tqcq/llama.cpp
0
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-08-15 04:33:06 -04:00
Code Issues Packages Projects Releases Wiki Activity
Files
863d341eeb81db104902b14b6f9413daa515e957
llama.cpp/CMakeLists.txt
Jonathan Graehl 5cdb27e091 finetune: SGD optimizer, more CLI args (#13873) 
* examples/finetune -opt SGD (stochastic gradient descent) memory opt

add unit tested GGML_OPT_OPTIMIZER_SGD to ggml - avoids allocating
m, v tensors.

support finetune.cpp arg -opt SGD (or sgd). (default adamw as before)

llama 3.2-1b-F32 result: observed 11gb gpu ram (41 sec/epoch)
when using SGD instead of 19gb (55 sec/epoch) using adamw.
(wikipedia 100 lines finetune)

(
using the same GPU memory, adamw can only do before OOM 512
batch/context, reaching:
train: [███████▉] data=0000140/0000140 loss=0.02575±0.00099 acc=99.52±0.03% t=00:00:47 ETA=00:00:00
val:   [███████▉] data=0000008/0000008 loss=4.76565±0.28810 acc=41.46±0.77% t=00:00:00 ETA=00:00:00

SGD is superior, though it converges slower, with max before OOM 1728
batch/context (esp see the better validation perf):
train: [███████▉] data=0000039/0000039 loss=0.00371±0.00010 acc=99.96±0.01% t=00:00:41 ETA=00:00:00
val:   [███████▉] data=0000003/0000003 loss=5.11406±0.76034 acc=48.01±0.69% t=00:00:01 ETA=00:00:00
)

note: when finetuning long enough (or w/ enough -lr),
validation accuracy *eventually* drops ('catastrophic forgetting')

-lr-half (halflife) option useful for SGD to avoid oscillation or
super slow underdamped learning (makes setting -lr more forgiving).
terminal -lr for now is set by lr-halvings i.e. if you want at most
1/8 the inital -lr you set -lr-halvings 3.

note: objective loss not directly comparable between adamw, sgd? -
check perplexity or accuracy or consider relative improvements
for convergence

new finetune args -wd 1e-9 to enable weight decay in sgd or adamw,
and max -epochs N (default 2 as before)

cache (1 - wd*alpha) in 'adamw' opt struct -
no noticeable perf benefit, disabled (still done
for new SGD though)

since opt. memory is pre-allocated, the ggml_opt_get_optimizer_params
would probably be able to change between SGD and AdamW with each epoch
but would need to use adamw for the first (unconfirmed - no cmdline arg
to set such a policy yet)

test-opt checks adamw as before and now sgd (except for a few disabled
tests for sgd only; probably just needs logging values and adding
alternate reference values);  tolerance on the 'regression'
test is broader for sgd (so we don't need many more epochs)

* Vulkan: Implement GGML_OP_OPT_STEP_SGD

* tests: Fix OPT_STEP_SGD test-backend-ops

* SGD op param store weight-decay and not 1-alpha*wd

* minor + cosmetic changes

* fix vulkan sgd

* try CI fix

---------

Co-authored-by: 0cc4m <picard12@live.de>
Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
2025-08-14 12:03:57 +02:00

267 lines
7.8 KiB
CMake
Raw Blame History

cmake_minimum_required(VERSION 3.14) # for add_link_options and implicit target directories.
project("llama.cpp" C CXX)
include(CheckIncludeFileCXX)
#set(CMAKE_WARN_DEPRECATED YES)
set(CMAKE_WARN_UNUSED_CLI YES)
set(CMAKE_EXPORT_COMPILE_COMMANDS ON)
if (NOT XCODE AND NOT MSVC AND NOT CMAKE_BUILD_TYPE)
set(CMAKE_BUILD_TYPE Release CACHE STRING "Build type" FORCE)
set_property(CACHE CMAKE_BUILD_TYPE PROPERTY STRINGS "Debug" "Release" "MinSizeRel" "RelWithDebInfo")
endif()
message("CMAKE_BUILD_TYPE=${CMAKE_BUILD_TYPE}")
# Add path to modules
list(APPEND CMAKE_MODULE_PATH "${CMAKE_CURRENT_SOURCE_DIR}/cmake/")
set(CMAKE_RUNTIME_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/bin)
set(CMAKE_LIBRARY_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/bin)
if (CMAKE_SOURCE_DIR STREQUAL CMAKE_CURRENT_SOURCE_DIR)
set(LLAMA_STANDALONE ON)
include(git-vars)
# configure project version
# TODO
else()
set(LLAMA_STANDALONE OFF)
endif()
option(LLAMA_USE_SYSTEM_GGML "Use system libggml" OFF)
if (EMSCRIPTEN)
set(BUILD_SHARED_LIBS_DEFAULT OFF)
option(LLAMA_WASM_SINGLE_FILE "llama: embed WASM inside the generated llama.js" ON)
else()
if (MINGW)
set(BUILD_SHARED_LIBS_DEFAULT OFF)
else()
set(BUILD_SHARED_LIBS_DEFAULT ON)
endif()
endif()
option(BUILD_SHARED_LIBS "build shared libraries" ${BUILD_SHARED_LIBS_DEFAULT})
if (WIN32)
add_compile_definitions(_CRT_SECURE_NO_WARNINGS)
endif()
if (MSVC)
add_compile_options("$<$<COMPILE_LANGUAGE:C>:/utf-8>")
add_compile_options("$<$<COMPILE_LANGUAGE:CXX>:/utf-8>")
add_compile_options("$<$<COMPILE_LANGUAGE:C>:/bigobj>")
add_compile_options("$<$<COMPILE_LANGUAGE:CXX>:/bigobj>")
endif()
#
# option list
#
# debug
option(LLAMA_ALL_WARNINGS "llama: enable all compiler warnings" ON)
option(LLAMA_ALL_WARNINGS_3RD_PARTY "llama: enable all compiler warnings in 3rd party libs" OFF)
# build
option(LLAMA_FATAL_WARNINGS "llama: enable -Werror flag" OFF)
# sanitizers
option(LLAMA_SANITIZE_THREAD "llama: enable thread sanitizer" OFF)
option(LLAMA_SANITIZE_ADDRESS "llama: enable address sanitizer" OFF)
option(LLAMA_SANITIZE_UNDEFINED "llama: enable undefined sanitizer" OFF)
# utils
option(LLAMA_BUILD_COMMON "llama: build common utils library" ${LLAMA_STANDALONE})
# extra artifacts
option(LLAMA_BUILD_TESTS "llama: build tests" ${LLAMA_STANDALONE})
option(LLAMA_BUILD_TOOLS "llama: build tools" ${LLAMA_STANDALONE})
option(LLAMA_BUILD_EXAMPLES "llama: build examples" ${LLAMA_STANDALONE})
option(LLAMA_BUILD_SERVER "llama: build server example" ${LLAMA_STANDALONE})
# 3rd party libs
option(LLAMA_CURL "llama: use libcurl to download model from an URL" ON)
option(LLAMA_LLGUIDANCE "llama-common: include LLGuidance library for structured output in common utils" OFF)
# Required for relocatable CMake package
include(${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 ${LLAMA_FATAL_WARNINGS})
# change the default for these ggml options
if (NOT DEFINED GGML_LLAMAFILE)
set(GGML_LLAMAFILE_DEFAULT ON)
endif()
if (NOT DEFINED GGML_CUDA_GRAPHS)
set(GGML_CUDA_GRAPHS_DEFAULT ON)
endif()
# transition helpers
function (llama_option_depr TYPE OLD NEW)
if (${OLD})
message(${TYPE} "${OLD} is deprecated and will be removed in the future.\nUse ${NEW} instead\n")
set(${NEW} ON PARENT_SCOPE)
endif()
endfunction()
llama_option_depr(FATAL_ERROR LLAMA_CUBLAS GGML_CUDA)
llama_option_depr(WARNING LLAMA_CUDA GGML_CUDA)
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)
llama_option_depr(WARNING LLAMA_RPC GGML_RPC)
llama_option_depr(WARNING LLAMA_SYCL GGML_SYCL)
llama_option_depr(WARNING LLAMA_SYCL_F16 GGML_SYCL_F16)
llama_option_depr(WARNING LLAMA_CANN GGML_CANN)
if (NOT MSVC)
if (LLAMA_SANITIZE_THREAD)
message(STATUS "Using -fsanitize=thread")
add_compile_options(-fsanitize=thread)
link_libraries (-fsanitize=thread)
endif()
if (LLAMA_SANITIZE_ADDRESS)
message(STATUS "Using -fsanitize=address")
add_compile_options(-fsanitize=address -fno-omit-frame-pointer)
link_libraries (-fsanitize=address)
endif()
if (LLAMA_SANITIZE_UNDEFINED)
message(STATUS "Using -fsanitize=undefined")
add_compile_options(-fsanitize=undefined)
link_libraries (-fsanitize=undefined)
endif()
endif()
#
# 3rd-party
#
if (LLAMA_USE_SYSTEM_GGML)
message(STATUS "Using system-provided libggml, skipping ggml build")
find_package(ggml REQUIRED)
add_library(ggml ALIAS ggml::ggml)
endif()
if (NOT TARGET ggml AND NOT LLAMA_USE_SYSTEM_GGML)
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
#
add_subdirectory(src)
#
# utils, programs, examples and tests
#
if (NOT LLAMA_BUILD_COMMON)
message(STATUS "LLAMA_BUILD_COMMON is OFF, disabling LLAMA_CURL")
set(LLAMA_CURL OFF)
endif()
if (LLAMA_BUILD_COMMON)
add_subdirectory(common)
endif()
if (LLAMA_BUILD_COMMON AND LLAMA_BUILD_TESTS AND NOT CMAKE_JS_VERSION)
include(CTest)
add_subdirectory(tests)
endif()
if (LLAMA_BUILD_COMMON AND LLAMA_BUILD_EXAMPLES)
add_subdirectory(examples)
add_subdirectory(pocs)
endif()
if (LLAMA_BUILD_COMMON AND LLAMA_BUILD_TOOLS)
add_subdirectory(tools)
endif()
#
# install
#
include(GNUInstallDirs)
include(CMakePackageConfigHelpers)
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")
set(LLAMA_PUBLIC_HEADERS
${CMAKE_CURRENT_SOURCE_DIR}/include/llama.h
${CMAKE_CURRENT_SOURCE_DIR}/include/llama-cpp.h)
set_target_properties(llama
PROPERTIES
PUBLIC_HEADER "${LLAMA_PUBLIC_HEADERS}")
install(TARGETS llama LIBRARY PUBLIC_HEADER)
configure_package_config_file(
${CMAKE_CURRENT_SOURCE_DIR}/cmake/llama-config.cmake.in
${CMAKE_CURRENT_BINARY_DIR}/llama-config.cmake
INSTALL_DESTINATION ${CMAKE_INSTALL_LIBDIR}/cmake/llama
PATH_VARS LLAMA_INCLUDE_INSTALL_DIR
LLAMA_LIB_INSTALL_DIR
LLAMA_BIN_INSTALL_DIR )
write_basic_package_version_file(
${CMAKE_CURRENT_BINARY_DIR}/llama-version.cmake
VERSION ${LLAMA_INSTALL_VERSION}
COMPATIBILITY SameMajorVersion)
install(FILES ${CMAKE_CURRENT_BINARY_DIR}/llama-config.cmake
${CMAKE_CURRENT_BINARY_DIR}/llama-version.cmake
DESTINATION ${CMAKE_INSTALL_LIBDIR}/cmake/llama)
install(
FILES convert_hf_to_gguf.py
PERMISSIONS
OWNER_READ
OWNER_WRITE
OWNER_EXECUTE
GROUP_READ
GROUP_EXECUTE
WORLD_READ
WORLD_EXECUTE
DESTINATION ${CMAKE_INSTALL_BINDIR})
configure_file(cmake/llama.pc.in
"${CMAKE_CURRENT_BINARY_DIR}/llama.pc"
@ONLY)
install(FILES "${CMAKE_CURRENT_BINARY_DIR}/llama.pc"
DESTINATION ${CMAKE_INSTALL_LIBDIR}/pkgconfig)
Reference in New Issue View Git Blame Copy Permalink