrix A
 #define BLOCK_SIZE_N 32 // 4 simdgroup matrices from matrix B
@@ -6802,6 +7191,27 @@ template [[host_name("kernel_get_rows_iq1_m")]]   kernel get_rows_q_t kernel_get
 template [[host_name("kernel_get_rows_iq4_nl")]]  kernel get_rows_q_t kernel_get_rows_q<block_iq4_nl,  2,     dequantize_iq4_nl>;
 template [[host_name("kernel_get_rows_iq4_xs")]]  kernel get_rows_q_t kernel_get_rows_q<block_iq4_xs,  QK_NL, dequantize_iq4_xs>;
 
+//
+// set rows
+//
+
+typedef decltype(kernel_set_rows_f<float>) set_rows_f_t;
+
+template [[host_name("kernel_set_rows_f32")]]  kernel set_rows_f_t kernel_set_rows_f<float>;
+template [[host_name("kernel_set_rows_f16")]]  kernel set_rows_f_t kernel_set_rows_f<half>;
+#if defined(GGML_METAL_USE_BF16)
+template [[host_name("kernel_set_rows_bf16")]] kernel set_rows_f_t kernel_set_rows_f<bfloat>;
+#endif
+
+typedef decltype(kernel_set_rows_q32<block_q8_0, quantize_q8_0>) set_rows_q32_t;
+
+template [[host_name("kernel_set_rows_q8_0")]]   kernel set_rows_q32_t kernel_set_rows_q32<block_q8_0,   quantize_q8_0>;
+template [[host_name("kernel_set_rows_q4_0")]]   kernel set_rows_q32_t kernel_set_rows_q32<block_q4_0,   quantize_q4_0>;
+template [[host_name("kernel_set_rows_q4_1")]]   kernel set_rows_q32_t kernel_set_rows_q32<block_q4_1,   quantize_q4_1>;
+template [[host_name("kernel_set_rows_q5_0")]]   kernel set_rows_q32_t kernel_set_rows_q32<block_q5_0,   quantize_q5_0>;
+template [[host_name("kernel_set_rows_q5_1")]]   kernel set_rows_q32_t kernel_set_rows_q32<block_q5_1,   quantize_q5_1>;
+template [[host_name("kernel_set_rows_iq4_nl")]] kernel set_rows_q32_t kernel_set_rows_q32<block_iq4_nl, quantize_iq4_nl>;
+
 //
 // matrix-matrix multiplication
 //
diff --git a/ggml/src/ggml-musa/mudnn.cuh b/ggml/src/ggml-musa/mudnn.cuh
index a63be5755..c30128561 100644
--- a/ggml/src/ggml-musa/mudnn.cuh
+++ b/ggml/src/ggml-musa/mudnn.cuh
@@ -1,7 +1,7 @@
 #pragma once
 
-#include "../include/ggml.h"
-#include "../ggml-cuda/common.cuh"
+#include "ggml-cuda/common.cuh"
+#include "ggml.h"
 
 // Asynchronously copies data from src tensor to dst tensor using the provided context.
 // Returns a musaError_t indicating success or failure.
diff --git a/ggml/src/ggml-opencl/CMakeLists.txt b/ggml/src/ggml-opencl/CMakeLists.txt
index 9f930c70b..45a488334 100644
--- a/ggml/src/ggml-opencl/CMakeLists.txt
+++ b/ggml/src/ggml-opencl/CMakeLists.txt
@@ -65,6 +65,7 @@ set(GGML_OPENCL_KERNELS
     gemv_noshuffle_general
     gemv_noshuffle
     get_rows
+    glu
     group_norm
     im2col_f32
     im2col_f16
@@ -80,6 +81,7 @@ set(GGML_OPENCL_KERNELS
     mul_mv_q4_0_f32_1d_8x_flat
     mul_mv_q4_0_f32_1d_16x_flat
     mul_mv_q6_k
+    mul_mv_id_q4_0_f32_8x_flat
     mul
     norm
     relu
@@ -95,6 +97,12 @@ set(GGML_OPENCL_KERNELS
     sub
     sum_rows
     transpose
+    concat
+    tsembd
+    upscale
+    tanh
+    pad
+    repeat
 )
 
 foreach (K ${GGML_OPENCL_KERNELS})
diff --git a/ggml/src/ggml-opencl/ggml-opencl.cpp b/ggml/src/ggml-opencl/ggml-opencl.cpp
index 5dbe97ab2..a9fc03903 100644
--- a/ggml/src/ggml-opencl/ggml-opencl.cpp
+++ b/ggml/src/ggml-opencl/ggml-opencl.cpp
@@ -231,160 +231,7 @@ static ggml_cl_compiler_version get_adreno_cl_compiler_version(const char *drive
     return { type, major, minor, patch };
 }
 
-struct ggml_backend_opencl_context;
-
-// backend device context
-struct ggml_backend_opencl_device_context {
-    cl_platform_id platform;
-    std::string platform_name;
-
-    cl_device_id   device;
-    std::string    device_name;
-    cl_device_type device_type;
-    std::string    device_version;
-
-    // Initialized by ggml_cl2_init().
-    ggml_backend_opencl_context * backend_ctx = nullptr;
-
-    // Initialized by ggml_backend_opencl_device_get_buffer_type()
-    ggml_backend_buffer_type buffer_type;
-
-    cl_context context = nullptr;
-};
-
-// backend context
-struct ggml_backend_opencl_context {
-    cl_device_id device;
-    std::string device_name;
-
-    std::string driver_version;
-
-    GPU_FAMILY gpu_family;
-    ADRENO_GPU_GEN adreno_gen;
-
-    cl_int alignment;
-    size_t max_alloc_size;
-    bool fp16_support;
-    bool has_vector_subgroup_broadcast;
-    ggml_cl_compiler_version adreno_cl_compiler_version;
-
-    int adreno_wave_size;
-
-    cl_bool non_uniform_workgroups;
-
-    cl_context context;
-    cl_command_queue queue;
-
-    cl_program program_add;
-    cl_program program_clamp;
-    cl_program program_cpy;
-    cl_program program_cvt;
-    cl_program program_diag_mask_inf;
-    cl_program program_gelu;
-    cl_program program_gemv_noshuffle_general;
-    cl_program program_gemv_noshuffle;
-    cl_program program_get_rows;
-    cl_program program_im2col_f16;
-    cl_program program_im2col_f32;
-    cl_program program_mul_mat_Ab_Bi_8x4;
-    cl_program program_mul_mv_q4_0_f32;
-    cl_program program_mul_mv_q4_0_f32_v;
-    cl_program program_mul_mv_q4_0_f32_8x_flat;
-    cl_program program_mul_mv_q4_0_f32_1d_8x_flat;
-    cl_program program_mul_mv_q4_0_f32_1d_16x_flat;
-    cl_program program_mul_mv_q6_K;
-    cl_program program_mul_mv_f16_f16;
-    cl_program program_mul_mv_f16_f32_1row;
-    cl_program program_mul_mv_f16_f32_l4;
-    cl_program program_mul_mv_f16_f32;
-    cl_program program_mul_mv_f32_f32;
-    cl_program program_mul;
-    cl_program program_div;
-    cl_program program_sub;
-    cl_program program_norm;
-    cl_program program_relu;
-    cl_program program_rms_norm;
-    cl_program program_group_norm;
-    cl_program program_rope;
-    cl_program program_scale;
-    cl_program program_silu;
-    cl_program program_sigmoid;
-    cl_program program_softmax_f32;
-    cl_program program_softmax_f16;
-    cl_program program_softmax_4_f32;
-    cl_program program_softmax_4_f16;
-    cl_program program_argsort_f32_i32;
-    cl_program program_sum_rows_f32;
-
-    cl_kernel kernel_add, kernel_add_row;
-    cl_kernel kernel_mul, kernel_mul_row;
-    cl_kernel kernel_div, kernel_div_row;
-    cl_kernel kernel_sub, kernel_sub_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_sigmoid_f32, kernel_sigmoid_f16;
-    cl_kernel kernel_clamp;
-    cl_kernel kernel_norm;
-    cl_kernel kernel_rms_norm;
-    cl_kernel kernel_group_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;
-    cl_kernel kernel_mul_mat_f16_f32_1row;
-    cl_kernel kernel_mul_mat_f16_f32;
-    cl_kernel kernel_mul_mat_f16_f32_l4;
-    cl_kernel kernel_mul_mat_q4_0_f32, kernel_mul_mat_q4_0_f32_v;
-    cl_kernel kernel_convert_block_q4_0, kernel_restore_block_q4_0;
-    cl_kernel kernel_mul_mat_q4_0_f32_8x_flat;
-    cl_kernel kernel_convert_block_q4_0_noshuffle;
-    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;
-    cl_kernel kernel_argsort_f32_i32;
-    cl_kernel kernel_sum_rows_f32;
-
-#ifdef GGML_OPENCL_USE_ADRENO_KERNELS
-    // Transpose kernels
-    cl_program program_transpose;
-
-    cl_kernel kernel_transpose_32;
-    cl_kernel kernel_transpose_32_16;
-    cl_kernel kernel_transpose_16;
-
-    cl_mem A_s_d_max;            // max scale buffer size for transpose
-    cl_mem A_q_d_max;            // max weight buffer size for transpose
-    cl_mem B_d_max;              // max activation buffer size for transpose
-
-    // Gemm and Gemv related programs, kernels, etc
-    cl_program program_CL_gemm;
-    cl_program program_CL_gemv_general;
-    cl_program program_CL_gemv_4096_1_11008;
-    cl_program program_CL_gemv_4096_1_4096;
-    cl_program program_CL_gemv_11008_1_4096;
-    cl_program program_CL_gemv_32000_1_4096;
-    cl_kernel CL_mul_mat_Ab_Bi_8x4;
-    cl_kernel CL_mul_mat_vec_q4_0_f32_1d_4x_flat_general;
-    cl_kernel CL_mul_mat_vec_q4_0_f32_1d_4x_flat_4096_1_11008;
-    cl_kernel CL_mul_mat_vec_q4_0_f32_1d_4x_flat_4096_1_4096;
-    cl_kernel CL_mul_mat_vec_q4_0_f32_1d_4x_flat_11008_1_4096;
-    cl_kernel CL_mul_mat_vec_q4_0_f32_1d_4x_flat_32000_1_4096;
-#endif // GGML_OPENCL_USE_ADRENO_KERNELS
-};
-
-// All registered devices with a default device in the front.
-static std::vector<ggml_backend_device> g_ggml_backend_opencl_devices;
-
 // Profiling
-#ifdef GGML_OPENCL_PROFILING
 struct ProfilingInfo {
     std::string op_name;
     std::string kernel_name;
@@ -416,8 +263,324 @@ struct ProfilingInfo {
     size_t output_size[4];
 };
 
-std::vector<ProfilingInfo> g_profiling_info;
+static void populateProfilingInfo(
+        ProfilingInfo& info, cl_event evt, cl_kernel kernel, cl_uint work_dim,
+        size_t global_size[3], size_t local_size[3],
+        const ggml_tensor * tensor) {
+    info.op_name     = tensor->name;
+    info.kernel      = kernel;
+    info.evt         = evt;
+
+    // 0 means not specified, e.g., 2D workgroup, or NULL for driver to choose
+    info.local_size[0] = 0;
+    info.local_size[1] = 0;
+    info.local_size[2] = 0;
+
+    info.global_size[0] = 0;
+    info.global_size[1] = 0;
+    info.global_size[2] = 0;
+
+    if (local_size) {
+        for (cl_uint i = 0; i < work_dim; ++i) {
+            info.local_size[i] = local_size[i];
+        }
+    }
+
+    for (cl_uint i = 0; i < work_dim; ++i) {
+        info.global_size[i] = global_size[i];
+    }
+
+    info.output_size[0] = tensor->ne[0];
+    info.output_size[1] = tensor->ne[1];
+    info.output_size[2] = tensor->ne[2];
+    info.output_size[3] = tensor->ne[3];
+}
+
+struct ggml_backend_opencl_context;
+
+// backend device context
+struct ggml_backend_opencl_device_context {
+    cl_platform_id platform;
+    std::string platform_name;
+
+    cl_device_id   device;
+    std::string    device_name;
+    cl_device_type device_type;
+    std::string    device_version;
+
+    // Initialized by ggml_cl2_init().
+    ggml_backend_opencl_context * backend_ctx = nullptr;
+
+    // Initialized by ggml_backend_opencl_device_get_buffer_type()
+    ggml_backend_buffer_type buffer_type;
+
+    cl_context context = nullptr;
+};
+
+// backend context
+struct ggml_backend_opencl_context {
+    int ref_count;
+
+    cl_device_id device;
+    std::string device_name;
+
+    std::string driver_version;
+
+    GPU_FAMILY gpu_family;
+    ADRENO_GPU_GEN adreno_gen;
+
+    cl_int alignment;
+    size_t max_alloc_size;
+    bool fp16_support;
+    bool has_vector_subgroup_broadcast;
+    ggml_cl_compiler_version adreno_cl_compiler_version;
+
+    int adreno_wave_size;
+
+    cl_bool non_uniform_workgroups;
+
+    cl_context context;
+    cl_command_queue queue;
+
+    cl_program program_add;
+    cl_program program_clamp;
+    cl_program program_cpy;
+    cl_program program_cvt;
+    cl_program program_diag_mask_inf;
+    cl_program program_gelu;
+    cl_program program_gemv_noshuffle_general;
+    cl_program program_gemv_noshuffle;
+    cl_program program_get_rows;
+    cl_program program_glu;
+    cl_program program_im2col_f16;
+    cl_program program_im2col_f32;
+    cl_program program_mul_mat_Ab_Bi_8x4;
+    cl_program program_mul_mv_q4_0_f32;
+    cl_program program_mul_mv_q4_0_f32_v;
+    cl_program program_mul_mv_q4_0_f32_8x_flat;
+    cl_program program_mul_mv_q4_0_f32_1d_8x_flat;
+    cl_program program_mul_mv_q4_0_f32_1d_16x_flat;
+    cl_program program_mul_mv_q6_K;
+    cl_program program_mul_mv_f16_f16;
+    cl_program program_mul_mv_f16_f32_1row;
+    cl_program program_mul_mv_f16_f32_l4;
+    cl_program program_mul_mv_f16_f32;
+    cl_program program_mul_mv_f32_f32;
+    cl_program program_mul;
+    cl_program program_div;
+    cl_program program_sub;
+    cl_program program_norm;
+    cl_program program_relu;
+    cl_program program_rms_norm;
+    cl_program program_group_norm;
+    cl_program program_rope;
+    cl_program program_scale;
+    cl_program program_silu;
+    cl_program program_sigmoid;
+    cl_program program_softmax_f32;
+    cl_program program_softmax_f16;
+    cl_program program_softmax_4_f32;
+    cl_program program_softmax_4_f16;
+    cl_program program_argsort_f32_i32;
+    cl_program program_sum_rows_f32;
+    cl_program program_repeat;
+    cl_program program_pad;
+    cl_program program_tanh;
+    cl_program program_upscale;
+    cl_program program_concat;
+    cl_program program_tsembd;
+    cl_program program_mul_mv_id_q4_0_f32_8x_flat;
+
+    cl_kernel kernel_add, kernel_add_row;
+    cl_kernel kernel_mul, kernel_mul_row;
+    cl_kernel kernel_div, kernel_div_row;
+    cl_kernel kernel_sub, kernel_sub_row;
+    cl_kernel kernel_scale;
+    cl_kernel kernel_silu, kernel_silu_4;
+    cl_kernel kernel_gelu, kernel_gelu_4;
+    cl_kernel kernel_gelu_erf, kernel_gelu_erf_4;
+    cl_kernel kernel_gelu_quick, kernel_gelu_quick_4;
+    cl_kernel kernel_relu;
+    cl_kernel kernel_sigmoid_f32, kernel_sigmoid_f16;
+    cl_kernel kernel_clamp;
+    cl_kernel kernel_geglu, kernel_reglu, kernel_swiglu, kernel_geglu_erf, kernel_geglu_quick,
+              kernel_geglu_f16, kernel_reglu_f16, kernel_swiglu_f16, kernel_geglu_erf_f16, kernel_geglu_quick_f16;
+    cl_kernel kernel_norm;
+    cl_kernel kernel_rms_norm;
+    cl_kernel kernel_group_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;
+    cl_kernel kernel_mul_mat_f16_f32_1row;
+    cl_kernel kernel_mul_mat_f16_f32;
+    cl_kernel kernel_mul_mat_f16_f32_l4;
+    cl_kernel kernel_mul_mat_q4_0_f32, kernel_mul_mat_q4_0_f32_v;
+    cl_kernel kernel_convert_block_q4_0, kernel_restore_block_q4_0;
+    cl_kernel kernel_mul_mat_q4_0_f32_8x_flat;
+    cl_kernel kernel_convert_block_q4_0_noshuffle;
+    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;
+    cl_kernel kernel_argsort_f32_i32;
+    cl_kernel kernel_sum_rows_f32;
+    cl_kernel kernel_repeat;
+    cl_kernel kernel_pad;
+    cl_kernel kernel_tanh_f32_nd;
+    cl_kernel kernel_tanh_f16_nd;
+    cl_kernel kernel_upscale;
+    cl_kernel kernel_upscale_bilinear;
+    cl_kernel kernel_concat_f32_contiguous;
+    cl_kernel kernel_concat_f32_non_contiguous;
+    cl_kernel kernel_timestep_embedding;
+    cl_kernel kernel_mul_mv_id_q4_0_f32_8x_flat;
+
+    std::vector<ProfilingInfo> profiling_info;
+
+    void write_profiling_info() {
+        FILE * fperf = fopen("cl_profiling.csv", "w");
+        if (!fperf) {
+            GGML_LOG_ERROR("Failed to open cl_profiling.csv\n");
+            return;
+        }
+
+        // Populate profiling info
+        for (ProfilingInfo & info : 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, 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 : profiling_info) {
+            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[1], info.local_size[2],
+                info.output_size[0], info.output_size[1], info.output_size[2], info.output_size[3]);
+        }
+        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 : 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);
+    }
+
+    void enqueue_ndrange_kernel(cl_kernel kernel, cl_uint work_dim, size_t *global_work_size, size_t *local_work_size, const ggml_tensor * tensor) {
+#ifdef GGML_OPENCL_PROFILING
+        cl_event evt;
+        CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, work_dim, NULL, global_work_size, local_work_size, 0, NULL, &evt));
+
+        profiling_info.emplace_back();
+        populateProfilingInfo(profiling_info.back(), evt, kernel, work_dim, global_work_size, local_work_size, tensor);
+#else
+        GGML_UNUSED(tensor);
+        CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, work_dim, NULL, global_work_size, local_work_size, 0, NULL, NULL));
 #endif
+    }
+
+#ifdef GGML_OPENCL_USE_ADRENO_KERNELS
+    // Transpose kernels
+    cl_program program_transpose;
+
+    cl_kernel kernel_transpose_32;
+    cl_kernel kernel_transpose_32_16;
+    cl_kernel kernel_transpose_16;
+
+    cl_mem A_s_d_max;            // max scale buffer size for transpose
+    cl_mem A_q_d_max;            // max weight buffer size for transpose
+    cl_mem B_d_max;              // max activation buffer size for transpose
+
+    // Gemm and Gemv related programs, kernels, etc
+    cl_program program_CL_gemm;
+    cl_program program_CL_gemv_general;
+    cl_program program_CL_gemv_4096_1_11008;
+    cl_program program_CL_gemv_4096_1_4096;
+    cl_program program_CL_gemv_11008_1_4096;
+    cl_program program_CL_gemv_32000_1_4096;
+    cl_kernel CL_mul_mat_Ab_Bi_8x4;
+    cl_kernel CL_mul_mat_vec_q4_0_f32_1d_4x_flat_general;
+    cl_kernel CL_mul_mat_vec_q4_0_f32_1d_4x_flat_4096_1_11008;
+    cl_kernel CL_mul_mat_vec_q4_0_f32_1d_4x_flat_4096_1_4096;
+    cl_kernel CL_mul_mat_vec_q4_0_f32_1d_4x_flat_11008_1_4096;
+    cl_kernel CL_mul_mat_vec_q4_0_f32_1d_4x_flat_32000_1_4096;
+#endif // GGML_OPENCL_USE_ADRENO_KERNELS
+
+    void free() {
+        ref_count--;
+        if (ref_count == 0) {
+#ifdef GGML_OPENCL_PROFILING
+            write_profiling_info();
+#endif
+        }
+    }
+};
+
+// All registered devices with a default device in the front.
+static std::vector<ggml_backend_device> g_ggml_backend_opencl_devices;
 
 inline std::string read_file(const std::string &path) {
   std::ifstream ifs(path);
@@ -574,11 +737,38 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
 
         CL_CHECK((backend_ctx->kernel_gelu         = clCreateKernel(backend_ctx->program_gelu, "kernel_gelu", &err), err));
         CL_CHECK((backend_ctx->kernel_gelu_4       = clCreateKernel(backend_ctx->program_gelu, "kernel_gelu_4", &err), err));
+        CL_CHECK((backend_ctx->kernel_gelu_erf     = clCreateKernel(backend_ctx->program_gelu, "kernel_gelu_erf", &err), err));
+        CL_CHECK((backend_ctx->kernel_gelu_erf_4   = clCreateKernel(backend_ctx->program_gelu, "kernel_gelu_erf_4", &err), err));
         CL_CHECK((backend_ctx->kernel_gelu_quick   = clCreateKernel(backend_ctx->program_gelu, "kernel_gelu_quick", &err), err));
         CL_CHECK((backend_ctx->kernel_gelu_quick_4 = clCreateKernel(backend_ctx->program_gelu, "kernel_gelu_quick_4", &err), err));
         GGML_LOG_CONT(".");
     }
 
+    // glu
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "glu.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("glu.cl");
+#endif
+        backend_ctx->program_glu =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_geglu           = clCreateKernel(backend_ctx->program_glu, "kernel_geglu", &err), err));
+        CL_CHECK((backend_ctx->kernel_reglu           = clCreateKernel(backend_ctx->program_glu, "kernel_reglu", &err), err));
+        CL_CHECK((backend_ctx->kernel_swiglu          = clCreateKernel(backend_ctx->program_glu, "kernel_swiglu", &err), err));
+        CL_CHECK((backend_ctx->kernel_geglu_erf       = clCreateKernel(backend_ctx->program_glu, "kernel_geglu_erf", &err), err));
+        CL_CHECK((backend_ctx->kernel_geglu_quick     = clCreateKernel(backend_ctx->program_glu, "kernel_geglu_quick", &err), err));
+        CL_CHECK((backend_ctx->kernel_geglu_f16       = clCreateKernel(backend_ctx->program_glu, "kernel_geglu_f16", &err), err));
+        CL_CHECK((backend_ctx->kernel_reglu_f16       = clCreateKernel(backend_ctx->program_glu, "kernel_reglu_f16", &err), err));
+        CL_CHECK((backend_ctx->kernel_swiglu_f16      = clCreateKernel(backend_ctx->program_glu, "kernel_swiglu_f16", &err), err));
+        CL_CHECK((backend_ctx->kernel_geglu_erf_f16   = clCreateKernel(backend_ctx->program_glu, "kernel_geglu_erf_f16", &err), err));
+        CL_CHECK((backend_ctx->kernel_geglu_quick_f16 = clCreateKernel(backend_ctx->program_glu, "kernel_geglu_quick_f16", &err), err));
+        GGML_LOG_CONT(".");
+    }
+
     // get_rows
     {
 #ifdef GGML_OPENCL_EMBED_KERNELS
@@ -1097,6 +1287,166 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
         GGML_LOG_CONT(".");
     }
 
+    // repeat
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "repeat.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("repeat.cl");
+#endif
+        if (!kernel_src.empty()) {
+            backend_ctx->program_repeat =
+                build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+            CL_CHECK((backend_ctx->kernel_repeat = clCreateKernel(backend_ctx->program_repeat, "kernel_repeat", &err), err));
+            GGML_LOG_CONT(".");
+        } else {
+            GGML_LOG_WARN("ggml_opencl: repeat kernel source not found or empty. Repeat operations will not be available.\n");
+            backend_ctx->program_repeat = nullptr;
+            backend_ctx->kernel_repeat = nullptr;
+        }
+    }
+
+    // pad
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "pad.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("pad.cl");
+#endif
+        if (!kernel_src.empty()) {
+            backend_ctx->program_pad =
+                build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+            CL_CHECK((backend_ctx->kernel_pad = clCreateKernel(backend_ctx->program_pad, "kernel_pad", &err), err));
+            GGML_LOG_CONT(".");
+        } else {
+            GGML_LOG_WARN("ggml_opencl: pad kernel source not found or empty. Pad operations will not be available.\n");
+            backend_ctx->program_pad = nullptr;
+            backend_ctx->kernel_pad = nullptr;
+        }
+    }
+
+    // tanh
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "tanh.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("tanh.cl");
+#endif
+        if (!kernel_src.empty()) {
+            backend_ctx->program_tanh =
+                build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+            CL_CHECK((backend_ctx->kernel_tanh_f32_nd = clCreateKernel(backend_ctx->program_tanh, "kernel_tanh_f32_nd", &err), err));
+            CL_CHECK((backend_ctx->kernel_tanh_f16_nd = clCreateKernel(backend_ctx->program_tanh, "kernel_tanh_f16_nd", &err), err));
+            GGML_LOG_CONT(".");
+        } else {
+            GGML_LOG_WARN("ggml_opencl: tanh kernel source not found or empty. Tanh operation will not be available.\n");
+            backend_ctx->program_tanh = nullptr;
+            backend_ctx->kernel_tanh_f32_nd = nullptr;
+            backend_ctx->kernel_tanh_f16_nd = nullptr;
+        }
+    }
+
+    // upscale
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "upscale.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("upscale.cl");
+#endif
+        if (!kernel_src.empty()) {
+            backend_ctx->program_upscale =
+                build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+            CL_CHECK((backend_ctx->kernel_upscale = clCreateKernel(backend_ctx->program_upscale, "kernel_upscale", &err), err));
+            if (backend_ctx->program_upscale) {
+                 cl_int err_bilinear;
+                 backend_ctx->kernel_upscale_bilinear = clCreateKernel(backend_ctx->program_upscale, "kernel_upscale_bilinear", &err_bilinear);
+                 if (err_bilinear != CL_SUCCESS) {
+                    GGML_LOG_WARN("ggml_opencl: kernel_upscale_bilinear not found in upscale.cl. Bilinear upscale will not be available. Error: %d\n", err_bilinear);
+                    backend_ctx->kernel_upscale_bilinear = nullptr;
+                 }
+            } else {
+                backend_ctx->kernel_upscale_bilinear = nullptr;
+            }
+            GGML_LOG_CONT(".");
+        } else {
+            GGML_LOG_WARN("ggml_opencl: upscale kernel source not found or empty. Upscale operations will not be available.\n");
+            backend_ctx->program_upscale = nullptr;
+            backend_ctx->kernel_upscale = nullptr;
+            backend_ctx->kernel_upscale_bilinear = nullptr;
+        }
+    }
+
+    // concat
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "concat.cl.h"
+        };
+#else
+
+        const std::string kernel_src = read_file("concat.cl");
+#endif
+        if (!kernel_src.empty()) {
+            backend_ctx->program_concat =
+                build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+            CL_CHECK((backend_ctx->kernel_concat_f32_contiguous = clCreateKernel(backend_ctx->program_concat, "kernel_concat_f32_contiguous", &err), err));
+            CL_CHECK((backend_ctx->kernel_concat_f32_non_contiguous = clCreateKernel(backend_ctx->program_concat, "kernel_concat_f32_non_contiguous", &err), err));
+            GGML_LOG_CONT(".");
+        } else {
+            GGML_LOG_WARN("ggml_opencl: concat kernel source not found or empty. Concat operations will not be available.\n");
+            backend_ctx->program_concat = nullptr;
+            backend_ctx->kernel_concat_f32_contiguous = nullptr;
+            backend_ctx->kernel_concat_f32_non_contiguous = nullptr;
+        }
+    }
+
+    // timestep_embedding
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "tsembd.cl.h"
+        };
+#else
+
+        const std::string kernel_src = read_file("tsembd.cl");
+#endif
+        if (!kernel_src.empty()) {
+            backend_ctx->program_tsembd =
+                build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+            CL_CHECK((backend_ctx->kernel_timestep_embedding = clCreateKernel(backend_ctx->program_tsembd, "kernel_timestep_embedding", &err), err));
+            GGML_LOG_CONT(".");
+        } else {
+            GGML_LOG_WARN("ggml_opencl: timestep_embedding kernel source not found or empty. This op will not be available.\n");
+            backend_ctx->program_tsembd = nullptr;
+            backend_ctx->kernel_timestep_embedding = nullptr;
+        }
+    }
+
+    // mul_mv_id_q4_0_f32_8x_flat
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "mul_mv_id_q4_0_f32_8x_flat.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("mul_mv_id_q4_0_f32_8x_flat.cl");
+#endif
+        backend_ctx->program_mul_mv_id_q4_0_f32_8x_flat =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_mul_mv_id_q4_0_f32_8x_flat = clCreateKernel(backend_ctx->program_mul_mv_id_q4_0_f32_8x_flat, "kernel_mul_mv_id_q4_0_f32_8x_flat", &err), err));
+        GGML_LOG_CONT(".");
+    }
+
     // Adreno kernels
 #ifdef GGML_OPENCL_USE_ADRENO_KERNELS
     // transpose
@@ -1492,6 +1842,12 @@ static ggml_backend_opencl_context * ggml_cl2_init(ggml_backend_dev_t dev) {
     backend_ctx->device     = dev_ctx->device;
     backend_ctx->gpu_family = GPU_FAMILY::UNKNOWN;
 
+    // ref_count get increased in ggml_backend_opencl_device_init
+    // This function is also used to retrieve backend context, so we don't want
+    // to increase ref_count for each call. We only want to increase ref_count
+    // when the associated device is initialized
+    backend_ctx->ref_count  = 0;
+
     if (strstr(dev_ctx->device_name.c_str(), "Adreno") ||
         strstr(dev_ctx->device_name.c_str(), "Qualcomm") ||
         strstr(dev_ctx->device_version.c_str(), "Adreno")) {
@@ -1664,93 +2020,22 @@ static ggml_backend_opencl_context * ggml_cl2_init(ggml_backend_dev_t dev) {
     return dev_ctx->backend_ctx;
 }
 
-static void ggml_cl2_free(void) {
-#ifdef GGML_OPENCL_PROFILING
-    FILE * fperf = fopen("cl_profiling.csv", "w");
-    if (!fperf) {
-        GGML_LOG_ERROR("Failed to open cl_profiling.csv\n");
-        return;
+static void ggml_cl2_free(ggml_backend_t backend) {
+    ggml_backend_opencl_context * ctx = (ggml_backend_opencl_context *) backend->context;
+    ctx->free();
+
+    // The CL context is shared by all backends, release it if all backends have been released
+    bool should_release_opencl = true;
+    for (auto device : g_ggml_backend_opencl_devices) {
+        ggml_backend_opencl_device_context * ctx_dev = (ggml_backend_opencl_device_context *) device.context;
+        if (ctx_dev->backend_ctx->ref_count > 0) {
+            should_release_opencl = false;
+        }
     }
 
-    // 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;
+    if (should_release_opencl) {
+        CL_CHECK(clReleaseContext(ctx->context));
     }
-
-    // Dump a csv
-    float total_kernel_time = 0;
-    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.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[1], info.local_size[2],
-            info.output_size[0], info.output_size[1], info.output_size[2], info.output_size[3]);
-    }
-    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
 }
 
 //------------------------------------------------------------------------------
@@ -1834,9 +2119,7 @@ static const char * ggml_backend_opencl_name(ggml_backend_t backend) {
 }
 
 static void ggml_backend_opencl_free(ggml_backend_t backend) {
-    ggml_cl2_free();
-
-    GGML_UNUSED(backend);
+    ggml_cl2_free(backend);
 }
 
 static void ggml_backend_opencl_set_tensor_async(ggml_backend_t backend, ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
@@ -1863,7 +2146,12 @@ static bool ggml_backend_opencl_cpy_tensor_async(ggml_backend_t backend, const g
 }
 
 static void ggml_backend_opencl_synchronize(ggml_backend_t backend) {
-    GGML_UNUSED(backend);
+    auto * backend_ctx = static_cast<ggml_backend_opencl_context *>(backend->context);
+
+    cl_event evt;
+    CL_CHECK(clEnqueueBarrierWithWaitList(backend_ctx->queue, 0, nullptr, &evt));
+    CL_CHECK(clWaitForEvents(1, &evt));
+    CL_CHECK(clReleaseEvent(evt));
 }
 
 // Syncronizes the 'backend_ctx's device with others so that commands
@@ -1906,7 +2194,7 @@ static ggml_status ggml_backend_opencl_graph_compute(ggml_backend_t backend, ggm
         //       dependencies.
         sync_with_other_backends(backend);
 
-        if (node->op == GGML_OP_RESHAPE || node->op == GGML_OP_TRANSPOSE || node->op == GGML_OP_VIEW || node->op == GGML_OP_PERMUTE || node->op == GGML_OP_NONE) {
+        if (ggml_is_empty(node) || node->op == GGML_OP_RESHAPE || node->op == GGML_OP_TRANSPOSE || node->op == GGML_OP_VIEW || node->op == GGML_OP_PERMUTE || node->op == GGML_OP_NONE) {
             continue;
         }
 
@@ -1941,6 +2229,12 @@ static bool ggml_opencl_supports_op(ggml_backend_dev_t dev, const struct ggml_te
                 default:
                     return false;
             }
+        case GGML_OP_SET_ROWS:
+            {
+                // TODO: add support
+                // ref: https://github.com/ggml-org/llama.cpp/pull/14274
+                return false;
+            } break;
         case GGML_OP_CPY:
         case GGML_OP_DUP:
         case GGML_OP_CONT:
@@ -1975,10 +2269,25 @@ static bool ggml_opencl_supports_op(ggml_backend_dev_t dev, const struct ggml_te
                 case GGML_UNARY_OP_GELU:
                 case GGML_UNARY_OP_SILU:
                 case GGML_UNARY_OP_RELU:
+                case GGML_UNARY_OP_GELU_ERF:
                 case GGML_UNARY_OP_GELU_QUICK:
-                    return ggml_is_contiguous(op->src[0]) && op->src[0]->type == GGML_TYPE_F32;
+                   return ggml_is_contiguous(op->src[0]) && op->src[0]->type == GGML_TYPE_F32;
                 case GGML_UNARY_OP_SIGMOID:
                     return ggml_is_contiguous(op->src[0]);
+                case GGML_UNARY_OP_TANH:
+                   return (op->src[0]->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32) ||
+                          (op->src[0]->type == GGML_TYPE_F16 && op->type == GGML_TYPE_F16);
+                default:
+                    return false;
+            }
+        case GGML_OP_GLU:
+            switch (ggml_get_glu_op(op)) {
+                case GGML_GLU_OP_GEGLU:
+                case GGML_GLU_OP_REGLU:
+                case GGML_GLU_OP_SWIGLU:
+                case GGML_GLU_OP_GEGLU_ERF:
+                case GGML_GLU_OP_GEGLU_QUICK:
+                    return ggml_is_contiguous_1(op->src[0]) && (op->type == GGML_TYPE_F32 || op->type == GGML_TYPE_F16);
                 default:
                     return false;
             }
@@ -1988,6 +2297,17 @@ static bool ggml_opencl_supports_op(ggml_backend_dev_t dev, const struct ggml_te
         case GGML_OP_NORM:
         case GGML_OP_RMS_NORM:
             return true;
+        case GGML_OP_REPEAT:
+            return op->src[0]->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32; // Assuming F32 for now, can be expanded
+        case GGML_OP_PAD:
+            return op->src[0]->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32 &&
+                   op->src[0]->ne[3] == 1 && op->ne[3] == 1;
+        case GGML_OP_UPSCALE:
+            return op->src[0]->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32;
+        case GGML_OP_CONCAT:
+            return op->src[0]->type == GGML_TYPE_F32 && op->src[1]->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32;
+        case GGML_OP_TIMESTEP_EMBEDDING:
+            return op->src[0]->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32;
         case GGML_OP_GROUP_NORM:
             return ggml_is_contiguous(op->src[0]);
         case GGML_OP_MUL_MAT:
@@ -2000,6 +2320,13 @@ static bool ggml_opencl_supports_op(ggml_backend_dev_t dev, const struct ggml_te
                 return op->src[1]->type == GGML_TYPE_F32 && ggml_is_contiguous(op->src[0]) && ggml_is_contiguous(op->src[1]);
             }
             return false;
+        case GGML_OP_MUL_MAT_ID:
+            if (op->src[0]->type == GGML_TYPE_Q4_0) {
+                if (op->src[1]->type == GGML_TYPE_F32) {
+                    return ggml_is_contiguous(op->src[0]) && ggml_is_contiguous(op->src[1]);
+                }
+            }
+            return false;
         case GGML_OP_RESHAPE:
         case GGML_OP_VIEW:
         case GGML_OP_PERMUTE:
@@ -2052,7 +2379,7 @@ static ggml_backend_i ggml_backend_opencl_i = {
     /* .set_tensor_async        = */ NULL,  /* ggml_backend_opencl_set_tensor_async */
     /* .get_tensor_async        = */ NULL,  /* ggml_backend_opencl_get_tensor_async */
     /* .cpy_tensor_async        = */ NULL,  /* ggml_backend_opencl_cpy_tensor_async */
-    /* .synchronize             = */ NULL,  /* ggml_backend_opencl_synchronize */
+    /* .synchronize             = */ ggml_backend_opencl_synchronize,
     /* .graph_plan_create       = */ NULL,
     /* .graph_plan_free         = */ NULL,
     /* .graph_plan_update       = */ NULL,
@@ -2696,6 +3023,8 @@ static void ggml_backend_opencl_device_get_props(ggml_backend_dev_t dev, struct
 
 static ggml_backend_t ggml_backend_opencl_device_init(ggml_backend_dev_t dev, const char * params) {
     ggml_backend_opencl_context * backend_ctx = ggml_cl2_init(dev);
+    // Getting a new reference to the backend, increase ref_count
+    backend_ctx->ref_count++;
 
     ggml_backend_t backend = new ggml_backend {
         /* .guid      = */ ggml_backend_opencl_guid(),
@@ -2886,7 +3215,7 @@ static void dump_tensor(ggml_backend_t backend, const struct ggml_tensor * tenso
 
     // Open file and dump.
     char fname[512];
-    sprintf(fname, "./tensor-dumps/%s.txt", tensor->name);
+    snprintf(fname, sizeof(fname), "./tensor-dumps/%s.txt", tensor->name);
     FILE * f = fopen(fname, "w");
     if (!f) {
         printf("Failed to open %s\n", fname);
@@ -2956,31 +3285,6 @@ static void dump_tensor(ggml_backend_t backend, const struct ggml_tensor * tenso
 #define dump_tensor(tensor)
 #endif
 
-//------------------------------------------------------------------------------
-// Profiling utility
-//------------------------------------------------------------------------------
-#ifdef GGML_OPENCL_PROFILING
-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) {
-    info.op_name     = tensor->name;
-    info.kernel      = kernel;
-    info.evt         = evt;
-
-    info.local_size[0]  = local_size[0];
-    info.local_size[1]  = local_size[1];
-    info.local_size[2]  = local_size[2];
-    info.global_size[0] = global_size[0];
-    info.global_size[1] = global_size[1];
-    info.global_size[2] = global_size[2];
-    info.output_size[0] = tensor->ne[0];
-    info.output_size[1] = tensor->ne[1];
-    info.output_size[2] = tensor->ne[2];
-    info.output_size[3] = tensor->ne[3];
-}
-#endif
-
 //------------------------------------------------------------------------------
 // Ops
 //------------------------------------------------------------------------------
@@ -3024,7 +3328,6 @@ static void ggml_cl_get_rows(ggml_backend_t backend, const ggml_tensor * src0, c
     const cl_ulong nb2  = dst  ?  dst->nb[2] : 0;
 
     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 * extra1 = (ggml_tensor_extra_cl *)src1->extra;
@@ -3068,15 +3371,7 @@ static void ggml_cl_get_rows(ggml_backend_t backend, const ggml_tensor * src0, c
     size_t global_work_size[] = {(size_t)ne10, (size_t)ne11, 1};
     size_t local_work_size[] = {1, 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
+    backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
 }
 
 static void ggml_cl_add(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
@@ -3118,7 +3413,6 @@ static void ggml_cl_add(ggml_backend_t backend, const ggml_tensor * src0, const
     const cl_ulong nb3  = dst ? dst->nb[3] : 0;
 
     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 * extra1 = (ggml_tensor_extra_cl *)src1->extra;
@@ -3193,29 +3487,13 @@ static void ggml_cl_add(ggml_backend_t backend, const ggml_tensor * src0, const
             local_work_size_ptr = nullptr;  // Let driver choose the work-group sizes.
         }
 
-#ifdef GGML_OPENCL_PROFILING
-        cl_event evt;
-        CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size_ptr, 0, NULL, &evt));
-
-        g_profiling_info.emplace_back();
-        populateProfilingInfo(g_profiling_info.back(), evt, kernel, global_work_size, local_work_size_ptr, dst);
-#else
-        CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size_ptr, 0, NULL, NULL));
-#endif
+        backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size_ptr, dst);
     } else {
         unsigned int nth = MIN(64, ne0);
         size_t global_work_size[] = {ne01*nth, (size_t)ne02, (size_t)ne03};
         size_t local_work_size[] = {nth, 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
+        backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
     }
 }
 
@@ -3258,7 +3536,6 @@ static void ggml_cl_mul(ggml_backend_t backend, const ggml_tensor * src0, const
     const cl_ulong nb3  = dst ? dst->nb[3] : 0;
 
     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 * extra1 = (ggml_tensor_extra_cl *)src1->extra;
@@ -3333,29 +3610,13 @@ static void ggml_cl_mul(ggml_backend_t backend, const ggml_tensor * src0, const
             local_work_size_ptr = nullptr;  // Let driver choose the work-group sizes.
         }
 
-#ifdef GGML_OPENCL_PROFILING
-        cl_event evt;
-        CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size_ptr, 0, NULL, &evt));
-
-        g_profiling_info.emplace_back();
-        populateProfilingInfo(g_profiling_info.back(), evt, kernel, global_work_size, local_work_size_ptr, dst);
-#else
-        CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size_ptr, 0, NULL, NULL));
-#endif
+        backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size_ptr, dst);
     } else {
         unsigned int nth = MIN(64, ne0);
         size_t global_work_size[] = {ne01*nth, (size_t)ne02, (size_t)ne03};
         size_t local_work_size[] = {nth, 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
+        backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
     }
 }
 
@@ -3395,7 +3656,6 @@ static void ggml_cl_div(ggml_backend_t backend, const ggml_tensor * src0, const
     const cl_ulong nb3  = dst->nb[3];
 
     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 * extra1 = (ggml_tensor_extra_cl *)src1->extra;
@@ -3458,29 +3718,13 @@ static void ggml_cl_div(ggml_backend_t backend, const ggml_tensor * src0, const
         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;
-        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
+        backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
     } else {
         unsigned int nth = MIN(64, ne0);
         size_t global_work_size[] = {ne01*nth, (size_t)ne02, (size_t)ne03};
         size_t local_work_size[] = {nth, 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
+        backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
     }
 }
 
@@ -3520,7 +3764,6 @@ static void ggml_cl_sub(ggml_backend_t backend, const ggml_tensor * src0, const
     const cl_ulong nb3  = dst->nb[3];
 
     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 * extra1 = (ggml_tensor_extra_cl *)src1->extra;
@@ -3583,29 +3826,13 @@ static void ggml_cl_sub(ggml_backend_t backend, const ggml_tensor * src0, const
         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;
-        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
+        backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
     } else {
         unsigned int nth = MIN(64, ne0);
         size_t global_work_size[] = {ne01*nth, (size_t)ne02, (size_t)ne03};
         size_t local_work_size[] = {nth, 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
+        backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
     }
 }
 
@@ -3618,7 +3845,6 @@ static void ggml_cl_gelu(ggml_backend_t backend, const ggml_tensor * src0, const
     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;
@@ -3645,15 +3871,45 @@ static void ggml_cl_gelu(ggml_backend_t backend, const ggml_tensor * src0, const
     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);
+    backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
+}
 
-    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_gelu_erf(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;
+
+    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_erf_4;
+        n /= 4;
+    } else {
+        kernel = backend_ctx->kernel_gelu_erf;
+    }
+
+    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};
+
+    backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
 }
 
 static void ggml_cl_gelu_quick(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
@@ -3665,7 +3921,6 @@ static void ggml_cl_gelu_quick(ggml_backend_t backend, const ggml_tensor * src0,
     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;
@@ -3692,15 +3947,7 @@ static void ggml_cl_gelu_quick(ggml_backend_t backend, const ggml_tensor * src0,
     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
+    backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
 }
 
 static void ggml_cl_silu(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
@@ -3712,7 +3959,6 @@ static void ggml_cl_silu(ggml_backend_t backend, const ggml_tensor * src0, const
     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;
@@ -3744,15 +3990,7 @@ static void ggml_cl_silu(ggml_backend_t backend, const ggml_tensor * src0, const
         local_work_size_ptr = nullptr;  // Let driver choose the work-group sizes.
     }
 
-#ifdef GGML_OPENCL_PROFILING
-    cl_event evt;
-    CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size_ptr, 0, NULL, &evt));
-
-    g_profiling_info.emplace_back();
-    populateProfilingInfo(g_profiling_info.back(), evt, kernel, global_work_size, local_work_size_ptr, dst);
-#else
-    CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size_ptr, 0, NULL, NULL));
-#endif
+    backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size_ptr, dst);
 }
 
 static void ggml_cl_relu(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
@@ -3764,7 +4002,6 @@ static void ggml_cl_relu(ggml_backend_t backend, const ggml_tensor * src0, const
     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;
@@ -3789,15 +4026,7 @@ static void ggml_cl_relu(ggml_backend_t backend, const ggml_tensor * src0, const
         local_work_size_ptr = nullptr;  // Let driver choose the work-group sizes.
     }
 
-#ifdef GGML_OPENCL_PROFILING
-    cl_event evt;
-    CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size_ptr, 0, NULL, &evt));
-
-    g_profiling_info.emplace_back();
-    populateProfilingInfo(g_profiling_info.back(), evt, kernel, global_work_size, local_work_size_ptr, dst);
-#else
-    CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size_ptr, 0, NULL, NULL));
-#endif
+    backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size_ptr, dst);
 }
 
 static void ggml_cl_sigmoid(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
@@ -3809,7 +4038,6 @@ static void ggml_cl_sigmoid(ggml_backend_t backend, const ggml_tensor * src0, co
     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;
@@ -3841,15 +4069,7 @@ static void ggml_cl_sigmoid(ggml_backend_t backend, const ggml_tensor * src0, co
         local_work_size_ptr = nullptr;  // Let driver choose the work-group sizes.
     }
 
-#ifdef GGML_OPENCL_PROFILING
-    cl_event evt;
-    CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size_ptr, 0, NULL, &evt));
-
-    g_profiling_info.emplace_back();
-    populateProfilingInfo(g_profiling_info.back(), evt, kernel, global_work_size, local_work_size_ptr, dst);
-#else
-    CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size_ptr, 0, NULL, NULL));
-#endif
+    backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size_ptr, dst);
 }
 
 static void ggml_cl_clamp(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
@@ -3861,7 +4081,6 @@ static void ggml_cl_clamp(ggml_backend_t backend, const ggml_tensor * src0, cons
     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;
@@ -3893,15 +4112,7 @@ static void ggml_cl_clamp(ggml_backend_t backend, const ggml_tensor * src0, cons
         local_work_size_ptr = nullptr;  // Let driver choose the work-group sizes.
     }
 
-#ifdef GGML_OPENCL_PROFILING
-    cl_event evt;
-    CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size_ptr, 0, NULL, &evt));
-
-    g_profiling_info.emplace_back();
-    populateProfilingInfo(g_profiling_info.back(), evt, kernel, global_work_size, local_work_size_ptr, dst);
-#else
-    CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size_ptr, 0, NULL, NULL));
-#endif
+    backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size_ptr, dst);
 }
 
 static void ggml_cl_norm(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
@@ -3913,7 +4124,6 @@ static void ggml_cl_norm(ggml_backend_t backend, const ggml_tensor * src0, const
     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;
@@ -3954,15 +4164,7 @@ static void ggml_cl_norm(ggml_backend_t backend, const ggml_tensor * src0, const
     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
-    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
+    backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
 }
 
 static void ggml_cl_rms_norm(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
@@ -3974,7 +4176,6 @@ static void ggml_cl_rms_norm(ggml_backend_t backend, const ggml_tensor * src0, c
     UNUSED(src1);
 
     ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
-    cl_command_queue queue = backend_ctx->queue;
 
     //ggml_backend_opencl_device_context * dev_ctx =
     //    (ggml_backend_opencl_device_context *)backend->device->context;
@@ -4038,15 +4239,7 @@ static void ggml_cl_rms_norm(ggml_backend_t backend, const ggml_tensor * src0, c
     // This is local memory - the size depends on subgroup size.
     CL_CHECK(clSetKernelArg(kernel, 12, sizeof(float)*nth/sgs,  NULL));
 
-#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
+    backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
 }
 
 static void ggml_cl_group_norm(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
@@ -4058,7 +4251,6 @@ static void ggml_cl_group_norm(ggml_backend_t backend, const ggml_tensor * src0,
     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;
@@ -4097,15 +4289,499 @@ static void ggml_cl_group_norm(ggml_backend_t backend, const ggml_tensor * src0,
     size_t global_work_size[] = {(size_t)n_groups*sgs, 1, 1};
     size_t local_work_size[] = {(size_t)sgs, 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));
+    backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
+}
 
-    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
+static void ggml_cl_tanh(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;
+
+    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_abs = extra0->offset + src0->view_offs;
+    cl_ulong offsetd_abs = extrad->offset + dst->view_offs;
+
+    cl_kernel kernel;
+    if (dst->type == GGML_TYPE_F32) {
+        kernel = backend_ctx->kernel_tanh_f32_nd;
+    } else if (dst->type == GGML_TYPE_F16) {
+        kernel = backend_ctx->kernel_tanh_f16_nd;
+    } else {
+        GGML_ASSERT(false && "Unsupported type for ggml_cl_tanh");
+    }
+    GGML_ASSERT(kernel != nullptr);
+
+    const int ne00 = src0->ne[0]; const int ne01 = src0->ne[1]; const int ne02 = src0->ne[2]; const int ne03 = src0->ne[3];
+    const cl_ulong nb00 = src0->nb[0]; const cl_ulong nb01 = src0->nb[1]; const cl_ulong nb02 = src0->nb[2]; const cl_ulong nb03 = src0->nb[3];
+
+    const int ne10 = dst->ne[0]; const int ne11 = dst->ne[1]; const int ne12 = dst->ne[2]; const int ne13 = dst->ne[3];
+    const cl_ulong nb10 = dst->nb[0]; const cl_ulong nb11 = dst->nb[1]; const cl_ulong nb12 = dst->nb[2]; const cl_ulong nb13 = dst->nb[3];
+
+    CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),   &extra0->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong), &offset0_abs));
+    CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem),   &extrad->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong), &offsetd_abs));
+
+    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), &nb00));
+    CL_CHECK(clSetKernelArg(kernel, 9, sizeof(cl_ulong), &nb01));
+    CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_ulong),&nb02));
+    CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_ulong),&nb03));
+
+    CL_CHECK(clSetKernelArg(kernel, 12, sizeof(int),     &ne10));
+    CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int),     &ne11));
+    CL_CHECK(clSetKernelArg(kernel, 14, sizeof(int),     &ne12));
+    CL_CHECK(clSetKernelArg(kernel, 15, sizeof(int),     &ne13));
+    CL_CHECK(clSetKernelArg(kernel, 16, sizeof(cl_ulong),&nb10));
+    CL_CHECK(clSetKernelArg(kernel, 17, sizeof(cl_ulong),&nb11));
+    CL_CHECK(clSetKernelArg(kernel, 18, sizeof(cl_ulong),&nb12));
+    CL_CHECK(clSetKernelArg(kernel, 19, sizeof(cl_ulong),&nb13));
+
+    size_t global_work_size[3];
+    if (ne10 == 0 || ne11 == 0 || ne12 == 0 || ne13 == 0) { // Handle case of 0 elements
+        return;
+    }
+    global_work_size[0] = (size_t)ne10;
+    global_work_size[1] = (size_t)ne11;
+    global_work_size[2] = (size_t)ne12;
+
+    size_t lws0 = 16, lws1 = 4, lws2 = 1;
+    if (ne10 < 16) lws0 = ne10;
+    if (ne11 < 4) lws1 = ne11;
+    if (ne12 < 1) lws2 = ne12 > 0 ? ne12 : 1;
+
+    while (lws0 * lws1 * lws2 > 256 && lws0 > 1) lws0 /= 2;
+    while (lws0 * lws1 * lws2 > 256 && lws1 > 1) lws1 /= 2;
+    while (lws0 * lws1 * lws2 > 256 && lws2 > 1) lws2 /= 2;
+
+
+    size_t local_work_size[] = {lws0, lws1, lws2};
+
+    size_t* local_work_size_ptr = local_work_size;
+    if (!backend_ctx->non_uniform_workgroups) {
+        if (global_work_size[0] % local_work_size[0] != 0 ||
+            global_work_size[1] % local_work_size[1] != 0 ||
+            global_work_size[2] % local_work_size[2] != 0) {
+            local_work_size_ptr = NULL;
+        }
+    }
+    if (global_work_size[0] == 0 || global_work_size[1] == 0 || global_work_size[2] == 0) return;
+
+    backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size_ptr, dst);
+}
+
+static void ggml_cl_repeat(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1_shape_def, ggml_tensor * dst) {
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src0->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+    GGML_ASSERT(dst->type == src0->type);
+
+    UNUSED(src1_shape_def);
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    if (backend_ctx->kernel_repeat == nullptr) {
+        GGML_LOG_WARN("%s: repeat kernel not available, skipping OpenCL execution.\n", __func__);
+        return;
+    }
+
+    ggml_tensor_extra_cl * extra_src0 = (ggml_tensor_extra_cl *)src0->extra;
+    ggml_tensor_extra_cl * extra_dst  = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong off_src0 = extra_src0->offset + src0->view_offs;
+    cl_ulong off_dst  = extra_dst->offset  + dst->view_offs;
+
+    const int src0_ne0 = src0->ne[0]; const int src0_ne1 = src0->ne[1]; const int src0_ne2 = src0->ne[2]; const int src0_ne3 = src0->ne[3];
+    const cl_ulong src0_nb0 = src0->nb[0]; const cl_ulong src0_nb1 = src0->nb[1]; const cl_ulong src0_nb2 = src0->nb[2]; const cl_ulong src0_nb3 = src0->nb[3];
+
+    const int dst_ne0 = dst->ne[0]; const int dst_ne1 = dst->ne[1]; const int dst_ne2 = dst->ne[2]; const int dst_ne3 = dst->ne[3];
+    const cl_ulong dst_nb0 = dst->nb[0]; const cl_ulong dst_nb1 = dst->nb[1]; const cl_ulong dst_nb2 = dst->nb[2]; const cl_ulong dst_nb3 = dst->nb[3];
+
+    cl_kernel kernel = backend_ctx->kernel_repeat;
+
+    CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),    &extra_src0->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem),    &extra_dst->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_ulong),  &off_src0));
+    CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong),  &off_dst));
+    CL_CHECK(clSetKernelArg(kernel, 4, sizeof(int),       &src0_ne0));
+    CL_CHECK(clSetKernelArg(kernel, 5, sizeof(int),       &src0_ne1));
+    CL_CHECK(clSetKernelArg(kernel, 6, sizeof(int),       &src0_ne2));
+    CL_CHECK(clSetKernelArg(kernel, 7, sizeof(int),       &src0_ne3));
+    CL_CHECK(clSetKernelArg(kernel, 8, sizeof(cl_ulong),  &src0_nb0));
+    CL_CHECK(clSetKernelArg(kernel, 9, sizeof(cl_ulong),  &src0_nb1));
+    CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_ulong), &src0_nb2));
+    CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_ulong), &src0_nb3));
+    CL_CHECK(clSetKernelArg(kernel, 12, sizeof(int),      &dst_ne0));
+    CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int),      &dst_ne1));
+    CL_CHECK(clSetKernelArg(kernel, 14, sizeof(int),      &dst_ne2));
+    CL_CHECK(clSetKernelArg(kernel, 15, sizeof(int),      &dst_ne3));
+    CL_CHECK(clSetKernelArg(kernel, 16, sizeof(cl_ulong), &dst_nb0));
+    CL_CHECK(clSetKernelArg(kernel, 17, sizeof(cl_ulong), &dst_nb1));
+    CL_CHECK(clSetKernelArg(kernel, 18, sizeof(cl_ulong), &dst_nb2));
+    CL_CHECK(clSetKernelArg(kernel, 19, sizeof(cl_ulong), &dst_nb3));
+
+    size_t gws0 = dst_ne1 > 0 ? (size_t)dst_ne1 : 1;
+    size_t gws1 = dst_ne2 > 0 ? (size_t)dst_ne2 : 1;
+    size_t gws2 = dst_ne3 > 0 ? (size_t)dst_ne3 : 1;
+
+    size_t global_work_size[] = { gws0, gws1, gws2 };
+
+    backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, NULL, dst);
+}
+
+static void ggml_cl_pad(ggml_backend_t backend, const ggml_tensor * src0, ggml_tensor * dst) {
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src0->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type == GGML_TYPE_F32);
+    GGML_ASSERT(src0->ne[3] == 1 && dst->ne[3] == 1);
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    if (backend_ctx->kernel_pad == nullptr) {
+        GGML_LOG_WARN("%s: pad kernel not available, skipping OpenCL execution.\n", __func__);
+        return;
+    }
+
+    ggml_tensor_extra_cl * extra_src0 = (ggml_tensor_extra_cl *)src0->extra;
+    ggml_tensor_extra_cl * extra_dst  = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong off_src0 = extra_src0->offset + src0->view_offs;
+    cl_ulong off_dst  = extra_dst->offset  + dst->view_offs;
+
+    const int s_ne0 = src0->ne[0];
+    const int s_ne1 = src0->ne[1];
+    const int s_ne2 = src0->ne[2];
+
+    const int d_ne0 = dst->ne[0];
+    const int d_ne1 = dst->ne[1];
+    const int d_ne2 = dst->ne[2];
+
+    cl_kernel kernel = backend_ctx->kernel_pad;
+
+    CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),    &extra_src0->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong),  &off_src0));
+    CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem),    &extra_dst->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong),  &off_dst));
+    CL_CHECK(clSetKernelArg(kernel, 4, sizeof(int),       &s_ne0));
+    CL_CHECK(clSetKernelArg(kernel, 5, sizeof(int),       &s_ne1));
+    CL_CHECK(clSetKernelArg(kernel, 6, sizeof(int),       &s_ne2));
+    CL_CHECK(clSetKernelArg(kernel, 7, sizeof(int),       &d_ne0));
+    CL_CHECK(clSetKernelArg(kernel, 8, sizeof(int),       &d_ne1));
+    CL_CHECK(clSetKernelArg(kernel, 9, sizeof(int),       &d_ne2));
+
+    size_t lws0 = 64;
+    size_t gws0 = (( (size_t)d_ne0 + lws0 - 1 ) / lws0) * lws0;
+
+    size_t global_work_size[] = { gws0, (size_t)d_ne1, (size_t)d_ne2 };
+    size_t local_work_size[]  = { lws0, 1, 1 };
+
+    size_t * local_work_size_ptr = local_work_size;
+     if (d_ne0 % lws0 != 0 && !backend_ctx->non_uniform_workgroups) {
+        local_work_size_ptr = nullptr;
+    }
+
+    backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size_ptr, dst);
+}
+
+static void ggml_cl_upscale(ggml_backend_t backend, const ggml_tensor * src0, ggml_tensor * dst) {
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src0->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type == GGML_TYPE_F32);
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    const int mode_flags        = (ggml_scale_mode) ggml_get_op_params_i32(dst, 0);
+    const ggml_scale_mode mode  = (ggml_scale_mode) (mode_flags & 0xFF);
+    cl_kernel kernel = nullptr;
+
+    if (mode == GGML_SCALE_MODE_NEAREST) {
+        kernel = backend_ctx->kernel_upscale;
+        if (kernel == nullptr) {
+            GGML_LOG_WARN("%s: nearest upscale kernel not available, skipping OpenCL execution.\n", __func__);
+            return;
+        }
+    } else if (mode == GGML_SCALE_MODE_BILINEAR) {
+        kernel = backend_ctx->kernel_upscale_bilinear;
+        if (kernel == nullptr) {
+            GGML_LOG_WARN("%s: bilinear upscale kernel not available, skipping OpenCL execution.\n", __func__);
+            return;
+        }
+    } else {
+        GGML_LOG_WARN("%s: unsupported upscale mode %d, skipping OpenCL execution.\n", __func__, mode);
+        return;
+    }
+
+    ggml_tensor_extra_cl * extra_src0 = (ggml_tensor_extra_cl *)src0->extra;
+    ggml_tensor_extra_cl * extra_dst  = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong off_src0 = extra_src0->offset + src0->view_offs;
+    cl_ulong off_dst  = extra_dst->offset  + dst->view_offs;
+
+    const cl_ulong nb00 = src0->nb[0];
+    const cl_ulong nb01 = src0->nb[1];
+    const cl_ulong nb02 = src0->nb[2];
+    const cl_ulong nb03 = src0->nb[3];
+
+    const int ne00 = src0->ne[0];
+    const int ne01 = src0->ne[1];
+    const int ne02 = src0->ne[2];
+    const int ne03 = src0->ne[3];
+
+    const int ne0 = dst->ne[0];
+    const int ne1 = dst->ne[1];
+    const int ne2 = dst->ne[2];
+    const int ne3 = dst->ne[3];
+
+    float sf0 = (float)ne0 / ne00;
+    float sf1 = (float)ne1 / ne01;
+    float sf2 = (float)ne2 / ne02;
+    float sf3 = (float)ne3 / ne03;
+
+    float pixel_offset = 0.5f;
+
+    CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),    &extra_src0->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong),  &off_src0));
+    CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem),    &extra_dst->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong),  &off_dst));
+    CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_ulong),  &nb00));
+    CL_CHECK(clSetKernelArg(kernel, 5, sizeof(cl_ulong),  &nb01));
+    CL_CHECK(clSetKernelArg(kernel, 6, sizeof(cl_ulong),  &nb02));
+    CL_CHECK(clSetKernelArg(kernel, 7, sizeof(cl_ulong),  &nb03));
+
+    if (mode == GGML_SCALE_MODE_NEAREST) {
+        CL_CHECK(clSetKernelArg(kernel, 8, sizeof(int),       &ne0));
+        CL_CHECK(clSetKernelArg(kernel, 9, sizeof(int),       &ne1));
+        CL_CHECK(clSetKernelArg(kernel, 10, sizeof(int),      &ne2));
+        CL_CHECK(clSetKernelArg(kernel, 11, sizeof(int),      &ne3));
+        CL_CHECK(clSetKernelArg(kernel, 12, sizeof(float),    &sf0));
+        CL_CHECK(clSetKernelArg(kernel, 13, sizeof(float),    &sf1));
+        CL_CHECK(clSetKernelArg(kernel, 14, sizeof(float),    &sf2));
+        CL_CHECK(clSetKernelArg(kernel, 15, sizeof(float),    &sf3));
+    } else if (mode == GGML_SCALE_MODE_BILINEAR) {
+        if (mode_flags & GGML_SCALE_FLAG_ALIGN_CORNERS) {
+            sf0 = (float)(ne0 - 1) / (ne00 - 1);
+            sf1 = (float)(ne1 - 1) / (ne01 - 1);
+            pixel_offset = 0.0f;
+        }
+
+        CL_CHECK(clSetKernelArg(kernel, 8, sizeof(int),       &ne00));
+        CL_CHECK(clSetKernelArg(kernel, 9, sizeof(int),       &ne01));
+        CL_CHECK(clSetKernelArg(kernel, 10, sizeof(int),      &ne0));
+        CL_CHECK(clSetKernelArg(kernel, 11, sizeof(int),      &ne1));
+        CL_CHECK(clSetKernelArg(kernel, 12, sizeof(int),      &ne2));
+        CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int),      &ne3));
+        CL_CHECK(clSetKernelArg(kernel, 14, sizeof(float),    &sf0));
+        CL_CHECK(clSetKernelArg(kernel, 15, sizeof(float),    &sf1));
+        CL_CHECK(clSetKernelArg(kernel, 16, sizeof(float),    &sf2));
+        CL_CHECK(clSetKernelArg(kernel, 17, sizeof(float),    &sf3));
+        CL_CHECK(clSetKernelArg(kernel, 18, sizeof(float),    &pixel_offset));
+    }
+
+
+    size_t dst_total_elements = (size_t)ne0 * ne1 * ne2 * ne3;
+    if (dst_total_elements == 0) {
+        return;
+    }
+    size_t global_work_size[] = { dst_total_elements, 1, 1 };
+    size_t local_work_size_pref = 256;
+    size_t local_work_size[] = { MIN(local_work_size_pref, dst_total_elements), 1, 1};
+
+    size_t * local_work_size_ptr = local_work_size;
+    if (dst_total_elements % local_work_size[0] != 0 && !backend_ctx->non_uniform_workgroups) {
+        local_work_size_ptr = nullptr;
+    }
+
+    backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size_ptr, dst);
+}
+
+static void ggml_cl_concat(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src0->extra);
+    GGML_ASSERT(src1);
+    GGML_ASSERT(src1->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT(src1->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type == GGML_TYPE_F32);
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+    cl_command_queue queue = backend_ctx->queue;
+
+    if (backend_ctx->kernel_concat_f32_contiguous == nullptr || backend_ctx->kernel_concat_f32_non_contiguous == nullptr) {
+        GGML_LOG_WARN("%s: concat kernels not available, skipping OpenCL execution.\n", __func__);
+        return;
+    }
+
+    ggml_tensor_extra_cl * extra0_cl = (ggml_tensor_extra_cl *)src0->extra;
+    ggml_tensor_extra_cl * extra1_cl = (ggml_tensor_extra_cl *)src1->extra;
+    ggml_tensor_extra_cl * extrad_cl = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong off_src0 = extra0_cl->offset + src0->view_offs;
+    cl_ulong off_src1 = extra1_cl->offset + src1->view_offs;
+    cl_ulong off_dst  = extrad_cl->offset + dst->view_offs;
+
+    const int32_t dim = ((const int32_t *) dst->op_params)[0];
+    GGML_ASSERT(dim >= 0 && dim <= 3);
+
+    if (ggml_is_contiguous(src0) && ggml_is_contiguous(src1) && ggml_is_contiguous(dst)) {
+        if (dim == 3) {
+
+            size_t nbytes_src0 = ggml_nbytes(src0);
+            size_t nbytes_src1 = ggml_nbytes(src1);
+
+            CL_CHECK(clEnqueueCopyBuffer(queue, extra0_cl->data_device, extrad_cl->data_device,
+                                         off_src0, off_dst, nbytes_src0, 0, NULL, NULL));
+            CL_CHECK(clEnqueueCopyBuffer(queue, extra1_cl->data_device, extrad_cl->data_device,
+                                         off_src1, off_dst + nbytes_src0, nbytes_src1, 0, NULL, NULL));
+        } else {
+
+            cl_kernel kernel = backend_ctx->kernel_concat_f32_contiguous;
+            size_t global_work_size[3];
+
+            for (int i3 = 0; i3 < dst->ne[3]; ++i3) {
+                cl_ulong current_off_src0 = off_src0 + (i3 * src0->nb[3]);
+                cl_ulong current_off_src1 = off_src1 + (i3 * src1->nb[3]);
+                cl_ulong current_off_dst  = off_dst  + (i3 * dst->nb[3]);
+
+                int d_ne00 = src0->ne[0]; int d_ne01 = src0->ne[1]; int d_ne02 = src0->ne[2];
+                int d_ne10 = src1->ne[0]; int d_ne11 = src1->ne[1]; int d_ne12 = src1->ne[2];
+                int d_ne0  = dst->ne[0];  int d_ne1  = dst->ne[1];  int d_ne2  = dst->ne[2];
+
+                CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),    &extra0_cl->data_device));
+                CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong),  &current_off_src0));
+                CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem),    &extra1_cl->data_device));
+                CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong),  &current_off_src1));
+                CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_mem),    &extrad_cl->data_device));
+                CL_CHECK(clSetKernelArg(kernel, 5, sizeof(cl_ulong),  &current_off_dst));
+                CL_CHECK(clSetKernelArg(kernel, 6, sizeof(int),       &d_ne00));
+                CL_CHECK(clSetKernelArg(kernel, 7, sizeof(int),       &d_ne01));
+                CL_CHECK(clSetKernelArg(kernel, 8, sizeof(int),       &d_ne02));
+                CL_CHECK(clSetKernelArg(kernel, 9, sizeof(int),       &d_ne10));
+                CL_CHECK(clSetKernelArg(kernel, 10, sizeof(int),      &d_ne11));
+                CL_CHECK(clSetKernelArg(kernel, 11, sizeof(int),      &d_ne12));
+                CL_CHECK(clSetKernelArg(kernel, 12, sizeof(int),      &d_ne0));
+                CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int),      &d_ne1));
+                CL_CHECK(clSetKernelArg(kernel, 14, sizeof(int),      &d_ne2));
+                CL_CHECK(clSetKernelArg(kernel, 15, sizeof(int),      &dim));
+
+                global_work_size[0] = d_ne0;
+                global_work_size[1] = d_ne1;
+                global_work_size[2] = d_ne2;
+
+                backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, NULL, dst);
+            }
+        }
+    } else {
+        cl_kernel kernel = backend_ctx->kernel_concat_f32_non_contiguous;
+
+        long ne00 = src0->ne[0], ne01 = src0->ne[1], ne02 = src0->ne[2], ne03 = src0->ne[3];
+        cl_ulong nb00 = src0->nb[0], nb01 = src0->nb[1], nb02 = src0->nb[2], nb03 = src0->nb[3];
+
+        cl_ulong nb10 = src1->nb[0], nb11 = src1->nb[1], nb12 = src1->nb[2], nb13 = src1->nb[3];
+
+        long d_ne0 = dst->ne[0], d_ne1 = dst->ne[1], d_ne2 = dst->ne[2], d_ne3 = dst->ne[3];
+        cl_ulong d_nb0 = dst->nb[0], d_nb1 = dst->nb[1], d_nb2 = dst->nb[2], d_nb3 = dst->nb[3];
+
+
+        CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),    &extra0_cl->data_device));
+        CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong),  &off_src0));
+        CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem),    &extra1_cl->data_device));
+        CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong),  &off_src1));
+        CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_mem),    &extrad_cl->data_device));
+        CL_CHECK(clSetKernelArg(kernel, 5, sizeof(cl_ulong),  &off_dst));
+
+        CL_CHECK(clSetKernelArg(kernel, 6, sizeof(long),      &ne00));
+        CL_CHECK(clSetKernelArg(kernel, 7, sizeof(long),      &ne01));
+        CL_CHECK(clSetKernelArg(kernel, 8, sizeof(long),      &ne02));
+        CL_CHECK(clSetKernelArg(kernel, 9, sizeof(long),      &ne03));
+        CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_ulong),    &nb00));
+        CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_ulong),    &nb01));
+        CL_CHECK(clSetKernelArg(kernel, 12, sizeof(cl_ulong),    &nb02));
+        CL_CHECK(clSetKernelArg(kernel, 13, sizeof(cl_ulong),    &nb03));
+
+        CL_CHECK(clSetKernelArg(kernel, 14, sizeof(cl_ulong),    &nb10));
+        CL_CHECK(clSetKernelArg(kernel, 15, sizeof(cl_ulong),    &nb11));
+        CL_CHECK(clSetKernelArg(kernel, 16, sizeof(cl_ulong),    &nb12));
+        CL_CHECK(clSetKernelArg(kernel, 17, sizeof(cl_ulong),    &nb13));
+
+        CL_CHECK(clSetKernelArg(kernel, 18, sizeof(long),     &d_ne0));
+        CL_CHECK(clSetKernelArg(kernel, 19, sizeof(long),     &d_ne1));
+        CL_CHECK(clSetKernelArg(kernel, 20, sizeof(long),     &d_ne2));
+        CL_CHECK(clSetKernelArg(kernel, 21, sizeof(long),     &d_ne3));
+        CL_CHECK(clSetKernelArg(kernel, 22, sizeof(cl_ulong),    &d_nb0));
+        CL_CHECK(clSetKernelArg(kernel, 23, sizeof(cl_ulong),    &d_nb1));
+        CL_CHECK(clSetKernelArg(kernel, 24, sizeof(cl_ulong),    &d_nb2));
+        CL_CHECK(clSetKernelArg(kernel, 25, sizeof(cl_ulong),    &d_nb3));
+        CL_CHECK(clSetKernelArg(kernel, 26, sizeof(int),      &dim));
+
+        size_t global_work_size_nc[] = { d_ne1 > 0 ? (size_t)d_ne1 : 1,
+                                         d_ne2 > 0 ? (size_t)d_ne2 : 1,
+                                         d_ne3 > 0 ? (size_t)d_ne3 : 1 };
+
+        backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size_nc, NULL, dst);
+    }
+}
+
+static void ggml_cl_timestep_embedding(ggml_backend_t backend, const ggml_tensor * src0, ggml_tensor * dst) {
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src0->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type == GGML_TYPE_F32);
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    if (backend_ctx->kernel_timestep_embedding == nullptr) {
+        GGML_LOG_WARN("%s: timestep_embedding kernel not available, skipping OpenCL execution.\n", __func__);
+        return;
+    }
+
+    ggml_tensor_extra_cl * extra_src0 = (ggml_tensor_extra_cl *)src0->extra;
+    ggml_tensor_extra_cl * extra_dst  = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong off_src0 = extra_src0->offset + src0->view_offs;
+    cl_ulong off_dst  = extra_dst->offset  + dst->view_offs;
+
+    const int logical_dim = dst->op_params[0];
+    const int max_period  = dst->op_params[1];
+    const int dst_nb1_bytes = dst->nb[1];
+
+    cl_kernel kernel = backend_ctx->kernel_timestep_embedding;
+
+    CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),    &extra_src0->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong),  &off_src0));
+    CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem),    &extra_dst->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong),  &off_dst));
+    CL_CHECK(clSetKernelArg(kernel, 4, sizeof(int),       &dst_nb1_bytes));
+    CL_CHECK(clSetKernelArg(kernel, 5, sizeof(int),       &logical_dim));
+    CL_CHECK(clSetKernelArg(kernel, 6, sizeof(int),       &max_period));
+
+    size_t gws0 = (size_t)(((logical_dim + 1) / 2) + 1);
+
+    size_t gws1 = (size_t)src0->ne[0];
+
+    size_t global_work_size[] = {gws0, gws1, 1};
+
+    backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, NULL, dst);
 }
 
 static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
@@ -4120,7 +4796,6 @@ static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, co
     const enum ggml_type src1t = src1 ? src1->type : GGML_TYPE_COUNT;
 
     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 * extra1 = (ggml_tensor_extra_cl *)src1->extra;
@@ -4325,15 +5000,7 @@ static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, co
                 static_cast<size_t>(padded_height_B)
             };
 
-            #ifdef GGML_OPENCL_PROFILING
-                cl_event evt;
-                CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 2, NULL, global_size_t, local_size_t, 0, NULL, &evt));
-
-                g_profiling_info.emplace_back();
-                populateProfilingInfo(g_profiling_info.back(), evt, kernel, global_size_t, local_size_t, dst);
-            #else
-                CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 2, NULL, global_size_t, local_size_t, 0, NULL, NULL));
-            #endif
+            backend_ctx->enqueue_ndrange_kernel(kernel, 2, global_size_t, local_size_t, dst);
         } else {
             // no need to transpose B in other cases
             // create an image for B from sub_buffer
@@ -4455,16 +5122,7 @@ 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();
-        populateProfilingInfo(g_profiling_info.back(), evt, kernel, global_work_size, local_work_size, dst);
-        // enqueue kernel without profiling
-    #else
-        CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, NULL));
-    #endif
+        backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
         // <--------------------------------------------> //
 
         // deallocate sub buffers and images
@@ -4544,15 +5202,7 @@ static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, co
                 global_work_size[2] = (size_t)ne12*ne13;
             }
 
-#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
+            backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
             return;
         }
 #else // GGML_OPENCL_SOA_Q
@@ -4782,15 +5432,7 @@ static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, co
         size_t global_work_size[] = {(size_t)(ne01 + ndst-1)/ndst*nth0, (size_t)ne11*nth1, (size_t)ne12*ne13};
         size_t local_work_size[] = {(size_t)nth0, (size_t)nth1, 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
+        backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
     } else if (src0t == GGML_TYPE_Q4_K) {
         GGML_ASSERT(false && "not implemented");
     } else if (src0t == GGML_TYPE_Q3_K) {
@@ -4801,33 +5443,138 @@ static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, co
         size_t global_work_size[] = {(size_t)(ne01+1)/2*nth0, (size_t)ne11*nth1, (size_t)ne12*ne13};
         size_t local_work_size[] = {(size_t)nth0, (size_t)nth1, 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
+        backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
     } else {
         int64_t ny = (ne11 + nrows - 1)/nrows;
 
         size_t global_work_size[] = {(size_t)ne01*nth0, (size_t)ny*nth1, (size_t)ne12*ne13};
         size_t local_work_size[] = {(size_t)nth0, (size_t)nth1, 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
+        backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
     }
 }
 
+static void ggml_cl_mul_mat_id(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src0->extra);
+    GGML_ASSERT(src1);
+    GGML_ASSERT(src1->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+
+    const ggml_tensor * src2 = dst->src[2];
+    GGML_ASSERT(src2);
+    GGML_ASSERT(src2->extra);
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    ggml_tensor_extra_cl * extra1 = (ggml_tensor_extra_cl *)src1->extra;
+    ggml_tensor_extra_cl * extra2 = (ggml_tensor_extra_cl *)src2->extra;
+    ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong offset1 = extra1->offset + src1->view_offs;
+    cl_ulong offset2 = extra2->offset + src2->view_offs;
+    cl_ulong offsetd = extrad->offset + dst->view_offs;
+
+#ifdef GGML_OPENCL_SOA_Q
+    ggml_tensor_extra_cl_q4_0 * extra0_q4_0 = (ggml_tensor_extra_cl_q4_0 *)src0->extra;
+#endif
+
+    const int ne00 = src0->ne[0];
+    const int ne01 = src0->ne[1];
+    const int ne02 = src0->ne[2];
+    const int ne03 = src0->ne[3];
+
+    const cl_ulong nb00 = src0->nb[0];
+    const cl_ulong nb02 = src0->nb[2];
+
+    const int ne10 = src1->ne[0];
+    const int ne11 = src1->ne[1];
+    const int ne12 = src1->ne[2];
+    const int ne13 = src1->ne[3];
+
+    const cl_ulong nb11 = src1->nb[1];
+    const cl_ulong nb12 = src1->nb[2];
+
+    const int ne20 = src2->ne[0];
+    const int ne21 = src2->ne[1];
+
+    const cl_ulong nb21 = src2->nb[1];
+
+    const int ne0 = dst->ne[0];
+    const int ne1 = dst->ne[1];
+
+    const int r2 = ne12/ne02;
+    const int r3 = ne13/ne03;
+    const int dst_rows = ne20*ne21; // ne20 = n_used_experts, ne21 = n_rows
+
+    GGML_ASSERT(ne00 == ne10);
+
+    int sgs   = 32; // subgroup size
+    int nsg   = 1;  // number of subgroups
+    int nrows = 1;  // number of row in src1
+    int ndst  = 4;  // number of values produced by each subgroup
+
+    cl_kernel kernel;
+
+    // subgroup mat vec
+    switch (src0->type) {
+        case GGML_TYPE_Q4_0: {
+            kernel = backend_ctx->kernel_mul_mv_id_q4_0_f32_8x_flat;
+
+            if (backend_ctx->gpu_family == INTEL) {
+                sgs  = 16;
+                nsg  = 1;
+                ndst = 8;
+            } else if (backend_ctx->gpu_family == ADRENO) {
+                sgs  = 64;
+                nsg  = 1;
+                ndst = 8;
+            } else {
+                GGML_ASSERT(false && "TODO: Unknown GPU");
+            }
+
+            CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0_q4_0->q));
+            CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_mem),   &extra0_q4_0->d));
+            CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &extra1->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &offset1));
+            CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),   &extra2->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &offset2));
+            CL_CHECK(clSetKernelArg(kernel,  6, sizeof(cl_mem),   &extrad->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  7, sizeof(cl_ulong), &offsetd));
+            CL_CHECK(clSetKernelArg(kernel,  8, sizeof(int),      &ne00));
+            CL_CHECK(clSetKernelArg(kernel,  9, sizeof(int),      &ne01));
+            CL_CHECK(clSetKernelArg(kernel, 10, sizeof(int),      &ne02));
+            CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_ulong), &nb00));
+            CL_CHECK(clSetKernelArg(kernel, 12, sizeof(cl_ulong), &nb02));
+            CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int),      &ne10));
+            CL_CHECK(clSetKernelArg(kernel, 14, sizeof(int),      &ne11));
+            CL_CHECK(clSetKernelArg(kernel, 15, sizeof(int),      &ne12));
+            CL_CHECK(clSetKernelArg(kernel, 16, sizeof(cl_ulong), &nb11));
+            CL_CHECK(clSetKernelArg(kernel, 17, sizeof(cl_ulong), &nb12));
+            CL_CHECK(clSetKernelArg(kernel, 18, sizeof(int),      &ne20));
+            CL_CHECK(clSetKernelArg(kernel, 19, sizeof(int),      &ne21));
+            CL_CHECK(clSetKernelArg(kernel, 20, sizeof(cl_ulong), &nb21));
+            CL_CHECK(clSetKernelArg(kernel, 21, sizeof(int),      &ne0));
+            CL_CHECK(clSetKernelArg(kernel, 22, sizeof(int),      &ne1));
+            CL_CHECK(clSetKernelArg(kernel, 23, sizeof(int),      &r2));
+            CL_CHECK(clSetKernelArg(kernel, 24, sizeof(int),      &r3));
+
+            break;
+        }
+        default:
+            GGML_ASSERT(false && "not implemented");;
+    }
+
+    int _ne1 = 1;
+    int ne123 = dst_rows;
+
+    size_t global_work_size[] = {(size_t)(ne01+ndst*nsg-1)/(ndst*nsg)*sgs, (size_t)(_ne1+nrows-1)/nrows*nsg, (size_t)ne123};
+    size_t local_work_size[] = {(size_t)sgs, (size_t)nsg, 1};
+
+    backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
+}
+
 static void ggml_cl_scale(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     GGML_ASSERT(src0);
     GGML_ASSERT(src0->extra);
@@ -4838,7 +5585,6 @@ static void ggml_cl_scale(ggml_backend_t backend, const ggml_tensor * src0, cons
     GGML_ASSERT(ggml_is_contiguous(src0));
 
     ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
-    cl_command_queue queue = backend_ctx->queue;
 
     float scale;
     memcpy(&scale, dst->op_params, sizeof(scale));
@@ -4867,15 +5613,7 @@ static void ggml_cl_scale(ggml_backend_t backend, const ggml_tensor * src0, cons
         local_work_size_ptr = nullptr;  // Let driver choose the work-group sizes.
     }
 
-#ifdef GGML_OPENCL_PROFILING
-    cl_event evt;
-    CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size_ptr, 0, NULL, &evt));
-
-    g_profiling_info.emplace_back();
-    populateProfilingInfo(g_profiling_info.back(), evt, kernel, global_work_size, local_work_size_ptr, dst);
-#else
-    CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size_ptr, 0, NULL, NULL));
-#endif
+    backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size_ptr, dst);
 }
 
 static void ggml_cl_cpy(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
@@ -4912,7 +5650,6 @@ static void ggml_cl_cpy(ggml_backend_t backend, const ggml_tensor * src0, const
     const enum ggml_type src1t = src1 ? src1->type : GGML_TYPE_COUNT;
 
     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 * extra1 = (ggml_tensor_extra_cl *)src1->extra;
@@ -4977,15 +5714,7 @@ static void ggml_cl_cpy(ggml_backend_t backend, const ggml_tensor * src0, const
     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
-    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, src1);
-#else
-    CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, NULL));
-#endif
+    backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, src1);
 }
 
 static void ggml_cl_dup(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
@@ -5008,7 +5737,6 @@ static void ggml_cl_diag_mask_inf(ggml_backend_t backend, const ggml_tensor * sr
     const int  ne02 = src0 ? src0->ne[2] : 0;
 
     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;
@@ -5032,15 +5760,7 @@ static void ggml_cl_diag_mask_inf(ggml_backend_t backend, const ggml_tensor * sr
         size_t global_work_size[] = {(size_t)ne00*ne01*ne02/8, 1, 1};
         size_t local_work_size[] = {64, 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
+        backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
     } else {
         kernel = backend_ctx->kernel_diag_mask_inf;
 
@@ -5060,15 +5780,7 @@ static void ggml_cl_diag_mask_inf(ggml_backend_t backend, const ggml_tensor * sr
             local_work_size_ptr = nullptr;  // Let driver choose the work-group sizes.
         }
 
-#ifdef GGML_OPENCL_PROFILING
-        cl_event evt;
-        CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size_ptr, 0, NULL, &evt));
-
-        g_profiling_info.emplace_back();
-        populateProfilingInfo(g_profiling_info.back(), evt, kernel, global_work_size, local_work_size_ptr, dst);
-#else
-        CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size_ptr, 0, NULL, NULL));
-#endif
+        backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size_ptr, dst);
     }
 }
 
@@ -5088,7 +5800,6 @@ static void ggml_cl_soft_max(ggml_backend_t backend, const ggml_tensor * src0, c
     }
 
     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;
@@ -5100,19 +5811,31 @@ static void ggml_cl_soft_max(ggml_backend_t backend, const ggml_tensor * src0, c
 
     cl_ulong offset1 = extra1 ? extra1->offset + src1->view_offs : offset0;
 
-    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 int ne00 = src0->ne[0];
+    const int ne01 = src0->ne[1];
+    const int ne02 = src0->ne[2];
+    const int ne03 = src0->ne[3];
+
+    const cl_long nb01 = src0->nb[1];
+    const cl_long nb02 = src0->nb[2];
+    const cl_long nb03 = src0->nb[3];
+
+    const int ne12 = src1 ? src1->ne[2] : 0;
+    const int ne13 = src1 ? src1->ne[3] : 0;
+
+    const cl_long nb11 = src1 ? src1->nb[1] : 0;
+    const cl_long nb12 = src1 ? src1->nb[2] : 0;
+    const cl_long nb13 = src1 ? src1->nb[3] : 0;
+
+    const cl_long nb1 = dst->nb[1];
+    const cl_long nb2 = dst->nb[2];
+    const cl_long nb3 = dst->nb[3];
 
     float scale, max_bias;
     memcpy(&scale,    dst->op_params + 0, sizeof(float));
     memcpy(&max_bias, dst->op_params + 1, sizeof(float));
 
-    const int nrows_x = ggml_nrows(src0);
-    const int nrows_y = src0->ne[1];
-
-    const int n_head      = nrows_x/nrows_y;
+    const int n_head      = src0->ne[2];
     const int n_head_log2 = 1u << (uint32_t) floorf(log2f((float) n_head));
 
     const float m0 = powf(2.0f, -(max_bias       ) / n_head_log2);
@@ -5157,26 +5880,27 @@ static void ggml_cl_soft_max(ggml_backend_t backend, const ggml_tensor * src0, c
     CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),   &extrad->data_device));
     CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &offsetd));
     CL_CHECK(clSetKernelArg(kernel,  6, sizeof(int),      &ne00));
-    CL_CHECK(clSetKernelArg(kernel,  7, sizeof(int),      &ne01));
-    CL_CHECK(clSetKernelArg(kernel,  8, sizeof(int),      &ne02));
-    CL_CHECK(clSetKernelArg(kernel,  9, sizeof(float),    &scale));
-    CL_CHECK(clSetKernelArg(kernel, 10, sizeof(float),    &max_bias));
-    CL_CHECK(clSetKernelArg(kernel, 11, sizeof(float),    &m0));
-    CL_CHECK(clSetKernelArg(kernel, 12, sizeof(float),    &m1));
-    CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int),      &n_head_log2));
+    CL_CHECK(clSetKernelArg(kernel,  7, sizeof(cl_ulong), &nb01));
+    CL_CHECK(clSetKernelArg(kernel,  8, sizeof(cl_ulong), &nb02));
+    CL_CHECK(clSetKernelArg(kernel,  9, sizeof(cl_ulong), &nb03));
+    CL_CHECK(clSetKernelArg(kernel, 10, sizeof(int),      &ne12));
+    CL_CHECK(clSetKernelArg(kernel, 11, sizeof(int),      &ne13));
+    CL_CHECK(clSetKernelArg(kernel, 12, sizeof(cl_ulong), &nb11));
+    CL_CHECK(clSetKernelArg(kernel, 13, sizeof(cl_ulong), &nb12));
+    CL_CHECK(clSetKernelArg(kernel, 14, sizeof(cl_ulong), &nb13));
+    CL_CHECK(clSetKernelArg(kernel, 15, sizeof(cl_ulong), &nb1));
+    CL_CHECK(clSetKernelArg(kernel, 16, sizeof(cl_ulong), &nb2));
+    CL_CHECK(clSetKernelArg(kernel, 17, sizeof(cl_ulong), &nb3));
+    CL_CHECK(clSetKernelArg(kernel, 18, sizeof(float),    &scale));
+    CL_CHECK(clSetKernelArg(kernel, 19, sizeof(float),    &max_bias));
+    CL_CHECK(clSetKernelArg(kernel, 20, sizeof(float),    &m0));
+    CL_CHECK(clSetKernelArg(kernel, 21, sizeof(float),    &m1));
+    CL_CHECK(clSetKernelArg(kernel, 22, sizeof(int),      &n_head_log2));
 
     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
-    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
+    backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
 }
 
 static void ggml_cl_rope(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
@@ -5188,7 +5912,6 @@ static void ggml_cl_rope(ggml_backend_t backend, const ggml_tensor * src0, const
     GGML_ASSERT(dst->extra);
 
     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 * extra1 = (ggml_tensor_extra_cl *)src1->extra;
@@ -5354,15 +6077,7 @@ static void ggml_cl_rope(ggml_backend_t backend, const ggml_tensor * src0, const
     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
-    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
+    backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
 }
 
 static void ggml_cl_im2col(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
@@ -5377,7 +6092,6 @@ static void ggml_cl_im2col(ggml_backend_t backend, const ggml_tensor * src0, con
     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;
@@ -5446,15 +6160,7 @@ static void ggml_cl_im2col(ggml_backend_t backend, const ggml_tensor * src0, con
     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
+    backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
 }
 
 static void ggml_cl_argsort(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
@@ -5469,7 +6175,6 @@ static void ggml_cl_argsort(ggml_backend_t backend, const ggml_tensor * src0, co
     GGML_ASSERT(ggml_is_contiguous(src0));
 
     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;
@@ -5501,15 +6206,7 @@ static void ggml_cl_argsort(ggml_backend_t backend, const ggml_tensor * src0, co
     size_t global_work_size[] = {(size_t)ne00_padded, (size_t)nrows, (size_t)1};
     size_t local_work_size[] = {(size_t)ne00_padded, 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
+    backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
 }
 
 static void ggml_cl_sum_rows(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
@@ -5523,7 +6220,6 @@ static void ggml_cl_sum_rows(ggml_backend_t backend, const ggml_tensor * src0, c
     GGML_ASSERT(ggml_is_contiguous(src0));
 
     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;
@@ -5564,15 +6260,106 @@ static void ggml_cl_sum_rows(ggml_backend_t backend, const ggml_tensor * src0, c
     size_t global_work_size[] = {(size_t)ne01, (size_t)ne02, (size_t)ne03};
     size_t local_work_size[] = {(size_t)64, 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));
+    backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
+}
 
-    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
+static void ggml_cl_glu(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);
+
+    GGML_ASSERT(ggml_is_contiguous_1(src0));
+
+    if (src1) {
+        GGML_ASSERT(src1);
+        GGML_ASSERT(src1->extra);
+        GGML_ASSERT(ggml_are_same_shape(src0, src1));
+    }
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    cl_kernel kernel;
+    switch (ggml_get_glu_op(dst)) {
+        case GGML_GLU_OP_GEGLU:
+            if (dst->type == GGML_TYPE_F32) {
+                kernel = backend_ctx->kernel_geglu;
+            } else {
+                kernel = backend_ctx->kernel_geglu_f16;
+            }
+            break;
+        case GGML_GLU_OP_REGLU:
+            if (dst->type == GGML_TYPE_F32) {
+                kernel = backend_ctx->kernel_reglu;
+            } else {
+                kernel = backend_ctx->kernel_reglu_f16;
+            }
+            break;
+        case GGML_GLU_OP_SWIGLU:
+            if (dst->type == GGML_TYPE_F32) {
+                kernel = backend_ctx->kernel_swiglu;
+            } else {
+                kernel = backend_ctx->kernel_swiglu_f16;
+            }
+            break;
+        case GGML_GLU_OP_GEGLU_ERF:
+            if (dst->type == GGML_TYPE_F32) {
+                kernel = backend_ctx->kernel_geglu_erf;
+            } else {
+                kernel = backend_ctx->kernel_geglu_erf_f16;
+            }
+            break;
+        case GGML_GLU_OP_GEGLU_QUICK:
+            if (dst->type == GGML_TYPE_F32) {
+                kernel = backend_ctx->kernel_geglu_quick;
+            } else {
+                kernel = backend_ctx->kernel_geglu_quick_f16;
+            }
+            break;
+        default:
+            GGML_ABORT("Unsupported glu op");
+    }
+
+    ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra;
+    ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
+
+    ggml_tensor_extra_cl * extra1 = src1 ? (ggml_tensor_extra_cl *)src1->extra : nullptr;
+
+    cl_ulong offset0 = extra0->offset + src0->view_offs;
+    cl_ulong offsetd = extrad->offset + dst->view_offs;
+
+    cl_ulong offset1 = extra1 ? extra1->offset + src1->view_offs : offset0;
+
+    const int ne0       = dst->ne[0];
+
+    const cl_ulong nb01 = src0->nb[1];
+    const cl_ulong nb11 = src1 ? src1->nb[1] : nb01;
+
+    const cl_ulong nb1  = dst->nb[1];
+
+    const int swp = ((const int32_t *) dst->op_params)[1];
+    const int ne00_off = src1 ? 0 : (swp ? ne0 : 0);
+    const int ne10_off = src1 ? 0 : (swp ? 0 : ne0);
+
+    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),   src1 ? &extra1->data_device : &extra0->data_device));
+    CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &offset1));
+    CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),   &extrad->data_device));
+    CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &offsetd));
+    CL_CHECK(clSetKernelArg(kernel,  6, sizeof(cl_ulong), &nb01));
+    CL_CHECK(clSetKernelArg(kernel,  7, sizeof(cl_ulong), &nb11));
+    CL_CHECK(clSetKernelArg(kernel,  8, sizeof(int),      &ne0));
+    CL_CHECK(clSetKernelArg(kernel,  9, sizeof(cl_ulong), &nb1));
+    CL_CHECK(clSetKernelArg(kernel, 10, sizeof(int),      &ne00_off));
+    CL_CHECK(clSetKernelArg(kernel, 11, sizeof(int),      &ne10_off));
+
+    const size_t nrows = ggml_nrows(src0);
+    size_t nth = 512;
+    size_t global_work_size[] = {nrows*nth, 1, 1};
+    size_t local_work_size[] = {nth, 1, 1};
+
+    backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
 }
 
 //------------------------------------------------------------------------------
@@ -5643,6 +6430,12 @@ bool ggml_cl_compute_forward(ggml_backend_t backend, struct ggml_tensor * tensor
                     }
                     func = ggml_cl_gelu;
                     break;
+                case GGML_UNARY_OP_GELU_ERF:
+                    if (!any_on_device) {
+                        return false;
+                    }
+                    func = ggml_cl_gelu_erf;
+                    break;
                 case GGML_UNARY_OP_GELU_QUICK:
                     if (!any_on_device) {
                         return false;
@@ -5667,9 +6460,21 @@ bool ggml_cl_compute_forward(ggml_backend_t backend, struct ggml_tensor * tensor
                     }
                     func = ggml_cl_sigmoid;
                     break;
+                case GGML_UNARY_OP_TANH:
+                    if (!any_on_device) {
+                        return false;
+                    }
+                    func = ggml_cl_tanh;
+                    break;
                 default:
                     return false;
             } break;
+        case GGML_OP_GLU:
+            if (!any_on_device) {
+                return false;
+            }
+            func = ggml_cl_glu;
+            break;
         case GGML_OP_CLAMP:
             if (!any_on_device) {
                 return false;
@@ -5694,12 +6499,48 @@ bool ggml_cl_compute_forward(ggml_backend_t backend, struct ggml_tensor * tensor
             }
             func = ggml_cl_group_norm;
             break;
+                case GGML_OP_REPEAT:
+             if (!any_on_device) {
+                return false;
+            }
+            func = ggml_cl_repeat;
+            break;
+        case GGML_OP_PAD:
+            if (!any_on_device) {
+                return false;
+            }
+            ggml_cl_pad(backend, tensor->src[0], tensor);
+            return true;
+        case GGML_OP_UPSCALE:
+            if (!any_on_device) {
+                return false;
+            }
+            ggml_cl_upscale(backend, tensor->src[0], tensor);
+            return true;
+        case GGML_OP_CONCAT:
+            if (!any_on_device) {
+                return false;
+            }
+            func = ggml_cl_concat;
+            break;
+        case GGML_OP_TIMESTEP_EMBEDDING:
+            if (!any_on_device) {
+                return false;
+            }
+            ggml_cl_timestep_embedding(backend, tensor->src[0], tensor);
+            return true;
         case GGML_OP_MUL_MAT:
             if (!any_on_device && !ggml_cl_can_mul_mat(tensor->src[0], tensor->src[1], tensor)) {
                 return false;
             }
             func = ggml_cl_mul_mat;
             break;
+        case GGML_OP_MUL_MAT_ID:
+            if (!any_on_device) {
+                return false;
+            }
+            func = ggml_cl_mul_mat_id;
+            break;
         case GGML_OP_SCALE:
             if (!any_on_device) {
                 return false;
diff --git a/ggml/src/ggml-opencl/kernels/concat.cl b/ggml/src/ggml-opencl/kernels/concat.cl
new file mode 100644
index 000000000..132758469
--- /dev/null
+++ b/ggml/src/ggml-opencl/kernels/concat.cl
@@ -0,0 +1,109 @@
+kernel void kernel_concat_f32_contiguous(
+    global const char * p_src0, ulong off_src0,
+    global const char * p_src1, ulong off_src1,
+    global char * p_dst, ulong off_dst,
+    int d_ne00, int d_ne01, int d_ne02, // src0->ne[0..2] for the slice
+    int d_ne10, int d_ne11, int d_ne12, // src1->ne[0..2] for the slice (d_ne1X must match d_ne0X on non-concat axes)
+    int d_ne0,  int d_ne1,  int d_ne2,  // dst->ne[0..2] for the slice
+    int dim
+) {
+    global const float * src0 = (global const float*)((global char*)p_src0 + off_src0);
+    global const float * src1 = (global const float*)((global char*)p_src1 + off_src1);
+    global float * dst        = (global float*)((global char*)p_dst + off_dst);
+
+    int i0 = get_global_id(0); // Index along dst's 0th dimension
+    int i1 = get_global_id(1); // Index along dst's 1st dimension
+    int i2 = get_global_id(2); // Index along dst's 2nd dimension
+
+    if (i0 >= d_ne0 || i1 >= d_ne1 || i2 >= d_ne2) {
+        return;
+    }
+
+    ulong dst_idx = (ulong)i2 * d_ne0 * d_ne1 + (ulong)i1 * d_ne0 + i0;
+    ulong src_idx;
+
+    if (dim == 0) {
+        if (i0 < d_ne00) { // Data from src0
+            src_idx = (ulong)i2 * d_ne00 * d_ne01 + (ulong)i1 * d_ne00 + i0;
+            dst[dst_idx] = src0[src_idx];
+        } else { // Data from src1
+            src_idx = (ulong)i2 * d_ne10 * d_ne11 + (ulong)i1 * d_ne10 + (i0 - d_ne00);
+            dst[dst_idx] = src1[src_idx];
+        }
+    } else if (dim == 1) {
+        if (i1 < d_ne01) { // Data from src0
+            src_idx = (ulong)i2 * d_ne00 * d_ne01 + (ulong)i1 * d_ne00 + i0;
+            dst[dst_idx] = src0[src_idx];
+        } else { // Data from src1
+            src_idx = (ulong)i2 * d_ne10 * d_ne11 + (ulong)(i1 - d_ne01) * d_ne10 + i0;
+            dst[dst_idx] = src1[src_idx];
+        }
+    } else if (dim == 2) {
+        if (i2 < d_ne02) { // Data from src0
+            src_idx = (ulong)i2 * d_ne00 * d_ne01 + (ulong)i1 * d_ne00 + i0;
+            dst[dst_idx] = src0[src_idx];
+        } else { // Data from src1
+
+            src_idx = (ulong)(i2 - d_ne02) * d_ne10 * d_ne11 + (ulong)i1 * d_ne10 + i0;
+            dst[dst_idx] = src1[src_idx];
+        }
+    }
+}
+
+kernel void kernel_concat_f32_non_contiguous(
+    global const char * p_src0, ulong off_src0,
+    global const char * p_src1, ulong off_src1,
+    global char * p_dst, ulong off_dst,
+
+    long ne00, long ne01, long ne02, long ne03,
+    ulong nb00, ulong nb01, ulong nb02, ulong nb03,
+
+    ulong nb10, ulong nb11, ulong nb12, ulong nb13, // Strides for src1
+
+    long d_ne0, long d_ne1, long d_ne2, long d_ne3,
+    ulong d_nb0, ulong d_nb1, ulong d_nb2, ulong d_nb3,
+    int dim
+) {
+    global const char * src0_base = p_src0 + off_src0;
+    global const char * src1_base = p_src1 + off_src1;
+    global char * dst_base        = p_dst + off_dst;
+
+    long current_i1 = get_global_id(0); // Index for dst_dim_1
+    long current_i2 = get_global_id(1); // Index for dst_dim_2
+    long current_i3 = get_global_id(2); // Index for dst_dim_3
+
+    if (current_i1 >= d_ne1 || current_i2 >= d_ne2 || current_i3 >= d_ne3) {
+        return;
+    }
+
+    global const float * x_val_ptr;
+    global float * y_val_ptr;
+
+    for (long current_i0 = 0; current_i0 < d_ne0; ++current_i0) {
+        bool use_src0;
+        long s_i0 = current_i0, s_i1 = current_i1, s_i2 = current_i2, s_i3 = current_i3;
+
+        if (dim == 0) {
+            use_src0 = (current_i0 < ne00);
+            if (!use_src0) { s_i0 = current_i0 - ne00; }
+        } else if (dim == 1) {
+            use_src0 = (current_i1 < ne01);
+            if (!use_src0) { s_i1 = current_i1 - ne01; }
+        } else if (dim == 2) {
+            use_src0 = (current_i2 < ne02);
+            if (!use_src0) { s_i2 = current_i2 - ne02; }
+        } else { // dim == 3
+            use_src0 = (current_i3 < ne03);
+            if (!use_src0) { s_i3 = current_i3 - ne03; }
+        }
+
+        if (use_src0) {
+            x_val_ptr = (global const float *)(src0_base + (ulong)s_i3*nb03 + (ulong)s_i2*nb02 + (ulong)s_i1*nb01 + (ulong)s_i0*nb00);
+        } else {
+            x_val_ptr = (global const float *)(src1_base + (ulong)s_i3*nb13 + (ulong)s_i2*nb12 + (ulong)s_i1*nb11 + (ulong)s_i0*nb10);
+        }
+
+        y_val_ptr = (global float *)(dst_base + (ulong)current_i3*d_nb3 + (ulong)current_i2*d_nb2 + (ulong)current_i1*d_nb1 + (ulong)current_i0*d_nb0);
+        *y_val_ptr = *x_val_ptr;
+    }
+}
diff --git a/ggml/src/ggml-opencl/kernels/gelu.cl b/ggml/src/ggml-opencl/kernels/gelu.cl
index 71c310cc9..1ab426c77 100644
--- a/ggml/src/ggml-opencl/kernels/gelu.cl
+++ b/ggml/src/ggml-opencl/kernels/gelu.cl
@@ -6,6 +6,7 @@
 #define GELU_COEF_A     0.044715f
 #define GELU_QUICK_COEF -1.702f
 #define SQRT_2_OVER_PI  0.79788456080286535587989211986876f
+#define SQRT_2_INV      0.70710678118654752440084436210484f
 
 kernel void kernel_gelu(
     global float * src0,
@@ -35,6 +36,32 @@ 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_erf(
+    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)] = 0.5f*x*(1.0f + erf(x*SQRT_2_INV));
+}
+
+kernel void kernel_gelu_erf_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)] = 0.5f*x*(1.0f + erf(x*SQRT_2_INV));
+}
+
 kernel void kernel_gelu_quick(
     global float * src0,
     ulong offset0,
diff --git a/ggml/src/ggml-opencl/kernels/glu.cl b/ggml/src/ggml-opencl/kernels/glu.cl
new file mode 100644
index 000000000..7cca16e6a
--- /dev/null
+++ b/ggml/src/ggml-opencl/kernels/glu.cl
@@ -0,0 +1,337 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#define GELU_COEF_A     0.044715f
+#define GELU_QUICK_COEF -1.702f
+#define SQRT_2_OVER_PI  0.79788456080286535587989211986876f
+#define SQRT_2_INV      0.70710678118654752440084436210484f
+
+//------------------------------------------------------------------------------
+// geglu
+//------------------------------------------------------------------------------
+kernel void kernel_geglu(
+    global char * src0,
+    ulong  offset0,
+    global char * src1,
+    ulong  offset1,
+    global char * dst,
+    ulong  offsetd,
+    ulong nb01,
+    ulong nb11,
+    int ne0,
+    ulong nb1,
+    int ne00_off,
+    int ne10_off
+) {
+    src0 = (global char*)((global char*)src0 + offset0);
+    src1 = (global char*)((global char*)src1 + offset1);
+    dst  = (global char*)((global char*)dst  + offsetd);
+
+    global float * src0_row = (global float *) ((global char *) src0 + get_group_id(0)*nb01) + ne00_off;
+    global float * src1_row = (global float *) ((global char *) src1 + get_group_id(0)*nb11) + ne10_off;
+    global float * dst_row  = (global float *) ((global char *) dst  + get_group_id(0)*nb1);
+
+    for (int i0 = get_local_id(0); i0 < ne0; i0 += get_local_size(0)) {
+        const float x0 = src0_row[i0];
+        const float x1 = src1_row[i0];
+
+        const float gelu = 0.5f*x0*(1.0f + tanh(SQRT_2_OVER_PI*x0*(1.0f + GELU_COEF_A*x0*x0)));
+
+        dst_row[i0] = gelu*x1;
+    }
+}
+
+kernel void kernel_geglu_f16(
+    global char * src0,
+    ulong  offset0,
+    global char * src1,
+    ulong  offset1,
+    global char * dst,
+    ulong  offsetd,
+    ulong nb01,
+    ulong nb11,
+    int ne0,
+    ulong nb1,
+    int ne00_off,
+    int ne10_off
+) {
+    src0 = (global char*)((global char*)src0 + offset0);
+    src1 = (global char*)((global char*)src1 + offset1);
+    dst  = (global char*)((global char*)dst  + offsetd);
+
+    global half * src0_row = (global half *) ((global char *) src0 + get_group_id(0)*nb01) + ne00_off;
+    global half * src1_row = (global half *) ((global char *) src1 + get_group_id(0)*nb11) + ne10_off;
+    global half * dst_row  = (global half *) ((global char *) dst  + get_group_id(0)*nb1);
+
+    for (int i0 = get_local_id(0); i0 < ne0; i0 += get_local_size(0)) {
+        const half x0 = src0_row[i0];
+        const half x1 = src1_row[i0];
+
+        const half gelu = 0.5f*x0*(1.0f + tanh(SQRT_2_OVER_PI*x0*(1.0f + GELU_COEF_A*x0*x0)));
+
+        dst_row[i0] = gelu*x1;
+    }
+}
+
+//------------------------------------------------------------------------------
+// reglu
+//------------------------------------------------------------------------------
+kernel void kernel_reglu(
+    global char * src0,
+    ulong  offset0,
+    global char * src1,
+    ulong  offset1,
+    global char * dst,
+    ulong  offsetd,
+    ulong nb01,
+    ulong nb11,
+    int ne0,
+    ulong nb1,
+    int ne00_off,
+    int ne10_off
+) {
+    src0 = (global char*)((global char*)src0 + offset0);
+    src1 = (global char*)((global char*)src1 + offset1);
+    dst  = (global char*)((global char*)dst  + offsetd);
+
+    global float * src0_row = (global float *) ((global char *) src0 + get_group_id(0)*nb01) + ne00_off;
+    global float * src1_row = (global float *) ((global char *) src1 + get_group_id(0)*nb11) + ne10_off;
+    global float * dst_row  = (global float *) ((global char *) dst  + get_group_id(0)*nb1);
+
+    for (int i0 = get_local_id(0); i0 < ne0; i0 += get_local_size(0)) {
+        const float x0 = src0_row[i0];
+        const float x1 = src1_row[i0];
+
+        dst_row[i0] = x0*x1*(x0 > 0.0f);
+    }
+}
+
+kernel void kernel_reglu_f16(
+    global char * src0,
+    ulong  offset0,
+    global char * src1,
+    ulong  offset1,
+    global char * dst,
+    ulong  offsetd,
+    ulong nb01,
+    ulong nb11,
+    int ne0,
+    ulong nb1,
+    int ne00_off,
+    int ne10_off
+) {
+    src0 = (global char*)((global char*)src0 + offset0);
+    src1 = (global char*)((global char*)src1 + offset1);
+    dst  = (global char*)((global char*)dst  + offsetd);
+
+    global half * src0_row = (global half *) ((global char *) src0 + get_group_id(0)*nb01) + ne00_off;
+    global half * src1_row = (global half *) ((global char *) src1 + get_group_id(0)*nb11) + ne10_off;
+    global half * dst_row  = (global half *) ((global char *) dst  + get_group_id(0)*nb1);
+
+    for (int i0 = get_local_id(0); i0 < ne0; i0 += get_local_size(0)) {
+        const half x0 = src0_row[i0];
+        const half x1 = src1_row[i0];
+
+        dst_row[i0] = x0*x1*(x0 > 0.0f);
+    }
+}
+
+//------------------------------------------------------------------------------
+// swiglu
+//------------------------------------------------------------------------------
+kernel void kernel_swiglu(
+    global char * src0,
+    ulong  offset0,
+    global char * src1,
+    ulong  offset1,
+    global char * dst,
+    ulong  offsetd,
+    ulong nb01,
+    ulong nb11,
+    int ne0,
+    ulong nb1,
+    int ne00_off,
+    int ne10_off
+) {
+    src0 = (global char*)((global char*)src0 + offset0);
+    src1 = (global char*)((global char*)src1 + offset1);
+    dst  = (global char*)((global char*)dst  + offsetd);
+
+    global float * src0_row = (global float *) ((global char *) src0 + get_group_id(0)*nb01) + ne00_off;
+    global float * src1_row = (global float *) ((global char *) src1 + get_group_id(0)*nb11) + ne10_off;
+    global float * dst_row  = (global float *) ((global char *) dst  + get_group_id(0)*nb1);
+
+    for (int i0 = get_local_id(0); i0 < ne0; i0 += get_local_size(0)) {
+        const float x0 = src0_row[i0];
+        const float x1 = src1_row[i0];
+
+        const float silu = x0 / (1.0f + exp(-x0));
+
+        dst_row[i0] = silu*x1;
+    }
+}
+
+kernel void kernel_swiglu_f16(
+    global char * src0,
+    ulong  offset0,
+    global char * src1,
+    ulong  offset1,
+    global char * dst,
+    ulong  offsetd,
+    ulong nb01,
+    ulong nb11,
+    int ne0,
+    ulong nb1,
+    int ne00_off,
+    int ne10_off
+) {
+    src0 = (global char*)((global char*)src0 + offset0);
+    src1 = (global char*)((global char*)src1 + offset1);
+    dst  = (global char*)((global char*)dst  + offsetd);
+
+    global half * src0_row = (global half *) ((global char *) src0 + get_group_id(0)*nb01) + ne00_off;
+    global half * src1_row = (global half *) ((global char *) src1 + get_group_id(0)*nb11) + ne10_off;
+    global half * dst_row  = (global half *) ((global char *) dst  + get_group_id(0)*nb1);
+
+    for (int i0 = get_local_id(0); i0 < ne0; i0 += get_local_size(0)) {
+        const half x0 = src0_row[i0];
+        const half x1 = src1_row[i0];
+
+        const half silu = x0 / (1.0f + exp(-x0));
+
+        dst_row[i0] = silu*x1;
+    }
+}
+
+//------------------------------------------------------------------------------
+// geglu_erf
+//------------------------------------------------------------------------------
+kernel void kernel_geglu_erf(
+    global char * src0,
+    ulong  offset0,
+    global char * src1,
+    ulong  offset1,
+    global char * dst,
+    ulong  offsetd,
+    ulong nb01,
+    ulong nb11,
+    int ne0,
+    ulong nb1,
+    int ne00_off,
+    int ne10_off
+) {
+    src0 = (global char*)((global char*)src0 + offset0);
+    src1 = (global char*)((global char*)src1 + offset1);
+    dst  = (global char*)((global char*)dst  + offsetd);
+
+    global float * src0_row = (global float *) ((global char *) src0 + get_group_id(0)*nb01) + ne00_off;
+    global float * src1_row = (global float *) ((global char *) src1 + get_group_id(0)*nb11) + ne10_off;
+    global float * dst_row  = (global float *) ((global char *) dst  + get_group_id(0)*nb1);
+
+    for (int i0 = get_local_id(0); i0 < ne0; i0 += get_local_size(0)) {
+        const float x0 = src0_row[i0];
+        const float x1 = src1_row[i0];
+
+        const float gelu_erf = 0.5f*x0*(1.0f + erf(x0*SQRT_2_INV));
+
+        dst_row[i0] = gelu_erf*x1;
+    }
+}
+
+kernel void kernel_geglu_erf_f16(
+    global char * src0,
+    ulong  offset0,
+    global char * src1,
+    ulong  offset1,
+    global char * dst,
+    ulong  offsetd,
+    ulong nb01,
+    ulong nb11,
+    int ne0,
+    ulong nb1,
+    int ne00_off,
+    int ne10_off
+) {
+    src0 = (global char*)((global char*)src0 + offset0);
+    src1 = (global char*)((global char*)src1 + offset1);
+    dst  = (global char*)((global char*)dst  + offsetd);
+
+    global half * src0_row = (global half *) ((global char *) src0 + get_group_id(0)*nb01) + ne00_off;
+    global half * src1_row = (global half *) ((global char *) src1 + get_group_id(0)*nb11) + ne10_off;
+    global half * dst_row  = (global half *) ((global char *) dst  + get_group_id(0)*nb1);
+
+    for (int i0 = get_local_id(0); i0 < ne0; i0 += get_local_size(0)) {
+        const half x0 = src0_row[i0];
+        const half x1 = src1_row[i0];
+
+        const half gelu_erf = 0.5f*x0*(1.0f + erf(x0*SQRT_2_INV));
+
+        dst_row[i0] = gelu_erf*x1;
+    }
+}
+
+//------------------------------------------------------------------------------
+// geglu_quick
+//------------------------------------------------------------------------------
+kernel void kernel_geglu_quick(
+    global char * src0,
+    ulong  offset0,
+    global char * src1,
+    ulong  offset1,
+    global char * dst,
+    ulong  offsetd,
+    ulong nb01,
+    ulong nb11,
+    int ne0,
+    ulong nb1,
+    int ne00_off,
+    int ne10_off
+) {
+    src0 = (global char*)((global char*)src0 + offset0);
+    src1 = (global char*)((global char*)src1 + offset1);
+    dst  = (global char*)((global char*)dst  + offsetd);
+
+    global float * src0_row = (global float *) ((global char *) src0 + get_group_id(0)*nb01) + ne00_off;
+    global float * src1_row = (global float *) ((global char *) src1 + get_group_id(0)*nb11) + ne10_off;
+    global float * dst_row  = (global float *) ((global char *) dst  + get_group_id(0)*nb1);
+
+    for (int i0 = get_local_id(0); i0 < ne0; i0 += get_local_size(0)) {
+        const float x0 = src0_row[i0];
+        const float x1 = src1_row[i0];
+
+        const float gelu_quick = x0*(1.0f/(1.0f + exp(GELU_QUICK_COEF*x0)));
+
+        dst_row[i0] = gelu_quick*x1;
+    }
+}
+
+kernel void kernel_geglu_quick_f16(
+    global char * src0,
+    ulong  offset0,
+    global char * src1,
+    ulong  offset1,
+    global char * dst,
+    ulong  offsetd,
+    ulong nb01,
+    ulong nb11,
+    int ne0,
+    ulong nb1,
+    int ne00_off,
+    int ne10_off
+) {
+    src0 = (global char*)((global char*)src0 + offset0);
+    src1 = (global char*)((global char*)src1 + offset1);
+    dst  = (global char*)((global char*)dst  + offsetd);
+
+    global half * src0_row = (global half *) ((global char *) src0 + get_group_id(0)*nb01) + ne00_off;
+    global half * src1_row = (global half *) ((global char *) src1 + get_group_id(0)*nb11) + ne10_off;
+    global half * dst_row  = (global half *) ((global char *) dst  + get_group_id(0)*nb1);
+
+    for (int i0 = get_local_id(0); i0 < ne0; i0 += get_local_size(0)) {
+        const half x0 = src0_row[i0];
+        const half x1 = src1_row[i0];
+
+        const half gelu_quick = x0*(1.0f/(1.0f + exp(GELU_QUICK_COEF*x0)));
+
+        dst_row[i0] = gelu_quick*x1;
+    }
+}
diff --git a/ggml/src/ggml-opencl/kernels/mul_mv_id_q4_0_f32_8x_flat.cl b/ggml/src/ggml-opencl/kernels/mul_mv_id_q4_0_f32_8x_flat.cl
new file mode 100644
index 000000000..7ccf41efb
--- /dev/null
+++ b/ggml/src/ggml-opencl/kernels/mul_mv_id_q4_0_f32_8x_flat.cl
@@ -0,0 +1,283 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_subgroups
+#pragma OPENCL EXTENSION cl_intel_subgroups : enable
+#else
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#endif
+
+#ifdef cl_intel_required_subgroup_size
+#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)
+#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")))
+#endif
+
+#define QK4_0                   32
+
+typedef char int8_t;
+typedef uchar uint8_t;
+typedef short int16_t;
+typedef ushort uint16_t;
+typedef int int32_t;
+typedef uint uint32_t;
+
+//------------------------------------------------------------------------------
+// block_q4_0
+//------------------------------------------------------------------------------
+struct block_q4_0
+{
+    half d;
+    uint8_t qs[QK4_0 / 2];
+};
+
+// This function requires the original shuffled weights.
+// As a reminder, the original weights are shuffled so that (q[0], q[16]) are
+// packed together in a byte, so are (q[1], q[17]) and so on.
+inline float block_q_4_0_dot_y_flat(
+        global uchar * x,
+        global half  * dh,
+        float sumy,
+        float16 yl,
+        int il
+) {
+    float           d   = *dh;
+    global ushort * qs  = ((global ushort *)x + il/2);
+    float           acc = 0.f;
+
+    acc += yl.s0 * (qs[0] & 0x000F);
+    acc += yl.s1 * (qs[0] & 0x0F00);
+    acc += yl.s8 * (qs[0] & 0x00F0);
+    acc += yl.s9 * (qs[0] & 0xF000);
+
+    acc += yl.s2 * (qs[1] & 0x000F);
+    acc += yl.s3 * (qs[1] & 0x0F00);
+    acc += yl.sa * (qs[1] & 0x00F0);
+    acc += yl.sb * (qs[1] & 0xF000);
+
+    acc += yl.s4 * (qs[2] & 0x000F);
+    acc += yl.s5 * (qs[2] & 0x0F00);
+    acc += yl.sc * (qs[2] & 0x00F0);
+    acc += yl.sd * (qs[2] & 0xF000);
+
+    acc += yl.s6 * (qs[3] & 0x000F);
+    acc += yl.s7 * (qs[3] & 0x0F00);
+    acc += yl.se * (qs[3] & 0x00F0);
+    acc += yl.sf * (qs[3] & 0xF000);
+
+    return d * (sumy * -8.f + acc);
+}
+
+//
+// This variant outputs 8 values.
+//
+#undef N_DST
+#undef N_SIMDGROUP
+#undef N_SIMDWIDTH
+
+#ifdef INTEL_GPU
+#define N_DST 8 // each SIMD group works on 8 rows
+#define N_SIMDGROUP 1 // number of SIMD groups in a thread group
+#define N_SIMDWIDTH 16 // subgroup size
+#elif defined (ADRENO_GPU)
+#define N_DST 8
+#define N_SIMDGROUP 1
+#define N_SIMDWIDTH 64
+#endif
+
+inline void mul_vec_q_n_f32_8x_flat(
+        global char  * src0_q,
+        global half  * src0_d,
+        global float * src1,
+        global float * dst,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne10,
+        int ne12,
+        int ne0,
+        int ne1,
+        int r2,
+        int r3
+) {
+    const ulong nb = ne00/QK4_0;
+
+    int r0 = get_group_id(0);
+    int r1 = get_group_id(1);
+    int im = 0;
+
+    int first_row = (r0 * N_SIMDGROUP + get_sub_group_id()) * N_DST;
+
+    int i12 = im%ne12;
+    int i13 = im/ne12;
+
+    // The number of scales is the same as the number of blocks.
+    ulong offset0_d = first_row * nb + (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
+    // Each block contains QK4_0/2 uchars, hence offset for qs is as follows.
+    ulong offset0_q = (first_row * nb + (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02)) * QK4_0/2;
+
+    global uchar * x = (global uchar *) src0_q + offset0_q;
+    global half  * d = (global half  *) src0_d + offset0_d;
+    global float * y = (global float *) src1   + r1*ne10 + im*ne00*ne1;
+
+    float16 yl;
+    float8 sumf = 0.f;
+
+    int ix = get_sub_group_local_id()/2;
+    int il = 8*(get_sub_group_local_id()%2);
+
+    global float * yb = y + ix*QK4_0 + il;
+
+    for (int ib = ix; ib < nb; ib += N_SIMDWIDTH/2) {
+        float sumy = 0.f;
+
+        sumy += yb[0];
+        sumy += yb[1];
+        sumy += yb[2];
+        sumy += yb[3];
+        sumy += yb[4];
+        sumy += yb[5];
+        sumy += yb[6];
+        sumy += yb[7];
+
+        sumy += yb[16];
+        sumy += yb[17];
+        sumy += yb[18];
+        sumy += yb[19];
+        sumy += yb[20];
+        sumy += yb[21];
+        sumy += yb[22];
+        sumy += yb[23];
+
+        yl.s0 = yb[0];
+        yl.s1 = yb[1]/256.f;
+
+        yl.s2 = yb[2];
+        yl.s3 = yb[3]/256.f;
+
+        yl.s4 = yb[4];
+        yl.s5 = yb[5]/256.f;
+
+        yl.s6 = yb[6];
+        yl.s7 = yb[7]/256.f;
+
+        yl.s8 = yb[16]/16.f;
+        yl.s9 = yb[17]/4096.f;
+
+        yl.sa = yb[18]/16.f;
+        yl.sb = yb[19]/4096.f;
+
+        yl.sc = yb[20]/16.f;
+        yl.sd = yb[21]/4096.f;
+
+        yl.se = yb[22]/16.f;
+        yl.sf = yb[23]/4096.f;
+
+        sumf.s0 += block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 0*nb*QK4_0/2, d + ib + 0*nb, sumy, yl, il);
+        sumf.s1 += block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 1*nb*QK4_0/2, d + ib + 1*nb, sumy, yl, il);
+        sumf.s2 += block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 2*nb*QK4_0/2, d + ib + 2*nb, sumy, yl, il);
+        sumf.s3 += block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 3*nb*QK4_0/2, d + ib + 3*nb, sumy, yl, il);
+
+        sumf.s4 += block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 4*nb*QK4_0/2, d + ib + 4*nb, sumy, yl, il);
+        sumf.s5 += block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 5*nb*QK4_0/2, d + ib + 5*nb, sumy, yl, il);
+        sumf.s6 += block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 6*nb*QK4_0/2, d + ib + 6*nb, sumy, yl, il);
+        sumf.s7 += block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 7*nb*QK4_0/2, d + ib + 7*nb, sumy, yl, il);
+
+        yb += QK4_0 * (N_SIMDWIDTH/2);
+    }
+
+    float8 tot = (float8)(
+        sub_group_reduce_add(sumf.s0), sub_group_reduce_add(sumf.s1),
+        sub_group_reduce_add(sumf.s2), sub_group_reduce_add(sumf.s3),
+        sub_group_reduce_add(sumf.s4), sub_group_reduce_add(sumf.s5),
+        sub_group_reduce_add(sumf.s6), sub_group_reduce_add(sumf.s7)
+    );
+
+    if (get_sub_group_local_id() == 0) {
+        if (first_row + 0 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 0] = tot.s0;
+        }
+        if (first_row + 1 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 1] = tot.s1;
+        }
+        if (first_row + 2 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 2] = tot.s2;
+        }
+        if (first_row + 3 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 3] = tot.s3;
+        }
+
+        if (first_row + 4 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 4] = tot.s4;
+        }
+        if (first_row + 5 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 5] = tot.s5;
+        }
+        if (first_row + 6 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 6] = tot.s6;
+        }
+        if (first_row + 7 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 7] = tot.s7;
+        }
+    }
+}
+
+#ifdef INTEL_GPU
+REQD_SUBGROUP_SIZE_16
+#elif defined (ADRENO_GPU)
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_mul_mv_id_q4_0_f32_8x_flat(
+        global char  *  src0_q,
+        global half  *  src0_d,
+        global float *  src1,
+        ulong           offset1,
+        global char  *  src2,
+        ulong           offset2,
+        global float *  dst,
+        ulong           offsetd,
+        int             ne00,
+        int             ne01,
+        int             ne02,
+        ulong           nb00,
+        ulong           nb02,
+        int             ne10,
+        int             ne11,
+        int             ne12,
+        ulong           nb11,
+        ulong           nb12,
+        int             ne20,
+        int             ne21,
+        ulong           nb21,
+        int             ne0,
+        int             ne1,
+        int             r2,
+        int             r3
+) {
+    src1 = (global float *)((global char *)src1 + offset1);
+    src2 = (global char  *)((global char *)src2 + offset2);
+    dst  = (global float *)((global char *)dst  + offsetd);
+
+    const int iid1 = get_group_id(2)/ne20;
+    const int idx  = get_group_id(2)%ne20;
+
+    const int i02 = ((global int *)(src2 + iid1*nb21))[idx];
+
+    const int i11 = idx%ne11;
+    const int i12 = iid1;
+
+    const int i1 = idx;
+    const int i2 = i12;
+
+    global char  * src0_q_cur = src0_q + (i02*nb02/nb00)*(QK4_0/2);
+    global half  * src0_d_cur = src0_d + (i02*nb02/nb00);
+    global float * src1_cur   = (global float *)((global char *) src1  + i11*nb11 + i12*nb12);
+    global float * dst_cur    = dst + i1*ne0 + i2*ne1*ne0;
+
+    mul_vec_q_n_f32_8x_flat(src0_q_cur, src0_d_cur, src1_cur, dst_cur, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3);
+}
diff --git a/ggml/src/ggml-opencl/kernels/pad.cl b/ggml/src/ggml-opencl/kernels/pad.cl
new file mode 100644
index 000000000..747fa7feb
--- /dev/null
+++ b/ggml/src/ggml-opencl/kernels/pad.cl
@@ -0,0 +1,30 @@
+kernel void kernel_pad(
+        global const void * src0_ptr,
+        ulong src0_offset,
+        global void * dst_ptr,
+        ulong dst_offset,
+        int s_ne0, int s_ne1, int s_ne2,
+        int d_ne0, int d_ne1, int d_ne2
+) {
+    global const float * src0 = (global const float *)((global const char *)src0_ptr + src0_offset);
+    global float * dst = (global float *)((global char *)dst_ptr + dst_offset);
+
+    int nidx   = get_global_id(0);
+    int idx_d1 = get_group_id(1);
+    int idx_d2 = get_group_id(2);
+
+    if (nidx >= d_ne0) {
+        return;
+    }
+
+    int dst_el_offset = nidx + idx_d1 * d_ne0 + idx_d2 * d_ne0 * d_ne1;
+
+    bool in_src_bounds = (nidx < s_ne0) && (idx_d1 < s_ne1) && (idx_d2 < s_ne2);
+
+    if (in_src_bounds) {
+        int src_el_offset = nidx + idx_d1 * s_ne0 + idx_d2 * s_ne0 * s_ne1;
+        dst[dst_el_offset] = src0[src_el_offset];
+    } else {
+        dst[dst_el_offset] = 0.0f;
+    }
+}
diff --git a/ggml/src/ggml-opencl/kernels/repeat.cl b/ggml/src/ggml-opencl/kernels/repeat.cl
new file mode 100644
index 000000000..079498f5a
--- /dev/null
+++ b/ggml/src/ggml-opencl/kernels/repeat.cl
@@ -0,0 +1,39 @@
+kernel void kernel_repeat(
+    global const char * src0_data_in,
+    global       char * dst_data_in,
+    ulong src0_offset,
+    ulong dst_offset,
+    int src0_ne0, int src0_ne1, int src0_ne2, int src0_ne3,
+    ulong src0_nb0, ulong src0_nb1, ulong src0_nb2, ulong src0_nb3,
+    int dst_ne0, int dst_ne1, int dst_ne2, int dst_ne3,
+    ulong dst_nb0, ulong dst_nb1, ulong dst_nb2, ulong dst_nb3
+) {
+    global const char * src0_data = src0_data_in + src0_offset;
+    global       char * dst_data  = dst_data_in + dst_offset;
+
+    const int d3 = get_global_id(2);
+    const int d2 = get_global_id(1);
+    const int d1 = get_global_id(0);
+
+    if (d3 >= dst_ne3 || d2 >= dst_ne2 || d1 >= dst_ne1) {
+        return;
+    }
+
+    const int s3 = d3 % src0_ne3;
+    const int s2 = d2 % src0_ne2;
+    const int s1 = d1 % src0_ne1;
+
+    const global char * p_src0_slice = src0_data + (ulong)s3*src0_nb3 + (ulong)s2*src0_nb2 + (ulong)s1*src0_nb1;
+    global char * p_dst_slice  = dst_data  + (ulong)d3*dst_nb3 + (ulong)d2*dst_nb2 + (ulong)d1*dst_nb1;
+
+    for (int d0 = 0; d0 < dst_ne0; ++d0) {
+        // Determine source index for dimension 0 based on tiling/broadcasting.
+        const int s0 = d0 % src0_ne0;
+
+        const global char * restrict current_src_el_ptr = p_src0_slice + (ulong)s0*src0_nb0;
+        global char * restrict current_dst_el_ptr  = p_dst_slice  + (ulong)d0*dst_nb0;
+        for (int k = 0; k < src0_nb0; ++k) {
+            current_dst_el_ptr[k] = current_src_el_ptr[k];
+        }
+    }
+}
diff --git a/ggml/src/ggml-opencl/kernels/softmax_4_f16.cl b/ggml/src/ggml-opencl/kernels/softmax_4_f16.cl
index 62c05369a..a6d8ede67 100644
--- a/ggml/src/ggml-opencl/kernels/softmax_4_f16.cl
+++ b/ggml/src/ggml-opencl/kernels/softmax_4_f16.cl
@@ -22,32 +22,45 @@
 REQD_SUBGROUP_SIZE_64
 #endif
 kernel void kernel_soft_max_4_f16(
-        global float * src0,
+        global char * src0,
         ulong offset0,
-        global half * src1,
+        global char * src1,
         ulong offset1,
-        global float * dst,
+        global char * dst,
         ulong offsetd,
         int ne00,
-        int ne01,
-        int ne02,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne12,
+        int ne13,
+        ulong nb11,
+        ulong nb12,
+        ulong nb13,
+        ulong nb1,
+        ulong nb2,
+        ulong nb3,
         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);
+    src0 = src0 + offset0;
+    src1 = src1 + offset1;
+    dst  = 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);
+    int i13 = i03%ne13;
+    int i12 = i02%ne12;
+    int i11 = i01;
+
+    global float4 * psrc4 = (global float4 *)(src0 + i01*nb01 + i02*nb02 + i03*nb03);
+    global half4  * pmask = src1 != src0 ? (global half4 *)(src1 + i11*nb11 + i12*nb12 + i13*nb13) : 0;
+    global float4 * pdst4 = (global float4 *)(dst  + i01*nb1 + i02*nb2 + i03*nb3);
 
     float slope = 1.0f;
 
diff --git a/ggml/src/ggml-opencl/kernels/softmax_4_f32.cl b/ggml/src/ggml-opencl/kernels/softmax_4_f32.cl
index d562774ea..35b5573b4 100644
--- a/ggml/src/ggml-opencl/kernels/softmax_4_f32.cl
+++ b/ggml/src/ggml-opencl/kernels/softmax_4_f32.cl
@@ -22,32 +22,45 @@
 REQD_SUBGROUP_SIZE_64
 #endif
 kernel void kernel_soft_max_4(
-        global float * src0,
+        global char * src0,
         ulong offset0,
-        global float * src1,
+        global char * src1,
         ulong offset1,
-        global float * dst,
+        global char * dst,
         ulong offsetd,
         int ne00,
-        int ne01,
-        int ne02,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne12,
+        int ne13,
+        ulong nb11,
+        ulong nb12,
+        ulong nb13,
+        ulong nb1,
+        ulong nb2,
+        ulong nb3,
         float scale,
         float max_bias,
         float m0,
         float m1,
         int n_head_log2
 ) {
-    src0 = (global float*)((global char*)src0 + offset0);
-    src1 = (global float*)((global char*)src1 + offset1);
-    dst = (global float*)((global char*)dst + offsetd);
+    src0 = src0 + offset0;
+    src1 = src1 + offset1;
+    dst  = 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 float4 * pmask = src1 != src0 ? (global float4 *)(src1 + i01*ne00) : 0;
-    global float4 * pdst4 = (global float4 *)(dst  + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00);
+    int i13 = i03%ne13;
+    int i12 = i02%ne12;
+    int i11 = i01;
+
+    global float4 * psrc4 = (global float4 *)(src0 + i01*nb01 + i02*nb02 + i03*nb03);
+    global float4 * pmask = src1 != src0 ? (global float4 *)(src1 + i11*nb11 + i12*nb12 + i13*nb13) : 0;
+    global float4 * pdst4 = (global float4 *)(dst  + i01*nb1 + i02*nb2 + i03*nb3);
 
     float slope = 1.0f;
 
diff --git a/ggml/src/ggml-opencl/kernels/softmax_f16.cl b/ggml/src/ggml-opencl/kernels/softmax_f16.cl
index d38d09967..9d292b574 100644
--- a/ggml/src/ggml-opencl/kernels/softmax_f16.cl
+++ b/ggml/src/ggml-opencl/kernels/softmax_f16.cl
@@ -22,32 +22,45 @@
 REQD_SUBGROUP_SIZE_64
 #endif
 kernel void kernel_soft_max_f16(
-        global float * src0,
+        global char * src0,
         ulong offset0,
-        global half * src1,
+        global char * src1,
         ulong offset1,
-        global float * dst,
+        global char * dst,
         ulong offsetd,
         int ne00,
-        int ne01,
-        int ne02,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne12,
+        int ne13,
+        ulong nb11,
+        ulong nb12,
+        ulong nb13,
+        ulong nb1,
+        ulong nb2,
+        ulong nb3,
         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);
+    src0 = src0 + offset0;
+    src1 = src1 + offset1;
+    dst  = 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;
+    int i13 = i03%ne13;
+    int i12 = i02%ne12;
+    int i11 = i01;
+
+    global float * psrc0 = (global float *)(src0 + i01*nb01 + i02*nb02 + i03*nb03);
+    global half  * pmask = src1 != src0 ? (global half *)(src1 + i11*nb11 + i12*nb12 + i13*nb13) : 0;
+    global float * pdst  = (global float *)(dst  + i01*nb1 + i02*nb2 + i03*nb3);
 
     float slope = 1.0f;
 
diff --git a/ggml/src/ggml-opencl/kernels/softmax_f32.cl b/ggml/src/ggml-opencl/kernels/softmax_f32.cl
index 001b587ab..7c53dfbe5 100644
--- a/ggml/src/ggml-opencl/kernels/softmax_f32.cl
+++ b/ggml/src/ggml-opencl/kernels/softmax_f32.cl
@@ -22,32 +22,45 @@
 REQD_SUBGROUP_SIZE_64
 #endif
 kernel void kernel_soft_max(
-        global float * src0,
+        global char * src0,
         ulong offset0,
-        global float * src1,
+        global char * src1,
         ulong offset1,
-        global float * dst,
+        global char * dst,
         ulong offsetd,
         int ne00,
-        int ne01,
-        int ne02,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne12,
+        int ne13,
+        ulong nb11,
+        ulong nb12,
+        ulong nb13,
+        ulong nb1,
+        ulong nb2,
+        ulong nb3,
         float scale,
         float max_bias,
         float m0,
         float m1,
         int n_head_log2
 ) {
-    src0 = (global float*)((global char*)src0 + offset0);
-    src1 = (global float*)((global char*)src1 + offset1);
-    dst = (global float*)((global char*)dst + offsetd);
+    src0 = src0 + offset0;
+    src1 = src1 + offset1;
+    dst  = 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 float * pmask = src1 != src0 ? src1 + i01*ne00 : 0;
-    global float * pdst  = dst  + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
+    int i13 = i03%ne13;
+    int i12 = i02%ne12;
+    int i11 = i01;
+
+    global float * psrc0 = (global float *)(src0 + i01*nb01 + i02*nb02 + i03*nb03);
+    global float * pmask = src1 != src0 ? (global float *)(src1 + i11*nb11 + i12*nb12 + i13*nb13) : 0;
+    global float * pdst  = (global float *)(dst  + i01*nb1 + i02*nb2 + i03*nb3);
 
     float slope = 1.0f;
 
diff --git a/ggml/src/ggml-opencl/kernels/tanh.cl b/ggml/src/ggml-opencl/kernels/tanh.cl
new file mode 100644
index 000000000..d9da86b14
--- /dev/null
+++ b/ggml/src/ggml-opencl/kernels/tanh.cl
@@ -0,0 +1,63 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_required_subgroup_size
+#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)
+#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")))
+#endif
+
+kernel void kernel_tanh_f32_nd(
+    global void * p_src0_base, ulong off_src0_abs,
+    global void * p_dst_base,  ulong off_dst_abs,
+    int ne00, int ne01, int ne02, int ne03,
+    ulong nb00, ulong nb01, ulong nb02, ulong nb03,
+    int ne10, int ne11, int ne12, int ne13,
+    ulong nb10, ulong nb11, ulong nb12, ulong nb13
+) {
+    int i0 = get_global_id(0);
+    int i1 = get_global_id(1);
+    int i2 = get_global_id(2);
+
+    if (i0 < ne10 && i1 < ne11 && i2 < ne12) {
+        for (int i3 = 0; i3 < ne13; ++i3) {
+            ulong src_offset_in_tensor = (ulong)i0*nb00 + (ulong)i1*nb01 + (ulong)i2*nb02 + (ulong)i3*nb03;
+            global const float *src_val_ptr = (global const float *)((global char *)p_src0_base + off_src0_abs + src_offset_in_tensor);
+
+            ulong dst_offset_in_tensor = (ulong)i0*nb10 + (ulong)i1*nb11 + (ulong)i2*nb12 + (ulong)i3*nb13;
+            global float *dst_val_ptr = (global float *)((global char *)p_dst_base + off_dst_abs + dst_offset_in_tensor);
+
+            *dst_val_ptr = tanh(*src_val_ptr);
+        }
+    }
+}
+
+kernel void kernel_tanh_f16_nd(
+    global void * p_src0_base, ulong off_src0_abs,
+    global void * p_dst_base,  ulong off_dst_abs,
+    int ne00, int ne01, int ne02, int ne03,
+    ulong nb00, ulong nb01, ulong nb02, ulong nb03,
+    int ne10, int ne11, int ne12, int ne13,
+    ulong nb10, ulong nb11, ulong nb12, ulong nb13
+) {
+    int i0 = get_global_id(0);
+    int i1 = get_global_id(1);
+    int i2 = get_global_id(2);
+
+    if (i0 < ne10 && i1 < ne11 && i2 < ne12) {
+        for (int i3 = 0; i3 < ne13; ++i3) {
+            ulong src_offset_in_tensor = (ulong)i0*nb00 + (ulong)i1*nb01 + (ulong)i2*nb02 + (ulong)i3*nb03;
+            global const half *src_val_ptr = (global const half *)((global char *)p_src0_base + off_src0_abs + src_offset_in_tensor);
+
+            ulong dst_offset_in_tensor = (ulong)i0*nb10 + (ulong)i1*nb11 + (ulong)i2*nb12 + (ulong)i3*nb13;
+            global half *dst_val_ptr = (global half *)((global char *)p_dst_base + off_dst_abs + dst_offset_in_tensor);
+
+            *dst_val_ptr = tanh(*src_val_ptr);
+        }
+    }
+}
diff --git a/ggml/src/ggml-opencl/kernels/tsembd.cl b/ggml/src/ggml-opencl/kernels/tsembd.cl
new file mode 100644
index 000000000..4b1107f70
--- /dev/null
+++ b/ggml/src/ggml-opencl/kernels/tsembd.cl
@@ -0,0 +1,48 @@
+kernel void kernel_timestep_embedding(
+    global const void * p_timesteps,
+    ulong off_timesteps,
+    global void * p_dst,
+    ulong off_dst,
+    int dst_nb1_bytes,
+    int logical_dim,
+    int max_period
+) {
+    int local_i;
+    int local_j;
+    int local_half_dim;
+    float local_timestep_val;
+    float local_freq;
+    float local_arg;
+    global float * local_embed_data_ptr;
+    global const float * local_timesteps_input_ptr;
+    global float * local_dst_output_base_ptr;
+
+    local_timesteps_input_ptr = (global const float *)((global char *)p_timesteps + off_timesteps);
+    local_dst_output_base_ptr = (global float *)((global char *)p_dst + off_dst);
+
+    local_i = get_global_id(1);
+    local_j = get_global_id(0);
+
+    local_half_dim = logical_dim / 2;
+    local_embed_data_ptr = (global float *)((global char *)local_dst_output_base_ptr + local_i * dst_nb1_bytes);
+
+    if (logical_dim % 2 != 0 && local_j == ((logical_dim + 1) / 2)) {
+        local_embed_data_ptr[logical_dim] = 0.0f;
+    }
+
+    if (local_j >= local_half_dim) {
+        return;
+    }
+
+    local_timestep_val = local_timesteps_input_ptr[local_i];
+
+    if (local_half_dim == 0) {
+        local_freq = 1.0f;
+    } else {
+        local_freq = exp(-log((float)max_period) * (float)local_j / (float)local_half_dim);
+    }
+
+    local_arg = local_timestep_val * local_freq;
+    local_embed_data_ptr[local_j] = cos(local_arg);
+    local_embed_data_ptr[local_j + local_half_dim] = sin(local_arg);
+}
diff --git a/ggml/src/ggml-opencl/kernels/upscale.cl b/ggml/src/ggml-opencl/kernels/upscale.cl
new file mode 100644
index 000000000..25c68351b
--- /dev/null
+++ b/ggml/src/ggml-opencl/kernels/upscale.cl
@@ -0,0 +1,120 @@
+kernel void kernel_upscale(
+    global const void * p_src0,
+    ulong off_src0,
+    global void * p_dst,
+    ulong off_dst,
+    ulong nb00,
+    ulong nb01,
+    ulong nb02,
+    ulong nb03,
+    int ne10,
+    int ne11,
+    int ne12,
+    int ne13,
+    float sf0,
+    float sf1,
+    float sf2,
+    float sf3
+) {
+    global const char * src_base = (global const char *)p_src0 + off_src0;
+    global float * dst_base = (global float *)((global char *)p_dst + off_dst);
+
+    int index = get_global_id(0);
+    int dst_total_elements = ne10 * ne11 * ne12 * ne13;
+
+    if (index >= dst_total_elements) {
+        return;
+    }
+
+    int i10 = index % ne10;
+    int i11 = (index / ne10) % ne11;
+    int i12 = (index / (ne10 * ne11)) % ne12;
+    int i13 = index / (ne10 * ne11 * ne12);
+
+    int i00 = (int)(i10 / sf0);
+    int i01 = (int)(i11 / sf1);
+    int i02 = (int)(i12 / sf2);
+    int i03 = (int)(i13 / sf3);
+
+    ulong offset_src_element = (ulong)i03 * nb03 + (ulong)i02 * nb02 + (ulong)i01 * nb01 + (ulong)i00 * nb00;
+    global const float * src_element_ptr = (global const float *)(src_base + offset_src_element);
+
+    dst_base[index] = *src_element_ptr;
+}
+
+kernel void kernel_upscale_bilinear(
+    global const void * p_src0,
+    ulong off_src0,
+    global void * p_dst,
+    ulong off_dst,
+    ulong nb00,
+    ulong nb01,
+    ulong nb02,
+    ulong nb03,
+    int ne00_src,
+    int ne01_src,
+    int ne10_dst,
+    int ne11_dst,
+    int ne12_dst,
+    int ne13_dst,
+    float sf0,
+    float sf1,
+    float sf2,
+    float sf3,
+    float pixel_offset
+) {
+    global const char * src_base = (global const char *)p_src0 + off_src0;
+    global float * dst_base = (global float *)((global char *)p_dst + off_dst);
+
+    int index = get_global_id(0);
+    int dst_total_elements = ne10_dst * ne11_dst * ne12_dst * ne13_dst;
+
+    if (index >= dst_total_elements) {
+        return;
+    }
+
+    int i10_dst = index % ne10_dst;
+    int i11_dst = (index / ne10_dst) % ne11_dst;
+    int i12_dst = (index / (ne10_dst * ne11_dst)) % ne12_dst;
+    int i13_dst = index / (ne10_dst * ne11_dst * ne12_dst);
+
+    int i02_src = (int)(i12_dst / sf2);
+    int i03_src = (int)(i13_dst / sf3);
+
+    float y_src_f = ((float)i11_dst + pixel_offset) / sf1 - pixel_offset;
+    long y0_src = (long)floor(y_src_f);
+    long y1_src = y0_src + 1;
+
+    y0_src = max(0L, min(y0_src, (long)ne01_src - 1));
+    y1_src = max(0L, min(y1_src, (long)ne01_src - 1));
+
+    float dy = y_src_f - (float)y0_src;
+    dy = max(0.0f, min(dy, 1.0f));
+
+    float x_src_f = ((float)i10_dst + pixel_offset) / sf0 - pixel_offset;
+    long x0_src = (long)floor(x_src_f);
+    long x1_src = x0_src + 1;
+
+    x0_src = max(0L, min(x0_src, (long)ne00_src - 1));
+    x1_src = max(0L, min(x1_src, (long)ne00_src - 1));
+
+    float dx = x_src_f - (float)x0_src;
+    dx = max(0.0f, min(dx, 1.0f));
+
+    global const float * p_a = (global const float *)(src_base + (ulong)x0_src * nb00 + (ulong)y0_src * nb01 + (ulong)i02_src * nb02 + (ulong)i03_src * nb03);
+    global const float * p_b = (global const float *)(src_base + (ulong)x1_src * nb00 + (ulong)y0_src * nb01 + (ulong)i02_src * nb02 + (ulong)i03_src * nb03);
+    global const float * p_c = (global const float *)(src_base + (ulong)x0_src * nb00 + (ulong)y1_src * nb01 + (ulong)i02_src * nb02 + (ulong)i03_src * nb03);
+    global const float * p_d = (global const float *)(src_base + (ulong)x1_src * nb00 + (ulong)y1_src * nb01 + (ulong)i02_src * nb02 + (ulong)i03_src * nb03);
+
+    const float val_a = *p_a;
+    const float val_b = *p_b;
+    const float val_c = *p_c;
+    const float val_d = *p_d;
+
+    float result = val_a * (1.0f - dx) * (1.0f - dy) +
+                   val_b * dx * (1.0f - dy) +
+                   val_c * (1.0f - dx) * dy +
+                   val_d * dx * dy;
+
+    dst_base[index] = result;
+}
diff --git a/ggml/src/ggml-quants.c b/ggml/src/ggml-quants.c
index 84ec6dfe3..9a7d1b22d 100644
--- a/ggml/src/ggml-quants.c
+++ b/ggml/src/ggml-quants.c
@@ -568,14 +568,14 @@ static float make_qkx2_quants(int n, int nmax, const float * GGML_RESTRICT x, co
     }
     float iscale = nmax/(max - min);
     float scale = 1/iscale;
-    float best_mad = 0;
+    float best_error = 0;
     for (int i = 0; i < n; ++i) {
         int l = nearest_int(iscale*(x[i] - min));
         L[i] = MAX(0, MIN(nmax, l));
         float diff = scale * L[i] + min - x[i];
         diff = use_mad ? fabsf(diff) : diff * diff;
         float w = weights[i];
-        best_mad += w * diff;
+        best_error += w * diff;
     }
     if (nstep < 1) {
         *the_min = -min;
@@ -601,18 +601,18 @@ static float make_qkx2_quants(int n, int nmax, const float * GGML_RESTRICT x, co
                 this_min = 0;
                 this_scale = sum_xl / sum_l2;
             }
-            float mad = 0;
+            float cur_error = 0;
             for (int i = 0; i < n; ++i) {
                 float diff = this_scale * Laux[i] + this_min - x[i];
                 diff = use_mad ? fabsf(diff) : diff * diff;
                 float w = weights[i];
-                mad += w * diff;
+                cur_error += w * diff;
             }
-            if (mad < best_mad) {
+            if (cur_error < best_error) {
                 for (int i = 0; i < n; ++i) {
                     L[i] = Laux[i];
                 }
-                best_mad = mad;
+                best_error = cur_error;
                 scale = this_scale;
                 min = this_min;
             }
@@ -2425,8 +2425,6 @@ void dequantize_row_iq1_m(const block_iq1_m * GGML_RESTRICT x, float * GGML_REST
     }
 }
 
-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 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k) {
     assert(k % QK4_NL == 0);
     const int64_t nb = k / QK4_NL;
diff --git a/ggml/src/ggml-rpc/ggml-rpc.cpp b/ggml/src/ggml-rpc/ggml-rpc.cpp
index 4f0abb5a6..f468f796d 100644
--- a/ggml/src/ggml-rpc/ggml-rpc.cpp
+++ b/ggml/src/ggml-rpc/ggml-rpc.cpp
@@ -53,6 +53,9 @@ struct socket_t {
     }
 };
 
+// macro for nicer error messages on server crash
+#define RPC_STATUS_ASSERT(x) if (!(x)) GGML_ABORT("Remote RPC server crashed or returned malformed response")
+
 // all RPC structures must be packed
 #pragma pack(push, 1)
 // ggml_tensor is serialized into rpc_tensor
@@ -425,7 +428,7 @@ static bool send_rpc_cmd(const std::shared_ptr<socket_t> & sock, enum rpc_cmd cm
 static bool check_server_version(const std::shared_ptr<socket_t> & sock) {
     rpc_msg_hello_rsp response;
     bool status = send_rpc_cmd(sock, RPC_CMD_HELLO, nullptr, 0, &response, sizeof(response));
-    GGML_ASSERT(status);
+    RPC_STATUS_ASSERT(status);
     if (response.major != RPC_PROTO_MAJOR_VERSION || response.minor > RPC_PROTO_MINOR_VERSION) {
         fprintf(stderr, "RPC server version mismatch: %d.%d.%d\n", response.major, response.minor, response.patch);
         return false;
@@ -481,7 +484,7 @@ static void ggml_backend_rpc_buffer_free_buffer(ggml_backend_buffer_t buffer) {
     ggml_backend_rpc_buffer_context * ctx = (ggml_backend_rpc_buffer_context *)buffer->context;
     rpc_msg_free_buffer_req request = {ctx->remote_ptr};
     bool status = send_rpc_cmd(ctx->sock, RPC_CMD_FREE_BUFFER, &request, sizeof(request), nullptr, 0);
-    GGML_ASSERT(status);
+    RPC_STATUS_ASSERT(status);
     delete ctx;
 }
 
@@ -493,7 +496,7 @@ static void * ggml_backend_rpc_buffer_get_base(ggml_backend_buffer_t buffer) {
     rpc_msg_buffer_get_base_req request = {ctx->remote_ptr};
     rpc_msg_buffer_get_base_rsp response;
     bool status = send_rpc_cmd(ctx->sock, RPC_CMD_BUFFER_GET_BASE, &request, sizeof(request), &response, sizeof(response));
-    GGML_ASSERT(status);
+    RPC_STATUS_ASSERT(status);
     ctx->base_ptr = reinterpret_cast<void *>(response.base_ptr);
     return ctx->base_ptr;
 }
@@ -545,7 +548,7 @@ static enum ggml_status ggml_backend_rpc_buffer_init_tensor(ggml_backend_buffer_
         request.tensor = serialize_tensor(tensor);
 
         bool status = send_rpc_cmd(ctx->sock, RPC_CMD_INIT_TENSOR, &request, sizeof(request), nullptr, 0);
-        GGML_ASSERT(status);
+        RPC_STATUS_ASSERT(status);
     }
     return GGML_STATUS_SUCCESS;
 }
@@ -560,7 +563,7 @@ static void ggml_backend_rpc_buffer_set_tensor(ggml_backend_buffer_t buffer, ggm
         request.hash = fnv_hash((const uint8_t*)data, size);
         rpc_msg_set_tensor_hash_rsp response;
         bool status = send_rpc_cmd(ctx->sock, RPC_CMD_SET_TENSOR_HASH, &request, sizeof(request), &response, sizeof(response));
-        GGML_ASSERT(status);
+        RPC_STATUS_ASSERT(status);
         if (response.result) {
             // the server has the same data, no need to send it
             return;
@@ -573,7 +576,7 @@ static void ggml_backend_rpc_buffer_set_tensor(ggml_backend_buffer_t buffer, ggm
     memcpy(input.data() + sizeof(rpc_tensor), &offset, sizeof(offset));
     memcpy(input.data() + sizeof(rpc_tensor) + sizeof(offset), data, size);
     bool status = send_rpc_cmd(ctx->sock, RPC_CMD_SET_TENSOR, input.data(), input.size());
-    GGML_ASSERT(status);
+    RPC_STATUS_ASSERT(status);
 }
 
 static void ggml_backend_rpc_buffer_get_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * tensor, void * data, size_t offset, size_t size) {
@@ -583,7 +586,7 @@ static void ggml_backend_rpc_buffer_get_tensor(ggml_backend_buffer_t buffer, con
     request.offset = offset;
     request.size = size;
     bool status = send_rpc_cmd(ctx->sock, RPC_CMD_GET_TENSOR, &request, sizeof(request), data, size);
-    GGML_ASSERT(status);
+    RPC_STATUS_ASSERT(status);
 }
 
 static bool ggml_backend_rpc_buffer_cpy_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * src, ggml_tensor * dst) {
@@ -601,7 +604,7 @@ static bool ggml_backend_rpc_buffer_cpy_tensor(ggml_backend_buffer_t buffer, con
     request.dst = serialize_tensor(dst);
     rpc_msg_copy_tensor_rsp response;
     bool status = send_rpc_cmd(ctx->sock, RPC_CMD_COPY_TENSOR, &request, sizeof(request), &response, sizeof(response));
-    GGML_ASSERT(status);
+    RPC_STATUS_ASSERT(status);
     return response.result;
 }
 
@@ -609,7 +612,7 @@ static void ggml_backend_rpc_buffer_clear(ggml_backend_buffer_t buffer, uint8_t
     ggml_backend_rpc_buffer_context * ctx = (ggml_backend_rpc_buffer_context *)buffer->context;
     rpc_msg_buffer_clear_req request = {ctx->remote_ptr, value};
     bool status = send_rpc_cmd(ctx->sock, RPC_CMD_BUFFER_CLEAR, &request, sizeof(request), nullptr, 0);
-    GGML_ASSERT(status);
+    RPC_STATUS_ASSERT(status);
 }
 
 static ggml_backend_buffer_i ggml_backend_rpc_buffer_interface = {
@@ -635,7 +638,7 @@ static ggml_backend_buffer_t ggml_backend_rpc_buffer_type_alloc_buffer(ggml_back
     rpc_msg_alloc_buffer_rsp response;
     auto sock = get_socket(buft_ctx->endpoint);
     bool status = send_rpc_cmd(sock, RPC_CMD_ALLOC_BUFFER, &request, sizeof(request), &response, sizeof(response));
-    GGML_ASSERT(status);
+    RPC_STATUS_ASSERT(status);
     if (response.remote_ptr != 0) {
         ggml_backend_buffer_t buffer = ggml_backend_buffer_init(buft,
             ggml_backend_rpc_buffer_interface,
@@ -650,7 +653,7 @@ static ggml_backend_buffer_t ggml_backend_rpc_buffer_type_alloc_buffer(ggml_back
 static size_t get_alignment(const std::shared_ptr<socket_t> & sock) {
     rpc_msg_get_alignment_rsp response;
     bool status = send_rpc_cmd(sock, RPC_CMD_GET_ALIGNMENT, nullptr, 0, &response, sizeof(response));
-    GGML_ASSERT(status);
+    RPC_STATUS_ASSERT(status);
     return response.alignment;
 }
 
@@ -662,7 +665,7 @@ static size_t ggml_backend_rpc_buffer_type_get_alignment(ggml_backend_buffer_typ
 static size_t get_max_size(const std::shared_ptr<socket_t> & sock) {
     rpc_msg_get_max_size_rsp response;
     bool status = send_rpc_cmd(sock, RPC_CMD_GET_MAX_SIZE, nullptr, 0, &response, sizeof(response));
-    GGML_ASSERT(status);
+    RPC_STATUS_ASSERT(status);
     return response.max_size;
 }
 
@@ -683,7 +686,7 @@ static size_t ggml_backend_rpc_buffer_type_get_alloc_size(ggml_backend_buffer_ty
 
         rpc_msg_get_alloc_size_rsp response;
         bool status = send_rpc_cmd(sock, RPC_CMD_GET_ALLOC_SIZE, &request, sizeof(request), &response, sizeof(response));
-        GGML_ASSERT(status);
+        RPC_STATUS_ASSERT(status);
 
         return response.alloc_size;
     } else {
@@ -761,7 +764,7 @@ static enum ggml_status ggml_backend_rpc_graph_compute(ggml_backend_t backend, g
     rpc_msg_graph_compute_rsp response;
     auto sock = get_socket(rpc_ctx->endpoint);
     bool status = send_rpc_cmd(sock, RPC_CMD_GRAPH_COMPUTE, input.data(), input.size(), &response, sizeof(response));
-    GGML_ASSERT(status);
+    RPC_STATUS_ASSERT(status);
     return (enum ggml_status)response.result;
 }
 
@@ -835,7 +838,7 @@ bool ggml_backend_is_rpc(ggml_backend_t backend) {
 static void get_device_memory(const std::shared_ptr<socket_t> & sock, size_t * free, size_t * total) {
     rpc_msg_get_device_memory_rsp response;
     bool status = send_rpc_cmd(sock, RPC_CMD_GET_DEVICE_MEMORY, nullptr, 0, &response, sizeof(response));
-    GGML_ASSERT(status);
+    RPC_STATUS_ASSERT(status);
     *free = response.free_mem;
     *total = response.total_mem;
 }
diff --git a/ggml/src/ggml-sycl/CMakeLists.txt b/ggml/src/ggml-sycl/CMakeLists.txt
index 2a0045bcc..efd78b912 100644
--- a/ggml/src/ggml-sycl/CMakeLists.txt
+++ b/ggml/src/ggml-sycl/CMakeLists.txt
@@ -142,7 +142,7 @@ else()
         FetchContent_Declare(
             ONEMATH
             GIT_REPOSITORY https://github.com/uxlfoundation/oneMath.git
-            GIT_TAG c255b1b4c41e2ee3059455c1f96a965d6a62568a
+            GIT_TAG 8efe85f5aaebb37f1d8c503b7af66315feabf142
         )
         FetchContent_MakeAvailable(ONEMATH)
         # Create alias to match with find_package targets name
diff --git a/ggml/src/ggml-sycl/binbcast.cpp b/ggml/src/ggml-sycl/binbcast.cpp
index 0a3883ae1..741630dba 100644
--- a/ggml/src/ggml-sycl/binbcast.cpp
+++ b/ggml/src/ggml-sycl/binbcast.cpp
@@ -225,9 +225,9 @@ struct bin_bcast_sycl {
                     dpct::has_capability_or_fail(stream->get_device(),
                                                  {sycl::aspect::fp16});
 
-                    stream->parallel_for(
-                        sycl::nd_range<3>(sycl::range<3>(1, 1, block_num) *
-                                              sycl::range<3>(1, 1, block_size),
+                    sycl_parallel_for(
+                        stream,
+                        sycl::nd_range<3>(sycl::range<3>(1, 1, block_num) * sycl::range<3>(1, 1, block_size),
                                           sycl::range<3>(1, 1, block_size)),
                         [=](sycl::nd_item<3> item_ct1) {
                             k_bin_bcast_unravel<bin_op>(
@@ -246,9 +246,8 @@ struct bin_bcast_sycl {
                 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_parallel_for(
+                    stream, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
                         k_bin_bcast<bin_op>(src0_dd, src1_dd, dst_dd, ne0, ne1,
                                             ne2, ne3, ne10, ne11, ne12, ne13,
                                             s1, s2, s3, s01, s02, s03, s11, s12, s13,
diff --git a/ggml/src/ggml-sycl/common.hpp b/ggml/src/ggml-sycl/common.hpp
index 15ee9dc69..4e7449d06 100644
--- a/ggml/src/ggml-sycl/common.hpp
+++ b/ggml/src/ggml-sycl/common.hpp
@@ -149,8 +149,6 @@ typedef sycl::float2 dfloat2;
 
 #define MMVQ_MAX_BATCH_SIZE  8
 
-static const int8_t kvalues_iq4nl[16]={-127, -104, -83, -65, -49, -35, -22, -10, 1, 13, 25, 38, 53, 69, 89, 113};
-
 static int g_all_sycl_device_count = -1;
 static bool g_ggml_backend_sycl_buffer_type_initialized = false;
 
@@ -201,7 +199,7 @@ struct sycl_device_info {
     // size_t  smpb;               // max. shared memory per block
     bool    vmm;                // virtual memory support
     size_t  total_vram;
-    sycl_hw_info hw_info;
+    //sycl_hw_info hw_info;     \\ device id and aarch, currently not used
     optimize_feature opt_feature;
 };
 
@@ -288,29 +286,6 @@ struct ggml_tensor_extra_gpu {
 
 void release_extra_gpu(ggml_tensor_extra_gpu * extra, std::vector<queue_ptr> 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;
@@ -515,9 +490,9 @@ constexpr size_t ceil_div(const size_t m, const size_t n) {
 
 bool gpu_has_xmx(sycl::device &dev);
 
-template <int N, class T> void debug_print_array(const std::string & prefix, const T array[N]) {
+template <int N, class T> std::string debug_get_array_str(const std::string & prefix, const T array[N]) {
     if (LIKELY(!g_ggml_sycl_debug)) {
-        return;
+        return "";
     }
     std::stringstream ss;
     ss << prefix << "=[";
@@ -528,29 +503,26 @@ template <int N, class T> void debug_print_array(const std::string & prefix, con
         ss << array[N - 1];
     }
     ss << "]";
-    GGML_SYCL_DEBUG("%s", ss.str().c_str());
+    return ss.str();
 }
 
-inline void debug_print_tensor(const std::string & prefix, const ggml_tensor * tensor,
-                               const std::string & suffix = "") {
-    if (LIKELY(!g_ggml_sycl_debug)) {
-        return;
-    }
-    GGML_SYCL_DEBUG("%s=", prefix.c_str());
+inline std::string debug_get_tensor_str(const std::string &prefix,
+        const ggml_tensor *tensor, const std::string &suffix = "") {
+    std::stringstream ss;
+    if (LIKELY(!g_ggml_sycl_debug)) { return ss.str(); }
+    ss << prefix.c_str() << "=";
     if (tensor) {
-        GGML_SYCL_DEBUG("'%s':type=%s", tensor->name, ggml_type_name(tensor->type));
-        debug_print_array<GGML_MAX_DIMS>(";ne", tensor->ne);
-        debug_print_array<GGML_MAX_DIMS>(";nb", tensor->nb);
-        if (!ggml_is_contiguous(tensor)) {
-            GGML_SYCL_DEBUG(";strided");
-        }
-        if (ggml_is_permuted(tensor)) {
-            GGML_SYCL_DEBUG(";permuted");
-        }
+        ss << "'" << tensor->name << "':type=" << ggml_type_name(tensor->type);
+        ss << debug_get_array_str<GGML_MAX_DIMS>(";ne", tensor->ne);
+        ss << debug_get_array_str<GGML_MAX_DIMS>(";nb", tensor->nb);
+
+        if (!ggml_is_contiguous(tensor)) { ss << ";strided"; }
+        if (ggml_is_permuted(tensor)) { ss << ";permuted"; }
     } else {
-        GGML_SYCL_DEBUG("nullptr");
+        ss << "nullptr";
     }
-    GGML_SYCL_DEBUG("%s", suffix.c_str());
+    ss << suffix;
+    return ss.str();
 }
 
 // Use scope_op_debug_print to log operations coming from running a model
@@ -566,10 +538,10 @@ struct scope_op_debug_print {
             return;
         }
         GGML_SYCL_DEBUG("[SYCL][OP] call %s%s:", func.data(), func_suffix.data());
-        debug_print_tensor(" dst", dst);
+        GGML_SYCL_DEBUG("%s", debug_get_tensor_str(" dst", dst).c_str());
         if (dst) {
             for (std::size_t i = 0; i < num_src; ++i) {
-                debug_print_tensor("\tsrc" + std::to_string(i), dst->src[i]);
+                GGML_SYCL_DEBUG("%s", debug_get_tensor_str("\tsrc" + std::to_string(i), dst->src[i]).c_str());
             }
         }
         GGML_SYCL_DEBUG("%s\n", suffix.data());
diff --git a/ggml/src/ggml-sycl/concat.cpp b/ggml/src/ggml-sycl/concat.cpp
index 7aa91c861..3501484a1 100644
--- a/ggml/src/ggml-sycl/concat.cpp
+++ b/ggml/src/ggml-sycl/concat.cpp
@@ -89,33 +89,24 @@ static void concat_f32_sycl(const float *x, const float *y, float *dst,
   sycl::range<3> gridDim(ne2, ne1, num_blocks);
   switch (dim) {
   case 0:
-    stream->parallel_for(
-        sycl::nd_range<3>(gridDim *
-                              sycl::range<3>(1, 1, SYCL_CONCAT_BLOCK_SIZE),
-                          sycl::range<3>(1, 1, SYCL_CONCAT_BLOCK_SIZE)),
-        [=](sycl::nd_item<3> item_ct1) {
-          concat_f32_dim0(x, y, dst, ne0, ne00, item_ct1);
-        });
-    break;
+      sycl_parallel_for(stream,
+                        sycl::nd_range<3>(gridDim * sycl::range<3>(1, 1, SYCL_CONCAT_BLOCK_SIZE),
+                                          sycl::range<3>(1, 1, SYCL_CONCAT_BLOCK_SIZE)),
+                        [=](sycl::nd_item<3> item_ct1) { concat_f32_dim0(x, y, dst, ne0, ne00, item_ct1); });
+      break;
   case 1:
-    stream->parallel_for(
-        sycl::nd_range<3>(gridDim *
-                              sycl::range<3>(1, 1, SYCL_CONCAT_BLOCK_SIZE),
-                          sycl::range<3>(1, 1, SYCL_CONCAT_BLOCK_SIZE)),
-        [=](sycl::nd_item<3> item_ct1) {
-          concat_f32_dim1(x, y, dst, ne0, ne01, item_ct1);
-        });
-    break;
+      sycl_parallel_for(stream,
+                        sycl::nd_range<3>(gridDim * sycl::range<3>(1, 1, SYCL_CONCAT_BLOCK_SIZE),
+                                          sycl::range<3>(1, 1, SYCL_CONCAT_BLOCK_SIZE)),
+                        [=](sycl::nd_item<3> item_ct1) { concat_f32_dim1(x, y, dst, ne0, ne01, item_ct1); });
+      break;
   // dim >=2 will be dispatched to the default path
   default:
-    stream->parallel_for(
-        sycl::nd_range<3>(gridDim *
-                              sycl::range<3>(1, 1, SYCL_CONCAT_BLOCK_SIZE),
-                          sycl::range<3>(1, 1, SYCL_CONCAT_BLOCK_SIZE)),
-        [=](sycl::nd_item<3> item_ct1) {
-          concat_f32_dim2(x, y, dst, ne0, ne02, item_ct1);
-        });
-    break;
+      sycl_parallel_for(stream,
+                        sycl::nd_range<3>(gridDim * sycl::range<3>(1, 1, SYCL_CONCAT_BLOCK_SIZE),
+                                          sycl::range<3>(1, 1, SYCL_CONCAT_BLOCK_SIZE)),
+                        [=](sycl::nd_item<3> item_ct1) { concat_f32_dim2(x, y, dst, ne0, ne02, item_ct1); });
+      break;
   }
 }
 
@@ -129,33 +120,29 @@ static void concat_f32_sycl_non_cont(
     int64_t ne2, int64_t ne3, uint64_t nb0, uint64_t nb1, uint64_t nb2,
     uint64_t nb3, int32_t dim) {
   sycl::range<3> gridDim(ne3, ne2, ne1);
-  stream->parallel_for(
-      sycl::nd_range<3>(gridDim, sycl::range<3>(1, 1, 1)),
-      [=](sycl::nd_item<3> item_ct1) {
-        int64_t i3 = item_ct1.get_group(0);
-        int64_t i2 = item_ct1.get_group(1);
-        int64_t i1 = item_ct1.get_group(2);
+  sycl_parallel_for(stream, sycl::nd_range<3>(gridDim, sycl::range<3>(1, 1, 1)), [=](sycl::nd_item<3> item_ct1) {
+      int64_t i3 = item_ct1.get_group(0);
+      int64_t i2 = item_ct1.get_group(1);
+      int64_t i1 = item_ct1.get_group(2);
 
-        int64_t o[4] = {0, 0, 0, 0};
-        o[dim] = dim == 0 ? ne00 : (dim == 1 ? ne01 : (dim == 2 ? ne02 : ne03));
+      int64_t o[4] = { 0, 0, 0, 0 };
+      o[dim]       = dim == 0 ? ne00 : (dim == 1 ? ne01 : (dim == 2 ? ne02 : ne03));
 
-        const float *x;
+      const float * x;
 
-        for (int i0 = item_ct1.get_local_id(2); i0 < ne0;
-             i0 += item_ct1.get_local_range(2)) {
+      for (int i0 = item_ct1.get_local_id(2); i0 < ne0; i0 += item_ct1.get_local_range(2)) {
           if (i0 < ne00 && i1 < ne01 && i2 < ne02 && i3 < ne03) {
-            x = (const float *)(src0 + (i3)*nb03 + (i2)*nb02 + (i1)*nb01 +
-                                (i0)*nb00);
+              x = (const float *) (src0 + (i3) *nb03 + (i2) *nb02 + (i1) *nb01 + (i0) *nb00);
           } else {
-            x = (const float *)(src1 + (i3 - o[3]) * nb13 + (i2 - o[2]) * nb12 +
-                                (i1 - o[1]) * nb11 + (i0 - o[0]) * nb10);
+              x = (const float *) (src1 + (i3 - o[3]) * nb13 + (i2 - o[2]) * nb12 + (i1 - o[1]) * nb11 +
+                                   (i0 - o[0]) * nb10);
           }
 
           float *y = (float *)(dst + i3 * nb3 + i2 * nb2 + i1 * nb1 + i0 * nb0);
 
           *y = *x;
-        }
-      });
+      }
+  });
 }
 
 void ggml_sycl_op_concat(ggml_backend_sycl_context & ctx, ggml_tensor *dst) {
diff --git a/ggml/src/ggml-sycl/conv.cpp b/ggml/src/ggml-sycl/conv.cpp
index 475bd34a2..c2f991e8d 100644
--- a/ggml/src/ggml-sycl/conv.cpp
+++ b/ggml/src/ggml-sycl/conv.cpp
@@ -59,16 +59,10 @@ static void conv_transpose_1d_f32_f32_sycl(
     const int num_blocks = (output_size + SYCL_CONV_TRANPOSE_1D_BLOCK_SIZE - 1) / SYCL_CONV_TRANPOSE_1D_BLOCK_SIZE;
     const sycl::range<3> block_dims(1, 1, SYCL_CONV_TRANPOSE_1D_BLOCK_SIZE);
     const sycl::range<3> block_nums(1, 1, num_blocks);
-    stream->parallel_for(
-        sycl::nd_range<3>(
-            block_nums * block_dims, block_dims),
-        [=](sycl::nd_item<3> item_ct1) {
-            conv_transpose_1d_kernel(
-                s0, output_size,
-                src0_ne0, src0_ne1, src0_ne2,
-                src1_ne0, dst_ne0,
-                src0, src1, dst, item_ct1);
-        });
+    sycl_parallel_for(stream, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
+        conv_transpose_1d_kernel(s0, output_size, src0_ne0, src0_ne1, src0_ne2, src1_ne0, dst_ne0, src0, src1, dst,
+                                 item_ct1);
+    });
 }
 
 void ggml_sycl_op_conv_transpose_1d(ggml_backend_sycl_context & ctx, ggml_tensor *dst) {
diff --git a/ggml/src/ggml-sycl/convert.cpp b/ggml/src/ggml-sycl/convert.cpp
index 75bac98e5..0ef567122 100644
--- a/ggml/src/ggml-sycl/convert.cpp
+++ b/ggml/src/ggml-sycl/convert.cpp
@@ -33,14 +33,11 @@ static void dequantize_block_sycl(const void *__restrict__ vx,
     {
         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_DEQUANTIZE_BLOCK_SIZE),
-                sycl::range<3>(1, 1, SYCL_DEQUANTIZE_BLOCK_SIZE)),
-            [=](sycl::nd_item<3> item_ct1) {
-                dequantize_block<qk, qr, dequantize_kernel>(vx, y, k, item_ct1);
-            });
+        sycl_parallel_for(
+            stream,
+            sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_DEQUANTIZE_BLOCK_SIZE),
+                              sycl::range<3>(1, 1, SYCL_DEQUANTIZE_BLOCK_SIZE)),
+            [=](sycl::nd_item<3> item_ct1) { dequantize_block<qk, qr, dequantize_kernel>(vx, y, k, item_ct1); });
     }
 }
 
@@ -53,24 +50,18 @@ static void dequantize_row_q2_K_sycl(const void *vx, dst_t *y, const int64_t k,
         dpct::has_capability_or_fail(stream->get_device(),
                                      {sycl::aspect::fp16});
 
-        stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
-                                                   sycl::range<3>(1, 1, 64),
-                                               sycl::range<3>(1, 1, 64)),
-                             [=](sycl::nd_item<3> item_ct1) {
-                                 dequantize_block_q2_K(vx, y, item_ct1);
-                             });
+        sycl_parallel_for(
+            stream, sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 64), sycl::range<3>(1, 1, 64)),
+            [=](sycl::nd_item<3> item_ct1) { dequantize_block_q2_K(vx, y, item_ct1); });
     }
 #else
     {
         dpct::has_capability_or_fail(stream->get_device(),
                                      {sycl::aspect::fp16});
 
-        stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
-                                                   sycl::range<3>(1, 1, 32),
-                                               sycl::range<3>(1, 1, 32)),
-                             [=](sycl::nd_item<3> item_ct1) {
-                                 dequantize_block_q2_K(vx, y, item_ct1);
-                             });
+        sycl_parallel_for(
+            stream, sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 32), sycl::range<3>(1, 1, 32)),
+            [=](sycl::nd_item<3> item_ct1) { dequantize_block_q2_K(vx, y, item_ct1); });
     }
 
 #endif
@@ -85,24 +76,18 @@ static void dequantize_row_q3_K_sycl(const void *vx, dst_t *y, const int64_t k,
         dpct::has_capability_or_fail(stream->get_device(),
                                      {sycl::aspect::fp16});
 
-        stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
-                                                   sycl::range<3>(1, 1, 64),
-                                               sycl::range<3>(1, 1, 64)),
-                             [=](sycl::nd_item<3> item_ct1) {
-                                 dequantize_block_q3_K(vx, y, item_ct1);
-                             });
+        sycl_parallel_for(
+            stream, sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 64), sycl::range<3>(1, 1, 64)),
+            [=](sycl::nd_item<3> item_ct1) { dequantize_block_q3_K(vx, y, item_ct1); });
     }
 #else
     {
         dpct::has_capability_or_fail(stream->get_device(),
                                      {sycl::aspect::fp16});
 
-        stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
-                                                   sycl::range<3>(1, 1, 32),
-                                               sycl::range<3>(1, 1, 32)),
-                             [=](sycl::nd_item<3> item_ct1) {
-                                 dequantize_block_q3_K(vx, y, item_ct1);
-                             });
+        sycl_parallel_for(
+            stream, sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 32), sycl::range<3>(1, 1, 32)),
+            [=](sycl::nd_item<3> item_ct1) { dequantize_block_q3_K(vx, y, item_ct1); });
     }
 #endif
 }
@@ -116,12 +101,9 @@ static void dequantize_row_q4_0_sycl(const void *vx, dst_t *y, const int64_t k,
         dpct::has_capability_or_fail(stream->get_device(),
                                      {sycl::aspect::fp16});
 
-        stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
-                                                   sycl::range<3>(1, 1, 32),
-                                               sycl::range<3>(1, 1, 32)),
-                             [=](sycl::nd_item<3> item_ct1) {
-                                 dequantize_block_q4_0(vx, y, nb32, item_ct1);
-                             });
+        sycl_parallel_for(
+            stream, sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 32), sycl::range<3>(1, 1, 32)),
+            [=](sycl::nd_item<3> item_ct1) { dequantize_block_q4_0(vx, y, nb32, item_ct1); });
     }
 }
 
@@ -135,13 +117,12 @@ static void dequantize_row_q4_0_sycl_reorder(const void *vx, dst_t *y, const int
     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);
-        });
-
+    sycl_parallel_for(stream,
+                      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 <typename dst_t>
@@ -153,12 +134,9 @@ static void dequantize_row_q4_1_sycl(const void *vx, dst_t *y, const int64_t k,
         dpct::has_capability_or_fail(stream->get_device(),
                                      {sycl::aspect::fp16});
 
-        stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
-                                                   sycl::range<3>(1, 1, 32),
-                                               sycl::range<3>(1, 1, 32)),
-                             [=](sycl::nd_item<3> item_ct1) {
-                                 dequantize_block_q4_1(vx, y, nb32, item_ct1);
-                             });
+        sycl_parallel_for(
+            stream, sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 32), sycl::range<3>(1, 1, 32)),
+            [=](sycl::nd_item<3> item_ct1) { dequantize_block_q4_1(vx, y, nb32, item_ct1); });
     }
 }
 
@@ -171,14 +149,13 @@ static void dequantize_row_q4_K_sycl(const void *vx, dst_t *y, const int64_t k,
         dpct::has_capability_or_fail(stream->get_device(),
                                      {sycl::aspect::fp16});
 
-        stream->submit([&](sycl::handler &cgh) {
+        sycl_launch(stream, [&](sycl::handler & cgh) {
             sycl::local_accessor<uint8_t, 1> scale_local_acc(sycl::range<1>(12), cgh);
-            cgh.parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
-                                                   sycl::range<3>(1, 1, 32),
-                                               sycl::range<3>(1, 1, 32)),
-                             [=](sycl::nd_item<3> item_ct1) {
-                                 dequantize_block_q4_K(vx, y, get_pointer(scale_local_acc), item_ct1);
-                             });
+            sycl_parallel_for(
+                cgh, sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 32), sycl::range<3>(1, 1, 32)),
+                [=](sycl::nd_item<3> item_ct1) {
+                    dequantize_block_q4_K(vx, y, get_pointer(scale_local_acc), item_ct1);
+                });
         });
     }
 }
@@ -191,13 +168,13 @@ static void dequantize_row_q4_K_sycl_reorder(const void * vx, dst_t * y, const i
 
     dpct::has_capability_or_fail(stream->get_device(), { sycl::aspect::fp16 });
 
-    stream->submit([&](sycl::handler & cgh) {
+    sycl_launch(stream, [&](sycl::handler & cgh) {
         sycl::local_accessor<uint8_t, 1> scale_local_acc(sycl::range<1>(12), cgh);
 
-        cgh.parallel_for(sycl::nd_range<1>(sycl::range<1>(global_size), sycl::range<1>(local_size)),
-                         [=](sycl::nd_item<1> item_ct1) {
-                             dequantize_block_q4_K_reorder(vx, y, get_pointer(scale_local_acc), item_ct1, nb);
-                         });
+        sycl_parallel_for<1>(cgh, sycl::nd_range<1>(sycl::range<1>(global_size), sycl::range<1>(local_size)),
+                             [=](sycl::nd_item<1> item_ct1) {
+                                 dequantize_block_q4_K_reorder(vx, y, get_pointer(scale_local_acc), item_ct1, nb);
+                             });
     });
 }
 
@@ -210,24 +187,18 @@ static void dequantize_row_q5_K_sycl(const void *vx, dst_t *y, const int64_t k,
         dpct::has_capability_or_fail(stream->get_device(),
                                      {sycl::aspect::fp16});
 
-        stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
-                                                   sycl::range<3>(1, 1, 64),
-                                               sycl::range<3>(1, 1, 64)),
-                             [=](sycl::nd_item<3> item_ct1) {
-                                 dequantize_block_q5_K(vx, y, item_ct1);
-                             });
+        sycl_parallel_for(
+            stream, sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 64), sycl::range<3>(1, 1, 64)),
+            [=](sycl::nd_item<3> item_ct1) { dequantize_block_q5_K(vx, y, item_ct1); });
     }
 #else
     {
         dpct::has_capability_or_fail(stream->get_device(),
                                      {sycl::aspect::fp16});
 
-        stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
-                                                   sycl::range<3>(1, 1, 32),
-                                               sycl::range<3>(1, 1, 32)),
-                             [=](sycl::nd_item<3> item_ct1) {
-                                 dequantize_block_q5_K(vx, y, item_ct1);
-                             });
+        sycl_parallel_for(
+            stream, sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 32), sycl::range<3>(1, 1, 32)),
+            [=](sycl::nd_item<3> item_ct1) { dequantize_block_q5_K(vx, y, item_ct1); });
     }
 
 #endif
@@ -242,29 +213,34 @@ static void dequantize_row_q6_K_sycl(const void *vx, dst_t *y, const int64_t k,
         dpct::has_capability_or_fail(stream->get_device(),
                                      {sycl::aspect::fp16});
 
-        stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
-                                                   sycl::range<3>(1, 1, 64),
-                                               sycl::range<3>(1, 1, 64)),
-                             [=](sycl::nd_item<3> item_ct1) {
-                                 dequantize_block_q6_K(vx, y, item_ct1);
-                             });
+        sycl_parallel_for(
+            stream, sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 64), sycl::range<3>(1, 1, 64)),
+            [=](sycl::nd_item<3> item_ct1) { dequantize_block_q6_K(vx, y, item_ct1); });
     }
 #else
     {
         dpct::has_capability_or_fail(stream->get_device(),
                                      {sycl::aspect::fp16});
 
-        stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
-                                                   sycl::range<3>(1, 1, 32),
-                                               sycl::range<3>(1, 1, 32)),
-                             [=](sycl::nd_item<3> item_ct1) {
-                                 dequantize_block_q6_K(vx, y, item_ct1);
-                             });
+        sycl_parallel_for(
+            stream, sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 32), sycl::range<3>(1, 1, 32)),
+            [=](sycl::nd_item<3> item_ct1) { dequantize_block_q6_K(vx, y, item_ct1); });
     }
 
 #endif
 }
 
+template <typename dst_t>
+static void dequantize_row_q6_K_sycl_reorder(const void * vx, dst_t * y, const int64_t k, dpct::queue_ptr stream) {
+    const int64_t nb = k / QK_K;
+
+    dpct::has_capability_or_fail(stream->get_device(), { sycl::aspect::fp16 });
+
+    sycl_parallel_for(stream,
+                      sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 64), sycl::range<3>(1, 1, 64)),
+                      [=](sycl::nd_item<3> item_ct1) { dequantize_block_q6_K_reorder(vx, y, item_ct1, nb); });
+}
+
 template <typename dst_t>
 static void dequantize_row_iq1_s_sycl(const void *vx, dst_t *y, const int64_t k,
                                         dpct::queue_ptr stream) {
@@ -273,15 +249,10 @@ static void dequantize_row_iq1_s_sycl(const void *vx, dst_t *y, const int64_t k,
         dpct::has_capability_or_fail(stream->get_device(),
                                      {sycl::aspect::fp16});
 
-        stream->submit([&](sycl::handler &cgh) {
-            cgh.parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
-                                                   sycl::range<3>(1, 1, 32),
-                                               sycl::range<3>(1, 1, 32)),
-                             [=](sycl::nd_item<3> item_ct1) {
-                                 dequantize_block_iq1_s(
-                                     vx, y, item_ct1, iq1s_grid_gpu
-                                     );
-                             });
+        sycl_launch(stream, [&](sycl::handler & cgh) {
+            sycl_parallel_for(
+                cgh, sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 32), sycl::range<3>(1, 1, 32)),
+                [=](sycl::nd_item<3> item_ct1) { dequantize_block_iq1_s(vx, y, item_ct1, iq1s_grid_gpu); });
         });
     }
 }
@@ -294,15 +265,10 @@ static void dequantize_row_iq1_m_sycl(const void *vx, dst_t *y, const int64_t k,
         dpct::has_capability_or_fail(stream->get_device(),
                                      {sycl::aspect::fp16});
 
-        stream->submit([&](sycl::handler &cgh) {
-            cgh.parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
-                                                   sycl::range<3>(1, 1, 32),
-                                               sycl::range<3>(1, 1, 32)),
-                             [=](sycl::nd_item<3> item_ct1) {
-                                 dequantize_block_iq1_m(
-                                     vx, y, item_ct1, iq1s_grid_gpu
-                                     );
-                             });
+        sycl_launch(stream, [&](sycl::handler & cgh) {
+            sycl_parallel_for(
+                cgh, sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 32), sycl::range<3>(1, 1, 32)),
+                [=](sycl::nd_item<3> item_ct1) { dequantize_block_iq1_m(vx, y, item_ct1, iq1s_grid_gpu); });
         });
     }
 }
@@ -315,15 +281,12 @@ static void dequantize_row_iq2_xxs_sycl(const void *vx, dst_t *y, const int64_t
         dpct::has_capability_or_fail(stream->get_device(),
                                      {sycl::aspect::fp16});
 
-        stream->submit([&](sycl::handler &cgh) {
-            cgh.parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
-                                                   sycl::range<3>(1, 1, 32),
-                                               sycl::range<3>(1, 1, 32)),
-                             [=](sycl::nd_item<3> item_ct1) {
-                                 dequantize_block_iq2_xxs(
-                                     vx, y, item_ct1, iq2xxs_grid,
-                                     ksigns_iq2xs, kmask_iq2xs);
-                             });
+        sycl_launch(stream, [&](sycl::handler & cgh) {
+            sycl_parallel_for(
+                cgh, sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 32), sycl::range<3>(1, 1, 32)),
+                [=](sycl::nd_item<3> item_ct1) {
+                    dequantize_block_iq2_xxs(vx, y, item_ct1, iq2xxs_grid, ksigns_iq2xs, kmask_iq2xs);
+                });
         });
     }
 }
@@ -336,15 +299,12 @@ static void dequantize_row_iq2_xs_sycl(const void *vx, dst_t *y, const int64_t k
         dpct::has_capability_or_fail(stream->get_device(),
                                      {sycl::aspect::fp16});
 
-        stream->submit([&](sycl::handler &cgh) {
-            cgh.parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
-                                                   sycl::range<3>(1, 1, 32),
-                                               sycl::range<3>(1, 1, 32)),
-                             [=](sycl::nd_item<3> item_ct1) {
-                                 dequantize_block_iq2_xs(
-                                     vx, y, item_ct1, iq2xs_grid,
-                                     ksigns_iq2xs, kmask_iq2xs);
-                             });
+        sycl_launch(stream, [&](sycl::handler & cgh) {
+            sycl_parallel_for(
+                cgh, sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 32), sycl::range<3>(1, 1, 32)),
+                [=](sycl::nd_item<3> item_ct1) {
+                    dequantize_block_iq2_xs(vx, y, item_ct1, iq2xs_grid, ksigns_iq2xs, kmask_iq2xs);
+                });
         });
     }
 }
@@ -357,13 +317,10 @@ static void dequantize_row_iq2_s_sycl(const void *vx, dst_t *y, const int64_t k,
         dpct::has_capability_or_fail(stream->get_device(),
                                      {sycl::aspect::fp16});
 
-        stream->submit([&](sycl::handler &cgh) {
-            cgh.parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
-                                                   sycl::range<3>(1, 1, 32),
-                                               sycl::range<3>(1, 1, 32)),
-                             [=](sycl::nd_item<3> item_ct1) {
-                                 dequantize_block_iq2_s(vx, y, item_ct1);
-                             });
+        sycl_launch(stream, [&](sycl::handler & cgh) {
+            sycl_parallel_for(
+                cgh, sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 32), sycl::range<3>(1, 1, 32)),
+                [=](sycl::nd_item<3> item_ct1) { dequantize_block_iq2_s(vx, y, item_ct1); });
         });
     }
 }
@@ -377,15 +334,12 @@ static void dequantize_row_iq3_xxs_sycl(const void *vx, dst_t *y, const int64_t
         dpct::has_capability_or_fail(stream->get_device(),
                                      {sycl::aspect::fp16});
 
-        stream->submit([&](sycl::handler &cgh) {
-            cgh.parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
-                                                   sycl::range<3>(1, 1, 32),
-                                               sycl::range<3>(1, 1, 32)),
-                             [=](sycl::nd_item<3> item_ct1) {
-                                 dequantize_block_iq3_xxs(
-                                     vx, y, item_ct1, iq3xxs_grid,
-                                     ksigns_iq2xs, kmask_iq2xs);
-                             });
+        sycl_launch(stream, [&](sycl::handler & cgh) {
+            sycl_parallel_for(
+                cgh, sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 32), sycl::range<3>(1, 1, 32)),
+                [=](sycl::nd_item<3> item_ct1) {
+                    dequantize_block_iq3_xxs(vx, y, item_ct1, iq3xxs_grid, ksigns_iq2xs, kmask_iq2xs);
+                });
         });
     }
 }
@@ -398,14 +352,10 @@ static void dequantize_row_iq3_s_sycl(const void *vx, dst_t *y, const int64_t k,
         dpct::has_capability_or_fail(stream->get_device(),
                                      {sycl::aspect::fp16});
 
-        stream->submit([&](sycl::handler &cgh) {
-            cgh.parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
-                                                   sycl::range<3>(1, 1, 32),
-                                               sycl::range<3>(1, 1, 32)),
-                             [=](sycl::nd_item<3> item_ct1) {
-                                 dequantize_block_iq3_s(
-                                     vx, y, item_ct1, kmask_iq2xs, iq3s_grid);
-                             });
+        sycl_launch(stream, [&](sycl::handler & cgh) {
+            sycl_parallel_for(
+                cgh, sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 32), sycl::range<3>(1, 1, 32)),
+                [=](sycl::nd_item<3> item_ct1) { dequantize_block_iq3_s(vx, y, item_ct1, kmask_iq2xs, iq3s_grid); });
         });
     }
 }
@@ -421,14 +371,11 @@ static void dequantize_row_iq4_xs_sycl(const void *vx, dst_t *y, const int64_t k
             dpct::has_capability_or_fail(stream->get_device(),
                                          {sycl::aspect::fp16});
 
-            stream->submit([&](sycl::handler &cgh) {
-                  cgh.parallel_for(
-                      sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
-                                            sycl::range<3>(1, 1, 32),
-                                        sycl::range<3>(1, 1, 32)),
-                      [=](sycl::nd_item<3> item_ct1) {
-                            dequantize_block_iq4_xs(vx, y, item_ct1);
-                      });
+            sycl_launch(stream, [&](sycl::handler & cgh) {
+                sycl_parallel_for(
+                    cgh,
+                    sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 32), sycl::range<3>(1, 1, 32)),
+                    [=](sycl::nd_item<3> item_ct1) { dequantize_block_iq4_xs(vx, y, item_ct1); });
             });
       }
 #endif
@@ -442,14 +389,11 @@ static void dequantize_row_iq4_nl_sycl(const void *vx, dst_t *y, const int64_t k
             dpct::has_capability_or_fail(stream->get_device(),
                                          {sycl::aspect::fp16});
 
-            stream->submit([&](sycl::handler &cgh) {
-                  cgh.parallel_for(
-                      sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
-                                            sycl::range<3>(1, 1, 32),
-                                        sycl::range<3>(1, 1, 32)),
-                      [=](sycl::nd_item<3> item_ct1) {
-                            dequantize_block_iq4_nl(vx, y, item_ct1);
-                      });
+            sycl_launch(stream, [&](sycl::handler & cgh) {
+                sycl_parallel_for(
+                    cgh,
+                    sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 32), sycl::range<3>(1, 1, 32)),
+                    [=](sycl::nd_item<3> item_ct1) { dequantize_block_iq4_nl(vx, y, item_ct1); });
             });
       }
 }
@@ -530,7 +474,11 @@ to_fp16_sycl_t ggml_get_to_fp16_sycl(ggml_type type, ggml_tensor * dst) {
         case GGML_TYPE_Q5_K:
             return dequantize_row_q5_K_sycl;
         case GGML_TYPE_Q6_K:
-            return dequantize_row_q6_K_sycl;
+            if (dst->src[0]->extra && ((ggml_tensor_extra_gpu *) dst->src[0]->extra)->optimized_feature.reorder) {
+                return dequantize_row_q6_K_sycl_reorder;
+            } else {
+                return dequantize_row_q6_K_sycl;
+            }
         case GGML_TYPE_IQ1_S:
             return dequantize_row_iq1_s_sycl;
         case GGML_TYPE_IQ1_M:
@@ -587,7 +535,11 @@ to_fp32_sycl_t ggml_get_to_fp32_sycl(ggml_type type, ggml_tensor *dst) {
         case GGML_TYPE_Q5_K:
             return dequantize_row_q5_K_sycl;
         case GGML_TYPE_Q6_K:
-            return dequantize_row_q6_K_sycl;
+            if (dst->src[0]->extra && ((ggml_tensor_extra_gpu *) dst->src[0]->extra)->optimized_feature.reorder) {
+                return dequantize_row_q6_K_sycl_reorder;
+            } else {
+                return dequantize_row_q6_K_sycl;
+            }
         case GGML_TYPE_IQ1_S:
             return dequantize_row_iq1_s_sycl;
         case GGML_TYPE_IQ1_M:
diff --git a/ggml/src/ggml-sycl/cpy.cpp b/ggml/src/ggml-sycl/cpy.cpp
index 44487c256..1ffd7f122 100644
--- a/ggml/src/ggml-sycl/cpy.cpp
+++ b/ggml/src/ggml-sycl/cpy.cpp
@@ -1,8 +1,12 @@
 #include "cpy.hpp"
 
 #include <float.h>
+#include <string>
 
 #include "dequantize.hpp"
+#include "ggml-sycl/common.hpp"
+#include "ggml-sycl/presets.hpp"
+#include "ggml.h"
 
 static __dpct_inline__ int best_index_int8(int n, const int8_t * val, float x) {
     if (x <= val[0]) {
@@ -116,6 +120,15 @@ static void cpy_blck_f32_q8_0(const char * cxi, char * cdsti) {
     }
 }
 
+/* quantized type same copy */
+template<typename T>
+static void cpy_blck_q_q(const char * cxi, char * cdsti) {
+    const T * xi = (const T *) cxi;
+    T * dsti = (T *) cdsti;
+    *dsti = *xi;
+}
+
+
 static void cpy_blck_q8_0_f32(const char * cxi, char * cdsti) {
     float * cdstf = (float *) (cdsti);
 
@@ -311,6 +324,34 @@ template <dequantize_kernel_t dequant, int qk> static void cpy_blck_q_f32(const
     }
 }
 
+
+template <typename T, int qk>
+static void cpy_q_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 / 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 / qk) * nb10 + i11 * nb11 + i12 * nb12 + i13 * nb13;
+
+    cpy_blck_q_q<T>(cx + x_offset, cdst + dst_offset);
+}
+
 template <cpy_kernel_t cpy_blck, int qk>
 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,
@@ -322,6 +363,7 @@ static void cpy_f32_q(const char * cx, char * cdst, const int ne, const int ne00
         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;
@@ -371,7 +413,8 @@ static void ggml_cpy_f16_f32_sycl(const char * cx, char * cdst, const int ne, co
     {
         dpct::has_capability_or_fail(stream->get_device(), { sycl::aspect::fp16 });
 
-        stream->parallel_for(
+        sycl_parallel_for(
+            stream,
             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) {
@@ -389,7 +432,8 @@ static void ggml_cpy_f32_f32_sycl(const char * cx, char * cdst, const int ne, co
     {
         dpct::has_capability_or_fail(stream->get_device(), { sycl::aspect::fp16 });
 
-        stream->parallel_for(
+        sycl_parallel_for(
+            stream,
             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) {
@@ -407,7 +451,8 @@ static void ggml_cpy_f32_f16_sycl(const char * cx, char * cdst, const int ne, co
     {
         dpct::has_capability_or_fail(stream->get_device(), { sycl::aspect::fp16 });
 
-        stream->parallel_for(
+        sycl_parallel_for(
+            stream,
             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) {
@@ -423,11 +468,11 @@ static void ggml_cpy_f32_q8_0_sycl(const char * cx, char * cdst, const int ne, c
                                    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<cpy_blck_f32_q8_0, QK8_0>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03,
-                                                                 ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
-                         });
+    sycl_parallel_for(stream, 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<cpy_blck_f32_q8_0, QK8_0>(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,
@@ -435,11 +480,11 @@ static void ggml_cpy_q8_0_f32_sycl(const char * cx, char * cdst, const int ne, c
                                    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<cpy_blck_q8_0_f32, QK8_0>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03,
-                                                                 ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
-                         });
+    sycl_parallel_for(stream, 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<cpy_blck_q8_0_f32, QK8_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_0_sycl(const char * cx, char * cdst, const int ne, const int ne00, const int ne01,
@@ -448,11 +493,11 @@ static void ggml_cpy_f32_q4_0_sycl(const char * cx, char * cdst, const int ne, c
                                    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<cpy_blck_f32_q4_0, QK4_0>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03,
-                                                                 ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
-                         });
+    sycl_parallel_for(stream, 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<cpy_blck_f32_q4_0, 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_q4_0_f32_sycl(const char * cx, char * cdst, const int ne, const int ne00, const int ne01,
@@ -460,8 +505,9 @@ static void ggml_cpy_q4_0_f32_sycl(const char * cx, char * cdst, const int ne, c
                                    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) {
+    sycl_parallel_for(
+        stream, 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<cpy_blck_q_f32<dequantize_q4_0, QK4_0>, QK4_0>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02,
                                                                      nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13,
                                                                      item_ct1);
@@ -474,11 +520,11 @@ static void ggml_cpy_f32_q4_1_sycl(const char * cx, char * cdst, const int ne, c
                                    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<cpy_blck_f32_q4_1, QK4_1>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03,
-                                                                 ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
-                         });
+    sycl_parallel_for(stream, 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<cpy_blck_f32_q4_1, 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_q4_1_f32_sycl(const char * cx, char * cdst, const int ne, const int ne00, const int ne01,
@@ -486,8 +532,9 @@ static void ggml_cpy_q4_1_f32_sycl(const char * cx, char * cdst, const int ne, c
                                    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) {
+    sycl_parallel_for(
+        stream, 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<cpy_blck_q_f32<dequantize_q4_1, QK4_1>, QK4_1>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02,
                                                                      nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13,
                                                                      item_ct1);
@@ -500,11 +547,11 @@ static void ggml_cpy_f32_q5_0_sycl(const char * cx, char * cdst, const int ne, c
                                    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<cpy_blck_f32_q5_0, QK5_0>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03,
-                                                                 ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
-                         });
+    sycl_parallel_for(stream, 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<cpy_blck_f32_q5_0, 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_q5_0_f32_sycl(const char * cx, char * cdst, const int ne, const int ne00, const int ne01,
@@ -512,8 +559,9 @@ static void ggml_cpy_q5_0_f32_sycl(const char * cx, char * cdst, const int ne, c
                                    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) {
+    sycl_parallel_for(
+        stream, 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<cpy_blck_q_f32<dequantize_q5_0, QK5_0>, QK5_0>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02,
                                                                      nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13,
                                                                      item_ct1);
@@ -526,11 +574,11 @@ static void ggml_cpy_f32_q5_1_sycl(const char * cx, char * cdst, const int ne, c
                                    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<cpy_blck_f32_q5_1, QK5_1>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03,
-                                                                 ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
-                         });
+    sycl_parallel_for(stream, 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<cpy_blck_f32_q5_1, 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_q5_1_f32_sycl(const char * cx, char * cdst, const int ne, const int ne00, const int ne01,
@@ -538,8 +586,9 @@ static void ggml_cpy_q5_1_f32_sycl(const char * cx, char * cdst, const int ne, c
                                    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) {
+    sycl_parallel_for(
+        stream, 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<cpy_blck_q_f32<dequantize_q5_1, QK5_1>, QK5_1>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02,
                                                                      nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13,
                                                                      item_ct1);
@@ -552,11 +601,11 @@ static void ggml_cpy_f32_iq4_nl_sycl(const char * cx, char * cdst, const int ne,
                                      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<cpy_blck_f32_iq4_nl, QK4_NL>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11,
-                                                   ne12, nb10, nb11, nb12, nb13, item_ct1);
-        });
+    sycl_parallel_for(stream, 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<cpy_blck_f32_iq4_nl, QK4_NL>(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,
@@ -567,7 +616,8 @@ static void ggml_cpy_f16_f16_sycl(const char * cx, char * cdst, const int ne, co
     {
         dpct::has_capability_or_fail(stream->get_device(), { sycl::aspect::fp16 });
 
-        stream->parallel_for(
+        sycl_parallel_for(
+            stream,
             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) {
@@ -586,7 +636,8 @@ static void ggml_cpy_i16_i16_sycl(const char * cx, char * cdst, const int ne, co
         // dpct::has_capability_or_fail(stream->get_device(),
         //                              {sycl::aspect::fp16});
 
-        stream->parallel_for(
+        sycl_parallel_for(
+            stream,
             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) {
@@ -605,7 +656,8 @@ static void ggml_cpy_i32_i32_sycl(const char * cx, char * cdst, const int ne, co
         // dpct::has_capability_or_fail(stream->get_device(),
         //                              {sycl::aspect::fp16});
 
-        stream->parallel_for(
+        sycl_parallel_for(
+            stream,
             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) {
@@ -615,10 +667,85 @@ static void ggml_cpy_i32_i32_sycl(const char * cx, char * cdst, const int ne, co
     }
 }
 
+static void ggml_cpy_q8_0_q8_0(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 = ceil_div(ne, SYCL_CPY_BLOCK_SIZE);
+    sycl_parallel_for(stream,
+                      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_q_q<block_q8_0, QK8_0>(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_q5_0(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 = ceil_div(ne, SYCL_CPY_BLOCK_SIZE);
+    sycl_parallel_for(stream,
+                      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_q_q<block_q5_0, 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_q5_1_q5_1(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 = ceil_div(ne, SYCL_CPY_BLOCK_SIZE);
+
+    sycl_parallel_for(stream,
+                      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_q_q<block_q5_1, 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_q4_0_q4_0(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 = ceil_div(ne, SYCL_CPY_BLOCK_SIZE);
+    sycl_parallel_for(stream,
+                      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_q_q<block_q4_0, 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_q4_1_q4_1(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 = ceil_div(ne, SYCL_CPY_BLOCK_SIZE);
+   sycl_parallel_for(stream,
+                     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_q_q<block_q4_1, QK4_1>(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 {
     // Unlike other operators ggml_sycl_cpy takes 2 distinct tensors instead of a dst ggml_tensor and rely on its src field
-    scope_op_debug_print scope_dbg_print(__func__, src1, /*num_src=*/0,
-                                         std::string(" src0 type=") + ggml_type_name(src0->type));
+    scope_op_debug_print scope_dbg_print(__func__, src1, /*num_src=*/0, debug_get_tensor_str("\tsrc0", src0));
     const int64_t ne = ggml_nelements(src0);
     GGML_ASSERT(ne == ggml_nelements(src1));
 
@@ -632,8 +759,10 @@ void ggml_sycl_cpy(ggml_backend_sycl_context & ctx, const ggml_tensor * src0, co
 
     char * src0_ddc = (char *) src0->data;
     char * src1_ddc = (char *) src1->data;
-
-    if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32) {
+    if ((src0->type == src1->type) && (ggml_is_contiguous(src0) && ggml_is_contiguous(src1))) {
+        GGML_SYCL_DEBUG("%s: memcpy path\n", __func__);
+        main_stream->memcpy(src1_ddc, src0_ddc, ggml_nbytes(src0));
+    } else 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) {
@@ -684,6 +813,16 @@ void ggml_sycl_cpy(ggml_backend_sycl_context & ctx, const ggml_tensor * src0, co
     } 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 if (src0->type == GGML_TYPE_Q8_0 && src1->type == GGML_TYPE_Q8_0) {
+        ggml_cpy_q8_0_q8_0(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_Q5_0) {
+        ggml_cpy_q5_0_q5_0(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_Q5_1) {
+        ggml_cpy_q5_1_q5_1(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_Q4_0) {
+        ggml_cpy_q4_0_q4_0(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_Q4_1) {
+        ggml_cpy_q4_1_q4_1(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));
diff --git a/ggml/src/ggml-sycl/dequantize.hpp b/ggml/src/ggml-sycl/dequantize.hpp
index 64e92f73f..540539bb2 100644
--- a/ggml/src/ggml-sycl/dequantize.hpp
+++ b/ggml/src/ggml-sycl/dequantize.hpp
@@ -538,6 +538,38 @@ static void dequantize_block_q6_K(const void * __restrict__ vx, dst_t * __restri
 #endif
 }
 
+template <typename dst_t>
+static void dequantize_block_q6_K_reorder(const void * __restrict__ vx, dst_t * __restrict__ yy,
+                                          const sycl::nd_item<3> & item_ct1, int64_t n_blocks) {
+    const int64_t ib = item_ct1.get_group(2);
+
+    const int64_t tid = item_ct1.get_local_id(2);
+    const int64_t ip  = tid / 32;       // ip is 0 or 1
+    const int64_t il  = tid - 32 * ip;  // 0...32
+    const int64_t is  = 8 * ip + il / 16;
+
+    const uint8_t *   base_ptr           = static_cast<const uint8_t *>(vx);
+    const auto        ql_offset          = ib * (QK_K / 2);
+    const auto        qh_offset          = (QK_K / 2) * n_blocks + (QK_K / 4) * ib;
+    const auto        base_scales_offset = (QK_K / 2) * n_blocks + (QK_K / 4) * n_blocks + (QK_K / 16) * ib;
+    const auto        base_d_offset      = ((QK_K / 2) + (QK_K / 4) + (QK_K / 16)) * n_blocks;
+    const uint8_t *   ql_ptr             = base_ptr + ql_offset;
+    const uint8_t *   qh_ptr             = base_ptr + qh_offset;
+    const uint8_t *   scales_ptr         = base_ptr + base_scales_offset;
+    const ggml_half * d                  = (const ggml_half *) (base_ptr + base_d_offset) + ib;
+
+    dst_t * y = yy + ib * QK_K + 128 * ip + il;
+
+    const uint8_t * ql = ql_ptr + 64 * ip + il;
+    const uint8_t   qh = *(qh_ptr + 32 * ip + il);
+    const int8_t *  sc = reinterpret_cast<const int8_t *>(scales_ptr + is);
+
+    y[0]  = *d * sc[0] * ((int8_t) ((ql[0] & 0xF) | (((qh >> 0) & 3) << 4)) - 32);
+    y[32] = *d * sc[2] * ((int8_t) ((ql[32] & 0xF) | (((qh >> 2) & 3) << 4)) - 32);
+    y[64] = *d * sc[4] * ((int8_t) ((ql[0] >> 4) | (((qh >> 4) & 3) << 4)) - 32);
+    y[96] = *d * sc[6] * ((int8_t) ((ql[32] >> 4) | (((qh >> 6) & 3) << 4)) - 32);
+}
+
 template<typename dst_t>
 static void dequantize_block_iq2_xxs(const void * __restrict__ vx, dst_t * __restrict__ yy,
                                      const sycl::nd_item<3> &item_ct1,
diff --git a/ggml/src/ggml-sycl/dmmv.cpp b/ggml/src/ggml-sycl/dmmv.cpp
index 4f2760110..70579c0c3 100644
--- a/ggml/src/ggml-sycl/dmmv.cpp
+++ b/ggml/src/ggml-sycl/dmmv.cpp
@@ -208,12 +208,10 @@ static void convert_mul_mat_vec_f16_sycl(const void *vx, const dfloat *y,
         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<1, 1, convert_f16>(vx, y, dst, ncols,
-                                                          nrows, item_ct1);
-            });
+        sycl_parallel_for(stream, 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<1, 1, convert_f16>(vx, y, dst, ncols, nrows, item_ct1);
+                          });
     }
 }
 
@@ -877,12 +875,11 @@ static void dequantize_mul_mat_vec_q4_0_sycl_reorder(const void *vx, const dfloa
         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<QK4_0, QR4_0, dequantize_q4_0_reorder>(
-                    vx, y, dst, ncols, nrows, item_ct1);
-            });
+        sycl_parallel_for(stream, 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<QK4_0, QR4_0, dequantize_q4_0_reorder>(vx, y, dst, ncols,
+                                                                                                    nrows, item_ct1);
+                          });
     }
 }
 
@@ -900,12 +897,10 @@ static void dequantize_mul_mat_vec_q4_0_sycl(const void *vx, const dfloat *y,
         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<QK4_0, QR4_0, dequantize_q4_0>(
-                    vx, y, dst, ncols, nrows, item_ct1);
-            });
+        sycl_parallel_for(stream, 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<QK4_0, QR4_0, dequantize_q4_0>(vx, y, dst, ncols, nrows, item_ct1);
+                          });
     }
 }
 
@@ -921,12 +916,10 @@ static void dequantize_mul_mat_vec_q4_1_sycl(const void *vx, const dfloat *y,
         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<QK4_1, QR4_1, dequantize_q4_1>(
-                    vx, y, dst, ncols, nrows, item_ct1);
-            });
+        sycl_parallel_for(stream, 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<QK4_1, QR4_1, dequantize_q4_1>(vx, y, dst, ncols, nrows, item_ct1);
+                          });
     }
 }
 
@@ -942,12 +935,10 @@ static void dequantize_mul_mat_vec_q5_0_sycl(const void *vx, const dfloat *y,
         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<QK5_0, QR5_0, dequantize_q5_0>(
-                    vx, y, dst, ncols, nrows, item_ct1);
-            });
+        sycl_parallel_for(stream, 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<QK5_0, QR5_0, dequantize_q5_0>(vx, y, dst, ncols, nrows, item_ct1);
+                          });
     }
 }
 
@@ -963,12 +954,10 @@ static void dequantize_mul_mat_vec_q5_1_sycl(const void *vx, const dfloat *y,
         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<QK5_1, QR5_1, dequantize_q5_1>(
-                    vx, y, dst, ncols, nrows, item_ct1);
-            });
+        sycl_parallel_for(stream, 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<QK5_1, QR5_1, dequantize_q5_1>(vx, y, dst, ncols, nrows, item_ct1);
+                          });
     }
 }
 
@@ -984,12 +973,10 @@ static void dequantize_mul_mat_vec_q8_0_sycl(const void *vx, const dfloat *y,
         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<QK8_0, QR8_0, dequantize_q8_0>(
-                    vx, y, dst, ncols, nrows, item_ct1);
-            });
+        sycl_parallel_for(stream, 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<QK8_0, QR8_0, dequantize_q8_0>(vx, y, dst, ncols, nrows, item_ct1);
+                          });
     }
 }
 
@@ -1002,11 +989,10 @@ static void dequantize_mul_mat_vec_q2_K_sycl(const void *vx, const float *y,
     const int block_num_y = (nrows + ny - 1) / ny;
     const sycl::range<3> block_nums(1, 1, block_num_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) [[sycl::reqd_sub_group_size(QK_WARP_SIZE)]] {
-            dequantize_mul_mat_vec_q2_k(vx, y, dst, ncols, nrows, item_ct1);
-        });
+    sycl_parallel_for(stream, sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                      [=](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);
+                      });
 }
 
 static void dequantize_mul_mat_vec_q3_K_sycl(const void *vx, const float *y,
@@ -1018,11 +1004,10 @@ static void dequantize_mul_mat_vec_q3_K_sycl(const void *vx, const float *y,
     const int block_num_y = (nrows + ny - 1) / ny;
     const sycl::range<3> block_nums(1, 1, block_num_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) [[sycl::reqd_sub_group_size(QK_WARP_SIZE)]] {
-            dequantize_mul_mat_vec_q3_k(vx, y, dst, ncols, nrows, item_ct1);
-        });
+    sycl_parallel_for(stream, sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                      [=](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);
+                      });
 }
 
 static void dequantize_mul_mat_vec_q4_K_sycl(const void *vx, const float *y,
@@ -1034,11 +1019,10 @@ static void dequantize_mul_mat_vec_q4_K_sycl(const void *vx, const float *y,
     const int block_num_y = (nrows + ny - 1) / ny;
     const sycl::range<3> block_nums(1, 1, block_num_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) [[sycl::reqd_sub_group_size(QK_WARP_SIZE)]] {
-            dequantize_mul_mat_vec_q4_k(vx, y, dst, ncols, nrows, item_ct1);
-        });
+    sycl_parallel_for(stream, sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                      [=](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);
+                      });
 }
 
 static void dequantize_mul_mat_vec_q5_K_sycl(const void *vx, const float *y,
@@ -1047,11 +1031,10 @@ static void dequantize_mul_mat_vec_q5_K_sycl(const void *vx, const float *y,
                                              dpct::queue_ptr stream) {
     GGML_ASSERT(ncols % QK_K == 0);
     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) [[sycl::reqd_sub_group_size(QK_WARP_SIZE)]] {
-            dequantize_mul_mat_vec_q5_k(vx, y, dst, ncols, item_ct1);
-        });
+    sycl_parallel_for(stream, 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(QK_WARP_SIZE)]] {
+                          dequantize_mul_mat_vec_q5_k(vx, y, dst, ncols, item_ct1);
+                      });
 }
 
 static void dequantize_mul_mat_vec_q6_K_sycl(const void *vx, const float *y,
@@ -1063,11 +1046,10 @@ static void dequantize_mul_mat_vec_q6_K_sycl(const void *vx, const float *y,
     const int block_num_y = (nrows + ny - 1) / ny;
     const sycl::range<3> block_nums(1, 1, block_num_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) [[sycl::reqd_sub_group_size(QK_WARP_SIZE)]] {
-            dequantize_mul_mat_vec_q6_k(vx, y, dst, ncols, nrows, item_ct1);
-        });
+    sycl_parallel_for(stream, sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                      [=](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);
+                      });
 }
 
 void ggml_sycl_op_dequantize_mul_mat_vec(
diff --git a/ggml/src/ggml-sycl/dpct/helper.hpp b/ggml/src/ggml-sycl/dpct/helper.hpp
index d538965b0..27c727860 100644
--- a/ggml/src/ggml-sycl/dpct/helper.hpp
+++ b/ggml/src/ggml-sycl/dpct/helper.hpp
@@ -13,10 +13,10 @@
 #ifndef GGML_SYCL_DPCT_HELPER_HPP
 #define GGML_SYCL_DPCT_HELPER_HPP
 
+#include <map>
 #include <sycl/sycl.hpp>
 #include <sycl/half_type.hpp>
 #include <syclcompat/math.hpp>
-#include <map>
 
 #ifdef GGML_SYCL_USE_INTEL_ONEMKL
 #include <oneapi/mkl.hpp>
@@ -118,6 +118,36 @@ inline auto get_onemath_backend(sycl::queue& queue)
 #endif
 }
 
+#ifdef SYCL_EXT_ONEAPI_ENQUEUE_FUNCTIONS
+    namespace syclex = sycl::ext::oneapi::experimental;
+#endif
+
+template <int NR, typename Func>
+__dpct_inline__ void sycl_parallel_for(sycl::handler & cgh, sycl::nd_range<NR> nd_range, Func && func) {
+#ifdef SYCL_EXT_ONEAPI_ENQUEUE_FUNCTIONS
+    syclex::nd_launch(cgh, nd_range, func);
+#else
+    cgh.parallel_for(nd_range, func);
+#endif
+}
+
+template <int NR, typename Func>
+__dpct_inline__ void sycl_parallel_for(sycl::queue * q, sycl::nd_range<NR> nd_range, Func && func) {
+#ifdef SYCL_EXT_ONEAPI_ENQUEUE_FUNCTIONS
+    syclex::nd_launch(*q, nd_range, func);
+#else
+    q->parallel_for(nd_range, func);
+#endif
+}
+
+template <typename Func> __dpct_inline__ void sycl_launch(sycl::queue * stream, Func && func) {
+#ifdef SYCL_EXT_ONEAPI_ENQUEUE_FUNCTIONS
+    syclex::submit(*stream, func);
+#else
+    stream->submit(func);
+#endif
+}
+
 namespace dpct
 {
     typedef sycl::queue *queue_ptr;
diff --git a/ggml/src/ggml-sycl/element_wise.cpp b/ggml/src/ggml-sycl/element_wise.cpp
index 5b7c4f0b4..0363b06a3 100644
--- a/ggml/src/ggml-sycl/element_wise.cpp
+++ b/ggml/src/ggml-sycl/element_wise.cpp
@@ -1,12 +1,19 @@
 #include "common.hpp"
+#include "ggml-sycl/presets.hpp"
 #include "ggml.h"
 #include "element_wise.hpp"
 
+#define SYCL_GLOBAL_ID_LOOP(K, ITEM) \
+    for (auto i = ITEM.get_global_id(0); i < (size_t)K; i += ITEM.get_global_range(0))
+
+#define SYCL_LOCAL_ID_CALC(ITEM, IDX) \
+    (ITEM.get_local_range(IDX) * ITEM.get_group(IDX) + ITEM.get_local_id(IDX))
+
+
 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) +
-                  item_ct1.get_local_id(2);
+    const int nb1, const int nb2, int offset, const sycl::nd_item<1> &item_ct1) {
+    const int i = SYCL_LOCAL_ID_CALC(item_ct1, 0);
     if (i >= ne) {
         return;
     }
@@ -21,248 +28,280 @@ static void acc_f32(const float * x, const float * y, float * dst, const int ne,
     }
 }
 
+/* Unary OP funcs */
 template<typename T>
-static void sgn(const T * x, T * dst, const int k, const sycl::nd_item<3> &item_ct1) {
-    for(auto i = item_ct1.get_global_id(2); i < (const size_t)k; i += item_ct1.get_global_range(2)) {
-        dst[i] = x[i] > static_cast<T>(0.f) ? static_cast<T>(1.f) : ((x[i] < static_cast<T>(0.f) ? static_cast<T>(-1.f) : static_cast<T>(0.f)));
-    }
+static __dpct_inline__ T op_sgn(T x) {
+    return x > static_cast<T>(0.f) ? static_cast<T>(1.f) : ((x < static_cast<T>(0.f) ? static_cast<T>(-1.f) : static_cast<T>(0.f)));
 }
 
 template<typename T>
-static void abs_op(const T * x, T * dst, const int k, const sycl::nd_item<3> &item_ct1) {
-    for(auto i = item_ct1.get_global_id(2); i < (const size_t)k; i += item_ct1.get_global_range(2)) {
-        dst[i] = sycl::fabs(x[i]);
-    }
+static __dpct_inline__ T op_abs(T x) {
+    return sycl::fabs(x);
 }
 
 template<typename T>
-static void elu_op(const T * x, T * dst, const int k, const sycl::nd_item<3> &item_ct1) {
-    for(auto i = item_ct1.get_global_id(2); i < (const size_t)k; i += item_ct1.get_global_range(2)) {
-        dst[i] = (x[i] > static_cast<T>(0.f)) ? x[i] : sycl::expm1(x[i]);
-    }
+static __dpct_inline__ T op_elu(T x) {
+    return (x > static_cast<T>(0.f)) ? x : sycl::expm1(x);
 }
 
 template<typename T>
-static void gelu(const T * x, T * dst, const int k,
-                     const sycl::nd_item<3> &item_ct1) {
+static __dpct_inline__ T op_gelu(T x) {
     const T GELU_COEF_A    = static_cast<T>(0.044715f);
     const T SQRT_2_OVER_PI = static_cast<T>(0.79788456080286535587989211986876f);
-    const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) +
-                  item_ct1.get_local_id(2);
-
-    if (i >= k) {
-        return;
-    }
-
-    float xi = x[i];
-    dst[i] = static_cast<T>(0.5f) * xi *
-             (static_cast<T>(1.0f) +
-              sycl::tanh(SQRT_2_OVER_PI * xi * (static_cast<T>(1.0f) + GELU_COEF_A * xi * xi)));
+    return static_cast<T>(0.5f) * x *
+           (static_cast<T>(1.0f) +
+            sycl::tanh(SQRT_2_OVER_PI * x * (static_cast<T>(1.0f) + GELU_COEF_A * x * x)));
 }
 
 template<typename T>
-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);
-
-    if (i >= k) {
-        return;
-    }
-    dst[i] = x[i] / (static_cast<T>(1.0f) + sycl::native::exp(-x[i]));
+static __dpct_inline__ T op_silu(T x) {
+    return x / (static_cast<T>(1.0f) + sycl::native::exp(-x));
 }
 
 template<typename T>
-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) +
-                  item_ct1.get_local_id(2);
-    if (i >= k) {
-        return;
-    }
-    dst[i] = x[i] * (static_cast<T>(1.0f) / (static_cast<T>(1.0f) + sycl::native::exp(GELU_QUICK_COEF * x[i])));
+static __dpct_inline__ T op_gelu_quick(T x) {
+    const T GELU_QUICK_COEF_LOCAL = static_cast<T>(-1.702f);
+    return x * (static_cast<T>(1.0f) / (static_cast<T>(1.0f) + sycl::native::exp(GELU_QUICK_COEF_LOCAL * x)));
 }
 
 template<typename T>
-static void gelu_erf(const T * x, T * dst, const int k, const sycl::nd_item<3> &item_ct1) {
+static __dpct_inline__ T op_gelu_erf(T x) {
     const T SQRT_2_INV = static_cast<T>(0.70710678118654752440084436210484f);
-    for(auto i = item_ct1.get_global_id(2); i < (const size_t)k; i += item_ct1.get_global_range(2)) {
-       auto x_i = x[i];
-        dst[i] = static_cast<T>(0.5f) * x_i * (static_cast<T>(1.0f) + sycl::erf(x_i * SQRT_2_INV));
+    return static_cast<T>(0.5f) * x * (static_cast<T>(1.0f) + sycl::erf(x * SQRT_2_INV));
+}
+
+template<typename T>
+static __dpct_inline__ T op_tanh(T x) {
+    return sycl::tanh(x);
+}
+
+template<typename T>
+static __dpct_inline__ T op_relu(T x) {
+    return sycl::fmax(x, static_cast<T>(0));
+}
+
+template<typename T>
+static __dpct_inline__ T op_sigmoid(T x) {
+    return static_cast<T>(1.0f) / (static_cast<T>(1.0f) + sycl::native::exp(-x));
+}
+
+template<typename T>
+static __dpct_inline__ T op_sqrt(T x) {
+    return sycl::sqrt(x);
+}
+
+template<typename T>
+static __dpct_inline__ T op_sin(T x) {
+    return sycl::sin(x);
+}
+
+template<typename T>
+static __dpct_inline__ T op_cos(T x) {
+    return sycl::cos(x);
+}
+
+template<typename T>
+static __dpct_inline__ T op_hardsigmoid(T x) {
+    return sycl::fmin(static_cast<T>(1.0f), sycl::fmax(static_cast<T>(0.0f), (x + static_cast<T>(3.0f)) / static_cast<T>(6.0f)));
+}
+
+template<typename T>
+static __dpct_inline__ T op_hardswish(T x) {
+    return x * sycl::fmin(static_cast<T>(1.0f), sycl::fmax(static_cast<T>(0.0f), (x + static_cast<T>(3.0f)) / static_cast<T>(6.0f)));
+}
+
+template<typename T>
+static __dpct_inline__ T op_exp(T x) {
+    return sycl::exp(x);
+}
+
+template<typename T>
+static __dpct_inline__ T op_log(T x) {
+    if (x <= static_cast<T>(0)) {
+        return neg_infinity<T>();
+    }
+    return sycl::log(x);
+}
+
+template<typename T>
+static __dpct_inline__ T op_neg(T x) {
+    return -x;
+}
+
+template<typename T>
+static __dpct_inline__ T op_step(T x) {
+    return (x > static_cast<T>(0.0f)) ? static_cast<T>(1.0f) : static_cast<T>(0.0f);
+}
+
+template<typename T>
+static __dpct_inline__ T op_leaky_relu(T x, float negative_slope) {
+    T neg_slope_T = static_cast<T>(negative_slope);
+    return sycl::fmax(x, static_cast<T>(0)) +
+           sycl::fmin(x, static_cast<T>(0.0f)) * neg_slope_T;
+}
+
+template<typename T>
+static __dpct_inline__ T op_sqr(T x) {
+    return x * x;
+}
+
+template<typename T>
+static __dpct_inline__ T op_clamp(T x, float min_val, float max_val) {
+    return x < static_cast<T>(min_val) ? static_cast<T>(min_val) : (x > static_cast<T>(max_val) ? static_cast<T>(max_val) : x);
+}
+
+template<typename T>
+static void unary_op_sgn_kernel(const T * x, T * dst, const int k, const sycl::nd_item<1> &item_ct1) {
+    SYCL_GLOBAL_ID_LOOP(k, item_ct1) {
+        dst[i] = op_sgn(x[i]);
     }
 }
 
 template<typename T>
-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((x[i]));
-}
-
-template<typename T>
-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);
-
-    if (i >= k) {
-        return;
-    }
-    dst[i] = sycl::fmax((x[i]), static_cast<T>(0));
-}
-
-template<typename T>
-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);
-
-    if (i >= k) {
-        return;
-    }
-    dst[i] = 1.0f / (static_cast<T>(1.0f) + sycl::native::exp(-x[i]));
-}
-
-template<typename T>
-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);
-
-    if (i >= k) {
-        return;
-    }
-    dst[i] = sycl::sqrt(x[i]);
-}
-
-template<typename T>
-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);
-
-    if (i >= k) {
-        return;
-    }
-    dst[i] = sycl::sin(x[i]);
-}
-
-template<typename T>
-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);
-
-    if (i >= k) {
-        return;
-    }
-    dst[i] = sycl::cos(x[i]);
-}
-
-template<typename T>
-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);
-
-    if (i >= k) {
-        return;
-    }
-    dst[i] = sycl::fmin(static_cast<T>(1.0f), sycl::fmax(static_cast<T>(0.0f), (x[i] + static_cast<T>(3.0f)) / static_cast<T>(6.0f)));
-}
-
-template<typename T>
-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);
-
-    if (i >= k) {
-        return;
-    }
-    dst[i] = x[i] * sycl::fmin(static_cast<T>(1.0f), sycl::fmax(static_cast<T>(0.0f), (x[i] + static_cast<T>(3.0f)) / static_cast<T>(6.0f)));
-}
-
-template<typename T>
-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);
-
-    if (i >= k) {
-        return;
-    }
-    dst[i] = sycl::exp(x[i]);
-}
-
-template<typename T>
-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);
-
-    if (i >= k) {
-        return;
-    }
-    T xi = x[i];
-    if (xi <= 0) {
-        dst[i] = neg_infinity<T>();
-    } else {
-        dst[i] = sycl::log(xi);
+static void unary_op_abs_kernel(const T * x, T * dst, const int k, const sycl::nd_item<1> &item_ct1) {
+    SYCL_GLOBAL_ID_LOOP(k, item_ct1) {
+        dst[i] = op_abs(x[i]);
     }
 }
 
 template<typename T>
-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);
-
-    if (i >= k) {
-        return;
+static void unary_op_elu_kernel(const T * x, T * dst, const int k, const sycl::nd_item<1> &item_ct1) {
+    SYCL_GLOBAL_ID_LOOP(k, item_ct1) {
+        dst[i] = op_elu(x[i]);
     }
-    dst[i] = -x[i];
 }
 
 template<typename T>
-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);
-
-    if (i >= k) {
-        return;
+static void unary_op_gelu_kernel(const T * x, T * dst, const int k, const sycl::nd_item<1> &item_ct1) {
+    SYCL_GLOBAL_ID_LOOP(k, item_ct1) {
+        dst[i] = op_gelu(x[i]);
     }
-    dst[i] = x[i] > static_cast<T>(0.0f);
 }
 
 template<typename T>
-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;
+static void unary_op_silu_kernel(const T * x, T * dst, const int k, const sycl::nd_item<1> &item_ct1) {
+    SYCL_GLOBAL_ID_LOOP(k, item_ct1) {
+        dst[i] = op_silu(x[i]);
     }
-    dst[i] = sycl::fmax((x[i]), static_cast<T>(0)) +
-             sycl::fmin((x[i]), static_cast<T>(0.0f)) * negative_slope;
 }
 
 template<typename T>
-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);
-
-    if (i >= k) {
-        return;
+static void unary_op_gelu_quick_kernel(const T * x, T * dst, const int k, const sycl::nd_item<1> &item_ct1) {
+    SYCL_GLOBAL_ID_LOOP(k, item_ct1) {
+        dst[i] = op_gelu_quick(x[i]);
+    }
+}
+
+template<typename T>
+static void unary_op_gelu_erf_kernel(const T * x, T * dst, const int k, const sycl::nd_item<1> &item_ct1) {
+    SYCL_GLOBAL_ID_LOOP(k, item_ct1) {
+        dst[i] = op_gelu_erf(x[i]);
+    }
+}
+
+template<typename T>
+static void unary_op_tanh_kernel(const T * x, T * dst, const int k, const sycl::nd_item<1> &item_ct1) {
+    SYCL_GLOBAL_ID_LOOP(k, item_ct1) {
+        dst[i] = op_tanh(x[i]);
+    }
+}
+
+template<typename T>
+static void unary_op_relu_kernel(const T * x, T * dst, const int k, const sycl::nd_item<1> &item_ct1) {
+    SYCL_GLOBAL_ID_LOOP(k, item_ct1) {
+        dst[i] = op_relu(x[i]);
+    }
+}
+
+template<typename T>
+static void unary_op_sigmoid_kernel(const T * x, T * dst, const int k, const sycl::nd_item<1> &item_ct1) {
+    SYCL_GLOBAL_ID_LOOP(k, item_ct1) {
+        dst[i] = op_sigmoid(x[i]);
+    }
+}
+
+template<typename T>
+static void unary_op_sqrt_kernel(const T * x, T * dst, const int k, const sycl::nd_item<1> &item_ct1) {
+    SYCL_GLOBAL_ID_LOOP(k, item_ct1) {
+        dst[i] = op_sqrt(x[i]);
+    }
+}
+
+template<typename T>
+static void unary_op_sin_kernel(const T * x, T * dst, const int k, const sycl::nd_item<1> &item_ct1) {
+    SYCL_GLOBAL_ID_LOOP(k, item_ct1) {
+        dst[i] = op_sin(x[i]);
+    }
+}
+
+template<typename T>
+static void unary_op_cos_kernel(const T * x, T * dst, const int k, const sycl::nd_item<1> &item_ct1) {
+    SYCL_GLOBAL_ID_LOOP(k, item_ct1) {
+        dst[i] = op_cos(x[i]);
+    }
+}
+
+template<typename T>
+static void unary_op_hardsigmoid_kernel(const T * x, T * dst, const int k, const sycl::nd_item<1> &item_ct1) {
+    SYCL_GLOBAL_ID_LOOP(k, item_ct1) {
+        dst[i] = op_hardsigmoid(x[i]);
+    }
+}
+
+template<typename T>
+static void unary_op_hardswish_kernel(const T * x, T * dst, const int k, const sycl::nd_item<1> &item_ct1) {
+    SYCL_GLOBAL_ID_LOOP(k, item_ct1) {
+        dst[i] = op_hardswish(x[i]);
+    }
+}
+
+template<typename T>
+static void unary_op_exp_kernel(const T * x, T * dst, const int k, const sycl::nd_item<1> &item_ct1) {
+    SYCL_GLOBAL_ID_LOOP(k, item_ct1) {
+        dst[i] = op_exp(x[i]);
+    }
+}
+
+template<typename T>
+static void unary_op_log_kernel(const T * x, T * dst, const int k, const sycl::nd_item<1> &item_ct1) {
+    SYCL_GLOBAL_ID_LOOP(k, item_ct1) {
+        dst[i] = op_log(x[i]);
+    }
+}
+
+template<typename T>
+static void unary_op_neg_kernel(const T * x, T * dst, const int k, const sycl::nd_item<1> &item_ct1) {
+    SYCL_GLOBAL_ID_LOOP(k, item_ct1) {
+        dst[i] = op_neg(x[i]);
+    }
+}
+
+template<typename T>
+static void unary_op_step_kernel(const T * x, T * dst, const int k, const sycl::nd_item<1> &item_ct1) {
+    SYCL_GLOBAL_ID_LOOP(k, item_ct1) {
+        dst[i] = op_step(x[i]);
+    }
+}
+
+template<typename T>
+static void unary_op_leaky_relu_kernel(const T * x, T * dst, const int k, float negative_slope, const sycl::nd_item<1> &item_ct1) {
+    SYCL_GLOBAL_ID_LOOP(k, item_ct1) {
+        dst[i] = op_leaky_relu(x[i], negative_slope);
+    }
+}
+
+template<typename T>
+static void unary_op_sqr_kernel(const T * x, T * dst, const int k, const sycl::nd_item<1> &item_ct1) {
+    SYCL_GLOBAL_ID_LOOP(k, item_ct1) {
+        dst[i] = op_sqr(x[i]);
+    }
+}
+
+template<typename T>
+static void unary_op_clamp_kernel(const T * x, T * dst, const int k, const sycl::nd_item<1> &item_ct1, float min_val, float max_val) {
+    SYCL_GLOBAL_ID_LOOP(k, item_ct1) {
+        dst[i] = op_clamp(x[i], min_val, max_val);
     }
-    dst[i] = x[i] * x[i];
 }
 
 template<typename  T>
@@ -281,10 +320,10 @@ static void upscale(const T  *x, T *dst, const int nb00, const int nb01,
     int i12 = (index / (ne10 * ne11)) % ne12;
     int i13 = (index / (ne10 * ne11 * ne12)) % ne13;
 
-    int i00 = i10 / sf0;
-    int i01 = i11 / sf1;
-    int i02 = i12 / sf2;
-    int i03 = i13 / sf3;
+    int i00 = static_cast<int>(i10 / sf0);
+    int i01 = static_cast<int>(i11 / sf1);
+    int i02 = static_cast<int>(i12 / sf2);
+    int i03 = static_cast<int>(i13 / sf3);
 
     dst[index] = *(const T *)((const char *)x + i03 * nb03 + i02 * nb02 + i01 * nb01 + i00 * nb00);
 }
@@ -292,8 +331,7 @@ static void upscale(const T  *x, T *dst, const int nb00, const int nb01,
 template <typename T>
 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);
+    int nidx = SYCL_LOCAL_ID_CALC(item_ct1, 2);
     if (nidx >= ne0) {
         return;
     }
@@ -310,313 +348,85 @@ static void pad(const T  *x, T *dst, const int ne0, const int ne00, const int ne
     }
 }
 
-
 template<typename T>
 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);
-
-    if (i >= k) {
-        return;
+                      const sycl::nd_item<1> &item_ct1) {
+    SYCL_GLOBAL_ID_LOOP(k, item_ct1) {
+        dst[i] = x[i] < static_cast<T>(min) ? static_cast<T>(min) : (x[i] > static_cast<T>(max) ? static_cast<T>(max) : x[i]);
     }
-
-    dst[i] = x[i] < static_cast<T>(min) ? static_cast<T>(min) : (x[i] > static_cast<T>(max) ? static_cast<T>(max) : x[i]);
 }
 
+template<typename T>
+static void gated_op_fused_geglu(const T * x, const T * g, T * dst, const uint64_t k, const uint64_t n, const uint64_t o0, const uint64_t o1, const sycl::nd_item<1> &item_ct1) {
+    SYCL_GLOBAL_ID_LOOP(k, item_ct1) {
+        const int64_t j0 = (i / n) * o0 + (i % n);
+        const int64_t j1 = o0 == o1 ? j0 : (i / n) * o1 + (i % n);
+        dst[i] = op_gelu(x[j0]) * g[j1];
+    }
+}
+
+template<typename T>
+static void gated_op_fused_reglu(const T * x, const T * g, T * dst, const uint64_t k, const uint64_t n, const uint64_t o0, const uint64_t o1, const sycl::nd_item<1> &item_ct1) {
+    SYCL_GLOBAL_ID_LOOP(k, item_ct1) {
+        const int64_t j0 = (i / n) * o0 + (i % n);
+        const int64_t j1 = o0 == o1 ? j0 : (i / n) * o1 + (i % n);
+        dst[i] = op_relu(x[j0]) * g[j1];
+    }
+}
+
+template<typename T>
+static void gated_op_fused_swiglu(const T * x, const T * g, T * dst, const uint64_t k, const uint64_t n, const uint64_t o0, const uint64_t o1, const sycl::nd_item<1> &item_ct1) {
+    SYCL_GLOBAL_ID_LOOP(k, item_ct1)  {
+        const int64_t j0 = (i / n) * o0 + (i % n);
+        const int64_t j1 = o0 == o1 ? j0 : (i / n) * o1 + (i % n);
+        dst[i] = op_silu(x[j0]) * g[j1];
+    }
+}
+
+template<typename T>
+static void gated_op_fused_geglu_erf(const T * x, const T * g, T * dst, const uint64_t k, const uint64_t n, const uint64_t o0, const uint64_t o1, const sycl::nd_item<1> &item_ct1) {
+    SYCL_GLOBAL_ID_LOOP(k, item_ct1) {
+        const int64_t j0 = (i / n) * o0 + (i % n);
+        const int64_t j1 = o0 == o1 ? j0 : (i / n) * o1 + (i % n);
+        dst[i] = op_gelu_erf(x[j0]) * g[j1];
+    }
+}
+
+template<typename T>
+static void gated_op_fused_geglu_quick(const T * x, const T * g, T * dst, const uint64_t k, const uint64_t n, const uint64_t o0, const uint64_t o1, const sycl::nd_item<1> &item_ct1) {
+    SYCL_GLOBAL_ID_LOOP(k, item_ct1) {
+        const int64_t j0 = (i / n) * o0 + (i % n);
+        const int64_t j1 = o0 == o1 ? j0 : (i / n) * o1 + (i % n);
+        dst[i] = op_gelu_quick(x[j0]) * g[j1];
+    }
+}
+
+namespace ggml_sycl_detail {
 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) {
-    int num_blocks = (n_elements + SYCL_ACC_BLOCK_SIZE - 1) / SYCL_ACC_BLOCK_SIZE;
-    stream->parallel_for(
-        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) *
-                              sycl::range<3>(1, 1, SYCL_ACC_BLOCK_SIZE),
-                          sycl::range<3>(1, 1, SYCL_ACC_BLOCK_SIZE)),
-        [=](sycl::nd_item<3> item_ct1) {
+    int num_blocks = ceil_div(n_elements, SYCL_ACC_BLOCK_SIZE);
+    sycl_parallel_for(stream,
+        sycl::nd_range<1>(sycl::range<1>(num_blocks) *
+                              sycl::range<1>(SYCL_ACC_BLOCK_SIZE),
+                          sycl::range<1>(SYCL_ACC_BLOCK_SIZE)),
+        [=](sycl::nd_item<1> item_ct1) {
             acc_f32(x, y, dst, n_elements, ne10, ne11, ne12, nb1, nb2, offset,
                     item_ct1);
         });
 }
 
-template<typename T>
-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(
-        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) *
-                              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(x, dst, k, item_ct1);
-        });
-}
-
-template<typename T>
-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(
-        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) *
-                              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(x, dst, k, item_ct1);
-        });
-}
-
-template<typename T>
-static void sgn_sycl(const T * x, T * dst, const int k, queue_ptr stream) {
-    // hard code for now
-    const int num_blocks = ceil_div(k, 256);
-    stream->parallel_for(
-            sycl::nd_range<3>((sycl::range<3>(1, 1, num_blocks) * sycl::range(1, 1, 256)), sycl::range(1, 1, 256)), [=](sycl::nd_item<3> item_ct1) {
-            sgn(x, dst, k, item_ct1);
-            });
-}
-
-template<typename T>
-static void abs_sycl(const T * x, T * dst, const int k, queue_ptr stream) {
-    // hard code for now
-    const int num_blocks = ceil_div(k, 256);
-    stream->parallel_for(
-            sycl::nd_range<3>((sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, 256)), sycl::range<3>(1, 1, 256)), [=](sycl::nd_item<3> item_ct1) {
-            abs_op(x, dst, k, item_ct1);
-            });
-}
-
-
-template<typename T>
-static void elu_sycl(const T * x, T * dst, const int k, queue_ptr stream) {
-    // hard code for now
-    const int num_blocks = ceil_div(k, 256);
-    stream->parallel_for(
-            sycl::nd_range<3>((sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, 256)), sycl::range<3>(1, 1, 256)), [=](sycl::nd_item<3> item_ct1) {
-            elu_op(x, dst, k, item_ct1);
-            });
-}
-
-template<typename T>
-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(
-        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) *
-                              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(x, dst, k, item_ct1);
-        });
-}
-
-
-template<typename T>
-static void gelu_erf_sycl(const T *x, T *dst, const int k,
-                                queue_ptr stream) {
-    const int num_blocks = ceil_div(k, SYCL_GELU_BLOCK_SIZE);
-    stream->parallel_for(
-        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) *
-                              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_erf(x, dst, k, item_ct1);
-        });
-}
-
-template<typename T>
-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(
-        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) *
-                              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(x, dst, k, item_ct1);
-        });
-}
-
-template<typename T>
-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(
-        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) *
-                              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(x, dst, k, item_ct1);
-        });
-}
-
-template<typename T>
-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(
-        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) *
-                              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(x, dst, k, item_ct1);
-        });
-}
-
-template<typename T>
-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(
-        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) *
-                              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(x, dst, k, item_ct1);
-        });
-}
-
-template<typename T>
-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(
-        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) *
-                              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(x, dst, k, item_ct1);
-        });
-}
-
-template<typename T>
-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(
-        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) *
-                              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(x, dst, k, item_ct1);
-        });
-}
-
-template<typename T>
-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(
-        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) *
-                              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(x, dst, k, item_ct1);
-        });
-}
-
-template<typename T>
-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(
-        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) *
-                              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(x, dst, k, item_ct1);
-        });
-}
-
-template<typename T>
-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(
-        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) *
-                              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(x, dst, k, item_ct1);
-        });
-}
-
-template<typename T>
-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(
-        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) *
-                              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(x, dst, k, item_ct1);
-        });
-}
-
-template<typename T>
-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(
-        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) *
-                              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(x, dst, k, item_ct1);
-        });
-}
-
-template<typename T>
-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(
-        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) *
-                              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(x, dst, k, item_ct1);
-        });
-}
-
-template<typename T>
-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;
-    stream->parallel_for(
-        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) *
-                              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(x, dst, k, negative_slope, item_ct1);
-        });
-}
-
-template<typename T>
-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(
-        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) *
-                              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(x, dst, k, item_ct1);
-        });
-}
-
 template<typename T>
 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) {
     int dst_size = ne10 * ne11 * ne12 * ne13;
-    int num_blocks = (dst_size + SYCL_UPSCALE_BLOCK_SIZE - 1) / SYCL_UPSCALE_BLOCK_SIZE;
+    int num_blocks = ceil_div(dst_size, SYCL_UPSCALE_BLOCK_SIZE);
     sycl::range<1> gridDim(num_blocks * SYCL_UPSCALE_BLOCK_SIZE);
-    stream->parallel_for(
-        sycl::nd_range<1>(gridDim, sycl::range<1>(SYCL_UPSCALE_BLOCK_SIZE)),
-        [=](sycl::nd_item<1> item_ct1) {
+    sycl_parallel_for<1>(
+        stream, sycl::nd_range<1>(gridDim, sycl::range<1>(SYCL_UPSCALE_BLOCK_SIZE)), [=](sycl::nd_item<1> item_ct1) {
             upscale(x, dst, nb00, nb01, nb02, nb03, ne10, ne11, ne12, ne13, sf0, sf1, sf2, sf3, item_ct1);
         });
 }
@@ -625,128 +435,16 @@ template<typename T>
 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;
+    int num_blocks = ceil_div(ne0, SYCL_PAD_BLOCK_SIZE);
     sycl::range<3> gridDim(ne2, ne1, num_blocks);
-    stream->parallel_for(
-        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(x, dst, ne0, ne00, ne01, ne02, item_ct1);
-        });
+    sycl_parallel_for(stream,
+                      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(x, dst, ne0, ne00, ne01, ne02, item_ct1); });
 }
 
-template<typename T>
-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_sgn(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<sycl::half>(dst);
-                sgn_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<float>(dst);
-                sgn_sycl(data_pts.src, data_pts.dst, ggml_nelements(dst->src[0]), main_stream);
-                break;
-            }
-        default:
-            GGML_ABORT("GGML tensor type not supported!\n");
-    }
-}
-
-inline void ggml_sycl_op_abs(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<sycl::half>(dst);
-                abs_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<float>(dst);
-                abs_sycl(data_pts.src, data_pts.dst, ggml_nelements(dst->src[0]), main_stream);
-                break;
-            }
-        default:
-            GGML_ABORT("GGML tensor type not supported!\n");
-    }
-}
-
-
-inline void ggml_sycl_op_elu(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<sycl::half>(dst);
-                elu_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<float>(dst);
-                elu_sycl(data_pts.src, data_pts.dst, ggml_nelements(dst->src[0]), main_stream);
-                break;
-            }
-        default:
-            GGML_ABORT("GGML tensor type not supported!\n");
-    }
-}
-
-inline void ggml_sycl_op_silu(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+template<typename KernelInvoker, typename... Args>
+static inline void dispatch_ggml_sycl_op_unary(ggml_backend_sycl_context & ctx, ggml_tensor * dst, KernelInvoker kernel_invoker, Args&&... args) {
 #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);
@@ -762,14 +460,14 @@ inline void ggml_sycl_op_silu(ggml_backend_sycl_context & ctx, ggml_tensor * dst
         case GGML_TYPE_F16:
             {
                 auto data_pts = cast_data<sycl::half>(dst);
-                silu_sycl(data_pts.src, data_pts.dst, ggml_nelements(dst->src[0]), main_stream);
+                kernel_invoker(data_pts.src, data_pts.dst, (int)ggml_nelements(dst->src[0]), main_stream, std::forward<Args>(args)...);
                 break;
             }
 #endif
         case GGML_TYPE_F32:
             {
                 auto data_pts = cast_data<float>(dst);
-                silu_sycl(data_pts.src, data_pts.dst, ggml_nelements(dst->src[0]), main_stream);
+                kernel_invoker(data_pts.src, data_pts.dst, (int)ggml_nelements(dst->src[0]), main_stream, std::forward<Args>(args)...);
                 break;
             }
         default:
@@ -777,7 +475,8 @@ inline void ggml_sycl_op_silu(ggml_backend_sycl_context & ctx, ggml_tensor * dst
     }
 }
 
-inline void ggml_sycl_op_gelu(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+template<typename KernelInvoker, typename... Args>
+static inline void dispatch_ggml_sycl_op_fused_glu(ggml_backend_sycl_context & ctx, ggml_tensor * dst, KernelInvoker kernel_invoker, Args&&... args) {
 #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);
@@ -788,19 +487,66 @@ inline void ggml_sycl_op_gelu(ggml_backend_sycl_context & ctx, ggml_tensor * dst
     GGML_ASSERT(dst->src[0]->type == dst->type);
     dpct::queue_ptr main_stream = ctx.stream();
     SYCL_CHECK(ggml_sycl_set_device(ctx.device));
+    const ggml_tensor * src0 = dst->src[0];
+    const ggml_tensor * src1 = dst->src[1];
+    const int64_t nc = src1 ? src0->ne[0] : src0->ne[0] / 2;;
+    GGML_ASSERT(dst->ne[0] == nc);
+    GGML_ASSERT(ggml_is_contiguous_1(dst->src[0]));
+    GGML_ASSERT(ggml_is_contiguous(dst));
+    const int32_t swapped = ((const int32_t *) dst->op_params)[1];
+    void * src0_d = src0->data;
+    void * src1_d = src1 ? src1->data : src0->data;
+    const int64_t src0_o = src0->nb[1];
+    const int64_t src1_o = src1 ? src1->nb[1] : src0->nb[1];
+    void * dst_d = dst->data;
+    if (src1) {
+        GGML_ASSERT(ggml_is_contiguous_1(src1));
+        GGML_ASSERT(src1->nb[0] == ggml_element_size(src1));
+        GGML_ASSERT(src1->ne[0] == nc);
+        GGML_ASSERT(src0->type == src1->type);
+    }
     switch (dst->type) {
 #if defined (GGML_SYCL_F16)
         case GGML_TYPE_F16:
             {
-                auto data_pts = cast_data<sycl::half>(dst);
-                gelu_sycl(data_pts.src, data_pts.dst, ggml_nelements(dst->src[0]), main_stream);
+                sycl::half * src0_p = (sycl::half *) src0_d;
+                sycl::half * src1_p = (sycl::half *) src1_d;
+
+                    if (!src1) {
+                        src0_p += swapped ? nc : 0;
+                        src1_p += swapped ? 0 : nc;
+                    }
+                kernel_invoker(src0_p,
+                               src1_p,
+                               (sycl::half *) dst_d,
+                               ggml_nelements(dst),
+                               nc,
+                               src0_o / sizeof(sycl::half),
+                               src1_o / sizeof(sycl::half),
+                               main_stream,
+                               std::forward<Args>(args)...);
                 break;
             }
 #endif
         case GGML_TYPE_F32:
             {
-                auto data_pts = cast_data<float>(dst);
-                gelu_sycl(data_pts.src, data_pts.dst, ggml_nelements(dst->src[0]), main_stream);
+                float * src0_p = (float *) src0_d;
+                float * src1_p = (float *) src1_d;
+
+                    if (!src1) {
+                        src0_p += swapped ? nc : 0;
+                        src1_p += swapped ? 0 : nc;
+                    }
+
+                kernel_invoker(src0_p,
+                               src1_p,
+                               (float *) dst_d,
+                               ggml_nelements(dst),
+                               nc,
+                               src0_o / sizeof(float),
+                               src1_o / sizeof(float),
+                               main_stream,
+                               std::forward<Args>(args)...);
                 break;
             }
         default:
@@ -808,511 +554,8 @@ inline void ggml_sycl_op_gelu(ggml_backend_sycl_context & ctx, ggml_tensor * dst
     }
 }
 
-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<sycl::half>(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<float>(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");
-    }
-}
-
-inline void ggml_sycl_op_gelu_erf(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<sycl::half>(dst);
-                gelu_erf_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<float>(dst);
-                gelu_erf_sycl(data_pts.src, data_pts.dst, ggml_nelements(dst->src[0]), main_stream);
-                break;
-            }
-        default:
-            GGML_ABORT("GGML tensor type not supported!\n");
-    }
-}
-
-
-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<sycl::half>(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<float>(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");
-    }
-}
-
-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));
-
-    switch (dst->type) {
-#if defined (GGML_SYCL_F16)
-        case GGML_TYPE_F16:
-            {
-                auto data_pts = cast_data<sycl::half>(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<float>(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");
-    }
-}
-
-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);
-
-    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<sycl::half>(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<float>(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");
-    }
-}
-
-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<sycl::half>(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<float>(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");
-    }
-}
-
-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<sycl::half>(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<float>(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");
-    }
-}
-
-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<sycl::half>(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<float>(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");
-    }
-}
-
-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<sycl::half>(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<float>(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");
-    }
-}
-
-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);
-
-    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<sycl::half>(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<float>(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");
-    }
-}
-
-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<sycl::half>(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<float>(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");
-    }
-}
-
-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<sycl::half>(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<float>(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");
-    }
-}
-
-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<sycl::half>(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<float>(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");
-    }
-}
-
-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<sycl::half>(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<float>(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");
-    }
-}
-
-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));
-    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<sycl::half>(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<float>(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");
-    }
-}
-
-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);
-#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<sycl::half>(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<float>(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");
-    }
-}
-
-inline void ggml_sycl_op_upscale(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+template<typename KernelInvoker, typename... Args>
+static inline void dispatch_ggml_sycl_op_upscale(ggml_backend_sycl_context & ctx, ggml_tensor * dst, KernelInvoker kernel_invoker, Args&&... args) {
 #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);
@@ -1334,18 +577,18 @@ inline void ggml_sycl_op_upscale(ggml_backend_sycl_context & ctx, ggml_tensor *
         case GGML_TYPE_F16:
             {
                 auto data_pts = cast_data<sycl::half>(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);
+                kernel_invoker(data_pts.src, data_pts.dst, (int)dst->src[0]->nb[0], (int)dst->src[0]->nb[1], (int)dst->src[0]->nb[2],
+                               (int)dst->src[0]->nb[3], (int)dst->ne[0], (int)dst->ne[1], (int)dst->ne[2], (int)dst->ne[3], sf0, sf1, sf2, sf3,
+                               main_stream, std::forward<Args>(args)...);
                 break;
             }
 #endif
         case GGML_TYPE_F32:
             {
                 auto data_pts = cast_data<float>(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);
+                kernel_invoker(data_pts.src, data_pts.dst, (int)dst->src[0]->nb[0], (int)dst->src[0]->nb[1], (int)dst->src[0]->nb[2],
+                               (int)dst->src[0]->nb[3], (int)dst->ne[0], (int)dst->ne[1], (int)dst->ne[2], (int)dst->ne[3], sf0, sf1, sf2, sf3,
+                               main_stream, std::forward<Args>(args)...);
                 break;
             }
         default:
@@ -1353,7 +596,8 @@ inline void ggml_sycl_op_upscale(ggml_backend_sycl_context & ctx, ggml_tensor *
     }
 }
 
-inline void ggml_sycl_op_pad(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+template<typename KernelInvoker, typename... Args>
+static inline void dispatch_ggml_sycl_op_pad(ggml_backend_sycl_context & ctx, ggml_tensor * dst, KernelInvoker kernel_invoker, Args&&... args) {
 #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);
@@ -1362,7 +606,7 @@ inline void ggml_sycl_op_pad(ggml_backend_sycl_context & ctx, ggml_tensor * dst)
     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
+    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) {
@@ -1370,16 +614,16 @@ inline void ggml_sycl_op_pad(ggml_backend_sycl_context & ctx, ggml_tensor * dst)
         case GGML_TYPE_F16:
             {
                 auto data_pts = cast_data<sycl::half>(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);
+                kernel_invoker(data_pts.src, data_pts.dst, (int)dst->src[0]->ne[0], (int)dst->src[0]->ne[1], (int)dst->src[0]->ne[2], (int)dst->ne[0],
+                               (int)dst->ne[1], (int)dst->ne[2], main_stream, std::forward<Args>(args)...);
                 break;
             }
 #endif
         case GGML_TYPE_F32:
             {
                 auto data_pts = cast_data<float>(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);
+                kernel_invoker(data_pts.src, data_pts.dst, (int)dst->src[0]->ne[0], (int)dst->src[0]->ne[1], (int)dst->src[0]->ne[2], (int)dst->ne[0],
+                               (int)dst->ne[1], (int)dst->ne[2], main_stream, std::forward<Args>(args)...);
                 break;
             }
         default:
@@ -1387,45 +631,320 @@ inline void ggml_sycl_op_pad(ggml_backend_sycl_context & ctx, ggml_tensor * dst)
     }
 }
 
-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
+} // namespace ggml_sycl_detail
 
-    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<sycl::half>(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<float>(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");
-    }
+
+static inline void ggml_sycl_op_sgn(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    ggml_sycl_detail::dispatch_ggml_sycl_op_unary(ctx, dst,
+        [](const auto* src, auto* dst_ptr, int k_elements, queue_ptr stream) {
+            const int num_blocks = ceil_div(k_elements, 256);
+            sycl_parallel_for(stream,
+                sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(256),
+                                  sycl::range<1>(256)),
+                [=](sycl::nd_item<1> item_ct1) {
+                    unary_op_sgn_kernel(src, dst_ptr, k_elements, item_ct1);
+                });
+        });
 }
 
-inline void ggml_sycl_op_acc(ggml_backend_sycl_context & ctx, ggml_tensor *dst) {
+static inline void ggml_sycl_op_abs(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    ggml_sycl_detail::dispatch_ggml_sycl_op_unary(ctx, dst,
+        [](const auto* src, auto* dst_ptr, int k_elements, queue_ptr stream) {
+            const int num_blocks = ceil_div(k_elements, 256);
+            sycl_parallel_for(stream,
+                sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(256),
+                                  sycl::range<1>(256)),
+                [=](sycl::nd_item<1> item_ct1) {
+                    unary_op_abs_kernel(src, dst_ptr, k_elements, item_ct1);
+                });
+        });
+}
 
+static inline void ggml_sycl_op_elu(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    ggml_sycl_detail::dispatch_ggml_sycl_op_unary(ctx, dst,
+        [](const auto* src, auto* dst_ptr, int k_elements, queue_ptr stream) {
+            const int num_blocks = ceil_div(k_elements, 256);
+            sycl_parallel_for(stream,
+                sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(256),
+                                  sycl::range<1>(256)),
+                [=](sycl::nd_item<1> item_ct1) {
+                    unary_op_elu_kernel(src, dst_ptr, k_elements, item_ct1);
+                });
+        });
+}
+
+static inline void ggml_sycl_op_silu(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    ggml_sycl_detail::dispatch_ggml_sycl_op_unary(ctx, dst,
+        [](const auto* src, auto* dst_ptr, int k_elements, queue_ptr stream) {
+            const int num_blocks = ceil_div(k_elements, SYCL_SILU_BLOCK_SIZE);
+            sycl_parallel_for(stream,
+                sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(SYCL_SILU_BLOCK_SIZE),
+                                  sycl::range<1>(SYCL_SILU_BLOCK_SIZE)),
+                [=](sycl::nd_item<1> item_ct1) {
+                    unary_op_silu_kernel(src, dst_ptr, k_elements, item_ct1);
+                });
+        });
+}
+
+static inline void ggml_sycl_op_gelu(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    ggml_sycl_detail::dispatch_ggml_sycl_op_unary(ctx, dst,
+        [](const auto* src, auto* dst_ptr, int k_elements, queue_ptr stream) {
+            const int num_blocks = ceil_div(k_elements, SYCL_GELU_BLOCK_SIZE);
+            sycl_parallel_for(stream,
+                sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(SYCL_GELU_BLOCK_SIZE),
+                                  sycl::range<1>(SYCL_GELU_BLOCK_SIZE)),
+                [=](sycl::nd_item<1> item_ct1) {
+                    unary_op_gelu_kernel(src, dst_ptr, k_elements, item_ct1);
+                });
+        });
+}
+
+static inline void ggml_sycl_op_gelu_quick(ggml_backend_sycl_context & ctx, ggml_tensor *dst) {
+    ggml_sycl_detail::dispatch_ggml_sycl_op_unary(ctx, dst,
+        [](const auto* src, auto* dst_ptr, int k_elements, queue_ptr stream) {
+            const int num_blocks = ceil_div(k_elements, SYCL_GELU_BLOCK_SIZE);
+            sycl_parallel_for(stream,
+                sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(SYCL_GELU_BLOCK_SIZE),
+                                  sycl::range<1>(SYCL_GELU_BLOCK_SIZE)),
+                [=](sycl::nd_item<1> item_ct1) {
+                    unary_op_gelu_quick_kernel(src, dst_ptr, k_elements, item_ct1);
+                });
+        });
+}
+
+static inline void ggml_sycl_op_gelu_erf(ggml_backend_sycl_context & ctx, ggml_tensor *dst) {
+    ggml_sycl_detail::dispatch_ggml_sycl_op_unary(ctx, dst,
+        [](const auto* src, auto* dst_ptr, int k_elements, queue_ptr stream) {
+            const int num_blocks = ceil_div(k_elements, SYCL_GELU_BLOCK_SIZE);
+            sycl_parallel_for(stream,
+                sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(SYCL_GELU_BLOCK_SIZE),
+                                  sycl::range<1>(SYCL_GELU_BLOCK_SIZE)),
+                [=](sycl::nd_item<1> item_ct1) {
+                    unary_op_gelu_erf_kernel(src, dst_ptr, k_elements, item_ct1);
+                });
+        });
+}
+
+static inline void ggml_sycl_op_tanh(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    ggml_sycl_detail::dispatch_ggml_sycl_op_unary(ctx, dst,
+        [](const auto* src, auto* dst_ptr, int k_elements, queue_ptr stream) {
+            const int num_blocks = ceil_div(k_elements, SYCL_TANH_BLOCK_SIZE);
+            sycl_parallel_for(stream,
+                sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(SYCL_TANH_BLOCK_SIZE),
+                                  sycl::range<1>(SYCL_TANH_BLOCK_SIZE)),
+                [=](sycl::nd_item<1> item_ct1) {
+                    unary_op_tanh_kernel(src, dst_ptr, k_elements, item_ct1);
+                });
+        });
+}
+
+static inline void ggml_sycl_op_relu(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    ggml_sycl_detail::dispatch_ggml_sycl_op_unary(ctx, dst,
+        [](const auto* src, auto* dst_ptr, int k_elements, queue_ptr stream) {
+            const int num_blocks = ceil_div(k_elements, SYCL_RELU_BLOCK_SIZE);
+            sycl_parallel_for(stream,
+                sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(SYCL_RELU_BLOCK_SIZE),
+                                  sycl::range<1>(SYCL_RELU_BLOCK_SIZE)),
+                [=](sycl::nd_item<1> item_ct1) {
+                    unary_op_relu_kernel(src, dst_ptr, k_elements, item_ct1);
+                });
+        });
+}
+
+static inline void ggml_sycl_op_hardsigmoid(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    ggml_sycl_detail::dispatch_ggml_sycl_op_unary(ctx, dst,
+        [](const auto* src, auto* dst_ptr, int k_elements, queue_ptr stream) {
+            const int num_blocks = ceil_div(k_elements, SYCL_HARDSIGMOID_BLOCK_SIZE);
+            sycl_parallel_for(stream,
+                sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(SYCL_HARDSIGMOID_BLOCK_SIZE),
+                                  sycl::range<1>(SYCL_HARDSIGMOID_BLOCK_SIZE)),
+                [=](sycl::nd_item<1> item_ct1) {
+                    unary_op_hardsigmoid_kernel(src, dst_ptr, k_elements, item_ct1);
+                });
+        });
+}
+
+static inline void ggml_sycl_op_hardswish(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    ggml_sycl_detail::dispatch_ggml_sycl_op_unary(ctx, dst,
+        [](const auto* src, auto* dst_ptr, int k_elements, queue_ptr stream) {
+            const int num_blocks = ceil_div(k_elements, SYCL_HARDSWISH_BLOCK_SIZE);
+            sycl_parallel_for(stream,
+                sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(SYCL_HARDSWISH_BLOCK_SIZE),
+                                  sycl::range<1>(SYCL_HARDSWISH_BLOCK_SIZE)),
+                [=](sycl::nd_item<1> item_ct1) {
+                    unary_op_hardswish_kernel(src, dst_ptr, k_elements, item_ct1);
+                });
+        });
+}
+
+static inline void ggml_sycl_op_exp(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    ggml_sycl_detail::dispatch_ggml_sycl_op_unary(ctx, dst,
+        [](const auto* src, auto* dst_ptr, int k_elements, queue_ptr stream) {
+            const int num_blocks = ceil_div(k_elements, SYCL_EXP_BLOCK_SIZE);
+            sycl_parallel_for(stream,
+                sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(SYCL_EXP_BLOCK_SIZE),
+                                  sycl::range<1>(SYCL_EXP_BLOCK_SIZE)),
+                [=](sycl::nd_item<1> item_ct1) {
+                    unary_op_exp_kernel(src, dst_ptr, k_elements, item_ct1);
+                });
+        });
+}
+
+static inline void ggml_sycl_op_log(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    ggml_sycl_detail::dispatch_ggml_sycl_op_unary(ctx, dst,
+        [](const auto* src, auto* dst_ptr, int k_elements, queue_ptr stream) {
+            const int num_blocks = ceil_div(k_elements, SYCL_EXP_BLOCK_SIZE); // Using EXP block size
+            sycl_parallel_for(stream,
+                sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(SYCL_EXP_BLOCK_SIZE),
+                                  sycl::range<1>(SYCL_EXP_BLOCK_SIZE)),
+                [=](sycl::nd_item<1> item_ct1) {
+                    unary_op_log_kernel(src, dst_ptr, k_elements, item_ct1);
+                });
+        });
+}
+
+static inline void ggml_sycl_op_neg(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    ggml_sycl_detail::dispatch_ggml_sycl_op_unary(ctx, dst,
+        [](const auto* src, auto* dst_ptr, int k_elements, queue_ptr stream) {
+            const int num_blocks = ceil_div(k_elements, SYCL_NEG_BLOCK_SIZE);
+            sycl_parallel_for(stream,
+                sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(SYCL_NEG_BLOCK_SIZE),
+                                  sycl::range<1>(SYCL_NEG_BLOCK_SIZE)),
+                [=](sycl::nd_item<1> item_ct1) {
+                    unary_op_neg_kernel(src, dst_ptr, k_elements, item_ct1);
+                });
+        });
+}
+
+static inline void ggml_sycl_op_step(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    ggml_sycl_detail::dispatch_ggml_sycl_op_unary(ctx, dst,
+        [](const auto* src, auto* dst_ptr, int k_elements, queue_ptr stream) {
+            const int num_blocks = ceil_div(k_elements, SYCL_NEG_BLOCK_SIZE); // Using NEG block size
+            sycl_parallel_for(stream,
+                sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(SYCL_NEG_BLOCK_SIZE),
+                                  sycl::range<1>(SYCL_NEG_BLOCK_SIZE)),
+                [=](sycl::nd_item<1> item_ct1) {
+                    unary_op_step_kernel(src, dst_ptr, k_elements, item_ct1);
+                });
+        });
+}
+
+static inline void ggml_sycl_op_sigmoid(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    ggml_sycl_detail::dispatch_ggml_sycl_op_unary(ctx, dst,
+        [](const auto* src, auto* dst_ptr, int k_elements, queue_ptr stream) {
+            const int num_blocks = ceil_div(k_elements, SYCL_SIGMOID_BLOCK_SIZE);
+            sycl_parallel_for(stream,
+                sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(SYCL_SIGMOID_BLOCK_SIZE),
+                                  sycl::range<1>(SYCL_SIGMOID_BLOCK_SIZE)),
+                [=](sycl::nd_item<1> item_ct1) {
+                    unary_op_sigmoid_kernel(src, dst_ptr, k_elements, item_ct1);
+                });
+        });
+}
+
+static inline void ggml_sycl_op_sqrt(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    ggml_sycl_detail::dispatch_ggml_sycl_op_unary(ctx, dst,
+        [](const auto* src, auto* dst_ptr, int k_elements, queue_ptr stream) {
+            const int num_blocks = ceil_div(k_elements, SYCL_SQRT_BLOCK_SIZE);
+            sycl_parallel_for(stream,
+                sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(SYCL_SQRT_BLOCK_SIZE),
+                                  sycl::range<1>(SYCL_SQRT_BLOCK_SIZE)),
+                [=](sycl::nd_item<1> item_ct1) {
+                    unary_op_sqrt_kernel(src, dst_ptr, k_elements, item_ct1);
+                });
+        });
+}
+
+static inline void ggml_sycl_op_sin(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    ggml_sycl_detail::dispatch_ggml_sycl_op_unary(ctx, dst,
+        [](const auto* src, auto* dst_ptr, int k_elements, queue_ptr stream) {
+            const int num_blocks = ceil_div(k_elements, SYCL_SIN_BLOCK_SIZE);
+            sycl_parallel_for(stream,
+                sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(SYCL_SIN_BLOCK_SIZE),
+                                  sycl::range<1>(SYCL_SIN_BLOCK_SIZE)),
+                [=](sycl::nd_item<1> item_ct1) {
+                    unary_op_sin_kernel(src, dst_ptr, k_elements, item_ct1);
+                });
+        });
+}
+
+static inline void ggml_sycl_op_cos(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    ggml_sycl_detail::dispatch_ggml_sycl_op_unary(ctx, dst,
+        [](const auto* src, auto* dst_ptr, int k_elements, queue_ptr stream) {
+            const int num_blocks = ceil_div(k_elements, SYCL_SIN_BLOCK_SIZE); // Using SIN block size
+            sycl_parallel_for(stream,
+                sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(SYCL_SIN_BLOCK_SIZE),
+                                  sycl::range<1>(SYCL_SIN_BLOCK_SIZE)),
+                [=](sycl::nd_item<1> item_ct1) {
+                    unary_op_cos_kernel(src, dst_ptr, k_elements, item_ct1);
+                });
+        });
+}
+
+static inline void ggml_sycl_op_leaky_relu(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    float negative_slope;
+    memcpy(&negative_slope, dst->op_params, sizeof(float));
+    ggml_sycl_detail::dispatch_ggml_sycl_op_unary(ctx, dst,
+        [](const auto* src, auto* dst_ptr, int k_elements, queue_ptr stream, float slope) {
+            const int num_blocks = ceil_div(k_elements, SYCL_RELU_BLOCK_SIZE);
+            sycl_parallel_for(stream,
+                sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(SYCL_RELU_BLOCK_SIZE),
+                                  sycl::range<1>(SYCL_RELU_BLOCK_SIZE)),
+                [=](sycl::nd_item<1> item_ct1) {
+                    unary_op_leaky_relu_kernel(src, dst_ptr, k_elements, slope, item_ct1);
+                });
+        }, negative_slope);
+}
+
+static inline void ggml_sycl_op_sqr(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    ggml_sycl_detail::dispatch_ggml_sycl_op_unary(ctx, dst,
+        [](const auto* src, auto* dst_ptr, int k_elements, queue_ptr stream) {
+            const int num_blocks = ceil_div(k_elements, SYCL_SQR_BLOCK_SIZE);
+            sycl_parallel_for(stream,
+                sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(SYCL_SQR_BLOCK_SIZE),
+                                  sycl::range<1>(SYCL_SQR_BLOCK_SIZE)),
+                [=](sycl::nd_item<1> item_ct1) {
+                    unary_op_sqr_kernel(src, dst_ptr, k_elements, item_ct1);
+                });
+        });
+}
+
+static inline void ggml_sycl_op_upscale(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    ggml_sycl_detail::dispatch_ggml_sycl_op_upscale(ctx, dst,
+        [](const auto* src, auto* dst_ptr, int nb00, int nb01, int nb02, int nb03,
+           int ne10, int ne11, int ne12, int ne13, float sf0, float sf1, float sf2, float sf3,
+           queue_ptr stream) {
+            ggml_sycl_detail::upscale_sycl(src, dst_ptr, nb00, nb01, nb02, nb03, ne10, ne11, ne12, ne13, sf0, sf1, sf2, sf3, stream);
+        });
+}
+
+static inline void ggml_sycl_op_pad(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    ggml_sycl_detail::dispatch_ggml_sycl_op_pad(ctx, dst,
+        [](const auto* src, auto* dst_ptr, int ne00, int ne01, int ne02, int ne0, int ne1, int ne2,
+           queue_ptr stream) {
+            ggml_sycl_detail::pad_sycl(src, dst_ptr, ne00, ne01, ne02, ne0, ne1, ne2, stream);
+        });
+}
+
+static inline void ggml_sycl_op_clamp(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    float min_val;
+    float max_val;
+    memcpy(&min_val, dst->op_params, sizeof(float));
+    memcpy(&max_val, (float *) dst->op_params + 1, sizeof(float));
+    ggml_sycl_detail::dispatch_ggml_sycl_op_unary(ctx, dst,
+        [](const auto* src, auto* dst_ptr, int k_elements, queue_ptr stream, float min_arg, float max_arg) {
+            const int num_blocks = ceil_div(k_elements, SYCL_CLAMP_BLOCK_SIZE);
+            sycl_parallel_for(stream,
+                sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(SYCL_CLAMP_BLOCK_SIZE),
+                                  sycl::range<1>(SYCL_CLAMP_BLOCK_SIZE)),
+                [=](sycl::nd_item<1> item_ct1) {
+                    clamp(src, dst_ptr, min_arg, max_arg, k_elements, item_ct1);
+                });
+        }, min_val, max_val);
+}
+
+static 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);
@@ -1441,7 +960,62 @@ inline void ggml_sycl_op_acc(ggml_backend_sycl_context & ctx, ggml_tensor *dst)
     // 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), dst->src[1]->ne[0], dst->src[1]->ne[1], dst->src[1]->ne[2], nb1, nb2, offset, main_stream);
+    ggml_sycl_detail::acc_f32_sycl(src0_dd, src1_dd, dst_dd, (int)ggml_nelements(dst), (int)dst->src[1]->ne[0], (int)dst->src[1]->ne[1], (int)dst->src[1]->ne[2], nb1, nb2, offset, main_stream);
+}
+
+static inline void ggml_sycl_op_geglu(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    ggml_sycl_detail::dispatch_ggml_sycl_op_fused_glu(ctx, dst,
+        [](const auto* x_ptr, const auto* g_ptr, auto* dst_ptr, uint64_t k, uint64_t n, uint64_t o0, uint64_t o1, queue_ptr main_stream) {
+            const uint32_t num_blocks = ceil_div(k, SYCL_GELU_BLOCK_SIZE);
+            sycl_parallel_for(main_stream,
+                    sycl::nd_range<1>((num_blocks * sycl::range<1>(SYCL_GELU_BLOCK_SIZE)), sycl::range<1>(SYCL_GELU_BLOCK_SIZE)), [=](sycl::nd_item<1> item_ct1) {
+                gated_op_fused_geglu(x_ptr, g_ptr, dst_ptr, k, n, o0, o1, item_ct1);
+            });
+        });
+}
+
+static inline void ggml_sycl_op_reglu(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    ggml_sycl_detail::dispatch_ggml_sycl_op_fused_glu(ctx, dst,
+        [](const auto* x_ptr, const auto* g_ptr, auto* dst_ptr, uint64_t k, uint64_t n, uint64_t o0, uint64_t o1, queue_ptr main_stream) {
+            const uint32_t num_blocks = ceil_div((uint32_t)k, SYCL_RELU_BLOCK_SIZE); // Using RELU block size for reglu
+            sycl_parallel_for(main_stream,
+                    sycl::nd_range<1>((num_blocks * sycl::range<1>(SYCL_RELU_BLOCK_SIZE)), sycl::range<1>(SYCL_RELU_BLOCK_SIZE)), [=](sycl::nd_item<1> item_ct1) {
+                gated_op_fused_reglu(x_ptr, g_ptr, dst_ptr, k, n, o0, o1, item_ct1);
+            });
+        });
+}
+
+static inline void ggml_sycl_op_swiglu(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    ggml_sycl_detail::dispatch_ggml_sycl_op_fused_glu(ctx, dst,
+        [](const auto* x_ptr, const auto* g_ptr, auto* dst_ptr, uint64_t k, uint64_t n, uint64_t o0, uint64_t o1, queue_ptr main_stream) {
+            const uint32_t num_blocks = ceil_div((uint32_t)k, SYCL_SILU_BLOCK_SIZE); // Using SILU block size for swiglu
+            sycl_parallel_for(main_stream,
+                    sycl::nd_range<1>((num_blocks * sycl::range<1>(SYCL_SILU_BLOCK_SIZE)), sycl::range<1>(SYCL_SILU_BLOCK_SIZE)), [=](sycl::nd_item<1> item_ct1) {
+                gated_op_fused_swiglu(x_ptr, g_ptr, dst_ptr, k, n, o0, o1, item_ct1);
+            });
+        });
+}
+
+static inline void ggml_sycl_op_geglu_erf(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    ggml_sycl_detail::dispatch_ggml_sycl_op_fused_glu(ctx, dst,
+        [](const auto* x_ptr, const auto* g_ptr, auto* dst_ptr, uint64_t k, uint64_t n, uint64_t o0, uint64_t o1, queue_ptr main_stream) {
+            const uint32_t num_blocks = ceil_div(k, SYCL_GELU_BLOCK_SIZE);
+            sycl_parallel_for(main_stream,
+                    sycl::nd_range<1>((num_blocks * sycl::range<1>(SYCL_GELU_BLOCK_SIZE)), sycl::range<1>(SYCL_GELU_BLOCK_SIZE)), [=](sycl::nd_item<1> item_ct1) {
+                gated_op_fused_geglu_erf(x_ptr, g_ptr, dst_ptr, k, n, o0, o1, item_ct1);
+            });
+        });
+}
+
+static inline void ggml_sycl_op_geglu_quick(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    ggml_sycl_detail::dispatch_ggml_sycl_op_fused_glu(ctx, dst,
+        [](const auto* x_ptr, const auto* g_ptr, auto* dst_ptr, uint64_t k, uint64_t n, uint64_t o0, uint64_t o1, queue_ptr main_stream) {
+            const uint32_t num_blocks = ceil_div(k, SYCL_GELU_BLOCK_SIZE);
+            sycl_parallel_for(main_stream,
+                    sycl::nd_range<1>((num_blocks * sycl::range<1>(SYCL_GELU_BLOCK_SIZE)), sycl::range<1>(SYCL_GELU_BLOCK_SIZE)), [=](sycl::nd_item<1> item_ct1) {
+                gated_op_fused_geglu_quick(x_ptr, g_ptr, dst_ptr, k, n, o0, o1, item_ct1);
+            });
+        });
 }
 
 
@@ -1569,3 +1143,28 @@ void ggml_sycl_elu(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
     scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/1);
     ggml_sycl_op_elu(ctx, dst);
 }
+
+void ggml_sycl_geglu(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/1);
+    ggml_sycl_op_geglu(ctx, dst);
+}
+
+void ggml_sycl_reglu(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/1);
+    ggml_sycl_op_reglu(ctx, dst);
+}
+
+void ggml_sycl_swiglu(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/1);
+    ggml_sycl_op_swiglu(ctx, dst);
+}
+
+void ggml_sycl_geglu_erf(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/1);
+    ggml_sycl_op_geglu_erf(ctx, dst);
+}
+
+void ggml_sycl_geglu_quick(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/1);
+    ggml_sycl_op_geglu_quick(ctx, dst);
+}
diff --git a/ggml/src/ggml-sycl/element_wise.hpp b/ggml/src/ggml-sycl/element_wise.hpp
index bd40113f0..50749e87d 100644
--- a/ggml/src/ggml-sycl/element_wise.hpp
+++ b/ggml/src/ggml-sycl/element_wise.hpp
@@ -3,27 +3,30 @@
 
 #include "common.hpp"
 #include "ggml.h"
-#include <limits.h>
+#include <limits> // For std::numeric_limits
 
 template <typename T>
 T neg_infinity() {
     return -std::numeric_limits<T>::infinity();
 }
 
-template<typename T>
+template<typename T_Dst, typename T_Src = T_Dst>
 struct typed_data {
-    const T * src;
-    T * dst;
+    const T_Src * src;
+    T_Dst * dst;
 };
 
-template<typename T>
-typed_data<T> cast_data(ggml_tensor * dst) {
+template<typename T_Dst, typename T_Src = T_Dst>
+typed_data<T_Dst, T_Src> cast_data(ggml_tensor * dst) {
     return {
-        /* .src = */ static_cast<const T *>(dst->src[0]->data),
-        /* .dst = */ static_cast<T *>(dst->data)
+        /* .src = */ static_cast<const T_Src *>(dst->src[0]->data),
+        /* .dst = */ static_cast<T_Dst *>(dst->data)
     };
 }
 
+const float GELU_QUICK_COEF = -1.702f;
+
+
 void ggml_sycl_sqrt(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
 
 void ggml_sycl_sin(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
@@ -73,5 +76,11 @@ void ggml_sycl_sgn(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
 void ggml_sycl_abs(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
 
 void ggml_sycl_elu(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
-#endif // GGML_SYCL_ELEMENTWISE_HPP
 
+void ggml_sycl_geglu(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
+void ggml_sycl_reglu(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
+void ggml_sycl_swiglu(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
+void ggml_sycl_geglu_erf(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
+void ggml_sycl_geglu_quick(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
+
+#endif // GGML_SYCL_ELEMENTWISE_HPP
diff --git a/ggml/src/ggml-sycl/gemm.hpp b/ggml/src/ggml-sycl/gemm.hpp
index 6cbc7e0f6..5efe03d36 100644
--- a/ggml/src/ggml-sycl/gemm.hpp
+++ b/ggml/src/ggml-sycl/gemm.hpp
@@ -65,6 +65,9 @@ public:
 
         dnnl::primitive_attr primitive_attr;
         primitive_attr.set_scratchpad_mode(dnnl::scratchpad_mode::user);
+#ifdef GGML_SYCL_F16
+        primitive_attr.set_fpmath_mode(dnnl::fpmath_mode::f16);
+#endif
 
         auto a_mem = dnnl::memory(a_in_md, eng, const_cast<void*>(a));
         auto b_mem = dnnl::memory(b_in_md, eng, const_cast<void*>(b));
diff --git a/ggml/src/ggml-sycl/getrows.cpp b/ggml/src/ggml-sycl/getrows.cpp
index 4a7712781..9c76ffeb9 100644
--- a/ggml/src/ggml-sycl/getrows.cpp
+++ b/ggml/src/ggml-sycl/getrows.cpp
@@ -60,54 +60,6 @@ static void k_get_rows(
     dst_row[iybs + iqs + y_offset] = v.y();
 }
 
-template<int qk, int qr, dequantize_kernel_t_reorder dequantize_kernel_recorder, typename dst_t>
-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<typename src0_t, typename dst_t>
 static void k_get_rows_float(
             const src0_t * src0, const int32_t * src1, dst_t * dst,
@@ -166,58 +118,15 @@ static void get_rows_sycl(ggml_backend_sycl_context & ctx, const ggml_tensor *sr
 
     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<qk, qr, dq>(
-                                 src0_dd, src1_dd, dst_dd, ne00, ne12, s1, s2,
-                                 s3, nb01, nb02, nb03, s10, s11, s12, item_ct1);
-                         });
+    sycl_parallel_for(stream, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
+        k_get_rows<qk, qr, dq>(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 <int qk, int qr, dequantize_kernel_t_reorder dq_reorder>
-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<qk, qr, dq_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 <typename src0_t>
 static void get_rows_sycl_float(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
                                 const ggml_tensor *src1, ggml_tensor *dst,
@@ -245,9 +154,8 @@ static void get_rows_sycl_float(ggml_backend_sycl_context & ctx, const ggml_tens
         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_parallel_for(
+            stream, 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);
             });
@@ -277,13 +185,8 @@ void ggml_sycl_op_get_rows(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
             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<QK4_0, QR4_0, dequantize_q4_0_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<QK4_0, QR4_0, dequantize_q4_0>(ctx, dst->src[0], dst->src[1], dst, (const float *)dst->src[0]->data,
-                src1_i32, (float *)dst->data, ctx.stream());
-            }
+            get_rows_sycl<QK4_0, QR4_0, dequantize_q4_0>(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<QK4_1, QR4_1, dequantize_q4_1>(ctx, dst->src[0], dst->src[1], dst, (const float *)dst->src[0]->data,
diff --git a/ggml/src/ggml-sycl/ggml-sycl.cpp b/ggml/src/ggml-sycl/ggml-sycl.cpp
index bcd2ea536..21c81e99a 100644
--- a/ggml/src/ggml-sycl/ggml-sycl.cpp
+++ b/ggml/src/ggml-sycl/ggml-sycl.cpp
@@ -83,9 +83,7 @@ static ggml_sycl_device_info ggml_sycl_init() {
 
         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.devices[i].opt_feature.reorder = device.ext_oneapi_architecture_is(syclex::arch_category::intel_gpu);
         info.max_work_group_sizes[i] = prop.get_max_work_group_size();
     }
 
@@ -195,7 +193,7 @@ static void ggml_check_sycl() try {
 
     if (!initialized) {
         g_ggml_sycl_debug = get_sycl_env("GGML_SYCL_DEBUG", 0);
-        g_ggml_sycl_disable_optimize= get_sycl_env("GGML_SYCL_DISABLE_OPT", 1);
+        g_ggml_sycl_disable_optimize = get_sycl_env("GGML_SYCL_DISABLE_OPT", 0);
         g_ggml_sycl_disable_graph = get_sycl_env("GGML_SYCL_DISABLE_GRAPH", 1);
         g_ggml_sycl_disable_dnn = get_sycl_env("GGML_SYCL_DISABLE_DNN", 0);
         g_ggml_sycl_prioritize_dmmv = get_sycl_env("GGML_SYCL_PRIORITIZE_DMMV", 0);
@@ -347,14 +345,15 @@ static enum ggml_status
 ggml_backend_sycl_buffer_init_tensor(ggml_backend_buffer_t buffer,
                                      ggml_tensor *tensor) try {
     GGML_SYCL_DEBUG("[SYCL] call %s", __func__);
-    debug_print_tensor(": tensor=", tensor, "\n");
+    GGML_SYCL_DEBUG("%s", debug_get_tensor_str(": tensor", tensor, "\n").c_str());
     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 GGML_STATUS_SUCCESS;
     }
-    if ((tensor->type == GGML_TYPE_Q4_0 || tensor->type == GGML_TYPE_Q4_K) && !g_ggml_sycl_disable_optimize) {
+    if ((tensor->type == GGML_TYPE_Q4_0 || tensor->type == GGML_TYPE_Q4_K || tensor->type == GGML_TYPE_Q6_K) &&
+        !g_ggml_sycl_disable_optimize) {
         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.
@@ -384,7 +383,7 @@ static void ggml_backend_sycl_buffer_set_tensor(ggml_backend_buffer_t buffer,
                                                 const void *data, size_t offset,
                                                 size_t size) try {
     GGML_SYCL_DEBUG("[SYCL] call %s", __func__);
-    debug_print_tensor(": tensor=", tensor);
+    GGML_SYCL_DEBUG("%s", debug_get_tensor_str(": tensor", tensor).c_str());
     GGML_SYCL_DEBUG(" size=%zu offset=%zu\n", size, offset);
     ggml_backend_sycl_buffer_context * ctx = ( ggml_backend_sycl_buffer_context *)buffer->context;
     ggml_sycl_set_device(ctx->device);
@@ -412,7 +411,7 @@ static void ggml_backend_sycl_buffer_get_tensor(ggml_backend_buffer_t buffer,
                                                 void *data, size_t offset,
                                                 size_t size) try {
     GGML_SYCL_DEBUG("[SYCL] call %s", __func__);
-    debug_print_tensor(": tensor=", tensor);
+    GGML_SYCL_DEBUG("%s", debug_get_tensor_str(": tensor", tensor).c_str());
     GGML_SYCL_DEBUG(" size=%zu offset=%zu\n", size, offset);
     ggml_backend_sycl_buffer_context * ctx = ( ggml_backend_sycl_buffer_context *)buffer->context;
 
@@ -443,8 +442,8 @@ ggml_backend_sycl_buffer_cpy_tensor(ggml_backend_buffer_t buffer,
                                     ggml_tensor *dst) try {
     bool is_cpy_supported = ggml_backend_buffer_is_sycl(src->buffer);
     GGML_SYCL_DEBUG("[SYCL] call %s", __func__);
-    debug_print_tensor(": dst=", dst);
-    debug_print_tensor(" src=", src);
+    GGML_SYCL_DEBUG("%s", debug_get_tensor_str(": dst", dst).c_str());
+    GGML_SYCL_DEBUG("%s", debug_get_tensor_str(" src", src).c_str());
     GGML_SYCL_DEBUG(" is_cpy_supported=%d\n", is_cpy_supported);
     if (is_cpy_supported) {
         ggml_backend_sycl_buffer_context * src_ctx = (ggml_backend_sycl_buffer_context *)src->buffer->context;
@@ -524,7 +523,7 @@ catch (sycl::exception const &exc) {
 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", __func__);
-    debug_print_tensor(": tensor=", tensor);
+    GGML_SYCL_DEBUG("%s", debug_get_tensor_str(": tensor", tensor).c_str());
     GGML_SYCL_DEBUG(" size=%zu offset=%zu value=%u\n", size, offset, value);
     ggml_backend_sycl_buffer_context * ctx = (ggml_backend_sycl_buffer_context *) buffer->context;
     SYCL_CHECK(ggml_sycl_set_device(ctx->device));
@@ -804,7 +803,7 @@ static enum ggml_status
 ggml_backend_sycl_split_buffer_init_tensor(ggml_backend_buffer_t buffer,
                                            ggml_tensor *tensor) try {
     GGML_SYCL_DEBUG("[SYCL] call %s", __func__);
-    debug_print_tensor(": tensor=", tensor, "\n");
+    GGML_SYCL_DEBUG("%s", debug_get_tensor_str(": tensor", tensor, "\n").c_str());
     GGML_ASSERT(tensor->view_src == nullptr); // views of split tensors are not supported
 
     ggml_backend_sycl_split_buffer_context * ctx = (ggml_backend_sycl_split_buffer_context *)buffer->context;
@@ -890,7 +889,7 @@ ggml_backend_sycl_split_buffer_set_tensor(ggml_backend_buffer_t buffer,
                                           ggml_tensor *tensor, const void *data,
                                           size_t offset, size_t size) try {
     GGML_SYCL_DEBUG("[SYCL] call %s", __func__);
-    debug_print_tensor(": tensor=", tensor);
+    GGML_SYCL_DEBUG("%s", debug_get_tensor_str(": tensor", tensor).c_str());
     GGML_SYCL_DEBUG(" size=%zu offset=%zu\n", size, offset);
     // split tensors must always be set in their entirety at once
     GGML_ASSERT(offset == 0);
@@ -946,7 +945,7 @@ ggml_backend_sycl_split_buffer_get_tensor(ggml_backend_buffer_t buffer,
                                           const ggml_tensor *tensor, void *data,
                                           size_t offset, size_t size) try {
     GGML_SYCL_DEBUG("[SYCL] call %s", __func__);
-    debug_print_tensor(": tensor=", tensor);
+    GGML_SYCL_DEBUG("%s", debug_get_tensor_str(": tensor", tensor).c_str());
     GGML_SYCL_DEBUG(" size=%zu offset=%zu\n", size, offset);
     // split tensors must always be set in their entirety at once
     GGML_ASSERT(offset == 0);
@@ -1434,6 +1433,59 @@ static void quantize_q8_1(const float * __restrict__ x, void * __restrict__ vy,
     reinterpret_cast<sycl::half &>(y[ib].ds.y()) = sum;
 }
 
+template <int ElementsPerWI>
+static __dpct_inline__ void quantize_and_reorder_q8_1(const float * __restrict__ x, void * reordered_q8_tensor,
+                                                      const int kx, const int kx_padded, const sycl::nd_item<1> & it) {
+    /*
+        Quantizes and reorders the resultant q8 tensor in a per row fashion
+        Each sub-group calculates one quant block. i.e. QK8_1 quant values and the d and sum values
+    */
+
+    auto subgroup_id = it.get_group(0);
+    auto wi_id       = it.get_local_id(0);
+
+    const int num_blocks_per_row = kx / QK8_1;
+    auto      row                = subgroup_id / num_blocks_per_row;
+    auto      col                = subgroup_id % num_blocks_per_row;
+
+    auto row_offset = row * (kx_padded / QK8_1) * sizeof(block_q8_1);
+    auto col_offset = QK8_1 * col + wi_id * ElementsPerWI;
+
+    auto quant_ptr = (int8_t *) ((char *) reordered_q8_tensor + row_offset + col_offset);
+    auto ds_ptr    = (sycl::half2 *) ((char *) reordered_q8_tensor + row_offset + kx + col * sizeof(sycl::half2));
+
+    sycl::vec<float, ElementsPerWI>  wi_f32_vals;
+    sycl::vec<int8_t, ElementsPerWI> quantized_values;
+
+    auto float_ptr_offset = subgroup_id * QK8_1 + ElementsPerWI * wi_id;
+    wi_f32_vals           = *reinterpret_cast<const sycl::vec<float, ElementsPerWI> *>(x + float_ptr_offset);
+
+    float sum  = 0.0f;
+    float amax = 0.0f;
+
+#pragma unroll(ElementsPerWI)
+    for (int i = 0; i < ElementsPerWI; i++) {
+        sum += wi_f32_vals[i];
+        amax                = sycl::fmax(amax, sycl::fabs(wi_f32_vals[i]));
+        quantized_values[i] = 0;
+    }
+    sum     = sycl::reduce_over_group(it.get_group(), sum, sycl::plus<float>());
+    amax    = sycl::reduce_over_group(it.get_group(), amax, sycl::maximum<float>());
+    float d = amax == 0 ? 1 : amax / 127;
+
+#pragma unroll(ElementsPerWI)
+    for (int i = 0; i < ElementsPerWI; i++) {
+        quantized_values[i] = sycl::round(wi_f32_vals[i] / d);
+    }
+
+    d = amax == 0 ? 0 : d;
+
+    *reinterpret_cast<sycl::vec<int8_t, ElementsPerWI> *>(quant_ptr) = quantized_values;
+    if (wi_id == 0) {
+        *ds_ptr = sycl::half2(sycl::half(d), sycl::half(sum));
+    }
+}
+
 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,
@@ -1718,23 +1770,30 @@ static  void pool2d_nchw_kernel(
         o_ptr[cur_oh * ow + cur_ow] = res;
 }
 
-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) {
-    const int block_num_x = (kx_padded + SYCL_QUANTIZE_BLOCK_SIZE - 1) / SYCL_QUANTIZE_BLOCK_SIZE;
-    const sycl::range<3> num_blocks(1, ky, block_num_x);
-    int constexpr QUANT_BLOCK_TILE = QK8_1 / WARP_SIZE;
-    static_assert(QK8_1 % WARP_SIZE == 0);
-    const sycl::range<3> block_size(1, 1, SYCL_QUANTIZE_BLOCK_SIZE / QUANT_BLOCK_TILE);
-    {
-        dpct::has_capability_or_fail(stream->get_device(),
-                                     {sycl::aspect::fp16});
+static void quantize_row_q8_1_sycl(const float * x, void * vy, const int kx, const int ky, const int kx_padded,
+                                   bool reorder_q8_tensor, queue_ptr stream) {
+    if (reorder_q8_tensor) {
+        auto local_range      = std::size_t(WARP_SIZE);
+        auto num_quant_blocks = ky * (kx / QK8_1);
+        auto global_range     = num_quant_blocks * local_range;
+        stream->parallel_for(sycl::nd_range<1>({ global_range }, { local_range }),
+                             [=](sycl::nd_item<1> it) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                                 quantize_and_reorder_q8_1<QK8_1 / WARP_SIZE>(x, vy, kx, kx_padded, it);
+                             });
+    } else {
+        const int            block_num_x = (kx_padded + SYCL_QUANTIZE_BLOCK_SIZE - 1) / SYCL_QUANTIZE_BLOCK_SIZE;
+        const sycl::range<3> num_blocks(1, ky, block_num_x);
+        int constexpr QUANT_BLOCK_TILE = QK8_1 / WARP_SIZE;
+        static_assert(QK8_1 % WARP_SIZE == 0);
+        const sycl::range<3> block_size(1, 1, SYCL_QUANTIZE_BLOCK_SIZE / QUANT_BLOCK_TILE);
+        {
+            dpct::has_capability_or_fail(stream->get_device(), { sycl::aspect::fp16 });
 
-        stream->parallel_for(
-            sycl::nd_range<3>(num_blocks * block_size, block_size),
-            [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                quantize_q8_1<QUANT_BLOCK_TILE>(x, vy, kx, kx_padded, item_ct1);
-            });
+            stream->parallel_for(sycl::nd_range<3>(num_blocks * block_size, block_size),
+                                 [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                                     quantize_q8_1<QUANT_BLOCK_TILE>(x, vy, kx, kx_padded, item_ct1);
+                                 });
+        }
     }
 }
 
@@ -1826,13 +1885,12 @@ static void argsort_f32_i32_sycl(const float *x, int *dst, const int ncols,
     const size_t shared_mem = ncols_pad * sizeof(int);
 
     if (order == GGML_SORT_ORDER_ASC) {
-        stream->submit([&](sycl::handler &cgh) {
+        sycl_launch(stream, [&](sycl::handler & cgh) {
             sycl::local_accessor<uint8_t, 1> dpct_local_acc_ct1(
                 sycl::range<1>(shared_mem), cgh);
 
-            cgh.parallel_for(
-                sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                [=](sycl::nd_item<3> item_ct1) {
+            sycl_parallel_for(
+                cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
                     k_argsort_f32_i32<GGML_SORT_ORDER_ASC>(
                         x, dst, ncols, ncols_pad, item_ct1,
                         dpct_local_acc_ct1.get_multi_ptr<sycl::access::decorated::no>()
@@ -1840,13 +1898,12 @@ static void argsort_f32_i32_sycl(const float *x, int *dst, const int ncols,
                 });
         });
     } else if (order == GGML_SORT_ORDER_DESC) {
-        stream->submit([&](sycl::handler &cgh) {
+        sycl_launch(stream, [&](sycl::handler & cgh) {
             sycl::local_accessor<uint8_t, 1> dpct_local_acc_ct1(
                 sycl::range<1>(shared_mem), cgh);
 
-            cgh.parallel_for(
-                sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                [=](sycl::nd_item<3> item_ct1) {
+            sycl_parallel_for(
+                cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
                     k_argsort_f32_i32<GGML_SORT_ORDER_DESC>(
                         x, dst, ncols, ncols_pad, item_ct1,
                         dpct_local_acc_ct1.get_multi_ptr<sycl::access::decorated::no>()
@@ -1864,50 +1921,47 @@ static void argmax_f32_i32_sycl(const float *x, int *dst, const int ncols,
     const sycl::range<3> block_nums(1, nrows, 1);
     const size_t shared_mem = 256 * sizeof(float);
 
-    stream->submit([&](sycl::handler &cgh) {
+    sycl_launch(stream, [&](sycl::handler & cgh) {
         sycl::local_accessor<float, 1> shared_data(
             sycl::range<1>(shared_mem/sizeof(float)), cgh);
         sycl::local_accessor<int, 1> shared_indices(
             sycl::range<1>(shared_mem/sizeof(float)), cgh);
 
-        cgh.parallel_for(
-            sycl::nd_range<3>(block_nums * block_dims, block_dims),
-            [=](sycl::nd_item<3> item_ct1) {
-                const int tid = item_ct1.get_local_id(2);
-                const int row = item_ct1.get_global_id(1);
+        sycl_parallel_for(cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
+            const int tid = item_ct1.get_local_id(2);
+            const int row = item_ct1.get_global_id(1);
 
-                float max_val = -INFINITY;
-                int max_idx = -1;
+            float max_val = -INFINITY;
+            int   max_idx = -1;
 
-                for (int col = tid; col < ncols; col += 256) {
-                    float val = x[row * ncols + col];
-                    if (val > max_val) {
-                        max_val = val;
-                        max_idx = col;
+            for (int col = tid; col < ncols; col += 256) {
+                float val = x[row * ncols + col];
+                if (val > max_val) {
+                    max_val = val;
+                    max_idx = col;
+                }
+            }
+
+            shared_data[tid]    = max_val;
+            shared_indices[tid] = max_idx;
+            item_ct1.barrier(sycl::access::fence_space::local_space);
+
+            for (int stride = 256 / 2; stride > 0; stride >>= 1) {
+                if (tid < stride) {
+                    float val1 = shared_data[tid];
+                    float val2 = shared_data[tid + stride];
+                    if (val2 > val1) {
+                        shared_data[tid]    = val2;
+                        shared_indices[tid] = shared_indices[tid + stride];
                     }
                 }
-
-                shared_data[tid] = max_val;
-                shared_indices[tid] = max_idx;
                 item_ct1.barrier(sycl::access::fence_space::local_space);
+            }
 
-                for (int stride = 256/2; stride > 0; stride >>= 1) {
-                    if (tid < stride) {
-                        float val1 = shared_data[tid];
-                        float val2 = shared_data[tid + stride];
-                        if (val2 > val1) {
-                            shared_data[tid] = val2;
-                            shared_indices[tid] = shared_indices[tid + stride];
-                        }
-                    }
-                    item_ct1.barrier(sycl::access::fence_space::local_space);
-                }
-
-
-                if (tid == 0) {
-                    dst[row] = shared_indices[0];
-                }
-            });
+            if (tid == 0) {
+                dst[row] = shared_indices[0];
+            }
+        });
     });
 }
 static void diag_mask_inf_f32_sycl(const float *x, float *dst,
@@ -2066,21 +2120,18 @@ inline void ggml_sycl_op_mul_mat_sycl(
         const sycl::half *src1_ptr = src1->type == GGML_TYPE_F16
                 ? (const sycl::half *)src1->data + src1_padded_row_size
                                          : src1_as_f16.get();
-        ggml_sycl_pool_alloc<sycl::half> dst_f16(ctx.pool(), row_diff * src1_ncols);
 
 #if GGML_SYCL_DNNL
         if (!g_ggml_sycl_disable_dnn) {
             DnnlGemmWrapper::row_gemm(ctx, src1_ncols, row_diff, ne10, src1_ptr,
                                       DnnlGemmWrapper::to_dt<sycl::half>(), src0_ptr, DnnlGemmWrapper::to_dt<sycl::half>(),
-                                      dst_f16.get(), DnnlGemmWrapper::to_dt<sycl::half>(), stream);
-            scope_op_debug_print scope_dbg_print(__func__, "/to_fp32_sycl", dst, /*num_src=*/2,
-                                                 " : converting dst to fp32");
-            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);
+                                      dst_dd_i, DnnlGemmWrapper::to_dt<float>(), stream);
         }
         else
 #endif
         {
+            ggml_sycl_pool_alloc<sycl::half> dst_f16(ctx.pool(), row_diff * src1_ncols);
+
             const sycl::half alpha_f16 = 1.0f;
             const sycl::half beta_f16  = 0.0f;
             SYCL_CHECK(CHECK_TRY_ERROR(dpct::gemm(
@@ -2446,9 +2497,10 @@ static void ggml_sycl_op_mul_mat(ggml_backend_sycl_context & ctx, const ggml_ten
             dev[i].src1_ddq = dev[i].src1_ddq_alloc.alloc(ctx.pool(i), nrows1*src1_padded_col_size*q8_1_ts/q8_1_bs);
 
             if (src1_on_device && src1_is_contiguous) {
+                bool reorder_q8_tensor = src0->extra && ((ggml_tensor_extra_gpu *)src0->extra)->optimized_feature.reorder;
                 scope_op_debug_print scope_dbg_print(__func__, "/quantize_row_q8_1_sycl", dst,
                                                      /*num_src=*/2, " : converting src1 to Q8_1");
-                quantize_row_q8_1_sycl(dev[i].src1_ddf, dev[i].src1_ddq, ne10, nrows1, src1_padded_col_size, stream);
+                quantize_row_q8_1_sycl(dev[i].src1_ddf, dev[i].src1_ddq, ne10, nrows1, src1_padded_col_size, reorder_q8_tensor, stream);
                 /*
                 DPCT1010:90: SYCL uses exceptions to report errors and does not
                 use the error codes. The call was replaced with 0. You need to
@@ -2554,7 +2606,7 @@ static void ggml_sycl_op_mul_mat(ggml_backend_sycl_context & ctx, const ggml_ten
                 if (convert_src1_to_q8_1 && !src1_is_contiguous) {
                     scope_op_debug_print scope_dbg_print(__func__, "/quantize_row_q8_1_sycl", dst,
                                                          /*num_src=*/2, " : converting src1 to Q8_1");
-                    quantize_row_q8_1_sycl(src1_ddf_i, src1_ddq_i, ne10, src1_ncols, src1_padded_col_size, stream);
+                    quantize_row_q8_1_sycl(src1_ddf_i, src1_ddq_i, ne10, src1_ncols, src1_padded_col_size, false, stream);
                     /*
                     DPCT1010:92: SYCL uses exceptions to report errors and does
                     not use the error codes. The call was replaced with 0. You
@@ -2893,7 +2945,7 @@ static void ggml_sycl_mul_mat_batched_sycl(ggml_backend_sycl_context & ctx, cons
                 void **       ptrs_dst_get = ptrs_dst.get();
                 size_t        nb12_scaled  = src1->type == GGML_TYPE_F16 ? nb12 : s12 * sizeof(sycl::half);
                 size_t        nb13_scaled  = src1->type == GGML_TYPE_F16 ? nb13 : s13 * sizeof(sycl::half);
-                cgh.parallel_for(sycl::nd_range<3>(block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
+                sycl_parallel_for(cgh, sycl::nd_range<3>(block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
                     k_compute_batched_ptrs(src0_f16, src1_f16, dst_ddf, ptrs_src_get, ptrs_dst_get, ne12, ne13, ne23, nb02,
                                            nb03, nb12_scaled, nb13_scaled, nbd2, nbd3, r2, r3, item_ct1);
                 });
@@ -2928,6 +2980,7 @@ inline bool ggml_sycl_supports_reorder_mul_mat_sycl(enum ggml_type type) {
         case GGML_TYPE_Q4_0:
             return true;
         case GGML_TYPE_Q4_K:
+        case GGML_TYPE_Q6_K:
             return !g_ggml_sycl_prioritize_dmmv;
         default:
             return false;
@@ -2947,6 +3000,7 @@ inline bool ggml_sycl_supports_reorder_mmvq(enum ggml_type type) {
     switch (type) {
         case GGML_TYPE_Q4_0:
         case GGML_TYPE_Q4_K:
+        case GGML_TYPE_Q6_K:
             return true;
         default:
             return false;
@@ -3031,6 +3085,50 @@ static void reorder_qw_q4_k(uint8_t * data_device, size_t size, size_t offset, d
     sycl::free(tmp_buf, *stream);
 }
 
+static void reorder_qw_q6_k(uint8_t * data_device, size_t size, size_t offset, dpct::queue_ptr stream) {
+    GGML_ASSERT(size % sizeof(block_q6_K) == 0);
+    GGML_ASSERT(offset % sizeof(block_q6_K) == 0);
+
+    const int nblocks = size / sizeof(block_q6_K);
+
+    auto * tmp_buf = sycl::malloc_shared<uint8_t>(size, *stream);
+    SYCL_CHECK(CHECK_TRY_ERROR((*stream).memcpy(tmp_buf, data_device, size).wait()));
+
+    auto *       ql_ptr     = data_device;
+    auto *       qh_ptr     = ql_ptr + (QK_K / 2) * nblocks;
+    auto *       scales_ptr = qh_ptr + (QK_K / 4) * nblocks;
+    sycl::half * dm_ptr     = (sycl::half *) (scales_ptr + (QK_K / 16) * nblocks);
+
+    stream
+        ->parallel_for(nblocks,
+                       [=](auto i) {
+                           const block_q6_K * x  = (const block_q6_K *) tmp_buf;
+                           const int          ib = i;
+
+                           const uint8_t * ql              = x[ib].ql;
+                           const uint8_t * qh              = x[ib].qh;
+                           uint8_t *       base_ql_ptr     = ql_ptr + (QK_K / 2) * ib;
+                           uint8_t *       base_qh_ptr     = qh_ptr + (QK_K / 4) * ib;
+                           uint8_t *       base_scales_ptr = scales_ptr + (QK_K / 16) * ib;
+
+                           for (int j = 0; j < QK_K / 2; ++j) {
+                               base_ql_ptr[j] = ql[j];
+                           }
+                           for (int j = 0; j < QK_K / 4; ++j) {
+                               base_qh_ptr[j] = qh[j];
+                           }
+
+                           for (int j = 0; j < QK_K / 16; ++j) {
+                               base_scales_ptr[j] = x[ib].scales[j];
+                           }
+
+                           dm_ptr[ib] = x[ib].d;
+                       })
+        .wait_and_throw();
+
+    sycl::free(tmp_buf, *stream);
+}
+
 static void reorder_qw(const ggml_tensor * src0, dpct::queue_ptr stream) {
     uint8_t * data_device = (uint8_t *) src0->data;
     size_t ncols = src0->ne[0];
@@ -3044,6 +3142,9 @@ static void reorder_qw(const ggml_tensor * src0, dpct::queue_ptr stream) {
         case GGML_TYPE_Q4_K:
             reorder_qw_q4_k(data_device, size, 0, stream);
             break;
+        case GGML_TYPE_Q6_K:
+            reorder_qw_q6_k(data_device, size, 0, stream);
+            break;
         default:
             GGML_ABORT("reorder_qw() called with unsupported type");
             break;
@@ -3348,7 +3449,7 @@ static void ggml_sycl_mul_mat_id(ggml_backend_sycl_context & ctx,
             {
                 sycl::range<3> block_dims(1, 1, std::min((unsigned int)ne10, 768u));
                 sycl::range<3> grid_dims(1, n_ids, ids->ne[1]);
-                stream->submit([&](sycl::handler &cgh) {
+                sycl_launch(stream, [&](sycl::handler & cgh) {
                     sycl::local_accessor<int, 0> src1_row_acc(cgh);
 
                     char *__restrict src1_contiguous_get =
@@ -3360,9 +3461,8 @@ static void ggml_sycl_mul_mat_id(ggml_backend_sycl_context & ctx,
                     size_t ids_nb_ct6 = ids->nb[1];
                     size_t ids_nb_ct7 = ids->nb[0];
 
-                    cgh.parallel_for(
-                        sycl::nd_range<3>(grid_dims * block_dims, block_dims),
-                        [=](sycl::nd_item<3> item_ct1) {
+                    sycl_parallel_for(
+                        cgh, sycl::nd_range<3>(grid_dims * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
                             k_copy_src1_to_contiguous(
                                 src1_original, src1_contiguous_get,
                                 dev_cur_src1_row_get,
@@ -3393,15 +3493,14 @@ static void ggml_sycl_mul_mat_id(ggml_backend_sycl_context & ctx,
             {
                 sycl::range<3> block_dims(1, 1, std::min((unsigned int)ne0, 768u));
                 sycl::range<3> grid_dims(1, 1, num_src1_rows);
-                stream->submit([&](sycl::handler &cgh) {
+                sycl_launch(stream, [&](sycl::handler & cgh) {
                     const char *__restrict dst_contiguous_get =
                         dst_contiguous.get();
                     const mmid_row_mapping *__restrict dev_row_mapping_get =
                         dev_row_mapping.get();
 
-                    cgh.parallel_for(
-                        sycl::nd_range<3>(grid_dims * block_dims, block_dims),
-                        [=](sycl::nd_item<3> item_ct1) {
+                    sycl_parallel_for(
+                        cgh, sycl::nd_range<3>(grid_dims * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
                             k_copy_dst_from_contiguous(dst_original,
                                                        dst_contiguous_get,
                                                        dev_row_mapping_get,
@@ -3577,6 +3676,27 @@ static bool ggml_sycl_compute_forward(ggml_backend_sycl_context & ctx, struct gg
                     return false;
             }
             break;
+        case GGML_OP_GLU:
+            switch (ggml_get_glu_op(dst)) {
+                case GGML_GLU_OP_REGLU:
+                    ggml_sycl_reglu(ctx, dst);
+                    break;
+                case GGML_GLU_OP_GEGLU:
+                    ggml_sycl_geglu(ctx, dst);
+                    break;
+                case GGML_GLU_OP_SWIGLU:
+                    ggml_sycl_swiglu(ctx, dst);
+                    break;
+                case GGML_GLU_OP_GEGLU_ERF:
+                    ggml_sycl_geglu_erf(ctx, dst);
+                    break;
+                case GGML_GLU_OP_GEGLU_QUICK:
+                    ggml_sycl_geglu_quick(ctx, dst);
+                    break;
+                default:
+                    return false;
+            }
+            break;
         case GGML_OP_NORM:
             ggml_sycl_norm(ctx, dst);
             break;
@@ -3755,7 +3875,7 @@ static void ggml_backend_sycl_set_tensor_async(ggml_backend_t backend,
                                                const void *data, size_t offset,
                                                size_t size) try {
     GGML_SYCL_DEBUG("[SYCL] call %s", __func__);
-    debug_print_tensor(": tensor=", tensor);
+    GGML_SYCL_DEBUG("%s", debug_get_tensor_str(": tensor", tensor).c_str());
     GGML_SYCL_DEBUG(" size=%zu offset=%zu\n", size, offset);
     ggml_backend_sycl_context * sycl_ctx = (ggml_backend_sycl_context *)backend->context;
     ggml_backend_buffer_t buf = tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
@@ -3776,7 +3896,7 @@ static void ggml_backend_sycl_get_tensor_async(ggml_backend_t backend,
                                                void *data, size_t offset,
                                                size_t size) try {
     GGML_SYCL_DEBUG("[SYCL] call %s", __func__);
-    debug_print_tensor(": tensor=", tensor);
+    GGML_SYCL_DEBUG("%s", debug_get_tensor_str(": tensor", tensor).c_str());
     GGML_SYCL_DEBUG(" size=%zu offset=%zu\n", size, offset);
     ggml_backend_sycl_context * sycl_ctx = (ggml_backend_sycl_context *)backend->context;
     ggml_backend_buffer_t buf = tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
@@ -3799,8 +3919,8 @@ static bool ggml_backend_sycl_cpy_tensor_async(ggml_backend_t backend,
     bool is_cpy_supported                = dst->buffer->buft == ggml_backend_sycl_buffer_type(sycl_ctx->device) &&
                             ggml_backend_buffer_is_sycl(src->buffer);
     GGML_SYCL_DEBUG("[SYCL] call %s", __func__);
-    debug_print_tensor(": dst=", dst);
-    debug_print_tensor(" src=", src);
+    GGML_SYCL_DEBUG("%s", debug_get_tensor_str(": dst", dst).c_str());
+    GGML_SYCL_DEBUG("%s", debug_get_tensor_str(" src", src).c_str());
     GGML_SYCL_DEBUG(" is_cpy_supported=%d\n", is_cpy_supported);
     if (is_cpy_supported) {
         /*
@@ -4113,6 +4233,18 @@ static bool ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, const g
                 default:
                     return false;
             }
+        case GGML_OP_GLU:
+            switch (ggml_get_glu_op(op)) {
+                case GGML_GLU_OP_REGLU:
+                case GGML_GLU_OP_GEGLU:
+                case GGML_GLU_OP_SWIGLU:
+                case GGML_GLU_OP_GEGLU_ERF:
+                case GGML_GLU_OP_GEGLU_QUICK:
+                    return ggml_is_contiguous_1(op->src[0]);
+                default:
+                    return false;
+            }
+            break;
         case GGML_OP_MUL_MAT:
         case GGML_OP_MUL_MAT_ID:
             {
@@ -4161,10 +4293,19 @@ static bool ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, const g
                         return false;
                 }
             }
+        case GGML_OP_SET_ROWS:
+            {
+                // TODO: add support
+                // ref: https://github.com/ggml-org/llama.cpp/pull/14274
+                return false;
+            } break;
         case GGML_OP_CPY:
             {
                 ggml_type src0_type = op->src[0]->type;
                 ggml_type src1_type = op->src[1]->type;
+                if (src0_type == src1_type && (ggml_is_contiguous(op->src[0]) && ggml_is_contiguous(op->src[1])) && src0_type != GGML_TYPE_BF16) {
+                    return true;
+                }
                 if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_F32) {
                     return true;
                 }
@@ -4210,6 +4351,21 @@ static bool ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, const g
                 if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_IQ4_NL) {
                     return true;
                 }
+                if(src0_type == GGML_TYPE_Q8_0 && src1_type == GGML_TYPE_Q8_0) {
+                    return true;
+                }
+                if(src0_type == GGML_TYPE_Q5_0 && src1_type == GGML_TYPE_Q5_0) {
+                    return true;
+                }
+                if(src0_type == GGML_TYPE_Q5_1 && src1_type == GGML_TYPE_Q5_1) {
+                    return true;
+                }
+                if(src0_type == GGML_TYPE_Q4_0 && src1_type == GGML_TYPE_Q4_0) {
+                    return true;
+                }
+                if(src0_type == GGML_TYPE_Q4_1 && src1_type == GGML_TYPE_Q4_1) {
+                    return true;
+                }
                 return false;
             }
         case GGML_OP_CONCAT:
@@ -4253,9 +4409,15 @@ static bool ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, const g
             return true;
         case GGML_OP_CONT:
             return op->src[0]->type != GGML_TYPE_BF16;
-        case GGML_OP_DIAG_MASK_INF:
         case GGML_OP_SOFT_MAX:
-            return true;
+            // TODO: support batching
+            if (op->src[0]->ne[3] != 1) {
+                return false;
+            }
+            // TODO: support broadcast
+            // ref: https://github.com/ggml-org/llama.cpp/pull/14435
+            return !op->src[1] || (op->src[1]->ne[2] == 1 && op->src[1]->ne[3] == 1);
+        case GGML_OP_DIAG_MASK_INF:
         case GGML_OP_ROPE:
         case GGML_OP_IM2COL:
             return true;
diff --git a/ggml/src/ggml-sycl/gla.cpp b/ggml/src/ggml-sycl/gla.cpp
index 879184fdd..b40cbf1f1 100644
--- a/ggml/src/ggml-sycl/gla.cpp
+++ b/ggml/src/ggml-sycl/gla.cpp
@@ -11,13 +11,13 @@ static void gated_linear_attn_f32_kernel(const dpct::queue_ptr stream, u_int B,
     const u_int n_seq_tokens = T / B;
     sycl::range<1> block_dims((C / H));
     sycl::range<1> grid_dims((B * H));
-    stream->submit([&](sycl::handler & cgh) {
+    sycl_launch(stream, [&](sycl::handler & cgh) {
         /* local memory accessors*/
         auto _k  = sycl::local_accessor<float, 1>(sycl::range<1>(head_size), cgh);
         auto _r  = sycl::local_accessor<float, 1>(sycl::range<1>(head_size), cgh);
         auto _td = sycl::local_accessor<float, 1>(sycl::range<1>(head_size), cgh);
 
-        cgh.parallel_for(sycl::nd_range<1>(grid_dims * block_dims, block_dims), [=](sycl::nd_item<1> item) {
+        sycl_parallel_for<1>(cgh, sycl::nd_range<1>(grid_dims * block_dims, block_dims), [=](sycl::nd_item<1> item) {
             u_int tid = item.get_local_id(0);
             u_int bid = item.get_group(0);
 
diff --git a/ggml/src/ggml-sycl/im2col.cpp b/ggml/src/ggml-sycl/im2col.cpp
index aa19c2527..52737cc74 100644
--- a/ggml/src/ggml-sycl/im2col.cpp
+++ b/ggml/src/ggml-sycl/im2col.cpp
@@ -70,7 +70,7 @@ static void im2col_sycl_internal(const float * x, T * dst, int64_t IW, int64_t I
 
     const int64_t CHW = IC * KH * KW;
 
-    stream->parallel_for(sycl::nd_range<3>(block_nums * local_range, local_range), [=](sycl::nd_item<3> item_ct1) {
+    sycl_parallel_for(stream, sycl::nd_range<3>(block_nums * local_range, local_range), [=](sycl::nd_item<3> item_ct1) {
         im2col_kernel<T>(x, dst, batch_offset, offset_delta, IC, IW, IH, OH, OW, KW, KH, parallel_elements, CHW, s0, s1,
                          p0, p1, d0, d1, item_ct1);
     });
diff --git a/ggml/src/ggml-sycl/mmq.cpp b/ggml/src/ggml-sycl/mmq.cpp
index ffb272aa2..c72fcd38e 100644
--- a/ggml/src/ggml-sycl/mmq.cpp
+++ b/ggml/src/ggml-sycl/mmq.cpp
@@ -1818,7 +1818,7 @@ static void ggml_mul_mat_q4_0_q8_1_sycl(const void *vx, const void *vy,
             dpct::has_capability_or_fail(stream->get_device(),
                                          {sycl::aspect::fp16});
 
-            stream->submit([&](sycl::handler &cgh) {
+            sycl_launch(stream, [&](sycl::handler & cgh) {
                 sycl::local_accessor<int, 1> tile_x_qs_q4_0_acc_ct1(
                     sycl::range<1>(mmq_y * (WARP_SIZE) + mmq_y), cgh);
                 sycl::local_accessor<float, 1> tile_x_d_q4_0_acc_ct1(
@@ -1829,9 +1829,8 @@ static void ggml_mul_mat_q4_0_q8_1_sycl(const void *vx, const void *vy,
                 sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
                     sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
 
-                cgh.parallel_for(
-                    sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                    [=](sycl::nd_item<3> item_ct1) {
+                sycl_parallel_for(
+                    cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
                         mul_mat_q4_0<need_check>(
                             vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
                             nrows_dst, item_ct1,
@@ -1853,7 +1852,7 @@ static void ggml_mul_mat_q4_0_q8_1_sycl(const void *vx, const void *vy,
             dpct::has_capability_or_fail(stream->get_device(),
                                          {sycl::aspect::fp16});
 
-            stream->submit([&](sycl::handler &cgh) {
+            sycl_launch(stream, [&](sycl::handler & cgh) {
                 sycl::local_accessor<int, 1> tile_x_qs_q4_0_acc_ct1(
                     sycl::range<1>(mmq_y * (WARP_SIZE) + mmq_y), cgh);
                 sycl::local_accessor<float, 1> tile_x_d_q4_0_acc_ct1(
@@ -1864,9 +1863,8 @@ static void ggml_mul_mat_q4_0_q8_1_sycl(const void *vx, const void *vy,
                 sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
                     sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
 
-                cgh.parallel_for(
-                    sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                    [=](sycl::nd_item<3> item_ct1) {
+                sycl_parallel_for(
+                    cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
                         mul_mat_q4_0<need_check>(
                             vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
                             nrows_dst, item_ct1,
@@ -1933,7 +1931,7 @@ static void ggml_mul_mat_q4_1_q8_1_sycl(const void *vx, const void *vy,
             dpct::has_capability_or_fail(stream->get_device(),
                                          {sycl::aspect::fp16});
 
-            stream->submit([&](sycl::handler &cgh) {
+            sycl_launch(stream, [&](sycl::handler & cgh) {
                 sycl::local_accessor<int, 1> tile_x_qs_q4_1_acc_ct1(
                     sycl::range<1>(mmq_y * (WARP_SIZE) + +mmq_y), cgh);
                 sycl::local_accessor<sycl::half2, 1> tile_x_dm_q4_1_acc_ct1(
@@ -1944,9 +1942,8 @@ static void ggml_mul_mat_q4_1_q8_1_sycl(const void *vx, const void *vy,
                 sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
                     sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
 
-                cgh.parallel_for(
-                    sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                    [=](sycl::nd_item<3> item_ct1) {
+                sycl_parallel_for(
+                    cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
                         mul_mat_q4_1<need_check>(
                             vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
                             nrows_dst, item_ct1,
@@ -1968,7 +1965,7 @@ static void ggml_mul_mat_q4_1_q8_1_sycl(const void *vx, const void *vy,
             dpct::has_capability_or_fail(stream->get_device(),
                                          {sycl::aspect::fp16});
 
-            stream->submit([&](sycl::handler &cgh) {
+            sycl_launch(stream, [&](sycl::handler & cgh) {
                 sycl::local_accessor<int, 1> tile_x_qs_q4_1_acc_ct1(
                     sycl::range<1>(mmq_y * (WARP_SIZE) + +mmq_y), cgh);
                 sycl::local_accessor<sycl::half2, 1> tile_x_dm_q4_1_acc_ct1(
@@ -1979,9 +1976,8 @@ static void ggml_mul_mat_q4_1_q8_1_sycl(const void *vx, const void *vy,
                 sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
                     sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
 
-                cgh.parallel_for(
-                    sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                    [=](sycl::nd_item<3> item_ct1) {
+                sycl_parallel_for(
+                    cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
                         mul_mat_q4_1<need_check>(
                             vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
                             nrows_dst, item_ct1,
@@ -2048,7 +2044,7 @@ static void ggml_mul_mat_q5_0_q8_1_sycl(const void *vx, const void *vy,
             dpct::has_capability_or_fail(stream->get_device(),
                                          {sycl::aspect::fp16});
 
-            stream->submit([&](sycl::handler &cgh) {
+            sycl_launch(stream, [&](sycl::handler & cgh) {
                 sycl::local_accessor<int, 1> tile_x_ql_q5_0_acc_ct1(
                     sycl::range<1>(mmq_y * (2 * WARP_SIZE) + mmq_y), cgh);
                 sycl::local_accessor<float, 1> tile_x_d_q5_0_acc_ct1(
@@ -2059,9 +2055,8 @@ static void ggml_mul_mat_q5_0_q8_1_sycl(const void *vx, const void *vy,
                 sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
                     sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
 
-                cgh.parallel_for(
-                    sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                    [=](sycl::nd_item<3> item_ct1) {
+                sycl_parallel_for(
+                    cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
                         mul_mat_q5_0<need_check>(
                             vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
                             nrows_dst, item_ct1,
@@ -2083,7 +2078,7 @@ static void ggml_mul_mat_q5_0_q8_1_sycl(const void *vx, const void *vy,
             dpct::has_capability_or_fail(stream->get_device(),
                                          {sycl::aspect::fp16});
 
-            stream->submit([&](sycl::handler &cgh) {
+            sycl_launch(stream, [&](sycl::handler & cgh) {
                 sycl::local_accessor<int, 1> tile_x_ql_q5_0_acc_ct1(
                     sycl::range<1>(mmq_y * (2 * WARP_SIZE) + mmq_y), cgh);
                 sycl::local_accessor<float, 1> tile_x_d_q5_0_acc_ct1(
@@ -2094,9 +2089,8 @@ static void ggml_mul_mat_q5_0_q8_1_sycl(const void *vx, const void *vy,
                 sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
                     sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
 
-                cgh.parallel_for(
-                    sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                    [=](sycl::nd_item<3> item_ct1) {
+                sycl_parallel_for(
+                    cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
                         mul_mat_q5_0<need_check>(
                             vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
                             nrows_dst, item_ct1,
@@ -2163,7 +2157,7 @@ static void ggml_mul_mat_q5_1_q8_1_sycl(const void *vx, const void *vy,
             dpct::has_capability_or_fail(stream->get_device(),
                                          {sycl::aspect::fp16});
 
-            stream->submit([&](sycl::handler &cgh) {
+            sycl_launch(stream, [&](sycl::handler & cgh) {
                 sycl::local_accessor<int, 1> tile_x_ql_q5_1_acc_ct1(
                     sycl::range<1>(mmq_y * (2 * WARP_SIZE) + mmq_y), cgh);
                 sycl::local_accessor<sycl::half2, 1> tile_x_dm_q5_1_acc_ct1(
@@ -2174,9 +2168,8 @@ static void ggml_mul_mat_q5_1_q8_1_sycl(const void *vx, const void *vy,
                 sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
                     sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
 
-                cgh.parallel_for(
-                    sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                    [=](sycl::nd_item<3> item_ct1) {
+                sycl_parallel_for(
+                    cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
                         mul_mat_q5_1<need_check>(
                             vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
                             nrows_dst, item_ct1,
@@ -2198,7 +2191,7 @@ static void ggml_mul_mat_q5_1_q8_1_sycl(const void *vx, const void *vy,
             dpct::has_capability_or_fail(stream->get_device(),
                                          {sycl::aspect::fp16});
 
-            stream->submit([&](sycl::handler &cgh) {
+            sycl_launch(stream, [&](sycl::handler & cgh) {
                 sycl::local_accessor<int, 1> tile_x_ql_q5_1_acc_ct1(
                     sycl::range<1>(mmq_y * (2 * WARP_SIZE) + mmq_y), cgh);
                 sycl::local_accessor<sycl::half2, 1> tile_x_dm_q5_1_acc_ct1(
@@ -2209,9 +2202,8 @@ static void ggml_mul_mat_q5_1_q8_1_sycl(const void *vx, const void *vy,
                 sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
                     sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
 
-                cgh.parallel_for(
-                    sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                    [=](sycl::nd_item<3> item_ct1) {
+                sycl_parallel_for(
+                    cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
                         mul_mat_q5_1<need_check>(
                             vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
                             nrows_dst, item_ct1,
@@ -2278,7 +2270,7 @@ static void ggml_mul_mat_q8_0_q8_1_sycl(const void *vx, const void *vy,
             dpct::has_capability_or_fail(stream->get_device(),
                                          {sycl::aspect::fp16});
 
-            stream->submit([&](sycl::handler &cgh) {
+            sycl_launch(stream, [&](sycl::handler & cgh) {
                 sycl::local_accessor<int, 1> tile_x_qs_q8_0_acc_ct1(
                     sycl::range<1>(mmq_y * (WARP_SIZE) + mmq_y), cgh);
                 sycl::local_accessor<float, 1> tile_x_d_q8_0_acc_ct1(
@@ -2289,9 +2281,8 @@ static void ggml_mul_mat_q8_0_q8_1_sycl(const void *vx, const void *vy,
                 sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
                     sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
 
-                cgh.parallel_for(
-                    sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                    [=](sycl::nd_item<3> item_ct1) {
+                sycl_parallel_for(
+                    cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
                         mul_mat_q8_0<need_check>(
                             vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
                             nrows_dst, item_ct1,
@@ -2313,7 +2304,7 @@ static void ggml_mul_mat_q8_0_q8_1_sycl(const void *vx, const void *vy,
             dpct::has_capability_or_fail(stream->get_device(),
                                          {sycl::aspect::fp16});
 
-            stream->submit([&](sycl::handler &cgh) {
+            sycl_launch(stream, [&](sycl::handler & cgh) {
                 sycl::local_accessor<int, 1> tile_x_qs_q8_0_acc_ct1(
                     sycl::range<1>(mmq_y * (WARP_SIZE) + mmq_y), cgh);
                 sycl::local_accessor<float, 1> tile_x_d_q8_0_acc_ct1(
@@ -2324,9 +2315,8 @@ static void ggml_mul_mat_q8_0_q8_1_sycl(const void *vx, const void *vy,
                 sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
                     sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
 
-                cgh.parallel_for(
-                    sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                    [=](sycl::nd_item<3> item_ct1) {
+                sycl_parallel_for(
+                    cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
                         mul_mat_q8_0<need_check>(
                             vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
                             nrows_dst, item_ct1,
@@ -2393,7 +2383,7 @@ static void ggml_mul_mat_q2_K_q8_1_sycl(const void *vx, const void *vy,
             dpct::has_capability_or_fail(stream->get_device(),
                                          {sycl::aspect::fp16});
 
-            stream->submit([&](sycl::handler &cgh) {
+            sycl_launch(stream, [&](sycl::handler & cgh) {
                 sycl::local_accessor<int, 1> tile_x_ql_q2_K_acc_ct1(
                     sycl::range<1>(mmq_y * (WARP_SIZE) + mmq_y), cgh);
                 sycl::local_accessor<sycl::half2, 1> tile_x_dm_q2_K_acc_ct1(
@@ -2406,9 +2396,8 @@ static void ggml_mul_mat_q2_K_q8_1_sycl(const void *vx, const void *vy,
                 sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
                     sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
 
-                cgh.parallel_for(
-                    sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                    [=](sycl::nd_item<3> item_ct1) {
+                sycl_parallel_for(
+                    cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
                         mul_mat_q2_K<need_check>(
                             vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
                             nrows_dst, item_ct1,
@@ -2431,7 +2420,7 @@ static void ggml_mul_mat_q2_K_q8_1_sycl(const void *vx, const void *vy,
             dpct::has_capability_or_fail(stream->get_device(),
                                          {sycl::aspect::fp16});
 
-            stream->submit([&](sycl::handler &cgh) {
+            sycl_launch(stream, [&](sycl::handler & cgh) {
                 sycl::local_accessor<int, 1> tile_x_ql_q2_K_acc_ct1(
                     sycl::range<1>(mmq_y * (WARP_SIZE) + mmq_y), cgh);
                 sycl::local_accessor<sycl::half2, 1> tile_x_dm_q2_K_acc_ct1(
@@ -2444,9 +2433,8 @@ static void ggml_mul_mat_q2_K_q8_1_sycl(const void *vx, const void *vy,
                 sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
                     sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
 
-                cgh.parallel_for(
-                    sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                    [=](sycl::nd_item<3> item_ct1) {
+                sycl_parallel_for(
+                    cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
                         mul_mat_q2_K<need_check>(
                             vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
                             nrows_dst, item_ct1,
@@ -2516,7 +2504,7 @@ static void ggml_mul_mat_q3_K_q8_1_sycl(const void *vx, const void *vy,
             dpct::has_capability_or_fail(stream->get_device(),
                                          {sycl::aspect::fp16});
 
-            stream->submit([&](sycl::handler &cgh) {
+            sycl_launch(stream, [&](sycl::handler & cgh) {
                 sycl::local_accessor<int, 1> tile_x_ql_q3_K_acc_ct1(
                     sycl::range<1>(mmq_y * (WARP_SIZE) + mmq_y), cgh);
                 sycl::local_accessor<sycl::half2, 1> tile_x_dm_q3_K_acc_ct1(
@@ -2531,9 +2519,8 @@ static void ggml_mul_mat_q3_K_q8_1_sycl(const void *vx, const void *vy,
                 sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
                     sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
 
-                cgh.parallel_for(
-                    sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                    [=](sycl::nd_item<3> item_ct1) {
+                sycl_parallel_for(
+                    cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
                         mul_mat_q3_K<need_check>(
                             vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
                             nrows_dst, item_ct1,
@@ -2557,7 +2544,7 @@ static void ggml_mul_mat_q3_K_q8_1_sycl(const void *vx, const void *vy,
             dpct::has_capability_or_fail(stream->get_device(),
                                          {sycl::aspect::fp16});
 
-            stream->submit([&](sycl::handler &cgh) {
+            sycl_launch(stream, [&](sycl::handler & cgh) {
                 sycl::local_accessor<int, 1> tile_x_ql_q3_K_acc_ct1(
                     sycl::range<1>(mmq_y * (WARP_SIZE) + mmq_y), cgh);
                 sycl::local_accessor<sycl::half2, 1> tile_x_dm_q3_K_acc_ct1(
@@ -2572,9 +2559,8 @@ static void ggml_mul_mat_q3_K_q8_1_sycl(const void *vx, const void *vy,
                 sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
                     sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
 
-                cgh.parallel_for(
-                    sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                    [=](sycl::nd_item<3> item_ct1) {
+                sycl_parallel_for(
+                    cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
                         mul_mat_q3_K<need_check>(
                             vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
                             nrows_dst, item_ct1,
@@ -2644,7 +2630,7 @@ static void ggml_mul_mat_q4_K_q8_1_sycl(const void *vx, const void *vy,
             dpct::has_capability_or_fail(stream->get_device(),
                                          {sycl::aspect::fp16});
 
-            stream->submit([&](sycl::handler &cgh) {
+            sycl_launch(stream, [&](sycl::handler & cgh) {
                 sycl::local_accessor<int, 1> tile_x_ql_q4_K_acc_ct1(
                     sycl::range<1>(mmq_y * (WARP_SIZE) + mmq_y), cgh);
                 sycl::local_accessor<sycl::half2, 1> tile_x_dm_q4_K_acc_ct1(
@@ -2657,9 +2643,8 @@ static void ggml_mul_mat_q4_K_q8_1_sycl(const void *vx, const void *vy,
                 sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
                     sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
 
-                cgh.parallel_for(
-                    sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                    [=](sycl::nd_item<3> item_ct1) {
+                sycl_parallel_for(
+                    cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
                         mul_mat_q4_K<need_check>(
                             vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
                             nrows_dst, item_ct1,
@@ -2682,7 +2667,7 @@ static void ggml_mul_mat_q4_K_q8_1_sycl(const void *vx, const void *vy,
             dpct::has_capability_or_fail(stream->get_device(),
                                          {sycl::aspect::fp16});
 
-            stream->submit([&](sycl::handler &cgh) {
+            sycl_launch(stream, [&](sycl::handler & cgh) {
                 sycl::local_accessor<int, 1> tile_x_ql_q4_K_acc_ct1(
                     sycl::range<1>(mmq_y * (WARP_SIZE) + mmq_y), cgh);
                 sycl::local_accessor<sycl::half2, 1> tile_x_dm_q4_K_acc_ct1(
@@ -2695,9 +2680,8 @@ static void ggml_mul_mat_q4_K_q8_1_sycl(const void *vx, const void *vy,
                 sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
                     sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
 
-                cgh.parallel_for(
-                    sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                    [=](sycl::nd_item<3> item_ct1) {
+                sycl_parallel_for(
+                    cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
                         mul_mat_q4_K<need_check>(
                             vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
                             nrows_dst, item_ct1,
@@ -2765,7 +2749,7 @@ static void ggml_mul_mat_q5_K_q8_1_sycl(const void *vx, const void *vy,
             dpct::has_capability_or_fail(stream->get_device(),
                                          {sycl::aspect::fp16});
 
-            stream->submit([&](sycl::handler &cgh) {
+            sycl_launch(stream, [&](sycl::handler & cgh) {
                 sycl::local_accessor<int, 1> tile_x_ql_q5_K_acc_ct1(
                     sycl::range<1>(mmq_y * (2 * WARP_SIZE) + mmq_y), cgh);
                 sycl::local_accessor<sycl::half2, 1> tile_x_dm_q5_K_acc_ct1(
@@ -2778,9 +2762,8 @@ static void ggml_mul_mat_q5_K_q8_1_sycl(const void *vx, const void *vy,
                 sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
                     sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
 
-                cgh.parallel_for(
-                    sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                    [=](sycl::nd_item<3> item_ct1) {
+                sycl_parallel_for(
+                    cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
                         mul_mat_q5_K<need_check>(
                             vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
                             nrows_dst, item_ct1,
@@ -2803,7 +2786,7 @@ static void ggml_mul_mat_q5_K_q8_1_sycl(const void *vx, const void *vy,
             dpct::has_capability_or_fail(stream->get_device(),
                                          {sycl::aspect::fp16});
 
-            stream->submit([&](sycl::handler &cgh) {
+            sycl_launch(stream, [&](sycl::handler & cgh) {
                 sycl::local_accessor<int, 1> tile_x_ql_q5_K_acc_ct1(
                     sycl::range<1>(mmq_y * (2 * WARP_SIZE) + mmq_y), cgh);
                 sycl::local_accessor<sycl::half2, 1> tile_x_dm_q5_K_acc_ct1(
@@ -2816,9 +2799,8 @@ static void ggml_mul_mat_q5_K_q8_1_sycl(const void *vx, const void *vy,
                 sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
                     sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
 
-                cgh.parallel_for(
-                    sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                    [=](sycl::nd_item<3> item_ct1) {
+                sycl_parallel_for(
+                    cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
                         mul_mat_q5_K<need_check>(
                             vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
                             nrows_dst, item_ct1,
@@ -2886,7 +2868,7 @@ static void ggml_mul_mat_q6_K_q8_1_sycl(const void *vx, const void *vy,
             dpct::has_capability_or_fail(stream->get_device(),
                                          {sycl::aspect::fp16});
 
-            stream->submit([&](sycl::handler &cgh) {
+            sycl_launch(stream, [&](sycl::handler & cgh) {
                 sycl::local_accessor<int, 1> tile_x_ql_acc_ct1(
                     sycl::range<1>(mmq_y * (2 * WARP_SIZE) + mmq_y), cgh);
                 sycl::local_accessor<sycl::half2, 1> tile_x_dm_acc_ct1(
@@ -2899,9 +2881,8 @@ static void ggml_mul_mat_q6_K_q8_1_sycl(const void *vx, const void *vy,
                 sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
                     sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
 
-                cgh.parallel_for(
-                    sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                    [=](sycl::nd_item<3> item_ct1) {
+                sycl_parallel_for(
+                    cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
                         mul_mat_q6_K<need_check>(
                             vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
                             nrows_dst, item_ct1,
@@ -2924,7 +2905,7 @@ static void ggml_mul_mat_q6_K_q8_1_sycl(const void *vx, const void *vy,
             dpct::has_capability_or_fail(stream->get_device(),
                                          {sycl::aspect::fp16});
 
-            stream->submit([&](sycl::handler &cgh) {
+            sycl_launch(stream, [&](sycl::handler & cgh) {
                 sycl::local_accessor<int, 1> tile_x_ql_acc_ct1(
                     sycl::range<1>(mmq_y * (2 * WARP_SIZE) + mmq_y), cgh);
                 sycl::local_accessor<sycl::half2, 1> tile_x_dm_acc_ct1(
@@ -2937,9 +2918,8 @@ static void ggml_mul_mat_q6_K_q8_1_sycl(const void *vx, const void *vy,
                 sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
                     sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
 
-                cgh.parallel_for(
-                    sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                    [=](sycl::nd_item<3> item_ct1) {
+                sycl_parallel_for(
+                    cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
                         mul_mat_q6_K<need_check>(
                             vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
                             nrows_dst, item_ct1,
diff --git a/ggml/src/ggml-sycl/mmvq.cpp b/ggml/src/ggml-sycl/mmvq.cpp
index cb70f83a4..c21929d51 100644
--- a/ggml/src/ggml-sycl/mmvq.cpp
+++ b/ggml/src/ggml-sycl/mmvq.cpp
@@ -29,24 +29,23 @@ static void mul_mat_vec_q_reorder(const void * __restrict__ vx, const void * __r
     static_assert(blocks_per_subgroup > 0);
     static_assert(block_elements_per_subgroup > 0);
 
-    const block_q8_1 * y = (const block_q8_1 *) vy;
-
     float partial_sum = 0.0f;
     for (int i = sg.get_local_linear_id() / block_elements_per_subgroup; i < blocks_per_row; i += blocks_per_subgroup) {
-        const int ibx       = row * blocks_per_row + i;  // x block index
-        // TODO: Generalize offsets, right now only works for quantizations that don't split high and low bits
-        const int bx_offset = block_type::get_block_offset(ibx);
-        const int d_offset  = block_type::get_d_offset(nrows, ncols, ibx);
+        const int ibx = row * blocks_per_row + i;  // x block index
 
+        const auto         bx_offset      = block_type::get_block_offset(ibx, nblocks);
+        const auto         d_offset       = block_type::get_d_offset(nrows, ncols, ibx);
         // Y block index that aligns with ibx
         const int iby = i * block_type::block_to_q8_1_ratio();
+        const int8_t* q8_1_quant_ptr = (const int8_t*)vy + iby * QK8_1;
+        const sycl::half2* q8_1_ds_ptr = (const sycl::half2*)((const char*)vy + ncols + iby * sizeof(sycl::half2));
 
 #pragma unroll
         for (int elem = 0; elem < block_elements_per_subgroup; elem += WARP_SIZE) {
             // x block quant index when casting the quants to int
             const int iqs = elem + block_traits::vdr_mmvq * (sg.get_local_linear_id() % block_elements_per_subgroup);
 
-            partial_sum += reorder_vec_dot_q_sycl()(vx, bx_offset, d_offset, &y[iby], iqs, nblocks);
+            partial_sum += reorder_vec_dot_q_sycl()(vx, bx_offset, d_offset, q8_1_quant_ptr, q8_1_ds_ptr, iqs);
         }
     }
 
@@ -545,12 +544,12 @@ static void reorder_mul_mat_vec_q4_0_q8_1_sycl(const void * vx, const void * vy,
     const sycl::range<3> global_size(1, GGML_SYCL_MMV_Y, (block_num_y * WARP_SIZE));
     const sycl::range<3> workgroup_size(1, GGML_SYCL_MMV_Y, num_subgroups * WARP_SIZE);
 
-    stream->submit([&](sycl::handler & cgh) {
-        cgh.parallel_for(sycl::nd_range<3>(global_size, workgroup_size),
-                         [=](sycl::nd_item<3> nd_item) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                             mul_mat_vec_q_reorder<reorder_vec_dot_q_sycl<GGML_TYPE_Q4_0>>(vx, vy, dst, ncols, nrows,
-                                                                                           nd_item);
-                         });
+    sycl_launch(stream, [&](sycl::handler & cgh) {
+        sycl_parallel_for(cgh, sycl::nd_range<3>(global_size, workgroup_size),
+                          [=](sycl::nd_item<3> nd_item) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                              mul_mat_vec_q_reorder<reorder_vec_dot_q_sycl<GGML_TYPE_Q4_0>>(vx, vy, dst, ncols, nrows,
+                                                                                            nd_item);
+                          });
     });
 }
 
@@ -562,12 +561,12 @@ static void mul_mat_vec_q4_0_q8_1_sycl(const void * vx, const void * vy, float *
     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) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                                 mul_mat_vec_q<QK4_0, QI4_0, block_q4_0, VDR_Q4_0_Q8_1_MMVQ, vec_dot_q4_0_q8_1>(
-                                     vx, vy, dst, ncols, nrows, item_ct1);
-                             });
+        sycl_launch(stream, [&](sycl::handler & cgh) {
+            sycl_parallel_for(cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                              [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                                  mul_mat_vec_q<QK4_0, QI4_0, block_q4_0, VDR_Q4_0_Q8_1_MMVQ, vec_dot_q4_0_q8_1>(
+                                      vx, vy, dst, ncols, nrows, item_ct1);
+                              });
         });
     }
 }
@@ -581,17 +580,12 @@ static void mul_mat_vec_q4_1_q8_1_sycl(const void *vx, const void *vy,
     const sycl::range<3> block_nums(1, 1, block_num_y);
     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)
-                    [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                        mul_mat_vec_q<QK4_0, QI4_1, block_q4_1,
-                                      VDR_Q4_1_Q8_1_MMVQ, vec_dot_q4_1_q8_1>(
-                            vx, vy, dst, ncols, nrows, item_ct1);
-                    });
+        sycl_launch(stream, [&](sycl::handler & cgh) {
+            sycl_parallel_for(cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                              [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                                  mul_mat_vec_q<QK4_0, QI4_1, block_q4_1, VDR_Q4_1_Q8_1_MMVQ, vec_dot_q4_1_q8_1>(
+                                      vx, vy, dst, ncols, nrows, item_ct1);
+                              });
         });
     }
 }
@@ -605,17 +599,12 @@ static void mul_mat_vec_q5_0_q8_1_sycl(const void *vx, const void *vy,
     const sycl::range<3> block_nums(1, 1, block_num_y);
     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)
-                    [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                        mul_mat_vec_q<QK5_0, QI5_0, block_q5_0,
-                                      VDR_Q5_0_Q8_1_MMVQ, vec_dot_q5_0_q8_1>(
-                            vx, vy, dst, ncols, nrows, item_ct1);
-                    });
+        sycl_launch(stream, [&](sycl::handler & cgh) {
+            sycl_parallel_for(cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                              [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                                  mul_mat_vec_q<QK5_0, QI5_0, block_q5_0, VDR_Q5_0_Q8_1_MMVQ, vec_dot_q5_0_q8_1>(
+                                      vx, vy, dst, ncols, nrows, item_ct1);
+                              });
         });
     }
 }
@@ -629,17 +618,12 @@ static void mul_mat_vec_q5_1_q8_1_sycl(const void *vx, const void *vy,
     const sycl::range<3> block_nums(1, 1, block_num_y);
     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)
-                    [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                        mul_mat_vec_q<QK5_1, QI5_1, block_q5_1,
-                                      VDR_Q5_1_Q8_1_MMVQ, vec_dot_q5_1_q8_1>(
-                            vx, vy, dst, ncols, nrows, item_ct1);
-                    });
+        sycl_launch(stream, [&](sycl::handler & cgh) {
+            sycl_parallel_for(cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                              [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                                  mul_mat_vec_q<QK5_1, QI5_1, block_q5_1, VDR_Q5_1_Q8_1_MMVQ, vec_dot_q5_1_q8_1>(
+                                      vx, vy, dst, ncols, nrows, item_ct1);
+                              });
         });
     }
 }
@@ -653,17 +637,12 @@ static void mul_mat_vec_q8_0_q8_1_sycl(const void *vx, const void *vy,
     const sycl::range<3> block_nums(1, 1, block_num_y);
     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)
-                    [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                        mul_mat_vec_q<QK8_0, QI8_0, block_q8_0,
-                                      VDR_Q8_0_Q8_1_MMVQ, vec_dot_q8_0_q8_1>(
-                            vx, vy, dst, ncols, nrows, item_ct1);
-                    });
+        sycl_launch(stream, [&](sycl::handler & cgh) {
+            sycl_parallel_for(cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                              [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                                  mul_mat_vec_q<QK8_0, QI8_0, block_q8_0, VDR_Q8_0_Q8_1_MMVQ, vec_dot_q8_0_q8_1>(
+                                      vx, vy, dst, ncols, nrows, item_ct1);
+                              });
         });
     }
 }
@@ -677,17 +656,12 @@ static void mul_mat_vec_q2_K_q8_1_sycl(const void *vx, const void *vy,
     const sycl::range<3> block_nums(1, 1, block_num_y);
     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)
-                    [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                        mul_mat_vec_q<QK_K, QI2_K, block_q2_K,
-                                      VDR_Q2_K_Q8_1_MMVQ, vec_dot_q2_K_q8_1>(
-                            vx, vy, dst, ncols, nrows, item_ct1);
-                    });
+        sycl_launch(stream, [&](sycl::handler & cgh) {
+            sycl_parallel_for(cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                              [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                                  mul_mat_vec_q<QK_K, QI2_K, block_q2_K, VDR_Q2_K_Q8_1_MMVQ, vec_dot_q2_K_q8_1>(
+                                      vx, vy, dst, ncols, nrows, item_ct1);
+                              });
         });
     }
 }
@@ -701,17 +675,12 @@ static void mul_mat_vec_q3_K_q8_1_sycl(const void *vx, const void *vy,
     const sycl::range<3> block_nums(1, 1, block_num_y);
     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)
-                    [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                        mul_mat_vec_q<QK_K, QI3_K, block_q3_K,
-                                      VDR_Q3_K_Q8_1_MMVQ, vec_dot_q3_K_q8_1>(
-                            vx, vy, dst, ncols, nrows, item_ct1);
-                    });
+        sycl_launch(stream, [&](sycl::handler & cgh) {
+            sycl_parallel_for(cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                              [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                                  mul_mat_vec_q<QK_K, QI3_K, block_q3_K, VDR_Q3_K_Q8_1_MMVQ, vec_dot_q3_K_q8_1>(
+                                      vx, vy, dst, ncols, nrows, item_ct1);
+                              });
         });
     }
 }
@@ -725,17 +694,12 @@ static void mul_mat_vec_q4_K_q8_1_sycl(const void *vx, const void *vy,
     const sycl::range<3> block_nums(1, 1, block_num_y);
     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)
-                    [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                        mul_mat_vec_q<QK_K, QI4_K, block_q4_K,
-                                      VDR_Q4_K_Q8_1_MMVQ, vec_dot_q4_K_q8_1>(
-                            vx, vy, dst, ncols, nrows, item_ct1);
-                    });
+        sycl_launch(stream, [&](sycl::handler & cgh) {
+            sycl_parallel_for(cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                              [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                                  mul_mat_vec_q<QK_K, QI4_K, block_q4_K, VDR_Q4_K_Q8_1_MMVQ, vec_dot_q4_K_q8_1>(
+                                      vx, vy, dst, ncols, nrows, item_ct1);
+                              });
         });
     }
 }
@@ -751,12 +715,12 @@ static void reorder_mul_mat_vec_q4_k_q8_1_sycl(const void * vx, const void * vy,
     const sycl::range<3> global_size(1, GGML_SYCL_MMV_Y, block_num_y * WARP_SIZE);
     const sycl::range<3> workgroup_size(1, GGML_SYCL_MMV_Y, num_subgroups * WARP_SIZE);
 
-    stream->submit([&](sycl::handler & cgh) {
-        cgh.parallel_for(sycl::nd_range<3>(global_size, workgroup_size),
-                            [=](sycl::nd_item<3> nd_item) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                                mul_mat_vec_q_reorder<reorder_vec_dot_q_sycl<GGML_TYPE_Q4_K>>(vx, vy, dst, ncols,
-                                                                                            nrows, nd_item);
-                            });
+    sycl_launch(stream, [&](sycl::handler & cgh) {
+        sycl_parallel_for(cgh, sycl::nd_range<3>(global_size, workgroup_size),
+                          [=](sycl::nd_item<3> nd_item) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                              mul_mat_vec_q_reorder<reorder_vec_dot_q_sycl<GGML_TYPE_Q4_K>>(vx, vy, dst, ncols, nrows,
+                                                                                            nd_item);
+                          });
     });
 }
 
@@ -770,21 +734,34 @@ static void mul_mat_vec_q5_K_q8_1_sycl(const void *vx, const void *vy,
     const sycl::range<3> block_nums(1, 1, block_num_y);
     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)
-                    [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                        mul_mat_vec_q<QK_K, QI5_K, block_q5_K,
-                                      VDR_Q5_K_Q8_1_MMVQ, vec_dot_q5_K_q8_1>(
-                            vx, vy, dst, ncols, nrows, item_ct1);
-                    });
+        sycl_launch(stream, [&](sycl::handler & cgh) {
+            sycl_parallel_for(cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                              [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                                  mul_mat_vec_q<QK_K, QI5_K, block_q5_K, VDR_Q5_K_Q8_1_MMVQ, vec_dot_q5_K_q8_1>(
+                                      vx, vy, dst, ncols, nrows, item_ct1);
+                              });
         });
     }
 }
 
+static void reorder_mul_mat_vec_q6_k_q8_1_sycl(const void * vx, const void * vy, float * dst, const int ncols,
+                                               const int nrows, dpct::queue_ptr stream) {
+    GGML_ASSERT(ncols % QK_K == 0);
+    const int        block_num_y   = ceil_div(nrows, GGML_SYCL_MMV_Y);
+    constexpr size_t num_subgroups = 16;
+    GGML_ASSERT(block_num_y % num_subgroups == 0);
+
+    const sycl::range<3> global_size(1, GGML_SYCL_MMV_Y, block_num_y * WARP_SIZE);
+    const sycl::range<3> workgroup_size(1, GGML_SYCL_MMV_Y, num_subgroups * WARP_SIZE);
+
+    sycl_launch(stream, [&](sycl::handler & cgh) {
+        sycl_parallel_for(cgh, sycl::nd_range<3>(global_size, workgroup_size),
+                          [=](sycl::nd_item<3> nd_item) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                              mul_mat_vec_q_reorder<reorder_vec_dot_q_sycl<GGML_TYPE_Q6_K>>(vx, vy, dst, ncols, nrows,
+                                                                                            nd_item);
+                          });
+    });
+}
 static void mul_mat_vec_q6_K_q8_1_sycl(const void *vx, const void *vy,
                                        float *dst, const int ncols,
                                        const int nrows,
@@ -794,17 +771,12 @@ static void mul_mat_vec_q6_K_q8_1_sycl(const void *vx, const void *vy,
     const sycl::range<3> block_nums(1, 1, block_num_y);
     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)
-                    [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                        mul_mat_vec_q<QK_K, QI6_K, block_q6_K,
-                                      VDR_Q6_K_Q8_1_MMVQ, vec_dot_q6_K_q8_1>(
-                            vx, vy, dst, ncols, nrows, item_ct1);
-                    });
+        sycl_launch(stream, [&](sycl::handler & cgh) {
+            sycl_parallel_for(cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                              [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                                  mul_mat_vec_q<QK_K, QI6_K, block_q6_K, VDR_Q6_K_Q8_1_MMVQ, vec_dot_q6_K_q8_1>(
+                                      vx, vy, dst, ncols, nrows, item_ct1);
+                              });
         });
     }
 }
@@ -819,14 +791,12 @@ static void mul_mat_vec_iq2_xxs_q8_1_sycl(const void *vx, const void *vy,
     const sycl::range<3> block_nums(1, 1, block_num_y);
     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)
-                    [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                        mul_mat_vec_q_iq2_xxs_q8_1<QK_K, QI2_XXS/2, block_iq2_xxs, 1>(
-                            vx, vy, dst, ncols, nrows, item_ct1);
-                    });
+        sycl_launch(stream, [&](sycl::handler & cgh) {
+            sycl_parallel_for(cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                              [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                                  mul_mat_vec_q_iq2_xxs_q8_1<QK_K, QI2_XXS / 2, block_iq2_xxs, 1>(vx, vy, dst, ncols,
+                                                                                                  nrows, item_ct1);
+                              });
         });
     }
 }
@@ -840,14 +810,12 @@ static void mul_mat_vec_iq2_xs_q8_1_sycl(const void *vx, const void *vy,
     const sycl::range<3> block_nums(1, 1, block_num_y);
     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)
-                    [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                        mul_mat_vec_q_iq2_xs_q8_1<QK_K, QI2_XS/2, block_iq2_xs, 1>(
-                            vx, vy, dst, ncols, nrows, item_ct1);
-                    });
+        sycl_launch(stream, [&](sycl::handler & cgh) {
+            sycl_parallel_for(cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                              [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                                  mul_mat_vec_q_iq2_xs_q8_1<QK_K, QI2_XS / 2, block_iq2_xs, 1>(vx, vy, dst, ncols,
+                                                                                               nrows, item_ct1);
+                              });
         });
     }
 }
@@ -861,15 +829,12 @@ static void mul_mat_vec_iq2_s_q8_1_sycl(const void *vx, const void *vy,
     const sycl::range<3> block_nums(1, 1, block_num_y);
     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)
-                    [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                        mul_mat_vec_q_iq2_s_q8_1<QK_K, QI2_S/2, block_iq2_s, 1>(
-                            vx, vy, dst, ncols, nrows, item_ct1);
-                    });
+        sycl_launch(stream, [&](sycl::handler & cgh) {
+            sycl_parallel_for(cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                              [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                                  mul_mat_vec_q_iq2_s_q8_1<QK_K, QI2_S / 2, block_iq2_s, 1>(vx, vy, dst, ncols, nrows,
+                                                                                            item_ct1);
+                              });
         });
     }
 }
@@ -883,15 +848,12 @@ static void mul_mat_vec_iq3_xxs_q8_1_sycl(const void *vx, const void *vy,
     const sycl::range<3> block_nums(1, 1, block_num_y);
     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)
-                    [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                        mul_mat_vec_q_iq3_xxs_q8_1<QK_K, QI3_XXS/2, block_iq3_xxs, 1>(
-                            vx, vy, dst, ncols, nrows, item_ct1);
-                    });
+        sycl_launch(stream, [&](sycl::handler & cgh) {
+            sycl_parallel_for(cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                              [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                                  mul_mat_vec_q_iq3_xxs_q8_1<QK_K, QI3_XXS / 2, block_iq3_xxs, 1>(vx, vy, dst, ncols,
+                                                                                                  nrows, item_ct1);
+                              });
         });
     }
 }
@@ -905,15 +867,12 @@ static void mul_mat_vec_iq3_s_q8_1_sycl(const void *vx, const void *vy,
     const sycl::range<3> block_nums(1, 1, block_num_y);
     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)
-                    [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                        mul_mat_vec_q_iq3_s_q8_1<QK_K, QI3_S/2, block_iq3_s, 1>(
-                            vx, vy, dst, ncols, nrows, item_ct1);
-                    });
+        sycl_launch(stream, [&](sycl::handler & cgh) {
+            sycl_parallel_for(cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                              [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                                  mul_mat_vec_q_iq3_s_q8_1<QK_K, QI3_S / 2, block_iq3_s, 1>(vx, vy, dst, ncols, nrows,
+                                                                                            item_ct1);
+                              });
         });
     }
 }
@@ -927,15 +886,12 @@ static void mul_mat_vec_iq1_s_q8_1_sycl(const void *vx, const void *vy,
     const sycl::range<3> block_nums(1, 1, block_num_y);
     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)
-                    [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                        mul_mat_vec_q_iq1_s_q8_1<QK_K, QI1_S, block_iq1_s, 1>(
-                            vx, vy, dst, ncols, nrows, item_ct1);
-                    });
+        sycl_launch(stream, [&](sycl::handler & cgh) {
+            sycl_parallel_for(cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                              [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                                  mul_mat_vec_q_iq1_s_q8_1<QK_K, QI1_S, block_iq1_s, 1>(vx, vy, dst, ncols, nrows,
+                                                                                        item_ct1);
+                              });
         });
     }
 }
@@ -949,14 +905,12 @@ static void mul_mat_vec_iq1_m_q8_1_sycl(const void *vx, const void *vy,
     const sycl::range<3> block_nums(1, 1, block_num_y);
     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)
-                    [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                        mul_mat_vec_q_iq1_m_q8_1<QK_K, QI1_S, block_iq1_m, 1>(
-                            vx, vy, dst, ncols, nrows, item_ct1);
-                    });
+        sycl_launch(stream, [&](sycl::handler & cgh) {
+            sycl_parallel_for(cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                              [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                                  mul_mat_vec_q_iq1_m_q8_1<QK_K, QI1_S, block_iq1_m, 1>(vx, vy, dst, ncols, nrows,
+                                                                                        item_ct1);
+                              });
         });
     }
 }
@@ -970,15 +924,12 @@ static void mul_mat_vec_iq4_nl_q8_1_sycl(const void *vx, const void *vy,
     const sycl::range<3> block_nums(1, 1, block_num_y);
     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)
-                    [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                        mul_mat_vec_q_iq4_nl_q8_1<QK4_NL, QI4_NL, block_iq4_nl, 2>(
-                            vx, vy, dst, ncols, nrows, item_ct1);
-                    });
+        sycl_launch(stream, [&](sycl::handler & cgh) {
+            sycl_parallel_for(cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                              [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                                  mul_mat_vec_q_iq4_nl_q8_1<QK4_NL, QI4_NL, block_iq4_nl, 2>(vx, vy, dst, ncols, nrows,
+                                                                                             item_ct1);
+                              });
         });
     }
 }
@@ -992,15 +943,12 @@ static void mul_mat_vec_iq4_xs_q8_1_sycl(const void *vx, const void *vy,
     const sycl::range<3> block_nums(1, 1, block_num_y);
     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)
-                    [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                        mul_mat_vec_q_iq4_xs_q8_1<QK_K, QI4_XS/4, block_iq4_xs, 1>(
-                            vx, vy, dst, ncols, nrows, item_ct1);
-                    });
+        sycl_launch(stream, [&](sycl::handler & cgh) {
+            sycl_parallel_for(cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                              [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                                  mul_mat_vec_q_iq4_xs_q8_1<QK_K, QI4_XS / 4, block_iq4_xs, 1>(vx, vy, dst, ncols,
+                                                                                               nrows, item_ct1);
+                              });
         });
     }
 }
@@ -1070,7 +1018,14 @@ void ggml_sycl_op_mul_mat_vec_q(ggml_backend_sycl_context & ctx, const ggml_tens
                 mul_mat_vec_q5_K_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream);
                 break;
             case GGML_TYPE_Q6_K:
-                mul_mat_vec_q6_K_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream);
+                if ((ggml_tensor_extra_gpu *) dst->src[0]->extra &&
+                    ((ggml_tensor_extra_gpu *) dst->src[0]->extra)->optimized_feature.reorder) {
+                    GGML_SYCL_DEBUG("Calling reorder_mul_mat_vec_q6_k_q8_1_sycl\n");
+                    reorder_mul_mat_vec_q6_k_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream);
+                } else {
+                    GGML_SYCL_DEBUG("Calling mul_mat_vec_q6_k_q8_1_sycl\n");
+                    mul_mat_vec_q6_K_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream);
+                }
                 break;
             case GGML_TYPE_IQ1_S:
                 mul_mat_vec_iq1_s_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream);
diff --git a/ggml/src/ggml-sycl/norm.cpp b/ggml/src/ggml-sycl/norm.cpp
index 4ec141684..79d846b41 100644
--- a/ggml/src/ggml-sycl/norm.cpp
+++ b/ggml/src/ggml-sycl/norm.cpp
@@ -254,14 +254,13 @@ static void norm_f32_sycl(const float * x, float * dst, const int ncols, const i
     GGML_ASSERT(ncols % WARP_SIZE == 0);
     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>(global_dims * block_dims, block_dims),
-                [=](sycl::nd_item<3> item_ct1)
-                [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                    norm_f32(x, dst, ncols, stride_row, stride_channel, stride_sample, eps, item_ct1, nullptr, WARP_SIZE);
-                });
-            });
+        sycl_launch(stream, [&](sycl::handler & cgh) {
+            sycl_parallel_for(cgh, sycl::nd_range<3>(global_dims * block_dims, block_dims),
+                              [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                                  norm_f32(x, dst, ncols, stride_row, stride_channel, stride_sample, eps, item_ct1,
+                                           nullptr, WARP_SIZE);
+                              });
+        });
     }
     else {
         const int work_group_size = ggml_sycl_info().max_work_group_sizes[device];
@@ -272,16 +271,15 @@ static void norm_f32_sycl(const float * x, float * dst, const int ncols, const i
         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_launch(stream, [&](sycl::handler & cgh) {
             sycl::local_accessor<sycl::float2, 1> s_sum_acc_ct1(
                             sycl::range<1>(work_group_size / WARP_SIZE), cgh);
-            cgh.parallel_for(
-                sycl::nd_range<3>(global_dims * block_dims, block_dims),
-                [=](sycl::nd_item<3> item_ct1)
-                [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                    norm_f32(x, dst, ncols, stride_row, stride_channel, stride_sample, eps, item_ct1, get_pointer(s_sum_acc_ct1), work_group_size);
-                });
-            });
+            sycl_parallel_for(cgh, sycl::nd_range<3>(global_dims * block_dims, block_dims),
+                              [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                                  norm_f32(x, dst, ncols, stride_row, stride_channel, stride_sample, eps, item_ct1,
+                                           get_pointer(s_sum_acc_ct1), work_group_size);
+                              });
+        });
     }
 }
 
@@ -290,18 +288,14 @@ static void group_norm_f32_sycl(const float* x, float* dst,
     const int ne_elements, queue_ptr stream, int device) {
     if (group_size < 1024) {
         const sycl::range<3> block_dims(1, 1, WARP_SIZE);
-        stream->submit([&](sycl::handler& cgh) {
+        sycl_launch(stream, [&](sycl::handler & cgh) {
             const float eps_ct4 = eps;
-            cgh.parallel_for(
-                sycl::nd_range<3>(sycl::range<3>(1, 1, num_groups) * block_dims,
-                    block_dims),
-                [=](sycl::nd_item<3> item_ct1)
-                [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                    group_norm_f32(
-                        x, dst, group_size, ne_elements, eps_ct4, item_ct1,
-                        nullptr, WARP_SIZE);
-                });
-            });
+            sycl_parallel_for(cgh, sycl::nd_range<3>(sycl::range<3>(1, 1, num_groups) * block_dims, block_dims),
+                              [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                                  group_norm_f32(x, dst, group_size, ne_elements, eps_ct4, item_ct1, nullptr,
+                                                 WARP_SIZE);
+                              });
+        });
     }
     else {
         const int work_group_size = ggml_sycl_info().max_work_group_sizes[device];
@@ -313,22 +307,18 @@ static void group_norm_f32_sycl(const float* x, float* dst,
         info::device::max_work_group_size. Adjust the work-group size if needed.
         */
 
-        stream->submit([&](sycl::handler& cgh) {
+        sycl_launch(stream, [&](sycl::handler & cgh) {
             sycl::local_accessor<float, 1> s_sum_acc_ct1(sycl::range<1>(work_group_size / WARP_SIZE),
                 cgh);
 
             const float eps_ct4 = eps;
 
-            cgh.parallel_for(
-                sycl::nd_range<3>(sycl::range<3>(1, 1, num_groups) * block_dims,
-                    block_dims),
-                [=](sycl::nd_item<3> item_ct1)
-                [[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);
-                });
-            });
+            sycl_parallel_for(cgh, sycl::nd_range<3>(sycl::range<3>(1, 1, num_groups) * block_dims, block_dims),
+                              [=](sycl::nd_item<3> item_ct1) [[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);
+                              });
+        });
     }
 }
 
@@ -340,14 +330,13 @@ static void rms_norm_f32_sycl(const float* x, float* dst, const int ncols, const
     const sycl::range<3> global_dims(nsamples, nchannels, nrows);
     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>(global_dims * block_dims, block_dims),
-                [=](sycl::nd_item<3> item_ct1)
-                [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                    rms_norm_f32(x, dst, ncols, stride_row, stride_channel, stride_sample, eps, item_ct1, nullptr, WARP_SIZE);
-                });
-            });
+        sycl_launch(stream, [&](sycl::handler & cgh) {
+            sycl_parallel_for(cgh, sycl::nd_range<3>(global_dims * block_dims, block_dims),
+                              [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                                  rms_norm_f32(x, dst, ncols, stride_row, stride_channel, stride_sample, eps, item_ct1,
+                                               nullptr, WARP_SIZE);
+                              });
+        });
     }
     else {
         const int work_group_size = ggml_sycl_info().max_work_group_sizes[device];
@@ -358,16 +347,15 @@ static void rms_norm_f32_sycl(const float* x, float* dst, const int ncols, const
         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_launch(stream, [&](sycl::handler & cgh) {
             sycl::local_accessor<float, 1> s_sum_acc_ct1(sycl::range<1>(work_group_size / WARP_SIZE),
                 cgh);
-            cgh.parallel_for(
-                sycl::nd_range<3>(global_dims * block_dims, block_dims),
-                [=](sycl::nd_item<3> item_ct1)
-                [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                    rms_norm_f32(x, dst, ncols, stride_row, stride_channel, stride_sample, eps, item_ct1, get_pointer(s_sum_acc_ct1), work_group_size);
-                });
-            });
+            sycl_parallel_for(cgh, sycl::nd_range<3>(global_dims * block_dims, block_dims),
+                              [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                                  rms_norm_f32(x, dst, ncols, stride_row, stride_channel, stride_sample, eps, item_ct1,
+                                               get_pointer(s_sum_acc_ct1), work_group_size);
+                              });
+        });
     }
 }
 
@@ -378,16 +366,12 @@ static void l2_norm_f32_sycl(const float* x, float* dst, const int ncols,
     // 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);
-                });
-            });
+        sycl_launch(stream, [&](sycl::handler & cgh) {
+            sycl_parallel_for(cgh, 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];
@@ -398,18 +382,15 @@ static void l2_norm_f32_sycl(const float* x, float* dst, const int ncols,
         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_launch(stream, [&](sycl::handler & cgh) {
             sycl::local_accessor<float, 1> 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);
-                });
-            });
+            sycl_parallel_for(cgh, 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);
+                              });
+        });
     }
 }
 
diff --git a/ggml/src/ggml-sycl/quants.hpp b/ggml/src/ggml-sycl/quants.hpp
index 88ec13ea2..8b952db43 100644
--- a/ggml/src/ggml-sycl/quants.hpp
+++ b/ggml/src/ggml-sycl/quants.hpp
@@ -14,12 +14,13 @@
 #ifndef GGML_SYCL_QUANTS_HPP
 #define GGML_SYCL_QUANTS_HPP
 
+#include <utility>
+
 #include "ggml-common.h"
 #include "ggml.h"
 
 namespace ggml_sycl_reordered {
 
-
 // The reordered block moves quants (qs) and  scales(d) to two
 // uniform regions of memory that is contiguous in the same tensor.
 // What this means is that instead of having:
@@ -32,7 +33,6 @@ namespace ggml_sycl_reordered {
 
 template <ggml_type type> struct block_q_t;
 
-
 // qk number of weights / quants in a block
 // qr number of weights in a byte (described as 'before dequantization')
 //    for quantization types that has low and high bits split, qr is calculated with
@@ -47,10 +47,12 @@ template <> struct block_q_t<GGML_TYPE_Q4_0> {
         static constexpr uint32_t vdr_mmvq = 2;
     };
 
-    static constexpr int get_block_offset(const int block_index) { return block_index * (traits::qk / traits::qr); }
+    static constexpr std::pair<int, int> get_block_offset(const int block_index, const int /* nblocks */) {
+        return { block_index * (traits::qk / traits::qr), 0 };
+    }
 
-    static constexpr int get_d_offset(int nrows, int ncols, const int block_index) {
-        return (ncols / traits::qr * nrows) + block_index * sizeof(ggml_half);
+    static constexpr std::pair<int, int> get_d_offset(int nrows, int ncols, const int block_index) {
+        return { (ncols / traits::qr * nrows) + block_index * sizeof(ggml_half), 0 };
     }
 
     static constexpr int block_to_q8_1_ratio() { return traits::qk / QK8_1; }
@@ -64,20 +66,46 @@ template <> struct block_q_t<GGML_TYPE_Q4_K> {
         static constexpr uint32_t vdr_mmvq = 2;
     };
 
-    static constexpr int get_block_offset(const int block_index) { return block_index * (traits::qk / traits::qr); }
+    static constexpr std::pair<int, int> get_block_offset(const int block_index, const int /* nblocks */) {
+        return { block_index * (traits::qk / traits::qr), 0 };
+    }
 
-    static constexpr int get_d_offset(int nrows, int ncols, const int block_index) {
+    static constexpr std::pair<int, int> get_d_offset(int nrows, int ncols, const int block_index) {
         auto nblocks = (nrows * (ncols / traits::qk));
-        return (nblocks * QK_K / 2) + (nblocks * K_SCALE_SIZE) + (block_index * sizeof(ggml_half2));
+        return { nblocks * (QK_K / 2),
+                 (nblocks * QK_K / 2) + (nblocks * K_SCALE_SIZE) + (block_index * sizeof(ggml_half2)) };
     }
 
     static constexpr int block_to_q8_1_ratio() { return traits::qk / QK8_1; }
 
     constexpr size_t get_total_qs_bytes(int nblocks) { return nblocks * QK_K / 2; }
-
-    constexpr size_t get_dm_offset(int nblocks) { return get_total_qs_bytes(nblocks) + nblocks * K_SCALE_SIZE; }
 };
 
+template <> struct block_q_t<GGML_TYPE_Q6_K> {
+    struct traits {
+        static constexpr uint32_t qk       = QK_K;
+        static constexpr uint32_t qi       = QI6_K;
+        static constexpr uint32_t qr       = QR6_K;
+        static constexpr uint32_t vdr_mmvq = 1;
+    };
+
+    static constexpr std::pair<int, int> get_block_offset(const int block_index, const int n_blocks) {
+        auto low_bits_index  = block_index * (traits::qk / traits::qr);
+        // the index of high bits it's after all low bits
+        auto high_bits_index = n_blocks * (QK_K / 2) + (block_index * (QK_K / 4));
+        return { low_bits_index, high_bits_index };
+    }
+
+    static constexpr std::pair<int, int> get_d_offset(int nrows, int ncols, const int block_index) {
+        auto nblocks        = (nrows * (ncols / traits::qk));
+        auto total_qs_bytes = nblocks * (QK_K / 2) + nblocks * (QK_K / 4);
+        auto block_scales   = total_qs_bytes + block_index * (QK_K / 16);
+        auto sb_scale       = total_qs_bytes + nblocks * (QK_K / 16);
+        return { block_scales, sb_scale };
+    }
+
+    static constexpr int block_to_q8_1_ratio() { return traits::qk / QK8_1; }
+};
 }  // namespace ggml_sycl_reordered
 
 #endif  // GGML_SYCL_QUANTS_HPP
diff --git a/ggml/src/ggml-sycl/rope.cpp b/ggml/src/ggml-sycl/rope.cpp
index 44473e1e5..e44c6b6ef 100644
--- a/ggml/src/ggml-sycl/rope.cpp
+++ b/ggml/src/ggml-sycl/rope.cpp
@@ -235,20 +235,22 @@ static void rope_norm_sycl(const T * x, T * dst, const int ne0, const int ne1, c
         the limit. To get the device limit, query
         info::device::max_work_group_size. Adjust the work-group size if needed.
         */
-        stream->parallel_for(sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
-            rope_norm<T, false>(x, dst, ne0, ne1, s1, s2, n_dims, pos, freq_scale, ext_factor, attn_factor, corr_dims,
-                                theta_scale, freq_factors, item_ct1);
-        });
+        sycl_parallel_for(stream, sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                          [=](sycl::nd_item<3> item_ct1) {
+                              rope_norm<T, false>(x, dst, ne0, ne1, s1, s2, n_dims, pos, freq_scale, ext_factor,
+                                                  attn_factor, corr_dims, theta_scale, freq_factors, item_ct1);
+                          });
     } else {
         /*
         DPCT1049:41: 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->parallel_for(sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
-            rope_norm<T, true>(x, dst, ne0, ne1, s1, s2, n_dims, pos, freq_scale, ext_factor, attn_factor, corr_dims,
-                               theta_scale, freq_factors, item_ct1);
-        });
+        sycl_parallel_for(stream, sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                          [=](sycl::nd_item<3> item_ct1) {
+                              rope_norm<T, true>(x, dst, ne0, ne1, s1, s2, n_dims, pos, freq_scale, ext_factor,
+                                                 attn_factor, corr_dims, theta_scale, freq_factors, item_ct1);
+                          });
     }
 }
 
@@ -267,15 +269,17 @@ static void rope_neox_sycl(const T * x, T * dst, const int ne0, const int ne1, c
     dpct::has_capability_or_fail(stream->get_device(), { sycl::aspect::fp16 });
 
     if (freq_factors == nullptr) {
-        stream->parallel_for(sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
-            rope_neox<T, false>(x, dst, ne0, ne1, s1, s2, n_dims, pos, freq_scale, ext_factor, attn_factor, corr_dims,
-                                theta_scale, freq_factors, item_ct1);
-        });
+        sycl_parallel_for(stream, sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                          [=](sycl::nd_item<3> item_ct1) {
+                              rope_neox<T, false>(x, dst, ne0, ne1, s1, s2, n_dims, pos, freq_scale, ext_factor,
+                                                  attn_factor, corr_dims, theta_scale, freq_factors, item_ct1);
+                          });
     } else {
-        stream->parallel_for(sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
-            rope_neox<T, true>(x, dst, ne0, ne1, s1, s2, n_dims, pos, freq_scale, ext_factor, attn_factor, corr_dims,
-                               theta_scale, freq_factors, item_ct1);
-        });
+        sycl_parallel_for(stream, sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                          [=](sycl::nd_item<3> item_ct1) {
+                              rope_neox<T, true>(x, dst, ne0, ne1, s1, s2, n_dims, pos, freq_scale, ext_factor,
+                                                 attn_factor, corr_dims, theta_scale, freq_factors, item_ct1);
+                          });
     }
 }
 
@@ -298,12 +302,12 @@ static void rope_multi_sycl(const T * x, T * dst, const int ne0, const int ne1,
     }
     // launch kernel
     if (freq_factors == nullptr) {
-        stream->parallel_for(nd_range, [=](sycl::nd_item<3> item_ct1) {
+        sycl_parallel_for(stream, nd_range, [=](sycl::nd_item<3> item_ct1) {
             rope_multi<T, false>(x, dst, ne0, ne1, ne2, s1, s2, n_dims, pos, freq_scale, ext_factor, attn_factor,
                                   corr_dims, theta_scale, freq_factors, sections, item_ct1);
         });
     } else {
-        stream->parallel_for(nd_range, [=](sycl::nd_item<3> item_ct1) {
+        sycl_parallel_for(stream, nd_range, [=](sycl::nd_item<3> item_ct1) {
             rope_multi<T, true>(x, dst, ne0, ne1, ne2, s1, s2, n_dims, pos, freq_scale, ext_factor, attn_factor,
                                  corr_dims, theta_scale, freq_factors, sections, item_ct1);
         });
@@ -333,12 +337,12 @@ static void rope_vision_sycl(const T * x, T * dst, const int ne0, const int ne1,
     }
     // launch kernel
     if (freq_factors == nullptr) {
-        stream->parallel_for(nd_range, [=](sycl::nd_item<3> item_ct1) {
+        sycl_parallel_for(stream, nd_range, [=](sycl::nd_item<3> item_ct1) {
             rope_vision<T, false>(x, dst, ne0, ne1, ne2, s1, s2, n_dims, pos, freq_scale, ext_factor, attn_factor,
                                   corr_dims, theta_scale, freq_factors, sections, item_ct1);
         });
     } else {
-        stream->parallel_for(nd_range, [=](sycl::nd_item<3> item_ct1) {
+        sycl_parallel_for(stream, nd_range, [=](sycl::nd_item<3> item_ct1) {
             rope_vision<T, true>(x, dst, ne0, ne1, ne2, s1, s2, n_dims, pos, freq_scale, ext_factor, attn_factor,
                                  corr_dims, theta_scale, freq_factors, sections, item_ct1);
         });
diff --git a/ggml/src/ggml-sycl/softmax.cpp b/ggml/src/ggml-sycl/softmax.cpp
index 52fcf4b3d..7b60c292e 100644
--- a/ggml/src/ggml-sycl/softmax.cpp
+++ b/ggml/src/ggml-sycl/softmax.cpp
@@ -127,11 +127,11 @@ static void soft_max_f32_submitter(const float * x, const T * mask, float * dst,
                                    const int nrows_y, const float scale, const float max_bias, const float m0,
                                    const float m1, uint32_t n_head_log2, sycl::range<3> block_nums, sycl::range<3> block_dims,
                                    const size_t n_local_scratch, queue_ptr stream) {
-    stream->submit([&](sycl::handler &cgh) {
+    sycl_launch(stream, [&](sycl::handler & cgh) {
         sycl::local_accessor<float, 1> local_buf_acc(n_local_scratch, cgh);
 
-        cgh.parallel_for(
-            sycl::nd_range<3>(block_nums * block_dims, block_dims),
+        sycl_parallel_for(
+            cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims),
             [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
                 soft_max_f32<vals_smem, ncols_template, block_size_template>(x, mask, dst, ncols_par,
                                                                              nrows_y, scale, max_bias, m0,
diff --git a/ggml/src/ggml-sycl/sycl_hw.cpp b/ggml/src/ggml-sycl/sycl_hw.cpp
index da121ffc2..704114003 100644
--- a/ggml/src/ggml-sycl/sycl_hw.cpp
+++ b/ggml/src/ggml-sycl/sycl_hw.cpp
@@ -1,6 +1,7 @@
 #include "sycl_hw.hpp"
 
-
+// TODO: currently not used
+/*
 sycl_hw_info get_device_hw_info(sycl::device *device_ptr) {
   sycl_hw_info res;
   int32_t id = device_ptr->get_info<sycl::ext::intel::info::device::device_id>();
@@ -11,3 +12,4 @@ sycl_hw_info get_device_hw_info(sycl::device *device_ptr) {
 
   return res;
 }
+*/
diff --git a/ggml/src/ggml-sycl/sycl_hw.hpp b/ggml/src/ggml-sycl/sycl_hw.hpp
index bf689450c..36b140bf0 100644
--- a/ggml/src/ggml-sycl/sycl_hw.hpp
+++ b/ggml/src/ggml-sycl/sycl_hw.hpp
@@ -10,6 +10,8 @@
 
 namespace syclex = sycl::ext::oneapi::experimental;
 
+// TODO: currently not used
+/*
 struct sycl_hw_info {
   syclex::architecture arch;
   int32_t device_id;
@@ -18,6 +20,7 @@ struct sycl_hw_info {
 bool is_in_vector(std::vector<int> &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/tsembd.cpp b/ggml/src/ggml-sycl/tsembd.cpp
index f6ca626ea..721c8fa6f 100644
--- a/ggml/src/ggml-sycl/tsembd.cpp
+++ b/ggml/src/ggml-sycl/tsembd.cpp
@@ -45,14 +45,9 @@ static void timestep_embedding_f32_sycl(
     int num_blocks = (half_ceil + SYCL_TIMESTEP_EMBEDDING_BLOCK_SIZE - 1) / SYCL_TIMESTEP_EMBEDDING_BLOCK_SIZE;
     sycl::range<3> block_dims(1, 1, SYCL_TIMESTEP_EMBEDDING_BLOCK_SIZE);
     sycl::range<3> gridDim(1, ne00, num_blocks);
-    stream->parallel_for(
-        sycl::nd_range<3>(
-            gridDim * block_dims, block_dims),
-        [=](sycl::nd_item<3> item_ct1) {
-            timestep_embedding_f32(
-                x, dst, nb1, dim, max_period, item_ct1
-            );
-        });
+    sycl_parallel_for(stream, sycl::nd_range<3>(gridDim * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
+        timestep_embedding_f32(x, dst, nb1, dim, max_period, item_ct1);
+    });
 }
 
 void ggml_sycl_op_timestep_embedding(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
diff --git a/ggml/src/ggml-sycl/vecdotq.hpp b/ggml/src/ggml-sycl/vecdotq.hpp
index ed3699313..0a5d49994 100644
--- a/ggml/src/ggml-sycl/vecdotq.hpp
+++ b/ggml/src/ggml-sycl/vecdotq.hpp
@@ -284,22 +284,23 @@ template <> struct reorder_vec_dot_q_sycl<GGML_TYPE_Q4_0> {
         return d4 * (sumi * ds8f.x() - (8 * q4_0_traits::vdr_mmvq / q4_0_traits::qi) * ds8f.y());
     }
 
-    __dpct_inline__ float operator()(const void * __restrict__ vbq, const int ibx_offset, const int d_offset,
-                     const block_q8_1 * __restrict__ bq8_1, const int & iqs, int /* nblocks */) {
-        const uint8_t * bq4_0 = static_cast<const uint8_t *>(vbq) + ibx_offset;
-        const ggml_half d     = *(reinterpret_cast<const ggml_half *>(static_cast<const uint8_t *>(vbq) + d_offset));
+    __dpct_inline__ float operator()(const void * __restrict__ vbq, const std::pair<int, int> ibx_offset,
+                                     const std::pair<int, int> d_offset, const int8_t * q8_1_quant_ptr,
+                                     const sycl::half2 * q8_1_ds, const int & iqs) {
+        const uint8_t * bq4_0 = static_cast<const uint8_t *>(vbq) + ibx_offset.first;
+        const ggml_half d = *(reinterpret_cast<const ggml_half *>(static_cast<const uint8_t *>(vbq) + d_offset.first));
         int             v[q4_0_traits::vdr_mmvq];
         int             u[2 * q4_0_traits::vdr_mmvq];
 
-#pragma unroll
 
+#pragma unroll
         for (size_t i = 0; i < q4_0_traits::vdr_mmvq; ++i) {
             v[i]         = get_int_from_uint8(bq4_0, iqs + i);
-            u[2 * i + 0] = get_int_from_int8_aligned(bq8_1->qs, iqs + i);
-            u[2 * i + 1] = get_int_from_int8_aligned(bq8_1->qs, iqs + i + q4_0_traits::qi);
+            u[2 * i + 0] = get_int_from_int8_aligned(q8_1_quant_ptr, iqs + i);
+            u[2 * i + 1] = get_int_from_int8_aligned(q8_1_quant_ptr, iqs + i + q4_0_traits::qi);
         }
 
-        return vec_dot_q4_0_q8_1_impl(v, u, d, bq8_1->ds);
+        return vec_dot_q4_0_q8_1_impl(v, u, d, *q8_1_ds);
     };
 };
 
@@ -346,24 +347,115 @@ template <> struct reorder_vec_dot_q_sycl<GGML_TYPE_Q4_K> {
     using q4_k_block  = ggml_sycl_reordered::block_q_t<GGML_TYPE_Q4_K>;
     using q4_k_traits = typename q4_k_block::traits;
 
-    float operator()(const void * __restrict__ vbq, const int ibx_offset, const int d_offset,
-                     const block_q8_1 * __restrict__ bq8_1, const int & iqs, int nblocks) {
-        const int ib = ibx_offset / (QK_K / 2);
+    __dpct_inline__ float operator()(const void * __restrict__ vbq, const std::pair<int, int> ibx_offset,
+                                     const std::pair<int, int> d_offset, const int8_t * q8_1_quant_ptr,
+                                     const sycl::half2 * q8_1_ds, const int & iqs) {
+        const int ib = ibx_offset.first / (QK_K / 2);
 
         const uint8_t *    base           = static_cast<const uint8_t *>(vbq);
-        const uint8_t *    qs             = base + ibx_offset;
-        const int          total_qs_bytes = nblocks * (QK_K / 2);
-        const uint8_t *    scs            = base + total_qs_bytes + ib * K_SCALE_SIZE;
-        const ggml_half2 * dms            = reinterpret_cast<const ggml_half2 *>(base + d_offset);
+        const uint8_t *    qs             = base + ibx_offset.first;
+        const uint8_t *    scs            = base + d_offset.first + ib * K_SCALE_SIZE;
+        const ggml_half2 * dms            = reinterpret_cast<const ggml_half2 *>(base + d_offset.second);
 
         const int        bq8_offset = QR4_K * ((iqs / 2) / (QI8_1 / 2));
         const int *      q4         = (const int *) (qs + 16 * bq8_offset + 4 * ((iqs / 2) % 4));
         const uint16_t * scales     = (const uint16_t *) scs;
 
-        return vec_dot_q4_K_q8_1_common(q4, scales, *dms, bq8_1, iqs);
+        int   v[2];
+        int   u[2 * QR4_K];
+        float d8[QR4_K];
+
+        v[0] = q4[0];
+        v[1] = q4[4];
+
+        uint16_t  aux[2];
+        const int j = (QR4_K * ((iqs / 2) / (QI8_1 / 2))) / 2;
+        if (j < 2) {
+            aux[0] = scales[j + 0] & 0x3f3f;
+            aux[1] = scales[j + 2] & 0x3f3f;
+        } else {
+            aux[0] = ((scales[j + 2] >> 0) & 0x0f0f) | ((scales[j - 2] & 0xc0c0) >> 2);
+            aux[1] = ((scales[j + 2] >> 4) & 0x0f0f) | ((scales[j - 0] & 0xc0c0) >> 2);
+        }
+
+        const uint8_t * sc = (const uint8_t *) aux;
+        const uint8_t * m  = sc + 2;
+
+        for (int i = 0; i < QR4_K; ++i) {
+            const int8_t* quant_base_ptr = q8_1_quant_ptr + (bq8_offset + i) * QK8_1;
+            sycl::half2 ds_values = *(q8_1_ds + bq8_offset + i);
+
+            d8[i]                   = ds_values[0];
+
+            const int * q8 = (const int *) quant_base_ptr + ((iqs / 2) % 4);
+            u[2 * i + 0]   = q8[0];
+            u[2 * i + 1]   = q8[4];
+        }
+
+        return vec_dot_q4_K_q8_1_impl_vmmq(v, u, sc, m, *dms, d8);
     }
 };
 
+template <> struct reorder_vec_dot_q_sycl<GGML_TYPE_Q6_K> {
+    static constexpr ggml_type gtype = GGML_TYPE_Q6_K;
+
+    using q6_k_block  = ggml_sycl_reordered::block_q_t<GGML_TYPE_Q6_K>;
+    using q6_k_traits = typename q6_k_block::traits;
+
+    __dpct_inline__ float vec_dot_q6_K_q8_1_impl_mmvq(const int vl, const int vh, const int * __restrict__ u,
+                                                      const int8_t * __restrict__ scales, const float d,
+                                                      const float * __restrict__ d8) {
+        float sumf = 0.0f;
+
+#pragma unroll
+        for (int i = 0; i < QR6_K; ++i) {
+            const int sc = scales[4 * i];
+
+            const int vil = (vl >> (4 * i)) & 0x0F0F0F0F;
+
+            const int vih = ((vh >> (4 * i)) << 4) & 0x30303030;
+
+            const int vi = dpct::vectorized_binary<sycl::char4>((vil | vih), 0x20202020,
+                                                                dpct::sub_sat());  // vi = (vil | vih) - 32
+
+            sumf += d8[i] * (dpct::dp4a(vi, u[i], 0) * sc);                        // SIMD dot product
+        }
+
+        return d * sumf;
+    }
+
+    __dpct_inline__ float operator()(const void * __restrict__ vbq, const std::pair<int, int> ibx_offset,
+                     const std::pair<int, int> d_offset, const int8_t * q8_1_quant_ptr, const sycl::half2 * q8_1_ds,
+                     const int iqs) {
+        const int ib = ibx_offset.first / (QK_K / 2);
+
+        const uint8_t *   base   = static_cast<const uint8_t *>(vbq);
+        const uint8_t *   ql     = base + ibx_offset.first;
+        const uint8_t *   qh     = base + ibx_offset.second;
+        const int8_t *    scales = reinterpret_cast<const int8_t *>(base + d_offset.first);
+        const ggml_half * d      = (const ggml_half *) (base + d_offset.second) + ib;
+
+        const int bq8_offset   = 2 * QR6_K * (iqs / (QI6_K / 2)) + (iqs % (QI6_K / 2)) / (QI6_K / 4);
+        const int scale_offset = (QI6_K / 4) * (iqs / (QI6_K / 2)) + (iqs % (QI6_K / 2)) / (QI6_K / 8);
+        const int vh_shift     = 2 * ((iqs % (QI6_K / 2)) / (QI6_K / 4));
+
+        const int vl = get_int_from_uint8(ql, iqs);
+        const int vh = get_int_from_uint8(qh, (QI6_K / 4) * (iqs / (QI6_K / 2)) + iqs % (QI6_K / 4)) >> vh_shift;
+
+        const int8_t * scs = scales + scale_offset;
+
+        int   u[QR6_K];
+        float d8[QR6_K];
+
+#pragma unroll
+        for (int i = 0; i < QR6_K; ++i) {
+            u[i] = get_int_from_int8_aligned(q8_1_quant_ptr + (bq8_offset + 2 * i) * QK8_1, iqs % QI8_1);
+            const sycl::half2 ds_values = *(q8_1_ds + bq8_offset + 2 * i);
+            d8[i]                       = ds_values[0];
+        }
+        return vec_dot_q6_K_q8_1_impl_mmvq(vl, vh, u, scs, *d, d8);
+    }
+};
 #define VDR_Q4_0_Q8_1_MMVQ 2
 #define VDR_Q4_0_Q8_1_MMQ  4
 
diff --git a/ggml/src/ggml-sycl/wkv.cpp b/ggml/src/ggml-sycl/wkv.cpp
index c10e2f764..3ed5bbf35 100644
--- a/ggml/src/ggml-sycl/wkv.cpp
+++ b/ggml/src/ggml-sycl/wkv.cpp
@@ -207,12 +207,11 @@ void ggml_sycl_op_rwkv_wkv6(ggml_backend_sycl_context& ctx, ggml_tensor* dst) {
 
     // Submit kernel
     if (C / H == WKV_BLOCK_SIZE) {
-        stream->submit([&](sycl::handler& cgh) {
+        sycl_launch(stream, [&](sycl::handler & cgh) {
             sycl::local_accessor<float, 1> 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) {
+            sycl_parallel_for(
+                cgh, sycl::nd_range<3>(grid_dims * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
                     rwkv_wkv6_f32_kernel<WKV_BLOCK_SIZE>(
                         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<sycl::access::decorated::no>().get()
@@ -220,12 +219,11 @@ void ggml_sycl_op_rwkv_wkv6(ggml_backend_sycl_context& ctx, ggml_tensor* dst) {
                 });
         });
     } else {
-        stream->submit([&](sycl::handler& cgh) {
+        sycl_launch(stream, [&](sycl::handler & cgh) {
             sycl::local_accessor<float, 1> 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) {
+            sycl_parallel_for(
+                cgh, sycl::nd_range<3>(grid_dims * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
                     rwkv_wkv6_f32_kernel<WKV_BLOCK_SIZE * 2>(
                         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<sycl::access::decorated::no>().get()
@@ -264,12 +262,11 @@ void ggml_sycl_op_rwkv_wkv7(ggml_backend_sycl_context& ctx, ggml_tensor* dst) {
 
     // Submit kernel
     if (C / H == WKV_BLOCK_SIZE) {
-        stream->submit([&](sycl::handler& cgh) {
+        sycl_launch(stream, [&](sycl::handler & cgh) {
             sycl::local_accessor<float, 1> 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) {
+            sycl_parallel_for(
+                cgh, sycl::nd_range<3>(grid_dims * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
                     rwkv_wkv7_f32_kernel<WKV_BLOCK_SIZE>(
                         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<sycl::access::decorated::no>().get()
@@ -277,12 +274,11 @@ void ggml_sycl_op_rwkv_wkv7(ggml_backend_sycl_context& ctx, ggml_tensor* dst) {
                 });
         });
     } else {
-        stream->submit([&](sycl::handler& cgh) {
+        sycl_launch(stream, [&](sycl::handler & cgh) {
             sycl::local_accessor<float, 1> 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) {
+            sycl_parallel_for(
+                cgh, sycl::nd_range<3>(grid_dims * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
                     rwkv_wkv7_f32_kernel<WKV_BLOCK_SIZE * 2>(
                         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<sycl::access::decorated::no>().get()
diff --git a/ggml/src/ggml-vulkan/CMakeLists.txt b/ggml/src/ggml-vulkan/CMakeLists.txt
index 4a88415f9..b97e7bf99 100644
--- a/ggml/src/ggml-vulkan/CMakeLists.txt
+++ b/ggml/src/ggml-vulkan/CMakeLists.txt
@@ -49,15 +49,7 @@ if (Vulkan_FOUND)
                              ../../include/ggml-vulkan.h
                             )
 
-    set(VULKAN_SHADER_GEN_CMAKE_ARGS
-        -DCMAKE_INSTALL_PREFIX=${CMAKE_BINARY_DIR}
-        -DCMAKE_RUNTIME_OUTPUT_DIRECTORY=${CMAKE_RUNTIME_OUTPUT_DIRECTORY}
-    )
-
-    set(VULKAN_SHADER_GEN_CMAKE_BUILD_ARGS "")
-    if (CMAKE_BUILD_TYPE AND CMAKE_BUILD_TYPE MATCHES "Debug|Release|MinSizeRel|RelWithDebInfo")
-        list(APPEND VULKAN_SHADER_GEN_CMAKE_BUILD_ARGS --config=${CMAKE_BUILD_TYPE})
-    endif()
+    set(VULKAN_SHADER_GEN_CMAKE_ARGS "")
 
     # Test all shader extensions
     test_shader_extension_support(
@@ -107,6 +99,7 @@ if (Vulkan_FOUND)
 
     if (GGML_VULKAN_SHADER_DEBUG_INFO)
         add_compile_definitions(GGML_VULKAN_SHADER_DEBUG_INFO)
+        list(APPEND VULKAN_SHADER_GEN_CMAKE_ARGS -DGGML_VULKAN_SHADER_DEBUG_INFO=ON)
     endif()
 
     if (GGML_VULKAN_VALIDATE)
@@ -136,42 +129,54 @@ if (Vulkan_FOUND)
         set(HOST_CMAKE_TOOLCHAIN_FILE "")
     endif()
 
-    # Always use ExternalProject_Add approach
     include(ExternalProject)
 
-    # Add toolchain file if cross-compiling
     if (CMAKE_CROSSCOMPILING)
         list(APPEND VULKAN_SHADER_GEN_CMAKE_ARGS -DCMAKE_TOOLCHAIN_FILE=${HOST_CMAKE_TOOLCHAIN_FILE})
         message(STATUS "vulkan-shaders-gen toolchain file: ${HOST_CMAKE_TOOLCHAIN_FILE}")
     endif()
 
-    # Native build through ExternalProject_Add
     ExternalProject_Add(
         vulkan-shaders-gen
         SOURCE_DIR ${CMAKE_CURRENT_SOURCE_DIR}/vulkan-shaders
-        CMAKE_ARGS ${VULKAN_SHADER_GEN_CMAKE_ARGS}
-        BUILD_COMMAND ${CMAKE_COMMAND} --build . ${VULKAN_SHADER_GEN_CMAKE_BUILD_ARGS}
-        INSTALL_COMMAND ${CMAKE_COMMAND} --install .
-        INSTALL_DIR ${CMAKE_BINARY_DIR}
+        CMAKE_ARGS -DCMAKE_INSTALL_PREFIX=${CMAKE_BINARY_DIR}/$<CONFIG>
+                   -DCMAKE_INSTALL_BINDIR=.
+                   -DCMAKE_BUILD_TYPE=$<CONFIG>
+                   ${VULKAN_SHADER_GEN_CMAKE_ARGS}
+
+        BUILD_COMMAND ${CMAKE_COMMAND} --build . --config $<CONFIG>
+        BUILD_ALWAYS  TRUE
+
+        # NOTE: When DESTDIR is set using Makefile generators and
+        # "make install" triggers the build step, vulkan-shaders-gen
+        # would be installed into the DESTDIR prefix, so it is unset
+        # to ensure that does not happen.
+
+        INSTALL_COMMAND ${CMAKE_COMMAND} -E env --unset=DESTDIR
+                        ${CMAKE_COMMAND} --install . --config $<CONFIG>
     )
-    ExternalProject_Add_StepTargets(vulkan-shaders-gen build install)
 
     set (_ggml_vk_host_suffix $<IF:$<STREQUAL:${CMAKE_HOST_SYSTEM_NAME},Windows>,.exe,>)
-    set (_ggml_vk_genshaders_cmd ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/vulkan-shaders-gen${_ggml_vk_host_suffix})
-    set (_ggml_vk_header     ${CMAKE_CURRENT_BINARY_DIR}/ggml-vulkan-shaders.hpp)
-    set (_ggml_vk_source     ${CMAKE_CURRENT_BINARY_DIR}/ggml-vulkan-shaders.cpp)
-    set (_ggml_vk_input_dir  ${CMAKE_CURRENT_SOURCE_DIR}/vulkan-shaders)
-    set (_ggml_vk_output_dir ${CMAKE_CURRENT_BINARY_DIR}/vulkan-shaders.spv)
+    set (_ggml_vk_genshaders_dir "${CMAKE_BINARY_DIR}/$<CONFIG>")
+    set (_ggml_vk_genshaders_cmd "${_ggml_vk_genshaders_dir}/vulkan-shaders-gen${_ggml_vk_host_suffix}")
+    set (_ggml_vk_header     "${CMAKE_CURRENT_BINARY_DIR}/ggml-vulkan-shaders.hpp")
+    set (_ggml_vk_source     "${CMAKE_CURRENT_BINARY_DIR}/ggml-vulkan-shaders.cpp")
+    set (_ggml_vk_input_dir  "${CMAKE_CURRENT_SOURCE_DIR}/vulkan-shaders")
+    set (_ggml_vk_output_dir "${CMAKE_CURRENT_BINARY_DIR}/vulkan-shaders.spv")
 
-    file(GLOB _ggml_vk_shader_deps "${_ggml_vk_input_dir}/*.comp")
-    set (_ggml_vk_shader_deps ${_ggml_vk_shader_deps} vulkan-shaders-gen)
+    file(GLOB _ggml_vk_shader_files CONFIGURE_DEPENDS "${_ggml_vk_input_dir}/*.comp")
 
-    # Add build and install dependencies for all builds
-    set(_ggml_vk_shader_deps ${_ggml_vk_shader_deps} vulkan-shaders-gen-build vulkan-shaders-gen-install)
+    # Because external projects do not provide source-level tracking,
+    # the vulkan-shaders-gen sources need to be explicitly added to
+    # ensure that changes will cascade into shader re-generation.
+
+    file(GLOB _ggml_vk_shaders_gen_sources
+              CONFIGURE_DEPENDS "${_ggml_vk_input_dir}/*.cpp"
+                                "${_ggml_vk_input_dir}/*.h")
 
     add_custom_command(
         OUTPUT ${_ggml_vk_header}
-                ${_ggml_vk_source}
+               ${_ggml_vk_source}
 
         COMMAND ${_ggml_vk_genshaders_cmd}
             --glslc      ${Vulkan_GLSLC_EXECUTABLE}
@@ -181,7 +186,10 @@ if (Vulkan_FOUND)
             --target-cpp ${_ggml_vk_source}
             --no-clean
 
-        DEPENDS ${_ggml_vk_shader_deps}
+        DEPENDS ${_ggml_vk_shader_files}
+                ${_ggml_vk_shaders_gen_sources}
+                vulkan-shaders-gen
+
         COMMENT "Generate vulkan shaders"
     )
 
diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
index 41d20aa5d..7b8faf178 100644
--- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp
+++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
@@ -78,7 +78,7 @@ static bool is_pow2(uint32_t x) { return x > 1 && (x & (x-1)) == 0; }
 #define VK_VENDOR_ID_INTEL 0x8086
 #define VK_VENDOR_ID_NVIDIA 0x10de
 
-#define VK_DEVICE_DESCRIPTOR_POOL_SIZE 32
+#define VK_DEVICE_DESCRIPTOR_POOL_SIZE 256
 
 #define GGML_VK_MAX_NODES 8192
 
@@ -102,25 +102,11 @@ static bool is_pow2(uint32_t x) { return x > 1 && (x & (x-1)) == 0; }
 
 struct ggml_backend_vk_context;
 
-struct vk_queue {
-    uint32_t queue_family_index;
-    vk::Queue queue;
-    vk::CommandPool pool;
-    uint32_t cmd_buffer_idx;
-    std::vector<vk::CommandBuffer> cmd_buffers;
-
-    vk::PipelineStageFlags stage_flags;
-
-    bool transfer_only;
-};
+#define MAX_PARAMETER_COUNT 8
 
 struct vk_pipeline_struct {
     std::string name;
     vk::ShaderModule shader_module;
-    vk::DescriptorSetLayout dsl;
-    std::vector<vk::DescriptorPool> descriptor_pools;
-    std::vector<vk::DescriptorSet> descriptor_sets;
-    uint32_t descriptor_set_idx;
     vk::PipelineLayout layout;
     vk::Pipeline pipeline;
     uint32_t push_constant_size;
@@ -167,6 +153,45 @@ struct ggml_backend_vk_buffer_type_context {
     vk_device device;
 };
 
+struct vk_queue;
+
+// Stores command pool/buffers. There's an instance of this
+// for each (context,queue) pair and for each (device,queue) pair.
+struct vk_command_pool {
+    void init(vk_device& device, vk_queue *q_);
+    void destroy(vk::Device& device);
+
+    vk::CommandPool pool;
+    uint32_t cmd_buffer_idx;
+    std::vector<vk::CommandBuffer> cmd_buffers;
+
+    vk_queue *q;
+};
+
+// Prevent simultaneous submissions to the same queue.
+// This could be per vk_queue if we stopped having two vk_queue structures
+// sharing the same vk::Queue.
+static std::mutex queue_mutex;
+
+struct vk_queue {
+    uint32_t queue_family_index;
+    vk::Queue queue;
+
+    vk_command_pool cmd_pool;
+
+    vk::PipelineStageFlags stage_flags;
+
+    bool transfer_only;
+
+    // copy everything except the cmd_pool
+    void copyFrom(vk_queue &other) {
+        queue_family_index = other.queue_family_index;
+        queue = other.queue;
+        stage_flags = other.stage_flags;
+        transfer_only = other.transfer_only;
+    }
+};
+
 static const char * ggml_backend_vk_buffer_type_name(ggml_backend_buffer_type_t buft);
 static ggml_backend_buffer_t ggml_backend_vk_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size);
 static size_t ggml_backend_vk_buffer_type_get_alignment(ggml_backend_buffer_type_t buft);
@@ -196,6 +221,22 @@ enum vk_device_architecture {
     AMD_RDNA1,
     AMD_RDNA2,
     AMD_RDNA3,
+    INTEL_XE2,
+};
+
+// HSK x HSV
+enum FaHeadSizes {
+    FA_HEAD_SIZE_64,
+    FA_HEAD_SIZE_80,
+    FA_HEAD_SIZE_96,
+    FA_HEAD_SIZE_112,
+    FA_HEAD_SIZE_128,
+    FA_HEAD_SIZE_192,
+    FA_HEAD_SIZE_192_128,
+    FA_HEAD_SIZE_256,
+    FA_HEAD_SIZE_576_512,
+    FA_HEAD_SIZE_UNSUPPORTED,
+    FA_HEAD_SIZE_COUNT = FA_HEAD_SIZE_UNSUPPORTED,
 };
 
 static vk_device_architecture get_device_architecture(const vk::PhysicalDevice& device) {
@@ -246,12 +287,40 @@ static vk_device_architecture get_device_architecture(const vk::PhysicalDevice&
             }
             return vk_device_architecture::AMD_RDNA2;
         }
+    } else if (props.vendorID == VK_VENDOR_ID_INTEL) {
+        const std::vector<vk::ExtensionProperties> ext_props = device.enumerateDeviceExtensionProperties();
+
+        bool subgroup_size_control = false;
+
+        for (const auto& properties : ext_props) {
+            if (strcmp("VK_EXT_subgroup_size_control", properties.extensionName) == 0) {
+                subgroup_size_control = true;
+            }
+        }
+
+        if (!subgroup_size_control) {
+            return vk_device_architecture::OTHER;
+        }
+
+        vk::PhysicalDeviceProperties2 props2;
+        vk::PhysicalDeviceSubgroupSizeControlPropertiesEXT subgroup_size_control_props;
+
+        props2.pNext = &subgroup_size_control_props;
+        device.getProperties2(&props2);
+
+        if (subgroup_size_control_props.minSubgroupSize == 16) {
+            // Xe2 architecture uses SIMD16 while previous Xe and Gen architecture uses SIMD8.
+            // Minimum subgroup size matches the SIMD width so we distinguish architecture by checking this value.
+            // https://www.intel.com/content/www/us/en/content-details/824434/2024-intel-tech-tour-xe2-and-lunar-lake-s-gpu.html
+            // https://www.intel.com/content/www/us/en/docs/oneapi/optimization-guide-gpu/2025-0/intel-xe-gpu-architecture.html
+            return vk_device_architecture::INTEL_XE2;
+        }
     }
     return vk_device_architecture::OTHER;
 }
 
 struct vk_device_struct {
-    std::mutex mutex;
+    std::recursive_mutex mutex;
 
     vk::PhysicalDevice physical_device;
     vk::PhysicalDeviceProperties properties;
@@ -312,6 +381,8 @@ struct vk_device_struct {
     // set to true to indicate that some shaders need to be compiled after the dryrun
     bool need_compiles {};
 
+    vk::DescriptorSetLayout dsl;
+
     vk_matmul_pipeline pipeline_matmul_f32 {};
     vk_matmul_pipeline pipeline_matmul_f32_f16 {};
     vk_matmul_pipeline pipeline_matmul_bf16 {};
@@ -369,17 +440,25 @@ 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_mul_f32;
     vk_pipeline pipeline_rms_norm_back_f32;
     vk_pipeline pipeline_l2_norm_f32;
 
     // [src/dst 0=fp32,1=fp16]
     vk_pipeline pipeline_gelu[2];
+    vk_pipeline pipeline_gelu_erf[2];
     vk_pipeline pipeline_gelu_quick[2];
     vk_pipeline pipeline_silu[2];
     vk_pipeline pipeline_relu[2];
     vk_pipeline pipeline_tanh[2];
     vk_pipeline pipeline_sigmoid[2];
 
+    vk_pipeline pipeline_geglu[2];
+    vk_pipeline pipeline_reglu[2];
+    vk_pipeline pipeline_swiglu[2];
+    vk_pipeline pipeline_geglu_erf[2];
+    vk_pipeline pipeline_geglu_quick[2];
+
     vk_pipeline pipeline_leaky_relu_f32;
     vk_pipeline pipeline_silu_back_f32;
     vk_pipeline pipeline_diag_mask_inf_f32;
@@ -396,6 +475,7 @@ struct vk_device_struct {
     vk_pipeline pipeline_count_equal_i32;
     vk_pipeline pipeline_im2col_f32, pipeline_im2col_f32_f16;
     vk_pipeline pipeline_timestep_embedding_f32;
+    vk_pipeline pipeline_conv_transpose_1d_f32;
     vk_pipeline pipeline_pool2d_f32;
     vk_pipeline pipeline_rwkv_wkv6_f32;
     vk_pipeline pipeline_rwkv_wkv7_f32;
@@ -404,31 +484,15 @@ struct vk_device_struct {
     vk_pipeline pipeline_conv2d_dw_cwhn_f32;
 
     // [2][2][2] is for {f16acc,f32acc}x{large,small_rows}x{unaligned, aligned}
-    vk_pipeline pipeline_flash_attn_f32_f16_D64_cm2[GGML_TYPE_COUNT][2][2][2];
-    vk_pipeline pipeline_flash_attn_f32_f16_D80_cm2[GGML_TYPE_COUNT][2][2][2];
-    vk_pipeline pipeline_flash_attn_f32_f16_D96_cm2[GGML_TYPE_COUNT][2][2][2];
-    vk_pipeline pipeline_flash_attn_f32_f16_D112_cm2[GGML_TYPE_COUNT][2][2][2];
-    vk_pipeline pipeline_flash_attn_f32_f16_D128_cm2[GGML_TYPE_COUNT][2][2][2];
-    vk_pipeline pipeline_flash_attn_f32_f16_D256_cm2[GGML_TYPE_COUNT][2][2][2];
+    vk_pipeline pipeline_flash_attn_f32_f16_cm2[GGML_TYPE_COUNT][FA_HEAD_SIZE_COUNT][2][2][2];
 
-    vk_pipeline pipeline_flash_attn_f32_f16_D64_cm1[GGML_TYPE_COUNT][2][2][2];
-    vk_pipeline pipeline_flash_attn_f32_f16_D80_cm1[GGML_TYPE_COUNT][2][2][2];
-    vk_pipeline pipeline_flash_attn_f32_f16_D96_cm1[GGML_TYPE_COUNT][2][2][2];
-    vk_pipeline pipeline_flash_attn_f32_f16_D112_cm1[GGML_TYPE_COUNT][2][2][2];
-    vk_pipeline pipeline_flash_attn_f32_f16_D128_cm1[GGML_TYPE_COUNT][2][2][2];
-    vk_pipeline pipeline_flash_attn_f32_f16_D256_cm1[GGML_TYPE_COUNT][2][2][2];
+    vk_pipeline pipeline_flash_attn_f32_f16_cm1[GGML_TYPE_COUNT][FA_HEAD_SIZE_COUNT][2][2][2];
 
-    vk_pipeline pipeline_flash_attn_f32_f16_D64[GGML_TYPE_COUNT][2][2][2];
-    vk_pipeline pipeline_flash_attn_f32_f16_D80[GGML_TYPE_COUNT][2][2][2];
-    vk_pipeline pipeline_flash_attn_f32_f16_D96[GGML_TYPE_COUNT][2][2][2];
-    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_f32_f16[GGML_TYPE_COUNT][FA_HEAD_SIZE_COUNT][2][2][2];
 
     vk_pipeline pipeline_flash_attn_split_k_reduce;
 
     std::unordered_map<std::string, vk_pipeline_ref> pipelines;
-    std::unordered_map<std::string, uint64_t> pipeline_descriptor_set_requirements;
 
     std::vector<std::tuple<void*, size_t, vk_buffer>> pinned_memory;
 
@@ -437,6 +501,8 @@ struct vk_device_struct {
 
     ggml_backend_buffer_type buffer_type;
 
+    bool disable_fusion;
+
 #ifdef GGML_VULKAN_MEMORY_DEBUG
     std::unique_ptr<vk_memory_logger> memory_logger;
 #endif
@@ -444,7 +510,7 @@ struct vk_device_struct {
     // for GGML_VK_PERF_LOGGER
     std::unique_ptr<vk_perf_logger> perf_logger;
     vk::QueryPool query_pool;
-    uint32_t num_queries;
+    int32_t num_queries;
 
     ~vk_device_struct() {
         VK_LOG_DEBUG("destroy device " << name);
@@ -453,10 +519,8 @@ struct vk_device_struct {
 
         ggml_vk_destroy_buffer(sync_staging);
 
-        device.destroyCommandPool(compute_queue.pool);
-        if (!single_queue) {
-            device.destroyCommandPool(transfer_queue.pool);
-        }
+        compute_queue.cmd_pool.destroy(device);
+        transfer_queue.cmd_pool.destroy(device);
 
         for (auto& pipeline : pipelines) {
             if (pipeline.second.expired()) {
@@ -468,10 +532,26 @@ struct vk_device_struct {
         }
         pipelines.clear();
 
+        device.destroyDescriptorSetLayout(dsl);
+
         device.destroy();
     }
 };
 
+void vk_command_pool::init(vk_device& device, vk_queue *q_) {
+    cmd_buffer_idx = 0;
+    q = q_;
+
+    vk::CommandPoolCreateInfo command_pool_create_info(vk::CommandPoolCreateFlags(VK_COMMAND_POOL_CREATE_TRANSIENT_BIT), q->queue_family_index);
+    pool = device->device.createCommandPool(command_pool_create_info);
+}
+
+void vk_command_pool::destroy(vk::Device& device) {
+    device.destroyCommandPool(pool);
+    pool = nullptr;
+    cmd_buffers.clear();
+}
+
 struct vk_buffer_struct {
     vk::Buffer buffer = VK_NULL_HANDLE;
     vk::DeviceMemory device_memory = VK_NULL_HANDLE;
@@ -557,6 +637,8 @@ struct vk_flash_attn_push_constants {
     uint32_t nev2;
     uint32_t nev3;
     uint32_t nem1;
+    uint32_t nem2;
+    uint32_t nem3;
 
     uint32_t nb01;
     uint32_t nb02;
@@ -567,14 +649,12 @@ struct vk_flash_attn_push_constants {
     uint32_t nb21;
     uint32_t nb22;
     uint32_t nb23;
-    uint32_t nb31;
 
     float scale;
     float max_bias;
     float logit_softcap;
 
-    uint32_t mask;
-    uint32_t n_head_log2;
+    uint32_t mask_n_head_log2;
     float m0;
     float m1;
 
@@ -582,6 +662,7 @@ struct vk_flash_attn_push_constants {
     uint32_t split_kv;
     uint32_t k_num;
 };
+static_assert(sizeof(vk_flash_attn_push_constants) <= 128, "sizeof(vk_flash_attn_push_constants) must be <= 128");
 
 struct vk_op_push_constants {
     uint32_t KX;
@@ -590,6 +671,13 @@ struct vk_op_push_constants {
     float param2;
 };
 
+struct vk_op_glu_push_constants {
+    uint32_t N;
+    uint32_t ne00;
+    uint32_t ne20;
+    uint32_t mode;  // 0: default, 1: swapped, 2: split
+};
+
 struct vk_op_unary_push_constants {
     uint32_t ne;
     uint32_t ne00; uint32_t ne01; uint32_t ne02; uint32_t ne03; uint32_t nb00; uint32_t nb01; uint32_t nb02; uint32_t nb03;
@@ -673,6 +761,14 @@ struct vk_op_rope_push_constants {
 struct vk_op_soft_max_push_constants {
     uint32_t KX;
     uint32_t KY;
+    uint32_t ne00;
+    uint32_t ne01;
+    uint32_t ne02;
+    uint32_t ne12;
+    uint32_t ne13;
+    uint32_t nb11;
+    uint32_t nb12;
+    uint32_t nb13;
     float scale;
     float max_bias;
     float m0;
@@ -706,6 +802,21 @@ struct vk_op_timestep_embedding_push_constants {
     uint32_t max_period;
 };
 
+struct vk_op_conv_transpose_1d_push_constants {
+    uint32_t Cout;
+    uint32_t Cin;
+    uint32_t K;
+    uint32_t L;
+    uint32_t KL;
+
+    uint32_t nb01;
+    uint32_t nb02;
+    uint32_t nb11;
+    uint32_t nb1;
+
+    int32_t s0;
+};
+
 struct vk_op_pool2d_push_constants {
     uint32_t IW; uint32_t IH;
     uint32_t OW; uint32_t OH;
@@ -774,7 +885,7 @@ struct vk_context_struct {
     std::vector<vk_staging_memcpy> in_memcpys;
     std::vector<vk_staging_memcpy> out_memcpys;
 
-    vk_queue * q;
+    vk_command_pool * p {};
 };
 typedef std::shared_ptr<vk_context_struct> vk_context;
 typedef std::weak_ptr<vk_context_struct> vk_context_ref;
@@ -885,6 +996,18 @@ struct ggml_backend_vk_context {
     vk_context_ref transfer_ctx;
 
     std::vector<vk_context_ref> tensor_ctxs;
+
+    std::vector<vk::DescriptorPool> descriptor_pools;
+    std::vector<vk::DescriptorSet> descriptor_sets;
+    uint32_t descriptor_set_idx {};
+    uint32_t pipeline_descriptor_set_requirements {};
+
+    vk_command_pool compute_cmd_pool;
+    vk_command_pool transfer_cmd_pool;
+
+    // number of additional consecutive nodes that are being fused with the
+    // node currently being processed
+    int num_additional_fused_ops {};
 };
 
 static void * const vk_ptr_base = (void *)(uintptr_t) 0x1000;  // NOLINT
@@ -948,6 +1071,14 @@ void vk_memory_logger::log_deallocation(vk_buffer_ref buf_ref) {
 struct vk_instance_t {
     vk::Instance instance;
 
+    bool debug_utils_support = false;  // VK_EXT_debug_utils enabled
+    PFN_vkSetDebugUtilsObjectNameEXT pfn_vkSetDebugUtilsObjectNameEXT = {};
+    PFN_vkQueueBeginDebugUtilsLabelEXT pfn_vkQueueBeginDebugUtilsLabelEXT = {};
+    PFN_vkQueueEndDebugUtilsLabelEXT   pfn_vkQueueEndDebugUtilsLabelEXT   = {};
+    PFN_vkCmdBeginDebugUtilsLabelEXT   pfn_vkCmdBeginDebugUtilsLabelEXT   = {};
+    PFN_vkCmdEndDebugUtilsLabelEXT pfn_vkCmdEndDebugUtilsLabelEXT = {};
+    PFN_vkCmdInsertDebugUtilsLabelEXT  pfn_vkCmdInsertDebugUtilsLabelEXT  = {};
+
     std::vector<size_t> device_indices;
     vk_device devices[GGML_VK_MAX_DEVICES];
 };
@@ -962,8 +1093,8 @@ static size_t vk_skip_checks;
 static size_t vk_output_tensor;
 
 static void ggml_vk_print_tensor(const ggml_tensor * tensor, const char * name);
-static void ggml_vk_check_results_0(ggml_tensor * tensor);
-static void ggml_vk_check_results_1(ggml_tensor * tensor);
+static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_cgraph * cgraph, int tensor_idx);
+static void ggml_vk_check_results_1(ggml_backend_vk_context * ctx, ggml_cgraph * cgraph, int tensor_idx);
 #endif
 
 typedef void (*ggml_vk_func_t)(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst);
@@ -1015,39 +1146,19 @@ static void ggml_vk_create_pipeline_func(vk_device& device, vk_pipeline& pipelin
                  ", (" << wg_denoms[0] << "," << wg_denoms[1] << "," << wg_denoms[2] << "), specialization_constants, " <<
                  disable_robustness << ", " << require_full_subgroups << ", " << required_subgroup_size << ")");
     GGML_ASSERT(parameter_count > 0);
+    GGML_ASSERT(parameter_count <= MAX_PARAMETER_COUNT);
     GGML_ASSERT(wg_denoms[0] > 0 && wg_denoms[1] > 0 && wg_denoms[2] > 0); // NOLINT
 
     vk::ShaderModuleCreateInfo shader_module_create_info({}, spv_size, reinterpret_cast<const uint32_t *>(spv_data));
     pipeline->shader_module = device->device.createShaderModule(shader_module_create_info);
 
-    std::vector<vk::DescriptorSetLayoutBinding> dsl_binding;
-    std::vector<vk::DescriptorBindingFlags> dsl_binding_flags;
-    for (uint32_t i = 0; i < parameter_count; i++) {
-        dsl_binding.push_back({i, vk::DescriptorType::eStorageBuffer, 1, vk::ShaderStageFlagBits::eCompute});
-        dsl_binding_flags.push_back({});
-    }
-
-    vk::DescriptorSetLayoutBindingFlagsCreateInfo dslbfci = { dsl_binding_flags };
-
     vk::PushConstantRange pcr(
         vk::ShaderStageFlagBits::eCompute,
         0,
         pipeline->push_constant_size
     );
 
-    vk::DescriptorSetLayoutCreateInfo descriptor_set_layout_create_info(
-        {},
-        dsl_binding);
-    descriptor_set_layout_create_info.setPNext(&dslbfci);
-    pipeline->dsl = device->device.createDescriptorSetLayout(descriptor_set_layout_create_info);
-
-    vk::DescriptorPoolSize descriptor_pool_size(vk::DescriptorType::eStorageBuffer, pipeline->parameter_count * VK_DEVICE_DESCRIPTOR_POOL_SIZE);
-    vk::DescriptorPoolCreateInfo descriptor_pool_create_info({}, VK_DEVICE_DESCRIPTOR_POOL_SIZE, descriptor_pool_size);
-    pipeline->descriptor_pools.push_back(device->device.createDescriptorPool(descriptor_pool_create_info));
-
-    pipeline->descriptor_set_idx = 0;
-
-    vk::PipelineLayoutCreateInfo pipeline_layout_create_info(vk::PipelineLayoutCreateFlags(), pipeline->dsl, pcr);
+    vk::PipelineLayoutCreateInfo pipeline_layout_create_info(vk::PipelineLayoutCreateFlags(), device->dsl, pcr);
     pipeline->layout = device->device.createPipelineLayout(pipeline_layout_create_info);
 
     std::vector<vk::SpecializationMapEntry> specialization_entries(specialization_constants.size());
@@ -1107,8 +1218,16 @@ static void ggml_vk_create_pipeline_func(vk_device& device, vk_pipeline& pipelin
     }
     pipeline->compiled = true;
 
+    if (vk_instance.debug_utils_support) {
+        vk::DebugUtilsObjectNameInfoEXT duoni;
+        duoni.objectType = vk::ObjectType::ePipeline;
+        duoni.pObjectName = pipeline->name.c_str();
+        duoni.objectHandle = reinterpret_cast<uint64_t>(static_cast<VkPipeline_T*>(pipeline->pipeline));
+        vk_instance.pfn_vkSetDebugUtilsObjectNameEXT(device->device, &static_cast<VkDebugUtilsObjectNameInfoEXT &>(duoni));
+    }
+
     {
-        std::lock_guard<std::mutex> guard(device->mutex);
+        std::lock_guard<std::recursive_mutex> guard(device->mutex);
         device->pipelines.insert({ pipeline->name, pipeline });
     }
 
@@ -1122,15 +1241,6 @@ static void ggml_vk_create_pipeline_func(vk_device& device, vk_pipeline& pipelin
 
 static void ggml_vk_destroy_pipeline(vk::Device& device, vk_pipeline& pipeline) {
     VK_LOG_DEBUG("ggml_pipeline_destroy_pipeline(" << pipeline->name << ")");
-    for (auto& pool : pipeline->descriptor_pools) {
-        device.destroyDescriptorPool(pool);
-    }
-    pipeline->descriptor_pools.clear();
-    pipeline->descriptor_sets.clear();
-    pipeline->descriptor_set_idx = 0;
-
-    device.destroyDescriptorSetLayout(pipeline->dsl);
-
     device.destroyPipelineLayout(pipeline->layout);
 
     device.destroyShaderModule(pipeline->shader_module);
@@ -1138,97 +1248,77 @@ static void ggml_vk_destroy_pipeline(vk::Device& device, vk_pipeline& pipeline)
     device.destroyPipeline(pipeline->pipeline);
 }
 
-static void ggml_pipeline_request_descriptor_sets(vk_device& device, vk_pipeline& pipeline, uint32_t n) {
+static void ggml_pipeline_request_descriptor_sets(ggml_backend_vk_context *ctx, vk_pipeline& pipeline, uint32_t n) {
     VK_LOG_DEBUG("ggml_pipeline_request_descriptor_sets(" << pipeline->name << ", " << n << ")");
-    device->pipeline_descriptor_set_requirements[pipeline->name] += n;
+    ctx->pipeline_descriptor_set_requirements += n;
     if (!pipeline->compiled) {
         pipeline->needed = true;
-        device->need_compiles = true;
+        ctx->device->need_compiles = true;
     }
 }
 
-static void ggml_pipeline_allocate_descriptor_sets(vk_device& device) {
-    std::lock_guard<std::mutex> guard(device->mutex);
+static void ggml_pipeline_allocate_descriptor_sets(ggml_backend_vk_context * ctx) {
 
-    for (auto& pair : device->pipeline_descriptor_set_requirements) {
-        vk_pipeline pipeline = device->pipelines.at(pair.first).lock();
-        const uint64_t n = pair.second;
+    if (ctx->descriptor_sets.size() >= ctx->pipeline_descriptor_set_requirements) {
+        // Enough descriptors are available
+        return;
+    }
 
-        VK_LOG_DEBUG("ggml_pipeline_allocate_descriptor_sets(" << pipeline->name << ", " << n << ")");
+    vk_device& device = ctx->device;
 
-        if (pipeline->descriptor_sets.size() >= pipeline->descriptor_set_idx + n) {
-            // Enough descriptors are available
-            continue;
+    uint32_t to_alloc = ctx->pipeline_descriptor_set_requirements - ctx->descriptor_sets.size();
+    uint32_t pool_remaining = VK_DEVICE_DESCRIPTOR_POOL_SIZE - ctx->descriptor_sets.size() % VK_DEVICE_DESCRIPTOR_POOL_SIZE;
+    uint32_t pool_idx = ctx->descriptor_sets.size() / VK_DEVICE_DESCRIPTOR_POOL_SIZE;
+
+    while (to_alloc > 0) {
+        const uint32_t alloc_count = std::min(pool_remaining, to_alloc);
+        to_alloc -= alloc_count;
+        pool_remaining = VK_DEVICE_DESCRIPTOR_POOL_SIZE;
+
+        if (pool_idx >= ctx->descriptor_pools.size()) {
+            vk::DescriptorPoolSize descriptor_pool_size(vk::DescriptorType::eStorageBuffer, MAX_PARAMETER_COUNT * VK_DEVICE_DESCRIPTOR_POOL_SIZE);
+            vk::DescriptorPoolCreateInfo descriptor_pool_create_info({}, VK_DEVICE_DESCRIPTOR_POOL_SIZE, descriptor_pool_size);
+            ctx->descriptor_pools.push_back(device->device.createDescriptorPool(descriptor_pool_create_info));
         }
 
-        uint32_t to_alloc = pipeline->descriptor_set_idx + n - pipeline->descriptor_sets.size();
-        uint32_t pool_remaining = VK_DEVICE_DESCRIPTOR_POOL_SIZE - pipeline->descriptor_sets.size() % VK_DEVICE_DESCRIPTOR_POOL_SIZE;
-        uint32_t pool_idx = pipeline->descriptor_sets.size() / VK_DEVICE_DESCRIPTOR_POOL_SIZE;
-
-        while (to_alloc > 0) {
-            const uint32_t alloc_count = std::min(pool_remaining, to_alloc);
-            to_alloc -= alloc_count;
-            pool_remaining = VK_DEVICE_DESCRIPTOR_POOL_SIZE;
-
-            if (pool_idx >= pipeline->descriptor_pools.size()) {
-                vk::DescriptorPoolSize descriptor_pool_size(vk::DescriptorType::eStorageBuffer, pipeline->parameter_count * VK_DEVICE_DESCRIPTOR_POOL_SIZE);
-                vk::DescriptorPoolCreateInfo descriptor_pool_create_info({}, VK_DEVICE_DESCRIPTOR_POOL_SIZE, descriptor_pool_size);
-                pipeline->descriptor_pools.push_back(device->device.createDescriptorPool(descriptor_pool_create_info));
-            }
-
-            std::vector<vk::DescriptorSetLayout> layouts(alloc_count);
-            for (uint32_t i = 0; i < alloc_count; i++) {
-                layouts[i] = pipeline->dsl;
-            }
-            vk::DescriptorSetAllocateInfo descriptor_set_alloc_info(pipeline->descriptor_pools[pool_idx], alloc_count, layouts.data());
-            std::vector<vk::DescriptorSet> sets = device->device.allocateDescriptorSets(descriptor_set_alloc_info);
-            pipeline->descriptor_sets.insert(pipeline->descriptor_sets.end(), sets.begin(), sets.end());
-
-            pool_idx++;
+        std::vector<vk::DescriptorSetLayout> layouts(alloc_count);
+        for (uint32_t i = 0; i < alloc_count; i++) {
+            layouts[i] = device->dsl;
         }
+        vk::DescriptorSetAllocateInfo descriptor_set_alloc_info(ctx->descriptor_pools[pool_idx], alloc_count, layouts.data());
+        std::vector<vk::DescriptorSet> sets = device->device.allocateDescriptorSets(descriptor_set_alloc_info);
+        ctx->descriptor_sets.insert(ctx->descriptor_sets.end(), sets.begin(), sets.end());
+
+        pool_idx++;
     }
 }
 
-static void ggml_pipeline_cleanup(vk_pipeline& pipeline) {
-    VK_LOG_DEBUG("ggml_pipeline_cleanup(" << pipeline->name << ")");
-    pipeline->descriptor_set_idx = 0;
-}
-
-static vk::CommandBuffer ggml_vk_create_cmd_buffer(vk_device& device, vk_queue& q) {
+static vk::CommandBuffer ggml_vk_create_cmd_buffer(vk_device& device, vk_command_pool& p) {
     VK_LOG_DEBUG("ggml_vk_create_cmd_buffer()");
-    std::lock_guard<std::mutex> guard(device->mutex);
 
-    if (q.cmd_buffers.size() > q.cmd_buffer_idx) {
+    if (p.cmd_buffers.size() > p.cmd_buffer_idx) {
         // Reuse command buffer
-        return q.cmd_buffers[q.cmd_buffer_idx++];
+        return p.cmd_buffers[p.cmd_buffer_idx++];
     }
 
     vk::CommandBufferAllocateInfo command_buffer_alloc_info(
-        q.pool,
+        p.pool,
         vk::CommandBufferLevel::ePrimary,
         1);
     const std::vector<vk::CommandBuffer> cmd_buffers = device->device.allocateCommandBuffers(command_buffer_alloc_info);
     auto buf = cmd_buffers.front();
 
-    q.cmd_buffers.push_back(buf);
-    q.cmd_buffer_idx++;
+    p.cmd_buffers.push_back(buf);
+    p.cmd_buffer_idx++;
 
     return buf;
 }
 
-static vk_submission ggml_vk_create_submission(vk_device& device, vk_queue& q, std::vector<vk_semaphore> wait_semaphores, std::vector<vk_semaphore> signal_semaphores) {
-    VK_LOG_DEBUG("ggml_vk_create_submission()");
-    vk_submission s;
-    s.buffer = ggml_vk_create_cmd_buffer(device, q);
-    s.wait_semaphores = std::move(wait_semaphores);
-    s.signal_semaphores = std::move(signal_semaphores);
-    return s;
-}
-
 static void ggml_vk_submit(vk_context& ctx, vk::Fence fence) {
     if (ctx->seqs.empty()) {
         if (fence) {
-            ctx->q->queue.submit({}, fence);
+            std::lock_guard<std::mutex> guard(queue_mutex);
+            ctx->p->q->queue.submit({}, fence);
         }
         return;
     }
@@ -1267,7 +1357,7 @@ static void ggml_vk_submit(vk_context& ctx, vk::Fence fence) {
             tl_signal_vals.push_back({});
             tl_signal_semaphores.push_back({});
             for (size_t i = 0; i < submission.wait_semaphores.size(); i++) {
-                stage_flags[idx].push_back(ctx->q->stage_flags);
+                stage_flags[idx].push_back(ctx->p->q->stage_flags);
                 tl_wait_vals[idx].push_back(submission.wait_semaphores[i].value);
                 tl_wait_semaphores[idx].push_back(submission.wait_semaphores[i].s);
             }
@@ -1297,7 +1387,8 @@ static void ggml_vk_submit(vk_context& ctx, vk::Fence fence) {
         }
     }
 
-    ctx->q->queue.submit(submit_infos, fence);
+    std::lock_guard<std::mutex> guard(queue_mutex);
+    ctx->p->q->queue.submit(submit_infos, fence);
 
     ctx->seqs.clear();
 }
@@ -1350,33 +1441,30 @@ static uint32_t ggml_vk_find_queue_family_index(std::vector<vk::QueueFamilyPrope
 
 static void ggml_vk_create_queue(vk_device& device, vk_queue& q, uint32_t queue_family_index, uint32_t queue_index, vk::PipelineStageFlags&& stage_flags, bool transfer_only) {
     VK_LOG_DEBUG("ggml_vk_create_queue()");
-    std::lock_guard<std::mutex> guard(device->mutex);
+    std::lock_guard<std::recursive_mutex> guard(device->mutex);
 
     q.queue_family_index = queue_family_index;
     q.transfer_only = transfer_only;
 
-    vk::CommandPoolCreateInfo command_pool_create_info_compute(vk::CommandPoolCreateFlags(VK_COMMAND_POOL_CREATE_TRANSIENT_BIT), queue_family_index);
-    q.pool = device->device.createCommandPool(command_pool_create_info_compute);
-
-    q.cmd_buffer_idx = 0;
+    q.cmd_pool.init(device, &q);
 
     q.queue = device->device.getQueue(queue_family_index, queue_index);
 
     q.stage_flags = stage_flags;
 }
 
-static vk_context ggml_vk_create_context(ggml_backend_vk_context * ctx, vk_queue& q) {
+static vk_context ggml_vk_create_context(ggml_backend_vk_context * ctx, vk_command_pool& p) {
     vk_context result = std::make_shared<vk_context_struct>();
     VK_LOG_DEBUG("ggml_vk_create_context(" << result << ")");
     ctx->gc.contexts.emplace_back(result);
-    result->q = &q;
+    result->p = &p;
     return result;
 }
 
-static vk_context ggml_vk_create_temporary_context(vk_queue& q) {
+static vk_context ggml_vk_create_temporary_context(vk_command_pool& p) {
     vk_context result = std::make_shared<vk_context_struct>();
     VK_LOG_DEBUG("ggml_vk_create_temporary_context(" << result << ")");
-    result->q = &q;
+    result->p = &p;
     return result;
 }
 
@@ -1409,15 +1497,29 @@ static vk::Event ggml_vk_create_event(ggml_backend_vk_context * ctx) {
     return ctx->gc.events[ctx->event_idx++];
 }
 
-static void ggml_vk_queue_cleanup(vk_device& device, vk_queue& q) {
-    VK_LOG_DEBUG("ggml_vk_queue_cleanup()");
-    std::lock_guard<std::mutex> guard(device->mutex);
+static void ggml_vk_command_pool_cleanup(vk_device& device, vk_command_pool& p) {
+    VK_LOG_DEBUG("ggml_vk_command_pool_cleanup()");
 
     // Requires command buffers to be done
-    device->device.resetCommandPool(q.pool);
-    q.cmd_buffer_idx = 0;
+    device->device.resetCommandPool(p.pool);
+    p.cmd_buffer_idx = 0;
 }
 
+static void ggml_vk_queue_command_pools_cleanup(vk_device& device) {
+    VK_LOG_DEBUG("ggml_vk_queue_command_pools_cleanup()");
+
+    // Arbitrary frequency to cleanup/reuse command buffers
+    static constexpr uint32_t cleanup_frequency = 10;
+
+    if (device->compute_queue.cmd_pool.cmd_buffer_idx >= cleanup_frequency) {
+        ggml_vk_command_pool_cleanup(device, device->compute_queue.cmd_pool);
+    }
+    if (device->transfer_queue.cmd_pool.cmd_buffer_idx >= cleanup_frequency) {
+        ggml_vk_command_pool_cleanup(device, device->transfer_queue.cmd_pool);
+    }
+}
+
+
 static uint32_t find_properties(const vk::PhysicalDeviceMemoryProperties* mem_props, vk::MemoryRequirements* mem_req, vk::MemoryPropertyFlags flags) {
     for (uint32_t i = 0; i < mem_props->memoryTypeCount; ++i) {
         vk::MemoryType memory_type = mem_props->memoryTypes[i];
@@ -1436,8 +1538,6 @@ static vk_buffer ggml_vk_create_buffer(vk_device& device, size_t size, vk::Memor
         throw vk::OutOfDeviceMemoryError("Requested buffer size exceeds device memory allocation limit");
     }
 
-    std::lock_guard<std::mutex> guard(device->mutex);
-
     vk_buffer buf = std::make_shared<vk_buffer_struct>();
 
     if (size == 0) {
@@ -1566,11 +1666,11 @@ static vk_subbuffer ggml_vk_subbuffer(vk_buffer& buf) {
 static void ggml_vk_sync_buffers(vk_context& ctx) {
     VK_LOG_DEBUG("ggml_vk_sync_buffers()");
 
-    const bool transfer_queue = ctx->q->transfer_only;
+    const bool transfer_queue = ctx->p->q->transfer_only;
 
     ctx->s->buffer.pipelineBarrier(
-        ctx->q->stage_flags,
-        ctx->q->stage_flags,
+        ctx->p->q->stage_flags,
+        ctx->p->q->stage_flags,
         {},
         { {
           { !transfer_queue ? (vk::AccessFlagBits::eShaderRead | vk::AccessFlagBits::eShaderWrite | vk::AccessFlagBits::eTransferRead | vk::AccessFlagBits::eTransferWrite) : (vk::AccessFlagBits::eTransferRead | vk::AccessFlagBits::eTransferWrite) },
@@ -1589,8 +1689,8 @@ static void ggml_vk_wait_events(vk_context& ctx, std::vector<vk::Event>&& events
 
     ctx->s->buffer.waitEvents(
         events,
-        ctx->q->stage_flags,
-        ctx->q->stage_flags,
+        ctx->p->q->stage_flags,
+        ctx->p->q->stage_flags,
         {},
         {},
         {}
@@ -1603,6 +1703,35 @@ enum FaCodePath {
     FA_COOPMAT2,
 };
 
+static FaHeadSizes fa_get_head_sizes(uint32_t hsk, uint32_t hsv) {
+    if (hsk != 192 && hsk != 576 && hsk != hsv) {
+        return FA_HEAD_SIZE_UNSUPPORTED;
+    }
+    switch (hsk) {
+    case 64: return FA_HEAD_SIZE_64;
+    case 80: return FA_HEAD_SIZE_80;
+    case 96: return FA_HEAD_SIZE_96;
+    case 112: return FA_HEAD_SIZE_112;
+    case 128: return FA_HEAD_SIZE_128;
+    case 192:
+        if (hsv == 192) {
+            return FA_HEAD_SIZE_192;
+        } else if (hsv == 128) {
+            return FA_HEAD_SIZE_192_128;
+        } else {
+            return FA_HEAD_SIZE_UNSUPPORTED;
+        }
+    case 256: return FA_HEAD_SIZE_256;
+    case 576:
+        if (hsv == 512) {
+            return FA_HEAD_SIZE_576_512;
+        } else {
+            return FA_HEAD_SIZE_UNSUPPORTED;
+        }
+    default: return FA_HEAD_SIZE_UNSUPPORTED;
+    }
+}
+
 // number of rows/cols for flash attention shader
 static constexpr uint32_t flash_attention_num_small_rows = 32;
 static constexpr uint32_t scalar_flash_attention_num_small_rows = 1;
@@ -1623,8 +1752,9 @@ static uint32_t get_fa_num_small_rows(FaCodePath path) {
     }
 }
 
-static std::array<uint32_t, 2> fa_rows_cols(FaCodePath path, uint32_t D, uint32_t clamp, ggml_type type, bool small_rows) {
+static std::array<uint32_t, 2> fa_rows_cols(FaCodePath path, uint32_t hsk, uint32_t hsv, uint32_t clamp, ggml_type type, bool small_rows) {
     GGML_UNUSED(clamp);
+    GGML_UNUSED(hsv);
 
     if (path == FA_SCALAR) {
         if (small_rows) {
@@ -1648,7 +1778,7 @@ static std::array<uint32_t, 2> fa_rows_cols(FaCodePath path, uint32_t D, uint32_
     }
 
     // small cols to reduce register count
-    if (ggml_is_quantized(type) || D == 256) {
+    if (ggml_is_quantized(type) || hsk >= 256) {
         return {64, 32};
     }
     return {64, 64};
@@ -1941,19 +2071,21 @@ static void ggml_vk_load_shaders(vk_device& device) {
                                       parameter_count, wg_denoms, specialization_constants, disable_robustness, require_full_subgroups, required_subgroup_size));
     };
 
-    auto const &fa_wg_denoms = [&](FaCodePath path, uint32_t D, uint32_t clamp, ggml_type type, bool small_rows) -> std::array<uint32_t, 3> {
-        return {fa_rows_cols(path, D, clamp, type, small_rows)[0], 1, 1};
+    auto const &fa_wg_denoms = [&](FaCodePath path, uint32_t hsk, uint32_t hsv, uint32_t clamp, ggml_type type, bool small_rows) -> std::array<uint32_t, 3> {
+        return {fa_rows_cols(path, hsk, hsv, clamp, type, small_rows)[0], 1, 1};
     };
 
-    auto const &fa_spec_constants = [&](FaCodePath path, uint32_t D, uint32_t clamp, ggml_type type, bool small_rows) -> std::vector<uint32_t> {
+    auto const &fa_spec_constants = [&](FaCodePath path, uint32_t hsk, uint32_t hsv, uint32_t clamp, ggml_type type, bool small_rows) -> std::vector<uint32_t> {
         // For large number of rows, 128 invocations seems to work best.
         // For small number of rows (e.g. N==1), 256 works better. But matrix granularity for 256 is 32, so we
         // can't use 256 for D==80.
         // For scalar, use 128 (arbitrary)
+        // The same D_split value is used for both HSK and HSV, so just base it on the union of the LSBs.
+        const uint32_t D = (hsk|hsv);
         uint32_t wg_size = (path == FA_SCALAR || path == FA_COOPMAT1)
                             ? scalar_flash_attention_workgroup_size
                             : ((small_rows && (D % 32) == 0) ? 256 : 128);
-        auto rows_cols = fa_rows_cols(path, D, clamp, type, small_rows);
+        auto rows_cols = fa_rows_cols(path, hsk, hsv, clamp, type, small_rows);
 
         // D_split can't be larger than a subgroup because we use subgroupShuffle to reduce it.
         // D_split can't be larger than the LSB of D divided by 4 due to vectorization in the shader.
@@ -1962,26 +2094,29 @@ static void ggml_vk_load_shaders(vk_device& device) {
 
         // 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, D_split};
+        return {wg_size, rows_cols[0], rows_cols[1], hsk, hsv, clamp, D_split};
     };
 
-#define CREATE_FA2(TYPE, NAMELC, FAPATH, SUFFIX, D) \
-        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16_D ## D ## SUFFIX[TYPE][0][0][0], "flash_attn_f32_f16_D" #D "_f16acc"         #NAMELC #SUFFIX,           flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _len,  flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _data,  "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, D,1,TYPE,false), fa_spec_constants(FAPATH, D,1,TYPE,false), 1,                                      true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? 32 : 0));     \
-        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16_D ## D ## SUFFIX[TYPE][0][0][1], "flash_attn_f32_f16_D" #D "_aligned_f16acc" #NAMELC #SUFFIX,           flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _len,  flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _data,  "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, D,0,TYPE,false), fa_spec_constants(FAPATH, D,0,TYPE,false), fa_rows_cols(FAPATH,D,0,TYPE,false)[1], true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? 32 : 0));     \
-        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16_D ## D ## SUFFIX[TYPE][1][0][0], "flash_attn_f32_f16_D" #D "_f32acc"         #NAMELC #SUFFIX,           flash_attn_f32_f16_ ## NAMELC ##     SUFFIX ## _len,         flash_attn_f32_f16_ ## NAMELC ##     SUFFIX ## _data,         "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, D,1,TYPE,false), fa_spec_constants(FAPATH, D,1,TYPE,false), 1,                                      true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? 32 : 0));     \
-        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16_D ## D ## SUFFIX[TYPE][1][0][1], "flash_attn_f32_f16_D" #D "_aligned_f32acc" #NAMELC #SUFFIX,           flash_attn_f32_f16_ ## NAMELC ##     SUFFIX ## _len,         flash_attn_f32_f16_ ## NAMELC ##     SUFFIX ## _data,         "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, D,0,TYPE,false), fa_spec_constants(FAPATH, D,0,TYPE,false), fa_rows_cols(FAPATH,D,0,TYPE,false)[1], true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? 32 : 0));     \
-        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16_D ## D ## SUFFIX[TYPE][0][1][0], "flash_attn_f32_f16_D" #D "_f16acc_smallrows"         #NAMELC #SUFFIX, flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _len,  flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _data,  "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, D,1,TYPE,true), fa_spec_constants(FAPATH, D,1,TYPE,true),   1,                                      true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? 32 : 0));     \
-        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16_D ## D ## SUFFIX[TYPE][0][1][1], "flash_attn_f32_f16_D" #D "_aligned_f16acc_smallrows" #NAMELC #SUFFIX, flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _len,  flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _data,  "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, D,0,TYPE,true), fa_spec_constants(FAPATH, D,0,TYPE,true),   fa_rows_cols(FAPATH,D,0,TYPE,true)[1],  true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? 32 : 0));     \
-        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16_D ## D ## SUFFIX[TYPE][1][1][0], "flash_attn_f32_f16_D" #D "_f32acc_smallrows"         #NAMELC #SUFFIX, flash_attn_f32_f16_ ## NAMELC ##     SUFFIX ## _len,         flash_attn_f32_f16_ ## NAMELC ##     SUFFIX ## _data,         "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, D,1,TYPE,true), fa_spec_constants(FAPATH, D,1,TYPE,true),   1,                                      true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? 32 : 0));     \
-        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16_D ## D ## SUFFIX[TYPE][1][1][1], "flash_attn_f32_f16_D" #D "_aligned_f32acc_smallrows" #NAMELC #SUFFIX, flash_attn_f32_f16_ ## NAMELC ##     SUFFIX ## _len,         flash_attn_f32_f16_ ## NAMELC ##     SUFFIX ## _data,         "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, D,0,TYPE,true), fa_spec_constants(FAPATH, D,0,TYPE,true),   fa_rows_cols(FAPATH,D,0,TYPE,true)[1],  true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? 32 : 0));     \
+#define CREATE_FA2(TYPE, NAMELC, FAPATH, SUFFIX, HSK, HSV, HEAD_SIZES) \
+        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16 ## SUFFIX[TYPE][FA_HEAD_SIZE_##HEAD_SIZES][0][0][0], "flash_attn_f32_f16_" #HEAD_SIZES "_f16acc"         #NAMELC #SUFFIX,           flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _len,  flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _data,  "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,1,TYPE,false), fa_spec_constants(FAPATH, HSK,HSV,1,TYPE,false), 1,                                            true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? 32 : 0));     \
+        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16 ## SUFFIX[TYPE][FA_HEAD_SIZE_##HEAD_SIZES][0][0][1], "flash_attn_f32_f16_" #HEAD_SIZES "_aligned_f16acc" #NAMELC #SUFFIX,           flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _len,  flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _data,  "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,0,TYPE,false), fa_spec_constants(FAPATH, HSK,HSV,0,TYPE,false), fa_rows_cols(FAPATH,HSK,HSV,0,TYPE,false)[1], true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? 32 : 0));     \
+        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16 ## SUFFIX[TYPE][FA_HEAD_SIZE_##HEAD_SIZES][1][0][0], "flash_attn_f32_f16_" #HEAD_SIZES "_f32acc"         #NAMELC #SUFFIX,           flash_attn_f32_f16_ ## NAMELC ##     SUFFIX ## _len,         flash_attn_f32_f16_ ## NAMELC ##     SUFFIX ## _data,         "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,1,TYPE,false), fa_spec_constants(FAPATH, HSK,HSV,1,TYPE,false), 1,                                            true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? 32 : 0));     \
+        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16 ## SUFFIX[TYPE][FA_HEAD_SIZE_##HEAD_SIZES][1][0][1], "flash_attn_f32_f16_" #HEAD_SIZES "_aligned_f32acc" #NAMELC #SUFFIX,           flash_attn_f32_f16_ ## NAMELC ##     SUFFIX ## _len,         flash_attn_f32_f16_ ## NAMELC ##     SUFFIX ## _data,         "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,0,TYPE,false), fa_spec_constants(FAPATH, HSK,HSV,0,TYPE,false), fa_rows_cols(FAPATH,HSK,HSV,0,TYPE,false)[1], true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? 32 : 0));     \
+        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16 ## SUFFIX[TYPE][FA_HEAD_SIZE_##HEAD_SIZES][0][1][0], "flash_attn_f32_f16_" #HEAD_SIZES "_f16acc_smallrows"         #NAMELC #SUFFIX, flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _len,  flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _data,  "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,1,TYPE,true), fa_spec_constants(FAPATH, HSK,HSV,1,TYPE,true),   1,                                            true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? 32 : 0));     \
+        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16 ## SUFFIX[TYPE][FA_HEAD_SIZE_##HEAD_SIZES][0][1][1], "flash_attn_f32_f16_" #HEAD_SIZES "_aligned_f16acc_smallrows" #NAMELC #SUFFIX, flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _len,  flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _data,  "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,0,TYPE,true), fa_spec_constants(FAPATH, HSK,HSV,0,TYPE,true),   fa_rows_cols(FAPATH,HSK,HSV,0,TYPE,true)[1],  true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? 32 : 0));     \
+        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16 ## SUFFIX[TYPE][FA_HEAD_SIZE_##HEAD_SIZES][1][1][0], "flash_attn_f32_f16_" #HEAD_SIZES "_f32acc_smallrows"         #NAMELC #SUFFIX, flash_attn_f32_f16_ ## NAMELC ##     SUFFIX ## _len,         flash_attn_f32_f16_ ## NAMELC ##     SUFFIX ## _data,         "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,1,TYPE,true), fa_spec_constants(FAPATH, HSK,HSV,1,TYPE,true),   1,                                            true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? 32 : 0));     \
+        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16 ## SUFFIX[TYPE][FA_HEAD_SIZE_##HEAD_SIZES][1][1][1], "flash_attn_f32_f16_" #HEAD_SIZES "_aligned_f32acc_smallrows" #NAMELC #SUFFIX, flash_attn_f32_f16_ ## NAMELC ##     SUFFIX ## _len,         flash_attn_f32_f16_ ## NAMELC ##     SUFFIX ## _data,         "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,0,TYPE,true), fa_spec_constants(FAPATH, HSK,HSV,0,TYPE,true),   fa_rows_cols(FAPATH,HSK,HSV,0,TYPE,true)[1],  true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? 32 : 0));     \
 
 #define CREATE_FA(TYPE, NAMELC, FAPATH, SUFFIX) \
-        CREATE_FA2(TYPE, NAMELC, FAPATH, SUFFIX, 64) \
-        CREATE_FA2(TYPE, NAMELC, FAPATH, SUFFIX, 80) \
-        CREATE_FA2(TYPE, NAMELC, FAPATH, SUFFIX, 96) \
-        CREATE_FA2(TYPE, NAMELC, FAPATH, SUFFIX, 112) \
-        CREATE_FA2(TYPE, NAMELC, FAPATH, SUFFIX, 128) \
-        CREATE_FA2(TYPE, NAMELC, FAPATH, SUFFIX, 256)
+        CREATE_FA2(TYPE, NAMELC, FAPATH, SUFFIX, 64, 64, 64) \
+        CREATE_FA2(TYPE, NAMELC, FAPATH, SUFFIX, 80, 80, 80) \
+        CREATE_FA2(TYPE, NAMELC, FAPATH, SUFFIX, 96, 96, 96) \
+        CREATE_FA2(TYPE, NAMELC, FAPATH, SUFFIX, 112, 112, 112) \
+        CREATE_FA2(TYPE, NAMELC, FAPATH, SUFFIX, 128, 128, 128) \
+        CREATE_FA2(TYPE, NAMELC, FAPATH, SUFFIX, 192, 192, 192) \
+        CREATE_FA2(TYPE, NAMELC, FAPATH, SUFFIX, 192, 128, 192_128) \
+        CREATE_FA2(TYPE, NAMELC, FAPATH, SUFFIX, 256, 256, 256) \
+        CREATE_FA2(TYPE, NAMELC, FAPATH, SUFFIX, 576, 512, 576_512)
 
     CREATE_FA(GGML_TYPE_F16, f16, FA_SCALAR, )
     CREATE_FA(GGML_TYPE_Q4_0, q4_0, FA_SCALAR, )
@@ -2585,7 +2720,8 @@ static void ggml_vk_load_shaders(vk_device& device) {
 
     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_unary_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", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {0, 0}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_rms_norm_mul_f32, "rms_norm_mul_f32", rms_norm_f32_len, rms_norm_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {0, 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);
 
@@ -2674,6 +2810,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
     ggml_vk_create_pipeline(device, device->pipeline_ ## name [1], #name "_f16", name ## _f16_len, name ## _f16_data, "main", 2, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);
 
     CREATE_UNARY(gelu)
+    CREATE_UNARY(gelu_erf)
     CREATE_UNARY(gelu_quick)
     CREATE_UNARY(silu)
     CREATE_UNARY(relu)
@@ -2681,6 +2818,17 @@ static void ggml_vk_load_shaders(vk_device& device) {
     CREATE_UNARY(sigmoid)
 #undef CREATE_UNARY
 
+#define CREATE_GLU(name)  \
+    ggml_vk_create_pipeline(device, device->pipeline_ ## name [0], #name "_f32", name ## _f32_len, name ## _f32_data, "main", 3, sizeof(vk_op_glu_push_constants), {512, 1, 1}, {}, 1, true);  \
+    ggml_vk_create_pipeline(device, device->pipeline_ ## name [1], #name "_f16", name ## _f16_len, name ## _f16_data, "main", 3, sizeof(vk_op_glu_push_constants), {512, 1, 1}, {}, 1, true);
+
+    CREATE_GLU(geglu)
+    CREATE_GLU(reglu)
+    CREATE_GLU(swiglu)
+    CREATE_GLU(geglu_erf)
+    CREATE_GLU(geglu_quick)
+#undef CREATE_GLU
+
     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_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);
 
@@ -2726,6 +2874,8 @@ static void ggml_vk_load_shaders(vk_device& device) {
 
     ggml_vk_create_pipeline(device, device->pipeline_timestep_embedding_f32, "timestep_embedding_f32", timestep_embedding_f32_len, timestep_embedding_f32_data, "main", 2, sizeof(vk_op_timestep_embedding_push_constants), {256, 1, 1}, {}, 1);
 
+    ggml_vk_create_pipeline(device, device->pipeline_conv_transpose_1d_f32, "conv_transpose_1d_f32", conv_transpose_1d_f32_len, conv_transpose_1d_f32_data, "main", 3, sizeof(vk_op_conv_transpose_1d_push_constants), {1, 1, 1}, {}, 1);
+
     ggml_vk_create_pipeline(device, device->pipeline_pool2d_f32, "pool2d_f32", pool2d_f32_len, pool2d_f32_data, "main", 2, sizeof(vk_op_pool2d_push_constants), {512, 1, 1}, {}, 1);
 
     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);
@@ -3322,6 +3472,22 @@ static vk_device ggml_vk_get_device(size_t idx) {
             }
         }
 
+
+        std::vector<vk::DescriptorSetLayoutBinding> dsl_binding;
+        std::vector<vk::DescriptorBindingFlags> dsl_binding_flags;
+        for (uint32_t i = 0; i < MAX_PARAMETER_COUNT; i++) {
+            dsl_binding.push_back({i, vk::DescriptorType::eStorageBuffer, 1, vk::ShaderStageFlagBits::eCompute});
+            dsl_binding_flags.push_back({});
+        }
+
+        vk::DescriptorSetLayoutBindingFlagsCreateInfo dslbfci = { dsl_binding_flags };
+
+        vk::DescriptorSetLayoutCreateInfo descriptor_set_layout_create_info(
+            {},
+            dsl_binding);
+        descriptor_set_layout_create_info.setPNext(&dslbfci);
+        device->dsl = device->device.createDescriptorSetLayout(descriptor_set_layout_create_info);
+
         ggml_vk_load_shaders(device);
 
         if (!device->single_queue) {
@@ -3329,7 +3495,8 @@ static vk_device ggml_vk_get_device(size_t idx) {
             ggml_vk_create_queue(device, device->transfer_queue, transfer_queue_family_index, transfer_queue_index, { vk::PipelineStageFlagBits::eTransfer }, true);
         } else {
             // TODO: Use pointer or reference to avoid copy
-            device->transfer_queue = device->compute_queue;
+            device->transfer_queue.copyFrom(device->compute_queue);
+            device->transfer_queue.cmd_pool.init(device, &device->transfer_queue);
         }
 
         device->buffer_type = {
@@ -3342,6 +3509,8 @@ static vk_device ggml_vk_get_device(size_t idx) {
 
         device->idx = idx;
 
+        device->disable_fusion = getenv("GGML_VK_DISABLE_FUSION") != nullptr;
+
         return device;
     }
 
@@ -3488,6 +3657,8 @@ static void ggml_vk_print_gpu_info(size_t idx) {
 static bool ggml_vk_instance_validation_ext_available(const std::vector<vk::ExtensionProperties>& instance_extensions);
 static bool ggml_vk_instance_portability_enumeration_ext_available(const std::vector<vk::ExtensionProperties>& instance_extensions);
 
+static bool ggml_vk_instance_debug_utils_ext_available(const std::vector<vk::ExtensionProperties> & instance_extensions);
+
 static void ggml_vk_instance_init() {
     if (vk_instance_initialized) {
         return;
@@ -3508,7 +3679,7 @@ static void ggml_vk_instance_init() {
 #ifdef __APPLE__
     const bool portability_enumeration_ext = ggml_vk_instance_portability_enumeration_ext_available(instance_extensions);
 #endif
-
+    const bool debug_utils_ext = ggml_vk_instance_debug_utils_ext_available(instance_extensions) && getenv("GGML_VK_DEBUG_MARKERS") != nullptr;
     std::vector<const char*> layers;
 
     if (validation_ext) {
@@ -3523,6 +3694,9 @@ static void ggml_vk_instance_init() {
         extensions.push_back("VK_KHR_portability_enumeration");
     }
 #endif
+    if (debug_utils_ext) {
+        extensions.push_back("VK_EXT_debug_utils");
+    }
     vk::InstanceCreateInfo instance_create_info(vk::InstanceCreateFlags{}, &app_info, layers, extensions);
 #ifdef __APPLE__
     if (portability_enumeration_ext) {
@@ -3546,13 +3720,24 @@ static void ggml_vk_instance_init() {
     vk_instance.instance = vk::createInstance(instance_create_info);
     vk_instance_initialized = true;
 
-    vk_perf_logger_enabled = getenv("GGML_VK_PERF_LOGGER") != nullptr;
+    if (debug_utils_ext) {
+        vk_instance.debug_utils_support              = true;
+        vk_instance.pfn_vkSetDebugUtilsObjectNameEXT = (PFN_vkSetDebugUtilsObjectNameEXT) vkGetInstanceProcAddr(vk_instance.instance, "vkSetDebugUtilsObjectNameEXT");
+        vk_instance.pfn_vkQueueBeginDebugUtilsLabelEXT = (PFN_vkQueueBeginDebugUtilsLabelEXT) vkGetInstanceProcAddr(vk_instance.instance, "vkQueueBeginDebugUtilsLabelEXT");
+        vk_instance.pfn_vkQueueEndDebugUtilsLabelEXT = (PFN_vkQueueEndDebugUtilsLabelEXT) vkGetInstanceProcAddr(vk_instance.instance, "vkQueueEndDebugUtilsLabelEXT");
+        vk_instance.pfn_vkCmdBeginDebugUtilsLabelEXT = (PFN_vkCmdBeginDebugUtilsLabelEXT) vkGetInstanceProcAddr(vk_instance.instance, "vkCmdBeginDebugUtilsLabelEXT");
+        vk_instance.pfn_vkCmdEndDebugUtilsLabelEXT =   (PFN_vkCmdEndDebugUtilsLabelEXT) vkGetInstanceProcAddr(vk_instance.instance, "vkCmdEndDebugUtilsLabelEXT");
+        vk_instance.pfn_vkCmdInsertDebugUtilsLabelEXT = (PFN_vkCmdInsertDebugUtilsLabelEXT) vkGetInstanceProcAddr(vk_instance.instance, "vkCmdInsertDebugUtilsLabelEXT");
 
-    size_t num_available_devices = vk_instance.instance.enumeratePhysicalDevices().size();
+    }
+
+    vk_perf_logger_enabled = getenv("GGML_VK_PERF_LOGGER") != nullptr;
 
     // Emulate behavior of CUDA_VISIBLE_DEVICES for Vulkan
     char * devices_env = getenv("GGML_VK_VISIBLE_DEVICES");
     if (devices_env != nullptr) {
+        size_t num_available_devices = vk_instance.instance.enumeratePhysicalDevices().size();
+
         std::string devices(devices_env);
         std::replace(devices.begin(), devices.end(), ',', ' ');
 
@@ -3568,9 +3753,9 @@ static void ggml_vk_instance_init() {
     } else {
         std::vector<vk::PhysicalDevice> devices = vk_instance.instance.enumeratePhysicalDevices();
 
-        // Make sure at least one device exists
+        // If no vulkan devices are found, return early
         if (devices.empty()) {
-            std::cerr << "ggml_vulkan: Error: No devices found." << std::endl;
+            GGML_LOG_INFO("ggml_vulkan: No devices found.\n");
             return;
         }
 
@@ -3653,9 +3838,20 @@ static void ggml_vk_instance_init() {
             }
         }
 
-        // If no dedicated GPUs found, fall back to GPU 0
+        // If no dedicated GPUs found, fall back to the first non-CPU device.
+        // If only CPU devices are available, return without devices.
         if (vk_instance.device_indices.empty()) {
-            vk_instance.device_indices.push_back(0);
+            for (size_t i = 0; i < devices.size(); i++) {
+                if (devices[i].getProperties().deviceType != vk::PhysicalDeviceType::eCpu) {
+                    vk_instance.device_indices.push_back(i);
+                    break;
+                }
+            }
+        }
+
+        if (vk_instance.device_indices.empty()) {
+            GGML_LOG_INFO("ggml_vulkan: No devices found.\n");
+            return;
         }
     }
     GGML_LOG_DEBUG("ggml_vulkan: Found %zu Vulkan devices:\n", vk_instance.device_indices.size());
@@ -3684,6 +3880,9 @@ static void ggml_vk_init(ggml_backend_vk_context * ctx, size_t idx) {
     ctx->fence = ctx->device->device.createFence({});
     ctx->almost_ready_fence = ctx->device->device.createFence({});
 
+    ctx->compute_cmd_pool.init(ctx->device, &ctx->device->compute_queue);
+    ctx->transfer_cmd_pool.init(ctx->device, &ctx->device->transfer_queue);
+
 #ifdef GGML_VULKAN_CHECK_RESULTS
     const char* skip_checks = getenv("GGML_VULKAN_SKIP_CHECKS");
     vk_skip_checks = (skip_checks == NULL ? 0 : atoi(skip_checks));
@@ -4004,6 +4203,7 @@ static void * ggml_vk_host_malloc(vk_device& device, size_t size) {
         return nullptr;
     }
 
+    std::lock_guard<std::recursive_mutex> guard(device->mutex);
     device->pinned_memory.push_back(std::make_tuple(buf->ptr, size, buf));
 
     return buf->ptr;
@@ -4014,6 +4214,8 @@ static void ggml_vk_host_free(vk_device& device, void* ptr) {
         return;
     }
     VK_LOG_MEMORY("ggml_vk_host_free(" << ptr << ")");
+    std::lock_guard<std::recursive_mutex> guard(device->mutex);
+
     vk_buffer buf;
     size_t index;
     for (size_t i = 0; i < device->pinned_memory.size(); i++) {
@@ -4036,6 +4238,7 @@ static void ggml_vk_host_free(vk_device& device, void* ptr) {
 }
 
 static void ggml_vk_host_get(vk_device& device, const void * ptr, vk_buffer& buf, size_t& buf_offset) {
+    std::lock_guard<std::recursive_mutex> guard(device->mutex);
     buf = nullptr;
     buf_offset = 0;
     for (size_t i = 0; i < device->pinned_memory.size(); i++) {
@@ -4049,9 +4252,9 @@ static void ggml_vk_host_get(vk_device& device, const void * ptr, vk_buffer& buf
     }
 }
 
-static vk_submission ggml_vk_begin_submission(vk_device& device, vk_queue& q, bool one_time = true) {
+static vk_submission ggml_vk_begin_submission(vk_device& device, vk_command_pool& p, bool one_time = true) {
     vk_submission s;
-    s.buffer = ggml_vk_create_cmd_buffer(device, q);
+    s.buffer = ggml_vk_create_cmd_buffer(device, p);
     if (one_time) {
         s.buffer.begin({ vk::CommandBufferUsageFlagBits::eOneTimeSubmit });
     } else {
@@ -4061,7 +4264,33 @@ 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<vk::DescriptorBufferInfo> const& descriptor_buffer_infos, size_t push_constant_size, const void* push_constants, std::array<uint32_t, 3> elements) {
+template <typename T> size_t push_constant_size(const T &t) {
+    static_assert(std::is_class<T>::value, "T must be a struct/class");
+    GGML_UNUSED(t);
+    return sizeof(T);
+}
+template <typename T> size_t push_constant_size(const std::vector<T> &t) {
+    GGML_UNUSED(t);
+    return sizeof(T) * t.size();
+}
+template <typename T, uint32_t N> size_t push_constant_size(const std::array<T, N> &t) {
+    GGML_UNUSED(t);
+    return sizeof(T) * N;
+}
+
+template <typename T> const T *push_constant_data(const T &t) {
+    static_assert(std::is_class<T>::value, "T must be a struct/class");
+    return &t;
+}
+template <typename T> const T *push_constant_data(const std::vector<T> &t) {
+    return t.data();
+}
+template <typename T, uint32_t N> const T *push_constant_data(const std::array<T, N> &t) {
+    return t.data();
+}
+
+template <typename T>
+static void ggml_vk_dispatch_pipeline(ggml_backend_vk_context* ctx, vk_context& subctx, vk_pipeline& pipeline, std::initializer_list<vk::DescriptorBufferInfo> const& descriptor_buffer_infos, const T &push_constants, std::array<uint32_t, 3> 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]);
     const uint32_t wg2 = CEIL_DIV(elements[2], pipeline->wg_denoms[2]);
@@ -4070,14 +4299,14 @@ static void ggml_vk_dispatch_pipeline(ggml_backend_vk_context* ctx, vk_context&
         std::cerr << "(" << buffer.buffer << ", " << buffer.offset << ", " << buffer.range << "), ";
     }
     std::cerr << "}, (" << wg0 << "," << wg1 << "," << wg2 << "))");
-    GGML_ASSERT(pipeline->descriptor_set_idx < pipeline->descriptor_sets.size());
-    GGML_ASSERT(descriptor_buffer_infos.size() == pipeline->parameter_count);
+    GGML_ASSERT(ctx->descriptor_set_idx < ctx->descriptor_sets.size());
+    GGML_ASSERT(descriptor_buffer_infos.size() <= MAX_PARAMETER_COUNT);
 
-    vk::DescriptorSet& descriptor_set = pipeline->descriptor_sets[pipeline->descriptor_set_idx++];
+    vk::DescriptorSet& descriptor_set = ctx->descriptor_sets[ctx->descriptor_set_idx++];
     vk::WriteDescriptorSet write_descriptor_set{ descriptor_set, 0, 0, pipeline->parameter_count, vk::DescriptorType::eStorageBuffer, nullptr, descriptor_buffer_infos.begin() };
     ctx->device->device.updateDescriptorSets({ write_descriptor_set }, {});
 
-    subctx->s->buffer.pushConstants(pipeline->layout, vk::ShaderStageFlagBits::eCompute, 0, push_constant_size, push_constants);
+    subctx->s->buffer.pushConstants(pipeline->layout, vk::ShaderStageFlagBits::eCompute, 0, push_constant_size(push_constants), push_constant_data(push_constants));
     subctx->s->buffer.bindPipeline(vk::PipelineBindPoint::eCompute, pipeline->pipeline);
     subctx->s->buffer.bindDescriptorSets(vk::PipelineBindPoint::eCompute,
                                 pipeline->layout,
@@ -4110,7 +4339,7 @@ static void ggml_vk_ctx_begin(vk_device& device, vk_context& subctx) {
         ggml_vk_ctx_end(subctx);
     }
 
-    subctx->seqs.push_back({ ggml_vk_begin_submission(device, *subctx->q) });
+    subctx->seqs.push_back({ ggml_vk_begin_submission(device, *subctx->p) });
     subctx->s = subctx->seqs[subctx->seqs.size() - 1].data();
 }
 
@@ -4311,7 +4540,9 @@ static void ggml_vk_buffer_write_2d(vk_buffer& dst, size_t offset, const void *
             memcpy((uint8_t *)dst->ptr + offset + i * width, (const uint8_t *) src + i * spitch, width);
         }
     } else {
-        vk_context subctx = ggml_vk_create_temporary_context(dst->device->transfer_queue);
+        std::lock_guard<std::recursive_mutex> guard(dst->device->mutex);
+
+        vk_context subctx = ggml_vk_create_temporary_context(dst->device->transfer_queue.cmd_pool);
         ggml_vk_ctx_begin(dst->device, subctx);
         ggml_vk_buffer_write_2d_async(subctx, dst, offset, src, spitch, width, height, true);
         ggml_vk_ctx_end(subctx);
@@ -4323,6 +4554,7 @@ static void ggml_vk_buffer_write_2d(vk_buffer& dst, size_t offset, const void *
         ggml_vk_submit(subctx, dst->device->fence);
         VK_CHECK(dst->device->device.waitForFences({ dst->device->fence }, true, UINT64_MAX), "vk_buffer_write_2d waitForFences");
         dst->device->device.resetFences({ dst->device->fence });
+        ggml_vk_queue_command_pools_cleanup(dst->device);
     }
 }
 
@@ -4399,7 +4631,9 @@ static void ggml_vk_buffer_read(vk_buffer& src, size_t offset, void * dst, size_
 
         memcpy(dst, (uint8_t *) src->ptr + offset, size);
     } else {
-        vk_context subctx = ggml_vk_create_temporary_context(src->device->transfer_queue);
+        std::lock_guard<std::recursive_mutex> guard(src->device->mutex);
+
+        vk_context subctx = ggml_vk_create_temporary_context(src->device->transfer_queue.cmd_pool);
         ggml_vk_ctx_begin(src->device, subctx);
         ggml_vk_buffer_read_async(subctx, src, offset, dst, size, true);
         ggml_vk_ctx_end(subctx);
@@ -4407,6 +4641,7 @@ static void ggml_vk_buffer_read(vk_buffer& src, size_t offset, void * dst, size_
         ggml_vk_submit(subctx, src->device->fence);
         VK_CHECK(src->device->device.waitForFences({ src->device->fence }, true, UINT64_MAX), "vk_buffer_read waitForFences");
         src->device->device.resetFences({ src->device->fence });
+        ggml_vk_queue_command_pools_cleanup(src->device);
 
         for (auto& cpy : subctx->out_memcpys) {
             memcpy(cpy.dst, cpy.src, cpy.n);
@@ -4426,15 +4661,17 @@ static void ggml_vk_buffer_copy_async(vk_context& ctx, vk_buffer& dst, size_t ds
 
 static void ggml_vk_buffer_copy(vk_buffer& dst, size_t dst_offset, vk_buffer& src, size_t src_offset, size_t size) {
     if (src->device == dst->device) {
+        std::lock_guard<std::recursive_mutex> guard(src->device->mutex);
         VK_LOG_DEBUG("ggml_vk_buffer_copy(SINGLE_DEVICE, " << size << ")");
         // Copy within the device
-        vk_context subctx = ggml_vk_create_temporary_context(src->device->transfer_queue);
+        vk_context subctx = ggml_vk_create_temporary_context(src->device->transfer_queue.cmd_pool);
         ggml_vk_ctx_begin(src->device, subctx);
         ggml_vk_buffer_copy_async(subctx, dst, dst_offset, src, src_offset, size);
         ggml_vk_ctx_end(subctx);
         ggml_vk_submit(subctx, src->device->fence);
         VK_CHECK(src->device->device.waitForFences({ src->device->fence }, true, UINT64_MAX), "vk_buffer_copy waitForFences");
         src->device->device.resetFences({ src->device->fence });
+        ggml_vk_queue_command_pools_cleanup(src->device);
     } else {
         VK_LOG_DEBUG("ggml_vk_buffer_copy(MULTI_DEVICE, " << size << ")");
         // Copy device to device
@@ -4459,7 +4696,8 @@ static void ggml_vk_buffer_memset_async(vk_context& ctx, vk_buffer& dst, size_t
 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 << ")");
 
-    vk_context subctx = ggml_vk_create_temporary_context(dst->device->transfer_queue);
+    std::lock_guard<std::recursive_mutex> guard(dst->device->mutex);
+    vk_context subctx = ggml_vk_create_temporary_context(dst->device->transfer_queue.cmd_pool);
     ggml_vk_ctx_begin(dst->device, subctx);
     subctx->s->buffer.fillBuffer(dst->buffer, offset, size, c);
     ggml_vk_ctx_end(subctx);
@@ -4467,6 +4705,7 @@ static void ggml_vk_buffer_memset(vk_buffer& dst, size_t offset, uint32_t c, siz
     ggml_vk_submit(subctx, dst->device->fence);
     VK_CHECK(dst->device->device.waitForFences({ dst->device->fence }, true, UINT64_MAX), "vk_memset waitForFences");
     dst->device->device.resetFences({ dst->device->fence });
+    ggml_vk_queue_command_pools_cleanup(dst->device);
 }
 
 static uint32_t ggml_vk_guess_split_k(ggml_backend_vk_context * ctx, int m, int n, int k, const vk_pipeline& pipeline) {
@@ -4540,7 +4779,7 @@ static void ggml_vk_matmul(
     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, padded_n };
-        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { a, b, d }, sizeof(vk_mat_mat_push_constants), &pc, { m, n, batch });
+        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { a, b, d }, pc, { m, n, batch });
         return;
     }
 
@@ -4548,10 +4787,10 @@ static void ggml_vk_matmul(
 
     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_dispatch_pipeline(ctx, subctx, pipeline, { a, b, split_k_buffer }, pc1, { (CEIL_DIV(m, pipeline->wg_denoms[0]) * pipeline->wg_denoms[0]) * split_k, n, batch });
     ggml_vk_sync_buffers(subctx);
     const std::array<uint32_t, 2> pc2 = { (uint32_t)(m * n * batch), split_k };
-    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 });
+    ggml_vk_dispatch_pipeline(ctx, subctx, ctx->device->pipeline_matmul_split_k_reduce, { split_k_buffer, d }, pc2, { m * n * batch, 1, 1 });
 }
 
 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) {
@@ -4599,7 +4838,7 @@ static void ggml_vk_matmul_id(
     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, 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 });
+    ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { a, b, d, ids }, pc, { m, nei1, n_as });
 }
 
 static bool ggml_vk_dim01_contiguous(const ggml_tensor * tensor) {
@@ -4684,9 +4923,17 @@ static vk_pipeline ggml_vk_get_cpy_pipeline(ggml_backend_vk_context * ctx, const
         // type size must be exactly 2 or 4.
         GGML_ASSERT(ggml_is_quantized(to) || ggml_type_size(src->type) == 2 || ggml_type_size(src->type) == 4);
         if ((ggml_type_size(src->type) % 4) == 0) {
-            return ctx->device->pipeline_contig_cpy_f32_f32;
+            if (contig) {
+                return ctx->device->pipeline_contig_cpy_f32_f32;
+            } else {
+                return ctx->device->pipeline_cpy_f32_f32;
+            }
         } else {
-            return ctx->device->pipeline_contig_cpy_f16_f16;
+            if (contig) {
+                return ctx->device->pipeline_contig_cpy_f16_f16;
+            } else {
+                return ctx->device->pipeline_cpy_f16_f16;
+            }
         }
     }
 
@@ -4720,7 +4967,7 @@ static void ggml_vk_cpy_to_contiguous(ggml_backend_vk_context * ctx, vk_context&
     };
     init_pushconst_fastdiv(pc);
     ggml_vk_sync_buffers(subctx);
-    ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { in, out }, sizeof(vk_op_unary_push_constants), &pc, elements);
+    ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { in, out }, pc, elements);
 }
 
 static vk_pipeline ggml_vk_get_quantize_pipeline(ggml_backend_vk_context * ctx, ggml_type type) {
@@ -4739,7 +4986,7 @@ static void ggml_vk_quantize_q8_1(ggml_backend_vk_context * ctx, vk_context& sub
     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 });
+    ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { in, out }, std::array<uint32_t, 1>{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) {
@@ -4747,7 +4994,7 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& sub
     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];
     std::cerr << "), (" << dst << ", name=" << dst->name << ", type=" << dst->type << ", ne0=" << dst->ne[0] << ", ne1=" << dst->ne[1] << ", ne2=" << dst->ne[2] << ", ne3=" << dst->ne[3] << ", nb0=" << dst->nb[0] << ", nb1=" << dst->nb[1] << ", nb2=" << dst->nb[2] << ", nb3=" << dst->nb[3];
     std::cerr << "), " << (dryrun ? "dryrun" : "") << ")");
-    GGML_ASSERT(ggml_vk_dim01_contiguous(src0) || src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16);  // NOLINT
+    GGML_ASSERT(ggml_vk_dim01_contiguous(src0) || src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || src0->type == GGML_TYPE_BF16);  // NOLINT
     GGML_ASSERT(ggml_vk_dim01_contiguous(src1) || src1->type == GGML_TYPE_F32 || src1->type == GGML_TYPE_F16);  // NOLINT
 
     const uint64_t ne00 = src0->ne[0];
@@ -4880,18 +5127,18 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& sub
         }
 
         // Request descriptor sets
-        ggml_pipeline_request_descriptor_sets(ctx->device, pipeline, 1);
+        ggml_pipeline_request_descriptor_sets(ctx, pipeline, 1);
         if (qx_needs_dequant) {
-            ggml_pipeline_request_descriptor_sets(ctx->device, to_fp16_vk_0, 1);
+            ggml_pipeline_request_descriptor_sets(ctx, to_fp16_vk_0, 1);
         }
         if (qy_needs_dequant) {
-            ggml_pipeline_request_descriptor_sets(ctx->device, to_fp16_vk_1, 1);
+            ggml_pipeline_request_descriptor_sets(ctx, to_fp16_vk_1, 1);
         }
         if (quantize_y) {
-            ggml_pipeline_request_descriptor_sets(ctx->device, to_q8_1, 1);
+            ggml_pipeline_request_descriptor_sets(ctx, to_q8_1, 1);
         }
         if (split_k > 1) {
-            ggml_pipeline_request_descriptor_sets(ctx->device, ctx->device->pipeline_matmul_split_k_reduce, 1);
+            ggml_pipeline_request_descriptor_sets(ctx, ctx->device->pipeline_matmul_split_k_reduce, 1);
         }
         return;
     }
@@ -4939,7 +5186,7 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& sub
     } else if (qx_needs_dequant) {
         const std::vector<uint32_t> pc = { (uint32_t)ne01, (uint32_t)ne10, (uint32_t)ne10, (uint32_t)ne10, (uint32_t)(ggml_nelements(src0)) };
         ggml_vk_sync_buffers(subctx);
-        ggml_vk_dispatch_pipeline(ctx, subctx, to_fp16_vk_0, { vk_subbuffer{ d_Qx, qx_buf_offset, qx_sz * ne02 * ne03 }, vk_subbuffer{ d_X, 0, x_sz * ne02 * ne03 } }, pc.size() * sizeof(uint32_t), pc.data(), { (uint32_t)(x_ne * ne02 * ne03), 1, 1});
+        ggml_vk_dispatch_pipeline(ctx, subctx, to_fp16_vk_0, { vk_subbuffer{ d_Qx, qx_buf_offset, qx_sz * ne02 * ne03 }, vk_subbuffer{ d_X, 0, x_sz * ne02 * ne03 } }, pc, { (uint32_t)(x_ne * ne02 * ne03), 1, 1});
     }
     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 });
@@ -4975,7 +5222,7 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context&
     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];
     std::cerr << "), (" << dst << ", name=" << dst->name << ", type=" << dst->type << ", ne0=" << dst->ne[0] << ", ne1=" << dst->ne[1] << ", ne2=" << dst->ne[2] << ", ne3=" << dst->ne[3] << ", nb0=" << dst->nb[0] << ", nb1=" << dst->nb[1] << ", nb2=" << dst->nb[2] << ", nb3=" << dst->nb[3];
     std::cerr << "), " << (dryrun ? "dryrun" : "") << "),)");
-    GGML_ASSERT(ggml_vk_dim01_contiguous(src0) || src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16);  // NOLINT
+    GGML_ASSERT(ggml_vk_dim01_contiguous(src0) || src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || src0->type == GGML_TYPE_BF16);  // NOLINT
     GGML_ASSERT(ggml_vk_dim01_contiguous(src1) || src1->type == GGML_TYPE_F32 || src1->type == GGML_TYPE_F16);  // NOLINT
 
     const uint64_t ne00 = src0->ne[0];
@@ -5073,12 +5320,12 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context&
 
         // Request descriptor sets
         if (qx_needs_dequant) {
-            ggml_pipeline_request_descriptor_sets(ctx->device, to_fp16_vk_0, 1);
+            ggml_pipeline_request_descriptor_sets(ctx, to_fp16_vk_0, 1);
         }
         if (qy_needs_dequant) {
-            ggml_pipeline_request_descriptor_sets(ctx->device, to_fp16_vk_1, 1);
+            ggml_pipeline_request_descriptor_sets(ctx, to_fp16_vk_1, 1);
         }
-        ggml_pipeline_request_descriptor_sets(ctx->device, dmmv, 1);
+        ggml_pipeline_request_descriptor_sets(ctx, dmmv, 1);
         return;
     }
 
@@ -5155,7 +5402,7 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context&
     ggml_vk_sync_buffers(subctx);
     ggml_vk_dispatch_pipeline(ctx, subctx, dmmv,
                               { vk_subbuffer{ d_X, x_buf_offset, x_sz * ne02 * ne03 }, vk_subbuffer{ d_Y, y_buf_offset, y_sz * ne12 * ne13 }, vk_subbuffer{ d_D, d_buf_offset, d_sz * ne22 * ne23} },
-                              sizeof(vk_mat_vec_push_constants), &pc, { groups_x, (uint32_t)(ne12 * ne13), groups_z });
+                              pc, { groups_x, (uint32_t)(ne12 * ne13), groups_z });
 }
 
 static void ggml_vk_mul_mat_vec_p021_f16_f32(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
@@ -5211,7 +5458,7 @@ static void ggml_vk_mul_mat_vec_p021_f16_f32(ggml_backend_vk_context * ctx, vk_c
 
     if (dryrun) {
         // Request descriptor sets
-        ggml_pipeline_request_descriptor_sets(ctx->device, ctx->device->pipeline_mul_mat_vec_p021_f16_f32[gqa_ratio - 1], 1);
+        ggml_pipeline_request_descriptor_sets(ctx, ctx->device->pipeline_mul_mat_vec_p021_f16_f32[gqa_ratio - 1], 1);
         return;
     }
 
@@ -5243,7 +5490,7 @@ static void ggml_vk_mul_mat_vec_p021_f16_f32(ggml_backend_vk_context * ctx, vk_c
     }
 
     ggml_vk_sync_buffers(subctx);
-    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 });
+    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 } }, 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) {
@@ -5300,7 +5547,7 @@ static void ggml_vk_mul_mat_vec_nc_f16_f32(ggml_backend_vk_context * ctx, vk_con
 
     if (dryrun) {
         // Request descriptor sets
-        ggml_pipeline_request_descriptor_sets(ctx->device, ctx->device->pipeline_mul_mat_vec_nc_f16_f32, 1);
+        ggml_pipeline_request_descriptor_sets(ctx, ctx->device->pipeline_mul_mat_vec_nc_f16_f32, 1);
         return;
     }
 
@@ -5326,7 +5573,7 @@ static void ggml_vk_mul_mat_vec_nc_f16_f32(ggml_backend_vk_context * ctx, vk_con
     const std::array<uint32_t, 9> pc = { (uint32_t)ne00, (uint32_t)ne01, row_stride_x, channel_stride_x, channel_stride_y, (uint32_t)(ne12 / ne02), (uint32_t)ne12, (uint32_t)(qy_shader_offset / ggml_type_size(src1->type)), (uint32_t)(d_shader_offset / ggml_type_size(dst->type)) };
     ggml_vk_sync_buffers(subctx);
     ggml_vk_dispatch_pipeline(ctx, subctx, ctx->device->pipeline_mul_mat_vec_nc_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 } }, 7 * sizeof(uint32_t), &pc, { 1, (uint32_t)ne01, (uint32_t)ne12 });
+        { 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 } }, pc, { 1, (uint32_t)ne01, (uint32_t)ne12 });
 }
 
 static void ggml_vk_mul_mat(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
@@ -5487,12 +5734,12 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context&
         }
 
         // Request descriptor sets
-        ggml_pipeline_request_descriptor_sets(ctx->device, pipeline, 1);
+        ggml_pipeline_request_descriptor_sets(ctx, pipeline, 1);
         if (qx_needs_dequant) {
-            ggml_pipeline_request_descriptor_sets(ctx->device, to_fp16_vk_0, 1);
+            ggml_pipeline_request_descriptor_sets(ctx, to_fp16_vk_0, 1);
         }
         if (qy_needs_dequant) {
-            ggml_pipeline_request_descriptor_sets(ctx->device, to_fp16_vk_1, 1);
+            ggml_pipeline_request_descriptor_sets(ctx, to_fp16_vk_1, 1);
         }
         return;
     }
@@ -5542,7 +5789,7 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context&
         const std::vector<uint32_t> pc = { (uint32_t)ne01, (uint32_t)ne10, (uint32_t)ne10, (uint32_t)ne10, (uint32_t)(ggml_nelements(src0)) };
         ggml_vk_sync_buffers(subctx);
         ggml_vk_dispatch_pipeline(ctx, subctx, to_fp16_vk_0,
-            { vk_subbuffer{ d_Qx, qx_buf_offset, qx_sz * ne02 * ne03 }, vk_subbuffer{ d_X, 0, x_sz * ne02 * ne03 } }, pc.size() * sizeof(uint32_t), pc.data(), { (uint32_t)(x_ne * ne02 * ne03), 1, 1});
+            { vk_subbuffer{ d_Qx, qx_buf_offset, qx_sz * ne02 * ne03 }, vk_subbuffer{ d_X, 0, x_sz * ne02 * ne03 } }, pc, { (uint32_t)(x_ne * ne02 * ne03), 1, 1});
     }
     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 });
@@ -5576,7 +5823,7 @@ static void ggml_vk_mul_mat_vec_id_q_f16(ggml_backend_vk_context * ctx, vk_conte
     std::cerr << "), (" << ids << ", name=" << ids->name << ", type=" << ids->type << ", ne0=" << ids->ne[0] << ", ne1=" << ids->ne[1] << ", ne2=" << ids->ne[2] << ", ne3=" << ids->ne[3] << ", nb0=" << ids->nb[0] << ", nb1=" << ids->nb[1] << ", nb2=" << ids->nb[2] << ", nb3=" << ids->nb[3];
     std::cerr << "), (" << dst << ", name=" << dst->name << ", type=" << dst->type << ", ne0=" << dst->ne[0] << ", ne1=" << dst->ne[1] << ", ne2=" << dst->ne[2] << ", ne3=" << dst->ne[3] << ", nb0=" << dst->nb[0] << ", nb1=" << dst->nb[1] << ", nb2=" << dst->nb[2] << ", nb3=" << dst->nb[3];
     std::cerr << "), " << (dryrun ? "dryrun" : "") << ")");
-    GGML_ASSERT(ggml_vk_dim01_contiguous(src0) || src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16);  // NOLINT
+    GGML_ASSERT(ggml_vk_dim01_contiguous(src0) || src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || src0->type == GGML_TYPE_BF16);  // NOLINT
     GGML_ASSERT(ggml_vk_dim01_contiguous(src1) || src1->type == GGML_TYPE_F32 || src1->type == GGML_TYPE_F16);  // NOLINT
     GGML_ASSERT(ids->type == GGML_TYPE_I32);
 
@@ -5681,12 +5928,12 @@ static void ggml_vk_mul_mat_vec_id_q_f16(ggml_backend_vk_context * ctx, vk_conte
 
         // Request descriptor sets
         if (qx_needs_dequant) {
-            ggml_pipeline_request_descriptor_sets(ctx->device, to_fp16_vk_0, 1);
+            ggml_pipeline_request_descriptor_sets(ctx, to_fp16_vk_0, 1);
         }
         if (qy_needs_dequant) {
-            ggml_pipeline_request_descriptor_sets(ctx->device, to_fp16_vk_1, 1);
+            ggml_pipeline_request_descriptor_sets(ctx, to_fp16_vk_1, 1);
         }
-        ggml_pipeline_request_descriptor_sets(ctx->device, dmmv, 1);
+        ggml_pipeline_request_descriptor_sets(ctx, dmmv, 1);
         return;
     }
 
@@ -5762,7 +6009,7 @@ static void ggml_vk_mul_mat_vec_id_q_f16(ggml_backend_vk_context * ctx, vk_conte
     ggml_vk_dispatch_pipeline(ctx, subctx, dmmv,
         { vk_subbuffer{ d_X, x_buf_offset, x_sz * ne02 * ne03 },
         vk_subbuffer{ d_Y, y_buf_offset, y_sz * ne12 * ne13 }, vk_subbuffer{ d_D, d_buf_offset, d_sz * ne22 * ne23}, vk_subbuffer{ d_ids, ids_buf_offset, ids_sz } },
-        sizeof(vk_mat_vec_id_push_constants), &pc, { groups_x, (uint32_t)nei0, groups_z });
+        pc, { groups_x, (uint32_t)nei0, groups_z });
 }
 
 static void ggml_vk_mul_mat_id(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) {
@@ -5770,28 +6017,74 @@ static void ggml_vk_mul_mat_id(ggml_backend_vk_context * ctx, vk_context& subctx
     if (src2->ne[1] == 1 && (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || ggml_is_quantized(src0->type))) {
         ggml_vk_mul_mat_vec_id_q_f16(ctx, subctx, src0, src1, src2, dst, dryrun);
     } else {
-        ggml_vk_mul_mat_id_q_f16(ctx, subctx, src0, src1, src2, dst, dryrun);
+        // Split based on number of ids, to fit in shared memory
+        const uint32_t nei0 = (uint32_t)src2->ne[0];
+        const uint32_t nei1 = (uint32_t)src2->ne[1];
+
+        GGML_ASSERT(nei0 <= 4096);
+        const uint32_t split_size = std::min(nei1, 4096u / nei0);
+
+        ggml_tensor src1_copy = *src1;
+        ggml_tensor src2_copy = *src2;
+        ggml_tensor dst_copy = *dst;
+
+        for (uint32_t token_start = 0; token_start < nei1; token_start += split_size) {
+            const uint32_t n_tokens = std::min(split_size, nei1 - token_start);
+
+            src1_copy.view_offs = src1->view_offs + token_start * src1_copy.nb[2];
+            src2_copy.view_offs = src2->view_offs + token_start * src2_copy.nb[1];
+            dst_copy.view_offs = dst->view_offs + token_start * dst_copy.nb[2];
+
+            src1_copy.ne[2] = n_tokens;
+            src2_copy.ne[1] = n_tokens;
+            dst_copy.ne[2] = n_tokens;
+
+            ggml_vk_mul_mat_id_q_f16(ctx, subctx, src0, &src1_copy, &src2_copy, &dst_copy, dryrun);
+        }
     }
 }
 
-static bool ggml_vk_flash_attn_coopmat_shmem_support(const vk_device& device, const uint32_t D, bool f32acc) {
+static bool ggml_vk_flash_attn_scalar_shmem_support(const vk_device& device, const uint32_t hsk, uint32_t hsv) {
     // Needs to be kept up to date on shader changes
+    GGML_UNUSED(hsv);
     const uint32_t wg_size = scalar_flash_attention_workgroup_size;
     const uint32_t Br = scalar_flash_attention_num_large_rows;
     const uint32_t Bc = scalar_flash_attention_Bc;
 
+    const uint32_t tmpsh = wg_size * sizeof(float);
+    const uint32_t tmpshv4 = wg_size * 4 * sizeof(float);
+
+    const uint32_t masksh = Bc * Br * sizeof(float);
+
+    const uint32_t Qf = Br * (hsk / 4 + 2) * 4 * sizeof(float);
+
+    const uint32_t total_size = tmpsh + tmpshv4 + masksh + Qf;
+    const bool supported = total_size <= device->properties.limits.maxComputeSharedMemorySize;
+
+    VK_LOG_DEBUG("ggml_vk_flash_attn_coopmat_shmem_support(HSK=" << hsk << ", HSV=" << hsv << ", total_size=" << total_size << ", supported=" << supported);
+
+    return supported;
+}
+
+static bool ggml_vk_flash_attn_coopmat_shmem_support(const vk_device& device, const uint32_t hsk, uint32_t hsv, bool f32acc) {
+    // Needs to be kept up to date on shader changes
+    GGML_UNUSED(hsv);
+    const uint32_t wg_size = scalar_flash_attention_workgroup_size;
+    const uint32_t Br = coopmat1_flash_attention_num_large_rows;
+    const uint32_t Bc = scalar_flash_attention_Bc;
+
     const uint32_t acctype = f32acc ? 4 : 2;
     const uint32_t f16vec4 = 8;
 
     const uint32_t tmpsh = wg_size * sizeof(float);
     const uint32_t tmpshv4 = wg_size * 4 * acctype;
 
-    const uint32_t Qf = Br * (D / 4 + 2) * f16vec4;
+    const uint32_t Qf = Br * (hsk / 4 + 2) * f16vec4;
 
-    const uint32_t sfshstride = (D <= 128) ? (Br + 8) : Br;
+    const uint32_t sfshstride = (hsk <= 128) ? (Br + 8) : Br;
     const uint32_t sfsh = Bc * sfshstride * acctype;
 
-    const uint32_t kshstride = D / 4 + 2;
+    const uint32_t kshstride = hsk / 4 + 2;
     const uint32_t ksh = Bc * kshstride * f16vec4;
 
     const uint32_t slope = Br * sizeof(float);
@@ -5799,7 +6092,7 @@ static bool ggml_vk_flash_attn_coopmat_shmem_support(const vk_device& device, co
     const uint32_t total_size = tmpsh + tmpshv4 + Qf + sfsh + ksh + slope;
     const bool supported = total_size <= device->properties.limits.maxComputeSharedMemorySize;
 
-    VK_LOG_DEBUG("ggml_vk_flash_attn_coopmat_shmem_support(D=" << D << ", f32acc=" << f32acc << ", total_size=" << total_size << ", supported=" << supported);
+    VK_LOG_DEBUG("ggml_vk_flash_attn_coopmat_shmem_support(HSK=" << hsk << ", HSV=" << hsv << ", f32acc=" << f32acc << ", total_size=" << total_size << ", supported=" << supported);
 
     return supported;
 }
@@ -5821,13 +6114,15 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
     GGML_TENSOR_LOCALS(size_t,  nb,  dst, nb)
 
     const uint32_t nem1 = mask ? mask->ne[1] : 0;
-    const uint32_t nbm1 = mask ? mask->nb[1] : 0;
+    const uint32_t nem2 = mask ? mask->ne[2] : 0;
+    const uint32_t nem3 = mask ? mask->ne[3] : 0;
 
-    const uint32_t D = neq0;
+    const uint32_t HSK = nek0;
+    const uint32_t HSV = nev0;
     uint32_t N = neq1;
     const uint32_t KV = nek1;
 
-    GGML_ASSERT(ne0 == D);
+    GGML_ASSERT(ne0 == HSV);
     GGML_ASSERT(ne2 == N);
 
     // input tensor rows must be contiguous
@@ -5835,12 +6130,9 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
     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(neq0 == HSK);
 
     GGML_ASSERT(neq1 == N);
-    GGML_ASSERT(nev0 == D);
 
     GGML_ASSERT(nev1 == nek1);
 
@@ -5861,7 +6153,7 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
         const bool coopmat_shape_supported = (dst->op_params[3] == GGML_PREC_F32 && ctx->device->coopmat_support_16x16x16_f32acc) ||
                                              (dst->op_params[3] != GGML_PREC_F32 && ctx->device->coopmat_support_16x16x16_f16acc);
 
-        const bool coopmat_shmem_supported = ggml_vk_flash_attn_coopmat_shmem_support(ctx->device, D, dst->op_params[3] == GGML_PREC_F32);
+        const bool coopmat_shmem_supported = ggml_vk_flash_attn_coopmat_shmem_support(ctx->device, HSK, HSV, dst->op_params[3] == GGML_PREC_F32);
 
         if (!coopmat_shape_supported || !coopmat_shmem_supported) {
             path = FA_SCALAR;
@@ -5891,7 +6183,7 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
     }
 
     if (N == 1 && qk_ratio > 1 && qk_ratio <= max_gqa &&
-        qk_ratio * nek2 == neq2 && nek2 == nev2 && neq3 == 1 && nek3 == 1 && nev3 == 1) {
+        qk_ratio * nek2 == neq2 && nek2 == nev2 && nem2 <= 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.
@@ -5914,47 +6206,25 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
         path = FA_SCALAR;
     }
 
+    // with large hsk/hsv, scalar path may need to use small_rows to fit in shared memory
+    if (path == FA_SCALAR &&
+        !ggml_vk_flash_attn_scalar_shmem_support(ctx->device, HSK, HSV)) {
+        small_rows = true;
+    }
+
     bool f32acc = path == FA_SCALAR || dst->op_params[3] == GGML_PREC_F32;
 
+    FaHeadSizes head_sizes = fa_get_head_sizes(k->ne[0], v->ne[0]);
+
     switch (path) {
     case FA_SCALAR:
-        switch (D) {
-        case 64: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D64[k->type][f32acc][small_rows][0]; break;
-        case 80: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D80[k->type][f32acc][small_rows][0]; break;
-        case 96: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D96[k->type][f32acc][small_rows][0]; break;
-        case 112: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D112[k->type][f32acc][small_rows][0]; break;
-        case 128: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D128[k->type][f32acc][small_rows][0]; break;
-        case 256: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D256[k->type][f32acc][small_rows][0]; break;
-        default:
-            GGML_ASSERT(!"unsupported D value");
-            return;
-        }
+        pipelines = &ctx->device->pipeline_flash_attn_f32_f16[k->type][head_sizes][f32acc][small_rows][0];
         break;
     case FA_COOPMAT1:
-        switch (D) {
-        case 64: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D64_cm1[k->type][f32acc][small_rows][0]; break;
-        case 80: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D80_cm1[k->type][f32acc][small_rows][0]; break;
-        case 96: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D96_cm1[k->type][f32acc][small_rows][0]; break;
-        case 112: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D112_cm1[k->type][f32acc][small_rows][0]; break;
-        case 128: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D128_cm1[k->type][f32acc][small_rows][0]; break;
-        case 256: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D256_cm1[k->type][f32acc][small_rows][0]; break;
-        default:
-            GGML_ASSERT(!"unsupported D value");
-            return;
-        }
+        pipelines = &ctx->device->pipeline_flash_attn_f32_f16_cm1[k->type][head_sizes][f32acc][small_rows][0];
         break;
     case FA_COOPMAT2:
-        switch (D) {
-        case 64: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D64_cm2[k->type][f32acc][small_rows][0]; break;
-        case 80: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D80_cm2[k->type][f32acc][small_rows][0]; break;
-        case 96: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D96_cm2[k->type][f32acc][small_rows][0]; break;
-        case 112: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D112_cm2[k->type][f32acc][small_rows][0]; break;
-        case 128: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D128_cm2[k->type][f32acc][small_rows][0]; break;
-        case 256: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D256_cm2[k->type][f32acc][small_rows][0]; break;
-        default:
-            GGML_ASSERT(!"unsupported D value");
-            return;
-        }
+        pipelines = &ctx->device->pipeline_flash_attn_f32_f16_cm2[k->type][head_sizes][f32acc][small_rows][0];
         break;
     default:
         GGML_ASSERT(0);
@@ -5984,7 +6254,7 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
     // Try to use split_k when KV is large enough to be worth the overhead
     if (workgroups_x == 1 && shader_core_count > 0 && KV >= 512) {
         // Try to run two workgroups per SM.
-        split_k = ctx->device->shader_core_count * 2 / workgroups_y;
+        split_k = shader_core_count * 2 / (workgroups_y * workgroups_z);
         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.
@@ -5994,9 +6264,9 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
         }
     }
 
-    // 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;
+    // Reserve space for split_k temporaries. For each split x batch, we need to store the O matrix (D x ne1)
+    // and the per-row m and L values (ne1 rows). We store all the matrices first, followed by the rows.
+    const uint64_t split_k_size = split_k > 1 ? (HSV * ne1 * sizeof(float) + ne1 * sizeof(float) * 2) * split_k * ne3 : 0;
     if (split_k_size > ctx->device->max_memory_allocation_size) {
         GGML_ABORT("Requested preallocation size is too large");
     }
@@ -6006,9 +6276,9 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
 
     if (dryrun) {
         // Request descriptor sets
-        ggml_pipeline_request_descriptor_sets(ctx->device, pipeline, 1);
+        ggml_pipeline_request_descriptor_sets(ctx, pipeline, 1);
         if (split_k > 1) {
-            ggml_pipeline_request_descriptor_sets(ctx->device, ctx->device->pipeline_flash_attn_split_k_reduce, 1);
+            ggml_pipeline_request_descriptor_sets(ctx, ctx->device->pipeline_flash_attn_split_k_reduce, 1);
         }
         return;
     }
@@ -6083,18 +6353,19 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
         }
     }
 
+    uint32_t mask_n_head_log2 = ((mask != nullptr) << 16) | n_head_log2;
+
     const vk_flash_attn_push_constants pc = { N, KV,
                                               (uint32_t)ne1, (uint32_t)ne2, (uint32_t)ne3,
                                               (uint32_t)neq2, (uint32_t)neq3,
                                               (uint32_t)nek2, (uint32_t)nek3,
                                               (uint32_t)nev2, (uint32_t)nev3,
-                                              nem1,
+                                              nem1, nem2, nem3,
                                               q_stride, (uint32_t)nbq2, (uint32_t)nbq3,
                                               k_stride, (uint32_t)nbk2, (uint32_t)nbk3,
                                               v_stride, (uint32_t)nbv2, (uint32_t)nbv3,
-                                              nbm1,
                                               scale, max_bias, logit_softcap,
-                                              mask != nullptr, n_head_log2, m0, m1,
+                                              mask_n_head_log2, m0, m1,
                                               gqa_ratio, split_kv, split_k };
 
     ggml_vk_sync_buffers(subctx);
@@ -6112,16 +6383,16 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
                                     // 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 });
+                                    pc, { workgroups_x * pipeline->wg_denoms[0], workgroups_y, workgroups_z });
 
         ggml_vk_sync_buffers(subctx);
-        const std::array<uint32_t, 3> pc2 = { D, (uint32_t)ne1, split_k };
+        const std::array<uint32_t, 4> pc2 = { HSV, (uint32_t)ne1, (uint32_t)ne3, 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 });
+                                    pc2, { (uint32_t)ne1, 1, (uint32_t)ne3 });
     } else {
         ggml_vk_dispatch_pipeline(ctx, subctx, pipeline,
                                     {
@@ -6131,7 +6402,7 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
                                         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 });
+                                    pc, { workgroups_x, workgroups_y, workgroups_z });
     }
 }
 
@@ -6262,7 +6533,7 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
         return nullptr;
     case GGML_OP_RMS_NORM:
         if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
-            return ctx->device->pipeline_rms_norm_f32;
+            return ctx->num_additional_fused_ops > 0 ? ctx->device->pipeline_rms_norm_mul_f32 : ctx->device->pipeline_rms_norm_f32;
         }
         return nullptr;
     case GGML_OP_RMS_NORM_BACK:
@@ -6287,6 +6558,8 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
                 return ctx->device->pipeline_silu[dst->type == GGML_TYPE_F16];
             case GGML_UNARY_OP_GELU:
                 return ctx->device->pipeline_gelu[dst->type == GGML_TYPE_F16];
+            case GGML_UNARY_OP_GELU_ERF:
+                return ctx->device->pipeline_gelu_erf[dst->type == GGML_TYPE_F16];
             case GGML_UNARY_OP_GELU_QUICK:
                 return ctx->device->pipeline_gelu_quick[dst->type == GGML_TYPE_F16];
             case GGML_UNARY_OP_RELU:
@@ -6299,6 +6572,28 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
                 break;
         }
         return nullptr;
+    case GGML_OP_GLU:
+        if ((src0->type != GGML_TYPE_F32 && src0->type != GGML_TYPE_F16) ||
+            (dst->type != GGML_TYPE_F32 && dst->type != GGML_TYPE_F16) ||
+            (src0->type != dst->type)) {
+            return nullptr;
+        }
+
+        switch (ggml_get_glu_op(dst)) {
+            case GGML_GLU_OP_GEGLU:
+                return ctx->device->pipeline_geglu[dst->type == GGML_TYPE_F16];
+            case GGML_GLU_OP_REGLU:
+                return ctx->device->pipeline_reglu[dst->type == GGML_TYPE_F16];
+            case GGML_GLU_OP_SWIGLU:
+                return ctx->device->pipeline_swiglu[dst->type == GGML_TYPE_F16];
+            case GGML_GLU_OP_GEGLU_ERF:
+                return ctx->device->pipeline_geglu_erf[dst->type == GGML_TYPE_F16];
+            case GGML_GLU_OP_GEGLU_QUICK:
+                return ctx->device->pipeline_geglu_quick[dst->type == GGML_TYPE_F16];
+            default:
+                break;
+        }
+        return nullptr;
     case GGML_OP_DIAG_MASK_INF:
         if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
             return ctx->device->pipeline_diag_mask_inf_f32;
@@ -6392,6 +6687,11 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
             return ctx->device->pipeline_timestep_embedding_f32;
         }
         return nullptr;
+    case GGML_OP_CONV_TRANSPOSE_1D:
+        if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
+            return ctx->device->pipeline_conv_transpose_1d_f32;
+        }
+        return nullptr;
     case GGML_OP_POOL_2D:
         if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
             return ctx->device->pipeline_pool2d_f32;
@@ -6566,7 +6866,7 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
     }
 
     if (dryrun) {
-        ggml_pipeline_request_descriptor_sets(ctx->device, pipeline, 1);
+        ggml_pipeline_request_descriptor_sets(ctx, pipeline, 1);
         return;
     }
 
@@ -6726,6 +7026,10 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
             uint32_t half_ceil = (dim + 1) / 2;
             elements = { half_ceil, (uint32_t)src0->ne[0], 1 };
         } break;
+    case GGML_OP_CONV_TRANSPOSE_1D:
+        {
+            elements = {uint32_t(src0->ne[1]), 1, 1}; // parallelize in {Cout, 1, 1}
+        } break;
     case GGML_OP_POOL_2D:
         {
             const uint32_t N = dst->ne[3];
@@ -6750,6 +7054,7 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
     case GGML_OP_CONCAT:
     case GGML_OP_UPSCALE:
     case GGML_OP_UNARY:
+    case GGML_OP_GLU:
     case GGML_OP_CONV_2D_DW:
         {
             uint32_t ne = ggml_nelements(dst);
@@ -6790,7 +7095,7 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
         }
     }
 
-    if (op == GGML_OP_SOFT_MAX) {
+    if (op == GGML_OP_SOFT_MAX || op == GGML_OP_GLU) {
         // Empty src1 is possible in soft_max, but the shader needs a buffer
         vk_subbuffer subbuf_y;
         if (use_src1) {
@@ -6800,7 +7105,7 @@ 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);
+        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 } }, pc, elements);
     } 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;
@@ -6811,26 +7116,26 @@ 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 }, vk_subbuffer{ d_Y, y_buf_offset, y_sz }, subbuf_z, vk_subbuffer{ d_D, d_buf_offset, d_sz } }, sizeof(PC), &pc, elements);
+        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 }, subbuf_z, vk_subbuffer{ d_D, d_buf_offset, d_sz } }, pc, elements);
     } else if (op == GGML_OP_IM2COL) {
         // 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);
+        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 } }, 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);
+        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 } }, 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);
+        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 } }, pc, elements);
     } else if (use_src1) {
         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);
+        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 } }, pc, elements);
     } else {
         ggml_vk_sync_buffers(subctx);
-        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_X, x_buf_offset, x_sz }, vk_subbuffer{ d_D, d_buf_offset, d_sz } }, sizeof(PC), &pc, elements);
+        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_X, x_buf_offset, x_sz }, vk_subbuffer{ d_D, d_buf_offset, d_sz } }, pc, elements);
     }
 }
 
@@ -6943,7 +7248,7 @@ static void ggml_vk_op_f32_wkv(ggml_backend_vk_context * ctx, vk_context& subctx
     GGML_ASSERT(pipeline != nullptr);
 
     if (dryrun) {
-        ggml_pipeline_request_descriptor_sets(ctx->device, pipeline, 1);
+        ggml_pipeline_request_descriptor_sets(ctx, pipeline, 1);
         return;
     }
 
@@ -6999,7 +7304,7 @@ static void ggml_vk_op_f32_wkv(ggml_backend_vk_context * ctx, vk_context& subctx
             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);
+        }, pc, elements);
     } else if (version == 7) {
         ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, {
             vk_subbuffer{ d_srcs[0], src_offsets[0], src_sizes[0] },
@@ -7010,7 +7315,7 @@ static void ggml_vk_op_f32_wkv(ggml_backend_vk_context * ctx, vk_context& subctx
             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);
+        }, pc, elements);
     } else {
         // shouldn't happen
         GGML_ASSERT(false);
@@ -7082,7 +7387,7 @@ static void ggml_vk_op_f32_opt_step_adamw(ggml_backend_vk_context * ctx, vk_cont
     GGML_ASSERT(pipeline != nullptr);
 
     if (dryrun) {
-        ggml_pipeline_request_descriptor_sets(ctx->device, pipeline, 1);
+        ggml_pipeline_request_descriptor_sets(ctx, pipeline, 1);
         return;
     }
 
@@ -7147,7 +7452,7 @@ static void ggml_vk_op_f32_opt_step_adamw(ggml_backend_vk_context * ctx, vk_cont
         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);
+    }, pc, elements);
 }
 
 static void ggml_vk_opt_step_adamw(ggml_backend_vk_context * ctx, vk_context& subctx, ggml_tensor * dst, bool dryrun = false) {
@@ -7353,18 +7658,18 @@ static void ggml_vk_group_norm(ggml_backend_vk_context * ctx, vk_context& subctx
     ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_GROUP_NORM, { group_size, 0, eps, 0.0f }, dryrun);
 }
 
-static void ggml_vk_rms_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;
+static void ggml_vk_rms_norm(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, float * op_params, 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<vk_op_unary_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_RMS_NORM, {
+    ggml_vk_op_f32<vk_op_binary_push_constants>(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_RMS_NORM, {
         (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) 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,
+        (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,
-        op_params[0], 0.0f,
-        0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+        op_params[0], 0.0f, 0,
     }, dryrun);
 }
 
@@ -7382,6 +7687,25 @@ static void ggml_vk_unary(ggml_backend_vk_context * ctx, vk_context& subctx, con
     ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_UNARY, { (uint32_t)ggml_nelements(src0), 0, 0.0f, 0.0f }, dryrun);
 }
 
+static void ggml_vk_glu(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
+    const bool swapped = (bool)dst->op_params[1];
+    const bool split = src1 != nullptr;
+
+    GGML_ASSERT(ggml_is_contiguous(src0));
+
+    if (!split) {
+        GGML_ASSERT(src0->ne[0] / 2 == dst->ne[0]);
+    } else {
+        GGML_ASSERT(src0->ne[0] == src1->ne[0]);
+        GGML_ASSERT(src0->ne[0] == dst->ne[0]);
+        GGML_ASSERT(src0->type == src1->type);
+    }
+
+    const uint32_t mode = split ? 2 : (swapped ? 1 : 0);
+
+    ggml_vk_op_f32<vk_op_glu_push_constants>(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_GLU, { (uint32_t)ggml_nelements(dst), (uint32_t)src0->ne[0], (uint32_t)dst->ne[0], mode }, dryrun);
+}
+
 static void ggml_vk_diag_mask_inf(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
     int32_t * op_params = (int32_t *)dst->op_params;
     ggml_vk_op_f32<vk_op_diag_mask_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_DIAG_MASK_INF, { (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], op_params[0] }, dryrun);
@@ -7397,7 +7721,13 @@ static void ggml_vk_soft_max(ggml_backend_vk_context * ctx, vk_context& subctx,
     const uint32_t nrows_x = (uint32_t)ggml_nrows(src0);
     const uint32_t nrows_y = (uint32_t)src0->ne[1];
 
-    const uint32_t n_head_kv   = nrows_x/nrows_y;
+    const uint32_t ne12 = src1 ? (uint32_t)(src1->ne[2]) : 0u;
+    const uint32_t ne13 = src1 ? (uint32_t)(src1->ne[3]) : 0u;
+    const uint32_t nb11 = src1 ? (uint32_t)(src1->nb[1] / src1->nb[0]) : 0u;
+    const uint32_t nb12 = src1 ? (uint32_t)(src1->nb[2] / src1->nb[0]) : 0u;
+    const uint32_t nb13 = src1 ? (uint32_t)(src1->nb[3] / src1->nb[0]) : 0u;
+
+    const uint32_t n_head_kv   = src0->ne[2];
     const uint32_t n_head_log2 = 1u << (uint32_t) floorf(log2f((float) n_head_kv));
 
     const float m0 = powf(2.0f, -(max_bias       ) / n_head_log2);
@@ -7406,6 +7736,9 @@ static void ggml_vk_soft_max(ggml_backend_vk_context * ctx, vk_context& subctx,
     ggml_vk_op_f32<vk_op_soft_max_push_constants>(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_SOFT_MAX, {
         ncols,
         src1 != nullptr ? nrows_y : (uint32_t)0,
+        (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2],
+        ne12, ne13,
+        nb11, nb12, nb13,
         scale, max_bias,
         m0, m1,
         n_head_log2,
@@ -7529,6 +7862,37 @@ static void ggml_vk_timestep_embedding(ggml_backend_vk_context * ctx, vk_context
     }, dryrun);
 }
 
+static void ggml_vk_conv_transpose_1d(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
+    // src0: (K, Cout, Cin, 1) -- kernel
+    // src1: (L, Cin, 1, 1) -- input
+    // dst: (*, Cout, 1, 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
+
+    GGML_ASSERT(nb00 == sizeof(float));
+    GGML_ASSERT(nb10 == sizeof(float));
+
+    const int32_t s0 = dst->op_params[0];
+
+    vk_op_conv_transpose_1d_push_constants p{};
+    p.Cout = static_cast<uint32_t>(ne01);
+    p.Cin = static_cast<uint32_t>(ne02);
+    p.K = static_cast<uint32_t>(ne00);
+    p.L = static_cast<uint32_t>(ne10);
+    p.KL = static_cast<uint32_t>(ne0);
+    p.nb01 = static_cast<uint32_t>(nb01 / nb00);
+    p.nb02 = static_cast<uint32_t>(nb02 / nb00);
+    p.nb11 = static_cast<uint32_t>(nb11 / nb10);
+    p.nb1 = static_cast<uint32_t>(nb1 / nb0);
+    p.s0 = static_cast<uint32_t>(s0);
+
+    ggml_vk_op_f32(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_CONV_TRANSPOSE_1D, std::move(p), dryrun);
+}
+
 static void ggml_vk_pool_2d(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
     uint32_t op = static_cast<uint32_t>(dst->op_params[0]);
     const int32_t k1 = dst->op_params[1];
@@ -7729,9 +8093,9 @@ static void ggml_vk_test_matmul(ggml_backend_vk_context * ctx, size_t m, size_t
         }
     }
 
-    ggml_pipeline_request_descriptor_sets(ctx->device, p, num_it);
+    ggml_pipeline_request_descriptor_sets(ctx, p, num_it);
     if (split_k > 1) {
-        ggml_pipeline_request_descriptor_sets(ctx->device, ctx->device->pipeline_matmul_split_k_reduce, num_it);
+        ggml_pipeline_request_descriptor_sets(ctx, ctx->device->pipeline_matmul_split_k_reduce, num_it);
 
         if (ctx->prealloc_split_k == nullptr || ctx->prealloc_split_k->size < sizeof(float) * d_ne * split_k) {
             // Resize buffer
@@ -7746,7 +8110,7 @@ static void ggml_vk_test_matmul(ggml_backend_vk_context * ctx, size_t m, size_t
         ggml_vk_load_shaders(ctx->device);
     }
 
-    ggml_pipeline_allocate_descriptor_sets(ctx->device);
+    ggml_pipeline_allocate_descriptor_sets(ctx);
 
     vk_buffer d_X = ggml_vk_create_buffer_check(ctx->device, sizeof(X_TYPE) * x_ne, vk::MemoryPropertyFlagBits::eDeviceLocal);
     vk_buffer d_Y = ggml_vk_create_buffer_check(ctx->device, sizeof(Y_TYPE) * y_ne, vk::MemoryPropertyFlagBits::eDeviceLocal);
@@ -7788,7 +8152,7 @@ static void ggml_vk_test_matmul(ggml_backend_vk_context * ctx, size_t m, size_t
     ggml_vk_buffer_write(d_X, 0, x, sizeof(X_TYPE) * k * m * batch);
     ggml_vk_buffer_write(d_Y, 0, y, sizeof(Y_TYPE) * k * n * batch);
 
-    vk_context subctx = ggml_vk_create_context(ctx, ctx->device->compute_queue);
+    vk_context subctx = ggml_vk_create_context(ctx, ctx->compute_cmd_pool);
     ggml_vk_ctx_begin(ctx->device, subctx);
     for (size_t i = 0; i < num_it; i++) {
         ggml_vk_matmul(
@@ -7804,6 +8168,7 @@ static void ggml_vk_test_matmul(ggml_backend_vk_context * ctx, size_t m, size_t
     ggml_vk_submit(subctx, ctx->fence);
     VK_CHECK(ctx->device->device.waitForFences({ ctx->fence }, true, UINT64_MAX), "ggml_vk_test_matmul waitForFences");
     ctx->device->device.resetFences({ ctx->fence });
+    ggml_vk_queue_command_pools_cleanup(ctx->device);
 
     auto end = std::chrono::high_resolution_clock::now();
     double time = std::chrono::duration_cast<std::chrono::microseconds>(end-begin).count() / 1000.0;
@@ -7905,16 +8270,13 @@ static void ggml_vk_test_matmul(ggml_backend_vk_context * ctx, size_t m, size_t
 
     free(d_chk);
 
-    ggml_vk_queue_cleanup(ctx->device, ctx->device->transfer_queue);
-    ggml_vk_queue_cleanup(ctx->device, ctx->device->compute_queue);
+    ggml_vk_command_pool_cleanup(ctx->device, ctx->compute_cmd_pool);
+    ggml_vk_command_pool_cleanup(ctx->device, ctx->transfer_cmd_pool);
 
     ggml_vk_destroy_buffer(d_X);
     ggml_vk_destroy_buffer(d_Y);
     ggml_vk_destroy_buffer(d_D);
 
-    ggml_pipeline_cleanup(p);
-    ggml_pipeline_cleanup(ctx->device->pipeline_matmul_split_k_reduce);
-
     free(x);
     free(y);
     free(d);
@@ -7992,20 +8354,20 @@ static void ggml_vk_test_dequant(ggml_backend_vk_context * ctx, size_t ne, ggml_
     ggml_vk_quantize_data(x, qx, ne, quant);
     ggml_vk_dequantize_data(qx, x_ref, ne, quant);
 
-    ggml_pipeline_request_descriptor_sets(ctx->device, p, 1);
+    ggml_pipeline_request_descriptor_sets(ctx, p, 1);
 
     if (ctx->device->need_compiles) {
         ggml_vk_load_shaders(ctx->device);
     }
 
-    ggml_pipeline_allocate_descriptor_sets(ctx->device);
+    ggml_pipeline_allocate_descriptor_sets(ctx);
 
     ggml_vk_buffer_write(qx_buf, 0, qx, qx_sz);
 
-    vk_context subctx = ggml_vk_create_context(ctx, ctx->device->compute_queue);
+    vk_context subctx = ggml_vk_create_context(ctx, ctx->compute_cmd_pool);
     ggml_vk_ctx_begin(ctx->device, subctx);
     const std::vector<uint32_t> pc = { 1, (uint32_t)ne, (uint32_t)ne, (uint32_t)ne, (uint32_t)ne };
-    ggml_vk_dispatch_pipeline(ctx, subctx, p, { vk_subbuffer{ qx_buf, 0, qx_sz }, vk_subbuffer{ x_buf, 0, x_sz_f16 } }, pc.size() * sizeof(int), pc.data(), { (uint32_t)ne, 1, 1});
+    ggml_vk_dispatch_pipeline(ctx, subctx, p, { vk_subbuffer{ qx_buf, 0, qx_sz }, vk_subbuffer{ x_buf, 0, x_sz_f16 } }, pc, { (uint32_t)ne, 1, 1});
     ggml_vk_ctx_end(subctx);
 
     auto begin = std::chrono::high_resolution_clock::now();
@@ -8013,6 +8375,7 @@ static void ggml_vk_test_dequant(ggml_backend_vk_context * ctx, size_t ne, ggml_
     ggml_vk_submit(subctx, ctx->fence);
     VK_CHECK(ctx->device->device.waitForFences({ ctx->fence }, true, UINT64_MAX), "ggml_vk_test_dequant waitForFences");
     ctx->device->device.resetFences({ ctx->fence });
+    ggml_vk_queue_command_pools_cleanup(ctx->device);
 
     auto end = std::chrono::high_resolution_clock::now();
 
@@ -8092,17 +8455,17 @@ static void ggml_vk_test_dequant(ggml_backend_vk_context * ctx, size_t ne, ggml_
 //
 //     vk_pipeline p = ggml_vk_get_quantize_pipeline(ctx, quant);
 //
-//     ggml_pipeline_request_descriptor_sets(ctx->device, p, 1);
+//     ggml_pipeline_request_descriptor_sets(ctx, p, 1);
 //
 //     if (ctx->device->need_compiles) {
 //         ggml_vk_load_shaders(ctx->device);
 //     }
 //
-//     ggml_pipeline_allocate_descriptor_sets(ctx->device);
+//     ggml_pipeline_allocate_descriptor_sets(ctx);
 //
 //     ggml_vk_buffer_write(x_buf, 0, x, x_sz);
 //
-//     vk_context subctx = ggml_vk_create_context(ctx, ctx->device->compute_queue);
+//     vk_context subctx = ggml_vk_create_context(ctx, ctx->compute_cmd_pool);
 //     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);
@@ -8112,6 +8475,7 @@ static void ggml_vk_test_dequant(ggml_backend_vk_context * ctx, size_t ne, ggml_
 //     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 });
+//     ggml_vk_queue_command_pools_cleanup(ctx->device);
 //
 //     auto end = std::chrono::high_resolution_clock::now();
 //
@@ -8251,9 +8615,9 @@ static void ggml_vk_test_dequant_matmul(ggml_backend_vk_context * ctx, size_t m,
         // y[i] = i % k;
     }
 
-    ggml_pipeline_request_descriptor_sets(ctx->device, p, num_it);
+    ggml_pipeline_request_descriptor_sets(ctx, p, num_it);
     if (split_k > 1) {
-        ggml_pipeline_request_descriptor_sets(ctx->device, ctx->device->pipeline_matmul_split_k_reduce, num_it);
+        ggml_pipeline_request_descriptor_sets(ctx, ctx->device->pipeline_matmul_split_k_reduce, num_it);
 
         if (ctx->prealloc_split_k == nullptr || ctx->prealloc_split_k->size < sizeof(float) * d_ne * split_k) {
             // Resize buffer
@@ -8264,19 +8628,19 @@ static void ggml_vk_test_dequant_matmul(ggml_backend_vk_context * ctx, size_t m,
         }
     }
     if (mmq) {
-        ggml_pipeline_request_descriptor_sets(ctx->device, ctx->device->pipeline_quantize_q8_1, num_it);
+        ggml_pipeline_request_descriptor_sets(ctx, 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);
+    ggml_pipeline_allocate_descriptor_sets(ctx);
 
     ggml_vk_buffer_write(qx_buf, 0, qx, qx_sz);
     ggml_vk_buffer_write(y_buf, 0, y, y_sz);
 
-    vk_context subctx = ggml_vk_create_context(ctx, ctx->device->compute_queue);
+    vk_context subctx = ggml_vk_create_context(ctx, ctx->compute_cmd_pool);
     ggml_vk_ctx_begin(ctx->device, subctx);
     if (mmq) {
         for (size_t i = 0; i < num_it; i++) {
@@ -8305,6 +8669,7 @@ static void ggml_vk_test_dequant_matmul(ggml_backend_vk_context * ctx, size_t m,
     ggml_vk_submit(subctx, ctx->fence);
     VK_CHECK(ctx->device->device.waitForFences({ ctx->fence }, true, UINT64_MAX), "ggml_vk_test_dequant waitForFences");
     ctx->device->device.resetFences({ ctx->fence });
+    ggml_vk_queue_command_pools_cleanup(ctx->device);
 
     auto end = std::chrono::high_resolution_clock::now();
 
@@ -8523,11 +8888,12 @@ 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, bool almost_ready);
+static bool ggml_vk_compute_forward(ggml_backend_vk_context* ctx, ggml_cgraph * cgraph, 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 almost_ready, bool submit){
+static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgraph, int node_idx, ggml_tensor *node_begin, int node_idx_begin, bool dryrun, bool last_node, bool almost_ready, bool submit){
+    ggml_tensor * node = cgraph->nodes[node_idx];
     if (ggml_is_empty(node) || !node->buffer) {
         return false;
     }
@@ -8552,6 +8918,7 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod
         switch (ggml_get_unary_op(node)) {
         case GGML_UNARY_OP_SILU:
         case GGML_UNARY_OP_GELU:
+        case GGML_UNARY_OP_GELU_ERF:
         case GGML_UNARY_OP_GELU_QUICK:
         case GGML_UNARY_OP_RELU:
         case GGML_UNARY_OP_TANH:
@@ -8561,6 +8928,18 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod
             return false;
         }
         break;
+    case GGML_OP_GLU:
+        switch (ggml_get_glu_op(node)) {
+        case GGML_GLU_OP_GEGLU:
+        case GGML_GLU_OP_REGLU:
+        case GGML_GLU_OP_SWIGLU:
+        case GGML_GLU_OP_GEGLU_ERF:
+        case GGML_GLU_OP_GEGLU_QUICK:
+            break;
+        default:
+            return false;
+        }
+        break;
     case GGML_OP_REPEAT:
     case GGML_OP_REPEAT_BACK:
     case GGML_OP_GET_ROWS:
@@ -8600,6 +8979,7 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod
     case GGML_OP_COUNT_EQUAL:
     case GGML_OP_IM2COL:
     case GGML_OP_TIMESTEP_EMBEDDING:
+    case GGML_OP_CONV_TRANSPOSE_1D:
     case GGML_OP_POOL_2D:
     case GGML_OP_CONV_2D_DW:
     case GGML_OP_RWKV_WKV6:
@@ -8618,7 +8998,7 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod
 
     if (!dryrun) {
         if (ctx->compute_ctx.expired()) {
-            compute_ctx = ggml_vk_create_context(ctx, ctx->device->compute_queue);
+            compute_ctx = ggml_vk_create_context(ctx, ctx->compute_cmd_pool);
             ctx->compute_ctx = compute_ctx;
             ggml_vk_ctx_begin(ctx->device, compute_ctx);
         } else {
@@ -8652,6 +9032,7 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod
         case GGML_OP_RMS_NORM_BACK:
         case GGML_OP_L2_NORM:
         case GGML_OP_UNARY:
+        case GGML_OP_GLU:
         case GGML_OP_DIAG_MASK_INF:
         case GGML_OP_SOFT_MAX:
         case GGML_OP_SOFT_MAX_BACK:
@@ -8664,6 +9045,7 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod
         case GGML_OP_COUNT_EQUAL:
         case GGML_OP_IM2COL:
         case GGML_OP_TIMESTEP_EMBEDDING:
+        case GGML_OP_CONV_TRANSPOSE_1D:
         case GGML_OP_POOL_2D:
         case GGML_OP_CONV_2D_DW:
         case GGML_OP_LEAKY_RELU:
@@ -8671,7 +9053,7 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod
                 // These operations all go through ggml_vk_op_f32, so short-circuit and
                 // do the only thing needed for the dryrun.
                 vk_pipeline pipeline = ggml_vk_op_get_pipeline(ctx, src0, src1, src2, node, node->op);
-                ggml_pipeline_request_descriptor_sets(ctx->device, pipeline, 1);
+                ggml_pipeline_request_descriptor_sets(ctx, pipeline, 1);
                 return false;
             }
         default:
@@ -8763,8 +9145,14 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod
 
         break;
     case GGML_OP_RMS_NORM:
-        ggml_vk_rms_norm(ctx, compute_ctx, src0, node, dryrun);
-
+        if (ctx->num_additional_fused_ops > 0) {
+            // fused rms_norm + mul
+            ggml_tensor *mul = cgraph->nodes[node_idx + 1];
+            ggml_tensor *other_src = mul->src[0] == node ? mul->src[1] : mul->src[0];
+            ggml_vk_rms_norm(ctx, compute_ctx, src0, other_src, mul, (float *)node->op_params, dryrun);
+        } else {
+            ggml_vk_rms_norm(ctx, compute_ctx, src0, src0, node, (float *)node->op_params, dryrun);
+        }
         break;
     case GGML_OP_RMS_NORM_BACK:
         ggml_vk_rms_norm_back(ctx, compute_ctx, src0, src1, node, dryrun);
@@ -8778,6 +9166,7 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod
         switch (ggml_get_unary_op(node)) {
         case GGML_UNARY_OP_SILU:
         case GGML_UNARY_OP_GELU:
+        case GGML_UNARY_OP_GELU_ERF:
         case GGML_UNARY_OP_GELU_QUICK:
         case GGML_UNARY_OP_RELU:
         case GGML_UNARY_OP_TANH:
@@ -8788,6 +9177,19 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod
             return false;
         }
         break;
+    case GGML_OP_GLU:
+        switch (ggml_get_glu_op(node)) {
+        case GGML_GLU_OP_GEGLU:
+        case GGML_GLU_OP_REGLU:
+        case GGML_GLU_OP_SWIGLU:
+        case GGML_GLU_OP_GEGLU_ERF:
+        case GGML_GLU_OP_GEGLU_QUICK:
+            ggml_vk_glu(ctx, compute_ctx, src0, src1, node, dryrun);
+            break;
+        default:
+            return false;
+        }
+        break;
     case GGML_OP_DIAG_MASK_INF:
         ggml_vk_diag_mask_inf(ctx, compute_ctx, src0, node, dryrun);
 
@@ -8835,6 +9237,10 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod
     case GGML_OP_TIMESTEP_EMBEDDING:
         ggml_vk_timestep_embedding(ctx, compute_ctx, src0, node, dryrun);
 
+        break;
+    case GGML_OP_CONV_TRANSPOSE_1D:
+        ggml_vk_conv_transpose_1d(ctx, compute_ctx, src0, src1, node, dryrun);
+
         break;
     case GGML_OP_POOL_2D:
         ggml_vk_pool_2d(ctx, compute_ctx, src0, node, dryrun);
@@ -8905,12 +9311,13 @@ 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, almost_ready);
+        bool ok = ggml_vk_compute_forward(ctx, cgraph, 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<ggml_unary_op>(node->op_params[0])) << ")" << std::endl;
-            }
-            else {
+            } else if (node->op == GGML_OP_GLU) {
+                std::cerr << __func__ << ": error: op not supported GLU " << node->name << " (" << ggml_glu_op_name(static_cast<ggml_glu_op>(node->op_params[0])) << ")" << std::endl;
+            } else {
                 std::cerr << __func__ << ": error: op not supported " << node->name << " (" << ggml_op_name(node->op) << ")" << std::endl;
             }
         }
@@ -8919,7 +9326,8 @@ 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, bool almost_ready = false) {
+static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_cgraph * cgraph, ggml_tensor * tensor, int tensor_idx, bool use_fence = true, bool almost_ready = false) {
+    GGML_UNUSED(cgraph);
     ggml_backend_buffer * buf = nullptr;
 
     switch (tensor->op) {
@@ -8963,6 +9371,7 @@ static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_tensor *
     case GGML_OP_COUNT_EQUAL:
     case GGML_OP_IM2COL:
     case GGML_OP_TIMESTEP_EMBEDDING:
+    case GGML_OP_CONV_TRANSPOSE_1D:
     case GGML_OP_POOL_2D:
     case GGML_OP_CONV_2D_DW:
     case GGML_OP_RWKV_WKV6:
@@ -8978,6 +9387,7 @@ static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_tensor *
         switch (ggml_get_unary_op(tensor)) {
         case GGML_UNARY_OP_SILU:
         case GGML_UNARY_OP_GELU:
+        case GGML_UNARY_OP_GELU_ERF:
         case GGML_UNARY_OP_GELU_QUICK:
         case GGML_UNARY_OP_RELU:
         case GGML_UNARY_OP_TANH:
@@ -8988,6 +9398,19 @@ static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_tensor *
             return false;
         }
         break;
+    case GGML_OP_GLU:
+        switch (ggml_get_glu_op(tensor)) {
+        case GGML_GLU_OP_GEGLU:
+        case GGML_GLU_OP_REGLU:
+        case GGML_GLU_OP_SWIGLU:
+        case GGML_GLU_OP_GEGLU_ERF:
+        case GGML_GLU_OP_GEGLU_QUICK:
+            buf = tensor->buffer;
+            break;
+        default:
+            return false;
+        }
+        break;
     case GGML_OP_MUL_MAT:
     case GGML_OP_MUL_MAT_ID:
     case GGML_OP_FLASH_ATTN_EXT:
@@ -9014,7 +9437,7 @@ static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_tensor *
     // Only run if ctx hasn't been submitted yet
     if (!subctx->seqs.empty()) {
 #ifdef GGML_VULKAN_CHECK_RESULTS
-        ggml_vk_check_results_0(tensor);
+        ggml_vk_check_results_0(ctx, cgraph, tensor_idx);
         use_fence = true;
 #endif
 
@@ -9034,7 +9457,7 @@ static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_tensor *
             ggml_vk_wait_for_fence(ctx);
         }
 #ifdef GGML_VULKAN_CHECK_RESULTS
-        ggml_vk_check_results_1(tensor);
+        ggml_vk_check_results_1(ctx, cgraph, tensor_idx);
 #endif
     }
 
@@ -9058,19 +9481,8 @@ static void ggml_vk_graph_cleanup(ggml_backend_vk_context * ctx) {
     }
     ctx->gc.temp_buffers.clear();
 
-    for (auto& dsr : ctx->device->pipeline_descriptor_set_requirements) {
-        vk_pipeline_ref plr = ctx->device->pipelines[dsr.first];
-
-        if (plr.expired()) {
-            continue;
-        }
-
-        vk_pipeline pl = plr.lock();
-        ggml_pipeline_cleanup(pl);
-    }
-
-    ggml_vk_queue_cleanup(ctx->device, ctx->device->compute_queue);
-    ggml_vk_queue_cleanup(ctx->device, ctx->device->transfer_queue);
+    ggml_vk_command_pool_cleanup(ctx->device, ctx->compute_cmd_pool);
+    ggml_vk_command_pool_cleanup(ctx->device, ctx->transfer_cmd_pool);
 
     for (size_t i = 0; i < ctx->gc.semaphores.size(); i++) {
         ctx->device->device.destroySemaphore({ ctx->gc.semaphores[i].s });
@@ -9091,7 +9503,8 @@ static void ggml_vk_graph_cleanup(ggml_backend_vk_context * ctx) {
 
     ctx->tensor_ctxs.clear();
     ctx->gc.contexts.clear();
-    ctx->device->pipeline_descriptor_set_requirements.clear();
+    ctx->pipeline_descriptor_set_requirements = 0;
+    ctx->descriptor_set_idx = 0;
 }
 
 // Clean up on backend free
@@ -9118,6 +9531,15 @@ static void ggml_vk_cleanup(ggml_backend_vk_context * ctx) {
 
     ctx->device->device.destroyFence(ctx->fence);
     ctx->device->device.destroyFence(ctx->almost_ready_fence);
+
+    for (auto& pool : ctx->descriptor_pools) {
+        ctx->device->device.destroyDescriptorPool(pool);
+    }
+    ctx->descriptor_pools.clear();
+    ctx->descriptor_sets.clear();
+
+    ctx->compute_cmd_pool.destroy(ctx->device->device);
+    ctx->transfer_cmd_pool.destroy(ctx->device->device);
 }
 
 static int ggml_vk_get_device_count() {
@@ -9325,6 +9747,12 @@ static size_t ggml_backend_vk_host_buffer_type_get_alignment(ggml_backend_buffer
     UNUSED(buft);
 }
 
+static size_t ggml_backend_vk_host_buffer_type_get_max_size(ggml_backend_buffer_type_t buft) {
+    return vk_instance.devices[0]->suballocation_block_size;
+
+    UNUSED(buft);
+}
+
 // Should be changed to return device-specific host buffer type
 // but that probably requires changes in llama.cpp
 ggml_backend_buffer_type_t ggml_backend_vk_host_buffer_type() {
@@ -9333,7 +9761,7 @@ ggml_backend_buffer_type_t ggml_backend_vk_host_buffer_type() {
             /* .get_name         = */ ggml_backend_vk_host_buffer_type_name,
             /* .alloc_buffer     = */ ggml_backend_vk_host_buffer_type_alloc_buffer,
             /* .get_alignment    = */ ggml_backend_vk_host_buffer_type_get_alignment,
-            /* .get_max_size     = */ NULL, // defaults to SIZE_MAX
+            /* .get_max_size     = */ ggml_backend_vk_host_buffer_type_get_max_size,
             /* .get_alloc_size   = */ ggml_backend_cpu_buffer_type()->iface.get_alloc_size,
             /* .is_host          = */ ggml_backend_cpu_buffer_type()->iface.is_host,
         },
@@ -9384,7 +9812,7 @@ static void ggml_backend_vk_set_tensor_async(ggml_backend_t backend, ggml_tensor
 
     if (ctx->transfer_ctx.expired()) {
         // Initialize new transfer context
-        transfer_ctx = ggml_vk_create_context(ctx, ctx->device->transfer_queue);
+        transfer_ctx = ggml_vk_create_context(ctx, ctx->transfer_cmd_pool);
         ctx->transfer_ctx = transfer_ctx;
         ggml_vk_ctx_begin(ctx->device, transfer_ctx);
     } else {
@@ -9407,7 +9835,7 @@ static void ggml_backend_vk_get_tensor_async(ggml_backend_t backend, const ggml_
 
     if (ctx->transfer_ctx.expired()) {
         // Initialize new transfer context
-        transfer_ctx = ggml_vk_create_context(ctx, ctx->device->transfer_queue);
+        transfer_ctx = ggml_vk_create_context(ctx, ctx->transfer_cmd_pool);
         ctx->transfer_ctx = transfer_ctx;
         ggml_vk_ctx_begin(ctx->device, transfer_ctx);
     } else {
@@ -9430,7 +9858,7 @@ static bool ggml_backend_vk_cpy_tensor_async(ggml_backend_t backend, const ggml_
 
         if (ctx->transfer_ctx.expired()) {
             // Initialize new transfer context
-            transfer_ctx = ggml_vk_create_context(ctx, ctx->device->transfer_queue);
+            transfer_ctx = ggml_vk_create_context(ctx, ctx->transfer_cmd_pool);
             ctx->transfer_ctx = transfer_ctx;
             ggml_vk_ctx_begin(ctx->device, transfer_ctx);
         } else {
@@ -9476,22 +9904,65 @@ static bool ggml_vk_is_empty(ggml_tensor * node) {
     return ggml_is_empty(node) || node->op == GGML_OP_NONE || node->op == GGML_OP_RESHAPE || node->op == GGML_OP_TRANSPOSE || node->op == GGML_OP_VIEW || node->op == GGML_OP_PERMUTE;
 }
 
+static bool ggml_vk_can_fuse(const struct ggml_cgraph * cgraph, int node_idx, std::initializer_list<enum ggml_op> ops) {
+    if (!ggml_can_fuse(cgraph, node_idx, ops)) {
+        return false;
+    }
+
+    if (ops.size() == 2 && ops.begin()[0] == GGML_OP_RMS_NORM && ops.begin()[1] == GGML_OP_MUL) {
+        // additional constraints specific to this fusion
+        const ggml_tensor *rms_norm = cgraph->nodes[node_idx];
+        const ggml_tensor *mul = cgraph->nodes[node_idx + 1];
+
+        GGML_ASSERT(rms_norm->src[0]->type == GGML_TYPE_F32);
+        GGML_ASSERT(rms_norm->type == GGML_TYPE_F32);
+        // rms_norm only supports f32
+        if (mul->src[0]->type != GGML_TYPE_F32 ||
+            mul->src[1]->type != GGML_TYPE_F32 ||
+            mul->type != GGML_TYPE_F32) {
+            return false;
+        }
+        // if rms_norm is the B operand, then we don't handle broadcast
+        if (rms_norm == mul->src[1] &&
+            mul->src[0]->ne[1] != rms_norm->ne[1]) {
+            return false;
+        }
+        // rms_norm shader assumes contiguous rows
+        if (!ggml_is_contiguous_rows(mul->src[0]) || !ggml_is_contiguous_rows(mul->src[1])) {
+            return false;
+        }
+    }
+    return true;
+}
+
 static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cgraph * cgraph) {
     VK_LOG_DEBUG("ggml_backend_vk_graph_compute(" << cgraph->n_nodes << " nodes)");
     ggml_backend_vk_context * ctx = (ggml_backend_vk_context *)backend->context;
 
+    if (vk_instance.debug_utils_support) {
+        vk::DebugUtilsLabelEXT dul = {};
+        dul.pLabelName = "ggml_backend_vk_graph_compute";
+        dul.color = std::array<float,4>{1.0f, 1.0f, 1.0f, 1.0f};
+        vk_instance.pfn_vkQueueBeginDebugUtilsLabelEXT(ctx->device->compute_queue.queue, reinterpret_cast<VkDebugUtilsLabelEXT*>(&dul));
+    }
+
     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, false);
+        if (!ctx->device->disable_fusion && ggml_vk_can_fuse(cgraph, i, { GGML_OP_RMS_NORM, GGML_OP_MUL })) {
+            ctx->num_additional_fused_ops = 1;
+        }
+        ggml_vk_build_graph(ctx, cgraph, 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]);
         }
+        i += ctx->num_additional_fused_ops;
+        ctx->num_additional_fused_ops = 0;
     }
     if (ctx->device->need_compiles) {
         ggml_vk_load_shaders(ctx->device);
     }
     ggml_vk_preallocate_buffers(ctx);
-    ggml_pipeline_allocate_descriptor_sets(ctx->device);
+    ggml_pipeline_allocate_descriptor_sets(ctx);
 
     int last_node = cgraph->n_nodes - 1;
 
@@ -9513,8 +9984,8 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
             if (ctx->device->query_pool) {
                 ctx->device->device.destroyQueryPool(ctx->device->query_pool);
             }
-            VkQueryPoolCreateInfo query_create_info = { VK_STRUCTURE_TYPE_QUERY_POOL_CREATE_INFO };
-            query_create_info.queryType = VK_QUERY_TYPE_TIMESTAMP;
+            vk::QueryPoolCreateInfo query_create_info;
+            query_create_info.queryType = vk::QueryType::eTimestamp;
             query_create_info.queryCount = cgraph->n_nodes + 100;
             ctx->device->query_pool = ctx->device->device.createQueryPool(query_create_info);
             ctx->device->num_queries = query_create_info.queryCount;
@@ -9523,7 +9994,7 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
         ctx->device->device.resetQueryPool(ctx->device->query_pool, 0, cgraph->n_nodes+1);
 
         GGML_ASSERT(ctx->compute_ctx.expired());
-        compute_ctx = ggml_vk_create_context(ctx, ctx->device->compute_queue);
+        compute_ctx = ggml_vk_create_context(ctx, ctx->compute_cmd_pool);
         ctx->compute_ctx = compute_ctx;
         ggml_vk_ctx_begin(ctx->device, compute_ctx);
         compute_ctx->s->buffer.writeTimestamp(vk::PipelineStageFlagBits::eAllCommands, ctx->device->query_pool, 0);
@@ -9547,24 +10018,31 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
             mul_mat_bytes += ggml_nbytes(cgraph->nodes[i]->src[0]);
         }
 
+        if (!ctx->device->disable_fusion && ggml_vk_can_fuse(cgraph, i, { GGML_OP_RMS_NORM, GGML_OP_MUL })) {
+            ctx->num_additional_fused_ops = 1;
+        }
+
         // 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) ||
+                      (i + ctx->num_additional_fused_ops == 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);
+        bool enqueued = ggml_vk_build_graph(ctx, cgraph, i, cgraph->nodes[submit_node_idx], submit_node_idx, false, i + ctx->num_additional_fused_ops == last_node, almost_ready, submit);
 
         if (vk_perf_logger_enabled) {
             if (ctx->compute_ctx.expired()) {
-                compute_ctx = ggml_vk_create_context(ctx, ctx->device->compute_queue);
+                compute_ctx = ggml_vk_create_context(ctx, ctx->compute_cmd_pool);
                 ctx->compute_ctx = compute_ctx;
                 ggml_vk_ctx_begin(ctx->device, compute_ctx);
             } else {
                 compute_ctx = ctx->compute_ctx.lock();
             }
-            compute_ctx->s->buffer.writeTimestamp(vk::PipelineStageFlagBits::eAllCommands, ctx->device->query_pool, i+1);
+            // If there are fused ops, just write out timestamps for all nodes to keep the accounting simple
+            for (int j = 0; j < ctx->num_additional_fused_ops + 1; ++j) {
+                compute_ctx->s->buffer.writeTimestamp(vk::PipelineStageFlagBits::eAllCommands, ctx->device->query_pool, i+j+1);
+            }
         }
 
         if (enqueued) {
@@ -9586,6 +10064,8 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
             }
             submit_count++;
         }
+        i += ctx->num_additional_fused_ops;
+        ctx->num_additional_fused_ops = 0;
     }
 
     if (vk_perf_logger_enabled) {
@@ -9600,7 +10080,7 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
 
         // Get the results and pass them to the logger
         std::vector<uint64_t> timestamps(cgraph->n_nodes + 1);
-        ctx->device->device.getQueryPoolResults(ctx->device->query_pool, 0, cgraph->n_nodes + 1, (cgraph->n_nodes + 1)*sizeof(uint64_t), timestamps.data(), sizeof(uint64_t), vk::QueryResultFlagBits::e64 | vk::QueryResultFlagBits::eWait);
+        VK_CHECK(ctx->device->device.getQueryPoolResults(ctx->device->query_pool, 0, cgraph->n_nodes + 1, (cgraph->n_nodes + 1)*sizeof(uint64_t), timestamps.data(), sizeof(uint64_t), vk::QueryResultFlagBits::e64 | vk::QueryResultFlagBits::eWait), "get timestamp results");
         for (int i = 0; i < cgraph->n_nodes; i++) {
             if (!ggml_vk_is_empty(cgraph->nodes[i])) {
                 ctx->device->perf_logger->log_timing(cgraph->nodes[i], uint64_t((timestamps[i+1] - timestamps[i]) * ctx->device->properties.limits.timestampPeriod));
@@ -9747,6 +10227,7 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
         case GGML_OP_UNARY:
             switch (ggml_get_unary_op(op)) {
                 case GGML_UNARY_OP_GELU:
+                case GGML_UNARY_OP_GELU_ERF:
                 case GGML_UNARY_OP_GELU_QUICK:
                 case GGML_UNARY_OP_SILU:
                 case GGML_UNARY_OP_RELU:
@@ -9760,15 +10241,36 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
                     return false;
             }
             break;
+        case GGML_OP_GLU:
+            switch (ggml_get_glu_op(op)) {
+                case GGML_GLU_OP_GEGLU:
+                case GGML_GLU_OP_REGLU:
+                case GGML_GLU_OP_SWIGLU:
+                case GGML_GLU_OP_GEGLU_ERF:
+                case GGML_GLU_OP_GEGLU_QUICK:
+                    return ggml_is_contiguous(op->src[0]) &&
+                           (op->src[0]->type == GGML_TYPE_F32 || op->src[0]->type == GGML_TYPE_F16) &&
+                           (op->type == GGML_TYPE_F32 || op->type == GGML_TYPE_F16) &&
+                           (op->src[0]->type == op->type);
+                default:
+                    return false;
+            }
+            break;
         case GGML_OP_MUL_MAT:
         case GGML_OP_MUL_MAT_ID:
             {
                 ggml_type src0_type = op->src[0]->type;
                 ggml_backend_vk_device_context * ctx = (ggml_backend_vk_device_context *)dev->context;
                 const vk_device& device = ggml_vk_get_device(ctx->device);
-                if (op->op == GGML_OP_MUL_MAT_ID && !device->mul_mat_id_s[src0_type] && !device->mul_mat_id_m[src0_type] && !device->mul_mat_id_l[src0_type]) {
-                    // If there's not enough shared memory for row_ids and the result tile, fallback to CPU
-                    return false;
+                if (op->op == GGML_OP_MUL_MAT_ID) {
+                    if (!device->mul_mat_id_s[src0_type] && !device->mul_mat_id_m[src0_type] && !device->mul_mat_id_l[src0_type]) {
+                        // If there's not enough shared memory for row_ids and the result tile, fallback to CPU
+                        return false;
+                    }
+                    // Check against size of shared memory variable
+                    if (op->src[2]->ne[0] > 4096) {
+                        return false;
+                    }
                 }
                 switch (src0_type) {
                     case GGML_TYPE_F32:
@@ -9826,19 +10328,8 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
                 ggml_backend_vk_device_context * ctx = (ggml_backend_vk_device_context *)dev->context;
                 auto device = ggml_vk_get_device(ctx->device);
                 bool coopmat2 = device->coopmat2;
-                switch (op->src[0]->ne[0]) {
-                case 64:
-                case 80:
-                case 96:
-                case 112:
-                case 128:
-                case 256:
-                    break;
-                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
+                FaHeadSizes head_sizes = fa_get_head_sizes(op->src[1]->ne[0], op->src[2]->ne[0]);
+                if (head_sizes == FA_HEAD_SIZE_UNSUPPORTED) {
                     return false;
                 }
                 if (op->src[0]->type != GGML_TYPE_F32) {
@@ -9918,6 +10409,12 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
                         return false;
                 }
             } break;
+        case GGML_OP_SET_ROWS:
+            {
+                // TODO: add support
+                // ref: https://github.com/ggml-org/llama.cpp/pull/14274
+                return false;
+            } break;
         case GGML_OP_CONT:
         case GGML_OP_CPY:
         case GGML_OP_DUP:
@@ -10008,6 +10505,7 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
         case GGML_OP_SCALE:
         case GGML_OP_PAD:
         case GGML_OP_DIAG_MASK_INF:
+            return true;
         case GGML_OP_SOFT_MAX:
         case GGML_OP_SOFT_MAX_BACK:
         case GGML_OP_ARGSORT:
@@ -10024,6 +10522,8 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
         case GGML_OP_LEAKY_RELU:
         case GGML_OP_OPT_STEP_ADAMW:
             return true;
+        case GGML_OP_CONV_TRANSPOSE_1D:
+            return op->src[0]->type == GGML_TYPE_F32 && op->src[1]->type == GGML_TYPE_F32;
         default:
             return false;
     }
@@ -10167,11 +10667,28 @@ static bool ggml_vk_instance_portability_enumeration_ext_available(const std::ve
     UNUSED(instance_extensions);
 }
 
+// Extension availability
+static bool ggml_vk_instance_debug_utils_ext_available(
+    const std::vector<vk::ExtensionProperties> & instance_extensions) {
+    // Check for portability enumeration extension for MoltenVK support
+    for (const auto & properties : instance_extensions) {
+        if (strcmp("VK_EXT_debug_utils", properties.extensionName) == 0) {
+            return true;
+        }
+    }
+
+    std::cerr << "ggml_vulkan: WARNING: Instance extension VK_EXT_debug_utils not found." << std::endl;
+    return false;
+
+    UNUSED(instance_extensions);
+}
+
 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
-        return false;
+        // Only allowing Xe2 GPU at the moment since Xe2 GPU can gain significant performance boost,
+        // while some older hardware (ex. Arc A770) has performance regressions
+        return arch == vk_device_architecture::INTEL_XE2;
     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
@@ -10278,11 +10795,21 @@ void * comp_result;
 size_t comp_size;
 size_t comp_nb[GGML_MAX_DIMS];
 size_t check_counter = 0;
-static void ggml_vk_check_results_0(ggml_tensor * tensor) {
+static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_cgraph * cgraph, int tensor_idx) {
+    ggml_tensor * tensor = cgraph->nodes[tensor_idx];
     if (tensor->op == GGML_OP_TRANSPOSE) {
         return;
     }
 
+    bool fused_rms_norm_mul = false;
+    int rms_norm_idx = -1;
+    if (ctx->num_additional_fused_ops == 1 &&
+        tensor->op == GGML_OP_RMS_NORM &&
+        cgraph->nodes[tensor_idx + 1]->op == GGML_OP_MUL) {
+        fused_rms_norm_mul = true;
+        tensor = cgraph->nodes[tensor_idx + 1];
+    }
+
     check_counter++;
     if (!(vk_output_tensor > 0 && vk_output_tensor == check_counter) && check_counter <= vk_skip_checks) {
         return;
@@ -10310,6 +10837,15 @@ static void ggml_vk_check_results_0(ggml_tensor * tensor) {
 
     for (int i = 0; i < 6; i++) {
         ggml_tensor * srci = tensor->src[i];
+        if (fused_rms_norm_mul) {
+            rms_norm_idx = tensor->src[0]->op == GGML_OP_RMS_NORM ? 0 : 1;
+            ggml_tensor *rms_norm = tensor->src[rms_norm_idx];
+            switch (i) {
+            case 0: srci = rms_norm->src[0]; break;
+            case 1: srci = tensor->src[1 - rms_norm_idx]; break;
+            default: continue;
+            }
+        }
         if (srci == nullptr) {
             continue;
         }
@@ -10367,7 +10903,12 @@ static void ggml_vk_check_results_0(ggml_tensor * tensor) {
     } 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, src_clone[0], src_clone[1]);
+        if (fused_rms_norm_mul) {
+            tensor_clone = ggml_rms_norm(ggml_ctx, src_clone[0], *(float *)tensor->src[rms_norm_idx]->op_params);
+            tensor_clone = ggml_mul(ggml_ctx, tensor_clone, src_clone[1 - rms_norm_idx]);
+        } else {
+            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, src_clone[0], src_clone[1]);
     } else if (tensor->op == GGML_OP_CONCAT) {
@@ -10455,6 +10996,9 @@ static void ggml_vk_check_results_0(ggml_tensor * tensor) {
         case GGML_UNARY_OP_GELU:
             tensor_clone = ggml_gelu(ggml_ctx, src_clone[0]);
             break;
+        case GGML_UNARY_OP_GELU_ERF:
+            tensor_clone = ggml_gelu_erf(ggml_ctx, src_clone[0]);
+            break;
         case GGML_UNARY_OP_GELU_QUICK:
             tensor_clone = ggml_gelu_quick(ggml_ctx, src_clone[0]);
             break;
@@ -10471,6 +11015,12 @@ static void ggml_vk_check_results_0(ggml_tensor * tensor) {
             std::cerr << "Missing vk_check_results OP: " << ggml_op_name(tensor->op) << std::endl;
             GGML_ABORT("fatal error");
         }
+    } else if (tensor->op == GGML_OP_GLU) {
+        if (src_clone[1] == nullptr) {
+            tensor_clone = ggml_glu(ggml_ctx, src_clone[0], (ggml_glu_op) tensor->op_params[0], tensor->op_params[1]);
+        } else {
+            tensor_clone = ggml_glu_split(ggml_ctx, src_clone[0], src_clone[1], (ggml_glu_op) tensor->op_params[0]);
+        }
     } else if (tensor->op == GGML_OP_CPY || tensor->op == GGML_OP_DUP) {
         if (src1 == nullptr) {
             tensor_clone = ggml_dup(ggml_ctx, src_clone[0]);
@@ -10515,6 +11065,11 @@ static void ggml_vk_check_results_0(ggml_tensor * tensor) {
         const int32_t dim = tensor->op_params[0];
         const int32_t max_period = tensor->op_params[1];
         tensor_clone = ggml_timestep_embedding(ggml_ctx, src_clone[0], dim, max_period);
+    } else if (tensor->op == GGML_OP_CONV_TRANSPOSE_1D){
+        const int32_t s0 = tensor->op_params[0];
+        const int32_t p0 = tensor->op_params[1];
+        const int32_t d0 = tensor->op_params[2];
+        tensor_clone = ggml_conv_transpose_1d(ggml_ctx, src_clone[0], src_clone[1], s0, p0, d0);
     } else if (tensor->op == GGML_OP_POOL_2D) {
         enum ggml_op_pool op = static_cast<ggml_op_pool>(tensor->op_params[0]);
         const int32_t k0 = tensor->op_params[1];
@@ -10544,10 +11099,10 @@ static void ggml_vk_check_results_0(ggml_tensor * tensor) {
         GGML_ABORT("fatal error");
     }
 
-    ggml_cgraph * cgraph = ggml_new_graph(ggml_ctx);
-    ggml_build_forward_expand(cgraph, tensor_clone);
+    ggml_cgraph * cgraph_cpu = ggml_new_graph(ggml_ctx);
+    ggml_build_forward_expand(cgraph_cpu, tensor_clone);
 
-    ggml_graph_compute_with_ctx(ggml_ctx, cgraph, 8);
+    ggml_graph_compute_with_ctx(ggml_ctx, cgraph_cpu, 8);
 
     if (vk_output_tensor > 0 && vk_output_tensor == check_counter) {
         ggml_vk_print_tensor(tensor_clone, "tensor_clone");
@@ -10570,10 +11125,19 @@ static void ggml_vk_check_results_0(ggml_tensor * tensor) {
     VK_LOG_DEBUG("END ggml_vk_check_results_0(" << tensor->name << ")");
 }
 
-static void ggml_vk_check_results_1(ggml_tensor * tensor) {
+static void ggml_vk_check_results_1(ggml_backend_vk_context * ctx, ggml_cgraph * cgraph, int tensor_idx) {
+    ggml_tensor * tensor = cgraph->nodes[tensor_idx];
     if (tensor->op == GGML_OP_TRANSPOSE) {
         return;
     }
+    bool fused_rms_norm_mul = false;
+    if (ctx->num_additional_fused_ops == 1 &&
+        tensor->op == GGML_OP_RMS_NORM &&
+        cgraph->nodes[tensor_idx + 1]->op == GGML_OP_MUL) {
+        fused_rms_norm_mul = true;
+        tensor = cgraph->nodes[tensor_idx + 1];
+    }
+
     if (!(vk_output_tensor > 0 && vk_output_tensor == check_counter) && check_counter <= vk_skip_checks) {
         return;
     }
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/CMakeLists.txt b/ggml/src/ggml-vulkan/vulkan-shaders/CMakeLists.txt
index e60e9d1e5..e1f613fb4 100644
--- a/ggml/src/ggml-vulkan/vulkan-shaders/CMakeLists.txt
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/CMakeLists.txt
@@ -19,21 +19,13 @@ if (GGML_VULKAN_BFLOAT16_GLSLC_SUPPORT)
     add_compile_definitions(GGML_VULKAN_BFLOAT16_GLSLC_SUPPORT)
     message(STATUS "Enabling bfloat16 glslc support")
 endif()
+if (GGML_VULKAN_SHADER_DEBUG_INFO)
+    add_compile_definitions(GGML_VULKAN_SHADER_DEBUG_INFO)
+    message(STATUS "Enabling shader debug info")
+endif()
 
 set(TARGET vulkan-shaders-gen)
 add_executable(${TARGET} vulkan-shaders-gen.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_compile_features(${TARGET} PRIVATE cxx_std_17)
 target_link_libraries(vulkan-shaders-gen PUBLIC Threads::Threads)
-
-# Configure output directories for MSVC builds
-if(MSVC)
-    # Get the main project's runtime output directory if possible
-    if(DEFINED CMAKE_RUNTIME_OUTPUT_DIRECTORY)
-        foreach(CONFIG ${CMAKE_CONFIGURATION_TYPES})
-            string(TOUPPER ${CONFIG} CONFIG)
-            set_target_properties(${TARGET} PROPERTIES
-                RUNTIME_OUTPUT_DIRECTORY_${CONFIG} ${CMAKE_RUNTIME_OUTPUT_DIRECTORY})
-        endforeach()
-    endif()
-endif()
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/conv_transpose_1d.comp b/ggml/src/ggml-vulkan/vulkan-shaders/conv_transpose_1d.comp
new file mode 100644
index 000000000..b17b4e83e
--- /dev/null
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/conv_transpose_1d.comp
@@ -0,0 +1,98 @@
+#version 450
+
+#include "types.comp"
+
+layout (binding = 0) readonly buffer A {A_TYPE data_a[];};   // src0 - kernel:    [K, Cout, Cin]
+layout (binding = 1) readonly buffer B {B_TYPE data_b[];};   // src1 - input:     [L, Cin]
+layout (binding = 2) writeonly buffer D {D_TYPE data_d[];};     // dst - result      [KL, Cout]
+
+layout(local_size_x = 128 , local_size_y = 1, local_size_z = 1) in;
+
+layout (push_constant) uniform parameter {
+    uint32_t Cout;
+    uint32_t Cin;
+    uint32_t K;
+    uint32_t L;
+    uint32_t KL;
+
+    uint32_t nb01;
+    uint32_t nb02;
+    uint32_t nb11;
+    uint32_t nb1;
+
+    int32_t s0;
+} p;
+
+
+uint32_t Cout_idx = gl_WorkGroupID.x;
+const uint32_t bs = gl_WorkGroupSize.x;
+uint32_t tid = gl_LocalInvocationID.x;
+// Code is more straightforward if we assume it is bs*s0+K instead of (bs-1)*s0+K.
+uint32_t tmp_len = bs*p.s0+p.K;
+shared D_TYPE tmp[4096];
+
+uint splitWork(uint workSize){
+    return (bs + workSize -1) / bs;
+}
+
+void main(){
+    for(uint32_t i = 0; i < splitWork(tmp_len); i++){
+        uint32_t idx = i*bs+tid;
+        if(idx < tmp_len){
+            tmp[idx] = 0.0;
+        }
+    }
+
+    uint32_t L_blocks = splitWork(p.L);
+    for(uint32_t L_block_id = 0; L_block_id < L_blocks; L_block_id++){
+        if(L_block_id > 0){
+            barrier();
+            // Shift values in tmp to the current processing window
+            for(int i = 0; i < splitWork(tmp_len); i++){
+                uint32_t idx = i*bs+tid;
+                if(idx >= bs*p.s0 && idx < tmp_len){
+                    tmp[idx-bs*p.s0] = tmp[idx];
+                    tmp[idx] = 0.0;
+                }else if(idx >= p.K && idx < bs*p.s0){
+                    tmp[idx] = 0.0;
+                }
+            }
+        }
+        barrier();
+
+        // Save contributions of the block to tmp
+        uint32_t L_idx = L_block_id*bs + tid;
+        for(uint32_t K_idx = 0; K_idx < p.K; K_idx++){
+            D_TYPE dp = 0.0;
+            for(uint32_t Cin_idx = 0; Cin_idx < p.Cin; Cin_idx++){
+                A_TYPE elemKrn = data_a[K_idx + Cout_idx * p.nb01 + Cin_idx * p.nb02];
+                if(L_idx < p.L){
+                    B_TYPE elemInp = data_b[L_idx + Cin_idx*p.nb11];
+                    dp = fma(elemKrn, elemInp, dp);
+                }
+            }
+            tmp[tid*p.s0 + K_idx] += dp;
+            barrier();
+        }
+
+        // Save the computed values except the last block that can have different size
+        uint32_t KLb_idx = L_block_id*bs*p.s0;
+        if(L_block_id < L_blocks-1){
+            for(uint32_t s0_idx = 0; s0_idx < p.s0; s0_idx++){
+                uint32_t sh_idx = p.s0*tid+s0_idx;
+                uint32_t KL_idx = KLb_idx+sh_idx;
+                if(KL_idx < p.KL){
+                    data_d[KL_idx + Cout_idx*p.nb1] = tmp[sh_idx];
+                }
+            }
+        }
+    }
+
+    for(uint32_t i = 0; i < splitWork(tmp_len); i++){
+        uint32_t idx = i*bs+tid;
+        uint32_t KL_idx = (L_blocks-1)*bs*p.s0+idx;
+        if(KL_idx < p.KL){
+            data_d[KL_idx + Cout_idx*p.nb1] = tmp[idx];
+        }
+    }
+}
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn.comp b/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn.comp
index ce230a8f7..45c6e7736 100644
--- a/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn.comp
@@ -11,7 +11,8 @@
 #include "types.comp"
 #include "flash_attn_base.comp"
 
-const uint32_t D_per_thread = D / D_split;
+const uint32_t HSK_per_thread = HSK / D_split;
+const uint32_t HSV_per_thread = HSV / D_split;
 
 const uint32_t cols_per_iter = WorkGroupSize / D_split;
 const uint32_t cols_per_thread = Bc / cols_per_iter;
@@ -29,7 +30,7 @@ layout (binding = 3) readonly buffer M {float16_t data_m[];};
 // 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)
 {
-    uint32_t offset = (iq2 + r) * D + c;
+    uint32_t offset = (iq2 + r) * HSV + c;
     data_o[o_offset + offset] = D_TYPE(elem);
     return elem;
 }
@@ -38,7 +39,7 @@ shared FLOAT_TYPE tmpsh[WorkGroupSize];
 shared vec4 tmpshv4[WorkGroupSize];
 
 shared float masksh[Bc][Br];
-shared vec4 Qf[Br][D / 4];
+shared vec4 Qf[Br][HSK / 4];
 
 void main() {
 #ifdef NEEDS_INIT_IQ_SHMEM
@@ -53,18 +54,18 @@ void main() {
 
     uint32_t q_offset = (iq2*p.nb02+iq3*p.nb03) / 4;
 
-    [[unroll]] for (uint32_t idx = 0; idx < Br * D / 4; idx += gl_WorkGroupSize.x) {
-        uint32_t d = (idx + tid) % (D / 4);
-        uint32_t r = (idx + tid) / (D / 4);
-        if (r < Br && d < D / 4 &&
+    [[unroll]] for (uint32_t idx = 0; idx < Br * HSK / 4; idx += gl_WorkGroupSize.x) {
+        uint32_t d = (idx + tid) % (HSK / 4);
+        uint32_t r = (idx + tid) / (HSK / 4);
+        if (r < Br && d < HSK / 4 &&
             i * Br + r < N) {
             Qf[r][d] = vec4(data_qv4[q_offset / 4 + (i * Br + r) * q_stride / 4 + d]) * p.scale;
         }
     }
     barrier();
 
-    vec4 Of[Br][D_per_thread / 4];
-    [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
+    vec4 Of[Br][HSV_per_thread / 4];
+    [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
         [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
             Of[r][d] = vec4(0.0);
         }
@@ -99,6 +100,10 @@ void main() {
     uint32_t k_offset = (ik2*p.nb12 + ik3*p.nb13) / 2;
     uint32_t v_offset = (iv2*p.nb22 + iv3*p.nb23) / 2;
 #endif
+    uint32_t m_offset = 0;
+    if (p.nem2 != 1 || p.nem3 != 1) {
+        m_offset = ((iq3 % p.nem3) * p.nem2 + (iq2 % p.nem2)) * p.nem1 * KV;
+    }
 
     [[dont_unroll]]
     for (uint32_t j = start_j; j < end_j; ++j) {
@@ -112,7 +117,7 @@ void main() {
 
 
         [[unroll]] for (uint32_t c = 0; c < cols_per_thread; ++c) {
-            [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
+            [[unroll]] for (uint32_t d = 0; d < HSK_per_thread / 4; ++d) {
 #if BLOCK_SIZE > 1
                 uint coord = (j * Bc + c * cols_per_iter + col_tid) * k_stride * BLOCK_SIZE + 4 * (d * D_split + d_tid);
                 uint ib = coord / BLOCK_SIZE;
@@ -144,13 +149,13 @@ void main() {
             }
         }
 
-        if (p.mask != 0) {
+        if ((p.mask_n_head_log2 & MASK_ENABLE_BIT) != 0) {
 
             [[unroll]] for (uint32_t idx = 0; idx < Bc * Br; idx += gl_WorkGroupSize.x) {
                 uint32_t c = (idx + tid) % Bc;
                 uint32_t r = (idx + tid) / Bc;
                 if (idx + tid < Bc * Br) {
-                    masksh[c][r] = float(data_m[(i * Br + r) * m_stride + (j * Bc + c)]);
+                    masksh[c][r] = float(data_m[m_offset + (i * Br + r) * m_stride + (j * Bc + c)]);
                 }
             }
             barrier();
@@ -191,14 +196,14 @@ void main() {
             Lf[r] = eMf[r]*Lf[r] + rowsumf[r];
         }
 
-        [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
+        [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
             [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
                 Of[r][d] = eMf[r] * Of[r][d];
             }
         }
 
         [[unroll]] for (uint32_t c = 0; c < cols_per_thread; ++c) {
-            [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
+            [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
 #if BLOCK_SIZE > 1
                 uint coord = (j * Bc + c * cols_per_iter + col_tid) * v_stride * BLOCK_SIZE + 4 * (d * D_split + d_tid);
                 uint ib = coord / BLOCK_SIZE;
@@ -255,7 +260,7 @@ void main() {
         Lf[r] = tmpsh[d_tid];
         barrier();
 
-        [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
+        [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
 
             Of[r][d] = eMf * Of[r][d];
             tmpshv4[tid] = Of[r][d];
@@ -277,11 +282,11 @@ void main() {
     // 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) {
-        uint32_t o_offset = D * p.ne1 * split_k_index;
+        uint32_t o_offset = HSV * p.ne1 * (split_k_index + iq3 * p.k_num);
 
         [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
             if (r < N) {
-                [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
+                [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
                     [[unroll]] for (uint32_t comp = 0; comp < 4; ++comp) {
                         perElemOpGqaStore(r, 4*(d * D_split + d_tid) + comp, Of[r][d][comp], o_offset, iq2, N);
                     }
@@ -289,7 +294,7 @@ void main() {
             }
         }
 
-        o_offset = D * p.ne1 * p.k_num + p.ne1 * split_k_index * 2;
+        o_offset = HSV * p.ne1 * p.ne3 * p.k_num + p.ne1 * (split_k_index + iq3 * p.k_num) * 2;
         [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
             if (r < N) {
                 perElemOpStoreCol0(r, 0u, ACC_TYPE(Lf[r]), o_offset, iq2, N);
@@ -305,18 +310,18 @@ void main() {
         Lfrcp[r] = 1.0 / Lf[r];
     }
 
-    [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
+    [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
         [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
             Of[r][d] *= Lfrcp[r];
         }
     }
 
-    uint32_t o_offset = iq3*p.ne2*p.ne1;
+    uint32_t o_offset = iq3*p.ne2*p.ne1*HSV;
 
     if (p.gqa_ratio > 1) {
         [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
             if (r < N) {
-                [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
+                [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
                     [[unroll]] for (uint32_t comp = 0; comp < 4; ++comp) {
                         perElemOpGqaStore(r, 4*(d * D_split + d_tid) + comp, Of[r][d][comp], o_offset, iq2, N);
                     }
@@ -326,9 +331,9 @@ void main() {
     } else {
         [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
             if (i * Br + r < N) {
-                [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
+                [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
                     [[unroll]] for (uint32_t comp = 0; comp < 4; ++comp) {
-                        data_o[o_offset + iq2 * D + (i * Br + r) * p.ne1 * D + 4*(d * D_split + d_tid) + comp] = D_TYPE(Of[r][d][comp]);
+                        data_o[o_offset + iq2 * HSV + (i * Br + r) * p.ne1 * HSV + 4*(d * D_split + d_tid) + comp] = D_TYPE(Of[r][d][comp]);
                     }
                 }
             }
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_base.comp b/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_base.comp
index 61d90e2d8..7defe72b4 100644
--- a/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_base.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_base.comp
@@ -4,10 +4,10 @@ layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
 layout (constant_id = 0) const uint32_t WorkGroupSize = 128;
 layout (constant_id = 1) const uint32_t Br = 1;
 layout (constant_id = 2) const uint32_t Bc = 32;
-layout (constant_id = 3) const uint32_t D = 32;
-layout (constant_id = 4) const uint32_t Clamp = 0;
-layout (constant_id = 5) const uint32_t D_split = 16;
-
+layout (constant_id = 3) const uint32_t HSK = 32;
+layout (constant_id = 4) const uint32_t HSV = 32;
+layout (constant_id = 5) const uint32_t Clamp = 0;
+layout (constant_id = 6) const uint32_t D_split = 16;
 
 layout (push_constant) uniform parameter {
     uint32_t N;
@@ -24,6 +24,8 @@ layout (push_constant) uniform parameter {
     uint32_t nev2;
     uint32_t nev3;
     uint32_t nem1;
+    uint32_t nem2;
+    uint32_t nem3;
 
     uint32_t nb01;
     uint32_t nb02;
@@ -34,14 +36,12 @@ layout (push_constant) uniform parameter {
     uint32_t nb21;
     uint32_t nb22;
     uint32_t nb23;
-    uint32_t nb31;
 
     float scale;
     float max_bias;
     float logit_softcap;
 
-    uint32_t mask;
-    uint32_t n_head_log2;
+    uint32_t mask_n_head_log2;
     float m0;
     float m1;
 
@@ -50,6 +50,9 @@ layout (push_constant) uniform parameter {
     uint32_t k_num;
 } p;
 
+#define MASK_ENABLE_BIT (1<<16)
+#define N_LOG2_MASK 0xFFFF
+
 layout (binding = 4) writeonly buffer O {D_TYPE data_o[];};
 
 #if defined(A_TYPE_PACKED16)
@@ -100,8 +103,10 @@ ACC_TYPE perElemOpComputeSlope(const in uint32_t r, const in uint32_t c, const i
 {
     const uint32_t h = iq2 + (r % p.gqa_ratio);
 
-    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);
+    uint32_t n_head_log2 = p.mask_n_head_log2 & N_LOG2_MASK;
+
+    const ACC_TYPE base = ACC_TYPE(h < n_head_log2 ? p.m0 : p.m1);
+    const int      exph = int(h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1);
 
     return ACC_TYPE(pow(base, ACC_TYPE(exph)));
 }
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm1.comp b/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm1.comp
index da478be24..486735fe8 100644
--- a/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm1.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm1.comp
@@ -13,7 +13,9 @@
 #include "types.comp"
 #include "flash_attn_base.comp"
 
-const uint32_t D_per_thread = D / D_split;
+const uint32_t HSK_per_thread = HSK / D_split;
+const uint32_t HSV_per_thread = HSV / D_split;
+
 const uint32_t row_split = 4;
 const uint32_t rows_per_thread = Br / row_split;
 const uint32_t cols_per_iter = gl_WorkGroupSize.x / D_split / row_split;
@@ -32,7 +34,7 @@ layout (binding = 3) readonly buffer M {float16_t data_m[];};
 // 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)
 {
-    uint32_t offset = (iq2 + r) * D + c;
+    uint32_t offset = (iq2 + r) * HSV + c;
     data_o[o_offset + offset] = D_TYPE(elem);
     return elem;
 }
@@ -44,14 +46,14 @@ const uint32_t MatBc = 16;
 shared FLOAT_TYPE tmpsh[gl_WorkGroupSize.x];
 shared ACC_TYPEV4 tmpshv4[gl_WorkGroupSize.x];
 
-const uint32_t qstride = D / 4 + 2; // in units of f16vec4
+const uint32_t qstride = HSK / 4 + 2; // in units of f16vec4
 shared f16vec4 Qf[Br * qstride];
 
-// Avoid padding for D==256 to make it fit in 48KB shmem.
-const uint32_t sfshstride = (D <= 128) ? (Br + 8) : Br;
+// Avoid padding for hsk==256 to make it fit in 48KB shmem.
+const uint32_t sfshstride = (HSK <= 128) ? (Br + 8) : Br;
 shared ACC_TYPE sfsh[Bc * sfshstride];
 
-const uint32_t kshstride = D / 4 + 2; // in units of f16vec4
+const uint32_t kshstride = HSK / 4 + 2; // in units of f16vec4
 shared f16vec4 ksh[Bc * kshstride];
 
 shared float slope[Br];
@@ -74,18 +76,18 @@ void main() {
 
     uint32_t q_offset = (iq2*p.nb02+iq3*p.nb03) / 4;
 
-    [[unroll]] for (uint32_t idx = 0; idx < Br * D / 4; idx += gl_WorkGroupSize.x) {
-        uint32_t d = (idx + tid) % (D / 4);
-        uint32_t r = (idx + tid) / (D / 4);
-        if (r < Br && d < D / 4 &&
+    [[unroll]] for (uint32_t idx = 0; idx < Br * HSK / 4; idx += gl_WorkGroupSize.x) {
+        uint32_t d = (idx + tid) % (HSK / 4);
+        uint32_t r = (idx + tid) / (HSK / 4);
+        if (r < Br && d < HSK / 4 &&
             i * Br + r < N) {
             Qf[r * qstride + d] = f16vec4(data_qv4[q_offset / 4 + (i * Br + r) * q_stride / 4 + d] * p.scale);
         }
     }
     barrier();
 
-    ACC_TYPEV4 Of[rows_per_thread][D_per_thread / 4];
-    [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
+    ACC_TYPEV4 Of[rows_per_thread][HSV_per_thread / 4];
+    [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
         [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
             Of[r][d] = ACC_TYPEV4(0.0);
         }
@@ -123,14 +125,18 @@ void main() {
     uint32_t k_offset = (ik2*p.nb12 + ik3*p.nb13) / 2;
     uint32_t v_offset = (iv2*p.nb22 + iv3*p.nb23) / 2;
 #endif
+    uint32_t m_offset = 0;
+    if (p.nem2 != 1 || p.nem3 != 1) {
+        m_offset = ((iq3 % p.nem3) * p.nem2 + (iq2 % p.nem2)) * p.nem1 * KV;
+    }
 
     [[dont_unroll]]
     for (uint32_t j = start_j; j < end_j; ++j) {
 
-        [[unroll]] for (uint32_t idx = 0; idx < Bc * D / 4; idx += gl_WorkGroupSize.x) {
-            uint32_t d = (idx + tid) % (D / 4);
-            uint32_t c = (idx + tid) / (D / 4);
-            if (c < Bc && d < D / 4) {
+        [[unroll]] for (uint32_t idx = 0; idx < Bc * HSK / 4; idx += gl_WorkGroupSize.x) {
+            uint32_t d = (idx + tid) % (HSK / 4);
+            uint32_t c = (idx + tid) / (HSK / 4);
+            if (c < Bc && d < HSK / 4) {
 #if BLOCK_SIZE > 1
                 uint coord = (j * Bc + c) * k_stride * BLOCK_SIZE + 4 * d;
                 uint ib = coord / BLOCK_SIZE;
@@ -145,14 +151,14 @@ void main() {
         }
         barrier();
 
-        // K * Q^T -> S^T: Bc x D * D x Br -> Bc x Br
-        // Bc split across workgroup (four subgroups), loop over D in chunks of 16: 16 x 16 * 16 x 16 -> 16 x 16
+        // K * Q^T -> S^T: Bc x HSK * HSK x Br -> Bc x Br
+        // Bc split across workgroup (four subgroups), loop over HSK in chunks of 16: 16 x 16 * 16 x 16 -> 16 x 16
         // This is written transposed in order to allow for N being 8 if implementations need it
         coopmat<ACC_TYPE, gl_ScopeSubgroup, MatBc, MatBr, gl_MatrixUseAccumulator> SfMat = coopmat<ACC_TYPE, gl_ScopeSubgroup, MatBc, MatBr, gl_MatrixUseAccumulator>(0);
         coopmat<float16_t, gl_ScopeSubgroup, MatBc, 16, gl_MatrixUseA> KMat;
         coopmat<float16_t, gl_ScopeSubgroup, 16, MatBr, gl_MatrixUseB> QMat;
 
-        for (uint32_t d = 0; d < D / 16; ++d) {
+        for (uint32_t d = 0; d < HSK / 16; ++d) {
             coopMatLoad(QMat, Qf, d * 16 / 4, qstride, gl_CooperativeMatrixLayoutColumnMajor);
 
             uint coord = (gl_SubgroupID * MatBc) * kshstride + d * 16 / 4;
@@ -176,12 +182,12 @@ void main() {
             barrier();
         }
 
-        if (p.mask != 0) {
+        if ((p.mask_n_head_log2 & MASK_ENABLE_BIT) != 0) {
             [[unroll]] for (uint32_t idx = 0; idx < Bc * Br; idx += gl_WorkGroupSize.x) {
                 uint32_t c = (idx + tid) % Bc;
                 uint32_t r = (idx + tid) / Bc;
                 if (idx + tid < Bc * Br || idx + gl_WorkGroupSize.x <= Bc * Br) {
-                    sfsh[c * sfshstride + r] += ACC_TYPE(slope[r] * float(data_m[(i * Br + r) * m_stride + (j * Bc + c)]));
+                    sfsh[c * sfshstride + r] += ACC_TYPE(slope[r] * float(data_m[m_offset + (i * Br + r) * m_stride + (j * Bc + c)]));
                 }
             }
             barrier();
@@ -202,7 +208,7 @@ void main() {
             eMf[r] = exp(Moldf - Mf[r]);
         }
 
-        [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
+        [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
             [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
                 Of[r][d] = float16_t(eMf[r]) * Of[r][d];
             }
@@ -217,7 +223,7 @@ void main() {
                 Pf[r] = exp(sfsh[tile_row(r) + (c * cols_per_iter + col_tid) * sfshstride] - Mf[r]);
                 Lf[r] += Pf[r];
             }
-            [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
+            [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
 #if BLOCK_SIZE > 1
                 uint coord = (j * Bc + c * cols_per_iter + col_tid) * v_stride * BLOCK_SIZE + 4 * (d * D_split + d_tid);
                 uint ib = coord / BLOCK_SIZE;
@@ -280,7 +286,7 @@ void main() {
     }
 
     [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
-        [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
+        [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
 
             Of[r][d] = float16_t(eMf[r]) * Of[r][d];
             tmpshv4[tid] = Of[r][d];
@@ -300,11 +306,11 @@ void main() {
     // 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) {
-        uint32_t o_offset = D * p.ne1 * split_k_index;
+        uint32_t o_offset = HSV * p.ne1 * (split_k_index + iq3 * p.k_num);
 
         [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
             if (tile_row(r) < N) {
-                [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
+                [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
                     [[unroll]] for (uint32_t comp = 0; comp < 4; ++comp) {
                         perElemOpGqaStore(tile_row(r), 4*(d * D_split + d_tid) + comp, float(Of[r][d][comp]), o_offset, iq2, N);
                     }
@@ -312,7 +318,7 @@ void main() {
             }
         }
 
-        o_offset = D * p.ne1 * p.k_num + p.ne1 * split_k_index * 2;
+        o_offset = HSV * p.ne1 * p.ne3 * p.k_num + p.ne1 * (split_k_index + iq3 * p.k_num) * 2;
         [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
             if (tile_row(r) < N) {
                 perElemOpStoreCol0(tile_row(r), 0u, ACC_TYPE(Lf[r]), o_offset, iq2, N);
@@ -328,18 +334,18 @@ void main() {
         Lfrcp[r] = 1.0 / Lf[r];
     }
 
-    [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
+    [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
         [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
             Of[r][d] *= float16_t(Lfrcp[r]);
         }
     }
 
-    uint32_t o_offset = iq3*p.ne2*p.ne1;
+    uint32_t o_offset = iq3*p.ne2*p.ne1*HSV;
 
     if (p.gqa_ratio > 1) {
         [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
             if (tile_row(r) < N) {
-                [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
+                [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
                     [[unroll]] for (uint32_t comp = 0; comp < 4; ++comp) {
                         perElemOpGqaStore(tile_row(r), 4*(d * D_split + d_tid) + comp, float(Of[r][d][comp]), o_offset, iq2, N);
                     }
@@ -349,9 +355,9 @@ void main() {
     } else {
         [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
             if (i * Br + tile_row(r) < N) {
-                [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
+                [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
                     [[unroll]] for (uint32_t comp = 0; comp < 4; ++comp) {
-                        data_o[o_offset + iq2 * D + (i * Br + tile_row(r)) * p.ne1 * D + 4*(d * D_split + d_tid) + comp] = D_TYPE(Of[r][d][comp]);
+                        data_o[o_offset + iq2 * HSV + (i * Br + tile_row(r)) * p.ne1 * HSV + 4*(d * D_split + d_tid) + comp] = D_TYPE(Of[r][d][comp]);
                     }
                 }
             }
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 6acf67a03..274f48fca 100644
--- a/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm2.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm2.comp
@@ -61,8 +61,8 @@ ACC_TYPE Max(const in uint32_t row, const in uint32_t col, const in ACC_TYPE ele
 // 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;
+    if (r < N && c < HSV) {
+        uint32_t offset = (iq2 + r) * HSV + c;
         data_o[o_offset + offset] = D_TYPE(elem);
     }
     return elem;
@@ -86,9 +86,9 @@ void main() {
     tensorLayoutV = setTensorLayoutBlockSizeNV(tensorLayoutV, 1, BLOCK_SIZE);
 #endif
 
-    tensorLayoutQ = setTensorLayoutDimensionNV(tensorLayoutQ, N, D);
-    tensorLayoutK = setTensorLayoutDimensionNV(tensorLayoutK, KV, D);
-    tensorLayoutV = setTensorLayoutDimensionNV(tensorLayoutV, KV, D);
+    tensorLayoutQ = setTensorLayoutDimensionNV(tensorLayoutQ, N, HSK);
+    tensorLayoutK = setTensorLayoutDimensionNV(tensorLayoutK, KV, HSK);
+    tensorLayoutV = setTensorLayoutDimensionNV(tensorLayoutV, KV, HSV);
 
     // hint to the compiler that strides are aligned for the aligned variant of the shader
     if (Clamp != gl_CooperativeMatrixClampModeConstantNV)
@@ -104,16 +104,16 @@ void main() {
     tensorLayoutK = setTensorLayoutStrideNV(tensorLayoutK, k_stride, 1);
     tensorLayoutV = setTensorLayoutStrideNV(tensorLayoutV, v_stride, 1);
 
-    coopmat<Q_TYPE, gl_ScopeWorkgroup, Br, D, gl_MatrixUseAccumulator> Q;
-    coopmat<float16_t, gl_ScopeWorkgroup, Br, D, gl_MatrixUseA> Qf16;
+    coopmat<Q_TYPE, gl_ScopeWorkgroup, Br, HSK, gl_MatrixUseAccumulator> Q;
+    coopmat<float16_t, gl_ScopeWorkgroup, Br, HSK, gl_MatrixUseA> Qf16;
 
     uint32_t q_offset = iq2*p.nb02+iq3*p.nb03;
-    coopMatLoadTensorNV(Q, data_q, q_offset, sliceTensorLayoutNV(tensorLayoutQ, i * Br, Br, 0, D));
+    coopMatLoadTensorNV(Q, data_q, q_offset, sliceTensorLayoutNV(tensorLayoutQ, i * Br, Br, 0, HSK));
 
-    Qf16 = coopmat<float16_t, gl_ScopeWorkgroup, Br, D, gl_MatrixUseA>(Q);
+    Qf16 = coopmat<float16_t, gl_ScopeWorkgroup, Br, HSK, gl_MatrixUseA>(Q);
     Qf16 *= float16_t(p.scale);
 
-    coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, D, gl_MatrixUseAccumulator> O = coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, D, gl_MatrixUseAccumulator>(0);
+    coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, HSV, gl_MatrixUseAccumulator> O = coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, HSV, gl_MatrixUseAccumulator>(0);
 
     coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator> L, M;
 
@@ -130,15 +130,20 @@ void main() {
         coopMatPerElementNV(slopeMat, slopeMat, perElemOpComputeSlope, iq2);
     }
 
+    uint32_t m_offset = 0;
+    if (p.nem2 != 1 || p.nem3 != 1) {
+        m_offset = ((iq3 % p.nem3) * p.nem2 + (iq2 % p.nem2)) * p.nem1 * KV * 2 /*sizeof(float16_t)*/;
+    }
+
     [[dont_unroll]]
     for (uint32_t j = start_j; j < end_j; ++j) {
 
         coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator> S = coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator>(0);
 
-        coopmat<float16_t, gl_ScopeWorkgroup, D, Bc, gl_MatrixUseB> K_T;
+        coopmat<float16_t, gl_ScopeWorkgroup, HSK, Bc, gl_MatrixUseB> K_T;
 
         uint32_t k_offset = ik2*p.nb12 + ik3*p.nb13;
-        coopMatLoadTensorNV(K_T, data_k, k_offset, sliceTensorLayoutNV(tensorLayoutK, j * Bc, Bc, 0, D), tensorViewTranspose DECODEFUNC);
+        coopMatLoadTensorNV(K_T, data_k, k_offset, sliceTensorLayoutNV(tensorLayoutK, j * Bc, Bc, 0, HSK), tensorViewTranspose DECODEFUNC);
         S = coopMatMulAdd(Qf16, K_T, S);
 
         if (p.logit_softcap != 0.0f) {
@@ -148,14 +153,14 @@ void main() {
             }
         }
 
-        if (p.mask != 0) {
+        if ((p.mask_n_head_log2 & MASK_ENABLE_BIT) != 0) {
             tensorLayoutNV<2, Clamp> tensorLayoutM = createTensorLayoutNV(2, Clamp);
             tensorLayoutM = setTensorLayoutDimensionNV(tensorLayoutM, p.nem1, KV);
             tensorLayoutM = setTensorLayoutStrideNV(tensorLayoutM, m_stride, 1);
 
             coopmat<float16_t, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator> mv;
 
-            coopMatLoadTensorNV(mv, data_m, 0, sliceTensorLayoutNV(tensorLayoutM, i * Br, Br, j * Bc, Bc));
+            coopMatLoadTensorNV(mv, data_m, m_offset, sliceTensorLayoutNV(tensorLayoutM, i * Br, Br, j * Bc, Bc));
 
             S += slopeMat*coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator>(mv);
         }
@@ -203,42 +208,42 @@ void main() {
         rowsum = coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator>(0.0);
         rowsum = coopMatMulAdd(P_A, One, rowsum);
 
-        coopmat<float16_t, gl_ScopeWorkgroup, Bc, D, gl_MatrixUseB> V;
+        coopmat<float16_t, gl_ScopeWorkgroup, Bc, HSV, gl_MatrixUseB> V;
         uint32_t v_offset = iv2*p.nb22 + iv3*p.nb23;
-        coopMatLoadTensorNV(V,  data_v, v_offset, sliceTensorLayoutNV(tensorLayoutV, j * Bc, Bc, 0, D) DECODEFUNC);
+        coopMatLoadTensorNV(V,  data_v, v_offset, sliceTensorLayoutNV(tensorLayoutV, j * Bc, Bc, 0, HSV) DECODEFUNC);
 
         L = eM*L + rowsum;
 
         // This is the "diagonal" matrix in the paper, but since we do componentwise
         // multiply rather than matrix multiply it has the diagonal element smeared
         // across the row
-        coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, D, gl_MatrixUseAccumulator> eMdiag;
+        coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, HSV, gl_MatrixUseAccumulator> eMdiag;
 
         // resize eM by using smear/reduce
         coopMatReduceNV(eMdiag, eM, gl_CooperativeMatrixReduceRowNV, smearReduce);
 
         // multiply with fp16 accumulation, then add to O.
-        coopmat<float16_t, gl_ScopeWorkgroup, Br, D, gl_MatrixUseAccumulator> PV = coopmat<float16_t, gl_ScopeWorkgroup, Br, D, gl_MatrixUseAccumulator>(0);
+        coopmat<float16_t, gl_ScopeWorkgroup, Br, HSV, gl_MatrixUseAccumulator> PV = coopmat<float16_t, gl_ScopeWorkgroup, Br, HSV, gl_MatrixUseAccumulator>(0);
         PV = coopMatMulAdd(P_A, V, PV);
 
-        O = eMdiag * O + coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, D, gl_MatrixUseAccumulator>(PV);
+        O = eMdiag * O + coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, HSV, gl_MatrixUseAccumulator>(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<D_TYPE, gl_ScopeWorkgroup, Br, D, gl_MatrixUseAccumulator> O_D = coopmat<D_TYPE, gl_ScopeWorkgroup, Br, D, gl_MatrixUseAccumulator>(O);
+        coopmat<D_TYPE, gl_ScopeWorkgroup, Br, HSV, gl_MatrixUseAccumulator> O_D = coopmat<D_TYPE, gl_ScopeWorkgroup, Br, HSV, gl_MatrixUseAccumulator>(O);
 
-        uint32_t o_offset = D * p.ne1 * split_k_index;
+        uint32_t o_offset = HSV * p.ne1 * (split_k_index + iq3 * p.k_num);
         coopMatPerElementNV(O_D, O_D, perElemOpGqaStore, o_offset, iq2, N);
 
-        o_offset = D * p.ne1 * p.k_num + p.ne1 * split_k_index * 2;
+        o_offset = HSV * p.ne1 * p.ne3 * p.k_num + p.ne1 * (split_k_index + iq3 * p.k_num) * 2;
         coopMatPerElementNV(L, L, perElemOpStoreCol0, o_offset, iq2, N);
         coopMatPerElementNV(M, M, perElemOpStoreCol0, o_offset + p.ne1, iq2, N);
         return;
     }
 
-    coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, D, gl_MatrixUseAccumulator> Ldiag;
+    coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, HSV, gl_MatrixUseAccumulator> Ldiag;
 
     // resize L by using smear/reduce
     coopMatReduceNV(Ldiag, L, gl_CooperativeMatrixReduceRowNV, smearReduce);
@@ -250,18 +255,18 @@ void main() {
 
     O = Ldiag*O;
 
-    uint32_t o_offset = iq3*p.ne2*p.ne1;
+    uint32_t o_offset = iq3*p.ne2*p.ne1*HSV;
 
-    coopmat<D_TYPE, gl_ScopeWorkgroup, Br, D, gl_MatrixUseAccumulator> O_D = coopmat<D_TYPE, gl_ScopeWorkgroup, Br, D, gl_MatrixUseAccumulator>(O);
+    coopmat<D_TYPE, gl_ScopeWorkgroup, Br, HSV, gl_MatrixUseAccumulator> O_D = coopmat<D_TYPE, gl_ScopeWorkgroup, Br, HSV, gl_MatrixUseAccumulator>(O);
     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);
+        tensorLayoutD = setTensorLayoutDimensionNV(tensorLayoutD, p.ne2, p.ne1, HSV);
 
         // 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);
+        coopMatStoreTensorNV(O_D, data_o, o_offset, sliceTensorLayoutNV(tensorLayoutD, i * Br, Br, iq2, N, 0, HSV), 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
index a7e395685..599cef072 100644
--- 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
@@ -12,6 +12,7 @@ layout (binding = 1) writeonly buffer D {float data_d[];};
 layout (push_constant) uniform parameter {
     uint D;
     uint N;
+    uint ne3;
     uint k_num;
 } p;
 
@@ -19,13 +20,14 @@ void main() {
     // Each workgroup handles a row
     const uint n = gl_WorkGroupID.x;
     const uint tid = gl_LocalInvocationID.x;
+    const uint iq3 = gl_WorkGroupID.z;
 
     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 l_offset = D * N * p.ne3 * k_num + N * iq3 * k_num * 2 + n;
+    uint m_offset = D * N * p.ne3 * k_num + N * iq3 * k_num * 2 + N + n;
     uint lm_stride = N * 2;
 
     // Compute the max m value for the row
@@ -49,11 +51,11 @@ void main() {
     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;
+            uint o_offset = D * N * (k + iq3 * k_num) + 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;
+        data_d[iq3 * D * N + D * n + d] = O;
     }
 }
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/geglu.comp b/ggml/src/ggml-vulkan/vulkan-shaders/geglu.comp
new file mode 100644
index 000000000..f4268ed24
--- /dev/null
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/geglu.comp
@@ -0,0 +1,13 @@
+#version 450
+
+#include "glu_head.comp"
+
+const float GELU_COEF_A    = 0.044715f;
+const float SQRT_2_OVER_PI = 0.79788456080286535587989211986876f;
+
+float op(float a, float b) {
+    const float val = SQRT_2_OVER_PI*a*(1.0f + GELU_COEF_A*a*a);
+    return 0.5f*a*(2.0f - 2.0f / (exp(2 * val) + 1)) * b;
+}
+
+#include "glu_main.comp"
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/geglu_erf.comp b/ggml/src/ggml-vulkan/vulkan-shaders/geglu_erf.comp
new file mode 100644
index 000000000..cbd4cb36b
--- /dev/null
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/geglu_erf.comp
@@ -0,0 +1,27 @@
+#version 450
+
+#include "glu_head.comp"
+
+// based on Abramowitz and Stegun formula 7.1.26 or similar Hastings' approximation
+// ref: https://www.johndcook.com/blog/python_erf/
+const float p_erf  = 0.3275911f;
+const float a1_erf = 0.254829592f;
+const float a2_erf = -0.284496736f;
+const float a3_erf = 1.421413741f;
+const float a4_erf = -1.453152027f;
+const float a5_erf = 1.061405429f;
+
+const float SQRT_2_INV = 0.70710678118654752440084436210484f;
+
+float op(float a, float b) {
+    const float a_div_sqr2 = a * SQRT_2_INV;
+    const float sign_x = sign(a_div_sqr2);
+    const float x = abs(a_div_sqr2);
+    const float t = 1.0f / (1.0f + p_erf * x);
+    const float y = 1.0f - (((((a5_erf * t + a4_erf) * t) + a3_erf) * t + a2_erf) * t + a1_erf) * t * exp(-x * x);
+    const float erf_approx = sign_x * y;
+
+    return 0.5f * a * (1.0f + erf_approx) * b;
+}
+
+#include "glu_main.comp"
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/geglu_quick.comp b/ggml/src/ggml-vulkan/vulkan-shaders/geglu_quick.comp
new file mode 100644
index 000000000..3a2a6897b
--- /dev/null
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/geglu_quick.comp
@@ -0,0 +1,11 @@
+#version 450
+
+#include "glu_head.comp"
+
+const float GELU_QUICK_COEF = -1.702f;
+
+float op(float a, float b) {
+    return a * (1.0f / (1.0f + exp(GELU_QUICK_COEF * a))) * b;
+}
+
+#include "glu_main.comp"
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/gelu_erf.comp b/ggml/src/ggml-vulkan/vulkan-shaders/gelu_erf.comp
new file mode 100644
index 000000000..5fd5a5e70
--- /dev/null
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/gelu_erf.comp
@@ -0,0 +1,39 @@
+#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() {
+    // based on Abramowitz and Stegun formula 7.1.26 or similar Hastings' approximation
+    // ref: https://www.johndcook.com/blog/python_erf/
+    const float p_erf  = 0.3275911f;
+    const float a1_erf = 0.254829592f;
+    const float a2_erf = -0.284496736f;
+    const float a3_erf = 1.421413741f;
+    const float a4_erf = -1.453152027f;
+    const float a5_erf = 1.061405429f;
+
+    const float SQRT_2_INV = 0.70710678118654752440084436210484f;
+    const uint i = gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
+
+    if (i >= p.KX) {
+        return;
+    }
+
+    const float a = float(data_a[i]);
+    const float a_div_sqr2 = a * SQRT_2_INV;
+    const float sign_x = sign(a_div_sqr2);
+    const float x = abs(a_div_sqr2);
+    const float t = 1.0f / (1.0f + p_erf * x);
+    const float y = 1.0f - (((((a5_erf * t + a4_erf) * t) + a3_erf) * t + a2_erf) * t + a1_erf) * t * exp(-x * x);
+    const float erf_approx = sign_x * y;
+
+    data_d[i] = D_TYPE(0.5f * a * (1.0f + erf_approx));
+}
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/glu_head.comp b/ggml/src/ggml-vulkan/vulkan-shaders/glu_head.comp
new file mode 100644
index 000000000..41a298890
--- /dev/null
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/glu_head.comp
@@ -0,0 +1,15 @@
+#extension GL_EXT_shader_16bit_storage : require
+
+layout(local_size_x = 512, 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 {A_TYPE data_b[];};
+layout (binding = 2) writeonly buffer D {D_TYPE data_d[];};
+
+layout (push_constant) uniform parameter
+{
+    uint N;
+    uint ne00;
+    uint ne20;
+    uint mode;
+} p;
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/glu_main.comp b/ggml/src/ggml-vulkan/vulkan-shaders/glu_main.comp
new file mode 100644
index 000000000..85cf65a9e
--- /dev/null
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/glu_main.comp
@@ -0,0 +1,29 @@
+void main() {
+    const uint i = gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
+
+    if (i >= p.N) {
+        return;
+    }
+
+    const uint row = i / p.ne20;
+    const uint col = i - row * p.ne20;
+
+    if (p.mode == 0) {
+        // Default
+        const uint offset = p.ne00 / 2;
+        const uint idx = row * p.ne00 + col;
+
+        data_d[row * offset + col] = D_TYPE(op(float(data_a[idx]), float(data_a[idx + offset])));
+    } else if (p.mode == 1) {
+        // Swapped
+        const uint offset = p.ne00 / 2;
+        const uint idx = row * p.ne00 + col;
+
+        data_d[row * offset + col] = D_TYPE(op(float(data_a[idx + offset]), float(data_a[idx])));
+    } else {
+        // Split
+        const uint idx = row * p.ne00 + col;
+
+        data_d[idx] = D_TYPE(op(float(data_a[idx]), float(data_b[idx])));
+    }
+}
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp
index 26163b167..888ce79f6 100644
--- a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp
@@ -500,10 +500,9 @@ void main() {
             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 uint ib = idx / 32;                  // 8 values per idx
+            const uint ib32 = (idx % 32) / 4;         // 0..7
+            const uint ib8 = idx % 32;
 
             const float d = float(data_a[ib].d);
             const uint qh = data_a[ib].qh[ib32];
@@ -512,22 +511,16 @@ void main() {
             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);
+            [[unroll]] for (int k = 0; k < 8; ++k) {
+                buf_a[buf_idx + k] = FLOAT_TYPE(dl * (bitfieldExtract(grid, 2 * k, 2) + delta));
+            }
 #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 ib = idx / 32;  // 8 values per idx
+            const uint ib8 = idx % 32;
             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;
@@ -538,21 +531,17 @@ void main() {
             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);
+            [[unroll]] for (int k = 0; k < 8; ++k) {
+                buf_a[buf_idx + k] = FLOAT_TYPE(dl * (bitfieldExtract(grid, 2 * k, 2) + delta));
+            }
 #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;
 
-            const uint ib = idx / 128;                  // 2 values per idx
-            const uint ib32 = (idx % 128) / 16;         // 0..7
-            const uint ib8 = (idx / 4) % 4;
+            const uint ib = idx / 32;                 // 8 values per idx
+            const uint ib32 = (idx % 32) / 4;         // 0..7
+            const uint ib8 = idx % 4;
 
             const float d = float(data_a[ib].d);
             const uint qs = data_a[ib].qs[8 * ib32 + ib8];
@@ -562,63 +551,81 @@ void main() {
                 data_a[ib].qs[8*ib32 + 6],
                 data_a[ib].qs[8*ib32 + 7]
             ));
-            const float db = d * 0.25 * (0.5 + (signs >> 28));
+            const FLOAT_TYPE db = FLOAT_TYPE(d * 0.25 * (0.5 + (signs >> 28)));
             const uint32_t sign7 = bitfieldExtract(signs, 7 * int(ib8), 7);
-            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); // vec4 used due to #12147
+            const uint sign = sign7 | (bitCount(sign7) << 7);
+            const uvec2 grid = iq2xxs_grid[qs];
+            const vec4 grid0 = vec4(unpack8(grid.x));
+            const vec4 grid1 = vec4(unpack8(grid.y));
 
-            buf_a[buf_idx    ] = FLOAT_TYPE(v.x);
-            buf_a[buf_idx + 1] = FLOAT_TYPE(v.y);
+            buf_a[buf_idx    ] = db * FLOAT_TYPE((sign &   1) != 0 ? -grid0.x : grid0.x);
+            buf_a[buf_idx + 1] = db * FLOAT_TYPE((sign &   2) != 0 ? -grid0.y : grid0.y);
+            buf_a[buf_idx + 2] = db * FLOAT_TYPE((sign &   4) != 0 ? -grid0.z : grid0.z);
+            buf_a[buf_idx + 3] = db * FLOAT_TYPE((sign &   8) != 0 ? -grid0.w : grid0.w);
+            buf_a[buf_idx + 4] = db * FLOAT_TYPE((sign &  16) != 0 ? -grid1.x : grid1.x);
+            buf_a[buf_idx + 5] = db * FLOAT_TYPE((sign &  32) != 0 ? -grid1.y : grid1.y);
+            buf_a[buf_idx + 6] = db * FLOAT_TYPE((sign &  64) != 0 ? -grid1.z : grid1.z);
+            buf_a[buf_idx + 7] = db * FLOAT_TYPE((sign & 128) != 0 ? -grid1.w : grid1.w);
 #elif defined(DATA_A_IQ2_XS)
             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 / 4) % 4;             // 0..3
+            const uint ib = idx / 32;            // 8 values per idx
+            const uint ib32 = (idx % 32) / 4;    // 0..7
+            const uint ib8 = idx % 4;            // 0..3
 
             const float d = float(data_a[ib].d);
             const uint scale = (data_a[ib].scales[ib32] >> (2 * (ib8 & 2))) & 0xf;
-            const float db = d * 0.25 * (0.5 + scale);
+            const FLOAT_TYPE db = FLOAT_TYPE(d * 0.25 * (0.5 + scale));
             const uint qs = data_a[ib].qs[4 * ib32 + ib8];
             const uint sign7 = qs >> 9;
-            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); // vec4 used due to #12147
+            const uint sign = sign7 | (bitCount(sign7) << 7);
+            const uvec2 grid = iq2xs_grid[qs & 511];
+            const vec4 grid0 = vec4(unpack8(grid.x));
+            const vec4 grid1 = vec4(unpack8(grid.y));
 
-            buf_a[buf_idx    ] = FLOAT_TYPE(v.x);
-            buf_a[buf_idx + 1] = FLOAT_TYPE(v.y);
+            buf_a[buf_idx    ] = db * FLOAT_TYPE((sign &   1) != 0 ? -grid0.x : grid0.x);
+            buf_a[buf_idx + 1] = db * FLOAT_TYPE((sign &   2) != 0 ? -grid0.y : grid0.y);
+            buf_a[buf_idx + 2] = db * FLOAT_TYPE((sign &   4) != 0 ? -grid0.z : grid0.z);
+            buf_a[buf_idx + 3] = db * FLOAT_TYPE((sign &   8) != 0 ? -grid0.w : grid0.w);
+            buf_a[buf_idx + 4] = db * FLOAT_TYPE((sign &  16) != 0 ? -grid1.x : grid1.x);
+            buf_a[buf_idx + 5] = db * FLOAT_TYPE((sign &  32) != 0 ? -grid1.y : grid1.y);
+            buf_a[buf_idx + 6] = db * FLOAT_TYPE((sign &  64) != 0 ? -grid1.z : grid1.z);
+            buf_a[buf_idx + 7] = db * FLOAT_TYPE((sign & 128) != 0 ? -grid1.w : grid1.w);
 #elif defined(DATA_A_IQ2_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 ib8 = (idx % 128) / 4; // 0..31
-            const uint ib32 = ib8 / 4;        // 0..7
+            const uint ib = idx / 32;  // 8 values per idx
+            const uint ib8 = idx % 32; // 0..31
+            const uint ib32 = ib8 / 4; // 0..7
 
             const uint scale = (data_a[ib].scales[ib32] >> (2 * (ib8 & 2))) & 0xf;
             const uint qs = data_a[ib].qs[ib8];
             const uint qh = data_a[ib].qh[ib32];
             const uint qhshift = 2 * (ib8 % 4);
-            const uint sign = data_a[ib].qs[QUANT_K / 8 + ib8] >> (2 * (idx % 4));
+            const uint sign = data_a[ib].qs[QUANT_K / 8 + ib8];
 
             const float d = float(data_a[ib].d);
-            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(uint32_t(grid)).xy); // vec4 used due to #12147
+            const FLOAT_TYPE db = FLOAT_TYPE(d * 0.25 * (0.5 + scale));
+            const uvec2 grid = iq2s_grid[qs | ((qh << (8 - qhshift)) & 0x300)];
+            const vec4 grid0 = vec4(unpack8(grid.x));
+            const vec4 grid1 = vec4(unpack8(grid.y));
 
-            buf_a[buf_idx    ] = FLOAT_TYPE(v.x);
-            buf_a[buf_idx + 1] = FLOAT_TYPE(v.y);
+            buf_a[buf_idx    ] = db * FLOAT_TYPE((sign &   1) != 0 ? -grid0.x : grid0.x);
+            buf_a[buf_idx + 1] = db * FLOAT_TYPE((sign &   2) != 0 ? -grid0.y : grid0.y);
+            buf_a[buf_idx + 2] = db * FLOAT_TYPE((sign &   4) != 0 ? -grid0.z : grid0.z);
+            buf_a[buf_idx + 3] = db * FLOAT_TYPE((sign &   8) != 0 ? -grid0.w : grid0.w);
+            buf_a[buf_idx + 4] = db * FLOAT_TYPE((sign &  16) != 0 ? -grid1.x : grid1.x);
+            buf_a[buf_idx + 5] = db * FLOAT_TYPE((sign &  32) != 0 ? -grid1.y : grid1.y);
+            buf_a[buf_idx + 6] = db * FLOAT_TYPE((sign &  64) != 0 ? -grid1.z : grid1.z);
+            buf_a[buf_idx + 7] = db * FLOAT_TYPE((sign & 128) != 0 ? -grid1.w : grid1.w);
 #elif defined(DATA_A_IQ3_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;
 
-            const uint ib = idx / 128;                  // 2 values per idx
-            const uint iqs = (idx % 128) / 2;           // 0..63
+            const uint ib = idx / 64;            // 4 values per idx
+            const uint iqs = idx % 64;           // 0..63
             const uint is = QUANT_K / 4 + 4 * (iqs / 8); // 8 values
 
             const float d = float(data_a[ib].d);
@@ -631,33 +638,36 @@ void main() {
             ));
             const float db = d * 0.5 * (0.5 + (signs >> 28));
             const uint32_t sign7 = bitfieldExtract(signs, 7 * (int(iqs / 2) % 4), 7);
-            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); // vec4 used due to #12147
+            const uint sign = (sign7 | (bitCount(sign7) << 7)) >> (4 * (idx % 2));
+            const uint grid = iq3xxs_grid[qs];
+            const vec4 v = db * vec4(unpack8(grid));
 
-            buf_a[buf_idx    ] = FLOAT_TYPE(v.x);
-            buf_a[buf_idx + 1] = FLOAT_TYPE(v.y);
+            buf_a[buf_idx    ] = FLOAT_TYPE((sign &   1) != 0 ? -v.x : v.x);
+            buf_a[buf_idx + 1] = FLOAT_TYPE((sign &   2) != 0 ? -v.y : v.y);
+            buf_a[buf_idx + 2] = FLOAT_TYPE((sign &   4) != 0 ? -v.z : v.z);
+            buf_a[buf_idx + 3] = FLOAT_TYPE((sign &   8) != 0 ? -v.w : v.w);
 #elif defined(DATA_A_IQ3_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 iqs = (idx % 128) / 2;           // 0..63
+            const uint ib = idx / 64;            // 4 values per idx
+            const uint iqs = idx % 64;           // 0..63
             const uint iqh = iqs / 8;
 
             const float d = float(data_a[ib].d);
             const uint qs = data_a[ib].qs[iqs];
             const uint qh = data_a[ib].qh[iqh];
-            const int8_t sign = int8_t(data_a[ib].signs[iqs / 2] >> (2 * (idx % 4)));
+            const int8_t sign = int8_t(data_a[ib].signs[iqs / 2] >> (4 * (idx % 2)));
             const uint scale = data_a[ib].scales[iqs / 16];
             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); // vec4 used due to #12147
+            const uint32_t grid = iq3s_grid[qs | ((qh << (8 - (iqs % 8))) & 256)];
+            const vec4 v = db * vec4(unpack8(grid));
 
-            buf_a[buf_idx    ] = FLOAT_TYPE(v.x);
-            buf_a[buf_idx + 1] = FLOAT_TYPE(v.y);
+            buf_a[buf_idx    ] = FLOAT_TYPE((sign &   1) != 0 ? -v.x : v.x);
+            buf_a[buf_idx + 1] = FLOAT_TYPE((sign &   2) != 0 ? -v.y : v.y);
+            buf_a[buf_idx + 2] = FLOAT_TYPE((sign &   4) != 0 ? -v.z : v.z);
+            buf_a[buf_idx + 3] = FLOAT_TYPE((sign &   8) != 0 ? -v.w : v.w);
 #elif defined(DATA_A_IQ4_XS)
             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;
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/reglu.comp b/ggml/src/ggml-vulkan/vulkan-shaders/reglu.comp
new file mode 100644
index 000000000..0073d8f76
--- /dev/null
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/reglu.comp
@@ -0,0 +1,9 @@
+#version 450
+
+#include "glu_head.comp"
+
+float op(float a, float b) {
+    return max(a, 0.0f) * b;
+}
+
+#include "glu_main.comp"
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/rms_norm.comp b/ggml/src/ggml-vulkan/vulkan-shaders/rms_norm.comp
index deb8ee996..6428ca7ba 100644
--- a/ggml/src/ggml-vulkan/vulkan-shaders/rms_norm.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/rms_norm.comp
@@ -1,11 +1,13 @@
 #version 450
 
-#include "generic_unary_head.comp"
+#include "generic_binary_head.comp"
 #include "types.comp"
 
 #extension GL_EXT_control_flow_attributes : enable
 #define BLOCK_SIZE 512
 
+layout (constant_id = 1) const bool do_multiply = false;
+
 layout(local_size_x = BLOCK_SIZE, local_size_y = 1, local_size_z = 1) in;
 
 shared FLOAT_TYPE sum[BLOCK_SIZE];
@@ -25,6 +27,7 @@ void main() {
     const uint stride_sample    = p.nb03;
 
     uint32_t a_offset = samp*stride_sample + channel*stride_channel + row*stride_row + get_aoffset();
+    uint32_t b_offset = src1_idx(0, row, channel, samp) + get_boffset();
     uint32_t d_offset = ((samp*nchannels + channel)*nrows + row)*ncols + get_doffset();
 
     sum[tid] = FLOAT_TYPE(0.0f); // partial sum for thread in warp
@@ -46,7 +49,13 @@ void main() {
     const FLOAT_TYPE mean = sum[0] / FLOAT_TYPE(ncols);
     const FLOAT_TYPE scale = inversesqrt(mean + FLOAT_TYPE(p.param1));
 
-    [[unroll]] for (uint col = tid; col < ncols; col += BLOCK_SIZE) {
-        data_d[d_offset + col] = D_TYPE(scale * FLOAT_TYPE(data_a[a_offset + col]));
+    if (do_multiply) {
+        [[unroll]] for (uint col = tid; col < ncols; col += BLOCK_SIZE) {
+            data_d[d_offset + col] = D_TYPE(scale * FLOAT_TYPE(data_a[a_offset + col]) * FLOAT_TYPE(data_b[b_offset + col]));
+        }
+    } else {
+        [[unroll]] for (uint col = tid; col < ncols; col += BLOCK_SIZE) {
+            data_d[d_offset + col] = D_TYPE(scale * FLOAT_TYPE(data_a[a_offset + col]));
+        }
     }
 }
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/soft_max.comp b/ggml/src/ggml-vulkan/vulkan-shaders/soft_max.comp
index 51fc2dc7e..5bcd3b1e3 100644
--- a/ggml/src/ggml-vulkan/vulkan-shaders/soft_max.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/soft_max.comp
@@ -6,6 +6,14 @@ layout (push_constant) uniform parameter
 {
     uint KX;
     uint KY;
+    uint ne00;
+    uint ne01;
+    uint ne02;
+    uint ne12;
+    uint ne13;
+    uint nb11;
+    uint nb12;
+    uint nb13;
     float scale;
     float max_bias;
     float m0;
@@ -31,7 +39,15 @@ shared FLOAT_TYPE vals[BLOCK_SIZE];
 void soft_max(uint num_iters) {
     const uint tid = gl_LocalInvocationID.x;
     const uint rowx = gl_WorkGroupID.z * 262144 + gl_WorkGroupID.y * 512 + gl_WorkGroupID.x;
-    const uint rowy = (p.KY > 0) ? (rowx % p.KY) : 0;
+
+    const uint32_t i03 = rowx / (p.ne01 * p.ne02);
+    const uint32_t i02 = (rowx - i03 * p.ne01 * p.ne02) / p.ne01;
+    const uint32_t i01 = rowx % p.ne01;
+
+    uint rowy_start = 0;
+    if (p.KY > 0) {
+        rowy_start = i01 * p.nb11 + (i02 % p.ne12) * p.nb12 + (i03 % p.ne13) * p.nb13;
+    }
 
     if (rowx >= p.nrows_x) {
         return;
@@ -41,7 +57,7 @@ void soft_max(uint num_iters) {
 
     // ALiBi
     if (p.max_bias > 0.0f) {
-        const uint h = rowx/p.KY; // head index
+        const uint h = (rowx / p.ne01) % p.ne02; // head index
 
         const float base = h < p.n_head_log2 ? p.m0 : p.m1;
         const uint   exp  = h < p.n_head_log2 ? h + 1 : 2*(h - p.n_head_log2) + 1;
@@ -67,7 +83,7 @@ void soft_max(uint num_iters) {
 
         FLOAT_TYPE b = FLOAT_TYPE(0);
         if (p.KY > 0 && col < p.KX) {
-            b = data_b[rowy * p.KX + col];
+            b = data_b[rowy_start + col];
         }
 
         FLOAT_TYPE v = a * p.scale + slope * b;
@@ -111,7 +127,7 @@ void soft_max(uint num_iters) {
         if (idx < DATA_CACHE_SIZE) {
             val = exp(data_cache[idx] - max_val);
         } else {
-            val = exp(FLOAT_TYPE(data_a[i]) * p.scale + (p.KY > 0 ? slope * FLOAT_TYPE(data_b[rowy * p.KX + col]) : FLOAT_TYPE(0.0f)) - max_val);
+            val = exp(FLOAT_TYPE(data_a[i]) * p.scale + (p.KY > 0 ? slope * FLOAT_TYPE(data_b[rowy_start + col]) : FLOAT_TYPE(0.0f)) - max_val);
         }
         sum += val;
         if (idx < DATA_CACHE_SIZE) {
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/swiglu.comp b/ggml/src/ggml-vulkan/vulkan-shaders/swiglu.comp
new file mode 100644
index 000000000..a28e7c6cc
--- /dev/null
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/swiglu.comp
@@ -0,0 +1,9 @@
+#version 450
+
+#include "glu_head.comp"
+
+float op(float a, float b) {
+    return a / (1.0f + exp(-a)) * b;
+}
+
+#include "glu_main.comp"
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 9361e2ac8..30f78fabb 100644
--- a/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp
@@ -360,9 +360,9 @@ void matmul_shaders(bool fp16, bool matmul_id, bool coopmat, bool coopmat2, bool
 
     for (const auto& tname : type_names) {
         std::string load_vec_quant = "2";
-        if ((tname == "q4_0") || (tname == "q4_1"))
+        if ((tname == "q4_0") || (tname == "q4_1") || (tname == "iq1_s") || (tname == "iq1_m") || (tname == "iq2_xxs") || (tname == "iq2_xs") || (tname == "iq2_s"))
             load_vec_quant = "8";
-        else if ((tname == "q5_0") || (tname == "q5_1") || (tname == "q8_0") || (tname == "iq4_nl"))
+        else if ((tname == "q5_0") || (tname == "q5_1") || (tname == "q8_0") || (tname == "iq3_xxs") || (tname == "iq3_s") || (tname == "iq4_nl"))
             load_vec_quant = "4";
 
         if (tname == "bf16") {
@@ -497,7 +497,7 @@ void process_shaders() {
     // 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_f32", "rms_norm.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_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"}}));
 
@@ -574,6 +574,8 @@ void process_shaders() {
 
     string_to_spv("gelu_f16",       "gelu.comp",        {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"}});
     string_to_spv("gelu_f32",       "gelu.comp",        {{"A_TYPE", "float"},       {"D_TYPE", "float"}});
+    string_to_spv("gelu_erf_f16",   "gelu_erf.comp",    {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"}});
+    string_to_spv("gelu_erf_f32",   "gelu_erf.comp",    {{"A_TYPE", "float"},       {"D_TYPE", "float"}});
     string_to_spv("gelu_quick_f16", "gelu_quick.comp",  {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"}});
     string_to_spv("gelu_quick_f32", "gelu_quick.comp",  {{"A_TYPE", "float"},       {"D_TYPE", "float"}});
     string_to_spv("silu_f16",       "silu.comp",        {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"}});
@@ -585,6 +587,17 @@ void process_shaders() {
     string_to_spv("sigmoid_f16",    "sigmoid.comp",     {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"}});
     string_to_spv("sigmoid_f32",    "sigmoid.comp",     {{"A_TYPE", "float"},       {"D_TYPE", "float"}});
 
+    string_to_spv("geglu_f16",      "geglu.comp",       {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"}});
+    string_to_spv("geglu_f32",      "geglu.comp",       {{"A_TYPE", "float"},       {"D_TYPE", "float"}});
+    string_to_spv("reglu_f16",      "reglu.comp",       {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"}});
+    string_to_spv("reglu_f32",      "reglu.comp",       {{"A_TYPE", "float"},       {"D_TYPE", "float"}});
+    string_to_spv("swiglu_f16",     "swiglu.comp",      {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"}});
+    string_to_spv("swiglu_f32",     "swiglu.comp",      {{"A_TYPE", "float"},       {"D_TYPE", "float"}});
+    string_to_spv("geglu_erf_f16",  "geglu_erf.comp",   {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"}});
+    string_to_spv("geglu_erf_f32",  "geglu_erf.comp",   {{"A_TYPE", "float"},       {"D_TYPE", "float"}});
+    string_to_spv("geglu_quick_f16","geglu_quick.comp", {{"A_TYPE", "float16_t"},   {"D_TYPE", "float16_t"}});
+    string_to_spv("geglu_quick_f32","geglu_quick.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("silu_back_f32",  "silu_back.comp",   {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}});
 
@@ -622,6 +635,8 @@ void process_shaders() {
 
     string_to_spv("timestep_embedding_f32", "timestep_embedding.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
 
+    string_to_spv("conv_transpose_1d_f32", "conv_transpose_1d.comp", {{"A_TYPE", "float"},  {"B_TYPE", "float"}, {"D_TYPE", "float"}});
+
     string_to_spv("pool2d_f32", "pool2d.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
 
     string_to_spv("rwkv_wkv6_f32", "wkv6.comp", merge_maps(base_dict, {{"A_TYPE", "float"}}));
diff --git a/ggml/src/ggml.c b/ggml/src/ggml.c
index 196b7b8f3..75fc1e707 100644
--- a/ggml/src/ggml.c
+++ b/ggml/src/ggml.c
@@ -61,9 +61,6 @@
 #define m512i(p) (__m512i)(p)
 #endif
 
-// precomputed f32 table for f16 (256 KB) (ggml-impl.h)
-float ggml_table_f32_f16[1 << 16];
-
 #if defined(__linux__) || \
     defined(__FreeBSD__) || defined(__NetBSD__) || defined(__OpenBSD__) || \
     (defined(__APPLE__) && !TARGET_OS_TV && !TARGET_OS_WATCH)
@@ -205,19 +202,34 @@ void ggml_print_backtrace(void) {
 }
 #endif
 
+static ggml_abort_callback_t g_abort_callback = NULL;
+
+// Set the abort callback (passing null will restore original abort functionality: printing a message to stdout)
+GGML_API ggml_abort_callback_t ggml_set_abort_callback(ggml_abort_callback_t callback) {
+    ggml_abort_callback_t ret_val = g_abort_callback;
+    g_abort_callback = callback;
+    return ret_val;
+}
+
 void ggml_abort(const char * file, int line, const char * fmt, ...) {
     fflush(stdout);
 
-    fprintf(stderr, "%s:%d: ", file, line);
+    char message[2048];
+    int offset = snprintf(message, sizeof(message), "%s:%d: ", file, line);
 
     va_list args;
     va_start(args, fmt);
-    vfprintf(stderr, fmt, args);
+    vsnprintf(message + offset, sizeof(message) - offset, fmt, args);
     va_end(args);
 
-    fprintf(stderr, "\n");
+    if (g_abort_callback) {
+        g_abort_callback(message);
+    } else {
+        // default: print error and backtrace to stderr
+        fprintf(stderr, "%s\n", message);
+        ggml_print_backtrace();
+    }
 
-    ggml_print_backtrace();
     abort();
 }
 
@@ -461,6 +473,14 @@ bool ggml_guid_matches(ggml_guid_t guid_a, ggml_guid_t guid_b) {
     return memcmp(guid_a, guid_b, sizeof(ggml_guid)) == 0;
 }
 
+const char * ggml_version(void) {
+    return GGML_VERSION;
+}
+
+const char * ggml_commit(void) {
+    return GGML_COMMIT;
+}
+
 //
 // timing
 //
@@ -888,12 +908,6 @@ struct ggml_context {
     struct ggml_object * objects_end;
 };
 
-struct ggml_context_container {
-    bool used;
-
-    struct ggml_context context;
-};
-
 //
 // data types
 //
@@ -942,6 +956,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
     "TRANSPOSE",
     "GET_ROWS",
     "GET_ROWS_BACK",
+    "SET_ROWS",
     "DIAG",
     "DIAG_MASK_INF",
     "DIAG_MASK_ZERO",
@@ -953,6 +968,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
     "CONV_TRANSPOSE_1D",
     "IM2COL",
     "IM2COL_BACK",
+    "CONV_2D",
     "CONV_2D_DW",
     "CONV_TRANSPOSE_2D",
     "POOL_1D",
@@ -961,6 +977,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
     "UPSCALE",
     "PAD",
     "PAD_REFLECT_1D",
+    "ROLL",
     "ARANGE",
     "TIMESTEP_EMBEDDING",
     "ARGSORT",
@@ -989,9 +1006,11 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
     "CROSS_ENTROPY_LOSS",
     "CROSS_ENTROPY_LOSS_BACK",
     "OPT_STEP_ADAMW",
+
+    "GLU",
 };
 
-static_assert(GGML_OP_COUNT == 82, "GGML_OP_COUNT != 82");
+static_assert(GGML_OP_COUNT == 86, "GGML_OP_COUNT != 86");
 
 static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
     "none",
@@ -1037,6 +1056,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
     "transpose(x)",
     "get_rows(x)",
     "get_rows_back(x)",
+    "set_rows(x)",
     "diag(x)",
     "diag_mask_inf(x)",
     "diag_mask_zero(x)",
@@ -1048,6 +1068,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
     "conv_transpose_1d(x)",
     "im2col(x)",
     "im2col_back(x)",
+    "conv_2d(x)",
     "conv_2d_dw(x)",
     "conv_transpose_2d(x)",
     "pool_1d(x)",
@@ -1056,6 +1077,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
     "upscale(x)",
     "pad(x)",
     "pad_reflect_1d(x)",
+    "roll(x)",
     "arange(start, stop, step)",
     "timestep_embedding(timesteps, dim, max_period)",
     "argsort(x)",
@@ -1084,9 +1106,11 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
     "cross_entropy_loss(x,y)",
     "cross_entropy_loss_back(x,y)",
     "adamw(x)",
+
+    "glu(x)",
 };
 
-static_assert(GGML_OP_COUNT == 82, "GGML_OP_COUNT != 82");
+static_assert(GGML_OP_COUNT == 86, "GGML_OP_COUNT != 86");
 
 static_assert(GGML_OP_POOL_COUNT == 2, "GGML_OP_POOL_COUNT != 2");
 
@@ -1112,6 +1136,17 @@ static const char * GGML_UNARY_OP_NAME[GGML_UNARY_OP_COUNT] = {
 static_assert(GGML_UNARY_OP_COUNT == 15, "GGML_UNARY_OP_COUNT != 15");
 
 
+static const char * GGML_GLU_OP_NAME[GGML_GLU_OP_COUNT] = {
+    "REGLU",
+    "GEGLU",
+    "SWIGLU",
+    "GEGLU_ERF",
+    "GEGLU_QUICK",
+};
+
+static_assert(GGML_GLU_OP_COUNT == 5, "GGML_GLU_OP_COUNT != 5");
+
+
 static_assert(sizeof(struct ggml_object)%GGML_MEM_ALIGN == 0, "ggml_object size must be a multiple of GGML_MEM_ALIGN");
 static_assert(sizeof(struct ggml_tensor)%GGML_MEM_ALIGN == 0, "ggml_tensor size must be a multiple of GGML_MEM_ALIGN");
 
@@ -1214,11 +1249,19 @@ const char * ggml_unary_op_name(enum ggml_unary_op op) {
     return GGML_UNARY_OP_NAME[op];
 }
 
+const char * ggml_glu_op_name(enum ggml_glu_op op) {
+    return GGML_GLU_OP_NAME[op];
+}
+
 const char * ggml_op_desc(const struct ggml_tensor * t) {
     if (t->op == GGML_OP_UNARY) {
         enum ggml_unary_op uop = ggml_get_unary_op(t);
         return ggml_unary_op_name(uop);
     }
+    if (t->op == GGML_OP_GLU) {
+        enum ggml_glu_op gop = ggml_get_glu_op(t);
+        return ggml_glu_op_name(gop);
+    }
     return ggml_op_name(t->op);
 }
 
@@ -1355,6 +1398,12 @@ bool ggml_is_contiguous_channels(const struct ggml_tensor * tensor) {
         tensor->nb[2] == ggml_type_size(tensor->type);
 }
 
+bool ggml_is_contiguous_rows(const struct ggml_tensor * tensor) {
+    return
+        tensor->ne[0] == ggml_blck_size(tensor->type) ||
+        tensor->nb[0] == ggml_type_size(tensor->type);
+}
+
 static inline bool ggml_is_padded_1d(const struct ggml_tensor * tensor) {
     static_assert(GGML_MAX_DIMS == 4, "GGML_MAX_DIMS is not 4 - update this function");
 
@@ -1426,14 +1475,6 @@ struct ggml_context * ggml_init(struct ggml_init_params params) {
         // initialize time system (required on Windows)
         ggml_time_init();
 
-        for (int i = 0; i < (1 << 16); ++i) {
-            union {
-                uint16_t u16;
-                ggml_fp16_t fp16;
-            } u = {i};
-            ggml_table_f32_f16[i] = GGML_COMPUTE_FP16_TO_FP32(u.fp16);
-        }
-
         is_first_call = false;
     }
 
@@ -1737,6 +1778,11 @@ enum ggml_unary_op ggml_get_unary_op(const struct ggml_tensor * tensor) {
     return (enum ggml_unary_op) ggml_get_op_params_i32(tensor, 0);
 }
 
+enum ggml_glu_op ggml_get_glu_op(const struct ggml_tensor * tensor) {
+    GGML_ASSERT(tensor->op == GGML_OP_GLU);
+    return (enum ggml_glu_op) ggml_get_op_params_i32(tensor, 0);
+}
+
 const char * ggml_get_name(const struct ggml_tensor * tensor) {
     return tensor->name;
 }
@@ -2616,6 +2662,156 @@ struct ggml_tensor * ggml_exp_inplace(
     return ggml_unary_inplace(ctx, a, GGML_UNARY_OP_EXP);
 }
 
+// ggml_glu
+
+static struct ggml_tensor * ggml_glu_impl(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        struct ggml_tensor  * b,
+        enum ggml_glu_op      op,
+        bool                  swapped) {
+    GGML_ASSERT(ggml_is_contiguous_1(a));
+
+    if (b) {
+        GGML_ASSERT(ggml_is_contiguous_1(b));
+        GGML_ASSERT(ggml_are_same_shape(a, b));
+        GGML_ASSERT(a->type == b->type);
+    }
+
+    int64_t ne[GGML_MAX_DIMS] = { a->ne[0] / 2 }; for (int i = 1; i < GGML_MAX_DIMS; i++) ne[i] = a->ne[i];
+    struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, GGML_MAX_DIMS, b ? a->ne : ne, NULL, 0);
+
+    ggml_set_op_params_i32(result, 0, (int32_t) op);
+    ggml_set_op_params_i32(result, 1, (int32_t) swapped);
+
+    result->op     = GGML_OP_GLU;
+    result->src[0] = a;
+    result->src[1] = b;
+
+    return result;
+}
+
+struct ggml_tensor * ggml_glu(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        enum ggml_glu_op      op,
+        bool                  swapped) {
+    return ggml_glu_impl(ctx, a, NULL, op, swapped);
+}
+
+struct ggml_tensor * ggml_glu_split(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        struct ggml_tensor  * b,
+        enum ggml_glu_op      op) {
+    return ggml_glu_impl(ctx, a, b, op, false);
+}
+
+// ggml_reglu
+
+struct ggml_tensor * ggml_reglu(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a) {
+    return ggml_glu_impl(ctx, a, NULL, GGML_GLU_OP_REGLU, false);
+}
+
+struct ggml_tensor * ggml_reglu_swapped(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a) {
+    return ggml_glu_impl(ctx, a, NULL, GGML_GLU_OP_REGLU, true);
+}
+
+struct ggml_tensor * ggml_reglu_split(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        struct ggml_tensor  * b) {
+    return ggml_glu_impl(ctx, a, b, GGML_GLU_OP_REGLU, false);
+}
+
+// ggml_geglu
+
+struct ggml_tensor * ggml_geglu(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a) {
+    return ggml_glu_impl(ctx, a, NULL, GGML_GLU_OP_GEGLU, false);
+}
+
+struct ggml_tensor * ggml_geglu_swapped(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a) {
+    return ggml_glu_impl(ctx, a, NULL, GGML_GLU_OP_GEGLU, true);
+}
+
+struct ggml_tensor * ggml_geglu_split(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        struct ggml_tensor  * b) {
+    return ggml_glu_impl(ctx, a, b, GGML_GLU_OP_GEGLU, false);
+}
+
+// ggml_swiglu
+
+struct ggml_tensor * ggml_swiglu(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a) {
+    return ggml_glu_impl(ctx, a, NULL, GGML_GLU_OP_SWIGLU, false);
+}
+
+struct ggml_tensor * ggml_swiglu_swapped(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a) {
+    return ggml_glu_impl(ctx, a, NULL, GGML_GLU_OP_SWIGLU, true);
+}
+
+struct ggml_tensor * ggml_swiglu_split(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        struct ggml_tensor  * b) {
+    return ggml_glu_impl(ctx, a, b, GGML_GLU_OP_SWIGLU, false);
+}
+
+// ggml_geglu_erf
+
+struct ggml_tensor * ggml_geglu_erf(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a) {
+    return ggml_glu_impl(ctx, a, NULL, GGML_GLU_OP_GEGLU_ERF, false);
+}
+
+struct ggml_tensor * ggml_geglu_erf_swapped(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a) {
+    return ggml_glu_impl(ctx, a, NULL, GGML_GLU_OP_GEGLU_ERF, true);
+}
+
+struct ggml_tensor * ggml_geglu_erf_split(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        struct ggml_tensor  * b) {
+    return ggml_glu_impl(ctx, a, b, GGML_GLU_OP_GEGLU_ERF, false);
+}
+
+// ggml_geglu_quick
+
+struct ggml_tensor * ggml_geglu_quick(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a) {
+    return ggml_glu_impl(ctx, a, NULL, GGML_GLU_OP_GEGLU_QUICK, false);
+}
+
+struct ggml_tensor * ggml_geglu_quick_swapped(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a) {
+    return ggml_glu_impl(ctx, a, NULL, GGML_GLU_OP_GEGLU_QUICK, true);
+}
+
+struct ggml_tensor * ggml_geglu_quick_split(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        struct ggml_tensor  * b) {
+    return ggml_glu_impl(ctx, a, b, GGML_GLU_OP_GEGLU_QUICK, false);
+}
+
 // ggml_norm
 
 static struct ggml_tensor * ggml_norm_impl(
@@ -3399,6 +3595,35 @@ struct ggml_tensor * ggml_get_rows_back(
     return result;
 }
 
+// ggml_set_rows
+
+struct ggml_tensor * ggml_set_rows(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        struct ggml_tensor  * b,
+        struct ggml_tensor  * c) {
+    GGML_ASSERT(a->ne[0] == b->ne[0]);
+    GGML_ASSERT(a->ne[2] == b->ne[2]);
+    GGML_ASSERT(a->ne[3] == b->ne[3]);
+    GGML_ASSERT(b->ne[1] == c->ne[0]);
+    GGML_ASSERT(b->ne[2] % c->ne[1] == 0);
+    GGML_ASSERT(b->ne[3] % c->ne[2] == 0);
+    GGML_ASSERT(c->ne[3] == 1);
+    GGML_ASSERT(b->type == GGML_TYPE_F32);
+    GGML_ASSERT(c->type == GGML_TYPE_I64);
+
+    GGML_ASSERT(ggml_is_contiguous_rows(a));
+    GGML_ASSERT(ggml_is_contiguous_rows(b));
+
+    struct ggml_tensor * result = ggml_view_tensor(ctx, a);
+
+    result->op     = GGML_OP_SET_ROWS;
+    result->src[0] = b;
+    result->src[1] = c;
+
+    return result;
+}
+
 // ggml_diag
 
 struct ggml_tensor * ggml_diag(
@@ -3493,9 +3718,10 @@ static struct ggml_tensor * ggml_soft_max_impl(
     if (mask) {
         GGML_ASSERT(mask->type == GGML_TYPE_F16 || mask->type == GGML_TYPE_F32);
         GGML_ASSERT(ggml_is_contiguous(mask));
-        GGML_ASSERT(ggml_is_matrix(mask));
         GGML_ASSERT(mask->ne[0] == a->ne[0]);
         GGML_ASSERT(mask->ne[1] >= a->ne[1]);
+        GGML_ASSERT(a->ne[2]%mask->ne[2] == 0);
+        GGML_ASSERT(a->ne[3]%mask->ne[3] == 0);
     }
 
     if (max_bias > 0.0f) {
@@ -4135,6 +4361,44 @@ struct ggml_tensor * ggml_conv_2d_dw_direct(
     return result;
 }
 
+// ggml_conv_2d_direct
+
+struct ggml_tensor * ggml_conv_2d_direct(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,   // convolution kernel [KW, KH, IC, OC]
+        struct ggml_tensor  * b,   // input data [W, H, C, N]
+        int                   s0,  // stride dimension 0
+        int                   s1,  // stride dimension 1
+        int                   p0,  // padding dimension 0
+        int                   p1,  // padding dimension 1
+        int                   d0,  // dilation dimension 0
+        int                   d1) {// dilation dimension 1
+
+    GGML_ASSERT(a->ne[2] == b->ne[2]);
+    //GGML_ASSERT(a->type == b->type);
+
+    int64_t ne[4];
+    ne[0] = ggml_calc_conv_output_size(b->ne[0], a->ne[0], s0, p0, d0);
+    ne[1] = ggml_calc_conv_output_size(b->ne[1], a->ne[1], s1, p1, d1);
+    ne[2] = a->ne[3];
+    ne[3] = b->ne[3];
+
+    struct ggml_tensor * result = ggml_new_tensor(ctx, b->type, 4, ne);
+
+    ggml_set_op_params_i32(result, 0, s0);
+    ggml_set_op_params_i32(result, 1, s1);
+    ggml_set_op_params_i32(result, 2, p0);
+    ggml_set_op_params_i32(result, 3, p1);
+    ggml_set_op_params_i32(result, 4, d0);
+    ggml_set_op_params_i32(result, 5, d1);
+
+    result->op = GGML_OP_CONV_2D;
+    result->src[0] = a;
+    result->src[1] = b;
+
+    return result;
+}
+
 // ggml_conv_transpose_2d_p0
 
 static int64_t ggml_calc_conv_transpose_output_size(int64_t ins, int64_t ks, int s, int p) {
@@ -4251,24 +4515,21 @@ struct ggml_tensor * ggml_pool_2d_back(
     return result;
 }
 
-// ggml_upscale
+// ggml_upscale / ggml_interpolate
 
-static struct ggml_tensor * ggml_upscale_impl(
+static struct ggml_tensor * ggml_interpolate_impl(
         struct ggml_context * ctx,
         struct ggml_tensor  * a,
-        int                   ne0,
-        int                   ne1,
-        int                   ne2,
-        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);
-    GGML_ASSERT(a->ne[3] <= ne3);
+        int64_t               ne0,
+        int64_t               ne1,
+        int64_t               ne2,
+        int64_t               ne3,
+        uint32_t              mode) {
+    GGML_ASSERT((mode & 0xFF) < GGML_SCALE_MODE_COUNT);
 
     struct ggml_tensor * result = ggml_new_tensor_4d(ctx, a->type, ne0, ne1, ne2, ne3);
 
-    ggml_set_op_params_i32(result, 0, mode);
+    ggml_set_op_params_i32(result, 0, (int32_t)mode);
 
     result->op     = GGML_OP_UPSCALE;
     result->src[0] = a;
@@ -4281,7 +4542,8 @@ struct ggml_tensor * ggml_upscale(
         struct ggml_tensor  * a,
         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);
+    GGML_ASSERT(scale_factor > 1);
+    return ggml_interpolate_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(
@@ -4292,7 +4554,18 @@ struct ggml_tensor * ggml_upscale_ext(
         int                   ne2,
         int                   ne3,
         enum ggml_scale_mode  mode) {
-    return ggml_upscale_impl(ctx, a, ne0, ne1, ne2, ne3, mode);
+    return ggml_interpolate_impl(ctx, a, ne0, ne1, ne2, ne3, mode);
+}
+
+struct ggml_tensor * ggml_interpolate(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        int64_t               ne0,
+        int64_t               ne1,
+        int64_t               ne2,
+        int64_t               ne3,
+        uint32_t              mode) {
+    return ggml_interpolate_impl(ctx, a, ne0, ne1, ne2, ne3, mode);
 }
 
 // ggml_pad
@@ -4347,6 +4620,34 @@ struct ggml_tensor * ggml_pad_reflect_1d(
     return result;
 }
 
+// ggml_roll
+
+struct ggml_tensor * ggml_roll(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        int                   shift0,
+        int                   shift1,
+        int                   shift2,
+        int                   shift3) {
+    GGML_ASSERT(a->nb[0] == ggml_type_size(a->type));
+    GGML_ASSERT(abs(shift0) < a->ne[0]);
+    GGML_ASSERT(abs(shift1) < a->ne[1]);
+    GGML_ASSERT(abs(shift2) < a->ne[2]);
+    GGML_ASSERT(abs(shift3) < a->ne[3]);
+
+    struct ggml_tensor * result = ggml_dup_tensor(ctx, a);
+
+    ggml_set_op_params_i32(result, 0, shift0);
+    ggml_set_op_params_i32(result, 1, shift1);
+    ggml_set_op_params_i32(result, 2, shift2);
+    ggml_set_op_params_i32(result, 3, shift3);
+
+    result->op     = GGML_OP_ROLL;
+    result->src[0] = a;
+
+    return result;
+}
+
 // ggml_arange
 
 struct ggml_tensor * ggml_arange(
@@ -4441,13 +4742,17 @@ struct ggml_tensor * ggml_flash_attn_ext(
     GGML_ASSERT(ggml_can_mul_mat(k, q));
     // TODO: check if vT can be multiplied by (k*qT)
 
+    GGML_ASSERT(q->ne[3] == k->ne[3]);
+    GGML_ASSERT(q->ne[3] == v->ne[3]);
+
     if (mask) {
         GGML_ASSERT(ggml_is_contiguous(mask));
-        GGML_ASSERT(mask->ne[2] == 1);
-        GGML_ASSERT(mask->ne[3] == 1);
         GGML_ASSERT(mask->ne[1] >= GGML_PAD(q->ne[1], GGML_KQ_MASK_PAD) &&
                 "the Flash-Attention kernel requires the mask to be padded to GGML_KQ_MASK_PAD and at least n_queries big");
         //GGML_ASSERT(ggml_can_repeat_rows(mask, qk));
+
+        GGML_ASSERT(q->ne[2] % mask->ne[2] == 0);
+        GGML_ASSERT(q->ne[3] % mask->ne[3] == 0);
     }
 
     if (max_bias > 0.0f) {
@@ -4575,7 +4880,6 @@ struct ggml_tensor * ggml_ssm_conv(
     const int64_t n_s     = sx->ne[2];
 
     // TODO: maybe support other strides than 1?
-    // FIXME: this is always true?
     GGML_ASSERT(sx->ne[0] == d_conv - 1 + n_t);
     GGML_ASSERT(sx->ne[1] == d_inner);
     GGML_ASSERT(n_t >= 0);
@@ -4598,36 +4902,49 @@ struct ggml_tensor * ggml_ssm_scan(
         struct ggml_tensor  * dt,
         struct ggml_tensor  * A,
         struct ggml_tensor  * B,
-        struct ggml_tensor  * C) {
+        struct ggml_tensor  * C,
+        struct ggml_tensor  * ids) {
     GGML_ASSERT(ggml_is_contiguous(s));
-    GGML_ASSERT(ggml_is_contiguous(x));
     GGML_ASSERT(ggml_is_contiguous(dt));
     GGML_ASSERT(ggml_is_contiguous(A));
-    GGML_ASSERT(ggml_is_matrix(A));
-    GGML_ASSERT(ggml_is_3d(B));
-    GGML_ASSERT(ggml_is_3d(s));
+    GGML_ASSERT(x->nb[0] == ggml_type_size(x->type));
     GGML_ASSERT(B->nb[0] == ggml_type_size(B->type));
     GGML_ASSERT(C->nb[0] == ggml_type_size(C->type));
-    GGML_ASSERT(ggml_are_same_shape(x, dt));
+    GGML_ASSERT(x->nb[1] == x->ne[0]*x->nb[0]);
+    GGML_ASSERT(B->nb[1] == B->ne[0]*B->nb[0]);
+    GGML_ASSERT(C->nb[1] == C->ne[0]*C->nb[0]);
     GGML_ASSERT(ggml_are_same_shape(B, C));
+    GGML_ASSERT(ids->type == GGML_TYPE_I32);
 
     {
         const int64_t d_state      = s->ne[0];
-        const int64_t d_inner      = s->ne[1];
-        const int64_t n_seq_tokens = x->ne[1];
-        const int64_t n_seqs       = x->ne[2];
+        const int64_t head_dim     = x->ne[0];
+        const int64_t n_head       = x->ne[1];
+        const int64_t n_seq_tokens = x->ne[2];
+        const int64_t n_seqs       = x->ne[3];
 
-        GGML_ASSERT(s->ne[2] == n_seqs);
-        GGML_ASSERT(x->ne[0] == d_inner);
-        GGML_ASSERT(A->ne[0] == d_state);
-        GGML_ASSERT(A->ne[1] == d_inner);
+        GGML_ASSERT(dt->ne[0] == n_head);
+        GGML_ASSERT(dt->ne[1] == n_seq_tokens);
+        GGML_ASSERT(dt->ne[2] == n_seqs);
+        GGML_ASSERT(ggml_is_3d(dt));
+        GGML_ASSERT(s->ne[1] == head_dim);
+        GGML_ASSERT(s->ne[2] == n_head);
         GGML_ASSERT(B->ne[0] == d_state);
-        GGML_ASSERT(B->ne[1] == n_seq_tokens);
-        GGML_ASSERT(B->ne[2] == n_seqs);
+        GGML_ASSERT(B->ne[2] == n_seq_tokens);
+        GGML_ASSERT(B->ne[3] == n_seqs);
+        GGML_ASSERT(ids->ne[0] == n_seqs);
+        GGML_ASSERT(ggml_is_vector(ids));
+        GGML_ASSERT(A->ne[1] == n_head);
+        GGML_ASSERT(ggml_is_matrix(A));
+
+        if (A->ne[0] != 1) {
+            // Mamba-1 has more granular decay factors
+            GGML_ASSERT(A->ne[0] == d_state);
+        }
     }
 
     // concatenated y + ssm_states
-    struct ggml_tensor * result = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, ggml_nelements(x) + ggml_nelements(s));
+    struct ggml_tensor * result = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, ggml_nelements(x) + s->ne[0]*s->ne[1]*s->ne[2]*ids->ne[0]);
 
     result->op   = GGML_OP_SSM_SCAN;
     result->src[0] = s;
@@ -4636,6 +4953,7 @@ struct ggml_tensor * ggml_ssm_scan(
     result->src[3] = A;
     result->src[4] = B;
     result->src[5] = C;
+    result->src[6] = ids;
 
     return result;
 }
@@ -5776,13 +6094,28 @@ static void ggml_compute_backward(
             }
             GGML_ASSERT(!src1_needs_grads && "backward pass for labels not implemented");
         } break;
+        case GGML_OP_GLU: {
+            switch (ggml_get_glu_op(tensor)) {
+                case GGML_GLU_OP_SWIGLU: {
+                    if (src0_needs_grads) {
+                        GGML_ASSERT(src1 && "backward pass only implemented for split swiglu");
+                        ggml_add_or_set(ctx, cgraph, isrc0, ggml_silu_back(ctx, ggml_mul(ctx, grad, src1), src0));
+                    }
+                    if (src1_needs_grads) {
+                        ggml_add_or_set(ctx, cgraph, isrc1, ggml_mul(ctx, ggml_silu(ctx, src0), grad));
+                    }
+                } break;
+                default: {
+                    GGML_ABORT("unsupported glu op for backward pass: %s", ggml_glu_op_name(ggml_get_glu_op(tensor)));
+                } //break;
+            }
+        } break;
         case GGML_OP_NONE: {
             // noop
         } break;
         case GGML_OP_COUNT:
         default: {
-            fprintf(stderr, "%s: unsupported ggml op for backward pass: %s\n", __func__, ggml_op_name(tensor->op));
-            GGML_ABORT("fatal error");
+            GGML_ABORT("%s: unsupported ggml op for backward pass: %s\n", __func__, ggml_op_name(tensor->op));
         } //break;
     }
 
@@ -5791,19 +6124,32 @@ static void ggml_compute_backward(
     GGML_ASSERT(!src2_needs_grads || ggml_are_same_shape(src2, cgraph->grads[isrc2]));
 }
 
-static void ggml_visit_parents(struct ggml_cgraph * cgraph, struct ggml_tensor * node) {
+static size_t ggml_visit_parents(struct ggml_cgraph * cgraph, struct ggml_tensor * node) {
     // check if already visited
-    if (ggml_hash_insert(&cgraph->visited_hash_set, node) == GGML_HASHSET_ALREADY_EXISTS) {
-        return;
+    size_t node_hash_pos = ggml_hash_find(&cgraph->visited_hash_set, node);
+    GGML_ASSERT(node_hash_pos != GGML_HASHSET_FULL);
+    if (!ggml_bitset_get(cgraph->visited_hash_set.used, node_hash_pos)) {
+        // This is the first time we see this node in the current graph.
+        cgraph->visited_hash_set.keys[node_hash_pos] = node;
+        ggml_bitset_set(cgraph->visited_hash_set.used, node_hash_pos);
+        cgraph->use_counts[node_hash_pos] = 0;
+    } else {
+        // already visited
+        return node_hash_pos;
     }
 
     for (int i = 0; i < GGML_MAX_SRC; ++i) {
         const int k =
             (cgraph->order == GGML_CGRAPH_EVAL_ORDER_LEFT_TO_RIGHT) ? i :
             (cgraph->order == GGML_CGRAPH_EVAL_ORDER_RIGHT_TO_LEFT) ? (GGML_MAX_SRC-1-i) :
-            /* unknown order, just fall back to using i*/ i;
-        if (node->src[k]) {
-            ggml_visit_parents(cgraph, node->src[k]);
+            /* unknown order, just fall back to using i */ i;
+
+        struct ggml_tensor * src = node->src[k];
+        if (src) {
+            size_t src_hash_pos = ggml_visit_parents(cgraph, src);
+
+            // Update the use count for this operand.
+            cgraph->use_counts[src_hash_pos]++;
         }
     }
 
@@ -5827,6 +6173,8 @@ static void ggml_visit_parents(struct ggml_cgraph * cgraph, struct ggml_tensor *
         cgraph->nodes[cgraph->n_nodes] = node;
         cgraph->n_nodes++;
     }
+
+    return node_hash_pos;
 }
 
 static void ggml_build_forward_impl(struct ggml_cgraph * cgraph, struct ggml_tensor * tensor, bool expand) {
@@ -5964,6 +6312,7 @@ static size_t ggml_graph_nbytes(size_t size, bool grads) {
     incr_ptr_aligned(&p, sizeof(struct ggml_cgraph), 1);
     incr_ptr_aligned(&p, size * sizeof(struct ggml_tensor *), sizeof(struct ggml_tensor *)); // nodes
     incr_ptr_aligned(&p, size * sizeof(struct ggml_tensor *), sizeof(struct ggml_tensor *)); // leafs
+    incr_ptr_aligned(&p, hash_size * sizeof(int32_t), sizeof(int32_t)); // use_counts
     incr_ptr_aligned(&p, hash_size * sizeof(struct ggml_tensor *), sizeof(struct ggml_tensor *)); // hash keys
     if (grads) {
         incr_ptr_aligned(&p, hash_size * sizeof(struct ggml_tensor *), sizeof(struct ggml_tensor *)); // grads
@@ -5993,11 +6342,12 @@ struct ggml_cgraph * ggml_new_graph_custom(struct ggml_context * ctx, size_t siz
 
     void * p = cgraph + 1;
 
-    struct ggml_tensor ** nodes_ptr     =         incr_ptr_aligned(&p, size      * sizeof(struct ggml_tensor *), sizeof(struct ggml_tensor *));
-    struct ggml_tensor ** leafs_ptr     =         incr_ptr_aligned(&p, size      * sizeof(struct ggml_tensor *), sizeof(struct ggml_tensor *));
-    struct ggml_tensor ** hash_keys_ptr =         incr_ptr_aligned(&p, hash_size * sizeof(struct ggml_tensor *), sizeof(struct ggml_tensor *));
-    struct ggml_tensor ** grads_ptr     = grads ? incr_ptr_aligned(&p, hash_size * sizeof(struct ggml_tensor *), sizeof(struct ggml_tensor *)) : NULL;
-    struct ggml_tensor ** grad_accs_ptr = grads ? incr_ptr_aligned(&p, hash_size * sizeof(struct ggml_tensor *), sizeof(struct ggml_tensor *)) : NULL;
+    struct ggml_tensor ** nodes_ptr      =         incr_ptr_aligned(&p, size      * sizeof(struct ggml_tensor *), sizeof(struct ggml_tensor *));
+    struct ggml_tensor ** leafs_ptr      =         incr_ptr_aligned(&p, size      * sizeof(struct ggml_tensor *), sizeof(struct ggml_tensor *));
+    int32_t             * use_counts_ptr =         incr_ptr_aligned(&p, hash_size * sizeof(int32_t), sizeof(int32_t));
+    struct ggml_tensor ** hash_keys_ptr  =         incr_ptr_aligned(&p, hash_size * sizeof(struct ggml_tensor *), sizeof(struct ggml_tensor *));
+    struct ggml_tensor ** grads_ptr      = grads ? incr_ptr_aligned(&p, hash_size * sizeof(struct ggml_tensor *), sizeof(struct ggml_tensor *)) : NULL;
+    struct ggml_tensor ** grad_accs_ptr  = grads ? incr_ptr_aligned(&p, hash_size * sizeof(struct ggml_tensor *), sizeof(struct ggml_tensor *)) : NULL;
 
     ggml_bitset_t * hash_used = incr_ptr_aligned(&p, ggml_bitset_size(hash_size) * sizeof(ggml_bitset_t), sizeof(ggml_bitset_t));
 
@@ -6012,6 +6362,7 @@ struct ggml_cgraph * ggml_new_graph_custom(struct ggml_context * ctx, size_t siz
         /*.grads        =*/ grads_ptr,
         /*.grad_accs    =*/ grad_accs_ptr,
         /*.leafs        =*/ leafs_ptr,
+        /*.use_counts   =*/ use_counts_ptr,
         /*.hash_table   =*/ { hash_size, hash_used, hash_keys_ptr },
         /*.order        =*/ GGML_CGRAPH_EVAL_ORDER_LEFT_TO_RIGHT,
     };
@@ -6038,7 +6389,8 @@ struct ggml_cgraph ggml_graph_view(struct ggml_cgraph * cgraph0, int i0, int i1)
         /*.grads            =*/ NULL, // gradients would need visited_hash_set
         /*.grad_accs        =*/ NULL,
         /*.leafs            =*/ NULL,
-        /*.visited_hash_set =*/ { 0, NULL, NULL },
+        /*.use_counts       =*/ cgraph0->use_counts,
+        /*.visited_hash_set =*/ cgraph0->visited_hash_set,
         /*.order            =*/ cgraph0->order,
     };
 
@@ -6065,7 +6417,8 @@ void ggml_graph_cpy(struct ggml_cgraph * src, struct ggml_cgraph * dst) {
     for (size_t i = 0; i < src->visited_hash_set.size; ++i) {
         // copy all hashset keys (tensors) that are in use
         if (ggml_bitset_get(src->visited_hash_set.used, i)) {
-            ggml_hash_insert(&dst->visited_hash_set, src->visited_hash_set.keys[i]);
+            size_t new_hash_pos = ggml_hash_insert(&dst->visited_hash_set, src->visited_hash_set.keys[i]);
+            dst->use_counts[new_hash_pos] = src->use_counts[i];
         }
     }
 
diff --git a/ggml/src/gguf.cpp b/ggml/src/gguf.cpp
index a0a318a29..5ffd12b8b 100644
--- a/ggml/src/gguf.cpp
+++ b/ggml/src/gguf.cpp
@@ -335,7 +335,11 @@ struct gguf_context * gguf_init_from_file_impl(FILE * file, struct gguf_init_par
 
         for (uint32_t i = 0; i < magic.size(); i++) {
             if (magic[i] != GGUF_MAGIC[i]) {
-                GGML_LOG_ERROR("%s: invalid magic characters: '%c%c%c%c', expected 'GGUF'\n", __func__, magic[0], magic[1], magic[2], magic[3]);
+                char c0 = isprint(magic[0]) ? magic[0] : '?';
+                char c1 = isprint(magic[1]) ? magic[1] : '?';
+                char c2 = isprint(magic[2]) ? magic[2] : '?';
+                char c3 = isprint(magic[3]) ? magic[3] : '?';
+                GGML_LOG_ERROR("%s: invalid magic characters: '%c%c%c%c', expected 'GGUF'\n", __func__, c0, c1, c2, c3);
                 gguf_free(ctx);
                 return nullptr;
             }
diff --git a/gguf-py/gguf/constants.py b/gguf-py/gguf/constants.py
index ae30129b3..6d3d596d6 100644
--- a/gguf-py/gguf/constants.py
+++ b/gguf-py/gguf/constants.py
@@ -118,6 +118,10 @@ class Keys:
         EMBEDDING_SCALE                   = "{arch}.embedding_scale"
         TOKEN_SHIFT_COUNT                 = "{arch}.token_shift_count"
         INTERLEAVE_MOE_LAYER_STEP         = "{arch}.interleave_moe_layer_step"
+        ACTIVATION_SPARSITY_SCALE         = "{arch}.activation_sparsity_scale"
+        ALTUP_ACTIVE_IDX                  = "{arch}.altup.active_idx"
+        ALTUP_NUM_INPUTS                  = "{arch}.altup.num_inputs"
+        EMBD_LENGTH_PER_LAYER_INP         = "{arch}.embedding_length_per_layer_input"
 
     class Attention:
         HEAD_COUNT                   = "{arch}.attention.head_count"
@@ -142,6 +146,8 @@ class Keys:
         SCALE                        = "{arch}.attention.scale"
         KEY_LENGTH_MLA               = "{arch}.attention.key_length_mla"
         VALUE_LENGTH_MLA             = "{arch}.attention.value_length_mla"
+        SHARED_KV_LAYERS             = "{arch}.attention.shared_kv_layers"
+        SLIDING_WINDOW_PATTERN       = "{arch}.attention.sliding_window_pattern"
 
     class Rope:
         DIMENSION_COUNT         = "{arch}.rope.dimension_count"
@@ -164,6 +170,7 @@ class Keys:
         INNER_SIZE     = "{arch}.ssm.inner_size"
         STATE_SIZE     = "{arch}.ssm.state_size"
         TIME_STEP_RANK = "{arch}.ssm.time_step_rank"
+        GROUP_COUNT    = "{arch}.ssm.group_count"
         DT_B_C_RMS     = "{arch}.ssm.dt_b_c_rms"
 
     class WKV:
@@ -198,6 +205,7 @@ class Keys:
         MASK_ID              = "tokenizer.ggml.mask_token_id"
         ADD_BOS              = "tokenizer.ggml.add_bos_token"
         ADD_EOS              = "tokenizer.ggml.add_eos_token"
+        ADD_SEP              = "tokenizer.ggml.add_sep_token"
         ADD_PREFIX           = "tokenizer.ggml.add_space_prefix"
         REMOVE_EXTRA_WS      = "tokenizer.ggml.remove_extra_whitespaces"
         PRECOMPILED_CHARSMAP = "tokenizer.ggml.precompiled_charsmap"
@@ -291,6 +299,7 @@ class MODEL_ARCH(IntEnum):
     BERT             = auto()
     NOMIC_BERT       = auto()
     NOMIC_BERT_MOE   = auto()
+    NEO_BERT         = auto()
     JINA_BERT_V2     = auto()
     JINA_BERT_V3     = auto()
     BLOOM            = auto()
@@ -313,12 +322,14 @@ class MODEL_ARCH(IntEnum):
     GEMMA            = auto()
     GEMMA2           = auto()
     GEMMA3           = auto()
+    GEMMA3N          = auto()
     STARCODER2       = auto()
     RWKV6            = auto()
     RWKV6QWEN2       = auto()
     RWKV7            = auto()
     ARWKV7           = auto()
     MAMBA            = auto()
+    MAMBA2           = auto()
     XVERSE           = auto()
     COMMAND_R        = auto()
     COHERE2          = auto()
@@ -344,6 +355,9 @@ class MODEL_ARCH(IntEnum):
     WAVTOKENIZER_DEC = auto()
     PLM              = auto()
     BAILINGMOE       = auto()
+    DOTS1            = auto()
+    ARCEE            = auto()
+    ERNIE4_5         = auto()
 
 
 class VISION_PROJECTOR_TYPE(IntEnum):
@@ -396,12 +410,29 @@ class MODEL_TENSOR(IntEnum):
     ATTN_Q_NORM          = auto()
     ATTN_K_NORM          = auto()
     LAYER_OUT_NORM       = auto()
+    PER_LAYER_TOKEN_EMBD = auto() # gemma3n
+    PER_LAYER_MODEL_PROJ = auto() # gemma3n
+    PER_LAYER_INP_GATE   = auto() # gemma3n
+    PER_LAYER_PROJ       = auto() # gemma3n
+    PER_LAYER_PROJ_NORM  = auto() # gemma3n
+    PER_LAYER_POST_NORM  = auto() # gemma3n
+    ALTUP_PROJ           = auto() # gemma3n
+    ALTUP_UNEMBD_PROJ    = auto() # gemma3n
+    ALTUP_CORRECT_COEF   = auto() # gemma3n
+    ALTUP_CORRECT_SCALE  = auto() # gemma3n
+    ALTUP_PREDICT_COEF   = auto() # gemma3n
+    ALTUP_ROUTER         = auto() # gemma3n
+    ALTUP_ROUTER_NORM    = auto() # gemma3n
+    LAUREL_L             = auto() # gemma3n
+    LAUREL_R             = auto() # gemma3n
+    LAUREL_POST_NORM     = auto() # gemma3n
     SSM_IN               = auto()
     SSM_CONV1D           = auto()
     SSM_X                = auto()
     SSM_DT               = auto()
     SSM_A                = auto()
     SSM_D                = auto()
+    SSM_NORM             = auto()
     SSM_OUT              = auto()
     TIME_MIX_W0          = auto()
     TIME_MIX_W1          = auto()
@@ -572,6 +603,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
     MODEL_ARCH.BERT:             "bert",
     MODEL_ARCH.NOMIC_BERT:       "nomic-bert",
     MODEL_ARCH.NOMIC_BERT_MOE:   "nomic-bert-moe",
+    MODEL_ARCH.NEO_BERT:         "neo-bert",
     MODEL_ARCH.JINA_BERT_V2:     "jina-bert-v2",
     MODEL_ARCH.JINA_BERT_V3:     "jina-bert-v3",
     MODEL_ARCH.BLOOM:            "bloom",
@@ -594,12 +626,14 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
     MODEL_ARCH.GEMMA:            "gemma",
     MODEL_ARCH.GEMMA2:           "gemma2",
     MODEL_ARCH.GEMMA3:           "gemma3",
+    MODEL_ARCH.GEMMA3N:          "gemma3n",
     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.MAMBA2:           "mamba2",
     MODEL_ARCH.XVERSE:           "xverse",
     MODEL_ARCH.COMMAND_R:        "command-r",
     MODEL_ARCH.COHERE2:          "cohere2",
@@ -625,6 +659,9 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
     MODEL_ARCH.WAVTOKENIZER_DEC: "wavtokenizer-dec",
     MODEL_ARCH.PLM:              "plm",
     MODEL_ARCH.BAILINGMOE:       "bailingmoe",
+    MODEL_ARCH.DOTS1:            "dots1",
+    MODEL_ARCH.ARCEE:            "arcee",
+    MODEL_ARCH.ERNIE4_5:         "ernie4_5",
 }
 
 VISION_PROJECTOR_TYPE_NAMES: dict[VISION_PROJECTOR_TYPE, str] = {
@@ -677,12 +714,29 @@ TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
     MODEL_TENSOR.FFN_UP_EXP:                "blk.{bid}.ffn_up_exps",
     MODEL_TENSOR.FFN_EXP_PROBS_B:           "blk.{bid}.exp_probs_b",
     MODEL_TENSOR.LAYER_OUT_NORM:            "blk.{bid}.layer_output_norm",
+    MODEL_TENSOR.PER_LAYER_TOKEN_EMBD:      "per_layer_token_embd",           # gemma3n
+    MODEL_TENSOR.PER_LAYER_MODEL_PROJ:      "per_layer_model_proj",           # gemma3n
+    MODEL_TENSOR.PER_LAYER_PROJ_NORM:       "per_layer_proj_norm",            # gemma3n
+    MODEL_TENSOR.ALTUP_UNEMBD_PROJ:         "altup_unembd_proj",              # gemma3n
+    MODEL_TENSOR.ALTUP_PROJ:                "altup_proj",                     # gemma3n
+    MODEL_TENSOR.PER_LAYER_INP_GATE:        "blk.{bid}.inp_gate",             # gemma3n
+    MODEL_TENSOR.PER_LAYER_PROJ:            "blk.{bid}.proj",                 # gemma3n
+    MODEL_TENSOR.PER_LAYER_POST_NORM:       "blk.{bid}.post_norm",            # gemma3n
+    MODEL_TENSOR.ALTUP_CORRECT_COEF:        "blk.{bid}.altup_correct_coef",   # gemma3n
+    MODEL_TENSOR.ALTUP_CORRECT_SCALE:       "blk.{bid}.altup_correct_scale",  # gemma3n
+    MODEL_TENSOR.ALTUP_PREDICT_COEF:        "blk.{bid}.altup_predict_coef",   # gemma3n
+    MODEL_TENSOR.ALTUP_ROUTER:              "blk.{bid}.altup_router",         # gemma3n
+    MODEL_TENSOR.ALTUP_ROUTER_NORM:         "blk.{bid}.altup_router_norm",    # gemma3n
+    MODEL_TENSOR.LAUREL_L:                  "blk.{bid}.laurel_l",             # gemma3n
+    MODEL_TENSOR.LAUREL_R:                  "blk.{bid}.laurel_r",             # gemma3n
+    MODEL_TENSOR.LAUREL_POST_NORM:          "blk.{bid}.laurel_post_norm",     # gemma3n
     MODEL_TENSOR.SSM_IN:                    "blk.{bid}.ssm_in",
     MODEL_TENSOR.SSM_CONV1D:                "blk.{bid}.ssm_conv1d",
     MODEL_TENSOR.SSM_X:                     "blk.{bid}.ssm_x",
     MODEL_TENSOR.SSM_DT:                    "blk.{bid}.ssm_dt",
     MODEL_TENSOR.SSM_A:                     "blk.{bid}.ssm_a",
     MODEL_TENSOR.SSM_D:                     "blk.{bid}.ssm_d",
+    MODEL_TENSOR.SSM_NORM:                  "blk.{bid}.ssm_norm",
     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",
@@ -1079,6 +1133,18 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.FFN_UP_EXP,
         MODEL_TENSOR.LAYER_OUT_NORM,
     ],
+    MODEL_ARCH.NEO_BERT: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_QKV,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+        MODEL_TENSOR.ENC_OUTPUT_NORM,
+        MODEL_TENSOR.CLS,
+        MODEL_TENSOR.CLS_OUT,
+    ],
     MODEL_ARCH.JINA_BERT_V2: [
         MODEL_TENSOR.TOKEN_EMBD,
         MODEL_TENSOR.TOKEN_EMBD_NORM,
@@ -1481,6 +1547,41 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.FFN_PRE_NORM,
         MODEL_TENSOR.FFN_POST_NORM,
     ],
+    MODEL_ARCH.GEMMA3N: [
+        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,
+        # altup / laurel
+        MODEL_TENSOR.PER_LAYER_TOKEN_EMBD,
+        MODEL_TENSOR.PER_LAYER_MODEL_PROJ,
+        MODEL_TENSOR.PER_LAYER_INP_GATE,
+        MODEL_TENSOR.PER_LAYER_PROJ,
+        MODEL_TENSOR.PER_LAYER_PROJ_NORM,
+        MODEL_TENSOR.PER_LAYER_POST_NORM,
+        MODEL_TENSOR.ALTUP_PROJ,
+        MODEL_TENSOR.ALTUP_UNEMBD_PROJ,
+        MODEL_TENSOR.ALTUP_CORRECT_COEF,
+        MODEL_TENSOR.ALTUP_CORRECT_SCALE,
+        MODEL_TENSOR.ALTUP_PREDICT_COEF,
+        MODEL_TENSOR.ALTUP_ROUTER,
+        MODEL_TENSOR.ALTUP_ROUTER_NORM,
+        MODEL_TENSOR.LAUREL_L,
+        MODEL_TENSOR.LAUREL_R,
+        MODEL_TENSOR.LAUREL_POST_NORM,
+    ],
     MODEL_ARCH.STARCODER2: [
         MODEL_TENSOR.TOKEN_EMBD,
         MODEL_TENSOR.OUTPUT_NORM,
@@ -1632,6 +1733,19 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.SSM_D,
         MODEL_TENSOR.SSM_OUT,
     ],
+    MODEL_ARCH.MAMBA2: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.SSM_IN,
+        MODEL_TENSOR.SSM_CONV1D,
+        MODEL_TENSOR.SSM_DT,
+        MODEL_TENSOR.SSM_A,
+        MODEL_TENSOR.SSM_D,
+        MODEL_TENSOR.SSM_NORM,
+        MODEL_TENSOR.SSM_OUT,
+    ],
     MODEL_ARCH.XVERSE: [
         MODEL_TENSOR.TOKEN_EMBD,
         MODEL_TENSOR.OUTPUT_NORM,
@@ -2058,6 +2172,59 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.FFN_DOWN_SHEXP,
         MODEL_TENSOR.FFN_UP_SHEXP,
     ],
+    MODEL_ARCH.DOTS1: [
+        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_EXP_PROBS_B,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.FFN_GATE,
+        MODEL_TENSOR.FFN_GATE_EXP,
+        MODEL_TENSOR.FFN_GATE_INP,
+        MODEL_TENSOR.FFN_GATE_SHEXP,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_DOWN_EXP,
+        MODEL_TENSOR.FFN_DOWN_SHEXP,
+        MODEL_TENSOR.FFN_UP,
+        MODEL_TENSOR.FFN_UP_EXP,
+        MODEL_TENSOR.FFN_UP_SHEXP,
+    ],
+    MODEL_ARCH.ARCEE: [
+        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_NORM,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+    ],
+    MODEL_ARCH.ERNIE4_5: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_Q,
+        MODEL_TENSOR.ATTN_K,
+        MODEL_TENSOR.ATTN_V,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.FFN_GATE,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+    ],
     # TODO
 }
 
@@ -2362,6 +2529,7 @@ KEY_SSM_CONV_KERNEL    = Keys.SSM.CONV_KERNEL
 KEY_SSM_INNER_SIZE     = Keys.SSM.INNER_SIZE
 KEY_SSM_STATE_SIZE     = Keys.SSM.STATE_SIZE
 KEY_SSM_TIME_STEP_RANK = Keys.SSM.TIME_STEP_RANK
+KEY_SSM_GROUP_COUNT    = Keys.SSM.GROUP_COUNT
 KEY_SSM_DT_B_C_RMS     = Keys.SSM.DT_B_C_RMS
 
 # tokenization
diff --git a/gguf-py/gguf/gguf_writer.py b/gguf-py/gguf/gguf_writer.py
index de6e45ae8..a7ecf3d31 100644
--- a/gguf-py/gguf/gguf_writer.py
+++ b/gguf-py/gguf/gguf_writer.py
@@ -271,7 +271,7 @@ class GGUFWriter:
 
     def add_key_value(self, key: str, val: Any, vtype: GGUFValueType, sub_type: GGUFValueType | None = None) -> None:
         if any(key in kv_data for kv_data in self.kv_data):
-            raise ValueError(f'Duplicated key name {key!r}')
+            logger.warning(f'Duplicated key name {key!r}, overwriting it with new value {val!r} of type {vtype.name}')
 
         self.kv_data[0][key] = GGUFValue(value=val, type=vtype, sub_type=sub_type)
 
@@ -672,6 +672,18 @@ class GGUFWriter:
     def add_decoder_start_token_id(self, id: int) -> None:
         self.add_uint32(Keys.LLM.DECODER_START_TOKEN_ID.format(arch=self.arch), id)
 
+    def add_embedding_length_per_layer_input(self, value: int) -> None:
+        self.add_uint32(Keys.LLM.EMBD_LENGTH_PER_LAYER_INP.format(arch=self.arch), value)
+
+    def add_altup_active_idx(self, val: int) -> None:
+        self.add_uint32(Keys.LLM.ALTUP_ACTIVE_IDX.format(arch=self.arch), val)
+
+    def add_altup_num_inputs(self, val: int) -> None:
+        self.add_uint32(Keys.LLM.ALTUP_NUM_INPUTS.format(arch=self.arch), val)
+
+    def add_activation_sparsity_scale(self, values: Sequence[float]) -> None:
+        self.add_array(Keys.LLM.ACTIVATION_SPARSITY_SCALE.format(arch=self.arch), values)
+
     def add_head_count(self, count: int | Sequence[int]) -> None:
         if isinstance(count, int):
             self.add_uint32(Keys.Attention.HEAD_COUNT.format(arch=self.arch), count)
@@ -702,6 +714,12 @@ class GGUFWriter:
     def add_clamp_kqv(self, value: float) -> None:
         self.add_float32(Keys.Attention.CLAMP_KQV.format(arch=self.arch), value)
 
+    def add_shared_kv_layers(self, value: int) -> None:
+        self.add_uint32(Keys.Attention.SHARED_KV_LAYERS.format(arch=self.arch), value)
+
+    def add_sliding_window_pattern(self, value: Sequence[bool]) -> None:
+        self.add_array(Keys.Attention.SLIDING_WINDOW_PATTERN.format(arch=self.arch), value)
+
     def add_logit_scale(self, value: float) -> None:
         self.add_float32(Keys.LLM.LOGIT_SCALE.format(arch=self.arch), value)
 
@@ -843,6 +861,9 @@ class GGUFWriter:
     def add_ssm_time_step_rank(self, value: int) -> None:
         self.add_uint32(Keys.SSM.TIME_STEP_RANK.format(arch=self.arch), value)
 
+    def add_ssm_group_count(self, value: int) -> None:
+        self.add_uint32(Keys.SSM.GROUP_COUNT.format(arch=self.arch), value)
+
     def add_ssm_dt_b_c_rms(self, value: bool) -> None:
         self.add_bool(Keys.SSM.DT_B_C_RMS.format(arch=self.arch), value)
 
@@ -891,6 +912,9 @@ class GGUFWriter:
     def add_add_eos_token(self, value: bool) -> None:
         self.add_bool(Keys.Tokenizer.ADD_EOS, value)
 
+    def add_add_sep_token(self, value: bool) -> None:
+        self.add_bool(Keys.Tokenizer.ADD_SEP, value)
+
     def add_add_space_prefix(self, value: bool) -> None:
         self.add_bool(Keys.Tokenizer.ADD_PREFIX, value)
 
@@ -935,6 +959,9 @@ class GGUFWriter:
     def add_eom_token_id(self, id: int) -> None:
         self.add_uint32(Keys.Tokenizer.EOM_ID, id)
 
+    def add_classifier_output_labels(self, labels: Sequence[str]) -> None:
+        self.add_array(Keys.Classifier.OUTPUT_LABELS.format(arch=self.arch), labels)
+
     # for vision models
 
     def add_clip_has_vision_encoder(self, value: bool) -> None:
diff --git a/gguf-py/gguf/tensor_mapping.py b/gguf-py/gguf/tensor_mapping.py
index 93dd1d802..51634ef6b 100644
--- a/gguf-py/gguf/tensor_mapping.py
+++ b/gguf-py/gguf/tensor_mapping.py
@@ -31,6 +31,7 @@ class TensorNameMap:
             "model.embeddings",                          # rwkv7
             "model.word_embeddings",                     # bailingmoe
             "language_model.model.embed_tokens",         # llama4
+            "encoder",                                   # neobert
         ),
 
         # Token type embeddings
@@ -134,6 +135,7 @@ class TensorNameMap:
             "rwkv.blocks.{bid}.ln1",                                # rwkv6
             "model.layers.{bid}.ln1",                               # rwkv7
             "model.layers.{bid}.input_layernorm",                   # llama4
+            "transformer_encoder.{bid}.attention_norm",             # neobert
         ),
 
         # Attention norm 2
@@ -161,6 +163,7 @@ class TensorNameMap:
             "model.layers.{bid}.self_attn.qkv_proj",                               # phi3
             "encoder.layers.{bid}.self_attention.query_key_value",                 # chatglm
             "transformer.layers.{bid}.attn.qkv_proj",                              # openelm
+            "transformer_encoder.{bid}.qkv",                                       # neobert
         ),
 
         # Attention query
@@ -236,6 +239,7 @@ class TensorNameMap:
             "transformer.layers.{bid}.attn.out_proj",                       # openelm
             "transformer.h.{bid}.attn.attention.out_proj",                  # exaone
             "model.layers.{bid}.self_attn.o_proj",                          # llama4
+            "transformer_encoder.{bid}.wo",                                 # neobert
         ),
 
         # Attention output norm
@@ -276,6 +280,7 @@ class TensorNameMap:
             "encoder.layers.{bid}.post_attention_layernorm",                 # chatglm
             "transformer.layers.{bid}.ffn_norm",                             # openelm
             "model.layers.{bid}.post_attention_layernorm",                   # llama4
+            "transformer_encoder.{bid}.ffn_norm",                            # neobert
         ),
 
         # Post feed-forward norm
@@ -305,7 +310,7 @@ class TensorNameMap:
         ),
 
         MODEL_TENSOR.FFN_EXP_PROBS_B: (
-            "model.layers.{bid}.mlp.gate.e_score_correction", # deepseek-v3
+            "model.layers.{bid}.mlp.gate.e_score_correction", # deepseek-v3 dots1
         ),
 
         # Feed-forward up
@@ -333,11 +338,14 @@ class TensorNameMap:
             "encoder.layers.{bid}.mlp.fc11",                          # nomic-bert
             "encoder.layers.{bid}.mlp.fc1",                           # nomic-bert-moe
             "model.layers.{bid}.mlp.c_fc",                            # starcoder2
-            "encoder.layer.{bid}.mlp.gated_layers_v",                 # jina-bert-v2
+            "encoder.layer.{bid}.mlp.gated_layers_v",                 # jina-bert-v2 (split up/gate, no longer used)
+            "encoder.layer.{bid}.mlp.gated_layers",                   # jina-bert-v2 (GEGLU)
+            "encoder.layer.{bid}.mlp.up_gated_layer",                 # jina-v2-code (GEGLU)
             "model.layers.{bid}.residual_mlp.w3",                     # arctic
             "encoder.layers.{bid}.mlp.dense_h_to_4h",                 # chatglm
             "transformer.h.{bid}.mlp.c_fc_1",                         # exaone
             "model.layers.{bid}.feed_forward.up_proj",                # llama4
+            "transformer_encoder.{bid}.ffn.w12",                      # neobert
         ),
 
         MODEL_TENSOR.FFN_UP_EXP: (
@@ -370,7 +378,7 @@ class TensorNameMap:
             "model.layers.layers.{bid}.mlp.gate_proj",    # plamo
             "model.layers.{bid}.feed_forward.w1",         # internlm2
             "encoder.layers.{bid}.mlp.fc12",              # nomic-bert
-            "encoder.layer.{bid}.mlp.gated_layers_w",     # jina-bert-v2
+            "encoder.layer.{bid}.mlp.gated_layers_w",     # jina-bert-v2 (split up/gate, no longer used)
             "transformer.h.{bid}.mlp.linear_1",           # refact
             "model.layers.{bid}.residual_mlp.w1",         # arctic
             "transformer.h.{bid}.mlp.c_fc_0",             # exaone
@@ -420,6 +428,7 @@ class TensorNameMap:
             "encoder.layers.{bid}.mlp.dense_4h_to_h",                 # chatglm
             "model.layers.h.{bid}.mlp.c_proj",                        # exaone
             "model.layers.{bid}.feed_forward.down_proj",              # llama4
+            "transformer_encoder.{bid}.ffn.w3",                       # neobert
         ),
 
         MODEL_TENSOR.FFN_DOWN_EXP: (
@@ -468,7 +477,71 @@ class TensorNameMap:
             "encoder.layers.{bid}.norm2",                   # nomic-bert
             "transformer.decoder_layer.{bid}.rms_norm_3",   # Grok
             "encoder.layer.{bid}.mlp.layernorm",            # jina-bert-v2
-            "encoder.layer.{bid}.layer_norm_2"              # jina-v2-code
+            "encoder.layer.{bid}.layer_norm_2",             # jina-v2-code
+        ),
+
+        MODEL_TENSOR.PER_LAYER_TOKEN_EMBD: (
+            "model.embed_tokens_per_layer",  # gemma3n
+        ),
+
+        MODEL_TENSOR.PER_LAYER_MODEL_PROJ: (
+            "model.per_layer_model_projection",  # gemma3n
+        ),
+
+        MODEL_TENSOR.PER_LAYER_PROJ_NORM: (
+            "model.per_layer_projection_norm",  # gemma3n
+        ),
+
+        MODEL_TENSOR.ALTUP_PROJ: (
+            "model.altup_projections",  # gemma3n
+        ),
+
+        MODEL_TENSOR.ALTUP_UNEMBD_PROJ: (
+            "model.altup_unembed_projections",  # gemma3n
+        ),
+
+        MODEL_TENSOR.PER_LAYER_INP_GATE: (
+            "model.layers.{bid}.per_layer_input_gate",  # gemma3n
+        ),
+
+        MODEL_TENSOR.PER_LAYER_PROJ: (
+            "model.layers.{bid}.per_layer_projection",  # gemma3n
+        ),
+
+        MODEL_TENSOR.PER_LAYER_POST_NORM: (
+            "model.layers.{bid}.post_per_layer_input_norm",  # gemma3n
+        ),
+
+        MODEL_TENSOR.ALTUP_CORRECT_COEF: (
+            "model.layers.{bid}.altup.correction_coefs",  # gemma3n
+        ),
+
+        MODEL_TENSOR.ALTUP_CORRECT_SCALE: (
+            "model.layers.{bid}.altup.correct_output_scale",  # gemma3n
+        ),
+
+        MODEL_TENSOR.ALTUP_PREDICT_COEF: (
+            "model.layers.{bid}.altup.prediction_coefs",  # gemma3n
+        ),
+
+        MODEL_TENSOR.ALTUP_ROUTER: (
+            "model.layers.{bid}.altup.modality_router",  # gemma3n
+        ),
+
+        MODEL_TENSOR.ALTUP_ROUTER_NORM: (
+            "model.layers.{bid}.altup.router_norm",  # gemma3n
+        ),
+
+        MODEL_TENSOR.LAUREL_L: (
+            "model.layers.{bid}.laurel.linear_left",  # gemma3n
+        ),
+
+        MODEL_TENSOR.LAUREL_R: (
+            "model.layers.{bid}.laurel.linear_right",  # gemma3n
+        ),
+
+        MODEL_TENSOR.LAUREL_POST_NORM: (
+            "model.layers.{bid}.laurel.post_laurel_norm",  # gemma3n
         ),
 
         MODEL_TENSOR.SSM_IN: (
@@ -501,6 +574,10 @@ class TensorNameMap:
             "backbone.layers.{bid}.mixer.D",
         ),
 
+        MODEL_TENSOR.SSM_NORM: (
+            "backbone.layers.{bid}.mixer.norm",  # mamba2
+        ),
+
         MODEL_TENSOR.SSM_OUT: (
             "model.layers.{bid}.out_proj",
             "backbone.layers.{bid}.mixer.out_proj",
@@ -830,12 +907,14 @@ class TensorNameMap:
         # TODO: these do not belong to block_mappings_cfg - move them to mappings_cfg
         MODEL_TENSOR.ENC_OUTPUT_NORM: (
             "encoder.final_layer_norm", # t5
+            "layer_norm",               # neobert
         ),
 
         MODEL_TENSOR.CLS: (
             "classifier",       # jina
             "classifier.dense", # roberta
             "pre_classifier",   # distillbert
+            "dense",            # neobert
         ),
 
         MODEL_TENSOR.CLS_OUT: (
diff --git a/gguf-py/gguf/vocab.py b/gguf-py/gguf/vocab.py
index cca097986..635fcef35 100644
--- a/gguf-py/gguf/vocab.py
+++ b/gguf-py/gguf/vocab.py
@@ -7,7 +7,10 @@ import os
 from pathlib import Path
 from typing import Any, Callable, Sequence, Mapping, Iterable, Protocol, ClassVar, runtime_checkable
 
-from sentencepiece import SentencePieceProcessor
+try:
+    from sentencepiece import SentencePieceProcessor
+except ImportError:
+    SentencePieceProcessor = None
 
 import gguf
 
@@ -116,6 +119,7 @@ class SpecialVocab:
         logger.warning(f'Special token type {typ}, id {tid} out of range, must be under {self.n_vocab} - skipping')
 
     def _try_load_from_tokenizer_json(self, path: Path) -> bool:
+        tokenizer = None
         tokenizer_file = path / 'tokenizer.json'
         if tokenizer_file.is_file():
             with open(tokenizer_file, encoding = 'utf-8') as f:
@@ -149,15 +153,110 @@ class SpecialVocab:
             added_tokens = tokenizer.get('added_tokens', {})
         else:
             added_tokens = {}
+        tokenizer_config = None
         tokenizer_config_file = path / 'tokenizer_config.json'
-        if not tokenizer_config_file.is_file():
+        if tokenizer_config_file.is_file():
+            with open(tokenizer_config_file, encoding = 'utf-8') as f:
+                tokenizer_config = json.load(f)
+        if tokenizer:
+            special_bos = (tokenizer_config or {}).get('bos_token')
+            special_cls = (tokenizer_config or {}).get('cls_token')
+            special_eos = (tokenizer_config or {}).get('eos_token')
+            special_sep = (tokenizer_config or {}).get('sep_token')
+            if not special_bos and special_cls and tokenizer_config:
+                tokenizer_config['bos_token'] = special_bos = special_cls
+            if not special_eos and special_sep and tokenizer_config:
+                tokenizer_config['eos_token'] = special_eos = special_sep
+            if post_processor := tokenizer.get('post_processor'):
+                for processor in post_processor.get('processors', [post_processor]):
+                    if processor.get('type') == 'RobertaProcessing':
+                        self.add_special_token['bos'] = True
+                        self.add_special_token['eos'] = True
+                        self.add_special_token['sep'] = True
+                        if not special_cls and tokenizer_config:
+                            special_cls = processor.get('cls', [special_bos])[0]
+                            tokenizer_config['cls_token'] = special_cls
+                        if not special_sep and tokenizer_config:
+                            special_sep = processor.get('sep', [special_eos])[0]
+                            tokenizer_config['sep_token'] = special_sep
+                        continue
+                    # Crude parsing of TemplateProcessing to determine if BOS/SEP/EOS should be added
+                    # Only works with simple templates, **will** get it wrong on unusual sequences
+                    if processor.get('type') == 'TemplateProcessing':
+                        tmpl_single = processor.get('single', [])
+                        tmpl_pair = processor.get('pair', [])
+                        special_first = None
+                        special_last = None
+                        if len(tmpl_single) > 1:
+                            if special_first := tmpl_single[0].get('SpecialToken', {}).get('id'):
+                                if not tokenizer_config:
+                                    special_bos = special_first
+                                self.add_special_token['bos'] = True if special_first in (special_bos, special_cls) else False
+                                if special_first not in (special_bos, special_cls):
+                                    logger.warning(f'Unknown leading special token {special_first!r} in TemplateProcessing<single>')
+                            if special_last := tmpl_single[-1].get('SpecialToken', {}).get('id'):
+                                if not tokenizer_config:
+                                    special_eos = special_last
+                                elif special_last != special_eos:
+                                    if 'eot' not in self.special_token_types:
+                                        self.special_token_types = tuple(self.special_token_types) + ('eot', )
+                                        tokenizer_config['eot_token'] = special_eos
+                                    elif 'eom' not in self.special_token_types:
+                                        self.special_token_types = tuple(self.special_token_types) + ('eom', )
+                                        tokenizer_config['eom_token'] = special_eos
+                                    else:
+                                        logger.warning(f'Overriding EOS token {special_eos!r} with {special_last!r} without EOT/EOM fallback!')
+                                    tokenizer_config['eos_token'] = special_eos = special_last
+                                self.add_special_token['eos'] = True if special_last == special_eos else False
+                                if special_last != special_eos:
+                                    logger.warning(f'Unknown trailing special token {special_last!r} in TemplateProcessing<single>')
+                        if tmpl_pair:
+                            seq_start = 1 if special_first and tmpl_pair[0].get('SpecialToken', {}).get('id') == special_first else 0
+                            seq_stop = -1 if special_last and tmpl_pair[-1].get('SpecialToken', {}).get('id') == special_last else None
+                            if (special_first and seq_start == 0) or (special_last and seq_stop is None):
+                                logger.warning('TemplateProcessing<single> leading/trailing special tokens do not match TemplateProcessing<pair>')
+                            if tmpl_pair := tmpl_pair[slice(seq_start, seq_stop)]:
+                                tmpl_a = tmpl_pair[0].get('Sequence', {}).get('id')
+                                tmpl_b = tmpl_pair[-1].get('Sequence', {}).get('id')
+                                if tmpl_a != 'A' or tmpl_b != 'B':
+                                    logger.warning(f'Unknown sequence {tmpl_a}...{tmpl_b} in TemplateProcessing<pair>')
+                                # A [sep] [eos] B
+                                if tmpl_a == 'A' and tmpl_b == 'B' and (tmpl_pair := tmpl_pair[1:-1]):
+                                    add_sep = False
+                                    if special_entry := tmpl_pair[0].get('SpecialToken', {}).get('id'):
+                                        if special_entry in (special_sep, special_eos) and not special_last:
+                                            add_sep = True
+                                        if special_entry not in (special_sep, special_eos):
+                                            logger.warning(f'Unknown separator token {special_entry!r} in TemplateProcessing<pair>')
+                                    else:
+                                        logger.warning(f'Unknown middle sequence {tmpl_pair[0]!r} in TemplateProcessing<pair>')
+                                    if len(tmpl_pair) == 2:
+                                        if special_entry := tmpl_pair[1].get('SpecialToken', {}).get('id'):
+                                            if special_entry in (special_sep, special_eos):
+                                                add_sep = True
+                                            if special_entry not in (special_sep, special_eos):
+                                                logger.warning(f'Unknown second separator token {special_entry!r} in TemplateProcessing<pair>')
+                                        else:
+                                            logger.warning(f'Unknown second middle sequence {tmpl_pair[1]!r} in TemplateProcessing<pair>')
+                                    self.add_special_token['sep'] = add_sep
+                                    if add_sep and not special_sep and tokenizer_config:
+                                        tokenizer_config['sep_token'] = special_eos
+                        continue
+        if not tokenizer_config:
             return True
-        with open(tokenizer_config_file, encoding = 'utf-8') as f:
-            tokenizer_config = json.load(f)
         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_json = path / 'chat_template.json'
+        chat_template_jinja = path / 'chat_template.jinja'
+        if chat_template_jinja.is_file():
+            with open(chat_template_jinja, encoding = 'utf-8') as f:
+                chat_template_alt = f.read()
+            if additional_templates := list((path / 'additional_chat_templates').glob('*.jinja')):
+                chat_template_alt = [{'name': 'default', 'template': chat_template_alt}]
+                for template_path in additional_templates:
+                    with open(template_path, encoding = 'utf-8') as fp:
+                        chat_template_alt.append({'name': template_path.stem, 'template': fp.read()})
+        elif chat_template_json.is_file():
+            with open(chat_template_json, 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)):
@@ -302,6 +401,9 @@ class SentencePieceVocab(Vocab):
     name = "spm"
 
     def __init__(self, base_path: Path):
+        if SentencePieceProcessor is None:
+            raise RuntimeError("sentencepiece is not installed")
+
         added_tokens: dict[str, int] = {}
         if (fname_tokenizer := base_path / 'tokenizer.model').exists():
             # normal location
diff --git a/gguf-py/pyproject.toml b/gguf-py/pyproject.toml
index f11351cba..0f3a1eeee 100644
--- a/gguf-py/pyproject.toml
+++ b/gguf-py/pyproject.toml
@@ -1,6 +1,6 @@
 [tool.poetry]
 name = "gguf"
-version = "0.17.0"
+version = "0.17.1"
 description = "Read and write ML models in GGUF for GGML"
 authors = ["GGML <ggml@ggml.ai>"]
 packages = [
@@ -22,7 +22,7 @@ python = ">=3.8"
 numpy = ">=1.17"
 tqdm = ">=4.27"
 pyyaml = ">=5.1"
-sentencepiece = ">=0.1.98,<=0.2.0"
+sentencepiece = { version = ">=0.1.98,<=0.2.0", optional = true }
 PySide6 = { version = "^6.9", python = ">=3.9,<3.14", optional = true }
 
 [tool.poetry.dev-dependencies]
diff --git a/include/llama.h b/include/llama.h
index da0f652cf..3eda9bc68 100644
--- a/include/llama.h
+++ b/include/llama.h
@@ -61,7 +61,10 @@ extern "C" {
     struct llama_model;
     struct llama_context;
     struct llama_sampler;
-    struct llama_kv_cache;
+
+    typedef struct llama_memory_i * llama_memory_t;
+
+    struct llama_kv_cache; // DEPRECATED (use llama_memory instead)
 
     typedef int32_t llama_pos;
     typedef int32_t llama_token;
@@ -240,18 +243,21 @@ extern "C" {
 
     typedef bool (*llama_progress_callback)(float progress, void * user_data);
 
-    // Input data for llama_decode
+    // Input data for llama_encode/llama_decode
     // A llama_batch object can contain input about one or many sequences
     // The provided arrays (i.e. token, embd, pos, etc.) must have size of n_tokens
     //
     // - token  : the token ids of the input (used when embd is NULL)
     // - embd   : token embeddings (i.e. float vector of size n_embd) (used when token is NULL)
     // - pos    : the positions of the respective token in the sequence
-    //            (if set to NULL, the token position will be tracked automatically by llama_decode)
+    //            (if set to NULL, the token position will be tracked automatically by llama_encode/llama_decode)
     // - seq_id : the sequence to which the respective token belongs
     //            (if set to NULL, the sequence ID will be assumed to be 0)
     // - logits : if zero, the logits (and/or the embeddings) for the respective token will not be output
-    //            (if set to NULL, only the logits for last token will be returned)
+    //            (if set to NULL:
+    //               - if embeddings: all tokens are output
+    //               - if not:        only the last token is output
+    //            )
     //
     typedef struct llama_batch {
         int32_t n_tokens;
@@ -259,8 +265,8 @@ extern "C" {
         llama_token  *  token;
         float        *  embd;
         llama_pos    *  pos;
-        int32_t      *  n_seq_id; // TODO: remove, should belong to only 1 sequence
-        llama_seq_id ** seq_id;   // TODO: become llama_seq_id * seq_id;
+        int32_t      *  n_seq_id;
+        llama_seq_id ** seq_id;
         int8_t       *  logits;   // TODO: rename this to "output"
     } llama_batch;
 
@@ -384,6 +390,7 @@ extern "C" {
         void * imatrix;                       // pointer to importance matrix data
         void * kv_overrides;                  // pointer to vector containing overrides
         void * tensor_types;                  // pointer to vector containing tensor types
+        void * prune_layers;                  // pointer to vector containing layer indices to prune
     } llama_model_quantize_params;
 
     typedef struct llama_logit_bias {
@@ -493,9 +500,11 @@ 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    struct llama_kv_cache * llama_get_kv_self (      struct llama_context * ctx);
+    LLAMA_API           llama_memory_t   llama_get_memory  (const 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
 
+    DEPRECATED(LLAMA_API struct llama_kv_cache * llama_get_kv_self(struct llama_context * ctx), "use llama_get_memory instead");
+
     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);
 
@@ -509,6 +518,13 @@ extern "C" {
     // Get the model's RoPE frequency scaling factor
     LLAMA_API float llama_model_rope_freq_scale_train(const struct llama_model * model);
 
+    // Returns the number of classifier outputs (only valid for classifier models)
+    // Undefined behavior for non-classifier models
+    LLAMA_API uint32_t llama_model_n_cls_out(const struct llama_model * model);
+
+    // Returns label of classifier output by index (<n_cls_out). Returns nullptr if no label provided
+    LLAMA_API const char * llama_model_cls_label(const struct llama_model * model, uint32_t i);
+
     LLAMA_API enum llama_vocab_type llama_vocab_type(const struct llama_vocab * vocab);
 
     LLAMA_API int32_t llama_vocab_n_tokens(const struct llama_vocab * vocab);
@@ -609,7 +625,81 @@ extern "C" {
                          int32_t   il_end);
 
     //
-    // KV cache
+    // Memory
+    //
+
+    // Clear the memory contents
+    // If data == true, the data buffers will also be cleared together with the metadata
+    LLAMA_API void llama_memory_clear(
+            llama_memory_t mem,
+                      bool data);
+
+    // Removes all tokens that belong to the specified sequence and have positions in [p0, p1)
+    // Returns false if a partial sequence cannot be removed. Removing a whole sequence never fails
+    // seq_id < 0 : match any sequence
+    // p0 < 0     : [0,  p1]
+    // p1 < 0     : [p0, inf)
+    LLAMA_API bool llama_memory_seq_rm(
+            llama_memory_t mem,
+              llama_seq_id seq_id,
+                 llama_pos p0,
+                 llama_pos p1);
+
+    // Copy all tokens that belong to the specified sequence to another sequence
+    // p0 < 0 : [0,  p1]
+    // p1 < 0 : [p0, inf)
+    LLAMA_API void llama_memory_seq_cp(
+            llama_memory_t mem,
+              llama_seq_id seq_id_src,
+              llama_seq_id seq_id_dst,
+                 llama_pos p0,
+                 llama_pos p1);
+
+    // Removes all tokens that do not belong to the specified sequence
+    LLAMA_API void llama_memory_seq_keep(
+            llama_memory_t mem,
+              llama_seq_id seq_id);
+
+    // Adds relative position "delta" to all tokens that belong to the specified sequence and have positions in [p0, p1)
+    // p0 < 0 : [0,  p1]
+    // p1 < 0 : [p0, inf)
+    LLAMA_API void llama_memory_seq_add(
+            llama_memory_t mem,
+              llama_seq_id seq_id,
+                 llama_pos p0,
+                 llama_pos p1,
+                 llama_pos delta);
+
+    // Integer division of the positions by factor of `d > 1`
+    // p0 < 0 : [0,  p1]
+    // p1 < 0 : [p0, inf)
+    LLAMA_API void llama_memory_seq_div(
+            llama_memory_t mem,
+              llama_seq_id seq_id,
+                 llama_pos p0,
+                 llama_pos p1,
+                       int d);
+
+    // Returns the smallest position present in the memory for the specified sequence
+    // This is typically non-zero only for SWA caches
+    // Note that all positions in the range [pos_min, pos_max] are guaranteed to be present in the memory
+    // Return -1 if the sequence is empty
+    LLAMA_API llama_pos llama_memory_seq_pos_min(
+            llama_memory_t mem,
+              llama_seq_id seq_id);
+
+    // Returns the largest position present in the memory for the specified sequence
+    // Note that all positions in the range [pos_min, pos_max] are guaranteed to be present in the memory
+    // Return -1 if the sequence is empty
+    LLAMA_API llama_pos llama_memory_seq_pos_max(
+            llama_memory_t mem,
+              llama_seq_id seq_id);
+
+    // Check if the memory supports shifting
+    LLAMA_API bool llama_memory_can_shift(llama_memory_t mem);
+
+    //
+    // KV cache for self-attention (TODO: deprecate in favor of llama_memory)
     //
 
     // Returns the number of tokens in the KV cache (slow, use only for debug)
@@ -622,86 +712,95 @@ extern "C" {
                "Use llama_kv_self_seq_pos_max() and llama_kv_self_seq_pos_min() instead (https://github.com/ggml-org/llama.cpp/issues/13793)");
 
     // Clear the KV cache - both cell info is erased and KV data is zeroed
-    LLAMA_API void llama_kv_self_clear(
-            struct llama_context * ctx);
+    DEPRECATED(LLAMA_API void llama_kv_self_clear(
+                struct llama_context * ctx),
+            "Use llama_memory_clear() instead");
 
     // Removes all tokens that belong to the specified sequence and have positions in [p0, p1)
     // Returns false if a partial sequence cannot be removed. Removing a whole sequence never fails
     // seq_id < 0 : match any sequence
     // p0 < 0     : [0,  p1]
     // p1 < 0     : [p0, inf)
-    LLAMA_API bool llama_kv_self_seq_rm(
+    DEPRECATED(LLAMA_API bool llama_kv_self_seq_rm(
             struct llama_context * ctx,
                     llama_seq_id   seq_id,
                        llama_pos   p0,
-                       llama_pos   p1);
+                       llama_pos   p1),
+            "Use llama_memory_seq_rm() instead");
 
     // Copy all tokens that belong to the specified sequence to another sequence
     // 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_self_seq_cp(
+    DEPRECATED(LLAMA_API void llama_kv_self_seq_cp(
             struct llama_context * ctx,
                     llama_seq_id   seq_id_src,
                     llama_seq_id   seq_id_dst,
                        llama_pos   p0,
-                       llama_pos   p1);
+                       llama_pos   p1),
+            "Use llama_memory_seq_cp() instead");
 
     // Removes all tokens that do not belong to the specified sequence
-    LLAMA_API void llama_kv_self_seq_keep(
+    DEPRECATED(LLAMA_API void llama_kv_self_seq_keep(
             struct llama_context * ctx,
-                    llama_seq_id   seq_id);
+                    llama_seq_id   seq_id),
+            "Use llama_memory_seq_keep() instead");
 
     // 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()
     // p0 < 0 : [0,  p1]
     // p1 < 0 : [p0, inf)
-    LLAMA_API void llama_kv_self_seq_add(
+    DEPRECATED(LLAMA_API void llama_kv_self_seq_add(
             struct llama_context * ctx,
                     llama_seq_id   seq_id,
                        llama_pos   p0,
                        llama_pos   p1,
-                       llama_pos   delta);
+                       llama_pos   delta),
+            "Use llama_memory_seq_add() instead");
 
     // 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()
     // p0 < 0 : [0,  p1]
     // p1 < 0 : [p0, inf)
-    LLAMA_API void llama_kv_self_seq_div(
+    DEPRECATED(void llama_kv_self_seq_div(
             struct llama_context * ctx,
                     llama_seq_id   seq_id,
                        llama_pos   p0,
                        llama_pos   p1,
-                             int   d);
+                             int   d),
+            "Use llama_memory_seq_div() instead");
 
     // Returns the smallest position present in the KV cache for the specified sequence
     // This is typically non-zero only for SWA caches
     // Note that all positions in the range [pos_min, pos_max] are guaranteed to be present in the KV cache
     // Return -1 if the sequence is empty
-    LLAMA_API llama_pos llama_kv_self_seq_pos_min(
+    DEPRECATED(LLAMA_API llama_pos llama_kv_self_seq_pos_min(
             struct llama_context * ctx,
-                    llama_seq_id   seq_id);
+                    llama_seq_id   seq_id),
+            "Use llama_memory_seq_pos_min() instead");
 
     // Returns the largest position present in the KV cache for the specified sequence
     // Note that all positions in the range [pos_min, pos_max] are guaranteed to be present in the KV cache
     // Return -1 if the sequence is empty
-    LLAMA_API llama_pos llama_kv_self_seq_pos_max(
+    DEPRECATED(LLAMA_API llama_pos llama_kv_self_seq_pos_max(
             struct llama_context * ctx,
-                    llama_seq_id   seq_id);
+                    llama_seq_id   seq_id),
+            "Use llama_memory_seq_pos_max() instead");
 
     // Defragment the KV cache
     // This will be applied:
     //   - lazily on next llama_decode()
-    LLAMA_API DEPRECATED(void llama_kv_self_defrag(struct llama_context * ctx),
+    DEPRECATED(LLAMA_API void llama_kv_self_defrag(struct llama_context * ctx),
             "simply remove this call, the context will automatically decide when to do a defragmentation based on 'defrag_thold'");
 
     // Check if the context supports KV cache shifting
-    LLAMA_API bool llama_kv_self_can_shift(const struct llama_context * ctx);
+    DEPRECATED(LLAMA_API bool llama_kv_self_can_shift(const struct llama_context * ctx),
+            "use llama_memory_can_shift() instead");
 
     // Apply the KV cache updates (such as K-shifts, defragmentation, etc.)
-    LLAMA_API DEPRECATED(void llama_kv_self_update(struct llama_context * ctx),
+    DEPRECATED(LLAMA_API void llama_kv_self_update(struct llama_context * ctx),
             "simply remove this call, updates are applied lazily on the next llama_decode()");
 
     //
@@ -709,7 +808,7 @@ extern "C" {
     //
 
     // Returns the *actual* size in bytes of the state
-    // (logits, embedding and kv_cache)
+    // (logits, embedding and memory)
     // Only use when saving the state, not when restoring it, otherwise the size may be too small.
     LLAMA_API size_t llama_state_get_size(struct llama_context * ctx);
     LLAMA_API DEPRECATED(size_t llama_get_state_size(struct llama_context * ctx),
@@ -765,12 +864,12 @@ extern "C" {
                           size_t   n_token_count),
         "use llama_state_save_file instead");
 
-    // Get the exact size needed to copy the KV cache of a single sequence
+    // Get the exact size needed to copy the state of a single sequence
     LLAMA_API size_t llama_state_seq_get_size(
             struct llama_context * ctx,
                     llama_seq_id   seq_id);
 
-    // Copy the KV cache of a single sequence into the specified buffer
+    // Copy the state of a single sequence into the specified buffer
     LLAMA_API size_t llama_state_seq_get_data(
             struct llama_context * ctx,
                          uint8_t * dst,
@@ -836,21 +935,23 @@ extern "C" {
     // For encode-decoder contexts, processes the batch using the encoder.
     // Can store the encoder output internally for later use by the decoder's cross-attention layers.
     //   0 - success
-    // < 0 - error. the KV cache state is restored to the state before this call
+    // < 0 - error. the memory state is restored to the state before this call
     LLAMA_API int32_t llama_encode(
             struct llama_context * ctx,
               struct llama_batch   batch);
 
     // Process a batch of tokens.
-    // Requires KV cache.
+    // Requires the context to have a memory.
     // For encode-decoder contexts, processes the batch using the decoder.
     // Positive return values does not mean a fatal error, but rather a warning.
-    // Upon non-zero return values, the KV cache state is restored to the state before this call
+    // Upon fatal-error or abort, the ubatches that managed to be been processed will remain in the memory state of the context
+    //   To handle this correctly, query the memory state using llama_memory_seq_pos_min() and llama_memory_seq_pos_max()
+    // Upon other return values, the memory state is restored to the state before this call
     //    0 - success
     //    1 - could not find a KV slot for the batch (try reducing the size of the batch or increase the context)
-    //    2 - aborted
+    //    2 - aborted     (processed ubatches will remain in the context's memory)
     //   -1 - invalid input batch
-    // < -1 - error
+    // < -1 - fatal error (processed ubatches will remain in the context's memory)
     LLAMA_API int32_t llama_decode(
             struct llama_context * ctx,
               struct llama_batch   batch);
@@ -866,8 +967,8 @@ extern "C" {
     // Get the number of threads used for prompt and batch processing (multiple token).
     LLAMA_API int32_t llama_n_threads_batch(struct llama_context * ctx);
 
-    // Set whether the model is in embeddings mode or not
-    // If true, embeddings will be returned but logits will not
+    // Set whether the context outputs embeddings or not
+    // TODO: rename to avoid confusion with llama_get_embeddings()
     LLAMA_API void llama_set_embeddings(struct llama_context * ctx, bool embeddings);
 
     // Set whether to use causal attention or not
@@ -916,7 +1017,7 @@ extern "C" {
 
     // Get the embeddings for a sequence id
     // Returns NULL if pooling_type is LLAMA_POOLING_TYPE_NONE
-    // when pooling_type == LLAMA_POOLING_TYPE_RANK, returns float[1] with the rank of the sequence
+    // when pooling_type == LLAMA_POOLING_TYPE_RANK, returns float[n_cls_out] with the rank(s) of the sequence
     // otherwise: float[n_embd] (1-dimensional)
     LLAMA_API float * llama_get_embeddings_seq(struct llama_context * ctx, llama_seq_id seq_id);
 
@@ -946,6 +1047,7 @@ extern "C" {
 
     LLAMA_API bool llama_vocab_get_add_bos(const struct llama_vocab * vocab);
     LLAMA_API bool llama_vocab_get_add_eos(const struct llama_vocab * vocab);
+    LLAMA_API bool llama_vocab_get_add_sep(const struct llama_vocab * vocab);
 
     LLAMA_API llama_token llama_vocab_fim_pre(const struct llama_vocab * vocab);
     LLAMA_API llama_token llama_vocab_fim_suf(const struct llama_vocab * vocab);
@@ -989,6 +1091,7 @@ extern "C" {
     /// @param tokens The tokens pointer must be large enough to hold the resulting tokens.
     /// @return Returns the number of tokens on success, no more than n_tokens_max
     /// @return Returns a negative number on failure - the number of tokens that would have been returned
+    /// @return Returns INT32_MIN on overflow (e.g., tokenization result size exceeds int32_t limit)
     /// @param add_special Allow to add BOS and EOS tokens if model is configured to do so.
     /// @param parse_special Allow tokenizing special and/or control tokens which otherwise are not exposed and treated
     ///                      as plaintext. Does not insert a leading space.
diff --git a/models/templates/Mistral-Small-3.2-24B-Instruct-2506.jinja b/models/templates/Mistral-Small-3.2-24B-Instruct-2506.jinja
new file mode 100644
index 000000000..19a3eaee4
--- /dev/null
+++ b/models/templates/Mistral-Small-3.2-24B-Instruct-2506.jinja
@@ -0,0 +1,124 @@
+{%- set today = strftime_now("%Y-%m-%d") %}
+{%- set default_system_message = "You are Mistral Small 3, a Large Language Model (LLM) created by Mistral AI, a French startup headquartered in Paris.\nYour knowledge base was last updated on 2023-10-01. The current date is " + today + ".\n\nWhen you're not sure about some information or when the user's request requires up-to-date or specific data, you must use the available tools to fetch the information. Do not hesitate to use tools whenever they can provide a more accurate or complete response. If no relevant tools are available, then clearly state that you don't have the information and avoid making up anything.
+
+If the user's question is not clear, ambiguous, or does not provide enough context for you to accurately answer the question, you do not try to answer it right away and you rather ask the user to clarify their request (e.g. \"What are some good restaurants around me?\" => \"Where are you?\" or \"When is the next flight to Tokyo\" => \"Where do you travel from?\").
+You are always very attentive to dates, and when asked about information at specific dates, you discard information that is at another date.
+You follow these instructions in all languages, and always respond to the user in the language they use or request.
+Next sections describe the capabilities that you have.
+
+# WEB BROWSING INSTRUCTIONS
+
+You cannot perform any web search or access internet to open URLs, links etc. If it seems like the user is expecting you to do so, you clarify the situation and ask the user to copy paste the text directly in the chat.
+
+# MULTI-MODAL INSTRUCTIONS
+
+You have the ability to read images, but you cannot generate images. You also cannot transcribe audio files or videos.
+You cannot read nor transcribe audio files or videos.
+
+# TOOL CALLING INSTRUCTIONS
+
+You may have access to tools that you can use to fetch information or perform actions. You must use these tools in the following situations:
+
+1. When the request requires up-to-date information.
+2. When the request requires specific data that you do not have in your knowledge base.
+3. When the request involves actions that you cannot perform without tools.
+
+Always prioritize using tools to provide the most accurate and helpful response. If tools are not available, inform the user that you cannot perform the requested action at the moment." %}
+
+{{- bos_token }}
+
+{%- set system_prompt = default_system_message %}
+{%- set loop_messages = messages %}
+
+{%- if not tools is defined %}
+    {%- set tools = none %}
+{%- endif %}
+
+{%- if messages|length > 0 and messages[0]['role'] == 'system' %}
+    {%- if messages[0]['content'] is string %}
+        {%- set system_prompt = messages[0]['content'] %}
+    {%- else %}
+        {%- set system_prompt = messages[0]['content'][0]['text'] %}
+    {%- endif %}
+    {%- set loop_messages = messages[1:] %}
+{%- endif %}
+
+{%- set user_messages = loop_messages | selectattr("role", "equalto", "user") | list %}
+
+{%- set ns = namespace(index=0) %}
+{%- for message in loop_messages %}
+    {%- if not (message.role == "tool" or (message.get('tool_calls'))) %}
+        {%- if (message["role"] == "user") != (ns.index % 2 == 0) %}
+            {{- raise_exception("After the optional system message, conversation roles must alternate user/assistant/user/assistant/...") }}
+        {%- endif %}
+        {%- set ns.index = ns.index + 1 %}
+    {%- endif %}
+{%- endfor %}
+
+{{- '[SYSTEM_PROMPT]' + system_prompt + '[/SYSTEM_PROMPT]' }}
+
+{%- for message in loop_messages %}
+    {%- if message['role'] == 'system' %}
+        {%- if message['content'] is string %}
+            {{- '[SYSTEM_PROMPT]' + message['content'] + '[/SYSTEM_PROMPT]' }}
+        {%- else %}
+            {{- '[SYSTEM_PROMPT]' + message['content'][0]['text'] + '[/SYSTEM_PROMPT]' }}
+        {%- endif %}
+    {%- elif message['role'] == 'user' %}
+        {%- if tools is not none and (message == user_messages[-1]) %}
+            {{- '[AVAILABLE_TOOLS]' + tools|tojson + '[/AVAILABLE_TOOLS]' }}
+        {%- endif %}
+        {{- '[INST]' }}
+        {%- if message['content'] is string %}
+            {{- message['content'] }}
+        {%- else %}
+            {%- for block in message['content'] %}
+                {%- if block['type'] == 'text' %}
+                    {{- block['text'] }}
+                {%- elif block['type'] in ['image', 'image_url'] %}
+                    {{- '[IMG]' }}
+                {%- else %}
+                    {{- raise_exception('Only text and image blocks are supported in message content!') }}
+                {%- endif %}
+            {%- endfor %}
+        {%- endif %}
+        {{- '[/INST]' }}
+    {%- elif message['role'] == 'assistant' %}
+        {%- if message.get('tool_calls') %}
+            {%- for tool_call in message.tool_calls %}
+                {{- '[TOOL_CALLS]' + tool_call.function.name }}
+                {%- if not tool_call.id is defined or tool_call.id is not string or tool_call.id|length != 9 %}
+                    {{- raise_exception("Tool call IDs should be alphanumeric strings with length 9!") }}
+                {%- endif %}
+                {{- '[CALL_ID]' + tool_call.id }}
+                {{- '[ARGS]' + tool_call['function']['arguments']|tojson }}
+            {%- endfor %}
+            {{- eos_token }}
+        {%- elif message['content'] is string %}
+            {{- message['content'] + eos_token }}
+        {%- else %}
+            {%- for block in message['content'] %}
+                {%- if block['type'] == 'text' %}
+                    {{- block['text'] }}
+                {%- elif block['type'] in ['image', 'image_url'] %}
+                    {{- '[IMG]' }}
+                {%- else %}
+                    {{- raise_exception('Only text and image blocks are supported in assistant content!') }}
+                {%- endif %}
+            {%- endfor %}
+            {{- eos_token }}
+        {%- endif %}
+    {%- elif message['role'] == 'tool_results' or message['role'] == 'tool' %}
+        {%- if message.content is defined and message.content.content is defined %}
+            {%- set content = message.content.content %}
+        {%- else %}
+            {%- set content = message.content %}
+        {%- endif %}
+        {%- if not message.tool_call_id is defined or message.tool_call_id is not string or message['tool_call_id']|length != 9 %}
+            {{- raise_exception("Tool call IDs should be alphanumeric strings with length 9!") }}
+        {%- endif %}
+        {{- '[TOOL_RESULTS]' + message.tool_call_id + '[TOOL_CONTENT]' + content|string + '[/TOOL_RESULTS]' }}
+    {%- else %}
+        {{- raise_exception('Only system, user, assistant, and tool roles are supported!') }}
+    {%- endif %}
+{%- endfor %}
diff --git a/requirements/requirements-compare-llama-bench.txt b/requirements/requirements-compare-llama-bench.txt
index e0aaa3204..d87e897e1 100644
--- a/requirements/requirements-compare-llama-bench.txt
+++ b/requirements/requirements-compare-llama-bench.txt
@@ -1,2 +1,3 @@
 tabulate~=0.9.0
 GitPython~=3.1.43
+matplotlib~=3.10.0
diff --git a/scripts/apple/validate-apps.sh b/scripts/apple/validate-apps.sh
index a571aa6fc..f0475758c 100755
--- a/scripts/apple/validate-apps.sh
+++ b/scripts/apple/validate-apps.sh
@@ -1,4 +1,4 @@
-#!/bin/bash
+#!/usr/bin/env bash
 ./scripts/apple/validate-ios.sh
 ./scripts/apple/validate-macos.sh
 ./scripts/apple/validate-visionos.sh
diff --git a/scripts/apple/validate-ios.sh b/scripts/apple/validate-ios.sh
index 7bda1b972..50800d84a 100755
--- a/scripts/apple/validate-ios.sh
+++ b/scripts/apple/validate-ios.sh
@@ -1,4 +1,4 @@
-#!/bin/bash
+#!/usr/bin/env bash
 # validate-ios.sh - Validate iOS Application with embedded llama.xcframework using SwiftUI
 
 # Authentication options (optional) (can be set via environment variables)
diff --git a/scripts/apple/validate-macos.sh b/scripts/apple/validate-macos.sh
index 6dc28e694..fa800ee68 100755
--- a/scripts/apple/validate-macos.sh
+++ b/scripts/apple/validate-macos.sh
@@ -1,4 +1,4 @@
-#!/bin/bash
+#!/usr/bin/env bash
 # validate-macos.sh - Validate macOS Application with embedded llama.xcframework using SwiftUI
 
 # Authentication options (optional) (can be set via environment variables)
diff --git a/scripts/apple/validate-tvos.sh b/scripts/apple/validate-tvos.sh
index 6120189e8..b4da69874 100755
--- a/scripts/apple/validate-tvos.sh
+++ b/scripts/apple/validate-tvos.sh
@@ -1,4 +1,4 @@
-#!/bin/bash
+#!/usr/bin/env bash
 # validate-tvos.sh - Validate tvOS Application with embedded llama.xcframework using SwiftUI
 
 # Authentication options (optional) (can be set via environment variables)
diff --git a/scripts/apple/validate-visionos.sh b/scripts/apple/validate-visionos.sh
index a18ddcce4..bbdec6602 100755
--- a/scripts/apple/validate-visionos.sh
+++ b/scripts/apple/validate-visionos.sh
@@ -1,4 +1,4 @@
-#!/bin/bash
+#!/usr/bin/env bash
 # validate-visionos.sh - Validate visionOS Application with embedded llama.xcframework using SwiftUI
 
 # Authentication options (optional) (can be set via environment variables)
diff --git a/scripts/check-requirements.sh b/scripts/check-requirements.sh
index 4c3b05f68..da2357d76 100755
--- a/scripts/check-requirements.sh
+++ b/scripts/check-requirements.sh
@@ -1,4 +1,4 @@
-#!/bin/bash
+#!/usr/bin/env bash
 set -euo pipefail
 
 #
diff --git a/scripts/ci-run.sh b/scripts/ci-run.sh
index 06b5d9c6e..5877a7eda 100755
--- a/scripts/ci-run.sh
+++ b/scripts/ci-run.sh
@@ -1,4 +1,4 @@
-#!/bin/bash
+#!/usr/bin/env bash
 set -euo pipefail
 this=$(realpath "$0"); readonly this
 cd "$(dirname "$this")"
diff --git a/scripts/compare-commits.sh b/scripts/compare-commits.sh
index 94a8eceb3..051a7a098 100755
--- a/scripts/compare-commits.sh
+++ b/scripts/compare-commits.sh
@@ -1,4 +1,4 @@
-#!/bin/bash
+#!/usr/bin/env bash
 
 if [ $# -lt 2 ]; then
     echo "usage: ./scripts/compare-commits.sh <commit1> <commit2> [additional llama-bench arguments]"
diff --git a/scripts/compare-llama-bench.py b/scripts/compare-llama-bench.py
index a1013c3b7..30e3cf864 100755
--- a/scripts/compare-llama-bench.py
+++ b/scripts/compare-llama-bench.py
@@ -19,6 +19,7 @@ except ImportError as e:
     print("the following Python libraries are required: GitPython, tabulate.") # noqa: NP100
     raise e
 
+
 logger = logging.getLogger("compare-llama-bench")
 
 # All llama-bench SQL fields
@@ -122,11 +123,15 @@ help_s = (
 parser.add_argument("--check", action="store_true", help="check if all required Python libraries are installed")
 parser.add_argument("-s", "--show", help=help_s)
 parser.add_argument("--verbose", action="store_true", help="increase output verbosity")
+parser.add_argument("--plot", help="generate a performance comparison plot and save to specified file (e.g., plot.png)")
+parser.add_argument("--plot_x", help="parameter to use as x axis for plotting (default: n_depth)", default="n_depth")
+parser.add_argument("--plot_log_scale", action="store_true", help="use log scale for x axis in plots (off by default)")
 
 known_args, unknown_args = parser.parse_known_args()
 
 logging.basicConfig(level=logging.DEBUG if known_args.verbose else logging.INFO)
 
+
 if known_args.check:
     # Check if all required Python libraries are installed. Would have failed earlier if not.
     sys.exit(0)
@@ -499,7 +504,6 @@ else:
 
 name_compare = bench_data.get_commit_name(hexsha8_compare)
 
-
 # If the user provided columns to group the results by, use them:
 if known_args.show is not None:
     show = known_args.show.split(",")
@@ -544,6 +548,14 @@ else:
             show.remove(prop)
         except ValueError:
             pass
+
+    # Add plot_x parameter to parameters to show if it's not already present:
+    if known_args.plot:
+        for k, v in PRETTY_NAMES.items():
+            if v == known_args.plot_x and k not in show:
+                show.append(k)
+                break
+
     rows_show = bench_data.get_rows(show, hexsha8_baseline, hexsha8_compare)
 
 if not rows_show:
@@ -600,6 +612,161 @@ if "gpu_info" in show:
 headers  = [PRETTY_NAMES[p] for p in show]
 headers += ["Test", f"t/s {name_baseline}", f"t/s {name_compare}", "Speedup"]
 
+if known_args.plot:
+    def create_performance_plot(table_data: list[list[str]], headers: list[str], baseline_name: str, compare_name: str, output_file: str, plot_x_param: str, log_scale: bool = False):
+        try:
+            import matplotlib.pyplot as plt
+            import matplotlib
+            matplotlib.use('Agg')
+        except ImportError as e:
+            logger.error("matplotlib is required for --plot.")
+            raise e
+
+        data_headers = headers[:-4] # Exclude the last 4 columns (Test, baseline t/s, compare t/s, Speedup)
+        plot_x_index = None
+        plot_x_label = plot_x_param
+
+        if plot_x_param not in ["n_prompt", "n_gen", "n_depth"]:
+            pretty_name = PRETTY_NAMES.get(plot_x_param, plot_x_param)
+            if pretty_name in data_headers:
+                plot_x_index = data_headers.index(pretty_name)
+                plot_x_label = pretty_name
+            elif plot_x_param in data_headers:
+                plot_x_index = data_headers.index(plot_x_param)
+                plot_x_label = plot_x_param
+            else:
+                logger.error(f"Parameter '{plot_x_param}' not found in current table columns. Available columns: {', '.join(data_headers)}")
+                return
+
+        grouped_data = {}
+
+        for i, row in enumerate(table_data):
+            group_key_parts = []
+            test_name = row[-4]
+
+            base_test = ""
+            x_value = None
+
+            if plot_x_param in ["n_prompt", "n_gen", "n_depth"]:
+                for j, val in enumerate(row[:-4]):
+                    header_name = data_headers[j]
+                    if val is not None and str(val).strip():
+                        group_key_parts.append(f"{header_name}={val}")
+
+                if plot_x_param == "n_prompt" and "pp" in test_name:
+                    base_test = test_name.split("@")[0]
+                    x_value = base_test
+                elif plot_x_param == "n_gen" and "tg" in test_name:
+                    x_value = test_name.split("@")[0]
+                elif plot_x_param == "n_depth" and "@d" in test_name:
+                    base_test = test_name.split("@d")[0]
+                    x_value = int(test_name.split("@d")[1])
+                else:
+                    base_test = test_name
+
+                if base_test.strip():
+                    group_key_parts.append(f"Test={base_test}")
+            else:
+                for j, val in enumerate(row[:-4]):
+                    if j != plot_x_index:
+                        header_name = data_headers[j]
+                        if val is not None and str(val).strip():
+                            group_key_parts.append(f"{header_name}={val}")
+                    else:
+                        x_value = val
+
+                group_key_parts.append(f"Test={test_name}")
+
+            group_key = tuple(group_key_parts)
+
+            if group_key not in grouped_data:
+                grouped_data[group_key] = []
+
+            grouped_data[group_key].append({
+                'x_value': x_value,
+                'baseline': float(row[-3]),
+                'compare': float(row[-2]),
+                'speedup': float(row[-1])
+            })
+
+        if not grouped_data:
+            logger.error("No data available for plotting")
+            return
+
+        def make_axes(num_groups, max_cols=2, base_size=(8, 4)):
+            from math import ceil
+            cols = 1 if num_groups == 1 else min(max_cols, num_groups)
+            rows = ceil(num_groups / cols)
+
+            # Scale figure size by grid dimensions
+            w, h = base_size
+            fig, ax_arr = plt.subplots(rows, cols,
+                                       figsize=(w * cols, h * rows),
+                                       squeeze=False)
+
+            axes = ax_arr.flatten()[:num_groups]
+            return fig, axes
+
+        num_groups = len(grouped_data)
+        fig, axes = make_axes(num_groups)
+
+        plot_idx = 0
+
+        for group_key, points in grouped_data.items():
+            if plot_idx >= len(axes):
+                break
+            ax = axes[plot_idx]
+
+            try:
+                points_sorted = sorted(points, key=lambda p: float(p['x_value']) if p['x_value'] is not None else 0)
+                x_values = [float(p['x_value']) if p['x_value'] is not None else 0 for p in points_sorted]
+            except ValueError:
+                points_sorted = sorted(points, key=lambda p: group_key)
+                x_values = [p['x_value'] for p in points_sorted]
+
+            baseline_vals = [p['baseline'] for p in points_sorted]
+            compare_vals = [p['compare'] for p in points_sorted]
+
+            ax.plot(x_values, baseline_vals, 'o-', color='skyblue',
+                    label=f'{baseline_name}', linewidth=2, markersize=6)
+            ax.plot(x_values, compare_vals, 's--', color='lightcoral', alpha=0.8,
+                    label=f'{compare_name}', linewidth=2, markersize=6)
+
+            if log_scale:
+                ax.set_xscale('log', base=2)
+                unique_x = sorted(set(x_values))
+                ax.set_xticks(unique_x)
+                ax.set_xticklabels([str(int(x)) for x in unique_x])
+
+            title_parts = []
+            for part in group_key:
+                if '=' in part:
+                    key, value = part.split('=', 1)
+                    title_parts.append(f"{key}: {value}")
+
+            title = ', '.join(title_parts) if title_parts else "Performance comparison"
+
+            ax.set_xlabel(plot_x_label, fontsize=12, fontweight='bold')
+            ax.set_ylabel('Tokens per second (t/s)', fontsize=12, fontweight='bold')
+            ax.set_title(title, fontsize=12, fontweight='bold')
+            ax.legend(loc='best', fontsize=10)
+            ax.grid(True, alpha=0.3)
+
+            plot_idx += 1
+
+        for i in range(plot_idx, len(axes)):
+            axes[i].set_visible(False)
+
+        fig.suptitle(f'Performance comparison: {compare_name} vs. {baseline_name}',
+                     fontsize=14, fontweight='bold')
+        fig.subplots_adjust(top=1)
+
+        plt.tight_layout()
+        plt.savefig(output_file, dpi=300, bbox_inches='tight')
+        plt.close()
+
+    create_performance_plot(table, headers, name_baseline, name_compare, known_args.plot, known_args.plot_x, known_args.plot_log_scale)
+
 print(tabulate( # noqa: NP100
     table,
     headers=headers,
diff --git a/scripts/debug-test.sh b/scripts/debug-test.sh
index c6c1e988a..7e9e8421b 100755
--- a/scripts/debug-test.sh
+++ b/scripts/debug-test.sh
@@ -1,4 +1,4 @@
-#!/bin/bash
+#!/usr/bin/env bash
 
 PROG=${0##*/}
 build_dir="build-ci-debug"
diff --git a/scripts/gen-authors.sh b/scripts/gen-authors.sh
index 3ef8391cc..73e7b386f 100755
--- a/scripts/gen-authors.sh
+++ b/scripts/gen-authors.sh
@@ -1,4 +1,4 @@
-#!/bin/bash
+#!/usr/bin/env bash
 
 printf "# date: $(date)\n" > AUTHORS
 printf "# this file is auto-generated by scripts/gen-authors.sh\n\n" >> AUTHORS
diff --git a/scripts/get-hellaswag.sh b/scripts/get-hellaswag.sh
index 4e1b1cc15..484e56fd8 100755
--- a/scripts/get-hellaswag.sh
+++ b/scripts/get-hellaswag.sh
@@ -1,4 +1,4 @@
-#!/bin/bash
+#!/usr/bin/env bash
 
 wget https://raw.githubusercontent.com/klosax/hellaswag_text_data/main/hellaswag_val_full.txt
 
diff --git a/scripts/get-pg.sh b/scripts/get-pg.sh
index b027793e1..f180bf834 100755
--- a/scripts/get-pg.sh
+++ b/scripts/get-pg.sh
@@ -1,4 +1,4 @@
-#!/bin/bash
+#!/usr/bin/env bash
 
 function usage {
     echo "usage: <n>$0"
diff --git a/scripts/get-wikitext-103.sh b/scripts/get-wikitext-103.sh
index 9c65fafbc..244a371ba 100755
--- a/scripts/get-wikitext-103.sh
+++ b/scripts/get-wikitext-103.sh
@@ -1,4 +1,4 @@
-#!/bin/bash
+#!/usr/bin/env bash
 
 wget https://s3.amazonaws.com/research.metamind.io/wikitext/wikitext-103-raw-v1.zip
 
diff --git a/scripts/get-wikitext-2.sh b/scripts/get-wikitext-2.sh
index 5f3845ef5..67b0b0118 100755
--- a/scripts/get-wikitext-2.sh
+++ b/scripts/get-wikitext-2.sh
@@ -1,4 +1,4 @@
-#!/bin/bash
+#!/usr/bin/env bash
 
 wget https://huggingface.co/datasets/ggml-org/ci/resolve/main/wikitext-2-raw-v1.zip
 unzip wikitext-2-raw-v1.zip
diff --git a/scripts/get-winogrande.sh b/scripts/get-winogrande.sh
index f1fc0e2d4..2b48b1175 100755
--- a/scripts/get-winogrande.sh
+++ b/scripts/get-winogrande.sh
@@ -1,4 +1,4 @@
-#!/bin/bash
+#!/usr/bin/env bash
 
 wget https://huggingface.co/datasets/ikawrakow/winogrande-eval-for-llama.cpp/raw/main/winogrande-debiased-eval.csv
 
diff --git a/scripts/hf.sh b/scripts/hf.sh
index b251925fa..e41b9053a 100755
--- a/scripts/hf.sh
+++ b/scripts/hf.sh
@@ -1,4 +1,4 @@
-#!/bin/bash
+#!/usr/bin/env bash
 #
 # Shortcut for downloading HF models
 #
diff --git a/scripts/qnt-all.sh b/scripts/qnt-all.sh
index bc43738a2..dc04670df 100755
--- a/scripts/qnt-all.sh
+++ b/scripts/qnt-all.sh
@@ -1,4 +1,4 @@
-#!/bin/bash
+#!/usr/bin/env bash
 
 qnt=(q8_0 q6_k q5_k q5_1 q5_0 q4_k q4_1 q4_0 q3_k q2_k)
 args=""
diff --git a/scripts/run-all-perf.sh b/scripts/run-all-perf.sh
index 6384e364d..b7de764ff 100755
--- a/scripts/run-all-perf.sh
+++ b/scripts/run-all-perf.sh
@@ -1,4 +1,4 @@
-#!/bin/bash
+#!/usr/bin/env bash
 
 qnt=(f16 q8_0 q6_k q5_k q5_1 q5_0 q4_k q4_1 q4_0 q3_k q2_k)
 args="-ngl 999 -n 64 -p 512"
diff --git a/scripts/run-all-ppl.sh b/scripts/run-all-ppl.sh
index e15f74f1b..918ecda27 100755
--- a/scripts/run-all-ppl.sh
+++ b/scripts/run-all-ppl.sh
@@ -1,4 +1,4 @@
-#!/bin/bash
+#!/usr/bin/env bash
 
 qnt=(f16 q8_0 q6_k q5_k q5_1 q5_0 q4_k q4_1 q4_0 q3_k q2_k)
 args="-ngl 999 -t 8"
diff --git a/scripts/sync-ggml-am.sh b/scripts/sync-ggml-am.sh
index 204354209..29d30e0a1 100755
--- a/scripts/sync-ggml-am.sh
+++ b/scripts/sync-ggml-am.sh
@@ -1,4 +1,4 @@
-#!/bin/bash
+#!/usr/bin/env bash
 #
 # Synchronize ggml changes to llama.cpp
 #
@@ -83,7 +83,6 @@ while read c; do
         src/ggml-cpu/* \
         src/ggml-cuda/* \
         src/ggml-hip/* \
-        src/ggml-kompute/* \
         src/ggml-metal/* \
         src/ggml-musa/* \
         src/ggml-opencl/* \
@@ -141,7 +140,6 @@ if [ -f $SRC_LLAMA/ggml-src.patch ]; then
     # src/ggml-cpu/*       -> ggml/src/ggml-cpu/*
     # src/ggml-cuda/*      -> ggml/src/ggml-cuda/*
     # src/ggml-hip/*       -> ggml/src/ggml-hip/*
-    # src/ggml-kompute/*   -> ggml/src/ggml-kompute/*
     # src/ggml-metal/*     -> ggml/src/ggml-metal/*
     # src/ggml-musa/*      -> ggml/src/ggml-musa/*
     # src/ggml-opencl/*    -> ggml/src/ggml-opencl/*
@@ -174,7 +172,6 @@ if [ -f $SRC_LLAMA/ggml-src.patch ]; then
         -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' \
diff --git a/scripts/sync-ggml.last b/scripts/sync-ggml.last
index aa0fb8fb0..4157e1f53 100644
--- a/scripts/sync-ggml.last
+++ b/scripts/sync-ggml.last
@@ -1 +1 @@
-94a83ba5a725ae2aee79df75dd99b2119d0478cc
+0405219965324e11a29b6aadfe22a6d66131978f
diff --git a/scripts/sync-ggml.sh b/scripts/sync-ggml.sh
index aa1a46b4b..9b98329e0 100755
--- a/scripts/sync-ggml.sh
+++ b/scripts/sync-ggml.sh
@@ -1,4 +1,4 @@
-#!/bin/bash
+#!/usr/bin/env bash
 
 cp -rpv ../ggml/CMakeLists.txt       ./ggml/CMakeLists.txt
 cp -rpv ../ggml/src/CMakeLists.txt   ./ggml/src/CMakeLists.txt
@@ -15,7 +15,6 @@ cp -rpv ../ggml/src/ggml-cann/*    ./ggml/src/ggml-cann/
 cp -rpv ../ggml/src/ggml-cpu/*     ./ggml/src/ggml-cpu/
 cp -rpv ../ggml/src/ggml-cuda/*    ./ggml/src/ggml-cuda/
 cp -rpv ../ggml/src/ggml-hip/*     ./ggml/src/ggml-hip/
-cp -rpv ../ggml/src/ggml-kompute/* ./ggml/src/ggml-kompute/
 cp -rpv ../ggml/src/ggml-metal/*   ./ggml/src/ggml-metal/
 cp -rpv ../ggml/src/ggml-musa/*    ./ggml/src/ggml-musa/
 cp -rpv ../ggml/src/ggml-opencl/*  ./ggml/src/ggml-opencl/
diff --git a/scripts/tool_bench.sh b/scripts/tool_bench.sh
index 6c7616a88..05b41d2f1 100755
--- a/scripts/tool_bench.sh
+++ b/scripts/tool_bench.sh
@@ -1,4 +1,4 @@
-#!/bin/bash
+#!/usr/bin/env bash
 set -euo pipefail
 
 cmake --build build -j
diff --git a/src/CMakeLists.txt b/src/CMakeLists.txt
index d20bd4fe2..8f9cd6524 100644
--- a/src/CMakeLists.txt
+++ b/src/CMakeLists.txt
@@ -20,11 +20,11 @@ add_library(llama
             llama-hparams.cpp
             llama-impl.cpp
             llama-io.cpp
-            llama-kv-cache.cpp
             llama-kv-cache-unified.cpp
             llama-kv-cache-unified-iswa.cpp
-            llama-kv-cache-recurrent.cpp
             llama-memory.cpp
+            llama-memory-hybrid.cpp
+            llama-memory-recurrent.cpp
             llama-mmap.cpp
             llama-model-loader.cpp
             llama-model-saver.cpp
diff --git a/src/llama-arch.cpp b/src/llama-arch.cpp
index 0ef81054b..cbec31954 100644
--- a/src/llama-arch.cpp
+++ b/src/llama-arch.cpp
@@ -20,6 +20,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_BERT,             "bert"             },
     { LLM_ARCH_NOMIC_BERT,       "nomic-bert"       },
     { LLM_ARCH_NOMIC_BERT_MOE,   "nomic-bert-moe"   },
+    { LLM_ARCH_NEO_BERT,         "neo-bert"         },
     { LLM_ARCH_JINA_BERT_V2,     "jina-bert-v2"     },
     { LLM_ARCH_JINA_BERT_V3,     "jina-bert-v3"     },
     { LLM_ARCH_BLOOM,            "bloom"            },
@@ -42,8 +43,10 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_GEMMA,            "gemma"            },
     { LLM_ARCH_GEMMA2,           "gemma2"           },
     { LLM_ARCH_GEMMA3,           "gemma3"           },
+    { LLM_ARCH_GEMMA3N,          "gemma3n"          },
     { LLM_ARCH_STARCODER2,       "starcoder2"       },
     { LLM_ARCH_MAMBA,            "mamba"            },
+    { LLM_ARCH_MAMBA2,           "mamba2"           },
     { LLM_ARCH_XVERSE,           "xverse"           },
     { LLM_ARCH_COMMAND_R,        "command-r"        },
     { LLM_ARCH_COHERE2,          "cohere2"          },
@@ -73,6 +76,9 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_WAVTOKENIZER_DEC, "wavtokenizer-dec" },
     { LLM_ARCH_PLM,              "plm"              },
     { LLM_ARCH_BAILINGMOE,       "bailingmoe"       },
+    { LLM_ARCH_DOTS1,            "dots1"            },
+    { LLM_ARCH_ARCEE,            "arcee"            },
+    { LLM_ARCH_ERNIE4_5,         "ernie4_5"         },
     { LLM_ARCH_UNKNOWN,          "(unknown)"        },
 };
 
@@ -145,6 +151,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
     { LLM_KV_ATTENTION_SCALE,                        "%s.attention.scale"                        },
     { LLM_KV_ATTENTION_KEY_LENGTH_MLA,               "%s.attention.key_length_mla"               },
     { LLM_KV_ATTENTION_VALUE_LENGTH_MLA,             "%s.attention.value_length_mla"             },
+    { LLM_KV_ATTENTION_LAYER_INDICES,                "%s.attention.layer_indices"                },
 
     { LLM_KV_ROPE_DIMENSION_COUNT,      "%s.rope.dimension_count"                 },
     { LLM_KV_ROPE_DIMENSION_SECTIONS,   "%s.rope.dimension_sections"              },
@@ -165,6 +172,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
     { LLM_KV_SSM_INNER_SIZE,     "%s.ssm.inner_size"     },
     { LLM_KV_SSM_STATE_SIZE,     "%s.ssm.state_size"     },
     { LLM_KV_SSM_TIME_STEP_RANK, "%s.ssm.time_step_rank" },
+    { LLM_KV_SSM_GROUP_COUNT,    "%s.ssm.group_count"    },
     { LLM_KV_SSM_DT_B_C_RMS,     "%s.ssm.dt_b_c_rms"     },
 
     { LLM_KV_WKV_HEAD_SIZE, "%s.wkv.head_size" },
@@ -195,13 +203,13 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
     { LLM_KV_TOKENIZER_MASK_ID,              "tokenizer.ggml.mask_token_id"            },
     { LLM_KV_TOKENIZER_ADD_BOS,              "tokenizer.ggml.add_bos_token"            },
     { LLM_KV_TOKENIZER_ADD_EOS,              "tokenizer.ggml.add_eos_token"            },
+    { LLM_KV_TOKENIZER_ADD_SEP,              "tokenizer.ggml.add_sep_token"            },
     { LLM_KV_TOKENIZER_ADD_PREFIX,           "tokenizer.ggml.add_space_prefix"         },
     { LLM_KV_TOKENIZER_REMOVE_EXTRA_WS,      "tokenizer.ggml.remove_extra_whitespaces" },
     { LLM_KV_TOKENIZER_PRECOMPILED_CHARSMAP, "tokenizer.ggml.precompiled_charsmap"     },
     { LLM_KV_TOKENIZER_HF_JSON,              "tokenizer.huggingface.json"              },
     { LLM_KV_TOKENIZER_RWKV,                 "tokenizer.rwkv.world"                    },
     { LLM_KV_TOKENIZER_CHAT_TEMPLATE,        "tokenizer.chat_template"                 },
-    { LLM_KV_TOKENIZER_CHAT_TEMPLATE_N,      "tokenizer.chat_template.%s"              },
     { LLM_KV_TOKENIZER_FIM_PRE_ID,           "tokenizer.ggml.fim_pre_token_id"         },
     { LLM_KV_TOKENIZER_FIM_SUF_ID,           "tokenizer.ggml.fim_suf_token_id"         },
     { LLM_KV_TOKENIZER_FIM_MID_ID,           "tokenizer.ggml.fim_mid_token_id"         },
@@ -245,6 +253,24 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_FFN_UP_EXPS,     "blk.%d.ffn_up_exps" },
         },
     },
+    {
+        LLM_ARCH_ARCEE,
+        {
+            { 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_NORM,        "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
+        },
+    },
     {
         LLM_ARCH_LLAMA4,
         {
@@ -496,6 +522,21 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_FFN_UP_EXPS,     "blk.%d.ffn_up_exps" },
         },
     },
+    {
+        LLM_ARCH_NEO_BERT,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+            { LLM_TENSOR_ATTN_NORM,       "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_QKV,        "blk.%d.attn_qkv" },
+            { 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_TENSOR_ENC_OUTPUT_NORM, "enc.output_norm" },
+            { LLM_TENSOR_CLS,             "cls" },
+            { LLM_TENSOR_CLS_OUT,         "cls.output" },
+        },
+    },
     {
         LLM_ARCH_JINA_BERT_V2,
         {
@@ -910,6 +951,42 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_FFN_POST_NORM,   "blk.%d.post_ffw_norm" },
         },
     },
+    {
+        LLM_ARCH_GEMMA3N,
+        {
+            { 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_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_TENSOR_PER_LAYER_TOKEN_EMBD, "per_layer_token_embd" },
+            { LLM_TENSOR_PER_LAYER_MODEL_PROJ, "per_layer_model_proj" },
+            { LLM_TENSOR_PER_LAYER_PROJ_NORM,  "per_layer_proj_norm" },
+            { LLM_TENSOR_ALTUP_UNEMBD_PROJ,    "altup_unembd_proj" },
+            { LLM_TENSOR_ALTUP_PROJ,           "altup_proj" },
+            { LLM_TENSOR_PER_LAYER_INP_GATE,   "blk.%d.inp_gate" },
+            { LLM_TENSOR_PER_LAYER_PROJ,       "blk.%d.proj" },
+            { LLM_TENSOR_PER_LAYER_POST_NORM,  "blk.%d.post_norm" },
+            { LLM_TENSOR_ALTUP_CORRECT_COEF,   "blk.%d.altup_correct_coef" },
+            { LLM_TENSOR_ALTUP_CORRECT_SCALE,  "blk.%d.altup_correct_scale" },
+            { LLM_TENSOR_ALTUP_PREDICT_COEF,   "blk.%d.altup_predict_coef" },
+            { LLM_TENSOR_ALTUP_ROUTER,         "blk.%d.altup_router" },
+            { LLM_TENSOR_ALTUP_ROUTER_NORM,    "blk.%d.altup_router_norm" },
+            { LLM_TENSOR_LAUREL_L,             "blk.%d.laurel_l" },
+            { LLM_TENSOR_LAUREL_R,             "blk.%d.laurel_r" },
+            { LLM_TENSOR_LAUREL_POST_NORM,     "blk.%d.laurel_post_norm" },
+        },
+    },
     {
         LLM_ARCH_STARCODER2,
         {
@@ -944,6 +1021,22 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_SSM_OUT,         "blk.%d.ssm_out" },
         },
     },
+    {
+        LLM_ARCH_MAMBA2,
+        {
+            { 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_SSM_IN,          "blk.%d.ssm_in" },
+            { LLM_TENSOR_SSM_CONV1D,      "blk.%d.ssm_conv1d" },
+            { LLM_TENSOR_SSM_DT,          "blk.%d.ssm_dt" },
+            { LLM_TENSOR_SSM_A,           "blk.%d.ssm_a" },
+            { LLM_TENSOR_SSM_D,           "blk.%d.ssm_d" },
+            { LLM_TENSOR_SSM_NORM,        "blk.%d.ssm_norm" },
+            { LLM_TENSOR_SSM_OUT,         "blk.%d.ssm_out" },
+        },
+    },
     {
         LLM_ARCH_XVERSE,
         {
@@ -1571,6 +1664,51 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_FFN_UP_SHEXP,       "blk.%d.ffn_up_shexp" },
         },
     },
+    {
+        LLM_ARCH_DOTS1,
+        {
+            { 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_UP,             "blk.%d.ffn_up" },
+            { LLM_TENSOR_FFN_DOWN,           "blk.%d.ffn_down" },
+            { 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_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_TENSOR_FFN_EXP_PROBS_B,    "blk.%d.exp_probs_b" },
+        }
+    },
+    {
+        LLM_ARCH_ERNIE4_5,
+        {
+            { 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_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_GATE,           "blk.%d.ffn_gate" },
+            { LLM_TENSOR_FFN_DOWN,           "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,             "blk.%d.ffn_up" },
+        },
+    },
     {
         LLM_ARCH_UNKNOWN,
         {
@@ -1656,6 +1794,7 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
     {LLM_TENSOR_SSM_CONV1D,                 {LLM_TENSOR_LAYER_REPEATING, GGML_OP_SSM_CONV}},
     {LLM_TENSOR_SSM_A,                      {LLM_TENSOR_LAYER_REPEATING, GGML_OP_SSM_SCAN}},
     {LLM_TENSOR_SSM_D,                      {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
+    {LLM_TENSOR_SSM_NORM,                   {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
     {LLM_TENSOR_TIME_MIX_LERP_X,            {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
     {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}},
@@ -1699,6 +1838,23 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
     {LLM_TENSOR_FFN_GATE_EXPS,              {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT_ID}},
     {LLM_TENSOR_FFN_UP_EXPS,                {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT_ID}},
     {LLM_TENSOR_FFN_EXP_PROBS_B,            {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ADD}},
+    // altup / laurel (gemma 3n)
+    {LLM_TENSOR_PER_LAYER_TOKEN_EMBD,       {LLM_TENSOR_LAYER_OUTPUT,    GGML_OP_GET_ROWS}},
+    {LLM_TENSOR_PER_LAYER_MODEL_PROJ,       {LLM_TENSOR_LAYER_OUTPUT,    GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_PER_LAYER_PROJ_NORM,        {LLM_TENSOR_LAYER_OUTPUT,    GGML_OP_MUL}},
+    {LLM_TENSOR_ALTUP_PROJ,                 {LLM_TENSOR_LAYER_OUTPUT,    GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_ALTUP_UNEMBD_PROJ,          {LLM_TENSOR_LAYER_OUTPUT,    GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_PER_LAYER_INP_GATE,         {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_PER_LAYER_PROJ,             {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_PER_LAYER_POST_NORM,        {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
+    {LLM_TENSOR_ALTUP_CORRECT_COEF,         {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_ALTUP_CORRECT_SCALE,        {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
+    {LLM_TENSOR_ALTUP_PREDICT_COEF,         {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_ALTUP_ROUTER,               {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_ALTUP_ROUTER_NORM,          {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
+    {LLM_TENSOR_LAUREL_L,                   {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_LAUREL_R,                   {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_LAUREL_POST_NORM,           {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
     // this tensor is loaded for T5, but never used
     {LLM_TENSOR_DEC_CROSS_ATTN_REL_B,       {LLM_TENSOR_LAYER_REPEATING, GGML_OP_NONE}},
     {LLM_TENSOR_CONV1D,                     {LLM_TENSOR_LAYER_INPUT,     GGML_OP_IM2COL}},
@@ -1722,8 +1878,14 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
 LLM_KV::LLM_KV(llm_arch arch, const char * suffix) : arch(arch), suffix(suffix) {}
 
 std::string LLM_KV::operator()(llm_kv kv) const {
-    return suffix ? ::format(LLM_KV_NAMES.at(kv), LLM_ARCH_NAMES.at(arch), suffix)
-        : ::format(LLM_KV_NAMES.at(kv), LLM_ARCH_NAMES.at(arch));
+    std::string name = ::format(LLM_KV_NAMES.at(kv), LLM_ARCH_NAMES.at(arch));
+
+    if (suffix != nullptr) {
+        name += ".";
+        name += suffix;
+    }
+
+    return name;
 }
 
 std::string LLM_TN_IMPL::str() const {
@@ -1762,3 +1924,26 @@ llm_arch llm_arch_from_string(const std::string & name) {
 const llm_tensor_info & llm_tensor_info_for(llm_tensor tensor) {
     return LLM_TENSOR_INFOS.at(tensor);
 }
+
+bool llm_arch_is_recurrent(const llm_arch & arch) {
+    switch (arch) {
+        case LLM_ARCH_MAMBA:
+        case LLM_ARCH_MAMBA2:
+        case LLM_ARCH_RWKV6:
+        case LLM_ARCH_RWKV6QWEN2:
+        case LLM_ARCH_RWKV7:
+        case LLM_ARCH_ARWKV7:
+            return true;
+        default:
+            return false;
+    }
+}
+
+bool llm_arch_is_hybrid(const llm_arch & arch) {
+    // TODO: There are currently no hybrid models! Once there are, this will be
+    //  the place to identify them
+    switch (arch) {
+        default:
+            return false;
+    }
+}
diff --git a/src/llama-arch.h b/src/llama-arch.h
index 1dcd4fa35..89608ca63 100644
--- a/src/llama-arch.h
+++ b/src/llama-arch.h
@@ -24,6 +24,7 @@ enum llm_arch {
     LLM_ARCH_BERT,
     LLM_ARCH_NOMIC_BERT,
     LLM_ARCH_NOMIC_BERT_MOE,
+    LLM_ARCH_NEO_BERT,
     LLM_ARCH_JINA_BERT_V2,
     LLM_ARCH_JINA_BERT_V3,
     LLM_ARCH_BLOOM,
@@ -46,8 +47,10 @@ enum llm_arch {
     LLM_ARCH_GEMMA,
     LLM_ARCH_GEMMA2,
     LLM_ARCH_GEMMA3,
+    LLM_ARCH_GEMMA3N,
     LLM_ARCH_STARCODER2,
     LLM_ARCH_MAMBA,
+    LLM_ARCH_MAMBA2,
     LLM_ARCH_XVERSE,
     LLM_ARCH_COMMAND_R,
     LLM_ARCH_COHERE2,
@@ -77,6 +80,9 @@ enum llm_arch {
     LLM_ARCH_WAVTOKENIZER_DEC,
     LLM_ARCH_PLM,
     LLM_ARCH_BAILINGMOE,
+    LLM_ARCH_DOTS1,
+    LLM_ARCH_ARCEE,
+    LLM_ARCH_ERNIE4_5,
     LLM_ARCH_UNKNOWN,
 };
 
@@ -149,6 +155,7 @@ enum llm_kv {
     LLM_KV_ATTENTION_SCALE,
     LLM_KV_ATTENTION_KEY_LENGTH_MLA,
     LLM_KV_ATTENTION_VALUE_LENGTH_MLA,
+    LLM_KV_ATTENTION_LAYER_INDICES,
 
     LLM_KV_ROPE_DIMENSION_COUNT,
     LLM_KV_ROPE_DIMENSION_SECTIONS,
@@ -169,6 +176,7 @@ enum llm_kv {
     LLM_KV_SSM_CONV_KERNEL,
     LLM_KV_SSM_STATE_SIZE,
     LLM_KV_SSM_TIME_STEP_RANK,
+    LLM_KV_SSM_GROUP_COUNT,
     LLM_KV_SSM_DT_B_C_RMS,
 
     LLM_KV_WKV_HEAD_SIZE,
@@ -191,13 +199,13 @@ enum llm_kv {
     LLM_KV_TOKENIZER_MASK_ID,
     LLM_KV_TOKENIZER_ADD_BOS,
     LLM_KV_TOKENIZER_ADD_EOS,
+    LLM_KV_TOKENIZER_ADD_SEP,
     LLM_KV_TOKENIZER_ADD_PREFIX,
     LLM_KV_TOKENIZER_REMOVE_EXTRA_WS,
     LLM_KV_TOKENIZER_PRECOMPILED_CHARSMAP,
     LLM_KV_TOKENIZER_HF_JSON,
     LLM_KV_TOKENIZER_RWKV,
     LLM_KV_TOKENIZER_CHAT_TEMPLATE,
-    LLM_KV_TOKENIZER_CHAT_TEMPLATE_N,
     LLM_KV_TOKENIZER_FIM_PRE_ID,
     LLM_KV_TOKENIZER_FIM_SUF_ID,
     LLM_KV_TOKENIZER_FIM_MID_ID,
@@ -266,12 +274,29 @@ enum llm_tensor {
     LLM_TENSOR_LAYER_OUT_NORM,
     LLM_TENSOR_POST_ATTN_NORM,
     LLM_TENSOR_POST_MLP_NORM,
+    LLM_TENSOR_PER_LAYER_TOKEN_EMBD, // gemma3n
+    LLM_TENSOR_PER_LAYER_MODEL_PROJ, // gemma3n
+    LLM_TENSOR_PER_LAYER_INP_GATE,   // gemma3n
+    LLM_TENSOR_PER_LAYER_PROJ,       // gemma3n
+    LLM_TENSOR_PER_LAYER_PROJ_NORM,  // gemma3n
+    LLM_TENSOR_PER_LAYER_POST_NORM,  // gemma3n
+    LLM_TENSOR_ALTUP_PROJ,           // gemma3n
+    LLM_TENSOR_ALTUP_UNEMBD_PROJ,    // gemma3n
+    LLM_TENSOR_ALTUP_CORRECT_COEF,   // gemma3n
+    LLM_TENSOR_ALTUP_CORRECT_SCALE,  // gemma3n
+    LLM_TENSOR_ALTUP_PREDICT_COEF,   // gemma3n
+    LLM_TENSOR_ALTUP_ROUTER,         // gemma3n
+    LLM_TENSOR_ALTUP_ROUTER_NORM,    // gemma3n
+    LLM_TENSOR_LAUREL_L,             // gemma3n
+    LLM_TENSOR_LAUREL_R,             // gemma3n
+    LLM_TENSOR_LAUREL_POST_NORM,     // gemma3n
     LLM_TENSOR_SSM_IN,
     LLM_TENSOR_SSM_CONV1D,
     LLM_TENSOR_SSM_X,
     LLM_TENSOR_SSM_DT,
     LLM_TENSOR_SSM_A,
     LLM_TENSOR_SSM_D,
+    LLM_TENSOR_SSM_NORM,
     LLM_TENSOR_SSM_OUT,
     LLM_TENSOR_TIME_MIX_W0,
     LLM_TENSOR_TIME_MIX_W1,
@@ -438,3 +463,6 @@ const char * llm_arch_name(llm_arch arch);
 llm_arch llm_arch_from_string(const std::string & name);
 
 const llm_tensor_info & llm_tensor_info_for(llm_tensor tensor);
+
+bool llm_arch_is_recurrent(const llm_arch & arch);
+bool llm_arch_is_hybrid   (const llm_arch & arch);
diff --git a/src/llama-batch.cpp b/src/llama-batch.cpp
index 6a19a2431..3bc8554e5 100644
--- a/src/llama-batch.cpp
+++ b/src/llama-batch.cpp
@@ -1,301 +1,69 @@
 #include "llama-batch.h"
 
+#include "llama-impl.h"
+#include "llama-vocab.h"
+#include "llama-memory.h"
+
 #include <cassert>
 #include <cstring>
 #include <algorithm>
+#include <sstream>
 
-llama_ubatch llama_sbatch::reserve_ubatch(size_t n_ubatch, bool has_embd) {
-    // clear empty sequences
-    // the previous ubatch is assumed to be gone,
-    // so nothing should refer to values in these sequences anymore.
-    for (size_t i = seq.size(); i-- > 0;) {
-        if (seq[i].length == 0) {
-            seq.pop_back();
-        } else {
-            break;
-        }
+llama_batch_allocr::llama_batch_allocr(uint32_t n_pos_per_embd) : n_pos_per_embd(n_pos_per_embd) {
+    const char * LLAMA_BATCH_DEBUG = getenv("LLAMA_BATCH_DEBUG");
+    debug = LLAMA_BATCH_DEBUG ? atoi(LLAMA_BATCH_DEBUG) : 0;
+
+    seq_pos.resize(LLAMA_MAX_SEQ);
+    seq_cpl.resize(LLAMA_MAX_SEQ);
+    for (auto & cur : seq_cpl) {
+        cur.resize(LLAMA_MAX_SEQ);
     }
 
-    udatas.push_back({});
-
-    auto & udata = udatas.back();
-
-    udata.token.resize(!has_embd ? n_ubatch : 0);
-    udata.embd.resize(has_embd ? n_embd * n_ubatch : 0);
-    udata.pos.resize(n_ubatch);
-    udata.n_seq_id.resize(n_ubatch);
-    udata.seq_id.resize(n_ubatch);
-    udata.output.resize(n_ubatch);
-
-    llama_ubatch ubatch = {
-        /*equal_seqs   =*/ true,
-        /*n_tokens     =*/ 0,
-        /*n_seq_tokens =*/ 0,
-        /*n_seqs       =*/ 0,
-        /*token        =*/ !has_embd ? udata.token.data() : nullptr,
-        /*embd         =*/ has_embd  ? udata.embd.data()  : nullptr,
-        /*pos          =*/ udata.pos.data(),
-        /*n_seq_id     =*/ udata.n_seq_id.data(),
-        /*seq_id       =*/ udata.seq_id.data(),
-        /*output       =*/ udata.output.data(),
-    };
-
-    return ubatch;
+    seq_idx.resize(LLAMA_MAX_SEQ, -1);
 }
 
-void llama_sbatch::add_seq_to_ubatch(llama_ubatch & ubatch, llama_sbatch_seq & seq, size_t length) {
-    GGML_ASSERT(batch != nullptr);
-    GGML_ASSERT(length <= seq.length);
-    // Can only add sequences of equal lengths to a batch,
-    // otherwise it isn't clear to which sequence a token belongs
-    GGML_ASSERT(seq.n_seq_id == 0 || ubatch.n_seqs == 0 || length == (size_t) ubatch.n_tokens / ubatch.n_seqs);
-    GGML_ASSERT((seq.n_seq_id != 0) == ubatch.equal_seqs);
-    // NOTE: loops are separated for cache-friendliness
-    if (batch->token) {
-        if (ubatch.equal_seqs) {
-            for (size_t i = 0; i < length; ++i) {
-                ubatch.token[ubatch.n_tokens + i] = batch->token[ids[seq.offset + i]];
-            }
-        } else {
-            // simple split
-            ubatch.token = batch->token + seq.offset;
-        }
-    } else {
-        ubatch.token = nullptr;
-    }
-    if (batch->embd) {
-        if (ubatch.equal_seqs) {
-            for (size_t i = 0; i < length; ++i) {
-                memcpy(
-                        ubatch.embd + (n_embd * (ubatch.n_tokens + i)),
-                        batch->embd + (n_embd * ids[seq.offset + i]),
-                        n_embd * sizeof(float)
-                      );
-            }
-        } else {
-            // simple split
-            ubatch.embd = batch->embd + (n_embd * seq.offset);
-        }
-    } else {
-        ubatch.embd = nullptr;
-    }
-    if (ubatch.equal_seqs) {
-        for (size_t i = 0; i < length; ++i) {
-            ubatch.pos[ubatch.n_tokens + i] = batch->pos[ids[seq.offset + i]];
-        }
-    } else {
-        // simple split
-        ubatch.pos = batch->pos + seq.offset;
-    }
-    if (ubatch.equal_seqs) {
-        ubatch.n_seq_id[ubatch.n_seqs] = seq.n_seq_id;
-        if (seq.seq_id) {
-            ubatch.seq_id[ubatch.n_seqs] = seq.seq_id;
-        }
-    } else {
-        // simple split
-        if (batch->n_seq_id) {
-            ubatch.n_seq_id = batch->n_seq_id + seq.offset;
-        } else {
-            for (size_t i = 0; i < length; ++i) {
-                ubatch.n_seq_id[ubatch.n_seqs + i] = 1;
-            }
-        }
-        if (batch->seq_id) {
-            ubatch.seq_id = batch->seq_id + seq.offset;
-        }
-    }
-    if (logits_all) {
-        for (size_t i = 0; i < length; ++i) {
-            ubatch.output[ubatch.n_tokens + i] = 1;
-            out_ids.push_back(ids[seq.offset + i]);
-        }
-    } else if (batch->logits) {
-        if (ubatch.equal_seqs) {
-            for (size_t i = 0; i < length; ++i) {
-                size_t id = ids[seq.offset + i];
-                int8_t is_output = batch->logits[id];
-                ubatch.output[ubatch.n_tokens + i] = is_output;
-                if (is_output) { out_ids.push_back(id); }
-            }
-        } else {
-            // simple split
-            ubatch.output = batch->logits + seq.offset;
-            for (size_t i = 0; i < length; ++i) {
-                if (ubatch.output[i] != 0) { out_ids.push_back(seq.offset + i); }
-            }
-        }
-    } else {
-        // only get last output
-        for (size_t i = 0; i < length; ++i) {
-            size_t id = ids[seq.offset + i];
-            int8_t is_last = id == ids.size() - 1;
-            ubatch.output[ubatch.n_tokens + i] = is_last;
-            if (is_last) { out_ids.push_back(id); }
-        }
-    }
-    if (ubatch.n_tokens == 0 && ubatch.n_seqs == 0) {
-        ubatch.n_seq_tokens = ubatch.equal_seqs ? length : 1;
-    }
-    ubatch.n_tokens += length;
-    ubatch.n_seqs += ubatch.equal_seqs ? 1 : length; // virtual sequences for simple splits
-    seq.offset += length;
-    seq.length -= length;
-    n_tokens -= length;
-    GGML_ASSERT(ubatch.n_tokens == ubatch.n_seq_tokens * ubatch.n_seqs);
-}
+bool llama_batch_allocr::init(
+        const llama_batch & batch_inp,
+        const llama_vocab & vocab,
+        const llama_memory_i * memory,
+        uint32_t n_embd,
+        bool output_all) {
+    clear();
 
-llama_ubatch llama_sbatch::split_simple(size_t n_ubatch) {
-    n_ubatch = n_tokens < n_ubatch ? n_tokens : n_ubatch;
-    llama_ubatch ubatch = reserve_ubatch(n_ubatch, /* has_embd */ batch->embd != nullptr);
-    ubatch.equal_seqs = false;
-    if (!seq.empty()) {
-        llama_sbatch_seq & s = seq[0];
-        size_t length = s.length < n_ubatch ? s.length : n_ubatch;
-        GGML_ASSERT(seq.size() == 1 && s.n_seq_id == 0); // don't mix with other splits
-        add_seq_to_ubatch(ubatch, s, length);
-    }
-    return ubatch;
-}
+    batch = batch_inp;
 
-llama_ubatch llama_sbatch::split_equal(size_t n_ubatch) {
-    n_ubatch = n_tokens < n_ubatch ? n_tokens : n_ubatch;
-    llama_ubatch ubatch = reserve_ubatch(n_ubatch, /* has_embd */ batch->embd != nullptr);
-    if (!seq.empty()) {
-        size_t length = 0;
-        size_t n_tokens_in_ubatch = 0;
-        GGML_ASSERT(seq[0].n_seq_id > 0); // should not be mixed with simple splits
-                                          // smallest first, because it's easier to split this way;
-                                          // starting from the end to pop in constant time.
-        for (size_t i = seq.size(); i-- > 0;) {
-            llama_sbatch_seq & s = seq[i];
-            GGML_ASSERT(s.length > 0);
-            if (length == 0) {
-                length = s.length < n_ubatch ? s.length : n_ubatch;
-            }
-            add_seq_to_ubatch(ubatch, s, length);
-            n_tokens_in_ubatch += length;
-            // shared prompts can't be mixed with any of their sequences,
-            // so it's safer to compute them in their own ubatch
-            if (s.n_seq_id > 1) { break; }
-            // stop when there isn't enough space for another sequence
-            if (length + n_tokens_in_ubatch > n_ubatch) { break; }
-        }
-    }
-    return ubatch;
-}
+    this->vocab = &vocab;
 
-llama_ubatch llama_sbatch::split_seq(size_t n_ubatch) {
-    n_ubatch = n_tokens < n_ubatch ? n_tokens : n_ubatch;
-    llama_ubatch ubatch = reserve_ubatch(n_ubatch, /* has_embd */ batch->embd != nullptr);
-    if (!seq.empty()) {
-        llama_sbatch_seq & s = seq[seq.size() - 1];
-        size_t length = s.length < n_ubatch ? s.length : n_ubatch;
-        GGML_ASSERT(s.n_seq_id > 0); // should not be mixed with simple splits
-        add_seq_to_ubatch(ubatch, s, length);
-    }
-    return ubatch;
-}
-
-llama_sbatch::llama_sbatch(const llama_batch & batch, size_t n_embd, bool simple_split, bool logits_all) {
-    GGML_ASSERT(batch.n_tokens >= 0);
-    this->batch = &batch;
-    this->n_embd = n_embd;
-    this->logits_all = logits_all;
-
-    n_tokens = batch.n_tokens;
-    ids.resize(n_tokens);
-    out_ids.clear();
-    // TODO: reserve out_ids and seq
-
-    for (size_t i = 0; i < n_tokens; ++i) {
-        ids[i] = i;
-    }
-
-    if (simple_split) {
-        seq.resize(1);
-        llama_sbatch_seq & s = seq[0];
-        s.n_seq_id = 0;
-        s.seq_id = nullptr;
-        s.offset = 0;
-        s.length = n_tokens;
-        return;
-    }
-
-    std::sort(ids.begin(), ids.end(),
-            [&batch](size_t a, size_t b) {
-                int32_t n_seq_a = batch.n_seq_id ? batch.n_seq_id[a] : 1;
-                int32_t n_seq_b = batch.n_seq_id ? batch.n_seq_id[b] : 1;
-                // sort by seq_id, then by pos
-                if (n_seq_a == n_seq_b) {
-                    if (batch.seq_id) {
-                        for (int32_t i = 0; i < n_seq_a; ++i) {
-                            llama_seq_id seq_id_a = batch.seq_id[a][i];
-                            llama_seq_id seq_id_b = batch.seq_id[b][i];
-                            // smaller seq_ids go first
-                            if (seq_id_a != seq_id_b) {
-                                return seq_id_a < seq_id_b;
-                            }
-                        }
-                    }
-                    // when all else is equal, sort by pos
-                    if (batch.pos) {
-                        return batch.pos[a] < batch.pos[b];
-                    }
-                    // no pos, sort by id
-                    return a < b;
-                }
-                // shared prompts go first
-                return n_seq_a > n_seq_b;
-            }
-    );
-
-    // init seq
-    llama_sbatch_seq * last_seq = nullptr;
-
-    for (size_t i = 0; i < n_tokens; ++i) {
-        const size_t bi = ids[i];
-        const int32_t n_seqs = batch.n_seq_id[bi];
-        llama_seq_id * seq_ids = batch.seq_id[bi];
-        if (last_seq != nullptr) {
-            bool same = n_seqs == last_seq->n_seq_id;
-            for (int32_t j = 0; same && j < n_seqs; ++j) {
-                if (seq_ids[j] != last_seq->seq_id[j]) {
-                    same = false;
-                }
-            }
-            if (same) {
-                last_seq->length += 1;
-                continue;
-            }
-        }
-        llama_sbatch_seq new_seq = {n_seqs, seq_ids, i, 1};
-        seq.push_back(new_seq);
-        last_seq = &seq.back();
-    }
-
-    // keep shared prompts first at the end, then sort by length descending.
-    std::sort(seq.begin(), seq.end(),
-            [](llama_sbatch_seq & a, llama_sbatch_seq & b) {
-                if (a.n_seq_id == b.n_seq_id) {
-                    return a.length > b.length;
-                }
-                return a.n_seq_id < b.n_seq_id;
-            }
-            );
-}
-
-llama_batch_allocr::llama_batch_allocr(struct llama_batch in_batch, llama_pos p0) {
-    batch = in_batch;
     GGML_ASSERT(batch.n_tokens > 0);
-    if (!batch.pos) {
-        assert(p0 >= 0);
-        pos.resize(batch.n_tokens);
-        for (int32_t i = 0; i < batch.n_tokens; i++) {
-            pos[i] = p0 + i;
+
+    //
+    // validate input batch
+    //
+
+    if (batch.token) {
+        for (int32_t i = 0; i < batch.n_tokens; ++i) {
+            if (batch.token[i] < 0 || (uint32_t) batch.token[i] >= vocab.n_tokens()) {
+                LLAMA_LOG_ERROR("%s: invalid token[%d] = %d\n", __func__, i, batch.token[i]);
+                return false;
+            }
         }
-        batch.pos = pos.data();
     }
+
+    if (batch.seq_id) {
+        for (int32_t i = 0; i < batch.n_tokens; ++i) {
+            for (int32_t s = 0; s < batch.n_seq_id[i]; ++s) {
+                if (batch.seq_id && (batch.seq_id[i][s] < 0 || batch.seq_id[i][s] >= LLAMA_MAX_SEQ)) {
+                    LLAMA_LOG_ERROR("%s: invalid seq_id[%d][%d] = %d > %d\n", __func__, i, s, batch.seq_id[i][s], LLAMA_MAX_SEQ);
+                    return false;
+                }
+            }
+        }
+    }
+
+    //
+    // auto-generate missing fields
+    //
+
     if (!batch.n_seq_id) {
         n_seq_id.resize(batch.n_tokens);
         for (int32_t i = 0; i < batch.n_tokens; i++) {
@@ -303,6 +71,7 @@ llama_batch_allocr::llama_batch_allocr(struct llama_batch in_batch, llama_pos p0
         }
         batch.n_seq_id = n_seq_id.data();
     }
+
     if (!batch.seq_id) {
         seq_id.resize(batch.n_tokens + 1);
         seq_id[batch.n_tokens] = NULL;
@@ -311,10 +80,729 @@ llama_batch_allocr::llama_batch_allocr(struct llama_batch in_batch, llama_pos p0
         }
         batch.seq_id = seq_id.data();
     }
+
+    if (!batch.pos) {
+        pos.resize(batch.n_tokens);
+
+        // initialize the starting position for each sequence based on the positions in the memory
+        llama_pos p0[LLAMA_MAX_SEQ];
+        for (int32_t s = 0; s < LLAMA_MAX_SEQ; ++s) {
+            if (!memory) {
+                // if no memory -> start from 0
+                p0[s] = 0;
+            } else {
+                p0[s] = memory->seq_pos_max(s) + 1;
+            }
+        }
+
+        for (int32_t i = 0; i < batch.n_tokens; i++) {
+            const llama_seq_id seq_id = batch.seq_id[i][0];
+
+            pos[i] = p0[seq_id];
+
+            // update the starting position for all sequences that are assigned to the this token
+            for (int32_t s = 0; s < batch.n_seq_id[i]; ++s) {
+                const llama_seq_id seq_id = batch.seq_id[i][s];
+
+                p0[seq_id] = pos[i] + 1;
+            }
+        }
+
+        batch.pos = pos.data();
+    }
+
     if (!batch.logits) {
-        logits.resize(batch.n_tokens);
-        logits[logits.size() - 1] = true;
-        batch.logits = logits.data();
+        if (output_all) {
+            // return the output for all tokens
+            output.resize(batch.n_tokens, true);
+        } else {
+            // return the output only for the last token
+            output.resize(batch.n_tokens, false);
+            output[output.size() - 1] = true;
+        }
+
+        batch.logits = output.data();
+    } else if (output_all) {
+        bool warn = false;
+
+        for (int32_t i = 0; i < batch.n_tokens; ++i) {
+            if (batch.logits[i] == 0) {
+                warn = true;
+            }
+        }
+
+        if (warn) {
+            LLAMA_LOG_WARN("%s: embeddings required but some input tokens were not marked as outputs -> overriding\n", __func__);
+
+            output.resize(batch.n_tokens, true);
+            batch.logits = output.data();
+        }
+    }
+
+    //
+    // compute stats
+    //
+
+    this->n_embd = n_embd;
+
+    // count the outputs in this batch
+    for (int32_t i = 0; i < batch.n_tokens; ++i) {
+        n_outputs += batch.logits[i] != 0;
+    }
+
+    // determine coupled sequences
+    // these are pairs of sequences that have at least one token in the input batch that is assigned to both of them
+    for (int32_t i = 0; i < batch.n_tokens; ++i) {
+        const llama_seq_id s0 = batch.seq_id[i][0];
+
+        for (int32_t s = 0; s < batch.n_seq_id[i]; ++s) {
+            const llama_seq_id s1 = batch.seq_id[i][s];
+
+            seq_pos[s1].insert(batch.pos[i]);
+
+            if (s > 0) {
+                // mark that sequence s1 is coupled to s0
+                seq_cpl[s1][s0] = true;
+
+                // note: tracking the other way around is not necessary for now
+                //seq_cpl[s0][s1] = true;
+
+                has_cpl = true;
+            }
+        }
+    }
+
+    // precompute the sequence sets for each token and determine the unique sequence ids that participate in the batch
+    {
+        seq_set_t seq_set_unq;
+
+        for (int32_t i = 0; i < batch.n_tokens; ++i) {
+            seq_set_t cur;
+            for (int32_t s = 0; s < batch.n_seq_id[i]; ++s) {
+                const llama_seq_id seq_id = batch.seq_id[i][s];
+
+                cur        .set(seq_id);
+                seq_set_unq.set(seq_id);
+            }
+
+            seq_set.push_back(cur);
+            seq_set_map[cur].push_back(i);
+        }
+
+        for (int32_t s = 0; s < LLAMA_MAX_SEQ; ++s) {
+            if (seq_set_unq.test(s)) {
+                seq_idx[s] = seq_id_unq.size();
+                seq_id_unq.push_back(s);
+            }
+        }
+    }
+
+    if (debug > 0) {
+        LLAMA_LOG_DEBUG("%s: input batch info:\n", __func__);
+
+        llama_ubatch ubatch {
+            /*.equal_seqs   =*/ false,
+            /*.n_tokens     =*/ (uint32_t) batch.n_tokens,
+            /*.n_seq_tokens =*/ (uint32_t) 1,
+            /*.n_seqs       =*/ (uint32_t) batch.n_tokens,
+            /*.n_seqs_unq   =*/ (uint32_t) this->seq_id_unq.size(),
+            /*.token        =*/ batch.token,
+            /*.embd         =*/ batch.embd,
+            /*.pos          =*/ batch.pos,
+            /*.n_seq_id     =*/ batch.n_seq_id,
+            /*.seq_id       =*/ batch.seq_id,
+            /*.seq_id_unq   =*/ this->seq_id_unq.data(),
+            /*.seq_idx      =*/ this->seq_idx.data(),
+            /*.output       =*/ batch.logits,
+        };
+
+        ubatch_print(ubatch, debug);
+
+        LLAMA_LOG_DEBUG("%s:   seq       = [\n", __func__);
+        for (int s0 = 0; s0 < (int) seq_pos.size(); ++s0) {
+            if (seq_pos[s0].empty()) {
+                continue;
+            }
+
+            std::stringstream ss;
+            for (int s1 = 0; s1 < (int) seq_cpl[s0].size(); ++s1) {
+                if (seq_cpl[s0][s1]) {
+                    ss << s1 << " ";
+                }
+            }
+
+            LLAMA_LOG_DEBUG("%s:  %4d: pos = [%4d, %4d], cpl = %s\n",
+                    __func__, s0, seq_pos_min(s0), seq_pos_max(s0), ss.str().empty() ? "-" : ss.str().c_str());
+        }
+        LLAMA_LOG_DEBUG("%s:   ]\n", __func__);
+    }
+
+    //
+    // consistency checks
+    //
+
+    for (int32_t s = 0; s < LLAMA_MAX_SEQ; ++s) {
+        if (seq_pos[s].empty()) {
+            continue;
+        }
+
+        const llama_pos p0 = memory ? memory->seq_pos_max(s) : -1;
+
+        if (p0 >= 0) {
+            bool ok = true;
+
+            if (batch.token) {
+                if (seq_pos_min(s) != p0 + 1) {
+                    ok = false;
+                }
+            } else {
+                assert(batch.embd);
+
+                // for embeddings (typically used as vision input), we allow them to have repeating positions
+                // ref: https://github.com/ggml-org/llama.cpp/issues/13694#issuecomment-2983871762
+                if (seq_pos_min(s) != p0 && seq_pos_min(s) != p0 + 1) {
+                    ok = false;
+                }
+            }
+
+            if (!ok) {
+                LLAMA_LOG_ERROR(
+                        "%s: the tokens of sequence %d in the input batch have inconsistent sequence positions:\n"
+                        " - the last position stored in the memory module of the context (i.e. the KV cache) for sequence %d is X = %d\n"
+                        " - the tokens for sequence %d in the input batch have a starting position of Y = %d\n"
+                        " it is required that the sequence positions remain consecutive: Y = X + 1\n",
+                        __func__, s, s, p0, s, seq_pos_min(s));
+
+                return false;
+            }
+        }
+
+        if (seq_pos_max(s) - seq_pos_min(s) + 1 > (int) seq_pos[s].size()) {
+            LLAMA_LOG_ERROR("%s: sequence %d positions are not continuous\n", __func__, s);
+            return false;
+        }
+    }
+
+    if (memory) {
+        for (int32_t s0 = 0; s0 < LLAMA_MAX_SEQ; ++s0) {
+            for (int32_t s1 = 0; s1 < LLAMA_MAX_SEQ; ++s1) {
+                if (seq_cpl[s0][s1]) {
+                    if (memory->seq_pos_min(s0) != memory->seq_pos_min(s1) ||
+                        memory->seq_pos_max(s0) != memory->seq_pos_max(s1)) {
+                        LLAMA_LOG_ERROR("%s: sequence %d is coupled to %d in the input batch, but have divereged\n", __func__, s0, s1);
+                        return false;
+                    }
+                }
+            }
+        }
+    }
+
+    // disallow partial sequence sub-sets:
+    //
+    // invalid:          x
+    //            i: 0 1 2 ...
+    // ---------------------------------------
+    // seq_id[i][0]: 0 0 1
+    // seq_id[i][1]: 1 1 2
+    // seq_id[i][2]: 2
+    //
+    // disallow decreasing sequence positions:
+    //
+    // invalid:                  x
+    //            i: 0 1 2 3 4 5 6 ...
+    // ---------------------------------------
+    //       pos[i]: 4 5 0 1 6 2 3
+    // seq_id[i][0]: 0 0 1 1 0 1 0
+    //
+    {
+        seq_set_t cur_seq_set[LLAMA_MAX_SEQ];
+        for (int32_t s = 0; s < LLAMA_MAX_SEQ; ++s) {
+            cur_seq_set[s].set();
+        }
+
+        llama_pos cur_seq_pos[LLAMA_MAX_SEQ];
+        for (int32_t s = 0; s < LLAMA_MAX_SEQ; ++s) {
+            cur_seq_pos[s] = -1;
+        }
+
+        for (int32_t i = 0; i < batch.n_tokens; ++i) {
+            const llama_pos pos = batch.pos[i];
+
+            for (int32_t s = 0; s < batch.n_seq_id[i]; ++s) {
+                const llama_seq_id seq_id = batch.seq_id[i][s];
+
+                cur_seq_set[seq_id] &= seq_set[i];
+
+                if (cur_seq_set[seq_id].none()) {
+                    LLAMA_LOG_ERROR("%s: sequence %d belongs to incompatible sequence sets (not allowed)\n", __func__, seq_id);
+                    return false;
+                }
+
+                if (pos < cur_seq_pos[seq_id]) {
+                    LLAMA_LOG_ERROR("%s: sequence %d positions are decreasing (not allowed)\n", __func__, seq_id);
+                    return false;
+                }
+            }
+        }
+    }
+
+    split_reset();
+
+    return true;
+}
+
+llama_ubatch llama_batch_allocr::ubatch_reserve(uint32_t n_seq_tokens, uint32_t n_seqs) {
+    const uint32_t n_tokens = n_seq_tokens*n_seqs;
+
+    clear();
+    split_reset();
+
+    ubatches.emplace_back();
+
+    auto & ubatch = ubatches.back();
+
+    ubatch.token     .resize(n_tokens);
+    ubatch.embd      .clear();
+    ubatch.pos       .resize(n_tokens);
+    ubatch.n_seq_id  .resize(n_tokens);
+    ubatch.seq_id    .resize(n_tokens);
+    ubatch.seq_id_unq.resize(0);
+    ubatch.seq_idx   .resize(LLAMA_MAX_SEQ, -1);
+    ubatch.output    .resize(n_tokens);
+
+    for (uint32_t s = 0; s < n_seqs; ++s) {
+        ubatch.seq_idx[s] = s;
+        ubatch.seq_id_unq.push_back(s);
+    }
+
+    llama_ubatch res {
+        /*.equal_seqs   =*/ true,
+        /*.n_tokens     =*/ n_tokens,
+        /*.n_seq_tokens =*/ n_seq_tokens,
+        /*.n_seqs       =*/ n_seqs,
+        /*.n_seqs_unq   =*/ n_seqs,
+
+        /*.token        =*/ ubatch.token.data(),
+        /*.embd         =*/ nullptr,
+        /*.pos          =*/ ubatch.pos.data(),
+        /*.n_seq_id     =*/ ubatch.n_seq_id.data(),
+        /*.seq_id       =*/ ubatch.seq_id.data(),
+        /*.seq_id_unq   =*/ ubatch.seq_id_unq.data(),
+        /*.seq_idx      =*/ ubatch.seq_idx.data(),
+        /*.output       =*/ ubatch.output.data(),
+    };
+
+    return res;
+}
+
+const llama_batch & llama_batch_allocr::get_batch() const {
+    return batch;
+}
+
+uint32_t llama_batch_allocr::get_n_tokens() const {
+    return batch.n_tokens;
+}
+
+uint32_t llama_batch_allocr::get_n_outputs() const {
+    return n_outputs;
+}
+
+uint32_t llama_batch_allocr::get_n_used() const {
+    return n_used;
+}
+
+std::vector<int32_t> & llama_batch_allocr::get_out_ids() {
+    return out_ids;
+}
+
+llama_pos llama_batch_allocr::seq_pos_min(llama_seq_id seq_id) const {
+    return seq_pos[seq_id].empty() ? -1 : *seq_pos[seq_id].begin();
+}
+
+llama_pos llama_batch_allocr::seq_pos_max(llama_seq_id seq_id) const {
+    return seq_pos[seq_id].empty() ? -1 : *seq_pos[seq_id].rbegin();
+}
+
+void llama_batch_allocr::split_reset() {
+    out_ids.clear();
+
+    n_used = 0;
+
+    used.clear();
+    used.resize(get_n_tokens(), false);
+
+    ubatches.clear();
+}
+
+llama_ubatch llama_batch_allocr::split_simple(uint32_t n_ubatch) {
+    // find the first unused token
+    uint32_t cur_idx = 0;
+    while (cur_idx < used.size() && used[cur_idx]) {
+        ++cur_idx;
+    }
+
+    // we are done
+    if (cur_idx >= used.size()) {
+        return {};
+    }
+
+    std::vector<int32_t> idxs;
+
+    while (true) {
+        idxs.push_back(cur_idx);
+
+        used[cur_idx] = true;
+        ++n_used;
+
+        ++cur_idx;
+
+        if (cur_idx >= used.size()) {
+            break;
+        }
+
+        if (idxs.size() >= n_ubatch) {
+            break;
+        }
+    }
+
+    return ubatch_add(idxs, idxs.size(), false);
+}
+
+llama_ubatch llama_batch_allocr::split_equal(uint32_t n_ubatch, bool sequential) {
+    if (sequential && has_cpl) {
+        LLAMA_LOG_ERROR("%s: sequential split is not supported when there are coupled sequences in the input batch\n", __func__);
+
+        return {};
+    }
+
+    std::vector<seq_set_t> cur_seq_set;
+
+    llama_seq_id last_seq_id = -1;
+
+    // determine the non-overlapping sequence sets participating in this ubatch
+    for (int32_t i = 0; i < batch.n_tokens; ++i) {
+        if (used[i]) {
+            continue;
+        }
+
+        bool add = true;
+
+        for (uint32_t s = 0; s < cur_seq_set.size(); ++s) {
+            // no overlap with existing sequence sets:
+            if (!(cur_seq_set[s] & seq_set[i]).none()) {
+                add = false;
+                break;
+            }
+        }
+
+        // accept only increasing sequence ids
+        if (sequential) {
+            add = add && (cur_seq_set.empty() || batch.seq_id[i][0] == last_seq_id + 1);
+        }
+
+        if (add) {
+            cur_seq_set.push_back(seq_set[i]);
+
+            last_seq_id = batch.seq_id[i][0];
+
+            if (cur_seq_set.size() > n_ubatch) {
+                break;
+            }
+        }
+    }
+
+    const uint32_t n_seqs = cur_seq_set.size();
+
+    // we are done
+    if (n_seqs == 0) {
+        return {};
+    }
+
+    // the current batch index of each sequence set
+    std::vector<int32_t> cur_idx(n_seqs, 0);
+
+    for (uint32_t s = 0; s < n_seqs; ++s) {
+        while (used[seq_set_map[cur_seq_set[s]][cur_idx[s]]]) {
+            ++cur_idx[s];
+        }
+    }
+
+    // the list of batch indices for each sequence set
+    // at the end we will concat these to get the final ubatch
+    std::vector<idx_vec_t> idxs_per_seq(n_seqs);
+
+    while (true) {
+        // we can only add new n_seq_tokens tokens if all the sequence sets have at least one more unused token and
+        //   if we haven't reached n_ubatch
+        bool can_expand = true;
+
+        for (uint32_t s = 0; s < n_seqs; ++s) {
+            if (cur_idx[s] >= (int32_t) seq_set_map[cur_seq_set[s]].size()) {
+                can_expand = false;
+                break;
+            }
+        }
+
+        if (!can_expand) {
+            break;
+        }
+
+        for (uint32_t s = 0; s < n_seqs; ++s) {
+            const int32_t idx = seq_set_map[cur_seq_set[s]][cur_idx[s]];
+
+            idxs_per_seq[s].push_back(idx);
+
+            used[idx] = true;
+            ++n_used;
+
+            ++cur_idx[s];
+        }
+
+        if  ((idxs_per_seq[0].size() + 1)*n_seqs > n_ubatch) {
+            break;
+        }
+    }
+
+    // concat the per-sequence-set lists
+    std::vector<int32_t> idxs;
+
+    for (uint32_t s = 0; s < n_seqs; ++s) {
+        idxs.insert(idxs.end(), idxs_per_seq[s].begin(), idxs_per_seq[s].end());
+    }
+
+    return ubatch_add(idxs, n_seqs, true);
+}
+
+llama_ubatch llama_batch_allocr::split_seq(uint32_t n_ubatch) {
+    // find the first unused token
+    uint32_t cur_idx = 0;
+    while (cur_idx < used.size() && used[cur_idx]) {
+        ++cur_idx;
+    }
+
+    // we are done
+    if (cur_idx >= used.size()) {
+        return {};
+    }
+
+    // this is the starting sequence set
+    // we allow adding tokens only if their sequence set is a subset of the current sequence set
+    auto cur_seq_set = seq_set[cur_idx];
+
+    std::vector<int32_t> idxs;
+
+    while (true) {
+        idxs.push_back(cur_idx);
+
+        used[cur_idx] = true;
+        ++n_used;
+
+        if (idxs.size() >= n_ubatch) {
+            break;
+        }
+
+        do {
+            ++cur_idx;
+        } while (cur_idx < get_n_tokens() && (used[cur_idx] || ((cur_seq_set & seq_set[cur_idx]) != seq_set[cur_idx])));
+
+        if (cur_idx == get_n_tokens()) {
+            break;
+        }
+
+        cur_seq_set = seq_set[cur_idx];
+    }
+
+    return ubatch_add(idxs, 1, true);
+}
+
+void llama_batch_allocr::clear() {
+    n_outputs = 0;
+
+    batch = {};
+
+    pos       .clear();
+    n_seq_id  .clear();
+    seq_id    .clear();
+    seq_id_unq.clear();
+    output    .clear();
+
+    for (auto & cur : seq_pos) {
+        cur.clear();
+    }
+
+    for (auto & cur : seq_cpl) {
+        std::fill(cur.begin(), cur.end(), false);
+    }
+
+    seq_set.clear();
+
+    seq_set_map.clear();
+
+    std::fill(seq_idx.begin(), seq_idx.end(), -1);
+}
+
+llama_ubatch llama_batch_allocr::ubatch_add(const std::vector<int32_t> & idxs, uint32_t n_seqs, bool equal_seqs) {
+    const uint32_t n_tokens = idxs.size();
+
+    assert(n_tokens%n_seqs == 0);
+
+    ubatches.emplace_back();
+
+    auto & ubatch = ubatches.back();
+
+    const int32_t n_pos_cur = batch.embd ? n_pos_per_embd : 1;
+
+    const int64_t n_embd_all = batch.embd ? (int64_t) n_tokens*n_embd : 0;
+    const int64_t n_pos_all  =              (int64_t) n_tokens*n_pos_cur;
+
+    ubatch.token     .resize(n_tokens);
+    ubatch.embd      .resize(n_embd_all);
+    ubatch.pos       .resize(n_pos_all);
+    ubatch.n_seq_id  .resize(n_tokens);
+    ubatch.seq_id    .resize(n_tokens);
+    ubatch.seq_id_unq.resize(0);
+    ubatch.seq_idx   .resize(LLAMA_MAX_SEQ, -1);
+    ubatch.output    .resize(n_tokens);
+
+    seq_set_t seq_set_unq;
+
+    for (size_t i = 0; i < idxs.size(); ++i) {
+        if (batch.token) {
+            ubatch.token[i] = batch.token[idxs[i]];
+        }
+
+        if (batch.embd) {
+            memcpy(ubatch.embd.data() + i*n_embd, batch.embd + (int64_t) idxs[i]*n_embd, n_embd*sizeof(float));
+        }
+
+        for (int j = 0; j < n_pos_cur; ++j) {
+            ubatch.pos[j*n_tokens + i] = batch.pos[j*batch.n_tokens + idxs[i]];
+        }
+
+        ubatch.n_seq_id[i] = batch.n_seq_id[idxs[i]];
+        ubatch.seq_id[i]   = batch.seq_id[idxs[i]];
+        ubatch.output[i]   = batch.logits[idxs[i]];
+
+        for (int s = 0; s < ubatch.n_seq_id[i]; ++s) {
+            seq_set_unq.set(ubatch.seq_id[i][s]);
+        }
+
+        if (ubatch.output[i]) {
+            out_ids.push_back(idxs[i]);
+        }
+    }
+
+    for (int32_t s = 0; s < LLAMA_MAX_SEQ; ++s) {
+        if (seq_set_unq.test(s)) {
+            ubatch.seq_idx[s] = ubatch.seq_id_unq.size();
+            ubatch.seq_id_unq.push_back(s);
+        }
+    }
+
+    llama_ubatch res {
+        /*.equal_seqs   =*/ equal_seqs,
+        /*.n_tokens     =*/ n_tokens,
+        /*.n_seq_tokens =*/ n_tokens/n_seqs,
+        /*.n_seqs       =*/ n_seqs,
+        /*.n_seqs_unq   =*/ (uint32_t) ubatch.seq_id_unq.size(),
+
+        /*.token        =*/ batch.token ? ubatch.token.data() : nullptr,
+        /*.embd         =*/ batch.embd ? ubatch.embd.data() : nullptr,
+        /*.pos          =*/ ubatch.pos.data(),
+        /*.n_seq_id     =*/ ubatch.n_seq_id.data(),
+        /*.seq_id       =*/ ubatch.seq_id.data(),
+        /*.seq_id_unq   =*/ ubatch.seq_id_unq.data(),
+        /*.seq_idx      =*/ ubatch.seq_idx.data(),
+        /*.output       =*/ ubatch.output.data(),
+    };
+
+    if (debug > 0) {
+        LLAMA_LOG_DEBUG("%s: added ubatch %d to split:\n", __func__, (int) ubatches.size() - 1);
+
+        ubatch_print(res, debug);
+    }
+
+    return res;
+}
+
+void llama_batch_allocr::ubatch_print(const llama_ubatch & ubatch, int debug) {
+    if (debug > 0) {
+        LLAMA_LOG_DEBUG("%s:   equal_seqs   = %d\n", __func__, ubatch.equal_seqs);
+        LLAMA_LOG_DEBUG("%s:   n_tokens     = %d\n", __func__, ubatch.n_tokens);
+        LLAMA_LOG_DEBUG("%s:   n_seq_tokens = %d\n", __func__, ubatch.n_seq_tokens);
+        LLAMA_LOG_DEBUG("%s:   n_seqs       = %d\n", __func__, ubatch.n_seqs);
+        LLAMA_LOG_DEBUG("%s:   n_seqs_unq   = %d\n", __func__, ubatch.n_seqs_unq);
+
+        std::stringstream ss_seq_id_unq;
+        std::stringstream ss_seq_idx;
+
+        ss_seq_id_unq << "[ ";
+        ss_seq_idx << "[";
+
+        for (uint32_t s = 0; s < ubatch.n_seqs_unq; ++s) {
+            ss_seq_id_unq << ubatch.seq_id_unq[s] << " ";
+        }
+
+        for (uint32_t s = 0; s < LLAMA_MAX_SEQ; ++s) {
+            if (ubatch.seq_idx[s] >= 0) {
+                ss_seq_idx << ubatch.seq_idx[s]%10;
+            } else {
+                ss_seq_idx << ".";
+            }
+        }
+
+        ss_seq_id_unq << "]";
+        ss_seq_idx    << "]";
+
+        LLAMA_LOG_DEBUG("%s:   token      = %p\n", __func__, (void *) ubatch.token);
+        LLAMA_LOG_DEBUG("%s:   embd       = %p\n", __func__, (void *) ubatch.embd);
+        LLAMA_LOG_DEBUG("%s:   pos        = %p\n", __func__, (void *) ubatch.pos);
+        LLAMA_LOG_DEBUG("%s:   n_seq_id   = %p\n", __func__, (void *) ubatch.n_seq_id);
+        LLAMA_LOG_DEBUG("%s:   seq_id     = %p\n", __func__, (void *) ubatch.seq_id);
+        LLAMA_LOG_DEBUG("%s:   seq_id_unq = %s\n", __func__, ss_seq_id_unq.str().c_str());
+        LLAMA_LOG_DEBUG("%s:   seq_idx    = %s\n", __func__, ss_seq_idx.str().c_str());
+        LLAMA_LOG_DEBUG("%s:   output     = %p\n", __func__, (void *) ubatch.output);
+        LLAMA_LOG_DEBUG("%s:   n_outputs  = %d\n", __func__, n_outputs);
+
+        if (debug > 1) {
+            int seq_id_max = 0;
+            for (uint32_t i = 0; i < ubatch.n_tokens; ++i) {
+                for (int s = 0; s < ubatch.n_seq_id[i]; ++s) {
+                    for (int s = 0; s < ubatch.n_seq_id[i]; ++s) {
+                        seq_id_max = std::max(seq_id_max, ubatch.seq_id[i][s]);
+                    }
+                }
+            }
+            ++seq_id_max;
+
+            LLAMA_LOG_DEBUG("%s:   token     = [\n", __func__);
+            for (uint32_t i = 0; i < ubatch.n_tokens; ++i) {
+                std::vector<int8_t> seq_id(seq_id_max);
+
+                for (int s = 0; s < ubatch.n_seq_id[i]; ++s) {
+                    seq_id[ubatch.seq_id[i][s]] = 1;
+                }
+
+                std::stringstream ss;
+                for (int s = 0; s < seq_id_max; ++s) {
+                    if (seq_id[s]) {
+                        ss << s%10;
+                    } else {
+                        ss << ".";
+                    }
+                }
+
+                if (ubatch.token) {
+                    LLAMA_LOG_DEBUG("%s:  %4d: id = %6d (%16s), pos = %4d, n_seq_id = %2d, seq_id = [%s], output = %d\n",
+                            __func__, i, ubatch.token[i], vocab->token_to_piece(ubatch.token[i]).c_str(),
+                            ubatch.pos[i], ubatch.n_seq_id[i], ss.str().c_str(), ubatch.output[i]);
+                } else {
+                    LLAMA_LOG_DEBUG("%s:  %4d: [embd], pos = %4d, n_seq_id = %2d, seq_id = [%s], output = %d\n",
+                            __func__, i, ubatch.pos[i], ubatch.n_seq_id[i], ss.str().c_str(), ubatch.output[i]);
+                }
+            }
+            LLAMA_LOG_DEBUG("%s:   ]\n", __func__);
+        }
     }
 }
 
@@ -326,25 +814,25 @@ struct llama_batch llama_batch_get_one(
              llama_token * tokens,
                  int32_t   n_tokens) {
     return {
-        /*n_tokens       =*/ n_tokens,
-        /*tokens         =*/ tokens,
-        /*embd           =*/ nullptr,
-        /*pos            =*/ nullptr,
-        /*n_seq_id       =*/ nullptr,
-        /*seq_id         =*/ nullptr,
-        /*logits         =*/ nullptr,
+        /*n_tokens =*/ n_tokens,
+        /*tokens   =*/ tokens,
+        /*embd     =*/ nullptr,
+        /*pos      =*/ nullptr,
+        /*n_seq_id =*/ nullptr,
+        /*seq_id   =*/ nullptr,
+        /*logits   =*/ nullptr,
     };
 }
 
 struct llama_batch llama_batch_init(int32_t n_tokens_alloc, int32_t embd, int32_t n_seq_max) {
     llama_batch batch = {
-        /*n_tokens       =*/ 0,
-        /*tokens         =*/ nullptr,
-        /*embd           =*/ nullptr,
-        /*pos            =*/ nullptr,
-        /*n_seq_id       =*/ nullptr,
-        /*seq_id         =*/ nullptr,
-        /*logits         =*/ nullptr,
+        /*n_tokens =*/ 0,
+        /*tokens   =*/ nullptr,
+        /*embd     =*/ nullptr,
+        /*pos      =*/ nullptr,
+        /*n_seq_id =*/ nullptr,
+        /*seq_id   =*/ nullptr,
+        /*logits   =*/ nullptr,
     };
 
     if (embd) {
diff --git a/src/llama-batch.h b/src/llama-batch.h
index b8260b94f..3420803ff 100644
--- a/src/llama-batch.h
+++ b/src/llama-batch.h
@@ -2,93 +2,153 @@
 
 #include "llama.h"
 
+#include "llama-cparams.h"
+
 #include <array>
 #include <vector>
+#include <set>
+#include <bitset>
+#include <unordered_map>
 
-// very similar to llama_batch,
-// but has more metadata about sequences
+// keep this struct lightweight
+// it points to data in `llama_batch_allocr`
 struct llama_ubatch {
     bool equal_seqs;
     // TODO: whole_seqs for embeddings?
 
     uint32_t n_tokens;     // total tokens (n_seq_tokens * n_seqs)
-    uint32_t n_seq_tokens; // tokens per sequence
-    uint32_t n_seqs;
+    uint32_t n_seq_tokens; // tokens per sequence set
+    uint32_t n_seqs;       // sequence sets in the ubatch
+    uint32_t n_seqs_unq;   // unique sequence ids in the ubatch
 
-    llama_token  *  token;    // [n_tokens]
-    float        *  embd;     // [n_embd, n_tokens]
-    llama_pos    *  pos;      // [n_tokens]
-    int32_t      *  n_seq_id; // [n_seqs] // TODO: remove, should belong to only 1 sequence
-    llama_seq_id ** seq_id;   // [n_seqs] // TODO: become llama_seq_id * seq_id;
-    int8_t       *  output;   // [n_tokens]
+    // seq_id_unq: unique sequence ids in the ubatch
+    // seq_idx:    indices of the unique sequence ids in the ubatch in [0, n_seqs_unq)
+    //             used for extracting sequence pooled embeddings
+
+    //                          // size               | idx | val
+    llama_token  *  token;      // [n_tokens]         | i   | id, token
+    float        *  embd;       // [n_embd, n_tokens] | i   | embd
+    llama_pos    *  pos;        // [n_tokens]         | i   | pos
+    int32_t      *  n_seq_id;   // [n_tokens]         | i   | -
+    llama_seq_id ** seq_id;     // [n_tokens]         | s   | s0, s1, seq_id
+    llama_seq_id *  seq_id_unq; // [n_seqs_unq]       | s   | seq_id
+    int32_t      *  seq_idx;    // [LLAMA_MAX_SEQ]    | -   | seq_idx
+    int8_t       *  output;     // [n_tokens]         | i   | -
 };
 
-struct llama_sbatch_seq {
-    int32_t n_seq_id;
+// a helper for sanitizing, fulfilling and splitting a batch
+class llama_batch_allocr {
+public:
+    llama_batch_allocr(uint32_t n_pos_per_embd);
 
-    llama_seq_id * seq_id;
+    // sanitize and auto-gen missing data in the input batch
+    // memory is optional. if provided will be used to check for sequence continuity and to determine the positions
+    bool init(
+            const llama_batch & batch_inp,
+            const llama_vocab & vocab,
+            const llama_memory_i * memory,
+            uint32_t n_embd,
+            bool output_all);
 
-    size_t offset;
-    size_t length;
-};
+    const llama_batch & get_batch() const;
 
-// sequence-length-aware batch splitting
-struct llama_sbatch {
-    // tokens left in this batch
-    size_t n_tokens;
+    uint32_t get_n_tokens()  const;
+    uint32_t get_n_outputs() const;
+    uint32_t get_n_used()    const;
 
-    size_t n_embd;
+    // the array of output indices in the order they were encountered during the ubatch splitting
+    std::vector<int32_t> & get_out_ids();
 
-    bool logits_all; // TODO: remove once lctx.logits_all is removed too
+    // min/max positions of each sequence in the current ubatch
+    llama_pos seq_pos_min(llama_seq_id seq_id) const;
+    llama_pos seq_pos_max(llama_seq_id seq_id) const;
+
+    // call once before splitting the batch to reset the internal state
+    void split_reset();
+
+    // simple split, unknown number of sequence sets of unequal lengths
+    llama_ubatch split_simple(uint32_t n_ubatch);
+
+    // make ubatches of equal-length sequences sets
+    // if sequential == true, the tokens in the ubatch will have increasing sequential sequence ids
+    llama_ubatch split_equal(uint32_t n_ubatch, bool sequential);
+
+    // sequence-set-wise split - each ubatch contains a single sequence-set
+    llama_ubatch split_seq(uint32_t n_ubatch);
+
+    // a helper method for creating a well-defined ubatch of tokens
+    // TODO: support embeddings if needed in the future
+    llama_ubatch ubatch_reserve(uint32_t n_seq_tokens, uint32_t n_seqs);
+
+private:
+    void clear();
+
+    // create the next ubatch based on the provided batch indices (idxs) and the number of sequence sets (n_seqs)
+    // return llama_ubatch.n_tokens == 0 if the entire batch was consumed
+    llama_ubatch ubatch_add(const std::vector<int32_t> & idxs, uint32_t n_seqs, bool equal_seqs);
+
+    // for debugging, start with LLAMA_BATCH_DEBUG=2
+    void ubatch_print(const llama_ubatch & ubatch, int debug);
+
+    llama_batch batch;
+
+    // only for debugging purposes
+    const llama_vocab * vocab;
+
+    // TODO: this is more of a temporary solution until we have a better way to handle multiple positions per token/embd
+    //       ref: https://github.com/ggml-org/llama.cpp/issues/13694#issuecomment-2983871762
+    const uint32_t n_pos_per_embd;
+
+    uint32_t n_embd;
+    uint32_t n_outputs;
+
+    std::array<llama_seq_id, 1> seq_id_0 = { 0 }; // default sequence id
+
+    std::vector<llama_pos>      pos;
+    std::vector<int32_t>        n_seq_id;
+    std::vector<llama_seq_id *> seq_id;
+    std::vector<llama_seq_id>   seq_id_unq;
+    std::vector<int32_t>        seq_idx;
+    std::vector<int8_t>         output;
+
+    using pos_set_t = std::set<llama_pos>;
+    using seq_cpl_t = std::vector<bool>;
+
+    // helper flag to quickly determine if there are any coupled sequences in the batch
+    bool has_cpl;
+
+    std::vector<pos_set_t> seq_pos; // seq_pos[s]: the set of positions in sequence s
+    std::vector<seq_cpl_t> seq_cpl; // seq_cpl[s0][s1]: if sequence s0 is coupled to sequence s1
+
+    using idx_vec_t = std::vector<int32_t>;
+    using seq_set_t = std::bitset<LLAMA_MAX_SEQ>;
+
+    std::vector<seq_set_t> seq_set; // seq_set[i]: the sequence set of token i
+
+    std::unordered_map<seq_set_t, idx_vec_t> seq_set_map; // the indices at which the sequence set appears
 
-    // sorted indices into the batch
-    std::vector<int64_t> ids;
     // batch indices of the output
-    std::vector<int64_t> out_ids;
-    std::vector<llama_sbatch_seq> seq;
+    std::vector<int32_t> out_ids;
 
-    const llama_batch * batch = nullptr;
+    uint32_t n_used;
 
-    // buffers for the ubatches
-    // TODO: very hacky, this needs a complete rework
-    struct ubatch_data {
+    // used[i] indicates if token i has already been used in a previous ubatch
+    std::vector<bool> used;
+
+    // llama_ubatch points to this data:
+    struct ubatch {
         std::vector<llama_token>    token;
         std::vector<float>          embd;
         std::vector<llama_pos>      pos;
         std::vector<int32_t>        n_seq_id;
         std::vector<llama_seq_id *> seq_id;
+        std::vector<llama_seq_id>   seq_id_unq;
+        std::vector<int32_t>        seq_idx;
         std::vector<int8_t>         output;
     };
 
-    std::vector<ubatch_data> udatas;
+    // current splitting state:
+    std::vector<ubatch> ubatches;
 
-    llama_ubatch reserve_ubatch(size_t n_ubatch, bool has_embd = false);
-
-    void add_seq_to_ubatch(llama_ubatch & ubatch, llama_sbatch_seq & seq, size_t length);
-
-    // simple split, unknown number of sequences of unequal lengths
-    llama_ubatch split_simple(size_t n_ubatch);
-
-    // make batches of equal-length sequences
-    llama_ubatch split_equal(size_t n_ubatch);
-
-    // sequence-wise split
-    llama_ubatch split_seq(size_t n_ubatch);
-
-    llama_sbatch() = default;
-    llama_sbatch(const llama_batch & batch, size_t n_embd, bool simple_split = false, bool logits_all = false);
-};
-
-// temporary allocate memory for the input batch if needed
-struct llama_batch_allocr {
-    struct llama_batch batch;
-
-    std::array<llama_seq_id, 1> seq_id_0 = { 0 }; // default sequence id
-    std::vector<llama_pos>      pos;
-    std::vector<int32_t>        n_seq_id;
-    std::vector<llama_seq_id *> seq_id;
-    std::vector<int8_t>         logits;
-
-    // optionally fulfill the batch returned by llama_batch_get_one
-    llama_batch_allocr(struct llama_batch in_batch, llama_pos p0);
+    int debug;
 };
diff --git a/src/llama-chat.cpp b/src/llama-chat.cpp
index d12743e6b..5d317f4ee 100644
--- a/src/llama-chat.cpp
+++ b/src/llama-chat.cpp
@@ -183,6 +183,8 @@ llm_chat_template llm_chat_detect_template(const std::string & tmpl) {
         return LLM_CHAT_TEMPLATE_BAILING;
     } else if (tmpl_contains("<|header_start|>") && tmpl_contains("<|header_end|>")) {
         return LLM_CHAT_TEMPLATE_LLAMA4;
+    } else if (tmpl_contains("<|endofuserprompt|>")) {
+        return LLM_CHAT_TEMPLATE_DOTS1;
     }
     return LLM_CHAT_TEMPLATE_UNKNOWN;
 }
@@ -331,7 +333,7 @@ int32_t llm_chat_apply_template(
             std::string role(message->role);
             if (role == "system") {
                 // there is no system message for gemma, but we will merge it with user prompt, so nothing is broken
-                system_prompt = trim(message->content);
+                system_prompt += trim(message->content);
                 continue;
             }
             // in gemma, "assistant" is "model"
@@ -353,7 +355,7 @@ int32_t llm_chat_apply_template(
             std::string role(message->role);
             if (role == "system") {
                 // there is no system message support, we will merge it with user prompt
-                system_prompt = message->content;
+                system_prompt += message->content;
                 continue;
             } else if (role == "user") {
                 ss << "Human: ";
@@ -526,12 +528,17 @@ int32_t llm_chat_apply_template(
         }
     } else if (tmpl == LLM_CHAT_TEMPLATE_RWKV_WORLD) {
         // this template requires the model to have "\n\n" as EOT token
-        for (auto message : chat) {
-            std::string role(message->role);
-            if (role == "user") {
-                ss << "User: " << message->content << "\n\nAssistant:";
-            } else {
-                ss << message->content << "\n\n";
+        for (size_t i = 0; i < chat.size(); i++) {
+            std::string role(chat[i]->role);
+            if (role == "system") {
+                ss << "System: " << trim(chat[i]->content) << "\n\n";
+            } else if (role == "user") {
+                ss << "User: " << trim(chat[i]->content) << "\n\n";
+                if (i == chat.size() - 1) {
+                    ss << "Assistant:";
+                }
+            } else if (role == "assistant") {
+                ss << "Assistant: " << trim(chat[i]->content) << "\n\n";
             }
         }
     } else if (tmpl == LLM_CHAT_TEMPLATE_GRANITE) {
@@ -643,6 +650,21 @@ int32_t llm_chat_apply_template(
         if (add_ass) {
             ss << "Assistant:";
         }
+    } else if (tmpl == LLM_CHAT_TEMPLATE_DOTS1) {
+        // dots.llm1.inst (DOTS1)
+        for (auto message : chat) {
+            std::string role(message->role);
+            if (role == "system") {
+                ss << "<|system|>" << message->content << "<|endofsystem|>";
+            } else if (role == "user") {
+                ss << "<|userprompt|>" << message->content << "<|endofuserprompt|>";
+            } else {
+                ss << "<|response|>" << message->content << "<|endofresponse|>";
+            }
+        }
+        if (add_ass) {
+            ss << "<|response|>";
+        }
     } else {
         // template not supported
         return -1;
diff --git a/src/llama-chat.h b/src/llama-chat.h
index db24ade21..38800010a 100644
--- a/src/llama-chat.h
+++ b/src/llama-chat.h
@@ -43,6 +43,7 @@ enum llm_chat_template {
     LLM_CHAT_TEMPLATE_BAILING,
     LLM_CHAT_TEMPLATE_LLAMA4,
     LLM_CHAT_TEMPLATE_SMOLVLM,
+    LLM_CHAT_TEMPLATE_DOTS1,
     LLM_CHAT_TEMPLATE_UNKNOWN,
 };
 
diff --git a/src/llama-context.cpp b/src/llama-context.cpp
index 4ab574387..06e93b19c 100644
--- a/src/llama-context.cpp
+++ b/src/llama-context.cpp
@@ -1,10 +1,11 @@
 #include "llama-context.h"
 
 #include "llama-impl.h"
+#include "llama-batch.h"
 #include "llama-io.h"
+#include "llama-memory.h"
 #include "llama-mmap.h"
 #include "llama-model.h"
-#include "llama-kv-cache.h"
 
 #include <cinttypes>
 #include <cstring>
@@ -18,7 +19,8 @@
 llama_context::llama_context(
         const llama_model & model,
               llama_context_params params) :
-    model(model) {
+    model(model),
+    balloc(std::make_unique<llama_batch_allocr>(model.hparams.n_pos_per_embd())) {
     LLAMA_LOG_INFO("%s: constructing llama_context\n", __func__);
 
     t_start_us = model.t_start_us;
@@ -27,8 +29,8 @@ llama_context::llama_context(
     const auto & hparams = model.hparams;
 
     cparams.n_seq_max = std::max(1u, params.n_seq_max);
-    if (cparams.n_seq_max > LLAMA_MAX_PARALLEL_SEQUENCES) {
-        throw std::runtime_error("n_seq_max must be <= " + std::to_string(LLAMA_MAX_PARALLEL_SEQUENCES));
+    if (cparams.n_seq_max > LLAMA_MAX_SEQ) {
+        throw std::runtime_error("n_seq_max must be <= " + std::to_string(LLAMA_MAX_SEQ));
     }
 
     cparams.n_threads        = params.n_threads;
@@ -123,7 +125,7 @@ llama_context::llama_context(
                 __func__, n_ctx_per_seq, hparams.n_ctx_train);
     }
 
-    if (!params.swa_full && cparams.n_seq_max > 1) {
+    if (!params.swa_full && cparams.n_seq_max > 1 && hparams.is_swa_any()) {
         LLAMA_LOG_WARN("%s: requested n_seq_max (%u) > 1, but swa_full is not enabled -- performance may be degraded: %s\n",
                 __func__, cparams.n_seq_max, "https://github.com/ggml-org/llama.cpp/pull/13845#issuecomment-2924800573");
     }
@@ -277,10 +279,9 @@ llama_context::llama_context(
         int n_nodes_tg  = -1;
 
         // simulate full KV cache
-        llama_kv_cache * kv_self = static_cast<llama_kv_cache *>(memory.get());
 
-        const auto kv_state = kv_self->init_full();
-        if (!kv_state) {
+        const auto mctx = memory->init_full();
+        if (!mctx) {
             throw std::runtime_error("failed to initialize KV cache");
         }
 
@@ -288,7 +289,7 @@ llama_context::llama_context(
 
         // reserve pp graph first so that buffers are only allocated once
         {
-            auto * gf = graph_reserve(n_tokens, n_seqs, n_tokens, kv_state.get());
+            auto * gf = graph_reserve(n_tokens, n_seqs, n_tokens, mctx.get());
             if (!gf) {
                 throw std::runtime_error("failed to allocate compute pp buffers");
             }
@@ -299,7 +300,7 @@ llama_context::llama_context(
 
         // reserve with tg graph to get the number of splits and nodes
         {
-            auto * gf = graph_reserve(1, 1, 1, kv_state.get());
+            auto * gf = graph_reserve(1, 1, 1, mctx.get());
             if (!gf) {
                 throw std::runtime_error("failed to allocate compute tg buffers");
             }
@@ -310,7 +311,7 @@ llama_context::llama_context(
 
         // reserve again with pp graph to avoid ggml-alloc reallocations during inference
         {
-            auto * gf = graph_reserve(n_tokens, n_seqs, n_tokens, kv_state.get());
+            auto * gf = graph_reserve(n_tokens, n_seqs, n_tokens, mctx.get());
             if (!gf) {
                 throw std::runtime_error("failed to allocate compute pp buffers");
             }
@@ -419,40 +420,68 @@ uint32_t llama_context::n_threads_batch() const {
     return cparams.n_threads_batch;
 }
 
-llama_kv_cache * llama_context::get_kv_self() {
-    llama_kv_cache * kv_self = static_cast<llama_kv_cache *>(memory.get());
-    return kv_self;
+llama_memory_t llama_context::get_memory() const {
+    return memory.get();
 }
 
-const llama_kv_cache * llama_context::get_kv_self() const {
-    llama_kv_cache * kv_self = static_cast<llama_kv_cache *>(memory.get());
-    return kv_self;
+// deprecated
+void llama_context::kv_self_defrag_sched() {
+    if (!memory) {
+        return;
+    }
+
+    memory_force_optimize = true;
 }
 
-bool llama_context::kv_self_update() {
+// deprecated
+bool llama_context::kv_self_update(bool optimize) {
     if (!memory) {
         return false;
     }
 
-    llama_kv_cache * kv_self = static_cast<llama_kv_cache *>(memory.get());
+    {
+        // TODO: remove in the future
+        optimize |= memory_force_optimize;
+        memory_force_optimize = false;
 
-    if (!kv_self->update(*this)) {
-        // no updates have been performed
-        return false;
+        const auto mctx = memory->init_update(this, optimize);
+        switch (mctx->get_status()) {
+            case LLAMA_MEMORY_STATUS_SUCCESS:
+                {
+                    // noop
+                } break;
+            case LLAMA_MEMORY_STATUS_NO_UPDATE:
+                {
+                    // no updates need to be performed
+                    return false;
+                }
+            case LLAMA_MEMORY_STATUS_FAILED_PREPARE:
+            case LLAMA_MEMORY_STATUS_FAILED_COMPUTE:
+                {
+                    LLAMA_LOG_ERROR("%s: failed to prepare memory update\n", __func__);
+                    return false;
+                }
+        }
+
+        if (!mctx->apply()) {
+            LLAMA_LOG_ERROR("%s: failed to apply memory update\n", __func__);
+        }
     }
 
-    // if the KV cache did any computation, we have to reserve a new worst-case graph
-    const auto kv_state = kv_self->init_full();
-    if (!kv_state) {
-        throw std::runtime_error("failed to initialize KV cache");
-    }
+    // if the memory module did any computation, we have to reserve a new worst-case graph
+    {
+        const auto mctx = memory->init_full();
+        if (!mctx) {
+            throw std::runtime_error("failed to initialize memory context");
+        }
 
-    const uint32_t n_seqs   = cparams.n_seq_max;
-    const uint32_t n_tokens = std::min(cparams.n_ctx, cparams.n_ubatch);
+        const uint32_t n_seqs   = cparams.n_seq_max;
+        const uint32_t n_tokens = std::min(cparams.n_ctx, cparams.n_ubatch);
 
-    auto * gf = graph_reserve(n_tokens, n_seqs, n_tokens, kv_state.get());
-    if (!gf) {
-        LLAMA_LOG_ERROR("%s: failed to reserve graph after the KV cache update\n", __func__);
+        auto * gf = graph_reserve(n_tokens, n_seqs, n_tokens, mctx.get());
+        if (!gf) {
+            LLAMA_LOG_ERROR("%s: failed to reserve graph after the memory update\n", __func__);
+        }
     }
 
     return true;
@@ -467,7 +496,7 @@ float * llama_context::get_logits() {
 }
 
 float * llama_context::get_logits_ith(int32_t i) {
-    int32_t j = -1;
+    int64_t j = -1;
 
     try {
         if (logits == nullptr) {
@@ -490,7 +519,7 @@ float * llama_context::get_logits_ith(int32_t 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));
+            throw std::runtime_error(format("corrupt output buffer (j=%" PRId64 ", n_outputs=%d)", j, n_outputs));
         }
 
         return logits + j*model.vocab.n_tokens();
@@ -509,7 +538,7 @@ float * llama_context::get_embeddings() {
 }
 
 float * llama_context::get_embeddings_ith(int32_t i) {
-    int32_t j = -1;
+    int64_t j = -1;
 
     try {
         if (embd == nullptr) {
@@ -532,7 +561,7 @@ float * llama_context::get_embeddings_ith(int32_t 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));
+            throw std::runtime_error(format("corrupt output buffer (j=%" PRId64 ", n_outputs=%d)", j, n_outputs));
         }
 
         return embd + j*model.hparams.n_embd;
@@ -649,9 +678,9 @@ bool llama_context::apply_adapter_cvec(
     return cvec.apply(model, data, len, n_embd, il_start, il_end);
 }
 
-llm_graph_result_ptr llama_context::process_ubatch(const llama_ubatch & ubatch, llm_graph_type gtype, llama_memory_state_i * mstate, ggml_status & ret) {
-    if (mstate && !mstate->apply()) {
-        LLAMA_LOG_ERROR("%s: failed to apply memory state\n", __func__);
+llm_graph_result_ptr llama_context::process_ubatch(const llama_ubatch & ubatch, llm_graph_type gtype, llama_memory_context_i * mctx, ggml_status & ret) {
+    if (mctx && !mctx->apply()) {
+        LLAMA_LOG_ERROR("%s: failed to apply memory context\n", __func__);
         ret = GGML_STATUS_FAILED;
         return nullptr;
     }
@@ -663,7 +692,7 @@ llm_graph_result_ptr llama_context::process_ubatch(const llama_ubatch & ubatch,
         return nullptr;
     }
 
-    auto res = graph_build(ctx_compute.get(), gf, ubatch, gtype, mstate);
+    auto res = graph_build(ctx_compute.get(), gf, ubatch, gtype, mctx);
     if (!res) {
         LLAMA_LOG_ERROR("%s: failed to build graph\n", __func__);
         ret = GGML_STATUS_FAILED;
@@ -692,62 +721,47 @@ llm_graph_result_ptr llama_context::process_ubatch(const llama_ubatch & ubatch,
     return res;
 }
 
-int llama_context::encode(llama_batch & inp_batch) {
-    if (inp_batch.n_tokens == 0) {
+int llama_context::encode(const llama_batch & batch_inp) {
+    GGML_ASSERT((!batch_inp.token && batch_inp.embd) || (batch_inp.token && !batch_inp.embd)); // NOLINT
+
+    if (batch_inp.n_tokens == 0) {
         LLAMA_LOG_ERROR("%s: n_tokens == 0\n", __func__);
         return -1;
     }
 
-    // temporary allocate memory for the input batch if needed
-    // note: during encode, we always pass the full sequence starting from pos = 0
-    llama_batch_allocr batch_allocr(inp_batch, inp_batch.pos ? -1 : 0);
-
-    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
+    const int64_t n_embd = hparams.n_embd;
 
-    // TODO: move the validation to the llama_batch_allocr
-    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;
-            }
-
-            if (batch.seq_id && (batch.seq_id[i][0] < 0 || batch.seq_id[i][0] >= LLAMA_MAX_PARALLEL_SEQUENCES)) {
-                LLAMA_LOG_ERROR("%s: invalid seq_id[%d] = %d > %d\n", __func__, i, batch.seq_id[i][0], LLAMA_MAX_PARALLEL_SEQUENCES);
-                throw -1;
-            }
-        }
+    // note: during encode, we always pass the full sequence starting from pos = 0
+    if (!balloc->init(batch_inp, model.vocab, nullptr, n_embd, true)) {
+        LLAMA_LOG_ERROR("%s: failed to initialize batch\n", __func__);
+        return -1;
     }
 
+    const uint32_t n_tokens = balloc->get_n_tokens();
+
+    const llama_ubatch ubatch = balloc->split_simple(n_tokens);
+
     // 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");
+    GGML_ASSERT(cparams.n_ubatch >= n_tokens && "encoder requires n_ubatch >= n_tokens");
 
     if (t_compute_start_us == 0) {
         t_compute_start_us = ggml_time_us();
     }
 
+    // TODO: this clear of the buffer can easily be forgotten - need something better
     embd_seq.clear();
 
     n_queued_tokens += n_tokens;
 
-    const int64_t n_embd = hparams.n_embd;
-
-    llama_sbatch sbatch = llama_sbatch(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) {
+    for (uint32_t i = 0; i < n_tokens; ++i) {
         output_ids[i] = i;
     }
 
@@ -799,31 +813,28 @@ int llama_context::encode(llama_batch & inp_batch) {
                 {
                     // extract sequence embeddings
                     auto & embd_seq_out = embd_seq;
-                    embd_seq_out.clear();
 
-                    GGML_ASSERT(!ubatch.equal_seqs); // TODO: handle equal splits
+                    for (uint32_t s = 0; s < ubatch.n_seqs_unq; ++s) {
+                        const llama_seq_id seq_id  = ubatch.seq_id_unq[s];
+                        const int32_t      seq_idx = ubatch.seq_idx[seq_id];
 
-                    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));
+                        ggml_backend_tensor_get_async(backend_embd, t_embd, embd_seq_out[seq_id].data(), (n_embd*seq_idx)*sizeof(float), n_embd*sizeof(float));
                     }
                 } break;
             case LLAMA_POOLING_TYPE_RANK:
                 {
-                    // extract the rerank score - a single float per sequence
+                    // extract the rerank score - n_cls_out floats 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));
+                    const uint32_t n_cls_out = hparams.n_cls_out;
+
+                    for (uint32_t s = 0; s < ubatch.n_seqs_unq; ++s) {
+                        const llama_seq_id seq_id  = ubatch.seq_id_unq[s];
+                        const int32_t      seq_idx = ubatch.seq_idx[seq_id];
+
+                        embd_seq_out[seq_id].resize(n_cls_out);
+                        ggml_backend_tensor_get_async(backend_embd, t_embd, embd_seq_out[seq_id].data(), (n_cls_out*seq_idx)*sizeof(float), n_cls_out*sizeof(float));
                     }
                 } break;
             case LLAMA_POOLING_TYPE_UNSPECIFIED:
@@ -848,12 +859,16 @@ int llama_context::encode(llama_batch & inp_batch) {
         cross.v_embd.resize(cross.n_embd*cross.n_enc);
         memcpy(cross.v_embd.data(), embd, ggml_nbytes(t_embd));
 
+        const auto & batch = balloc->get_batch();
+
         // 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++) {
+        for (uint32_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];
+
+            for (int s = 0; s < batch.n_seq_id[i]; s++) {
+                const llama_seq_id seq_id = batch.seq_id[i][s];
+
                 cross.seq_ids_enc[i].insert(seq_id);
             }
         }
@@ -862,53 +877,42 @@ int llama_context::encode(llama_batch & inp_batch) {
     return 0;
 }
 
-int llama_context::decode(llama_batch & inp_batch) {
+int llama_context::decode(const llama_batch & batch_inp) {
+    GGML_ASSERT((!batch_inp.token && batch_inp.embd) || (batch_inp.token && !batch_inp.embd)); // NOLINT
+
     if (!memory) {
         LLAMA_LOG_DEBUG("%s: cannot decode batches with this context (calling encode() instead)\n", __func__);
-        return encode(inp_batch);
+        return encode(batch_inp);
     }
 
-    if (inp_batch.n_tokens == 0) {
+    if (batch_inp.n_tokens == 0) {
         LLAMA_LOG_ERROR("%s: n_tokens == 0\n", __func__);
         return -1;
     }
 
-    if (!inp_batch.pos) {
-        if (inp_batch.seq_id) {
-            LLAMA_LOG_ERROR("%s: pos == NULL, but seq_id != NULL\n", __func__);
-            return -1;
-        }
-    }
-
-    llama_kv_cache * kv_self = static_cast<llama_kv_cache *>(memory.get());
-
-    // temporary allocate memory for the input batch if needed
-    llama_batch_allocr batch_allocr(inp_batch, inp_batch.pos ? -1 : kv_self->seq_pos_max(0) + 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_embd  = hparams.n_embd;
 
-    const int64_t n_tokens_all = batch.n_tokens;
-    const int64_t n_embd       = hparams.n_embd;
+    // when computing embeddings, all tokens are output
+    const bool output_all = cparams.embeddings;
 
-    GGML_ASSERT((!batch.token && batch.embd) || (batch.token && !batch.embd)); // NOLINT
+    if (!balloc->init(batch_inp, vocab, memory.get(), n_embd, output_all)) {
+        LLAMA_LOG_ERROR("%s: failed to initialize batch\n", __func__);
+        return -1;
+    }
 
-    // TODO: move the validation to the llama_batch_allocr
-    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]);
-                return -1;
-            }
+    const uint32_t n_tokens_all  = balloc->get_n_tokens();
+    const uint32_t n_outputs_all = balloc->get_n_outputs();
 
-            if (batch.seq_id && (batch.seq_id[i][0] < 0 || batch.seq_id[i][0] >= LLAMA_MAX_PARALLEL_SEQUENCES)) {
-                LLAMA_LOG_ERROR("%s: invalid seq_id[%" PRId64 "] = %d >= %d\n", __func__, i, batch.seq_id[i][0], LLAMA_MAX_PARALLEL_SEQUENCES);
-                return -1;
-            }
+    if (output_all) {
+        // require that all tokens are output
+        if (n_outputs_all != n_tokens_all) {
+            LLAMA_LOG_ERROR("%s: pooled embedding requires that all tokens are output (n_outputs_all = %d, n_tokens_all = %d)\n",
+                    __func__, n_outputs_all, n_tokens_all);
+            return -1;
         }
     }
 
@@ -921,61 +925,52 @@ int llama_context::decode(llama_batch & inp_batch) {
     }
     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;
-
+    // TODO: this clear of the buffer can easily be forgotten - need something better
     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 (embd_pooled) {
-        n_outputs_all = n_tokens_all;
-    } else {
-        // keep last output only
-        n_outputs_all = 1;
-    }
+    bool did_optimize = false;
 
     // handle any pending defrags/shifts
-    kv_self_update();
+    kv_self_update(false);
 
-    llama_memory_state_ptr kv_state;
-
-    bool did_defrag = false;
+    llama_memory_context_ptr mctx;
 
     while (true) {
-        kv_state = kv_self->init_batch(batch, cparams.n_ubatch, embd_pooled, /* logits_all */ n_outputs_all == n_tokens_all);
-        if (!kv_state) {
+        mctx = memory->init_batch(*balloc, cparams.n_ubatch, output_all);
+        if (!mctx) {
             return -2;
         }
 
-        switch (kv_state->get_status()) {
+        switch (mctx->get_status()) {
             case LLAMA_MEMORY_STATUS_SUCCESS:
                 {
                 } break;
+            case LLAMA_MEMORY_STATUS_NO_UPDATE:
+                {
+                    LLAMA_LOG_ERROR("%s: unexpected memory context status: %d\n", __func__, mctx->get_status());
+
+                    return -2;
+                }
             case LLAMA_MEMORY_STATUS_FAILED_PREPARE:
                 {
-                    if (!did_defrag) {
-                        did_defrag = true;
+                    if (!did_optimize) {
+                        did_optimize = true;
 
-                        kv_self->defrag_sched(-1.0f);
-                        if (kv_self_update()) {
-                            LLAMA_LOG_DEBUG("%s: failed to init batch of size %d, retrying after defrag\n", __func__, batch.n_tokens);
+                        if (kv_self_update(true)) {
+                            LLAMA_LOG_DEBUG("%s: retrying batch size %d after cache optimization\n", __func__, balloc->get_n_tokens());
 
                             continue;
                         }
                     }
 
-                    LLAMA_LOG_WARN("%s: failed to find KV cache slot for batch of size %d\n", __func__, batch.n_tokens);
+                    LLAMA_LOG_WARN("%s: failed to find a memory slot for batch of size %d\n", __func__, balloc->get_n_tokens());
 
                     return 1;
                 }
             case LLAMA_MEMORY_STATUS_FAILED_COMPUTE:
                 {
+                    LLAMA_LOG_ERROR("%s: compute failed while preparing batch of size %d\n", __func__, balloc->get_n_tokens());
+
                     return -2;
                 }
         }
@@ -985,23 +980,22 @@ int llama_context::decode(llama_batch & inp_batch) {
 
     // 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);
+        LLAMA_LOG_ERROR("%s: could not reserve space for batch with %d outputs\n", __func__, n_outputs_all);
         return -2;
     };
 
     int64_t n_outputs_prev = 0;
 
     do {
-        const auto & ubatch = kv_state->get_ubatch();
+        const auto & ubatch = mctx->get_ubatch();
 
-        // count the outputs in this u_batch
+        // count the outputs in this ubatch
         {
             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);
                 }
@@ -1015,11 +1009,14 @@ int llama_context::decode(llama_batch & inp_batch) {
         ggml_backend_sched_set_eval_callback(sched.get(), cparams.cb_eval, cparams.cb_eval_user_data);
 
         ggml_status status;
-        const auto res = process_ubatch(ubatch, LLM_GRAPH_TYPE_DECODER, kv_state.get(), status);
+        const auto res = process_ubatch(ubatch, LLM_GRAPH_TYPE_DECODER, mctx.get(), status);
 
         if (!res) {
             // the last ubatch failed or was aborted -> remove all positions of that ubatch from the KV cache
-            llama_pos pos_min[LLAMA_MAX_PARALLEL_SEQUENCES] = { std::numeric_limits<llama_pos>::max() };
+            llama_pos pos_min[LLAMA_MAX_SEQ];
+            for (int s = 0; s < LLAMA_MAX_SEQ; ++s) {
+                pos_min[s] = std::numeric_limits<llama_pos>::max();
+            }
 
             for (uint32_t i = 0; i < ubatch.n_tokens; ++i) {
                 const auto & seq_id = ubatch.seq_id[i][0];
@@ -1027,14 +1024,14 @@ int llama_context::decode(llama_batch & inp_batch) {
                 pos_min[seq_id] = std::min(pos_min[seq_id], ubatch.pos[i]);
             }
 
-            for (int s = 0; s < LLAMA_MAX_PARALLEL_SEQUENCES; ++s) {
+            for (int s = 0; s < LLAMA_MAX_SEQ; ++s) {
                 if (pos_min[s] == std::numeric_limits<llama_pos>::max()) {
                     continue;
                 }
 
                 LLAMA_LOG_WARN("%s: removing KV cache entries for seq_id = %d, pos = [%d, +inf)\n", __func__, s, pos_min[s]);
 
-                llama_kv_self_seq_rm(this, s, pos_min[s], -1);
+                memory->seq_rm(s, pos_min[s], -1);
             }
 
             switch (status) {
@@ -1050,7 +1047,7 @@ int llama_context::decode(llama_batch & inp_batch) {
         //    ggml_graph_dump_dot(gf, NULL, "llama.dot");
         //}
 
-        auto * t_logits = cparams.embeddings ? nullptr         : res->get_logits();
+        auto * t_logits = res->get_logits();
         auto * t_embd   = cparams.embeddings ? res->get_embd() : nullptr;
 
         if (t_embd && res->get_embd_pooled()) {
@@ -1097,27 +1094,27 @@ int llama_context::decode(llama_batch & inp_batch) {
                         // 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;
-                            }
+                        for (uint32_t s = 0; s < ubatch.n_seqs_unq; ++s) {
+                            const llama_seq_id seq_id  = ubatch.seq_id_unq[s];
+                            const int32_t      seq_idx = ubatch.seq_idx[seq_id];
+
                             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));
+                            ggml_backend_tensor_get_async(backend_embd, t_embd, embd_seq_out[seq_id].data(), (n_embd*seq_idx)*sizeof(float), n_embd*sizeof(float));
                         }
                     } break;
                 case LLAMA_POOLING_TYPE_RANK:
                     {
-                        // extract the rerank score - a single float per sequence
+                        // extract the rerank score - n_cls_out floats 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));
+                        const uint32_t n_cls_out = hparams.n_cls_out;
+
+                        for (uint32_t s = 0; s < ubatch.n_seqs_unq; ++s) {
+                            const llama_seq_id seq_id  = ubatch.seq_id_unq[s];
+                            const int32_t      seq_idx = ubatch.seq_idx[seq_id];
+
+                            embd_seq_out[seq_id].resize(n_cls_out);
+                            ggml_backend_tensor_get_async(backend_embd, t_embd, embd_seq_out[seq_id].data(), (n_cls_out*seq_idx)*sizeof(float), n_cls_out*sizeof(float));
                         }
                     } break;
                 case LLAMA_POOLING_TYPE_UNSPECIFIED:
@@ -1128,20 +1125,20 @@ int llama_context::decode(llama_batch & inp_batch) {
         }
 
         n_outputs_prev += n_outputs;
-    } while (kv_state->next());
+    } while (mctx->next());
 
     // set to total number of outputs in the batch, for use in llama_get_logits_ith
     n_outputs = n_outputs_all;
 
     // set output mappings
-    {
+    if (n_outputs > 0) {
         bool sorted_output = true;
 
-        auto & out_ids = kv_state->out_ids();
+        auto & out_ids = balloc->get_out_ids();
 
-        GGML_ASSERT(out_ids.size() == (size_t) n_outputs_all);
+        GGML_ASSERT(out_ids.size() == (size_t) n_outputs);
 
-        for (int64_t i = 0; i < n_outputs_all; ++i) {
+        for (int64_t i = 0; i < n_outputs; ++i) {
             int64_t out_id = out_ids[i];
             output_ids[out_id] = i;
             if (out_id != i) {
@@ -1153,20 +1150,22 @@ int llama_context::decode(llama_batch & inp_batch) {
         // note: this is mostly relevant for recurrent models atm
         if (!sorted_output) {
             const uint32_t n_vocab = model.vocab.n_tokens();
-            const uint32_t n_embd  = model.hparams.n_embd;
+            const uint64_t n_embd  = model.hparams.n_embd;
 
             GGML_ASSERT((size_t) n_outputs == out_ids.size());
 
             // TODO: is there something more efficient which also minimizes swaps?
             // selection sort, to minimize swaps (from https://en.wikipedia.org/wiki/Selection_sort)
-            for (int32_t i = 0; i < n_outputs - 1; ++i) {
-                int32_t j_min = i;
-                for (int32_t j = i + 1; j < n_outputs; ++j) {
+            for (uint32_t i = 0; i < n_outputs - 1; ++i) {
+                uint32_t j_min = i;
+                for (uint32_t j = i + 1; j < n_outputs; ++j) {
                     if (out_ids[j] < out_ids[j_min]) {
                         j_min = j;
                     }
                 }
-                if (j_min == i) { continue; }
+                if (j_min == i) {
+                    continue;
+                }
                 std::swap(out_ids[i], out_ids[j_min]);
                 if (logits_size > 0) {
                     for (uint32_t k = 0; k < n_vocab; k++) {
@@ -1179,8 +1178,10 @@ int llama_context::decode(llama_batch & inp_batch) {
                     }
                 }
             }
+
             std::fill(output_ids.begin(), output_ids.end(), -1);
-            for (int32_t i = 0; i < n_outputs; ++i) {
+
+            for (uint32_t i = 0; i < n_outputs; ++i) {
                 output_ids[out_ids[i]] = i;
             }
         }
@@ -1189,11 +1190,6 @@ int llama_context::decode(llama_batch & inp_batch) {
     // 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.defrag_thold > 0.0f) {
-        kv_self->defrag_sched(cparams.defrag_thold);
-    }
-
     // 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());
@@ -1205,7 +1201,7 @@ int llama_context::decode(llama_batch & inp_batch) {
 // output
 //
 
-int32_t llama_context::output_reserve(int32_t n_outputs) {
+uint32_t llama_context::output_reserve(int32_t n_outputs) {
     const auto & hparams = model.hparams;
     const auto & vocab   = model.vocab;
 
@@ -1215,9 +1211,8 @@ int32_t llama_context::output_reserve(int32_t n_outputs) {
     const auto n_vocab = vocab.n_tokens();
     const auto n_embd  = hparams.n_embd;
 
-    // TODO: use a per-batch flag for logits presence instead
-    bool has_logits = !cparams.embeddings;
-    bool has_embd   =  cparams.embeddings && (cparams.pooling_type == LLAMA_POOLING_TYPE_NONE);
+    bool has_logits = true;
+    bool has_embd   = cparams.embeddings;
 
     // TODO: hacky enc-dec support
     if (model.arch == LLM_ARCH_T5) {
@@ -1271,8 +1266,7 @@ int32_t llama_context::output_reserve(int32_t n_outputs) {
     // set all ids as invalid (negative)
     std::fill(output_ids.begin(), output_ids.end(), -1);
 
-    this->n_outputs     = 0;
-    this->n_outputs_max = n_outputs_max;
+    this->n_outputs = 0;
 
     return n_outputs_max;
 }
@@ -1297,11 +1291,11 @@ ggml_cgraph * llama_context::graph_init() {
     return ggml_new_graph_custom(ctx_compute.get(), graph_max_nodes(), false);
 }
 
-ggml_cgraph * llama_context::graph_reserve(uint32_t n_tokens, uint32_t n_seqs, uint32_t n_outputs, const llama_memory_state_i * mstate) {
+ggml_cgraph * llama_context::graph_reserve(uint32_t n_tokens, uint32_t n_seqs, uint32_t n_outputs, const llama_memory_context_i * mctx) {
     LLAMA_LOG_DEBUG("%s: reserving a graph for ubatch with n_tokens = %4u, n_seqs = %2u, n_outputs = %4u\n", __func__, n_tokens, n_seqs, n_outputs);
 
     if (n_tokens % n_seqs != 0) {
-        n_tokens = (n_tokens / n_seqs) * n_seqs;
+        n_tokens = ((n_tokens + (n_seqs - 1)) / n_seqs) * n_seqs; // round to next multiple of n_seqs
         n_outputs = std::min(n_outputs, n_tokens);
 
         LLAMA_LOG_DEBUG("%s: making n_tokens a multiple of n_seqs - n_tokens = %u, n_seqs = %u, n_outputs = %u\n", __func__, n_tokens, n_seqs, n_outputs);
@@ -1313,11 +1307,11 @@ ggml_cgraph * llama_context::graph_reserve(uint32_t n_tokens, uint32_t n_seqs, u
 
     this->n_outputs = n_outputs;
 
-    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};
+    llama_batch_allocr balloc(model.hparams.n_pos_per_embd());
+    llama_ubatch ubatch = balloc.ubatch_reserve(n_tokens/n_seqs, n_seqs);
 
     auto * gf = graph_init();
-    auto res = graph_build(ctx_compute.get(), gf, ubatch, LLM_GRAPH_TYPE_DEFAULT, mstate);
+    auto res = graph_build(ctx_compute.get(), gf, ubatch, LLM_GRAPH_TYPE_DEFAULT, mctx);
 
     this->n_outputs = save_n_outputs;
 
@@ -1338,11 +1332,11 @@ ggml_cgraph * llama_context::graph_reserve(uint32_t n_tokens, uint32_t n_seqs, u
 }
 
 llm_graph_result_ptr llama_context::graph_build(
-                    ggml_context * ctx,
-                     ggml_cgraph * gf,
-              const llama_ubatch & ubatch,
-                  llm_graph_type   gtype,
-      const llama_memory_state_i * mstate) {
+                      ggml_context * ctx,
+                       ggml_cgraph * gf,
+                const llama_ubatch & ubatch,
+                    llm_graph_type   gtype,
+      const llama_memory_context_i * mctx) {
     return model.build_graph(
             {
                 /*.ctx         =*/ ctx,
@@ -1354,7 +1348,7 @@ llm_graph_result_ptr llama_context::graph_build(
                 /*.backend_cpu =*/ backend_cpu,
                 /*.cvec        =*/ &cvec,
                 /*.loras       =*/ &loras,
-                /*.mstate      =*/ mstate,
+                /*.mctx        =*/ mctx,
                 /*.cross       =*/ &cross,
                 /*.n_outputs   =*/ n_outputs,
                 /*.cb          =*/ graph_get_cb(),
@@ -1763,14 +1757,12 @@ size_t llama_context::state_write_data(llama_io_write_i & io) {
 
         std::vector<int32_t> 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];
+            int64_t pos = output_ids[i];
             if (pos >= 0) {
                 GGML_ASSERT(pos < n_outputs);
                 w_output_pos[pos] = i;
@@ -1810,11 +1802,9 @@ size_t llama_context::state_write_data(llama_io_write_i & io) {
         }
     }
 
-    llama_kv_cache * kv_self = static_cast<llama_kv_cache *>(memory.get());
-
-    if (kv_self != nullptr) {
+    if (memory != nullptr) {
         LLAMA_LOG_DEBUG("%s: - writing KV self\n", __func__);
-        kv_self->state_write(io);
+        memory->state_write(io);
     }
 
     return io.n_bytes();
@@ -1901,9 +1891,7 @@ size_t llama_context::state_read_data(llama_io_read_i & io) {
     if (memory) {
         LLAMA_LOG_DEBUG("%s: - reading KV self\n", __func__);
 
-        llama_kv_cache * kv_self = static_cast<llama_kv_cache *>(memory.get());
-
-        kv_self->state_read(io);
+        memory->state_read(io);
     }
 
     return io.n_bytes();
@@ -1913,9 +1901,7 @@ size_t llama_context::state_seq_write_data(llama_io_write_i & io, llama_seq_id s
     GGML_UNUSED(seq_id);
 
     if (memory) {
-        llama_kv_cache * kv_self = static_cast<llama_kv_cache *>(memory.get());
-
-        kv_self->state_write(io, seq_id);
+        memory->state_write(io, seq_id);
     }
 
     return io.n_bytes();
@@ -1925,9 +1911,7 @@ size_t llama_context::state_seq_read_data(llama_io_read_i & io, llama_seq_id seq
     GGML_UNUSED(seq_id);
 
     if (memory) {
-        llama_kv_cache * kv_self = static_cast<llama_kv_cache *>(memory.get());
-
-        kv_self->state_read(io, seq_id);
+        memory->state_read(io, seq_id);
     }
 
     return io.n_bytes();
@@ -2032,9 +2016,7 @@ void llama_context::opt_epoch_iter(
     const uint32_t n_batch  = std::min(this->n_batch(),  n_ctx);
     const uint32_t n_ubatch = std::min(this->n_ubatch(), n_batch);
 
-    llama_kv_cache * kv_self = static_cast<llama_kv_cache *>(memory.get());
-
-    kv_self->clear();
+    memory->clear(true);
 
     for (uint32_t pos_ctx = 0; pos_ctx < n_ctx; pos_ctx += n_batch) {
         batch.n_tokens = n_batch;
@@ -2046,42 +2028,44 @@ void llama_context::opt_epoch_iter(
             batch.logits  [pos_batch]    = true;
         }
 
-        const auto n_tokens_all = batch.n_tokens;
+        if (!balloc->init(batch, model.vocab, nullptr, model.hparams.n_embd, true)) {
+            LLAMA_LOG_ERROR("%s: failed to initialize batch\n", __func__);
+            return;
+        }
+
+        const uint32_t n_tokens_all = balloc->get_n_tokens();
 
         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 = n_tokens_all;
+        uint32_t n_outputs_all = n_tokens_all;
 
-        auto kv_state = kv_self->init_batch(batch, cparams.n_ubatch, embd_pooled, /* logits_all */ true);
-        if (!kv_state || kv_state->get_status() != LLAMA_MEMORY_STATUS_SUCCESS) {
+        auto mctx = memory->init_batch(*balloc, cparams.n_ubatch, true);
+        if (!mctx || mctx->get_status() != LLAMA_MEMORY_STATUS_SUCCESS) {
             LLAMA_LOG_ERROR("%s: could not initialize batch\n", __func__);
             break;
         }
 
         // 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);
+            LLAMA_LOG_ERROR("%s: could not reserve space for batch with %d outputs\n", __func__, n_outputs_all);
             GGML_ABORT("TODO: handle this error");
         };
 
         uint32_t pos_batch = 0;
         do {
-            const auto & ubatch = kv_state->get_ubatch();
+            const auto & ubatch = mctx->get_ubatch();
 
             n_outputs = ubatch.n_tokens;
 
-            if (!kv_state->apply()) {
-                LLAMA_LOG_ERROR("%s: failed to update the memory state\n", __func__);
+            if (!mctx->apply()) {
+                LLAMA_LOG_ERROR("%s: failed to update the memory context\n", __func__);
                 break;
             }
 
             auto * gf = graph_init();
-            auto res = graph_build(ctx_compute.get(), gf, ubatch, LLM_GRAPH_TYPE_DEFAULT, kv_state.get());
+            auto res = graph_build(ctx_compute.get(), gf, ubatch, LLM_GRAPH_TYPE_DEFAULT, mctx.get());
 
             struct ggml_context * ctx_compute_opt;
             {
@@ -2116,7 +2100,7 @@ void llama_context::opt_epoch_iter(
             ggml_free(ctx_compute_opt);
 
             pos_batch += ubatch.n_tokens;
-        } while (kv_state->next());
+        } while (mctx->next());
     }
 }
 
@@ -2277,13 +2261,14 @@ const llama_model * llama_get_model(const llama_context * ctx) {
     return &ctx->get_model();
 }
 
+// deprecated
 llama_kv_cache * llama_get_kv_self(llama_context * ctx) {
-    return ctx->get_kv_self();
+    return dynamic_cast<llama_kv_cache *>(ctx->get_memory());
 }
 
 // deprecated
 void llama_kv_self_update(llama_context * ctx) {
-    ctx->kv_self_update();
+    ctx->kv_self_update(false);
 }
 
 enum llama_pooling_type llama_pooling_type(const llama_context * ctx) {
@@ -2398,13 +2383,118 @@ int32_t llama_apply_adapter_cvec(
     return res ? 0 : -1;
 }
 
+//
+// memory
+//
+
+llama_memory_t llama_get_memory(const struct llama_context * ctx) {
+    return ctx->get_memory();
+}
+
+void llama_memory_clear(llama_memory_t mem, bool data) {
+    if (!mem) {
+        return;
+    }
+
+    mem->clear(data);
+}
+
+bool llama_memory_seq_rm(
+        llama_memory_t mem,
+          llama_seq_id seq_id,
+             llama_pos p0,
+             llama_pos p1) {
+    if (!mem) {
+        return true;
+    }
+
+    return mem->seq_rm(seq_id, p0, p1);
+}
+
+void llama_memory_seq_cp(
+        llama_memory_t mem,
+          llama_seq_id seq_id_src,
+          llama_seq_id seq_id_dst,
+             llama_pos p0,
+             llama_pos p1) {
+    if (!mem) {
+        return;
+    }
+
+    mem->seq_cp(seq_id_src, seq_id_dst, p0, p1);
+}
+
+void llama_memory_seq_keep(
+        llama_memory_t mem,
+          llama_seq_id seq_id) {
+    if (!mem) {
+        return;
+    }
+
+    mem->seq_keep(seq_id);
+}
+
+void llama_memory_seq_add(
+        llama_memory_t mem,
+          llama_seq_id seq_id,
+             llama_pos p0,
+             llama_pos p1,
+             llama_pos delta) {
+    if (!mem) {
+        return;
+    }
+
+    mem->seq_add(seq_id, p0, p1, delta);
+}
+
+void llama_memory_seq_div(
+        llama_memory_t mem,
+          llama_seq_id seq_id,
+             llama_pos p0,
+             llama_pos p1,
+                   int d) {
+    if (!mem) {
+        return;
+    }
+
+    mem->seq_div(seq_id, p0, p1, d);
+}
+
+llama_pos llama_memory_seq_pos_min(
+        llama_memory_t mem,
+          llama_seq_id seq_id) {
+    if (!mem) {
+        return -1;
+    }
+
+    return mem->seq_pos_min(seq_id);
+}
+
+llama_pos llama_memory_seq_pos_max(
+        llama_memory_t mem,
+          llama_seq_id seq_id) {
+    if (!mem) {
+        return -1;
+    }
+
+    return mem->seq_pos_max(seq_id);
+}
+
+bool llama_memory_can_shift(llama_memory_t mem) {
+    if (!mem) {
+        return false;
+    }
+
+    return mem->get_can_shift();
+}
+
 //
 // kv cache
 //
 
 // deprecated
 int32_t llama_kv_self_n_tokens(const llama_context * ctx) {
-    const auto * kv = ctx->get_kv_self();
+    const auto * kv = llama_get_memory(ctx);
     if (!kv) {
         return 0;
     }
@@ -2426,7 +2516,7 @@ int32_t llama_kv_self_n_tokens(const llama_context * ctx) {
 // deprecated
 // note: this is the same as above - will be removed anyway, so it's ok
 int32_t llama_kv_self_used_cells(const llama_context * ctx) {
-    const auto * kv = ctx->get_kv_self();
+    const auto * kv = llama_get_memory(ctx);
     if (!kv) {
         return 0;
     }
@@ -2445,115 +2535,119 @@ int32_t llama_kv_self_used_cells(const llama_context * ctx) {
     return res;
 }
 
+// deprecated
 void llama_kv_self_clear(llama_context * ctx) {
-    auto * kv = ctx->get_kv_self();
+    auto * kv = llama_get_memory(ctx);
     if (!kv) {
         return;
     }
 
-    kv->clear();
+    llama_memory_clear(kv, true);
 }
 
+// deprecated
 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();
+    auto * kv = llama_get_memory(ctx);
     if (!kv) {
         return true;
     }
 
-    return kv->seq_rm(seq_id, p0, p1);
+    return llama_memory_seq_rm(kv, seq_id, p0, p1);
 }
 
+// deprecated
 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();
+    auto * kv = llama_get_memory(ctx);
     if (!kv) {
         return;
     }
 
-    kv->seq_cp(seq_id_src, seq_id_dst, p0, p1);
+    llama_memory_seq_cp(kv, seq_id_src, seq_id_dst, p0, p1);
 }
 
+// deprecated
 void llama_kv_self_seq_keep(llama_context * ctx, llama_seq_id seq_id) {
-    auto * kv = ctx->get_kv_self();
+    auto * kv = llama_get_memory(ctx);
     if (!kv) {
         return;
     }
 
-    kv->seq_keep(seq_id);
+    llama_memory_seq_keep(kv, seq_id);
 }
 
+// deprecated
 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();
+    auto * kv = llama_get_memory(ctx);
     if (!kv) {
         return;
     }
 
-    kv->seq_add(seq_id, p0, p1, delta);
+    llama_memory_seq_add(kv, seq_id, p0, p1, delta);
 }
 
+// deprecated
 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();
+    auto * kv = llama_get_memory(ctx);
     if (!kv) {
         return;
     }
 
-    kv->seq_div(seq_id, p0, p1, d);
+    llama_memory_seq_div(kv, seq_id, p0, p1, d);
 }
 
+// deprecated
 llama_pos llama_kv_self_seq_pos_min(llama_context * ctx, llama_seq_id seq_id) {
-    const auto * kv = ctx->get_kv_self();
+    auto * kv = llama_get_memory(ctx);
     if (!kv) {
         return -1;
     }
 
-    return kv->seq_pos_min(seq_id);
+    return llama_memory_seq_pos_min(kv, seq_id);
 }
 
+// deprecated
 llama_pos llama_kv_self_seq_pos_max(llama_context * ctx, llama_seq_id seq_id) {
-    const auto * kv = ctx->get_kv_self();
+    auto * kv = llama_get_memory(ctx);
     if (!kv) {
         return -1;
     }
 
-    return kv->seq_pos_max(seq_id);
+    return llama_memory_seq_pos_max(kv, seq_id);
 }
 
 // deprecated
 void llama_kv_self_defrag(llama_context * ctx) {
-    auto * kv = ctx->get_kv_self();
-    if (!kv) {
-        return;
-    }
-
     // force defrag
-    kv->defrag_sched(-1.0f);
+    ctx->kv_self_defrag_sched();
 }
 
+// deprecated
 bool llama_kv_self_can_shift(const llama_context * ctx) {
-    const auto * kv = ctx->get_kv_self();
+    auto * kv = llama_get_memory(ctx);
     if (!kv) {
         return false;
     }
 
-    return kv->get_can_shift();
+    return llama_memory_can_shift(kv);
 }
 
 // llama state API
diff --git a/src/llama-context.h b/src/llama-context.h
index 3b880286b..9ce05715a 100644
--- a/src/llama-context.h
+++ b/src/llama-context.h
@@ -1,7 +1,6 @@
 #pragma once
 
 #include "llama.h"
-#include "llama-batch.h"
 #include "llama-cparams.h"
 #include "llama-graph.h"
 #include "llama-adapter.h"
@@ -13,13 +12,13 @@
 #include <vector>
 
 struct llama_model;
-struct llama_kv_cache;
+class llama_batch_allocr;
 
 class llama_io_read_i;
 class llama_io_write_i;
 
-class llama_memory_i;
-class llama_memory_state_i;
+struct llama_memory_i;
+struct llama_memory_context_i;
 
 struct llama_context {
     // init scheduler and compute buffers, reserve worst-case graphs
@@ -47,12 +46,12 @@ struct llama_context {
     uint32_t n_threads()       const;
     uint32_t n_threads_batch() const;
 
-          llama_kv_cache * get_kv_self();
-    const llama_kv_cache * get_kv_self() const;
+    llama_memory_t get_memory() const;
 
     // return true of the KV cache was updated
     // TODO: remove
-    bool kv_self_update();
+    bool kv_self_update(bool optimize);
+    void kv_self_defrag_sched();
 
     enum llama_pooling_type pooling_type() const;
 
@@ -94,17 +93,17 @@ struct llama_context {
                 int32_t   il_end);
 
     // process a single ubatch with a specific graph type
-    // if memory_state is provided, it will be applied first to the context's memory
+    // if memory_context is provided, it will be applied first to the context's memory
     // ret contains the status of the graph computation
     // returns nullptr only if ret != GGML_STATUS_SUCCESS
     llm_graph_result_ptr process_ubatch(
-              const llama_ubatch & ubatch,
-                  llm_graph_type   gtype,
-            llama_memory_state_i * mstate,
-                     ggml_status & ret);
+                const llama_ubatch & ubatch,
+                    llm_graph_type   gtype,
+            llama_memory_context_i * mctx,
+                       ggml_status & ret);
 
-    int encode(llama_batch & inp_batch);
-    int decode(llama_batch & inp_batch);
+    int encode(const llama_batch & batch_inp);
+    int decode(const llama_batch & batch_inp);
 
     //
     // state save/load
@@ -182,7 +181,7 @@ private:
 
     // 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);
+    uint32_t output_reserve(int32_t n_outputs);
 
     //
     // graph
@@ -198,15 +197,15 @@ public:
     ggml_status graph_compute(ggml_cgraph * gf, bool batched);
 
     // reserve a graph with a dummy ubatch of the specified size
-    ggml_cgraph * graph_reserve(uint32_t n_tokens, uint32_t n_seqs, uint32_t n_outputs, const llama_memory_state_i * mstate);
+    ggml_cgraph * graph_reserve(uint32_t n_tokens, uint32_t n_seqs, uint32_t n_outputs, const llama_memory_context_i * mctx);
 
 private:
     llm_graph_result_ptr graph_build(
-                    ggml_context * ctx,
-                     ggml_cgraph * gf,
-              const llama_ubatch & ubatch,
-                  llm_graph_type   gtype,
-      const llama_memory_state_i * mstate);
+                      ggml_context * ctx,
+                       ggml_cgraph * gf,
+                const llama_ubatch & ubatch,
+                    llm_graph_type   gtype,
+      const llama_memory_context_i * mctx);
 
     llm_graph_cb graph_get_cb() const;
 
@@ -231,6 +230,9 @@ private:
 
     std::unique_ptr<llama_memory_i> memory;
 
+    // TODO: temporary, until the llama_kv_self_defrag() API is removed
+    bool memory_force_optimize = false;
+
     // decode output (2-dimensional array: [n_outputs][n_vocab])
     size_t  logits_size = 0; // capacity (of floats) for logits
     float * logits      = nullptr;
@@ -244,8 +246,10 @@ private:
     // populated only when pooling_type != LLAMA_POOLING_TYPE_NONE
     std::map<llama_seq_id, std::vector<float>> embd_seq;
 
-    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
+    // reuse the batch_allocr to avoid unnecessary memory allocations
+    std::unique_ptr<llama_batch_allocr> balloc;
+
+    uint32_t n_outputs = 0; // number of actually-used outputs in the current ubatch or last logical batch
 
     std::vector<int32_t> output_ids; // map batch token positions to ids of the logits and embd buffers
 
diff --git a/src/llama-cparams.cpp b/src/llama-cparams.cpp
index f7b36590f..a3e7a37ee 100644
--- a/src/llama-cparams.cpp
+++ b/src/llama-cparams.cpp
@@ -1,5 +1,5 @@
 #include "llama-cparams.h"
 
 size_t llama_max_parallel_sequences(void) {
-    return LLAMA_MAX_PARALLEL_SEQUENCES;
+    return LLAMA_MAX_SEQ;
 }
diff --git a/src/llama-cparams.h b/src/llama-cparams.h
index 2871031ef..118615d5b 100644
--- a/src/llama-cparams.h
+++ b/src/llama-cparams.h
@@ -4,7 +4,7 @@
 
 #include <cstdint>
 
-#define LLAMA_MAX_PARALLEL_SEQUENCES 64
+#define LLAMA_MAX_SEQ 64
 
 struct llama_cparams {
     uint32_t n_ctx;           // context size used during inference
diff --git a/src/llama-graph.cpp b/src/llama-graph.cpp
index 727e119e3..7f0e8c67f 100644
--- a/src/llama-graph.cpp
+++ b/src/llama-graph.cpp
@@ -6,7 +6,8 @@
 
 #include "llama-kv-cache-unified.h"
 #include "llama-kv-cache-unified-iswa.h"
-#include "llama-kv-cache-recurrent.h"
+#include "llama-memory-hybrid.h"
+#include "llama-memory-recurrent.h"
 
 #include <cassert>
 #include <cmath>
@@ -86,41 +87,33 @@ void llm_graph_input_pos_bucket::set_input(const llama_ubatch * ubatch) {
 
 void llm_graph_input_pos_bucket_kv::set_input(const llama_ubatch * ubatch) {
     if (pos_bucket) {
-        kv_state->set_input_pos_bucket(pos_bucket, ubatch);
+        mctx->set_input_pos_bucket(pos_bucket, ubatch);
     }
 }
 
 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");
+    GGML_ASSERT(out_ids);
 
-        if (!out_ids) {
-            LLAMA_LOG_WARN("%s: 'out_ids' is not created\n", __func__);
-        } else {
-            const int64_t n_tokens = ubatch->n_tokens;
+    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;
+    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);
-            }
+    if (n_outputs == n_tokens) {
+        for (int i = 0; i < n_tokens; ++i) {
+            data[i] = i;
+        }
+
+        return;
+    }
+
+    GGML_ASSERT(ubatch->output);
+
+    int n_outputs = 0;
+
+    for (int i = 0; i < n_tokens; ++i) {
+        if (ubatch->output[i]) {
+            data[n_outputs++] = i;
         }
     }
 }
@@ -129,139 +122,114 @@ 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;
+        const int64_t n_seqs_unq   = ubatch->n_seqs_unq;
 
         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));
+        memset(mean->data, 0, n_tokens*n_seqs_unq*ggml_element_size(mean));
 
-        std::vector<uint64_t> sum(n_tokens, 0);
+        std::vector<uint64_t> sums(n_seqs_unq, 0);
+        for (int i = 0; i < n_tokens; i += n_seq_tokens) {
+            for (int s = 0; s < ubatch->n_seq_id[i]; ++s) {
+                const llama_seq_id seq_id  = ubatch->seq_id[i][s];
+                const int32_t      seq_idx = ubatch->seq_idx[seq_id];
 
-        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<float> 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);
+                sums[seq_idx] += ubatch->n_seq_tokens;
             }
         }
 
-        for (int s = 0; s < n_seqs; ++s) {
-            const llama_seq_id seq_id = ubatch->seq_id[s][0];
+        std::vector<float> div(n_seqs_unq, 0.0f);
+        for (int s = 0; s < n_seqs_unq; ++s) {
+            const uint64_t sum = sums[s];
+            if (sum > 0) {
+                div[s] = 1.0f/float(sum);
+            }
+        }
 
-            for (int i = 0; i < n_seq_tokens; ++i) {
-                data[seq_id*n_tokens + s*n_seq_tokens + i] = div[seq_id];
+        for (int i = 0; i < n_tokens; i += n_seq_tokens) {
+            for (int s = 0; s < ubatch->n_seq_id[i]; ++s) {
+                const llama_seq_id seq_id  = ubatch->seq_id[i][s];
+                const int32_t      seq_idx = ubatch->seq_idx[seq_id];
+
+                for (int j = 0; j < n_seq_tokens; ++j) {
+                    data[seq_idx*n_tokens + i + j] = div[seq_idx];
+                }
             }
         }
     }
 }
 
 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;
+    const int64_t n_tokens     = ubatch->n_tokens;
+    const int64_t n_seq_tokens = ubatch->n_seq_tokens;
+    const int64_t n_seqs_unq   = ubatch->n_seqs_unq;
 
+    if (cparams.embeddings && (
+            cparams.pooling_type == LLAMA_POOLING_TYPE_CLS ||
+            cparams.pooling_type == LLAMA_POOLING_TYPE_RANK
+        )) {
         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));
+        memset(cls->data, 0, n_seqs_unq*ggml_element_size(cls));
 
-        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_tokens; i += n_seq_tokens) {
+            for (int s = 0; s < ubatch->n_seq_id[i]; ++s) {
+                const llama_seq_id seq_id  = ubatch->seq_id[i][s];
+                const int32_t      seq_idx = ubatch->seq_idx[seq_id];
 
-            // 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;
-                }
+                data[seq_idx] = 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));
+        memset(cls->data, 0, n_seqs_unq*ggml_element_size(cls));
 
-        std::vector<int> last_pos(n_tokens, -1);
-        std::vector<int> last_row(n_tokens, -1);
+        std::vector<int> last_pos(n_seqs_unq, -1);
+        std::vector<int> last_row(n_seqs_unq, -1);
 
-        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_tokens; ++i) {
+            const llama_pos pos = ubatch->pos[i];
 
-            // 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 s = 0; s < ubatch->n_seq_id[i]; ++s) {
+                const llama_seq_id seq_id  = ubatch->seq_id[i][s];
+                const int32_t      seq_idx = ubatch->seq_idx[seq_id];
 
-            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;
+                if (pos >= last_pos[seq_idx]) {
+                    last_pos[seq_idx] = pos;
+                    last_row[seq_idx] = i;
                 }
             }
         }
 
-        for (int i = 0; i < n_tokens; ++i) {
-            if (last_row[i] >= 0) {
-                data[i] = last_row[i];
+        for (int s = 0; s < n_seqs_unq; ++s) {
+            if (last_row[s] >= 0) {
+                data[s] = last_row[s];
             }
         }
     }
 }
 
-void llm_graph_input_s_copy::set_input(const llama_ubatch * ubatch) {
+void llm_graph_input_rs::set_input(const llama_ubatch * ubatch) {
     GGML_UNUSED(ubatch);
 
-    const int64_t n_kv = kv_state->get_n_kv();
+    const int64_t n_rs = mctx->get_n_rs();
 
     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) {
-            data[i] = kv_state->s_copy(i);
-        }
-    }
-}
-
-void llm_graph_input_s_mask::set_input(const llama_ubatch * ubatch) {
-    GGML_UNUSED(ubatch);
-
-    const int64_t n_kv = kv_state->get_n_kv();
-
-    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) {
-            data[i] = kv_state->s_mask(i);
+        for (uint32_t i = 0; i < n_rs; ++i) {
+            data[i] = mctx->s_copy(i);
         }
     }
 }
@@ -277,141 +245,122 @@ void llm_graph_input_cross_embd::set_input(const llama_ubatch * ubatch) {
 }
 
 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;
+    const int64_t n_kv     = ubatch->n_tokens;
+    const int64_t n_tokens = ubatch->n_tokens;
 
-            GGML_ASSERT(ggml_backend_buffer_is_host(kq_mask->buffer));
-            float * data = (float *) kq_mask->data;
+    GGML_ASSERT(kq_mask);
+    GGML_ASSERT(ggml_backend_buffer_is_host(kq_mask->buffer));
 
-            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];
+    float * data = (float *) kq_mask->data;
 
-                    for (int j = 0; j < n_seq_tokens; ++j) {
-                        const int32_t tj = s1*n_seq_tokens + j;
+    for (int h = 0; h < 1; ++h) {
+        for (int i1 = 0; i1 < n_tokens; ++i1) {
+            const llama_seq_id s1 = ubatch->seq_id[i1][0];
 
-                        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 i0 = 0; i0 < n_tokens; ++i0) {
+                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;
-                                    }
-                                }
+                for (int s = 0; s < ubatch->n_seq_id[i0]; ++s) {
+                    const llama_seq_id s0 = ubatch->seq_id[i0][0];
 
-                                data[h*(n_kv*n_tokens) + tj*n_kv + ti] = f;
-                            }
+                    // TODO: reimplement this like in llama_kv_cache_unified
+                    if (s0 == s1 && (!cparams.causal_attn || ubatch->pos[i0] <= ubatch->pos[i1])) {
+                        if (hparams.use_alibi) {
+                            f = -std::abs(ubatch->pos[i0] - ubatch->pos[i1]);
+                        } else {
+                            f = 0.0f;
                         }
+                        break;
                     }
                 }
-            }
-        } 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;
-                        }
-                    }
-                }
+                data[h*(n_kv*n_tokens) + i1*n_kv + i0] = f;
             }
         }
     }
 }
 
 void llm_graph_input_attn_kv_unified::set_input(const llama_ubatch * ubatch) {
-    if (self_kq_mask) {
-        kv_state->set_input_kq_mask(self_kq_mask, ubatch, cparams.causal_attn);
-    }
+    mctx->set_input_k_idxs(self_k_idxs, ubatch);
+    mctx->set_input_v_idxs(self_v_idxs, ubatch);
+
+    mctx->set_input_kq_mask(self_kq_mask, ubatch, cparams.causal_attn);
 }
 
 void llm_graph_input_attn_kv_unified_iswa::set_input(const llama_ubatch * ubatch) {
-    if (self_kq_mask) {
-        kv_state->get_base()->set_input_kq_mask(self_kq_mask, ubatch, cparams.causal_attn);
-    }
+    mctx->get_base()->set_input_k_idxs(self_k_idxs, ubatch);
+    mctx->get_base()->set_input_v_idxs(self_v_idxs, ubatch);
 
-    if (self_kq_mask_swa) {
-        kv_state->get_swa()->set_input_kq_mask(self_kq_mask_swa, ubatch, cparams.causal_attn);
-    }
+    mctx->get_base()->set_input_kq_mask(self_kq_mask, ubatch, cparams.causal_attn);
+
+    mctx->get_swa()->set_input_k_idxs(self_k_idxs_swa, ubatch);
+    mctx->get_swa()->set_input_v_idxs(self_v_idxs_swa, ubatch);
+
+    mctx->get_swa()->set_input_kq_mask(self_kq_mask_swa, ubatch, cparams.causal_attn);
 }
 
 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(cross_kq_mask);
 
-        GGML_ASSERT(ggml_backend_buffer_is_host(cross_kq_mask->buffer));
-        GGML_ASSERT(!ubatch->equal_seqs); // TODO: use ubatch->n_seqs instead of failing
+    const int64_t n_enc    = cross_kq_mask->ne[0];
+    const int64_t n_tokens = ubatch->n_tokens;
 
-        float * data = (float *) cross_kq_mask->data;
+    GGML_ASSERT(ggml_backend_buffer_is_host(cross_kq_mask->buffer));
+    GGML_ASSERT(!ubatch->equal_seqs); // TODO: use ubatch->n_seqs instead of failing
 
-        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;
-                        }
+    float * data = (float *) cross_kq_mask->data;
+
+    for (int h = 0; h < 1; ++h) {
+        for (int i = 0; i < n_tokens; ++i) {
+            for (int j = 0; j < n_enc; ++j) {
+                float f = -INFINITY;
+
+                for (int s = 0; s < ubatch->n_seq_id[i]; ++s) {
+                    const llama_seq_id seq_id = ubatch->seq_id[i][s];
+
+                    if (cross->seq_ids_enc[j].find(seq_id) != cross->seq_ids_enc[j].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;
-                }
+                data[h*(n_enc*n_tokens) + i*n_enc + j] = 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;
             }
         }
     }
 }
 
+void llm_graph_input_mem_hybrid::set_input(const llama_ubatch * ubatch) {
+    mctx->get_attn()->set_input_k_idxs(self_k_idxs, ubatch);
+    mctx->get_attn()->set_input_v_idxs(self_v_idxs, ubatch);
+
+    mctx->get_attn()->set_input_kq_mask(self_kq_mask, ubatch, cparams.causal_attn);
+
+    const int64_t n_rs = mctx->get_recr()->get_n_rs();
+
+    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_rs; ++i) {
+            data[i] = mctx->get_recr()->s_copy(i);
+        }
+    }
+}
+
+void llm_graph_input_one::set_input(const llama_ubatch * ubatch) {
+    GGML_UNUSED(ubatch);
+    GGML_ASSERT(one && ggml_nelements(one) == 1);
+    float f_one = 1.0f;
+    ggml_backend_tensor_set(one, &f_one, 0, sizeof(float));
+}
+
 //
 // llm_graph_context
 //
@@ -451,16 +400,12 @@ llm_graph_context::llm_graph_context(const llm_graph_params & params) :
     backend_cpu      (params.backend_cpu),
     cvec             (params.cvec),
     loras            (params.loras),
-    mstate           (params.mstate),
+    mctx             (params.mctx),
     cross            (params.cross),
     cb_func          (params.cb),
     res              (std::make_unique<llm_graph_result>()) {
     }
 
-int64_t llm_graph_context::n_pos_per_embd() const {
-    return hparams.rope_type == LLAMA_ROPE_TYPE_MROPE ? 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);
@@ -620,12 +565,20 @@ ggml_tensor * llm_graph_context::build_ffn(
 
     switch (type_op) {
         case LLM_FFN_SILU:
-            {
+            if (gate && type_gate == LLM_FFN_PAR) {
+                cur = ggml_swiglu_split(ctx0, cur, tmp);
+                cb(cur, "ffn_swiglu", il);
+                type_gate = LLM_FFN_SEQ;
+            } else {
                 cur = ggml_silu(ctx0, cur);
                 cb(cur, "ffn_silu", il);
             } break;
         case LLM_FFN_GELU:
-            {
+            if (gate && type_gate == LLM_FFN_PAR) {
+                cur = ggml_geglu_split(ctx0, cur, tmp);
+                cb(cur, "ffn_geglu", il);
+                type_gate = LLM_FFN_SEQ;
+            } else {
                 cur = ggml_gelu(ctx0, cur);
                 cb(cur, "ffn_gelu", il);
                 if (act_scales != NULL) {
@@ -634,7 +587,11 @@ ggml_tensor * llm_graph_context::build_ffn(
                 }
             } break;
         case LLM_FFN_RELU:
-            {
+            if (gate && type_gate == LLM_FFN_PAR) {
+                cur = ggml_reglu_split(ctx0, cur, tmp);
+                cb(cur, "ffn_reglu", il);
+                type_gate = LLM_FFN_SEQ;
+            } else {
                 cur = ggml_relu(ctx0, cur);
                 cb(cur, "ffn_relu", il);
             } break;
@@ -648,16 +605,18 @@ ggml_tensor * llm_graph_context::build_ffn(
             } 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);
+                cur = ggml_swiglu(ctx0, cur);
+                cb(cur, "ffn_swiglu", il);
+            } break;
+        case LLM_FFN_GEGLU:
+            {
+                cur = ggml_geglu(ctx0, cur);
+                cb(cur, "ffn_geglu", il);
+            } break;
+        case LLM_FFN_REGLU:
+            {
+                cur = ggml_reglu(ctx0, cur);
+                cb(cur, "ffn_reglu", il);
             } break;
     }
 
@@ -769,9 +728,8 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
     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]
+        // repeat cur to [n_embd, n_expert_used, n_tokens]
+        ggml_tensor * repeated = ggml_repeat_4d(ctx0, cur, n_embd, n_expert_used, n_tokens, 1);
         cur = ggml_mul(ctx0, repeated, weights);
         cb(cur, "ffn_moe_weighted", il);
     }
@@ -789,12 +747,18 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
 
     switch (type_op) {
         case LLM_FFN_SILU:
-            {
+            if (gate_exps) {
+                cur = ggml_swiglu_split(ctx0, cur, up);
+                cb(cur, "ffn_moe_swiglu", il);
+            } else {
                 cur = ggml_silu(ctx0, cur);
                 cb(cur, "ffn_moe_silu", il);
             } break;
         case LLM_FFN_GELU:
-            {
+            if (gate_exps) {
+                cur = ggml_geglu_split(ctx0, cur, up);
+                cb(cur, "ffn_moe_geglu", il);
+            } else {
                 cur = ggml_gelu(ctx0, cur);
                 cb(cur, "ffn_moe_gelu", il);
             } break;
@@ -802,11 +766,6 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
             GGML_ABORT("fatal error");
     }
 
-    if (gate_exps) {
-        cur = ggml_mul(ctx0, cur, up); // [n_ff, n_expert_used, n_tokens]
-        cb(cur, "ffn_moe_gate_par", il);
-    }
-
     experts = build_lora_mm_id(down_exps, cur, selected_experts); // [n_embd, n_expert_used, n_tokens]
     cb(experts, "ffn_moe_down", il);
 
@@ -891,11 +850,11 @@ ggml_tensor * llm_graph_context::build_inp_embd(ggml_tensor * tok_embd) const {
 }
 
 ggml_tensor * llm_graph_context::build_inp_pos() const {
-    auto inp = std::make_unique<llm_graph_input_pos>(n_pos_per_embd());
+    auto inp = std::make_unique<llm_graph_input_pos>(hparams.n_pos_per_embd());
 
     auto & cur = inp->pos;
 
-    cur = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens*n_pos_per_embd());
+    cur = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, (int64_t)n_tokens*hparams.n_pos_per_embd());
     ggml_set_input(cur);
 
     res->add_input(std::move(inp));
@@ -918,6 +877,14 @@ ggml_tensor * llm_graph_context::build_inp_attn_scale() const {
 }
 
 ggml_tensor * llm_graph_context::build_inp_out_ids() const {
+    // note: when all tokens are output, we could skip this optimization to spare the ggml_get_rows() calls,
+    //       but this would make the graph topology depend on the number of output tokens, which can interere with
+    //       features that require constant topology such as pipline parallelism
+    //       ref: https://github.com/ggml-org/llama.cpp/pull/14275#issuecomment-2987424471
+    //if (n_outputs < n_tokens) {
+    //    return nullptr;
+    //}
+
     auto inp = std::make_unique<llm_graph_input_out_ids>(hparams, cparams, n_outputs);
 
     auto & cur = inp->out_ids;
@@ -935,7 +902,7 @@ ggml_tensor * llm_graph_context::build_inp_mean() const {
 
     auto & cur = inp->mean;
 
-    cur = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_tokens, n_tokens);
+    cur = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_tokens, ubatch.n_seqs_unq);
     ggml_set_input(cur);
 
     res->add_input(std::move(inp));
@@ -948,41 +915,7 @@ ggml_tensor * llm_graph_context::build_inp_cls() const {
 
     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 auto * kv_state = static_cast<const llama_kv_cache_recurrent_state *>(mstate);
-
-    auto inp = std::make_unique<llm_graph_input_s_copy>(kv_state);
-
-    const auto n_kv = kv_state->get_n_kv();
-
-    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 auto * kv_state = static_cast<const llama_kv_cache_recurrent_state *>(mstate);
-
-    auto inp = std::make_unique<llm_graph_input_s_mask>(kv_state);
-
-    const auto n_kv = kv_state->get_n_kv();
-
-    auto & cur = inp->s_mask;
-
-    cur = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, 1, n_kv);
+    cur = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, ubatch.n_seqs_unq);
     ggml_set_input(cur);
 
     res->add_input(std::move(inp));
@@ -1028,11 +961,11 @@ ggml_tensor * llm_graph_context::build_inp_pos_bucket_enc() const {
 }
 
 ggml_tensor * llm_graph_context::build_inp_pos_bucket_dec() const {
-    const auto * kv_state = static_cast<const llama_kv_cache_unified_state *>(mstate);
+    const auto * mctx_cur = static_cast<const llama_kv_cache_unified_context *>(mctx);
 
-    auto inp = std::make_unique<llm_graph_input_pos_bucket_kv>(hparams, kv_state);
+    auto inp = std::make_unique<llm_graph_input_pos_bucket_kv>(hparams, mctx_cur);
 
-    const auto n_kv = kv_state->get_n_kv();
+    const auto n_kv = mctx_cur->get_n_kv();
 
     auto & cur = inp->pos_bucket;
 
@@ -1059,6 +992,35 @@ ggml_tensor * llm_graph_context::build_pos_bias(ggml_tensor * pos_bucket, ggml_t
     return pos_bias;
 }
 
+llm_graph_input_mem_hybrid * llm_graph_context::build_inp_mem_hybrid() const {
+    const auto * mctx_cur = static_cast<const llama_memory_hybrid_context *>(mctx);
+
+    auto inp = std::make_unique<llm_graph_input_mem_hybrid>(hparams, cparams, mctx_cur);
+
+    {
+        GGML_ASSERT(hparams.swa_type == LLAMA_SWA_TYPE_NONE && "Hybrid recurrent is not supported with SWA attention layers");
+
+        const auto n_kv = inp->mctx->get_attn()->get_n_kv();
+
+        inp->self_k_idxs = mctx_cur->get_attn()->build_input_k_idxs(ctx0, ubatch);
+        inp->self_v_idxs = mctx_cur->get_attn()->build_input_v_idxs(ctx0, ubatch);
+
+        inp->self_kq_mask = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD), 1, 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;
+    }
+
+    {
+        const auto n_rs = mctx_cur->get_recr()->get_n_rs();
+
+        inp->s_copy = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_rs);
+        ggml_set_input(inp->s_copy);
+    }
+
+    return (llm_graph_input_mem_hybrid *) res->add_input(std::move(inp));
+}
+
 ggml_tensor * llm_graph_context::build_attn_mha(
          ggml_cgraph * gf,
          ggml_tensor * q,
@@ -1180,8 +1142,7 @@ llm_graph_input_attn_no_cache * llm_graph_context::build_attn_inp_no_cache() con
     auto inp = std::make_unique<llm_graph_input_attn_no_cache>(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);
+    inp->kq_mask = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_tokens, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD), 1, 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;
@@ -1234,17 +1195,19 @@ ggml_tensor * llm_graph_context::build_attn(
 }
 
 llm_graph_input_attn_kv_unified * llm_graph_context::build_attn_inp_kv_unified() const {
-    const auto * kv_state = static_cast<const llama_kv_cache_unified_state *>(mstate);
+    const auto * mctx_cur = static_cast<const llama_kv_cache_unified_context *>(mctx);
 
-    auto inp = std::make_unique<llm_graph_input_attn_kv_unified>(hparams, cparams, kv_state);
+    auto inp = std::make_unique<llm_graph_input_attn_kv_unified>(hparams, cparams, mctx_cur);
 
     {
         GGML_ASSERT(hparams.swa_type == LLAMA_SWA_TYPE_NONE && "Use llama_kv_cache_unified_iswa for SWA");
 
-        const auto n_kv = kv_state->get_n_kv();
+        const auto n_kv = mctx_cur->get_n_kv();
 
-        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);
+        inp->self_k_idxs = mctx_cur->build_input_k_idxs(ctx0, ubatch);
+        inp->self_v_idxs = mctx_cur->build_input_v_idxs(ctx0, ubatch);
+
+        inp->self_kq_mask = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD), 1, 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;
@@ -1271,19 +1234,22 @@ ggml_tensor * llm_graph_context::build_attn(
     ggml_build_forward_expand(gf, k_cur);
     ggml_build_forward_expand(gf, v_cur);
 
-    const auto * kv_state = static_cast<const llama_kv_cache_unified_state *>(mstate);
+    const auto * mctx_cur = static_cast<const llama_kv_cache_unified_context *>(mctx);
 
     // store to KV cache
     {
-        ggml_build_forward_expand(gf, kv_state->cpy_k(ctx0, k_cur, il));
-        ggml_build_forward_expand(gf, kv_state->cpy_v(ctx0, v_cur, il));
+        const auto & k_idxs = inp->get_k_idxs();
+        const auto & v_idxs = inp->get_v_idxs();
+
+        ggml_build_forward_expand(gf, mctx_cur->cpy_k(ctx0, k_cur, k_idxs, il));
+        ggml_build_forward_expand(gf, mctx_cur->cpy_v(ctx0, v_cur, v_idxs, il));
     }
 
     const auto & kq_mask = inp->get_kq_mask();
 
     ggml_tensor * q = q_cur;
-    ggml_tensor * k = kv_state->get_k(ctx0, il);
-    ggml_tensor * v = kv_state->get_v(ctx0, il);
+    ggml_tensor * k = mctx_cur->get_k(ctx0, il);
+    ggml_tensor * v = mctx_cur->get_v(ctx0, il);
 
     ggml_tensor * cur = build_attn_mha(gf, q, k, v, kq_b, kq_mask, v_mla, kq_scale);
     cb(cur, "kqv_out", il);
@@ -1303,36 +1269,6 @@ ggml_tensor * llm_graph_context::build_attn(
     return cur;
 }
 
-llm_graph_input_attn_kv_unified_iswa * llm_graph_context::build_attn_inp_kv_unified_iswa() const {
-    const auto * kv_state = static_cast<const llama_kv_cache_unified_iswa_state *>(mstate);
-
-    auto inp = std::make_unique<llm_graph_input_attn_kv_unified_iswa>(hparams, cparams, kv_state);
-
-    {
-        const auto n_kv = kv_state->get_base()->get_n_kv();
-
-        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;
-    }
-
-    {
-        GGML_ASSERT(hparams.swa_type != LLAMA_SWA_TYPE_NONE && "Use llama_kv_cache_unified for non-SWA");
-
-        const auto n_kv = kv_state->get_swa()->get_n_kv();
-
-        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_iswa *) res->add_input(std::move(inp));
-}
-
 ggml_tensor * llm_graph_context::build_attn(
         llm_graph_input_attn_kv_unified_iswa * inp,
         ggml_cgraph * gf,
@@ -1348,26 +1284,39 @@ ggml_tensor * llm_graph_context::build_attn(
     // 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 * kv_state_iswa = static_cast<const llama_kv_cache_unified_iswa_state *>(mstate);
+    if (k_cur) {
+        ggml_build_forward_expand(gf, k_cur);
+    }
+
+    if (v_cur) {
+        ggml_build_forward_expand(gf, v_cur);
+    }
+
+    const auto * mctx_iswa = static_cast<const llama_kv_cache_unified_iswa_context *>(mctx);
 
     const bool is_swa = hparams.is_swa(il);
 
-    const auto * kv_state = is_swa ? kv_state_iswa->get_swa() : kv_state_iswa->get_base();
+    const auto * mctx_cur = is_swa ? mctx_iswa->get_swa() : mctx_iswa->get_base();
 
-    // store to KV cache
-    {
-        ggml_build_forward_expand(gf, kv_state->cpy_k(ctx0, k_cur, il));
-        ggml_build_forward_expand(gf, kv_state->cpy_v(ctx0, v_cur, il));
+    // optionally store to KV cache
+    if (k_cur) {
+        const auto & k_idxs = is_swa ? inp->get_k_idxs_swa() : inp->get_k_idxs();
+
+        ggml_build_forward_expand(gf, mctx_cur->cpy_k(ctx0, k_cur, k_idxs, il));
+    }
+
+    if (v_cur) {
+        const auto & v_idxs = is_swa ? inp->get_v_idxs_swa() : inp->get_v_idxs();
+
+        ggml_build_forward_expand(gf, mctx_cur->cpy_v(ctx0, v_cur, v_idxs, il));
     }
 
     const auto & kq_mask = is_swa ? inp->get_kq_mask_swa() : inp->get_kq_mask();
 
     ggml_tensor * q = q_cur;
-    ggml_tensor * k = kv_state->get_k(ctx0, il);
-    ggml_tensor * v = kv_state->get_v(ctx0, il);
+    ggml_tensor * k = mctx_cur->get_k(ctx0, il);
+    ggml_tensor * v = mctx_cur->get_v(ctx0, il);
 
     ggml_tensor * cur = build_attn_mha(gf, q, k, v, kq_b, kq_mask, v_mla, kq_scale);
     cb(cur, "kqv_out", il);
@@ -1392,7 +1341,7 @@ llm_graph_input_attn_cross * llm_graph_context::build_attn_inp_cross() const {
 
     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));
+    inp->cross_kq_mask = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_enc, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD), 1, 1);
     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;
@@ -1442,56 +1391,181 @@ ggml_tensor * llm_graph_context::build_attn(
     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 auto * kv_state = static_cast<const llama_kv_cache_recurrent_state *>(mstate);
+ggml_tensor * llm_graph_context::build_attn(
+        llm_graph_input_mem_hybrid * 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,
+        ggml_tensor * v_mla,
+            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 n_kv    = kv_state->get_n_kv();
-    const auto kv_head = kv_state->get_head();
+    const auto * mctx_cur = static_cast<const llama_memory_hybrid_context *>(mctx)->get_attn();
 
-    ggml_tensor * states = ggml_reshape_2d(ctx0, s, n_state, kv_state->get_size());
+    // store to KV cache
+    {
+        const auto & k_idxs = inp->get_k_idxs();
+        const auto & v_idxs = inp->get_v_idxs();
+
+        ggml_build_forward_expand(gf, mctx_cur->cpy_k(ctx0, k_cur, k_idxs, il));
+        ggml_build_forward_expand(gf, mctx_cur->cpy_v(ctx0, v_cur, v_idxs, il));
+    }
+
+    const auto & kq_mask = inp->get_kq_mask();
+
+    ggml_tensor * q = q_cur;
+    ggml_tensor * k = mctx_cur->get_k(ctx0, il);
+    ggml_tensor * v = mctx_cur->get_v(ctx0, il);
+
+    ggml_tensor * cur = build_attn_mha(gf, q, k, v, kq_b, kq_mask, v_mla, kq_scale);
+    cb(cur, "kqv_out", il);
+
+    if (wo) {
+        cur = build_lora_mm(wo, cur);
+        if (arch == LLM_ARCH_GLM4) {
+            // GLM4 seems to have numerical issues with half-precision accumulators
+            ggml_mul_mat_set_prec(cur, GGML_PREC_F32);
+        }
+    }
+
+    if (wo_b) {
+        cur = ggml_add(ctx0, cur, wo_b);
+    }
+
+    return cur;
+}
+
+llm_graph_input_attn_kv_unified_iswa * llm_graph_context::build_attn_inp_kv_unified_iswa() const {
+    const auto * mctx_cur = static_cast<const llama_kv_cache_unified_iswa_context *>(mctx);
+
+    auto inp = std::make_unique<llm_graph_input_attn_kv_unified_iswa>(hparams, cparams, mctx_cur);
+
+    {
+        const auto n_kv = mctx_cur->get_base()->get_n_kv();
+
+        inp->self_k_idxs = mctx_cur->get_base()->build_input_k_idxs(ctx0, ubatch);
+        inp->self_v_idxs = mctx_cur->get_base()->build_input_v_idxs(ctx0, ubatch);
+
+        inp->self_kq_mask = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD), 1, 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;
+    }
+
+    {
+        GGML_ASSERT(hparams.swa_type != LLAMA_SWA_TYPE_NONE && "Use llama_kv_cache_unified for non-SWA");
+
+        const auto n_kv = mctx_cur->get_swa()->get_n_kv();
+
+        inp->self_k_idxs_swa = mctx_cur->get_swa()->build_input_k_idxs(ctx0, ubatch);
+        inp->self_v_idxs_swa = mctx_cur->get_swa()->build_input_v_idxs(ctx0, ubatch);
+
+        inp->self_kq_mask_swa = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD), 1, 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_iswa *) res->add_input(std::move(inp));
+}
+
+ggml_tensor * llm_graph_context::build_rs(
+        ggml_cgraph * gf,
+        ggml_tensor * s,
+        ggml_tensor * state_copy,
+            int32_t   state_size,
+            int32_t   n_seqs,
+           uint32_t   n_kv,
+           uint32_t   kv_head,
+           uint32_t   kv_size,
+            int32_t   rs_zero,
+        const llm_graph_get_rows_fn & get_state_rows) const {
+
+    ggml_tensor * states = ggml_reshape_2d(ctx0, s, state_size, kv_size);
+
+    // Clear a single state which will then be copied to the other cleared states.
+    // Note that this is a no-op when the view is zero-sized.
+    ggml_tensor * state_zero = ggml_view_1d(ctx0, states, state_size*(rs_zero >= 0), rs_zero*states->nb[1]*(rs_zero >= 0));
+    ggml_build_forward_expand(gf, ggml_scale_inplace(ctx0, state_zero, 0));
 
     // 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);
+    // {state_size, kv_size} -> {state_size, n_seqs}
+    ggml_tensor * output_states = get_state_rows(ctx0, states, ggml_view_1d(ctx0, state_copy, n_seqs, 0));
+    ggml_build_forward_expand(gf, output_states);
 
-    // 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)
+    // copy extra states which won't be changed further (between n_seqs and n_kv)
+    ggml_tensor * states_extra = ggml_get_rows(ctx0, states, ggml_view_1d(ctx0, state_copy, n_kv - n_seqs, n_seqs*state_copy->nb[0]));
     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))));
+            states_extra,
+            ggml_view_1d(ctx0, s, state_size*(n_kv - n_seqs), (kv_head + n_seqs)*state_size*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);
+    return output_states;
+}
+
+llm_graph_input_rs * llm_graph_context::build_rs_inp() const {
+    const auto * mctx_cur = static_cast<const llama_memory_recurrent_context *>(mctx);
+
+    auto inp = std::make_unique<llm_graph_input_rs>(mctx_cur);
+
+    const auto n_rs = mctx_cur->get_n_rs();
+
+    inp->s_copy = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_rs);
+    ggml_set_input(inp->s_copy);
+
+    return (llm_graph_input_rs *) res->add_input(std::move(inp));
+}
+
+ggml_tensor * llm_graph_context::build_rs(
+        llm_graph_input_rs * inp,
+        ggml_cgraph * gf,
+        ggml_tensor * s,
+            int32_t   state_size,
+            int32_t   n_seqs,
+        const llm_graph_get_rows_fn & get_state_rows) const {
+    const auto * kv_state = static_cast<const llama_memory_recurrent_context *>(mctx);
+
+    return build_rs(gf, s, inp->s_copy, state_size, n_seqs, kv_state->get_n_rs(), kv_state->get_head(), kv_state->get_size(), kv_state->get_rs_z(), get_state_rows);
+}
+
+ggml_tensor * llm_graph_context::build_rs(
+        llm_graph_input_mem_hybrid * inp,
+        ggml_cgraph * gf,
+        ggml_tensor * s,
+            int32_t   state_size,
+            int32_t   n_seqs,
+        const llm_graph_get_rows_fn & get_state_rows) const {
+    const auto * kv_state = static_cast<const llama_memory_hybrid_context *>(mctx)->get_recr();
+
+    return build_rs(gf, s, inp->s_copy, state_size, n_seqs, kv_state->get_n_rs(), kv_state->get_head(), kv_state->get_size(), kv_state->get_rs_z(), get_state_rows);
 }
 
 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,
+    llm_graph_input_rs * inp,
+           ggml_cgraph * gf,
+    const llama_ubatch & ubatch,
                  int   il) const {
-    const auto * kv_state = static_cast<const llama_kv_cache_recurrent_state *>(mstate);
+    const auto * mctx_cur = static_cast<const llama_memory_recurrent_context *>(mctx);
 
     const auto token_shift_count = hparams.token_shift_count;
 
     const int64_t n_seqs  = ubatch.n_seqs;
 
-    ggml_tensor * token_shift_all = kv_state->get_k_l(il);
+    ggml_tensor * token_shift_all = mctx_cur->get_r_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);
+    ggml_tensor * token_shift = build_rs(
+            inp, gf, token_shift_all,
+            hparams.n_embd_r(), n_seqs);
 
     token_shift = ggml_reshape_3d(ctx0, token_shift, hparams.n_embd, token_shift_count, n_seqs);
 
@@ -1502,19 +1576,19 @@ ggml_tensor * llm_graph_context::build_rwkv_token_shift_store(
          ggml_tensor * token_shift,
   const llama_ubatch & ubatch,
                  int   il) const {
-    const auto * kv_state = static_cast<const llama_kv_cache_recurrent_state *>(mstate);
+    const auto * mctx_cur = static_cast<const llama_memory_recurrent_context *>(mctx);
 
     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_state->get_head();
+    const auto kv_head = mctx_cur->get_head();
 
     return ggml_cpy(
         ctx0,
         ggml_view_1d(ctx0, token_shift, n_embd * n_seqs * token_shift_count, 0),
-        ggml_view_1d(ctx0, kv_state->get_k_l(il), hparams.n_embd_k_s()*n_seqs, hparams.n_embd_k_s()*kv_head*ggml_element_size(kv_state->get_k_l(il)))
+        ggml_view_1d(ctx0, mctx_cur->get_r_l(il), hparams.n_embd_r()*n_seqs, hparams.n_embd_r()*kv_head*ggml_element_size(mctx_cur->get_r_l(il)))
     );
 }
 
@@ -1565,23 +1639,30 @@ void llm_graph_context::build_pooling(
                 ggml_tensor * inp_cls = build_inp_cls();
                 inp = ggml_get_rows(ctx0, inp, inp_cls);
 
-                if (cls != nullptr && cls_b != nullptr) {
+                if (cls) {
                     // classification head
                     // https://github.com/huggingface/transformers/blob/5af7d41e49bbfc8319f462eb45253dcb3863dfb7/src/transformers/models/roberta/modeling_roberta.py#L1566
-                    cur = ggml_add(ctx0, ggml_mul_mat(ctx0, cls, inp), cls_b);
+                    cur = ggml_mul_mat(ctx0, cls, inp);
+                    if (cls_b) {
+                        cur = ggml_add(ctx0, cur, 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);
+                        cur = ggml_mul_mat(ctx0, cls_out, cur);
+                        if (cls_out_b) {
+                            cur = ggml_add(ctx0, cur, cls_out_b);
+                        }
                     }
                 } else if (cls_out) {
                     // Single layer classification head (direct projection)
                     // https://github.com/huggingface/transformers/blob/f4fc42216cd56ab6b68270bf80d811614d8d59e4/src/transformers/models/bert/modeling_bert.py#L1476
-                    GGML_ASSERT(cls_out_b != nullptr);
-                    cur = ggml_add(ctx0, ggml_mul_mat(ctx0, cls_out, inp), cls_out_b);
+                    cur = ggml_mul_mat(ctx0, cls_out, inp);
+                    if (cls_out_b) {
+                        cur = ggml_add(ctx0, cur, cls_out_b);
+                    }
                 } else {
                     GGML_ABORT("RANK pooling requires either cls+cls_b or cls_out+cls_out_b");
                 }
diff --git a/src/llama-graph.h b/src/llama-graph.h
index d1c5dd1bf..7bdf65676 100644
--- a/src/llama-graph.h
+++ b/src/llama-graph.h
@@ -17,11 +17,12 @@ struct ggml_tensor;
 struct llama_ubatch;
 struct llama_cparams;
 
-class llama_memory_state_i;
+struct llama_memory_context_i;
 
-class llama_kv_cache_unified_state;
-class llama_kv_cache_unified_iswa_state;
-class llama_kv_cache_recurrent_state;
+class llama_kv_cache_unified_context;
+class llama_kv_cache_unified_iswa_context;
+class llama_memory_recurrent_context;
+class llama_memory_hybrid_context;
 
 // certain models (typically multi-modal) can produce different types of graphs
 enum llm_graph_type {
@@ -36,6 +37,8 @@ enum llm_ffn_op_type {
     LLM_FFN_RELU,
     LLM_FFN_RELU_SQR,
     LLM_FFN_SWIGLU,
+    LLM_FFN_GEGLU,
+    LLM_FFN_REGLU,
 };
 
 enum llm_ffn_gate_type {
@@ -93,14 +96,14 @@ public:
 
 class llm_graph_input_pos : public llm_graph_input_i {
 public:
-    llm_graph_input_pos(int64_t n_pos_per_embd) : n_pos_per_embd(n_pos_per_embd) {}
+    llm_graph_input_pos(uint32_t n_pos_per_embd) : n_pos_per_embd(n_pos_per_embd) {}
     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_embd = 1;
+    const uint32_t n_pos_per_embd = 1;
 };
 
 // temperature tuning, used by llama4
@@ -134,7 +137,7 @@ 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_state * kv_state) : hparams(hparams), kv_state(kv_state) {}
+            const llama_kv_cache_unified_context * mctx) : hparams(hparams), mctx(mctx) {}
     virtual ~llm_graph_input_pos_bucket_kv() = default;
 
     void set_input(const llama_ubatch * ubatch) override;
@@ -142,7 +145,8 @@ public:
     ggml_tensor * pos_bucket = nullptr; // I32 [n_kv, n_batch]
 
     const llama_hparams & hparams;
-    const llama_kv_cache_unified_state * kv_state;
+
+    const llama_kv_cache_unified_context * mctx;
 };
 
 class llm_graph_input_out_ids : public llm_graph_input_i {
@@ -187,28 +191,16 @@ public:
     const llama_cparams & cparams;
 };
 
-class llm_graph_input_s_copy : public llm_graph_input_i {
+class llm_graph_input_rs : public llm_graph_input_i {
 public:
-    llm_graph_input_s_copy(const llama_kv_cache_recurrent_state * kv_state) : kv_state(kv_state) {}
-    virtual ~llm_graph_input_s_copy() = default;
+    llm_graph_input_rs(const llama_memory_recurrent_context * mctx) : mctx(mctx) {}
+    virtual ~llm_graph_input_rs() = default;
 
     void set_input(const llama_ubatch * ubatch) override;
 
     ggml_tensor * s_copy; // I32 [kv_size]
 
-    const llama_kv_cache_recurrent_state * kv_state;
-};
-
-class llm_graph_input_s_mask : public llm_graph_input_i {
-public:
-    llm_graph_input_s_mask(const llama_kv_cache_recurrent_state * kv_state) : kv_state(kv_state) {}
-    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_recurrent_state * kv_state;
+    const llama_memory_recurrent_context * mctx;
 };
 
 class llm_graph_input_cross_embd : public llm_graph_input_i {
@@ -236,8 +228,8 @@ public:
 
     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]
+    ggml_tensor * kq_mask     = nullptr; // F32 [n_tokens, n_batch, 1, 1]
+    ggml_tensor * kq_mask_cnv = nullptr; //     [n_tokens, n_batch, 1, 1]
 
     const llama_hparams & hparams;
     const llama_cparams & cparams;
@@ -248,24 +240,30 @@ public:
     llm_graph_input_attn_kv_unified(
             const llama_hparams & hparams,
             const llama_cparams & cparams,
-            const llama_kv_cache_unified_state * kv_state) :
+            const llama_kv_cache_unified_context * mctx) :
         hparams(hparams),
         cparams(cparams),
-        kv_state(kv_state) {
+        mctx(mctx) {
     }
     ~llm_graph_input_attn_kv_unified() = default;
 
     void set_input(const llama_ubatch * ubatch) override;
 
+    ggml_tensor * get_k_idxs() const { return self_k_idxs; }
+    ggml_tensor * get_v_idxs() const { return self_v_idxs; }
+
     ggml_tensor * get_kq_mask() const { return self_kq_mask_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_k_idxs = nullptr; // I64 [n_batch]
+    ggml_tensor * self_v_idxs = nullptr; // I64 [n_batch]
+
+    ggml_tensor * self_kq_mask     = nullptr; // F32 [n_kv, n_batch, 1, 1]
+    ggml_tensor * self_kq_mask_cnv = nullptr; //     [n_kv, n_batch, 1, 1]
 
     const llama_hparams & hparams;
     const llama_cparams & cparams;
 
-    const llama_kv_cache_unified_state * kv_state;
+    const llama_kv_cache_unified_context * mctx;
 };
 
 class llm_graph_input_attn_kv_unified_iswa : public llm_graph_input_i {
@@ -273,27 +271,37 @@ public:
     llm_graph_input_attn_kv_unified_iswa(
             const llama_hparams & hparams,
             const llama_cparams & cparams,
-            const llama_kv_cache_unified_iswa_state * kv_state) :
+            const llama_kv_cache_unified_iswa_context * mctx) :
         hparams(hparams),
         cparams(cparams),
-        kv_state(kv_state) {
+        mctx(mctx) {
     }
     ~llm_graph_input_attn_kv_unified_iswa() = default;
 
     void set_input(const llama_ubatch * ubatch) override;
 
+    ggml_tensor * get_k_idxs()     const { return self_k_idxs; }
+    ggml_tensor * get_v_idxs()     const { return self_v_idxs; }
+    ggml_tensor * get_k_idxs_swa() const { return self_k_idxs_swa; }
+    ggml_tensor * get_v_idxs_swa() const { return self_v_idxs_swa; }
+
     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]
+    ggml_tensor * self_k_idxs     = nullptr; // I64 [n_batch]
+    ggml_tensor * self_v_idxs     = nullptr; // I64 [n_batch]
+    ggml_tensor * self_k_idxs_swa = nullptr; // I64 [n_batch]
+    ggml_tensor * self_v_idxs_swa = nullptr; // I64 [n_batch]
+
+    ggml_tensor * self_kq_mask         = nullptr; // F32 [n_kv, n_batch, 1, 1]
+    ggml_tensor * self_kq_mask_cnv     = nullptr; //     [n_kv, n_batch, 1, 1]
+    ggml_tensor * self_kq_mask_swa     = nullptr; // F32 [n_kv, n_batch, 1, 1]
+    ggml_tensor * self_kq_mask_swa_cnv = nullptr; //     [n_kv, n_batch, 1, 1]
 
     const llama_hparams & hparams;
     const llama_cparams & cparams;
 
-    const llama_kv_cache_unified_iswa_state * kv_state;
+    const llama_kv_cache_unified_iswa_context * mctx;
 };
 
 class llm_graph_input_attn_cross : public llm_graph_input_i {
@@ -305,12 +313,56 @@ public:
 
     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]
+    ggml_tensor * cross_kq_mask     = nullptr; // F32 [n_outputs_enc, n_batch, 1, 1]
+    ggml_tensor * cross_kq_mask_cnv = nullptr; // F32 [n_outputs_enc, n_batch, 1, 1]
 
     const llama_cross * cross = nullptr;
 };
 
+class llm_graph_input_mem_hybrid : public llm_graph_input_i {
+public:
+    llm_graph_input_mem_hybrid(
+            const llama_hparams & hparams,
+            const llama_cparams & cparams,
+            const llama_memory_hybrid_context * mctx) :
+        hparams(hparams),
+        cparams(cparams),
+        mctx(mctx) {
+    }
+    virtual ~llm_graph_input_mem_hybrid() = default;
+
+    void set_input(const llama_ubatch * ubatch) override;
+
+    ggml_tensor * s_copy; // I32 [kv_size]
+
+    ggml_tensor * get_k_idxs() const { return self_k_idxs; }
+    ggml_tensor * get_v_idxs() const { return self_v_idxs; }
+
+    ggml_tensor * get_kq_mask() const { return self_kq_mask_cnv; }
+
+    ggml_tensor * self_k_idxs = nullptr; // I64 [n_batch]
+    ggml_tensor * self_v_idxs = nullptr; // I64 [n_batch]
+
+    ggml_tensor * self_kq_mask     = nullptr; // F32 [n_kv, n_batch, 1, 1]
+    ggml_tensor * self_kq_mask_cnv = nullptr; //     [n_kv, n_batch, 1, 1]
+
+    const llama_hparams & hparams;
+    const llama_cparams & cparams;
+
+    const llama_memory_hybrid_context * mctx;
+};
+
+// TODO: remove this when ggml_scale_add is implemented
+class llm_graph_input_one : public llm_graph_input_i {
+public:
+    llm_graph_input_one() {}
+    virtual ~llm_graph_input_one() = default;
+
+    void set_input(const llama_ubatch * ubatch) override;
+
+    ggml_tensor * one = nullptr; // F32
+};
+
 //
 // llm_graph_result
 //
@@ -384,16 +436,19 @@ struct llm_graph_params {
     ggml_backend_sched_t sched;
     ggml_backend_t backend_cpu;
 
-    const llama_adapter_cvec   * cvec;
-    const llama_adapter_loras  * loras;
-    const llama_memory_state_i * mstate;
-    const llama_cross          * cross;
+    const llama_adapter_cvec     * cvec;
+    const llama_adapter_loras    * loras;
+    const llama_memory_context_i * mctx;
+    const llama_cross            * cross;
 
-    int32_t n_outputs;
+    uint32_t n_outputs;
 
     const llm_graph_cb & cb;
 };
 
+// used in build_rs to properly order writes and avoid unnecessary copies
+using llm_graph_get_rows_fn = std::function<ggml_tensor * (ggml_context *, ggml_tensor * states, ggml_tensor * ids)>;
+
 struct llm_graph_context {
     const llm_arch arch;
 
@@ -423,8 +478,8 @@ struct llm_graph_context {
     const float norm_eps;
     const float norm_rms_eps;
 
-    const int32_t n_tokens;
-    const int32_t n_outputs;
+    const int64_t n_tokens;
+    const int64_t n_outputs;
     const int32_t n_ctx_orig; // yarn
 
     const enum llama_pooling_type pooling_type;
@@ -436,18 +491,17 @@ struct llm_graph_context {
 
     ggml_backend_t 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_state_i * mstate;
-    const llama_cross          * cross;
+    const llama_adapter_cvec     * cvec;
+    const llama_adapter_loras    * loras;
+    const llama_memory_context_i * mctx;
+    const llama_cross            * cross;
 
     const llm_graph_cb & cb_func;
 
     std::unique_ptr<llm_graph_result> res;
 
     llm_graph_context(const llm_graph_params & params);
-
-    int64_t n_pos_per_embd() const;
+    virtual ~llm_graph_context() = default;
 
     void cb(ggml_tensor * cur, const char * name, int il) const;
 
@@ -519,14 +573,14 @@ struct llm_graph_context {
     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;
 
+    llm_graph_input_mem_hybrid * build_inp_mem_hybrid() const;
+
     //
     // attention
     //
@@ -573,14 +627,15 @@ struct llm_graph_context {
 
     llm_graph_input_attn_kv_unified_iswa * build_attn_inp_kv_unified_iswa() const;
 
+    // note: if k_cur or v_cur are not provided, they will not be stored in the memory
     ggml_tensor * build_attn(
             llm_graph_input_attn_kv_unified_iswa * 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 * k_cur, // [n_embd_head_k, n_head_k, n_tokens] optional
+            ggml_tensor * v_cur, // [n_embd_head_v, n_head_v, n_tokens] optional
             ggml_tensor * kq_b,
             ggml_tensor * v_mla, // [n_embd_head_v_mla, n_embd_head_v, n_head_v]
                   float   kq_scale,
@@ -601,23 +656,62 @@ struct llm_graph_context {
                   float   kq_scale,
                     int   il) const;
 
+    ggml_tensor * build_attn(
+            llm_graph_input_mem_hybrid * 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,
+            ggml_tensor * v_mla, // [n_embd_head_v_mla, n_embd_head_v, n_head_v]
+                  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;
+    // TODO: avoid notion of "kv"
+    // TODO: move this implementation to llama_memory_recurrent.
+    //       this is analogous to llama_kv_cache_unified::cpy_k / cpy_v
+    //       when moving, avoid passing `ggml_cgraph` - only pass `ggml_context`. would likely need to split the
+    //         implementation in 2 separate methods. the goal is to avoid calling `ggml_build_forward_expand` in
+    //         `llama_memory_recurrent`
+    ggml_tensor * build_rs(
+            ggml_cgraph * gf,
+            ggml_tensor * s,
+            ggml_tensor * state_copy,
+                int32_t   state_size,
+                int32_t   n_seqs,
+               uint32_t   n_kv,
+               uint32_t   kv_head,
+               uint32_t   kv_size,
+                int32_t   rs_zero,
+            const llm_graph_get_rows_fn & get_state_rows = ggml_get_rows) const;
+
+    llm_graph_input_rs * build_rs_inp() const;
+
+    ggml_tensor * build_rs(
+            llm_graph_input_rs * inp,
+            ggml_cgraph * gf,
+            ggml_tensor * s,
+                int32_t   state_size,
+                int32_t   n_seqs,
+            const llm_graph_get_rows_fn & get_state_rows = ggml_get_rows) const;
+
+    ggml_tensor * build_rs(
+            llm_graph_input_mem_hybrid * inp,
+            ggml_cgraph * gf,
+            ggml_tensor * s,
+                int32_t   state_size,
+                int32_t   n_seqs,
+            const llm_graph_get_rows_fn & get_state_rows = ggml_get_rows) const;
 
     ggml_tensor * build_rwkv_token_shift_load(
-             ggml_cgraph * gf,
-             ggml_tensor * state_copy,
-             ggml_tensor * state_mask,
-      const llama_ubatch & ubatch,
+        llm_graph_input_rs * inp,
+               ggml_cgraph * gf,
+        const llama_ubatch & ubatch,
                      int   il) const;
 
     ggml_tensor * build_rwkv_token_shift_store(
diff --git a/src/llama-hparams.cpp b/src/llama-hparams.cpp
index 1499eb08a..86c814d51 100644
--- a/src/llama-hparams.cpp
+++ b/src/llama-hparams.cpp
@@ -65,7 +65,7 @@ uint32_t llama_hparams::n_embd_v_gqa(uint32_t il) const {
     return n_embd_head_v * n_head_kv;
 }
 
-uint32_t llama_hparams::n_embd_k_s() const {
+uint32_t llama_hparams::n_embd_r() const {
     if (wkv_head_size != 0) {
         // for RWKV models
         return token_shift_count * n_embd;
@@ -73,10 +73,11 @@ uint32_t llama_hparams::n_embd_k_s() const {
 
     // TODO: maybe support other convolution strides than 1
     // NOTE: since the first column of the conv_state is shifted out each time, it's not actually needed
-    return (ssm_d_conv > 0 ? ssm_d_conv - 1 : 0) * ssm_d_inner;
+    // Corresponds to Mamba's conv_states size
+    return (ssm_d_conv > 0 ? ssm_d_conv - 1 : 0) * (ssm_d_inner + 2*ssm_n_group*ssm_d_state);
 }
 
-uint32_t llama_hparams::n_embd_v_s() const {
+uint32_t llama_hparams::n_embd_s() const {
     if (wkv_head_size != 0) {
         // corresponds to RWKV's wkv_states size
         return n_embd * wkv_head_size;
@@ -86,6 +87,14 @@ uint32_t llama_hparams::n_embd_v_s() const {
     return ssm_d_state * ssm_d_inner;
 }
 
+bool llama_hparams::is_recurrent(uint32_t il) const {
+    return recurrent_layer_arr[il];
+}
+
+uint32_t llama_hparams::n_pos_per_embd() const {
+    return rope_type == LLAMA_ROPE_TYPE_MROPE ? 4 : 1;
+}
+
 bool llama_hparams::is_swa(uint32_t il) const {
     if (il < n_layer) {
         return swa_layers[il];
diff --git a/src/llama-hparams.h b/src/llama-hparams.h
index b2bcb8b01..476d0a5ea 100644
--- a/src/llama-hparams.h
+++ b/src/llama-hparams.h
@@ -114,6 +114,10 @@ struct llama_hparams {
     uint32_t ssm_d_inner = 0;
     uint32_t ssm_d_state = 0;
     uint32_t ssm_dt_rank = 0;
+    uint32_t ssm_n_group = 0;
+
+    // for hybrid state space models
+    std::array<bool, LLAMA_MAX_LAYERS> recurrent_layer_arr;
 
     bool ssm_dt_b_c_rms = false;
 
@@ -140,6 +144,12 @@ struct llama_hparams {
     uint32_t n_attn_temp_floor_scale = 8192;
     float    f_attn_temp_scale       = 0.1;
 
+    // gemma3n altup
+    uint32_t n_altup      = 4; // altup_num_inputs
+    uint32_t i_altup_act  = 0; // altup_active_idx
+    uint32_t laurel_rank  = 64;
+    uint32_t n_embd_altup = 256;
+
     // 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;
@@ -181,10 +191,15 @@ struct llama_hparams {
 
     // dimension of the rolling state embeddings
     // corresponds to Mamba's conv_states size or RWKV's token_shift states size
-    uint32_t n_embd_k_s() const;
+    uint32_t n_embd_r() const;
 
     // dimension of the recurrent state embeddings
-    uint32_t n_embd_v_s() const;
+    uint32_t n_embd_s() const;
+
+    // whether or not the given layer is recurrent (for hybrid models)
+    bool is_recurrent(uint32_t il) const;
+
+    uint32_t n_pos_per_embd() const;
 
     bool is_swa(uint32_t il) const;
 };
diff --git a/src/llama-kv-cache-unified-iswa.cpp b/src/llama-kv-cache-unified-iswa.cpp
index 0eb045634..fe207ad53 100644
--- a/src/llama-kv-cache-unified-iswa.cpp
+++ b/src/llama-kv-cache-unified-iswa.cpp
@@ -52,9 +52,9 @@ llama_kv_cache_unified_iswa::llama_kv_cache_unified_iswa(
             hparams.n_swa, hparams.swa_type);
 }
 
-void llama_kv_cache_unified_iswa::clear() {
-    kv_base->clear();
-    kv_swa ->clear();
+void llama_kv_cache_unified_iswa::clear(bool data) {
+    kv_base->clear(data);
+    kv_swa ->clear(data);
 }
 
 bool llama_kv_cache_unified_iswa::seq_rm(llama_seq_id seq_id, llama_pos p0, llama_pos p1) {
@@ -95,54 +95,93 @@ llama_pos llama_kv_cache_unified_iswa::seq_pos_max(llama_seq_id seq_id) const {
     return kv_swa->seq_pos_max(seq_id);
 }
 
-llama_memory_state_ptr llama_kv_cache_unified_iswa::init_batch(const llama_batch & batch, uint32_t n_ubatch, bool embd_pooled, bool logits_all) {
-    GGML_UNUSED(embd_pooled);
+llama_memory_context_ptr llama_kv_cache_unified_iswa::init_batch(llama_batch_allocr & balloc, uint32_t n_ubatch, bool embd_all) {
+    GGML_UNUSED(embd_all);
 
-    // TODO: if we fail with split_simple, we should attempt different splitting strategies
+    // first try simple split
+    do {
+        balloc.split_reset();
+
+        std::vector<llama_ubatch> ubatches;
+        while (true) {
+            auto ubatch = balloc.split_simple(n_ubatch);
+
+            if (ubatch.n_tokens == 0) {
+                break;
+            }
+
+            ubatches.push_back(std::move(ubatch)); // NOLINT
+        }
+
+        if (balloc.get_n_used() < balloc.get_n_tokens()) {
+            // failed to find a suitable split
+            break;
+        }
+
+        auto sinfos_base = kv_base->prepare(ubatches);
+        if (sinfos_base.empty()) {
+            break;
+        }
+
+        auto sinfos_swa = kv_swa->prepare(ubatches);
+        if (sinfos_swa.empty()) {
+            break;
+        }
+
+        assert(sinfos_base.size() == sinfos_swa.size());
+
+        return std::make_unique<llama_kv_cache_unified_iswa_context>(
+                this, std::move(sinfos_base), std::move(sinfos_swa), std::move(ubatches));
+    } while (false);
+
+    // if it fails, try equal split
+    do {
+        balloc.split_reset();
+
+        std::vector<llama_ubatch> ubatches;
+        while (true) {
+            auto ubatch = balloc.split_equal(n_ubatch, false);
+
+            if (ubatch.n_tokens == 0) {
+                break;
+            }
+
+            ubatches.push_back(std::move(ubatch)); // NOLINT
+        }
+
+        if (balloc.get_n_used() < balloc.get_n_tokens()) {
+            // failed to find a suitable split
+            break;
+        }
+
+        auto sinfos_base = kv_base->prepare(ubatches);
+        if (sinfos_base.empty()) {
+            break;
+        }
+
+        auto sinfos_swa = kv_swa->prepare(ubatches);
+        if (sinfos_swa.empty()) {
+            break;
+        }
+
+        assert(sinfos_base.size() == sinfos_swa.size());
+
+        return std::make_unique<llama_kv_cache_unified_iswa_context>(
+                this, std::move(sinfos_base), std::move(sinfos_swa), std::move(ubatches));
+    } while (false);
+
+    // TODO: if we fail again, we should attempt different splitting strategies
     //       but to do that properly, we first have to refactor the batches to be more flexible
 
-    auto sbatch = llama_sbatch(batch, hparams.n_embd, true, logits_all);
-
-    std::vector<llama_ubatch> ubatches;
-
-    while (sbatch.n_tokens > 0) {
-        auto ubatch = sbatch.split_simple(n_ubatch);
-
-        ubatches.push_back(ubatch);
-    }
-
-    auto heads_base = kv_base->prepare(ubatches);
-    if (heads_base.empty()) {
-        return std::make_unique<llama_kv_cache_unified_iswa_state>(LLAMA_MEMORY_STATUS_FAILED_PREPARE);
-    }
-
-    auto heads_swa = kv_swa->prepare(ubatches);
-    if (heads_swa.empty()) {
-        return std::make_unique<llama_kv_cache_unified_iswa_state>(LLAMA_MEMORY_STATUS_FAILED_PREPARE);
-    }
-
-    assert(heads_base.size() == heads_swa.size());
-
-    return std::make_unique<llama_kv_cache_unified_iswa_state>(LLAMA_MEMORY_STATUS_SUCCESS,
-            this, std::move(sbatch), std::move(heads_base), std::move(heads_swa), std::move(ubatches));
+    return std::make_unique<llama_kv_cache_unified_iswa_context>(LLAMA_MEMORY_STATUS_FAILED_PREPARE);
 }
 
-llama_memory_state_ptr llama_kv_cache_unified_iswa::init_full() {
-    return std::make_unique<llama_kv_cache_unified_iswa_state>(LLAMA_MEMORY_STATUS_SUCCESS, this);
+llama_memory_context_ptr llama_kv_cache_unified_iswa::init_full() {
+    return std::make_unique<llama_kv_cache_unified_iswa_context>(this);
 }
 
-bool llama_kv_cache_unified_iswa::update(llama_context & lctx) {
-    bool res = false;
-
-    res = res | kv_base->update(lctx);
-    res = res | kv_swa ->update(lctx);
-
-    return res;
-}
-
-void llama_kv_cache_unified_iswa::defrag_sched(float thold) {
-    kv_base->defrag_sched(thold);
-    kv_swa ->defrag_sched(thold);
+llama_memory_context_ptr llama_kv_cache_unified_iswa::init_update(llama_context * lctx, bool optimize) {
+    return std::make_unique<llama_kv_cache_unified_iswa_context>(this, lctx, optimize);
 }
 
 bool llama_kv_cache_unified_iswa::get_can_shift() const {
@@ -168,40 +207,46 @@ llama_kv_cache_unified * llama_kv_cache_unified_iswa::get_swa() const {
 }
 
 //
-// llama_kv_cache_unified_iswa_state
+// llama_kv_cache_unified_iswa_context
 //
 
-llama_kv_cache_unified_iswa_state::llama_kv_cache_unified_iswa_state(llama_memory_status status) : status(status) {}
+llama_kv_cache_unified_iswa_context::llama_kv_cache_unified_iswa_context(llama_memory_status status) : status(status) {}
 
-llama_kv_cache_unified_iswa_state::llama_kv_cache_unified_iswa_state(
-        llama_memory_status status,
-        llama_kv_cache_unified_iswa * kv) : status(status) {
-    state_base.reset(new llama_kv_cache_unified_state(status, kv->get_base()));
-    state_swa .reset(new llama_kv_cache_unified_state(status, kv->get_swa ()));
+llama_kv_cache_unified_iswa_context::llama_kv_cache_unified_iswa_context(
+        llama_kv_cache_unified_iswa * kv) :
+    ctx_base(kv->get_base()->init_full()),
+    ctx_swa (kv->get_swa ()->init_full()),
+    status(llama_memory_status_combine(ctx_base->get_status(), ctx_swa->get_status())) {
 }
 
-llama_kv_cache_unified_iswa_state::llama_kv_cache_unified_iswa_state(
-        llama_memory_status status,
+llama_kv_cache_unified_iswa_context::llama_kv_cache_unified_iswa_context(
         llama_kv_cache_unified_iswa * kv,
-        llama_sbatch sbatch,
-        std::vector<uint32_t> heads_base,
-        std::vector<uint32_t> heads_swa,
-        std::vector<llama_ubatch> ubatches)
-    : status(status),
-    sbatch(std::move(sbatch)),
-    ubatches(std::move(ubatches)) {
-        // note: here we copy the ubatches. not sure if this is ideal
-        state_base.reset(new llama_kv_cache_unified_state(status, kv->get_base(), {}, std::move(heads_base), this->ubatches));
-        state_swa .reset(new llama_kv_cache_unified_state(status, kv->get_swa (), {}, std::move(heads_swa),  this->ubatches));
-    }
+        llama_context * lctx,
+        bool optimize) :
+    ctx_base(kv->get_base()->init_update(lctx, optimize)),
+    ctx_swa (kv->get_swa ()->init_update(lctx, optimize)),
+    status(llama_memory_status_combine(ctx_base->get_status(), ctx_swa->get_status())) {
+}
 
-llama_kv_cache_unified_iswa_state:: ~llama_kv_cache_unified_iswa_state() = default;
+llama_kv_cache_unified_iswa_context::llama_kv_cache_unified_iswa_context(
+        llama_kv_cache_unified_iswa * kv,
+        slot_info_vec_t sinfos_base,
+        slot_info_vec_t sinfos_swa,
+        std::vector<llama_ubatch> ubatches) :
+    ubatches(std::move(ubatches)),
+    // note: here we copy the ubatches. not sure if this is ideal
+    ctx_base(new llama_kv_cache_unified_context(kv->get_base(), std::move(sinfos_base), this->ubatches)),
+    ctx_swa (new llama_kv_cache_unified_context(kv->get_swa (), std::move(sinfos_swa),  this->ubatches)),
+    status(llama_memory_status_combine(ctx_base->get_status(), ctx_swa->get_status())) {
+}
 
-bool llama_kv_cache_unified_iswa_state::next() {
+llama_kv_cache_unified_iswa_context:: ~llama_kv_cache_unified_iswa_context() = default;
+
+bool llama_kv_cache_unified_iswa_context::next() {
     assert(status == LLAMA_MEMORY_STATUS_SUCCESS);
 
-    state_base->next();
-    state_swa ->next();
+    ctx_base->next();
+    ctx_swa ->next();
 
     if (++i_next >= ubatches.size()) {
         return false;
@@ -210,40 +255,35 @@ bool llama_kv_cache_unified_iswa_state::next() {
     return true;
 }
 
-bool llama_kv_cache_unified_iswa_state::apply() {
-    assert(status == LLAMA_MEMORY_STATUS_SUCCESS);
+bool llama_kv_cache_unified_iswa_context::apply() {
+    assert(!llama_memory_status_is_fail(status));
 
     bool res = true;
 
-    res = res & state_base->apply();
-    res = res & state_swa ->apply();
+    res = res & ctx_base->apply();
+    res = res & ctx_swa ->apply();
 
     return res;
 }
 
-std::vector<int64_t> & llama_kv_cache_unified_iswa_state::out_ids() {
-    assert(status == LLAMA_MEMORY_STATUS_SUCCESS);
-
-    return sbatch.out_ids;
-}
-
-llama_memory_status llama_kv_cache_unified_iswa_state::get_status() const {
+llama_memory_status llama_kv_cache_unified_iswa_context::get_status() const {
     return status;
 }
 
-const llama_ubatch & llama_kv_cache_unified_iswa_state::get_ubatch() const {
+const llama_ubatch & llama_kv_cache_unified_iswa_context::get_ubatch() const {
     assert(status == LLAMA_MEMORY_STATUS_SUCCESS);
+
     return ubatches[i_next];
 }
 
-const llama_kv_cache_unified_state * llama_kv_cache_unified_iswa_state::get_base() const {
+const llama_kv_cache_unified_context * llama_kv_cache_unified_iswa_context::get_base() const {
     assert(status == LLAMA_MEMORY_STATUS_SUCCESS);
 
-    return state_base.get();
+    return static_cast<const llama_kv_cache_unified_context *>(ctx_base.get());
 }
 
-const llama_kv_cache_unified_state * llama_kv_cache_unified_iswa_state::get_swa()  const {
+const llama_kv_cache_unified_context * llama_kv_cache_unified_iswa_context::get_swa()  const {
     assert(status == LLAMA_MEMORY_STATUS_SUCCESS);
 
-    return state_swa.get();
+    return static_cast<const llama_kv_cache_unified_context *>(ctx_swa.get());
 }
diff --git a/src/llama-kv-cache-unified-iswa.h b/src/llama-kv-cache-unified-iswa.h
index 8b067da03..23205d826 100644
--- a/src/llama-kv-cache-unified-iswa.h
+++ b/src/llama-kv-cache-unified-iswa.h
@@ -11,7 +11,7 @@
 // utilizes two instances of llama_kv_cache_unified
 //   the first instance is for the non-SWA layers of the model and the second instance is for the SWA layers
 
-class llama_kv_cache_unified_iswa : public llama_kv_cache {
+class llama_kv_cache_unified_iswa : public llama_memory_i {
 public:
     llama_kv_cache_unified_iswa(
             const llama_model & model,
@@ -31,7 +31,18 @@ public:
     // llama_memory_i
     //
 
-    void clear() override;
+    llama_memory_context_ptr init_batch(
+            llama_batch_allocr & balloc,
+            uint32_t n_ubatch,
+            bool embd_all) override;
+
+    llama_memory_context_ptr init_full() override;
+
+    llama_memory_context_ptr init_update(llama_context * lctx, bool optimize) override;
+
+    bool get_can_shift() const override;
+
+    void clear(bool data) 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;
@@ -42,24 +53,6 @@ public:
     llama_pos seq_pos_min(llama_seq_id seq_id) const override;
     llama_pos seq_pos_max(llama_seq_id seq_id) const override;
 
-    //
-    // llama_kv_cache
-    //
-
-    llama_memory_state_ptr init_batch(
-            const llama_batch & batch,
-            uint32_t n_ubatch,
-            bool embd_pooled,
-            bool logits_all) override;
-
-    llama_memory_state_ptr init_full() override;
-
-    bool update(llama_context & lctx) override;
-
-    void defrag_sched(float thold) override;
-
-    bool get_can_shift() const override;
-
     // state write/load
 
     void state_write(llama_io_write_i & io, llama_seq_id seq_id = -1) const override;
@@ -79,58 +72,59 @@ private:
     std::unique_ptr<llama_kv_cache_unified> kv_swa;
 };
 
-class llama_kv_cache_unified_iswa_state : public llama_memory_state_i {
+class llama_kv_cache_unified_iswa_context : public llama_memory_context_i {
 public:
-    // used for errors
-    llama_kv_cache_unified_iswa_state(llama_memory_status status);
+    using slot_info_vec_t = llama_kv_cache_unified::slot_info_vec_t;
 
-    // used to create a full-cache state
-    llama_kv_cache_unified_iswa_state(
-            llama_memory_status status,
+    // used for errors
+    llama_kv_cache_unified_iswa_context(llama_memory_status status);
+
+    // used to create a full-cache context
+    llama_kv_cache_unified_iswa_context(
             llama_kv_cache_unified_iswa * kv);
 
-    // used to create a state from a batch
-    llama_kv_cache_unified_iswa_state(
-            llama_memory_status status,
+    // used to create an update context
+    llama_kv_cache_unified_iswa_context(
             llama_kv_cache_unified_iswa * kv,
-            llama_sbatch sbatch,
-            std::vector<uint32_t> heads_base,
-            std::vector<uint32_t> heads_swa,
+            llama_context * lctx,
+            bool optimize);
+
+    // used to create a batch processing context from a batch
+    llama_kv_cache_unified_iswa_context(
+            llama_kv_cache_unified_iswa * kv,
+            slot_info_vec_t sinfos_base,
+            slot_info_vec_t sinfos_swa,
             std::vector<llama_ubatch> ubatches);
 
-    virtual ~llama_kv_cache_unified_iswa_state();
+    virtual ~llama_kv_cache_unified_iswa_context();
 
     //
-    // llama_memory_state_i
+    // llama_memory_context_i
     //
 
     bool next()  override;
     bool apply() override;
 
-    std::vector<int64_t> & out_ids() override;
-
     llama_memory_status  get_status() const override;
     const llama_ubatch & get_ubatch() const override;
 
     //
-    // llama_kv_cache_unified_iswa_state specific API
+    // llama_kv_cache_unified_iswa_context specific API
     //
 
-    const llama_kv_cache_unified_state * get_base() const;
-    const llama_kv_cache_unified_state * get_swa()  const;
+    const llama_kv_cache_unified_context * get_base() const;
+    const llama_kv_cache_unified_context * get_swa()  const;
 
 private:
-    const llama_memory_status status;
-
     //llama_kv_cache_unified_iswa * kv;
 
-    llama_sbatch sbatch;
-
     // the index of the next ubatch to process
     size_t i_next = 0;
 
     std::vector<llama_ubatch> ubatches;
 
-    std::unique_ptr<llama_kv_cache_unified_state> state_base;
-    std::unique_ptr<llama_kv_cache_unified_state> state_swa;
+    const llama_memory_context_ptr ctx_base;
+    const llama_memory_context_ptr ctx_swa;
+
+    const llama_memory_status status;
 };
diff --git a/src/llama-kv-cache-unified.cpp b/src/llama-kv-cache-unified.cpp
index a81715476..d3129cc53 100644
--- a/src/llama-kv-cache-unified.cpp
+++ b/src/llama-kv-cache-unified.cpp
@@ -1,6 +1,7 @@
 #include "llama-kv-cache-unified.h"
 
 #include "llama-impl.h"
+#include "llama-io.h"
 #include "llama-model.h"
 #include "llama-context.h"
 
@@ -32,13 +33,19 @@ llama_kv_cache_unified::llama_kv_cache_unified(
 
     GGML_ASSERT(kv_size % n_pad == 0);
 
+    // TODO: this is temporary until we support passing reuse layer filters [KV_REUSE]
+    auto n_layer_cache = hparams.n_layer;
+    if (model.arch == LLM_ARCH_GEMMA3N) {
+        n_layer_cache = 20;
+    }
+
     // create a context for each buffer type
     std::map<ggml_backend_buffer_type_t, ggml_context *> 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()) {
             ggml_init_params params = {
-                /*.mem_size   =*/ size_t(2u*hparams.n_layer*ggml_tensor_overhead()),
+                /*.mem_size   =*/ size_t(2u*n_layer_cache*ggml_tensor_overhead()),
                 /*.mem_buffer =*/ NULL,
                 /*.no_alloc   =*/ true,
             };
@@ -61,14 +68,14 @@ llama_kv_cache_unified::llama_kv_cache_unified(
 
     cells.resize(kv_size);
 
-    for (uint32_t il = 0; il < hparams.n_layer; il++) {
+    for (uint32_t il = 0; il < n_layer_cache; il++) {
         if (filter && !filter(il)) {
             LLAMA_LOG_DEBUG("%s: layer %3d: skipped\n", __func__, il);
             continue;
         }
 
-        const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa(il) + hparams.n_embd_k_s();
-        const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il) + hparams.n_embd_v_s();
+        const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa(il);
+        const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il);
 
         const char * dev_name = "CPU";
 
@@ -101,6 +108,26 @@ llama_kv_cache_unified::llama_kv_cache_unified(
         layers.push_back({ il, k, v });
     }
 
+    // TODO: this is temporary until we support passing reuse layer filters [KV_REUSE]
+    if (model.arch == LLM_ARCH_GEMMA3N) {
+        LLAMA_LOG_DEBUG("%s: GEMMA3N: reuse layers [%d, %d]\n", __func__, n_layer_cache, hparams.n_layer - 1);
+
+        for (uint32_t il = n_layer_cache; il < hparams.n_layer; il++) {
+            if (filter && !filter(il)) {
+                LLAMA_LOG_DEBUG("%s: layer %3d: skipped\n", __func__, il);
+                continue;
+            }
+
+            const bool     is_swa   = hparams.is_swa(il);
+            const uint32_t il_reuse = n_layer_cache - (is_swa ? 2 : 1);
+
+            GGML_ASSERT(map_layer_ids.find(il_reuse) != map_layer_ids.end());
+            map_layer_ids[il] = map_layer_ids[il_reuse];
+
+            LLAMA_LOG_DEBUG("%s: layer %3d: reuse layer %d, isw = %d\n", __func__, il, il_reuse, is_swa);
+        }
+    }
+
     // allocate tensors and initialize the buffers to avoid NaNs in the padding
     for (auto it : ctx_map) {
         auto * buft = it.first;
@@ -126,15 +153,27 @@ llama_kv_cache_unified::llama_kv_cache_unified(
                 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));
     }
+
+    const char * LLAMA_KV_CACHE_DEBUG = getenv("LLAMA_KV_CACHE_DEBUG");
+    debug = LLAMA_KV_CACHE_DEBUG ? atoi(LLAMA_KV_CACHE_DEBUG) : 0;
+
+    const char * LLAMA_SET_ROWS = getenv("LLAMA_SET_ROWS");
+    supports_set_rows = LLAMA_SET_ROWS ? atoi(LLAMA_SET_ROWS) : 0;
+
+    if (!supports_set_rows) {
+        LLAMA_LOG_WARN("%s: LLAMA_SET_ROWS=0, using old ggml_cpy() method for backwards compatibility\n", __func__);
+    }
 }
 
-void llama_kv_cache_unified::clear() {
+void llama_kv_cache_unified::clear(bool data) {
     cells.reset();
 
     head = 0;
 
-    for (auto & buf : bufs) {
-        ggml_backend_buffer_clear(buf.get(), 0);
+    if (data) {
+        for (auto & buf : bufs) {
+            ggml_backend_buffer_clear(buf.get(), 0);
+        }
     }
 }
 
@@ -149,12 +188,27 @@ bool llama_kv_cache_unified::seq_rm(llama_seq_id seq_id, llama_pos p0, llama_pos
         p1 = std::numeric_limits<llama_pos>::max();
     }
 
-    for (uint32_t i = 0; i < cells.size(); ++i) {
-        if (!cells.pos_in(i, p0, p1)) {
-            continue;
-        }
+    if (seq_id >= 0) {
+        for (uint32_t i = 0; i < cells.size(); ++i) {
+            if (!cells.pos_in(i, p0, p1)) {
+                continue;
+            }
+
+            if (cells.seq_has(i, seq_id) && cells.seq_rm(i, seq_id)) {
+                if (new_head == cells.size()) {
+                    new_head = i;
+                }
+            }
+        }
+    } else {
+        // match any sequence
+        for (uint32_t i = 0; i < cells.size(); ++i) {
+            if (!cells.pos_in(i, p0, p1)) {
+                continue;
+            }
+
+            cells.rm(i);
 
-        if (cells.seq_has(i, seq_id) && cells.seq_rm(i, seq_id)) {
             if (new_head == cells.size()) {
                 new_head = i;
             }
@@ -286,39 +340,87 @@ llama_pos llama_kv_cache_unified::seq_pos_max(llama_seq_id seq_id) const {
     return cells.seq_pos_max(seq_id);
 }
 
-llama_memory_state_ptr llama_kv_cache_unified::init_batch(
-            const llama_batch & batch,
+llama_memory_context_ptr llama_kv_cache_unified::init_batch(
+            llama_batch_allocr & balloc,
             uint32_t n_ubatch,
-            bool embd_pooled,
-            bool logits_all) {
-    GGML_UNUSED(embd_pooled);
+            bool embd_all) {
+    GGML_UNUSED(embd_all);
 
-    auto sbatch = llama_sbatch(batch, hparams.n_embd, true, logits_all);
+    do {
+        balloc.split_reset();
 
-    std::vector<llama_ubatch> ubatches;
-    while (sbatch.n_tokens > 0) {
-        ubatches.push_back(sbatch.split_simple(n_ubatch));
-    }
+        std::vector<llama_ubatch> ubatches;
+        while (true) {
+            auto ubatch = balloc.split_simple(n_ubatch);
 
-    auto heads = prepare(ubatches);
-    if (heads.empty()) {
-        return std::make_unique<llama_kv_cache_unified_state>(LLAMA_MEMORY_STATUS_FAILED_PREPARE);
-    }
+            if (ubatch.n_tokens == 0) {
+                break;
+            }
 
-    return std::make_unique<llama_kv_cache_unified_state>(LLAMA_MEMORY_STATUS_SUCCESS,
-            this, std::move(sbatch), std::move(heads), std::move(ubatches));
+            ubatches.push_back(std::move(ubatch)); // NOLINT
+        }
+
+        if (balloc.get_n_used() < balloc.get_n_tokens()) {
+            // failed to find a suitable split
+            break;
+        }
+
+        auto sinfos = prepare(ubatches);
+        if (sinfos.empty()) {
+            break;
+        }
+
+        return std::make_unique<llama_kv_cache_unified_context>(
+                this, std::move(sinfos), std::move(ubatches));
+    } while (false);
+
+    return std::make_unique<llama_kv_cache_unified_context>(LLAMA_MEMORY_STATUS_FAILED_PREPARE);
 }
 
-llama_memory_state_ptr llama_kv_cache_unified::init_full() {
-    return std::make_unique<llama_kv_cache_unified_state>(LLAMA_MEMORY_STATUS_SUCCESS, this);
+llama_memory_context_ptr llama_kv_cache_unified::init_full() {
+    return std::make_unique<llama_kv_cache_unified_context>(this);
 }
 
-std::vector<uint32_t> llama_kv_cache_unified::prepare(const std::vector<llama_ubatch> & ubatches) {
-    std::vector<uint32_t> res;
+llama_memory_context_ptr llama_kv_cache_unified::init_update(llama_context * lctx, bool optimize) {
+    bool do_shift = get_has_shift();
+
+    defrag_info dinfo;
+
+    // see if we need to defrag
+    {
+        bool do_defrag = optimize;
+
+        const auto thold = lctx->get_cparams().defrag_thold;
+
+        if (!do_defrag && thold > 0.0f) {
+            const auto n_kv = cells.used_max_p1();
+
+            // - do not defrag small contexts (i.e. < 2048 tokens)
+            // - count the padding towards the number of used tokens
+            const float fragmentation = n_kv >= 2048 ? std::max(0.0f, 1.0f - (float(cells.get_used() + n_pad)/n_kv)) : 0.0f;
+
+            if (fragmentation > thold) {
+                LLAMA_LOG_DEBUG("%s: fragmentation: %.2f - requesting defrag\n", __func__, fragmentation);
+
+                do_defrag = true;
+            }
+        }
+
+        if (do_defrag) {
+            dinfo = defrag_prepare(lctx->graph_max_nodes());
+        }
+    }
+
+    return std::make_unique<llama_kv_cache_unified_context>(this, lctx, do_shift, std::move(dinfo));
+}
+
+llama_kv_cache_unified::slot_info_vec_t llama_kv_cache_unified::prepare(const std::vector<llama_ubatch> & ubatches) {
+    llama_kv_cache_unified::slot_info_vec_t res;
 
     struct state {
         uint32_t head_old; // old position of the head, before placing the ubatch
-        uint32_t head_new; // new position of the head, after placing the ubatch
+
+        slot_info sinfo; // slot info for the ubatch
 
         llama_kv_cells_unified cells; // copy of the old cells, before placing the ubatch
     };
@@ -329,26 +431,29 @@ std::vector<uint32_t> llama_kv_cache_unified::prepare(const std::vector<llama_ub
     bool success = true;
 
     for (const auto & ubatch : ubatches) {
+        // non-continuous slots require support for ggml_set_rows()
+        const bool cont = supports_set_rows ? false : true;
+
         // only find a suitable slot for the ubatch. don't modify the cells yet
-        const int32_t head_new = find_slot(ubatch);
-        if (head_new < 0) {
+        const auto sinfo_new = find_slot(ubatch, cont);
+        if (sinfo_new.empty()) {
             success = false;
             break;
         }
 
         // remeber the position that we found
-        res.push_back(head_new);
+        res.push_back(sinfo_new);
 
         // store the old state of the cells in the recovery stack
-        states.push_back({head, (uint32_t) head_new, cells.cp(head_new, ubatch.n_tokens)});
+        states.push_back({head, sinfo_new, cells.cp(sinfo_new.idxs)});
 
         // now emplace the ubatch
-        apply_ubatch(head_new, ubatch);
+        apply_ubatch(sinfo_new, ubatch);
     }
 
     // iterate backwards and restore the cells to their original state
     for (auto it = states.rbegin(); it != states.rend(); ++it) {
-        cells.set(it->head_new, it->cells);
+        cells.set(it->sinfo.idxs, it->cells);
         head = it->head_old;
     }
 
@@ -359,12 +464,12 @@ std::vector<uint32_t> llama_kv_cache_unified::prepare(const std::vector<llama_ub
     return res;
 }
 
-bool llama_kv_cache_unified::update(llama_context & lctx) {
+bool llama_kv_cache_unified::update(llama_context * lctx, bool do_shift, const defrag_info & dinfo) {
     bool updated = false;
 
-    auto * sched = lctx.get_sched();
+    auto * sched = lctx->get_sched();
 
-    if (cells.get_has_shift()) {
+    if (do_shift) {
         if (!get_can_shift()) {
             GGML_ABORT("The current KV cache / model configuration does not support K-shift");
         }
@@ -375,9 +480,9 @@ bool llama_kv_cache_unified::update(llama_context & lctx) {
         if (hparams.rope_type != LLAMA_ROPE_TYPE_NONE) {
             ggml_backend_sched_reset(sched);
 
-            auto * gf = lctx.graph_init();
+            auto * gf = lctx->graph_init();
 
-            auto res = build_graph_shift(lctx.get_cparams(), lctx.get_ctx_compute(), gf);
+            auto res = build_graph_shift(lctx->get_cparams(), lctx->get_ctx_compute(), gf);
             if (!res) {
                 LLAMA_LOG_ERROR("%s: failed to build graph for K-shift\n", __func__);
                 return updated;
@@ -390,7 +495,7 @@ bool llama_kv_cache_unified::update(llama_context & lctx) {
 
             res->set_inputs(nullptr);
 
-            if (lctx.graph_compute(gf, false) != GGML_STATUS_SUCCESS) {
+            if (lctx->graph_compute(gf, false) != GGML_STATUS_SUCCESS) {
                 LLAMA_LOG_ERROR("%s: failed to compute K-shift\n", __func__);
                 return updated;
             }
@@ -401,57 +506,56 @@ bool llama_kv_cache_unified::update(llama_context & lctx) {
         cells.reset_shift();
     }
 
-    if (do_defrag) {
+    if (!dinfo.empty()) {
         LLAMA_LOG_DEBUG("%s: defragmenting KV cache\n", __func__);
 
-        if (defrag_prepare(lctx.graph_max_nodes())) {
-            ggml_backend_sched_reset(sched);
+        // apply moves:
+        {
+            const auto n_kv = dinfo.ids.size();
 
-            auto * gf = lctx.graph_init();
+            for (uint32_t i = 0; i < n_kv; ++i) {
+                assert(dinfo.ids[i] <= n_kv);
 
-            auto res = build_graph_defrag(lctx.get_cparams(), lctx.get_ctx_compute(), gf);
-            if (!res) {
-                LLAMA_LOG_ERROR("%s: failed to build graph for defrag\n", __func__);
-                return updated;
+                if (dinfo.ids[i] == n_kv || dinfo.ids[i] == i) {
+                    continue;
+                }
+
+                cells.mv(i, dinfo.ids[i]);
             }
 
-            if (!ggml_backend_sched_alloc_graph(sched, gf)) {
-                LLAMA_LOG_ERROR("%s: failed to allocate compute graph for defrag\n", __func__);
-                return updated;
-            }
-
-            res->set_inputs(nullptr);
-
-            if (lctx.graph_compute(gf, false) != GGML_STATUS_SUCCESS) {
-                LLAMA_LOG_ERROR("%s: failed to compute defrag\n", __func__);
-                return updated;
-            }
-
-            updated = true;
+            // reset the head so we can find the first free slot during the next ubatch
+            head = 0;
         }
 
-        do_defrag = false;
+        ggml_backend_sched_reset(sched);
+
+        auto * gf = lctx->graph_init();
+
+        auto res = build_graph_defrag(lctx->get_cparams(), lctx->get_ctx_compute(), gf, dinfo);
+        if (!res) {
+            LLAMA_LOG_ERROR("%s: failed to build graph for defrag\n", __func__);
+            return updated;
+        }
+
+        if (!ggml_backend_sched_alloc_graph(sched, gf)) {
+            LLAMA_LOG_ERROR("%s: failed to allocate compute graph for defrag\n", __func__);
+            return updated;
+        }
+
+        res->set_inputs(nullptr);
+
+        if (lctx->graph_compute(gf, false) != GGML_STATUS_SUCCESS) {
+            LLAMA_LOG_ERROR("%s: failed to compute defrag\n", __func__);
+            return updated;
+        }
+
+        updated = true;
     }
 
     return updated;
 }
 
-void llama_kv_cache_unified::defrag_sched(float thold) {
-    const auto n_kv = cells.used_max_p1();
-
-    // - do not defrag small contexts (i.e. < 2048 tokens)
-    // - count the padding towards the number of used tokens
-    const float fragmentation = n_kv >= 2048 ? std::max(0.0f, 1.0f - (float(cells.get_used() + n_pad)/n_kv)) : 0.0f;
-
-    // queue defragmentation for next llama_kv_cache_update
-    if (fragmentation > thold) {
-        LLAMA_LOG_DEBUG("%s: fragmentation: %.2f - requesting defrag\n", __func__, fragmentation);
-
-        do_defrag = true;
-    }
-}
-
-int32_t llama_kv_cache_unified::find_slot(const llama_ubatch & ubatch) const {
+llama_kv_cache_unified::slot_info llama_kv_cache_unified::find_slot(const llama_ubatch & ubatch, bool cont) const {
     const uint32_t n_tokens = ubatch.n_tokens;
 
     uint32_t head_cur = this->head;
@@ -462,131 +566,202 @@ int32_t llama_kv_cache_unified::find_slot(const llama_ubatch & ubatch) const {
         head_cur = 0;
     }
 
-    // otherwise, one cell per token.
-
     if (n_tokens > cells.size()) {
         LLAMA_LOG_ERROR("%s: n_tokens = %d > size = %u\n", __func__, n_tokens, cells.size());
-        return -1;
+        return { };
     }
 
-//#define FIND_SLOT_DEBUG 1
-#if FIND_SLOT_DEBUG
-    LLAMA_LOG_WARN("begin: n = %5d, used = %5d, head = %5d, n_swa = %5d\n", cells.used_max_p1(), cells.get_used(), head, n_swa);
+    if (debug > 0) {
+        LLAMA_LOG_DEBUG("%s: n = %5d, used = %5d, head = %5d, size = %5d, n_swa = %5d\n", __func__, cells.used_max_p1(), cells.get_used(), head, get_size(), n_swa);
 
-    // for debugging
-    {
-        std::string ss;
-        if (n_swa > 0) {
+        if ((debug == 2 && n_swa > 0) || debug > 2) {
+            std::string ss;
             for (uint32_t i = 0; i < cells.size(); ++i) {
                 if (cells.is_empty(i)) {
                     ss += '.';
                 } else {
-                    ss += std::to_string(cells.seq_get(i));
+                    assert(cells.seq_count(i) >= 1);
+
+                    if (cells.seq_count(i) == 1) {
+                        ss += std::to_string(cells.seq_get(i));
+                    } else {
+                        ss += 'M';
+                    }
                 }
                 if (i%256 == 255) {
+                    ss += " *";
                     ss += '\n';
                 }
             }
-        }
-        LLAMA_LOG_WARN("\n%s\n", ss.c_str());
-    }
-
-    for (int s = 0; s < LLAMA_MAX_PARALLEL_SEQUENCES; ++s) {
-        if (cells.seq_pos_min(s) < 0) {
-            continue;
+            LLAMA_LOG_DEBUG("\n%s\n", ss.c_str());
         }
 
-        LLAMA_LOG_WARN("kv_cells: n_swa = %4d, min[%d] = %5d, max[%d] = %5d\n", n_swa, s, cells.seq_pos_min(s), s, cells.seq_pos_max(s));
+        if ((debug == 2 && n_swa > 0) || debug > 2) {
+            std::string ss;
+            for (uint32_t i = 0; i < cells.size(); ++i) {
+                std::string cur;
+                if (cells.is_empty(i)) {
+                    cur = '.';
+                } else {
+                    cur = std::to_string(cells.pos_get(i));
+                }
+                const int n = cur.size();
+                for (int j = 0; j < 5 - n; ++j) {
+                    cur += ' ';
+                }
+                ss += cur;
+                if (i%256 == 255) {
+                    ss += " *";
+                }
+                if (i%64 == 63) {
+                    ss += '\n';
+                }
+            }
+            LLAMA_LOG_DEBUG("\n%s\n", ss.c_str());
+        }
+
+        for (int s = 0; s < LLAMA_MAX_SEQ; ++s) {
+            if (cells.seq_pos_min(s) < 0) {
+                continue;
+            }
+
+            LLAMA_LOG_DEBUG("%s: min[%d] = %5d, max[%d] = %5d\n", __func__, s, cells.seq_pos_min(s), s, cells.seq_pos_max(s));
+        }
     }
-#endif
 
     uint32_t n_tested = 0;
 
+    // for continuous slots, we test that all tokens in the ubatch fit, starting from the current head
+    // for non-continuous slots, we test the tokens one by one
+    const uint32_t n_test = cont ? n_tokens : 1;
+
+    slot_info res;
+
+    auto & idxs = res.idxs;
+
+    idxs.reserve(n_tokens);
+
     while (true) {
-        if (head_cur + n_tokens > cells.size()) {
+        if (head_cur + n_test > cells.size()) {
             n_tested += cells.size() - head_cur;
             head_cur = 0;
             continue;
         }
 
-        // keep track of what the minimum sequence positions would be if we accept the ubatch
-        llama_seq_id seq_pos_min[LLAMA_MAX_PARALLEL_SEQUENCES];
-        for (int s = 0; s < LLAMA_MAX_PARALLEL_SEQUENCES; ++s) {
-            seq_pos_min[s] = cells.seq_pos_min(s);
-        }
+        for (uint32_t i = 0; i < n_test; i++) {
+            const auto idx = head_cur;
 
-        bool found = true;
-        for (uint32_t i = 0; i < n_tokens; i++) {
-            const llama_pos    pos    = ubatch.pos[i];
-            const llama_seq_id seq_id = ubatch.seq_id[i][0];
+            //const llama_pos    pos    = ubatch.pos[i];
+            //const llama_seq_id seq_id = ubatch.seq_id[i][0];
 
             // can we use this cell? either:
             //  - the cell is empty
             //  - the cell is occupied only by one sequence:
-            //    - mask causally, if the sequence is the same as the one we are inserting
+            //    - (disabled) mask causally, if the sequence is the same as the one we are inserting
             //    - mask SWA, using current max pos for that sequence in the cache
             //                always insert in the cell with minimum pos
-            bool can_use = cells.is_empty(head_cur + i);
+            bool can_use = cells.is_empty(idx);
 
-            if (!can_use && cells.seq_count(head_cur + i) == 1) {
-                const llama_pos pos_cell = cells.pos_get(head_cur + i);
+            if (!can_use && cells.seq_count(idx) == 1) {
+                const llama_pos pos_cell = cells.pos_get(idx);
 
-                // causal mask
-                if (cells.seq_has(head_cur + i, seq_id)) {
-                    can_use = pos_cell >= pos;
-                }
+                // (disabled) causal mask
+                // note: it's better to purge any "future" tokens beforehand
+                //if (cells.seq_has(idx, seq_id)) {
+                //    can_use = pos_cell >= pos;
+                //}
 
                 if (!can_use) {
-                    const llama_seq_id seq_id_cell = cells.seq_get(head_cur + i);
+                    const llama_seq_id seq_id_cell = cells.seq_get(idx);
 
                     // SWA mask
-                    // note: we insert only in the cell with minimum pos in order to preserve the invariant that
-                    //       all positions between [pos_min, pos_max] for each sequence will be present in the cache
-                    //       ref: https://github.com/ggml-org/llama.cpp/pull/13746#issuecomment-2916057092
-                    if (pos_cell == seq_pos_min[seq_id_cell] &&
-                        is_masked_swa(pos_cell, cells.seq_pos_max(seq_id_cell) + 1)) {
-                        seq_pos_min[seq_id_cell]++;
+                    if (is_masked_swa(pos_cell, cells.seq_pos_max(seq_id_cell) + 1)) {
                         can_use = true;
                     }
                 }
             }
 
-            if (!can_use) {
-                found = false;
-                head_cur += i + 1;
-                n_tested += i + 1;
+            head_cur++;
+            n_tested++;
+
+            if (can_use) {
+                idxs.push_back(idx);
+            } else {
                 break;
             }
         }
 
-        if (found) {
+        if (idxs.size() == n_tokens) {
             break;
         }
 
+        if (cont) {
+            idxs.clear();
+        }
+
         if (n_tested >= cells.size()) {
             //LLAMA_LOG_ERROR("%s: failed to find a slot for %d tokens\n", __func__, n_tokens);
-            return -1;
+            return { };
         }
     }
 
-    return head_cur;
+    // we didn't find a suitable slot - return empty result
+    if (idxs.size() < n_tokens) {
+        res.clear();
+    }
+
+    return res;
 }
 
-void llama_kv_cache_unified::apply_ubatch(uint32_t head_cur, const llama_ubatch & ubatch) {
+void llama_kv_cache_unified::apply_ubatch(const slot_info & sinfo, const llama_ubatch & ubatch) {
+    // keep track of the max sequence position that we would overwrite with this ubatch
+    // for non-SWA cache, this would be always empty
+    llama_seq_id seq_pos_max_rm[LLAMA_MAX_SEQ];
+    for (int s = 0; s < LLAMA_MAX_SEQ; ++s) {
+        seq_pos_max_rm[s] = -1;
+    }
+
+    assert(ubatch.n_tokens == sinfo.idxs.size());
+
     for (uint32_t i = 0; i < ubatch.n_tokens; ++i) {
-        if (!cells.is_empty(head_cur + i)) {
-            cells.rm(head_cur + i);
+        const auto idx = sinfo.idxs.at(i);
+
+        if (!cells.is_empty(idx)) {
+            assert(cells.seq_count(idx) == 1);
+
+            const llama_seq_id seq_id = cells.seq_get(idx);
+            const llama_pos    pos    = cells.pos_get(idx);
+
+            seq_pos_max_rm[seq_id] = std::max(seq_pos_max_rm[seq_id], pos);
+
+            cells.rm(idx);
         }
 
-        cells.pos_set(head_cur + i, ubatch.pos[i]);
+        cells.pos_set(idx, ubatch.pos[i]);
 
-        for (int32_t j = 0; j < ubatch.n_seq_id[i]; j++) {
-            cells.seq_add(head_cur + i, ubatch.seq_id[i][j]);
+        for (int32_t s = 0; s < ubatch.n_seq_id[i]; s++) {
+            cells.seq_add(idx, ubatch.seq_id[i][s]);
+        }
+    }
+
+    // note: we want to preserve the invariant that all positions between [pos_min, pos_max] for each sequence
+    //       will be present in the cache. so we have to purge any position which is less than those we would overwrite
+    //       ref: https://github.com/ggml-org/llama.cpp/pull/13746#issuecomment-2916057092
+    for (int s = 0; s < LLAMA_MAX_SEQ; ++s) {
+        if (seq_pos_max_rm[s] == -1) {
+            continue;
+        }
+
+        if (cells.seq_pos_min(s) <= seq_pos_max_rm[s]) {
+            LLAMA_LOG_DEBUG("%s: purging positions [%d, %d] of sequence %d from KV cache\n",
+                    __func__, cells.seq_pos_min(s), seq_pos_max_rm[s], s);
+
+            seq_rm(s, cells.seq_pos_min(s), seq_pos_max_rm[s] + 1);
         }
     }
 
     // move the head at the end of the slot
-    head = head_cur + ubatch.n_tokens;
+    head = sinfo.idxs.back() + 1;
 }
 
 bool llama_kv_cache_unified::get_can_shift() const {
@@ -597,6 +772,10 @@ uint32_t llama_kv_cache_unified::get_size() const {
     return cells.size();
 }
 
+bool llama_kv_cache_unified::get_has_shift() const {
+    return cells.get_has_shift();
+}
+
 uint32_t llama_kv_cache_unified::get_n_kv() const {
     return std::min(cells.size(), std::max(n_pad, GGML_PAD(cells.used_max_p1(), n_pad)));
 }
@@ -635,56 +814,140 @@ ggml_tensor * llama_kv_cache_unified::get_v(ggml_context * ctx, int32_t il, uint
             0);
 }
 
-ggml_tensor * llama_kv_cache_unified::cpy_k(ggml_context * ctx, ggml_tensor * k_cur, int32_t il, uint32_t head_cur) const {
+ggml_tensor * llama_kv_cache_unified::cpy_k(ggml_context * ctx, ggml_tensor * k_cur, ggml_tensor * k_idxs, int32_t il, const slot_info & sinfo) const {
     const int32_t ikv = map_layer_ids.at(il);
 
     auto * k = layers[ikv].k;
 
+    const int64_t n_embd_k_gqa = k->ne[0];
     const int64_t n_tokens = k_cur->ne[2];
 
+    k_cur = ggml_reshape_2d(ctx, k_cur, k->ne[0], n_tokens);
+
+    if (k_idxs && supports_set_rows) {
+        return ggml_set_rows(ctx, k, k_cur, k_idxs);
+    }
+
+    // TODO: fallback to old ggml_cpy() method for backwards compatibility
+    //       will be removed when ggml_set_rows() is adopted by all backends
+
     ggml_tensor * k_view = ggml_view_1d(ctx, k,
-            n_tokens*hparams.n_embd_k_gqa(il),
-            ggml_row_size(k->type, hparams.n_embd_k_gqa(il))*head_cur);
+            n_tokens*n_embd_k_gqa,
+            ggml_row_size(k->type, n_embd_k_gqa)*sinfo.head());
 
     return ggml_cpy(ctx, k_cur, k_view);
 }
 
-ggml_tensor * llama_kv_cache_unified::cpy_v(ggml_context * ctx, ggml_tensor * v_cur, int32_t il, uint32_t head_cur) const {
+ggml_tensor * llama_kv_cache_unified::cpy_v(ggml_context * ctx, ggml_tensor * v_cur, ggml_tensor * v_idxs, int32_t il, const slot_info & sinfo) const {
     const int32_t ikv = map_layer_ids.at(il);
 
     auto * v = layers[ikv].v;
 
+    const int64_t n_embd_v_gqa = v->ne[0];
     const int64_t n_tokens = v_cur->ne[2];
 
-    v_cur = ggml_reshape_2d(ctx, v_cur, hparams.n_embd_v_gqa(il), n_tokens);
+    v_cur = ggml_reshape_2d(ctx, v_cur, n_embd_v_gqa, n_tokens);
+
+    if (v_idxs && supports_set_rows) {
+        if (!v_trans) {
+            return ggml_set_rows(ctx, v, v_cur, v_idxs);
+        }
+
+        // the row becomes a single element
+        ggml_tensor * v_view = ggml_reshape_3d(ctx, v, 1, v->ne[1], v->ne[0]);
+
+        // note: the V cache is transposed when not using flash attention
+        v_cur = ggml_permute(ctx, ggml_reshape_3d(ctx, v_cur, v_cur->ne[0], 1, v_cur->ne[1]), 2, 0, 1, 3);
+
+        // note: we can be more explicit here at the cost of extra cont
+        //       however, above we take advantage that a row of single element is always continuous regardless of the row stride
+        //v_cur = ggml_transpose(ctx, v_cur);
+        //v_cur = ggml_cont_3d(ctx, v_cur, 1, v_cur->ne[0], v_cur->ne[1]);
+
+        // we broadcast the KV indices n_embd_v_gqa times
+        // v      [1,        n_kv,     n_embd_v_gqa]
+        // v_cur  [1,        n_tokens, n_embd_v_gqa]
+        // v_idxs [n_tokens, 1,        1]
+        return ggml_set_rows(ctx, v_view, v_cur, v_idxs);
+    }
+
+    // TODO: fallback to old ggml_cpy() method for backwards compatibility
+    //       will be removed when ggml_set_rows() is adopted by all backends
 
     ggml_tensor * v_view = nullptr;
 
     if (!v_trans) {
         v_view = ggml_view_1d(ctx, v,
-                n_tokens*hparams.n_embd_v_gqa(il),
-                ggml_row_size(v->type, hparams.n_embd_v_gqa(il))*head_cur);
+                n_tokens*n_embd_v_gqa,
+                ggml_row_size(v->type, n_embd_v_gqa)*sinfo.head());
     } else {
-        // note: the V cache is transposed when not using flash attention
-        v_view = ggml_view_2d(ctx, v, n_tokens, hparams.n_embd_v_gqa(il),
-                (v->ne[1])*ggml_element_size(v),
-                (head_cur)*ggml_element_size(v));
-
         v_cur = ggml_transpose(ctx, v_cur);
+
+        v_view = ggml_view_2d(ctx, v, n_tokens, n_embd_v_gqa,
+                (v->ne[1]    )*ggml_element_size(v),
+                (sinfo.head())*ggml_element_size(v));
     }
 
     return ggml_cpy(ctx, v_cur, v_view);
 }
 
+ggml_tensor * llama_kv_cache_unified::build_input_k_idxs(ggml_context * ctx, const llama_ubatch & ubatch) const {
+    const uint32_t n_tokens = ubatch.n_tokens;
+
+    ggml_tensor * k_idxs = ggml_new_tensor_1d(ctx, GGML_TYPE_I64, n_tokens);
+
+    ggml_set_input(k_idxs);
+
+    return k_idxs;
+}
+
+ggml_tensor * llama_kv_cache_unified::build_input_v_idxs(ggml_context * ctx, const llama_ubatch & ubatch) const {
+    const uint32_t n_tokens = ubatch.n_tokens;
+
+    ggml_tensor * v_idxs = ggml_new_tensor_1d(ctx, GGML_TYPE_I64, n_tokens);
+
+    ggml_set_input(v_idxs);
+
+    return v_idxs;
+}
+
+void llama_kv_cache_unified::set_input_k_idxs(ggml_tensor * dst, const llama_ubatch * ubatch, const slot_info & sinfo) const {
+    if (!supports_set_rows) {
+        return;
+    }
+
+    const uint32_t n_tokens = ubatch->n_tokens;
+
+    GGML_ASSERT(ggml_backend_buffer_is_host(dst->buffer));
+    int64_t * data = (int64_t *) dst->data;
+
+    for (int64_t i = 0; i < n_tokens; ++i) {
+        data[i] = sinfo.idxs.at(i);
+    }
+}
+
+void llama_kv_cache_unified::set_input_v_idxs(ggml_tensor * dst, const llama_ubatch * ubatch, const slot_info & sinfo) const {
+    if (!supports_set_rows) {
+        return;
+    }
+
+    const uint32_t n_tokens = ubatch->n_tokens;
+
+    GGML_ASSERT(ggml_backend_buffer_is_host(dst->buffer));
+    int64_t * data = (int64_t *) dst->data;
+
+    for (int64_t i = 0; i < n_tokens; ++i) {
+        data[i] = sinfo.idxs.at(i);
+    }
+}
+
 void llama_kv_cache_unified::set_input_kq_mask(ggml_tensor * dst, const llama_ubatch * ubatch, bool causal_attn) const {
-    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 uint32_t n_tokens = ubatch->n_tokens;
 
     GGML_ASSERT(ggml_backend_buffer_is_host(dst->buffer));
     float * data = (float *) dst->data;
 
-    const auto n_kv = dst->ne[0];
+    const int64_t n_kv = dst->ne[0];
 
     // 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.
@@ -698,49 +961,47 @@ void llama_kv_cache_unified::set_input_kq_mask(ggml_tensor * dst, const llama_ub
     //      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 (uint32_t h = 0; h < 1; ++h) {
+        for (uint32_t i = 0; i < n_tokens; ++i) {
+            const llama_seq_id seq_id = ubatch->seq_id[i][0];
 
-            for (int j = 0; j < n_seq_tokens; ++j) {
-                const llama_pos p1 = ubatch->pos[s*n_seq_tokens + j];
+            const llama_pos p1 = ubatch->pos[i];
 
-                for (uint32_t i = 0; i < n_kv; ++i) {
-                    float f = 0.0f;
+            for (uint32_t j = 0; j < n_kv; ++j) {
+                float f = 0.0f;
 
-                    bool masked = false;
+                bool masked = false;
 
-                    if (cells.is_empty(i)) {
-                        masked = true;
-                    } else {
-                        const llama_pos p0 = cells.pos_get(i);
+                if (cells.is_empty(j)) {
+                    masked = true;
+                } else {
+                    const llama_pos p0 = cells.pos_get(j);
 
-                        // mask the token if not the same sequence
-                        masked = masked || (!cells.seq_has(i, seq_id));
+                    // mask the token if not the same sequence
+                    masked = masked || (!cells.seq_has(j, seq_id));
 
-                        // mask future tokens
-                        masked = masked || (causal_attn && p0 > p1);
+                    // mask future tokens
+                    masked = masked || (causal_attn && p0 > p1);
 
-                        // apply SWA if any
-                        masked = masked || (is_masked_swa(p0, p1));
+                    // apply SWA if any
+                    masked = masked || (is_masked_swa(p0, p1));
 
-                        if (!masked && hparams.use_alibi) {
-                            f = -std::abs(p0 - p1);
-                        }
+                    if (!masked && hparams.use_alibi) {
+                        f = -std::abs(p0 - p1);
                     }
-
-                    if (masked) {
-                        f = -INFINITY;
-                    }
-
-                    data[h*(n_kv*n_tokens) + s*(n_kv*n_seq_tokens) + j*n_kv + i] = f;
                 }
+
+                if (masked) {
+                    f = -INFINITY;
+                }
+
+                data[h*(n_kv*n_tokens) + i*n_kv + j] = f;
             }
         }
 
         // mask padded tokens
         if (data) {
-            for (int i = n_tokens; i < GGML_PAD(n_tokens, GGML_KQ_MASK_PAD); ++i) {
+            for (uint32_t i = n_tokens; i < GGML_PAD(n_tokens, GGML_KQ_MASK_PAD); ++i) {
                 for (uint32_t j = 0; j < n_kv; ++j) {
                     data[h*(n_kv*n_tokens) + i*n_kv + j] = -INFINITY;
                 }
@@ -770,12 +1031,12 @@ void llama_kv_cache_unified::set_input_pos_bucket(ggml_tensor * dst, const llama
     const int32_t n_kv = dst->ne[0];
 
     for (int h = 0; h < 1; ++h) {
-        for (int j = 0; j < n_tokens; ++j) {
-            for (int i = 0; i < n_kv; ++i) {
+        for (int i = 0; i < n_tokens; ++i) {
+            for (int j = 0; j < n_kv; ++j) {
                 // the position when the cells is empty is irrelevant - it will be masked out later in the attention
-                const llama_pos p0 = cells.is_empty(i) ? -1 : cells.pos_get(i);
+                const llama_pos p0 = cells.is_empty(j) ? -1 : cells.pos_get(j);
 
-                data[h*(n_kv*n_tokens) + j*n_kv + i] = llama_relative_position_bucket(p0, ubatch->pos[j], hparams.n_rel_attn_bkts, false);
+                data[h*(n_kv*n_tokens) + i*n_kv + j] = llama_relative_position_bucket(p0, ubatch->pos[i], hparams.n_rel_attn_bkts, false);
             }
         }
     }
@@ -890,11 +1151,9 @@ llm_graph_result_ptr llama_kv_cache_unified::build_graph_shift(
     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<llm_graph_input_k_shift>(this);
 
-    inp->k_shift = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, cparams.n_ctx);
+    inp->k_shift = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, cells.size());
     ggml_set_input(inp->k_shift);
 
     for (const auto & layer : layers) {
@@ -926,12 +1185,13 @@ llm_graph_result_ptr llama_kv_cache_unified::build_graph_shift(
 }
 
 llm_graph_result_ptr llama_kv_cache_unified::build_graph_defrag(
-        const llama_cparams & cparams,
-               ggml_context * ctx,
-                ggml_cgraph * gf) const {
+                const llama_cparams & cparams,
+                       ggml_context * ctx,
+                        ggml_cgraph * gf,
+                  const defrag_info & dinfo) const {
     auto res = std::make_unique<llm_graph_result>();
 
-    const auto & ids = defrag_info.ids;
+    const auto & ids = dinfo.ids;
 
 #if 0
     // CPU defrag
@@ -1072,7 +1332,7 @@ llm_graph_result_ptr llama_kv_cache_unified::build_graph_defrag(
     return res;
 }
 
-bool llama_kv_cache_unified::defrag_prepare(int32_t n_max_nodes) {
+llama_kv_cache_unified::defrag_info llama_kv_cache_unified::defrag_prepare(int32_t n_max_nodes) const {
     const uint32_t n_layer = layers.size();
 
     const uint32_t n_kv   = cells.used_max_p1();
@@ -1093,14 +1353,9 @@ bool llama_kv_cache_unified::defrag_prepare(int32_t n_max_nodes) {
     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;
+    defrag_info res;
+    auto & ids = res.ids;
 
-    ids.clear();
     ids.resize(n_kv, n_kv);
 
     for (uint32_t i0 = 0; i0 < n_used; ++i0) {
@@ -1164,11 +1419,6 @@ bool llama_kv_cache_unified::defrag_prepare(int32_t n_max_nodes) {
             // this cell goes to (i0 + nf)
             ids[i1] = i0 + nf;
 
-            // move the cell meta data
-            cells.mv(i1, i0 + nf);
-
-            head = n_used;
-
             if (!cont) {
                 n_moves++;
                 cont = true;
@@ -1191,14 +1441,14 @@ bool llama_kv_cache_unified::defrag_prepare(int32_t n_max_nodes) {
     }
 
     if (n_moves == 0) {
-        return false;
+        return {};
     }
 
     LLAMA_LOG_DEBUG("%s: (tmp log) KV defrag cell moves: %u\n", __func__, n_moves);
 
     LLAMA_LOG_DEBUG("%s: expected gf nodes: %u\n", __func__, 6*n_moves*n_layer);
 
-    return true;
+    return res;
 }
 
 bool llama_kv_cache_unified::is_masked_swa(llama_pos p0, llama_pos p1) const {
@@ -1276,7 +1526,7 @@ void llama_kv_cache_unified::state_read(llama_io_read_i & io, llama_seq_id seq_i
 
     if (!res) {
         if (seq_id == -1) {
-            clear();
+            clear(true);
         } else {
             seq_rm(seq_id, -1, -1);
         }
@@ -1324,7 +1574,7 @@ void llama_kv_cache_unified::state_write_data(llama_io_write_i & io, const std::
     for (const auto & layer : layers) {
         const uint32_t il = layer.il;
 
-        const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa(il) + hparams.n_embd_k_s();
+        const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa(il);
 
         // Write key type
         const int32_t k_type_i = (int32_t)layer.k->type;
@@ -1346,7 +1596,7 @@ void llama_kv_cache_unified::state_write_data(llama_io_write_i & io, const std::
         for (const auto & layer : layers) {
             const uint32_t il = layer.il;
 
-            const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il) + hparams.n_embd_v_s();
+            const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il);
 
             // Write value type
             const int32_t v_type_i = (int32_t)layer.v->type;
@@ -1370,7 +1620,7 @@ void llama_kv_cache_unified::state_write_data(llama_io_write_i & io, const std::
         for (const auto & layer : layers) {
             const uint32_t il = layer.il;
 
-            const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il) + hparams.n_embd_v_s();
+            const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il);
 
             // Write value type
             const int32_t v_type_i = (int32_t)layer.v->type;
@@ -1403,10 +1653,9 @@ bool llama_kv_cache_unified::state_read_meta(llama_io_read_i & io, uint32_t cell
 
         seq_rm(dest_seq_id, -1, -1);
 
-        llama_sbatch sbatch;
-        llama_ubatch batch = sbatch.reserve_ubatch(cell_count, /* has_embd */ false);
+        llama_batch_allocr balloc(hparams.n_pos_per_embd());
 
-        batch.n_tokens = cell_count;
+        llama_ubatch ubatch = balloc.ubatch_reserve(cell_count, 1);
 
         for (uint32_t i = 0; i < cell_count; ++i) {
             llama_pos pos;
@@ -1426,18 +1675,20 @@ bool llama_kv_cache_unified::state_read_meta(llama_io_read_i & io, uint32_t cell
                 io.read_to(&seq_id, sizeof(seq_id));
             }
 
-            batch.pos[i]      = pos;
-            batch.n_seq_id[i] = n_seq_id;
-            batch.seq_id[i]   = &dest_seq_id;
+            ubatch.pos[i]      = pos;
+            ubatch.n_seq_id[i] = n_seq_id;
+            ubatch.seq_id[i]   = &dest_seq_id;
         }
 
-        const auto head_cur = find_slot(batch);
-        if (head_cur < 0) {
+        const auto sinfo = find_slot(ubatch, true);
+        if (sinfo.empty()) {
             LLAMA_LOG_ERROR("%s: failed to find available cells in kv cache\n", __func__);
             return false;
         }
 
-        apply_ubatch(head_cur, batch);
+        apply_ubatch(sinfo, ubatch);
+
+        const auto head_cur = sinfo.head();
 
         // keep the head at the old position because we will read the KV data into it in state_read_data()
         head = head_cur;
@@ -1445,8 +1696,8 @@ bool llama_kv_cache_unified::state_read_meta(llama_io_read_i & io, uint32_t cell
         // DEBUG CHECK: head_cur should be our first cell, head_cur + 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_cur + cell_count <= cells.size());
-        GGML_ASSERT(cells.pos_get(head_cur)                  == batch.pos[0]);
-        GGML_ASSERT(cells.pos_get(head_cur + cell_count - 1) == batch.pos[cell_count - 1]);
+        GGML_ASSERT(cells.pos_get(head_cur)                  == ubatch.pos[0]);
+        GGML_ASSERT(cells.pos_get(head_cur + cell_count - 1) == ubatch.pos[cell_count - 1]);
         GGML_ASSERT(cells.seq_has(head_cur,                  dest_seq_id));
         GGML_ASSERT(cells.seq_has(head_cur + cell_count - 1, dest_seq_id));
     } else {
@@ -1457,7 +1708,7 @@ bool llama_kv_cache_unified::state_read_meta(llama_io_read_i & io, uint32_t cell
             return false;
         }
 
-        clear();
+        clear(true);
 
         for (uint32_t i = 0; i < cell_count; ++i) {
             llama_pos pos;
@@ -1513,7 +1764,7 @@ bool llama_kv_cache_unified::state_read_data(llama_io_read_i & io, uint32_t cell
     for (const auto & layer : layers) {
         const uint32_t il = layer.il;
 
-        const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa(il) + hparams.n_embd_k_s();
+        const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa(il);
 
         // Read type of key
         int32_t k_type_i_ref;
@@ -1543,7 +1794,7 @@ bool llama_kv_cache_unified::state_read_data(llama_io_read_i & io, uint32_t cell
         for (const auto & layer : layers) {
             const uint32_t il = layer.il;
 
-            const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il) + hparams.n_embd_v_s();
+            const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il);
 
             // Read type of value
             int32_t v_type_i_ref;
@@ -1573,7 +1824,7 @@ bool llama_kv_cache_unified::state_read_data(llama_io_read_i & io, uint32_t cell
         for (const auto & layer : layers) {
             const uint32_t il = layer.il;
 
-            const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il) + hparams.n_embd_v_s();
+            const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(il);
 
             // Read type of value
             int32_t v_type_i_ref;
@@ -1615,99 +1866,121 @@ bool llama_kv_cache_unified::state_read_data(llama_io_read_i & io, uint32_t cell
 }
 
 //
-// llama_kv_cache_unified_state
+// llama_kv_cache_unified_context
 //
 
-llama_kv_cache_unified_state::llama_kv_cache_unified_state(llama_memory_status status) : status(status) {}
+llama_kv_cache_unified_context::llama_kv_cache_unified_context(llama_memory_status status) : status(status) {}
 
-llama_kv_cache_unified_state::llama_kv_cache_unified_state(
-            llama_memory_status status,
-            llama_kv_cache_unified * kv) : status(status), kv(kv) {
-        n_kv = kv->get_size();
-        head = 0;
+llama_kv_cache_unified_context::llama_kv_cache_unified_context(
+        llama_kv_cache_unified * kv) : status(LLAMA_MEMORY_STATUS_SUCCESS), kv(kv) {
+    n_kv = kv->get_size();
+
+    // create a dummy slot info - the actual data is irrelevant. we just need to build the graph
+    sinfos.resize(1);
+    sinfos[0].idxs.resize(1);
+    sinfos[0].idxs[0] = 0;
+}
+
+llama_kv_cache_unified_context::llama_kv_cache_unified_context(
+        llama_kv_cache_unified * kv,
+        llama_context * lctx,
+        bool do_shift,
+        defrag_info dinfo) : status(LLAMA_MEMORY_STATUS_SUCCESS), kv(kv), lctx(lctx), do_shift(do_shift), dinfo(std::move(dinfo)) {
+    if (!do_shift && this->dinfo.empty()) {
+        status = LLAMA_MEMORY_STATUS_NO_UPDATE;
     }
+}
 
-llama_kv_cache_unified_state::llama_kv_cache_unified_state(
-            llama_memory_status status,
-            llama_kv_cache_unified * kv,
-            llama_sbatch sbatch,
-            std::vector<uint32_t> heads,
-            std::vector<llama_ubatch> ubatches)
-            : status(status),
-              kv(kv),
-              sbatch(std::move(sbatch)),
-              heads(std::move(heads)),
-              ubatches(std::move(ubatches)) {
-    }
+llama_kv_cache_unified_context::llama_kv_cache_unified_context(
+        llama_kv_cache_unified * kv,
+        llama_kv_cache_unified::slot_info_vec_t sinfos,
+        std::vector<llama_ubatch> ubatches) : status(LLAMA_MEMORY_STATUS_SUCCESS), kv(kv), sinfos(std::move(sinfos)), ubatches(std::move(ubatches)) {
+}
 
-llama_kv_cache_unified_state::~llama_kv_cache_unified_state() = default;
+llama_kv_cache_unified_context::~llama_kv_cache_unified_context() = default;
 
-bool llama_kv_cache_unified_state::next() {
+bool llama_kv_cache_unified_context::next() {
     assert(status == LLAMA_MEMORY_STATUS_SUCCESS);
 
-    if (++i_next >= ubatches.size()) {
+    if (++i_cur >= ubatches.size()) {
         return false;
     }
 
     return true;
 }
 
-bool llama_kv_cache_unified_state::apply() {
-    assert(status == LLAMA_MEMORY_STATUS_SUCCESS);
+bool llama_kv_cache_unified_context::apply() {
+    assert(!llama_memory_status_is_fail(status));
 
-    kv->apply_ubatch(heads[i_next], ubatches[i_next]);
+    // no ubatches -> this is a KV cache update
+    if (ubatches.empty()) {
+        kv->update(lctx, do_shift, dinfo);
+
+        return true;
+    }
+
+    kv->apply_ubatch(sinfos[i_cur], ubatches[i_cur]);
 
     n_kv = kv->get_n_kv();
-    head = heads[i_next];
 
     return true;
 }
 
-std::vector<int64_t> & llama_kv_cache_unified_state::out_ids() {
-    assert(status == LLAMA_MEMORY_STATUS_SUCCESS);
-
-    return sbatch.out_ids;
-}
-
-llama_memory_status llama_kv_cache_unified_state::get_status() const {
+llama_memory_status llama_kv_cache_unified_context::get_status() const {
     return status;
 }
 
-const llama_ubatch & llama_kv_cache_unified_state::get_ubatch() const {
+const llama_ubatch & llama_kv_cache_unified_context::get_ubatch() const {
     assert(status == LLAMA_MEMORY_STATUS_SUCCESS);
 
-    return ubatches[i_next];
+    return ubatches[i_cur];
 }
 
-uint32_t llama_kv_cache_unified_state::get_n_kv() const {
+uint32_t llama_kv_cache_unified_context::get_n_kv() const {
     return n_kv;
 }
 
-ggml_tensor * llama_kv_cache_unified_state::get_k(ggml_context * ctx, int32_t il) const {
+ggml_tensor * llama_kv_cache_unified_context::get_k(ggml_context * ctx, int32_t il) const {
     return kv->get_k(ctx, il, n_kv);
 }
 
-ggml_tensor * llama_kv_cache_unified_state::get_v(ggml_context * ctx, int32_t il) const {
+ggml_tensor * llama_kv_cache_unified_context::get_v(ggml_context * ctx, int32_t il) const {
     return kv->get_v(ctx, il, n_kv);
 }
 
-ggml_tensor * llama_kv_cache_unified_state::cpy_k(ggml_context * ctx, ggml_tensor * k_cur, int32_t il) const {
-    return kv->cpy_k(ctx, k_cur, il, head);
+ggml_tensor * llama_kv_cache_unified_context::cpy_k(ggml_context * ctx, ggml_tensor * k_cur, ggml_tensor * k_idxs, int32_t il) const {
+    return kv->cpy_k(ctx, k_cur, k_idxs, il, sinfos[i_cur]);
 }
 
-ggml_tensor * llama_kv_cache_unified_state::cpy_v(ggml_context * ctx, ggml_tensor * v_cur, int32_t il) const {
-    return kv->cpy_v(ctx, v_cur, il, head);
+ggml_tensor * llama_kv_cache_unified_context::cpy_v(ggml_context * ctx, ggml_tensor * v_cur, ggml_tensor * v_idxs, int32_t il) const {
+    return kv->cpy_v(ctx, v_cur, v_idxs, il, sinfos[i_cur]);
 }
 
-void llama_kv_cache_unified_state::set_input_k_shift(ggml_tensor * dst) const {
+ggml_tensor * llama_kv_cache_unified_context::build_input_k_idxs(ggml_context * ctx, const llama_ubatch & ubatch) const {
+    return kv->build_input_k_idxs(ctx, ubatch);
+}
+
+ggml_tensor * llama_kv_cache_unified_context::build_input_v_idxs(ggml_context * ctx, const llama_ubatch & ubatch) const {
+    return kv->build_input_v_idxs(ctx, ubatch);
+}
+
+void llama_kv_cache_unified_context::set_input_k_shift(ggml_tensor * dst) const {
     kv->set_input_k_shift(dst);
 }
 
-void llama_kv_cache_unified_state::set_input_kq_mask(ggml_tensor * dst, const llama_ubatch * ubatch, bool causal_attn) const {
+void llama_kv_cache_unified_context::set_input_k_idxs(ggml_tensor * dst, const llama_ubatch * ubatch) const {
+    kv->set_input_k_idxs(dst, ubatch, sinfos[i_cur]);
+}
+
+void llama_kv_cache_unified_context::set_input_v_idxs(ggml_tensor * dst, const llama_ubatch * ubatch) const {
+    kv->set_input_v_idxs(dst, ubatch, sinfos[i_cur]);
+}
+
+void llama_kv_cache_unified_context::set_input_kq_mask(ggml_tensor * dst, const llama_ubatch * ubatch, bool causal_attn) const {
     kv->set_input_kq_mask(dst, ubatch, causal_attn);
 }
 
-void llama_kv_cache_unified_state::set_input_pos_bucket(ggml_tensor * dst, const llama_ubatch * ubatch) const {
+void llama_kv_cache_unified_context::set_input_pos_bucket(ggml_tensor * dst, const llama_ubatch * ubatch) const {
     kv->set_input_pos_bucket(dst, ubatch);
 }
 
diff --git a/src/llama-kv-cache-unified.h b/src/llama-kv-cache-unified.h
index 1f1d44b97..b8b0356e8 100644
--- a/src/llama-kv-cache-unified.h
+++ b/src/llama-kv-cache-unified.h
@@ -2,8 +2,8 @@
 
 #include "llama-batch.h"
 #include "llama-graph.h"
-#include "llama-kv-cache.h"
 #include "llama-kv-cells.h"
+#include "llama-memory.h"
 
 #include <unordered_map>
 #include <vector>
@@ -17,13 +17,50 @@ struct llama_context;
 // llama_kv_cache_unified
 //
 
-class llama_kv_cache_unified : public llama_kv_cache {
+class llama_kv_cache_unified : public llama_memory_i {
 public:
     static uint32_t get_padding(const llama_cparams & cparams);
 
     // this callback is used to filter out layers that should not be included in the cache
     using layer_filter_cb = std::function<bool(int32_t il)>;
 
+    struct defrag_info {
+        bool empty() const {
+            return ids.empty();
+        }
+
+        // contains information about which cell moves where:
+        //  - cell i moves to ids[i]
+        //  - if ids[i] == i || ids[i] == ids.size(), then cell i is not moved
+        std::vector<uint32_t> ids;
+    };
+
+    // for each ubatch, create a slot_info that contains information about where the ubatch should be inserted in the
+    //   KV cells. for example, cell indices for each token, such that: token[i] -> goes to cells[idxs[i]]
+    struct slot_info {
+        // data for ggml_set_rows
+        using idx_vec_t = std::vector<uint32_t>;
+
+        idx_vec_t idxs;
+
+        uint32_t head() const {
+            return idxs.at(0);
+        }
+
+        bool empty() const {
+            return idxs.empty();
+        }
+
+        void clear() {
+            idxs.clear();
+        }
+
+        // TODO: implement
+        //std::vector<idx_vec_t> seq_idxs;
+    };
+
+    using slot_info_vec_t = std::vector<slot_info>;
+
     llama_kv_cache_unified(
             const llama_model &  model,
               layer_filter_cb && filter,
@@ -43,7 +80,18 @@ public:
     // llama_memory_i
     //
 
-    void clear() override;
+    llama_memory_context_ptr init_batch(
+            llama_batch_allocr & balloc,
+            uint32_t n_ubatch,
+            bool embd_all) override;
+
+    llama_memory_context_ptr init_full() override;
+
+    llama_memory_context_ptr init_update(llama_context * lctx, bool optimize) override;
+
+    bool get_can_shift() const override;
+
+    void clear(bool data) 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;
@@ -54,24 +102,6 @@ public:
     llama_pos seq_pos_min(llama_seq_id seq_id) const override;
     llama_pos seq_pos_max(llama_seq_id seq_id) const override;
 
-    //
-    // llama_kv_cache
-    //
-
-    llama_memory_state_ptr init_batch(
-            const llama_batch & batch,
-            uint32_t n_ubatch,
-            bool embd_pooled,
-            bool logits_all) override;
-
-    llama_memory_state_ptr init_full() override;
-
-    bool update(llama_context & lctx) override;
-
-    void defrag_sched(float thold) override;
-
-    bool get_can_shift() const override;
-
     // state write/load
 
     void state_write(llama_io_write_i & io, llama_seq_id seq_id = -1) const override;
@@ -83,6 +113,8 @@ public:
 
     uint32_t get_size() const;
 
+    bool get_has_shift() const;
+
     //
     // graph_build API
     //
@@ -94,28 +126,37 @@ public:
     ggml_tensor * get_v(ggml_context * ctx, int32_t il, uint32_t n_kv) const;
 
     // store k_cur and v_cur in the cache based on the provided head location
-    ggml_tensor * cpy_k(ggml_context * ctx, ggml_tensor * k_cur, int32_t il, uint32_t head_cur) const;
-    ggml_tensor * cpy_v(ggml_context * ctx, ggml_tensor * v_cur, int32_t il, uint32_t head_cur) const;
+    ggml_tensor * cpy_k(ggml_context * ctx, ggml_tensor * k_cur, ggml_tensor * k_idxs, int32_t il, const slot_info & sinfo) const;
+    ggml_tensor * cpy_v(ggml_context * ctx, ggml_tensor * v_cur, ggml_tensor * v_idxs, int32_t il, const slot_info & sinfo) const;
 
     //
     // preparation API
     //
 
-    // find places for the provided ubatches in the cache, returns the head locations
+    // find places for the provided ubatches in the cache, returns the slot infos
     // return empty vector on failure
-    std::vector<uint32_t> prepare(const std::vector<llama_ubatch> & ubatches);
+    slot_info_vec_t prepare(const std::vector<llama_ubatch> & ubatches);
 
-    // return the cell position where we can insert the ubatch
-    // return -1 on failure to find a contiguous slot of kv cells
-    int32_t find_slot(const llama_ubatch & ubatch) const;
+    bool update(llama_context * lctx, bool do_shift, const defrag_info & dinfo);
 
-    // emplace the ubatch context into slot: [head_cur, head_cur + ubatch.n_tokens)
-    void apply_ubatch(uint32_t head_cur, const llama_ubatch & ubatch);
+    // find a slot of kv cells that can hold the ubatch
+    // if cont == true, then the slot must be continuous
+    // return empty slot_info on failure
+    slot_info find_slot(const llama_ubatch & ubatch, bool cont) const;
+
+    // emplace the ubatch context into slot: [sinfo.idxs[0...ubatch.n_tokens - 1]]
+    void apply_ubatch(const slot_info & sinfo, const llama_ubatch & ubatch);
 
     //
-    // set_input API
+    // input API
     //
 
+    ggml_tensor * build_input_k_idxs(ggml_context * ctx, const llama_ubatch & ubatch) const;
+    ggml_tensor * build_input_v_idxs(ggml_context * ctx, const llama_ubatch & ubatch) const;
+
+    void set_input_k_idxs(ggml_tensor * dst, const llama_ubatch * ubatch, const slot_info & sinfo) const;
+    void set_input_v_idxs(ggml_tensor * dst, const llama_ubatch * ubatch, const slot_info & sinfo) const;
+
     void set_input_kq_mask   (ggml_tensor * dst, const llama_ubatch * ubatch, bool causal_attn) const;
     void set_input_k_shift   (ggml_tensor * dst) const;
     void set_input_pos_bucket(ggml_tensor * dst, const llama_ubatch * ubatch) const;
@@ -133,8 +174,7 @@ private:
         ggml_tensor * v;
     };
 
-    bool do_defrag = false;
-    bool v_trans   = true;  // the value tensor is transposed
+    bool v_trans = true;  // the value tensor is transposed
 
     // the current index from where we start searching for a free slot in the ring buffer of KV cells (see find_slot())
     // note: this is not part of the KV state and it's only used to speed-up the find_slot() method
@@ -148,6 +188,13 @@ private:
     // SWA
     const uint32_t n_swa = 0;
 
+    // env: LLAMA_KV_CACHE_DEBUG
+    int debug = 0;
+
+    // env: LLAMA_SET_ROWS (temporary)
+    // ref: https://github.com/ggml-org/llama.cpp/pull/14285
+    int supports_set_rows = false;
+
     const llama_swa_type swa_type = LLAMA_SWA_TYPE_NONE;
 
     std::vector<ggml_context_ptr>        ctxs;
@@ -160,13 +207,8 @@ private:
     // model layer id -> KV cache layer id
     std::unordered_map<int32_t, int32_t> map_layer_ids;
 
-    // defrag
-    struct {
-        std::vector<uint32_t> ids;
-    } defrag_info;
-
-    // return true if cells have been moved
-    bool defrag_prepare(int32_t n_max_nodes);
+    // return non-empty vector if cells have been moved
+    defrag_info defrag_prepare(int32_t n_max_nodes) const;
 
     size_t total_size() const;
 
@@ -192,7 +234,8 @@ private:
     llm_graph_result_ptr build_graph_defrag(
             const llama_cparams & cparams,
                    ggml_context * ctx,
-                    ggml_cgraph * gf) const;
+                    ggml_cgraph * gf,
+              const defrag_info & dinfo) const;
 
     void state_write_meta(llama_io_write_i & io, const std::vector<std::pair<uint32_t, uint32_t>> & cell_ranges, llama_seq_id seq_id = -1) const;
     void state_write_data(llama_io_write_i & io, const std::vector<std::pair<uint32_t, uint32_t>> & cell_ranges) const;
@@ -201,40 +244,46 @@ private:
     bool state_read_data(llama_io_read_i & io, uint32_t cell_count);
 };
 
-class llama_kv_cache_unified_state : public llama_memory_state_i {
+class llama_kv_cache_unified_context : public llama_memory_context_i {
 public:
-    // used for errors
-    llama_kv_cache_unified_state(llama_memory_status status);
+    // some shorthands
+    using slot_info_vec_t = llama_kv_cache_unified::slot_info_vec_t;
+    using defrag_info     = llama_kv_cache_unified::defrag_info;
 
-    // used to create a full-cache state
-    llama_kv_cache_unified_state(
-            llama_memory_status status,
+    // used for errors
+    llama_kv_cache_unified_context(llama_memory_status status);
+
+    // used to create a full-cache context
+    llama_kv_cache_unified_context(
             llama_kv_cache_unified * kv);
 
-    // used to create a state from a batch
-    llama_kv_cache_unified_state(
-            llama_memory_status status,
+    // used to create an update context
+    llama_kv_cache_unified_context(
             llama_kv_cache_unified * kv,
-            llama_sbatch sbatch,
-            std::vector<uint32_t> heads,
+            llama_context * lctx,
+            bool do_shift,
+            defrag_info dinfo);
+
+    // used to create a batch procesing context from a batch
+    llama_kv_cache_unified_context(
+            llama_kv_cache_unified * kv,
+            slot_info_vec_t sinfos,
             std::vector<llama_ubatch> ubatches);
 
-    virtual ~llama_kv_cache_unified_state();
+    virtual ~llama_kv_cache_unified_context();
 
     //
-    // llama_memory_state_i
+    // llama_memory_context_i
     //
 
     bool next()  override;
     bool apply() override;
 
-    std::vector<int64_t> & out_ids() override;
-
     llama_memory_status  get_status() const override;
     const llama_ubatch & get_ubatch() const override;
 
     //
-    // llama_kv_cache_unified_state specific API
+    // llama_kv_cache_unified_context specific API
     //
 
     uint32_t get_n_kv() const;
@@ -244,25 +293,42 @@ public:
     ggml_tensor * get_v(ggml_context * ctx, int32_t il) const;
 
     // store k_cur and v_cur in the cache based on the provided head location
-    ggml_tensor * cpy_k(ggml_context * ctx, ggml_tensor * k_cur, int32_t il) const;
-    ggml_tensor * cpy_v(ggml_context * ctx, ggml_tensor * v_cur, int32_t il) const;
+    ggml_tensor * cpy_k(ggml_context * ctx, ggml_tensor * k_cur, ggml_tensor * k_idxs, int32_t il) const;
+    ggml_tensor * cpy_v(ggml_context * ctx, ggml_tensor * v_cur, ggml_tensor * v_idxs, int32_t il) const;
 
-    void set_input_k_shift(ggml_tensor * dst) const;
+    ggml_tensor * build_input_k_idxs(ggml_context * ctx, const llama_ubatch & ubatch) const;
+    ggml_tensor * build_input_v_idxs(ggml_context * ctx, const llama_ubatch & ubatch) const;
 
+    void set_input_k_idxs(ggml_tensor * dst, const llama_ubatch * ubatch) const;
+    void set_input_v_idxs(ggml_tensor * dst, const llama_ubatch * ubatch) const;
+
+    void set_input_k_shift   (ggml_tensor * dst) const;
     void set_input_kq_mask   (ggml_tensor * dst, const llama_ubatch * ubatch, bool causal_attn) const;
     void set_input_pos_bucket(ggml_tensor * dst, const llama_ubatch * ubatch) const;
 
 private:
-    const llama_memory_status status;
+    llama_memory_status status;
 
     llama_kv_cache_unified * kv;
+    llama_context * lctx;
 
-    llama_sbatch sbatch;
+    //
+    // update context
+    //
 
-    // the index of the next ubatch to process
-    size_t i_next = 0;
+    bool do_shift = false;
+
+    defrag_info dinfo;
+
+    //
+    // batch processing context
+    //
+
+    // the index of the cur ubatch to process
+    size_t i_cur = 0;
+
+    slot_info_vec_t sinfos;
 
-    std::vector<uint32_t> heads;
     std::vector<llama_ubatch> ubatches;
 
     //
@@ -272,7 +338,4 @@ private:
     // a heuristic, to avoid attending the full cache if it is not yet utilized
     // as the cache gets filled, the benefit from this heuristic disappears
     int32_t n_kv;
-
-    // the beginning of the current slot in which the ubatch will be inserted
-    int32_t head;
 };
diff --git a/src/llama-kv-cache.cpp b/src/llama-kv-cache.cpp
deleted file mode 100644
index aefd23e32..000000000
--- a/src/llama-kv-cache.cpp
+++ /dev/null
@@ -1 +0,0 @@
-#include "llama-kv-cache.h"
diff --git a/src/llama-kv-cache.h b/src/llama-kv-cache.h
deleted file mode 100644
index 2d04705f2..000000000
--- a/src/llama-kv-cache.h
+++ /dev/null
@@ -1,44 +0,0 @@
-#pragma once
-
-#include "llama.h"
-#include "llama-io.h"
-#include "llama-memory.h"
-
-struct llama_kv_cache : public llama_memory_i {
-    virtual ~llama_kv_cache() = default;
-
-    // split the input batch into a set of ubatches and verify that they can fit into the cache
-    // return a state object containing the ubatches and KV cache state required to process them
-    // check the llama_memory_state_i::get_status() for the result
-    virtual llama_memory_state_ptr init_batch(
-            const llama_batch & batch,
-            uint32_t n_ubatch,
-            bool embd_pooled,
-            bool logits_all) = 0;
-
-    // simulate full cache, used for allocating worst-case compute buffers
-    virtual llama_memory_state_ptr init_full() = 0;
-
-    // process any pending defrag/shift/etc. operations
-    // optionally call once before processing a new batch
-    // return true if any operations were performed
-    virtual bool update(llama_context & lctx) = 0;
-
-    // schedule a defrag if the fragmentation threshold is exceeded. otherwise, do nothing
-    // TODO: change to
-    //   llama_memory_state_ptr init_defrag(float thold) = 0;
-    //
-    virtual void defrag_sched(float thold) = 0;
-
-    // getters
-    virtual bool get_can_shift() const = 0;
-
-    bool get_can_edit() const override { return get_can_shift(); }
-
-    //
-    // state write/read
-    //
-
-    virtual void state_write(llama_io_write_i & io, llama_seq_id seq_id = -1) const = 0;
-    virtual void state_read (llama_io_read_i  & io, llama_seq_id seq_id = -1) = 0;
-};
diff --git a/src/llama-kv-cells.h b/src/llama-kv-cells.h
index 9e2c4d927..0d0dd316f 100644
--- a/src/llama-kv-cells.h
+++ b/src/llama-kv-cells.h
@@ -7,6 +7,7 @@
 #include <cassert>
 #include <vector>
 #include <set>
+#include <map>
 
 // meta information about KV cells that can be part of multiple sequences at the same time
 // TODO: add unit tests
@@ -23,7 +24,7 @@ public:
 
         used.clear();
 
-        for (uint32_t s = 0; s < LLAMA_MAX_PARALLEL_SEQUENCES; ++s) {
+        for (uint32_t s = 0; s < LLAMA_MAX_SEQ; ++s) {
             seq_pos[s].clear();
         }
     }
@@ -80,6 +81,9 @@ public:
         assert(isrc < pos.size());
         assert(idst < pos.size());
 
+        assert(pos[idst] == -1);
+        assert(pos[isrc] != -1);
+
         pos  [idst] = pos  [isrc];
         shift[idst] = shift[isrc];
         seq  [idst] = seq  [isrc];
@@ -101,10 +105,30 @@ public:
         res.resize(n);
 
         for (uint32_t j = 0; j < n; ++j) {
-            res.pos[j] = pos[i + j];
-            res.seq[j] = seq[i + j];
+            const auto idx = i + j;
 
-            assert(shift[i + j] == 0);
+            res.pos[j] = pos[idx];
+            res.seq[j] = seq[idx];
+
+            assert(shift[idx] == 0);
+        }
+
+        return res;
+    }
+
+    // copy the state of cells [idxs[0], idxs[1], ..., idxs[idxs.size() - 1])
+    llama_kv_cells_unified cp(const std::vector<uint32_t> & idxs) const {
+        llama_kv_cells_unified res;
+
+        res.resize(idxs.size());
+
+        for (uint32_t j = 0; j < idxs.size(); ++j) {
+            const auto idx = idxs[j];
+
+            res.pos[j] = pos[idx];
+            res.seq[j] = seq[idx];
+
+            assert(shift[idx] == 0);
         }
 
         return res;
@@ -115,26 +139,58 @@ public:
         assert(i + other.pos.size() <= pos.size());
 
         for (uint32_t j = 0; j < other.pos.size(); ++j) {
-            if (pos[i + j] == -1 && other.pos[j] != -1) {
+            const auto idx = i + j;
+
+            if (pos[idx] == -1 && other.pos[j] != -1) {
                 used.insert(i + j);
             }
 
-            if (pos[i + j] != -1 && other.pos[j] == -1) {
+            if (pos[idx] != -1 && other.pos[j] == -1) {
                 used.erase(i + j);
             }
 
-            if (pos[i + j] != -1) {
+            if (pos[idx] != -1) {
                 seq_pos_rm(i + j);
             }
 
-            pos[i + j] = other.pos[j];
-            seq[i + j] = other.seq[j];
+            pos[idx] = other.pos[j];
+            seq[idx] = other.seq[j];
 
-            if (pos[i + j] != -1) {
+            if (pos[idx] != -1) {
                 seq_pos_add(i + j);
             }
 
-            assert(shift[i + j] == 0);
+            assert(shift[idx] == 0);
+        }
+    }
+
+    // set the state of cells [idxs[0], idxs[1], ..., idxs[idxs.size() - 1])
+    void set(const std::vector<uint32_t> & idxs, const llama_kv_cells_unified & other) {
+        assert(idxs.size() == other.pos.size());
+
+        for (uint32_t j = 0; j < other.pos.size(); ++j) {
+            const auto idx = idxs[j];
+
+            if (pos[idx] == -1 && other.pos[j] != -1) {
+                used.insert(idx);
+            }
+
+            if (pos[idx] != -1 && other.pos[j] == -1) {
+                used.erase(idx);
+            }
+
+            if (pos[idx] != -1) {
+                seq_pos_rm(idx);
+            }
+
+            pos[idx] = other.pos[j];
+            seq[idx] = other.seq[j];
+
+            if (pos[idx] != -1) {
+                seq_pos_add(idx);
+            }
+
+            assert(shift[idx] == 0);
         }
     }
 
@@ -144,9 +200,10 @@ public:
         assert(pos[i] != -1);
 
         seq_pos_rm(i);
+        seq[i].reset();
 
         pos[i] = -1;
-        seq[i].reset();
+        shift[i] = 0;
 
         used.erase(i);
     }
@@ -160,10 +217,11 @@ public:
         assert(seq_id >= 0);
 
         seq[i].reset(seq_id);
-        seq_pos[seq_id].erase(pos[i]);
+        seq_pos_dec(seq_id, pos[i]);
 
         if (seq[i].none()) {
             pos[i] = -1;
+            shift[i] = 0;
 
             used.erase(i);
 
@@ -182,7 +240,7 @@ public:
             seq[i].reset();
 
             seq[i].set(seq_id);
-            seq_pos[seq_id].insert(pos[i]);
+            seq_pos_inc(seq_id, pos[i]);
 
             return false;
         }
@@ -192,6 +250,7 @@ public:
             seq[i].reset();
 
             pos[i] = -1;
+            shift[i] = 0;
 
             used.erase(i);
 
@@ -226,7 +285,7 @@ public:
         assert(!seq[i].test(seq_id));
 
         seq[i].set(seq_id);
-        seq_pos[seq_id].insert(pos[i]);
+        seq_pos_inc(seq_id, pos[i]);
     }
 
     // return the sequence id of this cell
@@ -234,7 +293,7 @@ public:
     llama_seq_id seq_get(uint32_t i) const {
         assert(seq[i].count() == 1);
 
-        for (int s = 0; s < LLAMA_MAX_PARALLEL_SEQUENCES; ++s) {
+        for (int s = 0; s < LLAMA_MAX_SEQ; ++s) {
             if (seq[i].test(s)) {
                 return s;
             }
@@ -247,26 +306,30 @@ public:
     // return -1 if the sequence is not present
     llama_pos seq_pos_min(llama_seq_id seq_id) const {
         assert(seq_id >= 0);
-        assert(seq_id < LLAMA_MAX_PARALLEL_SEQUENCES);
+        assert(seq_id < LLAMA_MAX_SEQ);
 
         if (seq_pos[seq_id].empty()) {
             return -1;
         }
 
-        return *seq_pos[seq_id].begin();
+        assert(seq_pos[seq_id].begin()->second > 0);
+
+        return seq_pos[seq_id].begin()->first;
     }
 
     // the maximum position of sequence seq_id currently present in any of the cells
     // return -1 if the sequence is not present
     llama_pos seq_pos_max(llama_seq_id seq_id) const {
         assert(seq_id >= 0);
-        assert(seq_id < LLAMA_MAX_PARALLEL_SEQUENCES);
+        assert(seq_id < LLAMA_MAX_SEQ);
 
         if (seq_pos[seq_id].empty()) {
             return -1;
         }
 
-        return *seq_pos[seq_id].rbegin();
+        assert(seq_pos[seq_id].rbegin()->second > 0);
+
+        return seq_pos[seq_id].rbegin()->first;
     }
 
     // note: call only if the cell is not empty
@@ -317,21 +380,20 @@ public:
         pos[i]   += d;
         shift[i] += d;
 
-        seq_pos_add(i);
-
         has_shift = true;
 
         if (pos[i] < 0) {
-            seq_pos_rm(i);
-
             seq[i].reset();
             pos[i] = -1;
+            shift[i] = 0;
 
             used.erase(i);
 
             return true;
         }
 
+        seq_pos_add(i);
+
         return false;
     }
 
@@ -379,31 +441,50 @@ private:
     //
     std::vector<llama_pos> shift;
 
-    using bits_t = std::bitset<LLAMA_MAX_PARALLEL_SEQUENCES>;
+    using seq_set_t = std::bitset<LLAMA_MAX_SEQ>;
 
     // the bitset seq[i] tells us which sequences are currently occupying the i-th cell
-    std::vector<bits_t> seq;
+    std::vector<seq_set_t> seq;
 
-    // the set seq_pos[s] tells us which positions are currently present for sequence s
+    // the set seq_pos[s][p] tells us how many times the position p is currently present for sequence s
+    // if the position p is not present, seq_pos[s][p] is not set
     // this way seq_pos[s].begin() and seq_pos[s].rbegin() give us the min/max positions currently in the cache
-    std::set<llama_pos> seq_pos[LLAMA_MAX_PARALLEL_SEQUENCES];
+    //
+    // note that we cannot a use an std::set because in some cases a position can occur more than once for the same seq:
+    //  - during performing a cache reuse via (rm + add)
+    //  - some vision models have input embeddings with repeating positions
+    //
+    std::map<llama_pos, int> seq_pos[LLAMA_MAX_SEQ];
 
     // helper functions for updating `seq_pos`, once cell at a time:
 
+    void seq_pos_dec(llama_seq_id s, llama_pos p) {
+        auto it = seq_pos[s].find(p);
+        assert(it != seq_pos[s].end());
+
+        if (--it->second == 0) {
+            seq_pos[s].erase(it);
+        }
+    }
+
+    void seq_pos_inc(llama_seq_id s, llama_pos p) {
+        seq_pos[s][p]++;
+    }
+
     // remove cell i
     void seq_pos_rm(uint32_t i) {
-        for (int s = 0; s < LLAMA_MAX_PARALLEL_SEQUENCES; ++s) {
+        for (int s = 0; s < LLAMA_MAX_SEQ; ++s) {
             if (seq[i].test(s)) {
-                seq_pos[s].erase(pos[i]);
+                seq_pos_dec(s, pos[i]);
             }
         }
     }
 
     // add cell i
     void seq_pos_add(uint32_t i) {
-        for (int s = 0; s < LLAMA_MAX_PARALLEL_SEQUENCES; ++s) {
+        for (int s = 0; s < LLAMA_MAX_SEQ; ++s) {
             if (seq[i].test(s)) {
-                seq_pos[s].insert(pos[i]);
+                seq_pos_inc(s, pos[i]);
             }
         }
     }
diff --git a/src/llama-memory-hybrid.cpp b/src/llama-memory-hybrid.cpp
new file mode 100644
index 000000000..6cd10db06
--- /dev/null
+++ b/src/llama-memory-hybrid.cpp
@@ -0,0 +1,251 @@
+#include "llama-memory-hybrid.h"
+
+#include "llama-impl.h"
+#include "llama-model.h"
+#include "llama-context.h"
+
+//
+// llama_memory_hybrid
+//
+
+llama_memory_hybrid::llama_memory_hybrid(
+    const llama_model & model,
+                         /* attn */
+            ggml_type    type_k,
+            ggml_type    type_v,
+                 bool    v_trans,
+             uint32_t    kv_size,
+             uint32_t    n_pad,
+             uint32_t    n_swa,
+       llama_swa_type    swa_type,
+                         /* recurrent */
+            ggml_type    type_r,
+            ggml_type    type_s,
+             uint32_t    rs_size,
+                         /* common */
+             uint32_t    n_seq_max,
+                 bool    offload,
+                         /* layer filters */
+      layer_filter_cb && filter_attn,
+      layer_filter_cb && filter_recr) :
+    hparams(model.hparams),
+    mem_attn(new llama_kv_cache_unified(
+        model,
+        filter_attn == nullptr ?
+            [&](int32_t il) { return !hparams.is_recurrent(il); }
+            : filter_attn,
+        type_k,
+        type_v,
+        v_trans,
+        offload,
+        kv_size,
+        n_seq_max,
+        n_pad,
+        n_swa,
+        swa_type
+    )),
+    mem_recr(new llama_memory_recurrent(
+        model,
+        filter_recr == nullptr ?
+            [&](int32_t il) { return hparams.is_recurrent(il); }
+            : filter_recr,
+        type_r,
+        type_s,
+        offload,
+        rs_size,
+        n_seq_max
+    )) {}
+
+llama_memory_context_ptr llama_memory_hybrid::init_batch(llama_batch_allocr & balloc, uint32_t n_ubatch, bool embd_all) {
+    do {
+        balloc.split_reset();
+
+        // follow the recurrent pattern for creating the ubatch splits
+        std::vector<llama_ubatch> ubatches;
+
+        while (true) {
+            llama_ubatch ubatch;
+
+            if (embd_all) {
+                // if all tokens are output, split by sequence
+                ubatch = balloc.split_seq(n_ubatch);
+            } else {
+                ubatch = balloc.split_equal(n_ubatch, false);
+            }
+
+            if (ubatch.n_tokens == 0) {
+                break;
+            }
+
+            ubatches.push_back(std::move(ubatch)); // NOLINT
+        }
+
+        if (balloc.get_n_used() < balloc.get_n_tokens()) {
+            // failed to find a suitable split
+            break;
+        }
+
+        // prepare the recurrent batches first
+        if (!mem_recr->prepare(ubatches)) {
+            // TODO: will the recurrent cache be in an undefined context at this point?
+            LLAMA_LOG_ERROR("%s: failed to prepare recurrent ubatches\n", __func__);
+            return std::make_unique<llama_memory_hybrid_context>(LLAMA_MEMORY_STATUS_FAILED_PREPARE);
+        }
+
+        // prepare the attention cache
+        auto heads_attn = mem_attn->prepare(ubatches);
+        if (heads_attn.empty()) {
+            LLAMA_LOG_ERROR("%s: failed to prepare attention ubatches\n", __func__);
+            return std::make_unique<llama_memory_hybrid_context>(LLAMA_MEMORY_STATUS_FAILED_PREPARE);
+        }
+
+        return std::make_unique<llama_memory_hybrid_context>(
+                this, std::move(heads_attn), std::move(ubatches));
+    } while(false);
+
+    return std::make_unique<llama_memory_hybrid_context>(LLAMA_MEMORY_STATUS_FAILED_PREPARE);
+}
+
+llama_memory_context_ptr llama_memory_hybrid::init_full() {
+    return std::make_unique<llama_memory_hybrid_context>(this);
+}
+
+llama_memory_context_ptr llama_memory_hybrid::init_update(llama_context * lctx, bool optimize) {
+    return std::make_unique<llama_memory_hybrid_context>(this, lctx, optimize);
+}
+
+bool llama_memory_hybrid::get_can_shift() const {
+    // Shifting is trivially supported for recurrent
+    return mem_attn->get_can_shift();
+}
+
+void llama_memory_hybrid::clear(bool data) {
+    mem_attn->clear(data);
+    mem_recr->clear(data);
+}
+
+bool llama_memory_hybrid::seq_rm(llama_seq_id seq_id, llama_pos p0, llama_pos p1) {
+    // Try removing from the recurrent cache first since it may fail. If it does
+    // fail, the cache will not have been mutated.
+    if (!mem_recr->seq_rm(seq_id, p0, p1)) {
+        return false;
+    }
+    return mem_attn->seq_rm(seq_id, p0, p1);
+}
+
+void llama_memory_hybrid::seq_cp(llama_seq_id seq_id_src, llama_seq_id seq_id_dst, llama_pos p0, llama_pos p1) {
+    mem_attn->seq_cp(seq_id_src, seq_id_dst, p0, p1);
+    mem_recr->seq_cp(seq_id_src, seq_id_dst, p0, p1);
+}
+
+void llama_memory_hybrid::seq_keep(llama_seq_id seq_id) {
+    mem_attn->seq_keep(seq_id);
+    mem_recr->seq_keep(seq_id);
+}
+
+void llama_memory_hybrid::seq_add(llama_seq_id seq_id, llama_pos p0, llama_pos p1, llama_pos shift) {
+    mem_attn->seq_add(seq_id, p0, p1, shift);
+    mem_recr->seq_add(seq_id, p0, p1, shift);
+}
+
+void llama_memory_hybrid::seq_div(llama_seq_id seq_id, llama_pos p0, llama_pos p1, int d) {
+    mem_attn->seq_div(seq_id, p0, p1, d);
+    mem_recr->seq_div(seq_id, p0, p1, d);
+}
+
+llama_pos llama_memory_hybrid::seq_pos_min(llama_seq_id seq_id) const {
+    // the min of the total cache is the max of the two caches' min values
+    return std::max(mem_attn->seq_pos_min(seq_id), mem_recr->seq_pos_min(seq_id));
+}
+
+llama_pos llama_memory_hybrid::seq_pos_max(llama_seq_id seq_id) const {
+    // the max of the total cache is the min of the two caches' max values
+    return std::min(mem_attn->seq_pos_max(seq_id), mem_recr->seq_pos_max(seq_id));
+}
+
+void llama_memory_hybrid::state_write(llama_io_write_i & io, llama_seq_id seq_id) const {
+    mem_attn->state_write(io, seq_id);
+    mem_recr->state_write(io, seq_id);
+}
+
+void llama_memory_hybrid::state_read(llama_io_read_i & io, llama_seq_id seq_id) {
+    mem_attn->state_read(io, seq_id);
+    mem_recr->state_read(io, seq_id);
+}
+
+llama_kv_cache_unified * llama_memory_hybrid::get_mem_attn() const {
+    return mem_attn.get();
+}
+
+llama_memory_recurrent * llama_memory_hybrid::get_mem_recr() const {
+    return mem_recr.get();
+}
+
+llama_memory_hybrid_context::llama_memory_hybrid_context(llama_memory_status status) : status(status) {}
+
+llama_memory_hybrid_context::llama_memory_hybrid_context(llama_memory_hybrid * mem) :
+    ctx_attn(mem->get_mem_attn()->init_full()),
+    ctx_recr(mem->get_mem_recr()->init_full()),
+    status(llama_memory_status_combine(ctx_attn->get_status(), ctx_recr->get_status())) {
+}
+
+llama_memory_hybrid_context::llama_memory_hybrid_context(
+        llama_memory_hybrid * mem,
+              llama_context * lctx,
+                       bool   optimize) :
+    ctx_attn(mem->get_mem_attn()->init_update(lctx, optimize)),
+    ctx_recr(mem->get_mem_recr()->init_update(lctx, optimize)),
+    status(llama_memory_status_combine(ctx_attn->get_status(), ctx_recr->get_status())) {
+}
+
+llama_memory_hybrid_context::llama_memory_hybrid_context(
+              llama_memory_hybrid * mem,
+                  slot_info_vec_t   sinfos_attn,
+        std::vector<llama_ubatch>   ubatches) :
+    ubatches(std::move(ubatches)),
+    // note: here we copy the ubatches. not sure if this is ideal
+    ctx_attn(new llama_kv_cache_unified_context(mem->get_mem_attn(), std::move(sinfos_attn), this->ubatches)),
+    ctx_recr(new llama_memory_recurrent_context(mem->get_mem_recr(),                        this->ubatches)),
+    status(llama_memory_status_combine(ctx_attn->get_status(), ctx_recr->get_status())) {
+}
+
+bool llama_memory_hybrid_context::next() {
+    assert(status == LLAMA_MEMORY_STATUS_SUCCESS);
+
+    ctx_attn->next();
+    ctx_recr->next();
+
+    if (++i_next >= ubatches.size()) {
+        return false;
+    }
+
+    return true;
+}
+
+bool llama_memory_hybrid_context::apply() {
+    assert(!llama_memory_status_is_fail(status));
+
+    bool res = true;
+
+    res = res & ctx_attn->apply();
+    res = res & ctx_recr->apply();
+
+    return res;
+}
+
+llama_memory_status llama_memory_hybrid_context::get_status() const {
+    return status;
+}
+
+const llama_ubatch & llama_memory_hybrid_context::get_ubatch() const {
+    assert(status == LLAMA_MEMORY_STATUS_SUCCESS);
+    return ubatches[i_next];
+}
+
+const llama_kv_cache_unified_context * llama_memory_hybrid_context::get_attn() const {
+    return static_cast<const llama_kv_cache_unified_context *>(ctx_attn.get());
+}
+
+const llama_memory_recurrent_context * llama_memory_hybrid_context::get_recr() const {
+    return static_cast<const llama_memory_recurrent_context *>(ctx_recr.get());
+}
diff --git a/src/llama-memory-hybrid.h b/src/llama-memory-hybrid.h
new file mode 100644
index 000000000..4ac318175
--- /dev/null
+++ b/src/llama-memory-hybrid.h
@@ -0,0 +1,140 @@
+#pragma once
+
+#include "llama-batch.h"
+#include "llama-graph.h"
+#include "llama-kv-cache-unified.h"
+#include "llama-memory.h"
+#include "llama-memory-recurrent.h"
+
+#include <memory>
+#include <vector>
+
+//
+// llama_memory_hybrid
+//
+
+// utilizes instances of llama_memory_recurrent and llama_kv_cache_unified to
+//   support models where each layer may be either attention-based or recurrent
+
+class llama_memory_hybrid : public llama_memory_i {
+public:
+
+    // this callback is used to filter out layers that should not be included in the cache
+    using layer_filter_cb = std::function<bool(int32_t il)>;
+
+    llama_memory_hybrid(
+        const llama_model & model,
+                            /* attn */
+                ggml_type    type_k,
+                ggml_type    type_v,
+                     bool    v_trans,
+                 uint32_t    kv_size,
+                 uint32_t    n_pad,
+                 uint32_t    n_swa,
+           llama_swa_type    swa_type,
+                             /* recurrent */
+                ggml_type    type_r,
+                ggml_type    type_s,
+                 uint32_t    rs_size,
+                             /* common */
+                 uint32_t    n_seq_max,
+                     bool    offload,
+                             /* layer filters */
+          layer_filter_cb && filter_attn = nullptr,
+          layer_filter_cb && filter_recr = nullptr);
+
+    ~llama_memory_hybrid() = default;
+
+    //
+    // llama_memory_i
+    //
+
+    llama_memory_context_ptr init_batch(
+            llama_batch_allocr & balloc,
+            uint32_t n_ubatch,
+            bool embd_all) override;
+
+    llama_memory_context_ptr init_full() override;
+
+    llama_memory_context_ptr init_update(llama_context * lctx, bool optimize) override;
+
+    bool get_can_shift() const override;
+
+    void clear(bool data) 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 shift) override;
+    void seq_div (llama_seq_id seq_id,                              llama_pos p0, llama_pos p1, int d) override;
+
+    llama_pos seq_pos_min(llama_seq_id seq_id) const override;
+    llama_pos seq_pos_max(llama_seq_id seq_id) const override;
+
+    // state write/load
+
+    void state_write(llama_io_write_i & io, llama_seq_id seq_id = -1) const override;
+    void state_read (llama_io_read_i  & io, llama_seq_id seq_id = -1)       override;
+
+    //
+    // llama_memory_hybrid specific API
+    //
+
+    llama_kv_cache_unified * get_mem_attn() const;
+    llama_memory_recurrent * get_mem_recr() const;
+
+private:
+    const llama_hparams & hparams;
+
+    const std::unique_ptr<llama_kv_cache_unified> mem_attn;
+    const std::unique_ptr<llama_memory_recurrent> mem_recr;
+};
+
+class llama_memory_hybrid_context : public llama_memory_context_i {
+public:
+    using slot_info_vec_t = llama_kv_cache_unified::slot_info_vec_t;
+
+    // init failure
+    explicit llama_memory_hybrid_context(llama_memory_status status);
+
+    // init full
+    explicit llama_memory_hybrid_context(llama_memory_hybrid * mem);
+
+    // init update
+    explicit llama_memory_hybrid_context(
+        llama_memory_hybrid * mem,
+              llama_context * lctx,
+                       bool   optimize);
+
+    // init success
+    llama_memory_hybrid_context(
+              llama_memory_hybrid * mem,
+                  slot_info_vec_t   sinfos_attn,
+        std::vector<llama_ubatch>   ubatches);
+
+    ~llama_memory_hybrid_context() = default;
+
+    bool next()  override;
+    bool apply() override;
+
+    llama_memory_status  get_status() const override;
+    const llama_ubatch & get_ubatch() const override;
+
+    //
+    // llama_memory_hybrid_context
+    //
+
+    const llama_kv_cache_unified_context * get_attn() const;
+    const llama_memory_recurrent_context * get_recr() const;
+
+private:
+    // the index of the next ubatch to process
+    size_t i_next = 0;
+
+    std::vector<llama_ubatch> ubatches;
+
+    const llama_memory_context_ptr ctx_attn;
+    const llama_memory_context_ptr ctx_recr;
+
+    const llama_memory_status status;
+};
diff --git a/src/llama-kv-cache-recurrent.cpp b/src/llama-memory-recurrent.cpp
similarity index 57%
rename from src/llama-kv-cache-recurrent.cpp
rename to src/llama-memory-recurrent.cpp
index 641eab2f3..4b90dac7a 100644
--- a/src/llama-kv-cache-recurrent.cpp
+++ b/src/llama-memory-recurrent.cpp
@@ -1,6 +1,7 @@
-#include "llama-kv-cache-recurrent.h"
+#include "llama-memory-recurrent.h"
 
 #include "llama-impl.h"
+#include "llama-io.h"
 #include "llama-batch.h"
 #include "llama-model.h"
 
@@ -11,27 +12,28 @@
 #include <stdexcept>
 
 //
-// llama_kv_cache_recurrent
+// llama_memory_recurrent
 //
 
-llama_kv_cache_recurrent::llama_kv_cache_recurrent(
-        const llama_model & model,
-                ggml_type   type_k,
-                ggml_type   type_v,
-                     bool   offload,
-                 uint32_t   kv_size,
-                 uint32_t   n_seq_max) : hparams(model.hparams), n_seq_max(n_seq_max) {
+llama_memory_recurrent::llama_memory_recurrent(
+        const llama_model &  model,
+          layer_filter_cb && filter,
+                ggml_type    type_r,
+                ggml_type    type_s,
+                     bool    offload,
+                 uint32_t    mem_size,
+                 uint32_t    n_seq_max) : hparams(model.hparams), n_seq_max(n_seq_max) {
     const int32_t n_layer = hparams.n_layer;
 
-    LLAMA_LOG_INFO("%s: kv_size = %u, n_seq_max = %u, type_k = '%s', type_v = '%s', n_layer = %d\n",
-            __func__, kv_size, n_seq_max, ggml_type_name(type_k), ggml_type_name(type_v), n_layer);
+    LLAMA_LOG_INFO("%s: mem_size = %u, n_seq_max = %u, type_r = '%s', type_s = '%s', n_layer = %d\n",
+            __func__, mem_size, n_seq_max, ggml_type_name(type_r), ggml_type_name(type_s), n_layer);
 
     head = 0;
-    size = kv_size;
+    size = mem_size;
     used = 0;
 
     cells.clear();
-    cells.resize(kv_size);
+    cells.resize(mem_size);
 
     // create a context for each buffer type
     std::map<ggml_backend_buffer_type_t, ggml_context *> ctx_map;
@@ -58,12 +60,14 @@ llama_kv_cache_recurrent::llama_kv_cache_recurrent(
         return it->second;
     };
 
-    k_l.reserve(n_layer);
-    v_l.reserve(n_layer);
+    r_l.resize(n_layer);
+    s_l.resize(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();
+        if (filter && !filter(i)) {
+            LLAMA_LOG_DEBUG("%s: layer %3d: skipped\n", __func__, i);
+            continue;
+        }
 
         const char * dev_name = "CPU";
 
@@ -83,12 +87,12 @@ llama_kv_cache_recurrent::llama_kv_cache_recurrent(
             throw std::runtime_error("failed to create ggml context for kv cache");
         }
 
-        ggml_tensor * k = ggml_new_tensor_1d(ctx, type_k, n_embd_k_gqa*kv_size);
-        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);
-        k_l.push_back(k);
-        v_l.push_back(v);
+        ggml_tensor * r = ggml_new_tensor_1d(ctx, type_r, hparams.n_embd_r()*mem_size);
+        ggml_tensor * s = ggml_new_tensor_1d(ctx, type_s, hparams.n_embd_s()*mem_size);
+        ggml_format_name(r, "cache_r_l%d", i);
+        ggml_format_name(s, "cache_s_l%d", i);
+        r_l[i] = r;
+        s_l[i] = s;
     }
 
     // allocate tensors and initialize the buffers to avoid NaNs in the padding
@@ -106,32 +110,35 @@ llama_kv_cache_recurrent::llama_kv_cache_recurrent(
     }
 
     {
-        const size_t memory_size_k = size_k_bytes();
-        const size_t memory_size_v = size_v_bytes();
+        const size_t memory_size_r = size_r_bytes();
+        const size_t memory_size_s = size_s_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));
+        LLAMA_LOG_INFO("%s: KV self size  = %7.2f MiB, R (%s): %7.2f MiB, S (%s): %7.2f MiB\n", __func__,
+                (float)(memory_size_r + memory_size_s) / (1024.0f * 1024.0f),
+                ggml_type_name(type_r), (float)memory_size_r / (1024.0f * 1024.0f),
+                ggml_type_name(type_s), (float)memory_size_s / (1024.0f * 1024.0f));
     }
 }
 
-void llama_kv_cache_recurrent::clear() {
+void llama_memory_recurrent::clear(bool data) {
     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);
+    if (data) {
+        for (auto & buf : bufs) {
+            ggml_backend_buffer_clear(buf.get(), 0);
+        }
     }
 }
 
-bool llama_kv_cache_recurrent::seq_rm(llama_seq_id seq_id, llama_pos p0, llama_pos p1) {
+bool llama_memory_recurrent::seq_rm(llama_seq_id seq_id, llama_pos p0, llama_pos p1) {
     uint32_t new_head = size;
 
     if (p0 < 0) {
@@ -150,7 +157,7 @@ bool llama_kv_cache_recurrent::seq_rm(llama_seq_id seq_id, llama_pos p0, llama_p
     if (0 <= seq_id) {
         int32_t & tail_id = cells[seq_id].tail;
         if (tail_id >= 0) {
-            const kv_cell & cell = cells[tail_id];
+            const auto & cell = cells[tail_id];
             // partial intersection is invalid
             if ((0 < p0 && p0 <= cell.pos) || (0 < p1 && p1 <= cell.pos)) {
                 return false;
@@ -198,7 +205,7 @@ bool llama_kv_cache_recurrent::seq_rm(llama_seq_id seq_id, llama_pos p0, llama_p
     return true;
 }
 
-void llama_kv_cache_recurrent::seq_cp(llama_seq_id seq_id_src, llama_seq_id seq_id_dst, llama_pos p0, llama_pos p1) {
+void llama_memory_recurrent::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;
     }
@@ -212,11 +219,11 @@ void llama_kv_cache_recurrent::seq_cp(llama_seq_id seq_id_src, llama_seq_id seq_
     }
 
     if ((uint32_t) seq_id_dst < size && (uint32_t) seq_id_src < size) {
-        kv_cell & tail_src = cells[seq_id_src];
-        kv_cell & tail_dst = cells[seq_id_dst];
+        auto & tail_src = cells[seq_id_src];
+        auto & tail_dst = cells[seq_id_dst];
         if (tail_dst.tail >= 0) {
             // clear destination seq_id if it wasn't empty
-            kv_cell & cell_dst = cells[tail_dst.tail];
+            auto & cell_dst = cells[tail_dst.tail];
 
             cell_dst.seq_id.erase(seq_id_dst);
             tail_dst.tail = -1;
@@ -227,7 +234,7 @@ void llama_kv_cache_recurrent::seq_cp(llama_seq_id seq_id_src, llama_seq_id seq_
             }
         }
         if (tail_src.tail >= 0) {
-            kv_cell & cell_src = cells[tail_src.tail];
+            auto & cell_src = cells[tail_src.tail];
 
             cell_src.seq_id.insert(seq_id_dst);
             tail_dst.tail = tail_src.tail;
@@ -235,7 +242,7 @@ void llama_kv_cache_recurrent::seq_cp(llama_seq_id seq_id_src, llama_seq_id seq_
     }
 }
 
-void llama_kv_cache_recurrent::seq_keep(llama_seq_id seq_id) {
+void llama_memory_recurrent::seq_keep(llama_seq_id seq_id) {
     uint32_t new_head = size;
 
     for (uint32_t i = 0; i < size; ++i) {
@@ -267,7 +274,7 @@ void llama_kv_cache_recurrent::seq_keep(llama_seq_id seq_id) {
     }
 }
 
-void llama_kv_cache_recurrent::seq_add(llama_seq_id seq_id, llama_pos p0, llama_pos p1, llama_pos shift) {
+void llama_memory_recurrent::seq_add(llama_seq_id seq_id, llama_pos p0, llama_pos p1, llama_pos shift) {
     if (shift == 0) {
         return;
     }
@@ -289,7 +296,7 @@ void llama_kv_cache_recurrent::seq_add(llama_seq_id seq_id, llama_pos p0, llama_
     if (0 <= seq_id && seq_id < (int64_t) size) {
         const int32_t tail_id = cells[seq_id].tail;
         if (tail_id >= 0) {
-            kv_cell & cell = cells[tail_id];
+            auto & cell = cells[tail_id];
             if (cell.has_seq_id(seq_id) && p0 <= cell.pos && cell.pos < p1) {
                 cell.pos += shift;
             }
@@ -297,7 +304,7 @@ void llama_kv_cache_recurrent::seq_add(llama_seq_id seq_id, llama_pos p0, llama_
     }
 }
 
-void llama_kv_cache_recurrent::seq_div(llama_seq_id seq_id, llama_pos p0, llama_pos p1, int d) {
+void llama_memory_recurrent::seq_div(llama_seq_id seq_id, llama_pos p0, llama_pos p1, int d) {
     if (d == 1) {
         return;
     }
@@ -319,7 +326,7 @@ void llama_kv_cache_recurrent::seq_div(llama_seq_id seq_id, llama_pos p0, llama_
     if (0 <= seq_id && seq_id < (int64_t) size) {
         const int32_t tail_id = cells[seq_id].tail;
         if (tail_id >= 0) {
-            kv_cell & cell = cells[tail_id];
+            auto & cell = cells[tail_id];
             if (cell.has_seq_id(seq_id) && p0 <= cell.pos && cell.pos < p1) {
                 cell.pos /= d;
             }
@@ -327,7 +334,7 @@ void llama_kv_cache_recurrent::seq_div(llama_seq_id seq_id, llama_pos p0, llama_
     }
 }
 
-llama_pos llama_kv_cache_recurrent::seq_pos_min(llama_seq_id seq_id) const {
+llama_pos llama_memory_recurrent::seq_pos_min(llama_seq_id seq_id) const {
     llama_pos result = std::numeric_limits<llama_pos>::max();
 
     for (uint32_t i = 0; i < size; ++i) {
@@ -343,7 +350,7 @@ llama_pos llama_kv_cache_recurrent::seq_pos_min(llama_seq_id seq_id) const {
     return result;
 }
 
-llama_pos llama_kv_cache_recurrent::seq_pos_max(llama_seq_id seq_id) const {
+llama_pos llama_memory_recurrent::seq_pos_max(llama_seq_id seq_id) const {
     llama_pos result = -1;
 
     for (uint32_t i = 0; i < size; ++i) {
@@ -355,38 +362,51 @@ llama_pos llama_kv_cache_recurrent::seq_pos_max(llama_seq_id seq_id) const {
     return result;
 }
 
-llama_memory_state_ptr llama_kv_cache_recurrent::init_batch(const llama_batch & batch, uint32_t n_ubatch, bool embd_pooled, bool logits_all) {
-    GGML_UNUSED(embd_pooled);
+llama_memory_context_ptr llama_memory_recurrent::init_batch(llama_batch_allocr & balloc, uint32_t n_ubatch, bool embd_all) {
+    do {
+        balloc.split_reset();
 
-    auto sbatch = llama_sbatch(batch, hparams.n_embd, false, logits_all);
+        std::vector<llama_ubatch> ubatches;
+        while (true) {
+            llama_ubatch ubatch;
 
-    std::vector<llama_ubatch> ubatches;
+            if (embd_all) {
+                // if all tokens are output, split by sequence
+                ubatch = balloc.split_seq(n_ubatch);
+            } else {
+                ubatch = balloc.split_equal(n_ubatch, false);
+            }
 
-    while (sbatch.n_tokens > 0) {
-        llama_ubatch ubatch;
+            if (balloc.get_n_used() < balloc.get_n_tokens()) {
+                // failed to find a suitable split
+                break;
+            }
 
-        if (embd_pooled) {
-            // Pooled embeddings cannot be split across ubatches (yet)
-            ubatch = sbatch.split_seq(n_ubatch);
-        } else {
-            ubatch = sbatch.split_equal(n_ubatch);
+            ubatches.push_back(std::move(ubatch)); // NOLINT
         }
 
-        ubatches.push_back(ubatch);
-    }
+        if (!prepare(ubatches)) {
+            break;
+        }
 
-    if (!prepare(ubatches)) {
-        return std::make_unique<llama_kv_cache_recurrent_state>(LLAMA_MEMORY_STATUS_FAILED_PREPARE);
-    }
+        return std::make_unique<llama_memory_recurrent_context>(this, std::move(ubatches));
+    } while (false);
 
-    return std::make_unique<llama_kv_cache_recurrent_state>(LLAMA_MEMORY_STATUS_SUCCESS, this, std::move(sbatch), std::move(ubatches));
+    return std::make_unique<llama_memory_recurrent_context>(LLAMA_MEMORY_STATUS_FAILED_PREPARE);
 }
 
-llama_memory_state_ptr llama_kv_cache_recurrent::init_full() {
-    return std::make_unique<llama_kv_cache_recurrent_state>(LLAMA_MEMORY_STATUS_SUCCESS, this);
+llama_memory_context_ptr llama_memory_recurrent::init_full() {
+    return std::make_unique<llama_memory_recurrent_context>(this);
 }
 
-bool llama_kv_cache_recurrent::prepare(const std::vector<llama_ubatch> & ubatches) {
+llama_memory_context_ptr llama_memory_recurrent::init_update(llama_context * lctx, bool optimize) {
+    GGML_UNUSED(lctx);
+    GGML_UNUSED(optimize);
+
+    return std::make_unique<llama_memory_recurrent_context>(LLAMA_MEMORY_STATUS_NO_UPDATE);
+}
+
+bool llama_memory_recurrent::prepare(const std::vector<llama_ubatch> & ubatches) {
     // simply remember the full state because it is very small for this type of cache
     // TODO: optimize
     auto org_cells = cells;
@@ -395,21 +415,12 @@ bool llama_kv_cache_recurrent::prepare(const std::vector<llama_ubatch> & ubatche
 
     bool success = true;
 
-    // TODO: here we have to verify that all ubatches can fit in the cells
-    //       however, the current implementation is broken because it relies on s_copy() and s_mask() to update the cells
-    //         during the compute of each ubatch. to reproduce, uncomment the following loop and run:
-    //
-    //           $ llama-parallel -m ./mamba-130m/ggml-model-f16.gguf -np 5 -ns 8
-    //
-    //       recovery from failures when the batch does not fit in the KV cache will not work correctly until this is fixed
-    //
-    GGML_UNUSED(ubatches);
-    //for (const auto & ubatch : ubatches) {
-    //    if (!find_slot(ubatch)) {
-    //        success = false;
-    //        break;
-    //    }
-    //}
+    for (const auto & ubatch : ubatches) {
+        if (!find_slot(ubatch)) {
+            success = false;
+            break;
+        }
+    }
 
     // restore the original state
     cells = std::move(org_cells);
@@ -419,26 +430,13 @@ bool llama_kv_cache_recurrent::prepare(const std::vector<llama_ubatch> & ubatche
     return success;
 }
 
-bool llama_kv_cache_recurrent::update(llama_context & lctx) {
-    GGML_UNUSED(lctx);
-    // noop
-    return false;
-}
-
-void llama_kv_cache_recurrent::defrag_sched(float thold) {
-    GGML_UNUSED(thold);
-    // noop
-}
-
-bool llama_kv_cache_recurrent::find_slot(const llama_ubatch & ubatch) {
-    const uint32_t n_tokens = ubatch.n_tokens;
-    const uint32_t n_seqs   = ubatch.n_seqs;
-
+bool llama_memory_recurrent::find_slot(const llama_ubatch & ubatch) {
     const uint32_t n_seq_tokens = ubatch.n_seq_tokens;
+    const uint32_t n_seqs       = ubatch.n_seqs;
 
     // 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*n_tokens) {
+    if (head > used + 2*n_seqs) {
         head = 0;
     }
 
@@ -454,9 +452,11 @@ bool llama_kv_cache_recurrent::find_slot(const llama_ubatch & ubatch) {
 
     // 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];
+        const uint32_t i = s*n_seq_tokens; // first token of sequence set s
+        const uint32_t n_seq_id = ubatch.n_seq_id[i];
+
         for (uint32_t j = 0; j < n_seq_id; ++j) {
-            const llama_seq_id seq_id = ubatch.seq_id[s][j];
+            const llama_seq_id seq_id = ubatch.seq_id[i][j];
 
             if (seq_id < 0 || (uint32_t) seq_id >= size) {
                 // too big seq_id
@@ -465,9 +465,9 @@ bool llama_kv_cache_recurrent::find_slot(const llama_ubatch & ubatch) {
                 return false;
             }
             if (j > 0) {
-                kv_cell & seq = cells[seq_id];
+                auto & seq = cells[seq_id];
                 if (seq.tail >= 0) {
-                    kv_cell & cell = cells[seq.tail];
+                    auto & 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);
@@ -487,7 +487,7 @@ bool llama_kv_cache_recurrent::find_slot(const llama_ubatch & ubatch) {
         std::vector<int32_t> tails_verif;
         tails_verif.assign(size, -1);
         for (uint32_t i = 0; i < size; ++i) {
-            kv_cell & cell = cells[i];
+            auto & 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]);
@@ -508,42 +508,43 @@ bool llama_kv_cache_recurrent::find_slot(const llama_ubatch & ubatch) {
 
     for (uint32_t i = 0; i < size; ++i) {
         if (next_empty_cell >= size) { next_empty_cell -= size; }
-        kv_cell & cell = cells[next_empty_cell];
+        auto & 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];
-        kv_cell & seq_meta = cells[seq_id];
+        const uint32_t i = s*n_seq_tokens;
+        const llama_seq_id seq_id = ubatch.seq_id[i][0];
+        auto & seq_meta = cells[seq_id];
         bool has_cell = false;
         if (seq_meta.tail >= 0) {
-            kv_cell & cell = cells[seq_meta.tail];
+            auto & 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) {
-            kv_cell & empty_cell = cells[next_empty_cell];
+            auto & empty_cell = cells[next_empty_cell];
             GGML_ASSERT(empty_cell.is_empty());
             // copy old tail into the empty cell
             if (seq_meta.tail >= 0) {
-                kv_cell & orig_cell = cells[seq_meta.tail];
+                auto & 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
+                GGML_ASSERT(!orig_cell.is_empty()); // has at least one remaining seq_id
             }
             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; }
-                    kv_cell & cell = cells[next_empty_cell];
-                    if (cell.is_empty()) { break; }
+                for (uint32_t j = 0; j < size; ++j) {
                     next_empty_cell += 1;
+                    if (next_empty_cell >= size) { next_empty_cell -= size; }
+                    auto & cell = cells[next_empty_cell];
+                    if (cell.is_empty()) { break; }
                 }
             }
         }
@@ -553,102 +554,99 @@ bool llama_kv_cache_recurrent::find_slot(const llama_ubatch & ubatch) {
 
     // 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;
+        const uint32_t i = s*n_seq_tokens;
+        const int32_t dst_id = s + min;
+        const int32_t src_id = cells[ubatch.seq_id[i][0]].tail;
         if (dst_id != src_id) {
-            kv_cell & dst_cell = cells[dst_id];
-            kv_cell & src_cell = cells[src_id];
+            auto & dst_cell = cells[dst_id];
+            auto & 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;
+            // swap tails
+            for (uint32_t j = 0; j < size; ++j) {
+                int32_t & tail = cells[j].tail;
+                if (tail == src_id) {
+                    tail = dst_id;
+                } else if (tail == dst_id) {
+                    tail = src_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;
-        kv_cell & cell = cells[cell_id];
+        const uint32_t i = s*n_seq_tokens;
+        const llama_pos last_pos = ubatch.pos[i + n_seq_tokens - 1];
+        const int32_t cell_id = s + min;
+        auto & 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);
+                __func__, last_pos, cell.pos, ubatch.seq_id[i][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];
+        for (int32_t j = 0; j < ubatch.n_seq_id[i]; ++j) {
+            const llama_seq_id seq_id = ubatch.seq_id[i][j];
             cell.seq_id.insert(seq_id);
             cells[seq_id].tail = cell_id;
         }
     }
 
+    // Find first cell without src refs, to use as the zero-ed state
+    {
+        // TODO: bake-in src refcounts in the cell metadata
+        std::vector<int32_t> refcounts(size, 0);
+        for (size_t i = 0; i < size; ++i) {
+            const int32_t src = cells[i].src;
+            if (src >= 0) {
+                refcounts[src] += 1;
+            }
+        }
+
+        rs_z = -1;
+        for (int i = min; i <= max; ++i) {
+            if (refcounts[i] == 0) {
+                rs_z = i;
+                break;
+            }
+        }
+
+        for (int i = min; i <= max; ++i) {
+            if (cells[i].src < 0) {
+                GGML_ASSERT(rs_z >= 0);
+                cells[i].src0 = rs_z;
+            } else {
+                // Stage the source ids for all used cells to allow correct seq_* behavior
+                // and still make these values available when setting the inputs
+                cells[i].src0 = cells[i].src;
+            }
+            cells[i].src = i; // avoid moving or clearing twice
+        }
+    }
+
     // 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 kv_cell & cell){ return !cell.is_empty(); });
+        [](const mem_cell & cell){ return !cell.is_empty(); });
 
     // sanity check
     return n >= n_seqs;
 }
 
-bool llama_kv_cache_recurrent::get_can_shift() const {
-    return false;
+bool llama_memory_recurrent::get_can_shift() const {
+    // shifting the pos is trivial for recurrent models
+    return true;
 }
 
-int32_t llama_kv_cache_recurrent::s_copy(int i) const {
-    const uint32_t cell_id = i + head;
-
-    //////////////////////////////////////////////
-    // TODO: this should not mutate the KV cache !
-    kv_cell & cell = const_cast<kv_cell &>(cells[cell_id]);
-
-    // prevent out-of-bound sources
-    if (cell.src < 0 || (uint32_t) cell.src >= size) {
-        cell.src = cell_id;
-    }
-
-    int32_t res = cell.src;
-
-    // TODO: do not mutate the KV cache
-    // ensure copy only happens once
-    if (cell.src != (int32_t) cell_id) {
-        cell.src = cell_id;
-    }
-
-    return res;
-}
-
-float llama_kv_cache_recurrent::s_mask(int i) const {
-    const uint32_t cell_id = i + head;
-
-    //////////////////////////////////////////////
-    // TODO: this should not mutate the KV cache !
-    kv_cell & cell = const_cast<kv_cell &>(cells[cell_id]);
-
-    float res = (float) (cell.src >= 0);
-
-    // only clear once
-    if (cell.src < 0) {
-        cell.src = cell_id;
-    }
-
-    return res;
-}
-
-size_t llama_kv_cache_recurrent::total_size() const {
+size_t llama_memory_recurrent::total_size() const {
     size_t size = 0;
     for (const auto & buf : bufs) {
         size += ggml_backend_buffer_get_size(buf.get());
@@ -657,27 +655,31 @@ size_t llama_kv_cache_recurrent::total_size() const {
     return size;
 }
 
-size_t llama_kv_cache_recurrent::size_k_bytes() const {
-    size_t size_k_bytes = 0;
+size_t llama_memory_recurrent::size_r_bytes() const {
+    size_t size_r_bytes = 0;
 
-    for (const auto & k : k_l) {
-        size_k_bytes += ggml_nbytes(k);
+    for (const auto & r : r_l) {
+        if (r != nullptr) {
+            size_r_bytes += ggml_nbytes(r);
+        }
     }
 
-    return size_k_bytes;
+    return size_r_bytes;
 }
 
-size_t llama_kv_cache_recurrent::size_v_bytes() const {
-    size_t size_v_bytes = 0;
+size_t llama_memory_recurrent::size_s_bytes() const {
+    size_t size_s_bytes = 0;
 
-    for (const auto & v : v_l) {
-        size_v_bytes += ggml_nbytes(v);
+    for (const auto & s : s_l) {
+        if (s != nullptr) {
+            size_s_bytes += ggml_nbytes(s);
+        }
     }
 
-    return size_v_bytes;
+    return size_s_bytes;
 }
 
-void llama_kv_cache_recurrent::state_write(llama_io_write_i & io, llama_seq_id seq_id) const {
+void llama_memory_recurrent::state_write(llama_io_write_i & io, llama_seq_id seq_id) const {
     std::vector<std::pair<uint32_t, uint32_t>> cell_ranges; // ranges, from inclusive, to exclusive
     uint32_t cell_count = 0;
 
@@ -715,7 +717,7 @@ void llama_kv_cache_recurrent::state_write(llama_io_write_i & io, llama_seq_id s
     state_write_data(io, cell_ranges);
 }
 
-void llama_kv_cache_recurrent::state_read(llama_io_read_i & io, llama_seq_id seq_id) {
+void llama_memory_recurrent::state_read(llama_io_read_i & io, llama_seq_id seq_id) {
     uint32_t cell_count;
     io.read_to(&cell_count, sizeof(cell_count));
 
@@ -726,7 +728,7 @@ void llama_kv_cache_recurrent::state_read(llama_io_read_i & io, llama_seq_id seq
 
     if (!res) {
         if (seq_id == -1) {
-            clear();
+            clear(true);
         } else {
             seq_rm(seq_id, -1, -1);
         }
@@ -734,7 +736,7 @@ void llama_kv_cache_recurrent::state_read(llama_io_read_i & io, llama_seq_id seq
     }
 }
 
-void llama_kv_cache_recurrent::state_write_meta(llama_io_write_i & io, const std::vector<std::pair<uint32_t, uint32_t>> & cell_ranges, llama_seq_id seq_id) const {
+void llama_memory_recurrent::state_write_meta(llama_io_write_i & io, const std::vector<std::pair<uint32_t, uint32_t>> & 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];
@@ -753,98 +755,93 @@ void llama_kv_cache_recurrent::state_write_meta(llama_io_write_i & io, const std
     }
 }
 
-void llama_kv_cache_recurrent::state_write_data(llama_io_write_i & io, const std::vector<std::pair<uint32_t, uint32_t>> & cell_ranges) const {
-    const uint32_t v_trans = 0;
+void llama_memory_recurrent::state_write_data(llama_io_write_i & io, const std::vector<std::pair<uint32_t, uint32_t>> & cell_ranges) const {
+    const uint32_t s_trans = 0;
     const uint32_t n_layer = hparams.n_layer;
 
-    io.write(&v_trans, sizeof(v_trans));
-    io.write(&n_layer, sizeof(n_layer));
+    io.write(&s_trans, sizeof(s_trans));
+    io.write(&n_layer,   sizeof(n_layer));
 
     std::vector<uint8_t> 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));
+        const int32_t r_type_i = (int32_t)r_l[il]->type;
+        io.write(&r_type_i, sizeof(r_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));
+        const uint64_t r_size_row = ggml_row_size(r_l[il]->type, hparams.n_embd_r());
+        io.write(&r_size_row, sizeof(r_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);
+            const size_t buf_size = range_size * r_size_row;
+            io.write_tensor(r_l[il], range.first * r_size_row, buf_size);
         }
     }
 
-    if (!v_trans) {
+    if (!s_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));
+            const int32_t s_type_i = (int32_t)s_l[il]->type;
+            io.write(&s_type_i, sizeof(s_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));
+            const uint64_t s_size_row = ggml_row_size(s_l[il]->type, hparams.n_embd_s());
+            io.write(&s_size_row, sizeof(s_size_row));
 
-            // Read each range of cells of v_size length each into tmp_buf and write out
+            // Read each range of cells of s_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);
+                const size_t buf_size = range_size * s_size_row;
+                io.write_tensor(s_l[il], range.first * s_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;
+        const uint32_t mem_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();
+            const uint32_t n_embd_s = hparams.n_embd_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));
+            const int32_t s_type_i = (int32_t)s_l[il]->type;
+            io.write(&s_type_i, sizeof(s_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));
+            const uint32_t s_size_el = ggml_type_size(s_l[il]->type);
+            io.write(&s_size_el, sizeof(s_size_el));
 
             // Write GQA embedding size
-            io.write(&n_embd_v_gqa, sizeof(n_embd_v_gqa));
+            io.write(&n_embd_s, sizeof(n_embd_s));
 
             // For each row, we get the element values of each cell
-            for (uint32_t j = 0; j < n_embd_v_gqa; ++j) {
+            for (uint32_t j = 0; j < n_embd_s; ++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);
+                    const size_t src_offset = (range.first + j * mem_size) * s_size_el;
+                    const size_t buf_size = range_size * s_size_el;
+                    io.write_tensor(s_l[il], src_offset, buf_size);
                 }
             }
         }
     }
 }
 
-bool llama_kv_cache_recurrent::state_read_meta(llama_io_read_i & io, uint32_t cell_count, llama_seq_id dest_seq_id) {
+bool llama_memory_recurrent::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);
+        llama_batch_allocr balloc(hparams.n_pos_per_embd());
 
-        batch.n_tokens = cell_count;
-        batch.n_seq_tokens = cell_count;
-        batch.n_seqs = 1;
+        llama_ubatch ubatch = balloc.ubatch_reserve(cell_count, 1);
 
         for (uint32_t i = 0; i < cell_count; ++i) {
             llama_pos pos;
@@ -858,12 +855,12 @@ bool llama_kv_cache_recurrent::state_read_meta(llama_io_read_i & io, uint32_t ce
                 return false;
             }
 
-            batch.pos[i] = pos;
+            ubatch.pos[i] = pos;
         }
-        batch.n_seq_id[0] = 1;
-        batch.seq_id[0] = &dest_seq_id;
+        ubatch.n_seq_id[0] = 1;
+        ubatch.seq_id[0] = &dest_seq_id;
 
-        if (!find_slot(batch)) {
+        if (!find_slot(ubatch)) {
             LLAMA_LOG_ERROR("%s: failed to find available cells in kv cache\n", __func__);
             return false;
         }
@@ -871,8 +868,8 @@ bool llama_kv_cache_recurrent::state_read_meta(llama_io_read_i & io, uint32_t ce
         // 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].pos == ubatch.pos[0]);
+        GGML_ASSERT(cells[head + cell_count - 1].pos == ubatch.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 {
@@ -883,10 +880,10 @@ bool llama_kv_cache_recurrent::state_read_meta(llama_io_read_i & io, uint32_t ce
             return false;
         }
 
-        clear();
+        clear(true);
 
         for (uint32_t i = 0; i < cell_count; ++i) {
-            kv_cell & cell = cells[i];
+            auto & cell = cells[i];
 
             llama_pos pos;
             uint32_t  n_seq_id;
@@ -900,7 +897,7 @@ bool llama_kv_cache_recurrent::state_read_meta(llama_io_read_i & io, uint32_t ce
                 llama_seq_id seq_id;
                 io.read_to(&seq_id, sizeof(seq_id));
 
-                // TODO: llama_kv_cache_recurrent should have a notion of max sequences
+                // TODO: llama_memory_recurrent 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));
@@ -932,10 +929,10 @@ bool llama_kv_cache_recurrent::state_read_meta(llama_io_read_i & io, uint32_t ce
     return true;
 }
 
-bool llama_kv_cache_recurrent::state_read_data(llama_io_read_i & io, uint32_t cell_count) {
-    uint32_t v_trans;
+bool llama_memory_recurrent::state_read_data(llama_io_read_i & io, uint32_t cell_count) {
+    uint32_t s_trans;
     uint32_t n_layer;
-    io.read_to(&v_trans, sizeof(v_trans));
+    io.read_to(&s_trans, sizeof(s_trans));
     io.read_to(&n_layer, sizeof(n_layer));
 
     if (n_layer != hparams.n_layer) {
@@ -946,102 +943,100 @@ bool llama_kv_cache_recurrent::state_read_data(llama_io_read_i & io, uint32_t ce
         LLAMA_LOG_ERROR("%s: not enough cells in kv cache to restore state (%u > %u)\n", __func__, cell_count, size);
         return false;
     }
-    if (false != (bool) v_trans) {
-        LLAMA_LOG_ERROR("%s: incompatible V transposition\n", __func__);
+    if (false != (bool) s_trans) {
+        LLAMA_LOG_ERROR("%s: incompatible s 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);
+        int32_t r_type_i_ref;
+        io.read_to(&r_type_i_ref, sizeof(r_type_i_ref));
+        const int32_t r_type_i = (int32_t) r_l[il]->type;
+        if (r_type_i != r_type_i_ref) {
+            LLAMA_LOG_ERROR("%s: mismatched r type (%d != %d, layer %d)\n", __func__, r_type_i, r_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);
+        uint64_t r_size_row_ref;
+        io.read_to(&r_size_row_ref, sizeof(r_size_row_ref));
+        const size_t r_size_row = ggml_row_size(r_l[il]->type, hparams.n_embd_r());
+        if (r_size_row != r_size_row_ref) {
+            LLAMA_LOG_ERROR("%s: mismatched r row size (%zu != %zu, layer %d)\n", __func__, r_size_row, (size_t) r_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);
+            ggml_backend_tensor_set(r_l[il], io.read(cell_count * r_size_row), head * r_size_row, cell_count * r_size_row);
         }
     }
 
-    if (!v_trans) {
+    if (!s_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);
+            int32_t s_type_i_ref;
+            io.read_to(&s_type_i_ref, sizeof(s_type_i_ref));
+            const int32_t s_type_i = (int32_t)s_l[il]->type;
+            if (s_type_i != s_type_i_ref) {
+                LLAMA_LOG_ERROR("%s: mismatched s type (%d != %d, layer %d)\n", __func__, s_type_i, s_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);
+            uint64_t s_size_row_ref;
+            io.read_to(&s_size_row_ref, sizeof(s_size_row_ref));
+            const size_t s_size_row = ggml_row_size(s_l[il]->type, hparams.n_embd_s());
+            if (s_size_row != s_size_row_ref) {
+                LLAMA_LOG_ERROR("%s: mismatched s row size (%zu != %zu, layer %d)\n", __func__, s_size_row, (size_t) s_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);
+                ggml_backend_tensor_set(s_l[il], io.read(cell_count * s_size_row), head * s_size_row, cell_count * s_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();
+            const uint32_t n_embd_s = hparams.n_embd_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);
+            int32_t s_type_i_ref;
+            io.read_to(&s_type_i_ref, sizeof(s_type_i_ref));
+            const int32_t s_type_i = (int32_t)s_l[il]->type;
+            if (s_type_i != s_type_i_ref) {
+                LLAMA_LOG_ERROR("%s: mismatched s type (%d != %d, layer %d)\n", __func__, s_type_i, s_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);
+            uint32_t s_size_el_ref;
+            io.read_to(&s_size_el_ref, sizeof(s_size_el_ref));
+            const size_t s_size_el = ggml_type_size(s_l[il]->type);
+            if (s_size_el != s_size_el_ref) {
+                LLAMA_LOG_ERROR("%s: mismatched s element size (%zu != %zu, layer %d)\n", __func__, s_size_el, (size_t) s_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);
+            // Read state embedding size
+            uint32_t n_embd_s_ref;
+            io.read_to(&n_embd_s_ref, sizeof(n_embd_s_ref));
+            if (n_embd_s != n_embd_s_ref) {
+                LLAMA_LOG_ERROR("%s: mismatched s embedding size (%u != %u, layer %d)\n", __func__, n_embd_s, n_embd_s_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);
+                for (uint32_t j = 0; j < n_embd_s; ++j) {
+                    const size_t dst_offset = (head + j * size) * s_size_el;
+                    ggml_backend_tensor_set(s_l[il], io.read(cell_count * s_size_el), dst_offset, cell_count * s_size_el);
                 }
             }
         }
@@ -1051,25 +1046,22 @@ bool llama_kv_cache_recurrent::state_read_data(llama_io_read_i & io, uint32_t ce
 }
 
 //
-// llama_kv_cache_recurrent_state
+// llama_memory_recurrent_context
 //
 
-llama_kv_cache_recurrent_state::llama_kv_cache_recurrent_state(llama_memory_status status) : status(status) {}
+llama_memory_recurrent_context::llama_memory_recurrent_context(llama_memory_status status) : status(status) {}
 
-llama_kv_cache_recurrent_state::llama_kv_cache_recurrent_state(
-        llama_memory_status status,
-        llama_kv_cache_recurrent * kv) : status(status), kv(kv), is_full(true) {
+llama_memory_recurrent_context::llama_memory_recurrent_context(
+        llama_memory_recurrent * mem) : status(LLAMA_MEMORY_STATUS_SUCCESS), mem(mem), is_full(true) {
 }
 
-llama_kv_cache_recurrent_state::llama_kv_cache_recurrent_state(
-        llama_memory_status status,
-        llama_kv_cache_recurrent * kv,
-        llama_sbatch sbatch,
-        std::vector<llama_ubatch> ubatches) : status(status), kv(kv), sbatch(std::move(sbatch)), ubatches(std::move(ubatches)) {}
+llama_memory_recurrent_context::llama_memory_recurrent_context(
+        llama_memory_recurrent * mem,
+        std::vector<llama_ubatch> ubatches) : status(LLAMA_MEMORY_STATUS_SUCCESS), mem(mem), ubatches(std::move(ubatches)) {}
 
-llama_kv_cache_recurrent_state::~llama_kv_cache_recurrent_state() = default;
+llama_memory_recurrent_context::~llama_memory_recurrent_context() = default;
 
-bool llama_kv_cache_recurrent_state::next() {
+bool llama_memory_recurrent_context::next() {
     assert(status == LLAMA_MEMORY_STATUS_SUCCESS);
 
     if (++i_next >= ubatches.size()) {
@@ -1079,54 +1071,56 @@ bool llama_kv_cache_recurrent_state::next() {
     return true;
 }
 
-bool llama_kv_cache_recurrent_state::apply() {
-    assert(status == LLAMA_MEMORY_STATUS_SUCCESS);
+bool llama_memory_recurrent_context::apply() {
+    assert(!llama_memory_status_is_fail(status));
 
-    kv->find_slot(ubatches[i_next]);
+    // no ubatches -> this is an update
+    if (ubatches.empty()) {
+        // recurrent cache never performs updates
+        assert(status == LLAMA_MEMORY_STATUS_NO_UPDATE);
+
+        return true;
+    }
+
+    mem->find_slot(ubatches[i_next]);
 
     return true;
 }
 
-std::vector<int64_t> & llama_kv_cache_recurrent_state::out_ids() {
-    assert(status == LLAMA_MEMORY_STATUS_SUCCESS);
-
-    return sbatch.out_ids;
-}
-
-llama_memory_status llama_kv_cache_recurrent_state::get_status() const {
+llama_memory_status llama_memory_recurrent_context::get_status() const {
     return status;
 }
 
-const llama_ubatch & llama_kv_cache_recurrent_state::get_ubatch() const {
+const llama_ubatch & llama_memory_recurrent_context::get_ubatch() const {
     assert(status == LLAMA_MEMORY_STATUS_SUCCESS);
 
     return ubatches[i_next];
 }
 
-uint32_t llama_kv_cache_recurrent_state::get_n_kv() const {
-    return is_full ? kv->size : kv->n;
+uint32_t llama_memory_recurrent_context::get_n_rs() const {
+    return is_full ? mem->size : mem->n;
 }
 
-uint32_t llama_kv_cache_recurrent_state::get_head() const {
-    return is_full ? 0 : kv->head;
+uint32_t llama_memory_recurrent_context::get_head() const {
+    return is_full ? 0 : mem->head;
 }
 
-uint32_t llama_kv_cache_recurrent_state::get_size() const {
-    return kv->size;
+int32_t llama_memory_recurrent_context::get_rs_z() const {
+    return is_full ? 0 : mem->rs_z;
 }
 
-ggml_tensor * llama_kv_cache_recurrent_state::get_k_l(int32_t il) const {
-    return kv->k_l[il];
+uint32_t llama_memory_recurrent_context::get_size() const {
+    return mem->size;
 }
 
-ggml_tensor * llama_kv_cache_recurrent_state::get_v_l(int32_t il) const {
-    return kv->v_l[il];
+ggml_tensor * llama_memory_recurrent_context::get_r_l(int32_t il) const {
+    return mem->r_l[il];
 }
 
-int32_t llama_kv_cache_recurrent_state::s_copy(int i) const {
-    return kv->s_copy(i);
+ggml_tensor * llama_memory_recurrent_context::get_s_l(int32_t il) const {
+    return mem->s_l[il];
 }
 
-float llama_kv_cache_recurrent_state::s_mask(int i) const {
-    return kv->s_mask(i);
+int32_t llama_memory_recurrent_context::s_copy(int i) const {
+    return  mem->cells[i + mem->head].src0;
 }
diff --git a/src/llama-kv-cache-recurrent.h b/src/llama-memory-recurrent.h
similarity index 57%
rename from src/llama-kv-cache-recurrent.h
rename to src/llama-memory-recurrent.h
index a178ae85c..4d094f9a0 100644
--- a/src/llama-kv-cache-recurrent.h
+++ b/src/llama-memory-recurrent.h
@@ -2,34 +2,48 @@
 
 #include "llama-batch.h"
 #include "llama-graph.h"
-#include "llama-kv-cache.h"
+#include "llama-memory.h"
 
 #include <set>
 #include <vector>
 
 //
-// llama_kv_cache_recurrent
+// llama_memory_recurrent
 //
 
-// TODO: extract the KV cache state used for graph computation into llama_kv_cache_recurrent_state_i
-//       see the implementation of llama_kv_cache_unified_state_i for an example how to do it
-class llama_kv_cache_recurrent : public llama_kv_cache {
+// TODO: extract the cache state used for graph computation into llama_memory_recurrent_context_i
+//       see the implementation of llama_kv_cache_unified_context_i for an example how to do it
+class llama_memory_recurrent : public llama_memory_i {
 public:
-    llama_kv_cache_recurrent(
-            const llama_model & model,
-                    ggml_type   type_k,
-                    ggml_type   type_v,
-                         bool   offload,
-                     uint32_t   kv_size,
-                     uint32_t   n_seq_max);
 
-    ~llama_kv_cache_recurrent() = default;
+    // this callback is used to filter out layers that should not be included in the cache
+    using layer_filter_cb = std::function<bool(int32_t il)>;
+
+    llama_memory_recurrent(
+            const llama_model &  model,
+              layer_filter_cb && filter,
+                    ggml_type    type_r,
+                    ggml_type    type_s,
+                         bool    offload,
+                     uint32_t    mem_size,
+                     uint32_t    n_seq_max);
+
+    ~llama_memory_recurrent() = default;
 
     //
     // llama_memory_i
     //
 
-    void clear() override;
+    llama_memory_context_ptr init_batch(
+            llama_batch_allocr & balloc,
+            uint32_t n_ubatch,
+            bool embd_all) override;
+
+    llama_memory_context_ptr init_full() override;
+
+    llama_memory_context_ptr init_update(llama_context * lctx, bool optimize) override;
+
+    void clear(bool data) 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;
@@ -40,33 +54,13 @@ public:
     llama_pos seq_pos_min(llama_seq_id seq_id) const override;
     llama_pos seq_pos_max(llama_seq_id seq_id) const override;
 
-    //
-    // llama_kv_cache
-    //
-
-    llama_memory_state_ptr init_batch(
-            const llama_batch & batch,
-            uint32_t n_ubatch,
-            bool embd_pooled,
-            bool logits_all) override;
-
-    llama_memory_state_ptr init_full() override;
-
-    bool update(llama_context & lctx) override;
-
-    void defrag_sched(float thold) override;
-
     bool prepare(const std::vector<llama_ubatch> & ubatches);
 
-    // find a contiguous slot of kv cells and emplace the ubatch there
+    // find a contiguous slot of memory cells and emplace the ubatch there
     bool find_slot(const llama_ubatch & ubatch);
 
     bool get_can_shift() const override;
 
-    // TODO: temporary methods - they are not really const as they do const_cast<>, fix this
-    int32_t s_copy(int i) const;
-    float   s_mask(int i) const;
-
     // state write/load
 
     void state_write(llama_io_write_i & io, llama_seq_id seq_id = -1) const override;
@@ -79,10 +73,14 @@ public:
     // computed before each graph build
     uint32_t n = 0;
 
+    // first zero-ed state
+    int32_t rs_z = -1;
+
     // TODO: optimize for recurrent state needs
-    struct kv_cell {
+    struct mem_cell {
         llama_pos pos  = -1;
-        int32_t   src  = -1; // used to copy states
+        int32_t   src  = -1; // used to know where states should be copied from
+        int32_t   src0 = -1; // like src, but only used when setting the inputs (allowing to copy once)
         int32_t   tail = -1;
 
         std::set<llama_seq_id> seq_id;
@@ -95,15 +93,16 @@ public:
             return seq_id.empty();
         }
 
-        bool is_same_seq(const kv_cell & other) const {
+        bool is_same_seq(const mem_cell & other) const {
             return seq_id == other.seq_id;
         }
     };
 
-    std::vector<kv_cell> cells;
+    std::vector<mem_cell> cells;
 
-    std::vector<ggml_tensor *> k_l; // per layer
-    std::vector<ggml_tensor *> v_l;
+    // per layer
+    std::vector<ggml_tensor *> r_l;
+    std::vector<ggml_tensor *> s_l;
 
 private:
     //const llama_model & model;
@@ -116,8 +115,8 @@ private:
 
     size_t total_size() const;
 
-    size_t size_k_bytes() const;
-    size_t size_v_bytes() const;
+    size_t size_r_bytes() const;
+    size_t size_s_bytes() const;
 
     void state_write_meta(llama_io_write_i & io, const std::vector<std::pair<uint32_t, uint32_t>> & cell_ranges, llama_seq_id seq_id = -1) const;
     void state_write_data(llama_io_write_i & io, const std::vector<std::pair<uint32_t, uint32_t>> & cell_ranges) const;
@@ -126,57 +125,50 @@ private:
     bool state_read_data(llama_io_read_i & io, uint32_t cell_count);
 };
 
-class llama_kv_cache_recurrent_state : public llama_memory_state_i {
+class llama_memory_recurrent_context : public llama_memory_context_i {
 public:
     // used for errors
-    llama_kv_cache_recurrent_state(llama_memory_status status);
+    llama_memory_recurrent_context(llama_memory_status status);
 
-    // used to create a full-cache state
-    llama_kv_cache_recurrent_state(
-            llama_memory_status status,
-            llama_kv_cache_recurrent * kv);
+    // used to create a full-cache or update context
+    llama_memory_recurrent_context(
+            llama_memory_recurrent * mem);
 
-    // used to create a state from a batch
-    llama_kv_cache_recurrent_state(
-            llama_memory_status status,
-            llama_kv_cache_recurrent * kv,
-            llama_sbatch sbatch,
+    // used to create a batch processing context from a batch
+    llama_memory_recurrent_context(
+            llama_memory_recurrent * mem,
             std::vector<llama_ubatch> ubatches);
 
-    virtual ~llama_kv_cache_recurrent_state();
+    virtual ~llama_memory_recurrent_context();
 
     //
-    // llama_memory_state_i
+    // llama_memory_context_i
     //
 
     bool next()  override;
     bool apply() override;
 
-    std::vector<int64_t> & out_ids() override;
-
     llama_memory_status  get_status() const override;
     const llama_ubatch & get_ubatch() const override;
 
     //
-    // llama_kv_cache_recurrent_state specific API
+    // llama_memory_recurrent_context specific API
     //
 
-    uint32_t get_n_kv() const;
+    uint32_t get_n_rs() const;
     uint32_t get_head() const;
+    int32_t  get_rs_z() const;
     uint32_t get_size() const;
 
-    ggml_tensor * get_k_l(int32_t il) const;
-    ggml_tensor * get_v_l(int32_t il) const;
+    ggml_tensor * get_r_l(int32_t il) const;
+    ggml_tensor * get_s_l(int32_t il) const;
 
     int32_t s_copy(int i) const;
-    float   s_mask(int i) const;
 
 private:
     const llama_memory_status status;
 
-    llama_kv_cache_recurrent * kv;
-
-    llama_sbatch sbatch;
+    llama_memory_recurrent * mem;
 
     size_t i_next = 0;
 
diff --git a/src/llama-memory.cpp b/src/llama-memory.cpp
index 10173253e..ca6844c32 100644
--- a/src/llama-memory.cpp
+++ b/src/llama-memory.cpp
@@ -1 +1,59 @@
 #include "llama-memory.h"
+
+llama_memory_status llama_memory_status_combine(llama_memory_status s0, llama_memory_status s1) {
+    bool has_update = false;
+
+    switch (s0) {
+        case LLAMA_MEMORY_STATUS_SUCCESS:
+            {
+                has_update = true;
+                break;
+            }
+        case LLAMA_MEMORY_STATUS_NO_UPDATE:
+            {
+                break;
+            }
+        case LLAMA_MEMORY_STATUS_FAILED_PREPARE:
+        case LLAMA_MEMORY_STATUS_FAILED_COMPUTE:
+            {
+                return s0;
+            }
+    }
+
+    switch (s1) {
+        case LLAMA_MEMORY_STATUS_SUCCESS:
+            {
+                has_update = true;
+                break;
+            }
+        case LLAMA_MEMORY_STATUS_NO_UPDATE:
+            {
+                break;
+            }
+        case LLAMA_MEMORY_STATUS_FAILED_PREPARE:
+        case LLAMA_MEMORY_STATUS_FAILED_COMPUTE:
+            {
+                return s1;
+            }
+    }
+
+    // if either status has an update, then the combined status has an update
+    return has_update ? LLAMA_MEMORY_STATUS_SUCCESS : LLAMA_MEMORY_STATUS_NO_UPDATE;
+}
+
+bool llama_memory_status_is_fail(llama_memory_status status) {
+    switch (status) {
+        case LLAMA_MEMORY_STATUS_SUCCESS:
+        case LLAMA_MEMORY_STATUS_NO_UPDATE:
+            {
+                return false;
+            }
+        case LLAMA_MEMORY_STATUS_FAILED_PREPARE:
+        case LLAMA_MEMORY_STATUS_FAILED_COMPUTE:
+            {
+                return true;
+            }
+    }
+
+    return false;
+}
diff --git a/src/llama-memory.h b/src/llama-memory.h
index b3799d66e..e8ba336e8 100644
--- a/src/llama-memory.h
+++ b/src/llama-memory.h
@@ -3,10 +3,14 @@
 #include "llama.h"
 
 #include <memory>
-#include <vector>
 
 struct llama_ubatch;
 
+class llama_batch_allocr;
+
+class llama_io_write_i;
+class llama_io_read_i;
+
 struct llama_memory_params {
     // kv cache
     ggml_type type_k;
@@ -16,13 +20,76 @@ struct llama_memory_params {
     bool swa_full;
 };
 
+enum llama_memory_status {
+    LLAMA_MEMORY_STATUS_SUCCESS = 0,
+    LLAMA_MEMORY_STATUS_NO_UPDATE,
+    LLAMA_MEMORY_STATUS_FAILED_PREPARE,
+    LLAMA_MEMORY_STATUS_FAILED_COMPUTE,
+};
+
+// helper function for combining the status of two memory contexts
+// useful for implementing hybrid memory types (e.g. iSWA)
+llama_memory_status llama_memory_status_combine(llama_memory_status s0, llama_memory_status s1);
+
+// helper function for checking if a memory status indicates a failure
+bool llama_memory_status_is_fail(llama_memory_status status);
+
+// the interface for managing the memory context during batch processing
+// this interface is implemented per memory type. see:
+//   - llama_kv_cache_unified_context
+//   - llama_kv_cache_unified_iswa_context
+//   ...
+//
+// the only method that should mutate the memory and the memory context is llama_memory_i::apply()
+struct llama_memory_context_i {
+    virtual ~llama_memory_context_i() = default;
+
+    // consume the current ubatch from the context and proceed to the next one
+    // return false if we are done
+    virtual bool next() = 0;
+
+    // apply the memory state for the current ubatch to the memory object
+    // return false on failure
+    virtual bool apply() = 0;
+
+    // get the current ubatch
+    virtual const llama_ubatch & get_ubatch() const = 0;
+
+    // get the status of the memory context - used for error handling and checking if any updates would be applied
+    virtual llama_memory_status get_status() const = 0;
+};
+
+using llama_memory_context_ptr = std::unique_ptr<llama_memory_context_i>;
+
 // 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:
+struct llama_memory_i {
     virtual ~llama_memory_i() = default;
 
-    virtual void clear() = 0;
+    // split the input batch into a set of ubatches and verify that they can fit into the cache
+    // return a context object containing the ubatches and memory state required to process them
+    // check the llama_memory_context_i::get_status() for the result
+    virtual llama_memory_context_ptr init_batch(
+            llama_batch_allocr & balloc,
+            uint32_t n_ubatch,
+            bool embd_all) = 0;
+
+    // simulate full cache, used for allocating worst-case compute buffers
+    virtual llama_memory_context_ptr init_full() = 0;
+
+    // prepare for any pending memory updates, such as shifts, defrags, etc.
+    // status == LLAMA_MEMORY_STATUS_NO_UPDATE if there is nothing to update
+    virtual llama_memory_context_ptr init_update(llama_context * lctx, bool optimize) = 0;
+
+    // getters
+    virtual bool get_can_shift() const = 0;
+
+    //
+    // ops
+    //
+
+    // if data == true, the data buffers will also be cleared together with the metadata
+    virtual void clear(bool data) = 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;
@@ -33,44 +100,17 @@ public:
     virtual llama_pos seq_pos_min(llama_seq_id seq_id) const = 0;
     virtual llama_pos seq_pos_max(llama_seq_id seq_id) const = 0;
 
-    virtual bool get_can_edit() const = 0;
+    //
+    // state write/read
+    //
+
+    virtual void state_write(llama_io_write_i & io, llama_seq_id seq_id = -1) const = 0;
+    virtual void state_read (llama_io_read_i  & io, llama_seq_id seq_id = -1) = 0;
 };
 
-enum llama_memory_status {
-    LLAMA_MEMORY_STATUS_SUCCESS = 0,
-    LLAMA_MEMORY_STATUS_FAILED_PREPARE,
-    LLAMA_MEMORY_STATUS_FAILED_COMPUTE,
+using llama_memory_ptr = std::unique_ptr<llama_memory_i>;
+
+// TODO: temporary until the llama_kv_cache is removed from the public API
+struct llama_kv_cache : public llama_memory_i {
+    virtual ~llama_kv_cache() = default;
 };
-
-// the interface for managing the memory state during batch processing
-// this interface is implemented per memory type. see:
-//   - llama_kv_cache_unified_state
-//   - llama_kv_cache_unified_iswa_state
-//   ...
-//
-// the only method that can mutate the memory and the memory state is llama_memory_i::apply()
-//
-// TODO: rename to llama_memory_context_i ?
-class llama_memory_state_i {
-public:
-    virtual ~llama_memory_state_i() = default;
-
-    // consume the current ubatch from the state and proceed to the next one
-    // return false if we are done
-    virtual bool next() = 0;
-
-    // apply the memory state for the current ubatch to the memory object
-    // return false on failure
-    virtual bool apply() = 0;
-
-    // TODO: this might get reworked in the future when refactoring llama_batch
-    virtual std::vector<int64_t> & out_ids() = 0;
-
-    // get the current ubatch
-    virtual const llama_ubatch & get_ubatch() const = 0;
-
-    // get the status of the memory state
-    virtual llama_memory_status get_status() const = 0;
-};
-
-using llama_memory_state_ptr = std::unique_ptr<llama_memory_state_i>;
diff --git a/src/llama-mmap.cpp b/src/llama-mmap.cpp
index 9da97f1bc..47497cf95 100644
--- a/src/llama-mmap.cpp
+++ b/src/llama-mmap.cpp
@@ -401,7 +401,7 @@ struct llama_mmap::impl {
                 }
             }
 #else
-            throw std::runtime_error("PrefetchVirtualMemory unavailable");
+            LLAMA_LOG_DEBUG("skipping PrefetchVirtualMemory because _WIN32_WINNT < 0x602\n");
 #endif
         }
     }
diff --git a/src/llama-model-loader.cpp b/src/llama-model-loader.cpp
index ddb1b0367..bd9e6da88 100644
--- a/src/llama-model-loader.cpp
+++ b/src/llama-model-loader.cpp
@@ -288,9 +288,10 @@ namespace GGUFMeta {
 
     template<typename T>
     bool llama_model_loader::get_arr(const std::string & key, std::vector<T> & result, bool required) {
-        const int kid = gguf_find_key(meta.get(), key.c_str());
+        const gguf_context * ctx = meta.get();
+        const int kid = gguf_find_key(ctx, key.c_str());
 
-        if (kid < 0 || gguf_get_kv_type(meta.get(), kid) != GGUF_TYPE_ARRAY) {
+        if (kid < 0 || gguf_get_kv_type(ctx, kid) != GGUF_TYPE_ARRAY) {
             if (required) {
                 throw std::runtime_error(format("array key not found in model: %s", key.c_str()));
             }
@@ -298,28 +299,40 @@ namespace GGUFMeta {
         }
 
         struct GGUFMeta::ArrayInfo arr_info =
-            GGUFMeta::GKV<GGUFMeta::ArrayInfo>::get_kv(meta.get(), kid);
+            GGUFMeta::GKV<GGUFMeta::ArrayInfo>::get_kv(ctx, kid);
 
         switch (arr_info.gt) {
             case GGUF_TYPE_UINT32:
-            case GGUF_TYPE_INT32:   GGML_ASSERT((std::is_same<T,  int32_t>::value) ||
-                                                (std::is_same<T, uint32_t>::value)); break;
-            case GGUF_TYPE_FLOAT32: GGML_ASSERT((std::is_same<T,    float>::value)); break;
+            case GGUF_TYPE_INT32:   GGML_ASSERT((std::is_same<T,     int32_t>::value) ||
+                                                (std::is_same<T,    uint32_t>::value)); break;
+            case GGUF_TYPE_FLOAT32: GGML_ASSERT((std::is_same<T,       float>::value)); break;
+            case GGUF_TYPE_STRING:  GGML_ASSERT((std::is_same<T, std::string>::value)); break;
             default:
-                throw std::runtime_error(format("%s is not a float32/uint32/int32 array", key.c_str()));
+                throw std::runtime_error(format("%s is not a string/float32/uint32/int32 array", key.c_str()));
         }
 
-        result.resize(arr_info.length);
-        result.assign((const T*)arr_info.data, (const T *)arr_info.data + arr_info.length);
+        if constexpr (std::is_same<T, std::string>::value) {
+            const size_t n_items = gguf_get_arr_n(ctx, kid);
+            result.clear();
+
+            for (size_t i = 0; i < n_items; i++) {
+                const T value = gguf_get_arr_str(ctx, kid, i);
+                result.emplace_back(value);
+            }
+        } else {
+            result.resize(arr_info.length);
+            result.assign((const T*)arr_info.data, (const T *)arr_info.data + arr_info.length);
+        }
 
         return true;
     }
 
     template<typename T, size_t N_MAX>
     bool llama_model_loader::get_arr(const std::string & key, std::array<T, N_MAX> & result, bool required) {
-        const int kid = gguf_find_key(meta.get(), key.c_str());
+        const gguf_context * ctx = meta.get();
+        const int kid = gguf_find_key(ctx, key.c_str());
 
-        if (kid < 0 || gguf_get_kv_type(meta.get(), kid) != GGUF_TYPE_ARRAY) {
+        if (kid < 0 || gguf_get_kv_type(ctx, kid) != GGUF_TYPE_ARRAY) {
             if (required) {
                 throw std::runtime_error(format("array key not found in model: %s", key.c_str()));
             }
@@ -327,22 +340,32 @@ namespace GGUFMeta {
         }
 
         struct GGUFMeta::ArrayInfo arr_info =
-            GGUFMeta::GKV<GGUFMeta::ArrayInfo>::get_kv(meta.get(), kid);
+            GGUFMeta::GKV<GGUFMeta::ArrayInfo>::get_kv(ctx, kid);
 
         switch (arr_info.gt) {
             case GGUF_TYPE_UINT32:
-            case GGUF_TYPE_INT32:   GGML_ASSERT((std::is_same<T,  int32_t>::value) ||
-                                                (std::is_same<T, uint32_t>::value)); break;
-            case GGUF_TYPE_FLOAT32: GGML_ASSERT((std::is_same<T,    float>::value)); break;
+            case GGUF_TYPE_INT32:   GGML_ASSERT((std::is_same<T,     int32_t>::value) ||
+                                                (std::is_same<T,    uint32_t>::value)); break;
+            case GGUF_TYPE_FLOAT32: GGML_ASSERT((std::is_same<T,       float>::value)); break;
+            case GGUF_TYPE_STRING:  GGML_ASSERT((std::is_same<T, std::string>::value)); break;
             default:
-                throw std::runtime_error(format("%s is not a float32/uint32/int32 array", key.c_str()));
+                throw std::runtime_error(format("%s is not a string/float32/uint32/int32 array", key.c_str()));
         }
 
         if (arr_info.length > N_MAX) {
             throw std::runtime_error(format("array length %u for key %s exceeds max %u", (uint32_t) arr_info.length, key.c_str(), (uint32_t) N_MAX));
         }
 
-        std::copy((const T*)arr_info.data, (const T *)arr_info.data + arr_info.length, result.begin());
+        if constexpr (std::is_same<T, std::string>::value) {
+            const size_t n_items = gguf_get_arr_n(ctx, kid);
+
+            for (size_t i = 0; i < n_items; i++) {
+                const T value = gguf_get_arr_str(ctx, kid, i);
+                result[i] = value;
+            }
+        } else {
+            std::copy((const T*)arr_info.data, (const T *)arr_info.data + arr_info.length, result.begin());
+        }
 
         return true;
     }
@@ -352,6 +375,8 @@ namespace GGUFMeta {
         return get_arr(llm_kv(kid), result, required);
     }
 
+    template bool llama_model_loader::get_arr<std::vector<std::string>>(enum llm_kv kid, std::vector<std::string> & result, bool required);
+
     template<typename T>
     bool llama_model_loader::get_key(const std::string & key, T & result, bool required) {
         auto it = kv_overrides.find(key);
diff --git a/src/llama-model-saver.cpp b/src/llama-model-saver.cpp
index a70b98923..563823dc3 100644
--- a/src/llama-model-saver.cpp
+++ b/src/llama-model-saver.cpp
@@ -228,6 +228,7 @@ void llama_model_saver::add_kv_from_model() {
     // add_kv(LLM_KV_TOKENIZER_MASK_ID,                 ???);
     add_kv(LLM_KV_TOKENIZER_ADD_BOS,                 vocab.get_add_bos());
     add_kv(LLM_KV_TOKENIZER_ADD_EOS,                 vocab.get_add_eos());
+    add_kv(LLM_KV_TOKENIZER_ADD_SEP,                 vocab.get_add_sep());
     add_kv(LLM_KV_TOKENIZER_ADD_PREFIX,              vocab.get_add_space_prefix());
     add_kv(LLM_KV_TOKENIZER_REMOVE_EXTRA_WS,         vocab.get_remove_extra_whitespaces());
     add_kv(LLM_KV_TOKENIZER_PRECOMPILED_CHARSMAP,    vocab.get_precompiled_charsmap());
diff --git a/src/llama-model.cpp b/src/llama-model.cpp
index 9ac88c27c..e15a6cb83 100644
--- a/src/llama-model.cpp
+++ b/src/llama-model.cpp
@@ -8,7 +8,8 @@
 
 #include "llama-kv-cache-unified.h"
 #include "llama-kv-cache-unified-iswa.h"
-#include "llama-kv-cache-recurrent.h"
+#include "llama-memory-hybrid.h"
+#include "llama-memory-recurrent.h"
 
 #include "ggml-cpp.h"
 
@@ -47,6 +48,7 @@ const char * llm_type_name(llm_type type) {
         case LLM_TYPE_558M:          return "558M";
         case LLM_TYPE_770M:          return "770M";
         case LLM_TYPE_780M:          return "780M";
+        case LLM_TYPE_0_3B:          return "0.3B";
         case LLM_TYPE_0_5B:          return "0.5B";
         case LLM_TYPE_0_6B:          return "0.6B";
         case LLM_TYPE_1B:            return "1B";
@@ -81,6 +83,7 @@ const char * llm_type_name(llm_type type) {
         case LLM_TYPE_40B:           return "40B";
         case LLM_TYPE_65B:           return "65B";
         case LLM_TYPE_70B:           return "70B";
+        case LLM_TYPE_142B:          return "142B";
         case LLM_TYPE_236B:          return "236B";
         case LLM_TYPE_290B:          return "290B";
         case LLM_TYPE_314B:          return "314B";
@@ -102,6 +105,8 @@ const char * llm_type_name(llm_type type) {
         case LLM_TYPE_17B_128E:      return "17Bx128E (Maverick)";
         case LLM_TYPE_30B_A3B:       return "30B.A3B";
         case LLM_TYPE_235B_A22B:     return "235B.A22B";
+        case LLM_TYPE_E2B:           return "E2B";
+        case LLM_TYPE_E4B:           return "E4B";
         default:                     return "?B";
     }
 }
@@ -204,23 +209,27 @@ static bool weight_buft_supported(const llama_hparams & hparams, ggml_tensor * w
             } break;
         case GGML_OP_SSM_CONV:
             {
-                // FIXME
-                ggml_tensor * conv_x = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, 12345, w->ne[1], 6789);
+                const int64_t n_seq_tokens = 512;
+                const int64_t n_seqs       = 3;
+                ggml_tensor * conv_x = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, w->ne[0] - 1 + n_seq_tokens, w->ne[1], n_seqs);
                 op_tensor = ggml_ssm_conv(ctx, conv_x, w);
             } break;
         case GGML_OP_SSM_SCAN:
             {
-                // FIXME
-                const int64_t d_state      = w->ne[0];
-                const int64_t d_inner      = w->ne[1];
+                // w is ssm_a, which is used to distinguish Mamba-1 and Mamba-2
+                const int64_t d_state      = w->ne[0] == 1 ? hparams.ssm_d_state : w->ne[0];
+                const int64_t n_head       = w->ne[1];
+                const int64_t head_dim     = hparams.ssm_d_inner / n_head;
+                const int64_t n_group      = hparams.ssm_n_group ? hparams.ssm_n_group : 1;
                 const int64_t n_seq_tokens = 512;
-                const int64_t n_seqs       = 1;
-                ggml_tensor * s  = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, d_state, d_inner, n_seqs);
-                ggml_tensor * x = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, d_inner, n_seq_tokens, n_seqs);
-                ggml_tensor * dt = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, d_inner, n_seq_tokens, n_seqs);
-                ggml_tensor * B = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, d_state, n_seq_tokens, n_seqs);
-                ggml_tensor * C = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, d_state, n_seq_tokens, n_seqs);
-                op_tensor = ggml_ssm_scan(ctx, s, x, dt, w, B, C);
+                const int64_t n_seqs       = 3;
+                ggml_tensor * s   = ggml_new_tensor_4d(ctx, GGML_TYPE_F32, d_state, head_dim, n_head, n_seqs);
+                ggml_tensor * x   = ggml_new_tensor_4d(ctx, GGML_TYPE_F32, head_dim, n_head, n_seq_tokens, n_seqs);
+                ggml_tensor * dt  = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, n_head, n_seq_tokens, n_seqs);
+                ggml_tensor * B   = ggml_new_tensor_4d(ctx, GGML_TYPE_F32, d_state, n_group, n_seq_tokens, n_seqs);
+                ggml_tensor * C   = ggml_new_tensor_4d(ctx, GGML_TYPE_F32, d_state, n_group, n_seq_tokens, n_seqs);
+                ggml_tensor * ids = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, n_seqs);
+                op_tensor = ggml_ssm_scan(ctx, s, x, dt, w, B, C, ids);
             } break;
         case GGML_OP_RWKV_WKV6:
             {
@@ -470,6 +479,10 @@ void llama_model::load_hparams(llama_model_loader & ml) {
     std::fill(hparams.n_head_arr.begin(),    hparams.n_head_arr.end(),    0);
     std::fill(hparams.n_head_kv_arr.begin(), hparams.n_head_kv_arr.end(), 0);
     std::fill(hparams.n_ff_arr.begin(),      hparams.n_ff_arr.end(),      0);
+    std::fill(
+        hparams.recurrent_layer_arr.begin(),
+        hparams.recurrent_layer_arr.end(),
+        llm_arch_is_recurrent(ml.get_arch()));
 
     std::fill(hparams.rope_sections.begin(), hparams.rope_sections.end(), 0);
 
@@ -544,6 +557,12 @@ void llama_model::load_hparams(llama_model_loader & ml) {
     uint32_t n_vocab = 0;
     ml.get_key(LLM_KV_VOCAB_SIZE, n_vocab, false) || ml.get_arr_n(LLM_KV_TOKENIZER_LIST, n_vocab, false);
 
+    // for classifier models
+    ml.get_arr(LLM_KV_CLASSIFIER_OUTPUT_LABELS, classifier_labels, false);
+    if (!classifier_labels.empty()) {
+        hparams.n_cls_out = classifier_labels.size();
+    }
+
     // arch-specific KVs
     switch (arch) {
         case LLM_ARCH_LLAMA:
@@ -593,6 +612,16 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                     hparams.use_kq_norm = false;
                 }
             } break;
+        case LLM_ARCH_ARCEE:
+            {
+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+
+                // Arcee uses the same structure as Llama
+                switch (hparams.n_layer) {
+                    case 36: type = LLM_TYPE_4B; break;
+                    default: type = LLM_TYPE_UNKNOWN;
+                }
+            } break;
         case LLM_ARCH_DECI:
             {
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
@@ -687,7 +716,6 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS,    hparams.f_norm_eps);
                 ml.get_key(LLM_KV_ATTENTION_CAUSAL,           hparams.causal_attn);
                 ml.get_key(LLM_KV_POOLING_TYPE,               hparams.pooling_type, false);
-                ml.get_arr_n(LLM_KV_CLASSIFIER_OUTPUT_LABELS, hparams.n_cls_out, false);
 
                 switch (hparams.n_layer) {
                     case 3:
@@ -746,6 +774,16 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                     }
                 }
             } break;
+        case LLM_ARCH_NEO_BERT:
+            {
+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+                ml.get_key(LLM_KV_ATTENTION_CAUSAL,            hparams.causal_attn);
+                ml.get_key(LLM_KV_POOLING_TYPE,                hparams.pooling_type);
+
+                if (hparams.n_layer == 28) {
+                    type = LLM_TYPE_250M;
+                }
+            } break;
         case LLM_ARCH_BLOOM:
             {
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps);
@@ -969,6 +1007,11 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                     case 46: type = LLM_TYPE_27B; break;
                     default: type = LLM_TYPE_UNKNOWN;
                }
+
+                // ref: https://github.com/google/gemma_pytorch/blob/014acb7ac4563a5f77c76d7ff98f31b568c16508/gemma/config.py#L173
+                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_GEMMA3:
             {
@@ -989,10 +1032,29 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                     default: type = LLM_TYPE_UNKNOWN;
                 }
 
+                // ref: https://github.com/google/gemma_pytorch/blob/014acb7ac4563a5f77c76d7ff98f31b568c16508/gemma/config.py#L289
                 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_GEMMA3N:
+            {
+                hparams.swa_type = LLAMA_SWA_TYPE_STANDARD;
+                hparams.set_swa_pattern(5);
+
+                hparams.rope_freq_base_train_swa  = 10000.0f;
+                hparams.rope_freq_scale_train_swa = 1.0f;
+                hparams.f_attention_scale         = 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 30: type = LLM_TYPE_E2B; break;
+                    case 35: type = LLM_TYPE_E4B; break;
+                    default: type = LLM_TYPE_UNKNOWN;
+                }
+            } break;
         case LLM_ARCH_STARCODER2:
             {
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps);
@@ -1036,6 +1098,38 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                     default: type = LLM_TYPE_UNKNOWN;
                 }
             } break;
+        case LLM_ARCH_MAMBA2:
+            {
+                ml.get_key(LLM_KV_SSM_CONV_KERNEL,    hparams.ssm_d_conv);
+                ml.get_key(LLM_KV_SSM_INNER_SIZE,     hparams.ssm_d_inner);
+                ml.get_key(LLM_KV_SSM_STATE_SIZE,     hparams.ssm_d_state);
+                ml.get_key(LLM_KV_SSM_TIME_STEP_RANK, hparams.ssm_dt_rank);
+                ml.get_key(LLM_KV_SSM_GROUP_COUNT,    hparams.ssm_n_group);
+
+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+
+                switch (hparams.n_layer) {
+                    case 24:
+                        switch (hparams.n_embd) {
+                            case 768: type = LLM_TYPE_SMALL; break;
+                            default: type = LLM_TYPE_UNKNOWN;
+                        } break;
+                    case 48:
+                        switch (hparams.n_embd) {
+                            case 1024: type = LLM_TYPE_MEDIUM; break;
+                            case 1536: type = LLM_TYPE_LARGE; break;
+                            case 2048: type = LLM_TYPE_XL; break;
+                            default: type = LLM_TYPE_UNKNOWN;
+                        } break;
+                    case 64:
+                        switch (hparams.n_embd) {
+                            case 2560: type = LLM_TYPE_3B; break;
+                            case 4096: type = LLM_TYPE_7B; break;
+                            default: type = LLM_TYPE_UNKNOWN;
+                        } break;
+                    default: type = LLM_TYPE_UNKNOWN;
+                }
+            } break;
         case LLM_ARCH_XVERSE:
             {
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
@@ -1446,6 +1540,28 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                     default: type = LLM_TYPE_UNKNOWN;
                 }
             } break;
+        case LLM_ARCH_DOTS1:
+            {
+                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);
+                ml.get_key(LLM_KV_EXPERT_GATING_FUNC,          hparams.expert_gating_func, false);
+                switch (hparams.n_layer) {
+                    case 62: type = LLM_TYPE_142B; break;
+                    default: type = LLM_TYPE_UNKNOWN;
+                }
+            } break;
+        case LLM_ARCH_ERNIE4_5:
+            {
+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+                switch (hparams.n_layer) {
+                    case 18: type = LLM_TYPE_0_3B; break;
+                    default: type = LLM_TYPE_UNKNOWN;
+                }
+            } break;
         default: throw std::runtime_error("unsupported model architecture");
     }
 
@@ -2188,6 +2304,32 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                         layer.layer_out_norm_b = create_tensor(tn(LLM_TENSOR_LAYER_OUT_NORM, "bias", i),   {n_embd}, 0);
                     }
                 } break;
+            case LLM_ARCH_NEO_BERT:
+                {
+                    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD,  "weight"), {n_embd, n_vocab}, 0);
+
+                    cls   = create_tensor(tn(LLM_TENSOR_CLS, "weight"), {n_embd, n_embd}, TENSOR_NOT_REQUIRED);
+                    cls_b = create_tensor(tn(LLM_TENSOR_CLS, "bias"),   {n_embd},         TENSOR_NOT_REQUIRED);
+
+                    cls_out   = create_tensor(tn(LLM_TENSOR_CLS_OUT, "weight"), {n_embd, hparams.n_cls_out}, TENSOR_NOT_REQUIRED);
+                    cls_out_b = create_tensor(tn(LLM_TENSOR_CLS_OUT, "bias"),   {hparams.n_cls_out},         TENSOR_NOT_REQUIRED);
+
+                    output_norm_enc = create_tensor(tn(LLM_TENSOR_ENC_OUTPUT_NORM, "weight"), {n_embd}, 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.wqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}, 0);
+                        layer.wo   = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);
+
+                        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
+
+                        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd, n_ff*2}, 0);
+                        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, 0);
+                    }
+                } break;
             case LLM_ARCH_JINA_BERT_V2:
                 {
                     tok_embd  = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD,  "weight"), {n_embd, n_vocab}, 0); // word_embeddings
@@ -2225,8 +2367,8 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                         layer.attn_norm_2   = create_tensor(tn(LLM_TENSOR_ATTN_NORM_2, "weight", i), {n_embd}, TENSOR_NOT_REQUIRED);
                         layer.attn_norm_2_b = create_tensor(tn(LLM_TENSOR_ATTN_NORM_2, "bias",   i), {n_embd}, TENSOR_NOT_REQUIRED);
 
-                        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd, n_ff}, 0);
-                        layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, 0);
+                        layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, TENSOR_NOT_REQUIRED);
+                        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd, layer.ffn_gate ? n_ff : n_ff * 2}, 0);
 
                         layer.ffn_down   = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, 0);
                         layer.ffn_down_b = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "bias",   i), {n_embd}, 0);
@@ -2885,6 +3027,62 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                         layer.ffn_post_norm = create_tensor(tn(LLM_TENSOR_FFN_POST_NORM, "weight", i), {n_embd}, 0);
                     }
                 } break;
+            case LLM_ARCH_GEMMA3N:
+                {
+                    const int64_t n_altup      = hparams.n_altup;
+                    const int64_t laurel_rank  = hparams.laurel_rank;
+                    const int64_t n_embd_altup = hparams.n_embd_altup;
+
+                    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);
+                    }
+
+                    tok_embd           = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD,           "weight"), {n_embd, n_vocab}, 0);
+                    tok_embd_per_layer = create_tensor(tn(LLM_TENSOR_PER_LAYER_TOKEN_EMBD, "weight"), {n_embd_altup * n_layer, n_vocab}, 0);
+
+                    altup_proj           = create_tensor(tn(LLM_TENSOR_ALTUP_PROJ,           "weight"), {n_embd, n_embd, n_altup - 1}, 0);
+                    altup_unembd_proj    = create_tensor(tn(LLM_TENSOR_ALTUP_UNEMBD_PROJ,    "weight"), {n_embd, n_embd, n_altup - 1}, 0);
+                    per_layer_model_proj = create_tensor(tn(LLM_TENSOR_PER_LAYER_MODEL_PROJ, "weight"), {n_embd, n_embd_altup * n_layer}, 0);
+                    per_layer_proj_norm  = create_tensor(tn(LLM_TENSOR_PER_LAYER_PROJ_NORM,  "weight"), {n_embd_altup}, 0);
+
+                    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 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_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_q_norm    = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM,    "weight", i), {n_embd_head_k}, 0);
+                        layer.attn_k_norm    = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM,    "weight", i), {n_embd_head_k}, 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);
+
+                        // altup & laurel
+                        layer.per_layer_inp_gate   = create_tensor(tn(LLM_TENSOR_PER_LAYER_INP_GATE,  "weight", i), {n_embd, n_embd_altup}, 0);
+                        layer.per_layer_proj       = create_tensor(tn(LLM_TENSOR_PER_LAYER_PROJ,      "weight", i), {n_embd_altup, n_embd}, 0);
+                        layer.per_layer_post_norm  = create_tensor(tn(LLM_TENSOR_PER_LAYER_POST_NORM, "weight", i), {n_embd}, 0);
+                        layer.altup_correct_coef   = create_tensor(tn(LLM_TENSOR_ALTUP_CORRECT_COEF,  "weight", i), {n_altup, n_altup}, 0);
+                        layer.altup_correct_scale  = create_tensor(tn(LLM_TENSOR_ALTUP_CORRECT_SCALE, "weight", i), {n_embd}, 0);
+                        layer.altup_predict_coef   = create_tensor(tn(LLM_TENSOR_ALTUP_PREDICT_COEF,  "weight", i), {n_altup, n_altup * n_altup}, 0);
+                        layer.altup_router         = create_tensor(tn(LLM_TENSOR_ALTUP_ROUTER,        "weight", i), {n_embd, n_altup}, 0);
+                        layer.altup_router_norm    = create_tensor(tn(LLM_TENSOR_ALTUP_ROUTER_NORM,   "weight", i), {n_embd}, 0);
+                        layer.laurel_l             = create_tensor(tn(LLM_TENSOR_LAUREL_L,            "weight", i), {n_embd, laurel_rank}, 0);
+                        layer.laurel_r             = create_tensor(tn(LLM_TENSOR_LAUREL_R,            "weight", i), {laurel_rank, n_embd}, 0);
+                        layer.laurel_post_norm     = create_tensor(tn(LLM_TENSOR_LAUREL_POST_NORM,    "weight", i), {n_embd}, 0);
+                    }
+                } break;
             case LLM_ARCH_STARCODER2:
                 {
                     tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
@@ -2970,6 +3168,54 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                         layer.ssm_a = create_tensor(tn(LLM_TENSOR_SSM_A, i), {d_state, d_inner}, 0);
                         layer.ssm_d = create_tensor(tn(LLM_TENSOR_SSM_D, i), {d_inner}, 0);
 
+                        // out_proj
+                        layer.ssm_out = create_tensor(tn(LLM_TENSOR_SSM_OUT, "weight", i), {d_inner, n_embd}, 0);
+                    }
+                } break;
+            case LLM_ARCH_MAMBA2:
+                {
+                    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 n_head  = hparams.ssm_dt_rank;
+                    const int64_t n_group = hparams.ssm_n_group;
+                    const int64_t d_in_proj = 2*d_inner + 2*n_group*d_state + n_head;
+
+                    // only an expansion factor of 2 is supported for now
+                    GGML_ASSERT(2 * n_embd == d_inner);
+
+                    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}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                        // if output is NULL, init from the input tok embed, duplicated to allow offloading
+                        if (output == NULL) {
+                            output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_DUPLICATED);
+                        }
+                    }
+
+                    for (int i = 0; i < n_layer; ++i) {
+                        auto & layer = layers[i];
+
+                        // norm
+                        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+
+                        layer.ssm_in = create_tensor(tn(LLM_TENSOR_SSM_IN, "weight", i), {n_embd, d_in_proj}, 0);
+
+                        layer.ssm_conv1d = create_tensor(tn(LLM_TENSOR_SSM_CONV1D, "weight", i), {d_conv, d_inner + 2*n_group*d_state}, 0);
+                        layer.ssm_conv1d_b = create_tensor(tn(LLM_TENSOR_SSM_CONV1D, "bias", i), {d_inner + 2*n_group*d_state}, 0);
+
+                        layer.ssm_dt_b = create_tensor(tn(LLM_TENSOR_SSM_DT, "bias", i), {n_head}, 0);
+
+                        // no "weight" suffix for these
+                        layer.ssm_a = create_tensor(tn(LLM_TENSOR_SSM_A, i), {1, n_head}, 0);
+                        layer.ssm_d = create_tensor(tn(LLM_TENSOR_SSM_D, i), {1, n_head}, 0);
+
+                        layer.ssm_norm = create_tensor(tn(LLM_TENSOR_SSM_NORM, "weight", i), {d_inner / n_group, n_group}, 0);
+
                         // out_proj
                         layer.ssm_out = create_tensor(tn(LLM_TENSOR_SSM_OUT, "weight", i), {d_inner, n_embd}, 0);
                     }
@@ -4124,6 +4370,123 @@ 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_exp * n_expert_shared}, 0);
                     }
                 } break;
+            case LLM_ARCH_DOTS1:
+                {
+                    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_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_head_k * n_head}, 0);
+                        layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V,   "weight", i), {n_embd, n_embd_head_k * n_head}, 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);
+
+                        if (i < (int) hparams.n_layer_dense_lead) {
+                            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);
+                        } else {
+                            layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}, 0);
+                            layer.ffn_exp_probs_b = create_tensor(tn(LLM_TENSOR_FFN_EXP_PROBS_B, "bias", i), {n_expert}, TENSOR_NOT_REQUIRED);
+
+                            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");
+                            }
+
+                            // MoE branch
+                            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 branch
+                            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;
+            case LLM_ARCH_ARCEE:
+                {
+                    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.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));
+
+                        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_ERNIE4_5:
+                {
+                    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);
+
+                        // optional bias tensors
+                        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.bo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd},     TENSOR_NOT_REQUIRED);
+
+                        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;
             default:
                 throw std::runtime_error("unknown architecture");
         }
@@ -4363,11 +4726,24 @@ void llama_model::print_info() const {
         LLAMA_LOG_INFO("%s: freq_scale_train = %g\n",     __func__, hparams.rope_freq_scale_train);
         LLAMA_LOG_INFO("%s: n_ctx_orig_yarn  = %u\n",     __func__, hparams.n_ctx_orig_yarn);
         LLAMA_LOG_INFO("%s: rope_finetuned   = %s\n",     __func__, hparams.rope_finetuned ? "yes" : "unknown");
+    }
+
+    if (arch == LLM_ARCH_MAMBA || arch == LLM_ARCH_MAMBA2) {
         LLAMA_LOG_INFO("%s: ssm_d_conv       = %u\n",     __func__, hparams.ssm_d_conv);
         LLAMA_LOG_INFO("%s: ssm_d_inner      = %u\n",     __func__, hparams.ssm_d_inner);
         LLAMA_LOG_INFO("%s: ssm_d_state      = %u\n",     __func__, hparams.ssm_d_state);
         LLAMA_LOG_INFO("%s: ssm_dt_rank      = %u\n",     __func__, hparams.ssm_dt_rank);
+        LLAMA_LOG_INFO("%s: ssm_n_group      = %u\n",     __func__, hparams.ssm_n_group);
         LLAMA_LOG_INFO("%s: ssm_dt_b_c_rms   = %d\n",     __func__, hparams.ssm_dt_b_c_rms);
+
+        if (!classifier_labels.empty()) {
+            LLAMA_LOG_INFO("%s: n_cls_out        = %u\n", __func__, hparams.n_cls_out);
+
+            size_t i = 0;
+            for (auto label : classifier_labels) {
+                LLAMA_LOG_INFO("%s: cls_label[%2zu]    = %s\n", __func__, i++, label.c_str());
+            }
+        }
     }
 
     LLAMA_LOG_INFO("%s: model type       = %s\n",     __func__, type_name().c_str());
@@ -4550,6 +4926,8 @@ struct llm_build_llama : public llm_graph_context {
 
         const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
 
+        ggml_tensor * inp_out_ids = build_inp_out_ids();
+
         for (int il = 0; il < n_layer; ++il) {
             ggml_tensor * inpSA = inpL;
 
@@ -4612,9 +4990,7 @@ struct llm_build_llama : public llm_graph_context {
                 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();
+            if (il == n_layer - 1 && inp_out_ids) {
                 cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
                 inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
             }
@@ -4710,6 +5086,8 @@ struct llm_build_llama_iswa : public llm_graph_context {
 
         const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
 
+        ggml_tensor * inp_out_ids = build_inp_out_ids();
+
         for (int il = 0; il < n_layer; ++il) {
             ggml_tensor * inpSA = inpL;
 
@@ -4786,9 +5164,7 @@ struct llm_build_llama_iswa : public llm_graph_context {
                 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();
+            if (il == n_layer - 1 && inp_out_ids) {
                 cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
                 inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
             }
@@ -4888,6 +5264,9 @@ struct llm_build_deci : public llm_graph_context {
         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;
+
+        ggml_tensor * inp_out_ids = build_inp_out_ids();
+
         for (int il = 0; il < n_layer; ++il) {
             ggml_tensor * inpSA = inpL;
             const int64_t n_head_kv = hparams.n_head_kv(il);
@@ -4961,9 +5340,7 @@ struct llm_build_deci : public llm_graph_context {
                         Qcur, Kcur, Vcur, nullptr, nullptr, kq_scale, il);
             }
 
-            if (il == n_layer - 1) {
-                // skip computing output for unused tokens
-                ggml_tensor * inp_out_ids = build_inp_out_ids();
+            if (il == n_layer - 1 && inp_out_ids) {
                 cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
                 inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
             }
@@ -5042,6 +5419,8 @@ struct llm_build_baichuan : public llm_graph_context {
 
         auto * inp_attn = build_attn_inp_kv_unified();
 
+        ggml_tensor * inp_out_ids = build_inp_out_ids();
+
         for (int il = 0; il < n_layer; ++il) {
             ggml_tensor * inpSA = inpL;
 
@@ -5093,9 +5472,7 @@ struct llm_build_baichuan : public llm_graph_context {
                         Qcur, Kcur, Vcur, nullptr, 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();
+            if (il == n_layer - 1 && inp_out_ids) {
                 cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
                 inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
             }
@@ -5164,6 +5541,8 @@ struct llm_build_xverse : public llm_graph_context {
 
         auto * inp_attn = build_attn_inp_kv_unified();
 
+        ggml_tensor * inp_out_ids = build_inp_out_ids();
+
         for (int il = 0; il < n_layer; ++il) {
             ggml_tensor * inpSA = inpL;
 
@@ -5208,9 +5587,7 @@ struct llm_build_xverse : public llm_graph_context {
                         Qcur, Kcur, Vcur, nullptr, 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();
+            if (il == n_layer - 1 && inp_out_ids) {
                 cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
                 inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
             }
@@ -5278,6 +5655,8 @@ struct llm_build_falcon : public llm_graph_context {
 
         auto * inp_attn = build_attn_inp_kv_unified();
 
+        ggml_tensor * inp_out_ids = build_inp_out_ids();
+
         for (int il = 0; il < n_layer; ++il) {
             ggml_tensor * attn_norm;
 
@@ -5333,9 +5712,7 @@ struct llm_build_falcon : public llm_graph_context {
                         Qcur, Kcur, Vcur, nullptr, 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();
+            if (il == n_layer - 1 && 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);
@@ -5404,6 +5781,8 @@ struct llm_build_grok : public llm_graph_context {
 
         auto * inp_attn = build_attn_inp_kv_unified();
 
+        ggml_tensor * inp_out_ids = build_inp_out_ids();
+
         for (int il = 0; il < n_layer; ++il) {
             ggml_tensor * inpSA = inpL;
 
@@ -5463,9 +5842,7 @@ struct llm_build_grok : public llm_graph_context {
                         Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f, il);
             }
 
-            if (il == n_layer - 1) {
-                // skip computing output for unused tokens
-                ggml_tensor * inp_out_ids = build_inp_out_ids();
+            if (il == n_layer - 1 && inp_out_ids) {
                 cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
                 inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
             }
@@ -5564,6 +5941,8 @@ struct llm_build_dbrx : public llm_graph_context {
 
         auto * inp_attn = build_attn_inp_kv_unified();
 
+        ggml_tensor * inp_out_ids = build_inp_out_ids();
+
         for (int il = 0; il < n_layer; ++il) {
             ggml_tensor * inpSA = inpL;
 
@@ -5614,9 +5993,7 @@ struct llm_build_dbrx : public llm_graph_context {
                         Qcur, Kcur, Vcur, nullptr, 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();
+            if (il == n_layer - 1 && inp_out_ids) {
                 cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
                 inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
             }
@@ -5696,6 +6073,8 @@ struct llm_build_starcoder : public llm_graph_context {
         inpL = ggml_add(ctx0, inpL, pos);
         cb(inpL, "inpL", -1);
 
+        ggml_tensor * inp_out_ids = build_inp_out_ids();
+
         for (int il = 0; il < n_layer; ++il) {
             cur = build_norm(inpL,
                     model.layers[il].attn_norm,
@@ -5728,9 +6107,7 @@ struct llm_build_starcoder : public llm_graph_context {
                         Qcur, Kcur, Vcur, nullptr, 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();
+            if (il == n_layer - 1 && inp_out_ids) {
                 cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
                 inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
             }
@@ -5795,6 +6172,8 @@ struct llm_build_refact : public llm_graph_context {
 
         auto * inp_attn = build_attn_inp_kv_unified();
 
+        ggml_tensor * inp_out_ids = build_inp_out_ids();
+
         for (int il = 0; il < n_layer; ++il) {
             ggml_tensor * inpSA = inpL;
 
@@ -5827,9 +6206,7 @@ struct llm_build_refact : public llm_graph_context {
                         Qcur, Kcur, Vcur, nullptr, 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();
+            if (il == n_layer - 1 && inp_out_ids) {
                 cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
                 inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
             }
@@ -5915,78 +6292,79 @@ struct llm_build_bert : public llm_graph_context {
 
         auto * inp_attn = build_attn_inp_no_cache();
 
-        // iterate layers
+        ggml_tensor * inp_out_ids = build_inp_out_ids();
+
         for (int il = 0; il < n_layer; ++il) {
             ggml_tensor * cur = inpL;
 
-            ggml_tensor * Qcur;
-            ggml_tensor * Kcur;
-            ggml_tensor * Vcur;
+            {
+                ggml_tensor * Qcur;
+                ggml_tensor * Kcur;
+                ggml_tensor * Vcur;
 
-            // self-attention
-            if (model.layers[il].wqkv) {
-                cur = build_lora_mm(model.layers[il].wqkv, cur);
-                cb(cur, "wqkv", il);
+                // self-attention
+                if (model.layers[il].wqkv) {
+                    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 (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)));
+                } else {
+                    Qcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wq, cur), model.layers[il].bq);
+                    Kcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wk, cur), model.layers[il].bk);
+                    Vcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wv, cur), model.layers[il].bv);
                 }
 
-                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, cur), model.layers[il].bq);
-                Kcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wk, cur), model.layers[il].bk);
-                Vcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wv, cur), model.layers[il].bv);
+                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);
+                }
+
+                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);
+                }
+
+                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);
+
+                // RoPE
+                if (model.arch == LLM_ARCH_NOMIC_BERT || model.arch == LLM_ARCH_NOMIC_BERT_MOE || model.arch == LLM_ARCH_JINA_BERT_V3) {
+                    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, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                cb(cur, "kqv_out", il);
             }
 
-            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);
-            }
-
-            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);
-            }
-
-            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);
-
-            // RoPE
-            if (model.arch == LLM_ARCH_NOMIC_BERT || model.arch == LLM_ARCH_NOMIC_BERT_MOE || model.arch == LLM_ARCH_JINA_BERT_V3) {
-                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, 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();
+            if (il == n_layer - 1 && inp_out_ids) {
                 cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
                 inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
             }
@@ -6035,7 +6413,7 @@ struct llm_build_bert : public llm_graph_context {
                         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);
+                        model.layers[il].ffn_gate ? LLM_FFN_GELU : LLM_FFN_GEGLU, LLM_FFN_PAR, il);
                 cb(cur, "ffn_out", il);
             } else {
                 cur = build_ffn(cur,
@@ -6066,6 +6444,118 @@ struct llm_build_bert : public llm_graph_context {
     }
 };
 
+struct llm_build_neo_bert : public llm_graph_context {
+    llm_build_neo_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 = build_inp_pos();
+
+        // construct input embeddings (token, type, position)
+        inpL = build_inp_embd(model.tok_embd);
+        cb(inpL, "inp_embd", -1);
+
+        auto * inp_attn = build_attn_inp_no_cache();
+
+        ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+        for (int il = 0; il < n_layer; ++il) {
+            ggml_tensor * cur = inpL;
+
+            // pre-norm
+            cur = build_norm(inpL,
+                    model.layers[il].attn_norm, NULL,
+                    LLM_NORM_RMS, il);
+
+            {
+                ggml_tensor * Qcur;
+                ggml_tensor * Kcur;
+                ggml_tensor * Vcur;
+
+                // self-attention
+                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);
+
+                // RoPE
+                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, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                cb(cur, "kqv_out", il);
+            }
+
+            if (il == n_layer - 1 && 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);
+
+            ggml_tensor * ffn_inp = cur;
+            cb(ffn_inp, "ffn_inp", il);
+
+            // pre-norm
+            cur = build_norm(ffn_inp,
+                    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, NULL, NULL, NULL,
+                    model.layers[il].ffn_down,
+                    NULL, NULL, NULL,
+                    LLM_FFN_SWIGLU, LLM_FFN_SEQ, il);
+
+            // attentions bypass the intermediate layer
+            cur = ggml_add(ctx0, cur, ffn_inp);
+
+            // input for next layer
+            inpL = cur;
+        }
+
+        cur = inpL;
+
+        cur = build_norm(cur,
+                model.output_norm_enc, NULL,
+                LLM_NORM_RMS, -1);
+
+        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;
@@ -6086,6 +6576,8 @@ struct llm_build_bloom : public llm_graph_context {
                 LLM_NORM, -1);
         cb(inpL, "inp_norm", -1);
 
+        ggml_tensor * inp_out_ids = build_inp_out_ids();
+
         for (int il = 0; il < n_layer; ++il) {
             cur = build_norm(inpL,
                     model.layers[il].attn_norm,
@@ -6118,9 +6610,7 @@ struct llm_build_bloom : public llm_graph_context {
                         Qcur, Kcur, Vcur, nullptr, 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();
+            if (il == n_layer - 1 && inp_out_ids) {
                 cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
                 inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
             }
@@ -6197,6 +6687,8 @@ struct llm_build_mpt : public llm_graph_context {
             cb(inpL, "inpL", -1);
         }
 
+        ggml_tensor * inp_out_ids = build_inp_out_ids();
+
         for (int il = 0; il < n_layer; ++il) {
             ggml_tensor * attn_norm;
 
@@ -6259,9 +6751,7 @@ struct llm_build_mpt : public llm_graph_context {
                         Qcur, Kcur, Vcur, nullptr, 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();
+            if (il == n_layer - 1 && inp_out_ids) {
                 cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
                 inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
             }
@@ -6330,6 +6820,8 @@ struct llm_build_stablelm : public llm_graph_context {
 
         auto * inp_attn = build_attn_inp_kv_unified();
 
+        ggml_tensor * inp_out_ids = build_inp_out_ids();
+
         for (int il = 0; il < n_layer; ++il) {
             // norm
             cur = build_norm(inpL,
@@ -6405,9 +6897,7 @@ struct llm_build_stablelm : public llm_graph_context {
                         Qcur, Kcur, Vcur, nullptr, 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();
+            if (il == n_layer - 1 && 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);
@@ -6482,6 +6972,8 @@ struct llm_build_qwen : public llm_graph_context {
 
         auto * inp_attn = build_attn_inp_kv_unified();
 
+        ggml_tensor * inp_out_ids = build_inp_out_ids();
+
         for (int il = 0; il < n_layer; ++il) {
             ggml_tensor * inpSA = inpL;
 
@@ -6528,9 +7020,7 @@ struct llm_build_qwen : public llm_graph_context {
                         Qcur, Kcur, Vcur, nullptr, 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();
+            if (il == n_layer - 1 && inp_out_ids) {
                 cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
                 inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
             }
@@ -6599,6 +7089,8 @@ struct llm_build_qwen2 : public llm_graph_context {
 
         auto * inp_attn = build_attn_inp_kv_unified();
 
+        ggml_tensor * inp_out_ids = build_inp_out_ids();
+
         for (int il = 0; il < n_layer; ++il) {
             ggml_tensor * inpSA = inpL;
 
@@ -6648,9 +7140,7 @@ struct llm_build_qwen2 : public llm_graph_context {
                         Qcur, Kcur, Vcur, nullptr, 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();
+            if (il == n_layer - 1 && inp_out_ids) {
                 cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
                 inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
             }
@@ -6720,6 +7210,8 @@ struct llm_build_qwen2vl : public llm_graph_context {
         int sections[4];
         std::copy(std::begin(hparams.rope_sections), std::begin(hparams.rope_sections) + 4, sections);
 
+        ggml_tensor * inp_out_ids = build_inp_out_ids();
+
         for (int il = 0; il < n_layer; ++il) {
             ggml_tensor * inpSA = inpL;
 
@@ -6769,9 +7261,7 @@ struct llm_build_qwen2vl : public llm_graph_context {
                         Qcur, Kcur, Vcur, nullptr, 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();
+            if (il == n_layer - 1 && inp_out_ids) {
                 cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
                 inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
             }
@@ -6838,6 +7328,8 @@ struct llm_build_qwen2moe : public llm_graph_context {
 
         auto * inp_attn = build_attn_inp_kv_unified();
 
+        ggml_tensor * inp_out_ids = build_inp_out_ids();
+
         for (int il = 0; il < n_layer; ++il) {
             ggml_tensor * inpSA = inpL;
 
@@ -6896,9 +7388,7 @@ struct llm_build_qwen2moe : public llm_graph_context {
                         Qcur, Kcur, Vcur, nullptr, 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();
+            if (il == n_layer - 1 && inp_out_ids) {
                 cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
                 inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
             }
@@ -6997,6 +7487,8 @@ struct llm_build_qwen3 : public llm_graph_context {
 
         auto * inp_attn = build_attn_inp_kv_unified();
 
+        ggml_tensor * inp_out_ids = build_inp_out_ids();
+
         for (int il = 0; il < n_layer; ++il) {
             ggml_tensor * inpSA = inpL;
 
@@ -7049,9 +7541,7 @@ struct llm_build_qwen3 : public llm_graph_context {
                         Qcur, Kcur, Vcur, nullptr, 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();
+            if (il == n_layer - 1 && inp_out_ids) {
                 cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
                 inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
             }
@@ -7118,6 +7608,8 @@ struct llm_build_qwen3moe : public llm_graph_context {
 
         auto * inp_attn = build_attn_inp_kv_unified();
 
+        ggml_tensor * inp_out_ids = build_inp_out_ids();
+
         for (int il = 0; il < n_layer; ++il) {
             ggml_tensor * inpSA = inpL;
 
@@ -7170,9 +7662,7 @@ struct llm_build_qwen3moe : public llm_graph_context {
                         Qcur, Kcur, Vcur, nullptr, 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();
+            if (il == n_layer - 1 && inp_out_ids) {
                 cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
                 inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
             }
@@ -7248,6 +7738,8 @@ struct llm_build_phi2 : public llm_graph_context {
 
         auto * inp_attn = build_attn_inp_kv_unified();
 
+        ggml_tensor * inp_out_ids = build_inp_out_ids();
+
         for (int il = 0; il < n_layer; ++il) {
             attn_norm_output = build_norm(inpL,
                     model.layers[il].attn_norm,
@@ -7310,9 +7802,7 @@ struct llm_build_phi2 : public llm_graph_context {
                         Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f, il);
             }
 
-            if (il == n_layer - 1) {
-                // skip computing output for unused tokens
-                ggml_tensor * inp_out_ids = build_inp_out_ids();
+            if (il == n_layer - 1 && 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);
@@ -7384,6 +7874,8 @@ struct llm_build_phi3 : public llm_graph_context {
             inp_attn = build_attn_inp_kv_unified();
         }
 
+        ggml_tensor * inp_out_ids = build_inp_out_ids();
+
         for (int il = 0; il < n_layer; ++il) {
             auto * residual = inpL;
 
@@ -7447,9 +7939,7 @@ struct llm_build_phi3 : public llm_graph_context {
                         Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f, il);
             }
 
-            if (il == n_layer - 1) {
-                // skip computing output for unused tokens
-                ggml_tensor* inp_out_ids = build_inp_out_ids();
+            if (il == n_layer - 1 && inp_out_ids) {
                 cur      = ggml_get_rows(ctx0, cur,      inp_out_ids);
                 residual = ggml_get_rows(ctx0, residual, inp_out_ids);
             }
@@ -7535,15 +8025,16 @@ struct llm_build_plamo : public llm_graph_context {
 
         auto * inp_attn = build_attn_inp_kv_unified();
 
-        for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inp_out_ids = build_inp_out_ids();
 
+        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;
+            ggml_tensor * sa_inp = cur;
 
             // self-attention
             {
@@ -7581,18 +8072,17 @@ struct llm_build_plamo : public llm_graph_context {
                         model.layers[il].wo, NULL,
                         Qcur, Kcur, Vcur, nullptr, 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();
+            if (il == n_layer - 1 && inp_out_ids) {
                 cur    = ggml_get_rows(ctx0,    cur, inp_out_ids);
-                sa_out = ggml_get_rows(ctx0, sa_out, inp_out_ids);
+                sa_inp = ggml_get_rows(ctx0, sa_inp, inp_out_ids);
                 inpL   = ggml_get_rows(ctx0,   inpL, inp_out_ids);
             }
 
+            ggml_tensor * sa_out = cur;
+
+            cur = sa_inp;
+
             // feed-forward network
             {
                 cur = build_ffn(cur,
@@ -7657,6 +8147,8 @@ struct llm_build_gpt2 : public llm_graph_context {
         inpL = ggml_add(ctx0, inpL, pos);
         cb(inpL, "inpL", -1);
 
+        ggml_tensor * inp_out_ids = build_inp_out_ids();
+
         for (int il = 0; il < n_layer; ++il) {
             cur = build_norm(inpL,
                     model.layers[il].attn_norm,
@@ -7689,9 +8181,7 @@ struct llm_build_gpt2 : public llm_graph_context {
                         Qcur, Kcur, Vcur, nullptr, 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();
+            if (il == n_layer - 1 && inp_out_ids) {
                 cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
                 inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
             }
@@ -7761,6 +8251,8 @@ struct llm_build_codeshell : public llm_graph_context {
 
         auto * inp_attn = build_attn_inp_kv_unified();
 
+        ggml_tensor * inp_out_ids = build_inp_out_ids();
+
         for (int il = 0; il < n_layer; ++il) {
             cur = build_norm(inpL,
                     model.layers[il].attn_norm,
@@ -7805,9 +8297,7 @@ struct llm_build_codeshell : public llm_graph_context {
                         Qcur, Kcur, Vcur, nullptr, 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();
+            if (il == n_layer - 1 && inp_out_ids) {
                 cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
                 inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
             }
@@ -7861,128 +8351,128 @@ struct llm_build_codeshell : public llm_graph_context {
 
 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;
+        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_ASSERT(n_embd_head == hparams.n_embd_head_k);
+        GGML_ASSERT(n_embd_head == hparams.n_rot);
 
-    ggml_tensor * cur;
-    ggml_tensor * inpL;
+        ggml_tensor * cur;
+        ggml_tensor * inpL;
 
-    inpL = build_inp_embd(model.tok_embd);
+        inpL = build_inp_embd(model.tok_embd);
 
-    // inp_pos - contains the positions
-    ggml_tensor * inp_pos = build_inp_pos();
+        // inp_pos - contains the positions
+        ggml_tensor * inp_pos = build_inp_pos();
 
-    auto * inp_attn = build_attn_inp_kv_unified();
+        auto * inp_attn = build_attn_inp_kv_unified();
 
-    for (int il = 0; il < n_layer; ++il) {
-        ggml_tensor * inpSA = inpL;
+        ggml_tensor * inp_out_ids = build_inp_out_ids();
 
-        // norm
-        cur = build_norm(inpL,
-                model.layers[il].attn_norm, model.layers[il].attn_norm_b,
-                LLM_NORM, il);
-        cb(cur, "attn_norm", il);
+        for (int il = 0; il < n_layer; ++il) {
+            ggml_tensor * inpSA = 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);
-            // if (model.layers[il].bq) {
-            //     Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
-            //     cb(Qcur, "Qcur", 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 * 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);
-            // }
+            // 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 * 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);
-            // }
+                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);
+                // }
 
-            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 * 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_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
-            );
+                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);
 
-            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
-            );
+                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);
-            cb(Kcur, "Kcur", il);
-            cb(Vcur, "Vcur", 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
+                        );
 
-            cur = build_attn(inp_attn, gf,
-                    model.layers[il].wo, NULL,
-                    Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+                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, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+            }
+
+            if (il == n_layer - 1 && 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;
         }
 
-        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);
-        }
+        cur = inpL;
 
-        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
-        cb(ffn_inp, "ffn_inp", il);
+        cur = build_norm(cur,
+                model.output_norm, model.output_norm_b,
+                LLM_NORM, -1);
 
-        // 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);
+        cb(cur, "result_norm", -1);
+        res->t_embd = cur;
 
-        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);
+        // lm_head
+        cur = build_lora_mm(model.output, cur);
 
-        cur = ggml_add(ctx0, cur, ffn_inp);
+        cb(cur, "result_output", -1);
+        res->t_logits = cur;
 
-        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);
+        ggml_build_forward_expand(gf, cur);
     }
 };
 
@@ -8003,6 +8493,8 @@ struct llm_build_internlm2 : public llm_graph_context {
 
         auto * inp_attn = build_attn_inp_kv_unified();
 
+        ggml_tensor * inp_out_ids = build_inp_out_ids();
+
         for (int il = 0; il < n_layer; ++il) {
             ggml_tensor * inpSA = inpL;
 
@@ -8061,9 +8553,7 @@ struct llm_build_internlm2 : public llm_graph_context {
                         Qcur, Kcur, Vcur, nullptr, 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();
+            if (il == n_layer - 1 && inp_out_ids) {
                 cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
                 inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
             }
@@ -8139,6 +8629,8 @@ struct llm_build_minicpm3 : public llm_graph_context {
 
         auto * inp_attn = build_attn_inp_kv_unified();
 
+        ggml_tensor * inp_out_ids = build_inp_out_ids();
+
         for (int il = 0; il < n_layer; ++il) {
             ggml_tensor * inpSA = inpL;
 
@@ -8258,15 +8750,13 @@ struct llm_build_minicpm3 : public llm_graph_context {
                         q_states, k_states, v_states, nullptr, nullptr, kq_scale, il);
             }
 
-            if (il == n_layer - 1) {
-                // skip computing output for unused tokens
-                ggml_tensor * inp_out_ids = build_inp_out_ids();
+            if (il == n_layer - 1 && 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));
+            const float scale_res = scale_depth/sqrtf(float(n_layer)); // TODO: is this correct?
             cur = ggml_scale(ctx0, cur, scale_res);
             cb(cur, "hidden_scaled", il);
 
@@ -8343,6 +8833,8 @@ struct llm_build_gemma : public llm_graph_context {
 
         auto * inp_attn = build_attn_inp_kv_unified();
 
+        ggml_tensor * inp_out_ids = build_inp_out_ids();
+
         for (int il = 0; il < n_layer; ++il) {
             // norm
             cur = build_norm(inpL,
@@ -8388,9 +8880,7 @@ struct llm_build_gemma : public llm_graph_context {
                         Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f, il);
             }
 
-            if (il == n_layer - 1) {
-                // skip computing output for unused tokens
-                ggml_tensor * inp_out_ids = build_inp_out_ids();
+            if (il == n_layer - 1 && inp_out_ids) {
                 cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
                 inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
             }
@@ -8459,6 +8949,8 @@ struct llm_build_gemma2_iswa : public llm_graph_context {
 
         auto * inp_attn = build_attn_inp_kv_unified_iswa();
 
+        ggml_tensor * inp_out_ids = build_inp_out_ids();
+
         for (int il = 0; il < n_layer; ++il) {
             // norm
             cur = build_norm(inpL,
@@ -8496,32 +8988,23 @@ struct llm_build_gemma2_iswa : public llm_graph_context {
                 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);
+                Qcur = ggml_scale(ctx0, Qcur, hparams.f_attention_scale);
 
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, NULL,
                         Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f, il);
             }
 
+            if (il == n_layer - 1 && inp_out_ids) {
+                cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
+                inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
+            }
+
             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);
 
@@ -8600,6 +9083,8 @@ struct llm_build_gemma3_iswa : public llm_graph_context {
         // TODO: is causal == true correct? might need some changes
         auto * inp_attn = build_attn_inp_kv_unified_iswa();
 
+        ggml_tensor * inp_out_ids = build_inp_out_ids();
+
         for (int il = 0; il < n_layer; ++il) {
             const float freq_base_l  = model.get_rope_freq_base (cparams, il);
             const float freq_scale_l = model.get_rope_freq_scale(cparams, il);
@@ -8644,9 +9129,17 @@ struct llm_build_gemma3_iswa : public llm_graph_context {
                 cb(Kcur, "Kcur", il);
                 cb(Vcur, "Vcur", il);
 
+                // ref: https://github.com/google/gemma_pytorch/blob/014acb7ac4563a5f77c76d7ff98f31b568c16508/gemma/model.py#L315
+                Qcur = ggml_scale(ctx0, Qcur, hparams.f_attention_scale);
+
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, NULL,
-                        Qcur, Kcur, Vcur, nullptr, nullptr, hparams.f_attention_scale, il);
+                        Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f, il);
+            }
+
+            if (il == n_layer - 1 && inp_out_ids) {
+                cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
+                inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
             }
 
             cur = build_norm(cur,
@@ -8654,13 +9147,6 @@ struct llm_build_gemma3_iswa : public llm_graph_context {
                     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);
 
@@ -8713,6 +9199,442 @@ struct llm_build_gemma3_iswa : public llm_graph_context {
     }
 };
 
+struct llm_build_gemma3n_iswa : public llm_graph_context {
+    const llama_model & model;
+    ggml_cgraph * gf;
+
+    const int64_t n_embd_head;
+    const int64_t n_embd_altup;
+    const int64_t n_altup;
+    const int     i_altup_act;
+    const int     n_layer_kv = 20; // number of layers having KV [KV_REUSE]
+    const int     n_layer_sparsity = 10; // number of layers using activation sparsity
+    const float   f_sparsity_std_mul = 1.6448533535003662f; // std_multiplier = normal_dist.icdf(0.95)
+
+    ggml_tensor * one; // containing single element 1.0f
+
+    llm_build_gemma3n_iswa(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf)
+            : llm_graph_context(params),
+              model(model),
+              gf(gf),
+              n_embd_head(model.hparams.n_embd_head_k),
+              n_embd_altup(model.hparams.n_embd_altup),
+              n_altup(model.hparams.n_altup),
+              i_altup_act(model.hparams.i_altup_act) {
+        ggml_tensor * cur;
+        ggml_tensor * inpL;
+
+        // TODO: remove this when ggml_scale_add is implemented
+        one = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, 1);
+        {
+            auto inp = std::make_unique<llm_graph_input_one>();
+            inp->one = one;
+            res->add_input(std::move(inp));
+        }
+
+        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_iswa();
+
+        // inp_per_layer shape: [n_embd_altup, n_tokens, n_layer]
+        ggml_tensor * inp_per_layer = project_per_layer_inputs(inpL, get_per_layer_inputs());
+
+        // inpL now has only 1 altup, project it to the rest of the altups
+        // these "added" altups will be concat to the last dim of inpL
+        {
+            ggml_tensor * target_magnitude = calc_magnitude(inpL);
+            ggml_tensor * inp_repeated = ggml_repeat_4d(ctx0, inpL, n_embd, n_tokens, n_altup - 1, 1);
+            ggml_tensor * altup_added = ggml_mul_mat(ctx0, model.altup_proj, inp_repeated); // shape: [n_embd, n_tokens, n_altup - 1]
+            ggml_tensor * new_magnitude = calc_magnitude(altup_added);
+            altup_added = ggml_div(ctx0,
+                                ggml_mul(ctx0, altup_added, target_magnitude),
+                                new_magnitude);
+            inpL = ggml_concat(ctx0, inpL, altup_added, 2); // shape: [n_embd, n_tokens, n_altup]
+            cb(inpL, "inp_stacked", -1);
+        }
+
+        // inpL now has shape:          [n_embd,       n_tokens, n_altup]
+        // inp_per_layer now has shape: [n_embd_altup, n_tokens, n_layer]
+
+        for (int il = 0; il < n_layer; ++il) {
+            // this block is made to be closely resemble Gemma3p5DecoderLayer on python code
+            const bool has_kv = (il < n_layer_kv);
+
+            const float freq_base_l  = model.get_rope_freq_base (cparams, il);
+            const float freq_scale_l = model.get_rope_freq_scale(cparams, il);
+
+            ggml_tensor * cur = inpL; // [n_embd, n_tokens, n_altup]
+            ggml_tensor * predictions = altup_predict(cur, il); // [n_embd, n_tokens, n_altup]
+
+            // predicted value will go through self-attention and laurel
+            ggml_tensor * active_prediction = view_2d_slice(predictions, i_altup_act); // [n_embd, n_tokens]
+            cur = active_prediction;
+            cb(cur, "active_prediction", il);
+
+            // norm
+            cur = build_norm(cur, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
+            cb(cur, "attn_norm", il);
+
+            // laurel
+            ggml_tensor * laurel_out = laurel(cur, il); // [n_embd, n_tokens]
+
+            // self-attention
+            if (has_kv) {
+                // 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);
+                Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
+                Vcur = ggml_rms_norm(ctx0, Vcur, hparams.f_norm_rms_eps);
+
+                cb(Qcur, "Qcur_normed", il);
+                cb(Kcur, "Kcur_normed", il);
+                cb(Vcur, "Vcur_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 = 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_pos", il);
+                cb(Kcur, "Kcur_pos", il);
+
+                cur = build_attn(inp_attn, gf,
+                        model.layers[il].wo, NULL,
+                        Qcur, Kcur, Vcur, nullptr, nullptr, hparams.f_attention_scale, il);
+            } else {
+                // no KV layers
+                ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+                cb(Qcur, "Qcur", il);
+                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, 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);
+                cb(Qcur, "Qcur_pos", il);
+
+                cur = build_attn(inp_attn, gf,
+                    model.layers[il].wo, NULL,
+                    Qcur, nullptr, nullptr, nullptr, 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);
+
+            cur = ggml_add(ctx0, cur, active_prediction); // [n_embd, n_tokens]
+            cb(cur, "attn_gated", il);
+
+            ggml_tensor * attn_laurel = ggml_scale(ctx0,
+                                            ggml_add(ctx0, cur, laurel_out),
+                                            1.0f / sqrtf(2.0f)); // [n_embd, n_tokens]
+            cb(attn_laurel, "attn_laurel", il);
+
+            cur = build_norm(attn_laurel,
+                    model.layers[il].ffn_norm, NULL,
+                    LLM_NORM_RMS, il);
+            cb(cur, "ffn_norm", il);
+
+            // feed-forward network
+            {
+                ggml_tensor * up_proj   = build_lora_mm(model.layers[il].ffn_up,   cur);
+                ggml_tensor * gate_proj = build_lora_mm(model.layers[il].ffn_gate, cur);
+
+                if (il < n_layer_sparsity) {
+                    // apply activation sparsity
+                    gate_proj = gaussian_topk(gate_proj);
+                }
+                gate_proj = ggml_gelu(ctx0, gate_proj);
+
+                cur = ggml_mul(ctx0, up_proj, gate_proj);
+                cur = build_lora_mm(model.layers[il].ffn_down, cur);
+                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", il);
+
+            ggml_tensor * attn_ffw_laurel_gated = ggml_add(ctx0, cur, attn_laurel); // [n_embd, n_tokens]
+            cb(attn_ffw_laurel_gated, "attn_ffw_laurel_gated", il);
+
+            ggml_tensor * corrected = altup_correct(predictions, attn_ffw_laurel_gated, il); // [n_embd, n_tokens, n_altup]
+
+            ggml_tensor * first_prediction; // [n_embd, n_tokens]
+            {
+                first_prediction = view_2d_slice(corrected, i_altup_act); // [n_embd, n_tokens]
+                first_prediction = ggml_mul(ctx0, first_prediction, model.layers[il].altup_correct_scale);
+                first_prediction = build_lora_mm(model.layers[il].per_layer_inp_gate, first_prediction);
+                first_prediction = ggml_gelu(ctx0, first_prediction); // [n_embd_altup, n_tokens]
+                cb(first_prediction, "first_prediction_gated", il);
+                ggml_tensor * inp_this_layer = view_2d_slice(inp_per_layer, il); // [n_embd_altup, n_tokens]
+                first_prediction = ggml_mul(ctx0, first_prediction, inp_this_layer); // [n_embd_altup, n_tokens]
+                cb(first_prediction, "first_prediction_scaled", il);
+
+                first_prediction = build_lora_mm(model.layers[il].per_layer_proj, first_prediction); // [n_embd, n_tokens]
+                first_prediction = build_norm(first_prediction,
+                        model.layers[il].per_layer_post_norm, NULL,
+                        LLM_NORM_RMS, il);
+                cb(first_prediction, "first_prediction_out", il);
+            }
+
+            // equivalent to python code: corrected_predictions[1:] += first_prediction
+            {
+                ggml_tensor * slice_first = view_2d_slice(corrected, 0);
+                ggml_tensor * slice_rest  = ggml_view_3d(ctx0, corrected, n_embd, n_tokens, n_altup - 1,
+                                                    ggml_row_size(corrected->type, n_embd),
+                                                    ggml_row_size(corrected->type, n_embd*n_tokens),
+                                                    n_embd*n_tokens*ggml_element_size(corrected));
+                ggml_tensor * tmp = ggml_add(ctx0, slice_rest, first_prediction); // [n_embd, n_tokens, n_altup - 1]
+                corrected = ggml_concat(ctx0, slice_first, tmp, 2); // [n_embd, n_tokens, n_altup]
+            }
+
+            cur = corrected; // [n_embd, n_tokens, n_altup]
+            cur = build_cvec(cur, il);
+            cb(cur, "l_out", il);
+
+            // input for next layer
+            inpL = cur;
+        }
+
+        cur = inpL; // [n_embd, n_tokens, n_altup]
+
+        // cur now has multiple altup(s), we want to merge them back to 1 altup
+        {
+            ggml_tensor * target_magnitude = calc_magnitude(view_2d_slice(cur, i_altup_act)); // [n_embd, n_tokens]
+            // do a view to skip the first slice (active altup)
+            ggml_tensor * alt_slice = ggml_view_3d(ctx0, cur, n_embd, n_tokens, n_altup - 1,
+                                                    ggml_row_size(cur->type, n_embd),
+                                                    ggml_row_size(cur->type, n_embd*n_tokens),
+                                                    n_embd*n_tokens*ggml_element_size(cur));
+            ggml_tensor * altup_unembd = ggml_mul_mat(ctx0, model.altup_unembd_proj, alt_slice); // shape: [n_embd, n_tokens, n_altup - 1]
+            ggml_tensor * new_magnitude = calc_magnitude(altup_unembd);
+            altup_unembd = ggml_div(ctx0,
+                                ggml_mul(ctx0, altup_unembd, target_magnitude),
+                                new_magnitude);
+            cb(altup_unembd, "altup_unembd", -1);
+
+            // equivalent to torch.mean(hidden_states, dim=0)
+            cur = view_2d_slice(cur, 0); // [n_embd, n_tokens]
+            for (int i = 0; i < n_altup - 1; ++i) {
+                cur = ggml_add(ctx0, cur, view_2d_slice(altup_unembd, i));
+            }
+            cur = ggml_scale(ctx0, cur, 1.0f / float(n_altup)); // [n_embd, n_tokens]
+            cb(cur, "unembd_merged", -1);
+        }
+
+        // cur now has shape: [n_embd, n_tokens]
+
+        // TODO: move this to right after the last KV layer
+        {
+            // skip computing output for unused tokens
+            ggml_tensor * inp_out_ids = build_inp_out_ids();
+            cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
+        }
+
+        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);
+
+        {
+            // 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);
+    }
+
+    ggml_tensor * calc_magnitude(ggml_tensor * x) {
+        return ggml_sqrt(ctx0, ggml_sum_rows(ctx0, ggml_sqr(ctx0, x)));
+    }
+
+    // get 2D slice view from a 3D tensor, the idx corresponds to the 3rd dim
+    ggml_tensor * view_2d_slice(ggml_tensor * x, int idx) {
+        GGML_ASSERT(idx < (int)x->ne[2]);
+        return ggml_view_2d(ctx0, x, x->ne[0], x->ne[1],
+                            ggml_row_size(x->type, x->ne[0]),
+                            idx * x->ne[0] * x->ne[1] * ggml_element_size(x));
+    }
+
+    // equivalent to get_per_layer_inputs() in python code
+    // output shape: [n_embd_altup, n_layer, n_tokens]
+    ggml_tensor * get_per_layer_inputs() {
+        auto inp = std::make_unique<llm_graph_input_embd>();
+        ggml_tensor * inp_per_layer;
+        if (ubatch.token) {
+            inp->tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, ubatch.n_tokens);
+            ggml_set_input(inp->tokens);
+            res->t_tokens = inp->tokens;
+            inp_per_layer = ggml_get_rows(ctx0, model.tok_embd_per_layer, inp->tokens);
+            inp_per_layer = ggml_reshape_3d(ctx0, inp_per_layer, n_embd_altup, n_layer, n_tokens);
+            inp_per_layer = ggml_scale(ctx0, inp_per_layer, sqrtf((float)n_embd_altup));
+            cb(inp_per_layer, "inp_per_layer_selected", -1);
+        } else {
+            GGML_ABORT("TODO: support embd input");
+        }
+        res->add_input(std::move(inp));
+        return inp_per_layer;
+    }
+
+    // equivalent to project_per_layer_inputs() in python code
+    // this calculates the per-layer inputs, so the final tensor shape will have n_layer as the last dim
+    // output shape: [n_embd_altup, n_tokens, n_layer]
+    ggml_tensor * project_per_layer_inputs(ggml_tensor * inputs_embeds, ggml_tensor * inp_per_layer) {
+        const float per_layer_projection_scale = 1.0f / sqrtf((float)n_embd);
+        const float per_layer_input_scale      = 1.0f / sqrtf(2.0f);
+
+        ggml_tensor * per_layer_proj = ggml_mul_mat(ctx0, model.per_layer_model_proj, inputs_embeds);
+        per_layer_proj = ggml_scale(ctx0, per_layer_proj, per_layer_projection_scale);
+        per_layer_proj = ggml_reshape_3d(ctx0, per_layer_proj, n_embd_altup, n_layer, n_tokens);
+        per_layer_proj = build_norm(per_layer_proj,
+                                    model.per_layer_proj_norm, NULL,
+                                    LLM_NORM_RMS, -1); // [n_embd_altup, n_layer, n_tokens]
+        cb(per_layer_proj, "per_layer_proj", -1);
+
+        inp_per_layer = ggml_add(ctx0, inp_per_layer, per_layer_proj);
+        inp_per_layer = ggml_scale(ctx0, inp_per_layer, per_layer_input_scale);
+        cb(inp_per_layer, "inp_per_layer", -1);
+
+        // permute to shape: [n_embd_altup, n_tokens, n_layer]
+        inp_per_layer = ggml_cont(ctx0, ggml_permute(ctx0, inp_per_layer, 0, 2, 1, 3));
+        return inp_per_layer;
+    }
+
+    // input cur shape: [n_altup, n_tokens]
+    // output    shape: [n_altup, n_tokens]
+    ggml_tensor * laurel(ggml_tensor * cur, int il) {
+        ggml_tensor * tmp = cur;
+        tmp = build_lora_mm(model.layers[il].laurel_l, tmp);
+        tmp = build_lora_mm(model.layers[il].laurel_r, tmp);
+        tmp = build_norm(tmp, model.layers[il].laurel_post_norm, NULL, LLM_NORM_RMS, il);
+        tmp = ggml_add(ctx0, tmp, cur);
+        cb(tmp, "laurel_out", il);
+        return tmp;
+    }
+
+    // input x shape: [n_embd, n_tokens]
+    // output  shape: [n_embd, n_tokens]
+    ggml_tensor * gaussian_topk(ggml_tensor * x) {
+        ggml_tensor * mean = ggml_mean(ctx0, x);
+        ggml_tensor * std  = ggml_sqrt(ctx0, ggml_scale(ctx0,
+            ggml_sum_rows(ctx0, ggml_sqr(ctx0, ggml_sub(ctx0, x, mean))),
+            1.0f / (float)(x->ne[0] - 1)
+        ));
+        ggml_tensor * cutoff_x = ggml_add(ctx0, mean, ggml_scale(ctx0, std, f_sparsity_std_mul));
+        return ggml_relu(ctx0, ggml_sub(ctx0, x, cutoff_x));
+    }
+
+    //
+    // altup functions
+    //
+
+    // equivalent to compute_router_modalities() in python code
+    // input x shape: [n_embd,  n_tokens]
+    // output  shape: [n_altup, n_tokens]
+    ggml_tensor * altup_compute_router_modalities(ggml_tensor * x, int il) {
+        ggml_tensor * router_inputs = build_norm(x,
+            model.layers[il].altup_router_norm, NULL,
+            LLM_NORM_RMS, il);
+
+        // router_input_scale
+        router_inputs = ggml_scale(ctx0, router_inputs, 1.0f / (float)n_embd);
+
+        ggml_tensor * output = ggml_mul_mat(ctx0, model.layers[il].altup_router, router_inputs);
+        return ggml_tanh(ctx0, output); // [n_altup, n_tokens]
+    }
+
+    // input cur shape: [n_embd, n_tokens, n_altup]
+    // output    shape: [n_embd, n_tokens, n_altup]
+    ggml_tensor * altup_predict(ggml_tensor * cur, int il) {
+        ggml_tensor * activated = view_2d_slice(cur, i_altup_act); // [n_embd, n_tokens]
+        ggml_tensor * modalities = altup_compute_router_modalities(activated, il); // [n_altup, n_tokens]
+        cb(modalities, "modalities", il);
+
+        ggml_tensor * all_coefs = build_lora_mm(model.layers[il].altup_predict_coef, modalities);
+        cb(all_coefs, "all_coefs", il);
+        // first dim now having n_altup^2 elements, we reshape it to 2D (so we end up with 3D tensor)
+        all_coefs = ggml_reshape_3d(ctx0, all_coefs, n_altup, n_altup, n_tokens);
+
+        // permute to [n_altup, n_embd, n_tokens]
+        ggml_tensor * cur_permuted = ggml_cont(ctx0, ggml_permute(ctx0, cur, 1, 2, 0, 3));
+        ggml_tensor * predictions = ggml_mul_mat(ctx0, cur_permuted, all_coefs); // [n_altup, n_embd, n_tokens]
+
+        // final shape must be the same as cur: [n_embd, n_tokens, n_altup]
+        predictions = ggml_cont(ctx0, ggml_permute(ctx0, predictions, 0, 2, 1, 3));
+        predictions = ggml_add(ctx0, predictions, cur);
+        cb(predictions, "predictions", il);
+
+        return predictions;
+    }
+
+    // input predictions       shape: [n_embd, n_tokens, n_altup]
+    // input activated         shape: [n_embd, n_tokens]
+    // output                  shape: [n_embd, n_tokens, n_altup]
+    ggml_tensor * altup_correct(ggml_tensor * predictions, ggml_tensor * activated, int il) {
+        ggml_tensor * modalities = altup_compute_router_modalities(activated, il); // [n_altup, n_tokens]
+        cb(modalities, "modalities", il);
+
+        ggml_tensor * active_prediction = view_2d_slice(predictions, i_altup_act);
+        ggml_tensor * innovation = ggml_sub(ctx0, activated, active_prediction); // [n_embd, n_tokens]
+        cb(innovation, "innovation", il);
+
+        ggml_tensor * all_coefs = build_lora_mm(model.layers[il].altup_correct_coef, modalities); // [n_altup, n_tokens]
+        all_coefs = ggml_add(ctx0, all_coefs, one);
+        cb(all_coefs, "all_coefs", il);
+        all_coefs = ggml_cont(ctx0, ggml_transpose(ctx0, all_coefs)); // [n_tokens, n_altup]
+        all_coefs = ggml_reshape_3d(ctx0, all_coefs, 1, n_tokens, n_altup); // [1, n_tokens, n_altup]
+
+        innovation = ggml_repeat_4d(ctx0, innovation, n_embd, n_tokens, n_altup, 1);
+        ggml_tensor * corrected = ggml_mul(ctx0, innovation, all_coefs); // [n_embd, n_tokens, n_altup]
+        corrected = ggml_add(ctx0, corrected, predictions); // [n_embd, n_tokens, n_altup]
+        cb(corrected, "corrected", il);
+
+        return corrected;
+    }
+};
+
 // 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) {
@@ -8731,6 +9653,8 @@ struct llm_build_starcoder2 : public llm_graph_context {
 
         auto * inp_attn = build_attn_inp_kv_unified();
 
+        ggml_tensor * inp_out_ids = build_inp_out_ids();
+
         for (int il = 0; il < n_layer; ++il) {
             ggml_tensor * inpSA = inpL;
 
@@ -8789,9 +9713,7 @@ struct llm_build_starcoder2 : public llm_graph_context {
                         Qcur, Kcur, Vcur, nullptr, 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();
+            if (il == n_layer - 1 && inp_out_ids) {
                 cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
                 inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
             }
@@ -8843,17 +9765,16 @@ struct llm_build_starcoder2 : public llm_graph_context {
 };
 
 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) {
+    llm_build_mamba(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) {
         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();
+        auto * rs_inp = build_rs_inp();
+
+        ggml_tensor * inp_out_ids = build_inp_out_ids();
 
         for (int il = 0; il < n_layer; ++il) {
             // norm
@@ -8862,12 +9783,13 @@ struct llm_build_mamba : public llm_graph_context {
                     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 (model.arch == LLM_ARCH_MAMBA2) {
+                cur = build_mamba2_layer(rs_inp, gf, cur, model, ubatch, il);
+            } else {
+                cur = build_mamba_layer(rs_inp, gf, cur, model, ubatch, il);
+            }
 
-            if (il == n_layer - 1) {
-                // skip computing output for unused tokens
-                ggml_tensor * inp_out_ids = build_inp_out_ids();
+            if (il == n_layer - 1 && inp_out_ids) {
                 cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
                 inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
             }
@@ -8899,22 +9821,23 @@ struct llm_build_mamba : public llm_graph_context {
         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 auto * kv_state = static_cast<const llama_kv_cache_recurrent_state *>(mstate);
+        llm_graph_input_rs * inp,
+               ggml_cgraph * gf,
+               ggml_tensor * cur,
+         const llama_model & model,
+        const llama_ubatch & ubatch,
+                       int   il) const {
+        const auto * mctx_cur = static_cast<const llama_memory_recurrent_context *>(mctx);
 
-        const auto kv_head = kv_state->get_head();
+        const auto kv_head = mctx_cur->get_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_head  = d_inner;
+        const int64_t head_dim = 1;
         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;
@@ -8927,18 +9850,11 @@ struct llm_build_mamba : public llm_graph_context {
         GGML_ASSERT(ubatch.equal_seqs);
         GGML_ASSERT(ubatch.n_tokens == n_seq_tokens * n_seqs);
 
-        ggml_tensor * conv_states_all = kv_state->get_k_l(il);
-        ggml_tensor * ssm_states_all  = kv_state->get_v_l(il);
+        ggml_tensor * conv_states_all = mctx_cur->get_r_l(il);
+        ggml_tensor * ssm_states_all  = mctx_cur->get_s_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);
+        ggml_tensor * conv = build_rs(inp, gf, conv_states_all, hparams.n_embd_r(), 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);
@@ -8987,8 +9903,8 @@ struct llm_build_mamba : public llm_graph_context {
             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));
+            ggml_tensor * B  = ggml_view_4d(ctx0, x_db, d_state, /* n_group */ 1, n_seq_tokens, n_seqs, d_state*x_db->nb[0], x_db->nb[1], x_db->nb[2], ggml_element_size(x_db)*dt_rank);
+            ggml_tensor * C  = ggml_view_4d(ctx0, x_db, d_state, /* n_group */ 1, n_seq_tokens, n_seqs, d_state*x_db->nb[0], 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) {
@@ -9001,23 +9917,36 @@ struct llm_build_mamba : public llm_graph_context {
             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);
+            cur = x;
+            x = ggml_reshape_4d(ctx0, x, head_dim, n_head, n_seq_tokens, n_seqs);
+
+            ggml_tensor * A = model.layers[il].ssm_a;
+
+            // use the states and the indices provided by build_recurrent_state
+            // (this is necessary in order to properly use the states before they are overwritten,
+            //  while avoiding to make unnecessary copies of the states)
+            auto get_ssm_rows = [&](ggml_context * ctx, ggml_tensor * states, ggml_tensor * ids) {
+                ggml_tensor * ssm = ggml_reshape_4d(ctx, states, d_state, head_dim, n_head, mctx_cur->get_size());
+
+                // 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}
+                return ggml_ssm_scan(ctx, ssm, x, dt, A, B, C, ids);
+            };
+
+            ggml_tensor * y_ssm = build_rs(inp, gf, ssm_states_all, hparams.n_embd_s(), ubatch.n_seqs, get_ssm_rows);
 
             // 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, y_ssm, d_state*d_inner*n_seqs, x->nb[3]*x->ne[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);
+            ggml_tensor * y = ggml_view_3d(ctx0, y_ssm, d_inner, n_seq_tokens, n_seqs, x->nb[2], x->nb[3], 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_add(ctx0, y, ggml_mul(ctx0, cur, 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}
@@ -9026,7 +9955,136 @@ struct llm_build_mamba : public llm_graph_context {
 
         // {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);
+        // cb(cur, "mamba_out", il);
+
+        return cur;
+    }
+
+    ggml_tensor * build_mamba2_layer(
+        llm_graph_input_rs * inp,
+             ggml_cgraph * gf,
+             ggml_tensor * cur,
+       const llama_model & model,
+      const llama_ubatch & ubatch,
+                     int   il) const {
+        const auto * mctx_cur = static_cast<const llama_memory_recurrent_context *>(mctx);
+
+        const auto kv_head = mctx_cur->get_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 n_head  = hparams.ssm_dt_rank;
+        const int64_t head_dim = d_inner / n_head;
+        const int64_t n_group = hparams.ssm_n_group;
+        const int64_t n_seqs  = ubatch.n_seqs;
+
+        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 = mctx_cur->get_r_l(il);
+        ggml_tensor * ssm_states_all  = mctx_cur->get_s_l(il);
+
+        ggml_tensor * conv = build_rs(inp, gf, conv_states_all, hparams.n_embd_r(), n_seqs);
+        conv = ggml_reshape_3d(ctx0, conv, d_conv - 1, d_inner + 2*n_group*d_state, 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);
+
+        // d_in_proj = 2 * self.d_inner + 2 * self.ngroups * self.d_state + self.nheads
+
+        // {n_embd, d_in_proj} @ {n_embd, n_seq_tokens, n_seqs} => {d_in_proj, n_seq_tokens, n_seqs}
+        ggml_tensor * zxBCdt = build_lora_mm(model.layers[il].ssm_in, cur);
+
+        // split the above in three
+        ggml_tensor * z = ggml_view_4d(ctx0, zxBCdt, head_dim, n_head, n_seq_tokens, n_seqs, head_dim*zxBCdt->nb[0], zxBCdt->nb[1], zxBCdt->nb[2], 0);
+        ggml_tensor * xBC = ggml_view_3d(ctx0, zxBCdt, d_inner + 2*n_group*d_state, n_seq_tokens, n_seqs, zxBCdt->nb[1], zxBCdt->nb[2], d_inner*ggml_element_size(zxBCdt));
+        ggml_tensor * dt = ggml_view_3d(ctx0, zxBCdt, n_head, n_seq_tokens, n_seqs, zxBCdt->nb[1], zxBCdt->nb[2], (2*d_inner + 2*n_group*d_state)*ggml_element_size(zxBCdt));
+
+        // conv
+        {
+            // => {d_conv - 1 + n_seq_tokens, d_inner + 2*n_group*d_state, n_seqs}
+            ggml_tensor * conv_x = ggml_concat(ctx0, conv, ggml_transpose(ctx0, xBC), 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 + 2*n_group*d_state, 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 + 2*n_group*d_state)*(n_seqs),
+                        kv_head*(d_conv - 1)*(d_inner + 2*n_group*d_state)*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.
+            xBC = ggml_ssm_conv(ctx0, conv_x, model.layers[il].ssm_conv1d);
+
+            // bias
+            xBC = ggml_add(ctx0, xBC, model.layers[il].ssm_conv1d_b);
+
+            xBC = ggml_silu(ctx0, xBC);
+        }
+
+        // ssm
+        {
+            // These correspond to V K Q in SSM/attention duality
+            ggml_tensor * x = ggml_view_4d(ctx0, xBC, head_dim, n_head, n_seq_tokens, n_seqs, head_dim*xBC->nb[0], xBC->nb[1], xBC->nb[2], 0);
+            ggml_tensor * B = ggml_view_4d(ctx0, xBC, d_state, n_group, n_seq_tokens, n_seqs, d_state*xBC->nb[0], xBC->nb[1], xBC->nb[2], d_inner*ggml_element_size(xBC));
+            ggml_tensor * C = ggml_view_4d(ctx0, xBC, d_state, n_group, n_seq_tokens, n_seqs, d_state*xBC->nb[0], xBC->nb[1], xBC->nb[2], (d_inner + n_group*d_state)*ggml_element_size(xBC));
+
+            // {n_head, n_seq_tokens, n_seqs}
+            dt = ggml_add(ctx0, ggml_cont(ctx0, dt), model.layers[il].ssm_dt_b);
+
+            ggml_tensor * A = model.layers[il].ssm_a;
+
+            // use the states and the indices provided by build_recurrent_state
+            // (this is necessary in order to properly use the states before they are overwritten,
+            //  while avoiding to make unnecessary copies of the states)
+            auto get_ssm_rows = [&](ggml_context * ctx, ggml_tensor * states, ggml_tensor * ids) {
+                ggml_tensor * ssm = ggml_reshape_4d(ctx, states, d_state, head_dim, n_head, mctx_cur->get_size());
+
+                // TODO: use semistructured matrices to implement state-space duality
+                // => {d_inner, n_seq_tokens, n_seqs} and {d_state, d_inner, n_seqs}
+                return ggml_ssm_scan(ctx, ssm, x, dt, A, B, C, ids);
+            };
+
+            ggml_tensor * y_ssm = build_rs(inp, gf, ssm_states_all, hparams.n_embd_s(), ubatch.n_seqs, get_ssm_rows);
+
+            // store last states
+            ggml_build_forward_expand(gf,
+                ggml_cpy(ctx0,
+                    ggml_view_1d(ctx0, y_ssm, d_state*d_inner*n_seqs, ggml_nelements(x)*x->nb[0]),
+                    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_4d(ctx0, y_ssm, head_dim, n_head, n_seq_tokens, n_seqs, x->nb[1], n_head*x->nb[1], n_seq_tokens*n_head*x->nb[1], 0);
+
+            // TODO: skip computing output earlier for unused tokens
+
+            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)));
+
+            // grouped RMS norm
+            y = ggml_reshape_4d(ctx0, y, d_inner / n_group, n_group, n_seq_tokens, n_seqs);
+            y = build_norm(y, model.layers[il].ssm_norm, NULL, LLM_NORM_RMS, il);
+            y = ggml_reshape_3d(ctx0, y, d_inner, n_seq_tokens, n_seqs);
+
+            // {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;
     }
@@ -9050,13 +10108,15 @@ struct llm_build_command_r : public llm_graph_context {
 
         auto * inp_attn = build_attn_inp_kv_unified();
 
-        for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inp_out_ids = build_inp_out_ids();
 
+        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
@@ -9124,9 +10184,7 @@ struct llm_build_command_r : public llm_graph_context {
                         Qcur, Kcur, Vcur, nullptr, 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();
+            if (il == n_layer - 1 && 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);
@@ -9197,6 +10255,8 @@ struct llm_build_cohere2_iswa : public llm_graph_context {
 
         auto * inp_attn = build_attn_inp_kv_unified_iswa();
 
+        ggml_tensor * inp_out_ids = build_inp_out_ids();
+
         for (int il = 0; il < n_layer; ++il) {
             const bool is_swa = hparams.is_swa(il);
 
@@ -9259,9 +10319,7 @@ struct llm_build_cohere2_iswa : public llm_graph_context {
                         Qcur, Kcur, Vcur, nullptr, 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();
+            if (il == n_layer - 1 && 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);
@@ -9332,6 +10390,8 @@ struct llm_build_olmo : public llm_graph_context {
 
         auto * inp_attn = build_attn_inp_kv_unified();
 
+        ggml_tensor * inp_out_ids = build_inp_out_ids();
+
         for (int il = 0; il < n_layer; ++il) {
             ggml_tensor * inpSA = inpL;
 
@@ -9390,9 +10450,7 @@ struct llm_build_olmo : public llm_graph_context {
                         Qcur, Kcur, Vcur, nullptr, 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();
+            if (il == n_layer - 1 && inp_out_ids) {
                 cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
                 inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
             }
@@ -9460,6 +10518,8 @@ struct llm_build_olmo2 : public llm_graph_context {
 
         auto * inp_attn = build_attn_inp_kv_unified();
 
+        ggml_tensor * inp_out_ids = build_inp_out_ids();
+
         for (int il = 0; il < n_layer; ++il) {
             ggml_tensor * inpSA = inpL;
 
@@ -9510,18 +10570,16 @@ struct llm_build_olmo2 : public llm_graph_context {
                         Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
             }
 
+            if (il == n_layer - 1 && inp_out_ids) {
+                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+            }
+
             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);
 
@@ -9589,6 +10647,8 @@ struct llm_build_olmoe : public llm_graph_context {
 
         auto * inp_attn = build_attn_inp_kv_unified();
 
+        ggml_tensor * inp_out_ids = build_inp_out_ids();
+
         for (int il = 0; il < n_layer; ++il) {
             ggml_tensor * inpSA = inpL;
 
@@ -9643,9 +10703,7 @@ struct llm_build_olmoe : public llm_graph_context {
                         Qcur, Kcur, Vcur, nullptr, 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();
+            if (il == n_layer - 1 && inp_out_ids) {
                 cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
                 inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
             }
@@ -9715,6 +10773,8 @@ struct llm_build_openelm : public llm_graph_context {
 
         auto * inp_attn = build_attn_inp_kv_unified();
 
+        ggml_tensor * inp_out_ids = build_inp_out_ids();
+
         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);
@@ -9776,11 +10836,9 @@ struct llm_build_openelm : public llm_graph_context {
                         Qcur, Kcur, Vcur, nullptr, 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();
+            if (il == n_layer - 1 && inp_out_ids) {
                 residual = ggml_get_rows(ctx0, residual, inp_out_ids);
-                cur = ggml_get_rows(ctx0, cur, inp_out_ids);
+                cur      = ggml_get_rows(ctx0, cur,      inp_out_ids);
             }
 
             ggml_tensor * ffn_inp = ggml_add(ctx0, residual, cur);
@@ -9846,6 +10904,8 @@ struct llm_build_gptneox : public llm_graph_context {
 
         auto * inp_attn = build_attn_inp_kv_unified();
 
+        ggml_tensor * inp_out_ids = build_inp_out_ids();
+
         for (int il = 0; il < n_layer; ++il) {
             cur = build_norm(inpL,
                     model.layers[il].attn_norm,
@@ -9890,9 +10950,7 @@ struct llm_build_gptneox : public llm_graph_context {
                         Qcur, Kcur, Vcur, nullptr, 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();
+            if (il == n_layer - 1 && inp_out_ids) {
                 cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
                 inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
             }
@@ -9994,6 +11052,8 @@ struct llm_build_arctic : public llm_graph_context {
 
         auto * inp_attn = build_attn_inp_kv_unified();
 
+        ggml_tensor * inp_out_ids = build_inp_out_ids();
+
         for (int il = 0; il < n_layer; ++il) {
             ggml_tensor * inpSA = inpL;
 
@@ -10040,9 +11100,7 @@ struct llm_build_arctic : public llm_graph_context {
                         Qcur, Kcur, Vcur, nullptr, 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();
+            if (il == n_layer - 1 && inp_out_ids) {
                 cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
                 inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
             }
@@ -10134,6 +11192,8 @@ struct llm_build_deepseek : public llm_graph_context {
 
         const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
 
+        ggml_tensor * inp_out_ids = build_inp_out_ids();
+
         for (int il = 0; il < n_layer; ++il) {
             ggml_tensor * inpSA = inpL;
 
@@ -10195,14 +11255,11 @@ struct llm_build_deepseek : public llm_graph_context {
                         Qcur, Kcur, Vcur, nullptr, nullptr, kq_scale, il);
             }
 
-            if (il == n_layer - 1) {
-                // skip computing output for unused tokens
-                ggml_tensor * inp_out_ids = build_inp_out_ids();
+            if (il == n_layer - 1 && 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);
 
@@ -10310,6 +11367,8 @@ struct llm_build_deepseek2 : public llm_graph_context {
 
         auto * inp_attn = build_attn_inp_kv_unified();
 
+        ggml_tensor * inp_out_ids = build_inp_out_ids();
+
         for (int il = 0; il < n_layer; ++il) {
             ggml_tensor * inpSA = inpL;
 
@@ -10459,9 +11518,7 @@ struct llm_build_deepseek2 : public llm_graph_context {
                 }
             }
 
-            if (il == n_layer - 1) {
-                // skip computing output for unused tokens
-                ggml_tensor * inp_out_ids = build_inp_out_ids();
+            if (il == n_layer - 1 && inp_out_ids) {
                 cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
                 inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
             }
@@ -10557,6 +11614,8 @@ struct llm_build_bitnet : public llm_graph_context {
 
         auto * inp_attn = build_attn_inp_kv_unified();
 
+        ggml_tensor * inp_out_ids = build_inp_out_ids();
+
         for (int il = 0; il < n_layer; ++il) {
             ggml_tensor * inpSA = inpL;
 
@@ -10639,9 +11698,7 @@ struct llm_build_bitnet : public llm_graph_context {
                 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();
+            if (il == n_layer - 1 && inp_out_ids) {
                 cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
                 inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
             }
@@ -10716,6 +11773,8 @@ struct llm_build_t5_enc : public llm_graph_context {
 
         auto * inp_attn = build_attn_inp_no_cache();
 
+        ggml_tensor * inp_out_ids = build_inp_out_ids();
+
         for (int il = 0; il < n_layer; ++il) {
             ggml_tensor * inpSA = inpL;
 
@@ -10749,9 +11808,7 @@ struct llm_build_t5_enc : public llm_graph_context {
                 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();
+            if (il == n_layer - 1 && inp_out_ids) {
                 cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
                 inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
             }
@@ -10822,6 +11879,8 @@ struct llm_build_t5_dec : public llm_graph_context {
         auto * inp_attn_self  = build_attn_inp_kv_unified();
         auto * inp_attn_cross = build_attn_inp_cross();
 
+        ggml_tensor * inp_out_ids = build_inp_out_ids();
+
         for (int il = 0; il < n_layer; ++il) {
             ggml_tensor * inpSA = inpL;
 
@@ -10913,11 +11972,8 @@ struct llm_build_t5_dec : public llm_graph_context {
                 //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();
+            if (il == n_layer - 1 && 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);
             }
 
@@ -10987,6 +12043,8 @@ struct llm_build_jais : public llm_graph_context {
 
         auto * inp_attn = build_attn_inp_kv_unified();
 
+        ggml_tensor * inp_out_ids = build_inp_out_ids();
+
         for (int il = 0; il < n_layer; ++il) {
             cur = build_norm(inpL,
                     model.layers[il].attn_norm,
@@ -11019,9 +12077,7 @@ struct llm_build_jais : public llm_graph_context {
                         Qcur, Kcur, Vcur, nullptr, 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();
+            if (il == n_layer - 1 && inp_out_ids) {
                 cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
                 inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
             }
@@ -11085,6 +12141,8 @@ struct llm_build_chatglm : public llm_graph_context {
 
         auto * inp_attn = build_attn_inp_kv_unified();
 
+        ggml_tensor * inp_out_ids = build_inp_out_ids();
+
         for (int il = 0; il < n_layer; ++il) {
             ggml_tensor * inpSA = inpL;
 
@@ -11151,9 +12209,7 @@ struct llm_build_chatglm : public llm_graph_context {
                         Qcur, Kcur, Vcur, nullptr, 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();
+            if (il == n_layer - 1 && inp_out_ids) {
                 cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
                 inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
             }
@@ -11218,6 +12274,8 @@ struct llm_build_glm4 : public llm_graph_context {
 
         auto * inp_attn = build_attn_inp_kv_unified();
 
+        ggml_tensor * inp_out_ids = build_inp_out_ids();
+
         for (int il = 0; il < n_layer; ++il) {
             ggml_tensor * inpSA = inpL;
 
@@ -11284,9 +12342,7 @@ struct llm_build_glm4 : public llm_graph_context {
                         Qcur, Kcur, Vcur, nullptr, 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();
+            if (il == n_layer - 1 && inp_out_ids) {
                 cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
                 inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
             }
@@ -11369,6 +12425,8 @@ struct llm_build_nemotron : public llm_graph_context {
 
         auto * inp_attn = build_attn_inp_kv_unified();
 
+        ggml_tensor * inp_out_ids = build_inp_out_ids();
+
         for (int il = 0; il < n_layer; ++il) {
             ggml_tensor * inpSA = inpL;
 
@@ -11428,9 +12486,7 @@ struct llm_build_nemotron : public llm_graph_context {
                         Qcur, Kcur, Vcur, nullptr, 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();
+            if (il == n_layer - 1 && inp_out_ids) {
                 cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
                 inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
             }
@@ -11498,6 +12554,8 @@ struct llm_build_exaone : public llm_graph_context {
 
         auto * inp_attn = build_attn_inp_kv_unified();
 
+        ggml_tensor * inp_out_ids = build_inp_out_ids();
+
         for (int il = 0; il < n_layer; ++il) {
             ggml_tensor * inpSA = inpL;
 
@@ -11559,9 +12617,7 @@ struct llm_build_exaone : public llm_graph_context {
                         Qcur, Kcur, Vcur, nullptr, 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();
+            if (il == n_layer - 1 && inp_out_ids) {
                 cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
                 inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
             }
@@ -11648,14 +12704,13 @@ struct llm_build_rwkv6_base : public llm_graph_context {
     }
 
     ggml_tensor * build_rwkv6_time_mix(
+            llm_graph_input_rs * inp,
             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 auto * kv_state = static_cast<const llama_kv_cache_recurrent_state *>(mstate);
+        const auto * mctx_cur = static_cast<const llama_memory_recurrent_context *>(mctx);
 
         const auto n_tokens = ubatch.n_tokens;
         const auto n_seqs = ubatch.n_seqs;
@@ -11665,7 +12720,7 @@ struct llm_build_rwkv6_base : public llm_graph_context {
         const auto n_head = n_embd / head_size;
         const auto n_head_kv = hparams.n_head_kv(il);
 
-        const auto kv_head = kv_state->get_head();
+        const auto kv_head = mctx_cur->get_head();
 
         const auto & layer = model.layers[il];
 
@@ -11776,9 +12831,9 @@ struct llm_build_rwkv6_base : public llm_graph_context {
             k = ggml_sub(ctx0, k, ggml_mul(ctx0, k, w));
         }
 
-        ggml_tensor * wkv_state = build_copy_mask_state(
-                gf, kv_state->get_v_l(il), state_copy, state_mask,
-                hparams.n_embd_v_s(), n_seqs);
+        ggml_tensor * wkv_state = build_rs(
+                inp, gf, mctx_cur->get_s_l(il),
+                hparams.n_embd_s(), n_seqs);
 
         ggml_tensor * wkv_output;
         if (is_qrwkv) {
@@ -11796,9 +12851,9 @@ struct llm_build_rwkv6_base : public llm_graph_context {
                     wkv_state,
                     ggml_view_1d(
                         ctx0,
-                        kv_state->get_v_l(il),
-                        hparams.n_embd_v_s() * n_seqs,
-                        hparams.n_embd_v_s() * kv_head * ggml_element_size(kv_state->get_v_l(il))
+                        mctx_cur->get_s_l(il),
+                        hparams.n_embd_s() * n_seqs,
+                        hparams.n_embd_s() * kv_head * ggml_element_size(mctx_cur->get_s_l(il))
                         )
                     )
                 );
@@ -11832,20 +12887,19 @@ struct llm_build_rwkv6 : public llm_build_rwkv6_base {
         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();
+        auto * rs_inp = build_rs_inp();
 
         const auto n_embd = hparams.n_embd;
         const auto n_seq_tokens = ubatch.n_seq_tokens;
         const auto n_seqs = ubatch.n_seqs;
 
+        ggml_tensor * inp_out_ids = build_inp_out_ids();
+
         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 * token_shift = build_rwkv_token_shift_load(rs_inp, gf, 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));
@@ -11860,7 +12914,7 @@ struct llm_build_rwkv6 : public llm_build_rwkv6_base {
                     1
                     );
 
-            cur = build_rwkv6_time_mix(gf, att_norm, x_prev, state_copy, state_mask, ubatch, il);
+            cur = build_rwkv6_time_mix(rs_inp, gf, att_norm, x_prev, ubatch, il);
 
             ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL);
             cb(ffn_inp, "ffn_inp", il);
@@ -11882,13 +12936,16 @@ struct llm_build_rwkv6 : public llm_build_rwkv6_base {
                     );
             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);
+            ffn_inp  = ggml_reshape_2d(ctx0, ffn_inp,  n_embd, n_tokens);
+            ffn_norm = ggml_reshape_2d(ctx0, ffn_norm, n_embd, n_tokens);
+            x_prev   = ggml_reshape_2d(ctx0, x_prev,   n_embd, n_tokens);
+            cur      = ggml_reshape_2d(ctx0, cur,      n_embd, n_tokens);
+
+            if (il == n_layer - 1 && inp_out_ids) {
+                ffn_inp  = ggml_get_rows(ctx0, ffn_inp,  inp_out_ids);
+                ffn_norm = ggml_get_rows(ctx0, ffn_norm, inp_out_ids);
+                x_prev   = ggml_get_rows(ctx0, x_prev,   inp_out_ids);
+                cur      = ggml_get_rows(ctx0, cur,      inp_out_ids);
             }
 
             cur = build_rwkv6_channel_mix(layer, ffn_norm, x_prev, LLM_ARCH_RWKV6);
@@ -11923,27 +12980,26 @@ struct llm_build_rwkv6 : public llm_build_rwkv6_base {
 // 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_ASSERT(n_embd == hparams.n_embd_r());
 
         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();
+        auto * rs_inp = build_rs_inp();
 
         const auto n_embd = hparams.n_embd;
         const auto n_seq_tokens = ubatch.n_seq_tokens;
         const auto n_seqs = ubatch.n_seqs;
 
+        ggml_tensor * inp_out_ids = build_inp_out_ids();
+
         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 * token_shift = build_rwkv_token_shift_load(rs_inp, gf, 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);
@@ -11955,7 +13011,7 @@ struct llm_build_rwkv6qwen2 : public llm_build_rwkv6_base {
                     1
                     );
 
-            cur = build_rwkv6_time_mix(gf, att_norm, x_prev, state_copy, state_mask, ubatch, il);
+            cur = build_rwkv6_time_mix(rs_inp, gf, att_norm, x_prev, 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));
@@ -11963,11 +13019,12 @@ struct llm_build_rwkv6qwen2 : public llm_build_rwkv6_base {
             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);
+            cur     = ggml_reshape_2d(ctx0, cur,     n_embd, n_tokens);
+            ffn_inp = ggml_reshape_2d(ctx0, ffn_inp, n_embd, n_tokens);
+
+            if (il == n_layer - 1 && inp_out_ids) {
+                cur     = ggml_get_rows(ctx0, cur,     inp_out_ids);
+                ffn_inp = ggml_get_rows(ctx0, ffn_inp, inp_out_ids);
             }
 
             // feed-forward network
@@ -12043,15 +13100,14 @@ struct llm_build_rwkv7_base : public llm_graph_context {
     }
 
     ggml_tensor * build_rwkv7_time_mix(
+            llm_graph_input_rs * inp,
             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 auto * kv_state = static_cast<const llama_kv_cache_recurrent_state *>(mstate);
+        const auto * mctx_cur = static_cast<const llama_memory_recurrent_context *>(mctx);
 
         const auto n_tokens = ubatch.n_tokens;
         const auto n_seqs = ubatch.n_seqs;
@@ -12060,7 +13116,7 @@ struct llm_build_rwkv7_base : public llm_graph_context {
         const auto head_count = n_embd / head_size;
         const auto n_seq_tokens = ubatch.n_seq_tokens;
 
-        const auto kv_head = kv_state->get_head();
+        const auto kv_head = mctx_cur->get_head();
 
         const auto & layer = model.layers[il];
 
@@ -12130,9 +13186,9 @@ struct llm_build_rwkv7_base : public llm_graph_context {
         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_state->get_v_l(il), state_copy, state_mask,
-                hparams.n_embd_v_s(), n_seqs);
+        ggml_tensor * wkv_state = build_rs(
+                inp, gf, mctx_cur->get_s_l(il),
+                hparams.n_embd_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);
@@ -12145,9 +13201,9 @@ struct llm_build_rwkv7_base : public llm_graph_context {
                     wkv_state,
                     ggml_view_1d(
                         ctx0,
-                        kv_state->get_v_l(il),
-                        hparams.n_embd_v_s() * n_seqs,
-                        hparams.n_embd_v_s() * kv_head * ggml_element_size(kv_state->get_v_l(il))
+                        mctx_cur->get_s_l(il),
+                        hparams.n_embd_s() * n_seqs,
+                        hparams.n_embd_s() * kv_head * ggml_element_size(mctx_cur->get_s_l(il))
                         )
                     )
                 );
@@ -12188,20 +13244,19 @@ struct llm_build_rwkv7 : public llm_build_rwkv7_base {
         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();
+        auto * rs_inp = build_rs_inp();
 
         const auto n_embd = hparams.n_embd;
         const auto n_seq_tokens = ubatch.n_seq_tokens;
         const auto n_seqs = ubatch.n_seqs;
 
+        ggml_tensor * inp_out_ids = build_inp_out_ids();
+
         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 * token_shift = build_rwkv_token_shift_load(rs_inp, gf, 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));
@@ -12216,7 +13271,7 @@ struct llm_build_rwkv7 : public llm_build_rwkv7_base {
                     1
                     );
 
-            cur = build_rwkv7_time_mix(gf, att_norm, x_prev, state_copy, state_mask, v_first, ubatch, il);
+            cur = build_rwkv7_time_mix(rs_inp, gf, att_norm, x_prev, v_first, ubatch, il);
 
             ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL);
             cb(ffn_inp, "ffn_inp", il);
@@ -12238,12 +13293,14 @@ struct llm_build_rwkv7 : public llm_build_rwkv7_base {
                     );
             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);
+            ffn_inp  = ggml_reshape_2d(ctx0, ffn_inp,  n_embd, n_tokens);
+            ffn_norm = ggml_reshape_2d(ctx0, ffn_norm, n_embd, n_tokens);
+            x_prev   = ggml_reshape_2d(ctx0, x_prev,   n_embd, n_tokens);
+
+            if (il == n_layer - 1 && inp_out_ids) {
+                ffn_inp  = ggml_get_rows(ctx0, ffn_inp,  inp_out_ids);
+                ffn_norm = ggml_get_rows(ctx0, ffn_norm, inp_out_ids);
+                x_prev   = ggml_get_rows(ctx0, x_prev,   inp_out_ids);
             }
 
             cur = build_rwkv7_channel_mix(layer, ffn_norm, x_prev, LLM_ARCH_RWKV7);
@@ -12274,7 +13331,7 @@ struct llm_build_rwkv7 : public llm_build_rwkv7_base {
 
 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_ASSERT(n_embd == hparams.n_embd_r());
 
         ggml_tensor * cur;
         ggml_tensor * inpL;
@@ -12282,20 +13339,19 @@ struct llm_build_arwkv7 : public llm_build_rwkv7_base {
 
         inpL = build_inp_embd(model.tok_embd);
 
-        ggml_tensor * state_copy = build_inp_s_copy();
-        ggml_tensor * state_mask = build_inp_s_mask();
+        auto * rs_inp = build_rs_inp();
 
         const auto n_embd = hparams.n_embd;
         const auto n_seq_tokens = ubatch.n_seq_tokens;
         const auto n_seqs = ubatch.n_seqs;
 
+        ggml_tensor * inp_out_ids = build_inp_out_ids();
+
         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 * token_shift = build_rwkv_token_shift_load(rs_inp, gf, 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);
@@ -12307,7 +13363,7 @@ struct llm_build_arwkv7 : public llm_build_rwkv7_base {
                     1
                     );
 
-            cur = build_rwkv7_time_mix(gf, att_norm, x_prev, state_copy, state_mask, v_first, ubatch, il);
+            cur = build_rwkv7_time_mix(rs_inp, gf, att_norm, x_prev, 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));
@@ -12315,11 +13371,12 @@ struct llm_build_arwkv7 : public llm_build_rwkv7_base {
             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);
+            cur     = ggml_reshape_2d(ctx0, cur,     n_embd, n_tokens);
+            ffn_inp = ggml_reshape_2d(ctx0, ffn_inp, n_embd, n_tokens);
+
+            if (il == n_layer - 1 && inp_out_ids) {
+                cur     = ggml_get_rows(ctx0, cur,     inp_out_ids);
+                ffn_inp = ggml_get_rows(ctx0, ffn_inp, inp_out_ids);
             }
 
             // feed-forward network
@@ -12388,6 +13445,9 @@ struct llm_build_granite : public llm_graph_context {
         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;
+
+        ggml_tensor * inp_out_ids = build_inp_out_ids();
+
         for (int il = 0; il < n_layer; ++il) {
             ggml_tensor * inpSA = inpL;
 
@@ -12450,9 +13510,7 @@ struct llm_build_granite : public llm_graph_context {
                 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();
+            if (il == n_layer - 1 && inp_out_ids) {
                 cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
                 inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
             }
@@ -12571,6 +13629,8 @@ struct llm_build_chameleon : public llm_graph_context {
 
         auto * inp_attn = build_attn_inp_kv_unified();
 
+        ggml_tensor * inp_out_ids = build_inp_out_ids();
+
         for (int il = 0; il < n_layer; ++il) {
             ggml_tensor * inpSA = inpL;
 
@@ -12647,21 +13707,19 @@ struct llm_build_chameleon : public llm_graph_context {
                 cur = build_attn(inp_attn, gf,
                         model.layers[il].wo, nullptr,
                         Qcur, Kcur, Vcur, nullptr, 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();
+            if (il == n_layer - 1 && inp_out_ids) {
                 cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
                 inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
             }
 
+            if (hparams.swin_norm) {
+                cur = build_norm(cur,
+                        model.layers[il].attn_norm, NULL,
+                        LLM_NORM_RMS, il);
+            }
+
             ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
             cb(ffn_inp, "ffn_inp", il);
 
@@ -12902,6 +13960,8 @@ struct llm_build_plm : public llm_graph_context {
 
         auto * inp_attn = build_attn_inp_kv_unified();
 
+        ggml_tensor * inp_out_ids = build_inp_out_ids();
+
         for (int il = 0; il < n_layer; ++il) {
             ggml_tensor * inpSA = inpL;
 
@@ -13005,9 +14065,7 @@ struct llm_build_plm : public llm_graph_context {
                         q_states, k_states, v_states, nullptr, nullptr, kq_scale, il);
             }
 
-            if (il == n_layer - 1) {
-                // skip computing output for unused tokens
-                ggml_tensor * inp_out_ids = build_inp_out_ids();
+            if (il == n_layer - 1 && inp_out_ids) {
                 cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
                 inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
             }
@@ -13067,6 +14125,8 @@ struct llm_build_bailingmoe : public llm_graph_context {
 
         auto * inp_attn = build_attn_inp_kv_unified();
 
+        ggml_tensor * inp_out_ids = build_inp_out_ids();
+
         for (int il = 0; il < n_layer; ++il) {
             ggml_tensor * inpSA = inpL;
 
@@ -13128,9 +14188,7 @@ struct llm_build_bailingmoe : public llm_graph_context {
                         Qcur, Kcur, Vcur, nullptr, 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();
+            if (il == n_layer - 1 && inp_out_ids) {
                 cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
                 inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
             }
@@ -13199,70 +14257,506 @@ struct llm_build_bailingmoe : public llm_graph_context {
     }
 };
 
+struct llm_build_dots1 : public llm_graph_context {
+    llm_build_dots1(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();
+
+        ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+        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, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+            }
+
+            if (il == n_layer - 1 && 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);
+
+            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 {
+                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);
+
+                {
+                    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_ernie4_5 : public llm_graph_context {
+    llm_build_ernie4_5(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
+            {
+                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, 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_arcee : public llm_graph_context {
+    llm_build_arcee(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;
+
+        ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+        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 = model.get_rope_factors(cparams, 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, nullptr, kq_scale, il);
+                cb(cur, "attn_out", il);
+            }
+
+            if (il == n_layer - 1 && 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
+            // ARCEE uses relu^2 instead of silu
+            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);
+            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);
+    }
+};
+
 llama_memory_i * llama_model::create_memory(const llama_memory_params & params, llama_cparams & cparams) const {
     llama_memory_i * res;
 
     switch (arch) {
+        // Models that need specific instantiation should be handled in the
+        // switch statement
         case LLM_ARCH_BERT:
         case LLM_ARCH_JINA_BERT_V2:
         case LLM_ARCH_JINA_BERT_V3:
         case LLM_ARCH_NOMIC_BERT:
         case LLM_ARCH_NOMIC_BERT_MOE:
+        case LLM_ARCH_NEO_BERT:
         case LLM_ARCH_WAVTOKENIZER_DEC:
             {
                 res = nullptr;
             } break;
-        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_recurrent(
-                        *this,
-                        GGML_TYPE_F32,
-                        GGML_TYPE_F32,
-                        cparams.offload_kqv,
-                        std::max((uint32_t) 1, cparams.n_seq_max),
-                        cparams.n_seq_max);
-            } break;
+        // Models that need standard caching should rely on recurrent/hybrid
+        // checks
         default:
             {
-                const auto padding = llama_kv_cache_unified::get_padding(cparams);
-
-                cparams.n_ctx = GGML_PAD(cparams.n_ctx, padding);
-
-                LLAMA_LOG_DEBUG("%s: n_ctx = %u (padded)\n", __func__, cparams.n_ctx);
-
-                if (hparams.swa_type != LLAMA_SWA_TYPE_NONE) {
-                    GGML_ASSERT(hparams.is_swa_any());
-
-                    res = new llama_kv_cache_unified_iswa(
-                            *this,
-                            params.type_k,
-                            params.type_v,
-                            !cparams.flash_attn,
-                            cparams.offload_kqv,
-                            params.swa_full,
-                            cparams.n_ctx,
-                            cparams.n_seq_max,
-                            cparams.n_ubatch,
-                            padding);
-                } else {
-                    GGML_ASSERT(!hparams.is_swa_any());
-
-                    res = new llama_kv_cache_unified(
+                if (llm_arch_is_recurrent(arch)) {
+                    res = new llama_memory_recurrent(
                             *this,
                             nullptr,
-                            params.type_k,
-                            params.type_v,
-                            !cparams.flash_attn,
+                            GGML_TYPE_F32,
+                            GGML_TYPE_F32,
                             cparams.offload_kqv,
-                            cparams.n_ctx,
-                            cparams.n_seq_max,
-                            padding,
-                            hparams.n_swa,
-                            hparams.swa_type);
+                            std::max((uint32_t) 1, cparams.n_seq_max),
+                            cparams.n_seq_max);
+                } else if (llm_arch_is_hybrid(arch)) {
+                    const auto padding = llama_kv_cache_unified::get_padding(cparams);
+
+                    cparams.n_ctx = GGML_PAD(cparams.n_ctx, padding);
+
+                    res = new llama_memory_hybrid(
+                        /* model             */ *this,
+                        /* attn_type_k       */ params.type_k,
+                        /* attn_type_v       */ params.type_v,
+                        /* attn_v_trans      */ !cparams.flash_attn,
+                        /* attn_kv_size      */ cparams.n_ctx,
+                        /* attn_n_pad        */ padding,
+                        /* attn_n_swa        */ hparams.n_swa,
+                        /* attn_swa_type     */ hparams.swa_type,
+                        /* recurrent_type_k  */ GGML_TYPE_F32,
+                        /* recurrent_type_v  */ GGML_TYPE_F32,
+                        /* recurrent_kv_size */ std::max((uint32_t) 1, cparams.n_seq_max),
+                        /* n_seq_max         */ cparams.n_seq_max,
+                        /* offload           */ cparams.offload_kqv);
+                } else {
+                    const auto padding = llama_kv_cache_unified::get_padding(cparams);
+
+                    cparams.n_ctx = GGML_PAD(cparams.n_ctx, padding);
+
+                    LLAMA_LOG_DEBUG("%s: n_ctx = %u (padded)\n", __func__, cparams.n_ctx);
+
+                    if (hparams.swa_type != LLAMA_SWA_TYPE_NONE) {
+                        GGML_ASSERT(hparams.is_swa_any());
+
+                        res = new llama_kv_cache_unified_iswa(
+                                *this,
+                                params.type_k,
+                                params.type_v,
+                                !cparams.flash_attn,
+                                cparams.offload_kqv,
+                                params.swa_full,
+                                cparams.n_ctx,
+                                cparams.n_seq_max,
+                                cparams.n_ubatch,
+                                padding);
+                    } else {
+                        GGML_ASSERT(!hparams.is_swa_any());
+
+                        res = new llama_kv_cache_unified(
+                                *this,
+                                nullptr,
+                                params.type_k,
+                                params.type_v,
+                                !cparams.flash_attn,
+                                cparams.offload_kqv,
+                                cparams.n_ctx,
+                                cparams.n_seq_max,
+                                padding,
+                                hparams.n_swa,
+                                hparams.swa_type);
+                    }
                 }
             }
     }
@@ -13317,6 +14811,10 @@ llm_graph_result_ptr llama_model::build_graph(
             {
                 llm = std::make_unique<llm_build_bert>(*this, params, gf);
             } break;
+        case LLM_ARCH_NEO_BERT:
+            {
+                llm = std::make_unique<llm_build_neo_bert>(*this, params, gf);
+            } break;
         case LLM_ARCH_BLOOM:
             {
                 llm = std::make_unique<llm_build_bloom>(*this, params, gf);
@@ -13402,11 +14900,16 @@ llm_graph_result_ptr llama_model::build_graph(
             {
                 llm = std::make_unique<llm_build_gemma3_iswa>(*this, params, gf);
             } break;
+        case LLM_ARCH_GEMMA3N:
+            {
+                llm = std::make_unique<llm_build_gemma3n_iswa>(*this, params, gf);
+            } break;
         case LLM_ARCH_STARCODER2:
             {
                 llm = std::make_unique<llm_build_starcoder2>(*this, params, gf);
             } break;
         case LLM_ARCH_MAMBA:
+        case LLM_ARCH_MAMBA2:
             {
                 llm = std::make_unique<llm_build_mamba>(*this, params, gf);
             } break;
@@ -13539,6 +15042,18 @@ llm_graph_result_ptr llama_model::build_graph(
             {
                 llm = std::make_unique<llm_build_bailingmoe>(*this, params, gf);
             } break;
+        case LLM_ARCH_DOTS1:
+            {
+                llm = std::make_unique<llm_build_dots1>(*this, params, gf);
+            } break;
+        case LLM_ARCH_ARCEE:
+            {
+                llm = std::make_unique<llm_build_arcee>(*this, params, gf);
+            } break;
+        case LLM_ARCH_ERNIE4_5:
+            {
+                llm = std::make_unique<llm_build_ernie4_5>(*this, params, gf);
+            } break;
         default:
             GGML_ABORT("fatal error");
     }
@@ -13614,6 +15129,18 @@ int32_t llama_model_n_swa(const llama_model * model) {
     return model->hparams.n_swa;
 }
 
+uint32_t llama_model_n_cls_out(const struct llama_model * model) {
+    return model->hparams.n_cls_out;
+}
+
+const char * llama_model_cls_label(const struct llama_model * model, uint32_t i) {
+    if (i < model->classifier_labels.size()) {
+        return model->classifier_labels[i].c_str();
+    }
+
+    return nullptr;
+}
+
 // deprecated
 int32_t llama_n_ctx_train(const llama_model * model) {
     return llama_model_n_ctx_train(model);
@@ -13643,6 +15170,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
         case LLM_ARCH_REFACT:
         case LLM_ARCH_BLOOM:
         case LLM_ARCH_MAMBA:
+        case LLM_ARCH_MAMBA2:
         case LLM_ARCH_JINA_BERT_V2:
         case LLM_ARCH_T5:
         case LLM_ARCH_T5ENCODER:
@@ -13676,6 +15204,9 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
         case LLM_ARCH_GRANITE_MOE:
         case LLM_ARCH_CHAMELEON:
         case LLM_ARCH_BAILINGMOE:
+        case LLM_ARCH_NEO_BERT:
+        case LLM_ARCH_ARCEE:
+        case LLM_ARCH_ERNIE4_5:
             return LLAMA_ROPE_TYPE_NORM;
 
         // the pairs of head values are offset by n_rot/2
@@ -13702,6 +15233,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
         case LLM_ARCH_GEMMA:
         case LLM_ARCH_GEMMA2:
         case LLM_ARCH_GEMMA3:
+        case LLM_ARCH_GEMMA3N:
         case LLM_ARCH_STARCODER2:
         case LLM_ARCH_OPENELM:
         case LLM_ARCH_GPTNEOX:
@@ -13710,6 +15242,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
         case LLM_ARCH_NEMOTRON:
         case LLM_ARCH_EXAONE:
         case LLM_ARCH_MINICPM3:
+        case LLM_ARCH_DOTS1:
             return LLAMA_ROPE_TYPE_NEOX;
 
         case LLM_ARCH_QWEN2VL:
@@ -13775,7 +15308,7 @@ uint64_t llama_model_size(const llama_model * model) {
 }
 
 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)
+    const auto key = name ? LLM_KV(model->arch, name)(LLM_KV_TOKENIZER_CHAT_TEMPLATE)
         : LLM_KV(model->arch)(LLM_KV_TOKENIZER_CHAT_TEMPLATE);
     const auto & it = model->gguf_kv.find(key);
     if (it == model->gguf_kv.end()) {
@@ -13783,7 +15316,7 @@ const char * llama_model_chat_template(const llama_model * model, const char * n
         // do not extend this list unless absolutely necessary
         // Mistral-Small-2503 does not have built-in chat template
         llama_vocab_pre_type pre_type = model->vocab.get_pre_type();
-        if (pre_type == LLAMA_VOCAB_PRE_TYPE_TEKKEN && model->layers.size() == 40) {
+        if (!name && pre_type == LLAMA_VOCAB_PRE_TYPE_TEKKEN && model->layers.size() == 40) {
             return "mistral-v7-tekken";
         }
 
@@ -13817,14 +15350,7 @@ llama_token llama_model_decoder_start_token(const 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;
-        case     LLM_ARCH_RWKV7:      return true;
-        case     LLM_ARCH_ARWKV7:     return true;
-        default: return false;
-    }
+    return llm_arch_is_recurrent(model->arch);
 }
 
 const std::vector<std::pair<std::string, ggml_tensor *>> & llama_internal_get_tensor_map(const llama_model * model) {
diff --git a/src/llama-model.h b/src/llama-model.h
index 77e36f12f..9a7d8727b 100644
--- a/src/llama-model.h
+++ b/src/llama-model.h
@@ -40,6 +40,7 @@ enum llm_type {
     LLM_TYPE_558M,
     LLM_TYPE_770M,
     LLM_TYPE_780M,
+    LLM_TYPE_0_3B,
     LLM_TYPE_0_5B,
     LLM_TYPE_0_6B,
     LLM_TYPE_1B,
@@ -74,6 +75,7 @@ enum llm_type {
     LLM_TYPE_40B,
     LLM_TYPE_65B,
     LLM_TYPE_70B,
+    LLM_TYPE_142B,
     LLM_TYPE_236B,
     LLM_TYPE_290B,
     LLM_TYPE_314B,
@@ -95,6 +97,8 @@ enum llm_type {
     LLM_TYPE_17B_128E, // llama4 Maverick
     LLM_TYPE_30B_A3B,
     LLM_TYPE_235B_A22B,
+    LLM_TYPE_E2B,
+    LLM_TYPE_E4B,
 };
 
 std::string llama_rope_scaling_type_name(llama_rope_scaling_type rope_scaling_type);
@@ -169,6 +173,7 @@ struct llama_layer {
     struct ggml_tensor * ffn_sub_norm    = nullptr;
     struct ggml_tensor * attn_norm_cross = nullptr;
     struct ggml_tensor * attn_norm_enc   = nullptr;
+    struct ggml_tensor * ssm_norm        = nullptr;
 
     // attention
     struct ggml_tensor * wq        = nullptr;
@@ -316,6 +321,19 @@ struct llama_layer {
     struct ggml_tensor * ffn_up_scale   = nullptr;
     struct ggml_tensor * ffn_down_scale = nullptr;
 
+    // altup & laurel
+    struct ggml_tensor * per_layer_inp_gate   = nullptr;
+    struct ggml_tensor * per_layer_proj       = nullptr;
+    struct ggml_tensor * per_layer_post_norm  = nullptr;
+    struct ggml_tensor * altup_correct_coef   = nullptr;
+    struct ggml_tensor * altup_correct_scale  = nullptr;
+    struct ggml_tensor * altup_predict_coef   = nullptr;
+    struct ggml_tensor * altup_router         = nullptr;
+    struct ggml_tensor * altup_router_norm    = nullptr;
+    struct ggml_tensor * laurel_l             = nullptr;
+    struct ggml_tensor * laurel_r             = nullptr;
+    struct ggml_tensor * laurel_post_norm     = nullptr;
+
     struct llama_layer_posnet posnet;
 
     struct llama_layer_convnext convnext;
@@ -330,6 +348,9 @@ struct llama_model {
     llama_hparams hparams = {};
     llama_vocab   vocab;
 
+    // for classifier models
+    std::vector<std::string> classifier_labels;
+
     struct ggml_tensor * tok_embd   = nullptr;
     struct ggml_tensor * type_embd  = nullptr;
     struct ggml_tensor * pos_embd   = nullptr;
@@ -351,6 +372,13 @@ struct llama_model {
     struct ggml_tensor * conv1d   = nullptr;
     struct ggml_tensor * conv1d_b = nullptr;
 
+    // gemma3n altup
+    struct ggml_tensor * tok_embd_per_layer   = nullptr;
+    struct ggml_tensor * altup_proj           = nullptr;
+    struct ggml_tensor * altup_unembd_proj    = nullptr;
+    struct ggml_tensor * per_layer_model_proj = nullptr;
+    struct ggml_tensor * per_layer_proj_norm  = nullptr;
+
     std::vector<llama_layer> layers;
 
     llama_model_params params;
diff --git a/src/llama-quant.cpp b/src/llama-quant.cpp
index 159b1307a..f4b5713d7 100644
--- a/src/llama-quant.cpp
+++ b/src/llama-quant.cpp
@@ -1,5 +1,4 @@
 #include "llama-quant.h"
-
 #include "llama-impl.h"
 #include "llama-model.h"
 #include "llama-model-loader.h"
@@ -27,6 +26,56 @@ static void zeros(std::ofstream & file, size_t n) {
     }
 }
 
+static std::string remap_layer(const std::string & orig_name, const std::vector<int> & prune, std::map<int, std::string> & mapped, int & next_id) {
+    if (prune.empty()) {
+        return orig_name;
+    }
+
+    static const std::regex pattern(R"(blk\.(\d+)\.)");
+    if (std::smatch match; std::regex_search(orig_name, match, pattern)) {
+        const int blk = std::stoi(match[1]);
+        std::string new_name = orig_name;
+
+        if (mapped.count(blk)) {
+            // Already mapped, do nothing
+        } else if (std::find(prune.begin(), prune.end(), blk) != prune.end()) {
+            mapped[blk] = "";
+        } else if (blk < prune.front()) {
+            mapped[blk] = std::to_string(blk);
+            next_id = blk + 1;
+        } else {
+            mapped[blk] = std::to_string(next_id);
+            ++next_id;
+        }
+
+        return mapped[blk].empty() ? mapped[blk] : new_name.replace(match.position(1), match.length(1), mapped[blk]);
+    }
+
+    return orig_name;
+}
+
+static std::string remap_imatrix (const std::string & orig_name, const std::map<int, std::string> & mapped) {
+    if (mapped.empty()) {
+        return orig_name;
+    }
+
+    static const std::regex pattern(R"(blk\.(\d+)\.)");
+    if (std::smatch match; std::regex_search(orig_name, match, pattern)) {
+        const std::string blk(match[1]);
+        std::string new_name = orig_name;
+
+        for (const auto & p : mapped) {
+            if (p.second == blk) {
+                LLAMA_LOG_DEBUG("(blk.%d imatrix) ", p.first);
+                return new_name.replace(match.position(1), match.length(1), std::to_string(p.first));
+            }
+        }
+        GGML_ABORT("\n%s: imatrix mapping error for %s\n", __func__, orig_name.c_str());
+    }
+
+    return orig_name;
+}
+
 struct quantize_state_impl {
     const llama_model                 & model;
     const llama_model_quantize_params * params;
@@ -174,7 +223,7 @@ static ggml_type llama_tensor_get_type(quantize_state_impl & qs, ggml_type new_t
                 new_type = GGML_TYPE_Q6_K;
             }
         }
-    } else if (name == "token_embd.weight") {
+    } else if (name == "token_embd.weight" || name == "per_layer_token_embd.weight") {
         if (qs.params->token_embedding_type < GGML_TYPE_COUNT) {
             new_type = qs.params->token_embedding_type;
         } else {
@@ -568,6 +617,11 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
     const size_t align = GGUF_DEFAULT_ALIGNMENT;
     gguf_context_ptr ctx_out { gguf_init_empty() };
 
+    std::vector<int> prune_list = {};
+    if (params->prune_layers) {
+        prune_list = *static_cast<const std::vector<int> *>(params->prune_layers);
+    }
+
     // copy the KV pairs from the input file
     gguf_set_kv     (ctx_out.get(), ml.meta.get());
     gguf_set_val_u32(ctx_out.get(), "general.quantization_version", GGML_QNT_VERSION); // TODO: use LLM_KV
@@ -585,7 +639,8 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
             if (o.tag == LLAMA_KV_OVERRIDE_TYPE_FLOAT) {
                 gguf_set_val_f32(ctx_out.get(), o.key, o.val_f64);
             } else if (o.tag == LLAMA_KV_OVERRIDE_TYPE_INT) {
-                gguf_set_val_i32(ctx_out.get(), o.key, o.val_i64);
+                // Setting type to UINT32. See https://github.com/ggml-org/llama.cpp/pull/14182 for context
+                gguf_set_val_u32(ctx_out.get(), o.key, (uint32_t)abs(o.val_i64));
             } else if (o.tag == LLAMA_KV_OVERRIDE_TYPE_BOOL) {
                 gguf_set_val_bool(ctx_out.get(), o.key, o.val_bool);
             } else if (o.tag == LLAMA_KV_OVERRIDE_TYPE_STR) {
@@ -596,12 +651,32 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
         }
     }
 
+    std::map<int, std::string> mapped;
+    int blk_id = 0;
+    int pruned_attention_w = 0;
+
     // make a list of weights
     std::vector<const llama_model_loader::llama_tensor_weight *> tensors;
     tensors.reserve(ml.weights_map.size());
     for (const auto & it : ml.weights_map) {
+        const std::string remapped_name(remap_layer(it.first, prune_list, mapped, blk_id));
+        if (remapped_name.empty()) {
+            if (it.first.find("attn_v.weight") != std::string::npos ||
+                it.first.find("attn_qkv.weight") != std::string::npos ||
+                it.first.find("attn_kv_b.weight") != std::string::npos) {
+                    pruned_attention_w++;
+            }
+            LLAMA_LOG_DEBUG("%s: pruning tensor %s\n", __func__, it.first.c_str());
+            continue;
+        } else if (remapped_name != it.first) {
+            ggml_set_name(it.second.tensor, remapped_name.c_str());
+            LLAMA_LOG_DEBUG("%s: tensor %s remapped to %s\n", __func__, it.first.c_str(), ggml_get_name(it.second.tensor));
+        }
         tensors.push_back(&it.second);
     }
+    if (!prune_list.empty()) {
+        gguf_set_val_u32(ctx_out.get(), ml.llm_kv(LLM_KV_BLOCK_COUNT).c_str(), blk_id);
+    }
 
     // keep_split requires that the weights are sorted by split index
     if (params->keep_split) {
@@ -639,7 +714,7 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
         if (llama_model_has_encoder(&model)) {
             n_attn_layer *= 3;
         }
-        GGML_ASSERT((qs.n_attention_wv == n_attn_layer) && "n_attention_wv is unexpected");
+        GGML_ASSERT((qs.n_attention_wv == n_attn_layer - pruned_attention_w) && "n_attention_wv is unexpected");
     }
 
     size_t total_size_org = 0;
@@ -680,7 +755,7 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
         for (size_t i = 0; i < ctx_outs.size(); ++i) {
             gguf_set_val_u16(ctx_outs[i].get(), ml.llm_kv(LLM_KV_SPLIT_NO).c_str(), i);
             gguf_set_val_u16(ctx_outs[i].get(), ml.llm_kv(LLM_KV_SPLIT_COUNT).c_str(), n_split);
-            gguf_set_val_i32(ctx_outs[i].get(), ml.llm_kv(LLM_KV_SPLIT_TENSORS_COUNT).c_str(), ml.n_tensors);
+            gguf_set_val_i32(ctx_outs[i].get(), ml.llm_kv(LLM_KV_SPLIT_TENSORS_COUNT).c_str(), (int32_t)tensors.size());
         }
     }
 
@@ -755,6 +830,13 @@ 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("ffn_gate_inp.weight") == std::string::npos;
 
+        // these are very small (e.g. 4x4)
+        quantize &= name.find("altup")  == std::string::npos;
+        quantize &= name.find("laurel") == std::string::npos;
+
+        // these are not too big so keep them as it is
+        quantize &= name.find("per_layer_model_proj") == std::string::npos;
+
         // do not quantize positional embeddings and token types (BERT)
         quantize &= name != LLM_TN(model.arch)(LLM_TENSOR_POS_EMBD,    "weight");
         quantize &= name != LLM_TN(model.arch)(LLM_TENSOR_TOKEN_TYPES, "weight");
@@ -831,7 +913,7 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
 
             const float * imatrix = nullptr;
             if (imatrix_data) {
-                auto it = imatrix_data->find(tensor->name);
+                auto it = imatrix_data->find(remap_imatrix(tensor->name, mapped));
                 if (it == imatrix_data->end()) {
                     LLAMA_LOG_INFO("\n====== %s: did not find weights for %s\n", __func__, tensor->name);
                 } else {
@@ -946,6 +1028,7 @@ llama_model_quantize_params llama_model_quantize_default_params() {
         /*.imatrix                     =*/ nullptr,
         /*.kv_overrides                =*/ nullptr,
         /*.tensor_type                 =*/ nullptr,
+        /*.prune_layers                =*/ nullptr
     };
 
     return result;
diff --git a/src/llama-vocab.cpp b/src/llama-vocab.cpp
index 00d26ff85..29e8acf5c 100644
--- a/src/llama-vocab.cpp
+++ b/src/llama-vocab.cpp
@@ -9,16 +9,16 @@
 
 #include <algorithm>
 #include <cassert>
+#include <cctype>
 #include <cfloat>
-#include <climits>
 #include <cstdarg>
 #include <cstring>
 #include <forward_list>
+#include <limits>
 #include <map>
 #include <queue>
 #include <set>
 #include <unordered_map>
-#include <cctype>
 
 //
 // helpers
@@ -1269,6 +1269,7 @@ struct llama_vocab::impl {
     bool add_space_prefix           = false;
     bool add_bos                    = false;
     bool add_eos                    = false;
+    bool add_sep                    = false;
     bool ignore_merges              = false;
     bool clean_spaces               = false;  // clean_up_tokenization_spaces
     bool remove_extra_whitespaces   = false;
@@ -1421,6 +1422,8 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
             special_sep_id  = 102;
             special_pad_id  = 0;
             special_mask_id = 103;
+
+            add_sep = true;
         } else if (tokenizer_model == "gpt2") {
             type = LLAMA_VOCAB_TYPE_BPE;
 
@@ -1550,12 +1553,15 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                     tokenizer_pre == "jina-es" ||
                     tokenizer_pre == "jina-de" ||
                     tokenizer_pre == "gigachat"   ||
-                    tokenizer_pre == "jina-v1-en" ||
                     tokenizer_pre == "jina-v2-es" ||
-                    tokenizer_pre == "jina-v2-de" ||
+                    tokenizer_pre == "jina-v2-de") {
+                pre_type = LLAMA_VOCAB_PRE_TYPE_GPT2;
+            } else if (
+                    tokenizer_pre == "jina-v1-en" ||
                     tokenizer_pre == "jina-v2-code" ||
                     tokenizer_pre == "roberta-bpe") {
                 pre_type = LLAMA_VOCAB_PRE_TYPE_GPT2;
+                add_sep = true;
             } else if (
                     tokenizer_pre == "refact") {
                 pre_type = LLAMA_VOCAB_PRE_TYPE_REFACT;
@@ -1665,6 +1671,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
             clean_spaces = true;
             add_bos = true;
             add_eos = false;
+            add_sep = true;
         } else if (type == LLAMA_VOCAB_TYPE_UGM) {
             pre_type = LLAMA_VOCAB_PRE_TYPE_DEFAULT;
             add_bos = false;
@@ -1801,7 +1808,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
             }
         }
 
-        // Handle add_bos and add_eos
+        // Handle add_bos, add_eos and add_sep
         {
             bool temp = true;
 
@@ -1811,6 +1818,9 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
             if (ml.get_key(LLM_KV_TOKENIZER_ADD_EOS, temp, false)) {
                 add_eos = temp;
             }
+            if (ml.get_key(LLM_KV_TOKENIZER_ADD_SEP, temp, false)) {
+                add_sep = temp;
+            }
         }
 
         // auto-detect special tokens by text
@@ -1987,6 +1997,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                     || t.first == "<|eom_id|>"
                     || t.first == "<EOT>"
                     || t.first == "_<EOT>"
+                    || t.first == "<|end_of_text|>"
                ) {
                 special_eog_ids.insert(t.second);
                 if ((id_to_token[t.second].attr & LLAMA_TOKEN_ATTR_CONTROL) == 0) {
@@ -2059,9 +2070,9 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
     //NOTE: Per token attributes are missing from the GGUF file.
     //TODO: Extract attributes from GGUF file.
     {
-        auto _contains_any = [] (const std::string & str, const std::vector<std::string> & substrs) -> bool {
+        auto _contains_any = [] (const std::string & str, const std::vector<std::string_view> & substrs) -> bool {
             for (const auto & substr : substrs) {
-                if (str.find(substr) < std::string::npos) {
+                if (str.find(substr) != std::string::npos) {
                     return true;
                 }
             }
@@ -2098,7 +2109,11 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                 || _contains_any(tokenizer_pre, {"jina-v2-de", "jina-v2-es", "jina-v2-code"})
                 || _contains_any(general_arch, {"nomic-bert-moe", "jina-bert-v3"})
            ) {
-            _set_token_attr("<mask>", LLAMA_TOKEN_ATTR_LSTRIP, true);
+            if (token_to_id.count("<mask>") == 0) {
+                LLAMA_LOG_WARN("%s: Mask token is missing in vocab, please reconvert model!\n", __func__);
+            } else {
+                _set_token_attr("<mask>", LLAMA_TOKEN_ATTR_LSTRIP, true);
+            }
         } else if (_contains_any(model_name, {"phi-3", "phi3"})) {
             for (auto id : cache_special_tokens) {
                 _set_tokenid_attr(id, LLAMA_TOKEN_ATTR_RSTRIP, true);
@@ -2568,6 +2583,10 @@ int32_t llama_vocab::impl::token_to_piece(llama_token token, char * buf, int32_t
     // copy piece chars to output text buffer
     // skip up to 'lstrip' leading spaces before copying
     auto _try_copy = [=] (const char * token, size_t size) -> int32_t {
+        if (size >= static_cast<size_t>(std::numeric_limits<int32_t>::max())) {
+            GGML_ABORT("invalid token size: %zu exceeds int32_t limit", size);
+        }
+
         for (int32_t i = 0; i < lstrip && size && *token == ' '; ++i) {
             token++;
             size--;
@@ -2764,26 +2783,26 @@ void llama_vocab::impl::print_info() const {
     LLAMA_LOG_INFO("%s: n_merges         = %u\n",     __func__, (uint32_t) bpe_ranks.size());
 
     // special tokens
-    if (special_bos_id  != LLAMA_TOKEN_NULL)    { LLAMA_LOG_INFO( "%s: BOS token        = %d '%s'\n", __func__, special_bos_id,     id_to_token[special_bos_id].text.c_str() );  }
-    if (special_eos_id  != LLAMA_TOKEN_NULL)    { LLAMA_LOG_INFO( "%s: EOS token        = %d '%s'\n", __func__, special_eos_id,     id_to_token[special_eos_id].text.c_str() );  }
-    if (special_eot_id  != LLAMA_TOKEN_NULL)    { LLAMA_LOG_INFO( "%s: EOT token        = %d '%s'\n", __func__, special_eot_id,     id_to_token[special_eot_id].text.c_str() );  }
-    if (special_eom_id  != LLAMA_TOKEN_NULL)    { LLAMA_LOG_INFO( "%s: EOM token        = %d '%s'\n", __func__, special_eom_id,     id_to_token[special_eom_id].text.c_str() );  }
-    if (special_unk_id  != LLAMA_TOKEN_NULL)    { LLAMA_LOG_INFO( "%s: UNK token        = %d '%s'\n", __func__, special_unk_id,     id_to_token[special_unk_id].text.c_str() );  }
-    if (special_sep_id  != LLAMA_TOKEN_NULL)    { LLAMA_LOG_INFO( "%s: SEP token        = %d '%s'\n", __func__, special_sep_id,     id_to_token[special_sep_id].text.c_str() );  }
-    if (special_pad_id  != LLAMA_TOKEN_NULL)    { LLAMA_LOG_INFO( "%s: PAD token        = %d '%s'\n", __func__, special_pad_id,     id_to_token[special_pad_id].text.c_str() );  }
-    if (special_mask_id != LLAMA_TOKEN_NULL)    { LLAMA_LOG_INFO( "%s: MASK token       = %d '%s'\n", __func__, special_mask_id,    id_to_token[special_mask_id].text.c_str() ); }
+    if (special_bos_id  != LLAMA_TOKEN_NULL)    { LLAMA_LOG_INFO( "%s: BOS token        = %d '%s'\n", __func__, special_bos_id,     id_to_token.at(special_bos_id).text.c_str() );  }
+    if (special_eos_id  != LLAMA_TOKEN_NULL)    { LLAMA_LOG_INFO( "%s: EOS token        = %d '%s'\n", __func__, special_eos_id,     id_to_token.at(special_eos_id).text.c_str() );  }
+    if (special_eot_id  != LLAMA_TOKEN_NULL)    { LLAMA_LOG_INFO( "%s: EOT token        = %d '%s'\n", __func__, special_eot_id,     id_to_token.at(special_eot_id).text.c_str() );  }
+    if (special_eom_id  != LLAMA_TOKEN_NULL)    { LLAMA_LOG_INFO( "%s: EOM token        = %d '%s'\n", __func__, special_eom_id,     id_to_token.at(special_eom_id).text.c_str() );  }
+    if (special_unk_id  != LLAMA_TOKEN_NULL)    { LLAMA_LOG_INFO( "%s: UNK token        = %d '%s'\n", __func__, special_unk_id,     id_to_token.at(special_unk_id).text.c_str() );  }
+    if (special_sep_id  != LLAMA_TOKEN_NULL)    { LLAMA_LOG_INFO( "%s: SEP token        = %d '%s'\n", __func__, special_sep_id,     id_to_token.at(special_sep_id).text.c_str() );  }
+    if (special_pad_id  != LLAMA_TOKEN_NULL)    { LLAMA_LOG_INFO( "%s: PAD token        = %d '%s'\n", __func__, special_pad_id,     id_to_token.at(special_pad_id).text.c_str() );  }
+    if (special_mask_id != LLAMA_TOKEN_NULL)    { LLAMA_LOG_INFO( "%s: MASK token       = %d '%s'\n", __func__, special_mask_id,    id_to_token.at(special_mask_id).text.c_str() ); }
 
-    if (linefeed_id != LLAMA_TOKEN_NULL)        { LLAMA_LOG_INFO( "%s: LF token         = %d '%s'\n", __func__, linefeed_id,        id_to_token[linefeed_id].text.c_str() ); }
+    if (linefeed_id != LLAMA_TOKEN_NULL)        { LLAMA_LOG_INFO( "%s: LF token         = %d '%s'\n", __func__, linefeed_id,        id_to_token.at(linefeed_id).text.c_str() ); }
 
-    if (special_fim_pre_id != LLAMA_TOKEN_NULL) { LLAMA_LOG_INFO( "%s: FIM PRE token    = %d '%s'\n", __func__, special_fim_pre_id, id_to_token[special_fim_pre_id].text.c_str() ); }
-    if (special_fim_suf_id != LLAMA_TOKEN_NULL) { LLAMA_LOG_INFO( "%s: FIM SUF token    = %d '%s'\n", __func__, special_fim_suf_id, id_to_token[special_fim_suf_id].text.c_str() ); }
-    if (special_fim_mid_id != LLAMA_TOKEN_NULL) { LLAMA_LOG_INFO( "%s: FIM MID token    = %d '%s'\n", __func__, special_fim_mid_id, id_to_token[special_fim_mid_id].text.c_str() ); }
-    if (special_fim_pad_id != LLAMA_TOKEN_NULL) { LLAMA_LOG_INFO( "%s: FIM PAD token    = %d '%s'\n", __func__, special_fim_pad_id, id_to_token[special_fim_pad_id].text.c_str() ); }
-    if (special_fim_rep_id != LLAMA_TOKEN_NULL) { LLAMA_LOG_INFO( "%s: FIM REP token    = %d '%s'\n", __func__, special_fim_rep_id, id_to_token[special_fim_rep_id].text.c_str() ); }
-    if (special_fim_sep_id != LLAMA_TOKEN_NULL) { LLAMA_LOG_INFO( "%s: FIM SEP token    = %d '%s'\n", __func__, special_fim_sep_id, id_to_token[special_fim_sep_id].text.c_str() ); }
+    if (special_fim_pre_id != LLAMA_TOKEN_NULL) { LLAMA_LOG_INFO( "%s: FIM PRE token    = %d '%s'\n", __func__, special_fim_pre_id, id_to_token.at(special_fim_pre_id).text.c_str() ); }
+    if (special_fim_suf_id != LLAMA_TOKEN_NULL) { LLAMA_LOG_INFO( "%s: FIM SUF token    = %d '%s'\n", __func__, special_fim_suf_id, id_to_token.at(special_fim_suf_id).text.c_str() ); }
+    if (special_fim_mid_id != LLAMA_TOKEN_NULL) { LLAMA_LOG_INFO( "%s: FIM MID token    = %d '%s'\n", __func__, special_fim_mid_id, id_to_token.at(special_fim_mid_id).text.c_str() ); }
+    if (special_fim_pad_id != LLAMA_TOKEN_NULL) { LLAMA_LOG_INFO( "%s: FIM PAD token    = %d '%s'\n", __func__, special_fim_pad_id, id_to_token.at(special_fim_pad_id).text.c_str() ); }
+    if (special_fim_rep_id != LLAMA_TOKEN_NULL) { LLAMA_LOG_INFO( "%s: FIM REP token    = %d '%s'\n", __func__, special_fim_rep_id, id_to_token.at(special_fim_rep_id).text.c_str() ); }
+    if (special_fim_sep_id != LLAMA_TOKEN_NULL) { LLAMA_LOG_INFO( "%s: FIM SEP token    = %d '%s'\n", __func__, special_fim_sep_id, id_to_token.at(special_fim_sep_id).text.c_str() ); }
 
     for (const auto & id : special_eog_ids) {
-        LLAMA_LOG_INFO( "%s: EOG token        = %d '%s'\n", __func__, id, id_to_token[id].text.c_str() );
+        LLAMA_LOG_INFO( "%s: EOG token        = %d '%s'\n", __func__, id, id_to_token.at(id).text.c_str() );
     }
 
     LLAMA_LOG_INFO("%s: max token length = %d\n", __func__, max_token_len);
@@ -2991,6 +3010,10 @@ bool llama_vocab::get_add_eos() const {
     return pimpl->add_eos;
 }
 
+bool llama_vocab::get_add_sep() const {
+    return pimpl->add_sep;
+}
+
 bool llama_vocab::get_ignore_merges() const {
     return pimpl->ignore_merges;
 }
@@ -3051,6 +3074,11 @@ int32_t llama_vocab::tokenize(
                         bool   add_special,
                         bool   parse_special) const {
     auto res = tokenize(std::string(text, text_len), add_special, parse_special);
+    if (res.size() >= static_cast<size_t>(std::numeric_limits<int32_t>::max())) {
+        LLAMA_LOG_ERROR("%s: tokenization result size %zu exceeds int32_t limit\n", __func__, res.size());
+        return std::numeric_limits<int32_t>::min();
+    }
+
     if (n_tokens_max < (int) res.size()) {
         // LLAMA_LOG_ERROR("%s: too many tokens\n", __func__);
         return -((int) res.size());
@@ -3182,6 +3210,10 @@ bool llama_vocab_get_add_eos(const struct llama_vocab * vocab) {
     return vocab->get_add_eos();
 }
 
+bool llama_vocab_get_add_sep(const struct llama_vocab * vocab) {
+    return vocab->get_add_sep();
+}
+
 llama_token llama_vocab_fim_pre(const struct llama_vocab * vocab) {
     return vocab->token_fim_pre();
 }
diff --git a/src/llama-vocab.h b/src/llama-vocab.h
index daa6cf308..40e4d1c05 100644
--- a/src/llama-vocab.h
+++ b/src/llama-vocab.h
@@ -74,6 +74,7 @@ struct llama_vocab {
     bool get_add_space_prefix          () const;
     bool get_add_bos                   () const;
     bool get_add_eos                   () const;
+    bool get_add_sep                   () const;
     bool get_ignore_merges             () const;
     bool get_clean_spaces              () const;
     bool get_remove_extra_whitespaces  () const;
diff --git a/src/llama.cpp b/src/llama.cpp
index 2f06e0f8c..34906cdb6 100644
--- a/src/llama.cpp
+++ b/src/llama.cpp
@@ -198,14 +198,18 @@ static struct llama_model * llama_model_load_from_file_impl(
 
     // if using single GPU mode, remove all except the main GPU
     if (params.split_mode == LLAMA_SPLIT_MODE_NONE) {
-        if (params.main_gpu < 0 || params.main_gpu >= (int)model->devices.size()) {
-            LLAMA_LOG_ERROR("%s: invalid value for main_gpu: %d (available devices: %d)\n", __func__, params.main_gpu, (int)model->devices.size());
-            llama_model_free(model);
-            return nullptr;
+        if (params.main_gpu < 0) {
+            model->devices.clear();
+        } else {
+            if (params.main_gpu >= (int)model->devices.size()) {
+                LLAMA_LOG_ERROR("%s: invalid value for main_gpu: %d (available devices: %zu)\n", __func__, params.main_gpu, model->devices.size());
+                llama_model_free(model);
+                return nullptr;
+            }
+            ggml_backend_dev_t main_gpu = model->devices[params.main_gpu];
+            model->devices.clear();
+            model->devices.push_back(main_gpu);
         }
-        ggml_backend_dev_t main_gpu = model->devices[params.main_gpu];
-        model->devices.clear();
-        model->devices.push_back(main_gpu);
     }
 
     for (auto * dev : model->devices) {
diff --git a/src/unicode.cpp b/src/unicode.cpp
index e63bb4ab0..43a4581b9 100644
--- a/src/unicode.cpp
+++ b/src/unicode.cpp
@@ -204,12 +204,17 @@ static inline std::wstring unicode_wstring_from_utf8(const std::string & s) {
     // disable C++17 deprecation warning for std::codecvt_utf8
 #    pragma clang diagnostic push
 #    pragma clang diagnostic ignored "-Wdeprecated-declarations"
+#elif defined(__GNUC__)
+#    pragma GCC diagnostic push
+#    pragma GCC diagnostic ignored "-Wdeprecated-declarations"
 #endif
 
     std::wstring_convert<std::codecvt_utf8<wchar_t>> conv;
 
 #if defined(__clang__)
 #    pragma clang diagnostic pop
+#elif defined(__GNUC__)
+#    pragma GCC diagnostic pop
 #endif
 
     return conv.from_bytes(s);
diff --git a/tests/CMakeLists.txt b/tests/CMakeLists.txt
index 83f7d1a45..fc1557a2d 100644
--- a/tests/CMakeLists.txt
+++ b/tests/CMakeLists.txt
@@ -42,6 +42,34 @@ function(llama_test target)
     set_property(TEST ${TEST_NAME} PROPERTY LABELS ${LLAMA_TEST_LABEL})
 endfunction()
 
+function(llama_test_cmd target)
+    include(CMakeParseArguments)
+    set(options)
+    set(oneValueArgs NAME LABEL WORKING_DIRECTORY)
+    set(multiValueArgs ARGS)
+    cmake_parse_arguments(LLAMA_TEST "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
+
+    if (NOT DEFINED LLAMA_TEST_LABEL)
+        set(LLAMA_TEST_LABEL "main")
+    endif()
+    if (NOT DEFINED LLAMA_TEST_WORKING_DIRECTORY)
+        set(LLAMA_TEST_WORKING_DIRECTORY .)
+    endif()
+    if (DEFINED LLAMA_TEST_NAME)
+        set(TEST_NAME ${LLAMA_TEST_NAME})
+    else()
+        set(TEST_NAME ${target})
+    endif()
+
+    add_test(
+        NAME ${TEST_NAME}
+        WORKING_DIRECTORY ${LLAMA_TEST_WORKING_DIRECTORY}
+        COMMAND ${target}
+        ${LLAMA_TEST_ARGS})
+
+    set_property(TEST ${TEST_NAME} PROPERTY LABELS ${LLAMA_TEST_LABEL})
+endfunction()
+
 # Builds and runs a test source file.
 # Optional args:
 # - NAME: name of the executable & test target (defaults to the source file name without extension)
@@ -83,29 +111,35 @@ endfunction()
 # build test-tokenizer-0 target once and add many tests
 llama_build(test-tokenizer-0.cpp)
 
-llama_test(test-tokenizer-0 NAME test-tokenizer-0-bert-bge          ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-bert-bge.gguf)
-llama_test(test-tokenizer-0 NAME test-tokenizer-0-command-r         ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-command-r.gguf)
-llama_test(test-tokenizer-0 NAME test-tokenizer-0-deepseek-coder    ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-deepseek-coder.gguf)
-llama_test(test-tokenizer-0 NAME test-tokenizer-0-deepseek-llm      ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-deepseek-llm.gguf)
-llama_test(test-tokenizer-0 NAME test-tokenizer-0-falcon            ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-falcon.gguf)
-llama_test(test-tokenizer-0 NAME test-tokenizer-0-gpt-2             ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-gpt-2.gguf)
-llama_test(test-tokenizer-0 NAME test-tokenizer-0-llama-bpe         ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-llama-bpe.gguf)
-llama_test(test-tokenizer-0 NAME test-tokenizer-0-llama-spm         ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-llama-spm.gguf)
-llama_test(test-tokenizer-0 NAME test-tokenizer-0-mpt               ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-mpt.gguf)
-llama_test(test-tokenizer-0 NAME test-tokenizer-0-phi-3             ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-phi-3.gguf)
-llama_test(test-tokenizer-0 NAME test-tokenizer-0-qwen2             ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-qwen2.gguf)
-llama_test(test-tokenizer-0 NAME test-tokenizer-0-refact            ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-refact.gguf)
-llama_test(test-tokenizer-0 NAME test-tokenizer-0-starcoder         ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-starcoder.gguf)
-
-# TODO: missing HF tokenizer for this model in convert_hf_to_gguf_update.py, see https://github.com/ggml-org/llama.cpp/pull/13847
-# llama_test(test-tokenizer-0 NAME test-tokenizer-0-nomic-bert-moe    ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-nomic-bert-moe.gguf)
-
-if (LLAMA_LLGUIDANCE)
-    llama_build_and_test(test-grammar-llguidance.cpp ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-llama-bpe.gguf)
-endif ()
+llama_test(test-tokenizer-0 NAME test-tokenizer-0-bert-bge          ARGS ${PROJECT_SOURCE_DIR}/models/ggml-vocab-bert-bge.gguf)
+llama_test(test-tokenizer-0 NAME test-tokenizer-0-command-r         ARGS ${PROJECT_SOURCE_DIR}/models/ggml-vocab-command-r.gguf)
+llama_test(test-tokenizer-0 NAME test-tokenizer-0-deepseek-coder    ARGS ${PROJECT_SOURCE_DIR}/models/ggml-vocab-deepseek-coder.gguf)
+llama_test(test-tokenizer-0 NAME test-tokenizer-0-deepseek-llm      ARGS ${PROJECT_SOURCE_DIR}/models/ggml-vocab-deepseek-llm.gguf)
+llama_test(test-tokenizer-0 NAME test-tokenizer-0-falcon            ARGS ${PROJECT_SOURCE_DIR}/models/ggml-vocab-falcon.gguf)
+llama_test(test-tokenizer-0 NAME test-tokenizer-0-gpt-2             ARGS ${PROJECT_SOURCE_DIR}/models/ggml-vocab-gpt-2.gguf)
+llama_test(test-tokenizer-0 NAME test-tokenizer-0-llama-bpe         ARGS ${PROJECT_SOURCE_DIR}/models/ggml-vocab-llama-bpe.gguf)
+llama_test(test-tokenizer-0 NAME test-tokenizer-0-llama-spm         ARGS ${PROJECT_SOURCE_DIR}/models/ggml-vocab-llama-spm.gguf)
+llama_test(test-tokenizer-0 NAME test-tokenizer-0-mpt               ARGS ${PROJECT_SOURCE_DIR}/models/ggml-vocab-mpt.gguf)
+llama_test(test-tokenizer-0 NAME test-tokenizer-0-phi-3             ARGS ${PROJECT_SOURCE_DIR}/models/ggml-vocab-phi-3.gguf)
+llama_test(test-tokenizer-0 NAME test-tokenizer-0-qwen2             ARGS ${PROJECT_SOURCE_DIR}/models/ggml-vocab-qwen2.gguf)
+llama_test(test-tokenizer-0 NAME test-tokenizer-0-refact            ARGS ${PROJECT_SOURCE_DIR}/models/ggml-vocab-refact.gguf)
+llama_test(test-tokenizer-0 NAME test-tokenizer-0-starcoder         ARGS ${PROJECT_SOURCE_DIR}/models/ggml-vocab-starcoder.gguf)
 
 if (NOT WIN32)
-    # these tests are disabled on Windows because they use internal functions not exported with LLAMA_API
+    llama_test_cmd(
+        ${CMAKE_CURRENT_SOURCE_DIR}/test-tokenizers-repo.sh
+        NAME test-tokenizers-ggml-vocabs
+        WORKING_DIRECTORY ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}
+        ARGS https://huggingface.co/ggml-org/vocabs ${PROJECT_SOURCE_DIR}/models/ggml-vocabs
+    )
+endif()
+
+if (LLAMA_LLGUIDANCE)
+    llama_build_and_test(test-grammar-llguidance.cpp ARGS ${PROJECT_SOURCE_DIR}/models/ggml-vocab-llama-bpe.gguf)
+endif ()
+
+if (NOT WIN32 OR NOT BUILD_SHARED_LIBS)
+    # these tests are disabled on Windows because they use internal functions not exported with LLAMA_API (when building with shared libraries)
     llama_build_and_test(test-sampling.cpp)
     llama_build_and_test(test-grammar-parser.cpp)
     llama_build_and_test(test-grammar-integration.cpp)
@@ -113,8 +147,8 @@ if (NOT WIN32)
     llama_build_and_test(test-chat.cpp)
     # TODO: disabled on loongarch64 because the ggml-ci node lacks Python 3.8
     if (NOT ${CMAKE_SYSTEM_PROCESSOR} MATCHES "loongarch64")
-        llama_build_and_test(test-json-schema-to-grammar.cpp   WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}/..)
-        target_include_directories(test-json-schema-to-grammar PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/../tools/server)
+        llama_build_and_test(test-json-schema-to-grammar.cpp   WORKING_DIRECTORY ${PROJECT_SOURCE_DIR})
+        target_include_directories(test-json-schema-to-grammar PRIVATE ${PROJECT_SOURCE_DIR}/tools/server)
     endif()
 
     if (NOT GGML_BACKEND_DL)
@@ -127,20 +161,20 @@ if (NOT WIN32)
     llama_build(test-tokenizer-1-bpe.cpp)
 
     # TODO: disabled due to slowness
-    #llama_test(test-tokenizer-1-bpe NAME test-tokenizer-1-aquila    ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-aquila.gguf)
-    #llama_test(test-tokenizer-1-bpe NAME test-tokenizer-1-falcon    ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-falcon.gguf)
-    #llama_test(test-tokenizer-1-bpe NAME test-tokenizer-1-gpt-2     ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-gpt-2.gguf)
-    #llama_test(test-tokenizer-1-bpe NAME test-tokenizer-1-gpt-neox  ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-gpt-neox.gguf)
-    #llama_test(test-tokenizer-1-bpe NAME test-tokenizer-1-llama-bpe ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-llama-bpe.gguf --ignore-merges)
-    #llama_test(test-tokenizer-1-bpe NAME test-tokenizer-1-mpt       ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-mpt.gguf)
-    #llama_test(test-tokenizer-1-bpe NAME test-tokenizer-1-refact    ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-refact.gguf)
-    #llama_test(test-tokenizer-1-bpe NAME test-tokenizer-1-starcoder ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-starcoder.gguf)
+    #llama_test(test-tokenizer-1-bpe NAME test-tokenizer-1-aquila    ARGS ${PROJECT_SOURCE_DIR}/models/ggml-vocab-aquila.gguf)
+    #llama_test(test-tokenizer-1-bpe NAME test-tokenizer-1-falcon    ARGS ${PROJECT_SOURCE_DIR}/models/ggml-vocab-falcon.gguf)
+    #llama_test(test-tokenizer-1-bpe NAME test-tokenizer-1-gpt-2     ARGS ${PROJECT_SOURCE_DIR}/models/ggml-vocab-gpt-2.gguf)
+    #llama_test(test-tokenizer-1-bpe NAME test-tokenizer-1-gpt-neox  ARGS ${PROJECT_SOURCE_DIR}/models/ggml-vocab-gpt-neox.gguf)
+    #llama_test(test-tokenizer-1-bpe NAME test-tokenizer-1-llama-bpe ARGS ${PROJECT_SOURCE_DIR}/models/ggml-vocab-llama-bpe.gguf --ignore-merges)
+    #llama_test(test-tokenizer-1-bpe NAME test-tokenizer-1-mpt       ARGS ${PROJECT_SOURCE_DIR}/models/ggml-vocab-mpt.gguf)
+    #llama_test(test-tokenizer-1-bpe NAME test-tokenizer-1-refact    ARGS ${PROJECT_SOURCE_DIR}/models/ggml-vocab-refact.gguf)
+    #llama_test(test-tokenizer-1-bpe NAME test-tokenizer-1-starcoder ARGS ${PROJECT_SOURCE_DIR}/models/ggml-vocab-starcoder.gguf)
 
     # build test-tokenizer-1-spm target once and add many tests
     llama_build(test-tokenizer-1-spm.cpp)
 
-    llama_test(test-tokenizer-1-spm  NAME test-tokenizer-1-llama-spm ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-llama-spm.gguf)
-    #llama_test(test-tokenizer-1-spm  NAME test-tokenizer-1-baichuan  ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-baichuan.gguf)
+    llama_test(test-tokenizer-1-spm  NAME test-tokenizer-1-llama-spm ARGS ${PROJECT_SOURCE_DIR}/models/ggml-vocab-llama-spm.gguf)
+    #llama_test(test-tokenizer-1-spm  NAME test-tokenizer-1-baichuan  ARGS ${PROJECT_SOURCE_DIR}/models/ggml-vocab-baichuan.gguf)
 
     # llama_build_and_test(test-double-float.cpp) # SLOW
 endif()
@@ -151,6 +185,8 @@ llama_build_and_test(test-json-partial.cpp)
 llama_build_and_test(test-log.cpp)
 llama_build_and_test(test-regex-partial.cpp)
 
+llama_build_and_test(test-thread-safety.cpp ARGS -hf ggml-org/models -hff tinyllamas/stories15M-q4_0.gguf -ngl 99 -p "The meaning of life is" -n 128 -c 256 -ub 32 -np 4)
+
 # this fails on windows (github hosted runner) due to curl DLL not found (exit code 0xc0000135)
 if (NOT WIN32)
     llama_build_and_test(test-arg-parser.cpp)
diff --git a/tests/test-backend-ops.cpp b/tests/test-backend-ops.cpp
index 543db9340..652856a35 100644
--- a/tests/test-backend-ops.cpp
+++ b/tests/test-backend-ops.cpp
@@ -24,10 +24,12 @@
 #include <array>
 #include <cfloat>
 #include <cinttypes>
+#include <cstdarg>
 #include <cstdint>
 #include <cstdio>
 #include <cstdlib>
 #include <cstring>
+#include <ctime>
 #include <future>
 #include <memory>
 #include <random>
@@ -317,6 +319,538 @@ enum test_mode {
     MODE_GRAD,
 };
 
+// Output format support similar to llama-bench
+enum output_formats { CONSOLE, SQL };
+
+static const char * output_format_str(output_formats format) {
+    switch (format) {
+        case CONSOLE:
+            return "console";
+        case SQL:
+            return "sql";
+        default:
+            GGML_ABORT("invalid output format");
+    }
+}
+
+static bool output_format_from_str(const std::string & s, output_formats & format) {
+    if (s == "console") {
+        format = CONSOLE;
+    } else if (s == "sql") {
+        format = SQL;
+    } else {
+        return false;
+    }
+    return true;
+}
+
+// Test result structure for SQL output
+struct test_result {
+    std::string test_time;
+    std::string build_commit;
+    std::string backend_name;
+    std::string op_name;
+    std::string op_params;
+    std::string test_mode;
+    bool        supported;
+    bool        passed;
+    std::string error_message;
+    double      time_us;
+    double      flops;
+    double      bandwidth_gb_s;
+    size_t      memory_kb;
+    int         n_runs;
+
+    test_result() {
+        // Initialize with default values
+        time_us        = 0.0;
+        flops          = 0.0;
+        bandwidth_gb_s = 0.0;
+        memory_kb      = 0;
+        n_runs         = 0;
+        supported      = false;
+        passed         = false;
+
+        // Set test time
+        time_t t = time(NULL);
+        char   buf[32];
+        std::strftime(buf, sizeof(buf), "%FT%TZ", gmtime(&t));
+        test_time = buf;
+
+        // Set build info
+        build_commit = ggml_commit();
+    }
+
+    test_result(const std::string & backend_name, const std::string & op_name, const std::string & op_params,
+                const std::string & test_mode, bool supported, bool passed, const std::string & error_message = "",
+                double time_us = 0.0, double flops = 0.0, double bandwidth_gb_s = 0.0, size_t memory_kb = 0,
+                int n_runs = 0) :
+        backend_name(backend_name),
+        op_name(op_name),
+        op_params(op_params),
+        test_mode(test_mode),
+        supported(supported),
+        passed(passed),
+        error_message(error_message),
+        time_us(time_us),
+        flops(flops),
+        bandwidth_gb_s(bandwidth_gb_s),
+        memory_kb(memory_kb),
+        n_runs(n_runs) {
+        // Set test time
+        time_t t = time(NULL);
+        char   buf[32];
+        std::strftime(buf, sizeof(buf), "%FT%TZ", gmtime(&t));
+        test_time = buf;
+
+        // Set build info
+        build_commit = ggml_commit();
+    }
+
+    static const std::vector<std::string> & get_fields() {
+        static const std::vector<std::string> fields = {
+            "test_time", "build_commit",  "backend_name", "op_name", "op_params",      "test_mode", "supported",
+            "passed",    "error_message", "time_us",      "flops",   "bandwidth_gb_s", "memory_kb", "n_runs"
+        };
+        return fields;
+    }
+
+    enum field_type { STRING, BOOL, INT, FLOAT };
+
+    static field_type get_field_type(const std::string & field) {
+        if (field == "supported" || field == "passed") {
+            return BOOL;
+        }
+        if (field == "memory_kb" || field == "n_runs") {
+            return INT;
+        }
+        if (field == "time_us" || field == "flops" || field == "bandwidth_gb_s") {
+            return FLOAT;
+        }
+        return STRING;
+    }
+
+    std::vector<std::string> get_values() const {
+        return { test_time,
+                 build_commit,
+                 backend_name,
+                 op_name,
+                 op_params,
+                 test_mode,
+                 std::to_string(supported),
+                 std::to_string(passed),
+                 error_message,
+                 std::to_string(time_us),
+                 std::to_string(flops),
+                 std::to_string(bandwidth_gb_s),
+                 std::to_string(memory_kb),
+                 std::to_string(n_runs) };
+    }
+};
+
+// Printer classes for different output formats
+enum class test_status_t { NOT_SUPPORTED, OK, FAIL };
+
+struct test_operation_info {
+    std::string   op_name;
+    std::string   op_params;
+    std::string   backend_name;
+    test_status_t status = test_status_t::OK;
+    std::string   failure_reason;
+
+    // Additional information fields that were previously in separate structs
+    std::string error_component;
+    std::string error_details;
+
+    // Gradient info
+    int64_t     gradient_index = -1;
+    std::string gradient_param_name;
+    float       gradient_value = 0.0f;
+
+    // MAA error info
+    double maa_error     = 0.0;
+    double maa_threshold = 0.0;
+
+    // Flags for different types of information
+    bool has_error            = false;
+    bool has_gradient_info    = false;
+    bool has_maa_error        = false;
+    bool is_compare_failure   = false;
+    bool is_large_tensor_skip = false;
+
+    test_operation_info() = default;
+
+    test_operation_info(const std::string & op_name, const std::string & op_params, const std::string & backend_name,
+                        test_status_t status = test_status_t::OK, const std::string & failure_reason = "") :
+        op_name(op_name),
+        op_params(op_params),
+        backend_name(backend_name),
+        status(status),
+        failure_reason(failure_reason) {}
+
+    // Set error information
+    void set_error(const std::string & component, const std::string & details) {
+        has_error       = true;
+        error_component = component;
+        error_details   = details;
+        if (status == test_status_t::OK) {
+            status = test_status_t::FAIL;
+        }
+    }
+
+    // Set gradient information
+    void set_gradient_info(int64_t index, const std::string & param_name, float value) {
+        has_gradient_info   = true;
+        gradient_index      = index;
+        gradient_param_name = param_name;
+        gradient_value      = value;
+        if (status == test_status_t::OK) {
+            status = test_status_t::FAIL;
+        }
+    }
+
+    // Set MAA error information
+    void set_maa_error(double error, double threshold) {
+        has_maa_error = true;
+        maa_error     = error;
+        maa_threshold = threshold;
+        if (status == test_status_t::OK) {
+            status = test_status_t::FAIL;
+        }
+    }
+
+    // Set compare failure
+    void set_compare_failure() {
+        is_compare_failure = true;
+        if (status == test_status_t::OK) {
+            status = test_status_t::FAIL;
+        }
+    }
+
+    // Set large tensor skip
+    void set_large_tensor_skip() { is_large_tensor_skip = true; }
+};
+
+struct test_summary_info {
+    size_t tests_passed;
+    size_t tests_total;
+    bool   is_backend_summary = false;  // true for backend summary, false for test summary
+
+    test_summary_info() = default;
+
+    test_summary_info(size_t tests_passed, size_t tests_total, bool is_backend_summary = false) :
+        tests_passed(tests_passed),
+        tests_total(tests_total),
+        is_backend_summary(is_backend_summary) {}
+};
+
+struct testing_start_info {
+    size_t device_count;
+
+    testing_start_info() = default;
+
+    testing_start_info(size_t device_count) : device_count(device_count) {}
+};
+
+struct backend_init_info {
+    size_t      device_index;
+    size_t      total_devices;
+    std::string device_name;
+    bool        skipped = false;
+    std::string skip_reason;
+    std::string description;
+    size_t      memory_total_mb = 0;
+    size_t      memory_free_mb  = 0;
+    bool        has_memory_info = false;
+
+    backend_init_info() = default;
+
+    backend_init_info(size_t device_index, size_t total_devices, const std::string & device_name, bool skipped = false,
+                      const std::string & skip_reason = "", const std::string & description = "",
+                      size_t memory_total_mb = 0, size_t memory_free_mb = 0, bool has_memory_info = false) :
+        device_index(device_index),
+        total_devices(total_devices),
+        device_name(device_name),
+        skipped(skipped),
+        skip_reason(skip_reason),
+        description(description),
+        memory_total_mb(memory_total_mb),
+        memory_free_mb(memory_free_mb),
+        has_memory_info(has_memory_info) {}
+};
+
+struct backend_status_info {
+    std::string   backend_name;
+    test_status_t status;
+
+    backend_status_info() = default;
+
+    backend_status_info(const std::string & backend_name, test_status_t status) :
+        backend_name(backend_name),
+        status(status) {}
+};
+
+struct overall_summary_info {
+    size_t backends_passed;
+    size_t backends_total;
+    bool   all_passed;
+
+    overall_summary_info() = default;
+
+    overall_summary_info(size_t backends_passed, size_t backends_total, bool all_passed) :
+        backends_passed(backends_passed),
+        backends_total(backends_total),
+        all_passed(all_passed) {}
+};
+
+struct printer {
+    virtual ~printer() {}
+
+    FILE * fout = stdout;
+
+    virtual void print_header() {}
+
+    virtual void print_test_result(const test_result & result) = 0;
+
+    virtual void print_footer() {}
+
+    virtual void print_operation(const test_operation_info & info) { (void) info; }
+
+    virtual void print_summary(const test_summary_info & info) { (void) info; }
+
+    virtual void print_testing_start(const testing_start_info & info) { (void) info; }
+
+    virtual void print_backend_init(const backend_init_info & info) { (void) info; }
+
+    virtual void print_backend_status(const backend_status_info & info) { (void) info; }
+
+    virtual void print_overall_summary(const overall_summary_info & info) { (void) info; }
+};
+
+struct console_printer : public printer {
+    void print_test_result(const test_result & result) override {
+        if (result.test_mode == "test") {
+            print_test_console(result);
+        } else if (result.test_mode == "perf") {
+            print_perf_console(result);
+        }
+    }
+
+    void print_operation(const test_operation_info & info) override {
+        printf("  %s(%s): ", info.op_name.c_str(), info.op_params.c_str());
+        fflush(stdout);
+
+        // Handle large tensor skip first
+        if (info.is_large_tensor_skip) {
+            printf("skipping large tensors for speed \n");
+            return;
+        }
+
+        // Handle not supported status
+        if (info.status == test_status_t::NOT_SUPPORTED) {
+            if (!info.failure_reason.empty()) {
+                printf("not supported [%s]\n", info.failure_reason.c_str());
+            } else {
+                printf("not supported [%s]\n", info.backend_name.c_str());
+            }
+            return;
+        }
+
+        // Handle errors and additional information
+        if (info.has_error) {
+            if (info.error_component == "allocation") {
+                fprintf(stderr, "failed to allocate tensors [%s] ", info.backend_name.c_str());
+            } else if (info.error_component == "backend") {
+                fprintf(stderr, "  Failed to initialize %s backend\n", info.backend_name.c_str());
+            } else {
+                fprintf(stderr, "Error in %s: %s\n", info.error_component.c_str(), info.error_details.c_str());
+            }
+        }
+
+        // Handle gradient info
+        if (info.has_gradient_info) {
+            printf("[%s] nonfinite gradient at index %" PRId64 " (%s=%f) ", info.op_name.c_str(), info.gradient_index,
+                   info.gradient_param_name.c_str(), info.gradient_value);
+        }
+
+        // Handle MAA error
+        if (info.has_maa_error) {
+            printf("[%s] MAA = %.9f > %.9f ", info.op_name.c_str(), info.maa_error, info.maa_threshold);
+        }
+
+        // Handle compare failure
+        if (info.is_compare_failure) {
+            printf("compare failed ");
+        }
+
+        // Print final status
+        if (info.status == test_status_t::OK) {
+            printf("\033[1;32mOK\033[0m\n");
+        } else {
+            printf("\033[1;31mFAIL\033[0m\n");
+        }
+    }
+
+    void print_summary(const test_summary_info & info) override {
+        if (info.is_backend_summary) {
+            printf("%zu/%zu backends passed\n", info.tests_passed, info.tests_total);
+        } else {
+            printf("  %zu/%zu tests passed\n", info.tests_passed, info.tests_total);
+        }
+    }
+
+    void print_backend_status(const backend_status_info & info) override {
+        printf("  Backend %s: ", info.backend_name.c_str());
+        if (info.status == test_status_t::OK) {
+            printf("\033[1;32mOK\033[0m\n");
+        } else {
+            printf("\033[1;31mFAIL\033[0m\n");
+        }
+    }
+
+    void print_testing_start(const testing_start_info & info) override {
+        printf("Testing %zu devices\n\n", info.device_count);
+    }
+
+    void print_backend_init(const backend_init_info & info) override {
+        printf("Backend %zu/%zu: %s\n", info.device_index + 1, info.total_devices, info.device_name.c_str());
+
+        if (info.skipped) {
+            printf("  %s\n", info.skip_reason.c_str());
+            return;
+        }
+
+        if (!info.description.empty()) {
+            printf("  Device description: %s\n", info.description.c_str());
+        }
+
+        if (info.has_memory_info) {
+            printf("  Device memory: %zu MB (%zu MB free)\n", info.memory_total_mb, info.memory_free_mb);
+        }
+
+        printf("\n");
+    }
+
+    void print_overall_summary(const overall_summary_info & info) override {
+        printf("%zu/%zu backends passed\n", info.backends_passed, info.backends_total);
+        if (info.all_passed) {
+            printf("\033[1;32mOK\033[0m\n");
+        } else {
+            printf("\033[1;31mFAIL\033[0m\n");
+        }
+    }
+
+  private:
+    void print_test_console(const test_result & result) {
+        printf("  %s(%s): ", result.op_name.c_str(), result.op_params.c_str());
+        fflush(stdout);
+
+        if (!result.supported) {
+            printf("not supported [%s] ", result.backend_name.c_str());
+            printf("\n");
+            return;
+        }
+
+        if (result.passed) {
+            printf("\033[1;32mOK\033[0m\n");
+        } else {
+            printf("\033[1;31mFAIL\033[0m\n");
+        }
+    }
+
+    void print_perf_console(const test_result & result) {
+        int len = printf("  %s(%s): ", result.op_name.c_str(), result.op_params.c_str());
+        fflush(stdout);
+
+        if (!result.supported) {
+            printf("not supported\n");
+            return;
+        }
+
+        // align while also leaving some margin for variations in parameters
+        int align = 8;
+        int last  = (len + align - 1) / align * align;
+        if (last - len < 5) {
+            last += align;
+        }
+        printf("%*s", last - len, "");
+
+        printf("    %8d runs - %8.2f us/run - ", result.n_runs, result.time_us);
+
+        if (result.flops > 0) {
+            auto format_flops = [](double flops) -> std::string {
+                char buf[256];
+                if (flops >= 1e12) {
+                    snprintf(buf, sizeof(buf), "%6.2f TFLOP", flops / 1e12);
+                } else if (flops >= 1e9) {
+                    snprintf(buf, sizeof(buf), "%6.2f GFLOP", flops / 1e9);
+                } else if (flops >= 1e6) {
+                    snprintf(buf, sizeof(buf), "%6.2f MFLOP", flops / 1e6);
+                } else {
+                    snprintf(buf, sizeof(buf), "%6.2f kFLOP", flops / 1e3);
+                }
+                return buf;
+            };
+            uint64_t op_flops_per_run = result.flops * result.time_us / 1e6;
+            printf("%s/run - \033[1;34m%sS\033[0m", format_flops(op_flops_per_run).c_str(),
+                   format_flops(result.flops).c_str());
+        } else {
+            printf("%8zu kB/run - \033[1;34m%7.2f GB/s\033[0m", result.memory_kb, result.bandwidth_gb_s);
+        }
+        printf("\n");
+    }
+};
+
+struct sql_printer : public printer {
+    static std::string get_sql_field_type(const std::string & field) {
+        switch (test_result::get_field_type(field)) {
+            case test_result::STRING:
+                return "TEXT";
+            case test_result::BOOL:
+            case test_result::INT:
+                return "INTEGER";
+            case test_result::FLOAT:
+                return "REAL";
+            default:
+                GGML_ABORT("invalid field type");
+        }
+    }
+
+    void print_header() override {
+        std::vector<std::string> fields = test_result::get_fields();
+        fprintf(fout, "CREATE TABLE IF NOT EXISTS test_backend_ops (\n");
+        for (size_t i = 0; i < fields.size(); i++) {
+            fprintf(fout, "  %s %s%s\n", fields[i].c_str(), get_sql_field_type(fields[i]).c_str(),
+                    i < fields.size() - 1 ? "," : "");
+        }
+        fprintf(fout, ");\n\n");
+    }
+
+    void print_test_result(const test_result & result) override {
+        fprintf(fout, "INSERT INTO test_backend_ops (");
+        std::vector<std::string> fields = test_result::get_fields();
+        for (size_t i = 0; i < fields.size(); i++) {
+            fprintf(fout, "%s%s", fields[i].c_str(), i < fields.size() - 1 ? ", " : "");
+        }
+        fprintf(fout, ") VALUES (");
+        std::vector<std::string> values = result.get_values();
+        for (size_t i = 0; i < values.size(); i++) {
+            fprintf(fout, "'%s'%s", values[i].c_str(), i < values.size() - 1 ? ", " : "");
+        }
+        fprintf(fout, ");\n");
+    }
+};
+
+static std::unique_ptr<printer> create_printer(output_formats format) {
+    switch (format) {
+        case CONSOLE:
+            return std::make_unique<console_printer>();
+        case SQL:
+            return std::make_unique<sql_printer>();
+    }
+    GGML_ABORT("invalid output format");
+}
+
 struct test_case {
     virtual ~test_case() {}
 
@@ -382,6 +916,8 @@ struct test_case {
         return 0;
     }
 
+    virtual bool run_whole_graph() { return false; }
+
     ggml_cgraph * gf = nullptr;
     ggml_cgraph * gb = nullptr;
 
@@ -432,7 +968,7 @@ struct test_case {
         return t;
     }
 
-    bool eval(ggml_backend_t backend1, ggml_backend_t backend2, const char * op_name) {
+    bool eval(ggml_backend_t backend1, ggml_backend_t backend2, const char * op_name, printer * output_printer) {
         mode = MODE_TEST;
 
         ggml_init_params params = {
@@ -449,29 +985,33 @@ struct test_case {
         add_sentinel(ctx);
 
         ggml_tensor * out = build_graph(ctx);
-
-        if (op_name != nullptr && op_desc(out) != op_name) {
+        std::string current_op_name = op_desc(out);
+        if (op_name != nullptr && current_op_name != op_name) {
             //printf("  %s: skipping\n", op_desc(out).c_str());
             ggml_free(ctx);
             return true;
         }
 
-        printf("  %s(%s): ", op_desc(out).c_str(), vars().c_str());
-        fflush(stdout);
-
         // check if the backends support the ops
         bool supported = true;
         for (ggml_backend_t backend : {backend1, backend2}) {
             for (ggml_tensor * t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) {
                 if (!ggml_backend_supports_op(backend, t)) {
-                    printf("not supported [%s] ", ggml_backend_name(backend));
                     supported = false;
                     break;
                 }
             }
         }
+
         if (!supported) {
-            printf("\n");
+            // Create test result for unsupported operation
+            test_result result(ggml_backend_name(backend1), current_op_name, vars(), "test",
+                             false, false, "not supported");
+
+            if (output_printer) {
+                output_printer->print_test_result(result);
+            }
+
             ggml_free(ctx);
             return true;
         }
@@ -574,26 +1114,26 @@ struct test_case {
             GGML_UNUSED(index);
         };
 
-        const bool cmp_ok = ggml_backend_compare_graph_backend(backend1, backend2, gf, callback, &ud);
-
-        if (!cmp_ok) {
-            printf("compare failed ");
-        }
+        const bool cmp_ok = ggml_backend_compare_graph_backend(backend1, backend2, gf, callback, &ud, run_whole_graph() ? out : nullptr);
 
         ggml_backend_buffer_free(buf);
 
         ggml_free(ctx);
 
-        if (ud.ok && cmp_ok) {
-            printf("\033[1;32mOK\033[0m\n");
-            return true;
+        // Create test result
+        bool        test_passed = ud.ok && cmp_ok;
+        std::string error_msg   = test_passed ? "" : (!cmp_ok ? "compare failed" : "test failed");
+        test_result result(ggml_backend_name(backend1), current_op_name, vars(), "test", supported, test_passed,
+                           error_msg);
+
+        if (output_printer) {
+            output_printer->print_test_result(result);
         }
 
-        printf("\033[1;31mFAIL\033[0m\n");
-        return false;
+        return test_passed;
     }
 
-    bool eval_perf(ggml_backend_t backend, const char * op_name) {
+    bool eval_perf(ggml_backend_t backend, const char * op_name, printer * output_printer) {
         mode = MODE_PERF;
 
         static const size_t graph_nodes = 8192;
@@ -606,30 +1146,26 @@ struct test_case {
         ggml_context_ptr ctx(ggml_init(params)); // smart ptr
         GGML_ASSERT(ctx);
 
-        ggml_tensor * out = build_graph(ctx.get());
-
-        if (op_name != nullptr && op_desc(out) != op_name) {
+        ggml_tensor * out             = build_graph(ctx.get());
+        std::string   current_op_name = op_desc(out);
+        if (op_name != nullptr && current_op_name != op_name) {
             //printf("  %s: skipping\n", op_desc(out).c_str());
             return true;
         }
 
-        int len = printf("  %s(%s): ", op_desc(out).c_str(), vars().c_str());
-        fflush(stdout);
-
         // check if backends support op
         if (!ggml_backend_supports_op(backend, out)) {
-            printf("not supported\n");
+            // Create test result for unsupported performance test
+            test_result result(ggml_backend_name(backend), current_op_name, vars(), "perf", false, false,
+                               "not supported");
+
+            if (output_printer) {
+                output_printer->print_test_result(result);
+            }
+
             return true;
         }
 
-        // align while also leaving some margin for variations in parameters
-        int align = 8;
-        int last = (len + align - 1) / align * align;
-        if (last - len < 5) {
-            last += align;
-        }
-        printf("%*s", last - len, "");
-
         // allocate
         ggml_backend_buffer_ptr buf(ggml_backend_alloc_ctx_tensors(ctx.get(), backend)); // smart ptr
 
@@ -713,40 +1249,24 @@ struct test_case {
             total_runs += n_runs;
         } while (total_time_us < 1000*1000); // run for at least 1 second
 
-        printf("    %8d runs - %8.2f us/run - ",
-            total_runs,
-            (double)total_time_us / total_runs);
+        // Create test result
+        double avg_time_us      = (double) total_time_us / total_runs;
+        double calculated_flops = (op_flops(out) > 0) ? (op_flops(out) * total_runs) / (total_time_us / 1e6) : 0.0;
+        double calculated_bandwidth =
+            (op_flops(out) == 0) ? total_mem / (total_time_us / 1e6) / 1024.0 / 1024.0 / 1024.0 : 0.0;
+        size_t calculated_memory_kb = op_size(out) / 1024;
 
-        if (op_flops(out) > 0) {
-            double flops_per_sec = (op_flops(out) * total_runs) / (total_time_us / 1e6);
-            auto format_flops = [](double flops) -> std::string {
-                char buf[256];
-                if (flops >= 1e12) {
-                    snprintf(buf, sizeof(buf), "%6.2f TFLOP", flops / 1e12);
-                } else if (flops >= 1e9) {
-                    snprintf(buf, sizeof(buf), "%6.2f GFLOP", flops / 1e9);
-                } else if (flops >= 1e6) {
-                    snprintf(buf, sizeof(buf), "%6.2f MFLOP", flops / 1e6);
-                } else {
-                    snprintf(buf, sizeof(buf), "%6.2f KFLOP", flops / 1e3);
-                }
-                return buf;
-            };
-            printf("%s/run - \033[1;34m%sS\033[0m",
-                format_flops(op_flops(out)).c_str(),
-                format_flops(flops_per_sec).c_str());
+        test_result result(ggml_backend_name(backend), current_op_name, vars(), "perf", true, true, "", avg_time_us,
+                           calculated_flops, calculated_bandwidth, calculated_memory_kb, total_runs);
 
-        } else {
-            printf("%8zu kB/run - \033[1;34m%7.2f GB/s\033[0m",
-                op_size(out) / 1024,
-                total_mem / (total_time_us / 1e6) / 1024.0 / 1024.0 / 1024.0);
+        if (output_printer) {
+            output_printer->print_test_result(result);
         }
-        printf("\n");
 
         return true;
     }
 
-    bool eval_grad(ggml_backend_t backend, const char * op_name) {
+    bool eval_grad(ggml_backend_t backend, const char * op_name, printer * output_printer) {
         mode = MODE_GRAD;
         const std::vector<float> expect = grad_expect();
 
@@ -764,42 +1284,47 @@ struct test_case {
         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());
             return true;
         }
 
-        printf("  %s(%s): ", op_desc(out).c_str(), vars().c_str());
-        fflush(stdout);
-
         if (out->type != GGML_TYPE_F32) {
-            printf("not supported [%s->type != FP32]\n", out->name);
+            output_printer->print_operation(test_operation_info(op_desc(out), vars(), ggml_backend_name(backend),
+                                                                test_status_t::NOT_SUPPORTED,
+                                                                out->name + std::string("->type != FP32")));
             return true;
         }
 
+        // Print operation info first
+        output_printer->print_operation(test_operation_info(op_desc(out), vars(), ggml_backend_name(backend)));
+
         // check if the backend supports the ops
-        bool supported = true;
-        bool any_params = false;
+        bool        supported  = true;
+        bool        any_params = false;
+        std::string failure_reason;
+
         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;
+                supported      = false;
+                failure_reason = ggml_backend_name(backend);
                 break;
             }
             if ((t->flags & GGML_TENSOR_FLAG_PARAM)) {
                 any_params = true;
                 if (t->type != GGML_TYPE_F32) {
-                    printf("not supported [%s->type != FP32] ", t->name);
-                    supported = false;
+                    supported      = false;
+                    failure_reason = std::string(t->name) + "->type != FP32";
                     break;
                 }
             }
         }
         if (!any_params) {
-            printf("not supported [%s] \n", op_desc(out).c_str());
-            supported = false;
+            supported      = false;
+            failure_reason = op_desc(out);
         }
+
         if (!supported) {
-            printf("\n");
+            output_printer->print_operation(test_operation_info(op_desc(out), vars(), ggml_backend_name(backend),
+                                                                test_status_t::NOT_SUPPORTED, failure_reason));
             return true;
         }
 
@@ -810,7 +1335,9 @@ struct test_case {
             }
         }
         if (ngrads > grad_nmax()) {
-            printf("skipping large tensors for speed \n");
+            test_operation_info info(op_desc(out), vars(), ggml_backend_name(backend));
+            info.set_large_tensor_skip();
+            output_printer->print_operation(info);
             return true;
         }
 
@@ -833,25 +1360,30 @@ struct test_case {
 
         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));
+                output_printer->print_operation(test_operation_info(op_desc(out), vars(), ggml_backend_name(backend),
+                                                                    test_status_t::NOT_SUPPORTED,
+                                                                    ggml_backend_name(backend)));
                 supported = false;
                 break;
             }
             if ((t->flags & GGML_TENSOR_FLAG_PARAM) && t->type != GGML_TYPE_F32) {
-                printf("not supported [%s->type != FP32] ", t->name);
+                output_printer->print_operation(test_operation_info(op_desc(out), vars(), ggml_backend_name(backend),
+                                                                    test_status_t::NOT_SUPPORTED,
+                                                                    std::string(t->name) + "->type != FP32"));
                 supported = false;
                 break;
             }
         }
         if (!supported) {
-            printf("\n");
             return true;
         }
 
         // allocate
         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));
+            test_operation_info info(op_desc(out), vars(), ggml_backend_name(backend));
+            info.set_error("allocation", "");
+            output_printer->print_operation(info);
             return false;
         }
 
@@ -889,7 +1421,9 @@ struct test_case {
             for (int64_t i = 0; i < ne; ++i) { // gradient algebraic
                 // check for nans
                 if (!std::isfinite(ga[i])) {
-                    printf("[%s] nonfinite gradient at index %" PRId64 " (%s=%f) ", ggml_op_desc(t), i, bn, ga[i]);
+                    test_operation_info info(op_desc(out), vars(), ggml_backend_name(backend));
+                    info.set_gradient_info(i, bn, ga[i]);
+                    output_printer->print_operation(info);
                     ok = false;
                     break;
                 }
@@ -957,7 +1491,9 @@ struct test_case {
 
             const double err = mean_abs_asymm(gn.data(), ga.data(), gn.size(), expect);
             if (err > max_maa_err()) {
-                printf("[%s] MAA = %.9f > %.9f ", ggml_op_desc(t), err, max_maa_err());
+                test_operation_info info(op_desc(out), vars(), ggml_backend_name(backend));
+                info.set_maa_error(err, max_maa_err());
+                output_printer->print_operation(info);
                 ok = false;
                 break;
             }
@@ -966,16 +1502,18 @@ struct test_case {
             }
         }
 
+        // Create final test result
+        test_operation_info final_info(op_desc(out), vars(), ggml_backend_name(backend));
         if (!ok) {
-            printf("compare failed ");
+            final_info.set_compare_failure();
         }
+        final_info.status = ok ? test_status_t::OK : test_status_t::FAIL;
+        output_printer->print_operation(final_info);
 
         if (ok) {
-            printf("\033[1;32mOK\033[0m\n");
             return true;
         }
 
-        printf("\033[1;31mFAIL\033[0m\n");
         return false;
     }
 };
@@ -1104,6 +1642,111 @@ struct test_unary : public test_case {
 
 };
 
+// GGML_OP_GLU
+struct test_glu : public test_case {
+    const ggml_glu_op op;
+    const ggml_type type;
+    const std::array<int64_t, 4> ne_a;
+    int v; // view (1 : non-contiguous a)
+    bool swapped;
+
+    std::string vars() override {
+        return VARS_TO_STR4(type, ne_a, v, swapped);
+    }
+
+    test_glu(ggml_glu_op op,
+            ggml_type type = GGML_TYPE_F32,
+            std::array<int64_t, 4> ne_a = {128, 2, 2, 2},
+            int v = 0,
+            bool swapped = false)
+        : op(op), type(type), ne_a(ne_a), v(v), swapped(swapped) {}
+
+    ggml_tensor * build_graph(ggml_context * ctx) override {
+        ggml_tensor * a;
+        if (v & 1) {
+            auto ne = ne_a; ne[0] *= 3;
+            a = ggml_new_tensor(ctx, type, 4, ne.data());
+            ggml_set_name(a, "a");
+
+            a = ggml_view_4d(ctx, a, ne_a[0], ne_a[1], ne_a[2], ne_a[3], a->nb[1], a->nb[2], a->nb[3], 0);
+            ggml_set_name(a, "view_of_a");
+        } else {
+            a = ggml_new_tensor(ctx, type, 4, ne_a.data());
+            ggml_set_name(a, "a");
+        }
+
+        ggml_tensor * out = ggml_glu(ctx, a, op, swapped);
+        ggml_set_name(out, "out");
+
+        return out;
+    }
+
+    void initialize_tensors(ggml_context * ctx) override {
+        for (ggml_tensor * t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) {
+            // test extended range of values to check for NaNs in GELU
+            init_tensor_uniform(t, -150.f, 150.f);
+        }
+    }
+};
+
+struct test_glu_split : public test_case {
+    const ggml_glu_op op;
+    const ggml_type type;
+    const std::array<int64_t, 4> ne_a;
+    int v; // view (1 : non-contiguous a)
+
+    std::string vars() override {
+        return VARS_TO_STR3(type, ne_a, v) + ",split";
+    }
+
+    test_glu_split(ggml_glu_op op,
+            ggml_type type = GGML_TYPE_F32,
+            std::array<int64_t, 4> ne_a = {128, 2, 2, 2},
+            int v = 0)
+        : op(op), type(type), ne_a(ne_a), v(v) {}
+
+    ggml_tensor * build_graph(ggml_context * ctx) override {
+        ggml_tensor * a;
+        ggml_tensor * b;
+        if (v & 1) {
+            auto ne = ne_a; ne[0] *= 3;
+            a = ggml_new_tensor(ctx, type, 4, ne.data());
+            ggml_set_param(a);
+            ggml_set_name(a, "a");
+
+            a = ggml_view_4d(ctx, a, ne_a[0], ne_a[1], ne_a[2], ne_a[3], a->nb[1], a->nb[2], a->nb[3], 0);
+            ggml_set_name(a, "view_of_a");
+
+            b = ggml_new_tensor(ctx, type, 4, ne.data());
+            ggml_set_param(b);
+            ggml_set_name(b, "b");
+
+            b = ggml_view_4d(ctx, b, ne_a[0], ne_a[1], ne_a[2], ne_a[3], b->nb[1], b->nb[2], b->nb[3], 0);
+            ggml_set_name(a, "view_of_b");
+        } else {
+            a = ggml_new_tensor(ctx, type, 4, ne_a.data());
+            ggml_set_param(a);
+            ggml_set_name(a, "a");
+
+            b = ggml_new_tensor(ctx, type, 4, ne_a.data());
+            ggml_set_param(b);
+            ggml_set_name(b, "b");
+        }
+
+        ggml_tensor * out = ggml_glu_split(ctx, a, b, op);
+        ggml_set_name(out, "out");
+
+        return out;
+    }
+
+    void initialize_tensors(ggml_context * ctx) override {
+        for (ggml_tensor * t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) {
+            // test extended range of values to check for NaNs in GELU
+            init_tensor_uniform(t, -150.f, 150.f);
+        }
+    }
+};
+
 // GGML_OP_GET_ROWS
 struct test_get_rows : public test_case {
     const ggml_type type;
@@ -1213,6 +1856,76 @@ struct test_get_rows_back : public test_case {
     }
 };
 
+// GGML_OP_SET_ROWS
+struct test_set_rows : public test_case {
+    const ggml_type type;
+    const std::array<int64_t, 4> ne;
+    const std::array<int, 2> nr23; // broadcast only dims 2 and 3
+    const int r; // rows to set
+    const bool v; // view (non-contiguous src1)
+
+    std::string vars() override {
+        return VARS_TO_STR5(type, ne, nr23, r, v);
+    }
+
+    test_set_rows(ggml_type type,
+            std::array<int64_t, 4> ne,
+            std::array<int, 2> nr23,
+            int r, bool v = false)
+        : type(type), ne(ne), nr23(nr23), r(r), v(v) {}
+
+    ggml_tensor * build_graph(ggml_context * ctx) override {
+        ggml_tensor * dst = ggml_new_tensor_4d(ctx, type,          ne[0], ne[1], ne[2]*nr23[0], ne[3]*nr23[1]);
+        ggml_set_name(dst, "dst");
+
+        ggml_tensor * src = ggml_new_tensor_4d(ctx, GGML_TYPE_F32, ne[0], r,     ne[2]*nr23[0], ne[3]*nr23[1]);
+        ggml_set_name(src, "src");
+
+        ggml_tensor * row_idxs = ggml_new_tensor_3d(ctx, GGML_TYPE_I64, r, ne[2], ne[3]);
+        ggml_set_name(row_idxs, "row_idxs");
+
+        if (v) {
+            src      = ggml_view_4d(ctx, src, ne[0], r/2, ne[2]*nr23[0], ne[3]*nr23[1], src->nb[1], src->nb[2], src->nb[3], 0);
+            row_idxs = ggml_view_3d(ctx, row_idxs, r/2, ne[2], ne[3], row_idxs->nb[1], row_idxs->nb[2], 0);
+            ggml_set_name(row_idxs, "view_of_rows");
+        }
+
+        ggml_tensor * out = ggml_set_rows(ctx, dst, src, row_idxs);
+        ggml_set_name(out, "out");
+
+        return out;
+    }
+
+    void initialize_tensors(ggml_context * ctx) override {
+        std::random_device rd;
+        std::default_random_engine rng(rd());
+        for (ggml_tensor * t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) {
+            if (t->type == GGML_TYPE_I64) {
+                if (ggml_is_view_op(t->op)) {
+                    continue;
+                }
+
+                for (int i2 = 0; i2 < t->ne[2]; i2++) {
+                    for (int i1 = 0; i1 < t->ne[1]; i1++) {
+                        // generate a shuffled subset of row indices
+                        std::vector<int64_t> data(ne[1]);
+                        for (int i = 0; i < ne[1]; i++) {
+                            data[i] = i;
+                        }
+                        std::shuffle(data.begin(), data.end(), rng);
+                        data.resize(t->ne[0]);
+
+                        const size_t offs = i1*t->nb[1] + i2*t->nb[2];
+                        ggml_backend_tensor_set(t, data.data(), offs, t->ne[0]*sizeof(int64_t));
+                    }
+                }
+            } else {
+                init_tensor_uniform(t);
+            }
+        }
+    }
+};
+
 // GGML_OP_ARGMAX
 struct test_argmax : public test_case {
     const ggml_type type;
@@ -1826,6 +2539,59 @@ struct test_rms_norm_back : public test_case {
     }
 };
 
+// GGML_OP_RMS_NORM + GGML_OP_MUL
+struct test_rms_norm_mul : public test_case {
+    const ggml_type type;
+    const std::array<int64_t, 4> ne;
+    const float eps;
+
+    std::string op_desc(ggml_tensor * t) override {
+        GGML_UNUSED(t);
+        return "RMS_NORM_MUL";
+    }
+
+    bool run_whole_graph() override { return true; }
+
+    std::string vars() override {
+        return VARS_TO_STR3(type, ne, eps);
+    }
+
+    test_rms_norm_mul(ggml_type type = GGML_TYPE_F32,
+            std::array<int64_t, 4> ne = {64, 5, 4, 3},
+            float eps = 1e-6f)
+        : 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_tensor * b = ggml_new_tensor(ctx, type, 4, ne.data());
+        ggml_set_param(a);
+        ggml_set_name(a, "a");
+        ggml_set_param(b);
+        ggml_set_name(b, "b");
+
+        // Use a and b early, so we don't end up with an OP_NONE between rms_norm and mul
+        a = ggml_add(ctx, a, b);
+        ggml_tensor * out = ggml_mul(ctx, ggml_rms_norm(ctx, a, eps), b);
+        ggml_set_name(out, "out");
+
+        return out;
+    }
+
+    void initialize_tensors(ggml_context * ctx) override {
+        for (ggml_tensor * t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) {
+            init_tensor_uniform(t, -10.f, 10.f);
+        }
+    }
+
+    float grad_eps() override {
+        return 1.0f;
+    }
+
+    bool grad_precise() override {
+        return true;
+    }
+};
+
 // GGML_OP_SSM_CONV
 struct test_ssm_conv : public test_case {
     const ggml_type type;
@@ -1854,28 +2620,58 @@ struct test_ssm_scan : public test_case {
     const ggml_type type;
 
     const int64_t d_state;
-    const int64_t d_inner;
+    const int64_t head_dim;
+    const int64_t n_head;
+    const int64_t n_group;
     const int64_t n_seq_tokens;
     const int64_t n_seqs;
 
     std::string vars() override {
-        return VARS_TO_STR5(type, d_state, d_inner, n_seq_tokens, n_seqs);
+        return VARS_TO_STR7(type, d_state, head_dim, n_head, n_group, n_seq_tokens, n_seqs);
     }
 
     test_ssm_scan(ggml_type type = GGML_TYPE_F32,
-            int64_t d_state = 32, int64_t d_inner = 32, int64_t n_seq_tokens = 32, int64_t n_seqs = 32)
-        : type(type), d_state(d_state), d_inner(d_inner), n_seq_tokens(n_seq_tokens), n_seqs(n_seqs) {}
+            int64_t d_state = 32,
+            int64_t head_dim = 1, // non-zero for Mamba-2
+            int64_t n_head  = 32,
+            int64_t n_group = 1,
+            int64_t n_seq_tokens = 32,
+            int64_t n_seqs = 32)
+        : type(type), d_state(d_state), head_dim(head_dim), n_head(n_head), n_group(n_group), n_seq_tokens(n_seq_tokens), n_seqs(n_seqs) {}
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
-        ggml_tensor * s   = ggml_new_tensor(ctx, type, 4, std::vector<int64_t>{ d_state, d_inner,      n_seqs, 1 }.data());
-        ggml_tensor * x   = ggml_new_tensor(ctx, type, 4, std::vector<int64_t>{ d_inner, n_seq_tokens, n_seqs, 1 }.data());
-        ggml_tensor * dt  = ggml_new_tensor(ctx, type, 4, std::vector<int64_t>{ d_inner, n_seq_tokens, n_seqs, 1 }.data());
-        ggml_tensor * A   = ggml_new_tensor(ctx, type, 4, std::vector<int64_t>{ d_state, d_inner,      1     , 1 }.data());
-        ggml_tensor * B   = ggml_new_tensor(ctx, type, 4, std::vector<int64_t>{ d_state, n_seq_tokens, n_seqs, 1 }.data());
-        ggml_tensor * C   = ggml_new_tensor(ctx, type, 4, std::vector<int64_t>{ d_state, n_seq_tokens, n_seqs, 1 }.data());
-        ggml_tensor * out = ggml_ssm_scan(ctx, s, x, dt, A, B, C);
+        ggml_tensor * s   = ggml_new_tensor_4d(ctx, type, d_state,  head_dim,     n_head,       n_seqs);
+        ggml_tensor * x   = ggml_new_tensor_4d(ctx, type, head_dim, n_head,       n_seq_tokens, n_seqs);
+        ggml_tensor * dt  = ggml_new_tensor_3d(ctx, type, n_head,   n_seq_tokens, n_seqs);
+        ggml_tensor * A   = ggml_new_tensor_2d(ctx, type, (head_dim > 1) ? 1 : d_state, n_head);
+        ggml_tensor * B   = ggml_new_tensor_4d(ctx, type, d_state,  n_group,      n_seq_tokens, n_seqs);
+        ggml_tensor * C   = ggml_new_tensor_4d(ctx, type, d_state,  n_group,      n_seq_tokens, n_seqs);
+        ggml_tensor * ids = ggml_new_tensor_1d(ctx, GGML_TYPE_I32,  n_seqs);
+        ggml_tensor * out = ggml_ssm_scan(ctx, s, x, dt, A, B, C, ids);
         return out;
     }
+
+    // similar to test_mul_mat_id
+    void initialize_tensors(ggml_context * ctx) override {
+        std::random_device rd;
+        std::default_random_engine rng(rd());
+        for (ggml_tensor * t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) {
+            if (t->type == GGML_TYPE_I32) {
+                if (ggml_is_view_op(t->op)) { continue; }
+                // ids
+                for (int64_t r = 0; r < ggml_nrows(t); r++) {
+                    std::vector<int32_t> data(t->ne[0]);
+                    for (int i = 0; i < t->ne[0]; i++) {
+                        data[i] = i;
+                    }
+                    std::shuffle(data.begin(), data.end(), rng);
+                    ggml_backend_tensor_set(t, data.data(), r * t->nb[1], t->ne[0] * sizeof(int32_t));
+                }
+            } else {
+                init_tensor_uniform(t);
+            }
+        }
+    }
 };
 
 // GGML_OP_RWKV_WKV6
@@ -2455,11 +3251,12 @@ struct test_soft_max : public test_case {
     const std::array<int64_t, 4> ne;
     const bool mask;
     const ggml_type m_prec;
+    const std::array<int64_t, 2> nr23; // broadcast only dims 2 and 3
     const float scale;
     const float max_bias;
 
     std::string vars() override {
-        return VARS_TO_STR6(type, ne, mask, m_prec, scale, max_bias);
+        return VARS_TO_STR7(type, ne, mask, m_prec, nr23, scale, max_bias);
     }
 
     // the 1024 test with bias occasionally fails:
@@ -2472,18 +3269,19 @@ struct test_soft_max : public test_case {
             std::array<int64_t, 4> ne = {10, 5, 4, 3},
             bool mask = false,
             ggml_type m_prec = GGML_TYPE_F32,
+            std::array<int64_t, 2> nr23 = {1, 1},
             float scale = 1.0f,
             float max_bias = 0.0f)
-        : type(type), ne(ne), mask(mask), m_prec(m_prec), scale(scale), max_bias(max_bias) {}
+        : type(type), ne(ne), mask(mask), m_prec(m_prec), nr23(nr23), scale(scale), max_bias(max_bias) {}
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
-        ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data());
+        ggml_tensor * a = ggml_new_tensor_4d(ctx, type, ne[0], ne[1], ne[2]*nr23[0], ne[3]*nr23[1]);
         ggml_set_param(a);
         ggml_set_name(a, "a");
 
         ggml_tensor * mask = nullptr;
         if (this->mask) {
-            mask = ggml_new_tensor_2d(ctx, m_prec, ne[0], ne[1]);
+            mask = ggml_new_tensor_4d(ctx, m_prec, ne[0], ne[1], ne[2], ne[3]);
             ggml_set_name(mask, "mask");
         }
 
@@ -2706,8 +3504,8 @@ struct test_conv_transpose_1d : public test_case {
         return VARS_TO_STR5(ne_input, ne_kernel, s0, p0, d0);
     }
 
-    test_conv_transpose_1d(std::array<int64_t, 4> ne_input = {197, 32, 1, 1}, // [input_width, input_height, input_channels, 1]
-                           std::array<int64_t, 4> ne_kernel = {16, 32, 32, 1}, // [kernel_width, kernel_height, input_channels, 1]
+    test_conv_transpose_1d(std::array<int64_t, 4> ne_input = {197, 32, 1, 1}, // [input_width, input_channels, 1 /* assert in cpu kernel*/, 1 (should be batch)]
+                           std::array<int64_t, 4> ne_kernel = {16, 32, 32, 1}, // [kernel_width, output_channels, input_channels, 1 (should be batch)]
                            int s0 = 1, int p0 = 0, int d0 = 1)
         : ne_input(ne_input), ne_kernel(ne_kernel), s0(s0), p0(p0), d0(d0) {}
 
@@ -2725,6 +3523,35 @@ struct test_conv_transpose_1d : public test_case {
     }
 };
 
+// GGML_OP_CONV_TRANSPOSE_2D
+struct test_conv_transpose_2d : public test_case {
+    const std::array<int64_t, 4> ne_input;
+    const std::array<int64_t, 4> ne_kernel;
+    const int stride;
+
+    std::string vars() override {
+        return VARS_TO_STR3(ne_input, ne_kernel, stride);
+    }
+
+    test_conv_transpose_2d(std::array<int64_t, 4> ne_input = {10, 10, 3, 1}, // [input_width, input_height, input_channels, 1]
+                           std::array<int64_t, 4> ne_kernel = {3, 3, 3, 1}, // [kernel_width, kernel_height, input_channels, 1]
+                           int stride = 1)
+        : ne_input(ne_input), ne_kernel(ne_kernel), stride(stride){}
+
+    ggml_tensor * build_graph(ggml_context * ctx) override {
+        ggml_tensor * input = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne_input.data());
+        ggml_set_name(input, "input");
+
+        ggml_tensor * kernel = ggml_new_tensor(ctx, GGML_TYPE_F16, 4, ne_kernel.data());
+        ggml_set_name(kernel, "kernel");
+
+        ggml_tensor * out = ggml_conv_transpose_2d_p0(ctx, kernel, input, stride);
+        ggml_set_name(out, "out");
+
+        return out;
+    }
+};
+
 // GGML_OP_IM2COL
 struct test_im2col : public test_case {
     const ggml_type type_input;
@@ -3037,28 +3864,28 @@ struct test_upscale : public test_case {
     }
 };
 
-// GGML_OP_UPSCALE (ext)
-struct test_upscale_ext : public test_case {
+// GGML_OP_UPSCALE (via ggml_interpolate)
+struct test_interpolate : public test_case {
     const ggml_type type;
     const std::array<int64_t, 4> ne;
     const std::array<int64_t, 4> ne_tgt;
-    const ggml_scale_mode mode = GGML_SCALE_MODE_NEAREST;
+    const uint32_t mode = GGML_SCALE_MODE_NEAREST;
 
     std::string vars() override {
         return VARS_TO_STR4(type, ne, ne_tgt, mode);
     }
 
-    test_upscale_ext(ggml_type type = GGML_TYPE_F32,
+    test_interpolate(ggml_type type = GGML_TYPE_F32,
             std::array<int64_t, 4> ne     = {2, 5,  7, 11},
             std::array<int64_t, 4> ne_tgt = {5, 7, 11, 13},
-            ggml_scale_mode mode = GGML_SCALE_MODE_NEAREST)
+            uint32_t 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], mode);
+        ggml_tensor * out = ggml_interpolate(ctx, a, ne_tgt[0], ne_tgt[1],ne_tgt[2], ne_tgt[3], mode);
         ggml_set_name(out, "out");
 
         return out;
@@ -3285,7 +4112,7 @@ struct test_flash_attn_ext : public test_case {
     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 std::array<int64_t, 2> nr23; // repeat in dim 2 and 3, tests for grouped-query attention
     const int64_t kv; // kv size
     const int64_t nb; // batch size
 
@@ -3299,7 +4126,7 @@ struct test_flash_attn_ext : public test_case {
     std::array<int32_t, 4> permute;
 
     std::string vars() override {
-        return VARS_TO_STR12(hsk, hsv, nh, nr, kv, nb, mask, max_bias, logit_softcap, prec, type_KV, permute);
+        return VARS_TO_STR12(hsk, hsv, nh, nr23, kv, nb, mask, max_bias, logit_softcap, prec, type_KV, permute);
     }
 
     double max_nmse_err() override {
@@ -3310,13 +4137,13 @@ struct test_flash_attn_ext : public test_case {
         GGML_UNUSED(t);
         // Just counting matmul costs:
         // 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;
+        return (2 * nh*nr23[0] * nb * (hsk + hsv) * kv)*nr23[1];
     }
 
-    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,
+    test_flash_attn_ext(int64_t hsk = 128, int64_t hsv = 128, int64_t nh = 32, std::array<int64_t, 2> nr23 = {1, 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<int32_t, 4> 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) {}
+        : hsk(hsk), hsv(hsv), nh(nh), nr23(nr23), 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 hsk_padded = GGML_PAD(hsk, ggml_blck_size(type_KV));
@@ -3335,18 +4162,18 @@ struct test_flash_attn_ext : public test_case {
             return t;
         };
 
-        ggml_tensor * q = create_permuted(GGML_TYPE_F32, hsk_padded, nb, nh*nr, 1);
+        ggml_tensor * q = create_permuted(GGML_TYPE_F32, hsk_padded, nb, nh*nr23[0], nr23[1]);
         ggml_set_name(q, "q");
 
-        ggml_tensor * k = create_permuted(type_KV,       hsk_padded, kv, nh,    1);
+        ggml_tensor * k = create_permuted(type_KV,       hsk_padded, kv, nh,         nr23[1]);
         ggml_set_name(k, "k");
 
-        ggml_tensor * v = create_permuted(type_KV,       hsv_padded, kv, nh,    1);
+        ggml_tensor * v = create_permuted(type_KV,       hsv_padded, kv, nh,         nr23[1]);
         ggml_set_name(v, "v");
 
         ggml_tensor * m = nullptr;
         if (mask) {
-            m = ggml_new_tensor_4d(ctx, GGML_TYPE_F16, kv, GGML_PAD(nb, GGML_KQ_MASK_PAD), 1, 1);
+            m = ggml_new_tensor_4d(ctx, GGML_TYPE_F16, kv, GGML_PAD(nb, GGML_KQ_MASK_PAD), nr23[0], nr23[1]);
             ggml_set_name(m, "m");
         }
 
@@ -3637,6 +4464,7 @@ struct test_llama : public test_llm {
     static constexpr float attn_factor = 1.0f;
     static constexpr float beta_fast = 32.0f;
     static constexpr float beta_slow = 1.0f;
+    bool fused;
 
     std::string op_desc(ggml_tensor * t) override {
         GGML_UNUSED(t);
@@ -3652,7 +4480,9 @@ struct test_llama : public test_llm {
         return 2e-3;
     }
 
-    test_llama(int n_tokens = 1)
+    bool run_whole_graph() override { return fused; }
+
+    test_llama(int n_tokens = 1, bool fused = false)
         : test_llm({
             /*n_vocab        =*/ 32000,
             /*n_embd         =*/ 3200,
@@ -3664,7 +4494,9 @@ struct test_llama : public test_llm {
             /*f_norm_eps     =*/ 0.f,
             /*f_norm_rms_eps =*/ 1e-5f,
             /*n_tokens       =*/ n_tokens,
-        }) {
+        })
+        , fused(fused)
+    {
     }
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
@@ -3931,6 +4763,21 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
         }
     }
 
+    // glu ops
+    for (ggml_type type : {GGML_TYPE_F16, GGML_TYPE_F32}) {
+        for (int v : {0, 1}) {
+            for (int op = 0; op < GGML_GLU_OP_COUNT; op++) {
+                for (bool swapped : {false, true}) {
+                    test_cases.emplace_back(new test_glu((ggml_glu_op) op, type, { 128, 2, 2, 2 }, v, swapped));
+                    test_cases.emplace_back(new test_glu((ggml_glu_op) op, type, { 5, 7, 11, 13 }, v, swapped));
+                }
+
+                test_cases.emplace_back(new test_glu_split((ggml_glu_op) op, type, { 128, 2, 2, 2 }, v));
+                test_cases.emplace_back(new test_glu_split((ggml_glu_op) op, type, { 5, 7, 11, 13 }, v));
+            }
+        }
+    }
+
     test_cases.emplace_back(new test_get_rows(GGML_TYPE_F32, 1, 8, 2, 1, false));
     for (ggml_type type : all_types) {
         for (int b : {1, 7}) {
@@ -3955,6 +4802,23 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
         test_cases.emplace_back(new test_get_rows_back(GGML_TYPE_I32, 256, 5, 4, 1, v));
     }
 
+    test_cases.emplace_back(new test_set_rows(GGML_TYPE_F32, { 1, 8, 1, 3 }, { 1, 1 }, 2, false));
+    for (ggml_type type : all_types) {
+        for (int b : {1, 7}) {
+            for (bool v : {false, true}) {
+                test_cases.emplace_back(new test_set_rows(type, { 256, 5,  b, 3 }, { 1, 1, }, 1, v));
+                test_cases.emplace_back(new test_set_rows(type, { 256, 11, 1, b }, { 2, 3, }, 7, v));
+
+                test_cases.emplace_back(new test_set_rows(type, { 3*ggml_blck_size(type), 3, b, 1 }, { 2, 3, }, 2, v));
+
+                if (ggml_blck_size(type) == 1) {
+                    test_cases.emplace_back(new test_set_rows(type, { 31, 3, b, 1 }, { 2, 3, }, 2, v));
+                    test_cases.emplace_back(new test_set_rows(type, { 33, 5, 1, b }, { 2, 3, }, 1, v));
+                }
+            }
+        }
+    }
+
     for (ggml_type type_input : {GGML_TYPE_F32}) {
         for (ggml_op_pool pool_type : {GGML_OP_POOL_AVG, GGML_OP_POOL_MAX}) {
             for (int k0 : {1, 3}) {
@@ -4029,6 +4893,18 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
     test_cases.emplace_back(new test_conv_2d_dw({32, 8, 64, 1}, {3, 3, 1, 64}, 2, 1, 1, false));
     test_cases.emplace_back(new test_conv_2d_dw({32, 8, 64, 1}, {3, 3, 1, 64}, 2, 1, 1, true));
 
+    for(uint32_t Cout : {1, 9}){
+        for(uint32_t Cin : {1, 7}){
+            for(uint32_t K : {1, 3, 1337}){
+                for(uint32_t L : {1, 2, 13}){
+                    for(uint32_t s0: {1, 2, 3}){
+                        test_cases.emplace_back(new test_conv_transpose_1d({L,Cin,1,1}, {K,Cout,Cin,1}, s0, 0, 1));
+                    }
+                }
+            }
+        }
+    }
+
     test_cases.emplace_back(new test_conv_transpose_1d());
     test_cases.emplace_back(new test_conv_transpose_1d({3,2,1,1}, {2,3,2,1}, 3, 0, 1));
     test_cases.emplace_back(new test_conv_transpose_1d({3,2,1,1}, {2,3,2,1}, 2, 0, 1));
@@ -4038,6 +4914,9 @@ static std::vector<std::unique_ptr<test_case>> 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_conv_transpose_2d({3, 2, 3, 1}, {2, 2, 1, 3}, 1));
+    test_cases.emplace_back(new test_conv_transpose_2d({10, 10, 9, 1}, {3, 3, 1, 9}, 2));
+
     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}));
 
@@ -4175,6 +5054,9 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
         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));
     }
+    for (float eps : {0.0f, 1e-6f, 1e-4f, 1e-1f, 1.0f}) {
+        test_cases.emplace_back(new test_rms_norm_mul(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));
 
@@ -4182,7 +5064,8 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
     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}));
 
-    test_cases.emplace_back(new test_ssm_scan(GGML_TYPE_F32, 16, 1024, 32, 4));
+    test_cases.emplace_back(new test_ssm_scan(GGML_TYPE_F32, 16, 1, 1024, 1, 32, 4)); // Mamba-1
+    test_cases.emplace_back(new test_ssm_scan(GGML_TYPE_F32, 128, 64, 16, 2, 32, 4)); // Mamba-2
 
     test_cases.emplace_back(new test_rwkv_wkv6(GGML_TYPE_F32, 32, 64, 1, 1));
     test_cases.emplace_back(new test_rwkv_wkv6(GGML_TYPE_F32, 32, 64, 32, 1));
@@ -4208,39 +5091,45 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
 #if 1
     for (ggml_type type_a : base_types) {
         for (ggml_type type_b : {GGML_TYPE_F32, GGML_TYPE_F16}) {
-            // test cases without permutation
-            test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16,  1, 256, {1, 1}, {1, 1}));
-            test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16,  1, 256, {1, 1}, {2, 1}));
-            test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16,  1, 256, {1, 1}, {1, 2}));
-            test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16,  1, 256, {3, 1}, {1, 1}));
-            test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16,  1, 256, {3, 1}, {2, 1}));
-            test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16,  1, 256, {3, 2}, {1, 1}));
-            test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16,  1, 256, {3, 2}, {2, 1}));
-            test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16,  1, 256, {3, 2}, {1, 2}));
-            test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16,  1, 256, {3, 2}, {2, 2}));
+            std::vector<int> ks = { 256 };
+            if (ggml_blck_size(type_a) == 1) {
+                ks.push_back(4);
+            }
+            for (auto k : ks) {
+                // test cases without permutation
+                test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16,  1, k, {1, 1}, {1, 1}));
+                test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16,  1, k, {1, 1}, {2, 1}));
+                test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16,  1, k, {1, 1}, {1, 2}));
+                test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16,  1, k, {3, 1}, {1, 1}));
+                test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16,  1, k, {3, 1}, {2, 1}));
+                test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16,  1, k, {3, 2}, {1, 1}));
+                test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16,  1, k, {3, 2}, {2, 1}));
+                test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16,  1, k, {3, 2}, {1, 2}));
+                test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16,  1, k, {3, 2}, {2, 2}));
 
-            test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 16, 256, {1, 1}, {1, 1}));
-            test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 16, 256, {1, 1}, {2, 1}));
-            test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 16, 256, {1, 1}, {1, 2}));
-            test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 16, 256, {3, 1}, {1, 1}));
-            test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 16, 256, {3, 1}, {2, 1}));
-            test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 16, 256, {3, 2}, {1, 1}));
-            test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 16, 256, {3, 2}, {2, 1}));
-            test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 16, 256, {3, 2}, {1, 2}));
-            test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 16, 256, {3, 2}, {2, 2}));
+                test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 16, k, {1, 1}, {1, 1}));
+                test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 16, k, {1, 1}, {2, 1}));
+                test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 16, k, {1, 1}, {1, 2}));
+                test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 16, k, {3, 1}, {1, 1}));
+                test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 16, k, {3, 1}, {2, 1}));
+                test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 16, k, {3, 2}, {1, 1}));
+                test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 16, k, {3, 2}, {2, 1}));
+                test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 16, k, {3, 2}, {1, 2}));
+                test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 16, k, {3, 2}, {2, 2}));
 
-            // test cases with permutation
-            test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16,  1, 256, {2, 3}, {1, 1}, {0, 2, 1, 3}));
-            test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16,  1, 256, {2, 3}, {1, 1}, {0, 1, 3, 2}));
-            test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16,  1, 256, {2, 3}, {1, 1}, {0, 3, 2, 1}));
+                // test cases with permutation
+                test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16,  1, k, {2, 3}, {1, 1}, {0, 2, 1, 3}));
+                test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16,  1, k, {2, 3}, {1, 1}, {0, 1, 3, 2}));
+                test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16,  1, k, {2, 3}, {1, 1}, {0, 3, 2, 1}));
 
-            test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16,  8, 256, {2, 3}, {1, 1}, {0, 2, 1, 3}));
-            test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16,  8, 256, {2, 3}, {1, 1}, {0, 1, 3, 2}));
-            test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16,  8, 256, {2, 3}, {1, 1}, {0, 3, 2, 1}));
+                test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16,  8, k, {2, 3}, {1, 1}, {0, 2, 1, 3}));
+                test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16,  8, k, {2, 3}, {1, 1}, {0, 1, 3, 2}));
+                test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16,  8, k, {2, 3}, {1, 1}, {0, 3, 2, 1}));
 
-            test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 16, 256, {2, 3}, {1, 1}, {0, 2, 1, 3}));
-            test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 16, 256, {2, 3}, {1, 1}, {0, 1, 3, 2}));
-            test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 16, 256, {2, 3}, {1, 1}, {0, 3, 2, 1}));
+                test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 16, k, {2, 3}, {1, 1}, {0, 2, 1, 3}));
+                test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 16, k, {2, 3}, {1, 1}, {0, 1, 3, 2}));
+                test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 16, k, {2, 3}, {1, 1}, {0, 3, 2, 1}));
+            }
 
             // test cases with large ne00/ne10 to cover stream-k fixup
             test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16,  1, 1024, {3, 2}, {1, 1}));
@@ -4288,8 +5177,10 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
         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));
+                    for (ggml_type type: {GGML_TYPE_F16, GGML_TYPE_BF16, GGML_TYPE_F32}) {
+                        test_cases.emplace_back(new test_mul_mat(type, GGML_TYPE_F32, 1056 + m, 1, 128 + k,  {bs,  1}, {nr, 1}, {0, 2, 1, 3}));
+                        test_cases.emplace_back(new test_mul_mat(type, GGML_TYPE_F32, 128 + m,  1, 1056 + k, {bs,  1}, {nr, 1}, {0, 1, 2, 3}, true));
+                    }
                 }
             }
         }
@@ -4301,6 +5192,11 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
     // this case is verified (pass) in Intel(R) Data Center GPU Max 1100 (sycl backend) and NV A30 (cuda backend)
     // test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F16, 512, 262144, 9216, {1, 1}, {1, 1}));
 
+    // test large experts*tokens
+    for (bool b : {false, true}) {
+        test_cases.emplace_back(new test_mul_mat_id(GGML_TYPE_F16, GGML_TYPE_F32, 16, 16, b, 32, 1024, 16));
+    }
+
     for (ggml_type type_a : base_types) {
         for (ggml_type type_b : {GGML_TYPE_F32 /*, GGML_TYPE_F16 */}) {
             for (int n_mats : {4, 8}) {
@@ -4387,26 +5283,31 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
                     for (int64_t ne1 : {16, 1024}) {
                         if (mask) {
                             for (ggml_type m_prec : {GGML_TYPE_F32, GGML_TYPE_F16}) {
-                                test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {ne0,   ne1,   1, 1}, mask, m_prec, scale, max_bias));
-                                test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {ne0-1, ne1-1, 1, 1}, mask, m_prec, scale, max_bias));
+                                test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {ne0,   ne1,   1, 1}, mask, m_prec, {1, 1}, scale, max_bias));
+                                test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {ne0-1, ne1-1, 1, 1}, mask, m_prec, {1, 1}, scale, max_bias));
+
+                                if (ne0 <= 32 && ne1 <= 32) {
+                                    test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {ne0,   ne1,   1, 3}, mask, m_prec, {3, 1}, scale, max_bias));
+                                    test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {ne0-1, ne1-1, 1, 1}, mask, m_prec, {2, 3}, scale, max_bias));
+                                }
                             }
                         } else {
                             /* The precision of mask here doesn't matter as boolean mask is false */
-                            test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {ne0,   ne1,   1, 1}, mask, GGML_TYPE_F32, scale, max_bias));
-                            test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {ne0-1, ne1-1, 1, 1}, mask, GGML_TYPE_F32, scale, max_bias));
+                            test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {ne0,   ne1,   1, 1}, mask, GGML_TYPE_F32, {1, 1}, scale, max_bias));
+                            test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {ne0-1, ne1-1, 1, 1}, mask, GGML_TYPE_F32, {1, 1}, scale, max_bias));
                         }
                     }
                 }
             }
         }
     }
-    test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {16, 2, 32, 1}, true, GGML_TYPE_F32,  0.1f, 0.0f));
-    test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {16, 2, 32, 1}, true, GGML_TYPE_F16,  0.1f, 0.0f));
-    test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {16, 2, 32, 1}, false, GGML_TYPE_F32, 0.1f, 0.0f));
-    test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {32, 2, 32, 1}, true, GGML_TYPE_F32,  0.1f, 0.0f));
-    test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {32, 2, 32, 1}, true, GGML_TYPE_F16,  0.1f, 0.0f));
-    test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {32, 2, 32, 1}, true, GGML_TYPE_F32,  0.1f, 8.0f));
-    test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {32, 2, 32, 1}, true, GGML_TYPE_F16,  0.1f, 8.0f));
+    test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {16, 2, 32, 1}, true,  GGML_TYPE_F32, {1, 1}, 0.1f, 0.0f));
+    test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {16, 2, 32, 1}, true,  GGML_TYPE_F16, {1, 1}, 0.1f, 0.0f));
+    test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {16, 2, 32, 1}, false, GGML_TYPE_F32, {1, 1}, 0.1f, 0.0f));
+    test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {32, 2, 32, 1}, true,  GGML_TYPE_F32, {1, 1}, 0.1f, 0.0f));
+    test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {32, 2, 32, 1}, true,  GGML_TYPE_F16, {1, 1}, 0.1f, 0.0f));
+    test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {32, 2, 32, 1}, true,  GGML_TYPE_F32, {1, 1}, 0.1f, 8.0f));
+    test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {32, 2, 32, 1}, true,  GGML_TYPE_F16, {1, 1}, 0.1f, 8.0f));
 
     for (float max_bias : {0.0f, 8.0f}) {
         for (float scale : {1.0f, 0.1f}) {
@@ -4477,8 +5378,10 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
     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_interpolate(GGML_TYPE_F32, {2, 5,  7, 11}, {5, 7, 11, 13}, mode));
+        test_cases.emplace_back(new test_interpolate(GGML_TYPE_F32, {5, 7, 11, 13}, {2, 5,  7, 11}, mode));
     }
+    test_cases.emplace_back(new test_interpolate(GGML_TYPE_F32, {2, 5,  7, 11}, {5, 7, 11, 13}, GGML_SCALE_MODE_BILINEAR | GGML_SCALE_FLAG_ALIGN_CORNERS));
 
     test_cases.emplace_back(new test_sum());
     test_cases.emplace_back(new test_sum_rows());
@@ -4504,20 +5407,23 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
                     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) {
+                            for (int nr3 : { 1, 3, }) {
+                                if (hsk > 64 && nr3 > 1) continue; // skip broadcast for large head sizes
+                                for (int nr2 : { 1, 4, 16 }) {
+                                    if (nr2 == 16 && hsk != 128) continue;
+                                    for (int kv : { 512, 1024, }) {
+                                        if (nr2 != 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, {0, 2, 1, 3}));
+                                                                hsk, hsv, nh, {nr2, nr3}, 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, {nr2, nr3}, kv, nb, mask, max_bias, logit_softcap, prec, type_KV, {0, 2, 1, 3}));
+                                                    }
                                                 }
                                             }
                                         }
@@ -4538,8 +5444,9 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
 
     test_cases.emplace_back(new test_opt_step_adamw(GGML_TYPE_F32, {10, 5, 4, 3}));
 
-    // these tests are disabled to save execution time, but they can be handy for debugging
 #if 0
+    // these tests are disabled to save execution time, sbut they can be handy for debugging
+    test_cases.emplace_back(new test_llama(2, true));
     test_cases.emplace_back(new test_llama(1));
     test_cases.emplace_back(new test_llama(2));
     test_cases.emplace_back(new test_falcon(1));
@@ -4560,13 +5467,14 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_perf() {
     test_cases.emplace_back(new test_cpy(GGML_TYPE_F32, GGML_TYPE_F32, {8192, 512, 2, 1}, {0, 2, 1, 3}));
     test_cases.emplace_back(new test_cpy(GGML_TYPE_F32, GGML_TYPE_F32, {3072, 512, 2, 1}, {0, 2, 1, 3}));
 
-    test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {4096, 4096, 5, 1}, false, GGML_TYPE_F32, 1.0f, 0.0f));
-    test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {77, 4096, 5, 1}, false, GGML_TYPE_F32, 1.0f, 0.0f));
-    test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {1024, 1024, 10, 1}, false, GGML_TYPE_F32, 1.0f, 0.0f));
-    test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {77, 1024, 10, 1}, false, GGML_TYPE_F32, 1.0f, 0.0f));
-    test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {256, 256, 20, 1}, false, GGML_TYPE_F32, 1.0f, 0.0f));
-    test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {64, 64, 20, 1}, false, GGML_TYPE_F32, 1.0f, 0.0f));
-    test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {77, 64, 20, 1}, false, GGML_TYPE_F32, 1.0f, 0.0f));
+    test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {4096, 4096, 5, 1}, false, GGML_TYPE_F32, {1, 1}, 1.0f, 0.0f));
+    test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {12888, 256, 5, 1}, false, GGML_TYPE_F32, {1, 1}, 1.0f, 0.0f));
+    test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {77, 4096, 5, 1}, false, GGML_TYPE_F32, {1, 1}, 1.0f, 0.0f));
+    test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {1024, 1024, 10, 1}, false, GGML_TYPE_F32, {1, 1}, 1.0f, 0.0f));
+    test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {77, 1024, 10, 1}, false, GGML_TYPE_F32, {1, 1}, 1.0f, 0.0f));
+    test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {256, 256, 20, 1}, false, GGML_TYPE_F32, {1, 1}, 1.0f, 0.0f));
+    test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {64, 64, 20, 1}, false, GGML_TYPE_F32, {1, 1}, 1.0f, 0.0f));
+    test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {77, 64, 20, 1}, false, GGML_TYPE_F32, {1, 1}, 1.0f, 0.0f));
 
     test_cases.emplace_back(new test_argmax(GGML_TYPE_F32, {32, 10, 1, 1}));
     test_cases.emplace_back(new test_argmax(GGML_TYPE_F32, {1024, 10, 1, 1}));
@@ -4598,7 +5506,7 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_perf() {
     for (int kv : { 4096, 8192, 16384, }) {
         for (int hs : { 64, 128, }) {
             for (int nr : { 1, 4, }) {
-                test_cases.emplace_back(new test_flash_attn_ext(hs, hs, 8, nr, kv, 1, true, 0, 0, GGML_PREC_F32, GGML_TYPE_F16));
+                test_cases.emplace_back(new test_flash_attn_ext(hs, hs, 8, {nr, 1}, kv, 1, true, 0, 0, GGML_PREC_F32, GGML_TYPE_F16));
             }
         }
     }
@@ -4606,10 +5514,15 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_perf() {
     test_cases.emplace_back(new test_conv_2d_dw({512, 512, 256, 1}, {3, 3, 1, 256}, 1, 1, 1, false));
     test_cases.emplace_back(new test_conv_2d_dw({512, 512, 256, 1}, {3, 3, 1, 256}, 1, 1, 1, true));
 
+    test_cases.emplace_back(new test_conv_transpose_2d({256, 256, 256, 1}, {3, 3, 16, 256}, 1));
+
+    test_cases.emplace_back(new test_mean(GGML_TYPE_F32, {256, 256, 3, 1}));
+
     return test_cases;
 }
 
-static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op_name, const char * params_filter) {
+static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op_name, const char * params_filter,
+                         printer * output_printer) {
     auto filter_test_cases = [](std::vector<std::unique_ptr<test_case>> & test_cases, const char * params_filter) {
         if (params_filter == nullptr) {
             return;
@@ -4632,17 +5545,19 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op
         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");
+            test_operation_info info("", "", "CPU");
+            info.set_error("backend", "Failed to initialize CPU backend");
+            output_printer->print_operation(info);
             return false;
         }
 
         size_t n_ok = 0;
         for (auto & test : test_cases) {
-            if (test->eval(backend, backend_cpu, op_name)) {
+            if (test->eval(backend, backend_cpu, op_name, output_printer)) {
                 n_ok++;
             }
         }
-        printf("  %zu/%zu tests passed\n", n_ok, test_cases.size());
+        output_printer->print_summary(test_summary_info(n_ok, test_cases.size(), false));
 
         ggml_backend_free(backend_cpu);
 
@@ -4654,11 +5569,11 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op
         filter_test_cases(test_cases, params_filter);
         size_t n_ok = 0;
         for (auto & test : test_cases) {
-            if (test->eval_grad(backend, op_name)) {
+            if (test->eval_grad(backend, op_name, output_printer)) {
                 n_ok++;
             }
         }
-        printf("  %zu/%zu tests passed\n", n_ok, test_cases.size());
+        output_printer->print_summary(test_summary_info(n_ok, test_cases.size(), false));
 
         return n_ok == test_cases.size();
     }
@@ -4667,7 +5582,7 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op
         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);
+            test->eval_perf(backend, op_name, output_printer);
         }
         return true;
     }
@@ -4676,16 +5591,18 @@ 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>] [-p <params regex>]\n", argv[0]);
+    printf("Usage: %s [mode] [-o <op>] [-b <backend>] [-p <params regex>] [--output <console|sql>]\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");
     printf("      - perf (performance evaluation)\n");
     printf("    op names for -o are as given by ggml_op_desc() (e.g. ADD, MUL_MAT, etc)\n");
+    printf("    --output specifies output format (default: console)\n");
 }
 
 int main(int argc, char ** argv) {
     test_mode mode = MODE_TEST;
+    output_formats output_format = CONSOLE;
     const char * op_name_filter = nullptr;
     const char * backend_filter = nullptr;
     const char * params_filter = nullptr;
@@ -4718,6 +5635,16 @@ int main(int argc, char ** argv) {
                 usage(argv);
                 return 1;
             }
+        } else if (strcmp(argv[i], "--output") == 0) {
+            if (i + 1 < argc) {
+                if (!output_format_from_str(argv[++i], output_format)) {
+                    usage(argv);
+                    return 1;
+                }
+            } else {
+                usage(argv);
+                return 1;
+            }
         } else {
             usage(argv);
             return 1;
@@ -4727,23 +5654,29 @@ int main(int argc, char ** argv) {
     // load and enumerate backends
     ggml_backend_load_all();
 
-    printf("Testing %zu devices\n\n", ggml_backend_dev_count());
+    // Create printer for output format
+    std::unique_ptr<printer> output_printer = create_printer(output_format);
+    if (output_printer) {
+        output_printer->print_header();
+    }
+
+    output_printer->print_testing_start(testing_start_info(ggml_backend_dev_count()));
 
     size_t n_ok = 0;
 
     for (size_t i = 0; i < ggml_backend_dev_count(); i++) {
         ggml_backend_dev_t dev = ggml_backend_dev_get(i);
 
-        printf("Backend %zu/%zu: %s\n", i + 1, ggml_backend_dev_count(), ggml_backend_dev_name(dev));
-
         if (backend_filter != NULL && strcmp(backend_filter, ggml_backend_dev_name(dev)) != 0) {
-            printf("  Skipping\n");
+            output_printer->print_backend_init(
+                backend_init_info(i, ggml_backend_dev_count(), ggml_backend_dev_name(dev), true, "Skipping"));
             n_ok++;
             continue;
         }
 
         if (backend_filter == NULL && ggml_backend_dev_type(dev) == GGML_BACKEND_DEVICE_TYPE_CPU && mode != MODE_GRAD) {
-            printf("  Skipping CPU backend\n");
+            output_printer->print_backend_init(backend_init_info(
+                i, ggml_backend_dev_count(), ggml_backend_dev_name(dev), true, "Skipping CPU backend"));
             n_ok++;
             continue;
         }
@@ -4758,36 +5691,35 @@ int main(int argc, char ** argv) {
             ggml_backend_set_n_threads_fn(backend, std::thread::hardware_concurrency());
         }
 
-        printf("  Device description: %s\n", ggml_backend_dev_description(dev));
-        size_t free, total; // NOLINT
+        size_t free, total;  // NOLINT
         ggml_backend_dev_memory(dev, &free, &total);
-        printf("  Device memory: %zu MB (%zu MB free)\n", total / 1024 / 1024, free / 1024 / 1024);
-        printf("\n");
+        output_printer->print_backend_init(backend_init_info(i, ggml_backend_dev_count(), ggml_backend_dev_name(dev),
+                                                             false, "", ggml_backend_dev_description(dev),
+                                                             total / 1024 / 1024, free / 1024 / 1024, true));
 
-        bool ok = test_backend(backend, mode, op_name_filter, params_filter);
+        bool ok = test_backend(backend, mode, op_name_filter, params_filter, output_printer.get());
 
-        printf("  Backend %s: ", ggml_backend_name(backend));
         if (ok) {
-            printf("\033[1;32mOK\033[0m\n");
             n_ok++;
-        } else {
-            printf("\033[1;31mFAIL\033[0m\n");
         }
-
-        printf("\n");
+        output_printer->print_backend_status(
+            backend_status_info(ggml_backend_name(backend), ok ? test_status_t::OK : test_status_t::FAIL));
 
         ggml_backend_free(backend);
     }
 
     ggml_quantize_free();
 
-    printf("%zu/%zu backends passed\n", n_ok, ggml_backend_dev_count());
+    if (output_printer) {
+        output_printer->print_footer();
+    }
+
+    output_printer->print_overall_summary(
+        overall_summary_info(n_ok, ggml_backend_dev_count(), n_ok == ggml_backend_dev_count()));
 
     if (n_ok != ggml_backend_dev_count()) {
-        printf("\033[1;31mFAIL\033[0m\n");
         return 1;
     }
 
-    printf("\033[1;32mOK\033[0m\n");
     return 0;
 }
diff --git a/tests/test-c.c b/tests/test-c.c
index 95ba73df3..a05071080 100644
--- a/tests/test-c.c
+++ b/tests/test-c.c
@@ -1,7 +1,3 @@
 #include "llama.h"
 
-#ifdef GGML_USE_KOMPUTE
-#include "ggml-kompute.h"
-#endif
-
 int main(void) {}
diff --git a/tests/test-chat.cpp b/tests/test-chat.cpp
index 1c9807921..6ebf1464d 100644
--- a/tests/test-chat.cpp
+++ b/tests/test-chat.cpp
@@ -7,6 +7,8 @@
 //
 #include "chat.h"
 
+#include "log.h"
+
 #include "../src/unicode.h"
 #include "../src/llama-grammar.h"
 
@@ -19,8 +21,8 @@
 using json = nlohmann::ordered_json;
 
 static std::ostream & operator<<(std::ostream & os, const common_chat_msg_diff & diff) {
-    // os << "reasoning_content_delta: " << diff.reasoning_content_delta << '\n';
     os << "{ content_delta: " << diff.content_delta << "; ";
+    os << "reasoning_content_delta: " << diff.reasoning_content_delta << "; ";
     if (diff.tool_call_index != std::string::npos) {
         os << "tool_call_index: " << diff.tool_call_index << "; ";
         os << "tool_call_delta.name: " << diff.tool_call_delta.name << "; ";
@@ -1428,6 +1430,8 @@ static void test_msg_diffs_compute() {
 }
 
 int main(int argc, char ** argv) {
+    common_log_set_verbosity_thold(999);
+
     // try {
 #ifndef _WIN32
         if (argc > 1) {
diff --git a/tests/test-lora-conversion-inference.sh b/tests/test-lora-conversion-inference.sh
index 1d1f4886c..0255494b8 100755
--- a/tests/test-lora-conversion-inference.sh
+++ b/tests/test-lora-conversion-inference.sh
@@ -1,4 +1,4 @@
-#!/bin/bash
+#!/usr/bin/env bash
 set -e
 
 # Array of models to iterate over
diff --git a/tests/test-thread-safety.cpp b/tests/test-thread-safety.cpp
new file mode 100644
index 000000000..d525b7430
--- /dev/null
+++ b/tests/test-thread-safety.cpp
@@ -0,0 +1,152 @@
+// thread safety test
+// - Loads a copy of the same model on each GPU, plus a copy on the CPU
+// - Creates n_parallel (--parallel) contexts per model
+// - Runs inference in parallel on each context
+
+#include <thread>
+#include <vector>
+#include <atomic>
+#include "llama.h"
+#include "arg.h"
+#include "common.h"
+#include "log.h"
+#include "sampling.h"
+
+int main(int argc, char ** argv) {
+    common_params params;
+
+    if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_COMMON)) {
+        return 1;
+    }
+
+    common_init();
+
+    llama_backend_init();
+    llama_numa_init(params.numa);
+
+    LOG_INF("%s\n", common_params_get_system_info(params).c_str());
+
+    //llama_log_set([](ggml_log_level level, const char * text, void * /*user_data*/) {
+    //    if (level == GGML_LOG_LEVEL_ERROR) {
+    //        common_log_add(common_log_main(), level, "%s", text);
+    //    }
+    //}, NULL);
+
+    auto cparams = common_context_params_to_llama(params);
+
+    int dev_count = ggml_backend_dev_count();
+    int gpu_dev_count = 0;
+    for (int i = 0; i < dev_count; ++i) {
+        auto * dev = ggml_backend_dev_get(i);
+        if (dev && ggml_backend_dev_type(dev) == GGML_BACKEND_DEVICE_TYPE_GPU) {
+            gpu_dev_count++;
+        }
+    }
+    const int num_models = gpu_dev_count + 1 + 1; // GPUs + 1 CPU model + 1 layer split
+    //const int num_models = std::max(1, gpu_dev_count);
+    const int num_contexts = std::max(1, params.n_parallel);
+
+    std::vector<llama_model_ptr> models;
+    std::vector<std::thread> threads;
+    std::atomic<bool> failed = false;
+
+    for (int m = 0; m < num_models; ++m) {
+        auto mparams = common_model_params_to_llama(params);
+
+        if (m < gpu_dev_count) {
+            mparams.split_mode = LLAMA_SPLIT_MODE_NONE;
+            mparams.main_gpu = m;
+        } else if (m == gpu_dev_count) {
+            mparams.split_mode = LLAMA_SPLIT_MODE_NONE;
+            mparams.main_gpu = -1; // CPU model
+        } else {
+            mparams.split_mode = LLAMA_SPLIT_MODE_LAYER;;
+        }
+
+        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.path.c_str());
+            return 1;
+        }
+
+        models.emplace_back(model);
+    }
+
+    for  (int m = 0; m < num_models; ++m) {
+        auto * model = models[m].get();
+        for (int c = 0; c < num_contexts; ++c) {
+            threads.emplace_back([&, m, c, model]() {
+                LOG_INF("Creating context %d/%d for model %d/%d\n", c + 1, num_contexts, m + 1, num_models);
+
+                llama_context_ptr ctx { llama_init_from_model(model, cparams) };
+                if (ctx == NULL) {
+                    LOG_ERR("failed to create context\n");
+                    failed.store(true);
+                    return;
+                }
+
+                std::unique_ptr<common_sampler, decltype(&common_sampler_free)> sampler { common_sampler_init(model, params.sampling), common_sampler_free };
+                if (sampler == NULL) {
+                    LOG_ERR("failed to create sampler\n");
+                    failed.store(true);
+                    return;
+                }
+
+                llama_batch batch = {};
+                {
+                    auto prompt = common_tokenize(ctx.get(), params.prompt, true);
+                    if (prompt.empty()) {
+                        LOG_ERR("failed to tokenize prompt\n");
+                        failed.store(true);
+                        return;
+                    }
+                    batch = llama_batch_get_one(prompt.data(), prompt.size());
+                    if (llama_decode(ctx.get(), batch)) {
+                        LOG_ERR("failed to decode prompt\n");
+                        failed.store(true);
+                        return;
+                    }
+                }
+
+                const auto * vocab = llama_model_get_vocab(model);
+                std::string result = params.prompt;
+
+                for (int i = 0; i < params.n_predict; i++) {
+                    llama_token token;
+                    if (batch.n_tokens > 0) {
+                        token = common_sampler_sample(sampler.get(), ctx.get(), batch.n_tokens - 1);
+                    } else {
+                        token = llama_vocab_bos(vocab);
+                    }
+
+                    result += common_token_to_piece(ctx.get(), token);
+
+                    if (llama_vocab_is_eog(vocab, token)) {
+                        break;
+                    }
+
+                    batch = llama_batch_get_one(&token, 1);
+                    if (llama_decode(ctx.get(), batch)) {
+                        LOG_ERR("Model %d/%d, Context %d/%d: failed to decode\n", m + 1, num_models, c + 1, num_contexts);
+                        failed.store(true);
+                        return;
+                    }
+                }
+
+                LOG_INF("Model %d/%d, Context %d/%d: %s\n\n", m + 1, num_models, c + 1, num_contexts, result.c_str());
+            });
+        }
+    }
+
+    for (auto & thread : threads) {
+        thread.join();
+    }
+
+    if (failed) {
+        LOG_ERR("One or more threads failed.\n");
+        return 1;
+    }
+
+    LOG_INF("All threads finished without errors.\n");
+    return 0;
+}
diff --git a/tests/test-tokenizer-0.sh b/tests/test-tokenizer-0.sh
index 4d2b83655..7ef009dc9 100755
--- a/tests/test-tokenizer-0.sh
+++ b/tests/test-tokenizer-0.sh
@@ -1,4 +1,4 @@
-#!/bin/bash
+#!/usr/bin/env bash
 #
 # Usage:
 #
diff --git a/tests/test-tokenizers-repo.sh b/tests/test-tokenizers-repo.sh
new file mode 100755
index 000000000..1158aebae
--- /dev/null
+++ b/tests/test-tokenizers-repo.sh
@@ -0,0 +1,36 @@
+#!/usr/bin/env bash
+
+if [ $# -lt 2 ]; then
+    printf "Usage: $0 <git-repo> <target-folder> [<test-exe>]\n"
+    exit 1
+fi
+
+if [ $# -eq 3 ]; then
+    toktest=$3
+else
+    toktest="./test-tokenizer-0"
+fi
+
+if [ ! -x $toktest ]; then
+    printf "Test executable \"$toktest\" not found!\n"
+    exit 1
+fi
+
+repo=$1
+folder=$2
+
+if [ -d $folder ] && [ -d $folder/.git ]; then
+    (cd $folder; git pull)
+else
+    git clone $repo $folder
+fi
+
+shopt -s globstar
+for gguf in $folder/**/*.gguf; do
+    if [ -f $gguf.inp ] && [ -f $gguf.out ]; then
+        $toktest $gguf
+    else
+        printf "Found \"$gguf\" without matching inp/out files, ignoring...\n"
+    fi
+done
+
diff --git a/tools/batched-bench/batched-bench.cpp b/tools/batched-bench/batched-bench.cpp
index 119df471b..a0a2e5ac5 100644
--- a/tools/batched-bench/batched-bench.cpp
+++ b/tools/batched-bench/batched-bench.cpp
@@ -57,6 +57,8 @@ int main(int argc, char ** argv) {
         return 1;
     }
 
+    auto * mem = llama_get_memory(ctx);
+
     const int32_t n_kv_max = llama_n_ctx(ctx);
 
     llama_batch batch = llama_batch_init(n_kv_max, 0, 1);
@@ -132,7 +134,7 @@ int main(int argc, char ** argv) {
 
                 const auto t_pp_start = ggml_time_us();
 
-                llama_kv_self_clear(ctx);
+                llama_memory_clear(mem, false);
 
                 if (!decode_helper(ctx, batch, ctx_params.n_batch)) {
                     LOG_ERR("%s: llama_decode() failed\n", __func__);
@@ -141,7 +143,7 @@ int main(int argc, char ** argv) {
 
                 if (is_pp_shared) {
                     for (int32_t i = 1; i < pl; ++i) {
-                        llama_kv_self_seq_cp(ctx, 0, i, -1, -1);
+                        llama_memory_seq_cp(mem, 0, i, -1, -1);
                     }
                 }
 
diff --git a/tools/cvector-generator/cvector-generator.cpp b/tools/cvector-generator/cvector-generator.cpp
index 2a9071550..d2d97e05c 100644
--- a/tools/cvector-generator/cvector-generator.cpp
+++ b/tools/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<llama_token> & tokens) {
-    llama_kv_self_clear(ctx);
+    llama_memory_clear(llama_get_memory(ctx), true);
     if (llama_decode(ctx, llama_batch_get_one(tokens.data(), tokens.size()))) {
         fprintf(stderr, "%s : failed to eval\n", __func__);
         return false;
diff --git a/tools/gguf-split/tests.sh b/tools/gguf-split/tests.sh
index 05a932227..c9ad85da0 100755
--- a/tools/gguf-split/tests.sh
+++ b/tools/gguf-split/tests.sh
@@ -1,4 +1,4 @@
-#!/bin/bash
+#!/usr/bin/env bash
 
 set -eu
 
diff --git a/tools/imatrix/imatrix.cpp b/tools/imatrix/imatrix.cpp
index 81d0404d6..daad44e59 100644
--- a/tools/imatrix/imatrix.cpp
+++ b/tools/imatrix/imatrix.cpp
@@ -498,7 +498,7 @@ static bool compute_imatrix(llama_context * ctx, const common_params & params) {
         const auto t_start = std::chrono::high_resolution_clock::now();
 
         // clear the KV cache
-        llama_kv_self_clear(ctx);
+        llama_memory_clear(llama_get_memory(ctx), true);
 
         llama_batch batch = llama_batch_init(n_batch, 0, 1);
 
diff --git a/tools/llama-bench/llama-bench.cpp b/tools/llama-bench/llama-bench.cpp
index 803630d26..b80e984d0 100644
--- a/tools/llama-bench/llama-bench.cpp
+++ b/tools/llama-bench/llama-bench.cpp
@@ -267,6 +267,7 @@ struct cmd_params {
     int                              delay;
     bool                             verbose;
     bool                             progress;
+    bool                             no_warmup;
     output_formats                   output_format;
     output_formats                   output_format_stderr;
 };
@@ -303,6 +304,7 @@ static const cmd_params cmd_params_defaults = {
     /* delay                */ 0,
     /* verbose              */ false,
     /* progress             */ false,
+    /* no_warmup            */ false,
     /* output_format        */ MARKDOWN,
     /* output_format_stderr */ NONE,
 };
@@ -325,6 +327,7 @@ static void print_usage(int /* argc */, char ** argv) {
            output_format_str(cmd_params_defaults.output_format_stderr));
     printf("  -v, --verbose                             verbose output\n");
     printf("  --progress                                print test progress indicators\n");
+    printf("  --no-warmup                               skip warmup runs before benchmarking\n");
     printf("\n");
     printf("test parameters:\n");
     printf("  -m, --model <filename>                    (default: %s)\n", join(cmd_params_defaults.model, ",").c_str());
@@ -425,6 +428,7 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
     params.prio                 = cmd_params_defaults.prio;
     params.delay                = cmd_params_defaults.delay;
     params.progress             = cmd_params_defaults.progress;
+    params.no_warmup            = cmd_params_defaults.no_warmup;
 
     for (int i = 1; i < argc; i++) {
         arg = argv[i];
@@ -798,6 +802,8 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
                 params.verbose = true;
             } else if (arg == "--progress") {
                 params.progress = true;
+            } else if (arg == "--no-warmup") {
+                params.no_warmup = true;
             } else {
                 invalid_param = true;
                 break;
@@ -1900,7 +1906,7 @@ int main(int argc, char ** argv) {
 
         test t(inst, lmodel, ctx);
 
-        llama_kv_self_clear(ctx);
+        llama_memory_clear(llama_get_memory(ctx), false);
 
         // cool off before the test
         if (params.delay) {
@@ -1925,30 +1931,32 @@ int main(int argc, char ** argv) {
         llama_attach_threadpool(ctx, threadpool, NULL);
 
         // warmup run
-        if (t.n_prompt > 0) {
-            if (params.progress) {
-                fprintf(stderr, "llama-bench: benchmark %d/%zu: warmup prompt run\n", params_idx, params_count);
+        if (!params.no_warmup) {
+            if (t.n_prompt > 0) {
+                if (params.progress) {
+                    fprintf(stderr, "llama-bench: benchmark %d/%zu: warmup prompt run\n", params_idx, params_count);
+                }
+                //test_prompt(ctx, std::min(t.n_batch, std::min(t.n_prompt, 32)), 0, t.n_batch, t.n_threads);
+                bool res = test_prompt(ctx, t.n_prompt, t.n_batch, t.n_threads);
+                if (!res) {
+                    fprintf(stderr, "%s: error: failed to run prompt warmup\n", __func__);
+                    exit(1);
+                }
             }
-            //test_prompt(ctx, std::min(t.n_batch, std::min(t.n_prompt, 32)), 0, t.n_batch, t.n_threads);
-            bool res = test_prompt(ctx, t.n_prompt, t.n_batch, t.n_threads);
-            if (!res) {
-                fprintf(stderr, "%s: error: failed to run prompt warmup\n", __func__);
-                exit(1);
-            }
-        }
-        if (t.n_gen > 0) {
-            if (params.progress) {
-                fprintf(stderr, "llama-bench: benchmark %d/%zu: warmup generation run\n", params_idx, params_count);
-            }
-            bool res = test_gen(ctx, 1, t.n_threads);
-            if (!res) {
-                fprintf(stderr, "%s: error: failed to run gen warmup\n", __func__);
-                exit(1);
+            if (t.n_gen > 0) {
+                if (params.progress) {
+                    fprintf(stderr, "llama-bench: benchmark %d/%zu: warmup generation run\n", params_idx, params_count);
+                }
+                bool res = test_gen(ctx, 1, t.n_threads);
+                if (!res) {
+                    fprintf(stderr, "%s: error: failed to run gen warmup\n", __func__);
+                    exit(1);
+                }
             }
         }
 
         for (int i = 0; i < params.reps; i++) {
-            llama_kv_self_clear(ctx);
+            llama_memory_clear(llama_get_memory(ctx), false);
 
             if (t.n_depth > 0) {
                 if (params.progress) {
diff --git a/tools/main/main.cpp b/tools/main/main.cpp
index 1bd2be2d9..516bf0965 100644
--- a/tools/main/main.cpp
+++ b/tools/main/main.cpp
@@ -147,6 +147,8 @@ int main(int argc, char ** argv) {
         return 1;
     }
 
+    auto * mem = llama_get_memory(ctx);
+
     const llama_vocab * vocab = llama_model_get_vocab(model);
     auto chat_templates = common_chat_templates_init(model, params.chat_template);
 
@@ -290,6 +292,7 @@ int main(int argc, char ** argv) {
 
             if (!params.system_prompt.empty() || !params.prompt.empty()) {
                 common_chat_templates_inputs inputs;
+                inputs.use_jinja = g_params->use_jinja;
                 inputs.messages = chat_msgs;
                 inputs.add_generation_prompt = !params.prompt.empty();
 
@@ -351,7 +354,7 @@ int main(int argc, char ** argv) {
         }
 
         // remove any "future" tokens that we might have inherited from the previous session
-        llama_kv_self_seq_rm(ctx, -1, n_matching_session_tokens, -1);
+        llama_memory_seq_rm(mem, -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",
@@ -599,8 +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_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);
+                    llama_memory_seq_rm (mem, 0, params.n_keep            , params.n_keep + n_discard);
+                    llama_memory_seq_add(mem, 0, params.n_keep + n_discard, n_past, -n_discard);
 
                     n_past -= n_discard;
 
@@ -623,9 +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_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);
+                    llama_memory_seq_add(mem, 0, ga_i,                n_past,              ib*bd);
+                    llama_memory_seq_div(mem, 0, ga_i + ib*bd,        ga_i + ib*bd + ga_w, ga_n);
+                    llama_memory_seq_add(mem, 0, ga_i + ib*bd + ga_w, n_past + ib*bd,      dd);
 
                     n_past -= bd;
 
@@ -914,10 +917,19 @@ int main(int argc, char ** argv) {
                     embd_inp.insert(embd_inp.end(), line_inp.begin(), line_inp.end());
                     embd_inp.insert(embd_inp.end(), line_sfx.begin(), line_sfx.end());
 
+                    if (params.verbose_prompt) {
+                        LOG_INF("%s: number of tokens in prompt = %zu\n", __func__, embd_inp.size() - original_size);
+                    }
+
                     for (size_t i = original_size; i < embd_inp.size(); ++i) {
                         const llama_token token = embd_inp[i];
+                        const std::string token_str = common_token_to_piece(ctx, token);
                         output_tokens.push_back(token);
-                        output_ss << common_token_to_piece(ctx, token);
+                        output_ss << token_str;
+
+                        if (params.verbose_prompt) {
+                            LOG_INF("%6d -> '%s'\n", token, token_str.c_str());
+                        }
                     }
 
                     // reset assistant message
diff --git a/tools/mtmd/clip.cpp b/tools/mtmd/clip.cpp
index c25bacc17..9146c9e9c 100644
--- a/tools/mtmd/clip.cpp
+++ b/tools/mtmd/clip.cpp
@@ -187,7 +187,7 @@ struct clip_hparams {
     float eps = 1e-6;
     float rope_theta = 0.0;
 
-    std::vector<int32_t> image_grid_pinpoints;
+    std::vector<clip_image_size> image_res_candidates; // for llava-uhd style models
     int32_t image_crop_resolution;
     std::unordered_set<int32_t> vision_feature_layer;
     int32_t attn_window_size = 0;
@@ -1405,8 +1405,7 @@ struct clip_graph {
                 ggml_tensor * x = embeddings;
                 embeddings = ggml_mul_mat(ctx0, model.mm_model_mlp_2_w, embeddings);
                 x = ggml_mul_mat(ctx0, model.mm_model_mlp_1_w,x);
-                embeddings = ggml_silu_inplace(ctx0, embeddings);
-                embeddings = ggml_mul(ctx0, embeddings,x);
+                embeddings = ggml_swiglu_split(ctx0, embeddings, x);
                 embeddings = ggml_mul_mat(ctx0, model.mm_model_mlp_3_w, embeddings);
             }
             // arrangement of BOI/EOI token embeddings
@@ -1502,15 +1501,8 @@ struct clip_graph {
                 cur = ggml_mul_mat(ctx0, model.mm_1_w, cur);
 
                 // swiglu
-                {
-                    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)));
-
-                    // see SwiGLU in ultravox_model.py, the second half passed through is silu, not the first half
-                    x1 = ggml_silu(ctx0, x1);
-                    cur = ggml_mul(ctx0, x0, x1);
-                }
+                // see SwiGLU in ultravox_model.py, the second half passed through is silu, not the first half
+                cur = ggml_swiglu_swapped(ctx0, cur);
 
                 // mid-norm
                 cur = ggml_rms_norm(ctx0, cur, 1e-6);
@@ -1769,35 +1761,42 @@ private:
             cur = tmp;
         }
 
+        // we only support parallel ffn for now
         switch (type_op) {
             case FFN_SILU:
-                {
+                if (gate) {
+                    cur = ggml_swiglu_split(ctx0, cur, tmp);
+                    cb(cur, "ffn_swiglu", il);
+                } else {
                     cur = ggml_silu(ctx0, cur);
                     cb(cur, "ffn_silu", il);
                 } break;
             case FFN_GELU:
-                {
+                if (gate) {
+                    cur = ggml_geglu_split(ctx0, cur, tmp);
+                    cb(cur, "ffn_geglu", il);
+                } else {
                     cur = ggml_gelu(ctx0, cur);
                     cb(cur, "ffn_gelu", il);
                 } break;
             case FFN_GELU_ERF:
-                {
+                if (gate) {
+                    cur = ggml_geglu_erf_split(ctx0, cur, tmp);
+                    cb(cur, "ffn_geglu_erf", il);
+                } else {
                     cur = ggml_gelu_erf(ctx0, cur);
-                    cb(cur, "ggml_gelu_erf", il);
+                    cb(cur, "ffn_gelu_erf", il);
                 } break;
             case FFN_GELU_QUICK:
-                {
+                if (gate) {
+                    cur = ggml_geglu_quick_split(ctx0, cur, tmp);
+                    cb(cur, "ffn_geglu_quick", il);
+                } else {
                     cur = ggml_gelu_quick(ctx0, cur);
-                    cb(cur, "ffn_relu", il);
+                    cb(cur, "ffn_gelu_quick", il);
                 } break;
         }
 
-        // we only support parallel ffn for now
-        if (gate) {
-            cur = ggml_mul(ctx0, cur, tmp);
-            cb(cur, "ffn_gate_par", il);
-        }
-
         if (down) {
             cur = ggml_mul_mat(ctx0, down, cur);
         }
@@ -2109,8 +2108,7 @@ struct clip_model_loader {
             if (is_vision) {
                 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);
+                get_u32(KEY_IMAGE_CROP_RESOLUTION, hparams.image_crop_resolution, false);
                 get_i32(KEY_MINICPMV_VERSION, hparams.minicpmv_version, false); // legacy
 
             } else if (is_audio) {
@@ -2120,6 +2118,20 @@ struct clip_model_loader {
                 GGML_ASSERT(false && "unknown modality");
             }
 
+            // for pinpoints, we need to convert it into a list of resolution candidates
+            {
+                std::vector<int> pinpoints;
+                get_arr_int(KEY_IMAGE_GRID_PINPOINTS, pinpoints, false);
+                if (!pinpoints.empty()) {
+                    for (size_t i = 0; i < pinpoints.size(); i += 2) {
+                        hparams.image_res_candidates.push_back({
+                            pinpoints[i],
+                            pinpoints[i+1],
+                        });
+                    }
+                }
+            }
+
             // default warmup value
             hparams.warmup_image_size = hparams.image_size;
 
@@ -2198,6 +2210,9 @@ struct clip_model_loader {
                     {
                         hparams.rope_theta = 10000.0f;
                         hparams.warmup_image_size = hparams.patch_size * 8;
+                        // Mistral Small 2506 needs 1024x1024 image size cap to prevent OOM
+                        // ref: https://github.com/ggml-org/llama.cpp/issues/14310
+                        hparams.image_size = 1024;
                         get_u32(KEY_SPATIAL_MERGE_SIZE, hparams.spatial_merge_size, false);
                     } break;
                 case PROJECTOR_TYPE_GEMMA3:
@@ -2231,16 +2246,7 @@ struct clip_model_loader {
                     {
                         hparams.rope_theta = 10000.0f;
                         get_u32(KEY_PROJ_SCALE_FACTOR, hparams.proj_scale_factor);
-
-                        // borrowed from llava-1.6
-                        const int isize = hparams.image_size;
-                        hparams.image_grid_pinpoints = {
-                            isize,   isize*2, // 336, 672
-                            isize*2, isize,   // 672, 336
-                            isize*2, isize*2, // 672, 672
-                            isize*3, isize,   // 1008, 336
-                            isize,   isize*3, // 336, 1008
-                        };
+                        set_llava_uhd_res_candidates(model, 3);
                     } break;
                 case PROJECTOR_TYPE_ULTRAVOX:
                 case PROJECTOR_TYPE_QWEN2A:
@@ -2674,6 +2680,21 @@ struct clip_model_loader {
             output[i] = values[i];
         }
     }
+
+    void set_llava_uhd_res_candidates(clip_model & model, const int max_patches_per_side) {
+        auto & hparams = model.hparams;
+        for (int x = 1; x <= max_patches_per_side; x++) {
+            for (int y = 1; y <= max_patches_per_side; y++) {
+                if (x == 1 && y == 1) {
+                    continue; // skip the first point
+                }
+                hparams.image_res_candidates.push_back(clip_image_size{
+                    x*hparams.image_size,
+                    y*hparams.image_size,
+                });
+            }
+        }
+    }
 };
 
 struct clip_init_result clip_init(const char * fname, struct clip_context_params ctx_params) {
@@ -3028,36 +3049,41 @@ struct llava_uhd {
         bool padding_refined = false;  // if true, refine image will be padded to the grid size (e.g. llava-1.6)
     };
 
-    static int get_max_slices(struct clip_ctx * ctx) {
-        if (clip_is_minicpmv(ctx)) {
-            return 9;
-        }
-        return 0;
-    }
-
     static slice_instructions get_slice_instructions(struct clip_ctx * ctx, const clip_image_size & original_size) {
         slice_instructions res;
         const int patch_size      = clip_get_patch_size(ctx);
         const int slice_size      = clip_get_image_size(ctx);
-        const int max_slice_nums  = get_max_slices(ctx);
         const int original_width  = original_size.width;
         const int original_height = original_size.height;
-        const float log_ratio = log((float)original_width / original_height);
-        const float ratio = (float)original_width * original_height / (slice_size * slice_size);
-        const int multiple = fmin(ceil(ratio), max_slice_nums);
-        const bool has_slices = (multiple > 1);
-        const bool has_pinpoints = !ctx->model.hparams.image_grid_pinpoints.empty();
+
+        const bool has_slices    = original_size.width > slice_size || original_size.height > slice_size;
+        const bool has_pinpoints = !ctx->model.hparams.image_res_candidates.empty();
+
+        if (!has_slices) {
+            // skip slicing logic
+            res.overview_size = clip_image_size{slice_size, slice_size};
+            res.refined_size  = clip_image_size{0, 0};
+            res.grid_size     = clip_image_size{0, 0};
+
+            return res;
+        }
 
         if (has_pinpoints) {
             // has pinpoints, use them to calculate the grid size (e.g. llava-1.6)
             auto refine_size = llava_uhd::select_best_resolution(
-                ctx->model.hparams.image_grid_pinpoints,
-                original_size);
+                original_size,
+                ctx->model.hparams.image_res_candidates);
             res.overview_size   = clip_image_size{slice_size, slice_size};
             res.refined_size    = refine_size;
             res.grid_size       = clip_image_size{0, 0};
             res.padding_refined = true;
 
+            LOG_DBG("%s: using pinpoints for slicing\n", __func__);
+            LOG_DBG("%s: original size: %d x %d, overview size: %d x %d, refined size: %d x %d\n",
+                    __func__, original_width, original_height,
+                    res.overview_size.width, res.overview_size.height,
+                    res.refined_size.width,  res.refined_size.height);
+
             for (int y = 0; y < refine_size.height; y += slice_size) {
                 for (int x = 0; x < refine_size.width; x += slice_size) {
                     slice_coordinates slice;
@@ -3066,13 +3092,16 @@ struct llava_uhd {
                     slice.size.width  = std::min(slice_size, refine_size.width  - x);
                     slice.size.height = std::min(slice_size, refine_size.height - y);
                     res.slices.push_back(slice);
-                    if (x == 0) {
-                        res.grid_size.width++;
-                    }
+                    LOG_DBG("%s: slice %d: x=%d, y=%d, size=%dx%d\n",
+                            __func__, (int)res.slices.size() - 1,
+                            slice.x, slice.y, slice.size.width, slice.size.height);
                 }
-                res.grid_size.height++;
             }
 
+            res.grid_size.height = refine_size.height / slice_size;
+            res.grid_size.width  = refine_size.width  / slice_size;
+            LOG_DBG("%s: grid size: %d x %d\n", __func__, res.grid_size.width, res.grid_size.height);
+
             return res;
         }
 
@@ -3081,17 +3110,23 @@ struct llava_uhd {
         auto best_size    = get_best_resize(original_size, slice_size, patch_size, !has_slices);
         res.overview_size = best_size;
 
-        if (!has_slices) {
-            // skip slicing logic
-            res.refined_size = clip_image_size{0, 0};
-            res.grid_size    = clip_image_size{0, 0};
+        {
+            const int max_slice_nums = 9; // TODO: this is only used by minicpmv, maybe remove it
+            const float log_ratio = log((float)original_width / original_height);
+            const float ratio = (float)original_width * original_height / (slice_size * slice_size);
+            const int multiple = fmin(ceil(ratio), max_slice_nums);
 
-        } else {
             auto best_grid   = get_best_grid(max_slice_nums, multiple, log_ratio);
             auto refine_size = get_refine_size(original_size, best_grid, slice_size, patch_size, true);
             res.grid_size    = best_grid;
             res.refined_size = refine_size;
 
+            LOG_DBG("%s: original size: %d x %d, overview size: %d x %d, refined size: %d x %d, grid size: %d x %d\n",
+                    __func__, original_width, original_height,
+                    res.overview_size.width, res.overview_size.height,
+                    res.refined_size.width, res.refined_size.height,
+                    res.grid_size.width, res.grid_size.height);
+
             int width  = refine_size.width;
             int height = refine_size.height;
             int grid_x = int(width  / best_grid.width);
@@ -3108,7 +3143,9 @@ struct llava_uhd {
                     slice.size.width  = grid_x;
                     slice.size.height = grid_y;
                     res.slices.push_back(slice);
-                    // LOG_INF("slice %d: %d %d %d %d\n", ic, patches_i, patches_j, grid_x, grid_y);
+                    LOG_DBG("%s: slice %d: x=%d, y=%d, size=%dx%d\n",
+                            __func__, (int)res.slices.size() - 1,
+                            slice.x, slice.y, slice.size.width, slice.size.height);
                 }
             }
         }
@@ -3166,48 +3203,55 @@ private:
         return res;
     }
 
+    static clip_image_size resize_maintain_aspect_ratio(const clip_image_size & orig, const clip_image_size & target_max) {
+        float scale_width  = static_cast<float>(target_max.width)  / orig.width;
+        float scale_height = static_cast<float>(target_max.height) / orig.height;
+        float scale = std::min(scale_width, scale_height);
+        return clip_image_size{
+            static_cast<int>(orig.width  * scale),
+            static_cast<int>(orig.height * scale),
+        };
+    }
+
     /**
      * Selects the best resolution from a list of possible resolutions based on the original size.
      *
+     * For example, when given a list of resolutions:
+     *  - 100x100
+     *  - 200x100
+     *  - 100x200
+     *  - 200x200
+     *
+     * And an input image of size 111x200, then 100x200 is the best fit (least wasted resolution).
+     *
      * @param original_size The original size of the image
      * @param possible_resolutions A list of possible resolutions
      * @return The best fit resolution
      */
     static clip_image_size select_best_resolution(const clip_image_size & original_size, const std::vector<clip_image_size> & possible_resolutions) {
-        int original_width = original_size.width;
-        int original_height = original_size.height;
         clip_image_size best_fit;
+        int min_wasted_area = std::numeric_limits<int>::max();
         int max_effective_resolution = 0;
-        int min_wasted_resolution = std::numeric_limits<int>::max();
 
-        for (const auto & resolution : possible_resolutions) {
-            int width  = resolution.width;
-            int height = resolution.height;
-            float scale = std::min(static_cast<float>(width) / original_width, static_cast<float>(height) / original_height);
-            int downscaled_width  = static_cast<int>(original_width * scale);
-            int downscaled_height = static_cast<int>(original_height * scale);
-            int effective_resolution = std::min(downscaled_width * downscaled_height, original_width * original_height);
-            int wasted_resolution = (width * height) - effective_resolution;
-            // LOG_INF("resolution: %d %d, scale: %f, downscaled: %d %d, effective: %d, wasted: %d\n", width, height, scale, downscaled_width, downscaled_height, effective_resolution, wasted_resolution);
-            if (effective_resolution > max_effective_resolution || (effective_resolution == max_effective_resolution && wasted_resolution < min_wasted_resolution)) {
+        for (const clip_image_size & candidate : possible_resolutions) {
+            auto target_size = resize_maintain_aspect_ratio(original_size, candidate);
+            int effective_resolution = std::min(
+                target_size.width * target_size.height,
+                original_size.width * original_size.height);
+            int wasted_area = (candidate.width * candidate.height) - effective_resolution;
+
+            if (effective_resolution > max_effective_resolution || (effective_resolution == max_effective_resolution && wasted_area < min_wasted_area)) {
                 max_effective_resolution = effective_resolution;
-                min_wasted_resolution = wasted_resolution;
-                best_fit = resolution;
+                min_wasted_area = wasted_area;
+                best_fit = candidate;
             }
+
+            LOG_DBG("%s: candidate: %d x %d, target: %d x %d, wasted: %d, effective: %d\n", __func__, candidate.width, candidate.height, target_size.width, target_size.height, wasted_area, effective_resolution);
         }
 
         return best_fit;
     }
 
-    // used by llava 1.6 with custom list of pinpoints
-    static clip_image_size select_best_resolution(const std::vector<int32_t> & pinpoints, const clip_image_size & original_size) {
-        std::vector<clip_image_size> possible_resolutions; // TODO @ngxson : construct this inside hparams, not here
-        for (size_t i = 0; i < pinpoints.size(); i += 2) {
-            possible_resolutions.push_back(clip_image_size{pinpoints[i], pinpoints[i+1]});
-        }
-        return select_best_resolution(original_size, possible_resolutions);
-    }
-
     static int ensure_divide(int length, int patch_size) {
         return std::max(static_cast<int>(std::round(static_cast<float>(length) / patch_size) * patch_size), patch_size);
     }
@@ -3331,7 +3375,7 @@ bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, str
         return true;
 
     } else if (ctx->proj_type() == PROJECTOR_TYPE_LLAMA4) {
-        GGML_ASSERT(!params.image_grid_pinpoints.empty());
+        GGML_ASSERT(!params.image_res_candidates.empty());
         auto const inst = llava_uhd::get_slice_instructions(ctx, original_size);
         std::vector<clip_image_u8_ptr> imgs = llava_uhd::slice_image(img, inst);
 
@@ -3371,7 +3415,7 @@ bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, str
         res_imgs->entries.push_back(std::move(res));
         return true;
 
-    } else if (!params.image_grid_pinpoints.empty()) {
+    } else if (!params.image_res_candidates.empty()) {
         // "spatial_unpad" with "anyres" processing for llava-1.6
         auto const inst = llava_uhd::get_slice_instructions(ctx, original_size);
         std::vector<clip_image_u8_ptr> imgs = llava_uhd::slice_image(img, inst);
@@ -3431,17 +3475,6 @@ const char * clip_patch_merge_type(const struct clip_ctx * ctx) {
     return ctx->model.hparams.mm_patch_merge_type == PATCH_MERGE_SPATIAL_UNPAD ? "spatial_unpad" : "flat";
 }
 
-const int32_t * clip_image_grid(const struct clip_ctx * ctx) {
-    if (ctx->model.hparams.image_grid_pinpoints.size()) {
-        return &ctx->model.hparams.image_grid_pinpoints.front();
-    }
-    return nullptr;
-}
-
-size_t get_clip_image_grid_size(const struct clip_ctx * ctx) {
-    return ctx->model.hparams.image_grid_pinpoints.size();
-}
-
 int clip_n_output_tokens_x(const struct clip_ctx * ctx, struct clip_image_f32 * img) {
     const auto & params = ctx->model.hparams;
     const int n_total = clip_n_output_tokens(ctx, img);
diff --git a/tools/mtmd/clip.h b/tools/mtmd/clip.h
index cb2eb261f..08f3efb7b 100644
--- a/tools/mtmd/clip.h
+++ b/tools/mtmd/clip.h
@@ -46,9 +46,6 @@ int32_t clip_get_hidden_size(const struct clip_ctx * ctx);
 // TODO: should be enum, not string
 const char * clip_patch_merge_type(const struct clip_ctx * ctx);
 
-const int32_t * clip_image_grid(const struct clip_ctx * ctx);
-size_t get_clip_image_grid_size(const struct clip_ctx * ctx);
-
 int clip_n_output_tokens(const struct clip_ctx * ctx, struct clip_image_f32 * img);
 
 // for M-RoPE, this will be the number of token positions in X and Y directions
diff --git a/tools/mtmd/mtmd-cli.cpp b/tools/mtmd/mtmd-cli.cpp
index 508a64c58..599e682e0 100644
--- a/tools/mtmd/mtmd-cli.cpp
+++ b/tools/mtmd/mtmd-cli.cpp
@@ -70,6 +70,7 @@ struct mtmd_cli_context {
     llama_model       * model;
     llama_context     * lctx;
     const llama_vocab * vocab;
+    common_sampler    * smpl;
     llama_batch         batch;
     int                 n_batch;
 
@@ -89,8 +90,9 @@ struct mtmd_cli_context {
         model = llama_init.model.get();
         lctx = llama_init.context.get();
         vocab = llama_model_get_vocab(model);
+        smpl = common_sampler_init(model, params.sampling);
         n_threads = params.cpuparams.n_threads;
-        batch = llama_batch_init(params.n_batch, 0, 1);
+        batch = llama_batch_init(1, 0, 1); // batch for next token generation
         n_batch = params.n_batch;
 
         if (!model || !lctx) {
@@ -118,6 +120,11 @@ struct mtmd_cli_context {
         }
     }
 
+    ~mtmd_cli_context() {
+        llama_batch_free(batch);
+        common_sampler_free(smpl);
+    }
+
     void init_vision_context(common_params & params) {
         const char * clip_path = params.mmproj.path.c_str();
         mtmd_context_params mparams = mtmd_context_params_default();
@@ -153,7 +160,7 @@ struct mtmd_cli_context {
     }
 };
 
-static int generate_response(mtmd_cli_context & ctx, common_sampler * smpl, int n_predict) {
+static int generate_response(mtmd_cli_context & ctx, int n_predict) {
     llama_tokens generated_tokens;
     for (int i = 0; i < n_predict; i++) {
         if (i > n_predict || !g_is_generating || g_is_interrupted) {
@@ -161,9 +168,9 @@ static int generate_response(mtmd_cli_context & ctx, common_sampler * smpl, int
             break;
         }
 
-        llama_token token_id = common_sampler_sample(smpl, ctx.lctx, -1);
+        llama_token token_id = common_sampler_sample(ctx.smpl, ctx.lctx, -1);
         generated_tokens.push_back(token_id);
-        common_sampler_accept(smpl, token_id, true);
+        common_sampler_accept(ctx.smpl, token_id, true);
 
         if (llama_vocab_is_eog(ctx.vocab, token_id) || ctx.check_antiprompt(generated_tokens)) {
             LOG("\n");
@@ -261,7 +268,6 @@ int main(int argc, char ** argv) {
 
     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
@@ -300,7 +306,7 @@ int main(int argc, char ** argv) {
         if (eval_message(ctx, msg, true)) {
             return 1;
         }
-        if (!g_is_interrupted && generate_response(ctx, smpl, n_predict)) {
+        if (!g_is_interrupted && generate_response(ctx, n_predict)) {
             return 1;
         }
 
@@ -336,7 +342,7 @@ int main(int argc, char ** argv) {
             }
             if (line == "/clear") {
                 ctx.n_past = 0;
-                llama_kv_self_seq_rm(ctx.lctx, 0, 1, -1); // keep BOS
+                llama_memory_seq_rm(llama_get_memory(ctx.lctx), 0, 1, -1); // keep BOS
                 LOG("Chat history cleared\n\n");
                 continue;
             }
@@ -366,7 +372,7 @@ int main(int argc, char ** argv) {
                 return 1;
             }
             if (g_is_interrupted) break;
-            if (generate_response(ctx, smpl, n_predict)) {
+            if (generate_response(ctx, n_predict)) {
                 return 1;
             }
             content.clear();
diff --git a/tools/mtmd/mtmd-helper.cpp b/tools/mtmd/mtmd-helper.cpp
index 64f03fd1e..686f42f39 100644
--- a/tools/mtmd/mtmd-helper.cpp
+++ b/tools/mtmd/mtmd-helper.cpp
@@ -311,6 +311,7 @@ int32_t mtmd_helper_eval_chunk_single(mtmd_context * ctx,
         GGML_ABORT("chunk type not supported");
     }
 
+    llama_batch_free(text_batch);
     return 0;
 }
 
diff --git a/tools/mtmd/mtmd.cpp b/tools/mtmd/mtmd.cpp
index 8573f1143..e38297383 100644
--- a/tools/mtmd/mtmd.cpp
+++ b/tools/mtmd/mtmd.cpp
@@ -501,7 +501,10 @@ struct mtmd_tokenizer {
                 || ctx->slice_tmpl == MTMD_SLICE_TMPL_MINICPMV_2_6
                 || ctx->slice_tmpl == MTMD_SLICE_TMPL_LLAMA4
             ) {
+                const int n_col = batch_f32.grid_x;
+                const int n_row = batch_f32.grid_y;
                 // split batch into chunks of single images
+                // NOTE: batch_f32 will be invalidated after this call
                 auto chunks = split_batch_to_chunk(std::move(batch_f32), bitmap->id);
                 GGML_ASSERT(chunks.size() > 0);
 
@@ -521,8 +524,7 @@ struct mtmd_tokenizer {
 
                 // add slices (or tiles)
                 if (!chunks.empty()) {
-                    const int n_col = batch_f32.grid_x;
-                    const int n_row = batch_f32.grid_y;
+                    GGML_ASSERT((int)chunks.size() == n_row * n_col);
                     if (ctx->tok_slices_start != LLAMA_TOKEN_NULL) {
                         add_text({ctx->tok_slices_start});
                     }
diff --git a/tools/mtmd/tests.sh b/tools/mtmd/tests.sh
index aa0019893..b25024c2f 100755
--- a/tools/mtmd/tests.sh
+++ b/tools/mtmd/tests.sh
@@ -1,4 +1,4 @@
-#!/bin/bash
+#!/usr/bin/env bash
 
 # make sure we are in the right directory
 SCRIPT_DIR=$( cd -- "$( dirname -- "${BASH_SOURCE[0]}" )" &> /dev/null && pwd )
diff --git a/tools/perplexity/perplexity.cpp b/tools/perplexity/perplexity.cpp
index b5cdf5beb..189dcb3d7 100644
--- a/tools/perplexity/perplexity.cpp
+++ b/tools/perplexity/perplexity.cpp
@@ -361,7 +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_self_clear(ctx);
+        llama_memory_clear(llama_get_memory(ctx), true);
 
         llama_batch batch = llama_batch_init(n_batch, 0, 1);
 
@@ -547,7 +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_self_clear(ctx);
+        llama_memory_clear(llama_get_memory(ctx), true);
 
         for (int j = 0; j < num_batches; ++j) {
             const int batch_start = start + j * n_batch;
@@ -924,7 +924,7 @@ static void hellaswag_score(llama_context * ctx, const common_params & params) {
             return;
         }
 
-        llama_kv_self_clear(ctx);
+        llama_memory_clear(llama_get_memory(ctx), true);
 
         // decode all tasks [i0, i1)
         if (!decode_helper(ctx, batch, batch_logits, n_batch, n_vocab)) {
@@ -1217,7 +1217,7 @@ static void winogrande_score(llama_context * ctx, const common_params & params)
             return;
         }
 
-        llama_kv_self_clear(ctx);
+        llama_memory_clear(llama_get_memory(ctx), true);
 
         // decode all tasks [i0, i1)
         if (!decode_helper(ctx, batch, batch_logits, n_batch, n_vocab)) {
@@ -1592,7 +1592,7 @@ static void multiple_choice_score(llama_context * ctx, const common_params & par
             return;
         }
 
-        llama_kv_self_clear(ctx);
+        llama_memory_clear(llama_get_memory(ctx), true);
 
         // decode all tasks [i0, i1)
         if (!decode_helper(ctx, batch, batch_logits, n_batch, n_vocab)) {
@@ -1782,7 +1782,7 @@ static void kl_divergence(llama_context * ctx, const common_params & params) {
         }
 
         // clear the KV cache
-        llama_kv_self_clear(ctx);
+        llama_memory_clear(llama_get_memory(ctx), true);
 
         llama_batch batch = llama_batch_init(n_batch, 0, 1);
 
diff --git a/tools/quantize/quantize.cpp b/tools/quantize/quantize.cpp
index 3f54af7c5..8acc76517 100644
--- a/tools/quantize/quantize.cpp
+++ b/tools/quantize/quantize.cpp
@@ -107,13 +107,11 @@ static bool try_parse_ftype(const std::string & ftype_str_in, llama_ftype & ftyp
     return false;
 }
 
-// usage:
-//  ./llama-quantize [--allow-requantize] [--leave-output-tensor] [--pure] models/llama/ggml-model.gguf [models/llama/ggml-model-quant.gguf] type [nthreads]
-//
 [[noreturn]]
 static void usage(const char * executable) {
-    printf("usage: %s [--help] [--allow-requantize] [--leave-output-tensor] [--pure] [--imatrix] [--include-weights] [--exclude-weights] [--output-tensor-type]\n", executable);
-    printf("       [--token-embedding-type] [--tensor-type] [--keep-split] [--override-kv] model-f32.gguf [model-quant.gguf] type [nthreads]\n\n");
+    printf("usage: %s [--help] [--allow-requantize] [--leave-output-tensor] [--pure] [--imatrix] [--include-weights]\n", executable);
+    printf("       [--exclude-weights] [--output-tensor-type] [--token-embedding-type] [--tensor-type] [--prune-layers] [--keep-split] [--override-kv]\n");
+    printf("       model-f32.gguf [model-quant.gguf] type [nthreads]\n\n");
     printf("  --allow-requantize: Allows requantizing tensors that have already been quantized. Warning: This can severely reduce quality compared to quantizing from 16bit or 32bit\n");
     printf("  --leave-output-tensor: Will leave output.weight un(re)quantized. Increases model size but may also increase quality, especially when requantizing\n");
     printf("  --pure: Disable k-quant mixtures and quantize all tensors to the same type\n");
@@ -124,6 +122,8 @@ static void usage(const char * executable) {
     printf("  --token-embedding-type ggml_type: use this ggml_type for the token embeddings tensor\n");
     printf("  --tensor-type TENSOR=TYPE: quantize this tensor to this ggml_type. example: --tensor-type attn_q=q8_0\n");
     printf("      Advanced option to selectively quantize tensors. May be specified multiple times.\n");
+    printf("  --prune-layers L0,L1,L2...comma-separated list of layer numbers to prune from the model\n");
+    printf("      Advanced option to remove all tensors from the given layers\n");
     printf("  --keep-split: will generate quantized model in the same shards as input\n");
     printf("  --override-kv KEY=TYPE:VALUE\n");
     printf("      Advanced option to override model metadata by key in the quantized model. May be specified multiple times.\n");
@@ -286,6 +286,32 @@ static bool parse_tensor_type(const char * data, std::vector<tensor_quantization
     return true;
 }
 
+static bool parse_layer_prune(const char * data, std::vector<int> & prune_layers) {
+    if (!data) {
+        printf("\n%s: no layer pruning ids provided\n\n", __func__);
+        return false;
+    }
+
+    const auto block_ids = string_split<std::string>(data, ',');
+    for (const auto & block_id : block_ids) {
+        int id;
+        try {
+            id = std::stoi(block_id);
+        } catch (...) {
+            id = -1;
+        }
+        if (id < 0) {
+            printf("\n%s: invalid layer id '%s'\n\n", __func__, block_id.c_str());
+            return false;
+        }
+        prune_layers.emplace_back(id);
+    }
+
+    sort(prune_layers.begin(), prune_layers.end());
+    prune_layers.erase(std::unique(prune_layers.begin(), prune_layers.end()), prune_layers.end());
+    return true;
+}
+
 int main(int argc, char ** argv) {
     if (argc < 3) {
         usage(argv[0]);
@@ -298,6 +324,7 @@ int main(int argc, char ** argv) {
     std::vector<std::string> included_weights, excluded_weights;
     std::vector<llama_model_kv_override> kv_overrides;
     std::vector<tensor_quantization> tensor_types;
+    std::vector<int> prune_layers;
 
     for (; arg_idx < argc && strncmp(argv[arg_idx], "--", 2) == 0; arg_idx++) {
         if (strcmp(argv[arg_idx], "--leave-output-tensor") == 0) {
@@ -324,6 +351,10 @@ int main(int argc, char ** argv) {
             if (arg_idx == argc-1 || !parse_tensor_type(argv[++arg_idx], tensor_types)) {
                 usage(argv[0]);
             }
+        } else if (strcmp(argv[arg_idx], "--prune-layers") == 0) {
+            if (arg_idx == argc-1 || !parse_layer_prune(argv[++arg_idx], prune_layers)) {
+                usage(argv[0]);
+            }
         } else if (strcmp(argv[arg_idx], "--override-kv") == 0) {
             if (arg_idx == argc-1 || !string_parse_kv_override(argv[++arg_idx], kv_overrides)) {
                 usage(argv[0]);
@@ -411,6 +442,9 @@ int main(int argc, char ** argv) {
     if (!tensor_types.empty()) {
         params.tensor_types = &tensor_types;
     }
+    if (!prune_layers.empty()) {
+        params.prune_layers = &prune_layers;
+    }
 
     llama_backend_init();
 
diff --git a/tools/quantize/tests.sh b/tools/quantize/tests.sh
index 70f7610f9..ba9616148 100644
--- a/tools/quantize/tests.sh
+++ b/tools/quantize/tests.sh
@@ -1,4 +1,4 @@
-#!/bin/bash
+#!/usr/bin/env bash
 
 set -eu
 
diff --git a/tools/run/run.cpp b/tools/run/run.cpp
index 4aef93863..6fe728c68 100644
--- a/tools/run/run.cpp
+++ b/tools/run/run.cpp
@@ -9,6 +9,9 @@
 #include <nlohmann/json.hpp>
 
 #if defined(_WIN32)
+#    ifndef NOMINMAX
+#        define NOMINMAX
+#    endif
 #    include <windows.h>
 #    include <io.h>
 #else
@@ -939,23 +942,36 @@ static int apply_chat_template(const struct common_chat_templates * tmpls, Llama
 // Function to tokenize the prompt
 static int tokenize_prompt(const llama_vocab * vocab, const std::string & prompt,
                            std::vector<llama_token> & prompt_tokens, const LlamaData & llama_data) {
-    const bool is_first = llama_kv_self_seq_pos_max(llama_data.context.get(), 0) == 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);
-    if (llama_tokenize(vocab, prompt.c_str(), prompt.size(), prompt_tokens.data(), prompt_tokens.size(), is_first,
-                       true) < 0) {
-        printe("failed to tokenize the prompt\n");
+    const bool is_first = llama_memory_seq_pos_max(llama_get_memory(llama_data.context.get()), 0) == -1;
+    int n_tokens = prompt.size() + 2 * is_first;
+    prompt_tokens.resize(n_tokens);
+    n_tokens = llama_tokenize(vocab, prompt.c_str(), prompt.size(),
+                              prompt_tokens.data(), prompt_tokens.size(),
+                              is_first, /*parse_special =*/true);
+    if (n_tokens == std::numeric_limits<int32_t>::min()) {
+        printe("tokenization failed: input too large\n");
         return -1;
     }
-
-    return n_prompt_tokens;
+    if (n_tokens < 0) {
+        prompt_tokens.resize(-n_tokens);
+        int check = llama_tokenize(vocab, prompt.c_str(), prompt.size(),
+                                   prompt_tokens.data(), prompt_tokens.size(),
+                                   is_first, /*parse_special =*/true);
+        if (check != -n_tokens) {
+            printe("failed to tokenize the prompt (size mismatch)\n");
+            return -1;
+        }
+        n_tokens = check;
+    } else {
+        prompt_tokens.resize(n_tokens);
+    }
+    return n_tokens;
 }
 
 // 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_kv_self_seq_pos_max(ctx.get(), 0);
+    const int n_ctx_used = llama_memory_seq_pos_max(llama_get_memory(ctx.get()), 0);
     if (n_ctx_used + batch.n_tokens > n_ctx) {
         printf(LOG_COL_DEFAULT "\n");
         printe("context size exceeded\n");
diff --git a/tools/server/README.md b/tools/server/README.md
index 06533c172..6f962664f 100644
--- a/tools/server/README.md
+++ b/tools/server/README.md
@@ -164,6 +164,7 @@ The project is under active development, and we are [looking for feedback and co
 | `--api-key-file FNAME` | path to file containing API keys (default: none) |
 | `--ssl-key-file FNAME` | path to file a PEM-encoded SSL private key<br/>(env: LLAMA_ARG_SSL_KEY_FILE) |
 | `--ssl-cert-file FNAME` | path to file a PEM-encoded SSL certificate<br/>(env: LLAMA_ARG_SSL_CERT_FILE) |
+| `--chat-template-kwargs STRING` | JSON object containing additional params for the json template parser. Example: `--chat_template_kwargs "{\"enable_thinking\":false}`"<br/>(env: LLAMA_CHAT_TEMPLATE_KWARGS) |
 | `-to, --timeout N` | server read/write timeout in seconds (default: 600)<br/>(env: LLAMA_ARG_TIMEOUT) |
 | `--threads-http N` | number of threads used to process HTTP requests (default: -1)<br/>(env: LLAMA_ARG_THREADS_HTTP) |
 | `--cache-reuse N` | min chunk size to attempt reusing from the cache via KV shifting (default: 0)<br/>[(card)](https://ggml.ai/f0.png)<br/>(env: LLAMA_ARG_CACHE_REUSE) |
@@ -187,6 +188,8 @@ The project is under active development, and we are [looking for feedback and co
 | `-devd, --device-draft <dev1,dev2,..>` | comma-separated list of devices to use for offloading the draft model (none = don't offload)<br/>use --list-devices to see a list of available devices |
 | `-ngld, --gpu-layers-draft, --n-gpu-layers-draft N` | number of layers to store in VRAM for the draft model<br/>(env: LLAMA_ARG_N_GPU_LAYERS_DRAFT) |
 | `-md, --model-draft FNAME` | draft model for speculative decoding (default: unused)<br/>(env: LLAMA_ARG_MODEL_DRAFT) |
+| `-ctkd, --cache-type-k-draft TYPE` | KV cache data type for K for speculative decoding model<br/>allowed values: f32, f16, bf16, q8_0, q4_0, q4_1, iq4_nl, q5_0, q5_1<br/>(default: f16)<br/>(env: LLAMA_ARG_CACHE_TYPE_K_DRAFT) |
+| `-ctvd, --cache-type-v-draft TYPE` | KV cache data type for V for speculative decoding model<br/>allowed values: f32, f16, bf16, q8_0, q4_0, q4_1, iq4_nl, q5_0, q5_1<br/>(default: f16)<br/>(env: LLAMA_ARG_CACHE_TYPE_V_DRAFT) |
 | `-mv, --model-vocoder FNAME` | vocoder model for audio generation (default: unused) |
 | `--tts-use-guide-tokens` | Use guide tokens to improve TTS word recall |
 | `--embd-bge-small-en-default` | use default bge-small-en-v1.5 model (note: can download weights from the internet) |
@@ -368,6 +371,8 @@ node index.js
 
 ### GET `/health`: Returns heath check result
 
+This endpoint is public (no API key check).
+
 **Response format**
 
 - HTTP status code 503
@@ -706,7 +711,7 @@ If the tokens are missing, then the extra context is simply prefixed at the star
 
 ### **GET** `/props`: Get server global properties.
 
-This endpoint is public (no API key check). By default, it is read-only. To make POST request to change global properties, you need to start server with `--props`
+By default, it is read-only. To make POST request to change global properties, you need to start server with `--props`
 
 **Response format**
 
@@ -1114,6 +1119,8 @@ See [OpenAI Chat Completions API documentation](https://platform.openai.com/docs
 
 The `response_format` parameter supports both plain JSON output (e.g. `{"type": "json_object"}`) and schema-constrained JSON (e.g. `{"type": "json_object", "schema": {"type": "string", "minLength": 10, "maxLength": 100}}` or `{"type": "json_schema", "schema": {"properties": { "name": { "title": "Name",  "type": "string" }, "date": { "title": "Date",  "type": "string" }, "participants": { "items": {"type: "string" }, "title": "Participants",  "type": "string" } } } }`), similar to other OpenAI-inspired API providers.
 
+`chat_template_kwargs`: Allows sending additional parameters to the json templating system. For example: `{"enable_thinking": false}`
+
 *Examples:*
 
 You can use either Python `openai` library with appropriate checkpoints:
diff --git a/tools/server/chat-llama2.sh b/tools/server/chat-llama2.sh
index 1fc79b7e1..450445f17 100755
--- a/tools/server/chat-llama2.sh
+++ b/tools/server/chat-llama2.sh
@@ -1,4 +1,4 @@
-#!/bin/bash
+#!/usr/bin/env bash
 
 API_URL="${API_URL:-http://127.0.0.1:8080}"
 
diff --git a/tools/server/chat.sh b/tools/server/chat.sh
index da0a6ca68..84cea2d56 100755
--- a/tools/server/chat.sh
+++ b/tools/server/chat.sh
@@ -1,4 +1,4 @@
-#!/bin/bash
+#!/usr/bin/env bash
 
 API_URL="${API_URL:-http://127.0.0.1:8080}"
 
diff --git a/tools/server/public/index.html.gz b/tools/server/public/index.html.gz
index f8e304342..53b71079c 100644
Binary files a/tools/server/public/index.html.gz and b/tools/server/public/index.html.gz differ
diff --git a/tools/server/server.cpp b/tools/server/server.cpp
index 4b92eeac9..d3f627193 100644
--- a/tools/server/server.cpp
+++ b/tools/server/server.cpp
@@ -88,6 +88,26 @@ enum error_type {
     ERROR_TYPE_NOT_SUPPORTED, // custom error
 };
 
+static bool server_task_type_need_embd(server_task_type task_type) {
+    switch (task_type) {
+        case SERVER_TASK_TYPE_EMBEDDING:
+        case SERVER_TASK_TYPE_RERANK:
+            return true;
+        default:
+            return false;
+    }
+}
+
+static bool server_task_type_need_logits(server_task_type task_type) {
+    switch (task_type) {
+        case SERVER_TASK_TYPE_COMPLETION:
+        case SERVER_TASK_TYPE_INFILL:
+            return true;
+        default:
+            return false;
+    }
+}
+
 struct slot_params {
     bool stream        = true;
     bool cache_prompt  = true; // remember the prompt to avoid reprocessing all prompt
@@ -233,6 +253,7 @@ struct server_task {
         slot_params defaults;
         defaults.sampling    = params_base.sampling;
         defaults.speculative = params_base.speculative;
+        defaults.n_keep      = params_base.n_keep;
 
         // enabling this will output extra debug information in the HTTP responses from the server
         params.verbose           = params_base.verbosity > 9;
@@ -360,7 +381,7 @@ struct server_task {
                 params.oaicompat_chat_syntax.format = defaults.oaicompat_chat_syntax.format;
             }
             params.oaicompat_chat_syntax.reasoning_format = params_base.reasoning_format;
-            params.oaicompat_chat_syntax.reasoning_in_content = params.stream;
+            params.oaicompat_chat_syntax.reasoning_in_content = params.stream && (params_base.reasoning_format == COMMON_REASONING_FORMAT_DEEPSEEK_LEGACY);
             params.oaicompat_chat_syntax.thinking_forced_open = json_value(data, "thinking_forced_open", false);
             params.oaicompat_chat_syntax.parse_tool_calls = json_value(data, "parse_tool_calls", false);
         }
@@ -1329,13 +1350,24 @@ struct server_slot {
         n_draft_accepted = 0;
     }
 
-    bool is_non_causal() const {
-        return task_type == SERVER_TASK_TYPE_EMBEDDING || task_type == SERVER_TASK_TYPE_RERANK;
+    bool need_embd() const {
+        return server_task_type_need_embd(task_type);
+    }
+
+    bool need_logits() const {
+        return server_task_type_need_logits(task_type);
+    }
+
+    // if the context does not have a memory module then all embeddings have to be computed within a single ubatch
+    // also we cannot split if the pooling would require any past tokens
+    bool can_split() const {
+        return
+            !need_embd() ||
+            (llama_get_memory(ctx) && llama_pooling_type(ctx) == LLAMA_POOLING_TYPE_LAST);
     }
 
     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);
+        return task_type == other_slot.task_type && are_lora_equal(lora, other_slot.lora);
     }
 
     bool has_budget(const common_params & global_params) {
@@ -1479,7 +1511,6 @@ struct server_slot {
             {"n_ctx",         n_ctx},
             {"speculative",   can_speculate()},
             {"is_processing", is_processing()},
-            {"non_causal",    is_non_causal()},
             {"params",        params.to_json()},
             {"prompt",        prompt_tokens.detokenize(ctx, true)},
             {"next_token",
@@ -1938,10 +1969,8 @@ struct server_context {
             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;
+            params_dft.cache_type_k = params_base.speculative.cache_type_k;
+            params_dft.cache_type_v = params_base.speculative.cache_type_v;
 
             llama_init_dft = common_init_from_params(params_dft);
 
@@ -2006,7 +2035,7 @@ struct server_context {
             }
         }
 
-        if (!llama_kv_self_can_shift(ctx)) {
+        if (!llama_memory_can_shift(llama_get_memory(ctx))) {
             if (params_base.ctx_shift) {
                 params_base.ctx_shift = false;
                 SRV_WRN("%s\n", "ctx_shift is not supported by this context, it will be disabled");
@@ -2055,6 +2084,7 @@ struct server_context {
             SLT_INF(slot, "new slot n_ctx_slot = %d\n", slot.n_ctx);
 
             slot.params.sampling = params_base.sampling;
+            slot.params.n_keep = params_base.n_keep;
 
             slot.callback_on_release = [this](int) {
                 queue_tasks.pop_deferred_task();
@@ -2080,6 +2110,7 @@ struct server_context {
             /* use_jinja             */ params_base.use_jinja,
             /* prefill_assistant     */ params_base.prefill_assistant,
             /* reasoning_format      */ params_base.reasoning_format,
+            /* chat_template_kwargs  */ params_base.default_template_kwargs,
             /* common_chat_templates */ chat_templates.get(),
             /* allow_image           */ mctx ? mtmd_support_vision(mctx) : false,
             /* allow_audio           */ mctx ? mtmd_support_audio (mctx) : false,
@@ -2137,7 +2168,8 @@ struct server_context {
 
         // find the slot that has been least recently used
         if (ret == nullptr) {
-            int64_t t_last = ggml_time_us();
+            int64_t t_last = -1;
+
             for (server_slot & slot : slots) {
                 // skip the slot if it is not available
                 if (slot.is_processing()) {
@@ -2145,7 +2177,7 @@ struct server_context {
                 }
 
                 // select the current slot if the criteria match
-                if (slot.t_last_used < t_last) {
+                if (!ret || slot.t_last_used <= t_last) {
                     t_last = slot.t_last_used;
                     ret = &slot;
                 }
@@ -2219,7 +2251,7 @@ struct server_context {
         SRV_DBG("%s", "clearing KV cache\n");
 
         // clear the entire KV cache
-        llama_kv_self_clear(ctx);
+        llama_memory_clear(llama_get_memory(ctx), true);
         clean_kv_cache = false;
     }
 
@@ -2727,6 +2759,7 @@ struct server_context {
                         queue_tasks.defer(std::move(task));
                         break;
                     }
+
                     if (slot->is_processing()) {
                         // if requested slot is unavailable, we defer this task for processing later
                         SRV_DBG("requested slot is unavailable, defer task, id_task = %d\n", task.id);
@@ -2905,7 +2938,7 @@ struct server_context {
 
                     // Erase token cache
                     const size_t n_erased = slot->cache_tokens.size();
-                    llama_kv_self_seq_rm(ctx, slot->id, -1, -1);
+                    llama_memory_seq_rm(llama_get_memory(ctx), slot->id, -1, -1);
                     slot->cache_tokens.clear();
 
                     auto res = std::make_unique<server_task_result_slot_erase>();
@@ -2980,8 +3013,8 @@ 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_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);
+                llama_memory_seq_rm (llama_get_memory(ctx), slot.id, n_keep            , n_keep + n_discard);
+                llama_memory_seq_add(llama_get_memory(ctx), slot.id, n_keep + n_discard, slot.n_past,        -n_discard);
 
                 // add generated tokens to cache
                 {
@@ -3089,7 +3122,14 @@ struct server_context {
                             continue;
                         }
 
-                        if (slot.is_non_causal()) {
+                        // TODO: support memory-less logits computation
+                        if (slot.need_logits() && !llama_get_memory(ctx)) {
+                            slot.release();
+                            send_error(slot, "the current context does not logits computation. skipping", ERROR_TYPE_SERVER);
+                            continue;
+                        }
+
+                        if (!slot.can_split()) {
                             if (slot.n_prompt_tokens > n_ubatch) {
                                 slot.release();
                                 send_error(slot, "input is too large to process. increase the physical batch size", ERROR_TYPE_SERVER);
@@ -3184,8 +3224,8 @@ struct server_context {
 
                                             const int64_t kv_shift = (int64_t) head_p - (int64_t) head_c;
 
-                                            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);
+                                            llama_memory_seq_rm (llama_get_memory(ctx), slot.id, head_p, head_c);
+                                            llama_memory_seq_add(llama_get_memory(ctx), slot.id, head_c, head_c + n_match, kv_shift);
 
                                             for (size_t i = 0; i < n_match; i++) {
                                                 slot.cache_tokens.set_token(head_p + i, slot.cache_tokens[head_c + i]);
@@ -3203,32 +3243,28 @@ struct server_context {
                                 }
                             } else {
                                 // if we don't cache the prompt, we have to remove the entire KV cache
-                                llama_kv_self_seq_rm(ctx, slot.id, 0, -1);
                                 slot.n_past = 0;
-                                slot.cache_tokens.clear(); // TODO: not needed, will be cleared later via "keep_first()"
                             }
 
                             if (slot.n_past > 0 && slot.n_past < (int) slot.cache_tokens.size()) {
-                                const auto pos_min = llama_kv_self_seq_pos_min(ctx, slot.id);
+                                const auto pos_min = llama_memory_seq_pos_min(llama_get_memory(ctx), slot.id);
                                 if (pos_min == -1) {
                                     SLT_ERR(slot, "n_past = %d, cache_tokens.size() = %d, seq_id = %d, pos_min = %d\n", slot.n_past, (int) slot.cache_tokens.size(), slot.id, pos_min);
                                     GGML_ABORT("pos_min == -1, but n_past > 0 - should not happen: https://github.com/ggml-org/llama.cpp/pull/13833#discussion_r2116181237");
                                 }
 
                                 const auto n_swa = llama_model_n_swa(model);
-                                if (pos_min > slot.n_past - n_swa) {
+                                if (pos_min > std::max(0, slot.n_past - n_swa)) {
                                     SLT_WRN(slot, "n_past = %d, cache_tokens.size() = %d, seq_id = %d, pos_min = %d, n_swa = %d\n", slot.n_past, (int) slot.cache_tokens.size(), slot.id, pos_min, n_swa);
                                     SLT_WRN(slot, "forcing full prompt re-processing due to lack of cache data (likely due to SWA, see %s)\n",
                                             "https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055");
-                                    llama_kv_self_seq_rm(ctx, slot.id, 0, -1);
                                     slot.n_past = 0;
                                 }
                             }
                         }
 
                         if (slot.n_past == slot.n_prompt_tokens && slot.n_past > 0) {
-                            // we have to evaluate at least 1 token to generate logits.
-                            SLT_WRN(slot, "need to evaluate at least 1 token to generate logits, n_past = %d, n_prompt_tokens = %d\n", slot.n_past, slot.n_prompt_tokens);
+                            SLT_WRN(slot, "need to evaluate at least 1 token for each active slot, n_past = %d, n_prompt_tokens = %d\n", slot.n_past, slot.n_prompt_tokens);
 
                             slot.n_past--;
                         }
@@ -3236,8 +3272,7 @@ struct server_context {
                         slot.n_prompt_tokens_processed = 0;
                     }
 
-                    // non-causal tasks require to fit the entire prompt in the physical batch
-                    if (slot.is_non_causal()) {
+                    if (!slot.can_split()) {
                         // cannot fit the prompt in the current batch - will try next iter
                         if (batch.n_tokens + slot.n_prompt_tokens > n_batch) {
                             continue;
@@ -3245,9 +3280,9 @@ struct server_context {
                     }
 
                     // keep only the common part
-                    if (!llama_kv_self_seq_rm(ctx, slot.id, slot.n_past, -1)) {
+                    if (!llama_memory_seq_rm(llama_get_memory(ctx), slot.id, slot.n_past, -1)) {
                         // could not partially delete (likely using a non-Transformer model)
-                        llama_kv_self_seq_rm(ctx, slot.id, -1, -1);
+                        llama_memory_seq_rm(llama_get_memory(ctx), slot.id, -1, -1);
 
                         // there is no common part left
                         slot.n_past = 0;
@@ -3259,8 +3294,7 @@ struct server_context {
                     slot.cache_tokens.keep_first(slot.n_past);
 
                     // check if we should process the image
-                    if (slot.n_past < slot.n_prompt_tokens
-                            && slot.prompt_tokens[slot.n_past] == LLAMA_TOKEN_NULL) {
+                    if (slot.n_past < slot.n_prompt_tokens && slot.prompt_tokens[slot.n_past] == LLAMA_TOKEN_NULL) {
                         // process the image
                         int32_t new_n_past;
                         int32_t res = slot.prompt_tokens.process_chunk(ctx, mctx, slot.n_past, slot.id, new_n_past);
@@ -3291,8 +3325,8 @@ struct server_context {
                             break; // end of text chunk
                         }
 
-                        // without pooling, we want to output the embeddings for all the tokens in the batch
-                        const bool need_embd = slot.task_type == SERVER_TASK_TYPE_EMBEDDING && llama_pooling_type(slot.ctx) == LLAMA_POOLING_TYPE_NONE;
+                        // embedding requires all tokens in the batch to be output
+                        const bool need_embd = server_task_type_need_embd(slot.task_type);
 
                         common_batch_add(batch, cur_tok, slot.n_past, { slot.id }, need_embd);
                         slot.cache_tokens.push_back(cur_tok);
@@ -3346,41 +3380,10 @@ struct server_context {
         SRV_DBG("decoding batch, n_tokens = %d\n", batch.n_tokens);
 
         if (slot_batched) {
-            // make sure we're in the right embedding mode
-            llama_set_embeddings(ctx, slot_batched->is_non_causal());
             // apply lora, only need to do it once per batch
             common_set_adapter_lora(ctx, slot_batched->lora);
-        }
 
-        const bool do_encode = (params_base.embedding || params_base.reranking);
-
-        // pad the batch so that batch.n_tokens >= n_slots
-        // TODO: temporary workaround for https://github.com/ggml-org/llama.cpp/issues/13689
-        if (do_encode) {
-            const int n_slots = slots.size();
-
-            if (batch.n_tokens < n_slots) {
-                std::set<llama_seq_id> seq_ids;
-                for (int j = 0; j < batch.n_tokens; ++j) {
-                    seq_ids.insert(batch.seq_id[j][0]);
-                }
-
-                // find unused sequence id
-                llama_seq_id seq_id = -1;
-                for (int i = 0; i < n_slots; ++i) {
-                    if (seq_ids.find(i) == seq_ids.end()) {
-                        seq_id = i;
-                    }
-                }
-
-                const int n_add = n_slots - batch.n_tokens;
-
-                SRV_WRN("adding %d dummy tokens to the batch, seq_id = %d\n", n_add, seq_id);
-
-                for (int j = 0; j < n_add; ++j) {
-                    common_batch_add(batch, 0, j, { seq_id }, false);
-                }
-            }
+            llama_set_embeddings(ctx, slot_batched->need_embd());
         }
 
         int32_t i_next = 0;
@@ -3416,9 +3419,12 @@ struct server_context {
                     }
 
                     if (ret < -1) {
+                        // TODO: update slot state based on llama_memory_seq_pos_min() and llama_memory_seq_pos_max()
                         err = "Compute error.";
                     }
 
+                    // TODO: handle ret == 2 (abort) when we start aborting
+
                     if (!err.empty()) {
                         SRV_ERR("%s, i = %d, n_batch = %d, ret = %d\n", err.c_str(), i, n_batch, ret);
                         for (auto & slot : slots) {
@@ -3553,9 +3559,6 @@ struct server_context {
                 const llama_tokens & cached_text_tokens = slot.cache_tokens.get_text_tokens();
                 llama_tokens draft = common_speculative_gen_draft(slot.spec, params_spec, cached_text_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);
@@ -3563,6 +3566,9 @@ struct server_context {
                     continue;
                 }
 
+                // keep track of total number of drafted tokens tested
+                slot.n_draft_total += draft.size();
+
                 // construct the speculation batch
                 common_batch_clear(slot.batch_spec);
                 common_batch_add  (slot.batch_spec, id, slot.n_past, { slot.id }, true);
@@ -3581,13 +3587,13 @@ struct server_context {
                 slot.n_past    += ids.size();
                 slot.n_decoded += ids.size();
 
-                // update how many tokens out of draft was accepted
+                // update how many tokens out of those tested were accepted
                 slot.n_draft_accepted += ids.size() - 1;
 
                 slot.cache_tokens.push_back(id);
                 slot.cache_tokens.insert({ids.begin(), ids.end() - 1});
 
-                llama_kv_self_seq_rm(ctx, slot.id, slot.n_past, -1);
+                llama_memory_seq_rm(llama_get_memory(ctx), slot.id, slot.n_past, -1);
 
                 for (size_t i = 0; i < ids.size(); ++i) {
                     completion_token_output result;
@@ -4174,11 +4180,6 @@ int main(int argc, char ** argv) {
             oaicompat_type oaicompat) -> void {
         GGML_ASSERT(type == SERVER_TASK_TYPE_COMPLETION || type == SERVER_TASK_TYPE_INFILL);
 
-        if (ctx_server.params_base.embedding) {
-            res_error(res, format_error_response("This server does not support completions. Start it without `--embeddings`", ERROR_TYPE_NOT_SUPPORTED));
-            return;
-        }
-
         auto completion_id = gen_chatcmplid();
         std::unordered_set<int> task_ids;
         try {
@@ -4433,12 +4434,8 @@ int main(int argc, char ** argv) {
             OAICOMPAT_TYPE_NONE); // infill is not OAI compatible
     };
 
-    const auto handle_chat_completions = [&ctx_server, &res_error, &handle_completions_impl](const httplib::Request & req, httplib::Response & res) {
+    const auto handle_chat_completions = [&ctx_server, &handle_completions_impl](const httplib::Request & req, httplib::Response & res) {
         LOG_DBG("request: %s\n", req.body.c_str());
-        if (ctx_server.params_base.embedding) {
-            res_error(res, format_error_response("This server does not support completions. Start it without `--embeddings`", ERROR_TYPE_NOT_SUPPORTED));
-            return;
-        }
 
         auto body = json::parse(req.body);
         std::vector<raw_buffer> files;
@@ -4566,13 +4563,18 @@ int main(int argc, char ** argv) {
     };
 
     const auto handle_embeddings_impl = [&ctx_server, &res_error, &res_ok](const httplib::Request & req, httplib::Response & res, oaicompat_type oaicompat) {
-        const json body = json::parse(req.body);
+        if (!ctx_server.params_base.embedding) {
+            res_error(res, format_error_response("This server does not support embeddings. Start it with `--embeddings`", ERROR_TYPE_NOT_SUPPORTED));
+            return;
+        }
 
         if (oaicompat != OAICOMPAT_TYPE_NONE && llama_pooling_type(ctx_server.ctx) == LLAMA_POOLING_TYPE_NONE) {
             res_error(res, format_error_response("Pooling type 'none' is not OAI compatible. Please use a different pooling type", ERROR_TYPE_INVALID_REQUEST));
             return;
         }
 
+        const json body = json::parse(req.body);
+
         // for the shape of input/content, see tokenize_input_prompts()
         json prompt;
         if (body.count("input") != 0) {
@@ -4662,8 +4664,8 @@ int main(int argc, char ** argv) {
     };
 
     const auto handle_rerank = [&ctx_server, &res_error, &res_ok](const httplib::Request & req, httplib::Response & res) {
-        if (!ctx_server.params_base.reranking || ctx_server.params_base.embedding) {
-            res_error(res, format_error_response("This server does not support reranking. Start it with `--reranking` and without `--embedding`", ERROR_TYPE_NOT_SUPPORTED));
+        if (!ctx_server.params_base.embedding || ctx_server.params_base.pooling_type != LLAMA_POOLING_TYPE_RANK) {
+            res_error(res, format_error_response("This server does not support reranking. Start it with `--reranking`", ERROR_TYPE_NOT_SUPPORTED));
             return;
         }
 
@@ -4878,7 +4880,9 @@ int main(int argc, char ** argv) {
     };
 
     bool was_bound = false;
+    bool is_sock = false;
     if (string_ends_with(std::string(params.hostname), ".sock")) {
+        is_sock = true;
         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
@@ -4956,7 +4960,9 @@ int main(int argc, char ** argv) {
     SetConsoleCtrlHandler(reinterpret_cast<PHANDLER_ROUTINE>(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);
+    LOG_INF("%s: server is listening on %s - starting the main loop\n", __func__,
+            is_sock ? string_format("unix://%s", params.hostname.c_str()).c_str() :
+                      string_format("http://%s:%d", params.hostname.c_str(), params.port).c_str());
 
     // this call blocks the main thread until queue_tasks.terminate() is called
     ctx_server.queue_tasks.start_loop();
diff --git a/tools/server/tests/tests.sh b/tools/server/tests/tests.sh
index 33fa8cc64..709b5841a 100755
--- a/tools/server/tests/tests.sh
+++ b/tools/server/tests/tests.sh
@@ -1,4 +1,4 @@
-#!/bin/bash
+#!/usr/bin/env bash
 
 # make sure we are in the right directory
 SCRIPT_DIR=$( cd -- "$( dirname -- "${BASH_SOURCE[0]}" )" &> /dev/null && pwd )
diff --git a/tools/server/tests/unit/test_chat_completion.py b/tools/server/tests/unit/test_chat_completion.py
index 1b5205f79..7ee9a1651 100644
--- a/tools/server/tests/unit/test_chat_completion.py
+++ b/tools/server/tests/unit/test_chat_completion.py
@@ -132,6 +132,28 @@ def test_chat_template():
     assert res.body["__verbose"]["prompt"] == "<s> <|start_header_id|>system<|end_header_id|>\n\nBook<|eot_id|><|start_header_id|>user<|end_header_id|>\n\nWhat is the best book<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n"
 
 
+@pytest.mark.parametrize("prefill,re_prefill", [
+    ("Whill", "Whill"),
+    ([{"type": "text", "text": "Wh"}, {"type": "text", "text": "ill"}], "Whill"),
+])
+def test_chat_template_assistant_prefill(prefill, re_prefill):
+    global server
+    server.chat_template = "llama3"
+    server.debug = True  # to get the "__verbose" object in the response
+    server.start()
+    res = server.make_request("POST", "/chat/completions", data={
+        "max_tokens": 8,
+        "messages": [
+            {"role": "system", "content": "Book"},
+            {"role": "user", "content": "What is the best book"},
+            {"role": "assistant", "content": prefill},
+        ]
+    })
+    assert res.status_code == 200
+    assert "__verbose" in res.body
+    assert res.body["__verbose"]["prompt"] == f"<s> <|start_header_id|>system<|end_header_id|>\n\nBook<|eot_id|><|start_header_id|>user<|end_header_id|>\n\nWhat is the best book<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n{re_prefill}"
+
+
 def test_apply_chat_template():
     global server
     server.chat_template = "command-r"
@@ -228,6 +250,7 @@ def test_completion_with_grammar(jinja: bool, grammar: str, n_predicted: int, re
     [{"role": "system", "content": 123}],
     # [{"content": "hello"}], # TODO: should not be a valid case
     [{"role": "system", "content": "test"}, {}],
+    [{"role": "user", "content": "test"}, {"role": "assistant", "content": "test"}, {"role": "assistant", "content": "test"}],
 ])
 def test_invalid_chat_completion_req(messages):
     global server
diff --git a/tools/server/tests/unit/test_tool_call.py b/tools/server/tests/unit/test_tool_call.py
index 610610749..20f048c6f 100755
--- a/tools/server/tests/unit/test_tool_call.py
+++ b/tools/server/tests/unit/test_tool_call.py
@@ -499,13 +499,12 @@ def do_test_calc_result(server: ServerProcess, result_override: str | None, n_pr
 
 
 @pytest.mark.slow
-@pytest.mark.parametrize("n_predict,reasoning_format,stream,expect_reasoning_content,expect_content,hf_repo,template_override", [
-    (128, 'deepseek',   CompletionMode.NORMAL,   None, "^The sum of 102 and 7 is 109[\\s\\S]*",                                       "bartowski/Phi-3.5-mini-instruct-GGUF:Q4_K_M",       None),
-    (128,  None,        CompletionMode.NORMAL,   None, "^The sum of 102 and 7 is 109[\\s\\S]*",                                       "bartowski/Phi-3.5-mini-instruct-GGUF:Q4_K_M",       None),
-    (1024, 'deepseek',  CompletionMode.NORMAL,   "I need to calculate the sum of 102 and 7[\\s\\S]*", "To find the sum of[\\s\\S]*",  "bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q4_K_M", None),
-    (1024, 'deepseek',  CompletionMode.STREAMED, None, "^<think>I need to calculate [\\s\\S]*?</think>To find the sum of [\\s\\S]*",  "bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q4_K_M", None),
-    (1024, 'deepseek',  CompletionMode.NORMAL,   "First, I [\\s\\S]*", "To find the sum of[\\s\\S]*",                                 "bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q4_K_M", ("llama-cpp-deepseek-r1", None)),
-    (1024, 'deepseek',  CompletionMode.STREAMED, None, "^<think>First, I [\\s\\S]*?</think>To find the sum of[\\s\\S]*",              "bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q4_K_M", ("llama-cpp-deepseek-r1", None)),
+@pytest.mark.parametrize("stream", [CompletionMode.NORMAL, CompletionMode.STREAMED])
+@pytest.mark.parametrize("n_predict,reasoning_format,expect_reasoning_content,expect_content,hf_repo,template_override", [
+    (128, 'deepseek',   None, "^The sum of 102 and 7 is 109[\\s\\S]*",                                       "bartowski/Phi-3.5-mini-instruct-GGUF:Q4_K_M",       None),
+    (128,  None,        None, "^The sum of 102 and 7 is 109[\\s\\S]*",                                       "bartowski/Phi-3.5-mini-instruct-GGUF:Q4_K_M",       None),
+    (1024, 'deepseek',  "I need to calculate the sum of 102 and 7[\\s\\S]*", "To find the sum of[\\s\\S]*",  "bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q4_K_M", None),
+    (1024, 'deepseek',  "First, I [\\s\\S]*", "To find the sum of[\\s\\S]*",                                 "bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q4_K_M", ("llama-cpp-deepseek-r1", None)),
     # (1024, 'none',      CompletionMode.NORMAL,   None, "^(<think>\\s*)?I need[\\s\\S]*?</think>\\s*To find[\\s\\S]*",                 "bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF:Q4_K_M", None),
     # (128,  'deepseek',  None, "^Okay, let me figure out the sum of 102 and 7[\\s\\S]*",                      "bartowski/Qwen_QwQ-32B-GGUF:Q4_K_M",                None),
 ])
diff --git a/tools/server/tests/utils.py b/tools/server/tests/utils.py
index f7e1b3b3b..bc547ca03 100644
--- a/tools/server/tests/utils.py
+++ b/tools/server/tests/utils.py
@@ -308,10 +308,12 @@ class ServerProcess:
         stream = data.get('stream', False)
         if stream:
             content: list[str] = []
+            reasoning_content: list[str] = []
             tool_calls: list[dict] = []
             finish_reason: Optional[str] = None
 
             content_parts = 0
+            reasoning_content_parts = 0
             tool_call_parts = 0
             arguments_parts = 0
 
@@ -322,6 +324,10 @@ class ServerProcess:
                     assert len(choice['delta']['content']) > 0, f'Expected non empty content delta!'
                     content.append(choice['delta']['content'])
                     content_parts += 1
+                if choice['delta'].get('reasoning_content') is not None:
+                    assert len(choice['delta']['reasoning_content']) > 0, f'Expected non empty reasoning_content delta!'
+                    reasoning_content.append(choice['delta']['reasoning_content'])
+                    reasoning_content_parts += 1
                 if choice['delta'].get('finish_reason') is not None:
                     finish_reason = choice['delta']['finish_reason']
                 for tc in choice['delta'].get('tool_calls', []):
@@ -349,8 +355,10 @@ class ServerProcess:
                         tool_call['function']['name'] = tool_call['function'].get('name', '') + fct['name']
                     if fct.get('arguments') is not None:
                         tool_call['function']['arguments'] += fct['arguments']
+                        arguments_parts += 1
+                    tool_call_parts += 1
 
-            print(f'Streamed response had {content_parts} content parts, {tool_call_parts} tool call parts incl. {arguments_parts} arguments parts')
+            print(f'Streamed response had {content_parts} content parts, {reasoning_content_parts} reasoning_content parts, {tool_call_parts} tool call parts incl. {arguments_parts} arguments parts')
             result = dict(
                 choices=[
                     dict(
@@ -359,6 +367,7 @@ class ServerProcess:
                         message=dict(
                             role='assistant',
                             content=''.join(content) if content else None,
+                            reasoning_content=''.join(reasoning_content) if reasoning_content else None,
                             tool_calls=tool_calls if tool_calls else None,
                         ),
                     )
diff --git a/tools/server/utils.hpp b/tools/server/utils.hpp
index f3e0392a4..6c2e91359 100644
--- a/tools/server/utils.hpp
+++ b/tools/server/utils.hpp
@@ -271,12 +271,20 @@ static llama_tokens format_rerank(const struct llama_vocab * vocab, const llama_
     }
 
     result.reserve(doc.size() + query.size() + 4);
-    result.push_back(llama_vocab_bos(vocab));
+    if (llama_vocab_get_add_bos(vocab)) {
+        result.push_back(llama_vocab_bos(vocab));
+    }
     result.insert(result.end(), query.begin(), query.end());
-    result.push_back(eos_token);
-    result.push_back(llama_vocab_sep(vocab));
+    if (llama_vocab_get_add_eos(vocab)) {
+        result.push_back(eos_token);
+    }
+    if (llama_vocab_get_add_sep(vocab)) {
+        result.push_back(llama_vocab_sep(vocab));
+    }
     result.insert(result.end(), doc.begin(), doc.end());
-    result.push_back(eos_token);
+    if (llama_vocab_get_add_eos(vocab)) {
+        result.push_back(eos_token);
+    }
 
     return result;
 }
@@ -571,6 +579,7 @@ struct oaicompat_parser_options {
     bool use_jinja;
     bool prefill_assistant;
     common_reasoning_format reasoning_format;
+    std::map<std::string,std::string> chat_template_kwargs;
     common_chat_templates * tmpls;
     bool allow_image;
     bool allow_audio;
@@ -748,6 +757,13 @@ static json oaicompat_chat_params_parse(
         llama_params["parse_tool_calls"] = true;
     }
 
+    // merge the template args provided from command line with the args provided in the user request
+    auto chat_template_kwargs_object = json_value(body, "chat_template_kwargs", json::object());
+    inputs.chat_template_kwargs = opt.chat_template_kwargs;
+    for (const auto & item : chat_template_kwargs_object.items()) {
+        inputs.chat_template_kwargs[item.key()] = item.value().dump();
+    }
+
     // if the assistant message appears at the end of list, we do not add end-of-turn token
     // for ex. this can be useful to modify the reasoning process in reasoning models
     bool prefill_assistant_message = !inputs.messages.empty() && inputs.messages.back().role == "assistant" && opt.prefill_assistant;
@@ -763,6 +779,11 @@ static json oaicompat_chat_params_parse(
 
         /* TODO: test this properly */
         inputs.reasoning_format = COMMON_REASONING_FORMAT_NONE;
+
+        if ( (!inputs.enable_thinking) || inputs.chat_template_kwargs.find("enable_thinking") != inputs.chat_template_kwargs.end()) {
+            throw std::runtime_error("Assistant response prefill is incompatible with enable_thinking.");
+        }
+
         inputs.add_generation_prompt = true;
     }
 
@@ -771,7 +792,13 @@ static json oaicompat_chat_params_parse(
 
     /* Append assistant prefilled message */
     if (prefill_assistant_message) {
-         chat_params.prompt += last_message.content;
+        if (!last_message.content_parts.empty()) {
+            for (auto & p : last_message.content_parts) {
+                chat_params.prompt += p.text;
+            }
+        } else {
+            chat_params.prompt += last_message.content;
+        }
     }
 
     llama_params["chat_format"]      = static_cast<int>(chat_params.format);
diff --git a/tools/server/webui/src/App.tsx b/tools/server/webui/src/App.tsx
index 02f1719d3..8dfcf4907 100644
--- a/tools/server/webui/src/App.tsx
+++ b/tools/server/webui/src/App.tsx
@@ -32,7 +32,7 @@ function AppLayout() {
     <>
       <Sidebar />
       <main
-        className="drawer-content grow flex flex-col h-screen w-screen mx-auto px-4 overflow-auto bg-base-100"
+        className="drawer-content grow flex flex-col h-screen mx-auto px-4 overflow-auto bg-base-100"
         id="main-scroll"
       >
         <Header />
diff --git a/tools/server/webui/src/components/Sidebar.tsx b/tools/server/webui/src/components/Sidebar.tsx
index a77cb83b4..b52a8df03 100644
--- a/tools/server/webui/src/components/Sidebar.tsx
+++ b/tools/server/webui/src/components/Sidebar.tsx
@@ -231,7 +231,7 @@ function ConversationItem({
       >
         {conv.name}
       </button>
-      <div className="dropdown dropdown-end h-5">
+      <div tabIndex={0} className="dropdown dropdown-end h-5">
         <BtnWithTooltips
           // on mobile, we always show the ellipsis icon
           // on desktop, we only show it when the user hovers over the conversation item
diff --git a/tools/server/webui/src/index.scss b/tools/server/webui/src/index.scss
index 64460b740..362db6e17 100644
--- a/tools/server/webui/src/index.scss
+++ b/tools/server/webui/src/index.scss
@@ -41,6 +41,10 @@ html {
   max-width: 900px;
 }
 
+.chat-bubble {
+  @apply break-words;
+}
+
 .chat-bubble-base-300 {
   --tw-bg-opacity: 1;
   --tw-text-opacity: 1;